text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Declarative vs Rule-based Control for Flocking Dynamics
Usama Mehmood
Department of Computer Science,
Stony Brook University, USA
Radu Grosu
arXiv:1710.10013v1 [cs.MA] 27 Oct 2017
Cyber-Physical Systems Group,
Technische Universitat Wien, Austria
Ashish Tiwari
SRI International, USA
Nicola Paoletti
Department of Computer Science,
Stony Brook University, USA
Shan Lin
Department of Electrical and
Computer Engineering, Stony Brook
University, USA
Junxing Yang
Department of Computer Science,
Stony Brook University, USA
ABSTRACT
The popularity of rule-based flocking models, such as Reynolds’
classic flocking model, raises the question of whether more declarative flocking models are possible. This question is motivated by
the observation that declarative models are generally simpler and
easier to design, understand, and analyze than operational models.
We introduce a very simple control law for flocking based on a
cost function capturing cohesion (agents want to stay together) and
separation (agents do not want to get too close). We refer to it as
declarative flocking (DF). We use model-predictive control (MPC)
to define controllers for DF in centralized and distributed settings. A
thorough performance comparison of our declarative flocking with
Reynolds’ classic flocking model, and with more recent flocking
models that use MPC with a cost function based on lattice structures, demonstrate that DF-MPC yields the best cohesion and least
fragmentation, and maintains a surprisingly good level of geometric regularity while still producing natural flock shapes similar to
those produced by Reynolds’ model. We also show that DF-MPC
has high resilience to sensor noise.
ACM Reference Format:
Usama Mehmood, Nicola Paoletti, Dung Phan, Radu Grosu, Shan Lin, Scott
D. Stoller, Ashish Tiwari, Junxing Yang, and Scott A. Smolka. 2018. Declarative vs Rule-based Control for Flocking Dynamics. In Proceedings of ACM/SIGAPP
Symposium On Applied Computing (SAC 2018). ACM, New York, NY, USA,
8 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn
1
INTRODUCTION
Flocking is a collective behavior exhibited by a large number of
interacting agents possessing a common group objective [7]. The
term is most commonly associated with birds, and more recently,
drones. Examples include foraging for food, executing a predatoravoidance maneuver, and engaging in migratory behavior.
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
SAC 2018, April 9-13, 2018, Pau, France
© 2018 Copyright held by the owner/author(s).
ACM ISBN 978-x-xxxx-xxxx-x/YY/MM.
https://doi.org/10.1145/nnnnnnn.nnnnnnn
Dung Phan
Department of Computer Science,
Stony Brook University, USA
Scott D. Stoller
Department of Computer Science,
Stony Brook University, USA
Scott A. Smolka
Department of Computer Science,
Stony Brook University, USA
With the introduction of Reynolds’ model [12, 13], rule-based
control became the norm in the flocking community. Specifically,
in this model, at each time-step, each agent executes a control law
given in terms of the weighted sum of three competing forces to
determine its next acceleration. Each of these forces has its own rule:
separation (keep a safe distance away from your neighbors), cohesion
(move towards the centroid of your neighbors), and alignment (steer
toward the average heading of your neighbors). As the descriptions
suggest, these rules are executed by each agent in a distributed
environment with limited-range sensing and no communication.
The popularity of Reynolds’ model and its many variants raises
the question: Is there a more abstract declarative form of control for
flocking? This question is important because declarative models
are generally simpler and easier to design, understand, and analyze
than operational models. This is analogous to declarative programs
(e.g., functional programs and logic programs) being easier to write
and verify than imperative programs.
We show that the answer to this question is indeed positive by
providing a very simple control law for flocking based on a cost
function comprising two main terms: cohesion (the average squared
distance between all pairs of agents) and separation (a sum of inverse squared distances, except this time between pairs of agents
within each other’s sensing range). That is it. For example, no term
representing velocity alignment is needed. The cost function specifies what we want as the goal, and is hence declarative. In contrast,
the update rules in Reynolds’ model aim to achieve an implicit goal
and hence are operational. Executing declarative control amounts
to finding the right balance between attracting and repelling forces
between agents. We refer to this approach as Declarative Flocking
(DF). We use MPC (model-predictive control) to define controllers
for DF, and refer to this approach as DF-MPC. We define a centralized version of DF-MPC, which requires communication, and a
distributed version, which does not.
Previous MPCs for flocking exist, e.g., [16–18]. Most of these
MPCs are designed to conform to the α-lattice model of flocking
proposed in [7]. α-lattices impose a highly regular structure on
flocks: all neighboring agents are distance d apart, for a specified
constant d. This kind of structure is seen in some settings, such as
beehives, but is not expected in many other natural and engineered
settings, and it is not imposed by Reynolds’ model.
SAC 2018, April 9-13, 2018, Pau, France
U. Mehmood et al.
In this paper, we show, via a thorough performance evaluation,
how centralized and distributed DF-MPC compare to Reynolds’ rulebased approach [12, 13], Olfati-Saber’s potential-based approach
[7], a variant of Zhan and Li’s centralized lattice-based MPC approach [15, 16], and Zhang et al.’s distributed lattice-based MPC
approach [17]. We consider performance measures that capture multiple dimensions of flocking behavior: number of sub-flocks (flock
fragmentation), maximum sub-flock diameter (cohesion), velocity
convergence, and a new parameter-free measure of the geometric
regularity of the formation.
Our experimental results demonstrate that DF-MPC yields the
best cohesion and least fragmentation, and produces natural flock
shapes like those produced by Reynolds’ model. Also, distributed
DF-MPC maintains a surprisingly good level of geometric regularity.
We also analyze the resiliency of DF-MPC and the lattice-based
MPC approaches by considering the impact of sensor noise. Our
results demonstrate a remarkably high level of resiliency on the
part of DF-MPC in comparison with these other approaches.
The rest of the paper is organized as follows. Section 2 presents
the rule-based, potential-based, and lattice-based MPC approaches
mentioned above. Section 3 defines our declarative flocking approach. Section 4 defines our performance measures for flocking
models. Section 5 presents our experimental results and performance evaluation. Section 6 discusses related work. Finally, Section 7 offers our concluding remarks and directions for future work.
2
MODELS OF FLOCKING BEHAVIOR
We consider a set of dynamic agents B = {1, . . . , n} that move
according to the following discrete-time equation of motion:
x i (k + 1) = x i (k) + dT · vi (k), vi (k) ∈ V
vi (k + 1) = vi (k) + dT · ai (k), ai (k) ∈ A,
(1)
(2)
where x i (k), vi (k), ai (k) ∈ Rm are respectively position, velocity
and acceleration of agent i ∈ B in the m-dimensional space at
step k, and dT ∈ R+ is the time step. We consider physical constraints on velocities and accelerations, described by the sets V
and A, respectively, which are defined by V = {v | |v | ≤ v̄} and
A = {a | |a| ≤ ā}, where v̄ and ā limit the allowed magnitude of
the velocity and acceleration vectors, respectively.
In most flocking models, agents update their motion by changing
their acceleration. In this sense, ai (k) represents the control input
for agent i.
The configuration of all agents is described by the vector x(k) =
[x 1T (k) . . . x nT (k)]T ∈ Rm ·n . Let v(k) = [vT1 (k) . . . vnT (k)]T ∈
Rm ·n , and a(k) = [aT1 (k) . . . aTn (k)]T ∈ Rm ·n . Then the equation
of motion for all agents can be expressed as
x(k + 1) = x(k) + dT · v(k),
v(k + 1) = v(k) + dT · a(k),
(3)
(4)
The local neighborhood of agent i is defined by the set of other
agents, called neighbors, within a given distance from i, mimicking
the agent’s visibility sphere. For an interaction radius r > 0 and
configuration x, the set of spatial neighbors of agent i, Ni (x) ⊆ B,
is given by:
Ni (x) = j ∈ B | j , i ∧ ∥x i − x j ∥ < r ,
(5)
where ∥ · ∥ denotes the Euclidean norm.
a)
b)
Figure 1: Examples of α-lattice a) and quasi α-lattice b). Solid
lines connect agents in the same neighborhood that have distance d. Dashed lines connect those with have distance d ± ϵ
for ϵ ≤ δ (the tolerance).
For configuration x ∈ Rm ·n , we define the associated proximity
net G(x) = (B, E(x)) as the graph that connects agents within their
interaction radius:
E(x) = (i, j) ∈ B × B | ∥x i − x j ∥ < r , i , j ,
(6)
To capture the regular geometry of flocks, Olfati-Saber introduced the notions of α-lattices, i.e. configurations where each agent
is equally distant from its neighbors, and quasi α-lattices, i.e. configurations that are α-lattices modulo a small error in the distances
[7]. The scale parameter d defines the ideal inter-agent distance.
Definition 2.1 (α-lattice [7]). A configuration x ∈ Rm ·n is called
α-lattice if for all i ∈ B and all j ∈ Ni (x), ∥x i − x j ∥ = d, where d ∈
R+ is the scale of the α-lattice. For tolerance δ ∈ R+ , a configuration
x ∈ Rm ·n is called a quasi α-lattice if for all i ∈ B and all j ∈ Ni (x),
|∥x i − x j ∥ − d | ≤ δ .
2.1
Sensing noise
We extend the classical equations of motion, Eqs. (1)–(2), with sensing noise affecting how each agent perceives positions and velocities
of its neighbors. Existing work has put little focus on flocking dynamics subject to noise, which is unfortunately unavoidable in
realistic natural and engineered flocks.
For actual positions x(k) and velocities v(k) at step k, let x̃(k)
and ṽ(k) denote their noisy counterparts sensed by a generic agent,
defined by:
x̃(k) = x(k) + nx(k) and ṽ(k) = v(k) + nv(k),
(7)
where nx(k) and nv(k) in
are vectors of independent and
identically distributed (i.i.d.) random variables. The position noise
nx(k) and velocity noise nv(k) are distributed according to Gaussian distributions with mean 0 and standard deviation σx and σv ,
respectively. We stress the dependency on k because noise variables
are independent across time steps.
In centralized flocking algorithms, where agent decisions are
computed by a single controller with information about the whole
population, we use Eq. 7 to define noisy measurements. In distributed algorithms, sensing noise is independent for each agent.
We denote the noisy measurement of agent i by x̃▷i (k) and ṽ▷i (k),
where positions and velocities are noisy for all agents except agent
i
Rm ·n
x̃▷i (k) =[x̃ 1T (k) . . . x iT (k) . . . x̃ nT (k)]T and
ṽ▷i (k) =[ṽT1 (k) . . . viT (k) . . . ṽnT (k)]T ,
(8)
(9)
Declarative vs Rule-based Control for Flocking Dynamics
SAC 2018, April 9-13, 2018, Pau, France
with x̃ 1 (k), . . . , x̃ n (k) and ṽ 1 (k), . . . , ṽn (k) defined as per (7); implicitly, for each agent i and each other agent j, the noise distribution is
sampled independently to compute the x̃ Tj (k) component of x̃▷i (k).
2.2 Reynolds’ rule-based model
In Reynolds’ rule-based distributed model [12, 13], agents follow
simple rules to compute their accelerations from the positions and
velocities of their neighbors. The rules are illustrated in Figure 2.
They do not explicitly specify the desired flocking formation as an
objective; rather, flocking emerges from the interaction rules.
Specifically, each agent i ∈ B updates its acceleration ai (k) at
step k by considering the following three components (adapted to
include sensing noise):
(1) Alignment: agents match their velocities with the average
velocity of nearby agents.
Õ
©©
ª
ª
= w al ·
·
ṽ j (k)® − vi (k)® (10)
|Ni (x̃▷i (k))|
▷i
j ∈N i (x̃ (k ))
««
¬
¬
(2) Cohesion: agents move towards the centroid of the agents in
the local neighborhood.
1
aial (k)
Õ
1
©©
ª
ª
aci (k) = wc ·
·
x̃ j (k)® − x i (k)® (11)
▷i
|Ni (x̃ (k))|
j ∈N i (x̃▷i (k ))
««
¬
¬
(3) Separation: agents move away from nearby neighbors.
Õ
x i (k) − x̃ j (k) ª
©
·
® (12)
|Ni
∥x i (k) − x̃ j (k)∥ 2
«j ∈Ni (x̃▷i (k ))
¬
The cohesion and alignment rules help form and maintain a closely
packed, flock-like formation. The separation rule prevents agents
from coming too close to each other, thus reducing crowding and
collisions.
Non-negative constants w al , wc and w s are the weights for each
acceleration component. Typically, a smaller interaction radius
(hence a smaller neighborhood) is used for the separation rule,
because it is significant only when agents are very close to each
other. The overall acceleration in Reynolds’ model is given by:
asi (k) = w s ·
1
2.4
MPC-based models
Model predictive control (MPC) [2] is a well-established control
technique that works as follows: at each time step k, it computes
the optimal control sequence (agents’ accelerations in our case)
that minimizes a given cost function with respect to a predictive
model of the controlled system and a finite prediction horizon of
length T , i.e., from step k + 1 to k + T . Then, the first control input
of the optimal sequence is applied (the remainder of the sequence
is unused), and the algorithm proceeds with a new iteration.
Two main kinds of MPC-based flocking models exist, centralized
and distributed. Centralized models assume that information about
positions and velocities of all agents is available to compute their
optimal accelerations. Formally, at each time step k, it solves the
following optimization problem:
min
a(k |k ), ...,a(k +T −1|k )∈A
(13)
2.3 Olfati-Saber’s potential-based model
In potential-based flocking models, the interaction between a pair
of agents is modeled by a potential field. It is assumed that an agent
is a point source, and it has a potential field around it, which exerts a force, equal to its gradient, on other agents in its range of
influence. The potential field has circular symmetry and hence is a
function of distance from the source. In the work of Olfati-Saber [7],
the potential function ψ α for a pair of agents has its minimum at
the desired inter-agent distance d of the desired α-lattice. Outside
the interaction radius r , the potential function is constant, so the
potential field exerts no force. The exact definition of ψ α is complicated: it is the definite integral of an “action function” ϕ α that
is the product of a “bump function” ρh and an uneven sigmoidal
function ϕ. The control law computes an agent’s acceleration based
on the sum of the forces from all other agents in its neighborhood
and a velocity alignment term.
TÕ
−1
t =0
∥a(k + t | k)∥ 2
(14)
where a(k + t | k) is the control input (accelerations) for all agents
at predicted time step k + t starting from step k. The first term J (k)
is the primary model-specific cost function that the controller seeks
to optimize within the prediction horizon; it is implicitly a function
of the predicted configurations during the prediction horizon for
time step k. The second term is standard for MPC problems and
penalizes large control inputs, with weight λ > 0.
In distributed flocking models, each agent computes its optimal
acceleration based only on information about its neighbors. Each
agent i solves an optimization problem of the form:
(x̃▷i (k))|
ai (k) = aial (k) + aci (k) + asi (k).
J (k) + λ ·
min
a i (k |k ), ...,a i (k +T −1|k )∈A
Ji (k) + λ ·
TÕ
−1
t =0
∥ai (k + t | k)∥ 2
(15)
where ai (k + t | k) is the acceleration for agent i at predicted time
step k + t starting from step k, and Ji (k) is the model-specific cost
function for agent i. In distributed MPC, an agent has no way to
know current or future control decisions of its neighbors, which
are needed to make accurate predictions about their behavior. To
address this problem, some approaches allow agents to communicate their local control decisions or future positions (e.g. [16, 18]),
or assume that neighbors follow some default motion law, e.g.,
they move with constant velocities. We adopt the second strategy,
because it does not require any communication.
The majority of existing MPC-based approaches to flocking are
designed to optimize the regularity of the flock, by penalizing configurations where neighboring agents are not exactly distance d
apart, i.e., configurations that differ from an α-lattice [15–18]. We
call these approaches lattice-based MPC. Next we describe representative centralized and distributed lattice-based MPC flocking
models, which we extend to account for sensing noise. The centralized model is a variant of a model by Zhan and Li [15, 16]. The
distributed model is by Zhang et al. [17].
2.4.1 Centralized lattice-based MPC flocking. The centralized
lattice-based MPC problem is defined as:
min
a(k |k ), ...,a(k+T −1|k )∈A
T
Õ
t =1
TÕ
−1
∥д (x(k + t | k)) ∥ 2 +λ·
t =0
∥a(k+t | k)∥ 2
(16)
SAC 2018, April 9-13, 2018, Pau, France
U. Mehmood et al.
(a) Alignment
(b) Cohesion
(c) Separation
Figure 2: Interaction rules for flocking behavior in Reynolds Model.
where x(k + t | k) is the configuration of the system at predicted
time step k + t starting from step k, following the dynamics:
x i (k | k) = x̃ i (k) vi (k | k) = ṽi (k)
x i (k + t + 1 | k) = x i (k + t | k) + dT · vi (k + t | k)
vi (k + t + 1 | k) = vi (k + t | k) + dT · ai (k + t | k),
where the initial state of the prediction window is given by noisy
measurements. For configuration x, д(x) captures the α-lattice irregularity as the total deviation between agent distances and d:
Õ
d · x ji 2
∥д(x)∥ 2 =
x ji −
, with x ji = x j − x i . (17)
∥x ji ∥
(i, j)∈ E(x)
This model is inspired by [15] and [16] but differs from both:
in [15] the cost function also contains a velocity alignment term,
which the same authors removed in their subsequent work, while
in [16], “impulsive MPC” is used, which means that agents directly
control their velocities (instead of accelerations), an abstraction
that allows physically unrealizable accelerations.
2.4.2 Distributed lattice-based MPC flocking. In the distributed
MPC flocking model of Zhang et al. [17], each agent i controls its
acceleration based on position and velocity measurements of the
neighbors and assumes they have constant velocity (zero acceleration) during the prediction horizon. Similarly, the set of neighbors
of i is assumed invariant during the prediction horizon, and we
denote it by Ni (k) = Ni (x(k)). The control law for agent i is:
min
a i (k |k ), ...,a i (k +T −1|k )∈A
T
Õ
t =1
2
∥дi (x(k + t | k)) ∥ +
λ·
TÕ
−1
t =0
∥ai (k + t | k)∥ 2 .
(18)
where the predicted future dynamics of i is determined by:
x i (k | k) = x̃ i (k) vi (k | k) = ṽi (k)
x i (k + t + 1 | k) = x i (k + t | k) + dT · vi (k + t | k)
x i (k + t + 1 | k) = x i (k + t | k) + dT · ai (k + t | k),
while i’s neighbors j ∈ Ni (k) have constant velocity:
x j (k | k) = x̃ j (k) x j (k + t + 1 | k) = x j (k + t | k) + dT · ṽ j (k).
For configuration x, дi (x) is defined in a similar way to Eq. (17)
and quantifies how much i’s neighborhood Ni (k) deviates from an
α-lattice:
Õ
d · x ji 2
∥дi (x)∥ 2 =
x ji −
.
(19)
∥x ji ∥
j ∈N i (k )
3
DECLARATIVE FLOCKING
This section introduces centralized and distributed versions of our
Declarative Flocking (DF) model, and presents a flocking algorithm
based on MPC. Our formulation is declarative in that it consists of
just two simple terms: (1) a cohesion term based on the average
squared distance between pairs of agents, to keep the flock together,
and (2) a separation term based on the inverse squared distances
between pairs of agents, to avoid crowding. These two terms represent opposing forces on agents, causing agents to move towards
positions in which these forces are balanced. Unlike the majority
of existing MPC-based approaches that are designed to optimize
conformance to an α-lattice, our design does not impose a specific
geometric structure.
3.1
Centralized DF model
The cost function J for our centralized DF model contains the two
terms described above, with the cohesion term considering all pairs
of agents, and the separation term considering only pairs of agents
that are neighbors. The weight ω of the separation term provides
control over the density of the flock.
Õ
Õ Õ
1
2
∥x i j ∥ 2 + ω ·
J C (x) =
·
2
|B| · (|B| − 1)
∥x
ij ∥
i ∈B j ∈B,i <j
The control law is Eq. (14) with J (k) equal to
3.2
Distributed DF model
ÍT
(i, j)∈E(x)
t =1
J C (x(k + t | k)).
The cost function J for our distributed DF model is similar to the
centralized one, except that both terms are limited to consider pairs
of agents that are neighbors.
Õ
Õ
1
1
·
∥x i j ∥ 2 + ω ·
(20)
JiD (x) =
|Ni (k)|
∥x i j ∥ 2
j ∈N i (k )
j ∈N i (k )
The control law for agent i is Eq. (15) with Ji (k) equal to
ÍT
D
t =1 Ji (x(k + t | k)).
4
MEASURES OF FLOCKING PERFORMANCE
We introduce four key measures of flocking performance. A single
measure is insufficient, because flocking is indeed characterized by
multiple desirable properties, such as aligned velocities and cohesion. Olfati-Saber introduces four main properties for flocking [7],
informally described as:
(1) the group of agents stays connected in a unique flock, i.e., no
sub-flocks and fragmentation should emerge;
(2) the group remains cohesive, in a close-knit formation;
Declarative vs Rule-based Control for Flocking Dynamics
SAC 2018, April 9-13, 2018, Pau, France
(3) the group moves in a coherent way as if it was a unique body,
i.e., agents’ velocities are aligned; and
(4) the group maintains a regular geometry (in the α-lattice
sense).
We introduce the following four measures to capture these four
requirements. An important concept in these definitions is a subflock, which is a set of interacting agents that is too far apart from
other agents to interact with them. Formally, a sub-flock in a configuration x corresponds to a connected component of the proximity
net G(x). Let CC(x) ⊆ 2 B be the set of connected components of
the proximity net G(x).
(1) The number of connected components of the proximity net
quantifies connectedness—or, equivalently, fragmentation—of the
flock. There is no fragmentation when |CC(x)| = 1. Fragmentation
exists when |CC(x)| > 1. Fragmentation may be temporary or, if
sub-flocks move in different directions, permanent.
(2) The maximum component diameter, denoted D(x), quantifies
cohesion. It is defined by
D(x) =
max
B ′ ∈CC(x)
(21)
D(x, B ′ )
where D(x, B ′ ) is the diameter of connected component B ′ :
max
D(x, B ′ ) =
(i, j)∈B ′ ×B ′
i,j
∥xi j ∥.
(22)
Ð
Note that when all agents are isolated, i.e., CC(x) = i ∈B {{i}},
D(x) = −∞ because the domain of the max function in Equation 22
is empty when B ′ is a singleton. Note that we consider the maximum diameter of a sub-flock in order to make this measure more
independent of connectedness. If we instead considered the overall
diameter of the entire (possibly fragmented) flock, any flocking
model that did poorly on connectedness would also do very poorly
on this measure.
(3) The velocity convergence measure, adopted from [17], quantifies the average discrepancy between each agent’s velocity and the
average velocity of the flock. In particular, we extend the measure
of [17] to average velocity convergence values across sub-flocks:
VC(x, v) =
Í
B ′ ∈CC(x)
Í
i ∈B ′ vi
−
Í
|CC(x)|
j ∈B ′ v j
| B′ |
2
|B ′ |
(23)
(4) To measure the regularity of the geometric structure of a flock,
as reflected in the inter-agent spacing, we introduce a parameterfree and model-independent irregularity measure I (x). For a connected component (sub-flock) B ′ , it is defined as the sample standard deviation of the distances between each agent in B ′ and its
closest neighbor. Thus, the measure penalizes configurations where
there is dispersion in inter-agent distances, while not imposing any
fixed distance between them (unlike α-lattices).
Ð
Let CC ′ (x) = CC(x) \ i ∈B {{i}} be the set of connected components where isolated agents are excluded. For |CC ′ (x)| = 0 (or
equivalently, |CC(x)| = |B|), i.e., all agents are isolated, we set the
irregularity I (x) = 0, which is the optimal value. This reflects the
fact that a single point is a regular structure on its own. Moreover,
such a configuration is already highly penalized by |CC(x)| and
VC(v). For |CC ′ (x)| > 0, the measure is defined by:
Í
Ò
B ′ ∈CC ′ σ
i ∈B ′ min j,i ∥x i j ∥
.
(24)
I (x) =
|CC ′ |
where σ (S) is the standard deviation of the multiset of samples S
Ò
and is the sum operator (or disjoint union) for multisets.
An α-lattice (see Def. 2.1) has the optimal value of I (x), i.e.,
I (x) = 0, since all neighboring agents are located at the same
distance d from each other, leading to zero standard deviation for the
term σ ({d, d, . . . , d }). This shows that I (x) captures the regularity
underlying the concept of α-lattice.
We introduce this measure because previous measures of regularity or irregularity, such as those in [7, 16, 17], measure deviations
from an α-lattice with a specified inter-agent distance d and are
therefore inapplicable to flocking models, such as Reynolds’ model
and our DF models, that are not based on α-lattices and do not
have a specified target inter-agent distance. Also, our irregularity
measure is more flexible than those based on α-lattices, because it
gives an optimal score to some configurations that are geometrically
regular but not α-lattices. For example, consider a configuration x
in which the agents are on the vertices of a grid with edge length e,
and the interaction radius is equal to the length of the diagonal of
a box in the grid. This configuration has an optimal value for our
irregularity measure, i.e., I (x) = 0, because the distance from every
agent to its nearest neighbor is e. This configuration is not an αlattice and hence does not nave an optimal value for the irregularity
measures used in prior work.
5
PERFORMANCE EVALUATION
We compare the performance of the models of Section 2 with the
newly introduced DF flocking models in the 2-dimensional setting.
In the first set of experiments (Section 5.1), we evaluate the performance measures illustrated in Section 4. In the second set of
experiments (Section 5.2), we analyze the resilience of the algorithms to sensor noise.
For consistency with the experimental settings of [17], the latticebased MPC problems are solved using the interior point method implemented in MATLAB’s fmincon function. Our DF-MPC problems
are solved using gradient descent optimization. Unless otherwise
specified, the population size is n = 30, the simulation length is
100, dT = 0.3, v̄ = 8, ā = 1, r = 8.4, d = 7, T = 3, and λ = 1.
These parameter values are the same ones reported in [17]. Following the settings in the OpenSteer project [11], the parameters for
Reynolds’ model are rc = 9, r s = 5, r al = 7.5, wc = 8, w s = 12,
and w al = 8. The weight of the separation term in our centralized
and distributed DF-MPC is ω = 50. As in [17], initial positions and
initial velocities of agents are uniformly sampled from [−15, 15]2
and [0, 2]2 , respectively.
5.1
Performance Comparison of Flocking
Algorithms
Fig. 3 shows examples of final formations for all flocking models.
In particular, we chose configurations where fragmentation did
not occur. We observe that the formations for lattice-based MPC
algorithms have spread-out, rigid structures, consistent with the
SAC 2018, April 9-13, 2018, Pau, France
design objective of maximizing the α-lattice regularity. On the other
hand, Reynolds and our DF MPC models result in more natural
flock shapes.
In Fig. 4, we compare the performance measures averaged over
100 runs for each flocking model. Regarding the number of connected components (sub-flocks), our centralized DF-MPC registers
the best behavior, rapidly stabilizing to an average of 1 component
(see plot a). Our distributed DF-MPC and Reynolds’ model have
comparable performance, reaching an average number of sub-flocks
below 1.4. The lattice-based MPCs and Olfati-Saber instead lead
to constant fragmentation, with more than 2 sub-flocks for the
distributed lattice-based MPC, 6 for the centralized lattice-based
MPC, and more than 8 for Olfati-Saber’s model.
This ranking is confirmed by the diameter measure (plot b),
where our centralized and distributed DF-MPC and Reynolds’ model
show the best cohesion, outperforming the lattice-based approaches.
Recall that this measure indicates the maximum diameter over all
sub-flocks, not the diameter of the entire population. As a consequence, fragmentation tends to improve diameter values since it
produces sub-flocks with fewer individuals. This explains why our
distributed DF-MPC performs better on this measure than the centralized version, and similarly why Olfati-Saber’s model has smaller
diameter measure than centralized lattice-based MPC, which in turn
has smaller diameter measure than the distributed variant.
As expected, Olfati-Saber’s model and the lattice-based MPCs
have very good performance for irregularity (plot c), since they are
designed to achieve the regular geometric formation of α-lattice.
Surprisingly, our distributed DF-MPC performs almost as well as
them on this measure. Centralized DF-MPC and Reynolds’ model
have the least regular formations.
For velocity convergence (plot d), we find that all models perform
comparably well and are able to achieve flocks with consistent
velocities fairly quickly after an initial spike.
5.2
Robustness to Sensing Noise
To evaluate the resiliency of the models to sensor noise, we performed 20 runs for each model at 10 noise levels. The noise levels
are numbered from 1 to 10, and noise level i has σx = 0.2i and
σv = 0.1i. For each performance metric, we averaged its final values over 20 runs for each noise level. The results are plotted in Fig. 5.
Of the six models, Olfati-Saber’s model is the most vulnerable to
sensing noise: the number of sub-flocks |CC | in Olfati-Saber’s model
quickly increases to nearly 30, rendering other metrics irrelevant.
The lattice-based MPC models also exhibit high fragmentation, leading to nominally good but largely irrelevant values for the other
performance metrics. Our distributed DF-MPC and Reynolds’ model
have the best resiliency to sensing noise, with both models exhibiting similar profiles in all metrics. While the irregularity and velocity
convergence measures increase with noise level, as expected, both
models remarkably maintain almost a single connected component
with a nearly constant component diameter for all 10 noise levels,
with DF-MPC achieving a smaller diameter than Reynolds’ model.
6
RELATED WORK
Reynolds [12] introduced the first rule-based approach for simulation of flocking behavior. With only three simple rules, his model is
U. Mehmood et al.
able to capture complex flocking behaviors of animals. Additional
rules can be added to the model to simulate specific behaviors, such
as leader following and predator avoidance. Pearce et al. [8] present
a rule-based strategy for flocking, where agents move to maximize
their view out of the flock. Cucker and Dong [3] present a rulebased flocking approach with proofs of convergence and collision
avoidance.
Cucker and Smale [4] introduced another popular rule-based
flocking model. The Cucker-Smale model is parameterized by a
constant β such that if β < 1/2, velocity convergence is guaranteed.
If β ≥ 1/2, velocity convergence can also be achieved under some
conditions on the initial positions and initial velocities of the agents.
Ahn and Ha [1] investigated the effects of multiplicative noise on
the long term dynamics of the Cucker-Smale model. Erban et al. [5]
extend the Cucker-Smale model to take into account stochasticity
(imperfections) of agent behavior and delay in agents’ responses to
changes in their environment.
Flocking models based on potential fields have been proposed
in several papers. Tanner et al. [14] propose a potential function
Ui j , given in Equation 25, where ||r i j || 2 is the distance between
agents i and j. For distances greater than R, the potential is set to a
constant value, C, indicating a zero force. In their control law, the
acceleration of agent i is based on the sum over all neighbors j of
the gradient of the potential function Ui j .
(
1
+ loд||r i j || 2 , ||r i j || 2 < R
| |r i j | | 2
(25)
Ui j =
C,
||r i j || 2 ≥ R
A similar potential function is also proposed by [10]. Furthermore,
potential-based solutions have been extended with additional behaviors such as obstacle avoidance and leader following. For example,
Ogren et.al. [9] use the motion of the leader to guide the motion of
the flock; the leader’s motion is independent, i.e., is not influenced
by other agents.
La and Sheng [6] propose an extension of Olfati-Saber’s model
designed for noisy environments. In addition to the terms found in
Olfati-Saber’s model, their control law contains feedback terms for
position and velocity, to make agents tend to stay close to the centroid of their neighborhood and minimizing the velocity mismatch
with their neighbors. They show that adding these feedback terms
to the control law helps bound the error dynamics of the system.
7
CONCLUSIONS
This paper presents an abstract declarative form of control for
flocking behavior and the results of a thorough comparison of
centralized and distributed versions of our MPC-based declarative
flocking with four other flocking models. Our simulation results
demonstrate that DF-MPC yields the best cohesion and least fragmentation, and produces natural flock shapes like those produced
by Reynolds’ rule-based model. Our resiliency analysis shows that
the distributed version of our DF-MPC is highly robust to sensor
noise.
As future work, we plan to study resilience of the flocking models
with respect to additional noisy scenarios such as actuation noise
(i.e., noise affecting acceleration) and faulty agents with deviant
behavior. We also plan to investigate smoothing techniques to
increase resilience to sensor noise.
Declarative vs Rule-based Control for Flocking Dynamics
75
SAC 2018, April 9-13, 2018, Pau, France
135
140
80
70
125
65
120
125
120
120
110
20
55
50
190
200
210
105
0
20
40
0
110
90
-20
(a) Reynolds
20
110
115
100
0
40
115
40
60
60
130
130
60
100
60
100
120
140
(b) Lattice-based central- (c)
Lattice-based
ized MPC
tributed MPC
dis-
90
100
-20
110
80
(d) DF centralized MPC
90
100
0
(e) DF distributed MPC
20
40
60
80
(f) Olfati-Saber
Figure 3: Examples of final formations for different flocking models. The red dots are the agent positions. The blue lines
denote the agent velocities; the line lengths are proportional to the speeds.
Lattice Distributed
Reynolds
DF Centralized
60
10
20
40
3
2
4
10
80
1
0.5
20
60
1.5
1
I
D
0
O-S
4
30
6
0
Lattice Centralized
40
5
|CC|
|CC|
50
8
DF Distributed
2
VC
10
0
0
0
20
40
Time
60
80
0
20
40
Time
2
(a) Number of connected components
|CC |
0
0
10
30
0
80
20
40
(c) Irregularity I
40
50
60
60
80
Time
Time
(b) Max component diameter D
20
60
(d) Velocity convergence V C
70
80
90
Figure 4: Comparison of performance measures obtained
with 100 runs for each flocking algorithm.
Time
Lattice Distributed
10
30
DF Centralized
DF Distributed
Lattice Centralized
O-S
1.5
1.5
1
1
40
8
30
I
VC
20
D
|CC|
Reynolds
50
20
|CC|
10
0
2
4
6
0.5
0.5
6
10
4
0
8
10
Noise Level
2
(a) Number of connected components
|CC |
0
0
10
0
0
2
4
6
8
10
2
4
Noise Level
30
8
2
10
40
(c) Irregularity I
50
60
70
4
6
8
10
Noise Level
Noise Level
(b) Max component diameter D
20
6
(d) Velocity convergence V C
80
90
Figure 5: Comparison of the final values of the performance measures
obtained with 20 runs for each flocking algorithm and
Time
for each noise level.
REFERENCES
[1] Shin Mi Ahn and Seung-Yeal Ha. 2010.
Stochastic flocking dynamics of the CuckerâĂŞSmale model with multiplicative white noises. J.
Math. Phys. 51, 10 (2010), 103301.
https://doi.org/10.1063/1.3496895
arXiv:http://dx.doi.org/10.1063/1.3496895
[2] E.F Camacho and C. Bordons. 2007. Model predictive control. Springer.
[3] Felipe Cucker and Jiu-Gang Dong. 2011. A general collision-avoiding flocking
framework. IEEE Trans. Automat. Control 56, 5 (2011), 1124–1129.
[4] F. Cucker and S. Smale. 2007. Emergent Behavior in Flocks. IEEE Trans. Automat.
Control 52, 5 (May 2007), 852–862. https://doi.org/10.1109/TAC.2007.895842
[5] Radek Erban, Jan Hasźkovec, and Yongzheng Sun. 2016. A Cucker–Smale Model
with Noise and Delay. SIAM J. Appl. Math. 76, 4 (July 2016), 1535–1557. https:
//doi.org/10.1137/15M1030467
[6] H. M. La and W. Sheng. 2010. Flocking control of multiple agents in noisy
environments. In 2010 IEEE International Conference on Robotics and Automation.
4964–4969. https://doi.org/10.1109/ROBOT.2010.5509668
[7] Reza Olfati-Saber. 2006. Flocking for multi-agent dynamic systems: Algorithms
and theory. IEEE Transactions on automatic control 51, 3 (2006), 401–420.
[8] Daniel J. G. Pearce, Adam M. Miller, George Rowlands, and Matthew S. Turner.
2014. Role of projection in the control of bird flocks. Proceedings of the National
Academy of Sciences 111, 29 (2014), 10422–10426. https://doi.org/10.1073/pnas.
1402202111 arXiv:http://www.pnas.org/content/111/29/10422.full.pdf
[9] Naomi Ehrich Leonard Peter Ogren. 2004. Cooperative control of mobile sensor networks:Adaptive gradient climbing in a distributed environment. IEEE
transactions on Automatic Control 49, 8 (2004).
[10] John H. Reif and Hongyan Wang. 1999. Social potential fields: A distributed
behavioral control for autonomous robots. Robotics and Autonomous Systems 27,
3 (1999), 171 – 194. https://doi.org/10.1016/S0921-8890(99)00004-4
[11] Craig Reynolds. 2004. OpenSteer, steering behaviors for autonomous characters.
(2004). http://opensteer.sourceforge.net/
[12] Craig W. Reynolds. 1987. Flocks, Herds and Schools: A Distributed Behavioral
Model. SIGGRAPH Comput. Graph. 21, 4 (Aug. 1987), 25–34. https://doi.org/10.
1145/37402.37406
[13] Craig W. Reynolds. 1999. Steering Behaviors For Autonomous Characters. In
Proceedings of Game Developers Conference 1999. 763–782.
SAC 2018, April 9-13, 2018, Pau, France
[14] H. G. Tanner, A. Jadbabaie, and G. J. Pappas. 2003. Stable flocking of mobile
agents part I: dynamic topology. In 42nd IEEE International Conference on Decision
and Control (IEEE Cat. No.03CH37475), Vol. 2. 2016–2021 Vol.2.
[15] Jingyuan Zhan and Xiang Li. 2011. Flocking of discrete-time multi-agent systems
with predictive mechanisms. IFAC Proceedings Volumes 44, 1 (2011), 5669–5674.
[16] Jingyuan Zhan and Xiang Li. 2013. Flocking of multi-agent systems via model
predictive control based on position-only measurements. IEEE Transactions on
Industrial Informatics 9, 1 (2013), 377–385.
[17] Hai-Tao Zhang, Zhaomeng Cheng, Guanrong Chen, and Chunguang Li. 2015.
Model predictive flocking control for second-order multi-agent systems with
input constraints. IEEE Transactions on Circuits and Systems I: Regular Papers 62,
6 (2015), 1599–1606.
[18] Lifeng Zhou and Shaoyuan Li. 2017. Distributed model predictive control for
multi-agent flocking via neighbor screening optimization. International Journal
of Robust and Nonlinear Control 27, 9 (2017), 1690–1705.
U. Mehmood et al.
| 3cs.SY
|
On Categorical Time Series Models With Covariates
arXiv:1709.09372v2 [math.ST] 2 Oct 2017
Konstantinos Fokianos∗
Lionel Truquet †‡
October 2017
Abstract
We study the problem of stationarity and ergodicity for autoregressive multinomial logistic time series
models which possibly include a latent process and are defined by a GARCH-type recursive equation.
We improve considerably upon the existing about stationarity and ergodicity conditions of those models. Proofs are based on theory which has been developed for chains with complete connections. A
useful coupling technique is employed for studying ergodicity of infinite order finite-state stochastic
processes which generalize finite-state Markov chains. In addition, for finite order Markov chains, we
discuss ergodicity properties of a model which includes strongly exogenous but not necessarily bounded
covariates.
2010 Mathematics Subject Classification: Primary 62M10; secondary 60G10, 60B12.
Keywords and Phrases: autoregression, categorical data, chains with complete connection, coupling, covariates, ergodicity, Markov chains.
∗
University of Cyprus, Department of Mathematics & Statistics, PO BOX 20537, 1678 Nicosia, Cyprus Email:
[email protected].
†
UMR 6625 CNRS IRMAR, University of Rennes 1, Campus de Beaulieu, F-35042 Rennes Cedex, France and
‡
ENSAI, Campus de Ker-Lann, rue Blaise Pascal, BP 37203, 35172 Bruz cedex, France. Email: [email protected].
1
1 Introduction
The goal of this article is to improve upon theoretical properties of regression based models for the analysis
of categorical time series that might include some covariates which are not assumed to be bounded. Binary
time series are particular cases of a categorical time series and the results we obtain apply to logistic autoregressive models. We take the point of view of generalized linear models theory; see McCullagh and Nelder
(1989). The conditional distribution of a categorical time series given its past is multinomial which obviously belongs to the multivariate exponential family of distributions. As such, the theory of generalized
linear models can be applied for modeling different types of categorical data; nominal, interval and scale.
We will be mostly concerned with nominal data and therefore the multinomial logistic model is the natural
candidate for model fitting; see Fahrmeir and Tutz (2001) and Kedem and Fokianos (2002), among other
references, for further discussion on modeling issues regarding categorical data. We emphasize that finite
state Markov chains provide a simple but prominent model of a categorical time series where lagged values
of the response affect the determination of its future states. However, Markov modeling in the context of
categorical time series, poses challenging problems. Indeed, as the order of the Markov chain increases
so does the number of free parameters; in fact, the number of free parameters increases exponentially fast.
Furthermore, the Markovian property requires simultaneous specification of the dynamics of the response
and any possible covariates observed jointly; such a specification might not be possible, in general.
We will be studying models for binary and, more generally, categorical time series, which are of infinite
order or they are driven by a latent process or a feedback mechanism. This type of models is quite analogous
to GARCH models -see Bollerslev (1986)- but they are defined in terms of conditional log–odds instead
of conditional variances. In particular, feedback models make possible low dimensional parametrization,
yet they can accommodate quite complicated data structures. Examples of feedback models, in the context of binary and categorical time series have been studied recently by Moysiadis and Fokianos (2014)
and Fokianos and Moysiadis (2017), among others. We will discuss these results and we will compare
them with our findings. Models and inference for binary time series are topics that have been studied
by several authors; see Kedem (1980) for an early treatment. Regression modeling, in this context, has
been studied by Cox (1981), Stern and Coe (1984), and Slud and Kedem (1994), among others; see also
Kedem and Fokianos (2002, Ch. 2-3) for other early references. Recently, binary time series data have
been increasingly popular in various financial applications (Breen et al. (1989), Butler and Malaikah (1992),
Christoffersen and Diebold (2006), Christoffersen et al. (2007), Startz (2008), Nyberg (2010, 2011, 2013),
2
Kauppi (2012) and Wu and Cui (2014)), but also to other scientific fields. Previous results related to theoretical properties of binary time series models were given by de Jong and Woutersen (2011).
Related work on categorical time series has been reported by Fahrmeir and Kaufmann (1987), Kaufmann
(1987), Fokianos and Kedem (2003) and Russell and Engle (1998, 2005) who proposed a categorical time
series model for financial transactions data. Alternative classes of models are based on the probit link function. Such autoregressive models have been considered by Zeger and Qaqish (1988), Rydberg and Shephard
(2003), Kauppi and Saikkonen (2008), among others. Several other classes of models for the analysis of categorical data have been studied; see the books by Joe (1997) and MacDonald and Zucchini (1997) and the
articles by Biswas and Song (2009) and Weiß (2011).
To prove the theoretical results, we will be assuming a contraction type condition; such conditions are
usually employed for the theoretical analysis of time series models. For instance, in the case of count time
series models, see Fokianos et al. (2009), Neumann (2011) and Doukhan et al. (2012). However, our work
is closely related to the modeling approach suggested by Fokianos and Tjøstheim (2011), because the main
idea is, essentially, to employ the so called canonical link process to model the observed data. Note that
de Jong and Woutersen (2011) have shown near epoch dependence for a binary time series models but these
authors have a different modeling point of view.
Likelihood based inference for the models we study can be developed along the lines of previous references. The proof of consistency and asymptotic normality is based on standard arguments concerning
convergence of the score function and the Hessian matrix. However, we mention that this work relaxes considerably previous results. For the case of a model with covariates we improve upon Kaufmann (1987) and
Kedem and Fokianos (2002, Ch.3) because we avoid any assumptions on the design of covariates. In addition, we will show that we can obtain ergodicity when unbounded covariates are included in a model. The
study of maximum likelihood estimation requires existence of appropriate moments for the covariate process though. Central limit theorems for the maximum likelihood estimators have been given in the previous
references and therefore we do not give any details.
The article is structured as follows: Section 2 discusses general categorical time series models by allowing the conditional probabilities to depend on the whole past of the series. In addition we will be giving
a result about the stationarity and ergodicity of chains with complete connections. These results will be
applied to the case of an infinite order autoregressive multinomial logistic model. Section 3 discuss models
which might include a latent process. The results obtained by Theorem 1 improve the results obtained by
3
Moysiadis and Fokianos (2014) and Fokianos and Moysiadis (2017). Finally, Section 4 discuss inclusion of
exogenous covariates to the autoregressive multinomial logistic model; Theorem 2 is the main result of this
section which discuss existence of such processes and their ergodic properties.
2 Time series autoregressive models for categorical data
2.1 A general approach
Let A be a finite set. For simplicity, we assume that A = {1, 2, . . . , N}, where N is a nonnegative integer.
Suppose that we observe a process with state space A and we are interested on modeling its dynamics.
For instance, consider modeling of a stock price change (0 for no change, 1 for positive change and -1 for
a negative change; see Russell and Engle (1998)) or sleep state status (see Fokianos and Kedem (2003)).
Towards this goal, define a (N − 1)-dimensional vector Yt = Y1t , Y2t , . . . , Y(N−1)t ′ , for 1 ≤ t ≤ n, such that
1, if the j’th category is observed at time t,
Y jt =
0, otherwise,
for all j = 1, 2, . . . , N −1. Throughout this work, consider a stochastic processes (Yt )t∈Z adapted to a filtration
(Ft )t∈Z which is defined through a vector of conditional "success" probabilities, say pt ≡ (p1t , p2t , . . . , p(N−1)t )′ .
In other words
p jt = P Y jt = 1|Ft−1 ,
1 ≤ j ≤ N − 1.
For the last category N, set YNt = 1 −
PN−1
j=1
(1)
Y jt and corresponding success probability pNt = 1 −
PN−1
j=1
p jt .
There are several possibilities for autoregressive modeling of processes that take values on a finite space.
For instance, assuming that d is a vector and A, B matrices of appropriate dimension, consider the following
linear model
pt = d + Apt−1 + BYt−1 ,
t ∈ Z,
(2)
which was studied by Russell and Engle (1998) and Qaqish (2003). Model (2) implies quite complex restrictions on the parameters d, A and B because each element of the vector pt has to belong in the interval
(0, 1). Such restrictions become even more involved when a covariate process is included in (2). To avoid
such subtle technicalities, we adapt the generalized linear models point of view by considering a canonical
link model; see Fokianos and Kedem (2003) for instance. For j = 1, 2, . . . , N − 1, define
λ jt = log p jt /pNt
4
and suppose that the vector process λt = λ1t , . . . , λ(N−1)t ′ is determined by the infinite order model
λt = g (Yt−1 , Yt−2 , . . .) ,
(3)
where g is a suitably defined function. Then, the process (Yt )t∈Z which satisfies (1) and (3), takes its values
in the set E = {e1 , e2 , . . . , eN−1 , 0} where {e1 , . . . , eN−1 } is the canonical basis of RN−1 and 0 is the null vector
of RN−1 . Furthermore, g : E N → RN−1 is a measurable function and the conditional distribution of Yt given
− ≡ (Y , Y , . . .) possibly depends on its infinite past. A useful example of such process
its past values Yt−1
t−1 t−2
is given by the linear process
λt = d +
X
(4)
A j Yt− j
j≥1
where d is a (N − 1)-dimensional vector and (A j ) j≥1 is a sequence of (N − 1) × (N − 1) matrices. Comparison
of (4) to (2) shows that unnecessary restrictions on the unknown coefficients can be circumvented since the
vector λt ∈ RN−1 . Furthermore, covariates can be easily included in (4) by including an additional additive
term. Other categorical type autoregressive models can be considered but (4) has been widely used in several
applications. In the case that N = 2, then (4) is a simple logistic regression model which has been studied
widely in the literature (see Cox and Snell (1970) for an early reference).
Processes, as those we consider in this work, are particular examples of a more general class of processes which are called chains with complete connections. Such processes have been widely studied in
applied probability; Doeblin and Fortet (1937), Harris (1955) and Iosifescu and Grigorescu (1990). Following the work of Bressaud et al. (1999), we discuss next a coupling technique related to chains with complete
connections.
2.2 Some results about chains with complete connection
Throughout this section, consider a finite state space E. For x, y ∈ E N and a positive integer m, we write
m
x = y if xi = yi for 0 ≤ i ≤ m − 1. Consider a probability kernel p(·|·) defined on E N , B E N and takes
values on (E, B(E)) which satisfies the following assumption:
Assumption (A) There exists a sequence (γm )m∈N which decreases to zero, as m → ∞, with γ0 < 1 and
such that for a ∈ E
infm
x,y:x=y
p (a|x)
≥ 1 − γm .
p (a|y)
5
A chain with complete connections is a stationary process satisfying Assumption (A).
For x ∈ E N , consider the chain Znx n∈Z which satisfies that Z−x j = x j for j ≥ 1 and
∞
Y
x
P Znx = a|Zn−
=
z
,
j
≥
1
=
p(a|z)
1z j =x j−n .
j
j
j=n+1
In addition, given a real-valued sequence (γn )n∈N , let the Markov chain S n(γ)
n∈N
taking values in N and
defined by
P S 0(γ) = 0 = 1,
For n ≥ 1, define the quantity
(γ)
P S n+1
= i + 1|S n(γ) = i = 1 − γi ,
(γ)
P S n+1
= 0|S n(γ) = i = γi .
(γ)
γn∗ = P S n = 0 ,
which plays a crucial rule for evaluating the mixing coefficients of the chain. The following result is given
by Bressaud et al. (1999, Prop. 1 and Lemma 1).
Proposition 1. For all x, y ∈ E N , there is a coupling
x,y
integer-valued process T n
defined by
x,y x,y
U n , Vn
n∈Z
of Znx
n∈Z
y
and Zn
n∈Z
such that the
n∈Z
n
o
x,y
x,y
x,y
T n = inf m ≥ 0 : Un−m , Vn−m ,
satisfies
P S n(γ) ≥ k ≤ P T nx,y ≥ k ∀k ∈ N.
Proposition 1 is proved by defining iteratively the pair Unx,y , Vnx,y using the maximal coupling of the
conditional distributions p ·|(un− j ) j≥1 and p ·|(vn− j ) j≥1 (i.e the coupling associated to the total variation
distance between these conditional distributions). Proposition 1 yields the following corollary (see also
Bressaud et al. (1999, Cor. 1) for a specific case of the following result).
Corollary 1. For all k ≥ 1, x, y ∈ E N and B ∈ B(E k ), we have
j−1
k Y
y
X
y
x
x
∗
(1 − γm ) γn+k−
P (Zn , . . . , Zn+k ) ∈ B − P (Zn , . . . , Zn+k ) ∈ B ≤
j.
j=0 m=0
Proof. Using Proposition 1, we obtain
y
y
x
) ∈ B − P (Zn , . . . , Zn+k ) ∈ B
P (Znx , . . . , Zn+k
x,y
x,y
x,y
x,y
≤ P (Un , . . . , Un+k ) , (Vn , . . . , Vn+k )
x,y
≤ P Tn ≤ k .
6
Proposition 1 implies that P T nx,y ≤ k ≤ P S n(γ) ≤ k . The result of the corollary now follows by bounding
(γ)
P S n ≤ k along the lines of the derivation of Bressaud et al. (1999, eq. (4.25)).
As pointed out in Bressaud et al. (1999), if limn→∞ γn∗ = 0, then Corollary 1 implies existence and
uniqueness of a stationary chain (Zn )n∈Z with complete connections and satisfying Assumption (A). Furthermore, Corollary 1 yields a bound for controlling the φ−mixing coefficients associated with (Zn )n∈Z . Indeed,
recall that for two σ−algebras A and B, their φ−mixing coefficients are defined by (see Doukhan (1994),
for instance)
φ (A, B) =
|P (B|A) − P(B)| .
sup
(A,B)∈A×B:P(A)>0
Then, the φ-mixing coefficients of the random sequence (Zn )n∈Z are given by
φ(n) = φ F−∞,0 , Fn,∞ = sup φ F−∞,0 , Fn,n+k ,
k∈N
because a Borel set on the infinite product can be approximated by a finite union of cylinder sets.
P
∗
n≥1 γn
< ∞. Then the infinite order stationary Markov chain (Zn )n∈Z , which
P
exists by Corollary 1, is φ−mixing with mixing coefficients satisfying φ(n) ≤ j≥n γ∗j .
Proposition 2. Suppose that
Proof. Suppose that µ denotes the probability distribution of (Zi )i≤−1 . Then
Z
y
y
x
x
x
P (Znx , . . . , Zn+k
) ∈ B − P (Zn , . . . , Zn+k ) ∈ B µ(dy)
P (Zn , . . . , Zn+k ) ∈ B − P ((Zn , . . . , Zn+k ) ∈ B) ≤
j−1
k Y
X
∗
(1 − γm ) γn+k−
≤
j.
j=0 m=0
But the last bound does not depend on x. Hence, we obtain
P (Zn , . . . , Zn+k ) ∈ B|F−∞,0 − P ((Zn , . . . , Zn+k ) ∈ B)
j−1
k Y
X
∗
(1 − γm ) γn+k−
≤
j
j=0 m=0
X
≤
γ∗j .
j≥n
The last bound, which does not depend on k and B, is also an upper bound for φ(n).
Remark 1. It has been shown in Bressaud et al. (1999, Prop. 2) that
X
X
γk < ∞ ⇒
γk∗ < ∞.
k≥1
k≥1
Moreover, if (γm )m decreases exponentially, then so does γn∗ n . Hence, the result of Prop. 2 follows again
and if (γm )m decreases to zero exponentially fast then so does (φ(n))n . Note also that the φ−mixing property
implies ergodicity of the process; see Bradley (2007, pp. 50-51).
7
2.3 Application to categorical time series
Recall the categorical time series model (Yt )t∈Z whose state space is E = {e1 , . . . , eN−1 , 0} and defined by (1)
and (3). From the results of the previous subsection, we deduce the following corollary. (k · k denotes the
Euclidian norm on RN−1 .)
Corollary 2. Assume model (3) and let a function g : E N → RN−1 be such that there exist a sequence (δ j ) j∈N
P P
which satisfies j∈N k≥ j δk < ∞ and
kg(x) − g(y)k ≤
X
(5)
δ j 1 x j ,y j .
j∈N
Then, there exists a unique stochastic process (Yt )t∈Z taking values in E such that
exp g j (Yt−1 , Yt−2 , . . .)
, 1 ≤ j ≤ N − 1.
P Yt = e j |Ft−1 =
P
1 + N−1
s=1 exp (g s (Yt−1 , Yt−2 , . . .))
(6)
Moreover (Yt )t∈Z is stationary and φ−mixing.
Proof. Denote by p(·|·) the probability kernel defined by
p(e j |x) = F j g (x0 , x1 , . . .) , 1 ≤ j ≤ N − 1,
where F j : RN−1 → [0, 1] is defined for z ∈ RN−1 by
F j (z) =
exp(z j )
,
PN−1
1 + s=1 exp(zs )
1 ≤ j ≤ N − 1.
−1
P
Because of (1), F N (z) = 1 + N−1
. The Lipschitz assumption (5) implies that the j’th component
s=1 exp(z s )
of g, say g j , is bounded, for j = 1, 2 . . . , N. Hence, there exists η > 0 such that, for all 1 ≤ j ≤ N − 1 and
x ∈ EN,
η ≤ p(e j |x),
η ≤ p(0|x).
m
Moreover, F ′j is bounded, for all j. Set M = max1≤ j≤N−1 supz∈RN−1 kF ′j (z)k. Then, if x = y and a ∈ E, we
have that
M
p(a|x)
≥1−
p(a|y)
γm
P
j≥m+1 δ j
η
.
P
Provided that m is large enough, choose γm = M j≥m+1 δ j /η. Hence, there exists an m such that
P
P
= 1 − η. Then we have k≥1 γk < ∞ and using Remark 1, we have also k≥1 γk∗ < ∞. Then from
Corollary 1 and Proposition 2, there exists a unique stationary solution (Yt )t∈Z satisfying (6) and the solution
is φ−mixing.
8
P
< ∞ is equivalent to the condition j∈N ja j < ∞. For the
P
infinite order linear model (4), Corollary 2 applies provided that j≥1 jkA j k < ∞ where kA j k denotes the
We note that the condition
P
j∈N
P
k≥ j ak
corresponding operator norm of the matrix A j . In particular, when N = 2, we obtain that the logistic
P
autoregressive model of infinite order is stationary and φ-mixing if j≥1 j|A j | < ∞, where (A j ) j≥1 denotes a
real valued sequence.
3 Categorical time series with a latent process
In this section, we consider some specific instances of chains with complete connections. Following the
methodology of GARCH models (see Engle (1982), Bollerslev (1986) and the book by Francq and Zakoïan
(2010) for instance), and recalling the notation introduced in (3) we model the latent process (λt )t∈Z to depend additionally on its past values. From a statistical perspective, such parametrization yields parsimony
and allows for more flexible structures that can accommodate various forms of autocorrelation. To be specific, suppose that p and q are two positive integers and Let f : R(N−1)p × E q → RN−1 be a function such
that
λt = f λt−1 , . . . , λt−p , Yt−1 , . . . , Yt−q ,
t ∈ Z.
(7)
We will say that the process ((Yt , λt ))t∈Z is a solution of the problem P f if (7) is satisfied and for each t ∈ Z,
λt is Ft−1 −measurable.
3.1 A general result
For y ∈ E q , define the mapping Gy : R(N−1)p → R(N−1)p by
′
Gy (x) = f (x1 , . . . , x p , y1 , . . . , yq )′ , x′1 , . . . , x′p−1 ,
where f (·) has been defined by (7). The main result of this section is the following.
Theorem 1. Suppose that there exist an integer k ≥ 1, κ ∈ (0, 1) and K > 0 such that
kGy (x) − Gy′ (x′ )k ≤ K kx − x′ k + 1y,y′ ,
and for all x, x′ , y , . . . , y ,
1
q
kGy ◦ · · · ◦ Gy (x) − Gy ◦ · · · ◦ Gy (x′ )k ≤ κkx − x′ k.
1
k
1
k
Then, the following hold true:
9
1. Let x be a vector of R(N−1)p and y
j j≥1
a sequence of elements of E q . Then the limit
lim Gy ◦ · · · ◦ Gy (x)
s→∞
1
s
exists and does not depend on x. Let H : (E q )N → R(N−1)p be the function defined by
H y , y , . . . = lim Gy ◦ · · · ◦ Gy (x).
1
2
s→∞
1
s
Then the function H is bounded. Moreover there exist C > 0 such that
X
kH y , y , . . . − H y′ , y′ , . . . k ≤ C
κ j/k 1y ,y′ .
1
2
1
2
j
j
j≥1
2. A process ((Yt , λt ))t∈Z is solution of the problem P f is and only if (Yt )t∈Z is a chain with complete
connection associated to a function g (see Corollary 2) defined by
g (Yt−1 , Yt−2 , . . .) = H1 (Vt , Vt−1 , . . .) ,
Vt = (Yt−1 , . . . , Yt−q ).
Here H1 denotes the N − 1 first coordinates of the function H defined previously.
3. There exists a unique strictly stationary solution to the equations (1) and (7). Moreover the process
(Yt )t∈Z is φ−mixing with geometrically decreasing mixing coefficients. This implies the ergodicity of
the joint process ((Yt , λt ))t∈Z .
Proof.
1. The first part of the assertion is a straightforward consequence of the assumption and is omitted. We focus on the proof of the Lipschitz property of the function H. For j ≥ 1, we set
( j)
Gy = Gy
( j−1)k+1
We have
◦ Gy
( j−1)k+2
◦ · · · ◦ Gy .
jk
(s)
H y , y , . . . = lim G(1)
y ◦ · · · ◦ G y (x).
1
2
s→∞
By the stated assumption, we obtain that
( j)
( j)
kGy (x) − Gy′ (x)k ≤
k
X
K ℓ 1y
( j−1)k+ℓ
,y′
( j−1)k+ℓ
.
ℓ=1
Hence
(s)
(1)
(s)
kG(1)
y ◦ · · · ◦ G y (x) − G y′ ◦ · · · ◦ G y′ (x)k ≤
∞
X
κj
j=1
By setting C = (K ∨ 1)k and letting x → ∞ we obtain the result.
10
k
X
ℓ=1
K ℓ 1y
( j−1)k+ℓ
,y′
( j−1)k+ℓ
.
2. From the first point of the theorem, the necessary condition follows easily. Now let us assume that
λt = H1 (Vt , Vt−1 , . . .). Setting λt = H (Vt , Vt−1 , . . .), note that the continuity of the function GVt
implies that
lim GVt ◦ GVt−1 ◦ · · · GVt−s (x)
= GVt lim GVt−1 ◦ · · · ◦ GVt−s (x)
s→∞
= GVt λt ,
λt =
s→∞
which proofs that (λt )t∈Z satisfies (7).
3. The third point is a straightforward consequence of the two first results and of Corollary 2. Moreover
the geometric decay of the φ−mixing coefficients has been discussed in the remarks made following
Proposition 2. Finally, it is well-known that φ−mixing implies ergodicity of the process (Yt )t∈Z and
then ergodicity of the process ((λt , Yt ))t∈Z ; see Samorodnitsky (2016, Ch. 2), for instance.
3.2 Linear models
Let A0 , A1 , . . . , A p , B1 , . . . , Bq be some real matrices of size (N − 1) × (N − 1). We assume that
f x1 , . . . , x p ; y1 , . . . , yq = A0 +
p
X
i=1
Ai xi +
q
X
Bi yi .
i=1
Then the above model can be written alternatively as
e + B,
GVt x = Ax
with
A1 · · · A p−1
e
A =
I(N−1)(p−1)
,
0
Ap
A + Pq B Y
0
i=1 i t−i
0
,
B =
.
..
0
where I(N−1)(p−1) denotes the identity matrix of order (N − 1)(p − 1). Then, the assumptions of Theorem
e is less than unity (and then the norm of A
ek is less than one if k is
1 are satisfied if the spectral radius of A
Pp
large enough) which also means that the roots of the polynomial P(z) = det IN−1 − i=1 Ai zi are outside
11
the unit disc; Lütkepohl (2005, Ch.2). For the case p = q = 1, this result improves the conditions proved
by Moysiadis and Fokianos (2014) since it does not require and additional assumption on the coefficient B1 .
In addition, the results answers in affirmative the question posed by Tjøstheim (2012) for the case of binary
autoregressive model. Compared with the work of Fokianos and Moysiadis (2017) we note again that for
the case of logistic autoregressive modeling with binary data, the obtained conditions simplify considerably
their conditions.
3.3 Non-linear models
Recall (7) and assume that there exists a norm k·k on RN−1 and some positive real numbers β1 , . . . , β p , α1 , . . . , αq
Pp
with α = i=1 αi < 1 and for all x1 , x′1 , . . . , x p , x′p , y1 , y′1 , . . . , yq , y′q ∈ RN−1 ,
k f x1 , . . . , x p ; y1 , . . . , yq − f
x′1 , . . . , x′p ; y′1 , . . . , y′q
k≤
p
X
αi kxi −
i=1
x′i k
+
q
X
βi kyi − y′i k.
i=1
It can be proved under this condition that, for a process (Zt )t∈Z taking values in E q , the random mapping
GZt (x) = f (x, Zt ), x1 , . . . , x p−1
is contracting, after iteration. Indeed, if x, y ∈ R p , σix = xi for 1 ≤ i ≤ p and
x
x
, Zt ,
σtx = f σt−1
, . . . , σt−p
t ≥ p + 1,
it follows by induction that
y
σtx − σt ≤ α
t−p
p
kx − yk,
t ≥ p + 1.
Hence, there exists an integer m ≥ 1, such that the mapping
Ht(m) (x) = GZt ◦ GZt−1 ◦ · · · ◦ GZt−m (x)
satisfies
Ht(m) (x) − Ht(m) (x′ ) ≤ κkx − x′ k
for some κ ∈ (0, 1). Therefore the assumption of Theorem 1 is satisfied. We note again that this condition improves upon the conditions obtained by Moysiadis and Fokianos (2014) and Fokianos and Moysiadis
(2017) since they require only that α < 1.
12
4 Inclusion of exogenous covariates
In this section, we study the problem of including a covariate process (Zt )t∈Z in an autoregressive categorical
time series model. We will be assuming that the covariate process is strongly exogenous and unbounded.
The assumption of exogeneity implies that for each time t, Yt is independent from (Z s ) s≥t+1 conditionally
− ,Z
to Yt−1 , Zt , Yt−2 , Zt−1 . . . and allows simple computation of the likelihood function. Indeed, if f ·|Yt−1
t
denotes the conditional density of Yt given Yt−1 , Zt , Yt−2 , Zt−1 . . ., then, for a bounded measurable function h,
n
Y
h
i Z
E h (Y1 , Y2 , . . . , Yn ) Z1 = z1 , . . . , Zn = zn , Y0− = y−0 =
h(y1 , . . . , yn )
f yi |y−i−1 , zi µ(dy1 ) · · · µ(dyn ),
i=1
where µ denotes a reference measure for the model. This type of exogeneity is also called Granger causality
or Sims causality in the literature (see, for instance, Gouriéroux and Monfort (1995, Sec. 1.5.2), for a
discussion of these different concepts). We will be restricting our study to the case of finite order Markov
chains, i.e. the parameter λt does not depend on its past values. The general case appears more difficult to
tackle and is will be considered in another communication.
We are not aware of any result which guarantees ergodicity of a model with covariate even in the simple
case of a finite-state Markov chain. As we will see, there is an interesting parallel between Markov chains
with exogeneous covariates and Markov chains in random environments which have been studied in probability theory. In the proof of Theorem 2 given below, we will use an approach discussed in Cogburn (1984)
for showing ergodicity of Markov processes in random environments.
4.1 A general result for finite state Markov chains with covariates
We will be discussing results concerning stationarity and ergodicity of a finite state Markov chain which can
be jointly observed with a covariate process. In what follows, denote by Z = (Zt )t∈Z a stationary process
with values in the space G = Rd and (Yt )t∈N a process which takes values in a finite set E. In addition,
conditionally on Z, (Yt )t∈N is a finite-state inhomogeneous Markov chain. More precisely, we assume that
n
o
there exist a family of transition matrices Pg : g ∈ G such that
P (Yt = y|Yt−1 = x; Z) = PZt (x, y),
(x, y) ∈ E 2 .
Throughout the section we will assume the following:
(E1) There exists an integer m ≥ 1 such that for all (z1 , z2 , . . . , zm , x, y) ∈ Gm × E 2 ,
Pz1 Pz2 · · · Pzm (x, y) > 0.
13
(8)
(E2) The process Z is mixing in the ergodic theory sense, i.e for all elements A and B of B GZ , we have
lim P Z ∈ A, τn Z ∈ B = P(Z ∈ A)P(Z ∈ B),
n→∞
where τ denotes the shift operator on F Z defined by τZ = Z j+1
j∈Z
.
Note that Assumption (E1) implies that a process (Yt )t∈Z satisfying (8) also satisfies
P (Yt+m = y|Yt = x, Z) > 0, a.s
(x, y, t) ∈ E × E × Z.
In addition, Assumption (E2) is stronger than assuming ergodicity of the process Z but weaker than the
classical strong mixing condition usually employed in the literature. A large number of useful stochastic processes are mixing, for instance the strong mixing processes and Bernoulli shifts defined by Zt =
H (εt , εt−1 , . . .) where H is a measurable function and (εt )t∈Z is an i.i.d sequence. For instance, Samorodnitsky
(2016, Ch.2), discusses several properties of the different types of mixing in ergodic theory of stationary processes. Assumption (E2) will be employed for obtaining ergodicity for the shift operator τm which is not
implied by the ergodicity of the shift operator τ. The main result of this section is given by the following
theorem.
Theorem 2. Suppose that (E1-E2) hold true. Then there exists a unique stochastic process (Yt )t∈Z satisfying
(8). Moreover the process ((Yt , Zt ))t∈Z is ergodic.
Proof. We first show that the almost sure limit lim s→∞ PZt−s · · · PZt (x, y) exists for each y ∈ E and does
not depend on x. For Markov chains, this condition is comparable to the weak ergodicity notion, but here
the limit is taken in the backward sense. See Seneta (2006) for several sufficient conditions ensuring weak
ergodicity properties of time-inhomogeneous Markov chains, using ergodicity coefficients. Recall that the
so-called Dobrushin’s contraction coefficient of a stochastic matrix P is defined by
c(P) =
1
sup kP(x, ·) − P(y, ·)kT V ,
2 x,y∈E
where for two probability mesures µ and ν on the finite set E, the total variation distance between µ and ν is
P
defined by kµ − νkT V = x∈E |µ(x) − ν(x)|. It is well known that we have the contraction
kµP − νPkT V ≤ c(P)kµ − νkT V
and for two stochastic matrices P and Q, we have c(PQ) ≤ c(P)c(Q).
14
Moreover c(P) ≤ 1 − |E| min x,y∈E P(x, y), where |E| denotes the cardinality of the set E. So Assumption (E1)
ensures that for all t ∈ Z, c PZt PZt+1 · · · PZt+m−1 < 1 a.s. Now let x , x′ ∈ E, t ∈ Z and s = km + ℓ. we obtain
by setting ρ = 1 − η|E|,
Y
kPZt−s+1 · · · PZt (x, ·) − PZt−s+1 · · · PZt (x′ , ·)kT V ≤ 2c PZt−km+1 · · · PZt ≤ 2
c PZt−( j+1)m+1 · · · PZt− jm .
k−1
j=0
From Assumption (E2), the covariate process Z is mixing. Then the process Zt− j
j∈Z
is also mixing. Indeed,
if θt and τ denote the mappings defined on GZ by θt x = (xt−i )i∈Z and τx = (xi+1 )i∈Z respectively, we have
τ ◦ θt = θt ◦ τ−1 . Then for two Borel sets Aand B, we get
P θt Z ∈ A, τn θt Z ∈ B = P Z ∈ θt−1 A, τ−n Z ∈ θt−1 B
= P τn Z ∈ θt−1 A, Z ∈ θt−1 B
→ P (θt Z ∈ A) P (θt Z ∈ B) .
Moreover, observe that the operator τm is ergodic for PZ . Indeed, if a Borel set A is such that τm A = A, we
obtain, using assumption E2,
P (Z ∈ A) = P τkm Z ∈ A, Z ∈ A →k→∞ P (Z ∈ A)2 .
Then, we conclude that PZ (A) ∈ {0, 1}, which shows that τm is ergodic. Now, using Assumption (E1), we
have E log c PZ1 · · · PZm < 0. Then, from the ergodic theorem, we get
k−1
k−1
X
Y
c PZt−( j+1)m+1 · · · PZt− jm = exp log c PZt−( j+1)m+1 · · · PZt− jm →k→∞ 0.
j=0
j=0
In addition, when n ≥ s, we deduce that
kPZt−s+1 · · · PZt (x, ·) − PZt−n · · · PZt (x, ·)kT V ≤ 2c PZt−s+1 · · · PZt .
This shows that the product of matrices PZt−s+1 · · · PZt converges, when s → ∞, to a stochastic matrix whose
rows are all equal. Then there exists a measurable function D : F N → E N with N = |E| such that
D (Zt , Zt−1 , . . .) = lim PZt−s+1 · · · PZt (x, ·) a.s.
s→∞
Setting Dt = D (Zt , Zt−1 , . . .), Dt is a random probability measure on E. For t ∈ Z, z ∈ F Z , k a non-negative
integer and y0 , y1 , . . . , yk ∈ E, we set
µt:t+k (z; y0 , y1 , . . . , yk ) =
k
Y
i=1
15
Pzt+i (yi−1 , yi )Dt (y0 ) .
From the Kolmogorov extension theorem, there exists for PZ −almost all values of z ∈ B GZ a unique
measure µ(z, ·) on E Z with marginals µt:t+k (z, ·). Hence, if ζ denotes the probability distribution of Z, the
measure γ defined by
γ(A × B) =
Z
µ(z, A)ζ(dz),
(A, B) ∈ B(E Z ) × B(F Z ),
B
is that of a couple from a stationary process (Y, Z) satisfying (8). To show uniqueness, let Yt′
t∈Z
be another
stochastic process satisfying (8). Then the distribution of Y ′ |Z = z is that of a non-homogeneous Markov
chain with transitions {Pzt : t ∈ Z}. As shown before, this conditional distribution is unique and equal to
µ(z, ·).
Next we show ergodicity of the process ((Yt , Zt ))t∈Z . To this end, we use an approach introduced in Cogburn
(1984) for the study of Markov processes in random environment. This type of argument is also used in
Sinn and Poupart (2011) for positive transition matrix PZt and we give here a more general and shorter
proof. The approach used in Cogburn (1984) consists in considering the Markov kernel Q on E × F defined
by
Q ((x, z), {y} × A) = Pz1 (x, y)1A (τz),
A ∈ B FZ .
If ν denotes the probability distribution (Yt , τt Z) which takes values in E × F, then ν is invariant for Q and the
process (Ht )t∈Z defined by Ht = (Yt , τt Z) is a Markov chain of transition kernel Q. Let C be a ν− invariant
set, i.e Q((x, z), C) = 1 for ν−almost every (x, z) ∈ C. Using Corollary 5.11 in Hairer (2006), the Markov
chain (Ht )t∈Z forms an ergodic process if and only if every ν−invariant set C is of measure 0 or 1. In our
case, we have C = ∪x∈E {x} × C x for some C x ∈ B F Z . To this end, we first note that if C is ν−invariant,
then
ν(C) = νQ(C) =
Z
dν(x, z)Q((x, z), C) +
C
Z
Cc
dν(x, z)Q((x, z), C) = ν(C) +
Z
dν(x, z)Q((x, z), C).
Cc
Then we get Q((x, z), C) = 0 for ν−almost every (x, z) ∈ C c , the complement of C in E × F. Hence, we
obtain Q((x, z), C) = Q1C (x, z) = 1C (x, z) for ν−almost every (x, z). But this also gives Qm 1C = 1C , ν a.e,
where m has been defined by assumption (E1). Moreover, we have that
Qm ((x, z), C) =
X
y∈E
1C (y, τm z) Pz1 · · · Pzm (x, y).
(9)
We write A = B ν−a.e. if ν (A∆B) = 0 where A∆B denotes the symmetric difference of the sets A and B,
i.e A∆B = (A ∩ Bc ) ∪ (Ac ∩ B). From assumption (E1), all the entries of the matrix Pz1 · · · Pzm are positive.
16
Then we deduce that for almost every (x, z) ∈ C, we have (y, τm z) ∈ C for all y ∈ E. We set D = ∩y∈E Cy .
Let us denote by ν1 and ν2 the marginals of ν. We first remark that for all A ∈ B(F), we have
Z
XZ
ν ({x} × A) =
P (Y0 = x|Z = z) ν2 (dx) =
P (Y0 = x|Y−m = y, Z = z) P (Yn = y|Z = z) ν2 (dz).
A
y∈E
A
Employing again assumption (E1), we get ν ({x} × A) ≥ ην2 (A). For x ∈ E, we set B x = τm C x \ D. As stated
above we have
ν ({(x, z) : z ∈ B x }) =
We conclude that ν2 (B x ) = 0 for all x ∈ E and then ν2
X
ν ({x} × B x ) = 0.
x∈E
(τmC
x
\ C x ) = 0. By stationarity, ν2 (τmC x ) = ν2 (C x ).
Therefore, for every x ∈ E, τmC x = C x , µ−a.e. But using assumption E2, we have that
ν2 (C x ) = ν2 τkm C x ∩ C x →k→∞ ν2 (C x )2 .
Then, we conclude that ν2 (C x ) ∈ {0, 1}. If ν2 (C x ) = 0 for every x, we easily get ν(C) = 0. Now if there exists
x ∈ E such that ν2 (C x ) = 1, we have, using the equality ν2 (B x ) = 0,
1 ≤ ν2 (C x ) = ν2 τmC x = ν2 τmC x ∩ D ≤ ν2 (D) ≤ min ν2 (Cy ).
y∈E
Then ν2 (Cy ) = 1 for each y ∈ E. Finally we obtain
ν(C) =
X
X
ν1 (y) = 1.
ν {y} × Cy =
y∈E
y∈E
Hence, we have shown that the process (Ht )t∈Z is ergodic and so is the process ((Yt , Zt )t∈Z .
4.2 Application to the multinomial logistic model with covariates
We assume here that conditionally to a covariate process (Zt )t∈Z taking values in Rd , the process (Yt )t∈Z is a
q−order Markov chain such that
P Yt = e j |Yt−1 , . . . , Yt−q , Z =
exp g j Yt−1 , . . . , Yt−q ; Zt
Y
,
e
:=
Q
t−q:t−1
j
Z
P
t
1 + N−1
s=1 exp g s Yt−1 , . . . , Yt−q ; Zt
for some measurable functions g j : E q × Rd , 1 ≤ j ≤ N − 1. Let us check that assumption E1 is satisfied for
′
′ , . . . , Y′
the conditional Markov chain (Xt )t∈Z defined by Xt = Yt′ , Yt−1
t−q+1 . Conditionally to Z, the process
(Xt )t∈Z defines a time-inhomogeneous Markov chains such that
P Zt
q−1
Y
1vs+1 =us .
(u1 , . . . , uq ), (v1 , . . . , vq ) : = QZt (u1 , . . . , uq ), v1
s=1
17
Since the transition QZt takes only positive values, the assumption E1 follows by taking m = q and γ = aq .
Then assuming E2 for the covariate process, Theorem 2 applies and guarantees the ergodicity of the process
((Yt , Zt ))t∈Z . These results show that our approach simplifies conditions required to obtain consistency and
asymptotic normality of the maximum likelihood estimator, even in the case of considering covariates.
However, existence of moments for the covariate process is still required to study large sample properties
the maximum likelihood estimator.
References
Biswas, A. and P. X.-K. Song (2009). Discrete-valued ARMA processes. Statistics & Probability Letters 79,
1884–1889.
Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics 31, 307–327.
Bradley, R. (2007). Introduction to Strong Mixing Conditions. Vol 1. Heber City: UT:Kendrick Press.
Breen, W., L. R. Glosten, and R. Jagannathan (1989). Economic significance of predictable variations in
stock index returns. The Journal of Finance 44, 1177–1189.
Bressaud, X., R. Fernández, and A. Galves (1999). Decay of correlations for non-Hölderian dynamics. a
coupling approach. Electron. J. Probab. 4, 1–19.
Butler, K. C. and S. Malaikah (1992). Efficiency and inefficiency in thinly traded stock markets: Kuwait and
Saudi Arabia. Journal of Banking & Finance 16, 197–210.
Christoffersen, P., F. X. Diebold, R. S. Mariano, A. Tay, and Y. K. Tse (2007). Direction-of-change forecasts
for asian equity markets based on conditional variance, skewness and kurtosis dynamics: Evidence from
hong kong and singapore. Journal of Financial Forecasting 1, 1–22.
Christoffersen, P. F. and F. X. Diebold (2006). Financial asset returns, direction-of-change forecasting, and
volatility dynamics. Management Science 52, 1273–1287.
Cogburn, R. (1984). The ergodic theory of Markov chains in randon environments. Z. Wahrscheinlichkeitstheory verw. Gebiete 66, 109–128.
18
Cox, D. R. (1981). Statistical analysis of time series: some recent developments. Scand. J. Statist. 8, 93–115.
Cox, D. R. and E. J. Snell (1970). The Analysis of Binary Data. London: Chapman & Hall.
de Jong, R. M. and T. Woutersen (2011). Dynamic time series binary choice. Econometric Theory 27,
673–702.
Doeblin, W. and R. Fortet (1937). Sur les chaînes à liaisons complètes. Bull. Soc. Math. France 65, 132–148.
Doukhan, P. (1994). Mixing: properties and examples. Number 85 in Lecture Notes in Statistics. New York:
Springer-Verlag.
Doukhan, P., K. Fokianos, and D. Tjøstheim (2012). On weak dependence conditions for Poisson autoregressions. Statist. Probab. Lett. 82, 942–948.
Engle, R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of the variance of United
Kingdom inflation. Econometrica 50, 987–1007.
Fahrmeir, L. and H. Kaufmann (1987). Regression models for nonstationary categorical time series. Journal
of Time Series Analysis 8, 147–160.
Fahrmeir, L. and G. Tutz (2001). Multivariate statistical modelling based on generalized linear models
(Second ed.). Springer Series in Statistics. New York: Springer-Verlag. With contributions by Wolfgang
Hennevogl.
Fokianos, K. and B. Kedem (2003). Regression theory for categorical time series. Statist. Sci. 18, 357–376.
Fokianos, K. and T. Moysiadis (2017). Binary time series friven by a latent process. Econometrics and
Statistics 2, 117–130.
Fokianos, K., A. Rahbek, and D. Tjostheim (2009). Poisson autoregression. J. Amer. Statist. Assoc. 104,
1430–1439.
Fokianos, K. and D. Tjøstheim (2011). Log-linear Poisson autoregression. J. Multivariate Anal. 102, 563–
578.
Francq, C. and J.-M. Zakoïan (2010). GARCH models: Stracture, Statistical Inference and Financial Applications. United Kingdom: Wiley.
19
Gouriéroux, C. and A. Monfort (1995). Statistics and Econometric Models. Volume 1. Cambridge University
Press.
Hairer, M. (2006).
Ergodic properties of Markov processes.
Lecture notes available at
http://www.hairer.org/notes/Markov.pdf.
Harris, T. (1955). On chains of infinite order. Pacific J. Math. 5, 707–724.
Iosifescu, M. and S. Grigorescu (1990). Dependence with Complete Connections and its Applications.
Cambridge University Press.
Joe, H. (1997). Multivariate Models and Dependence Concepts. London: Chapman & Hall.
Kaufmann, H. (1987). Regression models for nonstationary categorical time series: Asymptotic estimation
theory. Annals of Statistics 15, 79–98.
Kauppi, H. (2012). Predicting the direction of the Fed’s target rate. Journal of Forecasting 31, 47–67.
Kauppi, H. and P. Saikkonen (2008). Predicting US recessions with dynamic binary response models. The
Review of Economics and Statistics 90, 777–791.
Kedem, B. (1980). Binary Time Series. Marcel Dekker, New York.
Kedem, B. and K. Fokianos (2002). Regression models for time series analysis. Hoboken, NJ: Wiley.
Lütkepohl, H. (2005). New Introduction to Multiple Time Series Analysis (1st ed.). Berlin: Springer.
MacDonald, I. L. and W. Zucchini (1997). Hidden Markov and Other Models for Discrete–valued Time
Series. London: Chapman & Hall.
McCullagh, P. and J. A. Nelder (1989). Generalized Linear Models (2nd ed.). London: Chapman & Hall.
Moysiadis, T. and K. Fokianos (2014). On binary and categorical time series models with feedback. J.
Multivariate Anal. 131, 209–228.
Neumann, M. (2011). Absolute regularity and ergodicity of poisson count processes. Bernoulli 17, 1268–
1284.
Nyberg, H. (2010). Dynamic probit models and financial variables in recession forecasting. Journal of
Forecasting 29, 215–230.
20
Nyberg, H. (2011). Forecasting the direction of the us stock market with dynamic binary probit models.
International Journal of Forecasting 27, 561–578.
Nyberg, H. (2013). Predicting bear and bull stock markets with dynamic binary time series models. Journal
of Banking & Finance 37, 3351–3363.
Qaqish, B. F. (2003). A family of multivariate binary distributions for simulating correlated binary variables
with specified marginal means and correlations. Biometrika 90, 455–463.
Russell, J. R. and R. F. Engle (1998). Econometric Analysis of Discrete-Valued Irregularly-Spaced Financial
Transactions Data Using a New Autoregressive Conditional Multinomial Model. SSRN eLibrary.
Russell, J. R. and R. F. Engle (2005). A discrete-state continuous-time model of financial transactions prices
and times. Journal of Business and Economic Statistics 23, 166–180.
Rydberg, T. H. and N. Shephard (2003). Dynamics of trade-by-trade price movements: decomposition and
models. Journal of Financial Econometrics 1, 2–25.
Samorodnitsky, G. (2016). Stochastic Processes and Long Range Dependence. Springer.
Seneta, E. (2006). Non-negative matrices and Markov chains. Springer Series in Statistics. Springer, New
York. Revised reprint of the second (1981) edition [Springer-Verlag, New York; MR0719544].
Sinn, M. and P. Poupart (2011). Asymptotic theory for linear-chain conditional random fields. In G. Gordon,
D. Dunson, and M. Dudik (Eds.), Proceedings of the Fourteenth International Conference on Artificial
Intelligence and Statistics, Volume 15 of Proceedings of Machine Learning Research, Fort Lauderdale,
FL, USA, pp. 679–687.
Slud, E. V. and B. Kedem (1994). Partial likelihood analysis of logistic regression and autoregression.
Statist. Sinica 4, 89–106.
Startz, R. (2008). Binomial autoregressive moving average models with an application to US recessions.
Journal of Business & Economic statistics 26, 1–8.
Stern, R. D. and R. Coe (1984). A model fitting analysis of daily rainfall data. Journal of the Royal Statistical
Society. Series A (General) 147, pp. 1–34.
21
Tjøstheim, D. (2012). Rejoinder on: Some recent theory for autoregressive count time series. TEST 21,
469–476.
Weiß, C. (2011). Generalized choice models for categorical time series. J. Statist. Plann. Inference 141,
2849–2862.
Wu, R. and Y. Cui (2014). A parameter-driven logit regression model for binary time series. Journal of Time
Series Analysis 35, 462–477.
Zeger, S. L. and B. Qaqish (1988). Markov regression models for time series: a quasi-likelihood approach.
Biometrics 44, 1019–1031.
22
| 10math.ST
|
SOCRATES
A System For Scalable Graph Analytics
C. Savkli, R.Carr, M. Chapman, B. Chee, D. Minch
Johns Hopkins University, Applied Physics Laboratory
Mail Stop MP6-N312
Laurel, MD 20723-6099 USA
[email protected]
Abstract— A distributed semantic graph processing system
that provides locality control, indexing, graph query, and parallel
processing capabilities is presented.
Keywords—graph, semantic,
I. INTRODUCTION
Graphs provide a flexible data structure that facilitates fusion
of disparate data sets. Popularity of graphs has shown a steady
growth with the development of internet, cyber, and social
networks. While graphs provide a flexible data structure,
processing of large graphs remain a challenging problem.
Successful implementation of graph analytics revolves around
several key considerations: rapid data ingest and retrieval,
scalable storage, and parallel processing. In this paper we
present a graph analytics platform that is particularly focused
on facilitating large scale analysis of semantic graphs.
Recently, NoSQL systems such as Hadoop have become
popular for storing Big Data, however these systems face
several fundamental challenges that make analyzing that data
difficult:
•
•
•
Lack of secondary indexing, which leads to poor
performance of attribute queries (e.g. “What flights
have we seen moving faster than 500mph?”).
Lack of locality control, which can lead to
unnecessary movement of data.
Lack of schema, which makes database maintenance
challenging.
More traditional relational databases (a.k.a. RDBMS) do not
have these problems, but face their own set of challenges
when dealing with Big Data:
•
•
Table structures not flexible enough to support new
kinds of data easily.
Poor parallelization & scalability.
SOCRATES provides a solution that combines the key
features of these two approaches and avoids all of the above
problems. It features several advances in data management for
parallel computing and scalable distributed storage:
•
•
•
•
•
•
Graph API supports representation of different kinds
of data using a simple, extensible database schema.
Distributed storage & processing: Functionality to
execute algorithms on each local machine and results
are merged without moving data.
Indexing: Every attribute is indexed for fast random
access and analytics.
Distributed management: Nothing in the cluster is
centrally managed. This includes communication as
well as locating graph elements.
Locality control: Graph vertices can be placed on
specific machines, a feature essential for minimizing
data movement in graph analytics.
Platform independence: Built using Java Standards
and can run in heterogeneous hardware/software
configurations.
II. RELATED WORK
SOCRATES is designed for attribute rich data – having
many labels on both nodes and edges. Systems such as
Twitter’s Cassovary and FlockDB, GPS, Pregel and Pegasus
are fast at processing structural graphs, however, they do not
leverage or provide any ability to store or use labels on edges
or nodes besides edge weights [1][2][3][4][5]. Current graph
benchmarking tools such as HPC Scalable Graph Analysis
Benchmark (HPC-SGAB) generate tuples of data with the
form <StartVertex, EndVertex, Weight> with no other
attributes which plays well to the strengths of Cassovary or
Pregel [6]. However, such benchmarks tend to ignore the
functionality we specifically aim for within this work.
Other graph databases focus more on content or tend to
model specific relationship types such as ontology. These
databases include Jena, OpenLink Virtuoso, R2DF and other
commercial offerings use Resource Description Framework
(RDF), which was originally designed to represent metadata
[6][7][8]. RDF expressions consist of triples with a subject,
predicate and object that is stored and queried against. The
predicate in each triple represents a relationship between a
subject and object. Intuitively, a general set of RDF tuples
can be considered a graph, however, formally RDF is not
defined as a mathematical concept of a graph[9].
Neo4J, Titan, HypergraphDB and DEX offer similar
capabilities to SOCRATES. SOCRATES, Neo4J and Titan
for example make use of Blueprints API for interacting with
graphs [10]. However, SOCRATES has extended this API to
offer enhanced graph processing functionality that includes
locality control, additional graph methods facilitating graph
analytics, as well as a parallel processing capability.
SOCRATES is built upon a cluster of SQL databases,
similar to FaceBook’s TAO and Twitter’s deprecated
FlockDB [11][5]. FaceBook’s social graph is currently stored
in TAO, a data model and API specifically designed for social
graphs. FaceBook’s social graph is served via TAO which
was tailor to fit its workload needs: frequent reads with fewer
writes, most edge queries having empty results and node
connectivity and data size have distributions with long tails.
These long tails enable efficient cache implementations, for
example if a large portion of reads are potentially in cache
(people are interested in current events – time locality or
alternatively something becomes “viral” and is viewed by
many people). TAO enables relatively few queries and stores
its main data objects and associations as key-value pairs. This
is similar to how SOCRATES stores attributes utilizing many
tables in which each attribute is stored as value and a node or
edge’s id as the key. This schema removes joins as a potential
memory and performance bottleneck. TAO was developed
with specific constraints in mind such as global scaling and
time based social graphs, whereas SOCRATES is targeted at
more general graph structures.
while edge attributes are stored on the machine where edges
originate.
!
Fig. 2: The JGraph model of parallelism in Socrates. Users
submit a processing job that is run in parallel on each node in
the cluster, and given access to the subgraph stored on that
node.
Access to data is provided by a Graph API where users can
interact with the cluster using a simple graph interface. Graph
API includes an implementation of Blueprints API. Besides the
standard Blueprints methods SOCRATES graph API provides
additional methods that take advantage of the architecture for
efficient implementation of the analytics. One such method
involves adding vertices of the graph to specific machines. The
machine level access to graph allows a user to partition the
graph to minimize edge crossings between machines. The
following example, which is built using the Brightkite data set,
illustrates the benefit of locality control:
III. DESIGN
A. Data Management & Locality
SOCRATES uses the following conventions: Each graph
vertex exists only on 1 machine and each graph edge can exist
at most on 2 machines. Edges know the IDs of vertices they
connect as well as the machine those nodes reside on. There is
no central management of location information.
Fig. 1: Graph structure representation on the cluster.
Attributes of the graph are stored separately in 2 column tables
where each attribute can be independently indexed and queried.
This approach avoids the complexity typically associated with
table structure changes in relational databases. Attributes of
graph vertices are stored on the machine those vertices reside
Fig. 3: Ability to partition graph over the cluster nodes
minimizes movement of data.
The graph on the left represents 1 node of a 4 machine cluster
where data was archived without locality control. In this case
the probability that a neighbor resides on the same machine is
¼ which is reflected in the outcome where only about a quarter
of vertices have their neighbors local. Example on the right
shows what the partial graph on the same machine looks like
when archived using SOCRATES. Ability to place graph
nodes in particular machines can also be used to automatically
partition the data when attributes can be hashed to generate a
machine ID, such as partitioning of graph based on latitudelongitude attributes of vertices.
SOCRATES provides efficient implementation of
parallelized Graph query for sub-graph matching. A
particularly useful query involves finding joint neighbors of a
pair of vertices. This query is a key operation for a variety of
link discovery analysis and it is efficiently implemented
without moving data irrespective of where vertices are located.
Implementation of the joint neighbor finding relies on the data
structure maintained on each machine to identify joint
neighbors through a database query rather than iterating over
neighbors of each vertex on the client. Every vertex on the
graph knows the location and unique ID of vertices it is
connected to without having to query another machine. An
example of sub-graph query is provided in Figure 4. In this
figure triangle shown on the left represent a graph query and
matching result is illustrated on the right.
Fig. 4: A query example where a sub-graph with structure and
attribute constraints is found in a larger graph.
B. Parallel Processing
A key feature for handling large scale graphs is ability to
process graph in parallel without having to move data.
Primary challenge in parallel processing of graphs is
associated with the fact that, for most nontrivial problems,
analysis on each machine on the cluster requires access to data
on other machines to be able to produce a result. This is an
area where ability to control partitioning of the graph over the
cluster becomes critical.
Socrates supports three models of parallel processing over
the distributed graph: DGraph, JGraph, and Neighborhood.
Each type provides a different tradeoff between ease of
algorithm implementation, parallelism of client code, and
network communication.
DGraph: In the first model, Socrates provides clients with
access to the DGraph class, which implements the Blueprints
API and abstracts away the distributed nature of the
underlying graph. Methods of the DGraph class are
implemented with parallel calls to the underlying database
where possible, but all results are sent back to the client
machine and no client code runs on the Socrates cluster. This
model of parallelism is suitable for developing analytics that
need a global view of the graph and cannot benefit from being
parallelized across the entire cluster (examples?).
JGraph: In the second model, Socrates allows clients to create
processing jobs that can be submitted to the cluster to run in
parallel on each node; each job is given access to the JGraph
local to the node it is being run on (See Fig. X). A JGraph is
another implementation of the Blueprints API that represents
the partial graph stored on its local machine. Vertex iterators
used in parallel jobs on JGraphs iterate over vertices that are
local in that machine. However any questions asked about
local vertices, such as getNeighbors() operation, retrieves all
matching results independent of where they are located.
Therefore the implementation of parallel jobs is quite similar
to a regular standalone program. User has a clear view of the
boundaries of the local graph and can limit operations on the
local graph based on this information. This model of
parallelism is suitable for developing analytics that can make
use of a wide view of the graph as well as benefit from
parallelism, such as sub-graph isomorphism. It is also useful if
the graph can be partitioned into disjoint sub-graphs that are
small enough to fit on one machine; in this case, it is trivial to
use Socrates’ locality control features to make sure the entire
sub-graphs are placed on the same machine, allowing any
algorithm that runs on a DGraph in the previous model to be
parallelized.
Neighborhood: The final model of parallelism supported
by Socrates is intended for algorithms that perform local
computation only, such as centrality measurements or
PageRank. Socrates provides an interface that allows clients to
define a function that will be run in batch on every vertex in
the graph. When the function is called, its input is a
TinkerGraph (an in-memory implementation of Blueprints)
that contains one vertex labeled “root”, and may contain other
elements that the client specifies when the function processing
job is submitted. The client is able to specify whether the
TinkerGraph should contain the root vertex’s immediate
neighbors (or in/out neighbors only in the case of a directed
graph) and their edges with the root, as well as any properties
of vertices or edges that should be fetched. The client’s
function is then able to write out new property values for the
root node or any of its neighboring edges (future
implementations will allow new neighboring vertices and
edges to be added as well).
This model of parallelism is intended to make it easy to
write local-computation graph analytics that take full
advantage of the computing power of a cluster. Under the
hood, Socrates takes care of running the client function in
parallel on each node using as many threads as the hardware
on that node will support, optimizing SQL queries and inserts,
caching frequently-used values, and minimizing network
communication between nodes.
Communication for parallel processing is provided by Java
Messaging Service (JMS) using publish-subscribe method. In
order to eliminate centralized communication and potential
bottlenecks, each machine operates its own message broker.
Every machine on the cluster is therefore on an equal footing.
Parallelizing message handling also eliminates a potential
single point of failure in the system. Jobs are executed in
parallel on each machine and the results are, optionally,
returned back to the client submitting the job request.
IV. PERFORMANCE
In this section, we provide performance benchmarks
demonstrating the scalability of Socrates in terms of ingest and
parallelized analytics.
A. Cluster configuration
Our cluster consists of 16 servers which are equipped with
quad-core Intel Xeon E5-2609 2.5GHz processors, 64 GB
1600MHz DDR3 RAM, and two 4.0TB Seagate
Constellation HDDs in RAID 0. The servers are running
CentOS, and Socrates is using MySQL 5.5 with the
TokuDB storage engine as its data store.
http://dl.acm.org/citation.cfm?id=1248393] in place of BTrees, and seems to avoid this problem.
Figure 6 shows the ingest speeds on a logarithmic scale for
varying cluster sizes ingesting an E-R graph with 10
million vertices and 100 million edges.
B. Ingest
To measure our ingest speeds, we insert large randomlygenerated
Erdos-Renyi
(E-R)
graphs
[
http://www.citeulike.org/group/3072/article/1666220] into
the Socrates cluster from an external machine. The E-R
graphs we use here consist of 100-vertex connected
components with an average of 1000 edges each, however
the ingest speed of these graphs depends only on the
number of vertices and edges, and not on the underlying
structure of the graph.
We insert E-R graphs with a total size varying from
100,000 vertices and one million edges to 100 million
vertices and one billion edges. In order to see how well our
ingest speed scales as the size of a cluster grows, we have
repeated the ingest benchmarks using only 2, 4, and 8
nodes in addition to the full 16.
Figure 5 shows the ingest speeds, expressed as elements
inserted per second, for graphs and clusters of varying
sizes.
Fig. 5: Socrates insertion speeds for a randomly-generated
Erdos-Renyi graph with an average of 10 edges per vertex,
with varying cluster sizes.
We can see that, although the 8- and 16-node cluster does
not have time to reach its top speed when ingesting the
smallest graph tested (which takes them 18 and 13 seconds,
espectively), ingest speeds hold steady even as the input
graphs grow to over a billion elements. Part of the reason
for this is our use of TokuDB 1 as the storage engine for
MySQL. Previous versions of Socrates used the standard
InnoDB engine, and suffered from severe I/O bottlenecks
when the graph grew above roughly 200 million edges,
likely due to InnoDB no longer being able to fit the interior
nodes of the B-Tree in its in-memory buffer pool. TokuDB
uses
cache-oblivious
lookahead
arrays
[
1
http://www.tokutek.com/products/tokudb-for-mysql/
Fig. 6: Socrates insertion speeds (on a logarithmic scale)
for an E-R graph with 10 million vertices and 100 million
edges, with varying cluster sizes. On our cluster, Socrates
has exhibited approximately linear ingestion speed-up as
nodes are added.
We can see that Socrates ingest speed scales linearly to at
least 16 nodes.
C. Parallel Processing
To measure our parallel processing capability, we use a
naive connected component algorithm implemented in the
Neighborhood parallelism model described above. On its
initial iteration, the algorithm assigns each vertex a
component attribute equal to the smallest vertex id among
itself and its neighbors. On subsequent iterations, the
algorithm examines the component attribute of itself and its
neighbors, and updates its component to be the smallest
value in the examined set. The algorithm terminates when
no vertex’s component changes in an iteration.
Not that this benchmark is not necessarily the fastest
method for computing connected components, however it is
useful as a benchmark because Socrates must fetch each
vertex’s in and out neighbors along with property
information for each node. Thus, the speed at which an
iteration of this algorithm runs can give us an idea of the
batch processing capability of the Neighborhood
parallelism model.
We have run this algorithm on the same graphs that were
ingested in the previous experiment, i.e., Erdos-Renyi
graphs consisting of connected components with 100
vertices and an average of 1000 edges each, varying in total
size from 1.1 million to 1.1 billion elements. Once again,
we have repeated the experiments using only 2, 4, and 8
nodes in our cluster in addition to the full 16.
Figures 7 and 8 present our results of these experiments.
The number of vertices processed per second, averaged
over all iterations of the algorithm after the initial iteration,
is on the y-axis. Note that processing a single vertex
involves fetching its immediate neighborhood (an average
of 10 edges and vertices), as well as the component
property for each vertex in the neighborhood.
Visualizations in the paper were produced using Pointillist,
a graph visualization software developed by J. Cohen. We
acknowledge D. Silberberg and L. DiStefano for providing
feedback and support during the development of Socrates.
REFERENCES
[1]
Fig. 7: Average processing speeds for an iteration of the
connected component algorithm on an ER graph with an
average of 10 edges per vertex.
These results are qualitatively similar to the ingest results.
Figure 7 demonstrates that, once the graph is large enough
to allow the 8- and 16-node clusters to reach full speed,
processing speed holds steady up to graphs of over a billion
elements.
Fig. 8: Average processing speeds (on a logarithmic scale)
for an E-R graph with 10 million vertices and 100 million
edges, with varying cluster sizes. Again, we see
approximately linear processing speed up as nodes are
added.
Figure 8 demonstrates that processing speeds also scale in
an approximately linear fashion up to a cluster of 16 nodes.
ACKNOWLEDGMENT
P. Gupta, A. Goel, J. Lin, A. Sharma, D. Wang, and R. Zadeh, “WTF:
The Who to Follow Service at Twitter,” in Proceedings of the 22Nd
International Conference on World Wide Web, Republic and Canton of
Geneva, Switzerland, 2013, pp. 505–514.
[2] S. Salihoglu and J. Widom, “GPS: A Graph Processing System,” in
Proceedings of the 25th International Conference on Scientific and
Statistical Database Management, New York, NY, USA, 2013, pp.
22:1–22:12.
[3] G. Malewicz, M. H. Austern, A. J. . Bik, J. C. Dehnert, I. Horn, N.
Leiser, and G. Czajkowski, “Pregel: A System for Large-scale Graph
Processing,” in Proceedings of the 2010 ACM SIGMOD International
Conference on Management of Data, New York, NY, USA, 2010, pp.
135–146.
[4] U. Kang, C. E. Tsourakakis, and C. Faloutsos, “PEGASUS: A PetaScale Graph Mining System Implementation and Observations,” in
Ninth IEEE International Conference on Data Mining, 2009. ICDM ’09,
2009, pp. 229–238.
[5] P. Kalmegh and S. B. Navathe, “Graph Database Design Challenges
Using HPC Platforms,” in High Performance Computing, Networking,
Storage and Analysis (SCC), 2012 SC Companion:, 2012, pp. 1306–
1309.
[6] D. Dominguez-Sal, P. Urbón-Bayes, A. Giménez-Vañó, S. GómezVillamor, N. Martínez-Bazán, and J. L. Larriba-Pey, “Survey of Graph
Database Performance on the HPC Scalable Graph Analysis
Benchmark,” in Proceedings of the 2010 International Conference on
Web-age Information Management, Berlin, Heidelberg, 2010, pp. 37–
48.
[7] J. P. Cedeño and K. S. Candan, “R2DF Framework for Ranked Path
Queries over Weighted RDF Graphs,” in Proceedings of the
International Conference on Web Intelligence, Mining and Semantics,
New York, NY, USA, 2011, pp. 40:1–40:12.
[8] R. Angles and C. Gutierrez, “Survey of Graph Database Models,” ACM
Comput Surv, vol. 40, no. 1, pp. 1:1–1:39, Feb. 2008.
[9] J. Hayes and C. Gutierrez, “Bipartite Graphs as Intermediate Model for
RDF,” in The Semantic Web – ISWC 2004, S. A. McIlraith, D.
Plexousakis, and F. van Harmelen, Eds. Springer Berlin Heidelberg,
2004, pp. 47–61.
[10] S. Jouili and V. Vansteenberghe, “An Empirical Comparison of Graph
Databases,” in 2013 International Conference on Social Computing
(SocialCom), 2013, pp. 708–715.
[11] N. Bronson, Z. Amsden, G. Cabrera III, P. Chakka, P. Dimov, H. Ding,
J. Ferris, A. Giardullo, S. Kulkarni, H. Li, M. Marchukov, D. Petrov, L.
Puzar, Y. J. Song, and V. Venkataramani, “TAO: Facebook’s
Distributed Data Store for the Social Graph,” USENIX Annu. Tech.
Conf. ATC, 2013.
| 8cs.DS
|
A Two-Stage Method for Text Line Detection in Historical Documents
Tobias Grüninga,∗, Gundram Leiferta , Tobias Straußa , Roger Labahna
a Computational
Intelligence Technology Lab, Institute of Mathematics, University of Rostock, 18057 Rostock, Germany
arXiv:1802.03345v1 [cs.CV] 9 Feb 2018
Abstract
This work presents a two-stage text line detection method for historical documents. In a first stage, a
deep neural network called ARU-Net labels pixels to belong to one of the three classes: baseline, separator
or other. The separator class marks beginning and end of each text line. The ARU-Net is trainable from
scratch with manageably few manually annotated example images (less than 50). This is achieved by
utilizing data augmentation strategies. The network predictions are used as input for the second stage
which performs a bottom-up clustering to build baselines. The developed method is capable of handling
complex layouts as well as curved and arbitrarily oriented text lines. It substantially outperforms current
state-of-the-art approaches. For example, for the complex track of the cBAD: ICDAR2017 Competiton
on Baseline Detection the F-value is increased from 0.859 to 0.922. The framework to train and run the
ARU-Net is open source.
Keywords: baseline detection, text line detection, layout analysis, historical documents, U-Net, pixel
labeling, semantic segmentation, state estimation
1. Introduction
Accessibility of the valuable cultural heritage of historical documents is an important concern of
archives, libraries as well as certain companies, e.g., those specialized in genealogy. After years of digitization at an industrial scale to protect and preserve these valuable goods, millions over millions of
scanned pages are stored at servers all over the world [1]. The generic next step is to make the enormous
amount of content of these document images accessible and enable humanists, historians, genealogists as
well as ordinary people to efficiently work with these documents. Besides the cost- and time-consuming
process of manually annotating volumes [2], it is subject to current research and scientific discussion how
to automate this process [3].
Since 2009, tremendous progress in the field of Automated Text Recognition1 (ATR) [4, 5] as well
as Keyword Spotting (KWS) [6, 7, 8] was achieved. The performance of state-of-the-art systems reaches
∗ Corresponding
author
Email addresses: [email protected] (Tobias Grüning), [email protected]
(Gundram Leifert), [email protected] (Tobias Strauß), [email protected] (Roger Labahn)
1 Optical Character Recognition + Handwritten Text Recognition
Preprint submitted to Elsevier
February 12, 2018
character error rates below 10% for ATR [9] and mean average precisions above 0.9 for KWS [10] for
complex handwritten documents. Although efforts are made to develop systems working solely on the
rough input image without any a-priori segmentation [11, 12, 13], the best performing recognition systems
– with reference to recently hosted competitions – rely on segmented words or text lines as input. Entirely
segmentation-free approaches suffer either from an enormous training/inference time and/or, up to now,
did not demonstrate its applicability with competitive quality on challenging datasets [10]. Hence, a
workflow which involves a text line extraction followed by the transformation of pixel information into
textual information (ATR/KWS) is the widely used standard. This work deals with the first step of the
information retrieval pipeline, namely the text line extraction. This is a mandatory step since errors
directly effect the performance of the overall information retrieval process. The text line extraction is
still unsolved to a certain extent for historical documents due to difficulties such as physical degradations
(e.g., bleed-through, faded away characters, heterogeneous stroke intensity), image capture conditions
(e.g., scan curve, illumination issues), complex layouts (e.g., structured documents, marginalia, multicolumn layouts, varying font sizes), arbitrary orientations and curved text lines.
The results achieved by state-of-the-art approaches are not satisfying [14], especially if dealing with
heterogeneous data. Therefore, this work focuses on the extraction of text lines in arbitrary historical
documents. Since different ATR/KWS systems necessitate different text line representations, e.g., bounding boxes [15], x-height areas [16] or more precise polygonal representations following all ascenders and
descenders [7], there is not the one correct text line representation. Therefore we limit ourselves towards
the text line detection task by representing each text line by its baseline. The detected baselines allow
for an extraction of the text lines in an appropriate – with respect to the following method – way. The
problem of extracting a text line given its baseline can be tackled by applying, e.g., histogram approaches
to estimate the x-height [16] or by utilizing Dynamic Programming to calculate separating seams [17].
Besides the classical image processing based approaches, deep learning based methods became omnipresent in the document analysis community within the last years. Such techniques were recently used
to solve several different problems such as binarization [18], page boundary extraction [19], page segmentation [20] or text line detection [16]. The presented work to our knowledge is the first which uses a
two-stage method, combining deep learning strategies and state-of-the-art image processing based techniques. We propose an extension of the U-Net [21], the so-called ARU-Net. The fully convolutional U-Net
is extended by incorporating residual blocks [22] to increase its representative power. Furthermore, a spatial attention mechanism is developed which allows the ARU-Net to focus on image content at different
positions and scales. The network is designed to processes the entire, arbitrarily-sized image at once to
take account of all spatial context. The ARU-Net is universal in a way, that it could be used to tackle any
pixel labeling task. In this work, it is trained in a fully supervised fashion to classify each pixel to belong
to one of the following classes: baseline, separator or other. The separator class is introduced to explicitly
2
predict beginning and end of each text line and not just rely on the information implicitly given by the
baseline class. This is advantageous for text lines which are close together but have to be separated,
e.g., those belonging to different columns. The network output serves as input for an image processing
based bottom-up clustering approach. This approach utilizes so-called states of superpixels [23], which
encode local text orientation and interline distances. This second stage allows for an error correction of
the network output by incorporating domain knowledge based on assumptions, which hold for text lines in
general, see Sec. 3.3.3. Additionally, it is easily possible to incorporate the separator information, which
allows for an handling of documents with complex layouts, e.g., images containing tables or marginalia.
Each method relying on supervised deep learning and therefore relying on training data can suffer
from the need of an enormous amount of labeled training data. We demonstrate that the presented
approach achieves high quality results on the Bozen dataset [24] with less than 50 full-page training
samples by using data augmentation strategies. Along with an annotating effort of just a few minutes
per page the adaptation of the proposed method is easy and cheap. We demonstrate the applicability
of the proposed method for images with arbitrarily oriented as well as curved text lines by achieving
nearly as good results as for straight 0° oriented text lines. Finally, we show that the presented approach
outperforms state-of-the-art methods on three different datasets. A relative F-value [25] error (the gap to
1.0) reduction of at least 24% is achieved for the cBAD dataset [26]. This dataset is composed of images
of nine different archives and libraries from all over Europe and is therefore – in the opinion of the authors
– the most representative and heterogeneous freely available dataset. Especially, for the complex track,
which contains mostly documents with complex layouts, the average F-value is increased from 0.859 to
0.922.
The main contributions of this work are:
• introduction of a newly designed deep neural network (ARU-Net) for pixel labeling along with a
meaningful parametrization – the ARU-Net and its training framework are open source2 ,
• introduction of the new concept of learned separators to handle complex layouts instead of an a-priori
page segmentation or white-/blackrun calculation
• introduction of a state-of-the-art two-stage workflow which combines state-of-the-art deep learning
and image processing techniques – the entire workflow is freely usable via the Transkribus platform3 .
2. Related Work
A comprehensive survey of approaches for text line extraction in historical documents is given in [27]
and [28]. In this section, we will focus on approaches relevant for this work.
2 https://github.com/TobiasGruening/ARU-Net
3 https://transkribus.eu
3
In [17, 29, 30], the principle of Dynamic Programming is utilized to calculate cost optimal paths
passing the image from left to right to separate different text lines from each other. These methods
basically differ in the way the images are pre-processed and in the definition of the cost function. Garz
et al. [31] propose a method based on clustering of interest points (this is just another name for what
we call superpixel). Using a standard clustering technique, interest points in an area which exceeds a
certain density are clustered to form word clusters. Word clusters are separated to sub-word segments
and these are finally grouped to build text lines. Ryu et al. [23] propose an algorithm which uses certain
characteristics (so-called states) of extracted connected components to assign costs to certain clustering
results. These states encode local text orientation and interline distances and are introduced in Def. 3.3.4.
Subsequently using four different operations (merge, split, merge-split, merge-merge-split) on an initial
coarse clustering, the costs are minimized to obtain an optimal clustering, which leads to the final text line
segmentation. Ahn et al. [32] improve this approach by the introduction of a newly developed binarization
method and an improved clustering process. Grüning et al. [33] extended the approach of Ryu et al. so
that it is applicable for more general superpixels with a newly introduced clustering procedure which
does not rely on a coarse initial clustering. Besides these “classical” approaches, which are based on
image processing techniques, methods based on machine learning gained importance within the last two
years. Moysset et al. [34] propose a method based on a recurrent neural network. The network is trained
given only the number of lines in the image utilizing Connectionist Temporal Classification which was
introduced to train networks for handwriting text recognition and allows for ground truth data without
any alignment. The trained neural network predicts confidences for the vertical coordinates of the image
to belong either to the classes line or interline. Further post-processing of the neural network output
is performed to detect the text lines. In follow-up works, they formulated the problem as a regression
problem [35, 36]. The recurrent neural network directly predicts bounding boxes as well as the start of
each text line, respectively. Besides this regression based approach, classification based approaches were
proposed most recently. In contrast to the approach of Moysset et al., these methods perform a pixel
labeling to classify each image pixel (instead of classifying rows of pixels, only). For instance, Renton
et al. [16] propose a fully convolutional network (FCN) based on dilated (or atrous) convolutions to
classify pixels as text line main body or not. The classification results are utilized to extract the text line
information. These techniques are currently very popular, e.g., four of the five participants of the cBAD:
ICDAR2017 Competition on Baseline Detection [37] use methods relying on FCNs.
3. Methodology
In this section, we introduce the two-stage method for baseline detection, see Fig. 3.1. The first stage
relies on a deep neural network – the ARU-Net – and performs a pixel labeling. The pixel labeling
can be seen as some kind of goal-oriented binarization. Instead of detecting all foreground elements, it
4
Stage II
Superpixel Clustering
Superpixel Calculation
Stage I
State Estimation
ARU-Net
Input
Output
Figure 3.1: Two-stage workflow to detect baselines – The first stage utilizes a deep hierarchical neural network to
perform a pixel labeling. The result of Stage I is the input for an image processing based method in Stage II. This method
clusters superpixel to build baselines. The image is sampled from the cBad complex test set [25].
restricts itself to those elements which are of interest for the specific task. The second stage performs a
superpixel (SP) extraction on the first stage’s output. These SPs are further clustered to build baselines.
In the following, the problem of baseline detection is formulated. Afterwards, a detailed description of
the proposed ARU-Net is given. Finally, the SP extraction and clustering approach is described.
3.1. Problem Statement
We will introduce the problem of baseline detection in a formal way by defining all necessary termini
and notation. Within this work we follow the definition of a baseline given in [26]:
Definition 3.1.1 (baseline). A baseline is defined in the typographical sense as the virtual line where
most characters rest upon and descenders extend below.
h×w
Definition 3.1.2 (image, pixel, intensity). A matrix I ∈ [0, 1]
is called (gray-scale) image of
height h and width w. A pair p = (y, x) ∈ {1, ..., h} × {1, ..., w} is called pixel of I. The matrix value
I(p) = I(y, x) = Iy,x in row y and column x is called intensity of pixel p.
Image means gray-scale image for the rest of this work. Ih denotes the height of image I, Iw denotes the
width, analogously. For visualization purposes a pixel intensity value of 1 means white and 0 means black.
If the colored image is available, we usually use this one for visualization even though it is converted to
its gray-scale version for calculations.
Definition 3.1.3 (coordinate). Let p = (y, x) be a pixel. py = y and px = x denote the elements of
the first and second dimension of p and are called coordinates (y−coordinate and x−coordinate).
5
Definition 3.1.4 (image space). The set of all possible images
I=
[
[0, 1]
h×w
(3.1)
h,w∈N
is called image space.
Definition 3.1.5 (polygonal chain, closed). A polygonal chain of length n ∈ N is an n-tuple of pixels
P = (p1 , p2 , ..., pn ) .
(3.2)
A polygonal chain is called closed iff p1 = pn holds.
Taking into account Def. 3.1.1, each baseline can be represented by a polygonal chain.
Definition 3.1.6 (polygonal chain space). The infinite set P of all possible polygonal chains is called
polygonal chain space.
Definition 3.1.7 (baseline detector, baseline hypothesis). We call a function b : I → P(P) which
maps each image to a subset of P a baseline detector. The set of all baseline detectors is denoted by B.
The output of b for a certain image I is called baseline hypothesis.
Definition 3.1.8 (baseline ground truth). The set GI ⊂ P of polygonal chains representing the baselines of an image I (possibly annotated by a human operator) is called baseline ground truth (for image
I).
Def. 3.1.1 allows for some baseline variety. Hence, there is not the one unique and correct ground truth
for an image. Therefore, ground truth information is always biased by its creator. This has to be taken
into account for the evaluation process as well as for the baseline detector design.
Definition 3.1.9 (similarity score). A function h·, ·iµ : P × P → [0, 1] assigning a scalar value to each
pair of baseline ground truth and baseline hypothesis polygonal chain sets is called similarity score.
A value of 1.0 indicates that two polygonal chains are regarded as equal. Within this work we follow
the similarity score introduced in [25]: We measure the accuracy of a baseline detector in terms of the
F-value, see [25] for a detailed introduction.
The problem tackled in this work can now be formulated as follows: Suppose there are two sets of
images along with their baseline ground truth information
Ttrain = {(I, GI )i | i = 1, ..., n},
Ttest = {(I, GI )i | i = 1, ..., m}.
6
(3.3)
We aim for a design of a baseline detector b∗ given Ttrain which solves
b∗ = arg max
b∈B
X
hGI , b(I)iµ .
(3.4)
(I,GI )∈Ttest
In the design phase of b∗ the set Ttest is unknown and one is allowed to use solely Ttrain . Hence, one has
to ensure that b∗ generalizes well from Ttrain to Ttest .
Since the proposed design consists of two stages and the first stage relies on deep learning techniques,
an adaptation to a differently biased ground truth (produced by a different annotator) can be done easily
by retraining the first stage without any fine tuning done by experts.
3.2. Stage I: ARU-Net
Typically, layout analysis algorithms directly work on the input image I or on a binarized version of it
[17, 23, 29, 30, 31, 33]. Instead, we employ a more goal-oriented transformation of the input image utilizing
a neural network, which is trained in a supervised manner to assign a certain class to each pixel like in
[21, 38, 39]. This is often referred to as pixel labeling or semantic segmentation. We will introduce the
problem of pixel labeling utilizing hierarchical neural networks, followed by a description of the proposed
ARU-Net architecture.
3.2.1. Pixel Labeling – Problem Formulation
Definition 3.2.1 (neural pixel labeler). A neural pixel labeler (NPL) for the classes C = {c1 , ..., cn }
is a hierarchical neural network Φ( · ; w) : I → I |C| . The NPL is parametrized by w ∈ RN . For I ∈ I it
performs a prediction over all pixels and all possible classes
Φ(I; w) = C ∈ [0, 1]
Ih ×Iw ×|C|
subject to
|C|
X
C(y, x, c) = 1
∀y, x.
(3.5)
c=1
C(:, :, c) = C:,:,c ∈ I denotes the image which encodes the pixel-wise prediction (probability) for the c−th
class.
Definition 3.2.2 (pixel ground truth). A cartesian product GI ∈ I |C| is called pixel ground truth (for
image I) if it assigns exactly one class (one-hot-encoding) to each pixel such that
∀y, x ∃!c ∈ {1, ..., |C|} :
GI (y, x, c) = 1 ∧ GI (y, x, c̃) = 0 (∀c̃ 6= c).
(3.6)
Following the problem formulation of Section 3.1, we aim for an NPL, which was tuned on a training set
and optimally performs on a test set. Assume there are training and test sets in the style of Eq. (3.3) –
7
with pixel ground truth information instead of baseline ground truth information
Tetrain = {(I, GI )i | i = 1, ..., n},
Tetest = {(I, GI )i | i = 1, ..., m}.
(3.7)
The performance of an NPL is evaluated in terms of the cross-entropy between the predicted and the
ground truth distribution. The cross-entropy can also be motivated by a maximum likelihood estimation.
This results in the cross-entropy loss function.
Definition 3.2.3 (loss function). Let Te be a set of images along with their pixel ground truth and
Φ( · ; w) is an NPL. The performance of Φ on Te is evaluated in terms of the (cross-entropy) loss function
L(Φ, Te ) = −
X
|C|
Ih X
Iw X
X
GI (y, x, c) ln Φ(I; w)y,x,c .
(3.8)
(I,GI )∈Te y=1 x=1 c=1
To improve the performance of the NPL on the training set, one can calculate the loss function’s gradient
with respect to the model parameters using the well-known technique of backpropagation [40]. The
gradient is used to adapt the model parameters by gradient descent
w ←w−τ ·
∂L
(Φ, Tetrain )
∂w
(3.9)
with a learning rate τ . This is repeated to successively adapt the NPL. The process of adapting the
model by minimizing its loss is called training. Since one does not aim for a minimization of the loss
on the training set, the system has to generalize to achieve high quality results on the test set as well.
To stabilize training, avoid over-fitting, improve generalization, ... dozens of techniques to improve the
simple rule in Eq. (3.9) were introduced within the last years. Since the introduction of these is beyond
the scope of this work, we refer to [41]. Details on techniques used within this work are given in Sec. 4.
3.2.2. ARU-Net – Architecture
The ARU-Net is a special form of an NPL and is described in this section. We omit a formal introduction of the used neural network components and concepts and refer to the above mentioned literature.
Within the last few years, different architectures were proposed for the pixel labeling task. Most of them
are based on Convolutional Neural Networks (CNNs) [42]. A direct application of CNNs for semantic
segmentation is presented in [38]. The presented Fully Convolutional Network (FCN) combines local
features to produce more meaningful high level features using pooling layers. Pooling reduces the spatial
dimension. Thus, the result suffers from a coarse resolution. Noh et al. [39] tackle this problem by
applying a deconvolutional network on the subsampled output of the FCN. The U-Net proposed in [21]
furthermore introduces shortcuts between layers of the same spatial dimension. This allows for an easier
combination of local low level features and global higher-level features. Additionally, error propagation
8
for deep structures is facilitated and the so-called vanishing gradient problems [43] are reduced. The
U-Net is the basis for the proposed ARU-Net. We extend the U-Net by two more key concepts – spatial
attention (A) and depth (residual structure (R)) to be described below. Remarkably, in contrast to the
U-Net proposed in [21], we perform border padding. Hence, the spatial dimensions in each scale space of
the U-Net are all the same, see Fig. 3.2 for a schematic representation of an U-Net. The output of the
U-Net thus is a feature map (Z features in Fig. 3.2) of the same spatial dimension as the input. Hence, the
U-Net becomes an NPL as defined in Def. 3.2.1 by adding a convolutional (to get pixel-wise predictions)
softmax classifier on top which distinguishes between the different classes of C.
Remark 3.2.1. If the presented architectures are used for the pixel labeling task, it is implicitly assumed
that such a classifier is always added to generate per class probabilities at pixel level.
Z
2Z
Identity
2Z
Z
Z
Output
I
Z
Input
1
2Z
4Z
2Z
Identity
II
4Z
4Z
8Z
4Z
Identity
III
8Z
IV
Conv+Act
Max Pool
Deconv+Act
2 Layer CNN Block
Figure 3.2: U-Net – The input is an image of arbitrary spatial dimension. ”Act” is the activation function thus the
rectangles represent sets of activation maps. Each rectangle represents a 3-dim array (∈ Rh×w×z ). Within each scale space
(roman numbers) the feature map widths and heights are constant (encoded by the height of the rectangles). The number
of feature maps Z is pictured by the width of the rectangles. Between adjacent scale spaces the spatial dimension decreases
by a certain factor (2 in the figure) and the representative depth (number of feature maps) increases by the same factor.
He et al. [22] introduce very deep neural networks which are still trainable and yield state-of-the-art
results. This is achieved using so-called residual blocks. Residual blocks introduce shortcuts, which
enable the error backpropagation and identity propagation even for very deep structures. Hence, the
vanishing gradient problems are reduced [22]. There are various different forms of residual blocks. The
one used within this work is depicted in Fig. 3.3.
Definition 3.2.4 (RU-Net). An RU-Net is an U-Net with residual blocks.
That means, each of the 2 layer CNN blocks in Fig. 3.2 is replaced by a residual block as in Fig. 3.3.
To explicitly incorporate the potential to handle various font sizes, especially mixed font sizes on a
single page, we introduce a pixel-wise (spatial) attention mechanism. For this purpose, we introduce an
9
Z
Z
Z
+
Conv+Act
Conv
Identity
Act
Output
Z
Input
Z
Logits
Figure 3.3: Residual Block – The input is convolved and the resulting 3-dim array (the maps before passed through an
acitvation function are referred to as logits) is used twice. At the first branch it is passed through the activation function
and further processed by several convolution layers. At the second branch it is directly fed into a summation node. After
a point-wise summation of the two logit maps an activation function is applied. The shortcut enables for an easy identity
propagation and error backpropagation. Arbitrarily many inner layers are possible.
attention network (A-Net). The A-Net is a multi-layer CNN which generates a single output feature map.
The A-Net will be applied along with the RU-Net at different scales, the same network weights are used
on all scales (weight sharing). Specially, a scale pyramid is built by downscaling the input image I = I1
several times. The resulting (scaled) images I1 , I2 , I4 , I8 , ..., Is (subscripts denote the scaling factors)
are fed into the RU-Net and the A-Net. Trainable deconvolutional layers (of corresponding scales) are
applied on the outputs of the RU- and the A-Net to obtain feature maps of spatial dimensions equal to
the inputs. A1 , ..., As denote the up-sampled feature maps of the A-Net, RU1 , ..., RUs of the RU-Net,
respectively. After applying a pixel-wise softmax normalization for the attention maps
bi (y, x) = P
A
exp(Ai (y, x))
j∈{1,2,...,s} exp(Aj (y, x))
(3.10)
bi sum to one (pixel-wise). The feature maps RUi are combined following
the normalized attention maps A
ARU =
X
RUi
bi ,
A
(3.11)
i∈{1,2,...,s}
where
is the Hadamard product (element-wise multiplication). ARU is the input for the classifier to
build a NPL, see Rem. 3.2.1.
Definition 3.2.5 (ARU-Net). An RU-Net incorporating the described spatial attention mechanism is
called ARU-Net, see Fig. 3.4.
The point-wise multiplication combined with the pixel-wise attention maps allow the ARU-Net to pay
attention in different scales at different positions of the image. In Fig. 3.4 one can see that this behavior
was indeed learned by the network. It seems like the RU-Net is specialized on a certain font size and the
A-Net distinguishes between areas of different font sizes (bright and dark areas).
The ARU-Net as introduced can be used for any pixel labeling task, e.g., binarization, page detection
10
A-Net
A-Net
RU-Net
Classification
Softmax
RU-Net
+
Deconv (4)
A-Net
Deconv (2)
RU-Net
Figure 3.4: ARU-Net – The input image and its downscaled versions are fed into the A-Net and R-U-Net
accross different scales). The results for the lower resolutions are deconvolved. The attention maps are
a softmax normalization. The brighter the map at a certain position the more attention is paid to that
corresponding scale. The attention maps are point-wise muliplied with the feature maps of the RU-Net.
summed and a classification is performed.
(weight sharing
passed through
position at the
The results are
and page segmentation. The purpose of the ARU-Net is defined and fixed by the number of classes and
the ground truth data provided for training. In this work, we limit ourselves to the baseline detection
problem introduced in Sec. 3.1. For this purpose, we introduce three different classes: baseline (bl),
separator (sep) and other (∅). The separators mark beginning and end of each text line. Although, the
separator information is implicitly encoded by the baselines, it is advantageous to explicitly introduce
it as possible classification result. Especially, for baselines which are close together, e.g., such belonging
to two adjacent columns, this approach helps to avoid segmentation errors. Pixel ground truth for the
classes C = {bl, sep, ∅} is automatically generated by Alg. 1 given the baseline ground truth of Eq. (3.3).
A sample image with baseline ground truth along with its generated pixel ground truth is depicted in
Fig. 3.5. The prediction of a trained ARU-Net for this sample image is shown in Fig. 3.6a.
3.3. Stage II: Baseline Estimation
This subsection describes the second stage of the proposed approach. Baselines are estimated given
the output of the ARU-Net. This task consists of three steps: superpixel calculation, state estimation
and superpixel clustering, which are described in the following.
Ih ×Iw ×3
The trained ARU-Net generates an output C ∈ [0, 1]
for each image I ∈ I. In the following
B = C:,:,1 denotes the image encoding the confidence of each pixel belonging to a baseline and S = C:,:,2
is the separator image, see Fig.3.6a
3.3.1. Superpixel Calculation
The number of all pixels in an image often exceeds several millions. To reduce the dimensionality of
the problem (the number of pixels to be regarded for the baseline estimation), we limit ourselves to a
11
Algorithm 1: Pixel Ground Truth Generation
1
2
3
4
5
6
7
8
9
10
11
12
input : image I, corresponding baseline ground truth GI
output: pixel ground truth GI
B, S, N ← 0
B of dimension Ih × Iw
for P = (p1 , ..., pn ) ∈ GI do
θ ← local text orientation of P
B see Def. 3.3.2
d ← interline distance of P
B see Def. 3.3.3
P b ← polygonal chain of length d and orient. θ + 90° centered at p1
P e ← polygonal chain of length d and orient. θ + 90° centered at pn
draw P b and P e in S
B draw: follow the chain and set pixel values to 1.0
draw P in B
E ← 3 × 3 matrix of ones
S ←S⊕E
B ← B ⊕ E ∧ ¬S
N ← ¬S ∧ ¬B
return : GI ← [B, S, N ]
B ⊕ morphological dilation
subset of all pixels.
Definition 3.3.1 (superpixel). Let S = {p1 , ..., pN } be a subset of the image pixels of I (typically,
N Ih · Iw holds). An element of S is called superpixel (SP).
Basically, the definition of a superpixel does not introduce any new concept. A SP is just a normal pixel
which is somehow regarded to be of certain importance. Since it is frequently used term, we decided to
introduce it via a definition. It is easy to see that the choice of the set of SPs is crucial for the overall
performance. If there are no SPs for a baseline at all, this baseline will be missed. To calculate a suitable
set of SPs, we utilize the baseline map B generated by the ARU-Net.
In a first step B is binarized Bb = B > b by an element-wise comparison of B with a confidence
threshold b. The morphological skeleton Bs = SKE(Bb ) is calculated for Bb following Lantuéjoul’s formula
[44]. All foreground pixels (pixels with an intensity of 1) of Bs build an initial set of pixels {p1 , ..., pM }.
Its elements are sorted (π : N → N) in descending order w.r.t. their baseline confidences
pπ(1) , ..., pπ(M ) : B(pπ(i) ) ≥ B(pπ(j) ) ⇔ i ≤ j.
(3.12)
Finally, S is set up by iteratively adding pixels of the sorted list of Eq. (3.12) (beginning with the first
pixel). To keep the number of SPs small, a new pixel p is added to S only if
kp − qk2 > d ∀q ∈ S
(3.13)
holds, otherwise it is skipped. In Fig. 3.6b the set of resulting SPs is shown. These SPs build the basis
for the further clustering.
Remark 3.3.1. For all experiments, we have chosen fixed values of b = 0.2 (binarization threshold) and
12
(a) Baseline ground truth – The baselines are described by the red dots. For better clarity dots of the same baseline
were connected.
(b) Pixel ground truth produced by Alg. 1 – Green encodes the separator class, red the baseline class and black
the ”other” class.
Figure 3.5: Baseline and pixel ground truth – These are shown for the top snippet of the image of Fig. 3.1.
d = 10 (Eq. (3.13)). These demonstrated to be well suited for a wide range of different scenarios. Hence,
they are not regarded as free parameters of the system which have to be further tuned. This also holds
for the parameters which are fixed in Rem. 3.3.4 & 3.3.8.
3.3.2. Superpixel State Estimation
Assume we can assign each SP to a certain text line. The state of an SP should encode meaningful
characteristics of its text line. These characteristics will be defined and combined to build the state. This
work is based on previous work of [23, 33], but adapted to the characteristics of SPs extracted given the
ARU-Net output, e.g., easier calculation of the local text orientation as well as a different smoothing cost
formulation.
Definition 3.3.2 (local text orientation). The local text orientation θ of an SP p is the slope of its
text line’s baseline at the coordinates closest (w.r.t. the euclidean distance) to p.
Definition 3.3.3 (interline distance). The interline distance s of an SP p is the distance of its text
line’s baseline to the nearest other baseline. Distance means the distance which is orthogonal to the local
text direction of p.
Definition 3.3.4 (state). The state of an SP is the pair (θ, s) of its local text orientation and its interline
distance.
In the following, we will describe a method to estimate the states of all SPs. The local text orientation
will be calculated in a straightforward way utilizing solely the baseline image B and local information.
13
(a) ARU-Net output – The estimated baselines B (blue) and separators S (cyan) are shown.
(b) Superpixel and neighborhood system – The calculated SPs (blue) are shown along with the resulting Delaunay
neighborhood system N (yellow).
Figure 3.6: Baseline detection process – Two intermediate steps are shown for the top snippet of the image of Fig. 3.1.
On the other hand, the estimation of the interline distances combines local information of the text line’s
periodicity with the more global assumption that nearby SPs tend to have similar interline distances. For
these approaches the concepts of neighborhood and connectivity are mandatory and will be introduced.
Definition 3.3.5 (neighborhood system, edge, adjacent). We call a subset N ⊂ S×S neighborhood
system. An element of N is called edge and denoted by ep,q . N is not directed (ep,q = eq,p ). Two SPs
p, q are adjacent if ep,q ∈ N . ep,q \ p ∈ S denotes the SP q.
Remark 3.3.2. In the following the neighborhood system N for a set of SPs is always calculated by
Delaunay’s triangulation [45].
Definition 3.3.6 (connectivity function). The line segment g( · ; ep,q ) : [0, 1] → R2 defined by
g(τ ; ep,q ) := p + τ (q − p) connects the two pixels p, q of the edge ep,q . The function Γ : N × I → [0, 1]
defined by
R1
Γ(ep,q , I) =
0
I(g(τ ; ep,q ))dτ
kp − qk2
(3.14)
is called connectivity function. I(g(τ ; ep,q )) denotes the intensity of the pixel in I closest (w.r.t the
euclidean distance) to the real-valued coordinates g(τ ; ep,q ).
The connectivity function calculates the average intensity for a given image along the shortest path
connecting two pixels. The local text orientation of each SP is estimated by θp = LTO(p; N , B) utilizing
N and the baseline image B, see Alg. 2. The LTO algorithm picks the two neighbors of a SP p with the
largest baseline connectivity to p and determines the slope of the line passing through these neighbors.
14
Algorithm 2: Local Text Orientation of p
1
2
3
4
5
6
input : SP p, neighborhood system N , baseline image B
output: local text orientation θ of p
M ← {eq,r ∈ N | q = p ∨ r = p}
L ← sorted list of M
B sorted by means of Γ(eq,r , B)
if |L| == 1 then
eq,r ← L1
B Lk denotes the k-th element of L
else
eq,r ← (L1 \ p, L2 \ p)
r −q
return : θ ← arctan rxy −qy
x
The periodicity of text lines in document images is utilized to calculate the interline distances. We
determine the interline distance of a SP p by evaluating the regional text-line periodicity around p as
follows. For an SP p, a circular region of diameter d ∈ N around p, and a projection direction determined
by the local text orientation θp , let
p,d
d
hp,d = (hp,d
1 , ..., hd ) ∈ N
(3.15)
be the projection profile with respect to S, see Fig. 3.7. For the calculation of hp,d , only SPs with a
distance to p of less than
d
2
are taken into account.
Remark 3.3.3. The projection profile hp,d can be calculated very efficiently by utilizing the cross product
T
# » for q ∈ S with kp − qk ≤ d .
of the orientation vector o = (cos(θ ), sin(θ )) and the vectors pq
p
p
2
2
To extract the regional periodicity inherent in the projection profile hp,d , a Discrete Fourier Transformation (DFT) is applied to hp,d with resulting coefficients
H p,d = H1p,d , ..., Hdp,d ∈ Cd .
(3.16)
A coefficient Hkp,d , k ∈ {1, ..., d} corresponds to the portion of the signal with a period of
p,d
signal h
0
. In the simplest case, the index k of the dominant coefficient of H
distance s of p as s =
d
k0 .
p,d
d
k
to the entire
determines the interline
However, we may be forced to assign a different value to s due to additional
constraints to be discussed in a moment. Therefore, we introduce a data energy value for each possible
value
d
k
of the interline distance s of p. From energy, we then derive a data cost to be used within a cost
minimization framework for finding the optimal interline distance.
Definition 3.3.7 (data energy, data cost). The data energy of SP p and interline distance
15
d
k
is given
Figure 3.7: Interline distance estimation – Illustration of several projection profiles for a certain SP (red point). The
profiles for different diameters d ∈ {64, 128, 256, 512} and an orientation of 0° are shown in green. The winning period
(interline distance) is drawn as yellow curve. In blue a histogram for a wrong orientation (45°) is shown.
by
d
=
Ep
k
Hkp,d
H p,d
2
2.
(3.17)
2
The corresponding data cost is calculated by
Dp
d
d
= − log Ep
.
k
k
(3.18)
Remarkably, the data energy is normalized such that it sums (over k) up to 1.0 for arbitrary d ∈ N. To
cover a suitable range of different interline distances as well as to be robust against disturbances due
to close-by text regions of a different style, the projection profiles and DFTs are calculated for different
diameters d ∈ {64, 128, 256, 512} and k ∈ {3, 4, 5}. The choice of the values for d and k is application
driven and results in reasonable interline distances (sorted list) of
S := (170.7, 128.0, 102.4, 85.3, 64.0, 51.2, 42.7, 32.0, 25.6, 21.3, 16.0, 12.8) .
In the following, we write sp for the assigned interline distance s =
d
k
(3.19)
∈ S of SP p and say p is labeled
with sp . A labeling {sp }p∈S of S assigns an interline distance to each SP of S. Following a greedy labeling
strategy by assigning the interline distance with the highest energy defined by Eq. (3.17) to each SP leads
to a noisy result, see Fig. 3.8a. To reduce the noise effects, the influence of close-by SPs is taken into
account. It is reasonable to expect that neighboring SPs tend to have similar interline distances. This
expectation is encoded via a smoothing cost defined for adjacent SPs.
Definition 3.3.8 (smoothing cost). For each ep,q ∈ N (and assigned interline distances sp , sq ) the
16
(a) Greedy states – The SP states for a greedy labeling using the highest energy of Eq. (3.17) are shown.
(b) Smoothed states – The SP states for the final labeling after minimizing Eq. (3.21) are shown.
Figure 3.8: SPs with their assinged states – The local text orientation of each SP is visualized by the orientation of the
green lines (rotated by 90°). The length of the lines encode the interline distance of the corresponding SP.
smoothing cost is defined by
Vp,q (sp , sq ) =
σ
, hsp , sq is ≥ 4,
hsp , sq is
, else.
(3.20)
hsp , sq is is the index difference of sp and sq in the sorted list S of Eq. (3.19), e.g., h16.0, 42.7is = 4.
Thus, the smoothing cost Vp,q (sp , sq ) becomes large if interline distances of different size are assigned to
adjacent SPs. A maximum cost value of σ is used for huge differences in the interline distances. Setting
σ to a large value prevents neighboring SPs to differ to much in their interline distances.
Definition 3.3.9 (labeling cost). The labeling cost is given by
X
X
C {sp }p∈S = αd
Dp (sp ) + αs
Vp,q (sp , sq ).
(3.21)
ep,q ∈N
p∈S
The data cost and the smoothing costs are weighted by αd and αs , respectively, to form the labeling cost.
The graphcut algorithm [46] is utilized to minimize Eq. (3.21). The final labeling is shown in Fig. 3.8b.
Remark 3.3.4. For all experiments, we have chosen fixed values of σ = 25, αd = 1 and αs = 1.
3.3.3. Superpixel Clustering
In the previous subsections the calculation of SPs and their enrichment with state information was
described. In a final step, this state information is utilized to cluster the SPs to build baselines. There
17
will be a one-to-one assignment between clusters and baselines. In the following, we call a set of SPs
cluster.
In this subsection we formulate the clustering problem and introduce a greedy clustering procedure to
solve the problem. Two assumptions which hold for baselines in general constitute the conditions for the
clustering problem:
(I) Baselines should not exceed a certain curvilinearity value.
(II) Within the interline distance of a baseline there are no other baselines.
Basically, assumption (I) claims that a baseline can be approximated by a polynomial function of a certain
degree, see [23]. Assumption (II) is self-explanatory.
Remark 3.3.5. In the following, θ({p1 , ..., pn }) denotes the average orientation and s({p1 , ..., pn }) the
average interline distance of all SPs in {p1 , ..., pn }.
Definition 3.3.10 (curvilinearity value). Let deg ∈ N and S be a set of SPs. Assume pS,deg (t) ∈ P [t]
is the polynomial which solves the linear regression problem in the monomials t0 , t1 , ..., tdeg for the rotated
pixels
cos(−θ(S))
S0 =
sin(−θ(S))
− sin(−θ(S))
cos(−θ(S))
·p
p∈S
⊂ R2 .
(3.22)
The root-mean-square regression error normalized by s(S) is called curvilinearity value of S and is denoted
by cur(S, deg).
Remark 3.3.6. We fix deg = 3 and omit it in the following.
Def. 3.3.10 allows for an easy evaluation of (I). To test for (II) we will introduce the distance of two clusters.
Remarkably, only distances orthogonal to the text orientation should be taken into account. First, the
orthogonal component of the distance between two SPs is introduced. Afterwards, this is generalized for
two clusters of SPs.
Definition 3.3.11 (off-text distance). Given two SPs p, q and an orientation θ, the off-text distance
of p and q is the length of the component of p − q ∈ R2 which is orthogonal to θ. It is denoted by
kp − qkθ .
Remark 3.3.7. The off-text distance can be efficiently calculated by
kp − qkθ = (px − q x ) sin(θ) − py − q y cos(θ) .
18
(3.23)
Calculating the minimal pairwise off-text distance off all SPs of two clusters could result in a cluster
distance distorted by SP outliers. Therefore, SPs in each cluster will be projected onto the corresponding
regression curve obtained by the regression problem in Def. 3.3.10, before taking pairwise distances.
Definition 3.3.12 (regression curve). Let S, S 0 and pS (t) be of Def. 3.3.10. The spatial t-range of
S 0 is given by tmin = min{px | p ∈ S 0 } and tmax = max{px | p ∈ S 0 }. A curve cS : [0, 1] → R2 which
results from rotating the graph of pS (t) for t ∈ [tmin , tmax ] by θ(S) is called regression curve of S.
The SPs in S are projected onto cS (in direction θ(S) + π2 ). The resulting projected SPs are denoted by
S c . To achieve robust distance estimates even for curved and differently slanted text lines we focus on
SPs of the different clusters which are quite close to each other and furthermore take into account the
slope of the regression curve at the specific SP positions instead of averaging over the entire text line.
Definition 3.3.13 (cluster distance). Assume two clusters S1 , S2 with regression curves cS1 (t), cS2 (t)
and projected SPs S1c , S2c . The cluster distance is defined as
d(S1 , S2 ) =
min
p∈S1c ,q∈S2c :
kp−qk2 <4·s(S1 ∪S2 )
kp − qkθc (p,q) ,
(3.24)
where θc (p, q) is the average slope of the corresponding regression curves at p and q, respectively.
Since, it is now possible to evaluate conditions (I) & (II), we will use this to introduce feasible sets of
clusters. For this purpose, we will limit ourselves to partitions (a special kind of cluster sets) and require
the baseline clusters to be N -linked.
Definition 3.3.14 (partition). Let M be a set. We call a set of subsets {M1 , ..., MN } of M a partition
of M iff M = M1 ∪ · · · ∪ MN ∧ Mi 6= ∅ ∀i ∧ Mi ∩ Mj = ∅, i 6= j. The set of all partitions of M is
denoted by par(M).
Definition 3.3.15 (N -linked). Let S be a cluster and N be a neighborhood system. S is N -linked iff
∀p, q ∈ S ∃p0 , ..., pN ∈ S : p0 = p ∧ pN = q ∧ epi ,pi+1 ∈ N (0 ≤ i ≤ N − 1)
(3.25)
holds.
Definition 3.3.16 (feasible). For γ, δ ∈ R+ , L ∈ N, a set of SPs S and a neighborhood system N , we
call a set of clusters P = {S0 , ..., SL } feasible iff
1. P ∈ par(S)
2. ∀i > 0 : Si is N -linked
3. conditions (I) and (II) hold:
19
• cur(Si ) < γ
∀i > 0
• d(Si , Sj ) > δ · max{s(Si ), s(Sj )}
∀i, j > 0, i 6= j.
The set of feasible sets of clusters is denoted by f easN (S).
The clusters Si , i > 0 identify the baselines, S0 constitutes the clutter cluster containing SPs not belonging
to any baseline. We identify the baseline corresponding to Si with the polygonal chain of the projected
SPs Sic which follow the regression curve cSi (t), see Fig. 3.9. The number L ∈ N of baselines is (a-priori)
unknown. In the following, we will incorporate domain knowledge to promote SPs belonging to different
baselines not to be N -linked. Hence, clusterings with erroneously connected baselines are not feasible
anymore. This is done by a modification of the neighborhood system N .
Since baselines of different text orientations should not contribute to the same cluster, we adjust the
initial neighborhood system N by removing edges ep,q of SPs with substantially different local orientations:
|θp − θq | mod π >
π
4.
In addition, it is an ease to incorporate layout information by further adjusting
N . The layout information encoded by the separator image S (Fig. 3.6a) can be incorporated by taking
into account the connectivity of SPs in S. All edges ep,q ∈ N for which a separator is crossed, i.e.,
Γ(ep,q , S) > η or maxτ S(g(τ ; ep,q )) > 2 · η (g of Def. 3.3.6) holds, are removed, see Fig. 3.9b.
Finally, a common scenario is the baseline detection with given text regions. We assume that the
text regions are represented by closed polygonal chains R1 , ..., RN . This additional layout information (if
available) is easy to integrate. All edges for which
@Ri : Ri contains p, q.
(3.26)
holds are removed. Roughly speaking, a closed polygonal chain contains a SP if for all ”ways” from the
SP to the image border one have to cross the polygonal chain. Hence, SPs which are part of different nonoverlapping text regions are not N -linked any more. Thus, each baseline Si , i > 0 is entirely contained
in one text region for all feasible sets. The resulting neighborhood system is still denoted by N .
Remark 3.3.8. For all experiments, we have chosen fixed values of γ = 0.3, δ = 0.5 (Def. 3.3.16) and
η = 0.125.
After reducing the neighborhood system, we now introduce the total baseline energy. We will assign an
energy to all feasible sets and aim for an optimal one. This allows for the formulation of the clustering
problem to be solved.
Definition 3.3.17 (total baseline energy). Let B be a baseline image, N a neighborhood system and
P = {S0 , ..., SL } a set of clusters over S. With N (Si ) = {ep,q ∈ N | p, q ∈ Si } ⊂ N the total baseline
20
(a) Without separator information – The entire neighborhood system (yellow) is shown.
(b) With separator information – The neighborhood system was reduced by removing edges (cyan) with high
separator connectivity. The corresponding separator information is illustrated in Fig. 3.6a.
Figure 3.9: Influence of the separator information – The resulting baselines (blue lines) with and without taking into
account the separator information are shown.
energy is defined by
b(P) =
L
X
X
Γ(ep,q , B).
(3.27)
i=1 ep,q ∈N (Si )
Finally, the clustering problem can be formulated as
P∗ =
arg max b(P).
(3.28)
P∈f easN (S)
Because there could be a huge number of feasible sets of clusters for large S, we introduce a greedy clustering algorithm to solve Eq. (3.28). The proposed algorithm clusters edges of N instead of clustering SPs.
If an edge is assigned to a cluster (set) of edges, we assign both corresponding SPs to the corresponding
cluster of SPs. In a first step, the set of edges in N is sorted in decreasing order w.r.t.
1−
kp − qkθ({p,q})
kp − qk2
!
· Γ(ep,q , B).
(3.29)
The sorted list is denoted by N . Eq. (3.29) takes into account the B-connectivity value of an edge and
discounts it if ep,q is rather orthogonal to θ({p, q}). Discounted edges are less likely part of a baseline
and are therefore sorted to the end of the list. This avoids that these edges are falsely assigned to baseline
clusters which are composed of just a few correct edges (statistics of the cluster are not reliable, yet).
Given S and N , the proposed clustering process is shown in Alg. 3.
21
Algorithm 3: SP Clustering
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
input : Set of SPs S and sorted list of edges N
output: optimized partition P ∗
S0 ← S, P ∗ ← {S0 }, n ← 0
while |N | =
6 n do
n ← |N |
for ep,q ∈ N do
B Four possible cases dependent on the SPs
if ∃i > 0 : p, q ∈ Si then
B Case 1: add edge to existing cluster!
N ← N \ {ep,q }
else if p, q ∈ S0 then
B Case 2: create new cluster?
if kp − qkθ({p,q}) < δ · s({p, q}) then
B γ, δ of Rem. 3.3.8
N ← N \ {ep,q }, S|P ∗ | ← {p, q}, S0 ← S0 \ S|P ∗ |
P ∗ ← P ∗ ∪ {S|P ∗ | }
else if w.l.o.g. p ∈ S0 ∧ ∃i > 0 : q ∈ Si then
B Case 3:
if cur(Si ∪ {p}) < γ ∧ d(Si , {p}) < δ · s(Si ) then
if d(Si ∪ {p}, Sj ) > δ · s(Sj ) ∀j 6= i, j > 0 then
N ← N \ {ep,q }, Si ← Si ∪ {p}, S0 ← S0 \ {p}
extend cluster Si ?
else if ∃i, j > 0 (i 6= j) : p ∈ Si ∧ q ∈ Sj then
B Case 4: merge clusters Si , Sj ?
if cur(Si ∪ Sj ) < γ ∧ d(Si , Sj ) < δ · min (s(Si ), s(Sj )) then
N ← N \ {ep,q }, Si ← Si ∪ Sj
P ∗ ← P ∗ \ {Sj }
return : P ∗
4. Experiments
The experiment section is divided into 4 subsections. First, we investigate the influence of the training
set size as well as the influence of different data augmentation strategies. This is followed by an investigation of the performance of the proposed method if it is applied to images with curved or arbitrarily
oriented text lines. The third subsection presents and compares results of different versions of our proposed NPL architectures on the very heterogeneous and challenging cBAD dataset [25, 26]. We perform
statistical tests to show the statistical significance of the stated conclusion – the superiority of the proposed ARU-Net in a two-stage workflow over other architectures and a single-stage workflow. Finally, we
compare the proposed method against other state-of-the-art methods on the datasets of 3 recently hosted
competitions. As mentioned in Sec. 3 we will follow the similarity score of [25] (F-value) to measure the
quality of the baseline detection. The configuration for all experiments including the hyperparameters of
the network architecture as well as the training are summarized in Tab. 4.1. This configuration is the
result of an extensive search in the hyperparameter space and results in impressive results for various
scenarios/datasets.
Since no early stopping based on the loss for any validation set is used, we train on the entire training
set. The ARU-Net workflow for training and inference (Tensorflow code) as well as a trained network
22
Table 4.1: Hyperparameters – The architecture and training configuration which were used in this work are described.
Image pre-processing: input image I is downscaled by a factor of 2 for max{Ih , Iw } < 2000, 3
for 2000 ≤ max{Ih , Iw } < 4800 or 4 followed by a normalization to mean 0 and variance 1 (on pixel
intensity level)
RU-Net architecture, see Fig. 3.2 & 3.3: number of scale spaces: 6, initial feature depth: 8,
residual depth (activated layers in a residual block): 3, feature increasing and spatial decreasing factor:
2, activation function: ReLu, kernel size: 3 × 3, stride: 1
A-Net architecture: 4 layer CNN, activation function: ReLu, kernel size: 4×4, stride: 1, maxpooling
of size 2 × 2 after each convolution, feature number: 12, 16, 32, 1
ARU-Net architecture, see Fig. 3.4: number of image scales: 5, classifier: 4 × 4 convolution layer
with softmax activation
Training: weight initialization: Xavier, optimizer: RMSprop, learning rate: 0.001, learning rate
decay per epoch: 0.985, weight decay on the L2 norm: 0.0005, exponential moving average on the
model weights: 0.9995, mini batch size: 1 (due to memory limitations of the GPU), early stopping:
none (trained for a fixed number of epochs)
are freely available4 . The ARU-Net training takes 3 h to 24 h from scratch (dependent on the number
of epochs and samples per epoch) on a Titan X GPU. The inference time per image ranges from 2 s to
12 s per image on a dual core laptop (Intel Core i7-6600U with 16GiB RAM), this reduces to 0.5 s to 2 s
running the ARU-Net on the Titan X.
4.1. Influence of Training Sample Number and Data Augmentation
A major drawback of state-of-the-art approaches (Sec. 2) is the need for an extensive expert tuning
if confronted with scenarios which are not already covered. But the eligibility for an usage at industrial
scale depends on the possibility to easily adapt at reasonable cost. For approaches relying on machine
learning, this reduces to two questions:
• What about the amount of ground truth needed?
• What about the effort of ground truth production?
Concerning the second question, we refer to Alg. 1. The annotation of baselines for a document image is
quite easy and does not need remarkable expert knowledge compared to, e.g., ground truth production
for ATR systems for historical handwritings or even the text line annotation at surrounding polygon
level. The effort is reduced to several minutes per page by using platforms such as Transkribus5 . In the
following, we want to examine the first question.
The influence of training dataset size along with different data augmentation strategies is investigated
for the freely available Bozen dataset6 [24], see Fig. A.1. This dataset is a subset of documents from the
4 https://github.com/TobiasGruening/ARU-Net
5 https://transkribus.eu
6 https://zenodo.org/record/218236
23
Ratsprotokolle collection of Bozen composed of minutes of the council meetings held from 1470 to 1805
and consists of 400 pages. It is written in Early Modern German. Baseline ground truth information is
available in form of PAGE7 XML. The dataset is quite challenging concerning layout analysis issues. Most
of the pages consist of a single main text region with many difficulties for line detection and extraction,
e.g., bleed through, touching text lines and marginalia. For the following experiments, we have randomly
divided the Bozen set in a set of training samples T of size 350 and a test set of size 50. In a first step,
we randomly set up a chain of subsets of T
T1 ⊂ T3 ⊂ T5 ⊂ T10 ⊂ T30 ⊂ T50 ⊂ T100 ⊂ T200 ⊂ T350 ,
(4.1)
where Ti contains i training samples (pages and pixel ground truth). Since we expect an influence of the
choice of training samples (= sorting of T ), we repeat the mentioned procedure 4 times. Notably, the test
set remains untouched. Finally, we got 45 training sets – five of each quantity. For each set, we trained the
RU-Net for 100 epochs with 256 images per epoch. Therefore, we randomly choose samples of the training
set and remove them from the set. If each element of the training set was used for training once, we start
again with the initial training set. Hence, it does not matter whether the number of training samples per
epoch exceeds the size of the training set or not. This procedure guarantees the same amount of training
samples shown to the networks in training independent of the size of the training set. The RU-Net was
chosen instead of the ARU-Net, because of the homogeneity of the Bozen dataset concerning font size
and resolution. We trained the RU-Net from scratch on all 45 sets in 4 different scenarios. For training
purposes the image pre-processing mentioned in Tab. 4.1 is disabled. Instead, the training samples (I, GI )i
are pre-processed following one of the four strategies:
1. subsampled by a constant factor of 3 (no further data augmentation - one training sample per
element of the training set) – B
2. randomly subsampled by a factor s ∈ [2, 5] – S
3. S + random affine transformation (three corner points of the image are randomly shifted within a
circle of diameter 0.025 · max(Ih , Iw ) around there original position) – S + A
4. S + A + elastic transformation [47] – S + A + E
For the test set the images were sub-sampled by the constant factor of 3 in all scenarios. The results
of these 180 experiments are shown in Fig. 4.1. One can see that all 3 data augmentation strategies
significantly improve the performance compared to the base (B) strategy. Notably, for small numbers of
training samples the min-max difference is much larger than for higher number of training samples. Hence,
7 http://www.primaresearch.org/tools
24
B
S
S+A
S+A+E
F-Value on Test Set
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
1
3
5
10
30
50
Number of Training Samples
100
200
350
Figure 4.1: Influence of the number of training samples and of different data augmentation strategies – The
bar height represents the mean F-value. The error bars encode min-max values of the 5 experiments (not the standard
deviation). The dashed green line marks the maximum mean value of 0.975 achieved for 350 trainings samples. For a
description of the different augemantation strategies: B, S, S+A and S+A+E, see main text.
if just a few training samples are available, the choice of these is of importance. The best mean F-value
(0.975) is achieved for all 350 training samples with the S + A + E strategy. Nevertheless, there only is a
negligible loss in performance for 200 or 100 training samples. Even for 30 training samples, a F-value of
0.963 is achieved for the S+A strategy, which is sufficient for most applications, see Fig. A.1. This results
in a quite acceptable effort for ground truth production making the presented approach interesting even
for industrial production. The S + A data augmentation strategy will be the default for the rest of this
work.
Of course, the presented numbers are not directly transferable to collections with pages of entirely
different scenarios, e.g., census tables mixed with postal cards mixed with ... . One would expect
that more than 30 training samples are necessary for this kind of scenario. Nevertheless, the presented
experiment reflects a common situation: One has a robust baseline detector which was trained on very
heterogeneous data (see Sec. 4.4.3), but this detector does not work satisfyingly well for a certain (in
most cases quite homogeneous) collection. The numbers presented here give a hint concerning the effort
of ground truth production necessary in this scenario.
4.2. Curved and Oriented Text Lines
In this subsection, we demonstrate the ability of the introduced approach to handle curved or arbitrarily
oriented text lines. In a first experiment, the test set of the Bozen dataset was deformed to contain
arbitrarily curved text lines. For this purpose, we utilized trigonometric functions with random period to
simulate curved text lines in the test phase, see Fig. A.2. The RU-Net was trained (5 times) for 100 epochs
with 256 samples per epoch on the Bozen training set using the S + A + E augmentation strategy with
25
strong elastic deformations. We choose elastic transformations in training, because they simulate curves
of different amplitudes and frequencies in the same image. Furthermore, we increased the polynomial
degree (Def. 3.3.10) to 5 to enable the system to handle the curvatures present in the test set.
Remark 4.2.1. Different methods were used to deform the images during training and test phases.
Hence, the system had to learn the concept of curved text lines instead of an inversion of the image
degradation method used in the training phase.
In a second experiment, we have trained an RU-Net (5 times) on arbitrarily oriented samples of the
Bozen training set and evaluated the resulting networks on oriented pages of the test set. The results are
shown in Tab. 4.2 and a few sample images are shown in Fig. A.2. For the curved scenario the results
are as good as for the base scenario. In case of the oriented scenario the results are slightly worse, but
still excellent. This demonstrates the applicability for images with curved or oriented text lines without
remarkable adaptation of the workflow. Finally, we have trained five models with all degradations (affine,
elastic, rotation) and evaluated this model on the three different scenarios. The corresponding F-values
are depicted in Tab. 4.2. The system is worse than the experts for the base and curved scenarios, but for
the oriented scenario it even benefits from the additional elastic transformations.
Table 4.2: Results for the Bozen test set – The results in the Base, Curved and Oriented scenario are depicted. The Pand R-values are strongly related to the well known precision and recall measures, see [25]. Finally, the results for a single
system trained with all degradations are shown.
Scenario
∅ P-Val
∅ R-Val
∅ F-val [min,max]
Base
0.9765
0.9734
0.9750 [0.9693, 0.9770]
Curved
0.9802
0.9690
0.9745 [0.9725, 0.9760]
Oriented
0.9625
0.9655
0.9640 [0.9582, 0.9674]
∅ F-val (Base)
∅ F-val (Curved)
∅ F-val (Oriented)
0.9531
0.9573
0.9676
Combined
4.3. U-Net vs. ARU-Net vs. Single-Stage Workflow
In Sec. 3, we have introduced the ARU-Net in a two-stage workflow. In this section, we will investigate
its superiority over the classical U-Net as well as over a ”single-stage” workflow. For this purpose we have
trained the U-, RU-, ARU- and LARU-Net (each 5 times – random weight initialization and random
training sample order) on the recently introduced cBAD dataset8 [26]. The LARU-Net is an ARU-Net
with a separable MDLSTM9 layer at the lowest resolution to incorporate full spatial context. The details
of the dataset are described in [25]. In our opinion, this is the most challenging freely available dataset at
8 https://zenodo.org/record/257972
9A
separable MDLSTM layer is a concatenation of two (x- and y-direction) BLSTM layers
26
Table 4.3: Results for the cBAD test set – The results for different neural network architectures and the workflow
without Stage II (for the ARU-Net) are shown. Each architecture is trained 5 times on the cBAD train set. The results are
sorted with respect to computational effort. The last two columns indicate whether an architecture is superior to all before
mentioned ones in terms of disjunct confidence intervals and the Tukey-Duckworth test.
Method
∅ F-val [95% CI]
Simple Track
Complex Track
ARU I†
0.9627 [0.9615, 0.9636]
0.9081 [0.9071, 0.9095]
CI
T-D
U
0.9714 [0.9701, 0.9721]
0.9114 [0.9107, 0.9122]
3
3
RU
0.9756 [0.9744, 0.9766]
0.9182 [0.9165, 0.9203]
3
3
ARU
0.9781 [0.9772, 0.9789]
0.9223 [0.9214, 0.9230]
3
3
LARU
0.9772 [0.9765, 0.9780]
0.9233 [0.9217, 0.9249]
7
7
†
single-stage workflow – baseline estimation by basic image processing
methods (binarization of B followed by a CC analysis, no usage of S)
the moment. We have trained each network for 250 epochs, 1024 training samples each epoch using the
S + A data augmentation strategy. To assure the statistical significance of the posed superiority of the
newly introduced architecture, we follow [48] and provide the results of a statistical analysis. The choice of
appropriate statistical tests is quite limited since we can’t make any assumptions regarding the underlying
distribution. We utilize 95% confidence intervals (CI) provided by non-parametric bootstrapping [49] as
well as the Tukey-Duckworth test (level of significance: 5%) [50]. The results obtained are summarized
in Tab. 4.3. The ARU-Net performs significantly (last two columns) better than all architectures with
less computational effort. The LARU-Net could not prove its superiority and is therefore dismissed.
Furthermore, the results show that the introduction of the second stage is beneficial for the overall
performance. Hence, the ARU-Net together with the two-stage workflow has shown its superiority (which
is statistically significant) over the other systems and is used in the following. It has to be mentioned
that the above comparison is not fair concerning the number of trainable parameters – U - 2.16, RU 4.13, ARU - 4.14, LARU - 6.25 (in millions) – nor concerning the training or even inference time. The
comparison is about different architectures which, theoretically, have different capabilities, and whether
they make good use of them or not. For instance, the LARU-Net should be capable of incorporating a
more detailed spatial context, but in fact it does not benefit (in our settings) from this capability.
4.4. Comparison against the State of the Art
In this subsection, we compare the proposed framework against the state of the art. We have chosen
the 3 most recent competitions on text line detection for historical documents, namely: ICDAR 2015
competition on text line detection in historical documents [14], ICDAR2017 Competition on Layout
Analysis for Challenging Medieval Manuscripts (Task 2) [51] and cBAD: ICDAR2017 Competition on
Baseline Detection [37]. We will not further introduce the datasets or metrics used and refer to the
competition papers.
27
4.4.1. ICDAR 2015 Competition on Text Line Detection in Historical Documents (ANDAR-TL)
The ARU-Net was trained on the cBAD training set10 . This competition aims at the origin point
(OP) detection. An OP is roughly spoken the lower left ”corner” of a text line. Hence, we calculate the
left most point of each detected baseline. This is the output of our system for this competition. The
achieved results are shown in Tab. 4.4. Since the ARU-Net was not trained on the original training data,
it is hard to compare its results to the other ones. Nevertheless, we would like to stress the fact, that
trained systems usually perform better if training set and test set are sampled from the same distribution.
E.g., the ARU-Net trained on the cBAD training set achieves an average F-value of 0.9605 for the Bozen
test set, which is worse than the F-vlaue of 0.9750 of the system trained solely on the Bozen training set,
see Tab. 4.2. This indicates (but does not prove) the superiority of the presented method over the other
methods in Tab. 4.4.
Table 4.4: Origing Point (OP) detection results for the ANDAR-TL test set – Results for the dataset of [14] are
shown. #DF means the number of detection failures (no OP detected by the system), #DM means the number of detection
misses (detected OP far away from the ground truth OP) and #FP means the number of false positives
Method
#HYP
#COR
#DF
#DM
#FP
avg. cost
UNIFR
9301
2578
3022
6456
267
19.00
IA-2
11789
5655
407
6032
102
14.51
A2iA-3†
8967
6523
2490
2263
181
13.20
SNU[32]
10466
7741
948
2700
25
9.77
[33]
10896
8015
517
2860
21
8.19
proposed
11635
9610
358
1942
83
5.39
†
According to [28] this is an extension of [34].
4.4.2. ICDAR2017 Competition on Layout Analysis for Challenging Medieval Manuscripts (Task 2)
The ARU-Net was trained for 250 epochs 1024 samples per epoch on the competition training data11
provided by the competition organizers. This allows an entirely fair comparison to the participant’s results,
see Tab. 4.5. The proposed method substantially outperforms the winning one and reduces the error (the
gap to 1.0) by 43.26% (relatively). The specialty of this competition was, that the methods should focus
on a special kind of text, e.g., comments were not annotated as text. Hence, the ARU-Net had to learn
to distinguish between different types of text. The output of the ARU-Net and the detected baselines for
a sample image of the CSG18 subset of the test set are shown in Fig. 4.2. One can see, that the ARU-Net
entirely ignores all text entities not regarded (in this competition) as main text. Remarkably, no further
information besides the input image is provided to the ARU-Net.
10 The
competition training data was not available to the authors.
11 http://diuf.unifr.ch/main/hisdoc/diva-hisdb
28
Table 4.5: Results for the ICDAR2017 Competition on Layout Analysis for Challenging Medieval Manuscripts
– The F-values for Task 2 of all participants and the proposed method are shown for the different subsets of the test set.
Method
CB55
CSG18
CSG863
overall
CVML
0.9534
0.8734
0.9751
0.9340
BYU
0.9597
0.9879
0.9830
0.9768
CITlab
0.9896
0.9853
0.9716
0.9822
proposed
0.9980
0.9828
0.9889
0.9899
Figure 4.2: Results for an image of the CSG18 subset of the test set – The original image (only the main text lines
were ground truthed), the baseline image generated by the trained ARU-Net and the baselines detected by the proposed
method are shown (from left to right).
4.4.3. cBAD: ICDAR2017 Competition on Baseline Detection
We compare our average result for the ARU-Net (see Tab. 4.3) to the results presented in [37], see
Tab. 4.6. Our method performs considerably better in both tracks compared to all submissions. Especially,
the increase in performance for the complex track is massive. Remarkably, the winning team uses an UNet based system with task specific pre- and postprocessing. This indicates that the newly introduced
concepts and parametrization, which are presented in this work, significantly improve the capability of the
classical U-Net. Some results on chosen images of the cBAD test set are shown in Fig. A.3-A.5. Notably,
no further information besides the input image (and the text region information in the simple track) is
provided to the ARU-Net nor to the second stage of the workflow during inference.
5. Conclusion
In this work we presented a machine learning based method for text line detection in historical documents. The text lines are represented by their baselines. The problem and the proposed method were
29
Table 4.6: Results for the cBAD test set – The P-, R- and F-values of all participants and of the proposed method for
the simple and complex track of the cBAD: ICDAR2017 Competition on Baseline Detection are shown.
Method
Simple Track
Complex Track
P-Val
R-Val
F-val
P-Val
R-Val
F-val
LITIS
0.780
0.836
0.807
–
–
–
IRISA†
0.883
0.877
0.880
0.692
0.772
0.730
UPVLC
0.937
0.855
0.894
0.833
0.606
0.702
BYU
0.878
0.907
0.892
0.773
0.820
0.796
DMRZ
0.973
0.970
0.971
0.854
0.863
0.859
proposed
0.977
0.980
0.978
0.926
0.918
0.922
†
This method is based on the work presented in [16].
introduced thoroughly. The proposed ARU-Net, which is a universal pixel labeling approach, was trained
to predict the baseline position and the beginning and end of each text line. This enables the system to
handle documents with complex layouts, e.g., tables, marginalia, multi columns layouts. We have shown
that the system can be trained from scratch with manageably few training samples for a complex but
homogeneous collection. Remarkably, ground truth production is quite cheap. A ground truth sample
is just a page with annotated baselines, which can be done in a few minutes per page. Therefore, one
can expect that an adaptation on collections, which are not covered by the neural network, is possible
with quite reasonable ground truthing effort. The applicability of the proposed method was shown for
straight, curved and oriented text lines as well as for a combined scenario. The superiority of the proposed
ARU-Net in the two-stage workflow over the classical U-Net and over a simplified workflow was shown and
statistically verified. Finally, we showed that the proposed method substantially outperforms the previous
state of the art. Nevertheless, as one can see in Fig. A.3-A.5 there are still errors made by the system,
e.g., missed baselines (see Fig. A.4 – bottom right), segmentation errors (see Fig. A.5 – bottom left),
false positives (see Fig. A.3 – top left) or problems with strongly degraded documents (see Fig. A.4 – top
left). But these errors do not seem to follow a certain deterministic principle, which is not surprising for
a method based on machine learning. However, we plan to test newly introduced concepts like capsules,
memory augmentation and deeply supervised networks to further improve the system’s performance.
Acknowledgement
NVIDIA Corporation kindly donated a Titan X GPU used for this research. This work was partially
funded by the European Unions Horizon 2020 research and innovation programme under grant agreement
No 674943 (READ Recognition and Enrichment of Archival Documents). Finally, we would like to thank
Udo Siewert for his valuable comments and suggestions.
30
References
References
[1] A. Isaac, R. Clayphan, B. Haslhofer, Europeana: Moving to linked open data, Information Standards
Quarterly 24 (2/3).
[2] T. Causer, V. Wallace, Building a volunteer community: Results and findings from transcribe bentham, Digital Humanities Quarterley 6 (2) (2012) 1–28.
[3] J. A. Sánchez, G. Mühlberger, B. Gatos, P. Schofield, K. Depuydt, R. M. Davis, E. Vidal, J. de Does,
tranScriptorium: a european project on handwritten text recognition, in: Proceedings of the 2013
ACM symposium on Document engineering, ACM, 2013, pp. 227–228.
[4] A. Graves, J. Schmidhuber, Offline handwriting recognition with multidimensional recurrent neural
networks, Advances in Neural Information Processing Systems 21, NIPS’21 (2008) 545–552.
[5] G. Leifert, T. Strauß, T. Grüning, W. Wustlich, R. Labahn, Cells in multidimensional recurrent
neural networks, J. Mach. Learn. Res. 17 (1) (2016) 3313–3349.
[6] J. Puigcerver, A. H. Toselli, E. Vidal, Word-graph and character-lattice combination for kws in handwritten documents, in: 2014 14th International Conference on Frontiers in Handwriting Recognition,
IEEE, 2014, pp. 181–186.
[7] T. Strauß, T. Grüning, G. Leifert, R. Labahn, CITlab ARGUS for Keyword Search in Historical
Handwritten Documents: Description of CITlab’s System for the ImageCLEF 2016 Handwritten
Scanned Document Retrieval Task, CEUR Workshop Proceedings, Évora, Portugal, 2016.
[8] T. Strauß, G. Leifert, T. Grüning, R. Labahn, Regular expressions for decoding of neural network
outputs, Neural Networks 79 (2016) 1–11.
[9] J. A. Sanchez, V. Romero, A. H. Toselli, E. Vidal, ICFHR2014 Competition on Handwritten Text
Recognition on Transcriptorium Datasets (HTRtS), in: Proceedings of International Conference on
Frontiers in Handwriting Recognition, ICFHR, Vol. 2014-Decem, IEEE, 2014, pp. 785–790.
[10] I. Pratikakis, K. Zagoris, J. Puigcerver, A. H. Toselli, E. Vidal, ICFHR2016 Handwritten Keyword
Spotting Competition ( H-KWS 2016 ), in: Proceedings of International Conference on Frontiers in
Handwriting Recognition, ICFHR, IEEE, 2016, pp. 613–618.
[11] M. Rusiñol, D. Aldavert, R. Toledo, J. Lladós, Efficient segmentation-free keyword spotting in historical document collections, Pattern Recognition 48 (2) (2015) 545–555.
[12] T. Bluche, Joint line segmentation and transcription for end-to-end handwritten paragraph recognition, Advances in Neural Information Processing Systems (2016) 838–846.
[13] T. Konidaris, A. L. Kesidis, B. Gatos, A segmentation-free word spotting method for historical
printed documents, Pattern Analysis and Applications 19 (4) (2016) 963–976.
[14] M. Murdock, S. Reid, B. Hamilton, J. Reese, ICDAR 2015 competition on text line detection in
historical documents, in: Proceedings of the International Conference on Document Analysis and
Recognition, ICDAR, Vol. 2015-Novem, IEEE, 2015, pp. 1171–1175.
[15] S. Sudholt, G. A. Fink, Phocnet : A deep convolutional neural network for word spotting in handwritten documents, in: Proceedings of International Conference on Frontiers in Handwriting Recognition,
ICFHR, 2016, pp. 1–6.
[16] G. Renton, C. Chatelain, S. Adam, C. Kermorvant, T. Paquet, Handwritten text line segmentation
using Fully Convolutional Network, in: Proceedings of the International Conference on Document
Analysis and Recognition, ICDAR, 2017, pp. 5–9.
31
[17] N. Arvanitopoulos, S. Süsstrunk, Seam Carving for Text Line Extraction on Color and Grayscale
Historical Manuscripts, International Conference on Frontiers in Handwriting Recognition (ICFHR)
(2014) 726 – 731.
[18] Q. N. Vo, S. H. Kim, H. J. Yang, G. Lee, Binarization of degraded document images based on
hierarchical deep supervised network, Pattern Recogn. 74 (2018) 568–586.
[19] C. Tensmeyer, B. Davis, C. Wigington, I. Lee, B. Barrett, PageNet: Page Boundary Extraction in
Historical Handwritten Documents, in: Proceedings of the 4th International Workshop on Historical
Document Imaging and Processing, HIP ’1, ACM, New York, NY, USA, 2017, pp. 59–64.
[20] K. Chen, M. Seuret, J. Hennebert, R. Ingold, Convolutional Neural Networks for Page Segmentation of Historical Document Images, in: Proceedings of the International Conference on Document
Analysis and Recognition, ICDAR, 2017, pp. 965–970.
[21] O. Ronneberger, P. Fischer, T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, Miccai (2015) 234–241.
[22] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: 2016 IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778.
[23] J. Ryu, H. I. Koo, N. I. Cho, Language-independent text-line extraction algorithm for handwritten
documents, IEEE Signal Processing Letters 21 (9) (2014) 1115–1119.
[24] J. A. Sánchez, V. Romero, A. H. Toselli, E. Vidal, READ dataset Bozen (2016). doi:10.5281/
zenodo.218236.
[25] T. Grüning, R. Labahn, M. Diem, F. Kleber, S. Fiel, READ-BAD: A New Dataset and Evaluation
Scheme for Baseline Detection in Archival Documents, arXiv preprint arXiv:1705.03311.
[26] M. Diem, F. Kleber, S. Fiel, T. Grüning, B. Gatos, ScriptNet: ICDAR 2017 Competition on Baseline
Detection in Archival Documents (cBAD) (2017). doi:10.5281/zenodo.257972.
[27] A. Zahour, L. Likforman-Sulem, W. Boussalaa, B. Taconet, Text Line segmentation of historical Arabic documents, Proceedings of the International Conference on Document Analysis and Recognition,
ICDAR 1 (2-4) (2007) 138–142.
[28] S. Eskenazi, P. Gomez-Krämer, J.-M. Ogier, A comprehensive survey of mostly textual document
segmentation algorithms since 2008, Pattern Recognition 64 (2017) 1–14.
[29] A. Nicolaou, B. Gatos, Handwritten text line segmentation by shredding text into its lines, in:
Proceedings of the International Conference on Document Analysis and Recognition, ICDAR, 2009,
pp. 626–630.
[30] R. Saabni, A. Asi, J. El-Sana, Text line extraction for historical document images, Pattern Recognition Letters 35 (1) (2014) 23–33.
[31] A. Garz, A. Fischer, R. Sablatnig, H. Bunke, Binarization-free text line segmentation for historical
documents based on interest point clustering, in: Proceedings - 10th IAPR International Workshop
on Document Analysis Systems, DAS 2012, IEEE, 2012, pp. 95–99.
[32] B. Ahn, J. Ryu, H. I. Koo, N. I. Cho, Textline detection in degraded historical document images,
EURASIP Journal on Image and Video Processing 2017 (1) (2017) 82.
[33] T. Grüning, G. Leifert, T. Strauß, R. Labahn, A Robust and Binarization-Free Approach for Text Line
Detection in Historical Documents, in: Proceedings of the International Conference on Document
Analysis and Recognition, ICDAR, 2017, pp. 236–241.
32
[34] B. Moysset, C. Kermorvant, C. Wolf, J. Louradour, Paragraph text segmentation into lines with
Recurrent Neural Networks, Proceedings of the International Conference on Document Analysis and
Recognition, ICDAR 2015-Novem (2015) 456–460.
[35] B. Moysset, J. Louradour, C. Kermorvant, C. Wolf, Learning text-line localization with shared and
local regression neural networks, in: Proceedings of International Conference on Frontiers in Handwriting Recognition, ICFHR, 2017, pp. 1–6.
[36] B. Moysset, C. Kermorvant, C. Wolf, Full-Page Text Recognition: Learning Where to Start and When
to Stop, in: Proceedings of the International Conference on Document Analysis and Recognition,
ICDAR, 2017, pp. 871–876.
[37] M. Diem, F. Kleber, S. Fiel, B. Gatos, T. Grüning, cBAD: ICDAR2017 Competition on Baseline
Detection, in: Proceedings of the International Conference on Document Analysis and Recognition,
ICDAR, 2017, pp. 1355–1360.
[38] J. Long, E. Shelhamer, T. Darrell, Fully convolutional networks for semantic segmentation, in: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
Vol. 07-12-June, 2015, pp. 3431–3440.
[39] H. Noh, S. Hong, B. Han, Learning deconvolution network for semantic segmentation, in: Proceedings
of the IEEE International Conference on Computer Vision, Vol. 2015 Inter, 2015, pp. 1520–1528.
[40] D. E. Rumelhart, G. E. Hinton, R. J. Williams, Learning representations by back-propagating errors,
Nature 323 (6088) (1986) 533–536.
[41] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT Press, 2016.
[42] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to document recognition,
Proceedings of the IEEE 86 (11) (1998) 2278–2323.
[43] X. Glorot, Y. Bengio, Understanding the difficulty of training deep feedforward neural networks,
in: Y. W. Teh, M. Titterington (Eds.), Proceedings of the Thirteenth International Conference on
Artificial Intelligence and Statistics, Vol. 9 of Proceedings of Machine Learning Research, PMLR,
2010, pp. 249–256.
[44] J. Serra, Image Analysis and Mathematical Morphology, Vol. 1, 1982.
[45] B. Delaunay, Sur la sphere vide, Bulletin de l’Académie des Sciences de l’URSS 6 (1934) 793–800.
[46] Y. Boykov, O. Veksler, R. Zabih, Fast approximate energy minimization via graph cuts, IEEE Transactions on Pattern Analysis and Machine Intelligence 23 (11) (2001) 1222–1239.
[47] P. Simard, D. Steinkraus, J. C. Platt, Best Practices for Convolutional Neural Networks Applied to
Visual Document Analysis, Proceedings of the 7th International Conference on Document Analysis
and Recognition (2003) 958–963.
[48] J. Puigcerver, Are Multidimensional Recurrent Layers Really Necessary for Handwritten Text Recognition?, in: Proceedings of the International Conference on Document Analysis and Recognition,
ICDAR, 2017, pp. 67–72.
[49] B. Efron, Better bootstrap confidence intervals, Journal of the American Statistical Association
82 (397) (1987) 171–185.
[50] J. W. Tukey, A Quick Compact Two Sample Test To Duckworth’s Specifications, Technometrics 1 (1)
(1959) 31–48.
[51] F. Simistira, M. Bouillon, M. Seuret, M. Würsch, M. Alberti, R. Ingold, M. Liwicki, ICDAR2017
Competition on Layout Analysis for Challenging Medieval Manuscripts, in: Proceedings of the International Conference on Document Analysis and Recognition, ICDAR, 2017, pp. 1361–1370.
33
Figure A.1: Results for an image of the Bozen test set – Results for RU-Nets trained on 5, 30 and 350 training
samples (left to right) with different data augmentation strategies B, S+A and S+A+E (top to bottom) are shown.
34
Figure A.2: Results for an image of the Bozen test set – Results for two “degraded” images are shown. The images
were arbitrarily curved and rotated.
35
Figure A.3: Results for images of the cBAD test set – Only the images without any layout information were used.
36
Figure A.4: Results for images of the cBAD test set – Only the images without any layout information were used.
37
Figure A.5: Results for images of the cBAD test set – Only the images without any layout information were used.
38
| 1cs.CV
|
Stochastic Recursive Inclusions in two timescales with non-addtive
iterate dependent Markov noise
Vinayaka G. Yaji and Shalabh Bhatnagar,
Department of Computer Science and Automation,
Indian Institute of Science, Bangalore.
[email protected], [email protected]
arXiv:1611.05961v1 [cs.SY] 18 Nov 2016
February 20, 2018
Abstract
In this paper we study the asymptotic behavior of a stochastic approximation scheme on two
timescales with set-valued drift functions and in the presence of non-additive iterate-dependent Markov
noise. It is shown that the recursion on each timescale tracks the flow of a differential inclusion obtained
by averaging the set-valued drift function in the recursion with respect to a set of measures which take
into account both the averaging with respect to the stationary distributions of the Markov noise terms
and the interdependence between the two recursions on different timescales. The framework studied
in this paper builds on the works of A. Ramaswamy et al. by allowing for the presence of nonadditive iterate-dependent Markov noise. As an application, we consider the problem of computing
the optimum in a constrained convex optimization problem where the objective function and the
constraints are averaged with respect to the stationary distribution of an underlying Markov chain.
Further the proposed scheme neither requires the differentiability of the objective function nor the
knowledge of the averaging measure.
1
Introduction
Consider the standard two timescale stochastic approximation scheme given by,
(2)
Yn+1 − Yn − b(n)Mn+1 = b(n)h2 (Xn , Yn ),
Xn+1 − Xn −
(1)
a(n)Mn+1
= a(n)h1 (Xn , Yn ),
(1a)
(1b)
where n ≥ 0 denotes the iteration index, {Xn }n≥0 is a sequence of Rd1 -valued random variables, {Yn }n≥0 is
a sequence of Rd2 -valued random variables, for any i ∈ {1, 2}, hi : Rd1 +d2 → Rdi is a Lipschitz continuous
(i)
function and {Mn }n≥1 is sequence of Rdi -valued square integrable martingale difference sequence. The
step size sequences, {a(n)}n≥0 and {b(n)}n≥0 are sequences of positive real numbers chosen such that they
b(n)
b(n)
satisfy limn→∞ a(n)
= 0 in addition to the Monte Carlo step size conditions. The condition limn→∞ a(n)
=
0, ensures that after large number of iterations the time step of recursion (1a) is much smaller than that
of (1b). Thus the recursion (1a) appears to be static with respect to the recursion (1b). In [1], using the
dynamical systems approach studied in [2], the above intuition was shown to hold. More precisely, the
faster timescale recursion (1b), was shown to track the ordinary differential equation (o.d.e.) given by,
dx
= h1 (x, y0 ),
dt
(2)
for some y0 ∈ Rd2 and assuming that for every y ∈ Rd2 , o.d.e. (2) admits a unique globally asymptotically
stable equilibrium point, say λ(y), the slower timescale recursion (1a) was shown to track the o.d.e. given
by,
dy
= h2 (λ(y), y).
(3)
dt
Further, the map y → λ(y) was assumed to be Lipschitz continuous.
An important application of the above stochastic approximation scheme is in the computation of a
saddle point of a function. Given a function f : Rd1 × Rd2 → R, (x∗ , y ∗ ) ∈ Rd1 +d2 (x∗ ∈ Rd1 and y ∗ ∈ Rd2
1
respectively) is a saddle point of the function f (·) if,
inf
inf f (x, y) = f (x∗ , y ∗ ).
sup f (x, y) = sup
(4)
d
y∈Rd2 x∈R 1
x∈Rd1 y∈Rd2
From [3, Prop. 5.5.6], we know that the function f (·) admits a saddle point if for every (x, y) ∈ Rd ,
(1) −f (x, ·) and f (·, y) are convex functions,
(2) the sub level sets of functions x → supy∈Rd2 f (x, y) and y → − inf x∈Rd1 f (x, y) are compact sets.
Over the years significant effort has been devoted for developing algorithms to compute such points (see
[4, 5] and references therein). Most of the solutions proposed in literature require the computation of
partial derivatives of the function f (·). However in practice the closed form expressions of the partial
derivatives are often not known or are expensive to compute and in such cases one often estimates the
partial derivatives using values of the objective function (see [6] for one such estimation method). The
two timescale stochastic approximation scheme can be used to compute a saddle point with noisy partial
derivative values by setting h1 (·) := −∇x f (·) and h2 (·) := ∇y f (·) where ∇x and ∇y denote the partial
(1)
derivative operators with respect to x and y respectively. In this setting, the sequences {Mn }n≥1 and
(2)
{Mn }n≥1 denote the partial derivative estimation errors and the map λ(·) denotes correspondence between y ∈ Rd2 and the minimum of the function f (·, y). The vector field associated with o.d.e. (3) is
now given by ∇y f (x, y)|x=λ(y) which can be shown to be the same as ∇y f (λ(y), y) under some conditions
known as envelope theorem in mathematical economics (see [7]). Thus the slower timescale maximizes the
function y → inf x∈Rd1 f (x, y) = f (λ(y), y), there by in the limit the iterates of recursion (1) converge to a
saddle point of the function f (·).
In some cases the function whose saddle point needs to be computed is itself averaged with respect
to a certain probability measure. For example consider the function f : Rd1 +d2 × S → R where S is a
compact metric space and for some probability measure µ on S, one wishes Rto compute the saddle point
of the function fµ : Rd1 +d2 → R where for every (x, y) ∈ Rd1 +d2 , fµ (x, y) := S f (x, y, s)µ(ds). If one has
access to i.i.d. samples with probability measure µ, then the saddle point problem above can be solved
using recursion (1). But if access to such samples are not available and one uses Markov chain Monte
Carlo methods to sample from the measure µ, then the recursion (1) has a non-additive iterate-dependent
Markov noise component. The recursion now takes the form:
(2)
Yn+1 − Yn − b(n)Mn+1 = b(n)h2 (Xn , Yn , Sn(2) ),
Xn+1 − Xn −
(1)
(1)
a(n)Mn+1
=
a(n)h1 (Xn , Yn , Sn(1) ),
(5a)
(5b)
(2)
where {Sn }n≥1 and {Sn }n≥1 denote the Markov noise terms taking values in an appropriate state space.
The recursion (5) was studied in [8], under assumptions similar to those in [1] which include the Lipschitz
continuity of the maps h1 (·), h2 (·) and λ(·).
Often, the maps h1 (·) and h2 (·) in recursion (1) are not Lipschitz continuous and the map λ(·) is
not even single valued (that is the o.d.e. (2) has a globally asymptotically stable equilibrium set). This
motivates one to study the two timescale recursion with set-valued drift functions. The recursion now
takes the form:
(2)
Yn+1 − Yn − b(n)Mn+1 ∈ b(n)H2 (Xn , Yn ),
Xn+1 − Xn −
(1)
a(n)Mn+1
∈ a(n)H1 (Xn , Yn ),
(6a)
(6b)
where H1 (·) and H2 (·) are set-valued maps and other quantities have similar interpretation to those in
recursion (1). The above recursion was studied in [9] and further the map λ(·) was allowed to be set-valued
and upper semicontinuous.
1.1
Contributions of this paper and comparisons with the state of the art
In this paper we study the asymptotic behavior of the recursion given by,
(2)
Yn+1 − Yn − b(n)Mn+1 ∈ b(n)H2 (Xn , Yn , Sn(2) ),
Xn+1 − Xn −
(1)
a(n)Mn+1
2
∈
a(n)H1 (Xn , Yn , Sn(1) ),
(7a)
(7b)
(2)
(1)
where H1 (·) and H2 (·) are set-valued maps and {Sn }n≥0 and {Sn }n≥0 are the Markov noise terms taking
values in compact metric spaces S (1) and S (2) respectively. We show that the fast timescale recursion (7b),
tracks the flow of the differential inclusion (DI) given by,
Z
dx
∈ ∪µ∈D(1) (x,y0 )
H1 (x, y0 , s(1) )µ(ds(1) ),
(8)
dt
S (1)
for some y0 ∈ Rd2 , where D(1) (x, y) denotes the set of stationary distributions of the Markov noise terms
(1)
{Sn }n≥0 for every (x, y) ∈ Rd and the integral above denotes the integral of a set-valued map with respect
to measure µ. Further we assume that for every y ∈ Rd2 , the above DI admits a unique globally attracting
set λ(y). The map y → λ(y) is also assumed to be upper semicontinuous. The slower timescale recursion
(7a), is show to track the flow of the DI given by,
Z
dy
H2 (x, y, s(2) )µ(dx, ds(2) ),
(9)
∈ ∪µ∈D(y)
d
dt
(2)
1
R ×S
where y → D(y) denotes a set-valued map taking values in the space of probability measures on S 2 and
the map D(·) is defined such that it captures both the equilibration of the fast timescale iterates to λ(·)
(2)
and the averaging due to the Markov noise terms {Sn }n≥0 .
In comparison with the two timescale framework studied in [8], our work allows for the drift functions
(that is H1 (·) and H2 (·)) to be set-valued and further the map λ(·) is allowed to be set-valued and upper
semicontinuous which is much weaker than the requirement of single valued and Lipschitz continuity
imposed in [8]. The generalization to the set-valued case allows one to analyze recursion (5) when the
drift functions h1 (·) and h2 (·) are single valued and are just measurable, since graph of such a map can
be embedded in the graph of a upper semicontinuous set-valued map as in [10, ch. 5.3(iv)]. We refer the
reader to [10, ch. 5.3] for several other scenarios where the study of stochastic approximation scheme with
set-valued maps becomes essential.
Our work further generalizes the two timescale framework studied in [9] by allowing for the presence
of Markov noise terms. The analysis in this paper does not extend in an straight forward manner from
those in [9] and requires results from set-valued map approximation, parametrization, integration and the
use of probability measure valued functions. However the method of analysis adopted in this paper can be
adapted appropriately to obtain the same convergence guarantees as in [9] when the Markov noise terms
are absent.
1.2
Overview of the analysis and organization of the paper
It is known that continuous, convex and compact set-valued maps taking values in a finite dimensional
space admit a continuous single-valued parametrization. The properties of the set-valued drift function
ensure that the drift functions H1 (·) and H2 (·) are convex and compact set-valued maps and is upper
semicontinuous. However such maps do not admit a continuous parametrization. We can work around
this problem by enlarging the graph of the drift function since the graph of drift function can be embedded
in the graph of a continuous, convex and compact set-valued map which admit a continuous single-valued
parametrization. Thus a sequence of continuous single-valued maps can be obtained which approximate
the set-valued drift function from above. This enables us to write the inclusion (7) in the form of recursion
(5) with an additional parameter. The results needed to accomplish the above are stated in section 2.1.
Before proceeding further one needs to identify the mean fields that the recursion (7) is expected to
track. To this end we need some results from the theory of integration of set-valued maps which are
reviewed in section 2.2. Further the measurablility and integrability properties of the drift functions of
the recursion are investigated and the characterization of the integral of a continuous set-valued map in
terms of its parametrization is established.
In section 2.3 we compile some definitions and results from the theory of differential inclusions which
are needed later to characterize the asymptotic behavior of recursion (7). Further in section 2.4 we state
the assumptions and the main result of the analysis of single timescale stochastic recursive inclusions with
non-additive iterate dependent Markov noise from [11] and in section 2.5 we define and compile some
results needed from the space of probability measure valued functions.
In section 3 we state and motivate the assumptions under which the recursion (7) is analyzed. Using
the results from integration of set-valued maps reviewed in section 2.2 the mean fields are defined and
the main convergence result is stated. The mean fields defined in section 3 possess some properties which
ensure existence of solutions (of their associated differential inclusions). These properties are established
in section 4. In section 4 it is also shown that appropriate modifications of the continuous set-valued maps
3
which approximate the drift functions (obtained in section 2.1) also approximate the mean fields which
play an important role in the analysis later.
The analysis of recursion (7) consists of two parts. In section 5.1, the recursion (7) is analyzed along
the faster timescale. The recursion (7) when viewed along the faster timescale appears to be a single
timescale stochastic recursive inclusion with non-additive iterate dependent Markov noise. In section
5.1, we show that recursion (7) viewed along the faster timescale satisfies all the assumptions associated
with the single timescale recursion presented in section 2.4. Applying the main result of single timescale
analysis we conclude that the faster timescale iterates converge to λ(·) for some y ∈ Rd2 . In section 5.2, the
slower timescale recursion is analyzed. It is shown that the linearly interpolated sample path of the slower
timescale iterates (defined in section 5.2.1) tracks an appropriate DI. Continuous functions tracking the
flow of a dynamical system are known as asymptotic pseudotrajectories (see [12] for definition and related
results). The asymptotic pseudotrajectory argument in this paper presented in section 5.2.2 comprises of
the following steps:
(2)
(1) First step is to get rid of the additive noise terms, {Mn }n≥1 . This involves defining an o.d.e. with
an appropriate piecewise constant vector field and showing that the limit points of the shifted linearly
interpolated trajectory of the slower timescale iterates coincide with the limit points of the solutions
of this o.d.e. in the space of continuous functions on [0, ∞) taking values in Rd2 . Further a simple
argument gives us that the set of limit points of the shifted linearly interpolated trajectories of the
slower timescale iterates is non-empty.
(2) The second step is to show that the limit point obtained in the first step is in fact a solution of DI
(9). This is accomplished using probability measure valued functions reviewed in section 2.5. This
method has also been used in analyzing stochastic approximation schemes such as recursion (5) in
[8] and in [13]. But the analysis in these references made explicit use of the Lipschitz property of the
underlying drift functions. We observe that continuity is sufficient to carry out this analysis. This is
also where our analysis significantly differs from that in [9]. The equilibration of the faster timescale
is also accomplished using probability measures which simplifies the proof compared to that in [9].
In section 5.2.3, the limit sets of the slower timescale iterates are characterized in terms of the dynamics
of DI (9). In addition to the above, using the convergence of the faster timescale iterates to λ(·) obtained
in section 5.1, we obtain the main convergence result of this paper.
In section 6, as an application, we propose an algorithm to compute a solution of a constrained convex
optimization problem. The objective function and constraints are assumed to be convex and affine respectively. Further the optimization problem is obtained by averaging the quantities involved with respect to
the stationary distribution of an underlying Markov chain. Such problems arise in optimal control where
the controller must find an optimum parameter where the changes in state of the underlying system can
be modeled by a Markov chain. The cost function and system constraints are dependent on the state of
the system and the controller seeks to find the optimum of the long run average of cost function while
satisfying the long run average constraints. In such applications the stationary distribution of the system
states are not known, but one has access to a sample path of system state changes. We propose a two
timescale scheme which performs primal ascent along the faster timescale and dual descent along the slower
timescale with the knowledge of the current state at a given iteration. Using the theory presented in this
paper, it is shown that the limit set of the iterates of the proposed two timescale scheme are contained in
the set of Lagrangian saddle points of the underlying averaged constrained convex optimization problem.
Further the algorithm does not assume the differentiability of the objective function and requires only a
noisy estimate of the subgradient.
In section 7, we conclude by providing a few directions for future research and outline certain extensions
where we believe the analysis remains the same.
2
Background
In this section we shall briefly review some results needed from the theory of set-valued maps and differential
inclusions, present a brief outline of the analysis of the single timescale version of stochastic recursive
inclusions with non-additive iterate-dependent Markov noise and define the space of probability measure
valued functions with a metrizable topology which are needed later in the analysis of the two timescale
recursion.
Throughout this paper S denotes a compact metric space and the metric on S is denoted by dS .
Further let 1 ≤ d1 ∈ Z, 1 ≤ d2 ∈ Z, d := d1 + d2 and (x, y) denotes a generic element in Rd where x ∈ Rd1
and y ∈ Rd2 .
4
2.1
Upper semicontinuous set-valued maps and their approximation
First we shall recall the notions of upper semicontinuity, lower semicontinuity and continuity of set-valued
maps. These notions are taken from [14, ch. 1.1].
Definition A set valued map F : Rd × S → subsets of Rk is,
• Upper semicontinuous (u.s.c.) if, for every (x0 , y0 , s0 ) ∈ Rd × S, for every ǫ > 0, there exists δ > 0
(depending on (x0 , y0 , s0 ) and ǫ) such that,
k (x, y) − (x0 , y0 ) k< δ, dS (s, s0 ) < δ =⇒ F (x, y, s) ⊆ F (x0 , y0 , s0 ) + ǫU,
where U denotes the closed unit ball in Rk .
• Lower semicontinuous (l.s.c) if, for every (x0 , y0 , s0 ) ∈ Rd × S, for every z0 ∈ F (x0 , y0 , s0 ), for every
sequence {(xn , yn , sn )}n≥1 converging to (x0 , y0 , s0 ), there exists a sequence {zn ∈ F (xn , yn , sn )}
converging to z0 .
• Continuous if, it is both u.s.c. and l.s.c.
For set valued maps taking compact set values we have the above mentioned notion of u.s.c. to be
equivalent to the standard notion of u.s.c. (see [14, pg. 45]). In this paper we shall encounter set valued
maps which are compact set valued and hence we have chosen to state the above as the definition of upper
semicontinuity.
Set-valued maps studied later satisfy certain properties under which we will be able to approximate
them with a family of continuous single-valued maps with an additional parameter. These properties are
natural extensions of the properties imposed on maps studied in [12, 9] to the case of stochastic recursive
inclusions with Markov noise and we choose to call such maps stochastic approximation maps (SAM). The
definition of SAM is stated below.
Definition [SAM] A set-valued map F : Rd × S → subsets of Rk is a stochastic approximation map if,
(a) for every (x, y, s) ∈ Rd × S, F (x, y, s) is a convex and compact subset of Rk ,
(b) for every (x0 , y0 , s0 ) ∈ Rd × S, for every Rd × S sequence, say {(xn , yn , sn )}n≥1 converging to
(x0 , y0 , s0 ) and a sequence {zn ∈ F (xn , yn , sn )}n≥0 converging to z ∈ Rk , we have that z ∈ F (x0 , y0 , s0 ),
(c) there exists K > 0 such that for every (x, y, s) ∈ Rd × S, supz∈F (x,y,s) kzk ≤ K (1 + k(x, y)k).
For SAM appearing in two-timescale stochastic recursive inclusions the condition (c) stated above is
replaced by an equivalent condition,
(c)′ there exists K > 0 such that for every (x, y, s) ∈ Rd × S, supz∈F (x,y,s) kzk ≤ K (1 + kxk + kyk).
The condition (b) in the definition of SAM tells us that the graph of the set-valued map F , G (F ),
defined as
G (F ) := (x, y, s, z) : z ∈ F (x, y, s), (x, y, s) ∈ Rd × S ⊆ Rd × S × Rk ,
is closed and hence the said condition is known as the closed graph property. The condition (c) (or (c)′ ) is
known as the point-wise boundedness condition and it makes sure that the ‘size’ of the sets grow linearly
with the distance from the origin. This is the only condition where we differ from the conditions imposed
in [12, 9]. It is easy to show that, when the Markov noise component is absent, condition (c) (or (c)′ )
imposed in this paper is the same as the one in [12] ([9]).
As a consequence of the properties possessed by a SAM, F , one can show that the map F is u.s.c. This
claim follows from arguments similar to that in [14, ch. 1.1, Cor. 1] and is stated as a lemma below.
Lemma 2.1 [u.s.c.] A set-valued map F which is a SAM is u.s.c.
The graph of a convex and compact u.s.c. set-valued map can be embedded in the graph of a sequence
of decreasing continuous set-valued maps. The following statement is made precise in the following lemma.
Lemma 2.2 [continuous embedding] For any set-valued map F , a SAM,
there exists a sequence of set
valued maps F (l) l≥1 such that for every l ≥ 1, F (l) : Rd × S → subsets of Rk is continuous and
satisfies the following.
5
(i) For every (x, y, s) ∈ Rd × S, F (l) (x, y, s) is a convex and compact subset of Rk .
(ii) For every (x, y, s) ∈ Rd × S, F (x, y, s) ⊆ F (l+1) (x, y, s) ⊆ F (l) (x, y, s).
(iii) There exists K (l) > 0 such that for every (x, y, s) ∈ Rd × S, supz∈F (l) (x,y,s) kzk ≤ K (l) (1 + k(x, y)k).
(If the set-valued map F satisfies condition (c)′ instead of (c) in the definition of SAM, we have
supz∈F (l) (x,y,s) kzk ≤ K (l) (1 + kxk + kyk)).
Furthermore, for every (x, y, s) ∈ Rd × S, ∩l≥1 F (l) (x, y, s) = F (x, y, s).
The statement of the above lemma can be found in [14, pg. 39] and the proof is similar to the proof of
[14, ch. 1.13, Thm. 1] (a brief outline can be found in [11, Appendix A]). The following are some useful
observations from the proof of Lemma 2.2.
(1) supl≥1 K (l) is finite and let K̃ := supl≥1 K (l) ,
(2) for every (x, y, s) ∈ Rd × S, for every ǫ > 0, there exists L (depending on ǫ and (x, y, s)), such that
for every l ≥ L, F (l) (x, y, s) ⊆ F (x, y, s) + ǫU where U denotes the closed unit ball in Rk .
Continuous set-valued maps admit a parametrization by which we mean that a continuous single-valued
map can be obtained which represents the set-valued map in the sense made precise in the lemma below
which follows from [14, ch. 1.7, Thm. 2]
Lemma 2.3 [parametrization] Let F be a SAM and for every l ≥ 1, the set-valued map F (l) be as in
Lemma 2.2. Then for every l ≥ 1 there exists a continuous single-valued map f (l) : Rd × S × U → Rk
where U denotes the closed unit ball in Rk , such that,
(i) for every (x, y, s) ∈ Rd ×S, F (l) (x, y, s) = f (l) (x, y, s, U ) where f (l) (x, y, s, U ) = f (l) (x, y, s, u) : u ∈ U ,
(ii) for K (l) as in Lemma 2.2(iii), for every (x, y, s, u) ∈ Rd × S × U , we have that f (l) (x, y, s, u) ≤
K (l) (1 + k(x, y)k) (If the set-valued map F satisfies condition (c)′ instead of (c) in the definition of
SAM, we have f (l) (x, y, s, u) ≤ K (l) (1 + kxk + kyk)).
Throughout this paper we shall use U to denote the closed unit ball in Rk where the dimension k will
be made clear by the context.
Combining Lemma 2.2 and Lemma 2.3 we obtain the approximation theorem stated below.
Theorem 2.4 [approximation] For any SAM F , there exists a sequence of continuous functions f (l) l≥1
such that for every l ≥ 1, f (l) : Rd × S × U → Rk is such that,
(i) for every (x, y, s) ∈ Rd × S, F (x, y, s) ⊆ f (l+1) (x, y, s, U ) ⊆ f (l) (x, y, s, U ) and f (l) (x, y, s, U ) is a
convex and compact subset of Rk ,
(ii) for K (l) as in Lemma 2.2(iii), for every (x, y, s, u) ∈ Rd × S × U , we have that f (l) (x, y, s, u) ≤
K (l) (1 + k(x, y)k) (If the set-valued map F satisfies condition (c)′ instead of (c) in the definition of
SAM, we have f (l) (x, y, s, u) ≤ K (l) (1 + kxk + kyk)).
Furthermore, for every (x, y, s) ∈ Rd × S, F (x, y, s) = ∩l≥1 f (l) (x, y, s, U ).
2.2
Measurable set-valued maps and integration
In this section we shall review concepts of measurability and integration of set-valued maps. These concepts
will be needed to define the limiting differential inclusion which the recursion studied later in this paper
is expected to track.
Let (W, FW ) denote a measurable space and F : W → subsets of Rk be a set-valued map such that,
for every w ∈ W, F (w) is a non-empty closed subset of Rk . Throughout this subsection F refers to the
set-valued map as defined above.
Definition [measurable set-valued map] A set-valued map F is measurable if for every C ⊆ Rk , closed,
F −1 (C) := {w ∈ W : F (w) ∩ C 6= ∅} ∈ FW .
We refer the reader to [15, Thm. 1.2.3] for other notions of measurability and their relation to the definition
above.
6
Definition [measurable selection] A function f : W → Rk is a measurable selection of a set-valued map
F if, f is measurable and for every w ∈ W, f (w) ∈ F (w).
For a set-valued map F let, S (F ) denote the set of all measurable selections. The next lemma summarizes
some standard results about measurable set-valued maps and their measurable selections.
Lemma 2.5 For any measurable set-valued map F ,
(i) S (F ) 6= ∅.
(ii) (Castaing representation) there exists {fn }n≥1 ⊆ S (F ) such that, for every w ∈ W, F (w) =
cl({fn (w)}n≥1 ), where cl(·) denotes the closure of a set.
We refer the reader to [15, Thm. 1.2.6] and [15, Thm. 1.2.7] for the proofs of Lemma 2.5(i) and (ii)
respectively.
Definition [µ-integrable set-valued map] Let µ be a probability measure on (W, FW ). A measurable
set-valued map F is said to be µ-integrable if, there exists f ∈ S (F ) which is µ-integrable.
Definition [Aumann’s integral] Let µ be a probability measure on (W, FW ). The integral of a µ-integrable
set-valued map F is defined as,
Z
Z
F (w)µ(dw) :=
f (w)µ(dw) : f ∈ S (F ), f is µ − integrable .
W
W
The next lemma states a useful result on the properties of the integral of a set-valued map which is
convex and compact set valued.
Lemma 2.6 Let µ be a probability measure on (W, FRW ) and F a µ-integrable set-valued map such that,
for every w ∈ W, F (w) is convex and compact. Then, W F (w)µ(dw) is a convex and closed subset of Rk .
For a proof of the above lemma we refer the reader to [15, Thm. 2.2.2].
Now we shall briefly investigate the measurability properties of a SAM. First we shall define slices of
a SAM,
, for each (x, y) ∈ Rd and for each y ∈ Rd2 . As shown in Lemma 2.2, when F is a SAM there
F(l)
exists F
a sequence of continuous set-valued maps which approximate F and for every l ≥ 1, the
l≥1
set-valued map F (l) can be parametrized with single-valued maps f (l) as in Lemma 2.3. We shall define
similar slices of F (l) and f (l) as well.
Definition Let F : Rd × S → subsets of Rk be a SAM. Let F (l) l≥1 and f (l) l≥1 be as in Lemma
2.2 and Lemma 2.3 respectively.
(i) For every (x, y) ∈ Rd , define F(x,y) : S → subsets of Rk such that for every s ∈ S, F(x,y) (s) :=
F (x, y, s).
(l)
(ii) For every l ≥ 1, for every (x, y) ∈ Rd , define F(x,y) : S → subsets of Rk such that for every s ∈ S,
(l)
F(x,y) (s) := F (l) (x, y, s).
(l)
(iii) For every l ≥ 1, for every (x, y) ∈ Rd , define f(x,y) : S × U → Rk such that for every (s, u) ∈ S × U ,
(l)
f(x,y)(s, u) := f (l) (x, y, s, u).
(iv) For every y ∈ Rd2 , define Fy : Rd1 × S → subsets of Rk
Fy (x, s) := F (x, y, s).
(l)
(v) For every l ≥ 1, for every y ∈ Rd2 , define Fy
(l)
(x, s) ∈ Rd1 × S, Fy (x, s) := F (l) (x, y, s).
such that for every (x, s) ∈ Rd1 × S,
: Rd1 × S → subsets of Rk
such that for every
(l)
(vi) For every l ≥ 1, for every y ∈ Rd2 , define fy : Rd1 × S × U → Rk such that for every (x, s, u) ∈
(l)
Rd1 × S × U , fy (x, s, u) := f (l) (x, y, s, u).
The next two lemmas summarize properties that the slices inherit from the maps F, F (l) and f (l) . Let
B(S) denote the Borel sigma algebra associated with the metric space (S, dS ).
7
Lemma 2.7 Let F : Rd × S → subsets of Rk
be a SAM. Let F (l)
d
2.2 and Lemma 2.3 respectively. For every (x, y) ∈ R , let F(x,y) ,
Definition 2.2. Then for every (x, y) ∈ Rd ,
l≥1
(l)
F(x,y)
and f (l)
and
(l)
f(x,y)
l≥1
be as in Lemma
denote the slices as in
(i) F(x,y) is a measurable set-valued map and for every s ∈ S, F(x,y) (s) is a convex and compact subset of
Rk . Further, there exists C(x,y) = K(1 + k(x, y)k) > 0 such that for every s ∈ S, supz∈F(x,y) (s) kzk ≤
C(x,y) . (If F satisfies condition (c)′ instead of condition (c) in the definition of SAM, we have
C(x,y) = K(1 + kxk + kyk)).
(l)
(l)
(ii) for every l ≥ 1, F(x,y) is a measurable set-valued map and for every s ∈ S, F(x,y) (s) is a convex and
(l)
compact subset of Rk . Further, there exists C(x,y) = K (l) (1 + k(x, y)k) > 0 such that for every s ∈ S,
supz∈F (l)
(x,y)
(l)
(s)
kzk ≤ C(x,y) . (If F satisfies condition (c)′ instead of condition (c) in the definition of
(l)
SAM, we have C(x,y) = K (l) (1 + kxk + kyk)).
(iii) for any probability measure µ on (S, B(S)), every measurable selection of F(x,y) is µ-integrable and
hence F(x,y) is µ-integrable.
(l)
(iv) for every l ≥ 1, for any probability measure µ on (S, B(S)), every measurable selection of F(x,y) is
(l)
µ-integrable and hence F(x,y) is µ-integrable.
(l)
(l)
(l)
(l)
(v) for every l ≥ 1, f(x,y) is continuous and for every s ∈ S, f(x,y) (s, U ) = F(x,y) (s) and supu∈U f(x,y) (s, u) ≤
(l)
(l)
C(x,y) where C(x,y) is as in part (ii) of this lemma.
The proof of the above lemma is similar to that of [11, Lemma 4.1] and we shall provide a brief
outline here. Fix (x, y) ∈ Rd . In order to show that F(x,y) is measurable, one needs to establish that
−1
F(x,y)
(C) ∈ B(S) for any C ⊆ Rk closed. Using the closed graph property of F one can show that
−1
F(x,y) (C) is closed subset of S and hence is in B(S). The bound C(x,y) and the claim that F(x,y) (s)
is convex and compact for every s ∈ S follows from conditions (c) (or (c)′ ) and (a) in the definition of
SAM respectively. Since all measurable selections of F(x,y) are bounded, they are µ-integrable for any
probability measure µ on (S, B(S)). The arguments are exactly same for the claims associated with the
slices of approximating maps F (l) , for every l ≥ 1. Finally the part (v) of the above lemma follows from
the properties of maps f (l) stated in Lemma 2.3.
Let µ be a probability measure on (Rd1 × S, B(Rd1 × S)) where B(Rd1 × S) denotes the Borel sigma
algebra on metric space Rd1 × S with metric max {kx − x′ k , dS (s, s′ )} for every (x, s), (x′ , s′ ) ∈ Rd1 × S
(in fact B(Rd1 × S) is the same as the product sigma algebra B(Rd1 ) ⊗ S). The support of the measure
µ denoted by supp(µ) is defined as a closed subset of Rd1 × S such that,
(1) µ(supp(µ)) = 1,
(2) for any other closed set A ⊆ Rd1 × S such that µ(A) = 1, we have supp(µ) ⊆ A.
For any probability measure µ on Rd1 × S such a set always exists and is unique (see [16, ch. 2, Thm. 2.1]).
Lemma 2.8 Let F : Rd × S → subsets of Rk be a SAM satisfying condition (c)′ instead of condition
(l)
(c) in the definition of SAM. Let F
and f (l) l≥1 be as in Lemma 2.2 and Lemma 2.3 respectively.
l≥1
(l)
(l)
For every (x, y) ∈ Rd , let Fy , Fy and fy denote the slices as in Definition 2.2. Then, for every y ∈ Rd2 ,
(i) Fy is a measurable set-valued map and for every (x, s) ∈ Rd1 × S, Fy (x, s) is a convex and compact
subset of Rk . Further for every (x, s) ∈ Rd1 × S, supz∈Fy (x,s) kzk ≤ Ky (1 + kxk) where Ky :=
max {K, K kyk} and K is as in condition (c)′ in the definition of SAM.
(l)
(l)
(ii) for every l ≥ 1, Fy is a measurable set-valued map and for every (x, s) ∈ Rd1 × S, Fy (x, s) is a
(l)
convex and compact subset of Rk . Further for every (x, s) ∈ Rd1 ×S, supz∈Fy (x,s) kzk ≤ Ky (1+kxk)
(l)
where Ky := max K (l) , K (l) kyk and K (l) is as in Lemma 2.2(iii).
(iii) for every probability measure µ on (Rd1 × S, B(Rd1 × S)) such that supp(µ) is a compact subset of
Rd1 × S, every measurable selection of Fy is µ-integrable and hence Fy is µ-integrable.
8
(iv) for every l ≥ 1, for every probability measure µ on (Rd1 × S, B(Rd1 × S)) such that supp(µ) is
(l)
(l)
a compact subset of Rd1 × S, every measurable selection of Fy is µ-integrable and hence Fy is
µ-integrable.
(l)
(v) for every l ≥ 1, fy
(l)
fy (x, s, u)
supu∈U
(l)
(l)
is continuous and for every (x, s) ∈ Rd1 × S, fy (x, s, U ) = Fy (x, s) and
≤
(l)
Ky (1
+ kxk) where
(l)
Ky
is as in part (ii) of this lemma.
The proof of parts (i), (ii) and (v) of the above lemma are similar to the corresponding in Lemma 2.7.
We shall provide a proof of part (iii) and the proof of part (iv) is exactly the same.
Proof Fix y ∈ Rd2 .
(iii) Consider f ∈ S (Fy ). By part (i) of this lemma we have that kf (x, s)k ≤ Ky (1 + kxk). Since
supp(µ) is a compact subset of Rd1 × S, there exists M > 0 such that for every x ∈ Rd1 for which
R
there exists s ∈ S satisfying (x, s) ∈ supp(µ), we have kxk ≤ M . Hence Rd1 ×S f (x, s)µ(dx, ds) =
R
R
R
supp(µ) f (x, s)µ(dx, ds) ≤ supp(µ) kf (x, s)k µ(dx, ds) ≤ supp(µ) Ky (1 + kxk)µ(dx, ds) ≤ Ky (1 +
M ). Therefore every measurable selection f ∈ S (Fy ) is µ-integrable and hence Fy is µ-integrable.
(l)
By Lemma 2.7(iv) and (v) we know that F(x,y) is a µ-integrable set-valued map for any probability
(l)
(l)
measure µ on (S, B(S)) and f(x,y) is a continuous parametrization of F(x,y) for every l ≥ 1 and for every
(l)
(x, y) ∈ Rd . Similarly, by Lemma 2.8(iv) and (v) we know that Fy is µ-integrable for any probability
(l)
(l)
measure µ on (Rd1 × S, B(Rd1 × S)) with compact support and fy is a continuous parametrization of Fy
d2
for every l ≥ 1 and for every y ∈ R . A natural question to ask is about the relation between integral of
(l)
(l)
(l)
(l)
map F(x,y) (or Fy ) and the integral of its parametrization f(x,y) (or fy ). The next lemma answers this
question. Before stating the lemma we introduce the following notation which will be used throughout
this paper.
Let P(· · · ) denote the space of probability measures on a Polish space ‘ · · ·′ with the Prohorov topology
(also known as the topology of convergence in distribution,see [17, ch. 2] for details). For any probability
measure ν ∈ P(S × U ), let νS ∈R P(S) denote the image of measure ν under the projection S × U → S (that
is for any A ∈ B(S), νS (A) = A×U µ(ds, du)). Similarly, for any probability measure ν ∈ P(Rd1 × S × U ),
let νRd1 ×S , νS and νRd1 belonging to P(Rd1 × S), P(S) and P(Rd1 ) respectively denote the image of
measure ν under the projections Rd1 × S × U → Rd1 × S, Rd1 × S × U → S and Rd1 × S × U → Rd1
respectively.
Lemma 2.9 Let F : Rd × S → subsets of Rk be a SAM. Let F (l) l≥1 and f (l) l≥1 be as in Lemma
(l)
(l)
2.2 and Lemma 2.3 respectively. For every l ≥ 1, for every (x, y) ∈ Rd , let F(x,y) , f(x,y) and for every
(l)
(l)
y ∈ Rd2 let Fy , fy denote the slices as in Definition 2.2.
(i) For every l ≥ 1, for every (x, y) ∈ Rd , for any probability measure µ ∈ P(S),
Z
Z
(l)
(l)
F(x,y) (s)µ(ds) =
f(x,y) (s, u)ν(ds, du) : ν ∈ P(S × U ), νS = µ .
S
S×U
(ii) Suppose F satisfies condition (c)′ instead of condition (c) in the definition of SAM. Then for every
l ≥ 1, for every y ∈ Rd2 , for any probability measure µ ∈ P(Rd1 × S) with compact support,
Z
Rd1 ×S
Fy(l) (dx, ds)µ(dx, ds) =
Z
Rd1 ×S×U
fy(l) (x, s, u)ν(dx, ds, du) : ν ∈ P(Rd1 × S × U ), νRd1 ×S = µ .
Remark For any ν ∈ P(Rd1 × S × U ) with νRd1 ×S = µ, the support of the measure ν, is contained in
supp(µ) × U , since by [17, ch. 3, Cor.3.1.2] there exists a µ a.s. unique measurable
map q : Rd1 × S →
R
d1
P(U ) such that, ν(dx, ds, du) = q(x, s, du)µ(dx, ds) and 1 = ν(R × S × U ) = Rd1 ×S×U ν(dx, ds, du) =
R
R
R
R
U q(x, s, du) µ(dx, ds) = ν(supp(µ) × U ). Therefore when
U q(x, s, du) µ(dx, ds) = supp(µ)
Rd 1 × S
supp(µ) is a compact set the support of measure ν is also compact and by Lemma 2.8(v) it is easy to
(l)
deduce that for all measures ν ∈ P(Rd1 × S × U ) with compact support, for all y ∈ Rd2 , fy is ν-integrable
for all l ≥ 1.
9
The proof of part (i) of the above lemma is exactly same as [11, Lemma 4.2]. The proof of part (ii) is
similar but with minor technical modifications and is presented below.
(ii) Fix y ∈ Rd2 , l ≥ 1 and µ ∈ P(Rd1 × S) with compact support.
R
R
(l)
(l)
Consider z ∈ Rd1 ×S Fy (x, s)µ(dx, ds). Then there exists f ∈ S (Fy ) such that z = Rd1 ×S f (x, s)µ(dx, ds).
Proof
n
o
Let G : Rd1 ×S → {subsets of U } be such that for every (x, s) ∈ Rd1 ×S, G(x, s) = u ∈ U : f (x, s) = fy(l) (x, s, u) .
(l)
(l)
(l)
By the fact that fy (x, s, U ) = Fy (x, s) and since f (x, s) ∈ Fy (x, s) for every (x, s) ∈ Rd1 × S we
(l)
have that G(x, s) is nonempty. By the continuity of fy (x, s, ·) we have that G(x, s) is closed for every
−1
d1
(x, s) ∈ R × S. For any C ⊆ U closed, G (C) ∈ B(Rd1 × S) (for a proof see [11, Appendix B]) and
hence G is measurable. Since G is measurable, by Lemma 2.5(i) we have that S (G) 6= ∅. Let g ∈
S (G) and let ĝ : Rd1 ×S → Rd1 ×S×U be such that for every (x, s) ∈ Rd1 ×S, ĝ(x, s) := (x, s, g(x, s)).
R
(l)
Let ν = µĝ −1 (push-forward measure). Clearly νRd1 ×U = µ and Rd1 ×S×U fy (x, s, u)ν(dx, ds, du) =
R
R
R
(l)
(l)
−1
(dx, ds, du) = Rd1 ×S fy (x, s, g(x, s))µ(dx, ds) = Rd1 ×S f (x, s)µ(dx, ds) =
Rd1 ×S×U fy (x, s, u)µĝ
z. Therefore L.H.S. is contained in R.H.S.
Let ν ∈ P(Rd1 × S × U ) with νRd1 ×S = µ. By [17, Cor. 3.1.2], there exists a µ a.s. unique measurable
map q : Rd1 × S → P(U ) such that ν(dx, ds, du) = q(x, s, du)µ(dx, ds). Since µ has compact support,
(l)
(see remark following
Lemma 2.9). Therefore
ν has compact support and hence fy is ν- integrable
hR
i
R
R
(l)
(l)
f (x, s, u)ν(dx, ds, du) = Rd1 ×S U fy (x, s, u)q(x, s, du) µ(dx, ds). By Lemma 2.8(v)
Rd1 ×S×U y
(l)
(l)
we know that for every (x, s) ∈ Rd1 × S, fy (x, s, U ) = Fy (x, s) and hence f (l) (x, s, U ) is convex
R (l)
(l)
and compact subset of Rk . Therefore for every (x, s) ∈ Rd1 × S, U fy (x, s, u)q(x, s, du) ∈ Fy (x, s).
R
Let f : Rd1 × S → Rk be such that for every (x, s) ∈ Rd1 × S, f (x, s) := U f (l) (x, s, u)q(x, s, du).
R
(l)
(l)
Then clearly, f is measurable and f ∈ S (Fy ). Therefore, Rd1 ×S×U fy (x, s, u)ν(dx, ds, du) =
h
i
R
R
R (l)
R
(l)
U fy (x, s, u)q(x, s, du) µ(dx, ds) = Rd1 ×S f (x, s)µ(dx, ds) ∈ Rd1 ×S Fy (x, s)µ(dx, ds).
Rd1 ×S
The above gives us that R.H.S. is contained in L.H.S
2.3
Differential inclusions and their limit sets
In this section we shall review results from the theory of differential inclusions and state definitions of
limit sets associated with such dynamical systems which are used later in the paper. Most of the results
in this section are taken from [12].
First we shall define a set-valued map whose associated differential inclusion (DI) is known to admit
at-least one solution through every initial condition. Such set-valued maps are called Marchaud maps and
the definition of such a map is stated below.
Definition [Marchaud map] F : Rk → subsets of Rk is a Marchaud map if,
(i) for every z ∈ Rk , F (z) is a convex and compact subset of Rk ,
(ii) there exists K > 0 such that for every z ∈ Rk , supz′ ∈F (z) kz ′ k ≤ K(1 + kzk),
(iii) for every z ∈ Rk , for every Rk -valued sequence, {zn }n≥0 converging to z ∈ Rk , for every sequence
{zn′ ∈ F (zn )}n≥0 converging to z ′ ∈ Rk , we have that z ′ ∈ F (z).
Let F be a Marchaud map. Then the DI associated with the map F is given by,
dz
∈ F (z).
dt
(10)
Since F is a Marchaud map, it is known that the DI (10), admits at-least one solution through every initial
condition (see [12, sec. 1.2]). By a solution of DI (10) with initial condition z ∈ Rk , we mean a function
z : R → Rk such that z(·) is absolutely continuous, z(0) = z and for a.e. t ∈ R, dz(t)
dt ∈ F (z(t)).
Now we shall recall the notions of flow, invariant sets, attracting sets, attractors, basin of attraction
and internally chain transitive sets. All of these notions are taken from
[12].
The flow of DI (10) is given by the set-valued map Φ : Rk × R → subsets of Rk such that for every
(z, t) ∈ Rk × R,
Φ(z, t) := {z(t) : z is a solution of DI (10) with z(0) = z} .
For any set A ⊆ Rk , let Φ(A, t) := ∪z∈A Φ(z, t).
10
A closed set A ⊆ Rk is invariant for the flow Φ of DI (10) if for every z ∈ A, there exists a solution
z(·) of DI (10) such that, z(0) = z and for every t ∈ R, z(t) ∈ A.
A compact set A ⊆ Rk is an attracting set for the flow Φ of DI (10), if there exists an open neighborhood of A, say O, with the property that for every ǫ > 0, there exists T > 0 (depending only on ǫ > 0)
such that for every t ≥ T , Φ(O, t) ⊆ N ǫ (A), where N ǫ (A) stands for the ǫ-neighborhood of A.
A compact set A ⊆ Rk is an attractor for the flow Φ of DI (10), if A is an attracting set and is
invariant for the flow Φ of DI (10).
For any z ∈ Rk , ωΦ (z) := ∩t≥0 Φ(z, [t, ∞)) where Φ(z, [t, ∞)) := ∪q≥t Φ(z, q). For any set A ⊆ Rk , the
basin of attraction of set A is denoted by B(A) and is defined as,
B(A) := z ∈ Rk : ωΦ (z) ⊆ A .
If A ⊆ Rk is an attractor whose basin of attraction is the whole of Rk (i.e. B(A) = Rk ) then A is
called a global attractor.
Given a set A ⊆ Rk and z, z ′ ∈ A, for any ǫ > 0 and T > 0 there exists an (ǫ, T ) chain from z to z ′ for
DI(10) if there exists an integer n ∈ N, solutions z1 , . . . , zn to DI (10) and real numbers t1 , . . . , tn greater
than T such that
• for all i ∈ {1, . . . , n} and for all q ∈ [0, ti ], zi (q) ∈ A,
• for all i ∈ {1, . . . , n}, k zi (ti ) − zi+1 (0) k≤ ǫ,
• k z1 (0) − z k≤ ǫ and k zn (tn ) − z ′ k≤ ǫ.
A compact set A ⊆ Rd is said to be internally chain transitive if for every z, z ′ ∈ A, for every ǫ > 0
and for every T > 0, there exists (ǫ, T ) chain from z to z ′ for DI (10).
Suppose L ⊆ Rk is an invariant
set. Then the flow of DI (10) restricted to the invariant set L is a
set-valued map, ΦL : L × R → subsets of Rk such that for every (z, t) ∈ L × R,
ΦL (z, t) := {z(t) : z(·) is a solution of DI (10) with z(0) = z and f or every t ∈ R, z(t) ∈ L} .
2.4
(11)
Single timescale stochastic recursive inclusions with non-additive iteratedependent Markov noise
In this section we review results from the analysis of single timescale stochastic recursive inclusions with
non-additive iterate dependent Markov noise. All of the results presented here can be found in [11].
Let (Ω, F , P) be a probability space and {Zn }n≥0 be a sequence of Rd -valued random variables satisfying
Zn+1 − Zn − a(n)Mn+1 ∈ a(n)F (Zn , Sn ),
(12)
where the following assumptions hold:
S(A1) the map F : Rd × S → subsets of Rd where (S, dS ) is a compact metric space is such that,
(i) for every (z, s) ∈ Rd × S, F (z, s) is a convex and compact subset of Rd ,
(ii) there exists K > 0 such that for every (z, s) ∈ Rd × S, supz′ ∈F (z,s) kz ′ k ≤ K(1 + kzk),
(iii) for every z ∈ Rd , for every Rd × S valued sequence, say {(zn , sn )}n≥1 converging to (z, s) and
for any sequence {zn′ ∈ F (zn , sn )}n≥1 converging to z ′ we have z ′ ∈ F (z, s).
S(A2) {Sn }n≥0 is a sequence of S-valued measurable functions on Ω such that for every n ≥ 0, for every
A ∈ B(S), P(Sn+1 ∈ A|Sm , Zm , m ≤ n) = P(Sn+1 ∈ A|Sn , Zn ) = Π(Zn , Sn )(A) a.s., where Π :
Rd × S → P(S) is continuous.
S(A3) {a(n)}n≥0 is a sequence of positive real numbers satisfying,
(i) a(0) ≤ 1 and for every n ≥ 0, a(n) ≥ a(n + 1),
P∞
P∞
2
(ii)
n=0 a(n) = ∞ and
n=0 (a(n)) < ∞,
S(A4) {Mn }n≥1 is a sequence of Rd -valued random variables on Ω such that forna.s.(ω), for any T > 0, o
Pk
Pm−1
limn→∞ supn≤k≤τ (n,T )
m=n a(m)Mm+1 (ω) = 0 where τ (n, T ) := min m > n :
k=n a(k) ≥ T .
S(A5) P(supn≥0 kXn k < ∞) = 1.
11
A detailed motivation for each of these assumptions can be found in [11]. We shall briefly explain them
and their consequences.
Assumption S(A1) ensures that the set-valued map F is a SAM and assumption S(A2) is the iteratedependent Markov noise assumption. As a consequence of assumption S(A2), for every z ∈ Rd we know
that the Markov chain defined by the transition kernel Π(z, ·)(·), possesses the weak Feller property (see
[18]). In addition to the above since the state space is compact, the set of stationary distributions for
the Markov chain whose transition probability is given by Π(z, ·)(·) is non-empty for every z ∈ Rd . Let
D(z) ⊆ P(S) denote the set of stationary distributions of the Markov chain whose
transition kernel is
R
Π(z, ·)(·) (for any z ∈ Rd , µ ∈ D(z) if and only if for every A ∈ B(S), µ(A) = S Π(z, s)(A)µ(ds)). We
also know that for every z ∈ Rd , D(z) is a convex and compact subset of P(S) and the map z → D(z)
has closed graph (see [11] and references therein). Assumption S(A3) is the standard step-size assumption
and assumption S(A4) is the general additive noise assumption which ensures that the contribution of
the additive noise is eventually negligible (for various noise models satisfying S(A4) see [12]). Assumption
S(A5) is the stability assumption on the
iterate sequence.
The set-valued map, F̂ : Rd → subsets of Rd that serves as the vector filed for the differential
inclusion (DI) that the iterates are expected to track is defined as,
Z
Fz (s)µ(ds),
F̂ (z) := ∪µ∈D(z)
S
for every z ∈ Rd where for every z ∈ Rd , Fz denotes the slice as in Definition 2.2(i) of the set-valued map
F appearing in recursion (12). The set-valued map F̂ , is a Marchaud map (see [11, Lemma 4.7]) and the
associated DI given by,
dz
∈ F̂ (z)
(13)
dt
admits at-least one solution through every initial condition (see [12, sec. 1.2]). Let Σ(z0 ) denote the set of
solutions of DI (13) with initial condition z0 ∈ Rd and Σ := ∪z0 ∈Rd Σ(z0 ) (the set of all possible solutions).
For every z0 ∈ Rd , Σ(z0 ) is a subset of C(R, Rd ), the set of all Rd -valued continuous functions on R. The
set C(R, Rd ) is a complete metric space for the metric D defined by,
D(z, z′ ) :=
∞
X
1
′
min
kz
−
z
k
,
1
,
[−k,k]
2k
k=1
where kz − z′ k[−k,k] := supt∈[−k,k] kz(t) − z′ (t)k. As a consequence of [12, Lemma 3.1], we have that Σ
and for every z0 ∈ Rd , Σ(z0 ) are closed and compact subsets of C(R, Rd ) respectively.
Pn−1
Let t0 := 0 and for every n ≥ 1, t(n) := k=0 a(k). Define the stochastic process with continuous
sample paths, Z̄ : Ω × R → Rd as,
t − t(n)
t(n + 1) − t
Z̄(ω, t) :=
Zn+1 (ω) +
Zn (ω),
t(n + 1) − t(n)
t(n + 1) − t(n)
for every (ω, t) ∈ Ω × [0, ∞) where n is such that t ∈ [t(n), t(n + 1)) and for every (ω, t) ∈ Ω × (−∞, 0], let
Z̄(ω, t) := Z0 (ω). Then the main result from the analysis of recursion (12) in [11] is as follows.
Theorem 2.10 Under assumptions S(A1) − S(A5), for almost every ω ∈ Ω,
(i) the family of functions Z̄(ω, · + t) t≥0 is relatively compact in C(R, Rd ),
(ii) every limit point of Z̄(ω, · + t) t≥0 in C(R, Rd ) is a solution of DI (13), more formally,
lim D(Z̄(ω, · + t), Σ) = 0,
t→∞
(iii) the limit set denoted by L(Z̄(ω, ·)) defined as
L(Z̄(ω, ·)) := ∩t≥0 Z̄(ω, q + t) : q ≥ 0 ,
is non-empty, compact and internally chain transitive for the flow of DI (13).
For a proof of the above theorem see [11, Thm. 6.6 & 6.7].
12
2.5
Space of probability measure valued functions
In this section we shall define the space of probability measure valued measurable functions on [0, ∞). We
shall introduce an appropriate topology on this space and show that such a space is compact metrizable.
These spaces are used in the theory of optimal control of diffusions (see [19]) and also in analyzing stochastic
approximation schemes (see [11, 13]). Quantities defined in this section will serve as tools in analyzing the
stochastic recursions later.
Throughout this section U will denote the closed unit ball in Rd2 and for any r > 0, Br denotes the
closed ball of radius r in Rd1 centered at the origin. For every r > 0, let M(U × Br × S) denote the set of
all functions γ(·) on [0, ∞) taking values in P(U × Br × S) (space of probability measures on U × Br × S
equipped with the Prohorov topology), such that γ(·) is measurable. Formally,
M(U × Br × S) := {γ : [0, ∞) → P(U × Br × S) : γ(·) is measurable} .
Similarly for every r > 0, M(Br × S) (or M(Br )) denotes the set of all functions γ(·) on [0, ∞) taking
values in P(Br × S) (or P(Br )) such that γ(·) is measurable. Formally,
M(Br × S) := {γ : [0, ∞) → P(Br × S) : γ(·) is measurable} ,
M(Br ) := {γ : [0, ∞) → P(Br ) : γ(·) is measurable} .
For every r > 0, let τU×Br ×S
hR denote the coarsest topology on iM(U × Br × S) which renders continuous
RT
the functions, γ(·) → 0 g(t) U×Br ×S f (u, x, s)γ(t)(du, dx, ds) dt for every f ∈ C(U × Br × S, R), for
every g ∈ L2 ([0, T ], R) and for every T > 0.
Similarly, for every r > 0, let τBr ×S hdenote the coarsest topology
on M(Br × S) which renders
i
RT
R
continuous the functions, γ(·) → 0 g(t) Br ×S f (x, s)γ(t)(dx, ds) dt for every f ∈ C(Br × S, R), for
every g ∈ L2 ([0, T ], R) and for every T > 0.
Finally, for every r > 0,
hR let τBr denote the
i coarsest topology on M(Br ) which renders continuous the
RT
functions, γ(·) → 0 g(t) Br f (x)γ(t)(dx) dt for every f ∈ C(Br , R), for every g ∈ L2 ([0, T ], R) and for
every T > 0.
The following is a well known metrization lemma for the topological spaces defined above.
Lemma 2.11 [metrization]
(i) For every r > 0, the topological space (M(U × Br × S), τU×Br ×S ) is compact metrizable.
(ii) For every r > 0, the topological space (M(Br × S), τBr ×S ) is compact metrizable.
(iii) For every r > 0, the topological space (M(Br ), τBr ) is compact metrizable.
We refer the reader to [13, Lemma 2.1] for the proof of the above metrization lemma. The next lemma
provides continuous functions between the above defined metric spaces which are used later. The proof of
the lemma below is an extention of [11, Lemma 5.2] to the above defined metric spaces. Recall that for
any probability measure ν ∈ P(U × Br × S), νBr ×S ∈ P(Br × S) denotes the image of the
R measure ν under
the projection U × Br × S → Br × S (that is, for every A ∈ B(Br × S), νBr ×S (A) = U×A ν(du, dx, ds)).
Similarly, νBr ∈ P(Br ) denotes
R the image of measure ν under the projection U × Br × S → Br (that is, for
every A ∈ B(Br ), νBr (A) = U×A×S ν(du, dx, ds)). It is easy to see that νBr is also the image of νBr ×S
under the projection Br × S → Br .
Lemma 2.12 For every r > 0,
(i) the map θ1 : P(U × Br × S) → P(Br × S) such that for every ν ∈ P(U × Br × S), θ1 (ν) := νBr ×S
is continuous.
(ii) the map θ2 : P(Br × S) → P(Br ) such that for every ν ∈ P(Br × S), θ2 (ν) := νBr is continuous.
(iii) for any γ ∈ M(U × Br × S), we have that θ1 ◦ γ ∈ M(Br × S) where for every t ≥ 0, (θ1 ◦ γ)(t) =
θ1 (γ(t)).
(iv) for any γ ∈ M(Br × S), we have that θ2 ◦ γ ∈ M(Br ) where for every t ≥ 0, (θ2 ◦ γ)(t) = θ2 (γ(t)).
(v) the map Θ1 : M(U × Br × S) → M(Br × S) such that for every γ ∈ M(U × Br × S), Θ1 (γ) := θ1 ◦ γ
is continuous.
13
(vi) the map Θ2 : M(Br ×S) → M(Br ) such that for every γ ∈ M(Br ×S), Θ2 (γ) := θ2 ◦γ is continuous.
Proof Fix r > 0.
(i) Let {ν n }n≥1 be a sequence in P(U × Br × S) converging to ν ∈ P(U × Br × S) as n → ∞ and let
π : U × Br × S → Br × S denote the projection map such that for every (u, x, s) ∈ U × Br × S,
π(u, x, s) = (x, s). Clearly π is continuous and for any continuous function f ∈ C(Br × S, R), f ◦ π
is continuous. Since U × Br R× S is a compact metric space, from [17,RThm. 2.1.1(ii)], we get that
for every f ∈ C(Br × S, R), U×Br ×S (f ◦ π) (u, x, s)ν n (du, dx, ds) → U×Br ×S (f ◦ π) ν(du, dx, ds)
n
as n → ∞. By definition, we have that for every n ≥ 0, νB
= ν n π −1 (the push-forward
r ×S R
n
−1
(dx, ds) →
measure) and νBr ×S = νπ . Therefore for every f ∈ C(Br × S, R), Br ×S f (x, s)νB
r ×S
R
n
f
(x,
s)ν
(dx,
ds).
Hence
by
[17,
Thm.
2.1.1]
we
get
that
ν
→
ν
as
n → ∞ in
B
×S
B
×S
r
r
Br ×S
Br ×S
P(Br × S) which gives us continuity of θ1 (·).
(ii) Similar to part (i) of this lemma.
(iii) & (iv) Composition of measurable functions is measurable.
(v) Let {γn }n≥1 be a sequence in M(U × Br × S) converging to γ ∈ M(U × Br × S) as n → ∞.
Then weh know that for every f ∈ C(U × Bri× S, R), for every
hR T > 0 and for every g ∈ L2 ([0, iT ], R),
RT
RT
R
0 g(t)
U×Br ×S f (u, x, s)γn (t)(du, dx, ds) dt → 0 g(t)
U×Br ×S f (u, x, s)γ(t)(du, dx, ds) dt as
n → ∞. Let π denote the projection map as in part (i) of this lemma and we know that for any
f ∈ C(Br × S, R), f ◦ π ∈ C(U × Br × S, R). Thenh we have that for every f ∈ C(Br × S, R),
i for
R
RT
every T > 0 and for every g ∈ L2 ([0, T ], R), 0 g(t) U×Br ×S (f ◦ π) (u, x, s)γn (t)(du, dx, ds) dt →
hR
i
RT
g(t)
(f
◦
π)
(u,
x,
s)γ(t)(du,
dx,
ds)
dt as n → ∞. By arguments similar to part (i) of
0
U×Br ×S
this lemma,
for every
hR we have that for every f ∈ C(Bri× S, R),
hR T > 0 and for every g ∈ L2 ([0, Ti], R),
RT
RT
g(t)
f
(x,
s)
(θ
◦
γ
)
(t)(dx,
ds)
dt
→
g(t)
f (x, s) (θ1 ◦ γ) (t)(dx, ds) dt as
1
n
0
U×Br ×S
0
U×Br ×S
n → ∞. Therefore Θ1 (γn ) → Θ1 (γ) in M(Br × S) as n → ∞ which gives us continuity of Θ1 (·).
(vi) Similar to part (v) of this lemma.
3
Recursion and assumptions
In this section we shall formally define the two timescale recursion as well as state and motivate the
assumptions imposed ((A1)-(A10)).
Let (Ω, F , P) denote a probability space, {Xn }n≥0 be a sequence of Rd1 -valued random variables on
Ω and {Yn }n≥0 be a sequence of Rd2 -valued random variables on Ω which satisfy for every n ≥ 0,
(2)
(14a)
(1)
(14b)
Yn+1 − Yn − b(n)Mn+1 ∈ b(n)H2 (Xn , Yn , Sn(2) ),
Xn+1 − Xn − a(n)Mn+1 ∈ a(n)H1 (Xn , Yn , Sn(1) ),
where,
(A1) the map H1 : Rd × S (1) → subsets of Rd1
such that,
with S (1) a compact metric space with metric dS (1) , is
(i) for every (x, y, s(1) ) ∈ Rd × S (1) , H1 (x, y, s(1) ) is a convex and compact subset of Rd1 ,
(ii) there exists K > 0, such that, for every (x, y, s(1) ) ∈ Rd × S (1) , supx′ ∈H1 (x,y,s(1) ) kx′ k ≤ K(1 +
kxk + kyk),
n
o
(1)
(iii) for every (x, y, s(1) ) ∈ Rd × S (1) , for every Rd × S (1) -valued sequence, (xn , yn , sn )
n≥1
n
o
(1)
(1)
d
(1)
′
converging to (x, y, s ) ∈ R × S , for every sequence xn ∈ H1 (xn , yn , sn )
converging
n≥1
to x′ ∈ Rd1 , we have that x′ ∈ H1 (x, y, s(1) ).
(A2) the map H2 : Rd × S (2) → subsets of Rd2 with S (2) a compact metric space with metric dS (2) , is
such that,
(i) for every (x, y, s(2) ) ∈ Rd × S (2) , H2 (x, y, s(2) ) is a convex and compact subset of Rd2 ,
14
(ii) there exists K > 0, such that, for every (x, y, s(2) ) ∈ Rd × S (2) , supy′ ∈H2 (x,y,s(2) ) ky ′ k ≤ K(1 +
kxk + kyk),
n
o
(2)
(iii) for every (x, y, s(2) ) ∈ Rd × S (2) , for every Rd × S (2) -valued sequence, (xn , yn , sn )
n≥1
n
o
(2)
converging to (x, y, s(2) ) ∈ Rd × S (2) , for every sequence yn′ ∈ H2 (xn , yn , sn )
converging
n≥1
to y ′ ∈ Rd2 , we have that y ′ ∈ H2 (x, y, s(2) ).
n
o
(1)
(A3) Sn
is a sequence of S (1) -valued random variables on Ω, such that for every n ≥ 0, for
n≥0
(1)
(1)
(1)
(2)
(2)
(2)
1
every A ∈ B(S (1) ), P(Sn+1 ∈ A|Sm
, Xm , Ym , 0 ≤ m ≤ n) = P(Sn+1 ∈ A|Sn , Xn , Yn ) =
(1)
(1)
(1)
Π (Xn , Yn , Sn )(A) a.s., where Π : Rd × S (1) → P(S (1) ) is continuous.
n
o
(2)
(A4) Sn
is a sequence of S (2) -valued random variables on Ω, such that for every n ≥ 0, for
n≥0
2
every A ∈ B(S (2) ), P(Sn+1 ∈ A|Sm
, Xm , Ym , 0 ≤ m ≤ n) = P(Sn+1 ∈ A|Sn , Xn , Yn ) =
(2)
(2)
(2)
Π (Xn , Yn , Sn )(A) a.s., where Π : Rd × S (2) → P(S (2) ) is continuous.
(A5) {a(n)}n≥0 and {b(n)}n≥0 are two sequences of positive real numbers satisfying,
(i) a(0) ≤ 1 and for every n ≥ 0, a(n) ≥ a(n + 1),
(ii) b(0) ≤ 1 and for every n ≥ 0, b(n) ≥ b(n + 1),
b(n)
(iii) limn→∞ a(n)
= 0,
P∞
P∞
P∞
2
2
(iv)
< ∞.
n=0 a(n) =
n=0 b(n) = ∞ and
n=0 (a(n)) + (b(n))
(1)
(A6) {Mn }n≥1 is a sequence of Rd1 -valued random variables on Ω such that forna.e.(ω), for any T > 0, o
Pk
Pm−1
(1)
1
limn→∞ supn≤k≤τ 1 (n,T )
m=n a(m)Mm+1 (ω) = 0 where τ (n, T ) := min m > n :
k=n a(k) ≥ T .
(2)
(A7) {Mn }n≥1 is a sequence of Rd2 -valued random variables on Ω such that forna.e.(ω), for any T > 0, o
Pm−1
Pk
(2)
2
limn→∞ supn≤k≤τ 2 (n,T )
k=n b(k) ≥ T .
m=n b(m)Mm+1 (ω) = 0 where τ (n, T ) := min m > n :
(A8) P supn≥0 (kXn k + kYn k) < ∞ = 1.
Assumptions (A1) and (A2) ensure that H1 and H2 are SAMs. Assumptions (A3) and (A4) are
the iterate-dependent Markov noise assumptions. Under (A3), for every (x, y) ∈ Rd , the Markov chain
associated with the transition kernel given by Π(1) (x, y, ·)(·) possesses the weak Feller property (see [18]).
In addition to the above since S (1) is a compact metric space, the Markov chain associated with the
transition kernel Π(1) (x, y, ·)(·) has at least one stationary distribution for every (x, y) ∈ Rd (µ ∈ P(S (1) )
is stationary for the
chain associated with the transition kernel Π(1) (x, y, ·)(·) if for every A ∈
R Markov
(1)
(1)
B(S ), µ(A) = S (1) Π (x, y, s(1) )(A)µ(ds(1) )). For every (x, y) ∈ Rd , let D(1) (x, y) ⊆ P(S (1) ) denote
the set of stationary distributions of the Markov chain associated with the transition kernel Π(1) (x, y, ·)(·).
It can easily be shown that,
(i) for every (x, y) ∈ Rd , D(1) (x, y) is a convex and compact subset of P(S (1) ),
(ii) graph of the map (x, y) → D(1) (x, y) is closed, that is, the set
n
o
G(D(1) ) := (x, y, µ) ∈ Rd × P(S (1) ) : (x, y) ∈ Rd , µ ∈ D(1) (x, y) ,
is a closed subset of Rd × P(S (1) ).
The proofs of the above two statements are similar to that in [10, pg. 69]. Similarly under assumption
(A4), for every (x, y) ∈ Rd the set of stationary distributions (denoted by D(2) (x, y)) associated with the
Markov chain defined by the transition kernel Π(2) (x, y, ·)(·) is a non-empty, convex and compact subset
of P(S (2) ) and further the map (x, y) → D(2) (x, y) has a closed graph (that is the set G(D(2) ) defined in
an analogous manner as G(D(1) ) is a closed subset of Rd × P(S (2) )).
Assumption (A5) is the standard two timescale step size assumption. Assumption (A5)(iii) tells that
eventually the time step taken by recursion (14a) is smaller than the time step taken by recursion (14b).
Hence recursion (14a) is called the slower timescale recursion and the recursion (14b) is called the faster
15
timescale recursion. Assumptions (A6) and (A7) are the conditions that the additive noise terms satisfy.
These guarantee that the contribution of additive noise terms is eventually negligible. For various noise
models where these additive noise assumptions are satisfied we refer the reader to [12].
Assumption (A8) is the stability assumption which ensures that the iterates remain within a bounded
set. While this is a standard requirement in the study of such recursions, it is highly nontrivial. An
important future direction would be to provide sufficient conditions for verification of (A8).
The Markov noise terms in the faster timescale, in limit will average the drift function H1 w.r.t. the
stationary distributions given by the map (x, y) → D(1) (x, y). The appropriate set-valued map whose
associated DI the faster timescale recursion is expected to track is given by,
Z
H1,(x,y) (s(1) )µ(ds(1) ),
(15)
Ĥ1 (x, y) := ∪µ∈D(1) (x,y)
S (1)
for every (x, y) ∈ Rd where for every (x, y) ∈ Rd , H1,(x,y) denotes the slice as in Definition2.2(i) of the
set-valued map H1 in the recursion (14b). As a consequence of the step size assumption (A5), with respect
to the faster timescale (14b), the slower timescale recursion (14a) appears to be static and one would
expect that the family of DIs,
dx
∈ Ĥ1 (x, y0 ),
(16)
dt
obtained by fixing some y0 ∈ Rd2 to describe the behavior of the faster timescale recursion (14b). Before
we proceed, we need to ensure that for every y0 ∈ Rd2 , the DI (16) has solutions through every initial
condition. The next lemma states the map Ĥ1 (·, y0 ) is a Marchaud map for every y0 ∈ Rd2 , which ensures
that the DI (16) has solutions.
Lemma 3.1 For every y0 ∈ Rd2 , the set-valued map Ĥ1 (·, y0 ) : Rd1 → subsets of Rd1 is a Marchaud
map.
Proof of the above lemma is given in section 4. The next assumption will ensure that for every y0 ∈ Rd2 ,
the DI (16) has a global attractor to which one expects the faster time scale iterates {Xn }n≥0 to converge.
d2
(A9) For
a globally attracting set, Ay0 . The map λ : Rd2 →
every y0 d∈ R , the DI (16) admits
1
where for every y ∈ Rd2 , λ(y) := Ay is such that
subsets of R
(i) for every y ∈ Rd2 , supx∈λ(y) kxk ≤ K(1 + kyk),
(ii) for every y ∈ Rd2 , for every Rd2 -valued sequence, {yn }n≥1 converging to y ∈ Rd2 , for every
{xn ∈ λ(yn )}n≥0 converging x ∈ Rd1 , we have x ∈ λ(y).
With respect to the slower timescale recursion (14a), the faster time scale recursion will appear to have
equilibrated. Further the Markov noise terms average the set-valued drift function H2 with respect to the
stationary distributions. In what follows we construct the set-valued map that the slower timescale recursion is expected to track which captures both the equilibration of the faster timescale and the averaging
by the Markov noise terms.
Before we proceed recall that P(Rd1 × S (2) ) denotes the set of probability measures on Rd1 × S (2)
with the Prohorov topology. For any µ ∈ P(Rd1 × S (2) ), µRd1 ∈ P(Rd1 ) and µS (2) ∈ P(S (2) ) denote
the images of the probability measure µ underR projections Rd1 × S (2) → Rd1 and Rd1 × S (2) → S (2)
respectively (for any A ∈ B(Rd1 ), µRd1 (A) := A×S (2) µ(dx, ds(2) ) and similarly for every A ∈ B(S (2) ),
R
µS (2) (A) := Rd1 ×A µ(dx, ds(2) )).
Define the map D : Rd2 → subsets of P(Rd1 × S) such that for every y ∈ Rd2 ,
Z
(2)
d1
(2)
(2)
(2)
(2)
D(y) := µ ∈ P(R × S ) : supp(µRd1 ) ⊆ λ(y) and f or every A ∈ B(S ), µS (2) (A) = Π (x, y, s )(A)µ(dx, ds ) ,
S (2)
(17)
where supp(µRd1 ) denotes the support of measure µRd1 (that is supp(µRd1 ) ⊆ Rd1 is a closed set such
that µRd1 (supp(µRd1 )) = 1 and for every closed set A ⊆ Rd1 , with µRd1 (A) = 1 we have supp(µRd1 ) ⊆ A).
A natural question to ask is whether D(y) is non-empty for every y ∈ Rd2 and if it is non-empty, what
properties
n the
o map D(·) possesses and its relation to the stationary distributions of the Markov noise
terms
(2)
Sn
n≥0
. The lemma below answers these questions.
Lemma 3.2 The map D(·) defined in (17) satisfies,
16
(i) for every y ∈ Rd2 , D(y) is non-empty, convex and compact subset of P(Rd1 × S (2) ),
(ii) for every y ∈ Rd2 , for every Rd2 -valued sequence, {yn }n≥1 converging to y ∈ Rd2 , for every P(Rd1 ×
S (2) )-valued sequence {µn ∈ D(yn )}n≥1 converging to µ ∈ P(Rd1 × S (2) ), we have µ ∈ D(y).
¯ δx∗ ⊗ ν ∈ P(Rd1 × S (2) ) : x∗ ∈ λ(y), ν ∈ D(2) (x∗ , y) ⊆ D(y), where for
(iii) for every y ∈ Rd2 , co
any x ∈ Rd1 , δx denotes the Dirac measure.
Proof
(i) Fix y ∈ Rd2 . Consider the product measure µ := δx∗ ⊗ ν ∈ P(Rd1 × S (2) ) where, δx∗ ∈ P(Rd1 )
denotes the Dirac measure on some x∗ ∈ λ(y) (that is for every A ∈ B(Rd1 ), δx∗ (A) = 1 if
x∗ ∈ A, δx∗ (A) = 0 otherwise) and ν ∈ P(S (2) ) is such that ν ∈ D(2) (x∗ , y) (that is ν is a stationary measure of the Markov chain whose transition kernel is given by Π(2) (x∗ , y, ·)(·)). Then
∗
∈ λ(y), supp(µRd1 ) = {xR∗ } ⊆Rλ(y). Further µS (2) = ν and
µRd1 = δx∗ and
for every
R since x (2)
(2)
A ∈ B(S ), Rd1 ×S (2) Π (x, y, s)(A)µ(dx, ds) = S (2) Rd1 Π(2) (x, y, s(2) )(A)δx∗ (dx) ν(ds(2) ) =
R
Π(2) (x∗ , y, s(2) )(A)ν(ds(2) ) = ν(A) where the last equality follows from the fact that ν ∈
S (2)
D(2) (x∗ , y). Therefore δx∗ ⊗ ν ∈ D(y) and hence D(y) 6= ∅.
Let µ1 , µ2 ∈ D(y) and α ∈ (0, 1). Consider the measure µ := αµ1 + (1 − α)µ2 (that is for any A ∈
B(Rd1 ×S (2) ), µ(A) = αµ1 (A)+(1−α)µ2 (A)). Clearly µ ∈ P(Rd1 ×S (2) ), µRd1 = αµ1Rd1 +(1−α)µ2Rd1
and µS (2) = αµ1S (2) +(1−α)µ2S (2) . For i ∈ 1, 2, supp(µiRd1 ) ⊆ λ(y), from which we have µiRd1 (λ(y)) = 1
and hence µRd1 (λ(y)) = αµ1Rd1 (λ(y))+(1−α)µ2Rd1 (λ(y)) = 1. Therefore supp(µRd1 ) ⊆ λ(y). For every
R
R
A ∈ B(S (2) ), Rd1 ×S (2) Π(2) (x, y, s(2) )(A)µ(dx, ds(2) ) = α Rd1 ×S (2) Π(2) (x, y, s(2) )(A)µ1 (dx, ds(2) ) +
R
(1 − α) Rd1 ×S (2) Π(2) (x, y, s(2) )(A)µ2 (dx, ds(2) ) = αµ1S (2) (A) + (1 − α)µ2S (2) (A) = µS (2) (A). Therefore
µ := αµ1 + (1 − α)µ2 ∈ D(y) which gives us the convexity of D(y).
In order to show that D(y) is compact, we will first show that the set D(y) is a closed set. Consider {µn }n≥1 such that for every n ≥ 1, µn ∈ D(y) converging to µ ∈ P(Rd × S (2) ). Clearly
{µnRd1 }n≥1 converges to µRd1 in P(Rd1 ). Since for every n ≥ 1, since supp(µnRd1 ) ⊆ λ(y), we have
µn (λ(y)) = 1 for every n ≥ 1. By assumption (A9), λ(y) is a compact subset of Rd1 and by [17,
Thm. 2.1.1(iv)], we have lim supn→∞ µnRd1 (λ(y)) ≤ µRd1 (λ(y)). Therefore µRd1 (λ(y)) = 1 which gives
us that supp(µRd1 ) ⊆ λ(y). Clearly µnS (2) n≥1 converges to µS (2) in P(S (2) ). Since S (2) is a compact
R
metric space, by [17, Thm. 2.1.1(ii)] we know that for every f ∈ C(S (2) , R), S (2) f (s̃(2) )µnS (2) (ds̃(2) ) →
R
R
(2)
)µS (2) (ds̃(2) ) as n → ∞. Let ν n (ds̃2 ) := Rd1 ×S (2) Π(2) (x, y, s(2) )(ds̃(2) )µ(dx, ds(2) ) ∈
S (2) f (s̃
R
P(S (2) ) for every n ≥ 1 and ν(ds̃(2) ) := Rd1 ×S (2) Π(2) (x, y, s(2) )(ds̃(2) )µ(dx, ds(2) ). It is easy to see
R
R
R
that for any f ∈ C(S (2) , R), S (2) f (s̃(2) )ν n (ds̃(2) ) = Rd1 ×S (2) S (2) f (s̃(2) )Π(2) (x, y, s(2) )(ds̃(2) ) µn (dx, ds(2) ).
R
By assumption (A4), (x, s(2) ) → S (2) f (s̃(2) )Π(2) (x, y, s(2) )(ds̃(2) ) is continuous for any f ∈ C(S (2) , R).
R
R
Therefore as µn → µ in P(Rd1 ×S (2) ), we have Rd1 ×S (2) S (2) f (s̃(2) )Π(2) (x, y, s(2) )(ds̃(2) ) µn (dx, ds(2) ) →
R
R
R
R
f (s̃(2) )Π(2) (x, y, s(2) )(ds̃(2) ) µ(dx, ds(2) ) or S (2) f (s̃(2) )ν n (ds̃(2) ) → S (2) f (s̃(2) )ν(ds̃(2) ).
S (2)
Rd1 ×S (2)
R
R
Since for every n ≥ 1, µn ∈ D(y), we have S (2) f (s̃(2) )µnS (2) (ds̃(2) ) = S (2) f (s̃(2) )ν n (ds̃(2) ) for every
R
R
f ∈ C(S (2) , R). Thus for every f ∈ C(S (2) , R), we have S (2) f (s̃(2) )µS (2) (ds̃(2) ) = S (2) f (s̃(2) )ν(ds̃(2) ).
Therefore µS (2) = ν which establishes that µ ∈ D(y) and hence D(y) is closed. To establish compactness of D(y) it is now enough to show that the set D(y) is relatively compact in P(Rd1 × S (2) ). For
any measure µ ∈ D(y), the support of the measure µ, denoted by supp(µ) is contained in λ(y) × S (2)
which is a compact set independent of µ ∈ D(y). Thus the family of measures {µ : µ ∈ D(y)} is tight
and by Prohorov’s theorem (see [17, Thm. 2.3.1]) we have that the set of measures D(y) is relatively
compact in P(Rd1 × S (2) ). Therefore D(y) is closed and relatively compact and hence is compact.
(ii) Let yn → y in Rd2 and µn ∈ D(yn ) → µ in P(Rd1 ×S (2) ) as n → ∞. Let B1 denote the closed unit ball
in Rd1 . By assumption (A9), we have that the set-valued map y → λ(y) is u.s.c. Therefore for every
ǫ > 0, there exists δ > 0(depending on ǫ and y) such that for every y ′ ∈ Rd1 , satisfying ky ′ − yk < δ
we have λ(y ′ ) ⊆ λ(y) + ǫB1 . Since λ(y) is compact, λ(y) + ǫB1 is compact. Since yn → y, there
exists N such that for every n ≥ N , kyn − yk < δ. Then for all n ≥ N , λ(yn ) ⊆ λ(y) + ǫB1 . By the
above we have that lim supn→∞ µnRd1 (λ(y) + ǫ(B1 )) = 1 for every ǫ > 0. Since µn → µ, we have that
µnRd1 → µRd1 in P(Rd1 ) and by [17, Thm. 2.1.1(iv)], we have that for every ǫ > 0, µRd1 (λ(y)+ǫB1 ) =
1. Since λ(y) is compact, λ(y) = ∩n≥1 (λ(y) + n1 B1 ) and µRd1 (λ(y)) = limn→∞ µRd1 (λ(y) + n1 B1 ) =
R
1. Therefore supp(µRd1 ) ⊆ λ(y). Let ν n (ds̃(2) ) := Rd1 ×S (2) Π(2) (x, yn , s(2) )(ds̃2 )µn (dx, ds(2) ) ∈
17
R
P(S 2 ) and ν(ds̃(2) ) := Rd1 ×S (2) Π(2) (x, y, s(2) )(ds̃(2) )µ(dx, ds(2) ). Then for any f ∈ C(S (2) , R),
R
R
R
for any n ≥ 1, S (2) f (s̃(2) )ν n (s̃(2) ) = Rd1 ×S (2) S (2) f (s̃(2) )Π(2) (x, yn , s(2) )(ds̃(2) ) µn (dx, ds(2) ) and
R
R
R
(2)
)ν(s̃(2) ) = Rd1 ×S (2) S (2) f (s̃(2) )Π(2) (x, y, s(2) )(ds̃(2) ) µ(dx, ds(2) ). Since for every n ≥ 1,
S (2) f (s̃
µn ∈ D(yn ), we have that supp(µn ) ⊆ λ(yn ) × S 2 . By using the u.s.c. property of the map
λ(·) and the fact that yn → y, we get that for any ǫ > 0, there exists N such that for ev(2)
ery n ≥ N
⊆ (λ(y)R+ ǫB1 ) × S (2) .RTherefore for every f ∈ C(S (2), R), for every
R , λ(yn ) × S
n ≥ N , S (2) f (s̃(2) )ν n (ds̃(2) ) = (λ(y)+ǫB1 )×S (2) S (2) f (s(2) )Π(2) (x, yn , s(2) )(ds̃(2) ) µn (dx, ds(2) ).
R
By assumption (A4), the map (x, y, s(2) ) → S (2) f (tildes(2) )Π(2) (x, y, s(2) )(ds̃(2) ) is continuous and
hence its restriction to the compact set (λ(y) + ǫB1 ) × C × S (2) is uniformly continuous where
C ⊆ Rd2 is a compact set such that for every n ≥ 1, yn ∈ C. By the above we can conclude that
for
> 0, there exists N1 such that
n ≥ N1 , for every (x, s(2) ) ∈ (λ(y) + ǫB1 ) × S (2) ,
R for every
R any ǫ̃ (2)
(2)
(2)
(2)
(2)
(2)
f (s̃ )Π (x, yn , s )(ds̃ ) − S (2) f (s̃ )Π (x, y, s(2) )(ds̃(2) ) < ǫ̃. Therefore for every f ∈
S (2)
C(S (2) , R), there exists Ñ := max {N, N1 } such that for every n ≥ Ñ ,
Z
f (s̃(2) )ν n (ds̃(2) ) −
S (2)
Z
f (s̃(2) )ν(ds̃(2) )
S (2)
≤
ǫ̃ +
Z
Rd1 ×S (2)
Z
S (2)
Z
f (s̃(2) )Π(2) (x, y, s(2) )(ds̃(2) ) µn (ds, ds(2) ) −
f (s̃(2) )ν(ds̃(2) ) .
S (2)
The second term in the R.H.S. of the above inequality goes to zero as n → ∞ (use the definition of
ν(ds̃(2) ), assumption (A4) and [17, Thm. 2.1.1(ii)]). Therefore
taking limit in the
R above equation we
R
get that for any f ∈ C(S (2) , R), for every ǫ̃ > 0, limn→∞ S (2) f (s̃(2) )ν n (ds̃(2) ) − S (2) f (s̃(2) )ν(ds̃(2) ) ≤
ǫ̃. Hence ν n → ν in P(S (2) ) as n → ∞. Clearly µnS (2) → µS (2) as n → ∞. Therefore ν = µS (2) which
gives us that µ ∈ D(y).
(iii) Follows from part (i) of this lemma.
Define the set-valued map Ĥ2 : Rd2 → subsets of Rd1 such that for every y ∈ Rd2 ,
Z
H2,y (x, s(2) )µ(dx, ds(2) ),
Ĥ2 (y) := ∪µ∈D(y)
(18)
Rd1 ×S (2)
where for every y ∈ Rd2 , H2,y denotes the slice as in Definition 2.2(iv) of the set valued map H2 . Since for
every y ∈ Rd2 , for every µ ∈ D(y), supp(µ) is compact, by Lemma 2.8(iii) we know that the slices H2,y
are µ-integrable for every µ ∈ D(y). So the above set-valued map is well defined and we show later that
the slower timescale iterates track the DI given by,
dy
∈ Ĥ2 (y).
dt
The above DI is guaranteed to have solutions as a consequence of the lemma below.
Lemma 3.3 The set-valued map Ĥ2 : Rd2 → subsets of Rd2 is a Marchaud map.
(19)
Proof of the above lemma is given in section 4.
Remark In order to understand the DI (19) better, we consider the cases where the map λ(·) is singlevalued and the case where Markov noise terms are absent. These special cases also highlight the fact our
results are a significant generalization of the results in [9] and [8].
(1) When the map λ(·) is single-valued, for any µ ∈ D(y), since supp(µRd1 ) ⊆ λ(y), we have that
µRd1 = δλ(y) where δλ(y) ∈ P(Rd1 ) denotes the Dirac measure at λ(y). Therefore the measure µ = δλ(y) ⊗ µS (2) . Since µ ∈ D(y), we know that for every A ∈ B(S (2) ), µS (2) (A) =
R
R
R
(2)
(x, y, s(2) )(A)µ(dx, ds(2) ) = S (2) Rd1 Π(2) (x, y, s(2) )(A)δλ(y) (dx) µS (2) (ds(2) ) =
(2) Π
RRd1 ×S(2)
Π (λ(y), y, s(2) )(A)µS (2) (ds(2) ). Thus µS (2) ∈ D(2) (λ(y), y), where D(2) (λ(y), y) denotes the
S (2)
set of stationary measures of the Markov chain with transition kernel Π(2) (λ(y), y, ·)(·). Therefore
for every y ∈ Rd2 ,
Z
Z
(2)
(2)
H2,(λ(y),y) (s(2) )ν(ds(2) ),
H2,y (x, s )µ(dx, ds ) = ∪ν∈D(2) (λ(y),y)
Ĥ2 (y) = ∪µ∈D(y)
S (2)
Rd1 ×S (2)
where H2,(λ(y),y) denotes the slice as in Definition 2.2(i) of the set-valued map H2 . Therefore DI
(19) is nothing but the set-valued analogue of the slower timescale DI in [8].
18
(2) Suppose Markov noise terms are absent (for the analysis and definition of such a recursion see [9]).
Then such a recursion can be rewritten in the form of recursion (14), with Markov noise terms
taking values in a dummy state space S (1) = S (2) = {s∗ } with transition laws Π(1) (x, y, s∗ ) =
Π(2) (x, y, s∗ ) = δs∗ for every (x, y) ∈ Rd . Then it is easy to deduce that the stationary distribution
maps D(1) (x, y) = D(2) (x, y) = δs∗ for every (x, y) ∈ Rd . Then for every y ∈ Rd2 , any µ ∈ D(y) is of
the form µ = ν ⊗ δs∗ where ν ∈ P(Rd1 ) with supp(ν) ⊆ λ(y). Then for any y ∈ Rd2 ,
Z
Z
(2)
(2)
H2,y (x, s∗ )ν(dx)
H2,y (x, s )µ(dx, ds ) = ∪ν∈P(Rd1 )supp(ν)⊆λ(y)
Ĥ2 (y) = ∪µ∈D(y)
Rd 1
Rd1 ×S (2)
= co
¯ ∪x∈λ(y) H2 (x, y, s∗ ) ,
which is exactly the same slower timescale DI as in [9].
Suppose now that the following holds in addition:
(A10) DI (19) has a globally attracting set Y ⊆ Rd2 ,
then the main result of this paper states that for almost every ω, as n → ∞,
Xn (ω)
→ ∪y∈Y (λ(y) × {y}) .
Yn (ω)
4
Mean fields and their properties
In this section we prove that for every y ∈ Rd2 , the set-valued map Ĥ1 (·, y) and the set-valued map Ĥ2 (·)
defined in equations (15) and (18) respectively, are Marchaud maps.
Recall that by assumptions (A1) and (A2), the set-valued maps H1 and H2 are SAMs. For such setvalued
o2.2, we know that there exist sequences of continuous set-valued maps, denoted
o by Lemma
n
n maps,
(l)
(l)
which approximate H1 and H2 respectively. Further by Lemma 2.3, these
and H2
by H1
l≥1
l≥1
(l)
(l)
appromating
denoted by, h1 and h2 . Throughout this section
o
oa continuous
n parametrization
o admit
n
o
n maps
n
(l)
(l)
(l)
(l)
denote the maps as described above.
and h2
, H2 , h 1
H1
l≥1
l≥1
l≥1
Similar to the definition of the maps Ĥ1 and Ĥ2 , we define the maps obtained by averaging the set(l)
(l)
valued maps H1 and H2 for every l ≥ 1 with respect to measures given by the maps (x, y) → D(1) (x, y)
and y → D(y).
Definition Let the maps D(1) : Rd → subsets of P(S (1) ) and D : Rd2 → subsets of P(Rd1 × S (2) )
be as in section 3. For every l ≥ 1,
(l)
(i) for every (x, y) ∈ Rd , define Ĥ1 : Rd → subsets of Rd1 such that,
Z
(l)
(l)
H1,(x,y) (s(1) )µ(ds(1) ),
Ĥ1 (x, y) := ∪µ∈D(x,y)
S (1)
(l)
(l)
where H1,(x,y) denotes the slice (as in Defn.2.2(ii)) of the set-valued map H1 ,
(l)
(ii) for every y ∈ Rd2 , define Ĥ2 : Rd2 → subsets of Rd2 such that,
Z
(l)
(l)
H2,y (x, s(2) )µ(dx, ds(2) ),
Ĥ2 (y) := ∪µ∈D(y)
Rd1 ×S (2)
(l)
(l)
where H2,y denotes the slice (as in Defn.2.2(v)) of the set-valued map H2 .
(l)
(l)
In the lemma below we prove that for every y ∈ Rd2 , the maps Ĥ1 (·, y) and the map Ĥ2 (·) are
Marchaud maps for every l ≥ 1.
Lemma 4.1 For every l ≥ 1,
(l)
(i) the set-valued map Ĥ1 : Rd → subsets of Rd1
19
is such that,
(l)
(a) for every (x, y) ∈ Rd , Ĥ1 (x, y) is a non-empty, convex and compact subset of Rd1 ,
(b) for K (l) > 0 where K (l) is as in Lemma 2.2, for every (x, y) ∈ Rd , supx′ ∈Ĥ (l) (x,y) kx′ k ≤
1
K (l) (1 + kxk + kyk),
(c) for every (x, y) ∈ Rd , for Rd -valued sequence, {(xn , yn )}n≥1 converging to (x, y) ∈ Rd , for every
n
o
(l)
(l)
sequence x′n ∈ Ĥ1 (xn , yn ) converging to x′ ∈ Rd1 , we have that x′ ∈ Ĥ1 (x, y).
(l)
(ii) for every y ∈ Rd2 , the map Ĥ1 (·, y) is a Marchaud map,
(l)
(iii) the map Ĥ2 (·) is a Marchaud map.
Proof Fix l ≥ 1.
(l)
(i) For every (x, y) ∈ Rd , by Lemma 2.7(iv), H1,(x,y) is µ-integrable for every µ ∈ D(1) (x, y). Hence
(l)
(l)
for every (x, y) ∈ Rd , Ĥ1 (x, y) is non-empty. Let x1 , x2 ∈ Ĥ1 (x, y) and α ∈ (0, 1). Then by
Lemma 2.9(i), there exist ν 1 , ν 2 ∈ P(S (1) × U ), such that for i ∈ {1, 2}, νSi (1) ∈ D(1) (x, y) and
R
(l)
xi = S (1) ×U h1,(x,y)(s(1) , u)ν i (ds(1) , du) where U denotes the closed unit ball in Rd1 . Then αx1 +
R
(l)
(1 − α)x2 = S (1) ×U h1,(x,y) (s(1) , u)(αν 1 + (1 − α)ν 2 )(ds(1) , du). Clearly (αν 1 + (1 − α)ν 2 )S (1) =
ανS1 (1) + (1 − α)νS2 (1) ∈ D(1) (x, y) where the last inclusion follows from the fact that D(1) (x, y) is a
(l)
convex subset of P(S (1) ). By Lemma 2.9(i), we get that αx1 + (1 − α)x2 ∈ Ĥ1 (x, y). Therefore
(l)
Ĥ1 (x, y) is convex.
(l)
(l)
By Lemma 2.7(ii), for every (x, y) ∈ Rd , the set-valued map H1,(x,y) is bounded by C(x,y) :=
(l)
K (l) (1 + kxk + kyk). Therefore for every f ∈ S (H1,(x,y) ), for every s(1) ∈ S (1) , f (s(1) ) ≤ C(x,y) .
R
(l)
(l)
Thus for every x′ ∈ Ĥ1 (x, y), by definition, x′ = S (1) f (s(1) )µ(ds(1) ) for some f ∈ S (H1 ) and
R
(l)
(l)
some µ ∈ D(1) (x, y). Therefore for every x′ ∈ Ĥ1 (x, y), kx′ k ≤ S (1) f (s(1) ) µ(ds(1) ) ≤ C(x,y) =
K (l) (1 + kxk + kyk).
As a consequence of the arguments in the preceding paragraph for some (x, y) ∈ Rd in order
(l)
to show that Ĥo1 (x, y) is compact, it is enough to show that it is closed. Consider a sequence
n
xn ∈ Ĥ1l (x, y)
(l)
n≥1
converging to x∗ ∈ Rd1 . Then by definition of Ĥ1 (x, y) and by Lemma
2.9(i), for every n ≥ 1, there exists ν n ∈ P(S 1 × U ), such that νSn(1) ∈ D(1) (x, y) and xn =
R
(l)
(1)
, u)ν n (ds(1) , du). Since S (1) × U is a compact metric space, P(S (1) × U ) is comS (1) ×U h1,(x,y) (s
pact and hence there exists a subsequence {nk }k≥1 such that {ν nk }k≥1 converges to ν ∈ P(S (1) × U ).
k
R
(l)
Clearly νSn(1)
converges to νS (1) and by [17, Thm. 2.1.1(ii)], xnk = S (1) ×U h1,(x,y) (s(1) , u)ν nk (ds(1) , du) →
k≥1
R
(l)
(1)
k
, u)ν(ds(1) , du) = x∗ . Since for every k, νSn(1)
∈ D(1) (x, y) and by the fact that
S (1) ×U h1,(x,y) (s
R
(l)
D(1) (x, y) is closed we get that, νS (1) ∈ D(1) (x, y). Therefore x∗ = S (1) ×U h1,(x,y) (s(1) , du)ν(ds(1) , du)
(l)
(l)
and νS (1) ∈ D(1) (x, y). Thus x∗ ∈ Ĥ1 (x, y) which gives us that Ĥ1 (x, y) is closed.
o
n
(l)
be a sequence convergLet {(xn , yn )}n≥1 be a sequence converging to (x, y) and let x′n ∈ Ĥ1
n≥1
ing to x′ . Then by Lemma 2.9(i), for every n ≥ 1, there exists ν n ∈ P(S (1) × U ) such that νSn(1) ∈
R
(l)
D(1) (xn , yn ) and x′n = S (1) ×U h1,(x,y) (s(1) , u)ν n (ds(1) , du). Since S (1) × U is a compact metric space,
P(S (1) × U ) is a compact metric space and hence there exists a subsequence say {nk }k≥1 such that
k
{ν nk }k≥1 converges to ν ∈ P(S (1) × U ). Clearly νSn(1)
→ νS (1) in P(S (1) ) and by closed graph property of the map (x, y) → D(1) (x, y), we have that νS (1) ∈ D(1) (x, y). Using the continuity of the map
(l)
(l)
(l)
h1 (·) it is easy to show that limk→∞ sup(s(1) ,u)∈S (1) ×U h1,(xn ,yn ) (s(1) , u) − h1,(x,y) (s(1) , u) = 0.
k
k
R
R
(l)
(l)
Then, kx′ − S (1) ×U h1,(x,y) (s(1) , u)ν(ds(1) , du)k ≤ kx′ − S (1) ×U h1,(xn ,yn ) (s(1) , u)ν nk (ds(1) , du)k +
k
k
R
R
(l)
(l)
(l)
(1)
nk
(1)
(1)
(s(1) , u)ν nk (ds(1) , du)−
h
(s
,
u)kν
(ds
,
du)+k
(s
,
u)−h
kh
1,(x,y)
1,(xnk ,ynk )
S (1) ×U 1,(x,y)
S (1) ×U
R
(l)
(s(1) , u)ν(ds(1) , du)k. Now use [17, Thm. 2.1.1(ii)] in the above inequality to obh
S (∞) ×U 1,(x,y)
R
(l)
tain limk→∞ kx′ − S (1) ×U h1,(x,y) (s(1) , u)ν(ds(1) , du)k = 0. Then Lemma 2.9(i) gives us that x′ ∈
(l)
Ĥ1 (x, y).
20
(ii) Follows from part (i) of this lemma.
(iii) Proof is similar to part (i) of this lemma with minor modifications. First modification is the use
(l)
of Lemma 2.9(ii) instead of Lemma 2.9(i). For example in order to show that Ĥ2 (y) is closed
(l)
for some y ∈ Rd2 , fix sequence {yn′ }n≥1 ⊆ Ĥ2 (y) converging to y ′ . Use Lemma 2.9(ii) and the
(l)
definition of Ĥ2 (y), to obtain {ν n }n≥1 ⊆ P(Rd1 × S (2) × U ) where U denotes the closed unit
ball in Rd2 and the sequence {ν n }n≥1 is such that for every n ≥ 1, νRnd1 ×S (2) ∈ D(y) and yn′ =
R
(l)
(2)
, u)ν n (dx, ds(2) , du). By definition of D(y), for every n ≥ 1, supp(νRnd1 ×S (2) ) ⊆
Rd1 ×S (2) ×U h2,y (x, s
λ(y) × S (2) and hence supp(ν n ) ⊆ λ(y) × S (2) × U which is a compact subset of Rd1 × S (2) × U . Now
by Prohorov’s theorem the sequence {ν n }n≥1 is a relatively compact subset of P(Rd1 × S (2) × U )
and hence
subsequence. By Lemma 3.2(i), D(y) is compact and hence every limit
o
n has a convergent
n
point of νRd1 ×S (2) is in D(y). The rest of the argument is same as the corresponding in part (i)
of this lemma.
(l)
In order to show that Ĥ2 (·) has a closed graph, fix sequences {yn }n≥1 converging to y and
n
o
(l)
yn′ ∈ Ĥ2 (yn )
converging to y ′ . Use Lemma 2.9(ii), to obtain {ν n }n≥1 ⊆ P(Rd1 × S (2) × U )
n≥1
R
(l)
such that for every n ≥ 1, νRnd1 ×S (2) ∈ D(yn ) and yn′ := Rd1 ×S (2) ×U h2,y (x, s(2) , u)ν n (dx, ds(2) , du).
Then for every n ≥ 1, supp(ν n ) ⊆ λ(yn ) × S (2) × U . By assumption (A9), for any δ > 0, the
set L := {x ∈ λ(ỹ) : kỹ − yk ≤ δ} is a compact subset of Rd1 . Therefore there exists N large such
that for every n ≥ N , supp(ν n ) ⊆ L × S (2) × U . By Prohorov’s theorem the sequence of measures
{ν n }
is tight and has a convergent subsequence. Clearly by Lemma 3.2(ii), every limit point of
n n≥N o
n
νRd1 ×S (2)
is in D(y). Now the rest of the argument is same as the corresponding in part (i)
n≥N
of this lemma.
(l+1)
(l)
By Lemma 2.2 we know that for every l ≥ 1, for every (x, y) ∈ Rd , H1 (x, y) ⊆ H1
(x, y) ⊆ H1 (x, y)
(l+1)
(l)
(similarly H2 (x, y) ⊆ H2
(x, y) ⊆ H2 (x, y)). The next lemma states that the above is true for Ĥi and
(l)
Ĥi as well, for every i ∈ {1, 2}.
Lemma 4.2 :
(l+1)
(i) For every l ≥ 1, for every (x, y) ∈ Rd , Ĥ1 (x, y) ⊆ Ĥ1
(l+1)
(l)
(x, y) ⊆ Ĥ1 (x, y).
(l)
(ii) For every l ≥ 1, for y ∈ Rd2 , Ĥ2 (y) ⊆ Ĥ2
(y) ⊆ Ĥ2 (y).
R
(l)
(l)
(iii) For every (x, y) ∈ Rd , ∩l≥1 Ĥ1 (x, y) = ∪µ∈D(1) (x,y) ∩l≥1 S (1) H1,(x,y) (s(1) )µ(ds(1) ).
(l)
(iv) For every y ∈ Rd2 , ∩l≥1 Ĥ2 (y) = ∪µ∈D(y) ∩l≥1
(l)
R
(l)
Rd2 ×S (2)
H2,y (x, s(2) )µ(dx, ds(2) ).
(v) For every (x, y) ∈ Rd , Ĥ1 (x, y) = ∩l≥1 Ĥ1 (x, y).
(l)
(vi) For every y ∈ Rd2 , Ĥ2 (y) = ∩l≥1 Ĥ2 (y).
(l)
Proof The proofs of parts (i) and (ii) follow directly from the definition of Ĥi , Ĥi for every l ≥ 1 and
(l)
(l+1)
(x, y) ⊆ Hi (x, y) for every (x, y) ∈ Rd . The proof of
the fact that for every i ∈ {1, 2}, Hi (x, y) ⊆ Hi
part (iii) is similar to part (iv) and we present the proof of part (iv) below (the proof of part (iii) is in
fact the same as that of [11, Lemma 4.4(ii)]).
(l)
(iv) Fix y ∈ Rd2 . Then by definition of Ĥ2 (y), we have that for every l ≥ 1, for any µ ∈ D(y),
R
R
(l)
(l)
(l)
H2,y (x, s(2) )µ(dx, ds(2) ) ⊆ Ĥ2 (y). Therefore, ∪µ∈D(y) ∩l≥1 Rd1 ×S (2) H2,y (x, s(2) )µ(dx, ds(2) ) ⊆
Rd1 ×S (2)
(l)
∩l≥1 Ĥ2 (y).
(l)
(l)
Let y ′ ∈ ∩l≥1 Ĥ2 (y). Then for every l ≥ 1, there exists µl ∈ D(y) such that y ′ ∈ Rd1 ×S (2) H2,y
(x, s(2) )µl (dx, ds(2) ).
l
l
(2)
Since µ l≥1 is a subset of D(y), for every l ≥ 1, supp(µ ) ⊆ λ(y) × S . Hence the se
quence of probability measures µl l≥1 is tight and by Prohorov’s theorem has a limit say µ∗ ∈
R
P(Rd1 × S (2) ). Let {lk }k≥1 be a subsequence such that µlk → µ∗ as k → ∞ and by Lemma
3.2(i) we know that D(y) is compact which gives us µ∗ ∈ D(y). Since for every l ≥ 1, for ev(l )
(l)
ery k such that lk ≥ l, S (H2,yk ) ⊆ S (H2,y ) we get that for every l ≥ 1, for every k such that
21
R
(l)
lk ≥ l, Rd1 ×S (2) H2,y (x, s(2) )µlk (dx, ds(2) ) = y ′ . For every l ≥ 1, by Lemma 2.9(ii), we know
that for every k such that lk ≥ l, there exists ν (l,lk ) ∈ P(Rd1 × S (2) × U ) (U denotes the closed
R
(l)
(l,l )
unit ball in Rd2 ) such that y ′ = Rd1 ×S (2) ×U h2,y (x, s(2) , u)ν (l,lk ) (dx, ds(2) , du) and νRd1k×S (2) = µlk .
Further,
for every l ≥ 1, for every k such that lk ≥ l, supp(ν (l,lk ) ) ⊆ λ(y) × S (2) × U and
(l,l
hence ν k ) k:l ≥l is tight and by Prohorov’s theorem has a convergent subsequence. For evk
ery l ≥ 1, let ν (l) denote a limit point of the sequence ν (l,lk ) k:l ≥l . Since for every l ≥ 1,
k
o
n
(l)
(l,l )
and µlk → µ∗ as k → ∞ we have that νRd1 ×S (2) = µ∗ ∈ D(y). By
νRd1k×S (2) = µlk
k:lk ≥l
R
(l)
[17, Thm. 2.1.1(ii)], for every l ≥ 1 y ′ = Rd1 ×S (2) ×U h2,y (x, s(2) , u)ν (l) (dx, ds(2) , du) and hence
R
(l)
by Lemma 2.9(ii), y ′ ∈ Rd1 ×S (2) H2,y (x, s(2) )µ∗ (dx, ds(2) ) where µ∗ ∈ D(y). Therefore there exR
(l)
ists µ∗ ∈ D(y) such that for every l ≥ 1, y ′ ∈ Rd1 ×S (2) H2,y (x, s(2) )µ∗ (dx, ds(2) ). Hence y ′ ∈
R
(l)
∪µ∈D(y) ∩l≥1 Rd1 ×S (2) H2,y (x, s(2) )µ(dx, ds(2) ).
The proof of part (v) is similar to the proof of part (vi) and we present a proof of part (vi) below (the
proof of part (v) is exactly the same as that of [11, Lemma 4.4(iii)])
(l)
(vi) From part (ii) of this lemma we have that, for every y ∈ Rd2 , Ĥ2 (y) ⊆ ∩l≥1 Ĥ2 (y).
R
(l)
Fix y ∈ Rd2 and µ ∈ D(y). Let y ′ ∈ ∩l≥1 Rd1 ×S (2) H2,y (x, s(2) )µ(dx, ds(2) ). Then for every l ≥
R
(l)
1, there exists f (l) ∈ S (H2,y ) such that y ′ = Rd1 ×S (2) f (l) (x, s(2) )µ(dx, ds(2) ). Let d(ỹ, A) :=
d2
d2
inf {kỹ
R − zk : z ∈ A} for(2)every ỹ ∈(2)R and for every A ⊆ R compact. By Lemma 2.6, we have
that Rd1 ×S (2) H2,y (x, s )µ(dx, ds ) is compact and convex. Then,
′
d(y ,
Z
H2,y (x, s
Rd1 ×S (2)
(2)
)µ(dx, ds
(2)
Z
′
(2)
(2)
)
H2,y (x, s )µ(dx, ds
)) = inf ky − zk : z ∈
Rd1 ×S (2)
Z
=
inf
f (x, s(2) )µ(dx, ds(2) )
y′ −
f ∈S (H2,y )
Rd1 ×S (2)
Z
=
inf
f (l) (x, s(2) ) − f (x, s(2) ) µ(dx, ds(2) )
f ∈S (H2,y )
d1
(2)
Z R ×S
≤
inf
f (l) (x, s(2) ) − f (x, s(2) ) µ(dx, ds(2) )
f ∈S (H2,y ) Rd1 ×S (2)
Z
o
n
inf f (l) (x, s(2) ) − ỹ : ỹ ∈ H2,y (x, s(2) ) µ(dx, ds(2) ),
=
Rd1 ×S (2)
where the last equality follows from [15, Lemma 1.3.12]. By [11, Lemma 3.7], we know that for every
l ≥ 1, the map (x, s(2) ) → d(f (l) (x, s(2) ), H2,y (x, s(2) )) is measurable and from the last equality it
follows that for every l ≥ 1,
Z
Z
d(f (l) (x, s(2) ), H2,y (x, s(2) ))µ(dx, ds(2) ).
H2,y (x, s(2) )µ(dx, ds(2) )) ≤
d(y ′ ,
Rd1 ×S (2)
Rd1 ×S (2)
By observation (2) stated after Lemma 2.2 we have that for every (x, s(2) ) ∈ Rd2 × S (2) ,
liml→∞ d(f (l) (x, s(2) ), H2,y (x, s(2) )) = 0. Since µ ∈ D(y), supp(µ) ⊆ λ(y) × S (2) , for every l ≥ 1,
Z
Z
d(f (l) (x, s(2) ), H2,y (x, s(2) ))µ(dx, ds(2) ) =
d(f (l) (x, s(2) ), H2,y (x, s(2) )µ(dx, ds(2) )).
Rd1 ×S (2)
λ(y)×S (2)
Since λ(y) is compact, there exists M > 0 such that for every x ∈ λ(y), kxk ≤ M . By Lemma 2.8(ii),
(A2) and observation (1) stated below Lemma 2.2, we have that for every l ≥ 1, for every (x, s(2) ) ∈
(l)
λ(y) × S (2) , d(f (l) (x, s(2) ), H2,y (x, s(2) )) ≤ (Ky + K)(1 + kxk) ≤ (max{K̃, K̃kyk} + K)(1 + M ). By
bounded convergence theorem we have,
Z
Z
(2)
(2)
′
H2,y (x, s )µ(dx, ds )) ≤ lim
d(y ,
d(f (l) (x, s(2) ), H2,y (x, s(2) ))µ(dx, ds(2) ) = 0.
l→∞
Rd1 ×S (2)
Rd1 ×S (2)
Therefore d(y ′ , Rd1 ×S (2) H2,y (x, s(2) )µ(dx, ds(2) )) = 0 and by Lemma 2.6, we know that
R
R
H2,y (x, s(2) )µ(dx, ds(2) ) is a closed subset of Rd2 . Hence y ′ ∈ Rd1 ×S (2) H2,y (x, s(2) )µ(dx, ds(2) ).
Rd1 ×S (2)
R
22
From the arguments in the preceding paragraph, we have that for every y ∈ Rd2 , for every µ ∈ D(y),
R
R
(l)
∩l≥1 Rd1 ×S (2) H2,y (x, s(2) )µ(dx, ds(2) ) ⊆ Rd1 ×S (2) H2,y (x, s(2) )µ(dx, ds(2) ). Thus for every y ∈ Rd2 ,
∪µ∈D(y) ∩l≥1
Z
Rd1 ×S (2)
(l)
H2,y (x, s(2) )µ(dx, ds(2) )
⊆ ∪µ∈D(y)
Z
H2,y (x, s(2) )µ(dx, ds(2) ) = Ĥ2 (y).
Rd1 ×S (2)
(l)
By part (iv) of this lemma we get that for every y ∈ Rd2 , ∩l≥1 Ĥ2 (y) ⊆ Ĥ2 (y).
Lemma 4.3 The set-valued map Ĥ1 : Rd → subsets of Rd1
as defined in eqn.(15) is such that,
(i) for every (x, y) ∈ Rd , Ĥ1 (x, y) is a non-empty, convex and compact subset of Rd1 ,
(ii) there exists K > 0 (same as in (A1)(ii)), such that for every (x, y) ∈ Rd , supx′ ∈Ĥ1 (x,y) kx′ k ≤
K(1 + kxk + kyk),
(iii) for every (x, y) ∈ Rd , for every Rd -valued sequence, {(xn , yn )}n≥1 converging to (x, y) and for every
{x′n ∈ Ĥ1 (xn , yn )}n≥1 converging to x′ , we have x′ ∈ Ĥ1 (x, y).
Proof
(i) Fix (x, y) ∈ Rd . By Lemma 2.7(iii) H1,(x,y) is µ-integrable for every µ ∈ D(1) (x, y). Hence
(l)
Ĥ1 (x, y) is non-empty. For every l ≥ 1, by Lemma 4.1(i)(a) we know that Ĥ1 (x, y) is convex and
(l)
compact subset of Rd1 . By Lemma 4.2(v) we have that Ĥ1 (x, y) = ∩l≥1 Ĥ1 (x, y) and hence Ĥ1 (x, y)
is convex and compact.
(ii) Fix (x, y) ∈ Rd . For any x′ ∈ Ĥ1 (x, y), there exists µ ∈ D(1) (x, y) and f ∈ S (H1,(x,y) ) such that
R
x′ = S (1) f (s(1) )µ(ds(1) ). By Lemma 2.7(i), we know that for every s(1) ∈ S (1) , kf (s(1) )k ≤ C(x,y) =
R
K(1 + kxk + kyk). Therefore kx′ k = k S (1) f (s(1) )µ(ds(1) )k ≤ C(x,y) = K(1 + kxk + kyk).
(iii) Let {(xn , yn )}n≥1 be a sequence converging to (x, y) and {x′n ∈ Ĥ1 (xn , yn )}n≥1 be a sequence
converging to x′ . Then by Lemma 4.2(v), we have that for every l ≥ 1, for every n ≥ 1, x′n ∈
(l)
(l)
Ĥ1 (xn , yn ). By Lemma 4.1(i)(c), we have that for every l ≥ 1, x′ ∈ Ĥ1 (x, y). Thus by Lemma
′
4.2(v), we have x ∈ Ĥ1 (x, y).
Lemma 3.1 is an immediate consequence of the above lemma. Similarly, the proof of Lemma 3.3 follows
(l)
from the fact that {Ĥ2 }l≥1 are Marchaud maps (see Lemma 4.1(iii)) which approximate Ĥ2 (see Lemma
4.2(vi)) and the linear growth property of the map λ(·) (that is (A9(i))).
5
Recursion analysis
In this section we present the analysis of recursion (14). The analysis comprises of two parts.
The first part deals with the analysis of the faster timescale recursion where we show that the faster
timescale iterates {Xn }n≥1 converge almost surely to λ(y) (as in (A9)) for some y ∈ Rd2 .
The second part deals with the slower timescale recursion analysis where we show that the slower
timescale iterates {Yn }n≥1 track the flow of DI (19).
Throughout this section we assume that assumptions (A1) − (A9) are satisfied.
5.1
Faster timescale recursion analysis
For every ω ∈ Ω, for every n ≥ 0, the two timescale recursion (14) can be written as,
(2)
Yn+1 (ω) − Yn (ω) − b(n)Mn+1 (ω) = b(n)Vn2 (ω),
Xn+1 (ω) − Xn (ω) −
(1)
a(n)Mn+1 (ω)
= a(n)Vn1 (ω),
where for every n ≥ 0, Vn1 and Vn2 are such that, for every ω ∈ Ω,
Vn1 (ω) ∈ H1 (Xn (ω), Yn (ω), Sn(1) (ω)),
Vn2 (ω) ∈ H2 (Xn (ω), Yn (ω), Sn(2) (ω)).
23
(20a)
(20b)
The recursion (20) can be rewritten as,
b(n) 2
b(n) (2)
Vn (ω) +
Mn+1 (ω) ,
a(n)
a(n)
(1)
1
Xn+1 (ω) − Xn (ω) = a(n) Vn (ω) + Mn+1 (ω) ,
Yn+1 (ω) − Yn (ω) = a(n)
for every ω ∈ Ω and for every n ≥ 0. The above can be now written in the form of the single timescale
recursion (that is (12)):
Zn+1 − Zn − a(n)Mn+1 ∈ a(n)F (Zn , Sn(1) ),
(22)
where,
(1) for every n ≥ 0, Zn = (Xn , Yn ),
(2)
(1) b(n)
Vn2 + Mn+1 ,
(2) for n ≥ 0, Mn+1 = Mn+1 , a(n)
(3) F : Rd × S (1) →
(H1 (x, y, s(1) ), 0).
subsets of Rd
such that for every (x, y, s(1) ) ∈ Rd × S (1) , F (x, y, s(1) ) =
We now show that the quantities defined above satisfy the assumptions associated with the single
timescale recursion as in section 2.4. Clearly by assumption (A5), the step size sequence {a(n)}n≥0 satisfies
n
o
(1)
assumption S(A3) and by assumption (A3) the Markov noise terms, Sn
satisfy assumption S(A2).
n≥0
As a consequence of the stability assumption (A8), we have that P(supn≥0 kZn := (Xn , Yn )k < ∞) = 1
and hence assumption S(A5) is satisfied.
Consider the set-valued map F defined above. Clearly by assumption (A1)(i), for every (x, y, s(1) ) ∈
d
R × S (1) , F (x, y, s(1) ) is a non-empty, convex and compact subset of Rd . Further by assumption (A1)(ii),
we have that for every (x, y, s(1) ) ∈ Rd × S (1) , sup(x′ ,y′ )∈F (x,y,s(1) ) k(x′ , y ′ )k = supx′ ∈H1 (x,y,s(1) ) kx′ k ≤
K(1 + kxk + kyk) ≤ max {K, KC} (1 + k(x, y)k) where C > 0 is such that kxk + kyk ≤ Ck(x, y)k for every
(x, y) ∈ Rd (see [20, Thm. 4.3.26]). By assumption (A1)(iii), the map H1 has a closed graph and hence
the map F also has a closed graph. Therefore the set-valued mapn satisfies assumption S(A1).
o
Pm−1
Recall that for every T > 0, for every n ≥ 0, τ 1 (n, T ) := min m > n : k=n a(k) ≥ T . Let,
Ω1 := {ω ∈ Ω : (A6), (A7) and (A8) hold} .
(23)
It is clear that P(Ω1 ) = 1. Let ω ∈ Ω1 and fix T > 0. For any n ≥ 0,
sup
n≤k≤τ 1 (n,T )
k
X
a(m)Mm+1 (ω) ≤
m=n
sup
n≤k≤τ 1 (n,T )
k
X
(1)
a(m)Mm+1 (ω) +
m=n
sup
n≤k≤τ 1 (n,T )
k
X
a(m)
m=n
b(m) 2
(2)
Vn (ω) + Mm+1 (ω) .
a(m)
By assumption (A8), for every ω ∈ Ω1 , there exists r > 0, such that supn≥0 (kXn (ω)k + kYn (ω)k) ≤ r.
(2)
Since for every n ≥ 0, for every ω ∈ Ω1 , Vn2 (ω) ∈ H2 (Xn (ω), Yn (ω), Sn (ω)), by assumption (A2)(ii) we
have that, kVn2 (ω)k ≤ K(1 + kXn(ω)k + kYn (ω)k) ≤ K(1 + r) =: R < ∞. Further by assumption (A5)(iii)
ǫ
for every 0 < ǫ < T , there exists N such that for every n ≥ N , b(n) ≤ T +1
a(n). Therefore for every
Pτ 1 (n,T )−1
P
Pm−1
m−1
ǫ
b(k) ≤ ǫ and
n ≥ N , for every m > n, k=n b(k) ≤ T +1 k=n a(k). Thus for every n ≥ N , k=n
τ 1 (n, T ) ≤ τ 2 (n, T ). Therefore for every 0 < ǫ < T , for every n ≥ N ,
sup
n≤k≤τ 1 (n,T )
k
X
a(m)Mm+1 (ω) ≤
sup
n≤k≤τ 1 (n,T )
m=n
k
X
(1)
a(m)Mm+1 (ω) + ǫR +
m=n
sup
n≤k≤τ 2 (n,T )
k
X
(2)
b(m)Mm+1 (ω) .
m=n
Taking limit in the above equation and using assumptions (A6) and (A7) gives us that, for every 0 < ǫ < T ,
lim
sup
n→∞ n≤k≤τ 1 (n,T )
k
X
a(m)Mm+1 (ω) ≤ Rǫ.
m=n
Pk
Therefore for every ω ∈ Ω1 , for every T > 0, limn→∞ supn≤k≤τ 1 (n,T )
m=n a(m)Mm+1 (ω) = 0. Thus
the additive noise terms {Mn }n≥1 satisfy assumption S(A4).
Therefore quantities in recursion (22) satisfy assumptions S(A1) − S(A5) and we apply the main result
of the single timescale recursion (see Theorem 2.10(iii)) to conclude that,
24
Lemma 5.1 Under assumptions (A1) − (A8), for almost every ω, there exists a non-empty compact set
L ⊆ Rd (depending on ω) such that,
(i) (Xn (ω), Yn (ω)) → L as n → ∞, where {(Xn , Yn )}n≥0 is as in recursion (14).
(ii) the set L is internally chain transitive for the flow of the DI,
dx
Ĥ1 (x, y)
dt
∈
.
dy
0
dt
(24)
The set-valued map associated with the DI (24) is clearly a Marchaud map (use Lemma 4.3). Further
any solution (x(·), y(·)) of DI (24) is such that for every t ∈ R, y(t) = y(0) and x(·) is a solution to DI
(16) with y0 = y(0).
Fix ω ∈ Ω such that Lemma 5.1 holds. Let L ⊆ Rd be as in Lemma 5.1. Let
A := (x, y) ∈ Rd : x ∈ λ(y), y ∈ Rd2 ,
(25)
where for every y ∈ Rd2 , λ(y) is as in assumption (A9). Since L is internally chain transitive for the flow of
DI (24), by [12, Lemma 3.5] we know that it is invariant. Let (x∗ , y ∗ ) ∈ L and (x(·), y(·)) be a solution to DI
(24) with initial condition (x∗ , y ∗ ) and for all t ∈ R, (x(t), y(t)) ∈ L. Then for every t ∈ R, y(t) = y ∗ and
x(·) is a solution of DI (16) with y0 = y ∗ . By assumption (A9), there exists a compact subset λ(y ∗ ) ⊆ Rd1 ,
which is a globally attracting set for the flow of DI (16) with y0 = y ∗ . By definition of a globally attracting
set we have that ∩t≥0 {x(q + t) : q ≥ 0} ⊆ λ(y ∗ ). Therefore (x(t), y(t)) → λ(y ∗ ) × {y ∗ } ⊆ A and since for
every t ∈ R, (x(t), y(t)) ∈ L we get that L ∩ A 6= ∅. In fact for any closed set C ⊆ Rd invariant for the
flow of DI (24) the above argument gives us that C ∩ A 6= ∅. If we are able to show that L ∩ A = L, then
by Lemma 5.1(i) we obtain that (Xn (ω), Yn (ω)) → L ⊆ A as n → ∞. In this regard we need to impose
the following assumption.
(A11) For any compact set C ⊆ Rd , invariant for the flow of DI (24), for any open neighborhood O of
C ∩ A, there exists an open neighborhood O′ of C ∩ A, such that,
ΦC (O′ ∩ C, [0, ∞)) ⊆ O ∩ C,
where ΦC : C × R → subsets of Rd
(see section 2.3 for definition).
denotes the flow of DI (24) restricted to the invariant set C
Remark The above assumption is a weaker form of assumption (A1) imposed in [10, Ch. 6] (that implies
assumption (A11) above). The above assumption is basically the Lyapunov stability condition (see [12,
Defn. IX(ii)]) for the flow restricted to the invariant set C. We shall see that in the application studied
later the above assumption is satisfied.
Lemma 5.2 Under assumptions (A1) − (A9) and (A11), for almost every ω, L ⊆ Rd as in Lemma 5.1,
is such that
L ⊆ (x, y) ∈ Rd : x ∈ λ(y), y ∈ Rd2 .
Therefore (Xn (ω), Yn (ω)) → L ⊆ (x, y) ∈ Rd : x ∈ λ(y), y ∈ Rd2 as n → ∞, where {(Xn , Yn )}n≥0 is as
in recursion (14).
Proof We present a brief outline here to highlight where assumption (A11) is used.
Let A ⊆ Rd be as in equation (25). Fix ω ∈ Ω and obtain L as in Lemma 5.1. We know that L is
internally chain transitive for the flow of DI (24) and since it is also invariant L ∩ A 6= ∅. By assumption
(A9), for every (x∗ , y ∗ ) ∈ L, ωΦL ((x∗ , y ∗ )) ⊆ L ∩ A. Thus for every (x∗ , y ∗ ) ∈ L, for every solution of DI
(24), (x(·), y(·)) such that (x(0), y(0)) = (x∗ , y ∗ ) and for every t ∈ R, (x(t), y(t)) ∈ L, for every open
neighborhood O of L ∩ A, there exists t > 0 such that (x(t), y(t)) ∈ O ∩ L. By [12, Lemma 3.13], we get
that for every open neighborhood O of L ∩ A, there exists T > 0, for every (x∗ , y ∗ ) ∈ L, for every solution
of DI (24), (x(·), y(·)) such that (x(0), y(0)) = (x∗ , y ∗ ) and for every t ∈ R, (x(t), y(t)) ∈ L, for some
t ∈ [0, T ], (x(t), y(t)) ∈ O ∩ L.
Fix ǫ > 0. Then by assumption (A11) there exists on open neighborhood of L ∩ A, O such that,
ΦL (O ∩ L, [0, ∞)) ⊆ N ǫ (L ∩ A) ∩ L. From arguments in the previous paragraph we can find T > 0 such
that for every (x∗ , y ∗ ) ∈ L, for every solution of DI (24) with (x(0), y(0)) = (x∗ , y ∗ ) and (x(t), y(t)) ∈ L for
every t ∈ R, there exists t ∈ [0, T ], such that (x(t), y(t)) ∈ O∩L. Therefore ΦL (L, [T, ∞)) ⊆ N ǫ (L∩A)∩L.
Thus L ∩ A is an attracting set for ΦL . Now the claim follows from [12, Prop. 3.20].
25
5.2
Slower timescale recursion analysis
Before we present the analysis of slower timescale recursion we present
o where we shall
o preliminaries
n
n some
(l)
(l)
denote maps
and h2
define various quantities needed later. Throughout this section let H2
l≥1
l≥1
as in section 4. Further we shall allow assumptions (A1) − (A9) and (A11) to be satisfied. The slower
timescale recursion analysis is similar to the analysis of single timescale inclusion in [11] with minor
modifications arising due to the presence of faster timescale iterates. Throughout this section U denotes
the closed unit ball in Rd2 and Br denotes the closed ball of radius r > 0 in Rd1 centered at the origin.
5.2.1
Preliminaries
Pn−1
Let ts (0) := 0 and for every n ≥ 1, ts (n) := m=0 b(m). Define Ȳ : Ω × [0, ∞) → Rd2 , such that for every
(ω, t) ∈ Ω × [0, ∞),
s
t (n + 1) − t
t − ts (n)
Y
(ω)
+
Yn (ω),
Ȳ (ω, t) := s
n+1
t (n + 1) − ts (n)
ts (n + 1) − ts (n)
where n is such that t ∈ [ts (n), ts (n + 1)).
Consider the slower timescale recursion (14a) given by,
(2)
Yn+1 − Yn − b(n)Mn+1 ∈ b(n)H2 (Xn , Yn , Sn(2) ),
(2)
for every n ≥ 0. By Lemma 2.2, we have that for every l ≥ 1, for every n ≥ 0, H2 (Xn , Yn , Sn ) ⊆
(2)
(l)
H2 (Xn , Yn , Sn ). Therefore, for every l ≥ 1, the following recursion follows from the above (that is,
(14a) ):
(l)
(2)
Yn+1 − Yn − b(n)Mn+1 ∈ b(n)H2 (Xn , Yn , Sn(2) ).
(l)
By lemma 2.3, we know that for every l ≥ 1 the set-valued map H2 admits a continuous single valued
(l)
parametrization, h2 . The next lemma allows us to write the slower timescale inclusion in terms of the
(l)
parametrization of H2 and result follows from [11, Lemma 6.1] .
(l)
Lemma 5.3 For every l ≥ 1, for every n ≥ 0, there exists a U -valued random variable on Ω, Un , such
that for every ω ∈ Ω,
(2)
(l)
Yn+1 (ω) − Yn (ω) − b(n)Mn+1 (ω) = b(n)h2 (Xn (ω), Yn (ω), Sn(2) (ω), Un(l) (ω)).
For every l ≥ 1, define Γ(l) : Ω × [0, ∞) → P(U × Rd1 × S (2) ), such that for every (ω, t) ∈ Ω × [0, ∞),
Γ(l) (ω, t) := δU (l) (ω) ⊗ δXn (ω) ⊗ δS (2) (ω) ,
n
(26)
n
(l)
where δU (l) (ω) ∈ P(U ) denotes the Dirac measure at Un (ω) ∈ U (for every A ∈ B(U ), δU (l) (ω) (A) = 1
n
n
(l)
if Un (ω) ∈ A, 0 otherwise), δXn (ω) ∈ P(Rd1 ) denotes the Dirac measure at Xn (ω) ∈ Rd1 and similarly
(2)
δS (2) (ω) ∈ P(S (2) ) denotes the Dirac measure at Sn (ω) ∈ S (2) .
n
The lemma below provides an equicontinuity result used later.
Lemma 5.4 For every l ≥ 1, for every r > 0, the family of maps
Z
(l)
(2)
(2)
(2)
y ∈ rU →
h2 (x, y, s , u)ν(du, dx, ds ) : ν ∈ P(U × Br × S )
U×Br ×S (2)
is equicontinuous.
(l)
(l)
Proof Fix l ≥ 1. By Lemma 2.3, we know that the map h2 (·) is continuous. Hence the map h2 (·)
restricted to the compact set Br × rU × S (2) × U is uniformly continuous. Therefore for every ǫ > 0, there
exists δ > 0, such that for every (x, y, s(2) , u), (x, y ′ , s(2) , u) ∈ Br × rU × S (2) × U , satisfying ky − y ′ k < δ,
(l)
(l)
h2 (x, y, s(2) , u) − h2 (x, y ′ , s(2) , u) < ǫ. Therefore for δ > 0 as above, with ky − y ′ k < δ, for any ν ∈
R
R
(l)
(l)
P(U ×Br ×S (2) ), U×Br ×S (2) h2 (x, y, s(2) , u)ν(du, dx, ds(2) ) − U×Br ×S (2) h2 (x, y ′ , s(2) , u)ν(du, dx, ds(2) ) ≤
R
(l)
(l)
h2 (x, y, s(2) , u) − h2 (x, y ′ , s(2) , u) ν(du, dx, ds(2) ) ≤ ǫ.
U×Br ×S (2)
26
For every l ≥ 1, define G(l) : Ω × [0, ∞) → Rd2 such that, for every (ω, t) ∈ Ω × [0, ∞),
(l)
G(l) (ω, t) := h2 (Xn (ω), Yn (ω), Sn(2) (ω), Un(l) (ω)),
(27)
where n is such that t ∈ [ts (n), ts (n + 1)).
In what follows most of the arguments are sample path wise and we use smaller case symbols to denote
(l)
(2)
(2)
(l)
the above defined quantities along a particular sample path. For example xn , yn , un , mn+1 , sn , ȳ(t), γn (t)
(l)
(2)
(2)
(l)
and g (l) (t) denote Xn (ω), Yn (ω), Un (ω), Mn+1 (ω), Sn (ω), Ȳ (ω, t), Γn (ω, t) and G(l) (ω, t) respectively
for some ω fixed.
5.2.2
Main result-Asymptotic pseudotrajectory
For every ω ∈ Ω, for every l ≥ 1, for every t̃ ≥ 0, let ỹ (l) (·; t̃) denote the solution of the o.d.e.
ỹ˙ (l) (t; t̃) = g (l) (t + t̃),
(28)
for every t ≥ 0 with initial condition ỹ (l) (0; t̃) = ȳ(t̃).
Let Ω1 be as in (23). Then by assumptions (A6)-(A8), we have that P(Ω1 ) = 1. First we shall get rid
of the additive noise terms. In this regard we prove the lemma below which states that for every ω ∈ Ω1 ,
the family of functions {ȳ(· + t)}t≥0 and ỹ (l) (·; t) t≥0 have the same limit points in C([0, ∞), Rd2 ) for
every l ≥ 1. Proof of the lemma below is similar to [11, Lemma 6.3] and is given in appendix A.
Lemma 5.5 For almost every ω, for every l ≥ 1, for every T > 0,
lim sup
t→∞ 0≤q≤T
ȳ(q + t) − ỹ (l) (q; t) = 0.
The lemma below guarantees the existence of limit points for
proof is similar to [11, Lemma 6.4] and is given in appendix B.
(l)
ỹ (·; t)
Lemma 5.6 For almost every ω, for every l ≥ 1, the family of functions
compact in C([0, ∞), Rd2 ).
t≥0
in C([0, ∞), Rd2 ). The
(l)
ỹ (·; t)
t≥0
is relatively
As a consequence of Lemmas 5.5 and 5.6 we get that, for almost every ω,
(i) the family of functions {ȳ(· + t)}t≥0 is relatively compact in C([0, ∞), Rd ),
(ii) the linearly interpolated trajectory of the slower timescale iterates, ȳ(·), is uniformly continuous on
[0, ∞).
The next proposition states that every limit point of {ȳ(· + t)}t≥0 is a solution of DI (19) on [0, ∞).
The proof is along the lines of [11, Prop. 6.5] but with modifications arising due to the presence of faster
timescale iterates.
Proposition 5.7 For almost every ω, every limit point y ∗ (·) of {ȳ(· + t)}t≥0 in C([0, ∞), Rd2 ), satisfies
the following.
(i) For some r > 0, for every l ≥ 1, there exists γ̃ (l) ∈ M(U × Br × S (2) ) such that for every t ≥ 0,
Z t Z
(l)
∗
(2)
(l)
(2)
∗
∗
h2 (x, y (q), s , u)γ̃ (q)(du, dx, ds ) dq.
y (t) = y (0) +
0
U×Br ×S (2)
(ii) For every l ≥ 1, γ̃ (l) as in part (i) of this proposition is such that for almost every t ≥ 0,
Θ1 (γ̃ (l) )(t) ∈ D(y ∗ (t)).
(iii) y ∗ (·) is absolutely continuous and for almost every t ∈ [0, ∞),
dy ∗ (t)
∈ Ĥ2 (y ∗ (t)).
dt
Proof Let Ω2 := {ω ∈ Ω : Lemma 5.2 holds}. From the proof of [11, Thm. 6.6 ] it is clear that Ω1 ⊆ Ω2
and P(Ω2 ) = 1. Fix ω ∈ Ω2 and let tn → ∞ be such that ȳ(· + tn ) → y ∗ (·) in C([0, ∞), Rd2 ).
27
(i) Fix l ≥ 1. By assumption (A8) there exists r > 0, such that supn≥0 (kxn k + kyn k) ≤ r. Then for
any t ≥ 0, γ (l) (t) := δu(l) ⊗ δxn ⊗ δs(2) ∈ P(U × Br × S (2) ) where n is such that t ∈ [ts (n), ts (n + 1)).
n
n
Therefore γ (l) ∈ M(U × Br × S (2) ) and by Lemma 2.11(i) we get that the sequence γ (l) (· + tn ) n≥1
has a convergent subsequence in M(U × Br × S (2) ). Let γ̃ (l) be a limit point of γ (l) (· + tn ) n≥1
and without loss of generality assume γ (l) (· + tn ) → γ̃ (l) as n → ∞. By definition of ỹ (l) (·; tn ), we
have that for every n ≥ 1, for every t ≥ 0,
Z t
(l)
ỹ (t; tn ) = ȳ(tn ) +
g (l) (q + tn )dq
0
Z t
(l)
(l)
(2)
= ȳ(tn ) +
h2 (x[tn +q] , y[tn +q] , s[tn +q] , u[tn +q] )dq,
0
where for any t ≥ 0, [t] := max {n ≥ 0 : t ≥ ts (n)}. Using the defintion of γ (l) (· + tn )(see (26) and
recall that γ (l) (· + tn ) := Γ(l) (ω, · + tn )) in the above we get that for every n ≥ 1, for every t ≥ 0,
Z t Z
(l)
h2 (x, y[tn +q] , s(2) , u)γ (l) (q + tn )(du, dx, ds(2) ) dq.
ỹ (l) (t; tn ) = ȳ(tn ) +
0
U×Br ×S (2)
Since ȳ(· + tn ) → y ∗ (·) in C([0, ∞), Rd2 ), by Lemma 5.5 we have that ỹ (l) (·; tn ) → y ∗ (·) as n → ∞.
By taking limit in the above equation we get that for every t ≥ 0,
Z t Z
h
i
(l)
(l)
(2)
(l)
(2)
lim ỹ (t; tn ) − ȳ(tn ) = lim
h2 (x, y[tn +q] , s , u)γ (q + tn )(du, dx, ds ) dq,
n→∞
n→∞
y ∗ (t) − y ∗ (0) = lim
n→∞
0
U×Br ×S (2)
Z t Z
0
U×Br ×S (2)
(l)
h2 (x, y[tn +q] , s(2) , u)γ (l) (q + tn )(du, dx, ds(2) ) dq.
(29)
Since γ (l) (· + tn ) → γ̃ (l) (·) and by our choice of the topology for M(U × Br × S (2) ), we have,
Z t Z
U ×Br ×S (2)
0
Z t Z
f˜(q, u, x, s(2) )γ (l) (q + tn )(du, dx, ds(2) ) dq−
0
U ×Br ×S (2)
f˜(q, u, x, s(2) )γ̃ (l) (q)(du, dx, ds(2) ) dq → 0,
for all bounded continuous f˜ : [0, t] × U × Br × S (2) → R of the form,
f˜(q, u, x, s(2) ) =
N
X
am gm (q)fm (u, x, s(2) ),
m=1
for some N ≥ 1, scalars am and bounded continuous functions gm , fm on [0, t], U × Br × S (2)
respectively, for 1 ≤ m ≤ N . By the Stone-Weierstrass theorem, such functions can uniformly
approximate any function in C([0, t] × U × Br × S (2) , R). Thus the above convergence holds true for
all real valued continuous functions on [0, t] × U × Br × S (2) , implying that
t−1 γ (l) (q + tn )(du, dx, ds(2) )dq → t−1 γ̃ (l) (q)(du, dx, ds(2) )dq in P([0, t] × U × Br × S (2) ). Thus,
k
Z t Z
0
(l)
h2 (x, y ∗ (q), s(2) , u)γ (l) (q + tn )(du, dx, ds(2) ) dq−
U ×Br ×S (2)
Z t Z
0
(l)
h2 (x, y ∗ (q), s(2) , u)γ̃ (l) (q)(du, dx, ds(2) ) dqk → 0
U ×Br ×S (2)
(30)
as n → ∞. Since ȳ(· + tn )|[0,t] n≥1 converges uniformly to y ∗ (·)|[0,t] we have that, the function
q → y[tn +q] converges uniformly to y ∗ (·)|[0,t] on [0, t]. Using the above and Lemma 5.4 we have that
for every ǫ > 0, there exists N (depending on ǫ) such that for every n ≥ N , for every q ∈ [0, t],
k
Z
(l)
h2 (x, y[tn +q] , s(2) , u)γ (l) (q+tn )(du, dx, ds(2) )−
U ×Br ×S (2)
28
Z
(l)
h2 (x, y ∗ (q), s(2) , u)γ (l) (q+tn )(du, dx, ds(2) )k < ǫ.
U ×Br ×S (2)
(31)
Now,
k
Z tZ
0
k
(l)
h2 (x, y[tn +q] , s(2) , u)γ (l) (q
U ×Br ×S (2)
Z tZ
0
k
Z tZ
0
+ tn )(du, dx, ds
(2)
Z tZ
) dq −
0
≤
Z tZ
(l)
h2 (x, y[tn +q] , s(2) , u)γ (l) (q + tn )(du, dx, ds(2) )dq −
U ×Br ×S (2)
0
(l)
h2 (x, y ∗ (q), s(2) , u)γ (l) (q
U ×Br ×S (2)
+
Z tZ
(l)
h2 (x, y ∗ (q), s(2) , u)γ (l) (q + tn )(du, dx, ds(2) ) dq −
U ×Br ×S (2)
(l)
h2 (x, y ∗ (q), s(2) , u)γ̃ (l) (q)(du, dx, ds(2) )
U ×Br ×S (2)
0
dqk
+ tn )(du, dx, ds(2) )dqk
(l)
h2 (x, y ∗ (q), s(2) , u)γ̃ (l) (q)(du, dx, ds(2) ) dqk.
U ×Br ×S (2)
Taking limit as n → ∞ in the above equation and using (30) and (31) we get,
lim k
n→∞
Z t Z
(l)
h2 (x, y[tn +q] , s(2) , u)γ (l) (q
U ×Br ×S (2)
0
+ tn )(du, dx, ds
(2)
) dq−
Z t Z
0
(l)
h2 (x, y ∗ (q), s(2) , u)γ̃ (l) (q)(du, dx, ds(2) ) dqk ≤ ǫt,
U ×Br ×S (2)
for every ǫ > 0. Therefore, for every t ≥ 0,
lim
n→∞
Z
t
0
Z
Z
(l)
h2 (x, y[tn +q] , s(2) , u)γ (l) (q + tn )(du, dx, ds(2) ) dq =
U ×Br ×S (2)
t
0
Z
(l)
h2 (x, y ∗ (q), s(2) , u)γ̃ (l) (q)(du, dx, ds(2) ) dq.
U ×Br ×S (2)
Substituting the above limit in equation (29) we get that for every t ≥ 0,
Z t Z
(l)
y ∗ (t) − y ∗ (0) =
h2 (x, y ∗ (q), s(2) , u)γ̃ (l) (q)(du, dx, ds(2) ) dq.
U×Br ×S (2)
0
(ii) Fix l ≥ 1. Let µ(l) := Θ1 (γ (l) ). Then for every n ≥ 1, µ(l) (· + tn ) = Θ1 (γ (l) (· + tn )) and since
γ (l) (·+tn ) → γ̃ (l) in M(U ×Br ×S (2) ), by Lemma 2.12(v) we get that µ(l) (·+tn ) → µ̃(l) (·) =: Θ1 (γ̃ (l) )
in M(Br × S (2) ) as n → ∞. In order to prove that for almost every t ≥ 0, µ̃(l) (t) ∈ D(y ∗ (t)) we
need to show that for almost every t ≥ 0,
(l)
(1) supp(µ̃Br (t)) ⊆ λ(y ∗ (t)),
(l)
(2) for every A ∈ B(S (2) ), µ̃S (2) (t)(A) =
R
Rd1 ×S (2)
Π(2) (x, y ∗ (t), s(2) )(A)µ̃(l) (t)(dx, ds(2) ).
(l)
First we present a proof of the claim in (1) above. µBr := Θ2 (µ(l) ) ∈ M(Br ) and since as n → ∞,
(l)
(l)
µ(l) (· + tn ) → µ̃(l) , by Lemma 2.12(vi), we have that µBr (· + tn ) → µ̃Br in M(Br ). Since as n → ∞,
(l)
(l)
µBr (· + tn ) → µ̃Br in M(Br ), by proof of [10, Ch. 6, Lemma 3], we have that for almost every t ≥ 0,
there exists a subsequence {nk }k≥1 and a subsequence of natural numbers {cp }p≥1 such that
cp
1 X (l)
(l)
µBr (t + tnk ) → µ̃Br (t)
cp
(32)
k=1
(l)
in P(Br ) as p → ∞. Fix t ≥ 0 such that the above holds. By definition of µBr , we have that
(l)
for every k ≥ 1, µBr (t + tnk ) = δx[t+tn
k
]
where [t + tnk ] := max {m > n : t + tnk ≥ ts (m)}. Since
ȳ(·+tnk ) → y ∗ (t), using the uniform continuity of ȳ(·), we have that the function t̃ ∈ [0, ∞) → y[t̃+tn ]
k
converges uniformly on compacts to the function y ∗ (·). Therefore y[t+tnk ] → y ∗ (t) as k → ∞ and
by Lemma 5.2 we have that x[t+tnk ] → λ(y ∗ (t)). Further by definition of r > 0 as in part (i) of this
proposition, we get that supn≥1 kxn k ≤ r. Hence λ(y ∗ (t)) ∩ Br 6= ∅ and x[t+tnk ] → λ(y ∗ (t)) ∩ Br as
k → ∞. For every ǫ > 0, clearly (λ(y ∗ (t)) + Bǫ ) ∩ Br is compact and there exists K large such that
for every k ≥ K, x[t+tnk ] ∈ (λ(y ∗ (t)) + Bǫ ) ∩ Br and hence δx[t+tn ] ((λ(y ∗ (t)) + Bǫ ) ∩ Br ) = 1 for
k
Pcp
(l)
(l)
µBr (t + tnk ) → µ̃Br (t) in P(Br ) as p → ∞, by [17, Thm. 2.1.1(iv)] we
every k ≥ K. Since c1p k=1
have that, for every ǫ > 0,
lim sup
p→∞
cp
cp
1 X (l)
1 X
µBr (t + tnk )((λ(y ∗ (t)) + Bǫ ) ∩ Br ) = lim sup
δx[t+tn ] ((λ(y ∗ (t)) + Bǫ ) ∩ Br )
k
cp
p→∞ cp
k=1
k=1
=1
(l)
≤ µ̃Br (t)((λ(y ∗ (t)) + Bǫ ) ∩ Br ) ≤ 1.
29
(l)
(l)
Therefore for every ǫ > 0, µ̃Br (t)((λ(y ∗ (t)) + Bǫ ) ∩ Br ) = 1 which gives us that µ̃Br (t)(λ(y ∗ (t)) ∩
(l)
Br ) = 1 and hence supp(µ̃Br (t)) ⊆ λ(y ∗ (t)). Since (32) holds for almost every t ≥ 0, we have that
(l)
for almost every t ≥ 0, supp(µ̃Br (t)) ⊆ λ(y ∗ (t)).
The proof of the claimin (2) above is similar to the proof of [11, Prop. 6.5(ii)] and we provide a brief
outline for the sake of completeness. Let {fi }i≥1 ⊆ C(S (2) , R) be a convergence determining class
for P(S (2) ). By an appropriate affine transformation we can ensure that for every i ≥ 1, for every
s(2) ∈ S (2) , 0 ≤ fi (s(2) ) ≤ 1. Define,
ζni
:=
n−1
X
k=0
Z
2
b(k) fi (Sk+1 ) −
fi (s̃
(2)
(2)
)Π
S (2)
(Xk , Yk , Sk2 )(ds̃(2) )
,
for every n ≥ 1, for every i ≥ 1. For every i ≥ 1, ζni n≥1 is a square integrable martingale
i
P∞
− ζni )2 |Fn ≤
w.r.t. the filtration Fn := σ Xk , Yk , Sk2 : 0 ≤ k ≤ n n≥1 and further n=0 E (ζn+1
P
2
2 ∞
theorem (see [10, Appendix
C, Thm. 11]) we get
n=1 (b(n)) < ∞. By Martingale convergence
that for almost every ω, for every i ≥ 1, ζni n≥1 converges. Let Ωm := ω ∈ Ω : ∀i ≥ 1, {ζni } converges .
Define, Ω∗ := Ωm ∩ Ω2 and from the arguments above we get that P(Ω∗ ) = 1. Therefore for every
ω ∈ Ω∗ , for every i ≥ 1 for every T > 0,
τ 2 (n,T )
X
k=n
Z
(2)
b(k) fi (sk+1 ) −
fi (s̃
(2)
(2)
)Π
S (2)
(2)
(xk , yk , sk )(ds̃(2) )
→ 0,
as n → ∞. By our choice of {fi }i≥1 , the fact that the step size sequence {b(n)}n≥0 is non- increasing
and the definition of µ(l) , we get that for every ω ∈ Ω∗ , for every i ≥ 1, for every T > 0,
Z
Z TZ
fi (s(2) ) −
fi (s̃(2) )Π(2) (x, y[q+t] , s(2) )(ds̃(2) ) µ(l) (q + t)(dx, ds(2) )dq = 0,
lim
t→∞
S (2)
Br ×S (2)
0
where [q + t] := max {n ≥ 0 : t ≥Rts (n)}. By assumption (A4), we have that for every i ≥ 1, the
function (x, y, s(2) ) → fi (s(2) ) − S (2) fi (s̃(2) )Π(2) (x, y, s(2) )(ds̃(2) ) is continuous and hence the restriction of the above function to the compact set Br × rU × S (2) is uniformly continuous where
r > 0 is as in part (i) of this proposition. Using the uniform continuity above and the fact that
limt→∞ sup0≤q≤T ȳ(q + t) − y[q+t] = 0 (which follows from definition of ȳ(·) and uniform continuity of ȳ(·)) we get that for every ω ∈ Ω∗ , for every i ≥ 1, for every T > 0,
lim
t→∞
T
Z
Z
Br ×S (2)
0
fi (s(2) ) −
Z
S (2)
fi (s̃(2) )Π(2) (x, ȳ(q + t), s(2) )(ds̃(2) ) µ(l) (q + t)(dx, ds(2) )dq = 0.
(33)
We know that as n → ∞, ȳ(·+tn ) → y (·) in C([0, ∞), R ) and µ (·+tn ) → µ̃ (·) in M(Br ×S (2) ).
Further by arguments similar to Lemma 5.4, the family of functions
Z
Z
(2)
(2)
(2)
(2)
(2)
(2)
(2)
fi (s̃ )Π (x, y, s )(ds̃ ) ν(dx, ds ) : ν ∈ P(Br × S )
fi (s ) −
y ∈ rU →
∗
(l)
d2
(l)
S (2)
Br ×S (2)
is equicontinuous. Therefore, for every ω ∈ Ω∗ , for every T > 0,
lim
t→∞
Z TZ
0
Z
Br ×S (2)
Z
lim
t→∞
T
Z
0
Z
fi (s̃(2) )Π(2) (x, y ∗ (q), s(2) )(ds̃(2) )−
S (2)
Br ×S (2)
−
Z
T
0
Z
fi (s(2) ) −
Br ×S (2)
Z
S (2)
fi (s̃(2) )Π(2) (x, ȳ(q + t), s(2) )(ds̃(2) ) µ(l) (q + t)(dx, ds(2) )dq = 0,
S (2)
fi (s̃(2) )Π(2) (x, y ∗ (q), s(2) )(ds̃(2) ) µ(l) (q + t)(dx, ds(2) )dq
fi (s(2) ) −
Z
S (2)
fi (s̃(2) )Π(2) (x, y ∗ (q), s(2) )(ds̃(2) ) µ̃(l) (q)(dx, ds(2) )dq = 0.
Using the above and equation (33), we get that for every ω ∈ Ω∗ , for every T > 0,
Z
0
T
Z
Br ×S (2)
fi (s
(2)
)−
Z
fi (s̃
(2)
(2)
)Π
∗
(x, y (q), s
S (2)
30
(2)
)(ds̃
(2)
) µ̃(l) (q)(dx, ds(2) )dq = 0.
By applying Lebesgue’s differentiation theorem (see [10, Ch. 11.1.3]), we get that for every ω ∈ Ω∗ ,
for every i ≥ 1, for almost every t ≥ 0,
Z
Z
fi (s(2) ) −
fi (s̃(2) )Π(2) (x, y ∗ (t), s(2) )(ds̃(2) ) µ̃(l) (t)(dx, ds(2) ) = 0.
S (2)
Br ×S (2)
Since for every t ≥ 0, µ̃(l) (t) ∈ P(Br ×S (2) ) it is also an element of P(Rd1 × S (2) ) with supp(µ̃(l) (t)) ⊆
Br × S (2) . Therefore, for every ω ∈ Ω∗ , for every i ≥ 1, for almost every t ≥ 0,
Z
Z
fi (s̃(2) )Π(2) (x, y ∗ (t), s(2) )(ds̃(2) ) µ̃(l) (t)(dx, ds(2) ) = 0.
fi (s(2) ) −
S (2)
Rd1 ×S (2)
Since {fi }i≥1 is a convergence determining class for P(S (2) ), from the above it follows that for every
ω ∈ Ω∗ , for almost every t ≥ 0,
Z
(l)
Π(2) (x, y ∗ (t), s(2) )(ds̃(2) )µ̃(l) (t)(dx, ds(2) ).
µ̃S (2) (t)(ds̃(2) ) =
Rd1 ×S (2)
(iii) Fix l ≥ 1. Then by part (i) of this proposition, y ∗ (·) is clearly absolutely continuous and for almost
every t ≥ 0,
Z
dy ∗ (t)
(l)
h2 (x, y ∗ (t), s(2) , u)γ̃ (l) (t)(du, dx, ds(2) ).
=
dt
U×Rd1 ×S (2)
(l)
By part (ii) of this lemma we know that for almost every t ≥ 0, Θ1 (γ̃ (l) )(t) = γ̃Br ×S (2) (t) =
(l)
γ̃Rd1 ×S (2) (t) ∈ D(y ∗ (t)). Hence, by Lemma 2.9(ii), for almost every t ≥ 0,
Z
dy ∗ (t)
(l)
h2 (x, y ∗ (t), s(2) , u)γ̃ (l) (t)(du, dx, ds(2) )
=
d
dt
(2)
1
U×R ×S
Z
(l)
H2,y∗ (t) (x, s(2) )µ(dx, ds(2) )
∈ ∪µ∈D(y∗ (t))
Rd1 ×S (2)
=
(l)
Ĥ2 (y ∗ (t)).
Since the above holds for every l ≥ 1, we get that for almost every t ≥ 0,
dy ∗ (t)
(l)
∈ ∩l≥1 Ĥ2 (y ∗ (t)) = Ĥ2 (y ∗ (t)),
dt
where the last equality follows from Lemma 4.2(vi).
A continuous function y : R → Rd2 is said to be an asymptotic pseudotrajectory for the flow of DI
(19) if limt→∞ D(y(· + t), Σ2 ) = 0 where Σ2 ⊆ C(R, Rd2 ) denotes the set of solutions of DI (19). Fix
ω ∈ Ω∗ . Extend ȳ(·) to the whole of R by defining ȳ(t) = ȳ(0) for every t < 0. Then by assumption (A8)
and uniform continuity of ȳ(·) we have that the family of functions {ȳ(· + t)}t≥0 is relatively compact
in C(R, Rd2 ). Let y ∗ (·) be a limit point of the above family of functions. Then by Proposition 5.7(iii),
y ∗ (·)|[0,∞) is a solution of DI (19) on [0, ∞). Usually the negative time argument is omitted since it follows
from the positive time argument as follows:
Fix T > 0. Let tn → ∞ be such that ȳ(· + tn ) → y ∗ (·) in C(R, Rd2 ). Then ȳ(· + tn − T ) → y ∗ (· − T ). By
Proposition 5.7(iii), y ∗ (· − T )|[0,∞) is a solution of DI (19) on [0, ∞). Therefore y ∗ (·)|[−T,0] is absolutely
∗
continuous and for a.e. t ∈ [−T, 0], dydt(t) ∈ Ĥ2 (y ∗ (t)). Since T was arbitrary we have that the y ∗ (·)|(−∞,0]
is solution of DI (19) on (−∞, 0]. Therefore y ∗ (·) ∈ Σ2 and by [12, Thm. 4.1] we get the following result.
Theorem 5.8 [APT] Under assumptions (A1) − (A9) and (A11), for almost every ω, the linearly interpolated trajectory of the slower timescale recursion (14a), ȳ(·), is an asymptotic pseudotrajectory of DI
(19).
5.2.3
Characterization of limit sets
As a consequence of Theorem 5.8 for almost every ω, the limit sets of the slower timescale recursion, L(ȳ),
defined as,
(34)
L(ȳ) := ∩t≥0 {ȳ(q + t) : q ≥ 0},
can be characterized in terms of the dynamics of DI (19). Further using Lemma 5.2 we get the main result
of this paper stated below.
31
Theorem 5.9 [Limit set] Under assumptions (A1) − (A9) and (A11), for almost every ω,
(i) L(ȳ) is a non-empty, compact subset of Rd2 and is internally chain transitive for the flow of DI (19),
(ii) if assumption (A10) is satisfied then L(ȳ) ⊆ Y and as n → ∞,
xn
→ ∪y∈Y (λ(y) × {y}) .
yn
Proof Fix ω ∈ Ω∗ .
(i) By Theorem 5.8 we know that ȳ(·) is an asymptotic pseudotrajectory for the flow of DI (19). Now
the claim follows from [12, Thm. 4.3].
(ii) By part (i) of this theorem we know that L(ȳ) is internally chain transitive for the flow of DI(19).
Since Y is a globally attracting set for DI(19), by [12, Cor. 3.24] we get that L(ȳ) ⊆ Y. Therefore
yn → Y as n → ∞ and by Lemma 5.2, we get that, as n → ∞,
xn
→ ∪y∈Y (λ(y) × {y}) .
yn
6
Application: Constrained convex optimization
In this section we consideran application of the theory to a problem of constrained convex optimization.
Throughout this section we assume that S (1) = S (2) = S and |S| < ∞.
Let the objective function J : Rd1 × S → R be such that J(·) is continuous and for every s ∈ S, J(·, s)
is convex and coercive (that is for any M > 0, there exists r > 0, such that for any x ∈ Rd1 with kxk ≥ r,
we have that J(x, s) ≥ M ). The functions describing the constraints
are given by C : S → Rd2 ×d1 and
d2
d1
w : S → R . We assume that for any s ∈ S, the set X (s) := x ∈ R : C(s)x = d(s) is non empty. The
law of the Markov noise terms is given by Π : Rd1 × S → P(S) such that Π is continuous and let µ ∈ P(S)
denote the unique stationary distribution of the Markov chain given by the transition kernel Π(x, ·)(·), for
every x ∈ Rd1 .
Let ∂J(x, s) denote the set of subgradients of the convex function J(·, s) at the point x ∈ Rd1 . Formally,
∂J(x, s) := g ∈ Rd1 : ∀x′ ∈ Rd1 , J(x′ , s) ≥ J(x, s) + hg, x′ − xi .
Then it is easy to show that for every (x, s) ∈ Rd1 × S, ∂J(x, s) is convex and compact. Further the map
(x, s) → ∂J(x, s) possesses the closed graph property. We assume that the map (x, s) → ∂J(x, s) satisfies
the linear growth property, that is, there exists K > 0 such that supx′ ∈∂J(x,s) kx′ k ≤ K(1 + kxk).
R
for every x ∈ Rd1 , Jµ (x) := S J(x, s)µ(ds). Similarly define
LetRJµ : Rd1 → R be defined such that
R
Cµ := S C(s)µ(ds) ∈ Rd2 ×d1 and wµ := S w(s)µ(ds) ∈ Rd2 . The optimization problem that we wish to
solve is given by,
OP (µ) :
min Jµ (x), subject to :
x∈Rd1
Cµ x = wµ .
The standard approach in solving the optimization problem OP (µ) is the projected subgradient descent
algorithm whose recursion is given by,
Xn+1 = Pµ (Xn − a(n)(gn + Mn+1 )) ,
where gn ∈ ∂Jµ (Xn ), Mn+1 is the
subgradient estimation error and Pµ denotes the projection operation
onto the affine subspace Xµ := x ∈ Rd1 : Cµ x = wµ . Such a scheme cannot be implemented when µ is
not known. Such is the case in problems arising in optimal control.
The feasible set of the optimization problem OP (µ), given by Xµ is non empty since for every s ∈ S,
X (s) is non-empty. Further, since for every s ∈ S, J(·, s) is coercive, the function Jµ (·) is coercive and
hence bounded below. Therefore the optimization problem OP (µ) has at least one solution. Let the
solution set of the optimization problem OP (µ), be denoted by Z.
origin. For every s ∈ S,
For any r > 0, let Br denote the closed ball of radius r in Rd1 centered at the P
pick xs ∈ X (s), and compute M1 := max{J(xs , s′ ) : s, s′ ∈ S}. Then xµ := s∈S µ(s)xs ∈ Xµ and
Jµ (xµ ) ≤ M1 . Since |S| < ∞ and the functions J(·, s) are coercive, for some M > max{0, M1 }, there
32
exists r > max{kxs k : s ∈ S}, such that for every s ∈ S, for every x ∈ Brc , J(x, s) ≥ M and for every
s ∈ S, Br ∩ X (s) 6= ∅. Then Z ⊆ Br . Instead of OP (µ) we shall solve the following penalized/regularized
optimization problem given by,
˜ (µ) :
OP
n
o
ǫ
K +1
2
2
2
kxk
+
max
kxk
−
r
,
0
,
2r2
2
x∈Rd1
subject to : Cµ x = wµ ,
min Jµ (x) +
where, r > 0 is as determined above, K is the constant associated with the linear growth property of the
subgradient map ∂J(·, s) and ǫ > 0 is an arbitrary constant small in value. Then it is easy to show that
˜ (µ) has at least one solution and the set of solutions of OP
˜ (µ), denoted by Z̃ is such that Z̃ ⊆ Br .
OP
Further for any x̃ ∈ Z̃, for any x∗ ∈ Z, Jµ (x̃) − Jµ (x∗ ) ≤ ǫ.
˜ (µ) defined such that
Consider the Lagrangian L : Rd → R associated with optimization problem OP
d
for every (x, y) ∈ R ,
L(x, y) := Jµ (x) +
n
o
ǫ
K +1
kxk2 +
max kxk2 − r2 , 0 + hy, Cµ x − wµ i .
2
2r
2
ˆ s) := J(x, s) +
Let Jˆ : Rnd1 × S → R,obe defined such that for every (x, s) ∈ Rd1 × S, J(x,
2
K+1
2
d1
× S,
2 max kxk − r , 0 . Then, for every (x, s) ∈ R
n
o
K +1
ǫ
max kxk2 − r2 , 0 =
Jˆµ (x) := Jµ (x) + 2 kxk2 +
2r
2
Z
ǫ
2r 2
2
kxk +
ˆ s)µ(ds).
J(x,
S
When the transition law Π and hence µ is not known we propose the following recursion which performs
primal descent along the faster time scale (that is minimization of L(·, y) w.r.t. x) and dual ascent on the
slower timescale (that is maximization of L(x, ·) w.r.t. y). The recursion is given by,
Yn+1 − Yn = b(n)(C(Sn )Xn − w(Sn )),
1
ˆ n , Sn ) + C(Sn )T Yn ,
Xn+1 − Xn − a(n)Mn+1
∈ −a(n) ∂ J(X
(35a)
(35b)
where the step size sequences {a(n)}n≥0 and {b(n)}n≥0 are chosen such that they satisfy assumption
(A5) and Mn1 n≥1 denotes the subgradient estimation error which is assumed to satisfy assumption (A6)
( for example, when Mn1 n≥1 is i.i.d. zero mean with finite variance, assumption (A6) is satisfied. More
generally (A6) is satisfied if Mn1 n≥1 is a martingale difference sequence satisfying assumption (A3) in
[10, ch. 2.1]).
ˆ s) + C(s)T y and (x, y, s) → C(s)x − w(s)
It is easy to see that the maps (x, y, s) → − ∂ J(x,
satisfy
assumptions (A1)
and (A2) respectively. The linear growth property of the map (x, y, s) →
T
ˆ
− ∂ J(x, s) + C(s) y , follows from the linear growth property of x → ∂J(x, s). Further by [3, Prop. 5.4.6],
R
ˆ s) + C(s)T y µ(ds) = − ∂ Jˆµ (s) + C T y = −∂L(x, y).
we get that for every (x, y) ∈ Rd , − S ∂ J(x,
µ
o
n
For every y ∈ Rd2 , let λ(y) := x ∈ Rd1 : −CµT y ∈ ∂ Jˆµ (x) . Then for every y ∈ Rd2 , λ(y) is non-empty
since L(·, y) is convex and coercive. Further |λ(y)| = 1, that is λ(y) is a singleton since, L(·, y) is strictly
convex. For any y ∈ Rd2 , x′ ∈ λ(y) if and only if there exists g̃ ∈ Rd1 in the set of subgradients of the
ǫ ′
2
2
′
T
′
function Jµ (·) + K+1
2 max{k · k − r , 0} at x , such that g̃ + r 2 x + Cµ y = 0. So either kx k ≤ r or if
′
′ 2
2
′ 2
2
′
kx k > r, then max{kx k − r , 0} = kx k − r and hence for some g ∈ ∂Jµ (x ), g̃ = g + (K + 1)x′ , from
which we get that,
(K + 1 +
ǫ
)kx′ k = kg + CµT yk
2r2
≤ K + Kkx′ k + kCµT kkyk.
Thus for K ′ := max{K, r, kCµT k}, we get that for every y ∈ Rd2 , kλ(y)k ≤ K ′ (1 + kyk). The set λ(y)
T
ˆ
and by [14, Thm. 6] the map
is clearly globally attracting for the flow of DI dx
dt ∈ − ∂ Jµ (x) + Cµ y
y ∈ Rd2 → λ(y) is u.s.c. (since λ(·) is also single valued, it is continuous). Hence the map λ(·) satisfies
assumption (A9).
33
If the iterates are stable for a.e. ω (that is, (A8) is satisfied), the result in section 5.1 gives us that for
almost every ω, there exists a non-empty compact set A ⊆ Rd , such that (Xn (ω), Yn (ω)) → A as n → ∞
and A is internally chain transitive for the flow of DI,
dx
−∂ Jˆµ (x) − CµT y
dt
∈
.
(36)
dy
0
dt
By arguments in section 5.1, we have that A ∩ G(λ) 6= ∅, where G(λ) := (λ(y), y) : y ∈ Rd2 . Let
O ⊆ Rd be an open neighborhood of G(λ). Let O′ (δ) := {(x, y) ∈ Rd : L(x, y) − L(λ(y), y) < δ}. By [14,
ch. 1.2, Thm. 6], the map y ∈ Rd2 → L(λ(y), y) is continuous and hence O′ (δ) is an open neighborhood
of G(λ). Further it is easy to show that ∩δ>0 O′ (δ) = G(λ) and hence ∩δ>0 (O′ (δ) ∩ A) = G(λ) ∩ A. Since
A ⊆ Rd is compact, there exists δ ∗ > 0, such that O′ (δ ∗ ) ∩ A ⊆ O ∩ A. Consider any solution of DI
(36), (x(·), y(·)) starting at (x∗ , y ∗ ) ∈ O′ (δ ∗ ) ∩ A and satisfying for every t ∈ R, (x(t), y(t)) ∈ A. Recall
from section5.1 that (x(·), y(·)) as above is such that for every t ≥ 0, y(t) = y ∗ and x(·) is a solution of
T ∗
∗
∗
ˆ
DI, dx
dt ∈ −(∂ Jµ (x) + Cµ y ) = −∂L(x, y ) and hence descends along the potential L(x, y ). Therefore
′ ∗
A
′ ∗
the solution (x(·), y(·)) remains within O (δ ) ∩ A which gives us that Φ (O (δ ) ∩ A, [0, ∞)) ⊆ O ∩ A,
where ΦA denotes the flow of DI (36) restricted to the set A. Thus assumption (A11) is satisfied and from
Lemma 5.2 we get the following result.
Lemma 6.1 [Faster timescale convergence] For almost every ω, (Xn (ω), Yn (ω)) → G(λ) as n → ∞.
Theorem 5.8 gives us that the iterates {Yn }n≥0 in recursion (35a) track the flow of DI,
Z
dy
(C(s)x − w(s)) ν(dx, ds),
∈ ∪ν∈D(y)
dt
Rd1 ×S
(37)
where, for every y ∈ Rd2 , D(y) is as in equation (17). Since for every y ∈ Rd2 , λ(y) is a singleton and
since µ is the unique stationary distribution of the Markov chain given by transition kernel Π(·)(·), we get
that for every y ∈ Rd2 , D(y) = δλ(y) ⊗ µ. Therefore DI (37) takes the form,
dy
= Cµ λ(y) − wµ .
dt
(38)
In order to analyze the asymptotic behavior of o.d.e. (38), we need the following version of the envelope
theorem. The proof of the envelope theorem below is similar to that in [7].
Lemma 6.2 [envelope theorem] Let y : [0, T ] → Rd2 be an absolutely continuous function. Let L̃ :
Rd1 × [0, T ] → R be defined such that for every (x, t) ∈ Rd1 × [0, T ], L̃(x, t) := L(x, y(t)). Then,
] with Lebesgue measure
(i) for every x ∈ Rd1 , L̃(x, ·) is absolutely continuous and there exists D ⊆ [0, T D
E
∂ L̃(x,t)
d1 ∂ L̃(x,t)
,
C
x
−
w
.
T such that for every t ∈ D, for every x ∈ R , ∂t exists and ∂t = dy(t)
µ
µ
dt
(ii) the function V : [0, T ] → R where for every 0 ≤ t ≤ T , V (t) := inf x∈Rd1 L̃(x, t) is absolutely
continuous. Further for any 0 < t ≤ T ,
Z t
∂ L̃(x, q)
dq.
V (t) = V (0) +
∂q
0
x=λ(y(q))
Proof (i) Since y(·) is absolutely continuous, it is differentiable almost everywhere and let D ⊆ [0, T ]
be the set of t ∈ [0, T ], such that dy(t)
dt exists. Then clearly the Lebesgue measure of D is T .
Fix x ∈ Rd1 . Then L(x, ·) is a Lipschitz continuous function since for any y ′ , y ′′ ∈ Rd2 ,
|L(x, y ′ ) − L(x, y ′′ )| = |hy ′ − y ′′ , Cµ x − wµ i|
≤ ky ′ − y ′′ k kCµ x − wµ k
= βx ky ′ − y ′′ k ,
where βx := kCµ x − wµ k. Further L(x, ·) is differentiable (i.e. totally differentiable since it is linear
in y) and the total derivative is given by ∇y L(x, y ′ ) = (Cµ x − wµ ) for every y ′ ∈ Rd2 . Since L̃(x, ·),
is the composition of absolutely continuous function y(·) and a Lipschitz continuous function L(x, ·),
we have that L̃(x, ·) is absolutely
continuous.E By [21, Thm. 9.15], we have that for every t ∈ D,
D
∂ L̃(x,t)
dy(t ′ )
∂ L̃(x,t)
exists and dt =
dt
dt , Cµ x − wµ .
34
(ii) Since y(·) is absolutely continuous, there exists α > 0 such that supt∈[0,T ] ky(t)k ≤ α. Further by
assumption (A9), for every t ∈ [0, T ],
V (t) =
inf
x∈Rd1 :
kxk≤K ′ (1+ky(t)k)
L̃(x, t) =
inf
x∈Rd1 :
kxk≤K ′ (1+α)
L̃(x, t).
Therefore for every 0 ≤ t < t′ ≤ T ,
|V (t′ ) − V (t)| ≤
L̃(x, t′ ) − L̃(x, t)
sup
x∈Rd1 :
kxk≤K ′ (1+α)
≤
sup
x∈Rd1 :
kxk≤K ′ (1+α)
=
Z
t′
∂ L̃(x, q)
dq
∂t
t
|hy(t′ ) − y(t), Cµ x − wµ i|
sup
x∈Rd1 :
′
kxk≤K (1+α)
≤
sup
x∈Rd1 :
′
kxk≤K (1+α)
kCµ x − wµ k ky(t′ ) − y(t)k .
Now the absolute continuity of V (·) follows from absolute continuity of y(·). Since V (·) is absolutely
Rt
continuous, dVdq(q) exists for a.e. q ∈ [0, T ] and for any 0 < t ≤ T , V (t) = V (0) + 0 dVdq(q) dq. Let
q ∈ (0, T ) be such that
dV (q)
dq
exists and q ∈ D. Then for q ′ > q, V (q ′ ) − V (q) ≤ L̃(λ(y(q)), q ′ ) −
L̃(λ(y(q)), q). Therefore the right hand derivative of V (·) at q which is the same as
dV (q)
dq
≤ ∂ L̃(x,q)
|x=λ(y(q)) .
isfies,
∂q
∂ L̃(x,q)
dV (q)
|x=λ(y(q)) ≤ dq . Thus
∂q
dV (q)
dq
sat-
′
Considering q < q and repeating the above argument gives us,
for a.e. q ∈ [0, T ],
absolutely continuous, for any 0 < t ≤ T ,
Z t
∂ L̃(x, q)
V (t) = V (0) +
∂q
0
∂ L̃(x,q)
|x=λ(y(q))
∂q
=
dV (q)
dq
and since V (·) is
dq.
x=λ(y(q))
Let Qµ : Rd2 → R be defined such that for y ∈ Rd2 , Qµ (y) := inf x∈Rd1 L(x, y) = L(λ(y), y). The
˜ (µ) and is a concave
function Qµ (·) is the objective function of the dual of the optimization problem OP
function. By the strong duality theorem (see [3, Prop. 5.3.3]), the dual optimization problem given by,
maxy∈Rd2 Qµ (y) has at least one solution and let the set of solutions of the dual optimization problem be
denoted by Y. Further the strong duality theorem also gives us that for any y ∈ Y and for any x ∈ Z̃,
Qµ (y) = Jˆµ (x).
Let y : R → Rd2 be a solution of the o.d.e. (38) with initial condition y ∈ Y. Then y(·) is absolutely
continuous and for a.e. t ∈ [0, ∞), dy(t)
dt = Cµ λ(y(t)) − wµ . By Lemma 6.2(ii) we have that for any t ≥ 0,
t
∂ L̃(x, q)
dq,
∂q
0
x=λ(y(q))
Z t
dy(q)
, Cµ λ(y(q)) − wµ dq,
= V (0) +
dq
0
Z t
2
= V (0) +
kCµ λ(y(q)) − wµ k dq.
V (t) = V (0) +
Z
(39)
0
Since V (t) = Qµ (y(t)) and V (0) = Qµ (y), where y ∈ Y, we get that V (t) − V (0) ≤ 0. Hence for every
Rt
t ≥ 0, 0 kCµ λ(y(q)) − wµ k2 dq ≤ 0 which gives us that kCµ λ(y(t)) − wµ k = 0 for a.e. t ∈ [0, ∞). Thus
for any solution of o.d.e. (38), y(·), with initial
condition y ∈ Y, we have that Cµ λ(y) − wµ = 0 and
for every t ≥ 0, y(t) = y. Therefore Y ⊆ y ∈ Rd2 : Cµ λ(y) − wµ = 0 . Further by [3, Prop. 5.3.3(ii)],
any y ∈ Rd2 , such that Cµ λ(y) − wµ = 0, is a solution of the dual optimization problem and hence
Y = y ∈ Rd2 : Cµ λ(y) − wµ = 0 (from this it also follows that Y is closed).
In the theorem below we summarize the main convergence result associated with the recursion (35).
Theorem 6.3 [Convergence to Lagrangian saddle points]
35
(i) For any solution y(·) of the o.d.e. (38) with any initial condition y0 ∈ Rd2 which is bounded for
t ≥ 0 (that is supt≥0 ky(t )k < ∞), we have that as t → ∞, inf y∈Y ky(t) − yk → 0.
˜ (µ) (that is λ(y) ∈ Z̃).
(ii) For any y ∈ Y, λ(y) is a solution of the optimization problem OP
(iii) If the iterates remain stable for almost every ω (that is (A8) is satisfied), then, for almost every ω,
(a) Yn (ω) → Y as n → ∞,
λ(y)
Xn (ω)
⊆ Rd .
(b)
→ ∪y∈Y
y
Yn (ω)
Proof
(i) Let y(·) be a solution of the o.d.e. (38) with initial condition y0 ∈ Rd2 (assume y0 ∈
/ Y since, otherwise
we know that for every t ≥ 0, y(t) = y0 and hence the claim follows) such that supt≥0 ky(t )k ≤ M
for some M > 0. Then y(·)|[0,∞) is uniformly continuous since for any 0 ≤ t < t′ < ∞,
′
ky(t ) − y(t)k =
Z
t′
(Cµ y(q) − wµ ) dq
t
≤
Z
t′
(kCµ k ky(q)k + kwµ k) dq
t
≤ (kCµ k M + kwµ k) (t′ − t).
The function y ∈ Rd2 → kCµ y − wµ k is uniformly continuous and hence the function t ∈ [0, ∞) →
kCµ y(t ) − wµ k is uniformly continuous. Further by Lemma 6.2(ii), for any t > 0, 0 ≤ V (t) − V (0) ≤
Qµ (y) − V (0) < ∞ where y ∈ Y. The claim that as t → ∞, y(t) → Y is equivalent to the claim that
as t → ∞, kCµ y(t) − wµ k → 0.
Suppose there exists ǫ > 0, such that for every T > 0, there exists t ≥ T , such that kCµ y(t) − wµ k >
ǫ. From the uniform continuity of t → kCµ y(t) − wµ k, there exists δ > 0 such that for every
t, t′ ∈ [0, ∞) satisfying |t − t′ | < δ, |kCµ y(t) − wµ k − kCµ y(t′ ) − wµ k| < 2ǫ . Therefore we can obtain
a sequence {tn }n≥1 such that for every n ≥ 1, δ < tn < tn+1 − 2δ and for every t ∈ (tn − δ, tn + δ),
2(Qµ (y)−V (0))
< N where y ∈ Y. Then by Lemma 6.2(ii),
kCµ y(t) − wµ k > 2ǫ . Let N be such that
ǫ2 δ
we get that,
Qµ (y(tN +1 )) − V (0) = V (tN +1 ) − V (0)
Z tN +1
dy(q)
=
, Cµ λ(y(q)) − wµ dq
dq
0
Z tN +1
kCµ y(q) − wµ k2 dq
=
≥
0
N
X Z tn +δ
n=1
>N
kCµ y(q) − wµ k2 dq
tn −δ
ǫ2 δ
2
> Qµ (y) − V (0)
which contradicts the fact that V (t) − V (0) ≤ Qµ (y) − V (0). Therefore limt→∞ kCµ y(t ) − wµ k = 0 .
˜ (µ). By definition
(ii) Let y ∈ Y. Then we know that Cµ λ(y) − wµ = 0 and hence λ(y) is feasible for OP
d1
of λ(y), we have that for every x ∈ R , L(λ(y), y) ≤ L(x, y). Now the claim follows from [3,
Prop. 5.3.3(ii)].
(iii) Let ω be such that Theorem 5.9 holds.
(a) Then by Theorem 5.9(i) we know that there exists a non empty, compact set A ⊆ Rd2 such
that as n → ∞, Yn (ω) → A. Further A is internally chain transitive for the flow of o.d.e. (38)
and hence is invariant. Let y(·) be a solution to o.d.e. (38) with initial condition in A and
for every t ∈ R, y(t ) ∈ A. Since A is compact, supt≥0 ky(t )k < ∞ and hence by part (i) of
36
this theorem we get that y(t) → Y as t → ∞.
′ Sinced for every t ≥ 0, ′ y(t ) ∈ A, we get that
2
) < δ ∩ A = Y ∩ A.
Y ∩ A 6= ∅. Further for some y ∈ Y, ∩δ>0
′ y ∈ dR : Qµ (y) − Qµ (y
For any ǫ > 0, there exists δǫ > 0 such that y ∈ R 2 : Qµ (y) − Qµ (y ′ ) < δ ∩ A ⊆ N ǫ (Y ∩ A)
where N ǫ(·) denotes the ǫ-neighborhood of a set. By
using Lemma6.2(ii), it is easy to show
that ΦA y ′ ∈ Rd2 : Qµ (y) − Qµ (y ′ ) < δ ∩ A, [0, ∞) ⊆ N ǫ (Y ∩ A) where ΦA denotes the flow
of o.d.e. (38) restricted to set A (see section 2.3). Therefore Y ∩ A is an attracting set for the
flow ΦA . From [12, Prop. 3.20] we get that Y ∩ A = A. Therefore as n → ∞, Yn (ω) → Y.
(b) Follows from part (iii)(a) of this theorem and Lemma 6.1.
7
Conclusions and directions for future work
We have presented a detailed analysis of a two timescale stochastic recursive inclusion with set-valued drift
functions and in the presence of non-additive iterate dependent Markov noise with non-unique stationary
distributions. Analysis in section 5 shows us that the asymptotic behavior of the two timescale recursion
(14) is such that the faster timescale iterates in recursion (14b), track the flow of DI (16) for some
fixed value of the slower timescale variable and the slower timescale iterates track the flow of DI (19). The
assumptions under which the two timescale recursion is studied in this paper is weaker than those in current
literature. Recursions with such behavior are often required to solve nested minimization problems which
arise in machine learning and optimization. A special case of constrained convex optimization with linear
constraints is considered as an application where the objective function is not assumed to be differentiable
and further the objective function and constraints are averaged with respect to stationary distribution
of an underlying Markov chain. When the transition law and hence the stationary distribution is not
known in advance, a primal descent-dual ascent algorithm as in recursion (35) can be implemented with
the knowledge of the sample paths of the underlying Markov chain and the analysis presented in this paper
guarantees convergence to an ǫ-optimal solution for a user specified choice of ǫ.
We outline a few important directions for future work.
(1) For two timescale stochastic approximation schemes with set-valued mean fields, to the best of our
knowledge there are no sufficient conditions for stability in current literature. We believe extensions
of the stability result for single timescale stochastic approximation as in [22, 23], can be made to the
case of two timescale recursions. Another approach to stability could be along the lines of [24].
(2) In many applications the iterates are projected at each time step and are ensured to remain within a
compact, convex set. Such projections often arise due to inherent need of the application or is used
to ensure stability. Such projected schemes have a tendency to introduce spurious equilibrium points
at the boundary of the feasible set. Further complications arise due to the presence of Markov noise
terms since the projection map is most of the time not differentiable but only directional derivatives
are known to exist. Such projected stochastic approximation schemes for single-valued case without
Markov noise component are analyzed in [25] and should serve as a basis for analyzing more general
frameworks with projection.
(3) In some applications arising in reinforcement learning, the noise terms are not Markov by themselves,
but their lack of Markov property comes through the dependence on a control sequence. Under such
controlled Markov noise assumption, two timescale stochastic approximation scheme has been analyzed in [8] but with single-valued, Lipschitz continuous drift functions. Extending the analysis
presented in this paper to the case with set-valued drfit function and controlled Markov noise assumption is straightforward and requires no major change in the overall flow of the analysis. This
extension allows one to analyse the asymptotic behavior of a larger class of reinforcement learning
algorithms (see [26]).
(4) Several other applications, such as two timescale controlled stochastic approximation, two timescale
approximate drift problem also can be analyzed with the help of the results presented in this paper
(see [10, ch. 5.3] for definitions of the above).
A
Proof of Lemma 5.5
Fix ω ∈ Ω1 , l ≥ 1 and T > 0. We prove the claim along the sequence {ts (n)}n≥1 from which the claim of
Lemma 5.5 easily follows.
37
Fix n ≥ 0. Let τ 2 (n, T ) := min {m > n : ts (m) ≥ ts (n) + T }. Let q ∈ [0, T ]. Then, there exists k such
that ts (n) + q ∈ [ts (k), ts (k + 1)) and n ≤ k ≤ τ 2 (n, T ) − 1. By definition of ȳ(·) and ỹ (l) (·; ts (n)), we have
that, ȳ(ts (n) + q) = αyk + (1 − α)yk+1 and ỹ (l) (q; ts (n)) = αỹ (l) (ts (k) − ts (n); ts (n)) + (1 − α)ỹ (l) (ts (k +
s
s
(n)−q
(l)
1) − ts (n); ts (n)) where α = t t(k+1)−t
(·; ts (n)) is a solution of the o.d.e. (28), we have
s (k+1)−ts (k) . Since ỹ
P
(l)
(l)
(2)
k−1
that, for every k ≥ n, ỹ (l) (ts (k) − ts (n); ts (n)) = yn + j=n b(j)h2 (xj , yj , sj , uj ) and by Lemma 5.3, we
Pk−1
Pk−1
(2)
(l)
(l)
(2)
have that, for every k ≥ n, yk = ȳ(ts (k)) = yn + j=n b(j)h2 (xj , yj , sj , uj ) + j=n b(j)mj+1 . Thus,
kȳ(ts (n) + q) − ỹ (l) (q; ts (n))k ≤ kα
b(j)mj+1 + (1 − α)
k−1
X
b(j)mj+1 k + (1 − α)k
(2)
(2)
sup
k
X
(2)
b(j)mj+1 k
j=n
j=n
≤
(2)
b(j)mj+1 k
j=n
j=n
≤ αk
k
X
k−1
X
k
n≤k≤τ (n,T )
k
X
(2)
b(j)mj+1 k.
j=n
Since the r.h.s. of the above inequality is independent of q ∈ [0, T ], we have,
Pk
(2)
sup0≤q≤T kȳ(ts (n) + q) − ỹ (l) (q; ts (n))k ≤ supn≤k≤τ 2 (n,T ) k j=n b(j)mj+1 k. Therefore,
Pk
(2)
limn→∞ sup0≤q≤T kȳ(ts (n) + q) − ỹ (l) (q; ts (n))k ≤ limn→∞ supn≤k≤τ 2 (n,T ) k j=n b(j)mj+1 k. Now the
claim follows follows from assumption (A7).
B
Proof of Lemma 5.6
Fix l ≥ 1, ω ∈ Ω1 . By assumption (A8), we know that there exists r > 0 such that supn≥0 (kxn k+kynk) ≤ r
and hence supt≥0 ỹ (l) (0; t) = supt≥0 ȳ(t) ≤ r.
For any t ≥ 0, let [t] := max {n ≥ 0 : ts (n) ≤ t}. For every t ≥ 0 and q1 , q2 ∈ [0, ∞) (w.l.o.g. assume
q1 < q2 ) we have,
Z q2
(l)
(l)
(2)
(l)
(l)
h2 (x[t+q] , y[t+q] , s[t+q] , u[t+q] )dqk
kỹ (q1 ; t) − ỹ (q2 ; t)k = k
≤
≤
Z
q1
q2
q
Z 1q2
(l)
(2)
(l)
kh2 (x[t+q] , y[t+q] s[t+q] , u[t+q] )kdq
K (l) (1 + kx[t+q] k + ky[t+q] k)dq
q1
≤ C (l) (q2 − q1 ),
where C (l) := K (l) (1 + r) and r > 0 is such that, supn≥0 (kxn k + kyn k) ≤ r. Thus ỹ (l) (·; t)
equicontinuous family. Now the claim follows from Arzella-Ascoli theorem.
t≥0
is an
References
[1] V. S. Borkar, “Stochastic approximation with two time scales,” Systems & Control Letters, vol. 29,
no. 5, pp. 291–294, 1997.
[2] M. Benaim, “A dynamical system approach to stochastic approximations,” SIAM Journal on Control
and Optimization, vol. 34, no. 2, pp. 437–472, 1996.
[3] D. P. Bertsekas, Convex optimization theory.
Athena Scientific Belmont, 2009.
[4] A. Nedić and A. Ozdaglar, “Subgradient methods for saddle-point problems,” Journal of optimization
theory and applications, vol. 142, no. 1, pp. 205–228, 2009.
[5] M. Benzi, G. H. Golub, and J. Liesen, “Numerical solution of saddle point problems,” Acta numerica,
vol. 14, pp. 1–137, 2005.
[6] J. C. Spall, “Multivariate stochastic approximation using a simultaneous perturbation gradient approximation,” IEEE transactions on automatic control, vol. 37, no. 3, pp. 332–341, 1992.
38
[7] P. Milgrom and I. Segal, “Envelope theorems for arbitrary choice sets,” Econometrica, vol. 70, no. 2,
pp. 583–601, 2002. [Online]. Available: http://dx.doi.org/10.1111/1468-0262.00296
[8] P. Karmakar and S. Bhatnagar, “Two timescale stochastic approximation with controlled markov
noise and off-policy temporal difference learning,” arXiv preprint arXiv:1503.09105, 2015.
[9] A. Ramaswamy and S. Bhatnagar, “Stochastic recursive inclusion in two timescales with an application to the lagrangian dual problem,” arXiv preprint arXiv:1502.01956, 2015.
[10] V. S. Borkar, Stochastic approximation : a dynamical systems viewpoint.
Press, 2008.
Cambridge University
[11] V. Yaji and S. Bhatnagar, “Stochastic recursive inclusions with non-additive iterate-dependent markov
noise,” arXiv preprint arXiv:1607.04735, 2016.
[12] M. Benaı̈m, J. Hofbauer, and S. Sorin, “Stochastic approximations and differential inclusions,” SIAM
Journal on Control and Optimization, vol. 44, no. 1, pp. 328–348, 2005.
[13] V. S. Borkar, “Stochastic approximation with ‘controlled Markov’ noise,” Systems & control letters,
vol. 55, no. 2, pp. 139–145, 2006.
[14] J.-P. Aubin and A. Cellina, Differential inclusions: set-valued maps and viability theory.
Science & Business Media, 2012, vol. 264.
Springer
[15] S. Li, Y. Ogura, and V. Kreinovich, Limit theorems and applications of set-valued and fuzzy set-valued
random variables. Springer Science & Business Media, 2013, vol. 43.
[16] K. R. Parthasarathy, Probability measures on metric spaces. American Mathematical Soc., 1967, vol.
352.
[17] V. S. Borkar, Probability theory: an advanced course. Springer Science & Business Media, 2012.
[18] S. P. Meyn and R. L. Tweedie, Markov chains and stochastic stability. Springer Science & Business
Media, 2012.
[19] V. S. Borkar, “Optimal control of diffusion processes,” in Pitman Research Notes in Math., 203. 36
Borkar V.(2005):“Controlled diffusion processes”, Probability surveys. Citeseer, 1989.
[20] S. Kumaresan, Topology of metric spaces.
Alpha Science Int’l Ltd., 2005.
[21] W.
Rudin,
Principles
of
Mathematical
Analysis,
in
pure
and
applied
mathematics.
McGraw-Hill,
https://books.google.co.in/books?id=kwqzPAAACAAJ
ser.
1976.
International
series
[Online].
Available:
[22] V. S. Borkar and S. P. Meyn, “The ODE method for convergence of stochastic approximation and
reinforcement learning,” SIAM Journal on Control and Optimization, vol. 38, no. 2, pp. 447–469,
2000.
[23] A. Ramaswamy and S. Bhatnagar, “A generalization of the Borkar-Meyn theorem for stochastic
recursive inclusions,” arXiv preprint arXiv:1502.01953, 2015.
[24] C. Andrieu, É. Moulines, and P. Priouret, “Stability of stochastic approximation under verifiable
conditions,” SIAM Journal on control and optimization, vol. 44, no. 1, pp. 283–312, 2005.
[25] A. Nagurney and D. Zhang, Projected dynamical systems and variational inequalities with applications.
Springer Science & Business Media, 2012, vol. 2.
[26] S. Perkins, D. S. Leslie et al., “Asynchronous stochastic approximation with differential inclusions,”
Stochastic Systems, vol. 2, no. 2, pp. 409–446, 2012.
39
| 3cs.SY
|
arXiv:1804.00323v1 [math.GR] 1 Apr 2018
JORDAN PROPERTY FOR LIE GROUPS AND
AUTOMORPHISM GROUPS OF COMPLEX SPACES
VLADIMIR L. POPOV∗
Abstract. We prove that all connected finite-dimensional real Lie groups are Jordan. This implies that all algebraic groups (not necessarily
affine) over fields of characteristic zero and some transformation groups
of complex spaces and Riemannian manifods are Jordan.
1. Introduction. We recall the definition introduced in [Po 2011, Def. 2.1]:
Definition 1. Given a group G, put
JG := sup min [F : A],
A
F
where F runs over all finite subgroups of G and A runs over all normal
abelian subgroups of F . If JG 6= ∞, then G is called a Jordan group and JG
is called the Jordan constant of G. In this case, we also say that G enjoys
the Jordan property.
Informally, the Jordan property of G means that all finite subgroups of G
are “almost abelian” in the sense that they are extensions of abelian groups
by groups taken from only a finite list. Definition 1 is inspired by the classical
theorem of Jordan [Jo 1878] claiming that JGLn (ℓ) 6= ∞ holds for every n and
every field ℓ of characteristic zero. If ℓ is algebraically closed, then, for every
fixed n, the constant JGLn (ℓ) is independent of ℓ, so we denote it simply by
J(n). It has been computed in [Co2007]; in particular,
J(n) = (n + 1)! for all n > 71 and n = 63, 65, 67, 69.
For more examples of Jordan groups see [Po 20141 ].
Below variety means algebraic variety over a fixed algebraically closed
field k of characteristic zero; in particular, any algebraic group is defined
over k. If G is either an algebraic group or a topological group, G0 denotes
the identity component of G.
After being posed seven years ago in [Po 2011, Sect. 2] (see also [Po 20141 ,
Sect. 2]), the following problem was explored by a number of researchers (see
the most recent brief survey and references in [PS 2017, Sect. 1]):
Problem. Describe varieties X for which the group Aut(X) is Jordan.
∗
This work is supported by the RSF under a grant 14-50-00005.
1
2
VLADIMIR L. POPOV
At present (April 2018) it is still unknown whether there are varieties X
such that the group Aut(X) is non-Jordan (note that complex manifolds
whose automorphism groups are non-Jordan do exist, see below Remark 2).
On the other hand, by now for many types of varieties X it is shown that the
group Aut(X) is Jordan. In particular, S. Meng and D.-Q. Zhang recently
proved the following
Theorem 1 ([MZ 2015, Thm. 1.6]). For every projective variety X, the
group Aut(X) is Jordan.
Given a variety X, we denote by Aut(X)0 the identity component of
Aut(X) in the sense of [Ra 1964]; see also [Po 20142 ]. By [Ra 1964, Cor. 1],
if X is complete, then Aut(X)0 is a connected (not necessarily affine) algebraic group. Jordan’s theorem cited above implies the claim that every affine
algebraic group is Jordan; see [Po 20141 , Thm. 2]. The key ingredient of the
proof of Theorem 1 given in [MZ 2015] is the proof that the extension of
this claim to all (i.e., not necessarily affine) algebraic groups holds true. The
latter proof is rather involved.
In the present note we obtain, with a very short proof, a general result,
from which the above-mentioned extension immediately follows (see Theorem 4 below). Namely, we prove that every finite-dimensional connected real
Lie group is Jordan (the more precise and general statements are formulated
in Theorems 2, 3, and Corollary 3 below). Then in Sections 5–7 we apply
this to showing that some transformation groups of complex spaces and
Riemannian manifolds are Jordan (see Theorems 5, 7, 8, 9 below).
The question of whether the Lie groups are Jordan was posed to me
by A. M. Vershik (see [Po 20143 , 95:20]) whom I thank. I am grateful to
Yu. G. Zarhin for the valuable comments.
2. Lie groups. We now explore the Jordan property for finite-dimensional
real Lie groups G. Note that non-Jordan groups of this type do exist, because
every discrete group is a 0-dimensional real Lie group and there are nonJordan discrete groups (see [Po 20141 , 1.2.5]). Therefore, the Jordan property of G can be expected only under some constraint on the component
group G/G0 .
To formulate this restriction we recall the following definition introduced
in [Po 2011, Def. 2.9]:
Definition 2. Given a group H, put
bH := sup |F |,
F
where F runs over all finite subgroups of H. If bH 6= ∞, then the group H
is called bounded.
In particular, every finite group H is bounded and bH = |H|.
In Theorems 2, 3 and Corollary 4 below, we consider the class of finitedimensional real Lie groups G whose component group G/G0 is bounded.
JORDAN PROPERTY FOR LIE GROUPS AND COMPLEX SPACES
3
Note that every compact Lie group K belongs to this class, because K/K 0
is finite.
Theorem 2. Let G be a finite-dimensional real Lie group whose component
group G/G0 is bounded. Then G is Jordan.
Proof. By [Po 2011, Lem. 2.11] (or [Po 20141 , Thm. 5]), we may (and shall)
assume that G is connected. This assumption implies the existence of a
compact Lie subgroup K of G such that every compact subgroup of G is
conjugate to that of K (see, e.g., [Ho 1965, Chap. XV, Thm. 3.1(iii)]). In
particular, every finite subgroup of G is conjugate to that of K. This and
Definition 1 show that G is Jordan if and only if K is, and if they are, then
JG = JK .
(1)
Being compact, the group K admits a faithful finite-dimensional representation, i.e., is isomorphic to a subgroup of GLm (R) for some m (see,
e.g., [OV 1990, Chap. 5, §2, Thm. 10]). Since the latter group is Jordan, K
is Jordan as well (see [Po 20141 , Thm. 3(i)]. This completes the proof.
Corollary 1. For every finite-dimensional real Lie group G whose component group G/G0 is bounded, the set of isomorphism classes of all finite
simple subgroups of G is finite.
We now dwell on estimating the Jordan constants of Lie groups whose
component group is finite, with a view of proving that the class of such
groups enjoys a property stronger than that of all its members to be Jordan
(see Corollary 3 below). Seeking only this goal, we did not seek to improve
the estimates obtained.
Lemma 1. Let S be a simply connected simple affine algebraic group. Then
the minimum rdim S of dimensions of faithful linear algebraic representations of S is given by the following table:
type of
S
rdim S
Aℓ
ℓ>1
ℓ+1
Bℓ
ℓ>2
2ℓ
Cℓ
ℓ>2
2ℓ
Dℓ
ℓ>3, ℓ odd
2ℓ−1
Dℓ
ℓ>4, ℓ even
2ℓ + 2ℓ−1
E6 E7
E8
F4 G 2
27 56 248 26
7
Remark 1. In the proof of Lemma 1 below, a faithful representation of S
of dimension rdim S is explicitly specified for each type of S.
Proof of Lemma 1. By Lefschetz’s principle (see, e.g., [Si 1968, VI.6]), we
may (and shall) assume that k is C. We fix a maximal torus T of S. Let
∨
α1 , . . . , αℓ ∈ (Lie T )∗ , ̟1 , . . . , ̟ℓ ∈ (Lie T )∗ , and α∨
1 , . . . , αℓ ∈ Lie T be
respectively the system of simple roots, fundamental weights, and simple
coroots of Lie T with respect to a fixed Borel subalgebra of Lie S containing
Lie T ; we number them as in [OV 1990].
The center Z of S is a finite subgroup of T . Fix a subset Ze of Lie T whose
image under the exponential map Lie T → T is the set of all nonidentity
elements of Z.
4
VLADIMIR L. POPOV
For every dominant weight λ ∈ (Lie T )∗ , let R(λ) be an irreducible representation of Lie S with the highest weight λ. The dimension of R(̟i ) for
every i is specified in [OV 1990, Ref. Chap., §2, Table 5, pp. 299–305]. Note
that Weyl’s dimension formula implies
Pℓ
Pℓ
if λi > µi for every i.
(2)
dim R
i=1 µi ̟i
i=1 λi ̟i > dim R
Since S is simply connected, R(λ) is the differential of a finite-dimensional
linear algebraic representation R(λ) of S. Since S is simple,
for every finite
L
set D of nonzero dominant weights and R(D) :=
R(λ),
we have
λ∈D
ker R(D) ⊆ Z. Hence
e there is λ ∈ D with λ(x) ∈
R(D) is faithful ⇐⇒ for every x ∈ Z
/ Z. (3)
As is well known, dim R(̟1 ) is the minimum of dimensions of nonzero
finite-dimensional algebraic representations of S (see [OV 1990, pp. 299–
305]).
If S is of type E8 , F4 , or G2 , then Z is trivial; hence in this case R(̟1 ) is
faithful and therefore we have the equality
rdim S = dim R(̟1 ),
(4)
which proves the claim of Lemma 1 for these types.
If S is of type Aℓ or Cℓ , then S is respectively SLℓ+1 and Sp2ℓ . Since for
these groups R(̟1 ) is the tautological faithful representation, in this case
(4) holds as well, which proves the claim of Lemma 1 for these types.
e taken from [OV 1990,
For the other types, we apply (3) to the set Z
Ref. Chap., §2, Table 3, p. 298]. Below is used that for any λi , µi ∈ k,
P
P
Pℓ
the value of ℓi=1 λi ̟i ∈ (Lie T )∗ in ℓi=1 µi α∨
i ∈ Lie T is
i=1 λi µi . (5)
e consists of only one element ζ := (α∨ + α∨ +
If S is of type E7 , then Z
1
3
By (5), we have ̟1 (ζ) = 1/2 ∈
/ Z, so R(̟1 ) is faithful. Therefore, in
this case again (4) holds, which proves the claim of Lemma 1 for this type.
∨
∨
If S is of type E6 , then Ze consists of two elements ζ := (α∨
1 − α2 + α4 −
∨
α5 )/3 and 2ζ. Since ̟1 (z) = 1/3 ∈
/ Z, ̟1 (2z) = 2/3 ∈
/ Z, in this case again
R(̟1 ) is faithful; whence (4) holds. This proves the claim of Lemma 1 for
this type.
/2. This and (3),
If S is of type Bℓ , then Ze consists of only one element α∨
Pℓ
(5) imply that R(D) is faithful if and only if D contains ℓi=1 λi ̟i with odd
λℓ . Using (2), from this we infer that R(̟ℓ ) is the faithful representation
of minimal dimension. Hence rdim S = dim R(̟ℓ ). This proves the claim of
Lemma 1 for this type.
If S is of type Dℓ , ℓ > 3, ℓ odd, then Ze consists of three elements
α∨
7 )/2.
∨
∨
∨
∨
ζ := (α∨
1 + α3 + · · · + αℓ−2 )/2 + (αℓ−1 − αℓ )/4, 2ζ, 3ζ.
(6)
From (3), (5), (6) we infer that R(D) is faithful if and only if D contains
Pℓ
i=1 λi ̟i such that 4 is coprime to either λℓ−1 or λℓ . This and (2) show that
JORDAN PROPERTY FOR LIE GROUPS AND COMPLEX SPACES
5
R(̟ℓ ) is the faithful representation of minimal dimension. Hence rdim S =
dim R(̟ℓ ), proving the claim of Lemma 1 for this type.
e consists of three elements
If S is of type Dℓ , ℓ > 4, ℓ even, then Z
∨
∨
∨
∨
ζ1 := (α∨
(7)
1 + α3 + · · · + αℓ−1 )/2, ζ2 := (αℓ−1 + αℓ )/2, ζ1 + ζ2 .
Pℓ
Hence if R(D) is faithful, then D contains ı=1 λi ̟i with odd λℓ or λℓ−1
P
and ℓı=1 µi ̟i with odd µi for some odd i 6= ℓ − 1. On the other hand, since
in this case Z is not cyclic, Schur’s lemma implies that |D| > 2. From this it
is not difficult to deduce that R(̟1 ) ⊕ R(̟ℓ ) is the faithful representation
of minimal dimension. Hence rdim S = dim R(̟1 ) + dim R(̟ℓ ) = 2ℓ + 2ℓ−1 .
This completes the proof of Lemma 1.
Corollary 2. Every simply connected simple affine algebraic group of rank ℓ
admits a faithful linear algebraic representation of dimension at most 2ℓ +10.
Proof. Clearly if an algebraic group admits a faithful linear algebraic representation, then it admits a faithful linear algebraic representation of any
bigger dimension. In view of this, the claim follows from the inequality
rdim S 6 2ℓ + 10, which, in turn, follows from Lemma 1: indeed, the latter
shows that rdim S 6 2ℓ if the type of S differs from F4 and G2 , and that
rdim S = 2ℓ + 10 and 2ℓ + 3 respectively for the types F4 and G2 .
Theorem 3. Let G be an n-dimensional real Lie group whose component
group G/G0 is bounded. Then
b
(8)
JG 6 bG/G0 J n(2n + 10) G/G0 .
Proof. By [Po 2011, Lem. 2.11] (or [Po 20141 , Thm. 5]), we may (and shall)
assume that G is connected; in particular,
bG/G0 = 1.
(9)
We use the notation of the proof of Theorem 2. Since G is connected, K is
connected, too; see [Ho 1965, Chap. XV, Thm. 3.1(ii)]. Hence (see [Bo1982,
§1, Prop. 4]) there are
(i) the compact simply connected simple Lie groups K1 , . . . , Kd ;
(ii) a compact torus S;
(iii) a group epimorphism with finite kernel
e := K1 × · · · × Kd × S → K.
π: K
(10)
By [Po 20141 , Thm. 3(ii)], from (iii) we infer that
JK 6 JKe .
(11)
Every Ki is a real form of the corresponding simply connected simple
complex affine algebraic group. The rank ℓi of the latter is equal to that
of Ki . By Corollary 2 we then conclude that Ki admits an embedding in
e = dim K 6 n, this in turn implies that Ki
GL2ℓi +10 (C). Since ℓi 6 dim K
admits an embedding in GL2n +10 (C). Clearly, S admits an embedding in
e also in GL2n +10 (C).
GLdim S (C), and therefore, in view of dim S 6 dim K,
6
VLADIMIR L. POPOV
e (see (10)) show that K
e admits an embedding in
This and the definition of K
the direct product of d+1 copies of GL2n +10 (C), hence in GL(d+1)(2n +10) (C).
e from this we infer
In turn, since, in view of (10), we have d + 1 6 dim K,
e
that K admits an embedding in GLn(2n +10) (C); whence,
(12)
JKe 6 J n(2n + 10) .
Now, putting (1), (11), (12), (9) together, we complete the proof.
Recall from [MZ 2015] the following
Definition 3. A family F of groups is called uniformly Jordan if every
group in F is Jordan and there is an integer JF such that JG 6 JF for every
G ∈ F.
Corollary 3. Fix an integer n > 0. Let Ln be the family of all connected
n-dimensional real Lie groups. Then
(i) the family Ln is uniformly Jordan;
(ii) one can take JLn = J n(2n + 10) .
Proof. This follows from (8) because bG/G0 = 1 for every G ∈ Ln .
Corollary 4. For every integer n > 0, the set of isomorphism classes of
finite simple groups embeddable in n-dimensional connected real Lie groups
is finite.
3. Algebraic groups. We now consider several applications of Theorems
2 and 3. First, we apply them to algebraic groups, answering Question 1.2
in [MZ 2015]:
Theorem 4. Every (not necessarily affine) n-dimensional algebraic group
G over an algebraically closed field k of characteristic 0 is Jordan. Moreover,
[G:G0 ]
.
(13)
JG 6 [G : G0 ]J n(22n+1 + 20)
Proof. In this case, G/G0 is finite. By Lefschetz’s principle, we may (and
shall) assume that k is C. Then G has a structure of 2n-dimensional real
Lie group whose identity component is G0 . The claim then follows from
Theorem 3.
Statement (i) of the next corollary is one of the main results of [MZ 2015]:
Corollary 5. Fix an integer n > 0. Let An be the family of all (not necessarily affine) connected n-dimensional algebraic groups over an algebraically
closed field k of characteristic 0. Then
(i) ([MZ 2015, Thm. 1.3]) the family An is uniformly Jordan;
(ii) one can take JAn = J n(22n+1 + 20) .
Proof. This follows from (13).
JORDAN PROPERTY FOR LIE GROUPS AND COMPLEX SPACES
7
4. Automorphism groups of complex spaces. The next application is
to automorphism groups of complex spaces.
Let C be a (not necessarily reduced) complex space. There exists a topology on Aut(C) with respect to which Aut(C) is a topological group (see
[Ak 1995, 2.1]).
Theorem 5. For every compact complex space C, the group Aut(C)0 is
Jordan.
Proof. By [Ka 1965], the compactness of C implies that Aut(C) is a complex
Lie group. The claim then follows from Theorem 2.
We do not know whether the statement of Theorem 5 remains true if
Aut(C)0 is replaced by Aut(C). By [PS 2017, Thm. 1.5], the answer is affirmative if C is a connected compact two-dimensional complex manifold. By
Theorem 1, it is also affirmative if C is a projective variety. On the other
hand, we recall that by [CPS 2014] there are connected smooth compact
real manifolds whose diffeomorphism groups are non-Jordan (this disproves
Ghys’ conjecture).
Remark 2. There are connected noncompact complex manifolds, whose
automorphism groups are non-Jordan. Indeed, by [Wi 2002], for any countable group Γ, there is a noncompact Riemann surface M such that Aut(M )
is isomorphic to Γ; whence the claim because of the existence of countable
non-Jordan groups (see [Po 20141 , Sect. 1.2.5]).
In actual fact, using the idea exploited earlier in [Po 2015], one can prove
more than is said in Remark 2, showing the existence of connected complex
manifolds with monstrous automorphism groups, namely:
Theorem 6. There is a 3-dimensional simply connected noncompact complex manifold M such that
(i) the group Aut(M ) contains an isomorphic copy of every finitely presentable (in particular, every finite) group;
(ii) every such copy is a discrete transformation group of M acting freely.
Proof. It follows (see, e.g., [Ro 1995, Thm. 12.29]) from Higman’s embedding
theorem [Hi 1961] that there is a universal finitely presented group, i.e., a
finitely presented group U containing as a subgroup an isomorphic copy
of every finitely presented group. In turn, by [ABCKT 1996, Cor. 1.66] the
finite presentability of U implies the existence of a connected 3-dimensional
compact complex manifold B whose fundamental group is isomorphic to
e → B. Then B
e is a simply connected
U . Consider the universal cover π : B
noncompact 3-dimensional complex manifold and the deck transformation
e isomorphic to U , which acts on B
e freely.
group of π is a subgroup of Aut B
e
Hence one can take M = B.
Remark 3. For M from Theorem 6, the group Aut(M ) is non-Jordan,
because for every integer n, there is a finite simple group of order > n
(cf. [Po 20141 , Example 4].
8
VLADIMIR L. POPOV
5. Automorphism groups of hyperbolic complex manifolds. The
next application is to complex manifolds hyperbolic in the sense of Kobayashi
(in particular, to bounded domains in Cn ).
Theorem 7. Fix an integer n > 0. Let Hn be the family of groups Aut(M )0 ,
where M runs over all connected complex manifolds hyperbolic in the sense
of Kobayashi and of complex dimension n. Then
(i) the family Hn is uniformly Jordan;
2
(ii) one can take JHn = J (2n + n2 )(22n+n + 10) ;
(iii) for every point x ∈ M , the Aut(M )-stablizer Aut(M )x of x is Jordan
and JAut(M )x 6 J(n).
Proof. Let M be a connected complex manifolds hyperbolic in the sense
of Kobayashi and of complex dimension n. By [Ko 2005, Thms. 2.1, 2.6],
Aut(M ) is a real Lie group of dimension 6 2n + n2 ; whence (i) and (ii) by
Theorems 2 and 3. By [Ko 2005, Thm. 2.6], the isotropy representation of
Aut(M )x is faithful and its image is isomorphic to a subgroup of the unitary
group U(n); whence (iii).
Remark 4. The group Aut(M )0 in the formulation of Theorem 7 cannot be
replaced by Aut(M ). Indeed, it follows from the construction in [Wi 2002]
that the Riemann surface M in Remark 2 is hyperbolic in the sense of
Kobayashi. Therefore there are connected hyperbolic complex manifolds M
such that the group Aut(M ) is not Jordan.
However, as the next theorem shows, for complex hyperbolic manifolds
M of a special type, the Jordan property holds for the whole Aut(M ) rather
than only for Aut(M )0 .
Theorem 8. For every strongly pseudoconvex bounded domain M with
smooth boundary in Cn , the group Aut(M ) of all biholomorphic transformations of M is Jordan.
Proof. If the Lie group Aut(M ) is compact, then the claim follows from
Theorem 2. If the group Aut(M ) is non-compact, then, by the Rosey–Wong
theorem [Ro 1979], [Wo 1977], the domain M is biholomorphic to the unit
ball Bn in Cn . Since Aut(Bn ) is PU(n, 1) (see [Ak 1995, Sect. 2.7, Prop. 3]),
and the latter Lie group is connected (see [He 1962, Chap. IX, Lem. 4.4]),
the claim then follows from Theorem 2.
Corollary 6. For every strongly pseudoconvex bounded domain M with
smooth boundary in Cn , the set of isomorphism classes of all finite simple
groups of biholomorphic transformations of M is finite.
6. Isometry groups of Riemannian manifolds. The last application is
to isometry groups Iso(M ) of Riemannian manifolds M . They are topological
groups with respect to the compact-open topology [Ko 1995].
JORDAN PROPERTY FOR LIE GROUPS AND COMPLEX SPACES
9
Theorem 9. Fix an integer n > 0. Let Rn be the family of groups Iso(M )0 ,
where M runs over all connected n-dimensional Riemannian manifolds.
Then
(i) the family Rn is uniformly Jordan;
2
(ii) one can take JRn = J (n2 + n)(2(n +n−2)/2 + 5) ;
(iii) for every point x ∈ M , the Iso(M )-stabilizer Iso(M )x of x is Jordan;
(iv) if the manifold M is compact, then the group Iso(M ) is Jordan.
Proof. It is known (see, e.g., [Ko 1995, Chap. II, Thms. 1.2 and 3.1]) that
Iso(M ) is a real Lie group of dimension at most n(n + 1)/2, the group
Iso(M )x is compact for every x, and the group Iso(M ) is compact if the
manifold M is compact. The claims then follows from combining these facts
with Theorems 2 and 3.
Remark 5. The group Aut(M )0 in the formulation of Theorem 9 cannot be
replaced by Aut(M ). Indeed, it follows from the construction in [Wi 2002]
that the Riemann surface M in Remark 2 is a two-dimensional Riemannian
manifold and Aut(M ) = Iso(M ). Therefore there are connected Riemannian
manifolds M such that the group Iso(M ) is not Jordan.
7. Concluding remarks. In view of (1), computing the Jordan constants
of connected real Lie groups is reduced to that of compact such groups.
For instance, the results of [Co2007] may be interpreted as computing the
Jordan constants of all unitary groups:
JUn = J(n) for every n.
This leads to the following natural
Problem. Compute the Jordan constants of all simple compact connected
real Lie groups.
References
[Ak 1995]
D. N. Akhiezer, Lie Group Actions in Complex Analysis, Aspects of Mathematics, Vol. E 27, Vieweg, Braunschweig, 1995.
[ABCKT 1996] J. Amorós, M. Burger, K. Corlette, D. Kotschick, D. Toledo, Fundamental
Groups of Compact Kähler Manifolds, Mathematical Surveys and Monographs, Vol. 44, American Mathematical Society, Providence, RI, 1996.
[Bo1982]
N. Bourbaki, Groupes et Algèbres de Lie, Chap. 9, Groupes de Lie Réels
Compacts, Masson, Paris, 1982.
[Co2007]
M. J. Collins, On Jordan’s theorem for complex linear groups, J. Group
Theory 10 (2007), 411–423.
[CPS 2014]
B. Csikós, L. Pyber, E. Szabó, Diffeomorphism groups of compact 4-manifolds are not always Jordan, arXiv:1411.7524 (2014).
[He 1962]
S. Helgason, Differential Geometry and Symmetric Spaces, Academic
Press, New York, 1962.
[Hi 1961]
G. Higman, Subgroups of finitely presented groups, Proc. R. Soc. London
A 262 (1961), 455–475.
[Ho 1965]
G. Hochschild, The Structure of Lie Groups, Holden-Day, San Francisco,
1965.
10
[Jo 1878]
[Ka 1965]
[Ko 1995]
[Ko 2005]
[MZ 2015]
[OV 1990]
[Po 2011]
[Po 20141 ]
[Po 20142 ]
[Po 20143 ]
[Po 2015]
[PS 2017]
[Ra 1964]
[Ro 1979]
[Ro 1995]
[Si 1968]
[Wi 2002]
[Wo 1977]
VLADIMIR L. POPOV
C. Jordan, Mémoire sur les equations différentielle linéaire à intégrale algébrique, J. Reine Angew. Math. 84 (1878), 89–215.
W. Kaup, Infinitesimale Transformationsgruppen komplexer Räume,
Math. Ann. 160 (1965), 72–92.
S. Kobayashi, Transformation Groups in Differential Geometry, Springer,
Berlin, 1995.
S. Kobayashi, Hyperbolic Manifolds and Holomorphic Mappings. An Introduction, 2nd ed., World Scientific, New Jersey, 2005.
S. Meng, D.-Q. Zhang, Jordan property for non-linear algebraic groups
and projective varieties, to appear in Amer. J. Math., available at https://
preprint.press.jhu.edu/ajm/article/jordan-property-non-linearalgebraic-groups-and-projective-varieties.
A. L. Onishchik, È. B. Vinberg, Lie Groups and Algebraic Groups, Springer Series in Soviet Mathematics, Springer-Verlag, Berlin, 1990.
V. L. Popov, On the Makar-Limanov, Derksen invariants, and finite automorphism groups of algebraic varieties, in: Affine Algebraic Geometry:
The Russell Festschrift, CRM Proceedings and Lecture Notes, Vol. 54,
American Mathematical Society, Providence, RI, 2011, pp. 289–311.
V. L. Popov, Jordan groups and automorphism groups of algebraic varieties, in: Automorphisms in Birational and Affine Geometry (Levico Terme,
Italy, October 2012), Springer Proceedings in Mathematics & Statistics,
Vol. 79, Springer, Heidelberg, 2014, pp. 185–213.
V. L. Popov, On infinite dimensional algebraic transformation groups,
Transform. Groups 19 (2014), no. 2, 549–568.
V. L. Popov, Jordan groups, talk at the St. Petersburg Division of Steklov
Institute of Mathematics, Russian Academy of Sciences, December 18,
2014, videorecord available at http://www.mathnet.ru/php/seminars.
phtml?&presentid=10616&option lang=eng.
V. L. Popov, Finite subgroups of diffeomorphism groups, Proc. Steklov
Inst. Math. 289 (2015), 221–226.
Yu. Prokhorov, C. Shramov, Automorphism groups of compact complex
surfaces, arXiv:1708.03566 (2017).
C. P. Ramanujam, A note on automorphism groups of algebraic varieties,
Math. Annalen 156 (1964), 25–33.
J.-P. Rosay, Sur une caractérisation de la boule parmi les domaines de Cn
par son groupe dautomorphismes, Ann. Inst. Fourier (Grenoble) 29 (1979),
no. 4, 91–97.
J. J. Rotman, An Introduction to the Theory of Groups, 4th ed., Graduate
Texts in Mathematics, Vol. 148, Springer, New York, 1995.
J. H. Silverman, The Arithmetic of Elliptic Curves, 2nd ed., Graduate
Texts in Mathematics, Vol. 106, Springer, Dordrecht, 2009.
J. Winkelmann, Realizing countable groups as automorphism groups of
Riemann surfaces, Documenta Math. 7 (2002), 413–417.
B. Wong, Characterization of the unit ball in Cn by its automorphism
group, Invent. math. 41 (1977), 253–257.
Steklov Mathematical Institute, Russian Academy of Sciences, Gubkina 8,
Moscow, 119991, Russia
E-mail address: [email protected]
| 4math.GR
|
1
Interdependency of Transmission and
Distribution Pricing
Sina Parhizi and Amin Khodaei
Department of Electrical and Computer Engineering
University of Denver
Denver, Colorado
[email protected], [email protected]
Abstract— Distribution markets are among the prospect being
considered for the future of power systems. They would facilitate
integration of distributed energy resources (DERs) and
microgrids via a market mechanism and enable them to monetize
services they can provide. This paper follows the ongoing work in
implementing the distribution market operator (DMO) concept,
and its clearing and settlement procedures, and focuses on
investigating the pricing conducted by the DMO. The
distribution locational marginal prices (D-LMPs) and their
relationship with the transmission system locational marginal
prices (T-LMPs) are subject of this paper. Numerical simulations
on a test distribution system exhibit the benefits and drawbacks
of the proposed DMO pricing processes.
Index Terms—Power distribution, microgrids, power system
economics, electricity markets.
NOMENCLATURE
a
b
Cc
Cu
CΔ
f
g
m
D
PM
PD M
PL
PLmax
DX
PXmax
r
t
ΔP
λ
λD
Pt pos
Elements of the bus-line incidence matrix.
Load benefit.
Customer payments to the DMO.
DMO payment to the ISO.
DMO cost surplus.
Superscript for fixed loads.
Index for bid segments.
Penalty factor for the power deviation.
Index for distribution system buses.
Load demand.
Power transfer from the main grid.
Total assigned power from the main grid.
Line power flow.
Line power flow limit.
The amount of load awarded to each segment of the
bid.
Maximum capacity of each bid segment.
Superscript for responsive loads (proactive
customers).
Index for hours.
Power transfer deviation.
Transmission locational marginal price (T-LMP).
Distribution locational marginal price (D-LMP).
Variables used to linearize absolute value.
Pt neg
H
Variables used to linearize absolute value.
I. INTRODUCTION
IGH penetration of proactive customers that utilize
variety of distributed energy resources (DER), electric
vehicles, and flexible loads are frequently mentioned as the
characteristics of future power grids [1]–[7]. This growing
trend is advocated in order to handle challenges such as
growing demand and boost efficiency in the grid operation
while meeting the standards put in place by various
environmental regulations [8], [9]. Proactive customers in the
distribution system are not completely relying on the utility
grid to provide their demand; they can self-sustain, in many
cases of DER deployment, and even actively participate in the
electricity markets and provide several valuable services. In
order for proactive customers to be able to realize their full
capabilities, to monetize services they provide and to play a
role in system pricing and clearing decision-making processes,
it is necessary to modernize the existing distribution system
operation and enable these entities to participate in the
electricity market [10].
The direct control and dispatch of the proactive customers
by the independent system operator (ISO) would create
several problems such as potential violation of distribution
companies’ responsibilities and also creating complexities in
the market optimization as the participation of proactive
customers grows [11], [12]; hence, several proposals have
been offered trying to define an entity that establishes an
electricity market and facilitates market participation at the
distribution level. Distributed System Platform Provider
(DSPP) is introduced in New York via the Reforming the
Energy Vision program as one of the early efforts in this
direction [13]. DSPP is proposed to operate in accord with the
electric utility and be in charge of the market operations. It
would coordinate with the ISO, customers, and market
participants. According to this proposal, distribution system
platform (DSP) is a set of functions provided by the utilities to
enable broad market participation. Similar to this effort, the
Distribution System Operator (DSO) is introduced in
California as an entity to be in charge of providing reliable
distribution services to the customers and further ensuring
predictability for the ISO. It wouldbe coordinating the
physical transactions in the transmission-distribution interface
2
but not necessarily the financial aspects of those transactions
that
[12], [14]–[16] propose can still be among the
responsibilities of ISO. The role of the DSO introduced in
[17] is similar to that of an ISO but for the distribution system.
The DSO would not only be responsible for the reliable
operation of the distribution system but also for providing
demand response. The DSO would have transactions with the
wholesale market at the substation level on one side and
proactive customers on the other side. Depending on the
extent of the ISO’s responsibility in dispatching resources in
the distribution system, there would be different levels of the
DSO autonomy in operating the distribution system and the
degree of the ISO’s control over it. The DSO can only act as
an aggregator and deliver ISO’s dispatch commands or
operate a retail market at the distribution level, or something
between these two extremes. In [18] the prospective roles for
the DSOs are listed including “managing multidirectional
flows, organizing auctions, and offering other incentives” in
order to minimize the operation costs of the DSO. In [19]
control strategies to coordinate several microgrids within a
distribution system via a distribution network operator (DNO)
is proposed. The DNO can trade energy with the microgrids
and the high-voltage transmission system. The problem is
formulated as a two-level optimization where microgrids
optimize their energy cost at the lower level and the DNO
ensures operational constraints at the higher level of
optimization. In [20] it is asserted that the utilities should
transition their distribution system operation responsibilities to
an independent distribution system operator (IDSO) while
owning distribution assets. It is argued that operation of the
distribution system by the utilities poses a conflict of interest,
and as utilities tend to expand their assets it would reduce
greater market participation of the proactive customers. The
proposed IDSO would be in charge of the system reliability,
provide market mechanisms, and optimally schedule the
distribution level resources. It would make system
management easier for the utilities and determine the value of
services provided by proactive customers more objectively.
Most of these proposals, however, are conceptual and lack
a detailed rigorous analysis of distribution market operations.
It is necessary to clearly define the roles of different parties
involved in the distribution system undergoing market-based
evolution and further investigate the detailed market
processes. Among the concerns that needs to be addressed is
the structure of the distribution market operation, the way it is
settled [21], and how the procedures for distribution market
clearing and settlement are established. The study in [22]
points out the necessity of the DSOs to be capable of
providing settlements with the ISO for any resource in the
distribution system. Studies in [23], [24], furthermore present
a marginal pricing method for the DSO considering the
congestion problem in the distribution system, and optimizing
the social welfare in a system with high penetration of electric
vehicles.
The distribution market operator (DMO) is a rather similar
entity discussed by authors in [25][26]. The DMO function
focused in this work is to facilitate the establishment of
market mechanisms in distribution systems and it will be an
interface interacting with both the ISO and proactive
customers to enable participation of customers in the
wholesale market. The DMO receives the demand bids from
customers in the distribution system, aggregates them and
submits a single aggregated bid to the ISO. After market
clearing by the ISO, the DMO divides the assigned power
awarded to it between the participated customers. The DMO
can be part of the electric utility company or be formed as a
separate entity. In either case, it should be an independent
operator so as to guarantee the fairness in the market
operation. The DMO manages the financial transactions and
electric distribution company (EDC) carries out the physical
transactions.
The implementation of a distribution market and
establishment of the DMO offers several advantages for
customers and the system as a whole: (i) Under distribution
markets, the proactive customers’ demand is set by the DMO
and known with certainty on a day-ahead basis, which would
enable an efficient control of the peak demand, increase
operational reliability, and improve efficiency; (ii) The
proactive customers can participate in the electricity market as
a player and exchange power with the utility grid and other
customers. The DMO would facilitate market participation
and coordinate the proactive customers’ interactions with the
utility grid to minimize the associated operational risks and
uncertainties; (iii) There will be a considerable reduction in
the required communication in the system as the proactive
customers and the ISO only need to communicate with the
DMOs. Considering the listed advantages, and many more
that will be obtained once these deployments are more
widespread, distribution markets can be considered as both
beneficial and necessary components in modern power grids
which will help accommodate a large penetration of proactive
customers. In terms of the distribution market clearing and
settlement, the models proposed in [25], [26] were
investigated in detail in [27], in which constant and variable
power clearing schemes were studied. This paper focuses on
how the locational marginal prices in the transmission side of
the DMO (i.e., T-LMP) are reflected in the distribution system
locational marginal prices (i.e., D-LMP). Furthermore,
microgrids will be considered for studies as a representative of
proactive customers. Microgrids models, however, can be
simplified to model any other type of the proactive customer,
such as prosumers or responsive consumers.
The rest of the paper is organized as follows. The
formulation for the proposed DMO market clearing and
settlement is demonstrated in section III, numerical
simulations are presented in section IV, and the paper is
concluded in section V.
II. DISTRIBUTION MARKET MODEL
The DMO collects demand bids from the microgrids in the
distribution system, creates an aggregated bid, and submits
this bid to the ISO. A typical microgrid bid is shown in Fig. 1.
3
The ISO collects the demand bids (from the DMOs as well as
curtailment service provides) and generation bids (from
GENCOs), and determines the generation and load schedule.
Once the DMO is notified of the ISO’s clearing process
decisions, it determines the amount of power generation,
transfer and demands within the distribution system, and
settles the prices and costs among various participants in the
market, with the objective of ensuring an optimal operation
and a fair settlement.
Price
($/MWh)
bi1
bi2
bi3
DX1gmax
DX 2max
g
max
DX 3g
Load (MW)
Fig. 1 Demand bid curve for customer at bus i with a three-segment bid
III. DISTRIBUTION MARKET CLEARING AND SETTLEMENT
The DMO’s objective is to maximize the distribution
system social welfare (1), i.e., load benefits minus generation
cost (which is the cost of assigned power from the main grid).
This paper proposes to add the last term in the objective to
penalize violations in the assigned power.
max bmg DX mgt t Pt M | Pt M |
t
t m g
(1)
t
The power assigned to the DMO by the ISO is determined
via the wholesale market clearing, hence it is constant. The TLMP is also determined by the ISO. The penalty coefficient μ
is multiplied with the deviation to ensure that the assigned
power will be followed in the distribution network. The
objective is subject to distribution network and microgrids
prevailing constraints (2)-(6):
a lm PLlt Dmt
t , m (2)
a l 0 PLlt Pt M 0
t
Dmt DX mgt Dmtf
t , m (4)
l
(3)
l
g
max
0 DX mgt DX mg
t , m, g (5)
PLmax
PLlt PLmax
l
l
t , l
(6)
Pt
M
t
t
(7)
neg
t
(8)
t
(9)
Pt PD
M
M
Pt Pt
M
pos
| Pt | Pt
M
Pt
pos
Pt
neg
Pt
neg
0
t
(10)
Pt
pos
0
t
(11)
The nodal power balance is ensured in (2) where the power
injected to each bus from connected lines is equal to total bus
load. The power balance at transmission-distribution interface
is ensured by (3) in which the power transferred by the main
grid is distributed among lines connected to this bus (here the
bus number 0). The load of passive customers will be
constant, while that of proactive customers is variable and
defined by the associated segments (4). The scheduled load of
microgrids is determined based on the scheduled power
consumption in each bid segment, in which each segment is
limited by its associated maximum capacity (5). Line power
flows are limited by the line capacity limits (6). The added
penalty in the objective function, which is represented as the
absolute value of the deviation makes the problem nonlinear.
In order to linearize this term and to ensure a linear
programming problem, (8)-(11) are used. Pt neg and Pt pos are
two non-negative variables used to model the absolute value.
If the variable inside the absolute value is positive, Pt neg would
be equal to zero and when the value is negative Pt pos would be
equal to zero. This is guaranteed to happen since the problem
is formulated as a linear programming minimization solved by
the Simplex method.
The D-LMP in each bus is calculated as the dual variable
of the power balance equation in that bus (2), i.e., as a
byproduct of the proposed clearing problem. The relationship
between the D-LMP and the T-LMP depends on the value of
the penalty factor μ. In this paper, the distribution market
clearing is conducted in two ways based on the penalty factor:
grid-following clearing and grid-independent clearing, as
discussed further in the following:
A. Grid-following clearing
When μ=0 in the proposed formulation, the D-LMP at bus
0 of the distribution (point of connection to the transmission
network) will be equal to the T-LMP. In this case the DMO is
permitted to import power from the utility grid more/less than
the power assigned to it by the ISO, as there would be no
penalty. This results in the T-LMP to be reflected in the DLMPs within the distribution system. At down-stream buses,
however, D-LMPs will be determined based on the T-LMP,
marginal cost of dispatchable units, and possible distribution
line congestions.
B. Grid-independent clearing
As the amount of μ is increased, the DMO seeks to
minimize the deviation of the scheduled power with the
assigned power transfer set by the ISO. In this case the
dependency of D-LMPs to the T-LMP will be lowered. For
the very large values of the penalty factor, there would be no
violation of the power transfer schedule determined by the
ISO, while the D-LMPs will be merely functions of the
dispatchable units’ marginal price and possible distribution
line congestions.
4
C. Market Settlement
Using D-LMPs, obtained from either methods, the market
can be settled, i.e., the payments from customers and the
payments to the utility can be determined. The payment of
each customer is calculated as the D-LMP times the associated
load. The total customer payments is the summation of all
payments (12), in which Dmt includes both consumers and
microgrids. The payment to the utility is calculated as the TLMP times the assigned power by the ISO (13). Since losses
are ignored in the proposed market clearing model, the sum of
the distribution loads will be equal to the total power assigned
by the ISO (14).
D
Cc mt
Dmt
(12)
Cu t Pt M
(13)
t m
Case 1: In this case μ is assumed to be 0, and the distribution
market clearing results for a variety of T-LMPs is determined.
The results of this case are shown in Fig. 3, which illustrates
the effect different values of the T-LMP on the marginal price
of each bus in the distribution system. A scaling factor is used
to change T-LMPs. As the scaling factor increases the power
is to be purchased at a higher rate, resulting in a lower power
transfer from the ISO and more local generation. This results
in lower congestion as the prices of the all busses tend to be
equal at higher values of the scaling factor. At lower T-LMP
values, D-LMPs follow the T-LMPs and can also impact the
grid prices as they respond to the price variations by
modifying their power injections. At lower scaling factors,
congestion at line 3-8 results in a rise of D-LMPs in buses 612. These buses have the same D-LMPs as the lines in the
downstream of the feeder do not become constrained.
t
Pt M Dmt
t
(14)
m
Considering the payments, the DMO cost surplus can be
calculated as the difference between the two calculated
payments as in (15):
D
C Cc Cu (mt
t ) Dmt
(15)
t m
The obtained CΔ can be negative, positive, or zero. The cost
settlement is one of the challenges facing DMOs as they
operate radial networks, as opposed to the wholesale power
system operated by ISOs, and they should guarantee a fair
market participation by customers at different locations across
the feeders. This issue will be the topic of a future research by
authors.
IV. NUMERICAL RESULTS
The IEEE 13-bus test system [28] in used to investigate the
viability and the merits of the proposed processes. Fig. 2
depicts this system in which microgrids are located at buses 2,
3, 5-7, and 10-13. Each customer submits a four-segment
power demand bid of maximum 10 MW. A large capacity for
distribution lines is considered, however it is assumed that
lines 3-8 and 4-5 have smaller capacities and then subject to
potential congestions. Various cases have been studied
considering the various values for parameters in (1).
Fig. 3 Daily average LMP at each bus for different T-LMP values (using a
scaling factor)
Case 2: In this case the second term in (1) is assumed to be 0,
i.e., the T-LMP is negligible, and μ is varied, so the
independent operation of the distribution market can be
analyzed. The results of this case are shown in Fig. 4, which
illustrates the effect of raising μ on D-LMPs. As μ increases,
the DMO seeks to minimize the deviation from the assigned
power transfer. The prices, however, tend to approach the
prices when the scheduled power is used without any option
to deviate. When μ approaches infinity, the D-LMPs are
functions of marginal prices of the dispatchable units and
become independent of the T-LMPs. This would result in the
settlement costs of the distribution system be higher, lower, or
equal to the payments to the ISO depending on the marginal
costs.
T-LMP
1
2
3
4
6
7
8
9
11
12
Fig. 2 IEEE 13-bus standard test system.
5
10
Fig. 4 Daily average D-LMP as a function of μ.
5
Case 3: In this case the T-LMP and μ are both nonzero. The
results of this case are shown in Figs. 5 and 6. In Fig. 5, TLMP scaling factor varies between 0.1 and 0.9 while μ is kept
at 1. The corresponding difference between the customers’
payment to the DMO and payment to the ISO is depicted in
Fig. 6 (a deficit for the DMO). It means that while the DMO
tries to minimize the deviation of the power transfer with
respect to the scheduled power, a term exists in the objective
that seeks to minimize the power transfer (even at the expense
of higher deviation) to reduce the payments to the ISO.
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
Fig. 5 Daily average D-LMPs as a function of scaling factors while μ =1.
[14]
[15]
[16]
[17]
[18]
Fig. 6 Payment received by DMO minus payments to the ISO.
V. CONCLUSION
One of the challenges in the operation of distribution
markets is the clearing and settlement method they should use,
and the consequent impacts on the distribution market prices.
This paper studied the interdependencies between D-LMPs at
a distribution network with the T-LMP at the upstream node.
It was shown that by adding a penalty factor the D-LMPs
could be calculated in a way that the prices follow the
associated T-LMP or be determined independent of that. It
was further shown that these price would significantly change
the market settlement. The decision on the exact value for the
penalty coefficient, however, should be made by the DMO.
[19]
[20]
[21]
[22]
[23]
[24]
[25]
VI. REFERENCES
[1]
[2]
[3]
[4]
S. Parhizi, H. Lotfi, A. Khodaei, and S. Bahramirad, “State of the
Art in Research on Microgrids: A Review,” IEEE Access, vol. 3,
2015.
A. Khodaei, “Microgrid Optimal Scheduling With Multi-Period
Islanding Constraints,” IEEE Trans. Power Syst., vol. 29, no. 3, pp.
1383–1392, May 2014.
A. Khodaei and M. Shahidehpour, “Microgrid-Based CoOptimization of Generation and Transmission Planning in Power
Systems,” IEEE Trans. Power Syst., vol. 28, no. 2, pp. 1–9, 2012.
S. Bahramirad, A. Khodaei, J. Svachula, and J. R. Aguero,
[26]
[27]
[28]
“Building Resilient Integrated Grids: One neighborhood at a time.,”
IEEE Electrif. Mag., vol. 3, no. 1, pp. 48–55, Mar. 2015.
A. Khodaei, “Provisional Microgrids,” IEEE Trans. Smart Grid,
vol. 6, no. 3, pp. 1107–1115, 2015.
A. Khodaei, “Resiliency-Oriented Microgrid Optimal Scheduling,”
IEEE Trans. Smart Grid, vol. 5, no. 4, pp. 1584–1591, Jul. 2014.
A. Khodaei, S. Bahramirad, and M. Shahidehpour, “Microgrid
Planning Under Uncertainty,” IEEE Trans. Power Syst., vol. 30, no.
5, pp. 2417–2425, 2015.
A. Ipakchi and F. Albuyeh, “Grid of the future,” IEEE Power
Energy Mag., vol. 7, no. 2, pp. 52–62, Mar. 2009.
J. G. Kassakian, R. Schmalensee, G. Desgroseilliers, T. D. Heidel,
K. Afridi, a M. Farid, J. M. Grochow, W. W. Hogan, H. D. Jacoby,
J. L. Kirtley, and Others, “The future of the electric grid,”
Massachusetts Institute of Technology, Cambridge, MA, 2011.
O. Zinaman, A. M. Miller, and D. Arent, Power Systems of the
Future. NREL, 2015.
J. Taft and A. Becker-Dippmann, “Grid Architecture,” PNNL, 2015.
[Online].
Available:
http://energy.gov/sites/prod/files/2015/04/f22/QER Analysis - Grid
Architecture_0.pdf.
L. Kristov and P. De Martini, “21st Century Electric Distribution
System Operations,” Caltech. 2014.
New York State Department of Public Service, “Developing the
REV Market in New York: DPS Staff Straw Proposal on Track One
Issues,” 2014.
P. De Martini and L. Kristov, “Distribution Systems in a High
Distributed Energy Resources Future: Planning, Market Design,
Operation and Oversight,” no. 2. Lawrence Berkeley National
Laboratory, 2015.
E. Martinot, L. Kristov, and J. D. Erickson, “Distribution System
Planning and Innovation for Distributed Energy Futures,” Curr.
Sustain. Energy Reports, vol. 2, no. 2, pp. 47–54, Apr. 2015.
P. De Martini, “MORE THAN SMART: A Framework to Make the
Distribution Grid More Open, Efficient and Resilient.” Greentech
Leadership Group, Aug-2014.
F. Rahimi and S. Mokhtari, “From ISO to DSO: imagining new
construct--an independent system operator for the distribution
network,” Public Util. Fortn., vol. 152, no. 6, pp. 42–50, 2014.
M. Keay, J. Rhys, and D. Robinson, “Distributed Generation and its
Implications for the Utility Industry,” in Distributed Generation and
its Implications for the Utility Industry, Elsevier, 2014, pp. 165–187.
Z. Wang, B. Chen, J. Wang, M. M. Begovic, and C. Chen,
“Coordinated Energy Management of Networked Microgrids in
Distribution Systems,” IEEE Trans. Smart Grid, vol. 6, no. 1, pp.
45–53, Jan. 2015.
J. Tong and J. Wellinghoff, “Rooftop Parity: Solar for Everyone,
Including Utilities,” Public Util. Fortn., vol. 152, no. 7, pp. 18–23,
2014.
“New York State Embarks On Bold New Vision,” Electr. J., vol. 27,
no. 6, pp. 1–3, Jul. 2014.
L. Kristov and D. Hou, “Distributed Generation and its Implications
for the Utility Industry,” in Distributed Generation and its
Implications for the Utility Industry, Elsevier, 2014, pp. 359–377.
S. Huang, Q. Wu, S. S. Oren, R. Li, and Z. Liu, “Distribution
Locational Marginal Pricing Through Quadratic Programming for
Congestion Management in Distribution Networks,” IEEE Trans.
Power Syst., vol. 30, no. 4, pp. 2170–2178, Jul. 2015.
R. Li, Q. Wu, and S. S. Oren, “Distribution Locational Marginal
Pricing for Optimal Electric Vehicle Charging Management,” IEEE
Trans. Power Syst., vol. 29, no. 1, pp. 203–211, Jan. 2014.
S. Parhizi and A. Khodaei, “Market-based Microgrid Optimal
Scheduling,” in IEEE SmartGridCommConf., 2015.
S. Parhizi and A. Khodaei, “Investigating the Necessity of
Distribution Markets in Accomodating High Penetration
Microgrids,” in IEEE PES Transmission & Distribution Conference
& Exposition, 2016.
S. Parhizi, A. Khodaei, and S. Bahramirad, “Distribution Market
Clearing and Settelement,” in IEEE PES General Meeting, 2016.
W. H. Kersting, “Radial distribution test feeders,” in 2001 IEEE
Power Engineering Society Winter Meeting. Conference
Proceedings (Cat. No.01CH37194), 2001, vol. 2, pp. 908–912.
| 3cs.SY
|
Observation of the Kibble–Zurek Mechanism in Microscopic Acoustic
Crackling Noises
H.O. GHAFFARI 1 (A) , P. BENSON2, W. A. GRIFFTH1 , K.XIA3 , R.P.YOUNG 3
1
2
The University of Texas at Arlington, 500 Yates St. Arlington, TX 76019
Rock Mechanics Laboratory, School of Earth and Environmental Sciences, University of Portsmouth, Burnaby building,
Portsmouth, PO1 3QL, UK
3
Department of Civil Engineering and Lassonde Institute, University of Toronto, Toronto, 170 College Street
,M5s3e3,On, Canada
Characterizing the fast evolution of microstructure is key to
understanding “crackling” phenomena during the deformation of solid
brittle materials. For example, it has been proposed using atomistic
simulations of crack propagation in elastic materials that the formation of a
nonlinear hyperelastic zone around moving crack tips controls crack
velocity. To date, progress in understanding the physics of this critical zone
has been limited due to the lack of data describing the complex physical
processes that operate near microscopic crack tips. We show, by analyzing
acoustic emissions during rock deformation experiments, that the signature
of this nonlinear zone maps directly to crackling noises. In particular, we
use a functional network method to characterize a weakening zone that
forms near the moving crack tips, and we determine the scaling law between
the formation of defects and the traversal rate across the critical point of
transition to rapid weakening at the crack tip. Our results are in agreement
with the Kibble-Zurek Mechanism (KZM).
The spatio-temporal evolution of acoustic signals, known as crackling noise [1], is a direct
result of failing atomic bonds during material fracture. Such signals, if properly interpreted, may be
used to better understand the dynamics of rupture progress in the vicinity of crack tips over a broad
range of scales and conditions [2-4]. During crackling, part of the stored energy near the crack tip is
consumed in the breaking of molecular and atomic bonds, resulting in new crack surface(s). The key
to understanding the crackling process lies in the characterizing structure of the near-tip region,
where stresses amplitudes are large. Due to the microscopic size and high speeds encountered in the
area of the crack tip, direct measurements are difficult, and analysis typically relies on computational
techniques [5]. Because of the large strains present near crack tips, nonlinear elastic contributions
1
must occur. Recent work [2-3] suggests that non-linearity around the moving crack tip governs the
rupture velocity. Specifically, the local hyperelastic zone around the moving crack tip enhances
energy flow for stiffening systems, and reduces energy flow for softening systems, resulting
respectively in increases and decreases in the fracture velocity relative to the linear elastic case [2].
Furthermore, investigations using “slow” cracks in gels demonstrate a link between the spatial
energy flow around the rupture tip and the curvature of the tip [5]. This link is thought to be
responsible for inaccuracies in linear elastic analyses that are commonly used in material science for
simulating crack tip processes, with applications ranging from metal fatigue to earthquake
nucleation.
In this study we use a functional network method [6-7] applied to acoustic emission (AE)
data recorded from rock deformation and rock friction experiments (see methods section) to show
that moving micro-cracks contain signatures of non-linearity. We discover that the onset of the nonlinear stage prior to global failure coincides with the nucleation of kink instabilities, and we use
“network strings” to visualize spatio-temporal evolution of these topological defects. For the first
time, we show that emitted crackling noises hold the signature of the Kibble-Zurek mechanism
(KZM) [8-11] that provides an estimation of the defect density as a function of the traversing rate. A
key output of KZM relates faster ramp rates to higher defect density, as the result of a spontaneous
symmetry breaking process. We measure real-time evolution of a first-order correlation function of
the system (networks) and verify the main prediction of KZM, namely the power-law scaling of the
correlation length with the ramp rate. Moreover, we show that the correlation length near the
transition remain effectively frozen. This is the main underlying hypothesis behind the Kibble-Zurek
mechanism. In addition, using our laboratory data sets we illustrate that adiabatic-impulse transition,
as the core of the KZ hypothesis, plays a major role in approximating weakening rate (i.e., rupture
velocity[4]).
We analyse AE waveforms (the laboratory analogue of seismograms due to rock fracture or
earthquake rupture) under different simulated depth (pressure) conditions and loading paths, and on
two rock types: Westerly granite and Basalt from Mount Etna, Italy. (datasets Lab.EQ1, Lab.EQ2,
Lab.EQ3 and Lab.EQ4 - see methods section and supporting information). We apply tools from the
theory of complex networks to analyze emitted noises from microscopic cracks, where the acoustic
time series recorded at each sensor is represented as a node [SI-section 1; Ref.6,7]. These results
allow us to develop an interpretation of reordered multiple acoustic-crackling signals involving a
2
microsecond evolution of different dynamic crack tip phases as encoded in the network modularity
(Q) which we refer to as “Q-profiles” (see methods section). This evolution can be broken down into
three distinct phases (Fig. 1a, Fig. 1d-e): (1). The S-phase: an initial strengthening phase preceding
the critical point at which point weakening and catastrophic failure begins; (2) the W-phase: a fastslip or weakening phase; and (3) the D-phase: a slow slip or decelerating phase [6-7,12]. To better
understand the S-phase and how transition occurs over the critical point between S to W, we use the
reciprocal of modularity (R = Q-1) profiles (i.e., “R-profiles”) which closely resemble dynamic stress
profiles commonly used to characterize rock failure (Fig. 1). In the following discussion, we describe
the observation of defect formation prior to onset of the W phase. To proceed, we define the critical
zone onset, Rc, as the value of R at the time of the first impulse in the inverse of mean betweenness
centrality ( 1⁄log〈 . . 〉 ) profile, where 〈 〉 indicates the mean value of all nodes (Fig.2 & 3). In
Fig.2, we show 1⁄log〈 . . 〉 profiles for 6 acoustic events from our laboratory tests. For all events
this profile is characterized by a narrow impulse at the transition from the S to W phase.
Later, we will show that the first impulse corresponds to the first nucleated defect in
transition from S to W while the second spike indicates the defects’ formation in an inverse transition
(W to S). We define the duration, c , as the temporal length of this impulse zone in transition from
S to W. This also corresponds to the time between Rc and Rmax, the maximum value of R(t). Rmax is
the critical point, where the failure occurs and fast-weakening onsets. The duration of defect
formation (nucleation time) c varies between 0.25 µs to 8 µs.
In order to study the spatial variability of this impulse regime, we visualize the spatial
evolution of the degree ki of the ith node (see supplementary material and [4]), (in 2D) using polar
coordinates ( ri , i ) i 1,..., Nodes where
ri ki and i indicates the position of the node (which is fixed on
the outer circumference of the cylindrical sample; Figure 3c). We refer to these configurations as “Kstrings”, and the normal vector of the K-strings at each node indicates the local direction of
increasing ki with time. We evaluate the variation of
ri ki at each position (node) while we consider
the temporal evolution of each single event (Fig.3; Fig.S5-S6). In Fig 3, we show that the onset of
the impulse zone coincides with the folding of K-strings where the normal vectors are flipped at the
onset of the non-linear regime and form a local domain that we refer to as a “kink” (Fig. 3d).
Therefore, formation of domains in the course of the S-W transition results in a non-linear behaviour
in the S phase of the R-profile. Since the trend of R(t) does match with the mean value of k for all
3
nodes <k(t)> (Fig.S4):
k ( t ) R ( t ) ( t ) ,then
we infer
k (t ) (t ) .
Then, the non-linearity of
stress is well-connected to nucleation of these kinks prior to onset of the W-phase where fastweakening occurs. The W-phase corresponds to global instability of K-strings and “collapsing” of Kstring occurs where almost all nodes ki 0 (in other words, where all normal vectors of the Kt
strings point inward).
As emphasized by Polyakov [13] in his string theory, crumpled stings are analogous with the
Heisenberg paramagnet while undulations destroy long-range order in surface normals [13-14].
These defects are local defects in initially ordered structures and can be removed by global collapse
of K-strings and local bending or twisting around the defects cannot remove them (i.e., topological
defects)[15,16]. To analysis deformation of K-strings, we map them on the “Ising chain” where for
each node we assign si sign (
with negative
ki
) and then si 1 . With this mapping, defects are represented
t
si indicating flipped, inward-pointing normal vectors. A kink separating zones with up
and down nodes (in analogy with “spins”) functions as a locator for the change from one ground
state (S-phase) to another degenerate ground state (W-phase) [17].Then, we can study the real-time
evolution of a cracking noise with a typical Ising chain when the K-string goes through a phase
transition.
The critical point c is defined when m s 0 which coincides with Rmax (Fig.4, 5b). Here, m is
the order parameter of the K-strings and transition from S to W (and vice versa) occurs
continuously. Furthermore, we have calculated a correlation function for the K-strings including all
nodes approaching the critical point (Fig.4 a,b). We can fit a correlation function such as
x
x
G ( x ) (1 ) exp( ) where L is the total number of nodes, x is distance and is the correlation
L
length. Correlation length is the cut-off length of the correlation function where for shorter
distances than the correlation length, a power law function can be fitted. For a fully ordered state a
triangular function (i.e. fully coherent system and ) is given by green line in Figure 4c.
Adding defects destroys long range of order among nodes (Fig.4b-Fig.S14-15). Interestingly, when
approaching m 0, the correlation length becomes essentially frozen as shown in Fig.4d for
crackling event #255. This means as the system approaches the critical point at a finite rate, and
after some point in time the correlation length cannot follow its diverging equilibrium value and
4
transition occurs with much smaller than the system size. This correlation length sets the mean
finite-size of final domains. The formation of domains as well as observation of frozen correlation
length at the critical point of our acoustic networks indicate the signature of a spontaneous symmetry
breaking process, i.e., Kibble-Zurek mechanism (KZM, see Methods section) [8-10]. The core idea
behind the Kibble-Zurek mechanism is that near the transition, the freezing of the correlation length
is unavoidable. Based on this theory, the resulting density of defects left behind by continuous
transitions is dependent on the rate at which the critical point is traversed, and the rate with which the
system can adjust (the relaxation time). This mechanism is reflected in the density of defects and
“freeze-out” time which scales with ramp-rate. To test the KZM’s density of defect prediction, we
carefully measured the number of flipped nodes in the vicinity of the critical point where correlation
length is frozen (Fig. 6).
A key output of our analysis is that the number of defects (i.e., flipped nodes at the final
state) we observe is larger when the local ramp rate
is faster (Fig. 6a). We can fit a power-law
dR 0.35
) t t (Fig.6a). Here ˆ is the frozen correlation length and ( dR ) is the ramp
scaling as: ˆ (
c
dt
dt
t tc
rate (Fig.S4-6,S10). To measure the ramp rate, we measure the slope of R(t) prior to the impulse
zone where we have fully a coherent system (Fig. 3b); this will yield the ramp rate, which is
analogous to the local loading rate prior to the nucleation of kinks. The exponent of ∼0.35 ±0.06
obtained by fitting the laboratory data is in agreement (within experimental error) with the theoretical
value ν/(1 + νz) ≅ 0.34 where, for a mean field model for 1D-kinks, ν≅0.5 and z≅ 1 [18]. Here, the
parameters ν and z are spatial and dynamical critical exponents (see Methods section).
Furthermore, we can evaluate the time-reversal transition from the W to S phase while we
approach to the critical point from right (in other words, if we heal or reverse the failure process).
Approaching from the left (S-phase) or right (W-phase; time reversal or healing scenario) to the
critical point
c results in slightly different characteristics of the defects’ density (Fig.5b). The rate
of the S-ramp (linear strengthening rate or ramp rate) for most of the recorded events is higher than
the rate of W-ramp (i.e., linear weakening rate). Next we estimate the linear weakening rate. Using
s 12z 12z
[1
(
) ] , in
the KZM scaling law for defects, we obtain (Supplementary Information): w
0
0
5
which 0 is determined by the microscopic details of the system;
s and w are the time
characteristics of ramps in the S and W phases, respectively. Therefore, the linear weakening rate is
given by: (
2 1 z
dR
)w 0 [1 ( s )1 z ] 2 . A faster local S-ramp inversely scales with weakening rate.
dt
0
This is an important result since it has been shown that weakening rate is correlated with the global
rupture velocity of cracks [4,7]. To verify this prediction, we measured the rate of weakening from R
R Rmin
profiles R w max
. As we have shown in Fig.6b , our measurements qualitatively confirms
w
the aforementioned prediction.
We have shown, for the first time via laboratory data, the microcrack evolution over the
duration of only a few microseconds derived purely from multiple AEs event data. We mapped
acoustic excitations from crackling events to complex networks and then further to Ising-like chains.
Using these novel tools, we elucidated the transition to fast-weakening, and we showed that the
nature of this transition is governed by a defect mediated transition process. In particular, very close
to the critical point of failure, the correlation length is frozen, and the density of nucleated defects
scales with local stress ramp rate, implying that the Kibble-Zurek mechanism can be employed in
studying acoustic crackling noises. To determine the KZM exponent in the estimation of defect
density, we made extensive measurements over many reordered crackling noises from our
experiments. In addition, we found that dynamic stress-time profiles involve a double-KZM theory
where order-disorder-order transitions occur .With this interpretation, we could estimate the linear
rate of the weakening phase, a parameter closely to global rupture velocity. While all our focus on
this study was to characterize events with the definite continuous transition of S-W, nevertheless, we
could recognize events with an abrupt change in the order parameter that is characteristic of the first
order transitions (Fig.S18). Further study is needed to explore the frequency of these first-order
“laboratory earthquakes” and the boundary or environmental conditions which favor dominant firstorder crackling noises. Having multiple-acoustic cracking noises where each nodes sets in analogy
with spins, it would be interesting to study the effect of flipping nodes and the concept of “spinwaves” (i.e., crystal vibrations or phonons [19]) and their interaction in the course of the S-W
transition. Furthermore, future work should be focused on studying the thermal activation
6
mechanism of nucleated kinks prior to freeze-out time and the effects of weak-planes (cleavage) in
different crystals on polarization patterns of K-strings.
Methods
Laboratory Procedures: We use four sets of recorded acoustic emissions (labeled as Lab.EQ1, 2, 3 and 4)
from Westerly granite and Basalt rock samples (most of the analyzed events are precursor rupture fronts). The
Lab.EQ1and 2 are the recorded multi-stationary acoustic waveforms from evolution of frictional rock-interfaces of
Westerly Granite samples. The interfaces were in dry conditions with smooth (saw-cut) and naturally rough surfaces,
respectively [20]. The evaluated events are from different stages and position and are not limited to particular stage of the
tests. The Lab.EQ3 is the fast-loading experiment on a cylindrical sample of Westerly Granite (~10-5 s-1) at 50MPa
confining pressure (approximately 2km), which is about an order faster than Lab.EQ1 and 2 [21]. Lab.EQ4 are events
from Basalt samples; described in [22]. The global loading rate was 10-6 s-1. In all of the above experiments, we reordered
amplified events using 16 to 18 sensor networks in both short (discrete events) and long timescale recorders (AE
records). The resolution of each recorded interval during the life-time of a waveform was ~20-50ns.
Networks of Acoustic emission waveforms: The concept behind studying each single acoustic excitation event
is to determine the onset of the critical phase where the defects precipitate out in a matrix of elastic material. The
excitation of defects in the vicinity of a moving crack tip (such as dislocations) must be reflected in each recorded
acoustic event. To study each acoustic event, we use functional network theory to analyze multiple recorded waveforms.
The full details of the employed algorithm can be found in [4,6-7; Section 1 of SI]. In summary, the algorithm uses a
thresholded-closeness metric to establish functional networks where the nodes are the sensors and per each time step, a
network is assigned. Per each network a set of properties - nodes’ degree, modularity index and centrality measures - are
extracted (see the definition in SI-section 1). The temporal evolution of Q-values in the monitored time interval then
forms the Q-curves. Q(t) carries generic universal time scales, distinguishing details of micro-cracks over microsecond
timescale. Following [4,12], we found three generic classes, corresponding to the following phases in Q-profiles
(Fig.S.1): (1) S-phase: nucleation and main deformation phase with the signature of initial strengthening; (2) W-phase or
fast-slip and; (3) D-phase or slow slip stage. To distinguish the role of the S- phase, we define the parameter such as:
Q (Q0 is the initial or rest value of Q(t)). We also use a parameter indicating the mean of
R (Qnorm. ) 1 where Q
norm .
Q0
local energy flow for a given network: betweenness centrality (B.C). It characterizes the importance of a node using the
number of shortest paths from all nodes that pass through that node [23].
Kibble-Zurek Mechanism : The idea behind the KZM is to compare the relaxation time (or healing time of the
system in equilibrium) with the timescale of change of the control parameter (ε). Assuming a linear change of control
parameter in vicinity of the critical point ε(t) = t/τs, where τs is the ramp time in S-phase . The relaxation or healing time
we consider is an equilibrium (static or quasi-static) condition: ( )
0
vz
and z . This determines the reaction
time of the order parameter. Here, ν and z are spatial and dynamical critical exponents, and τ0 is a characteristic
timescale. The system can adiabatically follow the change imposed by the local stress ramp if relaxation time
1
characterized by τ(ε) is outside the interval tˆ ( 0 sz )1 z centered around the transition point ( tˆ is freeze-out time).
7
The system will cease to maintain with the imposed change at time
tˆ
before reaching the critical point and correlation
v
length of the system is effectively frozen. The correlation length is given by [9-10] : ˆ 0 ( s / 0 ) 1 vz . Topological
defects are formed with the density of one defect fragment per domain. An estimate for the resulting density of
2v
topological defects is given by [9-10]: 1 ( / ) 1 vz . In the frozen phase, one can define an effective control
0
s
2
0
ˆ
parameter ˆ Rˆ t ( 0 )1/1 . From plotting events in Rˆ Rc plane which hold Rˆ ( Rc )1/1 , we obtain z
s s
.
v
Then, we can estimate from ˆ ( Rc )1 vz where we measure the frozen correlation length for the given event with the
rate of transition Rc .
This procedure leads to estimate ν≅0.5
and z≅ 1 which does match with mean-field ‘s
approximation of the scaling coefficients of the Ising-model (Z2 symmetry breaking [18]). To test the KZ’s scaling
exponent, we analyzed events with definitive continuous transition of order parameter (Fig.S15-17) while we measure the
frozen correlation length in the Ising-like chains of functional acoustic networks (Fig.S.14).
Supplementary Information
Supplementary information accompanies this paper.
Acknowledgements
We would like to acknowledge and thank B.D. Thompson (Mine Design Engineering, Kingston, Canada), for providing part
of the employed data set in this work. The first author would like to acknowledge A.del Campo and W.Zurek for their encouragements
and points on the results.
Author Contributions
All authors contributed to the analysis and writing of the manuscript.
Competing interests statement
The authors declare no competing financial interests.
References
[1] Sethna, J. P., Dahmen, K. A., & Myers, C. R. Crackling noise. Nature, 410(6825), 242-250(2001)..
[2] Buehler, M. J., Abraham, F. F., & Gao, H. Hyperelasticity governs dynamic fracture at a critical length scale. Nature, 426(6963),
141-146 (2003)..
[3] Passelègue, F. X., Schubnel, A., Nielsen, S., Bhat, H. S. & Madariaga, R. From Sub-Rayleigh to Supershear Ruptures During
Stick-Slip Experiments on Crustal Rocks. Science. 340, 1208–1211 (2013).
8
[4] Ghaffari, H. O., & Young, R. P. Acoustic-friction networks and the evolution of precursor rupture fronts in laboratory earthquakes.
Scientific reports, 3. (2013).
[5] Livne, A., Bouchbinder, E., Svetlizky, I., & Fineberg, J. The near-tip fields of fast cracks. Science, 327(5971), 1359-1363(2010)..
[6] Ghaffari, H. O., Thompson, B. D., & Young, R. P. Complex networks and waveforms from acoustic emissions in laboratory
earthquakes. Nonlinear Processes in Geophysics, 21(4), 763-775(2014)..
[7] Ghaffari, H. O., Nasseri, M. H. B., & Young, R. PFaulting of Rocks in a Three-Dimensional Stress Field by Micro-Anticracks.
Scientific reports, 4. (2014)..
[8] Kibble, T. W. B. Topology of cosmic domains and strings. J. Phys. A: Math. Gen. 9, 1387–1398 (1976).
[9] Zurek, W. H. Cosmological experiments in superfluid helium? Nature 317, 505–508 (1985).
[10] del Campo, A., & Zurek, W. H. Universality of phase transition dynamics: Topological defects from symmetry breaking.
International Journal of Modern Physics A, 29(08) (2014).
[11] Kibble, T. Phase-transition dynamics in the lab and the universe. Phys. Today 60, 47–52 (September, 2007).
[12] Ben-David, O., Rubinstein, S. & Fineberg, J. Slip-Stick: The evolution of frictional strength. Nature. 463, 76 (2010).
[13] Polyakov, A. Fine structure of strings. Nuclear Physics B, 268(2), 406-412(1986).
[14] Kantor, Y., & Nelson, D. R. Phase transitions in flexible polymeric surfaces. Physical Review A, 36(8),(1987).
[15] Mermin, N. D. The topological theory of defects in ordered media. Rev. Mod. Phys. 51,591–648 (1979).
[16] Anderson, P. W. Basic Notions of Condensed Matter Physics (Benjamin/Cummings Pub. Co., Advanced Book Program, 1984).
[17] Dziarmaga, J., Zurek, W. H., & Zwolak, M. (2012). Non-local quantum superpositions of topological defects. Nature
Physics, 8(1), 49-53.
[18] Chaikin, P. M. & Lubensky, T. C. Principles of Condensed Matter Physics (Cambridge Univ. Press, 2000).
[19] Feynman, R. P. (1998). Statistical Mechanics: A Set of Lectures, Advanced Book Classics.
[20] Thompson, B. D., Young, R. P. & Lockner, D. A. Premonitory acoustic emissions and stick-slip in natural and smooth-faulted
Westerly granite. J Geophys Res. 114, B02205J (2009).
[21] Thompson, B. D., Young, R. P., & Lockner, D. A. Fracture in Westerly granite under AE feedback and constant strain rate
loading: nucleation, quasi-static propagation, and the transition to unstable fracture propagation. In Rock Damage and Fluid
Transport, (2006).
[22] Benson, P. M., Vinciguerra, S., Meredith, P. G., & Young, R. P. (2010). Spatio-temporal evolution of volcano seismicity: A
laboratory study. Earth and Planetary Science Letters, 297(1), 315-323.
[23] Newman, M. E. J. Networks: An Introduction (Oxford University Press, 2010).
9
Figures
Figure 1| Q-profiles representing dynamic crackling noises. (a) Three main stages of typical acoustic crackling noises
as shown in normalized Q-profiles: S-W-D phases correspond with strengthening, weakening and decelerating stages,
respectively. (b) A typical -R profile as generated from dataset Lab.EQ1. (c) A schematic representation of impulse zone
in transition from S to W phase. (d) five recorded stick-slip events with (dynamic) strain gauges measurements in
centimeter scales rock-interfaces [7,21].(e) An event from (d) where stresses dynamically drop about 25 MPa during a
stick-slip experiment.
10
Figure 2. (a,b) two typical acoustic emission events from cracking Granite samples (from Lab.EQ3). We have
shown scaled recorded acoustic waveforms in ~800µs and the corresponding normalized Q(t). c) Different events from
our tests with the signature of inverse of mean betweeness centrality which shows divergence of the parameter in vicinity
of the nucleation zone. Based on the resolution of our measurements, the total precipitate out time varies between 0.516µs
11
Figure 3| Nucleation of kinks and formation of domains . a) The mean number of edges (<k>) versus time for
event#24 from Lab.EQ4. b) Transition to nucleation zone is imprinted in diverging the inverse of mean betweeness
centrality (B.C). c) Schematic representation of sensors location (red filled circles) where the radius of the ring is
proportional with node’s degree (d) We have shown accumulated 2D spatio-temporal patterns of nodes’ degree in the
polar system for each time interval as in panel (a). Transition from the linear stage (1) to the non-linear regime (2+3) is
indicated by the onset of local defects (black arrows), inducing formation of “domains”. In the example reported, there
are three defect-zones. The arrows are normal to strings and crumpled strings destroy long-range order in string normal.
We have shown more examples in Fig.S7-S.8.
12
Figure 4. . Continuous Phase Transition from S to W. Results from a cracking noise #255 from Lab.EQ4 with
mapping on the Ising-chain (a) The real–time order parameter m versus time. m goes to zero when it approaches the
critical point c as the failure point. (b) Real-time cross-correlation function .Very close to m=0 in red, we can see
nearly overlapping patterns of G(x) indicating a frozen zone of correlation length (c) cross-correlation function versus
the normalized distance. For fully ordered state a triangular function (fully coherent) is given by green line. Approaching
m=0, we can fit a correlation function (red line), to determine correlation length. For this event, the frozen correlation
length is ˆ 0.086 . Blue points are the experimental measures of the correlation function. Before normalization, we had
L=300 nodes and ˆ 21 nodes. Approaching the point m=0, the correlation length becomes frozen as shown in (d). For
this event , the correlation length is roughly constant for ~0.6µs. See supplementary figures S13-S15 for more examples.
13
Figure 5. | Description of Structural Phase Transition in Acoustic Excitations. (a) Formation of domains occurs in
SII-stage while “failure” (i.e., fast-weakening regime) is characterized by undulation along the K-string. (b) Schematic
of order-disorder-order transition in R-profiles: approaching from Left (S-phase) or Right (W-phase ; time reversal) to
the critical point
c
results slightly different characteristics of defects. In particular, the rate of the S-ramp (linear
strengthening rate; approaching from left) is higher than the rate of W-ramp (i.e. linear weakening rate; approaching from
right). According to KZM, slower ramp rate marks longer freeze-out time and smaller defect density, so time-reversal
approaching to the critical point c yields longer KZM freeze-out time :
14
tˆR tˆL .
Figure 6| Dependency of frozen correlation length (kink density) on ramping rate. (a) Events with faster transition
to their critical points induce higher defect density (i.e., shorter correlation length). Here we show typical rupture fronts
from Lab.EQ3 and the size of the network is 300 nodes. We obtained b 0.35 in ˆ ( Rc ) b (dashed line), in agreement
with the mean-field model prediction (also see Fig.S14-S16) (b) effect of local stress-ramp rates on the fast-weakening
regime: Events with faster weakening rate scale with slower ramp rate. This is illustrated here via a log-log plot of the
normalized rate of weakening R w observed as a function of ramp-rate (events from Lab.EQ4).
15
| 5cs.CE
|
The Quadratic Minimum Spanning Tree Problem and its Variations
Ante Ćustić∗
Ruonan Zhang†
Abraham P. Punnen‡
arXiv:1603.04451v1 [cs.DS] 14 Mar 2016
March 16, 2016
Abstract
The quadratic minimum spanning tree problem and its variations such as the quadratic bottleneck
spanning tree problem, the minimum spanning tree problem with conflict pair constraints, and the
bottleneck spanning tree problem with conflict pair constraints are useful in modeling various real life
applications. All these problems are known to be NP-hard. In this paper, we investigate these problems
to obtain additional insights into the structure of the problems and to identify possible demarcation
between easy and hard special cases. New polynomially solvable cases have been identified, as well as
NP-hard instances on very simple graphs. As a byproduct, we have a recursive formula for counting the
number of spanning trees on a (k, n)-accordion and a characterization of matroids in the context of a
quadratic objective function.
Keywords: Quadratic spanning tree; complexity; tree enumeration; sparse graphs; row graded matrix;
matroids.
1
Introduction
Let G = (V, E) be an undirected graph with |E| = m. Costs ce and q(f, g) are given for each edge e ∈ E
and each pair of edges f, g ∈ E, f 6= g, respectively. Then the quadratic minimum spanning tree problem
(QMST) is formulated as follows:
P P
P
Minimize z(T ) =
q(e, f ) +
ce
e∈T f ∈T
f 6=e
e∈T
Subject to
T ∈ F,
where F is the family of all spanning trees of G. The associated cost matrix Qm×m have its (i, j)-th entry
as q(i, j) when i 6= j, and as ci when i = j.
The QMST can be viewed as a generalization of many well known optimization problems such as the
travelling salesman problem, the quadratic assignment problem, the maximum clique problem etc., and it can
be used in modeling various real life application areas such as telecommunication, transportation, irrigation
energy distribution, and so on. The problem was introduced by Assad and Xu [2], along with its special
case - the adjacent-only quadratic minimum spanning tree problem (AQMST), in which q(e, f ) = 0 if e and
f are not adjacent. The strong NP-hardness of both the QMST and AQMST was proved in [2] along with
ideas for solving the problem using exact and heuristic algorithms. The broad applications base and inherent
complexity of the QMST makes it an interesting topic for further research. Most of the works on QMST
∗ [email protected]. Department of Mathematics, Simon Fraser University Surrey, 250-13450 102nd AV, Surrey, British
Columbia, V3T 0A3, Canada
† [email protected]. Department of Mathematical Sciences, Xi’an Jiaotong-Liverpool University, 111 Ren’ai Road,
Suzhou, Jiangsu, 215123, China
‡ [email protected]. Department of Mathematics, Simon Fraser University Surrey, 250-13450 102nd AV, Surrey, British
Columbia, V3T 0A3, Canada
1
have been focussed on heuristic algorithms [4, 9, 13, 17–19, 24, 25, 28]. Ćustić and Punnen [5] provided a
characterization QMST instances that can be solved as a minimum spanning tree problem. Exact algorithm
for AQMST and QMST was studied by Pereira, Gendreau, and Cunha [2, 19–21]. A special case of QMST
with one quadratic term was studied by Buchheim and Klein [3], A. Fischer and F. Fischer [8].
There are many other problems which have been studied in the literature, that are closely related to the
QMST in terms of formulation and applications. We list below some of these variations of QMST that we
investigate in this paper.
The minimum spanning tree problem with conflict pairs (MSTC) [6, 7, 27]: Given a graph G with edge
costs ce and a set S ⊆ {{e, f } ⊆ E : e 6= f }, the MSTC is to find a spanning tree of G such that the tree
cost is minimized and for each edge pair {e, f } in S, at most one of them is included in the tree. The
feasibility problem of the MSTC, denoted by FSTC, is the problem of finding a feasible solution of the
MSTC, regardless of the costs. Given an FSTC instance, we construct a QMST on the same graph with
(
1 if {e, f } ∈ S
q(e, f ) =
0 otherwise.
Then the FSTC instance is feasible if and only if the above QMST has the optimal objective function value
0. Therefore, the FSTC reduces to the QMST.
The quadratic bottleneck spanning tree problem (QBST): By replacing the objective function of the
QMST with max{q(e, f ) : e, f ∈ T }, we obtain the QBST. The problem is introduced in [26] and shown to
be NP-hard even on a bipartite graph with 0-1 q(e, f ) values. FQBST, the feasibility version of the QBST, is
described as “Given a value µ, does there exist a spanning tree T of G such that max q(e, f ) ≤ µ?”. As the
e,f ∈T
FQBST is equivalent to the FSTC, [26] develops heuristic algorithms for the QBST using MSTC heuristics
as subroutines.
The bottleneck spanning tree problem with conflict pairs (BSTC): Similar to the relation between the
QBST and the QMST, the BSTC is defined by substituting the “min-sum” objective function in the MSTC
with a “min-max” objective function.
Furthermore, we define the “adjacent-only counterparts” for the above problems: AQBST, in which
q(e, f ) = 0 if e and f are not adjacent in the graph G; MSTAC, FSTAC, BSTAC, where the edges in the
conflict pairs are all restricted to be adjacent.
Even though the above problems are all proved to be NP-hard in general, exploring nicely solvable
special cases provide additional insights into the structure of these problems and opportunities for developing
effective heuristics [1,16]. The primary research question we focus in this paper is: To what extend the QMST
and its variations would retain its NP-hardness status, or become amenable for polynomial time solvability?
We consider restricting the structure of the graph G and that of the cost matrix Q to identify possible
demarkation between easy and hard instances.
The rest of the paper is organized as follows: Section 2 introduces the sparse graphs that we are investigating. These include fans, fan-stars, ladders, wheels and their generalizations, (k, n)-ladders and (k, n)accordions. A recursive formula is derived to count the number of spanning trees of a (k, n)-accordion, which
generalizes the well known sparring tree counting formulas for fans and ladders. In Section 3 we study the
complexity of QMST and its variations on these sparse graphs. It is shown that the problems are NP-hard,
except in the case of AQBST, MSTAC, BSTAC and AQBST on (k, n)-ladders and for these cases, we provide
O(kn) time algorithms. The problems on a general graph G but with specially structured cost matrices are
discussed in Section 4. In particular, we show that when Q is a permuted doubly graded matrix, QMST and
QBST are solvable in polynomial time. In this case the optimal solution value attains the Assad-Xu lower
bound [2]. This result is extended to the case of matroid bases and it provides a new characterization of
matroids.
We use the notations V (G) and E(G) to denote, respectively, the node and edge sets of a graph G. The
quadratic costs q(ei , ej ) for the edge-pair (ei , ej ) is sometimes denoted by q(i, j) for simplicity.
2
2
The (k, n)-accordion and the number of spanning trees
In this section we define the classes of graphs called (k, n)-ladders and (k, n)-accordions and study QMST
and its variations on these graphs. We also study the number of spanning trees on such graphs, that is, the
number of feasible solutions of the corresponding QMST.
v1
v1
v6
v2
v6
v3
v5
v9
v2
v1
v8
v2
v3
v7
v5
v3
v6
v4
v4
v4
v5
Figure 1: A fan F6 , a wheel W6 and a fan-star F S9
Given a path Pn = v1 − v2 − · · · − vn , a fan (Fn ) is obtained by introducing a new node u and edges (u, vi )
for i = 1, 2, . . . , n. If we add one more edge (v1 , vn ) to Fn , the resulting graph is called a wheel, denoted
by Wn . When n = 3k, deleting the edges (v3i , v3i+1 ), i = 1, . . . , k − 1, from Fn results in a fan-star (F Sn ).
Examples of a fan, a wheel and a fan-star are presented in Figure 1.
Let Pn1 = v1 − v2 − · · · − vn and Pn2 = u1 − u2 − · · · − un be two node-disjoint paths. Add n edges (vi , ui ),
i = 1, . . . , n, and the resulting graph is called a ladder (Ln ), see Figure 2.
v1
v2
v3
vi
u1
u2
u3
vn
vi+1
···
···
···
ui
ui+1
···
un
Figure 2: A ladder Ln
Let us now define a class of graphs that generalizes both fans and ladders. Given integers k ≥ 3 and
n ≥ 1, a (k, n)-accordion is a graph A that can be obtained by recursively fusing together n node-disjoint
k-cycles C1k , C2k , . . . , Cnk along an edge so that in the resulting graph, only two consecutive cycles have an
edge in common. More precisely, a (k, n)-accordion is a graph constructed using the following rules:
(i) Initialize the graph A1 as the k-cycle C1k . Embed A1 on the plane and designate all its edges as free
edges.
(ii) For i = 2 to n define Ai as follows: Choose a free edge (r, s) of Ai−1 . Introduce a k-cycle, say Cik to
Ai−1 using the edge (r, s) and k − 2 new nodes so that we get a planar embedding of the resulting
graph Ai . Designate any edge incident to a new node of Ai as a free edge.
Every so obtained graph An is a (k, n)-accordion, and the set of all (k, n)-accordions is denoted by A(k, n).
Figure 3 presents examples from A(5, 7) and A(4, 9).
Let us now define a subclass of A(k, n) which we call (k, n)-ladders. It is defined the same way as (k, n)accordions, except that in the construction scheme for a (k, n)-ladder, we designate an edge as ‘free’ if both
its end points are the new nodes introduced. (Note that for the construction of a (k, n)-accordion, an edge
is designated as free if at least one of its end points is a new node.) We denote by L(k, n) the set of all
(k, n)-ladders.
3
C75
C74
C65
C84
C94
C64
C55
C45
C54
C44
C35
C34
C25
C14
C15
(a)
C24
(b)
Figure 3: Examples of (5, 7)-accordion and (4, 9)-accordion
Note that the (5, 7)-accordion in Figure 3(a) is also a (5, 7)-ladder, while the (4, 9)-accordion in Figure 3(b)
is not a (4, 9)-ladder. It is easy to verify that a fan is a (3, n)-accordion but not a (3, n)-ladder, and Ln is a
(4, n)-ladder and it is unique.
The formulas for counting the number of spanning trees of Fn and Ln are already known [12,23]. Now we
generalize these results by deriving a formula for counting the number of spanning trees in (k, n)-accordions
and (k, n)-ladders.
Let τ (G) be the number of spanning trees of a graph G. Then for any e ∈ E(G) the following property
holds.
Lemma 1. τ (G) = τ (T − e) + τ (G/e), where G − e is the graph obtained by deleting e from G, and G/e is
obtained by coalescing the endpoints of e.
The proof is straightforward due to the fact that the total number of spanning trees in G is the number
of spanning trees that contain e, plus the number of spanning trees without e.
A recursive formula for the number of spanning trees of a (k, n)-accordion is given in the following
theorem.
Theorem 2. Every (k, n)-accordion has the same number of spanning trees. If we denote this number by
τ (A(k, n)), then for every integers k ≥ 3, n ≥ 3
τ (A(k, n)) = k · τ (A(k, n − 1)) − τ (A(k, n − 2)).
(1)
Proof. Let A be a (k, n)-accordion generated by k-cycles Cik , . . . , Cnk . Similarly let B and C be the cork
responding (k, n − 1) and (k, n − 2)-accordions generated respectively by the k-cycles C1k , . . . , Cn−1
and
k
k
k
k
C1 , . . . , Cn−2 . An edge e in A is called a ‘free edge’ if e ∈ Cn and e ∈
/ Cn−1 . Likewise, an edge e in B is
k
k
called a ‘free edge’ if e ∈ Cn−1
and e ∈
/ Cn−2
. Let A1 be the graph obtained by contracting a free edge of A
1
and B be the graph obtained by contracting a free edge of B. If e is a free edge of B then τ (B − e) = τ (C).
Thus from Lemma 1,
τ (B) = τ (C) + τ (B 1 ).
(2)
Note that any spanning tree of A either contains all free edges or does not contain exactly one free edge.
Further, any spanning tree that contains all free edges does not contain the edge e∗ which is common to
k
both Cnk and Cn−1
. Then the graph obtained from A by deleting e∗ and then contracting the path of free
edges is isomorphic to B 1 . Since A contains k − 1 free edges, we have
τ (A) = (k − 1)τ (B) + τ (B 1 ).
4
(3)
From (2) and (3) we have
τ (A) = kτ (B) − τ (C).
(4)
Recall that A is an arbitrary (k, n)-accordion with n ≥ 3, and B, C depend on A, and are (k, n−1)-accordion
and (k, n − 2)-accordion, respectively. Note that there is only one (k, 1)-accordion and all (k, 2)-accordions
are isomorphic. Hence, for a fixed k and n ≤ 2, every (k, n)-accordion has the same number of spanning
trees. Then from recursion (4) it follows that for every fixed k ≥ 3 and n ≥ 1, the number of spanning trees
for any (k, n)-accordion is the same, and hence its recursive formula is given by (1).
Theorem 2 gives an implicit formula for the number of spanning trees on general (k, n)-accordions. In
the case k = 3, it gives us τ (A(3, n)) = 3 · τ (A(3, n − 1)) − τ (A(3, n − 2)). By solving the recursion we obtain
√ !n+1
√ !n+1
1 3+ 5
3− 5
τ (A(3, n)) = √
−
.
(5)
2
2
5
When k = 4, τ (A(4, n)) = 4 · τ (A(3, n − 1)) − τ (A(4, n − 2)), from which it follows that
√
√ n+1
√ n+1
3
2+ 3
− 2− 3
.
τ (A(4, n)) =
6
(6)
Formulas (5) and (6) are consistent with the known spanning tree enumeration formulas for fans and ladders [12,23], moreover, they generalize them. Furthermore, Theorem 2 can be used to deduce explicit formulas
for the (k, n)-accordions for any fixed k. Since every element of A(k, n) contains an exponential number of
spanning trees, solving the QMST and its variations on this class of graphs by complete enumeration would
be computationally expensive.
3
Complexity of the QMST variations on some sparse graphs
We now investigate the complexity of the QMST and its variations on the sparse graphs discussed in Section 2.
3.1
Intractability results
Recall from Section 1 that FSTAC is the feasibility version of the adjacent only minimum spanning tree
problem with conflict pair constraints.
Theorem 3. The FSTAC on fan-stars is NP-complete.
Proof. We reduce the 3-SAT problem to the FSTAC on the fan star F Sn .
Let s = (x11 ∨x12 ∨x13 )∧(x21 ∨x22 ∨x23 )∧· · ·∧(xn1 ∨xn2 ∨xn3 ), be an instance of 3-SAT given in conjunctive normal form. From this instance, we construct a graph G with node set {u, v11 , v12 , v13 , . . . , vn1 , vn2 , vn3 }
and edge set E1 ∪ E2 , where E1 = {eij = (u, vij ) : i = 1, . . . , n, j = 1, 2, 3}, E2 = {(vij , vij+1 ) :
i = 1, . . . , n, j = 1, 2}. As shown in Figure 4, G is a fan-star. The conflict set S is defined as S =
{{eij , ekl } : xij and xkl are negations of each other}. Note that the edges in each conflict pair are adjacent.
If T is a solution of the FSTAC on G, then let XT = {xij : i = 1, . . . , n, j = 1, 2, 3} be such that xij is
true if eij ∈ T , and false otherwise. For each i = 1, . . . , n, at least one of xi1 , xi2 , xi3 must be true since T
contains at least one of ei1 , ei2 , ei3 . Moreover, eij and ekl cannot both be in T if {eij , ekl } ∈ S, so in XT
at most one of xij , xkl will be true if they are negations of each other. Hence XT is a true assignment for
the 3-SAT problem.
Conversely, suppose X is a true assignment of the 3-SAT problem. Then S = {eij : xij is true in X} is
an acyclic subgraph of G with at most one edge from each conflict pair. If S spans G, then it is a solution
of the FSTAC. Otherwise, we add necessary edges from E2 to S to form a spanning tree T , which gives us
a solution of the FSTAC.
The result now follows from the NP-completeness of 3-SAT.
5
u
ei1
ei2
ei3
vi1
vi2
vi3
Figure 4
Since FSTAC is the feasibility version of MSTAC, the MSTAC is also NP-hard on fan-stars. Fan-star
is a subgraph of a fan and a wheel, hence the MSTAC on fan-stars can be reduced to the MSTAC on fans
or wheels by assigning large costs on additional edges. That proves NP-hardness of the MSTAC on fans,
wheels and (k, n)-accordions. MSTAC is a special case of MSTC and AQMST, and MSTC is a special
case of QMST, hence all of those problems are NP-hard on fan-stars, fans, wheels and (k, n)-accordions.
Furthermore, from Theorem 3 it easily follows that all bottleneck versions of these problems are NP-hard on
fan-stars, fans, wheels and (k, n)-accordions. These observations are summarized in the following corollary.
Corollary 4. MSTAC, MSTC, AQMST, QMST, BSTAC, BSTC, AQBST and QBST are NP-hard on
fan-stars, fans, wheels and (k, n)-accordions.
Next we identify some intractability results for problems on ladders.
Theorem 5. The FSTC on ladders is NP-complete.
Proof. Again we reduce the 3-SAT problem to the FSTC on ladders.
Let s = (x11 ∨ x12 ∨ x13 ) ∧ (x21 ∨ x22 ∨ x23 ) ∧ · · · ∧ (xn1 ∨ xn2 ∨ xn3 ), be an instance of 3-SAT given
in conjunctive normal form. From this, we construct a ladder G shown in Figure 5. Let S = {{eij , ekl } :
xij and xkl are negations of each other}. Note that in S the conflict edges are not necessarily adjacent.
Then there exists a solution T of the FSTC on G, if and only if the 3-SAT instance has a true assignment.
Figure 5: A ladder constructed from a 3-SAT instance
The detailed proof is very similar to the one given in Theorem 3, and hence omitted.
Again, NP-completeness of FSTC is propagated to MSTC, QMST, BSTC and QBST, as is summarized
in the following corollary.
Corollary 6. MSTC, QMST, BSTC and QBST are NP-hard on ladders and (k, n)-ladders.
3.2
Polynomially Solvable Special Cases
Let us now examine the complexity of the QMST variations that are not covered in Section 3.1. We show
that these remaining problems are easy by proposing a linear time algorithm to solve the AQMST on (k, n)ladders. Let us first introduce some notations.
6
Recall that (k, n)-ladder is a graph which is a sequence of n k-cycles such that only consecutive k-cycles
k
(Cik and Ci−1
) intersect, and their intersection is an edge (and its corresponding two vertices). Let us label
k
k
the edges of Cik by ei1 , ei2 , . . . , eik , where ei1 is the edge that is also in Ci−1
, eik is the edge that is also in Ci+1
,
i+1
i
i
i
i
i
i
and ek−1 and ek−2 are adjacent to ek . Note that then ek = e1 . Furthermore, let v1 and v2 denote the
vertices incident to eik . See Figure 6 for an illustration.
eik−1
v1i
C16
6
Ci−1
ei1
Ci6
eik
6
Ci+1
v2i
eik−2
Cn6
Figure 6
Given a (k, n)-ladder L, we denote with Li the (k, i)-ladder determined by the first i k-cycles of L. Next
we define T1i , T2i , T3i and T4i to be the spanning trees of Li with minimum AQMST cost that satisfy additional
properties. In particular, let T1i be a spanning tree of Li with minimum AQMST cost that contains eik , eik−1
and does not contain eik−2 . Analogously, let T2i be the minimum spanning tree that contains eik , eik−2 and
does not contain eik−1 , let T3i be the minimum spanning tree that contains eik−1 , eik−2 and does not contain
eik , and let T4i be the minimum spanning tree that contains eik , eik−1 and eik−2 . Note that all possible
configurations of eik , eik−1 , eik−2 are covered.
Similarly, we define F1i , F2i and F3i to be the minimum cost spanning forests of Li made of exactly two
trees, such that one tree contains v1i and the other tree contains v2i . In particular, let F1i be a minimum cost
such forest that contains eik−1 and does not contain eik , eik−2 , let F2i be a minimum cost such forest that
contains eik−2 and does not contain eik , eik−1 , and let F3i be a minimum cost such forest that contains eik−1 ,
eik−2 and does not contain eik . Note that one tree in forests F1i and F2i are exactly the single vertices v1i or
v2i .
Next we show that for i ≥ 2, if z(Tji−1 ), Tji−1 , j = 1, . . . , 4 and z(Fji−1 ), Fji−1 , j = 1, 2, 3 are known,
then z(Tji ), Tji , j = 1, . . . , 4 and z(Fji ), Fji , j = 1, 2, 3 can be calculate in O(k) time. For two edge disjoint
graphs/edge sets S1 , S2 we define S1 + S2 to be the graph spanned by edges of S1 and S2 . Then it is easy
to verify that the following recursive relations hold:
n
o
z(T1i ) = min z Tji−1 + {Cik − ei1 − eik−2 } ,
(7)
j=1,...,4
n
o
z(T2i ) = min z Tji−1 + {Cik − ei1 − eik−1 } ,
(8)
j=1,...,4
n
o
z(T3i ) = min z Tji−1 + {Cik − ei1 − eik } ,
(9)
j=1,...,4
(
n
o
z(T4i ) = min
min z Fji−1 + {Cik − ei1 } ,
j=1,2,3
min
j=1,...,4
`=2,...,k−3
n
o
z Tji−1 + {Cik − ei1 − ei` }
n
o
z Tji−1 + {Cik − ei1 − eik − eik−2 } ,
j=1,...,4
n
o
i
z(F2 ) = min z Tji−1 + {Cik − ei1 − eik − eik−1 } ,
z(F1i ) = min
j=1,...,4
7
)
,
(10)
(11)
(12)
(
z(F3i )
= min
min
j=1,2,3
n
o
z Fji−1 + {Cik − ei1 − eik } ,
min
j=1,...,4
`=2,...,k−3
n
o
z Tji−1 + {Cik − ei1 − eik − ei` }
)
.
(13)
Since adjacencies between edges of Cik and graphs Tji−1 , Fji−1 are known, the minimization functions
above can be calculated easily. Function values in (10) and (13) can be calculated in O(k) time and the
remaining values in constant time, provided z(Tji−1 )’s and z(Fji−1 )’s are known.
Tj1 , j = 1, . . . , 4 and Fj1 , j = 1, 2, 3 are easily calculated, and for i ≥ 2, Tji ’s and Fji ’s and their costs can
incrementally be calculated from Tji−1 ’s and Fji−1 ’s along with their costs. When the value of i increases to
n, the optimal solution of the AQMST on L is obtained. We call this algorithm the AQMST-(k, n)-ladder
algorithm and is summarized as Algorithm 1.
Algorithm 1: AQMST-(k, n)-ladder
Input: A (k, n)-ladder graph L with costs ce for e ∈ E and q(e, f ) for e, f ∈ E, e 6= f ;
Calculate z(Tj1 ), j = 1, . . . , 4, and z(Fj1 ), j = 1, 2, 3;
for i = 2 to n do
Determine Tji , j = 1, . . . , 4, and Fji , j = 1, 2, 3, using (7)-(13);
end for
T n is the minimum cost tree from {T1n , T2n , T3n , T4n };
7: Output T n and z(T n );
1:
2:
3:
4:
5:
6:
Theorem 7. The AQMST-(k, n)-ladder algorithm solves AQMST on (k, n)-ladders in O(kn) time.
Proof. The correctness of the AQMST-(k, n)-ladder algorithm follows from the exploration of all the possible
cases resulting in recursion relations (7)-(13). There are O(n) iterations of line 4 of the algorithm, and each
iteration takes O(k) time to calculate the trees and corresponding costs as discussed above. Hence, the
overall complexity is O(kn).
The MSTAC can be easily reduced to the AQMST, and by calculating the maximum rather than the
summation, Algorithm 1 can be adapted to solve the bottleneck versions of the problems on (k, n)-ladders.
So without describing the detailed steps, the following corollary holds.
Corollary 8. MSTAC, AQBST and BSTAC on (k, n)-ladders can be solved in O(kn) time.
We have determined complexities of all problem variations on all graph classes investigated. The results
are summarized in Table 1, in which “×” represents NP-hardness and “X” means polynomially solvable.
fan-star
fan
wheel
ladder
(k, n)-ladder
(k, n)-accordion
MSTAC
×
×
×
X
X
×
MSTC
×
×
×
×
×
×
AQMST
×
×
×
X
X
×
QMST
×
×
×
×
×
×
BSTAC
×
×
×
X
X
×
BSTC
×
×
×
×
×
×
AQBST
×
×
×
X
X
×
Table 1: The polynomial solvability of QMST variations
8
QBST
×
×
×
×
×
×
4
The QMST with row graded cost matrix
In Section 3 it is shown that the QMST and its variations are mostly NP-hard even when restricted to very
simple classes of graphs on which a wide variety of hard optimization problems can be solved efficiently.
Therefore, we shift the focus from special graphs to specially structured cost matrices.
For each ei ∈ E(G) consider the minimum spanning tree problem MST(i, Q):
P
MST(i, Q):
Minimize
q(i, j)
ej ∈E(T )
Subject to
T ∈ F,
where F is the family of all spanning trees of G. Let z i be the optimal objective function value of MST(i, Q),
and consider the minimum spanning tree problem:
P
MST(Q):
Minimize
zj
ej ∈E(T )
Subject to
T ∈ F.
Let L̃ be the optimal objective function value of the MST(Q). It is shown in [2] that L̃ is a lower bound
for the optimal objective function value of QMST. We call L̃ the natural lower bound for the QMST. If
we could find a spanning tree of G with objective function value L̃, it is surely an optimal solution of the
QMST.
An m × m matrix Q = (q(i, j)) is said to be row graded, if q(i, 1) ≤ q(i, 2) ≤ · · · ≤ q(i, m) for all
i = 1, . . . , m. Furthermore, Q is called doubly graded if both Q and QT are row graded. Given an m × m
matrix Q and a permutation π on {1, 2, . . . , m}, we define π(Q) to be the m × m matrix which (i, j)-th entry
is q(π −1 (i), π −1 (j)). We say that Q is permuted row graded or permuted doubly graded if there exist a
permutation π such that π(Q) is row graded or doubly graded, respectively. Note that permuted row graded
and permuted doubly graded matrices are recognizable in polynomial time.
Let G be a graph with n vertices and edge set E(G) = {e1 , e2 , . . . , em }. Given a permutation on
{1, 2, . . . , m} π, a spanning tree T = {ei1 , ei2 , . . . , ein−1 } of G is called the π-critical spanning tree if the set
{π(i1 ), π(i2 ), . . . , π(in−1 )} is lexicographically smallest among all spanning trees of G. Recall that any MST
can be solved by a greedy algorithm, therefore the π-critical spanning tree is optimal if permuting the cost
vector with π makes it nondecreasing.
Lemma 9. Let Q be a cost matrix of the QMST on a graph G. If Q is permuted row graded such that
π(Q) P
is row graded, and if the π-critical spanning tree T 0 is a minimum spanning tree on G with edge costs
i
z := ej ∈T 0 q(i, j) for i = 1, . . . , m, then T 0 is an optimal solution of the QMST on Q.
Proof. Since π(Q) is row graded, the π-critical spanning tree T 0 is an optimal
P solution of the MST(i, Q)
for all i = 1, . . . , m with corresponding optimal objective function value z i = ej ∈T 0 q(i, j). As T 0 is also
optimal for the MST with edge costs z i , it is optimal for the MST(Q). Thus z(T 0 ) = L̃ and the optimality
of T 0 for the QMST is proved.
Using Lemma 9 we show that the QMST on permuted doubly graded matrices is polynomially solvable.
Theorem 10. If the cost matrix Q of the QMST on a graph G is permuted doubly graded such that π(Q) is
doubly graded, then the π-critical spanning tree T 0 is an optimal solution.
P
Proof. Let z i := ej ∈T 0 q(i, j) for i = 1, . . . , m. Since π(Q) is row graded, T 0 is optimal for MST(i, Q),
i = 1, . . . , m, with corresponding optimal objective function value
z i . Also π(Q)T isP
row graded, so for all
P
−1
−1
−1
j, q(π (1), j) ≤ q(π (2), j) ≤ · · · ≤ q(π (m), j). Then ej ∈T 0 q(π −1 (1), j) ≤ ej ∈T 0 q(π −1 (2), j) ≤
P
−1
−1
−1
· · · ≤ ej ∈T 0 q(π −1 (m), j), i.e. z π (1) ≤ z π (2) ≤ · · · ≤ z π (m) . Thus T 0 is a minimum spanning tree on
G with edge costs z i . From Lemma 9, T 0 is optimal for the QMST, with optimal objective function value
z(T 0 ) = L̃.
9
In the rest of this section, we extend the above results to a more general structure called matroid bases
and give a new characterization of matroids in terms of quadratic objective function. To the best of our
knowledge, no characterization of matroids is known that uses an optimization problem with a quadratic
objective function.
Let E = {1, 2, . . . , m} be a ground set and F be a family of subsets of E that we call bases, where |S| = s,
a constant for any S ∈ F. Let I = {X ⊆ S : S ∈ F}. We call the structure (E, I) an independence system
and the structure (E, F) a base system. An independence system (E, I) is called a matroid if and only if
for any S1 , S2 ∈ F and i ∈ S1 \ S2 , there exists j ∈ S2 \ S1 such that S1 \ {i} ∪ {j} ∈ F. For instance, when
E is the edge set of a graph G, F is the collection of all the spanning trees of G, and I is the collection of
all acyclic subgraphs of G, then (E, I) is called the graphic matroid.
Given a base system (E, F) and a weight w(i, j) for each i, j ∈ E, the quadratic minimum weight base
problem (QMWB) is formulated as follows:
P P
Minimize Π(S) =
w(i, j)
i∈S j∈S
Subject to
S ∈ F,
where W = (w(i, j))m×m is the associated cost matrix. For each i ∈ E we define a minimum weight base
problem as follows:
P
MWB(i, W ):
Minimize
w(i, j)
j∈S
Subject to
S ∈ F.
Let f i be the optimal objective function value of the MWB(i, W ), and similar to the case for QMST, the
optimal objective function value L̄ of the problem
P j
MWB(Q):
Minimize
f
j∈S
Subject to
S ∈ F,
is called the natural lower bound of the QMWB with cost matrix W .
Given a permutation π on {1, 2, . . . , m}, we say that a base B 0 = {i1 , i2 , . . . , is } is the π-critical base
if the set {π(i1 ), π(i2 ), . . . , π(is )} is lexicographically smallest among all sets {π(j1 ), π(j2 ), . . . , π(js )} where
{j1 , j2 , . . . , js } ∈ F.
Theorem 11. The following statements are equivalent:
(i) (E, I) is a matroid.
(ii) Let W be a cost matrix of the QMWB on a base system (E, F). If W is permuted row graded
P such that
π(W ) is row graded, and if the π-critical base S 0 is a minimum weight base for costs f i := j∈S 0 w(i, j)
for i = 1, . . . , |E|, then S 0 is an optimal solution of the QMWB with cost matrix W .
(iii) If the cost matrix W of the QMWB problem on a base system (E, F) is permuted doubly graded such
that π(W ) is doubly graded, then the π-critical base S 0 is an optimal solution. Moreover, the natural
lower bound is the optimal objective function value.
Proof. Using the fact that a minimum weight matroid can be found by the greedy algorithm, statements (i)
implies (ii) and (i) implies (iii) can be proved similarly as Lemma 9 and Theorem 10. Hence their proofs
are omitted.
To show (iii) implies (i), we assume (E, I) is not a matroid and we aim to show that in that case (iii) is
not true. Hence, we assume that there are S1 ∈ F, S2 ∈ F and K ∈ S1 \ S2 such that S1 \ {K} ∪ {i} ∈
/ F for
all i ∈ S2 \ S1 . Let π be a permutation on {1, 2, . . . , m} for which {π(i) : i ∈ S1 \ {K}} = {1, 2, . . . , s − 1},
10
{π(i) : i ∈ S2 \ S1 } = {s, s + 1, . . . , |S1 ∪ S2 | − 1}, π(K) = |S1 ∪ S2 | and {π(i) : i ∈ E \ (S1 ∪ S2 )} =
{|S1 ∪ S2 | + 1, . . . , m}. Note that S1 is the π-critical base, since S1 \ {K} ∪ {i} ∈
/ F for all i ∈ S2 \ S1 . Let
a cost matrix W be such that its entries are
1 if i or j is in E \ (S1 ∪ S2 ),
w(i, j) = 1 if i ∈ (S2 \ S1 ) ∪ {K} and j = K,
0 otherwise.
Clearly π(W ) is doubly graded, and hence W is permuted doubly graded, see Figure 7. Now let us consider
π(S1 \ {K}) π(S2 \ S1 ) π(K) π(E \ (S1 ∪ S2 ))
π(W ) =
0
0
0
1
0
0
1
1
0
0
1
1
1
1
1
1
π(S1 \ {K})
π(S2 \ S1 )
π(K)
π(E \ (S1 ∪ S2 ))
Figure 7
the objective function values of S1 and S2 in QMWB with cost matrix W . Π(S1 ) = 1 and Π(S2 ) = 0. Hence,
the π-critical base S1 is not an optimal solution of the QMWB on cost matrix W , which implies that (iii)
is not true.
Similarly, we prove that (ii) implies (i) by showing that if (E, I) is not a matroid, then (ii) is not true.
That will complete the proof of the theorem. So, assume that there are S1 ∈ F, S2 ∈ F and K ∈ S1 \ S2
such that S1 \ {K} ∪ {i} ∈
/ F for all i ∈ S2 \ S1 . Then again consider permutation π and the cost matrix
π(W )P
from Figure 7. π(W ) is row graded and the π-critical base S1 is a minimum weight base on the costs
f i = j∈S1 w(i, j), but S1 is not an optimal solution of the corresponding QMWB, S2 is.
We end this section by noting that Theorem 10 and Theorem 11 hold true also for the QBST and
quadratic bottleneck base problem, respectively. That is, when the sum in the objective value function is
replaced by the maximum. The same proofs work, since the greedy algorithm obtains an optimal solution
also for the linear bottleneck objective functions on a base system of a matroid [11].
Acknowledgments
This work was supported by an NSERC discovery grant and an NSERC discovery accelerator supplement
awarded to Abraham P. Punnen.
References
[1] R.K. Ahuja, O. Ergun, J.B. Orlin and A.P. Punnen, A survey of very large-scale neighborhood search
techniques, Discrete Applied Mathematics 123 (2002), 75–102.
[2] A. Assad and W. Xu, The quadratic minimum spanning tree problem, Naval Research Logistics 39
(1992), 399–417.
11
[3] C. Buchheim and L. Klein, Combinatorial optimization with one quadratic term: Spanning trees and
forests, Discrete Applied Mathematics 177 (2014), 34–52.
[4] R. Cordone and G. Passeri, Solving the quadratic minimum spanning tree problem, Applied Mathematics
and Computation 218 (2012), 11597–11612.
[5] A. Ćustić and A.P. Punnen, Characterization of the linearizable instances of the quadratic minimum
spanning tree problem, arXiv:1510.02197, 2015.
[6] A. Darmann, U. Pferschy and J. Schauer, Minimal spanning trees with conflict graphs, Optimization
online, 2009. http://www.optimization-online.org/DB_FILE/2009/01/2188.pdf
[7] A. Darmann, U. Pferschy, S. Schauer, and G.J. Woeginger, Paths, trees and matchings under disjunctive
constraints, Discrete Applied Mathematics 159 (2011), 1726–1735.
[8] A. Fischer and F. Fischer, Complete description for the spanning tree problem with one linearised
quadratic term, Operations Research Letters 41 (2013), 701–705.
[9] Z.-H. Fu and J.-K. Hao, A three-phase search approach for the quadratic minimum spanning tree
problem, Engineering Applications of Artificial Intelligence 46 (2015), 113–130.
[10] J. Gao. and M. Lu, Fuzzy quadratic minimum spanning tree problem, Applied Mathematics and Computation 164 (2005), 773–788.
[11] S.K. Gupta and A.P. Punnen, k-sum optimization problems, Operations Research Letters 9 (1990),
121–126.
[12] A.J.W. Hilton, Spanning trees and Fibonacci and Lucas numbers, The Fibonacci Quarterly 12 (1974),
259–262.
[13] M. Lozanoa, F. Glover, C. Garcı́a-Martı́nez, F. Javier Rodrı́guez and R. Martı́, Tabu search with
strategic oscillation for the quadratic minimum spanning tree, IIE Transactions 46 (2014), 414–428.
[14] S.M.D.M. Maia, E.F.G. Goldbarg and M.C. Goldbarg, On the biobjective adjacent only quadratic
spanning tree problem, Electronic Notes in Discrete Mathematics 41 (2013), 535–542.
[15] S.M.D.M. Maia, E.F.G. Goldbarg and M.C. Goldbarg, Evolutionary algorithms for the bi-objective
adjacent only quadratic spanning tree, International Journal of Innovative Computing and Applications
6 (2014), 63–72.
[16] S. Mitrovic-Minic and A.P. Punnen, Local search intensified: Very large-scale variable neighborhood
search for the multi-resource generalized assignment problem, Discrete Optimization 6 (2009), 370–377.
[17] T. Öncan and A.P. Punnen, The quadratic minimum spanning tree problem: A lower bounding procedure and an efficient search algorithm, Computers & Operations Research 37 (2010), 1762–1773.
[18] G. Palubeckis, D. Rubliauskas and A. Targamadzė, Metaheuristic approaches for the quadratic minimum
spanning tree problem, Information Technology and Control 29 (2010), 257–268.
[19] D.L. Pereira, M. Gendreau and A.S. da Cunha, Stronger lower bounds for the quadratic minimum
spanning tree problem with adjacency costs, Electronic Notes in Discrete Mathematics 41 (2013), 229–
236.
[20] D.L. Pereira, M. Gendreau and A.S. da Cunha, Branch-and-cut and Branch-and-cut-and-price algorithms for the adjacent only quadratic minimum spanning tree problem, Networks 64 (2015), 367–379.
[21] D.L. Pereira, M. Gendreau and A.S. da Cunha, Lower bounds and exact algorithms for the quadratic
minimum spanning tree problem, Computers and Operations Research 63 (2015), 149–160.
12
[22] A.P. Punnen and R. Zhang, Quadratic bottleneck problems, Naval Research Logistics 58 (2011), 153–
164.
[23] J. Sedláček, On the number of spanning trees of finite graphs, Časopis pro Pěstovánı́ Matematiky 94
(1969), 217–221.
[24] S.-M. Soak, D.W. Corne and B.-H. Ahn, A new evolutionary algorithm for spanning tree based communication network design, IEICE Transaction on Communication E88-B (2005), 4090–4093.
[25] S.-M. Soak, D.W. Corne and B.-H. Ahn, The edge-window-decoder representation for tree-based problems, IEEE transactions on Evolutionary Computation 10 (2006), 124–144.
[26] R. Zhang and A.P. Punnen, Quadratic bottleneck knapsack problems, Journal of Heuristics 19 (2013),
573–589.
[27] R. Zhang, S. Kabadi and A.P. Punnen, The minimum spanning tree problem with conflict constraints
and its variations, Discrete Optimization 8 (2011), 191–205.
[28] G. Zhou and M. Gen, An effective genetic algorithm approach to the quadratic minimum spanning tree
problem, Computers & Operations Research 25 (1998), 229–237.
13
| 8cs.DS
|
Towards Automated Deduction in Blackmail Case Analysis with
Forensic Lucid
Serguei A. Mokhov
Joey Paquet
Mourad Debbabi
Faculty of Engineering and Computer Science
Concordia University, Montréal, Québec, Canada,
{mokhov,paquet,debbabi}@encs.concordia.ca
arXiv:0906.0049v1 [cs.LO] 30 May 2009
Abstract
This work-in-progress focuses on the refinement of application of the intensional logic to cyberforensic analysis and its benefits are compared with the finite-state automata approach. This work
extends the use of the scientific intensional programming paradigm onto modeling and implementation of a cyberforensics investigation process with the backtrace of event reconstruction, modeling
the evidence as multidimensional hierarchical contexts, and proving or disproving the claims with
it in the intensional manner of evaluation. This is a practical, context-aware improvement over the
finite state automata (FSA) approach we have seen in the related works. As a base implementation
language model we use in this approach is a new dialect of the Lucid programming language, that
we call Forensic Lucid and in this paper we focus on defining hierarchical contexts based on the
intensional logic for the evaluation of cyberforensic expressions.
Keywords: intensional logic, intensional programming, cyberforensics, Forensic Lucid, Lucid, finitestate automata
1
Introduction
Problem Statement. The first formal approach for event reconstruction cyberforensic analysis appeared in two papers [1, 2] by Gladyshev et al. that relies on the finite-state automata (FSA) and their
transformation and operation to model evidence, witnesses, stories told by witnesses, and their possible
evaluation. One of the examples the papers present is the use-case for the proposed technique – Blackmail Investigation. We aim at the same case to model and implement it using the new approach, which
promises to be more friendly and usable in the actual investigator’s work and serve as a basis to further
development in the area.
Proposed Solution. We intend to show the intensional approach with a Lucid-based dialect to the
problem is an asset in the field of cyberforensics as it is promising to be more practical and usable
than the FSA. Since Lucid was originally designed and used to prove correctness of programming languages [3, 4], and is based on the temporal logic, functional and data-flow languages its implementation
to backtracking in proving or disproving the evidential statements and claims in the investigation process
as a evaluation of an expression that either evaluates to true or false given all the facts in the formally
specified context. We will also attempt to retain the generality of the approach vs. building a problemspecific FSA in the FSA approach that can suffer a state explosion problem.
From the logic perspective, it was shown one can model computations (the basic unit in the finite state
machines in [1, 2]) as logic [5]. When armed with contexts as first-class values and a demand-driven
model adopted in the implementation of the Lucid-family of languages [6, 7, 8, 9, 10] that limits the
scope of evaluation in a given set of dimensions, we come to the intensional logic and the corresponding
programming artifact. In the essence, we model our forensic computation unit in the intensional logic and
propose the ways to implement it in practice within an intensional programming platform [11, 6, 12, 8].
1
Automated Deduction in Blackmail Case Analysis with Forensic Lucid
Mokhov et al.
We see a lot of potential for this work to be successful and beneficial for cyberforensics as well as
intensional programming communities.
Based on the parameters and terms defined in the papers [1, 2], we have various pieces of evidence
and witnesses telling their own stories of the incident. The goal is to put them together to make the
description of the incident as precise as possible. To show that a certain claim may be true, an investigator
has to show that there are some explanations of evidence that agrees with the claim. To disprove the
claim, the investigator has to show there is no explanation of evidence that agree with the claim [1]. The
authors of the FSA approach did a proof-of-concept implementation of the algorithms in CMU Common
LISP [1] that we target to improve by re-writing in a Lucid dialect, that we call Forensic Lucid. In
this work we focus on the specification of hierarchical context expressions and the operators on them
when modeling the examples. LIPS, unlike Lucid, entirely lacks contexts build into its logic, syntax,
and semantics, thereby making the implementation of the cases more clumsy and inefficient (i.e. highly
sequential). Our system [11] (not discussed here) offers distributed demand-driven evaluation of Lucid
programs in a more efficient way and is more general than LISP’s compiler and run-time environment.
Lucid Overview. Lucid [3, 4, 13, 14, 15] is a dataflow intensional and functional programming language. In fact, it is a family of languages that are built upon intensional logic (which in turn can be understood as a multidimensional generalization of temporal logic) involving context and demand-driven
parallel computation model. A program written in some Lucid dialect is an expression that may have
subexpressions that need to be evaluated at certain context. Given the set of dimensions D = {dimi }
in which an expression varies, and a corresponding set of indexes or tags defined as placeholders over
each dimension, the context is represented as a set of <dimi : tagi > mappings and each variable in
Lucid, called often a stream, is evaluated in that defined context that may also evolve using context operators [7, 9, 10, 16]. The generic version of Lucid, the General Intensional Programming Language
(GIPL) [17], defines two basic operators @ and # to navigate (switch and query) in the contexts P. The
GIPL is the first1 generic programming language of all intensional languages, defined by the means of
only two intensional operators @ and #. It has been proven that other intensional programming languages
of the Lucid family can be translated into the GIPL [17].
1.1
General Intensional Programming System (GIPSY).
The GIPSY [18, 11, 19, 20, 12, 21, 22, 6, 8] is a platform implemented primarily in Java to investigate
properties of the Lucid family of languages and beyond. It executes Lucid programs following a demanddriven distributed generator-worker architecture, and is designed as a modular collection of frameworks
where components related to the development (RIPE2 ), compilation (GIPC3 ), and execution (GEE4 ) of
Lucid programs are separated allowing easy extension, addition, and replacement of the components.
This is a proposed testing and investigation platform for our Forensic Lucid language.
2
Forensic Lucid Overview
This section summarizes concepts and considerations in the design of the Forensic Lucid language,
which is being studied through another use-case than related works [23, 24]. The end goal is to define
our Forensic Lucid language where its constructs concisely express cyberforensic evidence as context
1 The
second being Lucx [9, 10, 16]
Integrated Programming Environment, implemented in gipsy.RIPE
3 General Intensional Programming Compiler, implemented in gipsy.GIPC
4 General Eduction Engine, implemented in gipsy.GEE
2 Run-time
2
Automated Deduction in Blackmail Case Analysis with Forensic Lucid
Mokhov et al.
of evaluations, which can be initial state of the case towards what we have actually observed as a final
state in the FSM. The implementing system [6, 11, 12, 8] backtraces intermediate results to provide the
corresponding event reconstruction path if it exists (which we do not discuss in this work). The result of
the expression in its basic form is either true or false, i.e. “guilty” or “not guilty” given the evidential
evaluation context per explanation with the backtrace. There can be multiple backtraces, that correspond
to the explanation of the evidence (or lack thereof).
Properties. We define Forensic Lucid to model the evidential statements and other expressions representing the evidence and observations as context. An execution trace of a running Forensic Lucid
program is designed to expose the possibility of the proposed claim with the events that lead to the conclusion. Forensic Lucid aggregates the features of multiple Lucid dialects mentioned earlier needed for
these tasks along with its own extensions. The addition of the context calculus from Lucx (stands for
“Lucid enriched with context” that promotes contexts as first-class values) for operators on simple contexts and context sets (union, intersection, etc.) are used to manipulate complex hierarchical context
spaces in Forensic Lucid. Additionally, Forensic Lucid inherits many of the properties of Lucx, Objective
Lucid, JOOIP (Java-embedded Object-Oriented Intensional Programming language), and their comprising dialects, where the former is for the context calculus, and the latter for the arrays and structural
representation of data for modeling the case data structures such as events, observations, and groupings
of the related data, and so on. (We eliminate the OO-related aspects from this work as well as some
others to conserve space and instead focus on the context hierarchies, syntax, and semantics.) Hierarchical contexts are also following the example of MARFL [25] using a dot operator and by overloading
@ and # to accept different types as their left and right arguments. One of the basic requirements is
that the final target definition of the syntax, and the operational semantics of Forensic Lucid should be
compatible with the basic Lucx and GIPL. This is necessary for compiler and and the run time system
within the implementing system, called General Intensional Programming System (GIPSY) [6, 11, 8].
The translation rules or equivalent are to be provided when implementing the language compiler within
GIPSY, and such that the run-time environment (General Eduction Engine, or GEE) can execute it with
minimal changes to GEE’s implementation.
Context. We need to provide an ability to encode the stories told by the evidence and witnesses. This
will constitute the context of evaluation. The return value of the evaluation would be a collection of
backtraces, which contain the “paths of truth”. If a given trace contains all truths values, it’s an explanation of a story. If there is no such a path, i.e. the trace, there is no enough supporting evidence of the
entire claim to be true.
The context for this task for simplicity of the prototype language can be expressed as integers or
strings, to which we attribute some meaning or description. The contexts are finite and can be navigated
through in both directions of the index, potentially allowing negative tags in our tag sets of dimensions.
Concurrently, our contexts can be a finite set of symbolic labels and their values that can internally be
enumerated. The symbolic approach is naturally more appropriate for humans and we have a machinery
to so in Lucx’s implementation in GIPSY [26, 10]. We define streams of observations as our context,
that can be a simple context or a context set. In fact, in Forensic Lucid we are defining higher-level
dimensions and lower-level dimensions. The highest-level one is the evidential statement, which is
a finite unordered set of observation sequences. The observation sequence is a finite ordered set of
observations. The observation is an “eyewitness” of a particular property along with the duration of the
observation. As in the FSA [2, 1], the observations are tuples of (P, min, opt) in their generic form. The
observations in this form, specifically, the property P can be exploded further into Lucx’s context set
and further into an atomic simple context [16, 7]. Context switching between different observations is
3
Automated Deduction in Blackmail Case Analysis with Forensic Lucid
Mokhov et al.
done naturally with the Lucid @ context switching operator. Consider some conceptual expression of a
storyboard in Listing 1 where anything in [ ... ] represents a story, i.e. the context of evaluation. foo
can be evaluated at multiple contexts (stories), producing a collection of final results (e.g. true or false)
for each story as well as a collection of traces.
foo @
{
[ f i n a l observed event , p o s s i b l e
[
],
[
]
}
i n i t i a l observed event ] ,
Listing 1: Intensional Storyboard Expression
While the [...] notation here may be confusing with respect to the notation of [dimension:tag]
in Lucid and more specifically in Lucx [16, 10, 7], it is in fact a simple syntactical extension to allow
higher-level groups of contexts where this syntactical sugar is later translated to the baseline context
constructs. The tentative notation of {[...],...,[...]} implies a notion similar to the notion of
the “context set” in [16, 7, 10] except with the syntactical sugar mentioned earlier where we allow
syntactical grouping of properties, observations, observation sequences, and evidential statements as our
context sets. The generic observation sequence [1] can be expanded into the context stream using the
min and opt values, where they will translate into index values. Thus, obs = (A, 3, 0)(B, 2, 0) expands
the property labels A and B into a finite stream of five indexed elements: AAABB. Thus, a Forensic Lucid
fragment in Listing 2 would return the third A of the AAABB context stream in the observation portion of
o. Therefore, possible evaluations to check for the properties can be as shown in Figure 1.
/ / G i v e me o b s e r v e d p r o p e r t y a t i n d e x 2 i n t h e o b s e r v a t i o n s e q u e n c e o b s
o @. o b s 2
where
/ / Higher −l e v e l d i m e n s i o n i n t h e f o r m o f ( P , min , o p t )
observation o ;
/ / Equivalent to w r i t i n g = { A , A , A , B , B };
o b s e r v a t i o n s e q u e n c e o b s = (A, 3 , 0 ) ( B , 2 , 0 ) ;
where
/ / P r o p e r t i e s A and B a r e a r r a y s o f c o m p u t a t i o n s
/ / o r any E x p r e s s i o n s
A = [ c1 , c2 , c3 , c4 ] ;
B = E;
...
end ;
end ;
Listing 2: Observation Sequence With Duration
The property values of A and B can be anything that context calculus allows. The observation
sequence is a finite ordered context tag set [10] that allows an integral “duration” of a given tag property.
This may seem like we allow duplicate tag values that are unsound in the classical Lucid semantics;
however, we find our way around little further in the text with the implicit tag index. The semantics of
the arrays of computations is not a part of either GIPL or Lucx; however, the arrays are provided by
JLucid and Objective Lucid. We need the notion of the arrays to evaluate multiple computations at the
same context. Having an array of computations is conceptually equivalent of running an a Lucid program
under the same context for each array element in a separate instance of the evaluation engine and then the
4
Automated Deduction in Blackmail Case Analysis with Forensic Lucid
Mokhov et al.
results of those expressions are gathered in one ordered storage within the originating program. Arrays in
Forensic Lucid are needed to represent a set of results, or explanations of evidential statements, as well
as denote some properties of observations. We will explore the notion of arrays in Forensic Lucid much
greater detail in the near future work. In the FSA approach computations ci correspond to the state q and
event i that enable transition. For Forensic Lucid, we can have ci as theoretically any Lucid expression
E.
Observed property (context): A A A B B
Sub-dimension index: 0 1 2 3 4
o
o
o
o
o
@.obs
@.obs
@.obs
@.obs
@.obs
0
1
2
3
4
=
=
=
=
=
A
A
A
B
B
To get the duration/index position:
o @.obs A = 0 1 2
o @.obs B = 3 4
Figure 1: Handling Duration of an Observed Property in the Context
In Figure 1 we are illustrating a possibility to query for the sub-dimension indices by raw property
where it persists that produces a finite stream valid indices that can be used in subsequent expressions,
or, alternatively by supplying the index we can get the corresponding raw property at that index. The
latter feature is still under investigation of whether it is safe to expose it to Forensic Lucid programmers
or make it implicit at all times at the implementation level. This is needed to remedy the problem of
“duplicate tags”: as previously mentioned, observations form the context and allow durations. This
means multiple duplicate dimension tags with implied subdimension indexes should be allowed as the
semantics of a traditional Lucid approaches do not allow duplicate dimension tags. It should be noted
however, that the combination of the tag and its index in the stream is still unique and can be folded into
the traditional Lucid semantics.
Transition Function. A transition function (described at length [1, 2] and the derived works) determines how the context of evaluation changes during computation. A general issue exists that we have to
address is that the transition function ψ is usually problem-specific. In the FSA approach, the transition
function is the labeled graph itself. In the first prototype, we follow the graph to model our Forensic
Lucid equivalent. In general, Lucid has already basic operators to navigate and switch from one context
to another, which represent the basic transition functions in themselves (the intensional operators such
as @, #, iseod, first, next, fby, wvr, upon, and asa as well as their inverse operators5 ). However,
a specific problem being modeled requires more specific transition function than just plain intensional
operators. In this case the transition function is a Forensic Lucid function where the matching state
transition modeled through a sequence of intensional operators. A question arises a of how to explicitly
model the transition function ψ and its backtrace Ψ−1 in the new language. A possible approach is to use
predefined macros in Lucid syntax [27]. In fact, the forensic operators are just pre-defined functions that
rely on the traditional and inverse Lucid operators as well as context switching operators that achieve
5 Defined
further
5
Automated Deduction in Blackmail Case Analysis with Forensic Lucid
Mokhov et al.
something similar to the transitions. At the implementation level, it is the GEE that actually does the
execution of ψ within GIPSY. In fact, the intensional operators of Lucid represent the basic building
blocks for ψ and Ψ−1 .
Operational Semantics. As previously mentioned, the operational semantics of Forensic Lucid for
the large part is viewed as a composition of the semantic rules of GIPL, Objective Lucid, and Lucx
along with the new operators and definitions. The explanation of the rules and the notation are given in
great detail in the cited works and are trimmed in this extended abstract due to shortage of space. The
Objective Lucid semantic rules were affected and refined by some of the semantic rules of JOOIP [28].
We also omit the Objective Lucid and JOOIP semantic rules due to space limitation and defer them
to another publication. The new rules of the operational semantics of Forensic Lucid cover the newly
defined operators primarily, including the reverse and logical stream operators as well as forensic-specific
operators. Refining the semantics of context set operators of Lucx, such as box and range are also a part
of this work. We use the same notation as the referenced languages to maintain consistency in defining
our rules.
3
Initial Blackmail Case Modeling
Figure 2: Cluster Data with the Blackmail Fragments
The case description in this section is from [2]. A managing director of some company, Mr. C,
was blackmailed. He contacted the police and handed them evidence in the form of a floppy disk that
contained a letter with a number of allegations, threats, and demands. The message was known to have
come from his friend Mr. A. The police officers went to interview Mr. A and found that he was on
holiday abroad. They seized the computer of Mr. A and interviewed him as soon as he returned into the
country. Mr. A admitted that he wrote the letter, but denied making threats and demands. He explained
that, while he was on holiday, Mr. C had access to his computer. Thus, it was possible that Mr. C added
the threats and demands into the letter himself to discredit Mr. A. One of the blackmail fragments was
found in the slack space of another letter unconnected with the incident. When the police interviewed the
person to whom that letter was addressed, he confirmed that he had received the letter on the day that Mr.
A had gone abroad on holiday. It was concluded that Mr. A must have added the threats and demands
into the letter before going on holiday, and that Mr. C could not have been involved. In Figure 2 is the
initial view of the incident as a diagram illustrating cluster data of the blackmail and unconnected letters.
Modeling the Investigation. In the blackmail example, the functionality of the last cluster of a file
was used to determine the sequence of events and, hence, to disprove Mr. A’s alibi. Thus, the scope of
the model can be restricted to the functionality of the last cluster in the unrelated file. The last cluster
model can store data objects of only three possible lengths: LENGT H = {0, 1, 2}. Zero length means
that the cluster is unallocated. The length of 1 means that the cluster contains the object of the size of
6
Automated Deduction in Blackmail Case Analysis with Forensic Lucid
Mokhov et al.
Simplified model of the cluster:
possible data lengths:
possible data values:
0
1
2
LEFT PART
RIGHT PART
u – unrelated
t1 – threats-obscured part t2 – threats in slack
o1 – other data left part o2 – other data right part
Observed final state:
LENGT H = 1
(u) unrelated
LEFT PART = {u,t1 , o1 }
RIGHT PART = {t2 , o2 }
Q = LENGT H × LEFT PART × RIGHT PART
(t2 ) threats in slack
Figure 3: Simplified Cluster Model
the unrelated letter tip. The length of 2 means that the cluster contains the object of the size of the data
block with the threats. In Figure 3 is, therefore, the simplified model of the investigation.
Events.
The state of the last cluster can be changed by three types of events:
1. Ordinary writes into the cluster:
W RIT E = {(u), (t1 ), (o1 ), (u,t2 ), (u, o2 ), (t1 ,t2 ), (t1 , o2 ), (o1 ,t2 ), (o1 , o2 )}
2. Direct writes into the file to which the cluster is allocated (bypassing the OS):
DIRECT W RIT E = {d(u,t2 ), d(u, o2 ), d(o1 ), d(t1 ,t2 ), d(t1 , o2 ), d(o1 ,t2 ), d(o1 , o2 )}
3. Deletion of the file which sets the length of the file to zero:
I = W RIT E
S
DIRECT W RIT E {del}
S
Formalization of the Evidence. The final state observed by the investigators is (1, u,t2 ). Let O f inal
denote the observation of this state. The entire sequence of observations is then os f inal = ($, O f inal ). The
observation sequence osunrelated says that the unrelated letter was created at some time in the past, and
that it was received by the person to whom it was addressed is osunrelated = ($, Ounrelated , $, (CT , 0, 0), $)
where Ounrelated denotes the observation that the “unrelated” letter tip (u) is being written into the cluster.
The evidential statement is then: esblackmail = (os f inal , osunrelated ).
Finding an Explanation of Mr. A’s Theory. Mr. A’s theory, encoded using the proposed notation, is
osMr.A = ($, Ounrelated−clean , $, Oblackmail , $), where Ounrelated−clean denotes the observation that the “unrelated” letter (u) is being written into the cluster and, at the same time, the cluster does not contain the
blackmail fragment; Oblackmail denotes the observation that the right part of the model now contains the
blackmail fragment (t2 ).
Explanations. There are two most logically possible explanations that can be represented by state
machine. See the corresponding state diagram for the blackmail case in Figure 4.
7
Automated Deduction in Blackmail Case Analysis with Forensic Lucid
Mokhov et al.
Figure 4: Blackmail Case State Machine
1. The first explanation:
(u)
(u,t2 )
(u)
. . . −→ (1, u, o2 ) −−−→ (2, u,t2 ) −→ (1, u,t2 )
• Finding the unrelated letter, which was written by Mr. A earlier;
• Adding threats into the last cluster of that letter by editing it “in-place” with a suitable text
editor (such as ViM [29]);
• Restoring the unrelated letter to its original content by editing it “in-place” again.
“To understand this sequence of events, observe that certain text editors (e.g. ViM
[29]) can be configured to edit text “in-place”. In this mode of operation, the
modified file is written back into the same disk blocks that were allocated to the
original file. As a result, the user can forge the file’s slack space by (1) appending
the desired slack space content to the end of the file, (2) saving it, (3) reverting the
file back to the original content, (4) saving it again.” [2]
2. The second explanation:
(u)
d(u,t2 )
. . . −→ (1, u, o2 ) −−−→ (1, u,t2 )
• The threats are added into the slack space of the unrelated letter by writing directly into the
last cluster using, for example, a low-level disk editor.
The blackmail case example of the initial implementation steps is in Listing 3.
4
Conclusion
The proposed practical approach in the cyberforensics field can also be used in a normal investigation
process involving crimes not necessarily associated with information technology. Combined with an
expert system (e.g. implemented in CLIPS [30]), it can also be used in training new staff in investigation
techniques. The focus on hierarchical contexts as first-class values brings more understanding of the
process to the investigators in cybercrime case management tools.
Future Work.
• Forensic Lucid Compiler and run-time environment.
• Prove equivalence to the FSA approach.
• Adapt/re-implement a graphical UI based on the data-flow graph tool [31] to simplify Forensic
Lucid programming for not very tech-savvy investigators.
8
Automated Deduction in Blackmail Case Analysis with Forensic Lucid
Mokhov et al.
MrA @ e s m r a
where
e v i d e n t i a l s t a t e m e n t e s m r a = { os mra , o s f i n a l , o s u n r e l a t e d } ;
o b s e r v a t i o n sequence os mra = ( $ , o u n r e l a t e d c l e a n , $ , o blackmail , $ ) ;
observation sequence o s f i n a l = ( $ , o f i n a l ) ;
o b s e r v a t i o n s e q u e n c e o s u n r e l a t e d = ( $ , o u n r e l a t e d , $ , ( Ct , 0 , 0 ) , $ ) ;
observation o f i n a l = (1 , ”u” , ” t2 ” ) ;
o b s e r v a t i o n o u n r e l a t e d c l e a n = ( 1 , ” u ” , ” o1 ” ) ;
// ...
i n v t r a n s (Q, e s m r a , o f i n a l ) = b a c k r a c e s
where
/ / l i s t of a l l possible dimensions
o b s e r v a t i o n Q = l e n g t h s box l e f t p a r t box r i g h t p a r t ;
/ / events
o b s e r v a t i o n l e n g t h s = unordered {0 , 1 , 2};
/ / symbolic
observation
” u ” =>
” t 1 ” =>
” o1 ” =>
};
l a b e l s map t o human d e s c r i p t i o n s
l e f t p a r t = unordered {
” unrelated ” ,
” t h r e a t s −o b s c u r e d p a r t ” ,
” other data ( l e f t part ) ”
observation r i g h t p a r t = unordered {
” t 2 ” => ” t h r e a t s i n s l a c k ” ,
” o2 ” => ” o t h e r d a t a ( r i g h t p a r t ) ”
};
b a c k t r a c e s = [ A, B , C , D, ] ;
where
...
end ;
end ;
end ;
Listing 3: Blackmail Case Modeling in Forensic Lucid
Acknowledgments. This research work was funded by the Faculty of Engineering and Computer Science of Concordia University, Montreal, Canada.
References
[1] Pavel Gladyshev and Ahmed Patel. Finite state machine approach to digital event reconstruction. Digital
Investigation Journal, 2(1), 2004.
[2] Pavel Gladyshev. Finite state machine analysis of a blackmail investigation. International Journal of Digital
Evidence, 4(1), 2005.
[3] Edward A. Ashcroft and William W. Wadge. Lucid - a formal system for writing and proving programs.
SIAM J. Comput., 5(3), 1976.
9
Automated Deduction in Blackmail Case Analysis with Forensic Lucid
Mokhov et al.
[4] Edward A. Ashcroft and William W. Wadge. Erratum: Lucid - a formal system for writing and proving
programs. SIAM J. Comput., 6((1):200), 1977.
[5] René Lalement. Computation as Logic. Prentice Hall, 1993. C.A.R. Hoare Series Editor. English translation
from French by John Plaice.
[6] Joey Paquet and Ai Hua Wu. GIPSY – a platform for the investigation on intensional programming languages. In Proceedings of the 2005 International Conference on Programming Languages and Compilers
(PLC 2005), pages 8–14, Las Vegas, USA, June 2005. CSREA Press.
[7] Joey Paquet, Serguei A. Mokhov, and Xin Tong. Design and implementation of context calculus in the GIPSY
environment. In Proceedings of the 32nd Annual IEEE International Computer Software and Applications
Conference (COMPSAC), pages 1278–1283, Turku, Finland, July 2008. IEEE Computer Society.
[8] Joey Paquet. A multi-tier architecture for the distributed eductive execution of hybrid intensional programs.
In Proceedings of 2nd IEEE Workshop in Software Engineering of Context Aware Systems (SECASA’09).
IEEE Computer Society, 2009. To appear.
[9] Kaiyu Wan, Vasu Alagar, and Joey Paquet. Lucx: Lucid enriched with context. In Proceedings of the 2005
International Conference on Programming Languages and Compilers (PLC 2005), pages 48–14, Las Vegas,
USA, June 2005. CSREA Press.
[10] Xin Tong. Design and implementation of context calculus in the GIPSY. Master’s thesis, Department of
Computer Science and Software Engineering, Concordia University, Montreal, Canada, April 2008.
[11] The GIPSY Research and Development Group. The General Intensional Programming System (GIPSY)
project. Department of Computer Science and Software Engineering, Concordia University, Montreal,
Canada, 2002–2008. http://newton.cs.concordia.ca/∼gipsy/, last viewed April 2008.
[12] Serguei A. Mokhov. Towards hybrid intensional programming with JLucid, Objective Lucid, and General Imperative Compiler Framework in the GIPSY. Master’s thesis, Department of Computer Science and Software
Engineering, Concordia University, Montreal, Canada, October 2005. ISBN 0494102934.
[13] William Wadge and Edward Ashcroft. Lucid, the Dataflow Programming Language. Academic Press, London, 1985.
[14] Edward Ashcroft, Anthony Faustini, Raganswamy Jagannathan, and William Wadge. Multidimensional,
Declarative Programming. Oxford University Press, London, 1995.
[15] Edward A. Ashcroft and William W. Wadge. Lucid, a nonprocedural language with iteration. Communication
of the ACM, 20(7):519–526, July 1977.
[16] Kaiyu Wan. Lucx: Lucid Enriched with Context. PhD thesis, Department of Computer Science and Software
Engineering, Concordia University, Montreal, Canada, 2006.
[17] Joey Paquet. Scientific Intensional Programming. PhD thesis, Department of Computer Science, Laval
University, Sainte-Foy, Canada, 1999.
[18] Joey Paquet and Peter Kropf. The GIPSY architecture. In Proceedings of Distributed Computing on the Web,
Quebec City, Canada, 2000.
[19] Bo Lu. Developing the Distributed Component of a Framework for Processing Intensional Programming
Languages. PhD thesis, Department of Computer Science and Software Engineering, Concordia University,
Montreal, Canada, March 2004.
[20] Joey Paquet, Peter Grogono, and Ai Hua Wu. Towards a framework for the general intensional programming
compiler in the GIPSY. In Proceedings of the 19th Annual ACM Conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA 2004), Vancouver, Canada, October 2004. ACM.
[21] Emil Vassev and Joey Paquet. A generic framework for migrating demands in the GIPSY’s demand-driven
execution engine. In Proceedings of the 2005 International Conference on Programming Languages and
Compilers (PLC 2005), pages 29–35, Las Vegas, USA, June 2005. CSREA Press.
[22] Ai Hua Wu and Joey Paquet. Object-oriented intensional programming in the GIPSY: Preliminary investigations. In Proceedings of the 2005 International Conference on Programming Languages and Compilers
(PLC 2005), pages 43–47, Las Vegas, USA, June 2005. CSREA Press.
[23] Serguei A. Mokhov and Joey Paquet. Formally specifying and proving operational aspects of Forensic Lucid
in Isabelle. Technical Report 2008-1-Ait Mohamed, Department of Electrical and Computer Engineering,
Concordia University, August 2008. In Theorem Proving in Higher Order Logics (TPHOLs2008): Emerging
10
Automated Deduction in Blackmail Case Analysis with Forensic Lucid
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
Mokhov et al.
Trends Proceedings.
Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Formally specifying operational semantics and
language constructs of Forensic Lucid. In Oliver Göbel, Sandra Frings, Detlef Günther, Jens Nedon, and
Dirk Schadt, editors, Proceedings of the IT Incident Management and IT Forensics (IMF’08), pages 197–
216, Mannheim, Germany, September 2008. GI. LNI140.
Serguei A. Mokhov. Towards syntax and semantics of hierarchical contexts in multimedia processing applications using MARFL. In Proceedings of the 32nd Annual IEEE International Computer Software and
Applications Conference (COMPSAC), pages 1288–1294, Turku, Finland, July 2008. IEEE Computer Society.
Xin Tong, Joey Paquet, and Serguei A. Mokhov. Context Calculus in the GIPSY. Unpublished, 2007.
Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Designing a language for intensional cyberforensic
analysis. Unpublished, 2007.
Aihua Wu, Joey Paquet, and Serguei A. Mokhov. Object-Oriented Intensional Programming: Intensional
Classes Using Java and Lucid. Submitted for publication to PPPJ’09, 2009.
Bram Moolenaar and Contributors. Vim the editor – Vi Improved. [online], 2007. http://www.vim.org/.
Gary Riley. CLIPS: A tool for building expert systems. [online], 2007. http://www.ghg.net/clips/
CLIPS.html, last viewed: December 2007.
Lei Tao. Warehouse and garbage collection in the GIPSY environment. Master’s thesis, Department of
Computer Science and Software Engineering, Concordia University, Montreal, Canada, 2004.
11
| 6cs.PL
|
Logical Methods in Computer Science
Vol. 6 (3:11) 2010, pp. 1–34
www.lmcs-online.org
Submitted
Published
Oct. 20, 2008
Sep. 1, 2010
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
THOMAS EHRHARD AND OLIVIER LAURENT
Preuves Programmes Systèmes, CNRS – Université Paris 7, France
e-mail address: [email protected], [email protected]
Abstract. We present a restriction of the solos calculus which is stable under reduction
and expressive enough to contain an encoding of the pi-calculus. As a consequence, it is
shown that equalizing names that are already equal is not required by the encoding of
the pi-calculus. In particular, the induced solo diagrams bear an acyclicity property that
induces a faithful encoding into differential interaction nets. This gives a (new) proof that
differential interaction nets are expressive enough to contain an encoding of the pi-calculus.
All this is worked out in the case of finitary (replication free) systems without sum,
match nor mismatch.
Introduction
The question of extending the Curry-Howard correspondence (between the λ-calculus
and intuitionistic logic) to concurrency theory is a long-standing open problem. We developed in a previous paper [EL10] a translation between the π-calculus [MPW92] and
differential interaction nets [ER06]. This has shown that differential linear logic — a logical system whose sequent calculus [EL10] has been obtained by the first author from a
precise analysis of some denotational models of linear logic based on vector spaces [Ehr02]
— is a reasonable candidate for a Curry-Howard correspondence with concurrent computation, since differential interaction nets appear to be expressive enough to represent key
concurrency primitives as provided by the π-calculus.
There are other works trying to relate variants of linear logic proof-nets with process
algebras (such as [Abr93, BS94, BM06, FM05, Maz05, HL10]) or to work on graphical
syntaxes for encoding the π-calculus (see [Mil94, LPV01, JM04] for example). However,
even if some of them try to build bridges with (linear) logic, they require some ad hoc
graphical constructions without logical foundations. Our goal is to build on differential
interaction nets coming with the logical justification of differential linear logic.
1998 ACM Subject Classification: F.3.2, F.4.1, F.1.2.
Key words and phrases: solos calculus, pi-calculus, prefix, typing, differential interaction nets.
This work has been partially funded by the French ANR projet blanc “Curry Howard pour la Concurrence”
CHOCO ANR-07-BLAN-0324.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-6 (3:11) 2010
CC
T. Ehrhard and O. Laurent
Creative Commons
2
T. EHRHARD AND O. LAURENT
Let us tell a bit more about the genesis of our translation. We discovered the notion of
communication areas as particular differential interaction nets able to represent some communication primitives. However, due to the asynchronous flavour of differential interaction
nets, it was not immediate that additional constructions such as prefixing could be easily
encoded. This led us to have a look at the solos calculus.
The solos calculus [LV03] has allowed to prove how action prefixes (thus sequentiality
constraints) can be encoded in a calculus without prefixes (results of this kind also appear
in [HY94] and are going much further than in the asynchronous π-calculus [HT91, Bou92]).
This is done by encoding the π-calculus into the solos calculus [LV03]. From this point of
view, the solos calculus can be seen as a low level concurrent language: even more basic
than the π-calculus.
The main goal of the present paper is to stress the very close relation which appears
between differential interaction nets [ER06] (the graphical syntax for differential linear
logic) and solo diagrams [LPV01] (the graphical syntax for the solos calculus).
Following the ingredients of [EL10], it is possible to give a simple representation of solo
diagrams as differential interaction nets (based on an implementation of nodes as communication areas). Such a “static” correspondence is not very interesting if there is no “dynamic”
counterpart in the reduction semantics. And this is where troubles come into the picture...
The reduction semantics almost match... There is a mismatch between the solos calculus
and communication areas with respect to the identification of an occurrence of name with
another occurrence of the same name during reduction (let us call this phenomenon, selfidentification). Note that such an identification never occurs in the π-calculus. Since name
passing is handled by the substitution of a bound name by another name. As in the fusion
calculus [PV98], the solos calculus defines communication by unification of names, so that
two occurrences of a name could be identified by reduction. If x has to be identified with
itself, it is just considered as a dummy operation in a unification setting since we already
have x = x. The situation is different with differential interaction nets since communication
areas keep track of such an identification through an explicit link connecting the communication area associated with x to itself. Reduction is not able to simplify such a link in
order to give back a communication area. One gets some more complicated structure.
In [EL10] we were directly working with the π-calculus (thus with name passing by
substitution, thus without self-identification). This way a direct interpretation of the whole
(finitary) π-calculus into differential interaction nets was possible. The approach we propose
here is different: we look for a restriction of the solos calculus (avoiding self-identification)
for which the static correspondence with differential interaction nets also works at the
dynamic level. This is the acyclic solos calculus. This requires us to restrict the source
calculus but allows us to deal with a very natural translation.
The acyclic solos calculus is obtained from a precise analysis of the translation of the πcalculus into the solos calculus. The main idea is to be able to isolate, inside the unification
mechanism which is a symmetric operation (when one unifies a name x with a name y), a
“flow” from one side to the other (from x to y for example) as it happens in a substitution
mechanism (in the reduction of uhxi.P | u(y).Q in the π-calculus, one can consider a flow
going from x to y, with the binding occurrence y “controlling” all the other occurrences
of y in Q). The first ingredient in our definition of the acyclic solos calculus is a simple
typing system allowing us to label all the object occurrences of names of a solos term with
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
3
two protocols (S (send) and R (receive)) in a kind of uniform way. The main consequence
of the typing system is to break symmetry in reduction/unification: when two occurrences
of names have to be unified one is an S-occurrence and the other one is an R-occurrence.
However this does not impose important structural or behavioural constraints on protocols.
This is given by the second part of the definition of the calculus: the acyclicity conditions
(AC1–AC5). The justification for these conditions is to mimic the structure of the input
prefix u(y).Q of the π-calculus inside the solos calculus: exactly one binding occurrence y
(this will be the R-occurrence of the name y in the solos term), all the other occurrences
(thus S-occurrences for object occurrences in acyclic solos) are coming “sequentially after”
the binding one, ... Informally, all this together leads to a name passing by unification
which is very close to a substitution mechanism if we understand the unification of an
S-occurrence with an R-occurrence as a substitution flow from the S-occurrence to the Roccurrence (and then all the other occurrences of this name in S-position). Formally, we
prove that self-identification never occurs in the acyclic solos calculus.
We also show that the translation of the (finitary) π-calculus into the solos calculus
following [LV03] has its range contained in the acyclic solos calculus. This proves that our
restriction of the solos calculus still has a reasonable expressive power. Moreover this gives
us an alternative proof (with respect to the work presented in [EL10]) of the existence of a
translation of the finitary π-calculus into differential interaction nets.
The first part of the paper is devoted to the introduction of the required calculi (πcalculus and solos calculus in Section 1, solo diagrams in Section 2, and differential interaction nets in Section 3 which is almost copy-pasted from [EL10]). In Section 4 we
elaborate on the material presented in [EL10] to define the simple static translation of solo
diagrams [LPV01] (the graphical syntax for the solos calculus) into differential interaction
nets. We discuss the problems arising at the dynamic level, and we give a sufficient condition
on solo diagrams for the translation into differential interaction nets to be a bisimulation:
an occurrence of a name should never be unified with another occurrence of the same name
(i.e. no self-identification).
The main technical contribution of the paper comes in the last part (Section 5). Since
the property that there is no self-identification in the redexes of a solos term is of course not
preserved under the reduction of solos, we have to find a more clever property. We define
the acyclic solos calculus through a typing system for assigning protocols and the acyclicity
conditions (AC1–AC5). We prove this restriction to be well behaved with respect to the
reduction of the solos calculus. We show that the translation of a π-term is always an acyclic
solos term, showing the expressiveness of the system. Finally we prove that the sufficient
condition introduced in Section 4 is fulfilled by solo diagrams corresponding to acyclic solos
terms, showing that we obtain a bisimulation with respect to differential interaction nets.
1. The π-calculus and the solos calculus
In this section we recall the definition of the π-calculus and of the solos calculus we are
going to use. We also recall the translation from π to solos given in [LV03].
In the π-calculus [MPW92], the two basic components of communication are the output
and input prefixes uhxi.P and v(y).Q (with occurrences of y bound in Q). Communication
can occur if u = v and in this case, it is obtained by substituting y by x (in Q). Moreover
these prefix constructions induce a sequential behaviour: the continuation P (resp. Q)
4
T. EHRHARD AND O. LAURENT
cannot interact with other agents before the output on u (resp. the input on v) has been
triggered.
The fusion calculus [PV98] provides a generalization with communication modelized
by unification. This makes sending and receiving perfectly symmetric and thus the distinction between u x.P and u x.P is purely formal. Interaction is just constrained to occur
between two dual entities u and u. The symmetry is broken when one translates π-terms
into fusion terms by: u(x).P 7→ (x) u x.P (and unification restrained to the image of this
translation becomes substitution by properties of the binding restriction (x) P ). Notice that
the sequentiality aspect of prefixing remains.
Based on the unification mechanism provided by the fusion calculus, the solos calculus [LV03] is free from explicit sequentiality constructions: the fusion prefixes are restricted
to dummy continuations u x.0 and u x.0 (where 0 is an inactive process). It is shown
in [LV03] that dyadic solos u xy and u xy are as expressive as the whole fusion calculus.
Comments on chosen calculi. Since our goal is to focus on prefixing and sequentiality, we
deal with calculi without replications nor recursive definitions, without match/mismatch
and without sums (see Conclusion for additional comments).
We do not want to spend time to deal with arbitrary arities in the calculi we consider.
This is why we only consider monadic π-terms. There are three reasons for that: it makes
the presentation simpler, it does not lead to a loss of expressiveness, and finally the polyadic
case has already been considered in [EL10]. As a consequence (see the translation in Section 1.3), we are led to consider a triadic solos calculus (all the names are of arity exactly
3) and triadic solo diagrams (all the multiedges are of arity exactly 3). The more general
case of arbitrary arities could easily be obtained by introducing the appropriate sortings on
the various calculi.
1.1. The π-calculus. The terms of the (monadic, finitary) π-calculus are given by:
P
::= 0 | u(x).P
| uhxi.P
| (P | P ) | νx.P
where both u(x).P and νx.P bind occurrences of x in P .
The structural congruence on π-terms is the least congruence containing α-equivalence
and:
0|P ≡P
P |Q≡Q|P
(P | Q) | R ≡ P | (Q | R)
νx.νy.P ≡ νy.νx.P
νx.0 ≡ 0
(νx.P ) | Q ≡ νx.(P | Q)
if x ∈
/ fn(Q)
The reduction semantics of the π-calculus is given by:
uhxi.P | u(y).Q → P | Q [x/y]
P →Q
P |R→Q|R
P →Q
νx.P → νx.Q
P ≡ P′
P ′ → Q′
P →Q
Q′ ≡ Q
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
5
1.2. The solos calculus. Introduced in [LV03], the goal of the solos calculus is to prove
the expressiveness of a calculus without prefix construction.
The terms of the (triadic) solos calculus are given by:
P
::= 0 | u x1 x2 x3 | u x1 x2 x3 | (P | P ) | (x) P
where (x) P binds occurrences of x in P .
The structural congruence is the least congruence containing α-equivalence and:
0|P ≡P
P |Q≡Q|P
(P | Q) | R ≡ P | (Q | R)
(x) (y) P ≡ (y) (x) P
(x) 0 ≡ 0
((x) P ) | Q ≡ (x) (P | Q)
if x ∈
/ fn(Q)
This equivalence allows us to present terms in the solos calculus in canonical forms: either
the 0 process or a bunch of scope constructions (x1 ) (x2 ) . . . followed by solos in parallel.
The reduction semantics of the solos calculus is given by (z̃ stands for z1 . . . zn ):
(z̃) (u x1 x2 x3 | u y1 y2 y3 | P ) → P σ
where σ is a most general unifier of x1 x2 x3 and y1 y2 y3 , such
that exactly the names w in z̃ are modified, i.e. satisfy
σ(w) 6= w (in particular, in each equivalence class of names
induced by unification, at most one name is free).
P →Q
P →Q
P ≡ P′
P ′ → Q′
Q′ ≡ Q
P |R→Q|R
(x) P → (x) Q
P →Q
An alternative (but equivalent) definition of this reduction semantics is given in [LV03]
together with additional explanations.
For example, in (x) (y) (z) (w) (u uxy | u zww | v zuy), the only possible reduction is
between u uxy and u zww. It induces the identifications u = z, x = w and y = w, thus
two equivalence classes {u, z} and {x, y, w}. In the first one, u is free and thus the only
possibility is to map z to u. In the second one, all the elements are bound, we choose one
of them: y for example (the other choices would lead to structurally congruent results). We
consider the unifier containing z 7→ u, x 7→ y, w 7→ y and which is the identity on the other
names. We obtain the reduct (y) v uuy.
1.3. From π-terms to solos. In [LV03], the authors give different translations of the
fusion calculus [PV98] into solos. We are going to focus on one of them (the one which does
not introduce matching). By pre-composing this translation with the canonical embedding
of the π-calculus into the fusion calculus: u(x).P 7→ (x) u x.P , we obtain the translation of
the π-calculus into solos that we present here.
6
T. EHRHARD AND O. LAURENT
A π-term P is translated as [P ]:
Cv := (z) v zzv
[0]v := 0
[u(x).P ]v := (w) (y) (v uwy | Cy | (x) (v ′ ) (w xvv ′ | [P ]v′ ))
[uhxi.P ]v := (w) (y) (v uwy | Cy | (v ′ ) (w xv ′ v | [P ]v′ ))
[P | Q]v := [P ]v | [Q]v
[νx.P ]v := (x) [P ]v
[P ] := (v) ([P ]v | Cv )
As shown in [LV03], this encoding is adequate with respect to weak barbed congruence ≈:
[P ] ≈ [Q] implies P ≈ Q.
The π-term νx.(uhxi.0 | x(y).0) | u(z).zhti.0 is translated as a solos term which is
structurally congruent to (v) ((x) (P | Q) | R | Cv ) with:
P = (w) (y) (v uwy | Cy | (v ′ ) w xv ′ v)
Q = (w) (y ′ ) (v xwy ′ | Cy′ | (y) (v ′ ) w yvv ′ )
R = (w) (y) (v uwy | Cy | (z) (v ′ ) (w zvv ′ | (w) (y) (v ′ zwy | Cy | (v ′′ ) w tv ′′ v ′ )))
By applying reduction, we obtain in particular the following reducts (up to structural congruence):
(v) ((x) ((Cv | (v ′ ) u xv ′ v) | Q) | R)
(v) ((x) (((v ′ ) u xv ′ v) | Q) | Cv | (z) (v ′ ) (u zvv ′ | (w) (y) (v ′ zwy | Cy | (v ′′ ) w tv ′′ v ′ )))
(v) (x) (Q | Cv | ((w) (y) (v xwy | Cy | (v ′′ ) w tv ′′ v)))
(v) (x) (Cv | (y) (v ′ ) x yvv ′ | ((w) (y) (v xwy | Cy | (v ′′ ) w tv ′′ v)))
(v) (x) ((y) (v ′ ) x yvv ′ | Cv | (v ′′ ) x tv ′′ v)
(v) Cv
2. Solo diagrams
To make the relation with differential interaction nets simpler, we go through the graphical syntax associated with the solos calculus: solo diagrams [LPV01].
A (triadic) solo diagram is given by a finite set of nodes and a finite multiset of (ternary)
multiedges (directed edges with a list of three nodes as source and one node as target). A
node is tagged as either free or bound. A multiedge is tagged as either input or output. Any
node must be a source or a target of a multiedge.
2.1. Reduction. The reduction of solo diagrams is given:
• by choosing two multiedges e1 and e2 of opposite polarities (one input and one output)
with the same target (we call them dual multiedges) and with respective sources [n1 , n2 , n3 ]
and [m1 , m2 , m3 ],
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
7
reduces to
Figure 1: A reduction in solo diagrams.
• and by identifying the two nodes n1 and m1 , the two nodes n2 and m2 , and the two nodes
n3 and m3 . The reduction is allowed to occur only if it does not identify two free nodes.
Moreover when a free node is identified with a bound node, the obtained node is free and
when two bound nodes are identified, the obtained node is bound.
The chosen multiedges are erased and nodes which are not anymore source or target of a
multiedge are also removed. This can be applied to both free and bound nodes.
During a reduction step, the number of multiedges decreases by two and the number
of nodes decreases or remains the same.
In the graphical representation, we draw free nodes as white dots, bound nodes as
black dots, output edges with an outgoing arrow and input edges with an ingoing arrow.
An example is given in Figure 1.
2.2. Relation with solos. A term of the solos calculus can easily be translated into a solo
diagram. Free names are translated as free nodes, bound names as bound nodes and solos
as multiedges:
• The 0 term is translated as the empty graph with no edge.
• The solo u x1 x2 x3 is translated as the graph with free nodes corresponding to elements of
{u, x1 , x2 , x3 }, and with one input multiedge with source [x1 , x2 , x3 ] and target u (where
xi is the node corresponding to the name xi and u is the node corresponding to the name
u).
• The solo u x1 x2 x3 is translated in the same way with an output multiedge.
• The parallel composition P | Q is obtained by “graph union”: union of the sets of nodes
(i.e. nodes corresponding to the same name are identified) and sum of the multisets of
multiedges.
• The restriction (x) P corresponds to turning the free node x corresponding to the name
x (if any) into a bound node.
As shown in [LPV01], two solos terms are structurally congruent if and only if the corresponding solo diagrams are isomorphic. Moreover reduction in solo diagrams reflects
faithfully the reduction of the solos calculus.
The two solo diagrams of Figure 1 correspond to (x) (y) (z) (w) (u uxy | u zww | v zuy),
respectively, (y) v uuy (the examples used in Section 1.2).
2.3. Labeled transition system. We follow the same methodology as in [EL10] by using
labels to distinguish the actions of a process. In a term syntax as in [EL10], labels are put
on prefixes (or solos). In the graphical syntax of solo diagrams, we put labels on multiedges
(since multiedges correspond to solos).
8
T. EHRHARD AND O. LAURENT
We fix a countable set L of labels to be used for all our labeled transition systems.
A solo diagram is labeled if its multiedges are equipped with different labels belonging
to L.
Definition 2.1 (The transition system SL ). The objects of the labeled transition system
SL are labeled solo diagrams and transitions are labeled by pairs of distinct elements of L.
lm
Let G and H be two labeled solo diagrams, G −→ H if H is obtained from G by applying a
reduction step to the input multiedge labeled l and to the output multiedge labeled m (the
labels of the remaining multiedges of H must be the same as for the corresponding ones in
G).
This is quite different from what is usually done for defining transition systems for a process algebra [MPW92]. The standard approach aims at analyzing the possible interactions
of a process with its environment. Then bisimulation is used to show that two processes
have the same “external” behaviour. Our goal is to use bisimulation to compare the internal behaviours of solo diagrams and of differential interaction nets. We want to illustrate
that differential interaction nets are sufficiently expressive for simulating concurrency and
mobility. This requires the transition systems we use for bisimulation to carry precise informations about reductions of processes, in particular about the sub-entities taking part
into the interactions.
2.4. Solo diagrams with identifications. In order to compare solo diagrams with differential interaction nets, it will be useful to decompose the reduction of solo diagrams by
introducing the notion of solo diagrams with identifications.
Definition 2.2 (Solo diagram with identifications). Solo diagrams with identifications are
solo diagrams equipped with a finite set of undirected edges (usual binary edges, not multiedges, which connects nodes of the solo diagram). These edges are called identification
edges.
We can decompose the reduction of solo diagrams we have presented above by using
identification edges. A reduction step for a solo diagram G is defined as follows:
(R1) choose two dual multiedges e1 and e2 (i.e. of opposite polarities and with the same
target) with respective sources [n1 , n2 , n3 ] and [m1 , m2 , m3 ];
(R2) build the solo diagram with identifications G[e1 , e2 ] obtained from G by erasing e1
and e2 and by introducing three identification edges: between n1 and m1 , between n2
and m2 , and between n3 and m3 ;
(R3) contract the graph G[e1 , e2 ] by (repeatedly) choosing an identification edge and by
identifying the two (or one) nodes it connects if at least one of them is bound.
The reduction succeeds (i.e. is a valid reduction of solo diagrams) if we reach a solo diagram
with no remaining identification edge. It means in particular that step (2) does not introduce
identification edges between two free names.
In the graphical representation, we draw identification edges as dashed edges.
If we refine the reduction of the example of Figure 1 by means of solo diagrams with
identifications, the results of step (2) and then of one application of step (3) are in Figure 2.
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
l
9
l
Figure 2: An (3) reduction step in solo diagrams with identifications.
3. Differential Interaction Nets
3.1. The general formalism of interaction nets. We recall the general (graphical)
syntax of interaction nets, as introduced in [Laf95]. See also [Maz06, ER06] for more
details.
Statics of interaction nets. We first define a general typed syntax of interaction nets.
Assume we are given a set of symbols and that an arity (a non-negative integer) and
a typing rule is associated with each symbol. This typing rule is a list (A0 , A1 , . . . , An ) of
types, where n is the arity associated with the symbol. Types are formulae of some system
of linear logic (in particular if A is a type, A⊥ is also a type and A⊥⊥ = A). An interaction
net is made of cells. With each cell γ is associated exactly one symbol and therefore an
arity n and a typing rule (A0 , A1 , . . . , An ). Such a cell γ has one principal port p0 and n
auxiliary ports p1 , . . . , pn . An interaction net has also a finite set of free ports. All these
ports (the free ports and the ports associated with cells) have to be pairwise distinct and
a set of wires is given. This wiring is a set of pairwise disjoint sets of ports of cardinality
2 (ordinary wires) or 0 (loops1), and the union of these wires must be equal to the set of
all ports of the interaction net. In other words, each port of the interaction net (free or
associated with a cell) is connected to exactly one other port (free or associated with a cell)
through a wire, and each such wire connects exactly two ports: ports cannot be shared.
The free ports of the interaction net are those which are not associated with a cell.
An oriented wire of the interaction net is an ordered pair (p1 , p2 ) where {p1 , p2 } is a
wire. In an interaction net, a type is associated with each oriented wire, in such a way
that if A is associated with (p1 , p2 ), then A⊥ is associated with (p2 , p1 ). Last, the typing
rules of the cells must be respected in the sense that for each cell γ of arity n, whose
ports are p0 , p1 , . . . , pn and typing rule is (A0 , A1 , . . . , An ), denoting by p′0 , p′1 , . . . , p′n the
ports of the interaction net uniquely defined by the fact that the sets {pi , p′i } are wires
(for i = 0, 1, . . . , n), then the oriented wires (p0 , p′0 ), (p′1 , p1 ),. . . ,(p′n , pn ) have types A0 ,
A1 ,. . . ,An respectively. The free ports of the interaction net constitute its interface. With
each free port p can be associated the type of the unique oriented wire whose endpoint is
p: this is the type of p in the interface of the interaction net.
Here is a typical example of a typed interaction net:
A
•
α
B
C
•
β
E
D
•
γ
F
p
q
1To be more precise, one has to specify the number of loops in the net, but this will not play any role in
the sequel.
10
T. EHRHARD AND O. LAURENT
with cells of symbols α, β and γ, of respective types (B, A⊥ , C ⊥ ), (B ⊥ , A, E ⊥ , D ⊥ ) and
(F, D, C). The interface is (p : E, q : F ). Cells are represented as triangles, with principal
port located at one of the angles and other ports on the opposite edge. We often draw a
black dot to locate the auxiliary port number 1.
Dynamics of interaction nets. In Lafont’s traditional interaction nets, a reduction rule associates, to each pair of cells connected through their principal ports, an interaction net
with the same interface as this two-cells net. Moreover at most one rule is given for each
pair of cells. The application of a reduction rule substitutes (inside an interaction net) a
pair of cells connected through their principal ports by the associated reduct.
These ideas are strongly related with key steps of cut elimination in logics, where two
rules interact through their principal formulas.
Extensions of interaction nets. Since interaction nets have strong confluence properties,
extensions are required to take concurrency and non-determinism into account [Ale99].
A possible way of extending interaction nets for more concurrent behaviours is to modify
statics by introducing cells with many principal ports [Ale99, Kha03, Maz05] (this breaks
simple logical interpretations by contradicting the usual property that a logical rule of the
sequent calculus introduces one principal formula). In differential interaction nets, we stick
to the traditional statics syntax (but act on the dynamics side).
A possible extension of the dynamics towards non-deterministic behaviours is to allow
a choice between different possible reduction rules for a given pair of cells [Ale99].
We make a choice in-between this proposal and Lafont’s one (i.e. between many possible
rules, and at most one rule). We consider only one reduction rule for each pair of cells but
we extend interaction nets to formal sums of interaction nets [ER06]. This allows us to
represent, as one reduction, a finite set of non-deterministic choices: when the reduct of the
interaction of two cells is a (proper) sum of interaction nets.
3.2. Presentation of the cells. We define differential interaction nets from the interaction nets defined above. We first present our specific cells.
The untyped lambda-calculus can be seen as a typed lambda-calculus equipped with
a recursive equation D = D ⇒ D. Following this idea, V. Danos and L. Regnier [Reg92]
defined a notion of untyped proof-nets based on a single type symbol o (the type of outputs),
subject to the following recursive equation o = ?o⊥ `o. Our nets will be typed using this
typing system. We set ι = o⊥ , so that ι = !o ⊗ ι and o = ?ι`o. The tensor connective is
used only with premises !o and ι and dually for the par, and therefore, the only types we
actually need are o, ι, !o and ?ι for typing our nets.
In the present setting, there are ten symbols: par (arity 2), bottom (arity 0), tensor
(arity 2), one (arity 0), dereliction (arity 1), weakening (arity 0), contraction (arity 2),
codereliction (arity 1), coweakening (arity 0) and cocontraction (arity 2). We present now
the various cell symbols, with their typing rules, in a pictorial way.
Multiplicative cells. The par and tensor cells, as well as their “nullary” versions bottom and
one are as follows:
?ι
o
•
`
o
!o
ι
•
⊗
ι
⊥
o
1
ι
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
11
Exponential cells. They are typed according to a strictly polarized discipline. Here are first
the why not cells, which are called dereliction, weakening and contraction:
ι
?
?ι
?
?ι
?ι
?
?ι
?ι
and then the bang cells, called codereliction, coweakening and cocontraction:
o
!
!o
!
!o
!o
!
!o
!o
Differential interaction nets. We define differential interaction nets from the cells above.
• A simple differential interaction net is a typed interaction net, which uses the multiplicative and exponential cells introduced above.
• A differential interaction net is a finite formal sum S = s1 + · · · + sn (n ≥ 0) of simple
differential interaction nets having all the same interface, and this interface is then considered as the interface of S. A particular case is the differential interaction net S = 0
(the empty sum), and this differential interaction net has to be given together with its
interface: there is a 0 net for each interface.
Labeled differential interaction nets. We now introduce labeled differential interaction nets,
which are differential interaction nets where particular cells are equipped with labels. The
labeled transition system of differential interaction nets will be defined using these labels in
Section 3.3.3.
Let Lτ be the countable set of labels obtained by adding a distinguished element τ (to
be understood as the absence of label) to L (introduced in Section 2.3). A labeled simple
differential interaction net is a simple differential interaction net where all dereliction and
codereliction cells are equipped with labels belonging to Lτ . Moreover, if l and l′ are two
labels appearing in a labeled simple differential interaction net, either l 6= l′ or l = l′ = τ .
In our pictures, the labels of dereliction and codereliction cells will be indicated, when
this label is different from τ . When its label is τ , a (co)dereliction cell will be drawn without
any label.
All the differential interaction nets we consider in this paper are labeled. In the sequel,
since no confusion with other kinds of interaction nets will be possible, we shall use “net”
for “labeled differential interaction net” and “simple net” for “labeled simple differential
interaction net”.
3.3. Reduction rules. We denote by ∆ the collection of all simple nets, ranged over by
the letters s, t, u, with or without subscripts or superscripts, and by Nh∆i the collection
of all nets (finite sums of simple nets with the same interface, including the empty sum),
ranged over by the letters S, T , U , with or without subscripts or superscripts. We consider
∆ as a subset of Nh∆i (s ∈ ∆ is identified with the sum made of exactly one copy of s).
A reduction rule is a subset R of ∆ × Nh∆i consisting of pairs (s, S) where s is a
simple net made of two cells connected by their principal ports and S is a net that has the
same interface as s. There are actually reduction rules which transform simple nets into
non-simple ones.
This set R can be finite or infinite. Such a set is easily extended to a relation between
arbitrary simple nets and nets and then to arbitrary nets and nets:
12
T. EHRHARD AND O. LAURENT
• We note s R T if there is (s0 , u1 + · · · + un ) ∈ R where s0 is a subnet of s, each ui is a
simple net and T = t1 + · · · + tn where ti is the simple net resulting from the replacement
of s0 by ui in s (n ≥ 0).
• The relation RΣ is defined by s1 +· · ·+sn RΣ T (where each si is simple) if T = T1 +· · ·+Tn
where, for each i, si R Ti or si = Ti (n ≥ 0).
Last, the relation R∗ between nets and nets is the transitive closure of RΣ (which is reflexive).
3.3.1. Defining the reduction. We give now the reduction rules of differential interaction
nets.
Multiplicative reduction. The first two rules concern the interaction of two multiplicative
cells of the same arity.
?ι
o
•
`
⊗
?ι
•
?ι
o
o
⊥
m
o
1
ǫ
m
o
where ǫ stands for the empty simple net (i.e. the simple net containing no cell, no port
and thus no wire, not to be confused with the net 0 ∈ Nh∆i, the empty sum, which is not
a simple net). The next two rules concern the interaction between a binary and a nullary
multiplicative cell.
?ι
o
1
`
?ι
!
!o
o
1
ι
⊗
m
o
ι
⊥
!o
?
ι
⊥
m
Communication reduction. Let R ⊆ Lτ . We have the following reductions if l, m ∈ R.
ι
?ι
?
l
!
ι
ι
c,R
m
The labels l and m disappear in this reduction step.
Non-deterministic reduction. Let R ⊆ Lτ . We have the following reductions if l ∈ R.
ι
?
?ι
!
l
o
!
?
l
nd,R
ι
?ι
!o
!
?
nd,R
!o
?ι
!
?
l
!o
l
ι
?
?ι
o
!
l
nd,R
0
?ι
ι
+
?ι
l
?
!o
o
+
!o
o
?
l
!
l
!o
?ι
!o
!o
!
!
?ι
?
nd,R
0
According to the definition of reduction for arbitrary simple nets given above, the application of one of the last two steps to a simple net erases the whole simple net and gives the
0 net (as everywhere else in the paper, this means the 0 net with appropriate interface).
These two steps make some labels disappear.
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
13
Structural reduction.
?ι
?
?ι
!
?ι
!
!o
!
s
?ι
?ι
!o
?
?ι
?
?ι
?ι
!
s
ǫ
?
?ι
?ι
?ι
!
?ι
We use ∼s for the symmetric and transitive closure of
!o
?
?ι
!
s
?ι
?
s
!o
!
!o
!
?ι
?ι
?ι
?
?
?ι
?ι
s.
Remark 3.1. One can check that we have provided reduction rules for all redexes compatible with our typing system: for any simple net s made of two cells connected through
their principal ports, there is a reduction rule whose left member is s. This rule is unique,
up to the choice of a set of labels, but this choice has no influence on the right member of
the rule.
3.3.2. Confluence.
Theorem 3.2. Let R, R′ ⊆ Lτ . Let R ⊆ ∆ × Nh∆i be the union of some of the reduction
relations c,R , nd,R′ , m and s . The relation R∗ is confluent on Nh∆i.
The proof is essentially trivial since the rewriting relation has no critical pair (see [ER06]).
Given R ⊆ Lτ , we consider in particular the following R-reduction: R = m ∪ c,{τ } ∪
s∪
nd,R . We set
d =
∅ (“d” for “deterministic”) and denote by ∼d the symmetric
and transitive closure of this relation. Observe that, if s and S are nets with s simple and
if s d S, then S is also simple.
Some of the reduction rules we have defined depend on a set of labels. This dependence
is clearly monotone in the sense that the relation becomes larger when the set of labels
increases.
The reduction of nets is not normalizing mainly because of the last structural reduction
rule. Additional comments can be found in [ER06].
3.3.3. A transition system of simple nets. Let l and m be distinct elements of L. We call
(l, m)-communication redex a communication redex whose codereliction cell is labeled by l
and whose dereliction cell is labeled by m.
Definition 3.3 (The transition system DL ). We define a labeled transition system DL
whose objects are simple nets, and transitions are labeled by pairs of distinct elements of
lm
L. Let s and t be simple nets, we have s −→ t if the following holds: s ∗{l,m} s0 + S where
s0 is a simple net which contains an (l, m)-communication redex and becomes t when one
reduces this redex.
Remark 3.4. The non-deterministic steps allowed in the reduction from s to s0 +S can only
involve the codereliction and dereliction labeled by l and m respectively. The communication
steps only involve τ -labeled derelictions and coderelictions. In the solos calculus, solos
communicate in one step through a parallel composition. This single step becomes here a
14
T. EHRHARD AND O. LAURENT
sequence of many elementary steps and our restriction allows us to avoid considering the
steps which have nothing to do with the communication we are interested in.
The net S contains other possible communications, so it corresponds to other branches
of non-deterministic choices.
In this paper, to compare transition systems, we only consider strong bisimulations [Par81]
as given by the following definition.
Definition 3.5 (Bisimulation). Given two labeled transition systems LTS1 = (S1 , T1 ) (with
T1 ⊆ S1 × L × S1 ) and LTS2 = (S2 , T2 ) (with T2 ⊆ S2 × L × S2 ) on the same set L of labels,
a relation R between S1 and S2 is a bisimulation between LTS1 and LTS2 if:
• for any (s1 , s2 ) ∈ R and (s1 , l, s′1 ) ∈ T1 , there exists s′2 ∈ S2 such that (s2 , l, s′2 ) ∈ T2 ;
• and for any (s1 , s2 ) ∈ R and (s2 , l, s′2 ) ∈ T2 , there exists s′1 ∈ S1 such that (s1 , l, s′1 ) ∈ T1 .
A useful particular case is LTS1 = LTS2 .
Lemma 3.6. The relation ∼d ⊆ ∆ × ∆ is a bisimulation on DL .
3.4. A toolbox for process calculi interpretation. We introduce now a few families of
simple nets, which are built using the previously introduced basic cells. They will be used
as basic modules for interpreting processes. All of these nets, but the communication areas,
can be considered as compound cells: in reduction, they behave in the same way as cells of
interaction nets.
One could have directly defined these compound cells as primitive constructions for the
representation of processes. However we prefer to stress the possibility of implementing
these compound objects with more basic ones coming from the logically founded theory of
differential linear logic, in order to emphasize the possible application of linear logic tools
to the image of our encoding.
Communication areas cannot be considered as cells in our setting since they would
require more than one principal port. Interaction nets with many principal ports have been
studied (in particular for representing concurrent behaviours [Maz05]) but their theory is
quite different from the “one principal port” case.
3.4.1. Compound cells.
Generalized contraction and cocontraction. A generalized contraction cell is a simple net t.
One of its free ports (which is not connected to an auxiliary port of its cells) is called the
principal port of t. The other free ports (which are finitely many and are not connected
to principal ports of cells of the generalized cell) are called the auxiliary ports of t. These
generalized contraction cells are inductively defined:
• a wire with type ?ι is a generalized contraction cell (we select the port with type ?ι as
the principal port of the generalized cell);
• a weakening cell gives a generalized contraction cell by connecting a wire to its principal
port (the only free port of the obtained net is the principal port of the generalized cell);
• given two generalized contraction cells, by connecting the two auxiliary ports of an additional contraction cell to the principal ports of the generalized cells and by connecting
the principal port of this new contraction cell to a free port, one obtains a generalized
contraction cell (the free port connected to the principal port of the added contraction
cell is the principal port of the generalized cell).
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
!o
!o
..
.
•
?⊗
!o
?ι
=
l
•
•
!o
!o
•
⊗
ι
⊗
15
?ι
?
l
⊗
1
Figure 3: Dereliction-tensor compound cell.
..
.
•
..
= .
!
l
?⊗
..
.
•
..
.
!`
l
•
..
= .
?
l
?⊗
!`
..
.
•
?⊗
l
!`
Figure 4: Input and output compound cells.
3
Figure 5: Area of order 3.
Generalized cocontraction cells are defined dually.
We use the same graphical notations for generalized (co)contraction cells as for ordinary
(co)contraction cells, with a “∗” in superscript to the “!” or “?” symbols to avoid confusions.
Observe that there are infinitely many generalized (co)contraction cells of any given arity.
Even if one could define a choice of one canonical (co)contraction cell for each arity, pluging
two canonical (co)contraction cells together would not always give us a canonical one. This
is why we do not define such a choice here.
The dereliction-tensor and the codereliction-par cells. Let n be a non-negative integer. We
define an n-ary ?⊗ compound cell as in Figure 3. It will be decorated by the label of its
dereliction cell (as mentioned in Section 3.2, we put no decoration if the label is τ ). The
number of tensor cells in this compound cell is equal to n (in particular the 0-ary derelictiontensor cell contains exactly two cells: a one cell and a dereliction cell). We define dually
the !` compound cell.
The prefix cells. Now we can define the compound cells which will play the main role in
the interpretation of solos. Thanks to the above defined cells, all the oriented wires of the
nets we shall define will have type ?ι or !o. Therefore, we adopt the following graphical
convention: the wires will have an orientation corresponding to the ?ι type.
Let n be a non-negative integer. The n-ary input cell and the n-ary output cell are
defined in Figure 4, they have n pairs of auxiliary ports (δ1+ , δ1− , . . . , δn+ , δn− ).
The label of a prefix cells is the one carried by its outermost ?⊗ or !` compound cell,
the other ?⊗ or !` compound cells are required to be unlabeled (that is, labeled by τ ).
16
T. EHRHARD AND O. LAURENT
?∗
!∗
?∗
!∗
?∗
!∗
!∗
?∗
?∗
!∗
!∗
?∗
Figure 6: Communication areas of order −1, 0 and 1.
..
.
n1
n2
..
.
∗
s
..
.
n1 + n2
..
.
Figure 7: Aggregation, with n1 , n2 ≥ −1.
3.4.2. Communication areas. Let n ≥ −2. We define a family of nets with 2(n + 2) free
ports, called communication areas of order n, that we shall draw using rectangles with
beveled angles. Figure 5 shows how we picture a communication area of order 3.
A communication area of order n is made of n + 2 pairs of (n + 1)-ary generalized
+
−
cocontraction and contraction cells (γ1+ , γ1− ), . . . , (γn+2
, γn+2
), with, for each i and j such
that 1 ≤ i < j ≤ n + 2, a wire from an auxiliary port of γi+ to an auxiliary port of γj−
−
and a wire from an auxiliary port of γi− to an auxiliary port of γj+ . We note p+
i and pi
−
the principal ports of γi+ and γi− , and we call (p+
i , pi ) a pair of associated ports of the
communication area.
So the communication area of order −2 is the empty net ǫ, and communication areas
of order −1, 0 and 1 are the structures shown in Figure 6.
3.4.3. Useful reductions. One of the nice properties of communication areas is that, when
one connects two such areas through a pair of wires, one gets another communication area.
Lemma 3.7 (Aggregation of communication areas). Let C1 be a communication area of
−
order n1 ≥ −1 and C2 be a communication area of order n2 ≥ −1, let (p+
1 , p1 ) be a pair of
−
associated ports of C1 and (p+
2 , p2 ) be a pair of associated ports of C2 , let s be the simple net
+
−
+
∗ C where C is a communication
obtained by connecting p1 to p2 and p−
s
1 to p2 , we have s
area of order n1 + n2 (see Figure 7).
Proof. We only prove the particular case where n1 = n2 = 1 and the communication areas
C1 and C2 are built with contraction cells (not generalized ones).
By connecting C1 and C2 , and by applying two structural reduction steps, we obtain a
communication area of order 2 (see Figure 8).
Let t be a simple net and p be a port of t. We say that p is forwarded in t if there is a
free port q of t such that t is of one of the two shapes given in Figure 9.
This allows us to describe the interaction of derelictions and coderelictions with communication areas: derelictions and coderelictions can meet each other, when connected to
a common communication area.
Lemma 3.8 (Communication and forwarding of derelictions and coderelictions in communication areas). Let p ≥ −2 be an integer, let l, m ∈ Lτ , and let us consider the simple net
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
!
?
?
!
!
!
?
?
!
∗
s
!
?
!
?
?
?
!
!
?
!
?
?
!
?
!
17
?
!
!
?
Figure 8: Aggregation of communication areas of order 1.
p
p
.. ..
. .
q
?∗
.. ..
. .
q
!∗
···
···
Figure 9: Port forwarding.
···
···
!
?
!
···
l
p+2
m
?
∗
{l,m}
!
l
p
m
?
+
N
X
i=1
!
l r
ti
m
r′
?
Figure 10: Dereliction and codereliction communicating through a communication area.
s obtained by connecting (the principal ports of ) a codereliction labeled l and a weakening
to a pair of associated ports of a communication area of order p + 2, and by connecting (the
principal ports of ) a dereliction labeled m and a coweakening to another pair of associated
ports of the same communication area. The {l, m}-reduction applied to s leads to a sum of
simple nets: one contains the codereliction l and the dereliction m connected through their
principal ports, together with a communication area of order p; in all the other summands,
the principal ports r and r ′ of the codereliction l and of the dereliction m are forwarded (see
Figure 10).
Proof. We consider the particular case where p = −1 and the communication area is built
with contraction cells (not generalized ones). The reduction is pictured in Figure 11.
We first apply all the possible structural reduction steps, then we focus on the codereliction labeled l: the reduction of the associated redex gives a sum of two simple nets (in
one of them the principal port of the codereliction l is forwarded). We now apply in each
simple net the reduction step involving the dereliction labeled m and we finally obtain a
sum of two simple nets. In the first one the principal ports of the dereliction and of the
codereliction are forwarded, in the second one they are connected to each other and the
remaining part of the net is a communication area of order −1.
The generalization to other communication areas of order 1 is easy. We can obtain the
case p 6= −1 by decomposing the communication area into two communication areas (one
of order 1 and one of order p + 1). Lemma 3.7 and Theorem 3.2 help to conclude.
We now consider the interaction of two prefix cells.
18
T. EHRHARD AND O. LAURENT
?
∗
s
!
?
!
?
!
!
?
!
?
l
∗
{l}
?∗
l
!
m
!∗
!
?
!∗
?
m
∗
{m}
?
!
!
m
!∗
l
?∗
+
!∗
!
?∗
?
l
?∗
+
!
?
m
l
!∗
?∗
m
!∗
!
?
l
m
Figure 11: Forwarding of dereliction and codereliction.
•
..
.
•
l
!
?
..
.
c,{l,m}
u
∗
∅
..
.
m
Figure 12: Prefix reduction.
Lemma 3.9 (Reduction of prefixes). Let l, m ∈ Lτ . If we connect an n-ary output prefix
labeled by m to a p-ary input prefix labeled by l through their principal ports, we obtain a
simple net which reduces by c,{l,m} to a net u which reduces by ∗{τ } to 0 if n 6= p and to
simple wires by ∗∅ , as in Figure 12, if n = p.
Proof. The key point is to check that the simple net, obtained by connecting an n-ary
dereliction-tensor compound cell to a p-ary codereliction-par cell through their principal
ports, reduces to min(n, p) wires with n − p weakening cells if n > p or p − n coweakening
cells if p > n.
4. From solo diagrams to differential interaction nets
4.1. The translation. Relying on the toolbox described above, we define a translation of
labeled solo diagrams with identifications into labeled simple differential interaction nets:
• A node which appears n times as a source of a multiedge, which is k times target of
a multiedge and which is p times member of an identification edge is translated as a
communication area of order n + k + p − 2.
For example, for a node which is target of one input multiedge and of one output
multiedge (thus k = 2), which is source of one multiedge (thus n = 1), and which is
member of one identification edge (thus p = 1), we have n + k + p − 2 = 2:
!
2
• An input multiedge with sources [x1 , x2 , x3 ] and target u is translated as a 3-ary input
cell with principal port u0 and auxiliary ports [x01 , x11 , x02 , x12 , x03 , x13 ]. Each pair x0i , x1i is
connected to a pair of associated ports of the communication area corresponding to xi .
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
19
0
−1
0
0
? l
!
−1
Figure 13: The translation of the second solo diagram of Figure 2.
u0 is connected to the port p− and a weakening cell is connected to p+ , where (p+ , p− )
is a pair of associated ports of the communication area corresponding to u. The label of
the prefix cell is the label of the multiedge.
l
l
!
!
?
• An output multiedge with sources [x1 , x2 , x3 ] and target u is translated as a 3-ary output
cell with principal port u0 and auxiliary ports [x01 , x11 , x02 , x12 , x03 , x13 ]. Each pair x0i , x1i is
connected to a pair of associated ports of the communication area corresponding to xi .
u0 is connected to the port p+ and a coweakening cell is connected to p− , where (p+ , p− )
is a pair of associated ports of the communication area corresponding to u. The label of
the prefix cell is the label of the multiedge.
l
l
!
?
!
• An identification edge connecting x1 and x2 is translated as a pair of wires connecting a
−
pair (p+
1 , p1 ) of associated ports of the communication area corresponding to x1 with a
− +
pair (p2 , p2 ) of associated ports of the communication area corresponding to x2 .
!
Since communication areas of order n are not uniquely defined (because generalized contraction and cocontraction cells of a given arity are not unique either), this translation from
solo diagrams to differential interaction nets is in fact a relation, which we denote G ! s
(with G a solo diagram and s a simple differential interaction net).
Remark 4.1. Let G be a solo diagram (without identification edges) and s be a differential
interaction net, if G ! s then s has no d redex, except some s redexes involving weakening or coweakening cells. Indeed, all the redexes in s are given by a labeled codereliction
cell (i.e. whose label is not τ ) or a coweakening cell facing a weakening cell, a contraction
cell or a labeled dereliction cell (or dually by a dereliction cell or a weakening cell facing a
coweakening or ...).
A differential interaction net associated with the second solo diagram with identifications of Figure 2 is given in Figure 13.
20
T. EHRHARD AND O. LAURENT
n1
C1
n2
C2
!
↓
↓
!
n
C
Figure 14: Proof of Lemma 4.2.
4.2. A bisimulation. Our goal is to establish a bisimulation between the labeled transition
systems SL (Section 2.3) and DL (Section 3.3.3). Unfortunately the translation ! does
not provide such a bisimulation. We are going to show what the problems are and how to
restrict SL to get a bisimulation.
4.2.1. Mismatch. The crucial point comes from step (3) of the reduction of solo diagrams:
the contraction of identification edges in solo diagrams corresponds through ! to the
aggregation of communication areas as given by the following lemma.
Lemma 4.2. If G1 and G2 are solo diagrams with identifications, if G2 is obtained from
G1 by contracting an identification edge (step (3)) connecting the nodes n1 and n2 with
n1 6= n2 and if G1 ! s1 then G2 ! s2 where s2 is obtained from s1 by aggregating the
communication areas corresponding to n1 and n2 .
Proof. See Figure 14: G2 is obtained from G1 by replacing the two distinct nodes n1 and
n2 by a unique node n. In s1 , we have a communication area C1 corresponding to n1 and
a communication area C2 corresponding to n2 . Since n1 and n2 are connected with an
identification edge, C1 and C2 are connected with a pair of wires. By aggregating C1 and
C2 (see Lemma 3.7), we obtain a communication area C in s2 which exactly corresponds to
n through !.
The hypothesis n1 6= n2 is crucial since two communication areas C1 and C2 connected
together with a pair of wires reduce to one communication area (Lemma 3.7), except if
C1 = C2 ! We are thus able to encode the reduction of the solos calculus only if we can
ensure that we never have to contract an identification edge connecting a node with itself.
A typical example of this problem is given by:
(u) (x) (y) (z) (y ′ ) (z ′ ) (a) (b) (c) (a′ ) (b′ ) (c′ ) (u xyz | u xy ′ z ′ | x abc | x a′ b′ c′ )
As shown in Figure 15, while in solo diagrams the loop created after the first step eventually
disappears, it remains in differential interaction nets. The consequence being that we have
transitions in SL while the last differential interaction net has no transition to a differential
interaction net which is the translation of a solo diagram.
4.2.2. Restriction. Since our goal is to make the bisimulation result true, we are going to
constrain the syntax of solo diagrams in such a way that the hypotheses of Lemma 4.2
are always valid. We want to restrict the reduction step (3) to the case where the edge
is not connecting a node with itself. This is equivalent to asking the sub-graph Gid [e1 , e2 ]
of G[e1 , e2 ], containing the identification edges only, to be acyclic. By definition, we call
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
−1
l
′
′
m
l′
−1
?
−1
!
21
−1
m′
!
−1
?
−1
2
l
l
m
−1
?
m
!
0
!
−1
−1
(2)
+
{l,m}
−−→
−1
l
′
−1
?
′
m
l′
−1
?
−1
!
−1
m′
!
−1
?
−1
2
−1
−1
−1
−1
(3)2
+
d
−−→
−1
l′
l′
m′
−1
?
−1
!
−1
m′
!
−1
?
−1
2
(3)
−−→
l′
m′
(2)
−−→
(3)3
−−→
∅
Figure 15: Reductions in solo diagrams and in differential interaction nets, corresponding
to the term:
(u) (x) (y) (z) (y ′ ) (z ′ ) (a) (b) (c) (a′ ) (b′ ) (c′ ) (u xyz | u xy ′ z ′ | x abc | x a′ b′ c′ ).
22
T. EHRHARD AND O. LAURENT
acyclic reduction step a reduction step associated with two dual multiedges e1 and e2 such
that Gid [e1 , e2 ] is acyclic.
Another problem comes from the constraint on the freeness of names in the contraction
of identification edges in solo diagrams. Our translation into differential interaction nets
is forgetting the fact that some nodes are free and others are bound. As a consequence a
reduction might happen in the differential interaction net’s side without being possible in
the solo diagram’s side. An example is given by the term u xyz | u x′ y ′ z ′ . To avoid this
situation we introduce the notion of acyclic redex.
Definition 4.3 (Acyclic redex). In a solo diagram, a pair of dual multiedges is an acyclic
redex if it is a redex (meaning that the induced freeness conditions are satisfied) and the
induced reduction step is an acyclic reduction step.
ac
Definition 4.4 (Sac
L labeled transition system). SL is the biggest labeled transition system
such that:
ac
• each object of Sac
L is an object of SL and each transition of SL is a transition of SL (but
some objects and transitions are lost),
ac
• for any object in Sac
L , all the transitions starting from it in SL belong to SL ,
• all pairs of dual multiedges in objects of Sac
L are acyclic redexes.
This means that an object G in SL belongs to Sac
L as soon as none of the paths starting
from G in SL allows us to reach an object with a non-acyclic redex. This is a decidable
property since there are finitely many such paths.
Definition 4.5 (Acyclic solo diagram). A solo diagram which is an object of Sac
L is called
an acyclic solo diagram.
lm
Proposition 4.6. If G −→ H in Sac
L and G ! s, then there are simple nets t and t0 such
lm
that s −→ t0 in DL , H ! t and t0
d
t.
Proof. By hypothesis, there is a communication area C in s with two dual prefixes labeled
with l and m connected to it. We apply the forwarding of derelictions and coderelictions
in the communication area C to them (Lemma 3.8). This gives us a sum of simple nets
containing a simple net s1 with an (l, m)-communication redex and other summands where
l and m have been forwarded.
Following Figure 16, let t0 be the simple net obtained by reducing the (l, m)-communication
redex of s1 , and t1 be the simple net obtained by applying to s1 the prefix reduction starting
with the reduction from s1 to t0 (see Lemma 3.9). It is easy to see that G[e1 , e2 ] ! t1
where e1 and e2 are the multiedges of G with labels l and m. By Lemma 4.2, each step
of acyclic contraction of an identification edge corresponds to the aggregation of the two
associated communication areas, thus H ! t with t1 ∗d t and we have t0 ∗d t.
Lemma 4.7 (Diving). Let G be a solo diagram (without identification), if G ! s and
∗
s
{l,m} t + T with t containing an (l, m)-communication redex, then the codereliction
labeled l and the dereliction labeled m are connected to the same communication area C in s
and the (l, m)-communication redex is generated by the dereliction/codereliction forwarding
through C (Lemma 3.8).
Proof. According to Remark 4.1, the only {l,m} redexes of s are involving the codereliction
l, the dereliction m, a weakening or a coweakening.
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
•
•
?⊗
..
.
`
!
`
?
l
⊗
•
⊗
m
•
!`
⊥
•
?⊗
..
.
∗
d
..
.
1
•
c,{l,m}
⊥
`
`
⊗
23
(s1 )
•
⊗
•
!`
1
..
.
..
.
∗
d
(t0 )
(t1 )
(t)
..
.
Figure 16: Simulation of a solos reduction step.
By confluence (Theorem 3.2), we can assume the reductions involving the codereliction
l are coming before those involving m (and those involving the weakeaning and coweakening cells arrive after them). We look at the port connected to the principal port of the
codereliction l:
(l1) If it is the auxiliary port of a prefix cell, there is no reduction for l, and no {l, m}reduction will change the situation (or this prefix cell is labeled m and changing the
situation means erasing m).
(l2) If it is the auxiliary port of a generalized cocontraction cell (which is not a wire), there
is no reduction for l.
(l3) If it is the principal port of a prefix cell labeled k (k 6= m), there is no {l, m}-reduction
for l, and no {l, m}-reduction will change the situation.
(l4) If it is the principal port of the prefix cell labeled m, we immediately have the result,
and C is a communication area of order 0 (with the part connecting l and m which is
just a simple wire).
(l5) If it is the principal port of a generalized contraction cell of arity 0, reductions involving
l will eventually erase it and we will never reach an (l, m)-communication redex, this
contradicts the hypotheses.
(l6) If it is the principal port of a generalized contraction cell of arity n > 0, in order to
reach an (l, m)-communication redex, the reductions involving l must start destroying
the generalized contraction. In order to eventually reach an (l, m)-communication
redex, we must have a simple net corresponding to case 6 again, or to case 2 or 4.
We now consider these last two configurations: 6 cases followed by 2 (see Figure 18) or 4
(see Figure 17). Note that we cannot have infinitely many consecutive 6 cases since the
induced reduction makes the size of the net decrease. If after 6 cases we reach case 4, l and
m where connected to the same communication area C and we have the result (6 cases give
a forwarding of l through C). If after 6 cases we reach case 2, we turn our attention to the
reduction steps involving m. We look at the port connected to the principal port of the
dereliction m:
24
T. EHRHARD AND O. LAURENT
?∗
!
l
..
.
?
m
∗
{l}
!
?
l
m
6
4
Figure 17: Proof of Lemma 4.7: sequence 6∗ -4.
!
l
?∗
..
.
..
.
!∗
∗
{l}
?
m
6
!
l
..
.
!∗
∗
{m}
?
m
2/5
!
?
l
m
5
Figure 18: Proof of Lemma 4.7: sequence 6∗ -2-5+ .
(m1) If it is the auxiliary port of a prefix cell, there is no reduction for m, and no {l, m}reduction will change the situation (or this prefix cell is labeled l and changing the
situation means erasing l).
(m2) If it is the auxiliary port of a generalized contraction cell (which is not a wire), there
is no reduction for m (there could be reductions involving a coweakening but which
would eventually erase m), and we do not reach an (l, m)-communication redex, this
contradicts the hypotheses.
(m3) If it is the principal port of a prefix cell labeled k (k 6= l), there is no {l, m}-reduction
for m, and we do not reach an (l, m)-communication redex, this contradicts the hypotheses.
(m4) If it is the principal port of a generalized cocontraction cell of arity 0, reductions
involving m will eventually erase it and we will never reach an (l, m)-communication
redex, this contradicts the hypotheses.
(m5) If it is the principal port of a generalized cocontraction cell of arity n > 0, in order to
reach an (l, m)-communication redex, the reductions involving m must start destroying the generalized cocontraction. In order to reach an (l, m)-communication redex
we must escape the four cases above. This means that we go to case 5 again until
we reach the simplest 5 case: m facing l (note that we cannot have infinitely many
consecutive 5 cases since the size of the net decreases).
We conclude that, in order to reach an (l, m)-communication redex, we must go through a
sequence of configurations 6∗ -4 (see Figure 17) or 6∗ -2-5+ (see Figure 18). For l and m to
face each other after such a sequence, they must be connected to the same communication
area C in s and the reduction sequence corresponds to the forwarding of l and m through
C.
lm
Proposition 4.8. If s −→ t0 in DL and G ! s with G ∈ Sac
L , then there are a solo diagram
lm
H and a simple net t such that G −→ H in Sac
L , H ! t and t0
d
lm
t.
Proof. Notations coincide with Figure 16. By definition of s −→ t0 , there exist a simple
net s1 containing an (l, m)-communication redex and a net S such that s ∗{l,m} s1 + S
and t0 is obtained from s1 by reducing this (l, m)-communication redex. By Lemma 4.7, G
contains two dual multiedges e1 and e2 connected to the same node and labeled l and m.
Let t1 be the simple net obtained by applying to s1 the prefix reduction starting with the
reduction from s1 to t0 (see Lemma 3.9), we first show that G[e1 , e2 ] ! t1 . By Lemma 4.7
again, the only way the (l, m)-communication redex can have been generated in s1 is by the
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
25
forwarding of the codereliction l and the dereliction m through the communication area (of
order p+2) they are both connected to. This gives a prefix redex and a communication area
of order p (Lemma 3.8). By firing this prefix redex, we generate pairs of wires connecting
communication areas which exactly correspond to the identification edges of G[e1 , e2 ].
Let H be the solo diagram (without identifications) obtained by firing the redex between
e1 and e2 in G (this is possible since it is an acyclic redex). We iteratively apply Lemma 4.2
starting from G[e1 , e2 ] and t1 until we reach H. This is possible since G ∈ Sac
L . Let t be the
simple net obtained this way and corresponding to H, we have t0 d t.
To get our bisimulation result, we define !d as the composition of ! and ∼d : G !d s
if there exists a simple net s0 such that G ! s0 and s0 ∼d s.
Theorem 4.9 (Bisimulation). !d is a bisimulation between Sac
L and DL .
lm
Proof. If G −→ H in Sac
L , G ! s0 and s0 ∼d s, by Proposition 4.6 there are simple nets t
lm
and t0 such that s0 −→ t0 in DL , H ! t and t0 d t. By Lemma 3.6, there exists a simple
lm
net t1 such that s −→ t1 and t0 ∼d t1 . As a consequence t ∼d t1 thus H !d t1 .
lm
Conversely, if s −→ t1 in DL , G ! s0 and s0 ∼d s, by Lemma 3.6 there exists a simple
lm
net t0 such that s0 −→ t0 and t1 ∼d t0 . By Proposition 4.8, there are a solo diagram H and
lm
a simple net t such that G −→ H in Sac
d t. As a consequence t ∼d t1
L , H ! t and t0
thus H !d t1 .
G !
s0
∼d
s
lm
lm
lm
H !
t0 ∼ d t1
≀d
t
The rest of the paper will be devoted to the definition of a sub-calculus of the solos
calculus whose translation into solo diagrams lives inside Sac
L.
5. The acyclic solos calculus
By following the approach of the previous section and by analyzing the translation of
the π-calculus into the solos calculus, we define a sub-system of the solos calculus called the
acyclic solos calculus. The key (informal) properties of this calculus are:
• expressiveness: it contains the image of the π-calculus through the translation into solos
(Sections 5.1.2 and 5.2.2).
• stability: it is well-defined with respect to the reduction of solos since it is stable under
reduction (Propositions 5.4 and 5.10).
• acyclicity: the solo diagram associated with a term of the acyclic solos calculus is an
acyclic solo diagram (and thus the bisimulation with differential interaction nets holds!)
(Section 5.3).
The definition of the acyclic solos calculus is given in two steps: first by means of a
typing system assigning Send/Receive polarities to occurrences of names, second by structural constraints on typed terms. Intuitively, we restrict terms to a forest-like structure
26
T. EHRHARD AND O. LAURENT
with reduction occurring only between roots of the trees of the forest (in fact we will deal
with more general structures than forests). This is an abstraction of the forest structure
induced by the sequentiality of a π-term on its translation into solos. The Send/Receive
polarities on their side represent the input/output asymmetry of the π-calculus in the solos
setting.
5.1. Types. We consider a system with only two types V and W . We use U for denoting
either V or W .
A typed term is a term with scopes decorated with types: (xU ) P .
A typing judgment is of the shape Γ ⊢ P where Γ contains typing declarations associating types with names and P is a typed term. The typing rules are:
Γ, x : U ⊢ P
Γ⊢P
Γ⊢Q
Γ⊢0
Γ⊢P |Q
Γ ⊢ (xU ) P
Γ, x : V, y : V, z : W, z ′ : W ⊢ x̂ zz ′ y
Γ, x : W, z : W, y : V, y ′ : V ⊢ x̂ zyy ′
where x̂ is either x or x .
An intuitive way of understanding the types V and W is given by these two mutually
recursive definitions:
V ::= W W V
W ::= W V V
(more generally, starting from other mutually recursive definitions of a finite set of types,
one could derive other typing systems of the same kind which would also satisfy the subject
reduction property for example).
The main purpose of this typing system is to be able to associate a communication
protocol with each object occurrence. Communication protocols are S (send) and R (receive).
If we add communication protocols as a decoration on the definitions of V and W :
V ::= W R W S V S
W ::= W R V S V R
then any object occurrence of a typed term in context can be decorated with a communication protocol: in an input solo whose subject has type U , the decoration is given according
to those of the components of the type, in an output solo it is given in a dual way: S 7→ R
and R 7→ S. More formally, we can directly enrich typing derivations in such a way that
they assign communication protocols to object occurrences of names in a typed term:
Γ, x : U ⊢ P
Γ⊢P
Γ⊢Q
Γ⊢0
Γ⊢P |Q
Γ ⊢ (xU ) P
Γ, x : V, y : V, z : W, z ′ : W ⊢ x z R z ′S y S
Γ, x : V, y : V, z : W, z ′ : W ⊢ x z S z ′R y R
Γ, x : W, z : W, y : V, y ′ : V ⊢ x z R y S y ′R
Γ, x : W, z : W, y : V, y ′ : V ⊢ x z S y R y ′S
Examples are given by the image of the translation of the π-calculus in Section 5.1.2.
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
27
The intuition behind S and R comes from the difference between the name-passing
mechanism in the π-calculus and in the solos calculus. Name-passing by substitution (as
in π) entails that a flow of information is given by interaction from an output prefix to
an input prefix (receiving names are substituted with sending names). Name-passing by
unification (as in fusion or solos) establishes a perfect symmetry between the agents trying to
communicate. Communication protocols S and R are used to trace the flow of information
coming from π after translation into solos. Even if communication is symmetric (inside
the solos calculus), it is reasonable for a term decorated with communication protocols
to understand the unification of xS with y R as y being substituted with x (this will be a
consequence of the constraints given in Definition 5.6).
5.1.1. Preservation by reduction. We first check elementary properties of the typing system
with respect to reduction.
Lemma 5.1 (Free names). If Γ ⊢ P then all the free names of P appear in Γ.
Lemma 5.2 (Weakening). If x ∈
/ Γ and x ∈
/ fn(P ), Γ ⊢ P if and only if Γ, x : U ⊢ P .
Lemma 5.3 (Substitution). If Γ, x : U, y : U ⊢ P then Γ, x : U ⊢ P [x/y].
Proof. These three lemmas are proved by simple inductions on the typing derivations.
Proposition 5.4 (Subject reduction).
(1) If P ≡ Q then Γ ⊢ P ⇔ Γ ⊢ Q.
(2) If P reduces to Q and Γ ⊢ P then Γ ⊢ Q.
Proof.
(1) We simply have to consider each case of the definition of ≡ (given in Section 1.2).
The only interesting one is ((x) P ) | Q ≡ (x) (P | Q) with x ∈
/ fn(Q). We assume
Γ ⊢ ((xU ) P ) | Q, this entails Γ ⊢ (xU ) P and Γ ⊢ Q and thus Γ, x : U ⊢ P . By
Lemma 5.2, we have Γ, x : U ⊢ Q thus Γ, x : U ⊢ P | Q and finally Γ ⊢ (xU ) (P | Q).
Conversely, if Γ ⊢ (xU ) (P | Q) then Γ, x : U ⊢ P | Q, Γ, x : U ⊢ P , Γ, x : U ⊢ Q thus,
by Lemma 5.2 with x ∈
/ fn(Q), Γ ⊢ Q and:
Γ, x : U ⊢ P
Γ ⊢ (xU ) P
Γ⊢Q
((xU ) P )
Γ⊢
|Q
(2) The only interesting case is (z̃) (u x1 x2 x3 | u y1 y2 y3 | P ) → P σ where z̃ contains exactly
the names such that σ(w) 6= w. From the hypothesis, we deduce Γ, z̃ : Ũ ⊢ u x1 x2 x3
and Γ, z̃ : Ũ ⊢ u y1 y2 y3 and Γ, z̃ : Ũ ⊢ P . By Lemma 5.3, we have Γ ⊢ P σ.
5.1.2. Typability of the translation. We prove the typability of the translation of any π-term
(see Section 1.3 for the definition of the translation).
If P is a π-term with free names in the finite set X = {x1 , . . . , xk }, we define ΓX to be
the environment x1 : W, . . . , xk : W and ΓX ,v to be the environment ΓX , v : V .
Proposition 5.5 (Typability). If P is a π-term with free names in the finite set X , ΓX ⊢
[P ].
28
T. EHRHARD AND O. LAURENT
Proof. We first show that ΓX ,v ⊢ [P ]v .
The process Cv is typable:
ΓX ,v , z : W ⊢ v z R z S v S
ΓX ,v ⊢ (z W ) v z R z S v S
The translations of prefixes are typable (see Figure 19).
The cases of P | Q and νx.P are immediate.
Finally we get the typability of [P ]:
ΓX ,v ⊢ [P ]v
ΓX ,v ⊢ Cv
ΓX ,v ⊢ [P ]v | Cv
ΓX ⊢ (v V ) ([P ]v | Cv )
We can sum up the induced communication protocols:
Cv = (z W ) v z R z S v S
[u(x).P ]v = (wW ) (y V ) (v uS wR y R | Cy | (xW ) (v ′V ) (w xR v S v ′R | [P ]v′ ))
[uhxi.P ]v = (wW ) (y V ) (v uS wR y R | Cy | (v ′V ) (w xS v ′R v S | [P ]v′ ))
[P | Q]v = [P ]v | [Q]v
[νx.P ]v = (xW ) [P ]v
[P ] = (v V ) ([P ]v | Cv )
5.2. Acyclicity. Relying on the decoration with communication protocols induced by the
typing system, we are now able to define our restriction of the solos calculus.
Given a typed term P , we first define some properties and relations between names and
solos occurring in P . If s is a solo, we write subj(s) for its subject.
x∈s
xX ∈ s
s⊳x
s⊳t
s⊥t
s is a root
:=
:=
:=
:=
:=
:=
x has an object occurrence in s
x has an object occurrence in s with communication protocol X
xR ∈ s
s ⊳ subj(t)
subj(s) = subj(t), one is output and the other is input
there is no t such that t ⊳ s
We denote by ⊳+ the transitive closure of ⊳ and by ⊳⋆ the reflexive transitive closure of ⊳.
Definition 5.6 (Acyclic solos calculus). Acyclic solos terms are typed terms which satisfy
the following five properties:
(AC1)
(AC2)
(AC3)
(AC4)
(AC5)
each name has at most one R-occurrence
s ⊥ t implies that s and t are roots
s ⊳ x and x ∈ t implies s ⊳⋆ t
xS ∈ s and s ⊳ x implies s is an input
there is no free name with an Roccurrence
ΓX ,v , y : V, w : W, v ′ : V, x : W ⊢ w xR v S v ′R
ΓX ,v , y : V, w : W, v ′ : V, x : W ⊢ [P ]v′
′
ΓX ,v , y : V, w : W, v : V, x : W ⊢ w xR v S v ′R | [P ]v′
ΓX ,v , y : V, w : W ⊢ Cy
ΓX ,v , y : V, w : W ⊢ v uS wR y R | Cy
ΓX ,v , y : V, w : W, x : W ⊢ (v ′V ) (w xR v S v ′R | [P ]v′ )
ΓX ,v , y : V, w : W ⊢ (xW ) (v ′V ) (w xR v S v ′R | [P ]v′ )
ΓX ,v , y : V, w : W ⊢ v uS wR y R | Cy | (xW ) (v ′V ) (w xR v S v ′R | [P ]v′ )
ΓX ,v , w : W ⊢ (y V ) v uS wR y R | Cy | (xW ) (v ′V ) (w xR v S v ′R | [P ]v′ )
ΓX ,v ⊢ (wW ) (y V ) v uS wR y R | Cy | (xW ) (v ′V ) (w xR v S v ′R | [P ]v′ )
ΓX ,v , y : V, w : W, v ′ : V ⊢ w xS v ′R v S
ΓX ,v , y : V, w : W ⊢ v uS wR y R
ΓX ,v , y : V, w : W ⊢ Cy
S
R R
ΓX ,v , y : V, w : W ⊢ v u w y | Cy
ΓX ,v , y : V, w : W, v ′ : V ⊢ [P ]v′
ΓX ,v , y : V, w : W, v ′ : V ⊢ w xS v ′R v S | [P ]v′
ΓX ,v , y : V, w : W ⊢ (v ′V ) (w xS v ′R v S | [P ]v′ )
ΓX ,v , y : V, w : W ⊢ v uS wR y R | Cy | (v ′V ) (w xS v ′R v S | [P ]v′ )
ΓX ,v , w : W ⊢ (y V ) v uS wR y R | Cy | (v ′V ) (w xS v ′R v S | [P ]v′ )
ΓX ,v ⊢ (wW ) (y V ) v uS wR y R | Cy | (v ′V ) (w xS v ′R v S | [P ]v′ )
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
ΓX ,v , y : V, w : W ⊢ v uS wR y R
Figure 19: Typing derivations for the translations of prefixes.
29
30
T. EHRHARD AND O. LAURENT
The acyclic solos calculus is given by restricting terms to acyclic solos terms. The structural
congruence and the reduction relation are those induced by the solos calculus.
(1)
(2)
(3)
(4)
(5)
We make a few remarks and state immediate consequences of this definition:
Let us consider the following definition of communication protocols for occurrences of
names in a π-term: if the occurrence is an object occurrence in an input prefix, the
communication protocol is R, otherwise it is S. Assuming that the names appearing in
a π-term are made as different as possible by α-conversion, we can remark that each
name has at most one R-occurrence, and no free name has an R-occurrence.
(AC2) entails that reduction only occurs between roots.
In the case where ⊳ induces a forest ordering, (AC3) means that all the S-occurrences
of x are bigger (with respect to this forest ordering) than its R-occurrence (if any).
(AC4) says that a name can have both an S-occurrence and an R-occurrence in a solo s
only if s is an input. This is a place where we are breaking the symmetry of the solos
calculus between output solos and input solos. This allows us to rule out processes
like x y S y R uR | x z R z S v S that would lead to twice the identification of y with z which is
almost the same as identifying y with y (and would break acyclicity of the associated
solo diagram).
(AC5) ensures (with (AC1)) that identification edges are always contractible (no freeness
problem).
5.2.1. Preservation by reduction. We want to prove that reducts of an acyclic solos term
under the reduction of the solos calculus are all acyclic solos terms (Proposition 5.10). We
consider a reduction from P to Q by an interaction between the solos s0 and t0 of P . By
definition of reduction (see Section 1.2), we have a substitution σ such that each solo t of
Q is of the shape sσ for some solo s of P . We say that t is a residue of s. The solos s0 and
t0 of P have no residue in Q while all the other solos of P have a unique residue in Q. In
this section, in order to simplify the notations, when P reduces to Q we will use the same
name s for a solo in P and for its unique residue in Q if it exists.
To make the difference between the relations ⊳ in different processes, we use the notation
⊳P for the ⊳ relation of the process P .
The following three lemmas hold assuming only conditions (AC1), (AC2) and (AC3)
on P .
Lemma 5.7. If P is a typed term of the solos calculus satisfying (AC1), (AC2) and (AC3)
and which reduces to Q, then no R-occurrence of Q has been substituted during reduction.
Proof. Let s0 and t0 be the solos destroyed during the reduction from P to Q, if a name x
involved in the reduction (i.e. taking part in the unification, i.e. occurring in s0 or t0 ) has
its R-occurrence in a solo t of P different from s0 and t0 then t ⊳P x and, by (AC3), either
+
t ⊳+
P s0 or t ⊳P t0 thus one of them is not a root of P contradicting (AC2).
Lemma 5.8. If P is a typed term of the solos calculus satisfying (AC1), (AC2) and (AC3)
and which reduces to Q, if s is a solo in P whose subject has been modified during this
reduction ( i.e. the subject of s in P is not the same as the subject of its residue in Q), we
have both:
• (the residue of ) s is a root in Q
• s is a root in P or s0 ⊳P s or t0 ⊳P s
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
31
Proof. Let x be the subject of s in P and y be its subject in Q, each of x and y appears in
either s0 or t0 . If s is not a root in Q then there is some t such that t ⊳Q s thus t ⊳P y (by
Lemma 5.7). If y ∈ s0 (and symmetrically if y ∈ t0 ) then t ⊳⋆P s0 (by (AC3)) with t 6= s0
since t belongs to Q thus s0 is not a root in P contradicting (AC2). So that s is a root in
Q.
If s is not a root in P then there is some t such that t ⊳P s. Since the subject of s is
modified, it means that an R-occurrence in t is also substituted thus t ∈
/ Q by Lemma 5.7
and t = s0 or t = t0 .
Lemma 5.9. If P is a typed term of the solos calculus satisfying (AC1), (AC2) and (AC3)
and which reduces to Q, if the reduction introduces an object occurrence of the name x in
Q then x has no R-occurrence in Q.
Proof. For the reduction to put some x in some t, x must occur in s0 or t0 . Assume that
x occurs in s0 and that x has an R-occurrence in s in Q (so that s ⊳Q x). By Lemma 5.7,
s ⊳P x and by (AC3), we have s ⊳⋆P s0 but, since s 6= s0 (because s0 ∈
/ Q), this would entail
that s0 is not a root in P contradicting (AC2).
We now turn to the preservation result.
Proposition 5.10 (Acyclicity preservation). If P is an acyclic solos term and P reduces
to Q then Q is an acyclic solos term.
Proof. We first prove the preservation of conditions (AC1), (AC2) and (AC3), independently
of conditions (AC4) and (AC5).
Preservation of condition (AC1) is given by Lemma 5.7.
We move to condition (AC2). If s ⊥Q s′ , note that if one of them is not a root then
none of them is since t ⊳ s ⇒ t ⊳ subj(s) ⇒ t ⊳ subj(s′ ) ⇒ t ⊳ s′ . Assume that t ⊳Q s and
t ⊳Q s′ , if the subject of neither s nor s′ has been modified during reduction, then t ⊳P s
and t ⊳P s′ (since the R-occurrence in t of the subject of s and s′ has not been modified, by
Lemma 5.7) and P contradicts condition (AC2). If one of the subjects of s or s′ has been
modified then it is a root in Q by Lemma 5.8.
Concerning condition (AC3), we assume s ⊳Q x and x ∈Q t. By Lemma 5.7, s ⊳P x and
by Lemma 5.9, x ∈P t. So that, by (AC3), s ⊳⋆P t. If s6 ⊳⋆Q t, then the path from s to t for the
relation ⊳P has been broken during reduction, this entails that a solo t′ 6= s in this path has
got his subject modified. By Lemma 5.8, s0 ⊳P t′ or t0 ⊳P t′ but this is impossible because,
since s0 and t0 are roots in P and are not in Q, s 6= s0 and s 6= t0 .
We now assume that (AC1), (AC2) and (AC3) hold. Condition (AC4) is preserved:
if, in Q, an output solo s contains both an R-occurrence and an S-occurrence of a name x
then, by Lemmas 5.7 and 5.9, it is also the case in P . Condition (AC5) is preserved by
Lemma 5.7.
All this shows that our acyclic solos calculus is well behaved with respect to the reduction of the solos calculus.
5.2.2. Acyclicity of the translation. In order to prove the expressiveness of the acyclic solos
calculus, we check that it contains the image of the translation of the π-calculus.
32
T. EHRHARD AND O. LAURENT
The ⊳ relation is the following for translated π-terms:
Cv = (z W ) v z R z S v S
[0]v = 0
[u(x).P ]v = (wW ) (y V ) (v uS wR y R | Cy | (xW ) (v ′V ) (w xR v S v ′R | [P ]v′ ))
[uhxi.P ]v = (wW ) (y V ) (v uS wR y R | Cy | (v ′V ) (w xS v ′R v S | [P ]v′ ))
[P | Q]v = [P ]v | [Q]v
[νx.P ]v = (xW ) [P ]v
[P ] = (v V ) ([P ]v | Cv )
where an arrow s
t represents the relation s ⊳ t.
Theorem 5.11 (Acyclicity for the π-calculus). If P is a π-term, [P ] is an acyclic solos
term.
Proof. By Proposition 5.5, [P ] is typable.
Let us mention a few additional facts (which can be easily checked by induction on the
definition of [P ]v ):
(i) The free names of [P ]v are v and the free names of P .
(ii) The free names of P have no R-occurrence in [P ]v .
(iii) The only subject with no R-occurrence (thus the only subject of a root) in [P ]v is v.
(iv) ⊳ is a forest on [P ]v (which is one of the reasons for the name “acyclic solos”).
From these points, conditions (AC1), (AC2), (AC3), (AC4) and (AC5) are easily verified:
• We first check the five conditions for [P ]v :
(1C1) By induction on [P ]v by using facts i, ii and iii which entail that no free name in
[P ]v has an R-occurrence in [P ]v .
(2C2) An easy induction with fact i shows that s ⊥ t entails subj(s) = subj(t) = v and
we conclude with fact iii.
(3C3) By induction, using facts iii and iv.
(4C4) Immediate induction.
(5C5) A consequence of facts ii, i and iii.
• It is then easy to check the case of [P ].
5.3. Back to acyclic solo diagrams. The last property we want to show is that the solo
diagram associated with a term of the acyclic solos calculus (Definition 5.6) is an acyclic
solo diagram (Definition 4.5).
Let G be such a solo diagram associated with the term P of the acyclic solos calculus,
we have to show that any pair of dual edges in G is an acyclic redex (Lemma 5.12). This
will be enough to show that G belongs to Sac
L (Theorem 5.14).
Lemma 5.12. Let G be the solo diagram associated with an acyclic solos term P and
let e1 and e2 be two dual multiedges of G, this pair of multiedges defines an acyclic redex
(Definition 4.3).
We want to show that the graph Gid [e1 , e2 ] (sub-graph of G[e1 , e2 ] with identification
edges only, as introduced in Section 4.2.2) is acyclic and does not generate any freeness
problem in the contraction of identification edges.
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
33
By construction, edges in Gid [e1 , e2 ] are connecting two nodes which correspond to
occurrences of names in the term P which are going to be unified. By definition of the
acyclic solos calculus, one of these occurrences is an S-occurrence and the other one is an Roccurrence. We consider the directed graph G′ obtained by orienting the edges of Gid [e1 , e2 ]
towards R-occurrences.
We first prove the following lemma about directed and non-directed graphs.
Lemma 5.13. Let g be a finite directed graph and gs be the underlying non-directed graph.
If gs is connected, if g has a root (that is a node without incoming edge) and if each node
of g has at most one incoming edge then gs is acyclic.
Proof. The function which maps each edge of g to its target is an injective function (by
hypotheses) and its image is the set of nodes of g which are not roots. As a consequence,
the number of edges of g (and thus of gs ) is less than or equal to the number of vertices of
g (and thus of gs ) minus one. Since gs is connected, we conclude that it is acyclic.
We can now prove Lemma 5.12:
Proof. We prove that the connected components of Gid [e1 , e2 ] equipped with the orientation
of G′ satisfy the hypotheses of Lemma 5.13 and thus that Gid [e1 , e2 ] is acyclic. Let us assume
e1 is the input multiedge in the redex between e1 and e2 in G (since solos in a solos term
are in one-to-one correspondence with multiedges in the corresponding solo diagram, we
use the notations introduced in the beginning of Section 5.2 with multiedges as well as with
solos), we have the following properties:
(i) if xR ∈ e1 then all the other occurrences of x in e1 and e2 are S-occurrences in e1 :
by (AC1), x cannot have another R-occurrence, moreover, if xS ∈ e2 then, by (AC3),
e1 ⊳+ e2 thus e2 is not a root of P , contradicting (AC2).
(ii) if xR ∈ e2 then x has no other occurrence neither in e1 nor in e2 : as for i, x cannot
have any other R-occurrence nor any S-occurrence in e1 and finally, by (AC4), it cannot
have an S-occurrence in e2 .
(iii) if xS ∈ e2 then x has no R-occurrence neither in e1 nor in e2 : otherwise we apply i
or ii.
We consider a connected component G0 of Gid [e1 , e2 ]. It is the underlying non-directed
graph of the appropriate sub-graph G′0 of G′ . G′0 is finite, G0 is connected and, according
to condition (AC1), each node has at most one incoming edge in G′0 . We just have to show
that G′0 has a root, i.e. that there is a name in P which has only S-occurrences in e1 and e2
for each connected component G0 . We first show that any R-occurrence in e2 is connected
to an S-occurrence of a name without R-occurrence: we start from this R-occurrence in e2 ,
the corresponding occurrence in e1 is an S-occurrence of a name x, if x has no R-occurrence
we are done, and if x has an R-occurrence then it is also in e1 (by ii) and the corresponding
occurrence in e2 is an S-occurrence (thus without R-occurrence by iii). Since any connected
component contains either an R-occurrence in e2 or an S-occurrence in e2 , we are done
(with iii again).
Finally by (AC5), G′ has the property that only roots can be free nodes (according
to the free/bound labeling of nodes in G). This property is preserved by contraction of
identification edges thanks to (AC1). Moreover this property entails that at most one
extremity of an identification edge can be free, and thus no freeness problem arises during
contraction of identification edges.
34
T. EHRHARD AND O. LAURENT
Theorem 5.14 (Acyclic diagrams from acyclic terms). The solo diagram associated with
an acyclic solos term (Definition 5.6) is an acyclic solo diagram (Definition 4.5).
Proof. We prove that the range of the translation of the acyclic solos calculus into solo diagrams satisfies the conditions of Definition 4.4. By Lemma 5.12, we know such a translation
contains only acyclic redexes. Now if G is the translation of the acyclic solos term P and
reduces to H, there exists a term Q which translates into H, and such that P reduces to Q
thus Q is an acyclic solos term (Proposition 5.10).
Conclusion
In the spirit of a Curry-Howard correspondence between a logical device and a concurrent one, we have shown how to stress a strong connection between differential interaction
nets and the solos calculus. Such links allow one to share methods from the two worlds,
with the possibility of giving concurrent interpretations to the cut-elimination procedure,
logical foundations to process calculi, of applying tools from concurrency theory such as
behavioural equivalences to formal proofs, or conversely denotational semantics from linear
logic to concurrent languages, ...
Let us discuss a few constraints or restrictions we have to face here.
Finitary calculi. We have only considered finitary calculi: without replications nor recursive
definitions. Nevertheless, exponential boxes are a natural device from linear logic proof-nets
(compatible with differential interaction nets) for representing replicable processes. We
have shown in [EL10] how to represent with these boxes a restricted (but expressive) form
of replication of the π-calculus. This restriction is related with difficulties for controling
reduction on the auxiliary ports of exponential boxes. By exploiting this kind of restriction
in the language of solos, it should be possible to represent a restricted form of replication
through our translation.
Logical correctness. Proof-nets of linear logic provide us with a graphical syntax for representing proofs of the sequent calculus [Gir96]. However this requires to impose a correctness
criterion to proof-nets in order to characterize exactly those which could be sequentialized
into a sequent calculus proof. There is such a correctness criterion for differential interaction
nets [ER06] (relating them with the sequent calculus of differential linear logic). However
the differential interaction nets obtained with our translation do not satisfy the correctness
criterion in general (this comes in particular from the unconstrained use of communication
areas). Some logical tools do not care about logical correctness and can be directly applied
in our setting (for example, the relational model is a denotational model of differential interaction nets even if they do not satisfy correctness). On the contrary, many important
results in logic only hold in the logically correct case. This is why it would be important
to understand how our translation interacts with correctness, or how expressive would be
a concurrent calculus whose translation into differential interaction nets satisfies the correctness criterion. In particular, what kind of communications could be represented with
communication areas in a logically correct setting?
ACYCLIC SOLOS AND DIFFERENTIAL INTERACTION NETS
35
Acyclicity. In trying to build a strong bridge between differential interaction nets and solo
diagrams, we have been able to define a suitable restriction of the solos calculus that can be
used as an intermediary step between the π-calculus and differential interaction nets. This
led to a new proof of the bisimulation result presented in [EL10].
The technical choices in the design of our acyclic solos calculus were completely determined by the just mentioned goal. Nevertheless, it would be interesting to study acyclic
solos for themselves. Their computational behaviour are in many ways similar to what happens in the π-calculus (“substitution-like” name passing in an “unification style” setting for
example). It would be interesting to see if they contain specific communication primitives
or if somehow the behaviour of an acyclic solos term always mimics the behaviour of a
π-term.
Another approach would be to extend our translation relation to take cycles (as in
Figure 15) into account. This requires us to understand the precise impact of these cycles
on the behaviour of differential interaction nets. In this way, it seems possible to extend
the bisimulation result to the whole solos calculus.
Following the link provided by the present work, the above mentioned points should
now be addressed to strengthen the idea of an underlying Curry-Howard correspondence.
However some possibilities are already open such as relational semantics for the π-calculus,
geometry of interaction for processes, ...
Acknowledgement
We would like to thank Cosimo Laneve for the helpful discussions we had about solos.
We also thank the anonymous referees for their useful comments.
References
[Abr93]
[Ale99]
[BM06]
[Bou92]
[BS94]
[Ehr02]
[EL10]
[ER06]
[FM05]
[Gir96]
[HL10]
Samson Abramsky. Computational interpretations of linear logic. Theoretical Computer Science,
111(1–2):3–57, 1993.
Vladimir Alexiev. Non-deterministic Interaction Nets. Ph.D. thesis, University of Alberta, 1999.
Emmanuel Beffara and François Maurel. Concurrent nets: a study of prefixing in process calculi.
Theoretical Computer Science, 356(3):356–373, May 2006.
Gérard Boudol. Asynchrony and the pi-calculus. Research Report 1702, INRIA, 1992.
Gianluigi Bellin and Philip J. Scott. On the pi-calculus and linear logic. Theoretical Computer
Science, 135(1):11–65, 1994.
Thomas Ehrhard. On Köthe sequence spaces and linear logic. Mathematical Structures in Computer Science, 12(5):579–623, 2002.
Thomas Ehrhard and Olivier Laurent. Interpreting a finitary pi-calculus in differential interaction
nets. Information and Computation, 208(6):606–633, June 2010.
Thomas Ehrhard and Laurent Regnier. Differential interaction nets. Theoretical Computer Science, 364(2):166–195, 2006.
Claudia Faggian and François Maurel. Ludics nets, a game model of concurrent interaction. In
Proceedings of the 20th Annual IEEE Symposium on Logic in Computer Science, pages 376–385.
IEEE Computer Society, 2005.
Jean-Yves Girard. Proof-nets: the parallel syntax for proof-theory. In Aldo Ursini and Paolo
Agliano, editors, Logic and Algebra, volume 180 of Lecture Notes In Pure and Applied Mathematics, pages 97–124, New York, 1996. Marcel Dekker.
Kohei Honda and Olivier Laurent. An exact correspondence between a typed pi-calculus and
polarised proof-nets. Theoretical Computer Science, 411(22–24):2223–2238, May 2010.
36
T. EHRHARD AND O. LAURENT
[HT91]
Kohei Honda and Mario Tokoro. An object calculus for asynchronous communication. In Proceedings of ECOOP’91, volume 512 of Lecture Notes in Computer Science, pages 133–147. SpringerVerlag, 1991.
[HY94]
Kohei Honda and Nobuko Yoshida. Combinatory representation of mobile processes. In POPL
’94: Proceedings of the 21st ACM SIGPLAN-SIGACT symposium on Principles of programming
languages, pages 348–360, New York, NY, USA, 1994. ACM.
[JM04]
Ole Jensen and Robin Milner. Bigraphs and mobile processes (revised). Technical report, Cambridge University Computer Laboratory, 2004.
[Kha03] Lionel Khalil. Généralisation des Réseaux d’Interaction avec amb, l’agent de McCarthy : propriétés et applications. Thèse de doctorat, Ecole Normale Supérieure de Paris, June 2003.
[Laf95]
Yves Lafont. From proof nets to interaction nets. In Jean-Yves Girard, Yves Lafont, and Laurent
Regnier, editors, Advances in Linear Logic, volume 222 of London Mathematical Society Lecture
Note Series, pages 225–247. Cambridge University Press, 1995.
[LPV01] Cosimo Laneve, Joachim Parrow, and Björn Victor. Solo diagrams. In Proceedings of the 4th
conference on Theoretical Aspects of Computer Science, TACS’01, number 2215 in Lecture Notes
in Computer Science, pages 127–144. Springer-Verlag, 2001.
[LV03]
Cosimo Laneve and Björn Victor. Solos in concert. Mathematical Structures in Computer Science,
13(5):657–683, 2003.
[Maz05] Damiano Mazza. Multiport interaction nets and concurrency. In Proceedings of CONCUR 2005,
number 3653 in Lecture Notes in Computer Science, pages 21–35. Springer-Verlag, 2005.
[Maz06] Damiano Mazza. Interaction Nets: Semantics and Concurrent Extensions. Thèse de doctorat,
Université Aix-Marseille II / Università degli Studi Roma Tre, 2006.
[Mil94]
Robin Milner. Pi-nets: A graphical form of pi-calculus. In ESOP ’94: Proceedings of the 5th
European Symposium on Programming, pages 26–42, London, UK, 1994. Springer-Verlag.
[MPW92] Robin Milner, Joachim Parrow, and David Walker. A calculus of mobile processes, I/II. Information and Computation, 100(1):1–77, 1992.
[Par81]
David Park. Concurrency and automata on infinite sequences. In Peter Deussen, editor, 5th GIconference on Theoretical Computer Science, volume 104 of Lecture Notes in Computer Science,
pages 167–183. Springer-Verlag, March 1981.
[PV98]
Joachim Parrow and Björn Victor. The fusion calculus: Expressiveness and symmetry in mobile
processes. In Proceedings of the thirteenth annual symposium on Logic In Computer Science, pages
176–185, Indianapolis, June 1998. IEEE, IEEE Computer Society Press.
[Reg92] Laurent Regnier. Lambda-Calcul et Réseaux. Thèse de doctorat, Université Paris VII, 1992.
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http://creativecommons.org/licenses/by-nd/2.0/ or send a
letter to Creative Commons, 171 Second St, Suite 300, San Francisco, CA 94105, USA, or
Eisenacher Strasse 2, 10777 Berlin, Germany
| 6cs.PL
|
Quantum Algorithms for Graph Connectivity
and Formula Evaluation
Stacey Jeffery1 and Shelby Kimmel2
1
QuSoft and CWI, Amsterdam, the Netherlands
2
Middlebury College, Middlebury, VT, USA
arXiv:1704.00765v2 [quant-ph] 16 Aug 2017
August 17, 2017
We give a new upper bound on the quantum query complexity of deciding st-connectivity on certain classes of planar graphs, and show the bound is
sometimes exponentially better than previous results. We then show Boolean
formula evaluation reduces to deciding connectivity on just such a class of
graphs. Applying the algorithm √
for st-connectivity to Boolean formula evaluation problems, we match the O( N ) bound on the quantum query complexity
of evaluating formulas on N variables, give a quadratic speed-up over the classical query complexity of a certain class of promise Boolean formulas, and show
this approach can yield superpolynomial quantum/classical separations. These
results indicate that this st-connectivity-based approach may be the “right” way
of looking at quantum algorithms for formula evaluation.
1 Introduction
Deciding whether two points are connected in a network is a problem of significant practical
importance. In this work, we argue that this problem, st-connectivity, is also important
as a quantum algorithmic primitive.
Dürr, Heiligman, Høyer, and Mhalla designed a quantum algorithm for deciding stconnectivity that requires O(|V |3/2 ) queries to the adjacency matrix of a graph on vertex
set V [13]. Belovs and Reichardt later discovered an especially elegant span-program-based
quantum algorithm for this problem, which is time-efficient and requires only logarithmic
space [4]. Belovs and Reichardt’s algorithm improves on the query complexity of Dürr et
al.’s algorithm when the connecting path is promised to be short (if it exists).
Belovs and Reichardt’s st-connectivity algorithm has already been adapted or been used
as a subroutine for deciding other graph problems, such as detecting certain subgraphs [4],
deciding whether a graph is a forest [8], and deciding whether a graph is bipartite [8].
In this work, we modify the span program algorithm used in [4], inheriting its space
and time efficiency, and we restrict to deciding st-connectivity on a class of planar graphs.
If the effective resistances of the set of graphs in question (and their planar duals) are
small, then we find the quantum algorithm requires far fewer queries than suggested by
the analysis in [4]. In fact, we obtain a polynomial to constant improvement in query
complexity for some classes of graphs.
Stacey Jeffery: [email protected], SJ completed parts of this work while at the Institute for Quantum Information
and Matter (IQIM), Caltech
Shelby Kimmel: [email protected], SK completed parts of this work while at the Joint Center for
Quantum Information and Computer Science (QuICS), University of Maryland
Accepted in
Quantum 2017-08-12, click title to verify
1
In addition to improving our understanding of the quantum query complexity of stconnectivity problems, we show that Boolean formula evaluation reduces (extremely naturally) to st-connectivity problems of the kind for which our improved analysis holds. Therefore, finding good algorithms for st-connectivity can lead to good algorithms for Boolean
formula evaluation. While one might not expect that such a reduction would produce good
algorithms, we find the reduction gives optimal performance for certain classes of Boolean
formulas.
Boolean formula evaluation is a fundamental class of problems with wide-reaching implications in algorithms and complexity theory. Quantum speed-ups for evaluating formulas like or [15] and the nand-tree [14] spurred interest in better understanding the
performance of quantum algorithms for Boolean formulas. This research culminated in
the development of span program algorithms [23, 24], which can have optimal quantum
query
√ complexity for any problem [20]. Using span program algorithms, it was shown that
O( N ) queries are sufficient for any read-once formula with N inputs [20, 22]. Classically,
the query complexity of evaluating nand-trees is Θ(N .753 ) [25] and the query complexity
of evaluating arbitrary read-once formulas is Ω(N .51 ) [16].
While there are simple bounds on the quantum query complexity of total formula
evaluation problems, promise versions are still not fully understood. Kimmel [19] showed
that for a certain promise version of nand-trees, called k-fault trees, the quantum query
complexity is O(2k ), while Zhan, Kimmel, and Hassidim [27] showed the classical query
complexity is Ω((log logkN )k ), giving a superpolynomial quantum speed-up for a range of
values of k. More general treatment of when promises on the inputs give superpolynomial
query speed-ups can be found in [1].
Since our analysis of st-connectivity shows that graphs with small effective resistance
can be decided efficiently, this in turn means that Boolean formula evaluation problems
with the promise that their inputs correspond to low resistance graphs can also be evaluated
efficiently. This result gives us new insight into the structure of quantum speed-ups for
promise Boolean formulas.
Contributions. We summarize the main results in this paper as follows:
• Improved quantum query algorithm for deciding st-connectivity when the input is a
subgraph of some graph G such that G ∪ {{s, t}} — G with an additional st-edge —
is planar.
– The analysis involves the effective resistance of the original graph and its planar
dual.
– We find families of graphs for which this analysis gives exponential and polynomial improvements, respectively, over the previous quantum analysis in [4].
• Algorithm for Boolean formula evaluation via reduction to st-connectivity.
– Using this reduction, we provide a simple proof of the fact that
√ read-once
Boolean formulas with N input variables can be evaluated using O( N ) queries.
– We show both a quadratic and a superpolynomial quantum-to-classical speed-up
using this reduction, for certain classes of promise Boolean formula evaluation
problems.
Open Problems. We would like to have better bounds on the classical query complexity
of evaluating st-connectivity problems, as this would provide a new approach to finding
separations between classical and quantum query complexity. Additionally, our reduction
Accepted in
Quantum 2017-08-12, click title to verify
2
from Boolean formula evaluation to st-connectivity could be helpful in the design of new
classical algorithms for formulas.
Another open problem concerns span programs in general: when can we view span
programs as solving st-connectivity problems? This could be useful for understanding when
span programs are time-efficient, since the time-complexity analysis of st-connectivity span
programs is straightforward (see Appendix A.1, [4, Section 5.3], [17, Appendix B]).
An important class of st-connectivity-related span programs are those arising from
the learning graph framework, which provides a means of designing quantum algorithms
that is much simpler and more intuitive than designing a general span program [3]. A
limitation of this framework is its one-sidedness with respect to 1-certificates: whereas a
learning graph algorithm is designed to detect 1-certificates, a framework capable of giving
optimal quantum query algorithms for any decision problem would likely treat 0- and
1-inputs symmetrically. In our analysis of st-connectivity, 1-inputs and 0-inputs are on
equal footing. This duality between 1- and 0-inputs in st-connectivity problems could give
insights into how to extend the learning graph framework to a more powerful framework,
without losing its intuition and relative simplicity.
Organization: Section 2 provides background information. In Section 3, we describe
our improved analysis of the span program algorithm for st-connectivity for subgraphs
of graphs G such that G ∪ {{s, t}} is planar. In Section 4, we show that every formula
evaluation problem is equivalent to an st-connectivity problem. In Section 5, we apply
these results to promise nand-trees, for which we are able to prove the most significant
classical/quantum separation using our approach. Also in Section 5, we use these ideas to
create an improved algorithm for playing the two-player game associated with a nand-tree.
2 Preliminaries
2.1 Graph Theory
For an undirected weighted multigraph G, let V (G) and E(G) denote the vertices and
edges of G respectively. In this work, we will only consider undirected multigraphs, which
we will henceforth often refer to as graphs. To refer to an edge in a multigraph, we will
specify the endpoints, as well as a label λ, so that an edge is written ({u, v}, λ). Although
the label λ will be assumed to uniquely specify the edge, we include the endpoints for
→
−
convenience. Let E (G) = {(u, v, λ) : ({u, v}, λ) ∈ E(G)} denote the set of directed edges
of G. For a planar graph G (with an implicit planar embedding) let F (G) denote the faces
of G. We call the infinite face of a planar graph the external face.
For any graph G with connected vertices s and t, we can imagine a fluid flowing into G
at s, and traveling through the graph along its edges, until it all finally exits at t. The fluid
will spread out along some number of the possible st-paths in G. Such a linear combination
of st-paths is called an st-flow. More precisely:
Definition 1 (Unit st-flow). Let G be an undirected weighted graph with s, t ∈ V (G), and
→
−
s and t connected. Then a unit st-flow on G is a function θ : E (G) → R such that:
→
−
1. For all (u, v, λ) ∈ E (G), θ(u, v, λ) = −θ(v, u, λ);
=
P
3. for all u ∈ V (G) \ {s, t},
P
2.
P
−
→ θ(s, v, λ)
v,λ:(s,v,λ)∈ E
Accepted in
−
→ θ(v, t, λ)
v,λ:(v,t,λ)∈ E
= 1; and
−
→ θ(u, v, λ)
v,λ:(u,v,λ)∈ E
Quantum 2017-08-12, click title to verify
= 0.
3
Definition 2 (Unit Flow Energy). Given a unit st-flow θ on a graph G, the unit flow
energy is
J(θ) =
X
θ(u, v, λ)2 .
(1)
({u,v},λ)∈E(G)
Definition 3 (Effective resistance). Let G be a graph with s, t ∈ V (G). If s and t are
connected in G, the effective resistance is Rs,t (G) = minθ J(θ), where θ runs over all unit
st-flows. If s and t are not connected, Rs,t (G) = ∞.
Intuitively, Rs,t (G) characterizes “how connected” the vertices s and t are. The more,
shorter paths connecting s and t, the smaller the effective resistance.
The effective resistance has many applications. In a random walk on G, Rs,t (G)|E(G)| is
equal to the commute time between s and t, or the expected time a random walker starting
from s takes to reach t and then return to s [2, 9]. If G models an electrical network in
which each edge e of G is a unit resistor and a potential difference is applied between s and
t, then Rs,t (G) corresponds to the resistance of the network, which determines the ratio
of current to voltage in the circuit (see [11]). We can extend these connections further by
considering weighted edges. A network consists of a graph G combined with a positive
real-valued weight function c : E(G) → R+ .
Definition 4 (Effective Resistance with weights). Let N = (G, c) be a network with
P
θ(u,v,λ)2
s, t ∈ V (G). The effective resistance of N is Rs,t (N ) = minθ ({u,v},λ)∈E(G) c({u,v},λ)
,
where θ runs over all unit st-flows.
In a random walk on a network, which models any reversible Markov chain, a walker
at vertex u traverses edge ({u, v}, λ) with probability proportional to c({u, v}, λ). Then
P
the commute time between s and t is Rs,t (N ) e∈E(G) c(e). When N models an electrical
network in which each edge e represents a resistor with resistance 1/c(e), then Rs,t (N )
corresponds to the resistance of the network.
When G is a single edge e = ({s, t}, λ) with weight c(e), then the resistance Rs,t (G) =
1/c(e). When calculating effective resistance, Rs,t , we use the rule that for edges in series
(i.e., a path), or more generally, graphs connected in series, resistances add. Edges in
parallel, or more generally, graphs connected in parallel, follow the rule that conductances
in parallel add, where the conductance of a graph is given by one over the resistance. (The
conductance of an edge e is equal to c(e), the weight of the edge.) More precisely, it is easy
to verify the following:
Claim 5. Let two networks N1 = (G1 , c1 ) and N2 = (G2 , c2 ) each have nodes s and t. If
we create a new graph G by identifying the s nodes and the t nodes (i.e. connecting the
graphs in parallel) and define c : E(G) → R+ by c(e) = c1 (e) if e ∈ E(G1 ) and c(e) = c1 (e)
if e ∈ E(G2 ), then
Rs,t (G, c) =
1
1
+
Rs,t (G1 , c1 ) Rs,t (G2 , c2 )
!−1
.
(2)
However, if we create a new graph G by identifying the t node of G1 with the s node of G2 ,
relabeling this node v 6∈ {s, t} (i.e. connecting the graphs in series) and define c as before,
then
Rs,t (G, c) = Rs,t (G1 , c1 ) + Rs,t (G2 , c2 ).
Accepted in
Quantum 2017-08-12, click title to verify
(3)
4
As a bit of foreshadowing, if we let Rs,t (G1 , c1 ) and Rs,t (G2 , c2 ) take values 0, representing
false, or ∞, representing true, then clearly (3) computes the function or, since 0+0 = 0,
and 0 + ∞ = ∞ + 0 = ∞ + ∞ = ∞. We also have that (2) computes the and function, if
1
we use 01 = ∞ and ∞
= 0.
Definition 6 (st-cut). Given a graph G with s, t ∈ V (G), if s and t are not connected, an
st-cut is a function κ : V (G) → {0, 1} such that κ(s) = 1, κ(t) = 0, and κ(v) − κ(u) = 0
whenever {u, v} ∈ E(G).
In other words, κ defines a subset S ⊂ V (G) such that s ∈ S, t 6∈ S, and there is no edge
of G with one endpoint in S, and one endpoint in S. An st-cut is a witness that s and t
are in different components of G, so no path exists between s and t.
Finally, we consider dual graphs:
Definition 7 (Dual Graph). Let G be a planar graph (with an implicit embedding). The
dual graph, G† , is defined as follows. For every face f ∈ F (G), G† has a vertex vf , and
any two vertices are adjacent if their corresponding faces share an edge, e. We call the edge
between two such vertices the dual edge to e, e† . By convention, e and e† will always have
the same label, so that if e = ({u, v}, λ), then e† = ({vf , vf 0 }, λ) for f and f 0 the faces of
G on either side of the edge e.
2.2 Span Programs and Quantum Query Algorithms
Span programs [18] were first introduced to the study of quantum algorithms by Reichardt
and Špalek [24]. They have since proven to be immensely important for designing quantum
algorithms in the query model.
Definition 8 (Span Program). A span program P = (H, U, τ, A) on {0, 1}N is made up of
(I) finite-dimensional inner product spaces H = H1 ⊕· · ·⊕HN , and {Hj,b ⊆ Hj }j∈[N ],b∈{0,1}
such that Hj,0 + Hj,1 = Hj , (II) a vector space U , (III) a non-zero target vector τ ∈ U ,
and (IV) a linear operator A : H → U . For every string x ∈ {0, 1}N , we associate the
subspace H(x) := H1,x1 ⊕ · · · ⊕ HN,xN , and an operator A(x) := AΠH(x) , where ΠH(x) is
the orthogonal projector onto H(x).
Definition 9 (Positive and Negative Witness). Let P be a span program on {0, 1}N and
let x be a string x ∈ {0, 1}N . Then we call |wi a positive witness for x in P if |wi ∈ H(x),
and A|wi = τ . We define the positive witness size of x as:
w+ (x, P ) = w+ (x) = min{k|wik2 : |wi ∈ H(x), A|wi = τ },
(4)
if there exists a positive witness for x, and w+ (x) = ∞ otherwise. Let L(U, R) denote the
set of linear maps from U to R. We call a linear map ω ∈ L(U, R) a negative witness for x
in P if ωAΠH(x) = 0 and ωτ = 1. We define the negative witness size of x as:
w− (x, P ) = w− (x) = min{kωAk2 : ω ∈ L(U, R), ωAΠH(x) = 0, ωτ = 1},
(5)
if there exists a negative witness, and w− (x) = ∞ otherwise. If w+ (x) is finite, we say that
x is positive (wrt. P ), and if w− (x) is finite, we say that x is negative. We let P1 denote
the set of positive inputs, and P0 the set of negative inputs for P . In this way, the span
program defines a partition (P0 , P1 ) of [N ].
For a function f : X → {0, 1}, with X ⊆ {0, 1}N , we say P decides f if f −1 (0) ⊆ P0
and f −1 (1) ⊆ P1 . We can use P to design a quantum query algorithm that decides f ,
given access to the input x ∈ X via queries of the form Ox : |i, bi 7→ |i, b ⊕ xi i.
Accepted in
Quantum 2017-08-12, click title to verify
5
Theorem 10 ([21]). Fix X ⊆ {0, 1}N and f : X → {0, 1}, and let P be a span program on {0, 1}N that decides f . Let W+ (f, P ) = maxx∈f −1 (1) w+ (x, P ) and W− (f, P ) =
maxx∈f −1 (0) w− (x, P ). Then there is a bounded error quantum algorithm that decides f
p
with quantum query complexity O( W+ (f, P )W− (f, P )).
2.3 Boolean Formulas
A read-once Boolean formula can be expressed as a rooted tree in which the leaves are
uniquely labeled by variables, x1 , . . . , xN , and the internal nodes are labeled by gates from
the set {∧, ∨, ¬}. Specifically, a node of degree 2 must be labeled by ¬ (not), whereas
higher degree nodes are labeled by ∧ (and) or ∨ (or), with the fan-in of the gate being
defined as the number of children. The depth of a Boolean formula is the largest distance
from the root to a leaf. We define an and-or formula (also called a monotone formula) as a
read-once Boolean formula for which every internal node is labeled by ∧ or ∨. Restricting to
and-or formulas does not lose much generality, since for any formula, there is an equivalent
formula in which all not-gates are at distance one from a leaf, and such not gates do not
affect the query complexity of the formula. Moreover, although we only consider read-once
formulas here, our techniques can be applied to more general formulas in which a single
variable may label multiple leaves, since this is equivalent to a larger read-once formula
with a promise on the input. Hereafter, when we refer to a formula, we will mean an
and-or read-once formula.
In a slight abuse of notation, at times xi will denote a Boolean variable, and at times,
it will denote a bit instantiating that variable. If x ∈ {0, 1}N is an instantiation of all
variables labeling the leaves of a formula φ, then φ(x) is the value of φ on that input,
defined as follows. If φ = xi has depth 0, then φ(x) = xi . If φ has depth greater than 0,
we can express φ recursively in terms of subformulas φ1 , . . . , φl , as φ = φ1 ∧ · · · ∧ φl , if the
root is labeled by ∧, or φ = φ1 ∨ · · · ∨ φl , if the root is labeled by ∨. In the former case, we
define φ(x) = φ1 (x) ∧ · · · ∧ φl (x), and in the latter case, we define φ(x) = φ1 (x) ∨ · · · ∨ φl (x).
A family of formulas φ = φN on N variables gives rise to an evaluation problem, Evalφ , in
which the input is a string x ∈ {0, 1}N , and the output is φN (x). If φ(x) = 0, we say x is a 0instance, and if φ(x) = 1, x is a 1-instance. By φ1 ◦φ2 , we mean φ1 composed with φ2 . That
is, if φ1 : {0, 1}N1 → {0, 1} and φ2 : {0, 1}N2 → {0, 1}, then φ1 ◦ φ2 : {0, 1}N1 N2 → {0, 1}
evaluates as φ1 ◦φ2 (x) = φ1 (φ2 (x1 ), . . . , φ2 (xN1 )), where x = (x1 , . . . , xN1 ) for xi ∈ {0, 1}N2 .
An important formula evaluation problem is nand-tree evaluation. A nand-tree is a
full binary tree of arbitrary depth d — that is, every internal node has two children, and
every leaf node is at distance d from the root — in which an internal node is labeled by
∨ if it is at even distance from the leaves, or ∧ if it is at odd distance from the leaves.
We use nandd to denote a nand-tree of depth d. While nandd is sometimes defined as a
Boolean formula of nand-gates composed to depth d, we will instead think of the formula
as alternating and-gates and or-gates — when d is even, these two characterizations are
identical. An instance of nandd is a binary string x ∈ {0, 1}N , where N = 2d . For example,
the formula nand2 (x1 , x2 , x3 , x4 ) = (x1 ∧ x2 ) ∨ (x3 ∧ x4 ) is a nand-tree of depth 2. nand0
denotes the single-bit identity function.
d
A nandd instance x ∈ {0, 1}2 can be associated with a two-player game on the rooted
binary tree that represents nandd , where the leaves take the values xi , as in Figure 5.
The game starts at the root node, which we call the current node. In each round of the
game, as long as the current node is not a leaf, if the current node is at even (respectively
odd) distance from the leaves, Player A (resp. Player B) chooses one of the current node’s
children to become the current node. When the current node is a leaf, if the leaf has value
1, then Player A wins, and if the leaf has value 0, then Player B wins. The sequence of
Accepted in
Quantum 2017-08-12, click title to verify
6
moves by the two players determines a path from the root to a leaf.
A simple inductive argument shows that if x is a 1-instance of nand-tree, then there
exists a strategy by which Player A can always win, no matter what strategy B employs;
and if x is a 0-instance, there exists a strategy by which Player B can always win. We say
an input x is A-winnable if it has value 1 and B-winnable if it has value 0.
3 Improved Analysis of st-connectivity Algorithm
In this section, we give an improved bound on the runtime of a quantum algorithm for
st-connectivity on subgraphs of G, where G ∪ {{s, t}} is planar.
Let st-connG,D be a problem parameterized by a family of multigraphs G, which takes
as input a string x ∈ D where D ⊆ {0, 1}E(G) . An input x defines a subgraph G(x) of
G by including the edge e if and only if xe = 1. For all x ∈ D, st-connG,D (x) = 1 if
and only if there exists a path connecting s and t in G(x). We write st-connG when
D = {0, 1}E(G) . A quantum algorithm for st-connG,D accesses the input via queries to a
standard quantum oracle Ox , defined Ox |ei|bi = |ei|b ⊕ xe i.
The authors of [4] present a quantum query algorithm for st-connG when G is a
complete graph, which is easily extended to any multigraph G. We further generalize their
algorithm to depend on some weight function c : E(G) → R+ (a similar construction is
also implicit in [3]). We call the following span program PG,c :
→
−
∀e ∈ E (G) : He,0 = {0},
→
−
H = span{|ei : e ∈ E (G)}
He,1 = span{|ei},
X
U = span{|ui : u ∈ V (G)}, τ = |si − |ti, A =
q
c({u, v}, λ)(|ui − |vi)hu, v, λ|.
−
→
(u,v,λ)∈ E (G)
(6)
For any choice of weight function c, this span program decides st-connG , but as we will
soon see, the choice of c may impact the complexity of the resulting algorithm.
Using PG,c with c({u, v}, λ) = 1 for all ({u, v}, λ) ∈ E(G), the authors of Ref. [4] show
that the query complexity of evaluating st-connG,D is
!
r
O
max
x∈D:s,t are connected
Rs,t (G(x)) × |E(G)| .
(7)
Their analysis was for the case where G is a complete graph, but it is easily seen to apply
to more general multigraphs G. In fact, it is straightforward to show that this bound can
be improved to
O
!
r
max
x∈D:s,t are connected
Rs,t (G(x)) ×
max
x∈D:s,t are not connected
(Cs,t (G(x))) .
(8)
where
Cs,t (G(x)) =
min
κ:κ is an st-cut of G(x)
X
|κ(u) − κ(v)|
if s and t not connected
({u,v},λ)∈E(G)
∞
otherwise.
(9)
In particular, when G is a complete graph on vertex set V , with the promise
that if
√
an st-path exists, it is of length at most k, Eq. (7) gives a bound of O
k|V | . In the
Accepted in
Quantum 2017-08-12, click title to verify
7
worst case, when k = |V |, the analysis of [4] does not improve on the previous quantum
algorithm of [13], which gives a bound of O(|V |3/2 ).
In this paper, we consider in particular multigraphs that are planar even when an
additional st-edge is added (equivalently, there exists a planar embedding in which s and
t are on the same face), as in graph G in Figure 1. (In the case of Figure 1, s and t are
both on the external face.) Given such a graph G, we define three other related graphs,
†
which we denote by G, G , and G0 .
We first define the graph G, which is the same as G, but with an extra edge labeled by
†
∅ connecting s and t. We then denote by G the planar dual of G. Because every planar
dual has one edge crossing each edge of the original graph, there exists an edge that is dual
to ({s, t}, ∅), also labeled by ∅. We denote by s0 and t0 the two vertices at the endpoints
†
of ({s, t}, ∅)† = ({s0 , t0 }, ∅). Finally, we denote by G0 the graph G except with the edge
({s0 , t0 }, ∅) removed.
s
s
s
1
1
3 4
2
∅
3 4
2
t
G
t
∅
s0
G
s
∅
1
t0
G
s0
3 4
t0
2
G0
†
†
Figure 1: Example of how to derive G, G , and G0 from a planar graph G where s and t are on the
†
same face. G is obtained from G by adding an edge ({s, t}, ∅). G is the planar dual of G. (In the
†
†
†
diagram labeled by G , G is the gray graph, while G is black). G0 is obtained from G by removing the
edge ({s, t}, ∅)† . Note that dual edges inherit their labels (in this case 1, 2, 3, 4, ∅) from the primal edge.
By construction, G0 will always have the same number of edges as G. Then as x defines
a subgraph G(x) of G by including the edge e if and only if xe = 1, we let G0 (x) be the
subgraph of G0 where we include the edge e† if and only if xe = 0.
If there is no path from s to t in G(x), there must be a cut between s and t. Note that
for any e ∈ E(G), e ∈ E(G(x)) if and only if e† 6∈ E(G0 (x)). Looking at Figure 1, one
can convince oneself that any s0 t0 -path in G0 (x) defines an st-cut in G(x): simply define
κ(v) = 1 for vertices above the path, and κ(v) = 0 for vertices below the path.
Let c be a weight function on E(G). Then we define a weight function c0 on E(G0 ) as
c0 (e† ) = 1/c(e). Then for every x there will be a path either from s to t in G(x) (and hence
Rs,t (G(x), c)) will be finite), or a path from s0 to t0 in G0 (x) (in which case Rs0 ,t0 (G0 (x), c0 )
will be finite).
We can now state our main lemma:
Lemma 11. Let G be a planar multigraph with s, t ∈ V (G) such that G ∪ {{s, t}} is also
planar, and let c be a weight function on E(G). Let x ∈ {0, 1}E(G) . Then w+ (x, PG,c ) =
1
0
0
0 0
2 Rs,t (G(x), c) and w− (x, PG,c ) = 2Rs ,t (G (x), c ).
Using Lemma 11 and Theorem 10, we immediately have the following:
Theorem 12. Let G be a planar multigraph with s, t ∈ V (G) such that G ∪ {{s, t}} is also
planar. Then the bounded error quantum query complexity of evaluating st-connG,D is
O min
c
r
!
max
x∈D:st-connG (x)=1
Rs,t (G(x), c) ×
max
x∈D:st-connG (x)=0
Rs0 ,t0 (G0 (x), c0 )
(10)
where the minimization is over all positive real-valued functions c on E(G).
Accepted in
Quantum 2017-08-12, click title to verify
8
While it might be difficult in general to find the optimal edge weighting c, any choice
of c will at least give an upper bound on the query complexity. However, as we will see,
sometimes the structure of the graph will allow us to efficiently find good weight functions.
The proof of Lemma 11 is in Appendix A. The positive witness result follows from
generalizing the proof in [4] to weighted multigraphs. The idea is that an st-path witnesses
that s and t are connected, as does any linear combination of such paths — i.e. an st-flow.
The effective resistance Rs,t (G(x), c) characterizes the size of the smallest possible st-flow.
Just as a positive witness is some linear combination of st-paths, similarly, a negative
witness turns out to be a linear combination of st-cuts in G(x). But as we’ve argued, every
st-cut corresponds to an s0 t0 -path in G0 (x). Using the correspondence between cuts and
paths, we have that a negative witness is a linear combination of s0 t0 -paths in G0 (x). This
allows us to show a correspondence between complexity-optimal negative witnesses and
minimal s0 t0 -flows, connecting w− (x, PG,c ) to Rs0 ,t0 (G0 (x), c0 ).
In Appendix A.1, we show that if a quantum walk step on the network (G, c) can be
implemented time efficiently, then this algorithm is not only query efficient, but also time
efficient. Let
1
UG,c : |ui|0i 7→ qP
q
X
−
→
c({u, v}, λ)
−
→
v,λ:(u,v,λ)∈ E (G)
v,λ:(u,v,λ)∈ E (G)
c({u, v}, λ)|ui|u, v, λi.
(11)
Then we show the following.
Theorem 13. Let PG,c = (H, U, A, τ ) be defined as in (6). Let SG,c be an upper bound on
the time complexity of implementing UG,c . If G has the property that G ∪ {{s, t}} is planar,
then the time complexity of deciding st-connG,D is at most
O min SG,c
c
r
!
max
x∈D:s,t are connected
Rs,t (G(x), c) ×
max
x∈D:s,t are not connected
Rs0 ,t0 (G0 (x), c0 ) .
(12)
In Appendix A.1, we also show that if the space complexity of implementing UG,c in
0 , the algorithm referred to in Theorem 13 has space complexity at most
time SG,c is SG,c
0 ).
O(max{log |E(G)|, log |V (G)|} + SG,c
3.1 Comparison to Previous Quantum Algorithm
When G ∪ {{s, t}} is planar, our algorithm always matches or improves on the algorithm
in [4]. To see this, we compare Eqs. (10) and (8), and choose c to have value 1 on all edges
of G. Then the first terms are the same in both bounds, so we only need to analyze the
second term. However, using the duality between paths and cuts, we have
Cs,t (G(x)) = shortest path length from s0 to t0 in G0 (x) ≥ Rs0 t0 (G0 (x)).
(13)
To obtain the inequality in Eq. (13), we create an s0 t0 -flow on G0 (x) that has value one on
edges on the shortest path from s0 to t0 and zero on all other edges. Such a flow has unit
flow energy equal to the shortest path. However, the true effective resistance can only be
smaller than this, because it is the minimum energy over all possible s0 t0 -flows.
We now present two simple examples where our algorithm and analysis do better than
that of [4]. In the first example, we highlight how the change from Cs,t (G(x)) to Rs0 ,t0 (G(x))
in the complexity gives us an advantage for some graphs. In the second example, we show
Accepted in
Quantum 2017-08-12, click title to verify
9
that being able to choose a non-trivial weight function c can give us an advantage for some
graphs.
Let G be an st-path of length N : i.e., N + 1 vertices arranged in a line so that each
vertex is connected to its neighbors to the left and right by a single edge, and s and t
are the vertices on either end of the line, as in Figure 2. For some h ∈ {1, . . . , N }, let
D = {1N } ∪ {x ∈ {0, 1}N : |x| ≤ N − h}, where 1N is the all-one string of length N , and
|x| is the hamming weight of the string x.
s0
s
t
...
...
G0
G
t0
Figure 2: Example of graph for which our analysis does better than the analysis of [4], even with c = 1
for all edges, under the promise that G0 (x) always contains at least h edges, if s0 and t0 are connected.
Then, choosing c to have value 1 on all edges of G, we have
max
x∈D:st-connG (x)=1
Rs,t (G(x)) = N
(14)
because the only x ∈ D such that s and t are connected in G(x) is x = 1N , in which case
the only unit flow has value 1 on each edge. This flow has energy N . However
max
x∈D:st-connG (x)=0
Rs0 ,t0 (G0 (x)) ≤ 1/h,
(15)
because when s and t are not connected in G(x), G(x) has at most N − h edges, so G0 (x)
has at least h edges. Thus we can define a unit flow with value 1/h on each of h parallel
edges in G0 (x), giving an energy of 1/h. On the other hand
max
x∈D:st-connG (x)=0
Cs,t (G(x)) = 1.
(16)
In fact, since Cs,t (G(x)) counts the minimum number of edges ({u, v}, λ) across any cut
(i.e. such that κ(u) = 1 and κ(v) = 0), it is always at least 1, for any G(x) in which an
0
st-cut exists, whereas
√ Rs0 ,t0 (G (x)) can be as small as 1/N for some G.
Choosing h = N in our example, and applying Eqs. (8) and (10), the analysis in
[4] gives a query complexity of O(N 1/2 ) while our analysis gives a query complexity of
O(N 1/4 ). In Section 4 we will show that this bound is tight.
Now consider the graph G in Figure 3. It consists of N edges in a line, connecting
vertices s, u1 , . . . , uN , and then N multi-edges between uN and t. We assign weights
c(e) = 1 for edges e on the path from s to uN , and c(e) = N −1 for all other edges.
Then,
max
x∈D:st-connG (x)=1
Rs,t (G(x), c) = 2N,
(17)
which occurs when only one of the multi-edges between uN and t is present. In that case,
the N edges {s, u1 }, {u1 , u2 }, . . . , {uN −1 , uN } each contribute 1 to the effective resistance,
1
and the final edge between uN and t contributes c(e)
= N . Also
max
x∈D:st-connG (x)=0
Accepted in
Rs0 ,t0 (G0 (x), c0 ) ≤ 1,
Quantum 2017-08-12, click title to verify
(18)
10
N −1
s0 N
N −1
1
s
1
u1
1
u2
.1
..
1 N −1
uN
t
N −1
1
N
1 1 . . .1
G0
t0 N
N
N
G
N −1
Figure 3: Example of graph for which our analysis does quadratically better than the analysis of [4] by
taking advantage of a non-trivial weight function c. The values of c for each edge of G, and of c0 for
each edge of G0 , are shown in boldface.
where the maximum occurs when there is only one path from s0 to t0 . (If it is the path with
N edges, each edge has weight N , and so contributes 1/N to the flow energy.) However
max
x∈D:st-connG (x)=0
Cs,t (G(x)) = N
(19)
for a cut across the multi-edges between uN and t, and
max
x∈D:st-connG (x)=1
Rs,t (G(x)) = N + 1,
(20)
which occurs when only one of the multi-edges between uN and t is present.
Thus, the analysis in [4] gives a query complexity of O(N ) while our analysis gives a
query complexity of O(N 1/2 ).
In Section 5 we will give an example where our analysis provides an exponential improvement over the analysis in [4].
4 AND-OR Formulas and st-Connectivity
In this section, we present a useful relationship between and-or formula evaluation problems and st-connectivity problems on certain graphs. As mentioned in Section 2, for
simplicity we will restrict our analysis to read-once formulas, but the algorithm extends
simply to “read-many” formulas. In this case, we will primarily be concerned with the
query complexity: the input x = (x1 , . . . , xN ) to a formula will be given via a standard
quantum oracle Ox , defined Ox |ii|bi = |ii|b ⊕ xi i.
Given an and-or formula φ with N variables, we will recursively construct a planar
multigraph Gφ , such that Gφ has two distinguished vertices labeled by s and t respectively,
and every edge of Gφ is uniquely labeled by a variable {xi }i∈[N ] . If φ = xi is just a single
variable, then Gφ is just a single edge with vertices labeled by s and t, and edge label xi .
That is E(Gφ ) = {({s, t}, xi )} and V (Gφ ) = {s, t}.
Otherwise, suppose φ = φ1 ∧ · · · ∧ φl . Then Gφ is the graph obtained from the graphs
Gφ1 , . . . , Gφl by identifying the vertex labeled t in Gφi with the vertex labeled s in Gφi+1 ,
for all i = 1, . . . , l − 1, and labeling the vertex labeled s in Gφ1 by s, and the vertex labeled
t in Gφl by t. That is, we connect the graphs Gφ1 , . . . , Gφl in series, as in Figure 4. (For
a formal definition of Gφ , see Appendix B).
The only other possibility is that φ = φ1 ∨ · · · ∨ φl . In that case, we construct Gφ
by starting with Gφ1 , . . . , Gφl and identifying all vertices labeled by s, and labeling the
Accepted in
Quantum 2017-08-12, click title to verify
11
resulting vertex with s, and identifying all vertices labeled by t, and labeling the resulting
vertex by t. That is, we connect Gφ1 , . . . , Gφl in parallel (see Figure 4). We note that graphs
constructed in this way are exactly the set of series-parallel graphs with two terminals (see
e.g. [26, Def. 3]), and are equivalent to graphs without a K4 minor [10, 12].
s
s
s
s
s
t
G φ1
t
G φ2
t
G φ3
t
Gφ1 ∨φ2 ∨φ3
t
Gφ1 ∧φ2 ∧φ3
Figure 4: Let φ1 = x1 ∧x2 , φ2 = x3 ∨(x4 ∧x5 ), and φ3 = x6 . Then we obtain Gφ1 ∧φ2 ∧φ3 by connecting
Gφ1 , Gφ2 , and Gφ3 in series, and Gφ1 ∨φ2 ∨φ3 by connecting them in parallel.
Note that for any φ, Gφ is planar, and furthermore, both s and t are always on the
same face. Thus, we can define G0φ , Gφ (x) and G0φ (x) as in Section 3. Then we can show
the following:
Lemma 14. Let φ be any and-or formula on N variables. For every x ∈ {0, 1}N ,
there exists a path from s to t in Gφ (x) if and only if φ(x) = 1. Furthermore, for every
x ∈ {0, 1}N , there exists a path from s0 to t0 in G0φ (x) if and only if φ(x) = 0.
We give a formal proof of Lemma 14 in Appendix B, but the intuition is that an or of
subformulas, φ1 ∨ · · · ∨ φl evaluates to true if any of the subformulas evaluates to true, and
likewise, if two vertices are connected by multiple subgraphs in parallel, the vertices are
connected if there is a path in any of the subgraphs. An and of subformulas φ1 ∧ · · · ∧ φl
evaluates to true only if every subformula evaluates to true, and likewise, if two vertices
are connected by multiple subgraphs in series, the vertices are only connected if there is a
path through every subgraph. Thus, we can show by induction that s and t are connected
in Gφ (x) if and only if φ(x) = 1. To see that s0 and t0 are connected in G0φ (x) if and only
if φ(x) = 0, we can use a similar argument, and make use of the fact that an s0 t0 -path in
G0φ (x) is an st-cut in Gφ (x).
Lemma 14 implies that we can solve a formula evaluation problem Evalφ by solving
the associated st-connectivity problem, in which the input is a subgraph of Gφ . By our
construction, Gφ will always be a planar graph with s and t on the external face, so
moreover, we can apply Theorem 12 to obtain the following.
Theorem 15. For any family φ of and-or formulas, the bounded error quantum query
complexity of Evalφ when the input is promised to come from a set D is
O min
c
r
!
max
x∈D:φ(x)=1
Rs,t (Gφ (x), c) ×
max
x∈D:φ(x)=0
Rs0 ,t0 (G0φ (x), c0 )
,
(21)
where the minimization is over all positive real-valued functions c on E(Gφ ).
Proof. By Lemma 14, the query complexity of Evalφ on D is at most the query complexity
of st-connGφ ,D . Since Gφ is planar, and has s and t on the same face, we can apply
Theorem 12, which immediately implies the result.
Accepted in
Quantum 2017-08-12, click title to verify
12
4.1 Comparison to Existing Boolean Formula Algorithms
Reichardt proved
√ that the quantum query complexity of evaluating any formula on N
variables is O( N ) [22, Corollary 1.6]. Our algorithm recovers this result:
Theorem 16. Let φ be a read-once formula on N variables. Then there exists a choice
of c on E(Gφ ) such that the quantum algorithm obtained from the span program PGφ ,c
√
computes Evalφ with bounded error in O( N ) queries.
We need the following claim, which we prove in Appendix B:
Claim 17. If φ = φ1 ∨ φ2 ∨ · · · ∨ φl , then G0φ (x) is formed by composing {G0φi (x)}i in series,
and if φ = φ1 ∧ φ2 ∧ · · · ∧ φl , then G0φ (x) is formed by composing {G0φi (x)}i in parallel.
The intuition behind Claim 17 is the following. Although G0φ is defined via the dual
of Gφ , which is constructed through a sequence of series and parallel compositions, G0φ
itself can also be built up through a sequence of series and parallel compositions. For any
and-or formula φ on N variables, we can define a formula φ0 on N variables by replacing
all ∨-nodes in φ with ∧-nodes, and all ∧-nodes in φ with ∨-nodes. By de Morgan’s law, for
all x ∈ {0, 1}N , φ(x) = ¬φ0 (x̄), where x̄ is the entrywise negation of x. A simple inductive
proof shows that Gφ0 = G0φ , and for all x, Gφ0 (x̄) = G0φ (x) (see Lemma 35 in Appendix B).
Proof of Theorem 16. We will make use of the following fact: for any network (G, c), and
any positive real number W :
Rs,t (G, c/W ) = min
θ
X θ(e)2
θ(e)2
= W min
= W Rs,t (G, c).
θ
c(e)/W
c(e)
e∈E(G)
e∈E(G)
X
(22)
We now proceed with the proof. For any formula φ in {∧, ∨, ¬}, by repeated applications
of de Morgan’s law, we can push all not-gates to distance-1 from a leaf. Since xi and ¬xi
can both be learned in one query, we can restrict our attention to and-or formulas.
If φ has only N = 1 variable, it’s easy to see that W+ (PGφ ,c )W− (PGφ ,c ) ≤ N for c
taking value 1 on the single edge in Gφ . We will prove by induction that this is true for
any φ, for some choice ofq
c, completing the proof,since the complexity of our algorithm
obtained from PGφ ,c is O
W+ (PGφ ,c )W− (PGφ ,c ) .
P
Suppose φ = φ1 ∧ · · · ∧ φl for formulas φi on Ni variables, so φ has N = i Ni
variables. For x ∈ {0, 1}N , we will let xi ∈ {0, 1}Ni denote the (N1 + · · · + Ni−1 + 1)-th
to (N1 + · · · + Ni )-th bits of x. For each Gφi , by the induction hypothesis, there is some
weight function ci on E(Gφi ) such that W+ (PGφi ,ci )W− (PGφi ,ci ) ≤ Ni .
Using our construction, Gφ is formed by composing {Gφi }i in series. Thus every edge
({u, v}, λ) ∈ E(Gφ ) corresponds to an edge ({u, v}, λ) ∈ E(Gφi ) for some i. We create a
weight function c : E(Gφ ) → R+ such that c({u, v}, λ) = Wci−({u,v},λ)
(PG ,c ) if ({u, v}, λ) is an
φi
i
edge originating from the graph Gφi . That is, our new weight function is the same as
combining all of the old weight functions, up to scaling factors {W− (PGφi ,ci )}i .
Using Lemma 11, Claim 5, and Eq. (22), for any 1-instance x,
l
1
1X
ci
w+ (x, PGp hi,c ) = Rs,t (Gφ (x), c) =
Rs,t Gφi (x),
2
2 i=1
W− (PGφi ,ci )
=
Accepted in
!
l
l
X
1X
W− (PGφi ,ci )Rs,t (Gφi (x), ci ) ≤
W− (φi , PGφi ,ci )W+ (PGφi ,ci ).
2 i=1
i=1
(23)
Quantum 2017-08-12, click title to verify
13
Thus
W+ (PGφ ,c ) ≤
l
X
W− (PGφi ,ci )W+ (PGφi ,ci ) ≤
i=1
l
X
Ni = N.
(24)
i=1
Recall that for a weight function c on Gφ , we define a weight function c0 on G0φ by
c0 (e† ) = 1/c(e). Then for an edge e ∈ E(Gφi (xi )), we have c0 (e† ) = W− (PGφi ,ci )/ci (e) =
W− (PGφi ,ci )c0i (e† ). By Claim 17, G0φ is formed by composing {G0φi }i in parallel, so by
Lemma 11, Claim 5, and Eq. (22):
w− (x, PGφ ,c ) =
2Rs0 ,t0 (G0φ (x), c0 )
=2
l
X
i=1
=2
=2
i=1
1
i=1
Rs0 ,t0 (G0φi (xi ), c0 )
1
Rs0 ,t0 (G0φi (xi ), W− (PGφi ,ci )c0i )
l
X
W− (PGφi ,ci )
!−1
l
X
!−1
=2
l
X
W− (PGφi ,ci )
i=1
Rs0 ,t0 (G0φi (xi ), c0i )
!−1
!−1
w− (xi , PGφi ,ci )
.
(25)
Whenever x is a 0-instance of φ, the set S ⊆ [l] of i such that xi is a 0-instance of φi is
non-empty. This is exactly the set of i such that w− (xi , PGφi ,ci ) < ∞. Continuing from
above, we have:
w− (x, PGφ ,c ) =
X W− (PGφ ,ci )
i
!−1
w (xi , PGφi ,ci )
i∈S −
≤
!
X W− (PGφ ,ci ) −1
i
i∈S
W− (PGφi ,ci )
=
1
≤ 1.
|S|
(26)
Thus W− (PGφi ,ci ) ≤ 1. Combining this with Eq. (24) we have W+ (PGφ ,c )W− (PGφ ,c ) ≤ N ,
as desired.
The proof for the case φ = φ1 ∨ · · · ∨ φl is similar.
An immediate corollary of Theorem 16 is the following.
Corollary 18. Deciding st-connectivity on
√ subgraphs of two-terminal series-parallel graphs
of N edges can be accomplished using O( N ) queries, if s and t are chosen to be the two
terminal nodes.
As with many results in this field, characterizing classical complexity seems to be more
difficult than quantum complexity. However, we show we can lower bound the classical
query complexity of a class of Boolean formulas in terms of the effective resistance of
their corresponding graphs, achieving a quadratic quantum/classical speed-up in query
complexity.
We consider and-or formulas on restricted domains. For N, h ∈ Z+ , let DN,h = {x ∈
0
{0, 1}N : |x| = N or |x| ≤ N − h} and let DN,h
= {x ∈ {0, 1}N : |x| = 0 or |x| ≥ h}.
We will analyze and-or formulas such that the input to every gate in the formula comes
0
from DN,h (in the case of and), which we denote and|DN,h and DN,h
(in the case of or),
0
which we denote or|DN,h
. These promises on the domains make it easier to evaluate both
functions. For example, if or evaluates to 1, we are promised that there will not be just
one input with value 1, but at least h.
Then using sabotage complexity [5] to bound the classical query complexity, we have
the following theorem, whose (somewhat long, but not technical) proof can be found in
Appendix C:
Accepted in
Quantum 2017-08-12, click title to verify
14
0
Theorem 19. Let φ = φ1 ◦ φ2 ◦ · · · ◦ φl , where for each i ∈ [l], φi = or|DN
or
,h
i
i
φi = and|DNi ,hi . Then the randomized bounded-error query complexity of evaluating φ
is Ω
O
Q
l
i=1 Ni /hi
Q
l
i=1
, and the bounded-error quantum query complexity of evaluating φ is
p
Ni /hi .
Note that in the above theorem, when we compose formulas with promises on the input,
we implicitly assume a promise on the input to the composed formula. More precisely, for φ1
on D1 ⊆ {0, 1}N1 and φ2 on D2 ⊆ {0, 1}N2 , φ = φ1 ◦φ2 is defined on all x = (x1 , . . . , xN1 ) ∈
{0, 1}N1 N2 such that xi ∈ D2 for all i ∈ [N1 ], and (φ2 (x1 ), . . . , φ2 (xN1 )) ∈ D1 .
Theorem 19 is proven by showing that
!
Ql
i=1 Ni
Ql
i=1 hi
=
max
x∈D:φ(x)=1
!
Rs,t (Gφ (x))
max
x∈D:φ(x)=0
Rs,t (G0φ (x))
,
(27)
and using sabotage complexity to show that this is a lower bound on the randomized
query complexity of φ. This gives us a quadratic separation between the randomized and
quantum query complexities of this class of formulas. For details, see Appendix C.
Using the composition lower bound for promise Boolean functions of [19], and the lower
bound for Grover’s search with multiple marked items [6], we have that the quantum query
complexity of Theorem 19 is tight. Additionally, in light of our reduction from Boolean
formula evaluation to st-connectivity, we see that our example from Figure 2 in Section 3
is equivalent to the problem of and|DN,h , so our query bound in that example is also tight.
Based on Theorem 19, one might guess that when evaluating formulas using the stconnectivity reduction, one can obtain at most a quadratic speed-up over classical randomized query complexity. However, it is in fact possible to obtain a superpolynomial speed-up
for certain promise problems using this approach, as we will discuss in Section 5.1.
5 NAND-tree Results
5.1 Query Separations
In this section, we prove two query separations that are stronger than our previous results.
These query separations rely on the nand-tree formula with a promise on the inputs. This
restriction, the k-fault promise, will be defined shortly. Let Fkd be the set of inputs to
nandd that satisfy the k-fault condition. Then the two results are the following:
Theorem 20. Using the st-connectivity approach to formula evaluation (Theorem 15),
d
one can solve Evalnandd when the input is promised to be from Flog
d with O(d) queries,
log
log(d)
while any classical algorithm requires Ω(d
) queries.
For a different choice of k, this example demonstrates the dramatic improvement our
st-connectivity algorithm can give over the analysis of [4] — in this case, an exponential
(or more precisely, a polynomial to constant) improvement:
Theorem 21. Consider the problem st-connGnand
d
,F1d .
The analysis of [4] gives a bound
of O(N 1/4 ) quantum queries to decide this problem (where N = 2d is the number of edges in
Gnandd ), while our analysis shows this problem can be decided with O(1) quantum queries.
We now define what we mean by k-fault nand-trees. In [27], Zhan et al. find a relationship between the difficulty of playing the two-player game associated with a nand-tree,
Accepted in
Quantum 2017-08-12, click title to verify
15
and the witness size of a particular span program for nandd . They find that trees with
fewer faults, or critical decisions for a player playing the associated two-player game, are
easier to evaluate on a quantum computer. We show that our algorithm does at least as
well as the algorithm of Zhan et al. for evaluating k-fault trees. To see this, we relate fault
complexity to effective resistances of Gnandd (x) or G0nandd (x).
d
Consider a nandd instance x ∈ {0, 1}2 , and recall the relationship between a nandtree instance and the two-player nand-tree game described in Section 2.3. We call the
sequence of nodes that Player A and Player B choose during the course of a game a path p
— this is just a path from the root of the nand-tree to a leaf node. If x is Z-winnable, we
call PZ (x) the set of paths where Player Z wins, and Player Z never makes a move that
would allow her opponent to win. That is, a path in PA (x) (resp. PB (x)) only encounters
nodes that are themselves the roots of A-winnable (resp. B-winnable) subtrees and never
passes through a node where Player B (resp. Player A) could make a decision to move
to a B-winnable (resp. A-winnable) subtree. Whether a node in the tree is the root of
an A-winnable or B-winnable subtree can be determined by evaluating the subformula
corresponding to that subtree. See Figure 5 for an example of PA . Let νZ (p) be the set
of nodes along a path p at which it is Player Z’s turn. Thus, νA (p) (resp. νB (p)) contains
those nodes in p at even (resp. odd) distance > 0 from the leaves.
Zhan et al. call a node v a fault if one child is the root of an A-winnable tree, while the
other child is the root of a B-winnable tree. Such a node constitutes a critical decision point.
If we let fZ (p) denote the number of faults in νZ (p), we can define the fault complexity
F(x) of input x as1 F(x) = min{FA (x), FB (x)}, where:
FZ (x) =
( max
2 p∈PZ (x) fZ (p)
if x is Z-winnable
otherwise.
∞
(28)
d
For k = 0, . . . , d/2, the set of k-fault trees, Fkd , are those instances x ∈ {0, 1}2 with
log F(x) ≤ k. In these trees, the winning player will encounter at most k fault nodes
on their path to a leaf. Kimmel [19] shows there exists a span program for evaluating
nand-trees whose witness size for an instance x is at most the fault complexity F(x).
∨
Player A’s First Turn
∧
∧
∨
∨
∧
∧
1 1
p1 p2
1
∧
0
0
∨
∧
0
Player B’s First Turn
1 1
p3 p4
∨
∧
0
∧
0
0
Player A’s Second Turn
∧
1
1
∧Player B’s Second Turn
1
0
1
Figure 5: Depiction of a depth-4 nand-tree as a two-player game. Let x be the input to nand4 shown
in the figure. This instance is A-winnable, and PA (x) consists of the paths {p1 , p2 , p3 , p4 }, shown using
solid lines. Fault nodes are those with double circles. Each path in PA (x) encounters two faults at
nodes where Player A makes decisions. Therefore, FA (x) = 4.
We first show a relationship between effective resistance of Gnandd (x) or G0nandd (x)
and F(x):
1
We have actually used the more refined definition of k-fault from [19].
Accepted in
Quantum 2017-08-12, click title to verify
16
d
Lemma 22. For any x ∈ {0, 1}2 , if d is even, then we have Rs,t (Gnandd (x)) ≤ FA (x)
and Rs0 ,t0 (G0nandd (x)) ≤ FB (x), while if d is odd, we have Rs,t (Gnandd (x)) ≤ 2FA (x) and
Rs0 ,t0 (G0nandd (x)) ≤ 2FB (x).
The proof of Lemma 22 can be found in Appendix D. An immediate corollary of Lemma
22 and Theorem 15 is the following.
Corollary 23. The span program PGφ for φ = nandd decides Evalnandd restricted to the
domain X in O(maxx∈X F(x)) queries. In particular, it decides k-fault trees (on domain
Fkd ) in O(2k ) queries.
Proof of Theorem 20. Theorem 20 is now an immediate consequence of Corollary 23, with k
set to log(d), along with the fact from [27] that the classical query complexity of evaluating
such formulas is Ω(dlog log(d) ).
We will use Corollary 23, along with the following claim, to prove Theorem 21:
Claim 24. Let two graphs G1 and G2 each have nodes s and t and let x1 ∈ {0, 1}E(G1 )
and x2 ∈ {0, 1}E(G2 ) . Suppose we create a new graph G by identifying the s nodes and the
t nodes (i.e. connecting the graphs in parallel), then
Cs,t (G(x1 , x2 )) = Cs,t (G1 (x1 )) + Cs,t (G2 (x2 ))
(29)
If we create a new graph G by identifying the t node of G1 with the s node of G2 and
relabeling this node v 6∈ {s, t} (i.e. connecting the graphs in series), then
Cs,t (G(x1 , x2 )) = min{Cs,t (G1 (x1 )), Cs,t (G2 (x2 ))}.
(30)
Proof of Theorem 21. Using Corollary 23, our analysis shows that st-connGnand ,F d can
d 1
be decided in O(1) queries.
We apply Eq. (8) to compare to the analysis of [4]. We must characterize the quantity
max{Cs,t (Gnandd (x)) : x ∈ F1d , and s, t are not connected}, since we already have
max
x∈F1d :s,t are connected
Rs,t (G(x)) = O(1).
(31)
d
We now prove that for every x ∈ {0, 1}2 such that nandd (x) = 0, Cs,t (Gnandd (x)) =
2bd/2c . Thus, for any promise D on the input, as long as there exists some x ∈ D such
that nandd (x) = 0, we have maxx∈D:nandd (x)=0 Cs,t (Gnandd (x)) = 2bd/2c . Intuitively, this
is because every st-cut on any subgraph of Gnandd cuts across 2bd/2c edges of Gnandd .
The proof is by induction on d. For the base even case, d = 0 and Gnand0 is a single
0
edge connecting s and t. The only input x ∈ {0, 1}2 such that nandd (x) = 0 is x = 0,
in which case, the st-cut is κ(s) = 1 and κ(t) = 0, so the cut is across the unique edge in
Gnand0 , so Cs,t (Gnand0 (x)) = 1.
For the induction step, we treat even and odd separately. Suppose d > 0 is odd.
Then nandd (x) = nandd−1 (x0 ) ∧ nandd−1 (x1 ), where x0 = (x1 , . . . , x2d−1 ) and x1 =
(x2d−1 +1 , . . . , x2d ). Thus Gnandd (x) involves composing two graphs Gnandd−1 (x0 ) and
Gnandd−1 (x1 ) in series. Since we are assuming s and t are not connected in Gnandd (x), at
least one of Gnandd−1 (x0 ) and Gnandd−1 (x1 ) must not be connected. Without loss of generality, suppose Gnandd−1 (x0 ) is not connected and Cs,t (Gnandd−1 (x0 )) ≤ Cs,t (Gnandd−1 (x1 )).
By induction, Cs,t (Gnandd−1 (x0 )) = 2(d−1)/2 . Thus using Eq. (30) in Claim 24,
Cs,t (Gnandd (x)) = 2(d−1)/2 = 2bd/2c .
Accepted in
Quantum 2017-08-12, click title to verify
(32)
17
Now suppose d > 0 is even. Then nandd (x) = nandd−1 (x0 ) ∨ nandd−1 (x1 ). Thus
Gnandd (x) involves composing two graphs Gnandd−1 (x0 ) and Gnandd−1 (x1 ) in parallel. Since
we are assuming s and t are not connected in Gnandd (x), both of Gnandd−1 (x0 ) and
Gnandd−1 (x1 ) must not be connected, and so by induction, we have Cs,t (Gnandd−1 (x0 )) =
Cs,t (Gnandd−1 (x1 )) = 2b(d−1)/2c = 2d/2−1 , since d is even. Thus using Eq. (29) in Claim 24,
Cs,t (Gnandd (x)) = 2d/2−1 + 2d/2−1 = 2d/2 = 2bd/2c .
(33)
Therefore, using Eq. (8), we have that the √
analysis of [4] for d-depth nand-trees with
d
inputs in F1 gives a query complexity of O( 2bd/2c ) = O(N 1/4 ), where N = 2d is the
number of input variables. Comparing with our analysis, which gives a query complexity
of O(1), we see there is a polynomial to constant improvement.
5.2 Winning the NAND-tree Game
In this section, we describe a quantum algorithm that can be used to help a player make
decisions while playing the nand-tree game. In particular, we consider the number of
queries to x needed by Player A to make decisions throughout the course of the game in
order to win with probability ≥ 2/3. (In this section, we focus on A-winnable trees, but
the case of B-winnable trees is similar.)
We first describe a naive strategy, which uses a quantum algorithm [21, 23] that decides
if a depth-d tree is winnable with bounded error in O(2d/2 log d) queries. If Player A must
decide to move to node v0 or v1 , she evaluates each subtree rooted at v0 and v1 , amplifying
the success probability to Ω(1/d) by using O(log d) repetitions, and moves to one that
evaluates to 1. Since Player A has O(d) decisions to make, this strategy succeeds with
bounded error, and since evaluating a nand-tree of depth r costs O(2r/2 ) quantum queries,
the total query complexity is:
d
2
d
√
X
2i log d = O 2 2 log d = O
= O 2
N log log N .
O 2
P
d
2
i=0 2
d−2i
2
log d
(34)
i=0
This strategy does not use the fact that some subtrees may be easier to win than others.
For example, if one choice leads to a subtree with all leaves labeled by 1, whereas the other
subtree has all leaves labeled by 0, the player just needs to distinguish these two disparate
cases. More generally, one of the subtrees might have a small positive witness size — i.e.,
it is very winnable — whereas the other has a large positive witness size — i.e., is not very
winnable.
Our strategy will be to move to the subtree whose formula corresponds to a graph with
smaller effective resistance, unless the two subtrees are very close in effective resistance, in
which case it doesn’t matter which one we choose. For a depth d game on instance x, we
show if Rs,t (Gnandd (x)) is small and Player B plays randomly, this strategy does better
than the naive strategy, on average.
We estimate the effective resistance of both subtrees of the current node using the
witness size estimation algorithm of [17]. In particular, in Appendix D.2 we prove:
Lemma 25 (Est Algorithm). Let φ be an and-or formula with constant fan-in l, ∨-depth
d∨ and ∧-depth d∧ . Then the quantum query
complexity of estimating Rs,t (Gφ (x)) (resp.
q
e
Rs,t (G0φ (x))) to relative accuracy is O
Accepted in
1
ε3/2
Quantum 2017-08-12, click title to verify
e
Rs,t (Gφ (x))ld∨ (resp. O
1
ε3/2
q
Rs,t (G0φ (x))ld∧ ).
18
Let Est(x) be the algorithm from Lemma 25 with ε = 13 , and φ = nandd , so l = 2, and
both d∨ and d∧ are at most dd/2e. While estimating the effective resistance of two subtrees,
we only care about which of the subtrees has the smaller effective resistance, so we do not
want to wait for both iterations of Est to terminate. Let p(d)qbe some polynomial function
in d such that Est(x) always terminates after at most p(d) Rs,t (Gφ (x))2d/4 queries, for
d
all x ∈ {0, 1}2 . We define a subroutine, Select(x0 , x1 ), that takes two instances, x0 , x1 ∈
d−1
{0, 1}2 , and outputs a bit b such that Rs,t (Gnandd−1 (xb )) ≤ 2Rs,t (Gnandd−1 (xb̄ )), where
b̄ = b ⊕ 1. Select works as follows. It runs Est(x0 ) and Est(x1 ) in parallel. If one of these
programs, say Est(xb ), outputs some estimate wb , then it terminates the other program
√
after p(d) wb 2d/4 steps. If only the algorithm running on xb has terminated after this
time, it outputs b. If both programs have terminated, it outputs a bit b such that wb ≤ wb̄ .
In Appendix D.3, we prove the following lemma.
d
Lemma 26. Let x0 , x1 ∈ {0, 1}2 be instances of nandd with at least one of them a
0
1
1-instance. Let N = 2d , and wmin
s,t
(Gnandd (x )), Rs,t (Gnandd (x ))}. Then
= min{R
√
e N 1/4 wmin queries to (x0 , x1 ) and outputs b such
Select(x0 , x1 ) terminates after O
that Rs,t (Gnandd (xb )) ≤ 2Rs,t (Gnandd (xb̄ )) with bounded error.
Using Lemma 26, we can prove the following (the inductive proof is in Appendix D.3):
Theorem 27. Let x ∈ {0, 1}N for N = 2d be an A-winnable input to nandd . At every
node v where Player A makes a decision, let Player A use the Select algorithm in the
following way. Let v0 and v1 be the two children of v, with inputs to the respective subtrees
of v0 and v1 given by x0 and x1 respectively. Then Player A moves to vb where b is the
outcome that occurs a majority of times when Select(x0 , x1 ) is run O(log d) times. Then
if Player B, at his decision nodes, chooses left and right with equal
Player A
q
probability,
1/4
e
will win the game with probability at least 2/3, and will use O N
Rs,t (Gnandd (x))
queries on average, where the average is taken over the randomness of Player B’s choices.
6 Acknowledgments
SK is funded by the US Department of Defense. SJ acknowledges funding provided by the
Institute for Quantum Information and Matter, an NSF Physics Frontiers Center (NFS
Grant PHY-1125565) with support of the Gordon and Betty Moore Foundation (GBMF12500028); and an NWO WISE Fellowship.
References
[1] S. Aaronson and S. Ben-David. Sculpting quantum speedups. In Proceedings of the
31st Conference on Computational Complexity, CCC ’16, pages 26:1–26:28, Germany,
2016. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik. ISBN 978-3-95977-008-8.
DOI: 10.4230/LIPIcs.CCC.2016.26. URL http://dl.acm.org/citation.cfm?id=2982445.
2982471.
[2] D. Aldous and J. Fill. Reversible Markov chains and random walks on graphs, 2002.
Unfinished monograph, recompiled 2014, available at http://www.stat.berkeley.edu/
~aldous/RWG/book.html.
[3] A. Belovs. Span programs for functions with constant-sized 1-certificates. In Proceedings of the 44th Symposium on Theory of Computing (STOC 2012), pages 77–84, 2012.
DOI: 10.1145/2213977.2213985.
Accepted in
Quantum 2017-08-12, click title to verify
19
[4] A. Belovs and B. W. Reichardt. Span programs and quantum algorithms for stconnectivity and claw detection. In Proceedings of the 20th European Symposium on
Algorithms (ESA 2012), pages 193–204, 2012. DOI: 10.1007/978-3-642-33090-2_18.
[5] S. Ben-David and R. Kothari. Randomized query complexity of sabotaged and composed functions. In Proceedings of the 43th International Colloquium on Automata,
Languages and Programming (ICALP 2016), volume 55, pages 60:1–60:14, 2016. DOI:
10.4230/LIPIcs.ICALP.2016.60.
[6] M. Boyer, G. Brassard, P. Høyer, and A. Tapp. Tight bounds on quantum
searching.
Fortschritte der Physik, 46(4-5):493–505, 1998.
ISSN 1521-3978.
DOI:
10.1002/(SICI)1521-3978(199806)46:4/5<493::AID-PROP493>3.0.CO;2P.
URL
http://dx.doi.org/10.1002/(SICI)1521-3978(199806)46:4/5<493::
AID-PROP493>3.0.CO;2-P.
[7] G. Brassard, P. Høyer, M. Mosca, and A. Tapp. Quantum amplitude amplification
and estimation. Contemporary Mathematics, 305:53–74, 2002.
[8] C. Cade, A. Montanaro, and A. Belovs. Time and space efficient quantum algorithms
for detecting cycles and testing bipartiteness, 2016. arXiv:1610.00581.
[9] A. K. Chandra, P. Raghavan, W. L. Ruzzo, R. Smolensky, and P. Tiwari. The electrical
resistance of a graph captures its commute and cover times. Computational Complexity,
6(4):312–340, 1996. DOI: 10.1007/BF01270385.
[10] G. A. Dirac. A property of 4-chromatic graphs and some remarks on critical graphs.
Journal of the London Mathematical Society, 1(1):85–92, 1952. DOI: 10.1112/jlms/s127.1.85.
[11] P. G. Doyle and J. L. Snell. Random Walks and Electrical Networks, volume 22 of The
Carus Mathematical Monographs. The Mathematical Association of America, 1984.
[12] R. J. Duffin.
Topology of series-parallel networks.
Journal of Mathematical Analysis and Applications, 10(2):303–318, 1965.
ISSN 0022-247X.
DOI:
http://dx.doi.org/10.1016/0022-247X(65)90125-3. URL http://www.sciencedirect.
com/science/article/pii/0022247X65901253.
[13] C. Dürr, M. Heiligman, P. Høyer, and M. Mhalla. Quantum query complexity of
some graph problems. SIAM Journal on Computing, 35(6):1310–1328, 2006. DOI:
10.1137/050644719.
[14] E. Farhi, J. Goldstone, and S. Gutmann.
A quantum algorithm for the
Hamiltonian NAND tree.
Theory of Computing, 4(8):169–190, 2008.
DOI:
10.4086/toc.2008.v004a008. arXiv:quant-ph/0702144.
[15] L. K. Grover. A fast quantum mechanical algorithm for database search. In Proceedings
of the 28th annual ACM Symposium on Theory of computing (STOC 1996), pages 212–
219. ACM, 1996. DOI: 10.1145/237814.237866.
[16] R. Heiman and A. Wigderson. Randomized vs. deterministic decision tree complexity
for read-once Boolean functions. Computational Complexity, 1(4):311–329, 1991. DOI:
10.1007/BF01212962.
[17] T. Ito and S. Jeffery. Approximate span programs. In Proceedings of the 43rd International Colloquium on Automata, Languages, and Programming (ICALP 2016), pages
12:1–12:14, 2016. DOI: 10.4230/LIPIcs.ICALP.2016.12. arXiv:1507.00432.
[18] M. Karchmer and A. Wigderson. On span programs. In Proceedings of the IEEE 8th
Annual Conference on Structure in Complexity Theory, pages 102–111, 1993. DOI:
10.1109/SCT.1993.336536.
[19] S. Kimmel. Quantum adversary (upper) bound. Chicago Journal of Theoretical Computer Science, 2013(4), 2011. DOI: 10.4086/cjtcs.2013.004.
Accepted in
Quantum 2017-08-12, click title to verify
20
[20] T. Lee, R. Mittal, B. W. Reichardt, R. Špalek, and M. Szegedy. Quantum query
complexity of state conversion. In Proceedings of the 52nd Annual IEEE Symposium on Foundations of Computer Science (FOCS 2011), pages 344–353, 2011. DOI:
10.1109/FOCS.2011.75.
[21] B. W. Reichardt. Span programs and quantum query complexity: The general adversary bound is nearly tight for every Boolean function. In Proceedings of the 50th
IEEE Symposium on Foundations of Computer Science (FOCS 2009), pages 544–551,
2009. DOI: 10.1109/FOCS.2009.55. arXiv:quant-ph/0904.2759.
[22] B. W. Reichardt. Span programs and quantum query algorithms. Electronic Colloquium on Computational Complexity (ECCC), 17:110, 2010.
[23] B. W. Reichardt. Reflections for quantum query algorithms. In Proceedings of the
22nd Annual ACM-SIAM Symposium on Discrete Algorithms (SODA 2011), pages
560–569. SIAM, 2011.
[24] B. W. Reichardt and R. Špalek.
Span-program-based quantum algorithm
for evaluating formulas.
Theory of Computing, 8(13):291–319, 2012.
DOI:
10.4086/toc.2012.v008a013.
[25] M. Saks and A. Wigderson. Probabilistic Boolean decision trees and the complexity of evaluating game trees. In Proceedings of the 27th Annual Symposium on
Foundations of Computer Science (FOCS 1986), pages 29–38. IEEE, 1986. DOI:
10.1109/SFCS.1986.44.
[26] J. Valdes, R. E. Tarjan, and E. L. Lawler. The recognition of series parallel digraphs.
In Proceedings of the 11th Annual ACM Symposium on Theory of Computing (STOC
1979), pages 1–12. ACM, 1979. DOI: 10.1145/800135.804393. URL http://doi.acm.
org/10.1145/800135.804393.
[27] B. Zhan, S. Kimmel, and A. Hassidim. Super-polynomial quantum speed-ups for
Boolean evaluation trees with hidden structure. In Proceedings of the 3rd Innovations
in Theoretical Computer Science Conference (ITCS 2012), pages 249–265, New York,
NY, USA, 2012. ACM. ISBN 978-1-4503-1115-1. DOI: 10.1145/2090236.2090258. URL
http://doi.acm.org/10.1145/2090236.2090258.
A Analysis of the Span Program for st-Connectivity
In this section, we analyze the complexity of our span-program-based algorithms, proving
Lemma 11, first stated in Section 3, which relates witness sizes of the span program PG,c
to the effective resistance of graphs related to G.
We need the concept of a circulation, which is like a flow but with no source and no
sink.
→
−
Definition 28 (Circulation). A circulation on a graph G is a function θ : E (G) → R
such that:
→
−
1. For all (u, v, λ) ∈ E (G), θ(u, v, λ) = −θ(v, u, λ);
2. for all u ∈ V (G),
P
−
→
θ(u, v, λ)
v,λ:(u,vλ)∈ E (G)
= 0.
The following easily verified observations will be useful in several of the remaining
proofs in this section.
Claim 29. Let θ be a unit st-flow in some multigraph G. We can consider the corresponding
P
→
vector |θi = (u,v,λ)∈−
θ(u, v, λ)|u, v, λi. Then |θi can be written as a linear combination
E (G)
Accepted in
Quantum 2017-08-12, click title to verify
21
of vectors corresponding to self-avoiding st-paths and cycles that are edge-disjoint from
these paths.
Let σ be a circulation on G. Then |σi can be written as a linear combination of cycles
in G. Furthermore, |σi can be written as a linear combination of cycles such that each
cycle goes around a face of G.
The next claim shows a direct correspondence between positive witnesses, and st-flows.
Claim 30. Fix a span program PG,c as in (6). Call |wi ∈ H a positive witness in
PG,c if A|wi = τ (note that such a |wi is not necessarily a positive witness for any
P
√θ(u,v,λ) |u, v, λi
→
particular input x). Then if θ is a unit st-flow in G, 12 (u,v,λ)∈−
E (G)
c({u,v},λ)
is a positivepwitness in PG,c , and furthermore, if |wi is a positive witness in PG,c , then
θ(u, v, λ) = c({u, v}, λ)(hw|u, v, λi − hw|v, u, λi) is a unit st-flow in G.
Proof. The proof is a straightforward calculation. Let θ be a unit st-flow on G. Then
1
A
=
=
1
2
2
X
−
→
(u,v,λ)∈ E (G)
X
p
θ(u, v, λ)
|u, v, λi
c({u, v}, λ)
θ(u, v, λ)(|ui − |vi)
−
→
(u,v,λ)∈ E (G)
X
1
2 u∈V (G)
X
θ(u, v, λ) |ui +
−
→
v,λ:(u,v,λ)∈ E (G)
X
1
2 v∈V (G)
X
θ(v, u, λ) |vi
−
→
u,λ:(v,u,λ)∈ E (G)
1
1
(|si − |ti) + (|si − |ti) = τ.
(35)
2
2
P
→
Above we have used that θ(u, v, λ) = −θ(v, u, λ), and v,λ:(u,v,λ)∈−
θ(u, v, λ) = 0 when
E (G)
u 6∈ {s, t}, 1 when u = s, and −1 when u = t.
To provepthe second half of the claim, let |wi be such that A|wi = τ , and define
θ(u, v, λ) = c({u, v}, λ)(hw|u, v, λi − hw|v, u, λi). We immediately see that θ(u, v, λ) =
−θ(v, u, λ) for all (u, v, λ). Furthermore, we have:
=
|si − |ti = A|wi =
X
u∈V (G)
−
q
X
c({u, v}, λ)hu, v, λ|wi |ui
−
→
(u,v,λ)∈ E (G)
X
v∈V (G)
X
q
c({u, v}, λ)hu, v, λ|wi |vi
−
→
(u,v,λ)∈ E (G)
=
X
u∈V (G)
=
X
u∈V (G)
X
q
c({u, v}, λ)(hu, v, λ|wi − hv, u, λ|wi) |ui
−
→
(u,v,λ)∈ E (G)
X
−
→
(u,v,λ)∈ E (G)
θ(u, v, λ) |ui.
→
Thus, for all u ∈ V (G(x))\{s, t}, (u,v,λ)∈−
θ(u, v, λ) = 0, and
E (G)
P
−
→
θ(v, t, λ) = 1. Thus, θ is a unit st-flow on G.
(v,t,λ)∈ E (G)
P
(36)
P
−
→
θ(s, v, λ)
(s,v,λ)∈ E (G)
=
The next claim shows a direct correspondence between negative witnesses, and s0 t0 -flows.
Accepted in
Quantum 2017-08-12, click title to verify
22
Claim 31. For a planar graph G, fix a span program PG,c as in (6). Call a linear function
ω : V (G) → R a negative witness if ωτ = 1. Then θ((u, v, λ)† ) = ω(u) − ω(v) is a unit
s0 t0 -flow on G0 , and furthermore, for every s0 t0 -flow θ on G0 there is a negative witness ω
→
−
such that θ((u, v, λ)† ) = ω(u) − ω(v) for all (u, v, λ) ∈ E (G).
w2
w1
w3
w4
v0
w6
u0
w5
Figure 6: The duality between a cycle and a star.
Proof. When we consider the edges of G as directed edges, we assign edge directions to
the dual by orienting each dual edge π/2 radians counter-clockwise from the primal edge.
Note that without loss of generality, if ω is a negative witness, we can assume ω(s) = 1
and ω(t) = 0. This is because kωAk and ωAΠH(x) are invariant under affine transformations of ω.
→
−
We first show that if ω is a negative witness in PG,c , then θ : E (G0 ) → R defined
θ((u, v, λ)† ) = ω(u) − ω(v) is a unit s0 t0 -flow on G0 . To begin with, we will define
→
− †
θ0 ((u, v, λ)† ) = ω(u) − ω(v) on E (G ), so θ0 agrees with θ everywhere θ is defined, and in
addition, θ0 (s0 , t0 , ∅) = θ0 ((s, t, ∅)† ) = ω(s) − ω(t) = 1, and θ0 (t0 , s0 , ∅) = −1. Then clearly
†
we have θ0 (u0 , v 0 , λ) = −θ0 (v 0 , u0 , λ) for all ({u0 , v 0 }, λ) ∈ E(G ).
†
Next, every v 0 ∈ V (G ) corresponds to a face fv0 of G, and the edges coming
0
out of v are dual to edges going clockwise around the face fv0 (see Figure 6). If
(w1 , w2 , λ1 ), . . . , (wk , wk+1 , λk ), for wk+1 = w1 , are the directed edges going clockwise
around fv0 , then we have:
0=
k
X
k
X
i=1
i=1
(ω(wi ) − ω(wi+1 )) =
θ0 ((wi , wi+1 , λi )† ) =
θ0 (v 0 , u0 , λ).
X
(37)
u0 ,λ:
0
({v 0 ,u0 },λ)∈E(G )
Thus, θ0 is a circulation. Then, since θ0 (s0 , t0 , ∅) = 1, if we remove the flow on this edge,
which recovers θ, we get a unit s0 t0 -flow on G0 .
Next we will show that if θ is a unit s0 t0 -flow on G0 , then there exists a negative witness
→
−
ω in PG,c such that for all (u, v, λ) ∈ E (G), θ((u, v, λ)† ) = ω(u) − ω(v).
†
Define θ0 to be the circulation on G obtained from defining θ0 (u0 , v 0 , λ) = θ(u0 , v 0 , λ)
→
−
for all (u0 , v 0 , λ) ∈ E (G0 ), and θ0 (s0 , t0 , ∅) = −θ0 (t0 , s0 , ∅) = 1. Then if we define |θ0 i =
P
−
→ † θ 0 (u, v, λ)|u, v, λi, we can express |θ 0 i as a linear combination of cycles around
(u,v,λ)∈ E (G )
P
P
→
−
←
−
†
the faces of G , |θ0 i = f ∈F (G† ) αf | C f i + f ∈F (G† ) αf0 | C f i, where if wk+1 = w1 and
P
→
−
(w1 , w2 , λ1 ), . . . , (wk , wk+1 , λk ) is a clockwise cycle around f , | C f i = ki=1 |wi , wi+1 , λi i
P
←
−
is the clockwise cycle around the face f , and | C f i = ki=1 |wi+1 , wi , λi i is the counterclockwise cycle around f . There is a one-to-one correspondance between vertices in
†
V (G) = V (G) and faces in F (G ), so we can define ω : V (G) → R by ω(vf ) = 12 (αf − αf0 ).
→
−
We claim that for all (u, v, λ) ∈ E (G), ω(u) − ω(v) = θ0 ((u, v, λ)† ). Let (u0 , v 0 , λ) be
→
− †
†
any edge in E (G ). This edge is part of a clockwise cycle around one face in G , call it f ,
Accepted in
Quantum 2017-08-12, click title to verify
23
†
and a counter clockwise cycle around one face in G , call it g. Since these are the only
two faces containing the edge (u0 , v 0 , λ), we must have θ0 (u0 , v 0 , λ) = hu0 , v 0 , λ|θ0 i = αf + αg0 .
Since θ0 (u0 , v 0 , λ) = −θ0 (v 0 , u0 , λ), we have αf + αg0 = −αf0 − αg . Thus:
ω(vf ) − ω(vg ) =
1 0 0 0
1
αf − αf0 − αg + αg0 =
θ (u , v , λ) − θ0 (v 0 , u0 , λ) = θ0 ((vf , vg , d)† ).
2
2
(38)
In particular, this means that ω(s) − ω(t) = θ0 ((s, t, ∅)† ) = θ0 (s0 , t0 , ∅) = 1, so ω is a negative
→
−
witness, and for all (u, v, λ) ∈ E (G), ω(u) − ω(v) = θ((u, v, λ)† ).
Now we can prove the main result of this section, Lemma 11:
Lemma 11. Let G be a planar multigraph with s, t ∈ V (G) such that G ∪ {{s, t}} is also
planar, and let c be a weight function on E(G). Let x ∈ {0, 1}E(G) . Then w+ (x, PG,c ) =
1
0
0
0 0
2 Rs,t (G(x), c) and w− (x, PG,c ) = 2Rs ,t (G (x), c ).
Proof. If x is a 1-instance, s and t are connected in G(x), so there exists a unit st-flow
→
−
on G(x), which is a unit st-flow on G that is supported only on E (G(x)). Let θ be
P
θ(u,v,λ)2
the flow on G(x) such that Rs,t (G(x), c) = ({u,v},λ)∈E(G(x)) c({u,v},λ)
. By Claim 30,
|wi = 12 (u,v,λ) √θ(u,v,λ) |u, v, λi is a positive witness in PG,c , and since θ is supported
c({u,v},λ)
→
−
on E (G(x)), |wi ∈ H(x), and so |wi is a positive witness for x in PG,c . Thus
P
w+ (x, PG,c ) ≤ k|wik2 =
1
4
X
−
→
(u,v,λ)∈ E (G(x))
1
c({u, v}, λ)θ(u, v, λ)2 = Rs,t (G(x), c).
2
(39)
p On the other hand, let |wi be an optimal positive witness for x. By Claim 30, θ(u, v, λ) =
c({u, v}, λ)(hu, v, λ|wi−hv, u, λ|wi) is a unit st-flow on G, and since |wi ∈ H(x), θ(u, v, λ)
→
−
is only non-zero on E (G(x)), so θ is a unit st-flow on G(x). Thus,
Rs,t (G(x), c) ≤
=
θ(u, v, λ)2
1
=
c({u, v}, λ)
2
({u,v},λ)∈E(G(x))
X
X
X
X
hu, v, λ|wi2 −
(hu, v, λ|wi − hv, u, λ|wi)2
−
→
(u,v,λ)∈ E (G(x))
hu, v, λ|wihv, u, λ|wi ≤ 2 k|wik2 (40)
−
→
(u,v,λ)∈ E (G(x))
−
→
(u,v,λ)∈ E (G(x))
where the last inequality is by Cauchy-Schwarz. Thus, w+ (x, PG,c ) = 12 Rs,t (G(x)).
Now we prove that w− (x, PG,c ) = 2Rs0 ,t0 (G0 (x), c0 ). Let x ∈ {0, 1}E(G) be such that s
and t are not connected in G(x). Fix an optimal negative witness ω for x. By Claim 31
→
−
the linear function θ : E (G0 ) → R defined by θ((u, v, λ)† ) = ω(u) − ω(v) is a unit s0 t0 -flow
on G0 . Since ω is a negative witness for x, we also have:
0 = ωAΠH(x)
2
=
X
c({u, v}, λ)(ω(u) − ω(v))2
−
→
(u,v,λ)∈ E (G(x))
=
c({u, v}, λ)θ((u, v, λ)† )2
X
−
→
(u,v,λ)∈ E (G(x))
X
=
−
→
−
→
(u0 ,v 0 ,λ)∈ E (G0 )\ E (G0 (x))
Accepted in
Quantum 2017-08-12, click title to verify
θ(u0 , v 0 , λ)2
,
c0 ({u0 , v 0 }, λ)
(41)
24
→
−
→
−
since (u, v, λ) ∈ E (G(x)) exactly when (u, v, λ)† 6∈ E (G0 (x)). So θ is only supported on
→
− 0
E (G (x)), and so it is a unit s0 t0 -flow on G0 (x). Thus
w− (x, PG,c ) = kωAk2 =
X
c({u, v}, λ)(ω(u) − ω(v))2
−
→
(u,v,λ)∈ E (G)
X
=
−
→
(u0 ,v 0 ,λ)∈ E (G0 (x))
θ(u0 , v 0 , λ)2
≥ 2Rs0 ,t0 (G0 (x), c0 ).
c0 ({u0 , v 0 }, λ)
(42)
For the other direction, let θ be an s0 t0 -flow in G0 (x) with minimal energy. By Claim 31,
there is a negative witness ω such that θ((u, v, λ)† ) = ω(u) − ω(v). Since θ is supported on
→
−
→
−
edges (u0 , v 0 , λ) ∈ E (G0 (x)), which are exactly those edges such that (u0 , v 0 , λ)† 6∈ E (G(x)),
we have
0=
c({u, v}, λ)θ((u, v, λ)† )2 =
X
(u,v,λ)
−
→
∈ E (G(x))
X
c({u, v}, λ)(ω(u) − ω(v))2 = ωAΠH(x)
2
,
(u,v,λ)
−
→
∈ E (G(x))
(43)
so ω is a negative witness for x in PG,c . Thus:
w− (x, PG,c ) ≤ kωAk2 =
X
c({u, v}, λ)(ω(u) − ω(v))2
−
→
(u,v,λ)∈ E (G)
=
X
−
→
(u0 ,v 0 ,λ)∈ E (G0 (x))
θ(u0 , v 0 , λ)2
= 2Rs0 ,t0 (G0 (x), c0 ),
c0 ({u0 , v 0 }, λ)
(44)
completing the proof.
A.1 Time and Space Analysis of the Span Program Algorithm for st-Connectivity
In this section, we will give an upper bound on the time complexity of st-connG in terms
of the time complexity of implementing a step of a discrete-time quantum walk on G. The
analysis follows relatively straightforwardly from [4, Section 5.3], but we include it here for
completeness. At the end of this section, we show the space complexity of the algorithm
is O(max{log |E(G)|, log |V (G)|}).
We first describe the algorithm that can be derived from a span program, following the
conventions of [17]. Throughout this section, we will let ΠS denote the orthogonal projector
onto an inner product space S. For a span program P = (H, U, A, τ ), the corresponding
algorithm performs phase estimation on the unitary (2ΠH(x) − I)(2Πker A − I) applied
to initial state |w0 i = A+ τ , where ΠH(x) denotes the orthogonal projector onto H(x),
and Πker A denotes the orthogonal projector onto the kernel of A, and A+ denotes the
pseudo-inverse of A. To decide
q a function f on domain D, it is sufficient to perform phase
estimation to precision O
maxx∈D:f (x)=1 w+ (x) × maxx∈D:f (x)=0 w− (x) .
In case of the st-connectivity span program PG,c in (6), it is a simple exercise to see
that 2ΠH(x) − I can be implemented in O(1) quantum operations, including 2 queries to
x. The reflection 2Πker A − I is independent of x, and so requires 0 queries to implement,
however, it could still require a number of gates that grows quickly with the size of G. We
will show that implementing 2Πker A − I can be reduced to implementing a discrete-time
quantum walk on G, a task which could be quite easy, depending on the structure of G
(for example, in the case that G is a complete graph on n vertices, this can be done in
O(log n) gates [4]).
Accepted in
Quantum 2017-08-12, click title to verify
25
For a multigraph G and weight function c, we define a quantum walk step on G to be
a unitary UG,c that acts as follows for any u ∈ V (G):
1
UG,c : |ui|0i 7→ qP
q
X
−
→
c({u, v}, λ)
−
→
v,λ:(u,v,λ)∈ E (G)
v,λ:(u,v,λ)∈ E (G)
c({u, v}, λ)|ui|u, v, λi.
(45)
Theorem 13. Let PG,c = (H, U, A, τ ) be defined as in (6). Let SG,c be an upper bound on
the time complexity of implementing UG,c . If G has the property that G ∪ {{s, t}} is planar,
then the time complexity of deciding st-connG,D is at most
O min SG,c
c
!
r
max
x∈D:s,t are connected
Rs,t (G(x), c) ×
max
x∈D:s,t are not connected
Rs0 ,t0 (G0 (x), c0 ) .
(12)
This theorem follows from Lemma 32, stated below, and Lemma 33, which deals with the
construction of the algorithm’s initial state.
Lemma 32. Let A be defined as in (6). Let SG,c be an upper bound on the time complexity
of implementing UG,c . Then 2Πker A − I can be implemented in time complexity O(SG,c ).
Proof. This analysis follows [4] (see also [17]). Let
X
d(u) =
c({u, v}, λ).
(46)
−
→
v,λ:(u,v,λ)∈ E (G)
Define spaces Z and Y as follows.
X
−
→
v,λ:(u,v,λ)∈ E (G)
p
Z = span |zu i :=
c({u, v}, λ)
p
(|0, u, u, v, λi + |1, u, v, u, λi) : u ∈ V (G)
2d(u)
(47)
→
−
|0, u, u, v, λi − |1, v, u, v, λi
√
Y = span |yu,v,λ i :=
: (u, v, λ) ∈ E (G)
2
Define isometries whose column-spaces are Z and Y respectively:
MZ =
X
|zu ihu| and
X
MY =
|yu,v,λ ihu, v, λ|.
(48)
(49)
−
→
(u,v,λ)∈ E (G)
u∈V (G)
Now we note that for any ({u, v}, λ) ∈ E(G), we have the following:
p
c({u, v}, λ)
p
,
2 d(u)
hzu |yu,v,λ i =
p
and
hzv |yu,v,λ i = −
c({u, v}, λ)
p
.
2 d(v)
(50)
Thus, we can calculate:
MZ† MY
=
=
−
→
(u,v,λ)∈ E (G)
X
|u0 ihu0 |
p
u0 ∈V (G)
=
Accepted in
!
q
|ui
|vi
p
− p
c({u, v}, λ)hu, v, λ|
2 d(u) 2 d(v)
X
2 d(u0 )
X
q
c({u, v}, λ)(|ui − |vi)hu, v, λ|
−
→
(u,v,λ)∈ E (G)
|u0 ihu0 |
p
A.
2 d(u0 )
u0 ∈V (G)
X
Quantum 2017-08-12, click title to verify
(51)
26
Note that the rows of MZ† MY are non-zero multiples of the rows of A, so row(MZ† MY ) =
row(A), and thus ker(MZ† MY ) = ker A.
Define W = (2ΠZ − I)(2ΠY − I). We now claim that MY maps ker A to the (−1)eigenspace of W , and (ker A)⊥ to the 1-eigenspace of W , so that 2Πker A − I = MY† W MY .
To see this, note that if |ψi ∈ ker A, then |ψi ∈ ker(MZ† MY ) so MY |ψi ∈ ker MZ† = Z ⊥ .
Thus MY |ψi ∈ Y ∩ Z ⊥ , which is in the (−1)-eigenspace of W .
Next, suppose |ψi ∈ (ker A)⊥ = (ker(MZ† MY ))⊥ , so since MY is an isometry, MY |ψi ∈
(ker MZ )⊥ = rowMZ = Z. Thus MY |ψi ∈ Y ∩ Z, which is in the 1-eigenspace of W .
Thus, we can implement 2Πker A − I by MY† W MY . It only remains to argue that each
of MY , 2ΠZ − I and 2ΠY − I can be implemented in time complexity at most O(SG,c ).
We first show that we can implement the isometry MY , or rather a unitary UY that
acts as |0i|0i|u, v, λi 7→ MY |u, v, λi = |yu,v,λ i. First, use HX on the first qubit to perform
the map:
|0i|0i|u, v, λi 7→ |−i|0i|u, v, λi.
(52)
Conditioned on the value of the first register, copy either u or v into the second register to
get:
1
√ (|0, u, u, v, λi − |1, v, u, v, λi) = |yu,v,λ i.
2
(53)
Thus, we can implement UY in the time it takes to write down a vertex of G, O(log |V (G)|),
which is at most O(SG,c ). Using the ability to implement UY , we can implement 2ΠY − I
as UY RY UY† , where RY is the reflection that acts as the identity on computational basis
states of the form |0i|0i|u, v, λi, and reflects computational basis states without this form.
Next, we implement a unitary UZ that acts as |0i|ui|0i 7→ MZ |ui = |zu i. First, use the
quantum walk step UG,c , which can be implemented in time SG,c , to perform:
1
|+i|ui|0i 7→ p
2 d(u)
X
q
c({u, v}, λ)(|0i + |1i)|ui|u, v, λi.
(54)
−
→
v,λ:(u,v,λ)∈ E (G)
Conditioned on the bit in the first register, swap the third and fourth registers, to get:
1
2 d(u)
p
X
q
c({u, v}, λ)(|0i|ui|u, v, λi + |1i|ui|v, u, λi) = |zu i.
(55)
−
→
v,λ:(u,v,λ)∈ E (G)
The total cost of implementing UZ is O(SG,c + log |V (G)|) = O(SG,c ). Thus, we can
implement 2ΠZ − I in O(SG,c ) quantum gates.
Lemma 33. Let A and τ be defined as in (6). Let SG,c be an upper bound on the complexity
|w0 i
of implementing UG,c . Then the initial state of the algorithm, k|w
where |w0 i = A+ τ ,
0 ik
can be approximated in time O(SG,c ).
Proof. Without loss of generality, we can assume that G includes the edge ({s, t}, ∅) (we
can simply not include it in any subgraph). Furthermore, we set c({s, t}, ∅) = 1/r, for
some positive r to be specified later, so that A|s, t, ∅i = r−1/2 τ . This has no effect on other
edges in G. Note that
Π(ker A)⊥ |s, t, ∅i = A+ A|s, t, ∅i = r−1/2 A+ τ = r−1/2 |w0 i,
Accepted in
Quantum 2017-08-12, click title to verify
(56)
27
so
|s, t, ∅i = r−1/2 |w0 i + |w0⊥ i
(57)
for some |w0⊥ i ∈ ker A. Thus, constant precision phase estimation on 2Πker A − I maps
|s, t, ∅i to
r−1/2 |0i|w0 i + |1i|w0⊥ i.
(58)
Using quantum amplitude amplification [7], we can amplify the amplitude on the |0i|w0 i
part of this arbitrarily close to 1 using a number of calls to 2Πker A − I proportional to
−1
r−1/2 |w0 i
.
In fact, it is straightforward to show that for any |µi ∈ rowA, the vector |νi with
smallest norm that satisfies A|νi = |µi, is A+ |µi [17]. Using this fact along with Claim 30
and Definition 9, we have k|w0 ik2 = Rs,t (G, c).
Let R = Rs,t (G\{({s, t}, ∅)}, c) be the effective resistance of G without the edge (s, t, ∅).
Now we can think of ({s, t}, ∅) and G \ {({s, t}, ∅)} as two graphs in parallel, so using
Claim 5, we have
k|w0 ik2 =
1
.
1/R + 1/r
(59)
Setting r = R, we have k|w0 ik2 = R/2 and kr−1/2 |w0 ik−1 = O(1). Thus, using O(1) calls
to 2Πker A − I, we can approximate the initial state |w0 i.
Finally, we note that the space required by the algorithm, in addition to whatever
auxiliary space we need to implement UG,c , is O(max{log |E(G)|, log |V (G)|}). UY and
UG,c each act on a Hilbert space of dimension less than 4|V (G)|2 |E(G)|, so can in principle be implemented on O(max{log |E(G)|, log |V (G)|}) qubits, however, a time-efficient
0
implementation of UG,c may also make use of some number SG,c
of auxiliary qubits. We
use these unitaries to perform phase estimation on (2ΠH(x) − I)(2Πker A − I) to precision
O min
c
r
!
max
x∈D:φ(x)=1
Rs,t (Gφ (x), c) ×
max
x∈D:φ(x)=0
Rs0 ,t0 (G0φ (x), c0 )
= O(|E(G)|).
(60)
Thus we need O(log(|E(G)|) qubits to store the output of the phase estimation. Putting
everything together gives the claimed space complexity.
B Formula Evaluation and st-Connectivity
In this section, we prove the correspondence between evaluating the formula φ, and solving
st-connectivity on the graph Gφ . We first give a formal definition of Gφ .
Definition 34 (Gφ ). If φ = xi is a single-variable formula, then V (Gφ ) = {s, t} and
E(Gφ ) = {({s, t}, xi )}.
If φ = φ1 ∧ · · · ∧ φl , then define V (Gφ ) = {(i, v) : i ∈ [l], v ∈ V (Gφi ) \ {s, t}} ∪
{s, s2 , . . . , sl , t} and, letting s1 = s and sl+1 = t, define:
E(Gφ ) = {({(i, u), (i, v)}, xj ) : i ∈ [l], u, v ∈ V (Gφi ) \ {s, t}, ({u, v}, xj ) ∈ E(Gφi )}
∪ {({(i, u), si }, xj ) : i ∈ [l], u ∈ V (Gφi ), ({s, u}, xj ) ∈ E(Gφi )}
∪ {({(i, u), si+1 }, xj ) : i ∈ [l], u ∈ V (Gφi ), ({t, u}, xj ) ∈ E(Gφi )}.
Accepted in
Quantum 2017-08-12, click title to verify
(61)
28
If φ = φ1 ∨ · · · ∨ φl define V (Gφ ) = {(i, v) : i ∈ [l], v ∈ V (Gφi ) \ {s, t}} ∪ {s, t} and
E(Gφ ) ={({(i, u), (i, v)}, xj ) : i ∈ [l], u, v ∈ V (Gφi ) \ {s, t}, ({u, v}, xj ) ∈ E(Gφi )}
∪ {({(i, u), s}, xj ) : i ∈ [l], u ∈ V (Gφi ), ({u, s}, xj ) ∈ E(Gφi )}
∪ {({(i, u), t}, xj ) : i ∈ [l], u ∈ V (Gφi ), ({u, t}, xj ) ∈ E(Gφi )}.
(62)
In order to prove Lemma 14, we will first prove Lemma 35:
Lemma 35. For an and-or formula φ on {0, 1}N , define φ0 to be the formula obtained by
replacing ∨-nodes with ∧-nodes and ∧-nodes with ∨-nodes in φ. Then for all x ∈ {0, 1}N ,
if x̄ denotes the bitwise complement of x, then φ0 (x) = ¬φ(x̄). Furthermore up to an
isomorphism that maps s to s0 , t to t0 , and an edge labeled by any label λ to an edge labeled
by λ, we have G0φ = Gφ0 and G0φ (x) = Gφ0 (x̄).
Proof. The first part of the proof is by induction. Suppose φ has depth 0, so φ = xi for
some variable xi . Then φ0 (x) = φ(x) = ¬(φ(x̄)). So suppose φ = φ1 ∧ · · · ∧ φl . Then
φ0 = φ01 ∨ · · · ∨ φ0l . Then by the induction hypothesis,
φ0 (x) = φ01 (x) ∨ · · · ∨ φ0l (x) = (¬φ1 (x̄)) ∨ · · · ∨ (¬φl (x̄)) = ¬(φ1 (x̄) ∧ · · · ∧ φl (x̄)) = ¬φ(x̄)
(63)
where the second to last equality is de Morgan’s law. The case φ = φ1 ∨ · · · ∨ φl is similar.
G0φ1
s
G φ1
G φ2
s0
G0φ2
G0φl
..
.
..
.
t0
G φl
t
Figure 7: Gφ shown in black, and its dual, Gφ0 , shown in grey. The thick lines represent graphs. Edges
in Gφi are dual to edges in G0φi , and the dotted edge ({s, t}, ∅) is dual to ({s0 , t0 }, ∅).
†
We will now prove that Gφ = Gφ0 , and furthermore, dual edges have the same label, by
induction on the depth of φ, from which the result follows immediately.
If φ = xi is a depth-0 formula, then φ0 = xi . In that case, Gφ is just an edge from s to
t, labeled by xi , and G0φ is just an edge from s0 to t0 labeled xi , so G0φ = Gφ0 .
For the inductive step, to show that Gφ and Gφ0 are dual, and therefore G0φ = Gφ0 . It
suffices to exhibit a bijection ζ : V (Gφ0 ) → F (Gφ ) such that ({u, v}, xj ) ∈ E(Gφ0 ) if and
only if the faces ζ(u) and ζ(v) are separated by an edge in E(Gφ ) with the label xj . We
first consider the case that φ = φ1 ∧ · · · ∧ φl , so φ0 = φ01 ∨ · · · ∨ φ0l . Then, Gφ consists of
the graphs Gφ1 , . . . , Gφl , chained together in series as in Figure 7, with an additional edge
from s to t, so the faces of Gφ are exactly all the interior faces of each Gφi , as well as the
0
0
two faces on either side of the st-edge ({s, t}, ∅), which we will denote by f s and f t . That
is, adding an i to the label of each internal face of Gφi :
0
0
0
0
F (Gφ ) = {(i, f ) : i ∈ [l], f ∈ F (Gφi ) \ {f s , f t }} ∪ {f s , f t },
Accepted in
Quantum 2017-08-12, click title to verify
(64)
29
0
0
since F (Gφi ) \ {f s , f t } = F (Gφi ) \ {f E }, where f E is the external face. Since φ0 =
φ01 ∨ · · · ∨ φ0l we also have
V (Gφ0 ) = V (Gφ0 ) = {(i, v) : i ∈ [l], v ∈ V (Gφ0i ) \ {s, t}} ∪ {s0 , t0 },
(65)
where we will use the labels s0 and t0 in anticipation of the isometry between G0φ and Gφ0 .
By the induction hypothesis, for each i ∈ [l], there exists a bijection ζi : V (Gφ0i ) →
F (Gφi ) such that for all u, v ∈ V (Gφ0i ) = V (Gφ0i ), ({u, v}, xj ) ∈ E(Gφ0i ) if and only if ζi (u)
and ζi (v) are faces separated by an edge with the label xj . We define ζ by ζ(i, v) = (i, ζi (v))
0
0
for all i ∈ [l] and v ∈ V (Gφ0i ) \ {s, t}, ζ(s0 ) = f s , and ζ(t0 ) = f t . By the induction
hypothesis, we can see that for any edge ({u, v}, xj ) ∈ E(Gφ0 ) \ ({s0 , t0 }, ∅), ζ(u) and ζ(v)
are separated by an edge labeled xj . This is because this edge is in one of the Gφ0i , and so
it has a dual edge in Gφi , by the induction hypothesis (see Figure 7). The only other edge
in Gφ0 is the edge ({s0 , t0 }, ∅), and ζ(s0 ) and ζ(t0 ) are exactly those faces on either side of
†
({s, t}, ∅) in Gφ , completing the proof that Gφ = Gφ0 .
If φ = φ1 ∨ · · · ∨ φl , then φ0 = φ01 ∧ · · · ∧ φ0l , and a nearly identical proof shows that
†
G φ = G φ0 .
Now that we have shown an isomorphism between G0φ and Gφ0 , note that G0φ (x) is the
subgraph of G0φ that includes all those edges where xe = 0. On the other hand Gφ0 (x) is
the graph that includes all those edges where xe = 1. Taking the bitwise negation of x, we
find that G0φ (x) = Gφ0 (x̄).
Lemma 35 allows us to prove Claim 17:
Claim 17. If φ = φ1 ∨ φ2 ∨ · · · ∨ φl , then G0φ (x) is formed by composing {G0φi (x)}i in series,
and if φ = φ1 ∧ φ2 ∧ · · · ∧ φl , then G0φ (x) is formed by composing {G0φi (x)}i in parallel.
Proof. If φ = φ1 ∨ · · · ∨ φl , then φ0 = φ01 ∧ · · · ∧ φ0l . From Lemma 35, G0φ (x) = Gφ0 (x̄), which
using Definition 34 is composed of {Gφ0i (x̄)}li=1 in series. But using the isomorphism of
Lemma 35 again, this is just {G0φi (x)}li=1 composed in series. The proof for φ = φ1 ∧ · · · ∧ φl
is similar.
Now we can prove Lemma 14, which relates the existence of a path in Gφ (x) or G0φ (x)
to the value of the function φ(x) :
Lemma 14. Let φ be any and-or formula on N variables. For every x ∈ {0, 1}N ,
there exists a path from s to t in Gφ (x) if and only if φ(x) = 1. Furthermore, for every
x ∈ {0, 1}N , there exists a path from s0 to t0 in G0φ (x) if and only if φ(x) = 0.
Proof. We will prove the statement by induction on the depth of φ. If φ = xj has depth 0,
then Gφ is just an edge ({s, t}, xj ), and G0φ is just an edge ({s0 , t0 }, xj ). Thus s and t are
connected in Gφ (x) if and only if xj = 1, in which case φ evaluates to 1, and s0 and t0 are
connected in Gφ0 if and only if xj = 0, in which case φ evaluates to 0.
If φ = φ1 ∧ · · · ∧ φl , then Gφ consists of Gφ1 , . . . , Gφl connected in series from s to t,
and moreover, Gφ (x) consists of Gφ1 (x), . . . , Gφl (x) connected in series from s to t. Thus
an st-path in Gφ (x) consists of an st-path in Gφ1 (x), followed by an st-path in Gφ2 (x), etc.,
up to an st-path in Gφl (x). Thus, s and t are connected in Gφ (x) if and only if s and t are
connected in each Gφ1 (x), . . . , Gφl (x), which happens if and only if φ1 (x) ∧ · · · ∧ φl (x) = 1.
On the other hand, by Claim 17, G0φ consists of G0φ1 , . . . , G0φl connected in parallel
between s0 and t0 . So any s0 t0 -path in G0φ (x) is an s0 t0 -path in one of the G0φi (x), which is
equivalent to an st-path in one of Gφ0i (x̄). Thus, by Lemma 35 s0 and t0 are connected in
Accepted in
Quantum 2017-08-12, click title to verify
30
G0φ (x) if and only if φ01 (x̄) ∨ · · · ∨ φ0l (x̄) = ¬φ1 (x) ∨ · · · ∨ ¬φl (x) = 1. By de Morgan’s law
is true if and only if φ(x) = φ1 (x) ∧ · · · ∧ φl (x) = 0.
The case when φ = φ1 ∨ · · · ∨ φl is similar.
C Classical Lower Bound on Class of Promise Boolean Formulas
In this section, we consider the query complexity of classical algorithms for and-or formulas, proving Theorem 19. To do this, we use a recent tool from Ben-David and Kothari [5].
They show that the bounded-error classical randomized query complexity of a function
f , denoted R(f ), satisfies R(f ) = Ω(RS(f )), where RS(f ) is the randomized sabotage
complexity, defined presently. Furthermore, they prove that for a composed function f ◦ g,
RS(f ◦ g) ≥ (RS(f )RS(g)).
If f : D → {0, 1}, with D ⊆ {0, 1}N , let fsab : Dsab → {0, 1}, where
Dsab = {x ∈ {0, 1, ∗}N ∪ {0, 1, †}N : x is consistent with y, y 0 ∈ D, s.t. f (y) 6= f (y 0 )}.
(66)
We say x ∈ {0, 1, ∗, †}N is consistent with y ∈ {0, 1}N if xi = yi for all i ∈ [N ] such
that xi ∈ {0, 1}. Then, fsab (x) = 1 if x ∈ {0, 1, ∗}N , and fsab (x) = 0 if x ∈ {0, 1, †}N .
Finally, the randomized sabotage complexity is given by RS(f ) = R0 (fsab ), where R0 (f ) is
the zero-error randomized query complexity of f . (For further classical query complexity
definitions, see [5].)
0
We first bound the sabotage complexity of and|DN,h and or|DN,h
:
0
Lemma 36. RS or|DN,h
= RS and|DN,h = Ω(N/h).
0
0
]sab to be consistent with y, y 0 ∈ DN,h
such that or(y) 6= or(y 0 ), we
Proof. For x ∈ [DN,h
N
N
must have that x ∈ {0, ∗} ∪ {0, †} . Furthermore, the number of ∗’s or †’s in x must be
at least h. Thus the sabotaged problem reduces to finding at least one marked item out
of n, promised there are at least h marked items. The randomized bounded-error query
complexity of this task is Ω(N/h), and so by Theorem 3 in [5],
0
0
0
RS or|DN,h
= R0 (or|DN,h
)sab = Ω R (or|DN,h
)sab
= Ω(N/h).
(67)
The proof for and is similar.
The next corollary follows immediately from Lemma 36 and the composition property of
sabotage complexity:
0
Corollary 37. Let φ = φ1 ◦ φ2 ◦ · · · ◦ φl , where for each i ∈ [l], φi = or|DN
or
,h
φi = and|DNi ,hi . Then R(φ) = Ω
Q
l
i=1 Ni /hi
i
i
.
Now that we understand the query complexity of symmetric composed and-or formulas, we can look at how this compares to the quantum query complexity of evaluating such
functions. We now prove the following lemma.
0
Lemma 38. Let φ = φ1 ◦ φ2 ◦ · · · ◦ φl , where for each i ∈ [l], φi = or|DN
or φi =
,h
i
i
and|DNi ,hi . Let D be the domain of φ. Then
i=1 Ni
Ql
i=1 hi
Accepted in
!
!
Ql
=
max
x∈D:φ(x)=1
Rs,t (Gφ (x))
Quantum 2017-08-12, click title to verify
max
x∈D:φ(x)=0
Rs,t (G0φ (x))
.
(68)
31
Proof. The proof follows by induction on the number of compositions. First suppose that
0
φ = or|DN,h
. Then Gφ consists of N edges connected in parallel between s and t, and G0φ
consists of N edges connected in series. The only input x such that φ(x) = 0 is the all
zeros input. Therefore maxx∈D:φ(x)=0 Rs,t (G0φ (x)) = N. Now notice (using Claim 5) that
Rs,t (Gφ (x)) = 1/|x|. However because of the domain of orNi ,hi , inputs x have |x| ≥ h, so
maxx∈D:φ(x)=1 Rs,t (Gφ (x)) = 1/h. Thus
!
N/h =
max
x∈D:φ(x)=1
Rs,t (Gφ (x))
!
max
x∈D:φ(x)=0
Rs,t (G0φ (x))
.
(69)
A similar analysis holds for the base case φ = and|DN,h .
Now for the inductive step, let φ = φ1 ◦ ξ for ξ = φ2 ◦ · · · ◦ φl , where for each i, φi is
0
either or|DN
or and|DNi ,hi . Let Dξ be the domain of ξ and let xj ∈ Dξ denote the bits
,h
i
i
0
of x that are input to the j th copy of ξ. Suppose first that φ1 = or|DN
1 ,h1
. G0φ is formed
by taking the N1 graphs G0ξ and connecting them in series. The only way φ(x) = 0 is if
the input xj ∈ Dξ to each of the ξ functions satisfies ξ(xj ) = 0, so by Claim 5
max
x∈D:φ(x)=0
Rs,t (G0φ (x)) = N1
max
y∈Dξ :ξ(y)=0
Rs,t (G0ξ (y)).
(70)
On the other hand, Gφ is formed by taking N1 graphs Gξ and connecting them in
parallel. Using Claim 5, if xj ∈ Dξ is the input to j th function ξ, we have
−1
N
X
1
Rs,t (Gφ (x)) =
R (Gξ (xj ))
j=1 s,t
.
(71)
Thus larger values for Rs,t (Gφ (x)) come from cases where Rs,t (Gξ (xj )) are large. Now
max Rs,t (Gξ (x)) = ∞,
(72)
x∈Dξ
which occurs when ξ(y) = 0. Because of the promise on the domain of φ1 , there must be
at least h1 of the N1 subformulas ξ that evaluate to 1. On each of those subformulas, we
want to have an input xj ∈ Dξ that maximizes the effective resistance of that subformula.
Therefore, we have
max
x∈D:φ(x)=1
Rs,t (Gφ (x)) =
!−1
h1
maxy∈Dξ :ξ(y)=1 Rs,t (Gξj (y)
=
maxy∈Dξ :ξ(y)=1 Rs,t (Gξj (y))
h1
.
(73)
Therefore, using the inductive assumption,
!
max
x∈D:φ(x)=1
N1
=
h1
max
Rs,t (Gφ (x))
y∈Dξ :ξ(y)=1
!
max
Rs,t (G0φ (x))
max
Rs,t (G0ξ (y))
x∈D:φ(x)=0
j
Rs,t (Gξj (y ))
y∈Dξ :ξ(y)=0
Ql
Ni
.
i=1 hi
= Qi=1
l
(74)
The inductive step for φ1 = and|DN,h is similar.
Corollary 37 and Lemma 38 give Theorem 19.
Accepted in
Quantum 2017-08-12, click title to verify
32
D NAND-tree Proofs
D.1 Relationship Between Faults and Effective Resistance
In this section, we prove Lemma 22:
d
Lemma 22. For any x ∈ {0, 1}2 , if d is even, then we have Rs,t (Gnandd (x)) ≤ FA (x)
and Rs0 ,t0 (G0nandd (x)) ≤ FB (x), while if d is odd, we have Rs,t (Gnandd (x)) ≤ 2FA (x) and
Rs0 ,t0 (G0nandd (x)) ≤ 2FB (x).
Proof. We will give a proof for FA (x); the case of FB (x) is similar.
First, Rs,t (Gnandd (x)) = ∞ if and only if s and t are not connected in Gnandd (x), which,
by Lemma 14, occurs if and only if x is a 0-instance. This means exactly that x is not
A-winnable, which, by Eq. (28), holds if and only if FA (x) = ∞. Thus, suppose this is not
the case, so FA (x) < ∞.
The rest of the proof is by induction. We need to look at both odd and even cases. For
0
the case of d = 0, the only A-winnable input in {0, 1}2 is x = 1. In that case, using Eq.
(28), FA (x) = 1, since there are no decision nodes for Player A, and since Gnand0 (x) is just
a single edge from s to t, Rs,t (Gnand0 (x)) = 1.
d
Let x ∈ {0, 1}2 be any A-winnable input with d > 1. We let x0 be the first 2d−1 bits
of x and x1 be the last 2d−1 bits of x, so x = (x0 , x1 ).
We first consider odd d > 1. Using the definition of Gφ from Section 4, and the fact
that for d odd, the root node is an ∧-node, we see that Gnandd (x) consists of Gnandd−1 (x0 )
and Gnandd−1 (x1 ) connected in series, so by Claim 5
Rs,t (Gnandd (x)) = Rs,t (Gnandd−1 (x0 )) + Rs,t (Gnandd−1 (x1 )).
(75)
Now the root can not be a fault, because it is a decision node for Player B, but we know
the tree is A-winnable, so no choice Player B makes would allow her to win the game.
Therefore, both subtrees connected to the root node must be A-winnable. Using Eq. (28)
we have
FA (x0 ) + FA (x1 ) ≤ max 2F A (xb ) = 2FA (x).
(76)
b∈{0,1}
Combining Eqs. (75) and (76) and the inductive assumption for even depth trees, we have
for odd d,
Rs,t (Gnandd (x)) ≤ 2FA (x).
(77)
Now we consider the case that d is even, so the root is a decision node for Player A.
Consequently, the root node is a ∨-node, so by Claim 5
Rs,t (Gnandd (x)) =
1
Rs,t (Gnandd−1 (x0 ))
+
1
Rs,t (Gnandd−1 (x1 ))
!−1
.
(78)
Suppose the root is a fault. Without loss of generality, let’s assume the subtree with
input x0 is not A-winnable. Then Rs,t (Gnandd−1 (x0 )) = ∞ so Eq. (78) becomes
Rs,t (Gnandd (x)) = Rs,t (Gnandd−1 (x1 )).
(79)
Using the inductive assumption for odd depth trees, Eq. (28), and the fact that the root is
a fault, we have
Rs,t (Gnandd (x)) ≤ 2F(x1 ) = F(x).
Accepted in
Quantum 2017-08-12, click title to verify
(80)
33
If the root is not a fault, then both Rs,t (Gnandd−1 (x0 )) and Rs,t (Gnandd−1 (x1 )) are finite,
so from (78), and using the inductive assumption, we have
1
max{Rs,t (Gnandd−1 (x0 )), Rs,t (Gnandd−1 (x1 ))}
2
≤ max{F(x0 ), F(x1 )} = F(x).
Rs,t (Gnandd (x)) ≤
(81)
A similar analysis for FB (x) completes the proof.
D.2 Estimating Effective Resistances
In this section, we will prove Lemma 25, which bounds the query complexity of estimating
the effective resistance of a graph corresponding to a Boolean formula. In [17], Ito and
Jeffery describe a quantum query algorithm to estimate the positive or negative witness
size of a span program given access to Ox . We will describe how to use this algorithm to
estimate the effective resistance of graphs Gφ (x) or G0φ (x).
Ref. [17] define the approximate positive and negative witness sizes, w̃+ (x, P ) and
w̃− (x, P ). These are similar to the positive and negative witness sizes, but with the
conditions |wi ∈ H(x) and ωAΠH(x) = 0 relaxed.
Definition 39 (Approximate Positive Witness). For any span program P on {0, 1}N and
x ∈ {0, 1}N , we define the positive error of x in P as:
e+ (x) = e+ (x, P ) := min
ΠH(x)⊥ |wi
2
: A|wi = τ .
(82)
We say |wi is an approximate positive witness for x in P if ΠH(x)⊥ |wi
A|wi = τ. We define the approximate positive witness size as
w̃+ (x) = w̃+ (x, P ) := min k|wik2 : A|wi = τ, ΠH(x)⊥ |wi
2
2
= e+ (x) and
= e+ (x) .
(83)
If x ∈ P1 , then e+ (x) = 0. In that case, an approximate positive witness for x is a positive
witness, and w̃+ (x) = w+ (x). For negative inputs, the positive error is larger than 0.
We can define a similar notion of approximate negative witnesses:
Definition 40 (Approximate Negative Witness). For any span program P on {0, 1}N and
x ∈ {0, 1}N , we define the negative error of x in P as:
e− (x) = e− (x, P ) := min
ωAΠH(x)
2
: ω ∈ L(U, R), ωτ = 1 .
(84)
2
Any ω such that ωAΠH(x) = e− (x, P ) is called an approximate negative witness for x
in P . We define the approximate negative witness size as
w̃− (x) = w̃− (x, P ) := min kωAk2 : ω ∈ L(U, R), ωτ = 1, ωAΠH(x)
2
= e− (x, P ) .
(85)
If x ∈ P0 , then e− (x) = 0. In that case, an approximate negative witness for x is a negative
witness, and w̃− (x) = w− (x). For positive inputs, the negative error is larger than 0.
Then Ito and Jeffery give the following result:
Accepted in
Quantum 2017-08-12, click title to verify
34
Theorem 41 ([17]). Fix X ⊆ {0, 1}N and f : X → R≥0 . Let P be a span program such
f− = W
f− (P, f ) = maxx∈X w̃− (x, P ).
that for all x ∈ X, f (x) = w+ (x, P ) and define W
There
a quantum
algorithm that estimates f to relative error ε and that uses
exists
q
1
ε3/2
e
O
f−
w+ (x)W
queries. Similarly, let P be a span program such that for all x ∈ X,
f+ = W
f+ (P, f ) = maxx∈X w̃+ (x, P ). Then there exists
f (x) = w− (x, P ) and define W
q
1
ε3/2
e
a quantum algorithm that estimates f to accuracy ε and that uses O
f+
w− (x)W
queries.
We will apply Theorem 41 to the span program PG,c defined in Eq. (6), with G = Gφ .
Throughout this section, we will always set the weight function c to take value one on all
edges of the graph G. In this, case, to simplify notation, we will denote the span program
f+ (PG ) and W
f− (PG ). We will
PG,c as PG . To apply Theorem 41, we need bounds on W
φ
φ
prove:
Lemma 42. For any formula φ, its ∧-depth is the largest number of ∧-labeled nodes on
any path from the root to a leaf. Let φ be any and-or formula with maximum fan-in l,
f+ (PG ) ≤ 1 ld∧ and W
f− (PG ) ≤ 2ld∨ .
∧-depth d∧ , and ∨-depth d∨ . Then W
φ
φ
2
Then, applying Lemma 42 and Theorem 41, we have the main result of this section,
which was first stated in Section 5.2:
Lemma 25 (Est Algorithm). Let φ be an and-or formula with constant fan-in l, ∨-depth
d∨ and ∧-depth d∧ . Then the quantum query
complexity of estimating Rs,t (Gφ (x)) (resp.
q
e
Rs,t (G0φ (x))) to relative accuracy is O
1
ε3/2
e
Rs,t (Gφ (x))ld∨ (resp. O
1
ε3/2
q
Proof of Lemma 25. By Theorem 41, since Rs,t (Gφ (x)) = 12 w+ (x, PGφ ) (Lemma 11), we
f− (PG ). By
can estimate this quantity using a number of queries that depends on W
φ
d
f− (PG ) ≤ 2l ∨ , so we can estimate w+ (x) = Rs,t (Gφ (x))
Lemma 42, we have that W
φ
e
in O
1
p
ε2/3
1/2
f
w+ (x)W
−
e
= O
1
ε2/3
q
Rs,t (Gφ (x))ld∨
f+ ≤
2w− (x, PGφ ) for all 0-instances, and W
e
O
1
ε2/3
q
1 d∧
2l ,
queries. Similarly, Rs,t (G0φ (x)) =
so we can estimate Rs,t (G0φ (x)) in
Rs,t (G0φ (x))ld∧ queries.
To prove Lemma 42, we will use the following observation, which gives an upper bound
on the length of the longest self-avoiding st-path in Gφ , in terms of the ∧-depth of φ. This
bound is not tight in general.
Claim 43. Let φ be an and-or formula with constant fan-in l. If φ has ∧-depth d∧ , then
the longest self-avoiding path connecting s and t in Gφ has length at most ld∧ .
Proof. We will prove the statement by induction. If φ has ∧-depth d∧ = 0, then it has no
∧-nodes. Thus, it is easy to see that Gφ has only two vertices, s and t, with some number
of edges connecting them, so every st-path has length 1.
Suppose φ has ∧-depth d∧ > 0. First, suppose φ = φ1 ∧ · · · ∧ φl . Then since Gφ
consists of Gφ1 , . . . , Gφl connected in series, any st-path in Gφ consists of an st-path in
Gφ1 , followed by an st-path in Gφ2 , etc. up to an st-path in Gφl , so if d∧ (φi ) is the ∧-depth
of φi , then the longest st-path in Gφ has length at most:
ld∧ (φ1 ) + · · · + ld∧ (φl ) ≤ lld∧ −1 = ld∧ .
Accepted in
Quantum 2017-08-12, click title to verify
Rs,t (G0φ (x))ld∧ ).
(86)
35
If φ = φ1 ∨ · · · ∨ φl , then maxi d∧ (φi ) = d∧ (φ) = d∧ , and Gφ consists of Gφ1 , . . . , Gφl ,
connected in parallel. Any self-avoiding st-path must include exactly one edge adjacent to
s and one edge adjacent to t. However, any path that includes an edge from Gφi and Gφj
for i 6= j must go through s or t, so it must have more than one edge adjacent to s, or more
than one edge adjacent to t, so such a path can never be a self-avoiding st-path. Thus, any
self-avoiding st-path must be contained completely in one of the Gφi . The longest such
path is thus the longest self-avoiding st-path in any of the Gφi , which, by induction, is
maxi ld∧ (φi ) = ld∧ .
Now we can prove Lemma 42:
f+ . Suppose |w̃i is
Proof of Lemma 42. To begin, we will prove the upper bound on W
an optimal approximate positive witness for x. By Claim 30, if |w̃i is an approximate
positive witness, then since A|w̃i = τ , and c has unit value on all edges of G, θ(u, v, λ) =
hu, v, λ|w̃i−hv, u, λ|w̃i is a unit flow on G. Since |w̃i is an approximate positive witness for x,
2
it has minimal error for x, so it minimizes ΠH(x)⊥ |w̃i , and since it is optimal, it minimizes
k|w̃ik2 over all approximate positive witnesses. Define |θi =
P
−
→
θ(u, v, λ)|u, v, λi,
(u,v,λ)∈ E (G)
so we know that 12 |θi also maps to τ under A, so is also a positive witness in PGφ .Then we
have
ΠH(x)⊥ |θi
2
=2
X
hu, v, λ|w̃i2 −2
(u,v,λ)∈
−
→
−
→
E (G)\ E (G(x))
X
hu, v, λ|w̃ihv, u, λ|w̃i ≤ 2ΠH(x)⊥ |w̃i
2
,
(u,v,λ)∈
−
→
−
→
E (G)\ E (G(x))
(87)
where the last inequality uses Cauchy-Schwarz, so 12 |θi is also an approximate positive
witness for x. Similarly,
k|θik2 ≤ k2|w̃ik2 ,
(88)
so 21 |θi is optimal.
By Claim 29, we can consider a decomposition of |θi into self-avoiding paths pi and
P
P 0
cycles ci such that all cycles are disjoint from all paths, |θi = ri=1 αi |pi i + ri=1 βi |ci i,
where for each i,
|pi i =
Li
X
(i)
(i)
|uj , uj+1 , λi,j i −
j=1
Li
X
Li
X
(89)
0
(i) (i)
|vj , vj+1 , λ0i,j i
j=1
(i)
(i)
j=1
0
|ci i =
(i)
|uj+1 , uj , λi,j i,
−
Li
X
(i)
(i)
|vj+1 , vj , λ0i,j i
(90)
j=1
(i)
where vL0 +1 = v1 and {λi,j }i,j ∩ {λ0i,j }i,j = ∅. It’s easy to see (in the case of unit edge
j
weights) that A|ci i = 0 for all i, so
A
Let |θ0 i =
Pr
i=1 αi |pi i.
r
1X
1
αi |pi i = A |θi = τ.
2 i=1
2
(91)
Then since ci and pj have no common edges, we have hci |pj i = 0,
and also hci |(I − ΠH(x) )|pj i = 0, so the error of 12 |θ0 i is
1
4
ΠH(x)⊥ |θ0 i
2
≤
1
4
2
ΠH(x)⊥ |θi ,
so 12 |θ0 i also has minimal error. Furthermore, k|θ0 ik2 ≤ k|θik2 , with equality if and only
if there are no cycles in the decomposition. By the optimality of 12 |θi as an approximate
Accepted in
Quantum 2017-08-12, click title to verify
36
positive witness for x, we can conclude that |θi =
P
i, and A|θi = 2τ , we have ri=1 αi = 1. Then
Pr
i=1 αi |pi i,
and since A|pi i = 2τ for all
k|θik2 ≤ max k|pi ik2 = max 2Li .
i
i
(92)
Since the longest self-avoiding st-path in Gφ has length at most ld∧ , and each Li is the length
f+ (PG ) ≤ 1 ld∧ .
of a self-avoiding path in Gφ , we have w̃+ (x, PGφ ) ≤ 41 2ld∧ = 12 ld∧ . Thus W
φ
2
f− . A min-error approximate negative witness for x in
Next we prove the bound on W
2
PGφ is a function ω : V (Gφ ) → R such that ωτ = ω(s) − ω(t) = 1, and ωAΠH(x) =
P
−
→
(ω(u) − ω(v))2 is minimized. By Claim 31, since ωτ = 1, the function
(u,v,λ)∈ E (Gφ (x))
→
−
θ : E (G0φ ) → R defined by θ((u, v, λ)† ) = ω(u) − ω(v) is a unit s0 t0 -flow on G0φ = Gφ0 , and
the witness complexity is
kωAk2 =
X
(ω(u) − ω(v))2 =
−
→
(u,v,λ)∈ E (Gφ )
X
θ(u0 , v 0 , λ)2 = k|θik2
(93)
−
→
(u0 ,v 0 ,λ)∈ E (G0φ )
where we create |θi from θ in the usual way. By an argument similar to the previous
argument, if ω is an optimal approximate negative witness for x, then k|θik2 is upper
bounded by twice the length of the longest self-avoiding s0 t0 -path in G0φ = Gφ0 . By Lemma
0
35 and Claim 43, this is upper bounded by 2ld∧ (φ ) = 2ld∨ (φ) , where d∧ (φ0 ) is the ∧-depth
f− ≤ 2ld∨ .
of φ0 , and d∨ = d∨ (φ) is the ∨-depth of φ. Thus w̃− (x, PGφ ) ≤ 2ld∨ , and so W
D.3 Winning the NAND-tree
We now analyze the algorithm for winning the game associated with a nand-tree, proving
Lemma 26 and Theorem 27.
d
Lemma 26. Let x0 , x1 ∈ {0, 1}2 be instances of nandd with at least one of them a
0
1
1-instance. Let N = 2d , and wmin
s,t
= min{R
(Gnandd (x )), Rs,t (Gnandd (x ))}. Then
√
e N 1/4 wmin queries to (x0 , x1 ) and outputs b such
Select(x0 , x1 ) terminates after O
that Rs,t (Gnandd (xb )) ≤ 2Rs,t (Gnandd (xb̄ )) with bounded error.
Proof. Since at least one of x0 and x1 is a 1-instance, using the description of Select in
Section 5.2, at least one of the programs will terminate. Suppose without loss of generality
that Est(x0 ) is the first to terminate, outputting w0 . Then there are two possibilities:
√
Est(x1 ) does not terminate after p(d) w0 N 1/4 steps, in which case, Rs,t (Gnandd (x0 )) ≤
√
2Rs,t (Gnandd (x1 )), and Select outputs 0; or Est(x1 ) outputs w1 before p(d) w0 N 1/4 steps
have passed and Select outputs b such that wb ≤ wb̄ .
We will prove the first case by contradiction. Suppose
2Rs,t (Gnandd (x1 )) < Rs,t (Gnandd (x0 )).
(94)
Then Est(x1 ) must terminate after
q
q
1
p(d) Rs,t (Gnandd (x1 ))N 1/4 ≤ √ p(d) Rs,t (Gnandd (x0 ))N 1/4
2
(95)
steps. In Select, we run Est to relative accuracy ε = 1/3, so we have
1
|w0 − Rs,t (Gnandd (x0 ))| ≤ Rs,t (Gnandd (x0 )),
3
Accepted in
Quantum 2017-08-12, click title to verify
(96)
37
and so
2
w0 ≥ Rs,t (Gnandd (x0 )).
3
(97)
q
Plugging Eq. (97) into Eq. (95), we have Est(x1 ) must terminate after √12 p(d) 32 w0 N 1/4 <
√
p(d) w0 N 1/4 steps, which is a contradiction.
Thus, Rs,t (Gnandd (x0 )) ≤ 2Rs,t (Gnandd (x1 )),
qso outputting 0 is correct. Furthermore,
√
1/4
e
since we terminate after p(d) w0 N
= O( Rs,t (Gnandd (x0 ))N 1/4 ) steps, and since
e N 1/4 √wmin .
Rs,t (Gnandd (x0 )) = O(Rs,t (Gnandd (x1 ))), the running time is at most O
We now consider the second case, in which both programs output estimates w0 and w1 ,
such that |wb − Rs,t (Gnandd (xb ))| ≤ εRs,t (Gnandd (xb )) for b = 0, 1. Suppose wb ≤ wb̄ . We
then have
Rs,t (Gnandd (xb ))
1+ε
Rs,t (Gnandd (xb ))
wb̄
4/3
≤
≤
=
= 2.
wb
1−ε
2/3
Rs,t (Gnandd (xb̄ ))
Rs,t (Gnandd (xb̄ ))
(98)
Thus Rs,t (Gnandd (xb )) ≤ 2Rs,t (Gnandd (xb̄ )), as required. Furthermore, the running time of
the algorithm is bounded by the running timeq
of Est(x1 ), the second to terminate. We
e
know that Est(x1 ) has running time at most O
Rs,t (Gnandd (x1 ))N 1/4 steps, and by asq
√
e
sumption, Est(x1 ) terminated after less than p(d) w0 N 1/4 = O
Rs,t (Gnandd (x0 ))N 1/4
e N 1/4 √wmin .
steps, so the total running time is at most O
Theorem 27. Let x ∈ {0, 1}N for N = 2d be an A-winnable input to nandd . At every
node v where Player A makes a decision, let Player A use the Select algorithm in the
following way. Let v0 and v1 be the two children of v, with inputs to the respective subtrees
of v0 and v1 given by x0 and x1 respectively. Then Player A moves to vb where b is the
outcome that occurs a majority of times when Select(x0 , x1 ) is run O(log d) times. Then
if Player B, at his decision nodes, chooses left and right with equal
Player A
q
probability,
1/4
e
will win the game with probability at least 2/3, and will use O N
Rs,t (Gnandd (x))
queries on average, where the average is taken over the randomness of Player B’s choices.
Proof. First note that Player A must make O(d) choices over the course of the game.
We amplify Player A’s probability of success by repeating Select at each decision node
O(log d) times and taking the majority. Then the probability that Player A chooses the
wrong direction at any node is O(1/d), and we ensure that her probability of choosing the
wrong direction over the course of the algorithm is < 1/3. From here on, we analyze the
error free case.
Let p(d) be a non-decreasing polynomial function
in d such that Select, on inputs
q
d
0
1
2
d/4
x , x ∈ {0, 1} , terminates in at most p(d)2
min{Rs,t (Gnandd (x0 )), Rs,t (Gnandd (x1 ))}
queries. Then we will prove that for trees of odd depth d, the expected
number of queries
q
d/4+5
by Player A over the course of the game is at most p(d)2
Rs,t (Gnandd (x)), while
q
for even depth trees, it is at most p(d)2d/4+11/2 Rs,t (Gnandd (x)), thus proving the main
result.
We prove the result by induction on the depth of the tree. For depth zero trees, there
are no decisions, N = Rs,t Gnand0 (x) = 1, so the result holds.
For the inductive case, we treat odd and even depth cases separately. First consider
an instance of nandd with d > 0, d odd. Thus nandd (x) = nandd−1 (x0 ) ∧ nandd−1 (x1 ),
Accepted in
Quantum 2017-08-12, click title to verify
38
where x = (x0 , x1 ). Because the root is at odd distance from the leaves, the root is a
decision node for Player B. Because we are in an A-winnable tree, no matter which choice
Player B makes, we will end up at an A-winnable subtree of depth d − 1, so the inductive
assumption holds for those trees. That is, the expected number of queries for Player A
must make to win the subtree with input xb (for b ∈ {0, 1}) averaged over Player B’s
choices is at most
q
p(d − 1)2(d−1)/4+11/2 Rs,t (Gnandd−1 (xb )).
(99)
We are assuming that Player B chooses left and right with equal probability. Thus, the
expected number of queries that Player A must make over Player B’s choices throughout
the game is at most
q
1
p(d − 1)2(d−1)/4+11/2 Rs,t (Gnandd−1 (x0 ))
2
q
+p(d − 1)2(d−1)/4+11/2 Rs,t (Gnandd−1 (x1 ))
r
1
Rs,t (Gnandd−1 (x0 )) + Rs,t (Gnandd−1 (x1 ))
r2
1
= p(d − 1)2(d−1)/4+11/2
Rs,t (Gnandd (x)) by Claim 5,
q2
≤ p(d − 1)2
(d−1)/4+11/2
by Jensen’s ineq.,
≤ p (d) 2d/4−1/4+11/2−1/2 Rs,t (Gnandd (x))
q
≤ p (d) 2d/4+5 Rs,t (Gnandd (x)),
(100)
proving the case for odd d.
Now consider an instance of nandd with d > 0, d even. Thus nandd (x) = nandd−1 (x0 )∨
nandd−1 (x1 ), where x = (x0 , x1 ). Because the root is at even distance from the leaves,
the root is a decision node for Player A. Player A runs Select(x0 , x1 ), which returns
b ∈ {0, 1} such that (by Lemma 26)
Rs,t (Gnandd−1 (xb )) ≤ 2Rs,t (Gnandd−1 (xb̄ )),
(101)
which requires at most
q
min p (d − 1) 2(d−1)/4 Rs,t (Gnandd−1 (xb∗ ))
∗
b ∈{0,1}
(102)
queries.
After making the choice to move to the subtree with input xb , by the inductive
assumption, the expected number of queries that Player A need to make throughout the
rest of the game (averaged over Player B’s choices) is
q
p (d − 1) 2d/4+5 Rs,t (Gnandd−1 (xb )).
(103)
There are two cases to consider. If Rs,t (Gnandd−1 (xb )) ≤ Rs,t (Gnandd−1 (xb̄ )), then
combining Eq. (102) and Eq. (103), we have that the total number of queries averaged
Accepted in
Quantum 2017-08-12, click title to verify
39
over Player B’s choices is
q
q
p (d − 1) 2(d−1)/4 Rs,t (Gnandd−1 (xb )) + p (d − 1) 2(d−1)/4+5 Rs,t (Gnandd−1 (xb ))
q
≤p (d − 1) 2(d−1)/4 Rs,t (Gnandd−1 (xb ))(1 + 25 )
q
≤p (d − 1) 2(d−1)/4+5+1/16 Rs,t (Gnandd−1 (xb ))
q
≤p (d − 1) 2(d−1)/4+5+1/16+1/2 Rs,t (Gnandd (x))
q
≤p (d) 2d/4+11/2 Rs,t (Gnandd (x))
where we’ve used Rs,t (Gnandd (x)) =
(104)
−1
Rs,t (Gnandd−1 (x0 ))
+ Rs,t (Gnandd−1 (x1 ))
−1 −1
from Claim 5 and the fact that Rs,t (Gnandd−1 (xb )) ≤ Rs,t (Gnandd−1 (xb̄ )) to bound the
value Rs,t (Gnandd−1 (xb )) by 2Rs,t (Gnandd (x)). This proves the even induction step for this
case.
The other case is if Rs,t (Gnandd−1 (xb )) > Rs,t (Gnandd−1 (xb̄ )). In that case, again
−1
using the fact that Rs,t (Gnandd (x)) = Rs,t (Gnandd−1 (x0 ))
we have
−1 −1
+ Rs,t (Gnandd−1 (x1 ))
Rs,t (Gnandd−1 (xb̄ ))
Rs,t (Gnandd−1 (x )) = Rs,t (Gnandd (x)) 1 +
Rs,t (Gnandd−1 (xb ))
,
!−1
2
≤ Rs,t (Gnandd (x)),
3
(105)
where the inequality follows from Eq. (101). Thus, the average total number of queries is
b̄
q
q
p (d − 1) 2(d−1)/4 Rs,t (Gnandd−1 (xb̄ )) + p (d − 1) 2(d−1)/4+5 Rs,t (Gnandd−1 (xb ))
≤p(d − 1)2(d−1)/4
≤p(d − 1)2
(d−1)/4
q
q
Rs,t (Gnandd−1 (xb̄ )) + 25 2Rs,t (Gnandd−1 (xb̄ ))
(1 + 2
5+1/2
r
)
2
Rs,t (Gnandd (x))
3
q
≤p(d)2d/4−1/4+5 Rs,t (Gnandd (x))
q
≤p(d)2d/4+5 Rs,t (Gnandd (x)).
(106)
This proves the induction step for the other case.
Accepted in
Quantum 2017-08-12, click title to verify
40
| 8cs.DS
|
arXiv:1510.03367v1 [cs.DS] 12 Oct 2015
Layered Heaps Beating Standard and Fibonacci
Heaps in Practice
P. Huggins
March 28, 2018
Abstract
We consider the classic problem of designing heaps. Standard binary
heaps run faster in practice than Fibonacci heaps but have worse time
guarantees. Here we present a new type of heap that runs faster in practice
than both standard binary and Fibonacci heaps, but has asymptotic insert
times arbitrarily better than O(log n), namely O((log n)1/m ) for arbitrary
positive integer m. Our heap is defined recursively and maximum run
time speed up occurs when a recursion depth of 1 is used, i.e. a heap of
heaps.
1
Layered Heaps
We will define M -layered heaps for arbitrary integer M ≥ 1. For M =
1 the M -ary heap is a standard binary heap stored in an array. For
kM −1 ∗(M −1)/M
M ≥ 2, a kM -ary heap is used, with kM = 2(log n)
. Then
by inductive hypothesis, insert operations on the children heaps will be
(log n)(M −1/M )∗(1/(M −1)) = (log n)1/M . Furthermore the height of the M ary heap will also be (log n)1/M . So insert operations take O((log n)1/M )
time for any M we want. Pop/delete functions take standard O(log n)
time, because we may need to do a children heap operation which takes
(log n)(M −1)/M time a total of (log n)1/M times, i.e. the height of the
M -ary layered heap.
The operations and running times for them are explained in the following pseudocode:
INSERT INTO M-ARY HEAP A: Before swapping, start by
placing the element at end of array (position n = N ). Then do the
following:
(M −1)/M
• Set k = 2(log N )
• While n > 0:
– If A[n] > A[n/k] then swap their values
– Else insert A[n] into children (M − 1)-ary layered heap that
contains position n in the array. (Recursive) Then BREAK.
1
– Set n := n/k
POP OUT OF M-ARY HEAP A:
(M −1)/M
• Set k = 2(log N )
• Store and remove root element of M -ary heap.
• Put the last item in the heap at the root.
• Swap downwards with top of children heap while top of children
heap is greater than element. (Recursively balance the (M − 1)ary
heap in time O((log n)(M −1)/M )).
• Break when element is greater than top of current children heap.
• Return popped top of heap
As can be seen, the running time for insert is O((((log n)(M −1)/M ))1/(M −1) )
which is O((log n)1/M . The running time for pop is O((log n)(M −1)/M +1/M =
O(log n).
2
Popular Competing Heaps
In [1] the Fibonacci heap is presented, which has (amortized) constant insert time, and standard O(log n) delete/pop time. The amortized running
times were later improved to strict running time bounds per operation in
a later publication. However, in practice, the constants associated with
various Fibonacci heaps are too large to outperform a standard binary
tree. Thus, due to its simplicity and faster running time, binary heaps
are traditionally what is taught and used.
3
Running time comparisons for insert/pop
To simulate situations where asymptotically faster insertions in heaps may
be better than traditional heaps, we did a 10 to 1 simulation where 1000
elements would be added and then 100 elements would be popped, where
the ith insert inserted the value i (and the heap is a max-heap), and
repeated over and over with running times being recorded as a function
of the size of the heap. Binary heaps are faster than Fibonacci heaps for
practical data sizes in practice. Furthermore, analysis of our recursively
defined M -ary layered heaps made it clear that the constants become too
large to be overcome in practice unless M = 2. Thus we compared the
2-layer heap to the traditional binary heap. Results are shown in the
figure, where N is the number of elements in the heap as the heap grows.
Results were computed out to N = 232 and then extrapolated to N = 240
to cover all feasible data sizes.
As the figure shows, despite the O(log n) time for both heaps when
a pop is performed, the 2-ary layered heap has good cache performance
when processing a children heap because it only has about 50 elements
and so usually they all fit into cache after one memory access into the
2
children heap. Thus the running time in practice
√ for inserts into the 2ary layered heap look more like an inflated O( log n). Memory use is
identical for both heaps.
4
Discussion
Although M -ary layered heaps are interesting from a theoretical point of
view for arbitrary M , giving asymptotic insert running time arbitrarily
closer and closer to constant, in practice the 2-ary layered heap is the
fastest in practice and can run up to 3-4 times faster than a binary heap
for reasonable data sizes.
In fact, 2-ary heaps are easy enough to describe and implement and
analyze directly (as opposed to using induction/recursion for M > 2),
they should probably be taught in data structures courses after standard
binary heaps are presented.
5
Bibliography
1. Fredman, Michael Lawrence and Tarjan, Robert E. (1987). ”Fibonacci heaps and their uses in improved network optimization algorithms”. Journal of the Association for Computing Machinery 34
(3): 596615. doi:10.1145/28869.28874.
3
| 8cs.DS
|
arXiv:1510.08292v1 [math.AC] 28 Oct 2015
THE STRUCTURE OF THE SALLY MODULE OF
INTEGRALLY CLOSED IDEALS
KAZUHO OZEKI AND MARIA EVELINA ROSSI
To Shiro Goto on the occasion of his seventieth birthday
Abstract. The first two Hilbert coefficients of a primary ideal play an important role
in commutative algebra and in algebraic geometry. In this paper we give a complete algebraic structure of the Sally module of integrally closed ideals I in a Cohen-Macaulay
local ring A satisfying the equality e1 (I) = e0 (I)−ℓA (A/I)+ℓA (I 2 /QI)+1, where Q is
a minimal reduction of I, and e0 (I) and e1 (I) denote the first two Hilbert coefficients
of I, respectively the multiplicity and the Chern number of I. This almost extremal
value of e1 (I) with respect classical inequalities holds a complete description of the
homological and the numerical invariants of the associated graded ring. Examples are
given.
1. Introduction and notation
Throughout this paper, let A denote a Cohen-Macaulay local ring with maximal ideal
m and positive dimension d. Let I be an m-primary ideal in A and, for simplicity, we
assume the residue class field A/m is infinite. Let ℓA (N) denote, for an A-module N,
the length of N. The integers {ei (I)}0≤i≤d such that the equality
n+d−1
n+d
n+1
+ · · · + (−1)d ed (I)
− e1 (I)
ℓA (A/I ) = e0 (I)
d−1
d
holds true for all integers n ≫ 0, are called the Hilbert coefficients of A with respect
to I. This polynomial, known as the Hilbert-Samuel polynomial of I and denoted by
HPI (n), encodes the asymptotic information coming from the Hilbert function HI (t)
of I which is defined as
HI (t) = ℓA (I t /I t+1 ).
The generating function of the numerical function HI (t) is the power series
X
HSI (z) =
HI (t)z t .
t≥0
This series is called the Hilbert series of I. It is well known that this series is rational
and that, even more, there exists a polynomial hI (z) with integers coefficients such that
hI (1) 6= 0 and
hI (z)
.
HSI (z) =
(1 − z)d
Key words and phrases: Cohen-Macaulay local ring, associated graded ring, Hilbert function, Hilbert
coefficient
2010 Mathematics Subject Classification: 13D40, 13A30, 13H10.
1
Notice that for all i ≥ 0, the Hilbert coefficients can be computed as it follows
(i)
h (1)
ei (I) := I
i!
(i)
where hI denotes the i−th derivative of hI (z) evaluated at i (h(0) = hI ).
Choose a parameter ideal Q of A which forms a reduction of I and let
R = R(I) := A[It] and T = R(Q) := A[Qt] ⊆ A[t]
denote, respectively, the Rees algebras of I and Q. Let
R′ = R′ (I) := A[It, t−1 ] ⊆ A[t, t−1 ]
and G = G(I) := R′ /t−1 R′ ∼
=
Following Vasconcelos [V], we consider
S = SQ (I) = IR/IT ∼
=
the Sally module of I with respect to Q.
M
M
I n /I n+1 .
n≥0
I n+1 /Qn I
n≥1
The notion of filtration of the Sally module was introduced by M. Vaz Pinto [VP]
as follows. We denote by E(α), for a graded T -module E and each α ∈ Z, the graded
T -module whose grading is given by [E(α)]n = Eα+n for all n ∈ Z.
Definition 1.1. ([VP]) We set, for each i ≥ 1,
C (i) = (I i R/I i T )(−i + 1) ∼
=
M
I n+1 /Qn−i+1 I i .
n≥i
L
and let L(i) = T [C (i) ]i . Then, because L(i) ∼
= n≥i Qn−i I i+1 /Qn−i+1 I i and C (i) /L(i) ∼
=
(i+1)
C
as graded T -modules, we have the following natural exact sequences of graded
T -modules
0 → L(i) → C (i) → C (i+1) → 0
for every i ≥ 1.
We notice that C (1) = S, and C (i) are finitely generated graded T -modules for all
i ≥ 1, since R is a module-finite extension of the graded ring T .
So, from now on, we set
C = CQ (I) = C (2) = (I 2 R/I 2 T )(−1)
and we shall explore the structure of C. Assume that I is integrally closed. Then, by
[EV, GR], the inequality
e1 (I) ≥ e0 (I) − ℓA (A/I) + ℓA (I 2 /QI)
holds true and the equality e1 (I) = e0 (I)−ℓA (A/I)+ℓA (I 2 /QI) holds if and only if I 3 =
QI 2 . When this is the case, the associated graded ring G of I is Cohen-Macaulay and the
behavior of the Hilbert-Samuel function ℓA (A/I n+1 ) of I is known (see [EV], Corollary
2.10). Thus the integrally closed ideal I with e1 (I) = e0 (I)−ℓA (A/I)+ℓA (I 2 /QI) enjoys
nice properties and it seems natural to ask what happens on the integrally closed ideal
I which satisfies the equality e1 (I) = e0 (I) − ℓA (A/I) + ℓA (I 2 /QI) + 1. The problem is
not trivial even if we consider d = 1.
2
We notice here that ℓA (I 2 /QI) = e0 (I) + (d − 1)ℓA (A/I) − ℓA (I/I 2 ) holds true (see
for instance [RV3]), so that ℓA (I 2 /QI) does not depend on a minimal reduction Q of I.
Let B = T /mT ∼
= (A/m)[X1 , X2 , · · · , Xd ] which is a polynomial ring with d indeterminates over the field A/m. The main result of this paper is stated as follows.
Theorem 1.2. Assume that I is integrally closed. Then the following conditions are
equivalent:
(1) e1 (I) = e0 (I) − ℓA (A/I) + ℓA (I 2 /QI) + 1,
(2) mC = (0) and rankB C = 1,
(3) C ∼
= (X1 , X2 , · · · , Xc )B(−1) as graded T -modules for some 1 ≤ c ≤ d, where
X1 , X2 , · · · , Xd are linearly independent linear forms of the polynomial ring B.
When this is the case, c = ℓA (I 3 /QI 2 ) and I 4 = QI 3 , and the following assertions hold
true:
(i) depth G ≥ d − c and depthT C = d − c + 1,
(ii) depth G = d − c, if c ≥ 2.
(iii) Suppose c = 1 < d. Then HPI (n) = ℓA (A/I n+1 ) for all n ≥ 0 and
e1 (I) − e0 (I) + ℓA (A/I) + 1
1
ei (I) =
0
if i = 2,
if i = 3 and d ≥ 3,
if 4 ≤ i ≤ d.
(iv) Suppose 2 ≤ c < d. Then HPI (n) = ℓA (A/I n+1) for all n ≥ 0 and
e1 (I) − e0 (I) + ℓA (A/I)
0
ei (I) =
(−1)c+1
if i = 2,
if i =
6 c + 1, c + 2, 3 ≤ i ≤ d
if i = c + 1, c + 2, 3 ≤ i ≤ d
(v) Suppose c = d. Then HPI (n) = ℓA (A/I n+1 ) for all n ≥ 2 and
ei (I) =
e1 (I) − e0 (I) + ℓA (A/I)
0
if i = 2 and d ≥ 2,
if 3 ≤ i ≤ d
(vi) The Hilbert series HSI (z) is given by
HSI (z) =
ℓA (A/I) + {e0 (I) − ℓA (A/I) − ℓA (I 2 /QI) − 1}z + {ℓA (I 2 /QI) + 1}z 2 + (1 − z)c+1 z
.
(1 − z)d
Let us briefly explain how this paper is organized. We shall prove Theorem 1.2 in
Section 3. In Section 2 we will introduce some auxiliary results on the structure of the
T -module C = CQ (I) = (I 2 R/I 2 T )(−1), some of them are stated in a general setting.
Our hope is that these information will be successfully applied to give new insights
in problems related to the structure of Sally’s module. In Section 4 we will introduce
some consequences of Theorem 1.2. In particular we shall explore the integrally closed
ideals I with e1 (I) ≤ e0 (I) − ℓA (A/I) + 3. In Section 5 we will construct a class of
Cohen-Macaulay local rings satisfying condition (1) in Theorem 1.2.
3
2. Preliminary Steps
The purpose of this section is to summarize some results on the structure of the
graded T -module C = CQ (I) = (I 2 R/I 2 T )(−1), which we need throughout this paper.
Remark that in this section I is an m-primary ideal not necessarily integrally closed.
Lemma 2.1. The following assertions hold true.
(1) mℓ C = (0) for integers ℓ ≫ 0; hence dimT C ≤ d.
(2) The homogeneous components {Cn }n∈Z of the graded T -module C are given by
(0)
if n ≤ 1,
Cn ∼
=
I n+1 /Qn−1 I 2
if n ≥ 2.
(3) C = (0) if and only if I 3 = QI 2 .
(4) mC = (0) if and only if mI n+1 ⊆ Qn−1 I 2 for all n ≥ 2.
(5) S = T C2 if and only if I 4 = QI 3 .
Proof. (1) Let u = t−1 and T ′ = R′ (Q). Notice that C = (I 2 R/I 2 T )(−1) ∼
=
(I 2 R′ /I 2 T ′ )(−1) as graded T -modules. We then have uℓ ·(I 2 R′ /I 2 T ′ ) = (0) for some
ℓ ≫ 0, because the graded T ′ -module I 2 R′ /I 2 T ′ is finitely generated and [I 2 R′ /I 2 T ′ ]n =
(0) for all√n ≤ 0. Therefore, mℓ C = (0) for ℓ ≫ 0, because Qℓ = (Qtℓ )uℓ ⊆ uℓ T ′ ∩ A
and m = Q.
(2) Since [I 2 R]n = (I n+2 )tn and [I 2 T ]n = (I 2 Qn )tn for all n ≥ 0, assertion (2) follows
from the definition of the module C = (I 2 R/I 2 T )(−1).
Assertions (3), (4), and (5) are readily follow from assertion (2).
In the following result we need that Q ∩ I 2 = QI holds true. This condition is
automatically satisfied in the case where I is integrally closed (see [H, I2]).
Proposition 2.2. Suppose that Q ∩ I 2 = QI. Then we have AssT C ⊆ {mT } so that
dimT C = d, if C 6= (0).
Let Q = (a1 , . . . , ad ) be a minimal reduction of I. In the proof of Proposition 2.2 we
need the following lemmata.
Lemma 2.3. Suppose that Q ∩ I 2 = QI. Then we have (a1 , a2 , · · · , ai ) ∩ Qn+1 I 2 =
(a1 , a2 , · · · , ai )Qn I 2 for all n ≥ 0 and 1 ≤ i ≤ d. Therefore, a1 t ∈ T is a nonzero
divisor on T /I 2 T .
Proof. We have only to show that (a1 , a2 , · · · , ai ) ∩ Qn+1 I 2 ⊆ (a1 , a2 , · · · , ai )Qn I 2 holds
true for all n ≥ 0. We proceed by induction on n and i. We may assume that i < d and
that our assertion holds true for i+1. Suppose n = 0 and take x ∈ (a1 , a2 , · · · , ai )∩QI 2 .
Then, by the hypothesis of induction on i, we have
(a1 , a2 , · · · , ai ) ∩ QI 2 ⊆ (a1 , a2 , · · · , ai , ai+1 ) ∩ QI 2 = (a1 , a2 , · · · , ai , ai+1 )I 2
Then we write x = y + ai+1 z with y ∈ (a1 , a2 , · · · , ai )I 2 and z ∈ I 2 . Since z ∈
[(a1 , a2 , · · · , ai ) : ai+1 ] ∩ I 2 = (a1 , a2 , · · · , ai ) ∩ I 2 ⊆ (a1 , a2 , · · · , ai ) ∩ QI. Notice that,
since Q/QI ∼
= (A/I)d , (a1 , a2 , · · · , ai ) ∩ QI = (a1 , a2 , · · · , ai )I holds true. Hence, we
get x = y + ai+1 z ∈ (a1 , a2 , · · · , ai )I 2 . Therefore we have (a1 , a2 , · · · , ai ) ∩ QI 2 =
(a1 , a2 , · · · , ai )I 2 .
4
Assume that n ≥ 1 and that our assertion holds true for n − 1. Take x ∈
(a1 , a2 , · · · , ai ) ∩ Qn+1 I 2 . Then, by the hypothesis of induction on i, we have
(a1 , a2 , · · · , ai ) ∩ Qn+1 I 2 ⊆ (a1 , a2 , · · · , ai , ai+1 ) ∩ Qn+1 I 2 = (a1 , a2 , · · · , ai , ai+1 )Qn I 2
Write x = y + ai+1 z with y ∈ (a1 , a2 , · · · , ai )Qn I 2 and z ∈ Qn I 2 . By the hypothesis of
induction on n, we have z ∈ [(a1 , a2 , · · · , ai ) : ai+1 ] ∩ Qn I 2 = (a1 , a2 , · · · , ai ) ∩ Qn I 2 =
(a1 , a2 , · · · , ai )Qn−1 I 2 . Therefore, we get x = y + ai+1 z ∈ (a1 , a2 , · · · , ai )Qn I 2 . Thus,
(a1 , a2 , · · · , ai ) ∩ Qn+1 I 2 = (a1 , a2 , · · · , ai )Qn I 2 as required.
Lemma 2.4. Suppose that Q ∩ I 2 = QI. Then Qn+1 ∩ Qn I 2 = Qn+1 I for all n ≥ 0.
Proof. We have P
only to show that Qn+1 ∩Qn I 2 ⊆ Qn+1 I for n ≥ 0. Take f ∈ Qn+1 ∩Qn I 2
and write f = |α|=n xα aα1 1 aα2 2 · · · aαd d with xα ∈ I 2 , where α = (α1 , α2 , · · · , αd ) ∈ Zd
Pd
P
αd n
α1 α2
n
and |α| =
i=1 αi . Then, we have f t =
|α|=n xα a1 a2 · · · ad t = 0 in G(Q) =
R′ (Q)/t−1 R′ (Q), where f tn and xα aα1 1 aα2 2 · · · aαd d tn = 0 denote the images of f tn and
xα aα1 1 aα2 2 · · · aαd d tn in G(Q) respectively. Because G(Q) ∼
= (A/Q)[a1 t, a2 t, · · · , ad t] is the
polynomial ring over the ring A/Q, we have xα ∈ Q ∩ I 2 = QI. Thus f ∈ Qn+1 I so
that Qn+1 ∩ Qn I 2 = Qn+1 I holds true.
Lemma 2.5. Suppose that Q ∩ I 2 = QI. Then the following sequences
a
(1) 0 → T /IT →1 T /I 2 T → T /[I 2 T + a1 T ] → 0, and
a1 t
(2) 0 → T /[I 2 T + QT ](−1) →
T /[I 2 T + a1 T ] → T /[I 2 T + a1 tT + a1 T ] → 0
of graded T -modules are exact.
Proof. (1) Let us consider the homomorphism
φ : T → T /I 2 T
of graded T -modules such that φ(f ) = a1 f for f ∈ T where a1 f denotes the image of
a1 f in T /I 2 T . Because φ(IT ) = (0) and Cokerφ = T /[I 2 T + a1 T ], we have only to show
that ker φ ⊆ IT . Take f ∈ [ker φ]n and write f = xtn with n ≥ 0 and x ∈ Qn . Then we
have a1 f = a1 xtn ∈ I 2 T so that a1 x ∈ Qn I 2 ∩ Qn+1 = Qn+1 I by Lemma 2.4. Because
a1 t ∈ T forms a nonzero divisor on T /IT (notice that T /IT ∼
= (A/I)[a1 t, a2 t, · · · , ad t]
is the polynomial ring over the ring A/I), we have (a1 ) ∩ Qn+1 I = a1 Qn I so that
x ∈ Qn I. Therefore f ∈ IT , and hence ker φ ⊆ IT . Thus, we get the first required
exact sequence.
(2) Let us consider the homomorphism
ϕ : T (−1) → T /[I 2 T + a1 T ]
of graded T -modules such that ϕ(f ) = a1 tf for f ∈ T where a1 tf denotes the image of
a1 tf in T /[I 2 T +a1 T ]. Because ϕ(I 2 T +QT ) = (0) and Cokerφ = T /[I 2 T +a1 tT +a1 T ],
we need to show that [ker ϕ]n ⊆ I 2 Tn−1 + QTn−1 for all n ≥ 1. Take f ∈ [ker ϕ]n and
write f = xtn−1 with n ≥ 1 and x ∈ Qn−1 . Then we have a1 tf = a1 xtn ∈ I 2 T + a1 T so
that a1 x ∈ Qn I 2 + a1 Qn . Write a1 x = y + a1 z with y ∈ Qn I 2 and z ∈ Qn . Then have
a1 (x − z) = y ∈ (a1 ) ∩ Qn I 2 = a1 Qn−1 I 2
by Lemma 2.3. Hence x − z ∈ Qn−1 I 2 so that x ∈ Qn−1 I 2 + Qn . Therefore f ∈
I 2 Tn−1 + QTn−1 and hence [ker ϕ]n ⊆ I 2 Tn−1 + QTn−1 . Consequently, we get the second
required exact sequence.
5
The following Lemma 2.6 is the crucial fact in the proof of Proposition 2.2.
Lemma 2.6. Assume that Q ∩ I 2 = QI. Then we have AssT (T /I 2 T ) = {mT }.
Proof. Take P ∈ AssT (T /I 2 T ) then we have mT ⊆ P , because mℓ (T /I 2 T ) = (0) for
ℓ ≫ 0. Assume that mT ( P then htT P ≥ 2, because mT is a height one prime ideal
in T . Consider the following exact sequences
0 → TP /ITP → TP /I 2 TP → TP /[I 2 TP + a1 TP ] → 0 (∗1 ) and
0 → TP /[I 2 TP + QTP ] → TP /[I 2 TP + a1 TP ] → TP /[I 2 TP + a1 tTP + a1 TP ] → 0 (∗2 )
of TP -modules, which follow from the exact sequences in Lemma 2.5. Then, since
depthTP TP /ITP > 0 (notice that T /IT ∼
= (A/I)[X1 , X2 , · · · , Xd ] is the polynomial
ring with d indeterminates over the ring A/I) and depthTP TP /I 2 TP = 0, we have
depthTP TP /[I 2 TP + a1 TP ] = 0 by the exact sequence (∗1 ). Notice that T /[I 2 T + QT ] ∼
=
2
(A/[I + Q])[X1 , X2 , · · · , Xd ] is the polynomial rings with d indeterminates over the
ring A/[I 2 + Q]. Hence, depthTP TP /[I 2 TP + a1 tTP + a1 TP ] = 0 by the exact sequence
(∗2 ). Then, because a1 t ∈ AnnT (T /[I 2 T + a1 tT + a1 T ]), we have a1 t ∈ P . Therefore
depthTP TP /I 2 TP > 0 by Lemma 2.3, however it is impossible. Thus, P = mT as
required.
Let us now give a proof of Proposition 2.2.
Proof of Proposition 2.2. Take P ∈ AssT C. Then we have mT ⊆ P , because mℓ C = (0)
for some ℓ ≫ 0 by Lemma 2.1 (1). Suppose that mT ( P then htT P ≥ 2, because mT
is a height one prime ideal in T . We look at the following exact sequences
0 → I 2 TP → I 2 RP → CP → 0 (∗3 ) and 0 → I 2 TP → TP → TP /I 2 TP → 0 (∗4 )
of TP -modules which follows from the canonical exact sequences
0 → I 2 T (−1) → I 2 R(−1) → C → 0 and 0 → I 2 T → T → T /I 2 T → 0
of T -modules. We notice here that depthTP I 2 RP > 0, because a1 ∈ A is a nonzero
divisor on I 2 R. Thanks to the depth lemma and the exact sequence (∗3 ), we get
depthTP I 2 TP = 1, because depthTP I 2 RP > 0 and depthTP CP = 0. Then, since
depthTP TP ≥ 2, we have depthTP TP /I 2 TP = 0 by the exact sequence (∗4 ). Therefore,
we have P = mT by Lemma 2.6, which is impossible. Thus, P = mT as required.
The following techniques are due
Vaz Pinto [VP, Section 2].
L to M.
(1)
n−1 2
∼
Let L = L = T S1 then L = n≥1 Q I /Qn I and S/L ∼
= C as graded T -modules.
Then there exist a canonical exact sequence
0 → L → S → C → 0 (†)
of graded T -modules (Definition 1.1). We set D = (I 2 /QI) ⊗A (T /AnnA (I 2 /QI)T ).
Notice here that D forms a graded T -module and T /AnnA (I 2 /QI)T ∼
=
2
(A/AnnA (I /QI))[X1 , X2 , · · · , Xd ] is the polynomial ring with d indeterminates over
the ring A/AnnA (I 2 /QI). Let
θ : D(−1) → L
P
denotes an epimorphism of graded T -modules such that θ( α xα ⊗ X1α1 X2α2 · · · Xdαd ) =
P
αd |α|+1
α1 α2
for xα ∈ I 2 and α = (α1 , α2 , · · · , αd ) ∈ Zd with αi ≥ 0
α xα a1 a2 · · · ad t
6
P
(1 ≤ i ≤ d), where |α| = di=1 αi , and xα and xα aα1 1 aα2 2 · · · aαd d t|α|+1 denote the images
of xα in I 2 /QI and xα aα1 1 aα2 2 · · · aαd d t|α|+1 in L.
Then we have the following lemma.
Lemma 2.7. Suppose that Q ∩ I 2 = QI. Then the map θ : D(−1) → L is an isomorphism of graded T -modules.
Proof. We have only to show that ker θ = (0). Assume that ker θ 6= (0) and let n ≥ 2
as the least integer so that [ker θ]n 6= (0) (notice that [ker θ]n = (0) for all n ≤ 1). Take
0 6= g ∈ [ker θ]n and we set
P
Γ = {(α1 , α2 , · · · , αd−1 , 0) ∈ Zd | αi ≥ 0 for 1 ≤ i ≤ d − 1 and d−1
i=1 αi = n − 1} and
P
Γ′ = {(β1 , β2 , · · · , βd ) ∈ Zd | βi ≥ 0 for 1 ≤ i ≤ d − 1, βd ≥ 1, and di=1 βi = n − 1}.
Then because
P
Γ ∪ Γ′ = {(α1 , α2 , · · · , αd ) ∈ Zd | αi ≥ 0 for 1 ≤ i ≤ d and di=1 αi = n − 1}
we may write
g =
X
α∈Γ∪Γ′
=
X
α∈Γ
xα ⊗ X1α1 X2α2 · · · Xdαd
α
d−1
+
xα ⊗ X1α1 X2α2 · · · Xd−1
X
β∈Γ′
xβ ⊗ X1β1 X2β2 · · · Xdβd ,
where xα , xβ denote the images of xα , xβ ∈ I 2 in I 2 /QI, respectively. We may assume
P
that β∈Γ′ xβ ⊗ X1β1 X2β2 · · · Xdβd 6= 0 in D. Then we have
X
X
αd−1 n
θ(g) =
xα aα1 1 aα2 2 · · · ad−1
t +
xβ aβ1 1 aβ2 2 · · · aβd d tn = 0
α∈Γ
so that
X
α∈Γ
β∈Γ′
α
d−1
+
xα aα1 1 aα2 2 · · · ad−1
X
β∈Γ′
xβ aβ1 1 aβ2 2 · · · aβd d ∈ Qn I.
Because Qn I = (a1 , a2 , · · · , ad−1 )n I + ad Qn−1 I and βd ≥ 1, we may write
!
X
X
αd−1
β1 β2
βd −1
α1 α2
= τ + ad ρ
xα a1 a2 · · · ad−1 + ad
xβ a1 a2 · · · ad
α∈Γ
β∈Γ′
with τ ∈ (a1 , a2 , · · · , ad−1 )n I and ρ ∈ Qn−1 I. Then because
!
X
X
αd−1
ad
xβ aβ1 1 aβ2 2 · · · aβd d −1 − ρ = τ −
xα aα1 1 aα2 2 · · · ad−1
∈ (a1 , a2 , · · · , ad−1 )n−1
β∈Γ′
α∈Γ
we have, by Lemma 2.4,
X
xβ aβ1 1 aβ2 2 · · · aβd d −1 − ρ ∈ (a1 , a2 , · · · , ad−1 )n−1 ∩ Qn−2 I 2 ⊆ Qn−1 I.
β∈Γ′
7
P
P
Therefore, β∈Γ′ xβ aβ1 1 aβ2 2 · · · aβd d −1 ∈ Qn−1 I and hence θ( β∈Γ′ xβ ⊗X1β1 X2β2 · · · Xdβd −1 ) =
P
P
0. Then, because β∈Γ′ xβ ⊗ X1β1 X2β2 · · · Xdβd −1 ∈ [ker θ]n−1 = (0), we have β∈Γ′ xβ ⊗
X1β1 X2β2 · · · Xdβd = 0, which is contradiction. Thus ker θ = (0). Consequently, the map
θ : D(−1) → L is an isomorphism.
Thanks to Lemma 2.7, we can prove the following result.
Proposition 2.8. Suppose that Q ∩ I 2 = QI. Then we have
n+d−1
n+d
2
n+1
− {e0 (I) − ℓA (A/I) + ℓA (I /QI)}
ℓA (A/I ) = e0 (I)
d−1
d
n+d−2
− ℓA (Cn )
+ℓA (I 2 /QI)
d−2
for all n ≥ 0.
Proof. We have, for all n ≥ 0,
ℓA (Sn ) = ℓA (Ln ) + ℓA (Cn )
n+d−2
2
+ ℓA (Cn )
= ℓA (I /QI)
d−1
n+d−2
n+d−1
2
2
+ ℓA (Cn )
− ℓA (I /QI)
= ℓA (I /QI)
d−2
d−1
by the exact sequence
0 → L → S → C → 0 (†)
∼
and the isomorphisms L = D(−1) ∼
= (I 2 /QI) ⊗A (A/AnnA (I 2 /QI))[X1 , X2 , · · · , Xd ] of
graded T -modules (see Lemma 2.7). Therefore we have, for all n ≥ 0,
n+d−1
n+d
n+1
− ℓA (Sn )
− {e0 (I) − ℓA (A/I)}
ℓA (A/I ) = e0 (I)
d−1
d
n+d−1
n+d
− {e0 (I) − ℓA (A/I)}
= e0 (I)
d−1
d
n+d−2
n+d−1
2
2
+ ℓA (Cn )
− ℓA (I /QI)
− ℓA (I /QI)
d−2
d−1
n+d−1
n+d
2
− {e0 (I) − ℓA (A/I) + ℓA (I /QI)}
= e0 (I)
d−1
d
n+d−2
− ℓA (Cn )
+ℓA (I 2 /QI)
d−2
by [GNO, Proposition 2.2 (2)].
The following result specifies [GNO, Proposition 2.2 (3)] and, by using Proposition
2.2 and 2.8, the proof takes advantage of the same techniques.
Proposition 2.9. Suppose that Q ∩ I 2 = QI. Let p = mT . Then we have
e1 (I) = e0 (I) − ℓA (A/I) + ℓA (I 2 /QI) + ℓTp (Cp ).
8
Combining Lemma 2.1 (3) and Proposition 2.9 we obtain the following result that
was proven by Elias and Valla [EV, Theorem 2.1] in the case where I = m.
Corollary 2.10. Suppose that Q ∩ I 2 = QI. Then we have e1 (I) ≥ e0 (I) − ℓA (A/I) +
ℓA (I 2 /QI). The equality e1 (I) = e0 (I) − ℓA (A/I) + ℓA (I 2 /QI) holds true if and only if
I 3 = QI 2 . When this is the case, e2 (I) = e1 (I) − e0 (I) + ℓA (A/I) if d ≥ 2, ei (I) = 0
for all 3 ≤ i ≤ d, and G is a Cohen-Macaulay ring.
Let us introduce the relationship between the depth of the module C and the associated graded ring G of I.
Lemma 2.11. Suppose that Q ∩ I 2 = QI and C 6= (0). Let s = depthT C. Then we
have depthG ≥ s − 1. In particular, we have depthG = s − 1, if s ≤ d − 2.
Proof. Notice that L ∼
= D(−1) by Lemma 2.7 and D is a d dimensional Cohen-Macaulay
T -module. Therefore, we have s ≤ depthT S and, if s ≤ d − 2 then s = depthT S by the
exact sequence
0→L→S→C→0
(†).
Because depthG ≥ depthT S − 1 and, if depthT S ≤ d − 1 then depthG = depthT S − 1
by [GNO, Proposition 2.2 (4)], our assertions follow.
3. Proof of Theorem 1.2
The purpose of this section is to prove Theorem 1.2. Throughout this section, let I
be an integrally closed m-primary ideal.
Theorem 3.1. Suppose that I is integrally closed. Then the following conditions are
equivalent:
(1) e1 (I) = e0 (I) − ℓA (A/I) + ℓA (I 2 /QI) + 1,
(2) mC = (0) and rankB C = 1,
(3) there exists a non-zero graded ideal a of B such that C ∼
= a(−1) as graded
T -modules.
To prove Theorem 3.1, we need the following bound on e2 (I).
Lemma 3.2. ([I1, Theorem 12], [S2, Corollary 2.5], [RV3, Corollary 3.1], ) Suppose
d ≥ 2 and let I be an integrally closed ideal, then e2 (I) ≥ e1 (I) − e0 (I) + ℓA (A/I).
Proof of Theorem 3.1. Let p = mT then we see that e1 (I) = e0 (I) − ℓA (A/I) +
ℓA (I 2 /QI) + ℓTp (Cp ) by Lemma 2.9 and AssT C = {p} by Proposition 2.2.
(1) ⇒ (2) Since ℓTp (Cp ) = 1 and AssT C = {p}, we have mC = (0) and rankB C = 1.
(2) ⇒ (1) This is clear, because assertion (1) is equivalent to saying that ℓTp (Cp ) = 1.
(3) ⇒ (2) This is obvious.
(2) ⇒ (3) Because AssT C = {p}, C is torsion free B-module. If C is B-free, then
we have C ∼
= B(−2) as graded B-module, because C2 6= (0) and Cn = (0) for n ≤ 1.
Hence C ∼
= X1 B(−1) as graded B-modules with 0 6= X1 ∈ B1 .
Suppose that C is not B-free. Then we have d = dim A ≥ 2. Because rankB C = 1,
there exists a graded ideal a of B such that
C∼
= a(r)
as graded B-module for some r ∈ Z. Since every height one prime in the polynomial
ring B is principal, we may choose a with htB a ≥ 2. Then since ar+2 ∼
= a(r)2 ∼
= C2 6= (0)
9
and an = (0) for all n ≤ 0, we have r + 2 ≥ 1 so that r ≥ −1. It is now enough to show
that r = −1. Applying the exact sequence
0 → C → B(r) → (B/a)(r) → 0
of graded B-modules, we have
ℓA (Cn ) = ℓA (Br+n ) − ℓA ([B/a]r+n )
n+r+d−1
− ℓA ([B/a]r+n )
=
d−1
n+d−2
n+d−1
+ (lower terms)
+r
=
d−2
d−1
for all n ≫ 0, because htB a ≥ 2. Therefore, thanks to Proposition 2.8, we have
n+d−1
n+d
2
n+1
− {e0 (I) − ℓA (A/I) + ℓA (I /QI)}
ℓA (A/I ) = e0 (I)
d−1
d
n+d−2
− ℓA (Cn )
+ℓA (I 2 /QI)
d−2
n+d−1
n+d
2
− {e0 (I) − ℓA (A/I) + ℓA (I /QI) + 1}
= e0 (I)
d−1
d
n+d−2
+ (lower terms)
+{ℓA (I 2 /QI) − r}
d−2
for all n ≫ 0. Therefore e2 (I) = ℓA (I 2 /QI) − r. Then, because e2 (I) ≥ e1 (I) −
e0 (I) + ℓA (A/I) = ℓA (I 2 /QI) + 1 by Lemma 3.2 we have r ≤ −1. Thus r = −1 and so
C∼
= a(−1) as graded B-modules.
As a direct consequence of Theorem 3.1 the following result holds true.
Proposition 3.3. Assume that I is integrally closed. Suppose that e1 (I) = e0 (I) −
ℓA (A/I) + ℓA (I 2 /QI) + 1 and I 4 = QI 3 and let c = ℓA (I 3 /QI 2 ). Then
(1) 1 ≤ c ≤ d and µB (C) = c.
(2) depth G ≥ d − c and depthT C = d − c + 1,
(3) depth G = d − c, if c ≥ 2.
(4) Suppose c = 1 < d. Then HPI (n) = ℓA (A/I n+1 ) for all n ≥ 0 and
if i = 2,
e1 (I) − e0 (I) + ℓA (A/I) + 1
1
if i = 3 and d ≥ 3,
ei (I) =
0
if 4 ≤ i ≤ d.
(5) Suppose 2 ≤ c < d. Then HPI (n) = ℓA (A/I n+1) for all n ≥ 0 and
if i = 2,
e1 (I) − e0 (I) + ℓA (A/I)
0
if i 6= c + 1, c + 2, 3 ≤ i ≤ d
ei (I) =
(−1)c+1
if i = c + 1, c + 2, 3 ≤ i ≤ d
10
(6) Suppose c = d. Then HPI (n) = ℓA (A/I n+1 ) for all n ≥ 2 and
e1 (I) − e0 (I) + ℓA (A/I)
if i = 2 and d ≥ 2,
ei (I) =
0
if 3 ≤ i ≤ d
(7) The Hilbert series HSI (z) is given by
HSI (z) =
ℓA (A/I) + {e0 (I) − ℓA (A/I) − ℓA (I 2 /QI) − 1}z + {ℓA (I 2 /QI) + 1}z 2 + (1 − z)c+1 z
.
(1 − z)d
Proof. (1) We have C = T C2 , since I 4 = QI 3 (c.f. Lemma 2.1 (5)). Therefore, thanks
to Theorem 3.1, C ∼
= a(−1) as graded T -modules, where a = (X1 , X2 , · · · , Xc )B is an
ideal generated by linear forms {Xi }1≤i≤c of B. Hence, we get 1 ≤ c ≤ d and µB (C) = c.
(4), (5), (6) Let us consider the exact sequence
0 → C → B(−1) → (B/a)(−1) → 0
(∗5 )
of graded B-modules. Then, we have
ℓA (Cn ) = ℓA (Bn−1 ) − ℓA ([B/a]n−1 )
n−1+d−c−1
n−1+d−1
−
=
d−c−1
d−1
n+d−c−2
n+d−c−1
n+d−2
n+d−1
+
−
−
=
d−c−2
d−c−1
d−2
d−1
for all n ≥ 0 (resp. n ≥ 2) if 1 ≤ c ≤ d − 1 (resp. c = d). Therefore, our assertions (4),
(5), and (6) follow by Proposition 2.8.
(7) We have
z − (1 − z)c z
HSC (z) = HSB (z)z − HSB/a (z)z =
(1 − z)d
by the above exact sequence (∗5 ), where HS∗ (z) denotes the Hilbert series of the graded
modules. We also have
{ℓA (I 2 /QI) + 1}z − (1 − z)c z
HSS (z) = HSL (z) + HSC (z) =
(1 − z)d
by the exact sequence
0 → L → S → C → 0 (†)
∼
and isomorphisms L = D(−1) ∼
= (I 2 /QI) ⊗A (A/AnnA (I 2 /QI))[X1 , X2 , · · · , Xd ](−1)
of graded T -modules (Lemma 2.7). Then, because
HSI (z) =
ℓA (A/I) + {e0 (I) − ℓA (A/I)}z
− (1 − z)HSS (z)
(1 − z)d
([VP], [RV3, Proposition 6.3]), we can get the required result.
We prove now (2), (3). We have depthT C = d − c + 1 by the exact sequence (∗5 )
so that depthG ≥ d − c and, if c ≥ 3, then depthG = d − c by Lemma 2.11. Let
us consider the case where c = 2 and we need to show depthG = d − 2. Assume
depthG ≥ d − 1, then S is a Cohen-Macaulay T -module by [GNO, Proposition 2.2].
Taking the local cohomology functors HiM (∗) of T with respect to the graded maximal
11
ideal M = mT + T+ to the above exact sequence (†) of graded T -module, we get a
monomonophism
d
Hd−1
M (C) ֒→ HM (L)
of graded T -module.
Because C ∼
= (X1 , X2 )B(−1), we have Hd−1 (C) ∼
=
M
d−1
Hd−2
M (B/(X1 , X2 )B)(−1) as graded T -modules so that [HM (C)]−d+3 6= (0) (notice
On the other hand, we have
that B/(X1 , X2 )B ∼
= (A/m)[X3 , X4 , · · · , Xd ]).
[HdM (L)]n = (0) for all n ≥ −d + 2, because L ∼
= (I 2 /QI) ⊗A
= D(−1) ∼
(A/AnnA (I 2 /QI))[X1 , X2 , · · · , Xd ](−1) by Lemma 2.7. However, it is impossible.
Therefore, depth G = d − 2 if c = 2.
We prove now Theorem 1.2. Assume assertion (1) in Theorem 1.2. Then we have an
isomorphism C ∼
= a(−1) as graded B-modules for a graded ideal a in B by Theorem
3.1. Once we are able to show I 4 = QI 3 , then, because C = T C2 by Lemma 2.1
(5), the ideal a is generated by linearly independent linear forms {Xi }1≤i≤c of B with
c = ℓA (I 3 /QI 2 ) (recall that a1 ∼
= I 3 /QI 2 by Lemma 2.1 (2)). Therefore, the
= C2 ∼
implication (1) ⇒ (3) in Theorem 1.2 follows. We also notice that, the last assertions
of Theorem 1.2 follow by Proposition 3.3.
Thus our Theorem 1.2 has been proven modulo the following theorem.
Theorem 3.4. Assume that I is integrally closed. Suppose that e1 (I) = e0 (I) −
ℓA (A/I) + ℓA (I 2 /QI) + 1. Then I 4 = QI 3 .
Proof. We proceed by induction on d.
P Suppose that d = 1. Then the result follows by
[HM, Proposition 4.6] since e1 (I) = i≥0 ℓA (I i+1 /QI i ).
Assume that d ≥ 2 and that our assertion holds true for d − 1. Since the residue
class field A/m of A is infinite, without loss of generality, we may assume that a1 is
superficial element of I and I/(a1 ) is integrally closed (c.f. [I1, page. 648], [RV3,
Proposition 1.1]). We set A′ = A/(a1 ), I ′ = I/(a1 ), Q′ = Q/(a1 ). We then have
e1 (I ′ ) = e0 (I ′ ) − ℓA′ (A′ /I ′ ) + ℓA′ (I ′2 /Q′ I ′ ) + 1 because ei (I ′ ) = ei (I) for 0 ≤ i ≤ d − 1,
ℓA′ (A′ /I ′ ) = ℓA (A/I), and since (a1 ) ∩ I 2 = a1 I, ℓA′ (I ′2 /Q′ I ′ ) = ℓA (I 2 /QI). Then the
inductive assumption on d says that I ′4 = Q′ I ′3 holds true. If depthG(I ′ ) > 0 then,
thanks to Sally’s technique (c.f. [S1], [HM, Lemma 2.2]), a1 t is a non-zero divisor on
G. Then we have I 4 = QI 3 .
Assume that depthG(I ′ ) = 0. Then because e1 (I ′ ) = e0 (I ′ ) − ℓA′ (A′ /I ′ ) +
ℓA′ (I ′2 /Q′ I ′ ) + 1 and I ′4 = Q′ I ′3 , we have ℓA′ (I ′ 3 /Q′ I ′ 2 ) = d − 1 by Proposition 3.3 (2).
Since I ′ 3 /Q′ I ′ 2 is a homomorphic image of I 3 /QI 2 , we have ℓA (C2 ) = ℓA (I 3 /QI 2 ) ≥
d − 1. Let us now take an isomorphism
ϕ : C → a(−1)
of graded B-modules, where a is a graded ideal of B (c.f. Theorem 3.1). Then since
ℓA (a1 ) = ℓA (C2 ) ≥ d − 1, we have Y1 , Y2 , · · · , Yd−1 ∈ a where Y1 , Y2 , · · · , Yd−1 denote
linearly independent linear forms of B, which we enlarge to a basis Y1 , Y2 , · · · , Yd−1 , Yd
of B1 . If a = (Y1 , Y2 , · · · , Yd−1 )B then, because C = T C2 , we have I 4 = QI 3 by Lemma
2.1 (5). Assume that a 6= (Y1 , Y2, · · · , Yd−1)B. Then, since B = k[Y1 , Y2 , · · · , Yd ], the
ideal a/(Y1 , Y2 , · · · , Yd−1 ) is principal so that a = (Y1 , Y2 , · · · , Yd−1 , Ydℓ )B for some ℓ ≥ 1.
We need the following.
Claim 1. We have ℓ = 1 or ℓ = 2.
12
Proof of Claim 1. Assume that ℓ ≥ 3. Then I 4 /[QI 3 + mI 4 ] ∼
= [C/MC]3 = (0),
where M = mT + T+ denotes the graded maximal ideal of T . Therefore I 4 = QI 3
by Nakayama’s lemma so that C = BC2 , which is impossible. Thus ℓ = 1 or ℓ = 2.
We have to show that ℓ = 1. Assume that ℓ = 2. Let us write, for each 1 ≤ i ≤ d,
Yi = bi t with bi ∈ Q, where bi t denotes the image of bi t ∈ T in B. We notice here that
Q = (b1 , b2 , · · · , bd ), because Y1 , Y2 , · · · , Yd forms a k-basis of B1 .
Let us choose elements fi ∈ C2 for 1 ≤ i ≤ d − 1 and fd ∈ C3 so that ϕ(fi ) = Yi
for 1 ≤ i ≤ d − 1 and ϕ(fd ) = Yd2 . Let zi ∈ I 3 for 1 ≤ i ≤ d − 1 and zd ∈ I 4 so that
{fi }1≤i≤d−1 and fd are, respectively, the images of {zi t2 }1≤i≤d−1 and zd t3 in C. Let us
now consider the relations Yd2 fi = Yi fd in C for 1 ≤ i ≤ d − 1, that is
b2d zi − bi zd ∈ Q3 I 2
for 1 ≤ i ≤ d − 1. Notice that
by Lemma 2.3 and write
(b1 , bd ) ∩ Q3 I 2 = (b1 , bd )Q2 I 2
b2d z1 − b1 zd = b1 τ1 + bd τd
(1)
with τ1 , τd ∈ Q2 I 2 . Then we have
bd (bd z1 − τd ) = b1 (zd + τ1 )
so that bd z1 − τd ∈ (b1 ) because b1 , bd forms a regular sequence on A. Since τd ∈
(b1 , bd ) ∩ Q2 I 2 = (b1 , bd )QI 2 by Lemma 2.3, there exist elements τ1′ , τd′ ∈ QI 2 such that
τd = b1 τ1′ + bd τd′ . Then by the equality (1) we have
b2d (z1 − τd′ ) = b1 (zd + τ1 + bd τ1′ )
(2)
so that we have z1 −τd′ ∈ (b1 ). Hence z1 ∈ QI 2 +(b1 ). The same argument works for each
1 ≤ i ≤ d − 1 to see zi ∈ QI 2 + (bi ). Therefore, because I 3 = QI 2 + (z1 , z2 , · · · , zd−1 ),
we have I 3 ⊆ bd I 2 + (b1 , b2 , · · · , bd−1 ) and hence
I 4 ⊆ b2d I 2 + (b1 , b2 , · · · , bd−1 ).
Then, because zd + τ1 + bd τ1′ ∈ I 4 , there exist elements h ∈ I 2 and η ∈ (b1 , b2 , · · · , bd−1 )
such that zd + τ1 + bd τ1′ = b2d h + η. Then we have
b2d (z1 − τd′ − b1 h) = b1 η
(3)
by the equality (2). Since b1 , b2 , · · · , bd is a regular sequence on A, η ∈ (b2d ) ∩
(b1 , b2 , · · · , bd−1 ) = b2d (b1 , b2 , · · · , bd−1 ). Write η = b2d η ′ with η ′ ∈ (b1 , b2 , · · · , bd−1 ),
then we have
b2d (z1 − τd′ − b1 h) = b1 b2d η ′
by the equality (3), so that z1 − τd′ − b1 h = b1 η ′ . Then we have b1 η ′ = z1 − τd′ − b1 h ∈
Q2 ∩I 3 = Q2 I, since Q2 ∩I 3 ⊆ Q2 ∩Q3 = Q2 Q = Q2 I (c.f. [H, I2]), where J denotes the
integral closure of an ideal J. Hence z1 ∈ QI 2 , because τ1′ , b1 h, b1 η ′ ∈ QI 2 . Therefore
f1 = 0 in C, which is impossible. Thus ℓ = 1 so that we have a = (X1 , X2 , · · · , Xd )B.
Therefore C = T C2 , that is I 4 = QI 3 . This completes the proof of Theorem 3.4 and
that of Theorem 1.2 as well.
13
4. Consequences
The purpose of this section is to present some consequences of Theorem 1.2. Let us
begin with the following which is exactly the case where c = 1 in Theorem 1.2.
Theorem 4.1. Assume that I is integrally closed. Then the following conditions are
equivalent.
(1) C ∼
= B(−2) as graded T -modules.
(2) e1 (I) = e0 (I) − ℓA (A/I) + ℓA (I 2 /IQ) + 1, and if d ≥ 2 then e2 (I) 6= e1 (I) −
e0 (I) + ℓA (A/I).
(3) ℓA (I 3 /QI 2 ) = 1 and I 4 = QI 3 .
When this is the case, the following assertions follow.
(i) depthG ≥ d − 1.
(ii) e2 (I) = e1 (I) − e0 (I) + ℓA (A/I) + 1 if d ≥ 2.
(iii) e3 (I) = 1 if d ≥ 3, and ei (I) = 0 for 4 ≤ i ≤ d.
(iv) The Hilbert series HSI (z) is given by
HSI (z) =
ℓA (A/I) + {e0 (I) − ℓA (A/I) − ℓA (I 2 /QI)}z + {ℓA (I 2 /QI) − 1}z 2 + z 3
.
(1 − z)d
Proof. For (1) ⇔ (2), (1) ⇒ (3) and the last assertions see Theorem 1.2 with c = 1.
(3) ⇒ (1): Since mI n+1 = Qn−1 I 2 for all n ≥ 2, we have mC = (0) by Lemma 2.1
(4). Then we have an epimorphism B(−2) → C → 0 of graded T -modules, which must
be an isomorphism because dimT C = d (Proposition 2.2).
S
S
Let Je = n≥1 [J n+1 :A J n ] = n≥1 J n+1 :A (an1 , an2 , · · · , and ) denote the Ratliff-Rush
closure of an m-primary ideal J in A, which is the largest m-primary ideal in A such
e = ei (J) for all 0 ≤ i ≤ d (c.f. [RR]).
that J ⊆ Je and ei (J)
Let us note the following remark.
Remark 4.2. Assume that I is integrally closed. Then, by [H, I1], Qn ∩ I n+1 =
Qn ∩ Qn+1 = Qn Q = Qn I holds true for n ≥ 1, where J denotes the integral closure of
n+1 = Qn I for all n ≥ 1.
an ideal J. Thus, we have Qn ∩ Ig
The following result correspond to the case where c = d in Theorem 1.2. In section
5 we give an example of the maximal ideal which satisfy assertion (1) in Theorem 4.3.
Theorem 4.3. Suppose that d ≥ 2 and assume that I is integrally closed. Then the
following conditions are equivalent:
(1) C ∼
= B+ (−1) as graded T -modules.
(2) e1 (I) = e0 (I) − ℓA (A/I) + ℓA (I 2 /QI) + 1, e2 (I) = e1 (I) − e0 (I) + ℓA (A/I), and
ei (I) = 0 for all 3 ≤ i ≤ d.
n+1 = QIen for all n ≥ 2.
(3) ℓA (Ie2 /I 2 ) = 1 and Ig
When this is the case, the associated graded ring G of I is a Buchsbaum ring with
depth G = 0 and the Buchsbaum invariant I(G) = d.
P
Proof. We set c = ℓA (I 3 /QI 2 ) and F = {Ien }n≥0 . Let R′ (F ) = n∈Z Ien t ⊆ A[t, t−1 ] and
G(F ) = R′ (F )/t−1R′ (F ). Let ei (F ) denote the ith Hilbert coefficients of the filtration
F for 0 ≤ i ≤ d.
14
(1) ⇒ (2) follows from Theorem 1.2, because c = ℓA (C2 ) = d.
(2) ⇒ (1) Because e2 (I) = e1 (I) − e0 (I) + ℓA (A/I) and ei (I) = 0 for all 3 ≤ i ≤ d,
we have c = d by Theorem 1.2. Therefore C ∼
= B+ (−1) as graded T -modules.
(1) ⇒ (3) Since c = d, we have depthG = 0 by Theorem 1.2 (ii). We apply local
cohomology functors HiM (∗) of T with respect to the graded maximal ideal M = mT +T+
of T to the exact sequences
0 → I 2 R(−1) → IR(−1) → G+ → 0 and 0 → I 2 T (−1) → I 2 R(−1) → C → 0
of graded T -modules and get derived monomorphisms
H0M (G+ ) ֒→ H1M (I 2 R)(−1) and H1M (I 2 R)(−1) ֒→ H1M (C)
because depthT IR > 0 and depthT I 2 T ≥ 2 (recall that T is a Cohen-Macaulay ring with
dim T = d + 1 and depthT T /I 2 T ≥ 1 by Lemma 2.3). We furthermore have H1M (C) ∼
=
0
(B/B+ )(−1) since C ∼
B
(−1).
Since
I
is
integrally
closed,
we
have
[H
(G)]
=
(0)
= +
0
M
so that H0M (G) ∼
= H0M (G+ ) 6= (0). Then because ℓA (B/B+ ) = 1, we have isomorphisms
H0 (G) ∼
= B/B+ (−1)
= H1 (C) ∼
= H1 (I 2 R)(−1) ∼
M
M
M
H0M (G)
[H0M (G)]1
of graded B-modules and hence
=
2
2
e
2
e
2
I /I we have ℓA (I /I ) = 1. Hence we have
∼
= B/B+ . Then since [H0M (G)]1 ∼
=
e1 (F ) = e1 (I) = e0 (I) − ℓA (A/I) + ℓA (I 2 /QI) + 1
= e0 (F ) − ℓA (A/I) + ℓA (Ie2 /QI) + 1 − ℓA (Ie2 /I 2 )
= e0 (F ) − ℓA (A/I) + ℓA (Ie2 /Q ∩ Ie2 )
because Ie = I, Q ∩ Ie2 = QI by Remark 4.2, and ei (F ) = ei (I) for i = 0, 1. Therefore,
n+1 = QIen for all n ≥ 2 by [GR, Theorem 2.2].
Ig
n+1 = QIen for all n ≥ 2, we
(3) ⇒ (2) Because Q ∩ Ie2 = QI by Remark 4.2 and Ig
have e1 (F ) = e0 (F ) − ℓA (A/I) + ℓA (Ie2 /QI) and G(F ) is a Cohen-Macaulay ring by
[GR, Theorem 2.2]. Then, since ℓA (Ie2 /I 2 ) = 1, we have e1 (I) = e0 (I) − ℓA (A/I) +
ℓA (I 2 /QI) + 1. We furthermore have e2 (I) = e2 (F ) = ℓA (Ie2 /QI) = ℓA (I 2 /QI) + 1 =
e1 (I) − e0 (I) + ℓA (A/I), and ei (I) = ei (F ) = 0 for 3 ≤ i ≤ d, because G(F ) is a
Cohen-Macaulay ring (c.f. [HM, Proposition 4.6]).
Assume one of the equivalent conditions. We have H0M (G) = [H0M (G)]1 by the proof
of the implication (1) ⇒ (3). Let n ≥ 3 be an integer. We then have
Ien /I n = Ien ∩ I n−1 /I n ∼
= [H0M (G)]n−1 = (0)
because Ien = Qn−2 Ie2 ⊆ I n−1 . Therefore, we have Ien = I n for all n ≥ 3.
We set W = R′ (F )/R′ and look at the exact sequence
0 → R′ → R′ (F ) → W → 0 (∗′ )
of graded R′ -modules. Since Ien = I n for all n 6= 2, we have W = W2 = Ie2 /I 2
so that ℓA (W ) = 1. Then, because G(F ) = R′ (F )/t−1R′ (F ) is a Cohen-Macaulay
ring, so is R′ (F ). Let M ′ = (m, R+ , t−1 )R′ be the unique graded maximal ideal in
R′ . Then applying local cohomology functors HiM ′ (∗) to the exact sequence (∗′ ) yields
15
HiM ′ (R′ ) = (0) for all i 6= 1, d + 1 and H1M ′ (R′ ) = W . Since mW = (0), we have
mH1M ′ (R′ ) = (0). Thus, R′ is a Buchsbaum ring with the Buchsbaum invariant
d
X
d
ℓA (HiM ′ (R′ )) = d
I(R ) =
i
i=0
′
and hence so is the graded ring G = R′ /t−1 R′ . This completes the proof of Theorem
4.3.
In the rest of this section, we explore the relationship between the inequality of
Northcott [N] and the structure of the graded module C of an integrally closed ideal.
It is well known that the inequality e1 (I) ≥ e0 (I) − ℓA (A/I) holds true ([N]) and the
equality holds if and only if I 2 = QI ([H, Theorem 2.1]). When this is the case, the
associated graded ring G of I is Cohen-Macaulay.
Suppose that I is integrally closed and e1 (I) = e0 (I) − ℓA (A/I) + 1 then, thanks to
[I1, Corollary 14], we have I 3 = QI 2 and the associated graded ring G of I is CohenMacaulay. Thus the integrally closed ideal I with e1 (I) ≤ e0 (I) − ℓA (A/I) + 1 seems
satisfactory understood. In this section, we briefly study the integrally closed ideals I
with e1 (I) = e0 (I) − ℓA (A/I) + 2, and e1 (I) = e0 (I) − ℓA (A/I) + 3.
Let us begin with the following.
Theorem 4.4. Assume that I is integrally closed. Suppose that e1 (I) = e0 (I) −
ℓA (A/I) + 2 and I 3 6= QI 2 . Then the following assertions hold true.
(1)
(2)
(3)
(4)
(5)
ℓA (I 2 /QI) = ℓA (I 3 /QI 2 ) = 1, and I 4 = QI 3 .
C∼
= B(−2) as graded T -modules.
depth G = d − 1.
e2 (I) = 3 if d ≥ 2, e3 (I) = 1 if d ≥ 3, and ei (I) = 0 for 4 ≤ i ≤ d.
The Hilbert series HSI (z) is given by
HSI (z) =
ℓA (A/I) + {e0 (I) − ℓA (A/I) − 1}z + z 3
.
(1 − z)d
Proof. Because I 3 6= QI 2 , it follows from Corollary 2.10 that
0 < ℓA (I 2 /QI) < e1 (I) − e0 (I) + ℓA (A/I) = 2.
Therefore, ℓA (I 2 /QI) = 1 and hence e1 (I) = e0 (I) − ℓA (A/I) + ℓA (I 2 /QI) + 1. Let
I 2 = QI + (xy) with x, y ∈ I\Q. Then I 3 = QI 2 + (x2 y), so that ℓA (I 3 /QI 2 ) = 1 since
I 3 6= QI 2 and mI 2 ⊆ QI. Thanks to Theorem 1.2, C ∼
= B(−2) as graded T -modules,
so that assertions (1), (2), (4), and (5) follow, and depth G ≥ d − 1 by Theorem 4.1.
Since I 3 ⊆ QI, G is not a Cohen-Macaulay ring, for otherwise I 3 = Q ∩ I 3 = QI 2 , so
that depthG = d − 1. This completes the proof of Theorem 4.4.
Notice that the following result also follows by [RV3, Theorem 4.6].
Corollary 4.5. Assume that I is integrally closed and suppose that e1 (I) = e0 (I) −
ℓA (A/I) + 2. Then depth G ≥ d − 1 and I 4 = QI 3 , and the graded ring G is CohenMacaulay if and only if I 3 = QI 2 .
16
Before closing this section, we briefly study the integrally closed ideal I with e1 (I) =
e0 (I) − ℓA (A/I) + 3. Suppose that e1 (I) = e0 (I) − ℓA (A/I) + 3 then we have
0 < ℓA (I 2 /QI) ≤ e1 (I) − e0 (I) + ℓA (A/I) = 3
by Corollary 2.10. If ℓA (I 2 /QI) = 1 then we have depthG ≥ d − 1 by [RV1, W]. If
ℓA (I 2 /QI) = 3 then the equality e1 (I) = e0 (I) − ℓA (A/I) + ℓA (I 2 /QI) holds true, so
that I 3 = QI 2 and the associated graded ring G of I is Cohen-Macaulay by Corollary
2.10. Thus we need to consider the following.
Theorem 4.6. Suppose that d ≥ 2. Assume that I is integrally closed and e1 (I) =
e0 (I) − ℓA (A/I) + 3 and ℓA (I 2 /QI) = 2. Let c = ℓA (I 3 /QI 2 ). Then the following
assertions hold true.
(1) Either C ∼
= B(−2) as graded T -modules or there exists an exact sequence
0 → B(−3) → B(−2) ⊕ B(−2) → C → 0
of graded T -modules.
(2) 1 ≤ c ≤ 2 and I 4 = QI 3 .
(3) Suppose c = 1 then depthG ≥ d − 1 and e2 (I) = 4, e3 (I) = 1
ei (I) = 0 for 4 ≤ i ≤ d.
(4) Suppose c = 2 then depthG = d − 2 and e2 (I) = 3, e3 (I) =
e4 (I) = −1 if d ≥ 4, and ei (I) = 0 for 5 ≤ i ≤ d.
(5) The Hilbert series HSI (z) is given by
ℓA (A/I) + {e0 (I) − ℓA (A/I) − 2}z + z 2 + z 3
,
if
(1 − z)d
HSI (z) =
ℓA (A/I) + {e0 (I) − ℓA (A/I) − 2}z + 3z 3 − z 4
if
(1 − z)d
if d ≥ 3, and
−1 if d ≥ 3,
c = 1,
c = 2.
Proof. Since ℓA (I 2 /QI) = 2 and e1 (I) = e0 (I) − ℓA (A/I) + 3, we have e1 (I) = e0 (I) −
ℓA (A/I) + ℓA (I 2 /QI) + 1. We also have 1 ≤ ℓA (I 3 /QI 2 ) ≤ 2 (see the proof of [RV2,
Proposition 2.1 and 2.2]). Then, thanks to Theorem 1.2, C ∼
=
= X1 B(−1) or C ∼
(X1 , X2 )B(−1) as graded T -modules, where X1 , X2 denote the linearly independent
linear forms of B. Thus all assertions follow by Theorem 1.2.
We remark that ℓA (I 2 /QI) measures how far is the multiplicity of I from the minimal
value, see [RV3, Corollary 2.1]. If ℓA (I 2 /QI) ≤ 1, then depth G ≥ d − 1, but it is still
open the problem whether depthG ≥ d − 2, assuming ℓA (I 2 /QI) = 2. Theorem 4.6
confirms the conjectured bound.
Corollary 4.7. Assume that I is integrally closed. Suppose that e1 (I) = e0 (I) −
ℓA (A/I) + 3. Then depth G ≥ d − 2.
Proof. We have 0 < ℓA (I 2 /QI) ≤ e1 (I) − e0 (I) + ℓA (A/I) = 3 by Corollary 2.10. If
ℓA (I 2 /QI) = 1 or ℓA (I 2 /QI) = 3 then we have depthG ≥ d − 1 as above. Suppose that
ℓA (I 2 /QI) = 2 then we have depthG ≥ d − 2 by Theorem 4.6 (3), (4). This completes
a proof of Corollary 4.7.
17
5. An Example
The goal of this section is to construct an example of Cohen-Macaulay local ring with
the maximal ideal m satisfying the equality in Theorem 1.2 (1). The class of examples
we exhibit includes an interesting example given by H.-J. Wang, see [RV3, Example
3.2].
Theorem 5.1. Let d ≥ c ≥ 1 be integers. Then there exists a Cohen-Macaulay local
ring (A, m) such that
d = dim A, e1 (m) = e0 (m) + ℓA (m2 /Qm), and c = ℓA (m3 /Qm2 )
for some minimal reduction Q = (a1 , a2 , · · · , ad ) of m.
To construct necessary examples we may assume that c = d. In fact, suppose
that 0 < c < d and assume that we have already chosen a certain Cohen-Macaulay
local ring (A0 , m0 ) such that c = dim A0 , e1 (m0 ) = e0 (m0 ) + ℓA0 (m20 /Q0 m0 ), and
c = ℓA0 (m30 /Q0 m20 ) with Q0 = (a1 , a2 , · · · , ac )A0 a minimal reduction of m0 . Let
n = d − c and let A = A0 [[X1 , X2 , · · · , Xn ]] be the formal power series ring over
the ring A0 . We set m = m0 A + (X1 , X2 , · · · , Xn )A and Q = Q0 A + (X1 , X2 , · · · , Xn )A.
Then A is a Cohen-Macaulay local ring with dim A = d and the maximal ideal
m = m0 A + (X1 , X2 , · · · , Xn )A. The ideal Q is a reduction of m and because
X1 , X2 , · · · , Xn forms a super regular sequence in A with respect to m (recall that
G(m) = G(m0 )[Y1 , Y2 , · · · , Yn ] is the polynomial ring, where Yi ’s are the initial forms
of Xi ’s), we have ei (m) = ei (m0 ) for i = 0, 1, m2 /Qm ∼
=
= m20 /Q0 m0 , and m3 /Qm2 ∼
2
2
3
m0 /Q0 m0 . Thus we have e1 (m) = e0 (m) + ℓA (m /Qm) and ℓA (m3 /Qm2 ) = c. This
observation allows us to concentrate our attention on the case where c = d.
Let m ≥ 0 and d ≥ 1 be integers. Let
D = k[[{Xj }1≤j≤m , Y, {Vi }1≤i≤d , {Zi }1≤i≤d ]]
be the formal power series ring with m + 2d + 1 indeterminates over an infinite field k,
and let
a = [(Xj | 1 ≤ j ≤ m) + (Y )] · [(Xj | 1 ≤ j ≤ m) + (Y ) + (Vi | 1 ≤ i ≤ d)]
+(Vi Vj | 1 ≤ i, j ≤ d, i 6= j) + (Vi3 − Zi Y | 1 ≤ i ≤ d).
We set A = D/a and denote
√ the images of Xj , Y , Vi , and Zi in A by xj , y, vi , and ai ,
respectively. Then, since a = (Xj | 1 ≤ j ≤ m) + (Y ) + (Vi | 1 ≤ i ≤ d), we have
dim A = d. Let m = (xj | 1 ≤ j ≤ m) + (y) + (vi | 1 ≤ i ≤ d) + (ai | 1 ≤ i ≤ d) be the
maximal ideal in A and we set Q = (ai | 1 ≤ i ≤ d). Then, m2 = Qm + (vi2 | 1 ≤ i ≤ d),
m3 = Qm2 + Qy, and m4 = Qm3 . Therefore Q is a minimal reduction of m, and
a1 , a2 , · · · , ad is a system of parameters for A.
We are now interested in the Hilbert coefficients ei (m)′ of the maximal ideal m as
well as the structure of the associated graded ring G(m) and the module CQ (m) of m.
Theorem 5.2. The following assertions hold true.
(1) A is a Cohen-Macaulay local ring with dim A = d.
(2) CQ (m) ∼
= B+ (−1) as graded T -modules. Therefore, ℓA (m3 /Qm2 ) = d.
(3) e0 (m) = m + 2d + 2, e1 (m) = m + 3d + 2.
(4) e2 (m) = d + 1 if d ≥ 2, and ei (m) = 0 for all 3 ≤ i ≤ d.
(5) G(m) is a Buchsbaum ring with depthG(m) = 0 and I(G(m)) = d.
18
(6) The Hilbert series HSm (z) of A is given by
P
j−1
1 + {m + d + 1}z + d+2
j=3 (−1)
HSm (z) =
(1 − z)d
d+1
j−1
zj
.
Notice that Wang’s example before quoted corresponds to the particular case m = 0
and d = 2.
Let us divide the proof of Theorem 5.2 into two steps. Let us begin with the following.
Proposition 5.3. Let p = (Xj | 1 ≤ j ≤ m) + (Y ) + (Vi | 1 ≤ i ≤ d) in D. Then
ℓDp (Ap ) = m + 2d + 2.
e = D[{ 1 }1≤i≤d ]. We set X ′ =
Proof. Let e
k = k[{Zi }1≤i≤d , { Z1i }1≤i≤d ] and D
j
Zi
1 ≤ j ≤ m, Vi′ = ZV1i for 1 ≤ i ≤ d, and Y ′ = ZY1 . Then we have
Xj
Z1
for
e =e
D
k[{Xj′ | 1 ≤ j ≤ m}, Y ′ , {Vi′ | 1 ≤ i ≤ d}],
e = [(Xj′ | 1 ≤ j ≤ m) + (Y ′ )] · [(Xj′ | 1 ≤ j ≤ m) + (Y ′ ) + (Vi′ | 1 ≤ i ≤ d)]
aD
Z12 ′ 3
V − Y ′ | 1 ≤ i ≤ d),
Zi i
are algebraically independent over e
k. Let
+(Vi′ Vj′ | 1 ≤ i, j ≤ d, i 6= j) + (
and {Xj′ }1≤j≤m , Y ′ , and {Vi′ }1≤i≤d
W =e
k[{Xj′ | 1 ≤ j ≤ m}, {Vi′ | 1 ≤ i ≤ d}]
e and
in D
3
b = [(Xj′ | 1 ≤ j ≤ m) + (V1′ )] · [(Xj′ | 1 ≤ j ≤ m) + (Vi′ | 1 ≤ i ≤ d)]
+(Vi′ Vj′ | 1 ≤ i, j ≤ d, i 6= j) + (
Z12 ′ 3
3
Vi − Z1 V1′ | 2 ≤ i ≤ d)
Zi
e we get the isomorphism
in W . Then substituting Y ′ with Z1 V1′ 3 in D,
e U
e∼
D/a
= W/b
e D
e corresponds to the prime ideal P/b of W/b,
of e
k algebras. Then the prime ideal pD/a
′
′
where P = W+ = (Xj | 1 ≤ j ≤ m) + (Vi | 1 ≤ i ≤ d). Then because
3
b + (V1′ ) = (Xj′ | 1 ≤ j ≤ m) · [(Xj′ | 1 ≤ j ≤ m) + (Vi′ | 1 ≤ i ≤ d)]
3
+(Vi′ Vj′ | 1 ≤ i, j ≤ d, i 6= j) + (Vi′ | 1 ≤ i ≤ d)
and ℓWP ([b + (V1′ 3 )]WP /bWP ) = 1, we get
3
3
ℓWP /bWP (WP /bWP ) = ℓWP (WP /[b + (V1′ )]WP ) + ℓWP ([b + (V1′ )]WP /bWP )
= (m + 2d + 1) + 1 = m + 2d + 2.
Thus ℓAp (Ap ) = ℓWP /bWP (WP /bWP ) = m + 2d + 2.
19
Thanks to the associative formula of multiplicity, we have
A /pA
e0 (Q) = ℓAp (Ap ) · e0 p p ([Q + pA]/pA) = m + 2d + 2,
√
because p = a and A/pA = D/p ∼
= k[[Zi | 1 ≤ i ≤ d]]. On the other hand, we have
A/Q ∼
= k[[{Xj }1≤j≤m, Y, {Vi }1≤i≤d ]]/c
where
c = ([(Xj | 1 ≤ j ≤ m) + (Y )] · [(Xj | 1 ≤ j ≤ m) + (Y ) + (Vi | 1 ≤ i ≤ d)]
+(Vi Vj | 1 ≤ i, j ≤ d, i 6= j) + (Vi 3 | 1 ≤ i ≤ d).
Therefore ℓA (A/Q) = m+2d+2. Thus e0 (Q) = ℓA (A/Q) so that A is a Cohen-Macaulay
local ring with e0 (Q) = m + 2d + 2.
Let K0 = A, K1 = m, and Kn = mn + ymn−2 for n ≥ 2, and we set K = {Kn }n≥0 .
Let ei (K) denote the ith Hilbert coefficients of the filtration K for 0 ≤ i ≤ d.
Lemma 5.4. The following assertions hold true.
(1) ℓA (K2 /m2 ) = 1 and Kn = mn for all n ≥ 3.
(2) Q ∩ K2 = QK1 and Kn+1 = QKn for all n ≥ 2. Therefore e1 (K) = e0 (K) −
ℓA (A/K1 ) + ℓA (K2 /QK1 ), e2 (K) = ℓA (K2 /QK1 ) if d ≥ 2, and ei (K) = 0 for
3 ≤ i ≤ d.
Proof. (1) Since K2 = m2 + (y), we have ℓA (K2 /m2 ) = 1. We have Kn = mn + ymn−2 =
mn for all n ≥ 3, because ym = (yvi | 1 ≤ i ≤ d) = (vi 3 | 1 ≤ i ≤ d) ⊆ m3 .
f2 . Therefore
(2) Since Kn = mn for all n ≥ 3 by assertion (1), we have K2 ⊆ m
f2 = Qm = QK1 by Remark 4.2. It is routine to check that Kn+1 = QKn
Q ∩ K2 ⊆ Q ∩ m
for all n ≥ 2. Thus e1 (K) = e0 (K) − ℓA (A/K1 ) + ℓA (K2 /QK1 ) by [GR, Theorem 2.2].
We also have e2 (K) = ℓA (K2 /QK1 ) if d ≥ 2, and ei (K) = 0 for 3 ≤ i ≤ d by [HM,
Proposition 4.6].
We prove now Theorem 5.2.
Proof of Theorem 5.2. Since Kn = mn for all n ≥ 3 by Lemma 5.4 (1), we have ei (K) =
ei (m) for 0 ≤ i ≤ d. Therefore, e1 (m) = e0 (m) + ℓA (m2 /Qm), e2 (m) = ℓA (m2 /Qm) + 1
if d ≥ 2, and ei (m) = 0 for all 3 ≤ i ≤ d, because ℓA (K2 /m2 ) = 1, e1 (K) = e0 (K) +
ℓA (K2 /QK1 ) −1, e2 (K) = ℓA (K2 /QK1 ) if d ≥ 2, and ei (K) = 0 for 3 ≤ i ≤ d by Lemma
5.4. Then we have e1 (m) = m + 3d + 2 and e2 (m) = d + 1 because e0 (m) = m + 2d + 2
and ℓA (m2 /Qm) = d. The ring G(m) is Buchsbaum ring with depthG(m) = 0 and
I(G(m)) = d by Theorem 4.3. This completes the proof of Theorem 5.2.
References
[EV] J. Elias and G. Valla, Rigid Hilbert functions, J. Pure and Appl. Algebra 71 (1991) 19–41.
[GNO] S. Goto, K. Nishida, and K. Ozeki, Sally modules of rank one, Michigan Math. J. 57 (2008)
359–381.
[GR] A. Guerrieri and M. E. Rossi, Hilbert coefficients of Hilbert filtrations, J. Algebra 199 (1998)
40–61.
[H]
C. Huneke, Hilbert functions and symbolic powers, Michigan Math. J. 34 (1987) 293–318.
[HM] S. Huckaba and T. Marley, Hilbert coefficients and the depth of associated graded rings, J.
London Math. Soc. (2) 56 (1997) 64–76.
20
[I1]
[I2]
[N]
S. Itoh, Hilbert coefficients of integrally closed ideals, J. Algebra 176 (1995) 638–652.
S. Itoh, Integral closures of ideals generated by regular sequence, J. Algebra 117 (1988) 390–401.
D. G. Northcott, A note on the coefficients of the abstract Hilbert function, J. London Math.
Soc. 35 (1960) 209–214.
[RR] L. J. Ratliff and D. Rush, Two notes on reductions of ideals, Indiana Univ. Math. J. 27 (1978)
929–934.
[RV1] M. E. Rossi and G. Valla A conjecture of J. Sally, Comm. Alg. 24 (13) (1996) 4249–4261.
[RV2] M. E. Rossi and G. Valla Cohen-Macaulay local rings of embedding dimension e + d − 3, Proc.
London Math. Soc. (3) 80 (2000) 107–126.
[RV3] M. E. Rossi and G. Valla Hilbert functions of filtered modules, UMI Lecture Notes 9, Springer
(2010).
[S1]
J. D. Sally, Hilbert coefficients and reduction number 2, J. Alg. Geo. and Sing. 1 (1992) 325–333.
[S2]
J. D. Sally, Ideals whose Hilbert function and Hilbert polynomial agree at n = 1, J. Algebra
157 (1993), 534–547.
[V]
W. V. Vasconcelos, Hilbert Functions, Analytic Spread, and Koszul Homology, Contemporary
Mathematics, Vol 159 (1994) 410–422.
[VP] M. Vaz Pinto, Hilbert functions and Sally modules, J. Algebra, 192 (1996) 504–523.
[W]
H.-J. Wang, Links of symbolic powers of prime ideals, Math. Z. 256 (2007) 749–756.
Department of Mathematical Science, Faculty of Science, Yamaguchi University,
1677-1 Yoshida, Yamaguchi 753-8512, Japan
E-mail address: [email protected]
Dipartimento di Matematica, Universita di Genova, Via Dodecaneso, 35-16146, Genova, Italy
E-mail address: [email protected]
21
| 0math.AC
|
Bernoulli Embeddings for Graphs
Vinith Misraα and Sumit Bhatiaβ ∗
Netflix Inc., Los Gatos, CA, USA
β
IBM India Research Laboratory, New Delhi, India
[email protected], [email protected]
arXiv:1803.09211v1 [cs.LG] 25 Mar 2018
α
March 25, 2018
Abstract
Just as semantic hashing [Salakhutdinov and Hinton2009] can accelerate information retrieval, binary
valued embeddings can significantly reduce latency in the retrieval of graphical data. We introduce
a simple but effective model for learning such binary vectors for nodes in a graph. By imagining the
embeddings as independent coin flips of varying bias, continuous optimization techniques can be applied
to the approximate expected loss. Embeddings optimized in this fashion consistently outperform the
quantization of both spectral graph embeddings and various learned real-valued embeddings, on both
ranking and pre-ranking tasks for a variety of datasets.
1
Introduction
Consider users — perhaps from the research, intelligence, or recruiting community — who seek to explore
graphical data — perhaps knowledge graphs or social networks. If the graph is small, it is reasonable
for these users to directly explore the data by examining nodes and traversing edges. For larger graphs,
or for graphs with noisy edges, it rapidly becomes necessary to algorithmically aid users. The problems
that arise in this setting are essentially those of information retrieval and recommendation for graphical
data, and are well studied [Hasan and Zaki2011, Blanco et al.2013]: identifying the most important edges,
predicting links that do not exist, and the like. The responsiveness of these retrieval systems is critical
[Gray and Boehm-Davis2000], and has driven numerous system designs in both hardware [Hong et al.2011]
and software [Low et al.2014]. An alternative is to seek algorithmic solutions to this latency challenge.
Models that perform link prediction and node retrieval can be evaluated across two axes: the relevance
of the retrieved nodes and the speed of retrieval. The gold standard in relevance is typically set by trained
models that rely on “observable” features that quantify the connectivity between two nodes, but these models
are often quite slow to evaluate due to the complexity of the features in question.
At the other extreme, binary-valued embeddings can accelerate the retrieval of graphical data, much in
the same manner that semantic hashing [Salakhutdinov and Hinton2009] can assist in the efficient retrieval
of text and image data. Roughly speaking, having a binary representation for each node allows one to search
for similar nodes in constant time directly in the binary embedding space — much faster than the alternatives
(Table 1).
The challenge is that efficient binary representations can be difficult to learn: for any reasonable metric of
accuracy, finding optimal binary representations is NP-hard. One solution is to lean on the large body of work
around learning continuous embeddings for graphs, and utilize modern quantization techniques to binarize
these continuous representations. The “catch” with this approach is that the continuous embeddings are not
optimized with their future binarization in mind, and this hurts the relevance of retrieved nodes (Sec. 4.4).
∗ Part
of this work was conducted while both the authors were at IBM Almaden Research Center.
1
Technique
Preprocess
Query
Observable features
Real embeddings
Binary embeddings
(sparse similarities)
Real embeddings
(quantized)
Binary embeddings
(dense similarities)
O(1) (slow)
O(ED)
O(N ) (slow)
O(N ) (fast)
O(ED)
O(1)
O(ED)
O(1)
O(N 2 )
O(1)
Table 1: Complexity of five different node retrieval approaches, ranked from highest to lowest accuracy (Table
2). N nodes, E edges, and D latent dimensions.
Our primary contribution, thus, is an end-to-end method for learning embeddings that are explicitly
optimized with both binarization and their use in link prediction/node retrieval in mind. More concretely: in a
manner similar to Skip-gram [Mikolov et al.2013], the likelihood of an edge between two nodes is modeled as
a function of the Hamming distance between their bit embeddings. Rather than directly optimizing the bit
embeddings eij for this task — an NP-hard task [Weiss et al.2009] — they are instead imagined as being
drawn from a matrix of independent Bernoulli random variables Eij , parametrized by their (independent)
probabilities of success pij . By minimizing expected loss over this (product) distribution of embeddings, and
by applying efficient approximations to the Hamming distance (Sec. 3.4), continuous optimization techniques
can be applied. For convenience, we refer to bit embeddings learned in this manner as Bernoulli embeddings.
Comparisons performed on five different graphical datasets are described in Section 4. Bernoulli embeddings
are found to achieve significantly higher test-set mean average precision than a variety of alternative
binary embedding options, including various quantizations of DeepWalk vectors [Perozzi et al.2014], Fiedler
embeddings [Hendrickson2007], and several other real-valued embeddings that we ourselves introduce (Table 2).
This is also found to hold for the reranking scenario, where binary hashes are used as a preprocessing step
to accelerate more computationally intensive algorithms. Further, node retrieval performed using binary
embeddings is orders of magnitude faster than other alternatives, especially for larger datasets (Table 4).
2
Related Work
Approaches to node retrieval, roughly categorized in Table 1, can be evaluated in terms of both relevance
and speed. Bernoulli embeddings occupy an unexplored but valuable niche in this spectrum: they are binary
embeddings (O(1) retrieval) appropriate for large graphs (more than a few thousand nodes) that are learned
directly from the adjacency matrix (higher relevance of retrieved nodes). In the following, we describe the
other categories represented in Table 1.
2.1
Observable features
Methods for link prediction and node retrieval on graphs typically rely on observable neighborhood features [Dong et al.2014]. However, computing node-node similarity using these tools can have a significant
computational cost [Low et al.2014].
Second order neighborhood features, such as the Jaccard index [Hasan and Zaki2011] and the AdamicAdar (AA) score [Adamic and Adar2003] either implicitly or explicitly involve operations over length-2
paths and degree-2 neighborhoods. This generally requires either one or more joins with the graph table
or multiplications with the graph’s adjacency matrix. Even with efficient sparsity-exploiting indexing, this
operation has complexity O(E) in the number of edges E in the graph.
Higher order path features, such as the Katz metric, rooted PageRank [Hasan and Zaki2011], and the
regression-based Path Ranking Algorithm [Dong et al.2014] involve even longer paths and even more such oper2
ations. State-of-the-art link prediction often harnesses dozens of such features in parallel [Cukierski et al.2011].
Offline precomputation of the (dense) node similarity matrix can dramatically help with latency, but its
quadratic complexity in the number of nodes leaves it an option only for smaller graphs.
2.2
Real-valued Embeddings
With unquantized real-valued embeddings (i.e. no use of LSH), node retrieval typically involves a brute-force
O(N ) search for the nearest Euclidean neighbors to a query. While such embeddings have appeared most
prominently in the context of text for reasons unrelated to retrieval, Perozzi et al [Perozzi et al.2014] apply
the word2vec machinery of Mikolov et al [Mikolov et al.2013] to “sentences” generated by random walks
on a graph, and Yang et al [Yang et al.2015] illustrate that a graph embedding can be a powerful tool in
the context of semisupervised learning. The world of knowledge graphs has been particularly welcoming to
embeddings, starting with the work of Hinton [Hinton1986], and continuing with the models of Bordes et
al. [Bordes et al.2011], Sutskever et al. [Sutskever et al.2009], Socher et al. [Socher et al.2013] and others.
Additionally, graph Laplacian eigenvectors, which find prominent use in spectral clustering, can be
interpreted as a latent embedding (“Fiedler embedding” [Hendrickson2007]) analogous to LSA. Knowledge
graphs, which are more naturally represented with tensors than with adjacency matrices, analogously suggest
the use of tensor factorizations and approximate factorizations [Nickel et al.2012].
2.3
Discrete Embeddings
Hinton and Salakhudtinov [Salakhutdinov and Hinton2009] introduce semantic hashing as the solution to a
very similar problem in a different domain. Instead of relying on indexed TF-IDF vectors for the retrieval of
relevant documents, a discrete embedding is learned for every document in the corpus. At query time, a user
can rapidly retrieve a shortlist of relevant documents simply by scanning the query’s neighborhood in the
(discrete) embedding space. If the embedding is sufficiently compact, the neighborhood will be nonempty and
the scan will be fast. If the embedding is very compact, this retrieval yields a pre-ranked list that may be
reranked using more computationally demanding algorithms. These results have fueled the development of a
variety of similar techniques, all seeking to learn compact binary encodings for a given dataset.
Quantized Real-valued Embeddings: The most popular approach, taken by Weiss et al. [Weiss et al.2009],
Gong and Lazebnik [Gong and Lazebnik2011] and others, is to assume that the data consists of short real
vectors. To apply these algorithms to graphical data, one must first learn a real-valued embedding for the
nodes of the graph. We compare against these baselines in Sec. 4.
Binary Embeddings from Similarity Matrix: In this approach, also known as “supervised hashing” ’ [Liu et al.2012, Kulis and Grauman2012, Norouzi et al.2012], a matrix of pairwise similarities between
all data points is supplied. The objective is to preserve these similarities in the discrete embedding space.
On the surface, this appears very similar to the graphical setting of interest to us. However, no sparsity
assumption is placed on the similarity matrix. As such, proposed solutions (typically, variations of coordinate
descent) fall victim to an O(N 2 ) complexity, and application is limited to graphs with a few thousand nodes.
Note that an embedding-learning approach specifically avoids this issue by exploiting the sparse structure of
most graphs.
Liu et al. [Liu et al.2014] also assume that one is supplied a matrix of similarities for the data. Rather
than directly attempting to replicate these similarities in the embedded space, they perform a constrained
optimization that forces the embedded bits to be uncorrelated and zero-mean. In the graph setting, this
acts as an approximation to the sign bits of the Fiedler embedding, which appears amongst our empirical
baselines.
3
3
3.1
Architecture
Bernoulli Embeddings
We consider a generic graph, consisting of N nodes X , {1, 2, . . . , N } and a binary-valued adjacency matrix
G ∈ {0, 1}N ×N .
The goal is formulated as learning a matrix of probabilities p = {pij : 1 ≤ i ≤ N , 1 ≤ j ≤ d}, from which
the node embeddings E are sampled as a matrix of independent Bernoulli random variables Eij ∼ β(pij ). For
convenience, one may reparameterize the embeddings as Eij = 1(pij > Θij ), where ΘN ×d is a matrix of iid
random thresholds distributed uniformly over the interval [0, 1].
3.2
Model and Objective
Recall the use case for short binary codes: to retrieve a shortlist of nodes Y ∈ X m similar to a query node
x ∈ X , one should merely have to look up entries indexed at nearby locations in the embedding space. As
such, we seek a model where the Hamming distance between embeddings monotonically reflects the likelihood
of a connection between the nodes.
For real-valued embeddings, a natural and simple choice — used for instance with Skip-gram [Mikolov et al.2013]
— is to treat the conditional link probability between nodes i and j as a softmax-normalized cosine distance
between embeddings:
exp (Ei · Ej )
P (j|i; p) = P|X |
(1)
= softmaxj EiT E .
k=1 exp (Ei · Ek )
To translate this to the setting of binary vectors, it is natural to substitute Hamming distance dH (Ei , Ej ) for
the cosine distance Ei · Ej in (1). As the distance dH is limited to taking values in the set {0, 1, . . . , d}, a
transformation is required. Empirically, we find that more complex transformations are unnecessary for the
purpose of learning a good embedding1 , and a simple linear scaling suffices
P (j|i; p) = softmaxj adH (EiT , E) ,
(2)
where dH (x, y) = xT (1 − y) + (1 − x)T y represents the Hamming distance, and a is the (potentially negative)
scaling parameter.
Given the model of (2), we seek to maximize the expected log likelihood of the observed graph edges:
X
−E log softmaxj aEiT E Θ .
L(G; p) =
(3)
(i,j)∈G
The expression in (3) unfortunately introduces two obstacles: (1) the softmax, which requires a summation
over all candidate nodes j, and (2) the expectation of a discrete-valued functional, which presents difficulties for
optimization. The first is addressed by means of noise contrastive estimation [Gutmann and Hyvärinen2010],
detailed in Sec. 3.3. The second is addressed via several approximation techniques, detailed in Sec. 3.4.
3.3
Noise Contrastive Estimation (NCE)
To sidestep the softmax summation, we follow in the steps of Mnih and Kavukcuoglu [Mnih and Kavukcuoglu2013]
and employ NCE [Gutmann and Hyvärinen2010], whose minimum coincides with that of (3). Specifically,
softmax normalization is replaced with a learnable parameter b, and one instead optimizes for the model’s
1 More specifically: while a complex choice of transformation can improve the optimization objective (4), we find no consistent
or significant improvement to the test set precision-recall metrics reported in Sec. 4.4. Essentially, beyond a point, parametrization
of the mapping appears to improve the probabilities produced in a forward pass through the network, but does not noticeably
improve the embedding-parameter-gradients it returns in the backwards pass.
4
effectiveness at distinguishing a true data point (i, j) ∈ G from randomly generated noise (i, Kij ). This
objective is given by
"
#
X
eadH (Ei ,Ej )+b
pK (Kij |i)
L(G) =
−E log ad (E ,E )+b
+ log ad (E ,E )+b
(4)
,
e H i j
+ pK (j|i)
e H i Kij
+ pK (Kij |i)
(i,j)∈G
θ
where Kij is a negative sample drawn from the conditional noise distribution pK (·|i).
Gutmann and Hyvärinen [Gutmann and Hyvärinen2010] argue that one should choose the noise distribution to resemble the data distribution as closely as possible. We experiment with distributions ranging
from powers of the unigram, to distributions over a node’s 2nd-degree neighborhood (in accordance with
the locally closed world assumption of Dong et al. [Dong et al.2014]), to mixtures thereof, to curricula that
transition from easily identified noise to more complex noise models. Empirically, we find that none of these
techniques outperform the uniform noise distribution either significantly or consistently.
3.4
Approximation of objective
The objective function in (4) presents a challenge to gradient-based optimization. The expectation over Θ is
difficult to evaluate analytically, and because the argument to the expectation is discrete, the reparameterization trick [Kingma and Welling2013] does not help. We introduce two continuous approximations to the
discrete random variable DH (Ei , Ej ) that maneuver around this difficulty.
First, according to the independent-Bernoulli model for the embedding matrix Eij = 1(pij > Θij ), the
normalized Hamming distance between two embeddings is the mean of d independent (but not identically
distributed) Bernoullis F1 , . . . , Fd :
d
d
l=1
l=1
1
1X
1X
DH (Ei , Ej ) =
Eil (1 − Ejl ) + (1 − Eil )Ejl =
Fl .
d
d
d
By Kolmogorov’s strong law [Sen and Singer1994], this quantity converges with d almost surely to its
expectation. Therefore, for sufficiently large d, the Θ-expectation in (4) is approximated by
Lmean (G) =
X
(i,j)∈G
− log
pK (Kij |i)
eadH (pi ,pj )+b
− log ad (p ,p )+b
,
ad
(p
,p
)+b
i
j
i
H
H
K
e
+ pK (Kij |i)
e
+ pK (j|i)
(5)
which is amenable to gradient-based optimization.
While the approximation of (5) is accurate for larger dimensionality d, recall that our goal is to learn
short binary codes. A sharper approximation is possible for smaller d by means of the central limit theorem
as follows.
1
2
DH (Ei , Ej ) ≈ N µij = DH (pi , pj ),σij
d
(6)
!
d
X
=
DH (pik , pjk )(1 − DH (pik , pjk )) .
k=1
Applying the reparametrization trick [Kingma and Welling2013], our objective takes the form
LCLT (G) =
X
(i,j)∈G
−E log
ea(µij +σij Z)+b
ea(µij +σij Z)+b + pK (j|i)
pK (Kij |i)
+ log a(µ +σ Z)+b
e ij ij
+ pK (Kij |i)
5
(7)
,
Z
where Z is a zero mean and unit variance Gaussian. Observe that the argument to the expectation is now
differentiable with respect to the parameters (p, a, b), as is the case with (5).
A common approach [Kingma and Welling2013] to optimizing an expected-reparameterized loss such
as LCLT is to use Monte Carlo integration to approximate the gradient over N samples of noise Z. This
yields an unbiased estimate for the gradient with variance that converges O N1 . However, Monte Carlo
integration is generally appropriate for approximating higher dimensional integrals. For a single dimensional
integral, numerical quadrature
— while deterministic and therefore biased — can have the significantly faster
error convergence of O N14 (midpoint rule). For small N , accuracy can be further improved by performing
quadrature with respect to the Gaussian CDF Φz . Letting f denote the intra-expectation computation in (7),
E [f (µij + σij Z)]Z =
Z
1
f (µij + σij z)dΦz
0
N
2n − 1
1 X
−1
≈
f µij + σij Φz
.
2N n=1
2N
(8)
Figure 1 compares the mean approximation with the quadrature normal approximation over the range of
embedding dimensionalities we consider. For smaller embedding dimensionalities, the greater accuracy of the
quadrature approximation leads to a lower test set log loss.
Figure 1: Comparison of Bernoulli models optimized for Lmean and for LCLR (N = 5 samples) on the KG
dataset. Left: Error (absolute) between approximate loss and true loss. Right: True loss.
3.5
Optimization and Discretization
The training set loss, as given by LCLT and approximated by (8), is minimized with stochastic gradient
descent with the diagonalized AdaGrad update rule [Duchi et al.2011]. To generate a discrete embedding
E = 1(p < Θ) from a Bernoulli matrix p, each entry is rounded to 0 or 1 (i.e. Θ = 1/2), in accordance with
maximum likelihood.
6
4
Experiments
4.1
Datasets
Results are evaluated on five datasets. Five percent of the edges of each dataset are held out for the test set,
and the remainder are used in training.
KG (115K entities, 1.3M directed edges)2 is a knowledge graph extracted from the Wikipedia corpus
using statistical relation extraction software [Castelli et al.2012]. Edges are filtered to those with more than
one supporting location in the text, and both entity and relation types are ignored. Despite this filtration, KG
possesses a large number of spurious edges and entities. Such a noisy dataset is representative of automatically
constructed knowledge graphs commonly encountered in enterprise settings [Bhatia et al.2016].
Wordnet (82K entities, 232K directed edges) is a comparatively low-noise, manually constructed graph
consisting of the noun-to-noun relations in the Wordnet dataset [Miller1995]. As with KG, we ignore the edge
type attribute.
Slashdot (82K entities, 948K directed edges), Flickr (81K entities, 5.9M undirected edges), and BlogCatalog (10K entities, 334K undirected edges) are standard social graph datasets consisting of links between
users of the respective websites [Leskovec et al.2009, Tang and Liu2009].
4.2
Baselines and Comparisons
While there is little work specifically in obtaining bit-valued embeddings for graphs, we compare Bernoulli
embeddings against quantizations of three classes of real-valued embeddings.
B1: Fiedler embeddings [Hendrickson2007] are computed using the unnormalized graph Laplacian,
symmetric graph Laplacian, and random-walk graph Laplacian.
B2: DeepWalk embeddings are computed using the Skip-gram-inspired model of Perozzi et al [Perozzi et al.2014].
B3: Real-valued distance embeddings (DistEmb) are obtained by modifying the Bernoulli embedding objective to predict link probabilities from the Hamming, `1 , `2 , or cosine distance between real-valued
embedded vectors. Note that the Hamming distance case is equivalent to using the Bernoulli probabilities p
as embeddings, and the cosine distance variety can be interpreted as DeepWalk modified with NCE and a
window size of 2.
Three different quantizations of the above embeddings are computed. Random-hyperplane LSH [Charikar2002]
is selected due to its explicit goal of representing cosine similarity — used by both the DeepWalk and the
cosine-distance variety of DistEmb. Spectral Hashing (SH) is chosen for its similarity in objective to Fiedler
embeddings. Iterative Quantization (IQ), another data-driven embedding, has been found to outperform SH
on several datasets [Gong and Lazebnik2011], and as such we consider it as well.
B4: Observable predictor. Additionally, for use in re-ranking,
we perform logistic regression with
P
several observable neighborhood features of the form s(x, y) = z∈Γ(x)∩Γ(y) f (Γ(z)), where Γ(x) indicates
the degree-1 neighborhood of x. Specifically, we compute the number of common neighbors (f (x) = x), the
Adamic-Adar (AA) score (f (x) = 1/log(x)), variations of AA (f (x) = x−0.5 , x−0.3 ), and transformations
T (s) = [s, log(s + 1), s0.5 , s0.3 , s2 ] of each score. Despite being far from state-of-the-art, we find that this
predictor can significantly improve performance when reranking results produced with binary embeddings.
Both 10- and 25-dimensional embeddings are trained: for the scale of graphs we consider, the (sparsely
populated) latter is useful for instant-retrieval via semantic-hashing, while the (densely populated) former is
useful for reranking. Furthermore, we find that the quantizations of the real embeddings B1-B3 perform best
when highly-informative 100 and 200 dimensional Fiedler/DeepWalk/DistEmb embeddings are quantized
down to the 10 and 25 bit vectors that are sought.
4.3
Evaluation metrics
Each of the methods considered is likely to excel in the metric it is optimized for: Bernoulli embeddings
for expected log loss, DeepWalk for Skip-gram context prediction, etc. Our interest, however, lies in node
2 http://sumitbhatia.net/source/datasets.html
7
retrieval, and more specifically in the ranked list of nodes returned by each algorithm. Mean Average Precision
(MAP) is a commonly used and relatively neutral criterion appropriate for this task.
A subtlety, however, lies in the choice of set on which to evaluate MAP. Document retrieval algorithms
commonly evaluate precision and recall on documents in the training set [Salakhutdinov and Hinton2009].
This does not lead to overfitting, as the algorithms typically only make use of the training set documents
and not their labeled categories/similarities. Embedding-based link-prediction algorithms, however, explicitly
learn from the labeled similarity information, as represented by the edge list / adjacency matrix. Alternatively
stated: rather than extrapolating similarities from input text, the goal is to generalize and predict additional
similarities from those that have already been observed.
As such, we measure generalization via “test set precision”: for a query node, a retrieved node is only
judged as relevant if an edge between it and the query appears in the test set. Observe that this is much
smaller than typically reported MAP, as all edges appearing in the training set are judged non-relevant. More
specifically, for small test sets, its value can rarely be expected to exceed the inverse of the average degree.
Furthermore, in reporting observed results, scores corresponding to “DistEmb”, “DeepWalk”, and “Fiedler”
embeddings are the best observed test set score amongst all variations of quantizer type (SH, LSH, and ITQ),
graph laplacian type (unnormalized, symmetric, random walk), and distance embedding (Hamming and `2 3 ).
These optimizations almost certainly represent test set overfitting, and present a challenging baseline for the
Bernoulli embeddings.
0.05
0.18
Bernoulli
0.14
Deepwalk (LSH)
Precision
Precision
Fiedler (IQ)
0.03
Observable
Bernoulli
0.16
DistEmb (`2,SH)
0.04
0.02
DistEmb (`2,SH)
0.12
Fiedler (IQ)
0.10
Deepwalk (IQ)
0.08
0.06
0.04
0.01
0.02
0.00
0.0
0.2
0.4
0.6
0.8
0.00
0.0
1.0
Recall
0.2
0.4
0.6
0.8
1.0
Recall
Figure 2: Left: 25-bit Ranking. Right: 10-bit Reranking. Mean precision/recall of binary embeddings on
the Flickr test set, averaged over 1000 random queries. Real embeddings are quantized to 25 bits from 100
dimensions, and only their highest test set MAP parameterizations are shown.
4.4
Empirical results
In the case of directly retrieving results from binary embeddings, Bernoulli embeddings are found to significantly
outperform the various alternative binary embeddings (Fig. 2 and Table 2). It is also interesting to compare
to the unquantized real embeddings (last four rows of Table 2). Despite their informational disadvantage,
Bernoulli embeddings are competitive.
On most datasets, the observable-feature predictor achieves significantly higher MAP than any of the
latent embedding models — real or binary. This is in line with our expectations, and one can in fact
expect even better such results from a state-of-the-art link predictor. To harness this predictive power
without the computational expense of computing observable features, a compelling option is to re-rank the
3 `1
and cos similarity are consistently outperformed.
8
Dataset
d (Dimensionality)
Q-DistEmb
Q-DeepWalk
Q-Fiedler
Bernoulli
Q-DistEmb
Reranked, Binary Q-DeepWalk
Q-Fiedler
Ei ∈ {0, 1}d
Bernoulli
Observables
DistEmb
Ranked, Real
Fiedler
d
Ei ∈ R
DeepWalk
Ranked, Binary
Ei ∈ {0, 1}d
KG
10
25
.0016
.0047
.0004
.0004
.0011
.0026
.0042 .0112
.0122
.0179
.0049
.0060
.0091
.0129
.0249 .0267
.0865
.0216
.0254
.0085
.0086
.0022
.0021
Wordnet
10
25
.0021
.0212
.0011
.0012
.0001
.0004
.0054 .1013
.0074
.0178
.0031
.0030
.0019
.0024
.0101 .0514
.0126
.0195
.1227
.0020
.0016
.0761
.0777
Slashdot
10
25
.0014
.0213
.0002
.0004
.0009
.0216
.0016 .0298
.0471 .0493
.0081
.0189
.0420
.0402
.0516 .0487
.0898
.0313
.0350
.0269
.0270
.0326
.0326
Flickr
10
25
.0051
.0054
.0014
.0017
.0021
.0039
.0087 .0161
.0092
.0084
.0056
.0112
.0044
.0074
.0240 .0487
.0349
.0196
.0256
.0065
.0077
.0037
.0037
BlogCatalog
10
25
.0073
.0094
.0048
.0053
.0065
.0080
.0131 .0181
.0098
.0198
.0196 .0210
.0139
.0172
.0181
.0241
.0524
.0255
.0282
.0118
.0116
.0074
.0074
Table 2: Test set MAP at both 10 and 25 dimensions. For DistEmb, DeepWalk, Fiedler, and quantizations
Q-* we present the maximum MAP across Laplacian varieties, choice of DistEmb distance function, and
choice of quantizer. Bold indicates category leaders.
highest-confidence nodes retrieved by binary embeddings [Salakhutdinov and Hinton2009]. Observe the two
computational bottlenecks at play here: searching in the embedded space to populate the pre-ranked list of
entities, and computing observable features for each of the pre-ranked entities.
We re-rank under constraints on both of these operations: no more than 10000 locations in the embedded
space may be searched, and no more than 1000 nodes can be subject to re-ranking. As illustrated in Fig. 2
and documented in the second four rows of Table 2, Bernoulli embeddings are again found to outperform the
alternatives.
Finally, note that while both the Bernoulli objective function and the test-set MAP evaluation center around
the prediction of unknown links, generalization with those metrics also translates into a more qualitatively
meaningful ranking of known links. Table 3 illustrates this for the KG dataset. Notice that the observed
quality of the retrieved entities roughly reflects the MAP scores of the respective predictors/embeddings.
Query
"New Delhi"
"Roger Federer"
"Bruce Wayne"
Rank
1st
2nd
3rd
1st
2nd
3rd
1st
2nd
3rd
DistEmb (`2 ,SH)
"Ministry of Education"
"Indonesian"
"Poet"
"Gael Monfils"
"Third Round"
"Second Round"
"Dick Grayson"
"Damian Wayne"
"Gotham"
Bernoulli
"Delhi"
"Mumbai"
"British India"
"Grand Slam Final"
"Maria Sharapova"
"Rafael Nadal"
"Joker"
"Batman"
"Arkham Asylum"
Bernoulli Reranked
"Delhi"
"India"
"Indian"
"Rafael Nadal"
"French Open"
"Novak Djokovic"
"Batman"
"Robin"
"Joker"
Observable
"Delhi"
"India"
"Alumni"
"Rafael Nadal"
"French Open"
"Wimbledon"
"Batman"
"Robin"
"Joker"
Table 3: Top-3 retrieved nodes (excluding the query node) in the KG dataset for several queries.
4.5
Efficiency Results
Training a 25-dimensional embedding on the KG dataset — the largest problem we consider — takes roughly
10s per epoch on a 2.2GHz Intel i7 with 16GB of RAM, and validation loss is typically minimized between 30
and 60 epochs.
Table 4 reports the average time taken by different methods to retrieve the node most similar to a query
node. By considering a d-dimensional binary embedding as an address/hash in the d-dimensional space,
retrieval reduces to a near-instant look up of the hashtable: 0.003 ms for 25-dimension embeddings (Table 4).
Note that this retrieval speed is independent of the embedding dimensionality d and the dataset size. At the
9
Method
Time (ms)
binary embeddings, hash based retrieval
ranking using observables (KG)
re-ranking using observables (KG)
binaryBruteForce (100d, 100K nodes)
binaryBruteForce (100d, 1M nodes)
binaryBruteForce (100d, 10M nodes)
binaryBruteForce (100d, 100M nodes)
realBruteForce (100d, 100K nodes)
realBruteForce (100d, 1M nodes)
realBruteForce (100d, 10M nodes)
realBruteForce (100d, 100M nodes)
0.003
2949
7.6
0.7
4.5
91.7
1147.9
23.5
181.6
2967.1
OOM
Table 4: Time taken in milliseconds by different methods for retrieving similar nodes given a query node.
Times reported are averaged over 50 runs, ran on a system running Ubuntu 14.10, with 32GB RAM and 16
Core 2.3 GHz Intel Xeon processor. OOM indicates Out of Memory.
other extreme, the high-performance observables model for the KG dataset (115K entities) takes 2949 ms
and scales linearly with the number of nodes in the dataset. Finally, the “Goldilocks” solution takes 7 ms: a
10-dimensional embedding is used to obtain a shortlist, which is then reranked using the observables model.
Hash based lookup is not the only way to exploit binary embeddings. If one replaces a brute force nearest
neighbor search amongst real valued embeddings with a brute force search amongst binary embeddings,
order-of-magnitude memory and speed advantages remain despite the O(N ) runtime: binary embeddings
take up 32 to 64 times less space than real embeddings, and Hamming distance can be computed much faster
than Euclidean distance. Table 4 illustrates this on synthetically generated embeddings for web-scale graphs
of up to 100 million nodes.
5
Conclusion
We introduce the problem of learning discrete-valued embeddings for graphs, as a graphical analog to
semantic hashing. To sidestep the difficulties in optimizing a discrete graph embedding, the problem is
reformulated as learning the continuous parameters of a distribution from which a discrete embedding is
sampled. Bernoulli embeddings correspond to the simplest such distribution, and we find that they are
computable with appropriate approximations and sampling techniques. On a variety of datasets, in addition to
being memory and time efficient, Bernoulli embeddings demonstrate significantly better test-set precision and
recall than the alternative of hashed real embeddings. This performance gap continues to hold when retrieval
with binary embeddings is followed by reranking with a more powerful (and cumbersome) link predictor. In
this latter case, precision and recall can rival or exceed that of the link predictor, while performing retrieval
in time more or less independent of the data set size.
References
[Adamic and Adar2003] Lada A. Adamic and Eytan Adar. Friends and neighbors on the web. Social networks,
25(3):211–230, 2003.
[Bhatia et al.2016] Sumit Bhatia, Alok Goel, Elizabeth Bowen, and Anshu Jain. Separating wheat from the
chaff - A relationship ranking algorithm. In The Semantic Web - ESWC 2016 Satellite Events,, pages
79–83, 2016.
10
[Blanco et al.2013] Roi Blanco, Berkant Barla Cambazoglu, Peter Mika, and Nicolas Torzec. Entity recommendations in web search. In The Semantic Web–ISWC 2013, pages 33–48. Springer, 2013.
[Bordes et al.2011] Antoine Bordes, Jason Weston, Ronan Collobert, and Yoshua Bengio. Learning structured
embeddings of knowledge bases. In Conference on Artificial Intelligence, 2011.
[Castelli et al.2012] Vittorio Castelli, Hema Raghavan, Radu Florian, Ding-Jung Han, Xiaoqiang Luo, and
Salim Roukos. Distilling and exploring nuggets from a corpus. In Proceedings of the 35th international
ACM SIGIR conference on Research and development in information retrieval, pages 1006–1006. ACM,
2012.
[Charikar2002] Moses S. Charikar. Similarity estimation techniques from rounding algorithms. In Proceedings
of the thiry-fourth annual ACM symposium on Theory of computing, pages 380–388. ACM, 2002.
[Cukierski et al.2011] William Cukierski, Benjamin Hamner, and Bo Yang. Graph-based features for supervised link prediction. In Neural Networks (IJCNN), The 2011 International Joint Conference on, pages
1237–1244. IEEE, 2011.
[Dong et al.2014] Xin Dong, Evgeniy Gabrilovich, Geremy Heitz, Wilko Horn, Ni Lao, Kevin Murphy,
Thomas Strohmann, Shaohua Sun, and Wei Zhang. Knowledge vault: A web-scale approach to probabilistic
knowledge fusion. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge
discovery and data mining, pages 601–610. ACM, 2014.
[Duchi et al.2011] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online
learning and stochastic optimization. The Journal of Machine Learning Research, 12:2121–2159, 2011.
[Gong and Lazebnik2011] Yunchao Gong and Svetlana Lazebnik. Iterative quantization: A procrustean
approach to learning binary codes. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE
Conference on, pages 817–824. IEEE, 2011.
[Gray and Boehm-Davis2000] Wayne D. Gray and Deborah A. Boehm-Davis. Milliseconds matter: An
introduction to microstrategies and to their use in describing and predicting interactive behavior. Journal
of Experimental Psychology: Applied, 6(4):322, 2000.
[Gutmann and Hyvärinen2010] Michael Gutmann and Aapo Hyvärinen. Noise-contrastive estimation: A
new estimation principle for unnormalized statistical models. In International Conference on Artificial
Intelligence and Statistics, pages 297–304, 2010.
[Hasan and Zaki2011] Mohammad Al Hasan and Mohammed J. Zaki. A Survey of Link Prediction in Social
Networks. In Social Network Data Analytics, pages 243–275. Springer US, 2011.
[Hendrickson2007] Bruce Hendrickson. Latent semantic analysis and Fiedler retrieval. Linear Algebra and its
Applications, 421(2):345–355, 2007.
[Hinton1986] Geoffrey E. Hinton. Learning distributed representations of concepts. In Proceedings of the
eighth annual conference of the cognitive science society, volume 1, page 12. Amherst, MA, 1986.
[Hong et al.2011] Sungpack Hong, Tayo Oguntebi, and Kunle Olukotun. Efficient parallel graph exploration
on multi-core CPU and GPU. In Parallel Architectures and Compilation Techniques (PACT), 2011
International Conference on, pages 78–88. IEEE, 2011.
[Kingma and Welling2013] Diederik P. Kingma and Max Welling.
arXiv:1312.6114 [cs, stat], December 2013. arXiv: 1312.6114.
Auto-Encoding Variational Bayes.
[Kulis and Grauman2012] Brian Kulis and Kristen Grauman. Kernelized locality-sensitive hashing. Pattern
Analysis and Machine Intelligence, IEEE Transactions on, 34(6):1092–1104, 2012.
11
[Leskovec et al.2009] Jure Leskovec, Kevin J. Lang, Anirban Dasgupta, and Michael W. Mahoney. Community
structure in large networks: Natural cluster sizes and the absence of large well-defined clusters. Internet
Mathematics, 6(1):29–123, 2009.
[Liu et al.2012] Wei Liu, Jun Wang, Rongrong Ji, Yu-Gang Jiang, and Shih-Fu Chang. Supervised hashing
with kernels. In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pages
2074–2081. IEEE, 2012.
[Liu et al.2014] Wei Liu, Cun Mu, Sanjiv Kumar, and Shih-Fu Chang. Discrete graph hashing. In Advances
in Neural Information Processing Systems, pages 3419–3427, 2014.
[Low et al.2014] Yucheng Low, Joseph E. Gonzalez, Aapo Kyrola, Danny Bickson, Carlos E. Guestrin,
and Joseph Hellerstein. Graphlab: A new framework for parallel machine learning. arXiv preprint
arXiv:1408.2041, 2014.
[Mikolov et al.2013] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S. Corrado, and Jeff Dean. Distributed
representations of words and phrases and their compositionality. In Advances in Neural Information
Processing Systems, pages 3111–3119, 2013.
[Miller1995] George A. Miller. WordNet: a lexical database for English. Communications of the ACM,
38(11):39–41, 1995.
[Mnih and Kavukcuoglu2013] Andriy Mnih and Koray Kavukcuoglu. Learning word embeddings efficiently
with noise-contrastive estimation. In Advances in Neural Information Processing Systems 26, pages
2265–2273. 2013.
[Nickel et al.2012] Maximilian Nickel, Volker Tresp, and Hans-Peter Kriegel. Factorizing YAGO: scalable
machine learning for linked data. In Proceedings of the 21st international conference on World Wide Web,
pages 271–280. ACM, 2012.
[Norouzi et al.2012] Mohammad Norouzi, Ali Punjani, and David J. Fleet. Fast search in hamming space
with multi-index hashing. In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference
on, pages 3108–3115. IEEE, 2012.
[Perozzi et al.2014] Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. DeepWalk: Online Learning of Social
Representations. arXiv:1403.6652 [cs], pages 701–710, 2014. arXiv: 1403.6652.
[Salakhutdinov and Hinton2009] Ruslan Salakhutdinov and Geoffrey Hinton. Semantic hashing. International
Journal of Approximate Reasoning, 50(7):969–978, July 2009.
[Sen and Singer1994] P.K. Sen and J.M. Singer. Large Sample Methods in Statistics: An Introduction with
Applications. Chapman & Hall/CRC Texts in Statistical Science. Taylor & Francis, 1994.
[Socher et al.2013] Richard Socher, Danqi Chen, Christopher D Manning, and Andrew Ng. Reasoning With
Neural Tensor Networks for Knowledge Base Completion. In Advances in Neural Information Processing
Systems 26, pages 926–934. 2013.
[Sutskever et al.2009] Ilya Sutskever, Joshua B. Tenenbaum, and Ruslan R Salakhutdinov. Modelling Relational Data using Bayesian Clustered Tensor Factorization. In Advances in Neural Information Processing
Systems 22, pages 1821–1828. Curran Associates, Inc., 2009.
[Tang and Liu2009] Lei Tang and Huan Liu. Relational learning via latent social dimensions. In Proceedings
of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining, pages
817–826. ACM, 2009.
[Weiss et al.2009] Yair Weiss, Antonio Torralba, and Rob Fergus. Spectral hashing. In Advances in neural
information processing systems, pages 1753–1760, 2009.
12
[Yang et al.2015] Huei-Fang Yang, Kevin Lin, and Chu-Song Chen. Supervised Learning of SemanticsPreserving Hashing via Deep Neural Networks for Large-Scale Image Search. arXiv:1507.00101 [cs], July
2015. arXiv: 1507.00101.
13
| 2cs.AI
|
UPDATING THE SILENT SPEECH CHALLENGE BENCHMARK
WITH DEEP LEARNING
Yan Ji, Licheng Liu, Hongcui Wang, Zhilei Liu, Zhibin Niu, Bruce Denby
Tianjin University, Tianjin, China
ABSTRACT
The 2010 Silent Speech Challenge benchmark is updated with new results obtained in a Deep Learning
strategy, using the same input features and decoding strategy as in the original article. A Word Error
Rate of 6.4% is obtained, compared to the published value of 17.4%. Additional results comparing new
auto-encoder-based features with the original features at reduced dimensionality, as well as decoding
scenarios on two different language models, are also presented. The Silent Speech Challenge archive has
been updated to contain both the original and the new auto-encoder features, in addition to the original
raw data.
Index Terms—silent speech interface, multimodal speech recognition, deep learning, language
model
1. INTRODUCTION
1.1.Silent speech interfaces and challenges
A Silent Speech Interface, or SSI, is defined as a device enabling speech processing in the absence of an
exploitable audio signal – for example, speech recognition obtained exclusively from video images of
the mouth, or from electromyographic sensors (EMA) glued to the tongue. Classic applications targeted
by SSIs include:
1)
Voice-replacement for persons who have lost the ability to vocalize through illness or an
accident, yet who retain the ability to articulate;
2)
Speech communication in environments where silence is either necessary or desired: responding
to cellphone in meetings or public places without disturbing others; avoiding interference in call
centers, conferences and classrooms; private communications by police, military, or business
personnel.
The SSI concept was first identified as an outgrowth of speech production research, in tandem with the
proliferation of the use of cellular telephones, in 2010 in a special issue of Speech Communication [1],
where SSIs based on seven different non-acoustic sensor types were presented:
1)
MHz range medical ultrasound (US) + video imaging of tongue and lips;
2)
Surface electromyography, sEMG, sensors applied to the face and neck;
3)
Electromagnetic articulography EMA sensors attached to tongue, lips, jaw;
4)
Vibration sensors placed on the head and neck;
5)
Non-audible murmur microphones, NAM, placed on the neck;
6)
Electro-encephalography, EEG, electrodes;
7)
Cortical implants for a “thought-driven” SSI.
Figure 1: Overview of an SSI, showing non-acoustic sensors and non-acoustic automatic speech
recognition, ASR, which can be followed by speech synthesis, or retained as a phonetic, text, or other
digital representation, depending on the application.
As a non-acoustic technology, SSIs initially stood somewhat apart from the main body of speech
processing, where the standard techniques are intrinsically associated with an audio signal. Nevertheless,
the novelty of the SSI concept and their exciting range of applications – perhaps aided by an accrued
interest in multi-modal speech processing – are gradually allowing SSI technology to join the speech
processing main stream. Activity in SSI research has remained strong since the publication of [1], which
received the ISCA/Eurasip Best Paper Award in 2015. A recent survey of the literature reveals dozens of
publications on SSI systems, using not only on the original seven non-acoustic technologies mentioned
above, but also two additional ones, namely, low frequency air-borne ultrasound; and micropower radar
[1-51].
Despite this activity, SSIs today remain for the most part specialized laboratory instruments. The
performance of any automatic speech recognition (ASR) system is most often characterized by a Word
Error Rate, or WER, expressed as a percentage of the total number of words appearing in a corpus. To
date, no SSI ASR system has been able to achieve WER parity with state-of-the-art acoustic ASR.
Indeed, a number of practical issues make SSI ASR systems considerably more involved to implement
than their acoustic counterparts:
1. Sensor handling. While in acoustic ASR this may amount to no more than routine microphone
protocol, SSIs’ non-acoustic sensors are often rather specialized (and expensive), and require
physical contact with, or at a minimum careful placement with respect to, the speech biosignalproducing organs. This introduces problems of invasiveness; non-portability; and non-repeatability
of sensor placement, bringing added complexity to SSI experiments.
2. Interference. An SSI should in principle be silent, but certain SSI modalities – vibration sensors,
radar, and low frequency air-borne ultrasound, for example – are actually associated with signals that
can propagate beyond the area of utilization of the SSI. The possibility of interference or interception
may limit the adoption of these modalities outside the laboratory.
3. Feature extraction. While easily calculated Mel Frequency Cepstral Coefficients, MFCC, have
been the acoustic ASR features of choice for decades, feature selection for the specialized sensors of
SSIs remains an open question, particularly since many SSI modalities – ultrasound imaging, or
EEG, for example – are of much higher intrinsic dimensionality than a simple acoustic signal.
Furthermore, while the identification of stable phonetic signatures in acoustic data is today a mature
field, the existence of salient landmarks in speech biosignals – arising from imaging modalities or
electromyography, for example – is less evident.
Medical US operating in the MHz frequency range does not propagate outside the body. It is a well
established [53] and documented [54] technique in speech production and speech pathology research,
whose first use in the context of SSIs was discussed in [55]. US is a also relatively non-invasive
modality, requiring only a transducer placed under the speaker’s chin, coupled with a small video
camera in front of the mouth to capture lip movement. These sensors can be easily accommodated in a
lightweight acquisition helmet, thus minimizing sensor placement issues. US tongue imaging, with
added lip video, is thus in many ways an attractive modality for building a practical SSI.
1.2. The Silent Speech Challenge benchmark
In 2010, an US + lip video SSI trained on the well-known TIMIT corpus achieved, with the aid of a
language model (LM), a single speaker WER of 17.4% (84.2% “correct” word rate) on an independent
test corpus [52], representing a promising early SSI result on a benchmark continuous speech
recognition task. Subsequently, the raw image data of [52], that is, the original tongue ultrasound and lip
videos, were made available online as the so-called Silent Speech Challenge, or SSC archive [56]. The
purpose of the archive is to provide a stable data set to which newly developed feature extraction and
speech recognition techniques can be applied. The SSC data will serve as the basis of all the experiments
reported in this article.
Although a 17.4% WER for an SSI trained on a mono-speaker TIMIT corpus is “promising”, it must be
remembered that standard acoustic ASR can obtain similar or superior scores after training on the full
multi-speaker acoustic TIMIT corpus, a much more challenging task. Further advances are thus still
necessary in order to truly put Silent Speech Recognition, SSR, on a par with acoustic ASR.
In the past several years, improvements in acoustic speech recognition using Deep Neural NetworkHidden Markov Model (DNN-HMM) systems, rather than the traditional Gaussian Mixture ModelHMM (GMM-HMM), have become common. In this approach, a deep learning strategy is used to
improve estimation of the emission probabilities of the HMM used for speech decoding. It is natural to
ask to what extent a DNN-HMM approach can improve SSR performance as well. Despite the SSI
implementation challenges outlined earlier, applications of deep learning techniques to SSR have indeed
begun to appear. In [57], for example, tests are reported of phonetic feature discrimination for an EMGbased SSI, without a LM, on a small, experiment-specific speech corpus. In [58], deep learning on an
EMA based SSI is explored, giving SSR phone error rates, PER, (which will be lower than WER) of
36%, when the Mocha-TIMIT corpus is used for training, testing, and the development of a specific
bigram LM. In [59], a DNN-HMM is applied to the SSC benchmark data, albeit with a 38% WER, in a
study comparing the efficacy of different feature extraction methods.
The present article reports the first application of the DNN-HMM approach to the SSC recognition
benchmark using the same input features and decoding strategy as those reported in [52], thus allowing a
direct comparison of performances. The SSR results obtained here are significantly improved compared
to the archive, giving, in the best scenario, a 6.4% WER (94.1% “correct” word recognition rate), or a
nearly threefold improvement over the benchmark value. In contrast to [57-58], furthermore, the LM
used in [52], also employed here, was developed on a completely independent speech corpus. In adition,
results with a second, less task-specific LM are included in the present article. Finally, tests of reduced
dimensionality feature vectors, as well as completely new input features created from raw SSC archive
data, are also reported here. All new features have been added to the SSC archive for future use by other
researchers.
In the remainder of the article, the details of the SSC data acquisition system and a description of the
available archive data are first summarized, in Section 2. Section 3 then describes the feature extraction
strategy developed for the present study; while full details of the DNN-HMM based speech recognition
procedure appear in Section 4. The results are summarized in Section 5, and some conclusions and
perspectives for future work outlined in the final section.
2. SSC DATA ACQUISITION AND ARCHIVE RESOURCES
The SSC data acquisition system consisted of an acquisition helmet holding a 128 element, 4-8 MHz US
probe for tongue imaging, and a black and white, infrared-illuminated video camera to capture the lips.
The 320×240 pixel tongue images and 640×480 pixel lip images created by the system were acquired in
a synchronized manner at 60 frames per second (fps) using the Ultraspeech multisensory acquisition
system [60].
The SSC training corpus consists of US and lip video data from a single native English speaker
pronouncing the 2342 utterances (47 lists of 50 sentences) of the TIMIT corpus, in the non-verbalized
punctuation manner. The speech was recorded silently, i.e., without any vocalization; there is therefore
no audio track. The test set is comprised of one hundred short sentences selected from the WSJ0 5000word corpus [61] read by the same speaker. The data are available at the web address indicated in [56].
The archive initially contained only the raw ultrasound and lip images of the training and test sets; the
original features used, as well as the reduced-length feature vectors and new features created for the
present article (see section 3), have now been appended to it. Speech recognition for the Challenge data
was carried out in a standard GMM-HMM scheme and made use of a LM, which is also included in the
archive. Further details appear in section 4.
3. FEATURE EXTRACTION
3.1. Introduction
As mentioned earlier, speech recognition from non-acoustic sensor data faces the problem of
discovering an effective feature recognition strategy, and US + lip video SSIs, although attractive in
many ways, share this drawback. Being based on images, their intrinsic input dimensionality may be of
the order of 1 million pixels. Some means of dimension-reducing feature extraction is thus critical. (The
following discussion is centered on tongue features. Lip features, which are much easier to handle, will
for overall coherence be extracted in the same way as tongue features.)
3.2. Contour extraction approach
Tongue contour extraction is a tempting choice for reducing dimensionality that retains visual
interpretability of the features. In ultrasound imaging of the tongue, the air-tissue boundary at the upper
surface of the tongue produces a bright, continuous contour, referred to in a side-looking scan as the
sagittal contour. Image processing tools for automatically extracting and characterizing this contour
make ultrasound imaging a powerful tool for the study of speech production [53], [54]. Unfortunately,
despite extensive literature on techniques for extracting tongue contours from ultrasound data (see [62][64] and references therein), tongue contour tracking remains an extremely challenging task. The high
level of speckle noise in ultrasound images (multiplicative noise arising from the coherent nature of the
ultrasound wave), coupled with variations in acoustic contact between the transducer and the speaker’s
skin; blocking of the ultrasound wave by the hyoid bone and jaw; poor reflectivity of muscle fibers in
certain orientations of the tongue; and the lack of a complete echogenic tissue path to all parts of the
tongue surface, in particular the tongue tip, often result in sagittal contours that are incomplete, contain
significant artifacts, or are even totally absent. While even imperfect automatically-extracted contours
remove the tedium of hand-scanning and are valuable for qualitative studies, it is difficult to integrate
such information in a coherent way into labeled training datasets intended for machine learning tasks
such as speech recognition. As a consequence, US-based SSIs have tended to use projective feature
extraction techniques rather than contour finding. In work performed thus far, Principal Component
Analysis, PCA, and the Discrete Cosine Transform, DCT, have been the methods of choice.
3.3. PCA and DCT approaches
In [8] and [10], PCA was used on ultrasound-based SSIs in an “Eigentongues” approach, wherein each
ultrasound image is represented as a linear combination of a set of orthogonal Eigen-images determined
on a training set of representative images. The first 30 or so Eigentongues were found sufficient to
represent the discriminative power contained in the ultrasound images.
The DCT, widely used for lossy image compression, is based on the notion that most of the information
in an image is concentrated in the lower spatial frequencies [65]. We note that the DCT, as a direct
multiplicative transform related to the Fourier transform, does not make use of a training set. The
technique for calculating the DCT coefficients will be presented in the next section. In [52], the article
on which the SSC archive is based, it was found that DCT features provided substantially better
recognition scores, as well as faster execution times, than the Eigentongue approach. This result leads
to the important consequence that the SSC benchmark refers to recognition scores obtained using
DCT features (we note in addition that the original Eigentongue features of [52] are no longer
available). Consequently, a quantitative comparison of a DNN-HMM approach to the GMM-HMM
analysis used in the original benchmark – which is the major impetus of this article – must make use of
the identical DCT features in its baseline result.
The SSC archive DCT features were constructed in the following way. First, fixed Regions of Interest
(ROI) of tongue and lip images were resized to 64 by 64 pixels. This resizing is necessary in order to
keep the number of DCT coefficients tractable. For an image matrix A of size N*N, the two-dimension
DCT is then computed as:
!!! !!!
𝐷!" = 𝑎! 𝑎!
𝐴!" cos
!!! !!!
𝜋 2𝑚 + 1 𝑖
𝜋 2𝑛 + 1 𝑗
cos
(1)
2𝑁
2𝑁
0 ≪ i, j ≤ N − 1
where
1
𝑎! , 𝑎! =
,
𝑖, 𝑗 = 0
2
,
𝑁
1 ≪ 𝑖, 𝑗 ≪ 𝑁 − 1
𝑁
Dimensionality reduction is achieved by retaining only the K lowest frequency DCT components. In
[52], a feature size of K = 30 was selected, based on performance comparisons. In acoustic speech
recognition, it is usual to concatenate the first derivative, or Δ, of the MFCC feature vector to the vector
itself. This procedure was also carried out for the DCT features of the archive, thus creating a 120component feature vector for each tongue + lip frame.
3.4. New features created with Deep Auto Encoder
Although DCT features have provided promising recognition results for US + lip video SSIs, it has been
necessary to make certain compromises in extracting them, notably: 1) resizing the original images
before calculating them; and 2) retaining only a small, fixed number of DCT coefficients. While
computational tractability issues prevent us, at present, from removing the first restriction, the presence
of the raw tongue and lip data in the SSC archive allows us to consider taking a closer look at the second
one.
It is first of all interesting to examine the appearance of tongue and lip images reconstructed using 30
DCT coefficients. An example result on 4 frames is given in figure 2.
Figure 2: Original lip (top row) and tongue (third row) images compared to their reconstructions (second
and fourth rows, respectively) using 30 DCT coefficients.
Although the lip reconstructions are sufficiently clear to distinguish an overall degree of mouth opening,
an acoustically pertinent quantity, the visual fidelity of the tongue images is rather poor. The
information in the tongue images necessary for distinguishing different acoustic configurations is,
evidently, coded by the DCT in a way that does not retain a high level of visual fidelity. It is tantalizing
to ask, however, whether one might do better by creating, from the original images present in archive, a
new feature representation that reduces dimensionality while explicitly preserving visual fidelity, rather
than relying on a somewhat arbitrarily placed cut in spatial frequency space, as was done for the DCT. A
Deep Auto Encoder, or DAE, was used to explore this possibility.
A Deep Auto-Encoder is a neural network used for reducing the dimensionality and learning a
representation of input data [66]. It contains an “encoder” and a “decoder” symmetrically arranged about
a “code” layer, as shown in Figure 3. The action of the encoder can be defined as:
𝑧 = 𝑓(𝑊𝑥 + 𝑏)
(2)
where 𝑓 is an activation function, such as sigmoid function, 𝑊 a weight matrix, and b a bias. The
decoder output is defined as:
𝑥 ! = 𝑓 𝑊 !𝑧 + 𝑏!
(3)
where 𝑥 ! is of the same dimension as 𝑥. The weight matrix 𝑊 ! is equal to 𝑊 ! . An autoencoder is
trained by minimizing the image reconstruction error, computed as:
𝐿 𝑥, 𝑥 ! = −
!
!
!!![𝑓! (𝑥! , 𝑥! )
+ 𝑓! (𝑥! , 𝑥!! )]
(4)
where
𝑓! 𝑥, 𝑥 ! = 𝑥 log 𝑥 !
𝑓! 𝑥, 𝑥 ! = 1 − 𝑥 log 1 − 𝑥 ! .
When training is complete, the code layer may be regarded as a compressed representation of the input,
and is suitable for use as a feature vector. Details of the DAE training procedure can be found in [67].
Figure 3. Architecture of DAE
To obtain the new features, ROIs were selected and resized, once again for computational tractability
purposes, via bi-cubic interpolation, to 50×70 (lip) and 50×60 (tongue) pixel arrays, which form the
inputs to the DAE. After tests with various architectures, a J-1000-500-250-K network was chosen, with
J the number of inputs (3500 for lips and 3000 for the tongue); K the desired dimensionality of the
created feature vector; and the intermediate figures the number of neurons in each layer. Features were
calculated for K = 30, 20, 10, and 5. The encoder and symmetric decoder networks were trained on 12
lists of images selected at random from the SSC TIMIT training corpus.
Reconstructed images (bottom row of each panel) for tongue and lips are compared with the original
images (top row of each panel) in Figure 4, where (a) and (b) show the results using 30 and 5 DAE
features respectively. The figure shows that remarkable visual fidelity can be obtained using only 5 DAE
features. This is in contrast to images reconstructed using DCT features shown previously, which are
barely recognizable even for the 30 dimensional features. Although one may ask to what extent the DAE
solution is similar to PCA, the K = 5 case, with, as will be seen later, the SSR results it allows to obtain,
is nonetheless intriguing.
(a) K = 30
(b) K = 5
Figure 4: Original (top row of each panel) and reconstructed (bottom row of each panel) images of
tongue and lips using two dimensionalities of DAE features.
3.5. Feature-free approaches
In the past few years, “feature-free” approaches to pattern recognition in speech, signal and image
processing, based on convolutional neural networks, CNN, have proven very effective [68][69][70][71].
The CNN is a multilayer neural network consisting of multiple sub-networks with shared weights and
overlapping receptive fields, alternated with “pooling” layers that reduce dimensionality by retaining
only a subset of the afferent inputs. The use of shared weights across different instances of the identical
sub-networks greatly reduces the number of weights to be learned, thus allowing the training of a CNN
to remain relatively tractable. CNNs are thought to be able to learn a hierarchy of features, of
progressively higher order as information pass from the input to the output of the network.
Recently, CNN have begun to make their entry into the field of SSI. In [72], which is actually a lipreading application, a CNN is trained to transform video frames from a large video database directly into
synthesized un-vocalized speech, using the video sound track to create source-filter type training labels.
In [73], a CNN is trained to recognize phonetic targets in US tongue and video lip images in a 488sentence single speaker database, using a phonetically labeled sound track as ground truth, for a speech
recognition task with an HMM-GMM. In [74], CNN are used to recognize tongue gestural targets in US
tongue images for an isolated phoneme + nonsense word recognition task. In the latter reference,
extensive use is made of data augmentation [70] to increase the size of the training set, often a concern
in using CNN, which require very large training sets to be effective, due to the large number of weights
that must be learned.
Conceivably, the CNN technique could be applied to the raw images of the SSC archive, to try to
improve on the ad-hoc DCT and DAE features tested thus far. As the archive contains no sound track,
however, pre-training of the CNN, as in [72] and [73], will not be feasible: the CNN training will have
to take place conjointly with that of the HMM probabilities. A study of this possibility will appear in an
upcoming article.
4. DNN–HMM SPEECH RECOGNITION
4.1. System overview
The Kaldi open-source Deep Learning toolkit [75][76] was used to build the SSR system, whose overall
architecture is illustrated in Figure 5. Features extracted from the archive data were first normalized to
have zero mean and unit variance (Mean Variance Normalization (MVN) in the figure).
.
Features
MVN
Monophone
DNN
Triphone
(LDA+MLLT)
Triphone (∆)
Figure 5: Overall SSR training procedure.
In the SSC benchmark, HTK was used to perform the speech recognition, using standard GMM-HMM
architecture. In order to ensure as meaningful a comparison as possible with the benchmark result,
without actually re-doing it with HTK, the recognition with Kaldi was performed first using a GMMHMM, and then a DNN-HMM. The procedures used for the non-acoustic ASR, adapted from standard
recipes in acoustic speech recognition and Deep Learning [77-81], are described below.
In the Kaldi GMM-HMM “acoustic” model training stage (the name “acoustic model” is retained even
though the input feature data used here are non-acoustic), a monophone model was first trained using
combined tongue and lip feature vectors, of type DCT or DAE, of dimension K = 30, 20, 10, and 5. The
alignment of the monophone is then used for the triphone1 training stage, where the Δ features are also
included. In the subsequent phase, the triphone2b model is created using the alignment result of
triphone1, and applying Linear Discriminant Analysis (LDA) and Maximum Likelihood Linear
Transformation (MLLT) methods to replace the Δ features appearing in triphone1 and produce a new
feature vector of dimension 40. The monophone, triphone1 and triphone2b acoustic models were trained
consecutively, each time using the previous model for alignment, while for the DNN-HMM training, the
alignment of triphone2b was used.
The Deep Belief Network (DBN) implemented is illustrated in Figure 6 [77-79], using the D
dimensional feature vectors as inputs. Restricted Boltzmann Machines (RBM) are cascaded by means of
the weight vector W in the figure. On the top layer of the DBN is a softmax output layer, and transition
probabilities of the HMM are trained in the previous GMM-HMM phase. The system parameters are
summarized in Table 1, including total numbers of Gaussians, tied state (regression tree), search space
(beam), and acoustic/LM weight (acwt) parameters. The DNN training operates in two phases. During
the pre-training phase, the RBMs are trained using the contrastive divergence (CD) algorithm. The 6
hidden layers of the RBMs are made up of Bernoulli-Bernoulli units (learning rate 0.4) except for the
first, which is Gaussian-Bernoulli (learning rate 0.01). In the second phase, 90% of the training set was
used to train the DNN, optimizing per-frame cross-entropy, and the remaining 10% of the training was
used to test [80]. The weights learned in pre-training phase are then used to initialize the DNN model.
The DNN architecture was implemented on a CUDA GPU machine.
Figure 6. DNN structure used for SSR
Table 1. SSR system parameters
Monophone
Tot_Gaussian
1700
Triphone1
Regression tree leaves
1800
Tot_Gaussian
9000
Regression tree leaves
3000
Tot_Gaussian
25000
Number of hidden layers
6
Units per hidden layer
1024
Number of hidden layers
4
Units per hidden layer
1024
beam
13.0
Lattice_beam
8.0
acwt
0.1
Triphone2b
DNN pretrain
DNN training
4.2.Language model and lexicon issues
The lm_wsj_5k_nvp_2gram LM [61] used in the decoding stage of the SSC benchmark, derived from a
fixed 5000-word subset of the Wall Street Journal (WSJ) text corpus, was also adopted in these tests.
Obtaining realistic WER scores on small corpora, however, can be problematic. Using a closed
vocabulary, as is the case here, tends towards underestimation of attainable WER. On the other hand, an
unbiased lexicon derived exclusively from a small training set might not contain all the words present in
the test set, thus leading to an overly pessimistic WER estimate. To help address these issues, a second
estimate of the achievable WER on these data was also made using another, less task-specific LM,
namely lm_csr_5k_nvp_2gram [61]. This LM contains newswire text from WSJ, the San José Meteor,
and the Associated Press, along with some spontaneous dictation by journalists of hypothetical news
articles. Results on both LM appear in the next section.
5. RESULTS AND ANALYSIS
Table 2 shows a comparison of the Kaldi DNN-HMM results, on the WSJ0 5000-word corpus, to those
of the SSC benchmark, using the same 30-dimensional DCT input features and decoding strategy as in
[52]. The formula used for WER is
𝑊𝐸𝑅 =
!!!!!
(8)
!
where I is the number of insertion error, D is number of deletions, S the number of substitutions, and N
the total number of words in the test set. Although a test with HTK itself was not repeated here, the fact
that quite similar results were obtained using a GMM-HMM in Kaldi (column 2) provides reassurance
that the figures obtained using Kaldi are reasonable. The Table shows that the DNN-HMM strategy has
reduced the WER by almost a factor of 3 as compared to the benchmark.
Table 2. Comparison with original HTK result of [52], using 30-element DCT features
Error
HTK
Kaldi
Kaldi
SSC
GMM-
DNN-
Benchmark
HMM
HMM
Insertion
17
41
6
Deletion
23
17
8
Substitution
138
120
52
Number of
1023
1023
1023
862
886
963
Correct rate
84.26%
86.61%
94.13%
WER
17.4%
17.4%
6.45%
words
Correct
words
To perform the LM tests proposed in section 4.2, the procedure was repeated using the alternate LM, as
shown in Table 3 for 30-element feature vectors of both types (DCT and DAE). One notes first of all
that the DCT and DAE features give similar performances, barring the monophone case. We will return
to this point in the discussion of Tables 4 and 5. Nonetheless, although the WER performance obtained
on the less specific LM is somewhat worse, as expected, it is still significantly better than the SSC
benchmark, for both types of features.
Table 3. Comparing results for the 2 different LM, for 30-element feature vectors of both types
WER (%)
LM
DCT
DAE
lm_csr_5k_
lm_wsj_5_
nvp_2gram
nvp_2gram
monophone
45.55
40.47
Triphone2b
17.40
12.71
DNN
11.44
6.45
monophone
58.55
59.92
Triphone2b
21.70
14.76
DNN
13.98
7.72
To further explore different types of input features, DCT and DAE feature vectors of dimension K = 20,
10, 5 for each visual modality were also tested. Results are given in Table 4 for the WSJ LM, and in
Table 5 for the CSR LM. Overall, higher scores are again obtained with the more task specific LM, as
expected. The tables also show that for both LM, similar results are obtained for the two types of
features, with the DCT being slightly better, when the dimensionality K of the input vectors is 10 or
more. For K = 5, however, while the DCT features are no longer salient, the DAE retains most of its
effectiveness. Thus, although the DAE has not been completely successful at simultaneously optimizing
saliency and low dimensionality, the results it furnishes are intriguing, and suggest that it may be
possible to do better with a more sophisticated approach.
Table 4. Recognition results with WSJ LM
WER of
DCT (%)
Dimension
30
20
10
5
Monophone
40.47
37.15
36.36
98.24
Triphone2b
13.00
14.76
12.32
100
DNN
6.45
6.35
7.43
99.51
WER of
DAE (%)
Dimension
30
20
10
5
Monophone
59.92
44.18
41.15
45.45
Triphone2b
14.76
14.96
15.54
17.79
DNN
7.72
7.72
8.80
10.07
Table 5. Recognition results with CSR LM
WER of
DCT (%)
Dimension
30
20
10
5
Monophone
45.55
40.86
39.78
98.34
Triphone2b
17.79
19.16
16.42
100
DNN
11.44
11.53
12.32
99.80
WER of
DAE (%)
Dimension
30
20
10
5
Monophone
58.55
52.00
49.76
54.25
Triphone2b
21.70
21.41
19.75
22.48
DNN
13.98
13.10
14.37
14.86
6. CONCLUSIONS AND PERSPECTIVES
A confrontation of the SSC recognition benchmark with DNN-HMM SSR techniques using the Kaldi
Deep Learning package has led to an improvement in WER of almost a factor of 3 in the most favorable
scenario, thus helping to establish US as a highly attractive SSI modality. Tests performed using both
the original WSJ LM and a less task-specific CSR LM give WER values that are on these data, using the
significantly improved compared to the benchmark. Before the DNN-HMM tests, Kaldi was also used to
test a GMM-HMM architecture, in order to demonstrate compatibility with the methods used in the
benchmark. New features derived from the raw benchmark data using a DAE give results only slightly
worse than those obtained with the original DCT features, while retaining their effectiveness even at
very low dimensionality. Both new and original features have now been appended to the SSC
benchmark data.
While these results are promising, SSR still remains somewhat less accurate than acoustic speech
recognition, and further work will be necessary. In the future, for the SSC benchmark, it will be
interesting to experiment with other feature extraction strategies, for example convolutional neural
networks, CNN, which might allow the image-resizing step, where information may be lost, to be
skipped. For SSI more generally, it will be interesting to accumulate much larger (if possible multispeaker) data sets, so that some of the mentioned problems associated with small speech data sets may
be avoided.
7. ACKNOWLEDGEMENTS
This research was supported by the National Nature Science Foundation of China (No. 61303109 and
No. 61503277) and 985 Foundation from China’s Ministry of Education (No. 060-0903071001).
8. REFERENCES
[1] B. Denby, T. Schultz, K. Honda, T. Hueber, J. M.Gilbert, and J.S. Brumberg, Silent speech
interfaces, Speech Communication 52(4), 2010, pp. 270-287.
[2] T. Hueber, E-L. Benaroya, G. Chollet, B. Denby, G. Dreyfus, M. Stone, Development of a silent
speech interface driven by ultrasound and optical images of the tongue and lips, Speech
Communication, Volume 52, Issue 4, April 2010, pp. 288–300.
[3] T. Schultz and M. Wand, Modeling Coarticulation in EMG-based Continuous Speech Recognition,
Speech Communication, Volume 52, Issue 4, April 2010, pp. 341-353.
[4] J.S. Brumberg, A. Nieto-Castanon, P.R. Kennedy, F.H. Guenther, Brain-Computer Interfaces for
Speech Communication, Speech Communication, Volume 52, Issue 4, April 2010.
[5] T. Hirahara, M. Otani, S. Shimizu, M. Toda, K. Nakamura, Y. Nakajima, K. Shikano, Silent-speech
enhancement
system
utilizing
body-conducted
vocal-tract
resonance
signals,
Speech
Communication, Volume 52, Issue 4, 2010.
[6] V.A. Tran, G. Bailly, H. Loevenbruck, T. Toda, Improvement to a NAM-captured whisper-tospeech system, Speech Communication, Volume 52, Issue 4, April 2010.
[7] S.A. Patil, J.H.L. Hansen, A competitive alternative for speaker assessment: Physiological
Microphone (PMIC), Speech Communication, Volume 52, Issue 4, April 2010.
[8] T. Hueber, G. Aversano, G. Chollet, B. Denby, G. Dreyfus, Y. Oussar, P. Roussel, M. Stone,
Eigentongue feature extraction for an ultrasound-based silent speech interface, Proceedings of
ICASSP (Honolulu, USA), pp. 1245-1248, 2007.
[9] T. Hueber, G. Chollet, B. Denby, M. Stone, Acquisition of ultrasound, video and acoustic speech
data for a silent-speech interface application, Proceedings of International Seminar on Speech
Production (Strasbourg, France), pp. 365-369, 2008.
[10] T. Hueber, G. Chollet, B. Denby, G. Dreyfus, M. Stone, Phone Recognition from Ultrasound and
Optical Video Sequences for a Silent Speech Interface, Proceedings of Interspeech (Brisbane,
Australia), pp. 2032-2035, 2008.
[11] C.S. DaSalla, H. Kambara, M. Sato, Y. Koike, Spatial filtering and single-trial classification of
EEG during vowel speech imagery, Proceedings of the 3rd International Convention on
Rehabilitation Engineering and Assistive Technology (i-CREATe 2009), Singapore.
[12] T. Hueber, G. Chollet, B. Denby, G. Dreyfus, M. Stone, (2009), Visuo-Phonetic Decoding using
Multi-Stream and Context-Dependent Models for an Ultrasound-based Silent Speech Interface,
Proceedings of Interspeech (Brighton, UK), pp. 640-643.
[13] A. R. Toth, M. Wand, T. Schultz, Synthesizing Speech from Electromyography using Voice
Transformation Techniques, Proceedings of Interspeech (Brighton, UK), 2009.
[14] M. Florescu, L. Crevier-Buchman, B. Denby, T. Hueber, A. Colazo-Simon, C. Pillot-Loiseau,
Silent vs Vocalized Articulation for a Portable Ultrasound-Based Silent Speech Interface,
Proceedings of Interspeech 2010, Makuhari, Chiba, Japan, September 26-30, 2010.
[15] E. Lopez-Larraz, O. M. Mozos, J. M. Antelis, J. Minguez, Syllable-Based Speech Recognition
Using EMG, Proceedings of the IEEE Engineering in Medicine and Biology Society 2010; pp.
4699-702.
[16] T. Hueber, P. Badin, G. Bailly, A. Ben-Youssef, F. Elisei, B. Denby, G. Chollet, Statistical
Mapping between Articulatory and Acoustic Data, Application to Silent Speech Interface and
Visual Articulatory Feedback, Proceedings of Interspeech, pp. 593-596, Firenze, Italy, 2011.
[17] M. Wand, T. Schultz, Session-Independent EMG-based Speech Recognition, International
Conference on Bio-inspired Systems and Signal Processing, Proceedings of Biosignals 2011, Rome,
Italy.
[18] B. Denby, J. Cai, P. Roussel, G. Dreyfus, L. Crevier-Buchman, C. Pillot-Loiseau, T. Hueber, G.
Chollet, Tests of an Interactive, Phrasebook-Style Post-Laryngectomy Voice-Replacement System,
Proceedings of ICPhS 2011, Hong Kong.
[19] B. Denby, J. Cai, T. Hueber, P. Roussel, G. Dreyfus, L. Crevier-Buchman, C. Pillot-Loiseau, G.
Chollet, S. Manitsaris, M. Stone, Towards a Practical Silent Speech Interface Based on Vocal Tract
Imaging, 9th International Seminar on Speech Production (ISSP 2011), Montréal, Canada, pp.89-94,
2011.
[20] J. Freitas, A. Teixeira, M. S. Dias, C. Bastos, Towards a Multimodal Silent Speech Interface for
European Portuguese, "Speech Technologies", Ivo Ipsic, Ed., Intechopen, cdn.intechopen.com, 2011.
[21] K. A. Yuksel, S. Buyukbas, S. H. Adali, Designing Mobile Phones using Silent Speech Input and
Auditory Feedback, Proceedings of MobileHCI 2011, Aug 30–Sept 2, 2011, Stockholm, Sweden.
[22] T. Hueber, G. Bailly, B. Denby, Continuous Articulatory-to-Acoustic Mapping using Phone-based
Trajectory HMM for a Silent Speech Interface, Proceedings of Interspeech, Portland, USA, 2012.
[23] C. C. Jorgensen, D. D. Lee, S. T. Agabon, United States National Aeronautics and Space
Association, Silent Speech Sub-audible speech recognition based on electromyographic signals, US
Patent number 8200486B1; 2012.
[24] J. Wang, A. Samal, J. R. Green, F. Rudzicz, Sentence Recognition from Articulatory Movements
for Silent Speech Interfaces, 2012 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), Kyoto, 25-30 March 2012 pp. 4985 – 4988.
[25] A.A. García, C. A. Reyes-García, L. Villaseñor-Pineda, Towards a Silent Speech Interface Based
on Unspoken Speech, Proceedings of Biosignals 2012 (BIOSTEC), Algarve, Portugal. pp 370-373.
SciTePress, 2012.
[26] A. Barbulescu T. Hueber, G. Bailly, R. Ronfard, Audio-Visual Speaker Conversion using Prosody
Features, Proceedings of Int. Conference of Audio-visual Speech Processing (AVSP), Annecy,
France, 2013.
[27] R. Hofe, S. R. Ell, M. J. Fagan, J. M. Gilbert, P. D. Green, R. K. Moore, S. I. Rybchenko, Smallvocabulary speech recognition using a silent speech interface based on magnetic sensing, Speech
Communication 2013, Vol. 55, No. 1, pp. 22-32.
[28] M. Wand, A. Himmelsbach, T. Heistermann, M. Janke, T. Schultz, Artifact removal algorithm for
an EMG-based Silent Speech Interface. Proceedings of the IEEE Engineering Medicine and Biology
Society Conference, 2013; pp. 5750-5753.
[29] J. Freitas, A. Teixeira, M. S. Dias, Multimodal Silent Speech Interface based on Video, Depth,
Surface Electromyography and Ultrasonic Doppler: Data Collection and First Recognition Results,
Proceedings of SPASR2013, Lyon, France.
[30] J. A. Gonzalez, L. A. Cheah, J. Bai, S. R. Ell, J. M. Gilbert, R. K. Moore, P. D. Green, Analysis of
Phonetic Similarity in a Silent Speech Interface based on Permanent Magnetic Articulography,
Proceedings of Interspeech, Singapore, Malaysia, 2014.
[31] D. Bocquelet, T. Hueber, L. Girin, P. Badin, B. Yvert, Robust articulatory speech synthesis using
deep neural networks for BCI applications, Proceedings of Interspeech, Singapore, Malaysia, 2014,
pp. 2288-2292.
[32] J. Freitas, A. J. Ferreira, M. A. Teles de Figueiredo, A. Teixeira, M. S. Dias, Enhancing
Multimodal Silent Speech Interfaces With Feature Selection, Proceedings of Interspeech, Singapore,
Malaysia, 2014, p. 1169-1173.
[33] J. Wang, A. Samal, J. R. Green, Preliminary Test of a Real Time, Interactive Silent Speech
Interface Based on Electromagnetic Articulograph, Proceedings of the 5th Workshop on Speech and
Language Processing for Assistive Technologies (SLPAT), pages 38–45, Baltimore, Maryland USA,
2014.
[34] M. Salama, L. ElSherif, H. Lashin, T. Gamal, Recognition of Unspoken Words Using Electrode
Electroencephalograhic Signals, Sixth International Conference on Advanced Cognitive
Technologies and Applications, Cognitive 2014 (IARIA), Venice, Italy, 2014.
[35] H. Sahni, A. Bedri, G. Reyes, P. Thukral, Z. Guo, T. Starner, M. Ghovanloo, The tongue and ear
interface: a wearable system for silent speech recognition,
Proceedings of the 2014 ACM
International Symposium on Wearable Computers, Seattle, 2014.
[36] J. Freitas, A. Teixeira, M. S. Dias, Multimodal Corpora for Silent Speech Interaction, Proceedings
of the Ninth International Conference on Language Resources and Evaluation (LREC'14),
Reykjavik, Iceland, 2014.
[37] D. Balwani, H. Brijwani, K. Daswani, S. Rastog, Talking Without Talking, Int. Journal of
Engineering Research and Applications, Vol. 4, Issue 4, April 2014, pp.55-56,
[38] M. Matsumoto, Brain Computer Interface using Silent Speech for Speech Assistive Device, PhD
Thesis, Doctoral Program in Information Science and Engineering, Graduate School of Science and
Technology, Niigata University, 2014.
[39] J. Freitas, A. Teixeira, S. Silva, C. Oliveira, M.S. Dias, Detecting Nasal Vowels in Speech
Interfaces Based on Surface Electromyography, PLoS ONE 10(6): 2015.
[40] A. Al Safi, L. Alhafadhi, Review of EMG-based Speech Recognition, International Journal of
Review in Electronics & Communication Engineering (IJRECE), Volume 3, Issue 3, June 2015.
[41] D. Bocquelet, T. Hueber, L. Girin, C. Savariaux, B. Yvert, Real-time Control of a DNN-based
Articulatory Synthesizer for Silent Speech Conversion: a pilot study, Proceedings of Interspeech,
Dresden, 2015, pp. 2405-2409.
[42] J. Wang, S. Hahm, T. Mau, Determining an Optimal Set of Flesh Points on Tongue, Lips, and Jaw
for Continuous Silent Speech Recognition, 6th Workshop on Speech and Language Processing for
Assistive Technologies, Dresden, Germany, 11 September, 2015.
[43] S. Hahm, J. Wang, Silent Speech Recognition from Articulatory Movements using Deep Neural
Network, Proceedings of ICPhS2015, Glasgow, 2015.
[44] A. Jaumard-Hakoun, K. Xu, P. Roussel-Ragot, G. Dreyfus, M. Stone, B. Denby, Tongue Contour
Extraction from Ultrasound Images Based on Deep Neural Network, Proceedings of ICPhS2015,
Glasgow, 2015.
[45] K. Xu, Y. Yang, A. Jaumard-Hakoun, C. Leboullenger, G. Dreyfus, P. Roussel-Ragot, M. Stone, B.
Denby, Development of a 3D tongue motion visualization platform based on ultrasound image
sequences, Proceedings of ICPhS2015, Glasgow, 2015.
[46] H. Yamaguchi, T. Yamazaki, K. Yamamoto, S. Ueno, A. Yamaguchi, T. Ito, S. Hirose, K. Kamijo,
H. Takayanagi, T. Yamanoi, S. Fukuzumi? Decoding Silent Speech in Japanese from Single Trial
EEGS: Preliminary Results, Journal of Computer Science and Systems Biology 8:5, 2015.
[47] T. Hueber, G. Bailly, Statistical Conversion of Silent Articulation into Audible Speech using FullCovariance HMM, Computer Speech and Language, vol. 36, pp. 274-293, 2016.
[48] W. Li, Silent Speech Interface Design Methodology and Case Study, Chinese Journal of
Electronics 2016, Vol. 25 Issue (CJE-1): 88-92.
[49] L. A. Cheah , J. M. Gilbert, J.A. Gonzalez, J. Bai, S. R. Ell, M. J. Fagan, R. K. Moore, P. D. Green,
S. I. Rychenko, Integrating User-Centred Design in the Development of a Silent Speech Interface
Based on Permanent Magnetic Articulography, Biomedical Engineering Systems and Technologies,
Volume 574 of the series Communications in Computer and Information Science pp. 324-337,
January 2016.
[50] P. Patil, G. Gujarathi, G. Sonawane, Different Approaches for Artifact Removal in
Electromyography based Silent Speech Interface, International Journal of Science, Engineering and
Technology Research (IJSETR), Volume 5, Issue 1, January 2016.
[51] T. Yamazaki, Silent Speech Brain-Computer Interface in Japanese, LAP Lambert Academic
Publishing, 2016.
[52] J. Cai, B. Denby, P. Roussel-Ragot, G. Dreyfus, L. Crevier-Buchmann, Recognition and real time
performance of a lightweight ultrasound based silent speech interface employing a language model,
Proceedings of Interspeech, Florence, Italy, pp. 1005-1008, 2011.
[53] M. Stone, B. Sonies, T. Shawker, G. Weiss, L. Nadel, Analysis of real-time ultrasound images of
tongue configuration using a grid-digitizing system. Journal of Phonetics, 11, 207-218, 1983.
[54] M. Stone, A guide to analysing tongue motion from ultrasound images, Clinical Linguistics and
Phonetics, 19, 455-501, 2005.
[55] B. Denby, M. Stone, Speech Synthesis from Ultrasound Images of the Tongue, Proceedings of
ICASSP, Montréal, Canada, 2004.
[56] B. Denby, T. Hueber, J. Cai, P. Roussel, L. Crevier-Buchman, S. Manitsaris, G. Chollet, M. Stone,
C. Pillot, The Silent Speech Challenge Archive, 2013, online: https://ftp.espci.fr/pub/sigma/.
[57] M. Wand, T. Schultz, Pattern learning with deep neural networks in EMG-based speech
recognition, Proceedings of EMBC14, IEEE, Chicago, USA, 2014, pp. 4200-4203.
[58] S. Hahm, J. Wang, and J. Friedman, Silent speech recognition from articulatory movements using
deep neural network, Proc. of the International Congress of Phonetic Sciences, Glasgow, Scotland,
UK, 2015, pp. 1-5.
[59] L. Liu, Y. Ji, H. Wang, B. Denby, Comparison of DCT and autoencoder-based features for DNNHMM multimodal silent speech recognition, Proceedings of ISCSLP, Tianjin, China, 2016.
[60] T. Hueber, G. Chollet, B. Denby, M. Stone, Acquisition of ultrasound, video and acoustic speech
data for a silent-speech interface application, Proceedings of the International Seminar on Speech
Production, Strasbourg, France, pp. 365-369, 2008.
[61] J.
Garofalo,
D.
Graff,
D.
Paul,
D.
Pallett,
CSR-I
(WSJ0)
Complete,
online:
https://catalog.ldc.upenn.edu/LDC93S6A
[62] M. Li, C. Kambhamettu, M. Stone, Automatic Contour Tracking in Ultrasound Images, Clinical
Linguistics and Phonetics, 19(6-7) 2005 pp. 545-554.
[63] L. Tang. G. Hamarneh, Graph-based tracking of the tongue contour in ultrasound sequences with
adaptive temporal regularization, Computer Vision and Pattern Recognition Workshops (CVPRW),
IEEE, June 2010.
[64] K. Xu, Y. Yang, M. Stone, A. Jaumard-Hakoun, C. Leboullenger, G. Dreyfus, P. Roussel, B.
Denby, Robust contour tracking in ultrasound tongue image sequences. Clinical Linguistics and
Phonetics, 30(3-5) 2016, pp. 313-27.
[65] K. R. Rao and P. Yip, “Discrete Cosine Transform: Algorithms, Advantages, Applications”,
Academic Press, Boston, 1990.
[66] G. Hinton, A practical guide to training restricted Boltzmann machines, Momentum 9.1, 2010, p.
92ff.
[67] G. Hinton, R. Salakhutdinov, Reducing the dimensionality of data with neural networks, Science,
2006, 3135786, pp. 504-507.
[68] L. Bottou, F. Fogelman Soulié, P. Blanchet. J.S. Lienard, Experiments with Time Delay Networks
and Dynamic Time Warping for speaker independent isolated digits recognition, Proceedings of
Eurospeech, Paris, France, 1989, ISCA Archive.
[69] Y. Le Cun, L. Bottou, L., Y. Bengio, P. Haffner, Gradient-based learning applied to document
recognition, Proceedings of the IEEE, 86(11), 2278-2324, 1998.
[70] A. Krizhevsky, I. Sutskever, G. Hinton, Imagenet classification with deep convolutional neural
networks, Advances in neural information processing systems 25 (pp. 1097-1105). Lake Tahoe,
USA: MIT Press, 2012.
[71] O. Abdel-Hamid, A.-R. Mohamed, H. Jiang, L. Deng, G. Penn, D. Yu, Convolutional Neural
Networks for Speech Recognition, IEEE/ACM Transactions on Audio, Speech, and Language
Processing, Vol. 22, No. 10, 2014.
[72] ID2Speech: Speech reconstruction from silent video, Ariel Ephrat and Shmuel Peleg, to appear in
Proceedings of ICASSP2017, New Orleans, USA, 2017.
[73] E. Tatulli, T. Hueber, Feature extraction using multimodal convolution neural networks for visual
speech recognition, to appear in Proceedings of ICASSP2017, New Orleans, USA, 2017.
[74] K. Xu, P. Roussel, T. Gábor Csapó, B. Denby, Convolutional neural network-based automatic
classification of midsagittal tongue gestural targets using B-mode ultrasound images, submitted to
JASA Express Letters, 2017.
[75] D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P.
Motlicek, Y. Qian, P. Schwarz, J. Silovsky, G. Stemmer, K. Vesely, The Kaldi Speech Recognition
Toolkit, IEEE Workshop on Automatic Speech Recognition and Understanding, Hawaii, USA,
2011.
[76] The Kaldi Toolkit, Online: http://kaldi-asr.org/
[77] G. Dahl, D. Yu, L. Deng, A. Acero, Context-dependent pre-trained deep neural networks for largevocabulary speech recognition, IEEE Transactions on Audio Speech and Language Processing,
2015, 20(1), pp. 30 - 42.
[78] G. Dahl, D. Yu, L. Deng, and A. Acero, Large vocabulary continuous speech recognition with
context-dependent DBN-HMMs, in Proceedings of ICASSP2011, Prague, Czech Republic, pp.
4688-4691, 2011.
[79] G. Hinton, L. Deng, D. Yu, G. Dahl, A-R. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P.
Nguyen, T. Sainath, B. Kingsbury, Deep neural networks for acoustic modeling in speech
recognition. IEEE Signal Processing Magazine, 2012, 29(6), pp. 82 - 97.
[80] B. Popović, S. Ostrogonac, E. Pakoci, N. Jakovljević, V. Delić, Deep neural network based
continuous speech recognition for Serbian using the Kaldi toolkit, Lecture Notes in Computer
Science: Speech and Computer 9319, 2015, pp. 186-192.
[81] S. P. Rath, D. Povey, K. Vesely, J. Cernocky, Improved feature processing for Deep Neural
Networks, Proceedings of Interspeech, Lyon, France, 2013.
| 1cs.CV
|
1
Amplify-and-Forward Two-Way Relaying System
over Free-Space Optics Channels
arXiv:1708.05083v1 [cs.IT] 16 Aug 2017
Jaedon Park Member, IEEE, Chan-Byoung Chae Senior Member, IEEE, and Giwan Yoon Member, IEEE
Abstract—In this paper, we analyze the performance of a twoway subcarrier intensity-modulated (SIM) amplify-and-forward
(AF) relaying system over free-space optics (FSO) communication
channels. The analysis takes into consideration attenuations due
to atmospheric turbulence, geometric spread and pointing errors
at the same time. We derive, in generalized infinite power series
expressions, the tight upper and lower bounds of the overall
outage probability and average probability of errors of the
system. The study finds that this two-way subcarrier intensitymodulated AF relaying system using a binary phase shift keying
(BPSK) modulation could be used for practical applications in
case of a weak turbulence regime in which the required SNR
is about 30 dB to obtain the average bit error probability of
10−6 . It is also noted that the pointing errors clearly degrade the
performance of the two-way subcarrier intensity-modulated AF
relaying system.
Keywords—Amplify-and-forward relay, atmospheric turbulence,
average probability of error, free-space optics, overall outage
probability, pointing errors, subcarrier intensity modulation, twoway relaying system, upper and lower bounds.
I.
INTRODUCTION
REE-space optics (FSO) systems are high-capacity and
cost-effective communication techniques that are free of
radio frequency spectrum regulations [1]–[5]. Because of this
they have attracted enormous amounts of scholarly attention. Although an intensity modulation and direct detection
(IM/DD) system using on/off keying (OOK) has been widely
used due to its simplicity [1]–[4], such a system is not appropriate for applications of amplify-and-forward (AF) relaying
systems. Indeed, an IM/DD system requires an adaptive decision threshold that, in practice, is very difficult to implement.
To circumvent this implementation difficulty, we consider in
this paper a subcarrier intensity modulation (SIM) scheme. The
scheme requires no adaptive decision threshold and ameliorates
turbulence-induced irradiance fluctuation [6]–[11]. Therefore,
the subcarrier intensity modulation scheme is suitable for AF
relaying systems over FSO channels.
F
This research was in part supported by the Agency for Defense Development (ADD) and the Basic Science Research Program through the National
Research Foundation of Korea (NRF) funded by the Ministry of Education
(Grant No. 2016R1D1A1B01007074).
J. Park is with the Agency for Defense Development (ADD), Daejeon,
34186, Korea. E-mail: [email protected].
G. Yoon is with the School of Electrical Engineering, Korea Advanced
Institute of Science and Technology (KAIST), Daejeon, 34141, Korea. E-mail:
[email protected].
C.-B. Chae is with the School of Integrated Technology, Yonsei University,
03722, Korea. E-mail: [email protected].
FSO systems are highly affected by attenuations caused by
atmospheric turbulence and geometric spread and pointing errors [12]–[16]. As a result of variations in the refractive index,
turbulence-induced fading, also known as scintillation, causes
irradiance fluctuations in the received signals intensity [12]–
[15]. Apart from the scintillation effects, pointing errors due to
building sway also cause significant performance degradations
in FSO systems [12]–[15].
In addition to the performance degradation caused by atmospheric turbulence and pointing errors, FSO communication
systems suffer a significant degradation in non-line-of-sight
environments. Overcoming such problems is left up to relaying technologies [17]–[24]. Relaying systems are classified
as either amplify-and-forward (AF) or decode-and-forward
(DF) relays [23], [25], [26]. DF relaying systems, also called
regenerative systems, decode the received signal fully and reencode it before retransmitting it to another hop [23], [25],
[26]. The AF relaying systems, also called nonregenerative
systems, just amplify the received signal and forward it to
another hop with less complexity than the DF relaying systems
[23], [25], [26]. Since AF relaying systems do not decode and
re-encode the received signal, they need less power and lower
system complexity than do DF relaying systems [23], [25]–
[30].
Recently, an enormous amount of research interest has
been devoted to the two-way relaying system techniques for
conventional RF applications [31]–[36]. Such interest is largely
due to the technique’s even more efficient signaling scheme,
where two nodes bi-directionally communicate in just two
phases via a half-duplex relay, resulting in the improvement of
the spectral efficiency [31]–[36]. Upadhyay et al. studied the
performance of a two-way opportunistic relaying system with
analog network coding over Nakagami-m fading channels [31].
Guo et al. analyzed the overall outage probability as well as the
symbol error probability of a two-way AF relaying system over
an exponential distribution [32]. Jang et al. have investigated
the performance of a multiuser two-way relay channel [33].
Han et al. have analyzed the tight upper and lower bounds
of the average sum rate of a two-way AF relaying system
with/without Alamouti’s orthogonal space time block code
(OSTBC) considering gamma distributions [34]. Yang et al.
analyzed the performance of two-way AF relaying system in
Nakagami-m fading channels [35]. Ikki et al. have analyzed
the performance of two-way AF relaying in the presence of cochannel interferences over Rayleigh fading channels [36]. As
is the case with a general relaying system, the two-way AF
relaying system is more attractive in practice than the twoway DF relaying system due to its very simple processing
at the relay terminal [34]. Also, similarly to conventional AF
2
relaying systems, the two-way AF relaying systems are able
to lower more effectively the system power consumption and
complexity than are two-way DF relaying systems.
In the case of the two-way relaying systems for FSO
applications, Tang et al. proposed a network-coded cooperation
relay scheme for optical DF two-way relay networks; they
investigated the optimal bit decision algorithm for a receiving
node over Gamma-Gamma fading distributions [37]. Puri et al.
investigated the performance of a two-way relay-assisted FSO
system considering a DF protocol over log-normal distribution
or gamma-gamma distribution [38], [39]. They also proposed
and analyzed relay selection protocols of two-way DF FSO
relays assuming Gamma-Gamma distribution with pointing
errors [40]. The same authors later analyzed two-way AF
relay selection and derived the achievable-rates [41] and outage
probability and average error probability [42] in closed-form
expressions. Abu-Almaalie et al. studied two-way DF FSO
relay systems using SIM-BPSK [43].
It is difficult to gain an immediate insight from the analyzed
performances of FSO systems [44] since the error probabilities
of FSO channels are usually expressed as a complex Meijer’s
G-function due to the modified Bessel function of the second
kind [17], [19], [37]. Moreover, performance analysis is a large
challenge, especially in the case of AF relaying systems over
a Gamma-Gamma fading distribution, because of the complex
form of the Meijer’s G-function. In a novel approach, [2],
[44] simplify the mathematical expressions by introducing a
generalized infinite power series representation of the modified
Bessel function of the second kind. With this representation,
we can express the error probabilities of FSO systems as power
series expansions, composed of only elementary and Gamma
functions; this offers readers an mathematical insight into FSO
systems.
In this paper, we analyze, for the first time, the performance
of a two-way AF relaying system using the subcarrier intensity modulation scheme [6]–[10] over Gamma-Gamma fading
environments with a simple generalized infinite power series
expression [2], [44]. We analyze FSO communication channels
considering attenuations due to atmospheric turbulence and
geometric spread and pointing errors [12]–[15]. The main
contributions of this paper are as follows:
• Overall outage probability analysis of a two-way AF
relaying system over FSO channels: We derive the upper
and lower bounds of the overall outage probability of a
two-way AF relaying system over Gamma-Gamma fading channels considering attenuations due to atmospheric
turbulence and geometric spread and pointing errors.
• Average probability of error analysis of a two-way AF
relaying system over FSO channels: Based on the derived results of the overall outage probability upper and
lower bounds, we further derive the average probability
of errors of the two-way subcarrier intensity-modulated
AF relaying system over the Gamma-Gamma fading
channels considering attenuations due to atmospheric
turbulence and geometric spread and pointing errors.
The rest of the paper is organized as follows: In Section II,
we discuss the system and channel model of a two-way
Relay
+
Photo detector
Photo detector
Amplifier
DC bias
OBPF
+
OBPF
Laser driver
Terminal A
Laser driver
DC bias
+
Terminal B
OBPF
OBPF
Laser driver
Photo detector
Photo detector
+
BPSK
demodulator
BPSK
demodulator
Output data
Output data
x1 (t )
BPSK
modulator
Input data
Fig. 1.
DC bias
x2 (t )
BPSK
modulator
Input data
A two-way SIM AF relaying system using a BPSK modulation.
subcarrier intensity-modulated AF relaying system over the
Gamma-Gamma fading distributions. In Section III, we derive,
in generalized infinite power series expressions, the upper and
lower bounds of the overall outage probability for the system
under consideration. In Section IV, we derive the average
probability of errors corresponding to the upper and lower
bounds of the overall outage probability given in Section III.
In Sections V and VI, we present the numerical results and
the conclusion.
II. S YSTEM AND CHANNEL MODEL
A. System Model
Consider a single-relay, two-way AF relaying system as
described in Fig. 1, where the two source terminals, A and
B, communicate along with a relay terminal R using the
optical subcarrier intensity modulation scheme [6]–[10] with a
binary phase shift keying (BPSK) modulation. For simplicity,
throughout this paper, we assume a BPSK modulation. Other
modulation schemes, however, can also be applicable. The
channels are assumed to be stationary with independent and
identically distributed (i.i.d.) intensity fading statistics. The
channel state information (CSI) is also assumed to be available
at the receiver.
At Terminal A, the source data is modulated onto the RF
subcarrier signal x1 (t) using a BPSK modulation. The RF
signal is added with a bias signal to drive the laser diode on the
positive values. At Relay R, the received optical signal radiated
from Terminal A passes through an optical band pass filter
(OBPF) to reject the background radiation noise. Finally, the
photodetector generates the photocurrent yR1 (t) proportional
to x1 (t). Similarly, the photodetector generates the photocurrent yR2 (t) proportional to x2 (t), the RF subcarrier signal
PARK et al.: TWO-WAY AMPLIFY-AND-FORWARD RELAYING SYSTEM ...
3
radiated from the source, Terminal B. After the two signals
yR1 (t) and yR2 (t) are added, the photocurrent can be given
by
yR (t) = ηI1 (1 + ξx1 (t)) + ηI2 (1 + ξx2 (t)) + nR (t).
where η is the photodetector responsiveness; I1 is the irradiance from Terminal A to Relay R; I2 is the irradiance from
Terminal B to Relay R; ξ is a constant to fultill |ξx1 (t)| ≤ 1,
and |ξx2 (t)| ≤ 1; nR (t) ∼ N (0, σ 2 ) is an additive white
Gaussian noise (AWGN) mainly due to thermal and/or background noise [6], [10], [11]. The channel state is considered to
be the product of two random factors, i.e., Ii = Iia Iip where
Iia is the attenuation due to atmospheric turbulence, which is
modeled as Gamma-Gamma fading distribution [1]–[3], [6],
[9], [44], and Iip is the attenuation due to geometric spread
and pointing errors [12]–[15], [20]–[22]. It is noted that the
channel model of random variable Ii , i = 1, 2, is described in
detail in the following section.
After the DC components of the received photocurrent yR (t)
are filtered out, the photocurrent can be given by
yR (t) = ηI1 ξx1 (t) + ηI2 ξx2 (t) + nR (t).
(1)
The relay amplifies the electrical signal by the gain, GR , which
is given by
s
PR
GR =
,
(2)
2
PS (ηI1 ξ) + PS (ηI2 ξ)2 + σ 2
to meet the average transmit power constraints [32]. The
amplified electrical signal is added with a bias signal to
drive the laser diode on the positive values. Finally, the relay
retransmits the optical signal to both terminals, A and B. Here,
PR and PS represent the subcarrier signal powers of the relay
and the terminal, respectively. At Terminal A, after passing
through an optical band pass filter, the received photocurrent
is given by
yBRA (t) = ηI1 (1 + ξGR yR (t)) + nA (t)
(3)
2
where nA (t) is the AWGN with ∼ N (0, σ ). Substituting (1)
for (3) and filtering out the DC component and subtracting the
self-interference parts [31]–[33], [37], the received photocurrent can be rewritten as
yBRA (t) = GR ηI1 ξηI2 ξx2 (t) + GR ηI1 ξnR (t) + nA (t). (4)
Substituting (2) into (4), without any loss of generality, if
we assume both PS = PR = P0 and P0 ≫ σ 2 , the received
SNR at Terminal A can be expressed as
ΓBRA =
(I1 I2 )2 2 2
η ξ γ0 ,
2I12 + I22
where γ0 = P0 /σ 2 is the SNR in the absence of turbulence
and pointing errors. Similarly, the received SNR at Terminal
B can be expressed as
ΓARB =
(I1 I2 )2 2 2
η ξ γ0 .
I12 + 2I22
B. Turbulence and Misalignment Fading Model
It is assumed that the random variable Iia , i = 1, 2, follows a
Gamma-Gamma distribution, and using the generalized power
series representation method, the probability density function
(PDF) and the cumulative distribution function (CDF) of the
random variable Iia can be expressed as [2], [44]
fIia (Iia ) = lim
J→∞
J
X
j=0
j+β−1
j+α−1
,
aj (α, β)Iia
+ aj (β, α)Iia
(5)
J
X
aj (α, β)
aj (β, α) j+α
,
I
j + α ia
j=0
(6)
π(αβ)j+β
, and α and
where aj (α, β) = sin[π(α−β)]Γ(α)Γ(β)Γ(j−α+β+1)j!
β are the atmospheric turbulence parameters [2], [3], [6], [9],
[44]
−1
0.49χ2
α = exp
−1
,
(1 + 0.18d2 + 0.56χ12/5 )7/6
FIia (Iia ) = lim
J→∞
j+β
j+β
Iia
+
−1
0.51χ2 (1 + 0.69χ12/5 )−5/6
−
1
,
β = exp
(1 + 0.9d2 + 0.62d2 χ12/5 )5/6
where χ2 = 0.5Cn2 k 7/6 z 11/6 and d = (kD2 /4z)1/2 . Here
k = 2π/λ is the optical wave number, λ is the wavelength,
D is the diameter of the receiver’s collecting lens aperture, z
is the link distance in meters and Cn2 is the altitude-dependent
index of refraction structure (see [3], [6], [9], [44] for more
details).
The PDF of Iip , by considering a circular detection aperture
of radius r and a Gausssian beam, is given by [12]–[15]
fIip (Iip ) =
γ2
2
I γ −1 ,
γ 2 ip
A0
0 ≤ Iip ≤ A0
(7)
where γ = wzeq /2σs is the ratio between the equivalent beam
radius at the receiver and the pointing error displacement standard deviation (jitter) at the receiver [12]–[15]. The
√
√ parameter
relation
v
=
wzeq can be calculated using the
πr/
2wz ,
√
A0 = [erf(v)]2 and wz2eq = wz2 πerf(v)/2v exp (−v 2 ) where
erf(·) is the error function and wz is the beam waist (radius
calculated at e−2 ) at distance z [12]–[15]. The combined PDF
of Ii = Iia Iip is given by [12]–[15]
Z
(8)
fIi (Ii ) = fIi |Iia (Ii |Iia )fIia (Iia )dIia ,
where fIi |Iia (Ii |Iia ) is the conditional probability given Iia
state and is expressed by
Ii
1
fIip
fIi |Iia (Ii |Iia ) =
Iia
Iia
γ 2 −1
Ii
γ2
, 0 ≤ Ii ≤ A0 Iia . (9)
= γ2
A0 Iia Iia
4
xY
xY
η Yξ Yγ W
b
a
Γth
xY = −Y xX +
Ψ (a, c )
c
O
Fig. 2.
xX
η Yξ Yγ W
YΓth
Rectangular integral region for X1 and X2 .
η Yξ Yγ W
A
E
AX
B
η Yξ Yγ W
Γth
D
F
xY = −
ZΓth
Substituting (5) and (9), (8) can be rewritten as
Z
γ 2 γ 2 −1 ∞
fIi (Ii ) = γ 2 Ii
×
A0
Ii /A0
J
X
j+β−1−γ 2
j+α−1−γ 2
lim
aj (α, β)Iia
+ aj (β, α)Iia
dIia .
J→∞
CX
O
Y
Fig. 3.
C
Y
η ξ γ W η Yξ Yγ W
ZΓth
xX η Yξ Yγ W
+
Y
YΓth
η Yξ Yγ W
YΓth
xX
Γth
Integral region of OABC for Φ(γ0 , Γth ).
j=0
(10)
The two random variables, I1 and I2 , are assumed to be
independent but not necessarily identically distributed GammaGamma fading. Hence, the above rectangular integral region
can be rewritten as
By calculating the definite integration, (10) is derived to
fIi (Ii ) =
J
X
j=0
γ2
2
I γ −1 lim ×
γ2 i
J→∞
A0
2
aj (α, β)(Ii /A0 )j+β−γ
aj (β, α)(Ii /A0 )j+α−γ
+
2
γ −j−β
γ2 − j − α
2
!
Ψ(a, c) = FX2 (a)FX1 (c).
.
J
X
γ2
lim
A0 J→∞ j=0
aj (α, β)Iij+β
(j + β)(γ 2 − j − β)A0j+β−1
!
aj (β, α)Iij+α
+
.
(12)
(j + α)(γ 2 − j − α)Aj+α−1
0
It is easily confirmed that when γ 2 goes to ∞ (for the nonpointing errors case), (11) and (12) converge to (5) and (6),
respectively.
III. OVERALL OUTAGE PROBABILITY ANALYSIS
Before we analyze the overall outage probability, we first
study the integral of a rectangular region as described in Fig. 2.
Lemma 1: Integral of a rectangular region for 1/I12 , 1/I22 .
The rectangular integral region can be expressed as
√
√
Ψ(a, c) = {1 − FI2 (1/ a)}{1 − FI1 (1/ c)}.
(13)
Proof: For a random variable Ii , i = 1, 2, we define a
new random variable Xi = 1/Ii2 . The integral of a rectangular
region composed of X1 and X2 , as shown in Fig. 2, can be
written as
Z c
Ψ(a, c) =
Pr[X2 < a|x1 ]fX1 (x1 )dx1
Z0 c Z a
=
fX2 (x2 )dx2 fX1 (x1 )dx1 .
0
0
Since Xi = 1/Ii2 , i = 1, 2, the CDF of Xi can be given by
(11)
By integrating the PDF in (11), the CDF can be obtained as
FIi (Ii ) =
(14)
FXi (x) = Pr(Xi < x) = Pr(Ii > I) = 1 − Pr(Ii < I)
√
(15)
= 1 − FIi (I) = 1 − FIi (1/ x).
Substituting (15) into (14), (13) is obtained.
For the two-way subcarrier intensity-modulated AF relaying
system, the overall outage probability is, then, defined as
(I1 I2 )2 2 2
(I1 I2 )2 2 2
<
Γ
η
ξ
γ
,
η
ξ
γ
Pout = Pr min
th
0
0
2I12 + I22
I12 + 2I22
2
2
(I1 I2 )
Γth
(I1 I2 )
Γth
= 1−Pr
> 2 2 , 2
> 2 2
2
2
2
2I1 + I2
η ξ γ0 I1 + 2I2
η ξ γ0
(16)
where, Γth is the threshold SNR. Let Xi = 1/Ii2 , i = 1, 2, the
outage probability can be rewritten as
Pout = 1 − Φ(γ0 , Γth )
(17)
where,
Φ(γ0 , Γth )
X1
η 2 ξ 2 γ0
η 2 ξ 2 γ0
= Pr X2 < −
.
, X2 < −2X1 +
+
2
2Γth
Γth
(18)
The upper probability Φ(γ0 , Γth ) is very difficult to derive in
a closed-form [32]. As described in Fig. 3, the integral region
for the calculation of Φ(γ0 , Γth ) is OABC. Here, we analyze
the upper and lower bounds instead of the exact form.
PARK et al.: TWO-WAY AMPLIFY-AND-FORWARD RELAYING SYSTEM ...
5
A. The upper bound of the overall outage probability
Theorem 1: The upper bound of the overall outage probability is given by
2 2
η ξ γ0 η 2 ξ 2 γ0
Pout < Pout,U = 1 − Ψ
.
(19)
,
3Γth
3Γth
Proof: The integral region OABC is lower bounded by
OA1 BC1 . Here, the integral region, OA1 BC1 , can be obtained
by Lemma 1 as follows:
2 2
η ξ γ0 η 2 ξ 2 γ0
OA1 BC1 = Ψ
.
(20)
,
3Γth
3Γth
Thus, the integral region of OABC, Φ(γ0 , Γth ), is lowerbounded by
2 2
η ξ γ0 η 2 ξ 2 γ0
Φ(γ0 , Γth ) > Ψ
.
(21)
,
3Γth
3Γth
Substituting (21) into (17), (19) is obtained.
U
Also,
the
CDF,
of
the
FΓ (x),
(I1 I2 )2 2 2
(I1 I2 )2 2 2
min 2I 2 +I 2 η ξ γ0 , I 2 +2I 2 η ξ γ0 is approximated from
1
2
1
2
the outage probability upper bound in (19) by the variable
change from Γth to x as
2 2
η ξ γ0 η 2 ξ 2 γ0
.
(22)
FΓU (x) = 1 − Ψ
,
3x
3x
Thus, the integral region of OABC, Φ(γ0 , Γth ), is upperbounded by
η 2 ξ 2 γ0 η 2 ξ 2 γ0
,
Φ(γ0 , Γth ) < Ψ
2Γth
3Γth
2 2
2 2
η ξ γ0 η 2 ξ 2 γ0
η ξ γ0 η 2 ξ 2 γ0
−Ψ
. (26)
+Ψ
,
,
3Γth
2Γth
3Γth
3Γth
Substituting (26) into (17), (23) is obtained.
L
Also,
the
CDF,
of
the
FΓ (x),
2
2
(I1 I2 )
(I1 I2 )
2 2
2 2
min 2I 2 +I 2 η ξ γ0 , I 2 +2I 2 η ξ γ0 is approximated from
1
2
1
2
the outage probability lower bound in (23) by the variable
change from Γth to x as
(
η 2 ξ 2 γ0 η 2 ξ 2 γ0
=1− Ψ
,
2x
3x
)
2 2
η 2 ξ 2 γ0 η 2 ξ 2 γ0
η ξ γ0 η 2 ξ 2 γ0
−Ψ
. (27)
,
,
+Ψ
3x
2x
3x
3x
FΓL (x)
IV.
AVERAGE P ROBABILITY
E RROR A NALYSIS
In this section, the average probability of error is derived for
the two-way subcarrier intensity-modulated AF relaying system. If we let P (e|Γ) denote the conditional error probability
in an AWGN channel, the average probability of error can be
expressed as
B. The lower bound of the overall outage probability
Theorem 2: The lower bound of the overall outage probability is given by
OF
Pe =
∞
Z
P (e|Γ)fΓ (Γ)dΓ
(28)
0
where, the conditional error probability can be given by [45]
(
η 2 ξ 2 γ0 η 2 ξ 2 γ0
,
Pout > Pout,L = 1 − Ψ
2Γth
3Γth
2 2
2 2
)
2 2
η ξ γ0 η ξ γ0
η ξ γ0 η 2 ξ 2 γ0
+Ψ
−Ψ
.
,
,
3Γth
2Γth
3Γth
3Γth
P (e|Γ) = Q
(23)
Proof: The integral region OABC is upper bounded
by OAEBF C. The integral region of OAEBF C can be
expressed as
OAEBF C = OAEC1 + OA1 F C − OA1 BC1 .
(24)
Here, each integral of all regions, OAEC1 , OA1 F C, and
OA1 BC1 , can be obtained by Lemma 1 as follows:
η 2 ξ 2 γ0 η 2 ξ 2 γ0
OAEC1 : Ψ
,
,
2Γ
3Γth
2 2 th
η ξ γ0 η 2 ξ 2 γ0
OA1 F C : Ψ
,
,
3Γ
2Γth
2 th2
η ξ γ0 η 2 ξ 2 γ0
OA1 BC1 : Ψ
.
,
3Γth
3Γth
(25)
√
δΓ
(29)
where, δ is 2 for BPSK modulation. Substituting (29) to (28),
(28) can be rewritten as [27]
1
Pe = √
2π
Z
∞
FΓ
0
t2
δ
t2
e− 2 dt
(30)
where, FΓ (Γ) is the CDF of the random variable Γ. After the
variable change of x = t2 , the average probability of error can
be given by [26]
1
Pe = √
2π
Z
0
∞
FΓ
x dx
e− 2 √ .
δ
2 x
x
(31)
Theorem 3: If we substitute (22) and (27) into (31), we
obtain the average probability of errors corresponding to the
upper and lower bounds of the overall outage probability as
(
2 2
)
√
η ξ γ0 δ η 2 ξ 2 γ0 δ
1
2π − A
, (32)
,
PU (e) = √
3
3
2 2π
6
0
10
AJ3 (u, v)
2 2
J
X
j+2β+1
γ
j + 2β + 1
(1p)
2
=
cj (u, v)2
lim
Γ
J→∞
A0
2
j=0
2 2
J
X (2p)
j+α+β+1
γ
j+α+β+1
2
+
Γ
cj (u, v)2
lim
J→∞
A0
2
j=0
2 2
J
X
j+α+β+1
γ
j+α+β+1
(3p)
+
Γ
cj (u, v)2 2
lim
J→∞
A0
2
j=0
2 2
J
X
j+2α+1
j + 2α + 1
γ
(4p)
cj (u, v)2 2 Γ
.
lim
+
J→∞
A0
2
j=0
−2
Overall Outage Probability
10
−4
10
−6
10
−8
Up, Str.
Low, Str.
Up, Mod.
Low, Mod.
Up, Weak
Low, Weak
Monte−Carlo sim.
10
−10
10
−12
10
0
10
20
γ0 (dB)
30
40
50
(37)
Fig. 4. Overall outage probability of a two-way SIM AF relaying system. J
is set to 100. Γth = 0 dB.
(
2 2
√
1
η ξ γ0 δ η 2 ξ 2 γ0 δ
PL (e) = √
2π − A
,
2
3
2 2π
)
2 2
η 2 ξ 2 γ0 δ η 2 ξ 2 γ0 δ
η ξ γ0 δ η 2 ξ 2 γ0 δ
+A
,
,
,
−A
3
2
3
3
(33)
where,
A(u, v) =
√
2π − AJ1 (u) − AJ2 (v) + AJ3 (u, v).
(34)
Here, AJ1 (u), AJ2 (v) and AJ3 (u, v) are given by
AJ1 (u) =
AJ2 (v) =
j+β
J
X
aj (α, β)u− 2
γ2
lim
A0 J→∞ j=0 (j + β)(γ 2 − j − β)A0j+β−1
j+β+1
j+β+1
×2 2 Γ
2
(35)
j+α
J
2
X
γ
aj (β, α)u− 2
+
lim
A0 J→∞ j=0 (j + α)(γ 2 − j − α)Aj+α−1
0
j+α+1
j+α+1
,
×2 2 Γ
2
j+β
J
X
γ2
aj (α, β)v − 2
lim
A0 J→∞ j=0 (j + β)(γ 2 − j − β)A0j+β−1
j+β+1
j+β+1
2
×2
Γ
2
(36)
j+α
J
2
X
aj (β, α)v − 2
γ
lim
+
A0 J→∞ j=0 (j + α)(γ 2 − j − α)Aj+α−1
0
j+α+1
j+α+1
×2 2 Γ
,
2
Proof: See Appendix A.
V. N UMERICAL RESULTS
In this section, we present the numerical results of the derived overall outage probability’s upper and lower bounds and
average probability of errors corresponding to the bounds of
the overall outage probability. A two-way subcarrier intensitymodulated AF relaying system was considered, using a BPSK
modulation over the FSO channel. The FSO channel was
modeled as a Gamma-Gamma fading distribution with the
atmospheric turbulence parameters of α = 4.2 and β = 1.4
for a strong turbulence regime, parameters of α = 4.0 and β =
1.9 for a moderate turbulence regime, and α = 8.5 and β = 6.7,
for a weak turbulence regime [2], [6], [9], [19]. For pointing
errors, the normalized jitter of σs /r = 0.02 ∼ 0.4 and the
normalized beamwidth of wz /r = 5, 10, 15 were considered.
For numerical evaluations, the parameter J used in the infinite
power series was truncated to 100.
Figs. 4, 5, and 6 show the overall outage probability
performances of the two-way subcarrier intensity-modulated
AF relaying system with respect to the SNR in the absence
of turbulence and pointing errors for the threshold SNRs of
0 dB, 5 dB, and 10 dB, respectively. Here, we do not consider
pointing errors (γ 2 is set to ∞ with A0 = 1).
According to Fig. 4, which is given for the threshold SNR
of 0 dB, to obtain the overall outage probability of 10−6 , the
required SNR values are upper bounded by 99.2 dB, 77.8 dB,
and 34.6 dB for the strong, moderate, and weak turbulence
regimes. It can also be seen in the figure that to obtain the
overall outage probability of 10−6 the required SNR values
are lower bounded by 97.6 dB, 76.2 dB, and 32.6 dB for the
strong, moderate, and weak turbulence regimes.
According to Fig. 5, which is given for the threshold SNR
of 5 dB, to obtain the overall outage probability of 10−6 , the
required SNR values are upper bounded by 104.2 dB, 82.8 dB,
and 39.6 dB for the strong, moderate, and weak turbulence
regimes. It can also be seen in the figure that to obtain the
overall outage probability of 10−6 the required SNR values
are lower bounded by 102.6 dB, 81.2 dB, and 37.6 dB for the
strong, moderate, and weak turbulence regimes.
PARK et al.: TWO-WAY AMPLIFY-AND-FORWARD RELAYING SYSTEM ...
7
0
0
10
10
−2
10
−2
Overall Outage Probability
Overall Outage Probability
10
−4
10
−6
10
Up, Str.
Low, Str.
Up, Mod.
Low, Mod.
Up, Weak
Low, Weak
Monte−Carlo sim.
−8
10
−10
10
0
10
20
−4
10
−6
10
−8
10
Up, Str.
Low, Str.
Up, Mod.
Low, Mod.
Up, Weak
Low, Weak
Monte−Carlo sim.
−10
10
−12
10
−14
γ0 (dB)
30
40
50
Fig. 5. Overall outage probability of a two-way SIM AF relaying system. J
is set to 100. Γth = 5 dB.
10
10
20
30
40
γ0 (dB)
50
60
70
80
Fig. 7. Overall outage probability of a two-way SIM AF relaying system
with pointing errors. J is set to 100. Γth = 0 dB. σs /r = 0.1. Black color:
wz /r = 15, Red color: wz /r = 10, Blue color: wz /r = 5.
0
10
Overall Outage Probability
−2
Overall Outage Probability
10
0.025
0.02
−4
10
0.015
−6
0.01
10
Up, Str.
Low, Str.
Up, Mod.
Low, Mod.
Up, Weak
Low, Weak
Monte−Carlo sim.
−8
10
−10
10
0
10
20
0.005
0
5
4
0.4
0.3
3
γ0 (dB)
30
40
50
w /r
z
Fig. 6. Overall outage probability of a two-way SIM AF relaying system. J
is set to 100. Γth = 10 dB.
According to Fig. 6, which is given for the threshold SNR
of 10 dB, to obtain the overall outage probability of 10−6 , the
required SNR values are upper bounded by 109.3 dB, 87.9 dB,
and 44.6 dB for the strong, moderate, and weak turbulence
regimes. It can also be seen in the figure that to obtain the
overall outage probability of 10−6 , the required SNR values
are lower bounded by 107.6 dB, 86.2 dB, and 42.6 dB for the
strong, moderate, and weak turbulence regimes.
It can be clearly seen in Figs. 4, 5, and 6 that the derived
upper and lower bounds, having about 1.6 ∼ 2.0 dB gaps, are
very tight for the various atmospheric turbulence regimes and
SNR scenarios. We also present the Monte-Carlo simulation
results of the exact outage probability to evaluate the tightness of the upper and lower bounds of the analyzed outage
probability, given in (19) and (23).
0.2
2
0.1
1
0
σs/r
Fig. 8. Overall outage probability for the different normalized beamwidth
and normalized jitter. Γth = 0 dB. γ0 = 60 dB. Upper bound for the Strong
regime.
Figs. 7 and 8 present the overall outage probability performances of the two-way subcarrier intensity-modulated AF
relaying system considering the pointing error effects. Fig. 7
shows the overall outage probability performances for various
values of the normalized beamwidth wz /r = 5, 10, 15. The
threshold SNR is set to 0 dB. For the pointing errors, the
normalized jitter σs /r is set to 0.1. Fig. 8 depicts the upper
bound performance of the overall outage probability for the
strong regime varying the normalized beamwidths and normalized jitters at the same time. The threshold SNR is set to
0 dB and the SNR is set to 60 dB. It is clearly seen in the
figures that a better overall outage probability performance is
achieved by using a narrow beamwidth and a small jitter.
8
10
0
10
10
γ : 0 dB
0
γ0: 10 dB
8
−2
10
γ0: 20 dB
γ : 30 dB
6
Average Probability of Error
Overall Outage Probability
10
0
10
γ0: 40 dB
γ0: 50 dB
4
10
2
10
0
10
−2
−6
10
−8
10
Up, Str.
Low, Str.
Up, Mod.
Low, Mod.
Up, Weak
Low, Weak
Monte−Carlo sim.
−10
10
−12
10
10
−4
10
−4
10
−14
0
20
40
60
80
100
10
20
30
40
50
γ0 (dB)
J
Fig. 9. Convergence characteristics of the overall outage probability with
respect to J without pointing errors. Γth = 0 dB. Upper bound for the Moderate
regime.
60
70
80
Fig. 11. Average probability of error of a two-way AF relaying system for
SIM BPSK with pointing errors. J is set to 100. σs /r = 0.1. Black color:
wz /r = 15, Red color: wz /r = 10, Blue color: wz /r = 5.
Average Probability of Error
0
10
−2
10
0.025
Average Probability of Error
−4
10
0.02
−6
10
0.015
−8
10
0.01
−10
0.005
10
−12
10
−14
10
−16
10
20
Up, Str.
Low, Str.
Up, Mod.
Low, Mod.
Up, Weak
Low, Weak
Monte−Carlo sim.
30
40
0
5
4
0.4
0.3
3
0.2
2
50
γ0 (dB)
60
70
80
wz/r
0.1
1
0
σ /r
s
Fig. 10. Average probability of error of a two-way AF relaying system for
SIM BPSK. J is set to 100.
Fig. 12. Average probability of error for the different normalized beamwidth
and normalized jitter. γ0 = 50 dB. Upper bound for the Strong regime.
Fig. 9 illustrates the convergence performance of the derived
overall outage probability with respect to the varying number
of power series J. The threshold SNR Γth is set to 0 dB. Here,
the upper bound of the outage probability is considered for the
moderate regime. It is observed in the figure that the outage
probability converges only with J = 2 in the high SNR of 50
dB.
Fig. 10 shows the average probability of error performances
of the two-way subcarrier intensity-modulated AF relaying
system with respect to the SNR in the absence of turbulence
and pointing errors. In this figure, for simplicity’s sake, we
do not consider pointing errors. As described in Section IV,
since the average probability of errors are derived based on the
CDF functions from the outage probability’s upper and lower
bounds, their performances also show the upper and lower
bounds. As can be seen in the figure, to obtain the average
probability of error of 10−6 for the BPSK subcarrier intensity
modulation scheme, the required SNR values are bounded by
89.5 ∼ 91.3 dB, 70.0 ∼ 71.8 dB, and 32.3 ∼ 34.1 dB for the
strong, moderate, and weak turbulence regimes. It can also
be clearly seen in the figure that the derived upper and lower
bounds have about 1.8 dB gaps for the various atmospheric
turbulence regimes and SNR scenarios, which is very tight
and similar to the results of the outage probabilities. It should
be noted that this two-way subcarrier intensity-modulated AF
relaying system using BPSK modulation could be used for
practical applications in the case of a weak turbulence regime,
in which the required SNR is about 30 dB to obtain the average
PARK et al.: TWO-WAY AMPLIFY-AND-FORWARD RELAYING SYSTEM ...
15
10
γ : 0 dB
0
γ0: 10 dB
γ0: 20 dB
Average Probability of Error
10
γ : 30 dB
10
0
γ0: 40 dB
γ0: 50 dB
9
SNR is about 30 dB to obtain the average probability of error
10−6 . It is also noted that the performance of the two-way
subcarrier intensity-modulated AF relaying system is clearly
degraded with the pointing errors. We believe that the proposed
system could be combined with 5G technologies [47]–[50] and
leave this for our future work.
Future work includes
5
10
A PPENDIX A
0
10
−5
10
0
20
40
60
80
100
J
Fig. 13. Convergence characteristics of the average probability of error with
respect to J without pointing errors. Upper bound for the Moderate regime.
symbol error probability of 10−6 .
Figs. 11 and 12 present the average probability of error
performances of the two-way subcarrier intensity-modulated
AF relaying system considering the pointing error effects.
Fig. 11 shows the average probability of error performances for
various values of the normalized beamwidth wz /r = 5, 10, 15.
For the pointing errors, the normalized jitter σs /r is set to
0.1. Fig. 12 depicts the upper bound performance of the
average probability of error for the strong regime varying the
normalized beamwidths and normalized jitters at the same
time. The SNR is set to 50 dB. It is clearly seen in the
figures that a better average probability of error performance
is achieved by using a narrow beamwidth and a small jitter.
Fig. 13 shows the convergence performance of the derived
average probability of error with respect to the varying number
of power series J. The pointing error was not considered.
Here the upper bound of the average probability of error is
considered for the moderate regime. It is observed in the figure
that the series term J = 2 is enough to make the average
probability of error converge in the high SNR of 50 dB.
VI. C ONCLUSION
In this paper, we derived the upper and lower bounds of
the overall outage probability and average error probability of
a two-way subcarrier intensity-modulated AF relaying system
over FSO channels considering attenuations caused by atmospheric turbulence and geometric spread and pointing errors at
the same time. The derived performances are based on generalized infinite power series expressions to model the GammaGamma fading distribution of the FSO channel. According to
the analysis results, the overall outage probability and average
probability of error have only about 2.0 dB and 1.8 dB gaps
between the upper and lower bounds. It is noted that this twoway subcarrier intensity-modulated AF relaying system using
a BPSK modulation could be used for practical applications
in the case of a weak turbulence regime in which the required
If we substitute (22) and (27) into (31), the average probability of errors corresponding to the upper and lower bounds of
the overall outage probability can be expressed, respectively,
as
Z ∞
x x dx
1
PU (e) = √
e− 2 √ ,
FΓU
(38)
δ
2 x
2π 0
Z ∞
x x dx
1
(39)
FΓL
e− 2 √ .
PL (e) = √
δ
2 x
2π 0
Using (3.361-2) of [46],
Z ∞
0
e−qx
√ dx =
x
r
π
,
q
(40)
the average probability of errors corresponding to the upper
and lower bounds of the overall outage probability can be
further expressed as
(
2 2
)
√
η ξ γ0 δ η 2 ξ 2 γ0 δ
1
, (41)
,
2π − A
PU (e) = √
3
3
2 2π
(
2 2
√
1
η ξ γ0 δ η 2 ξ 2 γ0 δ
PL (e) = √
2π − A
,
2
3
2 2π
)
2 2
η 2 ξ 2 γ0 δ η 2 ξ 2 γ0 δ
η ξ γ0 δ η 2 ξ 2 γ0 δ
+A
,
,
,
−A
3
2
3
3
(42)
where,
A(u, v) =
Z
∞
0
Ψ
u v e− x2
√ dx.
,
x x
x
If we substitute (13) into (43), (43) can be given by
√
A(u, v) = 2π − AJ1 (u) − AJ2 (v) + AJ3 (u, v)
where,
r − x
x e 2
√ dx,
=
FI2
u
x
0
r − x
Z ∞
x e 2
√ dx,
FI1
AJ2 (v) =
v
x
0
AJ1 (u)
AJ3 (u, v) =
Z
Z
∞
FI2
0
(43)
(44)
∞
r
r − x
x
x e 2
√ dx.
FI1
u
v
x
(45)
(46)
(47)
10
If we substitute (12) into (45) and (46), AJ1 (u) and AJ2 (v) can
be derived as, respectively,
AJ1 (u) =
J
X
2
γ
aj (α, β)u
lim
A0 J→∞ j=0 (j + β)(γ 2 − j − β)A0j+β−1
Z ∞
j+β+1
x
×
x 2 −1 e− 2 dx
J
X
(48)
− j+β
2
0
j+α
J
X
γ2
aj (β, α)v − 2
+
lim
A0 J→∞ j=0 (j + α)(γ 2 − j − α)Aj+α−1
0
Z ∞
j+α+1
x
×
x 2 −1 e− 2 dx.
(49)
0
If we apply the gamma function [35, eq. (3.381-4)] to (48) and
(49)
Z ∞
1
xn−1 e−mx dx = n Γ(n),
(50)
m
0
then, A1 (u) and A2 (v) can be finally obtained as given by
(35) and (36). If we substitute (12) into the product of CDFs
in (47), it can be expressed as
r
r
x
x
FI2
FI1
u
v
2 2
j+β
J
X
aj (α, β)u− 2
γ
=
lim
xj/2
2 )Aj+β−1
J→∞
A0
(j
+
β)(j
+
β
−
γ
0
j=0
×
+
2
γ
A0
×
+
J
X
j=0
2
γ
A0
×
j=0
2
2
J
X
j=0
aj (α, β)v −
j+β
2
(j + β)(j + β − γ 2 )A0j+β−1
lim
J→∞
J
X
j=0
aj (α, β)u
(j + α)(j + α −
lim
J→∞
J
X
j=0
xj/2 xβ
j/2
x
j+α
2
α+β
xj/2 x 2
j+α−1
γ 2 )A0
aj (β, α)u
− j+α
2
(j + α)(j + α − γ 2 )Aj+α−1
0
aj (α, β)v −
j+β
2
xj/2 x
(j + β)(j + β − γ 2 )A0j+β−1
bj y j
lim
J→∞
J
X
j=0
J
X
j=0
aj (β, α)u−
j+α
2
(j + α)(j + α − γ 2 )Aj+α−1
0
aj (β, α)v −
j+α
2
(j + α)(j + α − γ 2 )Aj+α−1
0
xj/2
xj/2 xα .
α+β
2
∞
X
cj y j , cj =
j=0
j
X
bk dj−k ,
(52)
k=0
r
r
x
x
FI2
FI1
u
v
2 2
J
X
j
γ
(1p)
cj (u, v)x 2 +β
=
lim
J→∞
A0
j=0
2 2
J
X
α+β
j
γ
(2p)
cj (u, v)x 2 + 2
+
lim
J→∞
A0
j=0
2 2
J
X
j
α+β
γ
(3p)
cj (u, v)x 2 + 2
+
lim
J→∞
A0
j=0
2 2
J
X
j
γ
(4p)
cj (u, v)x 2 +α
lim
+
J→∞
A0
j=0
(53)
where,
(1p)
cj
(u, v) =
j
X
k=0
×
ak (α, β)u−
k+β
2
(k + β)(k + β − γ 2 )Ak+β−1
0
aj−k (α, β)v −
j−k+β
2
(j − k + β)(j − k + β − γ 2 )A0j−k+β−1
k+β
j
X
ak (α, β)u− 2
(2p)
cj (u, v) =
k+β−1
2
k=0 (k + β)(k + β − γ )A0
j−k+α
2
aj−k (β, α)v −
(j − k + α)(j − k + α − γ 2 )A0j−k+α−1
k+α
j
X
ak (β, α)u− 2
(3p)
cj (u, v) =
(k + α)(k + α − γ 2 )Ak+α−1
0
k=0
aj−k (α, β)v −
j−k+β
2
(j − k + β)(j − k + β − γ 2 )A0j−k+β−1
k+α
j
X
ak (β, α)u− 2
(4p)
cj (u, v) =
(k + α)(k + α − γ 2 )Ak+α−1
0
k=0
×
(51)
dj y j =
the product of CDFs can be expressed as
×
xj/2
∞
X
j=0
×
− j+β
2
(j + β)(j + β − γ 2 )A0j+β−1
aj (β, α)v −
∞
X
j=0
γ
aj (α, β)v
lim
A0 J→∞ j=0 (j + β)(γ 2 − j − β)A0j+β−1
Z ∞
j+β+1
x
×
x 2 −1 e− 2 dx
J
X
2
If we apply the multiplication of power series [35, eq. (0.316)]
0
AJ2 (v) =
γ2
A0
×
j+α
J
X
γ2
aj (β, α)u− 2
lim
A0 J→∞ j=0 (j + α)(γ 2 − j − α)Aj+α−1
0
Z ∞
j+α+1
x
×
x 2 −1 e− 2 dx,
2
− j+β
2
0
+
+
aj−k (β, α)v −
j−k+α
2
(j − k + α)(j − k + α − γ 2 )A0j−k+α−1
.
(54)
PARK et al.: TWO-WAY AMPLIFY-AND-FORWARD RELAYING SYSTEM ...
11
If we substitute the product of CDFs (53) into (47), A3 (u, v)
can be derived as
[12]
H. G. Sandalidis, T. A. Tsiftsis, G. K. Karagiannidis, and M. Uysal, “Ber
performance of fso links over strong atmospheric turbulence channels
with pointing errors,” IEEE Comm. Lett., vol. 12, no. 1, pp. 44–46,
2008.
[13]
H. G. Sandalidis, T. A. Tsiftsis, and G. K. Karagiannidis, “Optical
wireless communications with heterodyne detection over turbulence
channels with pointing errors,” J. Lightwave Technol., vol. 27, no. 20,
pp. 4440–4445, 2009.
AJ3 (u, v)
2 2
Z ∞
J
X
j+2β+1
γ
x
(1p)
=
x 2 −1 e− 2 dx
cj (u, v)
lim
J→∞
A0
0
j=0
2 2
Z
J
∞
X (2p)
j+α+β+1
x
γ
x 2 −1 e− 2 dx
cj (u, v)
+
lim
J→∞
A0
0
j=0
2 2
Z
J
∞
X (3p)
j+α+β+1
x
γ
x 2 −1 e− 2 dx
cj (u, v)
+
lim
J→∞
A0
0
j=0
2 2
Z
J
∞
X (4p)
j+2α+1
γ
x
+
x 2 −1 e− 2 dx.
cj (u, v)
lim
J→∞
A0
0
j=0
[14]
A. A. Farid and S. Hranilovic, “Outage capacity optimization for freespace optical links with pointing errors,” J. Lightwave Technol., vol. 25,
no. 7, pp. 1702–1710, 2007.
[15]
W. Gappmair, “Further results on the capacity of free-space optical
channels in turbulent atmosphere,” IET Comm., vol. 5, no. 9, pp. 1262–
1267, 2011.
[16]
J. Park, E. Lee, C.-B. Chae, and G. Yoon, “Impact of pointing errors on
the performance of coherent free-space optical systems,” IEEE Photon.
Technol. Lett., vol. 28, no. 2, pp. 181–184, 2016.
[17]
If we apply the gamma function [35, eq. (3.381-4)] to (55),
A3 (u, v) can be finally obtained as given by (37).
T. A. Tsiftsis, H. G. Sandalidis, G. K. Karagiannidis, and N. C. Sagias,
“Multihop free-space optical communications over strong turbulence
channels,” in Proc. IEEE Int. Conf. on Comm., vol. 6, 2006, pp. 2755–
2759.
[18]
M. Safari and M. Uysal, “Relay-assisted free-space optical communication,” IEEE Trans. Wireless Comm., vol. 7, no. 12, pp. 5441–5449,
Dec. 2008.
[19]
E. Lee, J. Park, D. Han, and G. Yoon, “Performance analysis of the
asymmetric dual-hop relay transmission with mixed RF/FSO links,”
IEEE Photon. Technol. Lett., vol. 23, no. 21, pp. 1642–1644, Nov. 2011.
[20]
I. Ansari, F. Yilmaz, and M. Alouini, “Impact of pointing errors on the
performance of mixed RF/FSO dual-hop transmission systems,” IEEE
Wireless Comm. Lett., vol. 2, pp. 351–354, June 2013.
[21]
I. S. Ansari, M.-S. Alouini, and F. Yilmaz, “On the performance of
hybrid RF and RF/FSO fixed gain dual-hop transmission systems,” in
Proc. IEEE Electronics, Comm. and Photon. Conf. (SIECPC), Saudi
Int., 2013, pp. 1–6.
[22]
I. S. Ansari, F. Yilmaz, and M.-S. Alouini, “On the performance
of mixed RF/FSO variable gain dual-hop transmission systems with
pointing errors,” in Proc. IEEE Veh. Technol. Conf., 2013, pp. 1–5.
[23]
J. N. Laneman, D. N. Tse, and G. W. Wornell, “Cooperative diversity
in wireless networks: Efficient protocols and outage behavior,” IEEE
Trans. Inform. Theory, vol. 50, no. 12, pp. 3062–3080, Dec. 2004.
[24]
J. Park, E. Lee, C.-B. Chae, and G. Yoon, “Outage probability analysis
of a coherent fso amplify-and-forward relaying system,” IEEE Photon.
Technol. Lett., vol. 27, no. 11, pp. 1204–1207, 2015.
[25]
M. Hasna and M.-S. Alouini, “A performance study of dual-hop
transmissions with fixed gain relays,” IEEE Trans. Wireless Comm.,
vol. 3, no. 6, pp. 1963–1968, Nov. 2004.
[26]
J. Park, E. Lee, G. Park, B. Roh, and G. Yoon, “Performance analysis of
asymmetric RF/FSO dual-hop relaying systems for UAV applications,”
in Proc. IEEE Mil. Comm. Conf., 2013, pp. 1651–1656.
[27]
H. A. Suraweera, R. H. Y. Louie, Y. Li, G. K. Karagiannidis, and
B. Vucetic, “Two hop amplify-and-forward transmission in mixed
Rayleigh and Rician fading channels,” IEEE Comm. Lett., vol. 13, no. 4,
pp. 227–229, April 2009.
[28]
C.-B. Chae, T. Tang, R. W. Heath, Jr., and S. Cho, “MIMO relaying with
linear processing for multiuser transmission in fixed relay networks,”
IEEE Trans. Sig. Proc., vol. 56, no. 2, pp. 727–738, Feb. 2008.
[29]
X. Xia, D. Zhang, K. Xu, and Y. Xu, “A comparative study on
interference-limited two-way transmission protocols," IEEE/KICS Jour.
of Comm. and Networks, vol. 18, no. 3, pp. 351–363, Sep. 2016.
[30]
J. Park, I. Song, S. Lee, Y. Kim, “Average rate performance of twoway amplify-and-forward relaying in asymmetric fading channels,"
IEEE/KICS Jour. of Comm. and Networks, vol. 13, no. 3, pp. 250–256,
June 2011.
[31]
P. K. Upadhyay and S. Prakriya, “Performance of two-way opportunistic
(55)
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
H. E. Nistazakis, T. A. Tsiftsis, and G. S. Tombras, “Performance
analysis of free-space optical communication systems over atmospheric
turbulence channels,” IET Comm., vol. 3, no. 8, pp. 1402–1409, 2009.
J. Park, E. Lee, and G. Yoon, “Average bit error rate of the Alamouti
scheme in Gamma-gamma fading channels,” IEEE Photon. Technol.
Lett., vol. 23, no. 4, pp. 269–271, Feb. 2011.
M. Uysal, J. Li, and M. Yu, “Error rate performance analysis of coded
free-space optical links over Gamma-Gamma atmouspheric turbulence
channels,” IEEE Trans. Wireless Comm., vol. 5, no. 6, pp. 1229–1233,
June 2006.
T. A. Tsiftsis, H. G. Sandalidis, G. K. Karagiannidis, and M. Uysal,
“Optical wireless links with spatial diversity over strong atmospheric
turbulence channels,” IEEE Trans. Wireless Comm., vol. 8, no. 2, pp.
951–957, Feb. 2009.
J. Park, E. Lee, C.-B. Chae, and G. Yoon, “Performance analysis
of coherent free-space optical systems with multiple receivers,” IEEE
Photon. Technol. Lett., vol. 27, no. 9, pp. 1010–1013, 2015.
W. O. Popoola and Z. Ghassemlooy, “BPSK subcarrier intensity modulated free-space optical communications in atmouspheric turbulence,”
J. Lightwave Technol., vol. 27, no. 8, pp. 967–973, April 2009.
J. Li, J. Q. Liu, and D. P. Taylor, “Optical communication using
subcarrier PSK intensity modulation through atmospheric turbulaence
channels,” IEEE Trans. Comm., vol. 55, no. 8, pp. 1598–1606, Aug.
2007.
W. O. Popoola, Z. Ghassemlooy, H. Haas, E. Leitgeb, and V. Ahmadi,
“Error performance of terrestrial free space optical links with subcarrier
time diversity,” IET Comm., vol. 6, no. 5, pp. 499–506, 2012.
K. P. Peppas and C. K. Datsikas, “Average symbol error probability of
general-order rectangular quadrature amplitude modulation of optical
wireless communication systems over atmospheric turbulence channels,” IEEE J. Opt. Comm. and Networking, vol. 2, no. 2, pp. 102–110,
Feb. 2010.
H. Samimi and P. Azmi, “Subcarrier intensity modulated free-space
optical communications in K-distributed turbulence channels,” IEEE J.
Opt. Comm. and Networking, vol. 2, no. 8, pp. 625–632, July 2010.
X. Song and J. Cheng, “Optical communication using subcarrier intensity modulation in strong atmospheric turbulence,” J. Lightwave
Technol., vol. 30, no. 22, pp. 3484–3493, 2012.
12
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
relaying with analog network coding over Nakagami-fading,” IEEE
Trans. on Veh. Technol., vol. 60, no. 4, pp. 1965–1971, May 2011.
H. Guo, J. Ge, and H. Ding, “Symbol error probability of two-way
amplify-and-forward relaying,” IEEE Comm. Lett., vol. 15, no. 1, pp.
22–24, Jan. 2011.
Y.-U. Jang and Y. H. Lee, “Performance analysis of user selection
for multiuser two-way amplify-and-forward relay,” IEEE Comm. Lett.,
vol. 14, no. 11, pp. 1086–1088, Nov. 2010.
Y. Han, S. H. Ting, C. K. Ho, and W. H. Chin, “Performance bounds for
two-way amplify-and-forward relaying,” IEEE Trans. Wireless Comm.,
vol. 8, no. 1, pp. 432–439, 2009.
J. Yang, P. Fan, T. Q. Duong, and X. Lei, “Exact performance of
two-way af relaying in nakagami-m fading environment,” IEEE Trans.
Wireless Comm., vol. 10, no. 3, pp. 980–987, 2011.
S. S. Ikki and S. Aissa, “Performance analysis of two-way amplifyand-forward relaying in the presence of co-channel interferences,” IEEE
Trans. Comm., vol. 60, no. 4, pp. 933–939, 2012.
Y. Tang, X. Zhou, Z. Zhang, and Q. Tian, “Performance analysis of
a two-way network-coded free space optical relay scheme over strong
turbulence channels,” in Proc. IEEE Veh. Technol. Conf., 2011, pp. 1–5.
P. Puri, M. A. P. Garg, and P. K. Sharma, “Outage analysis of twoway relay assisted FSO systems over weak turbulence region,” in Proc.
IEEE India Conf. (INDICON), 2013, pp. 1–5.
P. Puri, P. Garg, and M. Aggarwal, “Outage and error rate analysis of
network-coded coherent twr-fso systems,” IEEE Photon. Technol. Lett.,
vol. 26, no. 18, pp. 1797–1800, 2014.
——, “Partial dual-relay selection protocols in two-way relayed fso
networks,” J. Lightwave Technol., vol. 33, no. 21, pp. 4457–4463, 2015.
P. Puri, N. D. Chatzidiamantis, P. Garg, M. Aggarwal, and G. K. Karagiannidis, “Two-way relay selection in multiple relayed fso networks,”
IEEE Wireless Comm. Lett., vol. 4, no. 5, pp. 485–488, 2015.
P. Puri, P. Garg, and M. Aggarwal, “Asymptotic analysis of TWR
assisted FSO links with partial dual-relay selection,” IEEE Comm. Lett.,
vol. 19, no. 5, pp. 879–882, 2015.
Z. Abu-Almaalie, Z. Ghassemlooy, H. Le-Minh, and N. Aslam, “Physical layer network coding with two-way relay free space optical communication link,” in Internet Technologies and Applications (ITA), 2015.
IEEE, 2015, pp. 292–297.
E. Bayaki, R. Schober, and R. K. Mallik, “Performance analysis of
MIMO free-space optical systems in Gamma-Gamma fading,” IEEE
Trans. Comm., vol. 57, no. 11, pp. 3415–3424, Nov. 2009.
M. K. Simon and M.-S. Alouini, Digital communication over fading
channels. John Wiley and Sons, Inc., 2000.
I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
Products, 7th ed., A. Jeffrey and D. Zwillinger, Eds. Academic Press,
2007.
M. Chung, M. S. Sim, J. Kim, D.-K. Kim, and C.-B. Chae, “Prototyping
real-time full duplex radios,” IEEE Comm. Mag. , vol. 53, no. 9, pp.
56–64, Sep. 2015.
Y.-G. Lim, C.-B. Chae, and G. Caire, “Performance analysis of massive
MIMO for cell-boundary users," IEEE Trans. Wireless Comm., vol. 14,
no. 12, pp. 6827–6842, Dec. 2015.
M. S. Sim, J. Park, C.-B. Chae, and R. W. Heath, Jr., “Compressed
channel feedback for correlated massive MIMO systems," IEEE/KICS
Jour. of Comm. and Networks, vol. 18, no. 1, pp. 95-104, Feb. 2016.
H. B. Yilmaz, B. Koo, S. Park, H. Park, J. Han, and C.-B. Chae, “Frequency assignment problem with net filter discrimination constraints,"
to appear in IEEE/KICS Jour. of Comm. and Networks, Aug. 2017
PLACE
PHOTO
HERE
Jaedon Park received the B.S. degree in Electronics Engineering from Hanyang University, Seoul,
Korea in 2000, and the M.S. and Ph.D. degrees
in the School of Electrical Engineering from the
Korea Advanced Institute of Science and Technology
(KAIST), Daejeon, Korea in 2002 and 2016, respectively. Currently, he is a Senior Researcher in the
Agency for Defense Development (ADD), Daejeon,
Korea. His research interests are in MIMO systems,
relay systems, and FSO systems.
Chan-Byoung Chae (S’06 - M’09 - SM’12) is the
Underwood Distinguished Professor, in the School
of Integrated Technology, College of Engineering,
Yonsei University, Korea. He was a Member of TechPLACE
nical Staff (Research Scientist) at Bell Laboratories,
PHOTO
Alcatel-Lucent, Murray Hill, NJ, USA from 2009 to
HERE
2011. Before joining Bell Laboratories, he was with
the School of Engineering and Applied Sciences at
Harvard University, Cambridge, MA, USA as a PostDoctoral Research Fellow. He received the Ph.D.
degree in Electrical and Computer Engineering from
The University of Texas (UT), Austin, TX, USA in 2008, where he was a
member of the Wireless Networking and Communications Group (WNCG).
Prior to joining UT, he was a Research Engineer at the Telecommunications
R&D Center, Samsung Electronics, Suwon, Korea, from 2001 to 2005. While
having worked at Samsung, he participated in the IEEE 802.16e (mobile
WiMAX) standardization, where he made several contributions and filed a
number of related patents from 2004 to 2005. His current research interests
include capacity analysis and interference management in energy-efficient
wireless mobile networks and nano (molecular) communications. He has
served/serves as an Editor for the IEEE C OMM . M AG ., the IEEE W IRELESS
C OMM . L ETTERS , the IEEE T RANS . ON W IRELESS C OMM ., the IEEE
T RANS . ON M OLECULAR , B IOLOGICAL , M ULTI - SCALE C OMM ., the IEEE
T RANS . ON S MART G RID, and the IEEE/KICS J OUR . C OMM . N ETS . He was
a Guest Editor for the IEEE J OUR . S EL . A REAS IN C OMM . (special issue on
molecular, biological, and multi-scale comm.). He is an IEEE Senior Member.
Dr. Chae was the recipient/co-recipient of the Outstanding Teaching Award
(2017) from Yonsei University, the Underwood Distinguished Professor Award
(2016), the Yonam Research Award from LG Yonam Foundation (2016), the
Best Young Professor Award from the College of Engineering, Yonsei University (2015), the IEEE INFOCOM Best Demo Award (2015), the IEIE/IEEE
Joint Award for Young IT Engineer of the Year (2014), the Haedong Young
Scholar Award (2013), the IEEE Signal Processing Magazine Best Paper
Award (2013), the IEEE ComSoc AP Outstanding Young Researcher Award
(2012), the IEEE VTS Dan. E. Noble Fellowship Award (2008), the Gold Prize
(1st) in the 14th/19th Humantech Paper Contests, and the KSEA-KUSCO
scholarship (2007). He also received the Korea Government Fellowship
(KOSEF) during his Ph.D. studies.
PARK et al.: TWO-WAY AMPLIFY-AND-FORWARD RELAYING SYSTEM ...
Giwan Yoon received the B.S. degree from Seoul
National University (SNU), Seoul, Korea, in 1983,
the M.S. degree from the Korea Advanced Institute
of Science and Technology (KAIST), Seoul, Korea,
PLACE
in 1985, and the Ph.D. degree from the University
PHOTO
of Texas at Austin, USA, in 1994. From 1985 to
HERE
1990, he was employed as an engineer at LG Group,
Seoul, Korea, where he worked for the development
of high-speed bipolar transistors. From 1994 to 1997,
he was employed as a senior engineer at Digital
Equipment Corporation (DEC), MA, USA, where
he developed oxynitride gate dielectric CMOS devices. From 1997 to 2009,
he was a faculty member of Information and Communications University,
Daejeon, Korea, where he developed high-frequency devices for RF and
wireless communications. Since 2009, he has been with the KAIST, where he
is currently a professor in the School of Electrical Engineering with teaching
and research activities in the areas of nano devices and integrated systems,
energy generation & harvesting devices, and flexible/wearable sensing devices
for healthcare, IOT and sensor networks applications. Dr. Yoon is a member
of the IEEE.
13
| 7cs.IT
|
Complexity Analysis
in Presence of Control Operators
and Higher-Order Functions
arXiv:1310.1763v1 [cs.LO] 7 Oct 2013
Ugo Dal Lago
Giulio Pellitta
Abstract
A polarized version of Girard, Scedrov and Scott’s Bounded Linear Logic is introduced and
its normalization properties studied. Following Laurent [26], the logic naturally gives rise to a
type system for the λµ-calculus, whose derivations reveal bounds on the time complexity of the
underlying term. This is the first example of a type system for the λµ-calculus guaranteeing
time complexity bounds for typable programs.
1
Introduction
Among non-functional properties of programs, bounds on the amount of resources (like computation time and space) programs need when executed are particularly significant. The problem
of deriving such bounds is indeed crucial in safety-critical systems, but is undecidable whenever
non-trivial programming languages are considered. If the units of measurement become concrete
and close to the physical ones, the problem becomes even more complicated and architecturedependent. A typical example is the one of WCET techniques adopted in real-time systems [32],
which not only need to deal with how many machine instructions a program corresponds to, but
also with how much time each instruction costs when executed by possibly complex architectures
(including caches, pipelining, etc.), a task which is becoming even harder with the current trend
towards multicore architectures.
A different approach consists in analysing the abstract complexity of programs. As an example,
one can take the number of instructions executed by the program as a measure of its execution time.
This is of course a less informative metric, which however becomes more accurate if the actual time
taken by each instruction is kept low. One advantage of this analysis is the independence from the
specific hardware platform executing the program at hand: the latter only needs to be analysed
once. A variety of complexity analysis techniques have been employed in this context, from abstract
interpretation [21] to type systems [22] to program logics [10] to interactive theorem proving.
Properties of programs written in higher-order functional languages are for various reasons wellsuited to be verified by way of type systems. This includes not only safety properties (e.g. welltyped programs do not go wrong), but more complex ones, including resource bounds [22, 5, 15, 7].
In this paper, we delineate a methodology for complexity analysis of higher-order programs
with control operators. The latter are constructs which are available in most concrete functional
programming languages (including Scheme and OCaml), and allow control to flow in non-standard
ways. The technique we introduce takes the form of a type system for de Groote’s λµ-calculus [12]
derived from Girard, Scedrov and Scott’s Bounded Linear Logic [19] (BLL in the following). We
prove it to be sound: typable programs can indeed be reduced in a number of steps lesser or equal
to a (polynomial) bound which can be read from the underlying type derivation. A similar result
can be given when the cost model is the one induced by an abstract machine. To the authors’
knowledge, this is the first example of a complexity analysis methodology coping well not only
with higher-order functions, but also with control operators.
In the rest of this section, we explain the crucial role Linear Logic has in this work, in the
meantime delineating its main features.
1
1.1
Linear Logic and Complexity Analysis
Linear Logic [16] is one of the most successful tools for characterizing complexity classes in a
higher-order setting, through the Curry-Howard correspondence. Subsystems of it can indeed be
shown to correspond to the polynomial time computable functions [19, 18, 23] or the logarithmic
space computable functions [30]. Many of the introduced fragments can then be turned into type
systems for the λ-calculus [5, 15], some of them being relatively complete in an intensional sense [7].
The reason for this success lies in the way Linear Logic decomposes intuitionistic implication
into a linear implication, which has low complexity, and an exponential modality, which marks those
formulas to which structural rules can be applied. This gives a proper status to proof duplication,
without which cut-elimination can be performed in a linear number of steps. By tuning the rules
governing the exponential modality, then, one can define logical systems for which cut-elimination
can be performed within appropriate resource bounds. Usually, this is coupled with an encoding
of all functions in a complexity class C into the system at hand, which makes the system a
characterization of C.
Rules governing the exponential modality ! can be constrained in (at least) two different
ways:
• On the one hand, one or more of the rules governing ! (e.g., dereliction or digging) can be
dropped or restricted. This is what happens, for example, in Light Linear Logic [18] or Soft
Linear Logic [23].
• On the other, the logic can be further refined and enriched so as to control the number of times
structural rules are applied. In other words, rules for ! are still all there, but in a refined form.
This is what happens in Bounded Linear Logic [19]. Similarly, one could control so-called
modal impredicativity by a system of levels [4].
The first approach corresponds to cutting the space of proofs with an axe: many proofs, and
among them many corresponding to efficient algorithms, will not be part of the system because
they require one of the forbidden logical principles. The second approach is milder in terms of the
class of good programs that are “left behind”: there is strong evidence that with this approach
one can obtain a quite expressive logical system [8, 7].
Not much is known about whether this approach scales to languages in which not only functions
but also first-class continuations and control operators are present. Understanding the impact of
these features to the complexity of programs is an interesting research topic, which however has
received little attention in the past.
1.2
Linear Logic and Control Operators
On the other hand, more than twenty years have passed since classical logic has been shown to be
amenable to the Curry-Howard paradigm [20]. And, interestingly enough, classical axioms (e.g.
Pierce’s law or the law of the Excluded Middle) can be seen as the type of control operators like
Scheme’s callcc. In the meantime, the various facets of this new form of proofs-as-programs
correspondence have been investigated in detail, and many extensions of the λ-calculus for which
classical logic naturally provides a type discipline have been introduced (e.g. [28, 6]).
Moreover, the decomposition provided by Linear Logic is known to scale up to classical
logic [17]. Actually, Linear Logic was known to admit an involutive notion of negation from
its very inception [16]. A satisfying embedding of Classical Logic into Linear Logic, however,
requires restricting the latter by way of polarities [24]: this way one is left with a logical system
with most of the desirable dynamical properties.
In this paper, we define BLLP, a polarized version of Bounded Linear Logic. The kind of
enrichment resource polynomials provide in BLL is shown to cope well with polarization. Following
the close relationship between polarized linear logic and the λµ-calculus [26], BLLP gives rise to
a type system for the λµ-calculus. Proofs and typable λµ-terms are both shown to be reducible
to their cut-free or normal forms in a number of steps bounded by a polynomial weight. Such
a result for the former translates to a similar result for the latter, since any reduction step in
λµ-terms corresponds to one or more reduction steps in proofs. The analysis is then extended to
2
the reduction of λµ-terms by a Krivine-style abstract machine [13].
2
Bounded Polarized Linear Logic as A Sequent Calculus
In this section, we define BLLP as a sequent calculus. Although this section is self-contained, some
familiarity with both bounded [19] and polarized [26] linear logic would certainly help.
2.1
Polynomials and Formulas
m
Y
xi
, where
A resource monomial is any (finite) product of binomial coefficients in the form
ni
i=1
the xi are distinct variables and the ni are non-negative integers. A resource polynomial is any
finite sum of resource monomials. Given resource polynomials p, q write p ⊑ q to denote that q − p
is a resource polynomial. If p ⊑ r and q ⊑ s then also q ◦ p ⊑ s ◦ r. Resource polynomials are
closed by addition, multiplication, bounded sums and composition [19].
A polarized formula is a formula (either positive or negative) generated by the following grammar
P ::= V (~
p)
|
N ::= V ⊥ (~
p)
|
1
|
∃V P
N `N
|
⊥
|
P ⊗P
|
|
∀V N
!x<p N ;
|
?x<p P.
where V ranges over a countable sets of atoms. Throughout this paper, formulas (but also terms,
contexts, etc.) are considered modulo α-equivalence. Formulas (either positive or negative) are
ranged over by metavariables like A, B. Formulas like V ⊥ are sometime denoted as X, Y .
In a polarized setting, contraction can be performed on any negative formula. As a consequence, we need the notion of a labelled1 formula [A]px , namely the labelling of the formula A with
respect to x and p. All occurrences of x in A are bound in [A]px . Metavariables for labellings of
positive (respectively, negative) formulas are P, Q, R (respectively, N, M, L). Labelled formulas
are sometimes denoted with metavariables A, B when their polarity is not essential. Negation, as
usual in classical linear system, can be applied to any (possibly labelled) formula, à la De Morgan.
When the resource variable x does not appear in A, then we do not need to mention it when
writing [A]px , which becomes [A]p . Similarly for !x<p N and ?x<p P .
Both the space of formulas and the space of labelled formulas can be seen as partial orders by
stipulating that two (labelled) formulas can be compared iff they have exactly the same skeleton
and the polynomials occurring in them can be compared. Formally,
V (p1 , . . . , pn ) ⊑ V (q1 , . . . , qn ) iff ∀i.pi ⊑ qi ;
V ⊥ (p1 , . . . , pn ) ⊑ V ⊥ (q1 , . . . , qn ) iff ∀i.qi ⊑ pi ;
1 ⊑ 1;
⊥ ⊑ ⊥;
P ⊗ Q ⊑ R ⊗ S iff P ⊑ R ∧ Q ⊑ S;
N ` M ⊑ O ` K iff N ⊑ O ∧ M ⊑ K;
!x<p N ⊑!x<q M iff q ⊑ p ∧ N ⊑ M ;
?x<p P ⊑?x<q Q iff p ⊑ q ∧ P ⊑ Q;
∀V.N ⊑ ∀V.M iff N ⊑ M ;
∃V.P ⊑ ∃V.Q iff P ⊑ Q.
1 Keep in mind that linear logic contains a subset of formulas which is isomorphic to (polarized) classical logic.
[N ]px (resp. [P ]px ) can be thought of roughly as ?x<p N ⊥ (resp. !x<p P ⊥ ), i.e., in a sense we can think of labelled
formulas as formulas hiding an implicit exponential modality.
3
In a sense, then, polynomials occurring next to atoms or to the whynot operator are in positive
position, while those occurring next to the bang operator are in negative position. In all the other
cases, ⊑ is defined component-wise, in the natural way, e.g. P ⊗ Q ⊑ R ⊗ S iff both P ⊑ R and
Q ⊑ S. Finally [N ]px ⊑ [M ]qx iff N ⊑ M ∧ p ⊒ q. And dually, [P ]px ⊑ [Q]qx iff N ⊑ M ∧ p ⊑ q.
Lemma 2.1 A ⊑ B iff B ⊥ ⊑ A⊥ . Moreover, A ⊑ B iff B⊥ ⊑ A⊥ .
Proof. A ⊑ B iff B ⊥ ⊑ A⊥ can be proved by induction on the structure of A. Consider the second
part of the statement, now. Suppose that A, B are positive, and call them P, Q respectively. Then
[P ]px ⊑ [Q]qx ⇔ P ⊑ Q ∧ p ⊑ q
⇔ Q⊥ ⊑ P ⊥ ∧ p ⊑ q
⇔ [Q⊥ ]qx ⊑ [P ⊥ ]px .
The case when A, B are negative is similar.
Certain operators on resource polynomials can be lifted to formulas. As an example, we want
to be able to sum labelled formulas provided they have a proper form:
[N ]px ⊎ [N {x/y + p}]qy := [N ]p+q
x .
We are assuming, of course, that x, y are not free in either p or q. This construction can be
generalized to bounded sums: suppose that a labelled formula is in the form
X
[M ]ry = [N {x/y +
r{z/u}}]ry ,
u<z
where y and u are not freePin N nor in r and z is not free in N . Then the labelled formula
P
z<q r
r
. See [19, §3.3] for more details about the above constructions.
z<q [M ]y is defined as [N ]x
An abstraction formula of arity n is simply a formula A, where the n resource variables
x1 , . . . , xn are meant to be bound. A{V := B} is the result of substituting a second order
abstraction term B (of arity n) for all free occurrences of the propositional variable V (of the
same arity) in A. This can be defined formally by induction on the structure of A, but the only
interesting clauses are the following two:
V (p1 , . . . , pn ){V := B} = B{x1 , . . . , x/p1 , . . . , pn }
V ⊥ (p1 , . . . , pn ){V := B} = B ⊥ {x1 , . . . , x/p1 , . . . , pn }
2.2
Sequents and Rules
The easiest way to present BLLP is to give a sequent calculus for it. Actually, proofs will be
structurally identical to proofs of Laurent’s LLP. Of course, only some of LLP proofs are legal
BLLP proofs — those giving rise to an exponential blow-up cannot be decorated according to the
principles of Bounded Linear Logic.
A sequent is an expression in the form ⊢ Γ, where Γ = A1 , . . . An is a multiset of labelled
formulas such that at most one among A1 , . . . , An is positive. If Γ only contains (labellings of)
negative formulas, we indicate it with metavariables like N , M. The operator ⊎ can be extended
to one on multi-sets of formulas component-wise, so we can write expressions like N ⊎ M: this
amounts to summing the polynomials occurring in N and those occurring in M. Similarly for
bounded sums.
The rules of the sequent calculus for BLLP are in Figure 1. Please observe that:
• The relation ⊑ is implicitly applied to both formulas and polynomials whenever possible in
such a way that “smaller” formulas can always be derived (see Section 2.3).
• As in LLP, structural rules can act on any negative formula, and not only on exponential ones.
Since all formulas occurring in sequents are labelled, however, we can still keep track of how
many times formulas are “used”, in the spirit of BLL.
4
N⊑M
M⊥ ⊒ P
Ax
⊢ N, P
⊢ Γ, [N ]px , [M ]qx
p⊑r q⊑r
`
⊢ Γ, [N ` M ]rx
P
⊢ N , [N ]px
M ⊑ y<q N
!
⊢ M, [!x<p N ]qy
⊢ Γ, N, M
⊢Γ
?w
⊢ Γ, N
⊢ Γ, N
⊢ N , [P ]px
⊢ N , N⊥
Cut
⊢ Γ, N
⊢ M, [Q]qx r ⊑ p r ⊑ q
⊗
⊢ N , M, [P ⊗ Q]rx
p{y/0}
⊢ N , [P {y/0}]x
N ⊑ [?x<p P ]1y
⊢ N,N
L⊑N⊎M
?c
⊢ Γ, L
⊢ Γ, [N ]px
V 6∈ FV (N )
∀
⊢ Γ, [∀V N ]px
⊢Γ
⊥
⊢ Γ, [⊥]px
⊢ [1]px
?d
1
⊢ N , [P {V := Q}]px
∃
⊢ N , [∃V P ]px
Figure 1: BLLP, Sequent Calculus Rules
• A byproduct of taking sequents as multisets of labeled formulas is that multiplicative rules
themselves need to deal with labels. As an example, consider rule ⊗: the resource polynomial
labelling the conclusion P ⊗ Q is anything smaller or equal to the polynomials labeling the two
premises.
The sequent calculus we have just introduced could be extended with additive logical connectives.
For the sake of simplicity, however, we have kept the language of formulas very simple here.
As already mentioned, BLLP proofs can be seen as obtained by decorating proofs from Laurent’s
LLP [26] with resource polynomials. Given a proof π, hπi is the LLP proof obtained by erasing all
resource polynomials occurring in π. If π and ρ are two BLLP proofs, we write π ∼ ρ iff hπi ≡ hρi,
i.e., iff π and ρ are two decorations of the same LLP proof.
Even if structural rules can be applied to all negative formulas, only certain proofs will be
copied or erased along the cut-elimination process, as we will soon realize. A box is any proof
which ends with an occurrence of the ! rule. In non-polarized systems, only boxes can be copied
or erased, while here the process can be applied to ⊗-trees, which are proofs inductively defined
as follows:
• Either the last rule in the proof is Ax or ! or 1;
• or the proof is obtained from two ⊗-trees by applying the rule ⊗.
A ⊗-tree is said to be closed if it does not contain any axiom nor any box having auxiliary doors
(i.e., no formula in the context of the ! rules).
2.3
Malleability
The main reason for the strong (intensional) expressive power of BLL [8] is its malleability: the
conclusion of any proof π can be modified in many different ways without altering its structure.
Malleability is not only crucial to make the system expressive, but also to prove that BLLP enjoys
cut-elimination. In this section, we give four different ways of modifying a sequent in such a way
as to preserve its derivability. Two of them are anyway expected and also hold in BLL, while the
other two only make sense in a polarized setting.
First of all, taking smaller formulas (i.e., more general — cf. [19, §3.3, p. 21]) preserves
derivability:
Lemma 2.2 (Subtyping) If π ⊲ ⊢ Γ, A and A ⊒ B, then there is ρ ⊲ ⊢ Γ, B such that π ∼ ρ.
Proof. By a simple induction on π. The crucial cases:
• If the last rule used is an axiom:
N⊑M
M⊥ ⊒ P
Ax
⊢ N, P
5
If B ⊑ N, then we know that B ⊑ N ⊑ M, from which it follows that B ⊑ M. We can thus
take ρ as
B⊑M
M⊥ ⊒ P
Ax
⊢ B, P
If B ⊑ P, then we know that B ⊑ P ⊑ M⊥ , from which it follows that B ⊑ M⊥ . We can thus
take ρ as
B⊑M
M⊥ ⊒ B
Ax
⊢ B, M
• Suppose the last rule used is !:
P
σ ⊲ ⊢ N , [N ]ry
M ⊑ x<q N
!
⊢ M, [!y<r N ]qx
If B ⊑ [!y<r N ]qx , then necessarily B = [!y<s M ]px , where N ⊒ M , q ⊒ p and s ⊒ r. Hence
[N ]ry ⊒ [M ]sy and, by induction hypothesis, there is P
λ such thatP
λ ∼ σ and λ ⊲ ⊢ N , [M ]sy . As
a consequence ρ can be simply defined as, because x<q N ⊑ x<p N :
P
λ ⊲ ⊢ N , [M ]sy
M ⊑ x<p N
!
⊢ M, [!y<s M ]qx
If B ⊑ N ∈ M, then we can just derive the thesis from transitivity of ⊑.
• If the last rule used is ?d:
r{x/0}
σ ⊲ ⊢ N , [P {x/0}]y
N ⊑ [?y<r P ]1x
?d
⊢ N,N
Then the induction hypothesis immediately yields the thesis.
This concludes the proof.
Substituting resource variables for polynomials itself preserves typability:
Lemma 2.3 (Substitution) Let π ⊲ ⊢ Γ. Then there is a proof π{x/p} of ⊢ Γ{x/p}. Moreover,
π{x/p} ∼ π.
Proof. By an easy induction on the structure of π.
Lemma 2.4 [A]px ⊒ [B]px ⇒ [A{x/y + q}]py ⊒ [B{x/y + q}]py .
Proof. [A]px ⊒ [B]px ⇒ A ⊒ B ⇒ A{x/y + q} ⊒ B{x/y + q} ⇒ [A{x/y + q}]py ⊒ [B{x/y + q}]py
As we have already mentioned, one of the key differences between ordinary Linear Logic and
its polarized version is that in the latter, arbitrary proofs can potentially be duplicated (and
erased) along the cut-elimination process, while in the former only special ones, namely boxes,
can. This is, again, a consequence of the fundamentally different nature of structural rules in
the two systems. Since BLLP is a refinement of LLP, this means that the same phenomenon is
expected. But beware: in a bounded setting, contraction is not symmetric, i.e., the two copies of
the proof π we are duplicating are not identical to π.
What we need to prove, then, is that proofs can indeed be split in BLLP: But preliminary to
that is the following technical lemma:
P
P
r
Lemma 2.5 (Shifting Sums) If z<q [M ]ry = [N ]y z<q , then the formula N = [M ]ry {z/z + q}
is such that
X
X P r{z/z+q}
N = [N {x/x +
r}]y z<p
z<p
z<q
6
Proof. The fact
P
r
z<q [M ]y
exists implies that there exist N, x, u such that
M = N {x/y +
X
r{z/u}}
u<z
and y, z ∈
/ FV (N ) and y ∈
/ FV (r). As a consequence:
X
[M ]ry {z/z + q} = [N {x/y +
r{z/u}}{z/z + q}]yr{z/z+q}
u<z
= [N {x/y +
X
r{z/u}}]r{z/z+q}
y
u<z+q
= [N {x/y +
X
r{z/u} +
X
r{z/u}}{x/y +
X
r{z/u}}{x/y +
u<q
= [N {x/x +
X
r{z/u + q}}]yr{z/z+q}
u<z
u<q
= [N {x/x +
r{z/u + q}}]yr{z/z+q}
X
r{z/z + q}{z/u}}]yr{z/z+q}
u<z
u<q
Call the last formula N. As a consequence,
[N {x/x +
X
X
u<z
P
z<p
N exists and is equal to
P
r{z/u}}]y
z<p
r{z/z+q}
.
u<q
This concludes the proof.
Lemma 2.6 (Splitting) If π ⊲ ⊢ N , [P ]px is a ⊗-tree and p ⊒ r + s then there exist M, O such
that ρ ⊲ ⊢ M, [P ]rx , σ ⊲ ⊢ O, [P {x/y + r}]sy . Moreover, N ⊑ M ⊎ O and ρ ∼ π ∼ σ.
Proof. By induction on π:
• If the last rule used is an axiom then it is in the form
[N1 ]qx1 ⊑ [M ]tx
[M ⊥ ]tx ⊒ [P ]px
Ax
⊢ [N ]qx , [P ]px
for some M, t. We know that
r + s ⊑ p ⊑ t ⊑ q.
Observe that, we can form the following derivations
[N1 ]rx ⊑ [M ]rx
[M ⊥ ]rx ⊒ [P ]rx
Ax
⊢ [N1 ]rx , [P ]rx
[M ⊥ {x/y + r}]sy ⊒ [P {x/y + r}]sy
[N {x/y + r}]sy ⊑ [M {x/y + r}]sy
⊢ [N {x/y +
r}]sy , [P {x/y
+ r}]sy
Ax
where in building the second one we made use, in particular, of Lemma 2.4.
• If the last rule used is ⊗ then we can write π as
λ1 ⊲ ⊢ N1 , [P1 ]px1
λ2 ⊲ ⊢ N2 , [P2 ]px2
⊗
⊢ N1 , N2 , [P1 ⊗ P2 ]px
where p ⊑ p1 and p ⊑ p2 . As a consequence, p1 ⊒ q + r and p2 ⊒ q + r, and we can thus apply
the induction hypothesis to λ1 , λ2 easily reaching the thesis.
• If the last rule used is promotion ! then π has the following shape:
P
λ ⊲ ⊢ N , [N ]qz
M ⊑ x<r N
!
⊢ M, [!z<q N ]px
Then ρ is simply
7
λ ⊲ ⊢ N , [N ]qz
P
!
⊢ x<r N , [!z<q N ]rx
About σ, observe that λ{x/y + r} has conclusion
⊢ N {x/y + r}, [N {x/y + r}]q{x/y+r}
z
P
By Lemma 2.5, it is allowed to form y<s N {x/y + r}. As a consequence, σ is
q{x/y+r}
⊢ N {x/y + r}, [N {x/y + r}]z
P
!
⊢ y<s N {x/y + r}, [(!z<q N ){x/y + r}]sy
Observe that the conclusions of ρ and σ are in the appropriate relation, again because of
Lemma 2.5.
This concludes the proof.
Observe that not every proof can be split, but only ⊗-trees can. A parametric version of splitting
is also necessary here:
P
Lemma 2.7 (Parametric Splitting)P
If π ⊲ ⊢ N , [P ]px , where π is a ⊗-tree and p ⊒ x<r s,
then there exists, ρ ⊲ ⊢ M, [P ]sx where x<r M ⊒ N . and ρ ∼ π.
While splitting allows to cope with duplication, parametric splitting implies that an arbitrary
⊗-tree proof can be modified so as to be lifted into a box through one of its auxiliary doors Please
observe that pπ continues to be such an upper bound even if any natural number is substituted for
any of its free variables, an easy consequence of Lemma 2.3. The following is useful when dealing
with cuts involving the rule ?d:
P
Lemma 2.8 If q ⊒ 1, then z<q [M ]ry ⊑ [M ]ry {z/0}.
P
Proof. By hypothesis, we have that z<q [M ]ry = [N ]py for some N, y, p. As a consequence
M ≡ N {x/y +
X
p{z/u}}.
u<z
Now:
[M ]ry {z/0} ≡ [N {x/y +
X
p{z/u}}]r{z/0}
y
u<0
≡
[N {x/y}]r{z/0}
y
This concludes the proof.
3
P
≡ [N ]r{z/0}
⊒ [N ]x
x
z<q
r
Cut Elimination
In this Section, we show how a cut-elimination procedure for BLLP can be defined. We start by
showing how logical cuts can be reduced, where a cut is logical when the two immediate subproofs
end with a rule introducing the formula involved in the cut. We describe how logical cuts can be
reduced in the critical cases in Figure 2, which needs to be further explained:
• When reducing multiplicative logical cuts, we extensively use the Subtyping Lemma.
• In the dereliction reduction step, π{x/0} (obtained through Lemma 2.3) has conclusion ⊢
P
p{x/0}
N {x/0}, [N {x/0}]y
. By Lemma 2.8, M ⊑ x<q N ⊑ N {x/0}, and as a consequence,
r{x/0}
p{x/0}
. From [?y<p N ⊥ ]qx ⊒ [?y<r M ⊥ ]1x , it follows that [M ⊥ {x/0}]y
there is σ ⊲ ⊢ M, [N {x/0}]y
p{x/0}
p{x/0}
⊒ [N {x/0}]y
, and there is a proof λ ⊲ ⊢ O, [N {x/0}]y
.
• In the contraction reduction step, we suppose that π is a ⊗-tree. Then we can apply Lemma
2.6 and Lemma 2.2, and obtain σ ⊲ ⊢ M, [O⊥ ]px and λ ⊲ ⊢ O, [O⊥ {x/y + p}]qy such that
M ⊎ O ⊒ N.
8
Multiplicatives
q
⊢ Γ, [N ]p
x , [M]x
⊢ Γ, [N ` M]tx
⊢ N , [N ⊥ ]rx
`
⊢ Γ, [M]tx , [N ]tx
⊢ M, [M ⊥ ]sx
⊢ N , M, [N ⊥ ⊗ M ⊥ ]tx
⊢ Γ, N , M
⊢ N , [N ⊥ ]tx
⊢ Γ, N , [M]tx
Cut
⊗
7−→
Cut
⊢ M, [M ⊥ ]tx
Cut
⊢ Γ, N , M
Dereliction
π⊲ ⊢
⊢
N , [N ]p
y
M, [!y<p N ]qx
ρ ⊲ ⊢ O, [M ⊥ {x/0}]r{x/0}
y
!
⊢ O, [?y<p N ⊥ ]qx
⊢ M, O
p{x/0}
σ ⊲ ⊢ M, [N {x/0}]y
?d
7−→
Cut
p{x/0}
λ ⊲ ⊢ O, [N {x/0}]y
⊢ M, O
Cut
Contraction
q
ρ ⊲ ⊢ Γ, [O]p
x , [O{x/y + p}]y
π⊲ ⊢
⊢ Γ, [N ]rx
N , [N ⊥ ]rx
⊢ N, Γ
7−→
σ ⊲ ⊢ M, [O⊥ ]p
x
⊥
λ ⊲ ⊢ O, [O {x/y +
p}]qy
?c
Cut
q
π ⊲ ⊢ Γ, [O]p
x , [O{x/x + p}]y
⊢ M, Γ, [O{x/x + p}]qy , [N {x/x + p}]qy
⊢ M, O, Γ
⊢ N, Γ
Cut
Cut
?c
Digging
ρ⊲ ⊢
π⊲ ⊢
⊢ M, [N ]rx , [!y<p M]qx
N , M, [!y<p M]qx
N , [N ⊥ ]rx
⊢
O, [O]sy , [M]p
y
!
σ ⊲ ⊢ K, [O⊥ ]sy
⊢ K, O, [M]p
y
7−→
Cut
ρ ⊲ ⊢ O, [O]sy , [M]p
y
⊢
N , M, [!y<p M]qx
Cut
!
Figure 2: Some Logical Cut-Elimination Steps
P
• In digging, by Lemma 2.7 from π we can find σ ⊲ ⊢ K, [O⊥ ]sy , where N ⊑ x<q K.
All instances of the Cut rule which are not logical are said to be commutative, and induce a relation
on proofs. As an example, the proof
π ⊲ ⊢ Γ, N, [N ]px , [M ]qx
`
⊢ Γ, N, [N ` M ]rx
ρ ⊲ ⊢ N , N⊥
Cut
⊢ Γ, N , [N ` M ]rx
is equivalent to
π ⊲ ⊢ Γ, N, [N ]px , [M ]qx
ρ ⊲ ⊢ N , N⊥
Cut
p
⊢ Γ, N , [N ]x , [M ]qx
`
⊢ Γ, N , [N ` M ]rx
This way we can define an equivalence relation ∼
= on the space of proofs. In general, not all cuts
in a proof are logical, but any cut can be turned into a logical one:
Lemma 3.1 Let π be any proof containing an occurrence of the rule Cut. Then, there are two
proofs ρ and σ such that π ∼
= ρ 7−→ σ, where ρ can be effectively obtained from π.
The proof of Lemma 3.1 goes as follows: given any instance of the Cut rule
9
π ⊲ ⊢ Γ, [N ]px
ρ ⊲ ⊢ N , [P ]px
Cut
⊢ Γ, N
consider the path (i.e., the sequence of formula occurrences) starting from [N ]px and going upward
inside π, and the path starting from [P ]px and going upward inside ρ. Both paths end either at
an Ax rule or at an instance of a rule introducing the main connective in N or P . The game to
play is then to show that these two paths can always be shortened by way of commutations, thus
exposing the underlying logical cut.
Lemma 3.1 is implicitly defining a cut-elimination procedure: given any instance of the Cut
rule, turn it into a logical cut by the procedure from Lemma 3.1, then fire it. This way we are
implicitly defining another reduction relation −→. The next question is the following: is this
procedure going to terminate for every proof π (i.e., is −→ strongly, or weakly, normalizing)? How
many steps does it take to turn π to its cut-free form?
Actually, −→ produces reduction sequences of very long length, but is anyway strongly normalizing. A relatively easy way to prove it goes as follows: any BLLP proof π corresponds to a
LLP sequent calculus proof hπi, and the latter itself corresponds to a polarized proof net hhπii [26].
Moreover, π −→ ρ implies that hhπii 7→ hhρii, where 7→ is the canonical cut-elimination relation on
polarized proof-nets. Finally, hhπii is identical to hhρii whenever π ∼
= ρ. Since 7→ is known to be
strongly normalizing, −→ does not admit infinite reduction sequences:
Proposition 3.2 (Cut-Elimination) The relation −→ is strongly normalizing.
This does not mean that cut-elimination can be performed in (reasonably) bounded time. Already
in BLL this can take hyperexponential time: the whole of Elementary Linear Logic [18] can be
embedded into it.
3.1
Soundness
To get a soundness result, then, we somehow need to restrict the underlying reduction relation
−→. Following [19], one could indeed define a subset of −→ just by imposing that in dereliction,
contraction, or box cut-elimination steps, the involved ⊗-trees are closed. Moreover, we could
stipulate that reduction is external, i.e., it cannot take place inside boxes. Closed and external
reduction, however, is not enough to simulate head-reduction in the λµ-calculus, and not being
able to reduce under the scope of µ-abstractions does not make much sense anyway. We are forced,
then, to consider an extension of closed reduction. The fact that this new notion of reduction still
guarantees polynomial bounds is technically a remarkable strengthening with respect to BLL’s
Soundness Theorem [19].
There is a quite natural notion of downward path in proofs: from any occurrence of a negative
formula N, just proceed downward until you either find (the main premise of) a Cut rule, or a
conclusion. In the first case, the occurrence of N is said to be active, in the second it is said to be
passive. Proofs can then be endowed with a new notion of reduction: all dereliction, contraction or
box digging cuts can be fired only if the negative formula occurrences in its rightmost argument are
all passive. In the literature, this is sometimes called a special cut (e.g. [3]). Moreover, reduction
needs to be external, as usual. This notion of reduction, as we will see, is enough to mimic head
reduction, and is denoted with =⇒.
The next step consists in associating a weight, in the form of a resource polynomial, to
every proof, similarly to what happens in BLL. The pre-weight π ⋄ of a proof π with conclusion
⊢ A1 , . . . , An consists in:
• a resource polynomial pπ .
• n disjoints sets of resource variables S1π , . . . , Snπ , each corresponding to a formula in A1 , . . . , An ;
if this does not cause ambiguity, the set of resource variables corresponding to a formula A
will be denoted by S π (A). Similarly for S π (Γ), where Γ is a multiset of formulas.
If π has pre-weight pπ , S1π , . . . , Snπ , then the weight q π of π is simply pπ where, however, all the
variables in S1π , . . . , Snπ are substituted with 0: pπ {∪ni=1 Siπ /0}. The pre-weight of a proof π is
defined by induction on the structure of π, following the rules in Figure 3. Please notice how any
10
π⋄
π
[M]qx ⊑ [N ]p
x
r
[N ⊥ ]p
x ⊒ [P ]x
ρ ⊲ ⊢ Γ, [N ]p
x
σ ⊲ ⊢ N , [N ⊥ ]p
x
⊢ Γ, N
q
ρ ⊲ ⊢ Γ, [N ]p
x , [M]x
p⊑r
{y}, ∅, y
Ax
⊢ [M]qx , [P ]rx
q⊑r
σ ⊲ ⊢ M, [Q]p
x
r⊑p
⊢ N , M, [P ⊗ Q]rz
ρ ⊲ ⊢ N1 , . . . , Nn , [M]p
x
⊢
Mi ⊑
P
r⊑q
y<q
N ⊑ [?x<p P ]1y
⊢ N, N
ρ⊲ ⊢Γ
⊢ Γ, N
σ
p
ρ
σ
S ρ (N ), S σ (M), S ρ ([P ]p
x ) ∪ S ([Q]x ), p + p
⊗
Ni
M1 , . . . , Mn , [!x<p M]qy
ρ ⊲ ⊢ N , [P {y/0}]p{y/0}
x
ρ
p
ρ
S ρ (Γ), S ρ ([N ]p
x ) ∪ S ([M]x ) ∪ {y}, p + y
`
⊢ Γ, [N ` M]rx
ρ ⊲ ⊢ N , [P ]p
x
σ
σ
⊥ p
S ρ (Γ), S σ (N ), pρ {S ρ ([N ]p
x )/1} + p {S ([N ]x )/1}
Cut
!
S ρ (N1 ) ∪ {y1 }, . . . , S ρ (Nn ) ∪ {yn }, p · pρ + y1 + . . . + yn
?d
S ρ (N ), S ρ ([P {y/0}]p{y/0}
) ∪ {y}, pρ + y
x
S ρ (Γ), {y}
?w
ρ ⊲ ⊢ Γ, N, M
L⊑N⊎M
?c
⊢ Γ, L
ρ⊲ ⊢Γ
S ρ (Γ), {y}, pρ + y
⊥
⊢ Γ, [⊥]p
x
⊢ [1]p
x
S ρ (Γ), S ρ (N) ∪ S ρ (M) ∪ {y}
1
∅, 0
Figure 3: Pre-weights for Proofs.
negative formula N in the conclusion of π is associated with some fresh variables, each accounting
for the application of a rule to it. When N is then applied to a cut, all these variables are set to
1. This allows to discriminate between the case in which rules can “produce” time complexity
along the cut-elimination, and the case in which they do not. Ultimately, this leads to:
Lemma 3.3 If π ∼
= ρ, then q π = q ρ . If π =⇒ ρ, then q ρ ⊏ q π .
The main idea behind Lemma 3.3 is that even if the logical cut we perform when going from π to
ρ is “dangerous” (e.g. a contraction) and the involved ⊗-tree is not closed, the residual negative
rules have null weight, because they are passive.
We can conclude that:
Theorem 3.4 (Polystep Soundness) For every proof π, if π =⇒n ρ, then n ≤ q π .
In a sense, then, the weight of any proof π is a resource polynomial which can be easily computed
from π (rules in Figure 3 are anyway inductively defined) but which is also an upper bound on the
number of logical cut-elimination steps separating π from its normal form. Please observe that q π
continues to be such an upper bound even if any natural number is substituted for any of its free
variables, an easy consequence of Lemma 2.3.
Why then, are we talking about polynomial bounds? In BLL, and as a consequence also in
BLLP, one can write programs in such a way that the size of the input is reflected by a resource
variable occurring in its type. Please refer to [19].
4
A Type System for the λµ-Calculus
We describe here a version of the λµ-calculus as introduced by de Groote [11]. Terms are as follows
t, u ::= x
|
λx.t
|
µα.t
11
|
[α]t
|
(t)t,
t→u
t →w u
t →w u
t →h u
t →w u
tv →w uv
t →h u
λx.t →h λx.u
t →w u
[α]t →w [α]u
t →h u
µα.t →h µα.u
Figure 4: Weak and Head Notions of Reduction
where x and α range over two infinite disjoint sets of variables (called λ-variables and µ-variables,
respectively). In contrast with the λµ-calculus as originally formulated by Parigot [28], µ-abstraction
is not restricted to terms of the form [α]t here.
4.1
Notions of Reduction
The reduction rules we consider are the following ones:
(λx.t)u →β t[u /x ];
(µα.t)u →µ µα.t[[α](v)u /[α]v ];
µα.[α]t →θ t;
where, as usual, →θ can be fired only if α 6∈ FV (t). In the following, → is just →βµθ . In so-called
weak reduction, denoted →w , reduction simply cannot take place in the scope of binders, while
head reduction, denoted →h , is a generalization of the same concept from pure λ-calculus [13].
Details are in Figure 4. Please notice how in head reduction, redexes can indeed be fired even if
they lie in the scope of λ-or-µ-abstractions, which, however, cannot themselves be involved in a
redex. This harmless restriction, which corresponds to taking the outermost reduction order, is
needed for technical reasons that will become apparent soon.
4.2
The Type System
Following Laurent [26], types are just negative formulas. Not all of them can be used as types,
however: in particular, N ` M is a legal type only if N is in the form ?x<p O⊥ , and we use the
following abbreviation in this case: N ⊸px M ≡ (?x<p N ⊥ ) ` M . In particular, if M is ⊥ then
N ⊸px ⊥ can be abbreviated as ¬px N . Typing formulas are negative formulas which are either
⊥, or X, or in the form N ⊸px M (where N and M are typing formulas themselves). A modal
formula is one in the form ?x<p N ⊥ (where N is a typing formula). Please observe that all the
constructions from Section 2.1 (including labellings, sums, etc.) easily apply to typing formulas.
Finally, we use the following abbreviation for labeled modal formulas: qy [N ]px ≡ [?y<q N ⊥ ]px .
A typing judgement is a statement in the form Γ ⊢ t : N | ∆, where:
• Γ is a context assigning labelled modal formulas to λ-variables;
• t is a λµ-term;
• N is a typing formula;
• ∆ is a context assigning labelled typing formulas to µ-variables.
The way typing judgments are defined allows to see them as BLLP sequents. This way, again,
various concepts from Section 2.2 can be lifted up from sequents to judgments, and this remarkably
includes the subtyping relation ⊑.
Typing rules are in Figure 5. The typing rule for applications, in particular, can be seen as
overly complicated. In fact, all premises except the first two are there to allow the necessary
degree of malleability for contexts, without which even subject reduction would be in danger.
Alternatively, one could consider an explicit subtyping rule, the price being the loss of syntax
directness. Indeed, all malleability results from Section 2.3 can be transferred to the just defined
type assignment system.
12
Γ, x : sz [N ]py ⊢ t : [M ]qy | ∆
1 ⊑ p, r{y/0} ⊑ q, M ⊑ N {y/0}
var
p
Γ, x : rz [N ]y ⊢ x : [M ]qz | ∆
Θ ⊢ t : [N ⊸px M ]qy | Ψ
Γ ⊢ λx.t : [N
Ξ ⊢ u : [N ]px | Φ
Γ ⊢ (t)u : [M ]ky | ∆
⊸sz
r ⊒ q, r ⊒ p
M ]ry
h⊒q
k⊒q
P
Γ ⊑ Θ⊎Υ
Υ ⊑ P b<h Ξ
∆ ⊑ Ψ⊎Π
Π ⊑ b<h Φ
Γ ⊢ t : N | α : M, ∆
L⊑N⊎M
µ-name
Γ ⊢ [α]t : [⊥]qz | α : L, ∆
abs
|∆
app
Γ ⊢ t : [⊥]qz | β : N, ∆
µ-abs
Γ ⊢ µβt : N | ∆
Figure 5: (Additive) Type Assignment Rules
4.3
Subject Reduction and Polystep Soundness
The aim of this Section is to show that head reduction preserves types, and as a corollary, that
the number of reduction steps to normal form is bounded by a polynomial, along the same lines as
in Theorem 3.4. Actually, the latter will easily follow from the former, because so-called Subject
Reduction will be formulated (and in a sense proved) with a precise correspondence between type
derivations and proofs in mind.
In order to facilitate this task, Subject Reduction is proved on a modified type-assignment
system, called BLLPmult
λµ which can be proved equivalent to BLLPλµ . The only fundamental difference between the two systems lies in how structural rules, i.e., contraction and weakening, are
reflected into the type system. As we have already noticed, BLLPλµ has an additive flavour, since
structural rules are implicitly applied in binary and 0-ary typing rules. This, in particular, makes
the system syntax directed and type derivations more compact. The only problem with this approach is that the correspondence between type derivations and proofs is too weak to be directly
lifted to a dynamic level (e.g., one step in →h could correspond to possibly many steps in =⇒).
In BLLPmult
λµ , on the contrary, structural rules are explicit, and turns it into a useful technical tool
to prove properties of BLLPλµ .
BLLPmult
λµ ’s typing judgments are precisely the ones of BLLPλµ . What changes are typing rules,
which are in Figure 6. Whenever derivability in one of the system needs to be distinguished from
derivability on the other, we will put the system’s name in subscript position (e.g. Γ ⊢BLLPmult
t:
λµ
N | ∆). Not so surprisingly, the two BLLPλµ and BLLPmult
λµ type exactly the same class of terms:
Lemma 4.1 Γ ⊢BLLPmult
t : N | ∆ iff Γ ⊢BLLPλµ t : N | ∆
λµ
Proof. The left-to-right implication follows from weakening and contraction lemmas for BLLPλµ ,
which are easy to prove. The right-to-left implication is more direct, since additive var and app
are multiplicatively derivable.
⋄
Given a BLLPmult
λµ type derivation π, one can define a BLLP proof π following the rules in
Figure 7, which work by induction on the structure of π.
This way one not only gets some
guiding principles for subject-reduction, but can also prove that the underlying transformation
process is nothing more than cut-elimination:
p
Lemma 4.2 (λ-Substitution) If π ⊲ Γ, x : sz [N ]y ⊢ t : [M ]qy | ∆ and ρ ⊲ Θ ⊢ u : [N ]sz | Ξ, then
for all h ⊑ q there is σh such that
X
X
σh ⊲ Γ,
⊢ t{x/u} : [M ]qy | ∆,
Ξ.
b<h
b<h
⋄
Moreover, the proof obtained by h-boxing ρ and cutting it against π ⋄ is guaranteed to =⇒-reduce
to σh .
13
p
Γ, x : sz [N ]y ⊢ t : [M ]qy |∆
1 ⊑ p, r{y/0} ⊑ q, M ⊑ N {y/0}
var
p
x : rz [N ]y ⊢ x : [M ]qz |
Γ ⊢ t : [N ⊸px M ]qy |∆
Γ ⊢ λx.t : [N
Θ ⊢ u : [N ]px |Ξ
Γ, Ψ ⊢ (t)u : [M ]ky |∆, Φ
Γ ⊢ t : N|∆
µ-name
Γ ⊢ [α]t : [⊥]qz |α : N, ∆
⊸sz
r ⊒ q, r ⊒ p
abs
M ]ry |∆
h ⊒ q,
Pk ⊒ q,
Ψ ⊑ P b<h Θ,
Φ ⊑ b<h Ξ
app
Γ ⊢ t : [⊥]qz |β : N, ∆
µ-abs
Γ ⊢ µβt : N|∆
Γ⊢t:N|∆
?wλ
Γ, y : M ⊢ t : N | ∆
Γ, x : N, y : M ⊢ t : O | ∆
L ⊑N⊎M
?cλ
Γ, z : L ⊢ t{x/z}{y/z} : O | ∆
Γ⊢t:N|∆
?wµ
Γ ⊢ t : N | ∆, α : M
Γ ⊢ t : O | ∆, α : N, β : M
L⊑N⊎M
?cµ
Γ ⊢ t{α/γ}{β/γ} : O | ∆, γ : L
Figure 6: (Multiplicative) Type Assignment Rules
π⋄
π
1 ⊑ p, r{y/0} ⊑ q, M ⊑ N {y/0}
x : rz [N ]py ⊢ x : [M ]qz |
⊢
ρ ⊲ Γ, x : sz [N ]py ⊢ t : [M ]qy |∆
Γ ⊢ λx.t : [N
⊸sz
M ]ry |∆
ρ ⊲ Γ ⊢ t : N|∆
Γ ⊢ [α]t :
ρ ⊲ Γ ⊢ t : [⊥]qz |β : N, ∆
ρ ⊲ Γ⊢t:N|∆
Γ, y : M ⊢ t : N | ∆
ρ ⊲ Γ⊢t:N|∆
Γ ⊢ t : N | ∆, α : M
⊢ Ψ, [!x<p N ]h
y, Φ
ρ⋄ ⊲ ⊢ Γ, [?x<p N ⊥ `
M ]qy , ∆
ρ⋄ ⊲ Γ, N, ∆
ρ⋄
⊲ Γ, [⊥]qz , N, ∆
ρ⋄ ⊲ ⊢ Γ, [⊥]qz , N, ∆
⊢ Γ, N, ∆
µ-abs
?w λ
ρ⋄ ⊲ ⊢ Γ, N, ∆
⊢ Γ, M, N, ∆
?w µ
ρ⋄ ⊲ ⊢ Γ, N, ∆
⊢ Γ, N, M, ∆
ρ ⊲ Γ, x : N, y : M ⊢ t : O | ∆
Γ, z : L ⊢ t{x/z}{y/z} : O | ∆
ρ ⊲ Γ ⊢ t : O | ∆, α : N, β : M
Γ ⊢ t{α/γ}{β/γ} : O | ∆, γ : L
⊢ [1]qz
?cλ
ρ⋄ ⊲ ⊢ Γ, N, M, O, ∆
⊢ Γ, L, O, ∆
?cµ
ρ⋄ ⊲ ⊢ Γ, O, N, M, ∆
⊢ Γ, O, L, ∆
Figure 7: Mapping of (multiplicative) derivations into BLLP proofs
14
⊢ [M ⊥ ]ky , [M ]ky
⊢ Ψ, [!x<p N ⊗ M ⊥ ]qy , Φ, [M ]ky
⊢ Γ, Ψ, [M ]ky , ∆, Φ
µ-name
: N, ∆
Γ ⊢ µβt : N|∆
⊢ Γ, [?z<s N ` M ]ry , ∆
σ⋄ ⊲ ⊢ Θ, [N ]px , Ξ
app
Γ, Ψ ⊢ (t)u : [M ]ky |∆, Φ
, [M ]qz
[?z<r N ⊥ ]py , [M ]qz
ρ⋄ ⊲ ⊢ Γ, [?z<s N ]py , [M ]qy , ∆
abs
ρ ⊲ Γ ⊢ t : [N ⊸px M ]qy |∆
σ ⊲ Θ ⊢ u : [N ]px |Ξ
[⊥]qz |α
r{y/0}
⊢ [N ⊥ {y/0}]z
var
Proof. As usual, this is an induction on the structure of π. We only need to be careful and
generalize the statement to the case in which a simultaneous substitution for many variables is
needed.
Lemma 4.3 (µ-Substitution) If π ⊲ Γ ⊢ t : [⊥]qy | ∆, α : [N ⊸sz M ]py and ρ ⊲ Θ ⊢ u : [N ]sz | Ξ,
then for all h ⊒ q there is σh such that
X
X
Γ,
Θ ⊢ t{[α]w/[α](w)u} : [⊥]qy | ∆, α : [M ]py ,
Ξ
b<h
b<h
Moreover, the proof obtained by h-boxing ρ⋄ , tensoring it with an axiom and cutting the result
against π ⋄ is guaranteed to =⇒-reduce to σh .
Theorem 4.4 (Subject Reduction) Let π ⊲ Γ ⊢ t : N | ∆ and suppose t →h u. Then there
is ρ ⊲ Γ ⊢ u : N | ∆. Moreover π ⋄ =⇒+ ρ⋄ .
Proof. By induction on the structure of π. Here are some interesting cases:
• If t is an application, reduction takes place inside t, and π is as follows
Γ ⊢ t : [N ⊸px M ]qy |∆
Θ ⊢ v : [N ]py |Ξ
h ⊒ q, k ⊒ q
app
P
P
Γ ⊎ b<h Θ ⊢ (t)v : [M ]kz |∆ ⊎ b<h Ξ
then ρ is
Γ ⊢ u : [N ⊸px M ]qy |∆
Θ ⊢ v : [N ]py |Ξ
h ⊒ q, k ⊒ q
app
P
P
Γ ⊎ b<h Θ ⊢ (u)v : [M ]kz |∆ ⊎ b<h Ξ
which exists by induction hypothesis. We omit the other trivial cases.
• If t is a β-redex, then π looks as follows:
p
Γ, x : sz [N ]y ⊢ t : [M ]qy |∆
r ⊒ q, r ⊒ p
abs
s
r
Γ ⊢ λx.t : [N ⊸y M ]u |∆
Θ ⊢ u : [N ]sz |Ξ
h ⊒ q, k ⊒ q
app
P
P
Γ, b<h Θ ⊢ (λx.t)u : [M ]ky |∆, b<h Ξ
Lemma 4.2 ensures that the required type derivation actually exists:
P
P
Γ, b<h Θ ⊢ t{x/u} : [M ]ky |∆, b<h Ξ
• If t is a µ-redex, then π looks as follows:
Γ ⊢ t : [⊥]qz |β : [N ⊸sz M ]py , ∆
µ-abs
Γ ⊢ µβt : [N ⊸sz M ]py |∆
Θ ⊢ u : [N ]sy |Ξ
h ⊒ p, k ⊒ p
app
P
P
k
Γ, b<h Θ ⊢ (µβ.t)u : [M ]z |∆, b<h Ξ
and Lemma 4.3 ensures us that ρ exists for
P
P
Γ ⊎ b<h Θ ⊢ µβ.t[[β](v)u /[β]v ] : [M ]kz |∆ ⊎ b<h Ξ
• If t is a θ-redex, then π looks as follows:
π ⊲ Γ ⊢ t : [N ]px | ∆
?wµ
Γ ⊢ t : [N ]px | ∆, α : [N ]qy
r ⊒p+q
µ-name
s
r
Γ ⊢ [α]t : [⊥]x | ∆, α : [N ]y
µ-abs
Γ ⊢ t : [N ]rx | ∆
Since r = p + q ⊒ p we know that
π S ⊲ Γ ⊢ t : [N ]rx | ∆
where π S is the derivation obtained from π, applying the Subtyping Lemma to the derivation
π.
This concludes the proof.
Observe how performing head reduction corresponds to =⇒, instead of the more permissive −→.
The following, then, is an easy corollary of Theorem 4.4 and Theorem 3.4:
Theorem 4.5 (Polystep Soundness for Terms) Let π ⊲ Γ ⊢ t : N | ∆ and let t →nh u.
Then n ≤ pπ⋄ .
15
var
y : s [X]1 ⊢ y : [X]s | α : [X]0 , β : [Y ]0
y : s [X]1 ⊢ [α]y : [⊥]0 | α : [X]s , β : [Y ]0
y : s [X]1 ⊢ µβ.[α]y : [Y ]0 | α : [X]s
⊢ λy.µβ.[α]y : [X ⊸s U]1 |α : [X]s
π
1
µ-name
µ-abs
abs
P
v<r s
app
x : rv [(X ⊸s Y ) ⊸1 X] ⊢ (x)λy.µβ.[α]y : [X]rv |α : [X]v
k
k
⊒
⊒
r+
1
P
v<r
s
0
x : rv [(X ⊸s Y ) ⊸1 X] ⊢ [α](x)λy.µβ.[α]y : [⊥]1 | α : [X]k
1
x : rv [(X ⊸sv Y ) ⊸1 X] ⊢ µα.[α](x)λy.µβ.[α]y : [X]k |
⊢ λx.µα.[α](x)λy.µβ.[α]y : [((X ⊸s Y ) ⊸1 X) ⊸rv X]k |
µ-name
µ-abs
abs
Figure 8: A Type Derivation for κ
5
Control Operators
In this section, we show that BLLPλµ is powerful enough to type (the natural encoding of) two
popular control operators, namely Scheme’s callcc and Felleisen’s C [2] [26].
Control operators change the evaluation context of an expression. This is simulated by the
operators µ and [·] which can, respectively, save and restore a stack of arguments to be passed to
subterms. This idea, by the way, is the starting point of an extension of Krivine’s machine for de
Groote’s λµ [13] (see Section 6).
An extension of de Groote’s calculus named Λµ-calculus [29] satisfies a Böhm separation theorem that fails for Parigot’s calculus [9]. Hence in an untyped setting the original λµ of Parigot
is strictly less expressive than de Groote’s calculus.
5.1
callcc
An encoding of callcc into the λµ-calculus could be, e.g., κ = λx.µα.[α](x)λy.µβ.[α]y. Does κ
have the operational behavior we would expect from callcc? First of all, it should satisfy the
following property (see [14]). If k 6∈ FV (e), then (κ)λk.e →∗ e. Indeed:
(λx.µα.[α](x)λy.µβ.[α]y)λk.e →h µα.[α](λk.e)λy.µβ.[α]y →h µα.[α]e →h e,
where the second β-reduction step replaces e{k/λy.µβ.[α]y} with e since k 6∈ FV (e) by hypothesis.
It is important to observe that the second step replaces a variable for a term with a free µ-variable,
hence weak reduction gets stuck. (Actually, our notion of weak reduction is even more restrictive
than the one proposed by de Groote in [13].) Head reduction, on the contrary, is somehow more
liberal. Moreover, it is also straightforward to check that the reduction of callcc in [28, §3.4] can
be simulated by head reduction on κ.
But is κ typable in BLLPλµ ? The answer is positive: a derivation typing it with (an instance
of) Pierce’s law is in Figure 8, where π is the obvious derivation of
1
x : r [(X ⊸s Y ) ⊸1 X] ⊢ x : [(X ⊸s Y ) ⊸1 X]r | α : [X]0 .
5.2
Felleisen’s C
The canonical way to encode Felleisen’s C as a λµ-term is as the term ℵ = λf.µα.(f )λx.[α]x.
Its behavior should be something like (ℵ)wt1 . . . tk → (w)λx.(x)t1 . . . tk , where x 6∈ F V (t1 , . . . ,
tk ), i.e., x is a fresh variable. Indeed
(ℵ)wt1 . . . tk →h (µα.(w)λx.[α](x))t1 . . . tk →kh µα.(w)λx.[α](x)t1 . . . tk .
16
var
x : r [X]1 ⊢ x : [X]r
µ-name
x : r [X]1 ⊢ [α]x : [⊥]0 | α : [X]r
σ
f :
1
h
1 r
v [¬ ¬ X]
abs
⊢ λx.[α]x : [¬r X]1 | α : [X]r
⊢ (f )λx.[α]x : [⊥]h
v | α :
1
1 r
f :h
v [¬ ¬ X] ⊢ µα.(f )λx.[α]x : [X]
P
[X] v<h r
P
v<h r
app
µ-abs
|
k
k
k
⊢ λf.µα.(f )λx.[α]x : [¬1 ¬r X ⊸h
v X]
⊒
⊒
1
P
v<h
r
abs
Figure 9: A Type Derivation for ℵ
A type derivation for ℵ is in Figure 9, where σ is a derivation for
1
f : hv [¬1 ¬r X] ⊢ f : [¬1 ¬r X]hv | α : [X]0 .
It is worth noting that weak reduction is strong enough to properly simulating the operational
behavior of C. It is not possible to type C in Parigot’s λµ, unless an open term is used. Alternatively, a free continuation constant must be used (obtaining yet another calculus [2]). This is one
of the reasons why we picked the version of λµ-calculus proposed by de Groote over other calculi.
See [12] for a discussion about λµ-and-λ-calculi and Felleisen’s C.
6
Abstract Machines
Theorem 4.5, the main result of this paper so far, tells us that the number of head-reduction steps
performed by terms typable in BLLPλµ is bounded by the weight of the underlying type derivation.
One may wonder, however, whether taking the number of reduction steps as a measure of term
complexity is sensible or not — substitutions involve arguments which can possibly be much bigger
than the original term. Recent work by Accattoli and the first author [1], however, shows that
in the case of λ-calculus endowed with head reduction, the unitary cost model is polynomially
invariant with respect to Turing machines. We conjecture that those invariance results can be
extended to the λµ-calculus.
In this Section, we show that BLLPλµ is polystep sound for another cost model, namely the one
induced by de Groote’s K, an abstract machine for the λµ-calculus. This will be done following a
similar proof for PCF typed with linear dependent types [7] and Krivine’s Abstract Machine (of
which K is a natural extension).
Configurations of K are built around environments, closures and stacks, which are defined
mutually recursively as follows:
• Environments are partial functions which makes λ-variables correspond to closures and µvariables correspond to stacks; metavariables for environments are E , F , etc.;
• Closures are pairs whose first component is a λµ-term and whose second component is an
environment; metavariables for closure are C, D, etc.
• Stacks are just finite sequences of closures; metavariables for stacks are S , T , etc.
Configurations are pairs whose first component is a closure and whose second component is a
stack, and are indicated with C, D, etc. Reduction rules for configurations are in Figure 10. The
K-machine is sound and complete with respect to head reduction [13], where however, reduction
can take place in the scope of µ-abstractions, but not in the scope of λ-abstractions.2
Actually, BLLPλµ can be turned into a type system for K’s configurations. We closely follow
Laurent [25] here. The next step is to assign a weight q π to every type derivation π ⊲ C : N,
2 The authors are aware of the work in [31], in which a Krivine machine for λµ is derived semantically rather
than syntactically (independently of de Groote). In the same paper there is also a further extension of the machine
which allows to reduce under µ- and even λ-abstractions. The paper is not essential for our purposes since the
abstract machine of de Groote is enough to work with control operators. Still, even though there are some important
differences with respect to our setting (the calculus considered is an untyped variant of Parigot’s λµ), it might be
worthwhile to investigate in the future.
17
((x, E ), S ) ֒→ (E (x), S );
((λx.t, E ), C · S ) ֒→ ((t, E {x/C}), S );
((tu, E ), S ) ֒→ ((t, E ), (u, E ) · S );
((µα.t, E ), S ) ֒→ ((t, E {α/S }), ε);
(([α]t, E ), ε) ֒→ ((t, E ), E (α)).
Figure 10: K-machine Transitions.
similarly to what we have done in type derivations for terms. The idea then is to prove that the
weight of (typable) configurations decreases at every transition step:
Lemma 6.1 If C ֒→ D, then q C ⊐ q D .
This allows to generalize polystep soundness to K:
Theorem 6.2 (Polystep Soundness for the K) Let π ⊲ ⊢ C : N and let C ֒→n D. Then
n ≤ qπ .
Please observe how Theorem 6.2 holds in particular when C is the initial configuration for a
typable term t, i.e., hht, ∅i, εi.
7
Conclusions
In this paper we have presented some evidence that the enrichment to intuitionistic linear logic
provided by bounded linear logic is robust enough to be lifted to polarized linear logic and the
λµ-calculus. This paves the way towards a complexity-sensitive type system, which on the one
hand guarantees that typable terms can be reduced to their normal forms in a number of reduction
steps which can be read from their type derivation, and on the other allows to naturally type useful
control operators.
Many questions have been purposely left open here: in particular, the language of programs is
the pure, constant-free, λµ-calculus, whereas the structure of types is minimal, not allowing any
form of polymorphism. We expect that endowing BLLP with second order quantification or BLLPλµ
with constants and recursion should not be particularly problematic, although laborious: the
same extensions have already been considered in similar settings in the absence of control [19, 7].
Actually, a particularly interesting direction would be to turn BLLPλµ into a type system for Ong
and Stewart’s µPCF [27], this way extending the linear dependent paradigm to a language with
control. This is of course outside the scope of this paper, whose purpose was only to delineate the
basic ingredients of the logic and the underlying type system.
As we stressed in the introduction, we are convinced this work to be the first one giving a
time complexity analysis methodology for a programming language with higher-order functions
and control.3 One could of course object that complexity analysis of λµ-terms could be performed
by translating them into equivalent λ terms, e.g. by way of a suitable CPS-transform [11]. This,
however, would force the programmer (or whomever doing complexity analysis) to deal with
programs which are structurally different from the original one. And of course, translations could
introduce inefficiencies, which are maybe harmless from a purely qualitative viewpoint, but which
could make a difference for complexity analysis.
3 Tatsuta
has investigated the maximum length of µ-reduction for a language without λ-abstractions (RTA 2007).
18
References
[1] Beniamino Accattoli and Ugo Dal Lago. On the invariance of the unitary cost model for head
reduction. In RTA, volume 15 of LIPIcs, pages 22–37, 2012.
[2] Zena M. Ariola and Hugo Herbelin. Minimal classical logic and control operators. In ICALP,
volume 2719 of LNCS, pages 871–885. Springer, 2003.
[3] Patrick Baillot, Paolo Coppola, and Ugo Dal Lago. Light logics and optimal reduction:
Completeness and complexity. Information and Computation, 209(2):118–142, 2011.
[4] Patrick Baillot and Damiano Mazza. Linear logic by levels and bounded time complexity.
Theoretical Computer Science, 411(2):470–503, 2010.
[5] Patrick Baillot and Kazushige Terui. Light types for polynomial time computation in lambdacalculus. In Information and Computation, volume 207, pages 41–62, 2009.
[6] Pierre-Louis Curien and Hugo Herbelin. The duality of computation. In ICFP, pages 233–243.
ACM, 2000.
[7] Ugo Dal Lago and Marco Gaboardi. Linear dependent types and relative completeness. Logical
Methods in Computer Science, 8(4), 2012.
[8] Ugo Dal Lago and Martin Hofmann. Bounded linear logic, revisited. In TLCA, volume 5608
of LNCS, pages 80–94. Springer, 2009.
[9] René David and Walter Py. λµ-calculus and Böhm’s theorem. Journal of Symbolic Logic,
pages 407–413, 2001.
[10] Jacobus W. de Bakker, Arie de Bruin, and Jeffrey Zucker. Mathematical theory of program
correctness. Prentice-Hall International Series in Computer Science. Prentice Hall, 1980.
[11] Philippe de Groote. A CPS-translation of the λµ-calculus. In CAAP, volume 787 of LNCS,
pages 85–99. Springer, 1994.
[12] Philippe de Groote. On the relation between the λµ-calculus and the syntactic theory of
sequential control. In Logic Programming and Automated Reasoning, pages 31–43. Springer,
1994.
[13] Philippe de Groote. An environment machine for the λµ-calculus. Mathematical Structures
in Computer Science, 8(6):637–669, 1998.
[14] Matthias Felleisen. On the expressive power of programming languages. In ESOP, volume
432 of LNCS, pages 134–151. Springer, 1990.
[15] Marco Gaboardi and Simona Ronchi Della Rocca. A soft type assignment system for lambdacalculus. In CSL, volume 4646 of LNCS, pages 253–267. Springer, 2007.
[16] Jean-Yves Girard. Linear logic. Theoretical Computer Science, 50(1):1–101, 1987.
[17] Jean-Yves Girard. A new constructive logic: Classical logic. Mathematical Structures in
Computer Science, 1(3):255–296, 1991.
[18] Jean-Yves Girard. Light linear logic. Information and Computation, 143(2):175–204, 1998.
[19] Jean-Yves Girard, Andre Scedrov, and Phil Scott. Bounded linear logic: a modular approach
to polynomial-time computability. Theoretical Computer Science, 97(1):1–66, 1992.
[20] Timothy Griffin. A formulae-as-types notion of control. In POPL, pages 47–58. ACM Press,
1990.
19
[21] Sumit Gulwani. Speed: Symbolic complexity bound analysis. In CAV, volume 5643 of LNCS,
pages 51–62. Springer, 2009.
[22] Steffen Jost, Kevin Hammond, Hans-Wolfgang Loidl, and Martin Hofmann. Static determination of quantitative resource usage for higher-order programs. In POPL, Madrid, Spain,
2010. ACM Press.
[23] Yves Lafont. Soft linear logic and polynomial time. Theoretical Computer Science, 318(1):163–
180, 2004.
[24] Olivier Laurent. Étude de la polarisation en logique. Thèse de doctorat, Université AixMarseille II, March 2002.
[25] Olivier Laurent. Krivine’s abstract machine and the λµ-calculus (an overview). Unpublished
note, September 2003.
[26] Olivier Laurent. Polarized proof-nets and λµ-calculus.
290(1):161–188, 2003.
Theoretical Computer Science,
[27] C.-H. Luke Ong and Charles A. Stewart. A Curry-Howard foundation for functional computation with control. In POPL, pages 215–227. ACM Press, 1997.
[28] Michel Parigot. λµ-calculus: an algorithmic interpretation of classical natural deduction. In
LPAR, volume 624 of LNCS, pages 190–201. Springer, 1992.
[29] Alexis Saurin. Separation with streams in the λµ-calculus. In LICS, pages 356–365. IEEE,
2005.
[30] Ulrich Schöpp. Stratified bounded affine logic for logarithmic space. In LICS, pages 411–420,
2007.
[31] Thomas Streicher and Bernhard Reus. Classical logic, continuation semantics and abstract
machines. Journal of functional programming, 8(6):543–572, 1998.
[32] Reinhard Wilhelm, Jakob Engblom, Andreas Ermedahl, Niklas Holsti, Stephan Thesing,
David Whalley, Guillem Bernat, Christian Ferdinand, Reinhold Heckmann, Tulika Mitra,
Frank Mueller, Isabelle Puaut, Peter Puschner, Jan Staschulat, and Per Stenström. The
worst case execution time problem - overview of methods and survey of tools. ACM Transactions on Embedded Computing Systems, 2008.
20
| 6cs.PL
|
1
A Discontinuous Galerkin Time Domain
Framework for Periodic Structures Subject To
Oblique Excitation
arXiv:1311.0790v2 [cs.CE] 17 Feb 2014
Nicholas C. Miller, Andrew D. Baczewski, John D. Albrecht,
and Balasubramaniam Shanker
Abstract—A nodal Discontinuous Galerkin (DG) method is derived for the analysis of time-domain (TD) scattering from doubly
periodic PEC/dielectric structures under oblique interrogation.
Field transformations are employed to elaborate a formalism that
is free from any issues with causality that are common when
applying spatial periodic boundary conditions simultaneously
with incident fields at arbitrary angles of incidence. An upwind
numerical flux is derived for the transformed variables, which
retains the same form as it does in the original Maxwell
problem for domains without explicitly imposed periodicity. This,
in conjunction with the amenability of the DG framework to
non-conformal meshes, provides a natural means of accurately
solving the first order TD Maxwell equations for a number of
periodic systems of engineering interest. Results are presented
that substantiate the accuracy and utility of our method.
Index Terms—Periodic structures, Discontinuous Galerkin
(DG) methods, time domain analysis.
I. I NTRODUCTION
Periodic structures play a significant role in electromagnetics and optics in generating unique spectral responses that can
be readily engineered. Applications of periodicity include frequency selective surfaces (FSS) [1], electromagnetic band gap
(EBG) structures [2], biomimetic structures and metamaterials
[3], [4], etc. Computational analysis of fields in increasingly
intricate periodic unit cells plays a significant role in their
design and optimization. In the frequency domain, Integral
Equation (IE) [5], [6], Finite Element (FE) [7], [8], and Discontinuous Galerkin (DG) [9] methods have been successfully
applied to a variety of periodic electromagnetic systems. Timedomain (TD) methods for studying periodic systems include
FE [10],[11], IE [12], and Finite Difference Time Domain
(FDTD) [13], while DG methods remain relatively unexplored.
TD analysis of periodic structures provides a number of
advantages, such as characterization of the broadband response of a structure in a single simulation, and treatment
of nonlinearities. Both integral and differential formulations
of the Maxwell problem have attendant disadvantages as well.
For integral formulations, discretization yields a dense linear
system. While fast and efficient [12], [14] methods have
been applied to these problems, stable formulations of TDIEs
remain a research problem, with much recent progress [15].
Recent work has also been presented on obtaining transient response using entire domain Laguerre polynomials that results a
system wherein the time variable is completely avoided [16].
Alternatively, while differential formulations of the problem
yield sparse linear systems and stability is better understood,
the proper imposition of boundary conditions (BCs) becomes
challenging. In particular, the asymptotic boundary condition
on the fields receding to infinity must be enforced approximately with an absorbing boundary condition (ABC) or a
perfectly matched layer (PML) [17]. Further, while periodic
BCs at the perimeter of the unit cell are trivial to enforce
for systems excited at normal incidence, there are well-known
issues associated with causality at oblique incidence [10].
A set of field transformations that mitigate causality issues
was introduced for FDTD in 1993 [18], and later adapted
to an FETD framework in a sequence of papers in the mid2000s [10], [11]. Here, the frequency domain Floquet-periodic
boundary condition is exploited, wherein fields at the unit
cell boundaries are related to one another by a phase shift
that depends on the exciting wave vector and lattice vectors.
The frequency domain Maxwell Equations are then posed in
terms of a set of transformed variables, into which this phase
shift is built, and an inverse transform is applied to return the
equations to the time domain. Additional terms then appear in
the TD Maxwell Equations for the transformed variables.
In this work, we will apply these field transformations
to a time domain Discontinuous Galerkin (DG) framework
for the conservation form of the Maxwell equations for the
first time. Time domain analysis of periodic structures with
DG methods has received relatively little attention, with a
few exceptions [9], [19]. The unique contributions of this
paper are extensions of a time domain DG framework that
permit the analysis of doubly periodic structures at oblique
incidence. First, the field transformations that are used to
remove causality issues are reviewed. We then demonstrate
that the form of the upwind flux utilized in discretizing the
transformed Maxwell Equations is invariant to whether or
not one is utilizing the original or transformed fields. Issues
addressing the use of non-conformal meshes across periodic
boundaries are discussed, and relevant implementation details
are provided. Finally, results are presented that validate the
accuracy and utility of our method for a number of doubly
periodic test cases.
II. M ATHEMATICAL F ORMULATION
Fig. 1: Illustration of the z-plane of a doubly periodic structure with periods
|~a1 | and |~a2 |. The ellipses indicate that the structure is periodic in the x- and
y-directions.
Consider a domain, Ω ⊂ R3 depicted in Fig. 2, where a
doubly periodic distribution of isotropic, lossless, dielectric
and/or PEC scatterers reside. The periodicity of the system is
2
described by a 2-lattice, L2 , defined as:
L2 = {~un = n1~a1 + n2~a2 |n1 , n2 ∈ Z}
(1)
Here, the subscript n is defined as a multi-index, and ~ai are
basis vectors for the lattice. These vectors will be orthogonal
in this work, but extensions to non-orthogonal basis vectors
are simply realized. Incident on the system is a planewave
~ i (~r, t), with a wavevector k̂i = sin θ cos φx̂ +
excitation E
sin θ sin φŷ + cos θẑ. The incident wavevector, k̂i , can be
k
further decomposed into k̂i and k̂⊥
i , which are within and
orthogonal to the span of L2 , respectively.
As is evident from Eqns. (4a) and (4b), using these auxiliary
field components is tantamount to zero phase propagation at
the boundaries, i.e., there is no delay in boundaries of the unit
cell. This is the time domain analog to cell-periodic Bloch
functions typical of frequency analysis.
Applying the field transformations to the first order time
domain Maxwell Equations yields
k
~ r, t) k̂
∂ ~S(~r, t)
∂ P(~
+ i ×
= ∇ × ~S(~r, t)
(5a)
∂t
c0
∂t
k
~ r, t)
k̂
∂ P(~
∂ ~S(~r, t)
~ r, t)
− i ×
+µ
= −∇ × P(~
(5b)
c0
∂t
∂t
It is these equations that we will now discretize within the DG
framework.
ε
III. T HE D ISCONTINUOUS G ALERKIN M ETHOD
A. Discretization
Fig. 2: Illustration of a single unit cell of a doubly periodic structure with
periods |~a1 | and |~a2 |.
The fields obey the following boundary conditions under
spatial translation by a lattice vector in L2 :
!
k
~
k̂
·
r
~ r, t) = E(~
~ r + ~un , t) ∗ δ t + i
E(~
(2a)
c0
!
k
k̂i · ~r
~
~
H(~r, t) = H(~r + ~un , t) ∗ δ t +
(2b)
c0
Direct implementation of these periodic boundary conditions
requires knowledge of future values of fields at one periodic
boundary in order to update fields at the other periodic
boundary. In the context of a time integration scheme in which
fields are updated in time based upon a sequence of their
previous values, this is not possible without extrapolation.
Alternatively, transformed fields can be identified for which
the periodic boundary conditions remain causal. As done in
~ r, ω)
[10],[18], we introduce delayed auxiliary variables, P(~
~
and S(~r, ω)
k
~ r, ω) = P(~
~ r, ω)e−j~ki ·~r
E(~
~ r, ω) = ~S(~r, ω)e
H(~
(3a)
k
−j~
ki ·~r
(3b)
To allow a seamless extension from previous DG formulations [20], [21], [22], we write Eqns. (5a) and (5b) in
conservation form:
∂~q(~r, t)
~ (~q(~r, t)) = 0
Q
+∇·F
(6)
∂t
Here, the periodic/materials matrix Q, field six-vector ~q(~r, t),
~ (~q(~r, t)) are defined as:
and flux matrix F
k
εI1
c−1
k̂i × I1
0
,
Q=
k
−c−1
k̂
×
I
µI
1
1
0
i
~q(~r, t) =
~ r, t)
P(~
~S(~r, t)
~ (~q(~r, t)) =
,F
−êi × ~S(~r, t)
~ r, t)
êi × P(~
here, êi represents the ith Cartesian unit vector, ε is the
isotropic permittivity, µ is the isotropic permeability, and I1
is the 3x3 identity matrix.
Solving this system of equations requires discretizing the
domain using k non-overlapping tetrahedra, where domains
are denoted Ωk with boundaries ∂Ωk that are equipped with
an outward pointing normal n̂. The vector unknowns are
expanded into a set of globally discontinuous nodal polynoNp
P
~qk (~ri , t) `ki (~r). We use the nodal basis
mials ~q (~r, t) ≈
i=1
functions defined in [20].
Following standard DG practice [20], a strong form of the
problem is obtained as:
ZZZ
∂~q(~r, t)
~
Q
+ ∇ · F (~q(~r, t)) `kj (~r)d~r
∂t
Ωk
ZZ
~ (~q(~r, t)) − F
~ ∗ (~q(~r, t)) `k (~r)d~r
=
~n · F
(7)
j
∂Ωk
~∗
It can be shown trivially that these transformed fields obey
~ r, t) = P(~
~ r + ~un , t)
P(~
~S(~r, t) = ~S(~r + ~un , t)
where F is called the numerical flux. We can rewrite the
semi-discrete problem in Eqn. (7) as:
(4a)
(4b)
h
i
∂~q(~r, t)
~ −F
~∗
= Q−1 M−1 S~q + M−1 F n̂ · F
(8)
∂t
3
~ −F
~ ∗ , defined on
with the function of nodal values n̂ · F
~ (~q(~r, t)),
the element boundaries, replacing the flux matrix F
the periodic/materials matrix Q re-defined as
Q=
εI2
0
0
0
0
0
εI2
0
0
0
0
0
εI2
κy I2
−κx I2
0
0
κy I2
µI2
0
0
0
−κx I2
0
µI2
−κy I2
κx I2
0
0
0
−κy I2
κx I2
0
0
0
µI2
k
where k̂i = κx x̂ + κy ŷ and I2 is the Np xNp identity matrix.
The mass matrix M, stiffness matrix S, and face matrix F
are defined as
ZZZ
Mij =
`ki (~r)`kj (~r)d~r
it is evident that this system has two distinct characteristic
−1/2
values, ± (εµ)
. This implies that only three RankineHugoniot jump conditions are needed to relate the fields
across discontinuities [20], [23]. Using the convention in [24],
integrating over a single element, and reducing integration
limits to the faces of the elements, we arrive at the jump
conditions for the equivalent transformed equations
k
i
h
~∗−P
~ − ∗ δ t − k̂i · ~r
Z − ~S∗ − ~S− + n̂ × P
c0
h
!
k
i
~ ∗∗ − P
~ + ∗ δ t − k̂i · ~r
Z ~S∗∗ − ~S+ + n̂ × P
c0
!
k
h
i
~ ∗∗ − P
~ ∗ ∗ δ t − k̂i · ~r = 0
n̂ × P
c0
!
k
h
i
~
k̂
·
r
∗∗
∗
i
n̂ × ~S − ~S
∗δ t−
=0
c0
+
=0
!
=0
k
Z ZΩZ
Sij =
`ki (~r)∇`kj (~r)d~r
Since these equations hold for all time, the periodic numerical
flux may now be written as [24]
Ωk
ZZ
Fij =
`ki (~r)`kj (~r)d~r
∂Ωk
B. Periodic Numerical Flux
~ −F
~ ∗ is at the heart of all
Choice of the nodal values n̂· F
DG formulations. Hesthaven and Warburton have proven that
an upwind flux is both stable and convergent for Maxwell’s
Equations [20]. For the non-periodic Maxwell’s Equations, the
upwind flux takes the form
hh ii
hh ii
~
−Z̄ −1 n̂ × Z + ~S − n̂ × P
∗
~
~
n̂ · F − F =
hh ii
hh ii
~ + n̂ × ~S
Ȳ −1 n̂ × Y + P
(11)
hh ii
+
−
~
~
~
In Eqn. 11, P = P − P is the jump in the nodal field
values at an element’s boundaries.
C. Boundary Conditions
TABLE I: Boundary Condition Jumps
hh ii
hh ii
~
~ − n̂ × E
H
−Z̄ n̂ × Z
∗
~
~
n̂ · F − F =
hh ii
hh ii
~ + n̂ × H
~
Ȳ −1 n̂ × Y + E
(10)
hh ii
+
−
~
~
~
Here, the jump E = E − E is defined in terms of nodal
field values at the element boundaries, and the impedance
Z̄ = Z + + Z − is twice the average impedance shared at
these boundaries. To derive the periodic numerical
flux
for
k
k̂i ·~
r
~
~
~
~
P(~r, t) and S(~r, t), we note that E = P ∗ δ t − c0
and
k
~ = ~S ∗ δ t − k̂i ·~r . Using these in the conservation form
H
c0
−1
+
of Maxwell’s equations
k
~ ∗ δ t − k̂i ·~r
P
c0
εI1
0
∂
k
∂t
~S ∗ δ t − k̂i ·~r
0 µI1
c0
k
k̂i ·~r
~
−êi × S ∗ δ t − c0
=0
+∇ ·
k
k̂
·~
r
~ ∗δ t− i
êi × P
c0
B.C.
hh ii
~
P
hh ii
~
S
PEC:
~−
−2P
0
ABC (TE):
~ − |cos θ|
−2P
−2~
S−
ABC (TM):
~−
−2P
−2~
S− |cos θ|
TF/SF:
~+ −P
~− ±P
~ inc
P
~
S+ − ~
S− ± ~
Sinc
Applying boundary conditions to the periodic
ii
hh ii system
hh of
~ and ~S
equations requires constraining the jumps P
across a face. We present a list of common DG jumps
first presented in [21]. Here, TF/SF denotes total fields and
scattered fields, respectively. The addition of the angle of
incidence in the jumps for the planewave ABC allows the
periodic numerical flux to satisfy the well-known SilverMüller condition for the transformed fields
~
Z n̂ × ~S = |cos θ| n̂ × n̂ × P
~ = − |cos θ| n̂ × n̂ × ~S
Y n̂ × P
for TE and TM polarization, respectively. Here, Z = 1/Y is
the impedance of the medium.
4
We must also consider boundary conditions on the interfaces
between unit cells. To implement Eqns. (4a) and (4b), a map
must be created between the periodic planes of the unit cell.
A natural first choice for creating these maps is to create a
meshed unit cell in whichhh the
ii periodic planes are conformal,
~ = P(~
~ r + ~un , t) − P(~
~ r, t) and
and set the jumps to be P
hh ii
~S = ~S(~r + ~un , t) − ~S(~r, t). Alternatively, it is significantly
easier to generate a meshed unit cell without meticulous constraints on the periodic planes. The nodes of the periodic plane
will not align, and information regarding the non-conformal
triangles is generated. This interface is first decomposed into a
list of the four different types of fragments: three-, four-, five-,
and six-vertex fragments. A polygon clipping algorithm [25]
is employed to generate this data. These fragments are defined
to facilitate the definition of quadrature rules for numerically
integrating surface terms.
IV. R ESULTS
To demonstrate the validity of our computational framework, we discuss several scattering results. In all cases, a
low-storage fourth order Runge-Kutta integration [26] is used
with a time step size determined by c∆t = hP −2 , where h
is the minimum edge length and P is the polynomial order.
Reflection or transmission data presented for each structure is
obtained from Eqn. (12).
~ r/t (f )
E
Pr/t (f ) =
~ i (f )
E
2
2
Fig. 4: Power reflected from a planewave obliquely incident on a nonmagnetic
and lossless dielectric slab, θ = 50◦ . Top: Power reflection over broadband
frequency range for TE polarization (top left) and TM polarization (top
right). Bottom: minimum edge length (h) and polynomial order (P ) error
convergence for TE polarization.
~ 00 (t) given as
fundamental coefficient A
1
~ 00 (t) =
A
|~a1 ||~a2 |
Z|~a2 | Z|~a1 |
~
P(x,
y, z = zRT ; t)dxdy
(13)
y=0 x=0
(12)
~ i (f ) is the Fourier transform of the planewave excitaHere, E
~ r/t (f ),
tion. The reflected and transmitted field, denoted by E
is calculated as the magnitude of the Fourier transform of the
Fig. 3: Reflection coefficient (in dB) of a planewave normally incident on
periodically arranged PEC Minkowski Fractals. The unit cell dimensions for
the fractal are |~a1 | = |~a2 | = 30cm. Dimensions of the fractal are shown
above. The ABC surfaces were placed 10cm away from the PEC fractal in
±z-direction. The electric field is x-polarized.
This coefficient is integrated over the z = zRT plane [10]
located either below or above the scattering structure for
reflection or transmission, respectively.
The first result is scattering of a plane wave normally
incident on a Minkowski fractal FSS. This result validates
our implementation at normal incidence, and serves as a
check of the non-conformal treatment of periodic boundary
conditions independent of the oblique incidence framework.
Fig. 3 displays an illustration of the fractal and its dimensions,
and the unit cell dimensions were |~a1 | = |~a2 | = 30cm. An air
Fig. 5: (left) Illustration of the PEC rods oriented in the y-direction. The unit
cell dimensions are |~a1 | = 8mm and |~a2 | = 2mm. The radius of both rods is
0.8mm. (right) Power reflected from a normally (top) and obliquely (bottom,
θ = 30◦ ) incident planewave. The electric field is y-polarized for both cases.
5
Fig. 6: (left) Illustration of the nonmagnetic and lossless dielectric slab
with periodically arranged PEC strips located at the center of the slab. The
slab has a thickness of 2mm, and the PEC strips are 2.5mm by 5mm, as
shown in the illustration. (right) Power reflected from a normally (top) and
obliquely (bottom, θ = 30◦ ) incident planewave on a nonmagnetic and
lossless dielectric slab with periodically arranged PEC strips residing at the
center of the slab’s thickness. The electric field is y-polarized for both cases.
box was placed above and below the PEC fractal with heights
of 10cm. The DG-TD numerical results are displayed in Fig.
3. Reference data for the Minkowski fractal was drawn from
[12].
The next structure is a simple dielectric slab of thickness
d = 1.0m and relative permittivity εr = 4.0. This slab
is lossless and nonmagnetic. The unit cell dimensions were
chosen arbitrarily to be |~a1 | = |~a2 | = 0.35m. The height
of the air box above and below the slab was chosen to be
1.0m. Fig. 4 displays the power reflected from the slab with
the angle of incidence θ = 50◦ . For this structure, we show
excellent agreement between the theoretical and numerical
power reflection coefficient across the frequency range. To
demonstrate the higher order accuracy of the computational
framework, Fig. 4 displays the average absolute error between
the numerically and theoretically calculated reflection over the
frequency band.
The next structure consists of two infinite PEC rods oriented
in the y-direction. The unit cell dimensions, displayed in Fig.
5, are 8mm by 2mm in the x- and y-direction, respectively.
Length of the structure in the y-direction was chosen to reduce
the number of unknowns, as it is infinite in the y-direction.
Fig. 7: (left) Illustration of a single unit cell of periodically arranged dielectric
slabs (outlined in black) in the x-direction with εr1 = 2.56 and εr2 =
1.44. The slab heights and widths were chosen based on the ratio h/d =
1.713 and d/2.0, respectively. (right) Reflected power of an obliquely incident
planewave (θ = 45◦ ). The electric field is y-polarized.
The air boxes above and below the rods are 11mm from the
centers of the rods, and the centers of the rods were placed
8mm apart. The radius of both rods is 0.8mm. Fig. 5 displays
the numerical results of the periodic DG-TD method compared
against the numerical results of the periodic FEM-TD method.
Our framework demonstrates excellent results compared to the
FEM-TD framework. The effect of the planewave ABC past
the next higher order Floquet mode is also captured.
Our next structure is an array of PEC strips embedded in a
dielectric slab. The dielectric slab is lossless and nonmagnetic,
and the dimensions are shown in Fig. 6. An air box was placed
above and below the dielectric slab with a height of 30mm in
the ±z-direction. Reference data [10] agrees very well with
the numerical results of the DG-TD code shown in Fig. 6.
Again we see the effect of the planewave ABC much like the
FEM-TD framework [10].
Our last validation structure consists of dielectric slabs with
alternating dielectric constants. The dielectric slabs are lossless
and nonmagnetic, and the unit cell is displayed in Fig. 7.
Slab heights h and width of the slabs d are set based on
the ratio h/d = 1.713, and each slab’s width was set to
0.5d. An air box was placed above and below the set of slabs
with an arbitrarily chosen height of 0.5d above and d below.
The relative permittivity of each slab was εr1 = 2.56 and
εr2 = 1.44. Results for this structure are shown in Fig. 7,
with reference data drawn from [27]. Our results show good
agreement with the reference data.
We have shown several cases which validate this DGTD
framework. The final topic of this work is addressing the
stability of the explicit time integrator with respect to the
planewave’s angle of incidence. The speed of Floquet modes
is proportional to cos−1 θ [10], and therefore the CFL bound
c∆t ≤ hP −2 is not sufficient for higher angles of incidence.
The simplest solution of this problem is to scale the CFL
−1
condition as c∆t = hP −2 VCF
L . Fig. 8 displays the smallest
stable time step scale with respect to angle of incidence for a
planewave passing through freespace. The unit cell dimensions
for the freespace mesh were |~a1 | = |~a2 | = λmin /2, the
smallest edge length was h = λmin /10, and the polynomial
order was P = 2. These parameters were held constant for
each angle of incidence. The unit cell mesh was conformal
with respect to the periodic boundaries. This simple result
provides empirical evidence that the explicit time integration
scheme is conditionally stable, even at near grazing angles of
incidence. Satisfying the CFL condition at near grazing angles,
Fig. 8: Angular dependence of time step scale VCF L . Angles less than θ =
20◦ required unity scaling for stability.
6
however, requires scales of two orders of magnitude and thus
increases the number of time steps accordingly.
V. C ONCLUSION AND F UTURE W ORK
In this paper, we have presented a higher-order threedimensional Time Domain Discontinuous Galerkin Method
for analyzing the interaction of obliquely incident planewaves
with doubly periodic structures. We employed a field transformation to provide a formulation free from the well-known
causality issues with periodic boundary conditions in time.
The field transformations were applied to the first order
Maxwell’s Equations, and a numerical flux was derived using
an equivalent set of transformed equations. The computational
framework was validated using existing results in the literature. While the particular examples elaborated in this paper
employed a planewave ABC, we are currently developing
an exact time domain Floquet radiation boundary condition.
Future applications include the optimization of photonic band
gap structures and complex frequency selective surfaces.
VI. ACKNOWLEDGMENT
This work was supported by the National Science Foundation through grant CCF:1018576. The authors would like
to thank General Electric (GE) for support, and acknowledge
computing support from the HPC Center at Michigan State
University, East Lansing.
R EFERENCES
[1] B. Munk. Frequency Selective Surfaces: Theory and Design. John Wiley
& Sons, 2005.
[2] F. Yang and Y. Rahmat-samii. Electromagnetic Band Gap Structures in
Antenna Engineering. 2007.
[3] B. Munk. Metamaterials: Critique and Alternatives. John Wiley &
Sons, 2009.
[4] F. Capolino. Theory and Phenomena of Metamaterials, volume 8. CRC
Press, 2010.
[5] A.D. Baczewski, D.L. Dault, and B. Shanker. Accelerated Cartesian
Expansions for the Rapid Solution of Periodic Multiscale Problems.
IEEE Trans. Antennas Propagat., 60(9):4281–4290, 2012.
[6] A.D. Baczewski, N.C. Miller, and B. Shanker. Rapid analysis of
scattering from periodic dielectric structures using accelerated Cartesian
expansions. JOSA. A, 29(4):531–40, April 2012.
[7] E.W. Lucas and T.P. Fontana. A 3-D hybrid finite element/boundary
element method for the unified radiation and scattering analysis of
general infinite periodic arrays. IEEE Trans. Antennas Propagat.,
43(2):145–153, 1995.
[8] P. Sotirelis and J.D. Albrecht. Numerical simulation of photonic crystal
defect modes using unstructured grids and Wannier functions. Phys.
Rev. B, 76(7):075123, August 2007.
[9] S. Chun. High-order Accurate Methods for solving Maxwell’s equations
and their applications. (May), 2008.
[10] L.E.R. Petersson and J.M. Jin. Analysis of Periodic Structures via a
Time-Domain Finite-Element Formulation With a Floquet ABC. IEEE
Trans. Antennas Propagat., 54(3):933–944, March 2006.
[11] L.E.R. Petersson and J.M. Jin. A Three-Dimensional Time-Domain
Finite-Element Formulation for Periodic Structures. IEEE Trans. Antennas Propagat., 54(1):12–19, January 2006.
[12] N.W. Chen, M. Lu, F. Capolino, B. Shanker, and E. Michielssen.
Floquet wave-based analysis of transient scattering from doubly periodic,
discretely planar, perfectly conducting structures. Radio Sci., 40(4),
August 2005.
[13] P. Harms and R. Mittra. Implementation of the periodic boundary condition in the finite-difference time-domain algorithm for FSS structures.
IEEE Trans. Antennas Propagat., 42(9):1317–1324, 1994.
[14] D.L. Dault, N. V. Nair, and B. Shanker. An O(NS Nt log 2 Nt ) method
for evaluating convolutions with the time domain periodic Green’s
function. In 2012 International Conference on Electromagnetics in
Advanced Applications, pages 141–143. IEEE, September 2012.
[15] A.J. Pray, N.V. Nair, and B. Shanker. Stability Properties of the Time
Domain Electric Field Integral Equation Using a Separable Approximation for the Convolution With the Retarded Potential. IEEE Trans.
Antennas Propagat., 60(8):3772–3781, August 2012.
[16] B.H. Jung, Z. Mei, and T.K. Sarkar. Transient Wave Propagation in a
General Dispersive Media Using the Laguerre Functions in a Marchingon-in-Degree (MOD) Methodology. Progress In Electromagnetics Research, 118:135–149, 2011.
[17] J.M. Jin. Theory and Computation of Electromagnetic Fields. John
Wiley & Sons, Inc., Hoboken, NJ, USA, November 2010.
[18] M.E. Veysoglu, R.T. Shin, and J.A. Kong. A Finite-Difference TimeDomain Analysis of Wave Scattering from Periodic Surfaces: Oblique
Incidence Case. Journal of Electromagnetic Waves and Applications,
7(12):1595–1607, January 1993.
[19] K. Sirenko, H. Bagci, and Y. Sirenko. Accurate Characterization of
3D Diffraction Gratings Using Time Domain Discontinuous Galerkin
Method with Exact Absorbing Boundary Conditions. IEEE AP-S/URSI
Conference, 2013.
[20] J.S. Hesthaven and T. Warburton. Nodal High-Order Methods on
Unstructured Grids. J. Computat. Phys., 181(1):186–221, September
2002.
[21] J. Niegemann, M. König, K. Stannigel, and K. Busch. Higher-order
time-domain methods for the analysis of nano-photonic systems. Photonics and Nanostructures - Fundamentals and Applications, 7(1):2–11,
February 2009.
[22] K. Busch, M. König, and J. Niegemann. Discontinuous Galerkin
methods in nanophotonics. Laser & Photonics Reviews, 5(6):773–809,
November 2011.
[23] R.J. LeVeque. Finite Volume Methods for Hyperbolic Problems. Cambridge University Press, 2002.
[24] A.H. Mohammadian, V. Shankar, and W.F. Hall. Computation of
electromagnetic scattering and radiation using a time-domain finitevolume discretization procedure. Computer Physics Communications,
68(1-3):175–196, November 1991.
[25] B.R. Vatti. A generic solution to polygon clipping. Communications of
the ACM, 35(7):56–63, July 1992.
[26] M.H. Carpenter and A. Kennedy. Fourth-Order Kutta Schemes. 1994.
[27] H.L. Bertoni and L.S. Cheo. Frequency Selective Reflection and
Transmission by a Periodic Dielectric Layer. 31(1):78–83, 1989.
| 5cs.CE
|
ROCKET: Robust Confidence Intervals via Kendall’s Tau for
Transelliptical Graphical Models
arXiv:1502.07641v3 [math.ST] 1 Sep 2017
Rina Foygel Barber and Mladen Kolar
September 1, 2017
Abstract
Understanding complex relationships between random variables is of fundamental importance
in high-dimensional statistics, with numerous applications in biological and social sciences. Undirected graphical models are often used to represent dependencies between random variables, where
an edge between two random variables is drawn if they are conditionally dependent given all the
other measured variables. A large body of literature exists on methods that estimate the structure
of an undirected graphical model, however, little is known about the distributional properties of
the estimators beyond the Gaussian setting. In this paper, we focus on inference for edge parameters in a high-dimensional transelliptical model, which generalizes Gaussian and nonparanormal
graphical models. We propose ROCKET, a novel procedure for estimating parameters in the
latent inverse covariance matrix. We establish asymptotic normality of ROCKET in an ultra
high-dimensional setting under mild assumptions, without relying on oracle model selection results.
? ROCKET requires the same number of samples that are known to be necessary for obtaining
a n consistent estimator of an element in the precision matrix under a Gaussian model. Hence,
it is an optimal estimator under a much larger family of distributions. The result hinges on a tight
control of the sparse spectral norm of the non-parametric Kendall’s tau estimator of the correlation
matrix, which is of independent interest. Empirically, ROCKET outperforms the nonparanormal
and Gaussian models in terms of achieving accurate inference on simulated data. We also compare
the three methods on real data (daily stock returns), and find that the ROCKET estimator is
the only method whose behavior across subsamples agrees with the distribution predicted by the
theory.
1
Introduction
Probabilistic graphical models [Lauritzen, 1996] have been widely used to explore complex system and
aid scientific discovery in areas ranging from biology and neuroscience to financial modeling and social
media analysis. An undirected graphical model consists of a graph G “ pV, Eq, where V “ t1, . . . , pu
is the set of vertices and E is the set of edges, and a p-dimensional random vector X “ pX1 , . . . , Xp qJ
that is Markov with respect to G. In particular, we have that Xa and Xb are conditionally independent
given the remaining variables tXc | c P t1, . . . , puzta, buu if and only if ta, bu R E. One of the central
questions in high-dimensional statistics is estimation of the undirected graph G given n independent
realizations of X, as well as quantifying uncertainty of the estimator.
In this paper we focus on (asymptotic) inference for elements in the latent inverse covariance
matrix under the semiparametric elliptical copula model [Embrechts et al., 2003, Klüppelberg et al.,
2008], also known as the transelliptical model [Liu et al., 2012b]. Let X1 , . . . , Xn be n independent
copies of the random vector X that follows a transelliptical distribution,
X „ TEpΣ, ξ; f1 , . . . , fp q,
(1.1)
where Σ P Rp is a correlation matrix (that is, Σjj “ 1 for j “ 1, . . . , p), ξ P R is a nonnegative
random variable with Ptξ “ 0u “ 0, and f1 , . . . , fp are univariate, strictly increasing functions. Recall
1
that X follows a transelliptical distribution if the marginal transformation pf1 pX1 q, . . . , fp pXp qq of X
follows a (centered) elliptically contoured distribution with covariance matrix Σ [Fang et al., 1990].
Let Ω “ Σ´1 be the inverse covariance matrix, also known as the precision matrix. Under a Gaussian
model, nonzero elements in Ω correspond to pairs of variables that are conditionally dependent, that
is, form an edge in the graph G; under an elliptical model, nonzero elements in Ω correspond to
variables that are conditionally correlated (but in general it is possible to have Ωab “ 0 where fa pXa q
and fb pXb q are conditionally uncorrelated, but not conditionally independent). Under the model in
(1.1), we construct an estimator for a fixed element of the precision matrix, Ωab , that is asymptotically
normal. Furthermore, we construct a confidence interval for the unknown parameter Ωab that is valid
and robust to model selection mistakes. Finally, we construct a uniformly valid hypothesis test for
the presence of an edge in the graphical model.
Our main theoretical result establishes that given
? initial estimates of the regression coefficients for
pfa pXa q, fb pXb q on pfj pXj qqj‰a,b , one can obtain a n-consistent and asymptotically normal estimator
for Ωab . These initial estimators need to converge at a sufficiently fast rate (see Section 3). In
particular, we note that we do not require strict sparsity in these regressions, and allow for an error
rate that is achievable by ?
known methods such as a nonconvex Lasso [Loh and Wainwright, 2013] (see
Section 3.1). To achieve n-consistent rate, our estimator requires the same scaling for the sample
size n as in the Gaussian case; this sample size scaling is minimax optimal [Ren et al., 2013].
Given accurate initial estimates, in order to construct the asymptotically normal estimator, we
prove a key result: that the vector signpXi ´ Xi1 q is subgaussian at the scale CpΣq (the condition
number of Σ), with dependence on the dimension p coming only through CpΣq (this problem was
initially posed by Han and Liu [2013], where subgaussianity was proved for some special cases). This
result allows us to construct an asymptotically normal estimator by combining the initial regression
coefficient estimates with the Kendall’s tau rank correlation matrix. In particular, the subgaussianity
result allows us to establish a new concentration result on the operator norm of the Kendall’s tau
correlation matrix that holds with exponentially high probability. This result allows us to uniformly
control deviations of quadratic forms involving the Kendall’s tau correlation matrix over approximately
sparse vectors. These results are of independent interest and could be used to extend recent results
of Mitra and Zhang [2014], Wegkamp and Zhao [2013] and Han and Liu [2013] to the elliptical copula
setting. Furthermore, subgaussianity of signpXi ´ Xi1 q, which in turn leads to a bound on the error of
the Kendall’s tau estimate of Σ in the sparse spectral norm, allows us to study properties of penalized
rank regression in high-dimensions.
We base our confidence intervals and hypothesis tests on the asymptotically normal estimator of
the element Ωab (see Section 2). We point out that our results hold under milder conditions than
those required in Ren
? et al. [2013], which treats the special case of Gaussian graphical models. Most
notably, we give a n-consistent estimator for elements in the precision matrix without requiring
strong parametric assumptions.
1.1
Relationship To Literature
Our work contributes to several areas. First, we contribute to the growing literature on graphical
model selection in high dimensions. There is extensive literature on the Gaussian graphical model,
where it is assumed that X „ N p0, Σq, in which case the edge set E of the graph G is encoded by the
non-zero elements of the precision matrix Ω [Meinshausen and Bühlmann, 2006, Yuan and Lin, 2007,
Rothman et al., 2008, Friedman et al., 2008, d’Aspremont et al., 2008, Fan et al., 2009, Lam and Fan,
2009, Yuan, 2010, Cai et al., 2011, Liu and Wang, 2012, Zhao and Liu, 2014]. Learning structure of
the Ising model based on the penalized pseudo-likelihood was studied in Höfling and Tibshirani [2009],
Ravikumar et al. [2010] and Xue et al. [2012]. More recently, Yang et al. [2013] studied estimation
of graphical models under the assumption that each of the nodes’ conditional distribution belongs to
an exponential family distribution. See also Guo et al. [2011a], Guo et al. [2011b], Lee and Hastie
[2012], Cheng et al. [2013], Yang et al. [2012] and Yang et al. [2014] who studied mixed graphical
models, where the nodes’ conditional distributions are not necessarily all from the same family (for
2
instance, there may be continuous-valued nodes as well as discrete-valued nodes). The parametric
Gaussian assumption was relaxed in Liu et al. [2009], where graph estimation was studied under a
Gaussian copula model. More recently, Liu et al. [2012a], Xue and Zou [2012] and Liu et al. [2012b]
show that the graph can be recovered in the Gaussian and elliptical semiparametric model class under
the same conditions on the sample size n, number of nodes p and the maximum node degree in the
graph ?
k as if the estimation was done under the Gaussian assumption. In our paper, we construct a
novel n-consistent estimator of an element in the precision matrix without requiring oracle model
selection properties.
Second, we contribute to the literature on high-dimensional inference. Recently, there has been
much interest on performing valid statistical inference in the high-dimensional setting. Zhang and
Zhang [2013], Belloni et al. [2013a], Belloni et al. [2013d], van de Geer et al. [2014], Javanmard
and Montanari [2014], Javanmard and Montanari [2013], and Farrell [2013] developed methods for
construction of confidence intervals for low dimensional parameters in high-dimensional linear and
generalized linear models, as well as hypothesis tests. These methods construct honest, uniformly
valid confidence intervals and hypothesis test based on the `1 -penalized estimator in the first stage.
Similar results were obtained in the context of the `1 -penalized least absolute deviation and quantile
regression [Belloni et al., 2013c,b]. Lockhart et al. [2014] study significance of the input variables
that enter the model along the lasso path. Lee et al. [2013] and Taylor et al. [2014] perform postselection inference
? conditional on the selected model. Liu [2013], Ren et al. [2013] and Chen et al.
[2013] construct n-consistent estimators for elements of the precision matrix Ω under a Gaussian
assumption. We extend these results to perform valid inference under semiparametric ellitical copula
models. In a recent independent work, Gu et al. [2015] propose a procedure for inference under a
nonparanormal model. We will provide a detailed comparison in Section 3 and Section 5.
1.2
Notation
Let rns denote the set t1, . . . , nu and let 1It¨u denote the indicator function. For a vector a P Rd , we
let supppaq “
řtj : aj ‰ 0u be the support set, and let ||a||q , for q P r1, 8q, be the `q -norm defined
as ||a||q “ p iPrns |ai |q q1{q with the usual extensions for q P t0, 8u, that is, ||a||0 “ |supppaq| and
||a||8 “ maxiPrns |ai |.
For a matrix A P Rn1 ˆn2 , for sets S Ă rn1 s and T Ă rn2 s, we write AST to denote the |S| ˆ |T |
submatrix of A obtained by extracting the appropriate rows and columns. The sets S and/or T can
be replaced by single indices, for example, for S Ă rn1 s and j P rn2 s, ASj is a |S|-length vector. If
A P Rnˆn is a square matrix, for any T Ă rns we may write AT to denote the square submatrix AT T .
For a matrix A P Rn1 ˆn2 , we use the notation vecpAq to denote the vector
in Rn1 n2 formed by
ř
2
stacking the columns of A. We denote the Frobenius norm of A by ||A||F “ iPrn1 s,jPrn2 s A2ij , and the
operator norm (spectral norm) by ||A||op , that ř
is, the largest singular value of A. The norms ||A||1
and ||A||8 are applied entrywise, with ||A||1 “ ij |Aij | and ||A||8 “ maxij |Aij |. We write CpAq to
denote the condition number of A, that is, the ratio between the largest and smallest singular values.
For two matrices A P Rnˆm and B P Rrˆs , A b B P Rnrˆms denotes the Kronecker product, with
pA b Bqik,jl “ Aij Bkl . For two matrices of the same size, A, B P Rnˆm , A ˝ B P Rnˆm denotes the
Hadamard product (that is, the entrywise product), with pA ˝ Bqij “ Aij Bij . Kronecker products and
Hadamard products are defined also for vectors, by treating a vector as a matrix with one column.
Throughout, Φp¨q denotes the cumulative distribution function of the standard normal distribution,
that is, Φptq “ PtN p0, 1q ď tu.
2
Preliminaries and method
Before introducing our method, we begin with some preliminary definitions and properties of the
transelliptical distribution, and related models.
3
Gaussian and nonparanormal graphical models Suppose that X “ pX1 , . . . , Xp q follows a
multivariate normal distribution, X „ N pµ, Σq. A Gaussian graphical model represents the structure
of the covariance matrix Σ with a graph, where an edge between nodes a and b indicates that Ωab ‰ 0,
where Ω “ Σ´1 is the precision (inverse covariance) matrix. This model can be generalized by
allowing for arbitrary marginal transformations on the variables X1 , . . . , Xp . Liu et al. [2009] study
the resulting distribution, the nonparanormal model (also known as a Gaussian copula), where we
write X „ NPNpΣ; f1 , . . . , fp q, if the marginally transformed vector pf1 pX1 q, . . . , fp pXp qq follows a
(centered) multivariate normal distribution,
pf1 pX1 q, . . . , fp pXp qq „ N p0, Σq .
The sparse structure of the underlying graphical model, representing the sparsity pattern in Ω “ Σ´1 ,
can then be recovered using similar methods as in the Gaussian case. Note that the Gaussian model
is a special case of the nonparanormal model (by setting f1 , . . . , fp each to be the identity function,
or to be linear functions if we would like a nonzero mean).
Elliptical and transelliptical graphical models The elliptical model is a generalization of the
Gaussian graphical model that allows for heavier-tailed dependence between variables. The random
vector X “ pX1 , . . . , Xp q follows an elliptical distribution with the mean vector µ P Rp , covariance
matrix Σ P Rpˆp , and a random variable (the “radius”) ξ ě 0, denoted by X „ Epµ, Σ, ξq, if we can
write X “ µ ` ξ ¨ A ¨ U , where AAJ “ Σ is a Cholesky decomposition of Σ, and where U P Rp is a
unit vector drawn uniformly at random (independently from the radius ξ). Note that the level sets of
this distribution are given by ellipses, centered at µ and with shape determined by Σ. The Gaussian
model is a special case of the elliptical model (by taking ξ „ χp ).
The transelliptical model (also known as an elliptical copula) combines the elliptical distribution
with marginal transformations, much as the nonparanormal distribution applies marginal transformations to a multivariate Gaussian. For a random vector X P Rp we write
X „ TEpΣ, ξ; f1 , . . . , fp q
to denote that the marginally transformed vector pf1 pX1 q, . . . , fp pXp qq follows a centered elliptical
distribution, specifically,
pf1 pX1 q, . . . , fp pXp qq „ Ep0, Σ, ξq .
Here the marginal transformation functions f1 , . . . , fp are assumed to be strictly increasing. Note that
the Gaussian, nonparanormal, and elliptical models are each special cases of this model.
Pearson’s rho and Kendall’s tau From this point on, we assume for each distribution that µ “ 0
and that Σ is a correlation matrix (that is, diagonal elements are equal to one, Σaa “ 1). In the
case of the Gaussian distribution X „ N p0, Σq, the entries of Σ are the (population-level) Pearson’s
correlation coefficients for each pair of variables, which in this case we can also write as Σab “ ErXa Xb s.
In this setting, we can estimate Σ with the sample covariance.
In the nonparanormal setting, X „ NPNpΣ; f1 , . . . , fp q, it is no longer the case that Σab is equal
to the (population-level) correlation CorrpXa , Xb q, due to the marginal transformations. However, we
can estimate f1 , . . . , fp by performing marginal empirical transformations of each Xa to the standard
normal distribution. After taking these empirical transformations, Σ can again be estimated via
the empirical covariances. Similarly, for the elliptical model X „ Ep0, Σ, ξq, after rescaling so that
Erξ 2 s “ p we also have Σab “ ErXa Xb s. We can therefore again estimate Σ via the empirical
covariance.
For the transelliptical distribution, in contrast, this is no longer possible. Taking scaling Erξ 2 s “ p
for simplicity, we generalize the calculations above to have Σab “ Erfa pXa qfb pXb qs. Therefore, if we
can estimate the marginal transformations f1 , . . . , fp , then we can estimate Σ using the empirical
4
covariance of the transformed data. However, unlike the nonparanormal model, estimating f1 , . . . , fp
is not straightforward. The reason is that, for the elliptical distribution Ep0, Σ, ξq, the marginal
distributions are not known unless the distribution of the radius ξ is known. Therefore, marginally
for each Xa , we cannot estimate fa because we do not know what should be the marginal distribution
after transformation, that is, what should be the marginal distribution of fa pXa q. (In contrast, in the
nonparanormal model, fa pXa q is marginally normal.)
As an alternative, Liu et al. [2012b] use the Kendall rank correlation coefficient (Kendall’s tau).
At the population level, it is given by
“
‰
τab :“ τ pXa , Xb q “ E signpXa ´ Xa1 q ¨ signpXb ´ Xb1 q ,
where X 1 is an i.i.d. copy of X. Unlike Pearson’s rho, the Kendall’s tau coefficient is invariant to
marginal transformations: since fa , fb are strictly increasing functions, we see that
signpfa pXa q ´ fa pXa1 qq ¨ signpfb pXb q ´ fb pXb1 qq “ signpXa ´ Xa1 q ¨ signpXb ´ Xb1 q .
At the sample level, Kendall’s tau can be estimated by taking a U-statistic comparing each pair of
distinct observations:
ÿ
1
signpXia ´ Xi1 a q ¨ signpXib ´ Xi1 b q .
(2.1)
τpab “ `n˘
2
1ďiăi1 ďn
When X follows an elliptical distribution, Theorem 2 of Lindskog et al. [2003] gives us the following
relationship between Kendall’s tau and the Pearson’s rho coefficients given by the covariance matrix
Σ:
´π ¯
Σab “ sin
τab for each a, b P rps .
(2.2)
2
Since Kendall’s tau is invariant to marginal transformations, this identity holds for the transelliptical
family as well. For this reason, Liu et al. [2012b] estimate the covariance matrix Σ by
´
¯
p ab “ sin π τpab .
Σ
(2.3)
2
p is not necessarily positive semidefinite.
Note, however, that Σ
While Spearman’s rho, like Kendall’s tau, is also invariant to marginal transformations, Liu et al.
[2012b] comment that there is no equivalence between Σ and the population-level Spearman’s rho
values (analogous to (2.2) for Kendall’s tau) which holds uniformly across the entire elliptical (or
transelliptical) family. Therefore, this type of estimator as in (2.3) could only be carried out with
Kendall’s tau.
p denotes the estimate given here in (2.3). The matrix of the
For the remainder of this paper, Σ
Kendall’s tau coefficients is denoted as T , with entries Tab :“ τab , and Tp denotes its empirical estimate
(with entries as in (2.1)).
Comparing models: tail dependence It is clear that, compared to a Gaussian graphical model,
the nonparanormal model allows for data that may be extremely heavy-tailed (in the marginal distributions). A more subtle consideration is the question of tail dependence between two or more of the
variables. In particular, the nonparanormal model does not allow for tail dependence between two
variables to be any stronger than in the Gaussian distribution itself. Specifically, consider pairwise
α-tail dependence between Xa and Xb , given by
`
(
(˘
Tailα pXa , Xb q :“ Corr 1I Xa ě qαXa , 1I Xb ě qαXb
,
where qαXa is the α-quantile of the marginal distribution of Xa , and same for Xb . Taking α Ñ 1, this
is a measure of the correlation between the extreme right tail of Xa and the extreme right tail of Xb .
(Of course, we can also consider the left tail of the distribution of Xa and/or Xb .)
5
Note that marginal transformations of each variable do not affect this measure, since the quantiles
qαXa , qαXb take these transformations into account. In particular, the nonparanormal distribution has
the same tail correlations Tailα pXa , Xb q as the multivariate Gaussian distribution (with the same Σ).
In contrast, an elliptical or transelliptical model can exhibit much higher tail correlations. Since real
data often exhibits heavy tail dependence between variables, the flexible transelliptical model may be
a better fit in many applications.
We demonstrate this behavior with a simple example in Figure 1. Take
? ˙
ˆ
1? 1{ 2
,
(2.4)
X “ pX1 , X2 q „ Ep0, Σ, ξq with Σ “
1{ 2
1
?
where ξ „ χ2 ¨ d{χd for d P t0.1, 1, 5, 10, 8u, corresponding to a multivariate t-distribution with d
degrees of freedom (note that d “ 8 is equivalent to taking X „ N p0, Σq). Note that at α “ 0.5,
the relevant quantiles are qαX1 “ qαX2 “ 0, and so the tail correlation Tailα pX1 , X2 q is equal to the
Kendall’s tau coefficient τ pX1 , X2 q “ π2 arcsinpΣ12 q “ 0.5 at any value of d. Figure 1 shows that, as
α Ñ 1, the tail correlation decreases towards zero for the normal distribution (d “ 8) but grows for
low values of d.
Therefore, the shift from a nonparanormal to a transelliptical model is important, since it allows
us to model variables with high tail dependence, that is, high dependence between their “extreme
events”.
1
0.9
0.8
d=0.1
Tail dependence
0.7
0.6
d=1
0.5
d=5
0.4
0.3
d=10
d=
(Gaussian)
0.2
0.1
0
0.5
0.6
0.7
0.8
Quantile
0.9
1
Figure 1: Tail dependence for normal and elliptical distributions on R2 . Data is generated as in (2.4).
The figure displays Tailα pX1 , X2 q, estimated empirically from a sample size n “ 20000.
2.1
ROCKET: an asymptotically normal estimator
Suppose that our data points Xi are drawn i.i.d. from a transelliptical distribution with covariance
matrix Σ. We would like to perform inference on a particular entry of the precision matrix Ω “ Σ´1 ,
specifically, we are interested in producing a confidence interval for Ωab where a ‰ b P t1, . . . , pu is a
prespecified node pair.
To move towards constructing a confidence interval, we introduce a few definitions and calculations.
First, let I “ t1, . . . , puzta, bu, and observe that by block-wise matrix inversion, we can calculate the
6
ta, bu ˆ ta, bu sub-block of Ω as follows:
`
˘´1
Ωab,ab “ Σab,ab ´ Σab,I Σ´1
.
I ΣI,ab
(2.5)
´1
Define γa “ Σ´1
I ΣIa and γb “ ΣI ΣIb . In the nonparanormal graphical model setting, these are the
regression coefficients when fa pXa q or fb pXb q is regressed on tfj pXj q : j P Iu; in the linear model
setting, this idea has been used in Sun and Zhang [2012b] and Belloni et al. [2013a]. We then have
J
J
J
Σab,I Σ´1
I ΣI,ab “ pγa γb q ΣI,ab “ ΣI,ab pγa γb q “ pγa γb q ΣI pγa γb q .
We can therefore rewrite (2.5) as follows (this somewhat redundant formulation will allow for a
favorable cancellation of error terms later on):
J
Θ :“ pΩab,ab q´1 “ Σab,ab ´ pγa γb qJ ΣI,ab ´ ΣJ
I,ab pγa γb q ` pγa γb q ΣI pγa γb q .
(2.6)
We abuse notation and index the entries of Θ with the indices a and b, that is, we denote Θ as lying
in Rta,buˆta,bu rather than R2ˆ2 .
Next, we define an oracle estimator of Θ, defined by plugging the true values of γa and γb and the
empirical estimate of Σ (given in (2.3)) into (2.6) above:
r “Σ
p ab,ab ´ pγa γb qJ Σ
p I,ab ´ Σ
p J pγa γb q ` pγa γb qJ Σ
p I pγa γb q .
Θ
I,ab
(2.7)
Later on (in Theorem 4.1), we will show that due to standard results on the theory of U-statistics,
r were known, then, we would have achieved our
this oracle estimator is asymptotically normal.
If¯ Θ
´
´1
e
e
weakly converges to a Normal random variable
goal for inference in this model, as Ωab “ Θ
ab
centered at Ωab with variance that scales as Op1{nq (we calculate this variance later).
Of course, in practice we do not know the true values of γa and γb , and must instead use some
available estimators, denoted by γ
qa and γ
qb (we discuss how to obtain these preliminary estimates later
on). Given the estimators of the regression vectors, we then define our estimator of Θ as follows:
q “Σ
p ab,ab ´ pq
p I,ab ´ Σ
p J pq
p I pq
γa γ
qb qJ Σ
qb q ` pq
γa γ
qb qJ Σ
γa γ
qb q .
Θ
I,ab γa γ
(2.8)
Since we are interested in Ωab rather than in the matrix Θ, as a final step we define our estimator
´
¯
q ab “ Θ
q ´1
Ω
.
(2.9)
ab
q ab , which is a function
In order to make inference about Ωab , we approximate the distribution of Ω
q We first treat the distribution of the corresponding entry in the oracle estimator Θ.
r To do so,
of Θ.
let u, v P Rpn be the vectors with entries
ua “ 1, ub “ 0, uI “ ´γa and va “ 0, vb “ 1, vI “ ´γb ,
and observe from (2.6) and (2.7) that Θab “ uJ sin
a linear approximation to sinp¨q, we can write
´ ¯
`π ˘
π p
J
r
2 T v while Θab “ u sin 2 T v. Now, taking
A
´
¯
E
r ab ´ Θab « uv J ˝ π cos π Tp , Tp ´ T
Θ
2
2
´π ¯
E
1 ÿA J π
“ ` n˘
uv ˝ cos
Tp , signpXi ´ Xi1 q signpXi ´ Xi1 qJ ´ T .
2
2
2 iăi1
To study the variability of this error, we consider the kernel
´
´ π ¯¯
T signpX ´ X 1 q.
gpX, X 1 q “ signpX ´ X 1 qJ uv J ˝ cos
2
7
We will see later on that understanding the behavior of this kernel will allow us to characterize the
r ab , and from there, our empirical estimator Θ
q ab and ultimately
distribution of the oracle estimator Θ
1
q
Ωab . Of course, gpX, X q itself depends on unknown quantities, namely u, v, and T , so we replace
these with their estimates in our empirical version of the kernel: define the (random) kernel
´
´ π ¯¯
gqpX, X 1 q “ signpX ´ X 1 qJ u
qvqJ ˝ cos
Tp signpX ´ X 1 q ,
2
where
qa “ 1, u
qb “ 0, u
qI “ ´q
u
γa and vqa “ 0, vqb “ 1, vqI “ ´q
γb .
q ab “ u
p v .) Then define
q and vq so that Θ
qJ Σq
(Note that we have defined u
g
˜
¸2
f
f1 ÿ
ÿ
1
π
e
gqpXi , Xi1 q ´ meanpq
gq
Sqab “
¨
q
n i n ´ 1 i1 ‰i
detpΘq
` ˘´1 ř
2
q ab
g q “ n2
qpXi , Xi1 q. We will see later on that Sqab
where meanpq
{n estimates the variance of Θ
iăi1 g
and that the expression above arises naturally from the theory of U-statistics.
?
q
Our main result, Theorem 3.5 below, will prove that n ¨ ΩabSq´Ωab follows a distribution that is
ab
approximately standard normal. Therefore, an approximate p1´α)-confidence interval for Ωab is given
by
q
q ab ˘ zα{2 ¨ S
?ab ,
(2.10)
Ω
n
(
where zα{2 is the appropriate quantile of the normal distribution, that is, P N p0, 1q ą zα{2 “ α{2.
Notation for fixed vs random quantities From this point on, as much as possible throughout
the main body of the paper, quantities that depend on the data and depend on the initial estimates
q Quantities that depend on the data, but do
γ
qa , γ
qb are denoted with a “check” accent, for example, Θ.
p Any quantities with neither a
qa , γ
qb , are denoted with a “hat” accent, for example, Σ.
not depend on γ
“hat” nor a “check” are population quantities, that is, they are not random. Two important exceptions
r which is of course data-dependent (but
are the data itself, X1 , . . . , Xn , and the oracle estimator, Θ,
does not depend on γ
qa , γ
qb ).
3
Main results
In this section, we give a theoretical result showing that the confidence interval constructed in (2.10)
has asymptotically the correct coverage probability, as long as we have reasonably accurate estimators
´1
of γa “ Σ´1
I ΣIa and γb “ ΣI ΣIb . Our asymptotic result considers a problem whose dimension pn ě 2
grows with the sample size n. We also allow for the sparsity level in the true inverse covariance matrix
Ω P Rpn ˆpn to grow.1 We use kn to denote an approximate bound on the sparsity in each column of
Ω (details given below).
We begin by stating several assumptions on the distribution of the data and on the initial estimators
qa and γ
γ
qb . All of the constants appearing in these assumptions should be interpreted as values that
do not depend on the dimensions pn, pn , kn q of the problem.
Assumption 3.1. The data points X1 , . . . , Xn P Rpn are i.i.d. draws from a transelliptical distribuiid
tion, Xi „ TEpΣ, ξ; f1 , . . . , fpn q, where f1 , . . . , fpn are any strictly monotone functions, ξ ě 0 is any
1 While Σ, Ω, etc, all depend on the sample size n since the dimension of the problem grows, we abuse notation and
do not write Σn , Ωn , etc; the dependence on n is implicit.
8
random variable with Ptξ “ 0u “ 0, and the covariance matrix Σ P Rpn ˆpn is positive definite, with
diagpΣq “ 1 and bounded condition number,
CpΣq “
λmax pΣq
ď Ccov ,
λmin pΣq
for some constant Ccov .
Assumption 3.2. The a-th and b-th columns of the true inverse covariance Ω, denoted by Ωa and
Ωb , are approximately kn -sparse, with
a
||Ωa ||1 _ ||Ωb ||1 ď Csparse kn ,
for some constant Csparse .
Assumption 3.3. For some constant Cest and for some δn ą 0, with probability at least 1 ´ δn , for
each c “ a, b, the preliminary estimate γ
qc of the vector γc satisfies
c
c
kn logppn q
kn2 logppn q
, ||q
γc ´ γc ||1 ď Cest
.
(3.1)
||q
γc ´ γc ||2 ď Cest
n
n
2
Assumption 3.4. Define the kernel hpX, X 1 q “ signpX ´ X 1 q b signpX ´ X 1 q P Rpn and let h1 pXq “
E rhpX, X 1 q | Xs. Define the total variance Σh “ VarphpX, X 1 qq and the conditional Σh1 “ Varph1 pXqq,
iid
where X, X 1 „ TEpΣ, ξ; f1 , . . . , fpn q. Then for some constant Ckernel ą 0,2
Ckernel ¨ Σh ĺ Σh1 ĺ Σh .
Assumption 3.1 assumes that the smallest and largest eigenvalues of the correlation matrix Σ are
bounded away from zero and infinity, respectively. This assumption is commonly assumed in the
literature on learning structure of probabilistic graphical models [Ravikumar et al., 2011, Liu et al.,
2009, 2012a]. Assumption 3.2 does not require that the precision matrix Ω be exactly sparse, which
is commonly assumed in the literature on exact graph recovery [see, for example, Ravikumar et al.,
2011], but only requires that rows Ωa and Ωb have an `1 norm that does not grow too fast. Note that
if Ωc , for c “ a, b, is kn -sparse vector, then
a
a
a
||Ωc ||1 ď kn ||Ωc ||2 ď kn λmax pΩq ď Ccov kn
and we could then set Csparse “ Ccov . Assumption 3.3 is a high-level condition, which assumes existence
of initial estimators of γa and γb that converge at a fast enough rate. In the next section, we will see that
Assumption 3.1 together with a stronger version of Assumption 3.2 are sufficient for Assumption 3.3 to
be satisfied with a specific estimator that is efficient to compute. Finally, Assumption 3.4 is imposed
q ab . While Assumption 3.1 depends only on the
to allow for estimation of the asymptotic variance Ω
correlation matrix Σ without reference to the distribution of the radius ξ, Assumption 3.4 depends on
both Σ and ξ, and therefore cannot be derived as a consequence of the choice of Σ.
We now state our main result.
Theorem 3.5. Under Assumptions 3.1, 3.2, 3.2, and 3.4, there exists a constant Cconverge , depending
on Ccov , Csparse , Cest , Ckernel but not on the dimensions pn, pn , kn q of the problem, such that
d
ˇ
ˇ #
+
ˇ
ˇ
q ab ´ Ωab
? Ω
kn2 log2 ppn q
1
ˇ
ˇ
n¨
`
` δn .
ď t ´ Φptqˇ ď Cconverge ¨
sup ˇP
q
ˇ
ˇ
n
p
n
tPR
Sab
2 Here we use the positive semidefinite ordering on matrices, that is, A ľ B if A ´ B ľ 0. Note that the second part
of the inequality, Σh1 ĺ Σh , is always true by the law of total variance.
9
We note that the result holds uniformly over a large class of data generating processes satisfying
Assumptions 3.1, 3.2, 3.3, and 3.4, which are relatively weak assumptions compared to much of the
sparse estimation and inference literature; we emphasize that the result holds without requiring exact
model selection or oracle properties, which hold only for restrictive sequences of data generating
processes. For example, we do not require the “beta-min” condition (that is, a lower bound on |Ωab |
for all true edges) or any incoherence conditions [Bühlmann and van de Geer, 2011], which may
be implausible in practice. Instead of requiring perfect model selection, we only require estimation
consistency as given in Assumption 3.3; our weaker assumptions would not be sufficient to guarantee
model selection consistency.
As an immediate corollary, we see that the confidence interval constructed in (2.10) is asymptotically correct:
Corollary 3.6. Under the assumptions and notation of Theorem 3.5, the p1 ´ αq-confidence interval
constructed in (2.10) fails to cover the true parameter Ωab with probability no higher than
fi
»
d
2 log2 pp q
k
1
n
n
`
` δn fl .
α ` 2 –Cconverge ¨
n
pn
Again this result holds uniformly over a large class of data generating distributions.
Theorem 3.5 is striking as it shows that we can form an asymptotically normal
` estimator˘ of Ωab
under the transelliptical distribution family with the sample complexity n “ Ω kn2 log2 ppn q . This
sample size requirement was shown to be optimal for obtaining an asymptotically normal estimator
of an element in a precision matrix from multivariate normal data [Ren et al., 2013]. More precisely,
let3
"
*
ř
Ω “ pΩab qa,bPrpn s : maxaPrpn s b‰a 1ItΩab ‰ 0u ď kn ,
G0 pc0 , c1 , kn q “
and c0 ď λmin pΩq ď λmax pΩq ď c1 .
where c0 , c1 ą 0 are constants. Then Theorem 1 in Ren et al. [2013] proves
ˇ
!ˇ
´
¯)
ˇq
ˇ
´1
inf inf
sup
P ˇΩ
kn logppn q _ n´1{2
ě 0
ab ´ Ωab ˇ ě 0 n
a,b Ω
q ab G0 pc0 ,c1 ,kn q
and, therefore, our estimator is rate optimal in terms of the sample size scaling. (Above, the infimum
q ab which is a measurable function of the data.) We can also consider
is taken over any estimator Ω
a related optimality question: whether the confidence interval we produce has the optimal (that is,
lowest possible) width, given the desired coverage level. In the Gaussian setting, Ren et al. [2013]’s
method produces an interval which has asymptotically minimal length at the given sample size, due to
the fact that the variance of their estimator matches the Fisher information. Our ROCKET method
does not enjoy this theoretical property, but empirically we observe that our confidence intervals are
only slightly wider than those produced by Ren et al. [2013]’s method, for Gaussian data.
At this point, it is also worth mentioning the result of Gu et al. [2015], who study inference under
Gaussian copula graphical models. They base their inference procedure on decorrelating a pseudo
score function for the parameter of interest and showing that it is normally distributed. Their main
result, stated in Theorem 4.10, requires the sample size to satisfy
˙3{2
ˆ
´
¯
logppn q
logppn q
` kn2 Mn3
“ o n´1{2
kn3 Mn6
n
n
?
ř
where Mn “ maxaPrpn s bPrpn s |Ωab |. As Mn can be potentially as large as kn , it is immediately
clear that our result achieves much better scaling on the sample size.
their work, the constants c0 , c1 are instead denoted by a constant M ě c´1
0 _ c1 ; we use different notation here
to distinguish from the M used in Gu et al. [2015] which plays a very different role, and which we denote by Mn as it
is not necessarily constant.
3 In
10
3.1
Initial estimators
qa and γ
qb ,
The validity of our inference method relies in part on the accuracy of the initial estimators γ
which are assumed to satisfy error bounds with high probability as stated in Assumption 3.3—that
is, with high probability, we have
c
c
kn logppn q
kn2 logppn q
||q
γc ´ γc ||2 ď Cest
, ||q
γc ´ γc ||1 ď Cest
,
n
n
for c “ a, b, where Cest is some constant. Below, we will prove that these required error rates can be
obtained, under an additional sparsity assumption, by the Lasso estimators
*
"
1 Jp
Jp
qc “
γ
argmin ?
γ ΣI γ ´ γ ΣIc ` λ||γ||1
(3.2)
γPRI ;||γ||1 ďCcov 2kn 2
for each c “ a, b, when the penalty parameter λ is chosen appropriately. In fact, these optimization
p I will not necessarily be positive semidefinite.
problems may not be convex, because Σ
We now turn to proving that any local minima for (3.2) for c “ a, b will satisfy the required error
rates of Assumption 3.3. To proceed, we will use the theoretical results of Loh and Wainwright [2013],
which gives a theory for local minimizers of nonconvex regularized objective functions. In particular,
any local minimizers of the two optimization problems will satisfy requirements of Assumption 3.3
and, therefore, we only need to be able to run optimization algorithms that find local minima. We
specialize their main result to our setting.
Theorem 3.7 (Adapted from Loh and Wainwright [2013, Theorem 1]). Consider any n, p ě 1, any
A P Rpˆp and z P Rp , and any k-sparse x‹ P Rp with ||x‹ ||1 ď R. Suppose that A satisfies restricted
strong convexity conditions
logppq
v J Av ě α1 ||v||22 ´ τ1 ||v||21 ¨
.
(3.3)
n
If
16R2 τ1 maxtα1 , τ1 u logppq
ně
(3.4)
α12
and
#
max 4||Ax‹ ´ z||8 , 4α1
c
logppq
n
+
ďλď
α1
6R
(3.5)
then for any x
q that is a local minimum of the objective function 12 xJ Ax ´ xJ z ` λ||x||1 over the set
d
tx P R : ||x||1 ď Ru, it holds that
?
6λk
1.5λ k
‹
x ´ x‹ ||1 ď
x ´ x ||2 ď
and ||q
.
||q
α1
α1
We apply Loh and Wainwright [2013]’s results, Theorem 3.7, to our problem of estimating γa
and γb in a setting where we assume exact sparsity. (It is likely that similar results would hold for
approximate sparsity, but here we use exact sparsity to fit the assumptions of this existing theorem.)
Corollary 3.8. Suppose that Assumption 3.1 holds. Assume additionally that the columns Ωa , Ωb
of the true inverse covariance Ω “ Σ´1 are kn -sparse. Then there exist constants Csample , CLasso ,
depending on Ccov but not on pn, kn , pn q, such that if n ě Csample kn logppn q then, with probability at
qa of the objective function
least 1 ´ 2p1n , any local minimizer γ
1 Jp
p Ia ` λ||γ||1
γ ΣI γ ´ γ J Σ
2
11
?
over the set tγ P RI : ||γ||1 ď Ccov 2kn u satisfies
a
a
?
γa ´ γa ||2 ď 3 2Ccov λ kn and ||q
γa ´ γa ||1 ď 24Ccov λ kn .
||q
b
nq
where we choose λ “ CLasso ¨ logpp
. The same holds for estimating γb .
n
Using this corollary, we see that a local minimizer of (3.2), γ
qc , satisfies Assumption 3.3 with δn “ p1n
and Cest “ 24Ccov CLasso . We remark that in practice, the constant
CLasso suggested by the theory is in
a
general unknown,
but
choosing
λ
to
be
a
small
multiple
of
logpp
n q{n generally performs well—for
a
instance λ “ 2.1 logppn q{n as we use in our simulations, where the choice of the constant 2.1 ensures
that the penalty term dominates the variance of the elements of the objective function’s derivative,
pIγ ´ Σ
p Ic , at the true solution γ “ γc .
that is, the elements of Σ
To prove that this corollary follows from Loh and Wainwright [2013]’s result (Theorem 3.7), it is
sufficient to check that the restricted strong convexity condition (3.3) holds with high probability for
p I , and then compute the necessary values for λ and the other parameters of Theorem 3.7.
the matrix Σ
The proof is technical and relies on novel results on concentration of the Kendall’s tau correlation
matrix. Details are given in Appendix E.
We have provided sufficient condition for a local minimizer of (3.2) to satisfy Assumption 3.3,
however, many other estimators can be used as initial estimators. For example, one could use the
Dantzig selector [Candès and Tao, 2007]. Potential benefits of the Dantzig selector over the optimizab I is not
tion program in (3.2) are twofold. First, the optimization program is convex even when Σ
positive semi-definite. Second, one does not need to know an upper bound R on the `1 norm of Ωc
for c “ a, b. Using the techniques similar to those used to prove Corollary 3.8, we can also prove that
Assumption 3.3 holds when the Dantzig selector is used as an initial estimator. For large problems,
however, Dantzig selector type methods are computationally much slower than Lasso type methods; in
our empirical results, we implement the Lasso rather than the Dantzig selector since we study graphs
with as many as 1000 nodes.
In practice, we have found that in simulations, using the Lasso for model selection, and then
refitting without a penalty, leads to better empirical performance. Specifically, for each c “ a, b, we
first fit
"
*
1 Jp
Lasso
Jp
qc
γ ΣI γ ´ γ ΣIa ` λ||γ||1 ;
γ
“ argmin
2
γPRI
or, more precisely, find a local minimum of this nonconvex optimization problem over the ball tγ :
||γ||1 ď Ru for a large radius R. (In practice, every iteration will stay inside this ball; therefore, as
long as we see convergence in our iterative algorithm for solving this nonconvex Lasso, we do not
concern ourselves with this theoretical boundedness constraint.)
We then extract the combined support of the two solutions, Jq “ supppq
γaLasso q Y supppq
γbLasso q, and
refit the coefficients using least-squares:
´ ¯´1
pq
p q for c “ a, b .
γ
qc “ Σ
Σ
J
Jc
Following the work of Belloni and Chernozhukov [2013] or Sun and Zhang [2012a], it can be shown that
the refitted estimators also satisfy the Assumption 3.3; in practice, refitting improves the accuracy of
these preliminary estimators by reducing shrinkage bias.
` ˘
Finally,`we˘ remark that if we would like to perform inference for all p2n potential edges, then we
require 2 ¨ p2n „ p2n many initial estimators to be computed; this is of course quite computationally
demanding. However, Ren et al. [2013] propose a simple modification that significantly reduces comqaall .
putation time: for each node a we can first regress Xa on all the other variables; call this solution γ
all
all
γa qb “ 0, then this solution γ
qa is already optimal for regressing node a on
Next for any b ‰ a, if pq
nodes I “ rpn szta, bu; this will be the case for most nodes b due to sparsity. With this modification,
the actual number of regressions required is far smaller—if each node a forms edges with at most kn
12
other nodes (that is, γ
qaall is kn -sparse), then we will require only pn pkn ` 1q many regressions in total
to form all of the initial estimators.
4
Main technical tools
In this section, we outline the proof of Theorem 3.5 (Section 4.1) and state the key technical result
that establishes the sign-subgaussianity property of a vector X following a transelliptical distribution
(Section 4.2). We also illustrate an application of this technical result to establishing a bound on
b ´ Σ (Section 4.3).
Σ
4.1
Sketch of proof for main result
r ab ,
The proof of Theorem 3.5 has two key steps. First, in Step 1, we prove that the distribution of Θ
the oracle estimator of Θab , is asymptotically normal, with
?
n¨
r ab ´ Θab
Θ
Ñ N p0, 1q
Sab detpΘq
e ab . (Explicit form of Sab is given in the proof of Theorem 4.1.)
where Sab is the asymptotic variance of Ω
q ´ Θ,
r
Next, in Step 2, we prove that the difference between the estimator and the oracle estimator, Θ
q
converges to zero at a fast rate, and that the variance estimator Sab converges to Sab at a fast rate.
q ab is an asymptotically normal estimator of Ωab . The detailed
Combining these steps we prove that Ω
proofs for each step are found in Appendix D. Here, we outline the main results for each step.
Step 1 establishes the Berry-Esseen type bound for the centered and normalized oracle estimator
?
r ´Θab
r ab by a linear function of the Kendall’s tau
n ¨ SΘabab¨detpΘq
. We approximate the oracle estimator Θ
statistic Tp, which is a U-statistic of the data. We prove that the variance of the linear approximation
is bounded away from zero and apply existing results on convergence of U-statistics. The following
result is proved in Appendix D
Theorem 4.1. Suppose that Assumptions 3.1, 3.2, and 3.4 hold. Then there exist constants Cnormal , Cvariance
depending on Ccov , Csparse , Ckernel but not on pn, pn , kn q, such that
ˇ
ˇ #
+
ˇ
ˇ
r ab ´ Θab
?
Θ
kn logppn q
1
ˇ
ˇ
?
n¨
ď t ´ Φptqˇ ď Cnormal ¨
`
,
sup ˇP
ˇ
ˇ
S
¨
detpΘq
n
2p
ab
n
tPR
where Sab is defined in the proof and satisfies Sab ¨ detpΘq ě Cvariance ą 0.
Step 2 contains the main challenge of this problem, since it requires strong results on the concentrap of the covariance matrix Σ. The main ingredient for
tion properties of the Kendall’s tau estimator Σ
this step is a new result on “sign-subgaussianity”, that is, proving that the signs vector signpXi ´ Xi1 q
is subgaussian for i.i.d. observations Xi , Xi1 . Our results on sign-subgaussianity are discussed in Secb around Σ is given in Section 4.3. Using these
tion 4.2 and their application to concentration of Σ
tools, we are able to prove the following theorem (proved in Appendix D):
Theorem 4.2. Suppose that Assumptions 3.1, 3.2, and 3.3 hold. Then there exists a constant Coracle ,
depending on Ccov , Csparse , Cest but not on pn, pn , kn q, such that, if 4 n ě 15kn logppn q, then, with
probability at least 1 ´ 2p1n ´ δn ,
q ´ Θ||
r 8 ď Coracle ¨ kn logppn q
||Θ
n
4 Note that the additional condition n ě 15k logpp q can be assumed to hold in our main result Theorem 3.5, since
n
n
if this inequality does not hold, then the claim in Theorem 3.5 is trivial.
13
and
4.2
c
ˇ
ˇ
kn2 logppn q
ˇ
ˇq
q
.
ˇSab ¨ detpΘq ´ Sab ¨ detpΘqˇ ď Coracle ¨
n
Sign-subgaussian random vectors
Recall the definition of a subgaussian random vector:
Definition
A random vector X P Rp is C-subgaussian if, for any fixed vector v P Rp , it holds
” J 4.3.
ı
2
that E ev X ď eC¨||v||2 {2 .
For graphical models
ř where the data points Xi come from
ř a subgaussian distribution, the sample
covariance matrix n1 i pXi ´ XqpXi ´ XqJ , with X “ n1 i Xi , is known to concentrate near the
population covariance, as measured by different norms. For example, elementwise convergence of the
sample covariance to the population covariance, that is, convergence in || ¨ ||8 , is sufficient to establish
rates of convergence for the graphical Lasso, CLIME or graphical Dantzig selector for estimating the
sparse inverse covariance [Ravikumar et al., 2011, Cai et al., 2011,aYuan, 2010]. Similar results can
b ´ Σ||8 ď
beaobtained also for the transelliptical family, since ||Tb ´ T ||8 ď C logppq{n and hence ||Σ
C logppq{n, as was shown in Liu et al. [2012a] and Liu et al. [2012b]. However, in order to construct
asymptotically normal estimators for the elements of the precision matrix, stronger results are needed
about the convergence of the sample covariance to the population covariance [Ren et al., 2013]. In
particular, a result on convergence in spectral norm, uniformly over all sparse submatrices, is required.
One can relate the convergence in the elementwise `8 norm to (sparse) spectral norm convergence,
however, this would lead to suboptimal sample size. One way to obtain a tight bound on the (sparse)
spectral norm convergence is by utilizing subgaussianity of the data points Xi . This is exactly what
we proceed to establish.
Recall from (2.3) the Kendall’s tau estimator of the covariance,
´
¯
ÿ
p “ sin π Tp where Tp “ `1˘
Σ
signpXi ´ Xi1 q signpXi ´ Xi1 qJ .
n
2
2 iăi1
Therefore, it is crucial to determine whether the vector signpXi ´ Xi1 q is itself subgaussian, at a
scale that does not depend heavily on the ambient dimension pn .5 Using past results on elliptical
distributions, we can reduce to a simpler case using the arguments of Lindskog et al. [2003] (proved
in Appendix F):
iid
Lemma 4.4. Let X, X 1 „ TEpΣ, ξ; f1 , . . . , fp q. Suppose that Σ is positive definite, and that ξ ą 0
with probability 1. Then signpX ´ X 1 q is equal in distribution to signpZq, where Z „ N p0, Σq.
Previous work has shown that a Gaussian random vector Z „ N p0, Σq is “sign-subgaussian”, that is,
signpZq is subgaussian with variance proxy that depends on pn only through CpΣq, for special cases
when the covariance Σ is identity or equicorrelation matrix [Han and Liu, 2013]. However, a result
for general covariance structures was previously unknown.
In the following lemma, we resolve this question, proving that Gaussian vectors are sign-subgaussian
(recall CpΣq is the condition number of Σ):
Lemma 4.5. Let Z „ N pµ, Σq. Then signpZq ´ E rsignpZqs is CpΣq-subgaussian.
This lemma is the primary tool for our main results in this paper—specifically, it is the key ingredient
q ´Θ
r and Sqab ¨ detpΘq
q ´ Sab ¨ detpΘq. Lemma 4.5
to the proof of Theorem 4.2, which bounds the errors Θ
is proved in Appendix C. We also use this result in establishing results in the following section.
5 Note that v J signpX ´X 1 q is obviously subgaussian for any distribution on X, as it is a sum of subgaussian random
i
i
variables (since signp¨q is bounded), however, its scale could grow linearly with pn .
14
4.3
p ´Σ
Deterministic and probabilistic bounds on Σ
p ´ Σ. Results given in this section
Lemma 4.5 is instrumental in obtaining probabilistic bounds on Σ
are crucial for establishing Theorem 4.2 and Corollary 3.8.
Let Sk be the set of k-sparse vectors in the unit ball,
Sk “ tu P Rp : ||u||2 ď 1, ||u||0 ď ku ,
and abusing notation, let || ¨ ||Sk denote the sparse spectral norm for matrices, that is, ||M ||Sk “
maxu,vPSk uJ M v.
The following lemma provides a bound on the error in Kendall’s tau, that is, on Tp ´ T , in this
sparse sectral norm (with the proof given in Appendix F).
Lemma 4.6. Suppose that k ě 1 and δ P p0, 1q satisfy logp2{δq ` 2k logp12pq ď n. Then with
probability at least 1 ´ δ it holds that
c
?
logp2{δq ` 2k logp12pq
p
.
||T ´ T ||Sk ď 32p1 ` 5qCpΣq ¨
n
b to Tb, with the following deterministic bound on the sparse spectral norm of the
Next, we relate Σ
p which is proven in Appendix F:
error of the covariance estimator Σ,
Lemma 4.7. The following bound holds deterministically: for any k ě 1,
π2
¨ k||Tp ´ T ||28 ` 2π||Tp ´ T ||Sk .
(4.1)
8
A result in de la Pena and Giné [1999, Theorem 4.1.8] bounds ||Tp ´ T ||8 with high probability
(details of this bound are given in Appendix D). Combining the bound on ||Tp ´ T ||8 with Lemmas 4.6
and 4.7, we immediately obtain the following corollary:
p ´ Σ||S ď
||Σ
k
Corollary 4.8. Take any δ1 , δ2 P p0, 1q and any k ě 1 such that logp2{δ2 q ` 2k logp12pq ď n. Then,
p ´ Σ holds:
with probability at least 1 ´ δ1 ´ δ2 , the following bound on Σ
p ´ Σ||S ď
||Σ
k
c
` ` ˘ ˘
?
4 log 2 p2 {δ1
π2
logp2{δ2 q ` 2k logp12pq
¨k¨
` 2π ¨ 32p1 ` 5qCpΣq ¨
. (4.2)
8
n
n
Finally, we use a result based on the work of Sun and Zhang [2012a], in order to extend this sparse
spectral norm bound to a bound holding for all approximately sparse vectors u and v:
Lemma 4.9 (Based on Proposition 5 of Sun and Zhang [2012a]). For any fixed matrix M P Rpˆp
and vectors u, v P Rp , and any k ě 1,
´
? ¯ ´
? ¯
|uJ M v| ď ||u||2 ` ||u||1 { k ¨ ||v||2 ` ||v||1 { k ¨ ||M ||Sk .
Results of Lemma 4.6 and Corollary 4.8 can be compared to Theorem 2 in Mitra and Zhang
[2014], which proves essentially the same result for the Kendall’s tau estimate of Σ, but only for
the nonparanormal (Gaussian copula) model; their technique does not extend immediately to the
transelliptical model. When CpΣq “ Op1q, we extend their result to the transelliptical model and, as
a special case, this provides an alternative proof for their result on the Gaussian copula model. We
note that their result does not depend on the condition number of the covariance matrix, but only on
the maximum eigenvalue. However, in the context of graphical models it is commonly assumed that
the smallest eigenvalue is a constant. Furthermore, our results in Lemma 4.6 and Corollary 4.8 can
also be compared with Theorem 4.10 of Han and Liu [2013], which give similar bounds on the spectral
p ´ Σ, but with a sign-subgaussianity assumption on the
norm of sparse submatrices of Tp ´ T and Σ
distribution. We rigorously establish the same bounds for all well-conditioned covariance matrices,
without explicitly making the sign-subgaussian assumption.
15
5
Simulation studies
In this section, we illustrate finite sample properties of ROCKET described in Section 2 on simulated
data. (A real data experiment, and some additional simulations, are presented in Appendix A.)
We use ROCKET to construct confidence intervals for edge parameters and report empirical
coverage probabilities as well as the length of constructed intervals. For comparison, we also construct
confidence intervals using the procedure of Ren et al. [2013], which is based on the Pearson correlation
matrix, a nonparanormal estimator of the correlation matrix (NPN) proposed in Liu et al. [2009], and
the pseudo score procedure of Gu et al. [2015]. For the first two methods, we use the plugin estimate
of the correlation matrix together with (2.8) to estimate Ωab . Recall that Liu et al. [2009] estimate
the correlation matrix based on the marginal transformation of the observed data. Let
$
’
if Fba pxq ă δn
& δn
Fea pxq “
Fba pxq if δn ď Fba pxq ď 1 ´ δn
’
% 1´δ
if Fba pxq ą 1 ´ δn ,
n
¯´1
´
a
ř
.
where Fba pxq “ n´1 i 1I tXia ă xu is the empirical CDF of Xa and δn “ 4n1{4 π logpnq
¯¯
¯ ´
´
¯
´ ´
z Φ Fea pXia q , Φ Feb pXib q .
b ab
b “ Σ
b ab “ Corr
The correlation matrix Σ
is then estimated as Σ
ab
Asymptotic variance of estimators
of Ωab¯ based on the Pearson or nonparanormal correlation matrix
´
2
´1 q
q
q
q 2 . Gu et al. [2015] estimate Ωab as
is obtained as Sab “ n
Ωaa Ωbb ` Ω
ab
´` ˘
` ˘ ¯ `
˘
bΣ
b
bΩ
b
b ab Ω
bΣ
bΩ
b
` Σ
Ω
´ Ω
ab
ab
ab
q PS “
Ω
,
` ˘
` ˘
ab
bΩ
b
bΣ
b
´1
Ω
` Σ
ab
ab
b is the Kendall’s tau estimator of the covariance matrix in (2.3) and Ω
b is an initial estimator
where Σ
PS
b
of the precision matrix. Under suitable conditions, Ωab is asymptotically normal with the asymptotic
variance that can be consistently estimated as in Corollary 4.12 of Gu et al. [2015]. Gu et al. [2015]
b however, we find that empirically
suggest using the CLIME estimator [Cai et al., 2011] to construct Ω,
the method performs better using lasso-with-refitting to estimate each row
a of Ω, similar to Sun and
Zhang [2012a]. For all simulations, we set the tuning parameter λ “ 2.1 logppn q{n, as suggested by
our theory—this constant is large enough so that the penalty dominates the variance of each element
of the score. All computations are carried out in Matlab.
Simulation 1. We generate data from the model X „ Ep0, Σ, ξq, where ξ follows a t-distribution
with 5 degrees of freedom. The inverse covariance matrix Ω encodes a grid where each node is
connected to its four nearest neighbors with the nonzero elements of Ω0 equal to ω “ 0.24. Diagonal
` ˘´1
` ˘
` ˘
element of Ω0 are equal to 1. Let Ω0
“ Σ0 . Then set Σ “ pdiag Σ0 q´1{2 Σ0 pdiag Σ0 q´1{2
and Ω “ Σ´1 . (Additional simulations in Appendix B show the same experiment on a chain graph
structure.)
We take a grid of size 30 ˆ 30 (so that pn “ 900) and take sample size n “ 400. Figure 2 shows
?
quantile-quantile (Q-Q) plots based on 1000 independent realizations of the test statistic error, n ¨
q
Ωab ´Ωab
, for the four methods together with the reference line showing quantiles of the standard normal
qab
S
? q
distribution. From this figure, we observe that the quantiles of the test statistic error n ΩabSq´Ωab based
ab
on ROCKET is closest to the quantiles of the standard normal random variable. We further quantify
these results in Table 1, which reports empirical coverage and width of the confidence intervals based
? q
on n ΩabSq´Ωab . From the table, we can observe that the coverage of the confidence intervals based on
ab
ROCKET and the pseudo score are closest to nominal coverage of 95%. The three node pairs displayed
in this figure and table, namely ωp2,2q,p2,3q , ωp2,2q,p3,3q , ωp2,2q,p10,10q , correspond to a true edge, a nonedge between nearby nodes that is therefore easy to mistake for an edge, and a non-edge between
distant nodes, respectively.
16
Quantiles of Ť(2,2),(2,3)
ROCKET
Pearson
4
4
2
2
4
1
0
0
2
0
-2
-1
0
-2
-4
-2
-3
-4
-4
-2
0
2
4
Standard Normal Quantiles
Quantiles of Ť(2,2),(3,3)
Pseudo Score
5
6
-4
-5
-4
-2
0
2
4
-2
0
2
4
Standard Normal Quantiles
6
3
-6
-4
Standard Normal Quantiles
4
4
-4
-2
0
2
4
Standard Normal Quantiles
4
6
2
4
0
2
2
2
1
0
0
-1
-2
0
-4
-2
-2
-2
-4
-3
-4
-6
-4
-2
0
2
4
Standard Normal Quantiles
Quantiles of Ť(2,2),(10,10)
Nonparanormal
8
3
-6
-4
-2
0
2
4
4
-2
0
2
4
Standard Normal Quantiles
6
3
-4
-4
Standard Normal Quantiles
-4
-2
0
2
4
Standard Normal Quantiles
5
8
4
6
2
2
1
0
4
0
-1
0
2
-2
0
-4
-2
-2
-3
-4
-6
-4
-2
0
2
Standard Normal Quantiles
4
-5
-4
-2
0
2
4
-4
-4
Standard Normal Quantiles
-2
0
2
4
Standard Normal Quantiles
-4
-2
0
2
4
Standard Normal Quantiles
?
q
Figure 2: Simulation 1 (transelliptical data). Q-Q plot of n ¨ ΩabSq´Ωab when Ω corresponds to a
ab
grid graph structure. Row 1 corresponds to an edge, row 2 to a close non-edge, and row 3 to a far
non-edge.
ωp2,2q,p2,3q “ 0.37
ωp2,2q,p3,3q “ 0
ωp2,2q,p10,10q “ 0
ROCKET
Pearson
NPN
Pseudo Score
94.6 (0.51)
94.3 (0.53)
94.9 (0.56)
36.6 (0.88)
81.0 (0.86)
78.3 (0.88)
82.4 (0.48)
88.3 (0.47)
89.1 (0.48)
92.2 (0.52)
94.8 (0.50)
95.5 (0.53)
Table 1: Simulation 1 (transelliptical data). Percent empirical coverage (average length) of 95%
confidence intervals based on 1000 independent simulation runs.
These results are not surprising, since neither the Pearson nor the nonparanormal correlation
matrix consistently estimate the true Σ. In contrast, both ROCKET and the pseudo score method
?
q ab
are able to construct a test statistic n ¨ Ω
qab that is asymptotically distributed as a normal random
S
variable. The asymptotic distribution provides a good approximation to the finite sample distribution
?
q
of n ¨ ΩabSq´Ωab .
ab
Simulation 2. We illustrate performance of ROCKET when data are generated from a normal
and nonparanormal distribution. We consider Ω corresponding to a grid as in Simulation 1, and
generate n “ 400 samples from N p0, Ω´1 q and NPNpΩ´1 ; fe1 , . . . , fep q, where fej “ fmodpj´1,5q`1 with
a
f1 pxq “ x, f2 pxq “ signpxq |x|, f3 pxq “ x3 , f4 pxq “ Φpxq, f5 pxq “ exppxq.
Table 2 summarizes results from the simulation. We observe that when data are multivariate
normal all methods perform well, with ROCKET and the pseudo score having slightly wider intervals,
but with similar coverage. When data are generated from a nonparanormal distribution, using the
Pearson correlation in (2.8) results in confidence intervals that do not have nominal coverage due to
the bias. In this setting, nonparanormal estimator, ROCKET and the pseudo score still have the
correct nominal coverage. Note however that when Kendall’s tau is equal to zero, Pearson correlation
17
Transf. Gaussian
Gaussian
ROCKET
Pearson
NPN
Pseudo Score
ωp2,2q,p2,3q “ 0.37
ωp2,2q,p3,3q “ 0
ωp2,2q,p10,10q “ 0
95.6 (0.33)
95.4 (0.35)
96.0 (0.36)
94.1 (0.32)
95.9 (0.34)
95.3 (0.35)
94.4 (0.32)
95.3 (0.34)
95.6 (0.35)
95.9 (0.33)
95.0 (0.35)
94.8 (0.36)
ωp2,2q,p2,3q “ 0.37
ωp2,2q,p3,3q “ 0
ωp2,2q,p10,10q “ 0
95.1 (0.35)
95.3 (0.35)
93.7 (0.36)
12.0 (0.33)
93.1 (0.32)
93.2 (0.32)
94.9 (0.33)
94.9 (0.34)
94.2 (0.35)
95.3 (0.35)
94.7 (0.34)
94.8 (0.39)
Table 2: Simulation 2 (Gaussian and nonparanormal data). Percent empirical coverage (average
length) of 95% confidence intervals based on 1000 independent simulation runs.
No Transform
Gaussian
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
T5
ROCKET
Pearson
Nonparanormal
Pseudo Score
0
0.1
0.2
0.3
0.4
0.5
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
ρ
Marginal Transform
T1
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.3
0.4
0.5
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.1
0.2
0.1
0.2
0.3
0.4
0.5
0.3
0.4
0.5
0.3
0.4
0.5
ρ
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0
ρ
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
ρ
0.3
ρ
0.4
0.5
0
0.1
0.2
ρ
Figure 3: Simulation 3. Power plots for simulated data generated from a Gaussian distribution, and
from a multivariate t distribution with 5 d.f. or with 1 d.f.
is also equal to zero, and coverage for Pearson improves. See, for example, coverage for ωp2,2q,p3,3q and
ωp2,2q,p10,10q .
? q ab
Simulation 3. In this simulation, we illustrate the power of a test based on the statistic n ¨ Ω
qab
S
to reject the null hypothesis H0,ab : Ωab “ 0. Samples are generated from the Ep0, Σ, ξq with ξ having
χpn , t5 , and t1 distribution and the covariance matrix is of the form Σ “ IP ` E where E12 “ E21 “ ρ
and all other entries zero, with pn “ 1000 and n “ 400. Note that ξ „ χpn implies that X is
multivariate normal. We also consider marginal transformation of X as described in Simulation 2.
Figure 3 plots empirical power curves based on 1000 independent simulation runs for different settings.
When the data follow a normal distribution all methods have similar power. For other distributions,
tests based on Pearson and nonparanormal correlation do not have correct coverage and are shown
for illustrative purpose only.
18
6
Discussion
We have proposed a novel procedure ROCKET for inference on elements of the latent inverse correlation matrix under high-dimensional elliptical copula models. Our paper has established a surprising
result, which states that ROCKET produces an asymptotically normal estimator for an element of
the inverse correlation matrix in an elliptical copula model with the same sample complexity that is
required to obtain an asymptotically normal estimator for an element in the precision matrix under a
multivariate normal distribution. Furthermore, this sample complexity is optimal [Ren et al., 2013].
The result is surprising as the family of elliptical copula models is much larger than the family of multivariate normal distributions. For example, it contains distributions with heavy tail dependence as
discussed in Section 2. ROCKET achieves the optimal requirement on the sample size without knowledge of the marginal transformation. Our result is also of significant practical importance. Since
normal distribution is only a convenient mathematical approximation to data generating process, we
recommend using ROCKET whenever making inference about inverse correlation matrix, instead of
methods that heavily rely on normality. From simulation studies, even when data are generated from
a normal distribution, ROCKET does not lose power compared to procedures that were specifically
developed for inference under normality.
The main technical tool developed in the paper establishes that the sign of normal random vector, taken elementwise, is itself a sub-Gaussian random variable with the sub-Gaussian parameter
depending on the condition number of the covariance matrix Σ (but not on the dimension pn ). Based
on this result, we were able to establish a tight tail bound on the deviation of sparse eigenvalues of
the Kendall’s tau matrix Tb. This result is of independent interest and it would allow us to improve
a number of recent results on sparse principal component analysis, factor models and estimation of
structured covariance matrices [Mitra and Zhang, 2014, Han and Liu, 2013, Fan et al., 2014]. The
sharpest result on the nonparametric estimation of correlation matrices in spectral norm under a
Gaussian copula model was established in [Mitra and Zhang, 2014]. Our results establish a similar
result for the family of elliptical copula models and provide an alternative proof for the Gaussian
copula model.
Acknowledgments
This work is partially supported by an IBM Corporation Faculty Research Fund at the University of
Chicago Booth School of Business, and an Alfred P. Sloan Fellowship. This work was completed in
part with resources provided by the University of Chicago Research Computing Center.
References
Richard G. Baraniuk, Mark A. Davenport, , and Michael B. Wakin. A simple proof of the restricted isometry property for random matrices. Constructive Approximation, 28(3):253–263, Jan
2008. ISSN 1432-0940. doi: 10.1007/s00365-007-9003-x. URL http://dx.doi.org/10.1007/
s00365-007-9003-x.
Alexandre Belloni and Victor Chernozhukov. Least squares after model selection in high-dimensional
sparse models. Bernoulli‘, 19(2):521–547, May 2013. ISSN 1350-7265. doi: 10.3150/11-bej410. URL
http://dx.doi.org/10.3150/11-BEJ410.
Alexandre Belloni, Victor Chernozhukov, and Christian B. Hansen. Inference on treatment effects after
selection amongst high-dimensional controls. Rev. Econ. Stud., 81(2):608–650, Nov 2013a. ISSN
1467-937X. doi: 10.1093/restud/rdt044. URL http://dx.doi.org/10.1093/restud/rdt044.
Alexandre Belloni, Victor Chernozhukov, and Kengo Kato. Robust inference in high-dimensional
approximately sparse quantile regression models. arXiv preprint arXiv:1312.7186, December 2013b.
19
Alexandre Belloni, Victor Chernozhukov, and Kengo Kato. Uniform post selection inference for lad
regression models. arXiv preprint arXiv:1304.0282, 2013c. URL http://arxiv.org/abs/1304.
0282.
Alexandre Belloni, Victor Chernozhukov, and Ying Wei. Honest confidence regions for logistic regression with a large number of controls. arXiv preprint arXiv:1304.3969, 2013d. URL
http://arxiv.org/abs/1304.3969.
Peter Bühlmann and Sara A. van de Geer. Statistics for high-dimensional data. Springer Series in
Statistics. Springer, Heidelberg, 2011. ISBN 978-3-642-20191-2. doi: 10.1007/978-3-642-20192-9.
URL http://dx.doi.org/10.1007/978-3-642-20192-9. Methods, theory and applications.
T. Tony Cai, W. Liu, and X. Luo. A constrained `1 minimization approach to sparse precision matrix
estimation. J. Am. Stat. Assoc., 106(494):594–607, 2011.
Herman Callaert and Paul Janssen. The Berry-Esseen theorem for U -statistics. Ann. Stat., 6(2):
417–421, 1978. ISSN 0090-5364.
Emmanuel J. Candès and T. Tao. The dantzig selector: Statistical estimation when p is much larger
than n. Ann. Stat., 35(6):2313–2351, 2007. ISSN 0090-5364. doi: 10.1214/009053606000001523.
URL http://dx.doi.org/10.1214/009053606000001523.
Mengjie Chen, Zhao Ren, Hongyu Zhao, and Harrison H. Zhou. Asymptotically normal and efficient
estimation of covariate-adjusted gaussian graphical model. arXiv preprint arXiv:1309.5923, 2013.
URL http://arxiv.org/abs/1309.5923.
Jie Cheng, Elizaveta Levina, and Ji Zhu. High-dimensional mixed graphical models. ArXiv e-prints,
arXiv:1304.2810, April 2013.
Gabor Csardi and Tamas Nepusz. The igraph software package for complex network research. InterJournal, Complex Systems:1695, 2006. URL http://igraph.org.
Alexandre d’Aspremont, Onureena Banerjee, and Laurent El Ghaoui. First-order methods for sparse
covariance selection. SIAM J. Matrix Anal. Appl., 30(1):56–66, 2008. ISSN 0895-4798. doi: 10.
1137/060670985. URL http://dx.doi.org/10.1137/060670985.
Victor de la Pena and Evarist Giné. Decoupling: from dependence to independence. Springer, 1999.
Paul Embrechts, Filip Lindskog, and Alexander McNeil. Modelling dependence with copulas
and applications to risk management. In S. T. Rachev, editor, Handbook of heavy tailed
distributions in finance, pages 329–384. Elsevier, 2003.
URL http://books.google.com/
books?hl=en&lr=&id=sv8jGSVFra8C&oi=fnd&pg=PA329&dq=Modelling+
dependence+with+copulas+and+applications+to+risk+management&ots=Yvg5prGUXz&
amp;sig=b3Qs3l2OC8JgSQKPxUnY6Lq2Ql8.
Jianqing Fan, Y. Feng, and Y. Wu. Network exploration via the adaptive lasso and scad penalties.
Ann. Appl. Stat., 3(2):521–541, 2009. ISSN 1932-6157. doi: 10.1214/08-AOAS215. URL http:
//dx.doi.org/10.1214/08-AOAS215.
Jianqing Fan, Fang Han, and Han Liu. Page: Robust pattern guided estimation of large covariance
matrix. Technical report, Technical report, Princeton University, 2014.
Kai Tai Fang, Samuel Kotz, and Kai Wang Ng. Symmetric multivariate and related distributions,
volume 36 of Monographs on Statistics and Applied Probability. Chapman and Hall, Ltd., London,
1990. ISBN 0-412-31430-4. doi: 10.1007/978-1-4899-2937-2. URL http://dx.doi.org/10.1007/
978-1-4899-2937-2.
20
Max H. Farrell. Robust inference on average treatment effects with possibly more covariates than
observations. arXiv preprint arXiv:1309.4686, September 2013.
Jerome H. Friedman, Trevor J. Hastie, and Robert J. Tibshirani. Sparse inverse covariance estimation
with the graphical lasso. Biostatistics, 9(3):432–441, 2008.
Q. Gu, Y. Cao, Y. Ning, and Han Liu. Local and global inference for high dimensional gaussian copula
graphical models. ArXiv e-prints, arXiv:1502.02347, February 2015.
J. Guo, Elizaveta Levina, G. Michailidis, and J. Zhu. Joint estimation of multiple graphical models.
Biometrika, 98(1):1–15, 2011a.
Jian Guo, Elizaveta Levina, George Michailidis, and Ji Zhu. Asymptotic properties of the joint neighborhood selection method for estimating categorical markov networks. Technical report, University
of Michigan, 2011b.
Fang Han and Han Liu. Optimal rates of convergence for latent generalized correlation matrix estimation in transelliptical distribution. ArXiv e-prints, arXiv:1305.6916, May 2013.
Holger Höfling and Robert J. Tibshirani. Estimation of sparse binary pairwise markov networks using
pseudo-likelihoods. J. Mach. Learn. Res., 10:883–906, 2009. URL http://dl.acm.org/citation.
cfm?id=1577101.
Adel Javanmard and Andrea Montanari. Nearly optimal sample size in hypothesis testing for highdimensional regression. arXiv preprint arXiv:1311.0274, November 2013.
Adel Javanmard and Andrea Montanari. Confidence intervals and hypothesis testing for highdimensional regression. J. Mach. Learn. Res., 15(Oct):2869–2909, 2014. URL http://jmlr.org/
papers/v15/javanmard14a.html.
Claudia Klüppelberg, Gabriel Kuhn, and Liang Peng. Semi-parametric models for the multivariate
tail dependence function–the asymptotically dependent case. Scand. J. Stat., 35(4):701–718, 2008.
URL http://onlinelibrary.wiley.com/doi/10.1111/j.1467-9469.2008.00602.x/full.
C. Lam and Jianqing Fan. Sparsistency and rates of convergence in large covariance matrix estimation.
Ann. Stat., 37:4254–4278, 2009.
S. L. Lauritzen. Graphical Models, volume 17 of Oxford Statistical Science Series. The Clarendon
Press Oxford University Press, New York, 1996. ISBN 0-19-852219-3. Oxford Science Publications.
Jason D. Lee and Trevor J. Hastie. Learning mixed graphical models. ArXiv e-prints, arXiv:1205.5012,
May 2012.
Jason D. Lee, Dennis L. Sun, Yuekai Sun, and Jonathan E. Taylor. Exact post-selection inference
with the lasso. ArXiv e-prints, arXiv:1311.6238, November 2013.
Filip Lindskog, Alexander McNeil, and Uwe Schmock. Kendall’s tau for elliptical distributions. Credit
Risk, pages 149–156, 2003. ISSN 1431-1933. doi: 10.1007/978-3-642-59365-9 8. URL http://dx.
doi.org/10.1007/978-3-642-59365-9_8.
Han Liu and Lie Wang. Tiger: A tuning-insensitive approach for optimally estimating gaussian
graphical models. ArXiv e-prints, arXiv:1209.2437, September 2012.
Han Liu, John D. Lafferty, and Larry A. Wasserman. The nonparanormal: Semiparametric estimation
of high dimensional undirected graphs. J. Mach. Learn. Res., 10:2295–2328, 2009.
21
Han Liu, Fang Han, Ming Yuan, John D. Lafferty, and Larry A. Wasserman. High-dimensional
semiparametric Gaussian copula graphical models. Ann. Stat., 40(4):2293–2326, 2012a. ISSN 00905364. doi: 10.1214/12-AOS1037. URL http://dx.doi.org/10.1214/12-AOS1037.
Han Liu, Fang Han, and Cun-Hui Zhang. Transelliptical graphical models. In Proc. of NIPS, pages
809–817. 2012b. URL http://books.nips.cc/papers/files/nips25/NIPS2012_0380.pdf.
Weidong Liu. Gaussian graphical model estimation with false discovery rate control. Ann. Stat., 41
(6):2948–2978, 2013. ISSN 0090-5364. doi: 10.1214/13-AOS1169. URL http://dx.doi.org/10.
1214/13-AOS1169.
Richard Lockhart, Jonathan E. Taylor, Robert J. Tibshirani, and Robert J. Tibshirani. A significance
test for the lasso. Ann. Stat., 42(2):413–468, 2014. ISSN 0090-5364. doi: 10.1214/13-AOS1175.
URL http://dx.doi.org/10.1214/13-AOS1175.
Po-Ling Loh and Martin J. Wainwright. Regularized m-estimators with nonconvexity: Statistical and
algorithmic theory for local optima. arXiv preprint arXiv:1305.2436, 2013. URL http://arxiv.
org/abs/1305.2436.
Pascal Massart. Concentration inequalities and model selection, volume 1896 of Lecture Notes in
Mathematics. Springer, Berlin, 2007. ISBN 978-3-540-48497-4; 3-540-48497-3. Lectures from the
33rd Summer School on Probability Theory held in Saint-Flour, July 6–23, 2003, With a foreword
by Jean Picard.
Nicolas Meinshausen and Peter Bühlmann. High dimensional graphs and variable selection with the
lasso. Ann. Stat., 34(3):1436–1462, 2006.
Ritwik Mitra and Cun-Hui Zhang. Multivariate analysis of nonparametric estimates of large correlation
matrices. ArXiv e-prints, arXiv:1403.6195, March 2014.
Thomas Peel, Sandrine Anthoine, and Liva Ralaivola. Empirical bernstein inequalities for u-statistics.
In J.D. Lafferty, C.K.I. Williams, J. Shawe-Taylor, R.S. Zemel, and A. Culotta, editors, Adv. Neural
Inf. Process. Syst. 23, pages 1903–1911. Curran Associates, Inc., 2010. URL http://papers.nips.
cc/paper/4081-empirical-bernstein-inequalities-for-u-statistics.pdf.
R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical
Computing, Vienna, Austria, 2012. URL http://www.R-project.org/. ISBN 3-900051-07-0.
P. Ravikumar, Martin J. Wainwright, G. Raskutti, and B. Yu. High-dimensional covariance estimation
by minimizing `1 -penalized log-determinant divergence. Electron. J. Stat., 5:935–980, 2011.
Pradeep Ravikumar, Martin J. Wainwright, and J. D. Lafferty. High-dimensional ising model selection
using `1 -regularized logistic regression. Ann. Stat., 38(3):1287–1319, 2010. ISSN 0090-5364. doi:
10.1214/09-AOS691. URL http://dx.doi.org/10.1214/09-AOS691.
Zhao Ren, Tingni Sun, Cun-Hui Zhang, and Harrison H. Zhou. Asymptotic normality and optimalities
in estimation of large gaussian graphical model. arXiv preprint arXiv:1309.6024, 2013. URL
http://arxiv.org/abs/1309.6024.
Adam J. Rothman, Peter J. Bickel, Elizaveta Levina, and J. Zhu. Sparse permutation invariant
covariance estimation. Electron. J. Stat., 2:494–515, 2008. ISSN 1935-7524. doi: 10.1214/08-EJS176.
URL http://dx.doi.org/10.1214/08-EJS176.
Nathan Srebro and Adi Shraibman. Rank, trace-norm and max-norm. In Learning theory, volume 3559
of Lecture Notes in Comput. Sci., pages 545–560. Springer, Berlin, 2005. doi: 10.1007/11503415 37.
URL http://dx.doi.org/10.1007/11503415_37.
22
Tingni Sun and Cun-Hui Zhang. Sparse matrix inversion with scaled lasso. February 2012a.
Tingni Sun and Cun-Hui Zhang. Comment: “minimax estimation of large covariance matrices under
`1 -norm”. Statist. Sinica, 22:1354–1358, 2012b.
Jonathan E. Taylor, Richard Lockhart, Robert J. Tibshirani, and Robert J. Tibshirani. Post-selection
adaptive inference for least angle regression and the lasso. arXiv preprint arXiv:1401.3889, January
2014.
Sara A. van de Geer, Peter Bühlmann, Ya’acov Ritov, and Ruben Dezeure. On asymptotically optimal
confidence regions and tests for high-dimensional models. Ann. Stat., 42(3):1166–1202, Jun 2014.
ISSN 0090-5364. doi: 10.1214/14-aos1221. URL http://dx.doi.org/10.1214/14-AOS1221.
Roman Vershynin. Introduction to the non-asymptotic analysis of random matrices. In Y. C. Eldar
and G. Kutyniok, editors, Compressed Sensing: Theory and Applications. Cambridge University
Press, 2012.
Marten Wegkamp and Yue Zhao. Adaptive estimation of the copula correlation matrix for semiparametric elliptical copulas. ArXiv e-prints, arXiv:1305.6526, May 2013.
Lingzhou Xue and Hui Zou. Regularized rank-based estimation of high-dimensional nonparanormal
graphical models. Ann. Stat., 40(5):2541–2571, 2012. ISSN 0090-5364. doi: 10.1214/12-AOS1041.
URL http://dx.doi.org/10.1214/12-AOS1041.
Lingzhou Xue, Hui Zou, and Tianxi Ca. Nonconcave penalized composite conditional likelihood
estimation of sparse ising models. Ann. Stat., 40(3):1403–1429, 2012. URL http://projecteuclid.
org/euclid.aos/1344610588.
Eunho Yang, Genevera I. Allen, Zhandong Liu, and Pradeep Ravikumar. Graphical models via generalized linear models. In Advances in Neural Information Processing Systems
25, pages 1358–1366. Curran Associates, Inc., 2012. URL http://papers.nips.cc/paper/
4617-graphical-models-via-generalized-linear-models.pdf.
Eunho Yang, Pradeep Ravikumar, Genevera I. Allen, and Zhandong Liu. On graphical models via
univariate exponential family distributions. ArXiv e-prints, arXiv:1301.4183, January 2013.
Eunho Yang, Yulia Baker, Pradeep Ravikumar, Genevera I. Allen, and Zhandong Liu. Mixed graphical
models via exponential families. In Proc. 17th Int. Conf, Artif. Intel. Stat., pages 1042–1050, 2014.
M. Yuan. High dimensional inverse covariance matrix estimation via linear programming. J. Mach.
Learn. Res., 11:2261–2286, 2010.
M. Yuan and Y. Lin. Model selection and estimation in the gaussian graphical model. Biometrika,
94(1):19–35, 2007.
Cun-Hui Zhang and Stephanie S. Zhang. Confidence intervals for low dimensional parameters in high
dimensional linear models. J. R. Stat. Soc. B, 76(1):217–242, Jul 2013. ISSN 1369-7412. doi:
10.1111/rssb.12026. URL http://dx.doi.org/10.1111/rssb.12026.
Tuo Zhao and Han Liu. Calibrated precision matrix estimation for high dimensional elliptical distributions. IEEE Trans. Inf. Theory, pages 1–1, 2014. ISSN 1557-9654. doi: 10.1109/tit.2014.2360980.
URL http://dx.doi.org/10.1109/TIT.2014.2360980.
Tuo Zhao, Han Liu, Kathryn Roeder, John D. Lafferty, and Larry A. Wasserman. huge: Highdimensional Undirected Graph Estimation, 2014. URL http://CRAN.R-project.org/package=
huge. R package version 1.2.6.
23
A
Real data experiment
In this section, we evaluate the performance of the ROCKET method on a real data set, and compare
with the Gaussian graphical model based approach of Ren et al. [2013] (using Pearson correlation)
and the nonparanormal estimator proposed in Liu et al. [2009] (details for these methods are given in
Section 5).
We use stock price closing data obtained via the R package huge [Zhao et al., 2014], which was
gathered from publicly available data from Yahoo Finance.6 The data consists of daily closing prices
of 452 S&P 500 companies over 1258 days. We transform the data to consider the log-returns,
˙
ˆ
Closing price of stock j on day i ` 1
.
Xij “ log
Closing price of stock j on day i
While in practice there is dependence across time in this data set, we treat each row of X as independent.
We perform two experiments on this data set. In Experiment 1, we test whether empirical results
agree with the asymptotic normality predicted by the theory for the three methods—we do this
by splittting the data into disjoint subsamples and comparing estimates across these subsamples. In
Experiment 2, we use the full sample size and compare the estimates and confidence intervals produced
by each of the three methods.
A.1
Experiment 1: checking asymptotic normality
In this real data example, there is no available “ground truth” to compare to—that is, we do not know
the true distribution of the data, and cannot compare our estimates to an exact true precision matrix
Ω. However, we can still check whether the estimators produced by these methods exhibit asymptotic
normality (as claimed in the theory), by splitting the data into many subsamples and considering the
empirical distribution of the estimators across these subsamples.
To construct our subsampled data, we randomly select L “ 25 disjoint sets of size n “ 50 from
t1, . . . , 1257u, denoted as I1 , . . . , IL . Due to this small sample size, we restrict our attention to
companies in the categories Materials and Consumer Staples, which consist of 29 and 35 companies,
respectively, for a total of pn “ 64 companies. For each ` “ 1, . . . , L, define the `th data set X p`q “
XI` ,S P Rnˆp , where S Ă t1, . . . , 452u identifies the p “ 64 stocks of interest.
q p`q and Sqp`q using
Next, for each pair pa, bq of stocks, and for each subsample `, we compute Ω
ab
ab
ROCKET. If the true distribution of the data follows the transelliptical model with precision matrix
?
q p`q ´ Ωab q{Sqp`q is approximately standard
Ω, then our main result, Theorem 3.5, implies that n ¨ pΩ
ab
ab
p`q
normal. Since Sqab concentrates near Sab (see Theorem 4.1), we should have
p`q
zab :“
?
n¨
q p`q
? Ωab
Ω
ab
« n¨
` N p0, 1q .
p`q
Sab
q
S
ab
p1q
pLq
In particular, this implies that the sample variance of the vector pzab , . . . , zab q should have expectation approximately 1.
` ˘
In Figure 4, we show a histogram of the sample variances SampleVarpzab q across all p2n “ 2016
pairs of variables. To compare to the Pearson and nonparanormal methods, we repeat this procedure
for the estimators (and estimated variances) produced by the other two methods as well, which are
also displayed in Figure 4. We see that ROCKET produces a mean sample variance « 0.98 (very
near to 1), while the other two methods give mean sample variances of « 1.28 (Pearson) and « 1.265
(nonparanormal), substantially higher than the theoretical value of 1. This indicates that the normal
approximation to the distribution of the estimator may be approximately valid for ROCKET, but
6 http://ichart.finance.yahoo.com
24
ROCKET
600
400
400
300
300
300
200
200
200
100
100
100
0
0.5
1
1.5
2
2.5
0
0
0.5
1
1.5
2
?
2.5
Mean=1.2651
500
400
0
Nonparanormal
600
Mean=1.2798
500
Mean=0.98105
500
Pearson
600
3
0
0
0.5
1
1.5
2
2.5
3
q ab {Sqab , for each pair of variables pa, bq,
Figure 4: Sample variances of the rescaled estimator, n ¨ Ω
using the subsampled stock data. The sample variances should be approximately 1 according to the
theory (see Section A.1).
does not have the correct scale (that is, the scale predicted by the theory) for the other two methods,
on this data set.
p1q
pLq
The vector pzab , . . . , zab q, in addition to having sample variance near 1, should also exhibit
Gaussian-like tails according to the theory. To check this, we calculate the proportion of values in
this vector lying near to the mean,
c
ˇ
1 )ˇˇ
1 ˇˇ! ˇˇ p`q
ˇ
ˇ ` : ˇzab ´ z ab ˇ ď 1.6449 1 ´
ˇ,
L
L
which should be approximately 90% according to the theory (using standard normal quantiles). The
results are:
ROCKET: 90.55% coverage
Pearson: 85.01% coverage
Nonparanormal: 85.18% coverage
We see that only the ROCKET method achieves the appropriate coverage.
A.2
Experiment 2: estimating a graph
In the second experiment, we use the full sample size n “ 1257 to estimate a sparse graph over the
pn “ 64 stocks selected for Experiment 1, using each of the three methods. To do this, for each method
we first produce a (approximate) p-value testing for the presence of an edge between each pair pa, bq
of variables. Recall that according to our main result, Theorem 3.5, if the pair of variables pa, bq does
? q q
not have an edge, then Ωab “ 0 and so n ¨ Ω
as a standardˇ¯normal
ab {Sab is approximately distributed
´ˇ?
ˇ
q ab {Sqab ˇˇ .
variable. Then, using a two-sided z-test, we calculate a p-value Pab “ 2 ´ 2Φ ˇ n ¨ Ω
In Figure 5, we show the resulting graphs when edge pa, bq is drawn whenever the p-value passes
the threshold Pab ă 0.00001 or whenever Pab ă 0.001. The number of edges selected for each
method is shown in the figures. Overall we see that ROCKET selects roughly the same number of
edges as the Pearson method but less than the nonparanormal method, on this data set. Since the
Pearson and nonparanormal methods do not exhibit approximately normal behavior across subsamples
(Experiment 1), this should not be interpreted as a power comparison between the methods; the
additional edges selected by the nonparanormal method, for instance, may not be as reliable since
the p-value calculation is based on approximating the distribution of the estimator using a theoretical
scaling that does not appear to hold for this method.
B
Additional simulations
Simulation 1 (chain graph). Here we repeat Simulation 1 from the paper with a chain graph
25
ROCKET (33 edges with p−value < 0.00001)
●
●
●
●
●
●
●
●
●
Pearson (33 edges with p−value < 0.00001)
Materials
Consumer Staples
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Nonparanormal (83 edges with p−value < 0.001)
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Pearson (61 edges with p−value < 0.001)
●
●
●
●
●
●
●
●
●
●
●
●
ROCKET (65 edges with p−value < 0.001)
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Nonparanormal (42 edges with p−value < 0.00001)
●
●
●
●
Figure 5: Estimated graph for the stock data, using the ROCKET, Pearson, and nonparanormal
methods (see Section A.2). An edge is displayed for each pair of variables pa, bq with p-value Pab ă
0.00001 (top row) and Pab ă 0.001 (bottom row). Graphs were drawn using the igraph package
[Csardi and Nepusz, 2006] in R [R Core Team, 2012].
ω10,11 “ 10.38
ω10,12 “ 0
ω10,20 “ 0
ROCKET
Pearson
NPN
Pseudo Score
94.5 (10.16)
96.3 (9.78)
95.9 (16.04)
56.1 (4.90)
62.2 (4.62)
62.3 (6.04)
31.8 (3.53)
74.0 (3.32)
81.9 (4.25)
93.5 (11.12)
93.8 (10.11)
95.8 (17.69)
Table 3: Simulation 1 (transelliptical data). Percent empirical coverage (average length) of 95%
confidence intervals based on 1000 independent simulation runs. Ω corresponds to a chain graph
structure.
structure instead of a grid graph structure. The inverse covariance matrix Ω now has a chain structure
with Ω0j,j`1 “ Ω0j`1,j “ 0.5. We set pn “ 1000 and take sample size n “ 400. Figures 6 and 7
show quantile-quantile (Q-Q) plots based on 1000 independent realizations of the test statistic error,
? Ωq ab ´Ωab
n ¨ Sq
, for the four methods together with the reference line showing quantiles of the standard
ab
normal distribution. The first row in the two figures illustrates actual performance of the methods,
while the second row illustrates performance of an oracle procedure that does not need to solve a
high-dimensional variable selection problem, but instead knows the sparsity pattern of Ω. From these
? q
two figures, we observe that the quantiles of the test statistic error n ΩabSq´Ωab based on ROCKET and
ab
the pseudo score estimators are closest to the quantiles of the standard normal random variable. We
further quantify these results in Table 3, which reports empirical coverage of the confidence intervals
? q
based on n ΩabSq´Ωab . From the table, we can observe that the coverage of the confidence intervals
ab
based on ROCKET and the pseudo score are closest to nominal coverage of 95%.
Simulation 4. In this additional simulation, we evaluate robustness of the procedures to three
26
ROCKET
Pearson
Estimated Neighborhood
4
Nonparanormal
10
Pseudo Score
10
4
3
3
5
5
2
2
0
1
1
0
0
-5
0
-5
-1
-1
-10
-2
-2
-10
-15
-3
-3
-4
-15
-4
-2
0
2
4
Standard Normal Quantiles
-20
-4
-2
0
2
4
Standard Normal Quantiles
4
-4
-4
-2
0
2
4
Standard Normal Quantiles
10
-4
-2
0
2
4
Standard Normal Quantiles
4
10
3
3
5
5
2
2
1
0
Oracle
1
0
0
0
-5
-1
-5
-1
-2
-10
-3
-2
-10
-4
-15
-3
-5
-4
-15
-4
-2
0
2
4
Standard Normal Quantiles
-20
-4
-2
0
2
4
-6
-4
Standard Normal Quantiles
-2
Estimated Neighborhood
ROCKET
Pearson
8
3
6
2
4
1
2
0
0
-1
-2
-2
-4
-3
-6
-4
-8
2
4
Standard Normal Quantiles
Figure 6: Simulation 1 (transelliptical data). Q-Q plot of
edge) when Ω corresponds to a chain graph structure.
4
0
?
n¨
-4
-2
0
2
4
Standard Normal Quantiles
q ab ´Ωab
Ω
qab
S
with a “ 10 and b “ 11 (true
Nonparanormal
Pseudo Score
6
4
3
4
2
2
1
0
0
-1
-2
-2
-4
-4
-2
0
2
4
Standard Normal Quantiles
-3
-6
-4
-2
0
2
4
4
10
3
8
-4
-4
Standard Normal Quantiles
-2
0
2
4
Standard Normal Quantiles
-4
-2
0
2
4
Standard Normal Quantiles
6
4
3
4
6
2
2
4
2
Oracle
1
1
2
0
0
0
0
-1
-1
-2
-2
-2
-2
-4
-4
-3
-6
-4
-8
-4
-2
0
2
Standard Normal Quantiles
4
-3
-6
-4
-2
0
2
4
-4
-4
Standard Normal Quantiles
-2
0
2
Standard Normal Quantiles
4
-4
-2
0
2
4
Standard Normal Quantiles
?
q
Figure 7: Simulation 1 (transelliptical data). Q-Q plot of n ¨ ΩabSq´Ωab with a “ 10 and b “ 12
ab
(non-edge close to a true edge) when Ω corresponds to a chain graph structure.
27
Gaussian
t5
r
ROCKET
Pearson
NPN
Pseudo Score
ωp2,2q,p2,3q “
0.37
0.01
0.05
0.1
94.9 (0.34)
94.1 (0.35)
95.1 (0.37)
93.6 (0.32)
93.8 (0.32)
92.5 (0.32)
93.4 (0.32)
94.4 (0.32)
92.2 (0.32)
96.0 (0.33)
95.4 (0.32)
96.7 (0.37)
ωp2,2q,p3,3q “ 0
0.01
0.05
0.1
95.3 (0.35)
93.7 (0.36)
96.2 (0.38)
95.3 (0.32)
92.6 (0.32)
92.9 (0.32)
95.4 (0.31)
92.7 (0.31)
93.4 (0.31)
95.1 (0.33)
95.0 (0.33)
94.5 (0.34)
ωp2,2q,p10,10q “ 0
0.01
0.05
0.1
95.8 (0.36)
94.7 (0.38)
94.7 (0.40)
95.9 (0.33)
93.8 (0.33)
93.5 (0.33)
96.1 (0.32)
94.0 (0.32)
93.0 (0.32)
95.5 (0.36)
95.3 (0.38)
94.6 (0.43)
ωp2,2q,p2,3q “
0.37
0.01
0.05
0.1
94.4 (0.38)
93.7 (0.39)
94.9 (0.40)
85.1 (0.35)
84.6 (0.36)
83.5 (0.36)
89.0 (0.32)
89.7 (0.32)
86.4 (0.31)
96.1 (0.39)
96.4 (0.39)
96.0 (0.44)
ωp2,2q,p3,3q “ 0
0.01
0.05
0.1
95.5 (0.39)
94.1 (0.41)
95.0 (0.42)
88.8 (0.35)
87.4 (0.35)
85.8 (0.36)
92.9 (0.31)
92.3 (0.31)
88.8 (0.31)
94.6 (0.34)
95.4 (0.37)
96.1 (0.38)
ωp2,2q,p10,10q “ 0
0.01
0.05
0.1
95.2 (0.41)
95.4 (0.42)
94.6 (0.44)
88.7 (0.36)
87.2 (0.36)
84.5 (0.37)
90.5 (0.32)
91.7 (0.32)
89.6 (0.32)
94.2 (0.41)
94.7 (0.42)
94.3 (0.45)
Table 4: Simulation 4 (Gaussian and transelliptical data; row corruption mechanism). Percent empirical coverage (average length) of 95% confidence intervals based on 1000 independent simulation
runs. Ω corresponds to a grid graph structure. r denotes the fraction of corrupted rows.
data contamination mechanisms: random row contamination, deterministic row contamination, and
element contamination mechanism. Let r P p0, 1q be the contamination level. As before we set n “ 400
and pn “ 900. The precision matrix Ω is chosen as the grid structure used in earlier simulations. For
both row contamination and element contamination, we consider two settings, where the (uncorrupted)
samples are drawn either from N p0, Ω´1 q, or from Ep0, Ω´1 , ξq with ξ „ t5 .
The row contamination mechanism corrupts tnru rows of the data matrix. In the random row
contamination mechanism, each corrupted row is filled with i.i.d. entries drawn from a t1.5 distribution.
These contaminated rows are heavy tailed and do not come from the same elliptical copula family
as the uncorrupted data. In the deterministic row contamination mechanism, each corrupted row is
equal to the vector p`5, ´5, `5, ´5, . . .q P Rp , where the numbers `5 and ´5 occur in an alternating
way. The element contamination mechanism selects tnpru elements and substitutes them with with a
value drawn from either N p3, 3q or N p´3, 3q with equal probability.
The results of the simulation are summarized in Tables 4, 5, and 6. The deterministic row corruption mechanism is the most malicious and hurts the non-robust procedures the most. From the
simulation, we observe that the procedures using the Kendall’s tau correlation matrix tend to be more
robust. Furthermore, coverage of non-zero elements tends to be more severely affected by corruption
than the coverage of zero-elements.
28
Gaussian
t5
r
ROCKET
Pearson
NPN
Pseudo Score
ωp2,2q,p2,3q “
0.37
0.01
0.05
0.1
94.4 (0.35)
80.6 (0.51)
53.8 (0.86)
90.2 (0.40)
40.1 (0.85)
23.6 (1.40)
86.7 (0.32)
85.4 (0.45)
67.1 (0.63)
93.6 (0.39)
75.7 (1.00)
51.6 (0.46)
ωp2,2q,p3,3q “ 0
0.01
0.05
0.1
94.6 (0.37)
93.6 (0.52)
93.1 (0.83)
93.7 (0.39)
92.8 (0.83)
88.2 (1.30)
84.7 (0.32)
93.6 (0.40)
92.6 (0.49)
93.5 (0.29)
95.2 (0.47)
96.0 (1.40)
ωp2,2q,p10,10q “ 0
0.01
0.05
0.1
94.4 (0.37)
93.4 (0.49)
92.0 (0.64)
94.8 (0.37)
83.4 (0.75)
79.4 (1.20)
88.2 (0.33)
97.8 (0.41)
87.3 (0.56)
93.1 (0.42)
92.9 (0.59)
91.3 (0.49)
ωp2,2q,p2,3q “
0.37
0.01
0.05
0.1
92.4 (0.39)
77.0 (0.61)
60.4 (1.00)
78.2 (0.43)
52.8 (0.74)
36.5 (1.10)
82.5 (0.32)
83.2 (0.47)
73.0 (0.65)
91.9 (0.47)
72.1 (1.20)
62.8 (0.75)
ωp2,2q,p3,3q “ 0
0.01
0.05
0.1
94.8 (0.41)
94.1 (0.63)
94.0 (1.00)
88.4 (0.43)
88.1 (0.73)
87.6 (1.10)
84.2 (0.32)
91.7 (0.42)
90.1 (0.51)
92.3 (0.35)
95.6 (0.63)
96.5 (2.30)
ωp2,2q,p10,10q “ 0
0.01
0.05
0.1
94.5 (0.42)
93.5 (0.59)
92.8 (0.76)
90.2 (0.42)
75.0 (0.66)
68.7 (0.97)
86.3 (0.33)
90.7 (0.42)
77.7 (0.58)
91.9 (0.51)
94.6 (0.62)
93.1 (0.57)
Table 5: Simulation 4 (Gaussian and transelliptical data; deterministic row corruption mechanism).
Percent empirical coverage (average length) of 95% confidence intervals based on 1000 independent
simulation runs. Ω corresponds to a grid graph structure. r denotes the fraction of corrupted rows.
29
Gaussian
t5
r
ROCKET
Pearson
NPN
Pseudo Score
ωp2,2q,p2,3q “
0.37
0.01
0.05
0.10
94.2 (0.32)
93.6 (0.29)
85.9 (0.27)
72.6 (0.28)
8.80 (0.23)
0.70 (0.21)
90.4 (0.30)
52.0 (0.27)
15.3 (0.24)
96.5 (0.35)
93.5 (0.40)
82.6 (0.52)
ωp2,2q,p3,3q “ 0
0.01
0.05
0.1
95.6 (0.34)
91.4 (0.30)
89.7 (0.28)
92.3 (0.28)
76.8 (0.23)
73.8 (0.20)
94.6 (0.30)
88.0 (0.26)
81.7 (0.24)
94.4 (0.32)
92.7 (0.33)
92.7 (0.34)
ωp2,2q,p10,10q “ 0
0.01
0.05
0.1
95.4 (0.35)
95.3 (0.31)
95.5 (0.28)
95.4 (0.29)
94.2 (0.23)
94.9 (0.20)
95.8 (0.31)
95.3 (0.27)
94.7 (0.24)
95.6 (0.35)
94.3 (0.32)
95.7 (0.30)
ωp2,2q,p2,3q “
0.37
0.01
0.05
0.1
94.4 (0.36)
93.1 (0.32)
94.8 (0.29)
80.9 (0.32)
21.3 (0.25)
4.0 (0.22)
85.2 (0.30)
51.4 (0.27)
20.9 (0.24)
95.8 (0.39)
92.5 (0.45)
83.8 (0.54)
ωp2,2q,p3,3q “ 0
0.01
0.05
0.1
94.2 (0.38)
93.1 (0.33)
94.4 (0.29)
87.3 (0.31)
79.6 (0.25)
75.7 (0.22)
89.9 (0.30)
86.3 (0.27)
85.0 (0.24)
95.0 (0.34)
94.2 (0.36)
93.7 (0.36)
ωp2,2q,p10,10q “ 0
0.01
0.05
0.1
94.6 (0.39)
94.6 (0.34)
96.3 (0.30)
90.6 (0.32)
92.9 (0.25)
94.8 (0.22)
92.2 (0.31)
92.5 (0.27)
94.7 (0.24)
95.6 (0.39)
94.9 (0.35)
94.6 (0.30)
Table 6: Simulation 4 (Gaussian and transelliptical data; element corruption mechanism). Percent
empirical coverage (average length) of 95% confidence intervals based on 1000 independent simulation
runs. Ω corresponds to a grid graph structure. r denotes the fraction of corrupted elements.
30
C
Gaussian vectors and the sign-subgaussian property
In this section we prove Lemma 4.5, which shows that that a Gaussian vector Z „ N pµ, Σq satisfies
the sign-subgaussianity property, that is, the centered sign vector signpZq ´ E rsignpZqs is itself subgaussian. (In this paper, we will apply this lemma only with µ “ 0, in which case E rsignpZqs “ 0 and
so signpZq is therefore subgaussian.)
Lemma 4.5. Let Z „ N pµ, Σq. Then signpZq ´ E rsignpZqs is CpΣq-subgaussian.
Proof of Lemma 4.5. Without loss of generality, rescale so that λmin pΣq “ 1 and then CpΣq “ λmax pΣq
(note that the mean µ is therefore rescaled as well). Write Σ “ AAJ ` Ip for some matrix A P Rnˆn .
iid
Then we can write Z “ µ ` X ` AY , where X, Y „ N p0, Ip q. Then, for any fixed vector v P Rp ,
” J
ı
” ” J
ıı
” ” J
ıı
E ev signpZq “ E E ev signpZq | Y “ E E ev signpµ`X`AY q | Y
«
ff
ı
ź ”
“E
E evi signpµi `Xi `pAY qi q | Y
,
i
where the last step holds because, conditional on Y , each of the terms signpµi ` Xi ` pAY qi q depends
on Xi only, and therefore these terms are conditionally independent. Next, observe that
“
`
˘
‰
E rsignpµi ` Xi ` pAY qi q | Y s “ E sign N p0, 1q ` µi ` pAY qi | Y
“ Φ pµi ` pAY qi q ´ Φ p´µi ´ pAY qi q “ ψ pµi ` pAY qi q ,
where we define ψpzq “ Φpzq ´ Φp´zq for z P R. Then, for each i,
”
ı
E evi signpµi `Xi `pAY qi q | Y
”
ı
“ E evi psignpµi `Xi `pAY qi q´ψpµi `pAY qi qq | Y ¨ evi ψpµi `pAY qi q
2
ď evi {2 ¨ evi ψpµi `pAY qi q ,
where the inequality is proved by applying Hoeffding’s Lemma (see, for example, Massart [2007,
Lemma 2.6]) to the bounded mean-zero random variable
vi psignpµi ` Xi ` pAY qi q ´ ψ pµi ` pAY qi qq
(where, since we are conditioning on Y , only X is treated as random). Combining the calculations so
far, we have
” J
ı
” J
ı
2
E ev signpZq ď e||v||2 {2 ¨ E ev ψpµ`AY q ,
where ψpµ ` AY q applies the function ψp¨q elementwise to the vector µ ` AY .
Next we show that y ÞÑ v J ψpµ ` Ayq is Lipschitz over y P Rn . Note that z ÞÑ ψpzq is 1Lipschitz over z P R since the density of the standard normal distribution is bounded uniformly as
φpzq ď ?12π ď 21 . For any y, y 1 P Rn , we have
ˇ J
ˇ ÿ
ˇ
`
˘ˇ
ˇv ψpµ ` Ayq ´ v J ψpµ ` Ay 1 qˇ ď
|vi | ¨ ˇψ ppµ ` Ayqi q ´ ψ pµ ` Ay 1 qi ˇ
ď
ÿ
i
i
ˇ
ˇ
|vi | ¨ ˇpµ ` Ayqi ´ pµ ` Ay 1 qi ˇ ď ||v||2 ¨ ||Apy ´ y 1 q||2
ď ||v||2 ¨
31
a
λmax pΣq ´ 1 ¨ ||y ´ y 1 ||2 ,
where the last step is true because
||A||op “
Therefore, y ÞÑ v J ψpµ ` Ayq is
b
a
||Σ ´ Ip ||op “ CpΣq ´ 1 .
´
¯
a
||v||2 ¨ CpΣq ´ 1 -Lipschitz in y. We can now apply standard
concentration results for Lipschitz functions of a Gaussian: by Massart [2007, Proposition 3.5],
” J
ı
2
E ev pψpµ`AY q´Erψpµ`AY qsq ď e||v||2 pCpΣq´1q{2 .
Therefore,
” J
ı
”
ı
2
J
2
J
E ev signpZq ď E e||v||2 {2`v ψpµ`AY q ď e||v||2 ¨CpΣq{2 ¨ ev Erψpµ`AY qs .
Finally, for each i,
E rψpµi ` pAY qi qs “ E rE rsignpµi ` Xi ` pAY qi q | Y ss
“ E rsignpµi ` Xi ` pAY qi s “ E rsignpZi qs ,
” J
ı
or in other words, E rψpµ ` AY qs “ E rsignpZqs. Combining everything, we have proved that E ev signpZq ď
2
e||v||2 ¨CpΣq{2 ¨ ev
D
D.1
J
ErsignpZqs
, as desired.
Proof of main result
Preliminaries
We first compute bounds on ||γc ||2 and ||γc ||1 for each c “ a, b, which we will use many times in the
proofs below. First, for c “ a, b note that
´1
´1
||γc ||2 “ ||Σ´1
¨ λmax pΣq ď Ccov
I ΣIc ||2 ď ||ΣI || ¨ ||ΣIc ||2 ď rλmin pΣqs
(D.1)
by Assumption 3.1. Next,
||γc ||1 “ ||Σ´1
I ΣIc ||1
“ || ´ ΩI,ab Θab,c ||1 pby matrix blockwise inversionq
ÿ
ÿ
|Ωj,ab Θab,c | ď
“
||Ωj,ab ||1 ||Θab,c ||8
jPI
ď Ccov
ÿ
jPI
jPI
||Ωj,ab ||1
since
||Θ||8 ď λmax pΘq “ pλmin pΩab,ab qq
´1
ď pλmin pΩqq
´1
“ λmax pΣq ď Ccov .
Therefore,
||γc ||1 ď Ccov p||Ωa ||1 ` ||Ωb ||1 q ď 2Ccov Csparse
by applying Assumption 3.2.
32
a
kn
(D.2)
D.2
Proof of Theorem 4.1: asymptotic normality of the oracle estimator
Theorem 4.1. Suppose that Assumptions 3.1, 3.2, and 3.4 hold. Then there exist constants Cnormal , Cvariance
depending on Ccov , Csparse , Ckernel but not on pn, pn , kn q, such that
ˇ #
ˇ
+
ˇ
ˇ
r ab ´ Θab
?
kn logppn q
Θ
1
ˇ
ˇ
?
sup ˇP
n¨
ď t ´ Φptqˇ ď Cnormal ¨
`
,
ˇ
ˇ
S
¨
detpΘq
n
2p
ab
n
tPR
where Sab is defined in the proof and satisfies Sab ¨ detpΘq ě Cvariance ą 0.
r ab ´Θab can be approximated by a linear function
Proof of Theorem 4.1. We first show that the error Θ
p
of the Kendall’s tau estimator T . Define vectors u, v P Rpn with entries
ua “ 1, ub “ 0, uI “ ´γa and va “ 0, vb “ 1, vI “ ´γb .
r ab “ uJ Σv
p and Θab “ uJ Σv, that is, the error is given by
Then by definition, we have Θ
r ab ´ Θab “ uJ pΣ
p ´ Σqv .
Θ
´ ¯
` ˘
p “ sin π Tp and Σ “ sin π T , we take a second-order Taylor expansion of sinp¨q to see
Next, since Σ
2
2
that, for some t P r0, 1s,
„
´
¯
r ab ´ Θab “ uJ π cos π T ˝ pTp ´ T q´
Θ
2
2
´π
¯
1 ´ π ¯2
¨
¨ sin
pt ¨ T ` p1 ´ tq ¨ Tpq ˝ pTp ´ T q ˝ pTp ´ T q v . (D.3)
2
2
2
Next, we rewrite this linear term. We have
ı
´
´ π ¯¯
”
´π ¯
1 ÿ
T ˝ Tp v “ `n˘
signpXi ´ Xi1 qJ uv J ˝ cos
T signpXi ´ Xi1 q ,
L :“ uJ cos
2
2
2 iăi1
which is a U-statistic of order 2 with respect to the data pX1 , . . . , Xn q. Note that
”π
´π ¯
ı
”π
´π ¯
ı
”π
´π ¯
ı
L ´ ErLs “ uJ
cos
T ˝ Tp v ´ uJ
cos
T ˝ ErTps v “ uJ
cos
T ˝ pTp ´ T q v .
2
2
2
2
2
2
` J
` π ˘˘
1
1 J
1
Define the kernel gpX, X q “ signpX´X q uv ˝ cos 2 T signpX´X q, and let g1 pXq “ E rgpX, X 1 q | Xs,
“
‰
iid
where X, X 1 „ TEpΣ, ξ; f1 , . . . , fp q. Let νg21 “ Varpg1 pXqq and ηg3 “ E |gpX, X 1 q|3 . By Callaert and
Janssen [1978, Section 2], we have
ˇ "?
ˇ
*
ˇ
ˇ
ηg3
npL ´ ErLsq
1
ˇ
sup ˇP
ď t ´ Φptqˇˇ ď CUstat ¨ 3 ¨ ? ,
(D.4)
2ν
ν
n
g1
tPR
g1
for a universal constant CUstat . Next we bound the ratio
in Appendix F.8.
ηg3
νg3
in the following lemma, which is proved
1
Lemma D.1. Suppose that Assumptions 3.1, 3.2 and 3.4 hold. Let gpX, X 1 q and g1 pXq be defined as
in the proof of Theorem 4.1. Then
νg21 “ Varpg1 pXqq ě
and
1 2
C
π 2 variance
“
‰
νg31 ď ηg3 “ E |gpX, X 1 q|3 ď Cmoment
where Cvariance , Cmoment are constants depending only on Ccov , Ckernel and not on pn, pn , kn q.
33
In particular, this lemma implies that Sab :“ πνg1 pdetpΘqq´1 ě Cvariance ¨ pdetpΘqq´1 .
r ab ´ Θab . Define
Next, the linear term L analysed here provides only an approximation to Θ
r ab ´ Θab ´ π pL ´ ErLsq .
∆“Θ
2
Then we have
ˇ „ ´ ¯
ˇ
´π
¯
ˇ
ˇ
1
π 2
|∆| “ ˇˇuJ
¨
¨ sin
pt ¨ T ` p1 ´ tq ¨ Tpq ˝ pTp ´ T q ˝ pTp ´ T q v ˇˇ
2
2
2
› ´ ¯
›
´π
¯
›1
›
π 2
ď ||u||1 ||v||1 ›› ¨
¨ sin
pt ¨ T ` p1 ´ tq ¨ Tpq ˝ pTp ´ T q ˝ pTp ´ T q››
2
2
2
8
π2
||u||1 ||v||1 ||Tp ´ T ||28
8
π2
ď
¨ kn ¨ p1 ` 2Ccov Csparse q2 ¨ ||Tp ´ T ||28 ,
8
ď
(D.5)
where the last inequality holds by (D.2).
Finally, the next lemma is proved in de la Pena and Giné [1999].
Lemma D.2 ([de la Pena and Giné, 1999, Theorem 4.1.8]). For any δ ą 0, with probability at least
1 ´ δ,
d
` `pn ˘ ˘
4
log
2 2 {δ
||Tp ´ T ||8 ď
.
n
Applying this lemma with δ “
least 1 ´ 2p1n .
1
2pn ,
we have ||Tp ´ T ||28 ď
4 logp2p3n q
n
ď
16 logppn q
n
with probability at
r ab ´ Θab “ π pL ´ ErLsq ` ∆, where (D.4) gives
To summarize the computations so far, we have Θ
2
an asymptotic normality result for the linear term pL ´ ErLsq, while (D.5) gives a bound on ∆. To
r ab ´ Θab is asymptotically normal, we will use the following lemma (proved in
prove therefore that Θ
Appendix F.1):
Lemma D.3. Let A, B, C be random variables such that
sup |P tA ď tu ´ Φptq| ď A
P t|B| ď δB , |C| ď δC u ě 1 ´ BC ,
and
tPR
where A , BC , δB , δC P p0, 1q. Then the variable pA ` Bq ¨ p1 ` Cq´1 converges to a standard normal
distribution with rate
ˇ
ˇ
(
δC
sup ˇP pA ` Bq ¨ p1 ` Cq´1 ď t ´ Φptqˇ ď δB `
` A ` BC .
1 ´ δC
tPR
We apply this lemma with A “
π
2
¨
?
n¨
L´ErLs
Sab ¨detpΘq
and B “
sup |P tA ď tu ´ Φptq| ď CUstat ¨ `
tPR
by (D.4) and Lemma D.1. Furthermore,
#
P |B| ď
?
n¨
π2
8
¨ kn ¨ p1 ` 2Ccov Csparse q2 ¨
Cvariance
16 logppn q
n
?
n¨
∆
Sab ¨detpΘq
Cmoment
˘1.5
1
2
π 2 Cvariance
“: δB
+
and C “ 0. We have
1
¨ ? “: A
n
"
*
16 logppn q
1
2
p
ď P ||T ´ T ||8 ď
ě1´
“: 1 ´ BC
n
2pn
34
by (D.5) and Lemmas D.1 and D.2. Noting that
Cnormal “
?
n¨
“ A ` B, and defining
2π 2 p1 ` 2Ccov Csparse q2
` CUstat ¨ `
Cvariance
we have proved the desired result.
D.3
r ab ´Θab
Θ
Sab
Cmoment
˘1.5
1
2
π 2 Cvariance
,
Proof of Theorem 4.2: gap between the estimator and the oracle
estimator, and estimation of the variance
Theorem 4.2. Suppose that Assumptions 3.1, 3.2, and 3.3 hold. Then there exists a constant Coracle ,
depending on Ccov , Csparse , Cest but not on pn, pn , kn q, such that, if 7 n ě 15kn logppn q, then, with
probability at least 1 ´ 2p1n ´ δn ,
q ´ Θ||
r 8 ď Coracle ¨ kn logppn q
||Θ
n
and
c
ˇ
ˇ
kn2 logppn q
ˇq
ˇ
q
.
ˇSab ¨ detpΘq ´ Sab ¨ detpΘqˇ ď Coracle ¨
n
q of Θ and
The first part of Theorem 4.2, which bounds the distance between our estimator Θ
r
b
the oracle estimator Θ, is established using bounds on Σ ´ Σ in Section 4.3. Details are given in
Appendix
D.3.1. The second
ˇ
ˇ part of Theorem 4.2, which bounds the error in estimating variance,
ˇq
ˇ
q
ˇSab ¨ detpΘq ´ Sab ¨ detpΘqˇ, is treated in Appendix D.3.2.
D.3.1
q ´Θ
r
Bounds on Θ
p ´ Σ, to derive a bound on the difference between our
We use our bounds on the covariance error, Σ
q
r of Θ. The bounds we give here are deterministic
empirical estimator Θ and the oracle estimator Θ
(given that our initial assumptions hold).
The following lemma is proved in Appendix F.7:
Lemma D.4. If Assumptions 3.1, 3.2, and 3.3 hold, then with probability at least 1 ´ δn ,
¸
˜
c
k
logpp
q
k
logpp
q
n
n
n
n
p ´ Σ||S ¨
q ´ Θ||
r 8 ď Csubmatrix
` ||Σ
||Θ
,
kn
n
n
where Csubmatrix is a constant depending on Ccov , Cest , and Csparse but not on pn, pn , kn q.
From this point on, we combine Corollary 4.8 and Lemma D.4 to obtain our probabilistic bound
q ´ Θ||
r 8 (Theorem 4.2). Looking first at Corollary 4.8, and setting δ1 “ δ2 “ 1 , we see that
on ||Θ
6pn
by the assumption pn ě 2, kn ě 1 and the assumption n ě 15kn logppn q stated in Theorem 4.2, the
conditions of Corollary 4.8 must hold. Then, with probability at least 1 ´ δ1 ´ δ2 “ 1 ´ 3p1n ,
p ´ Σ||S
||Σ
kn
c
`
` ˘˘
?
4 log 12pn p2n
π2
logp12pn q ` 2kn logp12pn q
ď
¨ kn ¨
` 2π ¨ 16p1 ` 5qCcov ¨
8
n
n
c
kn logppn q
1
ď Ccov ¨ C ¨
, (D.6)
n
7 Note that the additional condition n ě 15k logpp q can be assumed to hold in our main result Theorem 3.5, since
n
n
if this inequality does not hold, then the claim in Theorem 3.5 is trivial.
35
? ?
where we choose the universal constant C 1 “ 3π 2 ` 2π ¨ 16p1 ` 5q 15 which guarantees that the
last inequality holds (using the assumptions n ě kn logppn q, pn ě 2, and kn ě 1).
Now combining this result with Lemma D.4, we obtain
q ´ Θ||
r 8ď
||Θ
˘
`
kn logppn q
¨ Csubmatrix 1 ` Ccov ¨ C 1 .
n
Taking Coracle ě Csubmatrix p1 ` Ccov ¨ C 1 q, we have proved that the first bound in Theorem 4.2 holds
with probability at least 1 ´ 3p1n ´ δn .
D.3.2
Variance estimate
For the second part of the theorem, that is, bounding the error in the variance estimate Sqab , we state
this bound as a lemma and defer the proof to Appendix F.11, since we need to develop some additional
technical results before treating this bound.
Lemma D.5. Under the assumptions and definitions of Theorem 4.2, with probability at least 1´ 6p1n ,
if n ě kn2 logppn q, on the event that the bounds (3.1) in Assumption 3.3 hold,
c
ˇ
ˇ
kn2 logppn q
ˇq
ˇ
q
.
ˇSab ¨ detpΘq ´ Sab ¨ detpΘqˇ ď Coracle ¨
n
Combining this lemma with the work above, and using Assumption 3.3, we have proved that both
bounds stated in Theorem 4.2 hold with probability at least 1 ´ 2p1n ´ δn , as desired.
D.4
Proof of Theorem 3.5: main result
We now prove our main result, Theorem 3.5.
Proof of Theorem 3.5. Recall that our goal is to prove that
´1
? q
npΩab ´Ωab q
qab
S
converges to the N p0, 1q
distribution. Recalling that Θ “ pΩab,ab q and using the formula for a 2 ˆ 2 matrix inverse, we
separate this random variable into several terms:
?
q ab ´ Ωab q
npΩ
Sqab
¯
¯
? ´ ´Θ
? ´ q
q
q
detpΘq
´Θab
n detpab
´
n
´
Θ
`
Θ
¨
ab
ab
q
detpΘq
detpΘq
Θq
“
“
q
Sqab
Sqab ¨ detpΘq
´
¯¯
? ´
q
r ab ` Θ
r ab ´ Θ
q ab ´ Θab ¨ 1 ´ detpΘq
n Θab ´ Θ
detpΘq
“
q
q
Sab ¨ detpΘq
´
¯ fi
¯ ? ´
¯ ?
» ? ´
r ab ´ Θab
r ab ´ Θ
q ab
q
n Θ
n Θ
n ¨ Ωab ¨ detpΘq ´ detpΘq
fl
“ –´
`
`
Sab ¨ detpΘq
Sab ¨ detpΘq
Sab ¨ detpΘq
«
ff´1
q ´ Sab ¨ detpΘq
Sqab ¨ detpΘq
ˆ 1`
.
Sab ¨ detpΘq
?
q
´Ωab q
To show that npΩSab
converges to the standard normal distribution, we will can apply Lemma D.3
qab
(stated in Appendix D.2). In order to apply this lemma and obtain the desired result, we assemble
the following pieces:
? r
npΘab ´Θab q
q
? n `
First, the variable A :“ ´ Sab ¨detpΘq satisfies suptPR |P tA ď tu ´ Φptq| ď Cnormal ¨ kn logpp
n
36
1
2pn “: A ,
as shown in Theorem 4.1.
?
?
r ab ´Θ
q ab q
q q
npΘ
n¨Ωab ¨pdetpΘq´detpΘq
and C :“
Second, we define variables B :“ Sab ¨detpΘq `
Sab ¨detpΘq
and set
ˆ
˙
2
2
Coracle ` 4Ccov
Coracle ` 2Ccov Coracle
kn logppn q
?
δB “
¨
n
Cvariance
and
Coracle
¨
δC “
Cvariance
c
qab ¨detpΘq´S
q
S
ab ¨detpΘq
,
Sab ¨detpΘq
kn2 logppn q
.
n
We now show that, by Theorem 4.2, with probability at least 1 ´ 2p1n ´ δn “: 1 ´ BC it holds
that
|B| ď δB and |C| ďˇ δC . For the variable C, this is a trivial consequence of the bound on
ˇ
ˇq
q ´ Sab ¨ detpΘqˇˇ in Theorem 4.2 combined with the lower bound Sab ¨ detpΘq ě Cvariance
ˇSab ¨ detpΘq
from Theorem 4.1.
q ´ Θ||
r 8 ď Coracle ¨ kn logppn q
Now we turn to the bound on B. To prove this bound, observe that ||Θ
n
by Theorem 4.2 (with the stated probability). We also have
¯ˇ
ˇ? ´
ˇ n Θ
?
r ab ´ Θ
q ab ˇˇ
ˇ
?
1
n
q ´ Θ||
r 8ď
q ´ Θ||
r 8,
ˇď n¨
|B| “ ˇˇ
¨
||
Θ
¨ ||Θ
ˇ
S
¨
detpΘq
S
¨
detpΘq
C
ab
ab
variance
ˇ
ˇ
where the last step follows from Theorem 4.1. And,
ˇ
ˇ ˇ´
¯ `
˘ˇ
ˇ
q ´ detpΘqˇˇ “ ˇˇ Θ
q aa Θ
q bb ´ Θ
q 2 ´ Θaa Θbb ´ Θ2 ˇˇ
ˇdetpΘq
ab
ab
q ´ Θ||8 ` 2||Θ
q ´ Θ||2
ď 4||Θ||8 ||Θ
8
and
|Ωab | ď λmax pΩq “ pλmin pΣqq´1 ď Ccov .
Therefore,
´
¯ˇ
ˇ?
ˇ n ¨ Ω ¨ detpΘq ´ detpΘq
q ˇˇ
´
¯
ab
ˇ
?
|Ωab |
q ´ Θ||8 ` 2||Θ
q ´ Θ||2
ˇď n¨
ˇ
¨
4||Θ||
||
Θ
8
8
ˇ
ˇ
Sab ¨ detpΘq
Sab ¨ detpΘq
ˇ
ˇ
´
¯
?
Ccov
q ´ Θ||8 ` 2||Θ
q ´ Θ||2 ,
ď n¨
¨ 4Ccov ||Θ
8
Cvariance
where the last step follows from Theorem 4.1 along with the fact that
||Θ||8 ď λmax pΘq “ pλmin pΩab,ab qq
´1
ď pλmin pΩqq
´1
“ λmax pΣq ď Ccov .
Combining everything, we have
|B| ď
?
n¨
1
Cvariance
q ´ Θ||
r 8`
¨ ||Θ
´
¯
q ´ Θ||8 ` 2||Θ
q ´ Θ||2
¨ 4Ccov ||Θ
8
Cvariance
«
2
2
Coracle ` 2Ccov Coracle
¨
kn logppn q Coracle ` 4Ccov
?
ď
n
Cvariance
?
n¨
Ccov
kn logppn q
n
ff
.
If n ă kn logppn q, then the main result in Theorem 3.5 holds trivially. Assuming then that n ě
kn logppn q, we have proved the desired bound on |B|.
37
Given these convergence results, we apply Lemma D.3 to obtain the following result:
ˇ #?
ˇ
+
ˇ
ˇ
q ab ´ Ωab q
δC
npΩ
ˇ
ˇ
sup ˇP
ď t ´ Φptqˇ ď δB `
` A ` BC
q
ˇ
1 ´ δC
tPR ˇ
Sab
ˆ
˙
2
2
kn logppn q
Coracle ` 4Ccov
Coracle ` 2Ccov Coracle
?
“
¨
`
n
Cvariance
b
2 logpp q
kn
Coracle
n
kn logppn q
1
1
Cvariance ¨
n
b
?
` Cnormal ¨
`
`
` δn .
2 logpp q
n
2p
2p
k
n
n
n
n
oracle
¨
1 ´ CCvariance
n
b
2 logpp q
kn
n
oracle
If CCvariance
¨
ą 12 , then the result of Theorem 3.5 holds trivially, and so assuming that this is
n
not the case, we have
d
ˇ #?
ˇ
+
ˇ
ˇ
q ab ´ Ωab q
npΩ
kn2 log2 ppn q
1
ˇ
ˇ
sup ˇP
ď t ´ Φptqˇ ď Cconverge ¨
`
` δn ,
q
ˇ
ˇ
n
p
n
tPR
Sab
where
Cconverge :“
E
2
2
2Coracle
Coracle ` 4Ccov
Coracle ` 2Ccov Coracle
`
` Cnormal .
Cvariance
Cvariance
Accuracy of the initial Lasso estimator
Corollary 3.8. Suppose that Assumption 3.1 holds. Assume additionally that the columns Ωa , Ωb
of the true inverse covariance Ω “ Σ´1 are kn -sparse. Then there exist constants Csample , CLasso ,
depending on Ccov but not on pn, kn , pn q, such that if n ě Csample kn logppn q then, with probability at
qa of the objective function
least 1 ´ 2p1n , any local minimizer γ
1 Jp
p Ia ` λ||γ||1
γ ΣI γ ´ γ J Σ
2
?
over the set tγ P RI : ||γ||1 ď Ccov 2kn u satisfies
a
a
?
||q
γa ´ γa ||2 ď 3 2Ccov λ kn and ||q
γa ´ γa ||1 ď 24Ccov λ kn .
b
nq
where we choose λ “ CLasso ¨ logpp
. The same result holds for estimating γb .
n
Proof of Corollary 3.8. Define
pI, z “ Σ
p Ia , x‹ “ γa , p “ pn ´ 2, k “ kn .
A“Σ
Now we apply Theorem 3.7 to this sparse recovery problem. In order to do so, we need to check that
the conditions (3.3), (3.4), and (3.5) hold, and that γa is feasible under the condition ||γ||1 ď R. Once
these conditions are satisfied, the result of Theorem 3.7 can be applied to this setting.
?
Feasibility of γa Define R “ Ccov 2kn . As proved in (D.1), ||γa ||2 ď Ccov , and furthermore
||γa ||?
0 ď ||Ωa ||0 ` ||Ωb ||0 ď 2kn (this is true because γa “ ´ΩI,ab Θab,a by (D.2)). Therefore, ||γa ||1 ď
Ccov 2kn “ R.
38
Condition (3.3) (restricted strong convexity) Now we need to check that the restricted strong
p I . We will show that Corollary 4.8 implies that
convexity conditions (3.3) hold for our matrix A “ Σ
there exists a constant CRSC depending only on Ccov , such that if n ě 16 logppn q, then with probability
at least 1 ´ 8p1n , for all v P Rpn ,
ˆ
˙
ˇ ´
¯ ˇ
1
CRSC logppn q
ˇ J p
ˇ
||v||22 ` ||v||21 ¨
.
ˇv ΣI ´ ΣI v ˇ ď
2Ccov
n
n
, and apply Lemma 4.9 to obtain
To see this why this holds, set k “ CRSC logpp
nq
ˇ ´
¯ ˇ
?
ˇ
ˇ J p
p ´ Σ||S ď 2p||v||2 ` ||v||2 {kq||Σ
p ´ Σ||S .
ˇv ΣI ´ ΣI v ˇ ď p||v||2 ` ||v||1 { kq2 ||Σ
2
1
k
k
1
, we see that with probability
Then, applying Corollary 4.8 with this value of k and with δ1 “ δ2 “ 16p
n
1
at least 1 ´ 8pn ,
p ´ Σ||S ď 1 ,
||Σ
k
4Ccov
as long as we set the constant CRSC large enough.
Then, if this event holds, for all v P RI we have
ˇ ´
ˇ ´
¯ ˇ
¯ ˇ
p I v ě v J ΣI v ´ ˇˇv J Σ
p I ´ ΣI v ˇˇ ě C ´1 ¨ ||v||2 ´ ˇˇv J Σ
p I ´ ΣI v ˇˇ
vJ Σ
cov
2
ě
CRSC logppn q
1
¨ ||v||22 ´
¨
¨ ||v||21 .
2Ccov
2Ccov
n
Therefore, with probability at least 1 ´ 8p1n , the restricted strong convexity condition (3.3) holds with
α1 “
1
CRSC
and τ1 “
.
2Ccov
2Ccov
Below, we will prove that, with probability at least 1´ 8p3n ,
« ?
ff
c
c
a
2
1.5 3π 2 1 ` Ccov
π
logppn q
logppn q
a
`
¨
ď Cfeasible
, (E.1)
2
n
n
Csample
Condition (3.5) (penalty parameter)
p I γa ´ Σ
p Ia ||8
||Ax‹ ´ z||8 “ ||Σ
for a constant Cfeasible depending only on Ccov , as long as we set
”
ı2
a
?
2
Csample ě 16p1 ` 5qCcov 1 ` Ccov
.
Given that this is true, we now require that condition (3.5) holds, that is,
#
+
c
logppq
α1
‹
max 4||Ax ´ z||8 , 4α1
ďλď
.
n
6R
Define
CLasso
ff
+
a
?
2
1.5 3π 2 1 ` Ccov
2
π
a
“ max 4
Cfeasible `
,
,
2
Ccov
Csample
# «
Plugging in the bound (E.1), we see that the lower bound on λ is satisfied for λ “ CLasso
check the upper bound, we only need
c
logppn q
α1
1
?
λ “ CLasso
ď
“
.
2
n
6R
2Ccov
2kn
b
logppn q
.
n
To
Assuming that
2
4
n ě 8CLasso
Ccov
¨ kn logppn q ,
then this follows directly. Therefore, (3.5) is satisfied with probability at least 1 ´
39
(E.2)
3
8pn .
Condition (3.4) (sample size) To satisfy (3.4), by plugging in the definitions of R, α1 , and τ1
above, we see that it is sufficient to require
2
n ě 64Ccov
CRSC maxt1, 2CRSC u ¨ kn logppn q .
(E.3)
Conclusion Combining all of our work above, we see that the conditions (3.3), (3.4), and (3.5), and
the feasibility of γa , are all satisfied with probability at least 1 ´ 2p1n , as long as
n ě Csample kn logppn q
for
Csample
" ”
*
ı2
a
?
2
4
2
2
:“ max 16, 16p1 ` 5qCcov 1 ` Ccov , 8CLasso Ccov , 64Ccov CRSC maxt1, 2CRSC u .
Therefore, applying Theorem 3.7, if these high probability events hold, then then for any γ
qa that is a
local minimizer of
1 p
Jp
Lpxq “ γ J Σ
I γ ´ γ ΣIa ` λ||γ||1
2
?
over the set tγ P RI : ||γ||1 ď 2Ccov Csparse kn u, it holds that
?
a
a
?
6λ ¨ 2kn
1.5λ ¨ 2kn
γa ´ γa ||1 ď
||q
γa ´ γa ||2 ď
“ 3 2Ccov λ kn and ||q
“ 24Ccov λ kn .
α1
α1
By the same arguments, the same results hold for estimating γb .
Proving (E.1)
have
p I γa ´ Σ
p Ia ||8 . Since γa “ Σ´1 ΣIa , we
Now we consider the term ||Ax‹ ´ z||8 “ ||Σ
I
p I γa ´ Σ
p Ia ||8 “ ||pΣ
p I ´ ΣI qγa ´ pΣ
p Ia ´ ΣIa q||8 “ ||pΣ
p ´ Σqu||8 ,
||Σ
where u P Rpn is the fixed vector with
ua “ 1, ub “ 0, uI “ ´γa .
p ´ Σ (calculated as in (D.3)), we have
By the Taylor expansion of Σ
ˇ
ˇ
p ´ Σqu||8 “ max ˇˇeJ pΣ
p ´ Σquˇˇ
||pΣ
j
j
π
2
π
ď
2
ď
ˇ ´
´π ¯
¯ ˇ
ˇ
ˇ
max ˇeJ
cos
T ˝ pTp ´ T q uˇ `
j
j
2
ˇ ´
´π ¯
¯ ˇ
ˇ
ˇ
max ˇeJ
cos
T ˝ pTp ´ T q uˇ `
j
j
2
¯ ˇ
π 2 ˇˇ J ´ ´ π ¯
ˇ
T ˝ pTp ´ T q ˝ pTp ´ T q uˇ
ˇej sin
8
2
π2
||u||1 ||Tp ´ T ||28 .
8
(E.4)
Next we bound each term in this final
separately.
Beginning
a expression (E.4)
a the second
a
?
? with
2 ď
2 q,
term, by (D.1), we know that ||u||1 ď ||u||0 ||u||2 ď 1 ` 2kn ¨ 1 ` Ccov
kn ¨ 3p1 ` Ccov
where to bound ||u||0 we use the calculation ||γa ||0 ď 2kn from before. Furthermore, by Lemma D.2,
with probability at least 1 ´ 8p1n ,
c
c
12
logp8p
q
48 logppn q
n
||Tp ´ T ||8 ď
ď
,
n
n
using pn ě 2. Therefore, the second term in (E.4) is bounded as
« ?
ff
c
a
2
2a
2
a
6
3π
1
`
C
π2
π
48
logpp
q
logpp
q
n
n
cov
2
2 q
a
||u||1 ||Tp ´ T ||8 ď
kn ¨ 3p1 ` Ccov
ď
¨
, (E.5)
8
8
n
n
Csample
where we use the assumption n ě Csample logppn q.
Next we turn to the first term in (E.4). In order to bound this term, we begin by stating two
lemmas (proved in Appendix F.9):
40
Lemma E.1. There exist vectors
ř a1 , a2 , . . . and b1 , b2`, . . . ˘ with
ř||ar ||8 , ||br ||8 ď 1 for all r ě 1, and
a sequence t1 , t2 , ¨ ¨ ¨ ě 0 with r tr “ 4, such that cos π2 T “ rě1 tr ar bJ
r.
Lemma E.2. For fixed u, v with ||u||2 , ||v||2 ď 1, for any |t| ď 4p1`?n5qC ,
cov
¸
˜“
?
‰
2
”
´
¯ı
2
4p1 ` 5q t2 ¨ Ccov
J p
.
E exp t ¨ u pT ´ T qv ď exp
n
By Lemma E.1, we can write
cos
´π ¯ ÿ
T “
tr ¨ ar bJ
r ,
2
rě1
ř
where tr ě 0, r tr “ 4, and ||ar ||8 , ||br ||8 ď 1. Then
¯
A
´π ¯
E ÿ
´
´π ¯
T ˝ pTp ´ T q u “ cos
T ˝ ej uJ , Tp ´ T “
tr ¨ par ˝ ej qJ pTp ´ T qpbr ˝ uq .
eJ
cos
j
2
2
rě1
Note that
||ar ˝ ej ||2 ď ||ar ||8 ¨ ||ej ||2 ď 1
and, by (D.1),
Then for any |t| ď
? n
16p1` 5qCcov
||br ˝ u||2 ď ||br ||8 ¨ ||u||2 ď
?
,
2
1`Ccov
a
2 .
1 ` Ccov
”
!
´
´π ¯
¯ )ı
E exp t ¨ eJ
cos
T ˝ pTp ´ T q u
j
2
+ff
«
#
ı
ÿ ”
“ E exp
tr t ¨ par ˝ ej qJ pTp ´ T qpbr ˝ uq
ď
ÿ tr
r
4
rě1
”
! ”
ı)ı
E exp 4 t ¨ par ˝ ej qJ pTp ´ T qpbr ˝ uq
˜“
pby Jensen’s inequalityq
¸
? ‰2
2
2
q
p1 ` Ccov
5q 16t2 ¨ Ccov
pby Lemma E.2q
ď
exp
4
n
r
˜“
¸
? ‰2
2
2
4p1 ` 5q 16t2 ¨ Ccov
q
p1 ` Ccov
“ exp
.
n
a
? n ?
Observe that we can set t “ ˘ n logppn q, which satisfies |t| ď
as long as we set
2
16p1` 5qCcov 1`Ccov
”
ı
2
a
?
2
Csample ě 16p1 ` 5qCcov 1 ` Ccov
, due to the assumption n ě Csample logppn q. Then, we see that
for any C ą 0,
#
+
c
ˇ ´
¯ ˇ
´π ¯
logppn q
ˇ J
ˇ
p
T ˝ pT ´ T q uˇ ą C
P ˇej cos
2
n
„ ?
b
?
logppn q
π
p
n logppn q¨eJ
n logppn q¨C
j pcosp 2 T q˝pT ´T qqu´
n
ďE e
„ ?
b
?
logppn q
π
p
´ n logppn q¨eJ
n logppn q¨C
j pcosp 2 T q˝pT ´T qqu´
n
`E e
˜“
¸
#
+
c
a
? ‰2
2
2
a
p1 ` Ccov
q
4p1 ` 5q 16p n logppn qq2 ¨ Ccov
logppn q
ď 2 exp
¨ exp ´ n logppn q ¨ C
n
n
ÿ tr
4p1 `
ď
41
´
¯
?
2
2
2
´ C´r4p1` 5qs ¨16Ccov
p1`Ccov
q
2pn
“ 2p´5
n ď
1
,
4p2n
? ‰2
“
2
2
where we set C “ Cfeasible :“ 5 ` 4p1 ` 5q ¨ 16Ccov
p1 ` Ccov
q. Therefore,
#
+
c
ˇ ´
´π ¯
¯ ˇ
logpp
q
1
ˇ J
ˇ
n
P max ˇej cos
ď
T ˝ pTp ´ T q uˇ ą Cfeasible
.
j
2
n
4pn
(E.6)
Combining (E.5) and (E.6), and returning to (E.4), we have
« ?
ff
c
c
a
2
2
1.5
3π
1
`
C
logpp
q
logpp
q
π
n
n
cov
p I γa ´ Σ
p Ia ||8 ď Cfeasible
a
`
¨
,
||Σ
2
n
n
Csample
with probability at least 1 ´
F
3
8pn .
This proves (E.1).
Proofs of lemmas
F.1
Proof of the normal convergence lemma
Lemma D.3. Let A, B, C be random variables such that
sup |P tA ď tu ´ Φptq| ď A
P t|B| ď δB , |C| ď δC u ě 1 ´ BC ,
and
tPR
where A , BC , δB , δC P r0, 1q. Then the variable pA ` Bq ¨ p1 ` Cq´1 converges to a standard normal
distribution with rate
ˇ
ˇ
(
sup ˇP pA ` Bq ¨ p1 ` Cq´1 ď t ´ Φptqˇ ď δB `
tPR
δC
` A ` BC .
1 ´ δC
Proof of Lemma D.3. First, define truncated versions of B and C:
Then, for any t P R,
r “ signpBq ¨ mint|B|, δB u, C
r “ signpCq ¨ mint|C|, δC u .
B
ˇ
!
)ˇ
(
ˇ
r ¨ p1 ` Cq
r ´1 ď t ˇˇ
ˇP pA ` Bq ¨ p1 ` Cq´1 ď t ´ P pA ` Bq
!
)
r or C ‰ C
r ď BC .
ďP B‰B
r ď δB and |C|
r ď δC with probability 1.
Note that |B|
Next, fix any t ě 0 and suppose that A ď tp1 ´ δC q ´ δB . Then
r ¨ p1 ` Cq
r ´1 ď pptp1 ´ δC q ´ δB q ` δB q ¨ p1 ´ δC q´1 “ t ,
pA ` Bq
and so
!
)
r ¨ p1 ` Cq
r ´1 ď t ě P tA ď tp1 ´ δC q ´ δB u
P pA ` Bq
ě Φ ptp1 ´ δC q ´ δB q ´ A
“ Φptq ´ P ttp1 ´ δC q ´ δB ă N p0, 1q ă tp1 ´ δC qu ´ P ttp1 ´ δC q ă N p0, 1q ă tu ´ A
1
ď 1)
ě Φptq ´ δB ´ P ttp1 ´ δC q ă N p0, 1q ă tu ´ A (since the density of N p0, 1q is ď ?
2π
δC
ě Φptq ´ δB ´
´ A . pby applying Lemma F.1 stated belowq
1 ´ δC
42
r ¨ p1 ` Cq
r ´1 ď t. Then
To prove the reverse bound, suppose that pA ` Bq
r ¨ p1 ` Cq
r ´1 ¨ p1 ` Cq
r ´B
r ď tp1 ` δC q ` δB .
A “ pA ` Bq
Therefore,
!
)
r ¨ p1 ` Cq
r ´1 ď t ď P tA ď tp1 ` δC q ` δB u
P pA ` Bq
ď Φ ptp1 ` δC q ` δB q ` A
“ Φptq ` P ttp1 ` δC q ă N p0, 1q ă tp1 ` δC q ` δB u ` P tt ă N p0, 1q ă tp1 ` δC qu ` A .
˙
ˆ
1
ď1
ď Φptq ` δB ` P tt ă N p0, 1q ă tp1 ` δC qu ` A
since the density of N p0, 1q is ď ?
2π
ď Φptq ` δB ` δC ` A . pby applying Lemma F.1 stated belowq
Therefore, for all t ě 0,
ˇ !
ˇ
)
ˇ
r ¨ p1 ` Cq
r ´1 ď t ´ Φptqˇˇ ď δB `
ˇP pA ` Bq
δC
` A .
1 ´ δC
By identical arguments, we can prove the same for t ď 0.
Lemma F.1. For any 0 ď a ď b,
P ta ă N p0, 1q ă bu ď
ˆ
ˆ
˙
˙
b
b
1
ď
´1 ¨ ?
´1 .
a
a
2πe
Proof.
żb
2
1
? e´t {2 dt
2π
t“a
2
1
ď pb ´ aq ¨ ? e´a {2
2π
ˆ
˙
2
b
1
“
´ 1 ¨ ? ¨ a ¨ e´a {2
a
2π
˙
ˆ
1
b
´1 ¨ ?
,
ď
a
2πe
P ta ă N p0, 1q ă bu “
where the last step holds because suptą0 tt ¨ e´t
F.2
2
{2
u“
?1 .
e
Sign vector of a transelliptical distribution
iid
Lemma 4.4. Let X, X 1 „ TEpΣ, ξ; f1 , . . . , fp q. Suppose that Σ is positive definite, and that ξ ą 0
with probability 1. Then signpX ´ X 1 q is equal in distribution to signpZq, where Z „ N p0, Σq.
Proof of Lemma 4.4. First, since the fj ’s are strictly monotone, we see that signpX ´X 1 q has the same
distribution regardless of the choice of the fj ’s. Therefore it suffices to consider the case that the fj ’s
iid
are each the identity function, and so X, X 1 „ Ep0, Σ, ξq, that is, a zero-mean elliptical distribution.
In this case, by Lindskog et al. [2003, Lemma 1], X ´ X 1 „ Ep0, Σ, ζq where the distribution of the
random variable ζ ě 0 obeys ϕζ ptq “ ϕξ ptq2 , where ϕζ and ϕξ are the characteristic functions of ζ and
iid
ξ, respectively. Note that for two independent copies ξ1 , ξ2 „ ξ, we have ϕξ1 `ξ2 “ ϕξ1 ¨ϕξ2 “ ϕ2ξ “ ϕζ ,
D
and therefore, ζ “ ξ1 ` ξ2 . Since ξ ą 0 with probability 1, this proves that ζ ą 0 with probability 1.
43
Σ´1{2 Z
||Σ´1{2 Z||2
Next take Z „ N p0, Σq. Then
ζ ¨ Σ1{2 ¨
is uniformly distributed on the unit sphere, and so
Σ´1{2 Z
„ Ep0, Σ, ζq ,
||Σ´1{2 Z||2
which is the distribution of X ´ X 1 . Using the fact that ζ ą 0 with probability 1, we see that
signpX ´ X 1 q is equal in distribution to
ˆ
sign ζ ¨ Σ1{2 ¨
Σ´1{2 Z
||Σ´1{2 Z||2
˙
“ signpΣ1{2 ¨ Σ´1{2 Zq “ signpZq ,
as desired.
F.3
Proof of Lemma 4.7
Lemma 4.7. The following bound holds deterministically: for any k ě 1,
p ´ Σ||S ď
||Σ
k
π2
¨ k||Tp ´ T ||28 ` 2π||Tp ´ T ||Sk .
8
Proof of Lemma 4.7. By Taylor’s theorem,
´
¯ ´
¯
´
¯ ´
¯ ´
¯
2
p “ Σ ` π cos π T ˝ Tp ´ T ´ π sin π T̄ ˝ Tp ´ T ˝ Tp ´ T
Σ
2
2
8
2
where T̄ has entries τ̄ab “ p1 ´ tab qτab ` tab τpab , with tab P r0, 1s for each a, b. Taking any u, v P Sk ,
then,
ˇ
ˇ πˇ ”
¯ı ˇ π 2 ˇ ” ´ π ¯ ´
¯ ´
¯ı ˇ
´π ¯ ´
ˇ J p
ˇ
ˇ
ˇ
ˇ J
ˇ
T ˝ Tp ´ T v ˇ `
T̄ ˝ Tp ´ T ˝ Tp ´ T v ˇ .
ˇu pΣ ´ Σqv ˇ ď ˇuJ cos
ˇu sin
2
2
8
2
First, to bound the sinp¨q matrix term, note that
ˇ ” ´π ¯ ´
¯ ´
¯ı ˇ
ˇ J
ˇ
T̄ ˝ Tp ´ T ˝ Tp ´ T v ˇ
ˇu sin
2
› ´π ¯ ´
¯›
¯ ´
›
›
ď ||u||1 ||v||1 ›sin
T̄ ˝ Tp ´ T ˝ Tp ´ T › ď ||u||1 ||v||1 ||Tp ´ T ||28 ,
2
8
?
where the last step holds since the sinp¨q function lies in r´1, 1s. Furthermore, ||u||1 , ||v||1 ď k for
all u, v P Sk by definition.
` ˘
Next, we bound the cosp¨q matrix term. By Lemma E.1, we can express cos π2 T as a convex
combination,
´π ¯ ÿ
cos
T “
tr ¨ ar bJ
r ,
2
rě1
ř
where ar , br P Rp satisfy ||ar ||8 , ||br ||8 ď 1 for all r, and tr ě 0 satisfy r tr “ 4. Furthermore, for
u, v P Sk and for each r, note that u ˝ ar , v ˝ br P Sk due to the bound on ||ar ||8 , ||br ||8 . Then
ˇ ”
´π ¯ ´
¯ı ˇ ÿ ˇ ”
´
¯ı ˇ
ˇ J
ˇ
ˇ
p ´ T v ˇˇ
T ˝ Tp ´ T v ˇ ď
tr ˇuJ ar bJ
˝
T
ˇu cos
r
2
rě1
ˇ
ˇ
ˇ
ˇ
´
¯
ÿ ˇ
ˇ
ˇ
ˇ
“
tr ˇpu ˝ ar qJ Tp ´ T pv ˝ br qˇ ď 4 sup ˇu1J pTp ´ T qv 1 ˇ “ 4||Tp ´ T ||Sk .
u1 ,v 1 PSk
rě1
p ´ Σ||S “ maxu,vPS uJ pΣ
p ´ Σqv, this proves the lemma.
Using the definition ||Σ
k
k
44
F.4
Proof of Lemma 4.6
Lemma 4.6. Suppose that k ě 1 and δ P p0, 1q satisfy logp2{δq ` 2k logp12pq ď n. Then with
probability at least 1 ´ δ it holds that
c
?
logp2{δq ` 2k logp12pq
p
||T ´ T ||Sk ď 32p1 ` 5qCpΣq ¨
.
n
Proof of Lemma 4.6. This lemma is a straightforward combination of Lemma E.2 (stated in Appendix E) together with the following result:
Lemma F.2 (Adapted from Lemma 5.1 and Theorem 5.2 of Baraniuk et al. [2008]). Let A be a
random matrix satisfying
" 2*
(
c1 t
J
exp t ¨ u Au ď exp
for all |t| ď c0 n and all unit vectors u P Rp
(F.1)
n
for some constants c0 , c1 . Then for any k ě 1 and any δ P p0, 1q satisfying
logp2{δq ` k logp12pq ď nc20 c1 ,
with probability at least 1 ´ δ it holds that
c
16c1
J
plogp2{δq ` k logp12pqq for all k-sparse unit vectors u P Rp .
|u Au| ď
n
(F.2)
Combined, Lemma F.2 (applied with 2k in place of k, with c0 “ 4p1`?15qC , and c1 “ c´1
0 ) and
cov
Lemma E.2 immediately yield the bound
c
ˇ
ˇ
?
logp2{δq ` 2k logp12pq
ˇ J p
ˇ
,
sup ˇu pT ´ T quˇ ď 16p1 ` 5qCpΣq ¨
n
uPS2k
with probability at least 1 ´ δ, as long as logp2{δq ` 2k logp12pq ď nc20 c1 “ n. Next take any u, v P Sk .
u´v
Then u`v
2 , 2 P S2k .
uJ pTp ´ T qv “
ˆ
u`v
2
˙J
pTp ´ T q
ˆ
u`v
2
˙
´
ˆ
u´v
2
˙J
pTp ´ T q
ď 16p1 `
?
ˆ
u`v
2
c
5qCpΣq ¨
˙
logp2{δq ` 2k logp12pq
.
n
This proves Lemma 4.6, as desired.
We next turn to the proof of Lemma F.2.
Proof of Lemma F.2. (Adapted from Lemma
5.1 and Theorem 5.2 of Baraniuk et al. [2008].) First
b
16c1
fix any S Ă rps with |S| “ k. Let “
n plogp2{δq ` k logp12pqq. Following the same arguments
as in Baraniuk et al. [2008, Lemma 5.1], we can take a set U Ă RS of unit vectors, with |U| ď 12k ,
such that
ˇ
ˇ
ˇ J ˇ
r Ar
uˇ .
sup ˇuJ Auˇ ď 2 sup ˇu
u
rPU
unit uPRS
Furthermore, for any fixed u
r P U, for any 0 ă t ď c0 n,
(
“
‰
rJ Ar
P u
u ą {2 ď E t ¨ u
rJ Ar
u ´ t ¨ {2
ˆ 2
˙
c1 t
´ t ¨ {2
ď exp
n
ˆ
˙
n2
“ exp ´
,
16c1
45
where for the last step we set t “
n
4c1
ď c0 n. Similarly,
ˆ
˙
(
n2
r Ar
u ă ´{2 ď exp ´
P u
.
16c1
J
Therefore,
˙
"
*
ˆ
ˇ J ˇ
n2
k
ˇ
ˇ
r Ar
u ą {2 ď 2 ¨ 12 ¨ exp ´
,
P sup u
16c1
u
rPU
and so
P
Finally, taking all
F.5
`p˘
˙
*
ˆ
2
ˇ J ˇ
ˇu Auˇ ą ď 2 ¨ 12k ¨ exp ´ n
.
16c1
unit uPRS
"
sup
ď pk choices for S, we see that
"
˙
*
ˆ
(
n2
P sup uJ Au ď ě 1 ´ 2p12pqk ¨ exp ´
“1´δ .
16c1
uPSk
k
Proof of Corollary 4.8
Corollary 4.8. Take any δ1 , δ2 P p0, 1q and any k ě 1 such that logp2{δ2 q ` 2k logp12pq ď n. Then,
p ´ Σ holds:
with probability at least 1 ´ δ1 ´ δ2 , the following bound on Σ
p ´ Σ||S ď
||Σ
k
c
` ` ˘ ˘
?
4 log 2 p2 {δ1
π2
logp2{δ2 q ` 2k logp12pq
¨k¨
` 2π ¨ 32p1 ` 5qCpΣq ¨
.
8
n
n
Proof of Corollary 4.8. This proof is a straightforward combination of Lemmas D.2, 4.6, and 4.7. We
have
p ´ Σ||S
||Σ
k
π2
¨ k||Tp ´ T ||28 ` 2π||Tp ´ T ||Sk by Lemma 4.7
8
c
` ` ˘ ˘
?
4 log 2 p2n {δ1
logp2{δ2 q ` 2k logp12pq
π2
¨k
` 2π ¨ 32p1 ` 5qCpΣq ¨
,
ď
8
n
n
ď
where the last step holds by applying Lemma D.2 with δ “ δ1 and Lemma 4.6 with δ “ δ2 .
F.6
Proof of Lemma 4.9
Lemma 4.9 (Based on Proposition 5 of Sun and Zhang [2012a]). For any fixed matrix M P Rpˆp
and vectors u, v P Rp , and any k ě 1,
´
? ¯ ´
? ¯
|uJ M v| ď ||u||2 ` ||u||1 { k ¨ ||v||2 ` ||v||1 { k ¨ sup |u1J M v 1 |.
u1 ,v 1 PSk
Proof of Lemma 4.9. We first introduce two vector norms used in Proposition 5 of Sun and Zhang
[2012a],
||w||p2,kq “ sup ||wS ||2
and its dual norm
||w||˚p2,kq .
|S|ďk
Note that
||w||p2,kq “ sup |z J w|,
zPSk
46
from the norm definition. We then have
|uJ M v| ď ||u||˚p2,kq ||M v||p2,kq
by definition of the dual norm, and
J
||M v||2,k “ sup |w M v| ď sup
wPSk
wPSk
||v||˚p2,kq ||M J w||p2,kq
“ sup
wPSk
||v||˚p2,kq
ˆ
˙
sup |z M w| ,
J
J
zPSk
and so combining everything,
|uJ M v| ď ||u||˚p2,kq ||v||˚p2,kq ¨ sup |wJ M z| “ ||u||˚p2,kq ||v||˚p2,kq ¨ ||M ||Sk .
w,zPSk
Finally, Proposition 5(ii) of Sun and Zhang [2012a] (applied with q “ 2 and m “ k, in their notation),
proves that, for any w,
?
?
||w||˚2,k ď ||w||p2,4kq ` ||w||1 { k ď ||w||2 ` ||w||1 { k,
where the last inequality follows trivially from the norm definition. This proves the lemma.
F.7
Proof of Lemma D.4
Lemma D.4. If Assumptions 3.1, 3.2, and 3.3 hold, then with probability at least 1 ´ δn ,
˜
¸
c
k
logpp
q
k
logpp
q
n
n
n
n
q ´ Θ||
r 8 ď Csubmatrix
p ´ Σ||S ¨
||Θ
` ||Σ
,
kn
n
n
where Csubmatrix is a constant depending on Ccov , Cest , and Csparse but not on pn, pn , kn q.
Proof
of Lemma
D.4. Choose any c, d P ta, bu; we will bound the pc, dqth entry of the error, that is,
ˇ
ˇ
ˇq
ˇ
r
qc ´ γc for each c “ a, b. First, with probability at least 1 ´ δn , the bounds
ˇΘcd ´ Θcd ˇ. Write ∆c “ γ
on ∆c , ∆d in Assumption 3.3 all hold; assume that this event occurs from this point on.
We have
ˇ
ˇ
ˇq
r cd ˇˇ
ˇΘcd ´ Θ
ˇ´
¯ ´
¯ˇ
ˇ p
ˇ
Jp
p Id ´ Σ
pJ γ
pIγ
p cd ´ γ J Σ
p
pJ
qcJ Σ
qcJ Σ
qd ´ Σ
“ˇ Σ
cd ´ γ
Ic qd ` γ
c Id ´ ΣIc γd ` γc ΣI γd ˇ
ˇ
ˇ
ˇ Jp
p I γ d ´ ∆J Σ
p Id ´ Σ
p J ∆d ˇˇ
γ c ΣI γ
qd ´ γcJ Σ
“ ˇq
c
Ic
ˇ
ˇ
ˇ Jp
ˇ
Jp
p I ∆d ` ∆J Σ
p
pJ
“ ˇ∆c ΣI γd ` γcJ Σ
c I ∆d ´ ∆c ΣId ´ ΣIc ∆d ˇ
ˇ
ˇ
J
J
J
J
ˇ
ď ˇ ∆J
(F.3)
c ΣI γd ` γc ΣI ∆d ` ∆c ΣI ∆d ´ ∆c ΣId ´ ΣIc ∆d
ˇ
ˇ
ˇ J p
J p
J p
J p
J
p Ic ´ ΣIc q ∆d ˇˇ .
` ˇ∆c pΣI ´ ΣI qγd ` γc pΣI ´ ΣI q∆d ` ∆c pΣI ´ ΣI q∆d ´ ∆c pΣId ´ ΣId q ´ pΣ
Now we bound each of these terms. To bound the first term on the right-hand side of (F.3), we have
ˇ J
ˇ
J
J
ˇ∆c ΣI γd ` γcJ ΣI ∆d ` ∆J
ˇ
c ΣI ∆d ´ ∆c ΣId ´ ΣIc ∆d
ˇ J
ˇ
J ´1
J
J
J
ˇ
“ ˇ∆c ΣI Σ´1
I ΣId ` ΣIc ΣI ΣI ∆d ` ∆c ΣI ∆d ´ ∆c ΣId ´ ΣIc ∆d
ˇ J
ˇ
“ ˇ ∆c Σ I ∆d ˇ
ď ||∆c ||2 ¨ ||∆d ||2 ¨ ||Σ||op
ď ||∆c ||2 ¨ ||∆d ||2 ¨ CpΣq
ď Ccov pCest q2
kn logppn q
,
n
47
where the next-to-last step holds because
||ΣI || ď ||Σ|| “ λmin pΣq ¨ CpΣq ,
and we must have λmin pΣq ď 1 because diagpΣq “ 1, while the last step holds by Assumptions 3.1
and 3.3.
Finally, to bound the second term on the right-hand side of (F.3), we have
ˇ
ˇ
ˇ J p
p I ´ ΣI q∆d ` ∆J pΣ
p I ´ ΣI q∆d ´ ∆J pΣ
p Id ´ ΣId q ´ pΣ
p Ic ´ ΣIc qJ ∆d ˇˇ
ˇ∆c pΣI ´ ΣI qγd ` γcJ pΣ
c
c
ˇ
ˇ ˇ
ˇ ˇ
ˇ ˇ
ˇ
ˇ ˇ
ˇ
p Ic ´ ΣIc qJ ∆d ˇˇ
p Id ´ ΣId qˇˇ ` ˇˇpΣ
p I ´ ΣI q∆d ˇˇ ` ˇˇ∆J pΣ
p I ´ ΣI q∆d ˇˇ ` ˇˇ∆J pΣ
p I ´ ΣI qγd ˇˇ ` ˇˇγ J pΣ
ď ˇ∆J
p
Σ
c
c
c
c
ˇ
ˇ ˇ
ˇ ˇ
ˇ ˇ
ˇ
ˇ ˇ
ˇ
ˇ ˇ
ˇ ˇ J p
ˇ ˇ J p
ˇ J p
ˇ ˇ J p
J
p
“ ˇ∆c pΣI ´ ΣI qγd ˇ ` ˇγc pΣI ´ ΣI q∆d ˇ ` ˇ∆c pΣI ´ ΣI q∆d ˇ ` ˇ∆c pΣI ´ ΣI qed ˇ ` ˇeJ
c pΣI ´ ΣI q ∆d ˇ ,
where ec and ed are the basis vectors in RI corresponding to nodes c and d. Now, for each vector
w P tγc , γd , ec , ed u and each vector z P t∆c , ∆d u, we have
?
||w||2 ` ||w||1 { k ď Ccov ` 2Ccov Csparse ` 2
by (D.1) and (D.2), and
?
||z||2 ` ||z||1 { k ď 2Cest
c
kn logppn q
n
by Assumption 3.3, and so
c
ˇ
ˇ
ˇ J p
ˇ
p I ´ ΣI ||S ¨ pCcov ` 2Ccov Csparse ` 2q ¨ 2Cest kn logppn q
ˇw pΣI ´ ΣI qz ˇ ď ||Σ
kn
n
p I ´ ΣI ||S ď ||Σ
p ´ Σ||S trivially, the desired result of the
by applying Lemma 4.9. Noting that ||Σ
kn
kn
lemma follows trivially from these bounds by setting Csubmatrix appropriately.
F.8
Proof of Lemma D.1
Lemma D.1. Suppose that Assumptions 3.1, 3.2 and 3.4 hold. Let gpX, X 1 q and g1 pXq be defined as
in the proof of Theorem 4.1. Then
νg21 “ Varpg1 pXqq ě
and
1 2
C
π 2 variance
“
‰
νg31 ď ηg3 “ E |gpX, X 1 q|3 ď Cmoment
where Cvariance , Cmoment are constants depending only on Ccov , Ckernel and not on pn, pn , kn q.
Proof of Lemma D.1. First, we have
”
´
´ π ¯¯
ı
g1 pXq “E signpX ´ X 1 qJ uv J ˝ cos
T signpX ´ X 1 q | X
2
”`
´
´ π ¯¯
ı
˘J
1
“ E signpX ´ X q b signpX ´ X 1 q vec uv J ˝ cos
T
|X
2 ´
´
“`
˘
‰J
π ¯¯
1
1
J
“ E signpX ´ X q b signpX ´ X q | X vec uv ˝ cos
T
2
´
´ π ¯¯
“ h1 pXqJ vec uv J ˝ cos
T
,
2
48
where h1 pXq is defined in Assumption 3.4, and has variance Σh1 . Therefore,
νg21 “ Varpg1 pXqq
´
´ π ¯¯J
´
´ π ¯¯
“ vec uv J ˝ cos
T
¨ Σh1 ¨ vec uv J ˝ cos
T
2
2
´
´ π ¯¯J
´
´ π ¯¯
ě Ckernel ¨ vec uv J ˝ cos
T
¨ Σh ¨ vec uv J ˝ cos
T
pby Assumption 3.4q
2
2
ˆ
˙
´
´ π ¯¯J
“ Ckernel ¨ Var vec uv J ˝ cos
T
hpX, X 1 q
2
¯
´
´
´ π ¯¯
T signpX ´ X 1 q
“ Ckernel ¨ Var signpX ´ X 1 qJ uv J ˝ cos
´
´
´ π ¯¯2
¯
J
J
“ Ckernel ¨ Var signpZq uv ˝ cos
T signpZq
2
´
´ π ¯¯2
J
ě Ckernel ¨ Csigns ¨ uv ˝ cos
T
,
(F.4)
2
ab
where for the next-to-last step, we take Z „ N p0, Σq and apply Lemma 4.4 to see that signpX ´ X 1 q
has the same distribution as signpZq, and for the last step, we apply the following lemma (proved in
Appendix F.10).
Lemma F.3. Take any positive definite Σ P Rpˆp , any distinct a, b P t1, . . . , pu, and any matrix
M P Rpˆp with Mja “ 0 for all j. Let Z „ N p0, Σq. Then there exists a constant Csigns depending on
CpΣq only, such that
`
˘
2
Var signpZqJ M signpZq ě Csigns ¨ Mab
.
`
˘
π
J
Applying this lemma with M “ uv ˝ cos 2 T yields the lower bound (F.4) (since va “ 0 so that
Mja “ 0 for all j).
Finally, we have
´π
¯2
´
´ π ¯¯2
T
“ u2a vb2 cos
Tab ě pCcov q´2 ,
uv J ˝ cos
2
2
ab
where the last step holds because ua “ vb “ 1 and
´π
´π
¯ c
¯2 b
cos
Tab “ 1 ´ sin
Tab “ 1 ´ Σ2ab ě λmin pΣab,ab q ě pCcov q´1 .
2
2
To summarize, we have
νg21 ě
1 2
Ckernel Csigns
“: 2 Cvariance
.
2
Ccov
π
Next, we give an upper bound on νg21 :
νg21 “ Varpg1 pXqq
´
´ π ¯¯J
´
´ π ¯¯
T
¨ Σh1 ¨ vec uv J ˝ cos
T
“ vec uv J ˝ cos
2
2
´
´ π ¯¯J
´
´ π ¯¯
ď vec uv J ˝ cos
T
¨ Σh ¨ vec uv J ˝ cos
T
2
2
´
´
´ π ¯¯
¯
T signpX ´ X 1 q
pas for the lower bound earlierq
“ Var signpX ´ X 1 qJ uv J ˝ cos
2
`
˘
“
‰
“
‰
2{3
“ Var gpX, X 1 q ď E |gpX, X 1 q|2 ď E |gpX, X 1 q|3
“ ηg2 .
“
‰
Finally, we compute an upper bound on ηg3 “ E |gpX, X 1 q|3 . By Lemma E.1, there exists a
decomposition
´π ¯ ÿ
cos
T “
t r a r bJ
r
2
rě1
49
where tr ě 0,
ř
r tr
ď 4, and ||ar ||8 , ||br ||8 ď 1. Note that, by (D.1),
b
a
2
||u||2 “ 1 ` ||γa ||22 ď 1 ` Ccov
and similarly ||v||2 ď
a
2 . Then for each r,
1 ` Ccov
||u ˝ ar ||2 _ ||v ˝ br ||2 ď
Then we have
a
2 .
1 ` Ccov
„ˇ
ˇ3
´
´ π ¯¯
‰
“
ˇ
ˇ
E |gpX, X 1 q|3 “ E ˇsignpX ´ X 1 qJ uv J ˝ cos
T signpX ´ X 1 qˇ
2
»ˇ
ˇ3 fi
ˇÿ
ˇ
`
˘
ˇ
1 ˇ fl
tr ¨ signpX ´ X 1 qJ uv J ˝ ar bJ
“ E –ˇ
r signpX ´ X qˇ
ˇrě1
ˇ
”ˇ
ÿ tr
ˇ ı
`
˘
1 ˇ3
pby Jensen’s inequalityq
ď
¨ E ˇ4 signpX ´ X 1 qJ uv J ˝ ar bJ
r signpX ´ X q
4
rě1
”ˇ
ˇ ı
`
˘
1 ˇ3
ď 64 ¨ max E ˇsignpX ´ X 1 qJ uv J ˝ ar bJ
r signpX ´ X q
r
”ˇ
ˇ3 ˇ
ˇ3 ı
“ 64 ¨ max E ˇsignpX ´ X 1 qJ pu ˝ ar qˇ ¨ ˇsignpX ´ X 1 qJ pv ˝ br qˇ
r
c ”
ı c ”
ı
6
E |signpX ´ X 1 qJ pu ˝ ar q| ¨ E |signpX ´ X 1 qJ pv ˝ br q|
”ˇ
ˇ6 ı
ď 64||u ˝ ar ||32 ¨ ||v ˝ br ||32 ¨ max E ˇsignpX ´ X 1 qJ wˇ
||w||2 “1
”ˇ
ˇ6 ı
2 3
ď 64p1 ` Ccov q ¨ max E ˇsignpX ´ X 1 qJ wˇ
||w||2 “1
?
2 3
3
ď 64p1 ` Ccov q ¨ Ccov
¨ 6! ¨ 2 e “: Cmoment ,
“ 64 ¨ max
6
r
where the last inequality holds because signpX ´ X 1 q is Ccov -subgaussian by Lemmas 4.4 and 4.5.
F.9
Proofs of lemmas for the initial estimators
Lemma E.1. There exist vectors
ř a1 , a2 , . . . and b1 , b2`, . . . ˘ with
ř||ar ||8 , ||br ||8 ď 1 for all r ě 1, and
a sequence t1 , t2 , ¨ ¨ ¨ ě 0 with r tr “ 4, such that cos π2 T “ rě1 tr ar bJ
r.
Proof of Lemma E.1. We will use the matrix max norm, defined for a matrix M P Rd1 ˆd2 as
||M ||max “
min
"
max ||Apiq ||2 ¨ max ||Bpjq ||2 : r ě 1, A P R
1ďiďd1
1ďjďd2
d1 ˆr
,B P R
d2 ˆr
s.t. M “ A ¨ B
J
*
,
where Apiq and Bpjq denote the ith row of A and the jth row of B, respectively. The matrix max
norm satisfies several key properties that we will use here [Srebro and Shraibman, 2005]: first,
W ľ 0 ñ ||W ||max ď max Wii ;
i
second,
||W ||max ď 1 ñ
and finally,
(
W
P ConvexHull abJ : ||a||8 , ||b||8 ď 1 ;
2
||W ˝ puv J q||˚ ď ||W ||max for all unit vectors u, v and all matrices W ,
50
(F.5)
(F.6)
(F.7)
where recall that || ¨ ||˚` is the
˘ matrix nuclear norm (the sum of the singular values).
For our matrix cos π2 T , Wegkamp and Zhao [2013] show that
´ π ¯ ÿ ˆ1{2˙
T “
p´1qr Σ ˝2r Σ, and Σ ˝2r Σ ľ 0 for all r ,
cos
2
r
rě0
where Σ˝2r Σ is the matrix with entries given by elementwise powers of Σ, that is, pΣ˝2r Σqjk “ pΣjk q2r .
Then for each r ě 0, applying (F.5),
||Σ ˝2r Σ||max ď maxpΣ ˝2r Σqii “ maxpΣii q2r “ 1 ,
i
i
since Σ is a correlation matrix. Then
›
›
›
› ÿ ˆ1{2˙
›
´ π ¯›
›
›
›
›
r
p´1q Σ ˝2r Σ›
“›
T ›
›cos
›
›rě0 r
2
max
max
ˇ
ˇ
ÿ ˇˆ1{2˙ˇˇ
ÿ ˇˆ1{2˙ˇˇ
ˇ
ˇ
ˇ
ˇ
ď
ˇ r ˇ“2,
ˇ r ˇ ¨ ||Σ ˝2r Σ||max ď
rě0
rě0
where the last identity comes from Wegkamp and Zhao [2013]. Finally, by (F.6), we have
` ˘
(
cos π2 T
P ConvexHull abJ : ||a||8 , ||b||8 ď 1
4
` ˘
and so cos π2 T can be expressed as a convex combination as stated in the lemma.
Lemma E.2. For fixed u, v with ||u||2 , ||v||2 ď 1, for any |t| ď
”
´
¯ı
E exp t ¨ uJ pTp ´ T qv ď exp
n
?
,
4p1` 5qCcov
˜“
¸
? ‰2
2
4p1 ` 5q t2 ¨ Ccov
.
n
Proof of Lemma E.2. We start with a simple observation that
´
¯
1
1
uT Tp ´ T v “ pu ` vqJ pTp ´ T qpu ` vq ´ pu ´ vqJ pTp ´ T qpu ´ vq,
4
4
which gives us (via Cauchy-Schwartz)
”
´
¯ı
E exp t ¨ uJ pTp ´ T qv
„
ˆ
˙
1
1
J p
J p
“ E exp t ¨ pu ` vq pT ´ T qpu ` vq ´ t ¨ pu ´ vq pT ´ T qpu ´ vq
4
4
d „
ˆ
˙ d „
ˆ
˙
1
1
J
J
p
p
ď E exp t ¨ pu ` vq pT ´ T qpu ` vq
¨ E exp ´t ¨ pu ´ vq pT ´ T qpu ´ vq
.
2
2
Note that || 12 pu ` vq||2 _ || 21 pu ´ vq||2 ď 1. Therefore, it will be sufficient to show that for any
|t| ď 4p1`?n5qC and any unit vector w,
cov
”
´
¯ı
E exp 2t ¨ wJ pTp ´ T qw ď exp
˜“
4p1 `
¸
? ‰2 2
2
5q t ¨ Ccov
.
n
(F.8)
We will prove (F.8) using the Chernoff bounding technique. To that end, denote Sn the group of
permutations of rns, and for any i, let Xpiq denote the i-th row of X. For a fixed w P Rp , for each
i P rn{2s and σ P Sn , define
`
`
˘
˘
Zσ,i “ wJ sign pXpσpiqq ´ Xpσpi`n{2qq qpXpσpiqq ´ Xpσpi`n{2qq qJ ´ T w .
51
Observe that
´
¯
1 ÿ 2 ÿ
wT Tp ´ T w “
Zσ,i ,
n! σPS n
n
(F.9)
iPrn{2s
and that for any fixed σ P Sn , the Zσ,i ’s are i.i.d. for i “ 1, . . . , n{2, and are identically distributed as
˘
˘
`
`
Zr “ wJ sign pXp1q ´ Xp1`n{2q qpXp1q ´ Xp1`n{2q qJ ´ T w .
`
˘
Using Lemma 4.4 and Lemma 4.5, for any fixed unit vector w P Rp , wT sign Xpiq ´ Xpi`n{2q is a
Ccov -subgaussian random variable, and
”`
˘˘2
˘˘2 ı
`
`
`
.
Zr “ wT sign Xp1q ´ Xp1`n{2q
´ E wT sign Xp1q ´ Xp1`n{2q
Applying Lemma F.4 (stated below), for any |t| ď
˙
ˆ
”
´ ¯ı
2
32t2 Ccov
r
ď exp
E exp tZ ď exp
1 ´ 4tCcov
Then, referring back to (F.9), for 0 ă t ď
?1
,
2p1` 5qCcov
˜
2
32t2 Ccov
2?
1 ´ 1` 5
¸
´
¯
?
2
“ exp 8p1 ` 5q2 t2 Ccov
.
n
?
,
4p1` 5qCcov
˛fi
»
¨
”
´
´
¯ ¯ı
ÿ
ÿ
2
t
Zσ,i ‚fl
E exp twT Tp ´ T w “ E –exp ˝
n! σPS n
iPrn{2s
n
»
¨
˛fi
ÿ
ÿ
1
2t
ď
E –exp ˝
Zσ,i ‚fl
n! σPS
n
pby Jensen’s inequalityq
iPrn{2s
n
ˆ „
ˆ
˙˙n{2
1 ÿ
2t r
“
E exp
Z
,
n! σPS
n
n
since for any fixed σ, the Zσ,i ’s are i.i.d., and are each equal to Zr in distribution. Continuing from
this step, we have
ˆ
˙˙n{2
”
´
´
¯ ¯ı ˆ „
2t r
T
p
E exp tw T ´ T w “ E exp
Z
n
´
´
¯¯n{2
?
2
ď exp 8p1 ` 5q2 p2t{nq2 Ccov
˜“
¸
? ‰2
2
4p1 ` 5q t2 ¨ Ccov
“ exp
.
n
2
Lemma F.4. Suppose Z is C-subgaussian, that is, E rexp ptZqs ď eCt {2 for all t P R. Then
ˆ
˙
“
(‰
32t2 C 2
2
2
E exp tpZ ´ ErZ sq ď exp
1 ´ 4|t|C
for all |t| ă
1
4C .
We remark that it is well known that the square of a subgaussian random variable satisfies subgaussian
tails near to its mean [see, for example, Lemmas 5.5, 5.14, 5.15 in Vershynin, 2012], but here we obtain
small explicit constants.
52
Proof. The
“ first
‰ part of this proof follows the arguments in Vershynin [2012, Lemma 5.5]. First, we
bound E Z 2k for all integers k ě 1. We have
»˜c
«
#c
¸2k fi
+ff
k
k
“ 2k ‰
C
2k
2k
C
fl ď
“
E Z
E–
E p2kq! ¨ exp
¨Z
¨Z
p2kqk
C
p2kqk
C
$˜c ¸
,
2
&
. C k p2kq!ek
k
C
2k
p2kq!
¨
exp
¨
C{2
“
.
ď
%
p2kqk
C
p2kqk
Then, for any t ą 0,
«`
˘k ff
” 2ı
“ 2‰ ÿ
“ ‰ ÿ tk C k ek p2kq!
tZ 2
tZ
E
E e
“ 1 ` tE Z `
ď 1 ` tE Z 2 `
¨
k!
p2kqk
k!
kě2
kě2
“ ‰ ÿ tk C k ek e ¨ p2kq2k`1{2 ¨ e´2k
ď 1 ` tE Z 2 `
¨ ?
p2kqk
2π ¨ k k`1{2 ¨ e´k
kě2
ÿ
“ 2‰
e
k
p2tCq
“ 1 ` tE Z ` ?
π kě2
pby Stirling’s approximation for p2kq! and k!q
“ ‰
e
4t2 C 2
“ 1 ` tE Z 2 ` ? ¨
π 1 ´ 2tC
“ 2‰
8t2 C 2
,
ď 1 ` tE Z `
1 ´ 2tC
(F.10)
as long as 2tC ă 1 (we also use the fact ?eπ ď 2 to simplify). Next, trivially, for any k ě 2,
ˇ ”`
˘k ıˇˇ
“
‰
ˇ
1
,
ˇE Z 2 ´ ErZ 2 s ˇ ď 2k E Z 2k . Then we have, for |t| ă 4C
“
`
˘‰
E exp tpZ 2 ´ ErZ 2 sq
ÿ |t|k ˇˇ ”`
ÿ tk ”`
˘k ı
˘k ıˇˇ
E Z 2 ´ ErZ 2 s
ď1`
“1`
ˇE Z 2 ´ ErZ 2 s ˇ
k!
k!
kě2
kě2
˜
¸
ÿ 2k |t|k “
ÿ 2k |t|k “
‰
‰
“ ‰
2k
2k
ď1`
E Z
E Z
“
´ 2|t|E Z 2
k!
k!
kě2
kě0
“
`
˘‰
“ 2‰
2
“ E exp 2|t|Z
´ 2|t|E Z
“ ‰
“ ‰
32t2 C 2
´ 2|t|E Z 2
pby (F.10) with 2|t| in place of tq
ď 1 ` 2|t|E Z 2 `
1 ´ 4|t|C
"
*
32t2 C 2
ď exp
.
1 ´ 4|t|C
F.10
Lower bounds on variance for signs of a Gaussian
Lemma F.3. Take any positive definite Σ P Rpˆp , any distinct a, b P t1, . . . , pu, and any matrix
M P Rpˆp with Mja “ 0 for all j. Let Z „ N p0, Σq. Then there exists a constant Csigns depending on
CpΣq only, such that
`
˘
2
Var signpZqJ M signpZq ě Csigns ¨ Mab
.
Proof of Lemma F.3. Let p´aq denote the set rpsztau. By the law of total variance,
`
˘
“
`
˘‰
Var signpZqJ M signpZq ě E Var signpZqJ M signpZq | Zp´aq .
53
Let Mj,p´aq P Rp´1 denote the jth row of M with its ath entry removed, written as a column vector.
Then, recalling that Mja “ 0 for all j, we have
`
˘
Var signpZqJ M signpZq | Zp´aq
˜
¸
ÿ
J
J
“ Var signpZa q ¨ Ma,p´aq signpZp´aq q `
signpZj q ¨ Mj,p´aq signpZp´aq q | Zp´aq
j‰a
´
¯
J
“ Var signpZa q ¨ Ma,p´aq
signpZp´aq q | Zp´aq
¯2
`
˘ ´ J
“ Var signpZa q | Zp´aq ¨ Ma,p´aq
signpZp´aq q
´
¯ ´
¯2
J
J
“ Var signpZp´aq
βa ` N p0, νa2 qq ¨ Ma,p´aq
signpZp´aq q ,
J
where the last step holds since the distribution of Za conditional on Zp´aq is given by Zp´aq
βa `N p0, νa2 q
´1
´1
2
J
where βa “ Σp´aq Σp´aq,a and νa “ Σaa ´ Σp´aq,a Σp´aq Σp´aq,a . Continuing from this step, we have
˘
`
Var signpZqJ M signpZq | Zp´aq
ˆ
¯2
”
ı2 ˙ ´
J
J
¨ Ma,p´aq
signpZp´aq q
“ 1 ´ E signpZp´aq
βa ` N p0, νa2 qq
¨
¸2 ˛
˜ J
´
¯2
Zp´aq βa
J
‚¨ Ma,p´aq
signpZp´aq q ,
“ ˝1 ´ ψ
νa
where ψpzq “ Φpzq ´ Φp´zq P p´1, 1q.
ˆNow we˙will give a lower bound on the expectation of this quantity. First consider the term
ψ
J
Zp´aq
βa
νa
. Note that
J
Zp´aq
βa
νa
„N
˜
βaJ Σp´aq βa
0,
νa2
¸
“N
˜
0,
´1
ΣJ
p´aq,a Σp´aq Σp´aq,a
´1
Σaa ´ ΣJ
p´aq,a Σp´aq Σp´aq,a
¸
and this variance is bounded by CpΣq. Then, for any c P p0, 1q,
ˇ
#ˇ ˜ J
¸ˇ
+
#ˇ J
+
ˇ
ˇ
ˇZ
´a
¯
Zp´aq βa ˇˇ
ˇ
ˇ p´aq βa ˇ a
´1
´1
P ˇψ
CpΣq ¨ Φ p1 ´ c{2q
“P ˇ
ˇďψ
ˇ ď CpΣq ¨ Φ p1 ´ c{2q
ˇ
ˇ νa ˇ
ˇ
νa
!
)
a
ě P |N p0, CpΣqq| ď CpΣq ¨ Φ´1 p1 ´ c{2q “ 1 ´ c . (F.11)
`
˘
J
Next, note that Ma,p´aq
signpZp´aq q is ||Ma,p´aq ||22 ¨ CpΣq -subgaussian by Lemma 4.5, and
„´
¯2
J
E Ma,p´aq signpZp´aq q
ě ||Ma,p´aq ||22 ¨ λmin pT q ,
“
‰
` ˘
where T “ E signpZq signpZqJ (recall that Σ “ sin π2 T ). Furthermore, by Wegkamp and Zhao
[2013, Section 4.3], we have
2 ÿ
T “
gpkqΣ ˝k Σ ,
π kě1
where gpkq ě 0 are nonnegative scalars, gp1q “ 1, and Σ ˝k Σ is the k-fold Hadamard product, that
is, pΣ ˝k Σqij “ pΣij qk . Wegkamp and Zhao [2013, Section 4.3] show also that Σ ˝k Σ ľ 0 for all k.
Therefore,
2
2 ÿ
2
T “ Σ`
gpkqΣ ˝k Σ ľ Σ ,
π
π kě2
π
54
and so λmin pT q ě π2 λmin pΣq ě π2 pCpΣqq´1 . Applying Lemma F.5 (stated below),
"´
*
¯2
1
J
2
P Ma,p´aq signpZp´aq q ě ||Ma,p´aq ||2 ¨ λmin pT q{2 ě
,
2CpΣq{λ
min pT q
16e
and so,
P
Now set c “
¨
˝1 ´ ψ
˜
"´
J
Ma,p´aq
1
32eπCpΣq2
J
Zp´aq
βa
νa
¯2
signpZp´aq q ě ||Ma,p´aq ||22 ¨
1
πCpΣq
*
ě
1
16eπCpΣq2
in (F.11). Then, we see that with probability at least
¸2 ˛
´
¯2
J
‚¨ Ma,p´aq
signpZp´aq q ě
˜
ˆ
a
1´ψ
CpΣq ¨ Φ´1 1 ´
ˆ
1
64eπCpΣq2
˙˙2 ¸
.
1
32eπCpΣq2
,
¨ ||Ma,p´aq ||22 ¨
1
.
πCpΣq
Therefore, combining everything,
`
˘
Var signpZq M signpZq ě
J
1
32eπCpΣq2
¨
˜
ˆ
a
CpΣq ¨ Φ´1 1 ´
1´ψ
ˆ
1
64eπCpΣq2
˙˙2 ¸
¨
||Ma,p´aq ||22
.
πCpΣq
2
Noting that ||Ma,p´aq ||22 ě Mab
, this proves the desired result, where we define
˜
ˆ
ˆ
˙˙2 ¸
a
1
1
1
´1
Csigns “
.
¨ 1´ψ
CpΣq ¨ Φ
1´
¨
πCpΣq
32eπCpΣq2
64eπCpΣq2
Lemma F.5. Suppose that W P R is a random variable with ErW s “ 0, ErW 2 s ě C0 , and EretW s ď
2
eC1 t {2 for all t P R. Then
(
1
P W 2 ě C0 {2 ě
.
2C
16e 1 {C0
Proof of Lemma F.5.
C0 {2 ď ErW 2 s ´ C0 {2
(
(
“ ErW 2 ¨ 1I W 2 ě C0 {2 s ` ErW 2 ¨ 1I W 2 ă C0 {2 s ´ C0 {2
(
ď ErW 2 ¨ 1I W 2 ě C0 {2 s
?
?
( `
˘
ď C0 ErpeW { C0 ` e´W { C0 q ¨ 1I W 2 ě C0 {2 s since t2 ď et ` e´t for all t P R
?
?
(
(
“ C0 EreW { C0 ¨ 1I W 2 ě C0 {2 s ` C0 Ere´W { C0 ¨ 1I W 2 ě C0 {2 s
b
b
?
?
2
2
W
{
C
2
2
0
ď C0 Erpe
q s ¨ Er1I tW ě C0 {2u s ` C0 Erpe´W { C0 q2 s ¨ Er1I tW 2 ě C0 {2u s
b
b
?
?
“ C0 Ere2W { C0 s ¨ PtW 2 ě C0 {2u ` C0 Ere´2W { C0 s ¨ PtW 2 ě C0 {2u
b
b
2 {2
C
{C
¨2
2
1
0
ď C0 e
¨ PtW ě C0 {2u ` C0 eC1 {C0 ¨22 {2 ¨ PtW 2 ě C0 {2u ,
and rearranging terms we have proved the lemma.
55
F.11
Bounding the error in estimating the variance (Lemma D.5)
Lemma D.5. Under the assumptions and definitions of Theorem 4.2, with probability at least 1´ 6p1n ,
if n ě kn2 logppn q, on the event that the bounds (3.1) in Assumption 3.3 hold,
c
ˇ
ˇ
2
ˇ
ˇq
q ´ Sab ¨ detpΘqˇ ď Coracle ¨ kn logppn q .
ˇSab ¨ detpΘq
n
Proof of Lemma D.5. Recall from the proof of Theorem 4.1 that we have defined
´
´ π ¯¯
gpX, X 1 q “ signpX ´ X 1 qJ uv J ˝ cos
T signpX ´ X 1 q ,
2
and g1 pXq “ E rgpX, X 1 q | Xs, where
ua “ 1, ub “ 0, uI “ ´γa and va “ 0, vb “ 1, vI “ ´γb .
Recall from the proof of Lemma D.1, given in Appendix F.8, that we have
´
´ π ¯¯
´
´ π ¯¯J
T
¨ Σh1 ¨ vec uv J ˝ cos
T
,
νg21 “ Varpg1 pXqq “ vec uv J ˝ cos
2
2
where Σh1 “ Varph1 pXqq for
“
‰
2
h1 pXq “ E signpX ´ X 1 q b signpX ´ X 1 q | X P Rpn .
Note that ||Σh1 ||8 “ 1.
To estimate this variance, define vectors u
q and vq with entries
qa “ 1, u
u
qb “ 0, u
qI “ ´q
γa and vqa “ 0, vqb “ 1, vqI “ ´q
γb ,
and define
ph
Σ
1
˜
¸˜
¸J
ÿ
1 ÿp
1
1ÿ p
p
p
h1 pXi q ´
h1 pXi q ´
,
h1 pXi1 q
h1 pXi1 q
“
n i
n i1
n i1
where abusing notation, we write
p
h1 pXi q “
We then define
Writing
we then have
Define also
1 ÿ
1 ÿ
hpXi , Xi1 q “
signpXi ´ Xi1 q b signpXi ´ Xi1 q .
n ´ 1 i1 ‰i
n ´ 1 i1 ‰i
´
´ π ¯¯J
´
´ π ¯¯
J
p h ¨ vec u
q
qvqJ ˝ cos
νqg21 “ vec u
Tp
¨Σ
v
q
˝
cos
Tp
.
1
2
2
´
´ π ¯¯
´
´ π ¯¯
q “ vec u
qvqJ ˝ cos
x “ vec uv J ˝ cos
T
and x
Tp
,
2
2
ph x
qJ Σ
q.
νg21 “ xJ Σh1 x and νqg21 “ x
1
´
´ π ¯¯
qvqJ ˝ cos
x “ vec u
T
.
2
The following lemma, proved in Appendix F.12, carries out some elementary calculations on the norms
of these vectors x, x, x
q.
56
Lemma F.6. Define x, x, x
q as in the proof of Lemma D.5, and assume n ě kn2 logppn q. If the
bounds (3.1) in Assumption 3.3 hold then for constants C0 , C1 , C2 , C3 that depend only on Ccov , Csparse , Cest ,
||x||1 ď C0 kn ,
and with probability at least 1 ´
1
36pn ,
the following bounds all hold as well:
c
kn2 logppn q
,
n
c
kn3 logppn q
||q
x ´ x||1 ď C2
,
n
c
kn2 logppn q
||matpx ´ xq||`1 {`2 ď C3
,
n
x ´ x||1 ď C1
||q
2
pn
where matp¨q reshapes
ř a vector in R into a pn ˆ pn matrix, and where we define the matrix `1 {`2
norm as M`1 {`2 :“ j ||Mj ||2 , where Mj is the jth column of M .
We now continue bounding error in estimating νg1 . We have:
ˇ
ˇ ˇ
ˇ ˇ
ˇ 2
ˇ ˇ J
ˇ ˇ
p h xˇˇ .
p h ´ Σh qxˇˇ ` ˇˇx
ph x
ph x
ˇνqg ´ νg2 ˇ “ ˇˇq
q ´ xJ Σ
q ´ xJ Σh1 xˇ ď ˇxJ pΣ
qJ Σ
x Σ
1
1
1
1
1
1
1
(F.12)
We bound each term separately. For the first term in (F.12), we apply the following lemma (proved
in Appendix F.12 below):
Lemma F.7. Under the same assumptions and notation as Lemmas D.1 and D.5, for a universal
constant Cstudentized ,
+
#
c
ˇ
ˇ
kn2 logppn q
1
ˇ J p
ˇ
ě1´
.
P ˇx pΣh1 ´ Σh1 qxˇ ď Cstudentized
n
36pn
p h ľ 0 and so y ÞÑ
For the second term in (F.12), since Σ
1
triangle inequality,
b
p h y is a norm and must satisfy the
yJ Σ
1
ˇ ˇb
ˇ
b
b
ˇ ˇˇb
ˇ
ˇ ˇ
ˇ
ˇ ˇ
ˇ Jp
Jp
J
J
J
J
p
p
p
p
ˇ
ˇ
q ´ x Σh1 xˇ “ ˇ x
q ´ x Σh 1 x ˇ ¨ ˇ x
q ` x Σh1 xˇˇ
q Σh 1 x
q Σh1 x
x Σh1 x
ˇq
ˇb
ˇ2 ˇb
ˇ b
b
b
ˇ
ˇ
ˇ
ˇ
JΣ
JΣ
JΣ
ph x
p h xˇ ` ˇ x
ph x
p h xˇ ¨ 2 xJ Σ
ph x
qJ Σ
q
q
q
´
´
ď ˇˇ x
x
x
1
1 ˇ
1
1 ˇ
1
ˇ
ˇb
ˇ2 ˇb
ˇ c
b
b
ˇ
ˇ
ˇ
ˇ
ˇ
ˇ
Jp q ´
ph x
p h xˇ ` ˇ x
p h xˇ ¨ 2 xJ Σh x ` ˇˇxJ pΣ
p h ´ Σh qxˇˇ . (F.13)
qJ Σ
q ´ xJ Σ
ď ˇˇ x
xJ Σ
1
1 ˇ
1 ˇ
1
1
1
ˇ q Σh1 x
ˇ
ˇb
b
ˇ
ˇ
JΣ
ph x
p h xˇ which appears twice in the expression above,
qJ Σ
q
´
x
To bound the difference term ˇˇ x
1
1 ˇ
57
applying the triangle inequality several times, we have
ˇb
ˇ b
b
ˇ
ˇ
JΣ
JΣ
ph x
p h xˇ ď pq
p h pq
ˇ x
q
q
´
x
x ´ xqJ Σ
x ´ xq
1
1 ˇ
1
ˇ
cˇ
b
ˇ
ˇ
ˇ
J
p h ´ Σh qpq
x ´ xqJ pΣ
x ´ xqˇ
x ´ xq Σh1 pq
x ´ xq ` ˇpq
ď pq
1
1
b
b
p h ´ Σh ||8 ¨ ||q
x ´ xqJ Σh1 pq
ď pq
x ´ x||21
x ´ xq ` ||Σ
1
1
b
b
b
p h ´ Σh ||8 ¨ ||q
ď pq
x ´ xqJ Σh1 pq
x ´ xq ` px ´ xqJ Σh1 px ´ xq ` ||Σ
x ´ x||21
1
1
b
b
b
p h ´ Σh ||8 ¨ ||q
x ´ x||21 ` px ´ xqJ Σh1 px ´ xq ` ||Σ
x ´ x||21
ď ||Σh1 ||8 ||q
1
1
c
c
b
b
3
kn2 logppn q
p h ´ Σh ||8 ¨ C2 kn logppn q .
ď C1
` px ´ xqJ Σh1 px ´ xq ` ||Σ
1
1
n
n
(F.14)
(F.15)
(F.16)
(F.17)
(F.18)
(F.19)
Next, we state two lemmas, which are proved in Appendix F.12.
Lemma F.8. With probability at least 1 ´
1
9pn ,
p h ´ Σh ||8 ď 100
||Σ
1
1
c
logppn q
.
n
2
Lemma F.9. Let Σh1 be defined as in Assumption 3.4. For every z P Rpn ,
z J Σh1 z ď λmax pΣq ¨ ||matpzq||2`1 {`2 ,
where ||matpzq||`1 {`2 is defined as in the statement of Lemma F.6.
From this point on, we assume that the bounds derived in Lemmas F.6 and F.8 all hold (which the
lemmas have shown to be true with probability at least 1 ´ 6p1n , on the event that the bounds (3.1)
of Assumption 3.3 hold.) By Lemmas F.6 and F.9,
px ´ xqJ Σh1 px ´ xq ď Ccov ¨
˜
C3
c
kn2 logppn q
n
¸2
.
Applying this bound, along with the high probability events of Lemmas F.7 and F.8, we return
to (F.14) and obtain
ˇb
ˇ
b
ˇ
ˇ
J
J
p
p
ˇ x
q ´ x Σh1 xˇˇ ď
ˇ q Σh 1 x
g
˜ c
¸2 d c
f
c
c
2 logpp q
kn2 logppn q f
k
logppn q
kn3 logppn q
n
e
n
C1
` Ccov ¨ C3
` 100
¨ C2
n
n
n
n
˜
¸ c
c
c
2
a
kn2 logppn q
kn2 logppn q
4 kn logppn q
“
¨ C1 ` C3 Ccov ` 10C2
ď
¨ C4 ,
n
n
n
?
where for the last step we define C4 “ C1 ` C3 Ccov ` 10C2 and use the assumption n ě kn2 logppn q.
58
Next, returning to (F.13),
ˇ
ˇ
ˇ Jp
p h xˇˇ
x Σh1 x
q ´ xJ Σ
ˇq
1
ˇb
ˇ2 ˇb
ˇ c
b
b
ˇ
ˇ
ˇ
ˇ
ˇ
ˇ
JΣ
JΣ
JΣ
p h ´ Σh qxˇˇ
ph x
p h xˇ ` ˇ x
ph x
p h xˇ ¨ 2 xJ Σh x ` ˇˇxJ pΣ
qJ Σ
q
q
q
ď ˇˇ x
´
x
´
x
1
1
1
1 ˇ
1
1 ˇ
1
ˇ
d
c
c
2
kn2 logppn q
kn2 logppn q
2{3
2 kn logppn q
ď C4 ¨
` C4
¨ 2 Cmoment ` Cstudentized
,
n
n
n
where the last step applies the high probability event of LemmabF.7, and uses the fact that xJ Σh1 x “
2{3
2{3
νg21 ď Cmoment by Lemma D.1. Defining C5 “ C42 ` C4 ¨ 2 Cmoment ` Cstudentized , and using the
assumption n ě kn2 logppn q, we have
c
ˇ
ˇ
kn2 logppn q
ˇ
ˇ Jp
Jp
q ´ x Σh1 xˇ ď C5
x Σh1 x
.
ˇq
n
Finally, returning to (F.12) and applying Lemma F.8, we see that
c
c
2 logpp q
ˇ
ˇ 2
k
kn2 logppn q
n
n
2
ˇνqg ´ νg ˇ ď Cstudentized ¨
` C5
.
1
1
n
n
Next, we have
b
ˇ 2
ˇ
ˇ 2
ˇ
2 logpp q
kn
n
ˇνqg ´ νg2 ˇ
ˇνqg ´ νg2 ˇ
pC
`
C
q
studentized
5
n
1
1
1
1
|q
νg1 ´ νg1 | “
ď
,
ď
1
νqg1 ` νg1
νg1
π Cvariance
where for the denominator we apply Lemma D.1. Finally, since we know that Sab “ πνg1 ¨ pdetpΘqq´1
q ´1 , and then we have
and Sqab “ πq
νg ¨ pdetpΘqq
1
Defining
b
2 logpp q
kn
n
ˇ
ˇ
pC
`
C
q
studentized
5
n
ˇq
ˇ
q ´ Sab ¨ detpΘqˇ “ π ¨ |q
νg1 ´ νg1 | ď π ¨
.
ˇSab ¨ detpΘq
1
π Cvariance
Coracle ě π ¨
we see that
F.12
Cstudentized ` C5
,
1
π Cvariance
c
ˇ
ˇ
2
ˇq
ˇ
q ´ Sab ¨ detpΘqˇ ď Coracle ¨ kn logppn q .
ˇSab ¨ detpΘq
n
Calculations for the variance estimate (Lemma D.5)
q as in the proof of Lemma D.5, and assume n ě kn2 logppn q. If the
Lemma F.6. Define x, x, x
bounds (3.1) in Assumption 3.3 hold then for constants C0 , C1 , C2 , C3 that depend only on Ccov , Csparse , Cest ,
||x||1 ď C0 kn ,
59
and with probability at least 1 ´
1
36pn ,
the following bounds all hold as well:
c
kn2 logppn q
,
n
c
kn3 logppn q
||q
x ´ x||1 ď C2
,
n
c
kn2 logppn q
||matpx ´ xq||`1 {`2 ď C3
,
n
x ´ x||1 ď C1
||q
2
pn
where matp¨q reshapes
ř a vector in R into a pn ˆ pn matrix, and where we define the matrix `1 {`2
norm as M`1 {`2 :“ j ||Mj ||2 , where Mj is the jth column of M .
Proof of Lemma F.6. We calculate
›
›
´ π ¯›
´ π ¯›
›
›
›
›
T › ď ||uv J ||1 ¨ ›cos
T › ď ||u||1 ||v||1
||x||1 “ ›uv J ˝ cos
2
2
1
8
ď kn p1 ` 2Ccov Csparse q2 “: C0 kn ,
where for the last inequality we apply (D.2).
Next,
ˇ
´
´π ¯
´ π ¯¯›
ˇ
›
uvqJ ˝ cos
x ´ x||1 “ || ˇ|}q
Tp ´ cos
T ›
||q
2
2
1
›
´π ¯
´ π ¯›
›
›
ď p||u||1 ` ||q
u ´ u||1 q ¨ p||v||1 ` ||q
v ´ v||1 q ¨ ›cos
Tp ´ cos
T › .
2
2
8
Applying (D.2) and Assumption 3.3, and the fact that cosp¨q is 1-Lipschitz, if the bounds in Assumption 3.3 hold, we then have
||q
x ´ x||1 ď
˜
a
kn p1 ` 2Ccov Csparse q ` Cest
c
kn2 logppn q
n
Furthermore, applying Lemma D.2, with probability at least 1 ´
¸2
1
36pn ,
¨
π p
||T ´ T ||8 .
2
˜
¸2
c
c
a
kn2 logppn q
π 4 logp36p3n q
x ´ x||1 ď
||q
kn p1 ` 2Ccov Csparse q ` Cest
¨
.
n
2
n
Finally, since 4 logp36p3n q ď 36 logppn q ď 4n, where the last step holds by assumption in Theorem 4.2,
c
kn2 logppn q
2
¨ pp1 ` 2Ccov Csparse q ` Cest q ¨ π
||q
x ´ x||1 ď
n
c
kn2 logppn q
2
“ C1
for C1 :“ pp1 ` 2Ccov Csparse q ` Cest q ¨ π .
n
60
Next,
||q
x ´ x||1 ď ||q
x ´ x||1 ` ||x ´ x||1
c
kn2 logppn q
ď C1
` ||x ´ x||1
n
c
´ π ¯›
kn2 logppn q ›› J
›
uvq ´ uv J q ˝ cos
“ C1
` ›pq
T ›
n
2
1
c
›
›
´ π ¯›
´
¯›
2
kn logppn q ›
π ›
›
›
u ´ uqv J ˝ cos
upq
v ´ vqJ ˝ cos
ď C1
` ›q
T › ` ›pq
T ›
n
2
2
1
1
c
›
›
´
´
¯›
2
π ¯››
π ›
kn logppn q
›
›
ď C1
u ´ u||1 ||v||1 ›cos
` ||q
u||1 ||q
v ´ v||1 ›cos
T › ` ||q
T ›
n
2
2
8
8
c
2
kn logppn q
u||1 ||q
v ´ v||1 ` ||q
u ´ u||1 ||v||1
ď C1
` ||q
n
˜
¸
c
c
c
a
kn2 logppn q
kn2 logppn q
kn2 logppn q
`
kn p1 ` 2Ccov Csparse q ` Cest
¨ Cest
ď C1
n
n
n
c
a
kn2 logppn q
` kn p1 ` 2Ccov Csparse q ¨ Cest
n
c
kn3 logppn q
,
ď C2
n
where the next-to-last step applies (D.2) (assuming the bounds (3.1) in Assumption 3.3 hold), and
2
2
where for the last step, C2 “ C1 `2p1`2C
` covJCsparseJq¨Cest `C
` πest and
˘˘ we use the assumption n ě kn logppq.
uvq ´ uv q ˝ cos 2 T , we calculate the `1 {`2 norm of this
Finally, noting that x ´ x “ vec pq
matrix:
›
´ π ¯›
› J
›
uvq ´ uv J q ˝ cos
T ›
||matpx ´ xq||`1 {`2 “ ›pq
2
`1 {`2
´ π ¯ı ››
ÿ ››”
J
J
›
› pq
“
› uvq ´ uv q ˝ cos 2 T j ›
2
j
´ π ¯›
ÿ ››“
‰ › ››
›
J
J ›
qvq ´ uv j › ¨ ›cos
ď
T ›
› u
2
2
8
j
›
ÿ ››“
‰ ›
qvqJ ´ uv J j ›
ď
› u
2
j
ď
“
ÿ
j
u ¨ pq
vj ´ vj q||2 ` ||pq
u ´ uq ¨ vj ||2
||q
j
||q
u||2 ¨ |q
vj ´ vj | ` ||q
u ´ u||2 ¨ |vj |
ÿ
u||2 ||q
v ´ v||1 ` ||q
u ´ u||2 ||v||1 .
“ ||q
Next, applying (D.1), if the bounds (3.1) in Assumption 3.3 hold, we then have
||matpx ´ xq||`1 {`2
˜
¸
c
c
c
a
k
logpp
q
kn2 logppn q
kn logppn q a a
n
n
2
ď
1 ` Ccov ` Cest
¨ Cest ¨
` Cest ¨
¨ kn 1 ` Ccov
n
n
n
c
kn2 logppn q
ď C3
,
n
a
2
2
2 ¨C
where we define C3 “ 2 1 ` Ccov
est ` Cest and use the assumption that n ě kn logppn q.
61
Lemma F.7. Under the same assumptions and notation as Lemmas D.1 and D.5, for a universal
constant Cstudentized ,
+
#
c
ˇ
ˇ
1
kn2 logppn q
ˇ
ˇ J p
ě1´
P ˇx pΣh1 ´ Σh1 qxˇ ď Cstudentized
.
n
36pn
Proof of Lemma F.7. By definition, we have Σh1 “ Varph1 pXqq for
“
‰
2
h1 pXq “ E signpX ´ X 1 q b signpX ´ X 1 q | X P Rpn .
Therefore, since x is fixed,
xJ Σh1 x “ xJ Varph1 pXqqx “ VarpxJ h1 pXqq “ Varpg1 pXqq “ νg21 ,
where we recall that g1 pXq “ E rgpX, X 1 q | Xs where we define the kernel
´
´ π ¯¯
T signpX ´ X 1 q “ xJ hpX, X 1 q .
gpX, X 1 q “ signpX ´ X 1 qJ uv J ˝ cos
2
Define
gpX, X 1 qgpX, X 2 q ` gpX 1 , XqgpX 1 , X 2 q ` gpX 2 , XqgpX 2 , X 1 q
.
3
Note that γpX, X 1 , X 2 q is a U-statistic of order 3, and that
γpX, X 1 , X 2 q “
sup |gpX, X 1 q| ď ||x||1 sup ||hpX, X 1 q||8 “ ||x||1 ď C0 kn ,
X,X 1
X,X 1
where the last step applies Lemma F.6. So,
||γ||8 :“
sup
X,X 1 ,X 2
|γpX, X 1 , X 2 q| ď sup |gpX, X 1 q|2 ď C02 kn2 .
X,X 1
And,
“
‰
“
‰
Varpγq :“ VarpγpX, X 1 , X 2 qq ď E γpX, X 1 , X 2 q2 ď E |gpX, X 1 q|4
“
‰
ď E |gpX, X 1 q|3 ¨ C0 kn ď Cmoment ¨ C0 kn ,
where we use Lemma D.1 for the last bound.
Next, we have
“
‰
“
‰
“ “
‰‰
E γpX, X 1 , X 2 q “ E gpX, X 1 qgpX, X 2 q “ E E gpX, X 1 qgpX, X 2 q|X
“ “
‰ “
‰‰
“
‰
“ E E gpX, X 1 q|X E gpX, X 2 q|X “ E g1 pXq2 .
Therefore,
“
‰
“
‰
“
‰2
2
xJ Σh1 x “ νg21 “ Varpg1 pXqq “ E γpX, X 1 , X 2 q ´ E rg1 pXqs “ E γpX, X 1 , X 2 q ´ E gpX, X 1 q .
p h , we obtain
Next, examining the definition of Σ
1
1
x Σh 1 x “
npn ´ 1q2
Jp
«
ÿ
i‰i1 ‰i2
γpXi , Xi1 , Xi2 q `
62
ÿ
i‰i1
gpXi , Xi1 q
2
ff
´
˜
1 ÿ
` n˘
gpXi , Xi1 q
2
iăi1
¸2
.
Therefore, using the fact that |gpX, X 1 q| ď C0 kn always,
ˇ
ˇ
ˇ ˇ 1
ˇ
ˇ
ˇ ˇ` ˘ ÿ
ˇ Jp
J
1
2 ˇ
γpXi , Xi1 , Xi2 q ´ ErγpX, X , X qsˇ
ˇx Σh1 x ´ x Σh1 xˇ ď ˇ n
ˇ 3 iăi1 ăi2
ˇ
ˇ˜
ˇ
¸2
ˇ
ˇ
ˇ
C02 kn2 ˇˇ 1 ÿ
1 2ˇ
`
˘
`
gpXi , Xi1 q ´ ErgpX, X qs ˇ .
`
n ´ 1 ˇˇ n2 iăi1
ˇ
Now, using Bernstein’s inequality for U-statistics (Peel et al. [2010, Theorem 2]), for any δ ą 0,
ˇ d
#ˇ
+
ˇ 1
ˇ
ÿ
2Varpγq logp2{δq 2||γ||8 logp2{δq
ˇ` ˘
1
2 ˇ
P ˇ n
ďδ.
γpXi , Xi1 , Xi2 q ´ ErγpX, X , X qsˇ ą
`
ˇ 3 iăi1 ăi2
ˇ
pn{3q
3pn{3q
Therefore, with probability at least 1 ´
1
72pn ,
ˇ
ˇ
ˇ 1
ˇ
ˇ` ˘ ÿ
ˇ
γpXi , Xi1 , Xi2 q ´ ErγpX, X 1 , X 2 qsˇ ď
ˇ n
ˇ 3 iăi1 ăi2
ˇ
d
c
2Cmoment ¨ C0 kn logp2 ¨ 72pn q 2C02 kn2 logp2 ¨ 72pn q
kn2 logppn q
`
ď
¨ C1 ,
pn{3q
3pn{3q
n
where
C1 “
a
6Cmoment C0 log2 p144q ` 2C02 log2 p144q ,
and we use the assumption n ě kn2 logppn q and pn ě 2. And, again using Bernstein’s inequality for
1
U-statistics, and using the fact that |gpX, X 1 q| ď C0 kn always, with probability at least 1 ´ 72p
,
n
ˇ
ˇ
ˇ 1 ÿ
ˇ
ˇ` ˘
1 ˇ
gpXi , Xi1 q ´ ErgpX, X qsˇ ď
ˇ n
ˇ 2 iăi1
ˇ
d
c
2C02 kn2 logp2 ¨ 72pn q 2C0 kn logp2 ¨ 72pn q
kn2 logppn q
`
ď
¨ C2 ,
pn{2q
3pn{2q
n
where
2
C “
d
2C02 log2 p144q 2C0 log2 p144q
`
,
p1{2q
3{2
and we use the assumption n ě kn2 logppn q and pn ě 2. Therefore,
ˇ˜
ˇ ˇ
ˇ2
¸2
ˇ
ˇ ˇ
ˇ
ˇ 1 ÿ
ˇ ˇ 1 ÿ
ˇ
1
2
1
ˇ ` ˘
ˇ
`
˘
1
1
gpX
,
X
q
´
ErgpX,
X
qs
ď
gpX
,
X
q
´
ErgpX,
X
qs
ˇ
ˇ `
i
i
i
i
ˇ n
ˇ ˇ n
ˇ
ˇ 2 iăi1
ˇ
2 iăi1
ˇ
ˇ
c
ˇ 1 ÿ
ˇ
kn2 logppn q
ˇ` ˘
1 ˇ
1
3
1
gpXi , Xi q ´ ErgpX, X qsˇ ¨ 2|ErgpX, X qs| ď C
,
ˇ n
ˇ
ˇ 2 iăi1
n
where we set
C 3 “ C 22 ` 2C 2 ¨ Cmoment
1{3
and again use n ě kn2 logppn q, and apply Lemma D.1 to bound |E rgpX, X 1 qs |. Combining everything,
1
this proves that, with probability at least 1 ´ 36p
,
n
c
c
ˇ
ˇ
kn2 logppn q
C02 kn2
kn2 logppn q
ˇ Jp
ˇ
J
1
3
¨C `
`C
.
ˇx Σh1 x ´ x Σh1 xˇ ď
n
n´1
n
63
Setting
Cstudentized “ C 1 ` C 3 ` 2C02
and using the fact that n ě 2 and n ě kn2 logppn q, we have
c
ˇ
ˇ
kn2 logppn q
ˇ
ˇ Jp
J
.
ˇx Σh1 x ´ x Σh1 xˇ ď Cstudentized
n
Lemma F.8. With probability at least 1 ´
1
9pn ,
p h ´ Σh ||8 ď 100
||Σ
1
1
c
logppn q
.
n
Proof of Lemma F.8. From our definitions, we see that
“
‰
Σh1 “ Varph1 pXqq “ E h1 pXqh1 pXqJ ´ Erh1 pXqsErh1 pXqsJ ,
and
˜
¸˜
¸J
ÿ
ÿ
ÿ
1
1
1
J
p
p
p
ph “
.
h1 pXi qp
h1 pXi q ´
h1 pXi q
h1 pXi q
Σ
1
n i
n i
n i
“
‰
ř
First, we bound || n1 i p
h1 pXi qp
h1 pXi qJ ´ E h1 pXqh1 pXqJ ||8 . We have
›
›
›1 ÿ
“
‰››
›
J
J
p
h pX qp
h pX q ´ E h1 pXqh1 pXq › ď
›
›n i 1 i 1 i
›
8
›
›
›
›1 ÿ
1ÿ
›
J
J›
p
p
h1 pXi qh1 pXi q ´
h1 pXi qh1 pXi q › `
›
›
›n i
n i
8
›
›
›1 ÿ
›
“
‰
›
›
h1 pXi qh1 pXi qJ ´ E h1 pXqh1 pXqJ ›
›
›n i
›
. (F.20)
8
ř
ř
h1 pXi qp
h1 pXi qJ ´ n1 i h1 pXi qh1 pXi qJ ||8 .
We handle these two terms separately. First, we bound || n1 i p
ř
ř
For convenience we define A :“ n1 i p
h1 pXi qp
h1 pXi qJ and B :“ n1 i h1 pXi qh1 pXi qJ . Since A and B
are both positive semidefinite matrices with ones on the diagonal, we have
||A ´ B||8 “
2
ˇ J
ˇ
1
max2 ˇfjk
pA ´ Bqfjk ˇ ,
2 j,kPrpn s
(F.21)
where fjk P Rpn is the vector with pfjk qj “ 1, pfjk qk “ ´1, and zeros elsewhere. Next we have
b
b
ˇ ´b
¯
ˇ J
ˇ ˇb
J Af
J Bf
ˇfjk pA ´ Bqfjk ˇ “ ˇˇ f J Afjk ´ f J Bfjk ˇˇ ¨
f
`
f
jk
jk
jk
jk
jk
jk
ˇ
ˇ
ˇ
ˇ
d
d
b
ˇb
ˇ
ÿ
ÿ
ˇ
ˇ
4
ˇ
ˇ
J
J
J
2
J
2
p
ˇ
ď 4 ˇ fjk Afjk ´ fjk Bfjk ˇ “ ? ˇ
ph1 pXi q fjk q ´
ph1 pXi q fjk q ˇˇ
nˇ i
ˇ
i
d
¯2
ÿ´
4
ď?
pp
h1 pXi q ´ h1 pXi qqJ fjk ,
n
i
64
where the first inequality follows from the fact that ||fjk ||1 ď 2 while ||A||8 , ||B||8 ď 1, and the
second inequality follows from the triangle inequality. Next, for each i and each j, k, observe that
p
h1 pXi qJ fjk “
1 ÿ
J
psignpXi ´ Xi1 q b signpXi ´ Xi1 qq fjk ,
n ´ 1 i1 ‰i
which after conditioning on Xi , is a mean of pn ´ 1q i.i.d. variables, each taking values in r´2, 2s since
||fjk ||1 ď 2. Furthermore, conditioning on Xi , we have Erp
h1 pXi qs “ h1 pXi q. Therefore, applying
Hoeffding’s lemma [see, for example, Lemma 2.6 in Massart, 2007], for each i, j, k, for any t P R,
" 2 *
”
!
)ı
2t
.
(F.22)
E exp t ¨ pp
h1 pXi q ´ h1 pXi qqJ fjk ď exp
n´1
Applying Lemma F.10 (stated below), then,
#
+
¯2
1 ÿ´ p
80
1
J
5
P
.
ph1 pXi q ´ h1 pXi qq fjk ą
¨ p1 ` logp27pn qq ď
n i
n´1
27p5n
Taking a union bound over all j, k P rp2n s, and returning to (F.21), we then have
›
#›
+
c
›1 ÿ
›
ÿ
80
1
1
›
›
J
J
p
P ›
h1 pXi qp
h1 pXi q ´
¨ p1 ` logp27p5n qq ď
h1 pXi qh1 pXi q › ą 2
.
›n i
›
n i
n´1
27pn
8
Next we turn to the second term in (F.20). Since ||h1 pXq||8 ď 1 always, we see that for each
j, k P rpn s,
¸
˜
1ÿ
J
h1 pXi qh1 pXi q
n i
jk
is a mean of n i.i.d. terms, each taking values in r´1, 1s. Applying Hoeffding’s inequality, for each j, k,
,
$ˇ˜
¸ ˇˇ
.
&ˇˇ 1 ÿ
ˇ
ˇ ě t ď 2e´nt2 {2
P ˇˇ
h1 pXi qh1 pXi qJ ´ ErhpXqhpXqJ s
ˇ
%ˇ n i
ˇ
jk
for any t ě 0. Setting t “
b
2 logp54p3n q
,
n
and taking a union bound, we see that
›
#›
›
›1 ÿ
›
J
J ›
h1 pXi qh1 pXi q ´ ErhpXqhpXq s›
P ›
›
›n i
8
ě
c
2 logp54p3n q
n
+
ď
Returning to (F.20), then, with probability at least 1 ´
2p2n
´n
¨e
ˆb
54 logpp3
nq
n
˙2
{2
“
1
.
27pn
2
27pn ,
c
c
“
‰
1 ÿp
80
2 logp54p3n q
J
J
3
p
||
h1 pXi qh1 pXi q ´ E h1 pXqh1 pXq ||8 ď 2
¨ p1 ` logp27pn qq `
. (F.23)
n i
n´1
n
This proves a bound on (F.20).
Next, to complete the proof, we bound
›˜
›
¸˜
¸J
›
›
ÿ
› 1ÿ
1
J›
p
p
›
›
h
pX
q
h
pX
q
´
E
rh
pXqs
E
rh
pXqs
1
i
1
i
1
1
› n
›
n i
›
›
i
8
65
.
We have
˜
¸˜
¸J
1 ÿp
1 ÿp
J
h1 pXi q
h1 pXi q
´ E rh1 pXqs E rh1 pXqs
n i
n i
˜
¸˜
¸J ˜
¸
1 ÿp
1 ÿp
1 ÿp
J
“
h1 pXi q
h1 pXi q ´ E rh1 pXqs
´
h1 pXi q ´ E rh1 pXqs E rh1 pXqs
n i
n i
n i
ř
h1 pXi q||8 , ||E rh1 pXqs ||8 ď 1, we therefore have
and, since || n1 i p
›˜
›
ˇ
›
¸˜
¸J
›
›
ˇ 1ÿ
›
ÿ
› 1ÿ
1
ˇ
›
J›
p
p
p
›
›
|}
ď
2||
h
pX
q
h
pX
q
´
E
rh
pXqs
E
rh
pXqs
h
pX
q
´
E
rh
pXqs
ˇ
› .
1
i
1
i
1
1
1
i
1
› n
›
ˇ
›
n
n
›
›
i
i
i
8
8
2
For each sign s P t˘1u, for each j P rp2n s, writing ej to denote the jth basis vector in Rpn , we have
«
#
E exp t ¨ s ¨
eJ
j
˜
¸+ff
1 ÿp
h1 pXi q ´ E rh1 pXqs
n i
*
"
!
´
¯)ı
1ÿ ”
t2
J p
ď
,
E exp t ¨ s ¨ ej h1 pXi q ´ E rh1 pXqs
ď exp
n i
2pn ´ 1q
where the first inequality follows from the convexity of x ÞÑ ex , while the second applies Hoeffding’s
lemma, as in (F.22) above. Then,
#
+
´
¯ c 2 logp54p3 q
1
n
J p
ď
,
P s ¨ ej h1 pXi q ´ E rh1 pXqs ą
n´1
54p3
and therefore taking a union bound over each s P t˘1u and each j P rp2n s,
›
#›
+
c
›1 ÿ
›
3q
2
logp54p
1
›
›
n
p
P ›
h pX q ´ E rh1 pXqs› ą
ď
.
›n i 1 i
›
n´1
27pn
8
Therefore, combining this with (F.23), with probability at least 1 ´
p h ´ Σh ||8 ď 2
||Σ
1
1
c
80
¨ p1 ` logp27p3n qq `
n´1
c
1
9pn ,
c
c
2 logp54p3n q
2 logp54p3n q
logppn q
`2
ď 100
,
n
n´1
n
where the last step uses the fact that n, pn ě 2.
2
Lemma F.9. Let Σh1 be defined as in Assumption 3.4. For every z P Rpn ,
z J Σh1 z ď λmax pΣq ¨ ||matpzq||2`1 {`2 ,
where ||matpzq||`1 {`2 is defined as in the statement of Lemma F.6.
Proof of Lemma F.9. Since the statement is deterministic, we can treat M “ matpzq P Rpn ˆpn as fixed.
66
Then z “ vecpM q and
`
˘
vecpM qJ Σh1 vecpM q “ Var vecpM qJ h1 pXq
`
˘
“ Var vecpM qJ ErhpX, X 1 q | Xs
`
˘
“ Var ErvecpM qJ hpX, X 1 q | Xs
`
˘
pby the law of total varianceq
ď Var vecpM qJ hpX, X 1 q
‰
“
J
1 2
ď E pvecpM q hpX, X qq
“
`
˘ ‰
“ E pvecpM qJ signpX ´ X 1 q b signpX ´ X 1 q q2
”`
˘2 ı
“ E signpX ´ X 1 qJ M signpX ´ X 1 q
»˜
¸2 fi
ÿ
“ E–
signpX ´ X 1 qJ Mj ¨ signpXj ´ Xj1 q fl pwhere Mj is the jth column of M q
»˜
j
¸2 fi
ÿˇ
ˇ
ˇsignpX ´ X 1 qJ Mj ˇ fl
ď E–
“
ď
“
“
ď
ÿ
jk
ÿ
jk
c ”
ı c ”
ı
2
2
1
J
E |signpX ´ X q Mj | ¨ E |signpX ´ X 1 qJ Mk |
ÿb
jk
ÿb
jk
ÿb
jk
j
ˇ ˇ
ˇ‰
“ˇ
E ˇsignpX ´ X 1 qJ Mj ˇ ¨ ˇsignpX ´ X 1 qJ Mk ˇ
MjJ E rsignpX ´ X 1 q signpX ´ X 1 qJ s Mj ¨
MjJ T Mj ¨
b
MkJ T Mk
||Mj ||22 ¨ λmax pT q ¨
“ λmax pT q ¨
˜
ÿ
j
b
MkJ E rsignpX ´ X 1 q signpX ´ X 1 qJ s Mk
||Mj ||2
b
||Mk ||22 ¨ λmax pT q
¸2
.
Finally, by Wegkamp and Zhao [2013, Theorem 2.3], λmax pT q ď λmax pΣq.
Lemma F.10. Let v P Rp be a fixed vector and let Z1 , . . . , Zn P r´1, 1sp be random vectors, not
necessarily independent, such that v J pZi ´ E rZi sq is C-subgaussian for each i, that is,
Erexpttv J pZi ´ E rZi squs ď exppCt2 {2q.
Then for any δ P p0, 1q, with probability at least 1 ´ δ,
˘2
1 ÿ` J
v pZi ´ ErZi sq ď 20Cp1 ` logp1{δqq .
n i
Proof of Lemma F.10. For each i, by assumption,
„
"
*
" 2*
1
t
E exp t ¨ ? v J pZi ´ ErZi sq
ď exp
.
2
C
By Vershynin [2012, Lemma 5.5] (and tracking constants carefully in this Lemma), for each i,
„
"
*
˘2
1 ` J
E exp
¨ v pZi ´ ErZi sq
ďe.
20C
67
By the convexity of x ÞÑ ex , then,
«
#
+ff
„
"
*
˘2
˘2
1
1 ÿ` J
1ÿ
1 ` J
v pZi ´ ErZi sq
E exp
¨
ď
E exp
¨ v pZi ´ ErZi sq
ďe.
20C n i
n i
20C
Therefore, we have
P
#
˘2
1 ÿ` J
v pZi ´ ErZi sq ą t
n i
+
«
ď E exp
#
˘2
1 1 ÿ` J
1
v pZi ´ ErZi sq ´
t
20C n i
20C
+ff
*
"
1
t .
ď exp 1 ´
20C
Setting t “ 20Cp1 ` logp1{δqq, then, we have proved the desired result.
68
| 10math.ST
|
A Game-Theoretic Model of Demand Response
Aggregator Competition for Selling Stored Energy
in Regulated and Unregulated Power Markets
Mahdi Motalleb, Reza Ghorbani1
Abstract
This work is concerned with the application of game-theoretic principles to model competition between demand response
aggregators for selling excess energy stored in electrochemical storage devices directly to other aggregators in a power
market. This market framework is presented as an alternative to the traditional vertically-integrated market structure,
which may be better suited for developing demand response and smart-grid technologies, in addition to increasing
penetration of independent renewable energy generation devices. Demand for power generated by the utility through
combustion of fuel could be replaced, lowering emission of pollutants, when the energy used to charge the batteries is
produced sustainably and traded on smaller scales. The four variants of game are considered: both non-cooperative
(unregulated competition) and Stackelberg (regulations on transaction price and size), each with and without DR
scheduling. The Nash equilibrium is derived for each game variant in order to serve as a bid-price decision making criteria
which determines the optimal bidding strategy for an aggregator to sell in the market. The model is applied to a case study
involving completion for selling between two aggregators. Bidding strategy is dependent on parameters inherent to an
aggregator’s energy storage hardware, and the strategy selected by each aggregator does not vary with the variations in
the game conditions considered. Demand response scheduling offers greater payoff for aggregators who implement it,
compared with those who do not. Addition of transaction price and volume regulations to the market do not affect the
participants optimal bidding strategies (the Nash equilibrium), but lowers payoffs for all aggregators participating in the
market relative to unregulated competition.
Index Terms: Demand Response Aggregator (DRA), Game theory, Non-cooperative game, Stackelberg game, Demand
response scheduling, Nash equilibrium
1. Introduction
Development and implementation of smart grid technologies offers substantial advantages over traditional
vertically integrated electric utilities. Both parties stand to benefit from proliferation of these technologies. The endusers of electrical power primarily wish to minimize the cost they pay to the utility for the energy they require to
power their devices. Opposed to this, the utility company wishes to maximize its net profit and is concerned with load
scheduling as they must remain able to provide adequate supply. In this paper, a market framework has been developed
based on game-theoretical concepts where Demand Response Aggregators (DRAs) compete with each other to sell
energy stored in consumer’s battery systems. Our model determines the optimal bidding decision for each DRA to
maximize its own pay off under the conditions of an incomplete information game.
Game theory has been used in power system markets to interpret a participant’s behavior in deregulated
environments and to allocate costs among pool participants. Two different hybrid algorithms were presented [1, 2, 3]
for the Generation Expansion Planning (GEP) problem for a pool-based electric market where the modified-gametheoretic algorithms were divided into two programing levels: master and slave. A static computational game theoretic
model has also been developed [4] to investigate the impacts of competition on the wholesale price of electricity, the
1
Mahdi Motalleb and Reza Ghorbani, are with Renewable Energy Design Laboratory (REDLab) at the Department of Mechanical Engineering,
University of Hawai'i Manoa, Honolulu, Hawai'i, 96822. Phone: +1-808-956-2292, Fax: +1808-956-0767. E-mail: [email protected],
[email protected].
demand for electricity, the profits of firms, and levels of various polluting emissions. A Medium Run Electricity
Market Simulator (MREMS) based on game theory was presented in [5]. This simulator incorporated two different
games, one for the unit commitment of thermal units and one for strategic bidding and hourly market clearing. The
most common electricity bidding mechanisms in electricity auction markets were analyzed using signaling game
theory [6] and also a Swarm platform was used to develop a simulation model based for multiple agents. The role of
sustainable energy volatility was investigated in the context of market participant’s competitive expansion planning
problem in [7]. An incomplete information non-cooperative game-theoretic method where each generation company
(GENCO) perceives strategies of other market participants was applied to make decisions about strategic generation
capacity expansion. An agent-based game-theoretic formulation was presented in [8] to investigate the learning speed
of traders and their strategic collaboration in dynamic electricity markets. It was shown that the learning speed of
traders decreases during large fluctuations in the power exchange market.
The Authors of [9] expanded the application of an Equilibrium Problem with Equilibrium Constraints (EPEC)
model, typically used with multi-leader-follower games, to a spatial market. In that model, the prevalent market setting
was investigated in the international market for metallurgical coal between 2008 and 2010, whose market
characteristics provide arguments for a wide variety of alternate market structures. A two-sided agent-based
framework was proposed in [10] employing a game-theoretic/Particle Swarm Optimization (PSO) hybrid simulation
approach to provide a suitable platform for sustainable GEP. A Bayesian game model analyzing cross-border transfer
of electric power was presented in [11] to explain cross-border transfer within a market containing players controlling
unreliable marginal generators. A dynamic game-theoretic model was developed in [12] to analyze the impacts of
natural gas market reformation in promoting Natural Gas-Fired Electricity (NGFE) generation, in which hourly RealTime Pricing (RTP) was applied in both natural gas and electricity markets [53]. An innovative game theoretic
framework was proposed in [13] for a next-generation retail electricity market (“Energy Internet’’) with high
penetration of distributed residential electricity suppliers (‘‘Energy Cells’’). The envisioned Energy Internet proposes
a large number of distributed Renewable energy generation and energy storage systems connected to the grid through
plug-and-play interfacing devices. A game-theoretic framework for economic operations of future residential
distribution systems was presented in [14] for cases involving extensive participation of distributed electricity
prosumers which defined novel roles for both the utilities and “customers”. The game theoretic algorithms used to
determine the retail electricity market clearing price consider group coalition scenarios of multiple electricity
prosumers.
Game theory is a useful mathematical tool to handle problems related to demand side management (DSM) [15,
48-51]. Several game-theoretical demand response (DR) programs have been proposed with differing objectives such
as: determining the optimal hourly incentive to be offered to customers who sign up for load curtailment [16, 17, 52],
scheduling load usage by creating several possible tariffs for consumers such that net demand remains below some
threshold [18], adjusting demand to meet supply, as well as smoothing the aggregated load in the system [19], and
evaluating the impact of the response capability of smart-home consumers on promoting further distributed PV
penetration [20]. Game theory was used in [21] to present an optimization model to minimize load curtailment needed
to restore equilibrium to the operating point when the system is in a fault condition (e.g., loss of generation). An
optimal time-of-use pricing with an evolutionary game-theoretic perspective was proposed in [22] for urban gas
markets where a power structure demand response program was employed to simulate user demand response.
Decision-making (DM) processes are the principle product of game theory. Basically, Game theory is the formal
study of DM under competitive conditions where choices potentially affect the interests of the other players. In [23] a
DM approach for Liquefied Natural Gas (LNG) projects was presented. The approach was based on a consensus
algorithm addressing the consensus output over a common value using cost functions within a framework based on
game theory. In another work [24], a DM process for conceptual planning and project evaluation in the oil and gas
industry was presented where the set of strategic decisions is generated by a binary genetic algorithm. In [25] , DM
processes of industrial and environmental concerns was evaluated with a game theoretic approach where Industry and
Environment were considered as two players with conflicting interest to find optimal strategies in governing energy
policy. A bi-level, complete-information, matrix game-theoretic model was proposed in [26] to assess the economic
impact and make operational decisions in carbon- constrained restructured electricity markets. A game theoretic
modeling approach was performed in [27] to develop financial transmission rights bidding strategies for power
suppliers assuming that they have adequately forecast Locational Marginal Prices (LMPs). The game theoretic model
considered multiple participants as well as network contingencies. An evolutionary imperfect-information game
approach was proposed in [28] to analyze bidding strategies in electricity markets with price-elastic demand. The
research work presented in [29] characterized the impact of Long-term plans on short-term maintenance decisions of
GENCOs by applying the Cournot model, which has been used for strategic generation dispatch of generating units
in electricity markets.
In order to characterize the DM process, Nash equilibrium is used. A set of strategies is a Nash equilibrium if no
player can do better by unilaterally changing his or her strategy [30]. The authors of [31] suggested to use a theory of
evolutionary games and the concept of a “near Nash equilibrium” to simulate the electricity market in the presence of
more than two producers. In [32], a novel game-theoretic approach was presented to the Generation Maintenance
Scheduling (GMS) problem in electricity markets where a coordination procedure for an Independent Service Operator
(ISO) was modeled and the GENCOs GMS process was modeled as a non-cooperative dynamic game, and the
GENCO’s optimal strategy profile was determined by the Nash equilibrium of the game. In [33], the conventional
model of a dynamic Cournot electricity market game (where GENCOs are assumed to hold a uniform and accurate
belief concerning market dynamics) was replaced by a more realistic model with subjective demand errors coining a
new equilibrium concept termed Subjective Equilibrium (SE). System performance, equilibrium output, profit and
customer surplus, at SE were analyzed. The results of [33] suggested that the system equilibriums are strongly
influenced by the GENCOs’ knowledge about market demand.
In game theory, different types of games are utilized for analysis of different types of problems. The different
types of games are categorized by number of players involved, symmetry of the game, and whether or not, cooperation
among players is allowed. In the literature for power markets, the different game models used include: cooperative
[34, 35, 36, 37], non-cooperative [20, 38], Stackelberg [9, 39], multi-leader-follower [9], Forchheimer (one leader)
[5], and Bertrand games (all players are leaders) [5]. Beside the application in power markets, game theory has been
used in diverse, and related fields such as: analysis of Electric Vehicle (EV) charging station construction [40],
charging method for plugged in hybrid EVs [41], analysis of power grid vulnerability [42], performance evaluation of
thermal power plants [43], analysis of effects of higher domestic gas prices in Russia on the European gas market
[44], and interactive energy management of networked microgrid-based active distribution systems [45].
In the present work, a demand response market framework has been developed based on game-theoretic
considerations. In this market, the commodity is energy stored in consumers’ batteries. End users of electricity charge
their batteries when the electricity is cheap and sell it back to the grid (or other consumers) when the electricity is
more expensive. There are four basic concepts in game theory including: players, nodes, moves, and payoffs. In the
proposed market framework, the players are DRAs and each node belongs to one DRA. The possible moves of the
game are the player's decision–i.e.: any of the options chosen by a given DRA. A player's strategy will determine the
action the player will take at each stage of the game. In the proposed market structure, strategies are different biddings
of DRAs. The payoff to each player (DRA) is the difference between monetary gain upon selling its stored energy and
the cost of storing that energy. A comprehensive quadratic cost function has been proposed in this study for
discharging the energy stored in batteries. Also, it has been assumed that players are privy to no information about
other player’s moves. Therefore, our problem is cast as a real, incomplete information game (i-game). Two different
types of game are considered here: non-cooperative, and Stackelberg. In the non-cooperative game, DRAs compete
to sell the energies stored in the batteries without any limitation or constraints in transaction size or price; but in the
Stackelberg game, the utility (as game leader) controls the DR market through restricting both. In each game, two
different DR programs are considered: non-scheduled demand and price-sensitive scheduling of water heaters (or all
thermostatic devices). Dynamic Programing (DP) has been used for price-sensitive scheduling.
Section 2 explains some principles of game theory’s application to energy markets. In section 3, a cost function
and a payoff function have been developed in order to utilize in the proposed method. Section 4 describes the proposed
game-theoretic market framework including the alternate game types and DR schedules considered. The presented
market framework has been applied to a case study and its dissection and results have been provided in section 5.
Finally, conclusions and possible future works are presented in section 6.
2.
Principles and definitions of game theory in power market
First, let us briefly present the fundamental principles and definitions of the game-theoretic concepts we need to
analyze a power market. The following subsections cover: a proposed function we will use to model the aggregator’s
cost, the bidding process of DRAs, and the payoff function for DRAs.
2.1. Game theory in power market
In game theory, beliefs are formulated against risky alternatives in order to maximize the expected revenue
(payoff function) for aggregators. In the real competition, each player (DRA) remains unaware of the detailed data of
other players (for instance data of strategies and payoffs). Probability theory gives an expected value of the payoff
given the probability of events and Bay’s law is used to revise beliefs given new information. The objective of the
game-theoretic analysis is to find the Nash equilibrium – the best strategy from the set of possibilities in the sense that
no incentive exist to alter the equilibrium strategy despite the strategies chosen by the other players. Fig. 1 illustrates
a power market where some of the aggregators acting as buyers and others acting as sellers submit their bids to a
center (ISO) and this center sets a market spot price between aggregators.
Fig. 1: General structure of a power market including buyer DRAs, seller DRAs, and ISO
DRAs participate in the market to sell their available stored energy which is not consumed in powering local loads. In
a power transaction game, aggregators’ transactions are modeled as a game of strategies to maximize their payoffs.
Two types of games are considered in this research: non-cooperative and Stackelberg. A non-cooperative game takes
place when each player (DRA) is interested to maximize its own payoff and there is no cooperation between
aggregators to coordinate their strategies. In a non-cooperative game, some aggregators choose a strategy while other
aggregators try to identify their best response to that strategy. In a Stackelberg game, there is one leader (utility) to
limit the market price and power transferred to ensure all policies are respected in the market.
2.2. Aggregator cost function
This subsection shows how DRA’s cost function can be presented with a quadratic function. Since the only energy
resources available for sale by a given DRA is the the aggregation of stored energy in residential batteries, the
aggregator cost function is the battery cost function (cost of discharging the battery). For the battery, the function we
chose is a variant of the widely used logarithmic barrier function, used as a penalty function in interior point methods
[46]:
C hbat
E hi a hi log(1
i
E hi
B
)
bat
(1)
where C hi is cost of the battery in house i as a function of stored energy,
the utility to give higher prices during peak-hours,
ahi is a pricing coefficient determined by
Ehi is the total load (kWh), and B is a parameter that has been
used to give cost values very close to the values given by a quadratic one, it also serves as the maximum typical value
for
Ehi . The relation between the presented cost function (Eq. 1) and the quadratic one can be understood from its
Taylor expansion. Since Ehi
Cibat Ehi ahi log(1
ahi (
Ehi
B
Ehi
B 1 , the Taylor series expansion is:
Ehi
2
)
B2
)
B
ahi Ei
B
ahi Ei
2
(2)
B2
Updates in load data resolve every 15 minutes from measurement devices. We assume that the load is constant for
each 15 minutes between updates and the quadratic cost function of the battery can be rewritten as a function of power
( Phi ):
C
bat
hi
P
ahi (0.25Phi )
hi
B
In order to obtain the coefficients
ahi (0.25Phi ) 2
B2
(3)
ahi and B, two factors should be considered: the electricity price of the grid (for
charging the battery) and the capital and maintenance costs of the battery. Thus, the cost of the battery of house i for
charging the power by an amount
Phi is:
(4)
where C
P is the cost which house i pays to the utility to sell its electricity and to charge its own battery as
amount power of P and C
P is the contribution of capital and maintenance costs of the battery of house
Chbat
Phi Chgrid
Phi ChCi &M Phi
i
i
grid
hi
hi
C &M
hi
hi
i for power of
hi
Phi . The coefficients ahi and B can be obtained using Eq. 4. Since the individual house cost functions
are quadratic in form, aggregation of the houses will also produce quadratic cost functions for each DRA:
Cg Pg vg Pg ug Pg
where
and
2
(5)
Cg is the cost function of aggregator g for selling the stored power of Pg (in kW), and the coefficients of u g
vg are:
vg
0.25ag
(6)
B
(7)
a (0.25) 2
ug g 2
B
1 Nh
where ag
ah , and Nh is the number of houses under aggregator g.
N h i1 i
2.3. Bidding process of DRAs
This section explains how DRAs bid their price for selling the stored energy in batteries, based on the cost function
defined in section 2.2. From Eq. 5, the marginal cost of aggregator i is:
i
dC i
vi 2u i Pi
dPi
(8)
Since marginal cost is a linear function of stored power in batteries, we assume that aggregators’ bids are also linear
functions of stored power:
i 0i mi Pi
(9)
where i Is the bid price per unit of power at discharging power level Pi , 0i is the marginal cost of electricity at P0i
, and
mi is the slope of the bid curve. The market coordinator (ISO) receives sale bids from the DRAs given by Eq.
9 and matches the lowest bid with the buyer aggregators. In Fig. 2, the aggregator intends to increase the battery
discharging beyond
P0i if the spot market price is greater than 0i .Here, Ti is the net power interchanged; if Ti is
positive (negative), the aggregator is selling power to (buying power from) the other aggregators,
P0i Ti Pmax,i P0i . The maximum stored power in the aggregated storages of DRA i is Pmax,i , the minimum
of the local load of DRA i which should be fed by local batteries is
P0i (can be established as one of the governing
policies of aggregator i).
max,i
λi
λ0,i
λmin,i
slope: mi
Buy
Sell
P0,i
Ti
λ Pi
Pmax,i
Fig. 2: Aggregator g’s biding curve
2.4. Aggregator’s payoff function
The mathematical formulation for each DRA’s payoff function in the proposed game-theoretic framework is
described below. For a given spot market price,
Ri Ci Ti
, the ith DRAs payoff,
Ri , is:
(10)
where Ci is the difference between the discharging cost of the batteries in aggregator before and after power
transactions, and
Ti is the net power transacted.
Let T denote a power transaction; associated with each T is a pair of aggregator payoffs
( , ) for participants
[, ( , )] characterized by a region, ,
and a payoff point ( , ) in . Without any trading, the payoff is (0,0) ; a trade would take place if and only if
1 and 2. The Nash bargaining problem for two aggregators is denoted by
both aggregators agree upon a solution in which represents an acceptable deal for the participants involved. The
problem is formulated as follows: given
( 0 , 0 ) (0,0) in as the initial solution, find a different solution,
( 1 , 1 ) , that satisfies the conditions:
( 1 , 1 ) is a point of .
11 00
( 0 , 0 ) & 1 0, 1 0
The point ( 1 , 1 ) is the Nash equilibrium solution to the bargaining game [, ( , )] .
From the definition of the payoff function for DRA g in eq. 10, the objective function is:
maximize
R ( p ,T ) ,
kK pij ,Tij
k
ij
ij
ij
where
Rijk ( pij , Tij ) ijTij* [Ci ( Pgi Tij ) Ci ( Pgi )]
(11)
k
ij
R is the payoff of DRA i (as a seller) after transmitting the power of Tij to aggregator j (and receiving the power of
Tij* by aggregator j) in contract k. The difference between Tij and Tij* is due to transmission losses. With negligible
transmission losses Tij Tij . The parameter,
*
ij , is the transaction price and Pgi is the amount of the local loads of
the aggregator i which is fed by the local storages before power transaction. The first term in eq. 11 is the gross revenue
of the aggregator due to the transaction and the second bracketed term is the change in the aggregator’s batteries
discharging cost owing to the transaction. In the case of two aggregators shown in Fig.3 (aggregator 1 as seller and 2
as buyer), the objective function is:
maximize
{Rseller Rbuyer} ,
where
Rseller T T12* [C1 ( Pg1 T12 ) C1 ( Pg1 )]
(12)
Rbuyer T T12* [C2 ( Pg 2 T12* ) C2 ( Pg 2 )]
(13)
and
T is the tranction price and Rseller and Rbuyer are the aggregators payoffs. Each aggregator has a constrained
storage capacity, and payoffs are constrained by
Rseller 0 , and Rbuyer 0 , indicating that negative payoffs are
excluded. When there are no transactions, the payoff for each aggregator is zero.
Fig. 3: A system including two DRAs as seller and buyer
In real competition between aggregators, participants don’t have information about the other participants
such as cost function coefficients, availability of the batteries, and biding prices. The following section shows the
proposed game-theoretic market framework for competition between DRAs to sell energy stored in batteries in an
incomplete information game.
3.
Proposed game-theoretic market framework for DRAs
The proposed methodology is explained in this section. The first subsection includes details about the incomplete
information game to make a market competition among DRAs. The proposed market framework based on game theory
has been presented in the second subsection. All mathematical and probabilistic details for the developed market
framework (including two different game types each with and two demand scheduling programs) have been provided
in second subsection.
Fig. 4 shows a visual schematic of the proposed market framework. Each DRA manages a number of buildings
with control signals sent through a wireless network. Each house contains various devices possibly including Water
Heaters (WHs), Batteries, Air Conditioners (ACs), and Electric Vehicles (EVs). The DRAs decide how to bid for
buying and selling power in the market. A DP scheduler outputs signals for deciding when a WH should run-on/off
setting as a function of time-to satisfy the conditions in the game (social welfare in all games, and price-sensitive DR
WH scheduling in the games where it is implemented) given as inputs hot water usage data (in all games) and possibly
electricity price signals from the utility (only for the games with price-sensitive DR WH scheduling). The DRAs
communicate through a wireless network with servers which communicate with the market directly to arrange
transactions. In the Stackelberg games, the servers also mitigate size and price of transactions in accordance with
utility policy constraints which are transmitted over the wireless network.
Fig. 4: General schematic of the proposed market framework in two types of games and two DR scheduling program
3.1. Market competition between DRAs with incomplete information
In this section, a market model is proposed to model competition among aggregators in an electricity market
where participants have incomplete information. Based on eq. 10, each aggregator’s payoff is a function of storages
discharging costs, power transactions, and the spot market price. Hence, an individual aggregator’s payoff is a function
of bids offered by other aggregators. Each participant estimates the other participants’ bids in order to maximize its
own payoff. Each aggregator has complete information concerning its own payoff but lacks information critical to
predicting others actions precisely, this competition is known as an incomplete game (i-game). Here, an aggregator’s
unknown characteristics are modeled by classification of participant’s types. The DRA’s type contains information
concerning: its own payoff function, other aggregators’ pay off functions, grid electricity prices, availability of charged
batteries, etc.. An aggregator’s type corresponds to its battery discharging cost structure, that is, coefficients
ai and
B (Eq. 5 and 6). Each aggregator would have full knowledge of its own costs, but only an estimate of the remaining
aggregator’ costs. Each participant adjusts the slope of m in its bidding curve (Fig. 2) in order to maximize its payoff,
Ri , in Eq. 10. The choice of m corresponds with the bidding strategy (bid low, bid high, bid marginal cost) in the
game. The main goal of the game is to determine the optimal DM process used to determine what value of m to choose
in order to maximize payoff despite the unknown parameters of the other participants.
The DRAs use a Bayesian approach to deal with incomplete information. In this approach, a probability
distribution,
, represents the unknown parameters. The expected value of the payoffs of the aggregators are
maximized for
. In estimating the probability distribution, each aggregator uses only the information common to
all aggregators.
Consider a scenario described from DRA B’s perspective for competing with DRA A for selling the stored energy
of the batteries to aggregator C depicted in Fig. 5, below. We assume that the participants’ types are drawn at random
from hypothetical populations A and B containing types
t Am
and t B , respectively where m 1,..., M and
n
n 1,..., N index the possible types of DRA A and B, respectively. For instance, to model the uncertainty in A’s
discharging cost, B would assume that there are M possible types of A,
t Am
, each with its corresponding cost.
Generally, aggregator B knows n but does not know m, and hence does not know its opponent’s type (i.e. it does not
know A’s discharging costs). Each DRA estimates the probability distributions ,
, for the random variable of the
“type” of its competitors based on freely available published information such as electricity prices, demand curves,
water usage (from water heaters), availability of storages, and aggregators’ parameters.
Fig. 5: A grid including three DRAs: A and B as sellers and C as buyer
3.2. Game-theoretic market framework
For our purposes, the goal of the game is to calculate the expected payoffs for different strategies for aggregators
A and B (in the mentioned scenario) to find the Nash equilibrium and to choose the strategy that maximizes payoff:
N
EPAm Am ( n ).H Am ( s Am , s Bn , m, n )
(14)
n 1
EPAm is the expected payoff function of aggregator A if it is of type m M . n is the type of opponent
m
aggregator B ( n N ). The conditional probability A (n) is the probability DRA A is of type m for DRA B of type
where
n:
Am (n) prob(t Bn | t Am )
mn
N
n 1
where
(15)
mn
prob shows the conditional probability function, and mn is the basic probability distribution
corresponds to the probability that A is type m and B is type n. A DRA’s strategy (for bidding) is determined by its
sAm is a vector of strategies for A’s type m. For example, a vector of strategies may be: bid high, bid at marginal
m
m
cost, and bid low. H A is the conditional payoff of DRA A which depends on the strategies of DRA A, sA , and DRA
type.
B,
sBn . Aggregator A needs to know the opponent’s type to maximize H Am , and since the information is not available
in an i-game, aggregator A maximizes the expected value of the payoff
EPAm .
Similarly, for aggregator B, the expected payoff function is:
M
EPBn Bn ( m ).H Bn ( s Bn , s Am , n, m )
(16)
m 1
where
Bn (m) prob(t Am | t Bn )
mn
M
m 1
(17)
mn
3.2.1. Non-cooperative game
The i-game considered previous converts to a (M + N)-participant game as an imperfect information complete
game (c-game), where M and N are the number of different types of aggregators A and B, respectively, in the case that
each aggregator knows the others’ payoff functions and, basic probability distributions involved in the game but does
not know the opponent’s type. The Nash equilibrium is the solution. We consider two variants of (c-game): noncooperative and Stackelberg, in a scenario involving three DRAs competing to sell energy stored in their batteries. In
each game, the effect of price-sensitive demand response scheduling (for WHs and other similar thermostatic storages)
has been considered and the market model has been proposed with/without DR scheduling. The algorithm shown (Fig.
6) depicts the procedure used to calculate the requisite parameters used in obtaining the Nash equilibrium.
Fig. 6: steps of calculating required parameters in the game-theoretic market
The following subsections explain how to determine the parameters needed in the DM process for each game and
each demand response scheduling program.
3.2.1.1. Non-cooperative game without DR scheduling
Assuming that DRAs A, and B, know C’s payoff function, we may determine the nash equilibrium following
the procedure outlined in Fig. 6.
Step1: Define aggregators’ types:
In the first step, a set of different possible discharging cost function’s coefficients ( ai and
B in Eq. 5 to 7) is defined
for each aggregator A and B. The number of defined types are M and N for aggregators A and B, respectively. The
participant’s type is determined largely by electricity price and availability of the batteries for discharging. The
electricity price is significant because the batteries are charged directly from the grid. Availability of stored energy in
the batteries may be estimated by considering use patterns for high power consumption devices such as Water Heaters
(WHs) and other thermostatic devices. For example, when the WHs are consuming power (based on the water usage
and water temperature data), most of the energy stored in the batteries is being consumed locally to power these
devices. Considering these two factors we may classify an aggregator into one of four scenarios:
1 : {elecexp , WHoff }
2 : {elecexp , WHon}
3 : {elecch ,WHoff }
4 : {elecch ,WHon}
For a non-cooperative game without price-sensitive demand response, scheduling depends entirely on water usage
and is completely independent of the price of electricity since the thermostat functions only to maintain the social
welfare demands of the customer without considering price. When two events are independent, the probability of both
events occurring is the product of the probabilities of each event. For aggregator scenarios
1 , 2 , 3 , 4 , we have:
P ( 1 ) P (elecexp WH off ) P (elecexp ) * P (WH off )
P ( 2 ) P (elecexp WH on ) P (elecexp ) * P (WH on )
P ( 3 ) P (elecch WH off ) P (elecch ) * P (WH off )
P ( 4 ) P (elecch WH on ) P (elecch ) * P (WH on )
4
P (
f 1
f
) 1 , and P () is the probability of event . The probability of the WHs’ power state ( P(WHoff ) and
P(WHon ) ) are obtained from the DP scheduler block described in Fig. 4. In this case, the scheduler is determined
by typical water consumption profile and the only objective function is to maintain social welfare- i.e. that the WH
maintains temperatures in the range require by the customer to ensure sufficient hot water is available.
Let
Af and Bf be probability distributions modeling uncertainties in each aggregator’s discharging cost:
Af [Af (1), Af (2),..., Af (M )]
(18)
(19)
Bf [Bf (1), Bf (2),..., Bf ( N )]
Af (m) is the probability that aggregator A is of type m ( m M ) in scenario f and Bf (n) is the probability that
aggregator B is in type n ( n N ).
Step2: Define the basic probability distribution of the game:
The probability that (m, n) would represent participants A and B’s types, respectively, would depend on
electricity price and local demand (WHs’ conditions). The expected probability that A is type m and B is type n,
mn
, is defined as:
4
mn ( P ( f ). ( f ).Af ( m).Bf ( n))
f 1
Where
(20)
( f ) is a participation coefficient for scenario f: equal to one for participants selling energy in the market
and 0 otherwise. For example, ( 2 ) 0 because when the electricity is expensive and the water usage is high,
energy stored in electrochemical cells will be used preferably for powering local loads.
( 1 ) ( 3 ) ( 4 ) 1
because if the price of electricity is cheap, the stored energy is available for sale as
loads may be supplied directly from the grid at lower cost; whereas even if the price is high, absence of local demand
leaves energy available for sale. Given this scheme for predicting mn , the original i-game may be treated as a c-game
with imperfect information. Each aggregator now knows its cost function’s coefficients and may compute its own
discharging costs without knowing the opponent’s discharging cost. In this step
(n N , m M ) .
Am (n) and Bn (m) are calculated
Step 3: Define aggregators’ strategies
An aggregator’s type corresponds to a set of strategies defined by bid slopes (Fig. 2). We assume that each
aggregator has three strategies based on Eq. 9. The slope of the bid curve of gth aggregator (Fig. 2) is mg S u g
S is set to [ S1 , S 2 , S 3 ] for three different strategies. These strategies are: S1 2 for bidding less than
marginal cost, S 2 2 for biding at marginal cost, and S 3 2 for bidding more than marginal cost. Coefficient
where
u g was defined in Eq. 5. In this step sAm and sBn are calculated (n N , m M ) .
Step 4: Define aggregators’ conditional payoff
In this step the following system of equations has been solved for a combination of participant types and
strategies:
A B T
PA PB LC
(21)
where PA and PB are the power remaining in the storages cells of DRA A and B, after feeding their local loads. This
power is sold to aggregator C ( LC is the power purchased by aggregator C), and
T is the transaction price. After
determining PA , PB , and T , in eq. 21, the conditional payoffs of DRA A and B may be calculated with eq. 12 and
13. The conditional payoff matrixes take the form:
hA,11 hA,12 hA,13
H Am (n) hA, 21 hA, 22 hA, 23
(22)
hA,31 hA,32 hA,33
hB ,11 hB ,12 hB ,13
n
H B (m) hB , 21 hB , 22 hB , 23
(23)
hB ,31 hB ,32 hB ,33
m
In H A (n) , each row corresponds to a strategy of A for type m and each column corresponds to a strategy of B for
type n. For instance,
hA, 23 in H A2 (1) corresponds to A’s type 2 payoff if A decides to bid marginal cost against a
situation where B is type 1 and bids above marginal.
H Bn (m) is defined similarly.
Step 5: Define expected payoff matrices.
m
n
Expected values of the payoffs, EPA and EPB , are calculated using eq. 14, 15, 16, 22 and 23. The final
form of the
EPAm matrix is:
1
1
1
1
1
1
1
11
121 131 21
22
23
31
32
33
2
EPAm 112 122 132 21
222 232 312 322 332
3
113 123 133 21
223 233 313 323 333
m
Each column of EPA corresponds to the presumed strategy of participant A’s opponent. For instance, 32 column
EPA1 is A’s type 1 payoff when B uses strategy 3 against A’s type 1 and strategy 2 against A’s type 2. The same
n
notation applies to EPB .
in
Step 6: Obtain the Nash equilibrium of strategies
m
n
Utilizing EPA and EPB ,the Nash equilibrium pairs are obtained. We look for the collection of strategies
in which each aggregator’s strategy would be represented by the best response to other aggregator’s strategies. The
Nash equilibrium is a prediction of how the game will be played. An aggregator’s optimal bid is derived for this
equilibrium point. All aggregators predict that a particular Nash equilibrium will occur and there is no incentive to
play differently.
3.2.1.2. Non-cooperative game with DR scheduling
In this game, price-sensitive scheduling of water heaters is considered. The steps of this game are the same as
non-cooperative game without DR scheduling (section 3.2.1.1) with a single exception in the first step. For a noncooperative game without DR scheduling, WHs keep the water temperature in a given range to satisfy social welfare
constraints , and situation of WHs (On/Off) is independent of the electricity price. In non-cooperative game with DR
scheduling, the objectives are both social welfare and demand scheduling. The first step in finding the Nash
equilibrium of the non-cooperative game with DR scheduling requires knowledge of conditional probabilities since
the events are not independent.
Step1: Define aggregators’ types
As in section 3.2.1.1, a set of possible discharging cost function’s coefficients ( ai and
B in eq. 6 and 7) is
defined for each DRA acting as a seller (A and B). Similarly we define four scenarios ( 1 ,..., 4 ) depending on the
same variables two variables (electricity price and WHs’ condition), but in this game these events are dependent due
to price-sensitive DR scheduling. When two events are dependent, the probability of both occurring is defined using
conditional probability:
P ( 1 ) P (elecexp WH off ) P (elecexp ) * P (WH off / elecexp )
P ( 1 ) P (elecexp WH on ) P (elecexp ) * P (WH on / elecexp )
P ( 3 ) P (elecch WH off ) P (elecch ) * P (WH off / elecch )
P ( 4 ) P (elecch WH on ) P (elecch ) * P (WH on / elecch )
The second part of each equation is the conditional probability. For instance
P(WH off / elecexp ) is the probability
that the water heater is off given that electricity is expensive. These conditional probabilities are calculated using the
DP scheduler (mentioned in Fig. 4) considering two objective functions: social welfare and DR scheduling. In this
game more batteries are available to provide power because scheduling has disabled some loads, thus each DRA
stands to make a larger payoff then in the game without DR scheduling (3.2.1.1).
3.2.1.3. Stackelberg game without DR scheduling
In the previous sections while considering non-cooperative games, we assumed that the transaction price (and
transaction power) depend only on the DRAs bid prices (eq. 21). DRAs with storage capabilities will tend to charge
their batteries when energy is less expensive during off-peak hours, in order to sell it to other DRAs during peak hours
when the price is high thereby maximizing their gains per transaction. When enough DRAs do this, a phenomenon
termed “Reverse Peaks” becomes apparent where the shift of peak consumption is accompanied by a high tendency
of aggregators to sell their stored energy back during high price hours. This undesirable effect can be resolved by
allowing a market controlling center (i.e., utility or any other center controlled by the utility) to a play the role of
market controller in the game. The leader is allowed to set bounds on market transaction quantity, and price thereby
ameliorating the undesirable effects on the market associated with reverse peaks. Competition including a market
controller can be modeled as a Stackelberg game. A Nash equilibrium exist for the Stackelberg game-i.e. each rational
player will chose a particular predictable strategy which offers them no incentive to alter it despite the strategies
chosen by the competitors.
Finding the Nash equilibrium for the Stackelberg game follows the procedure outlined in fig. 6, similar to
the previous sections with the exception that in calculating the conditional payoff (step 4), the transaction powers (
PA , PB ) and transaction price ( T ) are controlled by the leader (utility). If these values, ( PAmax , PBmax , Tmax ), fall
within the range permitted by the controller, the transactions proceeds; but if the obtained values from eq. 21 are
higher than maximum allowable values, these maximum values are used to clear the market. Applying the utility’s
policies in Stackelberg game, prevents the huge effects of players (aggregators) in the market and also prevents reverse
peaks.
3.2.1.4. Stackleberg game with DR scheduling:
The analysis of the Stakleberg game with DR scheduling follows reasoning similar to one without DR
scheduling (section 3.2.1.3) with the exception that we must assume that electricity price and water usage are
dependent events. Conditional probabilities are used in calculating the the probabilities of the intersection of dependent
events for the possible aggregator scenarios in step 1 analogous to what was done in section 3.2.1.2 when considering
DR scheduling for a non-cooperative game.
4. Case study, Discussion and results
For illustration purposes, consider a system with three DRAs labeled A, B, and C as depicted in fig. 7. Load data
for these three DRAs was taken from a real grid model on the island of Maui (Hawaii-United States) [47].
Fig. 7: Case study including 3 agregators: A and B as sellers and C as buyer
DRA A manages 200 houses, all of which contain electrochemical storage cells. DRA B manages 240 house, but only
180 have storage devices. DRA C manages 240 houses of which only 180 are equipped with storage devices. All
houses have WHs assumed to draw the majority of the load (60%). The nominal power of each residential storage
device is 3.3 kW. Thus, the maximum generation capacity of any DRA is:
PAgen = 660kW, PBgen = 594kW, PCgen = 528 kW
In this case study we assume that the nominal power of each WH is 4.5 kW. Thus, the total power demand of A is:
PAdemand = 200 * 4.5 *10/6 PAgen 2160kW
Similarly, for other aggregators:
PBdemand = 2394kW , PCdemand = 2478kW
Since the marginal cost of DRA C is greater than for other aggregators, A and B will compete to sell power to C. We
a A = [4 , 4.3 ]
B = [25 , 23.5]
assume that M = 2 and N = 2 and that the DRAs discharge cost coefficients are: Seller_A
aC = 5
a = [4.2 , 4.5]
Seller_B B
Buyer_C
B = 21.5
B = [24 , 23.5]
For simplicity, it is assumed that both sellers, A and B, know the discharging cost coefficient of the buyer, C.
Fig. 8 shows the normalized price of electricity (in Maui Island) as a function of time over the course of a single day,
price updates resolve every fifteen minutes. The normalized average water consumption for the houses of aggregators
A and B is shown in Fig. 9 for the same time intervals.
Fig. 8: Normalized electricity price in one day of case study (resolution: 15minutes)
Normalized Water Consumption
1
0.8
0.6
0.4
0.2
0
0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time(hour)
Fig. 9: Normalized average water consumption for houses of seller aggregators of case study (resolution: 15minutes)
Results for non-cooperative and Stackelberg games, with and without DR scheduling are presented in the following
sections.
4.1. Results of non-cooperative game without DR scheduling
For our purposes, we take prices over half of the normalized price of electricity to be “expensive”. From the
normalized price data (Fig. 8), P(elecexp ) 0.7083, P(elecch ) 0.2917 . The status of the WHs (on/off) may be
determined from the water consumption data (Fig. 9). In order to keep water temperatures in the range of
110 to
130 F , a typical WH runs (consumes power) for 18 of the 96 fifteen minute intervals of the day. Fig. 10 is the output
WH ON/OFF Status
of DP scheduler and shows the status of the seller’s WHs in this game (non-cooperative without DR scheduling).
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time (hour)
Fig. 10: On/Off status of WHs in non-cooperative game without DR scheduling for the seller aggregators
The probabilities
P(WHon ) 0.1875 and P(WH off ) 0.8125 follow immediately from the known ratio of on
to off intervals. Using the equations from section 3.2.1.1, the probabilities of the four scenarios are:
P( 1 ) 0.5755, P( 2 ) 0.1328, P( 3 ) 0.2370, P( 4 ) 0.0547 .
Probability distributions for both DRAs A and B in the above mentioned scenarios are:
A1 [0.16 , 0.84], B1 [0.21 , 0.79], A2 [0.11, 0.89], B2 [0.18 , 0.82],
A3 [0.75 , 0.25], B3 [0.67 , 0.33], A4 [0.69 , 0.31], B4 [0.60 , 0.40].
The expected probabilities (eq. 20) are:
11 0.0193 , 12 0.0727 , 21 0.1015 , 22 0.3819.
Conditional probability vectors for DRAs A and B are calculated using eq. 15 and 17:
1A [0.21, 0.79], 1B [0.16,0.84], A2 [0.21, 0.79], B2 [0.16,0.84] .
We assume that each DRA choses one of the following strategies: biding 80% of marginal cost ( S1
marginal cost ( S 2
1.6) , biding
2) , and biding 120% of marginal cost ( S 3 2.4) . Therefore, the strategies are:
s [0.00064, 0.00080, 0.00096]
1
A
2
A
1
B
2
B
s [0.00078, 0.00097,0.00117]
s [0.00073, 0.00091, 0.00109]
s [0.00081, 0.00102, 0.00122]
Finally, from the equations in step 5 of section 3.2.1.1, the expected pay off matrixes are:
7.95
EP 17.34
20.25
10.88
22.36
26.5
32
4.12
EP 11.62
13.13
6.71
16.09
9.06
20.12
4.75
12.72
7.34
17.19
9.69
21.22
5.34
13.74
7.93
18.21
18.57
23.55
14.45
19.89
24.87
15.7
21.14
1
A
2
A
13.9
EPB1 21.23
23.32
53.57
EPB2 25.18
0.61
13.41 8.69
26.7 18.61
21.82
11.62
23.63
14.16
27.96
9.35
19.74
12.28
24.75
28.07
33.56
23.23
29.48
14.81
29.09
34.97
10.28
22.24
26.12
23.97
36.02
40.6
18.35
27.71
22.32
33.57
14.78
22.53
19.23
29.01
23.2
34.86
15.55
23.68
20
30.16
30.82
37.7
24.84
32.34
39.22
26.22
33.71
64.56
37.03
100.84 128
75.41 104.55
52.34
82.56
89.85
63.56
117.01
92.7
59.86
31.92
40.01
70.22
7.59
96.14
70.3
46.99
123.3
99.44
77.2
12.94
The DRAs chose their bidding strategies by finding the Nash equilibrium. By inspecting the expected pay off matrices
1
2
1
2
we find that for DRA A, EPA and EPA , we see that the bottom row of each matrix, s A (3) and s A (3) - i.e. bid above
marginal cost if A is of type 1 or type 2- results in the greatest expected payoff regardless of competitors strategies;
we say that this strategy dominates the the others. DRA B will also conclude that a rational player A will choose
strategy 3 (bid above marginal cost) regardless of its type. DRA B need only consider the best response to this strategy.
EPB1 and EPB2 . Which represents the payoffs of for B’s possible
1
strategies when competing against a player A who bids above marginal cost. Examining the last column of EPB and
This is done by considering the 9th column of
EPB2 tells us that: if B is type 1, it will receive the greatest payoff by bidding above marginal cost; and if B is type 2
it will receive the greatest payoff by bidding below marginal cost. Player A knows that B will play this way, but has
no incentive to change its strategy. This strategy is represented by column 7th of
EPA1 and EPA2 (column 31 of
EPA in step 5 of section 3.2.1.1). The pair of strategies “9th column of EPB ” and “7th column of EPA ” is the Nash
equilibrium of the game. The Nash equilibrium is a “consistent” prediction of how the game will be played by rational
players. All participants predict that a particular Nash equilibrium will occur and there is no incentive to play
differently. The strategy pairs in Nash equilibrium are the participants’ maximum strategies which maximize
participants’ conditional payoffs. That is, a participant could obtain at least the payoff at the equilibrium point (or it
may obtain more depending on his opponent’s strategy).
In this section, scheduling of WHs are based on only one objective function: social welfare. The situation of each WH
is based on water consumption only (Fig. 9) and is not effected by the electricity price (Fig. 8).
4.2. Results of non-cooperative game with price-sensitive WH scheduling
In this case, DP scheduler has been used to schedule the WHs of aggregator A, based on the electricity price
(Fig.8) with the joint objectives of social welfare (keeping water temperature in a given range [110,130 ] F using
Fig. 9) and DR scheduling using DP. Similar to section 4.1, P (elecexp ) 0.7083, P (elecch ) 0.2917 . Fig. 11
WH ON/OFF Status
depicts scheduling of the WHs of aggregator A using DP (with inputs of data shown in Fig. 8 and 9). The highlighted
areas show the times when the electricity is expensive (prices higher than half of maximum price).
0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time (hour)
Fig. 11: On/Off status of WHs in non-cooperative game with DR scheduling for the aggregator A
In this case, based on the Fig. 11 and equations of section 3.2.1.2:
P ( 1 ) P (elecexp ) * P (WH off / elecexp ) 0.7083 * 58 68 0.6041
P ( 2 ) P (elecexp ) * P (WH on / elecexp ) 0.7083 *10 68 0.1042
P ( 3 ) P (elecch ) * P (WH off / elecch ) 0.2917 * 22 28 0.2292
P ( 4 ) P (elecch ) * P (WH on / elecch ) 0.2917 * 6 28 0.0625
In aggregator A, Comparison of Fig. 10 with Fig. 11 shows DR scheduling can curtail the load of WHs as much as
= 0.888 * 200 * 4.5 *10/6 PA 1992kW
11.2%. Therefore the new demand of the aggregator A is: PA
. It is assumed here that aggregators B and C do not schedule their WHs. Therefore, the demands of these aggregators
demand
demand
are the same as in section 4.1: PB
gen
= 2394 kW , PCdemand = 2478kW . Because of the load curtailment,
aggregator A has more power stored in its batteries to sell to the aggregator C. With the same probability distributions
and same strategies as in section 4.1 and following the same game process, the expected payoff matrices are:
EP
2
EPA
9.6
19.11
12.68
24.26
22.23
28.63
34.21
5.7
13.35
8.49
18.02
10.99
22.19
15.09
20.75
EP
2
EPB
13.14
20.2
1
A
1
B
15.32
28.69
10.38
20.41
13.47
25.57
16.1
11.07
29.99 21.57
23.84
30.24
35.82
6.39
14.5
9.18
19.17
11.68
23.34
7.02
15.56
9.82
20.23
25.87
16.48
22.14
27.26
17.77
23.43
17.4
26.42
21.21
32.06
13.98
21.45
18.24
27.67
22.06
33.3
14.72
22.55
18.99
28.78
22.19
29.39
36.02
23.65
30.86
45.74
17.57
81.66
55.54
108.55
84.37
51.96
24.23
87.88
62.2
114.77
91.03
56.62
29.29
- 6.73
32.23
62.11
0.16
39.12
69
5.45
37.49
25.28
24.98
16.79
31.15
37.26
14.16
26.72
31.68
12.31
24.4
28.56
32.18
22.8
34.41
38.81
92.53 119.42
67.25
96.09
44.41
74.29
Following the same reasoning used in section 4.1 for finding the Nash equilibrium, aggregator A obtains a higher
EPB1 and EPB2 , we learn that a rational
1
2
participant B would bid sB (3) (above its marginal cost) when it is type 1 and sB (1) (below marginal cost) when it
1
2
is type 2. This strategy is represented by column 7th of EPA and EPA . The pair of strategies “9th column of EPB ”
payoff by bidding above its marginal cost. By inspecting the last column of
and “7th column of EPA ” is the Nash equilibrium of the game. The Nash equilibrium strategies are for A to bid higher
than marginal cost regardless of its type and for B to bid above, or below marginal cost depending if it is type 1 or
type 2, respectively.
4.3. Stackelberg game without DR scheduling
In non-cooperative game, there was no limitation power transactions or transaction prices for selling the stored
energy between the aggregators. The transaction price and transaction power are obtained from Eq. 21, but in real
markets there will exist a controller who limits the transaction power and transaction price. This independent controller
is empowered by the utility and communicates with the utility to control the market. In this case study, for the noncooperative games (section 4.1 and 4.2) the transaction prices (eq. 21) varied between $0.3342 and $0.4903. It is
assumed now that the controller limits the selling price of the stored power to be less than $0.4 for each time interval.
In the case of Stackelberg game without DR scheduling, the probabilities and strategies are the same as section 4.1,
and the final expected pay off matrixes are:
EP
2
EPA
7.95
17.34
10.88
20.35
13.41
19.92
8.69
18.61
11.62
21.62
14.16
21.19
19.42
20.51
20.87
20.19
21.27
21.64
4.12
10.36
5.03
9.67
2.67
8.56
4.75
10.55
5.66
9.86
3.29
8.75
4.5
10.34
5.41
9.65
9.97
10.39
10.2
10.2
10.62
10.42
10.23
10.65
EP
13.9
21.23
18.35
25.67
20.61
25.11
14.78
22.53
19.23
26.97
21.49
26.41
15.55
22.69
20
27.13
23.32
25.7
26.12
24.22
26.6
27.02
24.36
26.74
1
A
1
B
9.35
18.8
20.37
14.81
21.38
21.45 21.81
12.28
21.81
3.04
8.53
10.45
22.26
26.57
27.16
EP
2
B
53.57
23.5
87.11
54.03
107.5
75.74
59.86
29.05
93.4
59.58
113.79
81.29
63.85
32.81
97.38
63.33
- 4.67
26.81
49.41
0.8
32.28
54.88
4.72
36.2
117.78
85.05
58.8
Because of limitations for transaction power price, the expected payoff values have been decreased in this game in
comparison with non-cooperative, unscheduled game (section 4.1). The Nash equilibrium strategies are still the same,
by the same reasoning: for A to bid higher than marginal cost regardless of its type and for B to bid above, or below
marginal cost depending if it is type 1 or type 2, respectively. The pair of strategies “9th column of EPB ” and “7th
column of EPA ” is the Nash equilibrium of this game.
4.4. Stackelberg game with price-sensitive DR scheduling
In this case, aggregator A schedules its WHs based on the electricity price (Fig. 8) and water consumption (Fig. 9) and
limitations on the market due to the constraints imposed by the controller must be considered This situation is the
closest to reality in comparison with the other situations considered. The objective functions are both social welfare
(keeping the water temperature of the houses in range of [110,130 ] F ) and DR scheduling (based on electricity
price and water consumption). Also, the real market policies (with real limitations) have been applied on this situation.
In the game, the probabilities and strategies are the same as section 4.2 and the final expected pay off matrixes are:
EP
2
EPA
9.60 12.68 15.32 10.38 13.47 16.10 11.07 14.16 16.79
19.11 23.32 23.01 20.41 24.63 24.32 20.92 25.13 24.82
22.13 23.32 23.78 23.13 24.32 24.78 23.33 24.52 24.98
EP
2
EPB
13.14 17.40 21.01 13.98 18.24 21.86 14.72 18.99 22.60
20.20 25.53 25.11 21.45 26.77 26.36 21.86 27.18 26.77
22.19 25.36 25.88 23.25 26.42 26.94 23.41 26.58 27.10
1
A
1
B
5.70 7.90 5.72 6.39 8.58 6.40 6.47 8.67 6.49
12.75 12.22 11.25 13.17 12.64 11.67 12.99 12.46 11.49
12.34 12.88 12.79 12.59 13.13 13.05 12.65 13.19 13.11
45.74 80.47 100.96 51.96 86.69 107.18 56.26 90.98 111.47
17.01 47.56 69.36 22.78 53.33 75.13 26.55 57.100 78.89
- 10.75 20.73 43.39 - 5.28 26.20 48.86 - 1.36 30.12 52.78
The Nash equilibrium strategy is for A to bid higher than marginal cost regardless of its type and for B to bid above,
or below marginal cost depending if it is type 1 or type 2, respectively. Here, the pair of strategies “9th column of
EPB ” and “7th column of EPA ” is the Nash equilibrium of this game.
Because of the presence of a controller in the Stackelberg games enforcing more market restrictions in the game,
payoffs of aggregator A and B have been decreased in section 4.3 in comparison with 4.1 and also in section 4.4 in
comparison with 4.2.
5.
Conclusions
In this work we applied principles of elementary game theory to model competition between DRAs for selling
power stored in electrochemical storage cells, in a power market with other DRAs as buyers. The situation presents
itself to competitors as an incomplete information game as any DRA is unable to determine parameters associated
with cost of operating other player’s equipment. In order utilize the better developed theory of c-games with imperfect
information we approximated local power demands using known water usage data by assuming that it was the major
component of the local loads and using public statistical data and a Bayesian approach to derive probability
distributions for DRA types. We then examined four types of c-game to consider the effects of pure competition (noncooperative without WH scheduling) and regulated competition (Stackelberg without WH scheduling). We also
considered both types of game with the addition of price-sensitive WH scheduling to examine the effect of adding DR
to the games. The non-cooperative games are quite idealized and would not be practical to implement for various
reasons explained above, but the Stackelberg games are relatively realistic and provide a possible beneficial alternative
to currently popular market structures.
Application of our model to data compiled from the island of Maui served as a case study to analyze the results
of our model under likely conditions. The results showed that bidding strategies of the DRAs was dependent only on
parameters associated with hardware and not conditions of the games considered. Addition of price-sensitive DR WH
scheduling increased payoffs to those DRAs which implemented it compared with those which did not. Payoffs were
decreased in the regulated (Stackelberg) games compared with non-cooperative purer competition.
DR WH scheduling conserves energy and the resources spent acquiring it. Competition between DRAs for selling
power independently of the generating company provides an alternative to the near-monopoly market structure
currently prominent. Competition for selling power in a market decreases the market price of electricity and allows
for smaller independent energy produces to enter the market thereby lowering the demands placed on the generating
company. In our treatment, power stored in batteries is bought from the utility, but if the power was produced locally
by renewable energy technologies the treatment would be analogous. Increasing penetration of renewable energy
generating technologies such as photovoltaic systems will have a great effect on gaining independence from our
combustible fossil-fuel economy and reducing pollution. We believe that proliferation and optimization of DR smart
grid technologies will play a vital role in building a sustainable power economy.
The University of Hawaii’s Renewable Energy Design Lab (REDlab) is currently developing hardware systems
capable of implementing the presented model. Further work will be forthcoming exploring the implications of our
game-theoretic competition model as an alternative market structure in the state of Hawaii where we currently face
difficulties with integration of larger numbers of PV systems into the legacy grid. Much more work will be required
if we are to reach the states 100% renewable energy goals.
Acknowledgements
This project is sponsored by the U.S. National Science Foundation under award number: 1310709. Thanks to for
John Branigan’s helping to edit this paper.
Bibliography
[1] H. A. Shayanfar, A. Saliminia Lahiji, J. Aghaei, A. Rabiee, "Generation Expansion Planning in pool
market: A hybrid modified game theory and improved genetic algorithm," Energy Conversion and
Management, vol. 50, p. 1149–1156, 2009.
[2] S. M. Moghddas‐Tafreshi, H. A. Shayanfar, A. Saliminia Lahiji, A. Rabiee, J. Aghaei, "Generation
expansion planning in Pool market: A hybrid modified game theory and particle swarm
optimization," Energy Conversion and Management, vol. 52, pp. 1512‐1519, 2011.
[3] A. Zeinalzadeh, A. Alptekinoglu, and G. Arslan, "Distribution‐Free learning in inventory
competition," in The 4th World Congress of the Game Theory Society, Istanbul, Turkey, 2012.
[4] W. Lise, V. Linderhof, O. Kuik, C. Kemfert, R. Ostling, T. Heinzow, "A game theoretic model of the
Northwestern European electricity market—market power and the environment," Energy Policy,
vol. 34, p. 2123–2136, 2006.
[5] E. Bompard, E. Carpaneto, G. Ciwei, R. Napoli, M. Benini, M. Gallanti, G. Migliavacca, "A game
theory simulator for assessing the performances of competitive electricity markets," Electric Power
Systems Research, vol. 78, p. 217–227, 2008.
[6] Z. Liu, X. Zhang, J. Lieu, "Design of the incentive mechanism in electricity auction market based on
the signaling game theory," Energy, vol. 35, pp. 1813‐1819, 2010.
[7] S. Kamalinia, M. Shahidehpour, L. Wu, "Sustainable resource planning in energy markets," Applied
Energy, vol. 133, pp. 112‐120, 2014.
[8] T. Sueyoshi, "An agent‐based approach equipped with game theory: Strategic collaboration among
learning agents during a dynamic market change in the California electricity crisis," Energy
Economics, vol. 32, pp. 1009‐1024, 2010.
[9] S. Lorenczik, T. Panke, "Assessing market structures in resource markets — An empirical analysis of
the market for metallurgical coal using various equilibrium models," Energy Economics, vol. 59, p.
179–187, 2016.
[10] N. Neshat, M. R. Amin‐Naseri, "Cleaner power generation through market‐driven generation
expansion planning: an agent‐based hybrid framework of game theory and Particle Swarm
Optimization," Journal of Cleaner Production, vol. 105, pp. 206‐217, 2015.
[11] P. C. See, O. B. Fosso, K. Y. Wong, "Cross‐Border Transfer of Electric Power under Uncertainty: A
Game of Incomplete Information," Energy Procedia, vol. 53, p. 95 – 103, 2014.
[12] R. Tian, Q. Zhang, G. Wang, H. Li, S. Chen, Y. Li, Y. Tian, "Study on the promotion of natural gas‐
fired electricity with energy market reform in China using a dynamic game‐theoretic model,"
Applied Energy, p. DOI: 10.1016/j.apenergy.2015.11.079, 2015.
[13] W. Su, A. Q. Huang, "A game theoretic framework for a next‐generation retail electricity market
with high penetration of distributed residential electricity suppliers," Applied Energy, vol. 119, pp.
341‐350, 2014.
[14] N. Zhang, Y. Yan, W. Su, "A game‐theoretic economic operation of residential distribution system
with high participation of distributed electricity prosumers," Applied Energy, vol. 154, pp. 471‐479,
2015.
[15] A. Sheikhi, M. Rayati, S. Bahrami, A. M. Ranjbar, S. Sattari, "A cloud computing framework on
demand side management game in smart energy hubs," Electrical Power and Energy Systems, vol.
64, pp. 1007‐1016, 2015.
[16] N. I. Nwulu, X. Xia, "Multi‐objective dynamic economic emission dispatch of electric power
generation integrated with game theory based demand response programs," Energy Conversion
and Management, vol. 89, pp. 963‐974, 2015.
[17] N. I. Nwulu, X. Xia, "Implementing a model predictive control strategy on the dynamic economic
emission dispatch problem with game theory based demand response programs," Energy, pp. 404‐
419, 2015.
[18] S. S. Reka, V. Ramesh, "A demand response modeling for residential consumers in smart grid
environment using game theory based energy scheduling algorithm," Ain Shams Engineering
Journal, vol. 7, pp. 835‐845, 2016.
[19] M. Yu, S. H. Hong, "Supply–demand balancing for power management in smart grid: A Stackelberg
game approach," Applied Energy, vol. 164, pp. 702‐710, 2016.
[20] G. Wang, Q. Zhang, H. Li, B. C. McLellan, S. Chen, Y. Li, Y. Tian, "Study on the promotion impact of
demand response on distributed PV penetration by using non‐cooperative game theoretical
analysis," Applied Energy, DOI: 10.1016/j.apenergy.2016.01.016.
[21] A. M. Ospina, N. Quijano, "Distributed control of small‐scale power systems using noncooperative
games," Electrical Power and Energy Systems, vol. 82, pp. 535‐544, 2016.
[22] C. Gong, K. Tang, K. Zhu, A. Hailu, "An optimal time‐of‐use pricing for urban gas: A study with a
multi‐agent evolutionary game‐theoretic perspective," Applied Energy, vol. 163, pp. 283‐294,
2016.
[23] L. Castillo, C. A. Dorao, "Consensual decision‐making model based on game theory for LNG
processes," Energy Conversion and Management, vol. 64, pp. 387‐396, 2012.
[24] L. Castillo, C. A. Dorao, "Decision‐making in the oil and gas projects based on game theory:
Conceptual process design," Energy Conversion and Management, vol. 66, pp. 48‐55, 2013.
[25] H. S. Aplak, M. Z. Sogut, "Game theory approach in decisional process of energy management for
industrial sector," Energy Conversion and Management, vol. 74, pp. 70‐80, 2013.
[26] V. Nanduri, N. Kazemzadeh, "Economic impact assessment and operational decision making in
emission and transmission constrained electricity markets," Applied Energy, vol. 96, pp. 212‐221,
2012.
[27] T. K. Das, P. Rocha, C. Babayigit, "A matrix game model for analyzing FTR bidding strategies in
deregulated electric power markets," Electrical Power and Energy Systems, vol. 32, pp. 760‐768,
2010.
[28] J. Wang, Z. Zhou, A. Botterud, "An evolutionary game approach to analyzing bidding strategies in
electricity markets with elastic demand," Energy, vol. 36, pp. 3459‐3467, 2011.
[29] M. A. F. Ghazvini, B. Canizes, Z. Vale, H. Morais, "Stochastic short‐term maintenance scheduling of
GENCOs in an oligopolistic electricity market," Applied Energy, vol. 101, pp. 667‐677, 2013.
[30] A. Zeinalzadeh, A. Alptekinoglu, and G. Arslan, "On Existence of equilibrium in inventory
competition with fixed order costs," in Operations Research Letters, 2015.
[31] D. Menniti, A. Pinnarelli, N. Sorrentino, "Simulation of producers behaviour in the electricity
market by evolutionary games," Electric Power Systems Research, vol. 78, pp. 475‐483, 2008.
[32] C. G. Min, M. K. Kim, J. K. Park, Y. T. Yoon, "Game‐theory‐based generation maintenance
scheduling in electricity markets," Energy, vol. 55, pp. 310‐318, 2013.
[33] Z. Qiu, N. Gui, C. Yang, G. Deconinck, W. Gui, "Assessing impact of subjective demand beliefs on a
dynamic duopoly electricity market game," Electrical Power and Energy Systems, vol. 60, pp. 182‐
189, 2014.
[34] Y. Maali, "A multiobjective approach for solving cooperative n‐person games," Electrical Power and
Energy Systems, vol. 31, pp. 608‐610, 2009.
[35] C. L. Prete, B. F. Hobbs, "A cooperative game theoretic analysis of incentives for microgrids in
regulated electricity markets," Applied Energy, vol. 169, pp. 524‐541, 2016.
[36] H. Omrani, R. Gharizadeh Beiragh, S. Shafiei Kaleibari, "Performance assessment of Iranian
electricity distribution companies by an integrated cooperative game data envelopment analysis
principal component analysis approach," Electrical Power and Energy Systems, vol. 64, pp. 617‐
625, 2015.
[37] S. Rahmani Dabbagh, M. K. Sheikh‐El‐Eslami, "Risk‐based profit allocation to DERs integrated with
a virtual powerplant using cooperative Game theory," Electric Power Systems Research, vol. 121,
pp. 368‐378, 2015.
[38] D. J. Kang, B. H. Kim, D. Hur, "Supplier bidding strategy based on non‐cooperative game theory
concepts in single auction power pools," Electric Power Systems Research, vol. 77, pp. 630‐636,
2007.
[39] R. Wan, J. R. Boyceb, "Non‐renewable resource Stackelberg games," Resource and Energy
Economics, vol. 37, pp. 102‐121, 2014.
[40] T. Wu, L. Ma, Z. Mao, X. Ou, "Setting up charging electric stations within residential communities in
current China: Gaming of government agencies and property management companies," Energy
Policy, vol. 77, pp. 216‐226, 2015.
[41] A. Sheikhi, Sh. Bahrami, A. M. Ranjbar, H. Oraee, "Strategic charging method for plugged in hybrid
electric vehicles in smart grids; a game theoretic approach," Electrical Power and Energy Systems,
vol. 53, p. 499–506, 2013.
[42] M. X. Cheng, M. Crow, Q. Ye, "A game theory approach to vulnerability analysis: Integrating power
flows with topological analysis," Electrical Power and Energy Systems, vol. 82, pp. 29‐36, 2016.
[43] M. Jahangoshai Rezaee, A. Moini, A. Makui, "Operational and non‐operational performance
evaluation of thermal power plants in Iran: A game theory approach," Energy, vol. 38, pp. 96‐103,
2012.
[44] A. Orlov, "Effects of higher domestic gas prices in Russia on the European gas market: A game
theoretical Hotelling model q," Applied Energy, vol. 164, pp. 188‐199, 2016.
[45] T. Lv, Q. Ai, "Interactive energy management of networked microgrids‐based active distribution
system considering large‐scale integration of renewable energy resources," Applied Energy, vol.
163, pp. 408‐422, 2016.
[46] H. M. Soliman, A. Leon‐Garcia, "Game‐Theoretic Demand‐Side Management With Storage
Devicesforthe Future Smart Grid," IEEE Trans. Smart Grids, vol. 5, no. 3, pp. 1475‐1485, 2014.
[47] GE Energy Consulting, "Hawaii Solar Integration Study: Final Technical Report for Maui," Hawai‘i
Natural Energy Institute (HNEI), University of Hawai‘i, 2013.
[48] A. Zeinalzadeh, A. Alptekinoglu, and G. Arslan, "On Existence of equilibrium in inventory
competition with fixed order costs", submitted, Operations Research Letters, August 5, 2015.
[49] A. Zeinalzadeh, A. Alptekinoglu, and G. Arslan, "Learning in infinite‐horizon inventory competition
with total demand observations," Proceedings of the American Control Conference, Montreal Canada,
June 27‐29, 2012, pp. 1382‐1387.
[50] A. Zeinalzadeh, "Learning in inventory competition games," University of Hawaii at Manoa, 2013.
[51] A. Zeinalzadeh, A. Alptekinoglu and G. Arslan, "Inventory competition under fixed order costs,"
INFORMS MSOM (Manufacturing and Service Operations Management) Annual Conference, Technion,
Haifa, Israel, June 28‐29, 2010.
[52] A. Zeinalzadeh and Vijay Gupta, "Minimizing risk of load shedding and renewable energy
curtailment in a microgrid with energy storage," ArXiv e‐prints, arXiv: 1611.08000, Nov. 2016.
[53] A. Zeinalzadeh, and Vijay Gupta, "Pricing energy in the presence of renewables," ArXiv e‐prints,
arXiv:1611.08006, Nov 2016.
| 3cs.SY
|
Efficient Computation of the 8-point DCT via Summation by Parts
D. F. G. Coelho∗ R. J. Cintra† V. S. Dimitrov‡
arXiv:1801.05832v2 [cs.DS] 28 Mar 2018
Abstract
This paper introduces a new fast algorithm for the 8-point discrete cosine transform (DCT) based on the
summation-by-parts formula. The proposed method converts the DCT matrix into an alternative transformation
matrix that can be decomposed into sparse matrices of low multiplicative complexity. The method is capable of
scaled and exact DCT computation and its associated fast algorithm achieves the theoretical minimal multiplicative
complexity for the 8-point DCT. Depending on the nature of the input signal simplifications can be introduced and
the overall complexity of the proposed algorithm can be further reduced. Several types of input signal are analyzed:
arbitrary, null mean, accumulated, and null mean/accumulated signal. The proposed tool has potential application
in harmonic detection, image enhancement, and feature extraction, where input signal DC level is discarded and/or
the signal is required to be integrated.
Keywords
DCT, Fast Algorithms, Image Processing
1
Introduction
Discrete transforms play a central role in signal processing. Noteworthy methods include trigonometric transforms—
such as the discrete Fourier transform (DFT) [1], discrete Hartley transform (DHT) [1], discrete cosine transform
(DCT) [2], and discrete sine transform (DST) [2]—as well as the Haar and Walsh-Hadamard transforms [3]. Among
these methods, the DCT has been applied in several practical contexts: noise reduction [4], watermarking methods [5],
image/video compression techniques [2], and harmonic detection [2], to cite a few. In fact, when processing signals
modeled as a stationary Markov-1 type random process, the DCT behaves as the asymptotic case of the optimal
Karhunen–Loève transform in terms of data decorrelation [2]. This approximation holds true when the correlation
coefficient of the related stochastic process tends to the unit, which is the case for many real signals—specially
images [2]. Moreover, the recent increase in image/video processing demand for consumer electronics [6] and big data
manipulation [7] emphasizes the necessity for fast and efficient DCT computation [8].
As a consequence, the 8-point DCT is adopted in several image and video coding schemes [9], such as JPEG [10],
MPEG-1 [10], H.264 [11], HEVC [12], AVS China [13, p. 61], and VP-10 [13, p. 165]. Aiming at minimizing the
computational cost of the DCT evaluation, a number of fast algorithms for the 8-point DCT have been proposed,
including Chen’s DCT algorithm [14], Lee method [15], Loeffler algorithm [16], Feig-Winograd DCT factorization [17],
and the Arai DCT [18].
Multiplication operations as required by DCT and others discrete-time transforms can be implemented via long
sequences of additions, bit-shifting operation, and sign changes [19]. Thus, algorithms that require multiplications
often have higher computational costs [20]. Therefore, above-mentioned methods were developed in order to reduce
the overall number of multiplications [2]. The Arai DCT is particularly useful because it furnishes a scaled version of
the DCT spectrum. In some applications such as harmonic detection [21, 22] and JPEG-like image compression [10, 9],
the scaled DCT is often a sufficient tool. This is because in these contexts only the relative value of the spectrum is
necessary. Therefore, part of the cost of computing the DCT can be avoided [2].
∗
D. F. G. Coelho is with the Department of Electrical and Computer Engineering, University of Calgary, Calgary, Canada.
R. J. Cintra is with the Signal Processing Group, Departamento de Estatı́stica, Universidade Federal de Pernambuco and the
Department of Electrical and Computer Engineering, University of Calgary, Calgary, Canada. E-mail: [email protected]
‡
V. S. Dimitrov is with the Department of Electrical and Computer Engineering, University of Calgary, Calgary, Canada.
†
1
Among the fundamental mathematical tools, we separate the summation-by-parts technique [23, p. 54], which
is the discrete-time counterpart for the well-known integration-by-parts method [24, p. 144]. Although applied in
several contexts such as computational physics for approximate second derivatives [25], approximations of the linear
advection-diffusion equation in computational fluid dynamics [26], and rapid calculation of slow converging series
in electromagnetic problems [27], it has been particularly overlooked by the signal processing community. Early
attempts to employ it as a numerical analysis tool are due to Boudreaux-Bartels and collaborators in the context of
the DFT computation [28] and the evaluation of Fourier coefficients errors calculations [29].
The aim of this paper is to propose a new fast algorithm for the 8-point DCT computation based on the summationby-parts formula for periodic signals [23, 30]. The introduced method is sought to achieve the theoretical minimal
multiplicative complexity for the exact DCT computation [31, 32]. Moreover, to further minimize computational costs,
the proposed algorithm is also sought to provide a scaled version of the DCT spectrum [2]. The proposed algorithm
finds application in some important problems, such as feature detection, where DC level may not be relevant [33, 34].
Also, it can be applied to scenarios where input signal is natively accumulated (integrated) [1, p. 19]. This situation
occurs in face recognition problems, where usual algorithms require data to be integrated [35, 36].
This paper is organized as follows. In Section 2, we furnish the mathematical background for the summationby-parts technique and the DCT. Considering matrix formalism, we detail the proposed algorithm for the DCT in
Section 3. In Section 4, the introduced method is assessed in terms of its computational complexity and comparisons
with competing algorithms are shown. Section 5 brings final comments and remarks.
2
Mathematical Background
2.1
Summation-by-parts
The summation-by-parts technique is the discrete-time equivalent of the integration-by-parts method [23]. Let x[n]
and y[n] be two discrete-time signals. The summation-by-parts prescribes that [23, 28]:
N−1
X
n=0
x[n]y[n] =x[N ]y[N ] − x[0]y[0]
−
N−1
X
n=0
n−1
X
!
x[i]
i=0
· ∆y[n],
where ∆ denotes the forward difference operator given by ∆y[n] , y[n + 1] − y[n] [23]. Above expression can be
simplified with the assumption of the following additional weak conditions. Admitting that the considered signals
are periodic with period N , it was established in [30] that:
N−1
X
n=0
x[n] · y[n] = −
N−1
X
n=0
n−1
X
i=0
!
x[i]
· ∆y[n].
(1)
The above condition is not too restrictive. Indeed discrete-time Fourier analysis often assume that the input
signals are periodic [1, 2, 37]. In particular, the DCT can be obtained as the solution to the harmonic oscillation
problem [2].
The expression
PN−1
n=0
x[n] · y[n] can be interpreted as a discrete-time transformation. Let x[n] be the input
signal to be transformed and y[n] = ker[n, k] be a given discrete transformation kernel for the kth transform-domain
component. Therefore, we have that:
X[k] =
N−1
X
n=0
x[n] · ker[n, k],
2
k = 0, 1, . . . , N − 1,
(2)
Table 1: Common discrete transform kernels
Transform
ker[n, k]
DFT [1]
exp −j 2πnk
N
2πnk
DHT [1]
cas N
DCT [2]
cos π(2n+1)k
2N
π
DST [2]
sin N (k + 21 )(n + 12 )
x[n]
DC offset
removal
Accumulation z[n]
system
Summationby-parts
formula
X[k]
Figure 1: Block diagram of the proposed architecture.
where X[k] is the transformed output signal. Table 1 summarizes common transformation kernels. Therefore,
applying (1) into (2) yields the following expression for the transform-domain components:
X[k] = −
=−
where z[n] =
Pn
i=0
N−1
X
n=0
N−1
X
n=0
n
X
!
x[i]
i=0
· ∆ ker[n, k]
z[n] · ∆ ker[n, k],
k = 0, 1, . . . , N − 1,
(3)
x[i], for n = 0, 1, . . . , N − 1. Comparing (2) with (3), we notice that the original transform
expression was re-written into an alternative form where both the input data and the kernel function were processed.
Notice that z[n] is the output of an accumulator system for input signal x[n] [1]; whereas ∆ ker[n, k] derives from
a forward difference system for input signal ker[n, k] [1]. Although the forward difference system is not causal, this
fact poses no difficulty to above formalism. This is because ker[n, k] is not a random real-time sequence—but a
deterministic sequence whose values are known a priori [19, p. 7].
Moreover, if x[n] possesses null mean, then the following expression holds true:
z[N − 1] =
N−1
X
x[i] = 0
and
X[0] = 0.
i=0
For trigonometric transforms, above condition implies X[0] = 0 (null DC value). Therefore, (3) can be simplified and
written as:
X[k] = −
N−2
X
n=0
z[n] · ∆ ker[n, k],
k = 1, 2, . . . , N − 1.
(4)
Above summation ranges from 0 to N − 2. This means that the transformation matrix linked to (3) has dimension
(N − 1) × (N − 1). This fact contrasts with the original transformation matrix, which has size N × N . Thus, the
summation-by-parts effected a dimension reduction of transform computation. As a consequence, the computational
cost of associate algorithms is expected to be reduced.
Figure 1 depicts the overall diagram for the transform computation based on the summation-by-parts formula,
when x[n] is assumed to be an arbitrary signal. Notice that, if N is a power of two, both the DC removal block and
the accumulation system are multiplierless operations.
3
2.2
Discrete Cosine Transform
The DCT is a linear transformation that maps an N -point discrete-time signal x[n] into another N -point discrete-time
signal X[k] according to the following relationship [16]:
N−1
X
π(2n + 1)k
4
x[n] cos
X[k] = √ αk
,
2N
N
n=0
(5)
√
where k = 0, 1, . . . , N − 1, α0 = 1/ 2, and αk = 1, for k > 0. The above expression can be given a compact
format by means of matrix representation. Indeed, considering signals x[n] and X[k] in column vector format as
h
i⊤
h
i⊤
x = x[0] x[1] · · · x[N − 1]
and X = X[0] X[1] · · · X[N − 1] , we have that:
X = CN · x,
where CN is the DCT matrix, whose (k, n)-entry is given by
(6)
√4 αk
N
cos
transformation matrix:
C , C8 =
√
2·
c4
c1
c2
c3
c4
c5
c6
c7
c4 c4 c4
c3 c5 c7
c6 −c6 −c2
−c7 −c1 −c5
−c4 −c4 c4
−c1 c7 c3
−c2 c2 −c6
−c5 c3 −c1
c4
−c7
−c2
c5
c4
−c3
−c6
c1
π(2n+1)k
2N
c4 c4 c4
−c5 −c3 −c1
−c6 c6 c2
c1 c7 −c3
−c4 −c4 c4
−c7 c1 −c5
c2 −c2 c6
−c3 c5 −c7
. For N = 8, we have the following
,
where ck = cos(kπ/16), for k = 1, 2, . . . , 7. This particular definition of the DCT is adopted in [38, 39, 40], having been
√
2·c4 = 1; therefore the DC component
considered to derive the well-known Loeffler DCT algorithm [16]. Notice that
X[0] is evaluated without multiplications [16]. In [32], Heideman introduces an in-depth mathematical analysis of
the multiplicative complexity of major discrete transforms. A result from multiplicative complexity theory that is
relevant to the current work is the following. If the transform blocklength is a power of two, N = 2r , r = 1, 2, . . ., then
the minimum multiplicative complexity µ(N ) of the DCT has a general formula given by µ(2r ) = 2r+1 − r − 2 [32,
Theorem 6.3, p. 117]. For N = 8, we obtain µ(23 ) = 11.
3
DCT Computation via Summation-by-parts
In this section, we apply the summation-by-parts technique to propose an alternative computation of the 8-point
DCT. Next we analyze the resulting expressions in order to derive a fast algorithm by means of matrix factorization.
3.1
Matrix Formalism
To facilitate the development of the sought DCT fast algorithm, we re-cast the summation-by-part formula into
matrix formalism. But, first, the forward difference operator needs to be adapted to manipulate matrices. Let M be
a square matrix. Then ∆M is the matrix resulted from cyclically applying the forward difference operator to each
row of M.
Therefore, considering the summation-by-parts formula shown in (4), we have that (6) can be written according
to:
X = ∆C · z,
where the N -point vector X
=
h
X[0]
X[1]
···
4
X[N − 1]
i⊤
represents the DCT spectrum, z
=
h
z[0]
z[1]
···
z[N − 1]
i⊤
is the accumulated input signal, and
∆C =
√
2×
0
c3 −c1
c6 −c2
−c3 −c7
−2c4
−c1 −c5
−c2 −c6
−c5 −c7
0
0
0
0
c5 −c3 c7 −c5 −2c7 c7 −c5
−2c6 c6 −c2
0 c2 −c6
c7 −c1 c1 −c5 2c5 c1 −c5
0
2c4
0
−2c4
c1 +c7 c3 −c7 −2c3 c3 −c7
2c2 −c2 −c6
0 c2 +c6
c3 +c5 −c1 −c3 2c1 −c1 −c3
0
0
0
c5 −c3 c3 −c1 2c1
2c6 c2 −c6
0
c7 −c1 −c3 −c7 2c3
0
2c4
0.
c1 +c7 −c1 −c5 2c5
−2c2
c2 +c6
0
c3 +c5 −c5 −c7 2c7
Considering the sum-to-product identities [41, p. 72] and symmetry identities [23], the entries of the matrix ∆C
can be given a multiplicative form, as shown below:
0
0
0
0
0
0
0
s1 s4 s1 s6 s1 s1 s6 s1 s4 s1 s2
s2 s2 s4
0 −s2 s4
−s2 −s2 s4
s3 s4 −s3 s2 −s3 −s3 s2 s3 s4 s3 s6
0
−s4
0
s4
0
−s4
−s5 s4 −s5 s2 s5 −s5 s2 −s5 s4 s5 s6
s6 s4
−s6 s6 s4
0 −s6 s4
s6 −s6 s4
s7 s2 −s7 s4 s7 s6 −s7 s7 s6 −s7 s4 s7 s2
s1 s2
s2 s4
√
∆C = 2 2 · s3 ss64
s5 s6
0
s7
0
s5
0,
s3
0
s1
for sk = sin(kπ/16), k = 1, 2, . . . , 7.
If the input signal possesses null mean, then we have that X[0] = 0 and z[N − 1] = 0. Therefore, the first row
and last column of ∆C can be neglected. It follows that only the remaining 7×7 submatrix is necessary for the
computation of X. This particular matrix is given by:
s1 s2
s1 s4 s1 s6 s1 s1 s6 s1 s4 s1 s2
s2 s2 s4
0 −s2 s4
−s2 −s2 s4
s3 s4 −s3 s2 −s3 −s3 s2 s3 s4 s3 s6
0
−s4
0
s4
0
−s4
s5 s6 −s5 s4 −s5 s2 s5 −s5 s2 −s5 s4 s5 s6
s6 s4
−s6 s6 s4
0 −s6 s4
s6 −s6 s4
s7 s2 −s7 s4 s7 s6 −s7 s7 s6 −s7 s4 s7 s2
s2 s4
√ s s
e = 2 2 · 3 s46
C
.
e is sufficient for the computation of all DCT coefficients—except the DC level.
Notice that C
3.2
Scaling Matrix
e shows repeated multiplicands along its rows. Thus, the following factorization is
An examination of matrix C
obtained:
s
2
s4
s4
1
s6
s4
1 s6
0 −s4
s4 −1
s2 −s4
s4 s2
−1 −s4
s6
−1 ,
s6
s4 0 −s4
1 −s4
s6 −1 s6 −s4 s2
s s −s −1 −s s
e = S · 61 40 −12 0 21 40
C
s6 −s4 −s2 1 −s2 −s4
√
where S = 2 2 · diag(s1 , s2 , s3 , s4 , s5 , s6 , s7 ). The above expression tells us that the matrix S contributes only with
scaling factors to the actual DCT computation. When considering applications that require only a scaled version
of the DCT—such as harmonic detection [22] and color enhancement [42]—the computational cost of S can be
disregarded. Additionally, in the context of image compression, diagonal matrices can be directly absorbed into the
quantization matrix; representing no extra computation [43, 44].
√
Notice also that the scaling factor 2 2s4 = 2 is a trivial multiplication [20] that can be implemented via a
simple bit-shifting operation. Therefore, the computational cost of the scaling matrix S is actually only six—not
seven—multiplications.
5
3.3
Fast Algorithm
e Because C
e is a highly symmetrical matrix, factorNow our aim is to provide a sparse matrix factorization to C.
ization methods based on butterfly structures can be directly applied [1, 2, 20]. Therefore, we obtain the following
factorization:
e = S · P · M · A,
C
where
P=
A=
1
0
0
0
0
0
0
1
0
00
0
0
1
0
0
0
0
1
0
0
0
1
0
0
0
1
0
0
1
0
0
0
0
0
0
0
1
0
1
0
0
0 0 0 0
0 0 0 0
0 1 0 0
0 0 0 1,
0 0 0 0
1 0 0 0
0 0 1 0
0 0 0 1
0 0 1 0
0 1 0 0
1 0 0 0,
0 −1 0 0
0 0 −1 0
0 0 0 −1
and
M=
s
s4 s6 1 0 0 0
2
s6 s4 −s2 −1 0 0 0
s6 −s4 −s2 1 0 0 0
s2 −s4 s6 −1 0 0 0 .
0 0 0 0 s4 1 s4
0
0
0 0 −1 0 1
0
0
0 0 s4 −1 s4
Matrix P is a simple permutation matrix, which represents no computational cost. In terms of hardware, P translates
into simple wiring. Matrix A is an additive matrix consisting of the usual butterfly stage present in decimation-infrequency algorithms [20]. The remaining matrix M is block-diagonal and still contains mathematical redundancies
due to its symmetrical nature. Considering the matrix factorizations for fast algorithm design described in [20], the
following expression can be obtained:
M = M1 · M2 · M3 · M4 ,
where
M1 =
M2 =
M3 =
M4 =
1
0 1
0 1 0
0 1 0
1 0 −1
0 0 0
0 0 0
0 0 0
s2 s6
s6 −s2
0 0
0 0
0
0
0
0
0
0
1 0 0
0 0 1
0 s4 0
0 0 0
0 0 0
0 0 0
0 0 0
1 0 0
0 1 0
00 00 10
0 0 0
0 0 0
0 0 0
0 0
1 0
−1 0
0 0
0 1
0 0
0 0
0 0
0 0
1 1
1 −1
0 0
0 0
0 0
0 0
0 0
0 0
1 0
0 s4
0 0
0 0
0 0
0 0
0 0
1 0
0 1
0 0
0 1
The block-diagonal matrix M2 contains a rotation block
6
0
0
0
0,
0
1
0
0 0 0
0 0 0
0 0 0
0 0 0,
1 1 0
1 −1 0
0 0 −1
0 0
0 0
0 0
0 0,
0 0
1 0
0 1
0 0
0 0
0 0
0 0.
0 1
1 0
0 −1
s2
0
0
0
0
0
0
1
s6
s6 −s2
, which can be further decomposed [2]. Thus,
z[0]
√
2c2
√
2s2
X[1]
√
2 2s1
X[2]
√
2 2s2
X[3]
√
2 2s3
X[4]
2
X[5]
√
2 2s5
X[6]
√
2 2s6
X[7]
√
2 2s7
z[1]
z[2]
s4
s2
z[3]
z[4]
s4
z[5]
z[6]
Figure 2: SFG for the proposed algorithm. Dashed lines represents multiplication by −1.
1
8
x[0]
z[0]
−1
x[1]
z[1]
−1
x[2]
z[2]
−1
x[3]
z[3]
−1
x[4]
z[4]
−1
x[5]
z[5]
−1
x[6]
z[6]
−1
x[7]
DC oset removal
A umulation
System
Figure 3: SFG for pre-processing stage consisting of the DC component removal coupled with accumulation system.
we have:
M2 = R1 · R2 · R3 ,
where
R1 =
R3 =
0
s2 +s6
0
0
0
0
0
1 0
0 1
10 10
0 0
0 0
0 0
0 0
s6 −s2
0
0
0
0
0
0
0 0
0 0
0 0
1 1
1 −1
0 0
0 0
0 0
1 0 0 0 0 0
−1 0 0 0 0 0
0 1 0 0 0 0
0 0 1 0 0 0,
0 0 0 1 0 0
0 0 0 0 1 0
0 0 0 0 0 1
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0,
1 1 0
1 −1 0
0 0 −1
and R2 = diag(1, 1, s2 , 1, 1, 1, 1, 1).
√
By means of usual trigonometric manipulations, we notice that the required multiplicands in R1 satisfy: s2 + s6 =
√
2 s2 . Finally, the complete sparse matrix factorization is given by:
2 c2 and s6 − s2 =
e = S · P · M1 · R1 · R2 · R3 · M3 · M4 · A.
C
The signal flow graph (SFG) for the proposed algorithm is shown in Figure 2.
7
4
Computational Complexity Assessment and Comparison
In this section, we assess the computational complexity of the proposed algorithm as measured by its arithmetic cost.
For such, we evaluate the additive and multiplicative count required by the introduced method.
4.1
Results
We distinguish the following scenarios for the input data: (i) arbitrary signal; (ii) null mean signal; (iii) accumulated
signal; and (iv) null mean and accumulated signal. Scenario (i) is the most general case. Thus, there is less
redundancy to be exploited aiming at the minimization of computational cost. Scenario (i) is commonly found in
image compression applications [2, 37]. Scenario (ii) is pertinent in the context of feature detection, for instance,
where DC level may not be relevant [33, 34]. Scenario (iii) represents the case where the input signal is natively
accumulated (integrated). This situation occurs in face recognition problems, where usual algorithms require data
to be integrated [35, 36]. Scenario (iv) is a combination of above last two cases. Notice that Scenario (ii) does not
require the DC offset removal block (cf. Figure 1); Scenario (iii) does not require the accumulation system but needs
a DC removal block; and Scenario (iv) requires neither the DC offset removal block nor the accumulation system.
Table 2 compares the proposed method with several prominent techniques for the 8-point DCT evaluation under
all discussed scenarios. The details of the arithmetic complexity assessment of each considered technique is found
in [2]. For each scenario, we emphasized in bold the best results. Because Arai method [18] and the proposed
algorithm are scaled DCT methods, we show the multiplicative complexity of the non-scaled computation as well. In
parenthesis, when applicable, we furnish the full multiplicative complexity of the scaled methods, when the scaling
factors are considered. The multiplicative complexity for Loeffler, Lee, and Chen DCT algorithms are limited to the
non-scaled case, because these methods do not admit scaled computation. The proposed algorithm could outperform
all competing method under Scenarios (ii), (iii), and (iv). Although the proposed method requires five multiplications
to compute the scaled DCT—the same as the Arai method—it demands fewer scaling multiplications. Scenario (i)
is not ideally suitable for the proposed algorithm because it benefits of the DC level removal and accumulated input
signal. However, even in this case, the introduced algorithm could achieve the theoretical minimum of multiplicative
complexity.
4.2
Discussion
For Scenario (i), the DC offset removal block requires seven additions to compute the mean value and then seven
subtractions to remove the DC level from input samples x[0], x[1], . . . , x[6]. Thus, a total of 14 additions are necessary.
The accumulation system requires a total of six additions (cf. (4)). Thus, the DC removal block combined with
accumulation system requires a total of 20 additions. Figure 3 details the inner structure of each of the abovementioned blocks. Multiplicative costs are concentrated in matrices S, R1 , R2 , and M3 . They account for 11 nontrivial multiplications, which is the minimum multiplicative complexity for the 8-point DCT [31, 32]. Matrices M1 ,
R1 , R3 , M4 , and A contribute with 19 additions. Thus, under Scenario (i), the proposed algorithm requires a total
of 39 additions. As we demonstrate below, this is the only scenario where the proposed method cannot outperform
the most relevant method for the 8-point DCT computation. Nevertheless, even in this scenario, the proposed method
attained the theoretical minimal multiplicative complexity, which is a relevant figure of merit [20]. In comparison
with competing methods, the extra required computation consists of additions only.
Under Scenario (ii), the proposed algorithm does not demand the DC removal block but only one instantiation of
an accumulation system (Figure 3, rightmost subblock). Therefore, a total of 13 additions in savings is attained when
compared to Scenario (i). Consequently, considering Scenario (ii), the proposed algorithm requires only 25 additions.
Usual methods presented in literature can be adapted for Scenario (ii). However, when compared to the costs under
Scenario (i), such methods could save only three additions, instead of seven additions for the computation of the
8
Table 2: Comparison for non-trivial products and additions for 8-point DCT algorithms
Algorithm
Scaled?
Scenarios
Mult.
Additions
Loeffler [16]
No
(i)
(ii)
(iii)
(iv)
11
11
11
11
29
26
36
33
Lee [15]
No
(i)
(ii)
(iii)
(iv)
12
11
12
11
29
26
36
33
Chen et al. [14]
No
(i)
(ii)
(iii)
(iv)
13
12
13
12
26
23
33
30
Arai et al. [18]
Yes
(i)
(ii)
(iii)
(iv)
5
5
5
5
(13)
(12)
(13)
(12)
28
25
35
32
Proposed
Yes
(i)
(ii)
(iii)
(iv)
5
5
5
5
(11)
(11)
(11)
(11)
39
25
30
19
x[0]
x[1]
−1
x[2]
−1
x[3]
−1 −1
x[4]
−1
x[5]
−1
x[6]
−1 −1
x[7]
−1
1
2
−1
z[0]
z[0]
z[1]
z[1]
z[2]
z[2]
z[3]
z[3]
z[4]
z[4]
z[5]
z[5]
z[6]
z[6]
z[7]
1
2
1
2
(a)
x[0]
−1 x[1]
−1
−1
−1
−1
−1
−1
x[2]
x[3]
x[4]
x[5]
x[6]
x[7]
(b)
Figure 4: SFG of (a) the DC removal block and (b) the finite difference operator, both for accumulated input signal.
9
DC level. This is particularly true for both Loeffler [16] and Lee [15] algorithms which demand the same number of
multiplications as the proposed method when exact spectrum is required. Additionally, the proposed algorithm can
furnish a scaled version of the DCT spectrum at the same performance required by the Arai algorithm [18].
Under Scenario (iii), the proposed method does not demand the accumulation system, but still requires the DC
removal block. In this case, the DC removal block must be adapted to process accumulated input signals. Figure 4(a)
shows the modified DC removal block, which requires 10 additions to generate the sought sequence z[n]. In contrast,
traditional methods for DCT computation when faced with accumulated signals have to pre-process the input data
through a difference system. Figure 4(b) shows the SFG for a difference system, which is completely eliminated in
the proposed algorithm.
Scenario (iv) is a combination of Scenarios (ii) and (iii); being the most appropriate scenario for the proposed
method. In this case, the proposed algorithm does not require any pre-processing stage, being the input data directly
applicable. Thus, the proposed method is limited the operations described in the architecture shown in Figure 2.
Under this scenario, usual methods requires a pre-processing stage consisting of a difference system (Figure 4(b)),
which demands seven extra additions.
The proposed algorithm can be extended for different sequence sizes, which can benefit multiple-blocklength
encoding methods such as HEVC [12, 2]. Such extension can be derived from (3) assuming an arbitrary length N .
Considering the DCT kernel (Table 1) and applying the forward difference operator, we obtain:
∆ ker[n, k] = 2αk sin
kπ
2N
sin
kπ(n + 1)
N
,
where αk is given as in (5) and n, k = 1, 2, . . . , N −1. Despite the similarity with the DST kernel, the above expression
contains a post-multiplicative factor 2αk sin (kπ/2N ) that can be separated into a diagonal matrix. Thus the design
of the associate fast algorithm becomes equivalent to the problem of factorizing an (N − 1) × (N − 1) matrix with
entries given by sin (kπ(n + 1)/N ), n, k = 1, 2, . . . , N − 1. Such factorization can be accomplished by means of the
techniques detailed in [20] and the approach described in this paper, where the 8-point DCT is a showcase.
The proposed 1D DCT algorithm can be readily extended to the two dimensional (2D) case, thus being suitable
for image processing applications. Indeed, because of DCT kernel separability, the 2D DCT can be computed by
successive calls of the 1D DCT. This can be achieved in a two-step procedure: (i) computation of the 1D DCT of
each row of a given input image; and (ii) computation of the 1D DCT of each column of the image derived from
step (i). The resulting image is the 2D DCT of the original image. Therefore, in principle, any 1D DCT algorithm
can be immediately extended to the 2D case without any additional modification [2].
Regarding the DCT computation over 2D domain, different image representation schemes can be employed in
order to reduce the overall execution time. The works on [45, 46, 47, 48] consider a different image representation
based on slice intensity representation (ISR). Although this representation can be useful for some scenarios such
as pattern recongnition [49, 50, 51], we attain to the usual image representation adopted by the signal processing
community [20, 2, 3].
5
Remarks and Conclusion
In this paper, we proposed a new method for the 8-point DCT computation based on the summation-by-parts formula.
A matrix formalism was furnished and its arithmetic complexity was assessed.
The proposed method attained the theoretical multiplicative complexity lower bound for the computation of the
exact DCT as detailed in [32]. The theoretical minimum is consists of 11 multiplications. Per se, achieving the
minimum of 11 multiplications is not a trivial task. Not many DCT algorithms are capable of such, being the Loeffler
DCT the most popular method in literature to achieve the minimum [16].
Moreover, the introduced method could also compute the scaled DCT with only five multiplications, matching
10
the well-known Arai DCT [18]. Thus the proposed method can simultaneously match both Loeffler and Arai DCT in
terms of multiplicative complexity. Not only it could attain the multiplicative complexity minimum, but it could also
outperform several competing methods in three different scenarios when we also consider the additive complexity as
a secondary comparison metric. In fact, it outperformed several popular DCT algorithms, when the input signal is
assumed to possess null mean or/and it is an integrated (accumulated) signal.
The proposed design can be understood as a fundamental mathematical block to be considered for software and
hardware realizations. Additionally, the field of integer and approximate transforms can benefit of the proposed
scheme [52, 53]. Indeed, the design of extremely low-complexity approximation methods relies on exact algorithms.
Therefore, exact, algebraically precise methods can be a useful resource; even when an approximation could be
enough [54].
Particularly, the proposed method is an alternative to Arai algorithm [18] for selected scenarios. The suggested
method is suitable in the context of feature detection, where DC level may not be relevant [34] as well as in situations
where traditional algorithms require data to be integrated, such as in face recognition problems [35, 36]. In future
works, we aim at applying the summation-by-parts formula to different transforms at various blocklengths.
Acknowledgements
This work was partially funded by CNPq, CAPES, and FACEPE.
References
[1] A. V. Oppenheim, R. W. Schafer, M. T. Yoder, and W. T. Padgett, Discrete Time Signal Processing, vol. 1. Upper Saddle
River, NJ: Prentice-Hall, Inc., 3rd ed., Aug. 2009. 1, 2, 3, 6
[2] V. Britanak, P. Yip, and K. R. Rao, Discrete Cosine and Sine Transforms. Academic Press, 2007. 1, 2, 3, 6, 8, 10
[3] N. Ahmed and K. R. Rao, Orthogonal Transforms for Digital Signal Processing. New York,NY: Springer-Verlag, 1975. 1,
10
[4] S. K. Gupta, J. Jain, and R. Pachauri, “Improved noise cancellation in discrete cosine transform domain using adaptive
block LMS filter,” International Journal of Engineering Science and Advanced Technology, vol. 2, pp. 498–502, June 2012.
1
[5] S. An and C. Wang, “A computation structure for 2-D DCT watermarking,” in 52nd IEEE International Midwest Symposium on Circuits and Systems, pp. 577–580, Aug. 2009. 1
[6] A. A. Chien and V. Karamcheti, “Moore’s law: The first ending and a new beginning,” Computer, vol. 46, no. 12, pp. 48–53,
2013. 1
[7] M. R. Wigan and R. Clarke, “Big data’s big unintended consequences,” Computer, vol. 46, no. 6, pp. 46–53, 2013. 1
[8] L. Ji and L. X. Ming, “New DCT computation algorithm for VLIW architecture,” in 6th International Conference on
Signal Processing, vol. 1, pp. 41–44, 2002. 1
[9] V. Bhaskaran and K. Konstantinides, Image and Video Compression Standards. Kluwer Academic Publishers, June 1997.
1
[10] G. K. Wallace, “The JPEG still picture compression standard,” IEEE Transactions on Consumer Electronics, vol. 38,
pp. xviii–xxxiv, Feb. 1992. 1
[11] T. Wiegand, G. J. Sullivan, G. Bjontegaard, and A. Luthra, “Overview of the H.264/AVC video coding standard,” IEEE
Transactions on Circuit and Systems for Video Technology, vol. 13, pp. 560–576, July 2003. 1
[12] M. T. Pourazad, C. Doutre, M. Azimi, and P. Nasiopoulos, “HEVC: The new gold standard for video compression: How
does HEVC compare with H.264/AVC?,” IEEE Consumer Electronics Magazine, vol. 1, pp. 36–46, July 2012. 1, 10
[13] K. R. Rao, D. N. Kim, and J. J. Hwang, Video coding standards: AVS China, H.264/MPEG-4 PART 10, HEVC, VP6,
DIRAC and VC-1. Arlington, TX, USA: Springer, 2014. 1
[14] H. W.-H. Chen, C. Smith, and S. Fralick, “A fast computational algorithm for the discrete cosine transform,” IEEE
Transactions on Communications, vol. 25, pp. 1004–1009, Jan. 2003. 1, 9
[15] B. Lee, “A new algorithm to compute the discrete cosine transform,” IEEE Transactions on Acoustics, Speech and Signal
Processing, vol. 32, pp. 1243–1245, Dec. 1984. 1, 9, 10
[16] C. Loeffler, A. Ligtenberg, and G. S. Moschytz, “A practical fast 1-D DCT algorithms with 11 multiplications,” in
International Conference on Acoustics, Speech, and Signal Processing, vol. 2, pp. 988–991, May 1989. 1, 4, 9, 10
11
[17] E. Feig and S. Winograd, “Fast algorithms for the discrete cosine transform,” IEEE Transactions on Signal Processing,
vol. 40, pp. 2174–2193, Sept. 1992. 1
[18] Y. Arai, T. Agui, and M. Nakajima, “A fast DCT-SQ scheme for images,” IEICE Transactions, vol. E71, pp. 1095–1097,
Nov. 1988. 1, 8, 9, 10, 11
[19] R. W. Hamming, Digital Filters. Hertfordshire, UK: Prentice Hall International Ltd., third ed., 1989. 1, 3
[20] R. E. Blahut, Fast Algorithms for Digital Signal Processing. Cambridge University Press, 2010. 1, 5, 6, 8, 10
[21] X. Limin, X. Weisheng, and Y. Youling, “A fast harmonic detection method based on recursive DFT,” in Electronic
Measurement and Instruments, 2007. ICEMI ’07. 8th International Conference on, pp. 3–972, 2007. 1
[22] E. Zheng, Z. Liu, and L. Ma, “Study on harmonic detection method based on FFT and wavelet transform,” in 2nd
International Conference on Signal Processing Systems (ICSPS), vol. 3, 2010. 1, 5
[23] R. L. Graham, D. E. Knuth, and O. Patashnik, Concrete Mathematics. Reading, MA: Addison-Wesley Publishing Company,
1989. 2, 5
[24] T. M. Apostol, Mathematical Analysis. Reading, MA: Addison-Wesley Publishing Company, 1981. 2
[25] K. Mattssona and J. Nordstrma, “Summation by parts operators for finite difference approximations of second derivatives,”
Journal of Computational Physics, vol. 199, pp. 503–540, Sept. 2004. 2
[26] K. Mattsson, “Boundary procedures for summation-by-parts operators,” Journal of Scientific Computing, vol. 18, pp. 133–
153, Feb. 2003. 2
[27] J. R. Mosig and A. A. Melcon, “The summation-by-parts algorithm - A new efficient technique for the rapid calculation
of certain series arising in shielded planar structures,” IEEE Transactions on Microwave Theory and Techniques, vol. 50,
pp. 215–218, Jan. 2002. 2
[28] G. Boudreaux-Bartels and T. W. Parks, “Discrete Fourier transform using summation by parts,” in IEEE International
Conference on Acoustics, Speech, and Signal Processing, vol. 12, pp. 1827–1830, Apr. 1987. 2
[29] G. Boudreaux-Bartels, D. W. Tufts, P. Dhir, G. Sadasiv, and G. Fischer, “Analysis of errors in the computation of
Fourier coefficients using the arithmetic Fourier transform (AFT) and summation by parts (SBP),” in IEEE International
Conference on Acoustics, Speech, and Signal Processing, vol. 2, pp. 1011–1014, May 1989. 2
[30] R. J. Cintra, “Soma por partes para sinais periódicos e transformadas discretas,” in XXX Simpósio Brasileiro de Telecomunicações - SBrT12, (Brası́lia, DF), Sept. 2012. 2
[31] P. Duhamel and H. H’Mida, “New 2n DCT algorithms suitable for VLSI implementation,” in IEEE International Conference on Acoustics, Speech, and Signal Processing, vol. 12, pp. 1805–1808, 1987. 2, 8
[32] M. T. Heideman, Multiplicative Complexity, Convolution, and the DFT. New York: Springer-Verlag, 1988. 2, 4, 8, 10
[33] A. K. Jain, Fundamentals of Digital Image Processing. University of California, Davis: Prentice Hall Information and
System Sciences Series, 1989. 2, 8
[34] Y. Wang, C.-S. Chua, and Y.-K. Ho, “Facial feature detection and face recognition from 2D and 3D images,” Pattern
Recognition Letters, vol. 23, no. 10, pp. 1191–1202, 2002. 2, 8, 11
[35] Elboher and M. Werman, “Efficient and accurate Gaussian image filtering using running sums,” in 12th International
Conference on Intelligent Systems Design and Applications (ISDA), pp. 897–902, 2012. 2, 8, 11
[36] P. Viola and M. Jones, “Rapid object detection using a boosted cascade of simple features,” in Computer Vision and
Pattern Recognition, 2001. CVPR 2001. Proceedings of the 2001 IEEE Computer Society Conference on, vol. 1, 2001. 2,
8, 11
[37] R. C. Gonzalez and R. E. Woods, Digital Image Processing. Prentice-Hall, Inc., 3 ed., 2007. 2, 8
[38] H. El-Banna, A. A. El-Fattah, and W. Fakhr, “An efficient implementation of the 1D DCT using FPGA technology,” in
Engineering of Computer-Based Systems, 2004. Proceedings. 11th IEEE International Conference and Workshop on the,
pp. 356–360, 2004. 4
[39] S. J. Shan, C. Chen, and E. Yang, “High performance 2-D IDCT for image/video decoding based on FPGA,” in 2012
International Conference on Audio, Language and Image Processing (ICALIP), pp. 33–38, 2012. 4
[40] H. K. C.-K. Fong and W.-K. Cham, “LLM integer cosine transform and its fast algorithm,” IEEE Transactions on Circuits
and Systems for Video Technology, vol. 22, no. 6, pp. 844–854, 2012. 4
[41] M. Abramowitz and I. A. Stegun, eds., Handbook of Mathematical Functions: with Formulas, Graphs, and Mathematical
Tables. MIT Press, 1964. 5
[42] Mukherjee and S. K. Mitra, “Enhancement of color images by scaling the DCT coefficients,” IEEE Transactions on Image
Processing, vol. 17, no. 10, pp. 1783–1794, 2008. 5
[43] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, “A multiplication-free transform for image compression,” in 2nd
International Conference on Signals, Circuits and Systems, (Monastir, TN), pp. 1–4, 2008. 5
[44] V. Dimitrov and K. A. Wahid, “On the error-free computation of fast cosine transform,” Information Theories and
Applications, vol. 12, no. 4, pp. 321–327, 2005. 5
[45] G. A. Papakostas, D. E. Koulouriotis, and E. G. Karakasis, Image Processing, ch. 2, Efficient 2-D DCT Computation from
an Image Representation Point of View. INTECH, Sept. 2009. 10
12
[46] G. A. Papakostas, E. G. Karakasis, and D. E. Koulouriotis, “Efficient and accurate computation of geometric moments on
gray-scale images,” Pattern Recognition, vol. 41, no. 6, pp. 1895–1904, 2008. 10
[47] G. A. Papakostas, E. G. Karakasis, and D. E. Koulouriotis, “On accelerating the computation of 2-D discrete cosine
transform in image processing,” in International Conference on Signals and Electronic Systems, pp. 7–10, Sept. 2008. 10
[48] G. A. Papakostas, E. G. Karakasis, and D. E. Koulouriotis, “Exact and speedy computation of legendre moments on
binary images,” in Proceedings of the Eight International Workshop on Image Analysis for Multimedia Interactive Services
(WIAMIS), (Washington, DC, USA), pp. 48–48, IEEE Computer Society, 2007. 10
[49] M. Hassaballah, A. A. Abdelmgeid, and H. A. Alshazly, Image Feature Detectors and Descriptors, ch. Image Features
Detection, Description and Matching, pp. 11–45. Cham: Springer International Publishing, 2016. 10
[50] M. Sayyouri, A. Hmimid, and H. Qjidaa, “A fast computation of hahn moments for binary and gray-scale images,” in
IEEE International Conference on Complex Systems (ICCS), pp. 1–6, Nov. 2012. 10
[51] A. Hmimid, M. Sayyouri, and H. Qjidaa, “A fast method for reconstruction of binary and gray-scale images by the
tchebichef moments,” in Colloquium in Information Science and Technology, pp. 106–111, Oct. 2012. 10
[52] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, “Binary discrete cosine and hartley transforms,” IEEE Transactions
on Circuits and Systems I, vol. 60, pp. 989–1002, Apr. 2013. 11
[53] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, “A low-complexity parametric transform for image compression,” in
IEEE International Symposium on Circuits and Systems, (Rio de Janeiro, BR), pp. 2145–2148, May 2011. 11
[54] C. J. Tablada, F. M. Bayer, and R. J. Cintra, “A class of DCT approximations based on the Feig-Winograd algorithm,”
Signal Processing, 2015. 11
13
| 8cs.DS
|
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
REVIEW OF FACE DETECTION SYSTEMS BASED
ARTIFICIAL NEURAL NETWORKS ALGORITHMS
Omaima N. A. AL-Allaf
Assistant Professor,
Faculty of Sciences & IT,
Al-Zaytoonah University of Jordan, P.O. Box130, Amman, (11733), Jordan
ABSTRACT
Face detection is one of the most relevant applications of image processing and biometric systems.
Artificial neural networks (ANN) have been used in the field of image processing and pattern recognition.
There is lack of literature surveys which give overview about the studies and researches related to the using
of ANN in face detection. Therefore, this research includes a general review of face detection studies and
systems which based on different ANN approaches and algorithms. The strengths and limitations of these
literature studies and systems were included also.
KEYWORDS
Face Detection, Face Recognition, Artificial Neural Networks
1. INTRODUCTION
In the past few years, face recognition has received a significant attention and regarded as one of
the most successful applications in the field of image analysis [1]. The human faces represent
complex, multidimensional, meaningful visual stimulant. Developing a computational model for
face recognition is difficult [2]. Face detection can be regarded as fundamental part of face
recognition systems according to its ability to focus computational resources on the part of an
image containing a face. The process of face detection in images is complex because of
variability present across human faces such as: pose; expression; position and orientation; skin
color; presence of glasses or facial hair; differences in camera gain; lighting conditions; and
image resolution [3]. The analysis of facial expression was primarily a research field for
psychologists in the past years [4]. At the same time, advances in many domains such as: face
detection [5][6]; tracking[7]; and recognition [1]; pattern recognition; and image processing
contributed significantly to research in automatic facial expression recognition.
Face detection should be performed before recognition system. This is done to extract relevant
information for face and facial expression analysis. Two classes of techniques for face
representation and relevant information extraction. And geometrical feature extraction relies on
parameters of distinctive features such as eyes, mouth and nose. At the same time, a face is
represented as an array of pixel intensity values suitably pre-processed in appearance based
approaches (texture). This array is then compared with a face template using a suitable metric [4].
Research [8] compared the performances of these representation techniques in face recognition.
Therefore, according to the complexity of face detection process, many applications based on
human face detection have been developed recently such as surveillance systems, digital
monitoring, intelligent robots, notebook, PC cameras, digital cameras and 3G cell phones. These
DOI : 10.5121/ijma.2013.6101
1
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
applications play an important role in our life. Nevertheless, the algorithms of the applications are
complicated and hard to meet real-time requirements of specific frame-rate [9]. Over the past
decade, many approaches for improving the performance of face detection have been proposed
[9][10][11][12][13][14][15][16][17][18][19]. At the same time, many literature studies focused
on survey on face detection techniques [20][21][6][22][23].
Artificial neural networks (ANN) were used largely in the recent years in the fields of image
processing (compression, recognition and encryption) and pattern recognition. Many literature
researches used different ANN architecture and models for face detection and recognition to
achieve better compression performance according to: compression ratio (CR); reconstructed
image quality such as Peak Signal to Noise Ratio (PSNR); and mean square error (MSE). Few
literature surveys that give overview about researches related to face detection based on ANN.
Therefore, this research includes survey of literature studies related to face detection systems and
approaches which were based on ANN.
The rest of this paper is organized as follows: Section 2 includes the main steps of face detection
and recognition. Section 3 includes literature studies related to face detection systems based on
ANN. Section 4 includes comparisons between these literature studies. Section 5 includes
recommendations. Finally section 6 concludes this work.
2. FACE DETECTION AND RECOGNITION
A general face recognition system includes many steps: face detection; feature extraction; and
face recognition [14][24] as shown in Figure1. Face detection and recognition includes many
complementary parts, each part is a complement to the other. Depending on regular system each
part can work individually. Face detection is a computer technology that is based on learning
algorithms to allocate human faces in digital images [25].
Face detection takes images/video sequences as input and locates face areas within these images.
This is done by separating face areas from non-face background regions. Facial feature extraction
locates important feature (eyes, mouth, nose and eye-brows) positions within a detected face.
Feature extraction simplifies face region normalization where detected face aligned to coordinate
framework to reduce the large variances introduced by different face scales and poses. The
accurate locations of feature points sampling the shape of facial features provide input parameters
for the face identification. Other face analysis task: facial expression analysis [16]; face animation
and face synthesis can be simplified by accurate localization of facial features [16][24].
Face identification generates the final output of complete face-recognition system: the identity of
the given face image. Based on normalized face image and facial feature locations derived from
previous stages, a feature vector is generated from given face and compared with a database of
known faces. If a close match is found, the algorithm returns the associated identity. A main
problem in face identification is the large differences between face images from the same person
as compared to those from different persons. Therefore, it is important to choose a suitable face
classification technique that can provide a good separate ability between different persons. Face
identification has a wide range of applications. Because it offers a non-intrusive way for human
identification, the face is used as an important biometric in security applications.
Recently, face recognition has received wide interest in number of countries are integrating facial
information into electronic passport and to other biometrics (fingerprints and iris) [26]. In
addition to security and law enforcement, face recognition is also applied in entertainment and
consumer electronics as a means for a natural user interface. By recognizing the existence of the
user and his identity, consumer devices can offer customized services, thereby creating an
2
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
enhanced user experience. To achieve high-performance face recognition system, each processing
stage in the system has to be designed to satisfy application requirements [14]. Face recognition
involves comparing an image with a database of stored faces to identify individual in input image.
The related task of face detection has direct relevance to face recognition because images must be
analysed and faces identified, before they can be recognized. Detecting faces in an image can help
to focus the computational resources of the face recognition system, optimizing the systems speed
and performance [3].
Figure 1. Framework of a face-recognition system [14]
The main steps of face detection system are shown in Figure2. Face detection separate image
windows into two parts: one containing faces, and one containing the background. The process is
difficult because the: commonalities exist between faces (vary in terms of age, skin color and
facial expression); and also differing in: lighting conditions; image qualities; and geometries. The
face detector would be able to detect the presence of any face under any set of lighting conditions,
upon any background.
Figure 2. A general face detection system [3]
3. ARTIFICIAL NEURAL NETWORKS FOR FACE DETECTION
In the recent years, different architectures and models of ANN were used for face detection and
recognition. ANN can be used in face detection and recognition because these models can
simulate the way neurons work in the human brain. This is the main reason for its role in face
recognition. This research includes summery review of the researches related to face detection
based on ANN.
3
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
3.1. Retinal Connected Neural Network (RCNN)
Rowley, Baluja and Kanade (1996) [27] presented face detection system based on a retinal
connected neural network (RCNN) that examine small windows of an image to decide whether
each window contains a face. Figure 3 shows this approach. The system arbitrates between many
networks to improve performance over one network. They used a bootstrap algorithm as training
progresses for training networks to add false detections into the training set. This eliminates the
difficult task of manually selecting non-face training examples, which must be chosen to span the
entire space of non-face images. First, a pre-processing step, adapted from [28], is applied to a
window of the image. The window is then passed through a neural network, which decides
whether the window contains a face. They used three training sets of images. Test SetA collected
at CMU: consists of 42 scanned photographs, newspaper pictures, images collected from WWW,
and TV pictures (169 frontal views of faces, and require ANN to examine 22,053,124 20×20
pixel windows). Test SetB consists of 23 images containing 155 faces (9,678,084 windows). Test
SetC is similar to Test SetA, but contains images with more complex backgrounds and without
any faces to measure the false detection rate: contains 65 images, 183 faces, and 51,368,003
window. The detection ratio of this approach equal 79.6% of faces over two large test sets and
small number of false positives.
Figure 3. RCNN for face detection [27]
3.2. Rotation Invariant Neural Network (RINN)
Rowley, Baluja and Kanade (1997) [29] presented a neural network-based face detection system.
Unlike similar systems which are limited to detecting upright, frontal faces, this system detects
faces at any degree of rotation in the image plane. Figure 4 shows the RINN approach. The
system employs multiple networks; the first is a “router” network which processes each input
window to determine its orientation and then uses this information to prepare the window for one
or more detector networks. We present the training methods for both types of networks. We also
perform sensitivity analysis on the networks, and present empirical results on a large test set.
Finally, we present preliminary results for detecting faces which are rotated out of the image
plane, such as profiles and semi-profiles [29].
4
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
Figure 4. RINN for face detection [29]
3.3. Principal Component Analysis with ANN (PCA & ANN)
Jeffrey Norris (1999) [30] used using principal component analysis (PCA) with class specific
linear projection to detect and recognized faces in a real-time video stream. Figure 5 shows PCA
& ANN for face detection. The system sends commands to an automatic sliding door, speech
synthesizer, and touchscreen through a multi-client door control server. Matlab, C, and Java were
used for developing system. The system steps to search for a face in an image:
1.
2.
3.
4.
5.
Select every 20×20 region of input image.
Use intensity values of its pixels as 400 inputs to ANN.
Feed values forward through ANN, and
If the value is above 0.5, the region represents a face.
Repeat steps (1..4) several times, each time on a resized version of the original input
image to search for faces at different scales [30].
Figure 5. PCA & ANN for face detection [30]
3.4. Fast Neural Networks (FNN)
Hazem El-Bakry (2002) [31] proposed fast neural networks (FNN) approach to reduce the
computation time for locating human faces. Each image is divided into small sub images and then
each one is tested separately using a fast ANN. The experimental results of comparison with
conventional neural networks showed that the high speed is achieved when applying FNN.
5
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
3.5. Polynomial Neural Network (PNN)
Lin-Lin Huang, et al. (2003) [32] proposed face detection method using a polynomial neural
network (PNN). The local regions in multi scale sliding windows are classified by the PNN to
two classes (face and non-face) to locate the human faces in an image. The PNN takes the
binomials of projection of local image onto a feature subspace learned by principal component
analysis (PCA) as inputs. They investigated the influence of PCA on either the face samples or
the pooled face and non-face samples.
3.6. Convolutional Neural Network (CNN)
Masakazu Matsugu (2003) [33] described a rule-based algorithm for robust facial expression
recognition combined with face detection using a convolutional neural network (CNN). Figure 6
shows the CNN approach. The problem of subject independence and translation, rotation, and
scale invariance in facial expression recognition were addressed in this study.
3.7. Evolutionary Optimization of Neural Networks
Stefan, et al (2004) [34] used ANN to get decision whether a pre-processed image region
represents a human face or not. They described the optimization of this network by a hybrid
algorithm combining evolutionary computation and gradient-based learning. The evolved
solutions perform considerably faster than an expert-designed architecture without loss of
accuracy. The proposed hybrid algorithm tackles the problem of reducing the number of hidden
neurons of face detection network without loss of detection accuracy. The speed of classification
whether an image region corresponds to a face or not could be improved by approximately 30 %.
Figure 7 shows evolutionary optimization of ANN.
Figure 6. CNN for face detection [33]
6
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
(a)
(b)
Figure 7. (a): Hybrid algorithm. (b): Visualization of input and field connectivity [34]
.
3.8. Multilayer Perceptron (MLP)
According to Rowley work in [27], Marian Beszedes & Milos Oravec (2005) [35] presented a
neural network based face detection system to detect faces in an unprocessed input image. Figure
8 shows the MLP approach. They used image processing techniques such as normalization,
rotation and position, light conditions improvement on small windows extracted from the input
image. Multilayer perceptron (MLP) used to detect rotation of input window and also to decide
whether the window contains a face or not. This system is based on method to distribute the
decision among multiple sub networks and an algorithm is used to train this ANN and the result
of this system is in the form of a set containing locations of human faces. The training/testing set
for includes 250 face and 250 randomly generated non-face samples. The training/testing sets
were extended according to boosting method by 2000 and 2500 non- face samples. These samples
were obtained as false face detections from testing of ANN trained on basic sets on images
containing no faces. Their system succeeded in 90100 percent when they consider the number of
faces in input picture. Only 18 from this amount were not classified correctly. Only three faces
from the total number 38 were not found.
Shilbayeh and Al-Qudah (2008) [36] proposed a face detector based multi layer perceptron
(MLP) ANN and maximal rejection classifier (MRC) to improve efficiency of detection in
comparison with traditional ANN. The ANN was organized to reject a majority of non-face
patterns in image backgrounds by improving detection efficiency, reducing computation cost and
maintaining the detection accuracy.
7
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
Figure 8. MLP of face localization system [35]
3.9. Back Propagation Neural Networks (BPNN)
Zoran and Samcovic (2006) [37] used ANN for face detection for video surveillance. The ANN
is trained with multilayer back propagation neural networks (BPNN). Three face representations
were taken (pixel, partial profile and eigenfaces) representation. Three independent sub-detectors
are generated based on these three face representation. Figure 9 shows the BPNN approach.
Aamer Mohamed, et al (2008) [38] proposed face detection system based on BPNN via Gaussian
mixture model to segment image based on skin color. In this approach start from skin and non
skin face candidates' selection. After that the features are extracted from discrete cosine transform
(DCT) coefficients. Based on DCT feature coefficients in Cb and Cr color spaces, BPNN was
used to train and classify faces. The BPNN used to check if the image include face or not. DCT
feature values of faces that represent the data set of skin/non-skin face candidates obtained from
Gaussian mixture model are fed into BPNN to classify whether original image includes a face or
not.
3.10. Gabor Wavelet Faces with ANN
Sahoolizadeh et al (2008) [39] proposed hybrid approaches for face recognition based on
combined Gabor wavelet faces with ANN feature classifier. The Gabor wavelets used to represent
face image. The representation of face images using Gabor wavelets is effective for facial action
recognition and face identification.
They reduced dimensionality and linear discriminate analysis on down sampled Gabor wavelet
faces can increase the discriminate ability. Nearest feature space is extended to various similarity
measures shows good performance which achieve 93% recognition rate on ORL data set.
8
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
Figure 9. BPNN architecture for face detection [37]
Avinash and Raina (2010) [40] presented face detection approach with Gabor wavelets transform
& feed forward neural network for finding feature points and extracting feature vectors. Gabor
filter used for feature extraction for face detection. The classifier (FFNN) take the feature vectors
as input. The location of feature points contains information about the face in this approach. The
graph is constructed from the general face idea. Instead of fitting this graph, the feature points are
obtained from the characteristics of each face automatically. Facial features allow to make a
decision from face parts because the facial features are compared locally instead of using a
general structure. Two measures were used in this study : false negative and false positive. These
two measures can be calculated using Equation 1 and Equation 2 as follows [39]:
݁ݒ݅ݐܽ݃݁ܰ ݁ݏ݈ܽܨ
ܰݏ݁ܿܽܨ ݀݁ݏݏ݅ܯ ݂ ݎܾ݁݉ݑ
=
ܶݏ݁ܿܽܨ ݈ܽݑݐܿܣ ݂ ݎܾ݁݉ݑܰ ݈ܽݐ
…………………………………………(1)
݁ݒ݅ݐ݅ݏܲ ݁ݏ݈ܽܨ
ܰݏ݁ܿܽܨ ݀݁ݐܿ݁ݐ݁ܦ ݐܿ݁ݎݎܿ݊ܫ ݂ ݎܾ݁݉ݑ
=
ܶݏ݁ܿܽܨ ݈ܽݑݐܿܣ ݂ ݎܾ݁݉ݑܰ ݈ܽݐ
…………………………………………(2)
Mohammad Abadi, et al, (2011) [41] proposed approach based on ANN and Gabor wavelets to
detect desirable number of faces in fixed photo with gray background. They used correlation of a
window with a face with photo. Then they estimated areas of candidate of face presence. After
that, they used step algorithm and referred these areas and around them to section of extraction of
Gabor wavelets characteristics and neural network classifier. The resultant areas lead to detection
of face locations in photo. They examined the result of estimation of efficiency of this method by
different tests. The method is simulated in MATLAB. They used 70 face photos and 60 non face
photos in training phase. Every face photo, its mirror photo and with the angle of 5,10,15 degrees
in positive and negative directions and photos with one pixel shift in every 4 directions are placed
in training set for reducing network sensitivity. For no face photos also, their mirror and their 180
degrees transformation is placed in training data. They obtained 5% right answer, error limit of
0.0001, false negative error= 5%. And for tested image with size=254×600, the positive false=12,
and detection= 56 from 57. Also for test image of size = 50×100, the positive false=0, and
detection= 2 from 3.
9
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
Anissa Bouzalmat, et al (2011) [42] presented BPNN for face recognition. The BPNN input is
feature vector based on Fourier Gabor filters. They used an algorithm for detecting face regions in
images using the color of skin which presents overlooked in different background, accessory and
clothing. After that, they introduced Gabor filters with 8 orientations and 5 resolutions to get
maximum information (to extract the maximum of information by varying the resolution and
orientation). This is done to generate and extract the features vector of the whole face in image.
BPNN is then applied to perform the recognition task. This solution was implemented using Java
environment. Results indicate that the proposed method achieves good results. Figure 10 shows
the BPNN with Gabor Wavelet for face detection.
Figure 10. Description of the proposed solution architecture [42]
3.11. Skin Color and BPNN
The construction of robust real-time face detection system can be regarded as one of the most
practical applications under vigorous development. Kalavdekar Prakash (2010) [43] described
face detection system that process images based neural network to detect face images and
achieving high detection rates. Analyzing a video sequence is the current challenge since faces
are constantly in dynamic motion, presenting many different possible rotational and illumination
conditions. While solutions to the task of face detection have been presented, detection
performances of many systems are dependent upon environment. Their suggested system
includes: skin color filter; image filtering; MLP; and detection. Figure 11 shows Skin Color and
BPNN for face detection. Face detection attempt to locate all regions that contain a face in given
still image or an image sequence. There are two main solutions for face detection: feature-based
and image-based approaches.
Elmansori and Khairuddin (2011) [44] presented face detection method combines two algorithms:
Skin color based face detector and BPNN. The Skin color based face detector used modeling the
distribution of skin color to identify areas most likely to be regions of skin to identify potential
areas of skin by equalizing probability of likelihood. The problem space is linearly separable and
a linear threshold function is offered for the solution which is supported by a sparse feature
mapping architecture. BPNN is used to represent function using arbitrary decision surfaces by
utilizing nonlinear activation functions. Their experiments showed that the methods show closer
performances for the classification in face and non-face space, and the method has achieved high
detection rates and an acceptable number of false negatives and false positives. The system was
implemented using C#.
10
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
Figure 11. Skin Color and BPNN for Face Detection [43]
3.12 Cascaded Neural Network
Finally, Zuo & de With (2008) [45] proposed a fast face detector based on a hierarchical cascade
of neural network ensembles to enhance detection accuracy and efficiency. They used a number
of neural network classifiers to form a neural network ensemble. Each classifier is specialized in a
sub region in the face-pattern space. These classifiers complement each other to perform the
detection task. Then, they organized the neural network ensembles in a pruning cascade to reduce
the total computation cost of face detection.
In this stage, simpler and more efficient ensembles used at earlier stages in the cascade are able to
reject a majority of non face patterns in image backgrounds by improving the overall detection
efficiency while maintaining detection accuracy. Their results showed that the proposed neuralnetwork ensembles improve the detection accuracy as compared to traditional ANN. Their
approach reduced training and detection cost by achieving detection rate equal 94%.
4. COMPARISONS BETWEEN DIFFERENT ANN APPROACHES
Literature studies based on ANN for building face detection systems were described in this study.
Each one of these studies was based on special architecture of ANN for face detection. Many of
these studies were based on one architecture of ANN such as : Multilayer Perceptron (MLP);
BackPropagation Neural Networks (BPNN); Retinal Connected Neural Network (RCNN);
Rotation Invariant Neural Network (RINN); Fast Neural Networks (FNN); Convolutional Neural
Network; Polynomial Neural Network (PNN). Other studies were based on ANN on combination
with other techniques and methods such as Principal Component Analysis with ANN (PCA &
ANN); Evolutionary Optimization of Neural Networks; Gabor Wavelet Faces with ANN; and
finally Skin Color and BPNN.
All of these studies were based on ANN. Each one of these studies includes its own experiments
and based on different database for training and testing images. Many of these studies take
detection rate as performance measure; other studies take error rate as performance measure and
so on. Other studies did not talk exactly about the used database.
On the other hand, each of these literature studies has its own strengths and limitations. We
cannot exactly determine what the best topology that can be used for face detection system with
high performance. Table 1 shows information about topology, database of training/testing images
and performance of face detection systems which were taken from many of these literature
studies.
11
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
Table 1 shows that the literature studies used different data bases for image training and testing
set. Other studies were not included in table 1 because they weren’t use known database and just
used camera image samples. The number of samples is different from one study to another.as
example the research [38] take only 50 image samples. Whereas other studies such as [27][32]
used more than one image sets as samples. We can note from table 1 also that the studies adopted
different image dimensions: (384×384), (92×112) and so on.
Table 1. DB and Performance measures used in literature studies
Research
[27]
Topology
Retinal
connected
neural
network
[30]
PCA with
ANN
[32]
PNN
Data Base: Training & Testing
Performance
Three training sets of images.
Test SetA: 42 scanned photographs
Test SetB: 23 images contain 155 faces
Test SetC: 65 images, 183 faces (images with
more complex backgrounds and without faces to
measure false detection )
Detect 78.9% - 90.5% of
faces in a set of 130 test
images
Select 700 pictures in Kah-Kay Sung’s data set of
1488 faces to train ANN with 700 random noise
pictures as negative examples remaining 788
faces in Kah-Kay’s data set, followed by 788
random noise pictures
First set: 3257 images downloaded from several
websites (384×384), with one face in each image.
Second set: 130 images downloaded from
website of CMU
Acceptable number of
false detections.
1.2% error after training
for 50 epochs
1566 examples, 35
mis classifications made
(2.23% error).
Detection rate = 84.6%
False rate=3:51 × 10−6
CNN
training of CNN, the number of facial fragment
images used is 2900 for the FD2 layer,
5290 for the FD3, and 14,700 (face) for the FD4
layer, respectively.
Number of non-face images, also used for the
FD4 layer, is 137.
Recognition
rate
=
97.6% for 5600 still
images of more than 10
subjects
[37]
BPNN
Training set contains 12000 face images collected
from various face DBs.
These samples also include the scaled versions at
the same face with factor (0.8 - 1.12)
[39]
Gabor
wavelet
with ANN
ORL dataset: 400 frontal faces: 10 tightly cropped
(92×112) with 256 grey images of 40 individuals
with variations in pose, illumination, ..etc
Detection
rates
measured for separate
test set of 500 faces and
4000 non-faces.
Performance=94%.
Detect (77.9% - 90.3%)
of faces in a set of 130
test images
[33]
[36]
[38]
MLP and
MRC
BPNN
Training set: face images from MIT DB. Images
(scaled to 20×20)
Test set: 2000 face/non-face images from MIT
DB.
Non-face patterns generated at different locations
and scales.
50 real images taken under different lighting
conditions (digital camera images and web images
from several websites).
Detection rate = 91.6%
Error rate = 7.54%
Detect 97.3% of
faces in a set of 50 real
images.
Processing Time (s) of
image (63×180)= 3.8
Processing Time (s) of
image (200×219)= 6.2
12
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
Finally we can note from table 1 that the factors used as performance evaluations are different
from one study to other. Many studies used detection rate, others used false rate, and so on.
Figure 12 shows the detection rate of different ANN approaches from many studies. We can note
from Figure 12 that the highest face detection rate can be obtained from using the CNN [33]
approach. At the same time the BPNN approach adopted in [38] and [37] result in obtaining good
and high detection rate.
Figure 12. Detection rate for different studies
5. RECOMMENDATIONS FOR FACE DETECTION SYSTEM
Face detection is first step in face recognition systems to localize and extract the face region from
the image background. The literature studies related to face detection systems which were based
on ANN were described earlier in this research can be summarized as follows:
• The face detection techniques presented above were based on 2D image detection.
• Many of these literature researches did not give overview about the exactly used database for
system training and testing.
• Many of these researches did not give sufficient information about performance measures used
for face detection. There is lack of equations related to these performance measures.
• Most of these studies of face detection systems were adopted ANN in combination with other
approaches and algorithms to obtain better results for detection and improve the performance
of face detection system. But this may increase the system complexity, required memory and
time for face detection.
• Lack of using other significant ANN architectures such as self-organizing map, PatternNet
ANN, Fast BPNN and so on.
• Lack of literature related to face detection based on combination of ANN and genetic
algorithm (GA).
According to above points, many recommendations must be taken in our consideration when we
suggest to build a strong face detection system such as:
• Try to design a real time face detection system that is based on video taken in real time
camera.
• Give sufficient details about the exactly used database for system training and testing.
• Give sufficient details about performance measures and equations used for face detection.
13
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
• ANN can be adopted in combination with other algorithms to obtain better results for face
detection. At the same time, we must focus on how to simplify the combined algorithms steps
to reduce the memory required and processing time.
• Try to use other ANN architectures: self-organizing map, PatternNet, FitNet and Fast BPNN.
• Try to use different optimization ANN training algorithms such as: Levenberg-Marquardt
(TRAINLM); TRAINBFG; Bayesian regularization (TRAINBR); TRAINCGF algorithm;
Gradient descent (TRAINGD); and Gradient descent with momentum (TRAINGDM) [46] to
obtain best results for face detection system.
• Try to use genetic algorithm (GA) as an optimization algorithm to obtain the best values of
ANN algorithm parameters that result to optimal results.
6. CONCLUSION
This paper includes a summary review of literature studies related to face detection systems based
on ANNs. Different architecture, approach, programming language, processor and memory
requirements, database for training/testing images and performance measure of face detection
system were used in each study. Each study has its own strengths and limitations.
In future work, a face detection system will be suggested based on using Pattern Net and Back
propagation neural network (BPNN) with many hidden layers. Different network architectures
and parameters’ values of BPNN and PatternNet will be adopted to determine PatternNet
architecture that will result in best performance values of face detection system.
ACKNOWLEDGEMENTS
The author would like to thanks Al-Zaytoonah university of Jordan- Amman- Jordan for
supporting this research.
REFERENCES
[1]
W. Zhao, et al (2000) “Face recognition: a literature survey”, Technical Report CAR-TR-948,
University of Maryland, October 2000.
[2] Turk M & Pentland A (1991) “Eigenfaces for recognition”, Journal of Cognitive Neuroscience, Vol.3,
pp71–86.
[3] Phil Brimblecombe (2002) “Face Detection using Neural Networks”, H615 – Meng Electronic
Engineering, School of Electronics and Physical Sciences, URN: 1046063.
[4] Bouchra Abboud, et al (2004) “Facial expression recognition and synthesis based on an appearance
model”, Signal Processing: Image Communication, Vol. 19, Issue. 8, pp723-740.
[5] P. Viola & M.J. Jones (2001) “Robust real-time object detection”, Technical Report CRL/2001/01,
Cambridge Research Laboratory, USA, February 2001
[6] Ming-Hsuan Yan, et al (2002), “Detecting Faces in Images: A Survey”, IEEE Transactions on Pattern
Analysis and Machine Intelligence, Vol. 24, No. 1, pp34-58, January2002
[7] Minyoung Kim, et.al (2008) “Face tracking and recognition with visual constraints in real-world
videos”, IEEE Conference on Computer Vision and Pattern Recognition, pp23-28, June.
[8] Brunelli R & Poggio T (1993) “Face recognition: features versus templates”, IEEE Transaction
Pattern Analysis and Machine Intelligence, Vol. 15, No.10, pp1042–1052.
[9] Yao-Jiunn Chen & Yen-Chun Lin (2007) “Simple Face-detection Algorithm Based on Minimum
Facial Features”, The 33rd Annual Conference of the IEEE Industrial Electronics Society (IECON) 58 Nov 2007, Taipei, Taiwan, pp455-460.
[10] Sanjay Singh, et. Al (2003) “A Robust Skin Color Based Face Detection Algorithm”, Tamkang
Journal of Science and Engineering, Vol. 6, No. 4, pp.227-234 .
14
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
[11] Abdenour Hadid, Matti Pietik¨ainen & Timo Ahone (2004) “A Discriminative Feature Space for
Detecting and Recognizing Face”, Proceedings of the 2004 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, Vol. 2.
[12] Elise Arnaud et al (2005) “A Robust And Automatic Face Tracker Dedicated To Broadcast Videos,
IEEE International Conference On Image Processing.
[13] Zhonglong Zheng, Jie Yang & Yitan Zhu (2006) “Face Detection and Recognition using Colour
Sequential Images”, Journal of Research and Practice in Information Technology, Vol. 38, No. 2,
pp.135-149, May 2006.
[14] Fei Zuo (2006) Embedded Face Recognition Using Cascaded Structures, Thesis, Technische
Universiteit Eindhoven, China.
[15] Zuo F. and P.H.N. de With (2002) “Automatic Human Face Detection for a Distributed Video
Security System”, Proceedings of the Progress workshop on Embedded Systems, pp269–274, Oct.
[16] Stan Z. Li & Anil K. Jai (2005) Handbook of Face Recognition, Springer Science &Business Media.
[17] Jun-Su Jang & Jong-Hwan Kim (2008) “Fast and Robust Face Detection Using Evolutionary
Pruning”, IEEE Transactions on Evolutionary Computation, pp1-10.
[18] Bernd Menser & Michael Brunig (2000) “Face Detection and Tracking for Video Coding
Application”, Conference Record of the Thirty-Fourth Asilomar Conference on Signals, Systems and
Computers, 29Oct-1Nov2000, Pacific Grove, CA, USA
[19] Pedro Alexandre Dias Martin (2008) “Active Appearance Models for Facial Expression, Recognition
and Monocular Head Pose Estimation”, master thesis, Dept. of Electrical and Computer Eng., Faculty
of Sciences and Technology, University of Coimbra.
[20] Hjelmas and Low (2001) “Face Detection: A Survey,” Computer Vision and Image Understanding,
vol. 83, pp236-274, doi:10.1006/cviu.2001.0921, http://www.idealibrary.com
[21] Yongzhong Lu, Jingli Zhou & Shengsheng Yu (2003) “A Survey of Face Detection, Extraction and
Recognition”, Computing and Informatics, Vol. 22, pp.163-195.
[22] W. Zhao et al (2003) “Face recognition: A literature survey”, ACM Computing Surveys, Vol.35,
No.4, December 2003, pp. 399–458.
[23] Cha Zhang & Zhengyou Zhang (2010) “A Survey of Recent Advances in Face Detection”, Technical
Report, MSR-TR-2010-66, Microsoft Research, Microsoft Corporation One Microsoft Way,
Redmond, WA 98052, http://www.research.microsoft
[24] Abboud B, Davoine F & Dang M (2004) “Facial expression recognition and synthesis based on an
appearance model”, Signal Processing: Image Communication, Vol.19, No.8, pp723–740.
[25] Mohammad Alia, Abdelfatah Tamimi and Omaima Al-Allaf, "Integrated System For Monitoring And
Recognizing Students During Class Session", AIRCC’s: International Journal Of Multimedia & Its
Applications (IJMA), Vol.5, No.6, December 2013, pp:45-52. Airccse.org/journal/ijma.html
[26] Jain A, Ross A & Prabhakar S (2004) “An introduction to biometric recognition”, IEEE Transactions
on Circuits and Systems for Video Technology, Vol.14, No.1, pp4–20, Jan.
[27] Henry Rowley, Baluja S. & Kanade T. (1999) “Neural Network-Based Face Detection, Computer
Vision and Pattern Recognition”, Neural Network-Based Face Detection, Pitts-burgh, Carnegie
Mellon University, PhD thesis.
[28] KahKay Sung & Tomaso Poggio (1994) Example Based Learning For View Based Human Face
Detection, Massachusetts Institute of Technology Artificial Intelligence Laboratory and Center For
Biological And Computational Learning, Memo 1521, CBCL Paper 112, MIT, December.
[29] Henry A. Rowley, Shumeet Baluja &Takeo Kanade. (1997) Rotation Invariant Neural Network-Based
Face Detection, December, CMU-CS-97-201
[30] Jeffrey S. Norris (1999) Face Detection and Recognition in Office Environments, thesis, Dept. of
Electrical Eng. and CS, Master of Eng in Electrical Eng., Massachusetts Institute of Technology.
[31] Hazem M. El-Bakry (2002), Face Detection Using Neural Networks and Image Decomposition
Lecture Notes in Computer Science Vol. 22, pp:205-215.
[32] Lin-Lin Huang, et al (2003) “Face detection from cluttered images using a polynomial neural
network”, Neurocomputing, Vol.51, pp197 – 211.
[33] Masakazu Matsugu (2003) “Subject independent facial expression recognition with robust face
detection using a convolutional neural network”, Neural Networks, Vol.16, pp555–559.
[34] Stefan W., Christian I. & Uwe H (2004) “Evolutionary Optimization of Neural Networks for Face
Detection”, Proceedings of the 12th European Symposium on Artificial Neural Networks, Evere,
Belgium: d-side publications.
[35] Marian Beszedes & Milos Oravec (2005) “A System For Localization Of Human Faces In Images
Using Neural Networks”, Journal Of Electrical Engineering, Vol. 56, No 7-8, pp195–199.
15
The International Journal of Multimedia & Its Applications (IJMA) Vol.6, No.1, February 2014
[36] N. Shilbayeh and G. Al-Qudah (2008) “Face Detection System Based On MLP Neural Network”,
Recent Advances in Neural Networks, Fuzzy Systems & Evolutionary Computing, ISSN: 1790-5109,
ISBN: 978-960-474-195-3, pp 238-243.
[37] Zoran Bojkovic & Andreja Samcovic (2006) Face Detection Approach In Neural Network Based
Method For Video Surveillance, 8th Seminar on Neural Network Applications in Electrical
Engineering, Neurel, Faculty Of Electrical Eng., University Of Belgrade, Serbia, September 25-27.
[38] Aamer Mohamed, et al (2008) “Face Detection based Neural Networks using Robust Skin Color
Segmentation”, 5th International Multi-Conference on Systems, Signals and Devices, IEEE.
[39] Sahoolizadeh, Sarikhanimoghadam and Dehghani (2008) “Face Detection using Gabor Wavelets and
Neural Networks”, World Academy of Science, Engineering and Technology, Vol. 45, pp552- 554.
[40] Avinash Kaushal, J P S Raina (2010) “Face Detection using Neural Network & Gabor Wavelet
Transform”, International Journal of Computer Science and Technology (IJCST), Vol. 1, Issue.1,
pp58-63, September 2010, ISSN : 0976 - 8491.
[41] Mohammad Abadi, et al, (2011) “Face Detection with the Help of Gabor Wavelets Characteristics
and Neural Network Classifier”, American Journal of Scientific Research, Issue.36, pp67-76, ISSN
1450-223X, http://www.eurojournals.com/ajsr.htm
[42] Anissa Bouzalmat, et al (2011) “Face Detection And Recognition Using Back Propagation Neural
Network And Fourier Gabor Filters”, Signal & Image Processing: An International Journal (SIPIJ)
Vol. 2, No. 3, September 2011, DOI : 10.5121/sipij.2011.2302 15
[43] Kalavdekar Prakash N. (2010) “Face Detection using Neural Network”, International Journal of
Computer Applications (0975 – 8887),Vol.1, No.14, pp36-39.
[44] Mansaf M Elmansori & Khairuddin Omar (2011) “An Enhanced Face Detection Method Using Skin
Color and Back-Propagation Neural Network”, European Journal of Scientific Research, ISSN 1450216X, Vol.55 No.1, pp80-86, http://www.eurojournals.com/ejsr.html
[45] Zuo F & P.H.N. de With (2008) “Fast face detection using a cascade of neural network ensembles”,
EURASIP Journal on Advances in Signal Processing, Volume 2008, Article ID 736508, Hindawi
Publishing Corporation, pp1-13, doi:10.1155/2008/736508
[46] Hudson, Hagan and Demuth, Neural Network Toolbox™ User’s Guide R2012a, The MathWorks,
Inc., 3 Apple Hill Drive Natick, MA 01760-2098, 2012, www.mathworks.com
AUTHOR
Dr. Omaima N. A. Al-Allaf received the M.Sc. degree in CS from the Dept. of CS\
Faculty of Computers and Mathematical Sciences\ University of Mosul\ Mosul\ Iraq
in 1999. She received the Ph.D. degree in CIS from CIS department\ Faculty of IS
and Technology\ AABFS \ Jordan in 2008. Currently, she is an Assistant Professor
at CIS department\ Faculty of Sciences and IT\ AlZaytoonah University of Jordan\
Amman\ Jordan (from 2009). Her research interests include image compression and
recognition, Artificial Neural Networks, Genetic Algorithms. She is member of
International Association of Engineers (IAENG).
16
| 9cs.NE
|
RHO-ESTIMATORS FOR SHAPE RESTRICTED DENSITY
ESTIMATION
arXiv:1503.04427v4 [math.ST] 25 Apr 2016
Y. BARAUD AND L. BIRGÉ
Abstract. The purpose of this paper is to pursue our study of ρ-estimators built from
i.i.d. observations that we defined in Baraud et al. (2014). For a ρ-estimator based on
some model S (which means that the estimator belongs to S) and a true distribution of the
observations that also belongs to S, the risk (with squared Hellinger loss) is bounded by a
quantity which can be viewed as a dimension function of the model and is often related to
the “metric dimension” of this model, as defined in Birgé (2006). This is a minimax point
of view and it is well-known that it is pessimistic. Typically, the bound is accurate for
most points in the model but may be very pessimistic when the true distribution belongs
to some specific part of it. This is the situation that we want to investigate here. For
some models, like the set of decreasing densities on [0, 1], there exist specific points in the
model that we shall call extremal and for which the risk is substantially smaller than the
typical risk. Moreover, the risk at a non-extremal point of the model can be bounded by
the sum of the risk bound at a well-chosen extremal point plus the square of its distance
to this point. This implies that if the true density is close enough to an extremal point,
the risk at this point may be smaller than the minimax risk on the model and this actually
remains true even if the true density does not belong to the model. The result is based
on some refined bounds on the suprema of empirical processes that are established in
Baraud (2016).
1. Introduction
The present paper pursues the study of ρ-estimation, introduced in Baraud et al. (2014),
as a versatile estimation strategy based on models. We want here to explain some specific
property of these estimators that we shall call superminimaxity, a study which was motivated by a conference that Adityanand Guntuboyina gave in Cambridge in June 2014. His
talk was about Gaussian regression but we shall deal here with density estimation. Given
n i.i.d. observations X1 , . . . , Xn with an unknown density s with respect to some reference
measure µ and an estimator sb(X1 , . . . , Xn ) of s, we measure its performance using the loss
function h2 (s, sb) where h is the Hellinger distance. We shall focus here on ρ-estimators and
some of their properties that lead to superminimaxity.
The first of these properties is robustness. There exist various notions of robustness:
robustness to model contamination, robustness to possible outliers, etc. — see Huber (1981)
for some illustrations —. In some of these cases, the problem can be formulated in the
following way. If we know the performance of an estimator when the true density s = s
belongs to a model S, how does it deteriorate when s is actually of the form (1 − ε)s + εt
for some small ε ∈ (0, 1) and an arbitrary density t 6= s, that is, when a proportion ε of
the data actually corresponds to a sample of density t and not s. Since for such a density
s one can check that h2 (s, s) ≤ ε, it is natural to wonder what happens to the risk of the
estimator not only when s is a mixture of the form (1 − ε)s + εt as before but also, more
Date: April 26, 2016.
1
generally, when it belongs to a small Hellinger ball around s, which leads to the notion of
robustness with respect to Hellinger deviations that we shall use here.
To illustrate the problem of contamination, assume that we choose as our statistical
model S for the unknown density s the set of uniform densities on [0, θ] with 0 < θ ≤ 10, in
which case the MLE (maximum likelihood estimator) is the uniform density sb on 0, X(n) ,
where X(n) is the largest observation, with a risk Es h2 (s, sb) bounded by C/n when s
belongs to S, Et denoting the expectation when the true density is t. But what if the true
density s does not belong to S? Unfortunately, the situation may become quite different.
If s is the mixture s0 = (1 − 1/n)1l[0,1] + (1/n)(1l[9,10] ) for some n larger than 100, it is easy
to check that, with probability of order 1 − e−1 ≈ 0.63, at least
one of the Xi will be larger
than 9 and the MLE will be the uniform distribution on 0, X(n) with X(n) > 9, which
is a terrible estimator of s = s0 , although the√model S is quite good since the Hellinger
distance between s and S is not larger than 1/ n.
The previous example shows that the MLE is definitely not robust in our sense since it
may be very sensitive to small deviations from the model on the contrary to ρ-estimators.
To be more precise, let us consider some model of densities S and a ρ-estimator sb based on
S with a risk function on S bounded by R(s, n), that is,
(1)
Es h2 (s, sb) ≤ R(s, n) for all s ∈ S.
The robustness of ρ-estimators can be expressed by the following property, proven in Baraud et al. (2014): whatever the density s,
for all s ∈ S,
(2)
Es h2 (s, sb) ≤ C0 R(s, n) + h2 (s, s)
where C0 is a universal positive constant. This is a fondamental property of ρ-estimators
for the following reasons: if s is quite close to a simple density s in S which can be estimated
with a small risk bound R(s, n), the ρ-estimator will essentially behave as if the true density
were s and the risk bound at s will be that at s plus a small additional term that can be
viewed as a squared bias. Intuitively, a ρ-estimator based on a sample with density s and a
ρ-estimator based on a sample with density s will remain close. In our parametric example
based on uniform distributions, everything happens as if the ρ-estimator only considered
the data with values in [0, 1] and ignored the others. Consequently, its risk remains of order
1/n even when s = s0 instead of s = s = 1[0,1] . This notion of robustness is quite flexible
and shows that the risk of the estimator does not deteriorate much in a small Hellinger
neighbourhood of any point s of the model.
For many well-chosen models S, the risk can be uniformly bounded on S:
(3)
sup R(s, n) ≤ R(S, n),
s∈S
which corresponds to the minimax point of view, so that (3) leads to
whatever the density s.
(4)
Es h2 (s, sb) ≤ C0 R(S, n) + h2 (s, S)
It turns out that for some models S, there exists a subset V of S such that the risk bounds
R(s, n) are substantially smaller than R(S, n) for all s ∈ V . This is what we call superminimaxity. Although there exists some analogy in the denomination, the notion is
quite different from the one of superefficiency as described in the famous counterexample of
Hodges and the Theorem of Le Cam about points of superefficiency, apart from the fact that
it deals with the property that estimation is faster at some points. However superefficiency
2
is an asymptotic property at a point, while superminimaxity on V is definitely nonasymptotic and defined for a given value of the number n of observations. For a detailed study
of superefficiency, one could look at the paper by Brown, Low, and Zhao (1997). Moreover, superminimaxity on V together with the robustness of ρ-estimators has the following
consequence: if s is either in V or close enough to it, the risk bound at s,
Es h2 (s, sb) ≤ C0 inf R(s, n) + h2 (s, s) ,
s∈V
may be substantially smaller than the typical risk bound (4) leading to superminimaxity at
s. It is actually the combination of the robustness of ρ-estimators and the existence of local
risk bounds of the form (1) that lead to this phenomenon as also described, in a different
framework, by Chatterjee et al. (2015), a paper that strongly influenced our research in
this direction.
Showing that the risk Es h2 (s, sb) at some particular points s can be bounded from
above by some quantity R(s, n) which is of smaller order than the global minimax risk over
S requires some specific probabilistic tools that have been established in Baraud (2016).
These tools allow to bound the expectation of the supremum of an empirical process over
the neighbourhood of an element s ∈ S by some quantity which is of smaller order than
that one could get by using the global entropy of the class S as, for example, in van de
Geer (1993).
The existence of points in the model on which the estimator is superminimax was already
noticed for the Grenander estimator of a non-increasing density — see Grenander (1981)
and Groeneboom (1985) — on an interval [a, +∞) with a known value of a. It is shown
in Birgé (1989) that the L1 -risk of the Grenander estimator of apnon-increasing piecewise
constant density based on at most D intervals is bounded by c D/n, for some positive
universal constant c, and can therefore be of smaller order than the typical risk for nonincreasing densities which is of order n−1/3 . We shall see below that, for the same estimation
problem, the ρ-estimator will perform similarly (up to possible logarithmic factors) with the
same superminimaxity property on piecewise constant densities. Moreover, the ρ-estimator
does not need to know a and is robust with respect to the Hellinger distance.
The case of monotone densities on [a, +∞) is far from unique. There are many other
examples of families S of densities for which one can find a subset V of S on which the
rates of convergence of the ρ-estimator are faster than the rate at a “typical” point of S.
Moreover, it happens that the set V often possesses good approximation properties with
respect to the much larger space S. These approximation properties combined with the
robustness of ρ-estimators as expressed by (2) allow to derive non-asymptotic minimax risk
bounds over large subsets of S. Such sets are possibly non-compact and therefore neither
possess a finite metric dimension nor a finite entropy.
In view of illustrating this superminimaxity phenomenon, we shall consider in the present
paper models of densities S defined by some shape constraints, namely piecewise monotone,
piecewise convex or concave and log-concave densities. There is a large amount of literature
dealing with these density models and we shall content ourselves to mention a few references
only and refer the reader to the bibliography therein. For monotone densities we refer to
the books by Groeneboom and Wellner (1992) and van de Geer (2000). For the estimation
of a convex density, we mention Groeneboom et al. (2001) and refer to the papers of Doss
and Wellner (2015), Dümbgen and Rufibach (2009), Cule and Samworth (2010) for the
estimation of a log-concave density. In the regression setting, let us mention Guntuboyina
3
and Sen (2015) for estimating a convex regression function and Chatterjee et al. (2015)
for the isotonic regression. Recently, Bellec (2015) extended the results of Chatterjee et
al. (2015) about the properties of least-squares estimators over convex polyhedral cones
in the homoscedastic Gaussian regression framework, to general closed convex subsets of
Rn from which he also derived some results of superminimaxity in this specific Gaussian
framework. In these two papers the results are restricted to convex models. As opposed,
convexity does not play any special role in our presentation and the models we shall use
here are not necessarily convex which allows us to deal with more general shape constraints
like piecewise monotonicity or log-concavity.
The paper is organised as follows. The statistical setting, main notations and conventions
as well as a brief reminder of what a ρ-estimator is in the density estimation framework
can be found in Section 2. The introductory example of the model of monotone densities
in Section 3 gives a first flavour of the results we establish all along the paper. The main
result can be found in Section 4 and its applications to different density models (piecewise
constant, piecewise monotone, piecewise convex-concave and log-concave densities) are detailed in Section 5. The problem of model selection is addressed in Section 6 and Section 7
is devoted to the proofs.
2. The statistical setting
Let (X , A ) be a measurable set, µ a σ-finite measure on (X , A ), Pµ the set of all
probabilities on (X , A ) which are absolutely continuous with respect to µ. We shall denote
by L the set of real-valued functions
R from X to R and Lµ the subset of L consisting
of those functions t ≥ 0 satisfying t dµ = 1, that is the set of probability densities with
respect to µ. An element of Pµ with density t ∈ Lµ will be denoted by Pt . We turn Pµ into
a metric space via the Hellinger distance h. We recall from Le Cam (1973 or 1986) that
the Hellinger distance between two elements P and Q of Pµ is given by
Z
2 1/2
p
p
1
h(P, Q) =
.
dP/dµ − dQ/dµ dµ
2 X
For Pt , Pu ∈ Pµ with t, u ∈ Lµ we shall write h(t, u) for h(Pt , Pu ).
We observe n i.i.d. random variables X1 , . . . , Xn with values in (X , A ) and distribution
Ps for some density s ∈ Lµ . Although s might not be uniquely defined in Lµ as a density
with respect to µ of the distribution of the observations, we shall refer to s as “the” density
of Ps for simplicity. To avoid trivialities, we shall always assume, in the sequel, that n ≥ 3,
so that log n > 1. The estimators sb that we shall consider here will be based on models for
s defined as follows.
Definition 1. A density model, or a model (for short), is a subset of Lµ for which there
exists an at most countable subset S ⊂ S such that {Pt , t ∈ S} is dense in {Pt , t ∈ S}
with respect to the Hellinger distance. We shall then say that S is dense in S and that S
is separable with respect to the Hellinger distance.
A density model S should be chosen so that the corresponding probability model {Pt , t ∈
S} approximates the true distribution Ps (with respect to the Hellinger distance). The
model S may or may not contain s. Of course a model S is good only if the distance
h s, S is not too large, where we set, for A ⊂ Lµ , h(t, A) = inf u∈A h(t, u). Our aim in
this paper is to study the performance of a ρ-estimator sb of s built on S. The definition
4
and properties of ρ-estimators have been described in great details in Baraud et al. (2014)
and we only give below a brief account of what a ρ-estimator is.
2.1. What is a ρ-estimator? In the context of density estimation based on i.i.d. variables,
which is the one we consider here, a ρ-estimator provides a robust (in our sense) and an
(almost) rate optimal estimator over a model S of densities in all cases we know. In order
to avoid long developments we restrict ourselves to its construction in the specific situations
we shall encounter here, namely when the observations are i.i.d.
Let ψ be the increasing function from [0, +∞] onto [−1, 1] defined by
u−1
ψ(u) = √
1 + u2
for u ∈ [0, +∞)
and
ψ(+∞) = 1.
Given a model S of densities on (X , A , µ) and a countable and dense subset S of S, a
ρ-estimator sb of the density s on S is defined in the following way. For densities t, t′ ∈ Lµ
we set
!
r
n
′
′
′
X
1
t
+
t
t
+
t
t
n
ψ
(Xi )
h2 t,
− h2 t′ ,
+√
T(X, t, t′ ) =
2
2
2
t
2
i=1
and define sb as any (measurable) element of the closure of the set
′
(5)
t ∈ S Υ(S, t) ≤ inf
Υ(S, t ) + 35.7
with Υ(S, t) = sup T(X, t, t′ ).
′
t ∈S
t′ ∈S
In the calculation of T(X, t, t′ ), which involves the ratio t′ /t, we use the convention 0/0 = 1
and a/0 = +∞ for a > 0. The constant 35.7 in (5) has only been chosen for convenience
in the calibration of the numerical constants in the original paper Baraud et al. (2014) and
can be replaced by any positive number. It is clear from the construction that, given a
model S, there is not a unique ρ-estimator on S. However, the risk bounds we derived in
Baraud et al. (2014) are valid for any version of these ρ-estimators.
2.2. Notations, conventions and definitions. We set log+ (x) = max{log x, 1}, N∗ =
N \ {0}, a ∨ b = max(a, b), a ∧ b = min(a, b) and, for x ∈ R+ , ⌈x⌉ = inf{n ∈ N, n ≥ x};
|A| denotes the cardinality of the finite set A and C, C ′ , . . . numerical constants that may
vary from line to line. For a function f on R, f (x+) and f (x−) denote respectively the
right-hand and left-hand
Plimits of f at x whenever these limits exist. We shall also use the
following conventions:
∅ = 0, x/0 = +∞ if x > 0, x/0 = −∞ if x < 0 and 0/0 = 1.
Definition 2. A partition of the open interval (a, b) (−∞ ≤ a < b ≤ +∞) of size k + 1
with k ∈ N is either ∅ when k = 0 or a finite set I = {x1 , . . . , xk } of real numbers with
a < x1 < x2 < · · · < xk < b if k ≥ 1. We shall call endpoints of the partition I the numbers
xj , 1 ≤ j ≤ k, and intervals of the partition the open intervals Ij = (xj , xj+1 ), 0 ≤ j ≤ k
with x0 = a and xk+1 = b. A partition I will also be identified to the set of its intervals
and we shall equally write I = {I0 , . . . , Ik } or I = {x1 , . . . , xk }.
The set of all partitions of R with k endpoints or k + 1 intervals is denoted by J (k + 1)
and the length of Ij by ℓ(Ij ). If I = {x1 , . . . , xk } and I ′ = {x′1 , . . . , x′k′ }, I ∨ I ′ =
{x1 , . . . , xk } ∪ {x′1 , . . . , x′k′ } and I I ′ means that {x1 , . . . , xk } ⊃ {x′1 , . . . , x′k′ }.
5
3. Monotone densities
In view of illustrating the main result of this paper to be presented in Section 4, let
us consider the example of the model S consisting of all the densities with respect to the
Lebesgue measure µ that are non-increasing on some arbitrary interval of R which is open
on its left end and vanish elsewhere. In this case (X , A ) = (R, B(R)) and S is the set
of all densities of the form t = f 1l(x,+∞) with x ∈ R and f is a non-increasing and nonnegative
R +∞function on (x, +∞) (which may be unbounded in the neighbourhood of x) such
that x f (x) dx = 1. The results we get below would be similar for the set of all densities
which are non-decreasing on some interval of R and vanish elsewhere.
P
For D ∈ N∗ we define V (D) to be the set of all densities of the form D
j=1 aj 1l(xj ,xj+1 ]
with I = {x1 , . . . , xD+1 } ∈ J (D + 2) and aj ≥ 0 for 1 ≤ j ≤ D. Note that the densities
in V (D) take the value 0 on the two unbounded extremal intervals I0 and ID+1 of the
partition I. For instance, V (1) corresponds to the family of uniform densities on intervals,
that is
V (1) = t(·) = θ1−1 1l(0,θ1 ] (· − θ0 ), θ1 > 0, θ0 ∈ R .
In such a situation, we can prove the following result.
Theorem 1. Any ρ-estimator sb on S, as defined in Section 2.1, satisfies
n
D
for all s ∈ Lµ
(6)
CEs h2 (s, sb) ≤ inf h2 s, V (D) ∩ S + log3+
D≥1
n
D
and some universal constant C ∈ (0, 1].
Remark. Since C ≤ 1, the left-hand side is always bounded by one so that it is useless
to consider values of D that lead to a bound which is not smaller than one, in particular
D ≥ n, and (6) is actually equivalent to
n
D
2
3
2
.
CEs h (s, sb) ≤ inf
h s, V (D) ∩ S + log+
1≤D<n
n
D
Although we shall not repeat it systematically, the same remark will hold for all our subsequent results.
Bound (6) means that the risk function s 7→ Es h2 (s, sb) of sb over Lµ can be quite small
in the neighbourhood of some specific densities t ∈ S: if s belongs to V (D) ∩ S with D < n
or is close enough to some density t ∈ V (D) ∩ S, the risk of sb is of order D/n, up to
logarithmic factors. More precisely,
n
D
sup Es h2 (s, sb) ≤ C ′ log3+
for 1 ≤ D < n.
n
D
s∈V (D)∩S
When n becomes large and D remains fixed, the rate of convergence of sb towards an element
of V (D) ∩ S is therefore almost parametric.
Of particular interest are the densities t which are bounded, supported on a compact
interval [a, b] of R (for numbers a < b depending on t) and non-increasing on (a, b). Given
M ≥ 0, we introduce the set S(M ) of densities t of this form and for which
√
2
t = M (t) ≤ M ,
(7)
(b − a)V[a,b]
√
√
where the variation V[a,b]
t of the non-increasing function t on [a, b] is defined in the
following way:
6
Definition 3. Let the function f be defined on some interval I (with positive length) of R
and monotone on the interior ˚
I of I. Its variation on I is given by
(8)
VI (f ) = sup f (x) − inf f (x) ∈ [0, +∞].
x∈˚
I
x∈˚
I
Note that S(0) is the set of uniform densities on intervals, so that S(0) = V (1), and that
S(M ) is not compact and contains densities that can be arbitrarily large in sup-norm.
The
functional M remains invariant by translation and scaling: if u(·) = λt λ(· − τ ) with λ > 0
and τ ∈ R, then M (u) = M (t) which implies that S(M ) is also invariant by translation
and scaling. It turns out that the densities lying in S(M ) can be well approximated by
elements of V (D). More precisely, the following approximation result holds.
S
Proposition 1. For all D ∈ N∗ and t ∈ M ≥0 S(M ),
h2 t, V (D) ∩ S ≤ M (t)/(2D)2 ∧ 1.
Using the triangle inequality, the right-hand side of (6) can be bounded from above in
the following way: for all M ≥ 0 and t ∈ S(M ),
D
2
3 n
inf h s, V (D) ∩ S + log+
D≥1
n
D
n
D
3
2
2
≤ 2h (s, t) + inf 2h t, V (D) ∩ S + log+
D≥1
n
D
n
D
M
.
+ log3+
≤ 2h2 (s, t) + inf
2
D≥1 2D
n
D
Finally, since t is arbitrary in S(M ),
n
n
D
D
M
2
2
3
3
inf h s, V (D) ∩ S + log+
≤ 2h s, S(M ) + inf
.
+ log+
D≥1
D≥1 2D 2
n
D
n
D
Optimizing the right-hand side with respect to D and using the facts that M is arbitrary
and log n > 1, we derive the following corollary of Theorem 1.
Corollary 1. For all probabilities Ps in Pµ , any ρ-estimator sb of s on S satisfies, for some
constant C ∈ (0, 1],
_
h
i
1/3
.
n−1 (log n)3
(9)
CEs h2 (s, sb) ≤ inf h2 s, S(M ) + M n−2/3 (log n)2
M ≥0
In particular, if s ∈ S(M ) for some M ≥ n−1 (log n)3 , the risk bound of the estimator
1/3
is not larger (up to a universal constant) than M n−2/3 (log n)2 while for smaller values
of M it is bounded by n−1 (log n)3 . Up to logarithmic factors, this rate (with respect to
n) is optimal since it corresponds to the lower bound of order n−2/3 for the minimax risk
on the subset of S(M ) consisting of the non-increasing densities supported in [0, 1] and
bounded by M . This lower bound follows from the proof of Proposition 1 of Birgé (1987).
The result was actually stated in this paper for the L1 -distance but its proof shows that it
applies to the Hellinger distance as well. This property means that, although the set S(M )
is not compact because the support of the densities is unknown, the minimax risk on S(M )
is finite. We do not know any other estimator with the same performance which is also
robust with respect to Hellinger deviations.
7
Note that Corollary 1 can also be used to determine the rate of estimation for decreasing
densities s with possibly unbounded support and maximum value, provided
that we have
some assumption about the behaviour of the function M 7→ h s, S(M ) when M goes to
infinity.
4. The main result
Let us start with some definitions.
Definition 4. A class C of subsets of X is said to shatter a finite subset A = {x1 , . . . , xm }
of X if the class of subsets
(10)
C ∩ A = {C ∩ A, C ∈ C }
is equal to the class of all subsets of A or, equivalently, if |C ∩ A| = 2m . A non-empty
class C of subsets of X is a VC-class with dimension d ∈ N if there exists some integer m
such that no finite subset A ⊂ X with cardinality m can be shattered by C and d + 1 is the
smallest m with this property.
Definition 5. Let F be a non-empty class of functions on a set X with values in [−∞, +∞].
We shall say that F is weak VC-major with dimension d ∈ N if d is the smallest integer
k ∈ N such that, for all u ∈ R, the class
(11)
Cu (F ) = {x ∈ X , f (x) > u}, f ∈ F
is a VC-class of subsets of X with dimension not larger than k.
We may now introduce the main property to be used in this paper.
Definition 6. Let F be a class of real-valued functions on X . We shall say that an element
f ∈ F is extremal in F (or is an extremal point of F ) with degree d(f ) = d ∨ 1 ∈ N∗ if
the class of functions
(F /f ) = {f /f , f ∈ F },
is weak VC-major with dimension d.
Proposition 2. Let F be a class of nonnegative functions on X . The element f ∈ F is
extremal in F with degree not larger than 2d if for all λ ≥ 0,
C (F , f , λ) = {X } ∪ {x ∈ X | f (x) − λf (x) > 0}, f ∈ F
is a VC-class with dimension not larger than d ≥ 1.
Proof. Let us bound the VC dimension of Cu ((F /f )) according to the value of u ∈ R. If
u < 0, Cu ((F /f )) = {X } and is therefore VC with dimension not larger than 0 ≤ 2d. Let
us now assume that u ≥ 0 and set A = {f > 0}. Using Lemma 5 (in Section 7.1), it suffices
to prove that Cu ((F /f )) ∩ A and Cu ((F /f )) ∩ Ac are two VC-classes with dimensions not
larger than d. For all x ∈ A and f ∈ F
f (x)/f (x) > u
is equivalent to
f (x) > uf (x)
showing thus that Cu ((F /f )) ∩ A ⊂ C (F , f , u) ∩ A and is therefore VC with dimension
not larger than d. Let us now turn to the case where x 6∈ A, which means that f (x) = 0 so
that f (x)/f (x) is either 1 or +∞ (with our conventions). For u ≥ 1 and all x ∈ Ac ,
f (x)/f (x) > u
is equivalent to
8
f (x) > 0
and Cu ((F /f )) ∩ Ac ⊂ C (F , f , 0) ∩ Ac and is therefore VC with dimension not larger than
d. For u ∈ [0, 1), (f /f )(x) > u for all x ∈ Ac , hence Cu ((F /f )) ∩ Ac = {Ac } which is VC
with dimension 0 < d and this concludes the proof.
Let us now state our main result.
Theorem 2. Let S be a model with a non-void set Λ of extremal points. Any ρ-estimator
sb on S satisfies, for some universal constant C ∈ (0, 1],
d(s)
n
ξ
Ps Ch2 (s, sb) ≤ inf h2 (s, s) +
log3+
+
≥ 1 − e−ξ for all ξ > 0,
n
d(s)
n
s∈Λ
whatever the true distribution Ps ∈ Pµ . Consequently,
2
n
d(s)
3
2
log+
.
CEs h (s, sb) ≤ inf h (s, s) +
n
d(s)
s∈Λ
Note that the boundedness of h implies that values of d(s) ≥ n lead to a trivial bound so
that the infimum could be reduced to those s such that d(s) < n. We do not know to what
extend the log3 factor is necessary. We believe that it is not optimal although a log-factor
appears to be necessary in some situations as shown by the example of Section 5.1 below.
5. Applications
Throughout this section, (X , A ) = (R, B(R)) and µ is the Lebesgue measure on R. In
particular, we shall only consider densities with respect to the Lebesgue measure. We start
with the following useful lemma:
Lemma 1. If D is a class of subsets of R such that each element of D is the union of at
most k intervals, D is VC with dimension at most 2k.
Proof. Let x1 < x2 < . . . < x2k+1 be 2k + 1 points of R. It is easy to check that elements
of the form J1 ∪ · · · ∪ Jl ∈ D, where the Jj are disjointed intervals and l ≤ k, cannot pick
S
up the subset of points ki=0 {x2i+1 }.
5.1. Piecewise constant densities. Let us now consider the model V (D) of Section 3
to build a ρ-estimator. If f and f belong to V (D), for all λ ≥ 0, f − λf is of the form
Pk
j=1 aj 1l(xj ,xj+1 ] with k < 2(D + 1) so that {x ∈ X | f (x) − λf (x) > 0} is the union
of at most D + 1 disjointed intervals. Applying Lemma 1 and Proposition 2 to the sets
D = C (V (D), f , λ) with λ ≥ 0, we obtain that all the elements of V (D) are extremal in
V (D) and their degrees are not larger than 4(D + 1). We therefore deduce from Theorem 2
that
n
D
,
(12)
sup Es h2 (s, sb) ≤ C log3+
n
D
s∈V (D)
which, up to the logarithmic factor, corresponds to a parametric rate (with respect to n)
although the partition that defines s can be arbitrary in J (D + 2) and the support of s
is unknown. It follows from Birgé and Massart (1998), Proposition 2, that a lower bound
for the minimax risk on V (D) is of the form C ′ (D/n) log+ (n/D), which shows that some
power of log+ (n/D) is necessary in (12). We suspect that the power three for the logarithm
is not optimal.
9
5.2. Piecewise monotone densities. Let us now see how Theorem 2 can be applied in
the simple situation of piecewise monotone densities.
Definition 7. Given k ∈ N∗ and a partition I = {I0 , . . . , Ik−1 } ∈ J (k), a real-valued
function f on R will be called piecewise monotone (with k pieces) based on I if f is monotone
on each open interval Ij , 0 ≤ j ≤ k − 1. The set of all such functions will be denoted by
Gk . For k ≥ 2 (since no density is monotone on R), Fk is the set of densities (with respect
to the Lebesgue measure) that belong to Gk .
Clearly, Gk ⊂ Gl and Fk ⊂ Fl for all l > k.
Proposition 3. For D ≥ 1 and k ≥ 2, any element f of Fk ∩ V (D) is extremal in Fk
with degree not larger than 3(k + D + 1).
Proof. Let f be a piecewise monotone density on R based on a partition I0 ∈ J (k), therefore
with k−1 endpoints. Let f ∈ V (D)∩Fk be a piecewise constant density based on a partition
I1 ∈ J (D + 2) (with D + 1 endpoints) and let I2 = I1 ∨ I0 . It is a partition of R with at
most k + D endpoints, therefore at most k + D + 1 intervals and on each such interval f is
monotone and f is constant which implies that f − λf belongs to Gk+D+1 for all λ ≥ 0. It
then follows from Lemma 2 below that the sets {x ∈ X | f (x) − λf(x) > 0} are unions of at
most (3/2)(k + D + 1) intervals. The conclusion follows from Lemma 1 and Proposition 2
applied to D = C (Fk , f , λ) with λ ≥ 0.
Lemma 2. If f ∈ Gk , whatever a ∈ R the set {x ∈ X | f (x) > a} can be written as a
union of at most k + ⌈(k − 1)/2⌉ ≤ 3k/2 intervals and the set {x ∈ X | f (x) ≤ a} as well.
Proof. Let f ∈ Gk , I be the partition of R with k open intervals associated to f and
x1 , . . . , xk−1 the k − 1 endpoints of this partition. For Ij ∈ I, {x ∈ X | f (x) > a} ∩ Ij is
either ∅ or a non-void interval and
k
[ k−1
[
[
[{xj } ∩ {f > a}] .
[{f > a} ∩ Ij ]
(13)
{x ∈ X | f (x) > a} =
j=1
j=1
This decomposition shows that {x ∈ X | f (x) > a} is the union of at most 2k − 1 disjointed
intervals. Nevertheless, this bound can be refined as follows. If f (xj ) ≤ a, {xj } ∩ {f >
a} = ∅ and the only situation we need to consider is when f (xj ) > a in which case
{xj } ∩ {f > a} = {xj }. If xj belongs
to the closure of one of the intervals of the form
{f > a} ∩ Ij ′ , the set {f > a} ∩ Ij ′ ∪ {xj } only counts for one interval in (13). The only
situation for which {xj } adds an extra interval occurs when f (xj−1 +) > a, f (xj −) ≤ a,
f (xj ) > a, f (xj +) ≤ a and f (xj+1 −) > a. The number of such points xj is not larger
than ⌈(k − 1)/2⌉ and {x ∈ X | f (x) > a} is therefore the union of at most k + ⌈(k − 1)/2⌉
intervals. The proof for {x | f (x) ≤ a} is the same.
h
i
T S
An application of Theorem 2 with Λ = Fk
D≥1 V (D) , the elements of which are
extremal in Fk by Proposition 3, leads to the following result.
Corollary 2. For all k ≥ 2, any ρ-estimator on Fk satisfies, for all distributions Ps ∈ Pµ ,
"
#
2
n
k
+
D
inf
log3+
CEs h (s, sb) ≤ inf
h2 (s, s) +
T
D≥1 s∈Fk V (D)
n
k+D
where C ∈ (0, 1] is a universal constant.
10
Note that the bound is trivial for k ≥ n − 1 and that using D with k + D ≥ n also leads
to a trivial bound so that we should restrict ourselves to D < n − k when k < n − 1.
Since, for t ∈ Fk ,
"
#
inf
T
s∈Fk
V (D)
h(s, s) ≤ inf
t∈Fk
h(s, t) +
inf
T
s∈Fk
h(t, s) ,
V (D)
to go further with our analysis it will be necessary to evaluate inf s∈Fk T V (D) h(t, s) for
t ∈ Fk . In order to do this we shall use an approximation result based on the following
functional Mk .
Definition 8. Let t ∈ Fk for k ≥ 2 and I = {I0 , . . . , Ik−1 } ∈ J (k) a partition on which t
is based. Using the convention (+∞) × 0 = 0, we define
3
k−1 h
√ i1/3
X
≤ +∞,
t
(14)
Mk (t, I) =
ℓ(Ij )VI2j
j=0
where VIj
Fk as
√
√
t is the variation of t on Ij given by (8). The functional Mk is defined on
Mk (t) = inf Mk (t, I)
I
for all t ∈ Fk ,
where the infimum runs among all partitions I ∈ J (k) on which t is based. For 0 ≤ M <
+∞ and k ∈ N∗ , we denote by Fk+2 (M ) the subset of Fk+2 of those densities t such that
Mk+2 (t) ≤ M .
Note that with our convention, if Mk (t, I) < +∞, t is equal to zero on I0 ∪ Ik−1 , in which
case the summation in (14) can be restricted to 1 ≤ j ≤ k − 2, which requires that k > 2,
and that Fk+2 (0) is equal to V (k) (in the L1 sense). The functional Mk is translation and
scale invariant which means that it takes the same value at t and λ−1 t((· − τ )/λ) whatever
λ > 0 and τ ∈ R. Besides, it possesses the following property.
Lemma 3. For all l > k and t ∈ Fk ⊂ Fl , Ml (t) ≤ Mk (t).
Proof. Let I ∈ J (k) on which t is based. For all partitions I ′ ∈ J (l) satisfying I ′ I, t
can be viewed as an element of Fl based on I ′ and consequently it suffices to show that
Ml (t, I ′ ) ≤ Mk (t, I). In fact, it suffices to show that, when we simply divide an interval J
of length L of I into m intervals J1 , . . . , Jm of respective lengths L1 , . . . , Lm ,
m
m
m h
√
√
√
√ i1/3
X
X
X
VJ j
Lj = L and
t
t when
t ≤ VJ
t .
≤ LVJ2
Lj VJ2j
j=1
j=1
j=1
√
√
P
1/3 2/3
t = βj VJ
t , this amounts to show that m
≤1
Setting Lj = αj L and VJj
j=1 αj βj
which follows from Hölder’s Inequality.
The approximation of elements of Fk+2 (M ) by elements of V (D) is controlled in the
following way.
Proposition 4. Let k ≥ 1 and t ∈ Fk+2 with Mk+2 (t) < +∞. Then, for all D ≥ 1,
h2 t, V (D + k) ∩ Fk+2 ≤ (2D)−2 Mk+2 (t).
11
Applying Corollary 2 leads to the following bound which is valid for all t ∈ Fk+2 with
Mk+2 (t) < +∞ and whatever the distribution Ps of the observations:
2
Mk+2 (t) k + D
n
2
3
(15)
CEs h (s, sb) ≤ h (s, t) + inf
+
log+
.
D≥1
D2
n
k+D
A final optimization with respect to D leads to
CEs h2 (s, sb) ≤ h2 (s, t) + [Mk+2 (t)]1/3 n−2/3 (log n)2 + kn−1 (log n)3 .
Since this result is valid for all densities t ∈ Fk+2 , we can again optimize it with respect
to t which finally leads to:
Theorem 3. Any ρ-estimator sb based on the model Fk+2 for some k ≥ 1 satisfies
h
1/3 _
i
,
kn−1 log n
CEs h2 (s, sb) ≤ inf h2 s, Fk+2 (M ) + (log n)2 M n−2
M>0
for all distributions Ps ∈ Pµ . In particular
h
1/3 _
i
sup
kn−1 log n .
Es h2 (s, sb) ≤ C(log n)2 M n−2
s∈Fk+2 (M )
If we want to estimate a bounded unimodal density s with support of finite length L, we
may build a ρ-estimator on F4 . In such a case, M4 (s) can be bounded by 4Lksk∞ ≥ 4 (since
s is a density, Lksk∞ ≥ 1) and the performance of the ρ-estimator for such a unimodal
density s will be given by
1/3
(log n)2 .
Es h2 (s, sb) ≤ C Lksk∞ n−2
5.3. Piecewise concave-convex densities. In the previous sections we considered densities
t which were piecewise monotone or constant which implied the same properties for
√
t
but
it follows from Proposition 4 that it is actually the approximation properties of
√
t that matter. This derives from the fact that the Hellinger distance is an L2 -distance
between the square roots of the densities. When going to more sophisticated
properties
√
than monotonicity, it is no more the same to state them for t or for t which accounts for
the slightly more complicated structure of this section.
Definition 9. Let I ∈ J (k) be a partition with k intervals. A function f is piecewise
convex-concave based on I if it is either convex or concave on each (open) interval Ij of the
partition. The set of all such functions when I varies in J (k) will be denoted by Gk1 . For
P
D ∈ N∗ we denote by W 1 (D) the set of all functions γ of the form γ = D
j=1 γj 1l(xj ,xj+1 ]
with x1 < x2 < · · · < xD+1 where γj is an √
affine function for all j. The sets Fk1 and V 1 (D)
are the sets of those densities t such that t belongs to Gk1 and W 1 (D) respectively.
We recall that if f is either concave or convex on some open interval I, it is continuous
on I and admits on I a right-hand derivative f ′ which is monotone.
The following result will prove useful to find extremal points of Fk1 .
Lemma 4. For all k ∈ N∗ , Gk1 ⊂ G2k .
Proof. Since f ∈ Gk1 there exists I ∈ J (k) such that f is either convex or concave on each
open interval Ij of I. The right-hand derivative f ′ of f on Ij being monotone, the sets
{x ∈ Ij , f ′ ≤ 0} and {x ∈ Ij , fj′ > 0} are two disjointed subintervals of Ij on which f is
monotone.
12
Proposition 5. For all D, k ∈ N∗ , the elements f ∈ Fk1 ∩ V 1 (D) are extremal in Fk1 with
degrees not larger than 12(D + k + 1).
Proof. Let us consider g ∈ Gk1 and g ∈ Gk1 ∩ W 1 (D). There exists a partition I0 with k − 1
endpoints such that g is either convex or concave on each interval of I0 and a partition I1
with D + 1 endpoints such that g is affine on each interval of I1 . The partition I0 ∨ I1
contains at most k + D + 1 intervals and on each of these intervals g − λg is either convex
1
which is a
or concave for all λ ∈ R+ . Hence, the function g − λg belongs to Gk+D+1
subset of G2(k+D+1) by Lemma 4. It then follows from Lemma 2 that {g − λg > 0} is
the union of at most 3(k + D + 1) intervals. Since λ is arbitrary in R+ we conclude with
Lemma 1 that C (Gk1 , g) is VC with dimension not larger than 6(D + k + 1), which shows by
Proposition 2 that the elements g ∈ Gk1 ∩ W 1 (D) are extremal in Gk1 with degrees not larger
than 12(D + k + 1). The conclusion follows from an application of Lemma 5 of Section 7.1
with α = 1/2.
We may now apply Theorem 2 with Λ = Fk1
T hS
i
V
(D)
which consists of extremal
D≥1 1
points of Fk1 and deduce the following risk bound from Proposition 5:
Corollary 3. For all k ≥ 2, any ρ-estimator on Fk1 satisfies for all distributions Ps ∈ Pµ ,
"
#
2
k
+
D
n
h2 (s, s) +
log3+
inf
CEs h (s, sb) ≤ inf
D≥1 s∈V 1 (D)∩F 1
n
k+D
k
where C ∈ (0, 1] is a universal constant.
The control of the approximation term inf s∈V 1 (D)∩F 1 h(s, s) is analogue to the one we
k
derived in the previous section for inf s∈Fk T V (D) h(t, s) but is based on a new functional:
Definition 10. Let t ∈ Fk1 and I = {I0 , . . . , Ik−1 } ∈ J (k) a partition on which t is based,
√ ′
√
that is, t is either convex or concave with monotone right-hand derivative
t on each
Ij . Using the convention (+∞) × 0 = 0, we define
1/5 5
k−1
X
′
√
≤ +∞,
t
Mk,1 (t, I) =
[ℓ(Ij )]3 VI2j
j=0
√
√ ′
′
where VIj
t
is the variation of
t on Ij . The functional Mk,1 is defined on Fk1 as
Mk,1 (t) = inf Mk,1 (t, I)
I
for all t ∈ Fk1 ,
where the infimum runs among all partitions I ∈ J (k) on which t is based. For 0 ≤ M <
1 (M ) the subset of F 1
+∞ and k ∈ N∗ , we denote by Fk+2
k+2 of those densities t such that
Mk+2,1 (t) ≤ M .
Note that if Mk,1 (t, I) is finite, the density t is necessarily zero on the two extremal
(unbounded) intervals of the partition I and therefore k ≥ 3. An analogue of Lemma 3
holds for the functional Mk,1 (t) with a similar proof, saying that if l > k and t ∈ Fk1 ⊂ Fl1 ,
then Ml,1 (t) ≤ Mk,1 (t). We omit the details.
13
1
with Mk+2,1 (t) < +∞. Then, for D ≥ 1,
Proposition 6. Let k ≥ 1 and t ∈ Fk+2
1
h2 t, V 1 2(D + k) ∩ Fk+2
≤ (D/2)−4 Mk+2,1 (t).
Now arguing as we did in the previous section we derive from Corollary 3 and Proposition 6 our concluding result.
1
with k ≥ 1 satisfies, for all
Theorem 4. Any ρ-estimator sb based on the model Fk+2
Ps ∈ Pµ ,
_
h
1/5
i
1
CEs h2 (s, sb) ≤ inf h2 s, Fk+2
(log n)12/5
kn−1 (log n)3 .
(M ) + M n−4
M >0
1 (M ), then
If, in particular, s ∈ Fk+2
h
_
1/5
i
Es h2 (s, sb) ≤ C
M n−4
(log n)12/5
kn−1 (log n)3 .
5.4. Log-concave densities. We now want to investigate a situation which is close to
the previous one, the case of log-concave densities on the line. These are densities of the
form 1lI exp(g) for some open interval I of R, possibly of infinite length, and some concave
′
function g on I. Let us denote by F ′ the set of all such densities and by V (D) the subset
of F ′ of those densities for which g is piecewise affine on I with D pieces. For instance,
′
′
the exponential density belongs to V (1) while the Laplace density belongs to V (2). Also
note that if 1lI exp(g) is log-concave, the same holds for its square root 1lI exp(g/2).
′
Proposition 7. For all D ∈ N∗ , the elements of V (D) are extremal in F ′ with degrees
not larger than 12(D + 2) + 4.
′
Proof. Let us consider 1lI exp(g) ∈ F ′ and 1lJ exp(g) ∈ V (D). Then the set on which
1lI exp(g) > λ1lJ exp(g), with λ ≥ 0 is the subset of I on which
g > log λ + log 1lJ + g,
with the convention that log 0 = −∞. If λ = 0 it is the set I itself. Otherwise it is equal
to the union of I ∩ J c and I ∩ J ∩ {g − g > log λ}. Since on the interval I ∩ J, g is concave
and g piecewise affine with at most D pieces, the function h = (g − g)1lI∩J + (log λ)1l(I∩J)c
1
is piecewise concave on R with at most D + 2 pieces. Hence h belongs to GD+2
and by
Lemma 4 it also belongs to G2(D+2) and it follows from Lemma 2 that I ∩ J ∩ {g − g >
log λ} = {h > log λ} is the union of at most 3(D + 2) intervals. Consequently, the set
{1lI exp(g) > λ1lJ exp(g)} is the union of at most 3(D + 2) + 1 intervals and we derive from
Lemma 1 that the VC-dimension of C (F ′ , 1lJ exp(g)) is not larger than 6(D + 2) + 2. The
conclusion then follows from Proposition 2.
We may now apply Theorem 2 with Λ =
the following risk bound.
S
D≥1 V
′
(D) and use Proposition 7 to derive
Corollary 4. Any ρ-estimator on F ′ satisfies, for all distributions Ps ∈ Pµ ,
#
"
n
2
D
2
3
inf h (s, s) + log+
CEs h (s, sb) ≤ inf
D≥1 s∈V ′ (D)
n
D
where C ∈ (0, 1] is a universal constant.
14
′
In particular, if s ∈ V (D),
n
D
Es h2 (s, sb) ≤ C log3+
,
n
D
′
which means that the elements of V (D) can be estimated by the ρ-estimator at a parametric
rate, up to some (log n)3 factor. This is the case for all uniform densities, for exponential
densities and their translates and for the Laplace density among many others.
Remark. For simplicity we have restricted our study to log-concave densities but we could
as well handle the case of piecewise log-concave densities with several pieces, that is densities
P
of the form kj=1 1lIj exp(gj ) for concave functions gj . The extension would be similar to
that which leads from monotone to piecewise monotone and is straightforward.
6. Model selection
All results of Sections 5.1 to 5.3 were based on the use of a single model: V (D) in
Section 5.1, Fk+2 in Section 5.2 and Fk1 in Section 5.3, which implies that our risk bounds
depend on D in the first case and on k in the other cases. In order to get the best possible
value of either D or k for the unknown distribution Ps , we may use a selection procedure.
There are different ways to do this but we shall explain how to do it using Theorem 9 of
Birgé (2006, Section 9). To simplify the presentation, we assume that the number n of
observations is even with n = 2p and we split the sample X = (X1 , . . . , Xn ) into two parts
1 , k ≥ 1,
of size p, X 1 and X 2 . We also consider all the models Fj+2 , j ≥ 1, and Fk+2
simultaneously. For each of these models we fix a weight ∆(j) = j and ∆(k) = k. It follows
that
X
X
2
.
(16)
exp[−∆(j)] +
exp[−∆(k)] =
e−1
j≥1
k≥1
We may now use each of our models to build a ρ-estimator based on the sample X 1 .
This results in a family of estimators sbj (X 1 ), j ≥ 1, and sbk (X 1 ), k ≥ 1. The risks of
these estimators are bounded according to Theorems 3 and 4. In the second step, we
consider these preliminary estimators based on sample X 1 as a set of points in Lµ . We
may apply to them the selection procedure described in Section 9.1 of Birgé (2006) via a
T-estimator based on the second sample X 2 . Then Theorem 9 of that paper applies with
the parameters Σ = 2/(e − 1), λ = 1, q = 2, d = h, κ = 4 and a = p/4. It follows that the
selection procedure results in an estimator sb which satisfies
2
2
2
CEs h (s, sb) | X 1 ≤ min inf h (s, sbj (X 1 )) + (j/p) , inf h (s, sbk (X 1 )) + (k/p) .
j≥1
k≥1
We may then take the expectation with respect to X 1 and get
2
2
2
CEs h (s, sb) ≤ min inf Es h (s, sbj (X 1 )) + (j/p) , inf Es h (s, sbk (X 1 )) + (k/p) .
j≥1
k≥1
Now applying Theorems 3 and 4 in order to bound Es h2 (s, sbj (X 1 )) and Es h2 (s, sbk (X 1 ))
respectively we derive that the two following bounds hold simultaneously:
h
i_
h
1/3
i
h2 s, Fj+2 (M ) +
CEs h2 (s, sb) ≤ inf
M n−2
(log n)2
jn−1 (log n)3
j≥1, M >0
15
and
CEs h2 (s, sb) ≤
inf
k≥1, M >0
h
_
1/5
i
1
(M ) + M n−4
h2 s, Fk+2
(log n)12/5
kn−1 (log n)3 .
This is only a simple example and the same procedure could be applied to a larger family
of models and preliminary estimators but we shall not insisit on that here, the important
point being that we may easily extend the results we got for a single model to large families
of models and get a final bound corresponding to the best bound among all models involved
in the procedure.
An alternative selection procedure leading to the same result is described in Baraud (2011,
Section 6.2). It is also possible to avoid the splitting device by using all models simultaneously and a penalized ρ-estimator as indicated in Section 7 of Baraud et al. (2014). Again,
we would get in the end the same type of risk bounds. For simplicity, we shall not insist
on this other approach here.
7. Proofs
7.1. Preliminaries. In the sequel, we shall use the following elementary properties.
Lemma 5.
1) If C is a VC-class of subsets of X with dimension not larger than d and A ⊂ X ,
then the same holds for the class C ∩ A defined by (10).
2) Let G be a class of real-valued functions on a set X , g an extremal point of G with
degree d(g) and φ(x) = xα for some positive α. Let F be a class of non-negative functions
on X such that
i) φ(F ) = {φ(f ) | f ∈ F } ⊂ G ;
ii) there exists f ∈ F such that φ(f ) = g.
Then f is extremal in F with degree not larger than d(g).
3) Let C be a class of subsets of X and A1 , . . . , Ak be a partition of X . If for all
j ∈ {1, . . . , k}, C ∩ Aj is a VC-class with dimension not larger than dj then C is a VCP
class with dimension not larger than d = kj=1 dj .
Proof. Let B ⊂ X be a set with cardinality d + 1. Either B ⊂ A and C ∩ A ∩ B = C ∩ B
for all C ∈ C so that B cannot be shattered by C or B ∩ Ac is not empty and cannot be of
the form C ∩ A ∩ B, which proves our first statement. The second statement follows from
the fact that C (F , f ) = C (φ(F ), g) ⊂ C (G , g). For the third one, we argue as follows: if
C could shatter d + 1 points, there would exist some j ∈ {1, . . . , k} and dj + 1 points of Aj
that could be shattered by C and hence by C ∩ Aj . This would be contradictory with the
fact that C ∩ Aj is a VC-class with dimension not larger than dj .
7.2. Proof of Theorem 2. For d ∈ N∗ , let Λ(d) = {s ∈ Λ, d(s) = d}. Since S is assumed
to be separable, Λ(d) ⊂ S is also separable and we may therefore choose a countable and
dense subset Λ(d) of Λ(d) for each d ∈ N∗ . Let us now choose a countable and dense
S
S
subset S for S. Possibly changing S into
S, we may assume with no loss
d≥1 Λ(d)
∗
of generality that Λ(d) ⊂ S for all d ∈ N . Finally, we define our estimator as (any) ρestimator sb of s based on S following the construction described in Section 4.2 of Baraud
et al. (2014) as well as the notations of this paper.
16
For y ≥ 1 and s ∈ Λ, we set
B S (s, s, y) = t ∈ S h2 (s, t) + h2 (s, s) ≤ y 2 /n .
Note that B S (s, s, y) may be empty. We start our proof with the following lemma.
Lemma 6. For all y ≥ 1 and s ∈ Λ
n p
o
t/s , t ∈ B S (s, s, y)
F (S, s, y) = ψ
is a weak VC-major class with dimension not larger than d(s).
Proof.
√ Since (S/s) is weak VC-major with dimension not larger than d(s) and the map x 7→
ψ( x) is increasing from [0, +∞] to [−1, 1], it follows from Baraud (2016, Proposition 3)
that
n p
o
t/s , t ∈ S
F ′ (S, s, y) = ψ
is weak VC-major with dimension not larger than d(s) and so is F (S, s, y) ⊂ F ′ (S, s, y).
Let us now go on with the proof of Theorem 2. We fix y ≥ √
1, s ∈ Λ and d = d(s). It
follows from Baraud (2011, Proposition 3 on page 386 with ψ/ 2 in place of ψ) and the
definition of B S (s, s, y) that, for all t ∈ B S (s, s, y),
"
!#
r
2
t
6y
2
2
2
Es ψ
(X1 )
≤ 6 h (s, t) + h (s, s) ∧ 1 ≤
∧ 1.
s
n
Since S is countable and ψ bounded by 1, the family F (S, s, y) is also countable and its
elements are bounded by 1. Besides, Lemma 6 ensures that F (S, s, y) is a weak VCmajor class with dimension not larger than d ≥ 1. We may therefore apply Corollary 1 of
Baraud (2016) to the family F (S, s, y) with b = 1, σ 2 = (6y 2 /n) ∧ 1 and get
"
#
n
X
sup
f (Xi ) − Es [f (Xi )]
wS (s, s, y) = Es
f ∈F (S,s,y) i=1
with
q
e
≤ 4 2nΓ(d) × σ log
+ 16Γ(d)
σ
q
n
≤ 8 3Γ(d) × y log e
∨1
+ 16Γ(d),
6y 2
Γ(d) = log 2
d∧n
X
n
j=0
j
e
≤ Γ(d)
= log 2 + (d ∧ n) log
en
.
d∧n
q
e
e
e
In particular, if y 2 ≥ Γ(d)/6
≥ (d ∧ n)/6 then Γ(d)
≤ y 6Γ(d),
hence
q
q
h en
√ i
S
e
e
for y ≥ Γ(d)/6.
(17)
w (s, s, y) ≤ 8y 3Γ(d) log
+2 2
n∧d
We recall that the quantity D S (s, s) is defined in Section 4.3 of Baraud et al. (2014) by
√
2−1
2
S
2
S
and c0 = √ .
D (s, s) = y0 ∨ 1 with y0 = sup y ≥ 0 w (s, s, y) > c0 y
2 2
17
q
e
or
implies that either y < Γ(d)/6
q
h en
i
√
e
y ≤ (c0 y)−1 wS (s, s, y) ≤ 8c−1
3Γ(d)
log
+ 2 2 = B.
0
n∧d
n
o
e
Since in both cases, y 2 ≤ max Γ(d)/6;
B 2 = B 2 and d = d(s), we deduce that
en
n
2
3
3
S
≤ κd(s) log+
(18)
D (s, s) ≤ B ≤ κ[d(s) ∧ n] log
d(s) ∧ n
d(s)
for all s ∈ Lµ , s ∈ S and some positive numerical constant κ. We now use Theorem 1
in Baraud et al. (2014) for which weSrecall that the notation h2 (t, t′ ) defined for densities
t, t′ ∈ Lµ means nh2 (t, t′ ). Since d≥1 Λ(d) ⊂ S, we obtain that for all ξ > 0, with
probability at least 1 − e−ξ ,
n ξ
ξ
D S (s, s)
d
3
′ 2
2
2
+ .
(19) C h (s, sb) ≤ inf h (s, s) +
+ ≤ inf inf h (s, s) + κ log+
s∈S
n
n d≥1 s∈Λ(d)
n
d
n
It follows from (17) that c0
y2
<
wS (s, s, y)
Finally, Λ(d) being dense in Λ(d),
inf h2 (s, s) = inf h2 (s, s) for all d ∈ N∗
s∈Λ(d)
s∈Λ(d)
and the bracketed term on the right-hand side of (19) becomes
#
"
n
n
d
d(s)
3
2
= inf h2 (s, s) + κ
log3+
inf
.
inf h (s, s) + κ log+
d≥1 s∈Λ(d)
n
d
n
d(s)
s∈Λ
Our conclusion follows.
7.3. Proof of Theorem 1. Let D ≥ 1, λ ≥ 0 and s ∈ V (D) be based on the partition
I ∈ J (D + 2). For all t = f 1l(x,+∞) and I ∈ I, the positive part (t − λs)+ of t − λs is 0
on (−∞, x] ∩ I and is non-increasing on I ∩ (x, +∞). Consequently, {t − λs > 0} ∩ I =
{(t − λs)+ > 0} ∩ I ∩ (x, +∞) is a sub-interval of I (possibly empty) and C (S, s, λ) ∩ I
is therefore VC with dimension not larger than 2. By Lemma 5, C (S, s, λ) is VC with
dimension not larger than 2(D + 2) and by Proposition 2 the element s is extremal in S
with dimension not larger than 4(D + 2). Finally Theorem 1 follows from Theorem 2.
7.4. Proof of Proposition 1. It relies on a series of approximation lemmas that shall
also prove useful in the sequel.
Lemma 7. Let f be a monotone function with finite variation VI (f ) on some interval I of
finite length l. Then
Z
Z
2
l [VI (f )]2
1
with
f (x) dx
f (x) − f dx ≤
f=
4
l I
I
and the factor 1/4 is optimal.
Proof. Assuming, without loss of generality, that f is non-increasing, let us observe that
one can replace f by g with g(x) = f (x − c) − f where c is the left-hand point of I. This
amounts to assume that c = 0 = f . Let f (0+) = a, f (l−) = −b, a + b = VI (f ) = R and
λ = l−1 sup{x | f (x) > 0} ∈ (0, 1). Then
Z l
Z λl
f (x) dx = Al ≤ l min{aλ, b(1 − λ)} = l min{(R − b)λ, b(1 − λ)}.
f (x) dx = −
0
λl
18
A maximization with respect to b and λ shows that A ≤ R/4 and it follows that
Z l
Z λl
Z l
lR2
2
2
f 2 (x) dx ≤ (a + b)Al = RAl ≤
f (x) dx +
.
f (x) dx =
4
λl
0
0
The optimality follows by considering the case of f = (R/2) 1l(0,l/2] − 1l(l/2,l) .
Our next lemma involves the norm in L2 (R, B(R), dx) hereafter denoted by k·k.
Lemma 8. Let f be a non-increasing function on (a, b) with finite variation V(a,b) (f ) < R.
For all D ≥ 1, there exists a partition I of (a, b) into at most D intervals and a function
fI which is piecewise constant on each element of the partition I and non-increasing such
that f (b−) ≤ fI ≤ f (a+), fI 1l(a,b) ≤ f 1l(a,b) and
√
Z b
Z b
R b−a
f (x) dx,
(f − fI )1l(a,b) ≤
fI (x) dx =
.
(20)
2D
a
a
Besides, there exists a partition I ′ of (a, b) into at most 2D intervals of length not larger
than (b − a)/D such that for all I ∈ I ′ , VI (f ) ≤ RD −1 . The same results hold for nondecreasing functions on (a, b).
Proof. Clearly, the results remain valid if we replace f by g with g = f almost everywhere
(with respect to the Lebesgue measure), g(a+) = f (a+) and g(b−) = f (b−). Since f is
non-increasing on (a, b), for all x ∈ (a, b) f (x+) exists and f admits an at most countable
number of discontinuities. We may therefore assume that f is actually defined on [a, b],
right-continuous on [a, b) and left-continuous at b.
Starting from x0 = a, define recursively for all j ≥ 1,
xj = sup x ∈ [xj−1 , b], f (xj−1 ) − f (x) ≤ RD −1 .
If k ≥ 1 and xk < b, f (xk−1 ) − f (x) > RD −1 for all x > xk hence f (xk ) − f (xk−1 ) ≥ RD −1
since f is right-continuous. In particular for such a k, we necessarily have
R > f (a) − f (xk ) =
k
X
j=1
f (xj−1 ) − f (xj ) ≥ kRD −1 ,
which implies that k < D. The process therefore results in a finite number of distinct points
x0 = a < x1 < . . . < xK+1 = b with K + 1 ≤ D. It also follows from the definition of the
xj that f (xj−1 ) − f (xj −) ≤ RD −1 for 1 ≤ j ≤ K + 1. Let us now set
−1
f j = (xj − xj−1 )
Z
xj
f (x) dx
and
fI =
xj−1
K+1
X
f j 1l(xj−1 ,xj ] .
j=1
Rb
Rb
Note that f (b−) ≤ fI ≤ f (a+), a fI (x) dx = a f (x) dx and that fI is non-increasing
and piecewise constant on a partition of (a, b) into K + 1 intervals. Since, for all j, 0 ≤
f (xj−1 ) − f (xj −) ≤ RD −1 , it follows from Lemma 7 that
(f − fI )1l(a,b)
2
=
K+1
X Z xj
j=1
xj−1
f − fj
2
dx ≤
19
R
2D
2 K+1
X
j=1
(xj − xj−1 ) =
R
2D
2
(b − a).
Moreover Jensen’s Inequality implies that
!2 Z
Z xj
Z xj
xj
1
2
fI (x) dx = (xj − xj−1 )
f (x) dx
≤
f 2 (x) dx,
x
−
x
j
j−1 xj−1
xj−1
xj−1
which shows that fI 1l(a,b) ≤ f 1l(a,b) and proves the first part of the lemma.
For the second part, define I ′ as follows: for each element I ∈ I with length ℓ(I) larger
than (b − a)/D divide I into ⌈Dℓ(I)/(b − a)⌉ intervals of length not larger than (b − a)/D.
The process results in a new partition I ′ thinner than I and its cardinality is not larger
than
#
"
X Dℓ(I) X Dℓ(I)
D X
ℓ(I) + |I| ≤ 2D.
+1 ≤
≤
b−a
b−a
b−a
I∈I
I∈I
I∈I
RD −1
Since by construction VI (f ) ≤
for all I ∈ I, this property is also true for the elements
I of the partition I ′ which is thinner than I. For non-decreasing functions, change f to
−f .
Lemma 9. Given two probability densities t, u with respect to µ,
√
√
for all λ ∈ R.
t−λ u
(21)
h(t, u) ≤
2
In particular,
√ if f is a non-negative element in L2 (µ) such that kf k > 0 and u = (f /kf k) ,
h(t, u) ≤
t − f for any probability density t with respect to µ.
√
√
scalar
Proof. We notice
R √ that t and u are two vectors of norm one in L2 (µ) and their
√
product is X ut dµ = cos α for some α ∈ [0, π/2]. It implies that v = cos α u is the
√
√
orthogonal projection of t on the linear space generated by u, hence
√
√
√
√
inf
t−λ u =
t − cos α u = sin α.
λ∈R
Inequality (21) follows from the fact that h2 (t, u) = 1 − cos α ≤ sin2 α for all α ∈ [0, π/2].
The last result is obtained from (21) with λ = kf k.
√
p the proof of Proposition 1, we apply Lemma 8 with f = t and R >
p To complete
t(a+) − t(b−). The resulting function fI is then nonnegative, non-increasing on (a, b)
and satisfies 0 < kfI k. Setting sI = fI2 / kfI k2 , which is an element of V (D), we
√ may apply
s
)
≤
kf
−
f
k
≤
R
b − a/(2D).
the last part of Lemma 9 with f = fI which gives h(t,
I
√ I
The conclusion follows by letting R converge to V[a,b]
t .
√
7.5. Proof of Proposition 4. Let t be based on I = {I0 , . . . , Ik+1 }, Rj > VIj
t for
1 ≤ j ≤ k and let D1 , . . . , Dk be positive integers. On the intervals I0 and Ik+1 , t is equal
to 0 and, for all other intervals of I, one can apply Lemma 8 to find an approximation fj of
√
tj which is monotone, piecewise constant with Dj pieces on Ij and satisfies, according to
p
√
P
(20), fj − t 1lIj ≤ Rj ℓ(Ij )/(2Dj ). Therefore, if f = kj=1 fj 1lIj and u = (f /kf k)2 ,
we derive from Lemma 9 that
k
k
X
2
√ 2 X
√
ℓ(Ij )Rj2
2
h (t, u) ≤ f − t =
= M.
fj − t 1lIj ≤
4Dj2
j=1
j=1
20
Moreover, we can always assume (modifying it on a negligeable set if necessary) that u
Pk
belongs to V (D ′ ) ∩ Fk+2 with D ′ =
j=1 Dj . Given D, a formal minimization with
P
Pk
−2
2
respect to the xj > 0 of j=1 ℓ(Ij )Rj xj under the condition that kj=1 xj ≤ D leads to
xj = λ
1/3
ℓ(Ij )Rj2
with
k
X
xj = D,
j=1
1/3
P
. Taking into account the fact that the Dj should
so that λ−1 = D −1 kj=1 ℓ(Ij )Rj2
∗
belong to N , we finally set
−1
k
X
1/3
1/3
,
ℓ(Ij )Rj2
ℓ(Ij )Rj2
Dj =
D
j=1
which implies that
Pk
j=1 Dj
≤ D + k and
2
3
k
k
k
X
X
X
1
1/3
1/3
1/3
= 1
.
M≤
ℓ(Ij )Rj2
ℓ(Ij )Rj2
ℓ(Ij )Rj2
2
2
4D
4D
j=1
j=1
j=1
The corresponding function u belongs to V (D + k) ∩ Fk+2 so that
3
k
X
1/3
1
.
h2 t, V (D + k) ∩ Fk+2 ≤
ℓ(Ij )Rj2
2
4D
j=1
The conclusion follows by letting each Rj converge to VIj
√
t .
7.6. Proof of Proposition 6. It relies on the following approximation lemma.
Lemma 10. Let f be a continuous and either convex or concave function on [a, b] with
right-hand derivative f ′ on (a, b) satisfying V(a,b) (f ′ ) < +∞. The affine function g on [a, b]
defined by g(a) = f (a) and g(b) = f (b) satisfies
sup |f (x) − g(x)| ≤
a≤x≤b
b−a
V(a,b) (f ′ ).
4
The factor 1/4 is optimal.
Proof. Changing f into −f , we may assume that f is concave on [a, b]. In particular,
h(x) = f (x) − g(x) ≥ 0 for x ∈ [a, b] and since h is continuous on [a, b] and satisfies
h(a) = h(b) = 0, there exists some c ∈ (a, b) such that
Z c
Z b
f (b) − f (a)
′
sup h(x) = h(c) =
f (u) − ℓ du =
ℓ − f ′ (u) du with ℓ =
.
b−a
a≤x≤b
a
c
The function f ′ being non-increasing on (a, b),
h(c) ≤ (f ′ (a+) − ℓ)(c − a) ∧ (ℓ − f ′ (b−))(b − c)
21
and consequently,
b−c
c−a
′
′
f (a+) − ℓ (c − a)
h(c) ≤
+ ℓ − f (b−) (b − c)
b−a
b−a
(c − a)(b − c) ′
f (a+) − ℓ + ℓ − f ′ (b−)
=
b−a
′
c−a
b−a
c−a
= (b − a)
f (a+) − f ′ (b−) ≤
V(a,b) (f ′ ).
1−
b−a
b−a
4
The constant 1/4 cannot be improved since it is reached for f (x) = 1 − |x| on [−1, 1].
Let f ′ be the function of Lemma 10 and R > V(a,b) (f ′ ). By Lemma 8, one can partition
(a, b) into K ≤ 2D intervals Jj , 1 ≤ j ≤ K of length not larger than D −1 (b − a) with
VJj (f ′ ) < RD −1 . Using this partition to approximate f by a piecewise affine function gK
with K pieces and applying Lemma 10, we derive that
sup |f (x) − gK (x)| ≤ (1/4)RD −1 [(b − a)/D] = (R/4)(b − a)D −2 ,
a≤x≤b
hence
Z
b
a
|f (x) − gK (x)|2 dx ≤ (R/4)2 (b − a)3 D −4 .
Note that, by√construction, gK is concave on [a, b] if f is and gK is convex in the opposite
case. Since t satisfies the assumptions of Lemma 10 on each of the k non-extremal
intervals of the partition I that defines t and is zero on the two extremal intervals,√we may
use the previous approximation
method on each non-extremal interval with f = t to get
√
P
an approximation v of t with D ′ = 2 kj=1 Dj pieces and such that
√
2
t−v
k
≤
1 X 2
Rj [ℓ(Ij )]3 Dj−4
16
if
Rj > VI j
j=1
√ ′
t
for 1 ≤ j ≤ k.
Renormalizing v as in Lemma 9, we conclude that there exists u which belongs to V 1 (D ′ ) ∩
1
and
Fk+2
k
h2 (t, u) ≤ M =
1 X 2
Rj [ℓ(Ij )]3 Dj−4 .
16
j=1
We now mimic the proof of Proposition 4 to optimize the Dj and get
−1
k
X
1/5
1/5
[ℓ(Ij )]3 Rj2
[ℓ(Ij )]3 Rj2
Dj =
,
D
j=1
so that finally
Pk
j=1 Dj
≤ D + k and
4
5
k
k
k
X
X
1 X
1
1/5
1/5
1/5
.
=
[ℓ(Ij )]3 Rj2
[ℓ(Ij )]3 Rj2
[ℓ(Ij )]3 Rj2
M≤
16D 4
16D 4
j=1
j=1
j=1
22
1
The corresponding function u belongs to V 1 2(D + k) ∩ Fk+2
so that
5
k
X
1/5
1
1
.
[ℓ(Ij )]3 Rj2
h2 t, V 1 2(D + k) ∩ Fk+2
≤
4
16D
j=1
√
′
The conclusion follows by letting Rj converge to VIj
for each j.
t
References
Baraud, Y. (2011). Estimator selection with respect to Hellinger-type risks. Probab. Theory
Related Fields, 151(1-2):353–401.
Baraud, Y. (2016). Bounding the expectation of the supremum of an empirical process over
a (weak) VC-major class. To appear in Electron. J. Stat.
Baraud, Y., Birgé, L., and Sart, M. (2014). A new method for estimation and model
selection: ρ-estimation. http://arxiv.org/abs/1403.6057.
Bellec, P. C. (2015). Sharp oracle inequalities for least squares estimators in shape restricted
regression. http://arxiv.org/abs/1510.08029.
Birgé, L. (1987). Estimating a density under order restrictions: nonasymptotic minimax
risk. Ann. Statist., 15(3):995–1012.
Birgé, L. (1989). The Grenander estimator: a nonasymptotic approach. Ann. Statist.,
17(4):1532–1549.
Birgé, L. (2006). Model selection via testing: an alternative to (penalized) maximum
likelihood estimators. Ann. Inst. H. Poincaré Probab. Statist., 42(3):273–325.
Birgé, L. and Massart, P. (1998). Minimum contrast estimators on sieves: exponential
bounds and rates of convergence. Bernoulli, 4(3):329–375.
Brown, L. D., Low, M. G., and Zhao, L. H. (1997). Superefficiency in nonparametric
function estimation. Ann. Statist., 25(6):2607–2625.
Chatterjee, S., Guntuboyina, A., and Sen, B. (2015). On risk bounds in isotonic and other
shape restricted regression problems. Ann. Statist., 43(4):1774–1800.
Cule, M. and Samworth, R. (2010). Theoretical properties of the log-concave maximum
likelihood estimator of a multidimensional density. Electron. J. Stat., 4:254–270.
Doss, C. R. and Wellner, J. A. (2015). Global rates of convergence of the MLEs of logconcave and s-concave densities. to appear in Ann. Statist.
Dümbgen, L. and Rufibach, K. (2009). Maximum likelihood estimation of a log-concave
density and its distribution function: basic properties and uniform consistency. Bernoulli,
15(1):40–68.
Grenander, U. (1981). Abstract inference. John Wiley & Sons, Inc., New York. Wiley
Series in Probability and Mathematical Statistics.
Groeneboom, P. (1985). Estimating a monotone density. In Proceedings of the Berkeley
conference in honor of Jerzy Neyman and Jack Kiefer, Vol. II (Berkeley, Calif., 1983),
Wadsworth Statist./Probab. Ser., pages 539–555. Wadsworth, Belmont, CA.
Groeneboom, P., Jongbloed, G., and Wellner, J. A. (2001). Estimation of a convex function:
characterizations and asymptotic theory. Ann. Statist., 29(6):1653–1698.
Groeneboom, P. and Wellner, J. A. (1992). Information bounds and nonparametric maximum likelihood estimation, volume 19 of DMV Seminar. Birkhäuser Verlag, Basel.
Guntuboyina, A. and Sen, B. (2015). Global risk bounds and adaptation in univariate
convex regression. Probab. Theory Related Fields, 163(1-2):379–411.
23
Huber, P. J. (1981). Robust Statistics. John Wiley & Sons, Inc., New York. Wiley Series
in Probability and Mathematical Statistics.
Le Cam, L. (1973). Convergence of estimates under dimensionality restrictions. Ann.
Statist., 1:38–53.
Le Cam, L. (1986). Asymptotic Methods in Statistical Decision Theory. Springer Series in
Statistics. Springer-Verlag, New York.
van de Geer, S. (1993). Hellinger-consistency of certain nonparametric maximum likelihood
estimators. Ann. Statist., 21(1):14–44.
van de Geer, S. A. (2000). Applications of empirical process theory, volume 6 of Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press,
Cambridge.
Univ. Nice Sophia Antipolis, CNRS, LJAD, UMR 7351, 06100 Nice, France.
E-mail address: [email protected]
Sorbonne Universités, UPMC Univ. Paris 06, CNRS UMR 7599, LPMA, 75005 Paris, France.
E-mail address: [email protected]
24
| 10math.ST
|
Energy-efficient Delivery by Heterogeneous
Mobile Agents∗
Andreas Bärtschi1 , Jérémie Chalopin2 , Shantanu Das2 , Yann
Disser3 , Daniel Graf1 , Jan Hackfeld4 , and Paolo Penna1
1
ETH Zürich, Department of Computer Science, Switzerland
{ baertschi,daniel.graf,paolo.penna }@inf.ethz.ch
LIF, CNRS and Aix-Marseille Université, France
{ shantanu.das,jeremie.chalopin }@lif.univ-mrs.fr
2
arXiv:1610.02361v1 [cs.DS] 7 Oct 2016
3
4
TU Darmstadt, [email protected]
TU Berlin, Institut für Mathematik, Technische Universität Berlin, Germany
[email protected]
Abstract
We consider the problem of delivering m messages between specified source-target pairs in a
weighted undirected graph, by k mobile agents initially located at distinct nodes of the graph.
Each agent consumes energy proportional to the distance it travels in the graph and we are interested in optimizing the total energy consumption for the team of agents. Unlike previous related
work, we consider heterogeneous agents with different rates of energy consumption (weights wi ).
To solve the delivery problem, agents face three major challenges: Collaboration (how to work
together on each message), Planning (which route to take) and Coordination (how to assign
agents to messages).
We first show that the delivery problem can be 2-approximated without collaborating and that
this is best possible, i.e., we show that the benefit of collaboration is 2 in general. We also show
that the benefit of collaboration for a single message is 1/ ln 2 ≈ 1.44. Planning turns out to be
NP-hard to approximate even for a single agent, but can be 2-approximated in polynomial time
if agents have unit capacities and do not collaborate. We further show that coordination is NPhard even for agents with unit capacity, but can be efficiently solved exactly if they have uniform
wi
weights. Finally, we give a polynomial-time (4 max w
)-approximation for message delivery with
j
unit capacities.
1998 ACM Subject Classification F.2 Analysis of Algorithms and Problem Complexity, F.2.2
Nonnumerical Algorithms and Problems
Keywords and phrases message delivery, mobile agents, energy optimization, approximation
algorithms
Digital Object Identifier 10.4230/LIPIcs...Page
1
Introduction
Recent technological progress in robotics allows the mass production of inexpensive mobile
robots which can be used to perform a variety of tasks autonomously without the need
for human intervention. This gives rise to a variety of algorithmic problems for teams of
autonomous robots, hereafter called mobile agents. We consider here the delivery problem of
∗
This work was partially supported by the project ANR-ANCOR (anr-14-CE36-0002-01), the SNF
(project 200021L_156620) and the DFG Priority Programme 1736 “Algorithms for Big Data” (grant
SK 58/10-1).
© A. Bärtschi, J. Chalopin, S. Das, Y. Disser, D. Graf, J. Hackfeld, P. Penna;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
Page:2
Energy-efficient Delivery by Heterogeneous Mobile Agents
moving some objects or messages between various locations. A mobile agent corresponds to
an automated vehicle that can pick up a message at its source and deliver it to the intended
destination. In doing so, the agent consumes energy proportional to the distance it travels.
We are interested in energy-efficient operations by the team of agents such that the total
energy consumed is minimized.
In general the agents may not be all identical; some may be more energy efficient than
others if they use different technologies or different sources of power. We assume each agent
has a given weight which is the rate of energy consumption per unit distance traveled by this
agent. Moreover, the agents may start from distinct locations. Thus it may be sometimes
efficient for a agent to carry the message to some intermediate location and hand it over to
another agent which carries it further towards the destination. On the other hand, an agent
may carry several messages at the same time. Finding an optimal solution that minimizes
the total energy cost involves scheduling the moves of the agents and the points where they
pick up or handover the messages. We study this problem (called WeightedDelivery) for
an edge-weighted graph G which connects all sources and destinations. The objective is to
deliver m messages between specific source-target pairs using k agents located in arbitrary
nodes G. Note that this problem is distinct from the connectivity problems on graphs or
network flow problems since the initial location of the agents are in general different from the
sources where the messages are located, which means we need to consider the cost of moving
the agents to the sources in addition to the cost of moving the messages. Furthermore, there
is no one-to-one correspondence between the agents and the messages in our problem.
Previous approaches to energy-efficient delivery of messages by agents have focused on
a bottleneck where the agents have limited energy (battery power) which restricts their
movements [1, 7]. The decision problem of whether a single message can be delivered without
exceeding the available energy for any agent is known as the DataDelivery problem [8] or
the BudgetedDelivery problem [4] and it was shown to be weakly NP-hard on paths [8] and
strongly NP-hard on planar graphs [4].
Our Model. We consider an undirected edge-weighted graph G = (V, E). Each edge e ∈ E
has a cost (or length) denoted by le . The length of a simple path is the sum of the lengths
of its edges. The distance between nodes u and v is denoted by dG (u, v) and is equal to
the length of the shortest path from u to v in G. There are k mobile agents denoted by
a1 , . . . ak and having weights w1 , . . . wk . These agents are initially located on arbitrary nodes
p1 , . . . , pk of G. We denote by d(ai , v) the distance from the initial location of ai to node v.
Each agent can move along the edges of the graph. Each time an agent ai traverses an edge
e it incurs an energy cost of wi · le . Furthermore there are m pairs of (source, target) nodes
in G such that for 1 ≤ i ≤ m, a message has to be delivered from source node si to a target
node ti . A message can be picked up by an agent from any node that it visits and it can be
carried to any other node of G, and dropped there. The agents are given a capacity κ which
limits the number of messages an agent may carry simultaneously. There are no restrictions
on how much an agent may travel. We denote by dj the total distance traveled by the j-th
agent. WeightedDelivery is the optimization problem of minimizing the total energy
Pk
j=1 wj dj needed to deliver all messages.
A schedule S describes the actions of all agents as a sequence (ordered list) of pick-up
actions (aj , p, mi , +) and drop-off actions (aj , q, mi , −), where each such tuple denotes the
action of agent aj moving from its current location to node p (node q) where it picks up
message mi (drops message mi , respectively). A schedule S implicitly encodes all the pick-up
Pk
and drop-off times and it is easy to compute its total energy use of cost(S) := j=1 wj dj .
Bärtschi, Chalopin, Das, Disser, Graf, Hackfeld and Penna
m1
a1 , w1 =2 s1
2
2
m2
a2 , w2 =3 s2
u
t2
2
1
t1
An optimal, feasible schedule:
(a1 , s1 , m1 , +), (a2 , s2 , m2 , +), (a1 , u , m1 , −),
(a2 , u , m2 , −), (a1 , u , m2 , +), (a1 , t2 , m2 , −),
(a2 , u , m1 , +), (a2 , t1 , m1 , −)
Figure 1 Example of an optimal, feasible schedule for two messages and two agents.
We denote by S|aj the subsequence of all actions carried out by agent aj and by S|mi the
subsequence of all actions involving pick-ups or drop-offs of message mi . We call a schedule
feasible if every pick-up action (_, p, mi , +), p 6= si , is directly preceded by a drop-off action
(_, p, mi , −) in S|mi and if all the messages get delivered, see Figure 1.
Our Contribution. Solving WeightedDelivery naturally involves simultaneously solving
three subtasks, collaboration, individual planning, and coordination: First of all, if multiple
agents work on the same message, they need to collaborate, i.e., we have to find all intermediate
drop-off and pick-up locations of the message. Secondly, if an agent works on more than
one message, we have to plan in which order it wants to approach its subset of messages.
Finally, we have to coordinate which agent works on which subset of all messages (if they
do this without collaboration, the subsets form a partition, otherwise the subsets are not
necessarily pairwise disjoint). Even though these three subtasks are interleaved, we investigate
collaboration, planning and coordination separately in the next three sections. This leads us
to a polynomial-time approximation algorithm for WeightedDelivery, given in Section 5.
In Section 2 we consider the Collaboration aspect of WeightedDelivery. We first
present a polynomial time solution for WeightedDelivery when there is only a single
message (m = 1). The algorithm has complexity O(|V |3 ) irrespective of the number of
agents k. In general, we show that any algorithm that only uses one agent for delivering
every message cannot achieve an approximation ratio better than what we call the benefit of
m
collaboration (BoC) which is at least 1/ ln ((1 + 1/(2m)) (1 + 1/(2m + 1))). We show this
to be tight for m = 1 (where BoC ≥ 1/ ln 2) and m → ∞ (where BoC → 2).
In Section 3 we look at the Planning aspect of WeightedDelivery. Individual planning
by itself turns out to be NP-hard on planar graphs and NP-hard to approximate within a
factor of less than 367
366 . On the positive side, we give approximation guarantees for restricted
versions of WeightedDelivery which turn out to be useful for the analysis in Section 5.
In Section 4 we study the Coordination aspect of WeightedDelivery. Even if collaboration and planning are taken care of (i.e., a schedule is fixed except for the assignment of
agents to messages), Coordination also turns out to be NP-hard even on planar graphs. The
result holds for any capacity, including κ = 1. This setting, however, becomes tractable if
restricted to uniform weights of the agents.
In Section 5 we give a polynomial-time approximation algorithm for WeightedDelivery
wi
with an approximation ratio of 4 · max w
for κ = 1. Due to the limited space, some proofs
j
are deferred to the appendix.
Related Work. The problem of communicating or transporting goods between sources
and destinations in a graph has been well studied in a variety of models with different
optimization criteria. The problem of finding the smallest subgraph or tree that connects
multiple sources and targets in a graph is called the point-to-point connection problem and
is known to be NP-hard [25]. The problem is related to the more well-known generalized
Steiner tree problem [28] which is also NP-hard. Unlike these problems, the maximum flow
Page:3
Page:4
Energy-efficient Delivery by Heterogeneous Mobile Agents
problem in a network [13], puts a limit on the number of messages that can be transported
over an edge, which makes the problem easier allowing for polynomial time solutions. In all
these problems, however, there are no agents carrying the messages as in our problem.
For the case of a single agent moving in a graph, the task of optimally visiting all nodes,
called the Traveling salesman problem or visiting all edges, called the Chinese postman
problem have been studied before. The former is known to be NP-hard [2] while the latter
can be solved in O(|V |2 |E|) time [12]. For metric graphs, the traveling salesman problem has
a polynomial-time 32 -approximation for tours [9] and for paths with one fixed endpoint [18].
For multiple identical agents in a graph, Demaine et al. [11] studied the problem of moving
the agents to form desired configurations (e.g. connected or independent configurations) and
they provided approximation algorithms and inapproximability results. Bilo et al. [6] studied
similar problems on visibility graphs of simple polygons and showed many motion planning
problems to be hard to approximate.
Another optimization criteria is to minimize the maximum energy consumption by any
agent, which requires partitioning the given task among the agents. Frederickson et al. [16]
studied this for uniform weights and called it the k-stacker-crane problem and they gave
approximation algorithms for a single agent and multiple agents. Also in this minmax context,
the problem of visiting all the nodes of a tree using k agents starting from a single location
is known to be NP-hard [15]. Anaya et al. [1] studied the model of agents having limited
energy budgets. They presented hardness results (on trees) and approximation algorithms
(on arbitrary graphs) for the problem of transferring information from one agent to all others
(Broadcast) and from all agents to one agent (Convergecast). For the same model, message
delivery between a single s-t node pair was studied by Chalopin et al. [7, 8, 4] as mentioned
above. A recent paper [10] shows that these three problems remain NP-hard for general
graphs even if the agents are allowed to exchange energy when they meet.
2
Collaboration
In this section, we examine the collaboration of agents: Given for each message mi all the
agents ai1 , ai2 , . . . , aix which at some point carry the message, one needs to find all pick-up and
drop-off locations (handovers) h1 , . . . , hy for the schedule entries (ai1 , _, mi , +), (ai1 , _, mi , −),
. . . , (aix , _, mi , −). Note, that in general we can have more than two action quadruples
(aij , _, mi , +/−) per agent aij . When there is only a single message overall (m = 1), we will
use a structural result to tie together WeightedDelivery and Collaboration. For multiple
messages, however, this no longer holds: In this case, we analyze the benefit we lose if we
forgo collaboration and deliver each message with a single agent.
2.1
An Algorithm for WeightedDelivery of a Single Message
I Lemma 1. In any optimal solution to WeightedDelivery for a single message, if the
message is delivered by agents with weights w1 , w2 , . . . wk , in this order, then (i) wi ≥ wj
whenever i < j, and (ii) without loss of generality, wi 6= wj for i 6= j. Hence there is an
optimal schedule S in which no agent aj has more than one pair of pick-up/drop-off actions.
Proof. If agent ai hands the message over to agent aj with wi < wj in any solution, we
can construct a better solution by replacing aj ’s trajectory carrying the message with the
same trajectory using agent ai . By the same argument we may also assume without loss of
generality that the weights of the agents carrying the message in an optimum schedule are
strictly decreasing, since we can merge trajectories of equal weight.
J
Bärtschi, Chalopin, Das, Disser, Graf, Hackfeld and Penna
I Example 2. This example shows that Lemma 1 is not true for more than one message.
In the graph shown in Figure 2, we let one agent a1 with weight w1 = 1 start in vertex s2
and a second agent a2 with weight w2 = 1.5 start in vertex v1 . In the optimal schedule, the
message 2 with starting location s2 is first transported by a1 to v1 and from there by a2 to
its destination t2 . Thus, the weights of the agents transporting the message 2 are increasing
in this case.
s1
1
t1
1
s2
1
v1
1
t2
Figure 2 Example where the weights of the agents in the order they are transporting a message
is increasing.
I Theorem 3. An optimal solution of WeightedDelivery of a single message in a graph
G = (V, E) with k ≤ |V | agents can be found in O(|V |3 ) time.
Proof. We use the properties of Lemma 1 to create an auxiliary graph on which we run
Dijkstra’s algorithm for computing a shortest path from s to t. Given an original instance of
single-message WeightedDelivery consisting of the graph G = (V, E), with s, t ∈ V , we
obtain the auxiliary, directed graph G0 = (V 0 , E 0 ) as follows:
For each node v ∈ V and each agent ai , there is a node vai in G0 .
Furthermore G0 contains two additional vertices s and t.
For 1 ≤ i ≤ k, there is an arc (s, sai ) of cost wi · dG (pi , s) and an arc (tai , t) of cost 0.
For (u, v) ∈ E and 1 ≤ i ≤ k, there are two arcs (uai , vai ) and (vai , uai ) of cost wi · l(u,v) .
For u ∈ V and agents ai , aj with wi > wj , there is an arc (uai , uaj ) of cost wj · dG (pj , u).
Note that any solution to the WeightedDelivery that satisfies the properties of Lemma 1
corresponds to some s-t-path in G0 such that the cost of the solution is equal to the length of
this path in Ga and vice versa. This implies that the length of the shortest s-t path in G0 is
the cost of the optimal solution for WeightedDelivery in G. Assuming that k ≤ |V |, the
graph G0 has |V |·k+2 ∈ O(|V |2 ) vertices and at most 2k+(k 2 |V |+|V |2 k)/2−|V |·k ∈ O(|V |3 )
arcs. The graph G0 can be constructed in O(|V |3 ) time if we use the Floyd Warshall all pair
shortest paths algorithm [14, 27] in G. Finally, we compute the shortest path from s to t in
G0 in time O(|V |3 ), using Dijkstra’s algorithm with Fibonacci heaps.
J
Unfortunately, the structural properties of Lemma 1 do not extend to multiple messages.
In the next two subsections we investigate how the quality of an optimal solution changes if
we only allow every message to be transported by one agent. Different messages may still
be transported by different agents and one agent may also transport multiple messages at
the same time as long as the number of messages is at most the capacity κ. To this end we
define the Benefit of Collaboration as the cost ratio between an optimal schedule Opt and a
best-possible schedule without collaboration S, BoC = minS cost(S)/cost(Opt).
2.2
Lower Bound on the Benefit of Collaboration
I Theorem 4. On instances of WeightedDelivery with agent capacity κ and m messages,
an algorithm using one agent for delivering every message cannot achieve an approximation
r
ratio better than 1/ ln ((1 + 1/(2r)) (1 + 1/(2r + 1))), where r := min{κ, m}.
Page:5
Page:6
Energy-efficient Delivery by Heterogeneous Mobile Agents
Proof. Consider the graph G = (V, E) given in Figure 3, where the length le of every edge e
is 1/n. This means that G is a star graph with center v0,n and r + 1 paths of total length
1 each. We have r messages and message i needs to be transported from vi,0 to v0,2n for
2r
i = 1, . . . , r. There further is an agent ai,j with weight wi,j = 2r+j/n
starting at every vertex
vi,j for (i, j) ∈ {1, . . . , r} × {0, . . . , n − 1} ∪ {0} × {n, . . . , 2n}.
We first show the following: If any agent transports s messages i1 , . . . , is from vij ,0 to
v0,2n , then this costs at least 2s. Note that this implies that any schedule S for delivering
all messages by the agents such that every message is only carried by one agent satisfies
cost(S) ≥ 2r.
So let an agent ai,j transport s messages from the source to the destination v0,2n .
Without loss of generality let these messages be 1, . . . , s, which are picked up in this order.
By construction, agent ai,j needs to travel a distance of at least nj to reach message 1, then
distance 1 to to move back to v0,n , then distance 2 for picking up message i and going back
to v0,n for i = 2, . . . , s, and finally it needs to move distance 1 from v0,n to v0,2n . Overall,
agent ai,j therefore travels a distance of at least 2s + nj . The overall cost for agent ai,j to
deliver the s messages therefore is at least
(2s + nj ) · wi,j = (2s + nj ) ·
2r
2r+j/n
≥ (2s + nj ) ·
2s
2s+j/n
= 2s.
Now, consider a schedule Scol , where the agents collaborate, i.e., agent ai,j transports
message i from vi,j to vi,j+1 for i = 1, . . . , r, j = 0, . . . , n − 1, where we identify vi,n with
v0,n . Then agent a0,j transports all r messages from v0,j to v0,j+1 for j = n, . . . , 2n − 1. This
is possible because r ≤ κ by the choice of r. The total cost of this schedule is given by
Z 1
Z 2
cost(Scol ) = r ·
fstep (x)dx +
fstep (x)dx,
0
1
where fstep (x) is a step-function defined on [0, 2] giving the current cost of transporting
2r
the message, i.e., fstep (x) = 2r+j/n
on the interval [j/n, (j + 1)/n) for j = 0, . . . , 2n − 1.
The first integral corresponds the the first part of the schedule, where the r messages are
transported separately and therefore the cost of transporting message i from vi,j to vi,j+1 is
R (j+1)/n
2r
fstep (x)dx = n1 · 2r+j/n
exactly j/n
. The second part of the schedule corresponds to the
part, where all r messages are transported together by one agent at a time.
1
Observe that the function f (x) = 2r · 2r−1/n+x
satisfies f (x) ≥ fstep (x) on [0, 2], hence
Z
1
cost(Scol ) ≤ r
0
= 2r ln
2
1
2
f (x)dx = 2r r ln(2r − n1 + x) + ln(2r − n1 + x)
0
1
1
r
r
n→∞
2r−1/n+1
2r−1/n+2
1
1
→ 2r ln 1 + 2r
1 + 2r+1
.
2r−1/n
2r−1/n+1
Z
f (x)dx +
The best approximation ratio of an algorithm that transports every message by only one
agent compared to an algorithm that uses an arbitrary number of agents for every message
v1,0
v1,1
v1,n−1
v2,0
v2,1
v2,n−1
v0,n
vr,0
vr,1
v0,n+1
vr,n−1
Figure 3 Lower bound construction for the benefit of collaboration.
v0,2n
Bärtschi, Chalopin, Das, Disser, Graf, Hackfeld and Penna
Page:7
is therefore bounded from below by
BoC ≥ min cost(S)/cost(Scol ) ≥ 1/ ln
S
1+
1 r
2r
1+
1
2r+1
.
J
r
By observing that limr→∞ 1/ ln ((1 + 1/(2r)) (1 + 1/(2r + 1))) = 1/ ln e1/2 = 2, we
obtain the following corollary.
I Corollary 5. A schedule for WeightedDelivery where every message is delivered by
a single agent cannot achieve an approximation ratio better than 2 in general, and better
than 1/ ln 2 ≈ 1.44 for a single message.
2.3
Upper Bounds on the Benefit of Collaboration
We now give tight upper bounds for Corollary 5. The following theorem shows that the
benefit of collaboration is 2 in general. We remark that finding an optimal schedule in
which every message is transported from its source to its destination by one agent, is already
NP-hard, as shown in Theorem 9.
I Theorem 6. Let Opt be an optimal schedule for a given instance of WeightedDelivery.
Then there exists a schedule S such that every message is only transported by one agent and
cost(S) ≤ 2 · cost(Opt).
Proof. We can assume without loss of generality that in the optimal schedule Opt every
message i is transported on a simple path from its starting point si to its destination ti . This
can be easily achieved by letting agents drop the messages at intermediate vertices if they
would otherwise transport it in a cycle. We define the directed multigraph GS = (V, E ∪˙ E)
as follows:
V is the set of vertices of the original graph G.
For every time in the optimal schedule that an agent traverses an edge {u, v} from u to v
while carrying a set of messages M 0 , we add the arc e = (u, v) to E and ē = (v, u) to E.
We further label both edges with the set of messages M 0 and write Me = Mē = M 0 to
denote these labels. We call the edges in E original edges and the edges in E reverse
edges.
We say that the tour of an agent A satisfies the edge labels, if every original edge e ∈ E is
traversed at most once by A and only while carrying the exact set of messages Me , and every
reverse edge ē ∈ E is traversed by A at most once and without carrying any message.
We will show that there exists a Eulerian tour satisfying the edge labels of every connected
component of GS . We then let the cheapest agent in each connected component follow the
respective Eulerian tour. This agent traverses every edge exactly twice as often as the edge
is traversed in the optimal schedule Opt by all agents. As we choose the cheapest agent in
each connected component, we obtain a schedule S with cost(S) ≤ 2 · cost(Opt).
By only considering a subset of the messages and a subschedule of Opt, we may from
now on assume that GS is strongly connected (by construction, every connected component
of GS is strongly connected). Further, let amin be an agent with minimum cost among the
agents that move in Opt, let M (v) be the set of messages currently placed on vertex v,
and let M (amin ) be the set of messages currently transported by agent amin . We first show
that the procedure computeTour computes a closed tour for amin that satisfies the edge
labels, and afterwards we explain how we can iterate the procedure to obtain a Eulerian tour
satisfying the edge labels.
Claim 1: If the agent amin starts in a vertex v0 and follows the tour computed by
computeTour, it satisfies the edge labels in every step and returns to its starting location.
Page:8
Energy-efficient Delivery by Heterogeneous Mobile Agents
Algorithm 1 computeTour
function computeTour
drop all messages
if ∃ edge e ∈ E incident to current vertex then
if M (current vertex) ⊇ Me then
pickup messages Me , traverse e and delete it
else
let j ∈ Me \ M (current vertex)
fetchMessage(j, currentVertex)
else if ∃ edge e ∈ E then
traverse e and delete it
Algorithm 2 fetchMessage
function fetchMessage(i, v)
drop all messages
while i ∈
/ M (current vertex) do
if there is a edge ē ∈ E with i ∈ Me leaving the current vertex then
traverse ē and delete it
else
give up
while v 6= current vertex do
let e ∈ E be edge incident to current vertex with i ∈ Me
if M (current vertex) ∪ M (amin ) ⊇ Me then
pickup messages Me , drop all other messages, traverse e and delete it
else
let j ∈ Me \ (M (current vertex) ∪ M (amin ))
fetchMessage(j, currentVertex)
Both procedures computeTour and fetchMessage make sure that the agent traverses
every edge e ∈ E with label Me while carrying the exact set of messages Me and every edge
ē ∈ E while carrying no messages. So the first part of the claim is clear. We only need to
show that amin cannot get stuck at some vertex v ∗ before returning to v0 . As GS is Eulerian
(ignoring the labels) and edges are deleted once they are traversed, this can only happen if
some call fetchMessage(i, v) gives up at vertex v ∗ . This means that the current vertex v ∗
does not contain message i and has no edge ē ∈ E with a label containing message i.
Note that when amin is currently proceeding according the call fetchMessage(i, v),
then it will be on a vertex of the path that message i takes from its start si to its destination
ti in the optimal schedule Opt, and this path is simple by our initial assumption. Also
note that, since edge labels are obeyed, message i only ever moves forward along its path
in Opt. This means that if amin is stuck at vertex v ∗ , there must initially have been an
edge ē = (v ∗ , w) ∈ E incident to v ∗ with i ∈ Me that was taken by the agent earlier and
then deleted. The agent traverses edges in E only in the procedure fetchMessage. So
there must have been a call fetchMessage(j1 , v1 ) before, where the agent traversed the
edge ē. This call cannot have been completed as otherwise the original edge e = (w, v ∗ ) ∈ E
corresponding to ē would have been used by amin and the message i would have already
reached v ∗ , since i ∈ Me . This contradicts that the message path of i is simple and the agent
is currently proceeding according the call fetchMessage(i, v) at vertex v ∗ .
Bärtschi, Chalopin, Das, Disser, Graf, Hackfeld and Penna
As the call fetchMessage(j1 , v1 ) is not complete, there must be a vertex v2 and a
message j2 missing at this vertex to further carry j1 on its paths to the destination, and a
call fetchMessage(j2 , v2 ) which is also incomplete. By iterating this argument, we obtain
that the current stack of functions is fetchMessage(js , vs ), . . . , fetchMessage(j1 , v1 )
for some s ∈ N, where js = i and vs = v. In the optimal schedule Opt the message j2
needs to be transported to v2 before j1 can be further transported from v2 together with j2 .
Similarly, message jr needs to be transported to vr before message jr−1 can be transported
further together with message jr from vr for r = 2, . . . , s. Moreover, on the edge e = (w, v ∗ )
the messages j1 and js = i need to be transported together and in particular, message j1
needs to be transported together with js before js can be transported further. But in the
optimal schedule the messages must be transported in a certain sequence and it cannot be
that message i needs to be transported to v before messages j1 is transported to message v1
and vice versa. Thus. computeTour must terminate with amin returning to the starting
location v0 .
Claim 2: After completing a tour given by computeTour the following holds: Every
message i has either been transported to its destination or it is on a vertex vi such that there
is a path from vi to ti with edges in E containing i in their labels, and a path in the reverse
direction with edges in E containing i in their labels.
Every edge e ∈ E with label Me is only traversed if the agent amin carries the set of
messages Me . Thus at any time there is a path from the current location vi of message i to
its destination ti with edges containing i in the label. This shows the first part of the claim.
Observe that a completed call fetchMessage(i, v) yields a closed walk, as the agent
starts and ends in v. Moreover, it first traverses exactly all edges in E on the path from v
to the current position vi of message i and then all edges in E on the path from vi to v.
Inductively, this also holds for all levels of recursive calls of fetchMessage. Hence, for
every original edge e ∈ E also the corresponding reverse edge ē ∈ E is traversed in a call of
fetchMessage.
This fact also implies that for any edge E ∪˙ E traversed in the procedure computeTour
(and not in fetchMessage), the corresponding original/reverse edge cannot be traversed in
˙ are
a call of fetchMessage. Inductively, we can therefore argue that if e1 , e2 , . . . , es ∈ E ∪E
the edges traversed in the procedure computeTour in this order, such that the corresponding
original/reverse edges were not traversed, then es , . . . , e1 is a path from the current location
of amin to its starting vertex. This shows that at termination for every original edge e ∈ E
also the corresponding reverse edge ē ∈ E is traversed.
Claim 3: Combining the tours returned by multiple calls of computeTour yields a
Eulerian tour that satisfies the edge labels in every step.
Assume that the tour T resulting from a call of computeTour does not traverse all
edges of GS . Let v0 be the starting vertex of the tour, v be the last vertex on the tour that
is is incident to an edge which is not visited, and vi be the position of message i after the
tour T is finished. Further, let G0S be the graph GS after the call of computeTour, i.e.,
without all edges in T and message i at position vi instead of si . We want to show that we
can run computeTour on G0S with amin starting at v and then add the resulting tour T 0
to T as follows: First amin follows T until the last time it visits v, then it follows T 0 , and
finally the remaining part of T .
The graph G0S is a feasible input to computeTour by Claim 2. It corresponds to
the instance of WeightedDelivery, where all message transported in the schedule Opt,
Page:9
Page:10 Energy-efficient Delivery by Heterogeneous Mobile Agents
f (x)
wj ∗
b
x+1
x∗
1
x
b
Figure 4 Choosing the largest b such that x+1
is a lower bound on the step-function f representing
the weight of the agent currently transporting the message.
which are done in the tour T by amin , are completed, and the agent positions are adapted
accordingly. By Claim 1, computeTour will produce a tour that satisfies the edge labels.
The only problem that can occur when combining the tours T and T 0 therefore is that
during following the tour T 0 , fetchMessage(i, v) is called, but some message i is not yet
transported to vi because the tour T has not been completed. This means that vertex vi is
visited after the last time v is visited by the tour T . By the choice of v, all edges incident
to vi must be visited by the tour T , in particular, we must have that vi = ti and message i
is delivered to its destination by the tour T . But then G0S does not contain any edge with
label i by Claim 2.
By iterative applying the above argument, we obtain a Eulerian tour that satisfies the
edge labels in every step.
J
Single Message. For the case of a single message, we can improve the upper bound of 2 on
the benefit of collaboration from Theorem 6, to a tight bound of 1/ ln 2 ≈ 1.44.
I Theorem 7. There is a (1/ ln 2)-approximation algorithm using a single agent for WeightedDelivery with m = 1.
Proof. By using Dijkstra’s algorithm, we can determine the agent that can transport the
message from s to t with lowest cost. We need to show that this is at most 1/ ln(2) the cost
of an optimum using all agents.
Fix an optimum solution and let the agents a1 , a2 , . . . , ar be labeled in the order in which
they transport the message in this optimum solution (ignoring unused agents). We can
assume by Lemma 1 that w1 > w2 > . . . > wr . By scaling, we can further assume without
loss of generality that wr = 1 and that the total distance traveled by the message is 1.
Now, for each point x ∈ [0, 1] along the message path there is an agent aj with cost wj
carrying the message at this point in the optimum schedule and we can define a function
f with f (x) = wj . The function f is a step function that is monotonically decreasing by
Lemma 1 with f (0) = w1 and f (1) = wr = 1. We now choose the largest b ∈ [0, 1] such that
b
f (x) ≥ x+1
, see Figure 4.
b
Note that b ≥ 1 as f (x) ≥ 1 ≥ x+1
for b = 1 and all x ∈ [0, 1]. Further, let gj be the
Pr
distance traveled by agent aj without the message and g := j=1 gj wi the total cost for the
distances traveled by all agents without the message. We obtain the following lower bound
for an optimum solution
Z
1
Z
f (x) dx + g ≥
cost(Opt) =
0
0
1
b
dx + g = b ln(2) + g.
x+1
Bärtschi, Chalopin, Das, Disser, Graf, Hackfeld and Penna
Page:11
b
By the choice of b, the functions f (x) and x+1
coincide in at least one point in the interval
∗
[0, 1]. Let this point be x and aj ∗ be the agent carrying the message at this point. This
means that f (x∗ ) = x∗b+1 = wj ∗ . We will show that it costs at most (1/ ln(2))cost(Opt)
for agent aj ∗ to transport the message alone from s to t. The cost for agent aj ∗ to reach
s is bounded by gj ∗ wj ∗ + x∗ · wj ∗ and the cost for transporting the message from s to t is
bounded by wj ∗ . Thus, the cost of the algorithm using only one agent can be bounded by
cost(ALG) ≤ gj ∗ wj ∗ + x∗ · wj ∗ + wj ∗ = gj ∗ wj ∗ + (x∗ + 1) ·
By using that gj ∗ wj ∗ ≤ g, we finally obtain
cost(Alg)
cost(Opt)
≤
b+gj ∗ wj ∗
b ln(2)+g
b
= b + gj ∗ wj ∗ .
x∗ + 1
≤
b
b ln(2)
=
1
ln(2) .
J
No Intermediate Dropoffs. In the following we show that for κ ∈ {1, ∞} the upper bound
of 2 on the benefit of collaboration still holds, with the additional property that each message
is carried by its single agent without any intermediate dropoffs. We will make use of this
result later in the approximation algorithm for WeightedDelivery with κ = 1 (Section 5).
I Theorem 8. Let Opt be an optimal schedule for a given instance of WeightedDelivery
with κ ∈ {1, ∞}. Then there exists a schedule S such that (i) every message is only transported
by a single agent, with exactly one pick-up and one drop-off, (ii) cost(S) ≤ 2 · cost(Opt),
and (iii) every agent aj returns to its starting location pj .
Proof. For κ = ∞ this is a corollary of Theorem 6: an agent a with infinite capacity can just
as well keep a message mi once it was picked up, i.e. we can simply remove all the actions
for this message between the first pick-up (a, si , mi , +) and the last drop-off (a, ti , mi , −).
For κ = 1 we need a different analysis. Given an optimum schedule Opt, we look at how
the trajectories of the messages are connected by the agents. More precisely, we construct
an auxiliary multigraph G0 = (V 0 , E 0 ) as follows: The vertex set V 0 consists of all messages.
Then, for each agent a we look at its subsequence of the optimum schedule, Opt|a . Since a
has capacity 1, its subsequence consists of alternating drop-offs (a, p, mi , −) and pick-ups
(a, q, mj , +). For each drop-off followed by a pick-up action we add an edge (mi , mj ) of
length dG (p, q) to E 0 (Figure 5). These edges correspond to the portions of the optimum
schedule where the agent travels without carrying a message. We assume without loss of
a1
s1
amin
a2
t2
s2
s3
s4
a3
t3
t6
s6
s5
t5
a4
t4
amin
pmin
m1
t1
m2
s1
m6
m5
m3
m4
t1
t2
s2
s3
s4
t3
t6
s6
s5
t5
t4
Figure 5 (left) An optimal schedule, (center) the auxiliary graph, (right) the 2-approximation.
generality that G0 is connected and denote by amin the agent of minimum weight involved
in Opt (otherwise we can look at each connected component and its agent of minimum
weight separately). Assume that the first action of amin in Opt is to move from its starting
position pmin to a node p where it picks up message m1 (note that p can potentially lie
anywhere on the trajectory between s1 and t1 ). We model this by adding a node pmin to V 0
and connecting it to m1 by an edge of corresponding length dG (pmin , p). Now we can take a
minimum spanning tree of G0 and remove all redundant edges.
Note that the total length of the minimum spanning tree is a lower bound on the sum
of the distances traveled by all the agents in Opt without carrying a message. Thus any
Page:12 Energy-efficient Delivery by Heterogeneous Mobile Agents
schedule S of amin which move exactly twice along the trajectory of each message and twice
along the path corresponding to each edge of the minimum spanning tree of G0 has a cost of
at most cost(S) ≤ 2 · cost(Opt). The following tour satisfies this property and delivers
each message to its destination immediately after it is picked up:
We first let amin walk to p, from where amin proceeds towards s1 . When amin reaches s1 ,
it picks up m1 and delivers it to t1 along its trajectory in Opt. Once amin reaches t1 , we let it
return to p and from there back to its original position pmin . If, however, along its way from
p to s1 or from t1 to p the agent visits an endpoint of a path corresponding to an edge of the
minimum spanning tree, we first let amin serve the adjacent subtree recursively, see Figure 5.
It is easy to see that in the resulting schedule S every message is directly transported from
its source to its destination, and that the capacity is respected at all times.
J
3
Planning
We now look in isolation at the problem of ordering the messages within the schedule of an
agent, which we call Planning. Formally, the Planning aspect of WeightedDelivery is the
following task: Given a schedule S and one of its agents aj , reorder the actions in S|aj in
such a way that the schedule remains feasible and the costs are minimized.
Generally speaking, for a complex schedule with many message handovers, the reordering
options for a single agent aj might be very limited. First of all, we must respect the capacity
of aj , i.e., in every prefix of S|aj , the number of pick-up actions (aj , ∗, ∗, +) cannot exceed
the number of drop-off actions (aj , ∗, ∗, −) by more than κ. Even then, reordering S|aj might
render S infeasible because of conflicts with some other subschedule S|ax . But Planning also
includes the instances where a single agent delivers all the messages, one after the other
straight to the target, and where the only thing that has to be decided is the ordering. We
show now that in this setting, where there is no non-trivial coordination or collaboration
aspect, WeightedDelivery is already NP-hard.
I Theorem 9. Planning of WeightedDelivery problem is NP-hard for all capacities κ
even for a single agent on a planar graph.
Proof. We proceed by a reduction from Hamiltonian cycles on a grid graph, a problem shown
to be NP-hard by Itai et al. [21]. A similar reduction was used for a sorting problem by
Graf [17]. Given an unweighted grid graph H = (VH , EH ) with VH = {v1 , v2 , . . . , vn }. We
add to every vertex v a new vertex v 0 with an edge e = (v, v 0 ) and a message with start v and
target v 0 . Denote the new graph with G = (V, E), where V = VH ∪ VH0 , VH0 = {v 0 | v ∈ VH }
0
0
and E = EH ∪ EH
, where EH
= {(v, v 0 ) | v ∈ VH }.
We build an instance of WeightedDelivery by taking G and placing a single agent
on an arbitrary vertex p1 ∈ VH . We let w1 = 1 and set unit edge length le = 1 for all edges
0
p′1
1
0
x
1
1
1
H
G
1
p1
0
1
0
1
0
1
1
1
0
1
0
1
0
1
0
1
1
1
1
0
1
0
1
0
1
0
d1 = |V | + x
Figure 6 Finding a Hamiltonian cycle via WeightedDelivery with a single agent.
Bärtschi, Chalopin, Das, Disser, Graf, Hackfeld and Penna
in EH and edge length le = 0 for all edges in EH 0 except (p1 , p01 ), see Figure 6. The edge
(p1 , p01 ) gets length x = l(p1 ,p01 ) = |V | instead.
Now let the d1 be the length of the shortest path in G starting in p1 on which the agent
can deliver all messages. We now argue that there is a Hamiltonian cycle in H if and only if
d1 = 2|V |. To see that 2|V | is a lower bound for d1 , let us distinguish whether the agent ends
at p01 or not. If we end at p01 , we have to reach every v ∈ VH \ {p1 } at least once and also go
back to p1 before using (p1 , p01 ). This sums up to at least (|V | − 1) + 1 + |V | = 2|V |. If we
end somewhere else, we have to use (p1 , p01 ) twice, hence d1 ≥ 2|V |. So we get a schedule of
cost 2|V | if and only if we reach every vertex exactly once and end at p01 . When removing all
0
the EH
-steps, such a schedule directly corresponds to a Hamiltonian cycle as illustrated in
Figure 6.
J
Using similar ideas, we can use recent results for the approximation hardness of metric
TSP [22] to immediately show that Planning of WeightedDelivery can not be approximated arbitrarily well, unless P = NP.
I Theorem 10. It is NP-hard to approximate the Planning of WeightedDelivery to
within any constant approximation ratio less than 367/366.
Proof. We build on top of a result by Karpinski, Lampis and Schmied [22, Theorem 4] which
shows that the symmetric, metric traveling salesperson problem is hard to approximate with
ratio better than 123
122 . For a reduction, we take any metric, undirected graph H, duplicate
the vertices and put a zero-length edge and a single message between each of them, just
like in Theorem 9 / Figure 6. To find a suitable length x for the extra edge (p1 , p01 ) at the
(arbitrary) starting vertex we have to consider the following: In a traveling salesman tour
in H, we want the agent to come back to p1 in the end. Hence in WeightedDelivery
on G we want the agent to end at p01 . We achieve this by setting x large enough to avoid
traveling (p1 , p01 ) twice. Let M be the cost of a minimum spanning tree in H. Clearly, both
the optimum traveling salesman path and the optimum traveling salesman tour have cost
at least M but also at most 2M . Hence, setting the length x of the extra edge (p1 , p01 ) to
2M ensures that any schedule for WeightedDelivery on G which doesn’t end in p01 (and
thus uses (p1 , p01 ) twice) has cost at least 2 · 2M + M = 5M , while an optimum schedule has
delivery cost at most 2M + 2M = 4M . It remains to look at schedules which end in p01 : As
the extra edge contributes at most two thirds of the cost of the final schedule, at least one
1
third of the approximation gap is conserved, giving 1 + 3·122
= 367
J
366 .
3.1
Polynomial-time Approximation for Planning in Restricted Settings
Motivated by Theorem 8, we now look at the restricted setting of planning for a feasible
schedule S R of which we know that each message is completely transported by some agent
aj without intermediate drop-offs, i.e., for every message mi there must be an agent j with
S R |mi = (aj , si , mi , +), (aj , ti , mi , −). This allows us to give polynomial-time approximations
for planning with capacity κ ∈ {1, ∞}:
I Theorem 11. Let S R be a feasible schedule for a given instance of WeightedDelivery
with the restriction that ∀i ∃j : S R |mi = (aj , si , mi , +), (aj , ti , mi , −). Denote by Opt(S R )
a reordering of S R with optimal cost. There is a polynomial-time planning algorithm Alg
which gives a reordering Alg(S R ) such that cost(Alg(S R )) ≤ 2 · cost(Opt(S R )) if κ = 1
and cost(Alg(S R )) ≤ 3.5 · cost(Opt(S R )) if κ = ∞.
Page:13
Page:14 Energy-efficient Delivery by Heterogeneous Mobile Agents
Proof. By the given restriction, separate planning of each S R |aj independently maintains
feasibility of S R . We denote by mj1 , mj2 , . . . , mjx the messages appearing in SaRj . We
define a complete undirected auxiliary graph G0 = (V 0 , E 0 ) on the node set V 0 = {pj } ∪
{sj1 , sj2 , . . . , sjx } ∪ {tj1 , . . . , tjx } with edges (u, v) having weight dG (u, v).
For κ = 1, the schedule Opt(S R )|aj corresponds to a Hamiltonian path H in G0 of
minimum length, starting in pj , subject to the condition that for each message mji the visit
of its source sji is directly followed by a visit of its destination tji . We can lower bound
the length of H with the total length of a spanning tree T 0 = (V 0 , E(T )0 ) ⊆ G0 as follows:
Starting with an empty graph on V 0 we first add all edges (sji , tji ). Following the idea
of Kruskal [23], we add edges from {pj } × {sj1 , . . . , sjx } ∪ {tj1 , . . . , tjx } × {sj1 , . . . , sjx } in
increasing order of their lengths, disregarding any edges which would result in the creation of
a cycle. Now a DFS-traversal of T 0 starting from pj visits any edge (sji , tji ) in both directions.
Whenever we cross such an edge from sji to tji , we add (aj , sji , mji , +), (aj , tji , mji , −) as a
suffix to the current schedule Alg(S R )|aj . We get an overall cost of cost(Alg(S R )|aj ) ≤
P
P
2 · e∈E(T 0 ) le ≤ 2 · e∈H = 2 · cost(Opt(S R )|aj ).
For κ = ∞, the schedule Opt(S R )|aj corresponds to a Hamiltonian path H in G0 of
minimum length, starting in pj , subject to the condition that for each message mji its
source sji is visited before its destination tji . We approximate Opt(S R )|aj with a schedule
Alg(S R )|aj that first collects all messages mj1 , . . . , mjx before delivering all of them. We
start by computing a minimum spanning tree M ST 0 of the subgraph of G0 consisting
of all nodes {pj } ∪ {sj1 , . . . , sjx }. A DFS-traversal on M ST 0 starting from pj collects
P
all messages, returns back to pj and has cost 2 · e∈E(M ST 0 ) le ≤ 2 · cost(Opt(S R )|aj ).
Next we consider the subgraph of G0 consisting of all nodes {pj } ∪ {tj1 , . . . , tjx }. Using
Christofides’ heuristic for the metric TSP path version with fixed starting point pj and
arbitrary endpoint, due to Hoogeveen [9, 18], we can deliver all messages by additionally
J
paying at most 32 · cost(Opt(S R )|aj ). In total, this gives a 3.5-approximation.
I Remark. If we assume as an additional property that the agent returns to its starting
position pj (as for example in the result of Theorem 8), we can get a better approximation
for the case κ = 1. Instead of traversing a spanning tree twice, we can model this as the
stacker-crane problem for which a polynomial-time 1.8-approximation is known [15].
4
Coordination
In this section, we restrict ourselves to the Coordination aspect of WeightedDelivery.
We assume that collaboration and planning are taken care of. More precisely, we are given a sequence containing the complete fixed schedule S − of all actions (_, si , mi , +), . . . , (_, h, mi , −),
. . . , (_, tj , mj , −), but without an assignment of the agents to the actions. Coordination is
the task of assigning agents to the given actions. Even though coordination appears to have
the flavor of a matching problem, it turns out to be NP-hard to optimally match up agents
with the given actions. This holds for any capacity, in particular for κ = 1. The latter,
however, has a polynomial-time solution if all agents have uniform weight.
4.1
NP-Hardness for Planar Graphs
We give a reduction from planar 3SAT: From a given planar 3SAT formula F we construct
an instance of WeightedDelivery that allows a schedule S with “good” energy cost(S)
if and only if F is satisfiable.
Bärtschi, Chalopin, Das, Disser, Graf, Hackfeld and Penna
v1 ∨ v2
v2 ∨ v3 ∨ v4
v4
false
v1
v2
v3
v4
v1
true
H
v1 ∨ v2 ∨ v4
v2 ∨ v3 ∨ v4
Page:15
false
v2
true
true
v3
false
true
v4
false
G(F )
Figure 7 (left) A restricted plane embedding of a 3CNF F which is satisfied by (v1 , v2 , v3 , v4 ) =
(true, f alse, f alse, true). (right) Its transformation to the corresponding delivery graph.
Planar 3SAT. Let F be a three-conjunctive normal form (3CNF) with x boolean variables
V (F ) = {v1 , . . . , vx } and y clauses C(F ) = {c1 , . . . , cy }. Each clause is given by a subset
of at most three literals of the form l(vi ) ∈ {vi , vi }. We define a corresponding graph
H(F ) = (N, A) with a node set consisting of all clauses and all variables (N = V (F ) ∪ C(F )).
We add an edge between a clause c and a variable v, if v or v is contained in c. Furthermore
we add a cycle consisting of edges between all pairs
of consecutive variables, i.e., A = A1 ∪ A2 ,
where A1 = {{ci , vj } | vj ∈ ci or vj ∈ ci } , A2 = vj , v(j mod x)+1 | 1 ≤ j ≤ x . We call
F planar if there is a plane embedding of H(F ). The planar 3SAT problem of deciding
whether a given planar 3CNF F is satisfiable is known to be NP-complete. Furthermore the
problem remains NP-complete if at each variable node the plane embedding is required to
have all arcs representing positive literals on one side of the cycle A2 and all arcs representing
negative literals on the other side of A2 [26]. We will use this restricted version in our
reduction and assume without loss of generality that the graph H(F ) \ A2 is connected and
that H(F ) is a simple graph (i.e. each variable appears at most once in every clause).
Building the Delivery Graph. We first describe a way to transform any planar 3CNF graph
H(F ) into a planar delivery graph G = G(F ), see Figure 7. We transform the graph in five
steps: First we delete all edges of the cycle A2 , but we keep in mind that at each variable
node all positive literal edges lie on one side and all negative literal edges on the other
side. Secondly let degH(F ),A1 (v) denote the remaining degree of a variable node v in H and
surround each variable node by a variable box. A variable box contains two paths adjacent
to v on which internally we place degH(F ),A1 (v) copies of v: One path (called henceforth
the true-path) contains all nodes having an adjacent positive literal edge, the other path
(the false-path) contains all nodes having an adjacent negative literal edge. In a next step,
we add a single node between any pair of node copies of the previous step. As a fourth
step, we want all paths to contain the same number of nodes, hence we fill in nodes at the
end of each path such that every path contains exactly 2y ≥ 2 degH(F ),A1 (v) internal nodes.
Thus each variable box contains a variable node v, an adjacent true-path (with internal
nodes vtrue,1 , . . . , vtrue,2y−1 and a final node vtrue,2y ) and a respective false-path. Finally for
each clause node c we add a new node c0 which we connect to c. The new graph G(F ) has
polynomial size and all the steps can be implemented in such a way that G(F ) is planar.
Messages, Agents and Weights. We are going to place one clause message on each of the
y clause nodes and a literal message on each of the 2x paths in the variable boxes for a total
of 4xy messages. More precisely, on each original clause node c we place exactly one clause
message which has to be delivered to the newly created node c0 . Furthermore we place a
literal message on every internal node vtrue,i of a true-path and set its target to vtrue,i+1
(same for the false-path). We set the length of all edges connecting a source to its target to 1.
Next we describe the locations of the agents in each variable box. We place one variable
Page:16 Energy-efficient Delivery by Heterogeneous Mobile Agents
1
Clause c2 = (v2 ∨ v3 ∨ v4 )
1−2ε
1+2ε
Variable v2
false
ε
1
ε
true
1
1+yε
1
1
1+2ε
1−yε
1+yε
1−ε
1+ε
1
1+ε
1
true
1
1+yε
1
1
1+2ε
1
1
1+ε
1
ε
ε
false
1
1
1+yε
Variable v3 Variable v4
1
1+ε
1
true
1
1+yε
1
1+ε
1
ε
ε
false
1
1
1+yε
1+yε
1
1
1
Figure 8 Agent positions () and weights (in black); Messages (→) and edge lengths (in color).
agent of weight 1 on the variable node v. The length of the two adjacent edges are set to
ε, where ε := (8xy)−2 . Furthermore we place y literal agents on each path: The i-th agent
is placed on vtrue,2(y−i) (respectively vfalse,2(y−i) ) and gets weight 1 + iε. It remains to set
the length of edges between clause nodes and internal nodes of a path. By construction the
latter is the starting position of an agent of uniquely defined weight 1 + iε; we set the length
of the edge to 1−iε
1+iε . For an illustration see Figure 8, where each agent’s starting location is
depicted by a square and each message is depicted by a colored arrow.
Reduction. The key idea of the reduction is that for each variable u, the corresponding
variable box contains a variable agent who can either deliver all messages on the true-path
(thus setting the variable to true), or deliver all messages on the false-path (thus setting
the variable to false). Assume u is set to true. If u is contained in a clause c, then on the
adjacent node vtrue,i there is a (not yet used) literal agent. Intuitively, this agent was freed
by the variable agent and can thus be sent to deliver the clause-message. If u is contained
in c, the corresponding literal on the false-path can’t be sent to deliver the clause message,
since it needs to transport messages along the false-path.
I Lemma 12 (Energy cost of a 3SAT-solution). Given a satisfiable assignment (a solution)
for the variables of a 3CNF F there is a feasible schedule Sol of the agents in G(F ) which
has a total (energy) cost of cost(Sol) = 4xy + 2y + x(y 2 + y + 1)ε.
Proof. We are given x variables, y clauses and a satisfiable assignment of the variables. We
construct a schedule from the assignment as follows: Assume that variable v is set to true
and consider the corresponding variable box.
The variable agent (which has weight 1) delivers all messages on the full true-path (which
has length 2y + ε). The energy needed to do so is 2y + ε.
Each literal agent placed on a node vfalse,i transports two messages: the message with
source vfalse,i and the message with source vfalse,i+1 . Summing over all messages on the
false-path we need an energy of 2 · ((1 + yε) + . . . + (1 + ε)).
Hence for the messages in each of the x variable boxes we have an energy consumption of
Py
2y + ε + 2y + 2 i=1 iε = 4y + (y 2 + y + 1)ε. Furthermore, since we start from a satisfiable
assignment, the source of each clause message is connected to at least one not yet used agent
of weight 1 + iε. Such an agent is adjacent to the source of that clause message only and
can move to the source (distance 1−iε
1+iε ), pick it up and deliver it (distance 1), hence for each
clause we get an energy consumption of exactly (1 + iε)( 1−iε
1+iε + 1) = 2. Summedover all
variables and all clauses we get a total energy consumption of x · 4y + (y 2 + y + 1)ε + 2y =
4xy + 2y + x(y 2 + y + 1)ε.
J
Fixed sequence (schedule without agent assignment). We now fix a sequence S − that describes the schedule constructed in Lemma 12 but which does not allow us to infer a satisfiable
Bärtschi, Chalopin, Das, Disser, Graf, Hackfeld and Penna
assignment. In our sequence S − every pick-up action of one of the 4xy + y many messages
mi at its location si is immediately followed by its drop-off at its destination ti . Hence we
are restricted to a schedule (_, sπ(1) , mπ(1) , +), (_, tπ(1) , mπ(1) , −), (_, sπ(2) , mπ(2) , +), . . . ,
(_, tπ(4xy+y) , mπ(4xy+y) , −), where π can be any permutation on 1, . . . , 4xy + y that satisfies
the following property: If two messages mi , mj lie on the same true- or false-path originating
at a variable node v, then dG (F )(v, si ) < dG (F )(v, sj ) ⇒ π −1 (i) < π −1 (j) (meaning if mi
lies to the left of mj , it should precede mj in the schedule).
Energy Consumption of Optimal Schedules. In the next three Lemmata 13, 14 and 15 we
show that (i) the total energy consumption of any optimum schedule Opt is cost(Opt) ≥
cost(Sol), and (ii) in every optimum schedule with cost(Opt) = cost(Sol), each variable
agent delivers exactly all messages on either the true- or the false-path. We remark that (i)
is true independent of whether Opt adheres tho the fixed schedule without assignments S −
or not and holds for any capacity κ. In the case of (ii), Opt has exactly the properties as
described in the proof of Lemma 12. Since thus for each agent aj the subsequence Opt|aj
is uniquely determined, and since each message is transported by a single agent (without
intermediate drop-offs), these subsequences can be merged to match the prescribed fixed
order of the actions S − .
Lower Bound. To this end we first give an upper bound UB on cost(Sol) and a lower
bound LB on the total energy consumption of any feasible schedule S. First, since ε = (8xy)−2 ,
we define UB := 4xy + 2y + 0.25 > 4xy + 2y + x(y 2 + y + 1)ε = cost(Sol). For the lower
bound, note that every agent has weight at least 1. We double count the distance traveled by
the agents via the distance covered by the messages, hence we have to be careful to take into
account that an agent might carry two or more messages at the same time (we do not want
to count the energy used during that time twice or more). Hence for each of the 4xy + y
messages we count the last edge over which it is transported towards its target. All message
targets (and thus the distance traveled towards them) are disjoint and have adjacent edges
1
all of length at least 1−yε
1+yε > 1 − 2yε > 1 − 32xy . Additionally, before an agent can deliver a
clause message, it needs first to travel towards its source. Such edge crossings are not counted
yet, hence we can add an additional distance of 1−yε
1+yε for each clause. Overall we get a lower
1
bound on the total energy consumption of LB := 4xy + 2y − 0.25 < (4xy + y + y) · (1 − 32xy
).
I Lemma 13 (Independence of variable boxes). Any optimum schedule Opt, in which an
agent placed in a variable box u moves to a clause node and back into a variable box of some
(not necessarily different) variable v, has cost(Opt) > cost(Sol).
Proof. Assume for the sake of contradiction that an agent leaves its variable box, walks
to a clause node and later on back to a variable box. If this agent delivers the clause
message, it has to return and thus walks the corresponding clause message’s distance of 1
twice. If it doesn’t, then we haven’t included the travel towards the clause node yet. In
1
both cases, we add at least another 1 − 32xy
> 0.5 to the given lower bound LB, yielding
cost(Opt) > LB + 0.5 = UB > cost(Sol).
J
From now on we can restrict ourselves to look at feasible schedules where each each agent
either stays inside its variable box, or walks to deliver a clause message and stays at the
target of that clause message. Next we show that we can also assume that agents walk only
from left to right inside a true- or false-path:
Page:17
Page:18 Energy-efficient Delivery by Heterogeneous Mobile Agents
I Lemma 14 (Agents move from left to right). Any optimum schedule Opt, in which there
is an agent a that moves at some point in the schedule from right to left along a true- or
false-path, has cost(Opt) > cost(Sol).
Proof. By Lemma 13 we can restrict ourselves to schedules Opt where each message i has
to be transported over the edge connecting si and ti . Without loss of generality we assume
that no message i ever leaves the interval [si , ti ] and that it is transported from si to ti
monotonically from left to right (otherwise we could adjust the schedule accordingly by
keeping the trajectories of the agents but adapting pick-up and drop-off locations).
First assume for the sake of contradiction that in Opt there is an agent a whose trajectory
contains moves from right to left of total length at least 0.5. The energy needed to do this
is at least 0.5, and these moves are not yet included in the lower bound LB. As before, by
adding 0.5, we get cost(Opt) > LB + 0.5 = UB > cost(Sol). Hence in the following we
assume that each agent moves strictly less than 0.5 from right to left.
We are going to show that any such schedule Opt can be transformed into a schedule
of smaller cost, contradicting the optimality of Opt. Consider the longest consecutive
right-to-left move of agent a: Since a moves by less than 0.5 to the left, it must come from a
handover hi point lying inside an edge (si , ti ) or go to a handover point hj lying inside an
edge (sj , tj ).
In the first case, hi is closer to si than to ti and the previous action of a in the schedule
must have been (a, hi , i). The agent b picking up message i at hi must have its starting
position on or to the left of si . Hence we could replace a’s pick-up and drop-off of message i
with pick-up and drop-off by agent b, thus strictly decreasing the total distance da traveled
by a, contradicting the optimality of OP T . In the second case, hj is closer to tj than to
sj and agent a moves to the right after picking up message j. Let b denote the agent that
dropped off j at hj . By the previous remarks we know that b will not move to the left in
its next action (if any), b can’t reach sj and no other message is inside [si , ti ]. Furthermore
by the weights given in our hardness reduction we know wb < 2wa . Therefore we can move
hj by a small δ > 0 to the right, thus strictly decreasing wa · da + wb · db , contradicting the
optimality of Opt.
J
From now on we assume that agents do not move from right to left at all. Now we are
ready to prove the key relation between optimum schedules and Sol schedules:
I Lemma 15 (Energy cost of an optimum schedule). Any optimum schedule Opt either has
total energy consumption cost(Opt) > cost(Sol) or cost(Opt) = cost(Sol). In the
latter case, each variable agent either delivers exactly all messages on its true-path or exactly
all messages on its false-path. Furthermore the literal agents on the respective other path
deliver exactly two literal messages each. Finally, clause messages are only delivered by freed
up literal agents on the paths chosen by the variable agents.
Proof. By Lemmata 13 and 14 we may assume that no agent travels into another variable
box and that agents only move from left to right on any true- or false-path. Furthermore we
have seen in the proof of Lemma 14 that this implies that every literal message is carried from
its source to its target by a single agent in a continuous left-to-right motion. We now show
that if there was a variable agent a which does not deliver either all messages on its adjacent
true-path or its adjacent false-path, then we would get a contradiction to optimality:
Assume first for the sake of contradiction that a stays on its starting location. Then
we can move a to the first internal node of its true-path, which contributes an additional
ε-distance to the total energy consumption. Let b be the agent carrying the first literal
Bärtschi, Chalopin, Das, Disser, Graf, Hackfeld and Penna
message; we know that b must have weight 1 + yε. We replace b in the schedule by a,
saving at least yε energy already on the first literal message, contradicting the optimality of
the schedule. Now assume that a either deviated at some internal node vtrue,2(y−i) of the
path it entered (to deliver a clause message) or that it stopped at such an internal node
(vtrue,2(y−i) or vtrue,2(y−i)+1 ). Let b denote the agent carrying the message which was placed
on the specified node. The edge to the adjacent clause (if any) has length 1−iε
1+iε and b has
weight 1 + jε, with j ≥ i. Now we can switch a and b in the remainder of the schedule.
The potential increase of energy cost (on the clause message delivery) amounts to at most
((1 + jε) − 1) · ( 1−iε
1+iε + 1) < 2jε, while the gained energy on the next two literal messages is
at least jε · 2, again contradicting the optimality of the schedule.
Hence each variable agent delivers either all messages on its true-path or all messages
on its false-path. This allows us to give a new lower bound LB 0 on the energy consumption
cost(Opt): Each variable agent contributes an energy consumption of 2y + ε to the total.
Delivery of each message on the respective other path needs an agent with starting location
coinciding with or to the left of the message source, yielding an energy contribution of at
least 2 · ((1 + yε) + . . . + (1 + ε)) = 2y + y(y + 1)ε, with equality if and only if each literal
agent placed on vtrue,i delivers the message placed on vtrue,i and the consecutive message
with source on vtrue,i+1 . Finally the source of each clause message is reached by an agent of
weight 1 + jε over an edge of length 1−iε
1+iε , j ≥ i, hence the delivery of each clause message
needs an energy of at least (1 + jε)( 1−iε
1+iε + 1) ≥ 2, with equality if and only if j = i. Summing
over the clauses and variable boxes we get LB 0 = 4xy + 2y + x(y 2 + y + 1)ε = cost(Sol). J
It remains to note that the first literal agent (of weight 1 + yε) on the path chosen by the
variable agent can walk over to the other path and from there on to deliver a clause message.
However such a schedule has higher energy consumption than cost(Sol): the literal agent
has to cross two edges of length ε and the required energy to do so has not been used in the
estimated lower bound LB 0 = cost(Sol). Hence we conclude:
I Theorem 16. Coordination of WeightedDelivery is NP-hard on planar graphs for all
capacities κ, even if we are given prescribed collaboration and planning.
4.2
Polynomial-time Algorithm for Uniform Weights and Unit Capacity
Note that Coordination is NP-hard even for capacity κ = 1. Next we show that this setting
is approachable once we restrict ourselves to uniform weights.
I Theorem 17. Given collaboration and planning in the form of a complete schedule with
missing agent assignment, Coordination of WeightedDelivery with capacity κ = 1 and
agents having uniform weights can be solved in polynomial time.
Proof. As before, denote by S − = (_, si , mi , +), . . . , (_, h, mi , −), . . . , (_, tj , mj , −) the
prescribed schedule without agent assignments. Since all agents have the same uniform
Pk
weight w, the cost cost(S) of any feasible schedule S is determined by cost(S) = w· j=1 dj .
Hence at a pick-up action (_, q, mi , +) it is not so much important which agent picks up the
message as where / how far it comes from.
Because we have capacity κ = 1, we know that the agent has to come from either its starting position or from a preceding drop-off action (_, p, mj , −) ∈ S − . This allows us to model
the problem as a weighted bipartite matching, see Figure 9 (center). We build an auxiliary
graph G0 = (A ∪ B, E10 ∪ E20 ). A maximum matching in this bipartite graph will tell us for
every pick-up action in B, where the agent that performs the pick-up action comes from in A.
Let A := {p1 , . . . , pk }∪{(_, ∗, ∗, −)} and B := {(_, ∗, ∗, +)}. We add edges between all agent
Page:19
Page:20 Energy-efficient Delivery by Heterogeneous Mobile Agents
s3
2
1
4
1
p1
s1
G
p1
p2
t1
t3
1
2
3
2
p2 t2
s2
(
(
(
(
, s3 , m1 , −)
, t2 , m2 , −)
, t3 , m3 , −)
, t1 , m1 , −)
2
1
0
0
0
2
35
21
21
G′
( , s1 , m1 , +)
( , s2 , m2 , +)
( , s3 , m3 , +)
( , s3 , m1 , +)
t3
s3
t1
p1
s1
s2
p2 t2
Figure 9 Illustration of the coordination of the following schedule S = (_, s1 , m1 , +),
(_, s2 , m2 , +), (_, s3 , m1 , −), (_, s3 , m3 , +), (_, t2 , m2 , −), (_, t3 , m3 , −), (_, s3 , m1 , +), (_, t1 , m1 , −)
(left) Instance with 3 messages and 2 agents of uniform weight. (center) Equivalent weighted
bipartite matching problem G0 . (right) The resulting trajectories of the agents.
starting positions and all pick-ups, E10 := {p1 , . . . , pk } × {(_, q, m, +) | (_, q, m, +) ∈ B} of
weight dG (pi , q). Furthermore we also add edges between drop-offs and all subsequent pick-ups
E20 := {((_, p, mj , −), (_, q, mi , +)) | (_, p, mj , −) < (_, q, mi , +) in S − } of weight dG (p, q).
A maximum matching of minimum cost in G0 captures the optimal assignment of agents
to messages and can be found by solving the classic assignment problem, a special case
of the minimum cost maximum flow problem. Both of these problems can be solved in
polynomial time for instance using the Hungarian method [24] or the successive shortest path
algorithm [13], respectively. The cost of this optimum matching corresponds to the cost of the
agents moving around without messages. The cost of the agents while carrying the messages
can easily be added: Consider the schedule S − restricted to a message mi . This subsequence
S − |mi is a sequence of pairs of pick-up/drop-off actions ((_, q, mi , +), (_, p, mi , −)), and in
P
every pair the message is brought from q to p on the shortest path, so we add
dG (q, p).
Concatenating these piecewise shortest paths gives the trajectory of each agent in the
optimum solution, as illustrated in Figure 9 (right).
J
Our algorithm is remotely inspired by a simpler problem at the ACM ICPC world finals
2015 [19]. The official solution is pseudo-polynomial [20], Austrin and Wojtaszczyk [3] later
sketched a min-cost bipartite matching solution.
5
Approximation Algorithm
By putting the results of the previous sections together, we obtain the following approximation
algorithm.
I Theorem 18. There is a polynomial-time (4 max
WeightedDelivery with capacity κ = 1.
wi
wj )-approximation
algorithm for
Proof. We start by artificially enlarging the weight wj of every agent aj to max wi . In doing
wi
so, we increase the energy cost contribution of each agent aj by a factor of w
. Thus any
j
wi
ρ-approximation to this weight-enhanced problem will give us a ρ · max wj approximation
for the original problem.
From now on assume without loss of generality that all agents have a uniform weight
w := max wi . Let Opt be an optimal schedule for an instance of WeightedDelivery with
uniform agent weights w and capacities κ = 1. We call a feasible schedule S restricted if in S
every message mi is transported by a single agent from si to ti without any intermediate dropoffs. By Theorem 8 there exists a restricted schedule S R with cost(S R ) ≤ 2 · cost(Opt).
Let OptR be any optimal restricted schedule with total energy consumption cost(OptR ) ≤
cost(S R ) ≤ 2 · cost(Opt).
Bärtschi, Chalopin, Das, Disser, Graf, Hackfeld and Penna
We define a complete undirected auxiliary graph G0 = (V 0 , E 0 ) on all agent starting
positions as well as all message sources and destinations, V 0 = {p1 , . . . , pk } ∪ {s1 , . . . , sm } ∪
{t1 , . . . , tm }. Each edge e = (u, v) ∈ E 0 has length le := dG (u, v). OptR has a natural
correspondence to a vertex-disjoint path cover P C(OptR ) of G0 with exactly k simple paths
P1 , . . . , Pk that has the following properties:
Each path Pj contains exactly one agent starting position, namely pj ;
and pj is an endpoint of Pj .
Each destination node ti is adjacent to its source node si ;
and si lies between ti and the endpoint pj .
Each path (possibly of length 0) with endpoint at a starting position pj corresponds to
P
the (possibly empty) schedule OptR |aj and cost(OptR ) = e∈P C(OptR ) le . Now let T C ∗
P
denote a tree cover of minimum total length e∈T C ∗ le among all vertex-disjoint tree covers
T C of G0 with exactly k trees T1 , . . . , Tk that satisfy the following properties:
Each tree Tj contains exactly one agent starting position, namely pj .
Each destination node ti is adjacent to its source node si .
Since P C(OptR ) itself is a tree cover satisfying the two mentioned properties, we
P
P
R
immediately get e∈T C ∗ le ≤
e∈P C(OptR ) le = cost(Opt ). By Theorem 11 we can
use DFS-traversals in each component of the optimum tree cover T C ∗ to construct in
P
polynomial-time a schedule S ∗ of total energy consumption cost(S ∗ ) ≤ 2 · e∈T C ∗ le ≤
2 · cost(OptR ) ≤ 4 · cost(Opt).
It remains to show that the tree cover T C ∗ can be found in polynomial time: Analogously to Theorem 11 we start with an empty graph on V 0 to which we add all edges
{(si , ti ) | i ∈ {1, . . . , m}}. Then we add all other edges of E 0 in increasing order of their
lengths, disregarding any edges which would result either in the creation of a cycle or in a
join of two starting positions pi , pj into the same tree.
J
Page:21
Page:22 Energy-efficient Delivery by Heterogeneous Mobile Agents
References
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
Julian Anaya, Jérémie Chalopin, Jurek Czyzowicz, Arnaud Labourel, Andrzej Pelc, and
Yann Vaxès. Convergecast and broadcast by power-aware mobile agents. Algorithmica,
74(1):117–155, 2016. doi:10.1007/s00453-014-9939-8.
David L. Applegate, Robert E. Bixby, Vasek Chvatal, and William J. Cook. The Traveling
Salesman Problem: A Computational Study (Princeton Series in Applied Mathematics).
Princeton University Press, Princeton, NJ, USA, 2007.
Per Austrin and Jakub Onufry Wojtaszczyk. ACM ICPC World Finals 2015 solution
sketches. http://www.csc.kth.se/~austrin/icpc/finals2015solutions.pdf, 2015.
A. Bärtschi, J. Chalopin, S. Das, Y. Disser, B. Geissmann, D. Graf, A. Labourel, and
M. Mihalák. Collaborative Delivery with Energy-Constrained Mobile Robots. In 23rd
International Colloquium on Structural Information and Communication Complexity SIROCCO’16, 2016.
A. Bärtschi, J. Chalopin, S. Das, Y. Disser, B. Geissmann, D. Graf, A. Labourel, and
M. Mihalák. Collaborative Delivery with Energy-Constrained Mobile Robots, arXiv preprint, 2016. arXiv:1608.08500.
Davide Bilò, Yann Disser, Luciano Gualà, Matús Mihalák, Guido Proietti, and Peter Widmayer. Polygon-constrained motion planning problems. In 9th International Symposium
on Algorithms and Experiments for Sensor Systems, Wireless Networks and Distributed
Robotics ALGOSENSORS’13, pages 67–82, 2013. doi:10.1007/978-3-642-45346-5_6.
Jérémie Chalopin, Shantanu Das, Matús Mihalák, Paolo Penna, and Peter Widmayer.
Data delivery by energy-constrained mobile agents. In 9th International Symposium on
Algorithms and Experiments for Sensor Systems, Wireless Networks and Distributed Robotics ALGOSENSORS’13, pages 111–122, 2013. doi:10.1007/978-3-642-45346-5_9.
Jérémie Chalopin, Riko Jacob, Matús Mihalák, and Peter Widmayer. Data delivery by
energy-constrained mobile agents on a line. In 41st International Colloquium on Automata, Languages, and Programming ICALP’14, pages 423–434, 2014. doi:10.1007/
978-3-662-43951-7_36.
Nicos Christofides. Worst-case analysis of a new heuristic for the travelling salesman problem. Technical Report 388, Graduate School of Industrial Administration, Carnegie-Mellon
University, Pittsburgh, PA, February 1976.
Jurek Czyzowicz, Krzysztof Diks, Jean Moussi, and Wojciech Rytter. Communication problems for mobile agents exchanging energy. In 23rd International Colloquium on Structural
Information and Communication Complexity SIROCCO’16, 2016.
Erik D. Demaine, Mohammadtaghi Hajiaghayi, Hamid Mahini, Amin S. Sayedi-Roshkhar,
Shayan Oveisgharan, and Morteza Zadimoghaddam. Minimizing movement. ACM Transactions on Algorithms (TALG), 5(3):1–30, 2009. doi:10.1145/1541885.1541891.
Jack Edmonds and Ellis L. Johnson. Matching, euler tours and the chinese postman.
Mathematical Programming, 5(1):88–124, 1973. doi:10.1007/BF01580113.
Jack Edmonds and Richard M Karp. Theoretical improvements in algorithmic efficiency
for network flow problems. Journal of the ACM (JACM), 19(2):248–264, 1972.
Robert W Floyd. Algorithm 97: shortest path. Communications of the ACM, 5(6):345,
1962.
P. Fraigniaud, L. Gasieniec, D. Kowalski, and A. Pelc. Collective tree exploration. In 6th
Latin American Theoretical Informatics Symposium LATIN’04, pages 141–151, 2004.
Greg N. Frederickson, Matthew S. Hecht, and Chul E. Kim. Approximation algorithms for
some routing problems. In Proceedings of the 17th Annual Symposium on Foundations of
Computer Science (FOCS), 1976.
Daniel Graf. How to sort by walking on a tree. In 23rd European Symposium on Algorithms
ESA’15, pages 643–655, 2015. doi:10.1007/978-3-662-48350-3_54.
Bärtschi, Chalopin, Das, Disser, Graf, Hackfeld and Penna
18
19
20
21
22
23
24
25
26
27
28
J. A. Hoogeveen. Analysis of christofides’ heuristic: Some paths are more difficult
than cycles. Operations Research Letters, 10(5):291–295, July 1991. doi:10.1016/
0167-6377(91)90016-I.
ACM ICPC. World Finals 2015 Problems, Task C: Catering. https://icpc.baylor.edu/
worldfinals/problems, May 2015.
ICPCNews. ACM ICPC 2015 problem catering. https://www.youtube.com/watch?v=
WxCgcI18_d8, May 2015.
Alon Itai, Christos H Papadimitriou, and Jayme Luiz Szwarcfiter. Hamilton paths in grid
graphs. SIAM Journal on Computing, 11(4):676–686, 1982.
Marek Karpinski, Michael Lampis, and Richard Schmied. New inapproximability bounds
for tsp. Journal of Computer and System Sciences, 81(8):1665–1677, 2015.
Joseph B. Kruskal, Jr. On the shortest spanning subtree of a graph and the traveling
salesman problem. Proceedings of the American Mathematical Society, 7:48–50, 1956.
Harold W Kuhn. The hungarian method for the assignment problem. Naval research
logistics quarterly, 2(1-2):83–97, 1955.
Chung-Lun Li, S. Thomas McCormick, and David Simchi-Levi. The point-to-point delivery and connection problems: Complexity and algorithms. Discrete Applied Mathematics,
36(3):267–292, 1992.
David Lichtenstein. Planar formulae and their uses. SIAM Journal on Computing,
11(2):329–343, 1982. doi:10.1137/0211025.
Stephen Warshall. A theorem on boolean matrices. Journal of the ACM (JACM), 9(1):11–
12, 1962.
Pawel Winter. Steiner problem in networks: A survey. Networks, 17(2):129–167, 1987.
doi:10.1002/net.3230170203.
Page:23
| 8cs.DS
|
Fundamental groups of a strongly controllable
group system or group code (including group shift
and linear block code)
arXiv:1709.08265v4 [cs.IT] 25 Feb 2018
Kenneth M. Mackenthun Jr. (email: [email protected])
February 27, 2018
ABSTRACT
A finite group G defines a signature group if and only if there is a normal
chain of G and a complete set of coset representatives of the normal chain
that can be arranged in a triangular form, where representatives in the same
row can be put in 1-1 correspondence so that certain sets of subtriangles of
representatives form isomorphic groups, when subtriangle multiplication is
induced by multiplication in G. We show that any strongly controllable
time invariant group system (or group shift or group code) can be associated with a signature group. Conversely given a signature group, we can
construct a strongly controllable time invariant group system. These results
show that to some extent the study of strongly controllable time invariant
group systems can be reduced to the study of signature groups. For a general strongly controllable group system which may vary in time, we show
there is a signature sequence of groups with similar properties as a signature
group. The general strongly controllable group system can be restricted to a
finite interval, in which case it becomes a block group system. We describe
the additive structure of the generator vectors which comprise the block
group system. This new result applies to any mathematical structure defined on a finite number of coordinates which is closed under group addition
in each coordinate, such as any linear block code over a group, field, vector
space, ring, or module. Finally we give algorithms to construct any strongly
controllable group system or its time invariant version or block version.
1
1. INTRODUCTION
The idea of group shifts and group codes is important in several areas
of mathematics and engineering such as symbolic dynamics, linear systems
theory, and coding theory. Research in this area started with the work of
Kitchens [1], Willems [2], Forney and Trott [3], and Loeliger and Mittelholzer
[4].
Kitchens [1] introduced the idea of a group shift [14] and showed that a
group shift has finite memory, i.e., it is a shift of finite type [14]. Using the
work of Willems [2] on linear systems, Forney and Trott [3] describe the state
group and state code of a set of sequences with a group property, which they
term a group code C. A time invariant group code is essentially a group shift.
They show that any group code that is complete (any global constraints can
be determined locally, see [3]) can be wholely specified by a sequence of
connected labeled group trellis sections (which may vary in time) which
form a group trellis C. They explain the important idea of “shortest length
code sequences” or generators. A generator is a code sequence which is not
a combination of shorter sequences. In a strongly controllable group code,
the nontrivial portions of all generators have a bounded length. They give
an encoder whose inputs are generators and whose outputs are codewords
in the group code. At each time t, a finite set of generators is used to give
a symbol in the codeword.
Loeliger and Mittelholzer [4] obtain an analog of the derivation of Forney
and Trott starting with a group trellis C instead of the group of sequences
C. To derive their encoder, they use an intersection of paths which split and
merge to the identity path in the trellis, an analog for the group trellis C of
the quotient group of code sequences (granule) in C used in [3].
In addition, Loeliger and Mittelholzer characterize the properties of any
group that is the branch group of a strongly controllable time invariant
group code. The properties are given in terms of an input/output condition
and a set of normal subgroups of the branch group that are related to paths
which split and merge to the identity path of the trellis. In this paper,
we find another group, a signature group, which characterizes any strongly
controllable time invariant group code. We characterize the signature group
in terms of properties of a complete set of coset representatives of a normal
chain of the branch group. The signature group is the component at time
t of a group of tensors. We give a similar characterization of any general
strongly controllable group system, which may vary in time.
The Forney-Trott encoder and Loeliger-Mittelholzer analog are very
much a global approach to finding an encoder of C and C, respectively.
The group code C has a normal chain C0 ⊂ C1 ⊂ · · · ⊂ Ck ⊂ · · · ⊂ Cℓ = C,
where each Ck is a k-controllable subcode of C [3]. The generators are coset
representatives of the coset decomposition chain of the normal chain. The
encoder steps up through the k-controllable subcodes of C, where at each
step k, there is a product of generator representatives over all time. In [19],
we find an encoder which is based on a local per unit time approach. At
each time t, the encoder output is a product of generator representatives in
a coset decomposition chain at time t. Nevertheless, both enocders can use
the same generators [19]. And both encoders have a shift register structure.
The global encoders of Forney and Trott and Loeliger and Mittelholzer
do not have the form of a time convolution. The local encoder in [19] has
2
the form of a time convolution. For this reason, we call the encoder in [19] a
time domain encoder, and the Forney and Trott encoder a spectral domain
encoder. In [19], we compare the outputs of the time and spectral domain
encoders when the generator inputs are the same, when time runs forward
and backward. We show that in an abelian group system, all four cases
give the same output for the same input, but in a nonabelian group system,
the outputs are related by a symmetry of a tensor set U . We also compare
outputs of the four encoders for different bases of the group system. These
results show there is a time and harmonic theory of group systems.
This paper can be regarded as giving dual results to those in [3] and
[4]. For example, the work in [3, 4] gives results in the spectral domain;
this work gives results in the time domain. The work in [3, 4] uses kcontrollable subcodes. In this paper, we use quotient group systems whose
normal subgroups are k-controllable group systems.
We consider both time invariant group systems and time varying group
systems. Time invariant group systems are the same as group shifts [14];
therefore the results here also apply to group shifts. A general time varying
group system can be defined over a finite interval, in which case it is a
block code over a group [3]. Therefore the results here can be applied to
block codes, and this gives new information about their group structure and
potentially new search algorithms and decoding algorithms.
Forney and Trott suggest the term group system as an alternative to
group code. In this paper and [19], we generally use the term group system
rather than group code because some results have analogues in classical
systems theory and harmonic analysis.
Forney and Trott have shown that any group system C can be reduced
to a group trellis C whose vertices are the states of the group system. The
states are defined using a group theoretic construction as quotient groups.
Each component of the trellis is a trellis section, a collection of branches
which forms a branch group B t at time t. We call group trellis C the first
canonic form of the group system C. We only consider strongly controllable
group systems C, in which there is a fixed integer ℓ such that for any time
t, for any sequence on (−∞, t] there exists a valid path of length ℓ to any
sequence on [t + ℓ, ∞), or alternatively, each state can be reached from any
other state in ℓ branches [3]. This material is reviewed in Section 2.
The product of any two paths ḃ, b̈ ∈ C is another path ḃb̈ in C. Each
path b is a sequence of branches bt in branch group B t for t ∈ Z. At
each time t, the product ḃb̈ of any two paths is a product ḃt b̈t of branches
ḃt , b̈t ∈ B t in paths ḃ, b̈ ∈ C, respectively.
In group trellis C, the sequence of branches that split from the identity
path and merge to the identity path form two normal chains [4]. The Schreier
refinement theorem can be applied to these two normal chains to obtain
another normal chain, a refinement of the two chains that we call a Schreier
series [19]. The Schreier series is a normal chain of the branch group at time
t, B t , of the group trellis. The Schreier series can be written in the form
of a matrix, with rows and columns determined by branches of the splitting
and merging trellis paths. When the group trellis is strongly controllable,
the matrix reduces to a triangular form, called the static matrix. The static
matrix is an echo of matrix ideas used in classical linear systems analysis.
By taking column j of static matrices at times t + j, for j = 0, . . . , ℓ + 1,
3
we can form another matrix, called a shift matrix. The shift matrix is defined
over the time interval [t, t + ℓ + 1], and is also a triangular form. We show
the shift matrix has a natural shift property, and in fact the shift matrix is
a part of the group trellis, the truncation of the ray of paths splitting from
the identity path at time t. This is discussed in Section 3.
We show that the rows of the shift matrix can be used to form quotient
groups, and the nontrivial generator vectors of the generator sequences of
Forney and Trott are a transversal of the quotient groups. The coset representatives of the generator vectors in the transversal are also a triangular
form, a shift matrix which we call a generator matrix. At time t, the components of the generator vectors form a complete set of coset representatives
for the Schreier series decomposition of branch group B t .
The static matrices and generator matrices can be arranged to form a
tensor r. Each tensor r is a sequence of generator matrices. The set of
tensors r is a set R. At each time t, a component rt of the tensor r is a
matrix, the static matrix. Therefore rt is a triangular form, also denoted
▽0,0 (rt ). Each static matrix is formed by ℓ + 1 generator matrices at times
t−j, for j = 0, 1, . . . , ℓ. The representatives in each rt are the representatives
at time t in a set of generator vectors; they form a coset representative chain
of some branch bt ∈ B t . This is discussed in Section 4.
A selection of a set B t of generator vectors at each time t is necessary
and sufficient to generate R. The sequence of component bases B t forms
a basis B. Each basis B gives a unique tensor set R and vice versa. If
component basis B t is the same for all t ∈ Z, then B is a constant basis. If
C is time invariant, then a constant basis can always be found.
In [19], we have shown there is a time domain encoder of C such that each
path r ∈ R is encoded into a path b ∈ C. This gives a 1-1 correspondence
R ↔ C. This material is included in Section 5.
If ḃ, b̈ are two paths in C, then their product ḃb̈ is another path in C.
Let ṙ ↔ ḃ and r̈ ↔ b̈ under the 1-1 correspondence R ↔ C. We define
an operation ∗ on R by ṙ ∗ r̈ = r̄ if r̄ ↔ ḃb̈ under the 1-1 correspondence
R ↔ C. The set of tensors R with operation ∗ forms a group (R, ∗). We
regard (R, ∗) as a second canonic form of group system C. The second
canonic form is obtained from the first canonic form C. We believe (R, ∗)
can be more revealing of the structure of a group system than group trellis
C.
The components rt of tensors r ∈ R at time t form a set Rt . Then we
can think of Rt as a set of triangular forms, denoted ▽0,0 (Rt ). The set
of representatives in Rt forms a complete set of coset representatives for a
normal chain decomposition of branch group B t . The time domain encoder
encodes rt into some branch bt in B t , and gives a 1-1 correspondence between
Rt and B t , Rt ↔ B t . The operation ∗ in (R, ∗) induces an operation ∗t on
Rt , such that if ṙt ↔ ḃt and r̈t ↔ b̈t under the 1-1 correspondence Rt ↔ B t ,
then ṙt ∗t r̈t = r̄t if r̄t ↔ ḃt b̈t . The set of components Rt with operation ∗t
forms a group (Rt , ∗t ), called an induced group of B t . We believe (Rt , ∗t )
plays a central role in understanding group systems. The groups (R, ∗) and
(Rt , ∗t ) are discussed in Section 6.1.
t as entries, for
Each triangular form rt = ▽0,0 (rt ) has representatives rj,k
rows k such that 0 ≤ k ≤ ℓ, for columns j such that 0 ≤ j ≤ k. We have
seen that B t induces a group (Rt , ∗t ) on the set of triangular forms ▽0,0 (Rt ).
4
The upper coordinates of the set ▽0,0 (Rt ) are a set of subtriangles ▽k,k (Rt )
of ▽0,0 (Rt ). In any finite group G, the upper coordinates of a complete
set of coset representatives form a group induced by group multiplication
in G. This means the set of subtriangles of upper coordinates ▽k,k (Rt )
forms a group (▽k,k (Rt ), ⊛tk,k ) under multiplication induced by B t , for k
such that 0 ≤ k ≤ ℓ, where (▽0,0 (Rt ), ⊛t0,0 ) = (Rt , ∗t ). Shifts of these
subtriangles, ▽j,k (Rt−j ), for 0 ≤ j ≤ k, contain representatives from the
same set of generator vectors as representatives of ▽k,k (Rt ). This means
that a representative in ▽j,k (Rt ) can be put in 1-1 corrrespondence with a
representative in ▽k,k (Rt ) if they belong to time shifts of the same generator
vector. Under this 1-1 correspondence, the set of subtriangles ▽j,k (Rt )
forms a group (▽j,k (Rt ), ⊛tj,k ) isomorphic to (▽k,k (Rt ), ⊛tk,k ), for 0 ≤ j ≤
k. These results give necessary and sufficient conditions for any induced
group (Rt , ∗t ) to be a signature group. We say group (Rt , ∗t ) is a signature
group if for each k such that 0 ≤ k ≤ ℓ, the group (Rt , ∗t ) restricted to
▽j,k (Rt ) forms a group (▽j,k (Rt ), ⊛tj,k ), and these groups are isomorphic
for 0 ≤ j ≤ k under a 1-1 congruence of ▽j,k (Rt ) and ▽k,k (Rt ) similar to
that just discussed.
The branch group B t of any strongly controllable time invariant group
system C has an induced group (Rt , ∗t ) that is a signature group. Conversely
any finite group G that has an induced group (Rt , ∗t ) that is a signature
group can be used to construct a strongly controllable time invariant group
system having branch group G. Therefore, in some sense, the study of
strongly controllable time invariant group systems is the study of signature
groups. The signature group characterizes a group system in terms of its
generator vectors. If basis B is chosen to be constant, the signature group
contains a description of all the generator vectors of the group system. These
results are discussed in Section 6.3.
The time invariant group system is a special case of a general group
system which may vary in time. In Section 6.2, we give similar results for
the general case. We show that any general ℓ-controllable group system
can be characterized by a signature sequence of groups that are slightly
more general versions of the signature group used in an ℓ-controllable time
invariant group system.
A special case of the general time varying ℓ-controllable group system is
a group system which is only nontrivial on a finite time interval [t, t + ℓ]. We
call this a block group system. A block group system includes a linear block
code over a group. Forney and Trott show that any block group system can
be decomposed into a set of generator vectors which can be arranged in a
group trellis [3]. In this paper, we characterize the additive structure of the
group of generator vectors which comprise the block group system. A linear
block code over a group is embedded in any linear block code over a field,
vector space, ring, or module, and so the result on additive structure applies
to these cases as well. More generally, the result applies to any mathematical
structure defined on a finite number of coordinates which is closed under
group addition in each coordinate. We use a signature sequence which is
isomorphic to the signature sequence defined using generator vectors. We
use this sequence to define a group which describes the additive structure
of the generator vectors. An example is given for ℓ = 2. This is done in
Section 6.4 and 6.5.
5
Forney and Trott [3] show that a group code does not have a homomorphism from C into the input sequences of C in general. However in Section
6.1 we show there is a homomorphism from C to permutation groups on
sets of generator vectors or generators selected by representatives in the
subtriangles used to define the signature group and signature sequence. In
Section 6.5 we give an example of this homomorphism for the (8, 4, 4) extended Hamming code or first order Reed Muller code, where ℓ = 3.
In Section 6.6, we use the isomorphic version of the signature sequence
to give an algorithm to construct an ℓ-controllable group system, an ℓcontrollable time invariant group system, and a block group system. The
algorithms are essentially a construction of the generator vectors of the
group system. This appears to be the first construction algorithm given for
a general group system and possibly a general block group system.
In Section 6.7, we show that any group system can be reduced to a
trivial group system by a sequence of group systems, which are the dual
of the derived sequence of shifts used in symbolic dynamics [10] and the
sequence of derivative codes [3].
6
2. GROUP SYSTEMS
This section gives a very brief review of some fundamental concepts in
[3], and introduces some definitions used here. We follow the notation of
Forney and Trott as closely as possible. One significant difference is that
subscript k in [3] denotes time; we use t (an integer) in place of k. In any
notation, a superscript is used exclusively to indicate time; thus t always
appears as a superscript in any notation. We use notation 1 in a generic
sense as the identity of any group; the particular group should be clear by
context.
Forney and Trott study a collection of sequences with time axis defined
on the set of integers Z, whose components at are taken from an alphabet
group or alphabet At at each time t, t ∈ Z. The set of sequences is a group
under componentwise addition in At . We call this a group system or group
code C [3]. A sequence a in C is given by
a = . . . , at−1 , at , at+1 , . . . ,
(1)
where at ∈ At is the component at time t.
The group system C is assumed to be complete [2, 3]; an important
consequence is that local behavior is sufficient to describe global behavior.
Completeness is the same as closure in symbolic dynamics [14]. Therefore a
time invariant complete group system C is the same thing as a group shift
in symbolic dynamics. In this paper, we use the language associated with
group systems [3] rather than group shifts [14].
+
Define Ct to be the set of all codewords in C for which an = 1n for n < t,
−
where 1n is the identity of An . Define Ct to be the set of all codewords
in C for which an = 1n for n ≥ t. The group system satisfies the axiom of
state: whenever two sequences pass through the same state at a given time,
the concatenation of the past of either with the future of the other is a valid
sequence [3]. The canonic state space Σt at time t is defined to be
C
def
Σt =
Ct− Ct+
.
The canonic state space is unique.
−
❅Ct
❅
❅
❅
❅
+
Ct
t−1 t t+1
+
−
Figure 1: Definition of Ct and Ct .
The state σ t (a) of a system sequence a at time t is determined by the
natural map
+
−
σ t : C → C/(Ct Ct ) = Σt ,
a homomorphism. There is therefore a well defined state sequence σ(a) =
{σ t (a) : t ∈ Z} associated with each a ∈ C, and a well defined state code
7
σ(C) = {σ(a) : a ∈ C} associated with C. The canonic realization C of a
group system C is the set of all pairs of sequences (a, σ(a)):
{(a, σ(a)) : a ∈ C},
(2)
where σ(a) is the state sequence of C. The state spaces of the canonic
realization are Σt . The canonic realization is a minimal realization of a
group system.
An element of the canonic realization C is denoted
b = . . . , bt−1 , bt , bt+1 , . . . ,
(3)
where component bt is given by bt = (st , at , st+1 ), where st ∈ Σt is the
canonic state at time t, and st+1 ∈ Σt+1 is the canonic state at time t + 1.
We say st is the left state of bt , and use notation (bt )− = st . We say st+1 is
the right state of bt , and use notation (bt )+ = st+1 . For any path b, as given
in (3), it is clear that for bt = (st , at , st+1 ) and bt+1 = (ŝt+1 , at+1 , st+2 ), we
must have st+1 = ŝt+1 or equivalently (bt )+ = (bt+1 )− .
Let σ(C) be the state code of C, the sequences of states
. . . , st−1 , st , st+1 , . . . in each b ∈ C.
Theorem 1 There is a group isomorphism from C to C given by the 1-1
correspondence a ↔ b, where a ∈ C and b ∈ C. If
a = . . . , at−1 , at , at+1 , . . . ,
and
b = . . . , bt−1 , bt , bt+1 , . . . ,
then for each time t, at 7→ bt = (st , at , st+1 ) is the assignment of the group
isomorphism.
Proof. There is a well defined state sequence σ(a) associated with each
a ∈ C. This means each a ∈ C is assigned to a well defined b ∈ C by the
assignment at 7→ bt = (st , at , st+1 ) for each time t. This map is a bijection
since if a ∈ C and â ∈ C are both assigned to the same b ∈ C, then we must
have at = ât for each time t, so a and â are the same.
•
We will be interested in canonic realization C rather than group system C
in the remainder of the paper. There is no loss in generality in considering
C rather than C because of the above 1-1 correspondence and isomorphism.
The canonic realization can be described with a graph [3]. Any other
minimal realization is graph isomorphic to the canonic realization [3]. We
think of component bt as a branch in a trellis section T t or an element in
branch group B t . Trellis section T t is a bipartite graph where the left vertices
are states in Σt , the right vertices are states in Σt+1 , and the label of a branch
(st , at , st+1 ) between state st and state st+1 is at ∈ At . B t is the group of
branches bt , which is a subdirect product, a subgroup of the direct product
group Σt × At × Σt+1 . Clearly there is a branch (st , at , st+1 ) in T t , with
label at between two vertices st and st+1 , if and only if (st , at , st+1 ) ∈ B t .
Then C can be described by a group trellis, a connected sequence of trellis
sections, where T t and T t+1 are joined together using the common states in
8
Σt+1 [3]. We refer to this as group trellis C. We regard group trellis C as
the first canonic form of group system C.
The states of B t are Σt and Σt+1 . We now describe state groups of B t
isomorphic to Σt and Σt+1 . Consider the projection map πL : B t → Σt onto
the left states of B t , given by the assignment (st , at , st+1 ) 7→ st . This is a
homomorphism with kernel X0t , where X0t is the subgroup of all elements
of B t of the form (1t , at , st+1 ), where 1t is the identity of Σt . Then by
the first homomophism theorem B t /X0t ≃ Σt . Also consider the projection
map πR : B t → Σt+1 onto the right states of B t , given by the assignment
(st , at , st+1 ) 7→ st+1 . This is a homomorphism with kernel Y0t , where Y0t is
the subgroup of all elements of B t of the form (st , at , 1t+1 ), where 1t+1 is the
identity of Σt+1 . Then by the first homomophism theorem B t /Y0t ≃ Σt+1 .
Thus any branch bt ∈ B t is of the form bt = (st , at , st+1 ) where st ∈ Σt ≃
B t /X0t and st+1 ∈ Σt+1 ≃ B t /Y0t . These results show there is a state group
isomorphism B t /Y0t ≃ Σt+1 ≃ B t+1 /X0t+1 at each time t + 1.
Let C be a group trellis, and let b be a trellis path in C. We define
time intervals for C as in [3]. We say bt is a trellis path segment of length
one on time interval [t, t]. In general, we say bt , bt+1 , . . . , bt+k is a trellis
path segment of length k + 1 on time interval [t, t + k]. Using (3), define
the projection map at time t, χt : C → B t , by the assignment b 7→ bt .
Define the projection map χ[t1 ,t2 ] : C → B t1 × · · · × B t2 by the assignment
b 7→ (bt1 , . . . , bt2 ); trellis path segment (bt1 , . . . , bt2 ) has length t2 −t1 +1. We
say that codeword b has span t2 − t1 + 1 if bt1 6= 1t1 , bt2 6= 1t2 , and bn = 1n
for n < t1 and n > t2 , where 1t is the identity of B t . We define C [t1 ,t2 ] to be
the set of trellis paths which are the identity outside time interval [t1 , t2 ].
For any integer l > 0, we say a group trellis C is l-controllable if for
any time epoch t, and any pair of states st and st+l , where st ∈ Σt and
st+l ∈ Σt+l , there is a trellis path segment of length l connecting the two
states [4]. A group trellis C is strongly controllable if it is l-controllable
for some integer l. The least integer l for which a group trellis is strongly
controllable is denoted as ℓ. In this paper, we only study the case l = ℓ.
The preceding definition of controllable is in terms of states. There is an
equivalent and consistent definition of controllable in terms of a trellis path
segment connecting two half infinite paths in [3].
9
3. THE STATIC MATRIX AND SHIFT MATRIX
In this section, we write the coset decomposition chain of a group as a
matrix, and call this a matrix chain. We study a matrix chain called a static
matrix. There is another matrix of group elements called a shift matrix. We
use both matrices to construct a tensor.
+
In like manner to C, define C t to be the set of all codewords in C for
which bn = 1n for n < t, where 1n is the identity component of B n at time
−
n. Define C t to be the set of all codewords in C for which bn = 1n for
n ≥ t. For all integers j, define
def
+
Xjt = {χt (b) : b ∈ C (t−j) }.
(4)
Note that X0t is consistent with the definition previously given in Section 2.
We have Xjt = 1t for j < 0. For all integers i, define
def
−
Yit = {χt (b) : b ∈ C (t+i+1) }.
(5)
Note that Y0t is consistent with the definition previously given in Section 2.
We have Yit = 1t for i > 0. It is clear that Xjt ✁ B t , and Yit ✁ B t for any
time t and any integer j.
The groups Xjt and Yit were first introduced in [4]. The group intersect , for 0 ≤ j ≤ ℓ, are the groups used in [4] to give an abstract
tions Xjt ∩ Yℓ−j
characterization of the branch group of an ℓ-controllable group trellis.
For any set H t ∈ B t , define (H t )+ to be the set of right states of H t , or
t+1
{s
: bt = (st , at , st+1 ) ∈ H t }, and define (H t )− to be the set of left states
of H t , or {st : bt = (st , at , st+1 ) ∈ H t }.
For sets H1t ⊂ B t and H1t+1 ⊂ B t+1 such that (H1t )+ = (H2t )− , define the
concatenation of H1t and H2t+1 , H1t ∧ H2t+1 , to be all the (valid) trellis path
segments of length two with first component in H1t and second component
in H2t+1 .
t+1 −
t+1 −
) for all integers i, j.
) and (Yit )+ = (Yi−1
Note that (Xjt )+ = (Xj+1
t+1
t+1
t
t
Then Xj ∧ Xj+1 and Yi ∧ Yi−1 are sets of trellis path segments of length
two.
The next result follows directly from Proposition 7.2 of [4], using our
notation.
Proposition 2 The group trellis C is ℓ-controllable if and only if Xℓt = B t ,
or equivalently, if and only if Yℓt = B t , for each time t.
The group B t has two normal series (and chief series)
t
1t = X−1
✁ X0t ✁ X1t ✁ · · · ✁ Xℓt = B t ,
and
t
1t = Y−1
✁ Y0t ✁ Y1t ✁ · · · ✁ Yℓt = B t .
We denote these normal series by {Xjt } and {Yit }.
The Schreier refinement theorem used to prove the Jordan-Hölder theorem [12] shows how to obtain a refinement of {Xjt } by inserting {Yit }; we
call this the forward Schreier series of {Xjt } and {Yit }. Since {Xjt } and {Yit }
are chief series, the forward Schreier series of {Xjt } and {Yit } is a chief series.
10
In equation (8), we have written the forward Schreier series as a matrix of
ℓ + 1 columns and ℓ + 2 rows. Note that the terms in the bottom row form
t , X t , X t , . . . X t , X t , and the terms in the top row form
the sequence X−1
0
1
ℓ−1
ℓ−2
t , X t . Thus (8) is indeed a refinement of the
t
the sequence X0 , X1t , X2t , . . . Xℓ−1
ℓ
normal series {Xjt }. We call (8) the matrix chain of the forward Schreier
series of {Xjt } and {Yit }.
Proposition 3 If the group trellis C is ℓ-controllable, then
t
t
) = Xjt ,
Xj−1
(Xjt ∩ Yℓ−j
(6)
for each t, for j ≥ 0.
Proof. If the group trellis C is ℓ-controllable, then from Proposition 7.2 of
[4], in our notation,
t
t
) = Xjt
) · · · (Xjt ∩ Yℓ−j
(X0t ∩ Yℓt )(X1t ∩ Yℓ−1
(7)
for all j ≥ 0. This means we can rewrite (7) as (6). (In general, the work
in [4] considers the ℓ-controllable time invariant group trellis. But it can
be seen Proposition 7.2 of [4] and its proof apply also to any ℓ-controllable
group trellis.)
•
t (X t ∩ Y t ) for
The diagonal terms of the matrix chain (8) are Xj−1
j
ℓ−j
t (X t ∩
j = 0, . . . ℓ. Proposition 3 shows that the diagonal terms satisfy Xj−1
j
t ) = X t for j = 0, . . . ℓ, if the group trellis is ℓ-controllable. For j ∈ [1, ℓ],
Yℓ−j
j
this means all column terms above the diagonal term are the same as the
diagonal term. Then we can reduce the matrix chain to a triangular form as
shown in (9). A triangle can be formed in two ways, depending on whether
the columns in (8) are shifted up or not; we have shifted the columns up
since it is more useful here. We call (9) the X [t,t] static matrix. The X [t,t]
static matrix is defined on time interval [t, t]. A typical entry in the matrix
t (X t ∩ Y t ).
is Xj−1
j
k−j
Theorem 4 The X [t,t] static matrix is a description (normal chain and
chief series) of the branch group B t of an ℓ-controllable group trellis.
Proof. Both {Xjt } and {Yit } are normal chains of the branch group B t .
Then by the Schreier refinement theorem, the forward Schreier series is a
normal chain of B t .
•
For each t, we can replace B t in the group trellis C with X [t,t] . We denote
the resulting structure by x; note that x is a tensor. Since X [t,t] is a coset
decomposition chain of B t , then x is a description of the coset structure of
group trellis C. Each path b ∈ C traverses some sequence of cosets in x.
Note that the first column of (9) is a description of X0 , which we can
think of as an input. The remaining columns are a description of B t /X0t ,
which is isomorphic to the state Σt . Thus columns of the static matrix
contain information about the input and state. Therefore an isomorphic
copy of the state code σ(C) is embedded in x.
11
∪
t (X t
X−1
0
∪
∩
Yℓt )
X0t (X1t
∪
∩
Yℓt )
···
∪
∪
t
t
t )
t
X0 (X1 ∩ Yℓ−1
∩ Yℓ−1 )
∪
∪
t (X t ∩ Y t )
t (X t ∩ Y t )
X
X−1
1
0
0
ℓ−2
ℓ−2
∪
∪
···
···
∪
∪
t (X t ∩ Y t )
t (X t ∩ Y t )
X
X−1
1
0
0
ℓ−j
ℓ−j
∪
∪
t
t
t
t (X t ∩ Y t
X−1
0
ℓ−j−1 ) X0 (X1 ∩ Yℓ−j−1 )
∪
∪
···
···
∪
∪
t
t
t
t
X0 (X1t ∩ Y1t )
X−1 (X0 ∩ Y1 )
∪
∪
t
t
t
t
X0 (X1t ∩ Y0t )
X−1 (X0 ∩ Y0 )
∪
∪
t
t
t
t
t
t )
X−1 (X0 ∩ Y−1 )
X0 (X1 ∩ Y−1
t (X t
X−1
0
t (X t
Xj−1
j
∩ Yℓt )
···
t )
∩ Yℓ−1
···
∪
···
···
···
···
···
···
···
···
···
t (X t
Xj−1
j
∪
t (X t ∩ Y t )
Xj−1
j
ℓ−2
∪
···
∪
t (X t ∩ Y t )
Xj−1
j
ℓ−j
∪
t (X t ∩ Y t
Xj−1
j
ℓ−j−1 )
∪
···
∪
t
Xj−1 (Xjt ∩ Y1t )
∪
t
Xj−1 (Xjt ∩ Y0t )
∪
t
t )
Xj−1 (Xjt ∩ Y−1
···
···
···
···
···
···
···
···
t (X t ∩ Y t )
Xℓ−1
ℓ
ℓ
∪
t (X t ∩ Y t )
Xℓ−1
ℓ−1
ℓ
∪
t (X t ∩ Y t )
Xℓ−1
ℓ
ℓ−2
∪
···
∪
t (X t ∩ Y t )
Xℓ−1
ℓ−j
ℓ
∪
t (X t ∩ Y t
Xℓ−1
ℓ−j−1 )
ℓ
∪
···
∪
t
Xℓ−1 (Xℓt ∩ Y1t )
∪
t
Xℓ−1 (Xℓt ∩ Y0t )
∪
t )
t
Xℓ−1 (Xℓt ∩ Y−1
(8)
q
q
q
q
t (X t ∩ Y t ) X t (1t )
t (X t ∩ Y t )
t (X t ∩ Y t )
t (X t ∩ Y t )
·
·
·
X
·
·
·
X
X
X−1
0
1
0
0
j−1
j
ℓ
ℓ
ℓ−1
ℓ−j
ℓ−1
ℓ
∪
∪
∪
∪
t (1t )
t (X t ∩ Y t
t (X t ∩ Y t ) X t (X t ∩ Y t )
···
Xℓ−1
···
Xj−1
X−1
0
1
0
j
ℓ−j−1 )
ℓ−2
ℓ−1
∪
∪
∪
···
···
···
···
···
∪
∪
∪
t (X t ∩ Y t )
t (X t ∩ Y t )
t (X t ∩ Y t )
·
·
·
X
···
X
X−1
1
0
0
j−1
j
k−1
k−j
k
∪
∪
∪
···
···
···
···
···
∪
∪
∪
t (X t ∩ Y t )
···
···
···
Xj−1
Xjt (1t )
0
j
∪
∪
∪
t
···
···
···
Xj−1 (1t )
∪
∪
t
t
t
t
X0 (X1t ∩ Y0t ) X1t (1t )
X−1 (X0 ∩ Y1 )
∪
∪
t (1t )
t (X t ∩ Y t )
X
X−1
0
0
0
∪
t (1t )
X−1
(9)
12
t (X t ∩ Y t )
X−1
0
ℓ
∪
t (X t ∩ Y t )
X−1
0
ℓ−1
∪
···
∪
t (X t ∩ Y t )
X−1
0
k
∪
···
∪
···
∪
···
∪
t
X−1 (X0t ∩ Y1t )
∪
t
X−1 (X0t ∩ Y0t )
∪
t (1t )
X−1
t+1
X0t+1 (X1t+1 ∩ Yℓ−1
)
∪
t+1
)
X0t+1 (X1t+1 ∩ Yℓ−2
∪
···
∪
t+1
)
X0t+1 (X1t+1 ∩ Yk−1
∪
···
∪
···
∪
···
∪
t+1
t+1
X0 (X1 ∩ Y0t+1 )
∪
t+1 t+1
X0 (1 )
···
···
···
···
···
···
···
t+j
t+j
Xj−1
(Xjt+j ∩ Yℓ−j
)
∪
t+j
t+j
Xj−1
(Xjt+j ∩ Yℓ−j−1
)
∪
···
∪
t+j
t+j
Xj−1
(Xjt+j ∩ Yk−j
)
∪
···
∪
t+j
Xj−1
(Xjt+j ∩ Y0t+j )
∪
t+j t+j
Xj−1 (1 )
···
···
t+ℓ
Xℓ−1
(Xℓt+ℓ ∩ Y0t+ℓ ) Xℓt+ℓ+1 (1t+ℓ+1 )
∪
t+ℓ t+ℓ
(1 )
Xℓ−1
···
···
···
···
···
(10)
Number the columns of (9) as 0, . . . , ℓ + 1. By using column j of static
matrix X [t+j,t+j] , for j = 0, . . . , ℓ + 1, we form the matrix shown in (10), of
the same shape as (9). In (10), we no longer have the inclusion from one
column to the next as in (9). However the coset decomposition within each
column is preserved. We call (10) the X [t,t+ℓ+1] shift matrix. The X [t,t+ℓ+1]
shift matrix is a composite of a single column from each of ℓ + 2 static
matrices. Notice the shift matrix is defined on time interval [t, t + ℓ + 1]. A
t+j
t+j
typical entry in the matrix is Xj−1
(Xjt+j ∩ Yk−j
).
For j = 0, . . . , ℓ + 1, the j-th column of static matrix X [t,t] is the j-th
column of a shift matrix X [t−j,t−j+ℓ+1] at time t − j. Thus the static matrix
X [t,t] is a composite of columns of ℓ + 2 shift matrices.
We now show that the X [t,t+ℓ+1] shift matrix (10) has a kind of shift
property, after some preliminary results. The discussion will show that the
shift matrix has a physical interpretation as paths that split from the identity
path.
For any time t, for each branch bt ∈ B t , we define the following branch
set F(bt ) to be the set of branches that can follow bt at the next time epoch
t + 1 in valid trellis paths. In other words, branch bt+1 ∈ F(bt ) if and
only if (bt )+ = (bt+1 )− . Then the following branch set F(bt ) represents the
contraction, correspondence, and expansion given by
η
bt 7→ bt Y0t ←→ bt+1 X0t+1 ,
η
where η is the 1-1 correspondence B t /Y0t ←→ B t+1 /X0t+1 given by the state
group isomorphism B t /Y0t ≃ B t+1 /X0t+1 .
It is clear that bt ∈ B t and F(bt ) ⊂ B t+1 . However note that F is not
a function with domain B t and range B t+1 . But we can think of F as a
13
relation on B t × B t+1 . In this relation, we can think of F as an assignment
of set F(bt ) to branch bt , or F : bt 7→ F(bt ).
Proposition 5 If (bt )+ = (bt+1 )− , the following branch set F(bt ) of a
branch bt in B t is the coset bt+1 X0t+1 in B t+1 , or the assignment F : bt →
7
bt+1 X0t+1 .
Define the following branch set F : B t → B t+1 such that for any set
⊂ B t , the set F(H t ) is the union ∪bt ∈H t F(bt ). The set F(H t ) always
t+1
consists of cosets of X0t+1 . In particular, F(Xjt ) = Xj+1
for all integers
j ≥ −1.
For a set H t ⊂ B t and integer j > 0, define F j (H t ) to be the j-fold
composition F j (H t ) = F ◦ F ◦ · · · ◦ F(H t ). For j = 0, define F j (H t ) =
F 0 (H t ) to be just H t . If H t is a set of trellis branches at time epoch t, then
F j (H t ) is the set of trellis branches at time epoch t + j, such that for each
bt+j ∈ F j (H t ) there is a bt ∈ H t and a path in the trellis from bt to bt+j .
Note that Xjt+j = F j (X0t ).
For a set H t ⊂ B t and integer k ≥ 0, define F [0,k] (H t ) to be the set of
all trellis path segments (bt , . . . , bt+k ) on time interval [t, t + k] that start
with a branch bt ∈ H t .
Ht
Proposition 6 For any subsets Gt , H t of B t , we have (Gt H t )+ =
(Gt )+ (H t )+ , (Gt H t )− = (Gt )− (H t )− , and F(Gt H t ) = F(Gt )F(H t ).
Proof. It is clear that (Gt H t )+ = (Gt )+ (H t )+ .
F(Gt H t ) = F(Gt )F(H t ).
Then it follows that
•
Proposition 7 For any subsets Gt , H t of B t , we have (Gt ∩H t )+ = (Gt )+ ∩
(H t )+ , (Gt ∩ H t )− = (Gt )− ∩ (H t )− , and F(Gt ∩ H t ) = F(Gt ) ∩ F(H t ).
t+j
t+j
) and
The matrix (10) contains terms of the form Xj−1
(Xjt+j ∩ Yk−j
t+j t+j
t+j
t+j
Xj−1
(1 ). We consider that terms of the form Xj−1
(Xjt+j ∩ Yk−j
) are in
t (X t ∩ Y t ) is in
column j and row k. For example, the lower left term X−1
0
0
t+j t+j
column j = 0 and row k = 0. The terms Xj−1
(1 ) are shorthand for
t+j
t+j
) when k = j − 1, and these are in column j and row
(Xjt+j ∩ Yk−j
Xj−1
[t,t+ℓ+1]
j − 1. Thus the X
shift matrix has ℓ + 2 columns j, for j such that
0 ≤ j ≤ ℓ + 1, and ℓ + 2 rows k, for k such that −1 ≤ k ≤ ℓ. Note that row
k has length k + 2 terms.
We now show the X [t,t+ℓ+1] shift matrix preserves shifts, that is, it has
a shift property.
Proposition 8 Fix k, 0 ≤ k ≤ ℓ, and fix j, 0 ≤ j ≤ k. The shift matrix
t+j+1
t+j+1
has a shift property: the term Xjt+j+1 (Xj+1
∩ Yk−j−1
) in column j + 1 and
t+j
t+j
row k is a shift of the term Xj−1
(Xjt+j ∩ Yk−j
) in column j and row k, that
is
t+j+1
t+j+1
t+j
t+j
∩ Yk−j−1
).
(11)
)) = Xjt+j+1 (Xj+1
(Xjt+j ∩ Yk−j
F(Xj−1
14
Proof. Fix k, 0 ≤ k ≤ ℓ, and fix j, 0 ≤ j ≤ k. We have
t+j
t+j
t+j
t+j
F(Xj−1
(Xjt+j ∩ Yk−j
)) = F(Xj−1
)F(Xjt+j ∩ Yk−j
)
t+j
= Xjt+j+1 (F(Xjt+j ) ∩ F(Yk−j
))
t+j+1
t+j+1
= Xjt+j+1 (Xj+1
∩ X0t+j+1 Yk−j−1
)
t+j+1
t+j+1
∩ Yk−j−1
)
= Xjt+j+1 X0t+j+1 (Xj+1
t+j+1
t+j+1
= Xjt+j+1 (Xj+1
∩ Yk−j−1
),
where the fourth equality follows from the Dedekind Law (if H, K, and L
are subgroups of group G with H ⊂ L, then HK ∩ L = H(K ∩ L)).
•
The first column of the shift matrix (10) will be important to us so we
define ∆tk = X0t ∩ Ykt , for −1 ≤ k ≤ ℓ. We now show that row k of the shift
matrix is just F [0,k+1] (∆tk ), or just the trellis path segments in C on time
interval [t, t + k + 1] that start with a branch bt ∈ ∆tk .
Theorem 9 Fix k, 0 ≤ k ≤ ℓ. We have
t+j
t+j
F j (∆tk ) = Xj−1
(Xjt+j ∩ Yk−j
),
(12)
t+j
for k < j ≤ ℓ, F j (∆tk ) = Xℓt+j for
for 0 ≤ j ≤ k + 1. And F j (∆tk ) = Xj−1
j > ℓ. The k-th row of the shift matrix is just the terms in F [0,k+1] (∆tk ).
Proof. We prove (12) by induction. Assume it is true for j = n. Then use
(11) to show it is true for j = n + 1. Then (12) shows the k-th row of the
shift matrix is just the terms in F [0,k+1] (∆tk ).
•
We now find a matrix of quotient groups using the shift matrix (10).
Using Theorem 9, we can represent a quotient group of adjacent terms in
the same column of shift matrix (10) in two equivalent ways:
t+j
t+j
Xj−1
(Xjt+j ∩ Yk−j
)
F j (∆tk )
,
= t+j t+j
t
t+j
j
F (∆k−1 )
Xj−1 (Xj ∩ Yk−j−1 )
(13)
for 0 ≤ j ≤ k.
Proposition 10 F [0,k] (∆tk−1 ) and F [0,k] (∆tk ) are groups.
Proof. ∆tk = X0t ∩ Ykt is a group so the trellis path segments in F [0,k] (∆tk )
are a group.
•
Proposition 11 F [0,k] (∆tk−1 ) ✁ F [0,k] (∆tk ) if and only if ∆tk−1 ✁ ∆tk . Then
as a result F [0,k] (∆tk−1 ) ✁ F [0,k] (∆tk ).
Theorem 12 We have
F [0,k] (∆tk )
∆tk
≃
.
∆tk−1
F [0,k] (∆tk−1 )
15
Proof. The projection χt : F [0,k] (∆tk ) → ∆tk is onto. It is a homomorphism with kernel F [0,k] (1t ). The projection χt : F [0,k] (∆tk−1 ) → ∆tk−1 is
onto. It is a homomorphism with kernel F [0,k] (1t ). Therefore, by the first
homomorphism theorem,
F [0,k] (∆tk )
≃ ∆tk ,
F [0,k] (1t )
F [0,k] (∆tk−1 )
≃ ∆tk−1 .
F [0,k] (1t )
Now use the correspondence theorem and third isomorphism theorem to
complete the proof.
•
Proposition 13 F j (∆tk−1 ) and F j (∆tk ) are groups.
Proof. See (12) or note that F j (∆tk ) is the projection of group F [0,k] (∆tk )
on the time interval [t + j, t + j].
•
Proposition 14 F j (∆tk−1 ) ✁ F j (∆tk ).
Proof. See (13).
•
Theorem 15 For 0 ≤ j ≤ k, we have
F [0,k] (∆tk )
F j (∆tk )
≃
.
F j (∆tk−1 )
F [0,k] (∆tk−1 )
Proof. The projection χt+j : F [0,k] (∆tk ) → F j (∆tk ) is onto. It is a homomorphism with kernel Kk , the path segments in F [0,k] (∆tk ) that are the identity at time t + j. The projection χt+j : F [0,k] (∆tk−1 ) → F j (∆tk−1 ) is onto.
It is a homomorphism with kernel Kk−1 , the path segments in F [0,k] (∆tk−1 )
that are the identity at time t + j. Therefore, by the first homomorphism
theorem,
F [0,k] (∆tk )
≃ F j (∆tk ),
Kk
F [0,k] (∆tk−1 )
≃ F j (∆tk−1 ).
Kk−1
We now show Kk = Kk−1 ; we first show Kk ⊂ Kk−1 .
Let
(bt , . . . , bt+j , . . . , bt+k ) be a path segment in F [0,k] (∆tk ) that is the identity
t ) = ∆t
at time t + j, 0 ≤ j ≤ k. But then bt must be in (X0t ∩ Yj−1
j−1 . Since
t
t+j
t
t
t
t
j ≤ k, then ∆j−1 ⊂ ∆k−1 and b ∈ ∆k−1 . Then (b , . . . , b , . . . , bt+k ) ∈
F [0,k] (∆tk−1 ) and (bt , . . . , bt+j , . . . , bt+k ) ∈ Kk−1 . Therefore Kk ⊂ Kk−1 .
We now show Kk−1 ⊂ Kk . Let (bt , . . . , bt+j , . . . , bt+k ) be a path segment
in F [0,k] (∆tk−1 ) that is the identity at time t + j, 0 ≤ j ≤ k. But then bt
t ) = ∆t
t
t
t
t
must be in (X0t ∩ Yj−1
j−1 . Since j ≤ k, then ∆j−1 ⊂ ∆k and b ∈ ∆k .
Then (bt , . . . , bt+j , . . . , bt+k ) ∈ F [0,k] (∆tk ) and (bt , . . . , bt+j , . . . , bt+k ) ∈ Kk .
Therefore Kk−1 ⊂ Kk .
We have just shown Kk = Kk−1 . Now use the correspondence theorem
and third isomorphism theorem to complete the proof.
•
16
Note that the proof breaks down if we try to go further. In other words,
we cannot show that for 0 ≤ j ≤ k, we have
F j (∆tk )
F [0,k] (∆tk )
≃
.
F j (∆tk−2 )
F [0,k] (∆tk−2 )
Define
def
Λ[t,t+k] =
F [0,k] (∆tk )
.
F [0,k] (∆tk−1 )
Corollary 16 For 0 ≤ j ≤ k, the t + j-th components of a transversal of
Λ[t,t+k] are a transversal of
F j (∆tk )
.
(14)
F j (∆tk−1 )
Proof. Theorem 15 shows that the projection χt+j (Λ[t,t+k] ) gives a 1-1 correspondence between cosets of Λ[t,t+k] and cosets of (14). Therefore the
projection χt+j of a transversal of Λ[t,t+k] is a transversal of (14).
•
Corollary 17 For 0 ≤ k ≤ ℓ, and 0 ≤ j ≤ k, we have
∆tk
F j (∆t )
F [0,k] (∆t )
≃ [0,k] t k ≃ j t k .
t
∆k−1
F (∆k−1 )
F
(∆k−1 )
Remark: This result can be regarded as a rectangle criterion for a shift
matrix, with ∆tk , ∆tk−1 , F j (∆tk ), and F j (∆tk−1 ) as the corners of a rectangle
in (10). It is similar in spirit to a quadrangle criterion for a Latin square [15]
or a configuration theorem for a net [16]. In fact, the rectangle condition
can be generalized further by starting with groups ∆tk and ∆tk−m, for m > 1.
These more general results are not needed.
We can use (13) and Corollary 17 to create a tensor. Fix j such that
t+j
0 ≤ j ≤ ℓ, and define Xjt+j //Xj−1
to be the column vector of quotient groups
t+j
t+j
Xj−1
(Xjt+j ∩ Yk−j
)
,
t+j
t+j
Xj−1
(Xjt+j ∩ Yk−j−1
)
(15)
for k such that j ≤ k ≤ ℓ. This is the vector of quotient groups formed from
groups in the normal chain in the center column of (10). Then using (13),
t+j def
=
Xjt+j //Xj−1
F j (∆tℓ )
F j (∆tℓ−1 )
···
F j (∆tk )
F j (∆tk−1 )
···
F j (∆tj )
F j (∆tj−1 )
T
.
t+j
For j = 0, . . . , ℓ, we obtain the column vectors Xjt+j //Xj−1
, which can be
[t,t+ℓ]
used to form the shift matrix X//
[t,t+ℓ] def
,
t+j
t+ℓ
t
X1t+1 //X0t+1 · · · Xjt+j //Xj−1
· · · Xℓt+ℓ //Xℓ−1
X0t //X−1
.
(16)
This is a second example of a shift matrix. The k-th row of shift matrix
[t,t+ℓ]
, 0 ≤ k ≤ ℓ, is a shift vector
X//
X//
=
F 0 (∆tk )
F 0 (∆tk−1 )
F 1 (∆tk )
F 1 (∆tk−1 )
···
F j (∆tk )
F j (∆tk−1 )
17
···
F k (∆tk )
F k (∆tk−1 )
.
The shift vector is just all the components of
Λ[t,t+k] =
F [0,k] (∆tk )
.
F [0,k] (∆tk−1 )
[t,t+ℓ]
is defined on time interval [t, t + ℓ] and has columns
Shift matrix X//
j for 0 ≤ j ≤ ℓ, and rows k for 0 ≤ k ≤ ℓ. We can think of shift matrix
[t,t+ℓ]
as a matrix of quotient groups of shift matrix X [t,t+ℓ+1] . Corollary
X//
[t,t+ℓ]
preserves isomorphism of quotient groups,
17 shows the shift matrix X//
and each shift of a quotient group in a row gives the next quotient group in
[t,t+ℓ]
as the natural shift
the row. Therefore we can regard a shift matrix X//
structure of a strongly controllable group system.
t
Fix j such that 0 ≤ j ≤ ℓ, and define Xjt //Xj−1
to be the column vector
of quotient groups
t (X t ∩ Y t )
Xj−1
j
k−j
,
(17)
t
t
t
Xj−1 (Xj ∩ Yk−j−1
)
for k such that j ≤ k ≤ ℓ. This is the vector of quotient groups formed from
groups in the normal chain in the center column of (9). For j = 0, . . . , ℓ, we
t , which can be used to form the static
obtain the column vectors Xjt //Xj−1
[t,t]
matrix X// ,
[t,t] def
X//
=
t
t
t
· · · Xℓt //Xℓ−1
X1t //X0t · · · Xjt //Xj−1
X0t //X−1
.
(18)
t+j
t
Note that the definition of Xjt+j //Xj−1
and Xjt //Xj−1
is consistent since
t+j
t+j
t
t
Xj //Xj−1 is defined using (15) and Xj //Xj−1 is defined using (17), and
(15) and (17) are consistent. Note that
(t−j)+j
t
= Xj
Xjt //Xj−1
(t−j)+j
(t−j)+j
//Xj−1
,
(t−j)+j
′
′
t +j
and we can think of Xj
//Xj−1
as the definition Xjt +j //Xj−1
with
′
time t defined by the parentheses term (t − j). Then we can also think of
static matrix (18) as
[t,t] def
(t)
(t−1)+1
(t−1)+1
(t−j)+j
(t−j)+j
(t−ℓ)+ℓ
(t−ℓ)+ℓ
X// = X0(t) //X−1
.
X1
//X0
· · · Xj
//Xj−1
· · · Xℓ
//Xℓ−1
(19)
Now it is clear that each term in (19) is from one of ℓ + 1 different shift
matrices.
[t,t+ℓ]
[t,t]
using the
We can relate a static matrix X// to a shift matrix X//
tensor description shown in (20). Time increases as we move up the page.
The vectors in the shift matrix (16) are the vectors along the diagonal in
(20), and the vectors in the static matrix (19) are the vectors in a row of
(20). The superscript parentheses terms in (20), like (t − j), indicate terms
that all belong to the same shift matrix. For example, the diagonal terms
(t−j)
X0
(t−j)
(t−j)+1
//X−1 , X1
(t−j)+1
//X0
(t−j)+j
, . . . , Xj
(t−j)+j
//Xj−1
(t−j)+ℓ
, . . . , Xℓ
[(t−j),(t−j)+ℓ]
(t−j)+ℓ
//Xℓ−1
. The
all belong to the shift matrix starting at time t − j, X//
center row in (20) is (19), which reduces to (18), which is just the static
18
,
[t,t]
matrix X// .
..
.
(t)+1
(t)+1
···
X1
//X0
(t−1)+1
(t−1)+1
X (t) //X (t)
X1
//X0
0
−1
..
..
.
.
(t−j)+1
(t−j)+1
···
X1
//X0
(t−j)
(t−j)
X
//X−1
···
0
(t−j)+ℓ
(t−j)+ℓ
//Xℓ−1
Xℓ
..
.
(t−ℓ)+ℓ
(t−ℓ)+ℓ
Xℓ
//Xℓ−1
..
.
(t)+ℓ
Xℓ
···
···
···
Xj
(t)+j
//Xj−1
..
.
(t−j)+j
//Xj−1
..
.
Xj
···
..
.
(t)+j
···
(t−j)+j
···
(20)
We let x// denote the tensor in (20), and say x// is a chain tensor. For a
given group trellis C, there is only one chain tensor x// . The tensor x// is
a description of the coset structure of group trellis C. The tensor x// has
a dual nature of having both shift matrices and static matrices. The most
natural and important way to understand x// is to look at (20) along the
diagonals, in terms of shift matrices.
Theorem 18 For each time t, the diagonals of (20) are a description of
the quotient groups Λ[t,t+k] for k such that 0 ≤ k ≤ ℓ.
In the next section, we will show how to recover paths b ∈ C from
generators, which are representatives of the coset structure described by x// .
19
(t)+ℓ
//Xℓ−1
..
.
4. GENERATORS AND THE GENERATOR MATRIX
We now review some results from [3]. Forney and Trott [3] define the kcontrollable subcode Ck of a group code C. We can transcribe their approach
to the group trellis C used here. The k-controllable subcode Ck of a group
trellis C is defined as the set of combinations of code sequences of span k + 1
or less:
Y
Ck =
C [t,t+k] .
t
They show
C0 ⊂ C1 ⊂ . . . Ck−1 ⊂ Ck ⊂ . . . Cℓ = C
is a normal series. Then in their Code Granule Theorem [3], they show
Ck /Ck−1 is isomorphic to a direct product of quotient groups Γ[t,t+k] ,
Y
Ck /Ck−1 ≃
Γ[t,t+k] ,
(21)
t
where Γ[t,t+k] is defined by
def
Γ[t,t+k] =
C [t,t+k]
.
C [t,t+k) C (t,t+k]
Γ[t,t+k] is called a granule. A coset representative of Γ[t,t+k] is called a generator. The coset representative of C [t,t+k)C (t,t+k] is always taken to be the
identity sequence. In case Γ[t,t+k] is isomorphic to the identity group, the
identity sequence is the only coset representative. A nonidentity generator
is an element of C [t,t+k] but not of C [t,t+k) or of C (t,t+k] , so its span is exactly k + 1. Thus every nonidentity generator is a codeword that cannot be
expressed as a combination of shorter codewords [3].
If Q is any quotient group, let [Q] denote a set of coset representatives of
Q, a transversal ofQQ. Let [Γ[t,t+k]
] be a transversal of Γ[t,t+k] . It follows from
(21) that the set t Γ[t,t+k] is a set of coset representatives for the cosets
of Ck−1 in Ck . We know that the set of coset representatives of the granule
Γ[t,t+k] , or [Γ[t,t+k] ], is a set of generators {g[t,t+k] }. This means (Generator
Theorem [3]) that every code sequence b can be uniquely expressed as a
product
Y Y
b=
g[t,t+k]
(22)
0≤k≤ℓ t
of generators g[t,t+k] . Thus every code sequence b is a product of some
sequence of generators, and conversely, every sequence of generators corresponds to some code sequence b. A basis of C is a minimal set of shortest
length generators that is sufficient to generate the group system C [8]. It
follows that a basis is a set of coset representatives of Γ[t,t+k] , for 0 ≤ k ≤ ℓ,
for each t ∈ Z.
In this paper, for each time t, we let component basis B t be the set of
generators {g[t,t+k] ∈ [Γ[t,t+k] ] : 0 ≤ k ≤ ℓ} in all transversals at time t. The
sequence of component bases, . . . , B t , B t+1 , . . ., gives a basis B = {B t : t ∈
Z}. If component basis B t is the same for all t ∈ Z, then B is a constant
basis.
We now show that the projection χ[t,t+k] of generators in [Γ[t,t+k] ] is
also a transversal of Λ[t,t+k] . Therefore a basis B of C can be found using
representatives of either Γ[t,t+k] or Λ[t,t+k] .
20
Lemma 19 The set of paths formed by the concatenation of groups
t+j
t+j+1
t+j+1
)∧(Xj+1
∩Yk−j−1
)∧· · ·∧(Xkt+k ∩Y0t+k )∧1t+k+1 , 1t+k+2 , . . .
. . . , 1t−2 , 1t−1 ∧(X0t ∩Ykt )∧· · ·∧(Xjt+j ∩Yk−j
(23)
is C [t,t+k].
Proof. From the proof of Proposition 8, we have
t+j
t+j+1
t+j+1
F(Xjt+j ∩ Yk−j
) = X0t+j+1 (Xj+1
∩ Yk−j−1
).
This means the set of paths formed by the concatenation of groups in (23) is
t+j
well defined: for any branch bt+j ∈ Xjt+j ∩ Yk−j
, there is a branch bt+j+1 ∈
t+j+1
t+j+1
∩ Yk−j−1
such that (bt+j )+ = (bt+j+1 )− , and (bt+j , bt+j+1 ) is a trellis
Xj+1
path segment of length two. The paths in (23) consist of sequences which
split from the identity state at time t and merge to the identity state at time
t + k + 1. Therefore, any path in (23) must be in C [t,t+k].
Fix integer k such that 0 ≤ k ≤ ℓ. Let b be a sequence in C [t,t+k] . We
now show b is in (23). If b ∈ C [t,t+k] , then for each j, 0 ≤ j ≤ k, bt+j
t+j
must be in Xjt+j , but cannot be in Xm
, m > j. Similarly, bt+j must be
t+j
t+j
for all j ∈ [0, k]. Since (23) contains all
in Yk−j
. Then bt+j ∈ Xjt+j ∩ Yk−j
t+j
code sequences whose component bt+j ∈ Xjt+j ∩ Yk−j
for all j ∈ [0, k], then
b is in (23).
•
t+j
Lemma 20 For j, 0 ≤ j ≤ k, we have χt+j (C [t,t+k] ) = Xjt+j ∩ Yk−j
. For
t+k
t+k
t+k
[t,t+k]
t
t
[t,t+k]
t
).
) and Y0 ∩Xk = χ (C
example, this means X0 ∩Yk = χ (C
t+j
Proof. From (23), we know χt+j (C [t,t+k] ) ⊂ Xjt+j ∩ Yk−j
.
t+j
⊂ χt+j (C [t,t+k] ). The proof of Lemma 19
We now show Xjt+j ∩ Yk−j
t+j
shows that for any branch bt+j ∈ Xjt+j ∩ Yk−j
, there is a branch bt+j+1 ∈
t+j+1
t+j+1
Xj+1
∩ Yk−j−1
such that (bt+j , bt+j+1 ) is a trellis path segment of length
t+j+1
∩
two. We can continue this argument: for any branch bt+j+1 ∈ Xj+1
t+j+1
t+j+2
t+j+2
Yk−j−1
, there is a branch bt+j+2 ∈ Xj+2
∩Yk−j−2
such that (bt+j+1 , bt+j+2 )
is a trellis path segment of length two. Continuing the argument further
t+j
shows that for any branch bt+j ∈ Xjt+j ∩ Yk−j
, there is a trellis path segment
t+j
t+j+1
t+k
of length k −j +1, (b , b
, . . . , b ), which merges to the identity state
at time t+k+1. This argument works in reverse time as well: for any branch
t+j
t+j−1
t+j−1
bt+j ∈ Xjt+j ∩ Yk−j
, there is a branch bt+j−1 ∈ Xj−1
∩ Yk−j+1
such that
t+j−1
t+j
(b
, b ) is a trellis path segment of length two, and so on. Thus we see
t+j
that for any bt+j ∈ Xjt+j ∩ Yk−j
, there is a sequence b ∈ C [t,t+k] such that
t+j
χt+j (b) = bt+j . Thus we have shown Xjt+j ∩ Yk−j
⊂ χt+j (C [t,t+k] ).
•
Lemma 21 We have
χ[t,t+k] (C [t,t+k] ) ⊂ F [0,k] (∆tk ),
(24)
χ[t,t+k] (C [t,t+k) C (t,t+k] ) ⊂ F [0,k] (∆tk−1 ).
(25)
and
Proof. We have (24) holds if and only if χt (C [t,t+k] ) ⊂ ∆tk . But this follows
from Lemma 20. We have (25) holds if and only if χt (C [t,t+k) C (t,t+k] ) ⊂
•
∆tk−1 . But χt (C [t,t+k) C (t,t+k] ) = χt (C [t,t+k) ) = ∆tk−1 from Lemma 20.
21
Theorem 22 There is an isomorphism
µ
Γ[t,t+k] ≃ Λ[t,t+k] ,
where the 1-1 correspondence µ between cosets of Γ[t,t+k] and Λ[t,t+k] is given
by
µ : C [t,t+k) C (t,t+k] b 7→ F [0,k] (χt (C [t,t+k) C (t,t+k] b)).
(26)
Proof. Using Lemma 21, we have
χt (C [t,t+k) C (t,t+k] b) = χt (C [t,t+k) b)
= χt (C [t,t+k) )χt (b)
= ∆tk−1 bt .
Since Λ[t,t+k] = F [0,k] (∆tk )/F [0,k] (∆tk−1 ), this shows we can properly define
the 1-1 correspondence µ between cosets of Γ[t,t+k] and Λ[t,t+k] as given in
(26).
Forney and Trott [3] define an input chain F0t ⊂ F1t ⊂ · · · ⊂ Fℓt by the
def
projection Fkt = χt (C [t,t+k] ) for k = 0, 1, . . . , ℓ. Using Lemma 20, this gives
Fkt = ∆tk . In their Input Granule Theorem [3], Forney and Trott show that
t
for k such that 0 < k ≤ ℓ. Then we have
Γ[t,t+k] ≃ Fkt /Fk−1
t
= ∆tk /∆tk−1 .
Γ[t,t+k] ≃ Fkt /Fk−1
Combining this with Theorem 12 gives
Γ[t,t+k] ≃ ∆tk /∆tk−1 ≃ Λ[t,t+k] .
Then following the correspondences given in the Input Granule Theorem of
[3] and Theorem 12 shows that the isomorphism Γ[t,t+k] ≃ Λ[t,t+k] is given
by µ.
•
Theorem 23 Let [Γ[t,t+k] ] be a set of generators {g[t,t+k] } which is a
transversal of Γ[t,t+k] . Then {χ[t,t+k] (g[t,t+k] ) : g[t,t+k] ∈ [Γ[t,t+k] ]} is a
transversal of Λ[t,t+k] .
The above theorem shows that any set of Forney-Trott generators can
equally well be found from a tranversal of Λ[t,t+k] .
If Q is any quotient group, there is another way we denote a transversal
of Q besides [Q]. If {q} is a set of coset representatives of Q which is a
transversal of Q, we let [{q}] denote a transversal of Q.
Fix k such that 0 ≤ k ≤ ℓ. Let generator g[t,t+k] be a representative in
[t,t+k]
Γ
,
t+j
t+k t+k+1 t+k+2
t+1
t
,1
,1
,...,
, . . . , rj,k
, . . . , rk,k
, r1,k
g[t,t+k] = . . . , 1t−2 , 1t−1 , r0,k
(27)
t+j
t+j
t+j
From Lemma 20 we know component rj,k is an element of Xj ∩ Yk−j , and
t+j
from Theorem 23 and Corollary 16 we know rj,k
is a representative of
t+j
t+j
Xj−1
(Xjt+j ∩ Yk−j
)
F j (∆tk )
,
=
t
t+j
t+j
t+j
j
F (∆k−1 )
Xj−1 (Xj ∩ Yk−j−1 )
(28)
for j = 0, 1, . . . , k. If we pick a set of generators g[t,t+k] which is a transversal
t+j
of Γ[t,t+k] , [Γ[t,t+k] ], then [Γ[t,t+k] ] induces a transversal [{rj,k
}] of (28), for
j = 0, 1, . . . , k.
22
Theorem 24 Let [Γ[t,t+k] ] be a set of generators {g[t,t+k] } which is a
transversal of Γ[t,t+k] . For j = 0, . . . , k, the t + j-th component of generators
g[t,t+k] ∈ [Γ[t,t+k] ] forms a transversal
t
[{χt+j (g[t,t+k] )}] = [{rj,k
}]
of (28).
The Zassenhaus lemma [12] is a crucial step in the proof of the Schreier
refinement theorem. In the time invariant case, we can use the Zassenhaus
lemma to give a second proof of Theorem 24 [20], and Theorems 22-24 can
be viewed as a generalization of the Zassenhaus lemma.
Pick a generator g[t,t+k] in Γ[t,t+k] for each k, 0 ≤ k ≤ ℓ. We can arrange
the nontrivial components of these generators in a matrix as shown in (30),
which is called a shift matrix, or also a generator matrix, at time t, and
denoted R[t,t+ℓ] . The k-th row of matrix R[t,t+ℓ] , 0 ≤ k ≤ ℓ, is a shift vector,
also called a generator vector, denoted r[t,t+k] , where
def
t+j
t+1
t+k
t
r[t,t+k] = (r0,k
, r1,k
, . . . , rj,k
, . . . , rk,k
).
(29)
A generator vector r[t,t+k] is the nontrivial components of the generator
g[t,t+k] .
t
r0,ℓ
t
r0,ℓ−1
..
.
t+1
r1,ℓ
t+1
r1,ℓ−1
..
.
···
···
..
.
···
···
..
.
t+j
rj,ℓ
t+j
rj,ℓ−1
..
.
···
···
..
.
···
···
..
.
t
r0,k
..
.
t+1
r1,k
..
.
···
..
.
···
..
.
t+j
rj,k
..
.
···
..
.
t+k
rk,k
···
..
.
···
..
.
···
..
.
···
t+j
rj,j
t
r0,2
t
r0,1
t
r0,0
t+1
r1,2
t+1
r1,1
t+2
r2,2
···
···
..
.
t+ℓ−1
t+ℓ
rℓ−1,ℓ
rℓ,ℓ
t+ℓ−1
rℓ−1,ℓ−1
(30)
We define rt+j
to be a column vector in (30), for 0 ≤ j ≤ ℓ, where
j
def
=
rt+j
j
t+j
t+j
t+j
rj,ℓ
· · · rj,k
· · · rj,j
Then we can rewrite (30) as
T
.
t+j
t+ℓ
R[t,t+ℓ] = (rt0 , rt+1
1 , . . . , rj , . . . , rℓ ).
(31)
There is another related form, shown in (32), called the static matrix
t is just an element in X t ∩ Y t . As can be seen,
R[t,t] , where component rj,k
j
k−j
all components of the static matrix occur at time t. For a generator matrix,
the first column specifies the matrix completely. For a static matrix, the
23
first column does not determine the static matrix uniquely.
t
r0,ℓ
t
r0,ℓ−1
..
.
t
r1,ℓ
t
r1,ℓ−1
..
.
···
···
..
.
···
···
..
.
t
rj,ℓ
t
rj,ℓ−1
..
.
···
···
..
.
···
···
..
.
t
r0,k
..
.
t
r1,k
..
.
···
..
.
···
..
.
t
rj,k
..
.
···
..
.
t
rk,k
···
..
.
···
..
.
···
..
.
···
t
rj,j
t
r0,2
t
r0,1
t
r0,0
t
r1,2
t
r1,1
t
r2,2
···
···
..
.
t
t
rℓ−1,ℓ
rℓ,ℓ
t
rℓ−1,ℓ−1
(32)
We can rewrite (32) as
R[t,t] = (rt0 , rt1 , . . . , rtj , . . . , rtℓ ).
(33)
We can relate a static matrix R[t,t] to a generator matrix R[t,t+ℓ] using
the tensor description shown in (36). Time increases as we move up the
page. The vectors in the generator matrix (31) are the vectors along the
diagonal in (36), and the vectors in the static matrix (33) are the vectors
in a row of (36). The superscript parentheses terms in (36), like (t − j),
indicate terms that all belong to the same generator matrix. For example,
(t−j) (t−j)+1
(t−j)+j
(t−j)+ℓ
all belong to the
the diagonal terms r0
, r1
, . . . , rj
, . . . , rℓ
[(t−j),(t−j)+ℓ]
generator matrix starting at time t − j, R
. The center row in
(36) is
(t) (t−1)+1
(t−j)+j
(t−ℓ)+ℓ
),
(34)
(r0 , r1
, . . . , rj
, . . . , rℓ
where each entry is itself a column; this reduces to
(rt0 , rt1 , . . . , rtj , . . . , rtℓ ),
(35)
which is just the static matrix R[t,t] . Notice that each term in (34) and (35)
is from one of ℓ + 1 different shift matrices.
..
.
(t)+ℓ
rℓ
.
..
(t)+j
···
rj
···
..
(t−j)+ℓ
.
rℓ
..
(t)+1
(36)
···
r1
···
.
(t)
(t−1)+1
(t−j)+j
(t−ℓ)+ℓ
r
r1
· · · rj
···
rℓ
0
.
.
.
.
.
.
.
.
.
.
.
.
(t−j)+1
r1
···
···
(t−j)
r
···
0
..
.
24
Theorem 25 Fix time t. A finite sequence of ℓ + 1 generator matrices
R[(t−j),(t−j)+ℓ] at times t − j, for j = 0, . . . , ℓ, uniquely determines a static
matrix R[t,t] , where column j of generator matrix R[(t−j),(t−j)+ℓ] , denoted
(t−j)+j
rj
, is column j of static matrix R[t,t] , denoted rtj .
Proof. The center row in (36) is (34), which reduces to (35), which is just
(t−j)+j
static matrix R[t,t] . But entry rj
in (34) is just the (j + 1)-th column
[(t−j),(t−j)+ℓ]
of the generator matrix R
at time t − j.
•
We let r denote the tensor in (36), and say r is a representative tensor.
We can regard r in two different ways, as a sequence of static matrices or as
a sequence of shift matrices. In the first way we can write r as
r = . . . , rt , rt+1 , . . . ,
(37)
where each rt is a static matrix R[t,t] in the set of all static matrices, denoted
Rt . Therefore (37) is equivalent to
r = . . . , R[t,t] , R[t+1,t+1] , . . . .
We have just seen from Theorem 25 that each rt is determined by ℓ + 1
shift matrices. Then tensor r in (37) is also determined by a sequence of
shift matrices. We denote this interpretation of r using notation
r ∼ . . . , R[t,t+ℓ] , R[t+1,t+1+ℓ] , . . . ,
where each shift matrix R[t,t+ℓ] is in the set of all possible shift matrices,
denoted R[t,t+ℓ] . We define tensor set R to be the set of representative
tensors r determined by the Cartesian product of all possible shift matrices,
R∼
∞
Y
R[t,t+ℓ] .
t=−∞
We show that there is a 1-1 correspondence between tensors r ∈ R and
paths b ∈ C in Section 5.
Each tensor r ∈ R is a selection of one generator vector r[t,t+k] from a
generator g[t,t+k] in component basis B t of basis B, for 0 ≤ k ≤ ℓ, for each
t ∈ Z. The sequence of component bases B t forms basis B. Each basis B
gives a unique tensor set R and each tensor set R corresponds to a unique
basis B.
Each tensor r ∈ R gives one shift matrix R[t,t+ℓ] and one static matrix
R[t,t] at each time t. From Theorem 23, we can regard R[t,t+ℓ] as a selection
of one coset representative (generator vector) from a single coset of each of
[t,t+ℓ]
. Thus R[t,t+ℓ] has the
the quotient groups {Λ[t,t+k] : 0 ≤ k ≤ ℓ} in X//
[t,t+ℓ]
shift matrix. Similarly R[t,t] is
same form and time indices as the X//
a selection of one coset representative from a single coset of each quotient
[t,t]
[t,t]
group in X// . Thus R[t,t] has the same form and time indices as the X//
static matrix. This explains why tensor r in (36) has the same form as
tensor x// in (20).
A given r ∈ R produces a sequence of shift matrices R[t,t+ℓ] and a sequence of static matrices R[t,t] . Any sequence of shift matrices corresponds
25
to some r ∈ R and uniquely determines a sequence of static matrices. But an
arbitrary sequence of static matrices may not correspond to a valid sequence
of generator vectors and therefore an r ∈ R.
The following lemma is essentially a change of variable in Theorem 24.
Lemma 26 Fix j such that 0 ≤ j ≤ ℓ. Fix k such that j ≤ k ≤ ℓ.
Let [Γ[t−j,t−j+k]] be a set of generators {g[t−j,t−j+k] } which is a transversal of Γ[t−j,t−j+k]. The (t − j) + j-th component of generators g[t−j,t−j+k] ∈
[Γ[t−j,t−j+k]] forms a transversal
(t−j)+j
[{χ(t−j)+j (g[(t−j),(t−j)+k] )}] = [{rj,k
of
t (X t ∩ Y t )
Xj−1
j
k−j
t (X t ∩ Y t
Xj−1
j
k−j−1 )
t
}] = [{rj,k
}]
.
(38)
(39)
Proof. Fix j, where 0 ≤ j ≤ ℓ, and examine time t − j. Fix k such that
j ≤ k ≤ ℓ. Pick a set of generators {g[(t−j),(t−j)+k] } which is a transversal
of Γ[(t−j),(t−j)+k] , denoted [Γ[(t−j),(t−j)+k] ]. Then [Γ[(t−j),(t−j)+k] ] induces a
(t−j)+m
}] of
transversal [{rj,k
(t−j)+m
Xj−1
(t−j)+m
(Xj
(t−j)+m
(t−j)+m
Xj−1
(Xj
(t−j)+m
∩ Yk−j
∩
)
(t−j)+m
Yk−j−1 )
,
(t−j)+m
for m = 0, 1, . . . , k. Choose m = j. Then [{rj,k
(t−j)+j
}]
[{rj,k
=
t }]
[{rj,k
(40)
}] is a transversal
of (40) for m = j, which is the same as (39).
•
(t−j)+j
}] for k such that j ≤ k ≤ ℓ
Note that the set of transversals [{rj,k
are the coset representatives of all cosets in quotient groups in column j of
(t−j)+j
(t−j)+j
[(t−j),(t−j)+ℓ]
, which is column Xj
//Xj−1
. In other
shift matrix X//
t
words, the set of transversals [{rj,k }] for k such that j ≤ k ≤ ℓ are the
coset representatives of all cosets in quotient groups in column j of static
[t,t]
t . By selecting transversals [{r t }]
matrix X// , which is column Xjt //Xj−1
j,k
[t,t]
for j ≤ k ≤ ℓ, for each column j of static matrix X// , 0 ≤ j ≤ ℓ, we obtain
a complete set of coset representatives for the normal chain of B t given by
the X [t,t] static matrix. This gives the following result.
Theorem 27 For 0 ≤ j ≤ ℓ, for k such that j ≤ k ≤ ℓ, let [Γ[t−j,t−j+k]]
be a set of generators {g[t−j,t−j+k] } which is a transversal of Γ[t−j,t−j+k].
The (t − j) + j-th component of generators g[t−j,t−j+k] ∈ [Γ[t−j,t−j+k]] forms
a transversal (38) of (39) for 0 ≤ j ≤ ℓ, for j ≤ k ≤ ℓ. The set of
t }], for 0 ≤ j ≤ ℓ, for j ≤ k ≤ ℓ, forms a complete set
transversals, [{rj,k
of coset representatives for the normal chain of B t given by the X [t,t] static
matrix.
t from each transversal [{r t }],
By selecting one coset representative rj,k
j,k
for j ≤ k ≤ ℓ, for 0 ≤ j ≤ ℓ, we obtain a coset representative chain for the
26
normal chain of B t given by the X [t,t] static matrix. Any branch bt ∈ B t
can be written using elements of some coset representative chain as
ℓ
ℓ
Y
Y
t
bt =
rj,k
.
(41)
j=0
k=j
By the convention used here, equation (41) is evaluated as
t
t
t
t
t
t
t
t
t
t
t
t t
· · · r0,2
r0,1
r0,0
.
· · · r1,1
r0,ℓ
· · · rj,j
· · · r2,2
r1,ℓ
· · · rj,k
· · · rj,ℓ
rℓ−1,ℓ rℓ−1,ℓ−1
bt = rℓ,ℓ
(42)
Note that bt is the product of terms in some static matrix R[t,t] , where the
inner product in parentheses in (41) is just the product of terms in the j-th
column of R[t,t] . Using (38), (41) can be written in equivalent forms as
ℓ
ℓ
Y
Y
t
rj,k
bt =
(43)
j=0
=
ℓ
Y
j=0
=
ℓ
Y
j=0
k=j
ℓ
Y
k=j
ℓ
Y
k=j
(t−j)+j
rj,k
(44)
χt (g[t−j,t−j+k] ) .
(45)
We have just shown that for any time t, we can find any branch bt ∈ B t
using a selected set of generators at times t − j, for j = 0, . . . , ℓ. However
we have not shown we can construct any path in C this way. We show this
in the next section.
27
5. THE TIME DOMAIN ENCODER
For 0 ≤ j < ℓ, we know that in shift matrix R[t,t+ℓ] there is a column
vector
T
t+j
t+j
t+j
,
(46)
rt+j
= rj,ℓ
· · · rj,k
· · · rj,j
j
and a column vector
t+j+1
rj+1
=
t+j+1
t+j+1
t+j+1
rj+1,ℓ
· · · rj+1,k
· · · rj+1,j+1
T
.
(47)
t+j+1
Column rj+1
is completely determined by column rt+j
j . Then we can
t+j+1
t+j
t+j+1
think of rj+1
as a shift of rt+j
be the
j . For 0 ≤ j ≤ ℓ, let Rj , Rj+1
t+j t+j+1
[t,t+ℓ]
set of all columns rj , rj+1 in all possible shift matrices R
. For
t+j+1
0 ≤ j < ℓ, define a column shift map σ : Rt+j
→ Rj+1
by the assignment
j
t+j
t+j+1
t+j
t+j+1
σ : rj → rj+1 , where this assignment is given by σ : rj,k
7→ rj+1,k
for
t+j
t+j
j < k ≤ ℓ. Note that σ(rj,j
) is not defined since rj,j
“shifts out”. We
t+j
t+j
t+j
t+j
abbreviate σ(rj ) as σrj and σ(rj,k ) as σrj,k . (The notation σ and
σ should have a column index j as subscript and a time index t + j as
superscript, but these are left out for notational simplicity; any ambiguity
in σ or σ can be resolved by looking at their arguments.) Define
def
σrt = (σrt0 , σrt1 , . . . , σrtj , . . . , σrtℓ−1 ).
We can now characterize paths r ∈ R.
Theorem 28 Let w = . . . , rt , rt+1 , . . . be an arbitrary sequence, not necessarily a tensor in R, where rt ∈ Rt for each time t ∈ Z. Then w is a tensor
t
in R if and only if for each time t + 1, rt+1 = (rt+1
0 , σr ) where new input
t+1
t+1
r0 is any element of R0 and σrt is a shift of rt .
Proof. First assume w ∈ R. Then we know w is formed from a sequence
of shift matrices. Consider (rt , rt+1 ) where rt ∈ Rt and rt+1 ∈ Rt+1 . Fix
(t−j)+j
in shift matrix
0 ≤ j < ℓ. We know column rtj of rt is a column rj
[(t−j),(t−j)+ℓ]
R
. From the preceding discussion of shifts, we know
(t−j)+j
σrtj = σrj
(t−j)+j+1
= rj+1
= rt+1
j+1 ,
(t−j)+j+1
where rj+1
is a column in shift matrix R[(t−j),(t−j)+ℓ] and rt+1
j+1 is a
t+1
column in r . Then
t+1 t+1
t+1
t+1
rt+1 = (rt+1
0 , r1 , r2 , . . . , rj+1 , . . . , rℓ )
t
t
t
t
= (rt+1
0 , σr0 , σr1 , . . . , σrj , . . . , σrℓ−1 )
t
= (rt+1
0 , σr ),
where rt+1
∈ Rt+1 .
0
t
Conversely, assume w = . . . , rt , rt+1 , . . . and rt+1 = (rt+1
0 , σr ) for each
t+j+1
t + 1 ∈ Z. Fix t ∈ Z. Then we have rj+1
= σrt+j
for 0 ≤ j < ℓ. Then the
j
28
[t,t+ℓ] . Since this holds for
columns rt+j
j , 0 ≤ j ≤ ℓ, form a shift matrix R
each t ∈ Z, we have shown w is a sequence of shift matrices, and therefore
a tensor in R.
•
An encoder of the group trellis C is a finite state machine that, given
a sequence of inputs, can produce any path in the group trellis. Forney
and Trott [3] first showed that a group code C has an encoder with a shift
structure; their approach can be thought of as a spectral domain approach
[19]. We give an encoder here with a shift structure which uses a time
domain approach [19].
Theorem 28 shows the tensor set R has a natural shift structure. In the
remainder of this section, we show that any path b ∈ C is the encoding of
some r ∈ R, and this gives a 1-1 correspondence R ↔ C. The encoder has
a sliding block structure that uses generator vectors in R. The encoder is
given in (41) and (44)-(45). It is useful to think of (41) and (44)-(45) as
equivalent forms of the same encoder; each version is useful in the following
discussion.
Assume we have found a basis B. Then we have found generators
g[t,t+k] ∈ [Γ[t,t+k] ] for each t ∈ Z, for 0 ≤ k ≤ ℓ. Fix time t. The nontrivial components of the selected generators in encoder (45) form a generator
matrix R[(t−j),(t−j)+ℓ] , for j = 0, . . . , ℓ. From Theorem 25, these generator
matrices uniquely determine a static matrix R[t,t] , where column j of gen(t−j)+j
erator matrix R[(t−j),(t−j)+ℓ] , rj
, is column j of static matrix R[t,t] , rtj .
Then we can see that (36) has the form of a sliding block encoder. At each
time t, we select a new generator matrix R[(t),(t)+ℓ] whose column vectors
(t−j)+j
are shown along the diagonals in (36). The column vectors rj
of the
generator matrix at time t − j,
(t−j)
R[(t−j),(t−j)+ℓ] = (r0
(t−j)+1
, r1
(t−j)+j
, . . . , rj
(t−j)+ℓ
, . . . , rℓ
),
of the generator matrix at time t,
and column vectors rt+j
j
(t)
(t)+1
R[(t),(t)+ℓ] = (r0 , r1
(t)+j
, . . . , rj
(t)+ℓ
, . . . , rℓ
),
are shown along the diagonals of (36). As time increases, we slide along the
infinite matrix in (36) from left to right. At time t, the output branch bt of
the sliding block encoder is calculated from the static matrix
(t)
(t−1)+1
R[t,t] = (r0 , r1
(t−j)+j
, . . . , rj
(t−ℓ)+ℓ
, . . . , rℓ
)
= (rt0 , rt1 , . . . , rtj , . . . , rtℓ ),
whose terms are shown in the center row in (36). The first term in
(t)
the center row is the new input r0 , the first column vector of the new
[(t),(t)+ℓ]
generator matrix R
selected at time t, and the remaining terms
(t−1)+1
(t−j)+j
(t−ℓ)+ℓ
are from previous generator matrices ser1
, . . . , rj
, . . . , rℓ
lected at times t − 1, . . . , t − j, . . . , t − ℓ, respectively. To calculate branch bt
at time t, the sliding block encoder uses time window [t − ℓ, t], and therefore
the encoder is causal. We now show that we can use (36) to implement (41)
as a sliding block encoder. We need the following useful lemma.
t
Lemma 29 Let (rt , rt+1 ) ∈ Rt × Rt+1 such that rt+1 = (rt+1
0 , σr ). Then
encoder (41) encodes (rt , rt+1 ) into a trellis path segment (bt , bt+1 ) of length
2 in group trellis C. In other words, bt ∈ B t , bt+1 ∈ B t+1 , and bt+1 ∈ F(bt ).
29
Proof. Using (41), the encoding of (rt , rt+1 ) is
ℓ
ℓ
ℓ
ℓ
ℓ
ℓ
Y
Y
Y
Y
Y
Y
t
t
rj,k
,
r t+1 =
rj,k
,
j,k
j=0
j=0
k=j
j=0
k=j
=
ℓ
Y
j=0
k=j
ℓ
Y
k=j
t
rj,k
,
ℓ
Y
k=0
ℓ
Y
k=0
! ℓ ℓ
Y Y
r t+1
r t+1
0,k
j,k
j=1
k=j
! ℓ−1 ℓ
Y
Y
t+1
t+1
r0,k
rj+1,k
.
j=0
k=j+1
(48)
(t+1)
t
We know that rt+1 is of the form (rt+1
0 , σr ). Then for 0 ≤ j ≤ ℓ − 1, rj+1,k
(t)
is a shift of rj,k . Since
(t)
rj,k = χt (g[t−j,t−j+k] ),
then
(t+1)
rj+1,k = χt+1 (g[t−j,t−j+k] ).
This means that we can rewrite (48) in terms of generators (see (45)) as
! ℓ−1 ℓ
ℓ
ℓ
ℓ
Y
Y
Y
Y
Y
t+1
χt (g[t−j,t−j+k] ) ,
r0,k
χt+1 (g[t−j,t−j+k] ) .
j=0
k=j
j=0
k=0
k=j+1
(49)
Fix j such that 0 ≤ j ≤ ℓ − 1. Note that for k = j, we have
χt+1 (g[t−j,t−j+k] ) = χt+1 (g[t−j,t] ) = 1t+1 . Then we can change the limits
of the last double product in (49) as
ℓ−1
ℓ
Y Y
χt+1 (g[t−j,t−j+k] ) .
(50)
j=0
k=j
For j = ℓ, we have
ℓ
Y
χt+1 (g[t−j,t−j+k] ) = χt+1 (g[t−ℓ,t−ℓ+ℓ] )
k=j
= χt+1 (g[t−ℓ,t] )
= 1t+1 .
Then we can rewrite (50) as
ℓ
ℓ
Y
Y
χt+1 (g[t−j,t−j+k] ) .
j=0
k=j
With these changes (49) becomes
! ℓ ℓ
ℓ
ℓ
ℓ
Y
Y
Y
Y Y
χt (g[t−j,t−j+k] ) ,
r t+1
χt+1 (g[t−j,t−j+k] ) .
0,k
j=0
k=j
j=0
k=0
k=j
(51)
30
Note that the term
ℓ
Y
t+1
r0,k
k=0
!
involves generators from vector basis B t+1 , and the other terms involve generators from vector bases B t−j for j = 0, . . . , ℓ.
t
t+1 be the
First consider the case where rt+1 is r̂t+1 = (1t+1
0 , σr ). Let b̂
t+1
t+1
t+1
t+1
encoding of r̂ . Since r0 = 10 , then components r0,k are the identity
for 0 ≤ k ≤ ℓ. Then we can rewrite (51) as
ℓ
ℓ
ℓ
ℓ
Y
Y
Y
Y
(bt , b̂t+1 ) =
χt (g[t−j,t−j+k] ) ,
χt+1 (g[t−j,t−j+k] ) .
j=0
j=0
k=j
k=j
(52)
Note that (52) just involve generators from vector bases B t−j for j = 0, . . . , ℓ.
We can pair terms in (52) as
ℓ
ℓ
Y
Y
[χt (g[t−j,t−j+k] ), χt+1 (g[t−j,t−j+k] )] ,
(bt , b̂t+1 ) =
(53)
j=0
k=j
where the product multiplication in the inner square bracket is component
by component, i.e., [a, b] ∗ [c, d] = [a ∗ c, b ∗ d]. But note that
[χt (g[t−j,t−j+k] ), χt+1 (g[t−j,t−j+k] )]
is a valid trellis path segment of length 2, for 0 ≤ j ≤ ℓ, for j ≤ k ≤ ℓ.
Then (53) is a product of valid trellis path segments of length 2, and hence
by properties of the group trellis, (bt , b̂t+1 ) is a valid trellis path segment of
length 2. This means b̂t+1 ∈ F(bt ).
t
t+1 be the encoding
Now consider the case where rt+1 = (rt+1
0 , σr ). Let b
t+1
of r . Then using (51) and (52), we have
!
!
ℓ
Y
t+1
t+1
t t+1
t
.
(54)
r0,k b̂
(b , b ) = b ,
k=0
But
ℓ
Y
t+1
,
r0,k
k=0
is some branch b̃t+1 ∈ X0t+1 . Then
(bt , bt+1 ) = (bt , b̃t+1 b̂t+1 )
where b̃t+1 ∈ X0t+1 and b̂t+1 ∈ F(bt ). This means bt+1 ∈ F(bt ).
•
Notice that we can think of the encoder as an estimator. The encoding
t
t+1 where b̂t+1 ∈ F(bt ). Then
of r̂t+1 = (1t+1
0 , σr ) gives an initial estimate b̂
(t+1)
at time t + 1, we use new input r0
to find b̃t+1 ∈ X0t+1 to correct the
t+1
t+1
t+1
initial estimate b̂
so that b
= b̃ b̂t+1 and bt+1 ∈ F(bt ).
Theorem 30 Each tensor r ∈ R is encoded into a path b ∈ C, where b is
an encoding of r using (41) for each t ∈ Z.
31
t
Proof. By Theorem 28, if r ∈ R, then rt+1 = (rt+1
0 , σr ) for each t ∈ Z.
t+1
t
Then from Lemma 29, b
∈ F(b ) for each t ∈ Z. Then b is a path in C.
•
Lemma 31 Fix rt ∈ Rt . Let bt be an encoding of rt using (41). There is
a 1-1 correspondence between the set of pairs (rt , rt+1 ) ∈ Rt × Rt+1 where
t
t t+1 ) of length 2
rt+1 = (rt+1
0 , σr ) and the set of trellis path segments (b , b
in C. The 1-1 correspondence is given by (rt , rt+1 ) ↔ (bt , bt+1 ) if and only
if bt+1 is an encoding of rt+1 using (41).
Proof. Let Art be the set of pairs (rt , rt+1 ) ∈ Rt × Rt+1 where rt+1 =
t
t t+1 ) of length 2
(rt+1
0 , σr ), and let Abt be set of trellis path segments (b , b
in C. We know from Lemma 29 that (41) encodes (rt , rt+1 ) where rt+1 =
t
t t+1 ) of length 2 in C. Therefore (41)
(rt+1
0 , σr ) to a trellis path segment (b , b
maps Art into Abt . The map Art → Abt is 1-1 since if (rt , rt+1 ) encodes to
(bt , bt+1 ), and (rt , r̄t+1 ) encodes to (bt , bt+1 ), we must have rt+1 = r̄t+1 . We
t
know that bt+1 is uniquely determined by rt+1 , and specifically (rt+1
0 , σr ).
t+1
t+1 ∈ F(bt ). Therefore
There are |X0t+1 | possible rt+1
0 , and |X0 | possible b
the map from Art to Abt is 1-1 and onto.
•
Theorem 32 For any b ∈ C, there is some r ∈ R such that b is an
encoding of r using (41) for each t ∈ Z.
Proof. Fix b ∈ C. For each t ∈ Z, let (bt , bt+1 ) ∈ b. By Lemma 31, we
t
know there exists (rt , rt+1 ) ∈ Rt × Rt+1 such that rt+1 = (rt+1
0 , σr ) and
t
t+1
t
t+1
(b , b ) is an encoding of (r , r ) using (41). Then the sequence of pairs
. . . , (bt−1 , bt ), (bt , bt+1 ), (bt+1 , bt+2 ), (bt+2 , bt+3 ), . . .
in b is the encoding (41) of some sequence of pairs
. . . , (r̄t−1 , r̄t ), (rt , rt+1 ), (r̄t+1 , r̄t+2 ), (rt+2 , rt+3 ), . . . ,
where
t+1
t
t+2
), . . . .
= (r̄t+2
. . . , rt+1 = (rt+1
0 , σr̄
0 , σr ), r̄
But we must have
. . . , r̄t = rt , rt+1 = r̄t+1 , r̄t+2 = rt+2 , . . . ,
since both left and right term in each of the equalities encode to the same
branch in B t , e.g., r̄t and rt both encode to bt so r̄t = rt . Therefore the
sequence of pairs can be reduced to the sequence
. . . , rt , rt+1 , rt+2 , . . . ,
(55)
t
where rt+1 = (rt+1
0 , σr ) for each t ∈ Z, and b is an encoding of (55). But
this is a path in R by Theorem 28.
•
R is a tensor set that depends on choice of basis B. If basis B is fixed,
then R is fixed.
Theorem 33 Given basis B, R is fixed, and there is a 1-1 correspondence
R ↔ C given by rb ↔ b, where b is an encoding of rb using (41) for each
t ∈ Z.
32
Proof. By Theorem 30 and Theorem 32, the encoding (41) of r ∈ R is onto
C. We now show it is 1-1. Assume there exists r, r̄ ∈ R such that r 6= r̄ and
both encode to b using (41) for each t ∈ Z. But this is impossible, since if
rt 6= r̄t for any t ∈ Z, then bt 6= b̄t where bt is the encoding of rt and b̄t is
the encoding of r̄t .
•
33
6. THE SIGNATURE SEQUENCE, SIGNATURE GROUP,
AND GENERATOR GROUP
6.1 The groups (R, ∗) and (Rt , ∗t )
Let ḃ, b̈ ∈ C. Let ḃ ↔ ṙ and b̈ ↔ r̈ under the 1-1 correspondence
def
C ↔ R. Define an operation ∗ on tensor set R as ṙ ∗ r̈ = r̄ if r̄ ↔ ḃb̈ under
the 1-1 correspondence R ↔ C.
Theorem 34 R with operation ∗ is a group system (R, ∗).
Theorem 35 The group system (R, ∗) is isomorphic to group system C
under the 1-1 correspondence b ↔ rb in C ↔ R, C ≃ (R, ∗).
Proof. By definition of (R, ∗), the 1-1 correspondence C ↔ R gives an
isomorphism of C and (R, ∗).
•
We can now give a result on multiplication in C.
Theorem 36 The multiplication by b in bC corresponds to multiplication
by rb in rb ∗ R. This changes each generator vector r[t,t+k] in R to another,
at each time t and length k, 0 ≤ k ≤ ℓ.
Corollary 37 Let ḃ, b̈ ∈ C. Let ḃ ↔ ṙ and b̈ ↔ r̈ under the 1-1 correspondence C ↔ R. The multiplication by ḃ in ḃb̈ corresponds to multiplication
by ṙ in ṙ ∗ r̈. This changes each generator vector r̈[t,t+k] in r̈ to another, at
each time t and length k, 0 ≤ k ≤ ℓ.
We now consider component multiplication in C and (R, ∗). Component
multiplication in C is easy. For ḃ, b̈ ∈ C, product ḃb̈ is given by ḃt b̈t for
each t ∈ Z. We define a component multiplication on (R, ∗) that agrees with
the definition of operation ∗. Component multiplication in (R, ∗) gives more
insight into the structure of a group system than component multiplication
in C.
At time t, the component of R is Rt . We define a multiplication on Rt
as follows. Let rt ∈ Rt . Define a 1-1 correspondence Rt ↔ B t by rt ↔ bt
if branch bt is an encoding of rt using (41). To emphasize this, rt is a
matrix R[t,t] of the triangular form (32). The components of the marix are
t , for 0 ≤ k ≤ ℓ, for 0 ≤ j ≤ k. Each representative r t is
representatives rj,k
j,k
either nontrivial or the identity. We define the 1-1 correspondence Rt ↔ B t
by rt ↔ bt if
t
t
t
t
t
t
t
t
t t
· · · rj,j
· · · r0,2
r0,1
r0,0
,
· · · rj,k
· · · rj,ℓ
· · · rj+ℓ−k,ℓ
rℓ−1,ℓ rℓ−1,ℓ−1
bt = rℓ,ℓ
as in (42).
Let ṙt , r̈t ∈ Rt . Let branch ḃt be an encoding of ṙt using (41), and let
branch b̈t be an encoding of r̈t using (41). We define an operation ∗t on Rt
def
by ṙt ∗t r̈t = r̄t if r̄t ↔ ḃt b̈t under the 1-1 correspondence Rt ↔ B t . We
now show the operation ∗t defined on Rt is consistent with the operation ∗
defined on R.
34
Lemma 38 Let ṙt , r̈t ∈ Rt . We have
ṙt ∗t r̈t = (ṙ ∗ r̈)t
(56)
for each t ∈ Z.
Proof. The left hand side and right hand side of (56) are both elements
of Rt . We can prove the left hand side and right hand side are equal if
they both correspond to the same branch bt ∈ B t in the 1-1 correspondence
Rt ↔ B t .
Let ṙ, r̈ ∈ R. Let ṙ ↔ ḃ and r̈ ↔ b̈ under the 1-1 correspondence
R ↔ C. We have defined ṙ ∗ r̈ by the 1-1 correspondence ṙ ∗ r̈ ↔ ḃb̈ in
R ↔ C. But if ṙ ∗ r̈ ↔ ḃb̈ under the 1-1 correspondence R ↔ C, then from
Theorem 33, the time t component of ḃb̈ is an encoding of (ṙ ∗ r̈)t using
(41). But the time t component of ḃb̈ is just ḃt b̈t . Then
(ṙ ∗ r̈)t ↔ ḃt b̈t
in the 1-1 correspondence Rt ↔ B t .
Since ṙ ↔ ḃ and r̈ ↔ b̈, then from Theorem 33, ḃt is an encoding of ṙt
using (41), and b̈t is an encoding of r̈t using (41). Then from the definition
of ṙt ∗t r̈t , we know that
ṙt ∗t r̈t ↔ ḃt b̈t
in the 1-1 correspondence Rt ↔ B t . Then the left hand side and right
hand side of (56) are in 1-1 correspondence with the same branch in B t , and
therefore they must be the same.
•
Theorem 39 The set Rt with operation ∗t forms a group (Rt , ∗t ).
Theorem 40 (Rt , ∗t ) ≃ B t under the 1-1 correspondence Rt ↔ B t , that is
rtb ↔ bt where bt is an encoding of rtb using (41).
We say (Rt , ∗t ) is an induced group of B t .
We first consider general group systems which may vary with time.
Therefore the component basis B t may be different for each time t, and the
branch group B t may be different for each time t. Therefore the component
group (Rt , ∗t ) may be different for each time t.
Let ṙt , r̈t ∈ Rt . If ḃt is an encoding of ṙt using (41), then from (42),
t
t
t
t
t
t
t
t
t t
· · · ṙj,j
· · · ṙ0,2
ṙ0,1
ṙ0,0
.
· · · ṙj,k
· · · ṙj,ℓ
· · · ṙj+ℓ−k,ℓ
ṙℓ−1,ℓ ṙℓ−1,ℓ−1
ḃt = ṙℓ,ℓ
(57)
Similarly, if b̈t is an encoding of r̈t using (41), then from (42),
t t
t
t
t
t
t
t
t
t
b̈t = r̈ℓ,ℓ
r̈ℓ−1,ℓ r̈ℓ−1,ℓ−1
· · · r̈j+ℓ−k,ℓ
· · · r̈j,ℓ
· · · r̈j,k
· · · r̈j,j
· · · r̈0,2
r̈0,1
r̈0,0
.
(58)
Let r̄t = ṙt ∗t r̈t and assume b̄t is an encoding of r̄t using (41). Then by
definition of ṙt ∗t r̈t , we know
b̄t =ḃt b̈t
t
t
t
t
t
t
t
t
t t
· · · ṙj,j
· · · ṙ0,2
ṙ0,1
ṙ0,0
· · · ṙj,k
· · · ṙj,ℓ
· · · ṙj+ℓ−k,ℓ
ṙℓ−1,ℓ ṙℓ−1,ℓ−1
= ṙℓ,ℓ
t
t
t
t
t t
· · · r̈j,k
· · · r̈j,ℓ
· · · r̈j+ℓ−k,ℓ
r̈ℓ−1,ℓ r̈ℓ−1,ℓ−1
r̈ℓ,ℓ
35
t
· · · r̈j,j
t
t
t
· · · r̈0,2
r̈0,1
r̈0,0
(59)
. (60)
We now give a general result on products of the form (59)-(60). Consider
any finite group G with normal chain
1 = G−1 ⊂ G0 ⊂ G1 ⊂ · · · ⊂ Gm ⊂ · · · ⊂ Gℓ = G,
where each Gm is normal in G. Let {gm }ℓm=0 be any complete set of coset
representatives of the normal chain. Then any g ∈ G can be written using
a subset of this complete set as a coset representative chain
g = gℓ · · · gm · · · g1 g0 .
(61)
Lemma 41 Let f, g, h ∈ G and assume f g = h. Let g be as in (61), f =
′ · · · g ′ g ′ , and h = g ′′ · · · g ′′ · · · g ′′ g ′′ . Consider the product f g = h,
gℓ′ · · · gm
m
1 0
1 0
ℓ
or
′′
′
· · · g1′′ g0′′ ).
(gℓ′ · · · gm
· · · g1′ g0′ )(gℓ · · · gm · · · g1 g0 ) = (gℓ′′ · · · gm
′′ of h are solely
In the product f g = h, the upper coordinates gℓ′′ , . . . , gm
′
′
determined by the upper coordinates gℓ , . . . , gm of f and upper coordinates
gℓ , . . . , gm of g, for m such that 0 ≤ m ≤ ℓ.
Proof. Let m be such that 0 ≤ m ≤ ℓ. Since Gm−1 ✁ G, we know that
f Gm−1 gGm−1 = (f g)Gm−1
= hGm−1 .
′ )G
But f Gm−1 = (gℓ′ · · · gm
m−1 , gGm−1 = (gℓ · · · gm )Gm−1 , and hGm−1 =
′′
′′
(gℓ · · · gm )Gm−1 . Therfore
′′
′
)Gm−1 ,
)Gm−1 (gℓ · · · gm )Gm−1 = (gℓ′′ · · · gm
(gℓ′ · · · gm
or
′′
′
)Gm−1 .
)(gℓ · · · gm )Gm−1 = (gℓ′′ · · · gm
(gℓ′ · · · gm
But this last result means
′′
′′′
′
)(gm−1
· · · g1′′′ g0′′′ )
)(gℓ · · · gm ) = (gℓ′′ · · · gm
(gℓ′ · · · gm
′′′ , . . . , g ′′′ , g ′′′ in the complete set. Therefor some coset representatives gm−1
0
1
′′
′′ of h are solely determined by the upper
fore the upper coordinates gℓ , . . . , gm
′ of f and upper coordinates g , . . . , g of g.
coordinates gℓ′ , . . . , gm
•
m
ℓ
To apply this result to (59)-(60), we use the following notation. Note
that static matrix R[t,t] has the triangular form (32). We now introduce a
triangle notation to describe certain subsets of representatives in rt = R[t,t] .
For rt ∈ Rt , we let ▽j,k (rt ) be the representatives in rt specified by the
t and upper vertices r t and r t
triangle with lower vertex rj,k
j,ℓ
j+ℓ−k,ℓ . These
t
are the representatives rm,n where m, n satisfy k ≤ n ≤ ℓ and j ≤ m ≤ (j +
def
n − k). Let ▽j,k (Rt ) be the set of all possible triangles ▽j,k (rt ), ▽j,k (Rt ) =
{▽j,k (rt ) : rt ∈ Rt }. Define ▽
− j,k (rt ) to be the same as ▽j,k (rt ) except
t
missing representative rj,k , and likewise define ▽
− j,k (Rt ) to be the same
t }. We see that
as ▽j,k (Rt ) except missing the set of representatives {rj,k
▽0,0 (Rt ) = Rt and so the group (Rt , ∗t ) may also be written (▽0,0 (Rt ), ∗t ).
Now applying Lemma 41 to (59)-(60) gives the following.
36
Lemma 42 Fix time t. Fix k such that 0 ≤ k ≤ ℓ. Let ṙt , r̈t ∈ Rt . In the
product ṙt ∗t r̈t , the representatives in ▽k,k (ṙt ∗t r̈t ) are uniquely determined
by the representatives in ▽k,k (ṙt ) and ▽k,k (r̈t ).
Theorem 43 Let ṙ, r̈ ∈ R. Let r = ṙ ∗ r̈. Fix time t. Fix k such that
0 ≤ k ≤ ℓ. The generator vector r[t−k,t] in r is uniquely determined by the
representatives in ▽k,k (ṙt ) and ▽k,k (r̈t ). Also, the generator vector r[t−k,t]
in r is uniquely determined by the generator vectors with representatives in
▽k,k (ṙt ) and ▽k,k (r̈t ).
Proof. The generator vector r[t−k,t] is uniquely determined by representat . Now apply Lemma 42.
tive rk,k
•
Fix time t. Fix k such that 0 ≤ k ≤ ℓ. We know that ▽k,k (Rt ) =
{▽k,k (rt ) : rt ∈ Rt }. Let ṙt , r̈t ∈ Rt . Define an operation ⊛tk,k on ▽k,k (Rt )
by
def
(62)
▽k,k (ṙt ) ⊛tk,k ▽k,k (r̈t ) = ▽k,k (ṙt ∗t r̈t )
for ṙt , r̈t ∈ Rt .
Lemma 44 The operation ⊛tk,k on ▽k,k (Rt ) is well defined.
Proof. Let ṙt , r̈t ∈ Rt and r̀t , ŕt ∈ Rt . Let ▽k,k (r̀t ) = ▽k,k (ṙt ) and
▽k,k (ŕt ) = ▽k,k (r̈t ). To show the operation is well defined, we need to
•
show ▽k,k (r̀t ∗t ŕt ) = ▽k,k (ṙt ∗t r̈t ). But this is clear from Lemma 42.
Theorem 45 The set ▽k,k (Rt ) with operation ⊛tk,k forms a group
(▽k,k (Rt ), ⊛tk,k ).
Proof. First we show the operation ⊛tk,k is associative.
▽k,k (rt ), ▽k,k (ṙt ), ▽k,k (r̈t ) ∈ ▽k,k (Rt ). We need to show
Let
(▽k,k (rt )⊛tk,k ▽k,k (ṙt ))⊛tk,k ▽k,k (r̈t ) = ▽k,k (rt )⊛tk,k (▽k,k (ṙt )⊛tk,k ▽k,k (r̈t )).
But using (62) we have
(▽k,k (rt ) ⊛tk,k ▽k,k (ṙt )) ⊛tk,k ▽k,k (r̈t ) = ▽k,k (rt ∗t ṙt ) ⊛tk,k ▽k,k (r̈t )
= ▽k,k ((rt ∗t ṙt ) ∗t r̈t ),
and
▽k,k (rt ) ⊛tk,k (▽k,k (ṙt ) ⊛tk,k ▽k,k (r̈t )) = ▽k,k (rt ) ⊛tk,k ▽k,k (ṙt ∗t r̈t )
= ▽k,k (rt ∗t (ṙt ∗t r̈t )).
Therefore the operation ⊛tk,k is associative since the operation ∗t in group
(Rt , ∗t ) is associative.
Let 1t be the identity of (Rt , ∗t ). We show ▽k,k (1t ) is the identity of
(▽k,k (Rt ), ⊛tk,k ). But using (62) we have
▽k,k (1t ) ⊛tk,k ▽k,k (rt ) = ▽k,k (1t ∗t rt )
= ▽k,k (rt )
37
and
▽k,k (rt ) ⊛tk,k ▽k,k (1t ) = ▽k,k (rt ∗t 1t )
= ▽k,k (rt ).
Let r̄t be an element in (Rt , ∗t ) such that r̄t ∗t rt = 1t . We show ▽k,k (r̄t )
is an element in (▽k,k (Rt ), ⊛tk,k ) such that ▽k,k (r̄t )⊛tk,k ▽k,k (rt ) = ▽k,k (1t ).
But using (62) we have
▽k,k (r̄t ) ⊛tk,k ▽k,k (rt ) = ▽k,k (r̄t ∗t rt )
= ▽k,k (1t ).
Together these results show (▽k,k (Rt ), ⊛tk,k ) is a group.
•
In the last chapter of his Ph.D. thesis on group codes [5], Mitchell Trott
lists several open problems on group codes. Among them, he mentions that
“... surely there must be an orthogonal decomposition...” of group codes.
Ideally one would like the individual generator vectors themselves to form
an orthogonal basis. We do not find this to be true here. For the special
case in which the only nontrivial generator vectors are of length ℓ + 1, we
have an orthogonal decomposition using the groups (▽ℓ,ℓ (Rt ), ⊛tℓ,ℓ ) for each
t ∈ Z. However, in general for 0 ≤ k < ℓ, the groups (▽k,k (Rt ), ⊛tk,k ) show
a dependence among the generator vectors for each t ∈ Z and do not give
an orthogonal decomposition.
We can think of group (▽k,k (Rt ), ⊛tk,k ) in two different ways. In view of
(62), we can think of group (▽k,k (Rt ), ⊛tk,k ) as a multiplication in (Rt , ∗t )
of the upper coordinates ▽k,k (r̈t ) of r̈t with the upper coordinates ▽k,k (ṙt )
of ṙt to find the upper coordinates ▽k,k (ṙt ∗t r̈t ) of ṙt ∗t r̈t . In the second
interpretation, the group (▽k,k (Rt ), ⊛tk,k ) is devoid of any meaning as a
multiplication of representatives in (Rt , ∗t ). We think of (▽k,k (Rt ), ⊛tk,k )
as a permutation of the elements ▽k,k (rt ) in set ▽k,k (Rt ) with permutation
operation ⊛tk,k . Note that ▽k,k (rt ) is a subtriangle of ▽0,0 (rt ) in ▽0,0 (Rt ).
Thus (62) is a permutation of subtriangle ▽k,k (r̈t ) in set ▽k,k (Rt ) by subtriangle ▽k,k (ṙt ) in set ▽k,k (Rt ) to obtain subtriangle ▽k,k (ṙt ∗t r̈t ) in set
▽k,k (Rt ).
Note that group (▽0,0 (Rt ), ⊛t0,0 ) is the same as group (Rt , ∗t ), which is
isomorphic to B t . We can generalize this.
t
Theorem 46 For 0 ≤ k ≤ ℓ, (▽k,k (Rt ), ⊛tk,k ) ≃ B t /Xk−1
.
Proof. We know (▽0,0 (Rt ), ⊛t0,0 ) ≃ B t . Define a homomorphism B t →
t
t
. This gives a homomorphism
by assignment bt 7→ bt Xk−1
B t /Xk−1
t
t
t
t
t
. This
(▽0,0 (R ), ⊛0,0 ) → B /Xk−1 by assignment ▽0,0 (rt ) 7→ bt 7→ bt Xk−1
t
t
t
gives a 1-1 correspondence ▽k,k (R ) ↔ B /Xk−1 where each ▽k,k (rt ) is
t
. Then it can be seen this corresponassigned to a unique coset of B t /Xk−1
t
.
•
dence gives an isomorphism (▽k,k (Rt ), ⊛tk,k ) ≃ B t /Xk−1
Triangles ▽j,k (rt+j ) and ▽k,k (rt+k ) have congruent shapes, and we can
identify representatives in congruent positions in triangles ▽j,k (rt+j ) and
t+j
▽k,k (rt+k ). For example representative rj,k
of triangle ▽j,k (rt+j ) is in cont+k
of triangle ▽k,k (rt+k ).
gruent position with representative rk,k
38
Lemma 47 Fix j such that 0 ≤ j ≤ k. Fix r ∈ R. Representatives in
congruent positions of triangles ▽j,k (rt+j ) and ▽k,k (rt+k ) are from the same
generator vector.
Lemma 48 Fix time t. Fix k such that 0 ≤ k ≤ ℓ. Fix j such that
0 ≤ j ≤ k. Let ṙt+j , r̈t+j ∈ Rt+j . In the product ṙt+j ∗t+j r̈t+j , the representatives in ▽j,k (ṙt+j ∗t+j r̈t+j ) are uniquely determined by the representatives
in ▽j,k (ṙt+j ) and ▽j,k (r̈t+j ).
Proof. Fix time t. Fix k such that 0 ≤ k ≤ ℓ. Let ṙ, r̈ ∈ R. From Lemma
42, the representatives in ▽k,k (ṙt+k ∗t+k r̈t+k ) are uniquely determined by
the representatives in ▽k,k (ṙt+k ) and ▽k,k (r̈t+k ). This means that the generator vectors in ṙ ∗ r̈ with representatives in ▽k,k ((ṙ ∗ r̈)t+k ) are uniquely
determined by the generator vectors in ṙ with representatives in ▽k,k (ṙt+k )
and the generator vectors in r̈ with representatives in ▽k,k (r̈t+k ), where
we have used ▽k,k (ṙt+k ∗t+k r̈t+k ) = ▽k,k ((ṙ ∗ r̈)t+k ) by Lemma 38. But
for 0 ≤ j ≤ k, note that ▽j,k (ṙt+j ) and ▽k,k (ṙt+k ) are representatives
from the same set of generator vectors in ṙ. Similarly note that ▽j,k (r̈t+j )
and ▽k,k (r̈t+k ) are representatives from the same set of generator vectors
in r̈, and ▽j,k ((ṙ ∗ r̈)t+j ) and ▽k,k ((ṙ ∗ r̈)t+k ) are representatives from the
same set of generator vectors in ṙ ∗ r̈. Therefore the representatives in
▽j,k ((ṙ ∗ r̈)t+j ) are uniquely determined by the representatives in ▽j,k (ṙt+j )
and ▽j,k (r̈t+j ). Now use ▽j,k ((ṙ ∗ r̈)t+j ) = ▽j,k (ṙt+j ∗t+j r̈t+j ) by Lemma
38.
•
Theorem 49 Let ṙ, r̈ ∈ R. Let r = ṙ ∗ r̈. Fix time t. Fix k such that
0 ≤ k ≤ ℓ. The generator vector r[t,t+k] in r is uniquely determined by the
representatives in ▽0,k (ṙt ) and ▽0,k (r̈t ). Also, the generator vector r[t,t+k]
in r is uniquely determined by the generator vectors with representatives in
▽0,k (ṙt ) and ▽0,k (r̈t ).
Proof. The generator vector r[t,t+k] is uniquely determined by representat . Now apply Lemma 48.
•
tive r0,k
Fix time t. Fix j, k such that 0 ≤ k ≤ ℓ and 0 ≤ j ≤ k. We know that
▽j,k (Rt+j ) = {▽j,k (rt+j ) : rt+j ∈ Rt+j }. Let ṙt+j , r̈t+j ∈ Rt+j . Define an
t+j ) by
operation ⊛t+j
j,k on ▽j,k (R
def
t+j
▽j,k (ṙt+j ) ⊛t+j
) = ▽j,k (ṙt+j ∗t+j r̈t+j )
j,k ▽j,k (r̈
(63)
for ṙt+j , r̈t+j ∈ Rt+j .
t+j ) is well defined, for 0 ≤ j ≤
Lemma 50 The operation ⊛t+j
j,k on ▽j,k (R
k.
Proof. Let ṙt+j , r̈t+j ∈ Rt+j and r̀t+j , ŕt+j ∈ Rt+j . Let ▽j,k (r̀t+j ) =
▽j,k (ṙt+j ) and ▽j,k (ŕt+j ) = ▽j,k (r̈t+j ). To show the operation is well defined, we need to show ▽j,k (r̀t+j ∗t+j ŕt+j ) = ▽j,k (ṙt+j ∗t+j r̈t+j ). But this
is clear from Lemma 48.
•
Theorem 51 The set ▽j,k (Rt+j ) with operation ⊛t+j
j,k forms a group
(▽j,k (Rt+j ), ⊛t+j
j,k ), for 0 ≤ j ≤ k.
39
Proof. The proof is the same as for Theorem 45.
•
Once again, we can think of group (▽j,k (Rt+j ), ⊛t+j
j,k ) in two different
ways. In view of (63), we can think of group (▽j,k (Rt+j ), ⊛t+j
j,k ) as a multit+j
t+j
t+j
t+j
plication in (R , ∗ ). We multiply r̈
by ṙ
to obtain ṙt+j ∗t+j r̈t+j .
As product (59)-(60) illustrates, the coordinates ▽j,k (ṙt+j ∗t+j r̈t+j ) of
ṙt+j ∗t+j r̈t+j are a complicated function of the coordinates ▽j,k (r̈t+j ) of r̈t+j
and the coordinates ▽j,k (ṙt+j ) of ṙt+j . In addition, as Lemma 50 shows, for
any r̀t+j , ŕt+j ∈ Rt+j such that ▽j,k (r̀t+j ) = ▽j,k (ṙt+j ) and ▽j,k (ŕt+j ) =
▽j,k (r̈t+j ), we must have ▽j,k (r̀t+j ∗t+j ŕt+j ) = ▽j,k (ṙt+j ∗t+j r̈t+j ). These
two facts place severe restrictions on the structure of (Rt+j , ∗t+j ). For example we show in [19] that (Rt+j , ∗t+j ) must have a type of commutative
property. We do not consider this further here.
In the second interpretation, the group (▽j,k (Rt+j ), ⊛t+j
j,k ) is again
thought of as a permutation group. Thus (63) is a permutation of subtriangle ▽j,k (r̈t+j ) in set ▽j,k (Rt+j ) by subtriangle ▽j,k (ṙt+j ) in set ▽j,k (Rt+j )
to obtain subtriangle ▽j,k (ṙt+j ∗t+j r̈t+j ) in set ▽j,k (Rt+j ). This second interpretation will be shown to give a much more revealing explanation of the
group structure and henceforth we think of groups like (▽j,k (Rt+j ), ⊛t+j
j,k )
as permutation groups.
t+k ), ⊛t+k ).
We now relate the groups (▽j,k (Rt+j ), ⊛t+j
j,k ) and (▽k,k (R
k,k
Theorem 52 Fix k such that 0 ≤ k ≤ ℓ. Fix j such that 0 ≤ j ≤ k. Define
a 1-1 correspondence ▽j,k (Rt+j ) ↔ ▽k,k (Rt+k ) by the assignment
▽j,k (rt+j ) ↔ ▽k,k (rt+k )
(64)
for r ∈ R, where representatives in congruent positions are in 1-1 correspondence. Under this correspondence there is an isomorphism
t+k
(▽j,k (Rt+j ), ⊛t+j
), ⊛t+k
k,k ).
j,k ) ≃ (▽k,k (R
Proof. Fix r ∈ R. For j such that 0 ≤ j ≤ k, note that representatives in
congruent positions of ▽j,k (rt+j ) and ▽k,k (Rt+k ) are from the same generator vector.
•
We can restate Theorem 52 slightly in terms of shifts.
Corollary 53 Fix k such that 0 ≤ k ≤ ℓ. Fix j such that 0 ≤ j < k. Define
a 1-1 correspondence ▽j,k (Rt+j ) ↔ ▽j+1,k (Rt+j+1 ) by the assignment
▽j,k (rt+j ) ↔ ▽j+1,k (σrt+j ) = ▽j+1,k (rt+j+1 )
for r ∈ R, where representatives in congruent positions are in 1-1 correspondence. Under this correspondence there is a shift isomorphism
t+j+1
t+j+1
(▽j,k (Rt+j ), ⊛t+j
), ⊛j+1,k
).
j,k ) ≃ (▽j+1,k (R
Proof. Fix r ∈ R. For j such that 0 ≤ j < k, note that representatives
in congruent positions of ▽j,k (rt+j ) and ▽j+1,k (σrt+j ) are from the same
generator vector.
•
40
For k = 1, Corollary 53 gives ▽0,1 (rt ) ↔ ▽1,1 (σrt+1 ) and
t+1 ) can be viewed
(▽0,1 (Rt ), ⊛t0,1 ) ≃ (▽1,1 (Rt+1 ), ⊛t+1
1,1 ). Note that ▽1,1 (σr
as a shift of ▽0,1 (rt ), and (▽1,1 (Rt+1 ), ⊛t+1
1,1 ) can be viewed as a shift of
t+1 /X t+1 and
t
t
(▽0,1 (R ), ⊛0,1 ). It can be shown that (▽1,1 (Rt+1 ), ⊛t+1
1,1 ) ≃ B
0
(▽0,1 (Rt ), ⊛t0,1 ) ≃ B t /Y0t (see [19]). Therefore this result gives an explanation and restatement of the state group isomorphism B t /Y0t ≃ B t+1 /X0t+1
in a clearer form (see [19]).
A change of time variable in Theorem 51 gives the following result.
Theorem 54 Fix time t. Fix k such that 0 ≤ k ≤ ℓ. For j such that 0 ≤
j ≤ k, the set ▽j,k (Rt ) with operation ⊛tj,k forms a group (▽j,k (Rt ), ⊛tj,k ).
We now study a homomorphism related to the groups in Theorem 54.
Theorem 55 Fix time t. Fix k such that 0 ≤ k ≤ ℓ. For j such that
0 ≤ j ≤ k, the projection map ▽0,0 (Rt ) → ▽j,k (Rt ) given by the assignment ▽0,0 (rt ) 7→ ▽j,k (rt ) defines a homomorphism from (▽0,0 (Rt ), ⊛t0,0 ) to
(▽j,k (Rt ), ⊛tj,k ).
Proof. Let ṙt , r̈t ∈ Rt . Consider the projections p : ṙt 7→ ▽j,k (ṙt ) and
p : r̈t 7→ ▽j,k (r̈t ). By (63) and Lemma 50, we have
▽j,k (ṙt ∗t r̈t ) = ▽j,k (ṙt ) ⊛tj,k ▽j,k (r̈t ).
It follows that
p(ṙt ∗t r̈t ) = p(ṙt ) ⊛tj,k p(r̈t ).
Therefore p is a homomorphism from group (▽0,0 (Rt ), ⊛t0,0 ) to group
•
(▽j,k (Rt ), ⊛tj,k ).
Theorem 56 For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j ≤ k, for any
m, n such that ▽m,n (Rt ) ⊂ ▽j,k (Rt ), the projection map p : ▽j,k (Rt ) →
▽m,n (Rt ) given by the assignment p : ▽j,k (rt ) 7→ ▽m,n (rt ) defines a homomorphism from (▽j,k (Rt ), ⊛tj,k ) to (▽m,n (Rt ), ⊛tm,n ).
Proof. Define the projection map p : ▽j,k (Rt ) → ▽m,n (Rt ) by the assignment p : ▽j,k (rt ) 7→ ▽m,n (rt ). We have to show p is well defined. Let
r̀t , ŕt ∈ Rt . Suppose r̀t 6= ŕt and ▽j,k (r̀t ) = ▽j,k (ŕt ). Since r̀t and ŕt agree
on ▽j,k (Rt ) they must agree on ▽m,n (Rt ) because ▽m,n (Rt ) ⊂ ▽j,k (Rt ).
Therefore ▽m,n (r̀t ) = ▽m,n (ŕt ) and p is well defined.
Let ▽j,k (ṙt ), ▽j,k (r̈t ) ∈ ▽j,k (Rt ).
Consider the projections p :
t
t
t
▽j,k (ṙ ) 7→ ▽m,n (ṙ ) and p : ▽j,k (r̈ ) 7→ ▽m,n (r̈t ). We want to show
p(▽j,k (ṙt ) ⊛tj,k ▽j,k (r̈t )) = p(▽j,k (ṙt )) ⊛tm,n p(▽j,k (r̈t )).
But
p(▽j,k (ṙt ) ⊛tj,k ▽j,k (r̈t )) = p(▽j,k ((ṙ ∗ r̈)t ))
= ▽m,n ((ṙ ∗ r̈)t )
= ▽m,n (ṙt ) ⊛tm,n ▽m,n (r̈t )
= p(▽j,k (ṙt )) ⊛tm,n p(▽j,k (r̈t )).
•
41
Corollary 57 For each k such that 0 ≤ k < ℓ, for 0 ≤ j ≤ k, for m = j and
m = j + 1, the projection map ▽j,k (Rt ) → ▽m,k+1 (Rt ) given by the assignment ▽j,k (rt ) 7→ ▽m,k+1 (rt ) defines a homomorphism from (▽j,k (Rt ), ⊛tj,k )
to (▽m,k+1 (Rt ), ⊛tm,k+1 ).
Proof. Corollary 57 follows directly from Theorem 56.
•
We now study homomorphisms from (R, ∗).
Theorem 58 The projection map p : R → ▽j,k (Rt ) given by the assignment p : r 7→ ▽j,k (rt ) defines a homomorphism from (R, ∗) to
(▽j,k (Rt ), ⊛tj,k ).
Proof. Let ṙ, r̈ ∈ R. Consider the projections p : ṙ 7→ ▽j,k (ṙt ) and p : r̈ 7→
▽j,k (r̈t ). By (63) and Lemma 50, we have
▽j,k ((ṙ ∗ r̈)t ) = ▽j,k (ṙt ∗t r̈t )
= ▽j,k (ṙt ) ⊛tj,k ▽j,k (r̈t ).
It follows that
p(ṙ ∗ r̈) = p(ṙ) ⊛tj,k p(r̈).
Therefore p is a homomorphism from group (R, ∗) to group (▽j,k (Rt ), ⊛tj,k ).
•
We have the following generalization of this result.
Fix a sequence of indices i = . . . , (j, k), (j ′ , k′ ), . . . at times . . . , t, t′ , . . .,
where 0 ≤ k, k′ ≤ ℓ, 0 ≤ j ≤ k, 0 ≤ j ′ ≤ k′ , and where we may have j 6= j ′
and k 6= k′ . The sequence of times . . . , t, t′ , . . . may be the integers Z, or
may be a finite or infinite subset of Z. For each r ∈ R, consider the sequence
′
. . . , ▽j,k (rt ), ▽j ′ ,k′ (rt ), . . .
restricted to times . . . , t, t′ , . . .. Let R(i) be the set of all such sequences.
Define a projection map ∂ : R → R(i) by the assignment
′
∂ : r 7→ . . . , ▽j,k (rt ), ▽j ′ ,k′ (rt ), . . .
for r ∈ R.
Let ṙ, r̈ ∈ R(i), where
′
ṙ = . . . , ▽j,k (ṙt ), ▽j ′ ,k′ (ṙt ), . . . ,
and
′
ṙ = . . . , ▽j,k (ṙt ), ▽j ′ ,k′ (ṙt ), . . . .
Define an operation ∗R(i) on R(i) by
def
′
′
′
ṙ ∗R(i) r̈ = . . . , ▽j,k (ṙt ) ⊛tj,k ▽j,k (r̈t ), ▽j ′ ,k′ (ṙt ) ⊛tj ′ ,k′ ▽j ′ ,k′ (r̈t ), . . . .
Theorem 59 The set R(i) with operation ∗R(i) forms a group (R(i), ∗R(i) ).
42
Proof. We have to show that ṙ ∗R(i) r̈ ∈ R(i). Since ṙ ∈ R(i), there exists
ṙ ∈ R such that ∂(ṙ) = ṙ. Similarly, there exists r̈ ∈ R such that ∂(r̈) = r̈.
We know ṙ ∗ r̈ ∈ R and so ∂(ṙ ∗ r̈) ∈ R(i). By definition, the time t
component of ∂(ṙ ∗ r̈) is ▽j,k ((ṙ ∗ r̈)t ), and the time t′ component of ∂(ṙ ∗ r̈)
′
is ▽j ′ ,k′ ((ṙ ∗ r̈)t ). By (63) and Lemma 50, we know that
▽j,k ((ṙ ∗ r̈)t ) = ▽j,k (ṙt ∗t r̈t )
= ▽j,k (ṙt ) ⊛tj,k ▽j,k (r̈t ),
and
′
′
′
′
▽j ′ ,k′ ((ṙ ∗ r̈)t ) = ▽j ′ ,k′ (ṙt ∗t r̈t )
′
′
′
= ▽j ′ ,k′ (ṙt ) ⊛tj ′ ,k′ ▽j ′ ,k′ (r̈t ).
This means that ∂(ṙ ∗ r̈) = ṙ ∗R(i) r̈ and so ṙ ∗R(i) r̈ ∈ R(i). This last result
shows that
∂(ṙ ∗ r̈) = ∂(ṙ) ∗R(i) ∂(r̈).
(65)
The operation ∗R(i) is associative since the operation ⊛tj,k in group
(▽j,k (Rt ), ⊛tj,k ) is associative at each time t.
Let 1 be the identity of (R, ∗). We show ∂(1) is the identity of
(R(i), ∗R(i) ). Let r ∈ R(i) where
′
r = . . . , ▽j,k (rt ), ▽j ′ ,k′ (rt ), . . . .
Pick r such that ∂(r) = r. Using (65), we have
∂(1) ∗R(i) r = ∂(1) ∗R(i) ∂(r)
= ∂(1 ∗ r)
= ∂(r)
= r.
Let r ∈ R(i). We show r has an inverse. Pick r such that ∂(r) = r.
Using (65), we have ∂(r−1 ) ∗R(i) ∂(r) = ∂(r−1 ∗ r) = ∂(1). Therefore ∂(r−1 )
is an inverse of r. Together, these results show (R(i), ∗R(i) ) is a group.
•
Corollary 60 The set R(i) with operation ∗R(i) forms a group system
(R(i), ∗R(i) ).
Proof. We have (R(i), ∗R(i) ) is a group system since (R(i), ∗R(i) ) is a
group defined by a componentwise group addition at each time t (group
•
(▽j,k (Rt ), ⊛tj,k )).
Theorem 61 The projection map ∂ : R → R(i) given by the assignment
r 7→ ∂(r) defines a homomorphism from (R, ∗) to (R(i), ∗R(i) ).
Proof. Use (65).
•
Corollary 62 The projection map ∂ : C → R(i) given by the assignment
b 7→ rb 7→ ∂(rb ) defines a homomorphism from C to (R(i), ∗R(i) ), where
b ↔ rb is the 1-1 correspondence of the isomorphism C ≃ (R, ∗).
43
Consider a second sequence of indices i′ = . . . , (m, n), (m′ , n′ ), . . . at
times . . . , t, t′ , . . ., such that
′
′
. . . , ▽m,n (Rt ) ⊂ ▽j,k (Rt ), ▽m′ ,n′ (Rt ) ⊂ ▽j ′ ,k′ (Rt ), . . . .
For each r ∈ R, consider the sequence
′
. . . , ▽m,n (rt ), ▽m′ ,n′ (rt ), . . .
restricted to times . . . , t, t′ , . . .. Let R(i′ ) be the set of all such sequences.
Define a projection map ∂ ′ : R(i) → R(i′ ) by the assignment
′
′
∂ ′ : . . . , ▽j,k (rt ), ▽j ′ ,k′ (rt ), . . . 7→ . . . , ▽m,n (rt ), ▽m′ ,n′ (rt ), . . .
(66)
for r ∈ R.
Theorem 63 The projection map ∂ ′ : R(i) → R(i′ ) given by the assignment
(66) defines a homomorphism from (R(i), ∗R(i) ) to (R(i′ ), ∗R(i′ ) ).
Proof. The proof is similar to the proof of Theorem 56.
•
6.2 The signature sequence
A sequence of finite groups,
. . . , Gt , Gt+1 , . . . ,
(67)
is a group sequence. We give necessary and sufficient conditions under which
any group sequence (67) is a sequence of branch groups in a group trellis C
of an ℓ-controllable group system. From Sections 3 and 4, for Gt to be a
branch group in C, we know Gt must have a normal chain and a complete
set of coset representatives Rt of the normal chain that can be arranged into
the triangular form of the static matrix R[t,t] given in (32).
If C is ℓ-controllable, then there must be at least one generator of span ℓ
in R, g[t,t+ℓ] , for some t ∈ Z. This means that there must be static matrices
t+j
R[t+j,t+j] = Rt+j with a nontrivial representative rj,ℓ
, for j = 0, . . . , ℓ, for
some t ∈ Z. Additional requirements on Rt are given below. Aside from
these requirements, there are no restrictions on Rt . For example, Rt can
have rows of all identities. In addition, for m, n such that 0 ≤ m ≤ n ≤ k,
t } and {r t } in row k of Rt can be of different
the sets of representatives {rm,k
n,k
size.
Let ▽0,0 (Rt ) = Rt . Fix k such that 0 ≤ k ≤ ℓ. For j such that
t } in ▽ (Rt ) and the
0 ≤ j < k, consider the set of representatives {rj,k
0,0
t+1
t+1
set of representatives {rj+1,k } in ▽0,0 (R ). Suppose the two sets are
t }| = |{r t+1 }|. Then we can define a 1-1 correof the same size, |{rj,k
j+1,k
t+1
t
spondence τj,k : {rj,k } → {rj+1,k } between the two sets. Suppose we have
t }| = |{r t+1 }| for 0 ≤ k ≤ ℓ, for 0 ≤ j < k. Then we can define a 1-1
|{rj,k
j+1,k
correspondence τ between the sets of triangles ▽0,1 (Rt ) and ▽1,1 (Rt+1 ),
τ : ▽0,1 (Rt ) → ▽1,1 (Rt+1 ),
t } → {r t+1 } for 0 ≤ k ≤ ℓ, for 0 ≤ j < k.
such that τj,k : {rj,k
j+1,k
44
For j, k such that 0 ≤ k ≤ ℓ, 0 ≤ j ≤ k, let ▽j,k (Rt ) be the set of
all subtriangles ▽j,k (rt ) in ▽0,0 (Rt ), i.e., ▽j,k (Rt ) = {▽j,k (rt ) : rt ∈ Rt }.
Fix j, k such that 0 ≤ k ≤ ℓ, 0 ≤ j < k. Consider the set of triangles
▽j,k (Rt ) and ▽j+1,k (Rt+1 ). Let triangle ▽j,k (rt ) ∈ ▽j,k (Rt ) and triangle ▽j+1,k (r̂t+1 ) ∈ ▽j+1,k (Rt+1 ). Triangles in ▽j,k (Rt ) and ▽j+1,k (Rt+1 )
are congruent, and we can identify representatives in congruent positions in
t of tritriangles ▽j,k (rt ) and ▽j+1,k (r̂t+1 ). For example representative rj,k
t+1
angle ▽j,k (rt ) is in congruent position with representative r̂j+1,k
of triangle
t+1
t
t
▽j+1,k (r̂ ). Let triangle ▽j,k (r ) ∈ ▽j,k (R ). Using the 1-1 correspondence τ , find the unique triangle ▽j+1,k (r̄t+1 ) ∈ ▽j+1,k (Rt+1 ) such that
representatives in congruent positions are in 1-1 correspondence under τ .
t 7→ r̄ t+1 . We indicate this by saying
For example we have τj,k : rj,k
j+1,k
τ
▽j,k (rt )↔▽j+1,k (r̄t+1 ).
In this way, correspondence τ gives a 1-1 correspondence between the triangle elements in set ▽j,k (Rt ) and set ▽j+1,k (Rt+1 ), written as
τ
▽j,k (Rt )↔▽j+1,k (Rt+1 ).
We now show Gt induces a group (Rt , ⋆t ) where (Rt , ⋆t ) ≃ Gt , similar in
manner to that for B t and (Rt , ∗t ). There is a 1-1 correspondence Gt ↔ Rt
given by gt ↔ rt such that each gt ∈ Gt can be written using representatives
in coset representative chain rt ∈ Rt as
t
t
t
t
t
t
t
t
t
t
t
t t
· · · r0,2
r0,1
r0,0
.
· · · r1,1
r0,ℓ
· · · rj,j
· · · r2,2
r1,ℓ
· · · rj,k
· · · rj,ℓ
rℓ−1,ℓ rℓ−1,ℓ−1
gt = rℓ,ℓ
Let ġ t , g̈t ∈ Gt . Let ġ t ↔ ṙt and g̈ t ↔ r̈t under the 1-1 correspondence
def
Gt ↔ Rt . Define an operation ⋆t on Rt by ṙt ⋆t r̈t = r̄t if r̄t ↔ ġ t g̈t . It is
clear Rt with operation ⋆t is a group (Rt , ⋆t ) and (Rt , ⋆t ) ≃ Gt . We call
(Rt , ⋆t ) an induced group of Gt . Note that we have reused the same notation
Rt , ▽0,0 (Rt ), and ▽j,k (Rt ) as defined previously, but the operation ⋆t is not
necessarily the same as operation ∗t defined previously, and group (Rt , ⋆t )
is not necessarily the same as (Rt , ∗t ). Then like (67),
. . . , (Rt , ⋆t ), (Rt+1 , ⋆t+1 ), . . .
(68)
is also a group sequence. We say the sequence of induced groups (68) is an
induced sequence of (67).
Consider any group sequence (67) and any induced sequence (68) of (67).
We say the induced sequence (68) is a signature sequence if and only if it
satisfies the following three conditions, for each t ∈ Z:
(i)a. For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j ≤ k, there is an operation
t
⊙
⋆ j,k on set ▽j,k (Rt ) that is well defined by
def
▽j,k (ṙt )⊙
⋆ tj,k ▽j,k (r̈t ) = ▽j,k (ṙt ⋆t r̈t )
for ṙt , r̈t ∈ Rt , and this forms a group (▽j,k (Rt ),⊙
⋆ tj,k ). Note that
(▽0,0 (Rt ),⊙
⋆ t0,0 ) ≃ (Rt , ⋆t ).
(ii)a. There is a 1-1 correspondence τ as described above, such that
τ : ▽0,1 (Rt ) → ▽1,1 (Rt+1 ).
45
(iii)a. For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j < k, under the 1-1
correspondence of triangle elements ▽j,k (rt ) and ▽j+1,k (r̄t+1 ) in ▽j,k (Rt )
and ▽j+1,k (Rt+1 ), respectively, given by
τ
▽j,k (Rt )↔▽j+1,k (Rt+1 ),
there is an isomorphism
(▽j,k (Rt ),⊙
⋆ tj,k ) ≃ (▽j+1,k (Rt+1 ),⊙
⋆ t+1
j+1,k ).
Theorem 64 The sequence of branch groups . . . , B t , B t+1 , . . . of any ℓcontrollable group system C (or group shift or group code) has an induced
sequence . . . , (Rt , ∗t ), (Rt+1 , ∗t+1 ), . . . that is a signature sequence.
Proof. We know we can find a basis B for C. According to previous work in
this paper, basis B gives a tensor set R, and then we can find a group (R, ∗)
with component group (Rt , ∗t ) at each time t ∈ Z, where (Rt , ∗t ) ≃ B t . We
will show that
. . . , (Rt , ∗t ), (Rt+1 , ∗t+1 ), . . .
is a signature sequence. But condition (i)a follows from Theorem 54. We
can think of shift σ as a mapping σ : ▽0,1 (Rt ) → ▽1,1 (Rt+1 ). Then the 1-1
correspondence τ is given by σ. And condition (iii)a follows from Corollary
53. This proves the statement of the theorem.
•
Theorem 65 Given any group sequence . . . , Gt , Gt+1 , . . . that has a signature sequence . . . , (Rt , ⋆t ), (Rt+1 , ⋆t+1 ), . . ., we can construct an ℓcontrollable group system C, where . . . , Gt , Gt+1 , . . . is the sequence of branch
groups of C.
Proof. Form a Cartesian product of sets Rt of groups (Rt , ⋆t ) in the signature sequence . . . , (Rt , ⋆t ), (Rt+1 , ⋆t+1 ), . . .,
· · · × Rt × Rt+1 × · · · .
(69)
A sequence in (69) is of the form
. . . , rt , rt+1 , . . . ,
(70)
where rt ∈ Rt for each t ∈ Z. We now give conditions under which a
sequence (70) is a tensor r in a tensor set R.
From Theorem 28, we know a sequence (70) is a tensor r ∈ R if and
t+1
t
only if for each time t + 1, rt+1 = (rt+1
∈ Rt+1
is an input
0 , σr ) where r0
0
t
t
and σr is a shift of r . We now use conditions (i)a-(iii)a to show when rt+1
t
t t
t ⋆ t ). From conditions
is of the form (rt+1
0,0
0 , σr ). Let (R , ⋆ ) = (▽0,0 (R ),⊙
(i)a-(iii)a, we know there is an isomorphism
(▽0,1 (Rt ),⊙
⋆ t0,1 ) ≃ (▽1,1 (Rt+1 ),⊙
⋆ t+1
1,1 )
given by the 1-1 correspondence of triangles ▽0,1 (rt ) and ▽1,1 (r̄t+1 ),
τ
▽0,1 (rt )↔▽1,1 (r̄t+1 ),
46
where representatives in congruent positions satisfy the correspondence τ .
Therefore we define (70) to be a tensor r in tensor set R if and only if rt
and rt+1 satisfy the following for each t ∈ Z,
τ
▽0,1 (rt )↔▽1,1 (r̄t+1 ) = ▽1,1 (rt+1 ).
(71)
Then in this case rt+1 is of the form
t+1
rt+1 = (rt+1
))
0 , ▽1,1 (r
def
t
= (rt+1
0 , σ̄r ),
def
where rt+1
is an input and σ̄rt = ▽1,1 (rt+1 ) = ▽1,1 (r̄t+1 ) is thought of as
0
a shift of rt . Tensor set R is the set of all sequences (70) that satisfy (71)
for each t ∈ Z.
Let r1 , r2 ∈ R. We define an operation ⋆ on paths in R by r1 ⋆ r2 , where
def
r1 ⋆ r2 = . . . , rt1 ⋆t rt2 , rt+1
⋆t+1 rt+1
1
2 ,...
is defined by component by component multiplication rt1 ⋆t rt2 in group
(▽0,0 (Rt ),⊙
⋆ t0,0 ) = (Rt , ⋆t ) for each t ∈ Z. We now show R with operation ⋆ forms a group (R, ⋆). We have to show r1 ⋆ r2 ∈ R. If r1 ∈ R,
then
τ
t+1
(72)
▽0,1 (rt1 )↔▽1,1 (r̄t+1
1 ) = ▽1,1 (r1 )
by (71), for each t ∈ Z. If r2 ∈ R, then
τ
t+1
▽0,1 (rt2 )↔▽1,1 (r̄t+1
2 ) = ▽1,1 (r2 )
(73)
by (71), for each t ∈ Z. To show r1 ⋆ r2 ∈ R, it is sufficient to show
τ
t+1 t+1 t+1
⋆ r2 )
⋆t+1 r̄t+1
▽0,1 (rt1 ⋆t rt2 )↔▽1,1 (r̄t+1
2 ) = ▽1,1 (r1
1
(74)
by (71), for each t ∈ Z. But
▽0,1 (rt1 ⋆t rt2 ) = ▽0,1 (rt1 )⊙
⋆ t0,1 ▽0,1 (rt2 ).
From (72) and (73) and the isomorphism in condition (iii)a, we have
τ
t+1
▽0,1 (rt1 )⊙
⋆ t0,1 ▽0,1 (rt2 )↔▽1,1 (r̄t+1
⋆ t+1
1 )⊙
1,1 ▽1,1 (r̄2 ).
But
t+1
t+1 t+1 t+1
▽1,1 (r̄t+1
⋆ t+1
⋆ r̄2 ).
1 )⊙
1,1 ▽1,1 (r̄2 ) = ▽1,1 (r̄1
Therefore we have shown
τ
t+1
t+1
▽0,1 (rt1 ⋆t rt2 )↔▽1,1 (r̄t+1
⋆t+1 r̄t+1
⋆ t+1
1
2 ) = ▽1,1 (r̄1 )⊙
1,1 ▽1,1 (r̄2 ).
(75)
But using (72) and (73) we have
t+1
t+1
t+1
▽1,1 (r̄t+1
⋆ t+1
⋆ t+1
1 )⊙
1,1 ▽1,1 (r̄2 ) = ▽1,1 (r1 )⊙
1,1 ▽1,1 (r2 )
=
▽1,1 (rt+1
⋆t+1 rt+1
1
2 ).
Combining (75)-(77), we have shown (74).
47
(76)
(77)
•
Note that we need not start with the sequence of groups . . . , Gt , Gt+1 , . . ..
If we have any sequence of groups of the form (Rt , ⋆t ), then it is easy to find
the sequence of groups . . . , Gt , Gt+1 , . . . that induces it.
6.3 The signature group
We now specialize Section 6.2 to the time invariant case. We give necessary and sufficient conditions under which a finite group G is a branch
group in a group trellis C of an ℓ-controllable time invariant group system.
From Sections 3 and 4, for G to be a branch group in C, we know G must
have a normal chain and a complete set of coset representatives Rt of the
normal chain that can be arranged into the triangular form of the static
matrix R[t,t] given in (32).
Since C is time invariant, we may choose a constant basis B for C, and
henceforth we assume this to be the case. If basis B is constant, then Rt is
the same for all t ∈ Z, aside from time indexing. Since C is ℓ-controllable
and time invariant, there must be a generator of span ℓ in R, g[t,t+ℓ] , for each
t ∈ Z. This means that the top row of Rt contains nontrivial representatives
t = r (t−j)+j from generators g[t−j,t−j+ℓ] , for j = 0, . . . , ℓ. As in the general
rj,ℓ
j,ℓ
case, Rt can have rows of all identities. In contrast to the general case, we
will see that for m, n such that 0 ≤ m ≤ n ≤ k, the sets of representatives
t } and {r t } in row k of Rt must be the same size.
{rm,k
n,k
Let ▽0,0 (Rt ) = Rt . Fix k such that 0 ≤ k ≤ ℓ. For j such that
t } in ▽ (Rt ) and the
0 ≤ j < k, consider the set of representatives {rj,k
0,0
t
} in ▽0,0 (Rt ). Suppose the two sets are of
set of representatives {rj+1,k
t }| = |{r t
the same size, |{rj,k
j+1,k }|. Then we can define a 1-1 correspont
t
dence ξj,k : {rj,k } → {rj+1,k } between the two sets. Suppose we have
t }| = |{r t
|{rj,k
j+1,k }| for 0 ≤ k ≤ ℓ, for 0 ≤ j < k. Then we can define
a 1-1 correspondence ξ between the sets of triangles ▽0,1 (Rt ) and ▽1,1 (Rt ),
ξ : ▽0,1 (Rt ) → ▽1,1 (Rt ),
t } → {r t
such that ξj,k : {rj,k
j+1,k } for 0 ≤ k ≤ ℓ, for 0 ≤ j < k.
For j, k such that 0 ≤ k ≤ ℓ, 0 ≤ j ≤ k, let ▽j,k (Rt ) be the set of all subtriangles ▽j,k (rt ) in ▽0,0 (Rt ), i.e., ▽j,k (Rt ) = {▽j,k (rt ) : rt ∈ Rt }. Consider the set of triangles ▽j,k (Rt ) and ▽j+1,k (Rt ). Let triangle ▽j,k (rt ) ∈
▽j,k (Rt ) and triangle ▽j+1,k (r̂t ) ∈ ▽j+1,k (Rt ). Triangles in ▽j,k (Rt ) and
▽j+1,k (Rt ) are congruent, and we can identify representatives in congruent
positions in triangles ▽j,k (rt ) and ▽j+1,k (r̂t ). For example representative
t of triangle ▽ (rt ) is in congruent position with representative r̂ t
rj,k
j,k
k,k of
triangle ▽j+1,k (r̂t ). Let triangle ▽j,k (rt ) ∈ ▽j,k (Rt ). Using the 1-1 correspondence ξ, find the unique triangle ▽j+1,k (r̄t ) ∈ ▽j+1,k (Rt ) such that
representatives in congruent positions are in 1-1 correspondence under ξ.
t 7→ r̄ t
For example we have ξj,k : rj,k
j+1,k . We indicate this by saying
ξ
▽j,k (rt )↔▽j+1,k (r̄t ).
In this way, correspondence ξ gives a 1-1 correspondence between the triangle elements in set ▽j,k (Rt ) and set ▽j+1,k (Rt ), written as
ξ
▽j,k (Rt )↔▽j+1,k (Rt ).
48
We now show G induces a group (Rt , ⋆) where (Rt , ⋆) ≃ G, similar in
manner to that for B t and (Rt , ∗t ). There is a 1-1 correspondence G ↔ Rt
given by g ↔ rt such that each g ∈ G can be written using representatives
in coset representative chain rt ∈ Rt as
t t
t
t
t
t
t
t
t
t
t
t
t
g = rℓ,ℓ
rℓ−1,ℓ rℓ−1,ℓ−1
· · · rj,ℓ
· · · rj,k
· · · rj,j
· · · r2,2
r1,ℓ
· · · r1,1
r0,ℓ
· · · r0,2
r0,1
r0,0
.
Let ġ, g̈ ∈ G. Let ġ ↔ ṙt and g̈ ↔ r̈t under the 1-1 correspondence G ↔ Rt .
def
Define an operation ⋆ on Rt by ṙt ⋆r̈t = r̄t if r̄t ↔ ġg̈. It is clear Rt with
operation ⋆ is a group (Rt , ⋆) and (Rt , ⋆) ≃ G. We call (Rt , ⋆) an induced
group of G. Again, we have reused the same notation Rt , ▽0,0 (Rt ), and
▽j,k (Rt ) as defined previously, but the operation ⋆ is not necessarily the
same as operation ∗ defined previously, and group (Rt , ⋆) is not necessarily
the same as (Rt , ∗).
Consider any finite group G and any induced group (Rt , ⋆) of G. We
say the induced group (Rt , ⋆) is a signature group if and only if it satisfies
the following three conditions:
(i)b. For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j ≤ k, there is an operation
⊙
⋆ j,k on set ▽j,k (Rt ) that is well defined by
def
▽j,k (ṙt )⊙
⋆ j,k ▽j,k (r̈t ) = ▽j,k (ṙt ⋆ r̈t )
for ṙt , r̈t ∈ Rt , and this forms a group (▽j,k (Rt ),⊙
⋆ j,k ). Note that
t
t
(▽0,0 (R ),⊙
⋆ 0,0 ) ≃ (R , ⋆).
(ii)b. There is a 1-1 correspondence ξ as described above, such that
ξ : ▽0,1 (Rt ) → ▽1,1 (Rt ).
(iii)b. For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j < k, under the 1-1
correspondence of triangle elements ▽j,k (rt ) and ▽j+1,k (r̄t ) in ▽j,k (Rt ) and
▽j+1,k (Rt ), respectively, given by
ξ
▽j,k (Rt )↔▽j+1,k (Rt ),
there is an isomorphism
(▽j,k (Rt ),⊙
⋆ j,k ) ≃ (▽j+1,k (Rt ),⊙
⋆ j+1,k ).
Theorem 66 Any branch group B t of any ℓ-controllable time invariant
group system C (or group shift or group code) has an induced group (Rt , ∗)
that is a signature group.
Proof. By Theorem 64, any ℓ-controllable group system C has a
sequence of branch groups . . . , B t , B t+1 , . . . and signature sequence
. . . , (Rt , ∗t ), (Rt+1 , ∗t+1 ), . . .. Since C is time invariant, then B t = B t+1
for each t ∈ Z. Since C is time invariant, then we may choose a constant basis B. If B is constant, then (Rt , ∗t ) = (Rt+1 , ∗t+1 ) for each
t ∈ Z, and we may use operations ∗ and ⊛j,k in place of ∗t and ⊛tj,k for
any t ∈ Z. We now show (Rt , ∗) satisfies conditions (i)b-(iii)b of the definition of a signature group. By property (ii)a of a signature sequence,
there is a 1-1 correspondence τ such that τ : ▽0,1 (Rt ) → ▽1,1 (Rt+1 )
49
with assignment τ : ▽0,1 (rt ) 7→ ▽1,1 (r̄t+1 ). But we know by time invariance that (Rt , ∗) = (Rt+1 , ∗). Then we can define a 1-1 correspondence ξ : ▽0,1 (Rt ) → ▽1,1 (Rt ) by assignment ξ : ▽0,1 (rt ) 7→ ▽1,1 (r̄t ) if
τ : ▽0,1 (rt ) 7→ ▽1,1 (r̄t+1 ) and ▽1,1 (r̄t+1 ) = ▽1,1 (r̄t ). This verifies condition
(ii)b of the definition of a signature group.
By property (iii)a of a signature sequence, for each k such that 0 ≤ k ≤ ℓ,
for 0 ≤ j < k, under the 1-1 correspondence of triangle elements ▽j,k (rt )
and ▽j+1,k (r̄t+1 ) in ▽j,k (Rt ) and ▽j+1,k (Rt+1 ), respectively, given by
τ
▽j,k (Rt )↔▽j+1,k (Rt+1 ),
(78)
(▽j,k (Rt ), ⊛j,k ) ≃ (▽j+1,k (Rt+1 ), ⊛j+1,k ).
(79)
there is an isomorphism
Then using the definition of ξ with (78) and (79), under the 1-1 correspondence of triangle elements given by
ξ
▽j,k (Rt )↔▽j+1,k (Rt ),
there is an isomorphism
(▽j,k (Rt ), ⊛j,k ) ≃ (▽j+1,k (Rt ), ⊛j+1,k ).
(80)
This verifies condition (iii)b of the definition of a signature group. But
condition (i)b is satisfied trivially. Therefore (Rt , ∗) is an induced group of
B t having properties (i)b-(iii)b.
•
In a signature sequence, for m, n such that 0 ≤ m ≤ n ≤ k, the sets of
t } and {r t } in row k of Rt can be of different size. In
representatives {rm,k
n,k
a signature group, since (80) holds for 0 ≤ j < k, the sets of representatives
t } and {r t } in row k must be the same size.
{rm,k
n,k
Corollary 67 In Theorem 66, the signature group (Rt , ∗) contains a copy
of all the generator vectors of C.
Proof. As in Theorem 66, any ℓ-controllable group system C has a signature
sequence . . . , (Rt , ∗t ), (Rt+1 , ∗t+1 ), . . .. Since C is time invariant, then we
may choose a constant basis B. Then Rt = Rt+1 for each t ∈ Z. There is a
1-1 correspondence τ such that τ : ▽0,1 (Rt ) → ▽1,1 (Rt+1 ). We know that
τ can be viewed as a shift mapping σ : ▽0,1 (Rt ) → ▽1,1 (Rt+1 ).
t
in ▽0,0 (Rt );
Fix k such that 0 ≤ k ≤ ℓ. Consider representative r0,k
t , r t+1 , . . . , r t+k ).
t
is a representative in a generator vector r[t,t+k] = (r0,k
r0,k
1,k
k,k
Since τ is a shift σ, in the 1-1 correspondence τ : ▽0,1 (Rt ) → ▽1,1 (Rt+1 ),
t 7→ r t+1 . By definition of 1-1
component τ0,k gives the assignment τ0,k : r0,k
1,k
t
7→
correspondence ξ : ▽0,1 (Rt ) → ▽1,1 (Rt ), there is a mapping ξ0,k : r0,k
t+1
t+1
t
t
t
r̄1,k if τ0,k : r0,k 7→ r1,k and r1,k = r̄1,k . Therefore ξ0,k is essentially a
mapping of the first component of generator vector r[t,t+k] to the second.
t and the mappings ξ , for 0 ≤ j < k,
Continuing in this way shows that r0,k
j,k
give the same representatives as the components of r[t,t+k] . Since this same
t , for 0 ≤ k ≤ ℓ, ▽ (Rt ) contains a copy of all the
result holds for any r0,k
0,0
generator vectors of C.
•
50
Theorem 68 Given any finite group G that has a signature group (Rt , ⋆),
we can construct an ℓ-controllable time invariant group system C, where
. . . , G, G, . . . is the sequence of branch groups of C.
Proof. Consider the sequence of groups . . . , G, G, . . . and the sequence of
signature groups
. . . , (Rt , ⋆), (Rt+1 , ⋆), . . . ,
(81)
where (Rt+n , ⋆) = (Rt , ⋆) for each n ∈ Z. We show (81) satisfies conditions
(i)a-(iii)a of the definition of a signature sequence. By property (ii)b of a
signature group, there is a 1-1 correspondence ξ such that ξ : ▽0,1 (Rt ) →
▽1,1 (Rt ) with assignment ξ : ▽0,1 (rt ) 7→ ▽1,1 (r̄t ). But we know by time
invariance that (Rt , ⋆) = (Rt+1 , ⋆). Then we can define a 1-1 correspondence
τ : ▽0,1 (Rt ) → ▽1,1 (Rt+1 ) by assignment τ : ▽0,1 (rt ) 7→ ▽1,1 (r̄t+1 ) if
ξ : ▽0,1 (rt ) 7→ ▽1,1 (r̄t ) and ▽1,1 (r̄t ) = ▽1,1 (r̄t+1 ). This verifies condition
(ii)a of the definition of a signature sequence.
By property (iii)b of a signature group, for each k such that 0 ≤ k ≤ ℓ,
for 0 ≤ j < k, under the 1-1 correspondence of triangle elements ▽j,k (rt )
and ▽j+1,k (r̄t ) in ▽j,k (Rt ) and ▽j+1,k (Rt ), respectively, given by
ξ
▽j,k (Rt )↔▽j+1,k (Rt ),
(82)
(▽j,k (Rt ),⊙
⋆ j,k ) ≃ (▽j+1,k (Rt ),⊙
⋆ j+1,k ).
(83)
there is an isomorphism
Then using the definition of τ with (82) and (83), under the 1-1 correspondence of triangle elements given by
τ
▽j,k (Rt )↔▽j+1,k (Rt+1 ),
there is an isomorphism
(▽j,k (Rt ),⊙
⋆ j,k ) ≃ (▽j+1,k (Rt+1 ),⊙
⋆ j+1,k ).
This verifies condition (iii)a of the definition of a signature sequence. But
condition (i)a is satisfied trivially. Therefore . . . , G, G, . . . is a group sequence with a signature sequence (81). Then from Theorem 65, we can
construct an ℓ-controllable group system C, where . . . , G, G, . . . is the sequence of branch groups of C. But this means C is time invariant.
•
6.4 The block group system
In Section 6.3, we showed that when a general strongly controllable group
system C is a time invariant system Cs , the signature sequence of C reduces
to a single group, the signature group of Cs . In this section, we specialize the
general time varying ℓ-controllable group system C to a finite time interval
[t, t + ℓ + ν], ν ≥ 0. Let {bt+n } be the set of possible branches in C at
time t + n, n ∈ Z. We say an ℓ-controllable group system C has support on
[t, t + ℓ + ν] if {bt+n } is trivial for n 6= [0, ℓ + ν], and {bt+n } is nontrivial
for each n ∈ [0, ℓ + ν]. In this case we say C is a block group system Cb . A
block group system Cb is isomorphic to a set of finite sequences defined on
[t, t + ℓ + ν] with a componentwise group addition. This is usually referred
51
to in engineering literature and coding literature as a linear block code.
We only study the most important case of a block group system and linear
block code, the case ν = 0. The general case for ν > 0 is a straightforward
extension of the case for ν = 0.
The work of Forney and Trott shows that any block group system can
be decomposed into a set of generator vectors which can be arranged in
a group trellis [3]. In this section we describe the form of the signature
sequence . . . , (Rt , ⋆t ), (Rt+1 , ⋆t+1 ), . . . of Cb . In the next section, we find a
signature sequence isomorphic to . . . , (Rt , ⋆t ), (Rt+1 , ⋆t+1 ), . . .. We use this
to define a group which characterizes the additive structure of the group of
generator vectors in a block group system.
Assume an ℓ-controllable group system C has support on [t, t + ℓ]. From
the work of [3], we know the block group system can be decomposed into
generator vectors. The generator vectors of length ℓ must start at time t,
and be of the form r[t,t+ℓ] . By definition there must be at least one nontrivial generator vector of length ℓ. This means there must be a nontrivial
t+n
for 0 ≤ n ≤ ℓ. If there are generator vectors of length
representative rn,ℓ
ℓ − 1, they must start at times t and t + 1, and be of the form r[t,t+ℓ−1] and
r[t+1,t+ℓ] . In general, generator vectors of length k, 0 ≤ k ≤ ℓ, must start at
times t, t + 1, . . . , t + ℓ − k, and be of the form r[t+n,t+n+k] , for 0 ≤ n ≤ ℓ − k.
Knowing the decomposition into generator vectors, we can describe the form
of the signature sequence . . . , (Rt , ⋆t ), (Rt+1 , ⋆t+1 ), . . . of Cb . The matrices ▽0,0 (Rt+n ) are trivial (i.e., all entries are the identity) for n < 0 and
n > ℓ. Matrix ▽0,0 (Rt ) is trivial except for the first column rt0 ; then matrix ▽1,1 (Rt ) is trivial. Matrix ▽0,0 (Rt+1 ) is trivial except for the first two
t+1 ) is trivial. Column rt+1 is a
columns rt+1
and rt+1
0
1 ; then matrix ▽2,2 (R
1
t has been shifted out. Column rt+1 is
shift of column rt0 ; note that entry r0,0
0
t+1
a new input; note that entry r0,0
is trivial since there is only one generator
vector of length ℓ. In general, for 0 < n < ℓ, ▽0,0 (Rt+n ) is composed of two
trivial triangles and a nontrivial parallelogram. Triangle ▽0,ℓ−n+1 (Rt+n ) is
trivial and triangle ▽n+1,n+1 (Rt+n ) is trivial. The nontrivial entries form a
t+n
t+n t+n
,
parallelogram shape whose corners are the representatives r0,0
, r0,ℓ−n , rn,ℓ
t+n
t+n
t+n
and rn,n . In other words, columns r0 , . . . , rn are nontrivial except for
t+n
column entries in triangle ▽0,ℓ−n+1 (Rt+n ), and columns rt+n
are
n+1 , . . . , rℓ
t+n
trivial. Thus, for j such that 0 ≤ j ≤ n, column rj contains nontrivial
t+n
entries rj,k
for k such that j ≤ k ≤ ℓ − n + j. Lastly, matrix ▽0,0 (Rt+ℓ ) is
t+ℓ
trivial except for the last diagonal, the representatives rj,j
for 0 ≤ j ≤ ℓ.
Lemma 69 Any ℓ-controllable group system C has support on [t, t + ℓ] if
and only if the matrices ▽0,0 (Rt+n ) are trivial for n < 0 and n > ℓ, and
there is at least one nontrivial generator vector r[t,t+ℓ] . This means there
t+n
is a nontrivial representative rn,ℓ
for 0 ≤ n ≤ ℓ. For 0 ≤ n ≤ ℓ, any
nontrivial entries in ▽0,0 (Rt+n ) lie in a parallelogram shape whose corners
t+n
t+n
t+n
t+n (for n = 0 and n = ℓ,
, and rn,n
, rn,ℓ
are the representatives r0,0
, r0,ℓ−n
the parallelogram is trivially a line).
Using Lemma 69, we can restate Theorems 64 and 65 of Section 6.2 for
a block group system as follows.
52
Theorem 70 The sequence of branch groups . . . , B t , B t+1 , . . . of any ℓcontrollable group system C with support on [t, t+ℓ] has an induced sequence
. . . , (Rt , ∗t ), (Rt+1 , ∗t+1 ), . . . that is a signature sequence, where ▽0,0 (Rt+n ),
−∞ < n < ∞, has the form given in Lemma 69.
Theorem 71 Given any group sequence . . . , Gt , Gt+1 , . . . that has a signature sequence . . . , (Rt , ⋆t ), (Rt+1 , ⋆t+1 ), . . ., where ▽0,0 (Rt+n ), −∞ < n <
∞, has the form given in Lemma 69, we can construct an ℓ-controllable
group system C with support on [t, t + ℓ], where . . . , Gt , Gt+1 , . . . is the sequence of branch groups of C.
6.5 Isomorphic signature sequence and homomorphism to generators
Previously we have constructed a tensor set R using generator vectors
for 0 ≤ k ≤ ℓ, for t ∈ Z. A generator vector r[t,t+k] in r ∈ R is
t , . . . , r t+k ). The first component r t (or alternatively the last
given by (r0,k
0,k
k,k
t+k
) specifies the generator vector uniquely. We now replace
component rk,k
generator vector r[t,t+k] in r ∈ R with a shift vector u[t,t+k] , where u[t,t+k] is
t , . . . , r t ), where the first component r t of r[t,t+k] is repeated
given by (r0,k
0,k
0,k
k + 1 times. Under the assignment r[t,t+k] 7→ u[t,t+k] for 0 ≤ k ≤ ℓ, for each
t ∈ Z, a tensor r ∈ R becomes a tensor u, r 7→ u. Let U be the set of
tensors u obtained this way. There is a 1-1 correspondence R ↔ U given by
the assignment r 7→ u. The operation ∗ in (R, ∗) determines an operation ◦
on U by u̇ ◦ ü = ū if and only if ṙ ∗ r̈ = r̄, where ṙ 7→ u̇, r̈ 7→ ü, and r̄ 7→ ū.
Then (U , ◦) forms a group and (U , ◦) ≃ (R, ∗).
Under this replacement a component rt given in (32) becomes a component ut as shown in (84).
r[t,t+k]
t
r0,ℓ
t
r0,ℓ−1
..
.
t−1
r0,ℓ
t−1
r0,ℓ−1
..
.
···
···
..
.
···
···
..
.
t−j
r0,ℓ
t−j
r0,ℓ−1
..
.
···
···
..
.
···
···
..
.
t
r0,k
..
.
t−1
r0,k
..
.
···
..
.
···
..
.
t−j
r0,k
..
.
···
..
.
t−k
r0,k
···
..
.
···
..
.
···
..
.
···
t−j
r0,j
t
r0,2
t
r0,1
t
r0,0
t−1
r0,2
t−1
r0,1
t−2
r0,2
···
···
..
.
t−ℓ
t−ℓ+1
r0,ℓ
r0,ℓ
t−ℓ+1
r0,ℓ−1
(84)
For each k, 0 ≤ k ≤ ℓ, and each j, 0 ≤ j ≤ k, the representa(t−j)+j
t−j
t
in rt in (32) is replaced by representative r0,k
in
tive rj,k
= rj,k
(t−j)+j
t−j
ut in (84), where r0,k
and rj,k
are from the same generator vector
(t−j)+k
(t−j)+j
(t−j)
). Then in place of matrix Rt
· · · rk,k
= (r0,k · · · rj,k
t , we have a matrix Ut of coset representatives
of coset representatives rj,k
t−j
r0,k
. Then we have a 1-1 correspondence Rt ↔ Ut with assignment rt 7→ ut .
Matrix Ut has the same form as matrix Rt , and so in the following we reuse
r[t−j,t−j+k]
the same notation for Ut as for Rt . For example we have ▽0,0 (Ut ) = Ut .
53
For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j ≤ k, note that ▽j,k (ut ) are the
t−j
representatives in ut specified by the triangle with lower vertex r0,k
and
t−(j+ℓ−k)
t−j
.
upper vertices r0,ℓ
and r0,ℓ
t
t
t
t
Let ṙ 7→ u̇ and r̈ 7→ ü . Define an operation ◦t on set Ut by
def
u̇t ◦t üt = ūt
if
ṙt ⋆t r̈t = r̄t
where r̄t 7→ ūt in Rt ↔ Ut . This gives a group (Ut , ◦t ) which is isomorphic
to (Rt , ⋆t ). We say (Ut , ◦t ) is an induced group of G if (Rt , ∗t ) is an induced
group of G. Then
. . . , (Ut , ◦t ), (Ut+1 , ◦t+1 ), . . .
(85)
is a group sequence. We call (85) an induced sequence of (67). It is clear
the induced sequence (85) is an isomorphic version of the induced sequence
(68).
Consider any group sequence (67) and any induced sequence (85) of (67).
We say the induced sequence (85) is a signature sequence of (67) if and only
if it satisfies the following two conditions, for each t ∈ Z:
(i)c. For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j ≤ k, there is an operation
⊚tj,k on set ▽j,k (Ut ) that is well defined by
def
▽j,k (u̇t ) ⊚tj,k ▽j,k (üt ) = ▽j,k (u̇t ◦t üt )
for u̇t , üt ∈ Ut , and this forms a group (▽j,k (Ut ), ⊚tj,k ). Note that
(▽0,0 (Ut ), ⊚t0,0 ) ≃ (Ut , ◦t ).
(ii)c. For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j < k, under the 1-1
correspondence of triangle elements ▽j,k (ut ) and ▽j+1,k (ūt+1 ) in ▽j,k (Ut )
and ▽j+1,k (Ut+1 ), respectively, given by
▽j,k (ut ) = ▽j+1,k (ūt+1 ),
there is an isomorphism
(▽j,k (Ut ), ⊚tj,k ) ≃ (▽j+1,k (Ut+1 ), ⊚t+1
j+1,k ).
By using Ut in place of Rt , we have eliminated the need for correspondence τ in the definition of signature sequence (68). We have the following
restatement of Theorems 64 and 65 for signature sequence (85).
Theorem 72 The sequence of branch groups . . . , B t , B t+1 , . . .
of any ℓ-controllable group system C has an induced sequence
. . . , (Ut , ◦t ), (Ut+1 , ◦t+1 ), . . . that is a signature sequence.
Theorem 73 Given any group sequence . . . , Gt , Gt+1 , . . . that has a signature sequence . . . , (Ut , ◦t ), (Ut+1 , ◦t+1 ), . . ., we can construct an ℓcontrollable group system C, where . . . , Gt , Gt+1 , . . . is the sequence of branch
groups of C.
We have the following analogs of Theorems 55 and 56 and Corollary 57.
54
Theorem 74 Fix time t. Fix k such that 0 ≤ k ≤ ℓ. For j such that
0 ≤ j ≤ k, the projection map ▽0,0 (Ut ) → ▽j,k (Ut ) given by the assignment ▽0,0 (ut ) 7→ ▽j,k (ut ) defines a homomorphism from (▽0,0 (Ut ), ⊚t0,0 )
to (▽j,k (Ut ), ⊚tj,k ).
Theorem 75 For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j ≤ k, for any
m, n such that ▽m,n (Ut ) ⊂ ▽j,k (Ut ), the projection map p : ▽j,k (Ut ) →
▽m,n (Ut ) given by the assignment p : ▽j,k (ut ) 7→ ▽m,n (ut ) defines a homomorphism from (▽j,k (Ut ), ⊚tj,k ) to (▽m,n (Ut ), ⊚tm,n ).
Corollary 76 For each k such that 0 ≤ k < ℓ, for 0 ≤ j ≤ k, for
m = j and m = j + 1, the projection map ▽j,k (Ut ) → ▽m,k+1 (Ut ) given
by the assignment ▽j,k (ut ) 7→ ▽m,k+1 (ut ) defines a homomorphism from
(▽j,k (Ut ), ⊚tj,k ) to (▽m,k+1 (Ut ), ⊚tm,k+1 ).
Fix a sequence of indices i = . . . , (j, k), (j ′ , k′ ), . . . at times . . . , t, t′ , . . .,
where 0 ≤ k, k′ ≤ ℓ, 0 ≤ j ≤ k, 0 ≤ j ′ ≤ k′ , and where we may have j 6= j ′
and k 6= k′ . The sequence of times . . . , t, t′ , . . . may be the integers Z, or
may be a finite or infinite subset of Z. For each u ∈ U , consider the sequence
′
. . . , ▽j,k (ut ), ▽j ′ ,k′ (ut ), . . .
restricted to times . . . , t, t′ , . . .. Let U(i) be the set of all such sequences.
Define a projection map ∂ : U → U(i) by the assignment
′
∂ : u 7→ . . . , ▽j,k (ut ), ▽j ′ ,k′ (ut ), . . .
for u ∈ U .
Let u̇, ü ∈ U(i), where
′
u̇ = . . . , ▽j,k (u̇t ), ▽j ′ ,k′ (u̇t ), . . . ,
and
′
u̇ = . . . , ▽j,k (u̇t ), ▽j ′ ,k′ (u̇t ), . . . .
Define an operation ◦U(i) on U(i) by
def
′
′
′
u̇ ◦U(i) ü = . . . , ▽j,k (u̇t ) ⊛tj,k ▽j,k (üt ), ▽j ′ ,k′ (u̇t ) ⊛tj ′ ,k′ ▽j ′ ,k′ (üt ), . . . .
Theorem 77 The set U(i) with operation ◦U(i) forms a group (U(i), ◦U(i) ).
Proof. The proof is analogous to the proof of Theorem 59 or use the isomorphism (R, ∗) ≃ (U , ◦).
•
Corollary 78 The set U(i) with operation ◦U(i) forms a group system
(U(i), ◦U(i) ).
Proof. We have (U(i), ◦U(i) ) is a group system since (U(i), ◦U(i) ) is a
group defined by a componentwise group addition at each time t (group
•
(▽j,k (Ut ), ⊛tj,k )).
Lemma 79 The projection map ∂ : U → U(i) given by the assignment
u 7→ ∂(u) defines a homomorphism from (U , ◦) to (U(i), ◦U(i) ).
55
Proof. Use the result analogous to (65).
•
Theorem 80 The projection map κ : C → U(i) given by the assignment
b 7→ r 7→ u 7→ ∂(u) defines a homomorphism from C to (U(i), ◦U(i) ), where
b ↔ r is the 1-1 correspondence of the isomorphism C ≃ (R, ∗) and r ↔ u
is the 1-1 correspondence of the isomorphism (R, ∗) ≃ (U , ◦).
Consider a second sequence of indices i′ = . . . , (m, n), (m′ , n′ ), . . . at
times . . . , t, t′ , . . ., such that
′
′
. . . , ▽m,n (Ut ) ⊂ ▽j,k (Ut ), ▽m′ ,n′ (Ut ) ⊂ ▽j ′ ,k′ (Ut ), . . . .
For each u ∈ U , consider the sequence
′
. . . , ▽m,n (ut ), ▽m′ ,n′ (ut ), . . .
restricted to times . . . , t, t′ , . . .. Let U(i′ ) be the set of all such sequences.
Define a projection map ∂ ′ : U(i) → U(i′ ) by the assignment
′
′
∂ ′ : . . . , ▽j,k (ut ), ▽j ′ ,k′ (ut ), . . . 7→ . . . , ▽m,n (ut ), ▽m′ ,n′ (ut ), . . .
(86)
for u ∈ U .
Theorem 81 The projection map ∂ ′ : U(i) → U(i′ ) given by the assignment
(86) defines a homomorphism from (U(i), ◦U(i) ) to (U(i′ ), ◦U(i′ ) ).
Proof. The proof is similar to the proof of Theorem 56.
•
We now view the group of tensors (U , ◦) in a slightly different way. Note
that rt and rt+1 have no representatives in common. However note that
representatives ▽0,1 (ut ) in ut and representatives ▽1,1 (ut+1 ) in ut+1 are
the same. We can slide ut+1 past ut by one column so the representatives
in ▽0,1 (ut ) overlap. Fix u ∈ U . Let PU be the sliding operation on u that
slides ut+1 past ut by one column for each t ∈ Z. Under operation PU ,
sequence u becomes a sequence us , PU : u 7→ us . Let Us be the set of all
us formed this way. Then there is a 1-1 correspondence U ↔ Us given by
the mapping PU : U → Us formed by the assignment PU : u 7→ us . Each
t , . . . , r t ) in u appears
t
in a shift vector u[t,t+k] = (r0,k
representative r0,k
0,k
k + 1 times in u but only once in us . Therefore we have collapsed shift
t in u .
vector u[t,t+k] in u to a single point r0,k
s
Lemma 82 The set Us is a description of the shift vectors in U and generator vectors in R. If u ↔ r under the 1-1 correspondence U ↔ R, and
t
in
P : u 7→ us under the 1-1 correspondence U ↔ Us , then each point r0,k
us is a shift vector in u ∈ U and a generator vector in r ∈ R.
Note that us can also be thought of as the sequence of columns (or
inputs)
t−1
t
. . . , ut+1
0 , u0 , u0 , . . .
in reverse time order. Let uts be the portion of us shown in (84). Then
def
uts = ut if P : u 7→ us and ▽j,k (uts ) = ▽j,k (ut ). Define Uts = Ut and
def
▽0,0 (Uts ) = Uts .
56
Let PU : u 7→ us and consider ∂(u) ∈ U(i). Under the sliding operation
PU , the finite or infinite sequence of nonoverlapping triangles of the form
▽j,k (ut ) in u becomes the finite or infinite sequence of triangles of the form
▽j,k (uts ) in us . Triangles ▽j,k (uts ) in us may overlap if they share the same
representative; triangles that overlap form polygon shapes in us . Define
∂(us ) to be this sequence of isolated triangles and polygon shapes formed
by overlapping triangles. Let Us (i) be the set of all ∂(us ) formed this way.
Let PU be a mapping PU : U(i) → Us (i) given by the assignment PU : ∂(u) 7→
∂(us ). These definitions are shown in the commutative diagram Figure
2. Since there is a 1-1 correspondence U ↔ Us given by the assignment
PU : u 7→ us , there is a 1-1 correspondence U(i) ↔ Us (i) given by the
assignment PU : ∂(u) 7→ ∂(us ).
PU
✲ us
u
∂
❄
∂(u)
❄
PU
✲ ∂(us )
Figure 2: Definition of ∂(us ) and PU .
Theorem 83 The group (U(i), ◦U(i) ) acts on U(i) and therefore acts on Us (i)
by the 1-1 correspondence U(i) ↔ Us (i).
Proof. It is clear the group (U(i), ◦U(i) ) acts on U(i). The 1-1 correspondence U(i) ↔ Us (i) shows (U(i), ◦U(i) ) acts on Us (i).
•
t , . . . , r t ).
Let u[t,t+k] be a shift vector in u ∈ U , where u[t,t+k] = (r0,k
0,k
In the group action of (U , ◦) on U , let û ∈ U take u to ū. Then shift
vector u[t,t+k] in u is taken to shift vector ū[t,t+k] in ū, where ū[t,t+k] =
t , . . . , r̄ t ). In the group action of (U(i), ◦
(r̄0,k
U(i) ) on U(i), ∂(û) takes ∂(u)
0,k
t
t
to ∂(ū). Then any components r0,k in ∂(u) are taken to components r̄0,k
t . Then in the
in ∂(ū). In Us , there is one and only one representative r0,k
t
group action of (U(i), ◦U(i) ) on Us (i), ∂(û) takes singular representative r0,k
t in ∂(ū ). Therefore we can view the
in ∂(us ) to singular representative r̄0,k
s
action of group (U(i), ◦U(i) ) on Us (i) as equivalently an action on the set
t
in
of generator vectors which are indicated by singular representatives r0,k
Us (i). As well, there is equivalently an action on the set of generators which
t in U (i).
are indicated by representatives r0,k
s
Let SX be the group of all permutations of set X. An action of group G
on a set X induces a homomorphism π from G to SX , π : G → SX , given by
π : g 7→ gX where gX ∈ SX is a permutation of X associated with g. From
Theorem 83, the group (U(i), ◦U(i) ) acts on U(i). Let u ∈ U(i). Then there is
a homomorphism π from (U(i), ◦U(i) ) to SU(i) , π : (U(i), ◦U(i) ) → SU(i) , given
by assignment π : u 7→ uU(i), where uU(i) is a permutation of U(i) associated
with u, given by coset u ◦U(i) U(i) in group (U(i), ◦U(i) ). Under the bijection
U(i) ↔ Us (i), the permutation uU(i) of U(i) becomes a permutation uUs (i)
of Us (i). Define a mapping πs from (U(i), ◦U(i) ) to SUs (i) by assignment
57
πs : u 7→ uUs (i). Since π is a homomorphism, πs is a homomorphism. We
say πs is an extension of π using the 1-1 correspondence U(i) ↔ Us (i). This
gives the following.
Theorem 84 There is a homomorphism πs from (U(i), ◦U(i) ) to SUs (i) ,
πs : (U(i), ◦U(i) ) → SUs (i) , which is an extension of homomorphism π :
(U(i), ◦U(i) ) → SU(i) , where homomorphism π is given by assignment π :
u 7→ u ◦U(i) U(i).
In other words, there is a permutation group on the set of generator vectors indicated by Us (i), and a homomorphism from (U(i), ◦U(i) ) to this permutation group. Each generator vector corresponds to the generator that
contains it. Then equivalently, there is a permutation group on the set of
generators indicated by Us (i), and a homomorphism from (U(i), ◦U(i) ) to this
permutation group.
Theorem 85 There is a homomorphism from C to SUs (i) .
Proof. Combine Theorems 80 and 84.
•
We can use the preceding results to study the additive structure of a
block group system Cb in terms of its generator vectors. To do this it just
t−1
t
remains to describe the sequence of columns . . . , ut+1
0 , u0 , u0 , . . . in any
us ∈ Us of a block group system. For a block group system with support on
[t, t + ℓ], we know all columns are trivial except for columns ut+n
0 , 0 ≤ n ≤ ℓ.
For these columns, we have the following result using Lemma 69.
Lemma 86 For 0 < n ≤ ℓ, columns ut+n
of a block group system have the
0
form
T
t+n
t+n
t+n
t+n
t+n
·
·
·
r
r
r
·
·
·
1
1
,
ut+n
=
0,1
0,0
0,ℓ−n
0,ℓ−n+1
0,ℓ
0
t+n
where representatives 1t+n
0,ℓ , . . . , 10,ℓ−n+1 are the identity, and the remaining
representatives may be trivial or nontrivial. For n = 0, there must be a colt ; the remaining representatives
umn ut0 having a nontrivial representative r0,ℓ
may be trivial or nontrivial.
The representatives which may be nontrivial in any us ∈ Us are shown in
(87); all representatives outside this triangle shape or pyramid shape must
be trivial.
···
···
t+ℓ
r0,0
t+ℓ−2
···
r0,2
···
t+ℓ−1
t+ℓ−2
r0,1
r0,1
···
t+ℓ−1
t+ℓ−2
r0,0
r0,0
···
···
..
.
t+ℓ−k
r0,k
..
.
···
..
.
···
..
.
···
..
.
···
···
···
t+ℓ−k
r0,2
···
t+ℓ−k
r0,1
···
t+ℓ−k
r0,0
···
···
···
···
58
···
t+1
r0,ℓ−1
..
.
t
r0,ℓ
···
t
r0,ℓ−1 · · ·
..
.
t+1
r0,k
..
.
t
r0,k
..
.
···
t+1
r0,2
t+1
r0,1
t+1
r0,0
t
r0,2
t
r0,1
t
r0,0
···
···
···
(87)
Example 1. In this example, we study the block group system for the case
ℓ = 2. Then (87) becomes (88). We identify all nontrivial groups of the form
(U(i), ◦U(i) ). First we find the simplest groups by considering sequences i of
only one term. These are the groups (▽j,k (Ut ), ⊚tj,k ) on triangles ▽j,k (Ut )
for some (j, k) and t ∈ Z. For this case we know that ▽j,k (Ut ) = ▽j,k (Uts ).
From (88), the only sets of the form ▽j,k (Uts ) in the block group system
t
Cb with nontrivial elements are ▽0,2 (Uts ), ▽0,1 (Uts ), ▽0,1 (Ut+1
s ), ▽0,0 (Us ),
t+2
▽0,0 (Ut+1
s ), and ▽0,0 (Us ). These are listed in the first column of Table 1. These are the same as the sets ▽0,2 (Ut ), ▽0,1 (Ut ), ▽0,1 (Ut+1 ),
▽0,0 (Ut ), ▽0,0 (Ut+1 ), and ▽0,0 (Ut+2 ), respectively, shown in the third column of Table 1. These sets are associated with the groups (▽0,2 (Ut ), ⊚t0,2 ),
t+1 ), ⊚t+1 ),
t
t
(▽0,1 (Ut ), ⊚t0,1 ), (▽0,1 (Ut+1 ), ⊚t+1
0,0
0,1 ), (▽0,0 (U ), ⊚0,0 ), (▽0,0 (U
t+2
t+2
and (▽0,0 (U ), ⊚0,0 ), respectively, shown in the fourth column. Using
(88), we see that each group has the group elements shown in the second
column of Table 1. For example, set ▽0,1 (Ut ) or ▽0,1 (Uts ) is the set of
t , r t , 1t−1 ) in the Cartesian product of the sets of representatriples (r0,1
0,2 0,2
t } × {r t } × 1t−1 , where 1t−1 is the trivial set just containing the
tives {r0,1
0,2
0,2
0,2
identity. The groups in Table 1 are shown in (89), stacked according to the
representatives in (88) that define them.
···
···
t+2
r0,0
t
···
r0,2
···
t+1
t
r0,1 r0,1 · · ·
t+1
t
···
r0,0
r0,0
Set Us (i)
Representatives in set Us (i)
▽0,2 (Uts )
▽0,1 (Uts )
▽0,1 (Ut+1
s )
▽0,0 (Uts )
▽0,0 (Ut+1
s )
▽0,0 (Ut+2
s )
t }
{r0,2
t } × {r t } × 1t−1
{r0,1
0,2
0,2
t+1
t
{r0,1
} × 1t+1
×
{r
0,2 }
0,2
t } × {r t } × {r t } × 1t−1 × 1t−1 × 1t−2
{r0,0
0,2
0,1
0,1
0,2
0,2
t+1
t+1
t } × {r t } × 1t−1
{r0,0
} × {r0,1
} × 1t+1
×
{r
0,1
0,2
0,2
0,2
t+2
t+2
t+2
t+1
t+1
t
{r0,0 } × 10,1 × 10,2 × {r0,1 } × 10,2 × {r0,2 }
(88)
Set U(i)
t+2
H0,0
t
H0,2
t
H0,1
t
H0,0
▽0,1 (Ut )
t = (▽ (Ut ), ⊚t )
H0,1
0,1
0,1
▽0,1 (Ut+1 )
t+1
H0,1
= (▽0,1 (Ut+1 ), ⊚t+1
0,1 )
▽0,0 (Ut )
t = (▽ (Ut ), ⊚t )
H0,0
0,0
0,0
▽0,0 (Ut+1 )
t+1
H0,0
= (▽0,0 (Ut+1 ), ⊚t+1
0,0 )
▽0,0 (Ut+2 )
t+2
H0,0
= (▽0,0 (Ut+2 ), ⊚t+2
0,0 )
(89)
We now find the remaining sets in (88) that involve index sequences i of
two or more terms. These are polygon shapes formed from multiple triangles
of the form ▽j,k (Ut ) in U that overlap in Us . The triangles in U that form
nontrivial overlapping triangles in Us are shown in the second column of
Table 2. For example ▽0,1 (Ut+1 ) and ▽0,1 (Ut ) overlap in Us and form the
t+1
t−1
t
t
polygon shape whose representatives are {r0,1
}×1t+1
0,2 ×{r0,1 }×{r0,2 }×10,2 ,
a union of the representatives from ▽0,1 (Ut+1 ) and ▽0,1 (Ut ). The nontrivial
representatives in the polygon shape are shown in the first column of Table
2. The group (U(i), ◦U(i) ) that corresponds to U(i) is shown in the third
column of Table 2. For example, the group (U(i), ◦U(i) ) corresponding to
59
def
t = (▽ (Ut ), ⊚t )
H0,2
0,2
0,2
▽0,2
Table 1: Some sets and group actions of a block group system Cb for ℓ = 2.
t+1
H0,1
t+1
H0,0
Group (U(i), ◦U(i) )
(Ut )
def
def
def
def
def
t+1
t , and we say H t+1 ∪ H t
U(i) = ▽0,1 (Ut+1 ), ▽0,1 (Ut ) is denoted H0,1
∪ H0,1
0,1
0,1
t+1
t
is the union of H0,1 and H0,1 . Tables 1 and 2 give all nontrivial groups
(U(i), ◦U(i) ) up to isomorphism. The groups in Tables 1 and 2 may have
t+2
t+1
t
other descriptions which are isomorphic. For example, H0,0
∪ H0,0
∪ H0,1
t+2
t+1
is isomorphic to group H0,0 ∪ H0,0 listed in Table 2.
Nontrivial representatives in polygon set Us (i)
t+1
t } × {r t }
{r0,1
} × {r0,1
0,2
t+1
t } × {r t } × {r t }
{r0,1 } × {r0,0
0,2
0,1
t+1
t+1
t } × {r t } × {r t }
{r0,0
} × {r0,1
} × {r0,0
0,1
0,2
t+2
t+1
t } × {r t }
{r0,0
} × {r0,1
} × {r0,1
0,2
t+2
t+1
t } × {r t } × {r t }
{r0,0
} × {r0,1
} × {r0,0
0,2
0,1
t+2
t+1
t+1
t } × {r t }
{r0,0
} × {r0,0
} × {r0,1
} × {r0,1
0,2
t+2
t+1
t+1
t } × {r t } × {r t }
{r0,0
} × {r0,0
} × {r0,1
} × {r0,0
0,1
0,2
Set U(i)
▽0,1 (Ut+1 ), ▽0,1 (Ut )
▽0,1 (Ut+1 ), ▽0,0 (Ut )
▽0,0 (Ut+1 ), ▽0,0 (Ut )
▽0,0 (Ut+2 ), ▽0,1 (Ut )
▽0,0 (Ut+2 ), ▽0,0 (Ut )
▽0,0 (Ut+2 ), ▽0,0 (Ut+1 )
▽0,0 (Ut+2 ), ▽0,0 (Ut+1 ), ▽0,0 (Ut )
Table 2: Remaining sets and group actions of a block group system Cb for
ℓ = 2.
We have the following results.
t is a subgroup of every group in (89). Alternatively, in the top
1. H0,2
t } is contained in every group.
row of (88), the set {r0,2
t+n
2. In the bottom row of (88), each set {r0,0
}, for 0 ≤ n ≤ 2, is only
t+n
contained in one group in (89), H0,0 .
3. The groups in (89) contain other groups in (89) as subgroups. The
t+n
subgroups in of any group H0,m
in (89) can be found from triangle shape
t in (89) has subgroups H t , H t in
▽0,m (Ut+n ). For example group H0,0
0,2
0,1
(89).
4. If any group in (89) is a subgroup of another group, the projection
map from the later to the former defines a homomorphism. This is an
application of Theorem 75.
5. The intersection of any two groups in (89) is another group in (89),
which is a subgroup of both groups.
6. The union of any two groups in (89) at different times is a group.
t+1
t is a group. This is an application of Theorem 77.
For example, H0,1
∪ H0,1
These groups are shown in Table 2.
7. The union of any three groups in (89) at different times is a group. For
t+2
t+1
t is a group. This is an application of Theorem
example, H0,0
∪ H0,1
∪ H0,0
77. This group is shown in Table 2. Any other union of three groups at
t+2
different times is isomorphic to H0,0
or a union of two groups shown in
Table 2.
8. Any collection of groups in (89) is a group. For each time t + n,
t+n
n ∈ {0, 1, 2}, there is some largest group H0,m
which contains any other
t+n
groups H0,m′ at the same time as subgroups. Then the union of the largest
groups at different times is a group. Therefore any collection of groups is
isomorphic to a group listed in Table 1 or Table 2.
9. Each representative in a set in Table 1 determines a generator vector.
Therefore we can use (89) to study the additive structure of a block group
t is
system in terms of its generator vectors. It follows that the group H0,2
isomorphic to the permutation group of all generator vectors of length 3
t+1
t
(including the identity generator vector of length 3). The group H0,1
∪ H0,1
60
Group (U(i), ◦U(i) )
t+1
t
H0,1
∪ H0,1
t+1
t
H0,1 ∪ H0,0
t+1
t
H0,0
∪ H0,0
t+2
t
H0,0
∪ H0,1
t+2
t
H0,0 ∪ H0,0
t+2
t+1
H0,0 ∪ H0,0
t+2
t+1
t
H0,0
∪ H0,0
∪ H0,0
is isomorphic to the permutation group of all generator vectors of length 2
t+2
t+1
t is isomorphic to the permutation group
and 3. The group H0,0
∪H0,1
∪H0,0
of all generator vectors of length 1, 2, and 3. This last permutation group
is isomorphic to the block group system. These results are an application of
Theorem 84.
10. There is a homomorphism from (U , ◦) to the union of any groups.
This is an application of Lemma 79. There is a homomorphism from the
block group system Cb to the union of any groups (Theorem 80).
11. If any union of subgroups is a subgroup of another union of groups,
the projection map from the later to the former defines a homomorphism.
This is an application of Theorem 81.
The results in this example are not surprising; they are merely an extension to a block group system of what happens for a single group. For a
single group, the set of highest order coset representatives forms a permutation group. Next the set of highest order and second highest order coset
representatives forms a permutation group, and so on. In the block group
system, the set of longest generator vectors forms a permutation group.
There are two sets of second longest generator vectors, those starting at
time t and those starting at time t + 1. The set starting at time t forms a
permutation group with the set of longest generator vectors. Also the set
starting at time t + 1 forms a permutation group with the set of longest
generator vectors. The union of these two groups also forms a group, the
permutation group of longest and all second longest generator vectors. The
intersection of these two groups is also a group, the permutation group of
longest generator vectors. And so on. The groups in Table 1, Table 2, and
(89) give a simple and direct explanation of the group structure of a block
group system in terms of its generator vectors.
•
Example 2. The theory of group systems developed in [3] can be applied
to block codes. We study the binary (8, 4, 4) extended Hamming code or
first order Reed-Muller code used as Example 4 in [3] (p. 1503 and 1510).
This code is linear over Z2 × Z2 . The trellis diagram for this code is shown
in Figure 8 of [3]. The granule representatives or generators of this code are
as follows:
[Γ[t,t+3] ] = {(00, 00, 00, 00), (10, 10, 10, 10)},
(90)
[Γ[t,t+1] ] = {(00, 00, 00, 00), (11, 11, 00, 00)},
(91)
[Γ
[t+1,t+2]
] = {(00, 00, 00, 00), (00, 11, 11, 00)},
(92)
[Γ
[t+2,t+3]
] = {(00, 00, 00, 00), (00, 00, 11, 11)}.
(93)
For this code ℓ = 3. Then (87) becomes (94). There are generator
vectors of length 4 in (90), and generator vectors of length 2 in (91)-(93),
but no generator vectors of length 3 or 1. This is the reason for the two
rows of identities in (94). We identify all nontrivial groups of the form
(U(i), ◦U(i) ) in the same way as for Example 1. First we find the simplest
groups by considering sequences i of only one term. These are the groups
t def
= (▽j,k (Ut ), ⊚tj,k ) on triangles ▽j,k (Ut ) for some (j, k) and t ∈ Z.
Hj,k
An example is shown in Table 3. These simplest groups are shown in (95),
stacked according to the representatives in (94) that define them.
61
···
···
1t+3
0,0
···
t+2
r0,1
1t+2
0,0
···
1t+1
0,2
t+1
r0,1
1t+1
0,0
Set Us (i)
Representatives in set Us (i)
▽0,1 (Ut+1
s )
t+1
{r0,1
}
×
1t+1
0,2
×
1t+1
0,3
t
r0,3
1t0,2
t
r0,1
1t0,0
···
···
···
···
t }×
× 1t0,2 × {r0,3
1t−1
0,3
(94)
Set U(i)
Group (U(i), ◦U(i) )
▽0,1 (Ut+1 )
t+1
H0,1
= (▽0,1 (Ut+1 ), ⊚t+1
0,1 )
def
Table 3: A set and group action of the binary (8, 4, 4) extended Hamming
code.
t+3
H0,0
t+2
H0,1
t+2
H0,0
t+1
H0,2
t+1
H0,1
t+1
H0,0
t
H0,3
t
H0,2
t
H0,1
t
H0,0
(95)
We give an example to show the application of Theorems 80, 84, and 85
to the (8, 4, 4) code. Let i be the finite sequence (j, k) = (0, 1) at time t + 1.
For u ∈ U , this is just the finite sequence ▽0,1 (ut+1 ). Then
U(i) = {▽0,1 (ut+1 ) : ut+1 ∈ Ut+1 } = ▽0,1 (Ut+1 )
and
t+1
t+1
Us (i) = {▽0,1 (ut+1
∈ Ut+1
s ) : us
s } = ▽0,1 (Us ).
This example is shown in Table 3. Theorem 84 shows there is a homomorphism from group (U(i), ◦U(i) ) to a permutation group on Us (i). This
t+1 ) =
is self evident since (U(i), ◦U(i) ) = (▽0,1 (Ut+1 ), ⊚t+1
0,1 ) and ▽0,1 (U
▽0,1 (Ut+1
s ) = Us (i).
Theorem 80 shows there is a homomorphism from C to (U(i), ◦U(i) ). For
our example (U(i), ◦U(i) ) = (▽0,1 (Ut+1 ), ⊚t+1
0,1 ). We now verify Theorem 80
by showing there is a homomorphism from the (8, 4, 4) extended Hamming
code to (▽0,1 (Ut+1 ), ⊚t+1
0,1 ). Theorem 80 shows the homomorphism is given
by the map κ : C → U(i) using the assignment b 7→ r 7→ u 7→ ∂(u),
where b ↔ r is the 1-1 correspondence of the isomorphism C ≃ (R, ∗) and
r ↔ u is the 1-1 correspondence of the isomorphism (R, ∗) ≃ (U , ◦). For this
example, we have ∂(u) = ▽0,1 (ut+1 ), and therefore the assignment becomes
b 7→ r 7→ u 7→ ▽0,1 (ut+1 ).
To determine this assignment, we start with ▽0,1 (ut+1 ) and work left
to b. As shown in Table 3, the only representatives in ▽0,1 (Ut+1 ) that
correspond to nontrivial generators in the (8, 4, 4) code are from the sets
t+1
t }. For each choice of representatives from these sets, we
{r0,1
} and {r0,3
obtain all elements ▽0,1 (ut+1 ) ∈ ▽0,1 (Ut+1 ) as shown in column four of
t , . . . , r t ).
t
is in the shift vector u[t,t+3] = (r0,3
Table 4. Representative r0,3
0,3
t+1
t+1
t+1
[t+1,t+2]
= (r0,1 , r0,1 ). For this
Representative r0,1 is in the shift vector u
code we have r[t,t+3] = u[t,t+3] and r[t+1,t+2] = u[t+1,t+2] . These generator
vectors are shown in the third column of Table 4. The generator vectors in
column three correspond to the generators in column two. Any codeword
b in the (8, 4, 4) code can be decomposed by selecting one generator from
each of the four transversals [Γ[t,t+3] ], [Γ[t,t+1] ], [Γ[t+1,t+2] ], and [Γ[t+2,t+3] ]
62
shown in (90)-(93). For each choice of two generators in the second column
of Table 4, there are four codewords that have these generators in their
decomposition. This gives the assignment from the first column of Table
4 to the second. This completes the chain of assignments b 7→ r 7→ u 7→
▽0,1 (ut+1 ).
It remains to verify this assignment gives a homomorphism from Cb
to (▽0,1 (Ut+1 ), ⊚t+1
0,1 ). To verify this, it is sufficient to show the assignment from the first to second column of Table 4 gives a homomorphism
from Cb to a permutation group on the four pairs of generators shown in
the second column. To show this, it is sufficient to show the quartet of
4-tuples in the first column forms a quotient group with the first quartet {(00, 00, 00, 00), (00, 00, 11, 11), (11, 11, 00, 00), (11, 11, 11, 11)} as a normal subgroup. But this is easy to show using the linearity of the (8, 4, 4)
code.
We can verify Theorem 85 by concatenating the examples for Theorems
80 and 84.
Codeword b
(00, 00, 00, 00)
(00, 00, 11, 11)
(11, 11, 00, 00)
(11, 11, 11, 11)
(00, 11, 11, 00)
(00, 11, 00, 11)
(11, 00, 11, 00)
(11, 00, 00, 11)
(10, 10, 10, 10)
(10, 10, 01, 01)
(01, 01, 10, 10)
(01, 01, 01, 01)
(10, 01, 01, 10)
(10, 01, 10, 01)
(01, 10, 01, 10)
(01, 10, 10, 01)
t+1 t+1 t+1 t
t , 1t−1 )
(r0,1
, 10,2 , 10,3 , 10,2 , r0,3
0,3
t+1
= ▽0,1 (u ) ∈ ▽0,1 (Ut+1 )
(00, 00, 00, 00, 00, 00)
Selected generators in
decomposition of b
(00, 00, 00, 00), (00, 00, 00, 00)
Generator vectors
(00, 00, 00, 00), (00, 11, 11, 00)
r[t,t+3] = (00, 00, 00, 00)
r[t+1,t+2] = (11, 11)
(11, 00, 00, 00, 00, 00)
(10, 10, 10, 10), (00, 00, 00, 00)
r[t,t+3] = (10, 10, 10, 10)
r[t+1,t+2] = (00, 00)
(00, 00, 00, 00, 10, 00)
(10, 10, 10, 10), (00, 11, 11, 00)
r[t,t+3] = (10, 10, 10, 10)
r[t+1,t+2] = (11, 11)
(11, 00, 00, 00, 10, 00)
r[t,t+3] = (00, 00, 00, 00)
r[t+1,t+2] = (00, 00)
Table 4: Mapping of codewords of the binary (8, 4, 4) extended Hamming
code to ▽0,1 (Ut+1 ) for the homomorphism Cb → (▽0,1 (Ut+1 ), ⊚t+1
0,1 ).
t , H t+1 , H t+2 , H t+3 in the bottom row of
Note that the groups H0,0
0,0
0,0
0,0
(95) are the same as the groups (▽0,0 (Ut ), ⊚t0,0 ), (▽0,0 (Ut+1 ), ⊚t+1
0,0 ),
t+2
t+3
t+2
t+3
(▽0,0 (U ), ⊚0,0 ), (▽0,0 (U ), ⊚0,0 ), respectively, which are isomort+2 ),⊙
phic to the groups (▽0,0 (Rt ),⊙
⋆ t0,0 ), (▽0,0 (Rt+1 ),⊙
⋆ t+1
⋆ t+2
0,0 ), (▽0,0 (R
0,0 ),
t+3
t+3
t
(▽0,0 (R ),⊙
⋆ 0,0 ), respectively.
The sets ▽0,0 (R ), ▽0,0 (Rt+1 ),
t+2
t+3
▽0,0 (R ), ▽0,0 (R ) can be used to construct the trellis diagram of the
t , H t+1 , H t+2 , H t+3 give a de(8, 4, 4) code [19]. Therefore the groups H0,0
0,0
0,0
0,0
scription of the group structure of the trellis in Figure 8 of [3] at each time
epoch t, t + 1, t + 2, t + 3.
•
6.6 Construction
63
In this section, we construct an ℓ-controllable group system which may
vary in time, a time invariant ℓ-controllable group system, and an ℓcontrollable block group system. To construct an ℓ-controllable group system, it is sufficient to construct a signature sequence. For a time invariant
group system, it is sufficient to construct a signature group. For a block
group system, it is sufficient to construct a signature sequence with support
on [t, t + ℓ]. We first construct a signature sequence, and then specialize this
construction for a signature group and block group system. It is slightly easier to construct the isomorphic version of signature sequence (68), sequence
(85).
First we give an easy way to evaluate the operation ⊚t0,0 in group
(▽0,0 (Ut ), ⊚t0,0 ) in signature sequence (85), and then give a construction
of (▽0,0 (Ut ), ⊚t0,0 ) = (Ut , ◦t ). We use Corollary 76 to evaluate the operation ⊚t0,0 . Since Corollary 76 holds, then if
▽j,k (u̇t ) ⊚tj,k ▽j,k (üt ) = ▽j,k (ūt )
for 0 ≤ j ≤ k, then
▽j,k+1 (u̇t ) ⊚tj,k+1 ▽j,k+1 (üt ) = ▽j,k+1 (ūt )
for 0 ≤ j ≤ k + 1. This means we may evaluate
▽0,0 (u̇t ) ⊚t0,0 ▽0,0 (üt ) = ▽0,0 (ūt ),
(96)
recursively for k = ℓ, . . . , 0 in order, using triangles ▽j,k (u̇t ), ▽j,k (üt ) for
0 ≤ j ≤ k, at each step k. First for step k = ℓ, we evaluate
▽j,ℓ (u̇t ) ⊚tj,ℓ ▽j,ℓ (üt ) = ▽j,ℓ (ūt )
t−j
for 0 ≤ j ≤ k. This gives the elements r̄0,ℓ
in row ℓ of ▽0,0 (ūt ) for 0 ≤ j ≤ ℓ.
Continue in this way for steps k = ℓ, . . . , 0 in order. In general, for any step
k, we evaluate
▽j,k (u̇t ) ⊚tj,k ▽j,k (üt ) = ▽j,k (ūt )
(97)
for 0 ≤ j ≤ k. Note that the evaluation of (97) depends on the evaluation of
▽j,k+1 (u̇t ) ⊚tj,k+1 ▽j,k+1 (üt ) and ▽j+1,k+1 (u̇t ) ⊚tj+1,k+1 ▽j+1,k+1 (üt ), which
are already done in the previous step k + 1. Then the evaluation of (97) is
t−j
greatly simplified, since the only new element we need to find is r̄0,k
. So in
t−j
step k we are finding elements r̄0,k
in row k of ▽0,0 (ūt ) for 0 ≤ j ≤ k.
In general, we can use this same approach to construct group (Ut , ◦t ) in
signature sequence (85). Suppose we have found groups (▽j,k+1 (Ut ), ⊙tj,k+1 )
for 0 ≤ j ≤ k+1. Then we need to find groups (▽j,k (Ut ), ⊙tj,k ) for 0 ≤ j ≤ k
that have a homomorphism to groups (▽j,k+1 (Ut ), ⊙tj,k+1 ) for 0 ≤ j ≤ k +1.
This process begins with the simplest groups (▽j,ℓ (Ut ), ⊙tj,ℓ ), for 0 ≤ j ≤ ℓ,
and terminates with the group (Ut , ◦t ) = (▽0,0 (Ut ), ⊙t0,0 ) that we want to
find.
In the definition of a signature sequence, the operation ⊚tj,k on set
▽j,k (Ut ) is defined by
def
▽j,k (u̇t ) ⊚tj,k ▽j,k (üt ) = ▽j,k (u̇t ◦t üt )
64
for any u̇t , üt ∈ Ut . The definition of operation ⊚tj,k uses u̇t , üt ∈ Ut . By
contrast, in the construction of a signature sequence, any ut ∈ Ut is not
available a priori to define operation ⊙tj,k in group (▽j,k (Ut ), ⊙tj,k ). Therefore in construction, we interpret group (▽j,k (Ut ), ⊙tj,k ) in a different way.
There is no use of ut ∈ Ut to define operation ⊙tj,k in group (▽j,k (Ut ), ⊙tj,k ).
We simply define an operation ⊙tj,k on the set of subtriangles in ▽j,k (Ut )
that forms a group, e.g., on the set of subtriangles in ▽j,k (Ut ) formed by
t−(j+ℓ−k)
t−j
t−j
in (84). After the construction is complete
vertices r0,k
, r0,ℓ
, and r0,ℓ
and ut ∈ Ut is known, we show the operation ⊙tj,k in group (▽j,k (Ut ), ⊙tj,k )
has the same property as (i)c in the definition of ⊚tj,k .
Along these lines, note that the definition of ut in (84) requires knowlt−j
edge of representative r0,k
which is also not available a priori in construc-
t−j
tion. In construction, we interpret r0,k
as a point in the permutation group
t
t
(▽j,k (U ), ⊙j,k ) devoid of any meaning as a representative. For example
in construction, we can consider the elements in (84) to be any set of distinct integers, and (▽j,k (Ut ), ⊙tj,k ) is to be defined on a subtriangle of these
integers.
We now construct a group (Ut , ◦t ) that satisfies condition (i)c of the
definition of a signature sequence (85).
Algorithm 87 (to construct (Ut , ◦t ) for a signature sequence)
DO
FOR j such that 0 ≤ j ≤ ℓ,
define a group (▽j,ℓ (Ut ), ⊙tj,ℓ ) with operation ⊙tj,ℓ on ▽j,ℓ (Ut ).
ENDFOR
FOR k = ℓ − 1, . . . , 0 (counting down in order),
FOR j such that 0 ≤ j ≤ k,
define a group (▽j,k (Ut ), ⊙tj,k ) such that for m = j and m = j + 1,
the projection map ▽j,k (Ut ) → ▽m,k+1 (Ut ) given by the assignment
▽j,k (ut ) 7→ ▽m,k+1 (ut ) defines a homomorphism from (▽j,k (Ut ), ⊙tj,k ) to
(▽m,k+1 (Ut ), ⊙tm,k+1 ).
ENDFOR
ENDFOR
def
Define (Ut , ◦t ) = (▽0,0 (Ut ), ⊙t0,0 ).
ENDDO
Lemma 88 At the completion of Algorithm 87, for each k such that 0 ≤
k ≤ ℓ, for 0 ≤ j ≤ k, the projection map Ut → ▽j,k (Ut ) given by the
assignment ut 7→ ▽j,k (ut ) defines a homomorphism from (▽0,0 (Ut ), ⊙t0,0 )
to (▽j,k (Ut ), ⊙tj,k ).
Proof. The result is trivial for j = k = 0. Consider the remaining cases. Fix
j, k such that 0 ≤ k ≤ ℓ, 0 ≤ j ≤ k, and fix group (▽0,0 (Ut ), ⊙t0,0 ). There
are three cases to consider. For j = 0, by Algorithm 87 there is a group
(▽j,k−1 (Ut ), ⊙tj,k−1 ) such that the projection map ▽j,k−1 (Ut ) → ▽j,k (Ut )
defines a homomorphism from (▽j,k−1 (Ut ), ⊙tj,k−1 ) to (▽j,k (Ut ), ⊙tj,k ). For
65
j = k, by Algorithm 87 there is a group (▽j−1,k−1 (Ut ), ⊙tj−1,k−1 ) such that
the projection map ▽j−1,k−1 (Ut ) → ▽j,k (Ut ) defines a homomorphism from
(▽j−1,k−1 (Ut ), ⊙tj−1,k−1 ) to (▽j,k (Ut ), ⊙tj,k ). For 0 < j < k, by Algorithm
87 there are two groups, (▽j,k−1 (Ut ), ⊙tj,k−1 ) and (▽j−1,k−1 (Ut ), ⊙tj−1,k−1 ),
such that the projection map ▽j,k−1 (Ut ) → ▽j,k (Ut ) defines a homomorphism from (▽j,k−1 (Ut ), ⊙tj,k−1 ) to (▽j,k (Ut ), ⊙tj,k ), and the projection map ▽j−1,k−1(Ut ) → ▽j,k (Ut ) defines a homomorphism from
(▽j−1,k−1 (Ut ), ⊙tj−1,k−1 ) to (▽j,k (Ut ), ⊙tj,k ).
Continuing this process, we can construct a ladder of groups ending in
(▽0,0 (Ut ), ⊙t0,0 ), given by
(▽0,0 (Ut ), ⊙t0,0 ) → (▽m,1 (Ut ), ⊙tm,1 ) → (▽n,2 (Ut ), ⊙tn,2 ) → · · ·
· · · → (▽p,k−1(Ut ), ⊙tp,k−1 ) → (▽j,k (Ut ), ⊙tj,k ), (98)
where m = 0 or m = 1, n = m or n = m + 1, and p = j or p = j − 1, such
that in each contiguous pair
(▽r,q (Ut ), ⊙tr,q ) → (▽s,q+1 (Ut ), ⊙ts,q+1 ),
where r = s or r = s − 1, the projection map ▽r,q (Ut ) → (▽s,q+1 (Ut )
defines a homomorphism from (▽r,q (Ut ), ⊙tr,q ) to (▽s,q+1 (Ut ), ⊙ts,q+1 ). In
general there are many other ladders besides (98).
Fix the ladder shown in (98). By assumption, there is a projection map
and homomorphism for the pair
(▽0,0 (Ut ), ⊙t0,0 ) → (▽m,1 (Ut ), ⊙tm,1 ),
and for the pair
(▽m,1 (Ut ), ⊙tm,1 ) → (▽n,2 (Ut ), ⊙tn,2 ).
It follows there is a projection map ▽0,0 (Ut ) → ▽n,2 (Ut ) and homomorphism for the concatenation
(▽0,0 (Ut ), ⊙t0,0 ) → (▽n,2 (Ut ), ⊙tn,2 ).
Continuing this process down the ladder, we can show there is a projection
map ▽0,0 (Ut ) → ▽j,k (Ut ) and homomorphism for the concatenation
(▽0,0 (Ut ), ⊙t0,0 ) → (▽j,k (Ut ), ⊙tj,k ).
(99)
The projection map for the concatenation (99) is the same for any other
ladder besides (98). Therefore the homomorphism is the same for all ladders.
•
Theorem 89 For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j ≤ k, in the group
(▽j,k (Ut ), ⊙tj,k ), the operation ⊙tj,k on set ▽j,k (Ut ) satisfies
▽j,k (u̇t ) ⊙tj,k ▽j,k (üt ) = ▽j,k (u̇t ◦t üt )
for any u̇t , üt ∈ Ut .
66
Proof. Let u̇t , üt ∈ Ut . By Lemma 88, if u̇t 7→ ▽j,k (u̇t ) and üt 7→ ▽j,k (üt )
under the projection map Ut → ▽j,k (Ut ), then
u̇t ⊙t0,0 üt 7→ ▽j,k (u̇t ) ⊙tj,k ▽j,k (üt )
under the projection map Ut → ▽j,k (Ut ). But this means
▽j,k (u̇t ⊙t0,0 üt ) = ▽j,k (u̇t ) ⊙tj,k ▽j,k (üt )
or
▽j,k (u̇t ◦t üt ) = ▽j,k (u̇t ) ⊙tj,k ▽j,k (üt ).
•
Corollary 90 Algorithm 87 is sufficient to satisfy condition (i)c for group
(Ut , ◦t ) in signature sequence (85).
Proof. Theorem 89 shows Algorithm 87 is sufficient to satisfy condition
(i)c.
•
Algorithm 87 asks for a group (▽j,k (Ut ), ⊙tj,k ) with a homomorphism to
(▽j,k+1 (Ut ), ⊙tj,k+1 ) and (▽j+1,k+1(Ut ), ⊙tj+1,k+1 ). This makes the choices
for group (▽j,k (Ut ), ⊙tj,k ) very limited. The permutation of ▽
− j,k (Ut ) is
t
t
t
already determined by (▽j,k+1 (U ), ⊙j,k+1 ) and (▽j+1,k+1 (U ), ⊙tj+1,k+1 ).
Therefore to define the group (▽j,k (Ut ), ⊙tj,k ), we need only define the per-
t−j
t−j
mutation of r0,k
for each (r0,k
,▽
− j,k (ut )) ∈ ▽j,k (Ut ).
We now give an algorithm to find a signature sequence (85). We have
already given Algorithm 87 to find (Ut , ◦t ) for some arbitrary fixed time
t. Next we give Algorithm 91 to obtain a signature sequence by extending (Ut , ◦t ) on the right with the half infinite sequence (Ut+1 , ◦t+1 ), . . .,
and on the left with the half infinite sequence . . . , (Ut−1 , ◦t−1 ). To extend
(Ut , ◦t ) on the right, we shift (Ut , ◦t ) to obtain a group (▽1,1 (Ut+1 ), ⊚t+1
1,1 ).
t+1
t+1
Then to complete the definition of a group (▽0,0 (U ), ⊚0,0 ) which satisfies condition (i)c, we just need to define the groups (▽0,k (Ut+1 ), ⊚t+1
0,k )
along the first column. This can be done in the same way as the construction of (Ut , ◦t ). We repeat this process for each successive time epoch. To
extend (Ut , ◦t ) on the left, we find the inverse shift of (Ut , ◦t ) to obtain
a group (▽0,1 (Ut−1 ), ⊚t−1
0,1 ). Then to complete the definition of a group
(▽0,0 (Ut−1 ), ⊚t−1
)
which
satisfies condition (i)c, we just need to define the
0,0
t−1
t−1
groups (▽k,k (U ), ⊚k,k ) along the last diagonal. Again this can be done
in the same way as the construction of (Ut , ◦t ). We repeat this process for
each previous time epoch.
Algorithm 91 (to construct signature sequence (85))
DO
Construct (Ut , ◦t ) using Algorithm 87.
ENDDO
DO
(to extend (Ut , ◦t ) on the right)
67
FOR time t + 1,
t
t
Define group (▽1,1 (Ut+1 ), ⊙t+1
1,1 ) as the shift of (▽0,1 (U ), ⊙0,1 ). In other
words, under the 1-1 correspondence of triangle elements ▽0,1 (ut ) and
▽1,1 (ūt+1 ) in ▽0,1 (Ut ) and ▽1,1 (Ut+1 ), respectively, given by
▽0,1 (ut ) = ▽1,1 (ūt+1 ),
there is an isomorphism
(▽0,1 (Ut ), ⊙t0,1 ) ≃ (▽1,1 (Ut+1 ), ⊙t+1
1,1 ).
FOR k = ℓ − 1, . . . , 0 (counting down in order),
define a group (▽0,k (Ut+1 ), ⊙t+1
0,k ) such that for m = 0 and m = 1,
t+1
the projection map ▽0,k (U ) → ▽m,k+1 (Ut+1 ) given by the assignment ▽0,k (ut+1 ) 7→ ▽m,k+1 (ut+1 ) defines a homomorphism from
t+1 ), ⊙t+1
(▽0,k (Ut+1 ), ⊙t+1
0,k ) to (▽m,k+1 (U
m,k+1 ).
ENDFOR
def
Define (Ut+1 , ◦t+1 ) = (▽0,0 (Ut+1 ), ⊙t+1
0,0 ).
ENDFOR
Repeat for times t + 2, t + 3, . . ..
ENDDO
DO
(to extend (Ut , ◦t ) on the left)
FOR time t − 1,
t
t
Define group (▽0,1 (Ut−1 ), ⊙t−1
0,1 ) as the inverse shift of (▽1,1 (U ), ⊙1,1 ). In
other words, under the 1-1 correspondence of triangle elements ▽0,1 (ut−1 )
and ▽1,1 (ūt ) in ▽0,1 (Ut−1 ) and ▽1,1 (Ut ), respectively, given by
▽0,1 (ut−1 ) = ▽1,1 (ūt ),
there is an isomorphism
t
t
(▽0,1 (Ut−1 ), ⊙t−1
0,1 ) ≃ (▽1,1 (U ), ⊙1,1 ).
FOR k = ℓ − 1, . . . , 0 (counting down in order),
define a group (▽k,k (Ut−1 ), ⊙t−1
k,k ) such that for m = k and m = k + 1,
the projection map ▽k,k (Ut−1 ) → ▽m,k+1 (Ut−1 ) given by the assignment ▽k,k (ut−1 ) 7→ ▽m,k+1 (ut−1 ) defines a homomorphism from
t−1 ), ⊙t−1
(▽k,k (Ut−1 ), ⊙t−1
m,k+1 ).
k,k ) to (▽m,k+1 (U
ENDFOR
def
Define (Ut−1 , ◦t−1 ) = (▽0,0 (Ut−1 ), ⊙t−1
0,0 ).
ENDFOR
Repeat for times t − 2, t − 3, . . ..
ENDDO
68
We now specialize the construction of signature sequence (85) to construct a signature group. We define an isomorphic version of (Rt , ⋆) in a
similar way as for (Rt , ⋆t ). We use the 1-1 correspondence Rt ↔ Ut with
assignment rt 7→ ut defined previously. Let ṙt 7→ u̇t and r̈t 7→ üt . Define an
operation ◦ on set Ut by
def
u̇t ◦ üt = ūt
if
ṙt ⋆ r̈t = r̄t
where r̄t 7→ ūt in Rt ↔ Ut . This gives a group (Ut , ◦) isomorphic to (Rt , ⋆)
and (Ut , ◦) ≃ (Rt , ⋆). We also say (Ut , ◦) is an induced group of G. Again,
matrix Ut has the same form as matrix Rt , and so in the following we reuse
the same notation for Ut as for Rt .
Consider any finite group G and any induced group (Ut , ◦) of G. We
say the induced group (Ut , ◦) is a signature group if and only if it satisfies
the following two conditions:
(i)e. For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j ≤ k, there is an operation
⊚j,k on set ▽j,k (Ut ) that is well defined by
def
▽j,k (u̇t ) ⊚j,k ▽j,k (üt ) = ▽j,k (u̇t ◦ üt )
for u̇t , üt ∈ Ut , and this forms a group (▽j,k (Ut ), ⊚j,k ). Note that
(▽0,0 (Ut ), ⊚0,0 ) ≃ (Ut , ◦).
(ii)e. For each k such that 0 ≤ k ≤ ℓ, for 0 ≤ j < k, under the 1-1
correspondence of triangle elements ▽j,k (ut ) and ▽j+1,k (ūt ) in ▽j,k (Ut )
and ▽j+1,k (Ut ), respectively, given by
▽j,k (ut ) = ▽j+1,k (ūt ),
there is an isomorphism
(▽j,k (Ut ), ⊚j,k ) ≃ (▽j+1,k (Ut ), ⊚j+1,k ).
In Algorithm 87, we have already constructed a group (Ut , ◦t ) which
satisfies condition (i)c for a signature sequence. To construct a signature
group (Ut , ◦), we need to satisfy both conditions (i)e and (ii)e just above; (i)e
is the same as (i)c. The following algorithm is a modification of Algorithm
87 which includes condition (ii)e. This algorithm is sufficient to satisfy both
(i)e and (ii)e.
Algorithm 92 (to construct signature group (Ut , ◦))
DO
FOR j = ℓ,
define a group (▽j,ℓ (Ut ), ⊙j,ℓ ) with operation ⊙j,ℓ on ▽j,ℓ (Ut ).
FOR j = ℓ − 1, . . . , 0 (counting down in order),
define a group (▽j,ℓ (Ut ), ⊙j,ℓ ) such that (▽j,ℓ (Ut ), ⊙j,ℓ )
≃
t
(▽j+1,ℓ (U ), ⊙j+1,ℓ ). In other words, under the 1-1 correspondence of
triangle elements ▽j,ℓ (ut ) and ▽j+1,ℓ (ūt ) in ▽j,ℓ (Ut ) and ▽j+1,ℓ (Ut ),
respectively, given by
▽j,ℓ (ut ) = ▽j+1,ℓ (ūt ),
69
there is an isomorphism
(▽j,ℓ (Ut ), ⊙j,ℓ ) ≃ (▽j+1,ℓ (Ut ), ⊙j+1,ℓ ).
ENDFOR
ENDFOR
FOR k = ℓ − 1, . . . , 0 (counting down in order),
define a group (▽k,k (Ut ), ⊙k,k ) such that for m = k and m = k + 1,
the projection map ▽k,k (Ut ) → ▽m,k+1 (Ut ) given by the assignment
▽k,k (ut ) 7→ ▽m,k+1 (ut ) defines a homomorphism from (▽k,k (Ut ), ⊙k,k ) to
(▽m,k+1 (Ut ), ⊙m,k+1 ).
FOR j = k − 1, . . . , 0 (counting down in order),
define a group (▽j,k (Ut ), ⊙j,k ) such that (▽j,k (Ut ), ⊙j,k )
≃
t
(▽j+1,k (U ), ⊙j+1,k ).
In other words, under the 1-1 correspondence
of triangle elements ▽j,k (ut ) and ▽j+1,k (ūt ) in ▽j,k (Ut ) and ▽j+1,k (Ut ),
respectively, given by
▽j,k (ut ) = ▽j+1,k (ūt ),
there is an isomorphism
(▽j,k (Ut ), ⊙j,k ) ≃ (▽j+1,k (Ut ), ⊙j+1,k ).
ENDFOR
ENDFOR
def
Define (Ut , ◦) = (▽0,0 (Ut ), ⊙0,0 ).
ENDDO
Algorithm 92 can be simplified as follows.
Algorithm 93 (to construct signature group (Ut , ◦))
DO
FOR j = ℓ,
define a group (▽j,ℓ (Ut ), ⊙j,ℓ ) with operation ⊙j,ℓ on ▽j,ℓ (Ut ).
Define a group (▽ℓ−1,ℓ (Ut ), ⊙ℓ−1,ℓ ) such that (▽ℓ−1,ℓ (Ut ), ⊙ℓ−1,ℓ ) ≃
(▽ℓ,ℓ (Ut ), ⊙ℓ,ℓ ). In other words, under the 1-1 correspondence of triangle
elements ▽ℓ−1,ℓ (ut ) and ▽ℓ,ℓ (ūt ) in ▽ℓ−1,ℓ (Ut ) and ▽ℓ,ℓ (Ut ), respectively,
given by
▽ℓ−1,ℓ (ut ) = ▽ℓ,ℓ (ūt ),
there is an isomorphism
(▽ℓ−1,ℓ (Ut ), ⊙ℓ−1,ℓ ) ≃ (▽ℓ,ℓ (Ut ), ⊙ℓ,ℓ ).
ENDFOR
FOR k = ℓ − 1, . . . , 0 (counting down in order),
define a group (▽k,k (Ut ), ⊙k,k ) such that for m = k and m = k + 1,
the projection map ▽k,k (Ut ) → ▽m,k+1 (Ut ) given by the assignment
▽k,k (ut ) 7→ ▽m,k+1 (ut ) defines a homomorphism from (▽k,k (Ut ), ⊙k,k ) to
(▽m,k+1 (Ut ), ⊙m,k+1 ).
70
Define a group (▽k−1,k (Ut ), ⊙k−1,k ) such that (▽k−1,k (Ut ), ⊙k−1,k ) ≃
(▽k,k (Ut ), ⊙k,k ). In other words, under the 1-1 correspondence of triangle
elements ▽k−1,k (ut ) and ▽k,k (ūt ) in ▽k−1,k (Ut ) and ▽k,k (Ut ), respectively,
given by
▽k−1,k (ut ) = ▽k,k (ūt ),
there is an isomorphism
(▽k−1,k (Ut ), ⊙k−1,k ) ≃ (▽k,k (Ut ), ⊙k,k ).
ENDFOR
def
Define (Ut , ◦) = (▽0,0 (Ut ), ⊙0,0 ).
ENDDO
We now specialize the construction of signature sequence (85) to construct a block group system. To construct a block group system, it is sufficient to construct a signature sequence with support on [t, t + ℓ], where
▽0,0 (Ut+n ), 0 ≤ n ≤ ℓ, has the same form as given in Lemma 69 for
▽0,0 (Rt+n ). To do this, we can modify Algorithm 91 as follows. For time
t + n, n < 0, define (Ut+n , ◦t+n ) to be the identity group. We then extend (Ut−1 , ◦t−1 ) on the right in the same way as (Ut , ◦t ) is extended on
the right in Algorithm 91. For time t, define group (▽1,1 (Ut ), ⊙t1,1 ) as the
t
shift of (▽0,1 (Ut−1 ), ⊙t−1
0,1 ). Then include columns u0 of the form given in
t
t
Lemma 86 to form group (U , ◦ ). This procedure is the same as done in
Algorithm 91 for time t + 1. In general, for time t + n, 0 < n ≤ ℓ, extend
t+n−1
(Ut+n−1 , ◦t+n−1 ) on the right by shifting (▽0,1 (Ut+n−1 ), ⊙0,1
) to from
t+n
t+n
t+n
group (▽1,1 (U ), ⊙1,1 ), and then include columns u0 of the form given
in Lemma 86 to form group (Ut+n , ◦t+n ). For time t + n, n > ℓ, define
(Ut+n , ◦t+n ) to be the identity.
6.7 A sequence of group systems
In this section, we show that the group system (R, ∗) can be reduced to
a trivial group system by a sequence of group systems. We find the sequence
of group systems by using Theorem 59, Corollary 60, and Theorem 61. For
each k such that 0 ≤ k ≤ ℓ, define a sequence of indices ik by
ik = . . . , (k, k), (k, k), . . .
at times . . . , t, t + 1, . . .. For each r ∈ R, consider the sequence
. . . , ▽k,k (rt ), ▽k,k (rt+1 ), . . . .
Let R(ik ) be the set of all such sequences. Define a projection map ∂k :
R → R(ik ) by the assignment
∂k : r 7→ . . . , ▽k,k (rt ), ▽k,k (rt+1 ), . . .
for r ∈ R.
Let ṙk , r̈k ∈ R(ik ), where
ṙk = . . . , ▽k,k (ṙt ), ▽k,k (ṙt+1 ), . . . ,
71
and
ṙk = . . . , ▽k,k (ṙt ), ▽k,k (ṙt+1 ), . . . .
Define an operation ∗R(ik ) on R(ik ) by
def
t+1
), . . . .
ṙk ∗R(ik ) r̈k = . . . , ▽k,k (ṙt ) ⊛tk,k ▽j,k (r̈t ), ▽k,k (ṙt+1 ) ⊛t+1
k,k ▽k,k (r̈
We have the following analogs of Theorem 59, Corollary 60, and Theorem
61.
Theorem 94 R(ik ) with operation ∗R(ik ) forms a group (R(ik ), ∗R(ik ) ).
Corollary 95 R(ik )
(R(ik ), ∗R(ik ) ).
with
operation
∗R(ik )
forms
a
group
system
Theorem 96 The projection map ∂k : R → R(ik ) given by the assignment
r 7→ ∂k (r) defines a homomorphism from (R, ∗) to (R(ik ), ∗R(ik ) ).
Corollary 97 There is a homomorphism C → (R(ik ), ∗R(ik ) ) with mapping b 7→ ∂k (rb ) where b ↔ rb is given by the 1-1 correspondence in the
isomorphism C ≃ (R, ∗).
Proof. There is an isomorphism C ≃ (R, ∗). Now use Theorem 96.
•
Corollary 98 There is an isomorphism C/Ck−1 ≃ (R(ik ), ∗R(ik ) ) with correspondence bCk−1 ↔ ∂k (rb ) where b ↔ rb is given by the 1-1 correspondence in the isomorphism C ≃ (R, ∗).
Proof. From Theorem 96, we know there is a homomorphism (R, ∗) →
(R(ik ), ∗R(ik ) ) with mapping r 7→ ∂k (r). Then by the first homomorphism
theorem, there is an isomorphism
(R/KR , ∗) ≃ (R(ik ), ∗R(ik ) )
(100)
with correspondence rKR ↔ ∂k (r), where KR is the kernel of the homomorphism. The kernel KR of the homomorphism is the set of r ∈ R that
map to the identity of (R(ik ), ∗R(ik ) ). The identity of (R(ik ), ∗R(ik ) ) is the
sequence
. . . , ▽k,k (1t ), ▽k,k (1t+1 ), . . . .
Then the set of r ∈ R that maps to the identity of (R(ik ), ∗R(ik ) ) is the
set of r ∈ R such that ▽k,k (rt ) = ▽k,k (1t ) for t ∈ Z. But this condition
fixes all the generator vectors r[t,t+m] in r, k ≤ m ≤ ℓ, to be the identity.
An r ∈ R such that r[t,t+m] is the identity for k ≤ m ≤ ℓ corresponds
to a c ∈ C which is in the (k − 1)-controllable subcode Ck−1 of C, where
r ↔ c is the correspondence of the isomorphism C ≃ (R, ∗). Therefore
KR ↔ Ck−1 . Since C ≃ (R, ∗) and KR ↔ Ck−1 , there is an isomorphism
C/Ck−1 ≃ (R/KR , ∗) with correspondence bCk−1 ↔ rb KR where b ↔ rb
is the correspondence of the isomorphism C ≃ (R, ∗). Then from (100) there
is an isomorphism C/Ck−1 ≃ (R(ik ), ∗R(ik ) ) with correspondence bCk−1 ↔
∂k (rb ), where b ↔ rb is given by the 1-1 correspondence of the isomorphism
C ≃ (R, ∗).
•
72
It can be shown the group system C/Ck−1 is strongly controllable, in
fact (ℓ − (k − 1))-controllable. Then the group system C/Ck−1 has a branch
group sequence whose terms are the time t components of C/Ck−1 , and a
signature sequence whose terms are the time t components of (R(ik ), ∗R(ik ) ).
These results show that any group system (R, ∗) can
be associated with a natural sequence of group systems
(R(iℓ ), ∗R(iℓ ) ), . . . , (R(ik ), ∗R(ik ) ), . . . , (R(i0 ), ∗R(i0 ) ) = (R, ∗).
The sequence of k-controllable subcodes C0 , C1 , . . . , Ck , . . . , Cℓ = C of C,
denoted {Ck }ℓk=0 , is used in [3] to find a shift register encoder for C. The
work [3] and [4] also mention a related and complementary approach to
finding a shift register encoder, the sequence of derivative codes of C, given
in the Loeliger thesis1 . In the sequence of derivative codes, each code is the
state code of the previous code, and so the sequence terminates in the trivial
code. The sequence of derivative codes is essentially the same as the derived
sequence of shifts used in symbolic dynamics [10]. The derivative codes and
derived sequence are used in the construction of a time invariant group code
in [9] and a (time invariant) homogeneous shift in [10], respectively.
We may also consider the sequence of quotient groups systems
C/Cℓ−1 , . . . , C/Ck−1 , . . . , C/C1 , C/C0 , C/C−1 = C, where C−1 is defined to be the identity sequence.
We denote this sequence as
0
{C/Ck−1 }k=ℓ .
From Corollary 98, we know that C/Ck−1 ≃
(R(ik ), ∗R(ik ) ).
Therefore the sequence of quotient group systems
C/Cℓ−1 , . . . , C/Ck−1 , . . . , C/C−1 = C is isomorphic to the sequence of group
systems (R(iℓ ), ∗R(iℓ ) ), . . . , (R(ik ), ∗R(ik ) ), . . . , (R(i0 ), ∗R(i0 ) ) = (R, ∗). In
general, the terms in sequence {C/Ck−1 }0k=ℓ are not the same as terms
in sequence {Ck }ℓk=0 , and the two sequences are not the same and not isomorphic. We can think of sequences {C/Ck−1 }0k=ℓ and {(R(ik ), ∗R(ik ) )}0k=ℓ
as the dual of the sequence {Ck }ℓk=0 , and the dual of the derivative codes
and derived sequence of shifts. Then the construction of an ℓ-controllable
time invariant group system given in Algorithm 93 can be regarded as a dual
construction to [9]. There appears to be no dual result in the literature for
the contruction of a general ℓ-controllable group system given in Algorithm
91.
1
I was unable to obtain the Loeliger thesis from the Zurich library by written request
in English.
73
References
[1] B. Kitchens, “Expansive dynamics on zero-dimensional groups,” Ergodic Theory and Dynamical Systems 7, pp. 249-261, 1987.
[2] J. C. Willems, “Models for dynamics,” in Dynamics Reported, vol. 2,
U. Kirchgraber and H. O. Walther, Eds., New York: John Wiley, 1989.
[3] G. D. Forney, Jr. and M. D. Trott, “The dynamics of group codes: state
spaces, trellis diagrams, and canonical encoders,” IEEE Trans. Inform.
Theory, vol. 39, pp. 1491-1513, Sept. 1993.
[4] H.-A. Loeliger and T. Mittelholzer, “Convolutional codes over groups,”
IEEE Trans. Inform. Theory, Part I, vol. 42, pp. 1660-1686, Nov. 1996.
[5] M. D. Trott, “The algebraic structure of trellis codes,” Ph.D. thesis,
Stanford Univ., Aug. 1992.
[6] G. D. Forney, Jr., “Coset codes–Part II: Binary lattices and related
codes,” IEEE Trans. Inform. Theory, vol. 34, pp. 1152-1187, Sept. 1988.
[7] G. D. Forney, Jr., “Geometrically uniform codes,” IEEE Trans. Inform.
Theory, vol. 37, pp. 1241-1260, Sept. 1991.
[8] G. D. Forney, Jr., “Minimal realizations of linear systems: the “shortest
basis” approach,” arXiv:0910.4336v3 [cs.IT], 24 Aug 2010.
[9] J. P. Sarvis and M. D. Trott, “Useful groups for trellis codes,” in Proc.
IEEE Int. Symp. Inform. Theory, Whistler, BC, Canada, Sept. 17-22,
1995, p. 308.
[10] N. T. Sindhushayana, B. Marcus, and M. Trott, “Homogeneous shifts,”
IMA J. Math. Contr. Inform., vol. 14, pp. 255-287, 1997.
[11] R. F. Williams, “Classification of subshifts of finite type,” Annals of
Mathematics, vol. 98, pp. 120-153, 1973; “erratum,” Annals of Mathematics, vol. 99, pp. 380-381, 1974.
[12] J. J. Rotman, An Introduction to the Theory of Groups (4-th edition),
Springer, New York, 1995.
[13] M. Hall, Jr., The Theory of Groups, Chelsea, New York, 1959.
[14] D. Lind and B. Marcus, An Introduction to Symbolic Dynamics and
Coding, Cambridge Univ. Press, New York, 1995.
[15] J. Dénes and A. D. Keedwell, Latin squares and Their Applications,
Budapest, Akadémiai Kiado, 1974.
[16] D. Jungnickel, “Latin squares, their geometries and their groups. A survey,” in Coding Theory and Design Theory, Part II (D. Ray-Chaudhuri,
ed.), vol. 21 of IMA Volumes in Mathematics and its Applications, pp.
166-225, Springer, 1992.
[17] K. M. Mackenthun, Jr., “On groups with a shift structure: the Schreier
matrix and an algorithm,” in 41st Annual Conf. on Information Sciences and Systems, Baltimore, MD, March 14-16, 2007.
74
[18] K. M. Mackenthun, Jr., “Group codes and the Schreier matrix form,”
arXiv:1106.3981v1 [cs.IT], 20 Jun 2011.
[19] K. M. Mackenthun, Jr., “Time and harmonic study of strongly controllable group systems, group shifts, and group codes,” arXiv:1211.2340v3
[cs.IT], 3 May 2017.
[20] K. M. Mackenthun, Jr., “The fundamental group of a strongly controllable group system, group shift, and group code,” arXiv:1709.08265v1
[cs.IT], 24 Sept 2017.
75
| 7cs.IT
|
WAYLA - Generating Images from Eye Movements
Bingqing Yu
Centre for Intelligent Machines, McGill University
arXiv:1711.07974v1 [cs.CV] 21 Nov 2017
[email protected]
James J. Clark
Centre for Intelligent Machines, McGill University
[email protected]
Abstract
[6, 12].
One of the shortcomings of the aforementioned image
enhancement approaches is that, for most of them, they only
consider the problem of generating an image with higher
resolution using a given low-resolution image. However, in
many instances, one doesn’t even have a lower resolution
image. How can we generate an image in this case? Obviously, some additional information is needed. One approach
that is quite popular recently is to train a network to generate an image based on a verbal description. In this paper we
look at another situation, wherein we have someone looking
at an image, and want to generate an image similar, if not
identical, to the one that is being viewed. But we assume
that the only information we have available is the viewers’
gaze patterns or eye movement trajectories.
We present a method for reconstructing images viewed
by observers based only on their eye movements. By exploring the relationships between gaze patterns and image
stimuli, the “What Are You Looking At?” (WAYLA) system
learns to synthesize photo-realistic images that are similar
to the original pictures being viewed. The WAYLA approach
is based on the Conditional Generative Adversarial Network (Conditional GAN) image-to-image translation technique of Isola et al. [9]. We consider two specific applications - the first, of reconstructing newspaper images from
gaze heat maps, and the second, of detailed reconstruction
of images containing only text. The newspaper image reconstruction process is divided into two image-to-image translation operations, the first mapping gaze heat maps into image segmentations, and the second mapping the generated
segmentation into a newspaper image. We validate the performance of our approach using various evaluation metrics,
along with human visual inspection. All results confirm the
ability of our network to perform image generation tasks
using eye tracking data.
It is not immediately obvious whether there is enough
information in the eye tracking data to allow reconstruction
of what is being viewed. But, there is a well-examined link
between image content and gaze patterns. It has been stated
by many studies related to visual attention that viewers scan
the interesting locations of a given image by controlling
their attention, and the trajectory of the attention can be explored by tracking the eye movements of the viewers. Examples in the literature include the Clark and Ferrier computational model of saliency which was used in a robotic
system to demonstrate the relationships that exist between
image saliency and eye movements [3] and the Itti et al.
visual attention model [10], which also showed the ability of the eye fixation data to represent viewing behaviors
and image salience characteristics. In the context of attention tracking during text reading, researchers have carried
various studies and developed numerous theories and models in order to describe how the eyes move while reading.
For example, O’Regan et al. [15] demonstrated that for a
single word recognition task, there exists an initial fixation
location which minimizes the probability of refixation and
total fixation duration on a word. In the context of contin-
1. Introduction
Image generation has always been one of the primary
topics in the field of computer vision. Due to various limitations such as the lack of source image information, it often
happens that the available image datasets are either insufficient in quantity or defective in quality. Therefore, a variety
of image synthesis methods have been developed to generate images that are more useful and valuable for subsequent
image processing tasks.
In past years, various linear and nonlinear methods have
been explored to improve image quality via interpolation [8,
20, 26]. Recently, researchers have focused on using superresolution approaches to generating more detailed images
1
uous reading, this was extended to the notion of Preferred
Viewing Location, where the fixation locations depend on
the words being read [13]. McConkie and co-workers have
discovered that there is an optimal fixation position in sentence reading, which is slightly to the left of a word center. Its duration can be influenced by various factors such
as the complexity of a word. Moreover, it has been found
that the fixation durations are the longest if the target is at
the word center [22]. Grammatical factors also have an impact: functional words such as prepositions often have a
higher skipping rate [25]. In sum, reading behaviors and
fixation characteristics are strongly influenced by the target
locations and various lexical variables of the reading material. This suggests that the particular shape and structure of
eye movement trajectories constrain, to some unknown extent, the text being read. Similar conclusions can be drawn
with respect to general image viewing - that the eye trajectory information constrain to some extent the contents of
the image being viewed.
1.1. Proposed Method
To investigate the problem of generating images from
eye movements, we propose to use deep learning and create
a neural network which takes in eye fixation data or gaze
heat maps and generates as output images similar to the
original scenes that are displayed to the viewers. We call
our approach ”WAYLA”, which stands for ”What Are You
Looking At?”. The feasibility of this approach is enhanced
by the presence of numerous datasets of images with corresponding eye tracking annotations that have been published
in recent years [24].
Although it is possible to construct an image generation model from scratch, we benefit from the state-of-theart deep learning models to perform our eye movements to
image prediction task. We propose to use a network based
on the Conditional Generative Adversarial Network, the details of which will be presented in Section 2 [14]. Instead
of using solely the idea of the traditional GAN network [7]
to generate images by minimizing the Euclidean distance
between output images and ground truth stimuli, we use
the architecture of the Conditional GAN. In this way, our
network can learn to generate a synthetic image by conditioning on a corresponding eye fixation heat map. As
a result, the goal becomes to minimize the difference between a patch combining the eye fixation heat map with
the generated image and a patch combining the eye fixation heat map with the ground truth image. The architecture of the proposed WAYLA network uses the image-toimage translation model presented by Isola et al. [9] as a
pre-built model. This study focuses on its specific application, which is transforming an image from one version
to another; however, we focus on exploring it as a solution
to the eye-fixation-data-to-image generation problem. We
mainly focused on applying our model to two eye tracking
datasets [21, 4]. One contains various scanned images of
newspapers and magazines with associated gaze heat maps,
and the other contains eye tracking data obtained during text
reading. Our network takes the eye fixation heat maps provided by these datasets as input and produces two sorts of
output images. The first type of output is a simplified semantic segmentation, while the second type of output is a
detailed photorealistic image. In particular, our network is
able to synthesize a segmented version of newspaper appearances. This simplified version of realistic newspapers is
consistent with the eye-tracking data and filled with semantically labeled regions representing picture and text. Taking
a step further, our network is also effective at then generating text-embedded images that are at a higher detail level.
In this way, it is possible to obtain a more concrete representation of what people are looking at during reading. The
dataset provided by Vilkin et al. [21] contains numerous
highly detailed newspaper images and their corresponding
segmented versions. Since this database did not have any
associated eye track data, we used a state-of-the-art saliency
model, MLNET [5], to generate all the eye fixation heat
maps for the ground truth newspaper images. These heat
maps are used to model where people may look while viewing these newspapers, and serve as the eye fixation data that
is used as input for our network during training. We also
applied our WAYLA approach to another dataset, named
GECO, the Ghent Eye-Tracking Corpus [4]. The GECO
dataset contains eye fixation data collected during reading
of text on pages of novels displayed on a computer screen.
We employ a conditional generative adversarial network
in our implementation. Where the eye fixation data is used
to condition the operation of the generator. The discriminator serves to distinguish between “fake” segmentations
or images and “real” ones. Therefore, in the case of training on “real” examples, the discriminator input consists of
an image which is a combination of the eye fixation data
and the ground truth images. The ground truth images are
newspaper images at various detail levels for the Vilkin et
al. dataset and text-embedded images for the GECO dataset.
However, in the case of “fake” examples, the combination
is a concatenation of the eye fixation data with the output
images produced by the generator. In this way, by training the network on the examples described above, it learns
to generate synthetic images that are similar to the ground
truth images.
2. Architecture and Training
2.1. Data Preparation
For training our network to learn a mapping from eye
movement data to newspaper images provided by the Vilkin
et al. dataset, we needed to produce eye fixation data and
Figure 1. Illustration of our image generation pipeline. Our approach formulates the problem of generating detailed newspaper
images from eye fixation data as a two-phase process. The network needs to be fed with the eye fixation heat maps as input and
trained with the segmented images as ground truth. Then we can
utilize the segmented images, along with the eye fixation data, to
train the network with the highly detailed images as ground truth.
Figure 2. (Left) A ground truth text-embedded image generated
from the GECO dataset. (Right) The corresponding eye fixation
heat map generated from the fixation data of the GECO dataset.
feed them as input for our model. For this purpose we used
the state-of-the-art MLNET model [5] to build the eye fixation heat map dataset. To obtain the MLNET saliency predictions for the newspaper dataset, we fed the MLNET network with the original scanned newspapers used by Vilkin
et al., and used this saliency model as the corresponding
salience heat map for each input stimulus. This allows us to
get fixation heat maps that are strongly in agreement with
the real gaze locations. In this way, by using the generated eye fixation data and the image datasets provided by
the Vilkin dataset, we can train our model to output newspaper images at various detail levels. As shown in Figure 1,
we broke the end-to-end image generation process into two
phases. The first phase has the goal of generating semantic segmentations of the newspaper images, while the second phase is used to generate the detailed newspaper images
from the segmentations.
We also investigated the effectiveness of our model on
reconstructing the text images used in obtaining the GECO
eye movement dataset. This dataset, which contains records
about the eye fixation positions and durations of each individual for each reading session, enabled us to directly use
their eye fixation data as input to our model. Therefore, as
shown in Figure 2, we were able to generate the eye fixation
heat maps that correspond to different parts of the novel that
were read by participants. When generating grayscale eye
fixation heat maps for the GECO dataset, for each observer
and for each time a fixation is made on a specific position
of a specific word, we place a bright point in the grayscale
heat map at a position that corresponds to its recorded fixation location. The brightness of that point is modulated with
the recorded percentage of time spent on that specific location out of the total trial time completed by an observer. The
maximum value for this recorded percentage value is 0.17
percent, so all fixation points with fixation duration percentage less than this value will be represented with a less bright
point in our heat maps. The maximum pixel value for the
synthetic heat map is 255, corresponding to a fixation point
that has a duration value of 0.17 percent. If a fixation occupies 0.017 percent of the total trial time, its pixel value
becomes 25.5. It is possible that several fixations are made
within one word, in this case, distinct bright points corresponding to distinct fixation positions will be added into the
heat map. However, for the fixation points belonging to the
same word, we chose to modulate the brightness of all these
fixation points using the total percentage of trial time for
that specific word under the assumption that the global duration value is more useful for estimating the importance of
a word as compared to other words in the reading material.
As for the ground truth images that serve as targets for
training our network, no sophisticated data preparation process was undertaken. The ground truth segmented newspaper images and ground truth detailed newspaper images are
simply taken from the dataset provided by Vilkin et al. Regarding the GECO dataset, we chose to break the reading
material into parts for generating RGB images containing
printed text. Each text image is a (256 X 256) size RGB image, with its red channel encoding a constant background
and its green channel encoding text content. The blue channel is set to zero everywhere. It was found experimentally
that this 3-channel arrangement provided better training stability, reduces the possibility of divergence and allows faster
convergence, as compared with using a single channel containing only the text content. Each image contains 15 words
arranged into 3 rows and each row contains 5 words one following another. For generating our eye fixation heat maps
for the GECO dataset, the locations of all saliency points are
adapted to the locations of the generated text-embedded images. We ensured that the correspondence is in agreement
with the data provided by the GECO dataset.
2.2. Network Training
As stated earlier, the relationships between eye movements and viewing material have been examined in numerous studies, giving us confidence that we can develop an
Figure 3. Illustration of the Conditional GAN architecture used for
WAYLA.
image synthesis method that is able to synthesize images
from eye track data that are similar to some extent to the
images actually being viewed. In recent years, generative
adversarial networks and their extensions led to advances in
the field of photo-realistic image synthesis [7, 17, 14, 9].
Given enough training samples, they can produce photorealistic images that are very close to the ground truth pictures. Therefore, we chose to build our system based on the
architecture of Conditional GAN, which has proven to be
an efficient way to do image-to-image translation for various computer vision tasks such as image segmentation and
grayscale-to-RGB image translation [9]. The overall architecture based on which the WAYLA system is built is illustrated in Figure 3.
The input data was the gaze heat maps obtained from
the aforementioned data preparation steps, and we fed these
heat maps to the input layer of our neural network using the
architecture of Conditional GAN described in the work of
Isola P. et al. [9]. The modifications that we made to the
published conditional GAN structure are presented in the
remaining part of the section.
2.2.1
Individual Training of Two Phases for Newspaper Generation
We proposed WAYLA as a model to generate image content
based on eye fixation data. Since the Vilkin et al. dataset
provided us with both segmented and detailed newspaper
images, we formulated the image generation task as a twophase process. The first phase consists in training the network to do eye-movement-data-to-segmented-newspaperimage synthesis. The second phase consists in training the
network to generate newspaper images with a higher level
of detail from the image segmentation. In this section, we
will present how we generated newspaper images by training the WAYLA model separately and independently for the
two phases. In the next section, we will present a slightly
different approach that joins the two training phases to yield
an End-to-End implementation for generating newspaper
images.
When training our network for the first phase, the generator is fed with the eye fixation heat maps that we produced using MLNET. During training, the generator is op-
Figure 4. Illustration of the implementation which trains the network separately for generating segmented images and detailed images. The top part of the figure shows the training process of the
first phase. Our model takes the eye fixation data as input and is
trained with the segmented newspaper images as ground truth. The
bottom part of the figures shows the training process of the second
phase. Our model takes the segmented images from the Vilkin
et al. dataset as input and is trained with the detailed newspaper
images as ground truth.
timized to produce outputs that are as similar as possible
to the ground truth segmented newspaper images. The discriminator is fed with an image patch which concatenates
the input eye fixation heat maps with the generated images
produced from the generator. When receiving this kind of
patches, the discriminator is trained to recognize them as
“fake” images. In the “real” image case, the discriminator receives patches which concatenate the eye fixation heat
maps with the ground truth segmented newspaper images.
An illustration of this first phase training is shown in the
upper part of Figure 4.
For the second phase, we trained our network to synthesize detailed newspaper images based on the segmented
newspaper images. The lower part of Figure 4 illustrates
the input and output settings of this second phase training.
During the second phase, the input layer of the generator
is fed with the segmented images provided by the Vilkin et
al. dataset. Then, the generator is optimized to produce outputs that are as similar as possible to the ground truth highly
detailed newspaper images. In this case, the discriminator
is fed with image patches which concatenate the segmented
images and the detailed images, and its task is to distinguish
synthesized data from ground truth data.
2.2.2
End-to-End Design of Pipeline for Newspaper
Generation
We also explored the possibility of our network to learn to
generate highly detailed newspaper images when it only has
the eye fixation heat maps at its disposal, which is the more
interesting case.
To achieve this goal, we applied an end-to-end training
process to our system. We started by feeding the input layer
of generator with the eye fixation heat maps and training
the system to generate segmented images by using the segmented newspapers from the Vilkin et al. dataset as ground
Figure 5. Illustration of the end-to-end implementation. Initially
our model takes the eye fixation data as input and is trained with
the segmented newspaper images as ground truth. Then, the model
takes as input image patches which concatenate the eye fixation
data with the generated segmented images, and it is trained with
the highly detailed newspaper images as ground truth.
truth. At this stage, the operation is exactly the same as the
first phase training presented in Section 2.2.1. However, the
difference between the implementation mentioned in Section 2.2.1 and this end-to-end implementation is indicated
by the input and output settings that we used for the remaining part of the training process. After finishing training our
system to produce segmented images, we re-initialized the
system and fed the input layer of the generator with a new
kind of input, different from the segmented images that we
used as input for the second phase training presented in Section 2.2.1. At this time, the segmented images generated by
the generator that we just trained previously are concatenated with the eye fixation heat maps to form a new set of
input RGB images, which are then fed into our re-initialized
system. Now, the generator takes these inputs and is optimized to output images that are as similar as possible to the
ground truth highly detailed newspaper images provided by
the Vilkin et al. dataset. It is worth mentioning that the concatenation is done in such a way that the new red channel is
formed by adding the pixel values of the eye fixation heat
maps to the pixels values of the red channel of the generated segmented images. The new blue channel is formed
by taking the pixel values of the blue channel of the generated segmented images. The new green channel is formed
by setting all the values to 0, except for the locations where
all the three channels of the generated segmented are equal
to 255, in this case, the green channel pixel remains 255 to
form a white color along with the other two channels. The
design of the discriminator is the same as all the aforementioned designs. It receives image patches and distinguishes
whether they belong to “true image pairs” or “fake image
pairs”.
In this way, we explored the feasibility of generating
highly detailed newspaper images when only the eye fixation heat maps are available. An illustration of the end-toend design is shown in Figure 5.
2.2.3
Training on the GECO Dataset
We also applied WAYLA on the GECO dataset in order to
investigate the effectiveness of our model to generate text-
only images based on eye fixation data. The generator has
as input the eye fixation heat maps that we created from the
GECO dataset. While the generator is trained to produce
text-like images and has the ground truth text embedded images as target, the discriminator is trained to distinguish the
generator’s outputs as “fake” images. In the “fake image”
case, the discriminator receives as input a combination of
the eye fixation heat maps and the generator’s outputs; however, in the “real image” case, the discriminator receives the
eye fixation heat maps concatenated with the ground truth
text embedded images as input.
The loss functions used for our network are presented in
the following, and it is applied on all training phases and all
datasets involved in our study.
The discriminator, whose task is to classify between real
and fake pairs, uses the following binary cross entropy loss
as its loss function:
LD = Ex,y [logD(x, y)] + Ex ([1 − logD(x, G(x))] (1)
In Equation 1, x is the input of the generator, y represents all ground truth images that the generator has as target. As for the generator, since it is stated in [16] that mixing the GAN loss with another standard content loss such as
Euclidean loss can improve the training of deep neural networks, we chose to use the L1 distance as the additional loss
and combine it with the adversarial loss described above to
construct the loss function for our generator. The L1 distance represents the difference between the outputs of the
generator and the ground truth images. Thus, the overall
loss function of the generator is defined as:
LG = LD + λL1 (G)
(2)
We set the value of λ to 0.01, and this decision is taken
based on our observation from experiments and on the analysis made in [9], which states that when the L1 loss is
weighted 100 time larger than the GAN loss, there are fewer
artifacts produced in the output of the generator. All layers
of the network need to be trained from scratch. Weights
are randomly initialized using uniform distribution between
−0.05 to 0.05. We always preserve 20 percent of the total samples for testing. The network is trained by updating
the generator and the discriminator alternatively. The GAN
cross-entropy loss is backpropagated to the discriminator
to update its weights. Then, by keeping the discriminator
weights constant, we combine the cross-entropy loss with
the L1 loss and backpropagate this error to update the generator weights. Minibatches of 2 samples per batch are used
for training. The RMSProp optimizer is used to optimize
both the generator and the discriminator, with a learning
rate of 0.001, a decay rate of 0.9, a momentum of 0 and an
of 1 × 10−6 . Dropout layers and batch normalization are
used in our network to accelerate convergence.
Segmented newspapers
Detailed newspapers
(Individual Training)
Detailed newspapers
(End-to-End)
SSIM
0.83
0.54
0.53
Table 1. SSIM scores obtained by comparing our synthesized images with the ground truth.
3. Evaluation and Discussion
Among the large body of work done on evaluating the
perceptual quality of images, losses such as L1 and L2 distances have been the dominant performance metrics in the
domain of computer vision. However, in order to address
a drawback of the basic L1 and L2 metrics, which is their
inability to indicate structural information carried by images, we chose to use the structural similarity (SSIM) index to compute the perceived similarity between our generated images and our ground truth images. In this way,
the evaluation metric may provide a better understanding of
the overall quality of generated images, since the similarity
scores are not only in terms of pixel values but also in terms
of image structure [23, 1]. In Table 1, similarity scores
are reported for both the segmented and detailed newspaper image generation. By computing the scores between
our generated segmented newspaper images and the ground
truth segmented newspapers, we can evaluate the ability of
our network to do eye-movement-data-to-segmented-image
prediction. By computing the similarity scores between our
generated detailed newspaper images and the ground truth
detailed newspapers, we can evaluate the ability of our network to do the segmented-image-to-highly-detailed-image
transformation.
As there are no other methods for generating images
from eye movement data, to investigate the significance
of our contribution to synthesizing newspaper images, we
looked at the SSIM scores reported in various papers concerning photo-realistic image generation. In comparing
our SSIM scores with other studies, our SSIM scores indicate that our synthesized images are similar to the ground
truth images. Mihaela R. et al. stated that with their autoencoding GAN, they can generate images with a SSIM
mean value of 0.62 when compared with ground truth images [18]. In another paper, which uses a GAN structure to
implement image super-resolution, it is shown that a generated image with SSIM score of 0.6423 is very similar to the
original image to human perception [11]. In addition, when
Zhou et al. introduced the concept of SSIM, they presented
an image disturbed with Gaussian noise with a SSIM score
of 0.30. The noisy version still preserves most of the texture and structure characteristics of the original image, so
Figure 6. Example results of WAYLA obtained by training the two
phases separately and independently. The qualitative results of
WAYLA on the testing set are compared to ground truth.
Figure 7. Example results of WAYLA obtained by implementing the end-to-end training pipeline. The qualitative results of
WAYLA on the testing set are compared to ground truth.
it is still understandable and recognized by human inspection. The SSIM score obtained from our generated detailed
newspaper images using the end-to-end implementation is
situated slightly below the other SSIM scores presented in
Table 1; however, it is understandable that the end-to-end
implementation, which trains the network to output highly
detailed images without using the ground truth segmented
images, can only synthesize images of limited quality compared to the case when the two phases are trained individually. By visually inspecting the qualitative results and comparing them to the ground truth, it can be confirmed the
effectiveness of our model at synthesizing real images with
eye data.
Figures 6 and 7 allow a qualitatively evaluation of
the performance of our image generation pipeline. It can
be observed that the generated segmented images produce
text/picture patterns that visually appear close to the ground
truth. It is also worth noticing that when WAYLA is applied
Figure 10. Illustration of synthesized versus ground truth textembedded images, along with the segmentation process used to
obtain the word segment length distribution.
Figure 8. Qualitative results obtained by training WAYLA to generate text-embedded images using the GECO dataset.
Figure 9. Comparison between the word length distribution obtained from the ground truth images versus the word length distribution obtained from our generated images. The horizontal axis
is the word segment length with unit (pixel). The vertical axis
indicates the frequency of occurrence for different word segment
length values.
on a specific dataset, which in our case is the newspaper
dataset, and trained for generating segmented images, the
network is able to extract meaningful information from eye
movement data in such a way that the structure of the generated images fits the specific dataset very well. Taking the
first row of Figure 6 as an example, despite the fact that
there are numerous eye fixation points in the eye fixation
image, in the generated segmented image, only part of the
eye fixation locations are converted into picture areas. The
rest of the eye fixation locations are successfully identified
as text areas in the generated result. By visually inspecting the detailed images generated by WAYLA, the image
patterns of all the generated images can be easily related to
the image patterns of the ground truth scanned newspapers
presented in the original dataset.
As for evaluating the performance of our approach when
it is applied to the GECO dataset, we first provide some
example results of the generated text-embedded images in
Figure 8, from which a qualitative evaluation can be done.
It is important to remember that, although the network is
designed to perform image generation without using any
natural language processing, the training has proven to be
extremely successful in the sense that all the generated images display text-like areas. Furthermore, numerous valid
English words can be observed in the generated images.
We also utilized various text analysis metrics, in combination with human inspection, in order to evaluate the quality
of our outputs. Optical Character Recognition (OCR) was
used to extract all the text-like content from the synthetic
images generated by our network [19]. In total, 13270 alphabet characters are retrieved from 241 synthetic images.
There are only 72 occurrences for which OCR engine is not
able to convert a scanned character into a valid alphabet letter. In addition, as presented in Figure 10, we also used
text segmentation in attempt to divide the text content in
our synthesized images into segments with various lengths.
Each segment can be considered as an individual word that
the system generates based on the input eye fixation data.
It is worth mentioning that for illustration purpose, we
display the qualitative results of the text-embedded images
by using a black text color with white background color.
Although we used a red and green encoding method to facilitate the training of our network, the generated results can
be easily converted to black and white afterwards in order
to be more similar to the real image content that is viewed
by the readers.
Figure 9 compares the histogram that summarizes the
length distribution of all segments obtained from our synthetic images versus the histogram generated using the
ground truth images. It can be observed that the shape
of the two distributions are similar. Most segments have
lengths ranging from 23 to 26 pixels. This means that in
both ground truth and generated images, most words are
constructed using 3 to 4 characters. Thus we observe a close
similarity between the generated images and the ground
truth images. This serves as a demonstration of our network’s ability to generate highly text-like image content.
4. Conclusion
In this paper, we explored the possibility of inverting
the relationships between image stimuli and eye movements
and successfully developed an approach to synthesize the
content of viewed images based on eye tracking information. We presented WAYLA, a deep-learning technique
which utilizes Conditional Generative Adversarial Nets to
generate newspaper images with different detail levels from
eye fixation heat maps. Moreover, the system is able to
reconstruct images containing only text using solely eye
tracking data. Our work proved the feasibility of generating
image content from gaze data, which has never been demonstrated before. Our results showed that the deep convolutional neural network that we used is effective at performing
image generation tasks, achieving high similarity between
the generated results and the ground truth. Although further
improvements could be achieved to enhance the quality of
the generated images, the idea of inverting the path starting
from viewed content and ending with eye tracking information can be widely applied in various settings. For example, many previous works use Baysian Inference [2] as the
standard approach to infer visual task from eye movements.
The proposed WAYLA approach can be used to improve
the performance of these traditional methods. Furthermore,
previously most studies related to image generation using
Generative Adversarial Networks and its derivatives were
done without considering gaze information as input. By introducing the possibility of inferring image content based
on gaze data, the WAYLA approach could open doors for
more diversified image generation models in the future.
References
[1] F. Boulogne, J. D. Warner, and E. Neil Yager. scikit-image:
Image processing in python. 2014. 6
[2] R. Carpenter. The visual origins of ocular motility. Vision
and visual function, 8:1–10, 1991. 8
[3] J. J. Clark and N. J. Ferrier. Modal control of an attentive
vision system. In ICCV, pages 514–523, 1988. 1
[4] U. Cop, N. Dirix, D. Drieghe, and W. Duyck. Presenting
geco: An eyetracking corpus of monolingual and bilingual
sentence reading. Behavior research methods, 49(2):602–
615, 2017. 2
[5] M. Cornia, L. Baraldi, G. Serra, and R. Cucchiara. A
deep multi-level network for saliency prediction. In Pattern
Recognition (ICPR), 2016 23rd International Conference on,
pages 3488–3493. IEEE, 2016. 2, 3
[6] S. Farsiu, M. D. Robinson, M. Elad, and P. Milanfar. Fast
and robust multiframe super resolution. IEEE transactions
on image processing, 13(10):1327–1344, 2004. 1
[7] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in neural information
processing systems, pages 2672–2680, 2014. 2, 4
[8] H. Hou and H. Andrews. Cubic splines for image interpolation and digital filtering. IEEE Transactions on acoustics,
speech, and signal processing, 26(6):508–517, 1978. 1
[9] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Imageto-image translation with conditional adversarial networks.
arXiv preprint arXiv:1611.07004, 2016. 1, 2, 4, 5
[10] L. Itti, C. Koch, and E. Niebur. A model of saliency-based
visual attention for rapid scene analysis. IEEE Transactions
on pattern analysis and machine intelligence, 20(11):1254–
1259, 1998. 1
[11] C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang,
and W. Shi. Photo-realistic single image super-resolution
using a generative adversarial network. arXiv preprint
arXiv:1609.04802, 2016. 6
[12] S. Liu, J. Pan, and M.-H. Yang. Learning recursive filters for
low-level vision via a hybrid neural network. In European
Conference on Computer Vision, pages 560–576. Springer,
2016. 1
[13] G. W. McConkie, P. W. Kerr, M. D. Reddix, and D. Zola. Eye
movement control during reading: I. the location of initial
eye fixations on words. Vision research, 28(10):1107–1118,
1988. 2
[14] M. Mirza and S. Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2, 4
[15] J. K. O’Regan and A. Lévy-Schoen. Eye-movement strategy
and tactics in word recognition and reading. 1987. 1
[16] D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A.
Efros. Context encoders: Feature learning by inpainting.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 2536–2544, 2016. 5
[17] A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 4
[18] M. Rosca, B. Lakshminarayanan, D. Warde-Farley,
and S. Mohamed.
Variational approaches for autoencoding generative adversarial networks. arXiv preprint
arXiv:1706.04987, 2017. 6
[19] R. Smith. An overview of the tesseract ocr engine. In
Document Analysis and Recognition, 2007. ICDAR 2007.
Ninth International Conference on, volume 2, pages 629–
633. IEEE, 2007. 7
[20] M. Unser. Splines: A perfect fit for signal and image processing. IEEE Signal processing magazine, 16(6):22–38, 1999.
1
[21] A. Vilkin, I. Safonov, and M. Egorova. Algorithm for segmentation of documents based on texture features. Pattern
recognition and image analysis, 23(1):153–159, 2013. 2
[22] F. Vitu, G. W. McConkie, P. Kerr, and J. K. O’Regan. Fixation location effects on fixation durations during reading: An
[23]
[24]
[25]
[26]
inverted optimal viewing position effect. Vision Research,
41(25):3513–3533, 2001. 2
Z. Wang and A. C. Bovik. Mean squared error: Love it or
leave it? a new look at signal fidelity measures. IEEE signal
processing magazine, 26(1):98–117, 2009. 6
S. Winkler and R. Subramanian. Overview of eye tracking
datasets. In Quality of Multimedia Experience (QoMEX),
2013 Fifth International Workshop on, pages 212–217.
IEEE, 2013. 2
C. Wotschack. Eye movements in reading strategies: How
reading strategies modulate effects of distributed processing
and oculomotor control, volume 1. Universitätsverlag Potsdam, 2009. 2
L. Zhang and X. Wu. An edge-guided image interpolation algorithm via directional filtering and data fusion. IEEE transactions on Image Processing, 15(8):2226–2238, 2006. 1
| 1cs.CV
|
Continuous Circadian Phase Estimation Using Adaptive Notch Filter
∗
Wei Qiao∗ , Kyle Altman∗ , Agung Julius∗ , Bernard Possidente† , John T. Wen∗ ,
Smart Lighting Engineering Research Center, Rensselaer Polytechnic Institute, Troy, NY
† Biology Department, Skidmore College, Saratoga, NY
arXiv:1804.00115v1 [cs.SY] 31 Mar 2018
Abstract
Actigraphy has been widely used for the analysis of circadian rhythm. Current practice applies regression
analysis to data from multiple days to estimate the circadian phase. This paper presents a filtering method for online
processing of biometric data to estimate the circadian phase. We apply the proposed method on actigraphy data of
fruit flies (Drosophila melanogaster).
I. I NTRODUCTION
Circadian rhythms are unique features of life that terrestrial species have evolved in response to the 24 h cyclic
environment of the Earth’s surface, which reflects the natural light cycle. For example, in humans, molecular
oscillations of the biological clock drive rhythmic physiological and behavioral functions i.e., heart beat, blood
pressure, body temperature, sleep-wake cycle, metabolism and locomotor activity. The disruption of circadian
rhythm caused by lack of synchrony between the circadian clock in the brain and the external environment may
produce serious detriments in human health and well-being and result in various issues ranging from increased
sleepiness, decreased attention span and lower productivity, to long-term health problems such as increased risk for
cancer, diabetes, obesity, and cardiovascular disorders [1]–[5]. Such disruption may be often observed in people
with irregular sleep patterns, artificial deprivation of light, i.e., submariners or mine workers [6], [7], frequently
shifted sleep-wake cycles of night nurses [2], and travelers who cross multiple time zones [8]. Therefore, control
of circadian rhythm is of great importance to circadian research.
Among many organisms, Drosophila has been widely studied in the investigation of circadian rhythms since
the 1950’s starting with the frontier work of C. Pittendrigh. Pittendrigh’s studies mostly focused on the eclosion
rhythms of Drosophila: the emergence of larva from the pupa case; in recent decades, genetic analysis based biomolecular interaction studies attracted many researchers to investigate circadian rhythm of Drosophila experimentally
and led to unveiling the molecular interactions of PER, TIM proteins and its mRNA, based on which Drosophila
circadian models are proposed [9]. However, the high cost and complexity of these protocols prohibit the continuous
measurement of circadian rhythm outputs; thus another behavioral genotype, the locomotor activity, is often chosen
as the output of the circadian clock. Although not as accurate as genetic analysis, the advantages of locomotor
activity based circadian research are low-cost, continuous measurement and integration as a fully automated control
and measuring system. Different types of Drosophila activity recording system were developed in the past decades;
the most widely used method is the commercial TriKinetics (www.trikinetics.com), which enables a fully automated
recording system.
The traditional method of circadian phase estimation is mostly based on double-actogram, in which phase is
extracted by Onset time, Acrophase (peak of a fitted cosine curve) or an eye-fit line. Despite the inaccurate nature
of this method, the phase estimation is on a daily basis and does not provide continuous phase estimation, which
would be an obstacle in many practices other than obtaining phase response curve, especially in feedback control
problems, which relies on circadian phase, i.e., a jet-lag problem.
In this paper, we adopt a model-free circadian phase estimation scheme using a modified adaptive notch filter
(ANF) developed by our group previously [10] (inspired by [11]), and design a set of experiments to obtain the
phase response curve. With ANF, we can continuously estimate circadian phase based on the locomotor activity
of Drosophila. This tool, which enables continuous circadian phase estimation, will be of vital importance to any
circadian related problems, is especially a big step towards closed loop circadian rhythm control in Drosophila and
is a helpful tool for a more complex study of human circadian rhythm. We will show in this paper that the ANF
method not only estimates the circadian phase accurately and continuously, but also offers filtering effects that can
successfully estimate the circadian phase from a set of noise-corrupted data, which the traditional double-actogram
method fails in.
The paper is organized as follows: some preliminary results of circadian phase estimation with ANF are introduced
in Section II. The experiment setup and protocols are detailed presented in Section III. The experiment design in
obtaining PRC is introduced in Section IV followed by the results and discussion in Section V. The concluding
remarks are given in Section VI.
II. P RELIMINARIES
In this section, how Drosophila locomotor activity is affected by external light is discussed and some preliminary
results of circadian phase estimation are introduced.
A. External Light and Drosophila Locomotor Activity
Similar to that of humans, Drosophila’s circadian system is most sensitive to blue light and almost completely
insensitive to red light [12], [13]. CRY (cryptochromes) in Drosophila’s deep brain is the circadian photorecepter,
which regulates the impact of light on the circadian system. The circadian pacemaker, affected by CRY, regulates
the expression of neuropeptide PDF, which affects the locomotor activity [14]. Therefore, Drosophila locomotor
activity can be used as a prediction of its circadian phase.
However, activity is not solely controlled by its circadian pacemaker, but also affected by the visual pathway:
light perception by the compound eye and other eye structures of Drosophila. It can be observed in the experiments
that the locomotor activity peaks right after the light turns on or off even for red light.
to be continued...
B. Adaptive Notch Filter and Locomotor Activity based Circadian Rhythm Estimation
Measuring circadian phase experimentally in real time is challenging; the most reliable method is genetic analysis
for Drosophila and onset of melatonin secretion under dim light conditions (dim light melatonin onset, or DLMO)
for human through lab test on saliva or plasma samples. However such measurements are inconvenient, time
consuming, and expensive. Some circadian-related physiological signals such as activity, heart rate, and body
surface temperature, which are possible to be measured at high sampling rate and show rhythmic patterns, are more
suitable for real time circadian phase estimation. Therefore, in this paper we focus on Drosophila locomotor activity
as the circadian clock output. Although simple and easy to measure, such signal is usually noisy and corrupted;
therefore a filter is required to extract circadian phase from such noisy physiological measurements.
A model-free circadian phase estimation scheme using a modified adaptive notch filter (ANF) developed by
our group previously [10] (inspired by [11]), which can also accommodate the non-zero mean and non-sinusoidal
waveform of the circadian signal, is introduced as follows:
Assume the following form of the circadian signal y,
N
y(t) =
∑ ak sin(kω ∗t + φk ) + d + w(t)
k=1
where d is a constant bias, w is a zero-mean white noise. The proposed modified ANF is given by
ẋ = AANF (ω)x + BANF (ω)y,
ω̇ = γω fANF (y, ω, x).
(1)
(2)
where the parameters and tuning procedures can be found in [10]. The ANF produces an estimate for (ω ∗t + φ1 ),
which is the argument of the fundamental harmonic and is used as the circadian phase estimate. In the cited work,
the local stability and robustness properties of the modified ANF algorithm are established and its effectiveness is
demonstrated on both synthesized periodic signals and the Drosophila activity data (an example is shown in Figure
1).
Fig. 1: An example of ANF algorithm on estimating the circadian phase of Drosophila activity data. Top: Drosophila
activity (black) and ANF estimation (red). Middle (top): Periodic blue light. Middle (bottom): the ANF extracts
the estimated signal argument. Bottom: the estimated phase is obtained by subtracting the linearly increasing part
from the argument.
III. E XPERIMENT S ETUP AND P ROTOCOL
MATERIALS: Drosophila Culture Kits from the Carolina Biological Supply Company are used for breeding
and preparation of the experiments. Each kit includes 36 culture vials in which the flies are bred and stored, instant
Drosophila medium food that sustains the flies in the culture vials, and sorting brushes. A FlyNap Anesthetic Kit,
also from Carolina, is used for sedation of the flies. The kit includes a 100 mL vial of FlyNap anesthetic and
several anesthetic transfer containers.
Chemicals and their application: The Flynap anesthetic is a mixture of 50% triethylamine, 1.25% 2-propanol,
1.13% methanol, 25% neutralizer fragrance, and 22.63% ethanol. 3 mL of FlyNap poured into the transfer container
is enough to anesthetize the flies for at least 30 minutes.
Flies: Canton S wild type. All flies selected for inclusion in the experiment had undergone eclosion three days
before the experiment began.
EXPERIMENTAL PREPARATION: Experimental set-up: 189 flies (21 flies in each incubator) were placed
in tubes (one tube per fly) that contained sufficient food for the experiments duration on one side: this food was
standard agar food that did not contain pymetrozine. These tubes were then loaded into nine separate Trikinetics
Drosophila Activity Monitor (DAM) 5 activity monitors. The temperature is kept at 25 degree Celsius using a
temperature control system.
EQUIPMENT:
(1) Trikinetics Drosophila Activity Monitor (DAM) 5: The monitors shine an infra-red beam down the center of
each tube and count the number of times the beam is broken by the fly in order to calculate how many times
each fly move along the tube. The monitors sum the count data into 1 minute bins and send the data to a PC.
(2) Culturing Incubators: The culture of flies are placed in light-tight incubators with Philips Rebel LEDs of cool
white light under 12 h light 12 h dark (LD 12:12) cycles.
(3) Experiment Incubators: The DAM 5 activity monitors are placed in light-tight incubators with Philips Rebel
LEDs at different wavelengths as light sources and mirrors mounted on the inner surface of the incubators.
The positions of LEDs are optimized by the optical simulator, ZEMAX, to ensure the light uniformity on the
monitor.
(4) Temperature Control System: High pressure air through a humidifier is used for ventilation and keeping the
food from drying. Temperature sensors and heaters are mounted on each incubator to maintain the temperature
Light Stimulus (CP (hrs)):
Light Strategy:
N/A
Dark (DD)
0
L
3
L
6
L
9
L
12
L
15
L
20
L
N/A
DD
TABLE I: Blue light (λ = 470 nm) stimulus at the designated circadian time.
in each incubator at 25 degree Celsius.
(5) Light control System: LEDs at different wavelengths mounted on top of each incubators are controlled by a
PC with designated light patterns and controllable light intensity.
Software: We use MATLAB for most of data handling and analysis and implementation of our algorithms.
ActogramJ [15] is used to obtain the Acrophase of the raw data in order to compare the results with that based on
our own algorithm.
IV. E XPERIMENTAL D ESIGN AND DATA A NALYSIS
In this section, the experiment is designed in order to obtain PRC of Drosophila (canton-S string) using ANF
method as well as traditional double-actogram based method.
A. Experimental Design
During the experiment, the lab is securely locked to prevent any external stimulus (except the designated LED
light) i.e, noise and ceiling light to affect the results. The experiment is designed such that all the flies are entrained
to the same phase before the external light stimulus are applied:
(1) Entrainment to the same phase: In order to investigate the influence of light stimulus to the circadian phase,
we need to ensure that all the flies are entrained into the same phase before given the light stimulus. Therefore,
a three days of LD 12:12 is imposed on all 9 incubators (each loaded with 21 flies).
(2) Light stimulus: After the three days entrainment, all incubators are put into darkness (DD). Light stimulus is
programmed at the designated circadian time (4 lux blue light with duration of 1 hour) 1 (Table I).
Data Analysis: The raw data in the form of text files is transferred to PC which stores how many times each
fly move along the tube within 1 minute with data logging time and light intensity information. There are more
than 17280 data points for each channel (fly) with total 189 channels. The raw data is analyzed in the following
manner:
(1) Exclude the arrhythmic flies: We use periodogram analysis to detect the rhythmicity of a fly and examine the
activity plot to exclude inactive flies. In our hands, we detect on average about 45% flies (lowest incubator
show approximately 30%) are rhythmic and will be used in the next step to obtain phase response curve.
(2) Utilize ANF to estimate the circadian phase: Each incubator’s screened data are analyzed using ANF algorithm,
and the estimated circadian phase is used to determine phase shift.
(3) Comparison to the traditional method using Cosinor on double actogram to determine phase shift: The
Acrophase (peak of a cosine wave fitted to the raw data) is used as a comparison in determining the phase
shift (ActogramJ software [15]).
1 The
beginning of subjective night is defined as Circadian Phase (CP) 0.
0.8
Amplitude
Amplitude
15
10
5
10
20
30
40
Period (h)
50
60
70
0.4
0.2
0
0
80
60
15
40
10
Activity
Activity
0
0
0.6
20
0
0
1000
2000
3000
4000
5000
6000
Time (min)
7000
8000
9000
(a) The fly passed the periodogram analysis which derived
from activity records of individual flies in complete darkness. Top panel shows the periodogram and the bottom
panel shows the raw activity records.
20
30
40
Period (h)
50
4000
5000
6000
Time (min)
60
70
80
5
0
0
10000
10
1000
2000
3000
7000
8000
9000
10000
(b) The fly failed the periodogram analysis which derived
from activity records of individual flies in complete darkness. Top panel shows the periodogram and the bottom
panel shows the raw activity records.
Fig. 2: Periodogram analysis.
V. R ESULTS
The experiments designed in this paper are standard, which have been widely adopted and discussed in the
literature [13]. The intention of this paper is to demonstrate a new circadian phase estimation tool, ANF, that is
accurate, continuous and robust against noise. We will first demonstrate the validity of our experiment, from which
the PRC is obtained using ANF and compares with that results from literature as well as the traditional method
based PRC (double-actogram and Acrophase); finally we demonstrate the filtering effect of ANF in which noise is
added to the raw data and ANF successfully rejects the noise, while the traditional double-actogram method shows
large error.
A. Periodogram and the Exclusion of the arrhythmic flies
Raw activity data of each individual fly under the 3 days of LD entrainment and designated light stimulus (Table
I) are obtained and analyzed here in order to exclude arrhythmic and inactive flies.
Periodogram is first applied to exclude the flies with an irregular period. 2 The data sets analyzed by the
periodogram are the activity records of individual flies from day 6 to day 12 when the flies are in complete
darkness. An example is shown in Figure 2a and Figure 2b where flies with regular period and irregular period are
shown, respectively. The averaged inclusion ratio is about 45% in each incubator.
B. Comparison of ANF and Acrophase based Phase Estimation
By excluding the arrhythmic flies, we use the selected data set to obtain phase response curve. The raw activity
data from the selected flies of each incubator are first averaged and then analyzed by the ANF algorithm (Section
II-B) to determine the phase of the averaged flies in each incubator. Results of two incubators are shown in
Figure 3a and Figure 3b in which 3 days of LD entrainment followed by a free running and a light pulse at CP
0, respectively. Top panel shows raw activity (black) and the ANF estimation (red); 2nd panel shows the light
pattern; bottom panel shows the ANF estimation of circadian phase. Notice that ANF provides a continuous period
adaptation and circadian phase estimation, which will be a very useful tool in any circadian related research. Now
the only question remains is how accurate the ANF based method is. We will compare the ANF based method
with the traditional double-actogram based method using the same data set.
The traditional method in determining circadian phase is mostly based on double-actogram where the daily peak
of the activity is identified using eye fit or Acrophase (peak of a cosine fitted curve). Here we use ActogramJ
software [15] to obtain Acrophase based on the same raw activity data of the selected flies obtained from Section
V-A. The double-actogram of control incubator is shown in Figure 4a where blue triangles are the estimated daily
2 Periodogram
is a nonparametric estimate of the power spectral density (PSD) estimate of the signal and uses traditional Fourier techniques
such as fast Fourier transform (FFT).
(a) Incubator with the 3 days of LD 12:12 and then free
run.
(b) Incubator with the 3 days of LD 12:12 and then a
light pulse at CP 0 with 4 lux intensity and 1 h duration.
Fig. 3: ANF analysis of the averaged data of each incubator: top panel shows raw activity (bold black) with standard
deviation (light black) and the ANF estimation (red); 2nd panel shows the light pattern; bottom panel shows the
ANF period adaptation.
1467
(a) Double actogram and the
Acrophase
estimation
using
ActogramJ. Blue triangles are
the estimated daily Acrophase and
the blue lines is a regression fit.
(b) Periodogram after the last light off using ActogramJ.
Fig. 4: Obtaining Periodogram and Acrophase based on the double actogram using ActogramJ (3 days of LD 12:12
and then free run).
Acrophase (with a regression fitted line) and the periodogram from day 4-12 is shown in Figure 4b. Notice here
the free-running period is τ = 24.453 where the ANF based period estimation is trying to adapt shown in Figure
3a and 3b bottom panel.
With the analysis of both methods ready, we move on to obtaining phase response curve. Note that circadian
phase is a relative concept, especially when obtaining PRC (the free-running period is normally not exactly 24 hrs).
Therefore the control incubator with 3 days of LD entrainment followed by free-running is used as the reference
incubator (Figure 3a) and agrees with the standard analysis in the literature. Phase shift is then defined as the phase
3 Period
of another control incubator is τ = 24.3.
difference between the target incubator and the control incubator. The same analysis is applied to both ANF and
double-actogram (ActogramJ) based method.
Two PRCs of both methods are shown in Figure 54 based on the phase difference between target incubator and
control incubator on day 10 with light stimulus given in table I. Notice the results are almost identical, demonstrating
that the accuracy of the ANF based method meets the well-accepted standard. We also compare our results with
those obtained in the literature [13] (black stars in Figure 5). The ANF obtains very similar results compared to
what has been accepted by the literature.
ActoJ
ANF (errorbar)
Literature
Fig. 5: ActogramJ based PRC (red dots and interpolated red curve) and the PRC obtained by ANF (blue dots and
interpolated blue curve), PRC obtained in the [13] is shown as black stars.
C. The Filtering Effect of ANF
The validity of ANF based circadian phase estimation is demonstrated above, where the result is almost identical
to that obtained by double-actogram based method and matches the PRC obtained by other researchers very well
[13]. In other words, ANF not only meets the accepted standard in the literature but also provides a continuous
circadian phase estimation.
Yet another important feature of ANF is the robustness against higher order harmonics, or noise [10]. In order
to demonstrate this feature, we intentionally corrupt the raw activity data selected in Section V-A with zero-mean
Gaussian white noise w where variance Var(w) = 10. Two examples are shown in Figure 6a and 6b. The same
procedure is applied to obtain ANF based PRC and double-actogram based PRC (ActoJgramJ) and the result is
shown in Figure 7. Notice that the ANF based method successfully rejects the noise effect and the error against its
original PRC is very small (average error is 0.57 hrs), while the double-actogram based method shows large error
(average error is 1.84 hrs).
VI. C ONCLUSION
ACKNOWLEDGMENT
This work was supported primarily by the Army Research Office through Grant number W911NF-13-1-0265.
We also gratefully acknowledge the support from National Science Foundation (NSF) through the Smart Lighting
Engineering Research Center (EEC-0812056) and from the Center for Automation Technologies and Systems
(CATS), which under a block grant from the New York State Empire State Development Division of Science,
Technology and Innovation (NYSTAR) contract C090145. The authors would also like to thank Drs. Mark Rea,
Mariana Figueiro, and Andrew Bierman at the RPI Light Research Center for introducing this topic to the authors.
4 Standard deviation of the ANF based method on individual fly in each incubator is shown as error bar, which is calculated as the square
root of the sum of the variance of target and control incubator.
(a) Incubator with the 3 days of LD 12:12 and then free
run.
(b) Incubator with the 3 days of LD 12:12 and then a
light pulse at CP 0 with 4 lux intensity and 1 h duration.
Fig. 6: ANF analysis of the corrupted data with noise w of each incubator: top panel shows raw activity (bold black)
with standard deviation (light black) and the ANF estimation (red); 2nd panel shows the light pattern; bottom panel
shows the ANF estimation of circadian phase.
ActoJ (Noncorrupted Data)
ActoJ (corrupted Data)
ANF (Noncorrupted Data)
ANF (corrupted Data)
Fig. 7: PRC obtained through the corrupted data with noise w, by ANF (blue dots and interpolated blue curve) and
ActogramJ (red dots and interpolated red curve). The original (based on non-corrupted data) PRCs are shown as
dotted curves.
R EFERENCES
[1] D.F. Kripke, D.J. Mullaney, M. Atkinson, and S. Wolf. Circadian rhythm disorders in manic-depressives. Biological Psychiatry, 1978.
[2] A. Knutsson. Health disorders of shift workers. Occupational Medicine, 53(2):103–108, 2003.
[3] S. Sephton and D. Spiegel. Circadian disruption in cancer: A neuroendocrine-immune pathway from stress to disease? Brain, Behavior,
and Immunity, 17(5):321–328, 2003.
[4] R.G. Stevens. Circadian disruption and breast cancer: from melatonin to clock genes. Epidemiology, 16(2):254, 2005.
[5] M.S Rea, A. Bierman, M.G. Figueiro, and J.D. Bullough. A new approach to understanding the impact of circadian disruption on
human health. Journal of Circadian Rhythm, 6, 2008.
[6] T.L. Kelly, D.F. Neri, J.T. Grill, D. Ryman, P.D. Hunt, D.J. Dijk, T.L. Shanahan, and C.A. Czeisler. Nonentrained circadian rhythms
of melatonin in submariners scheduled to an 18-hour day. Journal of Biological Rhythms, 14(3):190, 1999.
[7] J.N. Mills. Circadian rhythms during and after three months in solitude underground. The Journal of Physiology, 174(2):217, 1964.
[8] M. Harrington. Location, location, location: important for jet-lagged circadian loops. Journal of Clinical Investigation, 120(7):2265–
2267, July 2010.
[9] J.C. Leloup, D. Gonze, and A. Goldbeter. Limit cycle models for circadian rhythms based on transcriptional regulation in drosophila
and neurospora. J. Biological Rhythms, 14(6):433–448, 1999.
[10] J. Zhang, J.T. Wen, and A. Julius. Adaptive circadian rhythm estimator and its application to circadian rhythm control. In American
Control Conference, Washington D.C, June 2013.
[11] L. Hsu, R. Ortega, and G. Damm. A globally convergent frequency estimator. IEEE Trans. Automat. Control, 44(4):698–713, 1999.
[12] K.D. Frank and W.F. Zimmerman. Action spectra for phase shifts of a circadian rhythm in drosophila. Science, 163(3868):688–689,
1969.
[13] A. Klarsfeld, J.C. Leloup, and F. Rouyer. Circadian rhythms of locomotor activity in drosophila. Behavioural Processes, 64:161–175,
2003.
[14] P. Emery, R. Stanewsky, and C.H. Forster. Drosophila cry is a deep brain circadian photoreceptor. Neuron, 26:493C504, 2000.
[15] B. Schmid, C. Helfrich-Forster, and T. Yoshii. A new imagej plug-in ”actogramj” for chronobiological analysis. Journal of Biological
Rhythms, 26(5):464–467, 2011.
| 3cs.SY
|
THE MAGIC OF UNIVERSAL QUANTUM
COMPUTING WITH PERMUTATIONS
arXiv:1701.06443v2 [quant-ph] 6 Apr 2017
MICHEL PLANAT† AND RUKHSAN-UL-HAQ‡
Abstract. The role of permutation gates for universal quantum
computing is investigated. The ‘magic’ of computation is clarified in the permutation gates, their eigenstates, the Wootters discrete Wigner function and state-dependent contextuality (following many contributions on this subject). A first classification of
resulting magic states in low dimensions d ≤ 9 is performed.
PACS: 03.67.-a, 02.20.-a, 03.65.Fd, 03.65.Aa, 02.10.Ox, 03.65.Ud
MSC codes: 81P68, 81P45, 20B05, 81P13
1. Introduction
Quantum physics and (universal) computing are now considered to
be deeply interrelated. This relatively new idea of science owes much to
R. Feynman [1], D. Deutsch [2], P. Shor [3], S. Bravyi and A. Kitaev [4],
to mention a few popular landmarks. The time line of quantum computing also includes other important marks with many of them, during
the last two decades, dedicated to quantum computing platforms 1. R.
Feynman already understood that the simulation of a quantum system on a classical computer would need exponential resources. Later
D. Deutsch proposed a universal quantum computer made of quantum
gates as a way of simulating a quantum system with at most a polynomial overhead. We know from P. Shor that a quantum computer is able
to factor large integers in polynomial time by exploiting the parallelism
in the implementation of quantum Fourier transform [3, p. 221] and
that quantum error correction may circumvent the undesirable effects
due to decoherence by the use of quantum error-correcting codes [3, p.
435]. S. Bravyi and A. Kitaev introduced the principle of ‘magic state
distillation’: universal quantum computation may be realized thanks
to the stabilizer formalism (Clifford group unitaries, preparations and
measurements [7]) and the ability to prepare an ancilla in an appropriate single qubit mixed state.
Within the frame of universal quantum computation based on the
stabilizer formalism, it is being actively discussed if there exist critical
1
Let us mention Intel and NTT for spin qubits in semiconductors, Google, IBM
and D-Wave for superconducting qubits, Lockheed and INFINEON for trapped
ions. Other hardware efforts are being accomplished at university laboratories with
linear optics, atoms and cavity QED, quantum dots and impurity spins in solids,
etc [5, 6].
1
2
MICHEL PLANAT† AND RUKHSAN-UL-HAQ‡
resources responsible for the power of quantum computation. Remarkably, in odd dimensions, contextuality of the magic states seems to be
the magic ingredient [8]. In addition, the contextuality of states is witnessed by the negative entries of a quasi-probability distribution - the
discrete Wigner function (DWT). For even dimensions, the situation
is more obscure since state-independent contextuality (correlated with
the negativity of DWT) occurs from stabilizer states. Filtering of the
quantum states involved in the computation based on full state tomography has been proposed [9]. Finally, according to [10], contextuality
is required not just for measurement procedures but for preparation
procedures as well, in order that the two notions of nonclassicality are
revealed to be equivalent.
Wigner function was recognized to contain permutation symmetry
in its structure [11]. Interestingly enough, the experimental implementation of a simple quantum algorithm for determining the parity of
a permutation was performed [12]. We claim in this paper that permutation symmetry can sometimes be considered lying beneath the
concepts of magic states and contextuality that are responsible for the
universality and efficiency of quantum computation. Permutation gates
in the Clifford group (whose important elements are the CNOT gate
and the Toffoli or CCNOT gate) reveal non-stabilizer states in their
eigenvectors. The focus of this contribution is the use of two-generator
permutation groups seen as sets of permutation gates and being the
source of quantum states (stabilizer or magic) arising from their eigenstates. The groups of interest are in general different from the ones
encountered in the previous investigation of the Kochen-Specker theorem [13]. Only state-dependent contextuality is involved in quantum
computational universality, magic state distillation through quantum
error-correcting codes and possibly computational speed-up, as already
emphasized [8].
In Sec. 2, we remind standard results about the generalized Pauli
group, the phase-point operators, Wootters discrete Wigner function
and its link to quantum contextuality. In Sec. 3, we derive several types
of magic states arising from permutation matrices in small dimensions
d ≤ 9, we explicit their Wigner function and obtain state-dependent
proofs of contextuality based on the existence of pentagons between
the appropriate states (stabilizer and magic). Sec. 4 describes open
vistas for further study.
From permutations to quantum gates. It is unusual to recognize
the relationship of permutations with quantum gates as we intend to
do in this work. A n-letter permutation admits a n × n binary matrix
representation with exactly one entry of 1 in each row and each column
and 0s elsewhere. ‘Magic’ permutation matrices are those showing one
THE MAGIC OF UNIVERSAL QUANTUM COMPUTING WITH PERMUTATIONS
3
entry of 1 on their main diagonals. Some well known permutation matrices/gates are the Pauli gate X = 01 10 ≡ (2, 1), I ⊗ X ≡ (2, 1)(4, 3),
1 0 0 0
CNOT= 00 10 00 01 ≡ (1, 2)(4, 3), CCNOT≡ (1, 2, 3, 4, 5, 6)(8, 7), that
0 0 1 0
acts on one, two or three qubits, respectively. Similarly permutation
0 1 0
gates may act on qudits as the shift gate X = 0 0 1 ≡ (2, 3, 1) acting
1 0 0
on qutrits.
In Sec. 3, we focus on magic groups generated by two magic permutation gates (they exist as soon as d ≥ 4 as in (11) of Sec. 3.3).
2. The generalized Pauli group, the discrete Wigner
function and contextuality
We remind standard tools useful for further calculations and discussion.
The generalized Pauli group. Let d be a prime number, the qudit
Pauli group is generated by the shift and clock operators as follows
(1)
X |ji = |j + 1 mod di
Z |ji = ω j |ji
with ω = exp(2iπ/d) a dth root of unity. In dimension d = 2, X and
Z are the Pauli spin matrices σx and σz .
A general Pauli (also called Heisenberg-Weyl) operator is of the form
(2)
T(m,j) =
ijm Z m X j
if d = 2
−jm/2 m j
ω
Z X if d 6= 2.
where (j, m) ∈ Zd × Zd . For N particules, one takes the Kronecker
product of qudit elements N times.
Stabilizer states are defined as eigenstates of the Pauli group.
The discrete Wigner function. Associated with each d-dimensional
Hilbert space (d a prime) is a discrete phase space, a d × d array of
points on Zd × Zd . A set of phase point operators on the discrete phase
space is defined as [14, 15] (see also [16])
(3)
d−1
1 X pj−qm+jm/2 j m
Aα =
ω
X Z , α = (q, p) ∈ Zd × Zd .
d j,m=0
Wootters relations (10) and (11) in [14] follow.
Phase point operators have been built to satisfy properties analogous
to those of the continuous phase
R space in the context of the continuous
Wigner function W (q, p) = ρ(q + x/2, q − x/2) exp(ipx)dx in which
p and q are position and momentum, and ρ(x, x′ ) = ψ ∗ (x)ψ(x) is a
density matrix for a particle of coordinate x in a pure state of wave
function ψ(x) [1, p. 477]. It is required that the operators Aα satisfy
(i) For each point α, Aα is Hermitian,
4
MICHEL PLANAT† AND RUKHSAN-UL-HAQ‡
(ii) For any two points α and β, tr(Aα Aβ ) = dδαβ ,
(iii)Taking any complete set
P of d parallel lines (called a striation),
construct the average Pλ = 1d α∈λ Aα on each line λ. The d operators
Pλ form a set of mutually orthogonal projectors the sum of which is
the identity operator.
The d2 phase point operators Aα are linearly independent and form
a basis for the space of Hermitian operators acting on a d-dimensional
Hilbert space. As a result, any density operator can be developed as
X
(4)
ρ=
Wρ (q, p)A(q, p),
q,p
in which the real coefficients are explicitely given by the Wootters discrete Wigner function
(5)
1
Wρ (q, p) = tr[ρA(q, p)].
d
Unlike the continuous Wigner function, the discrete Wigner function
is a quasi probability distribution that may take negative values. It is
shown in [17] that, on a Hilbert space of odd dimension, the only pure
states to possess a non-negative discrete Wigner function are stabilizer
states.
On the contrary, a non-stabilizer pure state will be called a magic
state. This definition follows from [18, Corollary 1] which establishes
that any single-qubit pure state not one of the six Pauli eigenstates,
together with Clifford group operations and Pauli eigenstate preparation and measurement, allows universal quantum computation. For
arbitrary prime dimensions, magic state distillation is investigated in
[19]-[22]. For multiple qubits see [18, Corollary 2] any pure state which
is not a stabilizer state allows universal quantum computation and [9].
State-dependent quantum contextuality. Quantum contextuality
forbids theories revealing pre-existing values of observables under test
if the specific experimental set-up for measuring such observables is
not taken into account. One way to characterize quantum contextuality is to use a no-go approach a la Kochen-Specker involving sets
of quantum observables (as the Mermin square) or special subsets of
their eigenstates shared by their mutually commuting operators [23].
The smallest proof of state-independent contextuality is not a KochenSpecker set, it needs 13 rays in the three-dimensional Hilbert space [24].
State-independent contextuality can be obtained within the stabilizer
formalism for multiple qubits but is not manifested with qudits (when
d 6= 2).
The starting point of a state-dependent proof of quantum contextuality consists of a set of k binary tests that are represented by k rank
one projectors {Πi = |vi i hvi | , i = 1..k}. The tests are compatible if
THE MAGIC OF UNIVERSAL QUANTUM COMPUTING WITH PERMUTATIONS
5
and only if the projectors are mutually orthogonal. To the tests is associated an (exclusivity) graph Γ wherein each vertex corresponds to
the projectors Πi and an edge corresponds to compatible projectors. A
witness operator ΣΓ is defined as follows
(6)
ΣΓ =
k
X
Πi .
i=1
It is required that a value of 1 is assigned to at most one projector
in each joint measurement of (compatible) observables located on a
selected edge. For a non-contextual hidden variable theory, one expects
that the results of tests is such that hΣΓ imax ≤ α(Γ), where α(Γ) is
the independence number of the graph Γ, the cardinality of the largest
set of vertices such that no two elements are connected by an edge.
But a quantum contextual theory may bypass this bound and be such
that α(Γ) ≤ hΣΓ imax ≤ θ(Γ), in which the upper bound is the Lovasz
number of the exclusivity graph [8, 25]. It is calculated from θ(G) =
P
max|ψi ki=1 |hψ|vi i|2 , where the maximum is taken over all unit vectors
|ψi.
The simplest state-dependent proof of quantum contextuality corresponds √
to the cyclic graph C5 (also called a pentagon) in which
θ(C5 ) = 5 > α(C5 ) = 2. It was originally obtained for a spin1 quantum system, or qutrit [26]. Each exclusivity graph Γ where
α(Γ) < θ(Γ) allows a state-dependent proof of quantum contextuality.
If hΣΓ imax > α(Γ) for every state, then the proof of contextuality is
state-independent. To summarize, graphs in which α(Γ) < θ(Γ) may be
considered a proof of quantum contextuality if for appropriately chosen
projectors Πi and a state ρ the following non-contextuality inequality
is violated [21].
(7)
Tr(ΣΓ ρ) ≤ α(Γ).
Is contextuality needed for universality and quantum computational
speed-up? For qudits (in odd dimension), only states lying outside the
stabilizer polytope [27] manifest the negativity of the Wigner function
and simultaneously violate the inequality (7) through appropriate twoqudit projectors, hence they display state-dependent contextuality [8,
Theorem 1]. Thus the answer is yes.
For even dimensions, non-stabilizer pure states of single or multiple qubits are magic. But neither the condition of negativity nor the
criterion of contextuality is sufficient to promote such states to computational universality. This is because there exist (well characterized)
Kochen-Specker sets of multi-qubits [13]. However, it is shown in a recent paper [9] that contextuality is needed whenever two conditions are
satisfied, (i) the contextuality is state-dependent and (ii) one retains a
filtered set of quantum states able to ensure a full state tomography.
6
MICHEL PLANAT† AND RUKHSAN-UL-HAQ‡
state ψ
eigenstate of
|0i
σz
√1 (|0i + |1i)
σ
x
2
√1 (|0i − i |1i)
σ
y
2
|Hi
H=
σx√
+σz
2
Wρ
1 1 1
2 0 0
1 1 0
2 1 0
1 0 1
1 0
√ 2
1√
1 1+ 2
4√ 1
2
1−
√
1+ 3/3 1+ 3/3
√
√
1+ 3/3
3−1
1
|T i
SH
4
Table 1. Wootters discrete Wigner function for a few
pure states ρ = |ψi hψ|.
3. Magic states and state-dependent quantum
contextuality from groups of permutation gates
We already know that permutation symmetry exists in the discrete
Wigner function [11]. Our goal in this section is to shift our attention
from the set of Clifford gates to a subset whose elements are permutation gates, leading either to stabilizer states or not. Classes of magic
states and the related (state-dependent) contextuality will be investigated having in mind the tools described in Sec. 2 but by restricting to
the existence of pentagons between the appropriate states (as in [26]).
For dimensions larger than 4, the eigenstates of permutation gates
under consideration are living in a field that may be different from the
cyclotomic field Q[exp 2iπ
]. In this paper, for d > 4, we restrict the
d
classification of magic states to those having entries 0 and ±1 corresponding to eigenvalues ±1.
3.1. Qubit magic states. Fault tolerant quantum computing protocols based on stabilizer states have to be complemented by magic
states to reach quantum universality. Two distillation protocols based
on single qubit magic states |Hi and |T i are first described in [4] where
(8)
(9)
|Hi = cos( π8 ) |0i + sin( π8 ) |1i ,
|T i = cos(β) |0i + exp ( iπ4 ) sin(β) |1i , cos(2β) =
√1 .
3
Magic state |Hi is the +1-eigenstate of the the Hadamard matrix H =
σx√
+σz
that belongs to single qubit the Clifford group. Magic state |T i
2
2iπ
)-eigenstate of the SH matrix, where the phase
is the ω3 = exp(
3
1
0
gate S = 0 i also belongs to the Clifford group. Being not stabilizer
states, they cannot be prepared by actions from the Clifford group but
they can be distilled with these actions from a (dirty) mixed state to
a (neat) pure state thanks to appropriate quantum codes [4, 18, 19].
Matrix elements of Wootters discrete Wigner function Wρ (q, p) for a
few single qubit pure states are shown in Table 1. The first three rows
correspond to +1-eigenstates of Pauli spin matrices and the last two
THE MAGIC OF UNIVERSAL QUANTUM COMPUTING WITH PERMUTATIONS
7
rows are for magic states. Magic state |Hi possesses a negative entry
unlike the magic state |T i. In the latter case, negativity of the Wigner
function does not coincide to universality.
3.2. Magic states from qutrit permutation gates. The smallest
dimension for the occurrence of magic states associated to permutations
is three. There are, up to isomorphism, two (non-trivial, i.e. with two
distinct generators) two-generator permutation groups on three letters.
The permutation group isomorphic to Z3 contains the permutation
matrices I, X, and X 2 of the Pauli group, where X is the shift matrix
in (1). The eigenstates are the mutually orthogonal stabilizer states
(1, 1, 1), (1, ω, ω 2) and (1, ω 2 , ω), with ω the third root of unity.
Figure 1. (a) Triples of mutually orthogonal rays arising from qutrit permutation gates, (b) three pentagons
originating from the rays. Big black bullets are for stabilizer states, small bullets are for magic states.
The permutation group isomorphic to the symmetric group S3 exists
in three copies. One of them is generated as S3 = h(1, 2, 3), (2, 3)i and
1 0 0
contains the elements found in Z3 and three extra ones 0 0 1 ≡ (2, 3),
0 1 0
0 1 0
0 0 1
0 1 0 ≡ (1, 3) and 1 0 0 ≡ (1, 2), that do not lie in Pauli group but
0 0 1
1 0 0
are parts of the Clifford group. The 5 mutually orthogonal triples
between the 12 eigenstates are pictured in Fig. 1a. Apart from the 6
stabilizer states, there are 6 magic states of type (0, 1, 1) and (0, 1, −1).
In [28], they are identified as the Norrell states and strange states,
8
MICHEL PLANAT† AND RUKHSAN-UL-HAQ‡
respectively, with Wigner function.
1 2 −1 −1
1 −2 1 1
(10)
W(0,1,1) =
1 1 1
1 1 1 .
, W(0,1,−1) =
6 1 1 1
6 1 11
As expected, all magic states contains some negative entries in their
Wigner matrix.
In Fig 1b, it is shown that three pentagons are part of the orthogonality relations. Taking the rank 1-projectors associated to the vertices/states, the exclusivity graph attached to each pentagon is also
a pentagon and allows a state-dependent proof of contextuality. One
observes that only magic states of the strange type are involved.
3.3. Magic states from two-qubit permutation gates. From now
we restrict to permutation groups whose two generators are magic
gates. This only happens for two groups both isomorphic to the alternating group A4 . One copy is as follows
1 0 0 0
0 1 0 0
0
0
0
1
0
0
1
0
∼
(11)
A4 =
, 1000
.
0 1 0 0
0 0 1 0
0 0 0 1
Figure 2. (a) Maximum cliques of orthogonal rays from
two-qubit permutation gates: 6 4-tuples (thin lines) and
4 triples (thick lines), ω = exp( iπ3 ), Missing coordinates
are straightforward to recover. (b) Maximum cliques of
orthogonal rays from 5-dit permutation gates: 7 5-tuples
(thin lines) and 5 4-tuples (thick lines). Big black bullets
are for stabilizer states, small black bullets are for magic
states.
THE MAGIC OF UNIVERSAL QUANTUM COMPUTING WITH PERMUTATIONS
9
Looking for joined eigenstates shared by at least two commuting
gates, a set of 20 states is derived whose orthogonality graph is pictured
in Fig. 2a.
Wigner functions for magic states are of two types
(12)
√ !
1 −1 2 2
−1
1
1
1+2√3
1 −1 1 0 0
1
1
−1
3
3−2√3
W(0,1,1,1) =
, W(0,1,−ω,ω−1) =
,
2 0 4 0
1√
3√
−1√ 3−2 3
12 2 0 0 0
24
1−2 3 3+2 3 3+2 3
3
t
W(0,1,−ω,ω−1)
.
with W(0,1,ω−1,−ω) =
Both types of Wigner matrices contain negative entries as for the qutrit case (while in this particular case,
the stabilizer states selected from permutation gates do not).
It can be shown that the orthogonality graph contains 24 pentagons
and thus 24 state-dependent proofs of contextuality. Surprisingly, the
vertices of such pentagons either are stabilizer states or magic states of
the second type, analogous to qutrit strange states.
3.4. Magic states from 5-dit permutation gates. We again restrict restrict to permutation groups whose two generators are magic
gates. This happens for permutation groups isomorphic to the semidirect product Z5 ⋊ Z4 or to the symmetric group S5 , respectively.
As for the first case, taking all triples of mutually commuting/compatible
permutation gates, one gets a set of 30 eigenstates. They are organized
into maximum cliques shown in Fig. 2b with 20 of them being magic.
Missing coordinates are in the cyclotomic field Q[exp( 2iπ
)].
4
(13)
W(0,1,1,1,1) =
4 −1 −1 −1 −1
1 3
3
20 3
−′
−′
+
+
−′ + + −′
−′ + + −′
3 + −′ −′ +
, W(0,1,−1,−1,1) = 1
20
4
−1
−1
−1
−1
−1
−′′
+′′
+′′
−′′
−1
+′′
−′′
−′′
+′′
−1
+′′
−′′
−′′
+′′
−1
−′′
+′′ ,
+′′
−′′
√
where
the
symbol
‘±’
means
±(1
+
5)/2, the symbol ‘±′ ’ means
√
√
±( 5 − 1)/2 and ‘±′′ ’ means (3 ± 5)/2.
As for the group S5 , one gets 50 eigenstates (shared by the 10 maximum cliques of 5 mutually compatible gates). We do not show the
organization of eigenstates but only mention that one arrives at other
types of magic states of the form
(14)
W(0,0,0,1,±1)
1
=
10
0
−2
0
1
1
0
±′
0
1
1
0
∓
0
1
1
0
∓
0
1
1
0
±′
0
1
1
!
, W(0,0,1,1,1)
1
=
15
2
2
1
3
1
+′
+′
1
−′
1
−
−
1
+
1
−
−
1
+
1
+′
+′
1
−′
1
!
.
As before, state-dependent contextuality may be revealed from the
pentagons that are built from the states.
3.5. Magic states from qubit-qutrit permutation gates. The
smallest permutation group generated by two magic permutation gates
10
MICHEL PLANAT† AND RUKHSAN-UL-HAQ‡
in dimension 6 is the alternating group A5 . There exist maximum
cliques of mutually compatible permutation gates whose size is 2, 3
and 4 giving rise to shared eigenstates (stabilizer and magic) defined
over the cyclotomic field Q[exp( 2iπ
)]. As expected for even dimensions,
6
the negativity in the entries of the Wigner function is not the sign of
magicity of the state, e.g. for the stabilizer state (1, −1, 1, −1, 1, −1)
one gets
0 0 0 0 0 0
2 2
1 03 00 00 −1
−1 2 2
.
(15)
W(1,−1,1,−1,1,−1) =
0 0 0 0 0 0
18 0 0 0 −1 2 2
3 0 0 −1 2 2
A representative magic state is (0, 1, 1, 1, 1, 1) with Wigner function
4 −2 −2 3 + −
3
(16)
W(0,1,1,1,1,1)
−
+
0
0
0
−′′
+′
1
1 −′ +′′
1 −1 −′′ +′ −1 +′
=
5 −1 −1 5 −′′
30
5 +′ −′′ 1 1
1 +′′ −′
,
√
√
′
where in (16) the notation‘±’
means
(3±
3)/2,
‘±
’
means
±(
3+1)/2
√
and ‘±′′ ’ means ±( 3 − 1)/2.
Another permutation group generated by two magic permutation
gates is the alternating group A6 giving rise to magic states such as
(0, 0, 1, 1, ±1, ±1). One gets
1 − +′ 1 +′ −
(17)
W(0,0,1,1,1,1)
2 2
1 −1
′
=
4 −
+′
24
4 1
2
+
−
1
2 + −′
−1
0
4
2
2
+
0
−
−′
−1
where
√ in (17) the notation ‘±’ means ±(1 +
±( 3 − 1)/2.
−′
0
+′ ,
+
−1
√
3)/2 and ‘±′ ’ means
3.6. Magic states from 7-dit permutation gates. The smallest
permutation group generated by two magic permutation gates in dimension 7 is isomorphic to Z7 ⋊ Z6 . A representative magic state is
(0, 1, 1, 1, 1, 1, 1) with Wigner function
6 −1 −1 −1 −1 −1 −1
5 b
(18)
W(0,1,1,1,1,1,1)
c −a
b
c
c
5 c −a b
5 b c −a
1 5 c −a
=
5 −a b
42 5 −a b
−a
b
c
c
b
−a
c
−a
b
b
−a
c
b
c
−a ,
−a
c
b
where a = 2 cos(2π/7), b = −2 cos(4π/7) and c = −2 cos(6π/7) are
positive so that a negative sign in front of the matrix entries is a negative entry of the Wigner function.
Next, another permutation group generated by two magic gates is
isomorphic to P SL(2, 7), of order 168. One finds three types of magic
THE MAGIC OF UNIVERSAL QUANTUM COMPUTING WITH PERMUTATIONS
11
states whose entries are 0 or ±1 with Wigner functions as follows
0 0 0 0 0 0 0
2
W(0,0,0,0,1,1,1) =
W(0,0,0,0,0,1,±1)
2
4
(19)W(0,0,0,1,1,±1,±1) =
1
28
2
1
1
1
1
a −b −c −c −b a
a −b −c −c −b a
0
0
0
0
0
0 ,
1
1
1
1
1
1
3 1−b 1−c 1+a 1+a 1−c 1−b
1 1
1
1
1
1
1
0 0 0 0 0 0 0
0 0 0 0 0 0 0
1 ±2 ±a ∓b ∓c ∓c ∓b ±a
= 14
,
0 0 0 0 0 0 0
0 0 0 0 0 0 0
1 1 1 1 1 1 1
1 1 1 1 1 1 1
−a
−b
−c
−c
−b
a
±(b−1) ±(c−1) ∓(1+a) ∓(1+a) ±(c−1) ±(b−1)
−a
−b
−c
−c
−b
a
,
1
1
1
1
1
1
b∓1
1∓c
1±a
1±a
1∓c
b∓1
b∓1
1∓c
1±a
1±a
1∓c
b∓1
1
1
1
1
1
1
1
21
2
01
where 1 − b > 0, 1 − c < 0 and 1 − a < 0.
3.7. Higher dimensions. Table 2 summarizes the magic states found
from permutation gates of small dimensions d ≤ 9. For dimensions
d ≥ 5, only magic states with entries 0 or ±1 are considered. Column
3 provides the sum of negative entries in the Wigner matrix (5). It
is shown in [28] that the absolute value sn(ρ) of the sum of negative
entries in the discrete Wigner matrix Wρ is a computable magic monotone. It is a quantum computing resource that does not increase under
stabilizer operations. Similarly the so-called mana M(ρ) = log[(2sn(ρ])
is an (easily computable) additive magic monotone. According to [28,
Theorem. 14], a stabilizer protocol succeeds probabilistically to produce m copies of the target state σ from at least m M(σ)
copies of the
M(ρ)
state ρ on average.
4. Conclusion
We described the leading role played by permutations in shaping a
type of universal quantum computation based on magic states -states
living outside the stabilizer polytope defined by the (generalized) Pauli
group. In Sec. 3, we derived the main magic states ρ defined from
eigenstates of gates in magic permutation groups (a subset of Clifford
group) of small dimensions d ≤ 9. We explicitly computed the sum
negativity of the discrete Wigner matrix Wρ for estimating the value of
a magic state as a resource for universal quantum computation [22]. We
observed that state dependent quantum contextuality (through building block pentagons) occurs in all dimensions d ≥ 3 from appropriate
sets of stabilizer and magic states. It is desirable to extend the calculations to higher dimensions d > 9 to check a possible asymptotic trend
between the dimension and the amount of magicity and contextuality,
to clarify the relation between magic permutation groups and unitary
2-designs (singled out in [11]) and finally to relate the derived magic
states to distillation procedures and error correcting codes. A later
12
dim
2
3
4
5
6
7
8
9
MICHEL PLANAT† AND RUKHSAN-UL-HAQ‡
magic state ρ
sum of negative
entries Wρ
√
|Hi
(1 − 2)/4 ∼ −0.1035
|T i
positive
(0, 1, 1)
-1/3
(0, 1, −1)
-1/3
(0, 1, 1, 1)
−1/6
√
(0, 1, −ω, ω − 1)
(2 − 3 3)/12
∼ −0.266
√
(0, 1, 1, 1, 1)
− 5/5 ∼ −0.447
(0, 1, −1, −1, 1)
−2/5
√
(0, 0, 0, 1, ±1)
−( 5 +
1)/10
∼
−0.324
√
(0, 0, 1, 1, 1)
−(1√
+ 3 5)/15 ∼ −0.514
(0, 1, 1, 1, 1, 1)
−(3 √
3 + 7)/30 ∼ −0.406
(0, 0, 1, 1, 1, 1)
−(
√ 3 + 1)/6 ∼ −0.455
(0, 0, 1, −1, −1, 1)
−( 3 + 4)/12 ∼ −0.478
(0, 1, 1, 1, 1, 1, 1)
−[1 + 4 cos(2π/7)]/7 ∼ −0.499
(0, 0, 0, 0, 1, 1, 1)
[2 + 8 cos(4π/7) + 12 cos(6π/7)]/21 ∼ −0.504
(0, 0, 0, 0, 0, 1, ±1)
2[cos(4π/7) + cos(6π/7)]/7 ∼ −0.321
(0, 0, 0, 1, 1, 1, 1)
[1 + 4 cos(4π/7) + 10 cos(6π/7)]/14 ∼ −0.636
(0, 0, 0, 1, 1, −1, −1)
[2 + 8 cos(4π/7) + 10 cos(6π/7)]/14 ∼ −0.628
(0, 1, 1, 1, 1, 1, 1, 1)
−37/112 ∼ −0.330
(0, 0, 0, 0, 0, 1, 1, 1, 1)
−1/2
(0, 1, 1, 1, 1, 1, 1, 1, 1)
−5/9 ∼ −0.555
(0, 0, 0, 0, 1, 1, 1, 1, 1)
−2/3
(0, 0, 0, 0, 0, 0, 0, 1, 1)
−1/3
(0, 0, 0, 1, −1, 1, −1, 1, −1)
−19/27 ∼ −0.704
Remark
[4]
[4]
Norrell [22]
strange [22]
A4
Z5 ⋊ Z4
S5
A5
A6
Z7 ⋊ Z6
P SL(2, 7)
Z32 ⋊ Z7
Z23 ⋊ Z4
Z23 ⋊ Z8
G144
Table 2. The magic states ρ (column 2) and the sum
of negative entries in the Wigner matrix Wρ (column 3).
For d > 3, column 4 provides the permutation group
under consideration. The group G144 is isomorphic to
Z23 ⋊ (Z2 ⋊ D4 ), where D4 is the eight element dihedral
group.
paper will be devoted to the POVMs obtained from the Pauli group
action on the magic states [29].
References
[1] R. P. Feynman, Simulating physics with computers, Int. J. Theor. Phys. 21
467-488 (1982).
[2] F. Deutsch, Quantum theory, the Church-Turing principle and the universal
quantum computer, Proceedings of the Royal Society of London A400 97117(1985).
[3] M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information (Cambridge University Press, 2007).
[4] S. Bravyi and A. Kitaev, Universal quantum computation with ideal Clifford
gates and noisy ancillas, Phys. Rev. A71 022316 (2005).
[5] Quantum information processing and communication: strategic report on current status, visions and goals for research in Europe, June 1005, pp. 29-43 [also
P. Zoller et al, Eur. Phys. J. D36, 203-228 (2005)].
THE MAGIC OF UNIVERSAL QUANTUM COMPUTING WITH PERMUTATIONS
13
[6] A Quantum Information Science and Technology Roadmap Part 1: Quantum
Computation Report of the Quantum Information Science and Technology
Experts Panel, available at qist.lanl.gov/pdfs/rm intro.pdf (2012).
[7] D. Gottesman, A theory of fault tolerant quantum computation, Phys. Rev.
A57 127 (1998).
[8] M. Howard, J. Wakkman, V. Veitch and J. Emerson, Nature 510 351-355
(2014).
[9] J. Bermejo-Vega, N. Delfosse, D. E. Browne, C. Okay and R. Raussendorf, Contextuality as a resource for qubit quantum computation, Preprint 1610.08529
(quant-ph).
[10] R. W. Spekkens, Negativity and contextuality are equivalent notions of nonclassicality, PRL 101 020401 (2008).
[11] H. Zhu, Permutation symmetry determines the discrete Wigner function, Phys.
Rev. Lett. 116 040501 (2016).
[12] Z. Gedik, I. A. Silva, B. Çakmak, G. Karpat, E. L. G. Vidoto, D. O. SoaresPinto, E. R. deAzevedo and F. F. Fanchini, Computational speed-up with a
single qudit, Sci. Rep. 5 14671 (2015).
[13] M. Planat, Geometry of contextuality from Grothendieck’s coset space, Quantum Information Processing, 14 2563-2575 (2015).
[14] W. K. Wootters, A Wigner-function formulation of finite-state quantum mechanics,Ann. Phys. 176 1-21 (1987).
[15] C. Ferrie, Quasi-probability representations of quantum theory with applications to quantum information science, Rep. Prog. Phys. 74 116001 (2011),
24pp.
[16] A. Mann, P. E. Mello and M. Revzen, A family of Weyl-Wigner transforms for
discrete variables defined in a finite-dimensional Hilbert space, Quant. Stud.:
Math. Found. (2016), in press, doi:10.1007/s40509-016-0096-4.
[17] D. Gross, Hudson’s theorem for finite dimensional quantum systems, J. Math.
Phys. 47 122107 (2006).
[18] B. W. Reichardt, Quantum universality from magic states distillation applied
to CSS codes, Quant. Inf. Proc. 4 251-264 (2005).
[19] A. M. Meier, B. Eastin and E. Knill, Magic-state distillation with the fourqubit code, Quant. Inf. Comp. 13 195-209 (2013).
[20] E. T. Campbell, H. Anwar and D. E. Browne, Magic-state distillation in all
prime dimensions using quantum Reed-Muller codes, Phys. Rev. X2 041021
(2012).
[21] M. Howard and J. Vala, Qudit versions of the π/8 gate, Phys. Rev. A86
022316(2012).
[22] V. Veitch, C. Ferrie, D. Gross and L. Emerson, Negative quasi-probability as
a resource for quantum computation, New J. Phys. 14 113011 (2012).
[23] M. Planat, On small proofs of Bell-Kochen-Specker theorem for two, three and
four qubits, Eur. Phys. J. Plus 8 86 (2012).
[24] A. Cabello, M. Kleinmann and J. R. Portillo, Quantum state-independent
contextuality requires 13 rays, J. Phys. A: Math. Theor. 49 38LT01 (2016).
[25] A. Cabello, S. Severini and A. Winter, (Non-) contextuality of physical theories
as an axiom, Preprint 1010.2163 (quant-ph).
[26] A. A. Klyachko, M. A. Can, S. Binicioglu and A. Shumovsky, Simple test for
hidden variables in spin-1 systems, PRL 101 02403 (2008).
[27] C. Cormick, E. F. Galvao, D. Gottesman, J. P. Paz and A. O. Pittenger,
Classicality in discrete Wigner function, Phys. Rev. A73 012301 (2006).
[28] V. Veitch, S. A. H. Mousavian, D. Gottesman and. J. Emerson, The resource
theory of stabilizer computation, New J. Phys. 16 013009 (2014).
14
MICHEL PLANAT† AND RUKHSAN-UL-HAQ‡
[29] M. Planat and Z. Gedik, Magic informationally complete POVMs with permutations, Preprint.
†Institut FEMTO-ST, CNRS and Université de Bourgogne/FrancheComté, 15 B Avenue des Montboucons, F-25044 Besançon, France.
E-mail address: [email protected]
‡Jawaharlal Nehru Centre for Advanced Scientific Research, Jakkur
Bengaluru, India.
E-mail address: [email protected]
| 4math.GR
|
Preprint submitted to Energy and Buildings, November 2013
Pseudo Dynamic Transitional Modeling of Building Heating Energy Demand Using Artificial
Neural Network
Subodh Paudel
a,b,c
b
c
, Mohamed Elmtiri , Wil L. Kling , Olivier Le Corre
a*
,
Bruno Lacarrière
a
a
Department of Energy System and Environment, Ecole des Mines, Nantes, GEPEA, CNRS, UMR
6144, France
b
Environnement Recherche et Innovation, Veolia, France
c
Department of Electrical Engineering, Technische Universiteit Eindhoven, Netherlands
*Corresponding author. Tel.: +33 2 51 85 82 57
E-mail: [email protected]
Abstract
This paper presents the building heating demand prediction model with occupancy profile and
operational heating power level characteristics in short time horizon (a couple of days) using artificial
neural network. In addition, novel pseudo dynamic transitional model is introduced, which consider
time dependent attributes of operational power level characteristics and its effect in the overall model
performance is outlined. Pseudo dynamic model is applied to a case study of French Institution
building and compared its results with static and other pseudo dynamic neural network models. The
results show the coefficients of correlation in static and pseudo dynamic neural network model of 0.82
and 0.89 (with energy consumption error of 0.02%) during the learning phase, and 0.61 and 0.85
during the prediction phase respectively. Further, orthogonal array design is applied to the pseudo
dynamic model to check the schedule of occupancy profile and operational heating power level
characteristics. The results show the new schedule and provide the robust design for pseudo dynamic
model. Due to prediction in short time horizon, it finds application for Energy Services Company
(ESCOs) to manage the heating load for dynamic control of heat production system.
Keywords: Building Energy Prediction; Short term building energy forecasting; Operational Heating
Characteristics; Occupancy Profile; Artificial Neural Network; Orthogonal Arrays
1. Introduction
The global concerns of climate change and regulation in energy emissions have drawn more
attention towards researchers and industries for the design and implementation of energy systems for
low energy buildings. According to IEA statistics [1], total energy use globally accounts for around
7200 Mtoe (Mega Tonnes Oil Equivalents). Residential and commercial buildings consume 40% of
final energy use in the world and European countries consume 76% of energy towards thermal comfort
in buildings. The small deviations in design parameters of buildings could bring large adverse effect in
the energy efficiency and which, additionally, results in huge emissions from the buildings. It is
estimated that improvement in energy efficiency of the buildings in European Union by 20% will result
in saving at least 60 billion Euro annually [2]. So, research is very active in driving towards the
sustainable/low energy buildings. In order to accomplish this and to ensure thermal comfort, it is
essential to know energy flows and energy demand of the buildings for the control of heating and
cooling energy production from plant systems. The energy demand of the building system, thus,
depends on physical and geometrical parameters of buildings, operational characteristics of heating
and cooling energy plant systems, weather conditions, appliances characteristics and internal gains.
There
are
various
approaches
to
predict
building energy demand based on physical methods
and data-driven methods (statistical and regression methods and artificial intelligence methods) as
mentioned by Zhao et al. [3]. Physical methods are based on physical engineering methods and uses
thermodynamics and heat transfer characteristics to determine the energy demand of the building.
There are numerous physical simulation tools developed as EnergyPlus [4], ESP-r [5], IBPT [6],
SIMBAD [7], TRNSYS [8], CARNOT [9] etc… to compute the building energy demand. A simplified
physical model based on physical, geometrical, climatic and occupant model was presented by
Duanmu et al. [10] to bridge the complexities of collecting more physical data required in simulation
tools. Other possible approaches for building energy prediction are semi-physical models like
response factor method, transfer function method, frequency analysis method and lumped method
[11]. Though methodologies adapted to estimate energy demand of buildings are different in physical
and semi-physical models, both are highly parameterized. In addition, physical parameters of buildings
are not always known or even sometimes data are missing. And also, these models are
computationally expensive for Energy Services Company (ESCOs) to manage heating and cooling
loads for control applications.
Other approaches to predict building energy demand with limited physical parameters are
data-driven methods, which strongly dependent on the measurements of historical data. Statistical and
regression methods seem more feasible to predict building energy demand with limited physical
parameters. The statistical approaches have been widely used by Girardin et al. [12] to determine the
best model parameters by fitting actual data. Different approaches (physical and behaviour
characteristics based on statistical data) were presented by Yao et al. [13] to bridge the gap between
semi-physical and statistical methods. In their work, statistical daily load profile was grounded on
energy consumption per capita and human behaviour factor, and semi-physical method was based on
thermal resistance capacitance network. Nevertheless, these statistical models used linear
characteristics of input and output variables to evaluate the building parameters and are not adapted
to non-linear energy demand behavior. Regression models [14-15] have also been used to predict the
energy demand, but, they are not accurate enough to represent short term horizon (couple of days)
with hourly (or couple of minutes) sampling time energy demand prediction. In order to find the best
fitting from the actual data, this kind of models requires significant effort and time.
In recent years, there is a growth in research work in the field of artificial intelligence (AI) like
artificial neural network [3, 16] and support vector machines [3, 17-18]. These methods are known for
solving the complex non-linear function of energy demand models with limited physical parameters.
Neural network method has shown better performances than physical, statistical and regression
methods. Authors [19-20] used static neural network to predict energy demand of the building and
compared results with physical models. For instance, Kalogirou et al. [19] used climate variables
(mean and maximum of solar radiation, wind speed, and other parameters as wall and roof type)
coupled with artificial neural network (ANN) to predict daily heating and cooling load of the buildings. In
their work, results obtained using ANN are similar to those given by the physical modelling tool
TRNSYS. Neto et al. [20] presented a comparison of neural network approach with physical simulation
tool EnergyPlus. In this work, authors used climate variables as external dry temperature, relative
humidity and solar radiation as input variables to predict daily consumption of the building. Results
showed that neural network is slightly more accurate than EnergyPlus when comparing with real data.
Static neural network model proposed by Shilin et al. [21] consider climate variables as dry bulb
temperature and information regarding schedule of holiday’s to predict cooling power of residential
buildings. Dong et al. [17] used support vector machine (SVM) to predict the monthly building energy
consumption using dry bulb temperature, relative humidity and global solar radiation. Performance of
SVM and neural network model wee compared and results show that SVM was better than neural
network in prediction.
Various authors [22-26] performed hourly building energy prediction using ANN. Mihalakakou
et al. [22] performed hourly prediction of residential buildings with solar radiation and multiple delays of
air temperature predictions as input variables. Ekici et al. [23] used building parameters (window’s
transmittivity, building’s orientation, and insulation thickness) and Dombayci [24] used time series
information of hour, day and month, and energy consumption of the previous hour to predict the hourly
heating energy consumptions. Gonzalez et al. [25] used time series information hour and day, current
energy consumption and predicted values of temperature as input variables to predict hourly energy
consumption of building system. Popescu et al. [26] used climate variables as solar radiation, wind
speed, outside temperature of previous 24 hours, and other variables as mass flow rate of hot water of
previous 24 hours and hot water temperature exit from plant system to predict the space hourly heat
consumptions of buildings. Li et al. [18] used SVM to predict hourly cooling load of office building using
climate variables as solar radiation, humidity and outdoor temperature. In their work, SVM was
compared with static neural network and result showed SVM better than static neural network in terms
of model performance. Dynamic neural network method which includes time dependence was
presented by Kato et al. [27] to predict heating load of district heating and cooling system based on
maximum and minimum air temperature. Kalogirou et al. [28] used Jordan Elman recurrent dynamic
network to predict energy consumption of a passive solar building system based on seasonal
information, masonry thickness and thermal insulation.
For many authors [29-31] occupancy profile has a significant impact on building energy
consumption. Sun et al. [29] mentioned that occupancy profile period has a significant impact on initial
temperature requirement in the building during morning. In their work, reference day (the targeted day
prediction which depends on previous day and beginning of following day based on occupancy and
non-occupancy profile period) was calculated based on occupancy profile period. In addition to this
value, correlated weather data and prediction errors of previous 2 hours were used as input variables
to predict hourly cooling load. Yun et al. [30] used ARX (autoregressive with exogeneous i.e., external,
inputs) time and temperature indexed model with occupancy profile to predict hourly heating and
cooling load of building system and compared this with results given by neural network. Results
showed that occupancy profile has a significant contribution in determination of auto regressive terms
during different intervals of time and further showed a variation of it in the building heating and cooling
energy consumption. The proposed ARX model showed similar performance with neural network.
Sensitivity analysis for heating, cooling, hot water, equipment and lighting energy consumption based
on occupancy profile was performed by Azar et al. [31] for different sizes of office buildings. In their
work, they found that heating energy consumption has the highest sensitivity compared to cooling, hot
water, equipment and lighting energy consumption for small size buildings. Also, results showed that
heating energy consumption is highly influenced by occupancy profile for medium and small buildings
during the occupancy period. Moreover, few literatures focused on operational power level
characteristics (schedule of heating and cooling energy to manage energy production from plant
system). For example, Leung et al. [32] used climate variables and operational characteristics of
electrical power demand (power information of lighting, air-conditioning and office equipment which
implicitly depends on occupancy schedule of electrical power demand) to predict hourly and daily
building cooling load using neural network.
In conclusion, it can be reiterated that physical and semi-physical models [4-11], though give
precise prediction of building energy, they are highly parameterized and are computationally expensive
to manage the energy for control applications for ESCOs. Data-driven methods which depend on
measurement historical data are not effective during the early stage of building operation and
construction since measurement data are not available at these stages. When building energy data
are available, data-driven methods can be considered if measurement data are accurate and reliable
as this kind of models can be sensitive on the quality of measured data. Sensitivity of the accuracy of
data driven models, thus, depends on the measurement data. Data-driven models based on statistical
and regression methods [12-15, 26] cannot precisely represent short time horizon (couple of days)
with hourly (or couple of minutes) sampling time prediction, though they perform prediction of energy
consumptions of buildings with limited physical parameters. They also require significant efforts and
time to compute the best fitting of the actual data. Static neural network models [19-21] are used for
daily prediction and [22-25] are used for hourly prediction of the buildings energy consumptions.
Though dynamic neural network model [27-28] gives better precision in compared to static neural
network, they do not consider occupancy profile and operational power level characteristics of the
plant system and therefore not adapted for the ESCOs to manage energy production for control
applications. The important features like transition and time dependent attributes of operational power
level characteristics of the plant system are still missing, though, authors [29-30] consider occupancy
profile and author [32] considers operational characteristics of electrical power demand. The detailed
variables and application of models developed in the literature reviews are summarized in Table 1.
Table
1
:
Summary of variables and application models in the literature
Input Variable of Model
Climate Variables
Author and Year
Type of Model
Outside Tempeature
Inner
Temperature
Ambient Dry Bulb Wet Bulb
Girardin et al. (2009)
Yao et al. (2005)
Catalina et al. (2008)
Wan et al. (2012)
Statistical
Thermal and
Statistical
Regression
Horizon of
Forecast
Annually
√ (2*)
Daily
√
Regression
√
√
Static NN
Mihalakakou et al. (2002)
Ekici et al. (2009)
Dombayci (2010)
Gonzalez et al. (2005)
Popescu et al. (2009)
Kato et al. (2008)
Kalogirou et al. (2000)
Li et al. (2010)
Static NN
Static NN
Static NN
Static NN
Static NN
Dynamic NN
Dynamic NN
SVM
√(4*)
Regression
√
Autoregressive
with exogeneous
√
√
√
√
√
√
√
√
√ (3*)
√
√
√
√
Monthly
Monthly &
Yearly
Monthly
Daily
Daily
Daily
√
√ (5*)
√ (6*)
√ (7*)
√ (8*)
√
√
√
√
√(9*)
√ (10*)
√(11*)
Static NN
Physical
Occupancy
Operational
Other
Profile
Characteristics Parameters
√ (1*)
Shilin et al. (2010)
Leung et al. (2012)
Duanmu et al. (2013)
Relative
Humidity
√
SVM
Static NN
Static NN
Yun et al. (2012)
Wind
Speed
√
Dong et al. (2005)
Kalogirou et al. (2001)
Neto et al. (2008)
Sun et al. (2013)
Global
Solar
Radiation
√(11*)
√
√
√
√
√
√
√
√
√
√
√
√
√
√ (12*)
√
√
√ (14*)
√
√ (13*)
Hourly
Hourly
Hourly
Hourly
Hourly
Hourly
Hourly
Hourly
Hourly
Hourly
√
√ (15*)
√ (16*)
Hourly &
Daily
Hourly
Type of Applications for Buildings
80 Residential
(heating and cooling)
Residential (space heating)
Residential
Office (heating and cooling)
4 Buildings (total energy consumptions)
9 Buildings (heating and cooling)
Office (3000 m2)
Residential (cooling power)
Residential (200 m2)
Heating Energy of Buildings
Residential (heating energy)
Electrical load
8 Buildings
District (heating energy)
Passive solar buildings
Office building and library
Cooling load for high rise
buildings (440,000 m2)
Small building for
heating load (464 m2)
Office (space electrical power
demand)
Cooling load of buildings
Remarks:
1*: Nominal Temperature of heating, cooling and hot water system; Threshold heating and cooling temperature
2*: Appliances Model
3*: Climate Index based on principal component
4*: Multiple lag output predictions of ambient air temperature
5*: Transmittivity, orientation and insulation thickness
6*: Heating degree hour method
7*: Predict value of temperature, present electricity load, hour and day
8*: Outside temperature and mass flow rate in previous 24 hour, hot water temperature
9*: Highest and Lowest open air temperature
10*:Season, insulation, wall thickness, heat transfer coefficient
11*: Multiple lag of dry bulb temperature and solar radiation
12*: Reference day of each day based on occupancy schedule
13*: Correlated weather data based on reference day and accuracy of calibrated prediction error of previous 2 hours
14*: Occupancy profile represented by space electrical power demand
15*: Clearness of sky, rainfall, cloudiness conditions
16*: Physical and geometrical parameters, hourly cooling load factor
None of these studies has evaluated the transition and time dependent effects of operational
power level characteristics of heating plant system and has predicted building heating energy demand
in short time horizon (a couple of days). This short term prediction is important to ESCOs for dynamic
control of heat plant system. This paper bridges the gap between static and dynamic neural network
methods with occupancy profile and operational power level characteristics of heating plant system. It
introduces novel pseudo dynamic model, which incorporates time dependent attributes of operational
power level characteristics. Their effects on neural network model performances are compared to
static neural network for building heating demand. Orthogonal arrays are applied to the proposed
pseudo dynamic model for robust design and confirmed the new schedule of occupancy profile and
operational heating power level characteristics obtained from ESCOs. The proposed method allows
short term horizon prediction (around 4 days with sampling interval of 15 minutes) to make decision
(e.g. management of wood power plant) for the ESCOs. The next section describes methodology
including scope of study, design of transitional and pseudo dynamic characteristics, neural network
model and orthogonal arrays. Finally, a case study is presented and results and discussion are drawn
to analyze the performance of different static and pseudo dynamic models along with robustness of
proposed pseudo dynamic model for heating demand prediction of the building.
2. Methodology
Collection of climate
data
Collection of Building
Heating Energy data
Operational heating
power level
characteristics
Settling and Steady
State time of the control
system
Transitional
and
Pseudo Dynamic
Model
Occupancy
Profiles
Artificial Neural Network
Static and Pseudo dynamic model for heating demand
prediction
The development and implementation of models proposed in this work are based on collection of
real building heating demand, operational heating power level characteristics, climate variables and
approximated occupancy profile data (see Appendix A for selection of relevant input variables). An
outline of the methodology presented in this paper is shown in figure (1). The input of this methodology
is in form of time-series climate and building heating energy data. The other inputs data are occupancy
profile and operational heating power level characteristics for working and off-days for 24 hours.
Dynamics of building heating demand is also an input to the methodology which includes settling and
steady state time and is estimated from real building data. Based on operational heating power level
and dynamics of building characteristics, transitional and pseudo dynamic models are designed.
Finally, neural networks for static and pseudo dynamic models are designed to predict heating
demand in short time horizon (couple of days). For the robustness of pseudo dynamic model,
occupancy profile and operational power level characteristics are analyzed for different time intervals
to confirm occupancy schedule profile and operation of plant system from the orthogonal arrays. The
pseudo dynamic model after optimum orthogonal arrays design is used for final prediction of the
building heating demand. Scope of this study, details of transitional and pseudo dynamic model,
neural network model and orthogonal arrays are described in section 2.1 - 2.4.
Figure 1: Outline of the proposed methodology on heating demand prediction
2.1 Scope of Study
The scope of this paper is heating demand prediction in short time horizon for the large building. The
overall objective is to make an energy services decisions (e.g. management of wood power plant) for
ESCOs. The assumptions carried for this study are highlighted as:
1. Winter period is studied.
2. Existing building is considered and space heating demand of this building is fed up from a heat
network to a central substation. Domestic hot water (DHW) is out of the scope.
3. The heating demand data was recorded in data acquisition system database and thermal
comfort inside the building was performed in this database. Thus, the effects of ventilation and
air-conditioning on heating are already included in this database.
4. Simple occupancy profile of building is anticipated approximately to assist the ESCOs to
schedule their heat production system. In such a system, individual occupant’s behavior or
precise occupancy profile is not considered. Thus, the modeling constraints are closer to the
operational condition of ESCOs to estimate the heat demand.
5. The wind speed and direction are not taken into consideration. This is due to the fact that
present weather variables data are taken from data acquisition system but future weather
variables values are coming from an atmospheric modeling system which mesh size can be
15 km (as ARPEGE, see [33]), 10 km (as ALADIN, see [34]) or 2.5 km (as AROME, see [35]).
In such a case, wind impact on heating demand prediction of a specific building located inside
the mesh is very difficult or even impossible to consider for precise effect. Further, heating
energy demand is highly dependent on outside temperature and other climate variables have
less significant impact on heat energy [36].
2.2 Transitional and Pseudo Dynamic Model
The operational heating power level characteristics gives operational features of the plant system,
however, they do not give abstract information about transition attributes of operational heating power
level which is illustrated through an example in figure (2). The y-axis represents set up power level
from the production system and x-axis represents operation schedule.
100
3' State 1
3
90
4
4'
7'
7
α 43
State 3
8'
8
α 87
40
Transition 3
Transition 2
Transition 1
Transition 0
Power Level (%)
75
State 2
5' 6'
5
α 65
6
25
1
10
1'
State 0
State 4
2'
2
α 21
0
9
6
12
Hour
14
20
9' α 109 10'
10
24
Figure 2: Operational heating power level characteristics of the plant system (for a day)
In figure (2), operational power levels are identified by different states and transition levels and
each level has its own significant effects on the operational power level characteristics. State means
consistency in the power level from one operation schedule to another and transition means change in
power level from one operation schedule to another in heat production system. The transition level 0,
1, 2 and 3 have similar feature of transitional power level characteristics on the overall operational
performance, however, power level required for transition from point 2 to 3, point 4 to 5, point 6 to 7
and point 8 to 9 is different for each level. If the power level of state 0, 1, 2, 3 and 4 in operational
heating power level characteristics is represented by the
from point
α uv , then
the power required for transition
v to point u can be represented as β uv in the transitional characteristics as shown in
figure (3). Thus, the power level transition in transitional characteristics corresponding to operational
characteristics can be written as:
β uv = β (u −2 )(v−2 ) + 2Δβ α uv − α (u −2 )(v−2 ) , ∀ u = 4,6,8....., v = 3,5,7...
where,
(1)
, v = 1, u = 2
β0
β 0 , Δβ
and
represents initial power level, step size of transition power level and absolute
values respectively. Each level ( β 21 ,
β 43 , β 65 , β 87
and
β 109 )
represents transitional level and
depends on the power level of operational characteristics.
9
β109
Transition Level
7
Transition Level
β87
β87
β109
10
8
5 β65 6
3
β43
β43
4
PDL
1
β21
0
β21
2
6
12
14
Hour
20
24
Figure 3: Transitional and Pseudo dynamic characteristics (for a day)
The transitional characteristics explicate the power transition level of operational
characteristics, however, dynamic information of power level attributes is still lacking. It means that
power content in operational characteristics of figure (2) of point 1-1’ is not equal to 2-2’; point 3-3’ is
not equal to the 4-4’; 5-5’ is not equal to 6-6’; 7-7’ is not equal to 8-8’ and 9-9’ is not equal to 10-10’.
Dynamic transition information, thus, is necessary in the model which considers dynamic
characteristics of the building. The simple first order dynamics of building characteristic is shown in
figure (4), where τ represents time constant.
Amplitude of Heating Power (%)
100
80
60
Tsteady
40
Ts
τ
20
delay
0
-20
0
τ
Time Constant
5τ
6τ
Figure 4: Dynamics of building characteristics
In figure (4), delay represents time it takes from plant system to reach the building for heating
operation and after this, power is sufficient to provide heating demand. The τ represents the 63% of
power transferred to the building heating system from plant system. Other dynamics to incorporate is
settling time ( Ts ), which is the time elapsed for heating power to reach and remain within the specified
error band and equal to [2 τ , 5 τ ] and have almost similar behavior like steady state time. The steady
state time corresponds to [3 τ , 6 τ ]. Thus, τ , settling time ( Ts ) and steady state time ( Tsteady ) gives
information about dynamic characteristics of heating demand. This dynamic information of building,
thus, depends on the transitional attributes of power level and this information is not totally dynamic
but pertaining to the appearance of dynamic behavior, so pseudo dynamic name is chosen. Thus,
pseudo dynamic is just a lag of transitional attribute information and further depends on time constant
τ or range between settling and steady state of the dynamic building heating characteristics. The
simplified pseudo dynamic lag (PDL) is calculated from equation (2), where, ts represents the
sampling time of building data and
Tu
represents the new unknown time which lies between settling
and steady state time. The concise value of
Tu depends on dynamics of the heating demand and
pseudo dynamic characteristics can be seen from figure (3), where PDL is pseudo dynamic lag.
Ts ≤ Tu ≤ Tsteady , where Ts ∈ [2τ ,5τ ]; Tsteady ∈ [3τ ,6τ ]
PDL ∈
τ
ts
[3,6]
(2)
2.3 Neural Network Model
The neural network consists of neurons to interconnect the inputs, model parameters and
activation function. Each interconnection between the neurons represents model parameters. Input-
output mapping in neural network is based on the linear and non-linear activation function. From input
and targeted data, model parameters are adjusted to minimize the error i.e. difference between actual
values and predicted values produced by the network. Learning/training of data are repeated until
there is no significant change in the model parameters and only stops the training. This type of
learning approach is called supervised learning since predicted value of the model is guided by actual
values.
There are numerous ANN model like Feed-forward Multilayer Perceptron (MLP), Radial Basis
Function (RBF) Network, Recurrent Network and Self-Organizing Maps (SOM) [37]. All of these
networks have their own learning algorithm to learn and generalize the network. In this paper, MLP is
taken as a neural network model since pseudo dynamic model is not fully dynamic (in time behavior).
There are two ways of learning mechanism in the neural network: sequential learning and batch
learning. In sequential learning, cost function is computed and model parameters are adjusted after
each input is applied to the network. In batch learning, all the inputs are fed to the network before
model parameters are updated. In batch learning, model parameter adjustment is done at the end of
epoch (one complete representation of the learning process) and for this paper, batch learning is
carried out.
MLP network consists of three layers: input layer, hidden layer and output layer and there can
exist more than one hidden layer. However, according to the Kolmogorov’s theorem [38], single hidden
layer is sufficient to map the function provided suitable hidden neurons and for this paper, single
hidden layer is used as shown in figure (5). The hidden layer assists to solve non-linear separable
problems.
Figure 5: Neural Network Architecture
In figure (5), xi , wk and
neuron which varies from
lag of 1 and z
-M
y represents input neuron which varies from i = 0 to i = q , hidden
k = 0 to k = p and output neuron respectively. The z-1 signifies transition
signifies transition lag corresponding to PDL, where maximum value of M ( M max )
equals to PDL i.e.
M max = {1,2,.....PDL}. The MLP uses logistic function or hyperbolic tangent as a
threshold function in the hidden layer. It has been identified empirically [39] that network using logistic
functions tends to converge slower than hyperbolic tangent activation function in the hidden layer
during the learning phase. Hyperbolic tangent activation functions is chosen in the hidden layer and
pure linear activation function is chosen in the output layer for this paper and hyperbolic tangent
T
function is shown in equation (3), where θ represents model parameter with transpose of matrix.
Division of input and output data into learning, validation and testing gives more generalization of
model. Learning data sets are used to learn the behavior of input data and to adjust the model
parameters. Validation data is used to minimize the overfitting. It is not used to adjust the model
parameter but it is used to verify if any increase in accuracy over learning dataset actually yields an
increase in accuracy over dataset that has not learned to the network before. Testing data sets are
used to confirm the actual prediction from neural network model which is unknown to neural network
before. For this paper, data is divided into learning, validation and testing sets. Normalization of input
data is also important for faster convergence to achieve desire performance goal. If input data are
poorly scaled during learning process, there is a risk of inaccuracy and slower convergence. It is, thus,
essential to standardize the input data before applying to neural network. There are various methods
for normalization of input and output variable, and for this paper, normalization with zero mean and
i
unit standard deviation is done as shown in equation (4). In equation (4), x , X and m represents
mean of input variable, overall vector of input variable and number of datasets respectively and thus,
applies similarly for output variable.
(
)
T
hθ ,x =
eθ
T
x
− e −θ
T
x
eθ
T
x
+ e −θ
T
x
(3)
xi − x
Xi =
(4)
1
(
xi − x )
∑
m −1 i
The cost function of MLP network is computed in equation (5):
1 m (l )
(l )
J (θ ) =
y − ya
∑
2m l =1
[
]
2
(5)
y , y a , l and J (θ ) represents predicted values produced from the network, actual values of
given datasets, individual data from m number of datasets and cost function of the neural network
model respectively. Further, y of the network is computed as:
where
p
⎛ q
⎞
y = ∑θ k h⎜⎜ ∑θ ki xi ⎟⎟
k =1
⎝ i =0
⎠
(6)
In order to update the model parameters for a higher degree approximation on unknown nonlinear function for learning process, there are different methods as – gradient descent, Newton’s
method and so on [37]. Gradient descent is too slow for the convergence, and it takes more time to
compute the hessian matrix in Newton’s method as well. Levenberg-Marquardt algorithm is used for
this paper which takes approximation of hessian matrix in the form of Newton’s method and model
parameter update equation
[
θ t +1 is given as:
]
θ t +1 = θ t − [LT L + µI ] LT J (θ )
−1
(7)
T
In equation (7), hessian matrix is approximated as [ L
L]
LT J (θ ),
parameter, µ is
and gradient is computed as
L is Jacobian matrix, J (θ ) is vector of cost function, θ t
is initial model
suitable chosen scalar and I is identity matrix. Update model parameter, thus, depends on the cost
where,
function and scalar value µ .
2.3.1
Stopping Criteria
There are different criteria for stopping the neural network model. For this paper, the stopping
criteria depend on number of epochs to learn the network, performance goal, maximum range of µ
and maximum failures in the validation. The performance goal (PG) is given as:
m
PG = 0.01 ∑ y a
(l )
(8)
l =1
The maximum failures in validation or accuracy over validation datasets is defined to stop the
learning process if the accuracy of learning datasets increase and validation accuracy stays same or
decrease.
2.3.2
Model Performance
Performances of models are characterized by mean square error (MSE) and coefficient of
2
2
correlation (R ). The MSE and R can be calculated as:
m
MSE =
∑ [y
− ya
(l )
l =1
]
(9)
m
m
R2 =
2
(l )
∑ [y ( ) − y
(l )
l
a
l =1
m
]
2
(10)
∑ (y )
(l )
2
a
l =1
2.3.3
Degree of Freedom Adjustment
One of the issues of neural network model is over learning of the network. With increase of
hidden neurons, model performance can be increased, but, it will lead neural network to over learning.
Validation accuracy and degree of freedom (DOF) adjustments are done in this paper to avoid over
fitting. Number of learning equations that model could deliver are given by equation (11), where
learning equations of the network and
Le is
L y is length of vector output neurons ( y ), and in this case
equal to 1 since there is only heating demand load.
Le = m * L y
(11)
The number of model parameters for a single hidden layer MLP neural network are given by
the equation (12), where
Lθ , L x
and
Lw represents
number of model parameters, vector length of
input neurons ( xi ) and vector length of hidden neurons ( wk ) respectively.
Lθ = (Lx + 1) * Lw + (Lw + 1) * L y
(12)
DOF of neural network model is the difference between number of learning equations and
number of model parameters in the network. It should be always >>1 and depends on the optimum
size of hidden neurons. DOF and maximum hidden neurons are given by equation (13) and (14),
where,
δ
represents the scalar constant value and depends on DOF required for design and
Wmax
is
the maximum hidden neurons.
DOF = Le − Lθ
Wmax ≅
1
(L
(13)
− Ly )
θ
(14)
δ (L x + L y + 1)
Modified performance goal according to degree of freedom adjustment is given as:
m
PG =
0.01 DOF ∑ y a
(l )
l =1
(15)
Le
Model performance is also further modified based on degree of freedom adjustment. The
2
modified MSE and R can be calculated as:
m
MSE mod ified =
l =1
(l )
]
2
(16)
DOF * m
m
R 2 mod ified =
[
Le ∑ y (l ) − y a
[
Le ∑ y (l ) − y a
l =1
m
]
2
(17)
( )
DOF ∑ y a
l =1
(l )
(l )
2
For each hidden neurons, optimal
MSE mod ified and maximum R 2 mod ified for learning and
validation are calculated from the different initialized random parameters. For different number of
hidden neurons,
R 2 mod ified
and
MSE mod ified for each model is performed for learning and validation,
and based on it, optimal configuration of model is identified for the final prediction.
2.4 Orthogonal Arrays
It is essential to know whether schedule of occupancy profile and operational characteristics
obtained from ESCOs is reliable for the robust design of pseudo dynamic model. Occupancy profile
and operational characteristics transition period, thus, plays an important role in the model
performance and if all these transition period are consider for finding the best robust model, it takes
long time to compute. Orthogonal arrays (OA) identify the main effects with minimum number of trials
to find the best design. These are applied in various fields: mechanical and aerospace engineering
[40], electromagnetic propagation [41] and signal processing [42] for the robust design model.
The orthogonal array allows the effect of several parameters to find best design with given
different levels of parameters. It can be defined as matrix with column representing number of
parameters with different settings to be studied and rows representing number of experiments. In
orthogonal arrays, parameters are called factors and parameter settings are called levels. In general,
OA(N , k , s, t ) is used to represent the orthogonal arrays, where N , k , s
and
t represents number
of experiments, number of design parameters, number of levels and strength. There are different
methods as Latin square [43]; Juxtaposition [44]; Finite geometries [45] etc... to create orthogonal
arrays with different strength and levels. Orthogonal arrays with different number of design parameter,
level, and strength are available from OA databases or libraries. The orthogonal arrays used for this
paper is taken from OA library [46].
3. Case Study
The methodology is applied for case study at Ecole des Mines de Nantes, French Institution.
2
The building has floor area of 25,000 m . It has 600 students and 200 employees. The building
consists of 120 research and administration rooms, 30 class rooms, 3 laboratories, and 8 seminar
halls. Class rooms have different sizes and can accommodate to 18 to 28 students. The 2 big seminar
halls can be occupied by 250 students and 6 small seminar halls can be occupied by 80 students.
2
Each floor area of the laboratory is 600 m .
The data is taken from data acquisition system and consists of day/month/time, solar radiation,
outside air temperature and heating demand from mid of January to February 2013 with sampling
interval of 15 minutes. The 70% of data (outside temperature, solar radiation and heating demand as
shown in figure 5) are used for learning phase i.e.
in mathematical equation in neural network, see
section 2.3, equivalent to 19 days with 15 minute sampling time, and each 15% of data (4 days with 15
minute sampling time) is used for validation and testing phase. Outside temperature taken for this
0
0
0
study has minimum, average and maximum value of 1.2 C, 8.95 C and 15.3 C respectively. Global
2
2
solar radiation has an average and maximum value of 7 W/m and 438 W/m respectively.
The simplified/theoretical occupancy profile and operational heating power level characteristics
for working and off-days for 24 hours is shown in figure (6) and (7).
1000
Working Day
Off Day
900
800
Number of Occupants
700
600
500
400
300
200
100
0
0
8
12 13:30
Hour
17:45
Figure 6: Occupancy profiles for working and off-day
24
Working Day
Off Day
100
90
Power Level (%)
75
40
25
10
0
6
12
Hour
14
20
24
Figure 7: Operational heating power level characteristics for working and off-day
Power demand and occupancy profile during working day is depicted from figure (8). From
figure (8), occupancy profile almost gives information about power demand characteristics, however,
from 18 hour onwards, power demand characteristics is not accordance with occupancy profile. Thus,
it further shows that simplified occupancy profile is not enough to characterize the heating demand.
Working Day
Off Day
100
90
Power Level (%)
75
40
25
10
0
6
12
14
Hour
20
24
Figure 8: Heating power demand and occupancy profile during working days
Different neural network models are designed based on climate variables (outside temperature
and solar radiation), work/off day information, occupancy profile and operational characteristics as
shown in figure (5). For this case study, 10 represent working day and 5 represent off day information
(work/off day) to the input of neural network model. Static neural network model 1 consists of
operational characteristics and occupancy profile, external temperature and solar radiation as input
variables and heating power demand as an output variable, and thus, vector length of input neurons
( L x ) in equation (12) equals to 5. Model 2 comprises additional transitional characteristics in model 1
and vector length of input neurons ( L x ) in equation (12) equal to 6. For this case study the sampling
time ( ts ) of real building data is 15 minutes, settling time ( Ts ) is estimated approximately 45 minutes
and steady state time ( Tsteady ) is approximately 1 hour. The PDL, thus, is calculated from equation (2),
where PDL corresponds to settling and steady state time is nearly equal to 3 and 4 respectively. Since
pseudo dynamic model depends on transition lag of operational heating power level and building
dynamic characteristics, PDL is varied from 3-4, and to understand the phenomena of pseudo dynamic
lag, PDL is varied from 1-4. Model 3 comprises model 2 with additional parameters of one PDL i.e. i.e.
L x equals to
7; model 4 consists model 2 with additional parameters of two PDL i.e.
model 5 includes model 2 with additional parameters of three PDL i.e.
L x equals
L x equals to
8;
to 9 and model 6
comprises model 2 with additional parameters of four PDL in the transitional characteristics i.e.
L x equals to 10. Transitional and
pseudo dynamic characteristic with four lags during working day is
shown in figure (9). Transition level in figure (9) is calculated from equation (1) and for this case study,
25 is chosen for each
β0
and
Δβ .
In figure (9), lag 0 means static model which contains transition
attributes, lag 1 means pseudo dynamic model with transition lag 1 (PDL=1), lag 2 means pseudo
dynamic model with transition lag 2 (PDL=2) and so on. Further, effects of transitional and pseudo
dynamic effects on the heating demand can be understood from figure (10). It is clear that the
information hidden in heating demand which climate variables could not answer can be justify from
transitional and pseudo dynamic attributes of operational characteristics. The summary of models is
shown in table (2).
No Lag
Lag 1
Lag 2
Lag 3
Lag 4
Transition Level
825
575
425
275
25
0
6
7
12 13 14 15
Hour
20 21
24
Figure 9: Transitional and pseudo dynamic characteristics during working day
Heating Demand
No Lag
Lag 1
Lag 2
Lag 3
Lag 4
1000
1000
Pseudo Dynamic
transitional effects
800
800
Heating Demand (kW)
Transition
effects
600
PDL
600
400
400
200
200
0
0
24
48
Hour
72
82
Figure 10: Pseudo dynamic transitional effects on heating demand
Table 2: Summary of models
Model No.
Type of
Model
Input Variables
Remarks
Model 1
Static
Climates, occupancy profile and operational characteristics
No Lag
Model 2
Static
Model 1 with transitional characteristics
No Lag
Model 3
Pseudo Dynamic
Model 2 with pseudo dynamic transition in dead band
Lag 1
Model 4
Pseudo Dynamic
Model 2 with pseudo dynamic transition in t
Lag 2
Model 5
Pseudo Dynamic
Model 2 with pseudo dynamic transition in settling time
Lag 3
Model 6
Pseudo Dynamic
Model 2 with pseudo dynamic transition in steady state time
Lag 4
For each model, cost function
J (θ ) in equation (5) is computed iteratively up to 1000 for each
of the minimum and maximum number of hidden neurons. The maximum number of hidden neurons is
calculated from equation (14), where δ is chosen 8 as it gives the flexibility in the degree of model
parameters. Thus, three minimum hidden neurons are chosen as 3 for this case study. Hidden
neurons length ( Lw ), thus, is varied from 3 to
Wmax . Performance of model at each iteration (number
of epochs) is computed from equation (16) and (17) and model parameters are updated based on
equation (7), where initial value of µ is chosen as 0.01 and its value is increased with a factor of 10
and decreased with a factor of 0.1. The maximum value of
µ
is chosen as 1e10. Neural network
model in this study will be stopped if the number of epochs reached to 1000 and performance goal
reached the value given by equation (15).
Under the scope of study (see subsection 2.1), the accuracy on the number of occupants are
not relevant, however, it is essential to know inside the sampling time, when the staff and students
come and leaves the buildings. It is necessary to check occupancy and operational power level
characteristics provided by ESCOs are right or not for robust design model. And, the main controlling
factors for robust design model are the transition schedule of occupancy and operational
characteristics. From figure (6), it is clear that there is no transition of occupancy during off-day, but
there is transition of occupancy during the interval at 8 hour, 12 hour, 13:30 hour and 17:45 hour and
these are represented by t1, t2, t3 and t4 factors respectively. Similarly, there is a transition of
operational characteristics for working and off day as shown in figure (7) and these transition factors
are represented by t5, t6, t7 and t8 for working day for 6 hour, 12 hour, 14 hour and 20 hour; t9 and
t10 for off day for 6 hour and 20 hour. Since the sampling interval taken for this case study is 15
minutes, three levels are used for orthogonal arrays so that the model will represent the 15 minutes
ahead and before from occupancy and operational characteristics schedule period. The summary of
control factors and their levels are shown in table (3), where OSW represents occupancy schedule at
work day, OCSW represents operational characteristics schedule at work day and OCSO represent
operational characteristics schedule at off day.
Table 3: Summary of control factors and their levels
Levels
Factors
1
2
3
OSW at 8 hour (f1)
t1-15 min
t1
t1+15 min
OSW at 12 hour (f2)
t2-15 min
t2
t2+15 min
OSW at 13:30 hour (f3)
t3-15 min
t3
t3+15 min
OSW at 17:45 hour (f4)
t4-15 min
t4
t4+15 min
OCSW at 6 hour (f5)
t5-15 min
t5
t5+15 min
OCSW at 12 hour (f6)
t6-15 min
t6
t6+15 min
OCSW at 14 hour (f7)
t7-15 min
t7
t7+15 min
OCSW at 20 hour (f8)
t8-15 min
t8
t8+15 min
OCSO at 6 hour (f9)
t9-15 min
t9
t9+15 min
OCSO at 20 hour (f10)
t10-15 min
t10
t10+15 min
Thus, there are 10 factors and 3 levels that govern the robustness of the model and if the full
10
factorials are used to generalize the model, it takes 3 = 59049 experiments. The orthogonal arrays
reduce the number of experiments to 729 with 5 strengths. OA (729,10,3,5) is applied to the proposed
pseudo dynamic model in this case study.
4. Result and Discussion
Optimal configuration of the model is based on maximum
R 2 mod ified
and minimum
MSE mod ified
from different random initialized parameters. For each hidden neurons in the model, five
random initialized parameters is assigned for learning phase and based on it, the neurons with
minimum
MSE mod ified and maximum R 2 mod ified for learning and validation are chosen from random
initialized parameters. Optimal configuration of each model is chosen from maximum
minimum
R 2 mod ified
and
MSE mod ified
model performance from learning and validation datasets for different hidden
neurons. Figure (11) and (12) shows
R 2 mod ified
and
MSE mod ified performance for learning, validation
and testing for different hidden neurons sizes of model 5 and from this optimal configuration is chosen
from the best performance model. It is clear from figure (11) and (12) that the maximum
and minimum
R 2 mod ified
MSE mod ified performance is achieved in hidden neuron size 13 and which is the optimal
2
configuration of the model. It can also be noticed that although R testing performance increases for
2
hidden neuron size 15, R for validation and learning does not increase optimally. The model 5 is just
an example and similarly, the process is repeated for each model to find the optimal configuration of
the neural network model. The optimal configurations of the different neural network model are
summarized in table (4).
Learning
Validation
Testing
0.95
Best Performance
Coefficient of Correlation
0.9
0.85
0.8
0.75
0.7
0.65
2
4
6
8
10
12
14
Hidden Neurons
16
18
Figure 11: Coefficient of correlation performance (Model 5)
20
22
Learning
Validation
Testing
0.55
0.5
0.45
Mean Square Error
0.4
0.35
0.3
Best Performance
0.25
0.2
0.15
0.1
2
4
6
8
10
12
14
Hidden Neurons
16
18
20
22
Figure 12: Mean Square Error performance (Model 5)
Table 4: Optimal configuration of models
Model
Hidden
Coefficient of Correlation
Neurons Learning Validation
Testing
Mean Square Error
Learning Validation
Testing
Model 1
10
0.82
0.81
0.61
0.18
0.18
0.40
Model 2
19
0.87
0.85
0.80
0.13
0.15
0.21
Model 3
7
0.88
0.86
0.75
0.12
0.14
0.25
Model 4
9
0.89
0.87
0.82
0.12
0.13
0.18
Model 5
13
0.89
0.87
0.83
0.11
0.13
0.18
Model 6
9
0.89
0.87
0.85
0.11
0.13
0.15
Table (4) shows that with static neural network model 1, best
R 2 mod ified
for learning and
validation can be obtained up to 0.82 and 0.81. From this, it is clear that occupancy profile and
operational characteristics are not enough to determine and generalize the unknown function of the
building heating demand. As transitional attributes of operational characteristic is introduced in model
2,
R 2 mod ified
model performance increases significantly from 0.82 to 0.87 for learning phase and from
MSE mod ified decreases in contrast to model 1.
Pseudo dynamic transitional attributes in model 3 and time constant τ in model 4 leads increase in
0.81 to 0.85 for validation phase and correspondingly
model performance. Further, dynamics of settling time and steady state plays an important role in
characterizing the neural network model. It is seen that
R 2 mod ified
performance increases from 0.87 to
0.89 for learning and 0.85 to 0.87 for validation in model 5 compare to model 2 although transition
attributes is introduce in model 2. In addition, hidden neuron size is also reduces from 19 to 13.
Moreover, it is distinguish that learning and validation performances remained the same in the model 6
compared to model 5. The optimal choice of the model, thus, lies in between settling and steady state
time.
It can be further view that model 5 and model 6 show reasonable and consistent model
performances. However, minimum hidden neuron size and maximum learning criteria is essential for
the overall network generalization. Since the hidden neurons size decreases from 13 to 9 and model
performance
R 2 mod ified
remained the same (0.89) in model 6 comparing to model 5, model 6 is
chosen as the best configuration of the overall models. The optimal choice of the model 5 and model 6
can be delineated by the error in percentage of energy consumption (kWh) in actual and prediction for
the learning and validation phase. Heating energy consumption error in actual and prediction in
learning phase in Model 6 is 0.02% compare to 0.32% in Model 5. For validation phase, heating
energy consumption error is 2.39% in Model 6 compare to 2.57% in Model 5. From this energy
consumption error, it is clear that there is a small heating energy consumption error in Model 6
compare to Model 5 during the learning and validation phase. So, one can conclude that Model 6 can
be chosen as optimal configuration of the overall model. The model 6, thus, bridges the gap between
static and dynamic neural network model in the sense that it is better than static model and increases
the performance comparable to dynamic neural network model.
For the robustness of pseudo dynamic model, orthogonal arrays are applied to determine the
highest coefficient of correlation for learning and validation for the optimum 9 hidden neuron size of
model 6. Table (5) shows OA(729,10,3,5) and coefficient of correlation for learning and validation
phase. It is clear from table (5) that the schedule taken from the ESCOs is from experiment 1 and from
the orthogonal arrays, the optimal schedule that fits the best for model 6 is experiment 398. The
orthogonal arrays, thus, ensures that there is transition in occupancy in 7:45 hour, 12 hour, 13:45 hour
and 18 hour instead of 8 hour, 12 hour, 13:30 hour and 17:45 hour period in the existing case
respectively. There is also a transition in 5:45 hour, 11:45 hour, 14 hour and 17:45 hour instead of 6
hour, 12 hour, 14 hour and 17:45 hour for working day; 5:45 hour and 20 hour instead of 6 hour and
20 hour in off days for operational characteristics. The coefficient of correlation after the orthogonal
array design is 0.90 for learning, 0.88 for validation and 0.86 for training phase. Nevertheless, other
issue of overall model is that it is difficult to increase the coefficient of correlation beyond 0.90 and this
is due to the sampling time of 15 minutes. With short sampling time, it is very difficult to learn the
datasets which changes in 15 minutes sample, nonetheless, for good generalization of the model,
R 2 mod ified
value of 0.90 during the learning phase is always acceptable.
Coefficient of correlation of linear regression obtained from neural network model in the actual
and prediction of heating demand for learning, validation and testing phase of Model 6 after optimum
orthogonal array design are 0.95, 0.95 and 0.93 respectively. The prediction of heating demand for
model 6 after optimum orthogonal array design during validation phase is shown in figure (13).
Prediction gives the power heating demand and the area under the curve gives the heating energy
demand. From figure (13), it is clear that heating demand tremendously increases approximately 990
kW during third and fourth day and pseudo dynamic model is able to predict and learn the behavior.
However, there is a fluctuation in the power demand in the morning for each consecutive 4 days and it
is difficult to learn datasets which transits rapidly in actual power demand. The prediction of heating
demand for model 6 during testing phase after optimum orthogonal array design is shown in figure
(14). It is vivid that pseudo dynamic model is able to predict heating demand, however during the third
day, the pseudo dynamic model is not able to meet 1.1 MW of heating demand. This is due to the fact
that neural network does not learn this threshold maximum heating demand in the learning phase as
this kind of information is not available in the database. This data, thus, needs to be improved in the
learning phase through feature extraction techniques. Nonetheless, pseudo dynamic model (model 6)
prediction is in accordance to the actual target except for some rapid transits in the actual target. To
sum up, pseudo dynamic transition attributes in model 6 after orthogonal array design leads best
prediction of heating demand.
Table 5: OA(729,10,3,5) and coefficient of correlation for learning and validation for model 6
Element
f1
f2
f3
f4
f5
f6
f7
f8
f9
f10
1
2
2
2
2
2
2
2
2
2
2
1
2
2
2
2
2
2
1
3
3
2
2
2
2
2
2
4
2
1
2
2
2
2
3
5
1
1
2
2
2
2
6
3
1
2
2
2
7
2
3
2
2
8
1
3
2
2
9
3
3
2
10
2
2
11
1
12
3
….
Experiment
Coefficient of Correlation
Learning
Validation
Testing
2
0.89
0.87
0.85
1
1
0.89
0.88
0.81
3
3
3
0.90
0.86
0.76
2
1
3
0.89
0.86
0.79
3
1
3
2
0.89
0.87
0.78
2
3
3
2
1
0.89
0.88
0.79
2
2
1
2
3
1
0.89
0.87
0.83
2
2
1
1
2
3
0.89
0.87
0.84
2
2
2
1
3
1
2
0.90
0.86
0.85
1
2
2
2
3
1
2
1
0.89
0.87
0.76
2
1
2
2
2
3
3
1
3
0.89
0.87
0.67
2
1
2
2
2
3
2
3
2
0.89
0.87
0.67
.
.
.
.
.
.
.
.
.
.
.
.
.
…
.
.
.
.
.
.
.
.
.
.
.
.
.
394
2
3
1
3
1
1
3
3
3
3
0.89
0.87
0.80
395
1
3
1
3
1
1
3
2
2
2
0.90
0.87
0.76
396
3
3
1
3
1
1
3
1
1
1
0.90
0.87
0.76
397
2
2
3
3
1
1
2
2
2
3
0.89
0.88
0.81
398
1
2
3
3
1
1
2
1
1
2
0.90
0.88
0.86
399
3
2
3
3
1
1
2
3
3
1
0.90
0.87
0.70
400
2
1
3
3
1
1
3
2
1
1
0.90
0.87
0.77
401
1
1
3
3
1
1
3
1
3
3
0.89
0.88
0.84
402
3
1
3
3
1
1
3
3
2
2
0.87
0.88
0.74
….
.
.
.
.
.
.
.
.
.
.
.
.
.
…
.
.
.
.
.
.
.
.
.
.
.
.
.
725
1
1
3
3
3
3
2
1
2
3
0.89
0.87
0.80
726
3
1
3
3
3
3
2
3
1
2
0.90
0.87
0.84
727
2
3
3
3
3
3
3
2
2
2
0.90
0.87
0.80
728
1
3
3
3
3
3
3
1
1
1
0.89
0.88
0.78
729
3
3
3
3
3
3
3
3
3
3
0.89
0.88
0.61
Validation Phase
1000
Actual
Predict
900
Heating Demand (kW)
800
700
600
500
400
300
0
24
48
Hour
72
96
Figure 13: Prediction of heating demand in model 6 during validation phase (after optimum orthogonal
array design)
Test Phase
1100
Actual
Predict
1000
Heating Demand (kW)
900
800
700
600
500
400
300
0
24
48
Hour
72
96
Figure 14: Prediction of heating demand in model 6 during testing phase (after optimum orthogonal
array design)
4. Conclusion
This paper introduces pseudo dynamic transitional model for the building heating demand
prediction in a short time horizon using artificial neural network. Occupancy profile and operational
heating power level characteristics are included in the model. Dynamic characteristic of the building is
included in the model for the determination of pseudo dynamic transition lag. Settling time and steady
state time of the heating demand give an increment in precision of the model, however, choice of
model depends on their actual time between settling and steady state. The results were based on case
study where occupancy profile is already known and results may vary for more fluctuating occupancy
buildings. Coefficient of correlation increases from 0.82 to 0.89 for learning, 0.81 to 0.87 for validation
and 0.61 to 0.85 for testing in pseudo dynamic comparing to static neural network model. Also, the
size of hidden neuron is further reduced, which reduces complexities and increases generalization of
the model. Moreover, minimum energy consumption error is achieved in pseudo dynamic model as
0.02% for learning and 2.57% for validation phase. Further, orthogonal array is applied to optimal
pseudo dynamic model to confirm the schedule of occupancy profile and operational level
characteristics, and robustness of the model. The orthogonal array design leads to the increases in
coefficient of correlation in pseudo dynamic model and confirmed the new schedule of the occupancy
profile and operational level characteristics. The major contribution of this paper, thus, is the
introduction of transition and novel time dependent attributes of operational heating power level
characteristics, which is the dominant factor for building heating demand. Also, orthogonal array
design in the model makes flexibility in cross checking the schedule of occupancy profile and
operational heating power level characteristics obtained from ESCOs to design the robust model. The
prediction is in short time horizon (4 days) with sampling interval of 15 minutes and thus useful for
dynamic control of building heating demand.
Further, research will be focused towards the feature extraction of data before learning phase of
the neural network so that abnormalities in the data can be corrected in the learning phase. Also
adaptive and real time learning criteria with seasonal behaviour will be studied.
Acknowledgement
This research has been done in collaboration with Ecole des Mines, Nantes, Technische Universiteit
Eindhoven and VEOLIA Environnement Recherche et Innovation, funded through Erasmus Mundus
Joint Doctoral Programme SELECT+, the support of which is gratefully acknowledged.
References
1. J. Laustsen, Energy efficiency requirements in building codes, energy efficiency policies for
new
buildings,
International
Energy
Agency,
OECD/IEA,
(March)
(2008).
(http://www.iea.org/publications/freepublications/publication/Building_Codes.pdf
)
2. X. Li, C.P. Bowers, T. Schnier, Classification of Energy Consumption in Buildings with outlier
detection, IEEE Transactions on Industrial Electronics 57 (2010) 3639-3644.
3. H. Zhao, F. Magoules, A review on the prediction of building energy consumption, Renewable
and Sustainable Energy Reviews 16 (2012) 3586-3592.
4. D.B. Crawley, L.K. Lawrie, F.C. Winkelmann, W.F. Buhl, Y.J. Huang, C.O. Pedersen, R.K.
Strand, R.J. Liesen, D.E. Fisher, M.J. Witte, J. Glazer, EnergyPlus: creating a new-generation
building energy simulation program, Energy and Buildings 33 (2001) 319-331.
5. S. Citherlet, Towards the holistic assessment of building performance based on integrated
simulation approach, PhD Thesis, Swiss Federal Institute of Technology (2001).
(http://www.esru.strath.ac.uk/Documents/PhD/citherlet_thesis.pdf)
6. A.S. Kalagasidis, Weitzmann, T.R. Nielsen, R. Peuhkuri, C. Hagentoft, Rode, The international
building physics toolbox in simulink, Energy and Buildings 39 (2007) 665-674.
7. A. Husaunndee, R. Lahrech, H. Vaezi-Nejad, J.C. Visier, SIMBAD: A simulation toolbox for the
design and test of HVAC control systems, International IBPSA Conference, Prague
(September) (1997) 269-276 p. (http://www.ibpsa.org/proceedings/BS1997/BS97_P022.pdf)
8. TRNSYS 17, a TRaNsient SYstem Simulation program. http://sel.me.wisc.edu/trnsys/features
(Access on: 30/12/2012)
9. CARNOT Blockset, User’s Guide, Solar-Institut Juelich (1999).
10. L. Duanmu, Z. Wang, Z.J. Zhai, X. Li, A simplified method to predict hourly building cooling
load for urban energy planning, Energy and Buildings, 58 (2013) 281-291.
11. C.P. Underwood, F.W.H. Yik, Modeling methods for energy in Buildings, Blackwell Science
(2004).
12. L. Girardin, F. Marechal, M. Dubuis, N. Calame-Darbellay, D. Favrat, EnerGIS: A geographical
information based system for the evaluation of integrated energy conversion systems in urban
areas, Energy 35 (2010) 830-840.
13. R. Yao, K. Steemers, A method of formulating energy load profile for domestic buildings in the
UK, Energy and Buildings 37 (2005) 663-671.
14. T. Catalina, J. Virgone, E. Blanco, Development and Validation of regression models to predict
monthly heating demand for residental buildings, Energy and Buildings 40 (2008) 1825-1832.
15. K.K.W Wan, D.H.W Li, D. Liu, J.C. Lam, Future trends of building heating and cooling loads
and energy consumption in different climates, Building and Environment 46 (2011) 223-234.
16. R. Yokoyama, T. Wakui, R. Satake, Prediction of energy demands using neural network with
model identification by global optimization, Energy Conversion and Management 50 (2009)
319-327.
17. B. Dong, C. Cao, S.E. Lee, Applying support vector machines to predict building energy
consumption in tropical region, Energy and Buildings 37 (2005) 545-553.
18. Q. Li, Q. Meng, Development and applications of hourly building cooling load prediction model,
International Conference on Advances in Energy Engineering, IEEE, China (June) (2010).
(http://dx.doi.org/10.1109/ICAEE.2010.5557536)
19. S. Kalogirou, G. Florides, C. Neocleous, C. Schizas, Estimation of daily heating and cooling
loads using artificial neural networks, 2001 World Congress, Napoli (September) (2001) .
http://ktisis.cut.ac.cy/bitstream/10488/883/1/C41-CLIMA2001.pdf (Access on: 13/11/2012)
20. A.H. Neto, F.A.S. Fiorelli, Comparison between detailed model simulation and artificial neural
network for forecasting building energy consumption, Energy and Buildings, 40 (2008) 21692176.
21. Q. Shilin, S. Zhifeng, BP neural network for the prediction of urban building energy
consumption based on Matlab and its application, International Conference on Computer
Modeling and Simulation, IEEE, China (January) 2010.
22. G. Mihalakakou, M. Santamouris, A. Tsangrassoulis, On the energy consumptions in the
residential buildings, Energy and Buildings 34 (2002) 727-736.
23. B.B Ekici, U.T. Aksoy, Prediction of building energy consumption by using artificial neural
network, Advances in Engineering Software 40 (2009) 356-362.
24. O.A. Dombayci, The prediction of heating energy consumption in a model house using artificial
neural networks in Denizli-Turkey, Advances in Engineering Software 41 (2010) 141-147.
25. P.A. Gonzalez, J.M. Zamarreno, Prediction of hourly energy consumption in buildings based
on feedback artificial neural network, Energy and Buildings 37 (2005) 595-601.
26. P. Popescu, F. Ungureanu, A. Hernàndez-Guerrero, Simulation models for the analysis of
space heat consumption of buildings, Energy 34 (2009) 1447-1453.
27. K. Kato, M. Sakawa, K. Ishimaru, S. Ushiro, T. Shibano, Heat load prediction through recurrent
neural network in district heating and cooling systems, International Conference on Systems,
Man and Cybernetics (SMC), IEEE, Singapore (October) (2008).
28. S.A. Kalogirou, M. Bojic, Artificial neural networks for the prediction of the energy consumption
of a passive solar building, Energy, 25 (2000) 479-491.
29. Y. Sun, S. Wang, F. Xiao, Development and Validation of a simplified online cooling load
prediction strategy for a super high-rise building in Hongkong, Energy Conversion and
Management 68 (2013) 20-27.
30. K. Yun, R. Luck, P.J. Mago, H. Cho, Building hourly thermal load predictions using an indexed
ARX model, Energy and Buildings 54 (2012) 225-233.
31. E. Azar, C.C. Menassa, A comprehensive analysis of the impact of occupancy parameters in
energy simulation of office buildings, Energy and Buildings 55 (2012) 841-853.
32. M.C. Leung, N.C.F Tse, L.L. Lai, T.T. Chow, The use of occupancy space electrical power
demand in building cooling load prediction, Energy and Buildings, 55 (2012) 151-163.
33. F. Bompay, Evaluation of the Meteo-France response in ETEX release 1, Atmospheric
Environment, 32 (1998) 4351-4357.
34. C. Voyant, M. Muselli, C. Paoli, M-L. Nivet, Numerical weather prediction (NWP) and hybrid
ARMA/ANN model to predict global radiation, Energy, 39 (2012) 341-355.
35. The AROME modeling system. http://www.cnrm.meteo.fr/arome/
36. D. Chen, X. Wang, Z. Ren, Selection of climate variables and time scales for future weather
preparation in building heating and cooling energy predictions, Energy and Buildings 51 (2012)
223-233.
37. S. Haykin, Neural networks, a comprehensive foundation, Second Edition, Pearson Education
Inc (2005).
th
38. W. Yu, H. He, N. Zhang, Advances in Neural Networks – ISNN, 6 International Symposium
on Neural Networks, Springer-Berlin Heidelberg, New York (2009).
39. W.W. Hsieh, Machine learning methods in the environmental sciences, neural networks and
kernels, Cambridge University Press (2009).
40. X. Wu, D.Y.C. Leung, Optimization of biodiesel production from camelina oil using orthogonal
experiment, Applied Energy 88 (2011) 3615-3624.
41. W.C. Weng, F. Yang, A.Z. Elsherbeni, Linear antenna arrays synthesis using tgauchi’s
methods: a novel optimization technique in electromagnetics, IEEE Transactions on Antennas
and Propagation 55 (2007) 723-730.
42. L. Franek, X. Jiang, Orthogonal design of experiments for parameter learning in image
segementation, Signal Processing, 93 (2013) 1694-1704.
43. M.V.M Nguyen, Some new constructions of strength 3 mixed orthogonal arrays, Journal of
Statistical Planning and Inference, 138 (2008) 220-233.
44. C.Y. Suen, Construction of mixed orthogonal arrays by juxtaposition, Statistics and
Proabability Letters, 65 (2003) 161-163.
45. C.Y. Suen, A. Dey, Construction of asymmetric orthogonal arrays through finite geometries,
Journal of Statistical Planning and Inference, 115 (2003) 623-635.
46. N.J.A. Sloane, A library of orthogonal arrays (Online).
http://www2.research.att.com/~njas/oadir/ (Access on : 28/04/2013)
Appendix A
The influence of input variables on the model output is evaluated based on the correlation analysis.
Correlation measures the strength and weakness of linear relationship between two variables. There
are several coefficients that measure the correlation degree and Pearson’s correlation coefficient is
used to determine the input variables relevance for this paper. Pearson’s correlation coefficient is
calculated by dividing covariance of two variables by product of their standard deviation as shown in
equation (A.1 – A.2), where r represents Pearson’s correlation coefficient. In equations (A.1-A.2),
cov(xy)
is
covariance which represents strength of linear relationship between two variables
x
y ; x and y are mean values of variables x and y ; s x and s y are standard deviations of
variables x and y ; and n is the number of data.
and
r=
cov(xy )
sx s y
cov(xy ) =
(A.1)
1 n
∑ xi − x y i − y
n − 1 i =1
(
)(
)
(A.2)
The correlation coefficients can range from -1 to +1:
r =1
: perfect positive linear correlation
r = -1
: perfect negative linear correlation
0.1<
r <0.25 : small positive linear correlation
0.25<
r <0.6
0.6<
r <1
-1< r <0
: medium positive linear correlation
: strong positive linear correlation
: negative linear correlation
Climatic conditions (outside temperature and solar radiation), operational power level characteristics
and approximate occupancy profile are used to evaluate the relevance variables that affect building
heat demand based on case study data. Other variables pseudo dynamic transitional attributes, which
signifies the dynamics of building characteristics is not consider for relevance variable determination
since it only signifies time and phase interval of heating power transition.
Results show the linear coefficient of correlation of outside air temperature, solar radiations,
occupancy profile and operational power level characteristics with the heat load are -0.84, -0.40, 0.32
and 0.35 respectively. Results, thus, signifies that climatic conditions (outside temperature and solar
radiations) are relevant input variables to predict the heat load. Also, it is clearer that occupancy profile
and operational power level characteristics has medium positive correlation with heat load and shows
relevance to characterize the heat demand behaviour.
| 5cs.CE
|
P IXEL D ECONVOLUTIONAL N ETWORKS
Hongyang Gao
Washington State University
[email protected]
Hao Yuan
Washington State University
[email protected]
Zhengyang Wang
Washington State University
[email protected]
arXiv:1705.06820v4 [cs.LG] 27 Nov 2017
Shuiwang Ji
Washington State University
[email protected]
A BSTRACT
Deconvolutional layers have been widely used in a variety of deep models for
up-sampling, including encoder-decoder networks for semantic segmentation and
deep generative models for unsupervised learning. One of the key limitations of
deconvolutional operations is that they result in the so-called checkerboard problem. This is caused by the fact that no direct relationship exists among adjacent
pixels on the output feature map. To address this problem, we propose the pixel
deconvolutional layer (PixelDCL) to establish direct relationships among adjacent
pixels on the up-sampled feature map. Our method is based on a fresh interpretation of the regular deconvolution operation. The resulting PixelDCL can be used
to replace any deconvolutional layer in a plug-and-play manner without compromising the fully trainable capabilities of original models. The proposed PixelDCL
may result in slight decrease in efficiency, but this can be overcome by an implementation trick. Experimental results on semantic segmentation demonstrate that
PixelDCL can consider spatial features such as edges and shapes and yields more
accurate segmentation outputs than deconvolutional layers. When used in image
generation tasks, our PixelDCL can largely overcome the checkerboard problem
suffered by regular deconvolution operations.
1
I NTRODUCTION
Deep learning methods have shown great promise in a variety of artificial intelligence tasks such
as image classification (Krizhevsky et al., 2012; Simonyan & Zisserman, 2014), semantic segmentation (Noh et al., 2015; Shelhamer et al., 2016; Ronneberger et al., 2015), and natural image generation (Goodfellow et al., 2014; Kingma & Welling, 2014; Oord et al., 2016). Some key network
layers, such as convolutional layers (LeCun et al., 1998), pooling layers, fully connected layers and
deconvolutional layers, have been frequently used to create deep models for different tasks. Deconvolutional layers, also known as transposed convolutional layers (Vedaldi & Lenc, 2015), are
initially proposed in (Zeiler et al., 2010; 2011). They have been primarily used in deep models that
require up-sampling of feature maps, such as generative models (Radford et al., 2015; Makhzani
& Frey, 2015; Rezende et al., 2014) and encoder-decoder architectures (Ronneberger et al., 2015;
Noh et al., 2015). Although deconvolutional layers are capable of producing larger feature maps
from smaller ones, they suffer from the problem of checkerboard artifacts (Odena et al., 2016). This
greatly limits deep model’s capabilities in generating photo-realistic images and producing smooth
outputs on semantic segmentation. To date, very little efforts have been devoted to improving the
deconvolution operation.
In this work, we propose a simple, efficient, yet effective method, known as the pixel deconvolutional
layer (PixelDCL), to address the checkerboard problem suffered by deconvolution operations. Our
method is motivated from a fresh interpretation of deconvolution operations, which clearly pinpoints
the root of checkerboard artifacts. That is, the up-sampled feature map generated by deconvolution
can be considered as the result of periodical shuffling of multiple intermediate feature maps computed from the input feature map by independent convolutions. As a result, adjacent pixels on the
output feature map are not directly related, leading to the checkerboard artifacts. To overcome this
1
Figure 1: Comparison of semantic segmentation results. The first and second rows are images
and ground true labels, respectively. The third and fourth rows are the results of using regular
deconvolution and our proposed pixel deconvolution PixelDCL, respectively.
problem, we propose the pixel deconvolutional operation to be used in PixelDCL. In this new layer,
the intermediate feature maps are generated sequentially so that feature maps generated in a later
stage are required to depend on previously generated ones. In this way, direct relationships among
adjacent pixels on the output feature map have been established. Sequential generation of intermediate feature maps in PixelDCL may result in slight decrease in computational efficiency, but we show
that this can be largely overcome by an implementation trick. Experimental results on semantic
segmentation (samples in Figure 1) and image generation tasks demonstrate that the proposed PixelDCL can effectively overcome the checkerboard problem and improve predictive and generative
performance.
Our work is related to the pixel recurrent neural networks (PixelRNNs) (Oord et al., 2016) and
PixelCNNs (van den Oord et al., 2016; Reed et al., 2017), which are generative models that consider the relationship among units on the same feature map. They belong to a more general class
of autoregressive methods for probability density estimation (Germain et al., 2015; Gregor et al.,
2015; Larochelle & Murray, 2011). By using masked convolutions in training, the training time
of PixelRNNs and PixelCNNs is comparable to that of other generative models such as generative
adversarial networks (GANs) (Goodfellow et al., 2014; Reed et al., 2016) and variational autoencoders (VAEs) (Kingma & Welling, 2014; Johnson et al., 2016). However, the prediction time of
PixelRNNs or PixelCNNs is very slow since it has to generate images pixel by pixel. In contrast,
our PixelDCL can be used to replace any deconvolutional layer in a plug-and-play manner, and the
slight decrease in efficiency can be largely overcome by an implementation trick.
2
P IXEL D ECONVOLUTIONAL L AYERS AND N ETWORKS
We introduce deconvolutional layers and analyze the cause of checkerboard artifacts in this section.
We then propose the pixel deconvolutional layers and the implementation trick to improve efficiency.
2.1
D ECONVOLUTIONAL L AYERS
Deconvolutional networks and deconvolutional layers are proposed in (Zeiler et al., 2010; 2011).
They have been widely used in deep models for applications such as semantic segmentation (Noh
et al., 2015) and generative models (Kingma & Welling, 2014; Goodfellow et al., 2014; Oord et al.,
2016). Many encoder-decoder architectures use deconvolutional layers in decoders for up-sampling.
One way of understanding deconvolutional operations is that the up-sampled output feature map is
obtained by periodical shuffling of multiple intermediate feature maps obtained by applying multiple
convolutional operations on the input feature maps (Shi et al., 2016).
2
kernel
In
1
2
1
2
3
4
1
2
3
4
1
2
1
2
3
4
kernel2
kernel1
3
4
1
2
3
4
1
3
+
3
4
1
1
In
1
1
1
2
4
3
2
3
3
2
2
In
3
2
3
2
Intermediate feature maps
4
4
4
4
4
Intermediate feature maps
Output feature map
Output feature map
Figure 2: Illustration of 1D deconvolutional operation. In this deconvolutional layer, a 4×1 feature
map is up-sampled to an 8×1 feature map. The left figure shows that each input unit passes through
an 1×4 kernel. The output feature map is obtained as the sum of values in each column. It can be
seen from this figure that the purple outputs are only related to (1, 3) entries in the kernel, while
the orange outputs are only related to (2, 4) entries in the kernel. Therefore, 1D deconvolution can
be decomposed as two convolutional operations shown in the right figure. The two intermediate
feature maps generated by convolutional operations are dilated and combined to obtain the final
output. This indicates that the standard deconvolutional operation can be decomposed into multiple
convolutional operations.
Output feature map
Intermediate feature maps
Input feature map
Figure 3: Illustration of 2D deconvolutional operation. In this deconvolutional layer, a 4×4 feature
map is up-sampled to an 8×8 feature map. Four intermediate feature maps (purple, orange, blue,
and red) are generated using four different convolutional kernels. Then these four intermediate
feature maps are shuffled and combined to produce the final 8×8 feature map. Note that the four
intermediate feature maps rely on the input feature map but with no direct relationship among them.
This interpretation of deconvolution in 1D and 2D is illustrated in Figures 2 and 3, respectively.
It is clear from these illustrations that standard deconvolutional operation can be decomposed into
several convolutional operations depending on the up-sampling factor. In the following, we assume
the up-sampling factor is two, though deconvolution operations can be applied to more generic
settings. Formally, given an input feature map Fin , a deconvolutional layer can be used to generate
an up-sampled output Fout as follows:
F1 = Fin ~ k1 ,
F2 = Fin ~ k2 ,
F3 = Fin ~ k3 ,
Fout = F1 ⊕ F2 ⊕ F3 ⊕ F4 ,
F4 = Fin ~ k4 ,
(1)
where ~ denotes the convolutional operation and ⊕ denotes the periodical shuffling and combination operation as in Figure 3, Fi is the intermediate feature map generated by the corresponding
convolutional kernel ki for i = 1, · · · , 4.
It is clear from the above interpretation of deconvolution that there is no direct relationship among
these intermediate feature maps since they are generated by independent convolutional kernels. Although pixels of the same position on intermediate feature maps depend on the same receptive field
of the input feature map, they are not directly related to each other. Due to the periodical shuffling
operation, adjacent pixels on the output feature map are from different intermediate feature maps.
3
Figure 4: Illustration of the checkerboard problem in semantic segmentation using deconvolutional
layers. The first and second rows are the original images and semantic segmentation results, respectively.
This implies that the values of adjacent pixels can be significantly different from each other, resulting in the problem of checkerboard artifacts (Odena et al., 2016) as illustrated in Figure 4. One way
to alleviate checkerboard artifacts is to apply post-processing such as smoothing (Li et al., 2001),
but this adds additional complexity to the network and makes the entire network not fully trainable.
In this work, we propose the pixel deconvolutional operation to add direct dependencies among
intermediate feature maps, thereby making the values of adjacent pixels close to each other and
effectively solving the checkerboard artifact problem. In addition, our pixel deconvolutional layers
can be easily used to replace any deconvolutional layers without compromising the fully trainable
capability.
2.2
P IXEL D ECONVOLUTIONAL L AYERS
To solve the checkerboard problem in deconvolutional layers, we propose the pixel deconvolutional
layers (PixelDCL) that can add dependencies among intermediate feature maps. As adjacent pixels
are from different intermediate feature maps, PixelDCL can build direct relationships among them,
thus solving the checkerboard problem. In this method, intermediate feature maps are generated
sequentially instead of simultaneously. The intermediate feature maps generated in a later stage are
required to depend on previously generated ones. The primary purpose of sequential generation is to
add dependencies among intermediate feature maps and thus adjacent pixels in final output feature
maps. Finally, these intermediate feature maps are shuffled and combined to produce final output
feature maps. Compared to Eqn. 1, Fout is obtained as follows:
F1 = Fin ~ k1 ,
F3 = [Fin , F1 , F2 ] ~ k3 ,
Fout = F1 ⊕ F2 ⊕ F3 ⊕ F4 ,
F2 = [Fin , F1 ] ~ k2 ,
F4 = [Fin , F1 , F2 , F3 ] ~ k4 ,
(2)
where [·, ·] denotes the juxtaposition of feature maps. Note that in Eqn. 2, ki denotes a set of kernels
as it involves convolution with the juxtaposition of multiple feature maps. Since the intermediate
feature maps in Eqn. 2 depend on both the input feature map and the previously generated ones,
we term it input pixel deconvolutional layer (iPixelDCL). Through this process, pixels on output
feature maps will be conditioned not only on input feature maps but also on adjacent pixels. Since
there are direct relationships among intermediate feature maps and adjacent pixels, iPixelDCL is
expected to solve the checkerboard problem to some extent. Note that the relationships among
intermediate feature maps can be very flexible. The intermediate feature maps generated later on
can rely on part or all of previously generated intermediate feature maps. This depends on the design
of pixel dependencies in final output feature maps. Figure 5 illustrates a specific design of sequential
dependencies among intermediate feature maps.
In iPixelDCL, we add dependencies among generated intermediate feature maps, thereby making
adjacent pixels on final output feature maps directly related to each other. In this process, the information of the input feature map is repeatedly used when generating intermediate feature maps.
When generating the intermediate feature maps, information from both the input feature map and
previous intermediate feature maps is used. Since previous intermediate feature maps already contain information of the input feature map, the dependencies on the input feature map can be removed.
4
Intermediate feature maps
Output feature map
Input feature map
Figure 5: Illustration of iPixelDCL and PixelDCL described in section 2.2. In iPixelDCL, there
are additional dependencies among intermediate feature maps. Specifically, the four intermediate
feature maps are generated sequentially. The purple feature map is generated from the input feature
map (blue). The orange feature map is conditioned on both the input feature map and the purple
feature map that has been generated previously. In this way, the green feature map relies on the input
feature map, purple and orange intermediate feature maps. The red feature map is generated based
on the input feature map, purple, orange, and green intermediate feature maps. We also propose
to move one step further and allow only the first intermediate feature map to depend on the input
feature map. This gives rise to PixelDCL. That is, the connections indicated by dashed lines are
removed to avoid repeated influence of the input feature map. In this way, only the first feature map
is generated from the input and other feature maps do not directly rely on the input. In PixelDCL,
the orange feature map only depends on the purple feature map. The green feature map relies on
the purple and orange feature maps. The red feature map is conditioned on the purple, orange, and
green feature maps. The information of the input feature map is delivered to other intermediate
feature maps through the first intermediate feature map (purple).
Removing such dependencies for some intermediate feature maps can not only improve the computational efficiency but also reduce the number of trainable parameters in deep models.
In this simplified pixel deconvolutional layer, only the first intermediate feature map will depend
on the input feature map. The intermediate feature maps generated afterwards will only depend on
previously generated intermediate feature maps. This will simplify the dependencies among pixels
on final output feature map. In this work, we use PixelDCL to denote this simplified design. Our
experimental results show that PixelDCL yields better performance than iPixelDCL and regular
deconvolution. Compared to Eqn. 2, Fout in PixelDCL is obtained as follows:
F1 = Fin ~ k1 ,
F3 = [F1 , F2 ] ~ k3 ,
Fout = F1 ⊕ F2 ⊕ F3 ⊕ F4 .
F2 = F1 ~ k2 ,
F4 = [F1 , F2 , F3 ] ~ k4 ,
(3)
PixelDCL is illustrated in Figure 5 by removing the connections denoted with dash lines. When
analyzing the relationships of pixels on output feature maps, it is clear that each pixel will still rely on
adjacent pixels. Therefore, the checkerboard problem can be solved with even better computational
efficiency. Meanwhile, our experimental results demonstrate that the performance of models with
these simplified dependencies is even better than that with complete connections. This demonstrates
that repeated dependencies on the input may not be necessary.
2.3
P IXEL D ECONVOLUTIONAL N ETWORKS
Pixel deconvolutional layers can be applied to replace any deconvolutional layers in various models involving up-sampling operations such as U-Net (Ronneberger et al., 2015), VAEs (Kingma &
Welling, 2014) and GANs (Goodfellow et al., 2014). By replacing deconvolutional layers with
pixel deconvolutional layers, deconvolutional networks become pixel deconvolutional networks
(PixelDCN). In U-Net for semantic segmentation, pixel deconvolutional layers can be used to upsample from low-resolution feature maps to high-resolution ones. In VAEs, they can be applied in
decoders for image reconstruction. The generator networks in GANs typically use deep model (Radford et al., 2015) and thus can employ pixel deconvolutional layers to generate large images. In our
5
4
1
5
3
2
3
5
Figure 6: An efficient implementation of the pixel deconvolutional layer. In this layer, a 4×4
feature map is up-sampled to a 8×8 feature map. The purple feature map is generated through a 3×3
convolutional operation from the input feature map (step 1). After that, another 3×3 convolutional
operation is applied on the purple feature map to produce the orange feature map (step 2). The
purple and orange feature maps are dilated and added together to form a larger feature map (step
3). Since there is no relationship between the last two intermediate feature maps, we can apply a
masked 3×3 convolutional operation, instead of two separate 3×3 convolutional operations (step 4).
Finally, the two large feature maps are combined to generate the final output feature map (step 5).
experiments, we evaluate pixel deconvolutional layers in U-Net and VAEs. The results show that the
performance of pixel deconvolutional layers outperforms deconvolutional layers in these networks.
In practice, the most frequently used up-sampling operation is to increase the height and width of
input feature maps by a factor of two, e.g., from 2×2 to 4×4. In this case, the pixels on output
feature maps can be divided into four groups as in Eqn. 1. The dependencies can be defined as
in Figure 5. When implementing pixel deconvolutional layers, we design a simplified version to
reduce sequential dependencies for better parallel computation and training efficiency as illustrated
in Figure 6.
In this design, there are four intermediate feature maps. The first intermediate feature map depends
on the input feature map. The second intermediate feature map relies on the first intermediate
feature map. The third and fourth intermediate feature maps are based on both the first and the
second feature maps. Such simplified relationships enable the parallel computation for the third
and fourth intermediate feature maps, since there is no dependency between them. In addition, the
masked convolutional operation can be used to generate the last two intermediate feature maps. As
has been mentioned already, a variety of different dependencies relations can be imposed on the
intermediate feature maps. Our simplified design achieves reasonable balance between efficiency
and performance. Our code is publicly available1
3
E XPERIMENTAL S TUDIES
In this section, we evaluate the proposed pixel deconvolutional methods on semantic segmentation
and image generation tasks in comparison to the regular deconvolution method. Results show that
the use of the new pixel deconvolutional layers improves performance consistently in both supervised and unsupervised learning settings.
3.1
S EMANTIC S EGMENTATION
Experimental Setup: We use the PASCAL 2012 segmentation dataset (Everingham et al., 2010)
and MSCOCO 2015 detection dataset (Lin et al., 2014) to evaluate the proposed pixel deconvolutional methods in semantic segmentation tasks. For both datasets, the images are resized to
256×256×3 for batch training. Our models directly predict the label for each pixel without any
post-processing. Here we examine our models in two ways: training from scratch and fine-tuning
from state-of-art model such as DeepLab-ResNet.
For the training from scratch experiments, we use the U-Net architecture (Ronneberger et al., 2015)
as our base model as it has been successfully applied in various image segmentation tasks. The
1
https://github.com/divelab/PixelDCN
6
Figure 7: Sample segmentation results on the PASCAL 2012 segmentation dataset using training
from scratch models. The first and second rows are the original images and the corresponding
ground truth, respectively. The third, fourth, and fifth rows are the segmentation results of models
using deconvolutional layers, iPixelDCL, and PixelDCL, respectively.
network consists of four blocks in the encoder path and four corresponding blocks in the decoder
path. Within each decoder block, there is a deconvolutional layer followed by two convolutional
layers. The final output layer is adjusted based on the number of classes in the dataset. The PASCAL
2012 segmentation dataset has 21 classes while the MSCOCO 2015 detection dataset has 81 classes.
As the MSCOCO 2015 detection dataset has more classes than the PASCAL 2012 segmentation
dataset, the number of feature maps in each layer for this dataset is doubled to accommodate more
output channels. The baseline U-Net model employs deconvolutional layers within the decoder
path to up-sample the feature maps. We replace the deconvolutional layers with our proposed pixel
deconvolutional layers (iPixelDCL) and their simplified version (PixelDCL) while keeping all other
variables unchanged. The kernel size in DCL is 6×6, which has the same number of parameters as
iPixelDCL with 4 sets of 3×3 kernels, and more parameters than PixelDCL with 2 sets of 3×3 and
1 set of 2×2 kernels. This will enable us to evaluate the new pixel deconvolutional layers against
the regular deconvolutional layers while controlling all other factors.
For the fine-tuning experiments, we fine-tune our models based on the architecture of DeepLabResNet (Chen et al., 2016). The DeepLab-ResNet model is fine-tuned from ResNet101 (He et al.,
2016) and also use external data for training. The strategy of using external training data and finetuning from classic ResNet101 greatly boosts the performance of the model on both accuracy and
mean IOU. The output of DeepLab-ResNet is eight times smaller than the input image on the height
and width dimensions. In order to recover the original dimensions, we add three up-sampling blocks,
each of which up-samples the feature maps by a factor of 2. For each up-sampling block, there is
a deconvolutional layer followed by a convolutional layer. By employing the same strategy, we
replace the deconvolutional layer by PixelDCL and iPixelDCL using kernels of the same size as in
the training from scratch experiments.
Analysis of Results: Some sample segmentation results of U-Net using deconvolutional layers
(DCL), iPixelDCL, and PixelDCL on the PASCAL 2012 segmentation dataset and the MSCOCO
2015 detection dataset are given in Figures 7 and 8, respectively. We can see that U-Net models using
iPixelDCL and PixelDCL can better capture the local information of images than the same base
model using regular deconvolutional layers. By using pixel deconvolutional layers, more spacial
features such as edges and shapes are considered when predicting the labels of adjacent pixels.
Moreover, the semantic segmentation results demonstrate that the proposed models tend to produce
smoother outputs than the model using deconvolution. We also observe that, when the training epoch
is small (e.g., 50 epochs), the model that employs PixelDCL has better segmentation outputs than
the model using iPixelDCL. When the training epoch is large enough (e.g., 100 epochs), they have
similar performance, though PixelDCL still outperforms iPixelDCL in most cases. This indicates
that PixelDCL is more efficient and effective, since it has much fewer parameters to learn.
Table 1 shows the evaluation results in terms of pixel accuracy and mean IOU on the two datasets.
The U-Net models using iPixelDCL and PixelDCL yield better performance than the same base
7
Figure 8: Sample segmentation results on the MSCOCO 2015 detection dataset using training from
scratch models. The first and second rows are the original images and the corresponding ground
truth, respectively. The third, fourth, and fifth rows are the segmentation results of models using
deconvolutional layers, iPixelDCL, and PixelDCL, respectively.
Table 1: Semantic segmentation results on the PASCAL 2012 segmentation dataset and MSCOCO
2015 detection dataset. We compare the same base U-Net model and fine-tuned DeepLab-ResNet
using three different up-sampling methods in decoders; namely regular deconvolution layer (DCL),
the proposed input pixel deconvolutional layer (iPixelDCL) and pixel deconvolutional layer (PixelDCL). The pixel accuracy and mean IOU are used as performance measures.
Dataset
Model
Pixel Accuracy Mean IOU
U-Net + DCL
0.816161
0.415178
PASCAL 2012
U-Net + iPixelDCL
0.817129
0.448817
U-Net + PixelDCL
0.822591
0.455972
U-Net + DCL
0.809327
0.349769
MSCOCO 2015 U-Net + iPixelDCL
0.809239
0.360216
U-Net + PixelDCL
0.811575
0.371805
DeepLab-ResNet + DCL
0.929562
0.727036
PASCAL 2012
DeepLab-ResNet + iPixelDCL
0.934493
0.738552
DeepLab-ResNet + PixelDCL
0.931287
0.735585
model using regular deconvolution. The model using PixelDCL slightly outperforms the model using iPixelDCL. For the models fine-tuned from Deeplab-ResNet, the models using iPixelDCL and
PixelDCL have better performance than the model using DCL, with iPixelDCL performs the best. In
semantic segmentation, mean IOU is a more accuracy evaluation measure than pixel accuracy (Everingham et al., 2010). The models using pixel deconvolution have better evaluation results on mean
IOU than the base model using deconvolution.
3.2
I MAGE G ENERATION
Experimental Setup: The dataset used for image generation is the celebFaces attributes (CelebA)
dataset (Liu et al., 2015). To avoid the influence of background, the images have been preprocessed
so that only facial information is retained. The image generation task is to reconstruct the faces
excluding backgrounds in training images. The size of images is 64 × 64 × 3. We use the standard
variational auto-encoder (VAE) (Kingma & Welling, 2014) as our base model for image generation.
The decoder part in standard VAE employs deconvolutional layers for up-sampling. We apply our
proposed PixelDCL to replace deconvolutional layers in decoder while keeping all other components
the same. The kernel size in DCL is 6×6, which has more parameters than PixelDCL with 2 sets of
3×3 and 1 set of 2×2 kernels.
Analysis of Results: Figure 9 shows the generated faces using VAEs with regular deconvolution
(baseline) and PixelDCL in decoders. Some images generated by the baseline model suffer from
8
Figure 9: Sample face images generated by VAEs when trained on the CelebA dataset. The first
two rows are images generated by a standard VAE with deconvolutional layers for up-sampling. The
last two rows generated by the same VAE model, but using PixelDCL for up-sampling.
Table 2: Training and prediction time on semantic segmentation using the PASCAL 2012 segmentation dataset on a Tesla K40 GPU. We compare the training time of 10 epochs and prediction time
of 2109 images for the same base U-Net model using three different methods for up-sampling in the
decoders; namely DCL, iPixelDCL, and PixelDCL.
Model
Training time Prediction time
U-Net + DCL
365m26s
2m42s
511m19s
4m13s
U-Net + iPixelDCL
U-Net + PixelDCL
464m31s
3m27s
apparent checkerboard artifacts, while none is found on the images generated by the model with
PixelDCL. This demonstrates that the proposed pixel deconvolutional layers are able to establish
direct relationships among adjacent pixels on generated feature maps and images, thereby effectively
overcoming the checkerboard problem. Our results demonstrate that PixelDCL is very useful for
generative models since it can consider local spatial information and produce photo-realistic images
without the checkerboard problem.
3.3
T IMING C OMPARISON
Table 2 shows the comparison of the training and prediction time of the U-Net models using DCL,
iPixelDCL, and PixelDCL for up-sampling. We can see that the U-Net models using iPixelDCL
and PixelDCL take slightly more time during training and prediction than the model using DCL,
since the intermediate feature maps are generated sequentially. The model using PixelDCL is more
efficient due to reduced dependencies and efficient implementation discussed in Section 2.3. Overall,
the increase in training and prediction time is not dramatic, and thus we do not expect this to be a
major bottleneck of the proposed methods.
4
C ONCLUSION
In this work, we propose pixel deconvolutional layers that can solve the checkerboard problem in
deconvolutional layers. The checkerboard problem is caused by the fact that there is no direct relationship among intermediate feature maps generated in deconvolutional layers. PixelDCL proposed
here try to add direct dependencies among these generated intermediate feature maps. PixelDCL
generates intermediate feature maps sequentially so that the intermediate feature maps generated in
a later stage are required to depend on previously generated ones. The establishment of dependencies
in PixelDCL can ensure adjacent pixels on output feature maps are directly related. Experimental
results on semantic segmentation and image generation tasks show that PixelDCL is effective in
overcoming the checkerboard artifacts. Results on semantic segmentation also show that PixelDCL
9
is able to consider local spatial features such as edges and shapes, leading to better segmentation
results. In the future, we plan to employ our PixelDCL in a broader class of models, such as the
generative adversarial networks (GANs).
R EFERENCES
Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille.
Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and
fully connected crfs. arXiv:1606.00915, 2016.
Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman.
The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):
303–338, 2010.
Mathieu Germain, Karol Gregor, Iain Murray, and Hugo Larochelle. Made: Masked autoencoder
for distribution estimation. In Proceedings of The 32nd International Conference on Machine
Learning, pp. 881–889, 2015.
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680, 2014.
Karol Gregor, Ivo Danihelka, Alex Graves, Danilo Rezende, and Daan Wierstra. Draw: A recurrent
neural network for image generation. In Proceedings of the 32nd International Conference on
Machine Learning (ICML-15), pp. 1462–1471, 2015.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.
770–778, 2016.
Matthew Johnson, David K Duvenaud, Alex Wiltschko, Ryan P Adams, and Sandeep R Datta. Composing graphical models with neural networks for structured representations and fast inference.
In Advances in Neural Information Processing Systems, pp. 2946–2954, 2016.
Diederik P Kingma and Max Welling. Stochastic gradient vb and the variational auto-encoder. In
Second International Conference on Learning Representations, ICLR, 2014.
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,
2012.
Hugo Larochelle and Iain Murray. The neural autoregressive distribution estimator. In International
Conference on Artificial Intelligence and Statistics, pp. 29–37, 2011.
Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to
document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
Q Li, GP Steven, and YM Xie. A simple checkerboard suppression algorithm for evolutionary
structural optimization. Structural and Multidisciplinary Optimization, 22(3):230–239, 2001.
Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr
Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European
Conference on Computer Vision, pp. 740–755. Springer, 2014.
Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild.
In Proceedings of International Conference on Computer Vision (ICCV), 2015.
Alireza Makhzani and Brendan J Frey. Winner-take-all autoencoders. In Advances in Neural Information Processing Systems, pp. 2791–2799, 2015.
Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In IEEE International Conference on Computer Vision, 2015.
10
Augustus Odena, Vincent Dumoulin, and Chris Olah. Deconvolution and checkerboard artifacts. Distill, 2016. doi: 10.23915/distill.00003. URL http://distill.pub/2016/
deconv-checkerboard.
Aaron Van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks.
In Proceedings of The 33rd International Conference on Machine Learning, pp. 1747–1756, 2016.
Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep
convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.
Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee.
Generative adversarial text to image synthesis. In Proceedings of The 33rd International Conference on Machine Learning, volume 3, 2016.
Scott Reed, Aäron van den Oord, Nal Kalchbrenner, Sergio Gómez Colmenarejo, Ziyu Wang, Dan
Belov, and Nando de Freitas. Parallel multiscale autoregressive density estimation. arXiv preprint
arXiv:1703.03664, 2017.
Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and
approximate inference in deep generative models. In Proceedings of The 31st International Conference on Machine Learning, pp. 1278–1286, 2014.
Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and
Computer-Assisted Intervention, pp. 234–241. Springer, 2015.
Evan Shelhamer, Jonathon Long, and Trevor Darrell. Fully convolutional networks for semantic
segmentation. IEEE transactions on pattern analysis and machine intelligence, 2016.
Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel
Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient
sub-pixel convolutional neural network. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pp. 1874–1883, 2016.
Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556, 2014.
Aaron van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelcnn decoders. In Advances in Neural Information Processing
Systems, pp. 4790–4798, 2016.
Andrea Vedaldi and Karel Lenc. Matconvnet: Convolutional neural networks for matlab. In Proceedings of the 23rd ACM international conference on Multimedia, pp. 689–692. ACM, 2015.
Matthew D Zeiler, Dilip Krishnan, Graham W Taylor, and Rob Fergus. Deconvolutional networks.
In Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference on, pp. 2528–2535.
IEEE, 2010.
Matthew D Zeiler, Graham W Taylor, and Rob Fergus. Adaptive deconvolutional networks for mid
and high level feature learning. In Computer Vision (ICCV), 2011 IEEE International Conference
on, pp. 2018–2025. IEEE, 2011.
11
| 9cs.NE
|
arXiv:1606.01633v1 [math.ST] 6 Jun 2016
Bernoulli 22(4), 2016, 1963–1978
DOI: 10.3150/15-BEJ716
Conditions for a Lévy process to stay positive
near 0, in probability
ROSS A. MALLER
School of Finance, Actuarial Studies and Statistics, Australian National University, Canberra,
ACT,
Australia. E-mail: [email protected]
A necessary and sufficient condition for a Lévy process X to stay positive, in probability, near 0,
which arises in studies of Chung-type laws for X near 0, is given in terms of the characteristics
of X.
Keywords: Lévy process; staying positive
1. Introduction
Let (Xt )t≥0 be a real valued Lévy process with canonical triplet (γ, σ 2 , Π), thus having
characteristic function EeiθXt = etΨ(θ) , t ≥ 0, θ ∈ R, with characteristic exponent
1
Ψ(θ) := iθγ − σ 2 θ2 +
2
Z
R\{0}
(eiθx − 1 − iθx1{|x|≤1} )Π(dx).
Here, γ ∈ R, σ 2 ≥ 0, and Π is a Borel measure on R \ {0} such that
∞.
The condition
R
R\{0}
(1.1)
(x2 ∧ 1)Π(dx) <
lim inf P (Xt ≤ 0) ∧ P (Xt ≥ 0) > 0
t↓0
(1.2)
was shown by Wee [10] to imply a Chung-type law at 0 for X. Attention is drawn to
this in a recent paper of Aurzada, Döring and Savov [2], who give extended and refined
versions of the Chung law using a quite different approach to that of Wee. The difference
between (1.2) and the conditions imposed by Aurzada et al. [2] is not at all clear, though
based on some examples they suggest that theirs are weaker than (1.2). Our aim in this
paper is to give necessary and sufficient conditions for X to stay positive near 0, or to
stay negative near 0, and hence to characterise (1.2).
This is an electronic reprint of the original article published by the ISI/BS in Bernoulli,
2016, Vol. 22, No. 4, 1963–1978. This reprint differs from the original in pagination and
typographic detail.
1350-7265
c
2016 ISI/BS
2
R.A. Maller
We need some more notation. The positive, negative and two-sided tails of Π are
+
−
Π (x) := Π{(x, ∞)},
+
Π (x) := Π{(−∞, −x)} and
−
Π(x) := Π (x) + Π (x),
(1.3)
x > 0.
The restriction of Π to (0, ∞) is denoted by Π(+) , and we define Π(−) on (0, ∞) by
Π(−) (dx) := −Π(−dx), for x > 0. We are only interested in small time behaviour of Xt ,
and we eliminate the compound Poisson case by assuming Π(R) = ∞ throughout.
Define truncated and Winsorised moments as
Z
ν(x) = γ −
yΠ(dy),
x<|y|≤1
+
−
A(x) = γ + Π (1) − Π (1) −
Z
(1.4)
1
+
−
(Π (y) − Π (y)) dy
x
and
2
V (x) = σ +
Z
2
y Π(dy),
2
U (x) = σ + 2
0<|y|≤x
Z
x
yΠ(y) dy,
x > 0.
(1.5)
0
R
These functions are defined and finite for all x > 0 by virtue of property 0<|y|≤1 y 2 Π(dy) <
∞ of the Lévy measure Π but only their behaviour as x ↓ 0 will be relevant for us. Integration by parts shows that
+
−
A(x) = ν(x) + x(Π (x) − Π (x)),
x > 0.
(1.6)
Doney [5], Lemma 9, gives the following version of the Itô decomposition of X which
caters for positive and negative jumps separately. Take constants h+ > 0 and h− > 0.
Then for t ≥ 0,
Xt = tγ − tν+ (h+ ) + tν− (h− )
(S,h+ ,+)
+ σZt + Xt
(S,h− ,−)
+ Xt
(B,h+ ,+)
+ Xt
where γ and σ are as in (1.1), and the functions ν± are
Z
Z
xΠ(dx) and ν− (h− ) :=
ν+ (h+ ) :=
(B,h− ,−)
+ Xt
(1.7)
,
xΠ(−) (dx).
(1.8)
(h− ,1]
(h+ ,1]
Again, only their behaviour for small values of h± will be relevant. We can keep h± ∈
(S,h ,+)
(0, 1). Note that ν(x) = γ − ν+ (x) + ν− (x). In (1.7), (Xt + )t≥0 is a compensated sum
of small positive jumps, that is,
X
Z
(S,h ,+)
Xt + = a.s. lim
xΠ(dx) ,
∆Xs 1{ε<∆Xs ≤h+ } − t
ε↓0
0<s≤t
ε<x≤h+
Lévy process staying positive near 0
3
(S,h− ,−)
(Xt
)t≥0 is a compensated sum of small negative jumps, that is,
X
Z
(S,h ,−)
Xt − = a.s. lim
∆Xs 1{−h− ≤∆Xs <−ε} − t
xΠ(dx) ,
ε↓0
−h− ≤x<−ε
0<s≤t
(B,h ,±)
where the almost sure limits exist; and (Xt ± )t≥0 are the processes of positive and
negative big jumps, thus,
X
X
(B,h ,+)
(B,h ,−)
Xt + =
∆Xs 1{∆Xs >h+ } and Xt − =
∆Xs 1{∆Xs <−h− } ,
t > 0.
0<s≤t
0<s≤t
Finally, (Zt )t≥0 is a standard Brownian motion independent of the jump processes, all
of which are independent from each other.
To motivate our approach, we quote part of a result due to Doney [5]. It gives an
equivalence for X to remain positive at small times, with probability approaching 1, in
terms of the functions A(x), U (x) and the negative tail of Π. The condition reflects the
positivity of X at small times in that the function A(x) remains positive for small values
of x, and dominates U (x) and the negative tail of Π in a certain way.
−
Theorem 1.1. Suppose Π(R) = ∞. (i) Suppose also that Π (0+) > 0. Then
lim P (Xt > 0) = 1
(1.9)
t↓0
if and only if
A(x)
= ∞.
lim q
−
x↓0
U (x)Π (x)
(1.10)
−
(ii) Suppose alternatively that X is spectrally positive, that is, Π (x) = 0 for all x > 0.
Then (1.9) is equivalent to
σ2 = 0
and
A(x) ≥ 0
for all small x,
(1.11)
and this happens if and only if X is a subordinator. Furthermore, we then have A(x) ≥ 0,
not only for small x, but for all x > 0.
Remarks. (i) Other equivalences for (1.9) are in Theorem 1 of Doney [5] (and his remark
following the theorem). He assumes a priori that σ 2 = 0 but this is not necessary as it
follows from the inequality:
−
A(x)
< ∞,
lim sup q
−
x↓0
Π (x)
whenever Π (0+) > 0, which is proved in Buchmann, Fan and Maller [4].
(1.12)
4
R.A. Maller
(ii) When Π(R) < ∞, X is compound Poisson and its behaviour near 0 is simply
determined by the sign of the shift constant γ. We eliminate this case throughout.
The next section contains our main result which is essentially a subsequential version
of Theorem 1.1.
2. Staying positive near 0, subsequential version
Denote the jump process of X by (∆Xt )t≥0 , where ∆Xt = Xt − Xt− , t > 0, with ∆X0 ≡ 0,
(1)
and define ∆Xt+ = max(∆Xt , 0), ∆Xt− = max(−∆Xt , 0), (∆X + )t = sup0<s≤t ∆Xs+ ,
(1)
(∆X − )t = sup0<s≤t ∆Xs− .
Theorem 2.1. Assume Π(R) = ∞.
−
(i) Suppose Π (0+) > 0. Then the following are equivalent:
there is a non-stochastic sequence tk ↓ 0 such that
P (Xtk > 0) → 1;
(2.1)
there is a non-stochastic sequence tk ↓ 0 such that
Xtk
(1)
(∆X − )tk
P
−→ ∞
as k → ∞;
A(x)
lim sup q
= ∞.
−
x↓0
U (x)Π (x)
(2.2)
(2.3)
−
(ii) Suppose alternatively that X is spectrally positive, that is, Π (x) = 0 for all x > 0.
Then (2.1) is equivalent to limt↓0 P (Xt > 0) → 1, thus to (1.11), equivalently, Xt is a
subordinator, and A(x) ≥ 0 for all x > 0.
−
P
(iii) Suppose Π (0+) > 0. Then Xtk /tk −→ ∞ for a non-stochastic sequence tk ↓ 0 if
and only if
lim sup
x↓0
−
A(x)
q
= ∞.
−
1 + U (x)Π (x)
(2.4)
Remarks. (i) When Π (0+) > 0, sup0<s≤t ∆Xs− > 0 a.s. for all t > 0, so the ratio in
(2.2) is well defined.
(ii) Sato [9], page 65, shows that P (Xt ≤ x) is a continuous function of x for all t > 0
when Π(R) = ∞. So P (Xt > 0) = P (Xt ≥ 0) for all t > 0 and P (Xtk > 0) can be replaced
by P (Xtk ≥ 0) in (2.1) without changing the result (and similarly in Theorem 1.1).
Lévy process staying positive near 0
5
+
−
(iii) Assuming Π (0+) = ∞ and Π (0+) > 0, the contrapositive of (2.1) shows that
there is no sequence tk ↓ 0 such that P (Xtk > 0) → 1, or, equivalently, lim inf t↓0 P (Xt ≤
0) > 0, if and only if
A(x)
< ∞.
lim sup q
−
x↓0
U (x)Π (x)
−
(2.5)
+
By a symmetrical argument, when Π (0+) = ∞ and Π (0+) > 0, then lim inf t↓0 P (Xt ≥
0) > 0 if and only if
A(x)
lim inf q
> −∞.
+
x↓0
U (x)Π (x)
(2.6)
Combining these gives the following.
+
−
Corollary 2.2. Assume Π (0+) = Π (0+) = ∞. Then (1.2) holds if and only if
A(x)
− ∞ < lim inf q
x↓0
+
U (x)Π (x)
+
and
A(x)
lim sup q
< ∞.
−
x↓0
U (x)Π (x)
(2.7)
−
When one of Π (0+) or Π (0+) is infinite but the other is zero, conditions for (1.2)
can also be read from Theorem 2.1.
(ii) A random walk version of Theorem 2.1 is in Kesten and Maller [8]. Andrew [1],
Theorem 4, has results related to Theorem 2.1, including the equivalence of (2.1) and
(2.2).
3. Some inequalities for the distribution of X
For the proof of Theorem 2.1, some lemmas are needed. The first gives a non-uniform
Berry–Esseen bound for a small jump component of X. The proof is rather similar to
that of Lemma 4.3 of Bertoin, Doney and Maller [3], so we omit details.
(−h ,h )
Lemma 3.1. Fix h− ≥ 0, h+ ≥ 0, h− ∨ h+ > 0. Let (Xt − + )t≥0 be the small jump
martingale obtained from X as the compensated sum of jumps with magnitudes in
(−h− , h+ ):
X
Z
(−h ,h )
Xt − + = a.s. lim
xΠ(dx) .
∆Xs 1{∆Xs ∈(−h− ,−ε)∪(ε,h+ )} − t
ε↓0
0<s≤t
x∈(−h− ,−ε)∪(ε,h+ )
(Interpret integrals over intervals of the form (0, −ε), and (ε, 0), ε > 0, as 0.) Define
R
(−h ,h )
absolute moments mk − + := −h− <x<h+ |x|k Π(dx), k = 2, 3, . . . , and assume σ 2 +
6
R.A. Maller
(−h− ,h+ )
m2
> 0. Then we have the non-uniform bound: for any x ∈ R, t > 0,
(−h ,h )
(−h ,h )
Cm3 − +
σZt + Xt − +
≤ x − Φ(x) ≤ √
, (3.1)
P q
(−h ,h )
(−h ,h )
t(σ 2 + m2 − + )3/2 (1 + |x|)3
t(σ 2 + m2 − + )
where C is an absolute constant and Φ(x) is the standard normal c.d.f.
Next, we use Lemma 3.1 to develop other useful bounds. Define
Z
Z
2
y 2 Π(dy),
y Π(dy) and V− (x) =
V+ (x) =
x > 0.
(3.2)
−x≤y<0
0<y≤x
+
In the next lemma, the “+” and “−” signs are to be taken together. When Π (0+) = 0
(S,d ,+)
we have V+ ≡ 0, and interpret (Xt + )t≥0 as 0; similarly with “−” replacing “+”
−
when Π (0+) = 0.
Lemma 3.2. (i) Suppose d± > 0, κ± > 0 and K± are constants satisfying
1
κ±
p
,
K± ≥ 4C max
,
Φ(−κ± ) Φ(−κ± ) 1 − Φ(−κ± )/2
where C is the absolute constant in (3.1). Then for each t > 0
p
(S,d ,±)
P (Xt ± ≤ K± d± − κ± tV± (d± )) ≥ Φ(−κ± )/2.
(3.3)
(3.4)
(ii) Suppose, for each t > 0, d± = d± (t) > 0 satisfy
+
tΠ (d+ ) ≤ c+
and
−
tΠ (d− −) ≥ c−
(3.5)
for some c+ > 0, c− > 0. Assume κ± > 0 and K± are constants satisfying (3.3).
+
(a) Suppose Π (0+) > 0. Then for each t > 0 and L ≥ 0
p
p
P (Xt ≤ tγ − tν+ (d+ ) + tν− (d− ) + K+ d+ − Ld− − κ+ tV+ (d+ ) − κ− tV− (d− ))
≥ e−c+ Φ(−κ+ )Φ(−κ− )P (N (c− ) ≥ K− + L)/8,
(3.6)
where N (c− ) is a Poisson rv with expectation c− .
+
(b) When Π (0+) = 0, (3.6) remains true with ν+ (d+ ) = V+ (d+ ) = d+ = c+ = 0.
−
+
(iii) Suppose 0 ≤ Π (0+) < ∞ = Π (0+) and, for t > 0, d+ = d+ (t) > 0 is such that
+
tΠ (d+ (t)) ≤ c+ . Suppose κ+ > 0 and K+ are constants satisfying (3.3). Then
p
(3.7)
P (Xt ≤ tγ − tν+ (d+ ) + tν− (0) + K+ d+ − κ+ tV+ (d+ )) ≥ e−c+ Φ(−κ+ )/4,
−
where ν− (0) ≡ 0 when Π (0+) = 0.
Lévy process staying positive near 0
7
Proof. (i) We give the proof just for the “+” signs. Fix t > 0 and take any constants
d+ > 0, κ+ > 0 and K+ , with κ+ and K+ satisfying (3.3).
(S,d ,+)
(a) Assume V+ (d+ ) > 0. Apply the bound (3.1) Rin Lemma 3.1 to Xt + , which has
Lévy measure Π restricted to (0, d+ ). Noting that 0<y≤x y 3 Π(dy) ≤ xV+ (x), x > 0, (3.1)
then gives, for each t > 0,
(S,d+ ,+)
sup|P (Xt
x∈R
Substitute x = −κ+ in this to get
(S,d+ ,+)
P (Xt
p
Cd+
≤ x tV+ (d+ )) − Φ(x)| ≤ p
.
tV+ (d+ )
≤ −κ+
p
Cd+
.
tV+ (d+ )) ≥ Φ(−κ+ ) − p
tV+ (d+ )
p
When 2Cd+ ≤ Φ(−κ+ ) tV+ (d+ ), this inequality implies
p
(S,d ,+)
P (Xt + ≤ −κ+ tV+ (d+ )) ≥ 21 Φ(−κ+ ).
p
When 2Cd+ > Φ(−κ+ ) tV+ (d+ ), we have
p
2κ+ tV+ (d+ ) < 4Cd+ κ+ /Φ(−κ+ ) ≤ K+ d+ ,
(3.9)
(S,d+,+)
since K+ satisfies (3.3). Apply Chebychev’s inequality, noting that Xt
0 and variance tV+ (d+ ), to get
(S,d+ ,+)
P (Xt
≤ K+ d+ − κ+
(3.8)
p
tV+ (d+ )) ≥ 1 −
≥1−
has mean
tV+ (d+ )
p
(K+ d+ − κ+ tV+ (d+ ))2
4tV+ (d+ )
.
2 d2
K+
+
p
Also when 2Cd+ > Φ(−κ+ ) tV+ (d+ ), by choice of K+ in (3.3) we have
4tV+ (d+ )
16C 2
Φ(−κ+ )
≤
≤1−
,
2
2
2
2
K+ d+
Φ (−κ+ )K+
2
giving
(S,d+,+)
p
(3.10)
tV+ (d+ )) ≥ 12 Φ(−κ+ ).
p
The same inequality holds when 2Cd+ ≤ Φ(−κ+ ) tV+ (d+ ), by (3.9), so it holds in
general.
(b) When V+ (d+ ) = 0, Π(·) has no mass in (0, d+ ), and (3.4) with a “+” sign remains
(S,d ,+)
valid in the sense that Xt + = 0 a.s. and the left-hand side of (3.4) equals 1. This
proves (3.4) with a “+” sign, and the same argument goes through with “−” in place of
“+”.
P (Xt
≤ K+ d+ − κ+
8
R.A. Maller
(ii) We use the Itô representation in (1.7). Fix t > 0 and take any constants d± > 0
satisfying (3.5). Let κ± > 0 be any constants and choose K± to satisfy (3.3). For the small
+
jump processes, we have the bounds in (3.4). Note that these remain true if Π (0+) = 0
−
or Π (0+) = 0. For the big positive jumps, we have
(B,d+ ,+)
P (Xt
= 0) ≥ P (no ∆Xs exceeds d+ up till time t)
= e−tΠ
+
≥ e−c+
(d+ )
(3.11)
(by (3.5)).
+
Equation (3.11) remains true with c+ = 0 when Π (0+) = 0. By (1.7), the probability
on the left-hand side of (3.6) is, for any L ≥ 0,
(S,d+ ,+)
(B,d+ ,+)
(S,d− ,−)
(B,d ,−)
+ Xt −
p
p
≤ K+ d+ − Ld− − κ+ tV+ (d+ ) − κ− tV− (d− ))
p
(S,d ,+)
(B,d ,+)
≥ P (Zt ≤ 0, Xt + ≤ K+ d+ − κ+ tV+ (d+ ), Xt + = 0,
p
(B,d ,−)
(S,d ,−)
Xt − ≤ K− d− − κ− tV− (d− ), Xt − ≤ −(K− + L)d− )
P (σZt + Xt
+ Xt
+ Xt
(B,d− ,−)
≥ e−c+ Φ(−κ+ )Φ(−κ− )P (Xt
(3.12)
≤ −(K− + L)d− )/8.
In the last inequality, we used (3.4) (twice; once with “+” and once with “−”), (3.11)
(B,d ,−)
(·)
is larger
and the independence of the Zt and the Xt processes. No jump in Xt −
(B,d− ,−)
−
−
than −d− , so we have the upper bound Xt
≤ −d− Nt (d− ), where Nt (d− ) is
the number of jumps of Xt less than or equal in size to −d− which occur by time t.
−
−
Nt− (d− ) is distributed as Poisson with expectation tΠ (d− −), and tΠ (d− −) ≥ c− by
−
(3.5). (Note that this implies Π (0+) > 0.) The Poisson distribution is stochastically
monotone in the sense that if N (µ1 ) and N (µ2 ) are Poisson rvs with means µ1 > µ2 ,
then P (N (µ1 ) ≥ x) ≥ P (N (µ2 ) ≥ x) for all x ≥ 0. So, letting N (c− ) be a Poisson rv with
expectation c− , we have
P (Nt− (d− ) ≥ K− + L) ≥ P (N (c− ) ≥ K− + L).
(3.13)
(B,d− ,−)
(3.14)
Then using
P (Xt
≤ −(K− + L)d− ) ≥ P (Nt− (d− ) ≥ K− + L)
+
and (3.12) we arrive at (3.6). When Π (0+) = 0, we can take all the “+” terms in (3.12)
as 0 to get (3.6) with all the “+” terms 0.
−
+
(iii) Assume 0 ≤ Π (0+) < ∞ = Π (0+). In this case, we do not define d− but still
+
have d+ = d+ (t) > 0 and assume tΠ (d+ ) ≤ c+ as in (3.5). From (1.7), write
(S,d+ ,+)
Xt = tγ − tν+ (d+ ) + tν− (0) + Xt
(B,d+ ,+)
+ Xt
(0,−)
+ Xt
,
(3.15)
Lévy process staying positive near 0
9
where the negative jump components have been amalgamated into
X
(0,−)
Xt
:=
∆Xs 1{∆Xs ≤0} ,
t > 0,
0<s≤t
which is a compound Poisson process comprised of non-positive jumps. This term and
−
the term tν− (0) are absent from (3.15) when Π (0+) = 0. Using (3.4), (3.11) and (3.15),
write
p
P (Xt ≤ tγ − tν+ (d+ ) + tν− (0+) + K+ d+ − κ+ tV+ (d+ ))
p
(S,d ,+)
(B,d ,+)
(0,−)
≥ P (Zt ≤ 0, Xt + ≤ K+ d+ − κ+ tV+ (d+ ), Xt + = 0, Xt
≤ 0)
(0,−)
≥ e−c+ Φ(−κ+ )P (Xt
≤ 0)/4 = e−c+ Φ(−κ+ )/4
and this gives (3.7).
4. Proof of Theorem 2.1
−
Part (i). Assume Π (0+) > 0 throughout this part.
−
−
(2.3) =⇒ (2.1): Assume (2.3). Π (0+) > 0 implies Π (x) > 0 in a neighbourhood of 0
−
so we can assume Π (x) > 0 for all 0 < x < 1. Choose 1 > xk ↓ 0 such that
A(xk )
q
→∞
−
U (xk )Π (xk )
as k → ∞. This implies σ 2 = 0 by (1.12) (because U (x) ≥ σ 2 ). It also means that A(xk ) >
0 for all large k, and without loss of generality we may assume it to be so for all k. Let
s
U (xk )
sk :=
,
−
Π (xk )A2 (xk )
then
−
sk Π (xk ) =
q
−
U (xk )Π (xk )
A(xk )
→0
−
and since Π (0+) > 0, also sk → 0 as k → ∞. In addition, we have
q
−
U (xk )Π (xk )
U (xk )
=
→0
sk A2 (xk )
A(xk )
and
sk A(xk )
=
xk
s
U (xk )
−
2
xk Π (xk )
≥ 1.
10
R.A. Maller
Set
tk :=
s
sk
−
Π (xk )
,
−
so tk /sk → ∞, but still tk Π (xk ) → 0, as k → ∞. Then
sk U (xk )
U (xk )
=
→ 0,
tk A2 (xk ) tk sk A2 (xk )
(4.1)
tk A(xk ) tk sk A(xk )
=
→ ∞,
xk
sk
xk
(4.2)
and
as k → ∞.
Recall (1.6) and use the Itô decomposition in (1.7) with σ 2 = 0 and h+ = h− = h > 0
to write
(S,h)
Xt = tA(h) + Xt
(S,h)
(B,h,+)
+ Xt
(S,h,+)
+
(B,h,−)
− thΠ (h) + Xt
−
+ thΠ (h),
t > 0.
(S,h,−)
(4.3)
(B,h,±)
Here, Xt
= Xt
+ Xt
is the compensated small jump process, and Xt
are the positive and negative big jump processes.
+
(B,h,+)
Case (a): Suppose Π (0+) > 0. Since each jump in Xt
is at least h, we have the
+
(B,h,+)
lower bound Xt
≥ hNt+ (h), where Nt+ (h) is Poisson with expectation tΠ (h) (and
+
variance tΠ (h)). Using this and substituting in (4.3) with t = tk and h = xk we get
(S,xk )
Xtk ≥ tk A(xk ) + Xtk
−
(B,xk ,−)
Since tk Π (xk ) → 0, we have P (Xtk
(S,xk )
P (Xtk
+
(B,xk ,−)
+ xk (Nt+k (xk ) − tk Π (xk )) + Xtk
.
(4.4)
= 0) → 1 as k → ∞. Also, for ε ∈ (0, 1),
+
+
+ xk (Nt+k (xk ) − tk Π (xk )) ≤ −εtk A(xk )) ≤
≤
tk V (xk ) + tk x2k Π (xk )
ε2 t2k A2 (xk )
U (xk )
→ 0,
ε2 tk A2 (xk )
as k → ∞ by (4.1), so
Xtk
tk A(xk )
P
≥ (1 − ε)
xk
xk
P
→ 1,
and hence, by (4.2), Xtk /xk −→ ∞ as k → ∞. Thus, (2.1) holds.
(4.5)
Lévy process staying positive near 0
11
+
Case (b): Alternatively, if Π (0+) = 0, we can omit the term containing Nt+k (xk ) −
+
tk Π (xk ) in (4.4) and in what follows it, and again obtain (4.5), and hence (2.1).1
−
(2.3) =⇒ (2.2): Continuing the previous argument, tk Π (xk ) → 0 implies
−
(1)
P ((∆X − )tk > xk ) = P sup ∆Xs− > xk = 1 − e−tk Π (xk ) → 0
as k → ∞,
0<s≤tk
−
so, using (4.5), (2.2) also holds when (2.3) holds and Π (0+) > 0.
−
(2.2) =⇒ (2.1): This is obvious when Π (0+) > 0.
−
(2.1) =⇒ (2.3): Assume Π(R) = ∞ as well as Π (0+) > 0, and that (2.1) holds. Suppose
(2.3) fails, so we can choose 1 < a < ∞, x0 > 0, such that
q
−
A(x) ≤ a U (x)Π (x),
(4.6)
for all 0 < x ≤ x0 . We will obtain a contradiction. Note that (2.1) implies σ 2 = 0, because
√ D
Xt / t −→ N (0, σ 2 ), a non-degenerate normal rv, when σ 2 > 0. So we assume σ 2 = 0 in
what follows. We consider 3 cases.
−
+
Case (a): Assume in fact that Π (0+) = ∞ = Π (0+). In this situation, we can introduce quantile versions for the d± in (3.5). Define the non-decreasing function
+
d+ (t) := inf{x > 0 : Π (x) ≤ t−1 },
t > 0,
(4.7)
+
and set d+ (0) = 0. Since Π (0+) = ∞, we have 0 < d+ (t) < ∞ for all t > 0, d+ (t) ↓ 0 as
t ↓ 0, and
+
+
tΠ (d+ (t)) ≤ 1 ≤ tΠ (d+ (t)−)
for all t > 0.
(4.8)
Analogously, define d− (0) = 0, and
−
d− (t) := inf{x > 0 : Π (x) ≤ t−1 },
t > 0,
(4.9)
−
having, since Π (0+) = ∞, 0 < d− (t) < ∞, d− (t) ↓ 0 as t ↓ 0, and
−
−
tΠ (d− (t)) ≤ 1 ≤ tΠ (d− (t)−).
(4.10)
With a as in (4.6), set κ+ = κ− = κ = 2a, then choose K± to satisfy (3.3). Then (2.1)
together with (3.6) shows that we must have
0 ≤ tk (γ − ν+ (d+ ) + ν− (d− )) + K+ d+ − Ld− − κ
1 Observe
q
tk (V+ (d+ ) + V− (d− )),
(4.11)
that the assumption Π(R) = ∞ was not used in this part of the proof. The trivial case,
Xt = tγ, γ > 0, when A(x) ≡ γ, is included if we interpret (2.3) as holding then.
12
R.A. Maller
for all√large√k. Here, d+ and d− are any positive numbers and we used the inequality
√
a + b ≥ a + b, a, b > 0, in (3.6). Take λ > 0 and set
d+ = d+ (λtk )
and d− = d− (tk ),
where d+ (·) and d− (·) are defined in (4.7) and (4.9). By (4.8) and (4.10), we then have
+
tΠ (d+ (λt)) ≤ λ−1
−
and tΠ (d− (t)−) ≥ 1,
so (3.5) holds with c+ = λ−1 and c− = 1. With tk as the sequence in (2.1), let d = d(tk ) :=
max(d+ (λtk ), d− (tk )).
Equation (4.11) implies
Z
Z
(−)
yΠ (dy)
yΠ(dy) +
tk γ −
(d− ,1]
(d+ ,1]
Adding the quantity
(4.12)
q
≥ −K+ d+ + Ld− + κ tk (V+ (d+ ) + V− (d− )).
tk
Z
(d+ ,d]
yΠ(dy) −
Z
(−)
yΠ
(d− ,d]
(dy)
to both sides of (4.12) gives tk ν(d) on the left, and a quantity no smaller than
q
+
+
−
−
tk d+ (Π (d+ )− Π (d)) − tk d(Π (d− )− Π (d)) − K+ d+ + Ld− + κ tk (V+ (d+ ) + V− (d− ))
+
−
on the right. Further adding tk d(Π (d) − Π (d)) to both sides gives tk A(d) on the left
(see (1.6)), and then after some cancellation we arrive at
q
+
−
tk A(d) ≥ tk d+ Π (d+ ) − tk dΠ (d− ) − K+ d+ + Ld− + κ tk (V+ (d+ ) + V− (d− )). (4.13)
+
At this stage, it is helpful to assume that Π (x) is a continuous function on (0, ∞).
+
−
It then follows from (4.8) that tk Π (d+ (λtk )) = 1/λ, while tk Π (d− (tk )) ≤ 1 by (4.10).
Also, d ≤ d+ + d− . Thus, we deduce
q
tk A(d) ≥ (1/λ − K+ − 1)d+ + (L − 1)d− + κ tk (V+ (d+ ) + V− (d− )).
(4.14)
Next, write
V+ (d+ ) + V− (d− ) = V+ (d) −
Z
(d+ ,d]
y 2 Π(dy) + V− (d) −
+
+
+
−
Z
y 2 Π(−) (dy)
(d− ,d]
−
−
≥ V (d) − d2 (Π (d+ ) − Π (d)) − d2 (Π (d− ) − Π (d))
= U (d) − d2 (Π (d+ ) + Π (d− )).
(4.15)
Lévy process staying positive near 0
13
So
tk (V+ (d+ ) + V− (d− )) ≥ tk U (d) − d2 (1/λ + 1)
giving
q
p
√
tk (V+ (d+ ) + V− (d− )) ≥ tk U (d) − d(1/ λ + 1).
Substituting into (4.14), we obtain
√
p
tk U (d) + (1/λ − K+ − 1 − κ/ λ − κ)d+
√
+ (L − 1 − κ/ λ − κ)d− .
tk A(d) ≥ κ
Choose λ small enough for the first expression in brackets on the right-hand side to
be positive. Then choose L large enough for the second expression in brackets on the
right-hand side to be positive. This gives
p
tk A(d) ≥ κ tk U (d),
(4.16)
for all large k, where d = d(tk ) = max(d+ (λtk ), d− (tk )) ↓ 0 as k → ∞. Inequality (4.16)
implies
A(d)
κ
κ
q
≥q
≥q
≥ κ,
−
−
−
U (d)Π (d)
tk Π (d)
tk Π (d− )
(4.17)
giving a contradiction with (4.6), since κ = 2a.
+
−
+
This proves (2.3) from (2.1) in case Π (0+) = Π (0+) = ∞ and Π (x) is continuous
for x > 0. To complete the proof of part (i), case (a), of the theorem we remove the
assumption of continuity made in deriving (4.14). This can be done using the following
lemma.
+
Lemma 4.1. Let Π be any Lévy measure with Π (0+) = ∞. Then there exists a sequence
of Lévy measures Πn , absolutely continuous with respect to Lebesgue measure, and having
+
v
strictly positive C ∞ -densities on R∗ := R \ {0}, satisfying Πn (0+) = ∞ and Πn −→ Π as
n → ∞.
v
Proof. (−→ refers to vague convergence in R∗ ; see, for example, Chapter 15 in Kallenberg [7].)R We extend Π to a Borel measure on R by setting Π({0}) := 0. Assume Π 6= 0,
so C := x2 Π(dx)/(1 + x2 ) ∈ (0, ∞). Observe that P (dx) := x2 Π(dx)/C(1 + x2 ) defines
a Borel probability measure on R. For all n ∈ N, the convolved probability measure
Pn := P ⋆ N (0, 1/n) admits a strictly positive C ∞ -Lebesgue density, when N (0, 1/n) is
a normal rv with expectation 0 and variance 1/n. Set Πn (dx) := C(1 + x2 )Pn (dx)/x2 ,
n = 1, 2, . . . . It is easily verified that (Πn )n∈N is a sequence of Lévy measures with the
desired properties.
14
R.A. Maller
Now to complete the proof of part (i), assume (2.1), that is, P (Xtk ≥ 0) → 1 for some
tk ↓ 0, for a general X with Lévy measure Π. Using Lemma 4.1, construct a sequence of
approximating Lévy measures Πn , converging vaguely to Π, such that their positive tails
+
+
−
−
Πn are continuous on (0, ∞) with Πn (0+) = ∞. On the negative side, let Πn (x) = Π (x),
x > 0. Let (Xt (n))t≥0 be Lévy processes with measures Πn and other characteristics the
same as for X. Define νn , An , Vn , Un , νn,± , Vn,± , as in (1.4), (1.5), (1.8) and (3.2), but
with Πn replacing Π. The subscript n functions converge to the original functions at
points of continuity of the latter. Xt (n) has characteristic exponent given by (1.1) with
D
Πn replacing Π, so as n → ∞ we have Xt (n) −→ Xt for each t > 0. Under assumption
(2.1), limn→∞ P (Xtk (n) > 0) = P (Xtk > 0) > 1 − δ for arbitrary δ ∈ (0, 1/2) and k large
enough. Thus, P (Xtk (n) > 0) > 1 − 2δ for n ≥ n0 (k) and k ≥ k0 . So (4.11) holds for the
subscript n quantities with probability 1 − 2δ. But (4.11) is deterministic so it holds in
fact for the subscript n quantities (with probability 1) whenever n ≥ n0 (k) and k ≥ k0 .
−
The proof using continuity of Πn then shows that (4.17) holds with A, U and Π replaced
−
by An , Un and Πn . Then letting n → ∞ shows that (4.17) itself holds as stated for k ≥ k0 .
Again we get a contradiction, and thus complete the proof that (2.1) implies (2.3) for
case (a).
+
−
Case (b): Assume that 0 ≤ Π (0+) < ∞ = Π (0+). As in the proof for case (a), we
take κ+ = κ− = κ > 2a, K± to satisfy (3.3), define d− (t) > 0 by (4.9), and write d− =
d− (t) for t > 0. But for d+ we set d+ (t) ≡ d− (t) > 0. We take c+ = 1 in (3.5) as we may
+
+
+
since tΠ (d+ ) = tΠ (d− ) ≤ tΠ (0+) → 0 as t ↓ 0. With this set-up, (4.11) is true (with
d− replacing d+ ) and we can follow the proof of case (a) through to get (4.13) with d
and d+ replaced by d− ; thus,
q
+
−
tk A(d− ) ≥ tk d− (Π (d− ) − Π (d− )) − K+ d− + Ld− + κ tk (V+ (d− ) + V− (d− )). (4.18)
Estimating V± along the lines of (4.15) we find the right-hand side of (4.18) is not smaller
than
p
κ tk U (d− ) + (L − K+ − 1 − κ)d− .
Choose L large enough in this to get (4.16), and hence (4.17) with d− in place of d, hence
(2.3) again.
−
+
Case (c): Assume that 0 < Π (0+) < ∞ = Π (0+). Define d+ (t) by (4.7), so we have
(4.8). Then (3.7) with c+ = 1 and κ+ = κ = 2a, together with (2.1), shows that we must
have
p
0 ≤ tk (γ − ν+ (d+ ) + ν− (0)) + K+ d+ − κ tk V+ (d+ ),
(4.19)
for all large k. Here again we write d+ = d+ (λtk ) for λ > 0. Inequality (4.19) implies
Z
Z
p
yΠ(−) (dy) ≥ −K+ d+ + κ tk V+ (d+ ).
yΠ(dy) +
tk γ −
(4.20)
(d+ ,1]
(0,1]
Lévy process staying positive near 0
Subtracting the quantity
Z
tk
15
−
(0,d+ ]
−
yΠ(−) (dy) ≤ tk d+ (Π (0+) − Π (d+ ))
from both sides of (4.20) gives tk ν(d+ ) on the left, and a quantity no smaller than
−
−
−tk d+ (Π (0+) − Π (d+ )) − K+ d+ + κ
+
−
p
tk V+ (d+ )
on the right. Further adding tk d+ (Π (d+ ) − Π (d+ )) to both sides gives (see (1.6))
+
−
tk A(d+ ) ≥ tk d+ Π (d+ ) − tk d+ Π (0+) − K+ d+ + κ
+
p
tk V+ (d+ ).
(4.21)
At this stage, as before, assume Π (x) is continuous. It then follows from (4.8) that
+
−
tk Π (d+ (λtk )) = 1/λ, while tk Π (0+) ≤ 1 for large k. Thus, from (4.21) we deduce
p
(4.22)
tk A(d+ ) ≥ (1/λ − K+ − 1)d+ + κ tk V+ (d+ ),
for large enough k. Further,
+
−
tk V+ (d+ ) = tk (U (d+ ) − V− (d+ ) − d2+ Π (d+ ) − d2+ Π (d+ ))
−
+
≥ tk U (d+ ) − tk d2+ (2Π (0+) + Π (d+ ))
≥ tk U (d+ ) − 4d2+ ,
−
using that V− (d+ ) ≤ d2+ Π (0+). So
p
p
tk V+ (d+ ) ≥ tk U (d+ ) − 2d+ .
Substituting into (4.22), we get
p
tk A(d+ ) ≥ κ tk U (d+ ) + (1/λ − K+ − 1 − 2κ)d+ ,
for large k. We can choose λ small enough for the expression in brackets on the right-hand
side to be positive. This gives
p
tk A(d+ ) ≥ κ tk U (d+ ),
(4.23)
−
which, since Π (0+) > 0 is assumed, implies
A(d+ )
κ
q
≥q
→∞
−
−
U (d+ )Π (d+ )
tk Π (0+)
as k → ∞,
a contradiction with (4.6). We can remove the continuity assumption as before. So (2.3)
−
+
is proved when 0 < Π (0+) < ∞ = Π (0+).
16
R.A. Maller
−
+
Part (ii). Now we will deal with the case when Π (0+) = 0 but Π (0+) = ∞. Again
+
assume Π (x) is continuous. The working in case (c) is still valid from (4.19) to (4.23).
The negative jump process is now absent from Xt and (4.19) gives, with d+ = d+ (λt),
0 ≤ tk (γ − ν+ (d+ )) + K+ d+
Z 1
+
+
+
= tk γ −
Π (y) dy − d+ Π (d+ ) + Π (1) + K+ d+
(4.24)
d+
Z
≤ tk γ −
1
d+
Π (y) dy + Π (1) − (1/λ − K+ )d+ ,
+
+
+
+
using tk Π (d+ (λtk )) = 1/λ in the last inequality (since Π (x) is continuous). Then choosing λ < 1/K+ we get
Z
1
d+
+
+
Π (y) dy ≤ γ + Π (1).
R1 +
Letting k → ∞ (so tk ↓ 0 and d+ = d+ (λtk ) ↓ 0) shows that 0 Π (y) dy < ∞. Since Xt
has no negative jumps, from this we deduce that X is of bounded variation with drift
dX = A(0+) (see Doney and Maller [6], Theorem 2.1 and Remark 1), which is nonnegative by (4.23). Thus, X is a subordinator with non-negative drift. It follows that
+
A(x) = γ + Π (1) −
Z
1
x
+
Π (y) dy = dX +
Z
x
+
Π (y) dy
0
is non-negative for all x > 0. This is proved assuming continuity of Πn but that assumption can be removed as before. Then A(x) ≥ 0 together with σ 2 = 0 implies
limt↓0 P (Xt > 0) → 1 as t ↓ 0 by Theorem 1.1, hence (2.1).
−
Part (iii). Finally, suppose Π (0+) > 0 and (2.4) holds. Then we can choose xk ↓ 0
such that
A(xk )
q
→ ∞ and A(xk ) → ∞,
−
U (xk )Π (xk )
P
as k → ∞. Following exactly the proof of part (i), we get (4.5), and this implies Xtk /tk −→
∞, since A(xk ) → ∞.
P
Conversely, suppose Xtk /tk −→ ∞ as k → ∞ for a non-stochastic sequence tk ↓ 0. Then
M
limk→∞ P (Xtk > 0) = 1 for every M > 0, where XtM is Lévy with triplet (γ − M, σ 2 , Π).
Consequently, (2.3) holds with A, U , Π replaced by AM (·) = A(·)− M , U M = U , ΠM = Π,
and this modified version implies (2.4). This completes part (iii), and the proof of the
theorem.
Lévy process staying positive near 0
17
Acknowledgements
I am grateful to a referee for a close reading of the paper and helpful suggestions, and to
Boris Buchmann for supplying Lemma 4.1.
Research partially supported by ARC Grant DP1092502.
References
[1] Andrew, P. (2008). On the limiting behaviour of Lévy processes at zero. Probab. Theory
Related Fields 140 103–127. MR2357672
[2] Aurzada, F., Döring, L. and Savov, M. (2013). Small time Chung-type LIL for Lévy
processes. Bernoulli 19 115–136. MR3019488
[3] Bertoin, J., Doney, R.A. and Maller, R.A. (2008). Passage of Lévy processes across
power law boundaries at small times. Ann. Probab. 36 160–197. MR2370602
[4] Buchmann, B., Fan, Y. and Maller, R.A. (2015). Distributional representations and
dominance of a Lévy process over its maximal jump processes. Bernoulli. To appear.
Available at arXiv:1409.4050.
[5] Doney, R.A. (2004). Small-time behaviour of Lévy processes. Electron. J. Probab. 9 209–
229. MR2041833
[6] Doney, R.A. and Maller, R.A. (2002). Stability and attraction to normality for Lévy
processes at zero and at infinity. J. Theoret. Probab. 15 751–792. MR1922446
[7] Kallenberg, O. (2002). Foundations of Modern Probability, 2nd ed. Probability and Its
Applications (New York). New York: Springer. MR1876169
[8] Kesten, H. and Maller, R.A. (1997). Divergence of a random walk through deterministic
and random subsequences. J. Theoret. Probab. 10 395–427. MR1455151
[9] Sato, K. (1999). Lévy Processes and Infinitely Divisible Distributions. Cambridge: Cambridge Univ. Press.
[10] Wee, I.S. (1988). Lower functions for processes with stationary independent increments.
Probab. Theory Related Fields 77 551–566. MR0933989
Received September 2014
| 10math.ST
|
Extending Prolog with
Incomplete Fuzzy Information
arXiv:cs/0508091v1 [cs.PL] 22 Aug 2005
Susana Munoz-Hernandez1 and Claudio Vaucheret2
1
Departamento de Lenguajes, Sistemas de la Información e Ingenierı́a del Software
Facultad de Informática - Universidad Politécnica de Madrid
Campus de Montegancedo 28660 Madrid, Spain
[email protected]
2
Departamento de Ciencias de la Computación
Facultad de Economı́a y Administración
Universidad Nacional del Comahue
Universidad Politécnica de Madrid
Buenos Aires 1400 (8300) Neuquen, Argentina
[email protected]
Abstract. Incomplete information is a problem in many aspects of actual environments. Furthermore, in many sceneries the knowledge is not
represented in a crisp way. It is common to find fuzzy concepts or problems with some level of uncertainty. There are not many practical systems which handle fuzziness and uncertainty and the few examples that
we can find are used by a minority. To extend a popular system (which
many programmers are using) with the ability of combining crisp and
fuzzy knowledge representations seems to be an interesting issue.
Our first work (Fuzzy Prolog) was a language that models B([0, 1])-valued
Fuzzy Logic. In the Borel algebra, B([0, 1]), truth value is represented
using unions of intervals of real numbers. This work was more general in
truth value representation and propagation than previous works.
An interpreter for this language using Constraint Logic Programming
over Real numbers (CLP(R)) was implemented and is available in the
Ciao system .
Now, we enhance our former approach by using default knowledge to represent incomplete information in Logic Programming. We also provide
the implementation of this new framework. This new release of Fuzzy
Prolog handles incomplete information, it has a complete semantics (the
previous one was incomplete as Prolog) and moreover it is able to combine crisp and fuzzy logic in Prolog programs. Therefore, new Fuzzy
Prolog is more expressive to represent real world.
Fuzzy Prolog inherited from Prolog its incompleteness. The incorporation of default reasoning to Fuzzy Prolog removes this problem and requires a richer semantics which we discuss.
Keywords Incomplete knowledge, Fuzzy Prolog, Modeling Uncertainty, Fuzzy
Logic Programming, Constraint Programming Application, Implementation of
Fuzzy Prolog.
1
Introduction
World information is not represented in a crisp way. Its representation is imperfect, fuzzy, etc., so that the management of uncertainty is very important
in knowledge representation. There are multiple frameworks for incorporating
uncertainty in logic programming:
–
–
–
–
fuzzy set theory [2, 35, 30],
probability theory [6, 17, 20, 22, 23],
multi-valued logic [5, 13, 14, 16, 18, 26],
possibilistic logic [4, 33, 34]
In [18] a general framework was proposed that generalizes many of the previous approaches. At the same time an analogous theoretical framework was
provided and a prototype for Prolog was implemented [31]. Basically, a rule is of
the form A ← B1 , . . . , Bn , where the assignment I of certainties is taken from a
certainty lattice, to the Bi s. The certainty of A is computed by taking the set of
the certainties I(Bi ) and then they are propagated using the function F that is
an aggregation operator. This is a very flexible approach and in [32, 9] practical
examples in a Prolog framework are presented.
In this work we extend the approach of [9] with arbitrary assignments of
default certainty values (non-uniform default assumptions). The usual semantics
of logic programs can be obtained through a unique computation method, but
using different assumptions in a uniform way to assign the same default truthvalue to all the atoms. The most well known assumptions are:
– the Closed World Assumption (CWA), which asserts that any atom whose
truth-value cannot be inferred from the facts and clauses of the program is
supposed to be false (i.e. certainty 0). It is used in stable models [7, 8] and
well-founded semantics [19, 20, 22],
– the Open World Assumption (OWA), which asserts that any atom whose
truth-value cannot be inferred from the facts and clauses of the program is
supposed to be undefined or unknown (i.e. certainty in [0, 1]). It is used in
[18].
There are also some approaches [33, 34] where both assumptions can be combined and some atoms can be interpreted assuming CWA while others follows
OWA. Anyway, what seems really interesting is not only to combine both assumptions but to generalize the use of a default value. The aim is working with
incomplete information with more guarantees.
The rest of the paper is organized as follows. Section 2 introduces the Fuzzy
Prolog language. A complete description of the new semantics of Fuzzy Prolog
is provided in Section 3. Section 4 completes the details about the improved
implementation using CLP(R) with the extension to handle default knowledge.
Some illustrating examples are provided in section 5. Finally, we conclude and
discuss some future work (Section 6).
This research was partly supported by the Spanish MCYT project TIC2003-01036.
2
Fuzzy Prolog
In this section we are going to summarize the main characteristics of the Fuzzy
Prolog that we proposed in [9] and that is the basis of the work presented here.
Fuzzy Prolog is more general than previous approaches to introduce fuzziness in
Prolog in some respects:
1. A truth value will be a finite union of closed sub-intervals on [0, 1]. This
is represented by Borel algebra, B([0, 1]), while the algebra E([0, 1]) only
considers intervals. An interval is a special case of the union of one element,
and a unique truth value is a particular case of having an interval with only
one element.
2. A truth value will be propagated through the rules by means of an aggregation operator. The definition of aggregation operator is general in the sense
that it subsumes conjunctive operators (triangular norms [15] like min, prod,
etc.), disjunctive operators [28](triangular co-norms, like max, sum, etc.), average operators (like arithmetic average, quasi-linear average, etc) and hybrid
operators (combinations of the above operators [25]).
3. The declarative and procedural semantics for Fuzzy Logic programs are given
and their equivalence is proven.
4. An implementation of the proposed language is presented. A fuzzy program
is a finite set of
– fuzzy facts (A ← v, where A is an atom and v, a truth value, is an
element in B([0, 1]) expressed as constraints over the domain [0, 1]), and
– fuzzy clauses (A ←F B1 , . . . , Bn , where A, B1 , . . . , Bn are atoms, and F
is an interval-aggregation operator, which induces a union-aggregation,
as by Definition 22, F of truth values in B([0, 1]) represented as constraints over the domain [0, 1]).
We obtain information from the program through fuzzy queries or fuzzy goals
(v ← A ? where A is an atom, and v is a variable, possibly instantiated, that
represents a truth value in B([0, 1])).
Programs are defined as usual but handling truth values in B([0, 1]) (the
Borel algebra over the real interval [0, 1] that deals with unions of intervals) represented as constraints. We refer, for example, to expressions as: (v ≥
S 0.5 ∧ v ≤
0.7) ∨ (v ≥ 0.8 ∧ v ≤ 0.9) to represent a truth value in [0.5, 0.7]
[0.8, 0.9].
A lot of everyday situations can only be represented by this general representation of truth value. There are some examples in [9].
The truth value of a goal will depend on the truth value of the subgoals
which are in the body of the clauses of its definition. Fuzzy Prolog [9] uses
aggregation operators [29] in order to propagate the truth value by means of the
fuzzy rules. Fuzzy sets aggregation is done using the application of a numeric
operator of the form f : [0, 1]n → [0, 1]. An aggregation operator must verify
f (0, . . . , 0) = 0 and f (1, . . . , 1) = 1, and in addition it should be monotonic and
continuous. If we deal with the definition of fuzzy sets as intervals it is necessary
to generalize from aggregation operators of numbers to aggregation operators of
intervals. Following the theorem proven by Nguyen and Walker in [24] to extend
T-norms and T-conorms to intervals, we propose the following definitions.
Definition 21 (interval-aggregation) Given an aggregation f : [0, 1]n →
[0, 1], an interval-aggregation F : E([0, 1])n → E([0, 1]) is defined as follows:
F ([xl1 , xu1 ], ..., [xln , xun ]) = [f (xl1 , ..., xln ), f (xu1 , ..., xun )].
Actually, we work with union of intervals and propose the definition:
Definition 22 (union-aggregation) Given an interval-aggregation
F : E([0, 1])n → E([0, 1]) defined over intervals, a union-aggregation
F : B([0, 1])n → B([0, 1]) is defined over union of intervals as follows:
F (B1 , . . . , Bn ) = ∪{F (E1 , ..., En ) | Ei ∈ Bi }.
In the presentation of the theory of possibility [36], Zadeh considers that
fuzzy sets act as an elastic constraint on the values of a variable and fuzzy
inference as constraint propagation.
In [9] (and furthermore in the extension that we presented in this paper),
truth values and the result of aggregations are represented by constraints. A
constraint is a Σ-formula where Σ is a signature that contains the real numbers,
the binary function symbols + and ∗, and the binary predicate symbols =, <
and ≤. If the constraint c has solution in the domain of real numbers in the
interval [0, 1] then c is consistent, and is denoted as solvable(c).
3
Semantics
This section contains a reformulation of the semantics of Fuzzy Prolog. This new
semantics is complete thanks to the inclusion of default value.
3.1
Least Model Semantics
The Herbrand universe U is the set of all ground terms, which can be made up
with the constants and function symbols of a program, and the Herbrand base B
is the set of all ground atoms which can be formed by using the predicate symbols
of the program with ground terms (of the Herbrand universe) as arguments.
Definition 31 (default value) We assume there is a function default which
implement the Default Knowledge Assumptions. It assigns an element of B([0, 1])
to each element of the Herbrand Base. If the Closed World Assumption is used,
then default (A) = [0, 0] for all A in Herbrand Base. If Open World Assumption
is used instead, default (A) = [0, 1] for all A in Herbrand Base.
Definition 32 (interpretation) An interpretation I = hBI , VI i consists of
the following:
1. a subset BI of the Herbrand Base,
2. a mapping VI , to assign
(a) a truth value, in B([0, 1]), to each element of BI , or
(b) default (A), if A does not belong to BI .
Definition 33 (interval inclusion ⊆II ) Given two intervals I1 = [a, b], I2 =
[c, d] in E([0, 1]), I1 ⊆II I2 if and only if c ≤ a and b ≤ d.
Definition 34 (Borel inclusion ⊆BI ) Given two unions of intervals U = I1 ∪
′
. . .∪IN , U ′ = I1′ ∪. . .∪IM
in B([0, 1]), U ⊆BI U ′ if and only if ∀Ii ∈ U , i ∈ 1..N ,
Ii can be partitioned in to intervals Ji1 , ..., JiL , i.e. Ji1 ∪...∪JiL = Ii , Ji1 ∩...∩JiL
is the set of the border elements of the intervals except the lower limit of Ji1 and
′
′
the upper limit of JiL ) and for all k ∈ 1..L, ∃Jjk
∈ U ′ . Jik ⊆II Jjk
where
jk ∈ 1..M .
The Borel algebra B([0, 1]) is a complete lattice under ⊆BI (Borel inclusion),
and the Herbrand base is a complete lattice under ⊆ (set inclusion) and so the
set of all interpretations forms a complete lattice under the relation ⊑ defined
as follows.
Notice that we have redefined interpretation and Borel inclusion with respect
to the definitions in [9]. We will also redefine the operational semantics and
therefore the internal implementation of the Fuzzy Prolog library. Sections below
are completely new too. For uniformity reasons we have kept the same syntax
that was used in [9] in fuzzy programs.
Definition 35 (interpretation inclusion ⊑) Let I = hBI , VI i and
I ′ = hBI ′ , VI ′ i be interpretations. I ⊑ I ′ if and only if BI ⊆ BI ′ and for all
B ∈ BI , VI (B) ⊆BI VI ′ (B).
Definition 36 (valuation) A valuation σ of an atom A is an assignment of
elements of U to variables of A. So σ(A) ∈ B is a ground atom.
In the Herbrand context, a valuation is the same as a substitution.
Definition 37 (model) Given an interpretation I = hBI , VI i,
– I is a model for a fuzzy fact A ← v, if for all valuations σ, σ(A) ∈ BI and
v ⊆BI VI (σ(A)).
– I is a model for a clause A ←F B1 , . . . , Bn when the following holds: for all
valuations σ, σ(A) ∈ BI and v ⊆BI VI (σ(A)), where v = F (VI (σ(B1 )), . . . ,
VI (σ(Bn ))) and F is the union aggregation obtained from F .
– I is a model of a fuzzy program, if it is a model for the facts and clauses
of the program.
Every program has a least model which is usually regarded as the intended
interpretation of the program since it is the most conservative model. Let ∩
(that appears in the following theorem) be the meet operator on the lattice of
interpretations (I, ⊑). We can prove the following result.
Theorem 31 (model intersection property) Let I1 = hBI1 , VI1 i, I2 = hBI1 , VI1 i
be models of a fuzzy program P . Then I1 ∩ I2 is a model of P .
Proof. Let M = hBM , VM i = I1 ∩ I2 . Since I1 and I2 are models of P , they are
models for each fact and clause of P . Then for all valuations σ we have
– for all facts A ← v in P ,
• σ(A) ⊆ BI1 and σ(A) ∈ BI2 , and so σ(A) ∈ BI1 ∩ BI2 = BM ,
• v ⊆BI VI1 (σ(A)) and v ⊆BI VI2 (σ(A)), and so hence
v ⊆BI VI1 (σ(A)) ∩ VI2 (σ(A)) = VM (σ(A))
therefore M is a model for A ← v
– and for all clauses A ←F B1 , . . . , Bn in P
• since σ(A) ∈ BI1 and σ(A) ∈ BI2 , hence σ(A) ∈ BI1 ∩ BI2 = BM .
• if v = F (VM (σ(B1 )), . . . , VM (σ(Bn ))), since F is monotonic, v ⊆BI
VI1 (σ(A)) and v ⊆BI VI2 (σ(A)), hence v ⊆BI VI1 (σ(A)) ∩ VI2 (σ(A)) =
VM (σ(A))
therefore M is a model for A ←F B1 , . . . , Bn
and M is model of P .
Remark 31 (Least model semantic) If we let M T
be the set of all models of
a program P , the intersection of all of these models, M, is a model and it is
the least model of P . We denote the least model of a program P by lm(P ).
3.2
Fixed-Point Semantics
The fixed-point semantics we present is based on a one-step consequence operator
TP . The least fixed-point lfp(TP ) = I (i.e. TP (I) = I) is the declarative meaning
of the program P , so is equal to lm(P ). We include it here for clarity reasons
although it is the same that in [9].
Let P be a fuzzy program and BP the Herbrand base of P ; then the mapping
TP over interpretations is defined as follows:
Let I = hBI , VI i be a fuzzy interpretation,
then TP (I) = I ′ , I ′ = hBI ′ , VI ′ i,
S
BI ′ = {A ∈ BP | Cond}, VI ′ (A) = {v ∈ B([0, 1]) | Cond}
where
Cond = (A ← v is a ground instance of a fact in P and solvable(v))
or
(A ←F A1 , . . . , An is a ground instance of a clause in P,
and solvable(v), v = F (VI (A1 ), . . . , VI (An ))).
/ BI ′ .
Note that since I ′ must be an interpretation, VI ′ (A) = default (A) for all A ∈
The set of interpretations forms a complete lattice, so that TP it is continuous.
Recall (from [9]) the definition of the ordinal powers of a function G over a
complete lattice X:
S
{G ↑ α′ | α′ < α}
if α is a limit ordinal,
G↑α=
G(G ↑ (α − 1))
if α is a successor ordinal,
and dually,
T
{G ↓ α′ | α′ < α}
G↓α=
G(G ↓ (α − 1))
if α is a limit ordinal,
if α is a successor ordinal,
Since the first limit ordinal is 0, it follows that in particular, G ↑ 0 = ⊥X
(the bottom element of the lattice X) and G ↓ 0 = ⊤X (the top element). From
Kleene’s fixed point theorem we know that the least fixed-point of any continuous
operator is reached at the first infinite ordinal ω. Hence lf p(TP ) = TP ↑ ω.
Lemma 31 Let P a fuzzy program. Then M is a model of P if and only if M
is a pre-fixpoint of TP , that is TP (M ) ⊑ M .
Proof. Let M = hBM , VM i and TP (M ) = hBTP , VTP i.
We first prove the “only if” (→) direction. Let A be an element of Herbrand
Base, if A ∈ BTP , then by definition of TP there exists a ground instance of a
fact of P , A ← v, or a ground instance of a clause of P , A ←F A1 , . . . , An where
{A1 , . . . , An } ⊆ BM and v = F (VM (A1 ), . . . , VM (An )). Since M is a model
of P , A ∈ BM , and each v ⊆BI VM (A), then VTP (A) ⊆BI VM (A) and then
TP (M ) ⊑ M . . If A ∈
/ BTP then VTP (A) = default (A) ⊆BI VM (A).
Analogously, for the “if” (←) direction, for each ground instance
v = F (VM (A1 ), . . . , VM (An )), A ∈ BTP and v ⊆BI VTP (A), but as TP (M ) ⊑ M ,
BTP ⊆ BM and VTP (A) ⊆BI VM (A). Then A ∈ BM and v ⊆BI VM (A) therefore
M is a model of P .
Given this relationship, it is straightforward to prove that the least model of
a program P is also the least fixed-point of TP .
Theorem 32 Let P be a fuzzy program. Then lm(P ) = lfp(TP ).
Proof.
T
lm(P ) = T{M | M is a model of P }
= {M | M is a pre-fixpoint of P } from lemma 31
= lfp(TP ) by the Knaster-Tarski Fixpoint Theorem [27]
3.3
Operational Semantics
The improvement of Fuzzy Prolog is remarkable in its new procedural semantics
that is interpreted as a sequence of transitions between different states of a
system. We represent the state of a transition system in a computation as a tuple
hA, σ, Si where A is the goal, σ is a substitution representing the instantiation
of variables needed to get to this state from the initial one and S is a constraint
that represents the truth value of the goal at this state.
When computation starts, A is the initial goal, σ = ∅ and S is true (if there
are neither previous instantiations nor initial constraints). When we get to a
state where the first argument is empty then we have finished the computation
and the other two arguments represent the answer.
Definition 38 (Transition) A transition in the transition system is defined
as:
1. hA ∪ a, σ, Si → hAθ, σ · θ, S ∧ µa = vi
if h ← v is a fact of the program P , θ is the mgu of a and h, µa is the truth
value for a and solvable(S ∧ µa = v).
2. hA ∪ a, σ, Si → h(A ∪ B)θ, σ · θ, S ∧ ci
if h ←F B is a rule of the program P , θ is the mgu of a and h, c is
the constraint that represents the truth value obtained applying the unionaggregation F to the truth values of B, and solvable(S ∧ c).
3. hA ∪ a, σ, Si → hA, σ, S ∧ µa = vi
if none of the above are applicable and solvable(S ∧ µa = v) where µa =
default (a).
Definition 39 (Success set) The success set SS(P ) collects the answers to
simple goals p(b
x). It is defined as follows:
SS(P ) = hB, V i
where B = {p(b
x)σ|hp(b
x), ∅, truei →∗ h∅, σ, Si} is the set of elements of the
Herbrand Base that are instantiated and that have succeeded; and V (p(b
x)) =
∪{v|hp(b
x), ∅, truei →∗ h∅, σ, Si, and v is the solution of S} is the set of truth values of the elements of B that is the union (got by backtracking) of truth values
that are obtained from the set of constraints provided by the program P while
query p(b
x) is computed.
In order to prove the equivalence between operational semantic and fixedpoint semantic, it is useful to introduce a type of canonical top-down evaluation
strategy. In this strategy all literals are reduced at each step in a derivation. For
obvious reasons, such a derivation is called breadth-first.
Definition 310 (Breadth-first transition) Given the following set of valid
transitions:
h{{A1 , . . . , An }, σ, Si →
h{{A1 , . . . , An }, σ, Si →
..
.
h{{A2 , . . . , An } ∪ B1 , σ · θ1 , S ∧ c1 i
h{{A1 , A3 . . . , An } ∪ B2 , σ · θ2 , S ∧ c2 i
h{{A1 , . . . , An }, σ, Si →
h{{A1 , . . . , An−1 } ∪ Bn , σ · θn , S ∧ cn i
a breadth-first transition is defined as
h{A1 , . . . , An }, σ, Si →BF
hB1 ∪ . . . ∪ Bn , σ · θ1 · . . . · θn , S ∧ c1 ∧ . . . ∧ cn i
in which all literals are reduced at one step.
Theorem 33 Given a ordinal number n and TP ↑ n = hBTPn , VTPn i. There is a
successful breadth-first derivation of lengh less or equal to n + 1 for a program P ,
h{A1 , . . . , Ak }, σ, S1 i →∗BF h∅, θ, S2 i iff Ai θ ∈ BTPn and solvable(S ∧ µAi = vi )
and vi ⊆BI VTPn (Ai θ).
Proof. The proof is by induction on n. For the base case, all the literals are
reduced using the first type of transitions or the last one, that is, for each literal
Ai , it exits a fact hi ← vi such that θi is the mgu of Ai and hi , and µAi is
the truth variable for Ai , and solvable(S1 ∧ µAi = vi ) or µAi = default (Ai ). By
definition of TP , each vi ⊆BI VTP1 (Ai θ) where hBTP1 , VTP1 i = TP ↑ 1.
For the general case, consider the successful derivation,
h{A1 , . . . , Ak }, σ1 , S1 i →BF hB, σ2 , S2 i →BF . . . →BF h∅, σn , Sn i
the transition h{A1 , . . . , Ak }, σ1 , S1 i →BF hB, σ2 , S2 i
When a literal Ai is reduced using a fact or there is not rule for Ai , the result
is the same as in the base case. Otherwise there is a clause hi ←F B1i , . . . , Bmi
in P such that θi is the mgu of Ai and hi ∈ Bσ2 and Bji θi ∈ Bσ2 , by the
induction hypothesis Bσ2 ⊆ BTPn−1 and solvable(S2 ∧ µBji = vji ) and vji ⊆BI
VTPn−1 (Bji σ2 ) then Bji θi ⊆ BTPn−1 and by definition of TP , Ai θi ∈ BTPn and
solvable(S1 ∧ µAi = vi ) and vi =⊆BI VTPn (Ai σ1 ).
Theorem 34 For a program P there is a successful derivation
hp(b
x), ∅, truei →∗ h∅, σ, Si
iff p(b
x)σ ∈ B and v is the solution of S and v ⊆BI V (p(b
x)σ) where lf p(TP ) =
hB, V i
Proof. It follows from the fact that lf p(TP ) = TP ↑ ω and from the Theorem
33.
Theorem 35 For a fuzzy program P the three semantics are equivalent, i.e.
SS(P ) = lf p(T P ) = lm(P )
Proof. the first equivalence follows from Theorem 34 and the second from Theorem 32.
4
Implementation and Syntax
4.1
CLP(R)
Constraint Logic Programming [11] began as a natural merging of two declarative paradigms: constraint solving and logic programming. This combination
helps make CLP programs both expressive and flexible, and in some cases, more
efficient than other kinds of logic programs. CLP(R) [12] has linear arithmetic
constraints and computes over the real numbers.
Fuzzy Prolog was implemented in [9] as a syntactic extension of a CLP(R)
system. CLP(R) was incorporated as a library in the Ciao Prolog system1 .
The fuzzy library (or package in the Ciao Prolog terminology) which implements the interpreter of our Fuzzy Prolog language has been modified to handle
default reasoning.
1
The Ciao system [1] including our Fuzzy Prolog implementation can be downloaded
from http://www.clip.dia.fi.upm.es/Software/Ciao.
4.2
Syntax
Let us recall, from [9], the syntax of Fuzzy Prolog. Each Fuzzy Prolog clause
has an additional argument in the head which represents its truth value in terms
of the truth values of the subgoals of the body of the clause. A fact A ← v
is represented by a Fuzzy Prolog fact that describes the range of values of v
with a union of intervals (which can be only an interval or even a real number
in particular cases). The following examples illustrate the concrete syntax of
programs:
S
youth(45) ← [0.2, 0.5]
[0.8, 1]
youth(45):∼ [0.2,0.5]v[0.8,1]
tall (john) ← 0.7
tall(john):∼ 0.7.
swift (john) ← [0.6, 0.8]
swift(john):∼ [0.6,0.8]
good player (X) ←min tall (X),
good player(X):∼min tall(X),
swift (X)
swift(X)
These clauses are expanded at compilation time to constrained clauses that
are managed by CLP(R) at run-time. Predicates . = ./2, . < ./2, . <= ./2, . > ./2
and . >= ./2 are the Ciao CLP(R) operators for representing constraint inequalities, we will use them in the code of predicates definitions (while we will use the
common operators =, <, ≤, >, ≥ for theoretical definitions). For example the
first fuzzy fact is expanded to these Prolog clauses with constraints
youth(45,V):youth(45,V):-
V .>=. 0.2, V .<=. 0.5.
V .>=. 0.8, V .<. 1.
And the fuzzy clause
good player(X) :∼ min tall(X), swift(X).
is expanded to
good_player(X,Vp) :- tall(X,Vq), swift(X,Vr), minim([Vq,Vr],Vp),
Vp .>=. 0, Vp .=<. 1.
The predicate minim/2 is included as run-time code by the library. Its function is adding constraints to the truth value variables in order to implement the
T-norm min.
minim([],_).
minim([X],X).
minim([X,Y|Rest],Min):- min(X,Y,M), minim([M|Rest],Min).
min(X,Y,Z):- X .=<. Y , Z .=. X.
min(X,Y,Z):- X .>. Y, Z .=. Y .
We have implemented several aggregation operators as prod, max, luka (Lukasievicz
operator), etc. and in a similar way any other operator can be added to the system without any effort. The system is extensible by the user simply adding the
code for new aggregation operators to the library.
5
5.1
Combining Crisp and Fuzzy Logic
Example: Teenager Student
In order to use definitions of fuzzy predicates that include crisp subgoals we
must define properly their semantics with respect to the Prolog Close World
Assumption (CWA) [3]. We are going to present a motivating example from [9].
Fuzzy clauses usually use crisp predicate calls as requirements that data have
to satisfy to verify the definition in a level superior to 0, i.e. crisp predicates are
ussually tests that data should satisfy in the body of fuzzy clauses. For example,
if we can say that a teenager student is a student whose age is about 15 then we
can define the fuzzy predicate teenager student/2 in Fuzzy Prolog as
teenager_student(X,V):~ student(X), age_about_15(X,V2).
Note that we can face the risk of unsoundness unless the semantics of crisp
and fuzzy predicates is properly defined. CWA means that all non-explicit information is false. E.g., if we have the predicate definition of student/1 as
student(john).
student(peter).
then we have that the goal student(X) succeeds with X = john or with X =
peter but fails with any other value different from these; i.e:
?- student(john).
yes
?- student(nick).
no
which means that john is a student and nick is not. This is the semantics
of Prolog and it is the one we are going to adopt for crisp predicates because
we want our system to be compatible with conventional Prolog reasoning. But
what about fuzzy predicates? According to human reasoning we should assume
OWA (non explicit information in unknown). Consider the following definition
of age about 15/2
age_about_15(john,1):~ .
age_about_15(susan,0.7):~ .
age_about_15(nick,0):~ .
The goal age about 15(X, V ) succeeds with X = john and V = 1 or with X =
susan and V = 0.7. Therefore we do not know if the age of peter is about 15 or
not; and we know that nick’s age is definitely not about 15.
Our way to introduce crisp subgoals into the body of fuzzy clauses is translating the crisp predicate into the respective fuzzy predicate. For our example
we obtain the following Prolog definition.
teenager_student(X,V):~ f_student(X,V1), age_about_15(X,V2).
Where the default truth value of a crisp predicate is 0.
f_student(X,1):- student(X).
:-default(f_student/2,0).
Nevertheless, we consider for age about 15/2 and teenager student/2 that the
default value is unknown (the whole interval [0, 1]).
:-default(age_about_15/2,[0,1]).
:-default(teenager_student/2,[0,1]).
Observe the following consults:
?- age_about_15(john,X).
X = 1
?- age_about_15(nick,X).
X = 0
?- age_about_15(peter,X).
X .>=. 0, X .<=. 1
This means john’s age is about 15, nick’s age is not about 15 and we have
no data about peter’s age.
We expect the same behavior with the fuzzy predicate teenager student/2,
ie:
?- teenager_student(john,V).
V .=. 1
?- teenager_student(susan,V).
V .=. 0
?- teenager_student(peter,V).
V .>=. 0, V .<=. 1
as john is a “teenager student” (he is a student and his age is about 15),
susan is not a “teenager student” (she is not a student) and we do not know
the value of maturity of peter as student because although he is a student, we
do not know if his age is about 15.
5.2
Example: Timetable Compatibility
Another real example could be the problem of compatibility of a couple of shifts
in a work place. For example teachers that work in different class timetables,
telephone operators, etc. Imagine a company where the work is divided in shifts
of 4 hours per week. Many workers have to combine a couple of shifts in the
same week and a predicate compatible/2 is necessary to check if two shifts are
Day
Hour M
8
T
W
T
F
Day
Hour M
8
T
W
T
F
Day
Hour
M
8
T
W
T
F
Day
Hour
M
8
9
9
9
9
10
10
10
10
11
11
11
11
12
12
12
12
13
13
13
13
14
14
14
14
15
15
15
15
16
16
16
16
17
17
17
17
18
18
18
18
T
W
T
F
Fig. 1. Timetable 1, 2, 3 and 4
compatible or to obtain which couples of shifts are compatible. Two shifts are
compatible when both are correct (working days from Monday to Friday, hours
between 8 a.m. and 18 p.m. and there are no repetitions of the same hour in a
shift) and in addition when the shifts are disjoint.
compatible(T1,T2):- correct_shift(T1), correct_shift(T2),
disjoint(T1,T2).
But there are so many compatible combinations of shifts that it would be
useful to define the concept of compatibility in a fuzzy way instead of in the crisp
way it is defined above. It would express that two shifts could be incompatible if
one of them is not correct or if they are not disjoint but when they are compatible,
they can be more or less compatible. They can have a level of compatibility.
Two shifts will be more compatible if the working hours are concentrated (the
employee has to go to work few days during the week). Also, two shifts will
be more compatible if there are few free hours between the busy hours of the
working days of the timetable.
Therefore, we are handling crisp concepts (correct shif t/1, disjoint/2) besides fuzzy concepts (without gaps/2, f ew days/2). Their definitions, represented in figure 2, are expressed in our language in this simple way (using the
operator “: #” for function definitions and the reserved word “fuzzy predicate”):
few_days :#
fuzzy_predicate([(0,1),(1,0.8),(2,0.6),(3,0.4),(4,0.2),(5,0)]).
without_gaps :#
fuzzy_predicate([(0,1),(1,0.8),(5,0.3),(7,0.1),(8,0)]).
A simple implementation in Fuzzy Prolog combining both types of predicates
could be:
few_days
without_gaps
1
1
0
0
1
2
3
4
5
days
0
1
2
3
4
5
6
7
8
hours
Fig. 2. Fuzzy predicates few days/2 and without gaps/2
compatible(T1,T2,V):~ min
f_correct_shift(T1,V1),
f_correct_shift(T2,V2),
f_disjoint(T1,T2,V3),
f_append(T1,T2,T,V4),
f_number_of_days(T,D,V5),
few_days(D,V6),
f_number_of_free_hours(T,H,V7),
without_gaps(H,V8).
Here append/3 gives the total weekly timetable of 8 hours from joining two
shifts, number of days/3 obtains the total number of working days of a weekly
timetable and number of f ree hours/2 returns the number of free one-hour
gaps that the weekly timetable has during the working days. The f_predicates
are the corresponding fuzzified crisp predicates. The aggregation operator min
will aggregate the value of V from V 6 and V 8 checking that V 1, V 2, V 3, V 4,
V 5 and V 7 are equal to 1, otherwise it fails. Observe the timetables in figure
1. We can obtain the compatibility between the couple of shifts, T1 and T2,
represented in each timetable asking the subgoal compatible(T 1, T 2, V ). The
result is V = 0.2 for the timetable 1, V = 0.6 for the timetable 2, and V = 0 for
the timetable 3 (because the shifts are incompatible).
Regarding compatibility of shifts in a weekly timetable, we are going to ask
some questions about the shifts T1 and T2 of timetable 4 of figure 1. One hour
of T2 is not fixed yet.
We can note: the days of the week as mo, tu, we, th and f r; the slice of time
of one hour as the time of its beginning from 8 a.m. till 17 p.m.; one hour of the
week timetable as a pair of day and hour and one shift as a list of 4 hours of the
week.
If we want to know how to complete the shift T2 given a level of compatibility
higher than 70 %, we obtain the slice from 10 to 11 p.m. at Wednesday or Monday
morning.
?- compatible([(mo,9), (tu,10), (we,8), (we,9)],
[(mo,8), (we,11), (we,12), (D,H)], V),
V .>. 0.7 .
V = 0.9, D = we, H = 10 ? ;
V = 0.75, D = mo, H = 10 ? ;
no
6
Conclusions and Future work
Extending the expressivity of programming systems is very important for knowledge representation. We have chosen a practical and extended language for
knowledge representation: Prolog.
Fuzzy Prolog presented in [9] is implemented over Prolog instead of implementing a new resolution system. This gives it a good potential for efficiency,
more simplicity and flexibility. For example aggregation operators can be added
with almost no effort. This extension to Prolog is realized by interpreting fuzzy
reasoning as a set of constraints [36], and after that, translating fuzzy predicates
into CLP(R) clauses. The rest of the computation is resolved by the compiler.
In this paper we propose to enrich Prolog with more expressivity by adding
default reasoning and therefore the possibility of handling incomplete information that is one of the most worrying characteristics of data (i.e. all information
that we need usually is not available but only one part of the information is
available) and anyway searches, calculations, etc. should be done just with the
information that we had.
We have developed a complete and sound semantics for handling incomplete
fuzzy information and we have also provided a real implementation based in our
former Fuzzy Prolog approach.
We have managed to combine crisp information (CWA) and fuzzy information (OWA or default) in the same program. This is a great advantage because
it lets us model many problems using fuzzy programs. So we have extended
the expressivity of the language and the possibility of applying it to solve real
problems in which the information can be defined, fuzzy or incomplete.
Presently we are working in several related issues:
– Obtaining constructive answers to negative goals.
– Constructing the syntax to work with discrete fuzzy sets and its applications
(recently published in [21]).
– Implementing a representation model using unions instead of using backtracking.
– Introducing domains of fuzzy sets using types. This seems to be an easy task
considering that we are using a modern Prolog [10] where types are available.
– Implementing the expansion over other systems. We are studing now the
advantages of an implementation in XSB system where tabling is used.
– Using our approach for the engine of robots in a RoboCup league in a joint
project between our universities.
References
1. D. Cabeza and M. Hermenegildo. A New Module System for Prolog. In CL2000,
number 1861 in LNAI, pages 131–148. Springer-Verlag, July 2000.
2. T.H. Cao. Annotated fuzzy logic programs. Fuzzy Sets and Systems, 113(2):277–
298, 2000.
3. K. L. Clark. Negation as failure. In H. Gallaire and J. Minker, editors, Logic and
Data Bases, pages 293–322, New York, NY, 1978. Plenum Press.
4. D. Dubois, J. Lang, and H. Prade. Towards possibilistic logic programming. In
Proc. of ICLP-91, pages 581–595. MIT Press, 1991.
5. M. Fitting. Bilattices and the semantics of logic programming. Journal of Logic
Programmig, 11:91–116, 1991.
6. N. Fuhr. Probabilistic datalog: Implementing logical information retrieval for advanced applications. Journal of the American Society for Information Science,
51(2):95–110, 2000.
7. M. Gelfond and V. Lifschitz. The stable model semantics for logic programming.
In Fifth International Conference and Symposium on Logic Programming, pages
1070–1080, 1988.
8. M. Gelfond and V. Lifschitz. Logic programs with classical negation. In Seventh
International Conference on Logic Programming, pages 579–597, Jerusalem, Israel, 1990. MIT Press. Extended abstract. (Complete version in ”New Generation
Computing” 9:365-387,1991).
9. S. Guadarrama, S. Munoz-Hernandez, and C. Vaucheret. Fuzzy Prolog: A new
approach using soft constraints propagation. Fuzzy Sets and Systems, FSS,
144(1):127–150, 2004. ISSN 0165-0114.
10. M. Hermenegildo, F. Bueno, D. Cabeza, M. Carro, M. Garcı́a de la Banda,
P. López-Garcı́a, and G. Puebla. The CIAO Multi-Dialect Compiler and System:
An Experimentation Workbench for Future (C)LP Systems. In Parallelism and
Implementation of Logic and Constraint Logic Programming, pages 65–85. Nova
Science, Commack, NY, USA, April 1999.
11. J. Jaffar and J. L. Lassez. Constraint Logic Programming. In ACM Symp. Principles of Programming Languages, pages 111–119. ACM, 1987.
12. J. Jaffar, S. Michaylov, P. J. Stuckey, and R. H. C. Yap. The clp(r) language and
system. ACM Transactions on Programming Languages and Systems, 14(3):339–
395, 1992.
13. M. Kifer and Ai Li. On the semantics of rule-based expert systems with uncertainty.
In Proc. of ICDT-88, number 326 in LNCS, pages 102–117, 1988.
14. M. Kifer and V.S. Subrahmanian. Theory of generalized annotated logic programming and its applications. Journal of Logic Programming, 12:335–367, 1992.
15. E.P. Klement, R. Mesiar, and E. Pap. Triangular norms. Kluwer Academic Publishers.
16. L. Lakshmanan. An epistemic foundation for logic programming with uncertainty.
LNCS, 880:89–100, 1994.
17. L. Lakshmanan and N. Shiri. Probabilistic deductive databases. Int. Logic Programming Symposium, pages 254–268, 1994.
18. L. Lakshmanan and N. Shiri. A parametric approach to deductive databases with
uncertainty. IEEE Transactions on Knowledge and Data Engineering, 13(4):554–
570, 2001.
19. Y. Loyer and U. Straccia. Uncertainty and partial non-uniform assumptions in
parametric deductive databases. In Proc. of JELIA-02, volume 2424 of LNCS,
pages 271–282, 2002.
20. T. Lukasiewicz. Fixpoint characterizations for many-valued disjunctive logic programs with probabilistic semantics. In Proc. of LPNMR-01, volume 2173, pages
336–350, 2001.
21. S. Munoz-Hernandez and J.M. Gómez Pérez. Solving collaborative fuzzy agents
problems with clp(fd). In Manuel Hermenegildo and Daniel Cabeza, editors, International Symposium on Practical Aspects of Declarative Languages, PADL 2005,
volume 3350 of LNCS, pages 187–202, Long Beach, CA (USA), 2005. SpringerVerlag.
22. R. Ng and V.S. Subrahmanian. Stable model semantics for probabilistic deductive
databases. In Proc. of ISMIS-91, number 542 in LNCS, pages 163–171, 1991.
23. R. Ng and V.S. Subrahmanian. Probabilistic logic programming. Information and
Computation, 101(2):150–201, 1993.
24. H. T. Nguyen and E. A. Walker. A first Course in Fuzzy Logic. Chapman &
Hall/Crc, 2000.
25. A. Pradera, E. Trillas, and T. Calvo. A general class of triangular norm-based
aggregation operators: quasi-linear t-s operators. International Journal of Approximate Reasoning, 30(1):57–72, 2002.
26. V.S. Subrahmanian. On the semantics of quantitative logic programs. In Proc. of
4th IEEE Symp. on Logic Programming, pages 173–182. Computer Society Press,
1987.
27. A. Tarski. A lattice-theoretical fixpoint theorem and its applications. Pacific
Journal of Mathematics, 5:285–309, 1955.
28. E. Trillas, S. Cubillo, and J. L. Castro. Conjunction and disjunction on ([0, 1], <=).
Fuzzy Sets and Systems, 72:155–165, 1995.
29. E. Trillas, A. Pradera, and S. Cubillo. A mathematical model for fuzzy connectives
and its application to operators behavioural study. In B. Bouchon-Meunier, R.R.
Yager, and L.A. Zadeh, editors, Information, Uncertainty and Fusion, volume 516,
pages 307–318. Kluwer Academic Publishers (Series: The Kluwer International
Series in Engineering and Computer Sciences), 1999.
30. M.H. van Emden. Quantitative duduction and its fixpoint theory. Journal of Logic
Programming, 4(1):37–53, 1986.
31. C. Vaucheret, S. Guadarrama, and S. Munoz-Hernandez. Fuzzy prolog: A simple
implementation using clp(r). In Constraints and Uncertainty, Paphos (Cyprus),
2001. CP’2001 workshop.
http://www.clip.dia.fi.upm.es/clip/papers/fuzzylpar02.ps.
32. C. Vaucheret, S. Guadarrama, and S. Munoz-Hernandez. Fuzzy prolog: A simple
general implementation using clp(r). In M. Baaz and A. Voronkov, editors, Logic
for Programming, Artificial Intelligence, and Reasoning, LPAR 2002, number 2514
in LNAI, pages 450–463, Tbilisi, Georgia, October 2002. Springer-Verlag.
33. G. Wagner. A logical reconstruction of fuzzy inference in databases and logic
programs. In Proc. of IFSA-97, Prague, 1997.
34. G. Wagner. Negation in fuzzy and possibilistic logic programs. In Logic programming and Soft Computing. Research Studies Press, 1998.
35. Ehud Y. and Shapiro. Logic programs with uncertainties: A tool for implementing
rule-based systems. In IJCAI, pages 529–532, 1983.
36. L. Zadeh. Fuzzy sets as a basis for a theory of possibility. Fuzzy Sets and Systems,
1(1):3–28, 1978.
| 6cs.PL
|
arXiv:1504.01066v3 [math.AC] 20 Oct 2017
ON THE GROWTH OF DEVIATIONS
ADAM BOOCHER, ALESSIO D’ALÌ, ELOÍSA GRIFO, JONATHAN MONTAÑO,
AND ALESSIO SAMMARTANO
Abstract. The deviations of a graded algebra are a sequence of integers that
determine the Poincaré series of its residue field and arise as the number of
generators of certain DG algebras. In a sense, deviations measure how far a
ring is from being a complete intersection. In this paper we study extremal
deviations among those of algebras with a fixed Hilbert series. In this setting,
we prove that, like the Betti numbers, deviations do not decrease when passing
to an initial ideal and are maximized by the Lex-segment ideal. We also prove
that deviations grow exponentially for Golod rings and for certain quadratic
monomial algebras.
1. Introduction
Let R = S/I, where S = k[T1 , . . . , Tn ] is a polynomial ring over a field k and
I ⊆ (T1 , . . . , Tn )2 is a homogeneous ideal. With the pair (S, R) we can associate
two important sets of Betti numbers, namely
βiS (R) = dimk TorSi (R, k)
and
βiR (k) = dimk TorR
i (k, k).
Equivalently, these numbers are given by the ranks of the free modules appearing in
minimal free resolutions of R over S and k over R respectively. To avoid confusion,
and to emphasize that the latter sequence
typically infinite, we will often refer
P is R
i
instead to the Poincaré series PkR (z) = ∞
i=1 βi (k)z . Much work has been devoted
to studying the growth of Betti numbers for various classes of rings. For instance,
by the Auslander-Buchsbaum-Serre Theorem, PkR (z) is a polynomial if and only
if R is regular, i.e., a polynomial ring. In the next simplest case, the sequence
{βiR (k)} has polynomial growth if R is a complete intersection (cf. [33]). Finally, if
R is not a complete intersection, then the βiR (k) have exponential growth (cf. [4]).
In this paper we study Betti numbers from the point of view of a related invariant:
the set of deviations of R. Since PkR (z) has integer coefficients and constant term
equal to 1, there exist uniquely determined integers εi = εi (R), called the deviations
of R, such that the following infinite product expansion holds (cf. [5, 7.1.1]):
(1.1)
PkR (z) =
∞
Y
(1 + z 2i−1 )ε2i−1
(1 − z 2i )ε2i
i=1
.
In the sense explained in Remark 2.1, the deviations measure the complexity
of R. They play a crucial role in Avramov’s proof that the complete intersection
property localizes (cf. [3]). In addition, the εi (R) appear naturally as the number of
2010 Mathematics Subject Classification. Primary: 13D02; Secondary: 16E45, 13D40, 16S37,
05C25, 05C38.
Key words and phrases. Deviations, DG algebra resolution, Poincaré series, Hilbert series,
Betti numbers, Koszul algebra, Golod ring, initial ideal, lex-segment ideal, edge ideal.
1
generators of degree i in an acyclic closure of k over R, as the number of generators
of degree i − 1 in a minimal model of R over S, and as the ranks of the components
of the homotopy Lie algebra of R, cf. Section 2.
Our first results concern the deviations of algebras with fixed Hilbert function.
In Theorem 3.1 we prove that, for any term order <, the algebra presented by
in< I has larger deviations than R. Moreover, we show in Theorem 3.4 that the
algebra presented by the lex-segment ideal has the largest deviations. This is a
generalization of a result of Peeva which states that the lex-segment ideal attains
the largest values of PkR (z) among all I with the same Hilbert function (cf. [27]).
Peeva’s theorem in turn relies on its analogue for βiS (R), which is due to Bigatti,
Hulett, and Pardue (cf. [10, 24, 26]). Our methods use their techniques to provide
a proof in all characteristics. In characteristic zero, we also present a simpler proof
using Golod rings. Golod rings are those rings R whose Poincaré series is maximal
among all ideals with a fixed embedding dimension and set of Betti numbers βiS (R);
see [25, 6.16] for a list of Golod rings. In particular, if R is a Golod ring, then PkR (z)
is a rational function of z determined by the βiS (R) and edim R.
In the second half of the paper we turn to an analysis of the asymptotic behavior
of deviations. A result of Babenko shows that the radius of convergence of the
generating function of the εi (R) coincides with that of PkR (z), in the non-trivial
case when εi (R) 6= 0 infinitely often (cf. [7, Theorem 1]). Furthermore, there exist
subsequences of {εi (R)} which are not too sparse and have an exponential lower
bound, provided that R is not a complete intersection (cf. [4, 15]). In Theorem 4.3
and Theorem 4.4 we prove that, for Golod rings and for certain Koszul algebras, the
i
sequence of deviations is asymptotically equal to { bρi }i>1 for some ρ > 1, b ∈ N.
2. Preliminaries
We now introduce the notions of acyclic closures and minimal models, and how
they encode the deviations. We set up some notation and generally follow [5]
throughout the paper. Although the results therein are proved for local rings, they
can be extended to the graded case with minor changes in the proofs (cf. [5, 8.3.7],
see also [6]).
Let A = ⊕i>0 Ai be a DG algebra such that A0 is a Noetherian standard graded
algebra over a field. Assume that A is bi-graded with internal and homological
degree; we will denote homological degree by | · | and internal degree by deg(·).
For the differential we have |∂| = −1 and deg(∂) = 0. It follows from the Leibniz rule for ∂ that the cycles form a bi-graded subalgebra Z(A) and the boundaries form a bi-graded two-sided ideal B(A) ⊆ Z(A), and therefore the homology
H(A) = Z(A)/B(A) also has a bi-graded algebra structure. If ζ ∈ Zi (A)j is a
bi-homogeneous cycle, let y be a new variable of bi-degree (i + 1, j) and denote
by Ahyi the unique, up to isomorphism, DG algebra extension of A such that the
differential satisfies ∂(y) = ζ, where y is an exterior variable if i is even and a
divided power variable if i is odd (cf. [5, 6.1.1]). For any proper bi-homogeneous
ideal J ⊆ A0 we can construct
a semi-free DG algebra resolution AhY i of the factor
S
ring A0 /J, where Y = i>1 Yi and Yi is a finite set of bi-homogeneous variables of
homological degree i such that ∂(Y1 ) minimally generates J modulo ∂(A1 ) and the
homology classes of each
S ∂(Yn+1 ) minimally generate the A0 -module Hn (AhY6n i).
Here Y6n is short for i6n Yi . A DG algebra AhY i obtained in this way is called
2
an acyclic closure of A0 /J over the DG algebra A and the variables Y are called
Γ-variables. Notice that the i-th homology of AhY6n i vanishes for 0 < i < n.
A theorem proved independently by Gulliksen and Schoeller (cf. [20], [31]) states
that an acyclic closure RhY i of k over the DG algebra A = A0 = R is in fact a
minimal resolution of k over R, from which one can easily deduce that the i-th
deviation is equal to the number of Γ-variables in homological degree i:
εi (R) = Card(Yi ).
It follows in particular that εi (R) > 0. Note that RhY1 i is the Koszul complex K R
of R with respect to m. Basic homological properties of R can be characterized
in terms of the vanishing of deviations, as illustrated in the following discussion,
which explains the use of the word “deviation” (cf. [5, 7.3]).
Remark 2.1. The first deviation ε1 (R) is equal to the embedding dimension of
R, thus the following conditions are equivalent: (i) R is a field, (ii) ε1 (R) = 0, (iii)
εi (R) = 0 for every i > 1. The exactness of the Koszul complex K R gives the
following equivalent conditions: (i) R is regular, (ii) ε2 (R) = 0, (iii) εi (R) = 0 for
every i > 2. Finally, theorems of Assmus [2], Halperin [21] and Tate [33] give the
following equivalent conditions: (i) R is a complete intersection, (ii) ε3 (R) = 0, (iii)
εi (R) = 0 for every i > 3, (iv) εi (R) = 0 for some i > 3.
Following the notation in the construction of acyclic closures, denote by p the
homogeneous maximal ideal of A0 . Given J ⊆ p2 a homogeneous ideal and starting
from A0 , we can build another bi-graded DG algebra free resolution A0 [X] of A0 /J
over A0 by following the same steps above, but adjoining polynomial variables
(instead of divided powers) to kill cycles in odd degrees. A DG algebra resolution
obtained in this way is called a minimal model of A0 /J over A0 . If we do not
require ∂(Xn+1 ) to generate Hn (A0 [X6n ]) minimally for each n > 1, then the
resolution obtained is simply called a model of A0 /J over A0 . Analogously to
minimal free resolutions, an equivalent
condition for a model to be minimal is that
Pn−1
∂(X1 ) ⊆ p2 and ∂(Xn+1 ) ⊆ pXn + i=1
Xi (A0 [X])n−i for n > 1 (cf. [5, 7.2.2]).
Minimal models always exist and are unique up to isomorphism (cf. [5, 7.2.4]).
If J is a complete intersection or if A0 is of equal characteristic zero, then A0 hY i
and A0 [X] are isomorphic as DG algebras. However they differ in general (cf. [5,
6.1.10]).
The following theorem due to Avramov shows that minimal models also carry
information about deviations [5, 7.2.5, 7.2.6]. We include here the statement in the
standard graded case:
Theorem 2.2 (Avramov). Let S = k[T1 , . . . , Tn ] and n = (T1 , . . . , Tn ). Let I be a
homogeneous ideal such that I ⊆ n2 and R = S/I. Let S[X] be a model of R over
S. Then Card(Xi ) > εi+1 (R) for every i > 1. Furthermore, equality occurs for
every i if and only if S[X] is a minimal model of R over S.
Hence, it is possible to compute εi (R) by computing a minimal model of R over
S. Some applications of this fact and a finer analysis of the sequence of deviations
for edge ideals of special graphs are given in [11].
Finally, a third context where deviations arise naturally is that of homotopy Lie
algebras. The homotopy Lie algebra of R is the graded Lie algebra over k
π(R) = H(DerγR (RhY i, RhY i))
3
where RhY i is an acyclic closure of k over R and DerγR (RhY i, RhY i) denotes the
DG module of R-linear Γ-derivations (cf.[5, 6.2.2]). Its universal enveloping algebra
is ExtR (k, k) and the dimension of the graded component π i (R) is εi (R). A result
due to Avramov and Löfwall states that R is Golod if and only if π >2 (R) is the
R
free graded Lie algebra generated by the vector space Homk (ΣH>1
, k), where H R
denotes the homology of the Koszul complex of R. We refer to [5, Chapter 10] for
more details on the subject.
3. Extremal deviations
The goal of this section is to study extremal deviations among ideals with a
given Hilbert series. The first main result of this section is Theorem 3.1, where
we show that the deviations of S/I are at most equal to those of S/in< (I) for any
term order <. The second main result is Theorem 3.4, where we show that among
all rings S/I with the same Hilbert series, the ring S/L has the largest deviations,
where L denotes the lex-segment ideal of S/I. We modify a standard deformation
argument and apply it to minimal models.
A non-negative integer function ω on the set {T1 , . . . , Tn } is called a weight for
S; we can extend it to arbitrary monomials in S by ω(T1v1 · · · Tnvn ) = v1 ω(T1 )+· · ·+
vn ω(Tn ). Given a polynomial f ∈ S, we denote the highest weight of a monomial
in the support of f by ω(f ) and the sum of the terms of f with weight equal to
ω(f ) by inω (f ). Let t be a new variable and define the ideal
Iω = tω(f ) f t−w(T1 ) T1 , . . . , t−w(Tn ) Tn | f ∈ I ⊆ S[t].
Notice that upon setting t = 0 in Iω we obtain inω I := (inω (f ) | f ∈ I) whereas
upon setting t = 1 we obtain I. If < is a term order, there is always a weight ω so
that inω I = in< I (cf. [22, 3.1.2]).
Theorem 3.1. Let S = k[T1 , . . . , Tn ], n = (T1 , . . . , Tn ), and ω be a weight for S. If
I ⊆ n2 is a homogeneous ideal, then we have εi (S/I) 6 εi (S/inω I) for every i > 1.
In particular, if < is a term order on S then εi (S/I) 6 εi (S/in< I) for every i > 1.
Proof. The result holds for i = 1 as ε1 (S/I) = n = ε1 (S/inω I). We extend ω to a
weight ω ′ for S[t] by setting ω ′ (Ti ) = ω(Ti ) and ω ′ (t) = 1; then Iω is a homogeneous
ideal of S[t] under the grading induced by ω ′ . Let S[t][X] be a minimal model of
S[t]/Iω over S[t]. Since t and t − 1 are regular on S[t]/Iω by [22, 3.2.5], it follows
from [22, 3.2.6] that S[t][X] ⊗S[t] S[t]/(t) and S[t][X] ⊗S[t] S[t]/(t − 1) are free Sresolutions of S/inω I and S/I respectively, and hence models over S, as they inherit
the DG algebra structure.
Since S is positively graded by the grading induced by ω, a direct adaptation of
the proof of [5, 7.2.2] allows us to conclude that
2
∂(X1 ) ⊆ (n, t) ,
∂(Xn+1 ) ⊆ (n, t)Xn +
n−1
X
Xi (S[t][X])n−i
if n > 1
i=1
hence the model S[t][X] ⊗S[t] S[t]/(t) is minimal as a model of S/inω I over S. By
Theorem 2.2, we conclude that for every i > 2
εi (S/inω I) = Card(Xi−1 ) > εi (S/I).
4
Serre showed that the following coefficientwise inequality of formal power series
holds:
(1 + z)n
Pn
PkR (z)
.
1 − i=1 βiS (R)z i+1
The ring R is called Golod if the equality is achieved; we refer the reader to [5,
Chapter 5] for a thorough treatment of Golod rings. If I, J ⊆ (T1 , . . . , Tn )2 are two
homogeneous ideals such that S/J is Golod and βiS (S/I) 6 βiS (S/J) for all i, the
S/I
S/J
inequality Pk (z) Pk (z) holds (cf. [27, proof of 1.3]). In the next proposition
we extend this result to the sequence of deviations.
Proposition 3.2. Let S = k[T1 , . . . , Tn ] and I, J ⊆ S be two homogeneous ideals.
If S/I and S/J are Golod rings such that βiS (S/I) 6 βiS (S/J) for all i, then
εi (S/I) 6 εi (S/J) for every i > 2.
Proof. Let R be a Golod ring, H R the homology of the Koszul complex of R,
and π(R) the homotopy Lie algebra of R. Then π >2 (R) is the free graded Lie
R
algebra generated by the vector space Homk (ΣH>1
, k), and εi (R) = dimk π i (R) for
S
R
every i (cf. Section 2). Since βi (R) = dimk Hi for every i, the desired inequality
follows.
Remark 3.3. If we remove the assumption that S/I and S/J are Golod rings then
the result is not true. Indeed, one can take I = (T12 , T1 T2 ) and J = (T12 , T22 ). While
these ideals share the same Betti numbers over S, ε3 (S/I) = 1 and ε3 (S/J) = 0.
Let < be a term order on S. It is well-known that there exists a non-empty
Zariski open subset U ⊆ GLn (k) such that the ideal in< (g · I) is the same for every
g ∈ U, where g · I denotes the image of I under the change of coordinates in S
defined by g (cf. [22, 4.1.2]). This ideal is called the generic initial ideal of I
with respect to < and is denoted by gin< I. An important property of gin< I is
that it is fixed by the action of the Borel subgroup of GLn (k); when char k = 0 this
means that gin< I is strongly stable (cf. [22, 4.2.1, 4.2.6]), thus S/gin< I is a Golod
ring by [23, Theorem 4].
Let HFI denote the Hilbert function of the ideal I. The lex-segment ideal of I
is the vector space spanned by the lexicographically first HFI (d) monomials of S in
each degree d. It was shown in [10], [24], [26] that for any ideal J with HFJ = HFL
and every i > 0 we have
(3.1)
βi (S/J) 6 βi (S/L).
Note that the ideal L is strongly stable, and therefore S/L is a Golod ring.
We present now the second main result of this section. We provide two proofs;
although the second one only works in characteristic zero, we present both in order
to show an application of Proposition 3.2.
Theorem 3.4. Let S = k[T1 , . . . , Tn ] and n = (T1 , . . . , Tn ). Let I ⊆ n2 be a
homogeneous ideal and L be the lex-segment ideal of I. Then, for every i > 1,
εi (S/I) 6 εi (S/L).
Proof 1. We follow the construction of Pardue in [26]; our approach is similar to
the one in the proof of the main result in [28]. Pardue’s proof shows that any
ideal in the Hilbert scheme is connected to the lex-segment ideal by a sequence of
deformations and degenerations. It suffices to show that at each step the deviations
do not decrease. The result is clear for i = 1, so let i > 2.
5
(1) Generic changes of coordinates. Since deviations depend only on the isomorphism class of R, generic changes of coordinates preserve deviations.
(2) Passing to an initial ideal. This follows by Theorem 3.1.
(3) Polarization and then factoring out generic hyperplane sections. It is well
known that an ideal I and its polarization are related via a regular sequence
of linear forms. Finally, the generic hyperplane sections that Pardue employs are always a regular sequence, so by [5, 7.1.6] the deviations do not
decrease from the second one on.
Proof 2 (char k = 0). We may assume that i > 1. Let g be a generic change of
coordinates. By Proposition 3.1, we have for all i
εi (S/I) = εi (S/(g · I)) 6 εi (S/in< (g · I)) = εi (S/gin< I).
Since S/(gin< I) and S/L are Golod rings and HFgin< I = HFI = HFL , the conclusion follows from (3.1) and Proposition 3.2.
S/I
Remark 3.5. We deduce from (1.1) that βi (k) can be written as a function
of the deviations εi (S/I) using only sums, products, and binomial coefficients, see
also [5, 7.1]. Thus, a pointwise inequality for the deviations implies a pointwise
inequality for the Betti numbers of k. In this way from Theorem 3.1 we recover the
well-known fact that the Poincaré series of the residue field can only grow larger
when passing to an initial ideal:
S/I
Pk
S/inω I
(z) Pk
(z)
whereas from Theorem 3.4 it follows that considering the lex-segment ideal gives
the largest Poincaré series among all ideals of S with the same Hilbert series, a fact
originally proved in [27]:
S/I
Pk
S/L
(z) Pk
(z).
However, the next example shows that a pointwise inequality for the Betti numbers
of k does not necessarily imply a pointwise inequality for the deviations.
Example 3.6. Let S = k[T1 , . . . , T5 ]. Denote by R6 and R8 the quotients of S by
6 and 8 generic quadrics, respectively; we determine their Poincaré series following
[30]. The two rings have the property L3 (cf. [30, pp. 480, 482]), and thus their
Poincaré series satisfy the equation
(3.2)
1 + 1z
HSR (−z)
1
−
=
HSA (z)
z
PkR (z)
where A is the subalgebra of the Yoneda algebra ExtR (k, k) generated by Ext1R (k, k).
A is the enveloping algebra of a graded Lie algebra η = ⊕i>1 η i generated in degree
1. Therefore, its Hilbert series is given by
(3.3)
HSA (z) =
∞
Y
(1 + z 2i−1 )e2i−1
(1 − z 2i )e2i
i=1
where ei = dimk η i (notice that these are not the deviations of R).
The Hilbert series of the two rings are HSR6 (z) = 1+5z+9z 2+5z 3 and HSR8 (z) =
1 + 5z + 7z 2 . In order to determine the ei , we proceed as in [30, p. 487]. We only
6
show the computations for the case of 6 quadrics, the other case being analogous.
The series HSA (z) is equal to the diagonal power series
X
R6
(3.4)
(k)z i = 1 + 5z + 16z 2 + 40z 3 + 86z 4 + 166z 5 + 296z 6 + · · · .
βii
i>0
Comparing (3.3) and (3.4) we find that e1 = 5, so we divide (3.4) by (1 + z)5
(3.5)
HSA (z)
= 1 + 6z 2 + 21z 4 + 56z 6 + · · ·
(1 + z)5
from which we find that e2 = 6. Now we multiply (3.5) by (1 − z 2 )6 and get
(3.6)
(1 − z 2 )6 HSA (z)
= 1 + 0z + 0z 2 + 0z 3 + · · · .
(1 + z)5
Hence η 3 = 0 and, since η is generated in degree 1, we conclude that η i = 0 for all
i > 3.
From (3.2) we obtain
PkR6 (z) =
1
,
(z − 1)(z 6 − 3z 5 + z 4 + 5z 3 − 5z 2 + 4z − 1)
and in the same way we obtain
PkR8 (z) =
−(z 2 − z + 1)5
,
d(z)
where d(z) = z 18 − 9z 17 + 37z 16 − 93z 15 + 160z 14 − 192z 13 + 136z 12 + 31z 11
− 270z 10 + 505z 9 − 664z 8 + 710z 7 − 636z 6 + 479z 5 − 294z 4 + 140z 3
− 47z 2 + 10z − 1.
Using these expressions we compute the first few deviations and find
ε4 (R6 ) = 16 > 9 = ε4 (R8 )
whereas we verify that βiR6 (k) 6 βiR8 (k) for i = 0, . . . , 9. Finally, using partial
fraction decompositions for PkR6 (z) and PkR8 (z) we obtain estimates for the Betti
numbers, showing that βiR6 (k) 6 βiR8 (k) for all i > 10, and thus
βiR6 (k) 6 βiR8 (k) for all i > 0.
4. Exponential Growth of Deviations
The goal of this section is to prove that, when R is not a complete intersection,
deviations grow exponentially if R is Golod (cf. Theorem 4.3) or if R is a Koszul
algebra with the same Hilbert series as a ring presented by the edge ideal of a
claw-free graph (cf. Theorem 4.4).
We now briefly introduce Koszul algebras, referring to [13] for a survey on the
topic. The ring R is a Koszul algebra if the minimal free R-resolution of k
R
is linear, i.e., βi,j
(k) = 0 whenever i 6= j. If R is Koszul then I is generated
by quadrics; however, these two conditions are not equivalent. For instance, if
R
I = (T12 , T22 , T32 , T42 , T1 T2 + T3 T4 ), then β3,4
(k) 6= 0. By a classical theorem of
Fröberg, R is Koszul if I is a quadratic
monomial
ideal (cf. [16]). Denote the
P
Hilbert series of R by HSR (z) = i dimk (Ri )z i ; then R is Koszul if and only if the
following identity occurs:
7
(4.1)
PkR (z) HSR (−z) = 1.
The following proposition provides a compact formula for deviations when the
Poincaré series has a certain form.
Proposition 4.1. Let S = k[T1 , . . . , Tn ] and R = S/I with I ⊆ n2 a homogeneous
ideal. Assume there exist c ∈ N and complex numbers {αj }16j6m so that PkR (z) =
c
Qm(1+z)
(1+αj z) . Then for every i > 2
j=1
εi (R) =
m
(−1)i X i X d
α
µ
i
d j=1 j
d|i
where µ is the Möbius function.
Proof. Set εi := εi (R). From (1.1) and the assumption we obtain
(4.2)
∞
Y
i+1
(1 − z)c
(1 − z i )(−1) εi = Qm
.
j=1 (1 − αj z)
i=1
Proceeding as in [29, p. 23], we apply natural logarithm at both sides of (4.2),
i
and
the coefficient
Pm of iz of the corresponding Maclaurin series to obtain
P compare
d
d|i (−1) dεd = −c +
j=1 αj . The result now follows by applying the Möbius
inversion
formula to the arithmetic function f (i) = (−1)i iεi and using the fact that
P
d|i µ(d) = 0 for every i > 2.
Note that the assumption of Proposition 4.1 is satisfied by several classes of rings,
including complete intersections (cf. [33]), monomial rings (cf. [8]), Golod rings
(cf. Section 3), and Koszul algebras, as it follows from (4.1). Given two numerical
sequences {ai }i>1 , {bi }i>1 , the expression ai ∼ bi stands for asymptotic equality,
i.e. limi→∞ abii = 1.
Example 4.2. Consider a Koszul algebra R with h-polynomial h(z) = 1 + mz for
some m ∈ N. This holds for instance for I = (T1 , . . . , Tm )2 ⊆ S = k[T1 , . . . , Tm ]
or I = (Ti Tj | 1 6 i < j 6 m + 1) ⊆ S = k[T1 , . . . , Tm+1 ] (the edge ideal of the
complete graph Km+1 ). By Proposition 4.1 for each i > 2 we have
mi
(−1)i X i
µ
.
(−m)d ∼
εi (R) =
i
d
i
d|i
We present next the first main result of this section. We show that deviations
grow exponentially for Golod rings that are not complete intersections. For similar
techniques applied to Betti numbers, see [32].
Theorem 4.3. Let S = k[T1 , . . . , Tn ] and n = (T1 , . . . , Tn ). Let I be a non-principal
homogeneous ideal such that I ⊆ n2 and R = S/I. If R is a Golod ring, then there
exists a real number ρ > 1 so that
εi (R) ∼
ρi
.
i
Proof. Let r be the radius of convergence of PkR (z). The ring R is not regular,
hence r 6 1 as the power series PkR (z) is infinite. Since R is Golod and not a
hypersurface, it is not a complete intersection (see [5, p. 47, Remark]), therefore
8
n
0 < r < 1 by [5, 4.2.3, 8.2.2]. Moreover, since R is Golod, we have PkR (z) = (1+z)
v(z)
Pn
inside BC (0, r), where v(z) = 1 − i=1 βiS (R)z i+1 .
Now, since PkR (z) is a rational function, it has singularities with absolute value
r, and since it has nonnegative real coefficients, we conclude that r is a singularity
of PkR (z). Hence, r is a root of v(z). We will show that r is a simple root of v(z)
and that it is the only root on the boundary of BC (0, r). The result will follow by
Proposition 4.1 as α1 = − 1r and |αj | < |α1 | otherwise.
Set βi := βiS (R). Since I is not
we have β1 , β2 > 0. Therefore r is a
Pprincipal,
n
simple root of v(z) as v ′ (r) = − i=1 (i + 1)βi ri < 0. If y 6= r and |y| = r, then
y 6∈ R>0 and hence
|β1 y 2 + β2 y 3 | = r2 |β1 + β2 y| < r2 (β1 + β2 r) = β1 r2 + β2 r3 .
Thus
n
X
βi y i+1 6 |β1 y 2 + β2 y 3 | +
n
X
βi y i+1 <
i=3
i=i
n
X
βi ri+1 = 1,
i=1
i.e. y is not a root of v(z). The proof is completed.
Let G be a graph with vertices {v1 , v2 , . . . , vn }. Denote by
I(G) = (Ti Tj | vi vj is an edge of G) ⊆ S
its edge ideal. By Fröberg’s theorem, an algebra R presented by an edge ideal is
Koszul.
A graph with four vertices and edges {v1 , v2 }, {v1 , v3 }, {v1 , v4 } is called a
claw (see Figure 1). A simple graph G is said to be claw-free if no claw appears
as an induced subgraph of G. Note that complete graphs are claw-free (cf. (4.2)).
① • ❋❋❋❋
①①
❋❋
①
①
❋
①①
•
•
•
Figure 1. A claw.
The following theorem shows that the asymptotic behavior of deviations observed
in Example 4.2 holds more generally.
Theorem 4.4. Let S = k[T1 , . . . , Tn ] and R = S/I(G) where G is a graph. If G is
claw-free then there exist b ∈ N and ρ ∈ R with ρ > 1 so that
bρi
.
i
Proof. Let r be the radius of convergence of PkR (z). By (4.1) we have that PkR (z) =
(1+z)c
h(−z) inside BC (0, r), where h(z) is the h-polynomial of R and c = dim(R). Since
R is not a complete intersection, we conclude as in the proof of Theorem 4.3 that
0 < r < 1 and −r is a root of h(z); let b be the multiplicity of this root. The result
will follow as in the proof of Theorem 4.3 via Proposition 4.1 once we show that
−r is the only root of h(z) on the boundary of BC (0, r).
Let ∆ be the independence complex of G, i.e., the simplicial complex having I(G)
as its Stanley-Reisner ideal. By [12, 1.1], f (z), the f -polynomial of ∆, has only real
roots because G is a claw-free graph; hence so does h(z) as these two polynomials
εi (R) ∼
9
z
) for every z 6= 1. Since r is not a root of
satisfy the relation h(z) = (1 − z)c f ( 1−z
h(z) (see [34, 4.8]), the result follows.
Remark 4.5. Theorem 4.3 provides another class of graphs G whose deviations
grow exponentially, namely those whose complementary graph is chordal. By [17],
these graphs G are precisely the ones such that I(G) has a linear resolution (note
that in this case the characteristic of the field k is not relevant), and thus S/I(G)
is Golod by [9, Theorem 7].
Example 4.6. We now examine in detail the deviations for paths Pn and cycles
Cn on n vertices, with n > 3 (see also [11, 2.6]). Consider the f -polynomial and
the h-polynomial of the independence complex of these graphs. The roots of f for
Cn and Pn are determined explicitly in [1] and are respectively
jnk
1
s
=
1,
2,
.
.
.
,
c(n)
=
−
s
2
2(1 + cos( 2s−1
n π))
j
1
n + 1k
p(n)
=
−
s
=
1,
2,
.
.
.
,
.
s
2sπ
2
))
2(1 + cos( n+2
From the relation between the f -polynomial and the h-polynomial we have
τ
is a root of h.
τ 6= −1 is a root of f ⇔
1+τ
(n)
(n)
Note that h(1) 6= 0. For each s, n we have cs , ps < 0 and in both cases s = 1 gives
the root τ of f with minimum modulus. Note that τ ∈ [− 21 , 0). Let σ 6∈ {τ, −1}
σ
τ
be another root of f . If σ < −1 then | 1+σ
| > 1 > | 1+τ
|, while if −1 < σ < τ
τ
x
σ
then | 1+σ | > | 1+τ | because the map x 7→ | 1+x | is strictly decreasing on (−1, 0]. It
follows that the root of minimum modulus of h is the image of the root of minimum
x
modulus of f via the map x 7→ 1+x
. For both Cn and Pn , this root converges to
1
− 3 as n goes to infinity. Applying Theorem 4.4 we get that
εi (S/I(Cn )) ∼
εi (S/I(Pn )) ∼
(2(1 + cos( πn )) − 1)i
,
i
2π
(2(1 + cos( n+2
)) − 1)i
i
.
As n goes to infinity, both expressions on the RHS approach
3i
i .
We observe that the integer b in Theorem 4.4 can be arbitrary.
Example 4.7. If G consists of m connected components, each of them being a
copy of P4 , then h(z) = (1 + 2z)m and hence b = m in Theorem 4.4.
Remark 4.8. Theorem 4.4 can also be used to produce examples of non-monomial
Koszul algebras with exponential growth of deviations: by (4.1), two Koszul algebras with the same h-polynomial must have the same deviations for i > 2.
We conclude by asking a combinatorial question.
Question 4.9. Let G be a graph and let h be the h-polynomial of S/I(G). Does h
have a unique (but not necessarily simple) root of minimum modulus?
Note that, since we already know that there exists a real negative root of h of
minimum modulus, an affirmative answer to Question 4.9 would imply the exponential growth of deviations of S/I(G) for any graph G such that S/I(G) is not a
10
complete intersection. Moreover, it is already known that the f -polynomial of the
independence complex of G admits a unique (but not necessarily simple) root of
minimum modulus (cf. [18], [14]).
Acknowledgements
We would like to thank the organizers, lecturers, and participants of Pragmatic
2014 where this project started. We thank especially Srikanth Iyengar and Aldo
Conca for introducing us to the topic and for the enlightening discussions. The last
four authors also thank their Ph.D. advisors, Aldo Conca, Craig Huneke, Bernd
Ulrich, and Giulio Caviglia, respectively. We acknowledge the use of the software
Macaulay2 [19]. The authors are also grateful to the referee for her or his helpful
suggestions.
References
[1] S. Alikhani, Y. H. Peng, Independence roots and independence fractals of certain graphs, J.
Appl. Math. Comput. 36 (2011), 89–100.
[2] E. F. Assmus, On the homology of local rings, Illinois J. Math. 3 (1959), 187–199.
[3] L. L. Avramov, Flat morphisms of complete intersections, Soviet Math. Dokl. 16 (1975),
1413–1417; [translated from:] Dokl. Akad. Nauk. SSSR 225 (1975), 11–14 [Russian].
[4] L. L. Avramov, Local algebra and rational homotopy, Homotopie algébrique et algèbre locale;
Luminy, 1982 (J. M. Lemaire, J. C. Thomas, eds.) Astérisque 113–114, Soc. Math. France,
Paris (1984), pp. 15–43.
[5] L. L. Avramov, Infinite free resolutions, Six lectures on commutative algebra (Bellaterra,
1996), 1–118, Progr. Math., 166, Birkhäuser, Basel, 1998.
[6] L. L. Avramov, A. Conca, S. B. Iyengar, Free resolutions over commutative Koszul algebras,
Math. Res. Lett. 17 (2010), 197–210.
[7] I. K. Babenko, Analytical properties of Poincaré series of a loop space, Math. Notes 27
(1980), 359–367; [translated from:] Mat. Zametki 27 (1980), 751–765 [Russian].
[8] J. Backelin, Les anneaux locaux à relations monomiales ont des séries de Poincaré-Betti
rationelles, C. R. Acad. Sc. Paris (1982), 607–610.
[9] J. Backelin, R. Fröberg, Koszul algebras, Veronese subrings, and rings with linear resolutions,
Rev. Roumaine Math. Pures. Appl. 30 (1985), 85–97.
[10] A. Bigatti, Upper bounds for the Betti numbers of a given Hilbert function, Comm. Algebra
21 (1993), 2317–2334.
[11] A. Boocher, A. D’Alı̀, E. Grifo, J. Montaño, A. Sammartano, Edge ideals and DG algebra
resolutions, Le Matematiche 70 (2015), 215–237.
[12] M. Chudnovsky, P. Seymour, The roots of the independence polynomial of a clawfree graph,
J. Combin. Theory Ser. B 97 (2007), 350–357.
[13] A. Conca, Koszul algebras and their syzygies, in: Combinatorial Algebraic Geometry, Lecture
Notes in Mathematics 2108, Springer (2014), 1–31.
[14] P. Csikvári, Note on the smallest root of the independence polynomial, Combin. Probab.
Comput. 22 (2013), 1–8.
[15] Y. Felix, J. C. Thomas, The radius of convergence of Poincaré series of loop spaces, Invent.
Math. 68 (1982), 257–274.
[16] R. Fröberg, Determination of a class of Poincaré series, Math. Scand. 37 (1975), 29–39.
[17] R. Fröberg, On Stanley-Reisner rings, Banach Cent. Publ. 26 (1990), 57–70.
[18] M. Goldwurm, M. Santini, Clique polynomials have a unique root of smallest modulus, Inform.
Process. Lett. 75 (2000), 127–132.
[19] D. R. Grayson, M. E. Stillman, Macaulay2, a software system for research in algebraic
geometry, available at www.math.uiuc.edu/Macaulay2/.
[20] T. H. Gulliksen, A proof of the existence of minimal R-algebra resolutions, Acta Math. 120,
1968, 53–58.
[21] S. Halperin, The nonvanishing of the deviations of a local ring, Comment. Math. Helv. 62
(1987), 646–653.
[22] J. Herzog, T. Hibi, Monomial ideals, Springer London (2011).
11
[23] J. Herzog, V. Reiner, V. Welker, Componentwise Linear Ideals and Golod Rings, Michigan
Math. J. 46 (1999), 211–223.
[24] H. A. Hulett, Maximum Betti numbers of homogeneous ideals with a given Hilbert function,
Comm. Algebra 21 (1993), 2335–2350.
[25] J. Mccullough, I. Peeva, Infinite graded free resolutions, to appear in Commutative Algebra and Noncommutative Algebraic Geometry (Eisenbud, Iyengar, Singh, Stafford, Van den
Bergh, Eds.), Math. Sci. Res. Inst. Publ., Cambridge University Press.
[26] K. Pardue, Deformation classes of graded modules and maximal Betti numbers, Illinois J.
Math. 40 (1996), 564–585.
[27] I. Peeva, 0-Borel fixed ideals, J. Algebra 184 (1996), 945–984.
[28] I. Peeva, Consecutive cancellations in Betti numbers, Proc. Amer. Math. Soc. 132 (2004),
3503–3507.
[29] A. Polishchuk, L. Positselski, Quadratic Algebras, Univ. Lecture Ser. 37, Amer. Math. Soc.,
Providence, RI (2005).
[30] J. E. Roos, On computer-assisted research in homological algebra, Mathematics and Computers in Simulation 42 (1996), 475–490.
[31] C. Schoeller, Homologie des anneaux locaux noethériens, C. R. Acad. Sci. Paris Sér. A 265
(1967), 768–771.
[32] L. C. Sun, Growth of Betti numbers of modules over local rings of small embedding codimension or small linkage number, J. Pure and Appl. Algebra 96 (1994), 57–71.
[33] J. Tate, Homology of Noetherian rings and local rings, Illinois J. Math. 1 (1957), 14–27.
[34] J. Uliczka, Remarks on Hilbert series of graded modules over polynomial rings, Manuscripta
Math. 132 (2010), 159–168.
Adam Boocher, School of Mathematics, University of Edinburgh, James Clerk Maxwell
Building, Mayfield Road, Edinburgh EH9 3JZ, Scotland
E-mail address: [email protected]
Alessio D’Alı̀, Dipartimento di Matematica, Università degli Studi di Genova, Via
Dodecaneso 35, 16146 Genova, Italy
E-mail address: [email protected]
Eloı́sa Grifo, Department of Mathematics, University of Virginia, 141 Cabell Drive,
Kerchof Hall, Charlottesville, VA 22904, USA
E-mail address: [email protected]
Jonathan Montaño, Department of Mathematics, University of Kansas, 405 Snow
Hall, 1460 Jayhawk Blvd, Lawrence, KS 66045
E-mail address: [email protected]
Alessio Sammartano, Department of Mathematics, Purdue University, 150 North University Street, West Lafayette, IN 47907, USA
E-mail address: [email protected]
12
| 0math.AC
|
Minimax signal detection under weak noise assumptions∗
arXiv:1611.09008v3 [math.ST] 1 Nov 2017
Clément Marteau,
Univ Lyon, Université Claude Bernard Lyon 1,
CNRS UMR 5208, Institut Camille Jordan,
43 blvd. du 11 novembre 1918,
F-69622 Villeurbanne cedex, France
Email: [email protected]
and
Theofanis Sapatinas,
Department of Mathematics and Statistics,
University of Cyprus,
P.O. Box 20537, CY 1678 Nicosia, Cyprus.
Email: [email protected]
November 2, 2017
Abstract
We consider minimax signal detection in the sequence model. Working with certain
ellipsoids in the space of square-summable sequences of real numbers, with a ball of positive
radius removed, we obtain upper and lower bounds for the minimax separation radius in the
non-asymptotic framework, i.e., for a fixed value of the involved noise level. We use very
weak assumptions on the noise (i.e., fourth moments are assumed to be uniformly bounded).
In particular, we do not use any kind of Gaussianity or independence assumption on the
noise. It is shown that the established minimax separation rates are not faster than the ones
obtained in the classical sequence model (i.e., independent standard Gaussian noise) but,
surprisingly, are of the same order as the minimax estimation rates in the classical setting.
Under an additional condition on the noise, the classical minimax separation rates are also
retrieved in benchmark well-posed and ill-posed inverse problems.
AMS 2000 subject classifications: 62G05, 62K20
Keywords and phrases: Ellipsoids; ill-posed inverse problems; minimax signal detection;
well-posed inverse problems.
∗
This work was supported by the LABEX MILYON (ANR-10-LABX-0070) of Université de Lyon, within the
program ‘Investissements d’Avenir” (ANR-11-IDEX- 0007) operated by the French National Research Agency
(ANR).
1
1
Introduction
We consider the following sequence model (SM),
y k = bk θ k + ε ξ k ,
k ∈ N,
(1.1)
where N can be either N = {1, 2, . . .} or Nn = {1, . . . , n} for some n ≥ 1, b = (bk )k∈N is a
known positive sequence, θ = (θk )k∈N ∈ l2 (N ) is the unknown signal of interest, ξ = (ξk )k∈N is
a sequence of random variables (the noise), and ε > 0 is a known parameter (the noise level).
The observations are given by the sequence y = {yk }k∈N from the SM (1.1) and their joint law is
denoted by Pθ,ξ . Here, l2 (N ) denotes the space of squared-summable sequence of real numbers,
i.e.,
(
)
X
2
N
2
2
l (N ) = θ ∈ R : kθk :=
θj < +∞ .
k∈N
Let C > 0 be a known fixed constant. Concerning the noise, we will assume that ξ ∈ Ξ, where
2
4
Ξ := Ξ(C) = ξ : E[ξk ] = 0, E[ξk ] = 1 ∀k ∈ N and sup E[ξk ] ≤ C < +∞ .
(1.2)
k∈N
The SM (1.1) arises in many well-known situations. Consider for instance the stochastic
differential equation
dZε (t) = Af (t) + ǫdU (t), t ∈ [0, 1],
where A is a known bounded linear operator acting on L2 ([0, 1]), f (·) ∈ L2 ([0, 1]) is the unknown
response function that one wants to detect or estimate, U (·) is a given stochastic process on [0, 1]
and ε > 0 is a known parameter (the noise level). For the sake of simplicity, we only consider
the case when A is injective (meaning that A has a trivial nullspace).
• Let U (·) = W (·) be the standard Wiener process. Then, if A is the identity operator, we
can retrieve the SM (1.1) in the Fourier domain with bk = 1 for all k ∈ N = N and the
ξk , k ∈ N , are independent standard Gaussian random variables (direct problem). If A is
a self-adjoint operator with an eigen-decomposition, we can retrieve the SM (1.1) where
bk > b0 for some b0 > 0 for all k ∈ N = N and the ξk , k ∈ N , are independent standard
Gaussian random variables (well-posed inverse problems). If A is a compact operator, we
can retrieve the SM (1.1) where bk > 0 for all k ∈ N (since A is injective) with bk → 0
as k → +∞ and the ξk , k ∈ N , are independent standard Gaussian random variables
(ill-posed inverse problems). For more details regarding all these models, we refer to, e.g.,
[5].
• Let U (·) = W−γ (·), γ ∈]0, 1/2[, be the truncated fractional Brownian motion and let A
be the identity operator. Then, we can retrieve the SM (1.1) in the spline domain with
bk = (πk)−2γ (1 + o(1)) as k → +∞ and the ξk , k ∈ N , are (non-independent) standard
Gaussian random variables. For more details, we refer to, e.g., [10], [4].
The non-parametric inverse regression problem also provides observations of the form (1.1).
Indeed, consider the model
1
i
+ √ ηi , i ∈ {1, . . . , n},
Zi = Af
n
n
where A is a known (injective) bounded linear operator acting on L2 ([0, 1]), f (·) ∈ L2 ([0, 1])
is the unknown response function that one wants to detect or estimate, and ηi , i ∈ Nn , is
a sequence of independent and identically distributed random variables with zero mean, variance one and finite fourth moment. Given any appropriate bases (or, even, a tight frame, see,
2
e.g., [13], p. 126), we can retrieve the SM (1.1) with bk = 1 for all k ∈ Nn when A is the
identity operator, see, e.g., [16], Chapter 1. When A is a compact operator, we can retrieve an
approximation of the SM (1.1) where (bk )k∈Nn is a fixed sequence that depends on A, see, e.g., [2].
Minimax signal detection has been considered in the literature over the last two decades.
We refer to, e.g., [1], [12], [8], [11], [7], [6], [14]. All these contributions consider the classical
Gaussian sequence model (1.1), i.e,. where the ξk , k ∈ N , are independent standard Gaussian
random variables. We refer to [15] for a survey on available results and a discussion on the link
between asymptotic (the noise level is assumed to tend to zero) and non-asymptotic (the noise
level is assumed to be fixed) approaches to minimax signal detection. The aim of this work is to
obtain upper and lower bounds on the minimax separation radius in the non-asymptotic framework, for the general model (1.1) under weak assumptions on the noise, i.e., when ξ ∈ Ξ, where
the set Ξ has been introduced in (1.2). In particular, we do not use any kind of Gaussianity
or independence assumption on the noise. We prove that the minimax separation rates are not
faster than the ones obtained in the classical sequence model (see, e.g., [1], [12], [8], [11], [7], [6],
[14]) but, surprisingly, are of the same order as the minimax estimation rates in the classical
setting. Moreover, under additional conditions on the noise, we show that the classical minimax
separation rates can be retrieved in benchmark well-posed and ill-posed inverse problems.
Throughout the paper, we use the following notations. Given two sequences (ck )k∈N and
(dk )k∈N of real numbers, ck ∼ dk means that there exist 0 < κ0 ≤ κ1 < ∞ such that κ0 ≤
ck /dk ≤ κ1 for all k ∈ N , while ck . dk (resp. ck & dk ) means ck ≤ c0 dk (resp. ck ≥ c0 dk ) for
some c0 > 0 for all k ∈ N . Also, x ∧ y := min(x, y), for all x, y ∈ R.
2
Minimax Signal Detection
Given observations from the SM (1.1), we consider the signal detection problem, i.e., our aim is
to test
H0 : θ = 0 versus H1 : θ ∈ Θa (rε ).
(2.1)
Given a non-decreasing sequence a = (ak )k∈N of positive real numbers, with ak → +∞ as
k → +∞ when N = N, and a radius rε > 0, the set Θa (rε ) is defined as
Θa (rε ) = {θ ∈ Ea , kθk ≥ rε } ,
where
Ea =
(
θ ∈ l2 (N ),
X
k∈N
(2.2)
)
a2k θk2 ≤ 1 .
The set Ea can be seen as a condition on the decay of θ. The cases where the sequence a increases
very fast correspond to the signal θ with small coefficients. In such a case, the corresponding
signal can be considered as being ‘smooth’. The sequence a being fixed, the main issue for the
minimax signal detection problem (2.1)-(2.2) is then to characterize the values of the radius
rε > 0 for which both hypotheses H0 (called the null hypothesis) and H1 (called the alternative
hypothesis) are ‘separable’.
In the following, a (non-randomized) test Ψ := Ψ(y) will be defined as a measurable function
of the observation y = (yk )k∈N from the SM (1.1) having values in the set {0, 1}. By convention,
H0 is rejected if Ψ = 1 and H0 is not rejected if Ψ = 0. Then, given a test Ψ, we can investigate
• the type I (first kind) error probability defined as
sup P0,ξ (Ψ = 1),
ξ∈Ξ
3
(2.3)
which measures the worst probability of rejecting H0 when H0 is true (i.e., θ = 0, ξ ∈ Ξ,
where Ξ is defined in (1.2)); it is often constrained as being bounded by a prescribed level
α ∈]0, 1[, and
• the type II (second kind) error probability defined as
sup
Pθ,ξ (Ψ = 0),
(2.4)
θ∈Θa (rε )
ξ∈Ξ
which measures the worst possible probability of not rejecting H0 when H0 is not true
(i.e., when θ ∈ Θa (rε ) and ξ ∈ Ξ, where Ξ is defined in (1.2)); one would like to ensure
that it is bounded by a prescribed level β ∈]0, 1[.
We emphasize that in the classical minimax signal detection problem, the protection against
all possible noise distributions (i.e., supξ∈Ξ ) is not required, since the noise distribution is completely known. However, in the more general setting that we consider, in order to produce some
kind of robustness, we have adapted the definitions of type I and type II error probabilities to
accommodate the (possible) uncertainty on the noise.
Let α, β ∈]0, 1[ be given, and let Ψα be an α-level test, i.e., Ψα is such that supξ∈Ξ P0,ξ (Ψα =
1) ≤ α.
Definition 2.1 The separation radius of the α-level test Ψα over the class Ea is defined as
rε (Ea , Ψα , β) := inf rε > 0 :
sup Pθ,ξ (Ψα = 0) ≤ β .
θ∈Θa (rε )
ξ∈Ξ
In some sense, the separation radius rε (Ea , Ψα , β) corresponds to the smallest possible value
of the available signal kθk for which H0 and H1 can be ‘separated’ by the α-level test Ψα with
prescribed type I and type II error probabilities, α and β, respectively.
Definition 2.2 The minimax separation radius r̃ε := r̃ε (Ea , α, β) > 0 over the class Ea is defined
as
(2.5)
r̃ε := inf rε (Ea , Ψ̃α , β),
Ψ̃α
where the infimum is taken over all α-level tests Ψ̃α .
The minimax separation radius r̃ε corresponds to the smallest radius rε > 0 such that there
exists some α-level test Ψ̃α for which the type II error probability is not greater than β.
It is worth mentioning that Definitions 2.1 and 2.2 are valid for any fixed ε > 0 (i.e., it is not
required that ε → 0). The performances of any given test Ψα is easy to handle in the sense that
the type I error probability is bounded by α (i.e., Ψα is an α-level test), and that the dependence
of the minimax separation radius r̃ε with respect to given α and β can be precisely described.
3
3.1
Control of the Upper and Lower bounds
The spectral cut-off test and control of the upper bound
We define below a spectral cut-off test for the SM model (1.1) with ξ ∈ Ξ, where Ξ is defined in
(1.2). First, we show that it is an α-level test and then we obtain an upper bound for its type
II error probability.
4
Given a bandwidth D ∈ N and α ∈]0, 1[, we consider the following spectral cut-off test
Ψα,D := 1{TD ≥t1−α,D } ,
where
TD =
D
X
k=1
(3.1)
2
2
b−2
k (yk − ε )
and t1−α,D denotes a threshold depending on α and D. It is easily seen that, for all D ∈ N ,
D
X
Eθ,ξ [TD ] =
θk2 ,
k=1
and
4
Var0,ξ (TD ) = ε
|
D
X
2
b−4
k E[(ξk
k=1
2
4
− 1) ] + ε
{z
}
:=R0 (D)
D
X
k,l=1
k6=l
−2
2
2
b−2
k bl E[(ξk − 1)(ξl − 1)],
|
{z
(3.2)
}
:=S0 (D)
where the assumption ξ ∈ Ξ guarantees that the above variance is finite for every D ∈ N .
Proposition 3.1 Let α ∈]0, 1[ be given. Consider the spectral cut-off test Ψα,D defined in (3.1).
Then, for all ε > 0,
sup P0,ξ (Ψα,D = 1) ≤ α
ξ∈Ξ
as soon as
1 p
R0 (D) + S0 (D).
t1−α,D ≥ √
α
(3.3)
The proof of this proposition is postponed to Section 7.1.
Remarks:
• Using simple bounds, it is easily seen that
R0 (D) + S0 (D) ≤ C1 ε4
X
1≤k≤D
where
4
b−4
k + C1 ε
X
1≤k≤D
2
≤ 2C1 ε4
b−2
k
X
1≤k≤D
C1 := sup sup E[(ξk2 − 1)2 ] < +∞,
2
, (3.4)
b−2
k
(3.5)
ξ∈Ξ k∈N
since
D
X
k=1
Hence, the choice
b−4
k
≤
( max b−2
)
1≤k≤D k
2
t1−α,D = K1 ε
D
X
b−2
k ,
k=1
D
X
k=1
b−2
k
≤
D
X
k=1
b−2
k
!2
.
√
2C1
where K1 = √ ,
α
(3.6)
ensures that (3.3) is satisfied and that the spectral cut-off test Ψα,D defined in (3.1) is an
α-level test.
• In the classical setting (i.e., independent Gaussian noise), the threshold t1−α,D can be
chosen as the (1 − α)-quantile of the variable TD under H0 . This is no more the case here
since only a uniform bound on the fourth moment of the sequence ξk , k ∈ N , is available.
5
Proposition 3.2 Let α, β ∈]0, 1[ be given. Consider the spectral cut-off test Ψα,D defined in
(3.1). Select the threshold t1−α,D as in (3.6). Then, for all ε > 0,
sup
θ∈Θa (rε,D )
ξ∈Ξ
Pθ,ξ (Ψα,D = 0) ≤ β,
for all radius rε,D > 0 such that
2
rε,D ≥ Cβ ε
D
X
−2
b−2
k + aD ,
k=1
where Cβ > 0 is the solution of the equation (7.11).
The proof of Proposition is postponed to Section 7.2.
Remark: For practical purposes, the solution Cβ > 0 of equation (7.11) can be chosen as
Cβ = 8K2 /β. In particular, there exists some β0 > 0 such that 1 − K1 Cβ−1 ≥ 1/2 for all β ≤ β0 ,
hence ensuring that (7.9) is satisfied for all β small enough.
3.2
Control of the lower bound
We propose below a lower bound on the minimax type II error probability for the SM (1.1) with
ξ ∈ Ξ, where Ξ is defined in (1.2). In the sequel, the term inf Ψα corresponds to an infimum
taken over all possible α-level tests.
Proposition 3.3 Let α ∈]0, 1[ and β ∈]0, 1 − α[ be fixed. Then, for all ε > 0
sup
inf
Ψα θ∈Θa (rε )
ξ∈Ξ
Pθ,ξ (Ψα = 0) ≥ β,
for all D ∈ N and rε > 0 such that
rε2
≤
X
D
1
−2
2
ln(Cα,β ) ε
b−2
k ∧ aD ,
4
k=1
where Cα,β = 1 + 4(1 − α − β)2 .
The proof of Proposition 3.3 is postponed to Section 7.3. The main difficulty is to construct an
appropriate distribution for ξ that will allow one to obtain the largest possible lower bound.
4
Minimax Separation Radius
The following theorem provides upper and lower bounds for the minimax separation radius
r̃ε > 0 in the SM (1.1) with ξ ∈ Ξ, where Ξ is defined in (1.2).
Theorem 4.1 Let α, β ∈ ]0, 1[ be given. Then, for all ε > 0, the minimax separation radius
r̃ε > 0 satisfies
#
"
#
"
X
D
D
X
1
−2
−2
ln(Cα,β ) ε2
≤ r̃ε2 ≤ inf Cβ ε2
,
(4.1)
b−2
b−2
sup
k ∧ aD
k + aD
D∈N
4
D∈N
k=1
k=1
where Cβ > 0 is the solution of the equation (7.11) and Cα,β = 1 + 4(1 − α − β)2 .
6
The proof of Theorem 4.1 is postponed to Section 7.4.
Remark: If both sequences a = (ak )k∈N and b−1 = (b−1
k )k∈N are non-decreasing and satisfy
a⋆ ≤
aD−1
≤ a⋆
aD
and b⋆ ≤
bD−1
≤ b⋆
bD
for all
D ∈ N \ {1},
(4.2)
for some constants 0 < a⋆ ≤ a⋆ < ∞ and 0 < b⋆ ≤ b⋆ < ∞, then, it is easily seen that both
upper and lower bounds on the minimax separation radius r̃ε > 0, established in Theorem 4.1,
are of the same order. This follows easily working along the same lines of the proof of Proposition 4.1 in [15]. We note also that the condition (4.2) is satisfied for various combinations of
interest, among them: (i) mildly ill-posed inverse problems (bk ≍ k−t , k ∈ N, for some t > 0)
with ordinary smooth functions (ak ≍ ks , k ∈ N, for some s > 0), (ii) severely ill-posed inverse
problems (bk ≍ e−kt , k ∈ N, for some t > 0) with ordinary smooth functions (ak ≍ ks , k ∈ N, for
some s > 0), and (iii) mildly ill-posed inverse problems (bk ≍ k−t , k ∈ N, for some t > 0) with
super-smooth functions (ak ≍ eks , k ∈ N, for some s > 0). Among the possible situations where
the condition (4.2) is not satisfied, one can mention, for instance, power-exponential behaviors
l
r
(ak ≍ ek s , j ∈ N, for some s > 0 and l > 1, or bk ≍ e−k t , k ∈ N, for some t > 0 and r > 1).
See also Remark 4.3 in [15].
Remark: Note that the upper and lower bounds on the minimax separation radius r̃ε > 0,
established in Theorem 4.1, are quite different compared to the classical minimax separation
radii available in the literature, obtained in the SM (1.1) with independent standard Gaussian
noise (see, e.g., [15]). Although the bias terms a−2
D coincide, the corresponding variance terms
differ. In particular,
in
the
SM
(1.1)
with
ξ
∈
Ξ,
where Ξ is defined in (1.2), the variance term
PD −2
2
Gausiian noise, the
is of order ε
k=1 bk , while
qPfor the SM (1.1) with independent standard
P
D
D
−2
−4
2
variance term is of order ε2
k=1 bk is not greater than
k=1 bk . We stress that the term ε
qP
D
−4
the termε2
k=1 bk , which entails that the minimax separation rates are not faster compared
to the ones obtained in the classical model. It is also worth mentioning that, surprisingly, the
bias and variance terms in the SM (1.1) with ξ ∈ Ξ, where Ξ is defined in (1.2), are of the same
order of the corresponding terms in the classical minimax estimation setting. In particular,
the minimax separation rates in our general setting coincide with the minimax estimation rates
obtained in the classical estimation setting. For illustrative purposes, the Table 4.1 (see also
Table 1 in [5]) provides these minimax separation rates for benchmark problems, i.e., well-posed,
mildly ill-posed and severely ill-posed problems for ellipsoids with ordinary smooth and supersmooth sequences.
Minimax separation
rate (r̃ε2 )
well-posed
bk ∼ 1
mildly ill-posed
bk ∼ k−t
severely ill-posed
bk ∼ exp{−kt}
ordinary-smooth
ak ∼ ks
ε4s/(2s+1)
super-smooth
ak ∼ exp{ks}
ε2 (ln ε−1 )
ε4s/(2s+2t+1)
ε2 (ln ε−1 )2t+1
(ln ε−1 )−2s
ε4s/(2s+2t)
Table 4.1: Minimax separation rates for the SM (1.1) with ξ ∈ Ξ, where Ξ is defined in (1.2).
Remark: If the supremum over all possible noise distributions ξ ∈ Ξ is not considered in
the definition of type I and type II error probabilities, then it is easily seen that upper bound
on the type II error probability obtained in Proposition 3.2 still holds true. However, the
7
corresponding lower bound obtained in Proposition 3.3 is only true under Gaussianity. This
implies that the minimax separation rates displayed in Table 4.1 are still valid in the SM (1.1)
with non-independent standard Gaussian noise ξ.
5
An Additional Condition on the Noise to Obtain the Classical
Minimax Separations Rates
In this section, it is demonstrated that, under an additional condition on the noise ξ ∈ Ξ in the
SM (1.1), one is able to retrieve the classical minimax separation rates in benchmark well-posed
and ill-posed inverse problems.
Recall from equation (7.2), displayed in the proof of Proposition 3.2, that the variance of TD
can be written as
Varθ,ξ (TD ) = Rθ (D) + Sθ (D),
where
Rθ (D) :=
D
X
k=1
and
Sθ (D) :=
D
X
k,l=1
k6=l
2
2
b−4
k Varθ,ξ (yk − ε )
2
2 2
2
−2
b−2
k bl Cov θ,ξ (yk − ε , yl − ε ).
In the classical setting (i.e., independent standard Gaussian noise ξ), Sθ (D) = 0 for all θ ∈ l2 (N ).
Hence, in order to retrieve the classical minimax separation rates in the SM (1.1) with ξ ∈ Ξ,
where Ξ is defined in (1.2), Sθ (D) needs to be of the order of Rθ (D). We achieve this separately
under the null and the alternative hypotheses, for benchmark problems, such as, well-posed,
mildly ill-posed and severely ill-posed inverse problems.
We stress that in this section, we will only deal with upper bounds. Indeed, the lower
bounds established previously in the literature (see, e.g., [15], Theorem 4.1) for the independent
standard Gaussian noise are still valid in our set-up.
5.1
Well-posed and mildly ill-posed inverse problems
We assume that
bk ∼ k−t
∀k ∈ N
for some t ≥ 0 (t = 0 refers to well-posed inverse problems while t > 0 refers to mildly ill-posed
inverse problems). We start our discussion under the null hypothesis. Recall from (3.2) that
Var0,ξ (TD )
=
4
ε
D
X
2
b−4
k E[(ξk
D
X
2
b−4
k Var(ξk )
k=1
=
4
ε
2
4
− 1) ] + ε
4
+ε
D
X
D
X
k,l=1
k6=l
−2
2
2
b−2
k bl E[(ξk − 1)(ξl − 1)]
2 2
−2
b−2
k bl Cov(ξk , ξl )
k,l=1
k6=l
k=1
:= R0 (D) + S0 (D).
Using simple calculations, we can see that
R0 (D) ∼ ε4
D
X
k=1
8
k4t ∼ ε4 D 4t+1 .
Our aim is to exhibit a condition for which S0 (D) is (at least) of the same order as R0 (D).
Assumption HD : Let ξ ∈ Ξ, where Ξ is defined in (1.2), and, for all k, l ∈ N , let (ξk , ξl )′ be a
bivariate Gaussian random vector. Moreover, there exists s > 0 such that
ρkl := |Cov(ξk , ξl )| .
1
|k − l|s
∀k, l ∈ N , k 6= l.
Due to the Isserlis Theorem (see, e.g., [9]), it can be seen that, thanks to Assumption HD ,
for all k, l ∈ N , with k 6= l,
1
|k − l|2s
Cov(ξk2 , ξl2 ) = 2Cov2 (ξk , ξl ) .
(5.1)
and
E[(ξk2 − 1)ξl ] = E[(ξk2 − 1)ξl ] = 0.
(5.2)
These results allow us to propose a sharp control of the variance of TD under the null hypothesis.
Proposition 5.1 Assume that Assumption HD holds with s > 1/2. Then,
S0 (D) = o(R0 (D))
D → +∞.
as
The proof of Proposition 5.1 is postponed to Section 7.5.
Now, we propose a similar analysis under the alternative hypothesis.
Proposition 5.2 Assume that Assumption HD holds with s > 1/2. Then, for all γ ∈]0, 1[,
!2
D
D
X
X
θk2 .
b−4
Varθ,ξ (TD ) . (1 + γ −1 )ε4
k +γ
k=1
k=1
The proof of Proposition 5.2 is postponed to Section 7.6.
Starting from (7.1), and using Propositions 5.1 and 5.2, we get
Pθ,ξ (Ψα,D = 0) ≤
.
Varθ,ξ (TD )
P
D
2
k=1 θk − t1−α,D
(1 + γ −1 )ε4
P
PD
−4
k=1 bk + γ
D
2
k=1 θk
≤ β
provided
D
X
k=1
which holds as soon as
2
− ε4
PD
P
D
2
k=1 θk
2
−4
k=1 bk
2
v
uD
uX
2t
2
θk & ε
b−4
k ,
k=1
v
uD
uX
−2
2t
2
b−4
kθk & aD + ε
k .
k=1
The last inequality provides a classical condition that has been already discussed in, e.g., [1],
[7] and [12], or in Theorem 4.1 of [15], in the specific case where the noise ξ in the SM (1.1) is
assumed to be independent standard Gaussian. This entails that the Assumption HD suffices
to retrieve the classical minimax separation rates for mildly ill-posed inverse problems.
9
5.2
Severely ill-posed inverse problems
We assume in this section that
bk ∼ e−kt
∀k ∈ N
for some t > 0. Since minimax estimation and minimax separation rates in the classical setting
are of the same order (see, e.g., Tables 2 and 3 in [7]), we stress that non-independence does not
deteriorate the classical minimax separation rates. In other words, the independent standard
Gaussian assumption on noise ξ is not needed to get the classical minimax separation rates for
severely ill-posed inverse problems.
6
Concluding Remarks
We have established minimax separation rates in a general Gaussian sequence model, i.e., the
noise need neither to be independent nor standard Gaussian. These rates are not faster than
the ones obtained in the classical setting (i.e., independent standard Gaussian noise) but, surprisingly, are of the same order as the minimax estimation rates in the classical setting. The
involved spectral cut-off test depends on the unknown smoothness parameter of the signal under
the alternative hypothesis. It is therefore of paramount importance in practical applications to
provide minimax testing procedures that do not explicitly depend on the associated smoothness
parameter. This is, usually, referred to as the ‘adaptation’ problem. However, such an investigation needs careful attention that is beyond the scope of the present work. In particular, the
dependency of the involved constant with respect to the level α has a more intricate form than
the one involved in the classical setting.
7
7.1
Appendix
Proof of Proposition 3.1
Let ξ ∈ Ξ be fixed. Using the Markov inequality, we get
P0,ξ (Ψα,D = 1) = P0,ξ (TD ≥ t1−α,D ) ≤
provided
7.2
Var0,ξ (TD )
R0 (D) + S0 (D)
≤
≤α
2
t1−α,D
t21−α,D
1 p
t1−α,D ≥ √
R0 (D) + S0 (D).
α
Proof of Proposition 3.2
Let ξ ∈ Ξ be fixed. Using the Markov inequality, we obtain
Pθ,ξ (Ψα,D = 0) = Pθ,ξ (TD < t1−α,D ),
TD − Eθ [TD ] < t1−α,D −
= Pθ,ξ
≤ Pθ,ξ
≤
|TD − Eθ [TD ]| ≥
Varθ,ξ (TD )
P
D
2
k=1 θk − t1−α,D
10
2 ,
D
X
k=1
θk2
D
X
k=1
θk2
!
− t1−α,D
!
(7.1)
where we have implicitly assumed that
D
X
θk2 > t1−α,D .
k=1
Now, we need an upper bound for the variance term. First remark that
!
D
X
2
2
b−2
Varθ (TD ) = Varθ,ξ
k (yk − ε )
k=1
=
D
X
|k=1
2
2
b−4
k Varθ,ξ (yk − ε ) +
{z
}
:=Rθ (D)
D
X
k,l=1
k6=l
2
2 2
2
−2
b−2
k bl Cov θ,ξ yk − ε , yl − ε .
|
{z
(7.2)
}
:=Sθ (D)
Calculation of Rθ (D): Using simple algebra, we get, for all k ∈ N ,
Varθ (yk2 − ε2 ) = Varθ,ξ (bk θk + ε)2 − ε2
= Varθ,ξ b2k θk2 + ε2 ξk2 + 2bk θk εξk − ε2
= Varθ,ξ ε2 (ξk2 − 1) + 2bk θk εξk
= ε4 E[(ξk2 − 1)2 ] + 4ε2 b2k θk2 + 4ε3 bk θk E[ξk3 ].
Hence, using the last equality, we obtain
Rθ (D) ≤ C1 ε4
≤ C1 ε4
D
X
2
b−4
k + 4ε
k=1
D
X
D
X
3
2
b−2
k θk + 4C2 ε
k=1
k=1
2
−2
b−4
k + 4ε ( max bk )
1≤k≤D
k=1
D
X
D
X
b−3
k |θk |
θk2 + 4C2 ε3
D
X
k=1
k=1
b−3
k |θk |,
(7.3)
where the constant C1 has been introduced in (3.5) and
C2 := sup sup |E[ξk3 ]| < +∞.
ξ∈Ξ k∈N
Note that, using first the Cauchy-Schwartz inequality and then the Peter-Paul inequality (see,
e.g., [3], p. 18), we get
ε3
D
X
k=1
b−3
k |θk | =
D
X
k=1
−1
ε2 b−2
k εbk |θk |
v
v
u D
u D
u X
u X
−4 t 2
t
2
4
ε
ε
≤
bk
b−2
k θk
k=1
k=1
D
D
k=1
D
X
k=1
≤
1 4 X −4 1 2 X −2 2
ε
bk + ε
bk θ k
2
2
≤
1 4
ε
2
b−4
k
D
X
1
θk2 .
+ ε2 ( max b−2
)
2 1≤k≤D k
(7.4)
k=1
k=1
Combining inequalities (7.3) and (7.4), we obtain
Rθ (D) ≤ (C1 + 2C2 )ε4
D
X
2
−2
b−4
k + (4 + 2C2 )ε ( max bk )
1≤k≤D
k=1
11
D
X
k=1
θk2 .
(7.5)
Calculation of Sθ (D): First, remark that, for all k ∈ N , on noting that Eθ,ξ [(yk2 − ε2 )] = b2k θk2
and E[ξk2 ] = 1, we get
Covθ,ξ yk2 − ε2 , yl2 − ε2
= Eθ,ξ (yk2 − ε2 − b2k θk2 ) (yl2 − ε2 − b2l θl2 )
= Eθ,ξ ((bk θk + εξk )2 − ε2 − b2k θk2 ) ((bl θl + εξl )2 − ε2 − b2l θl2 )
= ε4 E[(ξk2 − 1)(ξl2 − 1)] + 4ε2 bk bl θk θl E[ξk ξl ] + 2ε3 bl θl E[(ξk2 − 1)ξl ] + 2ε3 bk θk E[(ξl2 − 1)ξk ].
Hence,
Sθ (D) = 2ε4
D
X
k,l=1
k6=l
+8ε2
−2
2
2
b−2
k bl E[(ξk − 1)(ξl − 1)]
D
X
−1
b−1
k bl θk θl E[ξk ξl ]
D
X
2
−1
b−2
k bl θl E[(ξk − 1)ξl ]
D
X
−1
2
b−2
l bk θk E[(ξl − 1)ξk ].
k,l=1
k6=l
+4ε3
k,l=1
k6=l
+4ε3
k,l=1
k6=l
(7.6)
Using the Cauchy-Schwarz inequality in each expectation of the above expression, we obtain
4
Sθ (D) ≤ 2C1 ε
D
X
k=1
b−2
k
!2
2
+ 8ε
D
X
k=1
!2
b−1
k |θk |
+ 8ǫ
1/2
C1
3
D
X
k=1
!
D
X
b−1
k |θk |
k=1
b−2
k
!
.
Then, using first the Peter-Paul inequality and then the Cauchy-Schwarz inequality, we get
1/2
Sθ (D) ≤ 2(C1 + 2C1 )ε4
1/2
≤ 2(C1 + 2C1 )ε4
D
X
b−2
k
!2
+ 4(2 + C1 )ε2
b−2
k
!2
+ 4(2 + C1 )ε2
k=1
D
X
k=1
1/2
D
X
k=1
1/2
D
X
b−1
k |θk |
b−2
k
k=1
Varθ,ξ (TD ) ≤ (C1 + 2C2 )ε
+4(C1 +
D
X
b−4
k
k=1
1/2
2C1 )ε4
b−2
k
!2
D
X
b−2
k
D
X
k=1
1/2
+2(10 + C2 + 4C1 )ε2
k=1
12
D
X
k=1
D
X
k=1
Hence, combining (7.2), (7.5) and (7.7), we obtain
4
!2
θk2 .
θk2 .
(7.7)
For all γ ∈]0, 1[, using again the Peter-Paul inequality, we get
4
Varθ,ξ (TD ) ≤ (C1 + 2C2 )ε
+4(C1 +
D
X
b−4
k
k=1
D
X
1/2
2C1 )ε4
b−2
k
k=1
!2
D
X
1/2
4C1 )ε4
+γ −1 (10 + C2 +
b−2
k
k=1
!2
+ γ(10 + C2 +
1/2
4C1 )
D
X
θk2
k=1
!2
.
Hence, since γ ∈]0, 1[, it is easily seen that
Varθ,ξ (TD )
≤
γ(10 + C2 +
1/2
4C1 )
D
X
k=1
n
+ (C1 + 2C2 ) + 4(C1 +
≤
≤
θk2
γ(10 + C2 +
1/2
4C1 )
1/2
2C1 )
D
X
θk2
k=1
1/2
(10 + 5C1 + 2C2 + 12C1 )
:= K2
γ
D
X
k=1
θk2
!2
+γ
!2
!2
ε
(10 + C2 +
n
o
1/2
4C1 )
+ γ −1 10 + 5C1 + 2C2 +
γ
D
X
θk2
k=1
D
X
−1 4
+γ
−1
k=1
b−2
k
!2
!2
+ γ −1 ε4
D
X
k=1
.
4
ε
D
X
k=1
1/2
12C1
b−2
k
b−2
k
o
ε4
!2
!2
D
X
k=1
b−2
k
!2
(7.8)
Now, using (7.1) and (7.8), and choosing γ = Cβ−1 , we get
P
2
PD 2 2
D
−2
−1
4
+γ ε
K2 γ
k=1 bk
k=1 θk
Pθ,ξ (Ψα,D = 0) ≤
P
PD −2 2
D
2 − K ε2
θ
1
k=1 k
k=1 bk
n
o
K2 γ + γ −1 Cβ−2
≤
2
1 − K1 Cβ−1
≤
2K2 Cβ−1
(7.9)
2
1 − K1 Cβ−1
≤ β,
provided that
D
X
D
X
b−2
k
(7.10)
2 = β.
1 − K1 Cβ−1
(7.11)
k=1
θk2 ≥ Cβ ε2
and Cβ is the solution of the equation
2K2 Cβ−1
13
k=1
To conclude the proof, since
that
P
2
k>D θk
≤ a−2
D , remark that inequality (7.10) is satisfied provided
2
2
kθk ≥ Cβ ε
D
X
−2
b−2
k + aD .
k=1
7.3
Proof of Proposition 3.3
When ξ is Gaussian, we will write ΞG instead of Ξ and Pθ,Σ instead of Pθ,ξ , where Σ = (Σkl )k,l∈N
denotes the associated covariance matrix. We also define S = {Σ : Σkk = 1}. Below, Ψα refers
to an α-level test.
Let θ ⋆ ∈ Θa (rǫ ), ξ ⋆ ∈ ΞG and Σ⋆ ∈ S be fixed. Their values will be made precise later on.
Then
sup
inf
Ψα θ∈Θa (rǫ )
ξ∈Ξ
Pθ,ξ (Ψα = 0) ≥ inf
sup
Pθ,ξ (Ψα = 0)
= inf
sup
Pθ,Σ̃ (Ψα = 0)
≥ inf
sup
Pθ,Σ⋆ (Ψα = 0)
Ψα θ∈Θa (rǫ )
ξ∈ΞG
Ψα θ∈Θa (rǫ )
Σ̃∈S
Ψα θ∈Θa (rǫ )
≥ inf Pθ⋆ ,Σ⋆ (Ψα = 0)
Ψα
≥ 1−α−
1/2
1
E0 [L2θ⋆ ,Σ⋆ (Y )] − 1
,
2
where Lθ⋆ ,Σ⋆ (Y ) = dPθ⋆ ,Σ⋆ (Y )/dP0,Σ⋆ (Y ) is the likelihood ratio between the probability measures Pθ⋆ ,Σ⋆ and P0,Σ⋆ (for the last inequality, we refer to, e.g., (3.1) in [15]). In particular, if
we can find θ ⋆ and Σ⋆ such that
E0 [L2θ⋆ ,Σ⋆ (Y )] ≤ Cα,β
for some β ∈]0, 1 − α[, then
inf
sup
Ψα θ∈Θa (rǫ )
ξ∈Ξ
Pθ,ξ (Ψα = 0) ≥ β.
Let D ∈ N be fixed. Now, we impose the following conditions on θ ⋆ and Σ⋆ :
θk⋆ = 0
∀k > D
and
Σ⋆kl = 0 ∀k > D, l > D, k 6= l.
Let Σ⋆D = (Σ⋆kl )1≤k,l≤D be the remaining submatrix of Σ⋆ . With a slight abuse of notation, we
⋆ )′ . Then, by simple algebra,
denote below Y = (Y1 , . . . , YD )′ and bθ ⋆ = (b1 θ1⋆ , . . . , bD θD
exp − 2ε12 (Y − bθ ⋆ )′ (Σ⋆D )−1 (Y − bθ ⋆ )
Lθ⋆ ,Σ⋆ (Y ) =
exp − 2ε12 Y ′ (Σ⋆D )−1 Y
1
⋆ ′
⋆ −1
⋆ ′
⋆ −1 ⋆
= exp
2(bθ
)
(Σ
)
Y
−
(bθ
)
(Σ
)
bθ
.
D
D
2ε2
Hence,
E0 [L2θ⋆ ,Σ⋆ (Y
2
1
⋆ ′
⋆ −1 ⋆
⋆ ′
⋆ −1
.
)] = exp − 2 (bθ ) (ΣD ) bθ E0 exp 2 (bθ ) (ΣD ) Y
ε
ε
14
It is easily seen that
2
⋆ ′
⋆ −1
E0 exp 2 (bθ ) (ΣD ) Y
ε
Z
1 ′ ⋆ −1
2
1
⋆ ′
⋆ −1
exp 2 (bθ ) (ΣD ) y exp − 2 y (ΣD ) y dy
=
ε
2ǫ
(2πε2 )D/2 |Σ⋆D |1/2 RD
Z
1
1 ′ ⋆ −1
⋆ ′
⋆ −1
=
dy
exp − 2 y (ΣD ) y − 4(bθ ) (ΣD ) y
2ǫ
(2πε2 )D/2 |Σ⋆D |1/2 RD
Z
2
1
1
⋆ ′
⋆ −1 ⋆
⋆ ′
⋆ −1
⋆
= exp 2 (bθ ) (ΣD ) bθ
exp − 2 (y − 2bθ ) (ΣD ) (y − 2bθ ) dy
ε
2ǫ
(2πε2 )D/2 |Σ⋆D |1/2 RD
2
= exp 2 (bθ ⋆ )′ (Σ⋆D )−1 bθ ⋆ .
ε
Hence,
E0 [L2θ⋆ ,Σ⋆ (Y
)] = exp
Now, we select θ ⋆ as follows
θk⋆ =
1
⋆ ′
⋆ −1 ⋆
(bθ ) (ΣD ) bθ .
ε2
⋆
rε b−1
k (ΣD v)k
∀k ∈ {1, . . . , D} and
ρ
where
2
ρ =
D
X
⋆
2
b−2
k (ΣD v)k
k=1
θk⋆ = 0 ∀k > D,
1
and vk = √ ∀k ∈ {1, . . . , D}.
D
Now, define v = (v1 , . . . , vD )′ and note that kvk = 1. Then, it is easily seen that
2
rε ′ ⋆
2
E0 [Lθ⋆ ,Σ⋆ (Y )] = exp 2 2 v ΣD v .
ε ρ
We first construct a specific ξ ⋆ ∈ ΞG . Let
q
ξk⋆ = dk η0 + 1 − d2k ηk ∀k ∈ {1, . . . , D} and
ξk⋆ = ηk ∀k > D,
(7.12)
(7.13)
(7.14)
where (ηk )k∈N0 denotes a sequence of independent
√ standard Gaussian random variables and
d = (dk )1≤k≤D is a real sequence such that 1/ 2 ≤ dk < 1 for all 1 ≤ k ≤ D. Obviously,
ξ ⋆ ∈ ΞG , since ξ ⋆ is Gaussian,
E[ξk⋆ ] = 0, E[(ξk⋆ )2 ] = 1
and
∀k ∈ N
max E[(ξk⋆ )4 ] ≤ max 8d4k E[ξ04 ] + 8(1 − dk )2 E[ηk4 ] ≤ 16E[ξ04 ] = C < +∞,
1≤k≤D
1≤k≤D
sup E[(ξk⋆ )4 ] = 3.
k>D
Now, we need to bound the expression in (7.13). Using (7.14), we get
Σ⋆kl = E[ξk⋆ ξl⋆ ] = dk dl ≥ 1/2 ∀k, l ∈ {1, . . . , D} with k 6= l
Note also that, since v is a unit vector,
v ′ Σ⋆D v ≤ max a′ Σ⋆D a ≤ D
kak=1
15
and Σkl = 0 ∀k, l > D, k 6= l.
since the largest eigenvalue of Σ⋆D is smaller than D. Now, using (7.12), we get
ρ2 =
D
X
⋆
2
b−2
k (ΣD v)k =
k=1
D
D
X
X
k=1
Σ⋆kl vl
l=1
!2
since Σ⋆kl ≥ 1/2 for all k, l ∈ {1, . . . , D}. Hence,
E0 [L2θ⋆ ,Σ⋆ (Y
4rε2
1
PD −2
ε2
k=1 bk
)] ≤ exp
provided
2
kθk =
rε2
≤
D
1 X −2
bk
≥ D
4
k=1
!
≤ Cα,β
X
D
1
ln(Cα,β ) ε2
b−2
k .
4
k=1
To conclude the proof, we need to ensure that the constructed θ ⋆ belongs to Ea . Remark that,
since a is an increasing sequence,
X
a2k (θk⋆ )2
=
k∈N
D
X
a2k (θk⋆ )2
inf
sup
k=1
provided rε2 ≤ a−2
D . Hence,
Ψα θ∈Θa (rε )
ξ∈Ξ
as soon as
rε2 ≤
≤
a2D
D
X
k=1
a2k (θk⋆ )2 ≤ a2D rε2 ≤ 1
Pθ,ξ (Ψα = 0) ≥ β,
X
D
1
−2
ln(Cα,β ) ε2
b−2
k ∧ aD .
4
k=1
7.4
Proof of Theorem 4.1
In Proposition 3.2, we have proved that for all D ∈ N , there exists an α-level test Ψα,D such
that
sup Pθ,ξ (Ψα,D = 0) ≤ β,
θ∈Θa (rε,D )
ξ∈Ξ
for all radius rε,D > 0 satisfying
2
rε,D
≥ Cβ ε
D
X
"
2
2
Now, setting
†
D = arg inf
D∈N
−2
b−2
k + aD .
k=1
Cβ ε
D
X
b−2
k
+
a−2
D
k=1
#
and denoting by Ψα,D† the associated α-level test, we get
sup
θ∈Θa (rε⋆ )
ξ∈Ξ
Pθ,ξ (Ψα,D† = 0) ≤ β,
for all radius rε⋆ > 0 satisfying
(rε⋆ )2
≥ inf
D∈N
"
2
Cβ ε
16
D
X
k=1
b−2
k
+
a−2
D
#
.
Hence,
r̃ε2 ≤ inf
D∈N
"
Cβ ε2
Similarly, using Proposition 3.3,
sup
inf
Ψα θ∈Θa (rε )
ξ∈Ξ
D
X
#
−2
.
b−2
k + aD
k=1
Pθ,ξ (Ψα = 0) ≥ β,
for all radius rε > 0 such that
rε2
≤
X
D
1
−2
b−2
ln(Cα,β ) ε2
k ∧ aD .
4
k=1
This results occurs for all D ∈ N . Hence,
sup
inf
Ψα θ∈Θa (rε )
ξ∈Ξ
Pθ,ξ (Ψα = 0) ≥ β,
for all radius rε > 0 such that
rε2 ≤ sup
"
#
X
D
1
−2
ln(Cα,β ) ε2
.
b−2
k ∧ aD
4
r̃ε2
"
#
X
D
1
−2
−2
2
bk ∧ a D .
ln(Cα,β ) ε
4
D∈N
This entails that
≥ sup
D∈N
k=1
k=1
7.5
Proof of Proposition 5.1
Remark that, under Assumption HD ,
S0 (D)
ε4
∼
D
X
k,l=1
k6=l
= 2
≤ 2
Then, for all k ∈ N ,
k−1
X
l=1
k2t l2t |k − l|−2s
D
X
k=1
D
X
k
2t
k4t
k=1
k−1
X
l2t
(k − l)2s
l=1
1
.
(k − l)2s
l=1
k−1
X
k−1
X
1
1
=
∼ k1−2s .
2s
2s
(k − l)
m
m=1
In particular, the above sum is finite whatever the value of k provided s > 1/2. Hence, under
Assumption HD , we get
S0 (D) . ε4
D
X
k=1
k4t k1−2s ∼ ε4 D 4t−2s+2 = o(R0 (D))
as D → +∞.
17
7.6
Proof of Proposition 5.2
Recall from (7.2) that
Varθ,ξ (TD ) = Rθ (D) + Sθ (D),
where, using (7.5),
Rθ (D) =
D
X
2
2
b−4
k Varθ,ξ (yk − ε )
k=1
. ε4
D
X
−2
2
b−4
k + ε ( max bk )
1≤k≤D
k=1
Moreover, using (7.6),
Sθ (D)
D
X
=
k,l=1
k6=l
=
2ε
θk2 .
k=1
−2
2 2
2
2
b−2
k bl Covθ,ξ yk − ε , yl − ε
D
X
4
D
X
k,l=1
k6=l
2
+8ε
2
2
−2
b−2
k bl E[(ξk − 1)(ξl − 1)]
D
X
−1
b−1
k bl θk θl E[ξk ξl ]
D
X
−1
2
b−2
k bl θl E[(ξk − 1)ξl ]
D
X
−1
2
b−2
l bk θk E[(ξl − 1)ξk ]
k,l=1
k6=l
+4ε3
k,l=1
k6=l
+4ε3
k,l=1
k6=l
:= R1 + R2 + R3 + R4 .
Note that, using the above proposition, R1 = S0 (D) = o(Rθ (D)) as D → +∞. Then, using
(5.2), we can immediately see that R3 = R4 = 0. In order to conclude, using the Cauchy-Schwarz
and Peter-Paul inequalities, we get, for any γ ∈]0, 1[,
2
R2 = 8ε
D
X
−1
b−1
k bl θk θl Cov(ξk , ξl )
k,l=1
k6=l
. ε2
D
X
k,l=1
k6=l
−1
b−1
k bl |θk ||θl |
1
|k − l|s
v
v
u D
u
D
uX
u X
1
u
u 4
−2 −2
bk bl
≤ uε
θk2 θl2
u
|k − l|2s t
t
k,l=1
k6=l
≤ γ
D
X
∼ γ
D
X
θk2
k=1
k=1
θk2
k,l=1
k6=l
!2
+γ
!2
+ γ −1 R1 .
−1 4
ε
D
X
k,l=1
k6=l
18
−2
b−2
k bl
1
|k − l|2s
Summarizing all the above computations, we obtain, for any γ ∈]0, 1[,
Varθ,ξ (TD ) . (1 + γ −1 )ε4
D
X
−2
2
b−4
k + ε ( max bk )
1≤k≤D
k=1
. (1 + γ −1 )ε4
D
X
b−4
k +γ
D
X
k=1
k=1
θk2
D
X
θk2 + γ
k=1
!2
D
X
k=1
θk2
!2
,
where we have used again the Peter-Paul inequality (see, e.g., [3], p. 18).
References
[1] Y. Baraud. Non-asymptotic minimax rates of testing in signal detection.
8(5):577–606, 2002.
Bernoulli,
[2] N. Bissantz, G. Claeskens, H. Holzmann, and A. Munk. Testing for lack of fit in inverse
regression—with applications to biophotonic imaging. Journal of the Royal Statistical Society, Series B,, 71(1):25–48, 2009.
[3] R.E. Castillo and H. Rafeiro. An Introductory Course in Lebesgue Spaces. CMS Books in
Mathematics/Ouvrages de Mathématiques de la SMC. Springer, 2016.
[4] L. Cavalier. Estimation in a problem of fractional integration. Inverse Problems, 20(5):1445–
1454, 2004.
[5] L. Cavalier. Inverse problems in statistics. In Inverse Problems and High-Dimensional
Estimation, volume 203 of Lect. Notes Stat. Proc., pages 3–96. Springer, Heidelberg, 2011.
[6] Yu.I. Ingster, T. Sapatinas, and I.A. Suslina. Minimax nonparametric testing in a problem
related to the Radon transform. Mathematical Methods of Statistics, 20(4):347–364, 2011.
[7] Yu.I. Ingster, T. Sapatinas, and I.A. Suslina. Minimax signal detection in ill-posed inverse
problems. Annals of Statistics, 40:1524–1549, 2012.
[8] Yu.I. Ingster and I.A. Suslina. Nonparametric Goodness-of-Fit Testing Under Gaussian
Models, volume 169 of Lecture Notes in Statistics. Springer-Verlag, New York, 2003.
[9] L. Isserlis. On a formula for the product-moment coefficient of any order of a normal
frequency distribution in any number of variables. Biometrika, 12(1/2):134–139, 1918.
[10] I.M. Johnstone. Wavelet shrinkage for correlated data and inverse problems: adaptivity
results. Statistica Sinica, 9(1):51–83, 1999.
[11] B. Laurent, J.-M. Loubes, and C. Marteau. Testing inverse problems: a direct or an indirect
problem? Journal of Statistical Planning and Inference, 141(5):1849–1861, 2011.
[12] B. Laurent, J.-M. Loubes, and C. Marteau. Non asymptotic minimax rates of testing in
signal detection with heterogeneous variances. Electronic Journal of Statistics, 6:91–122,
2012.
[13] S. Mallat. A Wavelet Tour of Signal Processing. 2nd Edition. Academic Press, San Diego,
1999.
19
[14] C. Marteau and P. Mathé. General regularization schemes for signal detection in inverse
problems. Mathematical Methods of Statistics, 23(3):176–200, 2014.
[15] C. Marteau and T. Sapatinas. A unified treatment for non-asymptotic and asymptotic
approaches to minimax signal detection. Statistics Surveys, 9:253–297, 2015.
[16] A.B. Tsybakov. Introduction to Nonparametric Estimation. Springer Series in Statistics.
Springer, New York, 2009. Revised and extended from the 2004 French original, Translated
by Vladimir Zaiats.
20
| 10math.ST
|
Efficient and Robust Pedestrian Detection
using Deep Learning for Human-Aware Navigation
André Mateus, David Ribeiro, Pedro Miraldo, and Jacinto C. Nascimento
arXiv:1607.04441v2 [cs.RO] 28 Sep 2017
Instituto de Sistemas e Robótica (LARSyS),
Instituto Superior Técnico, Lisboa,
Torre Norte - 6 Piso Av.Rovisco Pais, 1 1049-001 Lisboa, Portugal.
Corresponding author: [email protected]
Abstract
This paper addresses the problem of Human-Aware Navigation (HAN), using multi camera sensors to implement a vision-based
person tracking system. The main contributions of this paper are as follows: a novel and efficient Deep Learning person detection
and a standardization of human-aware constraints. In the first stage of the approach, we propose to cascade the Aggregate Channel
Features (ACF) detector with a deep Convolutional Neural Network (CNN) to achieve fast and accurate Pedestrian Detection
(PD). Regarding the human awareness (that can be defined as constraints associated with the robot’s motion), we use a mixture
of asymmetric Gaussian functions, to define the cost functions associated to each constraint. Both methods proposed herein are
evaluated individually to measure the impact of each of the components. The final solution (including both the proposed pedestrian
detection and the human-aware constraints) is tested in a typical domestic indoor scenario, in four distinct experiments. The results
show that the robot is able to cope with human-aware constraints, defined after common proxemics and social rules.
Keywords: Pedestrian Detection, Convolutional Neural Network, Human-Aware Navigation
1. Introduction
For robots to interact naturally with humans in their social
environments, they must have the ability to plan their motion
accounting for typical social norms. In this paper, we address
the robot navigation in the presence of humans, resorting to
multi cameras (static outside and/or onboard cameras) for the
vision-based person tracking system.
One of the research focus in robotics is Human-Robot Interaction and its role in social environments. When people think of
a robot interacting with a person, what comes to mind is a robot
that can speak with her or hand over some object. However,
the motion itself is of great importance in a social context (e.g.,
when a robot is requested to fetch an item), or simply when
a normal navigation behavior needs to be adjusted according
to proxemics rules, so it does not disturb people. The study
of robot navigation in the presence of people is called HumanAware Navigation (HAN).
Most approaches to HAN in the literature use only sensors
onboard the robot. Even though those sensors bring the advantage of context-independence, it is useful to have other external
sensors, which can add more information about the environment, not only in terms of coverage space, but also in terms
of precision on the estimation of the person’s position. Hence,
in this paper, in addition to the onboard camera, external cameras were mounted on the ceiling and used for the pedestrian
detection (PD) task. This setup ensures a broader perception of
the environment, capable of seeing/detecting both humans and
robots at the same time.
Preprint submitted to Elsevier Robotics and Autonomous Systems
Furthermore, in this paper we address the importance of incorporating deep learning (DL) in an HAN based architecture.
The integration of DL provides both: (i) efficiency and (ii) robustness to the pedestrian detection task, as detailed next.
Traditionally, the detection task is usually accomplished
through the sliding window paradigm, based on a exhaustive search over the image to find the object positions (e.g.,
[1, 2, 3, 4]). During this search, features for each window location are extracted and computed (possibly at multiple scales),
and later evaluated by a classifier. However, such procedure can
easily become intractable due to the substantial number of windows and the complexity of the features processing. Thus, for
the use of richer and more expensive features, it is necessary to
constrain the computations to a restricted number of locations,
considering only certain regions of interest.
Accordingly, we propose a cascade of the ACF detector
[5] with a CNN for PD. The ACF allows to obtain a selective search that identifies promising image regions for the presence of pedestrians (i.e. proposals). This alleviates the CNN
task, since the number of windows (i.e. proposals) to classify is
now substantially reduced. The advantages of this cascade are
twofold. Firstly, we only use the more expensive computations
(i.e. the CNN classification) in the promising ACF proposals.
Consequently, operating under this regions of interest paradigm
[6, 7], allows to speed up the detection procedure and perform
pedestrian detection at real time requirements. The methodology previously described, contributes to the efficiency goal.
Secondly, the first stage of the cascade (i.e. the ACF) provides
October 24, 2017
a large number of false positives (FP). This number of FPs is
drastically reduced by the application of the CNN, while maintaining most of the true positives (TP). This means that the errors corresponding to these FPs provided by the ACF are solved
within the CNN processing. This approach allows to achieve
the robustness goal.
The CNN model results from first, pre-training with a large
object dataset [8], and then fine-tuning (i.e. re-training) with a
PD dataset [1]. This transfer learning procedure [9] is adopted
because it improves the final model. In order to guanratee independence from the particular application in the HAN context,
no further fine-tuning is performed (although the CNN could be
fine-tuned again with a specific HAN dataset).
Finally, we also propose a novel solution for HAN resorting
to multiple cameras (onboard and offboard), for people state
estimation, coupled with the aforementioned DL strategy. A
costmap is computed by combining several constraints associated with HAN. Each time the robot receives a new goal, it computes a path on that costmap. When compared with state-ofthe-art approaches, the main contributions presented in this paper are: (i) a novel and efficient technique for people detection;
and (ii) a standardization of human-aware constraints. Moreover, since all the computations are performed using CPU, this
methodology is an economically viable solution, that reaches
the intended runtime figures and accuracy requirements. Therefore, the mentioned approach can be integrated in robots that
only have onboard CPUs (no GPUs).
The solution is tested in simulated and in realistic scenarios.
The results show that the proposed solution fulfils the aforementioned goals.
This paper proposes novel extensions of the authors’ previous works, namely [10, 11]. More specifically, in [10] HAN
constraints are introduced, and in [11] a real-time PD algorithm is proposed. Comparing the present paper with [11], additional outside cameras are used, more HAN constraints are
introduced and tested, and additional details are included (e.g.
in the method’s evaluation and the related work).
classifier. The Locally Decorrelated Channel Features (LDCF)
[14] is a more accurate variant of ACF (but also slower), which
decorrelates the previously mentioned image channels resorting to linear filters. This idea of adding a filtering step for the
features, was further studied in the work of [15].
Nevertheless, the application of deep compositional architectures, namely, Convolutional Neural Networks, to the tasks
of image classification, localization and detection [8], have significantly boosted the state-of-the-art. As such, the application
of deep learning to PD arises as a natural forthcoming step. Indeed, deep learning based architectures learn hierarchical features [16] that make it possible to reach a better classification
performance than using the handcrafted ones.
To avoid overfitting during the training of CNN models, the
dataset must be of substantial size and contain the corresponding annotations. This is, however, an apparent limitation since
this problem can be addressed by transferring parameters from
an already trained CNN model (with datasets belonging to other
tasks, for example, generic object classification) to the model of
interest. This CNN concerning the model of interest, is then retrained (i.e. fine-tuned) with the dataset corresponding to the
specific problem.
The computations associated with the CNN are expensive
when compared to the ones required by methods using handcrafted features. Therefore, to improve the detector’s speed, an
hybrid solution can be adopted by cascading a faster and shallower method, based on handcrafted features, with a deep CNN.
The handcrafted approach generates proposals (i.e., promising
regions for the pedestrians locations), whose classification is refined by the CNN (i.e., the accuracy is enhanced by removing
false positives).
Currently, various PD deep learning approaches have been
developed in the literature, mainly by extending the main
pipeline of successful object detectors, such as: R-CNN [6],
Fast R-CNN [17] and Faster R-CNN [18]. This pipeline comprises a combination of: (i) proposal extraction and (ii) CNN
evaluation (being identical to the hybrid scheme, but more
generic). For example, in [19], the ACF is used in conjunction with the Fast R-CNN architecture, which has two CNN
branches for small and large scales. The work of [20] is able to
improve the proposal extraction module of the Faster R-CNN
scheme, resorting to diverse network outputs in order to detect
objects of different scales. The work presented in [21] uses the
region proposal network (CNN based) from Faster R-CNN and
refines the detections resorting to CNN features and boosted
decision trees.
In our approach, we adopt a similar architecture to [22, 23],
which follows the R-CNN pipeline. However, we do not apply
any bounding box regression, contrasting to [22]. Besides, and
since we are concerned with speed, we introduce an ACF score
rejection threshold. Below this value, the proposals are eliminated and are not processed by the CNN, allowing to achieve
faster running time figures. Our main focus is devoted to the integration of the Computer Vision module (PD method) with the
HAN module, in order to achieve an accurate and fast enough
system.
1.1. State-of-the-Art
To give a better comprehensive understanding of the contributions presented herein, we next review the related work on
both PD and HAN, respectively.
1.1.1. Pedestrian Detection:
One of the main goals in our framework is to achieve an accurate and fast Pedestrian Detection algorithm. This has been
one of the major topics addressed in the computer vision community, surveys are available in [12, 13].
Classically, the PD problem has been addressed by using
conventional handcrafted features (e.g., image gradient, HOG,
wavelets, etc) that have plateaued in recent years. Some of the
popular handcrafted methods for PD include: the Aggregate
Channel Features (ACF) [5], where individual pixel lookups are
extracted from the concatenation of the LUV, histogram of oriented gradients and gradient magnitude image channels. These
pixels serve as features to be applied in a boosted decision trees
2
1.1.2. Human-Aware Navigation:
Regarding the HAN, a planner approach can be found in
[24]. This work focuses on human comfort, which is addressed
by three criteria: preventing personal space invasions; navigating in the humans’ field of view (FOV); and preventing sudden
appearances in the FOV of humans. Those criteria are modeled as cost functions in a 2D costmap and path planning is
performed with A? algorithm. Even though the HAN planner
accounts for replanning if people move, it does not adapt their
personal space during the motion. With that in mind, two extensions to HAN are proposed:
• A prediction cost function which, by increasing the cost
in front of a moving human, decreases the probability of
the robot entering that area [25]; and
• The concept of compatible paths, which means that two
paths are compatible if both agents can follow their paths
(reaching the goal position), without any deadlocks [26].
Figure 1: Illustration of the proposed methodology, cascading the ACF (nondeep) detector and the deep CNN. First, the ACF detector performs selective
search identifying promising image regions that might contain pedestrians, i.e.
generates pedestrian proposals (see the green rectangles in the third image of
the figure). The candidate proposals, in the RGB feature map, are forwarded
through the CNN to be more accurately classified (see text).
An alternative approach was proposed by [27], which differs from HAN planner on the considered constraints and their
formulation. Instead of focusing simply on human comfort,
constraints concerning social rules (e.g., navigate on the right
side of narrow passages) and low-level human navigation behavior (e.g., face direction of movement) are also taken into
account. Another important issue related with human comfort,
in a social context, is the interference with humans interacting
with other humans and/or objects. This issue is tackled by [28]
where, besides considering proxemics and the back space of a
person, a constraint is included to model the space between interacting entities. Other important work was presented in [29].
The authors presented a framework for planning a smooth path
through a set of milestones. Those are added, deleted, and/or
modified, based on the static and dynamic components of the
environment.
More recently, [30] defined the three goals for HumanAware Navigation: human comfort (e.g., space that people keep
from each other in different contexts, known as the theory of
proxemics [31], and velocity that robots navigate close to humans [32]); respect social rules; and mimic low-level human
behavior.
present the results with the complete framework (PD + HAN).
Section 8 concludes the paper.
2. Vision-Based People Detection through Deep Learning
For PD, this paper follows the strategy mentioned in the
previous section, providing the following contributions. First,
we adopt a very deep learning based approach, by using pretrained models. Second, we are able to drastically decrease
the computational effort associated with the exhaustive search
performed during the sliding window process. To accomplish
this, we cascade the ACF (non-deep) detector [5] with a CNN1 .
This cascade strategy is twofold: first, it provides a selective
search approach that significantly improves the computational
efficiency, since only the output proposals of the ACF are taken
into account; then, by cascading with the CNN, we are able to
boost the performance of the ACF detector (i.e., improve the
classification accuracy of the ACF proposals by reducing the
number of false positives). Fig. 1 illustrates the proposed approach for the PD task.
1.2. Outline of the Paper
This paper is organized as follows. Section 2 introduces the
main stages of the proposed vision-based framework. Section
2.1 describes the PD methodology. Section 2.1.1 details the
CNN model and operations used for PD. Section 3 is related
with how the deep learning methodology is integrated in the
navigation setup. To accomplish this, the CNN training must
be performed (Section 3.1), as well as the adaptation of the
CNN (Section 3.2). Experimental evaluation is conducted in
Section 4, in which, we evaluate the performance of the proposed PD methodology in the INRIA dataset (Section 4.1) and
in two real scenarios comprising the “corridor” and “MBOT”
sequences (Section 4.2). Section 5 presents the HAN constraints that will be used in the proposed framework. In Section 6, we evaluate the HAN constraints, and in Section 7, we
2.1. Methodology for PD
In this section we formalize the adopted methodology for
PD. First, let us consider that we have available the following
|D|
training set D = {(x, y)i }i=1 , where x denotes the input image
3
with x : Ω → R with Ω denoting the image lattice2 of size
w × h × d, with d = 3; the class label is defined in y ∈ Y =
{0, 1}C that denotes the (absence) presence of the pedestrian in
the ith image xi (i.e., C = 2). The training dataset D is the input
for the ACF detector. For each input image xi , a detector (e.g.,
1 Recall
that there exist several non-deep detectors, such as Regionlets [33]
or LDCF [14]. We choose the ACF detector since it is fast.
2 In this paper, the RGB feature map is considered for the image x.
3
and bias βout ∈ RC (C is the number of classes under consideration).
The output of the CNN mentioned in (1), can be seen as an
approximation of the input data (represented by v? in equation
(1)). The convolution mentioned above is formally defined as:
ACF or LDCF) is used to provide the candidate windows, or
the proposals, along with the scores (confidences). This can be
|O|
formalized as the following output set O = {(x(B), S )i }i=1 .
|B|
In this set, B = {bk }k=1 represents the set of bounding boxes
coordinates, with bk = [xk , yk , wk , hk ] ∈ R4 denoting the top-left
point and width and height enclosing (or not) the pedestrian; we
denote as x(B), the content (i.e., the proposals) of the image
|S |
delimited by the bounding boxes B; S = {sk }k=1 are the ACF
detector confidence scores assigned to the proposals x(B).
As mentioned in [9], the generalization ability of the CNN
can be boosted resorting to pre-trained models, instead of using
random initialization. Therefore, we use the proposed VGG
CNN model [34]3 , pre-trained with Imagenet [8]. Formally,
e
|D|
e = {(e
we have a dataset to pre-train the CNN, i.e. D
x,e
y)n }n=1 ,
f = {0, 1}Ce, where Ce is the number
with e
x : Ω → R3 and e
y∈Y
of classes in the pre-trained model (in the Imagenet case, the
number of classes is Ce = 1000). In Section 3.2.2 we detail how
transfer learning is accomplished in our classification problem.
cl (v(l−1) ( j)) =
∑
v(l−1) (i) ? Wl (i, j) + β l ( j),
(2)
i∈Ω( j)
where ? stands for the convolution, Ω( j) is the input region
addresses and where the convolutional filters are represented
by the weight matrix Wl and the bias vector β l . Notice that
the input v(l−1) ( j) in (2), is obtained following the structure in
(1), i.e. convolution, activation and sub-sampling operations 4 ,
from the preceding layer, that is:
v(l−1) ( j) = sl−1 al−1 (cl−1 (v(l−2) ( j))) ,
(3)
where v(0) ( j) represents the input image.
The L convolutional layers are followed by a sequence of
fully connected layers, that perform a particular instance of
L
the convolution in (2) to the entire vectorised input vL ∈ R|v | ,
where |vL | denotes the length-vector vL . In the final stage,
the above fully connected layers are followed by a classification layer that is defined by a soft-max function as follows (see
[35]):
fout ( f fc ) = softmax(Wout f fc + βout ),
(4)
2.1.1. CNN model:
We now formalize the main ingredients of the CNN architecture. Basically, this type of deep networks comprises several
processing stages. Each stage is characterized by having two
types of layers, namely: a convolutional layer containing a nonlinear activation function, and a non-linear subsampling layer.
In the former, a convolutional filter is applied to the input. In
the latter, the size reduction of the input is achieved. These two
stages are typically followed by several fully connected layers,
and a multinomial logistic regression layer (see details in [35]).
Formally, the convolutional neural network can be analytically
represented by the following mapping f : X → Y , where X
represents the image space and Y represents the classification
space:
with the soft-max function defined as y(q)i =
exp(q(i))
∑ j exp(q( j))
and
[0, 1]C
fout ∈
represents the output from the inference process
that takes the input v, with C representing the number of classes.
In our case, the input are the proposals of the ACF non-deep
detector, i.e. x(B), and C represents the two output classes
(absence/presence of the pedestrian). Thus, (1) is written as
(similarly for (2)):
f (v; θ ) = v? = fout ◦ ffc ◦ sL ◦ aL ◦ cL ◦ ... ◦ s1 ◦ a1 ◦ c1 (v(0) ), (1)
f (x(B); θ ) = x(B)?
where ◦ denotes the composition operator, {ci (.)}Li=1
represents
a convolutional layer, θ represents the model parameters comprising the input weight matrices Wl ∈ Rkl ×kl ×nl ×nl−1
and bias vector βl ∈ Rnl for each layer l ∈ {1, ..., L}, and
with kl × kl representing the size of the nl filters in the l-th
layer, having nl−1 input channels ; al (.) represents a non-linear
activation layer (e.g. the Rectified Linear Unit (ReLU),
for more details see [35]) ; sl (.) is a sub-sampling layer,
that is, a function that allows to obtain vl =↓ v(l−1) , where
↓ denotes a subsampling function that pools (using the mean
or max functions) the values from a region of the input data
; ffc is a fully-connected layer containing the weights
{W fc ,k }Kk=1 (with W fc ,k ∈ Rn fc ,k−1 ×n fc ,k representing the connections between the k − 1-th and k-th fully connected layers), and biases {β fc ,k }Kk=1 (with β ∈ Rn fc ,k ), that also belong to
the model parameters θ ; fout is a multinomial logistic
regression layer containing the weights Wout ∈ Rn fc ,K ×C
= fout ◦ ffc ◦ sL ◦ aL ◦ cL ◦ ... ◦ s1 ◦ a1 ◦ c1 (x(B (0) )),
(5)
where the inputs are the proposals (i.e., the image content, in
the RGB feature map, delimited by the bounding boxes), here
denoted as x(B (0) ) (see Fig. 1). The main idea is to take the
proposals x(B), that will be processed by the CNN, and produce a classification probability that a given proposal contains
a pedestrian. The proposals classified as non pedestrians are
discarded, allowing to eliminate false positives. The ones regarded as pedestrians are kept, including the original ACF detector score.
In the PD case, the CNN prediction output can be formally
represented by:
(6)
f (x(B), θ ) = y? ,
4 These are the basic operations of the CNN, but the subsampling layer does
not necessarily need to be present in every case.
3 Details are also available at: http://www.robots.ox.ac.uk/
˜vgg/research/very_deep/.
4
which is trained using the binary cross-entropy loss over the
training set indexed by i, as follows:
training images, from which 1218 are negative images (i.e., not
containing pedestrians) and 614 are positive images (i.e., containing pedestrians). There are 288 test images. In our experimental setup, we have to build not only the training set but also
the validation set to fine-tune the pre-trained CNN model. To
obtain the positive set, we use the ground truth positive training bounding boxes (i.e., proposals corresponding to the image
content delimited by them) Bpos = 1237. Data augmentation
is then performed for the positive samples using the following
two steps:
1 |D|
∑ −y(i) × log(y? (i)) − (1 − y(i)) × log(1 − y? (i))
|D| i=1
(7)
Let the pre-trained CNN be represented by the model e
y=
f (e
x, θe), with θe = [θecn , θefc , θelr ]. The process of pre-training a
CNN is defined by the following three steps:
L =
1. Training M1 stages (see cyan/blue block in the CNN illustrated in Fig. 1) of convolutional and non-linear subsampling layers, that are represented by the parameters
θecn ; then
1. Horizontal flipping over the set Bpos , resulting in a new
(1)
set, Bpos = 2474 (including also Bpos ); and
2. Random deformations (including translation and scale),
in the range R = [0, 5] pixels (for the beginning and end)
(1)
applied on the previous set, Bpos . This allows us to ob(2)
tain a new set, Bpos = 4948.
2. Training M2 fully connected layers (see green block in
the CNN illustrated in Fig. 1), represented by the parameters θefc ; and
3. Training one multinomial logistic regression layer with
parameters (see red block in the CNN illustrated in Fig.
1) θelr , by minimizing the cross-entropy loss function [35]
e
over the dataset D.
To build the negative set Bneg , we extract negative windows
(i.e., proposals) from the negative images using the strategy
mentioned in [23] (i.e., employing a non-fully trained version
of LDCF). As a result, a set of 12552 negative windows is obtained, by defining an upper-bound of 18 negative proposals per
image. This strategy allows to acquire a total of 17500 proposals, from which 15751 are used for training (90%) and 1749 for
validation (10%).
It is worth to mention that, transferring a large number of
pre-trained layers and fine-tuning the CNN is the key to achieve
the best classification results in transfer learning problems [9].
Following this strategy, we first take the M1 layers to initialize
a new model (see [9]). Since we have changed the CNN input
size to reduce the computational expense, the M2 layers were
randomly initialized from a Gaussian distribution, so that the
dimensions are compatible and inference is possible. Finally,
we introduce a new binomial logistic regression layer, with parameters θlr (randomly initialized from a Gaussian distribution)
adapted for two classes (pedestrian and non-pedestrian). Afterwards, we fine-tune the CNN model by minimizing the crossentropy loss function in (7), using the pedestrian training set
D.
3.2. Adaptation of the pre-trained CNN model
As mentioned in Sec. 2.1, we use a pre-trained CNN model.
As such, some modifications/adaptations are necessary. In the
following, we describe the required steps, by first detailing the
model used and, then, describing the modifications needed.
3.2.1. Pre-trained model used:
For the pre-trained network, we select the VGG Very Deep
16 architecture (VGG-VD16), more specifically, the “D” configuration in [34]6 .
This CNN architecture receives 224 × 224 × 3 input images.
The network comprises 13 convolutional layers, three fully connected layers and a multinomial logistic regression layer (see
Sec. 2.1.1). There are five max-pooling operations (i.e., nonlinear subsampling), which operate on a 2x2 region, reducing
the input by a factor of two. These operations appear after layers: 2, 4, 7, 10 and 13. The Rectified Linear Unit (ReLU) is selected to be the non-linearity, applied to the convolutional and
fully connected layers. Furthermore, the size of the receptive
field is the same for all convolutional layers (i.e., 3x3). However, the number of filters is different as described below:
3. Materials and methods
This section addresses the implementation details of the
adopted deep learning methodology, to be integrated in the navigation setup. First, in Sec. 3.1 we describe the experimental
setup used to train the deep CNN. Then, Sec. 3.2 addresses
the adaptation of the pre-trained model to the PD task. The
outcome of these sections, i.e. the final CNN architecture finetuned for PD, is subsequently used for testing purposes, as described in Sec. 4.
3.1. CNN training in the INRIA dataset
To train the CNN model we use the INRIA dataset, which
is a common benchmark used for research work in the detection of pedestrians in images5 . This dataset comprises 1832
• There are 64 filters in each of the layers 1 and 2;
• There are 128 filters in each of the layers 3 and 4;
6 See additional details at: http://www.robots.ox.ac.uk/ vgg/
˜
research/very_deep/.
5 More
details can be found at: http://pascal.inrialpes.fr/
data/human/.
5
• There are 256 filters in each of the layers 5, 6 and 7;
Table 1: Logarithmic average miss rate (MR), frames per second (FPS), number of true positives (TP), false positives (FP), and false negatives (FN) on the
INRIA dataset, for the ACF detector and the ACF+CNN.
• There are 512 filters in each of the layers from 8 to 13;
• There are 4096 filters in each of the layers 14 and 15;
Dataset
• There are 1000 filters in layer 16. Each of these filters is
associated with a different ILSVRC [8] class. After this
layer, there is a soft-max function.
INRIA
Baseline
The pre-training of the mentioned model was performed with
Imagenet [8], having 1000 classes, 1.2 million training, 50
thousand validation and 100 thousand test images.
INRIA
Threshold
3.2.2. Adaptation of the deep network:
In order to decrease the computational effort required by the
VGG-VD16 model, the expected input size for the CNN was
reduced from 224 × 224 × 3 (i.e., the original) to 64 × 64 × 3.
Nevertheless, due to this change in the input size, inference is
not possible after the first fully connected layer. Moreover, the
PD task requires the use of only two classes (i.e. C = 2) in order
to represent the presence or absence of a pedestrian. Therefore,
the layer 16 and the soft-max must be adjusted accordingly.
In order to solve the aforementioned issues, we choose to randomly initialize the parameters of the three fully connected layers. More specifically, they are obtained from a Gaussian distribution, with zero mean and variance σ 2 = 0.01. The changed
network is fine-tuned using the previously mentioned (in Sec.
3.1) positive and negative sets, obtained from the INRIA pedestrian dataset. The process of pre-training the network acts as a
regularization procedure, similar to data augmentation. Regarding the fine-tuning hyperparameters, we use ε = 10 epochs, a
minibatch of ξ = 100 samples, a learning rate of β = 0.001,
and a momentum of µ = 0.9. No special effort was dedicated
to fine-tune the mentioned hyperparameters.
During test, we run the ACF detector in the 288 INRIA
test images in order to obtain the proposals (i.e., regions potentially containing pedestrians). Then, we run the proposals
through the CNN in order to classify them as pedestrians or
non-pedestrians.
All the experiments will be detailed next (see Sec. 4 and
Tabs. 1 and 2) and were obtained with Matlab, running on CPU
mode on 2.50 GHz Intel Core i7-4710 HQ with 12 GB of RAM
and 64 bit architecture. The Piotr’s Computer Vision Matlab
Toolbox [36] (2014, version 3.40) was employed to execute the
ACF method, and to perform the performance evaluation. Regarding the implementation of the CNN framework, the MatConvNet toolbox [37] was utilized.
Metrics
TP
FP
FN
MR%
FPS
TP
FP
FN
MR%
FPS
ACF Proposals
551
1284
38
16.83
15.18
531
112
58
17.48
14.81
ACF+CNN
546
249
43
15.13
3.38
528
54
61
16.48
6.62
• In Sec. 4.2, we evaluate the performance of the PD
method in real scenarios comprising two sequences
(termed herein as “corridor” and “MBOT”). Special attention is given to the details of how we achieve real-time
requirements in the detection.
4.1. Performance evaluation on the INRIA dataset
To test the accuracy of the PD module, to be integrated in
the navigation setup, we first assess its performance in the INRIA dataset. The adopted evaluation metric is the log average
miss rate (MR), as proposed in [12]. This metric is obtained
from nine values in the False Positives Per Image (FPPI) interval [10−2 , 100 ]. Since it is a miss rate, lower MR values correspond to superior performances (i.e. improved accuracies).
To demonstrate the effectiveness of the proposed cascade of
a non-deep detector (ACF) with the CNN architecture, we also
present in Tab. 1 (field ”INRIA Baseline”) the results of the
ACF alone and the results of the cascade, in order to notice the
improvement achieved. From the experiments conducted, we
achieve a log average miss rate of 16.83% for the ACF detector
alone. Cascading the ACF with the CNN we are able to reach
15.13%.7
Tab. 1 shows the number of true positives (TP), false positives (FP) and false negatives (FN), before and after the use of
the CNN. Notice that, the number of FP is significantly reduced,
while maintaining most of the TP. This indicates that the CNN
is successfully discarding FP, allowing to reach performance
improvements (and justifying the observed gain).
An important concern when building the PD algorithm, is
to verify and ensure that the runtime figures satisfy the realtime requirements. Tab. 1 shows the ACF and ACF+CNN running times in frames per second (FPS) obtained for the INRIA
dataset. From the results, we conclude that the ACF+CNN system is able to perform PD at 3.38 FPS, which should be improved. The introduction of the CNN in the pipeline after the
4. Evaluation of the People Detection
This section provides the testing results for the evaluation
of the proposed PD method. The section is divided into two
parts:
• In Sec. 4.1, we describe the performance evaluation,
concerning both accuracy and runtime figures of the PD
method, when the INRIA dataset is used; and
7 The approach proposed herein, as shown in Fig. 1, is general and can be
applied to any other non-deep detector (e.g. Regionlets [33], LDCF [14] or
Spatial Pooling [38] detectors).
6
ACF method, resulted in a significant runtime decrease from
15.18 FPS to 3.38 FPS.
To improve the mentioned runtime figures, while achieving
as much accuracy as possible, two strategies can be used: (a)
reduce the original images size, or (b) discard ACF proposals
with confidence score below a certain threshold.
In this work, the latter option (b) is adopted. Notice that,
reducing the size of the images may jeopardize the quality of
the detections. Furthermore, the confidence scores outputted by
the ACF detector constitute a relevant indicator to filter the proposals. This can be achieved by applying a threshold over the
proposals, where only the ones above a certain score are kept
and are further processed by the CNN. For our experiments, a
threshold value of 40 was set for the confidence score. This
threshold is in accordance with the value suggested in [39], in
which it is proposed a thresholding technique based on upper
and lower bounds on the confidence scores of the ACF.
Another important remark is that, the gain in speed results
from the fact that the CNN only has to classify a smaller portion of the ACF proposals, instead of all of them. Therefore,
the easier false positives should be discarded by the threshold
operation, while the harder false positives should be discarded
by the CNN. The threshold value controls the trade-off between
potential accuracy loss and speed gain.
By selecting the later case (i.e. the threshold operation),
we are able to improve the runtime of the overall detector (i.e.
ACF+CNN), when compared with the results obtained for the
baseline. These results are shown in Tab. 1 (field ”INRIA
Threshold”), where it can be seen that the ACF+CNN frame
rate increases from 3.38 FPS (baseline) to 6.62 FPS (threshold).
In order to assess if this speed metrics are suited for realtime applications, we perform further experiments in real HAN
scenarios in Sec. 4.2.
Table 2: Runtime figures before (top, ”Baseline”) and after (bottom, ”Threshold”) the threshold operation applied to the ACF proposals, when using the
overall PD method (i.e., ACF+CNN).
Dataset
Baseline
Threshold
Data seq. 1 (corridor)
Data seq. 2 (Mbot)
Total time = 0.1273 sec.
Total time = 0.2066 sec.
avg. no. det. ACF = 2.70
avg. no. det. ACF = 4.83
avg. no. det. ACF+CNN = 2.36
avg. no. det. ACF+CNN = 2.45
ACF time = 0.0326 sec.
ACF time = 0.0367 sec.
CNN time = 0.0947 sec.
CNN time = 0.17 sec.
Frame rate = 7.85 FPS
Frame rate = 4.84 FPS
Total time = 0.0961 sec.
Total time = 0.1026 sec.
avg. no. det. ACF = 1.82
avg. no. det. ACF = 1.78
avg. no. det. ACF+CNN = 1.81
avg. no. det. ACF+CNN = 1.73
ACF time = 0.0333 sec.
ACF time = 0.0381 sec.
CNN time = 0.0628 sec.
CNN time = 0.0645 sec.
Frame rate = 10.41 FPS
Frame rate = 9.74 FPS
number of detections (”avg. no. det.”) before (ACF only) and
after (ACF+CNN) the application of the CNN, and also the
CNN time for each data sequence (top, the two columns in the
field named ”Baseline”).
By applying the threshold operation (as described in Sec.
4.1), it possible to obtain approximately 10 FPS for each of
the datasets (as shown in Tab. 2, field ”Threshold”)), which is
suited for real-time applications.
5. Human-Aware Navigation
From the method described in the previous section, we get
bounding boxes on the images, representing pedestrians. Multiple instances of the proposed PD can be applied in images of
several cameras at the same time. Assuming that we have a set
of images (a single image can also be used), the goal of this section is to: firstly, project the position of the pedestrians from the
image into the world environment; fuse the information given
from different imaging sensors; and, then, define the HAN constraints to be included in any conventional path planer.
After computing the position of the pedestrian in the world
coordinate systems, including the estimation of the pedestrian’s
velocity (the methods used in this paper for this module are
shown in Sec. 5.1), the main goal of this section is to define
the robot’s path in an environment with humans that may interact with it. To be as Human-Aware as possible, three goals are
considered, namely: human comfort; respect social rules; and
naturalness. To fulfill these requirements, the following constraints are taken into account:
4.2. Performance evaluation on real scenarios
To perform experiments in real scenarios, we acquired two
indoor datasets to evaluate the PD task. Two datasets are considered in these experiments:
• The “corridor” dataset, containing 5556 images; and
• The “MBOT” dataset, containing 3966 images.
The size of the frame (i.e., the image) is 480 × 640 for both of
the sequences. Some of the results obtained in the two datasets
are shown in Fig. 2, where each bounding box has a score,
showing the confidence of containing a pedestrian.
The total time to perform PD in the dataset “corridor” (5556
frames) and “MBOT” (3966 frames), is roughly 707.27 seconds
and 839 seconds, respectively. The running time figures per
frame are shown in Tab. 2. In this table, the field “Baseline”
refers to the metrics of the final PD detector as detailed in Sec.
3.2. The algorithm reaches 7.85 FPS and 4.84 FPS, for the
”corridor” and the ”MBOT” sequences, respectively.
The runtime is longer for the ”Mbot” dataset, because the
number of ACF detections that have to be processed by the
CNN is larger, in comparison with the ”corridor” dataset. This
observation is depicted in Tab. 2, by comparing the average
1. Take least effort path (naturalness) – Sec. 5.2;
2. Keep a distance from static obstacles (naturalness) –
Sec. 5.2;
3. Respect personal spaces (human comfort) – Sec. 5.3.1;
7
(a) Pedestrian detection in the “corridor” sequence.
(b) Pedestrian detection in the “MBOT” sequence.
Figure 2: In real scenarios, there are cameras mounted on the ceiling and on the robot. To test our PD, we used two sequences of images acquired from both possible
real scenario camera locations. In Figs. (a) and (b) are shown three images of the “corridor” and “MBOT‘” sequences, respectively.
4. Avoid navigating behind sitting humans
(human comfort) – Sec. 5.3.2;
5. Create and/or remove signaled tracks.
The people tracking is performed in the world coordinate
system, instead of the image plane. For this purpose, two assumptions are considered: i) people are standing or walking upright8 ; ii) given i), a person’s feet will always be on the ground
plane and, thus, the point, which represents a person feet, is on
the line between the bottom left and right corners of the respective bounding box. Since the person is, most of the times, in
the center of the bounding box, a good estimate for the position
of the feet is in the middle point of the considered line segment. The projection is performed by transforming the selected
point with an homography computed a priori (transformation
from the image plane into the floor plane). These positions will
be associated with the targets and be used as measurements in
Kalman Filters (one for each person), which will perform the
tracking. A constant velocity motion model was considered for
the prediction step of the filter.
The association between measurements and targets is performed with the NNJPDA method [40]. This is a hard assignment method, in which only one measurement is assigned to the
target, using a Maximum A Posteriori (MAP) approach. When
the association is complete, there is a condition that checks if
the assignment distance is higher than a threshold. When that is
the case, the track and measurement are unassigned. After the
assignment, there are three cases to be considered:
5. Do not interfere with human-object interactions (human
comfort) – Sec. 5.3.3; and
6. Overtake people by the left (social rule) – Sec. 5.3.4;
The first two constraints are related to navigation problems (the
method used in this paper is described in Sec. 5.2), whereas
the remaining four constraints are about the HAN (more details
about each of these constraints are shown in Sec. 5.3).
5.1. People Tracking in the World
As explained in the previous sections, the PD module returns bounding boxes representing people in the scene. Those
are sent to this module which uses the Nearest Neighbour Joint
Probabilistic Data Association (NNJPDA), [40], and an array
of Kalman Filters to track people (one for each person). It performs the following steps:
1. Project the middle point, between the bottom left and
right corners of each bounding box, to the ground plane,
in world coordinates;
2. Predict the new state of each track;
3. Associate the measurements to existing tracks and determine if tracks need to be created or removed;
4. Update the prediction with the respective measurement;
and
8 In our experiments we also considered a seated person. Even though our
method was not designed for this type of cases, the results were satisfactory.
8
• ex and ey – are the variables that define the space around
the person; and
1. The assignment exists;
2. The track is not assigned; and/or
• px and py – represent the person position.
3. The measurement is not assigned.
A graphical representation of these parameters is shown in
Fig. 3(a). Then, the personal space of a walking person was
modelled as:
If there is a successful assignment, it proceeds to the correction step of the filter, with the assigned measurement. When a
track is not assigned, there are two possibilities: the track increases the inactivity flag and, if the inactivity threshold was
reached, the track is deleted. If a measurement is not assigned,
after some iterations, a new Kalman Filter is created for this
measurement.
Velocity estimates (required for the HAN constraints) are
given from the Kalman filter state estimates. We are considering a person state to be her position and velocity.
2 1
g1 = asymGauss(ex , ey , px , py , α p , βe, βe, βe),
3 2
where βe = max (ν, 0.8) , (9)
α p is the person’s orientation and ν her speed. A graphical representation of the person walking along y–axis direction with a
velocity of 1 [m/s] is presented in Fig. 3(a).
Regarding a walking person, it makes sense for the personal
space in front to be larger than in the back (to ensure that the
robot does not pass in front of the person, decreasing the risk
of collision). On the other hand, if a person is standing and we
consider the personal space defined using the previous formulation, the robot may pass behind too close to the person, causing
discomfort. Thus, for this case we propose that the personal
space to be modelled as a circular Gaussian:
!
(ex − px )2 (ey − py )2
−
,
(10)
g2 = exp −
2βx2
2βy2
5.2. Path planner and obstacle avoidance
The first HAN constraint is addressed by the path planner
(see the list in the beginning of Sec. 5). In this work the A∗
algorithm [41] was used, ensuring a minimum cost path as long
as the heuristic is admissible. The total cost of a node is given
by the sum of the cost of reaching that node, with the heuristic
cost. The latter was considered to be the Euclidean distance
from the initial to the goal position. Since the environment is
dynamic (people may appear walking in the scene), the planner
computes a path periodically.
The goal of the second constraint is to prevent the robot
from passing too close to obstacles. This is solved by attributing
a high cost to the area surrounding the obstacles, [42].
In the next subsection, we define the remaining constraints,
that will be included in the path planner, for the navigation to
be human-aware.
where βx and βy are the standard deviation in the x and y direction respectively. This formulation was also considered for
a seated person.
If an object hand over is required, the robot should be able
to enter the personal space, to be at “arm’s length”. However,
the robot cannot be allowed to enter from a random direction,
instead it should only be allowed to approach a person from the
front [43]. Thus, our solution is to open the region in front of
the person 45 degrees, to a distance of 0.6[m]. Personal space,
in a hand over scenario, is depicted in Figure 3(b).
5.3. Human-Aware Navigation Cost Functions
The Human-Aware Cost Functions used in this work are
based on previous state-of-the-art approaches. However, some
of them are reformulated (namely the cost functions associated
with constraints 4 and 5) in order to be better integrated in our
approach, and to standardize their formulation.
5.3.2. Visibility Constraint:
Constraint 4 concerns preventing discomfort from passing
behind a seated person. We reformulated this problem with an
asymmetric Gaussian:
5.3.1. Personal Space Cost Functions:
The third constraint in the list accounts for personal space,
which models will be presented next. We consider three different situations: when a person is standing; walking; or seated.
For the case of a walking person, we used the formulation proposed in [27], which the authors call asymmetric Gaussian:
g = asymGauss(ex , ey , px , py , α, βh , βs , βr ),
g3 = asymGauss(ex , ey , px , py , α p − π, 1.2, 0.8, 0.006). (11)
5.3.3. Interaction Constraint:
The fifth constraint prevents the robot from interfering with
a person interacting with an object. It is represented by an interaction set modelled as a circle:
γ
if (ex − pex )2 + (ey − pey )2 ≤ r
g4 =
,
(12)
0
otherwise
(8)
where
• α – orientation of the function;
• βh – variance in the α direction;
where the middle position between the interacting entities is
denoted as ( pex , pey ), and the radius r is half the distance between
the entities (only one-on-one interactions are considered). γ is
an importance factor, which varies from 0 to 1.
• βr – variance in the α − π direction;
• βs – variance in the α ± π2 direction;
9
βh
α
βs
βs
βr
(a) Personal space
(b) Object hand over
(c) Seated person
(d) Walking person
Figure 3: Representation of cost functions associated with different people postures: Fig. (a) represents the cost function for the personal space of a person walking
in the y direction at 1 [m/s]; Fig. (b) shows the cost function of a person standing, oriented in the x direction, during an object hand over; Fig. (c) represents the cost
function for the case where a person is seated; and Fig. (d) shows the total cost function of a walking person, including the social rule of overtaking her by the left
and her personal space.
(a) The robot was placed behind a person and a goal position was defined. Then the person starts walking, the robot should replan a path through the person’s left.
(b) The robot is requested to hand over an object to a person, who is across the room. However it needs to pass between a seated person and a TV. As it starts
moving, the TV is turned on. To prevent interference, the robot replans around that area. However, there is yet another person, walking behind the couch, who must
be taken into account.
Figure 4: Evaluation of the proposed navigation system using simulated environments. The environment was created using Gazebo and the results are shown in
Rviz (ROS package). Figs. (a) and (b) show sequences of images representing experiments 1 and 2 respectively (more details regarding each of the experiments are
given in the text). In all cases, it can be seen: the costmap, current and goal position, path, and the trajectory.
10
(a) Image of the robot platform used for the
experimental results in realistic scenarios.
(b) Image of Camera 1,
mounted on the ceiling.
(c) Image of Camera 2,
mounted on the ceiling.
(d) Image of Camera 3,
mounted on the ceiling.
(e) On-board robot camera’s
image.
(f) Depiction of the environment in
Rviz, showing: the robot’s position, a
pedestrian, and the positions of the cameras.
Figure 5: Representation of the setup used in the experiments in a realistic scenario. Fig. (a) shows the robot platform and Figs. (b), (c), (d), and (e) show the
images of the cameras that will be used to detect the pedestrians (as it can be seen, in these images we already show the bounding boxes identifying a person in the
environment). To conclude, Fig. (f) shows the environment (ROS Rviz package), with the position of all the cameras, the position of the robot and the pedestrian,
with the respective HAN constraint (in this case the pedestrian was standing).
The proposed system was implemented as an extension of the
ROS navigation stack, [42], and the cost functions, described
in the previous section, were implemented as plug-ins to the
costmap layered structure, [44]. The simulation environment
was Gazebo running on a machine with an Intel Core i7-3770T,
and 8GB of RAM.
The experiments performed were:
5.3.4. Overtake Constraint:
Constraint 6 represents the social rule of overtaking people
by the left (considered only for walking persons). This constraint is also represented using an asymmetric Gaussian:
g5 = asymGauss(ex , ey , px , py , α p −
π
, 1.5, 0.3, 0.0075), (13)
2
5.3.5. Cost Function Fusion:
For the three possible postures of a person (standing, seated
and walking), there are two where more than one cost function
is applied and they must be combined. Since the main goal of
the framework is to maximize the comfort of the humans, the
cost functions are combined by taking the maximum cost value
attributed to each point in space. The first case of multiple cost
functions affecting the same space, is a seated person, whose
personal space is given by:
g6 = max (g2 , g3 ) ,
Experiment 1: The robot is navigating when encounters a slow
walking person, which it must overtake. The goal is to verify if it respects constraints 3 and 6, Sec. 5.
Experiment 2: The robot is requested to hand over some object to a person across the room. As it starts moving, the
TV is turned on, and the robot replans its path. However,
another person is going across the room behind the seated
person, who must be overtaken for it to reach its goal. This
experiment was designed to evaluate the navigation module
regarding constraints 3, 4, 5, and 6, Sec. 5.
(14)
The results for these experiments are shown in Fig. 4.
Next, we present experimental results using both simulated
and realistic environments.
this cost function is depicted in Fig. 3(c). The second case concerns a walking person, where the personal space must be combined with the respective social rule (a person should be overtaken by the left):
g7 = max (g1 , g5 ) .
7. Results of the Complete Framework
(15)
In this section, we evaluate the proposed framework (using
both the proposed PD and HAN). For that purpose, we use a
MBOT mobile platform [45] (see Fig. 5(a)), in a typical indoor
work scenario, as shown in Fig. 5(f). For the PD, we are using five distinct cameras: one onboard and four cameras fixed
on the ceiling. Figures showing the images captured by the
onboard camera and three of the ceiling cameras are shown in
Figs. 5(b)-(e).
A graphical representation of this cost function is shown in
Fig. 3(d).
6. Evaluation of the Human-Aware Navigation Constraint
To evaluate the proposed constraints for HAN, two experiments were defined and tested in simulated environments.
11
For the validation, we consider the following experiments:
starts moving, the person turns on the TV and, thus, the robot
needs to replan its path, so it does not interfere with the interaction that just started. As can be seen on the second image of
Fig. 8(a) the pedestrian was partially occluded by the couch on
the onboard sensor. Thus the PD running on this sensor was not
able to detect the person in the environment. However, since we
were using another sensor (external to the robot), which was detecting the person, the robot knew that there was a person there
and, when the TV was turned on, the robot replaned the path
according with the respective human-aware constraint.
In the last experiment, we show a case where the robot is
asked to receive some object from a person. In this case, the
robot should identify the person’s position, which is given by
the proposed PD algorithm. Then, the robot should navigate
towards his position and enter in his personal space to be able
to receive the object. As it can be seen from Fig. 9, both the
proposed PD and the HAN work as expected, even in the case
where the pedestrian is not seen by the onboard sensor.
A video with these experiments is sent as supplementary
material and ROS packages for both the PD and the HAN will
be available to the community.
Experiment 1: In the first experiments, the robot must avoid
three standing people that are distributed along the main corridor of the environment, performing a slalom path to avoid
entering in their personal space;
Experiment 2: The robot is navigating when encounters a slow
walking person, which it must overtake. The goal is to verify if it respects constraints 3 and 6, Sec. 5.
Experiment 3: A person is seated on a couch, watching TV,
and the robot wants to go across the room. The goal is to
test if the robot respects constraints 4 and 5, Sec. 5.
Experiment 4: The robot navigates towards a person, to hand
over some object. The goal is to verify the modification of
the personal space, constraint 3, Sec. 5.
Regarding the HAN throughout the experiments, the robot
displayed a similar behaviour to the simulation in terms of trajectory execution. However, the parameters of the cost functions (9), (10), (11), and (13) needed to be adjusted. The values
presented previously were derived empirically, taking into account: the values in the literature; the space restrictions of the
real scenario, and our intuition of comfort distances.
In the next subsection we discuss the results obtained for
each of the experiments.
8. Conclusions
This work addresses the problem of Human-Aware Navigation for a robot in a social context. For that purpose, in this
paper we derive a robust and efficient solution for the PD, using deep learning. In addition, regarding HAN, we reformulate
some of the respective constraints in order to have a standardization of human-aware constraints.
To validate our contributions, we first use the INRIA dataset
to evaluate both the accuracy and runtime figures of the PD
method. Then, to evaluate the performance using real data,
we use two sequences of images (one was acquired using the
robot’s onboard camera and the second sequence was acquired
using an external camera). The results show that the method is
both robust and fast enough (up to 10 frames per second) for
robot navigation applications. To evaluate the standardization
of HAN constraints, we use a simulated environment.
In a realistic scenario, we test both modules in four distinct
scenarios. The results show that the robot has the correct behavior, which means that both PD and HAN are working properly.
7.1. Discussion of the experimental results
Let us start by the Experiment 1. From the results shown
in Fig. 6, firstly, one can see that the proposed PD correctly detected the pedestrians, including on the onboard sensor, which
was able to detect people that were in front of the robot. Using this information, the robot planned and executed the correct
path towards the given goal, performing a slalom path to avoid
entering in their personal space.
Regarding the second experiment, a different setup was
considered. Now there is only a person in the environment but,
when the robot starts to move, that person will start moving,
blocking the robot’s path. As it can be seen from Fig. 7, firstly,
the robot plans the path taking into account a person standing
in the environment. Then, when the person starts walking the
robot replan its path to overtake the pedestrian by the left. The
proposed PD was able to detect people from different sensors
including the onboard sensor. Since we are using multiple sensors for the detections, even when the onboard camera doesn’t
see the pedestrian, the robot continues with the right path towards the goal. From these results, one can conclude that the
robot was able to correctly detect the pedestrian in real-time and
was able to plan its path according with the defined HumanAware Navigation constraints.
When considering the third experiment, the goal was to
avoid passing in front of a person that was watching TV (a
person interacting with an object). These results are shown in
Fig. 8. Firstly, the robot plans its motion according with the
less cost path towards the given goal position, which would include passing between the person and the TV. When the robot
References
References
[1] N. Dalal, B. Triggs, Histograms of oriented gradients for human detection, IEEE Proc. Computer Vision and Pattern Recognition (CVPR).
[2] P. F. Felzenszwalb, R. B. Girshick, D. McAllester, D. Ramanan, Object
detection with discriminatively trained part based models, IEEE Proc.
Pattern Analysis and Machine Intelligence.
[3] P. Viola, M. J. Jones, Robust real-time face detection, Int’l J. Computer
Vision.
[4] L. Zhu, Y. Chen, A. Yuille, W. Freeman, Latent hierarchical structural
learning for object detection, IEEE Proc. Computer Vision and Pattern
Recognition (CVPR).
[5] P. Dollar, R. Appel, S. Belongie, P. Perona, Fast feature pyramids for
object detection, IEEE Trans. Pattern Analysis and Machine Intelligence.
12
(a)
(b)
(c)
(d)
(e)
(f)
Figure 6: In this experiment, people are standing during the robot’s movement. From top to bottom, we show: 1) the Rviz representation of the environment, people
& robot positions, and the path planned at each instant; 2) detection from the onboard camera sensor; and 3) three detections from three different cameras mounted
on the ceiling. From Figs. (a)-(f) one can see the robot navigation going towards a goal position and avoiding people, while respecting the personal space constraint.
13
(a)
(b)
(d)
(c)
(e)
(f)
Figure 7: Contrarily to the previous experiment, in this case only a person is in the environment and, instead of standing, he starts moving, blocking the robot’s
path. In this figure we show the environment and images taken from the cameras, similar to Fig. 6. From Figs. (a)-(f) we can see that firstly the person is identified
as standing, but when the robot starts moving, he also starts moving, thus the robot replans its path, in order to overtake the pedestrian by the left. Notice that, since
we are using multiple sensors for the PD, even when the person is outside the onboard camera’s FOV, the robot performs well because the pedestrian is identified
by the external sensors.
14
(a)
(b)
(c)
(d)
(e)
Figure 8: This figure shows the results for the third experiment. From top to bottom, we show: the representation of the environment (including people & robot
positions and the path planned); the detections from the onboard camera; and the detections from an external camera. Through the results shown in Figs. (a)-(e), it
is possible to observe the two paths planned and followed by the robot when the TV is on and off. Firstly, the robot plans its path passing between the couch and the
pedestrian. When the TV is turned on, the robot replans its path to go around the couch.
(a)
(b)
(c)
(d)
(e)
Figure 9: Similar to the scheme shown in Fig. 8, from the top to the bottom, we show: the environment representation (with the location of the pedestrian and the
robot, as well as the robot’s path); detections from the onboard camera; and the detections from an external camera. From Figs: (a)-(e), it is possible to observe that
the robot follows the path planned to the hand over position, and that the costmap of the personal space is changed so the robot can reach that position.
15
[6] R. Girshick, J. Donahue, T. Darrell, J. Malik, Rich feature hierarchies for
accurate object detection and semantic segmentation, IEEE Proc. Conf.
Computer Vision and Pattern Recognition (CVPR).
[7] J. Hosang, M. Omran, R. Benenson, B. Schiele, Taking a deeper look at
pedestrians, IEEE Proc. Conf. Computer Vision and Pattern Recognition
(CVPR).
[8] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang,
A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, L. Fei-Fei, ImageNet
Large Scale Visual Recognition Challenge, Int’l J. Computer Vision.
[9] J. Yosinski, J. Clune, Y. Bengio, H. Lipson, How transferable are features
in deep neural networks?, Proc. Neural Information Processing Systems
(NIPS).
[10] A. Mateus, P. Miraldo, P. U. Lima, J. Sequeira, Human-Aware Navigation
using External Omnidirectional Cameras, Iberian Robotics Conference.
[11] D. Ribeiro, A. Mateus, P. Miraldo, J. C. Nascimento, A real-time deep
learning pedestrian detector for robot navigation, in: Autonomous Robot
Systems and Competitions (ICARSC), 2017 IEEE International Conference on, IEEE, 2017, pp. 165–171.
[12] P. Dollar, C. Wojec, B. Schiele, P. Perona, Pedestrian detection: An evaluation of the state of the art, IEEE Trans. Pattern Analysis and Machine
Intelligence.
[13] R. Benenson, M. Omran, J. Hosang, B. Schiele, Computer Vision - ECCV
2014 Workshops: Zurich, Switzerland, September 6-7 and 12, 2014, Proceedings, Part II, Springer International Publishing, 2014, Ch. Ten Years
of Pedestrian Detection, What Have We Learned?
[14] W. Nam, P. Dollar, J. H. Han, Local decorrelation for improved pedestrian
detection, Proc. Neural Information Processing Systems (NIPS).
[15] S. Zhang, R. Benenson, B. Schiele, Filtered channel features for pedestrian detection.
[16] I. Goodfellow, Y. Bengio, A. Courville, Deep learning, book in preparation for MIT Press (2016).
[17] R. Girshick, Fast R-CNN.
[18] S. Ren, K. He, R. Girshick, J. Sun, Faster R-CNN: Towards real-time
object detection with region proposal networks.
[19] J. Li, X. Liang, S. Shen, T. Xu, S. Yan, Scale-aware fast R-CNN for
pedestrian detection, CoRR abs/1510.08160.
URL http://arxiv.org/abs/1510.08160
[20] Z. Cai, Q. Fan, R. S. Feris, N. Vasconcelos, A unified multi-scale deep
convolutional neural network for fast object detection (2016) 354–370.
[21] L. Zhang, L. Lin, X. Liang, K. He, Is Faster R-CNN Doing Well for
Pedestrian Detection?
[22] J. Hosang, M. Omran, R. Benenson, B. Schiele, Taking a deeper look at
pedestrians, in: Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), 2015, pp. 4073–4082.
[23] D. Ribeiro, J. C. Nascimento, A. Bernardino, G. Carneiro, Improving the
performance of pedestrian detectors using convolutional learning, Pattern
Recognition.
[24] E. A. Sisbot, F. M.-U. Luis, R. Alami, T. Simeon, A Human Aware Mobile Robot Motion Planner, IEEE Trans. Robotics.
[25] T. Kruse, A. Kirsch, E. A. Sisbot, R. Alami, Exploiting Human Cooperation in Human-Centered Robot Navigation, IEEE Int’l Symposium on
Robot and Human Interactive Communication (RO-MAN).
[26] T. Kruse, P. Basili, S. Glasauer, A. Kirsch, Legible Robot Navigation
in the Proximity of Moving Humans, IEEE Workshop on Advanced
Robotics and its Social Impacts (ARSO).
[27] R. Kirby, Social Robot Navigation, Ph.D. thesis, Robotics Institute,
Carnegie Mellon University (2010).
[28] L. Scandolo, T. Fraichard, An Anthropomorphic Navigation Scheme for
Dynamic Scenarios, IEEE Proc. Int’l Conf. on Robotics and Automation
(ICRA).
[29] A. Pandey, R. Alami, A Framework towards a Socially Aware Mobile Robot Motion in Human-Centered Dynamic Environment, IEEE/RSJ
Proc. Int’l Conf. Intelligent Robots and Systems (IROS).
[30] T. Kruse, A. K. Pandey, R. Alami, A. Kirsch, Human-aware robot navigation: A survey, Robotics and Autonomous Systems.
[31] E. Hall, The Hidden Dimension: Man’s Use of Space in Public and Private, Bodley Head, 1969.
[32] D. Shi, E. G. Collins Jr., B. Goldiez, A. Donate, X. Liu, D. Dunlap,
Human-Aware Robot Motion Planning with Velocity Constraints, Int’l
Symposium Collaborative Technologies and Systems (CTS).
[33] X. Wang, M. Yang, S. Zhu, Y. Lin, Regionlets for generic object detection,
IEEE Proc. Int’l Conf. Computer Vision (ICCV).
[34] K. Simonyan, A. Zisserman, Very deep convolutional networks for largescale image recognition, Int’l Conf. Learning Representations (ICLR).
[35] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with
deep convolutional neural networks, Proc. Neural Information Processing
Systems (NIPS).
[36] P. Dollár, Piotr’s Computer Vision Matlab Toolbox (PMT), http://
vision.ucsd.edu/˜pdollar/toolbox/doc/index.html
(2012).
[37] A. Vedaldi, K. Lenc, MatConvNet – Convolutional Neural Networks for
MATLAB, ACM Proc. Int’l Conf. on Multimedia.
[38] S. Paisitkriangkrai, C. Shen, A. van den Hengel, Strengthening the effectiveness of pedestrian detection with spatially pooled features, Proc.
European Conf. Computer Vision (ECCV).
[39] A. Verma, R. Hebbalaguppe, L. Vig, S. Kumar, E. Hassan, Pedestrian
detection via mixture of cnn experts and thresholded aggregated channel
features, ICCV Workshop.
[40] Y. Bar-Shalon, F. Daum, J. Huang, The probabilistic data association filter: Estimation in the presence of measurement origin uncertainty, IEEE
Control Systems Magazine.
[41] P. E. Hart, N. J. Nilsson, B. Raphael, A Formal Basis for the Heuristic
Determination of Minimum Cost Paths, IEEE Trans. Systems Science
and Cybernetics 4 (2).
[42] E. Marder Eppstein, E. Berger, T. Foote, B. Gerkey, K. Konolige, The
Office Marathon: Robust Navigation in an Indoor Office Environment,
IEEE Proc. Int’l Conf. Robotics and Automation (ICRA).
[43] K. L. Koay, E. A. Sisbot, D. S. Syrdal, M. L. Walters, K. Dautenhahn,
R. Alami, Exploratory Study of a Robot Approaching a Person in the
Context of Handing Over an Object., AAAI spring symposium: multidisciplinary collaboration for socially assistive robotics.
[44] D. V. Lu, D. Hershberger, W. D. Smart, Layered Costmaps for ContextSensitive Navigation, IEEE/RSJ Proc. Int’l Conf. Intelligent Robots and
Systems (IROS).
[45] J. Messias, R. Ventura, P. Lima, J. Sequeira, P. Alvito, C. Marques,
P. Carriço, A Robotic Platform for Edutainment Activities in a Pediatric
Hospital, IEEE Int’l Conf. Autonomous Robot Systems and Competitions
(ICARSC).
16
| 1cs.CV
|
Under consideration for publication in Theory and Practice of Logic Programming
1
arXiv:1210.1653v1 [cs.LO] 5 Oct 2012
An Improved Proof-Theoretic Compilation of Logic
Programs
Iliano Cervesato
Department of Computer Science
Carnegie Mellon University
E-mail: [email protected]
submitted 1 January 2003; revised 1 January 2003; accepted 1 January 2003
Abstract
In prior work, we showed that logic programming compilation can be given a proof-theoretic justification for generic abstract logic programming languages, and demonstrated this technique in the
case of hereditary Harrop formulas and their linear variant. Compiled clauses were themselves logic
formulas except for the presence of a second-order abstraction over the atomic goals matching their
head. In this paper, we revisit our previous results into a more detailed and fully logical justification that does away with this spurious abstraction. We then refine the resulting technique to support
well-moded programs efficiently.
To appear in Theory and Practice of Logic Programming.
KEYWORDS: Compilation, Abstract Logic Programming, Hereditary Harrop Formulas, Well-Moded
Logic Programs.
1 Introduction
In (Cervesato 1998), we presented a general methodology for developing a compiler and
associated intermediate language for any abstract logic programming language (ALPL)
(Miller et al. 1991) that satisfies some basic proof-theoretic properties. We applied it abstractly to the language of hereditary Harrop formulas and its linear variant, and also based
the concrete implementations of the Twelf (Pfenning and Schürmann 1999) and LLF (Cervesato and Pfenning 2002)
systems directly on it. This methodology identified right sequent rules that behave like the
left rules that can appear in a uniform proof and used the corresponding connectives as the
compilation targets of the constructs in program clauses. The intermediate language was
therefore just another ALPL and its abstract machine relied on proof-search, like the source
ALPL. Because the transformation was based on the proof-theoretic duality between left
and right rules, proving the correctness of the compilation process amounted to a simple
induction. Finally, for Horn clauses the connectives in the target ALPL corresponded to
key instructions in the Warren Abstract Machine (WAM) (Warren 1983). The WAM is an
essential component of commercial Prolog systems since many compiled programs run
over an order of magnitude faster than when interpreted.
Up to then, the notoriously procedural instruction set of the WAM was regarded as a
wondrous piece of engineering without any logical status, in sharp contrast with the deep
2
I. Cervesato
logical roots of Prolog. In the words of (Börger and Rosenzweig 1995) “[the WAM] resembles an intricate puzzle, whose many pieces fit tightly together in a miraculous way”.
As a result, understanding it was complex in spite of the availability of excellent tutorials (Aı̈t-Kaci 1991), proving its correctness was a formidable task (Börger and Rosenzweig 1995;
Russinoff 1992), and adapting it to other logic programming languages a major endeavor
— it was done for CLP(R) (Jaffar et al. 1992) and λProlog (Nadathur and Mitchell 1999).
By contrast, the methodology in (Cervesato 1998) is simple, (mostly) logic-based, easily
verifiable, and of general applicability.
The technique in (Cervesato 1998) had however one blemish: it made use of equality
over atomic formulas together with a second-order binder over atomic goals, which lacked
logical status. In this paper, we remedy this drawback by carefully massaging the head of
clauses. This allows us to replace those constructs with term-level equality and regular universal quantifications over the arguments of a clause head. The result is an improved prooftheoretic account of compilation for logic programs that sits squarely within logic. It also
opens the doors to specializing the compilation process to well-moded programs, which
brings out the potential of doing away with unification in favor of matching, a more efficient operation in many languages. We present these results for the language of hereditary
Harrop formulas and only at the highest level of abstraction. Just like (Cervesato 1998),
they are however general, both in terms of the source ALPL and of the level of the abstraction considered. We are indeed in the process of using them to implement a compiler
for CLF (Watkins et al. 2003; Cervesato et al. 2003), a higher-order concurrent linear logic
programming language that combines backward and forward chaining.
The paper is organized as follows: Section 2 recalls the compilation process of (Cervesato 1998).
In Section 3, we present our improved compilation process. In Section 4, we refine it to
support moded programs. We lay out future developments in Sections 5 and 6.
2 Background and Recap
In this section, we recall the compilation process presented in (Cervesato 1998). For succinctness, we focus on a smaller source language — it corresponds to the language underlying the Twelf system (Pfenning and Schürmann 1999), on which this technique was first
used. We will comment on larger languages, including those examined in (Cervesato 1998),
in Section 5.
2.1 Source Language
We take the language freely generated from atomic propositions (a), intuitionistic implication (⊃) and universal quantification (∀) as our source language. We expand the openended atomic propositions of (Cervesato 1998), into a predicate symbol p followed by zero
or more terms t. A program is a sequence of closed formulas. This language, which we call
Ls , is given by the following grammar:
Formulas: A ::= a | A1 ⊃ A2 | ∀x. A
Atoms: a ::= p | a t
Programs: Γ ::= · | Γ, A
As in (Cervesato 1998), we leave the language of terms open, but require that it be predicative (substituting a term for a variable cannot alter the outer structure of a formula). We
An Improved Proof-Theoretic Compilation of Logic Programs
3
Uniform provability
′
u
u
Γ, A, Γ −→ A ≫ a
u
c “new”
Γ, A1 −→ A2
Γ −→ [c/x]A
u imp
u atm
u
u
Γ, A, Γ′ −→ a
u all
u
Γ −→ A1 ⊃ A2
Γ −→ ∀x. A
Immediate entailment
u
u
u
Γ −→ [t/x]A ≫ a
Γ −→ A1 ≫ a Γ −→ A2
i imp
i atm
u
u
Γ −→ a ≫ a
i all
u
Γ −→ A2 ⊃ A1 ≫ a
Γ −→ ∀x. A ≫ a
Fig. 1. Uniform Deduction System for Ls .
will often write an atom a as p t, where p is its predicate symbol and t is the sequence of
terms it is applied to. We implicitly assume that a predicate symbol is consistently applied
to the same number of terms throughout a program — its arity. We write [t′ /x]t (resp.
[t′ /x]A) for the capture-avoiding substitution of term t′ for all free occurrences of variable
x in term t (resp. in formula A). Simultaneous substitution is denoted [t′ /x]t and [t′ /x]A.
Ls is an abstract logic programming language (Miller et al. 1991) and, for appropriate
choices of the term language, has indeed the same expressive power as λProlog (Miller and Nadathur 1986)
or Twelf (Pfenning and Schürmann 1999). It differs from the first language discussed in (Cervesato 1998)
for the omission of conjunction and truth (see Section 5).
The operational semantics of Ls is given by the two judgments
u
Γ −→ A
u
Γ −→ A ≫ a
A is uniformly provable from Γ
a is immediately entailed by A in Γ
Their defining rules, given in Figure 1, produce uniform proofs (Miller et al. 1991): the
uniform provability judgment includes the right sequent rules for Ls and, once the goal is
atomic, rule u atm calls the immediate entailment judgment, which focuses on a program
formula A and decomposes it as prescribed by the left sequent rules. This strategy is complete with respect to the traditional sequent rules of this logic (Miller et al. 1991). From a
logic programming perspective, the connectives appearing in the goal — handled by right
rules — are search directives, while the left rules carry out a run-time preparatory phase.
2.2 Target Language
In (Cervesato 1998), the target language of the compilation process distinguished compiled
goals (G) from compiled clauses (C). A compiled goal was either an atomic proposition,
or a hypothetical goal (a goal to be solved in the presence of an additional clause) or a
universal goal (a goal to be solved in the presence of a new constant). A compiled clause
had the form Λα. C, where the second-order variable α stood for the atomic goal to be
resolved against the present clause, while C could either match α with the head a of this
.
clause (a = α), invoke a goal (C ∧ G), or request that a variable x be instantiated with
a term (∃x. C). A compiled program Ψ was then a sequence of compiled clauses. The
grammar for the resulting language, which we call Lc0 , is as follows:
Goals: G ::= a | (Λα. C) ⊃ G | ∀x. G
.
Clauses: C ::= a = α | C ∧ G | ∃x. C
Programs: Ψ ::= · | Ψ, Λα. C
The operational semantics of a compiled program, as given by the above grammar, is
4
I. Cervesato
Goals
′
c0
c
0
Ψ, Λα. C −→
G
Ψ, Λα. C, Ψ −→ [a/α]C
g0 atm
c “new”
c0
Ψ −→ [c/x]G
g0 imp
c
g0 all
c
0
Ψ, Λα. C, Ψ′ −→
a
c
0
Ψ −→
(Λα. C) ⊃ G
0
Ψ −→
∀x. G
Clause instances
c
r0 eq
c
0
Ψ −→
C̃
c0
c
Ψ −→ [t/x]C̃
0
Ψ −→
G
r0 and
c
.
0
Ψ −→
a=a
0
Ψ −→
C̃ ∧ G
r0 exists
c
0
Ψ −→
∃x. C̃
Fig. 2. Search Semantics of Lc0 .
defined on the basis of the following two judgments:
c
0
G
Ψ −→
c0
Ψ −→ C̃
G is uniformly provable from Ψ
C̃ is uniformly provable from Ψ
Here, clause instances C̃ are C’s whose variable α has been instantiated with an atomic
formula a′ . The operational semantics of Lc0 is shown in Figure 2. Observe that, with
the partial exception of g0 atm, it consists solely of right rules. This means that every
connective is seen as a search directive: the dynamic clause preparations embodied by the
left rules has now been turned into right search rules through a static compilation phase.
2.3 Compilation
Compilation, the process that transforms a logic program in Ls into a compiled program
in Lc0 , is expressed by means of the following three judgments:
Γ ≫ Ψ
A ≫ α\C
A ≫ G
Program Γ is compiled to Ψ
Clause A with α is compiled to C
Goal A is compiled to G
These judgments are defined by the rules in Figure 3 — see (Cervesato 1998) for details.
As our ongoing example, consider the following two clauses, taken from a type checking
specification for a Church-style simply typed λ-calculus. For clarity, we write program
clauses Prolog-style, using the reverse implication ⊂ instead of ⊃ in positive formulas.
1.
∀E1 . ∀E2 . ∀T1 . ∀T2 .
of (app E1 E2 ) T2
⊂ of E1 (arr T1 T2 )
⊂ of E2 T1
≫
∀E. ∀T1 . ∀T2 .
of (lam T1 E) (arr T1 T2 )
⊂ (∀x. of x T1
⊃ of (E x) T2 )
≫
2.
Λα.
∃E1 . ∃E2 . ∃T1 . ∃T2 .
.
(of (app E1 E2 ) T2 ) = α
∧ of E1 (arr T1 T2 )
∧ of E2 T1
Λα.
∃E. ∃T1 . ∃T2 .
.
(of (lam T1 E) (arr T1 T2 )) = α
.
∧ (∀x. Λβ. ((of x T1 ) = β)
⊃ of (E x) T2 )
The compiled language Lc0 is sound and complete for Ls . See (Cervesato 1998) for the
formal statements. The proof of both directions proceeds by straightforward induction,
which contrasts greatly with the complex proofs of soundness and correctness previously
devised for the WAM (Börger and Rosenzweig 1995; Russinoff 1992).
An Improved Proof-Theoretic Compilation of Logic Programs
5
Programs
Γ ≫ Ψ A ≫ α\C
p0c empty
p0c clause
· ≫ ·
Γ, A ≫ Ψ, Λα. C
Clauses
B ≫ α\C
.
A ≫ G
c0c atm
a ≫ α\a = α
A ≫ α\C
c0c imp
c0c all
A ⊃ B ≫ α\C ∧ G
∀x. A ≫ α \ ∃x. C
Goals
A ≫ α\C
B ≫ G
g0c atm
a ≫ a
A ≫ C
g0c imp
g0c all
A ⊃ B ≫ (Λα. C) ⊃ G
∀x. A ≫ ∀x. C
Fig. 3. Compilation of Ls into Lc0 .
3 Fully Logical Compilation
Because clauses are compiled to expressions of the form Λα. C, the language Lc0 is not
fully logical. In this section we consider a different compilation target, the language Lc1 ,
which lies entirely within logic.
In the previous section, a generic Horn clause of the form
∀y. (p t ⊂ a1 ⊂ . . . ⊂ an )
(1)
.
was compiled into Λα. ∃y. (p t = α ∧ a1 ∧ . . . ∧ an ). During execution, rule c0 atm
.
reduced the current atomic goal a to the clause instance ∃y. (p t = a ∧ a1 ∧ . . . ∧ an ).
Note that t may depend on y, but a does not. We will now compile that Horn clause into
.
∀x. (p x ⊂ ∃y. (x = t ∧ a1 ∧ . . . ∧ an ))
(2)
where x is a sequence of fresh variables, all distinct from each other, and equal in number
.
to the arity of p, and x = t stands for a conjunction of equalities between each variable
xi in x and the term ti in t in the corresponding position (or ⊤ if the arity of p is zero).
Notice that the non-logical second-order binder “Λα.” is gone. At run time, formula (2)
.
will resolve an atomic goal p t′ into the clause p t′ ⊂ ∃y. (t′ = t ∧ a1 ∧ . . . ∧ an ), which
′ .
immediately reduces to ∃y. (t = t ∧ a1 ∧ . . . ∧ an ). Like earlier, t may depend on y,
but t′ does not. The variables x correspond directly to the “argument registers” (An) of the
WAM (Aı̈t-Kaci 1991), while the y’s are closely related to its “permanent variables” (Yn).
Formula (2) can be understood as an uncurried form of (1): outer implications are transformed into conjunctions and universals into existentials. Doing so literally would yield the
formula p t ⊂ ∃y. (a1 ∧ . . . ∧ an ), which is incorrect because occurrences of variables
in y within t have escaped their scope. Instead, formula (2) installs fresh variables x as the
.
arguments of the head predicate p and adds the equality constraints x = t in the body.
3.1 Target Language
We now generalize the above intuition to any formula in Ls , not just Horn clauses. Our
second target language, Lc1 , is given by the following grammar.
Goals: G ::= a | C ⊃ G | ∀x. G
Clauses: C ::= R ⊃ p x | ∀x. C
.
Residuals: R ::= x = t | ⊤ | R ∧ G | ∃x. R
Programs: Ψ ::= · | Ψ, C
6
I. Cervesato
Goals
c
c
1
Ψ, C, Ψ′ −→
C ≫ a
1
Ψ, C −→
G
g1 atm
c “new”
c1
Ψ −→ [c/x]G
g1 imp
c
g1 all
c
1
Ψ, C, Ψ′ −→
a
c
1
Ψ −→
C ⊃G
1
Ψ −→
∀x. G
Clauses
c1
c
Ψ −→ [t/x]C̃ ≫ a
1
Ψ −→
R̃
c1 imp
c1 all
c
c
1
Ψ −→
R̃ ⊃ a ≫ a
1
Ψ −→
∀x. C̃ ≫ a
Residuals
c
1
Ψ −→
R̃
r1 eq
c
c
.
1
Ψ −→
t=t
1
Ψ −→
⊤
c1
c
Ψ −→ [t/x]R̃
1
Ψ −→
G
r1 true
r1 and
c
1
Ψ −→
R̃ ∧ G
r1 exists
c
1
Ψ −→
∃x. R̃
Fig. 4. Search Semantics of Lc1 .
Compiled goals (G) are just like in Section 2.2: atoms, hypothetical goals, or universal
goals. Compiled clauses (C) have the form ∀x. (R ⊃ p x), i.e., a (possibly empty) outer
layer of universal quantifiers enclosing an implication R ⊃ p x whose head p x always
consists of a predicate name (p) applied to a (possibly empty) sequence of distinct variables
.
(x). Its body is a residual (R). A residual can be either an equality constraint (x = t), the
trivial constraint ⊤ (logical truth), or like in Section 2.2 a goal invocation or an instantiation
request. Notice that C is now the full result of compiling a clause.
The operational semantics of Lc1 is specified by the following three judgments:
c
1
G
Ψ −→
c1
Ψ −→ C̃ ≫ a
c1
R̃
Ψ −→
G is uniformly provable from Ψ
a is immediately entailed by C̃ in Ψ
R̃ is uniformly provable from Ψ
where C̃ and R̃ differ from C and R by the instantiation of some variables in a clause head
and on the left-hand side of equalities, respectively.
Their operational semantics is given in Figure 4. Goals are handled exactly in the same
way as uniform provability in Ls (top part of Figure 1). The operational reading of compiled clauses is an instance of that of immediate entailment: rule c1 imp is a special case
of i imp while c1 all is isomorphic to i all. Note that rule c1 imp reduces immediately
to the residual R if the head of the clause matches the atomic goal a being proved. The
rules for residuals correspond closely to the rules for clause instances for our original target language at the bottom of Figure 2: rule r1 eq requires that the two sides of an equality
be indeed equal and rule r1 true is always satisfied.
The rules in Figure 4 build uniform proofs (Miller et al. 1991), characteristic of abstract
logic programming languages: the operational semantics decomposes a goal to an atomic
formula (top segment of Figure 4), then selects a clause and focuses on it until it finds a
matching head (middle segment) and then decomposes its body (bottom segment), which
may eventually expose some goals, and the cycle repeats. In particular, once an atomic goal
p t has been exposed, a successful derivation will necessarily contain an instance of rule
g1 atm that picks a clause C with head p x, as many instances of rule c1 all as the arity
of p, and an instance of rule c1 imp. This necessary sequence of steps is captured by the
An Improved Proof-Theoretic Compilation of Logic Programs
7
following derived “macro-rule” (the backchaining rule):
c
1
[t/x]R
Ψ, ∀x. (R ⊃ p x), Ψ′ −→
c
g1 atm′
1
pt
Ψ, ∀x. (R ⊃ p x), Ψ′ −→
Replacing rules g1 atm, c1 all and c1 imp with rule g1 atm′ yields a system that is
equivalent to that in Figure 4. Taking it as primitive amounts to replacing the construction
for compiled clauses, ∀x. (R ⊃ px), with a synthetic connective, call it Λp x. R. Therefore,
by accounting for the structure of atomic propositions and proper quantification patterns,
Lc1 provides a fully logical justification for clause compilation that Lc0 ’s Λα. C lacked.
3.2 Compilation
Compilation transforms logic programs in Ls into compiled logic programs in Lc1 . In order
to define it, the auxiliary notion of pseudo clause will come handy:
Pseudo Clauses: C ::= ⊃ p x | ∀x. C
A pseudo clause retains the outer structure of a clause, but has a hole () in place of the
residual R. In general, a pseudo clause C has the form ∀x. ⊃ p x′ . In a fully compiled
clause, variables x will coincide with x′ .
Pseudo clauses are generated while processing the head of a clause. The hole then needs
to be replaced with the compiled body, a residual. We write this operation, pseudo clause
instantiation, as C[R]. It is formally defined as follows:
( ⊃ p x)[R] = R ⊃ p x
(∀x. C)[R]
= ∀x. (C[R])
As is often the case with such contextual operations, pseudo clause instantiation can, and
generally will, lead to variable capture: in (∀x. ⊃ p x)[R], there may be free occurrences of variables in x within R. In the result, these occurrences are bound by the outer
quantifiers.
Compilation is expressed by means of the following four judgments
Γ ≫ Ψ
x ⊢ a ≫ C\E
A ≫ C\R
A ≫ G
Program Γ is compiled to Ψ
Head a with x is compiled to C and E
Clause A is compiled to C and R
Goal A is compiled to G
and defined by the rules in Figure 5, where we wrote E for conjunctions of equalities.
The judgment A ≫ C \ R compiles an Ls clause A into a pseudo clause C and a residual
R. They are assembled into an Lc1 clause in rules p1c clause and g1c imp. Programs
and goals are otherwise compiled just as for Lc0 in Figure 3. Clause heads are handled
differently: rule c1c atm invokes the auxiliary head compilation judgment to compile the
.
goal p t into a pseudo clause ∀x. ⊃ p x and the equalities x = t, which will form the
seed of the clause’s residual.
Consider the first example clause in Section 2.3. Its head (of (app E1 E2 ) T2 ) is compiled into the pseudo clause ∀x1 . ∀x2 . ( ⊃ of x1 x2 ) and the equality constraints ⊤ ∧
.
.
(x1 = app E1 E2 ) ∧ (x2 = T2 ), where x1 and x2 are new variables. These core equalities
8
I. Cervesato
Programs
Γ ≫ Ψ A ≫ C\R
p1c empty
p1c clause
· ≫ ·
Γ, A ≫ Ψ, C[R]
Heads
x x ⊢ a ≫ C\E
x ⊢ p ≫ ⊃ p x\⊤
x “new”
h1c pt
h1c p
.
x ⊢ a t ≫ ∀x. C \ E ∧ x = t
Clauses
· ⊢ a ≫ C\E
A ≫ G B ≫ C\R
a ≫ C\E
A ≫ C\R
c1c imp
c1c atm
A ⊃ B ≫ C\R ∧ G
c1c all
∀x. A ≫ C \ ∃x. R
Goals
A ≫ C\R B ≫ G
g1c atm
a ≫ a
A ≫ C
g1c imp
A ⊃ B ≫ C[R] ⊃ G
g1c all
∀x. A ≫ ∀x. C
Fig. 5. Compilation of Ls into Lc1 .
are then extended with the compiled body of that clause, (of E1 (arr T1 T2 )) ∧ (of E2 T1 ),
and existential quantifications over the original variables of the clause, E1 , E2 , T1 and T2 ,
are finally wrapped around the result before embedding it in the hole of the pseudo clause.
The resulting Lc1 clause is displayed in the top part of Figure 3.2.
The target language Lc1 is sound and complete with respect to Ls . In order to show it,
we need the following auxiliary results. The first statement is proved by induction on the
structure of a. The second by induction on the given derivation.
Lemma 3.1
c1
• If x ⊢ a ≫ C \ E, then for all t of the same length as x and all Ψ we have Ψ −→
[t/x](C[E]) ≫ a t.
c1
c1
R.
C[R] ≫ a, then Ψ −→
• If Ψ −→
The statements of soundness and completeness are as follows. For each of them, the
proof proceeds by mutual induction on the first derivation in the antecedent.
Theorem 3.2 (Soundness of the compilation to Lc1 )
c1
u
G.
• If Γ −→ A, Γ ≫ Ψ and A ≫ G, then Ψ −→
u
c1
C[R] ≫ a.
• If Γ −→ A ≫ a, Γ ≫ Ψ and A ≫ C \ R, then Ψ −→
Theorem 3.3 (Completeness of the compilation to Lc1 )
c1
u
G, Γ ≫ Ψ and A ≫ G, then Γ −→ A.
• If Ψ −→
c1
u
• If Ψ −→ C ≫ a, Γ ≫ Ψ, C = C[R] and A ≫ C \ R, then Γ −→ A ≫ a.
We conclude this section by showing in Figure 3.2 the output of our compilation procedure for the two examples seen in Section 2.3. We stretch the source clauses (left) to align
corresponding atoms. As can be gleaned from these clauses, there are ample opportunities
.
for optimizations in our compilation process. In particular, a constraint x = y mentioning
variables on both sides can often be eliminated by replacing the existential variable y with
the universal variable x in the rest of the clause (and removing the existential quantifier)
— the exception is when there are multiple constraints of this form for the same y. The
leading logical constant ⊤ makes for a succinct presentation of the compilation process,
but plays no actual role: it can also be eliminated.
An Improved Proof-Theoretic Compilation of Logic Programs
1. ∀E1 . ∀E2 . ∀T1 . ∀T2 .
of (app E1 E2 ) T2
≫
⊂
⊂
of E1 (arr T1 T2 )
of E2 T1
2. ∀E. ∀T1 . ∀T2 .
of (lam T1 E) (arr T1 T2 )
⊂
(∀x.
≫
of x T1
⊃ of (E x) T2 )
9
∀x1 . ∀x2 .
of x1 x2
⊂ (∃E1 . ∃E2 . ∃T1 . ∃T2 . ⊤
.
∧ x1 = app E1 E2
.
∧ x 2 = T2
∧ of E1 (arr T1 T2 )
∧ of E2 T1 )
∀x1 . ∀x2 .
of x1 x2
⊂ (∃E. ∃T1 . ∃T2 . ⊤
.
∧ x1 = lam T1 E
.
∧ x2 = arr T1 T2
∧ (∀x. ∀x′1 . ∀x′2 . ⊤
.
∧ x′1 = x
′ .
∧ x 2 = T1
∧ of x′1 x′2 )
⊃ of (E x) T2 )
Fig. 6. Lc1 Compilation Example
It is interesting to rewrite these clauses using the synthetic connective Λp discussed
earlier (we have omitted occurrences of ⊤ for readability):
Λof x1 x2 . ∃E1 . ∃E2 . ∃T1 . ∃T2 .
.
.
x1 = app E1 E2 ∧ x2 = T2
∧ of E1 (arr T1 T2 ) ∧ of E2 T1
Λof x1 x2 . ∃E. ∃T1 . ∃T2 .
.
.
x1 = lam T1 E ∧ x2 = arr T1 T2
.
.
∧ ∀x. (Λof x′1 x′2 . x′1 = x ∧ x′2 = T1 ) ⊃ of (E x) T2
4 Support for Moded Programs
In this section, we will specialize the compilation process just outlined to the case where the
source program is well-moded. In a well-model program, the argument positions of each
predicate symbol are designated as either input or output. Input arguments are guaranteed
to be ground terms at the time a goal is called. Dually, output arguments are guaranteed to
have been made ground by the time the call returns.
There are operational benefits to working with well-moded programs: while an interpreter for a generic program must implement term-level unification, well-moded programs
can be executed by relying uniquely on pattern matching and variable instantiation. This is
desirable because matching often behaves better than general unification. For example, it
is more efficient for first-order term languages were it only because it does away with the
occurs-check, and it is decidable for higher-order term languages while general unification
is not (Stirling 2009).
The development in this section is motivated by well-moding, but is sound independently
10
I. Cervesato
of whether a program is well-moded or not. Statically enforcing well-moding brings the
operational advantages just discussed, but the results in this section do not depend on it.
4.1 Source Language
In this section, we assume that each predicate symbol in Ls comes with a mode which
declares each of its arguments as input, written ˇ, or output, written ˆ. For simplicity of
exposition, we decorate the actual arguments of all atomic propositions with these symbols,
so that a term t in input position in an atomic proposition is written ť (read “in t”). Similarly
t in output position is written t̂ (pronounced “out t”). This amounts to revising the grammar
of atomic propositions as follows:
Atoms: a ::= p | a ť | a t̂
Just like we assume that the arity of a predicate symbol p remains constant in a program,
we require that all atomic propositions for p have their input/output marks in the same
positions. This pattern is the mode of p — an actual language would rely on explicit mode
declarations.
For typographic convenience and without loss of generality, our examples assume that
input positions precede output positions so that an atomic formula a can be written as
p ť t̂ where ť and t̂ are the (possibly empty) sequences of terms in input (resp. output)
positions for p. To avoid notational proliferation, we use the markers ˇ and ˆ both as mode
designators and as symbol decorations (like primes and subscripts) when working with
generic terms. Therefore, ť and t̂ indicate possibly different terms in p ť t̂, and similarly for
term sequences, as in p ť t̂ above.
At our level of abstraction, the rules in Figure 1 capture the operational semantics of this
variant of Ls : mode annotations are simply ignored. However, moded execution requires
that two of the operational choices left open by those rules be resolved using some algorithmic strategy: the order in which rule i imp searches for derivations of its two premises,
and the substitution term that rule i all picks. For both, we will assume the same strategy
as Prolog: implement rule i imp left to right and implement rule i all lazily by replacing
each variable x with a “logical variable” X which is instantiated incrementally through
unification. This allows us to view an atomic goal as a (non-deterministic) procedure call.
In a well-moded program (Debray and Warren 1988), terms in input position are seen as
the actual arguments of this procedure, and terms in output position yield return values.
In this section, we will not formalize the notion of well-modedness — see (Debray and Warren 1988)
for Prolog and (Sarnat 2010) for Twelf — nor refine our operational semantics to make goal
evaluation order and unification explicit — see (Pientka 2003). We will instead refine our
compilation process to account for mode information and produce compiled programs that,
if well-moded, can be executed without appealing to unification.
4.2 Target Language
In Lc1 , a (well-moded) Horn clause ∀y. p ť t̂ ⊂ a1 ⊂ . . . ⊂ an was compiled into
.
.
∀x̌ x̂. (p x̌ x̂ ⊂ ∃y. (x̌ = ť ∧ x̂ = t̂ ∧ a1 ∧ . . . ∧ an )). Here, the left-to-right execution order forces us to guess the final values of the output variables x̂ before the goals
An Improved Proof-Theoretic Compilation of Logic Programs
11
.
in its body have been fully executed. In Lc2 , we will move the equality x̂ = t̂ after the
last goal an . Since x̂ appear nowhere else in the residual, this equality is no more than an
assignment of the computed instance of t̂ to x̂. Accordingly, we will write it as x̂ := t̂.
Furthermore, in a well-moded program, this clause will be invoked with ground terms in
.
input position, so that x̌ will be bound to ground terms. Then, the input equality x̌ = ť
will match the variables in ť with appropriate subterms. For this reason, we will write it as
x̌ =: ť. Expanding each goal ai into qi ťi t̂i , the above clause will be compiled (almost) as
follows, where the arrows represent the data flow of a well-moded execution (note that it
parallels the control flow):
O
x
∀x̌ x̂. ( p x̌ x̂
⊂ (∃y. x̌J =: ť
∧
q1 ť1 t̂1
K
∧
.M . .
∧
qn ťn t̂n
K
∧
x̂ := t̂K ))
When executing an atomic goal, it is desirable to separate the call from the verification
that the output terms returned by the caller match the expected output terms in this goal.
We will do so by rewriting any atomic goal q ť t̂ in a compiled clause into the formula
∃z. (q ť z ∧ z =: t̂) for fresh variables z. This transformation preserves the left-to-right
control and data flow. No special provision needs to be made for the input arguments of
q as variables in it will have been instantiated to ground terms at the moment the call is
made.
Next, we again generalize this intuition to any formula in Ls , not just Horn clauses. Our
third target language, Lc2 , is defined by the following grammar.
Goal Matches: M ::= ⊤ | M ∧ z =: t̂
Atomic Goals: F ::= p ť ẑ ∧ M | ∃z. F
Goals: G ::= F | C ⊃ G | ∀x. G
Programs: Ψ ::= · | Ψ, C
Clauses: C ::= R ⊃ p x̌ x̂ | ∀x. C
Residuals: R ::= x̌ =: t | x̂ := t | ⊤ | R ∧ G | ∃x. R
.
Residuals (R) refine the equality predicate x = t of Lc1 into a matching predicate x =: t
and an assignment predicate x := t. At our level of abstraction, they behave just like
equality. During well-moded execution, the match predicate will have the form tg =: tv
where tg is a ground term while tv may contain variables. It will bind these variables
to ground subterms of tg , thereby realizing matching. However, presented with programs
that are not well-moded, the terms tg cannot be assumed to be ground and =: performs
unification. The assignment predicate will be called as x := t where x is a variable and t a
term — a ground term for well-moded programs. It simply binds x to t. Compiled clauses
and programs are just like in Lc1 .
Following the motivations above, an atomic goal p ť t̂ is not compiled any more to itself
as in Lc1 , but to a formula F of the form ∃z. (q ť ẑ ∧ ž =: t̂). In the grammar above, we
isolated the match predicates ž =: t̂ as the non-terminal M .
12
I. Cervesato
Goals Matches
c2
−→ M
m2 true
m2 mtch
c2
c2
−→ ⊤
−→ M ∧ t =: t
Atomic Goals
c2
′
c2f
c2
Ψ, C, Ψ −→ C ≫ p ť t̂
Ψ −→ [t/z]R
−→ M
a2 atm
a2 exists
c2f
c2f
Ψ, C, Ψ′ −→ p ť t̂ ∧ M
Ψ −→ ∃z. R
Goals
c
c2f
Ψ −→ F
2
Ψ, C −→
G
c “new”
c2
Ψ −→ [c/x]G
g2 imp
g2 f
c
g2 all
c
2
Ψ −→
F
c
2
Ψ −→
C⊃G
2
Ψ −→
∀x. G
Clauses
c
2
Ψ −→
[t/x]C ≫ a
c2
Ψ −→ R
c2 imp
c2 all
c
c
2
Ψ −→
R⊃a ≫ a
2
Ψ −→
∀x. C ≫ a
Residuals
r2 assg
r2 mtch
c
r2 true
c
2
Ψ −→
t =: t
c
2
Ψ −→
t := t
c
2
Ψ −→
R
2
Ψ −→
⊤
c
2
Ψ −→
[t/x]R
c
2
Ψ −→
G
r2 and
c
2
Ψ −→
R ∧ G
r2 exists
c
2
Ψ −→
∃x. R
Fig. 7. Search Semantics of Lc2 .
We specify the operational semantics of Lc2 by means of the following five judgments:
c
2
−→
c2f
Ψ −→
c2
Ψ −→
c2
Ψ −→
c2
Ψ −→
M
F
G
C ≫ a
R
M is provable
F is uniformly provable from Ψ
G is uniformly provable from Ψ
a is immediately entailed by C in Ψ
R is uniformly provable from Ψ
which parallel the grammar just presented. The resulting operational semantics is shown
in Figure 7. The rules for clauses are unchanged with respect to Lc1 while that language’s
residual rule for equality has been duplicated into isomorphic rules for matching and assignment. The rules for compiled goals have instead proliferated due to our handling of
terms in output position in atomic goals. Observe that rule a2 atm is essentially a combination of rule g1 atm in Lc1 and the rule for conjunction. Rules a2 exists and m2 true
are just the standard rules for existential quantification and truth. Rule m2 mtch combines
the rules for conjunction and matching.
Just like in the case of Lc1 , the rules in Figure 7 construct proofs that are uniform (Miller et al. 1991),
which makes Lc2 an abstract logic programming language. In a successful derivation, this
operational semantics decomposes a goal to formulas of the form F = ∃z. (p ť ẑ ∧ ž =: t̂)
(rules in the “Goals” segment). Then, rules a2 exists, m2 mtch and m2 true necessarily reduce it in a few steps into the atomic formula p ť t̂. Similarly to Lc1 , the left premise
of rule a2 atm selects a clause and focuses on it until it finds a potentially matching head
(“Clauses” segment). It then proceeds to decomposing its body (“Residuals” segment) and
the cycle repeats with whatever goals it finds in there.
An Improved Proof-Theoretic Compilation of Logic Programs
13
As just noticed, any atomic goal F of the form ∃ẑ. (p ť ẑ ∧ ẑ =: t̂) is necessarily reduced
to p ť t̂ by as many applications of rule a2 exists as there are variables in ẑ, a pass-through
instance of a2 atm via its right branch, and a similar number of uses of rules m2 mtch
and m2 true respectively. This entails that the macro-rule a2 atm′ , on the left-hand side
of the following display, is derivable:
c
c
2
C ≫ p ť t̂
Ψ, C, Ψ′ −→
2
p ť t̂
Ψ −→
a2 atm′′
′
c2
a2 atm
′ c2f
Ψ −→ ∃z. (p ť z ∧ z =: t̂)
Ψ, C, Ψ −→ p ť t̂
Having factored rule a2 atm′ out, the work performed by a2 atm degenerates to rule
a2 atm′′ on the right-hand side of the above display, which is akin to u atm. The system
obtained by replacing the m2 ∗ and a2 ∗ rules as well as g2 f with rules a2 atm′ and
a2 atm′′ is indeed equivalent to the rule set in Figure 7.
Rule a2 atm′ entices us to interpret the compiled formula ∃z. (p ť z ∧ z =: t̂) for an
atomic goal p ť t̂ as a synthetic operator call p ť =: t̂ which invokes a clause for p with its
(ground) input arguments ť and matches the returned values against its terms t̂ in output
position.
Having recovered atomic goals p ť t̂ through rules a2 atm′ and a2 atm′′ , we can carry
out a sequence of reasoning steps similar to what led us to the backchaining rule for Lc1 .
Exposing the trailing assignments, a generic compiled clause C has the form ∀x̌x̂. (∃y. R ∧
x̂ := ŝ) ⊃ p x̌ x̂. In a successful derivation, all rule a2 atm′′ does is to pick such a
clause. Then, applications of rule c2 all will instantiate variables x̌ x̂ with the terms ť t̂,
and next rule c2 imp will invoke the instantiated residual [ť/x̌, t̂/x̂](∃y. R ∧ x̂ := ŝ).
Now, because x̂ does not occur in R and x̌ x̂ cannot appear in ŝ, this formula reduces to
∃y. ([ť/x̌]R ∧ t̂ := ŝ) by pushing the substitution in. Rule r2 exists will then instantiate
the variables y with terms u (which cannot mention variables x̌ x̂). Pushing this substitution
in yields the formula [ť/x̌, u/y]R ∧ t̂ := [u/y]ŝ since variables in y can occur in neither ť
nor t̂. Finally, by rule r2 assg, t̂ and [u/y]ŝ must be equal in a successful derivation. This
necessary sequence of steps is captured by the following derived backchaining macro-rule,
c
2
[ť/x̌, u/y]R
Ψ, C, Ψ′ −→
′
c2
Ψ, ∀x̌ x̂. (∃y. R ∧ x̂ := ŝ) ⊃ p x̌ x̂, Ψ −→ p ť [u/y]ŝ
|
{z
}
g2 atm′
C
where we have carried out the assignment t̂ := [u/y]ŝ in the conclusion. This rule can be
seen as a refinement of g1 atm′ in Lc1 that makes use of the trailing assignment in the
compiled clauses of Lc2 . With this derived inference, rules a2 atm′′ , c2 imp and c2 all
become unnecessary: the system consisting of rules a2 atm′ , g2 atm′ , the goal rules
for implication and universal quantification, and the residual rules is equivalent to that in
Figure 7.
Taking rule g2 atm′ as primitive amounts to replacing compiled clauses with the following synthetic connective, which refines Lc1 ’s Λp x. R.
∀x̌ x̂. p x̌ x̂ ⊂ ∃y. (R
{z
}
|
Λp x̌.
∃y. (R
∧
;
x̂ := t̂)
| {z }
return t̂)
14
I. Cervesato
Programs
Γ ≫ Ψ A ≫ C \R 0 O
p2c empty
p2c clause
· ≫ ·
Γ, A ≫ Ψ, C[R ∧ O]
Heads
h2c p
x⊢p ≫ ⊃px %⊤0⊤
xx⊢a ≫ C%I 0O
xx⊢a ≫ C%I 0O
x “new”
x “new”
h2c ot
h2c in
x ⊢ a ť ≫ ∀x. C % I ∧ x =: ť 0 O
x ⊢ a t̂ ≫ ∀x. C % I 0 x := t̂ ∧ O
Clauses
A ≫ G
·⊢a ≫ C%I 0O
B ≫ C\R 0 O
a ≫ C\I 0 O
A ≫ C \R 0 O
c2c imp
c2c atm
A ⊃ B ≫ C\R ∧ G 0 O
c2c all
∀x. A ≫ C \ ∃x. R 0 O
Atomic goals
t z ⊢ a ≫ C\M
ť t ⊢ a ≫ F \ M
a2c p
t ⊢ p ≫ p t ∧ \⊤
z “new”
a2c ot
a2c in
t ⊢ a ť ≫ F \ M
t ⊢ a t̂ ≫ ∃z. F \ z =: t̂ ∧ M
Goals
· ⊢ a ≫ F \M
A ≫ C\R 0 O
B ≫ G
g2c atm
a ≫ F [M]
A ≫ C
g2c imp
A ⊃ B ≫ C[R ∧ O] ⊃ G
g2c all
∀x. A ≫ ∀x. C
Fig. 8. Compilation of Ls into Lc2 .
The variables y are then interpreted as local variables for the execution of this clause. In
this, they are akin to the Yn permanent variables of the WAM (Aı̈t-Kaci 1991).
In a valid proof in this system, an occurrence of a2 atm′ is always immediately followed by an instance of g2 atm′ : the conclusion of the latter must match the premise of
the former. This fact realizes the requirement that, upon returning from a call, the output
terms, here [u/y]ŝ, must be checked against the terms in output position of the caller.
4.3 Compilation
Compilation transforms logic programs in Ls to compiled programs in Lc2 . The input does
not have to be well-moded at the level of detail considered here, but this would be operationally advantageous in a refinement of the semantics in Figure 7 that handles quantifiers
lazily. We will make use of two auxiliary notions in this section: pseudo clauses that we
encountered already in Section 3.2 and the analogous notion of pseudo atomic goal. They
are defined as follows:
Pseudo Clauses: C ::= ⊃ p x | ∀x. C
Pseudo Atomic Goals: F ::= p ť ẑ ∧ | ∃z. F
Just like pseudo clauses retain the outer structure of a clause replacing the embedded residual with a hole (), pseudo atomic goals have a hole in place of their trailing matches. The
general form of pseudo clauses and pseudo atomic formulas, accounting for input and output positions, are ∀x̌ x̂. ⊃ p x̌′ x̂′ and ∃ẑ. (p ť zˆ′ ∧ ). In Section 3.2, wrote C[R] for
the replacement of the hole of C with the residual R and noted that variable capture could
(and generally will) occur. Similarly, we write F [M ] for replacement of the hole of F with
matches M .
The compilation process is modeled by the following five judgments, which are reminiscent of the compilation judgments Lc1 . They are more complex because clause compilation
An Improved Proof-Theoretic Compilation of Logic Programs
15
now needs to handle both matching and assignment as opposed to a generic equality. Furthermore, a new judgment is needed to compile atomic goals.
Γ ≫ Ψ
x⊢a ≫ C%I 0O
A ≫ C\R 0 O
t ⊢ a ≫ F \M
A ≫ G
Program Γ is compiled to Ψ
Head a with x is compiled to C, I and O
Clause A is compiled to C, R and O
Atomic goal a with t is compiled to F and M
Goal A is compiled to G
We write I and O for a conjunction of matches (compilation of terms in input position) and
assignments (compilation of output terms), respectively, in the body of a compiled clause.
In compiled atomic goals, we write M for a conjunction of matches.
The rules for compilation, which define these judgments, are shown in Figure 8. Compiling a clause A, modeled by the judgment A ≫ C \ R 0 O, returns a pseudo clause
C, the residual R (inclusive of input matches) and the output assignments O that will fill
its hole. The rules in the “Clauses” segment build up this residual starting with the compilation of its head, which is displayed in the “Heads” segment. The rules therein differ
from the similar inference for Lc1 by the fact that they dispatch terms in input and output
positions in the I and O zones of the judgment as matches and assignments respectively.
Residuals and assignments are plugged in the hole of the pseudo clause once this clause
has been fully compiled, as can be seen in the “Programs” segment and in rule g2c imp.
The compilation of goals differs from Lc1 for the treatment of atomic formulas: upon
encountering an atom a, the compilation appeals to the new judgment · ⊢ a ≫ F \ M .
It generates a pseudo atomic formula F and matches M , which are integrated in rule
g2c atm. The zone to the left of the turnstile serves as an accumulator, very much like
when compiling heads.
Target language, Lc2 , is sound and complete with respect to Ls . The following lemma
collects some auxiliary results needed to prove this property. The first two statements are
proved by induction on the structure of a; the third by induction on the given derivation.
Lemma 4.1
• If x ⊢ a ≫ C % I 0 O, then for any term sequence t of the same length as x and
c2
[t/x](C[I ∧ O]) ≫ a t.
program Ψ we have Ψ −→
c2
F [M ] ≫ a t.
• If t ⊢ a ≫ F \ M , then for all Ψ we have Ψ −→
c1
c2
• If Ψ −→ C[R] ≫ a, then Ψ −→ R.
We have the following soundness and completeness theorems for Lc2 . In both cases, the
proof proceeds by mutual induction over the first derivation in the antecedent.
Theorem 4.2 (Soundness of the compilation to Lc2 )
c2
u
G.
• If Γ −→ A, Γ ≫ Ψ and A ≫ G, then Ψ −→
c2
u
C[R ∧ O] ≫ a.
• If Γ −→ A ≫ a, Γ ≫ Ψ and A ≫ C \ R 0 O, then Ψ −→
Theorem 4.3 (Completeness of the compilation to Lc2 )
c2
u
G, Γ ≫ Ψ and A ≫ G, then Γ −→ A.
• If Ψ −→
c2
• If Ψ −→ C ≫ a, Γ ≫ Ψ, C = C[R ∧ O] and A ≫ C \ R 0 O, then
u
Γ −→ A ≫ a.
To conclude this section, we revisit our ongoing examples. Here, we assume that the
16
I. Cervesato
1. ∀E1 . ∀E2 . ∀T1 . ∀T2 .
of (app E1 E2 ) T2
⊂
⊂
≫
of E1 (arr T1 T2 )
of E2 T1
2. ∀E. ∀T1 . ∀T2 .
of (lam T1 E) (arr T1 T2 )
⊂
(∀x.
≫
of x T1
⊃ of (E x) T2 )
∀x1 . ∀x2 .
of x1 x2
⊂ (∃E1 . ∃E2 . ∃T1 . ∃T2 . ⊤
∧ x1 =: app E1 E2
∧ ∃z1 . (of E1 z1 ∧ z1 =: arr T1 T2 ∧ ⊤)
∧ ∃z2 . (of E2 z2 ∧ z2 =: T1 ∧ ⊤)
∧ x2 := T2 ∧ ⊤)
∀x1 . ∀x2 .
of x1 x2
⊂ (∃E. ∃T1 . ∃T2 . ⊤
∧ x1 =: lam T1 E
∧ ∃z. ((∀x. ( ∀x′1 . ∀x′2 . ⊤
∧ x′1 =: x
∧ of x′1 x′2
∧ x′2 := T1 ∧ ⊤)
⊃ of (E x) z)
∧ z =: T2 ∧ ⊤)
∧ x2 := arr T1 T2 ∧ ⊤)
Fig. 9. Lc2 Compilation Example
mode of the predicate of is of ˇˆ — the first argument is input and the second output. The
result of compiling our two familiar clauses into Lc2 is shown in Figure 9. As in Section 3.2,
the moded compilation process offers ample opportunities for optimization: matches and
assignments with variables on both side and the corresponding existential quantification
can often be elided, and all occurrences of ⊤ can be optimized away.
It is instructive to rewrite these clauses with the two synthetic connectives introduced
earlier for Lc2 , again omitting ⊤ for readability:
Λof x1 . ∃E1 . ∃E2 . ∃T1 . ∃T2 . x1 =: app E1 E2
∧ call (of E1 ) =: (arr T1 T2 ) ∧ call (of E2 ) =: T1 ;
return T2
Λof x1 . ∃E. ∃T1 . ∃T2 . x1 =: lam T1 E
∧ ∀x. (Λof x′1 . x′1 =: x ; return T1 ) ⊃ call (of (E x)) =: T2 ;
return (arr T1 T2 )
5 Larger Source Languages
In (Cervesato 1998), we illustrated our original abstract logical compilation method on the
language of hereditary Harrop formulas. This language differs from Ls for the presence of
conjunction (formulas of the form A ∧ B) and truth (⊤). While our original treatment could
handle them easily (in a clause position, they were compiled to disjunctions and falsehood
respectively), the approach taken in Sections 3 and 4 does not support them directly. The
problem is that, as soon as we allow these connectives, clauses can have multiple heads (or
even none). Consider for example:
∀x. ∀y. q x y ⊃ (p1 x y ∧ (r x y ⊃ p2 x))
An Improved Proof-Theoretic Compilation of Logic Programs
17
This clause has two heads: p1 x y and p2 x. What should it be compiled to? To ensure
immediacy (embodied in the macro-rule g1 atm′ ), our compilation strategy produces a
pseudo clause applied to a residual, thereby exposing the (flattened) head of a compiled
clause as close to the top level as possible. How to achieve this now that there may be more
than one head?
One approach to dealing with this problem is to observe that ∧ distributes over (the
antecedent of) ⊃ and ∀. By doing so to the above example, we obtain the formula
(∀x. ∀y. q x y ⊃ p1 x y) ∧ (∀x. ∀y. q x y ⊃ r x y ⊃ p2 x)
Observe that it is a conjunction of Ls clauses. Each of them can now be compiled as
in Section 3 and the results can be combined by means of a disjunction. This approach
generalizes to the full language of hereditary Harrop formulas. It pushes the conjunctions
to the outside, leaving inner formulas resembling the clauses of Lc0 (conjunction and truth
in a goal position are left alone as they are not problematic). Clauses with no head (e.g.,
A ⊃ ⊤) are reduced to ⊤. These preprocessing steps can be implemented as a source-code
transformation or integrated in the compilation process.
The other abstract logic programming language examined in (Cervesato 1998) is the language of linear hereditary Harrop formulas, found at the core of Lolli (Hodas and Miller 1994)
and LLF (Cervesato and Pfenning 2002). The improved compilation process discussed in
this paper extends directly in the presence of linearity. Because linear hereditary Harrop formulas feature a form of conjunction and truth, the technical device just outlined
is needed to obtain workable compiled clauses.
6 Future Work
The discussion in Section 4 sets the stage for a nearly functional operational semantics
of well-moded programs. Indeed, given an atomic goal with ground terms in its input positions, proof search will instantiate its output positions to ground terms, if it succeeds.
Being in a logic programming setting, more than one answer could be returned. Indeed, for
well-moded programs, the clauses for a predicate implement a partial, non-deterministic
function. This observation informed the choice of the notation for the synthetic operators
we exposed: call p ť =: t̂ and Λp x̌. ∃y. (R; return t̂).
Now we believe that, in the case of well-moded programs, a more detailed operational
semantics that exposes variable manipulations using logical variables and explicit substitutions (and restricts the execution order) can bring this functional interpretation to the surface. This would provide a logical justification for the natural impulse to give well-moded
programs a semantics that is typical of functional programming languages, where atomic
predicates carry just input terms and from which the terms in output position emerge by a
process of reduction.
In future work, we intend to carry out this program by giving such a detailed operational
semantics to Ls as well as well-moding rules. The goal will then be to perform logical
transformations, akin to what we did in this paper, that expose this functional semantics
for well-moded programs. It would also allow us to prove formally that the operator =: of
Section 4 can indeed be implemented as matching rather than general unification.
18
I. Cervesato
Acknowledgments
This work was supported by the Qatar National Research Fund under grant NPRP 09-11071-168. We are grateful to Frank Pfenning, Carsten Schürmann, Robert J. Simmons and
Jorge Sacchini for the many fruitful discussions, as well as to the anonymous reviewers.
References
A ÏT-K ACI , H. 1991. Warren’s Abstract Machine: a Tutorial Reconstruction. MIT Press.
B ÖRGER , E. AND ROSENZWEIG , D. 1995. The WAM — definition and compiler correctness. In
Logic Programming: Formal Methods and Practical Applications, C. Beierle and L. Pluemer, Eds.
Computer Science and Artificial Intelligence, vol. 11. North-Holland, 21–90.
C ERVESATO , I. 1998. Proof-Theoretic Foundation of Compilation in Logic Programming Languages. In 1998 Joint International Conference and Symposium on Logic Programming — JICSLP’98, J. Jaffar, Ed. MIT Press, Manchester, UK, 115–129.
C ERVESATO , I. AND P FENNING , F. 2002. A Linear Logical Framework. Information & Computation 179, 1, 19–75.
C ERVESATO , I., P FENNING , F., WALKER , D., AND WATKINS , K. 2003. A Concurrent Logical
Framework II: Examples and Applications. Technical Report CMU-CS-02-102, Department of
Computer Science, Carnegie Mellon University, Pittsburgh, PA. March 2002, revised May.
D EBRAY, S. K. AND WARREN , D. S. 1988. Automatic mode inference for logic programs. Journal
of Logic Programming 5, 207–229.
H ODAS , J. S. AND M ILLER , D. 1994. Logic programming in a fragment of intuitionistic linear
logic. Information and Computation 110, 2, 327–365.
JAFFAR , J., M ICHAYLOV, S., S TUCKEY, P., AND YAP, R. 1992. An abstract machine for CLP(R).
In Proceedings of the SIGPLAN’92 Conference on Programming Language Design and Implementation — PLDI’92. San Francisco, CA.
M ILLER , D. AND NADATHUR , G. 1986. Higher-order logic programming. In Proceedings of the
Third International Logic Programming Conference, E. Shapiro, Ed. London, 448–462.
M ILLER , D., NADATHUR , G., P FENNING , F., AND S CEDROV, A. 1991. Uniform proofs as a foundation for logic programming. Annals of Pure and Applied Logic 51, 125–157.
NADATHUR , G. AND M ITCHELL , D. J. 1999. System description: Teyjus — a compiler and abstract machine based implementation of lambda prolog. In Sixteenth Conference on Automated
Deduction (CADE’99), H. Ganzinger, Ed. 287–291.
P FENNING , F. AND S CH ÜRMANN , C. 1999. System Description: Twelf — A Meta-Logical Framework for Deductive Systems. In Proceedings of the 16th International Conference on Automated
Deduction — CADE-16. Springer-Verlag LNAI 1632, Trento, Italy, 202–206.
P IENTKA , B. 2003. Tabled higher-order logic programming. Ph.D. thesis, Department of Computer
Science, Carnegie Mellon University.
RUSSINOFF , D. M. 1992. A verified Prolog compiler for the Warren abstract machine. Journal of
Logic Programming 13, 367–412.
S ARNAT, J. 2010. Syntactic finitism in the metatheory of programming languages. Ph.D. thesis,
Department of Computer Science, Yale University.
S TIRLING , C. 2009. Decidability of higher-order matching. Logical Methods in Computer Science 5, 3.
WARREN , D. H. D. 1983. An abstract Prolog instruction set. Technical Note 309, SRI International,
Menlo Park, CA. Oct.
WATKINS , K., C ERVESATO , I., P FENNING , F., AND WALKER , D. 2003. A Concurrent Logical
Framework I: Judgments and Properties. Technical Report CMU-CS-02-101, Department of Computer Science, Carnegie Mellon University, Pittsburgh, PA. March 2002, revised May.
| 6cs.PL
|
1
Power-Traffic Coordinated Operation for Bi-Peak Shaving and Bi-Ramp Smoothing
–A Hierarchical Data-Driven Approach
Huaiguang Jiang, Member, IEEE, Yingchen Zhang, Senior Member, IEEE, Yuche Chen, Changhong
Zhao, Member, IEEE, Jin Tan, Member, IEEE
arXiv:1712.00211v1 [cs.SY] 1 Dec 2017
National Renewable Energy Laboratory, Golden, CO 80401
Abstract—With the rapid adoption of distributed photovoltaics
(PVs) in certain regions, issues such as lower net load valley during the day and more steep ramping of the demand after sunset
start to challenge normal operations at utility companies. Urban
transportation systems also have high peak congestion periods
and steep ramping because of traffic patterns. We propose using
the emerging electric vehicles (EVs) and the charing/discharging
stations (CDSs) to coordinate the operation between power
distribution system (PDS) and the urban transportation system
(UTS), therefore, the operation challenges in each system can be
mitigated by utilizing the flexibility of the other system.
The proposed operation approach is designed hierarchically
consists of a higher and a lower level. In the higher level, we
assume integrated operation of both the PDS and UTS, and target
of the operation is to minimize the social cost. Meanwhile, the
target for the EVs and the CDSs as customers is to minimize their
own expenditures. Then, there exists an equilibrium between
two targets to determine the optimal charging/discharging price.
In the lower level, the temporal & spatial models of PDS
and UTS are developed to provide a detailed analysis of the
power-traffic system. Specifically, the PDS is built with a threephase unbalanced AC power flow model, the optimal power flow
(OPF) problem is relaxed with the semidefinite relaxation programming (SDP), and solved with alternating direction method
of multiplier (ADMM). A dynamic user equilibrium (DUE)
problem is formulated for the UTS, which is based on the static
user equilibrium (SUE) with additional constraints to ensure
a temporally continuous path of flow. The EVs and the CDSs
function as reserves for both the PDS and UTS, and the state of
charge (SOC) is considered to optimize the charging/discharging
schedule and reduce the impacts to the PDS. We conducted the
simulation and numerical analysis using the IEEE 8,500-bus for
the PDS and the Sioux Falls system with about 10,000 cars for
the UTS. Two systems are simulated jointly to demonstrate the
feasibility and effectiveness of the proposed approach.
Index terms— Power distribution system, duck curve, urban
transportation system, electrical vehicle, charging/discharging
station, state of charge, dynamic user equilibrium, traffic
congestion, traffic pattern, optimal power flow, distributed
computation
N OMENCLATURE
Functions
t
FUT
Y
t
FCSO
t
FCDS
fPt S
t
fUT
S
t
fW
T
QtP DS
QtDP DS
QtDUT S
FPt DS
fka (xtka )
Parameters
G UT S
Pru su
Cka
G P DS
Υ, Υ
γ1
γ2
Abbreviations
PDS
UTS
CDS
SUE
DUE
OPF
SDP
ADMM
EV
SOC
Power distribution system
Urban transportation system
Charging/discharging station
Static user equilibrium
Dynamic user equilibrium
Optimal power flow
Semidefinite relaxation programming
Alternating direction method of multiplier
Electrical vehicle
State of charge
Objective function of the utility and customer in time interval t, t ∈ Dt
Objective function of the customer in time
interval t, t ∈ Dt
Objective function of the smart charging/discharging in a CDS
Cost function of PDS in time interval t
Cost function of UTS in time interval t
Cost function of charging/discharging waiting time
Netload of PDS in time interval t
Total charging/discharging power of all
CDS
Total EVs for charging/discharging
Objective function of OPF in PDS
Travel time function on link ka in time
interval t, where traffic flow is xtka
̺1 , ̺2
̺3 , ̺4
χi1
The graph for UTS with a node set and
a link set: G UT S = [V UT S , E UT S ], where
V UT S = {1, 2, · · · , nV U T S }, and E UT S =
{1, 2, · · · , mE U T S }
A set of path used to connect the OD
S
and
pair, which is defined as ru ∈ VrUT
u
UT S
UT S
UT S
su ∈ Vsu , Vru and su ∈ Vsu are the
original node set and destination node set,
respectively
The traffic flow capacity on link ka
The graph for PDS with a node set and a
link set: G P DS = [V P DS , E P DS ], where
V P DS = {1, 2, · · · , nV P DS }, and E P DS =
{1, 2, · · · , mE P DS }
The upper bound and lower bound of SOC
Weight factor between PDS and UTS
Iteration step coefficient for gradient descent
The electrical power price and the congestion fee
The EV parking ratios to the CDSs
The capacity of CDS i1
Variables
θkt a
π ∗t
The traffic flow on link ka in time interval
t
The optimal charging/discharging price in
time interval t
qrdtu su
hdpr1u su
δpdr1ut su ka
Viφt
Iiφt
sφt
i
ziφ
Siφt
viφt , iφt
i
xt,k , y t,k , λt,k
t1
Cev,i
3
Number of vehicles from ru to su departing
in time interval d via any path, d ∈ Dt , Dt
is the time interval set
Number of vehicles assigned to path pru su
with departing time interval d1
A 0-1 variable to indicate in time interval t,
whether the trip from ru to su is assigned to
path pru su ka via link ka departing in time
interval d1
Complex voltage on bus i with phase φ, φ
⊆ {a, b, c}
Complex current on bus i with phase φ, φ
⊆ {a, b, c}
φt
Power injection on bus i, where sφt
i = pi
φt
+ iqi
The complex impedance matrix ziφ = riφ +
ixφi
The complex power from bus i to bus i1 ,
where bus i1 is the ancestor of bus i, Siφt
= Viφt (Iiφt )H , H indicates the hermitian
transpose
viφt = Viφt (Viφt )H , liφt = Iiφt (Iiφt )H
The kth iterative variables of ADMM for
distributed OPF computation
The discharging (positive) and charging
(negative) speed of EV i3 at time t1
and transportation information. In [18], based on locational
marginal pricing (LMP), an optimal deployment of charging
stations is proposed for EVs. In [19], an optimal traffic-power
flow is proposed with the wireless charging technology and
congestion toll manipulation.
In this paper, a two-step power-traffic coordinated operation approach is proposed to address the bi-peak and biramp problems for both PDS and UTS. It also provides a
multifunction platform for the researches such as emission
reduction, multisystem integration, and future city planning.
The main contributions of this paper are:
1) Considering the complexity of the power-traffic system,
a two-step coordinate approach is designed from the
higher level to lower level to operate the system in
spatial and temporal domain hierarchically. In the higher
level, both the PDS and the UTS are regulated and
treated together as an utility to minimize the social
cost. The EVs and CDSs are treated as customers to
minimize the expenditure. Then, an equilibrium exists
between the utility and the customers to determine the
operation variables such as optimal charging/discharging
price, total electrical demand, and total available EVs. In
the lower level, the detailed models of PDS, UTS, EVs
& CDS are considered to specifically determine these
variables in spatial and temporal domains.
2) In the lower level, the PDS is built with the branch
flow model, which is equivalent to the classic businjection model and more suitable for the PDS with
radial topology [20]. There are several developed approaches for solving an optimization problem associated
with a PDS, usually called OPF such as DC OPF [21],
genetic algorithm [22], [23], and OPF relaxation with the
second-order cone program (SOCP) [24]. Based on [25],
[26], we relax an OPF with the SDP for the threephase unbalanced PDS. Meanwhile, several distributed
algorithm are designed to solve the OPF such as dual
decomposition [27] and predictor corrector proximal
multiplier (PCPM) [28]. Based on [25], [29], the ADMM
is applied to decompose the OPF problem and reduce
computation time.
3) In the lower level, for the UTS, the DUE is widely
used to estimate the traffic volume, congestion time,
and travel cost for each link [30]. Based on the assumption that all travelers can choose the best route
to minimize the cost of transportation, in the Wardrop
user equilibrium (UE), the travel times of all paths are
equal for each original-destination (OD) pair and less
than any unused paths [31], [32]. In [19], a static user
equilibrium assignment is used to integrate with the
PDS. Considering that the EV behaviors can be impacted
by the charging/discharging prices, a DUE is applied to
keep the temporal domain continuous for the UTS.
4) In this paper, the EVs and the CDSs are designed as
the “reserves” for both PDS and UTS, respectively.
Considering the SOC [33], [34], an optimal charging/discharging schedule is designed to meet the requirements of PDS and UTS and reduce the impacts to
I. I NTRODUCTION
Rooftop photovaltaic (PV) gained a foothold in many power
systems because of its continuously declining cost [1]–[3].
It brings many benefit to costumers such as reduced energy
cost; however, it also presents unprecedented challenges to
power systems operation and control. The upper right-hand
plot in Fig. 1 shows the so-called “duck curve”, which has
been recorded in some high-distribution PV regions such as
California [4]–[6]. The peak solar generation in the middle
of the day sinks the netload to lower valley and then when
the peak load occurs right after sunset, a huge volume of
energy demand ramps up in a short time frame. This creates
the artificial peak and ramping that are costly to balance using
the current power system assets.
It is not surprising that other man-made complex systems
also suffer similar issues. As shown in the lower right-hand
plot in Fig. 1, the transportation system also has high peaks
and steep ramps due to fairly constrained traffic patterns at the
rush hours in urban areas [7]–[9]. Thanks to the spawn of EVs
and the widely located CDSs, the two originally independent
systems: the PDS and the UTS can be coupled together.
Specifically, the increasing number of EVs can be seen
as a significant amount of distributed and highly manipulatable small reserves that can be used to provide demand
response, enhance the system reliability, and provide other
services for the power systems [10]–[16]. In these studies,
the geographical information and transportation information
are ignored, because the EVs are treated as the aggregated
loads/reserves. In [17], the optimal placement of the charging stations is studied with the geographical information
2
Fig. 1. The main idea and the bi-peak and bi-ramp problems in a power distribution system and a transportation system in a urban area.
tion VIII.
PDS. The test bench consists of real systems, the PDS is
the IEEE 8,500-bus distribution system, and the UTS is
the Sioux Falls transportation system with about 10,000
EVs, to demonstrate the proposed approach.
Based on Fig. 1, we acknowledge that cybersecurity is
a critical aspect to consider in the proposed power-traffic
system operation. Also, the large volume of data generated
by the proposed power-traffic system requires a high-speed,
flexible and reliable communication network. Because the
proposed power-traffic system contains two complex systems,
many industrial communication network infrastructures for
PDS and UTS operations need to be attached, such as 3G,
4G, and WiFi [35]. The multi-network integration brings a
lot of challenge for the 1). real-time monitoring & anomaly
detection [36], [37], 2). data transmission, storage, & security,
3). attack analysis & mitigation [38]–[41]. In this paper, we
assume that all the messages such as electrical power price,
traffic congestion information, and various control signals can
be transmitted correctly in real-time without any cybersecurity
issues.
The paper is organized as follows. In Section II, the
flowchart of the proposed approach is introduced. In Section III, the equilibrium between the utility and the customers
is designed to determine the optimal system variables. In
Section IV, based on SUE, a DUE is applied to model the
UTS and compute the dynamic traffic flow. In Section V, based
on the branch flow model, the OPF problem is relaxed with
SDP and solved with ADMM. In Section VI, considering the
SOC and the EVs behaviors, an optimal charging/discharging
schedule is designed to reduce the impacts to the PDS. In
Section VII, the numerical results are presented to validate
the proposed approach. The conclusion is presented in Sec-
II. T HE F LOWCHART OF THE P ROPOSED A PPROACH
The proposed two-step power-traffic operation approach is
shown in Fig. 2. In the higher level, the power-traffic system
consists of two major parts: the utility part (taking the PDS
and the UTS as a whole) shown in green, and customer
part (taking the CDSs and related EVs) shown in yellow.
At time interval t, the duck curve of the netload is recorded
with the renewable energy generation data and end user load
data in PDS. Meanwhile, the travel data is collected with the
traffic monitoring data and congestion model in UTS. Then,
as shown in the two green blocks, the PDS and UTS are
treated as a regulated utility to minimize the social cost. In the
yellow block, the EVs and CDSs are treated as the customers
to minimize their expenditures. There exists an equilibrium
to determine the system operation variables such as optimal
charging/discharging prices, total demand electrical power, and
total required EVs. These variables are fitted into the lower
level as inputs.
In the lower level, the power-traffic system consists of three
major parts: the PDS part, the EVs & CDSs part, and the UTS
part, which determine the system operation variables in spatial
and temporal domains with detailed models. First, in the UTS,
to accurately simulate the dynamic traffic system in spatial
and temporal domain, the DUE is built based on SUE with
temporal continuous constraints. The proposed DUE problem
can be transfered into a convex problem and solved with
optimal results. Second, in the PDS, the objective function
FP DS is designed to minimize the cost of the PDS, which
is built with a three-phase unbalanced branch flow model.
Based on the SDP relaxation, FP DS can be minimized in an
3
Synchronized time and unified GIS information
Load
data
Net load
(Duck curve)
EV
EV
CDS
CDS
CDS
Utility PDS
Traffic Flow
Monitoring
Congestion
EV
Traffic flow curve
Upper Level
Renewable
energy
Utility UTS
Customers
An equilibrium between the utility and customers
SCADA
AMI
Demand electrical
power
Required power,
charging/discharging
price
PDS
information
SDP
relaxation
Three-phase unbalanced power
distribution system
(Branch flow model)
Power distribution system
Road
Battery
SOC
Dynamic traffic
assginment
(DUE Model)
Model of
EVs
Charging/
discharging power
Optimal Power Flow
(ADMM)
GIS
Charging/
discharging price
and schedule
Charging/
discharging
speed
GIS
Total available EVs
Bi-direction charging/
discharging and schedule
Congestion
model
Congestion, available
EVs
Temporal continuosly
constraints
User
Equilibrium
EVs & charging/discharging stations
Lower Level
Optimal charging/
discharging price
Urban transportation system
Fig. 2. Flowchart of the proposed power-traffic coordinate operation approach.
number of available EVs on CDS i1 . The constraints of (1)
are illustrated as follows:
OPF problem, which is solved with ADMM. Third, with the
specified electrical power demands and available EVs for each
CDS, an objective function FCDS is designed to minimize the
charging/discharging impacts to the PDS while considering the
SOC.
In addition, in this paper, it is assumed that all the drivers
know the traffic information very well, and can get the optimal
charing/discharging price through the wireless communication
immediately. The user-equilibrium in UTS network can be
reached in a very short time.
Pit1 ≤ Pit1 ≤ Pit1 ,
(2a)
0 ≤ γ1 ≤ γ1
(2b)
Nit1
(2c)
0≤
Nit1
≤
fPt S is defined as a convex function [42], which determine
the PDS cost in time interval t, QtP DS is the netload of PDS,
and QtDP DS (Pit1 ) is the total charging/discharging power of
all CDSs, which can be defined as
III. E QUILIBRIUM D ESIGN B ETWEEN U TILITIES AND
C USTOMERS (H IGHER L EVEL )
QtDP DS (Pit1 ) =
X
Pit1
(3)
i1 ∈NCDS
t
where NCDS is a set of CDS. fUT
S is defined as a convex
function, which determines the UTS cost in time interval
t, QtUT S is the traffic load, and QtDUT S (Nit1 ) is the total
t
EVs for charging/discharging. fW
T (τ1 ) is the time cost with
charging/discharging time τ1 . Considering the number of EVs
is very large, QtDUT S (Nit1 ) can be estimated as
A. Equilibrium Design
In this paper, we focus on a small city with tens of thousands
of cars, which means any single user behavior has a very weak
impact on the equilibrium design in the higher level. The utility
consists of PDS and UTS. A customer consists of a CDS and
related EVs.
Utility Objective (Minimize Social Cost)
t
As shown in (1), FUT
Y is the objective function of the
utility, which is designed to minimize the social cost with the
variables Pit1 and Nit1 , t is a time interval t ∈ Dt . Pit1 is the
electrical power charging (negative) or discharging (positive)
at CDS i1 . γ1 is the weight coefficient of UTS. Nit1 is the
QtDUT S (Nit1 ) =
X
i1 ∈NCDS
Nit1 =
X
i1 ∈NCDS
Pit1
C1t
(4)
where C1t is the average charging/discharging speed of all
t
the EVs. fW
T is the waiting cost when an EV is parking
for charging/discharging, which is also defined as a convex
4
t
FUT
Y = min
t
t
Pi ,Ni
1
1
t
t
t
t
t
t
t
))
))
+
f
(τ
Q
(N
(Q
−
Q
(N
fPt S (QtP DS − QtDP DS (Pit1 )) + γ1 fUT
1
WT
DUT S
i1
S
UT S
DUT S
i1
function. In summary, the proposed utility objective function
t
FUT
Y is convex, which indicate an optimal point existed.
Customer Objective (Minimize Own Expenditure) Based
on (4), the objective function of a customer (a CDS and related
EVs) can be designed as follows:
t
t
FCSO
= min
fW
T (τ1
t
Pi
X Pit
1
i1
1
C1t
) − π t Pit1
IV. DYNAMIC U SER E QUILIBRIUM A SSIGNMENT
(L OWER L EVEL )
Compared with the SUE, the DUE can accurately describe
the traffic dynamics in a set of successive time intervals. In
this paper, the DUE is based on the SUE with a temporal
generalization [32], [46].
A. Basic Model of UTS
(5b)
The UTS can be represented in a graph with a node
set and a link set: G UT S = [V UT S , E UT S ], where V UT S
= {1, 2, · · · , nV U T S }, and E UT S = {1, 2, · · · , mE U T S }. The
S
origin and destination (OD) pair can be defined as ru ∈ VrUT
u
UT S
and su ∈ Vsu , a set of paths Pru su is used to connect the
OD pair, and the traffic flow between the OD pair can be
represented as qrdu su with departing time dt ∈ Dt , where Dt
is a set of all time intervals. Then, the congestion model can
be presented as follows [47]:
θt
(8)
fka (θkt a ) = fk0a 1 + 0.15( ka )4
Cka
where the objective function of the customer consists of two
parts: the time consumption in the CDSs and the benefit/cost
in charging/discharging. π t is the electrical power price, and
π t Pit1 is the benefit (by discharging, positive) or cost (by
charging, negative) at CDS i1 . As discuss above, the objective
t
function of the customer FCSO
is also convex.
B. Solutions
exists, and the optimal price
An equilibrium π ∗t and Pi∗t
1
π can be computed as [43], [44]
∗t
where fka (θkt a ) is the travel time (travel impedance) when the
traffic flow at link ka ∈ KA is θkt a in time interval t ∈ Dt , fk0
is the initial travel time, and Cka is the traffic flow capacity.
X Pi∗t
γ1 ′t
t
1
+ t fUT
−
π =
S (QUT S −
t )
C
C
1
1
i1
i1
(6)
where the optimal price depends on the PDS part fP′tS and the
′t
UTS part fUT
S . Based on the gradient descent, a distributed
algorithm is proposed to optimize the utility and the customers
jointly the equilibrium. Specifically, the objective function of
each customer can be computed independently to the optimal
price π ∗t . At k1 -th iteration:
X
fP′tS (QtP DS
)
Pi∗t
1
B. Dynamic User Equilibrium
Based on the SUE, the DUE can be built as follows [46]:
t
JDUE = mint
ka ,θka
π k1 ,t =
X k ,t γ 1
X Pik1 ,t
1 ′t
′t
t
t
1
1
+ t fUT S QUT S −
Pi1
fP S QP DS −
t
C
C
1
1
i
i
1
1
s.t. θkt a =
(π
k1 ,t
=
Pik11 ,t
, Pik11 +1,t )
+ γ2
=
X Pit
τ1 ′t
k1 ,t
1
f
(τ
1
W
T
t )−π
C1t
C
1
i
[π k1 ,t , Pik11 +1,t ]̟
1
qrdu1 su =
fkta (w)dw
X
(9a)
(9b)
hdpr1u su
(9c)
fkta (θkt a )δpdr1ut su ka
(9d)
pru su ∈Pru su
bdpr1u su =
(7b)
θka
X X Z
ka ∈KA t∈Dt 0
X
X
hdpr1u su δpdr1ut su ka
pru su ∈Pru su d1 ∈Dt
(7a)
Pik11 +1,t
UTS
(5a)
s.t. (2a), π t ≤ π t ≤ π t
∗t
FOR
(1)
X X
t∈Dt ka ∈Ka
where hdpr1u su is the number of vehicles in path pru su with
departing time interval d1 , bdpr1u su is the travel time from ru
to su via path pru su , and δpdr1ut su ka is a 0-1 variable and can
be defined as
(7c)
where the optimization stepsize can be set as 0 ≤ γ2 , and the
algorithm can converge when γ2 is small enough [45]. [·]̟
means that the results are projected onto the set ̟ defined by
(5b).
In summary, in this section, the optimal price π ∗t can be
gives
computed in the higher level, and the optimal results Pi∗t
1
the upper bound of power injection in (20). In the following
sections, the models of the UTS, PDS, and CDS are built
to provide detailed information in both temporal and spatial
domains about how to operate the power-traffic system in the
lower level.
δpdr1ut su ka =
1
0
If in time interval t, link ka is in
path pru su with departing time d1
otherwise
(10)
In time interval t, the traffic flow θkt a on link ka is given as
in (9b), which equals to the sum of the traffic flows hdpr1u su ≥ 0
via link ka for any departing time interval t ∈ Dt and any path
pru su ∈ Pru su . The number of vehicles from ru to su is given
5
Algorithm 1 DUE solution
Step 1: Initialization. Set the iteration number n1 = 1, the
.
threshold ε1 . Update the new demand as qrdtu su = qrdt−1
u su
Then using all-or-nothing assignment [32], [48] to generate
θkt a .
as in (9c), which equals to the sum of traffic flows hdpr1u su on
any assigned path pru su ∈ Pru su . The total time consumption
of path pru su from ru to su is given as in (9d), which keeps the
temporal continuous traffic flow for all time intervals t ∈ Dt
with a temporal unique constraint:
X
δpdr1ut su ka = 1
1
1
),
= fkta (θkt,n
Step 2: Update the travel time with θkt a : fkt,n
a
a
which is the beginning of the loop (Step 2 to Step 6).
(11)
t∈Dt
which ensures that a vehicle cannot be assigned on two links
in a certain time interval t.
Step 3: Determine the descend direction. Find the shortest
1
using all-or-nothing assignment to generate
path with fkt,n
a
t,n1
φ , which is the shortest route pattern.
C. Solutions
Step 4: Determine the iteration step ζ n1 , which can be
designed as:
Considering the temporal continuous traffic flow, in time
interval t, take the initial state fk0a in (8) as fkt−1
. In this
a
paper, the test bench is based on Sioux Falls network, and the
departing time is the same during each time interval t. Then,
in time interval t, the DUE problem can be transfered as a
SUE problem. The SUE is based on (9a), (9b), and (9c), and
the Lagrangian can be generated as (12).
The Hessian matrix of (9a) is positive definite, which
indicate the convexity. The optimal travel flow θk∗ta can be
computed with the KKT conditions [32], [43]. Therefore, the
solution of the UTS in time interval t can be designed as in
Algorithm 1.
η
minζ n1
1
1
)
+ ζ n1 (φt,n1 − θkt,n
s.t. η = θkt,n
a
a
0≤ζ
=
viφt
−
j∈Ci
φ φt
2(ri Pi +
φ 2 φt
xφi Qφt
i ) + |zi | li ,
Step 6: Compare the result to the threshold ε1 as:
qP
t,n1 +1
1 2
)
− θkt,n
ka (θka
a
≤ ε1
P
t,n1
ka θka
(13b)
(13c)
(14)
(15)
where if (15) is fulfilled, the result of traffic flow in time
1 +1
; otherwise, go back to step 2, with
interval t is θkt,n
a
n1 = n1 + 1.
B. Objective Function and SDP Relaxation of OPF in PDS
The objective function contains two major parts: the generation cost and system line loss, which can be designed as
follows:
(16a)
(16b)
FPt DS =
φt
φt
where sφt
i = pi + iqi indicates the power injection on bus
φt
φt φt H
i, Si = Vi (Ii ) is the complex power from bus i to bus
i1 , viφt = Viφt (Viφt )H , liφt = Iiφt (Iiφt )H , ziφ = riφ + ixφi , Ci is
the set of buses that have a common ancestor bus i, and H
indicates the Hermitian transpose.
In addition, according to [50], the matrix Mt is defined as
" φt
#
vi
Siφt
t
M =
(17)
(Siφt )H liφt
t
≤1
1
1
1 +1
)
+ ζ n1 (φt,n1 − θkt,n
= θkt,n
θkt,n
a
a
a
The PDS can be represented in a radial graph as: G P DS =
[V
, E P DS ], where V P DS = {1, 2, · · · , nV P DS }, and E P DS
= {1, 2, · · · , mE P DS }. Similar as UTS, the PDS also contains
the temporal characteristics, the superscript (·)t , t ∈ Dt .
Considering the three-phase unbalanced system, the phase configuration is defined as φ ⊆ {a, b, c}. For a radial distribution
system, there is an unique i1 as the ancestor bus for bus i,
and the branch between bus i1 and i can be named as i [25],
[49]. The branch flow model in time interval t can be defined
as:
viφt
1
n1
1 +1
with
Step 5: Update θkt,n
a
P DS
(Sjφt − zjφt ljφt )),
(13a)
The Golden-section search is employed to determine ζ n1 in
a short time.
A. Branch Flow Model
X
fkta (ω)dω
ka 0
V. O PTIMAL P OWER F LOW OF P OWER D ISTRIBUTION
S YSTEM (L OWER L EVEL )
φt
sφt
i = diag(Si −
XZ
X X
φt
2
[α1 (pφt
i ) + β1 (pi )]
(18)
i∈V PDS φi
where α1 and β1 are two weight coefficients for the generation
cost and system line loss, respectively.
The constraints of the voltage viφt and current liφt are
illustrated as follows:
t
where in (17), M 0 ( M is positive semidefinite) and
Rank(Mt ) = 1.
6
viφ ≤ viφt ≤ viφ ,
(19a)
liφ ≤ liφt ≤ liφ ,
(19b)
t
L1 (θ, λ1 , µ1 ) =
θka
X X Z
ka ∈KA t∈Dt 0
X
X X
hdpr1u su (12)
hdpr1u su δpdr1ut su ka + µd1,ru su qrdu1 su −
fkta (w)dw + λt1,ka θkt a −
In this paper, the injection power sφt
i plays an important
role in active control and reactive control for the OPF.
sφi
≤
sφt
i
≤
Pi∗t ,
min
Sit ,sti ,vit ,lti
(20)
min
FPt DS
s.t. (16), (19), and Mt 0
FPt DS =
min
Sit ,sti ,vit ,lti
s.t. (16), (19), and M 0
where Pi∗t is the optimal power injection in (7b). Considering
the relationship with the UTS, the optimal injection power s∗t
i1
decides the number of EVs in CDS i1 (Here for CDS, i1 = i
for the PDS and UTS, which can be explained as Fig. 3 with
yellow circles).
According to the SDP relaxation introduced in [25], [50],
the constraint Rank(Mt ) = 1 of (17) can be removed, and
the relaxed objective problem is shown as follows:
φt φt
φt
sφt
i ,vi ,li ,Si
Pru su
Pru su d1 ∈Dt
φt,(x)
φt,(y) φt,(x)
, li,i1
Si,i1 = Si
φt,(y) φt,(x)
φt,(x)
, li
= Si
Si
φt,(y)
φt,(x)
= si
si
=
=
X
FPt DS,i
(24a)
i∈V PDS
φt,(y) φt,(x)
, vi,i1
li
φt,(y) φt,(x)
, vi
li
(24b)
=
=
φt,(y)
vi
φt,(y)
vi
(21a)
(21b)
(26a)
y t,k+1 ∈ arg min Lρ (xt,k+1 , y t , λt,k )
(26b)
λt,k+1 = arg min λt,k + ρ(xt,k+1 − y t,k+1 )
(26c)
y∈Ky
C. OPF Solution with ADMM
In the power-traffic system, the OPF computation of the
three-phase unbalanced PDS is the bottleneck of the whole
system operation. Firstly, because the real PDS usually contains a lot of feeders, which brings high computation loads
for the system. Secondly, the PDS is a highly dynamic
system, which requires a short computation time. Therefore,
the ADMM, a distributed algorithm is proposed to solve this
problem in a short time, which is based on Lagrange multiplier
with an additional penalty term (augmentation term).
The standard form of the ADMM can be formulated as
follows:
s.t. x ∈ Kx , y ∈ Ky , and x = y
(24e)
xt,k+1 ∈ arg min Lρ (xt , y t,k , λt,k )
x∈Kx
x,y
(24d)
where the variables (Sit , sti , vit , lit ) is maintained in bus i,
which can be seemed as an local agent. In the ADMM, the
x-update and y-update indicate as (·)(x) and (·)(y) .
Then, as in [25], [29] the augmented Lagrangian of (24)
can be derived as (25).
The detail formulation of fPt DS,i,y and fPt DS,i,x can be
found in [25], and at iteration k, the variables x, y, and λ
are illustrated as follows:
where (21) is a convex problem and can be solved with
ADMM.
min fA (x) + gA (y)
(24c)
x∈Kx
Considering the time intervals of the UTS, the ADMMbased distributed OPF computation architecture is built to
dramatically reduce the time consumption to the proposed
power-traffic coordinated operation approach.
Therefore, in addition to the designed user equilibrium in
the higher level, the detail models of UTS and PDS are build
to provide the spatial and temporal information in the lower
level. In the next section, a smart charging/discharging strategy
is proposed for the EVs and bidirectional CDSs to further
reduce the impact to PDS.
(22a)
VI. S MART EV C HARGING / DISCHARGING (L OWER
L EVEL )
(22b)
In this paper, the EVs and CDSs act as the reserves for both
PDS and UTS. In this section, we take CDS i1 as an example,
other CDSs contain the same performances. In time interval
t, nti1 EVs are parked in CDS i1 with parking time τ2 , which
is defined in (1). In real world, nti1 depends on many factors,
for example, different habits of the drivers, which are beyond
the scope of this paper. Here, nti1 is determined as follows:
where fA and gA are convex, and Kx and Ky are convex sets.
With an additional penalty term ρ/2||x − y||22 , the efficient
augmented Lagrangian function is shown as follows [29]:
ρ
Lρ (x, y, λ) = fA (x) + gA (y)+ < λ, x − y > + ||x − y||22 ,
2
(23)
where 0 ≤ ρ is a coefficient for convergence, and ρ2 ||x − y||22
is the quadratic penalty term to increase the converge speed.
Therefore, the relaxed objective function (21) can be formulated as
7
nti1 =
t t
∗t
≥ ̺1 + ̺2
̺3 ∗ min{s∗t
i1 /C1 , θka } If π
t t
∗t
̺4 ∗ min{si1 /C1 , θka } If ̺1 ≤ π ∗t ≤ ̺1 + ̺2
0
If ̺1 ≥ π ∗t
(27)
X
Ltρ (xt , y t , λt ) =
FPt DS,i + < λt , xti − yit > + < µti , xti,i1 − yit > +
X
FPt DS,i + < λt , xti − yit > +
X
where s∗t
i1 is the optimal power injection, which is determined in the OPF solution (26). C1t is the average discharging/charging speed of EVs. ̺1 is the electrical power price,
̺2 is the congestion fee, ̺3 is the ratio for EV parking to the
CDSs. nti1 < χi1 , χi1 is the capacity of CDS i1 .
During the parking time τ2 , the stochastic EVs charging/discharging can impact the stability of the PDS. Considering the SOC, a smart EVs charging/discharging approach is
proposed to reduce the impact to PDS as follows:
t
FCDS
= min
t
1
Cev,i
3 t1 ∈τ2
nt
t1
Cev,i
3
(28c)
i3
X
t1
Cev,i
, 0 ≤ i3 ≤ nti1
3
(28d)
t1 ∈τ2
Υ ≤ SOC ≤ Υ,
(28e)
t1
Cev,i
3
where
is the discharging (positive) and charging (negt1
t1
ative) speed of EV i3 at time t1 . In (28c), PCDS
(Cev,i
) is
3
the netload at CDS i1 , which equals to original load of PDS
QtP1DS minus the total charging/discharging power. In (28d),
2
Qτev,i
is the total demand of charging/discharging of EV i3 ,
3
which contains a SOC constraint as (28e), Υ is the ratio of
SOC.
VII. N UMERICAL S IMULATION
AND
(25b)
The simulation flowchart is shown in Fig. 4, and the
corresponding description is shown in Algorithm 2. In Step 2,
a distributed algorithm is used to reduce the time consumption
of the equilibrium computation. In Step 3, for each time
interval t, the DUE problem is transferred as a SUE problem.
The the Golden-section search is employed to determine the
iteration step ζ n1 in a short time for the UTS with about 10,000
EVs and 11 CDSs. In Step 4, because the test bench is based
on the IEEE 8,500-bus PDS, the ADMM is implemented to
computed the OPF in distributed manner with the constraints
from Step 3. In Step 5, for each CDS, the charging/discharging
are independent, the smart EVs charging/discharging can be
computed in parallel, which also helps to reduce the computation time. The computation time analysis is shown in Fig 5.
The two curves indicate the average computation times of
the proposed approach are less than 50 s with 100 scenarios.
Compared with the time interval 15 minutes, the proposed
approach is quick enough to support continuously operations
for the power-traffic systems.
(28b)
i1
X
(y)t
(25a)
A. Flowchart of the Simulation
(28a)
t1
t1
PCDS
(Cev,i
) = QtP1DS −
3
(x)t
ρ
> + fPt DS,i,y
2
ρ
> + fPt DS,i,x
2
̺1 = 45 $/M W h, and the congestion fee ̺2 = 2 $/h [54],
[55]. The EVs parking ratio ̺3 = 0.8, ̺4 = 0.3. The parking
capacity for each CDS is χ1 = 100. The simulations are
executed using a server with 3.60 GHz Intel Xeon CPU and 32
GB RAM, and the software are Python, MATLAB, MATLAB
global optimization toolbox, and parallel computing toolbox.
The communication network is good enough to transmit all
the information in real-time without any cybersecurity issues.
2
t1 +1
t1
t1
t1
PCDS (Cev,i3 ) − PCDS (Cev,i3 )
t1
≤ Cev
s.t. Cev ≤ Cev,i
3
2
Qτev,i
=
3
(y)t
< µtj , xtj,i − yjt > + < γit , vi1 ,i − vi1
j∈Ci
i∈V PDS
X
(x)t
< γjt , vi,j − vi
j∈Ci
i∈V PDS
=
X
B. Bi-Peak Shaving and Bi-Ramp Smoothing
R ESULTS
As shown in Fig. 6(a), the blue curve is the netload peak of
PDS from 17:00 to 22:00, and the orange curve is the result of
proposed approach. The peak load decreases from 78 MW to
72 MW with 15% EVs discharging. The yellow curve is the
traffic delay peak in UTS, and the purple curve is the reduced
traffic delay with the proposed approach. The total peak traffic
delay time of UTS decreases from about 4,100 hours to 2,900
hours. In Fig. 6(b), it is also clear that from 8:00 to 15:00,
the netload of PDS increases from 48 MW to 51 MW, and the
total traffic delay of UTS decreases from 3,500 hours to 2,800
hours. From Fig. 6, it is clear that the proposed approach can
benefit both PDS and UTS to reduce the bi-peak and smooth
the bi-ramp.
In Fig. 7, the detailed traffic congestion information of
a congestion scenario is shown with the traffic maps. The
OD pair of this scenario is shown in Table I. The red and
pink arrows indicate the traffic congestions in the roads. The
traffic congestion information without the proposed approach
As shown in Fig. 3, the test bench is based on the IEEE
8,500-bus PDS and the Sioux Falls UTS [51], [52], which
cover the similar geographical areas. It is assumed that a small
city with 10,000 EVs has the bi-peak and bi-ramp problem,
the ”duck curve” data is based on [53] and the traffic behavior
data is based on [7]. 11 CDSs are located in node 2, 3, 5, 8,
10, 11, 17, 18, 20, 22, 23, which are shown with yellow circles
and illustrated in both PDS and UTS, respectively. The time
interval t is 15 minutes for the power-traffic systems. The
PDS cost function fPt DS is defined as quadratic function as
t
in [42], and the UTS cost function fUT
S is defined similarly.
t
t
Both fP DS and fUT S are convex, and the quadratic form
indicates the utility cost increase fast with the increasing
of PDS load and UTS load. As discussed above, the utility
t
objective function FUT
Y is convex. The weight factor γ1
equals to 1, which indicates the same importance of PDS
t
and UTS. Similarly, the customer objective function FCSO
is also set as a convex function. The electrical power price
8
7
18
6
8
16
5
9
10
15
22
21
4
11
14
23
24
3
12
18
2
2
17
19
20
17
8
20
22
10
5
23
11
3
1
13
Fig. 3. The test bench consists of a PDS and a UTS.
Algorithm 2 Simulation Process
Step 1: Initialization and data collection. Collect the data
and parameters of the PDS, UTS, CDSs, and EVs, such
as the topology information, electrical power price ̺1 ,
congestion fee ̺2 , set the simulation time interval t = 15
min, etc.
Objective
function of Utility
(PDS and UTS)
Objective
function of
Customers (EVs)
Optimal Charging/
EV Information
Discharging Price
∗
∗
Step 2: Higher level: equilibrium computation for the
defined utility part (PDS & UTS) and customer part (CDSs
& EVs). Build the utility and customer objective function
as (1) and (5). Then, compute the equilibrium π ∗t and Pi∗t
1
with (6) and (7).
1
UTS (DUE)
Parking EVs
in CDSs
Step 3: Lower level: DUE in UTS. Receive the information
, build the DUE objective
such as π ∗t , Nit1 , qrdu1 tsu and Pi∗t
1
function with constraints as in (9), solve it with Algorithm
1, and generate the traffic information such as θkt a and bdru1 su .
1
Traffic
Data
Information Collection
(OD Pair)
Traffic Information
1
1
∗
CDSs (Smart Charging/discharging)
CDS 3
CDS 5
Optimal
Power Flow
Step 4: Lower level: OPF in PDS. Receive the information
such as π ∗t , θkt a , and Pi∗t
, build the OPF objective function
1
as in (18) with the constraints (19) and (20), relax it as (21),
solve it as (25) and (26). Then, generate the s∗t
i1 to CDS.
∗
Upper Bound of
Power Injection
∗
1
Step 5: Lower level: smart EVs charging/discharging. Re∗t
t
ceive the information such as s∗t
i1 , π , and Ni1 , compute
the parking EVs for each CDS as (27), build the smart
charging/discharging as (28), then generate the feedback
Pnt t1
for PDS.
information nti1 for UTS and i3i1 Cev,i
3
1
CDS 23
Optimal Charging/
Discharging Price
Optimal
Power
Injection
1
,3
PDS (OPF)
Demand
Load
Data
Collection
(Load)
Fig. 4. The test bench consists of a PDS and a UTS.
is shown in Fig. 7(a) in 18:30. The red arrows indicate that
the traffic delays on these roads are larger than 15 minutes.
As shown in Fig. 7(b), with the proposed approach, the pink
arrows indicate that the traffic delays on these roads are less
than 15 minutes. The number of pink arrows (6 pink arrows)
Step 6: The power-traffic system update with the new UTS
and PDS information, then back to Step 2 with t = t + 1.
9
6
50
6.5
7
7.5
EV (thousand)
8
8.5
9
9.5
10
7
18
6
8
16
5
9
10
15
22
21
4
11
14
23
24
3
12
Computation time (Traffic)
Computation time (Power)
48
46
2
17
19
20
Computation Time (S)
44
42
40
38
36
34
32
30
0.9
0.95
1
1.05
1.1
Power (P.U.)
1.15
1.2
1.25
1.3
1
Fig. 5.
Computation time comparison with different traffic and load factor.
78
13
(a)
7
18
6
8
16
5
9
10
15
22
21
4
11
14
23
24
3
12
4.2
Original Netload
Proposed Approach
Original Traffic Hours
Proposed Approach
66
1.4
Power (MW)
2.8
60
17
18
19
Time (h)
20
Total Traffic Delay (Kh)
2
72
(a)
20
1
13
3.5
(b)
2.8
57
2.1
54
1.4
51
0.7
8
9
10
11
12
13
14
Fig. 7. (a) The traffic congestion scenario in 18:30, (b) The traffic congestion
scenario with proposed approach in 18:30.
Total Traffic Delay (Kh)
Original Netload
Proposed Approach
Original Traffic Hours
Proposed Approach
60
Power (MW)
19
0
22
21
63
48
17
Nodes
1
2
3
5
6
10
0
15
Time (h)
(b)
Fig. 6. (a) The bi-peak shaving and bi-ramp smoothing for PDS and UTS
from 17:00 to 22:00, (b) The peak-shaving and ramp-smoothing for UTS and
over-generation compensation for PDS from 8:00 to 15:00.
TABLE I
T HE OD PAIR OF UTS
13
14
18
20
200
350
240
300
270
210
360
200
200
300
220
345
250
320
260
450
300
210
270
250
200
200
345
345
21
256
200
270
230
300
200
23
345
310
345
345
300
300
C. Social Cost Analysis
Fig. 8, the cost of the proposed approach is lower, and it is
clear that the proposed approach can benefit more for the social
cost with the increasing number of EVs. In Fig. 8, the red
curve and cyan curve indicate the social cost with different
power factors. The EV number is 8000 in the UTS. From
Fig. 8, it is also clear that the proposed approach can benefit
more for the social cost with the increasing power factor. In
summary, the prosed approach can benefit both PDS and UTS
with different traffic scenarios and load factors. In addition, the
social cost increases more with the increasing of PDS load or
EV number, which indicates the quadratic objective function
of PDS and UTS. This design benefits the power-traffic system
to reduce the total social cost.
In Fig. 8, the blue curve with circle and green curve with
square are in one group, which indicates the social costs with
different number of EVs on the road. The load of PDS is 1.0
P.U., which means the total load of PDS is 60 MW. From
D. Smart EVs Charging/Discharging
To test the robustness of the proposed approach, we select
two continuous intervals load data, which contains the biggest
is much less than the number of red arrows (16 red arrows),
which also indicates the alleviation of traffic congestion with
the proposed approach. The total congestion delay reduces
from 4100 hours to 2900 hours.
In summary, from Fig. 6 and Fig. 7, the bi-peak and biramp problems can be reduced and smoothed by the proposed
approach, which benefits both PDS and UTS.
10
6
15000
6.5
7
EV (thousand)
8
7.5
8.5
9
9.5
discharging speeds are controlled to meet the constraint (28b).
10
Original Cost 1 (Traffic)
Proposed Approach (Traffic)
Original Cost 2 (Power)
Proposed Approach (Power)
VIII. C ONCLUSION
In this paper, a hierarchical approach is proposed to shave
the bi-peak and smooth the bi-ramp problem in the powertraffic system from higher level to lower level. The higher
level gives an “overview” for the whole system, and the lower
level gives the detailed description for each part. The EVs
and CDSs function as reserves for both the PDS and UTS
to utilize the flexibility and optimize the operations of the
power-traffic system. In the higher level, the PDS and UTS are
treated together to minimize the social cost, and the EVs and
CDSs are treated as customers to minimize their expenditure.
Then, an equilibrium is designed to determine the optimal
charging/discharging prices, total demand electrical power, and
total required EVs. In the lower level, considering the spatial
and time domain, the detail models of PDS and UTS are
built to specifically determine the power injection and EVs
behaviors. In each CDS, a smart EVs charging/discharging
approach is proposed to further reduce the impacts to the PDS.
In the numerical results, the test bench consists of the IEEE
8,500-bus PDS and Sioux Falls UTS with about 10,000 EVs
and 11 CDSs, which are used to demonstrate the feasibility
and effectiveness of the proposed approach.
In real-world implementation, the weather, human behavior,
and social issues bring the stochastic impacts to the powertraffic system, which increase the uncertainties and bring
more challenges for system operation. In addition, the other
systems such as the natural gas delivery system and the water
system can also impact the proposed system and result in
nonnegligible consequences. In the next step, other factors
such as stochastic of renewable energies, the departure time of
EVs, cybersecurity, multi-energy system, and human behaviors
will be taken into consideration.
Social Cost ($)
10000
5000
0
0.9
0.95
1
1.05
1.1
Power (P.U.)
1.15
1.2
1.25
1.3
Fig. 8. The different social costs comparison with different traffic and load
factors.
1000
180
Original Netload
Original Netload+EV Discharging
Power (KW)
120
500
60
EV Discharging Power (KW)
750
250
0
0
5
10
15
20
0
25
30
Time (Minutes)
(a)
1000
180
Original Netload
Original Netload+EV Discharging
Power (KW)
120
500
60
EV Discharging Power (KW)
750
R EFERENCES
[1] C.-M. Cheng, S.-L. Tsao, and P.-Y. Lin, “Seeds: A solar-based energyefficient distributed server farm,” IEEE Transactions on Systems, Man,
and Cybernetics: Systems, vol. 45, no. 1, pp. 143–156, 2015.
[2] M. Cui, J. Zhang, A. Florita, B.-M. Hodge, D. Ke, and Y. Sun,
“Solar power ramp events detection using an optimized swinging door
algorithm,” in Proc. ASME Int. Design Eng. Tech. Conf. Comput. Inf.
Eng. Conf, 2015.
[3] Y. Gu, H. Jiang, Y. Zhang, J. J. Zhang, T. Gao, and E. Muljadi,
“Knowledge discovery for smart grid operation, control, and situation
awarenessła big data visualization platform,” in North American Power
Symposium (NAPS), 2016. IEEE, 2016, pp. 1–6.
[4] Y. A. Katsigiannis, P. S. Georgilakis, and G. J. Tsinarakis, “A novel colored fluid stochastic petri net simulation model for reliability evaluation
of wind/pv/diesel small isolated power systems,” IEEE Transactions on
Systems, Man, and Cybernetics-Part A: Systems and Humans, vol. 40,
no. 6, pp. 1296–1309, 2010.
[5] Y. Tian and C.-Y. Zhao, “A review of solar collectors and thermal energy
storage in solar thermal applications,” Applied energy, vol. 104, pp. 538–
553, 2013.
[6] D. Connolly, H. Lund, B. V. Mathiesen, and M. Leahy, “A review of
computer tools for analysing the integration of renewable energy into
various energy systems,” Applied Energy, vol. 87, no. 4, pp. 1059–1082,
2010.
[7] E. J. Gonzales and C. F. Daganzo, “The evening commute with cars and
transit: Duality results and user equilibrium for the combined morning
and evening peaks,” Procedia-Social and Behavioral Sciences, vol. 80,
pp. 249–265, 2013.
250
0
0
5
10
15
Time (Minutes)
20
25
30
0
(b)
Fig. 9. (a) The normal EV discharging in CDS 3 without considering the
impacts to the PDS, (b) smart EV discharging in CDS 3 considering the
impacts to the PDS.
deviations within the 30 minutes. In Fig. 9, EV discharging
is taken as an example and 50 EVs are employed to test
the proposed approach. As shown in Fig. 9(a), without the
proposed approach, the stochastic EV discharging behaviors
increase the deviations of the netload. In Fig. 9(b), with the
proposed smart charging/discharging approach, the deviations
of the netload are smoothed, which decreases the impacts to
the PDS during the EV discharging. At the same time, the
11
[8] Y.-H. Cheng, Y.-H. Chang, and I. Lu, “Urban transportation energy and
carbon dioxide emission reduction strategies,” Applied Energy, vol. 157,
pp. 953–973, 2015.
[9] Z. Huang, X. Xu, H. He, J. Tan, and Z. Sun, “Parameterized batch
reinforcement learning for longitudinal control of autonomous land vehicles,” IEEE Transactions on Systems, Man, and Cybernetics: Systems,
2017.
[10] A. Foley, B. Tyther, P. Calnan, and B. Ó. Gallachóir, “Impacts of electric
vehicle charging under electricity market operations,” Applied Energy,
vol. 101, pp. 93–102, 2013.
[11] C. Liu, J. Wang, A. Botterud, Y. Zhou, and A. Vyas, “Assessment
of impacts of PHEV charging patterns on wind-thermal scheduling by
stochastic unit commitment,” IEEE Transactions on Smart Grid, vol. 3,
no. 2, pp. 675–683, 2012.
[12] D. Wu, D. C. Aliprantis, and L. Ying, “Load scheduling and dispatch for
aggregators of plug-in electric vehicles,” IEEE Transactions on Smart
Grid, vol. 3, no. 1, pp. 368–376, 2012.
[13] E. Sortomme and M. A. El-Sharkawi, “Optimal scheduling of vehicle-togrid energy and ancillary services,” IEEE Transactions on Smart Grid,
vol. 3, no. 1, pp. 351–359, 2012.
[14] H. Jiang, Y. Zhang, J. J. Zhang, D. W. Gao, and E. Muljadi,
“Synchrophasor-based auxiliary controller to enhance the voltage stability of a distribution system with high renewable energy penetration,”
IEEE Transactions on Smart Grid, vol. 6, pp. 2107–2115, 2015.
[15] C. Heymans, S. B. Walker, S. B. Young, and M. Fowler, “Economic
analysis of second use electric vehicle batteries for residential energy
storage and load-levelling,” Energy Policy, vol. 71, pp. 22–30, 2014.
[16] X. Geng and L. Xie, “Learning the lmp-load coupling from data: A
support vector machine based approach,” IEEE Transactions on Power
Systems, vol. 32, no. 2, pp. 1127–1138, 2017.
[17] P. Sadeghi-Barzani, A. Rajabi-Ghahnavieh, and H. Kazemi-Karegar,
“Optimal fast charging station placing and sizing,” Applied Energy, vol.
125, pp. 289–299, 2014.
[18] F. He, D. Wu, Y. Yin, and Y. Guan, “Optimal deployment of public
charging stations for plug-in hybrid electric vehicles,” Transportation
Research Part B: Methodological, vol. 47, pp. 87–101, 2013.
[19] W. Wei, S. Mei, L. Wu, M. Shahidehpour, and Y. Fang, “Optimal
traffic-power flow in urban electrified transportation networks,” IEEE
Transactions on Smart Grid, vol. 8, no. 1, pp. 84–95, 2017.
[20] B. Subhonmesh, S. H. Low, and K. M. Chandy, “Equivalence of
branch flow and bus injection models,” in Communication, Control, and
Computing (Allerton), 2012 50th Annual Allerton Conference on. IEEE,
2012, pp. 1893–1899.
[21] B. Stott, J. Jardim, and O. Alsaç, “Dc power flow revisited,” IEEE
Transactions on Power Systems, vol. 24, no. 3, pp. 1290–1300, 2009.
[22] A. G. Bakirtzis, P. N. Biskas, C. E. Zoumas, and V. Petridis, “Optimal
power flow by enhanced genetic algorithm,” IEEE Transactions on
power Systems, vol. 17, no. 2, pp. 229–236, 2002.
[23] Y. Gu, H. Jiang, Y. Zhang, and D. W. Gao, “Statistical scheduling of
economic dispatch and energy reserves of hybrid power systems with
high renewable energy penetration,” in 2014 48th Asilomar Conference
on Signals, Systems and Computers, 2014, pp. 530–534.
[24] Q. Peng and S. H. Low, “Distributed algorithm for optimal power flow
on a radial network,” in 2014 IEEE 53rd Annual Conference on Decision
and Control (CDC). IEEE, 2014, pp. 167–172.
[25] ——, “Distributed algorithm for optimal power flow on an unbalanced
radial network,” in Decision and Control (CDC), 2015 IEEE 54th Annual
Conference on. IEEE, 2015, pp. 6915–6920.
[26] E. Dall’Anese, H. Zhu, and G. B. Giannakis, “Distributed optimal power
flow for smart microgrids,” IEEE Transactions on Smart Grid, vol. 4,
no. 3, pp. 1464–1475, 2013.
[27] A. Y. Lam, B. Zhang, and N. T. David, “Distributed algorithms for
optimal power flow problem,” in Decision and Control (CDC), 2012
IEEE 51st Annual Conference on. IEEE, 2012, pp. 430–437.
[28] N. Li, L. Chen, and S. H. Low, “Demand response in radial distribution
networks: Distributed algorithm,” in Signals, Systems and Computers
(ASILOMAR), 2012 Conference Record of the Forty Sixth Asilomar
Conference on. IEEE, 2012, pp. 1549–1553.
[29] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
optimization and statistical learning via the alternating direction method
of multipliers,” Foundations and Trends R in Machine Learning, vol. 3,
no. 1, pp. 1–122, 2011.
[30] M. Patriksson, The traffic assignment problem: models and methods.
Courier Dover Publications, 2015.
[31] N. Jiang, C. Xie, J. C. Duthie, and S. T. Waller, “A network equilibrium
analysis on destination, route and parking choices with mixed gasoline
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
12
and electric vehicular flows,” EURO Journal on Transportation and
Logistics, vol. 3, no. 1, pp. 55–92, 2014.
Y. Sheffy, “Urban transportation networks: equilibrium analysis with
mathematical programming methods,” Traffic engineering control.
Prentice-Hall, ISBN 0-13-93-972, 1985.
Y. He, X. Liu, C. Zhang, and Z. Chen, “A new model for state-of-charge
(soc) estimation for high-power li-ion batteries,” Applied Energy, vol.
101, pp. 808–814, 2013.
B. Pattipati, C. Sankavaram, and K. Pattipati, “System identification and
estimation framework for pivotal automotive battery management system
characteristics,” IEEE Transactions on Systems, Man, and Cybernetics,
Part C (Applications and Reviews), vol. 41, no. 6, pp. 869–884, 2011.
H. Sedjelmaci, S. M. Senouci, and N. Ansari, “A hierarchical detection
and response system to enhance security against lethal cyber-attacks in
uav networks,” IEEE Transactions on Systems, Man, and Cybernetics:
Systems, 2017.
H. Jiang, J. J. Zhang, W. Gao, and Z. Wu, “Fault detection, identification,
and location in smart grid based on data-driven computational methods,”
IEEE Transactions on Smart Grid, vol. 5, pp. 2947 – 2956, 2014.
H. Jiang, X. Dai, W. Gao, J. Zhang, Y. Zhang, and E. Muljadi, “Spatialtemporal synchrophasor data characterization and analytics in smart
grid fault detection, identification and impact causal analysis,” IEEE
Transactions on Smart Grid, vol. 7, no. 5, pp. 2525–2536, 2016.
C.-W. Ten, G. Manimaran, and C.-C. Liu, “Cybersecurity for critical
infrastructures: Attack and defense modeling,” IEEE Transactions on
Systems, Man, and Cybernetics-Part A: Systems and Humans, vol. 40,
no. 4, pp. 853–865, 2010.
K. Islam, W. Shen, and X. Wang, “Wireless sensor network reliability
and security in factory automation: A survey,” IEEE Transactions on
Systems, Man, and Cybernetics, Part C (Applications and Reviews),
vol. 42, no. 6, pp. 1243–1256, 2012.
X. Geng and L. Xie, “A data-driven approach to identifying system
pattern regions in market operations,” in Power & Energy Society
General Meeting, 2015 IEEE. IEEE, 2015, pp. 1–5.
H. Jiang, Y. Li, Y. Zhang, J. J. Zhang, D. W. Gao, E. Muljadi, and Y. Gu,
“Big data-based approach to detect, locate, and enhance the stability of
an unplanned microgrid islanding,” Journal of Energy Engineering, vol.
143, no. 5, p. 04017045, 2017.
J.-B. Park, K.-S. Lee, J.-R. Shin, and K. Y. Lee, “A particle swarm
optimization for economic dispatch with nonsmooth cost functions,”
IEEE Transactions on Power systems, vol. 20, no. 1, pp. 34–42, 2005.
S. Boyd and L. Vandenberghe, Convex optimization.
Cambridge
university press, 2004.
N. Li, L. Chen, and S. H. Low, “Optimal demand response based on
utility maximization in power networks,” in Power and Energy Society
General Meeting, 2011 IEEE. IEEE, 2011, pp. 1–8.
D. P. Bertsekas and J. N. Tsitsiklis, Parallel and distributed computation:
numerical methods. Prentice hall Englewood Cliffs, NJ, 1989, vol. 23.
B. N. Janson, “Dynamic traffic assignment for urban road networks,”
Transportation Research Part B: Methodological, vol. 25, no. 2, pp.
143–161, 1991.
T. A. Manual, “Bureau of public roads,” US Department of Commerce,
1964.
C. Chekuri, S. Khanna, and F. B. Shepherd, “The all-or-nothing multicommodity flow problem,” in Proceedings of the thirty-sixth annual
ACM symposium on Theory of computing. ACM, 2004, pp. 156–165.
S. H. Low, “Convex relaxation of optimal power flowłpart i: Formulations and equivalence,” IEEE Transactions on Control of Network
Systems, vol. 1, no. 1, pp. 15–27, 2014.
L. Gan and S. H. Low, “Convex relaxations and linear approximation for
optimal power flow in multiphase radial networks,” in Power Systems
Computation Conference (PSCC), 2014. IEEE, 2014, pp. 1–9.
Distribution
test
feeder.
[Online].
Available:
http://ewh.ieee.org/soc/pes/dsacom/testfeeders/index.html
W. H. Lam, H. Shao, and A. Sumalee, “Modeling impacts of adverse
weather conditions on a road network with uncertainties in demand and
supply,” Transportation research part B: methodological, vol. 42, no. 10,
pp. 890–910, 2008.
California ISO today’s outlook details. [Online]. Available:
http://www.caiso.com/Pages/Today’s-Outlook-Details.aspx
Eia
us
energy
information
administration
regional
wholesale
markets.
[Online].
Available:
https://www.eia.gov/electricity/monthly/update/wholesale markets.php
A. de Palma and R. Lindsey, “Traffic congestion pricing methodologies
and technologies,” Transportation Research Part C: Emerging Technologies, vol. 19, no. 6, pp. 1377–1399, 2011.
| 3cs.SY
|
arXiv:1802.00381v1 [math.ST] 1 Feb 2018
Signal-plus-noise matrix models:
eigenvector deviations and fluctuations
Joshua Cape and Minh Tang and Carey E. Priebe
Department of Applied Mathematics and Statistics,
Johns Hopkins University
February 2, 2018
Abstract
Estimating eigenvectors and low-dimensional subspaces is of central importance for numerous problems in statistics, computer science,
and applied mathematics. This paper characterizes the behavior of
perturbed eigenvectors for a range of signal-plus-noise matrix models
encountered in both statistical and random matrix theoretic settings.
We prove both first-order approximation results (i.e. sharp deviations)
as well as second-order distributional limit theory (i.e. fluctuations).
The concise methodology considered in this paper synthesizes tools
rooted in two core concepts, namely (i) deterministic decompositions
of matrix perturbations and (ii) probabilistic matrix concentration
phenomena. We illustrate our theoretical results via simple simulation examples involving stochastic block model random graphs.
Keywords: signal-plus-noise matrix models; eigenvector perturbation; asymptotic normality; random matrices
1
Eigenvector deviations and fluctuations
1
J. Cape, M. Tang, and C. E. Priebe
Introduction
This paper considers the setting where M and E are large n × n symmetric
real-valued matrices with M̂ := M+E representing an additive perturbation
of M by E. For n × r matrices U and Û whose columns are orthonormal
eigenvectors corresponding to the r n leading eigenvalues of M and M̂,
respectively, we ask:
Question 1. How entrywise close are the matrices of eigenvectors U and Û?
Under quite general structural assumptions on U, M, and E, our main
results address Question 1 both at the level of first-order sharp deviations
and at the level of second-order fluctuations. Theorems 1 and 2 quantify the
entrywise closeness of Û to U modulo a necessary orthogonal transformation
W. Theorem 3 subsequently presents a multivariate limit theorem for the
rows of the matrix Û − UW, properly transformed, provided M exhibits
additional kernel-type structure when viewed as an operator.
Numerous problems in statistics consider the eigenstructure of large symmetric matrices. Prominent examples of such problems include (spike) population and covariance matrix estimation (Johnstone, 2001; Silverstein, 1984,
1989) as well as principal component analysis (Jolliffe, 1986; Nadler, 2008),
problems which have received additional attention and windfall as a result
of advances in random matrix theory (Bai and Silverstein, 2010; BenaychGeorges and Nadakuditi, 2011). In addition to these examples, the problem
of community detection in the study of networks has led to widespread interest in understanding perturbations of (large) graph Laplacian and adjacency
matrices given the success of spectral clustering methodologies (Lei and Rinaldo, 2015; Rohe et al., 2011; Sarkar and Bickel, 2015; Tang and Priebe,
2016). Towards this end, recent ongoing and concurrent effort in the statistics, computer science, and mathematics communities has been devoted to
obtaining precise entrywise bounds on eigenvector perturbations (Abbe et al.,
2017; Cape et al., 2017; Eldridge et al., 2017; Fan et al., 2016; Mao et al.,
2017; Tang et al., 2017).
This paper distinguishes itself from the literature by presenting both deviation and fluctuation results within a concise yet flexible signal-plus-noise
matrix model framework amenable to statistical applications and inference.
We employ the two-to-infinity norm machinery and perturbation considerations introduced in (Cape et al., 2017) which have also been employed in
(Abbe et al., 2017; Tang et al., 2017) in order to obtain strong first-order
2
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
bounds. This paper extends the results in (Cape et al., 2017; Tang et al.,
2017), and we further demonstrate how careful analysis within a unified
framework leads to second-order results à la multivariate distributional limit
theory.
2
Preliminaries
For n × r real matrices with orthonormal columns, denoted by Û, U ∈ On,r ,
the columns of Û and U each form orthonormal bases for r-dimensional
subspaces of Rn . The distance between subspaces is commonly defined via
the notion of canonical angles and the C(osine)-S(ine) matrix decomposition
which crucially involve the singular values of the matrix U> Û. Specifically,
by writing the singular values of U> Û as σ1 ≥ σ2 ≥ · · · ≥ σr , then the r × r
diagonal matrix of canonical angles is defined to be Θ(Û, U)ii := arccos(σi )
for i ∈ [r] := {1, 2, . . . , r} (Bhatia, 1997).
One frequently encounters the entrywise-defined matrix sin Θ(Û, U) ∈
r×r
R , since for the commonly considered spectral and Frobenius matrix
norms, η ∈ {k·k, k·kF }, small values of η[sin Θ(Û, U)] indicate small angular
separation (i.e. distance) between the subspaces corresponding to Û and U.
Importantly, the canonical angle notion of distance between r-dimensional
subspaces holds up to row space alignment via orthogonal transformation,
i.e. right-multiplication by an r × r orthogonal matrix W ∈ Or,r ≡ Or . Letting f g denote that there exist absolute constants c1 , c2 > 0 such that
c1 g ≤ f ≤ c2 g, for each choice of the norm η above, (Cai and Zhang, 2016)
h
i
h
i
min η Û − UW η sin Θ(Û, U) .
W∈Or
In this paper, we also focus on such matrices of the form
Û − UW ∈ Rn×r ,
(1)
but we instead consider the two-to-infinity matrix norm which is defined for
any matrix M by kMk2→∞ := maxkxk=1 kMxk∞ , where here k · k and k · k∞
denote the `2 and `∞ vector norms, respectively. The quantity kMk2→∞
has the convenient interpretation of being the maximum Euclidean norm
of the rows of M and will be shown to serve as a useful surrogate for the
maximum absolute entry matrix norm, kMkmax := maxi,j |Mij |. Among
the advantages of working with the two-to-infinity norm is that it remains
3
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
invariant with respect to right-multiplication by orthogonal matrices; see
(Cape et al., 2017) for additional discussion. Our subsequent analysis will be
shown to be particularly meaningful when M exhibits low/bounded coherence
(Candès and Recht, 2009) (i.e. when M is assumed to be delocalized ) in the
sense that kUk2→∞ decays sufficiently quickly in n.
For tall, thin matrices T ∈ Rn×r with n r (e.g. the matrix Û − UW),
standard norm relations reveal that kTkmax and kTk2→∞ differ by at most
a (small) factor depending on r. The same well-known relationship holds
for the spectral and Frobenius norms, kTk and kTkF , respectively, since
necessarily rank(T) ≤ r. In contrast, kTk2→∞ may in certain cases be much
smaller than kTk by a factor depending on n, summarized as
n
r
r
kTkmax kTk2→∞ kTk kTkF .
We refer the reader to (Cape et al., 2017) for further, more general discussion
of the two-to-infinity norm and statistical applications.
3
3.1
Main Results
Setting
Let M ≡ Mn ∈ Rn×n be a symmetric matrix with block spectral decomposition given by
M ≡ [U|U⊥ ][Λ ⊕ Λ⊥ ][U|U⊥ ]> = UΛU> + U⊥ Λ⊥ U>
⊥,
(2)
where the diagonal matrix Λ ∈ Rr×r contains the r largest-in-magnitude
nonzero eigenvalues of M with |Λ11 | ≥ |Λ22 | ≥ · · · ≥ |Λrr | > 0, and U ∈
On,r is an n × r matrix whose orthonormal columns are the corresponding
eigenvectors of M. The diagonal matrix Λ⊥ ∈ R(n−r)×(n−r) contains the
remaining ordered n − r eigenvalues of M with the associated matrix of
orthonormal eigenvectors U⊥ ∈ On,(n−r) . Let E ∈ Rn×n be a symmetric
>
matrix, and write the perturbation of M by E as M̂ := M + E ≡ ÛΛ̂Û +
>
Û⊥ Λ̂⊥ Û⊥ .
Assumption 1. Let ρn denote a possibly n-dependent scaling parameter
such that (0, 1] 3 ρn → cρ ∈ [0, 1] as n → ∞, with nρn ≥ c1 (log n)c2 for some
constants c1 , c2 > 0.
4
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
Assumption 2. There exist constants C, c > 0 such that for all n ≥ n0 (C, c),
|Λrr | ≥ c(nρn ) and |Λ11 ||Λrr |−1 ≤ C, while Λ⊥ ≡ 0 for all n.
Assumption 3. Suppose that kEk = OP ((nρn )1/2 ), i.e. that E is concentrated in the sense that there exist constants C, c > 0 such that kEk ≤
C(nρn )1/2 with probability at least 1 − n−c for all n ≥ n0 (C, c).
Assumption 1 introduces a sparsity scaling factor ρn for additional flexibility. This paper considers the large-n regime, and the dependence of (sequences of) matrices on n is suppressed for notational convenience.
Assumption 2 specifies the magnitude of the leading (i.e. signal) eigenvalues corresponding to the leading eigenvectors of interest. For simplicity
and specificity, all leading eigenvalues are taken to be of the same prescribed
order, and the remaining eigenvalues are assumed to vanish.
Assumption 3 specifies that the random matrix E is concentrated in spectral norm in the classical probabilistic sense. Such concentration holds widely
for random matrix models where E[E] = 0, in which case M̂ has low-rank
expectation since then E[M̂] = M. The advantage of Assumption 3 when
coupled with Assumption 2 is that, together with an application of Weyl’s inequality (Bhatia, 1997), all implicit spectral-based signal-to-noise ratio terms
behave explicitly as kEk|Λii |−1 , kEk|Λ̂ii |−1 = OP ((nρn )−1/2 ) for i ∈ [r].
Below, Assumption 4 specifies an additional probabilistic concentration
requirement that arises in conjunction with the model flexibility introduced
via the sparsity scaling factor ρn in Assumption 1. The notation d·e is used
to denote the ceiling function.
Assumption 4. Suppose that there exist constants CE , ν > 0, ξ > 1, such
that for all integers 1 ≤ k ≤ k(n) := dlog n/ log(nρn )e, for each fixed standard
basis vector ei , and for any fixed unit vector u,
|hEk u, ei i| ≤ CEk (nρn )k/2 (log n)kξ kuk∞
(3)
with probability at least 1 − exp(−ν(log n)ξ ) provided n ≥ n0 (CE , ν, ξ).
Assumption 4 amounts to a higher-order concentration estimate motivated by Lemma 7.10 in (Erdős et al., 2013) that, along with variants thereof,
holds for a broad class of random symmetric matrices including Wigner matrices whose entries exhibit subexponential decay. See Definition 2.1, Remark
2.5, and Definition 2.6 therein for further technical discussion and accompanying material. In addition, taking a simple union bound collectively over
5
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
1 ≤ k ≤ k(n), the standard basis vectors in Rn , and the columns of U yields
an event that holds with probability at least 1 − n−c for some constant c > 0
for sufficiently large n.
The function k(n) is fundamentally model-dependent though its connection with the sparsity factor ρn and satisfies (nρn )−k(n)/2 ≤ n−1/2 for
n sufficiently large. In the case when ρn ≡ 1, then k(n) ≡ 1, and the
behavior reflected in Eq. (3) reduces to commonly-encountered Bernsteintype probabilistic concentration. In contrast, when ρn → 0 and, for example, (nρn ) = Θ(n ) for some ∈ (0, 1), then k(n) ≡ −1 . If instead
(nρn ) = Θ((log n)c2 ) for some c2 > 0, then k(n) = dlog n/(c2 log log n)e. We
remark that all regimes in which ρn → cρ > 0 functionally correspond to the
regime where ρn ≡ 1 by appropriate rescaling.
3.2
First-order approximation (deviations)
Under Assumptions 2 and 3 in Section 3.1, spectral norm subspace analysis
(e.g. via the Davis-Kahan sin Θ theorem (Bhatia, 1997)) yields that for n
sufficiently large, there exists W ∈ Or such that
kÛ − UWk = OP ((nρn )−1/2 ).
(4)
Equation (4) serves as a benchmark bound for the quantity kÛ − UWk2→∞
which is shown below to at times be of much smaller order.
Theorem 1. Suppose that Assumptions 1–4 hold in the context of Section 3.1. Suppose in addition that nρn = ω((log n)2ξ ) and r1/2 ≤ (log n)ξ .
Then for n sufficiently large, there exists W ∈ Or such that
kÛ − UWk2→∞ = OP (nρn )−1/2 × min r1/2 (log n)ξ kUk2→∞ , 1 . (5)
The bound obtained by two-to-infinity norm methods in Eq. (5) is demonstrably superior to the bound implied by Eq. (4) when r1/2 (log n)ξ kUk2→∞ →
0 as n → ∞, i.e. when kUk2→∞ → 0 sufficiently quickly.
The proof of Theorem 1 first proceeds by way of refined deterministic
matrix decompositions and then subsequently leverages the additional, aforementioned probabilistic concentration assumptions. Our proof framework
further extends to permit second-order analysis, culminating in Theorem 3
in Section 3.3. In the process of proving Theorem 3 we also prove Theorem 2,
an extension and refinement of Theorem 1. The theorems in this paper are
collectively proven in Section 4.
6
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
Theorem 2. Suppose that Assumptions 1–3 hold and that Assumption 4
holds up to k(n) + 1. Suppose in addition that nρn = ω((log n)2ξ ) and r1/2 ≤
(log n)ξ . Then for n sufficiently large, there exists W ∈ Or such that
Û − UW = EUΛ−1 W + R,
(6)
for some matrix R ∈ Rn×r satisfying
kRk2→∞ = OP (nρn )−1 × r × max (log n)2ξ , kU> EUk + 1 × kUk2→∞ .
Moreover,
kEUΛ−1 Wk2→∞ = OP (nρn )−1/2 × r1/2 (log n)ξ kUk2→∞ .
Theorem 2 can be summarized as saying that, with respect to the twoto-infinity norm (moreover, in an entrywise sense),
Û ≈ M̂UΛ−1 W = UW + EUΛ−1 W.
Equation (6) provides a collective eigenvector (subspace) characterization
of the relationship between the leading eigenvectors of M and M̂ via the
perturbation E.
Remark 1. It always holds that kU> EUk ≤ kEk, where “≤” can be replaced by “ ” upon invoking Hoeffding-type concentration or more generally (C, c, γ)-concentration (O’Rourke et al., 2013) for suitable choices of E.
Moreover, kRk2→∞ kEUΛ−1 Wk2→∞ in Theorem 2 for numerous regimes
involving nρn and kUk2→∞ .
N
3.3
Second-order limit theory (fluctuations)
In this section we specify additional structure on M and E for the purpose
of establishing the second-order limit theory. Here, M is assumed to have
strictly positive leading eigenvalues, reminiscent of a spike covariance or kernel population matrix setting. It is possible though more involved to obtain
similar second-order results when M is allowed to have both strictly positive
and strictly negative leading eigenvalues of the same order. Specifically, such
modifications would give rise to considerations involving structured orthogonal matrices as well as the indefinite orthogonal group.
7
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
Assumption 5. Suppose that M can be written as M ≡ ρn XX> ≡ UΛU>
with (n−1 X> X) → Ξ ∈ Rr×r (almost surely) as n → ∞ for some symmetric,
invertible matrix Ξ. Also suppose that for a fixed
index i, theiscaled ith row
hP
n
−1/2
−1/2
of EX, written as (nρn )
(EX)i = (nρn )
j=1 Eij Xj , converges in
distribution to a centered multivariate normal random variable Yi ≡ YX,E,i ∈
Rr with second moment matrix Γi ≡ ΓX,E,i ∈ Rr×r .
Theorem 3. Suppose that Assumptions 1–3 hold and that Assumption 4
holds up to k(n) + 1. Suppose in addition that nρn = ω((log n)2ξ ), r1/2 ≤
(log n)ξ , and
ρ−1/2
× r × max (log n)2ξ , kU> EUk + 1 × kUk2→∞ → 0
(7)
n
in probability as n → ∞. Let Ûi and Ui be column vectors denoting the
ith rows of Û and U, respectively. Then there exist sequences ofh orthogonali
1/2
matrices {W} and {WX } such that the random vector nρn W>
X WÛi − Ui
converges in distribution to a centered multivariate normal random variable
with covariance matrix Σi := Ξ−3/2 Γi Ξ−3/2 , i.e.
h
i
>
nρ1/2
W
W
Û
−
U
∼ Nr (0, Σi ).
(8)
i
i
n
X
n→∞
Equation (7) in Theorem 3 amounts to a mild additional regularity condition that holds, for example, when kUk2→∞ = O((log n)c3 n−1/2 ) for a suitable
constant c3 > 0. Then, in the context of this paper, the left-hand side can
often be shown to behave as OP ((log n)c4 (nρn )−1/2 ) for some c4 > 0, with
(log n)c4 (nρn )−1/2 → 0 as n → ∞.
Remark 2 (Example: M matrix with kernel-type structure). Let F be a
probability distribution defined on X ⊆ Rr , and let X1 , . . . , Xn ∼ F be independent random vectors with full-rank second moment matrix E[X1 X1> ] ≡
Ξ ∈ Rr×r . For X := [X1 | . . . |Xn ]> ∈ Rn×r , let M := ρn XX> ≡ UΛU> ,
so for each n there exists an r × r orthogonal matrix WX (depending on
1/2
n) such that ρn X = UΛ1/2 WX . The strong law of large numbers guarantees that (n−1 X> X) → Ξ almost surely as n → ∞, and so M has r
eigenvalues of order Θ(nρn ) asymptotically almost surely. Moreover, then
kUk2→∞ ≤ Cn−1/2 kXk2→∞ asymptotically almost surely for some constant
C > 0, where kXk2→∞ can be suitably controlled by imposing additional
assumptions (e.g. taking X to be bounded or imposing moment assumptions
on kX1 k). Conditioning on X yields a deterministic choice of M for the
purposes of Assumption 5.
N
8
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
Remark 3 (Example: E matrix and multivariate normality). To continue the
discussion from Remark 2, let all the entries of E be centered, independent
and identically distributed up to symmetry with common variance given
by σ 2 (E). Then, by the classical multivariate central limit theorem, the
asymptotic normality condition in Assumption 5 holds, and by a slight abuse
of notation, Theorem 3 yields
h
i
d
>
1/2
nρn WX WÛi − Ui → Nr (0, σ 2 (E)Ξ−2 ).
Similar behavior holds more generally when the entries of E are permitted
to have heterogeneous variances or different distributions.
N
3.4
Simulations
The K-block stochastic block model (SBM) (Holland et al., 1983) is a simple
yet ubiquitous random graph model in which vertices are assigned to one of K
possible communities (blocks) and where the probability of any two vertices
sharing an edge occurs conditionally independently on the vertices’ (latent)
community memberships. For (undirected) SBM graphs on n vertices, the
binary (symmetric) adjacency matrix A ∈ {0, 1}n×n can be viewed as an
additive perturbation of a (low rank) population edge probability matrix
P ∈ [0, 1]n×n , A = P+E, where for K-block SBMs the matrix P corresponds
to an appropriate dilation of the block edge probability matrix B ∈ [0, 1]K×K .
In the language of this paper, M̂ := A and M := P. It can be verified that
the aforementioned assumptions and hypotheses hold for the following SBM
examples.
Consider n-vertex graphs arising from the three-block stochastic block
model with balanced (equal) block sizes where the within-block and between
block Bernoulli edge probabilities are given by B11 = 0.5 and B12 = 0.3,
respectively. Then rank(M) = 3, and the second-largest eigenvalue of M has
multiplicity two. More specifically,
λ1 (M) = n3 (B11 + 2B12 ) > λ2 (M) = λ3 (M) = n3 (B11 − B12 ) 0.
For each n ∈ {600, 900, 1200, . . . , 3000}, Figure 1 plots the empirical mean
and 90% confidence interval for kÛ − UWk2→∞ computed from 100 independent realizations (i.e. simulated adjacency matrices). Figure 1 also plots
−1/2
the function φ(n) := (λ3 (M))(log n)n−1/2 , which for large n captures the
dominant behavior of the bound in Theorem 2.
9
J. Cape, M. Tang, and C. E. Priebe
0.03
0.00
0.01
0.02
norm value
0.04
0.05
0.06
Eigenvector deviations and fluctuations
600
900
1200
1500
1800
2100
2400
2700
3000
n
Figure 1: First-order approximation and simulations for the three-block
SBM. The x-axis corresponds to the number of vertices n. The y axis corresponds to values of kÛ − UWk2→∞ .
Table 1: The block-conditional marginal sample covariance matrices for each
block of the two-block SBM. The final column shows the theoretical marginal
covariance matrices obtained from Theorem 3.
n
1000
Σ̂1
Σ̂2
∞
2000
8.95 −17.26
−17.26
54.51
5.68 −11.97
−11.97
43.19
10.23 −19.91
−19.91
56.51
9.88 −19.32
−19.32
57.32
6.72 −14.29
−14.29
48.87
6.66 −14.51
−14.51
50.12
Alternatively, consider n-vertex graphs arising from the two-block stochastic block model with 40% of the vertices belonging to block one and where
the block edge probability matrix B has entries B11 = 0.5, B12 = B21 = 0.2,
and B22 = 0.2. Table 1 shows aggregate block-conditional sample covari1/2
ance matrix estimates for the centered random vectors nρn W>
X [WÛi − Ui ]
together with the theoretical covariance matrices specified by Theorem 3.
10
Eigenvector deviations and fluctuations
4
J. Cape, M. Tang, and C. E. Priebe
Appendix
Proof of Theorems 1, 2, and 3. We begin with several important observations, namely that
k(I − UU> )Ûk = k sin Θ(Û, U)k = O(kEk|Λrr |−1 ) = OP ((nρn )−1/2 ), (9)
and that there exists W ∈ Or depending on Û and U such that
kU> Û − Wk ≤ k sin Θ(Û, U)k2 = OP ((nρn )−1 ).
(10)
In particular, W can be taken to be the product of the left and right orthogonal factors in the singular value decomposition of U> Û. Additional details
concerning Eqs. (9–10) may be found, for example, in (Cape et al., 2017).
Importantly, the relation ÛΛ̂ = M̂Û = (M + E)Û yields the matrix
Sylvester equation ÛΛ̂ − EÛ = MÛ. The spectra of Λ̂ and E are disjoint
from one another with high probability by Assumptions 2 and 3, so it follows
that Û can be written as a von-Neumann matrix series (Bhatia, 1997)
Û =
∞
X
Ek MÛΛ̂−(k+1) =
∞
X
Ek UΛU> ÛΛ̂−(k+1) ,
(11)
k=0
k=0
where the second equality holds since rank(M) = r.
For any choice of W ∈ Or , the matrix Û − UW can be decomposed as
Û − UW = EÛΛ̂−1 + UΛ(U> ÛΛ̂−1 − Λ−1 U> Û) + U(U> Û − W)
(2)
:= EÛΛ̂−1 + R(1) + RW .
(2)
For RW := U(U> Û − W), it follows that for W satisfying Eq. (10), then
(2)
kRW k2→∞ ≤ kU> Û − WkkUk2→∞ = OP (nρn )−1 kUk2→∞ .
As for R(1) := UΛR(3) where R(3) := (U> ÛΛ̂−1 − Λ−1 U> Û) ∈ Rr×r , the
entries of R(3) satisfy
h
i
(3)
Rij = hui , ûj i (Λ̂jj )−1 − (Λii )−1 = hui , ûj i(Λii − Λ̂jj )(Λii )−1 (Λ̂jj )−1 .
Define the matrix H1 ∈ Rr×r entrywise according to (H1 )ij := (Λii )−1 (Λ̂jj )−1 .
Then, with ◦ denoting the Hadamard matrix product,
R(3) = −H1 ◦ (U> ÛΛ̂ − ΛU> Û).
11
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
The rightmost matrix factor can be expanded as
(U> ÛΛ̂ − ΛU> Û) = U> EÛ = U> EUU> Û + U> E(I − UU> )Û,
and is therefore bounded in spectral norm using Eq. (9) in the manner
kU> ÛΛ̂ − ΛU> Ûk ≤ kU> EUk + OP (1).
Combining the above observations together with properties of matrix norms
yields the following two-to-infinity norm bound on R(1) .
kR(1) k2→∞ = kUΛR(3) k2→∞ ≤ rkUk2→∞ kΛkkH1 kmax kU> ÛΛ̂ − ΛU> Ûk
= OP (r(nρn )−1 (kU> EUk + 1)kUk2→∞ ).
Assumptions 2 and 3 with an application of Weyl’s inequality (Bhatia,
1997) guarantee that there exist constants C1 , C2 > 0 such that kEk ≤
C1 (nρn )1/2 and kΛ̂−1 k ≤ C2 (nρn )−1 with high probability for n sufficiently
large. Therefore, by considering the appropriate von-Neumann matrix series,
−1
kEÛΛ̂ k2→∞ =
∞
X
Ek UΛU> ÛΛ̂−(k+1)
k=1
2→∞
k(n)
≤
X
−1 k+1
k
kE Uk2→∞ kΛkkΛ̂ k
k=1
+
∞
X
kEkk kΛkkΛ̂−1 kk+1
k=k(n)+1
= OP r1/2 (nρn )−1/2 (log n)ξ kUk2→∞ + (nρn )−1/2 kUk2→∞ ,
where we have used the fact that nρn = ω((log n)2ξ ), (nρn )−k(n)/2 ≤ n−1/2 ≤
kUk2→∞ for n sufficiently large, and that by Assumption 4, for each k ≤ k(n),
with high probability
kEk Uk2→∞ ≤ r1/2 max |hEk uj , ei i| ≤ r1/2 CEk (nρn )k/2 (log n)kξ kUk2→∞ .
i∈[n],j∈[r]
Since kU> EUk ≤ kEk and r1/2 ≤ (log n)ξ with nρn = ω((log n)2ξ ), then
(2)
kÛ − UWk2→∞ ≤ kEÛΛ̂−1 k2→∞ + kR(1) k2→∞ + kRW k2→∞
= OP r1/2 (nρn )−1/2 (log n)ξ kUk2→∞ .
This completes the proof of Theorem 1.
12
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
Next, we further decompose the matrix EÛΛ̂−1 by extending the above
proof techniques in order to obtain second-order fluctuations. In particular,
using the von-Neumann matrix series form in Eq. (11) yields
EÛΛ̂
−1
>
= EUΛU ÛΛ̂
−2
+
∞
X
Ek UΛU> ÛΛ̂−(k+1)
k=2
= EUΛ W + EUΛ(U> ÛΛ̂−2 − Λ−2 U> Û) + EUΛ−1 (U> Û − W)
∞
X
+
Ek UΛU> ÛΛ̂−(k+1)
−1
k=2
(1)
(2)
(∞)
:= EUΛ−1 W + R2 + R2,W + R2 .
The final term satisfies the bound
(∞)
kR2 k2→∞ = OP r1/2 (nρn )−1 (log n)2ξ kUk2→∞ ,
which follows from Assumption 4 holding up to k(n) + 1, namely
k(n)+1
(∞)
kR2 k2→∞
≤
X
−1 k+1
k
kE Uk2→∞ kΛkkΛ̂ k
k=2
+
∞
X
kEkk kΛkkΛ̂−1 kk+1
k=k(n)+2
= OP r
1/2
−1
(nρn ) (log n) kUk2→∞ + (nρn )−1 kUk2→∞ .
2ξ
(2)
On the other hand, modifying the previous analysis used to bound RW yields
(2)
kR2,W k2→∞ ≤ kEUk2→∞ kΛ−1 kkU> Û−Wk = OP (r1/2 (nρn )−3/2 (log n)ξ kUk2→∞ ).
(1)
We now bound R2 := EUΛ(U> ÛΛ̂−2 − Λ−2 U> Û) by extending the pre(1)
(3)
(3)
vious argument used to bound R(1) . For R2 := EUΛR2 where R2 :=
(3)
(U> ÛΛ̂−2 − Λ−2 U> Û) ∈ Rr×r , the entries of R2 satisfy
h
i
(3)
−2
−2
Rij = hui , ûj i (Λ̂jj ) − (Λii )
= hui , ûj i(Λ2ii − Λ̂2jj )(Λii )−2 (Λ̂jj )−2 .
Define the matrix H2 ∈ Rr×r entrywise according to (H2 )ij := (Λii )−2 (Λ̂jj )−2 .
Then, with ◦ denoting the Hadamard matrix product,
(3)
R2 = −H2 ◦ (U> ÛΛ̂2 − Λ2 U> Û).
13
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
The rightmost matrix factor can be written as
(U> ÛΛ̂2 − Λ2 U> Û) = U> (M̂)2 Û − U> M2 Û = U> (ME + EM)Û,
and has bounded spectral norm on the order of OP ((nρn )3/2 ). Hence,
(1)
(3)
kR2 k2→∞ = kEUΛR2 k2→∞ ≤ rkEUk2→∞ kΛkkH2 kmax kU> ÛΛ̂2 − Λ2 U> Ûk
= OP (r3/2 (nρn )−1 (log n)ξ kUk2→∞ ).
(2)
(1)
(2)
(∞)
For R := R(1) + RW + R2 + R2,W + R2 , we have therefore shown that
Û − UW = EUΛ−1 W + R,
(12)
where since r1/2 ≤ (log n)ξ , then the residual matrix R satisfies
kRk2→∞ = OP r × (nρn )−1 × max (log n)2ξ , kU> EUk + 1 × kUk2→∞ .
The leading term agrees with the order of the bound in Theorem 1, namely
kEUΛ−1 Wk2→∞ = OP (nρn )−1/2 × r1/2 (log n)ξ kUk2→∞ .
This proves Theorem 2 as an intermediate step in proving Theorem 3, which
we now proceed to finish.
By assumption, M = ρn XX> ≡ UΛU> , so there exists an orthogonal
1/2
matrix WX (depending on n) such that ρn X = UΛ1/2 WX , hence ρn X> X =
−1
W>
X ΛWX . The matrix EUΛ W can therefore be written as
>
−3/2
EUΛ−1 W = ρ−1
(W>
n EX(X X)
X W).
Multiplication on the right by (inverses of) orthogonal transformations in
Eq. (12) therefore yields the matrix relation
h
i
>
>
−3/2
ÛW WX − UWX = ρ−1
+ RW> WX .
n EX(X X)
For fixed i, let Ûi , Ui , and Ri be column vectors denoting the ith rows of Û, U,
1/2
and R, respectively. Equation (7) implies that nρn kRi k → 0 in probability.
>
In addition, (n−1 X X)−3/2 → Ξ−3/2 almost surely by Assumption 5 together
with the continuous mapping theorem. The scaled ith row of EX satisfies
d
(nρn )−1/2 (EX)i → Yi ∼ Nr (0, Γi ) by Assumption 5, so combining the above
14
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
observations together with Slutsky’s theorem yields that there exist sequences
of orthogonal matrices {W} and {WX } such that, as n → ∞,
h
i
−3/2
d
>
>
−1 >
nρ1/2
W
W
Û
−
U
(nρn )−1/2 (EX)i + nρ1/2
i
i = n X X
n
X
n WX WRi
d
→ Ξ−3/2 Yi + 0.
In particular,
>
nρ1/2
n WX
h
i
WÛi − Ui ∼ Nr (0, Σi ),
n→∞
with Σi := Ξ−3/2 Γi Ξ−3/2 . This completes the proof of Theorem 3.
References
Abbe, E., J. Fan, K. Wang, and Y. Zhong (2017). Entrywise eigenvector analysis of random matrices with low expected rank. arXiv preprint
arXiv:1709.09565 .
Bai, Z. and J. W. Silverstein (2010). Spectral analysis of large dimensional
random matrices, Volume 20. Springer.
Benaych-Georges, F. and R. R. Nadakuditi (2011). The eigenvalues and
eigenvectors of finite, low rank perturbations of large random matrices.
Advances in Mathematics 227 (1), 494–521.
Bhatia, R. (1997). Matrix Analysis, Volume 169 of Graduate Texts in Mathematics. Springer-Verlag, New York.
Cai, T. T. and A. Zhang (2016). Rate-optimal perturbation bounds for
singular subspaces with applications to high-dimensional statistics. arXiv
preprint arXiv:1605.00353, to appear in The Annals of Statistics.
Candès, E. J. and B. Recht (2009). Exact matrix completion via convex
optimization. Foundations of Computational Mathematics 9 (6), 717.
Cape, J., M. Tang, and C. E. Priebe (2017). The two-to-infinity norm and
singular subspace geometry with applications to high-dimensional statistics. arXiv preprint arXiv:1705.10735 .
Eldridge, J., M. Belkin, and Y. Wang (2017). Unperturbed: spectral analysis
beyond Davis-Kahan. arXiv preprint arXiv:1706.06516 .
15
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
Erdős, L., A. Knowles, H.-T. Yau, and J. Yin (2013). Spectral statistics
of Erdős–Rényi graphs I: Local semicircle law. The Annals of Probability 41 (3B), 2279–2375.
Fan, J., W. Wang, and Y. Zhong (2016). An `∞ eigenvector perturbation
bound and its application to robust covariance estimation. arXiv preprint
arXiv:1603.03516 .
Holland, P. W., K. B. Laskey, and S. Leinhardt (1983). Stochastic blockmodels: First steps. Social Networks 5 (2), 109–137.
Johnstone, I. M. (2001). On the distribution of the largest eigenvalue in
principal components analysis. The Annals of Statistics 29 (2), 295–327.
Jolliffe, I. T. (1986). Principal Component Analysis. Springer.
Lei, J. and A. Rinaldo (2015). Consistency of spectral clustering in stochastic
block models. The Annals of Statistics 43 (1), 215–237.
Mao, X., P. Sarkar, and D. Chakrabarti (2017). Estimating mixed memberships with sharp eigenvector deviations. arXiv preprint arXiv:1709.00407 .
Nadler, B. (2008). Finite sample approximation results for principal component analysis: A matrix perturbation approach. The Annals of Statistics 36 (6), 2791–2817.
O’Rourke, S., V. Vu, and K. Wang (2013). Random perturbation of low rank
matrices: Improving classical bounds. arXiv preprint arXiv:1311.2657 .
Rohe, K., S. Chatterjee, and B. Yu (2011). Spectral clustering and the
high-dimensional stochastic blockmodel. The Annals of Statistics 39 (4),
1878–1915.
Sarkar, P. and P. J. Bickel (2015). Role of normalization in spectral clustering
for stochastic blockmodels. The Annals of Statistics 43 (3), 962–990.
Silverstein, J. W. (1984). Some limit theorems on the eigenvectors of large
dimensional sample covariance matrices. Journal of Multivariate Analysis 15 (3), 295–324.
Silverstein, J. W. (1989). On the eigenvectors of large dimensional sample
covariance matrices. Journal of Multivariate Analysis 30 (1), 1–16.
16
Eigenvector deviations and fluctuations
J. Cape, M. Tang, and C. E. Priebe
Tang, M., J. Cape, and C. E. Priebe (2017). Asymptotically efficient estimators for stochastic blockmodels: the naive MLE, the rank-constrained
MLE, and the spectral. arXiv preprint arXiv:1710.10936 .
Tang, M. and C. E. Priebe (2016). Limit theorems for eigenvectors of the normalized Laplacian for random graphs. arXiv preprint arXiv:1607.08601,
to appear in The Annals of Statistics.
17
| 10math.ST
|
WSEAS TRANSACTIONS on SYSTEMS
P. Balasubramanian
System Reliability, Fault Tolerance and Design Metrics Tradeoffs in the
Distributed Minority and Majority Voting Based Redundancy Scheme
P. BALASUBRAMANIAN*
* School of Computer Engineering
Nanyang Technological University
50 Nanyang Avenue
SINGAPORE 639798
E-mail: [email protected]
Abstract: - The distributed minority and majority voting based redundancy (DMMR) scheme was recently
proposed as an efficient alternative to the conventional N-modular redundancy (NMR) scheme for the physical
design of mission/safety-critical circuits and systems. The DMMR scheme enables significant improvements in
fault tolerance and design metrics compared to the NMR scheme albeit at the expense of a slight decrease in the
system reliability. In this context, this paper studies the system reliability, fault tolerance and design metrics
tradeoffs in the DMMR scheme compared to the NMR scheme when the majority logic group of the DMMR
scheme is increased in size relative to the minority logic group. Some example DMMR and NMR systems were
realized using a 32/28nm CMOS process and compared. The results show that 5-of-M DMMR systems have a
similar or better fault tolerance whilst requiring similar or fewer function modules than their counterpart NMR
systems and simultaneously achieve optimizations in design metrics. Nevertheless, 3-of-M DMMR systems
have the upper hand with respect to fault tolerance and design metrics optimizations than the comparable NMR
and 5-of-M DMMR systems. With regard to system reliability, NMR systems are closely followed by 5-of-M
DMMR systems which are closely followed by 3-of-M DMMR systems. The verdict is 3-of-M DMMR systems
are preferable to implement higher levels of redundancy from a combined system reliability, fault tolerance and
design metrics perspective to realize mission/safety-critical circuits and systems.
Key-Words: - NMR, DMMR, Reliability, Fault tolerance, Figure of Merit, Digital design, ASIC, Standard cells
realizing specific portions of future generation
mission and safety-critical circuits and systems
which demand greater fault tolerance. In this
context, it was recently shown [5] that the
distributed minority and majority voting based
redundancy (DMMR) scheme forms an efficient and
viable alternative to the conventional NMR scheme
for implementing mission and safety-critical circuits
and systems.
In the NMR scheme, N identical function
modules are used and the equivalent outputs of N
identical function modules are supplied to a
1 Introduction to NMR and DMMR
N-modular redundancy (NMR) has been widely
used for the fault-tolerant design of mission and
safety-critical circuits and systems [1] – [3] which
are used in space, aerospace, nuclear, defence,
banking, financial and other industrial applications.
Reference [4] suggests that due to increasing
reliability and variability issues in the
nanoelectronics regime, a selective utilization of
higher levels of redundancy which involves the use
of several identical function modules called as
progressive modular redundancy may be needed for
E-ISSN: 2224-2678
59
Volume 15, 2016
WSEAS TRANSACTIONS on SYSTEMS
P. Balasubramanian
Referring to Fig 2a, in the 3-of-M DMMR
system, at least two function modules out of F1, F2,
F3 comprising the majority logic group should
maintain the correct operation and at least one
function module amongst F4 to FM in the minority
logic group should maintain the correct operation
concurrently. On the other hand, referring to Fig 2b,
in the 5-of-M DMMR system, at least three function
modules out of F1, F2, F3, F4, F5 comprising the
majority logic group should maintain the correct
operation and at least one function module amongst
F6 to FM in the minority logic group should maintain
the correct operation concurrently. In this article, we
succinctly study the system reliability, fault
tolerance and design metrics aspects of some
example 5-of-M DMMR systems in comparison
with those of counterpart NMR and 3-of-M DMMR
systems to understand the tradeoffs involved.
majority voter, which processes and produces the
NMR system output by performing majority voting
on the function modules outputs. The block
schematic of the NMR system is shown in Fig 1. In
the NMR scheme, N is odd, and at least (N+1)/2
function modules are required to satisfy the majority
logic, i.e. the faulty/failure state of at most (N–1)/2
function modules are tolerated by the NMR scheme.
Majority Logic Group
DMMR Voter
Fig 1. Block diagram of NMR system
In the DMMR scheme, the function modules are
split into two groups as the majority logic group and
the minority logic group. A majority of function
modules in the majority logic group and at least one
function module in the minority logic group should
maintain the correct operation to ensure the correct
operation of the DMMR system. Note that in a
generic K-of-M DMMR system, K identical
function modules constitute the majority logic group
and the remaining (M–K) identical function
modules constitute the minority logic group, where
K and M are positive integers with K < M and K ≥
3. The equivalent outputs produced by identical
function modules of the majority and minority logic
groups are combined by the DMMR voter, as shown
in Fig 2, to produce the DMMR system output.
Function
Module 1
F1
Function
Module 2
F2
Function
Module 3
F3
Identical
inputs
supplied
to the
function
modules
from
outside
world
Function
Module 4
F4
Function
Module 5
Function
Module M
MAJ
Majority
Voter
F5
3-of-M DMMR
System Output
MIN
FM
Minority Logic Group
(a)
Majority Logic Group
Identical
inputs
supplied
to the
function
modules
from
outside
world
2 3-of-M and 5-of-M DMMR Systems
Fig 2a shows the 3-of-M DMMR system topology
where the majority logic group comprises 3 function
modules and the minority logic group comprises
(M–3) function modules. Fig 2b shows the 5-of-M
DMMR system topology where the majority logic
group comprises 5 function modules and the
minority logic group comprises (M–5) function
modules. The corresponding DMMR voters are also
shown in Fig 2. The majority voting element present
in the DMMR voter of the 3-of-M DMMR system is
a 2-of-3 majority voter [6], and the majority voting
element present in the DMMR voter of the 5-of-M
DMMR system is a 3-of-5 majority voter [7].
E-ISSN: 2224-2678
Identical
inputs
supplied
to the
function
modules
from
outside
world
Identical
inputs
supplied
to the
function
modules
from
outside
world
Function
Module 1
F1
Function
Module 2
F2
Function
Module 3
F3
Function
Module 4
F4
Function
Module 5
F5
Function
Module 6
F6
Function
Module 7
Function
Module M
DMMR Voter
Majority
Voter
F7
MIN
MAJ
5-of-M DMMR
System Output
FM
Minority Logic Group
(b)
Fig 2. Block diagrams of (a) 3-of-M DMMR system
and (b) 5-of-M DMMR system
60
Volume 15, 2016
WSEAS TRANSACTIONS on SYSTEMS
P. Balasubramanian
system reliabilities of counterpart NMR and 3-of-M
DMMR systems.
Equations (1) and (2) give the respective system
reliability expressions of 5-of-7 and 5-of-8 DMMR
systems by assuming perfect DMMR voters. The
system reliability equations of counterpart NMR
(i.e. 7MR and 9MR) and 3-of-M DMMR (i.e. 3-of-6
DMMR and 3-of-7 DMMR) systems are given in
[5]. In (1) and (2), RM denotes the module reliability
and RS denotes the system reliability. Since identical
function modules are used, the modules reliabilities
are assumed to be equivalent. RM denotes the
reliability or the probability of correct working of a
function module, and (1–RM) denotes the nonreliability or probability of the faulty/failure state of
a function module.
3 Simulation Results and Conclusions
Some example NMR, 3-of-M DMMR and 5-of-M
DMMR systems especially targeting higher levels of
redundancy were implemented based on a 32/28nm
CMOS process [8] by considering the 4×4 array
multiplier [9] as a representative function module as
in [5]. The structural integrity of the different
redundant systems was preserved during technology
mapping to facilitate a legitimate comparison of
their corresponding design metrics post-physical
synthesis. The simulation mechanism, environment
(typical case PVT) and test benches (supplied at
time intervals of 4ns, i.e. 250MHz) were maintained
the same as that of [5] to ensure uniformity and to
enable a direct correspondence with the previous
results. The simulation results viz. average power,
critical path delay, and Silicon area of the different
redundant systems estimated are given in Table 1. A
figure-of-merit (FOM) was calculated as the inverse
of the power-delay-area product as in [5], which is
also given in Table 1. Since power, delay and area
are desirable to be minimized, a high value of FOM
indicates an optimized design [10] – [17], which
signifies reduction in cost.
RS5-of-7 DMMR = 20RM4 (1 – RM)3 + 20RM5 (1 – RM)2
+ 7RM6 (1 – RM) + RM7
(1)
RS5-of-8 DMMR = 30RM4 (1 – RM)4 + 45RM5 (1 – RM)3
+ 28RM6 (1 – RM)2 + 8RM7 (1 – RM)
+ RM8
(2)
7MR, 3-of-6 DMMR and 5-of-7 DMMR systems
would be able to accommodate the faulty or failure
state of a maximum of 3 function modules, while
9MR, 3-of-7 DMMR and 5-of-8 DMMR can cope
with the faulty or failure state of a maximum of 4
function modules. With respect to the former, the 3of-6 DMMR system requires one function module
less than the 7MR and 5-of-7 DMMR systems. With
respect to the latter, the 5-of-8 DMMR system
requires one function module less than the 9MR
system and one function module more than the 3-of7 DMMR system.
An analysis of the system reliabilities of 7MR,
9MR, 3-of-6 DMMR, 3-of-7 DMMR, 5-of-7
DMMR and 5-of-8 DMMR systems versus their
module reliabilities reveals that NMR systems
feature the highest reliability closely followed by
the reliabilities of 5-of-M DMMR systems which
are closely followed by the reliabilities of 3-of-M
DMMR systems. For example, for RM = 0.9, the
system reliabilities of 7MR and 9MR systems are
0.997272 and 0.99910908; the system reliabilities of
5-of-7 and 5-of-8 DMMR systems are 0.9815256
and 0.99044856, and the system reliabilities of 3-of6 and 3-of-7 DMMR systems are 0.971028 and
0.9719028 respectively. In general, 5-of-M DMMR
systems show improvements in system reliability
compared to the system reliabilities of 3-of-M
DMMR systems whilst featuring the same degree of
fault tolerance although requiring a similar or
greater number of function modules. Hence, the
system reliabilities of 5-of-M DMMR systems
happen to lie midway between the corresponding
E-ISSN: 2224-2678
Table 1. Power, delay and area metrics of various
NMR, 3-of-M DMMR and 5-of-M DMMR systems
FOM
NMR/DMMR Power Delay Area
System Type (µW)
(ns)
(µm2) (× 106)
7MR
191.2 1.12
865.11 5.40
3-of-6 DMMR 129.4 0.90
567.25 15.14
5-of-7 DMMR 164.1 0.99
730.92 8.42
9MR
278.5 1.23
1269.7 2.30
3-of-7 DMMR 151.2 0.91
661.79 10.98
5-of-8 DMMR 184.5 0.99
817.33 6.70
Compared to 7MR, the 3-of-6 DMMR and 5-of-7
DMMR systems implementations report respective
improvements in FOM by 180.4% and 56%. In
comparison with 9MR, the 3-of-7 DMMR and 5-of8 DMMR systems implementations report
corresponding enhancements in FOM by 377.4%
and 191.3%. The simulation results show that the
FOM of 5-of-M DMMR systems lies approximately
midway between the FOMs of counterpart NMR
and 3-of-M DMMR systems.
Hence it can be inferred that when giving
importance to system reliability alone, the NMR
system topology is preferable. However, to realize
higher levels of redundancy, the NMR scheme
61
Volume 15, 2016
WSEAS TRANSACTIONS on SYSTEMS
P. Balasubramanian
method,” Proc. IEEE International Conference
on Emerging Technologies, pp. 289-294, 2005.
[12] P. Balasubramanian, M.R. Lakshmi Narayana,
R. Chinndadurai, “Power optimized logic
circuit design with a novel synthesis
technique,”
Proc.
IEEE
International
Conference on Emerging Technologies, pp.
306-311, 2005.
[13] P. Balasubramanian, N.E. Mastorakis, “Power,
delay and area comparisons of majority voters
relevant to TMR architectures,” Proc. 10th
International Conference on Circuits, Systems,
Signal and Telecommunications, pp. 110-117,
2016.
[14] P. Balasubramanian, K. Prasad, N.E.
Mastorakis, “A fault tolerance improved
majority voter for TMR system architectures,”
WSEAS Transactions on Circuits and Systems,
vol. 15, pp. 108-122, 2016.
[15] P. Balasubramanian, “ASIC-based design of
NMR
system
health
monitor
for
mission/safety-critical
applications,”
SpringerPlus, vol. 5, no. 1, pages 16, 2016.
[16] P. Balasubramanian, N.E. Mastorakis, “ASICbased implementation of synchronous sectioncarry based carry lookahead adders,” Proc. 10th
International Conference on Circuits, Systems,
Signal and Telecommunications, pp. 58-64,
2016.
[17] P. Balasubramanian, N.E. Mastorakis, “Design
of synchronous section-carry based carry
lookahead adders with improved figure of
merit,” WSEAS Transactions on Circuits and
Systems, vol. 15, pp. 155-164, 2016.
would be a poor choice when taking into account
the system implementation cost. In comparison, the
5-of-M DMMR system topology would be a better
choice as it could considerably reduce the
implementation cost whilst featuring only slightly
less system reliability. Nevertheless, from a
combined system reliability, fault tolerance and
design metrics (i.e. cost and weight) perspective, the
3-of-M DMMR system is all the more preferable
than the rest whilst being associated with a
moderately less system reliability.
References:
[1] B.W. Johnson, Design and Analysis of Fault
Tolerant Digital Systems, Addison-Wesley
Longman Publishing Company, Boston, MA,
USA, 1988.
[2] I. Koren, C. Mani Krishna, Fault-Tolerant
Systems, 1st edition, Morgan Kaufmann,
California, USA, 2007.
[3] E. Dubrova, Fault-Tolerant Design, Springer,
New York, USA, 2013.
[4] T. Ban, L. Naviner, “Progressive modular
redundancy for fault-tolerant designs in
nanoelectronics,” Microelectronics Reliability,
vol. 51, no. 9-11, pp. 1489-1492, SeptemberNovember 2011.
[5] P. Balasubramanian, D.L. Maskell, “A
distributed minority and majority voting based
redundancy
scheme,”
Microelectronics
Reliability, vol. 55, no. 9-10, pp. 1373-1378,
August-September 2015.
[6] P. Balasubramanian, N.E. Mastorakis, “A
standard cell based voter for use in TMR
implementation,”
Proc.
5th
European
Conference of Circuits Technology and
Devices, pp. 115-124, 2014.
[7] P. Balasubramanian, H.R. Arabnia, “A standard
cell based power-delay-area efficient 3-of-5
majority voter design,” Proc. 13th International
Conference on Embedded Systems and
Applications, pp. 31-35, 2015.
[8] Synopsys SAED_EDK32/28_CORE Databook,
Version 1.0.0, 2012.
[9] M.M. Vai, VLSI Design, CRC Press, Florida,
USA, 2000.
[10] K. Yano, Y. Sasaki, K. Rikino, K. Seki, “Topdown pass-transistor logic design,” IEEE
Journal of Solid-State Circuits, vol. 31, no. 6,
pp. 792-803, June 1996.
[11] P. Balasubramanian, M.R. Lakshmi Narayana,
R. Chinnadurai, “Design of combinational logic
digital circuits using a mixed logic synthesis
E-ISSN: 2224-2678
62
Volume 15, 2016
| 3cs.SY
|
arXiv:1802.09464v2 [cs.LG] 10 Mar 2018
Multi-Goal Reinforcement Learning: Challenging
Robotics Environments and Request for Research
Matthias Plappert, Marcin Andrychowicz, Alex Ray, Bob McGrew,
Bowen Baker, Glenn Powell, Jonas Schneider, Josh Tobin,
Maciek Chociej, Peter Welinder, Vikash Kumar, and Wojciech Zaremba
OpenAI
Correspondence to {matthias, marcin}@openai.com
Abstract
The purpose of this technical report is two-fold. First of all, it introduces a suite
of challenging continuous control tasks (integrated with OpenAI Gym) based on
currently existing robotics hardware. The tasks include pushing, sliding and pick
& place with a Fetch robotic arm as well as in-hand object manipulation with
a Shadow Dexterous Hand. All tasks have sparse binary rewards and follow a
Multi-Goal Reinforcement Learning (RL) framework in which an agent is told
what to do using an additional input.
The second part of the paper presents a set of concrete research ideas for improving RL algorithms, most of which are related to Multi-Goal RL and Hindsight
Experience Replay.
1
Environments
All environments are released as part of OpenAI Gym1 (Brockman et al., 2016) and use the MuJoCo
(Todorov et al., 2012) physics engine for fast and accurate simulation. A video presenting the new
environments can be found at https://www.youtube.com/watch?v=8Np3eC_PTFo.
1.1
Fetch environments
The Fetch environments are based on the 7-DoF Fetch robotics arm,2 which has a two-fingered
parallel gripper. They are very similar to the tasks used in Andrychowicz et al. (2017) but we have
added an additional reaching task and the pick & place task is a bit different.3
In all Fetch tasks, the goal is 3-dimensional and describes the desired position of the object (or
the end-effector for reaching). Rewards are sparse and binary: The agent obtains a reward of 0
if the object is at the target location (within a tolerance of 5 cm) and −1 otherwise. Actions are
4-dimensional: 3 dimensions specify the desired gripper movement in Cartesian coordinates and
the last dimension controls opening and closing of the gripper. We apply the same action in 20
subsequent simulator steps (with ∆t = 0.002 each) before returning control to the agent, i.e. the
agent’s action frequency is f = 25 Hz. Observations include the Cartesian position of the gripper, its
linear velocity as well as the position and linear velocity of the robot’s gripper. If an object is present,
we also include the object’s Cartesian position and rotation using Euler angles, its linear and angular
velocities, as well as its position and linear velocities relative to gripper.
1
https://github.com/openai/gym
http://fetchrobotics.com/
3
In Andrychowicz et al. (2017) training on this task relied on starting some of the training episodes from a
state in which the box is already grasped. This is not necessary for successful training if the target position of
the box is sometimes in the air and sometimes on the table and we do not use this technique anymore.
2
Figure 1: The four proposed Fetch environments: FetchReach, FetchPush, FetchSlide, and
FetchPickAndPlace.
Reaching (FetchReach) The task is to move the gripper to a target position. This task is very easy
to learn and is therefore a suitable benchmark to ensure that a new idea works at all.4
Pushing (FetchPush) A box is placed on a table in front of the robot and the task is to move it to
a target location on the table. The robot fingers are locked to prevent grasping. The learned behavior
is usually a mixture of pushing and rolling.
Sliding (FetchSlide) A puck is placed on a long slippery table and the target position is outside
of the robot’s reach so that it has to hit the puck with such a force that it slides and then stops at the
target location due to friction.
Pick & Place (FetchPickAndPlace) The task is to grasp a box and move it to the target location
which may be located on the table surface or in the air above it.
1.2
Hand environments
These environments are based on the Shadow Dexterous Hand,5 which is an anthropomorphic robotic
hand with 24 degrees of freedom. Of those 24 joints, 20 can be can be controlled independently
whereas the remaining ones are coupled joints.
In all hand tasks, rewards are sparse and binary: The agent obtains a reward of 0 if the goal has been
achieved (within some task-specific tolerance) and −1 otherwise. Actions are 20-dimensional: We
use absolute position control for all non-coupled joints of the hand. We apply the same action in 20
subsequent simulator steps (with ∆t = 0.002 each) before returning control to the agent, i.e. the
agent’s action frequency is f = 25 Hz. Observations include the 24 positions and velocities of the
robot’s joints. In case of an object that is being manipulated, we also include its Cartesian position
and rotation represented by a quaternion (hence 7-dimensional) as well as its linear and angular
velocities. In the reaching task, we include the Cartesian position of all 5 fingertips.
Reaching (HandReach) A simple task in which the goal is 15-dimensional and contains the target
Cartesian position of each fingertip of the hand. Similarly to the FetchReach task, this task is
relatively easy to learn. A goal is considered achieved if the mean distance between fingertips and
their desired position is less than 1 cm.
Block manipulation (HandManipulateBlock) In the block manipulation task, a block is placed
on the palm of the hand. The task is to then manipulate the block such that a target pose is achieved.
The goal is 7-dimensional and includes the target position (in Cartesian coordinates) and target
rotation (in quaternions). We include multiple variants with increasing levels of difficulty:
• HandManipulateBlockRotateZ Random target rotation around the z axis of the block.
No target position.
• HandManipulateBlockRotateParallel Random target rotation around the z axis of the
block and axis-aligned target rotations for the x and y axes. No target position.
• HandManipulateBlockRotateXYZ Random target rotation for all axes of the block. No
target position.
4
That being said, we have found that is so easy that even partially broken implementations sometimes learn
successful policies, so no conclusions should be drawn from this task alone.
5
https://www.shadowrobot.com/products/dexterous-hand/
2
Figure 2:
The four proposed Shadow Dexterous Hand environments:
HandManipulateBlock, HandManipulateEgg, and HandManipulatePen.
HandReach,
• HandManipulateBlockFull Random target rotation for all axes of the block. Random
target position.
A goal is considered achieved if the distance between the block’s position and its desired position is
less than 1 cm (applicable only in the Full variant) and the difference in rotation is less than 0.1 rad.
Egg manipulation (HandManipulateEgg) The objective here is similar to the block task but
instead of a block an egg-shaped object is used. We find that the object geometry makes a significant
differences in how hard the problem is and the egg is probably the easiest object. The goal is again
7-dimensional and includes the target position (in Cartesian coordinates) and target rotation (in
quaternions). We include multiple variants with increasing levels of difficulty:
• HandManipulateEggRotate Random target rotation for all axes of the egg. No target
position.
• HandManipulateEggFull Random target rotation for all axes of the egg. Random target
position.
A goal is considered achieved if the distance between the egg’s position and its desired position is
less than 1 cm (applicable only in the Full variant) and the difference in rotation is less than 0.1 rad.
Pen manipulation (HandManipulatePen) Another manipulation, this time using a pen instead
of a block or an egg. Grasping the pen is quite hard since it easily falls off the hand and can easily
collide and get stuck between other fingers. The goal is 7-dimensional and includes the target position
(in Cartesian coordinates) and target rotation (in quaternions). We include multiple variants with
increasing levels of difficulty:
• HandManipulatePenRotate Random target rotation x and y axes of the pen and no target
rotation around the z axis. No target position.
• HandManipulatePenFull Random target rotation x and y axes of the pen and no target
rotation around the z axis. Random target position.
A goal is considered achieved if the distance between the pen’s position and its desired position is
less than 5 cm (applicable only in the Full variant) and the difference in rotation, ignoring the z
axis,6 is less than 0.1 rad.
1.3
Multi-goal environment interface
All environments use goals that describe the desired outcome of a task. For example, in the
FetchReach task, the desired target position is described by a 3-dimensional goal. While our
environments are fully compatible with the OpenAI Gym API, we slightly extend upon it to support
this new type of environment. All environments extend the newly introduced gym.GoalEnv.
Goal-aware observation space First, it enforces a constraint on the observation space. More
concretely, it requires that the observation space is of type gym.spaces.Dict space, with at least
the following three keys:
• observation: The actual observation of the environment, For example robot state and
position of objects.
6
The z axis of the pen is parallel to its body and goes through its tip to its opposite end.
3
FetchReach-v0
FetchPush-v0
1.0
Median Test Success Rate
Median Test Success Rate
1.0
0.8
0.6
0.4
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
0.2
0.0
0
10
20
30
40
0.8
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
0.6
0.4
0.2
0.0
50
0
10
20
Epoch
FetchSlide-v0
50
1.0
Median Test Success Rate
Median Test Success Rate
0.8
40
FetchPickAndPlace-v0
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
1.0
30
Epoch
0.6
0.4
0.2
0.0
0.8
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
0.6
0.4
0.2
0.0
0
10
20
30
40
50
0
Epoch
10
20
30
40
50
Epoch
Figure 3: Median test success rate (line) with interquartile range (shaded area) for all four Fetch
environments.
• desired_goal: The goal that the agent has to achieve. In case of FetchReach, this would
be the 3-dimensional target position.
• achieved_goal: The goal that the agent has currently achieved instead. In case of
FetchReach, this is the position of the robots end effector. Ideally, this would be the
same as desired_goal as quickly as possible.
Exposed reward function Second, we expose the reward function in a way that allows for recomputing the reward with different goals. This is a necessary requirement for HER-style algorithms
which substitute goals. A detailed example is available in Appendix A
Compatibility with standard RL algorithms Since OpenAI Gym is commonly supported in most
RL algorithm frameworks and tools like OpenAI Baselines (Dhariwal et al., 2017), we include a
simple wrapper that converts the new dictionary-based goal observation space into a more common
array representation. A detailed example is available in Appendix A.
1.4
Benchmark results
We evaluate the performance of DDPG with and without Hindsight Experience Replay (HER, Andrychowicz et al. (2017)) on all environments with all its variants. We compare
the following four configurations:
•
•
•
•
DDPG+HER with sparse rewards
DDPG+HER with dense rewards
DDPG with sparse rewards
DDPG with dense rewards
Detailed hyperparameters can be found in Appendix B.
For all environments, we train on a single machine with 19 CPU cores. Each core generates experience using two parallel rollouts and uses MPI for synchronization. For FetchReach, FetchPush,
FetchSlide, FetchPickAndPlace, and HandReach, we train for 50 epochs (one epoch consists of
19 · 2 · 50 = 1 900 full episodes), which amounts to a total of 4.75 · 106 timesteps. For the remaining
4
HandReach-v0
0.8
1.0
Median Test Success Rate
Median Test Success Rate
HandManipulateBlockRotateXYZ-v0
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
1.0
0.6
0.4
0.2
0.0
0.8
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
0.6
0.4
0.2
0.0
0
10
20
30
40
50
0
25
50
75
Epoch
HandManipulateEggFull-v0
150
175
200
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
1.0
Median Test Success Rate
Median Test Success Rate
0.8
125
HandManipulatePenRotate-v0
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
1.0
100
Epoch
0.6
0.4
0.2
0.0
0.8
0.6
0.4
0.2
0.0
0
25
50
75
100
125
150
175
200
0
Epoch
25
50
75
100
125
150
175
200
Epoch
Figure 4: Median test success rate (line) with interquartile range (shaded area) for all four Fetch
environments.
environments, we train for 200 epochs, which amounts to a total of 38 · 106 timesteps. We evaluate the
performance after each epoch by performing 10 deterministic test rollouts per MPI worker and then
compute the test success rate by averaging across rollouts and MPI workers. Our implementation is
available as part of OpenAI Baselines7 (Dhariwal et al., 2017). In all cases, we repeat an experiment
with 5 different random seeds and report results by computing the median test success rate as well as
the interquartile range.
Figure 3 depicts the median test success rate for all four Fetch environments. FetchReach is clearly
a very simple environment and can easily be solved by all four configurations. On the remaining
environments, DDPG+HER clearly outperforms all other configurations. Interestingly, DDPG+HER
performs best if the reward structure is sparse but is also able to successfully learn from dense rewards.
For vanilla DDPG, it is typically easier to learn from dense rewards with sparse rewards being more
challenging.
Figure 4 depicts the median test success rate for all four hand environments. Similar to the Fetch
environments, DDPG+HER significantly outperforms the DDPG baseline. In fact, the baseline often
is not able to learn the problem at all. Similar to before, the sparse reward structure works significantly
better than the dense reward when using HER. HER is able to learn partly successful policies on all
environments but especially HandManipulatePen is especially challenging and we are not able to
fully solve it. Note that we do not depict results for all variants of the four environments in this figure.
A complete set of plots for all environments and their variants can be found in Appendix C.
We believe the reason why DDPG+HER typically performs better with sparse rewards is mainly due
to the following two reasons:
• Learning the critic is much simpler for sparse rewards. In the dense case, the critic has
to approximate a highly non-linear function that includes the Euclidean distance between
positions and the difference between two quaternions for rotations. On the other hand,
learning the sparse return is much simpler since the critic only has to differentiate between
successful and failed states.
7
https://github.com/openai/baselines
5
• A dense reward biases the policy towards a specific strategy. For instance, it may be
beneficial to first grasp an object properly and then start rotating it towards the desired
goal. The dense reward however encourages the policy to chose a strategy that achieves the
desired goal directly.
2
Request for Research
Deciding which problem is worth working on is probably the hardest part of doing research. Below we
present a set of research problems which we believe can lead to widely-applicable RL improvements.
For each problem we propose at least one potential solution but solving many of them will require
inventing new ideas. To make tracking the progress of work on these ideas easier, we would like to
ask authors to cite this report when publishing related research.
Automatic hindsight goals generation In Andrychowicz et al. (2017) the goals used for HER
were generated using a hand-crafted heuristic, e.g. replaying with a goal which was achieved at a
random future timestep in the episode. Instead, we could learn which goals are most valuable for
replay. They could be chosen from the goals achieved or seen during training or generated by a
separate neural network given a transition as input. The biggest question is how to judge which goals
are most valuable for replay. One option would be to train the generator to maximize the Bellman
error. This bears a lot of similarity to Prioritized Experience Replay (Schaul et al., 2015b) and we
expect that some techniques from this paper may be useful here.
Unbiased HER HER changes the joint distribution of replayed (state, action, next_state, goal)
tuples in an unprincipled way. This could, in theory, make training impossible in extremely stochastic
environment albeit we have not noticed this in practice. Consider an environment in which there is a
special action which takes the agent to a random state and the episode ends after that. Such an action
would seem to be perfect in hindsight if we replay with the goal achieved by the agent in the future.
How to avoid this problem? One potential approach would be to use importance sampling to cancel
the sampling bias but this would probably lead to prohibitively high variance of the gradient.
HER+HRL Hierarchical Actor-Critic (Levy et al., 2017) showed some promising results in applying HER in Hierarchical RL setup. One possible extension of this work would be to replace in
hindsight not only goals, but also higher-level actions, e.g. if the higher level asked the lower level
to reach state A, but some other state B was reached, we could replay this episode replacing the
higher-level action with B. This could allow the higher level to learn even when the lower level policy
is very bad but is not very principled and could make training unstable.
Richer value functions UVFA (Schaul et al., 2015a) extended value functions to multiple goals,
while TDM (Pong et al., 2018) extended them to different time horizons. Both of these innovations
can make training easier, despite the fact that the learned function is more complicated. What else
could we fed to the value function to improve the sample-efficiency? How about discount factor or
success threshold for binary rewards?
Faster information propagation Most state-of-the-art off-policy RL algorithms use target networks to stabilize training (e.g. DQN (Mnih et al., 2015) or DDPG (Lillicrap et al., 2015)). This,
however, comes at a price of limiting the maximum learning speed of the algorithm as each target
network update sends the information about returns only one step backward in time (if one-step bootstrapping is used). We noticed that the learning speed of DDPG+HER in the early stages of training is
often proportional to the frequency of target network updates8 but excessive frequency/magnitude of
target network updates leads to unstable training and worse final performance. How can we adapt the
frequency of target network updates (or the moving average coefficient used to update the network)
to maximize the training speed? Are there better ways to update the target network than a simple
replacement or a moving average over time? Are there other ways to stabilize training which does
not limit the learning speed (e.g. clipped objective similar to the one used in PPO (Schulman et al.,
2017b))?
8
Or to 1/(1 − averaging_coefficient) if target networks is computed using a moving average of the main
network’s parameters.
6
HER + multi-step returns HER generates data which is extremely off-policy9 and therefore multistep returns can not be used unless we employ some correction factors like importance sampling.
While there are many solutions for dealing with off-policies of the data (e.g. Munos et al. (2016)),
it is not clear if they would perform well in the setup where the training data is so far from being
on-policy. Another approach would be to use multi-step optimality tightening inequalities (He et al.,
2016). Using multi-step returns can be beneficial because the decreased frequency of bootstraping
can lead to less biased gradients. Moreover, it accelerates the transfer of information about the returns
backwards in time which, accordingly to our experiment, is often the limiting factor in DDPG+HER
training (compare previous paragraph).
On-policy HER How to combine HER with state-of-the-art on-policy RL algorithms like
PPO (Schulman et al., 2017b)? Some preliminary results with vanilla Policy Gradients were presented
by Rauber et al. (2017), but this approach needs to be tested on more challenging environments like
the ones proposed in this report. One possible option would also be to to use techniques similar to the
ones employed in IPG (Gu et al., 2017).
Combine HER with recent improvements in RL It would be interesting to see how recent
improvements in RL perform while combined with HER. The list of potential improvements is long
e.g. Prioritized Experience Replay (Schaul et al., 2015b), distributional RL (Bellemare et al., 2017),
entropy-regularized RL (Schulman et al., 2017a), or reverse curriculum generation (Florensa et al.,
2017).
RL with very frequent actions RL algorithms are very sensitive to the frequency of taking actions
which is why frame skip technique is usually used on Atari (Mnih et al., 2015). In continuous control
domains, the performance goes to zero as the frequency of taking actions goes to infinity, which is
caused by two factors: inconsistent exploration and the necessity to bootstrap more times to propagate
information about returns backward in time. How to design a sample-efficient RL algorithm which
can retain its performance even when the frequency of taking actions goes to infinity? The problem
of exploration can be addressed by using parameters noise for exploration (Plappert et al., 2017) and
faster information propagation could be achieved by employing multi-step returns. Other approach
could be an adaptive and learnable frame skip.
References
Andrychowicz, M., Wolski, F., Ray, A., Schneider, J., Fong, R., Welinder, P., McGrew, B., Tobin, J., Abbeel,
O. P., and Zaremba, W. (2017). Hindsight experience replay. In Advances in Neural Information Processing
Systems, pages 5055–5065.
Bellemare, M. G., Dabney, W., and Munos, R. (2017). A distributional perspective on reinforcement learning.
arXiv preprint arXiv:1707.06887.
Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., and Zaremba, W. (2016). Openai
gym. arXiv preprint arXiv:1606.01540.
Dhariwal, P., Hesse, C., Klimov, O., Nichol, A., Plappert, M., Radford, A., Schulman, J., Sidor, S., and Wu, Y.
(2017). OpenAI Baselines. https://github.com/openai/baselines.
Florensa, C., Held, D., Wulfmeier, M., and Abbeel, P. (2017). Reverse curriculum generation for reinforcement
learning. arXiv preprint arXiv:1707.05300.
Gu, S., Lillicrap, T., Ghahramani, Z., Turner, R. E., Schölkopf, B., and Levine, S. (2017). Interpolated policy
gradient: Merging on-policy and off-policy gradient estimation for deep reinforcement learning. arXiv
preprint arXiv:1706.00387.
He, F. S., Liu, Y., Schwing, A. G., and Peng, J. (2016). Learning to play in a day: Faster deep reinforcement
learning by optimality tightening. arXiv preprint arXiv:1611.01606.
Kingma, D. and Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
9
Notice that normally off-policy data used by RL algorithms come from an earlier version of the policy and
therefores is relatively close to on-policy data. It is not the case for HER, because we completely replace the
goals which are fed to the network.
7
Levy, A., Platt, R., and Saenko, K. (2017). Hierarchical actor-critic. arXiv preprint arXiv:1712.00948.
Lillicrap, T. P., Hunt, J. J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D., and Wierstra, D. (2015).
Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller, M.,
Fidjeland, A. K., Ostrovski, G., et al. (2015). Human-level control through deep reinforcement learning.
Nature, 518(7540):529–533.
Munos, R., Stepleton, T., Harutyunyan, A., and Bellemare, M. (2016). Safe and efficient off-policy reinforcement
learning. In Advances in Neural Information Processing Systems, pages 1054–1062.
Plappert, M., Houthooft, R., Dhariwal, P., Sidor, S., Chen, R. Y., Chen, X., Asfour, T., Abbeel, P., and
Andrychowicz, M. (2017). Parameter space noise for exploration. arXiv preprint arXiv:1706.01905.
Pong, V., Gu, S., Dalal, M., and Levine, S. (2018). Temporal difference models: Model-free deep rl for
model-based control. International Conference on Learning Representations.
Rauber, P., Mutz, F., and Schmidhuber, J. (2017). Hindsight policy gradients. arXiv preprint arXiv:1711.06006.
Schaul, T., Horgan, D., Gregor, K., and Silver, D. (2015a). Universal value function approximators. In
Proceedings of the 32nd International Conference on Machine Learning (ICML-15), pages 1312–1320.
Schaul, T., Quan, J., Antonoglou, I., and Silver, D. (2015b). Prioritized experience replay. arXiv preprint
arXiv:1511.05952.
Schulman, J., Abbeel, P., and Chen, X. (2017a). Equivalence between policy gradients and soft q-learning. arXiv
preprint arXiv:1704.06440.
Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. (2017b). Proximal policy optimization
algorithms. arXiv preprint arXiv:1707.06347.
Todorov, E., Erez, T., and Tassa, Y. (2012). Mujoco: A physics engine for model-based control. In Intelligent
Robots and Systems (IROS), 2012 IEEE/RSJ International Conference on, pages 5026–5033. IEEE.
8
A
Goal-based API Examples
Exposed reward function The following example demonstrates how the exposed reward function
can be used to re-compute a reward with substituted goals. The info dictionary can be used to store
additional information that may be necessary to re-compute the reward but that is independent of the
goal, e.g. state derived from the simulation.
import gym
env = gym . make ( ’ F e t c h R e a c h −v0 ’ )
env . r e s e t ( )
obs , r e w a r d , done , i n f o = env . s t e p (
env . a c t i o n _ s p a c e . s a m p l e ( ) )
# The f o l l o w i n g a l w a y s h a s t o h o l d :
a s s e r t r e w a r d == env . c o m p u t e _ r e w a r d (
obs [ ’ a c h i e v e d _ g o a l ’ ] , obs [ ’ d e s i r e d _ g o a l ’ ] , i n f o )
# . . . b u t you can a l s o s u b s t i t u t e g o a l s :
s u b s t i t u t e _ g o a l = o b s [ ’ a c h i e v e d _ g o a l ’ ] . copy ( )
s u b s t i t u t e _ r e w a r d = env . c o m p u t e _ r e w a r d (
obs [ ’ a c h i e v e d _ g o a l ’ ] , s u b s t i t u t e _ g o a l , i n f o )
Compatibility with standard RL algorithms The following example demonstrates how
to wrap the new goal-based environments to make their observation spaces compatible
with existing implementations. To do so, simply wrap any goal-based environment with
gym.wrappers.FlattenDictWrapper and specify the desired keys of the dictionary that you
would like to use.
import gym
env = gym . make ( ’ F e t c h R e a c h −v0 ’ )
p r i n t ( t y p e ( env . r e s e t ( ) ) )
# p r i n t s "< c l a s s ’ d i c t ’ >"
env = gym . w r a p p e r s . F l a t t e n D i c t W r a p p e r (
env , [ ’ o b s e r v a t i o n ’ , ’ d e s i r e d _ g o a l ’ ] )
ob = env . r e s e t ( )
p r i n t ( t y p e ( ob ) , ob . s h a p e )
# p r i n t s "< c l a s s ’ numpy . n d a r r a y ’> ( 1 3 , ) "
B
Hyperparameters
To ensure a fair comparison, we perform a hyperparameter search over the following parameters:
• Actor learning rate: {1 · 10−4 , 3 · 10−4 , 6 · 10−4 , 1 · 10−3 , 3 · 10−3 , 6 · 10−3 , 1 · 10−2 }
• Critic learning rate: {1 · 10−4 , 3 · 10−4 , 6 · 10−4 , 1 · 10−3 , 3 · 10−3 , 6 · 10−3 , 1 · 10−2 }
• Polyak-averaging coefficient τ : {0.9, 0.93, 0.95, 0.97, 0.99}
• Batch size: {32, 64, 128, 256}
• Probability of random action: {0, 0.1, 0.2, 0.3, 0.4}
• Scale of additive Gaussian noise: σ: {0, 0.1, 0.2, 0.3, 0.4}
• Action L2 norm coefficient: {0, 0.01, 0.03, 0.1, 0.3, 0.6, 1.}
Since searching all possible combinations exhaustivley is intractable, we randomly sample 40 combinations and train a policy on the HandManipulateBlockRotateZ-v0 environment for all four
configurations (DDPG+HER sparse, DDPG+HER dense, DDPG sparse, DDPG dense). We picked
this environment since all configurations are capable of learning on this environment. For each
configuration and combination we train with 3 random seeds and average performance across this. To
9
select the best hyperparameter combination, we numerically compute the area under the test success
rate curve and select the combination that achieves the best performance across all tasks.
All experiments in this paper use the following hyperparameters, which have been found by the
aforementioned search:
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Actor and critic networks: 3 layers with 256 units each and ReLU non-linearities
Adam optimizer (Kingma and Ba, 2014) with 1 · 10−3 for training both actor and critic
Buffer size: 106 transitions
Polyak-averaging coefficient: 0.95
Action L2 norm coefficient: 1.0
Observation clipping: [−200, 200]
Batch size: 256
Rollouts per MPI worker: 2
Number of MPI workers: 19
Cycles per epoch: 50
Batches per cycle: 40
Test rollouts per epoch: 10
Probability of random actions: 0.3
Scale of additive Gaussian noise: 0.2
Probability of HER experience replay: 0.8
Normalized clipping: [−5, 5]
All hyperparameters are described in greater detail in Andrychowicz et al. (2017).
Full Benchmark Results
FetchPickAndPlace-v0
1.0
Median Test Success Rate
C
0.8
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
0.6
0.4
0.2
0.0
0
10
20
30
Epoch
10
40
50
FetchPush-v0
Median Test Success Rate
1.0
0.8
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
0.6
0.4
0.2
0.0
0
10
20
30
40
50
Epoch
FetchReach-v0
Median Test Success Rate
1.0
0.8
0.6
0.4
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
0.2
0.0
0
10
20
30
Epoch
11
40
50
FetchSlide-v0
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
Median Test Success Rate
1.0
0.8
0.6
0.4
0.2
0.0
0
10
20
30
40
50
Epoch
HandManipulateBlockFull-v0
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
Median Test Success Rate
1.0
0.8
0.6
0.4
0.2
0.0
0
25
50
75
100
Epoch
12
125
150
175
200
HandManipulateBlockRotateParallel-v0
Median Test Success Rate
1.0
0.8
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
0.6
0.4
0.2
0.0
0
25
50
75
100
125
150
175
200
Epoch
HandManipulateBlockRotateXYZ-v0
Median Test Success Rate
1.0
0.8
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
0.6
0.4
0.2
0.0
0
25
50
75
100
Epoch
13
125
150
175
200
HandManipulateBlockRotateZ-v0
Median Test Success Rate
1.0
0.8
0.6
0.4
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
0.2
0.0
0
25
50
75
100
125
150
175
200
175
200
Epoch
HandManipulateEggFull-v0
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
Median Test Success Rate
1.0
0.8
0.6
0.4
0.2
0.0
0
25
50
75
100
Epoch
14
125
150
HandManipulateEggRotate-v0
Median Test Success Rate
1.0
0.8
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
0.6
0.4
0.2
0.0
0
25
50
75
100
125
150
175
200
Epoch
HandManipulatePenFull-v0
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
Median Test Success Rate
1.0
0.8
0.6
0.4
0.2
0.0
0
25
50
75
100
Epoch
15
125
150
175
200
HandManipulatePenRotate-v0
DDPG+HER w/ sparse rewards
DDPG w/ sparse rewards
DDPG+HER w/ dense rewards
DDPG w/ dense rewards
Median Test Success Rate
1.0
0.8
0.6
0.4
0.2
0.0
0
25
50
75
100
Epoch
16
125
150
175
200
| 2cs.AI
|
1
Storage, Communication, and Load Balancing
Trade-off in Distributed Cache Networks
arXiv:1706.10209v1 [cs.IT] 30 Jun 2017
Mahdi Jafari Siavoshani, Member, IEEE, Ali Pourmiri, Seyed Pooya Shariatpanahi
Abstract—We consider load balancing in a network of caching
servers delivering contents to end users. Randomized load
balancing via the so-called power of two choices is a wellknown approach in parallel and distributed systems. In this
framework, we investigate the tension between storage resources,
communication cost, and load balancing performance. To this
end, we propose a randomized load balancing scheme which
simultaneously considers cache size limitation and proximity in
the server redirection process.
In contrast to the classical power of two choices setup, since the
memory limitation and the proximity constraint cause correlation
in the server selection process, we may not benefit from the power
of two choices. However, we prove that in certain regimes of
problem parameters, our scheme results in the maximum load
of order Θ(log log n) (here n is the network size). This is an
exponential improvement compared to the scheme which assigns
each request to the nearest available replica. Interestingly, the
extra communication cost incurred by our proposed scheme,
compared to the nearest replica strategy, is small. Furthermore,
our extensive simulations show that the trade-off trend does not
depend on the network topology and library popularity profile
details.
Keywords
Randomized Algorithms, Distributed Caching Servers, Request Routing, Load Balancing, Communication Cost, Ballsinto-Bins, Content Delivery Networks.
I. I NTRODUCTION
A. Problem Motivation
Advancement of technology leads to the spread of smart
multimedia-friendly communication devices to the masses
which causes a rapid growth of demands for data communication [2]. Although Telcos have been spending hugely
on telecommunication infrastructures, they cannot keep up
with this data demand explosion. Caching predictable data in
network off-peak hours, near end users, has been proposed as
a promising solution to this challenge. This approach has been
used extensively in content delivery networks (CDNs) such as
Akamai, Azure, Amazon CloudFront, etc. [3], [4], and mobile
video delivery [5]. In this approach, a cache network is usually
referred to as a set of caching servers that are connected over
a network, giving content delivery service to end users.
As a schematic view of a typical cache network see Figure 1. This figure shows caching servers connected through a
The authors’ names appear in alphabetical order.
This work has been presented in part at IPDPS 2017, [1].
M. Jafari Siavoshani is with the Department of Computer Engineering,
Sharif University of Technology, Tehran, Iran (email: [email protected]).
A. Pourmiri is with the Department of Computer Engineering, University
of Isfahan, Isfahan, Iran (email: [email protected]).
S. P. Shariatpanahi is with the School of Computer Science, Institute for Research in Fundamental Sciences (IPM), Tehran, Iran (email: [email protected]).
Assigned Requests
Incoming Requests
Backhaul Network
Request Redirection
Fig. 1: A general distributed cache network.
backhaul network. These servers are responsible for delivering
contents requested by customers. Each server should assign
the demand to a responding server (which could be itself)
via an assignment strategy. In every cache network we have
three critical parameters, namely, Storage Resource, Communication Cost, and Network Imbalance. Storage, or memory
characterizes what percentage of the total content library can
be cached at each server. Communication cost is the amount
of data transferred inside the backhaul network to satisfy
content requests. Finally, network imbalance characterizes how
uniformly the requested contents’ loads are distributed among
different responding servers. This is usually measured by
comparing the load of the busiest server with the average
load of all servers after request assignments. Every request
assignment strategy, in fact, leads to a trade-off between these
three quantities.
From a practical viewpoint, we are interested in distributed
server selection strategies which are scalable in large networks. Randomized load balancing via the so-called “power of
two choices” is a well-investigated paradigm in parallel and
distributed settings [6], [7], [8], [9]. In this approach, upon
arrival of a request, the corresponding user will query about the
current loads of two independently at random chosen servers,
and then allocates the request to the least loaded server.
Considering only the load balancing perspective, Berenbrink
et al., [10], showed that in this scheme after allocating m
balls (requests, tasks, etc.) to n bins (servers, machines, etc.)
the maximum number of balls assigned to any bin, called
maximum load, is at most m/n + O(log log n) with high
probability. This only deviates O(log log n) from the average
2
↵
load and the deviation depends on the number of servers.
Although the power of two choices strategy addresses the
load balancing issue in a distributed manner, it does not
consider the role of two other important quantities, namely,
memory and communication cost. Our goal in this paper is
to extend the power of two choices framework in order to
characterize the trade-off between these three quantities. Our
results show that the maximum load, communication cost, and
servers’ memory are three entangled parameters. Thus, the
previous studies are not sufficient for designing a practical
load balancing strategy in cache networks.
B. Problem Setting and Our Contributions
In this paper, we consider a general cache network model
that entails basic characteristics of many practical scenarios.
We consider a network of n servers and a library of size
K files. Each server can cache M files in network lowtraffic hours. Let us assume a popularity distribution P =
{p1 , . . . , pK } on the library. We assume that the cache content
placement at each server is proportional to this popularity
distribution. In high-traffic hours there are n sequential file
requests, from the library, distributed among servers uniformly
at random. Every server either serves its requests or redirects
them (via an assignment scheme) to other nodes which have
cached the files. We define the maximum load L of an assignment scheme as the maximum number of allocations to any
single server after assigning all requests. The communication
cost C is the average number of hops required to deliver the
requested file to its request origin.
As the baseline assignment scheme, we consider that each
request arrived at every server is dispatched to the nearest file
replica. This scheme results in the minimum communication
cost, while ignoring maximum load of servers. We show that,
for a grid topology and every constant 0 < α < 1/2, if
K = n, M = nα , and P is a uniform distribution, this
scheme will result in the maximum load L in the interval
[Ω(log n/ log log n), O(log n)] with high probability1 (w.h.p.).
Moreover, for every constant 0 < < 1, if K = n1− and
M = Θ(1), then the maximum load is Θ(log n) w.h.p. We also
investigate the communication cost incurred in this scheme for
Uniform and Zipf popularity distributions. In
p particular, we
derive the communication cost C of order Θ( K/M ) for the
Uniform distribution in a grid topology.
In contrast, we propose a new scheme which considers
memory, maximum load, and communication cost, simultaneously. For each request, this scheme chooses two random
candidate servers that have cached the request while putting
a constraint on their distance r to the requesting node (i.e.,
the proximity constraint). Due to cache size limitation and the
proximity constraint, current results in the balanced allocation
literature cannot be carried over to our setting. Basically, we
show that here the two chosen servers will become correlated
and this might diminish the power of two choices. Since this
correlation arises from both memory limitation and proximity
constraint, the main challenge we address in this paper is
1 With high probability refers to an event that happens with probability
1 − 1/nc , for some constant c > 0.
1
Power of two choices
is achievable
1/2
Fig. 2: Suppose M = nα and r = nβ . Then the shaded area
shows the region where the power of two choices is asymptotically achievable (for more details, refer to Theorem 4).
characterizing the regimes where we can benefit from the
power of two choices and at the same time have a low
communication cost.
In particular, suppose 0 < α, β < 1/2 be two constants and let K = n, M = nα , r = nβ , and P be
a Uniform distribution. Then, for grid topology, provided
α + 2β ≥ 1 + 2(log log n/ log n), the maximum load is
Θ(log log n) w.h.p., and the communication cost is Θ(r).
Therefore, if we set β = 1−α
2 +log log n/ log n we achieve the
power of two
choices
with
the communication cost of order
1−α
Θ(r) = Θ n 2 log n . This communication cost is only
log n factor above the communicationpcost achieved by the
1−α
nearest replica strategy, which is Θ( K/M ) = Θ(n 2 ).
Figure 2 shows the region of parameters α and β where the
power of two choices is asymptotically achievable. It should
be noted that while our theoretical results are derived for grid
networks, the main reason for assuming a grid is presentation
clarity and the results can be extended to other topologies.
While our theoretical results are derived for large networks
(i.e., asymptotic analysis), our simulation results show their
validity even for finite sized networks. Also, in simulations
we investigate the problem in more diverse settings such as
considering other network topologies.
C. Related Work
Load balancing has been the focus of many papers on cache
networks [11], [12], [13], among which distributed approaches
have attracted a lot of attention (e.g., see [14], [8], and [15]).
Randomized load balancing via the power of two choices, is a
popular approach in this direction [7]. Chen et al. [16] consider
the two choices selection process, where the second choice is
the next neighbor of the first choice. In [17], Xia et al. use the
length of common prefix (LCP)-based replication to arrive at a
recursive balls and bins problem. In [16] and [17], the authors
benefit from the metaphor of power of two choices to design
algorithms for randomized load balancing. In contrast to these
works, we follow a theoretical approach to derive provable
results for cache networks.
Theoretical works investigating the power of two choices
in cache networks all just consider the role of two parameters
among Memory Resource, Load Balancing, and Communica-
3
tion Cost. These works can be summarized in three categories
as follow.
Memory Resource vs. Communication Cost trade-off has
been investigated in many works such as [5], [18], [19],
[20], and [21]. Non of these papers have considered the load
balancing issue, in cache networks which is very important in
practice.
Load Balancing vs. Memory Resource trade-off has been
investigated in [22] and [23]. In [22] the authors consider
the supermarket model for performance evaluation of CDNs.
Although the work [22] considers the memory limitation into
account, it does not consider the proximity principle which is
a central issue in our paper. Liu et al. [23] study the setting
where the clients compare the servers in terms of hit-rate
(for web applications), or bit-rate (for video applications) to
choose their favourite ones. Their setup and objectives are
different from those we consider here. Moreover, they have
not considered the effect of their randomized load balancing
scheme on communication cost.
Communication Cost vs. Load Balancing trade-off has been
investigated in [24], [25], [26], [27], [28], [29], and [30], without considering the effect of cache size limitation. Although
the works [24], [25], and [26] have mentioned this trade-off,
non of them provides a rigorous analysis.
In contrast to the standard balls and bins model, the works
[27], [28], [29], and [30] introduced the effect of proximity
constraint to the ball and bins framework. In the standard
balls and bins model, each ball (request) picks two bins
(servers) independently and uniformly at random and it is then
allocated to the one with lesser load [6]. However, in many
settings, selecting any two random servers might be infeasible
or costly. In other words, this proximity constraint translates
to a correlation between the two choices, i.e., the balls and
bins model with related choices.
The most related work to our paper is [28]. Kenthapadi and
Panigrahi [28] proposed a model where n bins are connected
as a d-regular graph. Corresponding to each ball, a node
is chosen uniformly at random as the first candidate. Then,
one of its neighbours is chosen uniformly at random as the
second candidate and the ball is allocated to the one with
the minimum load. Under this assumption, they proved that
if the graph is sufficiently dense (i.e., the average degree is
nΩ(log log n/ log n) ), then after allocating n balls, the maximum
load is Θ(log log n) w.h.p.
Although the model used in [28] considers the proximity
principle by assigning each request to the origin neighbors,
it cannot be directly applied to our cache network setup.
First, they do not consider multi-hop communication, while
in practice the communication is done in a multi-hop fashion.
Second, the above model cannot accommodate the cache size
limitation of servers. Cache size limitation introduces the
notion of cache content placement which should be based on
the popularity profile. In addition, this limitation will introduce
a new source of correlation between choices which is not
considered in [28].
The organization of the paper is as follows. In Section II,
we present our notation and problem setup. Then, in Section
III the nearest replica strategy, is investigated as the baseline
scheme. In Section IV, we propose and analyze the proximityaware two choices strategy, which at the same time considers
memory limitation, proximity constraint, and benefits from
the power of two choices. In Section V the performance of
these two schemes are investigated via extensive simulations.
Finally, our discussions and concluding remarks are presented
in Section VI.
II. N OTATION AND P ROBLEM S ETTING
A. Notation
Throughout the paper, with high probability refers to an
event that happens with probability 1−1/nc , for some constant
c > 0. Let G = (V, E) be a graph with vertex set V and edge
set E where e(G) := |E|. For u ∈ V let d(u) denote for the
degree of u in G. For every pair of nodes u, v ∈ V , dG (u, v)
denotes the length of a shortest path from u to v in G. The
neighborhood of u at distance r is defined as
Br (u) := {v : dG (u, v) ≤ r and v ∈ V (G)} .
Finally, we use Po(λ) to denote for the Poisson distribution
with parameter λ.
B. Problem Setting
We consider a cache network consisting of n caching
servers (also called cache-enabled nodes)
√ and√edges connecting neighboring servers forming a n × n grid. Direct
communication is possible only between adjacent nodes, and
other communications should be carried out in a multi-hop
fashion.
Remark 1. Throughout the paper for the sake of presentation
clarity we may consider a torus with n nodes. This helps to
avoid boundary effects of grid and all the asymptotic results
hold for the grid as well.
Suppose that the cache network is responsible for handling
a library of K files W = {W1 , . . . , WK }, whereas the popularity profile follows a known distribution P = {p1 , . . . , pK }.
The network operates in two phases, namely, cache content placement and content delivery. In the cache content
placement phase each node caches M ≤ K files randomly
from the library according to their popularity distribution
P = {p1 , . . . , pK } with replacement, independent of other
nodes. Also note that, throughout the paper we assume that
M K, unless otherwise stated.
Consider a time block during which n files are requested
from the servers sequentially that are chosen uniformly at
random. Let Di denote the number of requests (demands)
arrived at server i. Then for large n we have Di ∼ Po(1)
for all 1 ≤ i ≤ n.
For library popularity profile P, we consider two probability
distributions, namely, Uniform and Zipf with parameter γ. In
the Uniform distribution we have
pi =
1
,
K
i = 1, . . . , K,
4
which considers equal popularity for all the files. In Zipf
distribution the request probability of the i-th popular file is
inversely proportional to its rank as follows
γ
pi =
1/i
,
K
P
γ
1/j
i = 1, . . . , K,
j=1
which has been confirmed to be the case in many practical
applications [31], [32].
For any given cache content placement, an assignment
strategy determines how each request is mapped to a server.
Let Ti denote the number of requests assigned to server i at
the end of mapping process.
Now, for each strategy we define the following metrics.
respect to u and some positive number r > 0 define four
areas as follows
A1 (u) := {(x, y) : 0 ≤ y ≤ x/2 and (x, y) ∈ Br (u)},
A2 (u) := {(x, y) : 0 ≤ −x ≤ y/2 and (x, y) ∈ Br (u)},
A3 (u) := {(x, y) : 0 ≤ −y ≤ −x/2 and (x, y) ∈ Br (u)},
A4 (u) := {(x, y) : 0 ≤ x ≤ −y/2 and (x, y) ∈ Br (u)},
which are shown in Fig. 3. It is easy to see that all four areas
have the same size, that is
|A1 (u)| =
•
The communication cost of a strategy is the average
number of hops between the requesting node and the
serving node, denoted by C.
The maximum load of a strategy is the maximum number
of requests assigned to a single node, denoted by L =
max1≤i≤n Ti .
III. N EAREST R EPLICA S TRATEGY
The simplest strategy for assigning requests to servers is
to allocate each request to the nearest node that has cached
the file. This strategy, formally defined below, leads to the
minimum communication cost, while does not try to reduce
maximum load.
Definition 2 (Strategy I: Nearest Replica Strategy). In this
strategy each request is assigned to the nearest node –in the
sense of the graph shortest path distance– which has cached
the requested file. If there are multiple choices ties are broken
randomly.
Consider the set of nodes that have cached file Wj , say
Sj . According to Strategy I, each demand from node u for
file Wj will be served by arg minv∈Sj dG (u, v). This induces
a Voronoi Tessellation on the torus corresponding to file
Wj which we denote by Vj . Then, alternatively, we can
define Strategy I as assigning each request of file Wj to the
corresponding Voronoi cell center.
In order to analyze the maximum load imposed on each
node, we should investigate the size of such Voronoi regions.
The following Lemma is in this direction.
Lemma 1. Under the Uniform popularity distribution, the
maximum cell size (number of nodes inside each cell) of Vj ,
1 ≤ j ≤ K, is at most O (K log n/M ) w.h.p. In particular,
every Voronoi cell centered at
any node is contained
in a sub
p
K log n/M . Furthermore,
grid of size r × r with r = O
if K = n1− , for some constant 0 < < 1, and M = Θ(1),
then there exists a Voronoi cell of size Θ (K log n/M ) w.h.p.
Proof of Lemma 1. Upper Bound – Fix a node u and w.l.o.g.
assume that u is denoted by pair (0, 0) in the torus. With
1
y=0 x=2y
br/3c
=
X
(r − 3y + 1)
y=0
Definition 1 (Communication Cost and Maximum Load).
•
br/3c r−y
X X
br/3c
≥
X
3y
y=0
2
≥ r /8.
(1)
Let us fix some arbitrary 1 ≤ j ≤ K and for every node u
define indicator random variable Xu,j taking value 1 if u has
cached file Wj and there is no node in A1 (u) that has cached
file Wj , and 0 otherwise. Then,
M !
M (|A1 (u)|−1)
1
1
Pr [Xu,j = 1] = 1 − 1 −
1−
,
K
K
where the first term determines the probability that u caches
Wj and the second one determines the probabilitypthat nodes in
A1 (u) \ {u} do not cache Wj . By setting r = 5 K log n/M
and applying Inequality (1) we have,
M (|A1 (u)|−1)
25 log n
1
1−
= e− 8 (1 + o(1))
K
= O(n−3 ),
(2)
where it follows from 1 − 1/K = e−1/K (1 + o(1)) and
M/K = o(1). Moreover, by using the approximation
1 − (1 − 1/K)M = M (1 + o(1))/K, we have
M (1 + o(1))
.
K · n3
Therefore applying the union bound over all n nodes and
K files implies that w.h.p. for every u there exists at least one
node in A1 (u) which shares
a common file with u, supported
p
that we choose r = 5 K log n/M . We can similarly prove
the same argument for A2 (u), A3 (u) and A4 (u).
Suppose that u has cached file Wj , and we want to find
an upper bound for the size of the Voronoi cell centered at u
corresponding to Wj . In order to do this let us define v 1 =
(vx1 , vy1 ) ∈ A1 (u) to be the nearest node to u with file Wj .
Similarly define v i ∈ Ai (u), 2 ≤ i ≤ 4. W.l.o.g. assume that
u is the origin and the nodes are in a xy-coordinate system.
Define
Pr [Xu,j = 1] ≤
B := {(x, y) : vx3 ≤ x ≤ vx1 and vy4 ≤ y ≤ vy2 }.
Now we show that the Voronoi cell of u is contained in B,
and thus the size of B is an upper bound to the size of the
5
Also, we know that
X
Var [Yj ] =
Cov(Yu,j , Yv,j )
u,v
=
X
E [Yu,j Yv,j ] − E [Yu,j ] E [Yv,j ]
u,v
=
X
Pr [Yu,j = 1, Yv,j = 1]
u,v
− Pr [Yu,j = 1] · Pr [Yv,j = 1] ,
(3)
where the last equality holds because Yu,j ’s are indicator
random variables. It is easy to see that for every u and v
with dG (u, v) > 2r, Cov(Yu,j , Yv,j ) = 0 as cache content
placement at different nodes are independent processes. So
we only consider pairs u and v, with dG (u, v) ≤ 2r. Then for
each pair of nodes three following cases should be considered:
• u = v: In this case we have
Pr [Yu,j = 1, Yv,j = 1] − Pr [Yu,j = 1] · Pr [Yv,j = 1]
Fig. 3: Demonstration of regions A1 (u), . . . , A4 (u) used in
the upper bound proof of Lemma 1.
Voronoi cell. Consider Fig. 3. Let us consider node w in the
complement of B with wy > vy2 and wx > 0. Assume that
Puw is a shortest path from w to u that passes node (0, vy2 ).
By definition of A2 (u), we know the length of a shortest path
from (0, vy2 ) to (vx2 , vy2 ) is |vx2 | ≤ vy2 /2. This shows that w
is closer to v 2 than u, and by definition it does not belong
to the Voronoi cell centered at u. We similarly can show that
each arbitrary node w ∈ B c is closer to either v i ’s rather than
u. Since we arbitrarily choose u and 1 ≤ j ≤ K, there is
sub-grid B that contains every Voronoi cell in Vj , centered at
any given u. So the size of any Voronoi cell centered at an
arbitrary node u is bounded from above by |B|, which is at
most 4r2 = O(K log n/M ).
Lower Bound – Let us define indicator random variable
Yu,j for every u and some fixed j taking value 1 if u has
cached file Wj and there is no v ∈ Br (u) that has cached file
Wj , and 0 otherwise. Note that |Br (u) \ {u}| = 2r(r + 1).
Then we have
M !
M [2r(r+1)]
1
1
Pr [Yu,j = 1] = 1 − 1 −
1−
.
K
K
p
By setting r = · K · log n/4M and using similar approximations used in (2) we have
M (1 + o(1))
p , Pr [Yu,j = 1] =
.
K · n/2
P
Let Yj = u Yu,j . Then, we have the following claim.
Claim 1. For every j we have Yj ≥ 1 with probability 1−o(1).
This claim shows that there exists at least a Voronoi cell of
size Θ(r2 ) = Θ(K log n/M ) which concludes the proof.
Now, in order to prove the claim note that
X
M (1 + o(1))
E [Yj ] =
E [Yu,j ] = n ·
= M n/2 (1 + o(1)).
/2
K
·
n
u
= Pr [Yu,j = 1] − p2 = p(1 − p).
(4)
•
0 < dG (u, v) ≤ r: By definition of indicator random
variables Yu,j ’s, we have
Pr [Yu,j = 1, Yv,j = 1] − Pr [Yu,j = 1] · Pr [Yv,j = 1]
= 0 − p2 .
•
(5)
r < dG (u, v) ≤ 2r: In this case we have
Pr [Yu,j = 1, Yv,j = 1] − Pr [Yu,j = 1] · Pr [Yv,j = 1]
= Pr [Yu,j = 1|Yv,j = 1] Pr [Yv,j = 1] − p2
2M
M (1 + o(1))
p − p2 ≤
p.
(6)
≤
K
K
Now let us split the summation (3) based on dG (u, v) as
follows
X
Var [Yj ] =
Cov(Yu,j , Yu,j )
u
+
X
X
Cov(Yu,j , Yv,j )
u v:0<dG (u,v)≤r
+
X
X
Cov(Yu,j , Yv,j ).
u v:r<dG (u,v)≤2r
Applying results (4-6) yields
Var [Yj ] ≤ np(1 − p) − n|Br (u)|p2 + n|B2r (u)|
≤ np + 4nr(2r + 1)
2M p
K
2M p
K
≤ np + 6np log n
≤ 7np log n,
where we use the fact that 4r(2r + 1) ≤ 9r2 ≤ 3K log n.
Applying Chebychev’s inequality leads to
4 Var [Yj ]
Pr [|Yj − E [Yj ] | ≥ E [Yj ] /2] ≤
2
E [Yj ]
28np log n
≤
n2 p 2
28 log n
=
= o(1).
np
6
Therefore, Yj concentrates around its mean, i.e., np =
Θ(n/2 ), which proves the claim.
K − |T |. Then probability that a node in Br (u) caches Wj is
at most M/(K − t). Hence, for every Wj ∈ S,
Now, we are ready to present our main results for this
section which characterize the maximum load of Strategy I,
for two different parameter regimes, in Theorems 1 and 2.
Pr [Xu,j = 1|{Xu,j 0 = xu,j 0 , Wj 0 ∈ S \ {Wj }}]
!2r(r+1)
M
P
≥ 1−
K − Wj0 ∈S\{Wj } xu,j 0
2r(r+1)
M
= e−Ω(1) = p,
≥ 1−
K −M +1
Theorem 1. Suppose that K = n1− , for some constant 0 <
< 1, and M = Θ(1). Then, under Uniform distribution P,
Strategy I achieves maximum load of L = Θ(log n) w.h.p.
Proof. Consider node u which has cached a set of distinct
files, say S, with |S| ≤ M . Applying Lemma 1 shows that all
Voronoi cells centered at u corresponding to cached files at
u are contained in a sub-grid of size at most O(K log n/M )
w.h.p. Also in each round, every arbitrary node requests for a
file in S with probability |S|/nK ≤ M/nK, as each request
randomly chooses its origin and type. Hence, by union bound,
a node in the sub-grid may request for a file in S with
probability at most O(K log n/M ) · (M/nK) = O(log n/n).
Since there are n requests, the expected number of requests
imposed to node u is O(log n). Now using a Chernoff bound
(e.g., see Appendix A) shows that w.h.p. u has to handle at
most O(log n) requests.
On the other hand, to establish a lower bound on the
maximum load we proceed as follows. Lemma 1 shows that
there exits a Voronoi cell in Vj , for some j, such that the center
node should handle the requests of at least Θ(K log n/M )
nodes w.h.p. Also each node in the cell may request for
file Wj with probability 1/nK. So on average there are
Θ(log n/M ) requests imposed on the cell center. Similarly,
by a Chernoff bound, one can see that this node experiences
the load Θ(log n/M ), which concludes the proof for constant
M.
Remark 2. It should be noted that the same result of Θ(log n)
for the maximum load can also be proved for the Zipf
distribution. That is because the content placement distribution
is chosen proportional to the file popularity distribution P, and
consequently this result is insensitive to P. However, the proof
involves lengthy technical discussions which we omit in this
paper. For further numerical investigation on this remark refer
to Section V.
Theorem 2. Suppose that K = n and M = nα , for some 0 <
α < 1/2. Then, under the Uniform distribution, the maximum
load is in the interval [Ω(log n/ log log n), O(log n)] w.h.p.
Proof of Theorem 2. To establish upper bound O(log n) for
the maximum load, we follow the first part of proof of
Theorem 1. To obtain a lower bound, consider an arbitrary
server u that has cached file set S with s distinct files. Note
that by Lemma 2, we have for every node u, s = Θ(M ) with
high probability. Let us define the indicator random variable
Xu,j , Wj ∈ S, taking p
1 if the nearest replica of Wj is outside
of Br (u), where r = K/2M and zero otherwise. It is easy
to see that Xu,j ’s are correlated. For example, consider the
set of files T = {Wj1 , Wj2 , . . . , Wjt } ⊂ S, where Xu,j 0 = 1
for every Wj 0 ∈ T . Then conditioned on this event, each node
in Br (u) has cached files from a subset of the library of size
where |Br (u)\{u}|P= 2r(r +1) = Θ(K/M ) and hence p is a
constant. Let Z = Wj ∈S Xu,j and then E [Z] ≥ s · p. Using
a Chernoff bound for moderately correlated indicator random
variables (e.g., see Lemma 4) implies that
Pr [Z < sp/2] = o(1/n2 ).
Therefore Br (u) does not contain any replica of at least p/2
fraction of files cached at u, namely S 0 . Using the union bound
over all nodes we deduce the similar statement for every node
with probability at least 1 − o(1/n). Therefore, for every u we
have,
Pr [u severs a request] ≥
|Br/2 (u)| |S 0 |
·
= Ω(1/n)
n
K
where it follows from |S 0 | = Θ(M ), |Br/2 (u)| = Θ(K/M ).
Since there are n requests, it is easy to see that the load of
each server is bounded from below by a Poisson distribution
Po(c), where c is a constant. On the other hand, it is known
that (e.g., see [33]) the maximum number taken by n Poisson
distribution Po(c) is Ω(log n/ log log n) w.h.p. and hence the
lower bound is proved.
Next, we investigate the communication cost of Strategy I
in the following theorem.
Theorem 3. Under the Uniform popularity distribution,
p Strategy I achieves the communication cost C = Θ( K/M ),
for every M K. Under Zipf popularity distribution with
M = Θ(1), it achieves
p
Θ
K/M
:
0 < γ < 1,
p
Θ
K/M log K
:
γ = 1,
√
1−γ/2
C=
Θ K
/ M
:
1 < γ < 2,
(7)
√
Θ log K/ M
:
γ = 2,
√
Θ 1/ M
:
γ > 2.
Proof of Theorem 3. Assume a request from an arbitrary node
u for file Wj . The probability that this file is cached at another
node v is qj := 1 − (1 − pj )M . Cache content placement at
different nodes is independent. Thus, the number of nodes
which should be probed is a geometric random variable with
success probability qj . This results in the average 1/qj trials
that leads to expected distance of
!
1
1
Θ √
=Θ p
.
(8)
qj
1 − (1 − pj )M
7
When averaged over different files we will have
!
K
X
1
C=
pj Θ p
.
1 − (1 − pj )M
j=1
cached the requested file. Then, the request is assigned to the
node with lesser load. Ties are broken randomly.
(9)
For the sake of illustration, first, we consider some examples
in the following.
•
For Uniform distribution we have pj = 1/K and then
p
C = Θ( K/M ).
(10)
•
For Zipf distribution with M = Θ(1) we have
!
K
X
1
C=
pj Θ p
1 − (1 − pj )M
j=1
!
K
X
1
=
pj Θ p
pj M
j=1
PK −γ/2
j
j=1
= Θ
.
PK −γ 1/2
M j=1 j
Example 1 (M = K and r = ∞2 ). In this example each node
can store all the library and there is no constraint on proximity.
As mentioned in Section I, the number of files that should be
handled by each node (i.e., Di for i = 1, . . . , n) will be a
Po(1) random variable. In this case, according to Strategy II,
two random nodes are chosen from all network nodes and the
request is assigned to the node with lesser load.
Therefore, in terms of maximum load, this problem is
reduced to the standard power of two choices model in the
balanced allocations literature [6]. In this model there are
n bins and n sequential balls which are randomly allocated
to bins. In every round each ball picks two random bins
uniformly, and it is then allocated to the bin with lesser load
[6]. Then it is shown that the maximum load of network is
L = maxi Ti = log log n(1 + o(1)) w.h.p. [6], which is an
exponential improvement compared to Strategy I.
(11)
PK
Define Λ(γ) := j=1 j −γ , for every γ. On the other hand it
is known that for every γ > 0 (e.g., see [34])
0 < γ < 1,
Θ K 1−γ ,
Θ (log K) ,
γ = 1,
Λ(γ) =
(12)
Θ(1),
γ > 1.
Now inserting the above equations into (11) completes the
proof.
Theorem 3 shows how non-uniform file popularity reduces
communication cost. The skew in file popularity is determined
by the parameter γ which will affect the communication cost.
For example, for γ < 1 communication cost is similar to the
Uniform distribution, while for γ > 2, it becomes independent
of K.
Since in Strategy I we have assigned each request to the
nearest replica, Theorem 3 characterizes the minimum communication cost one can achieve. However, Theorems 1 and 2
show a logarithmic growth for the maximum load as a function
of network size n. This imbalance in the network load is because in Strategy I each request assignment does not consider
the current load of servers. A natural question is whether, at
each request allocation, one can use a very limited information
of servers’ current load in order to reduce the maximum load.
Also one can ask how does this affect the communication cost.
We address these questions in the following section.
IV. P ROXIMITY-AWARE T WO C HOICES S TRATEGY
Strategy I introduced in the previous section will result in
the minimum communication cost, while, the maximum load
for that strategy is of order Ω (log n/ log log n). In this section
we investigate an strategy which will result in an exponential
decrease in the maximum load, i.e., reduces maximum load to
Θ (log log n), formally defined as follows.
Definition 3 (Proximity-Aware Two Choices Strategy). For
each request born at an arbitrary node u consider two
uniformly at random chosen nodes from Br (u), that have
However, in contrast to Example 1, in cache networks
usually each node can store only a subset of files, and this
makes the problem different from the standard balls and bins
model, considered in [6]. Here, due to the memory constraint
at each node, the choices are much more limited than the
M = K case. In other words here we have the case of
related choices. In the related choices scenario, the event of
choosing the second choice is correlated with the first choice;
this correlation may annihilate the effect of power of two
choices as demonstrated in Example 2.
Example 2 (K = n, M = Θ(1), and r = ∞). In this regime,
there is a subset of the library, say S, with |S| = Θ(n), whose
files are cached inside the network. On the other hand, each
file type is requested Po(1) times and hence w.h.p. there will
be a file in S which is requested Θ(log n/ log log n) times
(e.g., see [33]). Since each file in S is replicated at most M
times, requests for the file are distributed among at most M
nodes and thus the maximum load of the corresponding nodes
will be at least Θ(log n/ log log n)/M . Hence, due to memory
limitation we cannot benefit from the power of two choices.
Although Example 2 shows that memory limitation can
annihilate the power of two choices but this is not always the
case. Example 3 shows even for M = 1 for some scenarios
we can achieve L = O(log log n).
Example 3 (K = n1− for every constant 0 < < 1,
M
r = ∞). For any popularity distribution P where
PK= 1, and −c
= o(1), Strategy II achieves maximum load
j=1 (pj n)
L = O(log log n) w.h.p. Also, notice that Uniform and Zipf
,
1
distributions satisfy this requirement, whenever ∈ γ−1
γ
for γ ≥ 1, where γ is the Zipf parameter.
Roughly speaking, when M = 1, we may partition the
servers based on their cached file and hence we have K “dis√
r = ∞) is equivalent to
√should be noted that r ≥ n (including
√
r = n. Thus in this paper we use r = n and r = ∞ alternatively.
2 It
8
joint” subsets of servers. Similarly there are K request types
where each request should be addressed by the corresponding
subset of servers. Thus, here we have K disjoint Balls and
Bins sub-problems, and the sub-problem with maximum load
determines the maximum load of the original setup. The reason
that here, in contrast to Example 2, we can benefit from power
of two choices is the assumption of K n.
Proof of Example 3. It is easy to see that for M = 1, the
number of caching servers with a specific file, say Wj denoted
by Sj , is distributed as a Bin(n, pj ). Thus applying a Chernoff
bound for Sj (e.g., see Appendix A) implies that
Pr [|Sj − E [Sj ] | ≥ E [Sj ] /2] ≤ 2 exp(−pj n/12).
Moreover, let Rj denote the number of requests for file Wj ,
which is the sum of n i.i.d. Bin(n, pj ) random variables. Again
applying a Chernoff bound (e.g., see Appendix A) for Poisson
random variables yields that
Pr [|Rj − E [Rj ] | ≥ E [Rj ] /2] ≤ 2 exp(−pj n/12).
Notice that E [Sj ] = E [Rj ] = npj . Suppose that Aj denotes
the event that |Sj − E [Sj ] | ≤ E [Sj ] /2 and |Rj − E [Rj ] | ≤
E [Rj ] /2. Then we have that Pr [Aj ] ≥ 1 − 4 exp(−pj n/12).
Also define Ej to be the event that the two-choice model with
Sj bins (caching servers) and Rj balls (requests) achieves
maximum load Rj /Sj + Θ(log log Sj ). It is shown that this
event happens with probability 1−O(1/Sjc ), for every constant
c (e.g., see [6]). So we have that
Pr [Ej ] = Pr [Ej |Aj ] Pr [Aj ] + Pr [Ej |¬Aj ] Pr [¬Aj ]
> (1 − 2(pj n)−c )(1 − 4 exp(−pj n/12))
+ (1 − 2(pj n)−c )(4 exp(−pj n/12))
≥ 1 − 8(pj n)−c .
Since we have K disjoint subsystems, the union bound over all
subsystems shows that the two choice model P
does achieve the
K
desired maximum load with probability 1−8 j=1 (npj )−c =
1 − o(1) which concludes the proof due to example’s assumption on popularity profile.
Now we show that the Uniform and Zipf distributions satisfy
the example’s assumption. When P is the Uniform distribution
over K files, ∀j, pj · n = n . Now by setting c = 3/, we
have that
K
X
(npj )−c = K(1/n )c = K/n3 = o(1/n2 ).
•
0 < β < 1: By setting c = 2/γ and using the fact that
K < n, we have
c
(1−γ)c
Λ(γ)
K
≤ n(1−γ)c−c Λ(γc)
Λ(γc) =Θ
n
nc
=n−γc Λ(γc) = o(1),
where we applied (12).
Above examples bring to attention the following question.
Question 1. In view of the memory limitation at each server
in cache networks, what are the regimes (in terms of problem
parameters) one can benefit from the power of two choices to
balance out the load?
Addressing Question 1, for the general M > 1 case, is more
challenging than Example 3 and needs a completely different
approach. The simplicity of case M = 1 is that there is no
interaction between K Balls and Bins sub-problems. On the
other hand, consider M > 1. If a request, say Wj , should
be allocated to a server then the load of two candidate bins
that have cached Wj should be compared. However, load of
other file types will also be accounted for in this comparison.
So there is flow of load information between different subproblems which makes them entangled.
In all above examples, we have not considered the proximity
constraint, i.e., r = ∞, yet.√ This results in a fairly high
communication cost C = Θ ( n). However, in general since
parameter r controls the communication cost, it can √
be chosen
to be much less than the network diameter, i.e., Θ( n). This
proximity awareness introduces another source of correlation
(other than the memory limitation) between the two choices.
Thus, considering the proximity constraint may annihilate
the power of two choices even in large memory cases as
demonstrated in the following example.
Example 4 (M = K and r = 1). In this example, when a
request arrives at a server, the server chooses two random
choices among itself and its neighbours. Then the request
is allocated to the one with lesser load. Since there exists
a server at which maxi Di = Θ(log n/ log log n) requests
arrive, maximum load of network (i.e., L = maxi Ti ) will
be at least Θ(log n/ log log n)/5.
Thus, similar to Question 1 regarding the memory limitation
effect, one can pose the following question regarding proximity principle.
j=1
Also, for Zipf distribution we have
j −γ
j −γ
=
.
pj = PK
−γ
Λ(γ)
j=1 j
Depending on γ, we consider two cases:
• γ ≥ 1: For every c > 1 we have
c
c
Λ(γ)
log K
Λ(γc) = Θ
Λ(γc) = o(1),
n
nc
where we used K < n and (12).
Question 2. In view of the proximity constraint of Scheme II,
what are the regimes (in terms of problem parameters) one
can benefit from the power of two choices to balance out the
load?
In order to completely analyze load balancing performance
of Scheme II, one should consider both sources of correlation
simultaneously (which is not the case in above examples). To
this end, in the following, we investigate two memory regimes,
namely M = K and M = nα for some 0 < α < 1/2.
Our main result for M = nα is presented in the following
theorem.
9
Theorem 4. Suppose that 0 < α, β < 1/2 be two constants
and let K = n, M = nα , and r = nβ . Then if
α + 2β ≥ 1 + 2 log log n/ log n,
under the Uniform popularity distribution, Strategy II achieves
maximum load L = Θ(log log n) and communication cost
C = Θ(r) w.h.p.
Remark 3. To have a more accessible proof, in Theorem 4,
we have assumed that K = n. Note that the proof techniques
can also be extended to the case where K = O(n).
In order to prove the theorem, let us first present an
interesting result that was shown in [28] as follows.
Theorem 5 ([28]). Given an almost ∆-regular graph3 G
with e(G) edges and n nodes representing n bins, if n balls
are thrown into the bins by choosing a random edge with
probability at most O(1/e(G)) and placing into the smaller
of the two bins connected
by the edge,
then the maximum load
log n
+ O(1) w.h.p.
is Θ(log log n) + O log(∆/
log4 n)
Remark 4. Note that in the original theorem presented in
[28], it is assumed that each edge is chosen uniformly among
all edges of graph G. However, here we slightly generalize
the result so that each edge is chosen with probability at most
O(1/e(G)). The proof follows the original proof’s idea with
some modifications in computation parts, where due to lack of
space we omit.
In order to apply Theorem 5, we first need to define a new
graph H as follows.
Definition 4 (Configuration Graph). For the given parameter
r, configuration graph H is defined as a graph whose nodes
represent the servers and two nodes, say u and v, are
connected if and only if u and v have cached a common file
and d(u, v) ≤ 2r in the torus.
For every two servers u and v, let T (u, v) be the set of
distinct files that have been cached in both nodes u and v.
Also denote |T (u, v)| by t(u, v). Define t(u) to be the number
of distinct cached files in u. Now, let us define goodness of a
placement strategy as follows.
Definition 5 (Goodness Property). For every positive constant
δ ∈ [0, 1] and µ = O(1), we say the file placement strategy is
(δ, µ)-good, if for every u and v, t(u) ≥ δM and t(u, v) < µ.
Lemma 2. The proportional cache placement strategy introduced in Section II, is (δ, µ)-good w.h.p.
Proof. Clearly, every set of cached files in every node (with
replacement) can be one-to-one
PK mapped to a non-negative
integral solution of equation i=1 xi = M , where each xi
expresses the number of times that file i has been cached in
the node. A combinatorial
argument shows that, the equation
−1
has K+M
non-negative
integer solutions. So for each
M
1 ≤ s ≤ M , we have
K M −1
Pr [t(u) = s] =
3A
s
M −s
K+M −1
M
,
(13)
graph is said to be almost ∆-regular, if each vertex has degree Θ(∆).
where we first fixed a set of file indexes of size s, say
I = {i1 , i2 , . . . , is }, and P
then count the number of integral
solutions to the equation i∈I xi = M − s.
In order to
bound (13), we note that for every 1 ≤ a ≤ b,
(b/a)a ≤ ab ≤ ba and also ab ≤ 2b . Recall that we assumed
K = n and M = nα , 0 < α < 1/2. Hence for every 1 ≤ s ≤
δM , we have
Pr [t(u) = s] ≤
K s 2M
≤
K
M
K s 2M
= (2M )M K s−M
(K/M )M
≤ (2nα nδ−1 )M .
Thus, by choosing δ = (1 − α)/3, for every 1 ≤ s ≤ δM , we
have
Pr [t(u) = s] ≤ (2nα+δ−1 )M = (2n2α/3−2/3 )M
≤ (2n−1/3 )M = n−ω(1) ,
where the last equality follows from M = nα = ω(1). Now
the union bound over all 1 ≤ s ≤ δM and n nodes yields
Pr [∃u ∈ V : t(u) ≤ δM ] = n−ω(1) .
(14)
By a similar argument, for each 1 ≤ t ≤ M and every u and
v, we have
!2
K+M −t−1
K
M −t
Pr [t(u, v) ≥ t] =
.
K+M −1
t
M
Thus, for any constant µ ≥ 5/(1 − 2α), we can write
Pr [t(u, v) ≥ µ]
2
(K + M − µ − 1)!M !
µ
≤K
(K + M − 1)!(M − µ)!
µ 2
M
M 2µ
≤ Kµ
= n(2α−1)µ = O(1/n5 ).
≤
Kµ
Kµ
By applying the union bound over all pairs of servers, for
every u and v we have
Pr [t(u, v) ≥ µ] = O(1/n3 ).
(15)
Hence, t(u, v) < µ w.h.p. Putting (14) and (15) together
concludes the proof.
The following lemma presents some useful properties of H
and Strategy II.
Lemma 3. Conditioning on “goodness” of the file placement
and assuming K = n, M = nα and r = nβ with α + 2β ≥
1 + 2 log log n/ log n, we have
2 2
(a) W.h.p. H is almost ∆-regular with ∆ = Θ MKr .
(b) For each request, Strategy II samples an edge of H (two
servers) with probability O(1/e(H)).
Proof. Consider an arbitrary node u with s distinct files. Then
by definition of H, for every node v we have
M
K −s
ps := Pr [t(u, v) ≥ 1|t(u) = s] = 1 −
K
sM
=
(1 + o(1)),
K
10
where 1 ≤ s ≤ M . On the other hand u and v are connected in
H, if in addition we have dG (u, v) ≤ 2r. Therefore for every
given node u with s distinct cached files, d(u) in H (degree
of u in H) has a binomial distribution Bin(b2r (u), ps ), where
b2r (u) = |B2r (u)|. Hence applying a Chernoff bound implies
that with probability 1 − n−ω(1) , we have
d(u) =
sM b2r (u)
(1 + o(1)).
K
Conditioning on the goodness of file placement, s = t(u) =
Θ(M ). Also by symmetry of torus, we have b2r (u) = Θ(r2 ),
for every u. So, with high probability for every u, we have
d(u) = Θ M 2 r2 /K ,
where this concludes the proof of part (a).
Now it remains to show that Strategy II picks an edge of
H, with probability O(1/e(H)). First, notice that
e(H) = Θ nM 2 r2 /K = Θ(M 2 r2 ),
(16)
as K = n. Then recall that each file is cached in every node
with probability p = 1 − (1 − 1/K)M = M/K(1 + o(1)),
independently. For any given node u and file Wj , let Fj (u)
be the number of nodes at distance at most r that have cached
file Wj . Then Fj (u) has a binomial distribution Bin(br (u), p),
where br (u) = |Br (u)|. So
E [Fj (u)] = br (u) · p = Θ(r2 M/K),
where br (u)
=
Θ(r2 ) for every u. Since
α + 2β ≥ 1 + 2 log log n/ log n, we have E [Fj (u)] =
ω(log n), for every u and j. Now, applying a Chernoff bound
for Fj (u) implies that with probability 1 − n−ω(1) , Fj (u)
concentrates around its mean and hence, w.h.p., we have for
every u and j
Fj (u) = Θ(r2 M/K) = Θ(r2 M/n).
Consider an edge (u, v) ∈ E(H), with t(u, v) = t. Define
Su,v to be the set of nodes that may pick pair u and v randomly
in Strategy II. It is not hard to see that |Su,v | = O(r2 ). Now
we have,
Pr [(u, v) ∈ E(H) is picked by Strategy II|t(u, v) = t]
X 1 X 1
1
=
K
n Fj (w)
j∈T (u,v)
1
= 2
n
=
1
n2
X
w∈Su,v
X
2
1
Fj (w)
2
j∈T (u,v) w∈Su,v
X
X
Θ(n2 /r4 M 2 ).
(17)
j∈T (u,v) w∈Su,v
Conditioned on “goodness,” we have for every (u, v) ∈ E(H),
1 ≤ t(u, v) < µ. So (17) can be simplified as
Pr [(u, v) ∈ E(H) is picked by Strategy II]
≤ Θ(µ|Su,v |/r4 M 2 )
= O(1/r2 M 2 ) = O(1/e(H)),
where the last equality follows from (16).
Proof of Theorem 4. Applying Lemma 3 shows that w.h.p. the
configuration graph H is an almost ∆-regular graph where
∆ = M 2 r2 /n. Moreover, in each step, every edge of H is
chosen randomly with probability O(1/e(H)). Hence, we can
apply Theorem 5 and conclude that w.h.p. the maximum load
is at most
log n
= Θ(log log n) + O(1),
Θ(log log n) + O
log(∆/ log4 n)
where it follows because α + 2β ≥ 1 + 2 log log n/ log n and
hence ∆ = M 2 r2 /n = n2α+2β−1 > nα .
Now let us present our next result regarding to the M = K
regime.
Theorem 6. Suppose M = K and Uniform distribution P
over the file library. Then Strategy II achieves the maximum
load L = Θ (log log n) and communication cost C = Θ nβ
for any β = Ω(log log n/ log n).
Proof. Let us choose r = nβ , for some β =
Ω(log log n/ log n). By the assumption M = K, the configuration graph H (corresponding to r) is a graph in which two
nodes u and v are connected if and only if d(u, v) ≤ 2r. Since
our network is symmetric, for every u, |Br (u)| = Θ(r2 ) and
hence H is a regular graph with ∆ = Θ(r2 ). Also it is not
hard to see that Strategy II is equivalent to choosing an edge
uniformly from H. Applying Theorem 5 ([28]) to H results in
the maximum load of Θ(log log(n)). In addition, choosing two
random nodes in |Br (u)| = Θ(r2 ) results in communication
cost of C = Θ(r) = Θ nβ .
The main point of Theorem 6 is that we should just endure
C = Θ nβ , for β = Ω(log log n/ log n), to benefit from the
luxury of power of two choices, which is a very encouraging
result.
V. S IMULATIONS
In this section, we demonstrate the simulation results for
two strategies introduced in the previous sections, namely,
nearest replica and proximity aware two choices strategies.
Our simulations are implemented in Python where the code is
available online at [35].
Figure 4 shows the maximum load of Strategy I as a
function of the number of servers where different curves
correspond to different cache sizes. The network graph is a
torus, where 100 files with Uniform popularity are placed
uniformly at random in each node. Each point is an average
of 10000 simulation runs. This figure is in agreement with
the logarithmic growth of the maximum load, asymptotically
proved in Section III, even for the intermediate values of
n ≈ 100, . . . , 3000, which makes the results of Section III
more general. Comparing different curves reveals the fact that
in larger cache size setting, we have a more balanced network.
That happens because enlarging cache sizes results in a more
uniform Voronoi tessellation, i.e., having cells with smaller
variation in size.
Furthermore, Figure 5 shows the communication cost of
Strategy I as a function of cache size where different curves
8
9
7.5
8
7
7
Maximum load
maximum load
11
6.5
6
5.5
Cache size = 1
Cache size = 2
5
4
0
500
1000
1500
2000
2500
5
4
3
Cache size = 10
Cache size = 100
4.5
Strat. I, cache size = 1
Strat. II, cache size = 1
Strat. II, cache size = 2
Strat. II, cache size = 10
6
2
3000
3500
# of servers
1
0
1000
2000
3000
4000
5000
6000
# of servers
Fig. 4: The maximum load versus number of servers for
Strategy I. Here, the network topology is a torus, the file
popularity is Uniform, and we have K = 100.
average cost (# of hops)
25
Library size = 100
Library size = 1000
Library size = 2000
20
15
10
5
0
0
10
20
30
40
50
60
70
80
90
100
Cache size (# of files)
Fig. 5: The communication cost versus cache size for Strategy I. Here, the network topology is a torus of size 2025 and
the file popularity is Uniform.
correspond to different library sizes. Here, the network graph
is a torus of size 2025 and each point is an average of 10000
simulation runs. This figure is in agreement with the result of
Theorem 3.
In order to simulate Strategy II, first we set r = ∞ to
study the effect of cache size on the maximum load and
11
Cache size = 1
Cache size = 2
Cache size = 10
Cache size = 100
10
maximum load
9
8
7
6
5
4
3
0
2
4
6
# of servers
8
10
12
4
x 10
Fig. 6: The maximum load versus number of servers for
Strategy II. Here, the network topology is torus, the file
popularity is Uniform, and the library size is K = 2000.
Moreover, we assume r = ∞.
Fig. 7: The maximum load versus number of servers for
Strategies I and II. Here, the network topology is torus, the
file popularity is Uniform, and the library size is K = 100.
Moreover, we assume r = ∞.
communication cost, and then consider the effect of limited
r on the performance of the system. Figure 6 shows the
maximum load of the network versus number of servers where
each curve demonstrates a different cache size. The network
graph is a torus, where 2000 files with Uniform popularity
are placed uniformly at random in each node. Each point is
an average of 800 simulation runs. In each curve, since cache
size and number of files are fixed, increasing the number of
servers translates to increasing each file replication. Figure 6
demonstrates the system performance for large system sizes,
i.e., n ≈ 104 , . . . , 105 . However, to get a better understanding
of network behavior, let us compare the load balancing performance of Strategies I and II in Figure 7 where the file library
size is K = 100 and n ≈ 103 .
In Figure 7, when the file replication is low, due to high
correlation between the two choices of Strategy II, power of
two choices is not expected. This is reflected in Figure 7; for
example in the curve corresponding to M = 1 for n ≤ 400 we
have a fast growth in maximum load which mimics the load
balancing performance of Strategy I. We see the same trend
for the curve corresponding to M = 2 for n ≤ 200. However,
assuming M ≥ 2, for n > 1000, since there is enough file
replication in the network, the load balancing performance is
greatly improved due to the power of two choices. This is in
accordance with the lessons learned from Section IV. Also, in
between, we have a transition region where a mixed behavior
is observed. Observations made above from Figure 7 have an
important practical implication. Since employing Strategy II
is only beneficial in networks with high file replication, for
other situations with limited cache size, the less sophisticated
Strategy I is a more proper choice.
Figure 8 draws the communication cost versus number of
servers for various cache sizes for similar setting used in
Figure 6. Since in this figure there is no constraint on√the
proximity, the communication cost growth is of order Θ( n).
In simulations presented so far, we only considered the case
r = ∞. In order to investigate the effect of parameter r on
the performance of the system, in Figure 9, we have simulated
network operation for different values of r. This results in a
12
8
100
90
7
average cost (# of hops)
80
Maximum load
70
60
50
40
Cache size = 1
Cache size = 2
Cache size = 10
Cache size = 100
30
20
10
0
2
4
6
8
10
5
4
3
12
# of servers
Strat. I, Zipf (gamma=0.0)
Strat. II, Zipf (gamma=0.0)
Strat. I, Zipf (gamma=1.0)
Strat. II, Zipf (gamma=1.0)
Strat. I, Zipf (gamma=1.5)
Strat. II, Zipf (gamma=1.5)
6
2
4
x 10
0
1000
2000
3000
4000
5000
6000
# of servers
Fig. 8: The communication cost versus number of servers for
Strategy II. Here, the settings are similar to that of Figure 6.
9
Cache size = 1
Cache size = 2
Cache size = 5
Cache size = 10
Cache size = 20
Cache size = 50
Cache size = 200
7
16
6
14
5
12
4
3
0
2
4
6
8
10
12
14
16
18
20
average cost (# of hops)
Fig. 9: The trade-off between the maximum load and communication cost for Strategy II. Here, the network topology is
a torus of size 2025, the file popularity is Uniform, and the
library size is K = 500.
trade-off between the maximum load and communication cost,
as shown in Figure 9. Here we consider a torus with 2025
servers, where 500 files with Uniform popularity are placed
uniformly at random in each node. Each point is an average
of 5000 simulation runs.
In this figure, like before (i.e., Figure 7), we observe two
performance regimes based on the file replication in the network. In high memory regime, e.g., for curves corresponding
to M = 50 and M = 200, we can achieve the power of
two choices by sacrificing a negligible communication cost.
On the other hand, in low memory regime, i.e., M = 1, we
cannot decrease the maximum load even at the expense of high
communication cost values. For intermediate values of M , we
clearly observe the trade-off between the maximum load and
communication cost.
All above simulations investigated the performance for
networks with torus topology and Uniform file popularity
distribution, being in agreement with our theoretical results’
indications. However, one may ask how sensitive are our
findings to the network topology and file popularity choices.
Thus, in the following, we examine network performance
for Zipf file popularity profile and more practical network
topologies, namely, random geometric graph (RGG) and power
law random graph model [36].
Maximum load
maximum load
8
Fig. 10: The maximum load versus number of servers for
Strategies I and II. Here, the network topology is torus, the file
popularity is Zipf, the library size is K = 64, and the cache
size is M = 2. Moreover, we assume r = ∞. Each point is
an average over 1000 independent simulation runs.
Strat. I, Zipf (gamma=0.0)
Strat. II, Zipf (gamma=0.0)
Strat. I, Zipf (gamma=1.0)
Strat. II, Zipf (gamma=1.0)
Strat. I, Zipf (gamma=1.5)
Strat. II, Zipf (gamma=1.5)
10
8
6
4
2
0
1000
2000
3000
4000
5000
6000
# of servers
Fig. 11: The maximum load versus number of servers for
Strategies I and II. Here, the network topology is RGG. The
remaining settings are similar to that of Figure 10.
Figures 10, 11, and 12, show the maximum load versus the
number of servers for different network topologies (namely,
torus, RGG, and power law model) and Zipf distribution with
parameter4 γ ∈ {0, 1, 1.5}.
Also Figures 13, 14, and 15, demonstrate the performance
trade-off between maximum load and communication cost for
different network topologies (namely, torus, random RGG, and
power law model) and Zipf distribution with parameter γ ∈
{0, 1, 1.5, 2}. All these simulations show that the trends and
the trade-off we found in our theoretical results are also valid
for more practical network settings.
For convenience, a summary of simulation parameters are
stated in Table I.
VI. D ISCUSSION AND C ONCLUDING R EMARKS
In this section, we first discuss three important practical
issues related to our proposed scheme, then we will conclude
the paper.
4 Note that the Zipf distribution with γ = 0 corresponds to the Uniform
distribution.
13
16
90
Maximum load
70
60
50
maximum load (# of requests)
Strat. I, Zipf (gamma=0.0)
Strat. II, Zipf (gamma=0.0)
Strat. I, Zipf (gamma=1.0)
Strat. II, Zipf (gamma=1.0)
Strat. I, Zipf (gamma=1.5)
Strat. II, Zipf (gamma=1.5)
80
40
30
20
Gamma = 0.0
Gamma = 1.0
Gamma = 1.5
Gamma = 2.0
14
12
10
8
6
4
10
2
0
0
0
1000
2000
3000
4000
5000
6000
Fig. 12: The maximum load versus number of servers for
Strategies I and II. Here, the network topology is power law
random graph. The remaining settings are similar to that of
Figure 10.
maximum load (# of requests)
Gamma = 0.0
Gamma = 1.0
Gamma = 1.5
Gamma = 2.0
5.5
5
4.5
4
3.5
3
0
2
4
6
8
10
12
14
average cost (# of hops)
Fig. 13: The trade-off between the maximum load and communication cost for Strategy II. Here, the network topology
is a torus of size 2025, the file popularity is Zipf, the library
size is K = 500, and the cache size is M = 10. Each point
is an average over 5000 independent simulation runs.
maximum load (# of requests)
5.5
Gamma = 0.0
Gamma = 1.0
Gamma = 1.5
Gamma = 2.0
5
4.5
4
3.5
3
0
1
2
3
4
5
6
average cost (# of hops)
Fig. 14: The trade-off between the maximum load and communication cost for Strategy II. Here, the network topology
is a RGG of size 2025. The remaining settings are similar to
that of Figure 13.
2
3
4
5
6
7
Fig. 15: The trade-off between the maximum load and communication cost for Strategy II. Here, the network topology
is a power law random graph of size 2025. The remaining
settings are similar to that of Figure 13.
Fig. Number
4, 5
6, 8
7
9
10
11
12
13
14
15
6
1
average cost (# of hops)
# of servers
Net. Topology
Torus
Torus
Torus
Torus
Torus
RGG
Power Law
Torus
RGG
Power Law
Strategy
Strategy I
Strategy II (r = ∞)
Strategy I, II (r = ∞)
Strategy II (r varying)
Strategy I, II (r = ∞)
Strategy I, II (r = ∞)
Strategy I, II (r = ∞)
Strategy II (r varying)
Strategy II (r varying)
Strategy II (r varying)
Popularity
Uniform
Uniform
Uniform
Uniform
Zipf
Zipf
Zipf
Zipf
Zipf
Zipf
TABLE I: Summary of the simulation parameters for each
figure.
Our theoretical results in Sections III and IV are stated for
a 2D-Grid topology. The main reason for assuming this rather
unrealistic topology is developing the main idea of the paper
clearly. However, it should be noticed that our approach can
be extended to more general graph models at the expense of
lengthy proofs and calculations. For example, as mentioned
in Section IV, |Br (u)| is the size of the ball of radius r
around node u. The main feature of the 2D-Grid which affects
our results is that |Br (u)| = Θ(r2 ) for all u. Now suppose
that, instead of assuming a 2D-Grid, we consider a graph in
which |Br (u)| = Θ(rdim ), w.h.p. Then the parameter dim will
appear in our results instead of dim = 2 in the special case
of 2D-Grid. For example, the term α + 2β in the statement
of Theorem 4 would be generalized to α + dim × β. More
generally, even for |Br (u)| = Θ(f (r)), our basic technicalities
can be extended too. However, in this paper we investigate
other network topologies, such as Random Geometric Graphs
and Scale-Free (power law) networks via extensive simulations
in Section V. As we have discussed it there, the main trends
are also valid for these topologies as well. Thus, our findings
cover a wider class of graph models which are more similar
to real-world CDN network topologies.
The proposed proximity-aware two choices scheme can be
implemented in a distributed manner. To see why, notice
14
that upon arrival of each request at each server, this strategy
needs two kinds of information to redirect the request. This
information can be provided to the requesting server without
the need for a centralized authority in the following way. The
first one is the cache content of other users in its neighborhood
with radius r. Since, the cache content dynamic of servers is
much slower than the requests arrival, this can be done by
periodic polling of nearby servers without introducing much
overhead (also see Distributed Hash Table (DHT) schemes,
e.g., [37] and [38]). The second type of information is the
queue length information of two randomly chosen nodes inside
its neighborhood with radius r, which can also be efficiently
done in a distributed manner by polling or piggybacking.
In practice, request arrivals and servers’ operation happen
in continuous time which needs a queuing theory based
performance analysis. However, as shown in [7] and [39], the
behaviour of load balancing schemes in continuous time (i.e.,
known as the supermarket model) and static balls and bins
problems are closely related. Thus, we conjecture that our
proposed scheme will also have the same performance in the
queuing theory based model. We postpone a rigorous analysis
of such scenario to future work.
In summary, in this work, we have considered the problem
of randomized load balancing and its tension with communication cost and memory resources in cache networks. By
proposing two request assignment schemes, this trade-off has
been investigated analytically. Moreover, extensive simulation
results support our theoretical findings and provide practical
design guidelines.
ACKNOWLEDGMENT
The authors would like to thank Seyed Abolfazl Motahari,
Omid Etesami, Thomas Sauerwald and Farzad Parvaresh for
helpful discussions and feedback.
R EFERENCES
[1] M. Jafari Siavoshani, A. Pourmiri, and S. P. Shariatpanahi, “Proximityaware balanced allocations in cache networks,” in IEEE International
Parallel & Distributed Processing Symposium (IPDPS), 2017.
[2] Cisco, “Cisco visual networking index: global mobile data traffic forecast update, 20132018,” White Paper, 2014.
[3] G. Zhang, Y. Li, and T. Lin, “Caching in information centric networking:
A survey,” Computer Network, vol. 57, pp. 3128–3141, 2013.
[4] E. Nygren, R. K. Sitaraman, and J. Sun, “The akamai network: a
platform for high-performance internet applications,” Operating Systems
Review, vol. 44, no. 3, pp. 2–19, 2010.
[5] N. Golrezaei, K. Shanmugam, A. G. Dimakis, A. F. Molisch, and
G. Caire, “Femtocaching: Wireless video content delivery through
distributed caching helpers,” in INFOCOM, 2012 Proceedings IEEE,
March 2012, pp. 1107–1115.
[6] Y. Azar, A. Z. Broder, A. R. Karlin, and E. Upfal, “Balanced allocations,” SIAM J. Comput., vol. 29, no. 1, pp. 180–200, 1999.
[7] M. Mitzenmacher, “The power of two choices in randomized load
balancing,” IEEE Trans. Parallel Distrib. Syst., vol. 12, no. 10, pp. 1094–
1104, 2001.
[8] M. Adler, S. Chakrabarti, M. Mitzenmacher, and L. E. Rasmussen, “Parallel randomized load balancing,” Random Struct. Algorithms, vol. 13,
no. 2, pp. 159–188, 1998.
[9] C. Lenzen and R. Wattenhofer, “Tight bounds for parallel randomized
load balancing,” Distributed Computing, vol. 29, no. 2, pp. 127–142,
2016.
[10] P. Berenbrink, A. Czumaj, A. Steger, and B. Vöcking, “Balanced
allocations: The heavily loaded case,” SIAM J. Comput., vol. 35, no. 6,
pp. 1350–1385, 2006.
[11] G. Peng, “CDN: content distribution network,” CoRR, vol.
cs.NI/0411069, 2004. [Online]. Available: http://arxiv.org/abs/cs.NI/
0411069
[12] M. Roussopoulos and M. Baker, “Practical load balancing for
content requests in peer-to-peer networks,” Distributed Computing,
vol. 18, no. 6, pp. 421–434, 2006. [Online]. Available: http:
//dx.doi.org/10.1007/s00446-005-0150-7
[13] M. Leconte, M. Lelarge, and L. Massoulié, “Bipartite graph structures
for efficient balancing of heterogeneous loads,” in ACM SIGMETRICS/PERFORMANCE Joint International Conference on Measurement
and Modeling of Computer Systems, SIGMETRICS ’12, London, United
Kingdom, June 11-15, 2012, 2012, pp. 41–52.
[14] S. Manfredi, F. Oliviero, and S. P. Romano, “A distributed control law for
load balancing in content delivery networks,” IEEE/ACM Trans. Netw.,
vol. 21, no. 1, pp. 55–68, 2013.
[15] F. Xia, A. M. Ahmed, L. T. Yang, and Z. Luo, “Community-based event
dissemination with optimal load balancing,” IEEE Trans. Computers,
vol. 64, no. 7, pp. 1857–1869, 2015.
[16] C. Chen, Y. Ling, M. Pang, W. Chen, S. Cai, Y. Suwa, and O. Altintas,
“Scalable request routing with next-neighbor load sharing in multiserver environments,” in 19th International Conference on Advanced
Information Networking and Applications (AINA 2005), 28-30 March
2005, Taipei, Taiwan, 2005, pp. 441–446.
[17] Y. Xia, A. Dobra, and S. C. Han, “Multiple-choice random network for
server load balancing,” in INFOCOM 2007. 26th IEEE International
Conference on Computer Communications, Joint Conference of the IEEE
Computer and Communications Societies, 6-12 May 2007, Anchorage,
Alaska, USA, 2007, pp. 1982–1990.
[18] A. Meyerson, K. Munagala, and S. Plotkin, “Web caching using access
statistics,” in Proceedings of the Twelfth Annual ACM-SIAM Symposium
on Discrete Algorithms, ser. SODA ’01, 2001, pp. 354–363.
[19] S. Borst, V. Gupta, and A. Walid, “Distributed caching algorithms for
content distribution networks,” in 2010 Proceedings IEEE INFOCOM,
March 2010, pp. 1–9.
[20] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,”
IEEE Trans. Information Theory, vol. 60, no. 5, pp. 2856–2867, 2014.
[21] S. Gitzenis, G. S. Paschos, and L. Tassiulas, “Asymptotic laws for
joint content replication and delivery in wireless networks,” IEEE
Transactions on Information Theory, vol. 59, no. 5, pp. 2760–2776,
May 2013.
[22] V. Shah and G. de Veciana, “High-performance centralized content
delivery infrastructure: Models and asymptotics,” IEEE/ACM Trans.
Netw., vol. 23, no. 5, pp. 1674–1687, 2015.
[23] C. Liu, R. K. Sitaraman, and D. Towsley, “Go-with-the-winner: Performance based client-side server selection,” in 2016 IFIP Networking
Conference, Networking 2016 and Workshops, Vienna, Austria, May 1719, 2016, 2016, pp. 404–412.
[24] A.-M. K. Pathan, C. Vecchiola, and R. Buyya, “Load and proximity
aware request-redirection for dynamic load distribution in peering cdns,”
in On the Move to Meaningful Internet Systems: OTM 2008, OTM 2008
Confederated International Conferences, CoopIS, DOA, GADA, IS, and
ODBASE 2008, Monterrey, Mexico, November 9-14, 2008, Proceedings,
Part I, 2008, pp. 62–81.
[25] J. Tang, W.-P. Tay, and Y. Wen, “Dynamic request redirection and
elastic service scaling in cloud-centric media networks,” IEEE Trans.
Multimedia, vol. 16, no. 5, pp. 1434–1445, 2014.
[26] R. Stanojevic and R. Shorten, “Load balancing vs. distributed rate
limiting: An unifying framework for cloud control,” in Proceedings of
IEEE International Conference on Communications, ICC 2009, Dresden,
Germany, 14-18 June 2009, 2009, pp. 1–6.
[27] P. Berenbrink, A. Brinkmann, T. Friedetzky, and L. Nagel, “Balls into
bins with related random choices,” J. Parallel Distrib. Comput., vol. 72,
no. 2, pp. 246–253, 2012.
[28] K. Kenthapadi and R. Panigrahy, “Balanced allocation on graphs,” in
Proc. 17th Symp. Discrete Algorithms (SODA), 2006, pp. 434–443.
[29] B. Godfrey, “Balls and bins with structure: balanced allocations on
hypergraphs,” in Proc. 19th Symp. Discrete Algorithms (SODA), 2008,
pp. 511–517.
[30] A. Pourmiri, “Balanced allocation on graphs: A random walk approach,”
in Computing and Combinatorics - 22nd International Conference,
COCOON 2016, Ho Chi Minh City, Vietnam, August 2-4, 2016, Proceedings, 2016, pp. 330–341.
[31] L. Breslau, P. Cao, L. Fan, G. Phillips, and S. Shenker, “Web caching
and zipf-like distributions: evidence and implications,” in INFOCOM
’99. Eighteenth Annual Joint Conference of the IEEE Computer and
Communications Societies. Proceedings. IEEE, vol. 1, Mar 1999, pp.
126–134 vol.1.
15
[32] M. Cha, H. Kwak, P. Rodriguez, Y.-Y. Ahn, and S. Moon, “I tube, you
tube, everybody tubes: Analyzing the world’s largest user generated
content video system,” in Proceedings of the 7th ACM SIGCOMM
Conference on Internet Measurement, ser. IMC ’07. New York, NY,
USA: ACM, 2007, pp. 1–14.
[33] L. Devroye, “The expected length of the longest probe sequence for
bucket searching when the distribution is not uniform,” J. Algorithms,
vol. 6, no. 1, pp. 1–9, 1985.
[34] M. Ji, A. M. Tulino, J. Llorca, and G. Caire, “Order-optimal rate of
caching and coded multicasting with random demands,” CoRR, vol.
abs/1502.03124, 2015. [Online]. Available: http://arxiv.org/abs/1502.
03124
[35] [Online]. Available: https://github.com/mahdijafaris/Balls and Bins
[36] R. Albert and A.-L. Barabási, “Statistical mechanics of complex networks,” Reviews of modern physics, vol. 74, no. 1, p. 47, 2002.
[37] D. Bauer, P. Hurley, and M. Waldvogel, “Replica placement and location
using distributed hash tables,” in 32nd Annual IEEE Conference on Local
Computer Networks (LCN 2007), 15-18 October 2007, Clontarf Castle,
Dublin, Ireland, Proceedings, 2007, pp. 315–324.
[38] D. R. Karger, E. Lehman, F. T. Leighton, R. Panigrahy, M. S. Levine,
and D. Lewin, “Consistent hashing and random trees: Distributed
caching protocols for relieving hot spots on the world wide web,” in
Proceedings of the Twenty-Ninth Annual ACM Symposium on the Theory
of Computing, El Paso, Texas, USA, May 4-6, 1997, 1997, pp. 654–663.
[39] M. Mitzenmacher, A. W. Richa, and R. Sitaraman, “The power of two
random choices: A survey of technique and results,” In Handbook of
Randomized Computation Volume 1, pp. 255–312, 2001.
[40] D. P. Dubhashi and A. Panconesi, Concentration of Measure for the
Analysis of Randomized Algorithms. Cambridge University Press, 2009.
[41] A. Auger and B. Doerr, Theory of Randomized Search Heuristics:
Foundations and Recent Developments. River Edge, NJ, USA: World
Scientific Publishing Co., Inc., 2011.
A PPENDIX A
S OME TAIL B OUNDS
Theorem
7
(Chernoff
Bounds). Suppose
that
X1 , X2 , . . . , X
∈
{0,
1}
are
independent
random
variables
Pnn
and let X = i=1 Xi . Then for every δ ∈ (0, 1) the following
inequalities hold:
Pr [X ≥ (1 + δ) E [X]] ≤ exp(−δ 2 E [X] /2),
Pr [X ≤ (1 − δ) E [X]] ≤ exp(−δ 2 E [X] /3).
In particular,
Pr [|X − E [X] | ≥ δ E [X]] ≤ 2 exp(−δ 2 E [X] /3).
For a proof see [40].
To deal with moderate independency we can state the
following lemma.
Lemma 4 (Deviation bounds for moderate independency,
see [41, Lemma 1.18]). Let X1 , . . . , Xn be arbitrary binary
random variables. Let X1∗ , X2∗ , . . . , Xn∗ be binary random
variables that are mutually independent and such that for all
i, Xi is independent of X1 , . . . , Xi−1 . Assume that for all i
and all x1 , . . . , xi−1 ∈ {0, 1},
Pr [Xi = 1|X1 = x1 , . . . , Xi−1 = xi−1 ] ≥ Pr [Xi∗ = 1] .
Then for all k ≥ 0, we have
" n
#
" n
#
X
X
∗
Pr
Xi ≤ k ≤ Pr
Xi ≤ k
i=1
i=1
and the latter term can be bounded by any deviation bound
for independent random variables.
| 8cs.DS
|
Online Auctions and Multi-scale Online Learning
∗
Sébastien Bubeck† , Nikhil R. Devanur† , Zhiyi Huang‡ and Rad Niazadeh§
arXiv:1705.09700v1 [cs.GT] 26 May 2017
†
‡
Microsoft Research, Redmond,
Department of Computer Science, The University of Hong Kong,
§
Department of Computer Science, Cornell University.
May 30, 2017
Abstract
We consider revenue maximization in online auctions and pricing. A seller sells an identical
item in each period to a new buyer, or a new set of buyers. For the online posted pricing problem,
we show regret bounds that scale with the best fixed price, rather than the range of the values.
We also show regret bounds that are almost scale free, and match the offline sample complexity,
when comparing to a benchmark that requires a lower bound on the market share. These results
are obtained by generalizing the classical learning from experts and multi-armed bandit problems
to their multi-scale versions. In this version, the reward of each action is in a different range, and
the regret w.r.t. a given action scales with its own range, rather than the maximum range.
1
Introduction
Consider the following revenue maximization problem in a repeated setting, called the online posted
pricing problem. In each period, the seller has a single item to sell, and a new prospective buyer. The
seller offers to sell the item to the buyer at a given price; the buyer buys the item if and only if the price
is below his private valuation for the item. The private valuation of the buyer itself is never revealed
to the seller. How should a monopolistic seller iteratively set the prices if he wishes to maximize his
revenue? What if he also cares about market share?
Estimating price sensitivities and demand models in order to optimize revenue and market share
is the bedrock of econometrics. The emergence of online marketplaces has enabled sellers to costlessly
change prices, as well as collect huge amounts of data. This has renewed the interest in understanding
best practices for data driven pricing. The extreme case of this when the price is updated for each buyer
is the online pricing problem described above; one can always use this for less frequent price updates.
Moreover this problem is intimately related to classical experimentation and estimation procedures.
This problem has been studied from an online learning perspective, as a variant of the multiarmed bandit problem. The revenue of a pricing algorithm is compared to the revenue of the best
fixed posted price, in hindsight, and the difference between the two, called the regret, is analyzed. No
assumption is made on the distribution of values; the regret bounds are required to hold for the worst
case sequence of values. Blum et al. (2004) assume that the buyer valuations are in [1, h], and show
∗
Zhiyi Huang is supported by a RGC grant HKU17202115E. Rad Niazadeh is supported by a NSF grant CCF-1512964
and by a Google Ph.D. Fellowship. Part of this work were completed while Zhiyi was visiting and Rad was doing an
internship at Microsoft Research Redmond.
1
the following multiplicative + additive bound on the regret: for any ǫ ∈ (0, 1), the regret is at most
ǫ times the revenue of the optimal price, + O(ǫ−2 h log h log log h). Blum and Hartline (2005) show
that the additive factor can be made to be O(ǫ−3 h log log h), trading off a log h factor for an extra ǫ−1
factor.
An undesirable aspect of these bounds is that they scale linearly with h; this is particularly problematic when h is an estimate and we might set it to be a generous upper bound on the range of prices
we wish to consider. A typical use case is when the same algorithm is used for many different products,
with widely varying price ranges. We may not be able to manually tune the range for each product
separately.
This dependency on h seems unavoidable, as is reflected by the lower bounds for the problem.
(Lower bounds are discussed later in the introduction.) Yet, somewhat surprisingly, our first contribution in this paper is to show that we can replace h by the best fixed price1 (that is used
in the definition of the benchmark). In particular, we show that the additive bound can be made to
be O(ǫ−2 p∗ log h), where p∗ is the best fixed price in hindsight. This allows us to use a very generous
estimate for h; we only lose a log h factor. The algorithm balances exploration probabilities of different
prices carefully and automatically zooms in on the relevant price range. This does not violate known
lower bounds, since in those instances p∗ is close to h.
Bar-Yossef et al. (2002), Blum et al. (2004), and Blum and Hartline (2005) also consider the “full
information” version of the problem, or what we call the online auction problem, where the valuations
of the buyers are revealed to the algorithm after the buyer has made a decision. Such information may
be available in a context where the buyers have to bid for the items, and are awarded the item if their
bid is above a hidden price. In this case, the additive term can be improved to O(ǫ−1 h log(ǫ−1 )), which
is tight. Once again, we show that h can be replaced with p∗ ; in particular, we show that the additive
term can be made to be O(ǫ−1 p∗ log(hǫ−1 )).
1.1
Purely multiplicative bounds and sample complexity
The regret bounds mentioned above can be turned into a purely multiplicative factor in the following
way: for any ǫ > 0, the algorithm is guaranteed to get a 1 − O(ǫ) fraction of the best fixed price
revenue, provided the number of periods T ≥ E/ǫ, where E is the additive term in the regret bounds
above. This follows from the observation that a revenue of T is a lower bound on the best fixed price
revenue. Call the number of periods required to get a 1 − ǫ multiplicative approximation (as a function
of ǫ) as the convergence rate of the algorithm.
A 1 − ǫ multiplicative factor is also the target in the recent line of work on the sample complexity
of auctions started by Dhangwatnotai et al. (2014); Cole and Roughgarden (2014). (We give a more
comprehensive discussion of this line of work in Section 1.3.) Here, i.i.d. samples of the valuations are
given from a fixed but unknown distribution, and the goal is to find a price such that its revenue w.r.t.
the hidden distribution is a 1 − ǫ fraction of the optimum revenue for this distribution. The sample
complexity is the minimum number of samples needed to guarantee this (as a function of ǫ).
The sample complexity and the convergence rate (for the full information setting) are closely related
to each other. The sample complexity is always smaller than the convergence rate: the problem is easier
because of the following.
1. The valuations are i.i.d. in case of sample complexity whereas they can be arbitrary (worst case)
in case of convergence rate.
1
Standard bounds allow regret to depend on the loss of the best action instead of the worst case loss. However, even
such bounds still depend linearly on the range of the losses, and thus they would not allow to replace h by the best fixed
price.
2
2. Sample complexity corresponds to an offline problem: you get all the samples at once. Convergence rate corresponds to an online problem: you need to decide what to do on a given valuation
without knowing what valuations arrive in the future.
This is formalized in terms of an online to offline reduction [folklore] which shows that a convergence
rate upper bound can be automatically translated to a sample complexity upper bound. This lets
us convert sample complexity lower bounds into lower bounds on the convergence rate, and in turn
into lower bounds on the additive error E in an additive + multiplicative regret bound. E.g., the
additive error for the online auction problem (and hence also for the posted pricing problem2 ) cannot
be o(hǫ−1 ) (Huang et al., 2015). Moreover, it is insightful to compare convergence rates we show with
the best known sample complexity upper bound; proving better convergence rates would mean improving
these bounds as well.
A natural target convergence rate for a problem is therefore the corresponding sample complexity,
but achieving this is not always trivial. An interesting version of the sample complexity bound for
auctions did not have an analogous convergence rate bound. This version takes into account both
revenue and market share, and surprisingly, gets sample complexity bounds that are scale free; there
is no dependence on h, which means it works for unbounded valuations! For any δ ∈ (0, 1), the best
fixed price benchmark is relaxed to ignore those prices whose market share (or equivalently probability
of sale) is below a δ fraction; as δ increases the benchmark is lower. This is a meaningful benchmark
since in many cases revenue is not the only goal, even if you are a monopolist. A more reasonable goal
is to maximize revenue subject to the constraint that the market share is above a certain threshold.
What more, this gives a sample complexity of O(ǫ−2 δ−1 log(δ−1 ǫ−1 )) (Huang et al., 2015). In fact δ
can be set to h−1 without loss of generality, when the values are in [1, h],3 and the above bound then
matches the sample complexity w.r.t. the best fixed price revenue. In addition, this bound gives a
precise interpolation: as the target market share δ increase, the number of samples needed decreases
almost linearly.
The second contribution of this paper is to show a convergence rate that almost
matches the above sample complexity, for the full information setting. We have a mild
dependence on h; the rate is proportional to log log h. Further, we also show a near optimal convergence
rate for the posted pricing problem.4
Multiple buyers: All of our results in the full information (online auction) setting extend to the
multiple buyer model. In this model, in each time period, a new set of n buyers compete for a single
item. The seller runs a truthful auction that determines the winning buyer and his payment. The
benchmark here is the set of all “Myerson-type” mechanisms. These are mechanisms that are optimal
when each period has n buyers of potentially different types, and the value of each buyer is drawn
independently from a type dependent distribution. In fact, our convergence rates also imply new
sample complexity bounds for these problems (except that they are not computationally efficient).
The various bounds and comparison to previous work are summarized in Tables 1 and 2.
2
We conjecture that the lower bound for the posted pricing problem should be worse by a factor of ǫ−1 , since one
needs to explore about ǫ−1 differnet prices.
3
When the values are in [1, h], we can guarantee a revenue of T by posting a price of 1, and to beat this, any other
price (and in particular a price of h) would have to sell at least T /h times.
4
Unfortunately, we cannot yet guarantee that our online algorithm itself gets a market share of δ, although we
strongly believe that it does. Showing such bounds on the market share of the algorithm is an important avenue for
future research.
3
Lower bound
Online single buyer auction
Ω
h
ǫ2
∗
Ω max{ ǫh2 , ǫ13 }
Online posted pricing
Ω( ǫh2 ) ∗
Online multi buyer auction
Best known
(Sample complexity)
†
Õ ǫh2
∗§
-
O( nh
ǫ3 )
†
Huang et al. (2015);
Blum et al. (2004);
§
(2017);
Kleinberg and Leighton (2003).
‡
∗
‡
Upper bound
Best known
(Convergence rate)
Õ ǫh2 †
Õ ǫh3 †
-
This paper
(Thm. 2.5)
∗
Õ pǫ2
∗
Õ pǫ3
Õ nh
ǫ3
Devanur et al. (2016); Gonczarowski and Nisan
Table 1: Number of rounds/samples needed to get a 1 − ǫ approximation to the best offline
price/mechanism. Sample complexity is for the offline case with i.i.d. samples from an unknown
distribution. Convergence rate is for the online case with a worst case sequence. Sample complexity
is always no larger than the convergence rate. Lower bounds hold for sample complexity too, except
for the online posted pricing problem for which there is no sample complexity version. The additive +
multiplicative regret bounds are converted to convergence rates by dividing the additive error by ǫ. In
the last row, n is the number of buyers. In the last column, p∗ denotes the optimal price.
Lower bound
(Sample complexity)
Online single buyer auction
Online posted pricing
Online multi buyer auction
∗
Ω
1
ǫ2 δ
∗
Ω max{ ǫ21δ , ǫ13 }
Ω ǫ21δ ∗
Huang et al. (2015);
†
∗†
Upper bound
This paper
Best known
(Sample complexity) (Thm. 2.6)
Õ ǫ21δ ∗
Õ ǫ21δ
Õ ǫ41δ
Õ ǫn3 δ
Kleinberg and Leighton (2003).
Table 2: Sample complexity & convergence rate w.r.t. the opt mechanism/price with market share ≥ δ.
1.2
Multi-scale online learning
The main technical ingredients in our results are variants of the classical problems of learning from
expert advice and multi-armed bandit. We introduce the multi-scale versions of these problems, where
each action has its reward bounded in a different range. Our third contribution is to give an
algorithm for this problem whose regret w.r.t. a certain action scales with the range of
rewards for that particular action. To contrast, the regret bounds in the standard versions scale
with the maximum range. We expect such bounds to be of independent interest.
The multi-scale versions of these problems exhibit subtle variations that don’t appear in the standard versions. First of all, our applications to auctions and pricing has non-negative rewards, and this
actually makes a difference. For both the expert and the bandit versions, the minimax regret bounds
for non-negative rewards are provably better than those when rewards could be negative. Further, for
the bandit version, we can prove a better bound if we only require the bound to hold w.r.t. the best
action, rather than all actions (for non-negative rewards). The various regret bounds and comparison
to standard bounds are summarized in Tables 3.
We use algorithms based on online (stochastic) mirror descent (OSMD) (Bubeck, 2011), with a
weighted negative entropy as the Legendre function. This framework gives regret bounds in terms of a
“local norm” as well as an “initial divergence”, which we then bound differently for each version of the
problem. In the technical sections we highlight how the subtle variations arise as a result of different
4
Multi-scale bound (this paper)
Standard
regret bound O(·)
p
cmax T log(k)
Experts/non-negative
Bandits/non-negative
cmax
Experts/symmetric
Bandits/symmetric
∗
p
T k log(k)
p
cmax T log(k)
p
cmax T k log(k)
Upper bound O(·)
p
ci T log(kT )
∗
2
ci∗
∗
†
1
ci T 3 (k log(kT )) 3
†
p
T k log(k), i∗ is the best action
q
)
ci T log(k · ccmax
min
q
log(kT · ccmax
)
ci T k · ccmax
min
min
Freund and Schapire (1995);
†
Auer et al. (1995).
Lower bound Ω(·)
p
ci T log(k)
√
ci T K
p
ci T log(k)
q
ci T k · ccmax
min
Table 3: Pure-additive regret bounds for non-negative rewards, i.e. when reward of any action i at
any time is in [0, ci ], and symmetric range rewards, i.e. when reward of any action i at any time is in
[−ci , ci ] (suppose T is the time horizon, A is the actions set, and k is the number of actions).
techniques used to bound these two terms.
Foster et al. (2017) also consider the multi-scale online learning problem motivated by a model
selection problem. They consider additive bounds, for the symmetric case, for full information, but
not bandit feedback. Their regret bounds are not comparable to ours in general; our bounds are better
for the pricing/auction applications we consider, and their bounds are better for their application.
1.3
Other related work
The online pricing problem, also called dynamic pricing, is a much studied topic, across disciplines
such as operations research and management science (Talluri and Van Ryzin, 2006), economics (Segal,
2003), marketing, and of course computer science. The multi-armed bandit approach to pricing is
particularly popular. See den Boer (2015) for a recent survey on various approaches to the problem.
Kleinberg and Leighton (2003) consider the online pricing problem, under the assumption that the
values are in [0, 1], and considered purely additive factors. They showed that the minimax additive
regret is Θ̃(T 2/3 ), where T is the number of periods. This is similar in spirit to regret bounds that
scale with h, since one has to normalize the values so that they are in [0, 1]. The finer distinction about
the magnitude of the best fixed price is absent in this work. Recently, Syrgkanis (2017) also consider
the online auction problem, with an emphasis on a notion of “oracle based” computational efficiency.
They assume the values are all in [0, 1] and don’t consider the scaling issue that we do; this makes
their contribution orthogonal to ours.
Starting with Dhangwatnotai et al. (2014), there has been a spate of recent results analyzing the
sample complexity of pricing and auction problems. Cole and Roughgarden (2014) and Devanur et al.
(2016) consider multiple buyer auctions with regular distributions (with unbounded valuations) and
give sample complexity bounds that are polynomial in n and ǫ−1 , where n is the number of buyers.
Morgenstern and Roughgarden (2015) consider arbitrary distributions with values bounded by h, and
gave bounds that are polynomial in n, h, and ǫ−1 . Roughgarden and Schrijvers (2016); Huang et al.
(2015) give further improvements on the single- and multi-buyer versions respectively; tables 1 and 2
give a comparison of these results with our bounds, for the problems we consider. The dynamic pricing
problem has also been studied when there are a given number of copies of the item to sell (limited supply) (Agrawal and Devanur, 2014; Babaioff et al., 2015; Badanidiyuru et al., 2013; Besbes and Zeevi,
2009). There are also variants where the seller interacts with the same buyer repeatedly, and the buyer
can strategize to influence his utility in the future periods (Amin et al., 2013; Devanur et al., 2014).
5
2
Model and Main Results
We consider a variety of online algorithmic problems that are all parts of the multiscale online learning
framework. We start by defining this framework and expressing our results in terms of action-specific
regret bounds for this general problem. Next, we investigate different auction design problems that are
covered by this framework, and show how to get multiplicative cum additive approximations for these
problems by the help of the multi-scale learning framework. We then consider competing with δ-guarded
benchmarks and show how our algorithms get pure multiplicative approximations with respect to these
benchmarks. We can then translate the convergence rate of our online algorithms to sample complexity
of auctions to 1) generalize many sample complexity upper-bounds to the online adversarial auction
settings, 2) compare our bounds with the known sample complexity lower-bounds, and 3) design new
algorithms achieving near-optimal sample complexity bounds for the offline Bayesian auction problem.
2.1
Multi-scale online learning framework
Our multi-scale online learning framework is basically the classical learning from expert advice problem
(under full-information) or multi-armed bandit problem (under partial-information). The main difference is that the range of different experts/arms could be different. Suppose there is a set of actions A.
The problem proceeds in T rounds, and in each round t ∈ [T ] :5
• The algorithm picks an action it ∈ A
• The adversary picks a reward function g(t) simultaneously, where action i has reward gi (t).
• The algorithm gets the reward git (t).
• In the full information setting, the algorithm sees the entire reward function g(t). In the bandit
setting, the algorithm sees only its own reward, git (t).
The total reward of the algorithm is denoted by
P
Galg := Tt=1 git (t).
The standard “best fixed action” benchmark is
Gmax := maxi∈A
PT
t=1 gi (t).
We consider both full-information and the bandit setting:
• Multi-scale experts: The action set is countable. If the action set is finite of size k, we identify
A = [k]. The reward g(t) is such that for all i ∈ A, gi (t) ∈ [0, ci ]. The entire reward function
g(t) is revealed to the algorithm after round t.
• Multi-scale bandit learning: The same as before, in the bandit setting.
We prove action-specific regret bounds, which we call also multi-scale regret guarantees. Towards this
end, we define the following quantities.
P
Gi :=
(1)
t∈[T ] gi (t) ,
regreti := Gi − Galg .
5
We use the notation [n] := {1, 2, . . . , n}, for any n ∈ N.
6
(2)
The regret bound w.r.t. action i, i.e., an upper bound on E [regreti ], depends on the range ci , as well
as any prior distribution π over the action set A; this way, we can handle countably many actions.
Let cmin = inf i∈A ci and cmax = supi∈A ci (if applicable) be the minimum and the maximum range.
We first
√ state a version of the regret bound which is parameterized by ǫ > 0; such bounds are stronger
than T type bounds which are more standard.
Theorem 2.1. There exists an algorithm for the multi-scale experts problem that takes as input any
distribution π over A, the ranges ci , ∀ i ∈ A, and a parameter 0 < ǫ ≤ 1, and satisfies:
1
1
(3)
log
· ci
∀i ∈ A : E [regreti ] ≤ ǫ · Gi + O
ǫ
ǫπi
Compare this to what you get by using the standard analysis for the
experts problem (Arora et al.,
2012), where the second term in the regret bound is O 1ǫ log(k) · cmax . Choosing π to be the uniform
distribution in the above theorem gives O 1ǫ log kǫ · ci . Also, one can compare the pure-additive
p
version of this bound with the classic pure-additive regret bound O cmax · T log(k) for the experts
q
)
problem by setting ǫ = log(kT
(Corollary 2.2).
T
Corollary 2.2. There exists an algorithm for the multi-scale experts problem that takes as input the
ranges ci , ∀ i ∈ A, and satisfies:
p
(4)
∀i ∈ A : E [regreti ] ≤ O ci · T log(kT )
For the bandit version, we can get a similar regret guarantee, but only for the best action. If we
require the regret bound to hold for all actions, then we can only get a weaker bound, where the second
term has ǫ−2 instead of ǫ−1 . The difference between the bounds for the bandit and the full information
setting is essentially a factor of k, which is unavoidable.
Theorem 2.3. There exists an algorithm for the online multi-scale bandits problem that takes as input
the ranges ci , ∀ i ∈ A, and a parameter 0 < ǫ ≤ 1, and satisfies,
• for i∗ = arg maxi∈A Gi ,
• for all i ∈ A,
E [regreti∗ ] ≤ ǫ · Gi∗ + O
1
ǫ k log
k
ǫ
E [regreti ] ≤ ǫ · Gi + O
1
ǫ2 k log
k
ǫ
· ci∗ .
· ci .
(5)
(6)
q Also, one can compute the pure-additive versions of the bounds in Theorems 2.3 by setting ǫ =
k log(kT )
) 1
and ǫ = ( k log(kT
) 3 resepctively (Corollary 2.4), and compare with the pure-additive regret
T
T
√
bound O cmax · T k log k for the adversarial multi-armed bandit problem (Auer et al., 1995).
Corollary 2.4. There exist algorithms for the online multi-scale bandits problem that satisfies,
• For i∗ = arg maxi∈A Gi ,
• For all i ∈ A,
p
E [regreti∗ ] ≤ O ci∗ · T k log(kT )
1
2
E [regreti ] ≤ O ci · T 3 (k log(kT )) 3
7
(7)
(8)
2.2
Online auction design
The auction design problems that we consider are as follows.
• Online single buyer auction: The action set A = [1, h]. The reward function is such that the
adversary picks a value v(t) ∈ [1, h] and for any price i ∈ A, the reward gi (t) := p · 1(v(t) ≥ i).
This is the full information setting, where the value v(t) is revealed to the algorithm after round
t.
• Online posted pricing: The same as above, in the bandit setting. The algorithm only learns
the indicator function 1(v(t) ≥ it ) where it is the price it picks in round t.
• Online multi buyer auction: The action set is the set of all “Myerson-type” mechanisms for n
buyers, for some n ∈ N. (See Definition 5.1.) The adversary picks a valuation vector v(t) ∈ [1, h]n
and the reward of a mechanism M is its revenue when the valuation of the buyers is given by
v(t); this is denoted by revM (v(t)). The algorithm sees the full vector of valuations v(t).
We show how to get a multiplicative cum additive approximations for these problems with Gmax as
the benchmark, à la Blum et al. (2004); Blum and Hartline (2005). The main improvement over these
results is that the additive term scales with the best price rather than h. Let p∗ be the best fixed price
on hindsight, which is the price that achieves Gmax .
Theorem 2.5. There are algorithms for the online single buyer auction, online posted price auction,
and the online multi buyer auction problems that take as input a parameter ǫ > 0, and satsify Galg ≥
(1 − ǫ)Gmax − O(E), where respectively (for the three problems mentioned above)
p∗ log h log(log h/ǫ)
hn log h log(n log h/ǫ)
p∗ log(log h/ǫ)
,
,
and
.
ǫ
ǫ2
ǫ2
Even if h is not known upfront, we can still get the similar approximation guarantee for online single
buyer auction and online multi buyer auction with:
p∗ log(p∗ /ǫ)
hn log h log(n log h/ǫ)
E=
,
and
.
ǫ
ǫ2
Bounds on the sample complexity of auctions imply that the first bound in this theorem is tight up
to log factors: the lower bound is hǫ−1 in an instance where p∗ = h, and the best upper bound known
is hǫ−1 log(1/ǫ). We conjecture that our bound for the online posted pricing problem is tight up to log
factors, and leave resolving this as an open problem. The third bound is not comparable to the best
sample complexity for the multi buyer auction problem by Roughgarden and Schrijvers (2016); it is
better than theirs for large ǫ (when 1/ǫ ≤ o(nh)), and is worse for smaller ǫ (when 1/ǫ ≥ ω(nh)). Also,
compare these to the corresponding upper bounds for the first two problems by Blum et al. (2004);
Blum and Hartline (2005), which are respectively
h log h log log h h log log h
h log(1/ǫ)
, and min
,
.
ǫ
ǫ2
ǫ3
E=
2.3
Competing with δ-guarded benchmarks
For the single buyer auction/pricing problem, we define a δ-guarded benchmark, for any δ ∈ [0, 1]. This
benchmark is restricted to only those prices that sell the item in at least a δ fraction of the rounds.
nP
o
PT
T
Gmax (δ) := max
g
(t)
:
p
∈
A,
1(v
≥
p)
≥
δT
.
t
t=1 p
t=1
As observed in Footnote 3, one can replace δ with 1/h and get the corresponding guarantees for Gmax
rather than Gmax (δ). However, the main point of these results is to show a graceful improvement of
the bounds as δ is chosen to be larger.
8
Multiple buyers: For the multi buyer auction problem, we define the δ-guarded benchmark as
follows. For any sequence of value vectors v(1), v(2), . . . , v(T ), let V̄ denote the largest value such
that there are at least δT distinct v(t)’s with maxi∈[n] vi (t) ≥ V̄ . Define the δ-guarded benchmark to
be
P
Gmax (δ) = maxM Tt=1 RevM min(V̄ ~1, v(t))) ,
where the min is to be understood to be applied co-ordinate wise, and the max is over all Myerson-type
mechanisms.
We focus on purely multiplicative approximation factors when competing with Gmax (δ). In particular, for any given ǫ > 0, we are interested in a 1 − ǫ approximation. We state our results in terms of
the convergence rate. We say that T (ǫ, δ) is the convergence rate of an algorithm if for all time horizon
T ≥ T (ǫ, δ), we are guaranteed that Galg ≥ (1 − ǫ)Gmax (δ). Our main results are as follows.
Theorem 2.6. There are algorithms for the online single buyer auction, online posted pricing, and the
online multi buyer auction problems with convergence rates respectively of
log h
n log (1/ǫδ) log(n log(1/ǫδ)/ǫ) log (log h/ǫ)
log(log h/ǫ)
+
,
O
,
and O
.
O
ǫ2 δ
ǫ4 δ
ǫ3 δ
ǫ2 δ
Even if h is not known upfront, we can still get the following similar convergence rates for online single
buyer auction and online multi buyer auction respectively:
log(p∗ /ǫ)
n log (1/ǫδ) log(n log(1/ǫδ)/ǫ) log (h/ǫ)
+
O
,
and O
.
ǫ2 δ
ǫ3 δ
ǫ2 δ
Once again, we compare to the sample compexity bounds: our first is within a log log h factor
of the best sample complexity upper bound in Huang et al. (2015). The lower bound for the online
single buyer auction is Ω(δ−1 ǫ−2 ), which is also the best lower bound known for the pricing and the
multi-buyer problem.6 For the online posted pricing problem, we conjecture that the right dependence
on ǫ should be ǫ−3 . No sample complexity bounds for the multi-buyer problem were known before; in
fact we introduce the definition of a δ-guarded benchmark for this problem.
2.4
Multi-scale online learning with symmetric range
The standard analysis for the experts and the bandit problems holds even if the range of gi (t) is
[−ci , ci ], rather than [0, ci ] as we have assumed. In contrast, there are subtle differences on the best
acheivable multi-scale regret bounds between the non-negative and the symmetric range. We first show
the following upper bound for the full information setting when the range is symmetric. This bounds
follows the same style of action-specific regret bounds as in Theorem 2.1. More detailed discussion on
how the choice of initial distribution π affects the bound is deferred to the appendix, Section A.1.
Theorem 2.7. There exists an algorithm for the multi-scale experts problem with symmetric range
that takes as input any distribution π over A, the ranges ci , ∀ i ∈ A, and a parameter 0 < ǫ ≤ 1, and
satisfies:
i
hP
ci
1
1
∀i ∈ A : E [regreti ] ≤ ǫ · E
(9)
t∈[T ] gt (i) + O ǫ log πi · cmin · ci .
2.1, we can compute the pure-additive version of the bound in Theorem 2.7 by
Similar to
q Section
log(k· ccmax )
min
setting ǫ =
, as in Corollary 2.2.
T
6
Cole and Roughgarden (2014) show that at least a linear dependence on n is necessary when the values are drawn
from a regular distribution, but as is, their lower bound needs unbounded valuations. The lower bound probably holds
for “large enough h” but it is not clear if it holds for all h.
9
Corollary 2.8. There exists an algorithm for the online multi-scale experts problem with symmetric
range that takes as input the ranges ci , ∀ i ∈ A, and satisfies:
q
∀i ∈ A : E [regreti ] ≤ O ci · T log(k · ccmax
)
(10)
min
√
If we compare the above regret bound with the standard O(cmax T log k) regret boundqfor the experts problem, we see that we replace the dependency on cmax in the standard bound with ci log( ccmax
).
min
It is natural
to ask whether we could get rid of the dependence on log(ci /cmin ) and show regret bound
√
of O(ci T log k), like we did for non-negative rewards. However, the next theorem shows that this
dependence on log(ci /cmin ) in the above bound is necessary, in a weak sense: where the constant in
the O(·) is universal and does not depend on the ranges ci . This is because the lower bound only holds
for “small” values of the horizon T , which nonetheless grows with the {ci }s.7
Theorem 2.9. There exists an action set of size k, and ranges ci , ∀i ∈ [k], and time horizon T ,
such that for all algorithms for the online multi-scale experts problem with symmetric range, there is a
sequence of T gain vectors such that
q
∃i ∈ A : E [regreti ] > c4i · T log(k · ccmax
)
min
We then show the following upper bound for the bandit setting when the range is symmetric. This
bound also follows the same style of action-specific regret bounds as in Theorem 2.3.
Theorem 2.10. There exists an algorithm for the multi-scale bandits problem with symmetric range
that takes as input the ranges ci , ∀ i ∈ A, and a parameter 0 < ǫ ≤ 1/2, and satisfies:
k cmax
∀i ∈ A : E [regreti ] ≤ O ǫT + kǫ ccmax
log
· ci .
(11)
ǫ cmin
min
Also, similar
qto Section 2.1, we can compute the pure-additive version of the bound in Theorem 2.10
k cmax log(kT · cmax )
by setting ǫ = √cmin T cmin , as in Corollary 2.2. This bound is comparable to the standard regret
bound of O(cmax kT log k) (Auer et al., 1995) for the adversarial multi-armed bandits problem.
Corollary 2.11. There exists an algorithm for the online multi-scale bandits problem with symmetric
range that satisfies:
q
cmax
∀i ∈ A : E [regreti ] ≤ O ci · T k · ccmax
log(kT
·
)
.
(12)
cmin
min
Once again, for the bandit problem, the following theorem shows that this bound cannot be improved beyond log factors (to get a guarantee like that of Theorem 2.3, for instance).
Theorem 2.12. There exists an action set of size k, and ranges ci , ∀i ∈ [k], such that for all algorithms
for the online multi-scale bandit problem with symmetric range, for all sufficiently large time horizon
T , there is a sequence of T gain vectors such that
r
cmax
ci
∃i ∈ A : E [regreti ] > √ · T k ·
.
cmin
8 2
7
For this reason we chose not to include this bound in Table 3.
10
Organization We start in Section 3 by showing regret upper bounds for the multi-scale experts
problem with non-negative rewards (Theorem 2.1). The corresponding upper bounds for the bandit
version are in section 4 (Theorem 2.3). In Section 5 we show how the multi-scale regret bounds
(Theorems 2.1 and 2.3) imply the corresponding bounds for the auction/pricing problems (Theorems
2.5 and 2.6). Finally, the regret (upper and lower) bounds for the symmetric range are discussed in
Section 6 (Theorems 2.7, 2.9, 2.10, and 2.12).
3
Multi-Scale Online Learning with Full Information
In this section, we look at the full information multi-scale learning problem, in which different experts
have different ranges. We exploit this structure to achieve expert-specific regret bounds.
Here is a map of this section. In Section 3.1 we propose an algorithm that exploits the aforementioned structure, and later in Section 3.2 we show how this algorithm is an online mirror descent with
weighted negative entropy as the Legendre function. For reward-only instances, we prove the regret
bound without dependency on log(ci /cmax ) in Section 3.3.
3.1
Multi-Scale Multiplicative-Weight (MSMW) algorithm
We propose the “Multi-Scale Multiplicative-Weight ” (MSMW) algorithm as a multiplicative-weight
update style learning for our problem. The algorithm is presented in Algorithm 1. The main idea
behind this algorithm is taking into account different ranges for different experts, and therefore
1. Normalizing reward of each expert accordingly, i.e. dividing the reward of expert i by ci .
2. Projecting the updated weights accordingly, by performing a smooth multi-scale projection into
the simplex that will be described later.
Algorithm 1 MSMW
1: input initial distribution µ over A, learning rate 0 < η ≤ 1.
2: initialize p(1) such that pi (1) = µi for all i ∈ A.
3: for t = 1, . . . , T do
4:
Randomly pick an action drawn from p(t), and observe g(t).
).
5:
∀i ∈ A : wi (t + 1) ← pi (t) · exp(η · gic(t)
P i
∗
∗
6:
Find λ (e.g., binary search) s.t. i∈A wi (t + 1) · exp(− λci ) = 1.
∗
7:
∀i ∈ A : pi (t + 1) ← wi (t + 1) · exp(− λci ).
8: end for
3.2
Equivalence to Online Mirror Descent (OMD) with weighted negative entropy
While it is possible to analyze the regret of MSMW algorithm (Algorithm 1) by using first principles
(Look at the proof of Lemma 3.4 in the appendix, Section A.4), we take a different approach . We show
how this algorithm is indeed an instance of Online Mirror Descent (OMD) algorithm for a particular
choice of Legendre function.
11
3.2.1
Preliminaries on online mirror descent.
A
Fix an open convex set D and its closure D̄, which in our case are RA
>0 and R+ respectively, and a closedconvex action set A ⊂ D̄, which in our case is ∆A , i.e. the set of all probability distributions over experts
in A. At the heart of an OMD algorithm there is a Legendre function F : D̄ → R, i.e. a strictly convex
function that admits continuous first order partial derivatives on D and limx→D̄\D k∇F (x)k = +∞,
where ∇F (.) denotes the gradient map of F . One can think of OMD as a member of projected gradient
descent algorithms, where the gradient update happens in the dual space ∇F (D) rather than in primal
D, and the projection is defined by using the Bregman divergence associated with F rather than ℓ2 distance.
Definition 3.1 (Bregman Divergence (Bubeck, 2011)). Given a Legendre function F over ∆A , the
Bregman divergence associated with F , denoted as DF : ∆A × ∆A → R, is defined by
DF (x, y) = F (x) − F (y) − (x − y)T ∇F (y)
Definition 3.2 (Online Mirror Descent Bubeck (2011)). Suppose F is a Legendre function. At every
time t ∈ [T ], the online mirror descent algorithm with Legendre function F selects an expert drawn
from distribution p(t), and then updates w(t) and p(t) given rewards g(t) by:
Gradient update:
∇F (w(t + 1)) = ∇F (p(t)) + η · g(t) ⇒ w(t + 1) = (∇F )−1 (∇F (p(t)) + η · g(t))
(13)
p(t + 1) = argmin (DF (p, w(t + 1)))
(14)
Bregman projection:
p∈∆A
where η > 0 is called the learning rate of OMD.
We use the following standard regret bound of OMD (Refer to Bubeck (2011) for a thorough
discussion on OMD. For completeness, a proof is also provided in the appendix, Section A.5).
Lemma 3.1. For any learning rate parameter 0 < η ≤ 1 and any benchmark distribution q over A,
the OMD algorithm with Legendre function F (.) admits the following:
1P
P
1
(15)
t∈[T ] g(t) · q − p(t) ≤ η
t∈[T ] DF (p(t), w(t + 1)) + η DF (q, p(1))
3.2.2
MSMW algorithm as an OMD
For our application, we focus on a particular choice of Legendre function that captures different learning
rates proportional to c−1
for different experts, as we saw earlier in Algorithm 1. We start by defining
i
the weighted negative entropy function.
Definition 3.3. Given expert-ranges {ci }i∈A , the weighted negative entropy is defined by
P
F (x) = i∈A ci · xi ln(xi )
(16)
P
Corollary 3.2. It is straightforward to see F (x) =
i ) is a non-negative Legendre
i∈A ci · xi ln(x
P
xi
function over RA
.
Moreover,
∇F
(x)
=
c
(1
+
ln(x
))
and
D
(x,
y)
=
i
i
i
F
+
i∈A ci · (xi ln( yi ) − xi + yi ).
We now have the following lemma that shows Algorithm 1 is indeed an OMD algorithm.
1, is equivalent to an OMD algorithm associated
Lemma 3.3. The MSMW algorithm, i.e. Algorithm
P
with the weighted negative entropy F (x) = i∈A ci · xi ln(xi ) as its Legendre function.
12
Proof Look
P at the gradient update step of OMD, as in Equation (13), with Legendre transform
F (x) = i∈A ci · xi ln(xi ). By using Corollary 3.2 we have
∇F (w(t + 1)) = ∇F (p(t)) + η · g(t) ⇒ ci (1 + ln(wi (t + 1))) = ci (1 + ln(pi (t))) + η · gi (t) ,
and therefore, wi (t + 1) = pi (t) · exp(η ·
gi (t)
ci ).
Moreover, for Bregman projection step we have
!
X
pi
) − pi + wi (t + 1))
ci · (pi ln(
p(t + 1) = argmin (DF (p, w(t + 1))) = argmin
wi (t + 1)
p∈∆A
p∈∆A
(17)
i∈A
This is a convex-minimization overPa convex set. To find a closed form solution,
P we look at the
pi
Lagrangian dual function L(p, λ) , i∈A ci · (pi ln( wi (t+1)
) − pi + wi (t + 1)) + λ( i∈A pi − 1) and the
Karush-Kuhn-Tucker (KKT) conditions ∇L(p∗ , λ∗ ) = 0. We have
p∗i
λ∗
) + λ∗ = 0 ⇒ p∗i = wi (t + 1) · exp(− )
(18)
wi (t + 1)
ci
P
P
∗
As i∈A p∗i = 1, λ∗ should be unique number s.t. i∈A wi (t + 1) · exp(− λci ) = 1, and then pi (t + 1) =
∗
wi (t + 1) · exp(− λci ). So, Algorithm 1 is equivalent to OMD with weighted negative entropy as its
Legendre transform.
ci · ln(
Lemma 3.4. For any initial distribution µ over A, and any learning rate parameter 0 < η ≤ 1, and
any benchmark distribution q over A, the MSMW algorithm satisfies that:
X
XX
(gi (t))2
qi
1 X
qi · Gi − E [Galg ] ≤ η
pi (t)
+ ·
ci qi ln
− q i + µi .
ci
η
µi
i∈A
3.3
t∈[T ] i∈A
i∈A
Regret bound for non-negative rewards - proof of Theorem 2.1
Proof of Theorem 2.1 Suppose imin is an action with the minimum ci . Let µ = (1 − η) · 1imin + η · π,
and let q = (1 − η) · 1i + η · π in Lemma 3.4. If i 6= imin , we get that (note that µj = qj for any
j 6= i, imin ):
X
qi
1
− q i + µi
(1 − η) · Gi + η ·
πj · Gj − E [Galg ] ≤ η · E [Galg ] + · ci · qi ln
η
µi
j∈A
1
qimin
+ · cimin · qimin ln
− qimin + µimin
η
µimin
By 1 ≥ qi > µi ≥ ηπi , the 2nd term on the RHS is upper bounded as:
qi
1
1
1
· ci · qi ln
− qi + µi ≤ · ci · ln
η
µi
η
ηπi
Similarly, by 1 ≥ µimin > qimin ≥ 0, the 3rd term on the RHS is upper bounded as
1
qimin
1
1
· cimin · qimin ln
− qimin + µimin ≤ · cimin ≤ · ci
η
µimin
η
η
Finally, note that Gj ≥ 0 for all j ∈ A in reward-only instances. So the LHS is lower bounded by
(1 − η) · Gi − E [Galg ] = (1 − η) · regreti − η · E [Galg ] .
Putting together we get that
E [regreti ] ≤
1
1
1
1
2η
· E [Galg ] + O
ln
· ci ≤ 3η · E [Galg ] + O
ln
· ci .
1−η
η
ηπi
η
ηπi
The theorem then follows by choosing η =
ǫ
3
and rearranging terms.
13
4
Multi Scale Online Learning with Bandit Feedback
In this section, we look at the bandit feedback version of multi scale online learning. Inspired by online
stochastic mirror descent algorithm, we introduce Bandit-MSMW algorithm. Our algorithm follows
the standard bandit route of using unbiased estimators for the rewards in a full information strategy
(in this case MSMW). We also mix the MSMW distribution with an extra uniform exploration, and
use a tailored initial distribution for our multi-scale learning setting.
Here is a map of this section. In Section 4.1 we propose our bandit algorithm and prove its general
regret guarantee for non-negative rewards. Then in Section 4.2 we show how to get a multi-scale style
regret guarantee for the best arm ci∗ , and a weaker guarantee for all arms {ci }iA .
4.1
Bandit Multi-Scale Multiplicative Weight (Bandit-MSMW) algorithm
We present our Bandit algorithm (Algorithm 2) when the set of actions A is finite (with |A| = k). Let
η be the learning rate and γ be the exploration probability. We show the following regret bound.
Algorithm 2 Bandit-MSMW
1: input exploration parameter γ > 0, learning rate η > 0.
γ
2: initialize p(1) = (1 − γ)1imin + k 1, where imin is the arm with minimum range cimin .
3: for t = 1, . . . , T do
4:
Let p̃(t) = (1 − γ)p(t) + γk 1.
5:
Randomly pick an expert it drawn from p̃(t), and observe git (t).
6:
Let g̃(t) be such that
gi (t) if i = it ;
p̃i (t)
g̃i (t) =
0
otherwise.
∀i ∈ A : wi (t + 1) ← pi (t) · exp( cηi · g̃i (t)).
P
∗
8:
Find λ∗ (e.g., binary search) s.t. i∈A wi (t + 1) · exp(− λci ) = 1.
∗
9:
∀i ∈ A : pi (t + 1) ← wi (t + 1) · exp(− λci ).
10: end for
7:
Lemma 4.1. For any exploration probability 0 < γ ≤ 21 and any learning rate parameter 0 < η ≤ γk ,
the Bandit-MSMW algorithm achieves the following regret bound when the gains are non-negative :
P
k
1
∀i ∈ A : E [regreti ] ≤ O η log γ · ci + η j∈A Gj + γ · Gi
Proof We further define:
ealg ,
G
ej
G
,
P
t∈[T ] git (t)
=
t∈[T ] g̃j (t)
.
P
P
t∈[T ] p̃(t) ·
In expectation over the randomness of the algorithm, we have:
h
i
ealg ; and
1. E [Galg ] = E G
h i
ej for any j ∈ A.
2. Gj = E G
14
g̃(t) ,
h
i
ei − G
ealg .
Hence, to upper bound E [regreti ] = Gi − E [Galg ], it suffices to upper bound E G
By the definition of the probability that the algorithm picks each arm, i.e., p̃(t), we have:
h
i
ealg ≥ (1 − γ) P
E G
t∈[T ] p(t) · g̃(t) .
Hence, we have that for any initial distribution q over A:
h i
h
i
h
i
X
P
ej − E G
ealg ≤ E P
e
qj · E G
j∈A qj · Gj −
t∈[T ] p(t) · g̃(t) +
j∈A
≤ E
hP
e
j∈A qj · Gj −
P
γ
1−γ E
h
i
ealg
G
i
h
i
ealg .
p(t)
·
g̃(t)
+
2γE
G
t∈[T ]
(19)
Next, we upper bound the 1st term on the RHS. Note that p(t)’s are the probability of choosing
experts by MSMW when the experts have rewards g̃(t)’s. By Lemma 3.4, we have that for any
benchmark distribution q over S, the Bandit-MSMW algorithm satisfies that:
X
X
X X pj (t)
2 1 X
qj
e
cj qj ln
· g̃j (t) +
− qj + pj (1) . (20)
qj · Gj −
p(t) · g̃(t) ≤ η
cj
η
pj (1)
j∈A
t∈[T ]
j∈A
t∈[T ] j∈A
g (t)
For any t ∈ [T ] and any j ∈ A, by the definition of g̃j (t), it equals p̃jj (t) with probability p̃j (t), and
equals 0 otherwise. Thus, if we fix the random coin flips in the first t − 1 rounds and, thus, fix p̃(t),
and take expectation over the randomness in round t, we have that:
2
pj (t)
gj (t) 2 pj (t) (gj (t))2
pj (t)
E
· g̃j (t)
· p̃j (t) ·
.
=
=
cj
cj
p̃j (t)
p̃j (t) cj
1
gj (t) ≤
Further note that p̃j (t) ≥ (1 − γ)pj (t), and gj (t) ≤ cj , the above is upper bounded by 1−γ
γ
2gj (t). Putting together with (20), we have that for any 0 < η ≤ n :
X
XX
X
X
q
1
j
ej −
cj qj ln
− qj + pj (1)
2gj (t) +
qj · G
p(t) · g̃(t) ≤ η
E
η
pj (1)
j∈A
j∈A
j∈A
t∈[T ]
t∈[T ]
X
1X
qj
= 2η
Gj +
cj qj ln
− qj + pj (1)
η
pj (1)
j∈A
j∈A
Combining with (19), we have:
h i
h
i
h
i
X
X
1X
qj
e
e
ealg
qj · E Gj − E Galg ≤ 2η
Gj +
cj qj ln
− qj + pj (1) + 2γE G
η
pj (1)
j∈A
j∈A
j∈A
Let q = (1 − γ)1i + γk 1. Recall that p(1) = (1 − γ)1imin + γk 1 (recall imin is the arm with minimum
range cimin ). Similar to
the discussionfor the expert problem in Section 3.3, the 2nd term on the RHS
is upper bounded by O η1 log γk · ci . Hence, we have:
h
i
h i
h
i
X
X
k
1
ealg .
e
e
log
· ci + 2γE G
(21)
qj · E Gj − E Galg ≤ 2η
Gj + O
η
γ
j∈A
j∈A
Further, the LHS is lower bounded as:
h
i
h i
h
i
h i γX h i
e
e
e
e
e
E Gj − E Galg ≥ (1 − γ)E Gi − E Galg .
(1 − γ)E Gi +
k
j∈A
The lemma then follows by putting it back to (21) and rearranging terms.
15
4.2
Regret bounds for non-negative rewards - proof of Theorem 2.3
Proof of Theorem 2.3 Let γ = ǫ and η = γk = kǫ in Lemma 4.1, we get that the expected regret
w.r.t. an action i ∈ A is bounded by:
P
O ǫ · Gi + kǫ j∈A Gj + ci · kǫ ln kǫ
.
k
When i = i∗ (best arm), regret is bounded by O ǫ · Gi∗ + c∗i · kǫ ln
P ǫ , as desired.
For the regret w.r.t. an arbitrary action, note that E [Galg ] ≥ γk j∈A Gj . Thus, the regret bound
w.r.t. an action i ∈ A in Lemma 4.1 is further upper bounded by:
h
i
ealg
O η1 log γk · ci + ηk
+
γ
·
E
G
γ
The theorem then follows by letting γ = ǫ and η =
5
5.1
γ
k
=
ǫ2
k.
Auctions and Pricing
Auctions and pricing as multi-scale online learning problems
Online single buyer auction and posted pricing Recall that in each round, the algorithm chooses
an action, i.e., a price, pt ∈ [1, h]; the adversary picks a value v(t) ∈ [1, h]; and the algorithm collects
reward gpt (t) = pt · 1(v(t) ≥ pt ). In order to obtain a 1 − ǫ approximation of the optimal revenue, it
suffices to consider prices of the form (1 + ǫ)j for 0 ≤ j ≤ ⌊log1+ǫ h⌋ = O( logǫ h ). As a result, we reduce
the online single buyer auction problem and the online posted pricing problem to a multi-sclae online
learning problem with full information and bandit feedback respectively with k = O( logǫ h ) actions
whose ranges form a geometric sequence (1 + ǫ)j , 0 ≤ j < k.
Online multi buyer auction In multi buyer auctions, we consider the set of all discretized Myersontype auctions as the action space. We start by defining Myerson-type auctions:
Definition 5.1 (Myerson-type auctions). A Myerson-type auction is defined by n non-decreasing
virtual value mappings φ1 , . . . , φn : [1, h] 7→ [−∞, h]. Given a value profile v1 , . . . , vn , the item is given
to the bidder j with the largest virtual value φj (vj ). Then, bidder j pays the minimum value that
would keep him as the the winner.
Myerson (1981) shows that when the bidders’ values are drawn from independent (but not necessarily identical) distributions, the revenue-optimal auction is a Myerson-type auction. Devanur et al.
(2016, Lemma 5) observe that to obtain a 1 − ǫ approximation, it suffices to consider the set of discretized Myerson-type auctions that treat each bidder’s value as if it is equal to the closest power of
1 + ǫ from below. As a result, it suffices to consider the set of discretized Myerson-type auctions, each
of which is defined by the virtual values of (1 + ǫ)j ’s, i.e., by O(n log h/ǫ) real numbers φℓ ((1 + ǫ)j ),
for ℓ ∈ [n], and 0 ≤ j ≤ ⌊log1+ǫ h⌋. Devanur et al. (2016); Gonczarowski and Nisan (2017) further
note that a discretized Myerson-type auction is in fact completely characterized by the total ordering
of φℓ ((1 + ǫ)j )’s; their actual values do not matter. Indeed, both the allocation rule and the payment
rule are determined by the ordering of virtual values. As a result, our action space is a finite set with
at most O((n log h/ǫ)!) actions. The range of an action, i.e., a discretized Myerson-type auction, is the
largest price ever charged by the auction, i.e., the largest value v of the form (1 + ǫ)j such that there
exists ℓ ∈ [n], φℓ (v) > φℓ ((1 + ǫ)−1 v).
16
5.2
Proof of Theorem 2.5
Proof Online single buyer auction. Recall the above formulation of the problem as an online learning
problem with full information. The case when h is known then follows by Theorem 2.1, letting π be
the uniform distribution over the k = O(log h/ǫ) actions, i.e., discretized prices.
When h is not known upfront, we consider a countably infinite action space comprised of all prices
of the form (1 + ǫ)j , for j ≥ 0. Then, let the prior distribution π be such that for any price p = (1 + ǫ)j ,
ǫ
· p1 . The approximation guarantee then follows by Theorem 2.1.
πp = ǫ(1 + ǫ)−j−1 = 1+ǫ
Online posted pricing. Recall the above formulation of the problem as an online learning problem
with bandit feedback. This part then follows by Theorem 2.3 with k = O(log h/ǫ) actions.
Online multi buyer auction. Recall the above formulation of the problem as an online learning
problem with full information. The case when h is known then follows by Theorem 2.1, where we let
π be the uniform distribution over the k = O((n log h/ǫ)!) actions, i.e., Myerson-type auctions.
When h is not known upfront, we consider a countably infinite action space A as follows. For any
p = (1 + ǫ)j , j ≥ 0, let the kp = O((n log p/ǫ)!) Myerson-type auctions for values in [1, p] be in A; we
assume these auctions treat any values greater than p as if they were p. Further, we choose the prior
ǫ
· 1p · k1p .
distribution π such that the probability mass of each auction for range [1, p] is equal to 1+ǫ
The approximation guarantee then follows by Theorem 2.1.
5.3
Proof of Theorem 2.6
Proof Online single buyer auction. When h is known, by Theorem 2.1, letting π be the uniform
distribution over the k = O(log h/ǫ) actions, i.e., discretized prices, we have that for any price p (recall
that cp = p):
Galg ≥ (1 − ǫ) · Gp − O
log(log h/ǫ)
ǫ
·p
.
For the δ-guarded optimal price p∗ (i.e., subject
to selling in at least δT rounds), we have Gp∗ ≥ δT ·p∗ .
Therefore, when T ≥ O log(log h/ǫ)/ǫ2 δ , the additive term of the above approximation guarantee is
at most ǫ · Gp∗ . So the theorem holds.
The treatment for the case when h is not known upfront is essentially the same as in Theorem 2.5.
We consider a countably infinite action space comprised of all prices of the form (1 + ǫ)j , for j ≥ 0.
ǫ
Then, let the prior distribution π be such that for any price p = (1 + ǫ)j , πp = ǫ(1 + ǫ)−j−1 = 1+ǫ
· p1 .
Online posted pricing. Recall the above formulation of the problem as an online learning problem
with bandit feedback. By Theorem 2.3 with k = O(log h/ǫ) actions, we have that for any price p:
h/ǫ)
Galg ≥ (1 − ǫ) · Gp − O log h log(log
·
p
.
ǫ3
Again, for the δ-guarded optimal price p∗ (i.e., subject to selling in at least δT rounds), we have
Gp∗ ≥ δT · p∗ . Therefore, when T ≥ O log h log log h/ǫ /ǫ4 δ , the additive term of the above
approximation guarantee is at most ǫ · Gp∗ . So the theorem holds.
Online multi buyer auction. Suppose i∗ is the δ-guarded best Myerson-type auction. Recall that
V̄ is the largest value such that there are at least δT distinct v(t)’s with maxℓ∈[n] vℓ (t) ≥ V̄ . So we
may assume without loss of generality that i∗ does not distinguish values greater than V̄ . Hence:
ci∗ ≤ V̄ .
17
(22)
Further, note that running a 2nd-price auction with anonymous reserve V̄ is a Myerson-type auction
(e.g., mapping values less than V̄ to virtual value −∞ and values greater than or equal to V̄ to virtual
value V̄ ), and it gets revenue at least δT · V̄ . So we have that:
Gp∗ ≥ δT · V̄ .
(23)
Finally, the above implies that to obtain a 1 − ǫ approximation, it suffices to consider prices that
are at least ǫδV̄ . Hence, it suffices to consider Myerson-type auctions that, for a given V̄ , do not
distinguish among values greater than V̄ , and do not distinguish among values smaller than ǫδV̄ .
There are O(log h/ǫ) different values of V̄ . Further, given V̄ , there are only O(log(1/ǫδ)/ǫ) distinct
values to be considered and, thus, there are at most O((n log(1/ǫδ)/ǫ)!) distinct Myerson-type auctions
of this kind. Hence, the total number of distinct Myerson-type actions that we need to consider is at
most:
n log(1/ǫδ)
log h
·
! .
k=O
ǫ
ǫ
When h is known, letting π be the uniform distribution over the k actions in Theorem 2.1, we have
that (recall Eqn. (22)):
n log (1/ǫδ) log(n log(1/ǫδ)/ǫ) log (log h/ǫ)
+
· V̄ .
Galg ≥ (1 − ǫ) · Gi∗ − O
ǫ2
ǫ
log(1/ǫδ)/ǫ)
log (log h/ǫ)
When T ≥ O n log (1/ǫδ) log(n
, the additive term of the above approximation
+
3
2
ǫ δ
ǫ δ
guarantee is at most ǫ · Gi∗ due to Eqn. (23). So the theorem holds.
Again, the treatment for the case when h is not known upfront is similar to that in Theorem 2.5.
When h is not known upfront, we consider a countably infinite action space A as follows. For any
V̄ = (1 + ǫ)j , j ≥ 0, let the k′ = O((n log(1/ǫδ)/ǫ)!) Myerson-type auctions that do not distinguish
among values greater than V̄ , and do not distinguish among values smaller than ǫδV̄ be in A. Further,
we choose the prior distribution π such that the probability mass of each Myerson-type auction for
ǫ
· V̄1 · k1′ . The approximation guarantee then follows by Theorem 2.1 and
a given V̄ is equal to 1+ǫ
essentially the same argument as the known h case.
Remark Devanur et al. (2016) show that when the values are drawn from independent regular distributions, the ǫ-guarded optimal is a 1 − ǫ approximation of the unguarded optimal. So our convergence
rate for the online multi buyer auction problem in Theorem 2.1 implies a Õ(nǫ−4 ) sample complexity
modulo a mild log log h dependency on the range, almost matching the best known sample complexity
upper bound for regular distributions.
6
Multi-scale Online Learning with Symmetric Range
In this section, we consider multi-scale online learning when the rewards are in a symmetric range, i.e.
for all i ∈ A and t ∈ [T ], gi (t) ∈ [−ci , ci ]. We look at both full information and bandit setting, and
prove action-specific regret upper bounds. We defer the regret lower bound proofs to the appendix,
Sections A.2 and A.3.
6.1
Multi-scale expert problem with symmetric range
Recall the proof of Lemma 3.4. The proof only requires gi (t) ∈ [−ci , ci ] for all i ∈ A, t ∈ [T ]. Choosing
q to be 1i , a vector with a 1-entry in ith coordinate and 0-entries elsewhere for an action i ∈ A, and
18
noting that
P
t∈[T ]
P
i∈A pi (t)
(gi (t))2
ci
≤
P
t∈[T ]
P
i∈A pi (t)
· gi (t) ,
we get the following regret bound as a corollary of Lemma 3.4.
Corollary 6.1. For any initial distribution µ over A, and any learning rate parameter 0 < η ≤ 1, the
MSMW algorithm achieves the following regret bound:
i
hP
1P
1
1
(24)
c
·
log
+
∀i ∈ A : E [regreti ] ≤ η · E
g
(t)
i
i
j∈A µj cj
t∈[T ]
η
µi + η
Now, we can prove the multi-scale regret upper bound in Theorem 2.7 using Corollary 6.1.
Proof of Theorem 2.7 The proof follows by choosing an appropriate initial distribution µ in Corollary 6.1. By Corollary 6.1, we have:
i
hP
P
+ η1 ci · log( µ1i ) + η1 j∈A µj cj
E [regreti ] ≤ η · E
g
(t)
i
t∈[T ]
Let imin be an action with the minimumPrange cimin = cmin . Consider an initial distribution
µj = πj cmin
j6=imin µj , i.e., putting all remaining probability mass
cj for all j 6= imin , and µimin = 1 −
on action imin . Then, the 3rd term on the RHS is upper bounded by:
P
P
P
j6=imin πj cmin + µimin cmin ≤ 2cmin ≤ 2ci .
j∈A µj cj =
j6=imin µj cj + µimin cimin =
For i 6= imin , by the definition of µi , we have:
i
hP
ci
) + η1 · 2cmin
+ η1 ci · log( π1i · cmin
E [regreti ] ≤ η · E
g
(t)
t∈[T ] i
i
hP
ci
1
1
=η·E
.
t∈[T ] gi (t) + O η log πi · cmin · ci
So the theorem
η = ǫ. For i = imin , note that µj ≤ πj for all j 6= imin and, thus,
P
P follows by choosing
µimin = 1 − j6=imin µj ≥ 1 − j6=imin πj = πimin = πimin ccimin . The theorem then holds following the
min
same calculation as in the j 6= imin case.
6.2
Multi-scale bandit problem with symmetric range
We start by showing the following regret bound, whose proof is an alteration of that for Lemma 4.1
under symmetric range (and is deferred to the appendix, Section A.6). Next, we prove Theorem 2.10.
min
} and any learning rate 0 < η ≤
Lemma 6.2. For any exploration rate 0 < γ ≤ min{ 12 , ccmax
Bandit-MSMW algorithm (Algorithm 2) achieves the following regret bound:
1
k
∀i ∈ A : E [regreti ] ≤ O
log
· ci + γT · cmax
η
γ
min
and η =
Proof of Theorem 2.10 Let γ = ǫ ccmax
γcmax = ǫcmin ≤ ǫci .
19
γ
k
γ
k,
the
in Lemma 6.2. Theorem follows noting that
References
Shipra Agrawal and Nikhil R Devanur. 2014. Bandits with concave rewards and convex knapsacks. In
Proceedings of the fifteenth ACM conference on Economics and computation. ACM, 989–1006.
Kareem Amin, Afshin Rostamizadeh, and Umar Syed. 2013. Learning prices for repeated auctions with
strategic buyers. In Advances in Neural Information Processing Systems. 1169–1177.
Sanjeev Arora, Elad Hazan, and Satyen Kale. 2012. The Multiplicative Weights Update Method: a
Meta-Algorithm and Applications. Theory of Computing 8, 1 (2012), 121–164.
Peter Auer, Nicolo Cesa-Bianchi, Yoav Freund, and Robert E Schapire. 1995. Gambling in a rigged
casino: The adversarial multi-armed bandit problem. In Foundations of Computer Science, 1995.
Proceedings., 36th Annual Symposium on. IEEE, 322–331.
Moshe Babaioff, Shaddin Dughmi, Robert Kleinberg, and Aleksandrs Slivkins. 2015. Dynamic pricing
with limited supply. ACM Transactions on Economics and Computation 3, 1 (2015), 4.
Ashwinkumar Badanidiyuru, Robert Kleinberg, and Aleksandrs Slivkins. 2013. Bandits with knapsacks.
In Foundations of Computer Science (FOCS), 2013 IEEE 54th Annual Symposium on. IEEE, 207–
216.
Ziv Bar-Yossef, Kirsten Hildrum, and Felix Wu. 2002. Incentive-compatible online auctions for digital
goods. In Proceedings of the thirteenth annual ACM-SIAM symposium on Discrete algorithms. Society
for Industrial and Applied Mathematics, 964–970.
Omar Besbes and Assaf Zeevi. 2009. Dynamic pricing without knowing the demand function: Risk
bounds and near-optimal algorithms. Operations Research 57, 6 (2009), 1407–1420.
Avrim Blum and Jason D Hartline. 2005. Near-optimal online auctions. In Proceedings of the sixteenth annual ACM-SIAM symposium on Discrete algorithms. Society for Industrial and Applied
Mathematics, 1156–1163.
Avrim Blum, Vijay Kumar, Atri Rudra, and Felix Wu. 2004. Online learning in online auctions.
Theoretical Computer Science 324, 2-3 (2004), 137–146.
Sébastien Bubeck. 2011. Introduction to online optimization. Lecture Notes (2011), 1–86.
Richard Cole and Tim Roughgarden. 2014. The sample complexity of revenue maximization. In Symposium on Theory of Computing, STOC 2014, New York, NY, USA, May 31 - June 03, 2014.
243–252.
Arnoud V den Boer. 2015. Dynamic pricing and learning: historical origins, current research, and new
directions. Surveys in operations research and management science 20, 1 (2015), 1–18.
Nikhil R Devanur, Zhiyi Huang, and Christos-Alexandros Psomas. 2016. The sample complexity of
auctions with side information. In Proceedings of the 48th Annual ACM SIGACT Symposium on
Theory of Computing. ACM, 426–439.
Nikhil R Devanur, Yuval Peres, and Balasubramanian Sivan. 2014. Perfect bayesian equilibria in
repeated sales. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms. SIAM, 983–1002.
20
Peerapong Dhangwatnotai, Tim Roughgarden, and Qiqi Yan. 2014. Revenue maximization with a
single sample. Games and Economic Behavior (2014).
Dylan Foster, Satyen Kale, Mehryar Mohri, and Karthik Sridharan. 2017. Personal communication.
(2017).
Yoav Freund and Robert E Schapire. 1995. A desicion-theoretic generalization of on-line learning
and an application to boosting. In European conference on computational learning theory. Springer,
23–37.
Yannai A Gonczarowski and Noam Nisan. 2017. Efficient Empirical Revenue Maximization in SingleParameter Auction Environments. In Proceedings of the ACM STOC.
Zhiyi Huang, Yishay Mansour, and Tim Roughgarden. 2015. Making the most of your samples. In
Proceedings of the Sixteenth ACM Conference on Economics and Computation. ACM, 45–60.
Robert Kleinberg and Tom Leighton. 2003. The value of knowing a demand curve: Bounds on regret for
online posted-price auctions. In Foundations of Computer Science, 2003. Proceedings. 44th Annual
IEEE Symposium on. IEEE, 594–605.
Jamie H Morgenstern and Tim Roughgarden. 2015. On the pseudo-dimension of nearly optimal auctions. In Advances in Neural Information Processing Systems. 136–144.
Roger B. Myerson. 1981. Optimal Auction Design. Mathematics of Operations Research 6, 1 (1981),
58–73.
Tim Roughgarden and Okke Schrijvers. 2016. Ironing in the dark. In Proceedings of the 2016 ACM
Conference on Economics and Computation. ACM, 1–18.
Ilya Segal. 2003. Optimal pricing mechanisms with unknown demand. The American economic review
93, 3 (2003), 509–529.
Vasilis Syrgkanis. 2017. A Sample Complexity Measure with Applications to Learning Optimal Auctions. arXiv preprint arXiv:1704.02598 (2017).
Kalyan T Talluri and Garrett J Van Ryzin. 2006. The theory and practice of revenue management.
Vol. 68. Springer Science & Business Media.
21
A
Other Deferred Proofs and Discussions
A.1
Discussion on choice of π for bandit symmetric range
We now describe how the choice of initial distribution π affects the bound given in Theorem 2.7.
• When the action set is finite, we can choose π to be the uniform distribution to get the term
O
1
log(kci /cmin ) · ci
ǫ
This recovers the standard bound by setting ci = cmax for all i ∈ A.
P
• We can choose πi = P ci cj to get O 1ǫ log( j∈A cj /cmin ) · ci . In particular, if the ci ’s form an
j∈A
arithmetic progression with a constant difference then this is just O logǫ k · ci .
P
• If there are infinitely many experts but i∈A c−1
is convergent, e.g., ci = cmin · (1 + ǫ)i−1 , then
i
P
−1
c2i
c
cmin
we can choose πi = P i c−1 for all i ∈ A. This gives O η1 log
· ci .
j cj · c2
min
j∈A j
A.2
Log factor dependence for symmetric range - proof of Theorem 2.9
Proof of Theorem 2.9 We first show that for any online learning algorithm, and any sufficiently
large h > 1, there is an instance that has two experts with c1 = 1 and c2 = h with T = Θ(log h)
rounds, such that either
√
or
E [regret2 ] > 21 T h + 51 h log2 h .
E [regret1 ] > 21 T + h ,
We will construct this instance with T = 12 log2 h − 1 rounds adaptively that always has gain 0 for
action 1 and gain either h or −h for action 2. The proof of the theorem then follows as cmin = 1,
cmax = h, T = 12 log2 h − 1, and k = 2 in this instance. Let qt denote the probability that the algorithm
picks action 2 in round t after having the same rewards 1 and h for the two actions respectively in the
first t − 1 rounds. We will first show that (1) if the algorithm has small regret with respect to action
1, then qt must be upper bounded since the adversary may let action 2 have cost −h in any round t
in which qt is too large. Then, we will show that (2) since qt is upper bounded for any 1 ≤ t ≤ T , the
algorithm must have large regret with respect to action 2.
We proceed with the upper bounding qt ’s. Concretely, we will show the following lemma.
√
t
Lemma A.1. Suppose E [regret1 ] ≤ 12 T + h. Then, for any 1 ≤ t ≤ T , we have qt ≤ √2 h .
Proof of Lemma A.1 We will prove by induction on t. Consider the base case t = 1. Suppose for
contradiction that q1 > √2 . Then, consider an instance in which action 2 always has gain. In this
h
case, the expected gain of the algorithm
(even if it always correctly picks action 1 in the remaining
√
h.
This
is a contradiction to the assumption that E [regret1 ] ≤
instance)
is
at
most
q
·
(−h)
<
−2
1
√
√
1
2 T + h < 2 h.
Next, suppose the lemma holds for all rounds prior to round t. Then, the expected gain of algorithm
in the first t − 1 rounds if arm 2 has gain H is
t−1
X
ℓ=1
qℓ · h ≤
t−1
X
ℓ=1
√
√
2ℓ h = 2t − 2 h .
22
t
Suppose for contradiction that qt > √2 . Then, consider an instance in which action 2 has gain H in
h
the first t − 1 rounds and −H afterwards. In this case, the expected gain of the algorithm (even if it
always correctly picks action 1 after round t) is at most
√
√
√
√
2t − 2 h + qt (−h) < 2t − 2 h + 2t h < −2 h .
√
√
This is a contradiction to the assumption that E [regret1 ] ≤ 21 T + h < 2 h.
√
Consider an instance in which action 2 always has gain H. Suppose that E [regret1 ] ≤ 12 T + h.
As an immediate implication of the above lemma, the algorithm is that the expected gain of the
algorithm is upper bounded by:
T
X
t=1
T
X
√
√
2t h < 2T +1 h = h .
qt h ≤
t=1
Note that in this instance E [G2 ] = T · h. Thus, the regret w.r.t. action 2 is at least (T − 1)h, which
is greater than 12 · E [G2 ] + 15 h log2 h for sufficiently large h.
A.3
Regret lower-bound for symmetric range -proof of Theorem 2.12
Proof of Theorem 2.12 We first show that for any online multi-scale bandits algorithm problem,
and there is an instance thatqhas two arms with c1 = 1 and c2 = h for some sufficiently large h, a
sufficiently large T , and ǫ =
h
256T
, such that either
E [regret1 ] > ǫT +
1
256ǫ h
or
,
E [regret2 ] > ǫT h +
2
1
256ǫ h
We will prove the existence of this instance by looking at the stochastic setting, i.e., the gain vectors
g(t)’s are i.i.d. for 1 ≤ t ≤ T . We consider two instances, both of which admit a fixed gain of 0 for
action 1. In the first instance, the gain of action 2 is h with probability 21 − 2ǫ, and −h otherwise.
Hence, the expected gain of playing action 2 is −4ǫh per round in instance 1. In the second instance,
the gain of action 2 is h with probability 21 + 2ǫ, and −h otherwise. Hence, the expected gain of playing
action two is 4ǫh per round in instance 2. Note this proves the theorem, as cmin = 1, cmax = h, k = 2
h
and and T = 256ǫ
2.
Suppose for contradiction that the algorithm satisfies:
E [regret1 ] ≤ ǫT +
1
256ǫ h
=
1
128ǫ h
,
E [regret2 ] ≤ ǫhT +
2
1
256ǫ h
=
2
1
128ǫ h
.
Let N1 denote the expected number of times that the algorithm plays action 2 in instance 1.
Then, the expected regret with respect to action 1 in instance 1 is N1 · 4ǫh. By the assumption that
1
1
E [regret1 ] ≤ 128ǫ
h, we have N1 ≤ 512ǫ
2.
Next, by standard calculation, we get that the Kullback-Leibler (KL) divergence of the observed
rewards in a single round in the two instances is 0 if action 1 is played and is at most 64ǫ2 (for
0 < ǫ < 0.1) if action 2 is played. So the KL divergence of the observed reward sequences in the two
instances is at most 64ǫ2 · N1 ≤ 81 .
Then, we use a standard inequality
about KL divergences.
For any measurable
function ψ : X 7→
{1, 2}, we have PrX∼ρ1 ψ(X) = 2 + PrX∼ρ2 ψ(X) = 1 ≥ 21 exp − KL(ρ1 , ρ2 ) . For any 1 ≤ t ≤ T ,
let ρ1 and ρ2 be the distribution of observed rewards up to a round t in the two instances, and let ψ(X)
be the action played by the algorithm. By this inequality and the above bound on the KL divergence
between the observed rewards in the two instances, we get that in each round, the probability that the
23
algorithm plays action 2 in instance 1, plus the probability that the algorithm plays action 1 in instance
2, is at least 21 exp (− 81 ) > 25 in any round t. Thus, the expected number of times that the algorithm
plays action 1 in instance 2 from round 1 to T , denoted as N2 , is at least N2 ≥ 52 · T − N1 ≥ 31 · T ,
where the second inequality holds for sufficiently large h. Therefore, the expected regret w.r.t. action
1
h2 . This is a contradiction to our assumption that
2 in instance 2 is at least: 4ǫh · 13 · T = 34 ǫhT > 128ǫ
1
2
E [regret2 ] ≤ 128ǫ h .
A.4
Proof of Lemma 3.4
Proof of Lemma 3.4 We have:
X
i∈A
qi · Gi − E [Galg ] =
X
t∈[T ]
q · g(t) −
X
t∈[T ]
p(t) · g(t) =
X
t∈[T ]
g(t) · q − p(t)
(25)
By applying the regret bound of OMD (Lemma 3.1) to upper-bound the RHS, we have
X
i∈A
qi · Gi − E [Galg ] ≤
1
1 X
DF (p(t), w(t + 1)) + DF (q, p(1))
η
η
(26)
t∈[T ]
To bound the first term in regret, a.k.a local norm, we have:
DF (p(t), w(t + 1)) =
X
i∈A
=
X
i∈A
ci · (pi (t) ln(
pi (t)
) − pi (t) + wi (t + 1))
wi (t + 1)
ci · pi (t)(−η ·
gi (t)
gi (t)
− 1 + exp(η ·
))
ci
ci
(27)
gi (t)
ci
∈ [−1, 1] because gi (t) ∈ [−ci , ci ] and 0 < η ≤ 1. By exp(x) − x − 1 ≤ x2 for
P
2
−1 ≤ x ≤ 1 and that ηgi (t) ∈ [−ci , ci ], the above is upper bounded by η 2 i∈A pi (t) (gi c(t))
. We can
i
also rewrite the second term in regret. In fact, if we set p(1) = µ, then
1 X
qi
1
− q i + µi
· DF (q, p(1)) = ·
ci qi ln
η
η
µi
Note that η ·
i∈A
P
2
By summing the upper-bounds η 2 i∈A pi (t) (gi c(t))
on each term of local norm in (27) for t ∈ [T ] and
i
putting all the pieces together, we get the desired bound.
We also provide an elementary proof of this lemma using first principles.
Proof of Lemma 3.4 from first principles Based on the update rule of Algorithm 1, we have
24
gi (t) =
ci
η
log( wip(t+1)
) for any i ∈ A. Therefore:
i (t)
X
g(t) · q − p(t) =
gi (t) qi − pi (t)
i∈A
=
X ci
i∈A
1
=
η
=
1
η
η
· log
X
i∈S
X
wi (t + 1)
· qi − pi (t)
pi (t)
wk (t + 1) X
pi (t)
ci · qi · log
+
ci · pi (t) · log
pk (t)
wi (t + 1)
i∈A
!
wk (t + 1) X
pk (t + 1)
ci · qi · log
+
ci · qi · log
pk (t + 1)
pk (t)
i∈S
i∈S
X
pi (t)
+
ci · pi (t) · log
wi (t + 1)
(28)
i∈A
Now, note that due to the normalization step of Algorithm 1, for any i ∈ S we have:
ci · log(
X
X
λ
wj (t + 1)
wi (t + 1)
)=λ=
cj · pj (t + 1) ·
=
cj · pj (t + 1) · log(
)
pi (t + 1)
cj
pj (t + 1)
j∈A
j∈A
So the first summation in (28) is equal to:
X
i∈S
ci · qi · log
X
wj (t + 1)
wk (t + 1) X
)
=
qi ·
cj · pj (t + 1) · log(
pk (t + 1)
pj (t + 1)
i∈S
=
X
j∈A
=
X
i∈A
j∈A
cj · pj (t + 1) · log(
ci · pi (t + 1) · log(
wj (t + 1)
)
pj (t + 1)
wi (t + 1)
)
pi (t + 1)
(29)
Combining Eqn. (28) and (29), we have:
1X
pi (t)
wi (t + 1)
g(t) · q − p(t) =
ci · pi (t) · log(
) + pi (t + 1) · log(
)
η
wi (t + 1)
pi (t + 1)
i∈A
+
1X
pi (t + 1)
ci · qi · log
η
pi (t)
i∈S
The 2nd part is a telescopic sum when we sum over t. We will upper bound the 1st part as follows.
By log(x) ≤ (x − 1), we get that:
X
wi (t + 1)
pi (t)
) + pi (t + 1) · log(
)
ci · pi (t) · log(
wi (t + 1)
pi (t + 1)
i∈A
X
pi (t)
≤
ci · pi (t) · log(
) − pi (t + 1) + wi (t + 1)
wi (t + 1)
i∈A
X
X
pi (t)
=
ci · pi (t) − pi (t + 1) +
ci · pi (t) · log(
) − pi (t) + wi (t + 1)
wi (t + 1)
i∈A
i∈A
25
Again, the 1st part is a telescopic sum when we sum over t. We will further work on the 2nd part.
By the relation between wi (t + 1) and pi (t), we get that:
X
X
pi (t)
gi (t)
gi (t)
ci · pi (t) · log(
) − pi (t) + wi (t + 1) =
ci · pi (t) −η ·
− 1 + exp(η ·
)
wi (t + 1)
ci
ci
i∈A
i∈A
gi (t)
ci
∈ [−1, 1] because gi (t) ∈ [−ci , ci ] and 0 < η ≤ 1. By exp(x) − x − 1 ≤ x2 for
P
2
−1 ≤ x ≤ 1 and that ηgi (t) ∈ [−ci , ci ], the above is upper bounded by η 2 i∈A pi (t) (gi c(t))
. Putting
i
together, we get that:
X
1X
pi (t + 1)
(gi (t))2
ci · qi · log
g(t) · q − p(t) ≤
+ pi (t) − pi (t + 1) + η
pi (t)
η
pi (t)
ci
Note that η ·
i∈S
i∈A
Summing over t, we have:
XX
1X
pi (T + 1)
(gi (t))2
g(t) · q − p(t) ≤
ci · qi · log
+ pi (1) − pi (T + 1) + η
pi (t)
η
pi (1)
ci
i∈S
t∈[T ] i∈A
Finally, by log(x) ≤ (x − 1), we get that qi log
pi (T +1)
qi
≤ pi (T + 1) − qi . Hence, we have:
XX
1X
qi
(gi (t))2
ci · qi · log
+ pi (1) − qi + η
pi (t)
g(t) · q − p(t) ≤
η
pi (1)
ci
i∈S
t∈[T ] i∈A
The lemma then follows by our choice of the initial distribution.
A.5
Proof of OMD regret bound
In order to prove the OMD regret bound, we need some properties of Bregman divergence.
Lemma A.2 (Properties of Bregman divergence (Bubeck, 2011)). Suppose F (·) is a Legendre function
and DF (·, ·) is its associated Bregman divergence as defined in Definition 3.1. Then:
• DF (x, y) > 0 if x 6= y as F is strictly convex, and DF (x, x) = 0.
• DF (., y) is a convex function for any choice of y.
• ( Pythagorean theorem) If A is a convex set, a ∈ A, b ∈
/ A and c = argmin (DF (x, b)), then
x∈A
DF (a, c) + DF (c, b) ≤ DF (a, b)
Given Lemma A.2, we are now ready to prove Lemma 3.1.
Proof of Lemma 3.1 To obtain the OMD regret bound, we have:
1
(q − p(t)) · (∇F (w(t + 1)) − ∇F (p(t)))
η
1
= (DF (qb, p(t)) + DF (p(t), w(t + 1)) − DF (qb, w(t + 1)))
η
(1) 1
1
≤ DF (p(t), w(t + 1)) + (DF (q, p(t)) − DF (q, p(t + 1)))
η
η
q · g(t) − p(t) · g(t) =
26
(30)
where in (1) we use DF (p(t + 1), w(t + 1)) ≥ 0 and DF (q, p(t + 1)) + DF (p(t + 1), w(t + 1)) ≤
DF (q, w(t + 1)) due to Pythagorean theorem (Lemma A.2). By summing up both hand sides of
(30) for t = 1, · · · , T we have:
X
t∈[T ]
A.6
1 X
1
DF (p(t), w(t + 1)) + DF (q, p(1))
g(t) · q − p(t) ≤
η
η
(31)
t∈[T ]
Symmetric range bandit regret bound - proof of Lemma 6.2
Proof of Lemma 6.2 We further define:
P
ealg , P
G
t∈[T ] p̃(t) · g̃(t) ,
t∈[T ] git (t) =
P
ej ,
G
t∈[T ] g̃j (t) .
In expectation over the randomness of the algorithm, we have:
h
i
ealg ; and
1. E [Galg ] = E G
h i
ej for any j ∈ A.
2. Gj = E G
h
i
ei − G
ealg .
Hence, to upper bound E [regreti ] = Gi − E [Galg ], it suffices to upper bound E G
By the definition of the probability that the algorithm picks each arm, i.e., p̃(t), and that reward
of each round is at least −cmax , we have that:
h
i
X
ealg ≥ (1 − γ)
E G
p(t) · g̃(t) − γT cmax .
t∈[T ]
Hence, for any benchmark distribution q over A, we have that:
i
h
i
hP
i
h
h i
P
P
γ
γ
e
e
e
e
E
G
≤
E
q
·
G
−
p(t)
·
g̃(t)
+
−
E
G
q
·
E
G
alg + 1−γ T cmax
j
alg
j
j∈A j
t∈[T ]
j∈A j
1−γ
hP
i
h
i
P
e
e
≤ E
q
·
G
−
p(t)
·
g̃(t)
+
2γE
G
j
alg + 2γT cmax
j∈A j
t∈[T ]
hP
i
P
e
≤ E
(32)
j∈A qj · Gj −
t∈[T ] p(t) · g̃(t) + 4γT cmax .
where the 2nd inequality is due to γ ≤ 21 , and the 3rd inequality follows by that cmax is the largest
possible reward per round.
Next, we upper bound the 1st term on the RHS of (32). Note that p(t)’s are the probability of
choosing experts by MSMW when the experts have rewards g̃(t)’s. By Lemma 3.4, we have that for
any benchmark distribution q over S, the Bandit-MSMW algorithm satisfies that:
X
X
X X pj (t)
2 1 X
qj
e
qj · Gj −
p(t) · g̃(t) ≤ η
cj qj ln
· g̃j (t) +
− qj + pj (1) . (33)
cj
η
pj (1)
j∈A
t∈[T ]
j∈A
t∈[T ] j∈A
g (t)
For any t ∈ [T ] and any j ∈ A, by the definition of g̃j (t), it equals p̃jj (t) with probability p̃j (t), and
equals 0 otherwise. Thus, if we fix the random coin flips in the first t − 1 rounds and, thus, fix p̃(t),
and take expectation over the randomness in round t, we have that:
2
pj (t)
gj (t) 2 pj (t) (gj (t))2
pj (t)
=
E
· g̃j (t)
· p̃j (t) ·
.
=
cj
cj
p̃j (t)
p̃j (t) cj
27
1
|gj (t)| ≤
Further note that p̃j (t) ≥ (1 − γ)pj (t), and |gj (t)| ≤ cj , the above is upper bounded by 1−γ
γ
2|gj (t)| ≤ 2cmax . Putting together with (33), we have that for any 0 < η ≤ n :
XX
X
X
qj
1X
ej −
cj qj ln
− qj + pj (1)
2cmax +
qj · G
p(t) · g̃(t) ≤ η
E
η
pj (1)
j∈A
j∈A
t∈[T ] j∈A
t∈[T ]
1X
qj
= 2ηT kcmax +
cj qj ln
− qj + pj (1)
η
pj (1)
j∈A
Combining with (32), we have (recall that η ≤ γk ):
i
h
i
qj
1X
e
e
cj qj ln
qj · E Gj − E Galg ≤ 2ηT kcmax +
− qj + pj (1) + 4γT cmax
η
pj (1)
j∈A
j∈A
qj
1X
≤
cj qj ln
− qj + pj (1) + 6γT cmax
η
pj (1)
X
h
j∈A
Let q = (1 − γ)1i + γk 1. Recall that p(1) = (1 − γ)1imin + γk 1 (recall imin is the arm with minimum
range cimin ). Similar to the discussionfor the expert problem in Section 3.3, the 1st term on the RHS
is upper bounded by O η1 log γk · ci . Hence, we have:
X
j∈A
i
h
i
k
1
e
e
log
· ci + 6γT cmax .
qj · E Gj − E Galg ≤ O
η
γ
h
Further, the LHS is lower bounded as:
h i γX h i
h
i
h i
h
i
ei +
ej − E G
ealg ≥ (1 − γ)E G
ei − γT cmax − E G
ealg .
(1 − γ)E G
E G
k
j∈A
The lemma then follows by putting it back to (34) and rearranging terms.
28
(34)
| 8cs.DS
|
1
Semantically-Secured Message-Key Trade-off
over Wiretap Channels with Random
arXiv:1708.04283v1 [cs.IT] 14 Aug 2017
Parameters
Alexander Bunin, Ziv Goldfeld, Haim H. Permuter, Shlomo Shamai (Shitz), Paul Cuff and Pablo
Piantanida
Abstract
We study the trade-off between secret message (SM) and secret key (SK) rates, simultaneously achievable over
a state-dependent (SD) wiretap channel (WTC) with non-causal channel state information (CSI) at the encoder. This
model subsumes other instances of CSI availability as special cases, and calls for efficient utilization of the state
sequence for both reliability and security purposes. An inner bound on the semantic-security (SS) SM-SK capacity
region is derived based on a superposition coding scheme inspired by a past work of the authors. The region is
shown to attain capacity for a certain class of SD-WTCs. SS is established by virtue of two versions of the strong
soft-covering lemma. The derived region yields an improvement upon the previously best known SM-SK trade-off
result reported by Prabhakaran et al., and, to the best of our knowledge, upon all other existing lower bounds for
either SM or SK for this setup, even if the semantic security requirement is relaxed to weak secrecy. It is demonstrated
that our region can be strictly larger than those reported in the preceding works.
I. I NTRODUCTION
A. Background
Physical layer security (PLS) [1]–[3], rooted in information-theoretic (IT) principles, is an approach to provably
secure communication that dates back to Wyner’s celebrated 1975 paper on the wiretap channel (WTC) [4]. By
harnessing randomness from the noisy communication channel and combining it with proper physical layer coding,
PLS guarantees protection against computationally-unlimited eavesdroppers, with no requirement that the legitimate
parties share a secret key (SK) in advance. Two fundamental questions in the field of PLS regard finding the best
The work of Alexander Bunin and Shlomo Shamai was supported by the European Union’s Horizon 2020 Research and Innovation Programme,
grant agreement No. 694630. The work of Z. Goldfeld and H. H. Permuter was supported by the Israel Science Foundation (grant no. 684/11),
an ERC starting grant and the Cyber Security Research Grant at Ben-Gurion University of the Negev. The work of Paul Cuff was supported by
the National Science Foundation, grant CCF-1350595, and the Air Force Office of Scientific Research, grant FA9550-15-1-0180.
This work was presented in part at the second international Workshop on Communication Security (WCS) affiliated to EUROCRYPT 2017.
A. Bunin and S. Shamai are with the Department of Electrical Engineering, Technion – Israel Institute of Technology, Haifa, Israel
([email protected],[email protected]). Z. Goldfeld and H. H. Permuter are with the Department of Electrical and Computer
Engineering, Ben-Gurion University of the Negev, Beer-Sheva, Israel ([email protected], [email protected]). P. Cuff is with the Department
of Electrical Engineering, Princeton University, Princeton, NJ, USA ([email protected]). P. Piantanida is with the Laboratory of Signals and
Systems, CentraleSupélec-CNRS-Université, Paris-Sud, France ([email protected]).
2
achievable transmission rate of a secret message (SM) over a noisy channel, and the highest attainable SK rate that
distributed parties can agree upon based on correlated observations.
The base model for SM transmission is Wyner’s WTC [4], where two legitimate parties communicate over a
noisy channel in the presence of an eavesdropper. The secrecy capacity of the degraded WTC was derived in [4],
and the result was extended to the general case by Csiszár and Körner [5]. The security analyses in both [4] and
[5] relied on evaluating particular conditional entropy terms, named equivocation. This technique has been widely
adopted in the IT community ever since.
Recently, distribution approximation arguments emerged as the tool of choice for proving security. This approach
relies on a soft-covering lemma (SCL) that originated in another 1975 paper by Wyner [6]. The SCL states that
the distribution induced by randomly selecting a codeword from an appropriately chosen codebook and passing it
through a memoryless channel will be asymptotically indistinguishable from the distribution of random noise. The
SCL was further developed over the years and stricter proximity measures between distributions were achieved
[7]–[10]. Based on these more advanced versions, one can make the channel output observed by the eavesdropper
in the WTC seem like noise and, in particular, be approximately independent of the confidential data. This, in turn,
implies IT security. Notably, [11] and [12] focused on tight soft-covering exponents with respect to relative entropy
and total variation, respectively.
The study of SK agreement was pioneered by Maurer [13], and, independently, by Ahlswede and Csiszár [14],
who studied the achievable SK rates based on correlated observations at the terminals that can communicate via
a noiseless and rate unlimited public link. The SK capacity when only one-way public communication is allowed
was characterized in [14]. This result was generalized in [15] to the case where the public link has finite capacity.
The optimal random coding scheme for these cases is a combination of superposition coding and Wyner-Ziv coding
[16]. If the encoder controls its source (rather than just observing it), this source becomes a channel input and the
setup evolves to a WTC. This is a special case of the SK channel-type model that was also studied in [14].
B. Model and Contributions
A more general framework to consider is the state-dependent (SD) WTC with non-causal encoder channel
state information (CSI). This model combines the WTC and the Gelfand and Pinsker (GP) channel [17], and
is therefore sometimes referred to as the GP-WTC. The dependence of the channel’s transition probability on the
state sequence accounts for the possible availability of correlated sources at the terminals. The similarity between
the SM transmission and the SK agreement tasks makes their integration in a single model natural. Adhering to
the most general framework, we study the SM-SK rate pairs that are simultaneously achievable over a SD-WTC
with non-causal encoder CSI.
The scenario where there is only a SM was studied in [18], where an achievable SM rate formula was established.
This result was improved in [19] based on a novel superposition coding scheme1 . SK agreement over the GP-WTC
was the focus of [22], and, more recently, of [23] (see also references therein). The combined model was considered
1
The respective causal scenario was recently studied in [20], [21].
PSfrag replacements
3
WSn
S
M
Encoder fn
Y
X
Decoder φn
(M̂, K̂)
n
WY,Z|S,X
Z
Eavesdropper M, K
K
Fig. 1. The state-dependent wiretap channel with non-casual encoder channel state information, exploited for simultaneous secret message
transmission and secret key generation.
by Prabhakaran et al. [24], who derived a benchmark inner bound on the SM-SK capacity region. The result from
[24] is optimal for several classes of SD-WTCs.
We propose a novel superposition coding scheme for the combined model that subsumes all the aforementioned
achievability results as special cases. Specifically, [18], [19], [22]–[24], as well as all the other existing inner bounds
(on SM transmission, SK agreement or both) that are known to the authors, are captured. Furthermore, our inner
bound is shown to achieve strictly higher rates than each of these previous results.
The coding scheme used herein is inspired by [19]. Namely, an over-populated superposition codebook that
encodes the entire confidential message in its outer layer is utilized. Using the redundancies in the inner and outer
layers, the transmission is correlated with the state sequence by means of the likelihood encoder [25]. Although
the redundancy indices are chosen as part of the encoding process, we show that their true distribution is close to
uniform. Consequently, as long as a certain redundancy index is kept secret (along with the confidential message),
it may be declared as a SK. The security analysis is based on constructing the inner codebook such that it is better
observable by the eavesdropper, making the inner layer index decodable by him/her. This enhances the secrecy
resources that the legitimate parties can extract from the outer layer, which they use to secure the SM and part of
the redundancy index of the outer layer. The latter is declared as the SK.
Our results are derived under the strict metric of semantic-security (SS). The SS criterion is a cryptographic
gold standard that was adapted to the WTC framework (of computationally unbounded adversaries with a noisy
observation) in [26]. As was shown in [26], SS is equivalent to negligible mutual information (MI) between the
confidential information (in our case, the SM-SK pair) and the eavesdropper’s observations, when maximized over
all possible message distributions. The proof of SS relies on the strong SCL for superposition [19, Lemma 1] and
the heterogeneous SCL [10, Lemma 1].
Since the past secrecy results from [18], [22]–[24] were derived under the weak secrecy metric (i.e., a vanishing
normalized MI with respect to a uniformly distributed message-key pair), our achievability outperforms those
schemes, not only in terms of the achievable rate pairs, but also in the upgraded sense of security.
4
C. Organization
This paper is organized as follows. Section II establishes notation and definitions and sets up the SD-WTC
problem. Section III states our main result – an inner bound on the SM-SK optimal trade-off region. In Section IV
our inner bound is shown to be tight for a certain class of channels. In Section V we discuss past results captured
within the considered framework, and illustrate the improvement our result yields. The proof of the main result is
the content of Section VI. Finally, Section VII summarizes the main achievements and outlines the main insights
emerging from this work.
II. P RELIMINARIES
AND
P ROBLEM S ET-U P
A. Preliminaries
We use the following notations. As is customary, N is the set of natural numbers, while R are the reals. We
further define R+ = {x ∈ R|x ≥ 0}. Given two real numbers a, b, we denote by [a : b] the set of integers
n ∈ N a ≤ n ≤ b . Calligraphic letters denote sets, e.g., X , while |X | stands for the cardinality of X . X n
denotes the n-fold Cartesian product of X . An element of X n is denoted by xn = (x1 , x2 , . . . , xn ); whenever the
dimension n is clear from the context, vectors (or sequences) are denoted by boldface letters, e.g., x.
Let Ω, F, P be a probability space, where Ω is the sample space, F is the σ-algebra and P is the probability
measure. Random variables over Ω, F, P are denoted by uppercase letters, e.g., X, with conventions for random
vectors similar to those for deterministic sequences. The probability of an event A ∈ F is denoted by P(A), while
P(A B ) denotes the conditional probability of A given B. We use 1A to denote the indicator function of A ∈ F.
The set of all probability mass functions (PMFs) on a finite set X is denoted by P(X ), i.e.,
X
P(X ) = p : X → [0, 1]
p(x) = 1 .
(1)
x∈X
PMFs are denoted by letters such as p or q, with a subscript that identifies the random variable and its possible
conditioning. For example, for two discrete correlated random variables X and Y over the same probability space,
we use pX , pX,Y and pX|Y to denote, respectively, the marginal PMF of X, the joint PMF of (X, Y ) and the
conditional PMF of X given Y . In particular, pX|Y : Y → P(X ) represents the stochastic matrix whose elements
are given by pX|Y (x|y) = P X = x|Y = y . Expressions such as pX,Y = pX pY |X are to be understood as
pX,Y (x, y) = pX (x)pY |X (y|x), for all (x, y) ∈ X × Y. Accordingly, when three random variables X, Y and Z
satisfy pX|Y,Z = pX|Y , they form a Markov chain, which is denoted by X −− Y −− Z.
Any PMF q ∈ P(X ) gives rise to a probability measure on (X , 2X )2 , which we denote by Pq ; accordingly,
P
Pq A) = x∈A q(x) for every A ⊆ X . We use Eq to denote an expectation taken with respect to Pq . Similarly,
we use Hq and Iq to indicate that an entropy or a mutual information term are calculated with respect to the PMF
q. For a random vector X n , if the entries of X n are drawn in an independent and identically distributed (i.i.d.)
Qn
manner according to pX , then for every x ∈ X n we have pX n (x) = i=1 pX (xi ) and we write pX n (x) = pnX (x).
2 Here
2X stands for the power set of X .
5
Similarly, if for every (x, y) ∈ X n ×Y n we have pY n |X n (y|x) =
Qn
i=1
pY |X (yi |xi ), then we write pY n |X n (y|x) =
pnY |X (y|x). The conditional product PMF pnY |X given a specific sequence x ∈ X n is denoted by pnY |X=x .
Pn
, where N (x|x) = i=1 1{xi =x} . We use
The empirical PMF νx of a sequence x ∈ X n is νx (x) , N (x|x)
n
Tǫn (pX ) to denote the set of letter-typical sequences of length n with respect to the PMF pX and the non-negative
number ǫ, i.e., we have
n
o
Tǫn (pX ) = x ∈ X n νx (x) − pX (x) ≤ ǫpX (x), ∀x ∈ X .
(2)
Definition 1 (Total Variation) Let (Ω, F) be a measurable space and µ and ν be two probability measures on
that space. The total variation between µ and ν is
||µ − ν||TV = sup µ(A) − ν(A) .
(3a)
A∈F
If the sample space Ω is countable, p, q ∈ P(Ω) and Pp and Pq are the probability measures induced by p and q,
respectively, then (3a) reduces to
||Pp − Pq ||TV =
1X
p(x) − q(x) , ||p − q||TV .
2
(3b)
x∈Ω
B. Problem Setup
We study the SD-WTC with non-causal encoder CSI, for which we establish a novel achievable region of
semantically secured message-key rate pairs.
Let S, X , Y and Z be finite sets. The S, X , Y, Z, WS , WY,Z|S,X discrete and memoryless (DM) SD-WTC
with non-causal encoder CSI is shown in Fig. 1. A state sequence s ∈ S n is sampled in an i.i.d. manner according
to WS and revealed in a non-causal fashion to the sender. Independently of the observation of s, the sender chooses
a message m from the set 1 : 2nRM and maps the pair (s, m) onto a channel input sequence x ∈ X n and a
key index k ∈ 1 : 2nRK (the mapping may be random). The sequence x is transmitted over the SD-WTC with
transition probability WY,Z|S,X : S × X → P(Y × Z). The output sequences y ∈ Y n and z ∈ Z n are observed by
the receiver and the eavesdropper, respectively. Based on y, the receiver produces the pair (m̂, k̂), its estimates of
(m, k). The eavesdropper tries to glean whatever it can about the message-key pair from z.
Remark 1 (Most General Model) The considered model is the most general instance of a SD-WTC with noncausal CSI known at some or all of the terminals. Seemingly, the broadest model one may consider is when the
SD-WTC WỸ,Z̃|St ,Xt ,Sr ,Se is driven by a triple of correlated state random variables (St , Sr , Se ) ∼ WSt ,Sr ,Se ,
where St , Sr and Se are known to the transmitter, the receiver and the eavesdropper, respectively. However, setting
S = St , Y = (Ỹ, Sr ), Z = (Z̃, Se ) in a SD-WTC with non-causal encoder CSI and defining the channel’s transition
probability as
WY,Z|S,X = W(Ỹ,Sr ),(Z̃,Se )|S,Xt = WSr ,Se |St WỸ,Z̃|S,Xt ,Sr ,Se ,
6
one recovers the aforementioned SD-WTC from the model with non-causal CSI only at the encoder. Our model also
supports the existence of a public or a private bit-pipe (respectively, from the transmitter to the receiver and the
eavesdropper, or only to the receiver), in addition to, or instead of, the noisy channel.
Definition 2 (Code) An (n, RM , RK )-code cn for the SD-WTC with non-causal encoder CSI and a message set
Mn , 1 : 2nRM and a key set Kn , 1 : 2nRK is a pair of functions (fn , φn ) such that
1) fn : Mn × S n → P(Kn × X n ) is a stochastic encoder.
2) φn : Y n → Mn × Kn is the decoding function.
For any message distribution pM ∈ P(Mn ) and any (n, RM , RK )-code cn , the induced joint PMF is
n
(y, z|x, s)1
p(cn ) (s, m, k, x, y, z, m̂, k̂) = WSn (s)PM (m)fn (k, x|m, s)WY,Z|S,X
(m̂,k̂)=φn (y)
.
(4)
The probability measure induced by p(cn ) is Pp(cn ) . The performance of cn is evaluated in terms of its rate pair
(RM , RK ), its maximal decoding error probability, the key uniformity and independence metric, and the SS-metric.
Definition 3 (Error Probability) The error probability of an (n, RM , RK )-code cn is
e(cn ) , max em (cn ),
(5a)
m∈Mn
where for any m ∈ Mn
em (cn ) , Pp(cn ) M̂, K̂ 6= (m, K) M = m
X
X
WSn (s)fn (k, x|m, s)
WYn|S,X (y|x, s),
=
(5b)
y∈Y n :
φn (y)6=(m,k)
(s,x)
∈S n ×X n
and subscript p(cn ) denotes that the underlying PMF is (4).
Remark 2 (Operational Interpretation of the Error Probability) The error probability in (5a) is defined by
maximizing (5b) over the set of messages Mn . The maximization is only with respect to the message (rather
than with respect to the SM-SK pair) because, while the choice of M ∼ pM is independent of the code cn , the
distribution of the SK, K, and its estimate, K̂, is induced by the code (see (4)). A similar logic applies for the
subsequent definition of the key uniformity and independence metric.
Definition 4 (Key Uniformity and Independence Metric) The key uniformity and independence (of the message)
metric under the (n, RM , RK )-code cn is
δ(cn ) , max δm (cn ),
(6a)
m∈Mn
where for any m ∈ Mn
(c )
(U)
n
− pK n
δm (cn ) , pK|M=m
(U)
and pKn is the uniform PMF over Kn .
TV
(6b)
7
Definition 5 (Information Leakage and SS Metric) The information leakage to the eavesdropper under the
(n, RM , RK )-code cn and the message PMF pM ∈ P(Mn ) is ℓ(pM , cn ) , Ip(cn ) (M, K; Z), where Ip(cn ) denotes
that the MI is taken with respect to (4). The SS metric with respect to cn is
ℓSem (cn ) ,
max
pM ∈P(Mn )
ℓ(pM , cn ).
(7)
Definition 6 (Achievability) A pair (RM , RK ) ∈ R2+ is called an achievable SS message-key rate pair for the
SD-WTC with non-causal encoder CSI, if for every ǫ > 0 and sufficiently large n there exists an (n, RM , RK )-code
cn with
max e(cn ), δ(cn ), ℓSem (cn ) ≤ ǫ.
(8)
Definition 7 (SS-Capacity) The SS message-key capacity region CSem of the SD-WTC with non-causal encoder
CSI is the convex closure of the set of all achievable SS message-key rate pairs.
III. M AIN R ESULT
The main result of this work is a novel inner bound on the SS message-key capacity region of the SD-WTC
with non-causal encoder CSI. Our achievable region is at least as good as the best known achievability results for
the considered problem, and is strictly larger in some cases. To state our main result, let U and V be finite sets and
for any qU,V,X|S : S → P(U × V × X ) define
RA qU,V,X|S
RM ≤ I(U, V ; Y ) − I(U, V ; S),
2
, (RM , RK ) ∈ R+ RM + RK ≤ I(V ; Y |U ) − I(V ; Z|U ),
RM + RK ≤ I(U, V ; Y ) − I(V ; Z|U ) − I(U ; S)
,
(9)
where the MI terms are calculated with respect to the joint PMF WS qU,V,X|S WY,Z|S,X , under which (U, V ) −−
(S, X) −− (Y, Z) forms a Markov chain.
Theorem 1 (SS Message-Key Capacity Inner Bound) The following inclusion holds:
CSem ⊇ RA ,
[
qU,V,X|S
RA qU,V,X|S .
(10)
The proof of Theorem 1 is given in Section VI, and is based on a secured superposition coding scheme. An
over-populated two-layered superposition codebook is constructed (independently of the state sequence), in which
the entire secret message is encoded in the outer layer. Thus, no data is carried by the inner layer. The likelihood
encoder [25] uses the redundancies in the inner and outer codebooks to correlate the transmitted codewords with
the observed state sequence. Upon doing so, part of the correlation index from the outer layer is declared by the
encoder as the key. The inner layer is designed to utilize the part of the channel which is better observable by
the eavesdropper. This saturates the eavesdropper with redundant information and leaves him/her with insufficient
8
resources to extract any information on the SM-SK pair from the outer layer. The legitimate decoder, on the other
hand, decodes both layers of the codebook and declares the appropriate indices as the decoded message-key pair.
Remark 3 (Interpretation of Theorem 1) To get some intuitive understanding of the result of Theorem 1, we
examine RA (qU,V,X|S ) from two different perspectives: when the joint PMF WS qU,V,X|S WY,Z|S,X is such that
I(U ; Y ) ≥ I(U ; S), and when the opposite inequality holds.
If I(U ; Y ) ≥ I(U ; S), the third rate bound in RA (qU,V,X|S ) becomes redundant and the dominating bounds are
RM ≤ I(U, V ; Y ) − I(U, V ; S),
(11a)
RM + RK ≤ I(V ; Y |U ) − I(V ; Z|U ).
(11b)
The right-hand side (RHS) of (11a) is the total rate of reliable (secured and unsecured) communication that our
superposition codebook supports. This clearly bounds the rate of the SM that may be transmitted. For (11b), the MI
difference on the RHS is the total rate of secrecy resources that are produced by the outer layer of the codebook.
Since the security of our SM-SK pair comes entirely from that outer layer, this MI difference is an upper bound on
the sum of rates.
For the opposite case, if I(U ; Y ) < I(U ; S), then the second inequality in RA is inactive and we are left with
RM ≤ I(U, V ; Y ) − I(U, V ; S),
RM + RK
(12a)
≤ I(V ; Y |U ) − I(V ; Z|U ) − I(U ; S) − I(U ; Y ) .
(12b)
While the interpretation of (12a) remains as before, to understand (12b) consider the following. Since I(U ; S) is
approximately the rate of the inner codebook, I(U ; Y ) < I(U ; S) means that looking solely at the inner layer,
the decoder lacks the resolution to decode it. However, the success of our communication protocol relies on the
decoder reliably decoding both layers. Therefore, in this case, some of the rate from the outer layer is allocated
to convey the inner layer index. Recalling that our security analysis is based on revealing the inner layer to the
eavesdropper, this rate allocation effectively results in a loss of I(U ; S) − I(U ; Y ) in the secrecy resources of the
outer layer, giving rise to the rate bound from (12b).
Remark 4 (Alternative Representations of RA ) By defining Ṽ = (U, V ), we see that it suffices to restrict the
maximization in (10) to joint PMFs that satisfy the Markov chain U −− V −− (S, X) −− (Y, Z). Furthermore,
rewriting the the bounds on RM + RK from (9) as
RM + RK ≤ I(U, V ; Y ) − I(U, V ; Z) − max I(U ; Y ), I(U ; S) + I(U ; Z),
(13)
it is evident that maximizing only over joint PMFs satisfying I(U ; Z) ≥ max I(U ; Y ), I(U ; S) attains optimality.
Indeed, if the opposite inequality holds, one could always choose Ṽ = (U, V ) and Ũ = ∅ to achieve higher rates.
9
Adapting Theorem 1 to the Rate-Equivocation Framework
A confidential transmission of a SM requires channel resources for both reliability and security. The lesser of
the two resources, therefore, limits the feasible transmission rates. The main focus of this paper is utilization of
the residual secrecy resources that the SD-WTC offers. However, if secrecy is the lesser resource, the superior
capability of the channel to support reliable communication may be utilized by considering a Rate-Equivocation
framework.
Theorem 1 naturally extends to an inner bound on the rate-equivocation region of the considered SD-WTC [5],
[27]. Equivocation represents the portion of the message that can be secured from the eavesdropper. Intuitively, it
answers the question of how much information does the eavesdropper lack for decoding the entire message. The
rate-equivocation framework enables communicating at rates higher than the secrecy capacity, as long as full secrecy
is forfeited. Since equivocation has added value over full secrecy only when the channel offers more resources for
reliable communication than for security, for simplicity we assume RK = 0.
(c )
Formally, the equivocation rate of an (n, R)-code cn is RE n ,
1
n Hp(cn ) (M̃ |Z),
where p(cn ) is given in (4) and
M̃ is a uniformly distributed message. The achievability of a rate-equivocation pair (R, RE ) ∈ R2+ requires the
(c )
existence of a sequence of (n, R)-codes {cn }n∈N with a vanishing error probability and an equivocation rate RE n
(c )
that satisfies RE n ր RE as n → ∞.
An adaptation of the arguments from the proof of Theorem 1 (see Section VI) shows that any rate-equivocation
pair (R, RE ) ∈ R2+ satisfying
R ≤ I(U, V ; Y ) − I(U, V ; S),
(14a)
RE ≤ R,
(14b)
RE ≤ I(V ; Y |U ) − I(V ; Z|U ),
(14c)
RE ≤ I(U, V ; Y ) − I(V ; Z|U ) − I(U ; S),
(14d)
for some PMF qU,V,X|S that induces a joint distribution WS qU,V,X|S WY,Z|S,X is achievable.
To prove this inner bound we follow the derivation from Section VI, while replacing the message M therein
with a pair of uniformly distributed messages M̃ , (M, M ′ ) of rates RE and R − RE , respectively; the total
rate of communication is R. To ensure that the distribution approximation arguments from Lemma 1 and the error
probability analysis hold, it suffices that Inequalities (36) and (41) all hold with R = RE + (R − RE ) in the role of
RM . To satisfy the equivocation requirement, the security analysis is only carried out with respect to M . Therefore,
Inequality (54) is replaced by
R2 + R − RE > I(V ; Z|U ).
(15)
We conclude by noting that securing M implies the desired equivocation for M̃ :
1
1
1
1
1
H(M̃ |Z n ) ≥ H(M |Z n ) = H(M ) − I(M ; Z n ) ≥ H(M ) − ǫ,
n
n
n
n
n
(16)
where the multi-letter MI and entropy terms above are taken with respect to the distribution induced by the extracted
10
PSfrag replacements
WLn
L
K
M
Y
X
Encoder fn
Decoder φn
(M̂, K̂)
n
WY,Z|S,X
Z
Eavesdropper M, K
S
WSn
Fig. 2. The SD less-noisy-eavesdropper WTC with a key.
(reliable and secure) sequence of codes. Applying Fourier-Motzkin elimination to remove R1 and R2 produces (14).
IV. T IGHT S ECRECY C APACITY R ESULTS
An operationally appealing special case of the considered SD-WTC is the following. Assume that WY,Z|S,X is
such that the eavesdropper’s channel is less noisy than the main channel, but that the legitimate parties share a SK
L ∼ WLn (independent of the state sequence S ∼ WSn ), using which they secure the confidential data. The setup is
illustrated in Fig. 2.
Formally, let L, S, X , Y and Z be the alphabets of the key, the state, the channel input and the two channel
outputs, respectively. The considered instance is the S̃, X , Ỹ, Z, WS̃ , WỸ,Z|X,S̃ SD-WTC with S̃ = L × S,
Ỹ = L × Y, WS̃ = WL × WS , S̃ = (L, S), Ỹ = (L′ , Y ), and whose channel transition matrix factors as
WỸ,Z|X,S̃ = W(L′ ,Y ),Z|X,(L,S) = 1{L′ =L} WY,Z|S,X ,
(17)
where WY,Z|S,X is such that Z is less noisy than Y . A less noisy Z means that I(U ; Y ) ≤ I(U ; Z) for any random
variable U for which U −− (S, X) −− (Y, Z) forms a Markov chain. We refer to this special case as the SD
less-noisy-eavesdropper WTC with a key.
Theorem 1 applies here since the above case is a certain instance of a SD-WTC with non-causal encoder CSI. As
subsequently shown, the obtained inner bound is tight, thus characterizing the SS SM-SK secrecy capacity region
of the SD less-noisy-eavesdropper WTC with a key. The following corollary states the result.
Corollary 1 (SM-SK Capacity Region) The SS SM-SK capacity region of the SD less-noisy-eavesdropper WTC
with a key is the set of all SM-SK rate pairs (RM , RK ) ∈ R2+ satisfying
RM ≤ max [I(U ; Y ) − I(U ; S)] ,
qU,X|S
RK + RM ≤ H(L),
(18a)
(18b)
11
where the MI terms in (18a) are with respect to the joint PMF WS qU,X|S WY |S,X .
The proof of Corollary 1 is relegated to Appendix A. Note that while (18a) bounds the total communication rate as
a function only of the communication channel, (18b) bounds the total secrecy rate depending solely on the secret
source.
A direct consequence of Corollary 1 is that when no SK is to be established between the legitimate parties, i.e.,
RK = 0, the best attainable SM rate is
CSM = min
max I(U ; Y ) − I(U ; S) , H(L) .
qU,X|S
(19)
A simple separation-based coding scheme achieves the secrecy capacity from (19). Namely, using a capacity
achieving error correction code, the channel is effectively converted into a reliable bit-pipe. Each of the legitimate
parties compresses L, which results in a uniform random variable. The latter is used to encrypt the SM via a
one-time pad. The encrypted message is then transmitted over the reliable bit-pipe. Therefore, The achievable SM
rate is equal to the minimum of the capacity of the channel maxqU,X|S I(U ; Y ) − I(U ; S) and the rate of the key
H(L).
While this scheme may seem very natural, to the best of our knowledge, none of the past achievability results
for the SD-WTC with non-causal CSI prior to [19] attain its performance. In Section V-A1, a special case of this
setup is used to demonstrate the improvement of our result over the previous benchmark achievable SM-SK region
for the SD-WTC from [24].
V. P REVIOUS R ESULTS
AS
S PECIAL C ASES
We compare the result of Theorem 1 to those from related past works. The previously best known inner bound
on the SM-SK trade-off region attainable over the considered SD-WTC is [24, Theorem 1]. The next subsection
restates this inner bound and shows that Theorem 1 can strictly outperform it. Afterwards, we provide a comparison
to the best past achievability results for only SM transmission [19] or only SK agreement [23]. The achievability
result from [19] captures the previous lower bounds on the secrecy capacity of the SD-WTC from [18], [28], [29].
The SK achievability results from [23] subsume previous lower bounds on the SK generation rate, such as [22],
[30]. Relating to one another these three benchmarks that we use to evaluate the performance of Theorem 1, we
note that while [19] recovers [24] when there is only a SM (RK = 0), [23] and [24] do not imply one another.
Remark 5 Another result on SK generation over SD-WTCs with non-causal CSI is found in [31]. Theorem 1 therein,
which seemingly attains higher SK rates than both schemes from [23] and than our inner bound, is incorrect. The
region suggested in [31, Theorem 1] does not account for the secrecy-rate-loss when the inner layer codeword
cannot be decoded on its own by the legitimate decoder, i.e., when I(U ; S) > I(U ; Y ). (See the second case in
Remark 3 for a further explanation.) For this reason, we chose [23] as a benchmark for the SK generation problem.
Following the steps of the proof of [31, Theorem 1], it appears that another constraint was assumed without
being explicitly stated. Following the notations from [31], the missing constraint seems to be
Cp + I(W ; Y̌ ) > I(W ; S),
(20)
12
which would assure decodability of the inner code layer by the legitimate receiver without relying on the outer
layer. Taking the additional constraint into consideration, our inner bound from Theorem 1 recovers the amended
Theorem 1 from [31] as follows.
We use (Ũ, Ṽ, X̃, S̃, Ỹ, Z̃) to denote the inner layer, the outer layer, the channel input, the encoder CSI, and
the observations of the legitimate receiver and the eavesdropper, respectively, in Theorem 1 of [31]. These were
originally denoted, respectively, by W , U , X, S, Y̌ and Ž. To adjust our model to that of [31], we identify
X = (X̃, Φ), Y = (Ỹ, Φ), Z = (Z̃, Φ), S = S̃ in Theorem 1, where Φ is the random variable representing the
input (and the outputs) of the public communication link. In order to comply with the rate restriction on the public
link from [31], we restrict the distribution of Φ to have H(Φ) ≤ CP . Finally, we set:
1) RM = 0.
2) Φ independent of (Ũ, Ṽ, X̃, S̃, Ỹ, Z̃) with maximal entropy, i.e., such that H(Φ) = CP .
3) U = (Ũ, Φ), V = (Ũ, Ṽ, Φ).
With respect to the above, substituting (U, V, X, Y, Z, S) into (9) and maximizing only over distributions that satisfy
I(U ; Y ) − I(U ; S) > 0 produces the amended version of [31, Theorem 1].
To conclude the discussion of [31, Theorem 1] (in its original form), Appendix B provides a specific example
that shows the rates from that achievability formula to be exceeding the SK capacity.
A. SM-SK Trade-off Region
The result of Theorem 1 recovers the previously best known achievable SM-SK trade-off region over the SD-WTC
with non-causal encoder CSI [24]. In [24, Theorem 1] the following region was established:
RPER ,
[
qU qV,X|U,S
RPER qU qV,X|U,S ,
(21a)
where, for any qU ∈ P(U) and qV,X|U,S : U × S → P(V × X ),
RPER qU × qV,X|U,S
RM ≤ I(U, V ; Y ) − I(U, V ; S),
, (RM , RK ) ∈ R2+
,
RM + RK ≤ I(V ; Y |U ) − I(V ; Z|U )
(21b)
and the MI terms are taken with respect to WS qU qV,X|U,S WY,Z|S,X , i.e., U and S are independent and (U, V ) −
− (S, X) −− (Y, Z) forms a Markov chain.
First note that Theorem 1 recovers RPER by restricting U to be independent of S in RA . This is since for an
independent pair (U, S), we have I(U ; S) = 0, while I(U, V ; Y ) ≥ I(V ; Y |U ) always holds. Consequently, the
third rate bound in RA becomes redundant and RPER is recovered.
The result from [24] was derived under the weak secrecy metric (i.e., a vanishing normalized MI
1
n I(M, K; Z)
between the SM-SK pair and the eavesdropper’s observation sequence, where the message is assumed to be uniform).
Our achievability, on the other hand, ensures SS. Theorem 1, therefore, improves upon [24, Theorem 1] both in the
rates it achieves and in the sense of security it provides.
PSfrag replacements
13
WLn
L
M
Encoder
S
X
G
Memory
with
Stuck At
Faults
Y
BEC(ǫ)
S
X
S
WSn
Decoder
M̂
Eavesdropper
Fig. 3. Section V-A1 example setup.
1) Achieving Strictly Higher Rates: Since [24, Theorem 1] allows only inner layer coding random variables U
that are independent of the state, Gelfand-Pinsker coding [17], which generally requires correlating U with S, is
not supported in the inner layer. Instead, only Shannon’s Strategies coding [32], which operates with independent
U and S is allowed. The latter is optimal if the encoder observes the state causally, but is generally sub-optimal
when non-causal encoder CSI is available. To demonstrate the improvement of Theorem 1 over [24] we exploit the
aforementioned limitation of the scheme therein, along with the observation that it is beneficial to exploit any part
of a considered SD-WTC that is better observable by the eavesdropper to transmit the inner layer of the code.
Let X = G = L = E = {0, 1}, S = {0, 1, 2}, Y = {0, 1, ?}, where ? ∈
/ {0, 1} and Z = X × S. Consider the
SD less-noisy-eavesdropper WTC with a key (defined in Section IV) shown in Fig. 3, whose transition probability
WY,Z|S,X , key L ∼ WL and state S ∼ WS are defined by the three parameters λ, ǫ, σ ∈ (0, 0.5) as follows:
•
L, S and E are independent random variables with L ∼ Ber(λ), E ∼ Ber(ǫ) and
WS (0) = WS (1) =
σ
2
;
WS (2) = 1 − σ.
(22)
The joint distribution of (L, S, E) is denoted by WL,S,E = WL WS WE .
•
The Memory with Stuck-at-Faults (MSAF) [33] is a deterministic SD channel, driven by a ternary state S. The
binary input and output symbols X and G, respectively, are related through the function g : S × X → G given
by
g(s, x) =
•
s,
x,
s ∈ {0, 1}
.
(23)
s=2
The output of the MSAF channel is fed into a Binary Erasure Channel with erasure probability ǫ (abbreviated
as a BEC(ǫ)). The input G and the ternary output Y of the BEC(ǫ) are related by means of the erasure random
14
variable E through the function y : E × G → Y, where
g,
y(e, g) =
?,
•
e=0
.
(24)
e=1
Z = (S, X), i.e., the eavesdropper noiselessly observes the transmitted symbol X and the state random variable
S.
With respect to the above definitions, the transition matrix of the SD less-noisy-eavesdropper with channel
WY,Z|S,X is
WY,Z|S,X (y ′ , z|x, s) =
X
X
WE (e)WG,Y,Z|S,X,E (g ′ , y ′ , z|s, x, e),
(25a)
g′ ∈{0,1} e∈{0,1}
where
WG,Y,Z|S,X,E = 1{G=g(S,X)}∩{Y =y(E,G)}∩{Z=(S,X)} .
(25b)
A possible interpretation of this communication scenario is when the legitimate parties communicate through a
public database that has memory faults known to the transmitter, but not to the receiver. The database and the faults
are assumed to be known in full to the eavesdropper. To secure the communication the legitimate parties share a
SK.
For any λ, ǫ, σ ∈ (0, 0.5), we denote the secrecy-capacity of the corresponding channel by C(λ, ǫ, σ). Furthermore,
let RA (λ, ǫ, σ) and RPER (λ, ǫ, σ) denote the maximal achievable secrecy rates attained by (10) from Theorem 1
and (21b) from [24, Theorem 1], respectively. By virtue of Corollary 1 (and, more specifically, (19)), we have that
Theorem 1 is tight for the considered channel, i.e.,
C(λ, ǫ, σ) = RA (λ, ǫ, σ),
∀λ, ǫ, σ ∈ (0, 0.5).
(26)
As stated in the following proposition, RPER (λ, ǫ, σ) is strictly below capacity.
Proposition 1 There exist λ, ǫ, σ ∈ (0, 0.5) such that RA (λ, ǫ, σ) > RPER (λ, ǫ, σ).
Proposition 1 is proven in Appendix C. The proof relies on the observation that for RPER (λ, ǫ, σ), a full utilization
of the key L implies that RM is upper bounded by the capacity of the considered channel with causal CSI. In
turn, this capacity is further upper bounded by the capacity of the MSAF with causal CSI. Choosing the parameters
λ, ǫ, σ so that the secrecy-capacity of the channel is strictly above the causal MSAF capacity, the superiority of our
scheme compared to [24, Theorem 1] is established.
Remark 6 This example actually demonstrates that [19, Theorem 1] (which is a special case of Theorem 1, when
RK = 0) achieves strictly higher SM rates than [24, Theorem 1].
15
B. SM Transmission over SD-WTCs
In [19, Theorem 1] a lower bound was established on the SS-capacity of a SM transmission (i.e., when RK = 0)
(SM)
over the considered SD-WTC. The SS-capacity CSem of a SM transmission was lower bounded by
(SM)
CSem ≥ RGCP , max RGCP qU,V,X|S ,
(27a)
qU,V,X|S
where, for any qU,V,X|S : S → P(U × V × X ),
RGCP qU,V,X|S , min
I(U, V ; Y ) − I(U, V ; S),
,
I(V ; Y |U ) − I(V ; Z|U ),
I(U, V ; Y ) − I(V ; Z|U ) − I(U ; S)
(27b)
and the MI terms are taken with respect to WS qU,V,X|S WY,Z|S,X .
RGCP is the projection in the (RM , RK )-plane of RA from Theorem 1 to the RM axis when RK = 0. The main
difference between the coding scheme from [19] and our superposition code is the additional index k ∈ Kn in the
outer layer of the codebook (that also encodes the SM m ∈ Mn ). Along with the other redundancy indices, k is used
to correlate the transmission with the observed state sequence via the likelihood encoder [25]. Based on distribution
approximation arguments we show that K is approximately independent of the message M and approximately
uniform. The pair (M, K) is known to the transmitter and is reliably decoded by the receiver. Finally, by securing
K along with M in our analysis, it is established as a SK.
The intuition behind the SK construction is that, unlike the message, the key does not have to be independent of
the state sequence, nor is it chosen by the user. Therefore, the padding that ensures the correlation with the state
sequence is a valid key, as long as it is secured.
Observing that any portion of the SM can be allocated in favor of a SK implies that (27b) is also an achievable
SM-SK trade-off region, when RM above is replaced with RM + RK . RA outperforms RGCP , e.g., in settings where
an external random source L ∼ WLn is observed by both legitimate parties but not by the eavesdropper, while the
capacity of the communication channel is zero (say, Y = Z = 0). For such a setup, the legitimate parties may
use the random source to generate a SK of rate H(L). While Theorem 1 supports this strategy, RGCP nullifies in
this case. To see this, let S̃ , L and Ỹ , (L, Y ) = (L, 0) be the state and the channel output observed by the
legitimate receiver, respectively. Inserting S̃ and Ỹ into the first term inside the minimum from (27b) produces
I(U, V ; Ỹ ) − I(U, V ; S̃) = I(U, V ; Y ; L) − I(U, V ; L) = 0, for any qU,V,X|S̃ .
C. SK Agreement over SD-WTCs
In [23] two achievable schemes were proposed for SK agreement over a WTC when the terminals have access
to correlated sources. The results from [23] do not imply one another. The difference between them is that [23,
Theorem 2] is based on source and channel separation, while [23, Theorem 3] relies on joint coding.
The setup in [23] consists of three correlated sources Sx , Sy and Sz that are observed by the encoder, the decoder
and the eavesdropper, respectively, and a SD-WTC in which the triple (Sx , Sy , Sz ) plays the role of the state. Our
16
general framework is defined through the state distribution WS and the SD-WTC WỸ,Z̃|S,X . Setting S = Sx ,
Ỹ = (Sy , Y ) and Z̃ = (Sz , Z) recovers the model from [23] (see Remark 1).
The first scheme from [23, Theorem 2] operates under the assumption that the SD-WTC decomposes as
W(Sy ,Y ),(Sz ,Z)|Sx ,X = WSy ,Sz |Sx WY,Z|X into a product of two WTCs, one being independent of the state (given
the input), while the other one depends only on it. Thus, the legitimate receiver (respectively, the eavesdropper)
observes not only the output Y (respectively, Z) of the WTC WY,Z|X , but also Sy (respectively, Sz ) - a noisy
version of the state sequence drawn according to the corresponding conditional marginal of WSy ,Sz |Sx . This scheme
shows that the SK capacity CSK is lower bounded by
(Separate)
CSK ≥ RBPS
h
i
, max I(T ; Y |Q) − I(T ; Z|Q) + I(Ṽ ; Sy |Ũ ) − I(Ṽ ; Sz |Ũ ) ,
(28)
where the maximization is over all qṼ |Sx qŨ|Ṽ : Sx → P(Ṽ × Ũ) and qQ,T qX|T ∈ P(Q × T × X ) that give rise to a
joint PMF WSx ,Sy ,Sz qṼ |Sx qŨ|Ṽ ×qQ,T qX|T WY,Z|X satisfying I(Ũ ; Sx |Sy ) ≤ I(Q; Y ) and I(Ṽ ; Sx |Sy ) ≤ I(T ; Y ).
With respect to this distribution, (Sy , Sz ) −− Sx −− V −− U and Q −− T −− X −− (Y, Z) form Markov chains
and (Sy , Sz , Sx , V, U ) are independent of (Q, T, X, Y, Z). This independence is the essence of separation that uses
the channel for two purposes: carrying communication for SK agreement based on the sources, and securing part
of this communication using wiretap coding.
Setting RM = 0, U = (Q, Ũ ), V = (T, Ṽ ) in Theorem 1, and limiting the union to joint PMFs that satisfy
I(U ; Sy , Y ) ≥ I(U ; Sx ), while keeping the above distribution X, recovers (28).
The joint coding scheme from [23, Theorem 3] does rely on the aforementioned decomposition of the SD-WTC
W(Sy ,Y ),(Sz ,Z)|S,Xx . It lower bounds CSK as
(Joint)
CSK ≥ RBPS
h
i
, max I(Ṽ ; Sy , Y |Ũ ) − I(Ṽ ; Sz , Z|Ũ ) ,
(29)
where the maximization is over all qṼ,X|Sx qŨ|Ṽ : Sx → P(Ṽ × X × Ũ) that give rise to a joint PMF
WSx qṼ,X|Sx qŨ|Ṽ W(Sy ,Y ),(Sz ,Z)|Sx ,X satisfying I(Ũ ; Sx ) ≤ I(Ũ ; Sy , Y ) and I(Ṽ ; Sx |Ũ ) ≤ I(Ṽ ; Sy , Y |Ũ ). Setting
(Joint)
RM = 0 and (U, V ) = (Ũ, Ṽ ) in Theorem 1, where (Ũ, Ṽ ) is a valid auxiliary pair for RBPS , recovers (29).
It was shown in [23] that, in some cases, the separation-based scheme achieves strictly higher rates than the joint
(Separate)
coding scheme, i.e., that RBPS
(Joint)
> RBPS . As Theorem 1 captures both these results, it unifies the two schemes
(Joint)
from [23], and, in particular, it outperforms RBPS . Furthermore, since the results from [23] were derived under
the weak secrecy metric, Theorem 1 also upgrades them to SS.
VI. P ROOF
OF
T HEOREM 1
The subsequently presented proof follows lines similar to those from the proof of [19, Theorem 1]. Several claims
herein are recovered from corresponding assertions in [19] by identifying the index j in [19] with the pair (j, k)
in our scheme. The proofs of such claims are omitted, and the reader is referred to [19].
Fix ǫ > 0 and a conditional PMF qU,V,X|S : S → P(U ×V ×X ). For any n ∈ N, let pM ∈ P(Mn ) be the message
distribution. We first show that for any (RM , RK ) ∈ RA qU,V,X|S there exists a SS sequence of (n, RM , RK )codes with a key distribution that is approximately uniform conditioned on any message, and a vanishing average
17
error probability. We then use the expurgation technique [34, Theorem 7.7.1] to ensure a vanishing maximal error
probability. This is done without harming the SS and the statistical properties of the key, since they hold for each
message in the original message set.
Codebook C n : We use a superposition codebook where the outer layer carries both the SM and the SK. The
codebook is constructed independently of S, but has sufficient redundancy to enable correlating the transmission
with it.
(n)
Define the index sets In , 1 : 2nR1 and Jn , 1 : 2nR2 . Let BU , U(i)
codebook, which is a set of random vectors of length n that are i.i.d. according
(n)
denoted by BU , u(i) i∈In .
To describe the outer layer codebook, fix
(n)
BU
i∈In
n
to qU
.
be a random inner layer
∈
and, for every i
(n)
An outcome of BU
In
(n)
let BV (i)
is
,
be a collection of i.i.d. random vectors of length n with distribution qVn |U=u(i) .
(n)
(n)
(n)
For each i ∈ In , an outcome of BV (i) given BU is denoted by BV (i) , v(i, j, k, m) (j,k,m)∈J ×K ×M .
n
n
n
We also set BV = BV (i) i∈In and denote its realizations by BV . Finally, a random superposition codebook is
n
o
n
o
(n)
(n)
(n)
(n)
given by Bn = BU , BV , while Bn = BU , BV
denotes a fixed codebook.
V(i, j, k, m)
(j,k,m)∈Jn ×Kn ×Mn
Let Bn be the set of all possible outcomes of Bn . The above codebook construction induces a PMF µ ∈ P(Bn )
over the codebook ensemble. For every Bn ∈ Bn , we have
µ(Bn ) =
Y
i∈Ib
n
u(i)
qU
Y
î,j,k,m
∈In ×Jn ×Kn ×Mm
qVn |U v î, j, k, m u(î) .
(30)
The encoder and decoder are described next for any superposition codebook Bn ∈ Bn .
(Bn )
Encoder fn
: The encoding function is based on the likelihood-encoder [25], which, in turn, allows us to
approximate the induced joint distribution by a simple distribution that we use for the analysis. Given m ∈ Mn
and s ∈ S n , the encoder randomly chooses (i, j, k) ∈ In × Jn × Kn according to
n
s u(i), v(i, j, k, m)
qS|U,V
(Bn )
,
P
pLE (i, j, k|m, s) =
n
qS|U,V
s u(i′ ), v(i′ , j ′ , k ′ , m)
(31)
(i′ ,j ′ ,k′ )
∈In ×Jn ×Kn
where qS|U,V is the conditional marginal of qS,U,V defined by qS,U,V (s, u, v) =
P
x∈X
WS (s)qU,V,X|S (u, v, x|s),
for every (s, u, v) ∈ S × U × V. The encoder declares the chosen index k ∈ Kn as the key. The channel input
sequence is generated by feeding the chosen u- and v-codewords along with the state sequence into the DM channel
n
n
qX|U,V,S
, i.e., it is sampled from the random vector X ∼ qX|U=u(i),V
=v(i,j,k,m),S=s .
Accordingly, the (stochastic) encoding function fn : Mn × S n → P(Kn × X n ) is given by
X
fn(Bn ) (k, x|m, s) =
(i,j)∈In ×Jn
(Bn )
Decoder φn
(B )
n
x u(i), v(i, j, k, m), s .
pLEn (i, j, k|m, s)qX|U,V,S
(32)
: Upon observing y ∈ Y n , the decoder searches for a unique tuple (î, ĵ, k̂, m̂) ∈ In × Jn × Kn ×
Mn such that
u(î), v(î, ĵ, k̂, m̂), y ∈ Tǫn (qU,V,Y ).
(33)
18
(Bn )
If such a unique quadruple is found, then set φn
(Bn )
The quadruple (Mn , Kn , fn
(Bn )
, φn
(B )
(y) = m̂, k̂ ; otherwise, φn n (y) = (1, 1).
) defined with respect to the codebook Bn is an (n, RM , RK )-code cn .
For any message distribution pM ∈ P(Mn ) and codebook Bn ∈ Bn , the induced joint distribution p(Bn ) over
Mn × S n × In × Jn × Kn × U n × V n × X n × Y n × Z n × M̂n × K̂n is
(B )
p(Bn ) (m, s, i, j, k, u, v, x, y, z, m̂, k̂) =pM (m)WSn (s)pLEn (i, j, k|m, s)
× 1
n
qX|U,V,S
(x|u, v, s)
u=u(i) ∩ v=v(i,j,k,m)
n
× WY,Z|S,X
(y, z|x, s)1
(U)
.
n)
(y)
(m̂,k̂)=φ(B
n
(34)
If pM = pMn , i.e., the message distribution is uniform, we write p̄(Bn ) instead of p(Bn ) .
Approximating Distribution: We now show that with high probability p(Bn ) is close in total variation to another
distribution π (Bn ) , which lends itself for simpler reliability and security analyses. For any pM ∈ P(Mn ) and
Bn ∈ Bn , π (Bn ) is
π (Bn ) (m, i, j, k, u, v, s, x, y, z, m̂, k̂) = pM (m)
1
1
|In ||Jn ||Kn | u=u(i),v=v(i,j,k,m)
n
n
n
(y, z|x, s)1
(x|u, v, s)WY,Z|S,X
(s|u, v)qX|U,V,S
× qS|U,V
n)
(y)
(m̂,k̂)=φ(B
n
.
(35)
(U)
As before, π̄ (Bn ) stands for π (Bn ) when pM = pMn .
The following lemma states sufficient conditions for π (Bn ) to be a good approximation (in total variation) of
p(Bn ) with double-exponential certainty.
Lemma 1 (Sufficient Conditions for Approximation) If
R1 > I(U ; S),
(36a)
R1 + R2 + RK > I(U, V ; S),
(36b)
then there exist α1 , α2 > 0, such that for any n large enough
(Bn )
(Bn )
− πM,S,I,J,K,U,V,X,Y,Z,
pM,S,I,J,K,U,V,X,Y,Z,
max
Pµ
M̂,K̂
M̂,K̂
pM ∈P(Mn )
TV
>e
−nα1
nα2
≤ e−e
.
(37)
In particular, for any such n it also holds that
Eµ
(Bn )
p̄M,S,I,J,K,U,V,X,Y,Z,
M̂,K̂
−
(Bn )
π̄M,S,I,J,K,U,V,X,Y,Z,
M̂,K̂
TV
≤e
−nα1
+ n log
1
ξS
nα2
e−e
,
(38)
where ξS = mins∈supp(WS ) WS (s) > 0. The subscript µ in Pµ and Eµ indicates that the probability measure and
the expectation are taken with respect to the random codebook Bn ∼ µ.
Lemma 1 essentially restates [19, Lemma 7] with the index j therein replaced here with the pair (j, k). The
proof of Lemma 1 relies on the strong SCL for superposition codes and some basic properties of total variation.
Due to the similarity to [19, Lemma 7] we omit the proof and the reader is referred to [19].
19
Lemma 1 is key for analyzing the performance of the proposed code. The reliability analysis that is presented
next exploits the convergence of the expected value from (38) to show that the average error probability can be
made arbitrarily small. The expurgation method [34, Theorem 7.7.1] is used in a later stage of this proof to upgrade
to a vanishing maximal error probability.
Average Error Probability Analysis: The average error probability3 ē(Bn ) associated with a codebook Bn is
ē(Bn ) ,
X
1
em (Bn ) = PP̄ (Bn ) M̂, K̂ 6= (M, K) .
|Mn |
(39)
m∈Mn
Our next step is to establish that the expected value of ē(Bn ) over the codebook ensemble is approximately the
same under p̄ and π̄. Then, the expected average error probability under π̄ is analyzed and shown to converge to
zero as n → ∞. Due to the simple structure of π̄, this analysis requires nothing but standard typicality arguments.
To do so we use the two following lemmas.
Lemma 2 (Average Error Probability Under p̄ and π̄) The following relation holds:
Eµ Pp̄(Bn )
where Eµ p̄(Bn ) − π̄ (Bn )
M̂, K̂ =
6 (M, K) −Eµ Pπ̄(Bn ) M̂, K̂ 6= (M, K)
≤ Eµ p̄(Bn ) − π̄ (Bn )
(B )
TV
TV
,
(40)
(B )
n
n
is a shorthand for Eµ p̄M,S,I,J,K,U,V,X,Y,Z,
− π̄M,S,I,J,K,U,V,X,Y,Z,
M̂
M̂
TV
.
Lemma 3 (Average Error Probability Under π̄) If the rate tuple (RM , RK , R1 , R2 ) satisfies
RM + RK + R2 < I(V ; Y |U ),
(41a)
RM + RK + R1 + R2 < I(U, V ; Y ),
(41b)
then
Eµ Pπ̄(Bn )
M̂, K̂ =
6 (M, K) −−−−→ 0.
(42)
n→∞
The proof of Lemma 2 is found in the Average Error Probability Analysis part of Section VI-B in [19]. Lemma
3 is also proven in the same reference by standard typicality decoding arguments. We stress that the conditions in
(41) ensure reliable decoding of the four indices (i, j, k, m), and, in particular, of the SM-SK pair (m, k).
Combining the claims of Lemmas 2-3 with (38) from Lemma 1, we have that as long as (41) and (36) are
satisfied
Eµ ē(Bn ) −−−−→ 0.
(43)
n→∞
(B )
n
=
Key Analysis: The structure of π (Bn ) from (35) implies that for any Bn ∈ Bn and m ∈ Mn we have πK|M=m
(U)
pKn . Adopting the same abuse of notation we used for the reliability analysis, we use Lemma 1 to upper bound
3 We slightly abuse notation here because ē and e
m are actually functions of the code cn rather than the codebook Bn . However, since Bn
uniquely defines cn we prefer this presentation for the sake of simplicity.
20
the probability that δ(Bn ) does not decay exponentially fast to zero as n grows. Therefore, assuming (36) holds,
we have that there exists η1 , η2 > such that
(U)
(Bn )
> e−nη1
− pK n
Pµ δ(Bn ) > e−nη1 = Pµ max pK|M=m
m∈Mn
TV
(Bn )
(Bn )
−nη1
>e
= Pµ max pK|M=m − πK|M=m
m∈Mn
TV
(Bn )
(Bn )
−nη1
>e
pM,K − πM,K
max
≤ Pµ
TV
pM ∈P(Mn )
(a)
nη2
≤ e−e
,
(44)
where (a) is by (37) from Lemma 1. We proceed with the security analysis.
Security Analysis: This part mainly deals with analyzing the SS metric under the distribution π (Bn ) . The following
lemma explains the reason for doing so. It states that if SS is attained for a codebook Bn ∈ Bn under π (Bn ) then
it is also attained under p(Bn ) .
Lemma 4 (SS for Induced vs. Approximating Distribution) Let Bn ∈ Bn and β1 > 0, such that for all pM ∈
P(Mn ) and n sufficiently large (independent of pM )
(B )
(B )
n
n
− pM πK,Z|M
pM pK,Z|M
TV
≤ e−nβ1 .
(45)
Then, there exist β2 > 0 such that for all pM ∈ P(Mn ) and large enough values of n (independent of pM ), we
have
Ip(Bn ) (M, K; Z) − Iπ(Bn ) (M, K; Z) ≤ e−nβ2 ,
(46)
where the subscripts p(Bn ) and π (Bn ) indicate that a mutual information term is calculated with respect to the
corresponding PMF.
The proof of Lemma 4 extends that of [19, Lemma 8], and is provided in Appendix D.
The hypothesis from (45) essentially follows from Lemma 1. Thus, if (36) holds, then so does (46), on account
of which we have
ℓSem (Bn ) ,
max
PM ∈P(Mn )
Ip(Bn ) (M, K; Z) ≤
max
PM ∈P(Mn )
Iπ(Bn ) (M, K; Z) + e−nβ2 .
This implies that any codebook for which the RHS of (47) is small is SS.
Fix Bn and pM ∈ P(Mn ), and consider
Iπ(Bn ) (M, K; Z) ≤ Iπ(Bn ) (M, K; I, U, Z)
(Bn ) (Bn )
(Bn )
πZ,I,U
πM,K
= D πM,K,Z,I,U
(a)
(Bn ) (Bn ) (Bn )
(Bn ) (Bn ) (Bn )
πI,U πZ|I,U
πI,U πZ|M,K,I,U πM,K
= D πM,K
(b)
(Bn ) (Bn )
(Bn )
(Bn )
πM,K
πI,U
πZ|I,U
= D πZ|M,K,I,U
(47)
21
(c)
(B )
(Bn ) (Bn )
(Bn )
n
πI,U ,
πM,K
qZ|U
≤ D πZ|M,K,I,U
(B )
(48)
(B )
n
n
where (a) is because πM,K,I,U
= πM,K
πI,Un (see (35)), (b) is by the relative entropy chain rule, while (c) follows
from
(Bn ) (Bn )
(Bn )
(Bn ) (Bn )
(Bn )
(Bn ) (Bn )
(Bn )
(Bn )
n
n
πI,U
(49)
πM,K
qZ|U
πI,U − D πZ|I,U
πM,K
qZ|U
πI,U = D πZ|M,K,I,U
πM,K
πZ|I,U
D πZ|M,K,I,U
and the non-negativity of relative entropy. Here, qZ|U is the conditional marginal of the single-letter distribution
WS qU,V,X|S WY,Z|X,S .
Maximizing both sides of (48) over all message distributions pM ∈ P(Mn ), we further have
max
pM ∈P(Mn )
≤
=
≤
=
Iπ (M, K; Z)
max
pM ∈P(Mn )
(Bn ) (Bn )
(Bn )
n
πM,K
πI,U
qZ|U
D πZ|M,K,I,U
max
pM ∈P(Mn )
max
(B )
(Bn )
(Bn )
n
πI,Un
qZ|U
(m, k)D πZ|M=m,K=k,I,U
πM,K
X
n
(m, k)
πM,K
(m,k)∈Mn ×Kn
max
pM ∈P(Mn )
X
(B )
(m,k)∈Mn ×Kn
(m,k)∈Mn ×Kn
max
(m̃,k̃)∈M×Kn
(Bn )
(Bn )
n
π
q
D πZ|M=
Z|U
I,U
m̃,K=k̃,I,U
(B )
(Bn )
n
πI,Un .
qZ|U
D πZ|M=m,K=k,I,U
(50)
Inserting (50) into (47), we have that if (36) holds, then
ℓSem (Bn ) ≤
max
(m,k)∈Mn ×Kn
(B )
(Bn )
n
πI,Un + e−nβ2 .
qZ|U
D πZ|M=m,K=k,I,U
(51)
The two following lemmas state conditions under which the probability that the RHS of (51) vanishes
exponentially quickly with n is double-exponentially close to 1.
Lemma 5 (Total Variation Dominates Relative Entropy) Let X and Y be finite sets, and for any n ∈ N let
pX ∈ P(X n ), pY|X : X n → P(Y n ) and qY |X : X → P(Y). If pY|X=x ≪ qYn |X=x , for all x ∈ X n , i.e., pY|X=x
is absolutely continuous with respect to qYn |X=x , then
D pY|X
qYn |X
pX ≤
pX pY|X −pX qYn |X
TV
1
n log |Y| + log
pX pY|X − pX qYn |X
+ n log ξY |X
TV
!
, (52)
where ξY |X = min (x,y)∈X ×Y: qY |X (y|x).
qY |X (y|x)>0
(B )
n
n
≪ qZ|U=u
,
Lemma 5 is [19, Lemma 9] and its proof is omitted. It is readily verified that πZ|M=m,K=k,I=i,U=u
for each (m, i, k, u) ∈ Mn × In × Kn × U n . Combining Lemma 5 and (51), we see that if Bn ∈ Bn is a codebook
with rates satisfying (36), and there exists ζ1 > 0 such that
max
(m,k)∈Mn ×Kn
(B ) (B )
(B )
n
n
− πI,Un qZ|U
πI,Un πZ|M=m,K=k,I,U
TV
≤ e−nζ1 ,
(53a)
22
for any sufficiently large n, then there exists ζ2 > 0 for which
ℓSem (Bn ) ≤ e−nζ2
(53b)
as n grows.
Lemma 6 (Sufficient Conditions for SS) If the rate tuple (RM , RK , R1 , R2 ) ∈ R4+ satisfies (36a) and
R2 > I(V ; Z|U ),
(54)
then there exist γ1 , γ2 > 0, such that for n sufficiently large
Pµ
max
(m,k)∈Mn ×Kn
(B ) (Bn )
πI,Un πZ|M=m,K=k,I,U
−
(B ) n
πI,Un qZ|U
TV
>e
−nγ1
!
nγ2
≤ e−e
.
(55)
Lemma 6 follows by the security analysis from [19] with (M, K) = (m, k) in the role of M = m therein.
Combining the lemma with (53), we deduce that there exist τ1 , τ2 > 0 such that
nτ2
Pµ ℓSem (Bn ) > e−nτ1 ≤ e−e ,
(56)
for any sufficiently large n.
Code Extraction: The above derivation shows that if (36), (41) and (54) are simultaneously satisfied, then
EBn ē(Bn ) −−−−→ 0,
(57a)
nη2
Pµ δ(Bn ) > e−nη1 ≤ e−e ,
nτ2
Pµ ℓSem (Bn ) > e−nτ1 ≤ e−e .
(57b)
n→∞
and for sufficiently large n, we also have
(57c)
The Selection Lemma from [9, Lemma 5] implies the existence of a sequence of superposition codebooks
Bn n∈N (an outcome of the random codebook sequence Bn n∈N ), for which
ē(Bn ) −−−−→ 0,
(58a)
−−−−→ 0,
(58b)
−−−−→ 0.
(58c)
n→∞
1
δ(Bn )>e−nη1
1
ℓSem (Bn )>e−nτ1
n→∞
n→∞
Since the indicator functions in (58b)-(58c) take only the values 0 and 1, we have that for any n large enough
δ(Bn ) ≤ e−nη1 ,
(59a)
ℓSem (Bn ) ≤ e−nτ1 .
(59b)
23
On account of (57a) and (59), we have that {Bn }n∈N is SS, satisfies the target key statistics, and is reliable with
respect to the average error probability.
Our last step is to upgrade {Bn }n∈N to have a small maximal error probability. This is a standard step that uses
the expurgation technique (see, e.g., [34, Theorem 7.7.1]). Namely, pushing the average error probability below 2ǫ ,
at least half of the messages in Mn result in a probability of error that is at most ǫ. Throwing away the rest of the
messages ensures a maximal error probability that is at most ǫ, while inflicting a negligible rate loss. Discarding
those messages does not harm the SS or the key uniformity and independence metric, thus producing a new sequence
of codes that satisfies (8). Applying the Fourier-Motzkin Elimination on (36), (41) and (54) shows that any SM-SK
rate pair (RM , RK ) ∈ RA qU,V,X|S is achievable, which concludes the proof.
VII. S UMMARY
AND
C ONCLUDING R EMARKS
We studied the trade-off between the SM and SK rates that are simultaneously achievable over a SD-WTC with
non-causal encoder CSI. This model subsumes all other instances of CSI availability as special cases. An inner
bound on the SS message-key capacity region was derived based on a superposition coding scheme, the likelihood
encoder and soft-covering arguments inspired by [19].
We presented a class of SD-WTCs for which our inner bound achieves capacity, and demonstrated that for
this class, the previously best known SM-SK trade-off region by Prabhakaran et al. [24] is strictly sub-optimal.
Furthermore, we showed that the inner bound derived here recovers the best lower bounds on either the SM [19] or
the SK [23] rate achievable over the considered SD-WTC. Our derivations ensure SS, thus upgrading the security
standard from most of the past results, which were derived under the weak secrecy metric.
As the message-key capacity region for this setup remains an open problem, finding good outer bounds is of
particular interest. Extensions to multiple terminals, action dependent states [35], and source reconstruction models
should be examined as well.
A PPENDIX A
P ROOF
C OROLLARY 1
Recall that the SD less-noisy-eavesdropper WTC with a key is the S̃, X , Ỹ, Z, WS̃ , WỸ,Z|X,S̃ SD-WTC, where
OF
S̃ = L × S, Ỹ = L × Y, WS̃ = WL × WS , S̃ = (L, S), Ỹ = (L′ , Y ), whose transition matrix satisfies (17) and
the less-noisy condition.
A qU,X|S,L induces a joint distribution over L × S × U × X × Y × Z that is given by
qL,S,U,X,Y,Z , WL WS qU,X|S,L WY,Z|S,X .
(60)
We now proceed with the direct and the converse proofs.
Direct: Fix qU,X|S such that (U, X) −− S −− L. The structure of (60) further implies that (S, U, X, Y, Z) ⊥ L.
Evaluating the bounds from Theorem 1 with respect to (60), while setting V = (L, U ) and using S̃ = (L, S) and
Ỹ = (L, Y ), we have
RM ≤ I(U, V ; Ỹ ) − I(U, V ; S̃)
24
= I(L, U ; L, Y ) − I(L, U ; L, S)
= I(U ; Y |L) − I(U ; S|L)
(a)
= I(U ; Y ) − I(U ; S),
(61a)
where (a) is because (S, U, Y ) are independent of L. Combining the two bounds on the sum RM + RK in one, we
further have
+
RK + RM ≤ I(V ; Ỹ |U ) − I(V ; Z|U ) − I(U ; S̃) − I(U ; Ỹ )
+
= I(L; L, Y |U ) − I(L; Z|U ) − I(U ; L, S) − I(U ; L, Y )
+
(a)
= H(L) − I(U ; S) − I(U ; Y ) ,
(61b)
where, similarly to the above, (a) is implied by the independence of (S, U, Y, Z) and L. Finally, due to (61a),
any joint distribution that produces a non-zero achievable region satisfies I(U ; Y ) − I(U ; S) ≥ 0; hence, the term
+
I(U ; S) − I(U ; Y ) from (61b) is zero. Maximizing over all qU,X|S concludes the proof.
Converse: To get (18a), notice that the secret communication rate of the setup cannot exceed the total reliable
communication rate. Therefore, an upper bound on the secrecy capacity is given by the GP channel capacity formula
[17]:
h
i
max I(U ; Ỹ ) − I(U ; S̃) ,
qU,X|S̃
(62)
where, for each qU,X|S̃ , the underlying joint PMF is qU,X|S̃ WỸ |X,S̃ , with S̃ = (L, S) and Ỹ = (L, Y ). We thus
have
RM ≤ max
qU,X|L,S
= max
qU,X|L,S
(a)
=
max
qU,X|L,S
≤ max
qU,X|L,S
≤ max
qL,U,X|S
(b)
I(U ; L, Y ) − I(U ; L, S)
I(U ; Y |L) − I(U ; S|L)
I(U ; Y |L) − I(L, U ; S)
I(L, U ; Y ) − I(L, U ; S)
I(L, U ; Y ) − I(L, U ; S)
= max I(U ; Y ) − I(U ; S) ,
qU,X|S
(63)
where (a) follows because L and S are independent (see (60)), while (b) follows by recasting (L, U ) as U .
For the bound on RM + RK from (18b), we enhance the channel by allowing the encoder to control both the
state S and the secret source L, yet constraining it to the original statistics WS × WL . The obtained channel is
equivalent to a WTC WỸ,Z̃|X̃ with input X̃ = (L, S, X) and outputs Ỹ = (L, Y ) and Z̃ = Z at the legitimate
receiver and the eavesdropper, respectively. Therefore, the channel secrecy capacity [5] is an upper bound on the
25
sum of rates4 ; thus,
RM + RK ≤ max I(U ; Ỹ ) − I(U ; Z̃) .
qU,X̃
(64)
Now, note that for any qU,X̃ = qU,(L,S,X) we have
I(U ; Ỹ ) − I(U ; Z̃) = I(U ; L, Y ) − I(U ; Z)
= I(U ; L|Y ) + I(U ; Y ) − I(U ; Z)
(a)
≤ I(U ; L|Y )
≤ H(L),
(65)
where (a) is because the less-noisy property of the channel implies (U ; Y ) < I(U ; Z). This concludes the converse
proof.
A PPENDIX B
D EMONSTRATION OF I NCORRECTNESS
IN
[31, T HEOREM 1]
We first restate [31, Theorem 1] through the notations of this work. This theorem stipulates the following lower
bound on the SK capacity CSK of the SD-WTC with non-causal encoder CSI5 :
h
i
CSK ≥ RZib , max I(V ; Y |U ) − I(V ; Z|U ) ,
(66a)
where the maximization is over all conditional PMFs qU|V : V → P(U) and qV,X|S : S → P(V × X ) satisfying
I(V ; Y ) ≥ I(V ; S).
(66b)
All the above MI terms are taken with respect to the appropriate marginals of WS qU|V qV,X|S WY,Z|S,X , where
U −− V −− (S, X) −− (Y, Z) forms a Markov chain.
Now consider the following setup.
•
Let A, B and Q be three i.i.d. Ber( 12 ) random variables. Also, set An , B n and Qn as three n-fold random
vectors whose coordinates are i.i.d. copies of A, B and Q, respectively.
•
For each i ∈ [1 : n], let Ti = t(Ai , Bi , Qi ), where t : {0, 1}3 → {0, 1} is the deterministic function
a, q = 0
.
t(a, b, q) =
b, q = 1
(67)
T n represents the output sequence of the deterministic and memoryless channel T = t(A, B, Q), when it is
fed by An , B n and Qn .
4
5
For the WTC without state, no additional secrecy may be extracted in the form of a key. [14, Channel Model]
[31, Theorem 1] also incorporates a public communication link into the setup. We restate the theorem assuming that the public communication
rate is zero.
26
•
Let fn be the stochastic encoder and Ψn be the sequence that fn produces and transmits over a private binary
bit-pipe to the legitimate receiver.
•
The encoder observes (An , B n ) non-causally and determines the binary bit-pipe transmission Ψn .
•
The decoder observes (Qn , T n , Ψn ).
•
The eavesdropper observes An ⊕n B n , where ⊕n stands for bit-wise addition modulo 2. (At each time instance
the eavesdropper observes Ai + Bi (mod 2).)
Thus, at each channel use i ∈ [1 : n], the encoder observes two fair coin tosses, Ai and Bi . The decoder observes
only one of them, namely Ti , chosen at random (using a third fair coin Qi ). The decoder knows which coin it
observes, but the encoder does not. There is a private bit-pipe from the encoder to the decoder, which enables the
transmission of a single noiseless bit each time the coins are flipped. The legitimate parties wish to agree upon a
key that is kept secret from the eavesdropper (who observes only the modulo 2 addition of the two coins, Ai ⊕ Bi ,
each time they are flipped).
Denoting the SK generated by the legitimate parties by Kn , the induced joint PMF of the system is
qAn ,B n ,Qn ,T n ,Ψn ,Kn (an , bn , q n , tn , ψ n , kn ) = fn (kn , ψ n |an , bn )
n h
i
Y
WA (ai )WB (bi )WQ (qi )1{Ti =t(ai ,bi ,qi )} .
i=1
(68)
To comply with our notations, we identify S = (A, B), X = Ψ, Y = (T, Q, Ψ) and Z = A ⊕ B, while also
denoting by Ỹ , (T, Q) the output-CSI pair observed by the decoder.
A valid choice of random variables for (66) is
1) Ψ ∼ Ber( 21 ) independent of (A, B, Q),
2) U = Z = A ⊕ B,
3) V = (A, B, Ψ),
which achieves RZib = 2. Hence, by showing that the SK capacity of the proposed setup is strictly less than 2, we
contradict the achievability of RZib from [31, Theorem 1] as the SK rate for this setup. We do so by showing that
the vanishing average error probability and the weak secrecy of the SK, used in the definition of achievability in
[31], cannot coexist in this setup while a SK rate of 2 is attained.
Consider a sequence of codes {cn }n∈N achieving RZib = 2 for the above setup. We have that there exists a
sequence {ǫn }, with limn→∞ ǫn = 0, such that
where:
H(Kn ) ≥ 2n − nǫn ,
(69a)
H(Ψn ) ≤ n,
(69b)
H(Kn |Ψn , Ỹ n ) ≤ nǫn ,
(69c)
I(Kn ; Ψn , Z n ) ≤ nǫn ,
(69d)
27
(69a) follows by the definition of SK rate achievability6.
(69b) is because the alphabet of Ψn is of size 2n and since a uniform distribution maximizes discrete entropy.
(69c) is Fano’s inequality, following the requirement of vanishing decoding error.
(69d) is the weak secrecy requirement.
Lemma 7 For the considered setup, the SK capacity is upper bounded by 2 bits per channel use,
CSK ≤ 2.
(70)
Lemma 7 follows because the considered setup, but without an eavesdropper (i.e., when Z = 0), falls within the
framework of the common randomness (CR) problem in Model i from [36].
Proof: Theorem 4.1 in [36] shows that the CR capacity is upper bounded by
CCR ≤ R + I(S; Ỹ ),
(71)
where R is the rate of the communication link between the transmitter and the receiver. Evaluating the RHS of (71)
with respect to the considered setup shows that it equals 2 (CR bits per channel use). This upper bound remains
valid when a security requirement is introduced, since it can only reduce the admissible rates.
Lemma 7 guarantees the existence of a sequence {ǫ′n }, with limn→∞ ǫ′n = 0, such that the following condition
may be added to the set (69):
H(Kn ) ≤ 2n + nǫ′n .
(72)
Another technical lemma we need is stated next. Its proof is relegated to Appendix E.
Lemma 8 If (69a)-(69c) hold, then
H(An , B n |Kn ) ≤ 4nǫn .
(73)
Now, combining (72) and (73), we have
H(Kn |An , B n ) = H(Kn )−H(An , B n )+H(An , B n |Kn ) ≤ 2n+nǫ′n −2n+H(An , B n |Kn ) ≤ (4ǫn +ǫ′n )n. (74)
Using (74) we can finally lower bound the conditional information leakage term I(Kn ; Ψn , Z n ). To do so, first
consider
H(Kn |Z n ) ≤ H(Kn , An , B n |Z n )
= H(An , B n |Z n ) + H(Kn |An , B n , Z n )
≤ H(An , B n |Z n ) + H(Kn |An , B n )
(a)
≤ H(An , B n |Z n ) + (4ǫn + ǫ′n )n
(b)
= H(An , B n ) − H(Z n ) + (4ǫn + ǫ′n )n
6 In
fact, apparently in [31] a strong requirement was used, not including the vanishing term.
28
(c)
= (1 + 4ǫn + ǫ′n )n,
(75)
where (a) uses (74), (b) follows by the chain rule and because Z n is deterministically defined by (An , B n ) and(c)
is since An , B n and Z n = An ⊕n B n are all i.i.d. Ber 21 sequences, and because An and B n are independent.
Having (75), we conclude with
(a)
I(Kn ; Z n , Ψn ) ≥ I(Kn ; Z n ) = H(Kn ) − H(Kn |Z n ) ≥ 2n − nǫn − (1 + 4ǫn + ǫ′n )n ≥ (1 − 5ǫn − ǫ′n )n, (76)
where (a) uses (69a) and (75). Evidently, (76) contradicts (69d).
A PPENDIX C
P ROOF
OF
P ROPOSITION 1
Fix σ ∈ (0, 0.5) and set
ǫ=
i
1 h σ
h
−σ ,
2
2
(77a)
λ = h−1 (1 − σ − ǫ),
(77b)
where h : [0, 1] → [0, 1] and h−1 : [0, 1] → [0, 0.5] are the binary entropy function and the inverse of its restriction
to [0, 0.5], respectively. It is readily verified that ǫ, λ ∈ (0, 0.5). By virtue of (26), the inner bound from Theorem
1 attains the SM capacity, which is given by (see (19))
CSM = min CGP (WY |S,X ), H(L) ,
(78)
where CGP (WY |S,X ) = maxqU,X|S I(U ; Y ) − I(U ; S) is the GP capacity of the SD channel WY |S,X with state
distribution WS . By the corollary to Theorem 2 from [37] we find that CGP (WY |S,X ) = (1 − σ)(1 − ǫ). As
H(L) = 1 − σ − ǫ < (1 − σ)(1 − ǫ), we obtain
CSM = H(L) = 1 − σ − ǫ = 1 −
σ i
1h
σ+h
,
2
2
(79)
and, therefore,
RA (λ, ǫ, σ) = 1 − σ − ǫ = 1 −
σ i
1h
σ+h
.
2
2
(80)
The achievability of (80) may also be verified directly from Theorem 1 by substituting U = G, V = (U, L) and
X ∼ Ber 21 independent of (S, L) into (9).
We now show that RPER (λ, ǫ, σ) < 1 − 12 σ + h σ2 . Fix a joint distribution to evaluate the region from (21b)
with RK = 0, and S and Y replaced with S̃ = (L, S), Ỹ = (L, Y ). This distribution factors as
qL,S,U,V,X,W,E,Y,Z , WL WS qU qV,X|U,S,L 1{G=g(S,X)} WE 1{Y =y(E,G)} 1{Z=(S,X)}
× 1{S̃=(L,S)}∩{Ỹ =(L,Y )} .
Note that the independence of (L, S) and U is a restriction on the feasible joint distributions in (21a).
(81)
29
Now, assume in contradiction that evaluating (21b) with respect to q produces a rate that is at least as high as
(80). Specifically, assume that
I(U, V ; Ỹ ) − I(U, V ; S̃) ≥ H(L)
(82a)
I(V ; Ỹ |U ) − I(V ; Z̃|U ) ≥ H(L).
(82b)
and
Consider the following upper bound on (82b).
I(V ; Ỹ |U ) − I(V ; Z̃|U ) = I(V ; L, Y |U ) − I(V ; S, X|U )
= I(V ; Y |U ) + I(V ; L|U, Y ) − I(V ; S, X|U )
= I(V ; U, Y ) + I(V ; L|U, Y ) − I(V ; U, S, X)
(a)
= I(V ; L|U, Y ) + I(V ; U, Y ) − I(V ; U, S, X, Y )
= I(V ; L|U, Y ) − I(V ; S, X|U, Y )
= H(L|U, Y ) − H(L|U, V, Y ) − I(V ; S, X|U, Y )
≤ H(L),
(83)
where (a) uses the Markov relation V −− (S, U, X) −− Y , which follows because Y = y E, g(S, X) and E is
independent of (S, U, V, X) under the distribution from (81).
On account of (82b), the single inequality from (83) must hold with equality. For this to happen, the following
argument must hold.
1) The conditioning is removed from the first (positive) term, i.e.,
H(L) = H(L|U, Y ).
(84)
This implies that L is independent of (U, Y ).
2) The second (negative) term is zero, i.e.,
(a)
0 = H(L|U, V, Y ) = H(L|U, V, Y, E)
= (1 − ǫ) · H(L|U, V, Y, E = 0) + ǫ · H(L|U, V, Y, E = 1),
(85)
where (a) is because E is deterministically defined by Y . Now, since ǫ > 0, we have that H(L|U, V, Y, E =
1) = 0. Observing that conditioned on {E = 1}, Y =? is a constant, we further deduce
(a)
H(L|U, V, Y, E = 1) = H(L|U, V, E = 1) = H(L|U, V ) = 0,
(86)
where (a) relies on the independence of E and (L, U, V ). The last equality in (86) implies that there exists
a (deterministic) function ℓ : U × V → L such that L = ℓ(U, V ).
30
3) Expanding the third (negative) term with respect to E in a similar manner to that presented in the above 2nd
point, we obtain
I(V ; S, X|U, Y, E = 1) = I(V ; S, X|U, E = 1) = I(V ; S, X|U ) = 0,
(87)
which establishes V −− U −− (S, X) as a Markov chain.
Since S and U are independent under q from (81), the Markov relation from point 3) further implies that S
is independent of the pair (U, V ). Observe that this effectively means that the inability of the scheme from [24,
Theorem 1] to support GP coding in the inner layer implies that GP coding is not supported at all.
We proceed to analyze (82a) under the above deductions. Consider
I(U, V ; Ỹ ) − I(U, V ; S̃) = I(U, V ; L, Y ) − I(U, V ; L, S)
= I(U, V ; Y |L) − I(U, V ; S|L)
≤ I(U, V, L; Y )
(a)
≤ I(U, V, L; G)
(b)
= I(U, V ; G),
(88)
where (a) follows by the Data Processing Inequality (see, e.g., [34, Section 2.8]) and since (L, U, V ) −− G −− Y
forms a Markov chain, while (b) is because L = ℓ(U, V ).
Define T = (U, V ) and observe that T is independent of S (since the pair (U, V ) is) and that T −− (S, X) −− G
forms a Markov chain (since G = g(S, X)). We further upper bound the RHS of (88) with T = (U, V ) by
maximizing it over all conditional distributions that satisfy qT,X|S = qT qX|S,T . We thus have
I(U, V ; Ỹ ) − I(U, V ; S̃) ≤ I(T ; G) ≤
max I(T ; G).
qT qX|S,T
(89)
The expression on the RHS of (89) is the capacity of the MSAF with causal encoder knowledge of the state
sequence (cf., e.g., [38, p.5469]). However, the causal CSI is useless for the MSAF encoder, as demonstrated in
Section V-A of [38]. Omitting the availability of any CSI from the MSAF encoder, the channel is equivalent to a
binary symmetric channel with flip probability σ2 (see (23)), whose capacity equals 1 − h σ2 .
We conclude with
I(U, V ; Ỹ ) − I(U, V ; S̃) ≤ max I(T ; G)
qT qX|T S
=1−h
σ
2
σ i
1h
σ+h
< 1−
2
2
(a)
= H(L),
where (a) is because σ < h( σ2 ) for any σ ∈ (0, 0.5). This is a contradiction to (82a).
(90)
31
A PPENDIX D
P ROOF
OF
L EMMA 4
Fix pM ∈ P(Mn ) and, for simplicity of notation, abbreviate p(Bn ) and π (Bn ) as p and π, respectively. Consider:
Ip (M, K; Z) − Iπ (M, K; Z) = Hp (M, K) + Hp (Z) − Hp (M, K, Z) − Hπ (M, K) − Hπ (Z) + Hπ (M, K, Z)
(a)
≤ Hp (M, K) − Hπ (M, K) + Hp (Z) − Hπ (Z) + Hp (M, K, Z) − Hπ (M, K, Z)
(b)
≤
pM,K − πM,K
+ pZ − πZ
TV
log
|Mn | · |Kn |
pM,K − πM,K
|Z n |
log
TV
pZ − πZ
+ pM,K,Z − πM,K,Z
TV
TV
|Mn | · |Kn | · |Z n |
pM,K,Z − πM,K,Z TV
log
TV
(c)
h
i
≤ e−nβ1 n(RM + RK ) + n log |Z| + n (RM + RK + log |Z|)
− pM,K − πM,K
− pZ − πZ
TV
TV
log pM,K − πM,K
log pZ − πZ
− pM,K,Z − πM,K,Z
TV
TV
TV
log pM,K,Z − πM,K,Z
TV
,
(91)
where (a) is due to the Triangle Inequality, (b) uses [34, Theorem 17.3.3], while (c) follows from the assumption
in (45).
i
h
1
Note that the function x 7→ −x log x is monotone increasing for x ∈ 0, 2− ln 2 and that there exists an
h
i
1
ñ1 ∈ N such that e−nβ1 ∈ 0, 2− ln 2 , for all n > ñ1 . Finally, since pM,K − πM,K TV and pZ − πZ TV and
pM,K,Z − πM,K,Z
TV
are all upper bounded by e−nβ1 , we have that for all n > ñ1
− pM,K − πM,K
TV
log pM,K − πM,K
TV
− pZ − πZ
− pM,K,Z − πM,K,Z
TV
TV
log pZ − πZ
log pM,K,Z − πM,K,Z
TV
TV
≤ −3e−nβ1 log e−nβ1 .
(92)
Plugging (92) into (91) gives
Ip (M ; Z) − Iπ (M ; Z) ≤ ne
−nβ1
1
2 log |Z| + RM + RK + 3β1
.
ln 2
(93)
The bound on the RHS of (93) is uniform in pM ∈ P(Mn ) and decays exponentially fast to zero as n grows. The
result of Lemma 4 follows by maximizing both sides on (93) over all message distributions.
A PPENDIX E
P ROOF
OF
L EMMA 8
Using (69a)-(69c) and some basic information identities, we have
n + nǫn ≥ H(Ψn ) + H(Kn |Ψn , Ỹ n )
32
≥ I(Kn ; Ψn |Ỹ n ) + H(Kn |Ψn , Ỹ n )
= H(Kn |Ỹ n )
= H(Kn ) − I(Kn ; Ỹ n )
≥ 2n − nǫn − I(Kn ; Ỹ n ),
(94)
I(Kn ; Ỹ n ) ≥ (1 − 2ǫn )n.
(95)
1
I(Kn ; Ỹ n ) ≤ n − H(An , B n |Kn ),
2
(96)
which produces
We next show that
which, when combined with (95) gives
1
n − 2nǫn ≤ I(Kn ; Ỹ n ) ≤ n − H(An , B n |Kn ).
2
(97)
This further implies (73), as required. Thus, to complete the proof of Lemma 8 it suffices to show that (96) holds.
Consider the following steps:
I(Kn ; Ỹ n ) =
n
X
I(Kn ; Ỹi |Ỹ i−1 )
i=1
=
n
X
I(Kn ; Ti , Qi |Ỹ i−1 )
i=1
n h
i
X
I(Kn ; Ti |Qi , Ỹ i−1 ) + I(Kn ; Qi |Ỹ i−1 )
=
(a)
=
=
=
(b)
=
i=1
n
X
I(Kn ; Ti |Qi , Ỹ i−1 )
i=1
n h
X
1
2
1
2
1
2
i
I(Kn ; Ti |Qi = 0, Ỹ i−1 ) + I(Kn ; Ti |Qi = 1, Ỹ i−1 )
i=1
n h
X
i
I(Kn ; Ai |Qi = 0, Ỹ i−1 ) + I(Kn ; Bi |Qi = 1, Ỹ i−1 )
i=1
n h
X
i=1
i
I(Kn ; Ai |Ỹ i−1 ) + I(Kn ; Bi |Ỹ i−1 )
n
i
1 Xh
=
H(Ai |Ỹ i−1 ) − H(Ai |Kn , Ỹ i−1 ) + H(Bi |Ỹ i−1 ) − H(Bi |Kn , Ỹ i−1 )
2 i=1
≤
n
i
1 Xh
H(Ai ) + H(Bi ) − H(Ai |Kn , Ỹ i−1 ) − H(Bi |Kn , Ỹ i−1 )
2 i=1
(c)
= n−
n
i
1 Xh
H(Ai |Kn , Ỹ i−1 ) + H(Bi |Kn , Ỹ i−1 )
2 i=1
n
≤n−
1X
H(Ai , Bi |Kn , Ỹ i−1 )
2 i=1
33
n
≤n−
(d)
= n−
1X
H(Ai , Bi |Kn , Qn , Ai−1 , B i−1 , Ỹ i−1 )
2 i=1
n
1X
H(Ai , Bi |Kn , Qn , Ai−1 , B i−1 )
2 i=1
1
= n − H(An , B n |Kn , Qn )
2
1
(e)
= n − H(An , B n |Kn ),
2
(98)
where (a) and (b) follow since Qi is independent of (Kn , Ai , Bi , Ỹ i−1 ), for every i ∈ [1 : n], as evident from (68),
(c) is because Ai , Bi ∼ Ber 21 , (d) follows since Ỹ i−1 is a deterministic function of (Qi−1 , Ai−1 , B i−1 ) and (e)
uses the independence of Qn and (Kn , An , B n ) (see (68)).
34
R EFERENCES
[1] M. Bloch and J. Barros. Physical-Layer Security: From Information Theory to Security Engineering. Cambridge Univ. Press, Cambridge,
UK, Oct. 2011.
[2] Y. Liu, H. H. Chen, and L. Wang. Physical layer security for next generation wireless networks: Theories, technologies, and challenges.
IEEE Commun. Surv. Tut., 19(1):347–376, First quarter 2017.
[3] K. Zeng. Physical layer key generation in wireless networks: challenges and opportunities. IEEE Commun. Mag., 53(6):33–39, June 2015.
[4] A. D. Wyner. The wire-tap channel. Bell Sys. Techn., 54(8):1355–1387, Oct. 1975.
[5] I. Csiszár and J. Körner. Broadcast channels with confidential messages. IEEE Trans. Inf. Theory, 24(3):339–348, May 1978.
[6] A. D. Wyner. The common information of two dependent random variables. IEEE Trans. Inf. Theory, 21(2):163–179, Mar. 1975.
[7] T. Han and S.Verdú. Approximation theory of output statistics. IEEE Trans. Inf. Theory, 39(3):752–772, May 1993.
[8] J. Hou and G. Kramer. Informational divergence approximations to product distributions. In Proc. 13th Canadian Workshop Inf. Theory
(CWIT), Toronto, Ontario, Canada, Jun. 2013.
[9] Z. Goldfeld, P. Cuff, and H. H. Permuter. Semantic-security capacity for wiretap channels of type II. IEEE Trans. Inf. Theory, 62(7):3863–
3879, Jul. 2016.
[10] Z. Goldfeld, P. Cuff, and H. H. Permuter. Arbitrarily varying wiretap channels with type constrained states. IEEE Trans. Inf. Theory,
62(12):7216–7244, Dec. 2016.
[11] M. Bastani Parizi, E. Telatar, and N. Merhav. Exact random coding secrecy exponents for the wiretap channel. IEEE Transactions on
Information Theory, 63(1):509–531, Jan 2017.
[12] S. Yagli and P. W. Cuff. In Preparation.
[13] U. M. Maurer. Secret key agreement by public discussion from common information. IEEE Trans. Inf. Theory, 39(3):733–742, May 1993.
[14] R. Ahlswede and I. Csiszár. Common randomness in information theory and cryptography. part i: secret sharing. IEEE Trans. Inf. Theory,
39(4):1121–1132, Jul. 1993.
[15] I. Csiszár and P. Narayan. Common randomness and secret key generation with a helper. IEEE Trans. Inf. Theory, 46(2):344–366, Aug.
2000.
[16] A. D. Wyner and J. Ziv. The rate-distortion function for source coding with side information at the decoder. IEEE Trans. Inf. Theory,
1:1–10, 1976.
[17] S. I. Gelfand and M. S. Pinsker. Coding for channel with random parameters. Problemy Pered. Inform. (Problems of Inf. Trans.), 9(1):19–31,
1980.
[18] Y. Chen and A. J. Han Vinck. Wiretap channel with side information. IEEE Trans. Inf. Theory, 54(1):395–402, Jan. 2008.
[19] Z. Goldfeld, P. Cuff, and H. H. Permuter. Wiretap channel with random states non-causally available at the encoder. Submitted to IEEE
Trans. Inf. Theory, 2016. Available on ArXiv at https://arxiv.org/abs/1608.00743.
[20] H. Fujita. On the secrecy capacity of wiretap channels with side information at the transmitter. IEEE Transactions on Information Forensics
and Security, 11(11):2441–2452, Nov 2016.
[21] T. S. Han and M. Sasaki. Wiretap channels with causal state information: Strong secrecy. ArXiv preprint, Aug 2017. Available at
https://arxiv.org/abs/1708.00422.
[22] A. Khisti, S. N. Diggavi, and G. W. Wornell. Secret-key agreement with channel state information at the transmitter. IEEE Trans. Inf.
Forensics Security, 6(3):672–681, Mar. 2011.
[23] G. Bassi, P. Piantanida, and S. Shamai (Shitz). Secret key generation over noisy channels with common randomness. ArXiv preprint, Sep.
2016. Available at https://arxiv.org/abs/1609.08330.
[24] V. Prabhakaran, K. Eswaran, and K. Ramchandran. Secrecy via sources and channels. IEEE Trans. Inf. Theory, 85(11):6747–6765, Nov.
2012.
[25] E. Song, P. Cuff, and V. Poor. The likelihood encoder for lossy compression. IEEE Trans. Inf. Theory, 62(4):1836–1849, Apr. 2016.
[26] M. Bellare, S. Tessaro, and A. Vardy. A cryptographic treatment of the wiretap channe. In Proc. Adv. Crypto. (CRYPTO 2012), Santa
Barbara, CA, USA, Aug. 2012.
[27] Y. Liang and H. V. Poor S. Shamai. Information theoretic security. Foundations and Trends R in Commun. and Inf. Theory, 5(4-5):355–580,
2009.
[28] W. Liu and B. Chen. Wiretap channel with two-sided state information. In Proc. 41st Asilomar Conf. Signals, Syst. Comp, page 893897,
Pacific Grove, CA, US, Nov. 2007.
35
[29] Y.-K. Chia and A. El Gamal. Wiretap channel with causal state information. IEEE Trans. Inf. Theory, 58(5):2838–2849, May 2012.
[30] A. Khisti, S. N. Diggavi, and G. W. Wornell. Secret-key generation using correlated sources and channels. IEEE Trans. Inf. Theory,
58(2):652–670, Feb. 2012.
[31] A. Zibaeenejad. Key generation over wiretap models with non-causal side information. Information Forensics and Security, IEEE
Transactions on, 10(7):1456–1471, July 2015.
[32] C. E. Shannon. Channels with side information at the transmitter. IBM J. Res. Devel., 2(4):289–293, Oct. 1958.
[33] A. V. Kuznetsov and B. S. Tsybakov. Coding in a memory with defective cells. Problemy Pered. Inform. (Problems of Inf. Trans.),
10(2):52–60, 1974.
[34] T. M. Cover and J. A. Thomas. Elements of Information Theory. Wiley, New-York, 2nd edition, 2006.
[35] B. Dai, A. J. Han Vinck, Y. Luo, and X. Tang. Wiretap channel with action-dependent channel state information. Entropy, 15:445–473,
2013.
[36] R. Ahlswede and I. Csiszár. Common randomness in information theory and cryptography. ii. cr capacity. Information Theory, IEEE
Transactions on, 44(1):225–240, 1998.
[37] C. Heegaard and A. El Gamal. On the capasity of computer memories with defects. IEEE Trans. Inf. Theory, IT-29:731–739, Sept. 1983.
[38] S. A. Jafar. Channel capacity with causal and noncaudal side information - a unified view. IEEE Trans. Inform. Theory, 52(12):5468–5474,
Dec. 2006.
| 7cs.IT
|
Datalog: Bag Semantics via Set Semantics
∗
Leopoldo Bertossi1 , Georg Gottlob2,3 and Reinhard Pichler3
arXiv:1803.06445v1 [cs.DB] 17 Mar 2018
1
Carleton University, Ottawa, [email protected]
2
University of Oxford, [email protected]
3
TU Wien, [email protected]
ABSTRACT
Duplicates in data management are common and problematic. In this work, we present a translation of Datalog under
bag semantics into a well-behaved extension of Datalog (the
so-called warded Datalog± ) under set semantics. From a
theoretical point of view, this allows us to reason on bag
semantics by making use of the well-established theoretical
foundations of set semantics. From a practical point of view,
this allows us to handle the bag semantics of Datalog by
powerful, existing query engines for the required extension
of Datalog. Moreover, this translation has the potential for
further extensions – above all to capture the bag semantics
of the semantic web query language SPARQL.
1.
INTRODUCTION
Duplicates are a common feature in data management. They appear, for instance, in relational databases queried by means of SQL and RDF data queried by
means of SPARQL. However, the semantics of data operations and queries in the presence of duplicates is not
always clear, mostly related to the fact that duplicates
are handled by bags or multisets, whereas the common
logic-based semantics used in data management are settheoretical, making it difficult to tell apart duplicates
through the use of sets alone.
To address this problem, a bag semantics for Datalog programs was proposed in [9], what we call the
derivation-tree bag semantics (DTB semantics). Intuitively, two duplicates of the same tuple in an intentional predicate are accepted as such, if they have syntactically different derivation trees. Also, an equivalent
formulation was given in terms of semi-naive evaluation. The DTB semantics was used in [1] to provide a
bag semantics for SPARQL.
The DTB semantics has two major drawbacks: first,
it is operational – thus losing the declarative, logicbased semantics of Datalog; and second, it is defined
via new constructs (the DTBs) – thus leaving the world
of query languages and losing the applicability of the
large body of query optimization techniques.
The goal of this paper, instead, is to identify an
extension of Datalog which allows us to express the
bag semantics in terms of the classical set semantics,
and stay within the realm of query languages. To this
∗
L. Bertossi is member of the “Millenium Research Institute
on Foundations of Data”, Chile.
end, we show that the DTB semantics of a Datalog program can be represented by means of its transformation
into a Datalog± program [4], in such a way that the intended model of the former, including duplicates, can
be characterized as the result of the duplicate-free chase
instance for the latter. This is achieved by creating
the right tuple identifiers (tids) by means of existential
rules. Duplicates with different tids of the same tuple will be admissible and usual duplicates when falling
back to a bag semantics for the original Datalog program. We establish the correspondence between the
DTB semantics and ours.
The Datalog± programs required for this task belong
to the well-behaved class of warded Datalog± programs
[6], which properly extends Datalog, has a tractable
(conjunctive) query answering (CQA) problem, and has
been successfully applied to represent SPARQL under
the OWL 2 QL core entailment regime [5], with set
semantics though [6] (see also [2]). Warded Datalog±
looks promising as a general language for specifying different data management tasks [3].
We also show that Datalog with stratified negation
can be captured by a similar transformation into a wellbehaved class of Datalog±. In this way, we achieve a
fully declarative way of expressing the bag semantics
of an important query language; and we immediately
recover full relational algebra (including set difference)
with bag semantics in terms of a well-behaved query
language under set semantics. Moreover, the translation into Datalog± ensures polynomial-time CQA.
2. PRELIMINARIES
We assume the reader to be familiar with the relational data model. An n-ary predicate P has positions:
P [1], . . . , P [n]. With Pos(P ) we denote the set of positions of predicate P . Similarly, Pos(Π) denotes the set
of positions of (predicates in) a Datalog program Π.
Basic Multiset Operations. We follow [8, 9]. Consider multisets M and elements e (from some domain)
with non-negative integer multiplicities, mult(e, M ). By
definition, e ∈ M iff mult(e, M ) ≥ 1. To a multiset M ,
we associate a set: set (M ) := {e | e ∈ M }. Now consider multiset relations R, S. Unless stated otherwise,
we assume that R, S contain tuples of the same arity,
say n. Multiset R is (multi-)contained in S, denoted,
R ⊆m S iff for every e ∈ R, mult(e, R) ≤ mult(e, S).
mgu θ, generate a DT for Hθ with Hθ as the root
label and hτ1 , . . . , τk i as the children, and r as the
label for the edges from the root to the children.
Moreover, we define the following multiset operations:
The multiset union ∪m is defined by R ∪m S := T ,
with mult(e, T ) := mult(e, R) + mult(e, S). Multiset
m
selection σC
(R), with C a condition, is defined as the
multiset T containing all tuples in R that satisfy C with
multiplicities the same as in R.
The multiset projection πk̄m (R) is defined as follows:
let k̄ be a k-tuple hi1 , . . . , ik i of elements from {1, . . . , n};
accordingly, for an n-tuple t̄ = ht1 , . . . , tn i ∈ R, we
consider the k-tuple hti1 , . . . , tik i. Now, πk̄m (R) is the
multiset T containing k-tuples e = hti1 , . . . , tik i, with
mult(e, πk̄m (R)) defined as the sum of the multiplicities
in R of tuples t̄ producing e.
For the multiset (natural) join assume that tuples
have arity n in R and arity n′ in S. To simplify the presentation, assume that the natural join is via the last attribute of R and first of S. Then we define the following
multiset of (n + n′ − 1)-tuples: t̄ = ht1 , . . . , tn+n′ −1 i ∈
R ✶m S iff there are r̄ ∈ R and s̄ ∈ S, such that
rn = s1 , t̄|[1,...,n] = r̄ and t̄|[n,...,n+n′ −1] = s̄, with
X
mult(r̄, R)×mult(s̄, S).
mult(t̄, R ✶m S) =
r̄,s̄
For a DT τ , we define Atoms(τ ) := root-label or col −1
of the root, with the latter case when τ S
is a single-node
tree. For a set of DTs T: Atoms(T) := m {Atoms(τ ) |
τ ∈ T}, which is a multiset that multi-contains E.
If DT (Π, col (E)) is the set of (syntactically different)
DTs, the derivation-tree bag (DTB) semantics for Π is
the multiset:
DTBS (Π, E) := Atoms(DT (Π, col (E))).
(2)
Example 2.2. Consider program Π below and E =
{Q(1, 2, 3), Q(1, 2, 3), Q(1, 2, 5), T (4, 1, 2), T (4, 1, 2)}.
r1 : P (x, y) ← R(x, y), S(x, y)
r2 : R(x, y) ← Q(x, y, z)
r3 : S(x, y) ← T (z, x, y).
Here, col (E) = {Q(1, 2, 3) : 1, Q(1, 2, 3) : 2, Q(1, 2, 5) :
1, T (4, 1, 2):1, T (4, 1, 2):2}. We have the following DTs:
1. Five single-node trees: Q(1, 2, 3):1, etc.
2. Five depth-two trees: R(1, 2) − r2 − Q(1, 2, 3):1,
R(1, 2)−r2 −Q(1, 2, 3):2, R(1, 2)−r2 −Q(1, 2, 5):1,
S(1, 2)−r3 −T (4, 1, 2):1, S(1, 2)−r3 −T (4, 1, 2):2.
3. Six depth-three trees: etc.
In total, DT (Π, col (E)) contains 16 different trees,
and DTBS (Π, E) = E ∪ {R(1, 2), R(1, 2), R(1, 2), S(1, 2),
S(1, 2), P (1, 2), P (1, 2), P (1, 2), P (1, 2), P (1, 2), P (1, 2)}.
s.t. rn = s1
For the multiset difference, two definitions are considered: Rrm S := T , with mult(e, T ) := max {mult(e, R)−
mult(e, S), 0}. Alternatively, R ran S := T , with
mult(e, T ) := mult(e, R) if e ∈
/ S and mult (e, T ) := 0,
otherwise (“an” stands for all-or-nothing).
The multiset intersection ∩m is a contentious operation. Extending the derivation-tree-based semantics
from [9] to multiset intersection would treat ∩m as a
special case of the join, which may be counter-intuitive.
However, intersection is not treated or used in [1, 8, 9].
3. DATALOG± -BASED BAG SEMANTICS
We now provide a bag semantics for a Datalog program Π with a multiset EDB E via the transformation
into a Datalog± program Π+ over a set EDB E + obtained from E. For this, we assume w.l.o.g., that the set
of nulls for a Datalog± program is partitioned into two
infinite ordered sets T = {ι1 , ι2 , . . .}, for unique, global
tuple identifiers (tids), and N = {η1 , η2 , . . .}, for usual
nulls in Datalog± programs. Given a multiset EDB E
and a program Π, instead of using colors and syntactically different derivation trees, we will use elements of T
to identify both the elements of the EDB and the tuples
resulting from applications of the chase procedure.
For a predicate P (. . .) in the program’s schema, we
introduce a new version P (. ; . . .) with an extra, first argument (its 0-th position) to accommodate a tid, which
is a null from T . If an atom P (ā) appears in E as n
duplicates, we create the tuples P (ι′1 ; ā), . . . , P (ι′n ; ā),
with the ι′i pairwise different nulls from T as tids, and
not used to identify any other element of E. We obtain
a set EDB E + from the multiset EBD E.
Given a rule in Π, we introduce tid-variables (i.e. appearing in the 0-th positions) and existential quantifiers in the rule head, to formally generate fresh tids
when the rule applies. More precisely, a rule in Π of the
form r : H(x̄) ← A1 (x̄1 ), A2 (x̄2 ), . . . , Ak (x̄k ), with
k > 0, x̄ ⊆ ∪i x̄i , becomes the Datalog± rule:
r+ : ∃z H(z; x̄) ← A1 (z1 ; x̄1 ), A2 (z2 ; x̄2 ), . . . , Ak (zk ; x̄k ),
with fresh, different variables z, z1 , . . . , zk .
Derivation-tree bag semantics of Datalog. We
follow [9], where tuples are “colored” to tell apart duplicates of a same element in the extensional database
(EDB), via an infinite, ordered list C of colors c1 , c2 , . . ..
For a multiset M and e ∈ M , with mult(e, M ) = n > 0,
the n copies, e:1, . . . , e:n, of e are colored with c1 , . . . , cn ,
respectively. So, col (e) := {e:1,S
. . . , e:n} becomes a set.
For a multiset M , col (M ) := e∈M col (e), which is a
set. For a “colored” set S, col −1 (S) produces a multiset
by stripping tuples from their colors.
Example 2.1. For M = {a, a, a, b, b, c}, col (a) = {a:
1, a:2, a:3}, and col (M ) = S = {a:1, a:2, a:3, b:1, b:2, c:1}.
The inverse operation, the decoloration, gives: col −1 (a:
2) := a; and col −1 (S) := {a, a, a, b, b, c}, a multiset. ✷
We now consider Datalog programs Π with multiset
predicates, i.e. their extensions can be multisets, and
multiset EDBs E. A derivation tree (DT) for Π wrt. E
is a tree with labeled nodes and edges, as follows:
1. For an EDB predicate Q and h ∈ col (Q(E)), a DT
for col −1 (h) contains a single node with label h.
2. For each rule of the form
r : H ← A1 , A2 , . . . , Ak , k > 0,
(1)
and for each tuple hτ1 , . . . , τk i of DTs for the atoms
hd1 , . . . , dk i that unify with hA1 , A2 , . . . , Ak i with
2
The resulting Datalog± program Π+ can be evaluated
according to the usual set semantics on the set EDB
E + . For this, we use the classical chase, i.e., when the
instantiated body of rule r+ becomes true, say with
A1 (ι′1 ; ā1 ), A2 (ι′2 ; ā2 ), . . . , Ak (ι′k ; āk ), then the new tuple
H(ι; ā) is created, with ι the first (new) null from T that
has not been used yet. The new null stands for the tid
of the newly created atom. The chase variant assumed
here is the so-called oblivious chase [7], i.e., a tgd is
activated for every instantiation of the rule body that
makes the rule body true, but a rule is never activated
more than once with the same instantiated body.
The chase instance is obtained by collecting all atoms
obtained by applying the chase with E + and Π+ , i.e.
that are final atoms in a derivation sequence. More
precisely, a derivation sequence is of the form: D1 ❀r1
A1 ; . . . ; Dn ❀rn An with atoms Ai , such that the ri
are tgds, and each Dk is a set of atoms contained in
E + ∪ {A1 , . . . , Ak−1 } that makes the body of rk true,
then Ak is created.
Example 3.1. (ex. 2.2 cont.) Π becomes Π+ :
Multiset Relational Algebra (MRA) consists of the
following basic multiset operations (cf. Section 2): multiset union ∪m , multiset projection πk̄m multiset selecm
tion σC
, with C a condition, multiset (natural) join ✶m ,
and (all-or-nothing) multiset difference ran .
In positive MRA, multiset difference is excluded. Then,
from Theorem 3.1, we immediately obtain:
Proposition 3.1. The PBB semantics applied to positive relational algebra coincides with MRA.
To support full MRA, we have to cover also the (all-ornothing) multiset difference ran . The difficulty of this
step is shown by revisiting Example 2.2.
Example 3.2. (ex. 2.2 cont.) Consider the modified
program Π′ = {r1′ , r2 , r3 } , with rules r2 , r3 as before,
and rule r1 replaced by: r1′ : P (x, y) ← R(x, y), ¬S(x, y),
i.e., body atom S(x, y) is now negated.
Its rewriting into a Datalog± rule as in the positive
case yields r1′+ : ∃zP (z; x, y) ← R(z1 ; x, y), ¬S(z2 , x, y),
with variable z2 occurring in a negated body atom but
not in a positive one, making the rule no longer safe.
We thus introduce an auxiliary predicate Aux in order
to eliminate this variable:
r1+ : ∃zP (z; x, y) ← R(z1 ; x, y), S(z2 , x, y)
r2+ : ∃zR(z; x, y) ← Q(z1 ; x, y, u)
r3+ : ∃zS(z; x, y) ← T (z1 ; u, x, y).
The EDB E becomes E + = {Q(ι1 ; 1, 2, 3), Q(ι2 ; 1, 2, 3),
Q(ι3 ; 1, 2, 5), T (ι4; 4, 1, 2), T (ι5 ; 4, 1, 2)}.
Analogously to the depth-two and depth-three trees
in Example 2.2, the chase produces eleven new atoms.
In total, we get following chase result: Chase(Π+ , E + ) =
E + ∪ {S(ι6 ; 1, 2), S(ι7 ; 1, 2), R(ι8 ; 1, 2), R(ι9 ; 1, 2),
R(ι10 ; 1, 2), P (ι11 ; 1, 2), . . . , P (ι16 ; 1, 2)}.
✷
−1
Inspired by operation col
of Section 2, we introduce de-identification and multiset merging operations,
sometimes using double braces, {{. . .}}, to emphasize
that a construction is intended to produce a multiset .
Definition 3.1. For a set E of tuples with tids, DI(E)
and SP(E), for de-identification and set-projection, resp.,
are: (a) DI(E) := {{P (c̄) | P (t; c̄) ∈ E for some t}},
a multiset ; and (b) SP(E) := {P (c̄) | P (t; c̄) ∈ E
for some t}, a set .
Definition 3.2. Given a Datalog program Π and a
multiset EDB E, the program-based bag semantics (PBB
semantics) assigns to Π ∪ E the multiset:
r1′+ : ∃z P (z; x, y) ← R(z1 ; x, y), ¬Aux (x, y)
Aux (x, y) ← S(z2 ; x, y).
✷
This idea can be generalized to transform safe Datalog
programs with stratified negation (denoted Datalog¬s),
by rewriting a rule of the form r : H(x̄) ← A1 (x̄1 ), . . . ,
Ak (x̄k ), ¬N1 (x̄k+1 ), . . . , ¬Nj (x̄k+j ) into
r′ : ∃zH(z; x̄) ← A1 (z1 ; x̄1 ), . . . , Ak (zk ; x̄k ),
¬Aux 1 (x̄k+1 ), . . . , ¬Aux j (x̄k+j ),
Aux i (x̄k+i ) ← Ni (zi ; x̄k+i ), i = 1, . . . , j.
We keep denoting with PBBS (Π, E) the chase-based
bag semantics for Datalog with stratified and safe negation (referred to as Datalog¬s ) obtained via this Datalog±
rewriting with tids.
In [10], a bag semantics for Datalog¬s was introduced
via derivation-trees (DTs), extending the DTB semantics in [9] for (positive) Datalog. This extension applies
to Datalog programs with stratified negation that are
range-restricted and safe (i.e. a variable in a rule head
or a negative literal must appear in a positive literal
in the body of the same rule); and leads to an allor-nothing interpretation of negation, so as with the
ran operator. We keep using DTBS (Π, E) to denote
the new DTB semantics, and we assume, from now on,
that Datalog programs are stratified, range-restricted
and safe. The following extensions of Theorem 3.1 and
Proposition 3.1 can be obtained:
PBBS (Π, E) := DI(Chase(Π+ , E + )) =
{{P (ā) | P (t; ā) ∈ Chase(Π+ , E + )}}.
The PBB semantics coincides with the DTB semantics , as the following result states, which can be immediately obtained from Lemma 3.1.
Theorem 3.1. For a Datalog program Π with a multiset EDB E, DTBS (Π, E) = PBBS (Π, E).
Lemma 3.1. For a Datalog program Π with multiset
EDB E, there is a one-to-one correspondence between
DTs in DT (Π, col (E)) with root atoms P (c̄) and chasesequences that end in atoms P (ι, c̄), with ι ∈ T .
Corollary 3.1. Given a Datalog program Π and a
multiset EDB E, the instance SP(Chase(Π+ , E + )) is
the minimal model of the Datalog program Π ∪ SP(E).
Theorem 3.2. For a stratified Datalog program Π
with multiset EDB E, DTBS (Π, E) = PBBS (Π, E).
Proposition 3.2. The PBB semantics applied to relational algebra including all-or-nothing multiset difference ran coincides with MRA.
3
4.
∃Z t(Z; U, V, V ) ← t(U, V, V )
✷
We will work out in detail the application of our PBB
semantics to the evaluation of SPARQL with bag semantics and give first experimental results with its implementation in the VADALOG system [3] in the forthcoming full version of this paper. Recall that warded
Datalog± has been shown to be expressive enough to
capture SPARQL under the OWL 2 QL core entailment
regime [5], with set semantics though [2, 6]. We will
tackle this extension of SPARQL with the OWL 2 QL
core entailment regime under bag semantics with our
transformation to warded Datalog± .
We are currently also working on inexpressibility results related to the transformation presented here: recall that our translation into Datalog± does not cover
multiset intersection ∩m ; moreover, multiset difference
is only handled in the all-or-nothing form ran , while the
sometimes more natural form rm has been left out. We
conjecture that these two operations are not expressible
in Datalog± with set semantics. The verification of this
conjecture is a matter of ongoing work.
PROPERTIES OF PBB SEMANTICS
In [6], warded Datalog was introduced as a particularly well-behaved fragment of Datalog± , for which
CQA is tractable (cf. [6] for details). Actually, in the
journal version of [6] (currently under review), warded
Datalog is extended with stratified negation, preserving
the favorable properties of warded Datalog.
We can show that the Datalog± program Π+ obtained
from a Datalog or Datalog¬s program Π by our construction in Section 3 is indeed warded:
Theorem 4.1. Let Π be a program in Datalog or Datalog¬s and let Π+ be the Datalog± program that represents its PBB semantics. Then Π+ is warded.
5.
CONCLUSION AND FUTURE WORK
In this work, we have proposed and achieved a declarative and computable specification in Datalog± of the
multiset semantics of Datalog. Actually, we have extended this specification to Datalog¬s thus capturing
the full multiset relational algebra (MRA), including
difference. The resulting Datalog± programs have still
good computational properties, and provably allow for
tractable (conjunctive) query answering.
An immediate application of our results is to the evaluation of SPARQL with duplicates via the Datalog rewriting proposed in [1] as an intermediate step. This is
illustrated by the following example:
Example 5.1. Consider the following SPARQL query
from
https://www.w3.org/2009/sparql/wiki/
TaskForce:PropertyPaths#Use_Cases:
PREFIX foaf: <http://xmlns.com/foaf/0.1/>
SELECT ?name
WHERE
{ ?x foaf:mbox <mailto:alice@example> .
?x foaf:knows/foaf:name ?name }
6. REFERENCES
[1] R. Angles and C. Gutiérrez. The multiset
semantics of SPARQL patterns. In Proc. ISWC
2016, volume 9981 of LNCS, pages 20–36, 2016.
[2] M. Arenas, G. Gottlob, and A. Pieris. Expressive
languages for querying the semantic web. In Proc.
PODS’14, pages 14–26. ACM, 2014.
[3] L. Bellomarini, G. Gottlob, A. Pieris, and
E. Sallinger. Swift logic for big data and
knowledge graphs. In Proc. IJCAI 2017, pages
2–10. ijcai.org, 2017.
[4] A. Calı̀, G. Gottlob, and M. Kifer. Taming the
infinite chase: Query answering under expressive
relational constraints. J. Artif. Intell. Res.,
48:115–174, 2013.
[5] B. Glimm, C. Ogbuji, S. Hawke, I. Herman,
B. Parsia, A. Polleres, and A. Seaborne. SPARQL
1.1 Entailment Regimes.
W3C Recommendation 21 march 2013, W3C, 2013.
https://www.w3.org/TR/sparql11-entailment/.
[6] G. Gottlob and A. Pieris. Beyond SPARQL under
OWL 2 QL entailment regime: Rules to the
rescue. In Proc. IJCAI 2015, pages 2999–3007.
AAAI Press, 2015.
[7] D. S. Johnson and A. C. Klug. Testing
containment of conjunctive queries under
functional and inclusion dependencies. J. Comput.
Syst. Sci., 28(1):167–189, 1984.
[8] M. J. Maher and R. Ramakrishnan. Déjà vu in
fixpoints of logic programs. In Proc. NACLP
1989, pages 963–980. MIT Press, 1989.
[9] I. S. Mumick, H. Pirahesh, and R. Ramakrishnan.
The magic of duplicates and aggregates. In Proc.
VLDB 1990, pages 264–277. Morgan Kaufmann,
1990.
[10] I. S. Mumick and O. Shmueli. Finiteness
properties of database queries. In Proc. ADC ’93,
pages 274–288. World Scientific, 1993.
This query retrieves the names of people Alice knows.
In our translation into Datalog, we omit the prefix and
use the constant alice short for <mailto:alice@example>
to keep the notation simple. Moreover, we assume that
the RDF data over which the query has to be evaluated
is given by a relational database with a single ternary
predicate t , i.e., the database consists of triples t(·, ·, ·).
The following Datalog program with answer predicate
ans is equivalent to the above SPARQL query:
ans(N ) ← t(X, mbox , alice),
t(X, knows, Y ), t(Y, name, N )
We now apply our transformation into Datalog± from
Section 3. Suppose that the query has to be evaluated
over a multiset E of triples. Recall that in our transformation, we would first have to extend all triples in
E to quadruples by adding a null (the tid) in the 0position. Actually, this can be easily automatized by
the first rule in the program below. We thus get the
following Datalog± program:
∃Z ans(Z; N ) ← t(Z1 ; X, mbox , alice),
t(Z2 ; X, knows, Y ),
t(Z3 ; Y, name, N )
4
| 2cs.AI
|
HOROFUNCTIONS ON GRAPHS OF LINEAR GROWTH
arXiv:1608.03727v2 [math.MG] 31 Oct 2016
MATTHEW C. H. TOINTON AND ARIEL YADIN
Abstract. We prove that a linear growth graph has finitely many horofunctions. This provides a
short and simple proof that any finitely generated infinite group of linear growth is virtually cyclic.
Résumé. Nous montrons qu’un graphe à croissance linéaire admet un nombre fini d’horofonctions.
Celà donne une preuve courte et simple que chaque groupe infini de type fini à croissance linéaire
est virtuellement cyclique.
1. Introduction
Let Γ be an infinite, connected, locally finite graph, and denote by d(·, ·) the graph metric on Γ.
Let o ∈ Γ. Then Γ is said to have linear volume growth if the balls about o for the metric d grow
at most linearly in the radius. The graph Γ is said to have polynomial volume growth if these balls
grow at most polynomially.
Given an element z ∈ Γ we define the Busemann function bz : Γ → Z via bz (y) = d(z, y) − d(z, o).
Given a geodesic ray ω = (z1 , z2 , . . .) in Γ we define the horofunction fω : G → Z by fω (y) =
limn→∞ bzn (y). It is a well-known and simple fact that this limit exists. Note that fω is not
constant, and in fact fω (zn ) = −n, which shows that fω is unbounded.
Theorem 1.1. Let Γ be an infinite, connected, locally finite graph of linear volume growth. Then
the set of horofunctions on Γ is finite.
A finitely generated group G is said to have polynomial (respectively, linear) volume growth if
some (and hence every) Cayley graph of G has polynomial (respectively, linear) volume growth.
A remarkable theorem of Gromov’s states that a finitely generated group of polynomial volume
growth contains a nilpotent subgroup of finite index [2]. As an application of Theorem 1.1 we give
a short argument to prove the linear-growth case of Gromov’s theorem, as follows.
Theorem 1.2. Let G be a finitely generated infinite group of linear volume growth. Then G contains
a cyclic subgroup of finite index.
In fact, Gromov’s theorem implies that a group of subquadratic growth is virtually cyclic, and
this has also been proved by elementary methods by Justin [5], van den Dries & Wilkie [1] and
Imrich & Seifter [4], the last two of these giving bounds on the index of the cyclic subgroup in
terms of the volume growth. Nonetheless, the present proof is completely different to all of those
and rather short, so we record it here.
We prove Theorem 1.1 in Section 2. The proof of Theorem 1.2 is in Section 3.
Let us mention a related (probably much more difficult) question.
Conjecture 1.3. Let G be a Cayley graph of polynomial volume growth. Then the set of horofunctions on G is countable.
1
2
MATTHEW C. H. TOINTON AND ARIEL YADIN
Proving this conjecture would provide an alternative proof to Gromov’s theorem (by using a
variant of Lemma 3.1 below). This has been suggested by Karlsson [6]. One method to prove
this conjecture could be using the structure of finitely generated nilpotent groups, and relying on
Gromov’s theorem, but that would somehow miss the point. (For example, in [7] Walsh shows that
nilpotent groups always have a finite orbit in the space of horofunctions. It seems that this can
be extended to virtually nilpotent groups as well.) It would be interesting to prove this conjecture
even in the quadratic growth case without using Gromov’s theorem, since that would imply a new
proof of the characterization of recurrent groups (which are finite extensions of Z or Z2 ).
2. Horofunctions on a graph of linear growth
For
S a graph Γ, we say that Γ is N-partite on a sequence (Γn )n∈N of disjoint sets Γn if Γ has vertex
set n∈N Γn and the neighbours of every x ∈ Γn lie in Γn−1 ∪ Γn+1 . We call the sets Γn the partite
sets of Γ. We call a path in Γ monotone if it has at most one vertex in each Γn .
The proof of Theorem 1.1 essentially rests on the following graph-theoretic result.
Proposition 2.1. Let Γ be an N-partite graph whose partite sets all have cardinality k ≥ 1. Then
there exist monotone paths γ1 , . . . , γk in Γ such that every infinite monotone path in Γ has infinite
intersection with some γj .
Recall that in a bipartite graph on two finite sets X1 , X2 of equal cardinality, a matching is a
subgraph in which each element of X1 is connected to precisely one element of X2 , and vice versa.
Hall’s Marriage Theorem [3] states that if there is no matching then there exists some subset Y ⊂ X1
such that the neighbourhood of Y in X2 has strictly smaller cardinality than Y itself.
Lemma 2.2. If for each n ∈ N there is a matching in Γ between Γn and Γn+1 then Γ satisfies
Proposition 2.1.
Proof. It is easy to see that the existence of such matchings implies that the vertices of Γ may be
partitioned into k monotone paths, and that this is sufficient to satisfy the proposition.
GivenSan N-partite graph Γ and a sequence N = (nj )j ⊂ N, we may define a new graph on
ΓN := j Γnj by placing an edge between x ∈ Γnj and x′ ∈ Γnj+1 if and only if there exists a
monotone path between x and x′ in Γ. (Note that ΓN is an N-partite graph with partite sets Γnj .)
The following is then immediate.
Lemma 2.3. If there exists a sequence N = (nj )j such that ΓN satisfies Proposition 2.1, then the
conclusion of Proposition 2.1 holds for Γ as well.
Proof of Proposition 2.1. We proceed by induction on k. The base case k = 1 is easy, so we assume
that k > 1. We may also delete every element of Γ that does not lie in any infinite monotone path;
the only potential problem with this is that the Γn may no longer all have the same cardinality,
but, using Lemma 2.3, we may fix this by passing to a subsequence.
Let N = (nj )j be a sequence and consider the N-partite graph ΓN . If for every j there exists a
matching (in the graph ΓN ) between Γnj and Γnj+1 , then we are done by combining Lemmas 2.2
and 2.3.
Thus, we assume that a sequence as above does not exist. Specifically, by Hall’s Marriage
Theorem, there exists n such that for any m > n, there exist Um ⊂ Γn and Vm ⊂ Γm such that
1 ≤ |Vm | < |Um | ≤ k, and such that every monotone path from Um to Γm ends in Vm . Without loss
of generality we assume that the sets Vm are minimal with respect to these properties, and hence
HOROFUNCTIONS ON GRAPHS OF LINEAR GROWTH
3
that every element v ∈ Vm lies in some monotone path from Um to Γm . By passing to a subsequence
of m > n, we have an infinite sequence n < m1 < m2 < · · · such that |Vmj | = |Vm1 | all have the
same size and Umj = Um1 = U are all the same fixed subset.
Let M = (mj )j and consider the graph ΓM . We claim that ΓM satisfies Proposition 2.1, which
will suffice by Lemma 2.3. We move to proving this claim.
Every monotone path in Γ starting in U and ending in Γmj must end in Vmj . Thus, by minimality
of Vmj , any monotone path in Γ starting in Vmj and ending in Γmj+1 must end in Vmj+1 . In
particular, in the graph ΓM , any infinite monotone path γ must satisfy the following dichotomy:
either γ ∩ Vmj = ∅ for all j, or there exists j0 such that for all
S j > j0 we have γ ∩ Vmj 6= ∅.
Let ΓA be the induced subgraph of ΓM on the vertex set j Vmj , and ΓB the induced subgraph
S
on the vertex set j (Γmj \ Vmj ). Note that ΓA is N-partite with partite sets (ΓA )j = Vmj , all of
size v = |Vm1 | < k. Also, ΓB is N-partite with partite sets (ΓB )j = Γmj \ Vmj , which all have size
w = k − v < k. Thus, any infinite monotone path in ΓM induces an infinite monotone path in
either ΓA or in ΓB . By induction, there exist infinite monotone paths α1 , . . . , αv in ΓA such that
any infinite monotone path in ΓA must intersect one of these infinitely many times. Similarly, there
are such paths β1 , . . . , βw in ΓB . Thus, any infinite monotone path in ΓM must intersect one of
α1 , . . . , αv , β1 , . . . , βw infinitely many times. This completes the proof.
Proof of Theorem 1.1. Note that if Γ has linear growth then, writing Br for the ball of radius r about
o, there is some k ∈ N and an infinite increasing sequence m1 , m2 , . . . such that |Bmn \Bmn −1 | = k
for every n. Define Γn = Bmn \Bmn −1 , and define an N-partite graph Γ̂ on Γ1 , Γ2 , . . . by joining
x ∈ Γn to x′ ∈ Γn+1 if and only if there is a path in Γ from x to x′ of length mn+1 − mn .
Define a map α from the set of geodesic rays in Γ starting at o to the set of monotone paths in Γ̂
(in the sense of Proposition 2.1) in the natural way. Specifically, if ω is a geodesic ray in Γ starting
S
at o then α(ω) is the unique monotone path in Γ̂ passing through the same elements of n Γn as
ω. Note that α is surjective onto the set of monotone paths in Γ̂, and also that if α(ω) and α(ω ′ )
have infinite intersection then so do ω and ω ′ .
Let γ1 , . . . , γk be as given by Proposition 2.1, and pick, using the surjectivity of α, geodesic rays
ω1 , . . . , ωk in Γ starting at o such that α(ωi ) = γi . If β is a geodesic ray in Γ, the tail of β coincides
with the tail of some geodesic ray β ′ in Γ starting at o (see Lemma 2.4 below). However, α(β ′ )
has infinite intersection with some γi by Proposition 2.1, and so β has infinite intersection with ωi .
This implies in particular that fβ = fωi , and so fω1 , . . . , fωk is a complete set of horofunctions.
For completeness we include a short argument for the following standard lemma.
Lemma 2.4. If γ = (x0 , x1 , x2 , . . .) is a geodesic ray starting at x0 then there exists some N such
that (xN , xN +1 , . . .) coincides with the tail of a geodesic ray ω starting at o.
Proof. The sequence d(xn , o) − d(xn , x0 ) is non-increasing in n, since
(2.1)
d(xn+1 , x0 ) = d(xn , x0 ) + 1
and |d(xn+1 , o) − d(xn , o)| ≤ 1 for every n. The triangle inequality also implies that d(xn , o) −
d(xn , x0 ) is bounded below by −d(o, x0 ). The sequence (d(xn , o) − d(xn , x0 ))∞
n=1 is therefore eventually constant, say for n ≥ N . Combined with (2.1), this implies that d(xn+1 , o) = d(xn , o) + 1 for
n ≥ N . The infinite path ω having initial segment some geodesic path from o to xN , followed by
xN +1 , xN +2 , . . ., is therefore a geodesic ray starting at o.
4
MATTHEW C. H. TOINTON AND ARIEL YADIN
3. The linear-growth case of Gromov’s theorem
A group G acts on the space {f : G → R | f (1) = 0} by x·f (y) = f (x−1 y)− f (x−1 ). Note that for
a Busemann function bz we have x · bz = bxz , and hence for a horofunction fω we have x · fω = fxω .
The following observation we learned from Anders Karlsson.
Lemma 3.1. If the set of horofunctions on a group G contains a finite orbit then G has a finiteindex subgroup admitting a surjective homomorphism onto Z.
Proof. Letting G act on the finite orbit, G contains a finite-index subgroup H that fixes some
element fω of the orbit. Thus for h ∈ H, g ∈ G we have fω (g) = h−1 · fω (g) = fω (hg) − fω (h), which
implies that fω is a homomorphism H → Z and that fω (Hg) = fω (H) + fω (g) for every g ∈ G. In
particular, if fω (H) = {0} then fω is constant on the finitely many cosets of H, contradicting the
fact that horofunctions are unbounded. We conclude that the image fω (H) is a non-trivial subgroup
of Z, and thus admits a surjective homomorphism onto Z.
Remark 3.2. Essentially the same argument shows more generally that if Γ is a graph of linear
growth and G < Aut (Γ) acts transitively on the vertices of Γ then G has a finite-index subgroup
admitting a surjective homomorphism onto Z.
Proof of Theorem 1.2. Theorem 1.1 implies that G has a finite set of horofunctions. The set of
horofunctions is invariant, so in this case it contains a finite orbit. Lemma 3.1 therefore implies
that there exists N ⊳ G of finite index such that N admits a surjective homomorphism onto Z. Let
K ⊳ N be the kernel of this homomorphism. Since N is finite index in G it is finitely generated
of linear growth. Since N/K ∼
= Z, it must be that K is finite. Hence, N is finite-by-Z, which by
standard methods implies that N is also Z-by-finite. Thus, G contains a finite-index infinite cyclic
subgroup.
Acknowledgements
We are grateful to Emmanuel Breuillard, Anders Karlsson, Sébastien Martineau, Tom Meyerovitch,
Ville Salo and an anonymous referee for helpful comments and discussions.
MT is supported by ERC grant GA617129 ‘GeTeMo’. AY is supported by the Israel Science
Foundation (grant no. 1346/15).
References
[1]
[2]
[3]
[4]
[5]
[6]
L. van den Dries and A. J. Wilkie. An effective bound for groups of linear growth, Arch. Math. 42 (1984), 391–396.
M. Gromov. Groups of polynomial growth and expanding maps, Publ. Math. IHES 53 (1981), 53–73.
P. Hall. On Representatives of Subsets, J. London Math. Soc. 10 (1935), 26–30
W. Imrich and N. Seifter. A bound for groups of linear growth, Arch. Math. 48 (1987), 100–104.
J. Justin. Groupes et semi-groupes à croissance linéaire, C.R. Acad. Sci. Paris 273 (1971), 212–214.
A.
Karlsson.
Ergodic
theorems
for
noncommuting
random
products
(lecture
notes)
http://www.unige.ch/math/folks/karlsson/wroclawtotal.pdf
[7] C. Walsh. The action of a nilpotent group on its horofunction boundary has finite orbits, Groups Geom. Dyn. 5
(2011), 189–206.
MT: Laboratoire de Mathématiques d’Orsay, Univ. Paris-Sud, CNRS, Université Paris-Saclay, 91405
Orsay, France
E-mail address: [email protected]
AY: Department of Mathematics, Ben-Gurion University of the Negev, Beer-Sheva, Israel
E-mail address: [email protected]
| 4math.GR
|
Trajectory-based Radical Analysis Network for
Online Handwritten Chinese Character Recognition
Jianshu Zhang, Yixing Zhu, Jun Du and Lirong Dai
arXiv:1801.10109v1 [cs.CV] 22 Jan 2018
National Engineering Laboratory for Speech and Language Information Processing
University of Science and Technology of China, Hefei, Anhui, P. R. China
Email: [email protected], [email protected], [email protected], [email protected]
Abstract—Recently, great progress has been made for online
handwritten Chinese character recognition due to the emergence
of deep learning techniques. However, previous research mostly
treated each Chinese character as one class without explicitly
considering its inherent structure, namely the radical components
with complicated geometry. In this study, we propose a novel
trajectory-based radical analysis network (TRAN) to firstly
identify radicals and analyze two-dimensional structures among
radicals simultaneously, then recognize Chinese characters by
generating captions of them based on the analysis of their internal
radicals. The proposed TRAN employs recurrent neural networks
(RNNs) as both an encoder and a decoder. The RNN encoder
makes full use of online information by directly transforming
handwriting trajectory into high-level features. The RNN decoder
aims at generating the caption by detecting radicals and spatial
structures through an attention model. The manner of treating a
Chinese character as a two-dimensional composition of radicals
can reduce the size of vocabulary and enable TRAN to possess
the capability of recognizing unseen Chinese character classes,
only if the corresponding radicals have been seen. Evaluated
on CASIA-OLHWDB database, the proposed approach significantly outperforms the state-of-the-art whole-character modeling
approach with a relative character error rate (CER) reduction
of 10%. Meanwhile, for the case of recognition of 500 unseen
Chinese characters, TRAN can achieve a character accuracy of
about 60% while the traditional whole-character method has no
capability to handle them.
I. I NTRODUCTION
Machine recognition of handwritten Chinese characters has
been studied for decades [1]. It is a challenging problem due to
a large number of character classes and enormous ambiguities
coming from handwriting input. Although some conventional
approaches have obtained great achievements [2]–[6], they
only treated the character sample as a whole without considering the similarity and internal structures among different
characters. And they have no capability of dealing with unseen
character classes.
However, Chinese characters can be decomposed into a
few fundamental structure components, called radicals [7].
It is an intuitive way to first extract information of radicals
that is embedded in Chinese characters and then use this
knowledge for recognition. In the past few decades, lots of
efforts have been made for radical-based Chinese character
recognition. For example, [8] proposed a matching method for
radical-based Chinese character recognition. It first detected
radicals separately and then employed a hierarchical radical
matching method to compose radicals into a character. [9]
analysis
structures
radicals
扌舟
几又
扌
舟
left−right (a)
几
top−down (d)
又
left−right (a)
a {扌a {舟d { 几 又 } } }
handwriting input
Chinese character caption
Fig. 1. Illustration of TRAN to recognize Chinese characters by analyzing
the radicals and the corresponding structures.
tried to over-segment characters into candidate radicals while
the proposed way could only handle the left-right structure
and over-segmentation brings many difficulties. Recently, [10]
proposed a multi-label learning for radical-based Chinese
character recognition. It turned a character class into a combination of several radicals and spatial structures. Generally,
these approaches have difficulties when dealing with radical
segmentation and the analysis of structures among radicals is
not flexible. Besides, they did not focus on recognizing unseen
Chinese character classes.
In this paper, we propose a novel radical-based approach
to online handwritten Chinese character recognition, namely
trajectory-based radical analysis network (TRAN). Different
from above mentioned radical-based approaches, in TRAN the
radical segmentation and structure detection are automatically
addressed by an attention model which is jointly optimized
with the entire network. The main idea of TRAN is to
decompose a Chinese character into radicals and detect the
spatial structures among radicals. We then describe the analysis of radicals as a Chinese character caption. A handwritten
Chinese character is successfully recognized when its caption
matches ground-truth. To be more accessible, we illustrate the
TRAN learning way in Fig. 1. The online handwritten Chinese
character input is visualized at the bottom-left of Fig. 1. It is
composed of four different radicals. The handwriting input
is finally recognized as the bottom-right Chinese character
caption after the top-down and left-right structures among radicals are detected. Based on analysis of radicals, the proposed
TRAN possesses the capability of recognizing unseen Chinese
character classes if the radicals have been seen.
The proposed TRAN is an improved version of attention-
based encoder-decoder model [11] with RNN [12]. The
attention-based encoder-decoder model has been extensively
applied to many applications including machine translation [13], [14], image captioning [15], [16], speech recognition [17] and mathematical expression recognition [18],
[19]. The raw data of online handwritten Chinese character
input are variable-length sequence (xy-coordinates). TRAN
first employs a stack of bidirectional RNN [20] to encode input
sequence into high-level representations. Then a unidirectional
RNN decoder converts the high-level representations into
output character captions one symbol at a time. For each
predicted radical, a coverage based attention model [21] built
in the decoder scans the entire input sequence and chooses
the most relevant part to describe a segmented radical or
a two-dimensional structure between radicals. Our proposed
TRAN is related to our previous work [22] with two main
differences: 1) [22] focused on the application of RAN on
printed Chinese character recognition while this paper focuses
on handwritten Chinese character recognition. It is interesting
to investigate the performance of RAN on handwritten Chinese
character recognition as handwritten characters are much more
ambiguous due to the diversity of writing styles. 2) Instead of
transforming online handwritten characters into static images
and employing convolutional neural network [23] to encode
them, we choose to directly encode the raw sequential data
by employing an RNN encoder in order to fully exploit the
dynamic trajectory information that can not be recovered from
static images.
The main contributions of this study are as follows:
•
•
•
•
We propose TRAN for online handwritten Chinese character recognition.
The size of radical vocabulary is largely less than Chinese
character vocabulary, leading to decrease of redundancy
among output classes and improvement of recognition
performance.
TRAN possess the ability of recognizing unseen or newly
created Chinese characters, only if the radicals have been
seen.
We experimentally demonstrate how RAN performs on
online handwritten Chinese character recognition compared with state-of-the-arts and show its effectiveness on
recognizing unseen character classes.
II. D ESCRIPTION OF C HINESE CHARACTER CAPTION
In this section, we will introduce how we generate captions
of Chinese characters. The character caption is composed of
three key components: radicals, spatial structures and a pair
of braces (e.g. “{” and “}”). A radical represents a basic
part of Chinese character and it is often shared by different Chinese characters. Compared with enormous Chinese
character categories, the amount of radicals is quite limited.
It is declared in GB13000.1 standard published by National
Language Committee of China that nearly 500 radicals consist
of over 20,000 Chinese characters. As for the complicated
two-dimensional spatial structures among radicals, Fig. 2
single
a
d
stl
str
sbl
sl
sb
st
s
w
Fig. 2. Graphical representation of eleven common spatial structures among
radicals, different radicals are divided by internal line.
illustrates eleven common structures and the descriptions are
demonstrated as follows:
• single-element: sometimes a single radical represents a
Chinese character and therefore we can not find internal
structures in such characters
• a: left-right structure
• d: top-bottom structure
• stl: top-left-surround structure
• str: top-right-surround structure
• sbl: bottom-left-surround structure
• sl: left-surround structure
• sb: bottom-surround structure
• st: top-surround structure
• s: surround structure
• w: within structure
After decomposing Chinese characters into radicals and
internal spatial structures by following [24], we use a pair of
braces to constrain a single structure. Take “stl” as an example,
it is captioned as “stl { radical-1 radical-2 }”. The generation
of a Chinese character caption is finished when all radicals are
included in the caption.
III. T HE PROPOSED APPROACH
In this section, we elaborate the proposed TRAN framework, namely generating an underlying Chinese character
caption from a sequence of online handwritten trajectory
points, as illustrated in Fig. 3. First, we extract trajectory
information as the input feature from original trajectory points
(xy-coordinates). A stack of bidirectional RNNs are then
employed as the encoder to transform the input feature into
high-level representations. Since the original trajectory points
are a variable-length sequence, the extracted high-level representations are also a variable-length sequence. To associate the
variable-length representations with variable-length character
caption, we generate a fixed-length context vector via weighted
summing the high-level representations and a unidirectional
RNN decoder uses the fixed-length context vector to generate
the character caption one symbol at a time. We introduce
a
{
扌
{
a
舟
…
B. Encoder
Given the feature sequence (x1 , x2 , . . . , xN ), we employ
RNN as the encoder to encode them into high-level representations as RNN has shown its strength in processing sequential
signals. However, a simple RNN has revealed serious problems during training namely vanishing gradient and exploding
gradient [25], [26]. Therefore, an improved version of RNN
named gated recurrent units (GRU) [27] which can alleviate
these two problems is employed in this study as it utilizes an
update gate and a reset gate to control the flow of forward
information and backward gradient. The GRU hidden state ht
in encoder is computed by:
𝐘
…
Decoder
Encoder
𝐀
𝐗
𝐡1𝟑
𝐡𝟑𝟑
𝐡𝟑𝟐
𝐡𝟑𝟓
𝐡𝟑𝟒
ht = GRU (xt , ht−1 )
𝐡1𝟐
𝐡𝟐𝟐
𝐡𝟐𝟑
𝐡𝟐𝟒
𝐡𝟐𝟓
𝐡𝟐𝟔
𝐡𝟐𝟕
𝐡𝟐𝟖
𝐡𝟐𝟗
𝟐
𝐡1𝟎
𝐡11
𝐡1𝟐
𝐡1𝟑
𝐡1𝟒
𝐡1𝟓
𝐡1𝟔
𝐡1𝟕
𝐡1𝟖
𝐡1𝟗
𝐡1𝟏𝟎
𝐱1
𝐱2
𝐱3
𝐱4
𝐱5
𝐱6
𝐱7
𝐱8
and the GRU function can be expanded as follows:
𝐱 9 𝐱10
Fig. 3. Overall framework of TRAN for online handwritten Chinese character
recognition. It is composed of a bidirectional RNN encoder and a unidirectional RNN decoder.
an attention model to produce the weighting coefficients so
that the context vector can contain only useful trajectory
information at each decoding step.
A. Feature extraction
During the data acquisition of online handwritten Chinese
character, the pen-tip movements (xy-coordinates) and pen
states (pen-down or pen-up) are stored as variable-length
sequential data:
{[x1 , y1 , s1 ] , [x2 , y2 , s2 ] , . . . , [xN , yN , sN ]}
(1)
where N is the length of sequence, xi and yi are the xycoordinates of the pen movements and si indicates which
stroke the ith point belongs to.
To address the issue of non-uniform sampling by different
writing speed and the size variations of the coordinates on
different potable devices, the interpolation and normalization
to the original trajectory points are first conducted according
to [4]. Then we extract a 6-dimensional feature vector for each
point:
[xi , yi , ∆xi , ∆yi , δ(si = si+1 ), δ(si 6= si+1 )]
(3)
(2)
where ∆xi = xi+1 − xi , ∆yi = yi+1 − yi , and δ(·) = 1 when
the condition is true or zero otherwise. The last two terms are
flags which indicate the status of the pen, i.e., [1, 0] and [0, 1]
are pen-down and pen-up respectively. For convenience, in the
following sections, we use X = (x1 , x2 , . . . , xN ) to denote
the input sequence of the encoder, where xi ∈ Rd (d = 6).
zt = σ(Wxz xt + Uhz ht−1 )
(4)
rt = σ(Wxr xt + Uhr ht−1 )
(5)
h̃t = tanh(Wxh xt + Urh (rt ⊗ ht−1 ))
(6)
ht = (1 − zt ) ⊗ ht−1 + zt ⊗ h̃t
(7)
where σ denotes the sigmoid activation function, ⊗ denotes
an element-wise multiplication operator, zt , rt and h̃t are the
update gate, reset gate and candidate activation, respectively.
Wxz , Wxr , Wxh , Uhz , Uhr and Urh are related weight
matrices.
Nevertheless, even if the unidirectional GRU can have
access to the history of input signals, it does not have the
ability of modeling future context. Therefore we exploit the
bidirectional GRU by passing the input vectors through two
GRU layers running in opposite directions and concatenating their hidden state vectors so that the encoder can use
both history and future information. To obtain a high-level
representation, the encoder stacks multiple GRU layers on
top of each other as illustrated in Fig. 3. In this study,
our encoder consists of 4 bidirectional GRU layers. Each
layer has 250 forward and 250 backward GRU units. We
also add pooling over time axes in high-level GRU layers
because: 1) the high-level representations are overly precise
and contain much redundant information; 2) the decoder needs
to attend less if the number of encoder output reduces, leading
to improvement of performance; 3) the pooling operation
accelerates the encoding process. The pooling is applied to
the top GRU layer by dropping the even output over time.
Assuming the bidirectional GRU encoder produces a highlevel representation sequence A with length L. Because there
is one pooling operation in the bidirectional GRU encoder,
L = N2 . Each of these representations is a D-dimensional
vector (D = 500):
A = {a1 , . . . , aL } , ai ∈ RD
(8)
C. Decoder with attention model
After obtaining high-level representations A, the decoder
aims to make use of them to generate a Chinese character
caption. The output sequence Y is represented as a sequence
of one-hot encoded vectors:
Y = {y1 , . . . , yC } , yi ∈ RK
(9)
where K is the vocabulary size and C is the length of character
caption. Note that, both the length of representation sequence
(L) and the length of character caption (C) are variable.
To address the mapping from variable-length representation
sequence to variable-length character caption, we attempt to
compute an intermediate fixed-size vector ct that incorporates
useful information of representation sequence. The decoder
then utilizes this fixed-size vector to predict the character caption one symbol at a time. As ct contains overall information
of input sequence, we call it context vector. At each decoding
step, the probability of the predicted word is computed by
the context vector ct , current decoder state st and previous
predicted symbol yt−1 using a multi-layer perceptron:
p(yt |yt−1 , X) = g (Wo h(Eyt−1 + Ws st + Wc ct )) (10)
where g denotes a softmax activation function over all the
symbols in the vocabulary, h denotes a maxout activation function. Let m and n denote the dimensions of embedding and
m
decoder state, Wo ∈ RK× 2 , Ws ∈ Rm×n , Wc ∈ Rm×D ,
and E denotes the embedding matrix.
Since the context vector ct needs to be fixed-length, it is
an intuitive way to produce it by summing all representation
vectors ai at time step t. However, average summing is too
robust and leads to loss of useful information. Therefore,
we adopt weighted summing while the weighting coefficients
are called attention probabilities. The attention probability
performs as a description that tells which part of representation
sequence is useful at each decoding step. We compute the
decoder state st and context vector ct as follows:
ŝt = GRU (yt−1 , st−1 )
Xt−1
F=Q∗
αl
(11)
T
eti = νatt
tanh(Watt ŝt + Uatt ai + Uf fi )
exp(eti )
αti = PL
k=1 exp(etk )
XL
ct =
αti ai
(13)
st = GRU (ct , ŝt )
(16)
l=1
i=1
(12)
(14)
past attention probabilities so that the coverage vector contains
the information of alignment history as shown in Eq. (12).
We adopt the coverage vector in order to let the attention
model know which part of representation sequence has been
0
attended or not0 [28]. Let n 0 denote the attention
dimension.
0
Then νatt ∈ Rn , Watt ∈ Rn ×n and Uatt ∈ Rn ×D .
IV. T RAINING AND T ESTING D ETAILS
The training objective of the proposed model is to maximize
the predicted symbol probability as shown in Eq. (10) and we
use cross-entropy (CE) as the objective function:
XC
O=−
log p(wt |yt−1 , X)
(17)
t=1
where wt represents the ground truth word at time step t,
C is the length of output string. The implementation details
of GRU encoder has been introduced in Section III-B. The
decoder uses two layers with each using 256 forward GRU
units. The embedding dimension m, decoder state dimension n
and attention dimension n0 are all set to 256. The convolution
kernel size for computing coverage vector is set to (5×1) as it
is a one-dimensional convolution operation, while the number
of convolution filters is set to 256. We utilize the adadelta
algorithm [29] with gradient clipping for optimization. The
adadelta hyperparameters are set as ρ = 0.95, ε = 10−8 .
In the decoding stage, we aim to generate a most likely
character caption given the input trajectory:
ŷ = arg max log P (y|X)
(18)
y
However, different from the training procedure, we do not
have the ground truth of previous predicted word. To prevent
previous prediction errors inherited by next decoding step , a
simple left-to-right beam search algorithm [30] is employed to
implement the decoding procedure. Here, we maintained a set
of 10 partial hypotheses beginning with the start-of-sentence
< sos >. At each time step, each partial hypothesis in the
beam is expanded with every possible word and only the 10
most likely beams are kept. This procedure is repeated until
the output word becomes the end-of-sentence < eos >.
V. E XPERIMENTS
(15)
Here, we can see that the decoder adopts two unidirectional
GRU layers to calculate the decoder state st . The GRU
function is the same one in Eq. (3). ŝt denotes the current
decoder state prediction, eti denotes the energy of ai at time
step t conditioned on ŝt . The attention probability αti , which
is the ith element of αt , is computed by taking eti as input of
a softmax function. The context vector ct is then calculated
via weighted summing representation vectors ai with attention
probabilities employed as weighting coefficients. During the
computation of attention probability, we also append a coverage vector fi (the ith vector of F) in the attention model. The
coverage vector is computed based on the summation of all
In this section, we present experiments on recognizing seen
and unseen online handwritten Chinese character classes by
answering the following questions:
Q1
Is the TRAN effective when recognizing seen Chinese character classes?
Q2
Is the TRAN effective when recognizing unseen
Chinese character classes?
Q3
How does the TRAN analyze the radicals and spatial
structures?
A. Performance on recognition of seen Chinese character
classes (Q1)
In this section, we show the effectiveness of TRAN on
recognizing seen Chinese character classes. The set of character class is 3,755 commonly used Chinese characters. The
dataset used for training is the CASIA [31] dataset including
OLHWDB1.0 and OLHWDB1.1. There are totally 2,693,183
samples for training and 224,590 samples for testing. The
training and testing data were produced by different writers
with enormous handwriting styles across individuals. In TaTABLE I
R ESULTS ON CASIA DATASET OF ONLINE HANDWRITTEN C HINESE
CHARACTER RECOGNITION .
Methods
Human Performance
Traditional Benchmark
NET4
TRAN
Reference
Accuracy
[32]
[33]
[4]
–
95.19%
95.31%
96.03%
96.43%
ble I, the human performance on CASIA test set and the
previous benchmark are both listed. NET4 is the proposed
method in [4] which represents the state-of-the-art method
on CASIA dataset and it belongs to non-radical based methods. NET4 achieved an accuracy of 96.03% while TRAN
achieved an accuracy of 96.43%, revealing relative character
error rate reduction of 10%. To be fairly comparable, here
NET4 and TRAN both did not use the sequential dropout
trick as proposed in [4] so the performance of NET4 is
not as good as the best performance presented in [4]. As
explained in the contributions of this study in Section I, the
main difference between radical based method and non-radical
based method for Chinese character recognition is the size
of radical vocabulary is largely less than Chinese character
vocabulary, yielding decrease of redundancy among output
classes and improvement of recognition performance.
B. Performance on recognition of unseen Chinese character
classes (Q2)
The number of Chinese character classes is not fixed as
more and more novel characters are being created. Also,
the overall Chinese character classes are enormous and it is
difficult to train a recognition system that covers them all.
Therefore it is necessary for a recognition system to possess
the capability of recognizing unseen Chinese characters, called
zero-shot learning.
Obviously traditional non-radical based methods are incapable of recognizing these unseen characters since the objective character class has never been seen during training procedure. However TRAN is able to recognize unseen Chinese
characters only if the radicals composing unseen characters
have been seen. To validate the performance of TRAN on
recognizing unseen Chinese character classes, we divide 3755
common Chinese characters into 3255 classes and the other
500 classes. We choose handwritten characters belonging to
3255 classes from original training set as the new training set
and we choose handwritten characters belonging to the other
500 classes from original testing set as the new testing set. By
doing so, both the testing character classes and handwriting
variations have never been seen during training. We explore
different size of training set to train TRAN, ranging from
a { 车 d { 龴 工 } }
a
{
车
d
{
龴
工
}
}
eos
Fig. 4. Examples of attention visualization during the decoding procedure.
The red color on trajectory describes the attention probabilities namely the
lighter color denotes higher attention probabilities and the darker color denotes
lower attention probabilities.
500 to 3255 Chinese character classes and we make sure the
radicals of testing characters are covered in training set.
TABLE II
R ESULTS ON NEWLY DIVIDED TESTING SET BASED ON CASIA DATASET
OF ONLINE HANDWRITTEN UNSEEN C HINESE CHARACTER RECOGNITION .
Train classes
Train samples
Test Accuracy
500
1000
1500
2000
2755
3255
359,036
717,194
1,075,344
1,435,295
1,975,972
2,335,433
10.74%
26.02%
39.35%
50.45%
60.37%
We can see in Table II the recognition accuracy of unseen
Chinese character classes is not available when training set
only contains 500 Chinese character classes. We believe it
is difficult to train TRAN properly to accommodate large
handwriting variations when the number of character classes
is quite small. When the training set contains 3255 character
classes, TRAN achieves a character accuracy of 60.37% which
is a relatively pleasant performance compared with traditional
recognition systems as they can not recognize unseen Chinese
character classes which means their accuracies are definitely
0%. The performance of recognizing unseen Chinese character
classes is not as good as the performance presented in [22]
because the handwritten Chinese characters are much more
ambiguous compared with printed Chinese characters due to
the large handwriting variations.
C. Attention visualization (Q3)
In this section, we show through attention visualization
how TRAN is able to recognize internal radicals and analyze
the two-dimensional spatial structure among radicals. Fig. 4
illustrates an example of attention visualization. Above the
dotted line, there is one Chinese character class and its
corresponding character caption. Below dotted line, there are
images denoting the visualization of attention probabilities
during decoding procedure. We draw the trajectory of input
handwritten Chinese character in a two-dimensional greyscale
image to visualize attention. Below images there are corresponding symbols generated by decoder at each decoding step.
As we can see in Fig. 4, when encountering basic radicals,
the attention model generates the alignment well corresponding to the human intuition. Also, it mainly focus on the ending
of last radical and the beginning of next radical to detect a
spatial structure. Take “d” as an example, by attending to the
ending of last radical and the beginning of next radical, the
attention model detects a top-bottom direction, therefore a topbottom structure is analyzed. Immediately after generating a
spatial structure, the decoder produces a pair of braces “{}”,
which are employed to constrain the two-dimensional structure
in Chinese character caption.
VI. C ONCLUSION AND FUTURE WORK
In this study we introduce TRAN for online handwritten
Chinese character recognition. The proposed TRAN recognizes Chinese character by identifying internal radicals and
analyzing spatial structures among radicals. We show from
experimental results that TRAN outperforms the state-of-theart method on recognition of online handwritten Chinese
characters and possesses the capability of recognizing unseen
Chinese character categories. By visualizing learned attention
probabilities, we can observe the alignments of radicals and
analysis of structures correspond well to human intuition.
R EFERENCES
[1] C. Y. Suen, M. Berthod, and S. Mori, “Automatic recognition of
handprinted characters–the state of the art,” Proceedings of the IEEE,
vol. 68, no. 4, pp. 469–487, 1980.
[2] R. Plamondon and S. N. Srihari, “Online and off-line handwriting
recognition: a comprehensive survey,” IEEE Transactions on pattern
analysis and machine intelligence, vol. 22, no. 1, pp. 63–84, 2000.
[3] C.-L. Liu, S. Jaeger, and M. Nakagawa, “Online recognition of Chinese
characters: the state-of-the-art,” IEEE transactions on pattern analysis
and machine intelligence, vol. 26, no. 2, pp. 198–213, 2004.
[4] X.-Y. Zhang, F. Yin, Y.-M. Zhang, C.-L. Liu, and Y. Bengio, “Drawing
and recognizing Chinese characters with recurrent neural network,”
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017.
[5] W. Yang, L. Jin, D. Tao, Z. Xie, and Z. Feng, “Dropsample: A new
training method to enhance deep convolutional neural networks for largescale unconstrained handwritten chinese character recognition,” Pattern
Recognition, vol. 58, pp. 190–203, 2016.
[6] Z. Zhong, L. Jin, and Z. Xie, “High performance offline handwritten
chinese character recognition using googlenet and directional feature
maps,” in Document Analysis and Recognition (ICDAR), 2015 13th
International Conference on. IEEE, 2015, pp. 846–850.
[7] S.-k. Chang, “An interactive system for Chinese character generation
and retrieval,” IEEE Transactions on Systems, Man, and Cybernetics,
vol. 3, no. 3, pp. 257–265, 1973.
[8] A.-B. Wang and K.-C. Fan, “Optical recognition of handwritten Chinese
characters by hierarchical radical matching method,” Pattern Recognition, vol. 34, no. 1, pp. 15–35, 2001.
[9] L.-L. Ma and C.-L. Liu, “A new radical-based approach to online
handwritten Chinese character recognition,” in Pattern Recognition,
2008. ICPR 2008. 19th International Conference on. IEEE, 2008,
pp. 1–4.
[10] T.-Q. Wang, F. Yin, and C.-L. Liu, “Radical-based Chinese character
recognition via multi-labeled learning of deep residual networks,” in
Document Analysis and Recognition (ICDAR), 2017 14th International
Conference on. IEEE, 2017.
[11] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by
jointly learning to align and translate,” arXiv preprint arXiv:1409.0473,
2014.
[12] A. Graves et al., Supervised sequence labelling with recurrent neural
networks. Springer, 2012, vol. 385.
[13] K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares,
H. Schwenk, and Y. Bengio, “Learning phrase representations using
RNN encoder-decoder for statistical machine translation,” arXiv preprint
arXiv:1406.1078, 2014.
[14] M.-T. Luong, H. Pham, and C. D. Manning, “Effective approaches to attention-based neural machine translation,” arXiv preprint
arXiv:1508.04025, 2015.
[15] K. Xu, J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhudinov, R. Zemel,
and Y. Bengio, “Show, attend and tell: Neural image caption generation
with visual attention,” in International Conference on Machine Learning,
2015, pp. 2048–2057.
[16] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan, “Show and tell: A
neural image caption generator,” in Proceedings of the IEEE conference
on computer vision and pattern recognition, 2015, pp. 3156–3164.
[17] D. Bahdanau, J. Chorowski, D. Serdyuk, P. Brakel, and Y. Bengio, “Endto-end attention-based large vocabulary speech recognition,” in Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEE International
Conference on. IEEE, 2016, pp. 4945–4949.
[18] J. Zhang, J. Du, S. Zhang, D. Liu, Y. Hu, J. Hu, S. Wei, and
L. Dai, “Watch, attend and parse: An end-to-end neural network based
approach to handwritten mathematical expression recognition,” Pattern
Recognition, 2017.
[19] J. Zhang, J. Du, and L. Dai, “Multi-scale attention with dense encoder
for handwritten mathematical expression recognition,” arXiv preprint
arXiv:1801.03530, 2018.
[20] A. Graves, A.-r. Mohamed, and G. Hinton, “Speech recognition with
deep recurrent neural networks,” in Acoustics, speech and signal processing (icassp), 2013 ieee international conference on. IEEE, 2013,
pp. 6645–6649.
[21] J. Zhang, J. Du, and L. Dai, “A gru-based encoder-decoder approach with
attention for online handwritten mathematical expression recognition,” in
Document Analysis and Recognition (ICDAR), 2017 14th International
Conference on. IEEE, 2017.
[22] J. Zhang, Y. Zhu, J. Du, and L. Dai, “RAN: Radical analysis networks for zero-shot learning of Chinese characters,” arXiv preprint
arXiv:1711.01889, 2017.
[23] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105.
[24] A. Madlon-Kay, “cjk-decomp.” [Online]. Available: https://github.com/
amake/cjk-decomp
[25] Y. Bengio, P. Simard, and P. Frasconi, “Learning long-term dependencies
with gradient descent is difficult,” IEEE transactions on neural networks,
vol. 5, no. 2, pp. 157–166, 1994.
[26] J. Zhang, J. Tang, and L.-R. Dai, “RNN-BLSTM based multi-pitch
estimation.” in INTERSPEECH, 2016, pp. 1785–1789.
[27] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evaluation of
gated recurrent neural networks on sequence modeling,” arXiv preprint
arXiv:1412.3555, 2014.
[28] Z. Tu, Z. Lu, Y. Liu, X. Liu, and H. Li, “Modeling coverage for neural
machine translation,” arXiv preprint arXiv:1601.04811, 2016.
[29] M. D. Zeiler, “ADADELTA: an adaptive learning rate method,” arXiv
preprint arXiv:1212.5701, 2012.
[30] K. Cho, “Natural language understanding with distributed representation,” arXiv preprint arXiv:1511.07916, 2015.
[31] C.-L. Liu, F. Yin, D.-H. Wang, and Q.-F. Wang, “CASIA online and
offline Chinese handwriting databases,” in Document Analysis and
Recognition (ICDAR), 2011 International Conference on. IEEE, 2011,
pp. 37–41.
[32] F. Yin, Q.-F. Wang, X.-Y. Zhang, and C.-L. Liu, “ICDAR 2013 Chinese handwriting recognition competition,” in Document Analysis and
Recognition (ICDAR), 2013 12th International Conference on. IEEE,
2013, pp. 1464–1470.
[33] C.-L. Liu, F. Yin, D.-H. Wang, and Q.-F. Wang, “Online and offline handwritten Chinese character recognition: benchmarking on new
databases,” Pattern Recognition, vol. 46, no. 1, pp. 155–162, 2013.
| 1cs.CV
|
The Bivariate Lack-of-Memory Distributions
Gwo Dong Lin, Xiaoling Dou and Satoshi Kuriki
Academia Sinica, Taiwan, Waseda University, Japan
arXiv:1606.05097v5 [math.ST] 16 Dec 2017
and The Institute of Statistical Mathematics, Japan
Abstract. We treat all the bivariate lack-of-memory (BLM) distributions in a unified approach and develop some new general properties of the BLM distributions, including joint
moment generating function, product moments and dependence structure. Necessary and
sufficient conditions for the survival functions of BLM distributions to be totally positive of
order two are given. Some previous results about specific BLM distributions are improved. In
particular, we show that both the Marshall–Olkin survival copula and survival function are
totally positive of all orders, regardless of parameters. Besides, we point out that Slepian’s
inequality also holds true for BLM distributions.
2010 AMS Mathematics Subject Classifications. Primary: 62N05, 62N86, 62H20,
62H05.
Key words and phrases: Lack-of-memory property, bivariate lack-of-memory distributions, Marshall and Olkin’s BVE, Block and Basu’s BVE, Freund’s BVE, likelihood ratio
order, usual stochastic order, hazard rate order, survival function, copula, survival copula,
positive quadrant dependent, total positivity, Slepian’s lemma/inequality.
Postal addresses: Gwo Dong Lin, Institute of Statistical Science, Academia Sinica, Taipei
11529, Taiwan, R.O.C. (E-mail: [email protected])
Xiaoling Dou, Waseda University, 3-4-1 Ohkubo, Shinjuku, Tokyo 169-8555, Japan (E-mail:
[email protected])
Satoshi Kuriki, The Institute of Statistical Mathematics, 10-3 Midoricho, Tachikawa, Tokyo
190-8562, Japan (E-mail: [email protected])
1
1. Introduction
The classical univariate lack-of-memory (LM) property is a remarkable characterization
of the exponential distribution which plays a prominent role in reliability theory, queuing
theory and other applied fields (Feller 1965, Fortet 1977, Galambos and Kotz 1978). The
recent bivariate LM property is, however, shared by the famous Marshall and Olkin’s, Block
and Basu’s as well as Freund’s bivariate exponential distributions, among many others;
see, e.g., Chapter 10 of Balakrishnan and Lai (2009), Chapter 47 of Kotz et al. (2000) and
Kulkarni (2006). These bivariate distributions have been well investigated individually in the
literature. Our main purpose in this paper is, however, to develop in a unified approach some
new general properties of the bivariate lack-of-memory (BLM) distributions which share the
same bivariate LM property.
In Section 2, we first review the univariate and bivariate LM properties, and then summarize the important known properties of the BLM distributions. We derive in Section 3
some new general properties of the BLM distributions, including joint moment generating
function, product moments and stochastic inequalities. The dependence structures of the
BLM distributions are investigated in Section 4. We find necessary and sufficient conditions for the survival functions (and the densities if they exist) of BLM distributions to be
totally positive of order two. Some previous results about specific BLM distributions are
improved. In particular, we show that both the Marshall–Olkin survival copula and survival
function are totally positive of all orders, regardless of parameters. In Section 5, we study the
stochastic comparisons in the family of all BLM distributions and point out that Slepian’s
lemma/inequality for bivariate normal distributions also holds true for BLM distributions.
2. Lack-of-Memory Property
We first review the well-known univariate lack-of-memory property. Let X be a nonnegative random variable with distribution function F. Then F satisfies (multiplicative) Cauchy’s
functional equation
F (x + y) = F (x)F (y), x ≥ 0, y ≥ 0,
(1)
where F (x) = 1 − F (x) = Pr(X > x), if and only if F (0) = 1 (X degenerates at 0) or F (x) =
1−exp(−λx), x ≥ 0, for some constant λ > 0, denoted by X ∼ Exp(λ) (X has an exponential
2
distribution with positive parameter λ). If X is the lifetime of a system with positive survival
function F , then Eq. (1) is equivalent to
Pr(X > x + y| X > y) = Pr(X > x), x ≥ 0, y ≥ 0.
(2)
This means that the conditional probability of a system surviving to time x+y given surviving
to time y is equal to the unconditional probability of the system surviving to time x. Namely,
the failure performance of the system does not depend on the past, given its present condition.
In such a case (2), we say that the distribution F lacks memory at each point y. So Eq. (1)
is called the LM property or memoryless property of F.
For simplicity, we consider only positive random variable X ∼ F from now on. Then, the
LM property (1) holds true iff X ∼ Exp(λ) for some λ > 0.
We next consider the bivariate LM property. Let the positive random variables X and Y
have joint distribution H with marginals F and G. Namely, (X, Y ) ∼ H, X ∼ F, Y ∼ G.
Moreover, denote the survival function of H by
H(x, y) ≡ Pr(X > x, Y > y) = 1 − F (x) − G(y) + H(x, y), x, y ≥ 0.
An intuitive extension of the LM property (2) to the bivariate case is the strict BLM property:
Pr(X > x + s, Y > y + t| X > s, Y > t) = Pr(X > x, Y > y), x, y, s, t ≥ 0
(H lacks memory at each pair (s, t)), which is equivalent to
H(x + s, y + t) = H(x, y)H(s, t), ∀ x, y, s, t ≥ 0,
(3)
if the survival function H is positive. In a two-component system, this means as before that
the conditional probability of two components surviving to times (x+s, y +t) given surviving
to times (s, t) is equal to the unconditional probability of these two components surviving
to times (x, y). But Eq. (3) has only one solution (Marshall and Olkin 1967, p. 33), namely,
the independent bivariate exponential distribution with survival function
H(x, y) = exp[−(λx + δy)], x, y ≥ 0,
for some constants λ, δ > 0; in other words, X and Y are independent random variables and
X ∼ Exp(λ), Y ∼ Exp(δ) for some positive parameters λ, δ.
3
In their pioneering paper, Marshall and Olkin (1967) considered instead the weaker BLM
property (with s = t)
Pr(X > x + t, Y > y + t| X > t, Y > t) = Pr(X > x, Y > y), x, y, t ≥ 0
(H lacks memory at each equal pair (t, t)), and solved the functional equation
H(x + t, y + t) = H(x, y)H(t, t), ∀ x, y, t ≥ 0.
(4)
It turns out that for given (X, Y ) ∼ H with marginals F, G on (0, ∞), H satisfies the BLM
property (4) iff its survival function is of the form
(
e−θy F (x − y), x ≥ y ≥ 0
H(x, y) =
e−θx G(y − x), y ≥ x ≥ 0,
(5)
where θ is a positive constant (see also Barlow and Proschan 1981, p. 130).
For convenience, denote by BLM(F, G, θ) the BLM distribution H with marginals F, G,
parameter θ > 0 and survival function H in (5), and denote by BLM the family of all BLM
distributions, namely,
BLM = {H : H = BLM(F, G, θ), where θ > 0, and F, G are marginal distributions}.
Theorem 1 below summarizes some important known properties of the BLM distributions;
for more details, see Marshall and Olkin (1967), Block and Basu (1974), Block (1977), and
Ghurye and Marshall (1984). For convenience, denote a∨b = max{a, b} and a∧b = min{a, b}.
Theorem 1. Let (X, Y ) ∼ H = BLM(F, G, θ) ∈ BLM. Then the following statements are
true.
(i) The marginals F , G have densities f , g, respectively. Moreover, the right-hand derivatives
f (x) = limε→0+ [F (x+ ε) −F (x)]/ε and g(x) = limε→0+ [G(x+ ε) −G(x)]/ε exist for all x ≥ 0,
which are right-continuous and are of bounded variation on [0, ∞).
(ii) Pr(X − Y > t) = F (t) − f (t)/θ and Pr(Y − X > t) = G(t) − g(t)/θ for all t ≥ 0.
(iii) Both eθx f (x) and eθx g(x) are increasing (nondecreasing) in x ≥ 0.
(iv) F (x) + G(x) ≥ 1 − exp(−θx), x ≥ 0.
(v) X ∧ Y ∼ Exp(θ) and is independent of X − Y.
(vi) f (0) ∨ g(0) ≤ θ ≤ f (0) + g(0).
4
(vii) f ′ (x) + θf (x) ≥ 0, g ′ (x) + θg(x) ≥ 0, x ≥ 0, if f and g are differentiable.
Remark 1. Some of the above necessary conditions (i)–(vii) also play as sufficient conditions
for (X, Y ) to obey a BLM distribution. For example, in addition to the above conditions (vi)
and (vii), assume that the marginal densities are absolutely continuous, then the H in (5) is
a bona fide survival function. This is a slight modification of Theorem 5.1 of Marshall and
Olkin (1967) who required (vi′ ) [f (0) + g(0)]/2 ≤ θ ≤ f (0) + g(0) instead of (vi) above. Note
that conditions (vi) and (vi′ ) are different unless f (0) = g(0), and that (vi) is a consequence of
(iii) and (iv) (see Corollary 2(i) below and Ghurye and Marshall 1984, p. 792). On the other
hand, the condition (v) together with continuous marginals F, G also implies that (X, Y )
has a BLM distribution (Block 1977, p. 810). It is interesting to recall that for independent
nondegenerate random variables X and Y , the above independence of X ∧ Y and X − Y is
a characterization of the exponential/geometric distributions under suitable conditions (see
Ferguson 1964, 1965, Crawford 1966, and Rao and Shanbhag 1994). Namely, in general,
the BLM distributions share the same independence property of X ∧ Y and X − Y with
independent exponential/geometric random variables.
Remark 2. There are some more observations: (a) Pr(X = Y ) = [f (0) + g(0)]/θ − 1 by
the above (ii), (b) at least one of f (0) and g(0) is positive, (c) the survival function H in
(5) is purely singular (i.e., X = Y almost surely) iff θ = [f (0) + g(0)]/2 iff f (0) = g(0) = θ
(because f (0) 6= g(0) implies θ > [f (0) + g(0)]/2 by (vi)), and (d) H is absolutely continuous
(i.e., X 6= Y almost surely) iff the marginal densities together satisfy f (0) + g(0) = θ (see
Ghurye and Marshall 1984, p. 792). In view of the above results, the survival function (5) of
H = BLM(F, G, θ) can be rewritten as the convex combination of two extreme ones:
f (0) + g(0)
f (0) + g(0)
− 1 Hs (x, y), x, y ≥ 0,
H(x, y) = 2 −
Ha (x, y) +
θ
θ
where Ha is absolutely continuous and Hs is purely singular with survival function Hs (x, y) =
exp[−θ max{x, y}], x, y ≥ 0. Clearly, the parameter θ regulates H between Ha and Hs .
On the other hand, Ghurye and Marshall (1984, Section 3) gave an interesting random
decomposition of (X, Y ) ∼ H ∈ BLM and represented H as a Laplace–Stieltjes integral by
another bivariate survival function. See also Ghurye (1987) and Marshall and Olkin (2015)
for further generalizations of the BLM distributions.
5
Remark 3. Kulkarni (2006) proposed an interesting and useful approach to construct some
BLM distributions by starting with marginal failure rate functions. First, choose two realvalued functions r1 , r2 and a constant θ satisfying the following (modified) conditions:
(a) The functions ri , i = 1, 2, are absolutely continuous on [0, ∞) and θ > 0.
(b) 0 ≤ ri (x) ≤ θ, x ≥ 0, i = 1, 2.
R∞
(c) 0 ri (x)dx = ∞, i = 1, 2.
(d) ri (x)(θ − ri (x)) + ri′ (x) ≥ 0, x ≥ 0, i = 1, 2.
(e) r1 (0) + r2 (0) ≥ θ.
Then set F (x) = exp(−
Rx
0
r1 (t)dt), x ≥ 0, and G(y) = exp(−
Ry
0
r2 (t)dt), y ≥ 0. In this way,
the H defined through (5) is a bona fide BLM distribution because the above conditions
(a)–(e) together imply that conditions (vi) and (vii) in Theorem 1 hold true (see Remark
1). Conversely, under the above smoothness conditions on ri and the setting of F and G,
if H in (5) is a BLM distribution, then its marginal failure rate functions ri should satisfy
conditions (b)–(e) from which some properties in Theorem 1 follow immediately (Kulkarni
2006, Proposition 1).
We now recall three important BLM distributions in the literature. For more details, see,
e.g., Chapter 10 of Balakrishnan and Lai (2009).
Example 1. Marshall and Olkin’s (1967) bivariate exponential distribution (BVE)
If both marginals F and G are exponential, then the BLM(F, G, θ) ∈ BLM defined in (5)
reduces to the Marshall–Olkin BVE with survival function of the form
H(x, y) = exp[−λ1 x − λ2 y − λ12 max{x, y}]
λ12
λ1 + λ2
Ha (x, y) +
Hs (x, y), x, y ≥ 0,
≡
λ
λ
(6)
(7)
where λ1 , λ2 , λ12 are positive constants, λ = λ1 + λ2 + λ12 , and Ha , Hs (written explicitly
below) are absolutely continuous and singular bivariate distributions, respectively.
In practice, the Marshall–Olkin BVE arises from a shock model for a two-component
system. Formally, the lifetimes of two components are (X, Y ) = (X1 ∧ X3 , X2 ∧ X3 ), where
X1 ∼ Exp(λ1 ), X2 ∼ Exp(λ2 ) and X3 ∼ Exp(λ12 ) are independent. So they have a joint
survival function H defined in (6). The singular part in (7) is identified by the conditional
probability: Hs (x, y) = Pr(X > x, Y > y| X3 ≤ X1 ∧ X2 ) = exp[−λ max{x, y}], while the
6
absolutely continuous part Ha is calculated from H and Hs via (7) (see the next example).
Example 2. Block and Basu’s (1974) bivariate exponential distribution
The Block–Basu BVE is actually the absolute continuous part Ha of Marshall–Olkin BVE
in (7) and has a joint density of the form
( λ2 λ(λ1 +λ12 )
exp[−(λ1 + λ12 )x − λ2 y], x ≥ y > 0
λ1 +λ2
h(x, y) =
λ1 λ(λ2 +λ12 )
exp[−λ1 x − (λ2 + λ12 )y], y > x > 0,
λ1 +λ2
(8)
where λ1 , λ2 , λ12 > 0, and λ = λ1 + λ2 + λ12 . Its survival function is equal to
H(x, y) = Ha (x, y)
λ
λ12
=
exp[−λ1 x − λ2 y − λ12 max{x, y}] −
exp[−λ max{x, y}], x, y ≥ 0.
λ1 + λ2
λ1 + λ2
Note that in this case, the marginals are not exponential but rather negative mixtures
of two exponentials. Specifically, F (x) =
and G(y) =
λ
λ1 +λ2
exp[−(λ2 + λ12 )y] −
λ
λ1 +λ2
λ12
λ1 +λ2
exp[−(λ1 + λ12 )x] −
λ12
λ1 +λ2
exp(−λx), x ≥ 0,
exp(−λy), y ≥ 0.
Example 3. Freund’s (1961) bivariate exponential distribution
The Freund BVE has a joint density of the form
′
α β exp[−(α + β − α′ )y − α′x], x ≥ y > 0
h(x, y) =
αβ ′ exp[−(α + β − β ′ )x − β ′ y], y > x > 0,
(9)
where α, α′, β, β ′ > 0. If α + β > α′ ∨ β ′ , its survival function is equal to
(
β
α−α′
exp[−(α + β − α′ )y − α′ x] + α+β−α
x≥y≥0
′ exp[−(α + β)x],
α+β−α′
H(x, y) =
β−β ′
α
exp[−(α + β − β ′ )x − β ′ y] + α+β−β
y ≥ x ≥ 0.
′ exp[−(α + β)y],
α+β−β ′
It worths noting that by choosing α =
λ1 λ
,
λ1 +λ2
β=
λ2 λ
,
λ1 +λ2
α′ = λ1 + λ12 and β ′ = λ2 + λ12 ,
Freund’s BVE (9) reduces to Block and Basu’s BVE (8).
3. New General Properties of BLM Distributions
Let (X, Y ) ∼ H = BLM(F, G, θ) ∈ BLM with marginals F and G on (0, ∞), parameter
θ > 0 and survival function (5). Denote the Laplace-Stieltjes transform of X (Y, resp.) by
LX (LY , resp.), and that of (X, Y ) by L. Then we have
Theorem 2. The Laplace-Stieltjes transform of (X, Y ) ∼ H = BLM(F, G, θ) ∈ BLM is
L(s, t) ≡ E e−sX−tY =
θ
1
[(θ + s)LX (s) + (θ + t)LY (t)] −
, s, t > 0.
θ+s+t
θ+s+t
7
To prove Theorem 2, we need the following lemma due to Lin et al. (2016).
Lemma 1. Let (X, Y ) ∼ H defined on R2+ = [0, ∞) × [0, ∞). Then the Laplace-Stieltjes
transform of (X, Y ) is equal to
Z ∞Z ∞
L(s, t) = st
H(x, y)e−sx−ty dxdy − 1 + LX (s) + LY (t), s, t ≥ 0.
0
0
Proof of Theorem 2. We have to calculate the double integral
Z ∞Z ∞
ZZ
ZZ
−sx−ty
H(x, y)e
dxdy =
+
≡ A1 + A2 ,
0
0
x≥y
y≥x
where, by changing variables and by integration by parts,
Z ∞
Z ∞
Z ∞
Z ∞
−(θ+s+t)y
−(θ+t)y
−sx
e
e−sz F (z)dzdy
A1 =
e
e F (x − y)dxdy =
0
0
0
y
Z ∞
1
1
1
1
−sz
=
F (z)de
=
−
(1 − LX (s)) ,
θ+s+t
s 0
θ+s+t s
and similarly,
1
1
(1 − LY (t)) .
A2 =
θ+s+t t
Lemma 1 together with the above A1 and A2 completes the proof.
Denote the moment generating function (mgf) of X (Y, resp.) by MX (MY , resp.), and
that of (X, Y ) by M. Then we have the following general result.
Theorem 3. Let (X, Y ) ∼ H = BLM(F, G, θ) ∈ BLM and let r, s be real numbers such
that s + t < θ. Then the mgf of (X, Y ) is
M(s, t) ≡ E esX+tY =
1
θ
[(θ − s)MX (s) + (θ − t)MY (t)] −
,
θ−s−t
θ−s−t
provided the expectations (mgfs) exist.
To prove Theorem 3, we need instead the following lemma due to Lin et al. (2014).
Lemma 2. Let (X, Y ) ∼ H defined on R2+ . Let α and β be two increasing and left-continuous
functions on R+ . Then the expectation of the product α(X)β(Y ) is equal to
Z ∞Z ∞
E[α(X)β(Y )] =
H(x, y)dα(x)dβ(y) − α(0)β(0) + α(0)E[β(Y )] + β(0)E[α(X)],
0
0
provided the expectations exist.
8
Proof of Theorem 3. Case (i): s, t ≥ 0. Let α(x) = esx and β(y) = ety in Lemma 2, then
Z ∞Z ∞
M(s, t) = st
H(x, y)esx+ty dxdy − 1 + MX (s) + MY (t).
0
0
We have to calculate the double integral
Z ∞Z ∞
ZZ
sx+ty
H(x, y)e
dxdy =
0
0
+
x≥y
ZZ
≡ B1 + B2 ,
y≥x
where, as before,
1
1
1
1
B1 =
(−1 + MX (s)) and B2 =
(−1 + MY (t)) .
θ−s−t s
θ−s−t t
Lemma 2 together with the above B1 and B2 completes the proof of Case (i).
Case (ii): s ≥ 0, t < 0. To apply Lemma 2, set α(x) = esx and β(y) = 1 − ety . Then both α
and β are increasing functions on R+ and E[α(X)β(Y )] = MX (s) − M(s, t). Therefore,
Z ∞Z ∞
H(x, y)esx+ty dxdy.
M(s, t) = MX (s) − E[α(X)β(Y )] = MX (s) − 1 + MY (t) + st
0
0
As before, we carry out the above double integral and complete the proof of Case (ii).
Case (iii): s < 0, t ≥ 0. Set α(x) = 1 − esx and β(y) = ety in Lemma 2. The remaining
proof is similar to that of Case (ii) and is omitted.
Case (iv): s, t < 0. This case was treated in Theorem 1. The proof is completed.
Next, we consider the product moments of BLM distributions.
Theorem 4. For positive integers i and j, the product moment E[X i Y j ] of (X, Y ) ∼ H =
BLM(F, G, θ) ∈ BLM is of the form
i
j
E[X Y ] = i j
i−1
X
k=0
j−1
X
1
1
i − 1 Γ(j + k)
j − 1 Γ(i + k)
i−k
E[X ] + i j
E[Y j−k ],
i+k
k
k
i−k
θj+k
j
−
k
θ
k=0
provided the expectations exist.
The first product moment has a neat representation in terms of marginal means and the
parameter θ, from which we can calculate Pearson’s correlation of BLM distributions.
Corollary 1. E[XY ] = 1θ (E[X] + E[Y ]) provided the expectations exist.
To prove Theorem 4 above, we will apply the following lemma due to Lin et al. (2014).
9
Lemma 3. Let (X, Y ) ∼ H defined on R2+ , and let the expectations E[X r Y s ], E[X r ] and
E[Y s ] be finite for some positive real numbers r and s. Then the product moment
Z ∞Z ∞
r s
E[X Y ] = rs
H(x, y)xr−1 y s−1dxdy.
0
0
Proof of Theorem 4. We have to calculate the double integral
Z ∞Z ∞
ZZ
ZZ
i−1 j−1
H(x, y)x y dxdy =
+
≡ C1 + C2 ,
0
0
x≥y
y≥x
where, by changing variables and by integration by parts,
Z ∞
Z ∞
Z ∞
Z
−θy j−1
−θy j−1
i−1
C1 =
e y
e y
x F (x − y)dxdy =
0
i−1
X
0
y
∞
(y + z)i−1 F (z)dzdy
0
∞
∞
i−1
y j−1+k e−θy
z i−1−k F (z)dzdy
k
0
0
k=0
X
Z ∞
i−1
i−1
X
i − 1 Γ(j + k)
1
1
i − 1 Γ(j + k)
i−k
E[X i−k ],
=
=
F (z)dz
j+k
j+k
k
θ
i
−
k
i
−
k
θ
k
0
k=0
k=0
=
Z
Z
and similarly,
C2 =
j−1
X
k=0
1
j − 1 Γ(i + k)
E[Y j−k ].
k
j−k
θi+k
Finally, Lemma 3 together with the above C1 and C2 completes the proof.
For moment generating functions of some specific BLM distributions, see Chapter 47 of
Kotz et al. (2000), while for product moments of such distributions, see Nadarajah (2006).
For the next and later results, we need some notations in reliability theory. For random
variables X ∼ F and Y ∼ G, we say that X is smaller than Y in the usual stochastic order
(denoted by X ≤st Y ) if F (x) ≤ G(x) for all x, that X is smaller than Y in the hazard rate
order (denoted by X ≤hr Y ) if G(x)/F (x) is increasing in x, and that X is smaller than
Y in the reversed hazard rate order (denoted by X ≤rh Y ) if G(x)/F (x) is increasing in x.
Suppose F and G have densities f and g, respectively. Then we say that X is smaller than
Y in the likelihood ratio order (denoted by X ≤ℓr Y ) if g(x)/f (x) is increasing in x. For
more definitions of the related stochastic orders, see, e.g., Müller and Stoyan (2002), Shaked
and Shanthikumar (2007), Lai and Xie (2006) as well as Kayid et al. (2016). The latter
studied stochastic comparisons of the age replacement models.
10
On the other hand, for a distribution F itself we define the notions of increasing failure
rate (IFR), decreasing failure rate (DFR), increasing failure rate in average (IFRA), and
decreasing failure rate in average (DFRA) as follows. We say that
(a) F is IFR (DFR, resp.) if − log F (x) is convex (concave, resp.) in x ≥ 0, and
(b) F is IFRA (DFRA, resp.) if −(1/x) log F (x) is increasing (decreasing, resp.) in x > 0,
α
or, equivalently, F (x) ≤ (≥, resp.) F (αx) for all α ∈ (0, 1) and x ≥ 0. (See Barlow and
Proschan 1981, Chapters 3 and 4.)
The bivariate IFRA and bivariate DFRA distributions H can be defined similarly:
α
H is bivariate IFRA (DFRA, resp.) if H (x, y) ≤ (≥, resp.) H(αx, αy) for all α ∈ (0, 1)
and x, y ≥ 0 (see Block and Savits 1976, 1980). It worths mentioning that there are some
other definitions of bivariate IFRA distributions that all extend the univariate case (see, e.g.,
Esary and Marshall 1979 or Shaked and Shanthikumar 1988).
Using reliability language, we have the following useful results. Especially, Theorem 5(iii)
means that in the BLM family, positive bivariate aging plays in favor of positive univariate
aging in the sense of IFRA, and vice versa. This is in general not true even under the condition of positive dependence for lifetimes; see Bassan and Spizzichino (2005, Remark 6.8),
which analyzed the relations among univariate and bivariate agings and dependence.
Theorem 5. Let (X, Y ) ∼ H = BLM(F, G, θ) ∈ BLM and Z ∼ Exp(θ). Then
(i) Z ≤ℓr X and Z ≤ℓr Y ;
(ii) Z ≤st X and Z ≤st Y ; Z ≤hr X and Z ≤hr Y ; Z ≤rh X and Z ≤rh Y ;
(iii) (X, Y ) has a bivariate IFRA distribution iff both marginals F and G are IFRA;
(iv) (X, Y ) has a bivariate DFRA distribution iff both marginals F and G are DFRA.
Proof. Part (i) follows immediately from Theorem 1(iii) (see Ghurye and Marshall 1984),
while part (ii) follows from the fact that the likelihood ratio order is stronger than the
usual stochastic order, hazard rate order and reversed hazard rate order (Müller and Stoyan
α
2002, pp. 12–13). Part (iii) holds true by verifying that H(αx, αy) ≥ H (x, y) ∀α ∈ (0, 1), x, y
α
α
≥ 0, if, and only if, (a) F (αx) ≥ F (x) ∀α ∈ (0, 1), x ≥ 0, and (b) G(αy) ≥ G (y) ∀α ∈
(0, 1), y ≥ 0. The proof of part (iv) is similar.
Applying the above stochastic inequalities, we can simplify the proof of some previous
11
known results. For example, we have
Corollary 2. Let (X, Y ) ∼ H = BLM(F, G, θ) ∈ BLM. Then the following statements
are true.
(i) Both hazard rates of marginals F, G are bounded by θ and hence θ ≥ f (0) ∨ g(0).
(ii) Both the functions F (− 1θ log(1 − t)) and G(− θ1 log(1 − t)) are convex in t ∈ [0, 1), and
hence f ′ (x) + θf (x) ≥ 0, g ′ (x) + θg(x) ≥ 0, x ≥ 0, if f and g are differentiable.
(iii) Let SF , SG be the supports of marginals F, G with densities f, g, respectively. Then
SF = [aF , ∞), SG = [aG , ∞) for some nonnegative constants aF , aG with aF aG = 0, and f, g
are positive on (aF , ∞), (aG , ∞), respectively.
(iv) If H is not absolutely continuous, then f (0) > 0, g(0) > 0, and hence aF = aG = 0.
Proof. Part (i) follows from the facts Z ≤hr X and Z ≤hr Y , where Z ∼ Exp(θ), while
part (ii) is due to the probability-probability plot characterization for Z ≤ℓr X and Z ≤ℓr Y
(see Theorem 1.4.3 of Müller and Stoyan 2002). Part (iii) follows from the facts Z ≤ℓr X,
Z ≤ℓr Y and Theorem 1(iv), because the latter implies that at least one of the left extremities
aF and aG of marginal distributions should be zero. Finally, to prove part (iv), we note that
Pr(X − Y > 0) = 1 − f (0)/θ and Pr(Y − X > 0) = 1 − g(0)/θ by Theorem 1(ii) (see Ghurye
and Marshall 1984, p. 789). So if H is not absolutely continuous, Pr(X = Y ) > 0, and hence
f (0) = θ Pr(X ≤ Y ) > 0 and g(0) = θ Pr(Y ≤ X) > 0. The proof is complete.
4. Dependence Structures of BLM Distributions
Recall that a bivariate distribution H with marginals F and G is positively quadrant
dependent (PQD) if
H(x, y) ≥ F (x)G(y) ∀ x, y ≥ 0, or, equivalently, H(x, y) ≥ F (x)G(y) ∀ x, y ≥ 0,
which implies that H has a nonnegative covariance by Hoeffding representation for covariance
(see, e.g., Lin et al. 2014, p. 2). A stronger (positive dependence) property than the PQD is
the total positivity defined below. For a nonnegative function K on the rectangle (a, b)×(c, d)
(or on the product of two subsets of R), we say that K(x, y) is totally positive of order r
(TPr , r ≥ 2) in x and y if for each fixed s ∈ {2, 3, . . . , r} and for all a < x1 < x2 < · · · <
xs < b and c < y1 < y2 < · · · < ys < d, the determinant of the s × s matrix (K(xi , yj ))
12
is nonnegative. The function K is said to be TP∞ if it is TPr for any order r ≥ 2 (Karlin
1968).
The total positivity plays an important role on various concepts of bivariate dependence
(see, e.g., Shaked 1977 and Lee 1985). Moreover, applying total positivity of the bivariate
distribution or its survival function, we can derive some useful probability inequalities, among
many applications to applied fields including statistics, reliability and economics (see, e.g.,
Gross and Richards 1998, 2004, and Karlin and Proschan 1960). Especially, the latter studied
the totally positive kernels that arise from convolutions of Pólya type distributions.
We now characterize the TP2 property of the survival functions of BLM distributions.
Theorem 6. Let (X, Y ) ∼ H = BLM(F, G, θ) ∈ BLM. Then the survival function H
is TP2 iff the marginal distributions F and G are IFR and together satisfy F (x)G(x) ≤
exp(−θx), x ≥ 0.
Proof. We define the cross-product ratio of H:
r ≡ r(x1 , x2 ; y1 , y2 ) =
H(x1 , y1 )H(x2 , y2 )
, 0 < x1 < x2 , 0 < y1 < y2 .
H(x1 , y2 )H(x2 , y1 )
Then, by definition, H is TP2 iff r(x1 , x2 ; y1 , y2 ) ≥ 1 for all 0 < x1 < x2 , 0 < y1 < y2 .
(Necessity) Suppose that H is TP2 . Then for all 0 < x1 = y1 < x2 = y2 , we have
r = r(x1 , x2 ; x1 , x2 ) =
exp(−θ(x2 − x1 ))
H(x1 , x1 )H(x2 , x2 )
=
≥ 1.
H(x1 , x2 )H(x2 , x1 )
F (x2 − x1 )G(x2 − x1 )
This implies that F (x)G(x) ≤ exp(−θx), x ≥ 0. Next, we prove that the marginal
distribution G is IFR. Note that the following statements are equivalent:
(i) g(y)/G(y) is increasing in y ≥ 0,
G(y+t)
is decreasing in t ∈ (0, ∞) for each y ≥ 0 (Barlow and Proschan
G(t)
G(t)
is increasing in t ∈ (0, ∞) for each y ≥ 0,
(iii) G(y+t)
2)
(iv) G(y−x
is increasing in y > x2 for any fixed 0 < x1 < x2 ,
G(y−x1 )
1 −x1 )G(y2 −x2 )
∗
≡ G(y
≥ 1 for all 0 < x1 < x2 < y1 < y2 .
(v) the ratio rG
G(y2 −x1 )G(y1 −x2 )
(ii)
1981, p. 54),
∗
The latter is true because in this case rG
= r(x1 , x2 ; y1 , y2) ≥ 1 by (5) and the assumption.
Similarly, we can prove that F is IFR because the ratio
rF∗ ≡
F (x1 − y1 )F (x2 − y2 )
≥ 1 for all 0 < y1 < y2 < x1 < x2 .
F (x2 − y1 )F (x1 − y2 )
13
(Sufficiency) Suppose that the marginal distributions F and G are IFR and together
satisfy F (x)G(x) ≤ exp(−θx), x ≥ 0. Then we want to prove that H is TP2 , that is, for all
0 < x1 < x2 , 0 < y1 < y2 , the cross-product ratio r = r(x1 , x2 ; y1 , y2) ≥ 1. Without loss of
generality, we consider only three possible cases below,
(a) 0 < x1 ≤ x2 ≤ y1 ≤ y2 , (b) 0 < x1 ≤ y1 ≤ x2 ≤ y2 , (c) 0 < x1 ≤ y1 ≤ y2 ≤ x2 ,
because the remaining cases can be proved by exchanging the roles of F and G.
For case (a), we have r ≥ 1 by the equivalence relations shown in the necessity part and
by the continuity of H when x2 = y1 . For case (b), the cross-product ratio
r=
exp(−θx2 )G(y1 − x1 )G(y2 − x2 )
G(y1 − x1 )G(y2 − x2 )G(x2 − y1 )
≥
,
exp(−θy1 )G(y2 − x1 )F (x2 − y1 )
G(y2 − x1 )
because F (x2 − y1 )G(x2 − y1 ) ≤ exp(−θ(x2 − y1 )) by the assumption. Recall that any
IFR distribution is new better than used (Barlow and Proschan 1981, p. 159). Therefore,
G(x + y) ≤ G(x)G(y) for all x, y ≥ 0, and hence the last r ≥ 1. Similarly, for case (c),
r=
G(y1 − x1 )G(y2 − y1 ) F (y2 − y1 )F (x2 − y2 )
exp(−θy2 )G(y1 − x1 )F (x2 − y2 )
≥
×
≥ 1,
exp(−θy1 )G(y2 − x1 )F (x2 − y1 )
G(y2 − x1 )
F (x2 − y1 )
by the assumptions. This completes the proof.
Recall also that for any bivariate distribution H with marginals F and G, there exist a
copula C (a bivariate distribution with uniform marginals on [0, 1]) and a survival copula
Ĉ such that H(x, y) = C(F (x), G(y)) and H(x, y) = Ĉ(F (x), G(y)) for all x, y ∈ R ≡
(−∞, ∞). Namely, C links H and (F, G), while Ĉ links H and (F , G).
Corollary 3. Let (X, Y ) ∼ H = BLM(F, G, θ) ∈ BLM. Then the survival copula Ĉ of
H is TP2 iff the marginal distributions F and G are IFR and together satisfy F (x)G(x) ≤
exp(−θx), x ≥ 0.
Proof. Since the marginal F is absolutely continuous on the support [aF , ∞) with positive
density f on (aF , ∞) (see Corollary 2(iii) above), F is strictly increasing and continuous
on (aF , ∞). Similarly, the marginal G is strictly increasing and continuous on (aG , ∞). By
Theorem 6, it suffices to prove that H is TP2 on (aF , ∞) ×(aG , ∞) iff its survival copula Ĉ is
TP2 on (0, 1)2. Recall the facts (i) H(x, y) = Ĉ(F (x), G(y)), (x, y) ∈ (aF , ∞) × (aG , ∞), (ii)
Ĉ(u, v) = H(F
−1
(u), G
−1
(v)), u, v ∈ (0, 1), where F
14
−1
,G
−1
are inverse functions of F , G,
respectively, and (iii) all the functions F , G, F
−1
and G
−1
are decreasing. The required
result then follows immediately (see, e.g., Lemma 5(ii) below).
The counterpart of TP2 property is the reverse regular of order two (RR2 ). For a nonnegative function K on (a, b) × (c, d), we say that K is RR2 if the determinant of the 2 × 2
matrix (K(xi , yj )) is non-positive for all a < x1 < x2 < b and c < y1 < y2 < d (see, e.g.,
Esna–Ashari and Asadi 2016 for examples of RR2 joint densities and survival functions).
Mimicking the proof of Theorem 6, we conclude that for H = BLM(F, G, θ) ∈ BLM, the
survival function H is RR2 iff the survival copula Ĉ of H is RR2 iff the marginal distributions F and G are DFR and satisfy F (x)G(x) ≥ exp(−θx), x ≥ 0. To construct such a BLM
distribution with RR2 survival function, we first consider the Pareto Type II distribution
(or Lomax distribution) F with density function f (x) = (α/β)(1 + x/β)−(α+1) , x ≥ 0, and
survival function F (x) = (1 + x/β)−α , x ≥ 0, where α, β > 0. Then choose the parameters:
α ≥ 1, β > 0 and θ = (α + 1)/β. It can be checked that the H defined in (5) with G = F is
a bona fide survival function, and is RR2 if α = 1.
It is seen that all the conditions in Theorem 6 are satisfied by the Marshall–Olkin BVE.
Therefore, the survival function and survival copula of the Marshall–Olkin BVE are both
TP2 , regardless of parameters; a more general result will be given in Theorem 8 below.
We next characterize, by a different approach, the TP2 property of some joint densities of
absolutely continuous BLM distributions.
Theorem 7. Let H = BLM(F, G, θ) ∈ BLM be absolutely continuous and have joint
density function h. Suppose that the marginal density functions f and g are three times
differentiable on (0, ∞) and that θf (0+ ) + f ′ (0+ ) = θg(0+ ) + g ′(0+ ) is finite. Assume further
the functions
h1 (x|θ) ≡ θf (x) + f ′ (x) > 0, x > 0, and h2 (y|θ) ≡ θg(y) + g ′(y) > 0, y > 0.
Then the joint density function h is TP2 iff the marginal densities satisfy (i) (h′i (x|θ))2 ≥
h′′i (x|θ)hi (x|θ), x > 0, i = 1, 2, and (ii) h1 (x|θ)h2 (x|θ) ≤ h21 (0+ |θ) exp(−θx), x > 0.
To prove this theorem, we need the concept of local dependence function and the following lemma, in which part (ii) is essentially due to Holland and Wang (1987, p. 872). An
alternative (complete) proof of part (ii) is provided below. In their proof, Holland and Wang
15
(1987) assumed implicitly the integrability of the local dependence function, while Kemperman (1977, p. 329) gave without proof the same result under continuity (smoothness)
condition (see also Newman 1984). Wang (1993) proved that a positive continuous bivariate
density on a Cartesian product (a, b) × (c, d) is uniquely determined by its marginal densities
and local dependence function when the latter exists and is integrable. On the other hand,
Jones (1996, 1998) investigated the bivariate distributions with constant local dependence.
Lemma 4. Let K be a positive function on D = (a, b) × (c, d). Then we have
(i) K is TP2 on D iff log K is 2-increasing;
(ii) K is TP2 on D iff the local dependence function γK (x, y) ≡
∂2
∂x∂y
log K(x, y) ≥ 0 on D,
provided the second-order partial derivatives exist.
Proof. Part (i) is trivial by the definition of 2-increasing functions (see Nelsen 2006, p. 8),
and part (ii) follows from part (i) and the fact that under the smoothness assumption,
log K is 2-increasing iff the local dependence function γK (x, y) ≥ 0. To prove part (ii) directly, note that the following statements are equivalent: (a)
(b)
∂
∂y
∂2
∂x∂y
log K(x, y) ≥ 0 on D,
log[K(x2 , y)/K(x1 , y)] ≥ 0 for all y and for all x1 < x2 , (c) log[K(x2 , y)/K(x1 , y)] is
increasing in y for all x1 < x2 , (d) K(x2 , y)/K(x1, y) is increasing in y for all x1 < x2 , (e)
K(x2 , y2)/K(x1 , y2 ) ≥ K(x2 , y1 )/K(x1 , y1 ) for all y1 < y2 , x1 < x2 , (f) the cross-product
ratio of K satisfies: K(x1 , y1 )K(x2 , y2 )/[K(x1 , y2 )K(x2 , y1 )] ≥ 1 for all x1 < x2 , y1 < y2 ,
and (g) the function K is TP2 on D. The proof is complete.
Proof of Theorem 7. By the assumptions, the joint density function of H is of the form
−θy
e h1 (x − y|θ), x ≥ y
h(x, y) =
e−θx h2 (y − x|θ), x ≤ y,
where hi (0|θ) ≡ hi (0+ |θ), i = 1, 2. For x 6= y, the local dependence function of h is
[h′ (x−y|θ)]2 −h′′ (x−y|θ)h (x−y|θ)
1
1
1
, x>y
2
h21 (x−y|θ)
∂
γh (x, y) =
log h(x, y) =
[h′2 (y−x|θ)]2 −h′′2 (y−x|θ)h2 (y−x|θ) , x < y.
∂x∂y
h22 (y−x|θ)
Therefore, γh (x, y) ≥ 0 for all (x, y) with x 6= y iff the property (i) holds true.
(Necessity) If h is TP2 on (0, ∞)2 , then it is also TP2 on each rectangle (rectangular
area) in the region A1 = {(x, y) : x > y > 0} or in A2 = {(x, y) : y > x > 0}, and hence
the property (i) holds true by Lemma 4 and the above observation. Next, the property (ii)
16
follows from the fact that for all 0 < x1 = y1 < x2 = y2 , the cross-product ratio rh of h
satisfies
1 ≤ rh ≡ rh (x1 , x2 ; y1 , y2 ) =
exp(−θ(x2 − x1 ))h1 (0|θ)h2 (0|θ)
h(x1 , y1 )h(x2 , y2)
=
.
h(x1 , y2 )h(x2 , y1)
h1 (x2 − x1 |θ)h2 (x2 − x1 |θ)
This completes the proof of the necessity part.
(Sufficiency) Suppose 0 < x1 < x2 and 0 < y1 < y2 , then we want to prove the crossproduct ratio rh ≥ 1 under the assumptions (i) and (ii). If the rectangle with four vertices
Pi , i = 1, 2, 3, 4, where P1 = (x1 , y1 ), P2 = (x2 , y1 ), P3 = (x2 , y2), P4 = (x1 , y2 ), lies entirely in
the region A1 or A2 , then rh ≥ 1 by the assumption (i) and Lemma 4. If 0 < x1 = y1 < x2 =
y2 , then the assumption (ii) implies rh ≥ 1. For the remaining cases, we apply the technique
of factorization of the cross-product ratio if necessary. For example, if P∗ = (x1 , y∗ ) ∈ P1 P4
and P ∗ = (x∗ , y2 ) ∈ P4 P3 denote the intersection of the diagonal line x = y and boundary
of the rectangle, where x1 < x∗ < x2 and y1 < y∗ < y2 , then we split the original rectangle
into four sub-rectangles by adding the new point (x∗ , y∗ ) and calculate the ratio
rh (x1 , x2 ; y1 , y2 ) = rh (x1 , x∗ ; y1 , y∗)rh (x1 , x∗ ; y∗ , y2 )rh (x∗ , x2 ; y1 , y∗)rh (x∗ , x2 ; y∗ , y2 ) ≥ 1,
each factor being greater than or equal to one by the previous results. The proof is complete.
It is known that the Marshall–Olkin BVE (6) is PQD, so are its copula C and survival
copula Ĉ (see Barlow and Proschan 1981, p. 129). Moreover, H and Ĉ are TP2 due to
Theorem 6 and its corollary (see also Nelsen 2006, p. 163, for a direct proof) and both are
even TP∞ if λ1 = λ2 (Lin et al. 2016). We are now able to extend these results to the
following.
Theorem 8. The Marshall–Olkin survival function H and survival copula Ĉ are both TP∞ ,
regardless of parameters.
To prove this theorem, we need two more useful lemmas. Lemma 5 is well-known (see,
e.g., Marshall et al. 2011, p. 758), while Lemma 6 is essentially due to Gantmacher and Krein
(2002), pp. 78–79 (see also Karlin 1968, p. 112, for an alternative version).
Lemma 5. Let r ≥ 2 be an integer.
(i) If k(x, y) is TPr in x and y, and if both u and v are nonnegative functions, then the
product function K(x, y) = u(x) v(y) k(x, y) is TPr in x and y.
17
(ii) If k(x, y) is TPr in x and y, and if u and v are both increasing, or both decreasing, then
the composition function K(x, y) = k(u(x), v(y)) is TPr in x and y.
Lemma 6. Let φ and ψ be two positive functions on (a, b). Define the symmetric function
ψ(x) φ(y), a < y ≤ x < b
Ks (x, y) =
φ(x) ψ(y), a < x ≤ y < b.
If φ(x)/ψ(x) is nondecreasing in x ∈ (a, b), then the function Ks (x, y) is TP∞ in x and y.
Proof of Theorem 8. We prove first that the Marshall–Olkin survival function H is TP∞ .
Rewrite the survival function (6) as
exp[−(λ1 + λ12 )x − λ2 y], x ≥ y
H(x, y) =
exp[−(λ2 + λ12 )y − λ1 x], x ≤ y
= exp[−λ1 x − λ2 y]Ks (x, y),
where the symmetric function
Ks (x, y) =
exp(−λ12 x), x ≥ y
exp(−λ12 y), x ≤ y.
(10)
Let φ(x) = 1 and ψ(y) = exp(−λ12 y). Then by Lemma 6, we see that the function Ks in
(10) is TP∞ , so is H by Lemma 5(i). Next, recall that the Marshall–Olkin survival copula
Ĉ(u, v) = H(F
where F
−1
and G
−1
−1
(u), G
−1
(v)),
u, v ∈ (0, 1),
are the inverse (decreasing) functions of F (x) = exp[−(λ1 + λ12 )x] and
G(y) = exp[−(λ2 + λ12 )y], respectively. Therefore, Ĉ is TP∞ by Lemma 5(ii).
It is well known that if a bivariate distribution H has TP2 density, then both H and its
joint survival function H are TP2 (see, e.g., Balakrishnan and Lai 2009, p. 116). A more
general result is given as follows.
Theorem 9. If the bivariate distribution H has TPr density with r ≥ 2, then both H and
H are TPr . Consequently, if H has TP∞ density, then both H and H are TP∞ .
Proof. Let us consider first the TP∞ indicator functions K1 (x, y) = I(−∞,x] (y) and K2 (x, y) =
I[x,∞)(y), and then apply Theorem 3.5 of Gross and Richards (1998) restated below.
For example, to prove the TPr property of H, we have to claim that for all x1 < · · · < xr
and y1 < · · · < yr , the determinant of each s × s sub-matrix (H(xi , yj )) (with 2 ≤ s ≤ r)
18
is nonnegative. To prove this, let us recall that H(xi , yj ) = E[I(−∞,xi ] (X) I(−∞,yj ] (Y )] =
E[φ(i, X) ψ(j, Y )], where φ(i, x) = I(−∞,xi ] (x) is TPr in two variables i ∈ {1, 2, . . . , r} and
x ∈ R, and ψ(j, y) = I(−∞,yj ] (y) is TPr in two variables j ∈ {1, 2, . . . , r} and y ∈ R. Then
Gross and Richards’ Theorem applies and hence H is TPr . Similarly, the survival function
H is TPr . The proof is complete.
Gross and Richards’ (1998) Theorem. Let r ≥ 2 be an integer and let the bivariate
(X, Y ) ∼ H have TPr density. Assume further that both the functions φ(i, x) and ψ(i, x) are
TPr in two variables i ∈ {1, 2, . . . , r} and x ∈ R. Then the r × r matrix (E[φ(i, X) ψ(j, Y )])
is totally positive, that is, all its minors (of orders ≤ r) are nonnegative real numbers.
As mentioned in Balakrishnan and Lai (2009, p. 124), the Block–Basu BVE (8) is PQD
if λ1 = λ2 . We now extend this result to the following.
Theorem 10. (i) If λ1 = λ2 in (8), then the Block–Basu BVE has TP∞ density.
(ii) If α = β ≤ α′ = β ′ in (9), then the Freund BVE has TP∞ density.
Proof. Take φ(x) = c1 exp(−λ1 x) and ψ(y) = c2 exp[−(λ2 + λ12 )y] for some constants
c1 , c2 > 0. Then part (i) follows from (8) and Lemma 6. Part (ii) can be proved similarly.
Remark 4. The same approach applies to other bivariate (non-BLM) distributions like Li
and Pellerey’s (2011) generalized Marshall–Olkin bivariate distribution described below.
In Marshall and Olkin’s (1967) shock model: (X, Y ) = (X1 ∧ X3 , X2 ∧ X3 ), we assume
instead that X1 , X2 , X3 are independent general positive random variables (not limited to
exponential ones) and that Xi ∼ Fi , i = 1, 2, 3. Let Ri = − log Fi be the hazard function of
Xi . Then the generalized Marshall–Olkin bivariate distribution H has survival function
H(x, y) = Pr(X > x, Y > y) = Pr(X1 > x, X2 > y, X3 > max{x, y})
= exp[−R1 (x) − R2 (y) − R3 (max{x, y})],
x, y ≥ 0,
(11)
which is PQD (Li and Pellerey 2011). (For other related shock models, see Marshall and
Olkin 1967, Ghurye and Marshall 1984 as well as Aven and Jensen 2013, Section 5.3.4.) We
now extend this result and Theorem 8 as follows.
Theorem 11. Let H be the generalized Marshall–Olkin distribution defined in (11). Then
(i) the survival function H is TP∞ ;
19
(ii) the survival copula Ĉ of H is TP∞ , provided the functions F1 F3 and F2 F3 are both
strictly decreasing.
Proof. Write the survival function (11) as
exp[−(R1 (x) + R3 (x)) − R2 (y)], x ≥ y
H(x, y) =
exp[−(R2 (y) + R3 (y)) − R1 (x)], x ≤ y
= exp[−R1 (x) − R2 (y)]Ks (x, y),
where the symmetric function
Ks (x, y) =
exp[−R3 (x)], x ≥ y
exp[−R3 (y)], x ≤ y.
(12)
By taking φ(x) = 1 and ψ(y) = exp[−R3 (y)] in Lemma 6, we know that the function
Ks in (12) is TP∞ , and hence the survival function H is TP∞ by Lemma 5(i). This proves
part (i). To prove part (ii), we note that the marginal survival functions of H are F (x) =
exp[−R̃1 (x)], x ≥ 0, and G(y) = exp[−R̃2 (y)], y ≥ 0, where the two functions R̃1 (x) =
R1 (x) + R3 (x), x ≥ 0, and R̃2 (y) = R2 (y) + R3 (y), y ≥ 0, are strictly increasing by the
conditions on Fi , i = 1, 2, 3. This in turn implies that the marginal distribution functions F
and G are strictly increasing and hence the survival copula
Ĉ(u, v) = H(F −1 (1 − u), G−1 (1 − v)),
u, v ∈ (0, 1),
because F −1 (F (t)) = t, t ∈ (0, 1), where the quantile function F −1 (t) = inf{x : F (x) ≥
t}, t ∈ (0, 1) (see, e.g., Shorack and Wellner 1986, p. 6). Therefore, Ĉ is TP∞ by part (i)
and Lemma 5(ii). The proof is complete.
5. Stochastic Comparisons of BLM Distributions
To provide more information about BLM distributions, we can study stochastic comparisons in the BLM family. As usual, define the notions of the upper orthant order
(≤uo ), the concordance order (≤c ) and the Laplace transform order (≤Lt ) as follows. Let
(Xi , Yi ) ∼ Hi with marginals (Fi , Gi ), i = 1, 2, on R+ . Then denote (i) (X1 , Y1 ) ≤uo (X2 , Y2 )
if H1 (x, y) ≤ H2 (x, y) for all x, y ≥ 0, (ii) (X1 , Y1 ) ≤c (X2 , Y2 ) if (F1 , G1 ) = (F2 , G2 ) and
(X1 , Y1 ) ≤uo (X2 , Y2 ), and (iii) (X1 , Y1 ) ≤Lt (X2 , Y2 ) if L1 (s, t) ≥ L2 (s, t) for all s, t ≥ 0
(Müller and Stoyan 2002, Shaked and Shanthikumar 2007). We have, for example, the following results whose proofs are straightforward and are omitted.
20
Theorem 12. Let (Xi , Yi ) ∼ Hi = BLM(Fi , Gi , θi ) ∈ BLM, i = 1, 2. Then we have
(i) X1 ≤st X2 , Y1 ≤st Y2 and θ1 ≥ θ2 , iff (X1 , Y1 ) ≤uo (X2 , Y2 ), or, equivalently, E[K(X1 , Y1 )] ≤
E[K(X2 , Y2 )] for any bivariate distribution K on R2+ ;
(ii) F1 = F2 , G1 = G2 and θ1 ≥ θ2 , iff (X1 , Y1) ≤c (X2 , Y2 ), or, equivalently, E[k1 (X1 )k2 (Y1 )] ≤
E[k1 (X2 )k2 (Y2 )] for all increasing functions k1 , k2 , provided the expectations exist; and
(iii) if X1 ≤Lt X2 , Y1 ≤Lt Y2 and θ1 = θ2 , then (X1 , Y1 ) ≤Lt (X2 , Y2 ), or, equivalently,
E[k1 (X1 )k2 (Y1 )] ≥ E[k1 (X2 )k2 (Y2 )] for all completely monotone functions k1 , k2 , provided
the expectations exist.
When H1 and H2 have the same pair of marginals (F, G), Theorem 12(i) reduces, by
Corollary 1, to the following interesting result which is related to the famous Slepian’s
inequality for bivariate normal distributions (see the discussion in Remark 5 below).
Corollary 4. Let (Xi , Yi ) ∼ Hi = BLM(F, G, θi ) ∈ BLM, with correlation ρi , i = 1, 2.
Then ρ1 ≤ ρ2 iff H1 (x, y) ≤ H2 (x, y) for all x, y ≥ 0, or, equivalently, H1 (x, y) ≤ H2 (x, y)
for all x, y ≥ 0.
Remark 5. In Corollary 4 above, if we consider standard bivariate normal distributions
instead of BLM ones, then the conclusion also holds true and the necessary part is the socalled Slepian’s lemma/inequality; see Slepian (1962), Müller and Stoyan (2002), p. 97, and
Hoffmann-Jørgensen (2013) for more general results. In Wikipedia, it was said that while
this intuitive-seeming result is true for Gaussian processes, it is not in general true for other
random variables. However, as we can see in Corollary 4, there are infinitely many BLM
distributions sharing the same Slepian’s inequality with bivariate normal ones.
Remark 6. We finally compare the effects of the dependence structure of BLM distributions
in different coherent systems. Consider a two-component system and let the two components
have lifetimes (X, Y ) ∼ H = BLM(F, G, θ). Then the lifetime of a series system composed of
these two components is X ∧Y ∼ Exp(θ), while the lifetime of a parallel system composed of
the same components is X ∨Y obeying the distribution Hp (z) = e−θz −1+F (z)+G(z), z ≥ 0.
Therefore the mean times to failure of series and parallel systems are, respectively, E[X∧Y ] =
R∞
R∞
H(x,
x)dx
=
1/θ
(decreasing
in
θ)
and
E[X
∨
Y
]
=
E[X]
+
E[Y
]
−
H(x, x)dx =
0
0
E[X] + E[Y ] − 1/θ (increasing in θ). The latter further implies that θ ≥ (E[X] + E[Y ])−1
21
(compare with Theorem 1(vi)) and that E[XY ] ∈ [1/θ2 , (E[X] + E[Y ])2 ] by Corollary 1,
provided the expectations exist. See also Aven and Jensen (2013, Section 2.3) for special
cases with exponential marginals as well as Lai and Lin (2014) for more general results.
Acknowledgments. The authors would like to thank the Editor-in-Chief and two Referees
for helpful comments and constructive suggestions which improve the presentation of the
paper. The paper was presented at the Ibusuki International Seminar, Ibusuki Phoenix
Hotel, held from 6th to 8th March 2016 by Waseda University, Japan. The authors thank
the organizer Professor Masanobu Taniguchi for his kind invitation and the audiences for
their comments and suggestions.
References
Aven, T. and Jensen, U. (2013). Stochastic Models in Reliability, 2nd ed. Springer, New
York.
Balakrishnan, N. and Lai, C.-D. (2009). Continuous Bivariate Distributions, 2nd ed.
Springer, New York.
Barlow, R. E. and Proschan, F. (1981). Statistical Theory of Reliability and Life Testing:
Probability Models, To Begin With. Silver Spring, MD.
Bassan, B. and Spizzichino, F. (2005). Relations among univariate aging, bivariate aging
and dependence for exchangeable lifetimes. J. Multivariate Anal., 93, 313–330.
Block, H. W. (1977). A characterization of a bivariate exponential distribution. Ann.
Statist., 5, 808–812.
Block, H. W. and Basu, A. P. (1974). A continuous bivariate exponential extension. J.
Amer. Statist. Assoc., 69, 1031–1037.
Block, H. W. and Savits, T. H. (1976). The IFRA closure problem. Ann. Probab., 4,
1030–1032.
Block, H. W. and Savits, T. H. (1980). Multivariate increasing failure rate average distributions. Ann. Probab., 8, 793–801.
22
Crawford, G. B. (1966). Characterization of geometric and exponential distributions. Ann.
Math. Statist., 37, 1790–1795.
Esary, J. D. and Marshall, A. W. (1979). Multivariate distributions with increasing hazard
rate average. Ann. Probab., 7, 359–370.
Esna–Ashari, M. and Asadi, M. (2016). On additive–multiplicative hazards model. Statistics, 50, 1421–1433.
Feller, W. (1965). An Introduction to Probability Theory, Vol. I. Wiley, New York.
Ferguson, T. S. (1964). A characterization of the exponential distribution. Ann. Math.
Statist., 35, 1199–1207.
Ferguson, T. S. (1965). A characterization of the geometric distribution. Amer. Math.
Monthly, 72, 256–260.
Fortet, R. (1977). Elements of Probability Theory. Gordon and Breach, New York.
Freund, J. E. (1961). A bivariate extension of the exponential distribution. J. Amer.
Statist. Assoc., 56, 971–977.
Galambos, J. and Kotz, S. (1978). Characterizations of Probability Distributions. Springer,
New York.
Gantmacher, F. R. and Krein, M. G. (2002). Oscillation Matrices and Kernels and Small
Vibrations of Mechanical Systems, Revised edn. Translation based on the 1941 Russian
original, Providence, RI.
Ghurye, S. G. (1987). Some multivariate lifetime distributions. Adv. Appl. Probab., 19,
138–155.
Ghurye, S. G. and Marshall, A. W. (1984). Shock processes with aftereffects and multivariate lack of memory. J. Appl. Probab., 21, 768–801.
23
Gross, K. I. and Richards, D. St. P. (1998). Algebraic methods toward higher-order probability inequalities. In: Stochastic Processes and Related Topics (B. Rajput et al., eds.),
189–211, Birkhäuser, Boston.
Gross, K. I. and Richards, D. St. P. (2004). Algebraic methods toward higher-order probability inequalities, II. Ann. Probab., 32, 1509–1544.
Hoffmann-Jørgensen, J. (2013). Slepian’s inequality, modularity and integral orderings.
High Dimensional Probability VI, 19–53, Progress in Probability, 66, Springer, Basel.
Holland, P. W. and Wang, Y. J. (1987). Dependence function for continuous bivariate
densities. Comm. Statist. – Theory and Methods, 16, 863–876.
Jones, M. C. (1996). The local dependence function. Biometrika, 83, 899–904.
Jones, M. C. (1998). Constant local dependence. J. Multivariate Anal., 64, 148–155.
Karlin, S. (1968). Total Positivity, Vol. I. Stanford University Press, CA.
Karlin, S. and Proschan, F. (1960). Pólya type distributions of convolutions. Ann. Math.
Statist., 31, 721–736.
Kayid, M., Izadkhah, S. and Alshami, S. (2016). Laplace transform ordering of time to
failure in age replacement models. J. Korean Statist. Theory, 45, 101–113.
Kemperman, J. H. B. (1977). On the FKG–inequality for measures on a partially ordered
space. Indagationes Mathematicae, 80, 313–331.
Kotz, S., Balakrishnan, N. and Johnson, N. L. (2000). Continuous Multivariate Distributions, Vol. 1: Models and Applications, 2nd ed. Wiley, New York.
Kulkarni, H. V. (2006). Characterizations and modelling of multivariate lack of memory
property. Metrika, 64, 167–180.
Lai, C.-D. and Lin, G. D. (2014). Mean time to failure of systems with dependent components. Appl. Math. Comput., 246, 103–111.
24
Lai, C.-D. and Xie, M. (2006). Stochastic Ageing and Dependence for Reliability. Springer,
New York.
Lee, M.-L.T. (1985). Dependence by total positivity. Ann. Probab., 13, 572–582.
Li, X. and Pellerey, F. (2011). Generalized Marshall–Olkin distributions and related bivariate aging properties. J. Multivariate Anal., 102, 1399–1409.
Lin, G. D., Dou, X., Kuriki, S. and Huang, J. S. (2014). Recent developments on the
construction of bivariate distributions with fixed marginals. Journal of Statistical Distributions and Applications, 1: 14.
Lin, G. D., Lai, C.-D. and Govindaraju, K. (2016). Correlation structure of the Marshall–
Olkin bivariate exponential distribution. Statist. Methodology, 29, 1–9.
Marshall, A. W. and Olkin, I. (1967). A multivariate exponential distribution. J. Amer.
Statist. Assoc., 62, 30–44.
Marshall, A. W. and Olkin, I. (2015). A bivariate Gompertz–Makeham life distribution. J.
Multivariate Anal., 139, 219–226.
Marshall, A. W., Olkin, I. and Arnold, B. C. (2011). Inequalities: Theory of Majorization
and Its Applications, 2nd ed. Springer, New York.
Müller, A. and Stoyan, D. (2002). Comparison Methods for Stochastic Models and Risks.
Wiley, New York.
Nadarajah, S. (2006). Exact distributions of XY for some bivariate exponential distributions. Statistics, 40, 307–324.
Nelsen, R. B. (2006). An Introduction to Copulas, 2nd ed. Springer, New York.
Newman, C. (1984). Asymptotic independence and limit theorems for positively and negatively dependent random variables. In: Inequalities in Statistics and Probability (Y.
L. Tong, ed.), IMS Lecture Notes–Monograph Series, Vol. 5, 127–140.
25
Rao, C. R. and Shanbhag, D. N. (1994). Choquet–Deny Type Functional Equations with
Applications to Stochastic Models. Wiley, New York.
Shaked, M. (1977). A family of concepts of positive dependence for bivariate distributions.
J. Amer. Statist. Assoc., 72, 642–650.
Shaked, M. and Shanthikumar, J. G. (1988). Multivariate conditional hazard rates and the
MIFRA and MIFR properties. J. Appl. Probab., 25, 150–168.
Shaked, M. and Shanthikumar, J. G. (2007). Stochastic Orders. Springer, New Jersey.
Shorack, G. R. and Wellner, J. A. (1986). Empirical Processes with Applications to Statistics. Wiley, New York.
Slepian, D. (1962). The one-sided barrier problem for Gaussian noise. Bell System Technical
Journal, 41, 463-501.
Wang, Y. J. (1993). Construction of continuous bivariate density functions. Statist. Sinica,
3, 173–187.
26
| 10math.ST
|
Generating Representative Executions
Extended Abstract
Hendrik Maarand
Tarmo Uustalu
Dept. of Software Science, Tallinn University of Technology
Analyzing the behaviour of a concurrent program is made difficult by the number of possible executions. This problem can be alleviated by applying the theory of Mazurkiewicz traces to focus only
on the canonical representatives of the equivalence classes of the possible executions of the program.
This paper presents a generic framework that allows to specify the possible behaviours of the execution environment, and generate all Foata-normal executions of a program, for that environment, by
discarding abnormal executions during the generation phase. The key ingredient of Mazurkiewicz
trace theory, the dependency relation, is used in the framework in two roles: first, as part of the
specification of which executions are allowed at all, and then as part of the normality checking algorithm, which is used to discard the abnormal executions. The framework is instantiated to the relaxed
memory models of the SPARC hierarchy.
1
Introduction
Let us consider a fragment from Dekker’s mutual exclusion algorithm as an example.
Init: x = 0; y = 0;
P1
P2
(a) [x] := 1
(c) [y] := 1
(b) r1 := [y] (d) r2 := [x]
Observed? r1 = 0; r2 = 0;
This is a concurrent program for two processors, P1 and P2 , where x is the flag variable for P1 that is used
to communicate that P1 wants to enter the critical section and y is for P2 . A processor may enter the critical
section, if it has notified the other processor by setting its flag variable to 1, reading the flag variable of
the other processor and checking that it is 0. We are interested in whether it is possible, starting from an
initial state where both x and y are 0, that both processors see each others’ flag variables as 0, meaning
that both processors enter the critical section. Here we are interested in the mutual exclusion property,
that at most one processor can enter the critical section.
In the interleaving semantics of Sequential Consistency (SC), the above program can have the following executions: abcd, cdab, acbd, cabd, acdb, cadb. Out of these six, the four last executions are
actually equivalent (in the sense that from the same initial state they will reach the same final state) and
for our purposes it is enough to check the final state of only one of them. We can observe that the mutual
exclusion property is satisfied. The situation is different, if we consider the possible executions on a
real-world processor, like x86, which follows the Total Store Order (TSO) model [8]. Under TSO, it is
possible for writes to be reordered with later reads from the same processor, resulting in an execution
that is observable as bdac. This does not satisfy the mutual exclusion property.
In this paper, we seek to alleviate the difficulty analyzing the large numbers of executions concurrent
programs, especially on relaxed memories, generate, by applying the theory of Mazurkiewicz traces to
focus only on some type of canonical representatives of the equivalence classes of the possible executions
of the program. We present a generic framework for interpreting concurrent programs under different
V.T. Vasconcelos and P. Haller (Eds.): Workshop on Programming Language
Approaches to Concurrency- and Communication-cEntric Software (PLACES’17)
EPTCS 246, 2017, pp. 39–48, doi:10.4204/EPTCS.246.8
c H. Maarand & T. Uustalu
This work is licensed under the
Creative Commons Attribution License.
Generating Representative Executions
40
semantics, so that only executions in the Foata normal form (corresponding to maximal parallelism)
are generated. We instantiate the framework to the relaxed memory models of the SPARC hierarchy.
This work is in the vein of partial order reduction techniques for analysis of systems, which are widely
used especially in model checking and have also been applied to relaxed memories, e.g., by Zhang et
al. [13]. The novelties here are that the different memory models are modelled uniformly based on a
flexible notion of a backlog of shadow events, using a standard normal form from trace theory, and using
generalized traces (with a dynamic independency relation) to be able to define execution equivalence
more finely, resulting in bigger and fewer equivalence classes. The framework has been prototyped
in Haskell where one can easily separate the phases of generating the tree of symbolic executions of a
program, discarding the abnormal executions, and running the tree of symbolic executions from an initial
state. This separation can be made without a performance penalty thanks to lazy evaluation.
2
Mazurkiewicz Traces
An execution (or a run) of a sequential program can be represented as a sequence of symbols that record
the events caused by the program in the order that they occurred. Such a sequence is a string over
some (finite) alphabet Σ. An execution of a concurrent program can be represented as an interleaving
of the executions on the processors involved, thereby reducing concurrency to non-deterministic choice.
Mazurkiewicz traces [7] (or just traces) are a generalization of strings, where some of the letters in the
string are allowed to commute. This allows representation of non-sequential behaviour. In other words,
traces are equivalence classes of strings with respect to a congruence relation that allows to commute
certain pairs of letters.
A dependency relation D ⊆ Σ × Σ is a reflexive and symmetric binary relation. a D b if and only if
the events a and b can be causally related, meaning that the two events cannot happen concurrently. The
complement of the dependency relation, I = (Σ × Σ) \ D, is called the independency relation. If a I b,
then the strings sabt and sbat represent the same non-sequential behaviour. Two strings s,t ∈ Σ∗ are said
to be Mazurkiewicz equivalent, s ≡D t, if and only if s can be transformed to t by a finite number of
exchanges of adjacent, independent events. For example, if Σ = {a, b, c, d} and a I c and b I d then the
trace acbd represents the strings acbd, cabd, acdb and cadb.
For our purposes, standard Mazurkiewicz traces are not enough and therefore we turn to the generalized Mazurkiewicz traces of Sassone et al. [10]. In generalized Mazurkiewicz traces, the dependency
relation is dynamic, it depends on the current context, which is the partial execution that has been performed so far. The dependency relation for a prefix s will be denoted by Ds and the subscript is omitted,
if the relation is static. Besides Ds having to be reflexive and symmetric for any s, D must satisfy some
sanity conditions. Most importantly, if s ≡D t, then it must be the case that Ds = Dt . In this setting, the
strings sabt and sbat are considered equivalent, if a Is b.
Normal Forms As traces are equivalence classes, it is reasonable to ask what the canonical representative or normal form of a trace is. There are two well-known normal forms for traces, the lexicographic
and Foata [4] normal forms. We are going to look at Foata normal forms for our purposes.
A step is a subset s ⊆ Σ of pairwise independent letters. The Foata normal form of a trace is a
sequence s1 . . . sk of steps such that the individual steps s1 , . . . , sk are chosen from the left to the right
with maximal cardinality. Since each step consists of independent letters, a step can be executed in
parallel, meaning that the Foata normal form encodes a maximal parallel execution. For example, if
Σ = {a, b, c, d} and a I c and b I d, then the Foata normal form of acbd is (ac)(bd).
We are interested in checking whether a given string is in normal form according to a given depen-
H. Maarand & T. Uustalu
41
dency relation. As a convenience, we also assume to have an ordering ≺ on Σ that is total on events that
are independent. A string is in Foata normal form, if it can be split into a sequence of steps s1 , . . . , sk so
that concatenation of the steps gives the original string and the following conditions are satisfied:
1. for every a, b ∈ si , if a 6= b then a Ii b;
2. for every b ∈ si+1 , there is an a ∈ si such that a Di b;
3. for every step si , the letters in it are in increasing order wrt. ≺.
In these definitions, we consider Di to be the dependency relation for the context s0 . . . si−1 and similarly
for Ii . The first condition ensures that the events in a step can be executed in parallel. The second
condition ensures that every event appears in the earliest possible step, i.e., maximal parallelism. The
third condition picks a permutation of a step as a representative of the step. Notice that if a string is not
in normal form, then neither is any string with that string as a prefix in normal form. This means that
when checking a string for normality by scanning it from the left to the right, we can discard it as soon
as we discover an abnormal prefix.
3
Framework
We now proceed to describing our framework for generating representative executions of a program and
its instantiations to different memory models.
We are going to look at programs executing on a machine that consists of processors and a shared
memory. Each processor also has access to a local memory (registers). The executions that we investigate
are symbolic, in the sense that we do not look at the actual values propagating in the memory, but just
the abstract actions being performed. Still, our goal is to find the possible final states of a program from
a given initial state. The idea is that once the symbolic executions have been computed, the canonical
executions can be picked and the final state needs to be computed only for those. This can be done lazily,
meaning that the evaluation of a particular execution for the given initial state is cancelled immediately,
if it is discovered that the execution is not normal.
The language for our system consists of arithmetic and boolean expressions and commands. An arithmetic expression is either a numeral value, a register, or an arithmetic operation. A boolean expression
is either boolean constant, a boolean operation, or a comparison of arithmetic expressions. Commands
consist of assignments to registers, loads and stores to shared memory, and if and while constructs.
Our framework is defined on top of the events generated by the system. We think of events as
occurrences of (the phases of) the actions that executing the program can trigger. An event can be
thought of as a record (pid, eid, kind, act) where pid is the identifier of the processor that generated the
event, eid is the processor-local identifier of the event, kind defines whether it is a main or a shadow
event, and act is the action performed in this event. An action can be an operation between registers, a
load from or a store to a variable, or an assertion on registers. An assertion is used to record a decision
made in the unfolding of a control structure of the program, for example, that a particular execution is
one where the true branch of a conditional was taken. If an assertion fails when an execution is evaluated
from a given initial state, then this execution is not valid for that initial state.
Since we are interested in modelling different memory models, our framework is parameterized by an
architecture, which characterizes the behavioural aspects of the system. An architecture consists of four
components. A predicate shadows describes whether an action is executed in a single stage or two stages,
generating just one (main) event or two events (a main and a shadow event). An irreflexive-antisymmetric
relation sameDep describes which events from a processor must happen before which other events from
the same processor: it plays a role in determining the possible next events from this processor, but also
42
Generating Representative Executions
defines which events from it are dependent. A relation diffDep describes when two events from different
processors are dependent. Finally, a relation ≺ orders independent events. The relations sameDep (its
reflexive-symmetric closure) and diffDep together determine the dependency relation in the sense of
Mazurkiewicz traces and ≺ is the relation used to totally order the events within a step.
In the previous paragraph, we mentioned shadow events. These are the key ingredients of this framework for modelling more intricate behaviours, for example, when some actions are non-atomic and this
fact needs to be reflected in the executions by two events, a main event and a shadow event. TSO, for
example, can be described as a model where writes to memory first enter the processor’s write-buffer and
are later flushed from the write-buffer to memory. We consider the write to buffer to be the main event
of the write action and the flush event to be the shadow event of the write action. Of these two events,
the shadow event is globally observable.
Generating Normal Forms The process of generating normal-form executions of a program can be
divided into two stages: lazily generating all executions of the program and then discarding those not in
normal form.
The executions are generated as follows: if all processors have completed, then we have a complete
execution and we are done, otherwise we pick a processor that has not yet completed and allow it to
make a small step, then repeat the process. The local configuration of a processor consists of its residual
program, backlog, and the value of a counter to provide identifiers for the generated events. The small
step can either correspond to beginning the action of the next instruction according to the program—in
which case a new main event is generated and added to the execution—, or to completing an already
started action—in this case, a shadow event is removed from the processor’s backlog and added to the
execution. If the step is to start a new action, then the shadows predicate is used to check whether a
new shadow event should be added to the backlog (if not, the action is completed by the main event).
A side-condition for adding a new main event is that there are no shadow events in the backlog that
are dependent with it. An event can be removed from the backlog, if it is independent (according to
sameDep) of all of the older events in the backlog. Conditionals like if and while are expanded to a
choice between two programs, where the choices correspond to the branches of the conditional together
with an assertion of the condition. The generation of executions is described by the small step rules in
Appendix A.
The second stage of the procedure is to single out the normal forms among the generated executions.
This is done by checking the normality of the executions according to the three conditions given in
Section 2 for Foata normal forms. The rules for checking the normality of an execution by scanning it
from the left to the right are given in Appendix A.
Instead of generating a flat set of executions in the first stage, we actually generate a tree of executions, so that the prefixes of executions are shared. Since the process of selecting the canonical executions
(more precisely, discarding the non-canonical ones) according to the conditions of Foata normal forms
can be fused into the generation stage, we can discard a whole set of executions when we discover that
the current path down the tree violates the normality conditions. More precisely, walking down the tree,
we keep track of the current prefix (which must be in normal form) and at each node we check whether
the event associated with the node would violate the normality conditions when added to the prefix. Only
if the normality condition is not violated does the subtree starting from that node need to be computed
actually.
We require sameDep a b to hold at least when a and b are main events and eid a < eid b or when
they are a main event and its shadow event (in which case they have the same eid). We also require
that sameDep a b can only hold when eid a < eid b or when eid a = eid b and a is a main event and b
H. Maarand & T. Uustalu
43
the corresponding shadow event. Under these assumptions, we can prove that the total set of executions
captured in the generated tree is closed under equivalence. As the normality checking stage keeps all
normal forms and discards all non-normal forms, it follows that the pruned set of executions contains
exactly one representative for every execution of the program.
In the introduction, we noted that our example program has six executions under interleaving semantics, of which four are equivalent. The executions are depicted in Figure 1 and the four equivalent
executions acbd, acdb, cabd and cadb are the ones in the middle. For this program we have that a I c
and b I d. Our framework would only generate acbd out of these four, as this corresponds to the Foata
normal form (ac)(bd) and the other three would be discarded. More precisely, (ac)(d) is in normal form,
but it cannot be extended by b, as neither (ac)(db) nor (ac)(d)(b) is in normal form: the first one fails due
to condition 3 and the second one fails due to condition 2. The node b of this path is shaded in the picture
to highlight the place where the normality condition is violated. For cabd, we start checking normality
from (c), which is valid, but neither (ca) nor (c)(a) is in normal form and we can discard all executions
that start with ca, which includes both cabd and cadb. The subtree at node a is shaded to highlight this
fact.
c
a
c
b
a
d
c
b
d
b
d
a
d
d
b
d
b
b
Figure 1: SC executions of the example program.
4
Instantiation to Relaxed Memory Models
Sequential Consistency In the Sequential Consistency (SC) model [6], any execution of a concurrent
program is an interleaving of the program order executions of its component threads. SC can be specified
as an architecture in the following way:
shadows a = false
sameDep a b = eid a < eid b
diffDep x y a b = crxw a b
a ≺ b = pid a < pid b
crxw a b represents the concurrent-read-exclusive-write property, which returns true, if events a and b
access the same location and at least one of them is a write. diffDep also takes two arguments that are
ignored here, which represent the backlogs of the two processors from which the events a and b originate
from. This information can be recovered from the prefix of the execution and it is as much information
Generating Representative Executions
44
as we need about the prefix of the execution in the memory models we consider. We could also just take
the prefix of the execution itself and compute the necessary information. Setting shadows to be always
false means that all instructions execute atomically. Setting sameDep a b to require eid a < eid b means
that the events from the same processor must be generated in program order and cannot be reordered,
which reflects the definition of SC.
Total Store Order In the Total Store Order (TSO) model [11], it is possible for a write action to be
reordered with later reads, meaning that writes happen asynchronously, but at the same time the order of
write actions is preserved. TSO can be specified in the following way:
shadows a = isWrite a
sameDep a b = isMain a ∧ isMain b ∧ eid a < eid b
∨ isMain a ∧ isShadow b ∧ eid a == eid b
∨ isShadow a ∧ isShadow b ∧ eid a < eid b
diffDep x y a b = crxw′ x y a b
a ≺ b = pid a < pid b ∨ pid a == pid b ∧ eid a < eid b
crxw’ is like crxw, except that it considers shadow write events instead of main write events as the
global write events, and read events as global only if they access the memory. This is where we need
generalized Mazurkiewicz traces, since if there is a pending write to the location of the read, then the
read action would not read its value from memory and thus could not be dependent with events from
other processors.
We consider the main event of a write instruction to be the write to buffer and the shadow event to
be the flushing of the write from buffer to memory. TSO can be thought of as a model where every
processor has a shadow processor and all events on every main processor are in program order, all of the
events on the associated shadow processor are in program order and an event on the shadow processor
must happen after the corresponding event on the main processor. Our example from introduction has
the following traces in Foata normal form under TSO: (ac)(a’c’)(bd), (ac)(a’b)(c’d), (ac)(c’d)(a’b) and
(ac)(bd)(a’c’) where a’ stands for the shadow event of a. The last of these is the one rejected by SC.
Partial Store Order The Partial Store Order (PSO) model [11] allows the reorderings of TSO, but it is
also possible for a write to be reordered with a later write to a different location. This can be thought of
as having a separate write buffer for every variable. PSO can be specified as TSO with the exception of
the sameDep relation:
sameDep a b = isMain a ∧ isMain b ∧ eid a < eid b
∨ isMain a ∧ isShadow b ∧ eid a == eid b
∨ isShadow a ∧ isShadow b ∧ eid a < eid b ∧ var a == var b
Intuitively, this corresponds to PSO, since it is like TSO except for the dependency relation on events
from the same processor, where the shadow events are dependent only if they are to the same location,
which allows one to reorder writes to different locations.
Relaxed Memory Order The Relaxed Memory Order [11] (RMO) model only enforces program order on write-write and read-write instruction pairs to the same variable and on instruction pairs in dependency, where the first instruction is a read. Dependency on instruction pairs here means that there is
H. Maarand & T. Uustalu
45
data- or control-dependency between the instructions. We can specify RMO in the following way:
shadows a = true
sameDep a b = isMain a ∧ isMain b ∧ eid a < eid b
∨ isMain a ∧ isShadow b ∧ eid a == eid b
∨ isShadow a ∧ isShadow b ∧ eid a < eid b
∧ (var a == var b ∧ (isWrite a ∨ isRead a) ∧ isWrite b
∨ dataDep a b ∨ controlDep a b)
′′
diffDep x y a b = crxw x y a b
a ≺ b = pid a < pid b ∨ pid a == pid b ∧ eid a < eid b
crxw” is like crxw’ except that it considers shadow reads and shadow writes as the global read and write
events. As for TSO and PSO, a shadow read is considered global, if it actually reads its value from
memory, which in this model happens, if there is no older shadow write to the same location in the
backlog. We consider events a and b to be in data-dependency, if a reads a register that is written by b.
We consider two events to be in control-dependency, if the older one is a conditional and the newer one
is a write.
4.1 Fences
In models like TSO, PSO and RMO that allow the reordering of some events, it becomes necessary to
be able to forbid these reorderings in certain situations, to rule out relaxed behaviour. Our example from
introduction does not behave correctly on TSO, where it is possible for both processors to read the value
0. To avoid this situation, it is necessary to make sure that both processors first perform the write and
when the effects of the write operation have become globally visible they may perform the read. With
this restriction the program behaves correctly on TSO and the way to achieve this is to insert a fence
between the write and read instructions.
In our framework, fences are described by two parameters that can take the values store or load,
which indicate between which events the ordering is enforced. Under SC, the fence instructions can
be ignored since no reorderings are possible. To be able to restore sequentially consistent behaviour,
TSO requires store-load fences, PSO requires also store-store fences, and RMO requires all four kinds
of fences. For TSO, PSO, and RMO, the idea is that fences have shadow events and the sameDep
relation is modified to disallow unwanted reorderings. Our example program requires a store-load fence,
so that the read operations appearing after the fence cannot be performed before the write operations
appearing before the fence have completed. This means that sameDep must be modified to consider a
shadow store-load fence to be dependent with all older shadow write events and all newer read events.
Dependence with a shadow event prevents the fence event from being removed from the backlog until
the older dependent events have been removed and it also prevents removing the newer dependent events
until the fence has been removed from the backlog. Likewise, a new main read event cannot be added to
the execution, if there is a store-load fence event in the backlog. The idea is similar for the other types
of fences.
5
Related Work
Relaxed memory consistency models and their specification and verification tasks have been an extensive
research topic. Owens et al. [8] showed that x86 adheres to TSO model and they gave both operational
Generating Representative Executions
46
and axiomatic models. Alglave [2] defined a framework in an axiomatic style for working with relaxed
memory models, which is also generic in the sense that different memory models can be represented by
specifying which relations are considered global. Generating the possible executions in our framework
turns out to be quite similar to an executable specification for RMO given by Park and Dill [9], more
precisely, our notion of backlog seems to correspond to the reordering box used there. Boudol et al. [3]
defined a generic operational semantics that captures TSO, PSO and RMO and uses temporary stores that
again are similar to our backlogs; they did not however consider any partial order reduction of the set of
executions of a program. As mentioned before, due to the interest in exploring the full set of executions
by constructing it explicitly and the use of trace theory, which is the foundation for partial order reduction
[5], this work is also close to methods based on model checking, like Zhang et al.’s [13] and Abdulla
et al.’s [1]. An executable specification was also given by Yang et al. [12]. Their approach is based on
axiomatic specifications and an execution is found by searching for an instantiation that satisfies all of
the constraints, either by Prolog or a SAT solver.
6
Conclusion
We have presented a generic framework for finding canonical representatives of equivalence classes of
the possible executions of a program. The framework proceeds by lazily generating all executions of
the given program and discards all those that are not in Foata normal form. The framework allows to
uniformly represent the semantics of a certain class of relaxed memory models, which we have illustrated
by encoding the models from the SPARC hierarchy in terms of our framework. An instantiation of the
framework to a particular model specifies which executions can occur at all for the given program and
which of those are equivalent, i.e., correspond to one generalized Mazurkiewicz trace, representable by
its normal form.
We plan to continue this work by elaborating on the formal aspects of the framework. We have
formalized soundness and completeness of Foata normalization of (standard) traces in the dependently
typed functional language Agda—any string is equivalent to its normal form, and if a string is equivalent
to a normal form, it is that string’s normal form. This development can be scaled for generalized traces,
adapted to prove that the tree filtering algorithm keeps exactly one representative of each equivalence
class of executions, to then move on to formalization of specifications of memory models.
Acknowledgments This research was supported by the Estonian Ministry of Education and Research institutional research grant no. IUT33-13 and the ERDF funded CoE project EXCITE (20142020.4.01.15-0018).
References
[1] P. A. Abdulla, S. Aronis, M. F. Atig, B. Jonsson, C. Leonardsson & K. Sagonas (2015): Stateless Model
Checking for TSO and PSO. In: C. Baier & C. Tinelli, editors: Proc. of 21st Int. Conf. on Tools and
Algorithms for the Construction and Analysis of Systems, TACAS 2015, Lect. Notes in Comput. Sci. 9035,
Springer, pp. 353–367, doi:10.1007/978-3-662-46681-0_28.
[2] J. Alglave (2010): A Shared Memory Poetics. Ph.D. thesis, Université Paris 7. Available at http://www0.
cs.ucl.ac.uk/staff/J.Alglave/these.pdf.
[3] G. Boudol, G. Petri & Serpette G. (2012): Relaxed Operational Semantics of Concurrent Programming
Languages. In B. Luttik & M. A. Reniers, editors: Proc. of Combined 19th Wksh. on Expressiveness in
Concurrency and 9th Wksh. on Structural Operational Semantics, EXPRESS/SOS 2012, Electron. Proc. in
Theor. Comput. Sci. 89, pp. 19–33, doi:10.4204/eptcs.89.3.
H. Maarand & T. Uustalu
47
[4] P. Cartier & D. Foata (1969): Problemes combinatoires de commutation et réarrangements. Lect. Notes in
Math. 85, Springer, doi:10.1007/bfb0079468.
[5] P. Godefroid (1996): Partial-Order Methods for the Verification of Concurrent Systems: An Approach to the
State-Explosion Problem. Springer, doi:10.1007/3-540-60761-7.
[6] L. Lamport (1979): How to Make a Multiprocessor Computer That Correctly Executes Multiprocess Programs. IEEE Trans. on Comput. 28(9), pp. 690–691, doi:10.1109/tc.1979.1675439.
[7] A. Mazurkiewicz (1995): Introduction to Trace Theory. The Book of Traces, pp. 3–41, doi:10.1142/
9789814261456_0001.
[8] S. Owens, S. Sarkar & P. Sewell (2009): A Better x86 Memory Model: x86-TSO. In S. Berghofer,
T. Nipkow, C. Urban & M. Wenzel, editors: Proc. of 22nd Int. Conf. on Theorem Proving in Higher
Order Logics, TPHOLs 2009, Lect. Notes in Comput. Sci. 5674, Springer, pp. 391–407, doi:10.1007/
978-3-642-03359-9_27.
[9] S. Park & D. L. Dill (1995): An Executable Specification, Analyzer and Verifier for RMO (Relaxed Memory
Order). In: Proc. of 7th Ann. ACM Symp. on Parallel Algorithms and Architectures, SPAA ’95, ACM, pp.
34–41, doi:10.1145/215399.215413.
[10] V. Sassone, M. Nielsen & G. Winskel (1993): Deterministic Behavioural Models for Concurrency. In
A. M. Borzyszkowski & S. Sokolowski, editors: Proc. of 18th Int. Symp. on Mathematical Foundations
of Computer Science, MFCS ’93, Lect. Notes in Comput. Sci. 711, Springer, pp. 682–692, doi:10.1007/
3-540-57182-5_59.
[11] SPARC International Inc. & David L. Weaver (1994): The SPARC Architecture Manual. Prentice-Hall.
[12] Y. Yang, G. Gopalakrishnan, G. Lindstrom & K. Slind (2004): Nemos: A Framework for Axiomatic and
Executable Specifications of Memory Consistency Models. In: Proc. of 18th Int. Parallel and Distributed
Processing Symposium, IPDPS 2004, IEEE, pp. 31–40, doi:10.1109/ipdps.2004.1302944.
[13] N. Zhang, M. Kusano & C. Wang (2015): Dynamic Partial Order Reduction for Relaxed Memory Models.
In: Proc. of 36th ACM SIGPLAN Conf. on Principles of Language Design and Implementation, PLDI 2015,
ACM, pp. 250–259, doi:10.1145/2737924.2737956.
A
Semantic Rules
Small steps of a processor
[] I same e′
e I same e′ bklg I same e′
e : bklg I same e′
shadows(act)
bklg I same (eid, ◦, act)
(eid,◦,act)
(act : prg, bklg, eid) −−−−−→ (prg, (eid, •, act) : bklg, eid + 1)
¬shadows(act)
bklg I same (eid, ◦, act)
(eid,◦,act)
(act : prg, bklg, eid) −−−−−→ (prg, bklg, eid + 1)
older I same le
le
(prg, newer ++ (le : older), eid) −
→ (prg, newer ++ older, eid)
le
(prgi , bklg, eid) −
→c
le
(prg0 + prg1 , bklg, eid) −
→c
Generating Representative Executions
48
Small steps of the system
le
c(pid) = lc lc −
→ lc′
(pid,le)
c −−−−→ c[pid 7→ lc′ ]
Executions
∀pid. c(pid) = ([], [], )
[]
c=
⇒c
Normal executions
es
e
⇒ c′
c−
→ c′′ c′′ =
e:es ′
c =⇒ c
pid 6= pid ′ le Issdiff le′
le I same le′
(pid, le) Iss (pid, le′ ) (pid, le) Iss (pid ′ , le′ )
e ≺ e′
[e] ≺ e′
e ≺ e′
s : e ≺ e′
e Iss e′
[e] Iss e′
s Iss e′ e Iss e′
s : e Iss e′
ss ⊢ []
[] : [e] ⊢ es
[] ⊢ e : es
s I[] e s ≺ e [] : (s : e) ⊢ es
[] : s ⊢ e : es
¬(s Iss e) ss : s : [e] ⊢ es ¬(s Iss e) s′ Iss:s e s′ ≺ e ss : s : (s′ : e) ⊢ es
ss : s ⊢ e : es
ss : s : s′ ⊢ e : es
| 6cs.PL
|
Dual Control for Approximate Bayesian Reinforcement Learning
Dual Control
for Approximate Bayesian Reinforcement Learning
Edgar D. Klenske
[email protected]
arXiv:1510.03591v2 [stat.ML] 11 Aug 2016
Max-Planck-Institute for Intelligent Systems
Spemannstraße 38
72076 Tübingen, Germany
Philipp Hennig
[email protected]
Max-Planck-Institute for Intelligent Systems
Spemannstraße 38
72076 Tübingen, Germany
Editor: Manfred Opper
Abstract
Control of non-episodic, finite-horizon dynamical systems with uncertain dynamics poses a
tough and elementary case of the exploration-exploitation trade-off. Bayesian reinforcement
learning, reasoning about the effect of actions and future observations, offers a principled
solution, but is intractable. We review, then extend an old approximate approach from
control theory—where the problem is known as dual control —in the context of modern
regression methods, specifically generalized linear regression. Experiments on simulated
systems show that this framework offers a useful approximation to the intractable aspects
of Bayesian RL, producing structured exploration strategies that differ from standard RL
approaches. We provide simple examples for the use of this framework in (approximate)
Gaussian process regression and feedforward neural networks for the control of exploration.
Keywords:
inference
reinforcement learning, control, Gaussian processes, filtering, Bayesian
1. Introduction
The exploration-exploitation trade-off is a central problem of learning in interactive settings,
where the learner’s actions influence future observations. In episodic settings, where the
control problem is re-instantiated repeatedly with unchanged dynamics, comparably simple
notions of exploration can succeed. E.g., assigning an exploration bonus to uncertain options
(Macready and Wolpert, 1998; Audibert et al., 2009) or acting optimally under one sample
from the current probabilistic model of the environment (Thompson sampling, see Thompson,
1933; Chapelle and Li, 2011), can perform well (Dearden et al., 1999; Kolter and Ng, 2009;
Srinivas et al., 2010). Such approaches, however, do not model the effect of actions on future
beliefs, which limits the potential for the balancing of exploration and exploitation. This
issue is most drastic in the non-episodic case, the control of a single, ongoing trial. Here, the
controller cannot hope to be returned to known states, and exploration must be carefully
controlled to avoid disaster.
1
Klenske and Hennig
A principled solution to this problem is offered by Bayesian reinforcement learning (Duff,
2002; Poupart et al., 2006; Hennig, 2011): A probabilistic belief over the dynamics and
cost of the environment can be used not just to simulate and plan trajectories, but also to
reason about changes to the belief from future observations, and their influence on future
decisions. An elegant formulation is to combine the physical state with the parameters
of the probabilistic model into an augmented dynamical description, then aim to control
this system. Due to the inference, the augmented system invariably has strongly nonlinear
dynamics, causing prohibitive computational cost—even for finite state spaces and discrete
time (Poupart et al., 2006), all the more for continuous space and time (Hennig, 2011).
The idea of augmenting the physical state with model parameters was noted early, and
termed dual control, by Feldbaum (1960–1961). It seems both conceptual and—by the
standards of the time—computational complexity hindered its application. An exception is a
strand of several works by Meier, Bar-Shalom, and Tse (Tse et al., 1973; Tse and Bar-Shalom,
1973; Bar-Shalom and Tse, 1976; Bar-Shalom, 1981). These authors developed techniques
for limiting the computational cost of dual control that, from a modern perspective, can
be seen as a form of approximate inference for Bayesian reinforcement learning. While the
Bayesian reinforcement learning community is certainly aware of their work (Duff, 2002;
Hennig, 2011), it has not found widespread attention. The first purpose of this paper is to
cast their dual control algorithm as an approximate inference technique for Bayesian RL in
parametric Gaussian (general least-squares) regression. We then extend the framework with
ideas from contemporary machine learning. Specifically, we explain how it can in principle
be formulated non-parametrically in a Gaussian process context, and then investigate simple,
practical finite-dimensional approximations to this result. We also give a simple, small-scale
example for the use of this algorithm for dual control if the environment model is constructed
with a feedforward neural network rather than a Gaussian process.
2. Model and Notation
Throughout, we consider discrete-time, finite-horizon dynamic systems (POMDPs) of form
xk+1 = fk (xk , uk ) + ξk
yk = Cxk + γk
(state dynamics)
(observation model).
At time k ∈ {0, . . . , T }, xk ∈ Rn is the state, ξk ∼ N (0, Q) is a Gaussian disturbance. The
control input (continuous action) is denoted uk ; for simplicity we will assume scalar uk ∈ R
throughout. Measurements yk ∈ Rd are observations of xk , corrupted by Gaussian noise
γk ∼ N (0, R). The generative model thus reads p(xk+1 ∣ xk , uk ) = N (xk+1 ; fk (xk , uk ), Q)
and p(yk ∣ xk ) = N (yk ; Cxk , R), with a linear map C ∈ Rd×n . Trajectories are vectors
x = [x0 , . . . , xT ], and analogously for u, y. We will occasionally use the subset notation
y i∶j = [yi , . . . , yj ]. We further assume that dynamics fk are not known, but can be described
up to Gaussian uncertainty by a general linear model with nonlinear features φ ∶ Rn _ Rm
and uncertain matrices Ak , Bk .
xk+1 = Ak φ(xk ) + Bk uk + ξk ,
Ak ∈ Rn×m ; Bk ∈ Rn×1 .
(1)
To simplify notation, we reshape the elements of Ak and Bk into a parameter vector
θk = [vec(Ak ); vec(Bk )] ∈ R(m+1)n , and define the reshaping transformations A(θk ) ∶ θk ↦ Ak
2
Dual Control for Approximate Bayesian Reinforcement Learning
and B(θk ) ∶ θk ↦ Bk . At initialization, k = 0, the belief over states and parameters is
assumed to be Gaussian
x̂0
x
x
Σxx Σxθ
0 ]) .
p ([ 0 ]) = N ([ 0 ] ; [ ] , [ 0θx
θ0
θ0
Σ0 Σθθ
θ̂0
0
(2)
The control response Bk uk is linear, a common assumption for physical systems. Nonlinear
mappings can be included in a generic form φ(xk , uk ), but complicate the following derivations
and raise issues of identifiability. For simplicity, we also assume that the dynamics do not
change through time: p(θk+1 ∣ θk ) = δ(θk+1 − θk ). This could be relaxed to an autoregressive
model p(θk+1 ∣ θk ) = N (θk+1 ; Dθk , Ξ), which would give additive terms in the derivations
below. Throughout, we assume a finite horizon with terminal time T and a quadratic cost
function in state and control
T
T −1
k=0
k=0
L(x, u) = [ ∑ (xk − rk )⊺ Wk (xk − rk ) + ∑ u⊺k Uk uk ] ,
where r = [r0 , . . . , rT ] is a target trajectory. Wk and Uk define state and control cost, they
can be time-varying. The goal, in line with the standard in both optimal control and
reinforcement learning, is to find the control sequence u that, at each k, minimizes the
expected cost to the horizon
Jk (uk∶T −1 , p(xk )) = Exk [(xk − rk )⊺ Wk (xk − rk ) + u⊺k Uk uk + Jk+1 (uk+1∶T −1 , p(xk+1 )) ∣ p(xk )] ,
(3)
where past measurements y 1∶k , controls u1∶k−1 and prior information p(x0 ) are incorporated
into the belief p(xk ), relative to which the expectation is calculated. Effectively, p(xk ) serves
as a bounded rationality approximation to the true information state. Since the equation
above is recursive, the final element of the cost has to be defined differently, as
JT (p(xT )) = ExT [(xT − rT )⊺ WT (xT − rT ) ∣ p(xT )]
(that is, without control input and future cost). The optimal control sequence minimizing
this cost will be denoted u∗ , with associated cost
∗
Jk∗ (p(xk )) = min Exk [(xk − rk )⊺ Wk (xk − rk ) + u⊺k Uk uk + Jk+1
(p(xk+1 )) ∣ p(xk )] .
uk
(4)
This recursive formulation, if written out, amounts to alternating minimization and expectation steps. As uk influences xk+1 and yk+1 , it enters the latter expectation nonlinearly.
Classic optimal control is the linear base case (φ(x) = x) with known θ, where u∗ can be
found by dynamic programming (Bellman, 1961; Bertsekas, 2005).
3. Bayesian RL and Dual Control
Feldbaum (1960–1961) coined the term dual control to describe the idea now also known as
Bayesian reinforcement learning in the machine learning community: While adaptive control
only considers past observations, dual control also takes future observations into account.
This is necessary because all other ways to deal with uncertain parameters have substantial
3
Klenske and Hennig
drawbacks. Robust controllers, for example, sacrifice performance due to their conservative
design; adaptive controllers based on certainty equivalence (where the uncertainty of the
parameters is not taken into account but only their mean estimates) do not show exploration,
so that all learning is purely passive. For most systems it is obvious that more excitation
leads to better estimation, but also to worse control performance. Attempts at finding a
compromise between exploration and exploitation are generally subsumed under the term
“dual control” in the control literature. It can only be achieved by taking the future effect of
current actions into account.
It has been shown that optimal dual control is practically unsolvable for most cases (Aoki,
1967), with a few examples where solutions were found for simple systems (e.g., Sternby,
1976). Instead, a large number of approximate formulations of the dual control problem were
formulated in the decades since then. This includes the introduction of perturbation signals
(e.g., Jacobs and Patchell, 1972), constrained optimization to limit the minimal control
signal or the maximum variance, serial expansion of the loss function (e.g., Tse et al., 1973)
or modifications of the loss function (e.g., Filatov and Unbehauen, 2004). A comprehensive
overview of dual control methods is given by Wittenmark (1995). A historical side-effect of
these numerous treatments is that the meaning of the term “dual control” has evolved over
time, and is now applied both to the fundamental concept of optimal exploration, and to
methods that only approximate this notion to varying degree. Our treatment below studies
one such class of practical methods that aim to approximate the true dual control solution.
The central observation in Bayesian RL / dual control is that both the states x and
the parameters θ are subject to uncertainty. While part of this uncertainty is caused by
randomness, part by lack of knowledge, both can be captured in the same way by probability
distributions. States and parameters can thus be subsumed in an augmented state (Feldbaum,
1960–1961; Duff, 2002; Poupart et al., 2006) zk⊺ = (x⊺k θk⊺ ) ∈ R(m+2)n . In this notation,
the optimal exploration-exploitation trade-off—relative to the probabilistic priors defined
above—can be written compactly as optimal control of the augmented system with a new
observation model p(yk ∣ zk ) = N (yk ; C̃zk , R) using C̃ = [C 0] and a cost analogous to
Eq. (3).
Unfortunately, the dynamics of this new system are nonlinear, even if the original physical
system is linear. This is because inference is always nonlinear and future states influence
future parameter beliefs, and vice versa. A first problem, not unique to dual control, is
thus that inference is not analytically tractable, even under the Gaussian assumptions
above (Aoki, 1967). The standard remedy is to use approximations, most popularly the
linearization of the extended Kalman filter (e.g., Särkkä, 2013). This gives a sequence of
approximate Gaussian likelihood terms. But even so, incorporating these Gaussian likelihood
terms into future dynamics is still intractable, because it involves expectations over rational
polynomial functions, whose degree increases with the length of the prediction horizon. The
following section provides an intuition for this complexity, but also the descriptive power of
the augmented state space.
As an aside, we note that several authors (Kappen, 2011; Hennig, 2011) have previously
pointed out another possible construction of an augmented state: incorporating not the
actual value of the parameters θk in the state, but the parameters µk , Σk of a Gaussian belief
p(θk ∣ µk , Σk ) = N (θk ; µk , Σk ) over them. The advantage of this is that, if the state xk is
observed without noise, these belief parameters follow stochastic differential equations—more
4
Dual Control for Approximate Bayesian Reinforcement Learning
precisely, Σk follows an ordinary (deterministic) differential equation, while µk follows a
stochastic differential equation—and it can then be attempted to solve the control problem
for these differential equations more directly.
While it can be a numerical advantage, this formulation of the augmented state also has
some drawbacks, which is why we have here decided not to adopt it: First, the simplicity of
the directly formalizable SDE vanishes in the POMDP setting, i.e. if the state is not observed
without noise. If the state observations are corrupted, the exact belief state is not a Gaussian
process, so that the parameters µk and Σk have no natural meaning. Approximate methods
can be used to retain a Gaussian belief (and we will do so below), but the dynamics of µk ,
Σk are then intertwined with the chosen approximation (i.e. changing the approximation
changes their dynamics), which causes additional complication. More generally speaking, it
is not entirely natural to give differing treatment to the state xk and parameters θk : Both
state and parameters should thus be treated within the same framework; this also allows
extending the framework to the case where also the parameters do follow an SDE.
3.1 A Toy Problem
To provide an intuition for sheer complexity of optimal dual control, consider the perhaps
simplest possible example: the linear, scalar system
xk+1 = axk + buk + ξk ,
(5)
with target rk = 0 and noise-free observations (R = 0). If a and b are known, the optimal uk
to drive the current state xk to zero in one step can be trivially verified to be
u∗k,oracle = −
abxk
.
U + b2
Let now parameter b be uncertain, with current belief p(b) = N (b; µk , σk2 ) at time k. The
naı̈ve option of simply replacing the parameter with the current mean estimate is known as
certainty equivalence (CE) control in the dual control literature (e.g., Bar-Shalom and Tse,
1974). The resulting control law is
u∗k,ce = −
aµk xk
.
U + µ2k
It is used in many adaptive control settings in practice, but has substantial deficiencies: If
the uncertainty is large, the mean is not a good estimate, and the CE controller might apply
completely useless control signals. This often results in large overshoots at the beginning or
after parameter changes.
A slightly more elaborate solution is to compute the expected cost Eb [x2k+1 + U u2k ∣ µk , σk2 ]
and then optimize for uk . This gives optimal feedback (OF) or “cautious” control (Dreyfus,
1964)1 :
aµk xk
u∗k,of = −
.
(6)
U + σk2 + µ2k
1. Dreyfus used the term “open loop optimal feedback” for his approach, a term that is misleading to
modern readers, because it is in fact a closed-loop algorithm.
5
Klenske and Hennig
This control law reduces control actions in cases of high parameter uncertainty. This
mitigates the main drawback of the CE controller, but leads to another problem: Since
the OF controller decreases control with rising uncertainty, it can entirely prevent learning.
Consider the posterior on b after observing xk+1 , which is a closed-form Gaussian (because
uk is chosen by the controller and has no uncertainty):
2
2
) = N (b; µk+1 , σk+1
) = N (b;
p(b ∣ µk+1 , σk+1
σk2 uk (buk + ξk ) + µk Q
σk2 Q
,
)
u2k σk2 + Q
u2k σk2 + Q
(7)
(b shows up in the fully observed xk+1 = axk + buk + ξk ). The dual effect here is that the
2
updated σk+1
depends on uk . For large values of σk2 , according to (6), u∗k,of _ 0, and the
2
new uncertainty σk+1
_ σk2 . The system thus will never learn or act, even for large xk . This
is known as the “turn-off phenomenon” (Aoki, 1967; Bar-Shalom, 1981).
However, the derivation for OF control above amounts to minimizing Eq. (3) for the
myopic controller, where the horizon is only a single step long (T = 1). Therefore, OF control
is indeed optimal for this case. By the optimality principle (e.g., Bertsekas, 2005), this
means that Eq. (6) is the optimal solution for the last step of every controller. But since it
does not show any form of exploration or “probing” (Bar-Shalom and Tse, 1976), a myopic
controller is not enough to show the dual properties.
In order to expose the dual features, the horizon has to be at least of length T = 2. Since
the optimal controller follows Bellman’s equation, the solution proceeds backwards. The
solution for the second control action u1 is identical to the solution of the myopic controller
(6); but after applying the first control action u0 , the belief over the unknown parameter b
needs an update according to Eq. (7), resulting in
u∗1
⎡
2 ⎤−1
⎢
σ02 Q
σ02 u0 (bu0 + ξ0 ) + µ0 Q ⎥⎥
σ02 u0 (bu0 + ξ0 ) + µ0 Q
⎢
= − ⎢U + 2 2
+(
)
[a
x1 ] .
⎥
u0 σ0 + Q
u20 σ02 + Q
u20 σ02 + Q
⎢
⎥
⎣
⎦
(8)
Inserting into Eq. (4) gives
J0∗ (x0 ) = min Ex0 [W x20 + U u20 + min Ex1 [W x21 + U u21 + Ex2 [W x22 ]]]
u1
u0
=
min [W x20
u
0
+ U u20
+ Eξ0 ,b [W x21 + U (u∗1 )2 + Eξ1 ,b [W (x1 + bu∗1 + ξ1 )2 ∣ µ1 , σ1 ] ∣ µ0 , σ0 ]] .
(9)
Since u∗1 from Eq. (8) is already a rational function of fourth order in b0 , and shows up
quadratically in Eq. (9), the relevant expectations cannot be computed in closed form (Aoki,
1967). For this simple case though, it is possible to compute the optimal dual control by
performing the expectation through sampling b, ξ0 , ξ1 from the prior. Fig. 1 shows such
samples of L(u0 ) (in gray; one single sample highlighted in orange), and the empirical
expectation J(u0 ) in dashed green. Each sample is a rational function of even leading order.
In contrast to the CE cost, the dual cost is much narrower, leading to more cautious behavior
of the dual controller. The average dual cost has its minima not at zero, but to either side
of it, reflecting the optimal amount of exploration in this particular belief state.
While it is not out of the question that the Monte Carlo solution can remain feasible
for larger horizons, we are not aware of successful solutions for continuous state spaces
6
4
4
3
3
cost
cost
Dual Control for Approximate Bayesian Reinforcement Learning
2
1
0
−1
2
1
−0.5
0
u0
0.5
0
−1
1
−0.5
0
u0
0.5
1
Figure 1: Left: Computing the T = 2 dual cost for the simple system of Eq. (5). Costs
L(u0 ) under optimal control on u1 for sampled parameter b (thin gray; one sample
highlighted, orange). Expected dual cost J(u0 ) under u∗1 (dashed green). The
optimal u∗0 lies at the minimum of the dashed green line. Right: Comparison
of sampling (dashed green; thin gray: samples) to three approximations: CE
(red) and CE with Bayesian exploration bonus (blue). The solid green line is the
approximate dual control constructed in Section 4. See also Sec. 6.1 for details.
(however, see Poupart et al., 2006, for a sampling solution to Bayesian reinforcement learning
in discrete spaces, including notes on the considerable computational complexity of this
approach). The next section describes a tractable analytic approximation that does not
involve samples.
4. Approximate Dual Control for Linear Systems
In 1973, Tse et al. (1973) constructed theory and an algorithm (Tse and Bar-Shalom, 1973)
for approximate dual (AD) control, based on the series expansion of the cost-to-go. This is
related to differential dynamic programming for the control of nonlinear dynamic systems
(Mayne, 1966). It separates into three conceptual steps (described in Sec. 4.1–4.3), which
together yield what, from a contemporary perspective, amounts to a structured Gaussian
approximation to Bayesian RL:
¬ Find an optimal trajectory for the deterministic part of the system under the mean
model: the nominal trajectory under certainty equivalent control. For linear systems
this is easy (see below), for nonlinear ones it poses a nontrivial, but feasible nonlinear
model predictive control problem (Allgöwer et al., 1999; Diehl et al., 2009). It yields a
nominal trajectory, relative to which the following step constructs a tractable quadratic
expansion.
Around the nominal trajectory, construct a local quadratic expansion that approximates
the effects of future observations. Because the expansion is quadratic, an optimal
control law relative to the deterministic system—the perturbation control —can be
constructed by dynamic programming. Plugging this perturbation control into the
residual dynamics of the approximate quadratic system gives an approximation for the
cost-to-go. This step adds the cost of uncertainty to the deterministic control cost.
7
Klenske and Hennig
Compute
CE control
Initialize
predict state
and covariance
for given uk
® Compute next value
uk for the search
no
¬ Compute
CE trajectory
and its covariances
Make new
measurement
Evaluate
cost-to-go
Simulate or run
the system
yes
search over?
Apply the
control
Figure 2: Flow-chart of the approximate dual control algorithm to show the overall structure.
Adapted from Tse and Bar-Shalom (1973). The left cycle is the inner loop,
performing the nonlinear optimization.
® In the current time step k, perform the prediction for an arbitrary control input uk
(as opposed to the analytically computed control input for later steps). Optimize uk
numerically by repeated computation of steps ¬ and at varying uk to minimize the
approximate cost.
These three steps will be explained in detail in the subsequent sections. The interplay
between the different parts of the algorithm is shown in Figure 2.
The abstract introductory work Tse et al. (1973) is relatively general, but the explicit
formulation in Tse and Bar-Shalom (1973) only applies to linear systems. Since both works
are difficult to parse for contemporary readers, the following sections thus first provide a
short review, before we extend to more modern concepts. In this section, we follow the
more transparent case of a linear system from Tse and Bar-Shalom (1973), i.e. φ(x) = x
in Eq. (1). For the augmented state z, this still gives a nonlinear system, because θ and x
interact multiplicatively
zk+1 = (
xk+1
A(θk ) 0
B(θk )
ξ
)=(
)z + (
) uk + ( k ) =∶ f˜(zk , uk ).
θk+1
0
I k
0
0
(10)
The parameters θ are assumed to be deterministic, but not known to the controller. This
uncertainty is captured by the distribution p(θ) representing the lack of knowledge.
4.1 Certainty Equivalent Control Gives a Nominal Reference Trajectory
The certainty equivalent model is built on the assumption that the uncertain θ coincide with
their most likely value, the mean θ̂ of p(θ), and that the system propagates deterministically
without noise. This means that the nominal parameters θ̄ are the current mean values θ̂,
8
Dual Control for Approximate Bayesian Reinforcement Learning
which decouples θ entirely from x in Eq. (10), and the optimal control for the finite horizon
problem can be computed by dynamic programming (DP) (Aoki, 1967), yielding an optimal
linear control law
−1
ū∗j = − (B̄ ⊺ K̄j+1 B̄ + Uj ) B̄ ⊺ [K̄j+1 Āx̄j + p̄j+1 ] ,
where we have momentarily simplified notation to Ā = A(θ̄j ), B̄ = B(θ̄j ), ∀j, because the θ̄j
are constant. The K̄j and p̄j for j = k + 1, . . . , T are defined and computed recursively as
K̄j = Ā⊺ (K̄j+1 − K̄j+1 B̄ (B̄ ⊺ K̄j+1 B̄ + Uj )
−1
p̄j = Ā⊺ (p̄j+1 − K̄j+1 B̄ (B̄ ⊺ K̄j+1 B̄ + Uj )
−1
B̄ ⊺ K̄j+1 ) Ā + Wj
B̄ ⊺ K̄j+1 ) p̄j+1 − Wj rj
K̄T = WT
p̄T = −WT rT ,
where r is the reference trajectory to be followed. This CE controller gives the nominal
trajectory of inputs ūk∶T −1 and states x̄k∶T , from the current time k to the horizon T . The true
future trajectory is subject to stochasticity and uncertainty, but the deterministic nominal
trajectory x̄, with its optimal control ū∗ and associated nominal cost J¯k∗ = L(x̄k∶T , ū∗k∶T )
provides a base, relative to which an approximation will be constructed.
4.2 Quadratic Expansion Around the Nominal Defines Cost of Uncertainty
The central idea of AD control is to project the nonlinear objective Jk (uk∶T −1 , p(xk ))
of Eq. (3) into a quadratic, by locally linearizing around the nominal trajectory x and
maintaining a joint Gaussian belief.
To do so, we introduce small perturbations around nominal cost, states, and control:
∆Jj = Jj − J¯j , ∆zj = zj − z̄j , and ∆uj = uj − ūj . These perturbations arise from both the
stochasticity of the state and the parameter uncertainty. Note that a change in the state
results in a change of the control signal, because the optimal control signal in each step
depends on the state. Even though the origin of the uncertainties is different (∆x arises
from stochasticity and ∆θ from the lack of knowledge), both can be modeled in a joint
probability distribution.
Approximate Gaussian filtering ensures that beliefs over ∆z remain Gaussian:
Σxx Σxθ
∆x
∆x̂
j
p(∆zj ) = N [( j ) ; ( j ) , ( jθx
)] .
∆θj
0
Σj
Σθθ
j
Note that shifting the mean to the nominal trajectory does not change the uncertainty.
Note further that the expected perturbation in the parameters is nil. This is because the
parameters are assumed to be deterministic and are not affected by any state or input.
Calculating the Gaussian filtering updates is in principle not possible for future measurements, since it violates the causality principle (Glad and Ljung, 2000). Nonetheless, it is
possible to use the expected measurements to simulate the effects of the future measurements
on the uncertainty, since these effects are deterministic. This is sometimes referred to as
preposterior analysis (Raiffa and Schlaifer, 1961).
To second order around the nominal trajectory, the cost is approximated by
Jk (uk∶T −1 , p(xk )) = J¯k∗ + ∆Jk ≈ J¯k∗ + ∆J˜k ,
9
Klenske and Hennig
where J¯k∗ is the optimal cost for the nominal system and ∆J˜k is the approximate additional
cost from the perturbation:
⎤
⎡T
T −1
⎥
⎢
1
1
∆J˜k ∶= Exk∶T ⎢⎢ ∑ {(x̄j − rj )⊺ Wj ∆xj + ∆x⊺j Wj ∆xj } + ∑ {ū⊺j Uj ∆uj + ∆u⊺j Uj ∆uj }⎥⎥ .
2
2
⎥
⎢j=k
j=k
⎦
⎣
(12)
Although the uncertain parameters θ do not show up explicitly in the above equation, this
step captures dual effects: The uncertainty of the trajectory ∆x depends on θ via the
dynamics. Higher uncertainty over θ at time j − 1 causes higher predictive uncertainty over
∆xj (for each j), and thus increases the expectation of the quadratic term ∆x⊺j Wj ∆xj .
Control that decreases uncertainty in θ can lower this approximate cost, modeling the benefit
of exploration. For the same reason, Eq. (12) is in fact still not a quadratic function and has
no closed form solution. To make it tractable, Tse and Bar-Shalom (1973) make the ansatz
that all terms in the expectation of Eq. (12) can be written as gj + p⊺j ∆zj + 1/2∆zj⊺ Kj ∆zj .
This amounts to applying dynamic programming on the perturbed system. Expectations
over the cost under Gaussian beliefs on ∆z can then be computed analytically. Because all
∆θ have zero mean, linear terms in these quantities vanish in the expectation. This allows
analytic minimization of the approximate optimal cost for each time step
1
1
∆J˜j∗ (p(xj )) = min {(xj − rj )⊺ Wj ∆x̂j∣j + ∆x̂⊺j∣j Wj ∆x̂j∣j + u⊺j U ∆uj + ∆u⊺j U ∆uj
∆uj
2
2
1
˜∗
+ tr [Wj Σxx
j∣j ] + E [∆Jj+1 (y 1∶j+1 ) ∣ p(xj )]} , (13)
∆xj+1
2
which is feasible given an explicit description of the Gaussian filtering update. It is important
to note that, assuming extended Kalman filtering, the update to the mean from expected
future observations yj+1 is nil. This is because we expect to see measurements consistent
with the current mean estimate. Nonetheless, the (co-)variance changes depending on the
control input uj , which is the dual effect.
Following the dynamic programming equations for the perturbed problem, including the
additional cost from uncertainty, the resulting cost amounts to (Tse et al., 1973)
1
∆J˜k∗ (p(xk )) = g̃k+1 + p̃⊺k+1 ∆ẑk + ∆ẑk⊺ K̃k+1 ∆ẑk
2
⎧
⎫
T −1
⎪
1 ⎪
⎪
⎪
xx
+ tr ⎨WT Σxx
+
[W
Σ
+
(Σ
−
Σ
)
K̃
]
∑
j j∣j
j+1 ⎬
j+1∣j
j+1∣j+1
T
∣T
⎪
2 ⎪
⎪
⎪
j=k
⎩
⎭
(where we have neglected second-order effects of the dynamics). Recalling that ∆ẑ = 0 and
dropping the constant part, the dual cost can be approximated to be
Jkd
⎧
⎫
T −1
⎪
⎪
1 ⎪
⎪
xx
xx
= tr ⎨WT ΣT ∣T + ∑ [Wj Σj∣j + (Σj+1∣j − Σj+1∣j+1 )K̃j+1 ]⎬
⎪
2 ⎪
⎪
⎪
j=k
⎩
⎭
(= ∆J˜k∗ − const)
where the recursive equation
xx
K̃j = Ã⊺ (K̃j+1 − K̃j+1 B̃ (B ⊺ Kj+1
B + Uj )
10
−1
B̃ ⊺ K̃j+1 ) Ã + W̃j
K̃T = W̃T
Dual Control for Approximate Bayesian Reinforcement Learning
∂ ˜
∂ ˜
is defined for the augmented system (10), with à = ∂z
f , B̃ = ∂u
f and W̃j = blkdiag(Wj , 0).
∗
d
¯
The approximation to the overall cost is then Jk + Jk , which is used in the subsequent
optimization procedure.
4.3 Optimization of the Current Control Input Gives Approximate Dual
Control
The last step ® amounts to the outer loop of the overall algorithm. A gradient-free black-box
optimization algorithm is used to find the minimum of the dual cost function. In every step,
this algorithm proposes a control input uk for which the dual cost is evaluated.
Depending on uk , approximate filtering is carried out to the horizon. The perturbation
control is plugged into Eq. (13) to give an analytic, recursive definition for K̃j , and an
approximation for the dual cost Jkd , as a function of the current control input uk .
Nonlinear optimization—through repetitions of steps ¬ and for proposed locations
uk —then yields an approximation to the optimal dual control u∗k . Conceptually the simplest
part of the algorithm, this outer loop dominates computational cost, because for every
location uk the whole machinery of ¬ and has to be evaluated.
5. Extension to Contemporary Machine Learning Models
The preceding section reviewed the treatment of dual control in linear dynamical systems
from Tse and Bar-Shalom (1973). In this section, we extend the approach to inference on,
and dual control of, the dynamics of nonlinear dynamical systems. This extension is guided
by the desire to use a number of popular, standard regression frameworks in machine learning:
Parametric general least-squares regression, nonparametric Gaussian process regression, and
feedforward neural networks (including the base case of logistic regression).
5.1 Parametric Nonlinear Systems
We begin with the generalized linear model mentioned in Eq. (1). The nonlinear features
φ can in principle be any function (popular choices include sines and cosines, radial basis
functions, sigmoids, polynomials and others), with the caveat that their structure crucially
influences the properties of the model. From a modeling perspective, this approach is quite
standard for machine learning. However, the dynamical learning setting requires a few
adaptations: First, to allow the modeling of higher-order dynamical systems, the original
states must to be included. This gives features of the form φ(x)⊺ = (x⊺ ϕ(x)⊺ ), consisting
of the linear representation, augmented by general features ϕ.
The next challenge is that the optimal control for nonlinear dynamical systems cannot
be optimized in closed form using dynamic programming, not even for the deterministic
nominal system. Instead, we find the nominal reference trajectory using nonlinear model
predictive control (Allgöwer et al., 1999; Diehl et al., 2009). In our case, we begin with
dynamic programming on a locally linearized system, then optimize nonlinearly with a
numerical method across the trajectory. This adds computational cost, and requires some
care to achieve stable optimization performance for specific system setups.
Filtering from observations is also more involved in the case of nonlinear dynamics. In
the experiments reported below, we stayed within the extended Kalman filtering framework
11
Klenske and Hennig
to retain Gaussian beliefs over the states and parameters. Extensions of this approach to
more elaborate filtering methods are an interesting direction for future work. This includes
relatively standard options like unscented Kalman filtering (Uhlmann, 1995), but also more
recent developments in machine learning and probabilistic control, such as analytic moment
propagation where the features ϕ allow this (e.g., Deisenroth and Rasmussen, 2011).
The final problem is the generalization of the derivations from the preceding sections
to the nonlinear dynamics. We take a relatively simplistic approach, which nevertheless
turns out to work well. A linearization gives locally linear dynamics whose structure closely
matches Eq. (10):
x̄
+ ∆xk+1
A(θk ) 0 φ(x̄k + ∆xk )
B(θk )
ξ
zk+1 = ( k+1
)=(
)(
)+(
) uk + ( k )
θ̄k+1 + ∆θk+1
0
I
θ̄k + ∆θk
0
0
A(θ̄k ) 0 φ(x̄k )
B(θ̄k )
ξ
≈(
)(
)+(
) uk + ( k )
0
I
θ̄k
0
0
+(
A(θ̄k ) ∂x∂ k φ(x̄k )
0
∂
∂θk
(A(θ̄k )φ(x̄k ) + B(θ̄k )uk ) ∆xk
)(
).
∆θk
I
This essentially amounts to extended Kalman filtering on the augmented state. Using this
linearization, the approximation described in Sec. 4 can be applied analogously.
5.2 Nonparametric Gaussian Process Dynamics Models
The above treatment of parametric linear models makes it comparably easy to extend the
description from finitely many feature functions to an infinite-dimensional feature space
defining a Gaussian process (GP) dynamics model: Assume that the true dynamics function
f is a draw from a Gaussian process prior p(f ) = GP(f ; m, κ̄) with prior mean function
m ∶ Rn _ Rn , and prior covariance function (kernel) κ̄ ∶ Rn × Rn _ Rn × Rn . This is using
the widely used notion of “multi-output regression” (Rasmussen and Williams, 2006, § 9.1),
i.e. formulating the covariance as
cov(fi (x), fj (x′ )) = κ̄ij (x, x′ ).
To simplify the treatment, we will assume that the covariance factors between inputs and
outputs, i.e. κ̄ij (x, x′ ) = Vij κ(x, x′ ) with a univariate kernel κ ∶ Rn × Rn _ R and a positive
semi-definite matrix V ∈ Rn×n of output covariances. By Mercer’s theorem (e.g., König,
1986; Rasmussen and Williams, 2006), the kernel can be decomposed into a converging series
over eigenfunctions φ(x), as
∞
κ(x, x′ ) = ∑ λ` φ` (x)φ∗` (x′ ),
(14)
`=1
where φ` ∶ Rn _ R are functions that are orthonormal relative to some measure µ over Rn
(the precise choice of which is irrelevant for the time being), with the property
′
′
′
∫ κ(x, x )φ` (x ) dµ(x ) = λ` φ` (x).
12
Dual Control for Approximate Bayesian Reinforcement Learning
Precisely in this sense, Gaussian process regression can be written as “infinite-dimensional”
Bayesian linear regression. We will use the suggestive, and somewhat abusive notation
fk (xk ) = LΩk Φ(xk ) for this generative model, defined as
T
∞
j=1
`=1
fki (xk ) = ∑ Lij ∑ Ωj`
k φi (xk )
(15)
where L is a matrix satisfying LL⊺ = V (e.g., the Cholesky decomposition), and the elements
of Ω are draws from the “white” Gaussian process Ωj`
k ∼ N (0, λ` ). Because of Mercer’s
2
theorem above, Eq. (15) exists in µ expectation, and is well-defined in this sense. This
notation allows writing Eq. (2) as a nonparametric prior with mean θ̂0 and covariance
⊺
Σθθ
0 = V ⊗ (ΦΛΦ ) where Λ is an infinite diagonal matrix with diagonal elements Λ`` = λ`
(the matrix multiplication ΦΛΦ⊺ is here defined as in Eq. (15)).
Using this notation, a tedious but straightforward linear algebra derivation (see Appendix A) shows that the posterior over z ⊺ = (x⊺ θ⊺ ) after a number k of EKF-linearized
Gaussian observations is a tractable Gaussian process, for which the Gram matrix
G = P + Q + K + F −1 RF −⊺
consists of the parts
P 0
P =[ 0 ]
0 0
Q = (Q ⊗ I)
K = κ(y 1∶m , y 1∶m )
R = (R ⊗ I)
of appropriate size, depending on the current time k. The multi-step state transition matrix
⎡ I
0
0
⎢
⎢ A1
I
0
⎢
⎢
A2
I
F = ⎢ A2 A1
⎢
⎢
⋮
⎢
⎢A ⋯A A ⋯A ⋯
⎣ m
1
m
2
⋯ 0⎤⎥
⋯ 0⎥⎥
⎥
⋯ 0⎥
⎥
⋱ ⎥⎥
I ⎥⎦
with
⎡ I
0
0
⋯
⎢
⎢−A1
I
0
⋯
⎢
⎢
−A2 I
⋯
F −1 = ⎢ 0
⎢
⎢ ⋮
⋱ ⋱
⋱
⎢
⎢ 0
⋯
0
−A
⎣
m
0⎤⎥
0⎥⎥
⎥
0⎥
⎥
⎥
⎥
I ⎥⎦
is needed to account for the effect of the measurement noise R over time. The A-matrices
are the Jacobians ∇x f (x)∣xk .
The posterior mean now evaluates to
[
x̂k
x̂
Φ(x̂k−1 )ΛΦ(y 1∶k−1 )⊺ −1
] = [ k−1 ] + [
] G [Φ(y 1∶k−1 )ΛΦ(x̂k−1 )⊺ Φ(y 1∶k−1 )Λ] (16)
0
ΛΦ(y 1∶k−1 )⊺
θ̂k
and the posterior covariance is comprised of
xx ⊺
θx ⊺
xθ ⊺
θθ ⊺
Σ̄xx
k = Ak Σk Ak + Q + Φk Σk Ak + Ak Σk Φk + Φk Σ Φk
(17a)
xx
xx
xx
Σxx
k = Σ̄k−1 − Σ̄k−1 [Σ̄k−1 + R]
(17b)
−1
Σ̄xx
k−1
−1
Σxθ
k = Fk,∶ Φ(y)Λ − Fk,∶ [P + K + Q] G Φ(y)Λ
Σθx
k
Σθθ
k
=
⊺
(Σxθ
k )
(17c)
(17d)
= Λ − ΛΦ(y) G Φ(y)Λ.
⊺ −1
13
(17e)
Klenske and Hennig
This formulation, together with the expositions in the preceding sections, defines a
nonparametric dual control algorithm for Gaussian process priors. It is important to stress
that this posterior is indeed “tractable” in so far as it depends only on a Gram matrix of
size nT × nT , and the posterior over any f (x) can be computed in time O((nT )3 ), despite
the infinite-dimensional state space.
5.2.1 An Approximation of Constant Cost
In practical control applications, continuously rising inference cost is rarely acceptable. It is
thus necessary to project the GP belief onto a finite representation, replacing the infinite
sum in Eq. (14) with a finite one, to bound the computational cost of the matrix inversion in
Eqs. (16), (17c) and (17e). We do so by projecting into a pre-defined finite basis of functions
drawn from the eigen-spectrum of the kernel with respect to the Lebesgue measure. This
approach has been recently popular elsewhere in regression (Rahimi and Recht, 2008). For
readers unaware of this line of work, here is a short, self-contained introduction:
By Bochner’s theorem (e.g., Stein, 1999; Rasmussen and Williams, 2006), the covariance
function k(r) (with r = ∣x − x′ ∣) of a stationary µ2 continuous random process can be
represented as the Fourier transform of a positive finite measure and, if that measure has a
density S(s), as the Fourier dual of S:
κ(r) = ∫ S(s)e2πisr ds,
This means that the eigenfunctions of the kernel are trigonometric functions, and stationary
covariance functions, like the commonly used square exponential kernel
κse (x, x′ ) = exp (−
(x − x′ )2
),
2λ2
can be approximated by sine and cosine basis functions as
√
κ(x, x ) ≈ κ̃(x, x ) =
′
′
F
2 /2
′
′
∑ sin(ω2i−1 ∣x − x ∣) + cos(ω2i ∣x − x ∣),
F i=1
where the frequencies ωi of the feature functions is sampled from the power spectrum of
the process. An example of such kernel approximation is shown in Fig. 3. With increasing
number of features, the approximation can be chosen as closely to the true covariance
function as needed, while keeping the number of features in a range that is still feasible
within the time constraints of the control algorithm.
5.3 Dual Control of Feedforward Neural Networks
Another extension of the parametric linear models of Section 5.1 is to allow for a nonlinear
parametrization of the dynamics function:
F
f (x; θ) = ∑ θilin φi (x; θinonlin ).
i
14
Dual Control for Approximate Bayesian Reinforcement Learning
full kernel
prior
4
2
2
0
0
−2
−2
−4
−4
kernel appximation
posterior
4
−2
0
2
−4
−4
4
4
4
2
2
0
0
−2
−2
−4
−4
−2
0
2
−4
−4
4
kernel
1
0.5
−2
0
2
4
0
−4
−2
0
2
4
−2
0
2
4
1
0.5
−2
0
2
4
0
−4
Figure 3: Prior (left), posterior (middle) and kernel function (right) of both the full kernel
function (top row) and the approximate kernel (bottom row). The thick lines
represent the mean and the thin lines show two standard deviations. The dashed
lines are samples from the shown distributions.
x1k+1
x2k+1
xik+1 = f i (xk ) ≈ ∑j v ij zkj
x3k+1
x4k+1
zk4
zk5
zk6
zki = σ (∑j wij xjk + bi )
x3k
x4k
b
xk
v
zk1
zk2
zk3
w
x1k
x2k
Figure 4: A two-layer feedforward neural network. Sketch to illustrate the structure of
Eq. (18).
15
Klenske and Hennig
A particularly interesting example of this structure are multilayer perceptrons. Consider a
two-layer network with logistic link function
f (x) = ∑ v i σ(wi x + bi ),
(18)
i
where v are the weights from the latent to the output layer, w are the weights from input
to hidden units, and b are the biases of the hidden units (see Fig. 4).
Neural networks are used in control quite regularly, see e.g., Nguyen and Widrow (1990).
Instead of using backpropagation and stochastic gradient descent as in most applications
of neural networks (Rumelhart et al., 1986; Robbins and Monro, 1951), the EKF inference
procedure can be used to train the weights as well (Singhal and Wu, 1989). This is possible
because the EKF linearization can also be applied for the nonlinear link function, e.g., the
logistic function. Speaking in terms of feature functions, not only the weight of each feature
but also the shape (steepness) can be inferred. A limiting factor for this inference naturally
is the number of data points: the more features and parameters are introduced, the more
data points are necessary to learn.
Using the state augmentation z ⊺ = (x⊺ v ⊺ w⊺ ), and linearizing w.r.t. all parameters in
each step, the EKF inference on the neural network parameters allows us not only to apply
relatively cheap inference on them, but also to use the dual control framework to plan control
signals, accounting for the effect of future observations and the subsequent change in the
belief. This means the adaptive dual controller described in Sec. 4 can identify those parts
of the neural net that are relevant for applying optimal control to the problem at hand. In
Sec. 6.3, we show an experiment with these properties.
6. Experiments
A series of experiments on single-episode tasks with continuous state space highlights
qualitative differences between the adaptive dual (AD) controller and three other controllers:
An oracle controller with access to the true parameters, which provides an unattainable
lower bound (LB) on the achievable performance, a certainty equivalent (CE) controller
as described in Sec. 4.1, and a controller minimizing the sum of CE cost and the Bayesian
exploration bonus (BEB) (see Kolter and Ng, 2009):
`beb = τ [sqrt (diag (Σθθ ))] [sqrt (diag (Σθθ ))]
⊺
(τ is a scalar exploration weight). The additional cost term `beb is evaluated for the predicted
parameter covariance where the prediction time is chosen according to the order of the
system such that the effect of the current control signal shows up in the belief over the
parameters. This type of controller is sometimes also called dual control, while being referred
to as explicit dual control, where the dual features are obtained by a modified cost function
(Filatov and Unbehauen, 2000).
Every experiment was repeated 50 times with different random seeds, which were shared
across controllers for comparability. All systems presented below are very simple setups.
Their primary point is to show qualitative differences of the controllers’ behavior. The
experiments were done with different approximations from the preceding section to show
experimental feasibility for each of them.
16
Dual Control for Approximate Bayesian Reinforcement Learning
The feature set used for a specific application is part of the prior assumptions for that
application. Large uncertainty requires flexible models (which take longer to converge, and
require more exploration). Feature selection is important, but since it is independent of the
dual control framework itself and a broad topic on its own, it is beyond the scope of this
paper. In the following experiments, different feature sets are used both as examples for
the flexibility of the framework, but also to model different structural knowledge about the
problems at hand.
6.1 On a Simple Scalar System, AD Control Matches Exact Dual Control Well
For the noise-free linear system of Sec. 3.1, (a = 1 (known), b = 2, p(b) = N (b; 1, 10), Q = 10−1 ,
R = 0, W = 1, Λ = 1, T = 2), Fig. 1, right, compares the cost functions of the various
controllers and the approximately exact sampling solution (which is only available for this
very simple setup). All cost functions are shifted by an irrelevant constant. The CE cost is
quadratic and indifferent about zero. The BEB (τ = 0.1) gives additional structure near zero
that encourages learning. While qualitatively similar to the dual cost, its global minimum
is almost at the same location as that of CE. The dual control approximates the sampling
solution much closer.
6.2 Faced with Time-Varying State Cost, AD Control Holds Off Exploration
Until Suitable
A cart on a rail is a simple example for a dynamical system. Combined with a nonlinearly
varying slope, a simple but nonlinear system can be constructed. The dynamics, prior beliefs,
and true values for the parameters are chosen to be
1 0.4
0 0 ϕ1 (x1 )
0
xk+1 = [
] xk + [ 1 2 ] [ 2 k1 ] + [ ] uk
0 1
θ θ
ϕ (xk )
1
0
1 0
θ ∼ N ([ ] , [
])
0
0 1
θtrue = [
0.8
],
0.4
where superscripts denote vector elements. The nonlinear functions ϕ are shifted logistic
functions of the form
ϕ1 (x) = −
1
1 + e(x+5)
ϕ2 (x) =
1
1 + e−(x−5)
,
(19)
and disturbance/noise is chosen to be R = Q = 10−2 I. We use this setup as a testbed for
a time-structured exploration problem. The actual system and its dynamics are relatively
irrelevant here, as we will focus on a complication caused by the cost function: The reference
to be tracked is
0
r 0∶11 = [ ]
0
r 12∶14 = [
10
]
0
0
r 15 = [ ]
0
r 16∶18 = [
−10
]
0
0
r 19∶20 = [ ] ;
0
it is also shown in each plot of Fig. 5 as dashed orange line. The state weighting is
time-dependent
W 0∶5 = [
10 0
]
0 0
0 0
W 5∶10 = [
]
0 0
17
W 11∶20 = [
100 0
],
0 0
Klenske and Hennig
x1
x1
10
0
−10
x1
10
0
−10
10
0
−10
cost
10
0
−10
x1
state cost
6
4
2
0
0
2
high state cost
no state cost
4
6
8
10
12
14
16
18
20
Figure 5: Top four: Density estimate for 50 trajectories (first state). From top to bottom:
optimal oracle control (gray), certainty equivalent control (red), CE with Bayesian
exploration bonus (blue), approximate dual control (green). Reference trajectory
in dashed orange. Bottom: The mean cost per time step is shown in the bottom
plot, with colors matching the controllers noted above.
and control cost is relatively low: Λ = 10−3 . The task, thus, is to first keep the cart fixed in
the starting position to high precision, for the first 4 time steps. This is followed by a “loose”
period between time steps 5 and 10. Then, the cart has to be moved to one side, back to
the center, to the other side, and back again, all at high cost. A good exploration strategy
in this setting is to act cautiously for the first 5 time steps, then aggressively explore in the
“loose” phase, to finally be able to control the motion with high precision.
The inference model is a GP with approximated SE kernel, as described in Sec. 5.2.1.
We use 30 alternating sine and cosine features that are distributed according to the power
spectrum of the full SE kernel. Since the true nonlinearity of Eq. (19) is not of this form,
the approximation is out of model and the lower bound controller only represents a perfectly
learned, but still not exact, model.
Fig. 5 shows a density estimated from 50 state trajectories for the four different controllers.
The lower bound controller (top) controls precisely at times of high cost, and does nothing
for times with zero cost, controlling perfectly up to the measurement and state disturbances.
The certainty equivalent controller (second from top) never explores actively, it only learns
“accidentally” from observations arising during the run. Since the initial trajectory requires
little action, it is left with a bad model when the reference starts to move at time step 12.
The exploration bonus controller (second from bottom) continuously explores, because it
18
Dual Control for Approximate Bayesian Reinforcement Learning
has no way of knowing about the “loose” phase ahead. Of course, this strategy incurs a
higher cost initially. The dual controller (bottom) efficiently holds off exploration until it
reaches the “loose” phase, where it explores aggressively.
6.3 AD Control Distinguishes Necessary and Unnecessary Parameter
Exploration
The system including nonlinearities for this experiment is the same as before, although with
noise parameters R = Q = 10−3 I. The reference trajectory and state weighting are much
simpler, though:
0
10
0
r 0∶11 = [ ]
r 12∶18 = [ ]
r 18∶20 = [ ] ,
0
0
0
with the time-dependent weighting
0 0
W 0∶10 = [
]
0 0
W 11∶20 = [
10 0
],
0 0
allowing for identification in the beginning, while penalizing deviations of the first state in
later time steps.
Important to note here is that the reference trajectory only passes areas of the state
space where ϕ1 is strong, and ϕ2 is negligible. Good exploration thus will ignore θ2 , but
this can only be found through reasoning about future trajectories.
In this experiment, the learned model is of the neural network form described in Sec. 5.3.
We use 4 logistic features (see Eq. (18)) with two free parameters each (wi and vi ) and
equally spaced bi between −5 and 5, the locations of the true nonlinear features. This means
it is possible to learn the perfect model in this case.
Fig. 6 shows a density estimated from 50 state trajectories for the four different controllers.
Because of symmetry in the cost function and feature functions, BEB (with τ = 1) cannot
“decide” between the relevant θ1 and the irrelevant θ2 , choosing the exploration direction
stochastically. It thus sometimes reduces the uncertainty on θ2 , which does not help the
subsequent control. The AD controller ignores θ2 completely and only identifies θ1 in early
phases, leading to good control performance.
6.4 AD Control Maintains Only Useful Knowledge
The last experiment is again similar to Sec. 6.2, but uses a different set of nonlinear functions,
namely shifted Gaussian functions (a.k.a. radial basis functions)
ϕ1 (x) = e−
(x−2)2
2
ϕ2 (x) = e−
(x+2)2
2
1.0
θtrue = [ ] .
0.8
For this experiment, the model is learned with parametric linear regression, according to
Sec. 5.1. The fundamental difference to the other experimental setups is that the model now
assumes parameter drift. This results in growing uncertainty for the parameters over time.
(The true parameters are kept constant for simplicity.)
19
Klenske and Hennig
x1
x1
10
0
−10
x1
10
0
−10
10
0
−10
cost
10
0
−10
x1
no state cost
0.8
0.6
0.4
0.2
0
0
2
4
state cost
6
8
10
12
14
16
18
20
Figure 6: Top four: Density estimate for 50 trajectories (second state). From top to
bottom: optimal oracle control (gray), certainty equivalent control (red), CE with
Bayesian exploration bonus (blue), approximate dual control (green). Reference
trajectory in dashed orange. Bottom: The mean cost per time step is shown in
the bottom plot, with colors matching the controllers noted above.
The reference to be tracked passes through both nonlinear features but then stays at
one of them:
−5
r 0∶6 = [ ]
0
−4
r7 = [ ]
0
−2
r8 = [ ]
0
0
r9 = [ ]
0
2
r 10∶20 = [ ] .
0
The cost structure is
0 0
W 0∶5 = [
]
0 0
10 0
W 6∶20 = [
],
0 0
such that there is cost starting with the linear reference trajectory at time instant 6.
Fig. 7 shows the parameter belief and relevant state of a single run of this experiment
over time. It shows clearly that the in the beginning necessary parameter θ1 is learned early
by BEB and the AD controller, while CE learns only “accidentally”. The BEB controller
also learns the second parameter in the beginning, even though the knowledge will be lost
over time. When the trajectory reaches the zone of the second parameter, the BEB controller
tries to lower the growing uncertainty every now and then (visible by the drops in state
x1 ), incurring high cost. AD control completely ignores the growing uncertainty on θ1 after
reaching the area of θ2 , thus preventing unnecessary exploration.
20
Dual Control for Approximate Bayesian Reinforcement Learning
θ1
x1
θ2
2
2
0
0
−2
−2
2
2
0
0
−2
−2
2
2
0
0
−2
−2
0
−5
0
−5
0
0
5
10
15
20
−5
0
5
10
15
0
20
5
10
15
20
Figure 7: Parameter knowledge (left, middle) and state trajectory (right) for different
controllers. From top to bottom: certainty equivalent control (red), CE with
Bayesian exploration bonus (blue), approximate dual control (green). The true
parameters are the black lines.
Exp. 6.1
mean
std
Oracle
CE
CE-BEB
AD
0.67
0.84
0.99
0.77
0.23
0.73
0.92
0.43
Exp. 6.2
mean
std
1.75
2.49
2.64
1.96
0.33
0.74
0.37
0.34
Exp. 6.3
mean
std
Exp. 6.4
mean
std
7.15
15.72
20.88
14.33
0.66
1.76
84.91
1.62
3.85
5.20
6.74
5.40
0.51
0.90
6.77
0.56
Table 1: Average and standard deviation of costs in the experiments for 50 runs.
6.5 Quantitative Comparison
The above experiments aim to emphasize qualitative strengths of AD control over simpler
approximations. It is desirable for a controllers to deal with flexible models of many
parameters, many of which will invariably be superfluous. For reference, Table 6.5 also
shows quantitative results: Averages and standard deviations of the cost, from the 50 runs
for each controller. The AD controller shows good performance overall; interestingly, it also
has low variance. CE and BEB were more prone to instabilities.
7. Conclusion
Bayesian reinforcement learning, or dual control, offers an elegant answer to the explorationexploitation trade-off, relative to prior probabilistic beliefs. Its intricate, intractable structure
requires approximations to balance another kind of trade-off, between computation and
performance. This work investigated an old approximate framework from control, re-phrased
21
Klenske and Hennig
it in the language of reinforcement learning, and extended it to apply to contemporary
inference methods from machine learning, including approximate Gaussian process regression
and multi-layer networks. The result is a tractable approximation that captures notions of
structured exploration, like the value of waiting for future exploration opportunities, and
distinguishing relevant from irrelevant model parameters.
The dual control framework, in its now clearer form, offers interesting directions for
research in reinforcement learning, including its combination with recent new developments
in learning and planning. Following this conceptual work, the main challenge for further
development is the still comparably high (but tractable) numerical load of dual control,
particularly in problems of higher dimensionality.
22
Dual Control for Approximate Bayesian Reinforcement Learning
Appendix A. Nonparametric EKF Form
The standard Kalman filter (KF) can be found in many textbooks (e.g., Särkkä, 2013) and
therefore will not be restated here. Starting from the standard equations, we derive a general
multi-step formulation of the classic KF with
p(zk ) = N (zk , mk , Pk ).
From there, state augmentation with an infinite-dimensional weight vector gives the expected
result.
A.1 Derivation of the Multi-Step KF Formulation
Assuming that the result of the KF and the Gaussian process framework should be identical
under certain circumstances, we wish to transform the KF to a formulation with full Gram
matrix. Therefore, the prediction and update step have to be combined to
P1 = (A0 P0 A⊺0 + Q) − (A0 P0 A⊺0 + Q) H ⊺ S1−1 H (A0 P0 A⊺0 + Q)
⊺
S1 = H (A0 P0 A⊺0 + Q) H ⊺ + R,
which is pretty straightforward. We’re adopting a standard notation, where Pk is the
covariance at time step k, Ak is the Jacobian, Q is the drift and R is the measurement
covariance. The same can be done for the second time step, but it is beneficial introducing
a compact notation for the predictive covariance first
(A1 P1 A⊺1 + Q)
= (A1 [(A0 P0 A⊺0 + Q) − (A0 P0 A⊺0 + Q) H ⊺ S1−1 H (A0 P0 A⊺0 + Q) ] A⊺1 + Q)
⊺
= A1 (A0 P0 A⊺0 + Q) A⊺1 + Q − A1 (A0 P0 A⊺0 + Q) H ⊺ ( S1 )−1 H (A0 P0 A⊺0 + Q) A⊺1
¯
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
∶=g00
∶=g11
∶=g10
∶=g01
−1
= g11 − g10 H ⊺ g00
Hg01 .
Using the compact notation and defining S2 analogously to S1 , we can write the two-step
update as
P2 = (g11 − g10 H ⊺ S1−1 Hg01 ) − (g11 − g10 H ⊺ S1−1 Hg01 ) H ⊺ S2−1 H (g11 − g10 H ⊺ S1−1 Hg01 )
P2 = g11 − g10 H ⊺ S1−1 Hg01 − g11 H ⊺ S2−1 Hg11 − g10 H ⊺ S1−1 Hg01 H ⊺ S2−1 Hg10 H ⊺ S1−1 Hg01
+ g11 H ⊺ S2−1 Hg10 H ⊺ S1−1 Hg01 + g10 H ⊺ S1−1 Hg01 H ⊺ S2−1 Hg11
S −1 + S1−1 Hg01 H ⊺ S2−1 Hg10 H ⊺ S1−1 −S1−1 Hg01 H ⊺ S2−1 Hg01
P2 = g11 − [g10 H ⊺ g11 H ⊺ ] [ 1
][
].
Hg11
−S2−1 Hg10 H ⊺ S1−1
S2−1
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
∶=G−1
Application of Schur’s lemma gives
G=[
Hg00 H ⊺ + R
Hg01 H ⊺
].
⊺
Hg10 H
Hg11 H ⊺ + R
23
Klenske and Hennig
Assuming full state measurement (H = I) for compactness of notation, the two-step
update is
g +R
g01
⊺
⊺
] [ 00
g11
P2 = g11 − [g10
]
g10
g11 + R
−1
[
g01
]
g11
= A1 (A0 P0 A⊺0 + Q) A⊺1 + Q − [A1 (A0 P0 A⊺0 + Q) (A1 (A0 P0 A⊺0 + Q) A⊺1 + Q)]
(A0 P0 A⊺0 + Q) + R
(A0 P0 A⊺0 + Q) A⊺1
]
A1 (A0 P0 A⊺0 + Q) (A1 (A0 P0 A⊺0 + Q) A⊺1 + Q) + R
−1
⋅[
[
(A0 P0 A⊺0 + Q) A⊺1
],
(A1 (A0 P0 A⊺0 + Q) A⊺1 + Q)
(20)
which already looks similar to GP inference. We can now generalize this two-step result to
the general form by building the Gram matrix according to
G = FPF ⊺ + FQF ⊺ + R,
where the individual parts are
P =[
P0 0
]
0 0
Q = (Q ⊗ I)
R = (R ⊗ I)
of appropriate size, depending on the current time k. The multi-step state transition matrix
⎡ I
0
0
⎢
⎢ A1
I
0
⎢
⎢
A2
I
F = ⎢ A2 A1
⎢
⎢
⋮
⎢
⎢A ⋯A A ⋯A ⋯
⎣ m
1
m
2
⋯ 0⎤⎥
⋯ 0⎥⎥
⎥
⋯ 0⎥ ,
⎥
⋱ ⎥⎥
I ⎥⎦
with
⎡ I
0
0
⋯
⎢
⎢−A1
I
0
⋯
⎢
⎢
−A2 I
⋯
F −1 = ⎢ 0
⎢
⎢ ⋮
⋱ ⋱
⋱
⎢
⎢ 0
⋯
0
−A
⎣
m
0⎤⎥
0⎥⎥
⎥
0⎥ ,
⎥
⎥
⎥
I ⎥⎦
is also needed so shift the initial covariance and drift covariances through time. Put together,
this results in
⊺
Pk = Fk,∶ (P + Q)Fk,∶
− Fk,∶ (P + Q)F(FPF ⊺ + FQF ⊺ + R)−1 F ⊺ (P + Q)F∶,k .
A more compact notation can be achieved by using F −1 to obtain
⊺
Pk = Fk,∶ (P + Q)Fk,∶
− Fk,∶ (P + Q)(P + Q + F −1 RF −⊺ )−1 (P + Q)F∶,k .
(21)
Calculating the mean prediction is done analogously:
mk = Fk,0 m0 + Fk,∶ (P + Q)(P + Q + F −1 RF −⊺ )−1 y.
A.2 Augmenting the State
Instead of tracking only the state covariance, in the GP setting also the dynamics function
has to be inferred. The system equations of the nonlinear system are now
xk = f (xk−1 ) + qk−1
qk−1 ∼ N (0, Q)
yk = Hxk + rk
rk ∼ N (0, R),
24
Dual Control for Approximate Bayesian Reinforcement Learning
where f ∼ GP(0, k). The inference in this model can be done through the EKF with
augmented state. We adopt the weight-space view with f = Φ⊺ w (see Rasmussen and
Williams, 2006) to augment the state with the infinite-dimensional weight vector w:
x
z=( )
w
Σ=(
P
Σxw
)
Σwx Σww
∂ f¯
JA = ( ∂x
0
∂ f¯
∂w )
I
=(
A Φ
),
0 I
where, in (20), the original x is replaced by the augmented z, P by Σ and A by JA .
Choosing H = [I 0] so that H recovers the original states from the augmented state
vector, we obtain, after calculations similar to those above, a Gram matrix with additional
terms including feature functions and the prior on them:
⊺
⊺ ⊺
⊺
Φ0 Σww
Φ0 Σww
0 Φ0
0 (Φ0 A1 + Φ1 )
G⋆ = G + (
).
⊺ ⊺
ww ⊺
ww
(A1 Φ0 + Φ1 ) Σ0 Φ0 (A1 Φ0 + Φ1 ) Σ0 (Φ0 A1 + Φ⊺1 )
⊺
At this point it is important to note that the infinite inner product Φk Σww
0 Φk corresponds
to an evaluation of the kernel
⊺
ww
′
′
Φ(x)Σww
0 Φ(x) = ∑ φi (x)Σ0,ij φj (x ) = κ(x, x ).
This means we can write the Gram matrix as
G⋆ = G + (
κ00
κ00 A⊺1 + κ01
),
⊺
A1 κ00 + κ10 A1 κ00 A1 + κ10 A⊺1 + A1 κ01 + κ11
where we have written κ⋅⋅ for κ(⋅, ⋅) to save space. In total, the Gram matrix is then
G⋆ = FPF ⊺ + FQF ⊺ + FKF ⊺ + R,
with K = κ(y 1∶m , y 1∶m ). Since inference is more compact and numerically stable if we absorb
F into the Gram matrix as in Eq. (21), we define
G = P + Q + K + F −1 RF −⊺ .
Inference is done according to
⊺
Pk = Fk,∶ (P + Q + K)Fk,∶
− Fk,∶ (P + Q + K)G −1 (P + Q + K)F∶,k
xw ⊺
ww
−1
ww
Σwx
k = (Σk ) = Fk,∶ Φ(y)Σ0 − Fk,∶ (P + K + Q) G Φ(y)Σ0
ww
ww
⊺ −1
ww
Σww
k = Σ0 − Σ0 Φ(y) G Φ(y)Σ0
for the covariance and
mk = Fk,0 m0 + Fk,∶ (P + Q + K)G −1 y
⊺ −1
w̄k = w̄0 + Σww
0 Φ(y) G y
for the mean.
25
Klenske and Hennig
Appendix B. Gradients and Hessians of Dynamics Functions
B.1 Neural Network Basis Functions
The neural network dynamics function is
F
f (x) = ∑ vi σ(wi (x − bi )),
σ(a) =
i=1
1
,
1 + e−a
with the well-known derivatives of the logistic
∂2
∂
σ(a) = σ(a)(1 − σ(a)),
σ(a) = σ(a) (1 − σ(a) (3 − 2σ(a))) .
∂a
∂a2
The gradient of f (x) can easily found to be
⎡ ∑Fi=1 vi wi σ(wi (x − bi ))(1 − σ(wi (x − bi ))) ⎤
⎥
⎢
⎥
⎢
σ(w1 (x − b1 ))
⎥
⎢
⎥
⎢
⎥
⎢
⋮
⎥
⎢
⎥.
∇f (x) = ⎢⎢
σ(wF (x − bF ))
⎥
⎢ v (x − b )σ(w (x − b ))(1 − σ(w (x − b ))) ⎥
⎥
⎢ 1
1
1
1
1
1
⎥
⎢
⎥
⎢
⋮
⎥
⎢
⎢vF (x − bF )σ(wF (x − bF ))(1 − σ(wF (x − bF )))⎥
⎦
⎣
The Hessian, written in parts, using ai = wi x + bi , is:
F
∇2x f (x) = ∑ vi wi2 σ(ai )(1 − σ(ai )(3 − 2σ(ai )))
i=1
∇x ∇vi f (x) = wi σ(ai )(1 − σ(ai ))
∇x ∇wi f (x) = vi σ(ai )(1 − σ(ai )) + (x − bi )wi vi σ(ai )(1 − σ(ai ))((1 − 2σ(ai ))
∇vi ∇vi f (x) = 0
∇vi ∇wi f (x) = (x − bi )σ(ai )(1 − σ(ai ))
∇wi ∇wi f (x) = vi (x − bi )2 σ(ai )(1 − σ(ai )(3 − 2σ(ai ))).
B.2 Fourier Basis Functions
The Fourier approximation to the dynamics function has the form
√ F/2
2
f (x) =
∑ v2i−1 sin(ω2i−1 x) + v2i cos(ω2i x).
F i=1
The gradient of f (x) can easily verified to be
√
⎤
⎡ 2 F/2
⎢ F ∑i=1 v2i−1 ω2i−1 cos(ω2i−1 x) − v2i ω2i sin(ω2i x)⎥
⎥
⎢
√
⎢
⎥
2
⎥
⎢
sin(ω
x)
1
F
⎥
⎢
√
⎢
⎥
2
⎥
⎢
cos(ω
x)
2
⎢
⎥.
F
∇f (x) = ⎢
⎥
⎢
⎥
⋮
⎥
⎢
√
⎢
⎥
2
⎢
⎥
F sin(ωF −1 x)
⎢
⎥
√
⎢
⎥
2
⎢
⎥
cos(ω
x)
F
⎣
⎦
F
26
Dual Control for Approximate Bayesian Reinforcement Learning
The Hessian, written in parts, using c =
√
2/F
for normalization, is:
F/2
∇2x f (x)
2
2
= c ∑ −v2i−1 ω2i−1
sin(ω2i−1 x) − v2i ω2i
cos(ω2i x)
i=1
⎧
⎪
⎪cωi cos(ωi x)
∇x ∇vi f (x) = ⎨
⎪
⎪
⎩−cωi sin(ωi x)
∇vi ∇vi f (x) = 0.
i
i
odd
even
B.3 Radial Basis Functions
With radial basis function features, the dynamics function is
F
f (x) = ∑ vi exp (−
i=1
(x − ci )2
).
2λ2
The gradient of f (x) is
2
⎡ F
(ci −x) ⎤
i)
⎢∑i=1 vi exp (− (x−c
) 2λ
2 ⎥
2λ2
⎢
⎥
2
⎢
⎥
(x−c
)
1
⎢
⎥
exp
(−
)
2
⎢
⎥.
2λ
∇f (x) = ⎢
⎥
⎢
⎥
⋮
⎢
⎥
2
⎢
⎥
(x−cF )
exp (− 2λ2 )
⎢
⎥
⎣
⎦
The Hessian, written in parts, is:
F
∇2x f (x) = ∑ vi exp (−
i=1
∇x ∇vi f (x) = exp (−
ci − x 2 1
(x − ci )2
)
[(
) − 2]
2λ2
2λ2
λ
(x − ci )2 ci − x
)
2λ2
2λ2
∇vi ∇vi f (x) = 0
References
F. Allgöwer, T.A. Badgwell, J.S. Qin, J.B. Rawlings, and S.J. Wright. Nonlinear predictive
control and moving horizon estimationan introductory overview. In Advances in Control,
pages 391–449. Springer, 1999.
M. Aoki. Optimization of Stochastic Systems. Academic Press, New York - London, 1967.
J.-Y. Audibert, R. Munos, and C. Szepesvári. Exploration-exploitation tradeoff using variance
estimates in multi-armed bandits. Theoretical Computer Science, 410(19):1876–1902, 2009.
Y. Bar-Shalom. Stochastic dynamic programming: Caution and probing. IEEE Transactions
on Automatic Control, 26(5):1184–1195, 1981.
27
Klenske and Hennig
Y. Bar-Shalom and E. Tse. Dual effect, certainty equivalence, and separation in stochastic
control. IEEE Transactions on Automatic Control, 19(5):494–500, 1974.
Y. Bar-Shalom and E. Tse. Caution, probing, and the value of information in the control of
uncertain systems. Annals of Economic and Social Measurement, 5(3):323–337, 1976.
R.E. Bellman. Adaptive Control Processes: A Guided Tour. Princeton University Press,
1961.
D.P. Bertsekas. Dynamic Programming and Optimal Control. Athena Scientific, 3rd edition,
2005.
O. Chapelle and L. Li. An Empirical Evaluation of Thompson Sampling. In Advances in
Neural Information Processing Systems (NIPS), pages 2249–2257, 2011.
R. Dearden, N. Friedman, and D. Andre. Model based Bayesian exploration. In Uncertainty
in Artificial Intelligence (UAI), volume 15, pages 150–159, 1999.
M.P. Deisenroth and C.E. Rasmussen. PILCO: A Model-Based and Data-Efficient Approach
to Policy Search. In International Conference on Machine Learning (ICML), 2011.
M. Diehl, H.J. Ferreau, and N. Haverbeke. Efficient numerical methods for nonlinear MPC
and moving horizon estimation. In Nonlinear Model Predictive Control, volume 384, pages
391–417. Springer, 2009.
S.E. Dreyfus. Some types of optimal control of stochastic systems. Journal of the Society
for Industrial & Applied Mathematics, Series A: Control, 2(1):120–134, 1964.
M.O.G. Duff. Optimal Learning: Computational procedures for Bayes-adaptive Markov
decision processes. PhD thesis, University of Massachusetts, Amherst, 2002.
A.A. Feldbaum. Dual Control Theory I-IV. Avtomatika i Telemekhanika, 21(9), 21(11),
22(1), 22(2), 1960–1961.
N.M. Filatov and H. Unbehauen. Survey of adaptive dual control methods. IEEE Proceedings
on Control Theory and Applications, 147(1):118–128, 2000.
N.M. Filatov and H. Unbehauen. Adaptive Dual Control. Springer Verlag, Berlin, 2004.
T. Glad and L. Ljung. Control theory: Multivariable and Nonlinear Methods. Taylor and
Francis, New York, London, 2000.
P. Hennig. Optimal Reinforcement Learning for Gaussian Systems. In Advances in Neural
Information Processing Systems (NIPS), 2011.
O.L.R. Jacobs and J.W. Patchell. Caution and probing in stochastic control. International
Journal of Control, 16(1):189–199, 1972.
H.J. Kappen. Optimal control theory and the linear Bellman equation. Inference and
Learning in Dynamic Models, pages 363–387, 2011.
28
Dual Control for Approximate Bayesian Reinforcement Learning
J.Z. Kolter and A.Y. Ng. Near-Bayesian exploration in polynomial time. In International
Conference on Machine Learning (ICML), 2009.
H. König. Eigenvalue Distribution of Compact Operators. Birkhäuser, 1986.
W.G. Macready and D.H. Wolpert. Bandit problems and the exploration/exploitation
tradeoff. IEEE Transactions on Evolutionary Computation, 2(1):2–22, 1998.
D.Q. Mayne. A second-order gradient method for determining optimal trajectories of
non-linear discrete-time systems. International Journal of Control, 3(1):85–95, 1966.
D.H. Nguyen and B. Widrow. Neural networks for self-learning control systems. IEEE
Control Systems Magazine, 10(3):18–23, 1990.
P. Poupart, N. Vlassis, J. Hoey, and K. Regan. An analytic solution to discrete Bayesian
reinforcement learning. In International Conference on Machine Learning (ICML), 2006.
A. Rahimi and B. Recht. Random features for large-scale kernel machines. In Advances in
Neural Information Processing Systems (NIPS), pages 1177–1184, 2008.
H. Raiffa and R. Schlaifer. Applied statistical decision theory. Studies in managerial economics.
Harvard University, Boston, 1961.
C.E. Rasmussen and C.K.I. Williams. Gaussian Processes for Machine Learning. MIT, 2006.
H. Robbins and S. Monro. A stochastic approximation method. The Annals of Mathematical
Statistics, 22(3):400–407, Sep. 1951.
D.E. Rumelhart, G.E. Hinton, and R.J. Williams. Learning representations by backpropagating errors. Nature, 323(6088):533–536, 1986.
S. Särkkä. Bayesian filtering and smoothing. Cambridge University Press, 2013.
S. Singhal and L. Wu. Training multilayer perceptrons with the extended kalman algorithm.
In Advances in Neural Information Processing Systems (NIPS), pages 133–140, 1989.
N. Srinivas, A. Krause, S. Kakade, and M. Seeger. Gaussian Process Optimization in the
Bandit Setting: No Regret and Experimental Design. In International Conference on
Machine Learning (ICML), 2010.
M.L. Stein. Interpolation of spatial data: some theory for Kriging. Springer Verlag, 1999.
J. Sternby. A simple dual control problem with an analytical solution. IEEE Transactions
on Automatic Control, 21(6):840–844, 1976.
W.R. Thompson. On the Likelihood that one unknown probability exceeds another in view
of two samples. Biometrika, 25:275–294, 1933.
E. Tse and Y. Bar-Shalom. An actively adaptive control for linear systems with random
parameters via the dual control approach. IEEE Transactions on Automatic Control, 18
(2):109–117, 1973.
29
Klenske and Hennig
E. Tse, Y. Bar-Shalom, and L. Meier III. Wide-sense adaptive dual control for nonlinear
stochastic systems. IEEE Transactions on Automatic Control, 18(2):98–108, 1973.
J. Uhlmann. Dynamic Map Building and Localization: New Theoretical Foundations. PhD
thesis, University of Oxford, 1995.
B. Wittenmark. Adaptive dual control methods: An overview. In IFAC symposium on
Adaptive Systems in Control and Signal Processing, pages 67–72, 1995.
30
| 3cs.SY
|
Optimal Inference After Model Selection
William Fithian∗1 , Dennis L. Sun2 , and Jonathan Taylor3
arXiv:1410.2597v4 [math.ST] 18 Apr 2017
1
2
Department of Statistics, University of California Berkeley
Department of Statistics, California Polytechnic State University
3
Department of Statistics, Stanford University
April 19, 2017
Abstract
To perform inference after model selection, we propose controlling the selective type I
error; i.e., the error rate of a test given that it was performed. By doing so, we recover
long-run frequency properties among selected hypotheses analogous to those that apply in the
classical (non-adaptive) context. Our proposal is closely related to data splitting and has a
similar intuitive justification, but is more powerful. Exploiting the classical theory of Lehmann
and Scheffé (1955), we derive most powerful unbiased selective tests and confidence intervals
for inference in exponential family models after arbitrary selection procedures. For linear
regression, we derive new selective z-tests that generalize recent proposals for inference after
model selection and improve on their power, and new selective t-tests that do not require
knowledge of the error variance.
1
Introduction
A typical statistical investigation can be thought of as consisting of two stages:
1. Selection: The analyst chooses a statistical model for the data at hand, and formulates testing,
estimation, or other problems in terms of unknown aspects of that model.
2. Inference: The analyst investigates the chosen problems using the data and the selected model.
Informally, the selection stage determines what questions to ask, and the inference stage answers
those questions. Most statistical methods carry an implicit assumption that selection is nonadaptive — that is, choices about which model to use, hypothesis to test, or parameter to estimate,
are made before seeing the data. Adaptive selection (also known colloquially as “data snooping”)
violates this assumption, formally invalidating any subsequent inference.
In some cases, it is possible to specify the question prior to collecting the data—for instance,
if the data are governed by some known physical law. However, in most applications, the choice
of question is at least partially guided by the data. For example, we often perform exploratory
analyses to decide which predictors or interactions to include in a regression model or to check
whether the assumptions of a test are satisfied. The goal of this paper is to codify what it means
for inference to be valid in the presence of adaptive selection and to propose methods that achieve
this “selective validity.”
If we do not account properly for adaptive model selection, the resulting inferences can have
troubling frequency properties, as we now illustrate with an example.
∗ To
whom correspondence should be addressed
1
Example 1 (File Drawer Effect). Suppose one or more scientific research groups make n independent measurements of n quantities, Yi ∼ N (µi , 1). They focus only on the apparently large effects,
selecting (say) only the indices i for which |Yi | > 1, i.e.
Ib = {i : |Yi | > 1}.
Each scientist wishes to test H0,i : µi = 0 for his own i ∈ Ib at significance level α = 0.05.
Most practitioners intuitively recognize that the nominal test that rejects H0,i when |Yi | > 1.96 is
invalidated by the selection.
What exactly is “invalid” about this test? After all, the probability of falsely rejecting a given
H0,i is still P(|Yi | > 1.96) = 0.05, since H0,i is simply not tested at all most of the time. Rather,
the troubling feature is that the error rate among the hypotheses selected for testing is possibly
much higher than α. To be precise, let n0 be the number of true null effects and suppose n0 → ∞
as n → ∞. Then, in the long run, the fraction of errors among the true nulls we test is
X
1
b reject H0,i }
1{i ∈ I,
n0 i: H0,i true
# false rejections
=
X
# true nulls selected
1
b
1{i ∈ I}
n0 i: H0,i true
→
b reject H0,i )
PH0,i (i ∈ I,
b
PH (i ∈ I)
0,i
b
= PH0,i (reject H0,i | i ∈ I),
(1)
which for the nominal test is Φ(−1.96)/Φ(−1) ≈ .16.
Thus, we see that (1), the probability of a false rejection conditional on selection, is a natural
error criterion to control in the presence of selection. In this example, we can directly control (1)
at level α = 0.05 simply by finding the critical value c solving
PH0,i |Yi | > c |Yi | > 1 = 0.05.
In this case c = 2.41, which is more stringent than the nominal 1.96 cutoff.
This paper will develop a theory for inference after selection based on controlling the selective
type I error rate (1). Our guiding principle is:
The answer must be valid, given that the question was asked.
For all its simplicity, Example 1 can be regarded as a stylized model of science. Imagine that
each Yi represents an estimated effect size from a scientific study. However, only the large estimates
are ever published—a caricature which may not be too far from the truth, as recently demonstrated
by Franco et al. (2014). To compound the problem, there may be many reasonable methodologies
to choose from, even once the analyst has decided roughly what scientific question to address
(Gelman and Loken, 2013). Because of the resulting selection bias, the error rate among published
claims may be very high, leading even to speculation that “most published research findings are
false” (Ioannidis, 2005). Thus, selection effects may be a partial explanation for the replicability
crisis reported in the scientific community (Yong, 2012) and the popular media (Johnson, 2014).
The setting of Example 1 has been studied extensively in the literature of simultaneous and
selective inference, and several authors have proposed adjusting for selection by means of conditional inference. Zöllner and Pritchard (2007) and Zhong and Prentice (2008) construct selectionadjusted estimators and intervals for genome-wide association studies for genes that pass a fixed
2
initial significance threshold, based on a conditional Gaussian likelihood. Cohen and Sackrowitz
(1989) obtain unbiased estimates for the mean of the population whose sample mean is largest by
conditioning on the ordering of the observed sample means, and Sampson and Sill (2005) and Sill
and Sampson (2009) apply the same idea to obtain estimates for the best-performing drug in an
adaptive clinical trial design. Hedges (1984) and Hedges (1992) propose methods to adjust for the
file drawer effect in meta-analysis when scientists only publish significant results.
Another framework for selection adjustment is proposed by Benjamini and Yekutieli (2005),
who consider the problem of constructing intervals for a number R of parameters selected after
viewing the data. Letting V denote the number of non-covering intervals among those constructed,
they define the false coverage-statement rate (FCR) as the expected fraction V / max(R, 1) of noncovering intervals. Controlling the FCR at level α thus amounts to “coverage on the average,
among selected intervals.” As we will see further in Section 8, FCR control is closely related to
the selective error control criterion we propose. In fact, Weinstein et al. (2013) employ conditional
inference to construct FCR-controlling intervals in the context of Example 1. Rosenblatt and
Benjamini (2014) propose a similar method for finding correlated regions of the brain, also with a
view toward FCR control.
1.1
Conditioning on Selection
In classical statistical inference, the notion of “inference after selection” does not exist. The analyst
must specify the model, as well as the hypothesis to be tested, in advance of looking at the data.
A classical level-α test for a hypothesis H0 under model M must control the usual or nominal type
I error rate:
PM,H0 (reject H0 ) ≤ α.
(2)
The subscript in (2) reminds us that the probability is computed under the assumption that the
data Y are generated from model M , and H0 is true; if M is misspecified, there are no guarantees
on the rejection probability.
In most statistical practice, it is unrealistic to rule out model selection altogether: statisticians
are trained to check their models and to tweak them if they diagnose a problem (to a purist, even
model checking is suspect, since it leaves open the possibility that the model will change after we
see the data). We will argue that if the model and hypothesis are selected adaptively, we should
instead control the selective type I error rate
PM,H0 (reject H0 | (M, H0 ) selected) ≤ α.
(3)
One can argue that models and hypotheses are practically never truly fixed but are chosen
randomly, since they are based on the outcomes of previous experiments in the (random) scientific
process. Typically, we ignore the random selection and use classical tests that control (2), implicitly
assuming that the randomness in selecting M and H0 is independent of the data used for inference.
In that case,
PM,H0 (reject H0 | (M, H0 ) selected) = PM,H0 (reject H0 ).
(4)
While it may seem pedantic to point out that model selection is random if based on previous
experiments, this viewpoint justifies a common prescription for what to do when previous experiments do not dictate a model. If it is possible to split the data Y = (Y1 , Y2 ) with Y1 independent of
Y2 , then we can imitate the scientific process by setting aside Y1 for selection and Y2 for inference.
If selection depends on Y1 only, then any nominal level-α test based on the value of Y2 will satisfy
(4), so the nominal test based on Y2 also controls the selective error (3).
This meta-algorithm for generating selective procedures from nominal ones is called data splitting or sample splitting. The idea dates back at least as far as Cox (1975), and, despite the paucity
of literature on the topic, is common wisdom among practitioners. For example, it is customary in
3
genetics to use one cohort to identify loci of interest and a separate cohort to confirm them (Sladek
et al., 2007). Wasserman and Roeder (2009) and Meinshausen et al. (2009) discuss data-splitting
approaches to high-dimensional inference.
Data splitting owes much of its popularity to its transparent justification, which even a nonexpert can appreciate: if we imagine that Y1 is observed “first,” then we can proceed to analyze Y2
as though model selection took place “ahead of time.” Equation (4) guarantees that this temporal
metaphor will not lead us astray even if it does not describe how Y1 and Y2 were actually collected.
Data splitting elegantly solves the problem of controlling selective error, but at a cost. It
not only reduces the amount of data available for inference, but also reduces the amount of data
available for selection. Furthermore, it is not always possible to split the data into independent
parts, as in the case of autocorrelated spatial and time series data.
In this article, we propose directly controlling the selective error rate (3) by conditioning on the
event that (M, H0 ) is selected. As with data splitting, we treat the data as though it were revealed
in stages: in the first stage, we “observe” just enough data to resolve the decision of whether to
test (M, H0 ), after which we can treat the data (Y | (M, H0 ) selected) as “not yet observed” when
the second stage commences.
The intuition of the above paragraph can be expressed formally in terms of the filtration
F0
⊆
F (1A (Y ))
used for selection
⊆
F (Y ),
(5)
used for inference
where F (Z) denotes the σ-algebra generated by a random variable Z (informally, everything
we know about the data after observing Z), F0 is the trivial σ-algebra (representing complete
ignorance), and A is the selection event {(M, H0 ) selected}. We can think of “time” as progressing
from left to right in (5). In stage one, we learn just enough to decide whether to test (M, H0 ), and
no more, advancing our state of knowledge from F0 to F (1A (Y )). We then begin stage two, in
which we discover the actual value of Y , advancing our knowledge to F (Y ). Because our selection
decision is made at the end of stage one, everything revealed during stage two is fair game for
inference.
In effect, controlling the type I error conditional on A prevents us from appealing to the fact
that Y ∈ A as evidence against H0 . Even if Y ∈ A is extremely surprising under H0 , we still
will not reject unless we are surprised anew in the second stage. In this sense, conditioning on a
random variable discards the information it carries about any parameter or hypothesis of interest.
In contrast to data splitting, which can be viewed as conditioning on Y1 instead of 1A (Y1 ), we
advocate discarding as little information as possible and reserving the rest for stage two. This
frugality results in a more efficient division of the information carried by Y , which we call data
carving.
1.2
Outline
In Section 2 we formalize the problem of selective inference, discuss general properties of selective
error control, and address key conceptual questions. Conditioning on the selection event effectively
discards the information used for selection, but some information is left over for second-stage
inference. We will also see that a major advantage of selective error control is that it allows us to
consider only one model at a time when designing tests and intervals, even if a priori there are
many models under consideration.
If L(Y ), the law of random variable Y , follows an exponential family model, then for any event
A, L(Y | A) follows a closely related exponential family model. As a result, selective inference
dovetails naturally with the classical optimality theory of Lehmann and Scheffé (1955); Section 3
briefly reviews this theory and derives most powerful unbiased selective tests in arbitrary exponential family models after arbitrary model selection procedures. Because conditioning on more data
4
than is necessary saps the power of second-stage tests, data splitting yields inadmissible selective
tests under general conditions.
Section 5 gives some general strategies for computing rejection cutoffs for the tests prescribed
in Section 3, while Sections 4–6 derive selective tests in specific examples. Section 4 focuses on
the case of linear regression, generalizing the recent proposals of Tibshirani et al. (2014), Lee et al.
(2016), and others. We derive new, more powerful selective z-tests, as well as selective t-tests that
do not require knowledge of the error variance σ 2 .
Several simulations in Section 7 compare the post-lasso selective z-test with data splitting, and
illustrate a selection–inference tradeoff, between using more data in the initial stage and reserving
more information for the second stage. Section 8 compares and contrasts selective inference with
multiple inference, and Section 9 concludes.
2
2.1
The Problem of Selective Inference
Example: Regression and the Lasso
In the previous section, we motivated the idea of conditioning on selection. Arguably, the most
familiar example of this “selection” is variable selection in linear regression. In regression, the
observed data Y ∈ Rn is assumed to be generated from a multivariate normal distribution
Y ∼ Nn (µ, σ 2 In ).
(6)
The goal is to model the mean µ as a linear function of predictors Xj , j = 1, . . . , p. To obtain
a more parsimonious model (or simply an identifiable model when p > n), researchers will often
use only a subset M ⊆ {1, . . . , p} of the predictors. Each subset M leads to a different statistical
model corresponding to the assumption µ = XM β M , where XM denotes the matrix consisting of
M
columns Xj for j ∈ M . Then, it is customary to report tests of H0,j
: βjM = 0 for each coefficient
in the model. If M was chosen in a data-dependent way, then to control selective error we must
M
condition on having selected (M, H0,j
), which in this case is the same as conditioning on having
selected model M .
There are many data-driven methods for variable selection in linear regression, ranging from
AIC minimization to forward stepwise selection, cf. Hastie et al. (2009). We will consider one
procedure in particular, based on the lasso, mostly because selective inference in the context of the
lasso (Lee et al., 2016) was a main motivation for the present work. The lasso (Tibshirani, 1996)
provides an estimate of β ∈ Rp that solves
β̂ = argmin ||Y − Xβ||22 + λ||β||1 ,
(7)
β
where X is the “full” matrix consisting of all p predictors. The first term is the usual least-squares
objective, while the second term encourages many of the coefficients to be exactly zero. Because of
this property, it makes sense to define the model “selected” by the lasso to be the set of variables
with non-zero coefficients, i.e.,
c(Y ) = {j : β̂j 6= 0}.
M
c(Y ) can take on up to 2p possible values, one for each subset of {1, ..., p}. The
Notice that M
c(y) = M } form a partition of Rn into regions that correspond to each model.
regions AM = {y : M
To control the selective error after selecting a particular M , we must condition on the event that Y
landed in AM . The partition for a lasso problem with p = 3 variables in n = 2 dimensions is shown
in Figure 1. An explicit characterization of the lasso partition can be found in Lee et al. (2016); see
also Harris (2014) for an interactive visualization of the way the lasso partitions the sample space.
A different selection procedure would partition the sample space differently; characterizations of
5
Figure 1: An example of the lasso with n = 2 observations and p = 3 variables.Tests are based on
the distribution of Y , conditional on its landing in the highlighted region.
the partitions in forward stepwise selection and marginal screening can be found in Loftus and
Taylor (2014) and Lee and Taylor (2014), respectively.
c(Y ), but
Imagine that in stage one, we loaded the data into a software package and computed M
we remain otherwise ignorant of the value Y — that is, we have observed which of the regions Y
falls into but not where Y is in that region. Now that we have chosen the model, we will construct
M
tests of H0,j
: βjM = 0 for each of the selected variables. In the example shown in Figure 1, we
selected variables 1 and 3 and thus test the two hypotheses
{1,3}
H0,1
{1,3}
H0,3
{1,3}
: β1
:
{1,3}
β3
=0
= 0.
Notice that we have to be careful to always specify the model along with the coefficient, since
the coefficient for variable j does not necessarily have a consistent interpretation across different
models. Each regression coefficient summarizes the effect of that variable, adjusting for the other
variables in the model. For example, “What is the effect of IQ on salary?” is a genuinely different
question from “What is the effect of IQ on salary, after adjusting for years of education?” Both
questions are meaningful, but they are fundamentally different.1
Having chosen the model M and conditioned on the selection, we will base our tests on the
precise location of Y , which we do not know yet. Conditionally, Y is not Gaussian, but it does
follow an exponential family. As a result, we can appeal to the classical theory of Lehmann and
Scheffé (1955) to construct tests or confidence intervals for its natural parameters, which are β M
if σ 2 is known, and otherwise are (β M /σ 2 , 1/σ 2 ).
With this concrete example in mind, we will now develop a general framework of selective
inference that is much more broadly applicable. Because we are explicitly allowing models and
hypotheses to be random, it is necessary to carefully define our inferential goals. We first discuss
selective inference in the context of hypothesis testing. The closely related developments for
confidence intervals will follow in Section 2.4.
1 We use the word “effect” here informally to refer to a regression coefficient, recognizing that regression cannot
establish causal claims on its own.
6
2.2
Selective Hypothesis Tests
We now introduce notation that we will use for the remainder of the article. Assume that our
data Y lies in some measurable space (Y, F ), with unknown sampling distribution Y ∼ F . The
analyst’s task is to pose a reasonable probability model M — i.e., a family of distributions which
she believes contains F — and then carry out inference based on the observation Y .
Let Q denote the question space of inference problems q we might tackle. A hypothesis testing
problem is a pair q = (M, H0 ) of a model M and null hypothesis H0 , by which we mean a
submodel H0 ⊆ M .2 We write M (q) and H0 (q) for the model and hypothesis corresponding
to q. Without loss of generality, we assume H0 (q) is tested against the alternative hypothesis
H1 (q) = M (q) \ H0 (q). To avoid measurability issues, we will assume throughout that Q is
countable, although our framework can be extended to uncountable Q with additional care.
In Section 2.1 where we test each variable in a selected regression model, the question space is
M
Q = {(M, H0,j
) : M ⊆ {1, . . . , p}, j ∈ M }.
Note our slight abuse of notation in using M interchangeably
to refer both to a subset of variable
indices and to the corresponding probability model Nn (XM βM , σ 2 In ) : β ∈ R|M | .
We model selective inference as a process with two distinct stages:
b )⊆Q
1. Selection: From the collection Q of possible questions, the analyst selects a subset Q(Y
to test, based on the data.
2. Inference: The analyst performs a hypothesis test of H0 (q) against M (q) \ H0 (q) for each
b ).
q ∈ Q(Y
In the case of the simple regression example shown in Figure 1, where we selected variables 1
b would consist of the hypotheses for each of the two variables in the model:
and 3, Q
n
o
{1,3}
{1,3}
b )=
Q(Y
{1, 3}, H0,1
, {1, 3}, H0,3
.
A correctly specified model M is one that contains the true sampling distribution F . Importantly, we expressly do not assume that all — or any — of the candidate models are correctly
specified. Because the analyst must choose M without knowing F , she could choose poorly, in
which case there may be no formal guarantees on the behavior of the test she performs in stage
two. Some degree of misspecification is the rule rather than the exception in most real statistical
applications, whether models are specified adaptively or non-adaptively. Our analyst would be in
the same position if she were to select a (probably wrong) model using Y , then use that model
to perform a test on new data Y ∗ collected in a confirmatory experiment. See Section 2.6.2 for
further discussion of this issue.
For our purposes, a hypothesis test is a function φ(y) taking values in [0, 1], representing the
probability of rejecting H0 if Y = y. In most cases, the value of the function will be either 0 or 1,
but with discrete variables, randomization may be necessary to achieve exact level α.
To adjust for selection in testing q, we condition on the event that the question was asked,
which we describe by the selection event
b )},
Aq = {q ∈ Q(Y
(8)
i.e., the event that q is among the questions asked. In general, the selection events for different
M
questions are not disjoint. In the regression example, where we test H0,j
if and only if model M
2 We identify a “null hypothesis” like H
0 : µ(F ) = 0 with the corresponding subfamily or “null model”
{F ∈ M : µ(F ) = 0}. This should remind us that the error guarantees of a test do not necessarily extend beyond
the model it was designed for.
7
c. By convention we take
is selected, conditioning on Aq is equivalent to simply conditioning on M
φq (Y ) = 0 for Y ∈
/ Aq to reflect the idea that if a hypothesis is not tested then it is not rejected;
note this convention does not affect the selective properties of φq .
In selective inference, we are mainly interested in the properties of a test φq for a question q,
conditional on Aq . We say that φq controls selective type I error at level α if
EF [φq (Y ) | Aq ] ≤ α,
for all F ∈ H0 (q).
(9)
and define its selective power function as
Powφq (F | Aq ) = EF [φq (Y ) | Aq ].
(10)
Because Q is countable, the only relevant q are those for which P(Aq ) > 0.
Notice that only the model M (q) and hypothesis H0 (q) are relevant for defining the selective
level and power of a test φq . This means that in designing valid φq , we can concentrate on one q
at a time, even if there are many mutually incompatible candidate models in Q. As long as each
φq controls the selective error at level α given its selection event Aq , then a global error is also
controlled:
E [# false rejections]
≤ α,
(11)
E [# true nulls selected]
provided that the denominator is finite. Equation (11) holds for countable Q regardless of the
dependence structure across different q. The fact that we can design tests one q at a time makes
it much easier to devise selective tests in concrete examples, which we take up in Sections 3–6.
2.3
Comparison to Familywise Error Rate
Selective error control is neither weaker nor stronger than control of the familywise error rate
(FWER), which is the probability of rejecting any true null hypothesis:
FWER = PF (φq (Y ) = 1 for any q with H0 (q) 3 F ).
(12)
Although the FWER is usually considered the most conservative control guarantee, it does not
scale easily across different researchers: suppose that in Example 1, each observation Yi is collected
by a different scientific research team at a different university, with each team then publishing the
nominal level-α test if |Yi | > 1 and otherwise moving on to another project. At the level of a
single research group and experiment, FWER is controlled, but there is a major unaccounted-for
multiplicity problem if we consider the discipline as a whole.
By contrast, selective error control scales naturally across multiple research groups and requires
no coordination among groups. If each research team in a discipline controls the selective error
rate for each of its own experiments, then the discipline as a whole will achieve long-run control
of the type I error rate among true selected null hypotheses, just as they would if there were no
selection.
Proposition 1 (Discipline-Wide Error Control). Suppose there are n independently operating
research groups in a scientific discipline with a shared, countable question space Q. Research group
bi (Yi ) ⊆ Q, and carries out selective level-α tests
i collects data Yi ∼ Fi , applies selection rule Q
b
(φq,i (yi ), q ∈ Qi ). Assume each research group has probability at least δ > 0 of carrying out at
least one test of a true null, and for some common B < ∞,
h
i
bi (Yi )|2 ≤ B,
EFi |Q
for all i.
Then as n grows, the discipline as a whole achieves long-run control over the frequentist error rate
lim sup
n→∞
# false rejections a.s.
≤ α.
# true nulls selected
8
(13)
{Sq = s}
Y
Aq
Figure 2: Instead of conditioning on the selection event Aq that question q is asked, we can
condition on a finer event, the value of the random variable Sq . We call Sq the selection variable.
The proof is deferred to Appendix A.
There is no counterpart to Proposition 1 for other popular error rates such as the false discovery
rate (FDR) (Benjamini and Hochberg, 1995) or familywise error rate (FWER). Section 8 discusses
further the relationship between selective error control and other common error rates in multiple
inference.
2.4
Selective Confidence Intervals
If the goal is instead to form confidence intervals for a parameter θ(F ), it is more convenient to
think of Q as containing pairs q = (M, θ(·)) of a model and a parameter. By analogy to (9), we
will call a set C(Y ) a (1 − α) selective confidence set if
PF (θ(F ) ∈ C(Y ) | Aq ) ≥ 1 − α,
for all F ∈ M.
(14)
The next result establishes that selective confidence sets can be obtained by inverting selective
tests, as one would expect by analogy to the classical case.
Proposition 2 (Duality of Selective Tests and Confidence Sets). Suppose we form a confidence
interval for θ(F ) on the event Aq . Suppose also that on this event, we form a test φt of H0,t =
{F : θ(F ) = t} for all t. Let C(Y ) be the set of t for which φt does not (always) reject:
C(Y ) = {t : φt (Y ) < 1} .
(15)
If each φt is a selective level-α test, then C(Y ) is a selective (1 − α) confidence set.
Proof. The selective non-coverage probability is
PF (θ(F ) ∈
/ C(Y ) | Aq ) = PF (φθ(F ) (Y ) = 1 | Aq ) ≤ EF φθ(F ) (Y ) | Aq ≤ α.
2.5
Conditioning Discards Information
Because performing inference conditional on a random variable effectively disqualifies that variable
as evidence against a hypothesis, we will typically want to condition on as little data as possible in
stage two. Even so, some selective inference procedures condition on more than Aq . For example,
data splitting can be viewed as inference conditional on Y1 , the part of the data used for selection.
More generally, we say a selection variable is any variable Sq (Y ) whose level sets partition the
sample space more finely than Aq does; i.e., Aq ∈ F (Sq ). Informally, we can think of conditioning
on a finer partition of Aq , as shown in Figure 2.
9
We say φ controls the selective type I error with respect to Sq at level α if the error rate is less
than α given Sq = s for {Sq = s} ⊆ Aq . More formally,
a.s.
EF φ(Y )1Aq (Y ) | Sq ≤ α,
for all F ∈ H0 (q)
(16)
Taking Sq (y) = 1Aq (y), the coarsest possible selection variable, recovers the baseline selective type
I error in (9). The definition of a selective confidence set may be generalized in the same way.
Generalizing (5) to finer selection variables gives
F0
⊆
F (S(Y ))
used for selection
F (Y ),
⊆
(17)
used for inference
suggesting that the more we refine S(Y ), the less data we have left for second-stage inference.
Indeed, the finer S is, the more stringent is the requirement (16):
Proposition 3 (Monotonicity of Selective Error). Suppose F (S1 ) ⊆ F (S2 ). If φ controls the type
I error rate at level α for q = (M, H0 ) w.r.t. the finer selection variable S2 , then it also controls
the type I error rate at level α w.r.t. the coarser S1 .
Proof. If F ∈ H0 , then
EF [φ(Y )1A (Y ) | S1 ] = EF
EF φ(Y )1A (Y ) | S2 | S1
a.s.
≤
α.
Because S(y) = 1A (y) is the coarsest possible choice, a test controlling the type I error w.r.t.
any other selection variable also controls the selective error in (9). At the other extreme, if S(y) = y,
then we cannot improve on the trivial “coin-flip” test φ(y) ≡ α. Proposition 3 suggests that we
will typically sacrifice power as we move from coarser to finer selection variables. Even so, refining
the selection variable can be useful for computational reasons. For example, in the case of the
lasso, by conditioning additionally on the signs of the nonzero β̂j , the selection event becomes a
c
convex region instead of the union of up to 2|M | disjoint convex regions (Lee et al., 2016). Another
valid reason to refine Sq beyond 1Aq is to strengthen our inferential guarantees in a meaningful
way; for example, we can achieve achieve false coverage-statement rate (FCR) control by choosing
b )|) (see Section 8, Proposition 11).
Sq = (1Aq (Y ), |Q(Y
Data splitting corresponds to setting every selection variable equal to S = Y1 . As a result,
data splitting does not use all the information that remains after conditioning on A, as we see
informally in the three-stage filtration
F0
F (1A (Y1 ))
⊆
used for selection
⊆
wasted
F (Y1 )
⊆
F (Y1 , Y2 ).
(18)
used for inference
As we will see in Section 3.2, this waste of information means that data splitting is inadmissible
under fairly general conditions.
We can quantify the amount of leftover information in terms of the Fisher information that
remains in the conditional law of Y given S. In a smooth parametric model, we can decompose
the Hessian of the log-likelihood as
∇2 `(θ; Y ) = ∇2 `(θ; S) + ∇2 `(θ; Y | S)
(19)
The conditional expectation
IY
| S (θ; S)
= −E ∇2 `(θ; Y | S) | S
10
(20)
Selective Confidence Interval
10
5
0
µ
0.5
Selective CI
Nominal CI
−5
0.0
Information
1.0
Leftover Fisher Information
−2
0
2
4
6
8
0
µ
5
10
15
Observed Y
(a) Leftover Fisher information as a function of µ.
For µ 3, then there is very little information in
the conditional distribution, since Y is conditionally highly concentrated on 3. For µ 3, then
Pµ (A) ≈ 1 and virtually no information is lost.
(b) Confidence intervals from inverting the UMPU
tests of Section 3. For Y 3, the interval essentially coincides with the nominal interval Y ± 1.96.
For Y close to 3, the wide interval reflects potentially severe selection bias.
Figure 3: Univariate Gaussian. Y ∼ N (µ, 1) with selection event A = {Y > 3}.
is the leftover Fisher information after selection at S(Y ) (the leftover information is essentially
the same as the missing information of Orchard et al. (1972), but we find “leftover” to be a more
intuitive descriptor than missing in this context since the information is at our disposal).
Taking expectations in (19), we obtain
E IY | S (θ; S) = IY (θ) − IS (θ) IY (θ).
(21)
Thus, on average, the price of conditioning on S — the price of selection — is the information S
carries about θ.3 In some cases this loss may be quite small, which a simple example elucidates.
Example 2. Consider selective inference under the univariate Gaussian model
Y ∼ N (µ, 1),
(22)
after conditioning on the selection event A = {Y > 3}.
Figure 3a plots the leftover information as a function of µ. If µ 3, there is very little
information in the conditional distribution: whether µ = −10 or µ = −11, Y is conditionally
highly concentrated on 3. By contrast, if µ 3, then Pµ (A) ≈ 1, the conditional law is practically
no different from the marginal law, and virtually no information is lost in the conditioning.
Figure 3b shows the confidence intervals that result from inverting the tests described in Section 3. When Y 3, the interval essentially coincides with the nominal interval Y ± 1.96 because
there is hardly any selection bias and no real adjustment is necessary. By contrast, when Y is
close to 3 it is potentially subject to severe selection bias. This fact is reflected by the confidence
interval, which is both longer than the nominal interval and centered at a value significantly less
than Y .
3 Note
that we do not necessarily have IY | S (θ; S) IY (θ) for every S. In fact there are interesting counterexamples where IY | A (θ) IY (θ) for certain θ, but we will not take them up here.
11
2.6
Conceptual Questions
We now pause to address conceptual objections we have encountered when explaining our work.
These objections can be most easily expressed, and answered, in the setting where there is a single
b
selected model and a single selected hypothesis to test or confidence interval to construct (i.e., Q
is always a singleton).
There is a common theme in every one of the conceptual objections to follow: they are all
equally good grounds for objecting to data splitting, or for that matter, to selecting a model and
hypothesis based on a prior experiment whose outcome was random. Thus, a good exercise is to
ask ourselves how we would answer the same question if it were asked about data splitting; most
likely, the same answer applies equally well to data carving.
2.6.1
How can the model be random?
In our framework, inference is based on a statistical model M that is allowed to be chosen randomly,
based on the data Y . A common first reaction is that if the data are generated according to the
model, and the model is selected based on the data, then the whole business is circular and
nonsensical.
To resolve this conundrum, note that in our framework the true sampling distribution F is not
selected in any sense; it is entirely outside the analyst’s control. The only thing selected is the
working model, a tool the analyst uses to carry out inference, which may or may not include the
true F . Thus, the sampling distribution F comes first, then the data Y , then the model M .
Random selection of models is not new and should not trouble or confuse us: M would be
just as random if it were selected via data splitting, or for that matter if it were based on a prior
experiment. Thoughtful skeptics may find reasons for concern about all of these approaches, believing that statistical testing is only appropriate when a model can be based purely on convincing
theoretical considerations. We answer only that this point of view would rule out most scientific
inquiries for which statistics is ever used. However, for those who are comfortable with choosing
a random model using data splitting or a previous experiment, we see no special reason to be any
more concerned about choosing a random model using data carving.
In any case, M is by no means required to be random, and our conditional-inference framework applies in many interesting settings where M is always the same pre-specified parametric or
nonparametric model, but we adaptively choose which hypotheses to test or which parameters to
estimate. For example, in our clinical trial example of Section 6.1, the statistical model is always
the same but we choose which null hypotheses to test after inspecting the data. The same is true of
the conditional confidence intervals of Weinstein et al. (2013), the saturated-model selective z-test
proposed by Lee et al. (2016) and discussed in 4.2, and the rank verification methods proposed in
Hung and Fithian (2016).
2.6.2
What if the selected model is wrong?
If we were writing about a topic other than selective inference, we might have begun by stating
a formal mathematical assumption that the sampling distribution F belongs to a known model
M , and then devised a test φ that behaves well when F ∈ M . The same φ might not work well
at all for F ∈
/ M : for example, if we choose to apply the one-sample t-test of µ = 0 to a sample
Y1 , . . . , Yn whose observations are highly correlated, then the probability of rejection may be a
great deal larger than the nominal α, even if E[Yi ] = 0. This is not a mistake in the formal theory,
nor does it make the t-test an inherently invalid test; rather, the validity or invalidity of a test is
defined with respect to its behavior when F ∈ H0 ⊆ M .
In any given application, the analyst must choose from among many statistical methods knowing that each one is designed to work under a particular set of parametric or nonparametric
12
assumptions about F — i.e., under a particular model M . Because our theory encompasses both
the choice and the subsequent analysis, it would not be sensible to assume that the analyst is
infallible and always selects a correct model. Typically some candidate models M are correctly
specified, others are not, and the analyst can never know for sure which are which. Any model
selection procedure using data splitting, data carving, or a prior experiment always carries a risk
of selecting the wrong model, and in all cases the second-stage type I error guarantees are only in
force when the model is correct.
Of course, the possibility of misspecification is not restricted to adaptive procedures like data
carving and data splitting: selecting an inappropriate model after seeing the data leaves us no
better or worse off than if we had chosen the same inappropriate model before seeing the data.
The alternative to adaptive model selection is not infallible model selection, it is non-adaptive
model selection.
There is a separate question of robustness: if F ∈
/ M but is “close” in some sense, we may
still want our procedure to behave predictably. However, even if some model gives a reasonable
approximation to L(Y ), there is no guarantee that the induced model for L(Y | A) is reasonable,
since conditioning can introduce new robustness problems. For example, suppose that a test
statistic Zn (Y ) tends in distribution to N (0, 1) under H0 as n → ∞. In a non-selective setting,
we might be comfortable modeling it as Gaussian as a basis for hypothesis testing. In this case
it is also true that L(Zn | Zn > c) converges to a truncated Gaussian law for any fixed c ∈ R, but
the approximation may be much poorer for intermediate values of n. Worse, if we use increasing
thresholds cn → ∞ with n, the truncated Gaussian approximation may never become reasonable.
Understanding the interaction between selective inference and asymptotic approximations is an
area of active ongoing study; see Tian and Taylor (2017); Tibshirani et al. (2015); Tian and
Taylor (2015); Taylor and Tibshirani (2016) for subsequent works discussing asymptotics without
Gaussian assumptions.
2.6.3
Does the result have a marginal interpretation?
Consider an adaptive clinical trial in which we select the most promising subgroup of patients based
on some preliminary analysis, and then report a confidence interval for the average treatment effect
on that subgroup. For some realizations of the data, we might decide to return an interval for
the effect on men over the age of 45, and for other realizations we might decide to return an
interval for the effect on Hispanic women with high blood pressure. Let S(Y ) denote the selected
subpopulation, a random region of covariate space, let θ(s) denote the true average treatment
effect on a given subpopulation s, and let Cs (Y ) denote the confidence interval we construct for
θ(s) when S(Y ) = s.
We find that confusion often occurs when people attempt to interpret C(Y ) = CS(Y ) (Y ) as
a marginal confidence interval for θ(S(Y )), the treatment effect on a random subpopulation. If
we ran the experiment again with the same selection procedure, we might choose a completely
different S(Y ), giving C(Y ) a completely different meaning, and 100 realizations of the data might
produce 100 disjoint realizations of C(Y ), meant to cover 100 very different true parameter values.
While technically correct, the above interpretation is usually best avoided. Rather, we recommend thinking of each Cs as a different confidence interval for a different fixed parameter θ(s),
having nothing to do with Cs0 for s0 6= s. During the selection stage, we choose one Cs to construct
and leave the other intervals undefined.
In other words, the interval has no useful interpretation until the first stage is complete, after
which S(Y ) is fixed: it is pointless to try to interpret an answer before we even decide what
question to ask. It is true that we might have asked about a different parameter if the data had
looked different. By the same token, we might have performed an entirely different experiment if
our most recent grant application had been funded. Neither of these contingencies should be a
13
source of confusion because experiments not performed, or parameters not selected, are irrelevant
to the situation at hand.
2.7
Prior Work on Selective Inference
This article takes its main inspiration from a recent ferment of work on the problem of inference
in linear regression models after model selection. Lockhart et al. (2014) derive an asymptotic test
for whether the nonzero fitted coefficients at a given knot in the lasso path contain all of the true
nonzero coefficients. Tibshirani et al. (2014) provided an exact (finite-sample) version of this result
and extended it to the LARS path, while Lee et al. (2016), Loftus and Taylor (2014), and Lee
and Taylor (2014) used similar approaches to derive exact tests for the lasso with a fixed value of
regularization parameter λ, forward stepwise regression, and regression after marginal screening,
respectively. All of the above approaches are derived assuming that the error variance σ 2 is known
or an independent estimate is available.
The present work attempts to unify the above approaches under a common theoretical framework generalizing the classical optimality theory of Lehmann and Scheffé (1955), and elucidate
previously unexplored questions of power. It also lets us generalize the results to the case of unknown σ 2 , and to arbitrary exponential families after arbitrary selection events. Since the initial
appearance of this work, it has been applied in many other settings; see Taylor and Tibshirani
(2015) for a recent review.
Other works have viewed selective inference as a multiple inference problem. Recent work in
this vein can be found in Berk et al. (2013) and Barber and Candès (2015). Section 8 argues that
inference after model selection and multiple inference are distinct problems with different scientific
goals; see Benjamini (2010) for more discussion of this distinction. An empirical Bayes approach
for selection-adjusted estimation can be found in Efron (2011).
There has also recently been work on inference in high-dimensional linear regression models,
notably Belloni et al. (2011), Belloni et al. (2014), Zhang and Zhang (2014), Javanmard and
Montanari (2014), and Van de Geer et al. (2014); see Dezeure et al. (2015) for a review. These
works focus on approximate asymptotic inference for a fixed model with many variables, while we
consider finite-sample inference after selecting a smaller submodel to focus our inferential goals.
Leeb and Pötscher (2005, 2006, 2008) prove certain impossibility results regarding estimating
the distribution of post-selection estimators. These results do not apply to our framework; under
the statistical models we use, the post-selection distributions of our test statistics are known and
thus do not require estimation.
The foregoing works are frequentist, as is this work. Because Bayesian inference conditions on
the entire data set, conditioning first on a selection event typically has no operative effect on the
posterior: if p and π are respectively the marginal likelihood and prior, then p(Y | A, θ) · π(θ | A) ∝
p(Y | θ) · π(θ) for Y ∈ A (Dawid, 1994). Yekutieli (2012) argues that in certain cases it is more
appropriate to condition the likelihood on selection without changing the prior to reflect that
conditioning, resulting in a posterior proportional to p(Y | A, θ) · π(θ). The credible intervals
discussed in Yekutieli (2012) resemble the confidence intervals proposed in this article, and the
discussion therein presents a somewhat different perspective on how and why conditioning can
adjust for selection.
Though our goals are very different, our theoretical framework is in some respects similar to the
conditional confidence framework of Kiefer (1976), in which inference is made conditional on some
estimate of the confidence with which a decision can be made. See also Kiefer (1977); Brownie and
Kiefer (1977); Brown (1978); Berger et al. (1994).
Olshen (1973) discussed error control given selection in a two-stage multiple comparison procedure, in which an F -test is first performed, then Scheffé’s S-method applied if the F -test rejects.
For large enough rejection thresholds, simultaneous coverage in the second stage is less than 1 − α
14
conditional on rejection in stage one.
3
Selective Inference in Exponential Families
As discussed in Section 2.2, we can construct selective tests “one at a time” for each model–
hypothesis pair (M, H0 ), conditional on the corresponding selection event Aq and ignoring any
other models that were previously under consideration. This is because the other candidate models
and hypotheses are irrelevant to satisfying (9). For that reason, we suppress the explicit dependence
on q = (M, H0 ) except where it is necessary to resolve ambiguity.
Our framework for selective inference is especially convenient when M corresponds to a multiparameter exponential family
Y ∼ fθ (y) = exp{θ0 T (y) − ψ(θ)} f0 (y)
(23)
with respect to some dominating measure. Then, the conditional distribution given Y ∈ A for
any measurable A is another exponential family with the same natural parameters and sufficient
statistics but different carrier measure and normalizing constant:
(Y | Y ∈ A) ∼ exp{θ0 T (y) − ψA (θ)} f0 (y) 1A (y)
(24)
This fact lets us draw upon the rich theory of inference in multiparameter exponential families.
3.1
Conditional Inference and Nuisance Parameters
Classically, conditional inference in exponential families arises as a means for inference in the
presence of nuisance parameters, as in Model 4 below.
Model 4 (Exponential Family with Nuisance Parameters). Y follows a p-parameter exponential
family with sufficient statistics T (y) and U (y), of dimension k and p − k respectively:
Y ∼ fθ,ζ (y) = exp{θ0 T (y) + ζ 0 U (y) − ψ(θ, ζ)} f0 (y),
(25)
with (θ, ζ) ∈ Θ ⊆ Rp open.
Assume θ corresponds to a parameter of interest and ζ to an unknown nuisance parameter.
The conditional law L(T (Y ) | U (Y )) depends only on θ:
(T | U = u) ∼ gθ (t | u) = exp{θ0 t − ψg (θ | u)} g0 (t | u),
(26)
letting us eliminate ζ from the problem by conditioning on U . For k = 1 (i.e., for θ ∈ R), we
obtain a single-parameter family for T .
Consider testing the null hypothesis H0 : θ ∈ Θ0 ⊆ Θ against the alternative H1 : θ ∈ Θ1 =
Θ \ Θ0 . We say a level-α selective test φ(y) is selectively unbiased if
Powφ (θ | A) = Eθ [φ(Y ) | A] ≥ α,
for all θ ∈ Θ1 .
(27)
The condition (27) specializes to the usual definition of an unbiased test when there is no selection
(when A = Y). Unbiasedness rules out tests that privilege some alternatives to the detriment of
others, such as one-sided tests when the alternative is two-sided.
A uniformly most powerful unbiased (UMPU) selective level-α test is one whose selective power
is uniformly highest among all level-α tests satisfying (27). A selectively unbiased confidence region
is one that inverts a selectively unbiased test, and confidence regions inverting UMPU selective
15
tests are called uniformly most accurate unbiased (UMAU). All of the above specialize to the usual
definitions when A = Y.
See Lehmann and Romano (2005) or Brown (1986) for thorough reviews of the rich literature
on testing in exponential family models. In particular, the following classic result of Lehmann and
Scheffé (1955) gives a simple construction of UMPU tests in exponential family models.
Theorem 5 (Lehmann and Scheffé (1955)). Under Model 4 with k = 1, consider testing the
hypothesis
H0 : θ = θ0 against H1 : θ 6= θ0
(28)
at level α. There is a UMPU test of the form φ(Y ) = f (T (Y ), U (Y )) with
1 t < c1 (u) or t > c2 (u)
f (t, u) = γi t = ci (u)
0 c1 (u) < t < c2 (u)
(29)
where ci and γi are chosen to satisfy
Eθ0 [f (T, U ) | U = u] = α
(30)
Eθ0 [T f (T, U ) | U = u] = α Eθ0 [T | U = u] .
(31)
The condition (30) constrains the power to be α at θ = θ0 , and (31) is obtained by differentiating
the power function and setting its derivative to 0 at θ = θ0 .
Because L(Y | A) is an exponential family, we can simply apply Theorem 5 to the conditional
law L(Y | A) to obtain an analogous construction in the selective setting.
Corollary 6 (UMPU Selective Tests). Under Model 4 with k = 1, consider testing the hypothesis
H0 : θ = θ0
against
H1 : θ 6= θ0
(32)
at selective level α on selection event A. There is a UMPU selective test of the form φ(Y ) =
f (T (Y ), U (Y )) with
1 t < c1 (u) or t > c2 (u)
t = ci (u)
(33)
f (t, u) = γi
0
c1 (u) < t < c2 (u)
for which ci and γi solve
Eθ0 [f (T, U ) | U = u, Y ∈ A] = α
Eθ0 [T f (T, U ) | U = u, Y ∈ A] = α Eθ0 [T | U = u, Y ∈ A] .
(34)
(35)
We emphasize here that the test φ as defined above is not merely UMPU among selective tests
that condition on U , but rather it is UMPU among all selective level-α tests; see Lehmann and
Romano (2005) for more details. In some cases, it may be useful to interpret φ conditionally on
U = u, for example if the observed u leads to a more or less powerful test.
It is worth keeping in mind that unbiasedness is only one way to choose a test when there is
no completely UMP one. For example, another simple choice is to use the equal-tailed test from
the same conditional law (26). The equal-tailed level-α rejection region is simply the union of the
one-sided level-α/2 rejection regions. While the equal-tailed and UMPU tests choose ci and γi in
different ways, both tests take the form (29). In fact, as we will see next, all admissible tests are
of this form, which implies that data splitting tests are usually inadmissible.
16
3.2
Conditioning, Admissibility, and Data Splitting
A selective level-α test φ is inadmissible on selection event A if there exists another selective level-α
test φ∗ for which
Eθ,ζ [φ∗ (Y )|A] ≥ Eθ,ζ [φ(Y )|A], for all (θ, ζ) ∈ Θ1 ,
(36)
with the inequality strict for at least one (θ, ζ). In the main result of this section, we will show
that tests based on data splitting are nearly always inadmissible.
Let Y be an observation from Model 4, and suppose we wish to test
H0 : θ = θ0
against
H1 : θ 6= θ0 .
(37)
We will assume all tests are functions of the sufficient statistic and write (with some abuse of
notation) φ(T, U ) for φ(Y ). We can do this without loss of generality because any test φ(Y ) can
be Rao-Blackwellized, i.e.,
φ(T, U ) ≡ E[φ(Y )|T, U ],
to obtain a new test that is a function of (T, U ), with the same power function as the original.
Therefore, if φ(T, U ) is inadmissible, then so is the original test φ(Y ).
Now we can apply the following result of Matthes and Truax (1967).
Theorem 7 (Matthes and Truax, Theorem 3.1). Let Y be an observation from Model 4, and
suppose we wish to test
H0 : θ = θ0 against H1 : θ =
6 θ0 .
(38)
Let C denote the class of all level-α tests φ(T, U ) of the form
t ∈ int C(u)
0
φ(t, u) = γ(t, u) t ∈ ∂C(u) ,
1
t∈
/ C(u)
(39)
and C(u) is a convex set for every u. Then, for any φ ∈
/ C , there exists φ∗ ∈ C such that
Eθ,ζ [φ∗ (T, U )] ≥ Eθ,ζ [φ(T, U )],
for all (θ, ζ) ∈ Θ1 .
(40)
Notice that, if (40) holds with equality for all (θ, ζ), then by the completeness of (T, U ) we
a.s.
have φ = φ∗ . Hence, every admissible test is in C or almost surely equal to a test in C .
In order to apply this result to data splitting, we first introduce a generic exponential family
composed of two independent data sets governed by the same parameters:
Model 8 (Exponential Family with Data Splitting). Model independent random variables (Y1 , Y2 ) ∈
Y1 × Y2 as
Yi ∼ exp {θTi (y) + ζ 0 Ui (y) − ψi (θ, ζ)} f0,i (y), i = 1, 2,
(41)
with θ ∈ R and with the models for Yi both satisfying Model 4.
Model 8 would, for example, cover the case where Y1 and Y2 are the responses for two linear
regressions with different design matrices but the same regression coefficients.
For a selection event A = A1 × Y2 , we say φ is a data-splitting test if φ(Y ) = φ2 (Y2 ); that is, the
selection stage uses only Y1 and the inference stage uses only Y2 . Again, by Rao-Blackwellization,
we can assume without loss of generality that the test is of the form φ(T2 , U2 ).
17
Next, define the cutoff gap g ∗ (φ) as the largest g ≥ 0 for which the acceptance and rejection
regions are separated by a “cushion” of width g. If T2∗ is a conditionally independent copy of T2
given U2 , then
g ∗ (φ) = sup {g : Pθ,ζ (|T2 − T2∗ | < g, φ(T2 , U2 ) > 0, φ(T2∗ , U2 ) < 1) = 0}.
(42)
Note that the support of (T2 , T2∗ , U2 ) does not depend on θ or ζ; thus, neither does g ∗ . For most
tests, g ∗ (φ) = 0. For example, g ∗ = 0 if either cutoff is in the interior of supp(T2 | U2 ) with positive
probability, or if φ is a randomized test for discrete (T2 , U2 ).
Next we prove the main technical result of this section: φ is inadmissible unless T1 is determined
by U1 on A1 , within an amount g ∗ of variability.
Theorem 9. Let T1∗ denote a copy of T1 that is conditionally independent given U1 and Y1 ∈ A,
and let φ be a data-splitting test of (38) in Model 8. If
Pθ,ζ (|T1 − T1∗ | > g ∗ (φ) | Y1 ∈ A) > 0
then φ is inadmissible.
Proof. Construct conditionally independent copies Ti∗ with (T1 , T1∗ , U1 ) ⊥⊥ (T2 , T2∗ , U2 ), and assume that φ is of the form φ(T, U ) with T = T1 + T2 and U = U1 + U2 (otherwise we could
Rao-Blackwellize it). If φ is admissible, then by Matthes and Truax (1967), it must be a.s. equivalent to a test of the form (39). That is, there exist ci (U ) for which
Pθ,ζ (φ(T, U ) < 1, T ∈
/ [c1 (U ), c2 (U )] | A) = Pθ,ζ (φ(T, U ) > 0, c1 (U ) < T < c2 (U ) | A) = 0. (43)
Now, by assumption, there exists δ > g ∗ (φ) for which
B1 , {|T1 − T1∗ | > δ}
occurs with positive probability. By the definition of g ∗ (φ) in (42), the event
B2 , {|T2 − T2∗ | > δ, φ(T2 , U2 ) > 0, φ(T2∗ , U2 ) < 1}
also occurs with positive probability. Since the two events are independent, B = B1 ∩ B2 occurs
with positive probability.
Next, assume w.l.o.g. that the event in (42) can occur with T2∗ > T2 (otherwise we could
reparameterize with natural parameter ξ = −θ, for which −Ti would be the sufficient statistics for
Yi ). Then for some δ > g ∗ (φ), the event
B = {T1 + δ < T1∗ , T2 < T2∗ < T2 + δ, φ(T2 , U2 ) > 0, and φ(T2∗ , U2 ) < 1}
occurs with positive probability for all θ, ζ. On B,
T1 + T2 < T1 + T2∗ < T1∗ + T2 < T1 + T2∗ ,
but φ(T, U ) > 0 for T = T1 + T2 and T = T1∗ + T2 and φ(T, U ) < 1 for the other two, ruling out
the possibility of (43).
In the typical case g ∗ = 0 and we have
Corollary 10. Suppose φ is a data-splitting test of (38) in Model 8 with g ∗ (φ) = 0. Then φ is
inadmissible unless T1 is a function of U1 on A.
18
Expected CI Length
3
1
2
Interval Length
1
Information
4
2
Leftover Fisher Information
−2
0
2
Data Splitting
Data Carving
0
0
Data Splitting
Data Carving
4
6
8
−2
0
µ
2
4
6
8
µ
(a) Fisher information available for second-stage inference.
(b) Expected confidence interval length.
Figure 4: Contrast between data splitting and data carving in Example 3, in which Yi ∼ N (µ, 1)
independently for i = 1, 2. Data splitting discards Y1 entirely, while data carving uses the leftover
information in Y1 for the second-stage inference. When µ 3, data carving also uses about one
data point for inference since there is no information left over in Y1 . But when µ 3, conditioning
barely effects the law of Y1 and data carving has nearly two data points left over.
Example 3. To illustrate Theorem 9, consider a bivariate version of Example 2:
Yi ∼ N (µ, 1), i = 1, 2,
with Y1 ⊥⊥ Y2 ,
(44)
in which we condition on the selection event A = {Y1 > 3}.
With data splitting, we could construct a 95% confidence interval using only Y2 ; namely,
Y2 ± 1.96. This interval is valid but does not use all the information available. A more powerful
alternative is to construct an interval based on the law
Lµ Y1 + Y2 Y1 > 3 ,
(45)
which uses the leftover information in Y1 .
Figure 4a shows the Fisher information that is available to each test as a function of µ. The
Fisher information of data splitting is exactly 1 no matter what µ is, whereas the optimal selective
test has information approaching 2 as µ increases. Figure 4b shows the expected confidence interval
length of the equal tailed interval as a function of µ. For µ
√ 3, the data splitting interval is roughly
41% longer than it needs to be (in the limit, the factor is 2 − 1).
Together, the plots tell a consistent story: when the selection event is not too unlikely, discarding the first data set exacts an unnecessary toll on the power of our second-stage procedure.
4
Selective Inference for Linear Regression
For a concrete example of the exponential family framework discussed in Section 3, we now turn to
linear regression, which is one of the most important applications of selective inference. In linear
regression, the data arise from a multivariate normal distribution
Y ∼ Nn (µ, σ 2 In ),
19
(46)
where µ is modeled as
µ = XM β M .
(47)
To avoid trivialities, we will assume that XM has full column rank for all M under consideration,
so that β M is well-defined.
Depending on whether σ 2 is assumed known or unknown, hypothesis tests for coordinates βjM
generalize either the z-test or the t-test. In the non-selective case, z- and t-tests are based on
†
†
coordinates of the ordinary least squares (OLS) estimator β̂ = XM
Y , where XM
is the Moore0
Penrose pseudoinverse. For a particular j and M , it will be convenient to write β̂jM = ηjM Y
with
Xj·M
⊥
Xj
(48)
, where Xj·M = PX
ηjM =
M \j
kXj·M k2
is the remainder after adjusting Xj for the other columns of XM , and PXM \j denotes projection
⊥
onto the column space of XM \j . Letting σ̂ 2 = kPX
Y k2 /(n − |M |), the test statistics
M
0
Z=
ηjM Y
σkηjM k
0
and
Te =
ηjM Y
σ̂kηjM k
(49)
are respectively distributed as N (0, 1) and tn−|M | under H0 : βjM = 0. Henceforth, we will suppress
the subscript and superscript for ηjM , simply writing η when there is no ambiguity. The optimal
selective t- and z-tests are based on the same test statistics, but compared against different null
distributions.
We consider two distinct modeling frameworks: Section 4.1 concerns inference under the more
restrictive selected linear model, the family of distributions for which (47) and (46) both hold, while
Section 4.2 concerns inference under the more general saturated model which assumes only (46)
0
and performs inference on ηjM µ. As we will see, selected-model tests can be more powerful than
saturated-model tests, but the extra power comes at a price since the inferences are only valid under
more restrictive modeling assumptions. Section 4.3 compares and contrasts the two approaches.
4.1
Inference Under the Selected Model
Suppressing the superscript M in β M , the selected model has the form
1 0
1
0
2
2
Y ∼ exp
β
X
kyk
−
ψ(X
β,
σ
)
y
−
M
M
σ2
2σ 2
(50)
If σ 2 is known, the sufficient statistics are Xk0 Y for k ∈ M , and inference for βj is based on
Lβj Xj0 Y
XM \j 0 Y, A .
(51)
Otherwise, kY k2 represents another sufficient statistic and inference is based on
Lβj /σ2 Xj0 Y
XM \j 0 Y, kY k, A .
(52)
Decomposing
⊥
Xj0 Y = Xj0 PXM \j Y + Xj0 PX
Y
M \j
=
Xj0 PXM \j Y
2 0
+ kXj.M k η Y,
(53)
(54)
0
we see that Z = η 0 Y /σkηk is a fixed affine transformation of Xj0 Y once we condition on XM
\j Y .
2
If σ is known, then, we can equivalently base our selective test on
Lβj Z
XM \j 0 Y, A .
(55)
20
0
While Z is marginally independent of XM
\j Y , it is generically not conditionally independent given
0
A, so that the null distribution of Z generically depends on XM
\j Y .
2
If σ is unknown, we may observe further that
σ̂ 2 =
⊥
kY k2 − kPXM \j Y k2 − (η 0 Y )2 /kηk2
kPX
Y k2
M
=
.
n − |M |
n − |M |
(56)
Writing Z0 (Y ) = η 0 Y /kηk, we have Te(Y ) = (n − |M |) Z0 /(kY k2 − kPXM \j Y k2 − Z02 ), which is a
0
monotone function of η 0 Y after fixing kY k2 and XM
\j Y . Thus, our test is based on the appropriate
conditional law of
Lβ /σ2 Te XM \j 0 Y, kY k, A .
(57)
j
Note that, given A, σ̂ 2 in (56) is neither unbiased for σ 2 nor χ2 -distributed. We recommend against
viewing it as a serious estimate of σ 2 in the selective setting.
Constructing a selective t-interval is not as straightforward as the general case described in
Section 5.2 because βj is not a natural parameter of the selected model; rather, βj /σ 2 is. Testing
βj = 0 is equivalent to testing βj /σ 2 = 0, but testing βj = b for b 6= 0 does not correspond to any
point null hypothesis about βj /σ 2 . However, we can define
Ye = Y − bXj ∼ N (Xβ − bXj , σ 2 I).
(58)
Because (βj − b)/σ 2 is a natural parameter for Ye , we can carry out a UMPU selective t-test for
H0 : βj = b ⇐⇒ (βj − b)/σ 2 = 0 based on the law of Ye .
4.2
Inference Under the Saturated Model
Even if we do not take the linear model (47) seriously, there is still a well-defined best linear
predictor in the population for design matrix XM :
†
θM = arg min Eµ kY − XM θk2 = XM
µ,
(59)
θ
We call θM the least squares coefficients for M . According to this point of view, each θjM corre0
sponds to the linear functional ηjM µ.
This point of view is convenient because the least-squares parameters are well-defined under
the more general saturated model (6), leading to meaningful inference even if we do a poor job of
selecting predictors. In particular, Berk et al. (2013) adopt this perspective as a way of avoiding
the need to consider multiple candidate statistical models.
Several recent articles have tackled the problem of exact selective inference in linear regression
after specific selection procedures (Lee et al., 2016; Loftus and Taylor, 2014; Lee and Taylor, 2014).
These works, as well as Berk et al. (2013), assume the error variance is known, or that an estimate
may be obtained from independent data, and target least-squares parameters in the saturated
model.
Under the selected model, βjM = θjM = η 0 µ, whereas under the saturated model β M may not
exist (i.e., there is no β M such that µ = XM β M ). Compared to the selected model, the saturated
⊥
model has n − |M | additional nuisance parameters corresponding to PX
µ.
M
We can write the saturated model in exponential family form as
1
1 0
2
2
µ y − 2 kyk − ψ(µ, σ ) ,
(60)
Y ∼ exp
σ2
2σ
21
which has n + 1 natural paramaters if σ 2 is unknown and n otherwise. To perform inference on
some least-squares coefficient θjM = η 0 µ, we can rewrite (60) as
Y ∼ exp
1
1
1
0
0
⊥ 0
⊥
2
2
µ η η y + 2 (Pη µ) (Pη y) − 2 kyk − ψ(µ, σ ) .
σ 2 kηk2
σ
2σ
(61)
If σ 2 is known, inference for θjM after selection event A is based on the conditional law
LθjM η 0 Y
Pη⊥ Y, A , or equivalently LθjM Z
Pη⊥ Y, A .
If σ 2 is unknown, we must instead base inference on
Pη⊥ Y, kY k, A .
(62)
LθjM /σ2 η 0 Y
Unfortunately, the conditioning in (62) is too restrictive. The set
y : Pη⊥ y = Pη⊥ Y, kyk = kY k
(63)
is a line intersected with the sphere kY kS n−1 , and consists only of the two points {Y, Y −
2η 0 Y }, which are equally likely under the hypothesis θjM = 0. Thus, under the saturated model,
conditioning on kY k leaves insufficient information about θjM to carry out a meaningful test.
4.3
Saturated Model or Selected Model?
When σ 2 is known, we have a choice whether to carry out the z-test with test statistic Z =
η 0 Y /σkηk in the saturated or the selected model. In other words, we must choose either to assume
⊥
that PX
µ = 0 or to treat it as an unknown nuisance parameter. Writing
M
U = XM \j 0 Y,
and
⊥
V = PX
Y,
M
(64)
we must choose whether to condition on U and V (saturated model) or only U (selected model).
Conditioning on both U and V can never increase our power relative to conditioning only on U ,
and (unless the tests coincide) will lead to an inadmissible test per Theorem 9.
In the non-selective case, this choice makes no difference at all since T, U, and V are mutually
independent. In the selective case, however, the choice may be of major consequence as it can
lead to very different tests. In general, T, U, and V are not conditionally independent given A,
⊥
and PX
µ may play an important role in determining the conditional distribution of T . If we
M
needlessly condition on V , we may lose a great deal of power, whereas failing to condition on V
⊥
could lead us astray if PX
µ is large. A simple example can elucidate this contrast.
M
Example 4. Suppose that y ∼ N2 (µ, I2 ), with design matrix X = I2 , and we choose the bestfitting one-sparse model. That is, we choose M = {1} if |Y1 | > |Y2 |, and M = {2} otherwise.
Figure 5 shows one realization of this process with Y = (2.9, 2.5). |Y1 | is a little larger than |Y2 |,
so we choose M = {1}. The yellow highlighted region A = {|Y1 | > |Y2 |} is the chosen selection
event, and the selected model is
Y ∼ N2 ((µ1 , 0), I2 ) .
(65)
In this case, T = Y1 , V = Y2 , and there is no U since XM has only one column. The selectedmodel test is based on L(Y1 | A), whereas the saturated-model test is based on L(Y1 | Y2 , A). The
second conditioning set, a union of two rays, is plotted in brown. Under the hypothesis µ = 0,
the realized |Y1 | is quite large given A, giving p-value 0.007. By contrast, |Y1 | is not terribly large
given {Y2 = 2.5} ∩ A = {Y2 = 2.5, |Y1 | > 2.5}, leading to p-value 0.30.
22
Conditional Null Distributions
1.4
Conditioning Sets
1.2
0.0
−4
0.2
0.4
−2
0
A
0.6
Density
2
Y
0.8
1.0
4
●
Y2
Saturated Model
Selected Model
Observed Value
−4
−2
0
2
4
−6
Y1
−4
−2
0
2
4
6
Y1
(a) For Y = (2.9, 2.5), the selected-model conditioning set is A = {y : |y1 | > |y2 |}, a union
of quadrants, plotted in yellow. The saturatedmodel conditioning set is {y : y2 = 2.5} ∩ A =
{y : y2 = 2.5, |y1 | > 2.5}, a union of rays, plotted
in brown.
(b) Conditional distributions of Y1 under H0 : µ1 =
0. Under the hypothesis µ = 0, the realized |Y1 |
is quite large given A, giving p-value 0.007. By
contrast, |Y1 | is not too large given A ∩ {y : y2 =
Y2 }, giving p-value 0.3.
Figure 5: Contrast between the saturated-model and selected-model tests in Example 4, in which
we fit a one-sparse model with design matrix X = I2 . The selected-model test is based on
L0 (Y1 | A), whereas the saturated-model test is based on L0 (Y1 | Y2 , A).
The selected-model approach is especially well-suited for testing goodness of fit of a selected
linear model — in that case, we prefer the test not to have level α, but rather to reject with high
probability, when important variables are not selected. Fithian et al. (2015) consider sequential
goodness-of-fit testing in a “path” of increasingly complex models selected by a method like the
lasso or forward stepwise regression. As Example 4 illustrates, the saturated-model p-value is
especially large for “near ties” when |Y1 | is not much larger than |Y2 |. As a result, the selectedmodel test can be much more powerful in early steps of the path where multiple strong variables
compete to enter the model first. For more details see Fithian et al. (2015).
5
Computations
We saw in Section 3 that inference in the one-parameter exponential family requires knowing
the conditional law Lθ (T | U, A). In a few cases, such as in the saturated model viewpoint, this
conditional law can be determined fairly explicitly. In other cases, we will need to resort to Monte
Carlo sampling. In this section, we suggest some general strategies.
5.1
Gaussians Under the Saturated Model
As we discussed in Section 4.2, the previous papers by Lee et al. (2016); Loftus and Taylor (2014);
2
Lee and Taylor (2014) adopted the
saturated model viewpoint with known σ . In this case,
Lθ (T | U, A) = Lθ η 0 Y
Pη⊥ Y, A is a truncated univariate Gaussian, since η 0 Y is a Gaussian
random variable and Pη⊥ Y is independent of η 0 Y . If A is convex, then the truncation is to an
interval [V − (Y ), V + (Y )], where the endpoints represent the maximal extent one can move in the
23
Figure 6: Saturated-model inference for a generic convex selection set for Y ∼ N (µ, In ). After
conditioning on the yellow set A, V + is the largest η 0 Y can get while V − is the smallest it can
get. Under H0 : η 0 µ = 0, the test statistic η 0 Y takes on the distribution of a standard Gaussian
0
Y )−Φ(V − )
random variable truncated to the interval [V − , V + ]. As a result, W (Y ) = Φ(η
Φ(V + )−Φ(V − ) is uniformly
distributed.
η direction at a “height” of Pη⊥ Y , while still remaining inside A, i.e.,
V + (Y ) =
sup
η 0 (Y + tη)
(66)
η 0 (Y + tη).
(67)
{t: Y +tη ∈ A}
V − (Y ) =
inf
{t: Y +tη ∈ A}
The geometric intuition is illustrated in Figure 6.
When A is specifically a polytope, we can obtain closed-form expressions for V − and V + . The
generalization to regions A that are non-convex is straightforward (i.e., instead of truncating to a
single interval, we truncate to a union of intervals). For further discussion of these points, see Lee
et al. (2016).
5.2
Monte Carlo Tests and Intervals
In a more generic setting, we may not have an easy formula for conditional law of T . In that case,
there are several options for inference using Monte Carlo methods.
If we can obtain a stream of samples from Lθ (T | U, A) for any value of θ, then we can carry out
hypothesis tests and construct intervals. This can be done efficiently via rejection sampling if, for
example, we can sample efficiently from Lθ (Y | U ) and Pθ (Y ∈ A | U ) is not too small. Otherwise,
more specialized sampling approaches may be required. A little more abstractly, we now consider
constructing a test based on the statistic Z, which is distributed according to a one-parameter
exponential family
Z ∼ gθ (z) = eθz−ψ(θ) g0 (z).
(68)
Exact Monte Carlo Tests Suppose that, in addition to Z, we are given an independent sequence from the reference distribution
i.i.d.
Z1 , . . . , Zn ∼ g0 (z).
24
(69)
Then an exact Monte Carlo one-sided test of H0 : θ ≤ 0 rejects if the observed value Z is among
the (n + 1) α largest of Z, Z1 , . . . , Zn (Barnard, 1963).
Even if i.i.d. samples are not available, the same procedure has level α provided the law
of (Z, Z1 , . . . , Zn ) is exchangeable under H0 . Besag and Clifford (1989) propose an ingenious
procedure for obtaining such an exchangeable sequence when we only know how to run a Markov
chain with stationary distribution g0 : Beginning at Z, take k ≥ 1 steps backward in the chain to
e Then, run n independent chains k steps forward, beginning each chain at Z,
e and letting Zi
Z.
denote the end state of the ith chain. If Z ∼ g0 , the sequence is exchangeable.
Note that, while this test has level α for any k, n ≥ 0, using small values of k, n makes the test
more random, reducing its power. If the chain is irreducible, then as k, n → ∞, the test converges
to the deterministic right-tailed level-α test of H0 .
Approximate Monte Carlo Intervals By reweighting the samples, we can use (Z1 , . . . , Zn )
to test H0 : θ ≤ θ0 for any other θ0 . Denote the importance-weighted empirical expectation as
Pn
h(Z )eθZi
b
Pn iθZ
(70)
Eθ h(Z) = i=1
i
i=1 e
a.s.
−−→ Eθ h(Z)
as n → ∞ for integrable h.
(71)
In effect, we have put an exponential family “through” the empirical distribution of the Zi in the
manner of Efron et al. (1996); see also Besag (2001). The Monte Carlo one-sided cutoff for a test
bθ (Z > c2 ) ≤ α. The test rejects for Z > c2 and
of H0 : θ ≤ θ0 is the smallest c2 for which P
0
randomizes appropriately at Z = c2 .
The two-sided test of H0 : θ = θ0 is a bit more involved, but similar in principle. We can solve
for c1 , γ1 , c2 , γ2 for which
b θ φ(Z) = α
E
0
b θ [Zφ(Z)] = α E
b θ Z.
E
0
0
(72)
(73)
In Appendix B we discuss how (72–73) can be solved efficiently for fixed θ0 and inverted to obtain
a confidence interval. Monte Carlo inference as described above is computationally straightforward
once Z1 , . . . , Zn are obtained.
More generally, the Zi could represent importance samples with weights Wi , or steps in a
Markov chain with stationary distribution g0 (z). The same methods apply as long as we still have
Pn
W h(Zi ) eθZi
b
Pn i
(74)
Eθ h(Z) = i=1
θZi
i=1 Wi e
a.s.
−−→ Eθ h(Z),
for integrable h.
(75)
Numerical problems may arise in solving (72–73) for θ0 far away from the reference parameter
used for sampling. Combining appropriately weighted samples from several different reference
values can help to keep the effective sample size from getting too small for any θ0 . For further
references on Monte Carlo inference see Jockel (1986); Forster et al. (1996); Mehta et al. (2000).
5.3
Sampling Gaussians with Affine and Quadratic Constraints
In the case where Y is Gaussian, several simplifications are possible. For one, there are many ways
to sample from a truncated multivariate Gaussian distribution. In this paper, we use hit-and-run
Gibbs sampling algorithms, while Pakman and Paninski (2014) suggest another approach based
on Hamiltonian Monte Carlo.
25
Efficient sampling from multivariate Gaussian distributions under such constraints is the main
algorithmic challenge for most of the Gaussian selective tests proposed in this paper. The works
cited above use the saturated model exclusively which means they do not require any sampling.
In many cases, the sampling problem may be greatly facilitated by refining the selection variable
that we use. For example, Lee et al. (2016) propose conditioning on the variables selected by the
lasso as well as the signs of the fitted β̂j , leading to a selection event consisting of a single polytope
in Rn . If we condition only on the selected variables and not on the signs, the selection event is
a union of up to 2s polytopes, where s is the number of variables in the selected model (though
most of the polytopes might be excluded after conditioning on U ).
Refining the selection variable never impairs the selective validity of the procedure, but it
typically leads to a loss in power. However, this loss of power may be quite small if, for example,
the conditional law puts nearly all of its mass on the realized polytope. This price in power
is acceptable if it is the only way to obtain a tractable test. Quantifying the tradeoff between
computation and power is an interesting topic for further work.
When carrying out selective t-tests, it is necessary to condition further on the realized vector
length kY k, adding a quadratic equality constraint to the support. To deal with this, we sample
instead from a ball and project the samples onto the sphere using an importance sampling scheme.
Appendix C gives details.
6
Selective Inference in Non-Gaussian Settings
In this section we describe tests in two simple non-Gaussian settings, selective inference in a
binomial problem, and tests involving a scan statistic in Poisson process models. More generally,
we address the question of selective inference in generalized linear models.
6.1
Selective Clinical Trial
To illustrate the application of our approach in a simple non-Gaussian setting we discuss a selective
clinical trial with binomial data. The experiment discussed here is similar to an adaptive design
proposed by Sill and Sampson (2009).
Consider a clinical trial with m candidate treatments for heart disease. We give treatment j
to nj patients for 0 ≤ j ≤ m, with j = 0 corresponding to the placebo. The number of patients
on treatment j to suffer a heart attack during the trial is
pj
θ
j=0
ind.
=
,
(76)
Yj ∼ Binom(pj , nj ),
with log
θ − βj j > 0
1 − pj
so βj measures the efficacy of treatment j. The likelihood for Y is
m
m
m
X
Y
X
Yj
Y ∼ exp θ
yj −
βj yj − ψ(θ, β)
,
nj
j=0
j=1
j=0
(77)
an exponential family with m + 1 sufficient statistics. Define p̂j = Yj /nj , and let p̂(j) denote the
jth smallest order statistic.
After observing the data, we select the best k < m treatments in-sample, then construct a
confidence interval for each one’s odds ratio relative to placebo. If there are ties, we select all
treatments for which p̂j ≤ p̂(k) (so that we could possibly select more than k treatments).
26
For simplicity, assume that treatments 1, . . . , k are the ones selected. Inference for β1 is then
based on the conditional law
m
X
Lβ1 Y1
Yj , Y2 , . . . , YJ , {j = 1 selected}
(78)
j=0
Under this law, Y2 , . . . , Ym are fixed, as is Y0 + Y1 , with Y0 and Y1 the only remaining unknowns.
Before conditioning on selection, we have the two-by-two multinomial table
Control
Treatment
Y0
Y1
n0 − Y0
n1 − Y1
Heart attack
No heart attack
The margins are fixed, and conditioning on selection gives an additional constraint that Y1 ≤
n1 p̂(k) , where the right-hand side is known after conditioning on the other Yj . Rejecting for
conditionally extreme Y1 amounts to a selective Fisher’s exact test. Aside from the constraint on its
support, the distribution of Y1 is hypergeometric if β1 = 0 and otherwise noncentral hypergeometric
with noncentrality parameter β1 . We can use this family to construct an interval for β1 .
6.2
Poisson Scan Statistic
As a second simple example, consider observing a Poisson process Y = {Y1 , . . . , YN (Y ) } on the
interval [0, 1] with piecewise-constant intensity, possibly elevated in some unknown window [a, b].
That is, Y ∼ Poisson(λ(t)) with
(
eα+β t ∈ [a, b]
λ(t) =
(79)
eα
otherwise.
Our goal is to locate [a, b] by maximizing some scan statistic, then test whether β > 0 or construct
a confidence interval for it. Assume we always have [â, b̂] = [Yi , Yj ] for some i, j; this is true,
for example, if we use the multi-scale-adjusted likelihood ratio statistic proposed in Rivera and
Walther (2013).
The density of Y can be written in exponential family form as
Z 1
(y)
NX
Y ∼ exp
log λ(yi ) −
λ(s) ds
(80)
0
i=1
= exp α N (Y ) + β T (y) − ψ(α, β) ,
(81)
where
N (y)
T (y) =
X
1{yi ∈ [a, b]}
and
ψ(α, β) = eα (1 − b + a) + eα+β (b − a).
(82)
i=1
If A is the event that [a, b] is chosen, we carry out inference with respect to Lβ (T | N, A). Note
that under β = 0 and conditional on N , Y is an i.i.d. uniform random sample on [0, 1].
Once we condition on the event {a, b ∈ Y }, the other N − 2 values are uniform. Thus, we can
sample from Lβ (T | N, A) with β = 0 by taking Y to include a, b, and N − 2 uniformly random
points, then rejecting samples for which [a, b] is not the selected window.
27
6.3
Generalized Linear Models
Our framework extends to logistic regression, Poisson regression, or other generalized linear model
(GLM) with response Y and design matrix X, since the GLM model may be represented as an
exponential family of the form
Y ∼ exp {β 0 X 0 y − ψ(Xβ)} f0 (y).
(83)
As a result, we can proceed just as we did in the case of linear regression in the reduced model,
conditioning on U = XM \j 0 Y and basing inference on LβjM (Xj0 Y | U, A).
A difficulty may arise for logistic or Poisson regression due to the discreteness of the response
distribution Y . If some control variable X1 is continuous, then for almost every realization of X,
all configurations of Y yield unique values of U = X10 Y . In that case, conditioning on X10 Y means
conditioning on Y itself. No information is left over for inference, so that the best (and only) exact
level-α selective test is the trivial one φ(Y ) ≡ α. By contrast, if all of the control variables are
discrete variables like gender or ethnicity, then conditioning on U may not constrain Y too much.
Because X 0 Y is approximately a multivariate Gaussian random variable, a more promising
approach may be to base inference on the asymptotic Gaussian approximation as in Taylor and
Tibshirani (2016).
7
Simulation: High-Dimensional Regression
As a simple illustration, we compare selective inference in linear regression after the lasso for
n = 100, p = 200. Here, the rows of the design matrix X are drawn from an equicorrelated
multivariate Gaussian distribution with pairwise correlation ρ = 0.3 between the variables. The
columns are normalized to have length 1.
We simulate from the model
Y ∼ N (Xβ, In ),
(84)
with β 7-sparse and its non-zero entries set to 7. The magnitude of β was chosen so that data
splitting with half the data yielded a superset of the true variables on roughly 20% of instances.
For data splitting and carving, Y is partitioned into selection and inference data sets Y1 and Y2 ,
containing n1 and n2 = n − n1 data points respectively.
We assume the error variance is known and carry out the Lasso on Y1 with Lagrange parameter
λ = 2E(kX T k∞ ),
∼ N (0, In )
as described in (Negahban et al., 2012). We then compare two post-selection inference procedures:
Data Splitting after Lasso on Y1 (Splitn1 ): Use the lasso on Y1 to select the model, and use
Y2 for inference.
Data Carving after Lasso on Y1 (Carven1 ): Use the lasso on Y1 to select the model, and use
Y2 and whatever is left over of Y1 for inference.
For the data carving procedures, we use the selected-model z-test of Section 4.1. In addition, we
condition on the signs of the active lasso coefficients, so procedure Carve100 is the inference-afterlasso test proposed in Lee et al. (2016).4
We know from Theorem 9 that procedure Carven1 strictly dominates procedure Splitn1 for
any n1 , but there is a selection–inference tradeoff between data-carving procedures Carven and
4 Because of the form of the selection event when we use the lasso after n data points, the test statistic is
⊥ Y . Thus, there is no distinction between the saturated- and selected-model
conditionally independent of PX
M
z-tests after the lasso on all n data points.
28
pscreen
E[V ]
E[R − V ]
FDR
Power
Level
Carve100
0.99
8.13
6.99
0.54
0.80
0.05
Split50
0.09
9.13
4.74
0.66
0.93
0.06
Carve50
0.09
9.13
4.74
0.66
0.99
0.06
Split75
0.68
9.24
6.59
0.58
0.47
0.05
Carve75
0.68
9.24
6.59
0.58
0.97
0.06
Algorithm
Table 1: Simulation results. pscreen is the probability of successfully selecting all 7 true variables,
and Power is the power, conditional on successful screening, of tests on the true variables. The
more data we use for selection, the better the selected model’s quality is, but there is a cost in
second-stage power. Carve75 appears to be finding a good tradeoff between these competing goals.
Carven1 always outperforms Splitn1 , as predicted by Theorem 9.
pscreen
E[V ]
E[R − V ]
FDR
Power
Level
Carve100
0.97
8.11
6.97
0.54
0.80
0.04
Split50
0.09
9.20
4.77
0.66
0.93
0.05
Carve50
0.09
9.20
4.77
0.66
0.99
0.06
Algorithm
Table 2: Simulation results under misspecification. Here, errors are drawn independently from
Student’s t5 . Our conclusions are identical to Table 1.
Carven1 for n1 < n. Carven uses all of the data for selection, and is therefore likely to select a
superior model, whereas procedure Carven1 reserves more power for the second stage.
Let R be the size of the model selected and V the number of noise variables included. We
compare the procedures with respect to aspects of their selection performance:
• chance of screening, i.e. obtaining a correct model (P(R − V = 7) or pscreen ).
• expected number of noise variables selected (E[V ]),
• expected number of true variables selected (E[R − V ]),
• false discovery rate of true variables selected (E[V / max(R, 1)] or FDR),
Conditional on having obtained a correct model, we also compare them on aspects of their
second stage performance:
• probability of correctly rejecting the null for one of the true variables (Power),
• probability of incorrectly rejecting the null for a noise variable (Level).
The results, shown in Table 1, bear out the intuition of Section 3.2. Because procedure Carve100
uses the most information in the first stage, it performs best in terms of model selection, but pays
29
1.0
0.8
0.6
0.4
Screening x Power
0.0
0.2
1.0
0.8
0.6
0.4
0.0
0.2
Probability
Screening
Power, Carving
Power, Splitting
Data Carving
Data Splitting
0
20
40
60
80
100
0
# data points used for selection
20
40
60
80
100
# data points used for selection
(a) Probability of successful screening, and power
conditional on screening, for Splitn1 and Carven1 .
(b) Probability of successful screening times power
conditional on screening, for Splitn1 and Carven1 .
Figure 7: Tradeoff between power and model selection. As n1 increases and more data is used
in the first stage, we have a better chance of successful screening (picking all the true nonzero
variables). However, increasing n1 also leads to reduced power in the second stage. Data splitting
suffers much more than data carving, though both are affected.
a price in lower second-stage power relative to Split50 or Carve50 . The procedure Carve50 clearly
dominates Split50 , as expected. Increasing n1 from 50 to 75 improves pscreen for Split75 , but Split75
suffers a drop in power. Procedure Carve75 seems to strike a better compromise.
Figure 7 shows the tradeoff curve of model selection success (as measured by the probability
of successful screening) against second-stage power conditional on successful screening. As n1
increases, stage-one performance improves while stage-two performance declines, but the decline
is much slower for data carving. Surprisingly, Carve98 and Carve99 have much higher power than
Carve100 : 91%, 86%, and 80% respectively. We cannot explain why holding out just one or two
data points in the first stage improves power so dramatically. Better understanding this tradeoff
is an interesting topic of further work.
Finally, to check the robustness of data carving, we replace the Gaussian errors with independent errors drawn from Student’s t distribution with five degrees of freedom. The numbers barely
change at all; see Table 2. Tian and Taylor (2017) rigorously analyze the case of non-Gaussian
errors.
8
Conditioning as a Device for Multiple Inference
To this point we have argued for controlling selective type I error as a goal in its own right,
but it can also serve as a device for controlling more traditional multiple inference goals. In this
section we discuss two examples: confidence intervals for selected parameters that control the false
coverage-statement rate (FCR) and familywise error rate (FWER).
Suppose that θq , q = 1, . . . , m correspond to parameters of a common (fixed) model M . We
b )| of them as interesting and construct a confidence
adaptively designate a number R(Y ) = |Q(Y
b Benjamini and Yekutieli (2005) propose controlling the false
interval Cq (Y ) for each q ∈ Q.
30
coverage-statement rate (FCR)
V
,
E
max(R, 1)
where
V (Y ) =
n
o
b θq (F ) ∈
q : q ∈ Q,
/ Cq (Y )
(85)
is the number of non-covering intervals constructed.
Other authors have addressed inference after selection by proposing to control the FWER, the
chance that any selected test incorrectly rejects the null or any constructed confidence interval fails
to cover its parameter. For example, the “post-selection inference” (PoSI) method of Berk et al.
(2013) constructs simultaneous (1 − α) confidence intervals for the least-squares parameters of all
linear regression models that were ever under consideration. As a result, no matter how we choose
the model, the overall probability of constructing any non-covering interval is controlled at α.
By choosing appropriate selection variables Sq , we can control the FCR or FWER as desired
using intervals with selective coverage. Our proof generalizes and extends a result in Weinstein
et al. (2013), who also use conditional control to achieve FCR control in a specialized setting. Using
a similar proof, we also show that using an adaptive Bonferroni rule, which adjusts the test’s level
based on the (random) number of intervals actually constructed, can achieve FWER control.
Proposition 11 (FCR and FWER Control via Selective Error Control). Assume Q is countable
with each q ∈ Q corresponding to a different parameter θq for the same model M . Let R(Y ) =
b )| with R(Y ) < ∞ a.s., and define V (Y ) as in (85).
|Q(Y
If each Cq enjoys coverage at level 1 − α given Sq = 1Aq (Y ), R(Y ) , then the collection of
b controls the FCR at level α:
intervals (Cq , q ∈ Q)
V
V
R ≥ 1 ≤ α.
(86)
E
≤ E
max(R, 1)
R
b controls the FWER at
If each Cq enjoys coverage at level 1 − α/R(Y ) given Sq , then (Cq , q ∈ Q)
level α:
P [V ≥ 1] ≤ α.
(87)
n
o
P
b ), θq (F ) ∈
Proof. Let Vq (Y ) = 1 q ∈ Q(Y
/ Cq (Y ) , so that V =
q∈Q Vq . If Cq has level-α
selective coverage, then for R ≥ 1, and for any F ∈ M ,
X
X
EF [V | R] =
EF [Vq | R] ≤
α EF 1Aq (Y ) | R = αR,
(88)
q∈Q
q∈Q
hence E [V /R | R] = α for each R ≥ 1.
We can repeat the argument when Cq has level-α/R selective coverage, we obtain EF [V | R] ≤
α. Marginalizing each bound over R(Y ) gives the result.
However, the converse of Proposition 11 is not true: FWER control does not in general guarantee control of relevant selective error rates. For example, suppose that we construct an interval
for the effect of red meat consumption on heart disease (Q(Y ) = 1) with probability 0.9 and for
the effect of statins on heart disease (Q(Y ) = 2) otherwise. If C1 and C2 have selective error rates
α1 = 0.02 and α2 = 0.3 respectively, the overall FWER is still controlled at α = 0.05.
Does our conservatism when asking about smoking compensate for our anti-conservatism when
asking about coffee? Perhaps not; those readers who are primarily interested in statins will be
consistently misled, and readers who are primarily interested in red meat consumption will be see
unnecessarily conservative intervals. As such, averaging our error rates across the two questions,
with two different interpretations, seems inappropriate.
31
More problematically, if the different questions correspond to different and non-overlapping
models — for example, if we examine residuals to decide between a Poisson log-linear model and a
negative-binomial model — then it is especially unintuitive to focus on error rates averaged across
the different choices of model.
By contrast, if the different questions represent a bag of relatively anonymous, a priori undifferentiated hypotheses which we are prioritizing for follow-up research, such as in a genome-wise
association study, then an error rate like the FDR is likely a better proxy for our scientific goals.
9
Discussion
Selective inference concerns the properties of inference carried out after using a data-dependent
procedure to select which questions to ask. We can recover the same long-run frequency properties
among answers to selected questions that we would obtain in the classical non-adaptive setting, if
we follow the guiding principle of selective error control:
The answer must be valid, given that the question was asked.
Happily, living up to this principle can be a simple matter in exponential family models including linear regression, due to the rich classical theory of optimal testing in exponential family
models. Even if we are possibly selecting from a large menu of diverse and incompatible models,
we can still design tests one model at a time and control the selective error using the test designed
for the selected model. We generally pay a price for conditioning, so it is desirable to condition on
as little as possible. Data carving can dramatically improve on data splitting by using the leftover
information in Y1 , the data set initially designated for selection.
Many challenges remain. Deriving the cutoffs for sample carving tests can be computationally
difficult in general. In addition, the entire development of this article takes the model selection
b as given, when in reality we can choose Q.
b More work is needed to learn what model
procedure Q
selection procedures lead to favorable second-stage properties.
As data sets and research questions become more and more complex, we have less and less
hope of specifying adequate statistical models ahead of time. As such, a key challenge of complex
research is to balance the goal of choosing a realistic model against the goal of inference once we
have chosen it. We hope that the ideas in this article represent a step in the right direction.
Reproducibility
A git repository with code to generate the figures for this file is available at the first author’s
website.
Acknowledgements
William Fithian was supported by National Science Foundation VIGRE grant DMS-0502385
and the Gerald J. Lieberman Fellowship. Dennis Sun was supported in part by the Stanford
Genome Training Program (NIH/NHGRI T32 HG000044) and the Ric Weiland Graduate Fellowship. Jonathan Taylor was supported in part by National Science Foundation grant DMS-1208857
and Air Force Office of Sponsored Research grant 113039. We would like to thank Stefan Wager,
Trevor Hastie, Rob Tibshirani, Brad Efron, Yoav Benjamini, Larry Brown, Maxwell Grazier G’sell,
Subhabrata Sen, and Yuval Benjamini for helpful discussions.
32
References
Rina Foygel Barber and Emmanuel J Candès. Controlling the false discovery rate via knockoffs.
The Annals of Statistics, 43(5):2055–2085, 2015.
GA Barnard. Discussion of professor bartlett’s paper. Journal of the Royal Statistical Society,
1963.
Alexandre Belloni, Victor Chernozhukov, and Christian Hansen. Inference for high-dimensional
sparse econometric models. arXiv preprint arXiv:1201.0220, 2011.
Alexandre Belloni, Victor Chernozhukov, and Christian Hansen. Inference on treatment effects
after selection among high-dimensional controls. The Review of Economic Studies, 81(2):608–
650, 2014.
Yoav Benjamini. Simultaneous and selective inference: current successes and future challenges.
Biometrical Journal, 52(6):708–721, 2010.
Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: a practical and powerful
approach to multiple testing. Journal of the Royal Statistical Society. Series B (Methodological),
pages 289–300, 1995.
Yoav Benjamini and Daniel Yekutieli. False discovery rate–adjusted multiple confidence intervals
for selected parameters. Journal of the American Statistical Association, 100(469):71–81, 2005.
James O Berger, Lawrence D Brown, and Robert L Wolpert. A unified conditional frequentist and
bayesian test for fixed and sequential simple hypothesis testing. The Annals of Statistics, pages
1787–1807, 1994.
Richard Berk, Lawrence Brown, Andreas Buja, Kai Zhang, and Linda Zhao. Valid post-selection
inference. The Annals of Statistics, 41(2):802–837, 2013.
Julian Besag. Markov chain monte carlo for statistical inference. Center for Statistics and the
Social Sciences, 2001.
Julian Besag and Peter Clifford. Generalized monte carlo significance tests. Biometrika, 76(4):
633–642, 1989.
Lawrence D Brown. A contribution to kiefer’s theory of conditional confidence procedures. The
Annals of Statistics, pages 59–71, 1978.
Lawrence D Brown. Fundamentals of statistical exponential families with applications in statistical
decision theory. Lecture Notes-monograph series, pages i–279, 1986.
C Brownie and J Kiefer. The ideas of conditional confidence in the simplest setting. Communications in Statistics-Theory and Methods, 6(8):691–751, 1977.
Arthur Cohen and Harold B Sackrowitz. Two stage conditionally unbiased estimators of the
selected mean. Statistics & Probability Letters, 8(3):273–278, 1989.
DR Cox. A note on data-splitting for the evaluation of significance levels. Biometrika, 62(2):
441–444, 1975.
AP Dawid. Selection paradoxes of bayesian inference. Lecture Notes-Monograph Series, pages
211–220, 1994.
33
Ruben Dezeure, Peter Bühlmann, Lukas Meier, Nicolai Meinshausen, et al. High-dimensional
inference: Confidence intervals, p-values and r-software hdi. Statistical Science, 30(4):533–558,
2015.
Bradley Efron. Tweedies formula and selection bias. Journal of the American Statistical Association, 106(496):1602–1614, 2011.
Bradley Efron, Robert Tibshirani, et al. Using specially designed exponential families for density
estimation. The Annals of Statistics, 24(6):2431–2461, 1996.
William Fithian, Jonathan Taylor, Robert Tibshirani, and Ryan Tibshirani. Adaptive sequential
model selection. arXiv preprint arXiv:1512.02565, 2015.
Jonathan J Forster, John W McDonald, and Peter WF Smith. Monte carlo exact conditional tests
for log-linear and logistic models. Journal of the Royal Statistical Society. Series B (Methodological), pages 445–453, 1996.
Annie Franco, Neil Malhotra, and Gabor Simonovits. Publication bias in the social sciences:
unlocking the file drawer. Science, 2014.
Andrew Gelman and Eric Loken. The garden of forking paths: Why multiple comparisons can be a
problem, even when there is no “fishing expedition” or “p-hacking” and the research hypothesis
was posited ahead of time. Downloaded January, 30:2014, 2013.
Naftali Harris. Visualizing lasso polytope geometry, June 2014. URL http://www.naftaliharris.
com/blog/lasso-polytope-geometry/.
Trevor Hastie, Robert Tibshirani, Jerome Friedman, T Hastie, J Friedman, and R Tibshirani. The
elements of statistical learning, volume 2. Springer, 2009.
Larry V Hedges. Estimation of effect size under nonrandom sampling: The effects of censoring
studies yielding statistically insignificant mean differences. Journal of Educational and Behavioral Statistics, 9(1):61–85, 1984.
Larry V Hedges. Modeling publication selection effects in meta-analysis. Statistical Science, pages
246–255, 1992.
Kenneth Hung and William Fithian. Rank verification for exponential families. arXiv preprint
arXiv:1610.03944, 2016.
John PA Ioannidis. Why most published research findings are false. PLoS medicine, 2(8):e124,
2005.
Adel Javanmard and Andrea Montanari. Hypothesis testing in high-dimensional regression under
the gaussian random design model: Asymptotic theory. IEEE Transactions on Information
Theory, 60(10):6522–6554, 2014.
Karl-Heinz Jockel. Finite sample properties and asymptotic efficiency of monte carlo tests. The
annals of Statistics, pages 336–347, 1986.
George Johnson. New truths that only one can see. The New York Times, 2014.
Jack Kiefer. Admissibility of conditional confidence procedures. The Annals of Statistics, pages
836–865, 1976.
Jack Kiefer. Conditional confidence statements and confidence estimators. Journal of the American
Statistical Association, 72(360a):789–808, 1977.
34
Jason D Lee and Jonathan E Taylor. Exact post model selection inference for marginal screening.
In Advances in Neural Information Processing Systems, pages 136–144, 2014.
Jason D Lee, Dennis L Sun, Yuekai Sun, and Jonathan E Taylor. Exact post-selection inference,
with application to the lasso. The Annals of Statistics, 44(3):907–927, 2016.
Hannes Leeb and Benedikt M Pötscher. Model selection and inference: Facts and fiction. Econometric Theory, 21(01):21–59, 2005.
Hannes Leeb and Benedikt M Pötscher. Can one estimate the conditional distribution of postmodel-selection estimators? The Annals of Statistics, pages 2554–2591, 2006.
Hannes Leeb and Benedikt M Pötscher. Can one estimate the unconditional distribution of postmodel-selection estimators? Econometric Theory, 24(02):338–376, 2008.
EL Lehmann and Joseph P Romano. Testing statistical hypotheses. New York:. Springer, 2005.
EL Lehmann and Henry Scheffé. Completeness, similar regions, and unbiased estimation: Part ii.
Sankhyā: The Indian Journal of Statistics (1933-1960), 15(3):219–236, 1955.
Richard Lockhart, Jonathan Taylor, Ryan J Tibshirani, and Robert Tibshirani. A significance test
for the lasso (with discussion). The Annals of Statistics, 42(2):413–468, 2014.
Joshua R Loftus and Jonathan E Taylor. A significance test for forward stepwise model selection.
arXiv preprint arXiv:1405.3920, 2014.
Ted K Matthes and Donald R Truax. Tests of composite hypotheses for the multivariate exponential
family. The Annals of Mathematical Statistics, pages 681–697, 1967.
Cyrus R Mehta, Nitin R Patel, and Pralay Senchaudhuri. Efficient monte carlo methods for
conditional logistic regression. Journal of The American Statistical Association, 95(449):99–108,
2000.
Nicolai Meinshausen, Lukas Meier, and Peter Bühlmann. P-values for high-dimensional regression.
Journal of the American Statistical Association, 104(488), 2009.
Sahand N. Negahban, Pradeep Ravikumar, Martin J. Wainwright, and Bin Yu. A unified framework for high-dimensional analysis of MM-estimators with decomposable regularizers. Statistical Science, 27(4):538–557, November 2012. ISSN 0883-4237. doi: 10.1214/12-STS400. URL
http://projecteuclid.org/euclid.ss/1356098555.
Richard A Olshen. The conditional level of the ftest. Journal of the American Statistical Association, 68(343):692–698, 1973.
Terence Orchard, Max A Woodbury, et al. A missing information principle: theory and applications. In Proceedings of the 6th Berkeley Symposium on mathematical statistics and probability,
volume 1, pages 697–715. University of California Press Berkeley, CA, 1972.
Ari Pakman and Liam Paninski. Exact hamiltonian monte carlo for truncated multivariate gaussians. Journal of Computational and Graphical Statistics, 23(2):518–542, 2014.
Camilo Rivera and Guenther Walther. Optimal detection of a jump in the intensity of a poisson
process or in a density with likelihood ratio statistics. Scandinavian Journal of Statistics, 40(4):
752–769, 2013.
JD Rosenblatt and Yoav Benjamini. Selective correlations; not voodoo. NeuroImage, 103:401–410,
2014.
35
Allan R Sampson and Michael W Sill. Drop-the-losers design: Normal case. Biometrical Journal,
47(3):257–268, 2005.
Michael W Sill and Allan R Sampson. Drop-the-losers design: Binomial case. Computational
statistics & data analysis, 53(3):586–595, 2009.
Robert Sladek, Ghislain Rocheleau, Johan Rung, Christian Dina, Lishuang Shen, David Serre,
Philippe Boutin, Daniel Vincent, Alexandre Belisle, Samy Hadjadj, et al. A genome-wide association study identifies novel risk loci for type 2 diabetes. Nature, 445(7130):881–885, 2007.
Jonathan Taylor and Robert Tibshirani. Post-selection inference for l1-penalized likelihood models.
arXiv preprint arXiv:1602.07358, 2016.
Jonathan Taylor and Robert J Tibshirani. Statistical learning and selective inference. Proceedings
of the National Academy of Sciences, 112(25):7629–7634, 2015.
Xiaoying Tian and Jonathan Taylor. Asymptotics of selective inference. Scandinavian Journal of
Statistics, 2017.
Xiaoying Tian and Jonathan E Taylor. Selective inference with a randomized response. arXiv
preprint arXiv:1507.06739, 2015.
Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical
Society. Series B (Methodological), pages 267–288, 1996.
Ryan J Tibshirani, Jonathan Taylor, Richard Lockhart, and Robert Tibshirani. Exact postselection inference for sequential regression procedures. arXiv preprint arXiv:1401.3889, 2014.
Ryan J Tibshirani, Alessandro Rinaldo, Robert Tibshirani, and Larry Wasserman. Uniform asymptotic inference and the bootstrap after model selection. arXiv preprint arXiv:1506.06266, 2015.
Sara Van de Geer, Peter Bühlmann, Yaacov Ritov, Ruben Dezeure, et al. On asymptotically
optimal confidence regions and tests for high-dimensional models. The Annals of Statistics, 42
(3):1166–1202, 2014.
Larry Wasserman and Kathryn Roeder. High dimensional variable selection. Annals of statistics,
37(5A):2178, 2009.
Asaf Weinstein, William Fithian, and Yoav Benjamini. Selection adjusted confidence intervals with
more power to determine the sign. Journal of the American Statistical Association, 108(501):
165–176, 2013.
Daniel Yekutieli. Adjusted bayesian inference for selected parameters. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 74(3):515–541, 2012.
Ed Yong. Replication studies: Bad copy. Nature, 485(7398):298–300, 2012.
Cun-Hui Zhang and Stephanie S Zhang. Confidence intervals for low dimensional parameters in
high dimensional linear models. Journal of the Royal Statistical Society: Series B (Statistical
Methodology), 76(1):217–242, 2014.
Hua Zhong and Ross L Prentice. Bias-reduced estimators and confidence intervals for odds ratios
in genome-wide association studies. Biostatistics, 9(4):621–634, 2008.
Sebastian Zöllner and Jonathan K Pritchard. Overcoming the winners curse: estimating penetrance
parameters from case-control data. The American Journal of Human Genetics, 80(4):605–615,
2007.
36
A
Proof of Proposition 1
Proof. For group i, let Ri be the number of true nulls selected, i.e.,
n
o
bi (Yi ), Fi ∈ H0 ⊆ M ,
Ri = (M, H0 ) : (M, H0 ) ∈ Q
Pn
Pn
and let Vi denote the number of false rejections. If ZnV = i=1 Vi and ZnR = i=1 Ri , then we
need to show lim supn→∞ ZnV /ZnR ≤ α.
By design, 0 ≤ Vi ≤ Ri and E(Vi ) ≤ α E(Ri ). As a result, E[ZnV ]/E[ZnR ] ≤ α for every n, so we
just need to show that the two sums are not far from their expectations. Because
∞
X
Var(Ri )
i=1
i2
≤B
∞
X
1
< ∞,
2
i
i=1
we can apply Kolmogorov’s strong law of large numbers to the independent but non-identical
sequence R1 , R2 , . . . to obtain
1 R
a.s.
(Z − EZnR ) −−→ 0,
n n
so
δ
ZnR
a.s.
− 1 ≤ (ZnR − EZnR ) −−→ 0.
EZnR
n
As for ZnV , we have
1 V
a.s.
(Z − EZnV ) −−→ 0,
n n
so
ZnV
δ
a.s.
− α ≤ (ZnV − EZnV ) −−→ 0;
EZnR
n
a.s.
a.s.
in other words, ZnR /EZnR −−→ 1 and lim supn ZnV /EZnR ≤ α.
B
Monte Carlo Tests and Confidence Intervals: Details
Assume Z arises from a one-parameter exponential family
Z ∼ gθ (z) = eθz−ψ(θ) g0 (z).
(89)
We wish to compute (by Monte Carlo) the UMPU two-sided rejection region for the hypothesis
H0 : θ = θ0 . Let U ∼ Unif[0, 1] be an auxiliary randomization variable.
Define the dictionary ordering on [0, 1]:
(z1 , u1 ) ≺ (z2 , u2 ) ⇐⇒ z1 < z2 or (z1 = z2 and u1 < u2 ).
(90)
If Γ1 = (c1 , γ1 ) and Γ2 = (c2 , 1 − γ2 ), then the region
RΓ1 ,Γ2 = {(z, u) : (z, u) ≺ Γ1 or (z, u) Γ2 }
(91)
implements the rejection region for the test with cutoffs c1 , c2 and boundary randomization parameters γ1 , γ2 .
For Γ1 ≺ Γ2 , write
K1 (Γ1 , Γ2 ; θ) = Pθ (RΓ1 ,Γ2 ) − α
K2 (Γ1 , Γ2 ; θ) = Eθ (Z | (Z, U ) ∈
(92)
RΓC1 ,Γ2 )
− Eθ (Z),
(93)
so that the correct cutoffs Γi are those for which K1 (Γ1 , Γ2 ; θ) = K2 (Γ1 , Γ2 ; θ) = 0. For fixed θ,
K1 is decreasing in Γ1 and increasing in Γ2 , while K2 is increasing in both Γ1 and Γ2 .
37
Let (Z1 , W1 ), (Z2 , W2 ), . . . be a sequence of random variables for which
Pn
W h(Zi )eθZi
n
b
Pn i
Eθ h(Z) = i=1
θZi
i=1 Wi e
a.s.
−−→ Eθ h(Z).
(94)
(95)
for all integrable h. This would be true if (Zi , Wi ) are a valid i.i.d. sample or i.i.d. importance
sample from g0 , or if they come from a valid Markov Chain Monte Carlo algorithm.
b n are defined analogously to Ki for i = 1, 2, with Eθ and Pθ replaced with their importanceIf K
i
a.s.
b n and P
bn , then K
bn −
b n satisfy
weighted empirical versions E
−→ K pointwise as n → ∞, and K
i
i
θ
θ
the same monotonicity properties almost surely for each n. As a result, we have almost sure
b n, K
b n ):
convergence on compacta for (K
1
2
sup
(Γ1 ,Γ2 )∈G
b n (Γ1 , Γ2 ; θ) − Ki (Γ1 , Γ2 ; θ)
max K
i
i
(96)
2
for each θ, for compact G ∈ (R × [0, 1]) .
b n and K
b n , in effect defining the
We carry out our tests by solving for Γ1 and Γ2 which solve K
1
2
UMPU tests for a one-parameter exponential family through the approximating empirical measure.
Specifically, we can define
n
o
b 2 (Γ1 ; θ) = inf Γ2 : K
b n (Γ1 , Γ2 ; θ) = 0 ,
Γ
(97)
1
b 2 = ∞ if the set is empty. That is, for a given lower cutoff we define the upper cutoff to
with Γ
b n Γ1 , Γ
b 2 (Γ1 ; θ); θ is an increasing
obtain a level-α acceptance region if that is possible. Then, K
2
b
function and we can solve it using binary search. Let Rθ denote
the rejection region
so obtained.
bθ if and only if K
b n (z, u), Γ
b 2 ((z, u)); θ < 0. This fact,
Note that (z, u) is in the left-tail of R
2
paired with an analogous test for whether (z, u) is in the right tail, gives us a quick way to carry
out the test. It also allows us to quickly find the upper and lower confidence bounds for the
approximating empirical family, via binary search.
C
Sampling for the Selective t-Test: Details
Let C ⊆ Rk denote a set with nonempty interior and consider the problem of integrating some
integrable function h(y) against the uniform probability measure on C ∩ S k−1 , where S k−1 is the
unit sphere of dimension k − 1, assuming the intersection is non-empty. Assume we are given an
i.i.d. sequence of uniform samples Y1 , Y2 , . . . from C ∩ B k , where B k is the unit ball.
k−1
Let R ∼ r k , so that if Z ∼ Unif(S k−1 ), then Y = RZ ∼ Unif(B k ). Let
Z
W (Z) =
0
1
rk−1
1{rZ ∈ C}
dr
k
−1
(98)
We can use the Yi for which Zi = Yi /kYi k ∈ C as a sequence of importance samples with
38
weights W (Zi ), since
E(h (Z)1 {Y, Z ∈ C} W (Z))
Z
Z 1
rk−1
=
h(z)1 {z, rz ∈ C} W (z)
dr dz
k
S k−1 0
Z
=
h(z)1 {z ∈ C} dz
(99)
(100)
(101)
S k−1
= E (h(Z)1 {Z ∈ C}) .
To carry out the selective t-test of H0 : βj = 0, we need to sample from
L η0 Y
PXM \j Y, kY k, A .
(102)
(103)
⊥
Let U = PXM \j Y , and let Q ∈ Rn×(n−|M |−1) be such that QQ0 = PX
. Then L2 , kQ0 Y k2 =
M \j
2
2
kY k − kU k is fixed under the selection event. Let
C = {v : U + Qv ∈ A},
(104)
so that AU = U + QC, an (n − |M | − 1)-dimensional hyperplane intersected with A, is the event
we would sample from for the selective z-test.
Under H0 , Y is uniformly distributed on
(U + QC) ∩ kY kS n−1 = U + Q C ∩ LS n−|M |−2 .
(105)
Assume we can resample Y ∗ uniformly from AU ∩ (U + LB n−|M |−1 ), which is just sampling
from AU with an additional quadratic constraint. Then V ∗ = Q0 (Y ∗ − U ) is a sample from the
ball of radius L, intersected with C. We can turn V ∗ into an importance-weighted sample from
the sphere via the scheme outlined above; then, the same importance weight suffices to turn Y ∗
into a sample from the selective t-test conditioning set.
39
| 10math.ST
|
Population Synthesis via k -Nearest Neighbor
Crossover Kernel
Fujitsu Laboratories Ltd.
Kanagawa, Japan
Email: {hamada-naoki, km.homma, h-higuchi, h.kikuchi}@jp.fujitsu.com
Abstract—The recent development of multi-agent simulations
brings about a need for population synthesis. It is a task of
reconstructing the entire population from a sampling survey of
limited size (1% or so), supplying the initial conditions from
which simulations begin. This paper presents a new kernel density
estimator for this task. Our method is an analogue of the classical
Breiman-Meisel-Purcell estimator, but employs novel techniques
that harness the huge degree of freedom which is required
to model high-dimensional nonlinearly correlated datasets: the
crossover kernel, the k-nearest neighbor restriction of the kernel
construction set and the bagging of kernels. The performance
as a statistical estimator is examined through real and synthetic
datasets. We provide an “optimization-free” parameter selection
rule for our method, a theory of how our method works and
a computational cost analysis. To demonstrate the usefulness as
a population synthesizer, our method is applied to a household
synthesis task for an urban micro-simulator.
I.
I NTRODUCTION
In recent years, increasing computational power enables
us to conduct large-scale multi-agent simulations in highly
public subjects, e.g., urban planning [1], transportation [2],
energy management [3], disaster prevention [4] and welfare
engineering [5]. To carry out multi-agent simulations in such
highly public subjects, we face difficulties in collecting detailed survey data that supply realistic initial conditions to simulators. For example, modern urban micro-simulators require
disaggregate socio-demographics of the study area including
each and every household’s residential place, family structure,
car ownership and income as well as person’s age, gender,
job, daily activities, etc. The complete survey of such massive
and detailed information is usually impracticable for cost and
privacy reasons.
If only a sampling survey is available, we have to recover the entire population from the obtained sample. For
this purpose, Iterative Proportional Fitting (IPF) [6] and its
extensions such as Iterative Proportional Updating (IPU) [7]
are widely used in the aforementioned areas. However, the IPFlike approach that simply weights, or copies, the sample points
to synthesize the population cannot reproduce the diversity of
the original population which might be missed through the
sampling survey. Moreover, since this approach only accepts
categorical variables, numerical variables such as age and
income must be roughly discretized (usually into two to five).
One promising alternative is the statistical approach that
first estimates the probability distribution of sample points and
then resamples the population from the estimated distribution,
though it is not easy to model complicated distributions of data
$900,000
$800,000
$700,000
$600,000
Annual income
arXiv:1508.06483v1 [cs.NE] 26 Aug 2015
Naoki Hamada, Katsumi Homma, Hiroyuki Higuchi and Hideyuki Kikuchi
$500,000
$400,000
$300,000
$200,000
$100,000
$0
$-100,000
0
5
10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Age
Fig. 1. Age vs. annual income of 1% Washington state citizens from the
American Community Survey 2012 [8]. Note that values at very high ages or
incomes are “top-coded” in this survey in order to protect the confidentiality
of survey respondents.
for multi-agent simulations. Figure 1 shows the scatter plot of
person’s age and income in a real city, which are commonly
used attributes of person agents in simulations. We can see that
even this very simplified example defies parametric models.
The distribution has an income-degenerate part for under 15
non-working ages while high income earners in middle ages
constitute a one-sided long tail. Furthermore, some artifact
lines arise due to “top-coding”. Our problem is much harder.
The data often consist of 5 to 20 variables and these variables
are nonlinearly correlated all together. Parametric modeling is
a formidable task even for domain experts.
The kernel density estimation has been shown to possess
excellent flexibility to fit various complex distributions in
low-dimensional spaces [9]. “Fitting” in this method is to
select the bandwidth parameters of kernel functions so as to
minimize some goodness-of-fit measure. In high dimensions,
this optimization-intensive fitting strategy becomes ineffective.
Each sample point requires a different bandwidth matrix for
representing a local correlation which varies with locations in
a nonlinear manner. Due to the large number of parameters
in the model, the fitting procedure involves a large-scale
nonlinear optimization that requires huge computational efforts
and limits the applicable dimensionality and shape of the data
distribution.
Recently, a new kind of kernels for density estimation, the
crossover kernel, has been proposed [10], which employs a
crossover operator in genetic algorithm as a kernel function
in kernel density estimation. The crossover kernel does not
require bandwidth parameters to specify the shape of the
function; instead, it uses a subset of sample points called
the kernel construction set (KCS). Hence it has a potential
to obviate costly fitting procedures and to quickly resample
high-dimensional data. Unfortunately, there are merely two
studies on how to choose KCSs. One appeared in Sakuma
& Kobayashi’s original work: they proposed an EM algorithm
to calculate sample’s weights that describe the probability of
which sample point belongs to which KCS. Another is Kimura
& Matsumura’s [11]: they pointed out that the above method
does not fit non-Gaussian-mixture-like distributions no matter
how their weights are optimized, and proposed optimizing the
choice of KCSs itself rather than the weights. Both methods
rely on optimization, and thus involve heavy fitting procedures
again.
The aim of this paper is to develop an “optimizationfree” crossover kernel for fast and accurate resampling by
introducing the k-nearest neighbor restriction of KCS choice.
We will show that this simple idea works surprisingly well
for our problem and has a certain theoretical background. Our
contributions are summarized as follows.
•
We proposed the k-nearest neighbor restriction of
KCS choice, and showed that our method is not only
faster, but also more accurate than conventional Gaussian kernels and Kimura-Matsumura crossover kernel
through complicated datasets of 2 to 17 dimensions.
•
We examined the parameter sensitivity of our method,
and gave a rule of thumb for choosing the neighborhood size k and the KCS size m without optimization.
•
We found that setting m ≤ k decreases the generalization error in experiments, and showed its rationale
from the viewpoint of bagging [12].
•
We demonstrated that the simulation accuracy of UrbanSim [1] can be enhanced if our method supplies
initial households instead of IPU.
II.
P ROBLEM D EFINITION , N OTATIONS AND
O RGANIZATION OF THE PAPER
Throughout this paper,
the following assumptions are
made. A sample X = x1 , . . . , xn ∈ Rd of size n(∈ N)
is given, which is taken from an unknown population. The
population is a set of l(∈ N) i.i.d. points with an unknown
continuous density f (x). We are going to estimate f , say fˆ,
and resample l points from fˆ to recover the population. The
problem of population synthesis will be discussed under two
different situations below.
Case I – an unbiased sample: The sample X is uniformly drawn from the population. The goal is thus to estimate
the density of X which is identical to f . The subsequent three
sections of this paper discuss this case. Section III reviews
existing kernel density estimators and crossover kernels. Section IV presents our proposal and its properties. Section V
shows the effectiveness of proposed method by comparing
to conventional kernel density estimators on some benchmark
datasets. Mechanisms behind the method are discussed.
Case II – a biased sample with marginal frequencies:
In real-world applications, samples available are often biased
rather than uniform. This case requires extra information about
what bias is introduced. We assume the population’s binned
frequencies F = { fb ∈ N | b ∈ B } is given, but the binning B
is applied to marginal distributions of variables and incomplete
to recover the entire joint distribution. Thus, what we have
to do in this case is to correct the sampling bias with the
frequencies F and to extract the variable correlation from the
sample X, combining them to estimate the entire joint density
f . In Section VI, we will tackle an application of this type and
will present a bias correction technique. Incorporated with our
resampling method developed in Section IV, this technique is
applied to a task of generating initial population of households
for an urban land-use micro-simulation.
III.
C ONVENTIONAL M ETHODS
Now focusing on Case I, this section reviews existing
kernel density estimators and crossover kernels.
A. Kernel Density Estimation
Kernel density estimators can be classified into two versions [13]: the sample point estimator and the balloon estimator. For resampling purpose, our estimate must be a density
and easy to sample. The sample point estimator satisfies both
requirements if the kernel is a density which is easy to sample,
while the balloon estimator is usually not a density even when
the kernel is [14]. Thus we concentrate on the former of the
following form:
n
1X
fˆ(y) =
K(y|xi , H i )
n i=1
(1)
where K is a kernel function to be specified and H i is a
bandwidth matrix that is a d × d symmetric positive-definite
matrix of parameters that we are estimating. In this paper, we
consider the multivariate Gaussian kernel:
1
K(y|xi , H i ) = N(y|xi , H i ) = p
(2π)d det(H i )
1
(y
−
x
)
. (2)
· exp − (y − xi )⊤ H −1
i
i
2
As a measure of goodness-of-fit, the mean integrated
squared error (MISE) is most commonly used1 and its
bias/variance decomposition is given by
Z
2
MISE = E
fˆ(y) − f (y) dy
Z
2
E[fˆ(y)] − f (y) dy
=
|
{z
}
(3)
squared bias
Z
2
dy.
+ E fˆ(y) − E[fˆ(y)]
{z
}
|
variance
1 Various measures based on the L -errors, the Kullback-Leibler divergence
1
and the Hellinger distance are used for this purpose [15]. In any case, the
problem that we will point out in this section still arises.
In the case of fixed matrix bandwidths (H 1 = · · · = H n ),
bandwidth selectors using cross-validation [16] and plugin [17] have been developed. However, when it comes to
variable bandwidths, existing techniques are very restrictive.
The Breiman-Meisel-Purcell (BMP) estimator [18] is the most
classical one, which uses scalar bandwidths, i.e., H i = h2i I,
and determines hi as a constant multiple of the Euclidean
distance δik from xi to the kth nearest other sample point:
hi = hδik
for fixed h > 0.
(4)
Asymptotically, this is equivalent to choosing hi ∝ f (xi )−1/d
as n → ∞ [14]. Abramson [19] proposed the square root law,
i.e., choosing hi ∝ f (xi )−1/2 regardless of d, and showed
that it reduces bias and accelerates the rate of convergence. To
handle matrix bandwidths, Sain [9] studied a binning technique
which decreases the number of parameters by sharing the same
bandwidth matrix with sample points in the same bin. His
simulation results clarified that even for a low-dimensional
(d = 2) moderate sample size (n = 320), the unbiased crossvalidation tends to yield too small, namely, overfit bandwidth
matrices. Since the (unbinned) variable bandwidth matrix has
nd(d + 1)/2 parameters, the situation is much worse in higher
dimensions. Developing bandwidth selectors for this case is
currently an open question.
Another problem in high dimensions is that the relative
contribution of variance dominates bias in MISE. Sain [9]
pointed out that for the fixed scalar bandwidth, the ratio of
order of bias to order of variance is 4 : d. Therefore, as
dimensionality increases, bias reduction techniques get less
importance and variance reduction techniques including the
bagging [12] become the core of the estimator.
B. Crossover Kernels
In the genetic algorithm literature, Kita et al. [20] argued
the desirable behavior of crossover operators and suggested
a guideline, the preservation of statistics, which tells us that
the crossover’s parents and children should have the same
mean and covariance. Specifically, suppose that the parents
and children are realizations of random variate x and y in Rd ,
respectively, and then the following should be satisfied:
E[x] = E[y],
C[x] = C[y].
(5)
Sakuma & Kobayashi [10] pointed out that the parents
x1 , . . . , xm (realizations of x) implicitly determine the probability density function K of y and thus K can be considered
as a data-driven kernel function. In this context, the set of
parents X ′ = { x1 , . . . , xm } is called the kernel construction
set (KCS). If one chooses a crossover so that it satisfies (5) and
K is the Gaussian distribution, then the crossover coincides
with the maximum likelihood estimator (MLE) of K calculated
from the KCS. Sakuma & Kobayashi also noticed that UNDXm [20] fulfills these conditions when the KCS size is set to
be m = d + 1.
The remaining question is how to construct the entire density estimate from crossover kernels. Sakuma &
Kobayashi [10] just used the Gaussian mixture model fˆ(y) =
P
k
i=1 αi N(y|µi , Σi ) as the final estimate, but its parameters
αi , µi , Σi are optimized by their modified EM algorithm
in which each Gaussian component at the tth EM step
(t)
(t)
N(y|µi , Σi ) is replaced by the average of crossover kernels
L
1X
(t)
(t)
(t)
e
N(y|X
K(y|Xil ).
i1 , . . . , XiL ) =
L
(6)
l=1
(t)
Their KCSs Xil are chosen at random from X by the
probability proportional to weights w(t) (i|xj ) which are calculated by their E step and describe the possibility that the
ith component generates the jth sample point. The superiority
of their EM algorithm over the original one has been shown
through simulation studies.
Kimura & Matsumura [11] proposed a more direct way,
giving the entire estimate by
L
1X
fˆ(y) =
K(y|Xl ).
L
(7)
l=1
Their
Pn KCSsˆ Xl are chosen so as to maximize the log-likelihood
i=1 log f (xi ): starting from a random choice, KCSs are
randomly remade (one at a time) if it improves the loglikelihood. They showed the high flexibility of this approach
by applying it to a circularly distributed dataset.
IV.
k-N EAREST N EIGHBOR C ROSSOVER K ERNEL
As we have seen, the BMP estimator is fast but restricted to
the scalar bandwidth, whereas the crossover kernel is slow but
able to handle the covariance matrix. Our idea is to combine
the strengths of both estimators.
A. REX Kernel
As mentioned in Section III-B, UNDX-m coincides with
the Gaussian MLE only when the KCS size is m = d + 1.
In order to extend this property to arbitrary size, we employ
REX [21] for the kernel function. REX generates a resampling
point y from the KCS X ′ = { x1 , . . . , xm } as follows:
m
X
1
εi (xi − µX ′ ), εi ∼ N(0, )
y = µX ′ +
(8)
m
i=1
Pm
1
where µX ′ = m
i=1 xi and εi is a random number drawn
1
).
from the univariate Gaussian distribution N(0, m
Following Sakuma & Kobayashi’s [10] derivation of the
kernel function of UNDX-m, we get the probability density
function of y, namely, the REX kernel:
1
K(y|X ′ ) = N(y|µX ′ , ΣX ′ ) = p
d
(2π) det(ΣX ′ )
1
⊤ −1
(9)
· exp − (y − µX ′ ) ΣX ′ (y − µX ′ )
2
Pm
1
⊤
where ΣX ′ = m
i=1 (xi − µX ′ ) ⊗ (xi − µX ′ ) . Regardless
of m, thus always K(y|X ′ ) is the Gaussian MLE calculated
from X ′ . Comparing the REX kernel (9) to the Gaussian
kernel (2), we see that the covariance matrix ΣX ′ in (9) is
corresponding to the bandwidth matrix H i in (2). Resampling
from the Gaussian kernel with explicit MLE costs O(md2 +d3 )
time;2 the REX kernel with resampler (8) bypasses the estima2 Since the random number generation relies on the Cholesky decomposition
of ΣX ′ , it requires O(md2 ) time to calculate ΣX ′ from X ′ and O(d3 ) time
the Cholesky decomposition of ΣX ′ .
Algorithm 1 k-NN REX Kernel
1.5
0.7
sample
proposed
BMP
1
0.6
0.5
0.5
f(x)
x2
0.4
0
0.3
-0.5
0.2
-1
-1.5
-1.5
0.1
0
-1
-0.5
0
x1
0.5
1
1.5
Fig. 2. An example of local correlation. Two typical Σi ’s of the proposed
method (10) are shown by solid ellipsoids. The corresponding hi ’s of the BMP
estimator (4) with h = 1 are drawn by dashed circles. The sample points (the
dots) and the underlying density (the gray scale) are of the ring-like PDF [11].
tion process and enables us to generate a sample point from
K(y|X ′ ) in O(md) time.
B. k-Nearest Neighbor Restriction of KCS Choice
In order to adapt KCSs to data without optimization, we
choose KCSs from the k-nearest neighbor (k-NN) of each
sample point. This is a simple yet powerful trick to realize
a crossover kernel analogue of the BMP estimator. It is shown
in Appendix A that the covariance matrix Σi of k-NN points
of a sample point xi is asymptotically of the form:
Σi ≈
4
2
∇f ⊗ ∇f ⊤
δik
δik
(xi )
I−
2
d+2
(d + 2)
f2
as n → ∞. (10)
For sufficiently large n, we can expect δik ≪ 1 and thus
2
the first O(δik
) term in the r.h.s. overwhelms the second
4
O(δik ) term. This √
means that (10) approximates to a scalar
bandwidth of √
δik / d + 2, that is, the BMP bandwidth (4)
with h = 1/ d + 2. Such an asymptotic behavior has a
strength and a weakness: one hand, our method inherits rich
asymptotics from the BMP estimator, including the normality,
the consistency, the leading term of bias/variance, the rate of
convergence, etc. [22]; on the other hand, our matrix extension
of bandwidth gives no advantage over the BMP estimator in
asymptopia.
For small-to-moderate n, the second term in the r.h.s. of
(10) can be influential when nonlinear correlation presents.
Notice that the matrix-valued function (∇f ⊗ ∇f ⊤ )/f 2 in
the second term captures the local correlation around xi ,
which is not taken into account in the BMP estimator. It is
easy to check that the matrix (∇f ⊗ ∇f ⊤ )(x) is always of
rank one unless the gradient ∇f (x) vanishes, and its unique
2
nonzero eigenvalue equals k∇f (x)k and the corresponding
eigenvector is parallel to ∇f (x). Dividing (∇f ⊗∇f ⊤ )(x) by
f 2 (x) can be thought of as a relativization of the effect. Thus,
subtracting such a matrix from the BMP bandwidth as in (10)
makes the bandwidth shrink in the gradient direction. Figure
2 illustrates how this effect captures the local correlation.
In this example, the gradient is everywhere normal to the
Input:
X = { x1 , . . . , xn } ⊂ Rd
⊲ Sample (training data)
k ∈ { 0, . . . , n }
⊲ Neighborhood size
m ∈ { 1, . . . , k + 1 }
⊲ KCS size
l∈N
⊲ Population size
Output:
Y = { y 1 , . . . , y l } ⊂ Rd
⊲ Synthetic population
1: for all x ∈ X do
2:
calculate the k-NN of x, say X(x, k)
3: end for
4: Y ← ∅
5: repeat
6:
draw x1 from X at random
7:
draw x2 , . . . , xm (xi 6= xj ) from X(x1 , k) at random
8:
generate y by (8)
9:
Y ← Y ∪ {y}
10: until |Y | = l
11: return Y
circumference of the circular density and becomes very large
near the circumference. Thus, the bandwidth shrinkage does
occur in the normal directions and does not in the tangent
directions with respect to the density manifold, resulting in a
better fit shown in Fig. 2. Accordingly, our matrix extension of
bandwidth will gain some improvement in finite sample size.
Another advantage of our method over the BMP estimator
in finite sample size is that the BMP bandwidth depends on
only the kth nearest sample point and the other k−1 points are
ignored while our bandwidth is determined by all the k points.
This will make the bandwidth computation more stable.
C. Bagging of Kernels
Even our k-NN MLE is stabler than BMP’s k-NN distance,
the severe variability of high-dimensional samples will still
fluctuate the resulting kernel. In order to exploit the sample
information for further variance reduction, we employ the
bagging [12] of kernels. That is, our KCS is not the k-NN
itself, but a subset of it, and we change the KCS every time
a resampling point is generated. The resulting entire estimate
becomes as follows:
n
1X
fˆ(y) =
EXi′ [K(y|Xi′ )]
n i=1
(11)
where EXi′ [ · ] denotes the expectation over all possible choices
of KCS Xi′ of size m from the k-NN of xi .
Of course, we cannot compute the value of fˆ(y) since
EXi′ [K(y|Xi′ )] is a mixture of too many Gaussians, but we
can easily draw a sample from it by simply choosing xi and
its k-NN points at random and passing them to (8). We end this
section by presenting the resampling algorithm via the k-NN
REX kernel in Algorithm 1.
V.
E XPERIMENTS
In order to evaluate the proposed method, we conduct
numerical experiments on benchmark datasets. The population
reproducibility, the parameter sensitivity, the bagging effect
and the CPU time are studied.
TABLE I.
Name
PUMS Person in WA
Swiss Roll
3D Road Network
Letter Recognition
D ESCRIPTION OF DATASETS .
Dimension
2
3
4
17
Sample Size
69301
10000
434874
20000
URL
[8]
[24]
[25]
[25]
A. Methods and Parameter Settings
k-NN REX kernel: Algorithm 1 was used as the proposed
method. Its parameters were set to be as follows: the neighborhood size k = 0, . . . , 99, and the KCS size m = 1, . . . , k + 1
for each k.
KM REX kernel: As a state-of-the-art crossover kernel,
we combined the Kimura-Matsumura (KM) model (7) with
the REX kernel (9). Its parameters were set to be as follows:
the number of KCSs L = 1, . . . , 100 and the KCS size m =
d + 1, . . . , d + 10. The optimization of KCS choice ran until
consecutive 10000 iterations do not increase the log-likelihood.
BMP Gaussian kernel: (1), (2) with H i = h2i I and
(4) was used to see our performance progress owing to the
bandwidth matrization and the bagging. Its parameters were
set to be as follows: the neighborhood size k = 1, . . . , 99 and
the constant multiplier h = 0.01, . . . , 1.00.
Fixed Gaussian kernel: (1) and (2) with H i = h2 I was
used as a conventional estimator.3 Its parameter, the bandwidth,
was set to be h = 0.001, . . . , 0.100.
B. Datasets
To examine the effect of dimensionality and sample size,
four datasets shown in Table I were used. Every dataset was
affinely transformed by the whitening [23] so that its mean
and covariance can be the zero vector and identity matrix,
respectively. All the datasets and their detailed descriptions are
available online; consult the references in the URL column of
the table.
C. Criterion
In order to evaluate the accuracy of population reproduction, we measured the Hellinger distance [26] between test
data and a synthesized population. Each dataset was evenly
divided into 100 subsets,4 and then used for the inverted crossvalidation (ICV): one subset was used as training data and the
remaining 99 subsets constituted test data. Since we cannot
compute the value of the k-NN REX kernel’s density (11)
as stated in Section IV-C, a binned version of the Hellinger
distance should be used. For a reconstructed population Y and
a set of test data Z, our performance measure can be written
by
s
2
p
1 X p
H(Y, Z) =
|Yb | / |Y | − |Zb | / |Z|
(12)
2
b∈B
where B is the set of bins, Yb (resp. Zb ) is a subset of Y
(resp. Z) whose elements fall in the bin b, and | · | is the
3 This is virtually (but, not exactly) identical to using a fixed matrix
bandwidth, i.e., H 1 = · · · = H n in (2), because our datasets are normalized
by an affine transformation.
4 For datasets that cannot be divided by 100, the remainders were randomly
dropped.
set cardinality. When H(Y, Z) < H(Y ′ , Z) holds, the method
that generates Y has a better accuracy than the method that
generates Y ′ .
For each setting, the 100-fold ICV was done once, and
then the average and standard deviation over its 100 Hellinger
distances were calculated.
D. Results
Table II shows the Hellinger distances from the test data
of 99n points to the population of 99n points synthesized by
each method with n training data. The Hellinger distances to
the training data themselves are also displayed as a baseline.
Note that their values are attainable by simply copying or
bootstrapping the training data.
Every method much better reproduced the population than
the baseline for all datasets, except the KM REX kernel for 3D
Road Network. In this sense, the kernel density estimation is,
more or less, generally preferable to copying and bootstrapping
approaches for the purpose of population synthesis, especially
when the sample is small and complex. Among others, the kNN REX kernel gained further significant improvements in all
cases. For visual comparison, we provide the scatter plots of
synthetic populations in Appendix B.
A surprising fact can be found in the comparison of
their standard deviations. The k-NN REX kernel’s standard
deviations were less than the KM REX kernel’s and as low
as the other three cases. Remember that our model is a
matrix bandwidth extension of the BMP model and uses a
lot more kernels than the KM model; thus it seems natural to
expect our standard deviations have the greatest values. That
paradoxical result implies the success of our variance reduction
techniques. It holds good even in Letter Recognition, i.e., a
high-dimensional (d = 17), small sample size (n = 200) case.
E. Discussions
1) Parameter Sensitivity: Figure 3 gives a survey of all
results of the k-NN REX kernel. The filled area in each
plot indicates the parameters whose results surpass the best
performance of the fixed Gaussian kernel. Every filled area
has a horizontally long part, which means the performance of
the k-NN REX kernel is insensitive to the neighborhood size
k. Our observation is consistent with the sensitivity analysis
for k of the BMP estimator studied by Breiman et al. [18].
The figure also clarifies that the choice of the KCS size m
is critical to the performance. How should practitioners find
a suitable m? For PUMS Person in WA (d = 2), the optimal
value is m = 3. This agrees with a common practice in genetic
algorithm: if the search space is of d dimensions, then the
crossover should use m = d + 1 parents [21]. For Swiss Roll
(d = 3), 3D Road Network (d = 4) and Letter Recognition
(d = 17), the optimal m’s are 3, 2 and 4, respectively.
It looks inconsistent, but once their intrinsic dimensions are
considered, the same rule can be found. Swiss Roll is a 2dimensional spiraling band distribution. In 3D Road Network,
the data are observed along roads, and the intrinsic dimension
is therefore 1. The dimensionality of Letter Recognition can
be reduced to about 3 by nonlinear dimensionality reduction
methods [27]. These remind us of a practice in multiobjective
TABLE II.
H ELLINGER DISTANCE ( AVERAGE ± STANDARD DEVIATION OVER 100- FOLD ICV) WITH BEST PARAMETER SETTINGS . I N EACH ROW, THE
BOLDED AVERAGE IS SMALLER THAN ANY OTHER ONE WITH STATISTICAL SIGNIFICANCE p < 0.01 BY W ELCH ’ S t- TEST.
k-NN
REX Kernel
2.14e-1 ± 6.82e-3
(k = 66, m = 3)
3.41e-1 ± 1.48e-2
(k = 12, m = 3)
1.95e-1 ± 1.01e-3
(k = 30, m = 2)
5.07e-1 ± 1.00e-2
(k = 36, m = 4)
Dataset
(d dimensions, n training points)
PUMS Person in WA
(d = 2, n = 693)
Swiss Roll
(d = 3, n = 100)
3D Road Network
(d = 4, n = 4348)
Letter Recognition
(d = 17, n = 200)
KM
REX Kernel
2.87e-1 ± 2.06e-2
(L = 35, m = 4)
3.89e-1 ± 2.63e-2
(L = 66, m = 6)
3.74e-1 ± 4.20e-3
(L = 96, m = 5)
5.64e-1 ± 1.42e-2
(L = 68, m = 20)
BMP
Gaussian Kernel
2.46e-1 ± 9.74e-3
(k = 2, h = 0.99)
3.74e-1 ± 1.32e-2
(k = 22, h = 0.12)
1.98e-1 ± 3.33e-3
(k = 3, h = 0.18)
5.45e-1 ± 1.12e-2
(k = 2, h = 0.12)
PUMS Person in WA (d = 2, n = 693)
Fixed
Gaussian Kernel
3.39e-1 ± 1.09e-2
(h = 0.011)
4.02e-1 ± 1.24e-2
(h = 0.056)
2.36e-1 ± 3.11e-3
(h = 0.009)
5.47e-1 ± 1.08e-2
(h = 0.059)
Training Data
(Baseline)
5.11e-1 ± 5.73e-3
6.65e-1 ± 1.34e-2
2.78e-1 ± 3.80e-3
7.00e-1 ± 1.26e-2
Swiss Roll (d = 3, n = 100)
100
100
0.55
0.65
H < 0.3386 (4938/5050)
H ≥ 0.3386 (112/5050)
H < 0.4020 (233/5050)
H ≥ 0.4020 (4817/5050)
90
90
0.5
0.6
80
0.45
0.4
50
0.35
40
30
KCS size m
60
0.55
70
Hellinger distance H
60
0.5
50
0.45
40
30
0.3
20
0.4
20
0.25
0.35
10
10
0
0.2
0
10
20
30
40
50
60
Neighborhood size k
70
80
90
0
100
0.3
0
10
20
3D Road Network (d = 4, n = 4348)
30
40
50
60
Neighborhood size k
70
80
90
100
Letter Recognition (d = 17, n = 200)
100
0.34
100
0.72
H < 0.2358 (496/5050)
H ≥ 0.2358 (4554/5050)
H < 0.5471 (2209/5050)
H ≥ 0.5471 (2841/5050)
90
0.7
90
0.32
80
0.68
80
0.3
0.66
70
50
0.26
40
0.24
30
0.64
KCS size m
0.28
60
Hellinger distance H
KCS size m
70
60
0.62
50
0.6
40
0.58
30
0.56
0.22
20
20
0.2
10
0
0.18
0
10
20
30
40
50
60
Neighborhood size k
70
80
90
100
Hellinger distance H
KCS size m
70
Hellinger distance H
80
0.54
10
0.52
0
0.5
0
10
20
30
40
50
60
Neighborhood size k
70
80
90
100
Fig. 3. Parameter (k, m) vs. Hellinger distance of the k-NN REX kernel. The Hellinger distance is averaged over 100-fold ICV. The legend “H <
α (β/5050) ” means as follows: Hellinger distances at filled squares are better than the best tuned Gaussian kernel’s Hellinger distance α, and there are β
out of 5050 parameter settings that achieve better distances. The legend “H ≥ α (β/5050) ” should be read similarly.
genetic algorithm: when the Pareto solution set forms a d′ dimensional submanifold in the d-dimensional search space,
using m = d′ + 1 parents exhibits a better performance than
m = d + 1 [28].
As a consequence, we suggest a parameter selection rule
that determines k and m by the following steps:
1)
2)
3)
Choose k between 10 and 50.
Identify the intrinsic dimensionality d′ of the dataset
by a priori knowledge or dimensionality reduction.
Set m = d′ + 1, or cross-validate near d′ if possible.
2) Effect of Bagging: Our bagging is different from
Breiman’s original proposal in some points. Given a sample
of size k, Breiman’s bagging makes some particular replicates
of size k by random sampling with replacement. Our bagging
aims to implicitly achieve similar effect by changing the KCS
every time, which corresponds to taking all possible replicates
of size m by random sampling without replacement.
Despite those differences, Fig. 3 shows that to get better
results, users should set k much greater than m. If it is the case,
many possibilities of KCS choice arise and the averaging effect
of our bagging would be enhanced. Therefore, we can gain
much variance reduction that will also decrease the Hellinger
Algorithm 2 k-NN REX Kernel with Bias Correction
distance.
3) Computational Cost: To analyze the time complexity
of the k-NN REX kernel, we examine Algorithm 1. Lines
1–3 compute the k-NN of n points of dimension d. We
implemented this by a naive O(dn2 ) time algorithm that
calculates the Euclidean distances among points. Lines 5–10
compute l points, each of which is generated by (8) in O(dm)
time. Therefore, the overall complexity is O(dn2 + dlm). In a
similar way, we also get the complexity of the BMP estimator
as O(dn2 + dl). The difference between them is m in the
second term, which is related to resampling. Usually m is
small as it should be the intrinsic dimensionality plus one.
Thus, the additional computation cost for our extensions will
not be so significant in most cases.
The actual computation time is shown in Table III. This
was measured by running a single threaded C++ program on a
server machine: Intel Xeon 3.07 GHz, 64-bit, 30 GB RAM. For
each method, the time includes reading the dataset from a file,
computing the density, synthesizing the population and writing
it to a file under one particular parameter setting; excludes
selecting parameters and calculating the Hellinger distance.
The table shows that for every dataset, the time of the kNN REX kernel is comparable to the BMP Gaussian kernel.
This result supports the theoretical comparison of their time
complexities. Our method is approximately four times faster
than the KM REX kernel in all cases, which implies that in
KCS choice, our k-NN restriction has a certain advantage over
KM’s log-likelihood optimization in terms of speed. The table
also shows that the fixed Gaussian kernel is much faster than
our method. However, this is the result excluding the time of
bandwidth selection. In practice, our method has a parameter
selection rule presented in Section V-E1, which enable us
to find a good parameter setting within several parameter
examinations. The other methods need a kind of grid search,
typically surveying tens or hundreds of parameter settings.
For fairer comparison, if the time of parameter selection is
taken into consideration, our method would be the fastest.
The bottleneck of our method, and the BMP estimator, is the
computation of k-NN as their complexities imply. In fact, for
3D Road Network and Letter Recognition, it consumes more
than 90% of overall time. If one would like to accelerate it, a
variety of nearest neighbor search techniques are available.
VI.
A PPLICATION
TO
U RBAN M ICRO -S IMULATION
In the last section, we have seen that the kernel density
estimation generally enjoys significant improvements over
sample-copying or bootstrapping approaches, given an unbiased sample, i.e., Case I. Now we proceed to Case II, a biased
sample with marginal frequencies. Our interest in this section
is whether that is yet beneficial for multi-agent simulations.
A. UrbanSim
UrbanSim [1] is one of the most advanced urban microsimulators. It simulates location choice of business units and
households, development and pricing of land and real estate,
and govermental policies involved in the study area, and can
investigate their change over years in parcel level. There are
several applications in U.S. cities: Honolulu, Hawaii; EugeneSpringfield, Oregon; Detroit, Michigan; Salt Lake City, Utah;
Input:
X = { x1 , . . . , xn } ⊂ Rd
⊲ PUMS household microdata
F = { Fb ∈ N | b ∈ B }
⊲ SF household marginal frequencies
k ∈ { 0, . . . , n }
⊲ Neighborhood size
m ∈ { 1, . . . , k + 1 }
⊲ KCS size
l∈N
⊲ Number of households
Output:
Y = { y 1 , . . . , y l } ⊂ Rd
⊲ Synthetic households
1: for all x ∈ X do
2:
calculate the k-NN of x, say X(x, k)
3: end for
4: Y ← ∅
5: repeat
6:
find the most vacant bin b∗ = argmaxb Fb − |Yb |
7:
if Xb∗ 6= ∅ then
8:
draw x1 at random from Xb∗
9:
else
10:
generate x1 by sampling from the uniform distribution on b∗
11:
calculate the k-NN of x1
12:
end if
13:
draw x2 , . . . , xm (xi 6= xj ) from X(x1 , k) at random
14:
generate y by (8)
15:
Y ← Y ∪{y}
16:
if ∃b ∈ B : |Yb | > Fb then
17:
remove a point from Yb at random
18:
end if
19: until |Y | = l
20: return Y
and Seattle, Washington; and in Europe: Paris, France; Brussels, Belgium; and Zurich, Switzerland. UrbanSim requires
data that describe detailed attributes of every household in the
study area as a part of the initial condition for simulation. In
the United States, the U.S. Census Bureau publishes annual
survey results for a 1% or 5% sample of households living
in each area as Public Use Microdata Sample (PUMS). The
marginal distribution of each attribute of households is published by Census Summary Files (SF). Combining them, one
can generate households to run UrbanSim.
B. PopGen
To supply households of UrbanSim, PopGen [29] is available. This software implements Iterative Proportional Updating
(IPU) [7] that weights each sample point to synthesize the
population where the weights are determined so that the
population’s marginal distributions match to the given SF
marginals.
C. k-NN REX Kernel with Bias Correction
To apply the k-NN REX kernel to this task, we need a
little modification so that its outcome matches to SF marginals.
Algorithm 2 is the version to do so. In Algorithm 2, Fb denotes
the frequency in the bin b and Yb denotes the subset of Y whose
elements are in the bin b. Since UrbanSim requires integer
attributes, we round the attributes of real numbers generated
by this algorithm.
D. Experimental Settings
We used the PUMS 1% sample of households in Seattle
in 2000 and extracted six attributes from them: age of head,
annual income, number of people, number of kids, race,
number of workers. PopGen’s IPU and our method generated
258,469 households (the number of households in Seattle in
TABLE III.
CPU TIME IN SECONDS ( AVERAGE ± STANDARD DEVIATION OVER 100- FOLD ICV). B OLD FONTS MEAN AS IN TABLE II.
Dataset
(d dimensions, n training points)
PUMS Person in WA
(d = 2, n = 693)
Swiss Roll
(d = 3, n = 100)
3D Road Network
(d = 4, n = 4348)
Letter Recognition
(d = 17, n = 200)
k-NN
REX Kernel
1.67e+0 ± 2.99e-2
(k = 66, m = 3)
3.04e-1 ± 4.74e-3
(k = 12, m = 4)
2.01e+1 ± 3.07e-1
(k = 30, m = 2)
2.27e+0 ± 3.61e-3
(k = 36, m = 4)
TABLE IV.
H ELLINGER DISTANCE BETWEEN SYNTHESIZED
HOUSEHOLDS AND PUMS ( AVERAGE ± STANDARD DEVIATION OVER
TRIALS ). B OLD FONTS MEAN AS IN TABLE II.
Year
2000 (before UrbanSim)
2012 (after UrbanSim)
k-NN REX Kernel
3.94e-1 ± 1.99e-2
4.04e-1 ± 2.21e-2
KM
REX Kernel
6.67e+0 ± 2.13e-1
(L = 35, m = 4)
1.26e+0 ± 1.01e-1
(L = 66, m = 6)
7.41e+1 ± 2.47e+0
(L = 96, m = 5)
8.39e+0 ± 3.13e-1
(L = 68, m = 20)
10
PopGen’s IPU
4.66e-1 ± 2.13e-2
4.79e-1 ± 2.52e-2
2000). With the households generated by each method, we
ran UrbanSim from 2000 to 2012. We evaluated the Hellinger
distance between the simulated households and the actual ones
of PUMS 5% sample5 . PopGen 1.1 was used as a baseline,
but note that our settings do not involve any person attribute.
We used UrbanSim 4.4.0 and kept all settings, other than
initial households, their default values in the demo project file
“seattle parcel default.xml”. We used k = 50 and m = 7
for our method, according to our parameter selection rule in
Section V-E1.
E. Results and Discussions
The result is shown in Table IV. Before simulation, the
k-NN REX kernel’s households are closer to the PUMS 5%
sample than IPU’s. This tendency remains through the simulation, resulting in a better prediction. We guess this is because
the diversity of household attributes is better reproduced by
our method. IPU simply copies the given sample to synthesize
the population, which never includes households that exist in
a real-world city but are not present in the PUMS 5% sample.
In contrast, the k-NN REX kernel estimates the density of the
underlying population and resamples in a probabilistic manner,
which can generate households that are not present in the given
sample.
VII.
C ONCLUSIONS
In this paper, we proposed the k-nearest neighbor crossover
kernel for population synthesis. We showed that our method
outperforms, in accuracy and speed, the conventional fixed
kernel, the Breiman-Meisel-Purcell variable kernel and the
Kimura-Matsumura crossover kernel. We gave a parameter
selection rule and a rationale for our method. We demonstrated
the usefulness of our method for a household synthesis task in
urban simulation.
For future work, we plan to apply the proposed method
to other real-world tasks, which will reveal the generality
of the method. Especially, oversampling for imbalanced data
would be a fruitful application. Another direction is to further
interchange knowledge between kernel density estimation and
5 In
evaluation, the sampling bias on the PUMS was corrected by using
PUMS weights.
BMP
Gaussian Kernel
1.56e+0 ± 3.68e-2
(k = 2, h = 0.99)
2.87e-1 ± 4.58e-3
(k = 22, h = 0.12)
1.96e+1 ± 3.11e-1
(k = 3, h = 0.18)
2.26e+0 ± 3.99e-2
(k = 2, h = 0.12)
Fixed
Gaussian Kernel
5.19e-1 ± 3.29e-3
(h = 0.011)
1.56e-1 ± 2.76e-3
(h = 0.056)
3.16e+0 ± 3.04e-2
(h = 0.009)
2.06e-1 ± 3.34e-3
(h = 0.059)
genetic algorithm. Theories developed in the former community may help to understand why genetic algorithms work.
Algorithms developed in the latter community can provide
how to calculate large-scale estimators that are theoretically
too complex to derive.
A PPENDIX A
D ERIVATION OF E QUATION (10)
Loftsgaarden & Quesenberry [30] showed that in the consistency analysis of their k-NN density estimator, if k is chosen
as a non-decreasing sequence of positive integers such that
k → ∞ and k/n → 0 as n → ∞, then the k-NN ball tends to
contain arbitrarily many points while its volume converges to
zero in probability as n increases. Our analysis is motivated by
their result. However, to simplify the situation, we will evaluate
the analytic covariance matrix (instead of the sample one) of
a random variate on a ball whose radius deterministically goes
to zero. Thus, this is not a rigorous proof but still gives an
insight into the asymptotic behavior of the sample covariance
of k-NN points in our REX kernel as n → ∞. Assume the
density f is continuously differentiable and has a non-zero
value at the point of consideration.
We consider the first-order Taylor series approximation to
the density around any point x ∈ Rd to which the k-NN
converges. Since the covariance matrix is translation invariant,
we can fix x = 0 without loss of generality:
f (y) = f (0) + f ′ (0)y + O(y 2 ).
∂f
∂f
where f ′ = (∇f )⊤ = ∂y
,
.
.
.
,
∂yd .
1
Integration over a d-ball B with radius δ centered at the
origin is written by
Z
B
φ(y)dy =
Z
0
δ
Z
0
2π
Z
π
...
0
Z
π
0
φ(Φ(θ1 , . . . , θd−1 , r))J(θ1 , . . . , θd−1 , r)
dθ1 . . . dθd−1 dr
where
Φ(θ1 , . . . , θd−1 , r) =
j−1
Y
(r cos θ1 , r sin θ1 cos θ2 , . . . , r(
sin θi ) cos θj ,
i=1
d−1
Y
. . . , r(
i=1
sin θi ) cos θd−1 , r
d−1
Y
i=1
sin θi )
and
J(θ1 , . . . , θd−1 , r) = rd−1
d−2
Y
(sin θi )d−1−i .
i=1
To calculate the integral, there are useful formulas for
definite integrals of trigonometric functions: for any nonnegative integers p, q,
p+1 q+1
Z π
B 2 , 2
(q even),
p
q
sin θ cos θdθ =
0
(q odd),
0
Z 2π
p+1 q+1
,
(p, q even),
2B
2
2
sinp θ cosq θdθ =
0
(otherwise)
0
where
B(x, y) =
Γ(x)Γ(y)
Γ(x + y)
is the beta function and Γ is the gamma function which
satisfies: for any non-negative integer r,
Γ(r + 1) = r!,
1
[2r − 1]! √
(2r − 1)!! √
Γ(r + ) =
π = [2r−1]
π
2
2r
2
[r − 1]!
with R[x] = max(x, 0). Using these formulas and the property
that B φ(y)dy = 0 for any odd function φ(−y) = −φ(y),
we get the following equations:
Z
dy = V δ d ,
Z B
ydy = 0,
Z B
V δ d+2
y 2 dy =
I,
d+2
ZB
y 3 dy = 0
B
where
Suppose y is a random variate with density g. The mean
of y is
Z
g(y)ydy
E[y] =
BZ
1
=
f (y)ydy
z ZB
1
≈
f (0)y + (f ′ (0)y) ydy
z B
Z
Z
1
y 2 dy∇f (0)
ydy +
f (0)
=
z
B
B
1
V δ d+2
=
I · ∇f (0)
f (0) · 0 +
z
d+2
V δ d+2
1
·
∇f (0)
=
f (0)V δ d d + 2
δ 2 ∇f
(0).
=
(d + 2) f
The mean of y 2 = y ⊗ y ⊤ is
Z
E[y 2 ] =
g(y)y 2 dy
BZ
1
=
f (y)y 2 dy
z B
Z
1
f (0)y 2 + (f ′ (0)y) y 2 dy
≈
z B
Z
Z
1
f (0)
y 2 dy + f ′ (0)
y 3 dy
=
z
B
B
1
V δ d+2
′
=
f (0) ·
I + f (0) · 0
z
d+2
f (0)V δ d+2
1
·
I
=
f (0)V δ d
d+2
δ2
=
I.
d+2
Finally, the covariance of y is
V =
π d/2
Γ( d2 + 1)
(volume of unit ball),
y2 = y ⊗ y⊤
3
⊤
y =y⊗y ⊗y
(d × d matrix),
(d × d × d tensor).
Let us consider the density restricted to B:
1
g(y) = f (y|y ∈ B) = f (y)
z
where
z=
Z
f (y)dy
ZB
f (0) + f ′ (0)ydy
Z
Z
ydy
dy + f ′ (0)
= f (0)
≈
C[y] = E[y 2 ] − E[y]2
2
δ2
δ 2 ∇f
=
I−
(0)
d+2
(d + 2) f
∇f f ′
δ4
δ2
(0).
I−
=
d+2
(d + 2)2 f 2
This result implies that our k-NN REX kernel, more generally, generic k-NN crossover kernels such that their crossovers
are compliant to Kita’s preservation of statistics, has the same
order of bandwidth as the BMP estimator. Especially, if we
choose the KCS size of the k-NN REX kernel as m = k+1
√ and
the constant multiplier of the BMP estimator as h = 1/ d + 2,
then both asymptotically coincides. Consequently, the k-NN
REX kernel can be considered as a matrix bandwidth extension
of the BMP estimator.
B
B
B
= f (0) · V δ d + f ′ (0) · 0
= f (0)V δ d .
A PPENDIX B
V ISUAL C OMPARISON
Figure 4 shows the distribution of synthetic populations
for PUMS Person in WA. These plots intuitively support the
correctness of our numerical evaluation presented in Table II.
$900,000
k-NN REX kernel
$800,000
$800,000
$700,000
$700,000
$600,000
$600,000
Annual income
Annual income
$900,000
$500,000
$400,000
$300,000
$300,000
$200,000
$100,000
$100,000
$0
$0
[12]
$-100,000
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Age
$900,000
[13]
Age
$900,000
BMP Gaussian kernel
$800,000
$800,000
$700,000
$700,000
$600,000
$600,000
Annual income
Annual income
[11]
$400,000
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
$500,000
$400,000
$300,000
fixed Gaussian kernel
[14]
$500,000
[15]
$400,000
$300,000
$200,000
$200,000
$100,000
$100,000
$0
$0
$-100,000
[16]
$-100,000
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Age
$900,000
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Age
$900,000
training data
$800,000
$800,000
$700,000
$700,000
$600,000
$600,000
Annual income
Annual income
KM REX kernel
$500,000
$200,000
$-100,000
[10]
$500,000
$400,000
$300,000
[17]
[18]
$500,000
$400,000
$300,000
$200,000
$200,000
$100,000
$100,000
$0
$0
$-100,000
test data
[19]
$-100,000
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Age
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Age
[20]
Fig. 4. Synthetic populations, training data and test data for PUMS Person
in WA.
[21]
R EFERENCES
[1]
P. Waddell, “UrbanSim: Modeling urban development for land use,
transportation, and environmental planning,” Journal of the American
Planning Association, vol. 68, no. 3, pp. 297–314, 2002.
[2]
M. Milkovits, E. Huang, C. Antoniou, M. Ben-Akiva, and J. Lopes,
“DynaMIT 2.0: The next generation real-time dynamic traffic assignment system,” in Proceedings of the Second International Conference
on Advances in System Simulation (SIMUL 2010), 2010, pp. 45–51.
[3]
P. Vytelingum, S. D. Ramchurn, T. D. Voice, A. Rogers, and N. R.
Jennings, “Trading agents for the smart electricity grid,” in Proceedings of the 9th International Conference on Autonomous Agents and
Multiagent Systems (AAMAS 2010), 2010, pp. 897–904.
[4]
D. Massaguer, V. Balasubramanian, S. Mehrotra, and N. Venkatasubramanian, “Multi-agent simulation of disaster response,” in Workshop
on Agent Technology for Disaster Management (ATDM) at Autonomous
Agents and Multiagent Systems (AAMAS 2006), 2006.
[5]
U. Endriss and N. Maudet, “Welfare engineering in multiagent systems,”
in Engineering Societies in the Agents World IV, ser. Lecture Notes in
Computer Science (LNCS). Springer Berlin Heidelberg, 2004, vol.
3071, pp. 93–106.
[6]
W. E. Deming and F. F. Stephan, “On a least squares adjustment of a
sampled frequency table when the expected marginal totals are known,”
The Annals of Mathematical Statistics, vol. 11, no. 4, pp. 427–444,
1940.
[7]
X. Ye, K. Konduri, R. M. Pendyala, B. Sana, and P. Waddell, “A
methodology to match distributions of both household and person
attributes in the generation of synthetic populations,” in 88th Annual
Meeting of the Transportation Research Board, 2009.
[8]
U.
S.
C.
B.,
“Public
use
microdata
samples
(PUMS),”
2012.
[Online].
Available:
http://www.census.gov/acs/www/data documentation/pums data/
[9]
S. R. Sain, “Multivariate locally adaptive density estimation,” Computational Statistics & Data Analysis, vol. 39, no. 2, pp. 165–186, 2002.
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
J. Sakuma and S. Kobayashi, “Real-coded crossover as a role of kernel
density estimation,” in Proceedings of the 7th Annual Conference on
Genetic and Evolutionary Computation (GECCO 2005). New York,
NY, USA: ACM, 2005, pp. 703–710.
S. Kimura and K. Matsumura, “Density estimation using crossover
kernels and its application to a real-coded genetic algorithm,” in
Proceedings of the 2008 IEEE Congress on Evolutionary Computation
(CEC 2008), 2008, pp. 694–701.
L. Breiman, “Bagging predictors,” Machine Learning, vol. 24, no. 2,
pp. 123–140, 1996.
M. Jones, “Variable kernel density estimates and variable kernel density
estimates,” Australian Journal of Statistics, vol. 32, no. 3, pp. 361–371,
1990.
G. R. Terrell and D. W. Scott, “Variable kernel density estimation,”
Ann. Statist., vol. 20, no. 3, pp. 1236–1265, 1992.
J. S. Simonoff, Smoothing Methods in Statistics, ser. Springer Series in
Statistics. Springer, 1996.
T. Duong and M. L. Hazelton, “Cross-validation bandwidth matrices
for multivariate kernel density estimation,” Scandinavian Journal of
Statistics, vol. 32, no. 3, pp. 485–506, 2005.
J. Chacón and T. Duong, “Multivariate plug-in bandwidth selection with
unconstrained pilot bandwidth matrices,” TEST, vol. 19, no. 2, pp. 375–
398, 2010.
L. Breiman, W. Meisel, and E. Purcell, “Variable kernel estimates of
multivariate densities,” Technometrics, vol. 19, no. 2, pp. 135–144,
1977.
I. S. Abramson, “On bandwidth variation in kernel estimates–a square
root law,” The Annals of Statistics, vol. 10, no. 4, pp. 1217–1223, 1982.
H. Kita, I. Ono, and S. Kobayashi, “Multi-parental extension of the unimodal normal distribution crossover for real-coded genetic algorithms,”
in Proceedings of the 1999 Congress on Evolutionary Computation
(CEC 99), vol. 2, 1999, pp. 1581–1588.
S. Kobayashi, “The frontiers of real-coded genetic algorithms,” Journal
of Japanese Society for Artificial Intelligence, vol. 24, no. 1, pp. 128–
143, 2009.
B. W. Silverman, Density Estimation for Statistics and Data Analysis,
ser. Chapman & Hall/CRC Monographs on Statistics & Applied Probability. London, UK: Taylor & Francis, 1986, vol. 26.
K. Fukunaga, Introduction to Statistical Pattern Recognition, ser. Electrical Science. New York, NY, USA: Academic Press, 1972.
S.
Gerber,
“cems:
Conditional
expectation
manifolds,”
2011,
R
package
version
0.2.
[Online].
Available:
http://CRAN.R-project.org/package=cems
K. Bache and M. Lichman, “UCI machine learning repository,” 2013.
[Online]. Available: http://archive.ics.uci.edu/ml
A. Bhattacharyya, “On a measure of divergence between two statistical
populations defined by their probability distributions,” Bulletin of the
Calcutta Mathematical Society, vol. 35, pp. 99–109, 1943.
J. Zhang, S. Li, and J. Wang, “Manifold learning and applications in
recognition,” in Intelligent Multimedia Processing with Soft Computing,
ser. Studies in Fuzziness and Soft Computing, Y.-P. Tan, K. Yap, and
L. Wang, Eds. Springer Berlin Heidelberg, 2005, vol. 168, pp. 281–
300.
N. Hamada, J. Sakuma, S. Kobayashi, and I. Ono, “Functionalspecialization multi-objective real-coded genetic algorithm: FSMOGA,” in Parallel Problem Solving from Nature – PPSN X, ser.
Lecture Notes in Computer Science (LNCS).
Berlin, Heidelberg:
Springer-Verlag, 2008, vol. 5199, pp. 691–701.
K. Konduri, “PopGen: Population generator,” 2010. [Online]. Available:
http://urbanmodel.asu.edu/popgen.html
D. O. Loftsgaarden and C. P. Quesenberry, “A nonparametric estimate of
a multivariate density function,” The Annals of Mathematical Statistics,
vol. 36, no. 3, pp. 1049–1051, 1965.
| 9cs.NE
|
Towards Reversible Sessions∗
Francesco Tiezzi
Nobuko Yoshida
IMT Institute for Advanced Studies, Lucca, Italy
Imperial College, London, U.K.
[email protected]
[email protected]
In this work, we incorporate reversibility into structured communication-based programming, to allow parties of a session to automatically undo, in a rollback fashion, the effect of previously executed
interactions. This permits taking different computation paths along the same session, as well as
reverting the whole session and starting a new one. Our aim is to define a theoretical basis for
examining the interplay in concurrent systems between reversible computation and session-based
interaction. We thus enrich a session-based variant of π-calculus with memory devices, dedicated
to keep track of the computation history of sessions in order to reverse it. We discuss our initial
investigation concerning the definition of a session type discipline for the proposed reversible calculus, and its practical advantages for static verification of safe composition in communication-centric
distributed software performing reversible computations.
1
Introduction
Reversible computing aims at providing a computational model that, besides the standard (forward) executions, also permits backward execution steps in order to undo the effect of previously performed
forward computations. Reversibility is a key ingredient in different application domains since many
years and, recently, also in the design of reliable concurrent systems, as it permits understanding existing patterns for programming reliable systems (e.g., compensations, checkpointing, transactions) and,
possibly, improving them or developing new ones.
A promising line of research on this topic advocates reversible variants of well-established process
calculi, such as CCS and π-calculus, as formalisms for studying reversibility mechanisms in concurrent
systems. Our work incorporates reversibility into a variant of π-calculus equipped with session primitives supporting structured communication-based programming. A (binary) session consists in a series
of reciprocal interactions between two parties, possibly with branching and recursion. Interactions on
a session are performed via a dedicated private channel, which is generated when initiating the session.
Session primitives come together with a session type discipline offering a simple static checking framework to guarantee the correctness of communication patterns.
Practically, combining reversibility and sessions paves the way for the development of session-based
communication-centric distributed software intrinsically capable of performing reversible computations.
In this way, without further coding effort by the application programmer, the interaction among session
parties is relaxed so that, e.g., the computation can automatically go back, thus allowing to take different
paths when the current one is not satisfactory. As an application example, used in this paper for illustrating our approach, we consider a simple scenario involving a client and multiple providers offering the
same service (e.g., on-demand video streaming). The client connects to a provider to request a given service (specifying, e.g., title of a movie, video quality, etc.). The provider replies with a quote determined
according to the requested quality of service and to the servers status (current load, available bandwidth,
∗ This work has been partially supported by the COST Action BETTY (IC1201), by the EU project ASCENS (257414), and
by the Italian MIUR PRIN project CINA (2010LHT4KM).
Alastair F. Donaldson, Vasco Vasconcelos (Eds.): Proceedings
of the 7th Workshop on Programming Language Approaches
to Concurrency and Communication-cEntric Software (PLACES 2014)
EPTCS 155, 2014, pp. 17–24, doi:10.4204/EPTCS.155.3
c Tiezzi F. and Yoshida N.
This work is licensed under the
Creative Commons Attribution License.
18
Towards Reversible Sessions
etc.). Then, the client can either accept, negotiate or reject the quote. If a problem occurs during the
interaction between the client and the provider, the computation can be reverted, in order to allow the
client to automatically start a new session with (possibly) another provider.
The proposed reversible session-based calculus relies on memories to store information about interactions and their effects on the system, which otherwise would be lost during forward computations.
This data is used to enable backward computations that revert the effects of the corresponding forward
ones. Each memory is devoted to record data concerning a single event, which can correspond to the
taking place of a communication action, a choice or a thread forking. Memories are connected each
other, in order to keep track of the computation history, by using unique thread identifiers as links. Like
all other formalisms for reversible computing in concurrent settings, forward computations are undone
in a causal-consistent fashion, i.e. backtracking does not have to necessarily follow the exact order of
forward computations in reverse, because independent actions can be undone in a different order.
The resulting formalism offers a theoretical basis for examining the interplay between reversible
computations and session-based structured interactions. We notice that reversibility enables session
parties not only to partially undo the interactions performed along the current session, but also to automatically undo the whole session and restart it, possibly involving different parties. The advantage
of the reversible approach is that this behaviour is realised without explicitly implementing loops. On
the other hand, the session type discipline affects reversibility as it forces concurrent interactions to follow structured communication patterns. In fact, linearizing behaviours on sessions reduces the effect of
causal consistency, because concurrent interactions along the same session are forbidden and, hence, the
rollback along a session follows a single path. However, interactions along different sessions are still
concurrent and, therefore, they can be reverted as usual in a causal-consistent fashion. Notably, interesting issues concerning reversibility and session types are still open questions, especially for what concerns
the validity in the reversible setting of standard properties (e.g., progress enforcement) and possibly new
properties (e.g., reversibility of ongoing session history, irreversible closure of sessions).
2
Related work
We review here the most closely related works, which concern the definition of reversible process calculi;
we refer to [9] for a more detailed review of reversible calculi.
Reversible CCS (RCCS) [3] is the first proposal of reversible calculus, from which all subsequent
works drew inspiration. To each currently running thread is associated an individual memory stack
keeping track of past actions, as well as forks and synchronisations. Information pushed on the memory
stacks, upon doing a forward transition, can be then used for a roll-back. The memories also serve as
a naming scheme and yield unique identifiers for threads. When a process divides in two sub-threads,
each sub-thread inherits the father memory together with a fork number indicating which of the two sons
the thread is. A drawback of this approach is that the parallel operator does not satisfy usual structural
congruence rules as commutativity, associativity and nil process as neutral element.
CCS-R [4] is another reversible variant of CCS, which mainly aims at formalising biological systems.
Like RCCS, it relies on memory stacks for storing data needed for backtracking, which now also includes
events corresponding to unfolding of process definitions. Differently from RCCS, specific identifiers are
used to label threads, and a different approach is used for dealing with forking.
CCS with communication Keys (CCSK) [8] is a reversible process calculus obtained by applying a
general procedure to produce reversible calculi. A relevant aspect of this approach is that it does not
rely on memories for supporting backtracking. The idea is to maintain the structure of processes fixed
Tiezzi F. and Yoshida N.
19
throughout computations, thus avoiding to consume guards and alternative choices. To properly revert
synchronisations, the two threads have to agree on a key, uniquely identifying that communication.
ρπ [7] is a reversible variant of the higher-order π-calculus. It borrows from RCCS the use of
memories for keeping track of past actions, although in ρπ they are not stacks syntactically associated
to threads, but parallel terms each one dedicated to a single communication. The connection between
memories and threads is kept by resorting to identifiers, which resemble CCSK keys. Fork handling is
based on structured tags, used to connect the identifier of a thread with the identifiers of its sub-threads.
This approach to reversibility has been applied in [5] to a distributed tuple-based language.
Another reversible variant of π-calculus is Rπ [2]. Similarly to RCCS, this calculus relies on memory
stacks, now recording communication events and forking. Differently from ρπ, it considers standard πcalculus (without choice and replication) as a host calculus and its semantics is defined in terms of a
labelled transition relation.
Finally, reversible structures [1] is a simple computational calculus for modelling chemical systems.
Reversible structures does not exploit memories, but maintains the structure of terms and uses a special
symbol to indicate the next forward and backward operations that a term can perform.
In our work, we mainly take inspiration from the ρπ approach. In fact, all other approaches are based
on CCS and cannot be directly applied to a calculus with name-passing. Moreover, the ρπ approach is
preferable to the Rπ one because the former proposes a reduction semantics, which we are interested
in, while the latter proposes a labelled semantics, which would complicate our theoretical framework in
order to properly deal with scope extrusion.
3
Reversible Session-based π-calculus
In this section, we introduce a reversible extension of a π-calculus enriched with primitives for managing
binary sessions, i.e. structured interactions between two parties. We call ReSπ (Reversible Session-based
π-calculus) this formalism. Due to lack of space, some technical details about semantics and results have
been omitted; we refer the interested reader to [9] for a more complete account.
From π-calculus to ReSπ. Our approach to keep track of computation history in ReSπ is as follows: we
tag processes with unique identifiers (tagged processes are called threads) and use memories to store the
information needed to reverse each single forward reduction. Thus, the history of a reduction sequence
is stored in a number of small memories connected each other by using tags as links. In this way, ReSπ
terms can perform, besides forward computations (denoted by ), also backward computations (denoted
by ) that undo the effect of the former ones in a causal-consistent fashion.
π-calculus processes and expressions are given by the grammars in Figure 1. The synchronisation on
a shared channel a of processes ā(x).P and a(y).Q initiates a session along a fresh session channel s. This
channel consists in a pair of (dual) endpoints, denoted by s and s̄ (such that s̄¯ = s), each one dedicated
to one party to exchange values with the other. These endpoints replace variables x and y, by means
of a substitution application, in order to be used by P and Q, respectively, for later communications.
Primitives k!hei.P and k0 ?(x).Q denote output and input via session endpoints identified by k and k0 ,
respectively. These communication primitives realise the standard synchronous message passing, where
messages result from expressions evaluation and may contain endpoints (delegation). Constructs k / l.P
and k0 . {l1 : P1 , . . . , ln : Pn } denote label selection and branching (with l1 , . . . , ln pairwise distinct) via
k and k0 , respectively. The above interaction primitives are combined by conditional choice, parallel
composition, restriction, recursion and inaction.
20
Towards Reversible Sessions
Shared channels a, b, . . .
Labels l, l 0 , . . .
Channels c ::= a | s
Session channels s, s0 , . . .
Process variables X, Y , . . .
Names h ::= c | t
Session endpoints s, s̄, s0 , s̄0 , . . .
Tags t, t 0 , . . .
Session ids k ::= s | s̄ | x
Variables x, y, . . .
Shared ids u ::= a | x
Processes P ::= ū(x).P | u(x).P | k!hei.P | k?(x).P | k / l.P | k . {l1 : P1 , . . . , ln : Pn }
| if e then P else Q | P | Q | (νc) P | X | µX.P | 0
Expressions e ::= v | x | op(e1 , . . . , en )
Values v ::= true | false | 0, 1, . . . | a | s | s̄
ReSπ processes M ::= t : P | (νh) M | M | N | m | nil
Memories m ::= ht1 −A→ t2 ,t10 ,t20 i | ht, e? P : Q,t 0 i | ht ⇒ (t1 ,t2 )i
A ::= a(x)(y)(νs)PQ | khei(x)PQ | k / li P{l1 : P1 , . . . , ln : Pn }
Figure 1: ReSπ syntax
ReSπ processes are built upon π-calculus processes by labelling them with tags to uniquely identify
threads t : P. Uniqueness of tags is ensured by using the restriction operator and by only considering
reachable terms (Def. 1). Moreover, ReSπ extends π-calculus with three kinds of memories m. An action
memory ht1 −A→ t2 ,t10 ,t20 i stores an action event A together with the tag t1 of the active party of the action,
the tag t2 of the passive party, and the tags t10 and t20 of the new threads activated by the corresponding
reduction. An action event records information necessary to revert each kind of interactions, which can
be either a session initiation a(x)(y)(νs)PQ, a communication along an established session khei(x)PQ,
or a branch selection k / li P{l1 : P1 , . . . , ln : Pn }. A choice memory ht, e? P : Q,t 0 i stores a choice event
together with the tag t of the conditional choice and t 0 of the new activated thread. The event e? P : Q
records the evaluated expression e, and processes P and Q of the then- and else-branch, respectively. A
fork memory ht ⇒ (t1 ,t2 )i stores the tag t of a splitting thread, of the form t : (P | Q), and the tags t1
and t2 of the new activated threads t1 : P and t2 : Q; these memories are analogous to connectors in [5].
Threads and memories are composed by parallel composition and restriction operators.
Not all processes allowed by the syntax are semantically meaningful. In a general term, the history
stored in the memories may not be consistent, due to the use of non-unique tags or broken connections
between continuation tags within memories and corresponding threads. For example, given the choice
memory ht, e? P : Q,t 0 i, we have a broken connection when no thread tagged by t 0 exists in the ReSπ
process and no memory of the form ht 0 −A→ t2 ,t10 ,t20 i, ht1 −A→ t 0 ,t10 ,t20 i, ht 0 , e? P1 :P2 ,t1 i, and ht 0 ⇒ (t1 ,t2 )i
exists. Thus, as in [2], to ensure history consistency we only consider reachable processes, i.e. processes
obtained by means of forward and backward reductions from processes with unique tags and no memory.
Def. 1 (Reachable processes) The set of reachable ReSπ processes is the closure under (see below)
of the set of terms, whose threads have distinct tags, generated by M ::= t : P | (νc) M | M | N | nil.
ReSπ semantics. The ReSπ operational semantics is given in terms of a reduction relation , given
as the union of the forward and backward reduction relations. We report here, by way of examples, the
forward and backward rules for session initiation (we require s, s̄ fresh in P1 and P2 in the forward rule):
t1 : ā(x).P1 | t2 : a(y).P2 (νs,t10 ,t20 )(t10 : P1 [s̄/x] | t20 : P2 [s/y] | ht1 −a(x)(y)(νs)P1 P2 → t2 ,t10 ,t20 i)
(νs,t10 ,t20 )(t10 : P | t20 : Q | ht1 −a(x)(y)(νs)P1 P2 → t2 ,t10 ,t20 i)
t1 : ā(x).P1 | t2 : a(y).P2
Tiezzi F. and Yoshida N.
21
When two parallel threads synchronise to establish a new session, two fresh tags are created to uniquely
identify the continuations. Moreover, all relevant information is stored in the action memory: the tag t1
of the initiator (i.e., the thread executing a prefix of the form ā(·)), the tag t2 of the thread executing the
dual action, the tags t10 and t20 of their continuations, the shared channel a used for the synchronisation,
the replaced variables x and y, the generated session channel s, and the processes P1 and P2 to which
substitutions are applied. All such information is exploited in the backward rule to revert this reduction.
In particular, the corresponding backward reduction is triggered by the coexistence of the memory described above with two threads tagged t10 and t20 , all of them within the scope of the session channel s
and tags t10 and t20 generated by the forward reduction (which, in fact, are removed by the backward one).
When considering reachable processes, due to tag uniqueness, processes P and Q coincide with P1 [s̄/x]
and P2 [s/y]; indeed, as registered in the memory, these latter processes have been tagged with t10 and t20
by the forward reduction. Therefore, the fact that two threads tagged with t10 and t20 are in parallel with
the memory ensures that all actions possibly executed by the two continuations activated by the forward
computation have been undone and, hence, we can safely undone the forward computation itself.
Multiple providers scenario. The scenario involving a client and two providers informally introduced
in Section 1 is rendered in ReSπ as (t1 : Pclient | t2 : Pprovider1 | t3 : Pprovider2 ), where the client process
Pclient is
alogin (x). x!hsrv reqi. x?(yquote ). if accept(yquote ) then x / lacc . Pacc
else (if negotiate(yquote ) then x / lneg . Pneg else x / lre j . 0)
while Pprovider i is
alogin (y). y?(zreq ). y!hquotei (zreq )i. y . {lacc : Qacc , lneg : Qneg , lre j : 0}
If the client contacts the first provider and accepts the proposed quote, the system evolves to
M = (νs, . . . ,t10 ,t20 )(t10 : Pacc [s̄/x, quote/yquote ] | t20 : Qacc [s/y, srv req/zreq ] | m1 | . . . | m5 ) | Pprovider2
where memories mi keep track of the computation history. Now, if a problem occurs during the subsequent interactions, the computation can be reverted to allow the client to start a new session with
(possibly) another provider:
M
∗
t1 : Pclient | t2 : Pprovider1 | t3 : Pprovider2
Properties of ReSπ. We show here that ReSπ enjoys standard properties of reversible calculi.
First, we demonstrate that ReSπ is a conservative extension of the (session-based) π-calculus. In fact,
as most reversible calculi, ReSπ is only a decoration of its host calculus. This decoration can be erased
by means of the forgetful map φ , mapping ReSπ terms into π-calculus ones by removing memories, tag
annotations and tag restrictions. The following lemmas show that each forward reduction of a ReSπ
process corresponds to a reduction of the corresponding π-calculus process and vice versa.
Lemma 1 Let M and N be two ReSπ processes. If M N then φ (M) → φ (N).
Lemma 2 Let P and Q be two π-calculus processes. If P → Q then for any ReSπ process M such that
φ (M) = P there exists a ReSπ process N such that φ (N) = Q and M N.
Then, we show that ReSπ backward reductions are the inverse of the forward ones and vice versa.
22
Towards Reversible Sessions
Lemma 3 (Loop lemma) Let M and N be two reachable ReSπ processes. M N if and only if N
M.
We conclude with the causal consistency result stating that two sequences of reductions (called traces
and ranged over by σ ), with the same initial state (coinitial) and equivalent w.r.t. the standard notion of
causal equivalence (), lead to the same final state (cofinal). Thus, in this case, we can rollback to the
initial state by reversing any of the two traces.
Theorem 1 Let σ1 and σ2 be coinitial traces. Then, σ1 σ2 if and only if σ1 and σ2 are cofinal.
4
Discussion on a type discipline
A question that should be answered before defining a static type discipline for a reversible calculus is
“Should we type check the processes stored in the memories?”. The question arises from the fact that
we should be able to determine if any ReSπ process is well-typed or not. In our case the answer is
“Yes”, otherwise typability would not be preserved under reduction (i.e., Subject Reduction would not
be satisfied). It is indeed easy to define a ReSπ process (see [9]) containing a memory that, even if
consistent, triggers a backward reduction leading to an untypable term (by the type system defined in
[10] for the host calculus).
One could wonder now if it is possible to type ReSπ processes in a naı̈ve way by separately type
checking the term resulting from the application of φ and each single memory, by using the type system
in [10]. For each memory we would check the term that has triggered the forward reduction generating
the memory. In general, this approach does not work, because the term stored in a memory cannot
be type checked in isolation without taking into account its context. For example, consider a memory
corresponding to a communication along a session s typable under typing ∆ = s̄ :![int].end · s :?[int].end
and in parallel with (t1 : s̄!h1i | t2 : s?(x)). The term resulting from the corresponding backward reduction
is not typable, because the typings of its sub-terms are not composable (indeed, ∆ · ∆ is not defined).
Memory context can be considered by extending the type system in [10] with rules that permits
typing (processes stored in) memories and ignoring tag annotations and tag restrictions (see [9] for the
definition of this type system). In this way, during type checking, typings of memories and threads must
be composed by means of the rule for parallel composition. Thus, e.g., the ReSπ process mentioned
above is, rightly, untypable. This type system properly works only on a simplified setting, which permits avoiding to deal with dependencies among memories and the threads outside memories, that could
cause unwanted conflicts during type checking. Specifically, we consider the class of ReSπ processes
that, extending Def. 1, are obtained by means of forward and backward reductions from processes with
unique tags, no memory, no session initialised, no conditional choices and recursions at top-level, and no
delegation. The characteristic of these processes is that, for each memory inside a process, there exists
within the process an ancestor memory corresponding to the initialisation of the considered session. The
type system checks only this latter kind of memories, which significantly simplifies the theory.
Coming back now to the multiple providers scenario, we can verify that the initial process is welltyped. In particular, the channel alogin can be typed by the shared channel type
h ?[Request]. ![Quote]. &[lacc : αacc , lneg : αneg , lre j : end] i
where sorts Request and Quote are used to type requests and quotes, respectively. Let us consider now a
scenario where the client wills to concurrently submit two different requests to the same provider, which
would concurrently serve them. Consider in particular the following specification of the client
alogin (x). ( x!hsrv req 1i. P1 | x!hsrv req 2i. P2 )
Tiezzi F. and Yoshida N.
23
The new specification is clearly not well-typed, due to the use of parallel threads within the same session.
This permits avoiding mixing up messages related to different requests and wrongly delivering them. In
order to properly concurrently submit separate requests, the client must instantiate separate sessions with
the provider, one for each request.
5
Concluding remarks
To bring the benefits of reversible computing to structured communication-based programming, we have
defined a theoretical framework based on π-calculus that can be used as formal basis for studying the
interplay between (causal-consistent) reversibility and session-based structured interaction.
The type discipline for ReSπ is still subject of study. In fact, the type system mentioned in Section 4 is not completely satisfactory, because its use is limited to a restricted class of processes. To
consider a broader class, an appropriate static type checking approach for memories has to be devised.
For each memory, we would check a term composed of the threads stored in the memory and of a context
composed of threads that have not been generated by the execution of the memory threads.
Concerning the reversible calculus, we plan to investigate the definition of a syntactic characterisation of consistent terms, which statically enforces history consistency in memories (as in [7]). It is
worth noticing that the calculus is fully reversible, i.e. backward computations are always enabled. Full
reversibility provides theoretical foundations for studying reversibility in session-based π-calculus, but
it is not suitable for a practical use. In line with [6], we plan to enrich the language with mechanisms to
control reversibility. Moreover, we intend to enrich the framework with an irreversible action for committing the closure of sessions. In this way, computation would go backward and forward, allowing the
parties to try different interactions, until the session is successfully completed. For instance, the process
Pacc in our example could terminate by performing the irreversible action commit(x), which has to synchronise with action commit(y) in Qacc . Differently from sessions terminated by 0, a session terminated
by a commit synchronisation is unbacktrackable. The irreversibility is due to the fact that no backward
rule is defined to revert this interaction. The type theory should be tailored to properly deal with this kind
of session closure.
As longer-term goals, we intend to apply the proposed approach to other session-based formalisms,
which consider, e.g., asynchronous sessions and multiparty sessions. Moreover, we plan to investigate
implementation issues that may arise when incorporating the approach into standard programming languages, in particular in case of a distributed setting.
References
[1] Luca Cardelli & Cosimo Laneve (2011): Reversible structures.
doi:10.1145/2037509.2037529.
In: CMSB, ACM, pp. 131–140,
[2] I. Cristescu, J. Krivine & D. Varacca (2013): A Compositional Semantics for the Reversible p-Calculus. In:
LICS, IEEE, pp. 388–397, doi:10.1109/LICS.2013.45.
[3] Vincent Danos & Jean Krivine (2004): Reversible Communicating Systems. In: CONCUR, LNCS 3170,
Springer, pp. 292–307, doi:10.1007/978-3-540-28644-8 19.
[4] Vincent Danos & Jean Krivine (2007): Formal Molecular Biology Done in CCS-R. Electr. Notes Theor.
Comput. Sci. 180(3), pp. 31–49, doi:10.1016/j.entcs.2004.01.040.
[5] E. Giachino, I. Lanese, C.A. Mezzina & F. Tiezzi (2013): Causal-Consistent Reversibility in a Tuple-Based
Language. Technical Report. http://www.cs.unibo.it/~lanese/work/klaimrev-TR.pdf.
24
Towards Reversible Sessions
[6] I. Lanese, C.A. Mezzina, A. Schmitt & J. Stefani (2011): Controlling Reversibility in Higher-Order Pi. In:
CONCUR, LNCS 6901, Springer, pp. 297–311, doi:10.1007/978-3-642-23217-6 20.
[7] I. Lanese, C.A. Mezzina & J. Stefani (2010): Reversing Higher-Order Pi. In: CONCUR, LNCS 6269,
Springer, pp. 478–493, doi:10.1007/978-3-642-15375-4 33.
[8] Iain C. C. Phillips & Irek Ulidowski (2007): Reversing algebraic process calculi. J. Log. Algebr. Program.
73(1-2), pp. 70–96, doi:10.1016/j.jlap.2006.11.002.
[9] Francesco Tiezzi & Nobuko Yoshida (2014): Towards Reversible Sessions. Technical Report. http://cse.
lab.imtlucca.it/~tiezzi/papers/places2014_full.pdf.
[10] N. Yoshida & V.T. Vasconcelos (2007): Language Primitives and Type Discipline for Structured
Communication-Based Programming Revisited: Two Systems for Higher-Order Session Communication.
Electr. Notes Theor. Comput. Sci. 171(4), pp. 73–93, doi:10.1016/j.entcs.2007.02.056.
| 6cs.PL
|
arXiv:1601.07131v1 [math.GR] 26 Jan 2016
ON THE YANG-BAXTER EQUATION AND LEFT NILPOTENT
LEFT BRACES
FERRAN CEDÓ, TATIANA GATEVA-IVANOVA AND AGATA SMOKTUNOWICZ
Abstract. We study non-degenerate involutive set-theoretic solutions (X, r)
of the Yang-Baxter equation, we call them simply solutions. We show that
the structure group G(X, r) of a finite non-trivial solution (X, r) cannot be
an Engel group. It is known that the structure group G(X, r) of a finite
multipermutation solution (X, r) is a poly-Z group, thus our result gives a
rich source of examples of braided groups and left braces G(X, r) which are
poly-Z groups but not Engel groups.
We also show that a finite solution of the Yang-Baxter equation can be
embedded in a convenient way into a finite brace and into a finite braided
group.
For a left brace A, we explore the close relation between the multipermutation level of the solution associated with it and the radical chain A(n+1) =
A(n) ∗ A introduced by Rump.
1. Introduction
Braces were introduced by Rump [15] to study non-degenerate involutive settheoretic solutions of the Yang-Baxter equation.
Recall that a left brace is a set B with two operations, + and ·, such that (B, +)
is an abelian group, (B, ·) is a group and for every a, b, c ∈ B,
(1)
a · (b + c) + a = a · b + a · c.
Right braces are defined similarly, changing the property (1) by (a + b) · c + c =
a · c + b · c. A two-sided brace is a left brace which is also a right brace. In any
left brace (B, +, ·) one defines another operation ∗ by the rule
a ∗ b = a · b − a − b,
for a, b ∈ B. It is known that (B, +, ·) is a two-sided brace if and only if (B, +, ∗)
is a Jacobson radical ring. Conversely, if R is a Jacobson radical ring, then one
Date: January 27, 2016.
2010 Mathematics Subject Classification. Primary 16N80, 16P90, 16N40 .
Key words and phrases. brace, Yang-Baxter equation, nilpotent braces.
1
2
CEDÓ, GATEVA-IVANOVA, SMOKTUNOWICZ
defines a new operation ◦ on R by a ◦ b = ab + a + b and (R, ◦) is called the adjoint
group of the radical ring R. Then (R, +, ◦) is a two-sided brace. Hence the study
of two-sided braces is equivalent to the study of Jacobson radical rings.
In general the operation ∗ in a left brace B is not associative, but it is left
distributive with respect to the sum, that is
a ∗ (b + c) = a ∗ b + a ∗ c,
for a, b, c ∈ B.
Let B be a left brace. For a ∈ B, let La : B −→ B be the map defined by
La (b) = ab − a for all b ∈ B. It is known that La is an automorphism of the
additive group of the left brace B and the map L : (B, ·) −→ Aut(B, +), defined
by a 7→ La , is a morphism of groups. The kernel of this morphism is called the
socle of B,
Soc(B) := {a ∈ B | La = id} = {a ∈ B | ab = a + b, for all b ∈ B}.
In fact the socle of a left brace B is an ideal of B, that is, a normal subgroup of its
multiplicative group invariant by the maps La for all a ∈ B. In particular, Soc(B)
is also a subgroup of the additive group of B. Note that if a, b ∈ Soc(B), then
a − b = Lb (b−1 a) ∈ Soc(B). Therefore the quotient of the multiplicative group
B/ Soc(B) is also the quotient of the additive group and (B/ Soc(B), +, ·) is a left
brace, the left brace quotient of B modulo its ideal Soc(B).
Let X be a non-empty set. Recall that a map r : X × X −→ X × X is a
set-theoretic solution of the Yang-Baxter equation if
r12 r23 r12 = r23 r12 r23 ,
where r12 , r23 : X × X × X −→ X × X × X are the maps r12 = r × idX and
r23 = idX × r. We will write r(x, y) = (x y, xy ). The map r is non-degenerate
if for every x ∈ X the maps y 7→ x y and y 7→ y x are bijective, r is involutive if
r2 = idX 2 .
Convention. By a solution of the YBE (or shortly, a solution) we mean a nondegenerate involutive set-theoretic solution (X, r) of the Yang-Baxter equation.
Let (X, r) be a solution of the YBE. The structure group of (X, r) is the group
G(X, r) with presentation
G(X, r) = hX | xy = (x y)(xy ), x, y ∈ Xi.
(Some authors call G(X, r) the YB group of (X, r)).
NILPOTENT BRACES AND THE YANG–BAXTER EQUATION
3
It follows from the results in [8] that X is naturally embedded in G(X, r). One
can define a sum on G(X, r) such that (G(X, r), +) is a free abelian group with
basis X. Moreover, (G(X, r), +, ·) is a left brace such that x y = Lx (y) ∈ X for all
x, y ∈ X, see [5, 9].
We say that this is the canonical left brace structure on G(X, r). It is known
that the group G acts on the set X from the left (and from the right). Moreover,
the assignments x 7→ Lx extend to a group homomorphism L : G −→ SymX .
The image L(G) of this homomorphism is a subgroup of SymX called the permutation group of (X, r) and denoted by G(X, r). It is known that G(X, r) :=
hLx | x ∈ Xi, where Lx (y) =
x
y, for all x, y ∈ X. The group epimorphism
L : G(X, r) −→ G(X, r), x 7→ Lx has kernel Ker L = Soc(G(X, r)) (as sets). Thus
G(X, r) inherits a structure of a left brace via this natural isomorphism of groups,
we say that this is the canonical structure of a left brace on G(X, r). Moreover,
G(X, r)/ Soc(G(X, r)) ∼
= G(X, r) as symmetric groups (i.e. involutive braided
groups) and as left braces, [9, 5].
In this paper, we prove some general results about braces and apply these
to study the close relations between the properties of solutions (X, r) and their
associated left braces G(X, r). This is in the spirit of [9] and [5].
2. Some results on G(X, r)
The results of this section were motivated by a result from [9], which assures
that the structure group G(X, r) of a non-trivial solution (X, r) of the Yang Baxter
cannot be a two-sided brace.
Let (B, +, ·) be a left brace. As usual, for any a, b ∈ B and positive integer m,
ab will denote a · b and am will denote a · a · · · a (where a appears m times).
Lemma 1. Let B be a left brace whose additive group (B, +) is torsion-free.
Assume that a, b ∈ B and that there is an integer n(a, b) such that a ∗ (a ∗ (. . . a ∗
(a ∗ b) . . .)) = 0 (where a occurs n(a, b) times and b once in this equation). Assume
moreover that Lan = id for some integer n. Then a ∗ b = 0 or equivalently,
a · b = a + b.
Proof. Note that La (b) = a ∗ b + b, for a, b ∈ B. Let m be a positive integer. Let
e1 (a, b) = a ∗ b and em+1 (a, b) = a ∗ em (a, b), for a, b ∈ B. It can be proved by
induction on m that
L
am
(b) = b +
m
X
m
i=1
i
ei (a, b).
4
CEDÓ, GATEVA-IVANOVA, SMOKTUNOWICZ
Since Lan = id, we have
n
X
n
ei (a, b),
b = L (b) = b +
i
i=1
an
and thus
ne1 (a, b) = −
n
X
n
ei (a, b).
i
i=2
Hence
n
X
n
ei+k−1 (a, b),
nek (a, b) = −
i
i=2
(2)
for all positive integer k.
Suppose that e1 (a, b) = a ∗ b 6= 0. Let n(a, b) be the smallest positive integer
such that en(a,b) (a, b) = 0. Then, by (2),
n
X
n
ei+n(a,b)−2 (a, b) = 0.
nen(a,b)−1 (a, b) = −
i
i=2
Since (B, +) is torsion-free, we have that en(a,b)−1 (a, b) = 0, in contradiction with
the definition of n(a, b).
Let G be a group. Following the notation of [14], for g, h ∈ G, we denote by
[g, h] the element [g, h] = g −1 h−1 gh. Recall that the group G is an Engel group
if and only if for each g, h ∈ G there exists a positive integer n(g, h) such that
[[. . . [[g, h], h] . . . ], h] = 1, where h occurs n(g, h) times.
Theorem 2. Let B be a left brace such that its additive group (B, +) is torsionfree and [B : Soc(B)] = n < ∞. If the multiplicative group (B, ·) of the left brace
B is an Engel group, then B is a trivial brace, that is a · b = a + b, for all a, b ∈ B.
Proof. Let a ∈ B and c ∈ Soc(B). Note that
La (c) = ac − a = aca−1 a − a = aca−1 + a − a = aca−1 .
Hence
[a, c] = a−1 c−1 ac = a−1 c−1 a + c
= La−1 (c−1 ) + c = La−1 (−c) + c
= −La−1 (c) + c = −a−1 ∗ c − c + c
= −a−1 ∗ c = (a−1 ∗ c)−1 .
NILPOTENT BRACES AND THE YANG–BAXTER EQUATION
5
Hence [c, a] = a−1 ∗c. Therefore [[. . . [[c, a], a] . . . ], a] = a−1 ∗(. . . a−1 ∗(a−1 ∗c) . . . ).
Let b ∈ B. Since [B : Soc(B)] = n < ∞ and Soc(B) is an ideal of B, nb ∈ Soc(B).
Since (B, ·) is an Engel group, there exists a positive integer m such that
[[. . . [[(nb), a], a] . . . ], a] = 1,
(where a occurs m times). Hence
0 = 1 = a−1 ∗ (. . . a−1 ∗ (a−1 ∗ (nb)) . . . ) = n(a−1 ∗ (. . . a−1 ∗ (a−1 ∗ b) . . . )),
(where a−1 appears m times). But (B, +) is torsion free, hence
a−1 ∗ (. . . a−1 ∗ (a−1 ∗ b) . . . ) = 0,
where a−1 occurs m times.
By Lemma 1, a−1 ∗ b = 0. Therefore a ∗ b = 0, for all a, b ∈ B, or equivalently,
B is a trivial left brace.
We call a left brace B left nilpotent if B n = 0 for some n, where B n+1 = B ∗ B n
is the chain introduced by Rump in [15]. As a consequence of Lemma 1 and
Theorem 2, we have the following two results.
Theorem 3. Let (X, r) be a finite solution of the YBE. Assume that for each
a, b ∈ X there is a positive integer n = n(a, b) such that the equality a ∗ (a ∗ (. . . a ∗
(a ∗ b))) = 0 holds in G(X, r),(a occurs n times and b occurs once in this equality).
Then (X, r) is the trivial solution. In particular, if G(X, r) is a left nilpotent left
brace, then (X, r) is the trivial solution.
Proof. Since G(X, r)/ Soc(G(X, r)) ∼
= G(X, r) is a subgroup of the symmetric
group SymX of the finite set X, we have that [G(X, r) : Soc(G(X, r))] < ∞.
Hence, by Lemma 1, a b = La (b) = ab − a = a + b − a = b, for all a, b ∈ X. In
particular, (X, r) is the trivial solution.
It is known that any ordered abelian-by-finite group is abelian, see for example
[12, Section 4]. It is also known that any torsion-free nilpotent group is ordered
(see [13, Lemma 13.1.6]). Recall that if (X, r) is a finite solution of the YBE, then
G(X, r) is a torsion-free, solvable and abelian-by-finite group (see [8] and [10]).
Therefore, if G(X, r) is nilpotent, then it is abelian. In this case the canonical left
brace structure on G(X, r) is trivial and (X, r) is the trivial solution. We have the
following related result.
Theorem 4. Let (X, r) be a finite solution of the YBE. If the structure group
G(X, r) is an Engel group, then (X, r) is the trivial solution.
6
CEDÓ, GATEVA-IVANOVA, SMOKTUNOWICZ
Proof. This is a consequence of Theorem 2.
3. Right nilpotent left braces
Etingof, Schedler and Soloviev in [8] introduced the retract solution of a given
solution of the YBE. Let (X, r) be a solution of the YBE. The retract relation ∼
on the set X with respect to r is defined by x ∼ y if σx = σy , where σx (z) = x z.
Then the retraction of (X, r) is Ret(X, r) = ([X], r[X] ), where [X] = X/ ∼ and
r[X] ([x], [y]) = ([x y], [xy ]),
where [x] denotes the ∼-class of x ∈ X. We define Ret1 (X, r) = Ret(X, r) and
Retk (X, r) = Ret(Retk−1 (X, r)) for k > 1. A solution (X, r) of the YBE is called a
multipermutation solution of level m if m is the smallest nonnegative integer such
that the solution Retm (X, r) has cardinality 1; in this case we write mpl(X, r) = m.
Let B be a left brace. By B (m) we mean the chain of ideals introduced by Rump
in [15], so B (1) = B and B (n+1) = B (n) ∗ B. We say that B is right nilpotent if
there exists a positive integer n such that B (n) = 0.
Recall that if B is a left brace, then the map r : B × B −→ B × B defined by
r(a, b) = (La (b), L−1
La (b) (a)),
is a solution of the YBE. This is the solution of the YBE associated with the left
brace B (see [6]).
Proposition 5. Let B be a nonzero left brace and let (B, r) be its associated
solution of the YBE. Then the multipermutation level of (B, r) = m < ∞ if and
only if B (m+1) = 0 and B (m) 6= 0.
Proof. Note that Soc(B) = {b ∈ B | b ∗ a = 0 for every a ∈ B}.
First we shall prove the implication (mpl(B, r) = m) ⇒ (B (m+1) = 0 and
B (m) 6= 0). We use induction on m = mpl(B, r). Suppose mpl(B, r) = 1. Therefore, La (b) = a ∗ b + b = b which is equivalent to a ∗ b = 0 for all a, b ∈ B. It follows
that B ∗ B = 0, so B (2) = 0. But B is a nonzero left brace, hence B (1) = B 6= 0.
This gives the base for induction.
Suppose now that for all k, 1 ≤ k ≤ m − 1, the condition mpl(B, r) = k ≤ m − 1
implies B (k+1) = 0 and B (k) 6= 0. Assume that mpl(B, r) = m, then the retraction
Ret(B, r) = ([B], r[B] ) has multipermutation level m − 1.
Moreover, there is an isomorphism of left braces (or equivalently an isomorphism
of braided groups) B/ Soc(B) ∼
= [B] and Ret(B, r) is isomorphic to the solution
of the YBE associated with B/ Soc(B) ([15], [6], [9]). Hence by the inductive
NILPOTENT BRACES AND THE YANG–BAXTER EQUATION
7
assumption (B/ Soc(B))(m) = 0 and (B/ Soc(B))(m−1) 6= 0. This implies B (m) ⊆
Soc(B) and that B (m−1) is not a subset of Soc(B). Therefore B (m+1) = 0 and
B (m) 6= 0.
Now we prove the inverse implication: (B (m+1) = 0 and B (m) 6= 0) ⇒ (mpl(B, r) =
m).
The base for the induction is clear. Assume that for all k ≤ m the implication is
true. Suppose that B is a left brace such that B (m+2) = 0 and B (m+1) 6= 0. Recall
that B (m+2) = B (m+1) ∗ B, therefore (B/ Soc(B))(m+1) = 0. On the other hand
B (m+1) 6= 0 and B (m+1) = B (m) ∗ B imply (B/ Soc(B))(m) 6= 0. By the inductive
assumption mpl(Ret(B, r)) = m, and therefore, mpl(B, r) = m + 1. This proves
the proposition.
4. Embedding solutions and groups into finite braces and finite
rings
In this section we will show that a finite solution of the YBE can be embedded
(in an explicit way) into a finite left brace. Recall that it was shown in [9] that
there is a canonical one-to-one correspondence between left braces and symmetric
groups (in the sense of Takeuchi [17]). Therefore Proposition 6 also shows explicitly
how to embed a finite solution of the YBE into a finite symmetric group.
Proposition 6. Let (X, r) be a finite solution of the YBE. Then there is a finite
left brace (B, +, ·) such that X ⊆ B generates the additive group of B. Moreover,
(X, r) is a subsolution of the solution (B, σ) associated canonically with the left
brace B, that is, X is σ-invariant and r = σ|X×X is the restriction of σ on X × X.
Proof. Let G = G(X, r) be the structure group of the finite solution (X, r). We
know that the additive group of the left brace (G, +, .) associated with G is free
abelian with basis X and a b = La (b) = ab − a and ab = L−1
La (b) (a), for all a, b ∈ X.
Since X is finite, [G : Soc(G)] < ∞, say [G : Soc(G)] = n. Consider the set
I = {ng | g ∈ G(X, r)}. We claim that I is an ideal of the brace G, that is, I is a
normal subgroup of the multiplicative group (G, .) which is invariant with respect
to the left actions by elements of G, [5]. It is clear that I is an additive subgroup of
(G, +) and I ⊆ Soc(G(X, r)). Then for u, v ∈ I one has uv = u + u v = u + v ∈ I,
u−1 = −u ∈ I, so I is a subgroup of G. Let g, h ∈ G. Then
h(ng)h−1 = (h (ng))(hng )(h−1 ) = (h (ng))(hh−1 ) = h (ng) = n(hg) ∈ I.
Thus, I is a normal subgroup of (G, .) which is also invariant under the left action
by elements of G. Therefore I is an ideal of the left brace (G, +, .). It is not difficult
8
CEDÓ, GATEVA-IVANOVA, SMOKTUNOWICZ
to show that brace quotient B = G/I is a finite left brace of order nm , where m is
the cardinality of X. Observe also that for any two elements x, y ∈ X, x 6= y, one
has x − y ∈
/ I, since the additive group of (G, +) is free abelian with a basis X.
Now the restriction of the natural map G → G/I = B on the set X is injective.
The proposition has been proved.
At a conference in Porto Cesareo, B. Amberg mentioned that he and his collaborators first became interested in Jacobson radical rings because they gave them
a way to construct examples of triply-factorizable groups. Later, they found more
ways of constructing such examples. Triply factorized groups can be also used to
define braces; see [16, Theorem 18]. Interesting results on triply factorized groups
can be found in [1, 2, 3, 4, 16, 11]. Triply factorized groups are for example useful
for investigating the structure of normal subgroups of a group G = AB which is a
product of two subgroups. Several authors investigated connections between triply
factorized groups and nearrings [16], [11]. It might be interesting to investigate
the connections between nearrings and braces. We would like to pose a related
open question:
Question 1. Investigate whether there is any relation between nearrings and solutions of the YBE?
The multiplicative group of a brace A is also called an adjoint group of brace A.
Observe that [7, Corollary 3.6] asserts that every finite solvable group is a subgroup
of an adjoint group of some left brace. We also make the following simple remark
which follows from [5, Lemma 8.1] and [7, Corollary 3.8].
Remark. (Related to [7, Corollary 3.8] and [6, Lemma 8.1]) Every finite nilpotent
group is a subgroup of the adjoint group of a finite nilpotent ring.
Let p be a prime. By [6, Lemma 8.1], every finite p-group is isomorphic to
a subgroup of the adjoint group of a finite nilpotent ring R such that R has
cardinality a power of p. Let G be a finite nilpotent group. Let p1 , . . . pm be the
the distinct prime divisors of the order of G. Let Pi be the Sylow pi -subgroup of
G. Then Pi is isomorphic to a subgroup of the adjoint group of a finite nilpotent
∼ P1 × · · · × Pm , it is clear that G is isomorphic to a subgroup
ring Ri . Since G =
of the adjoint group of the finite nilpotent ring R1 × · · · × Rm . If R is a ring, then
the adjoint semigroup of R is defined by a ◦ b = a + b + a · b.
By following the technique of the proof of [6, Lemma 8.1] we get the following
result.
NILPOTENT BRACES AND THE YANG–BAXTER EQUATION
9
Proposition 7. Let G be a group and let R be a ring (with unit). Then G is
isomorphic to a subgroup of the adjoint semigroup of the group ring R[G].
Proof. Let f : G −→ R[G] be the map defined by f (g) = g − 1, for g ∈ G. Clearly
f is injective. Let g, h ∈ G. We have that
f (gh) = gh − 1 = (g − 1)(h − 1) + g − 1 + h − 1 = (g − 1) ◦ (h − 1) = f (g) ◦ f (h).
Therefore f is an injective homomorphism of semigroups from G into the adjoint
semigroup of R[G].
Acknowledgments
The research of the first author was partially supported by grants DGI MICIIN
MTM2011-28992-C02-01 and MINECO MTM2014-53644-P. The research of the
second author was partially supported by Grant I 02/18 ”Computational and Combinatorial Methods in Algebra and Applications” of the Bulgarian National Science
Fund, by The Abdus Salam International Centre for Theoretical Physics (ICTP),
Trieste, and by Max-Planck Institute for Mathematics, Bonn. The research of the
third author was supported with ERC grant 320974.
References
[1] B. Amberg, Triply factorized groups in the Groups St Andrews 1989, Vol 1, London Mathematical Society Lecture Note Series, Cambridge University Press 1991, 1–13.
[2] B.Amberg, S. Franciosi, F. de Giovanni, Triply factorized groups, Communications in Algebra Volume 18, Issue 3, 1990, 789–809.
[3] B. Amberg, L. Kazarin, Nilpotent p-algebras and factorized p-groups, Proceedings of
”Groups St. Andrews 2005” , London Math. Society Lecture Note Series 339, Vol 1, Cambridge Univ. Press (2007), 130–147.
[4] B. Amberg, Ya.P. Sysak, Radical rings and products of groups, Groups St Andrews 1997 in
Bath: Edited by C. M. Campbell, 1–19.
[5] Ferran Cedó, Eric Jespers, Jan Okniński, Braces and the Yang-Baxter Equation,
arXiv:1205.3587v1 [math.RA].
[6] Ferran Cedó, Eric Jespers, Jan Okniński, Braces and the Yang-Baxter Equation, Communications in Mathematical Physics April 2014, Volume 327, Issue 1, pp 101–116.
[7] Ferran Cedó, Eric Jespers, Ángel de Rı́o, Involutive Yang-Baxter groups, Trans. Amer.
Math. Soc. 362 (2010), 2541–2558.
[8] P. Etingof, T. Schedler and A. Soloviev, Set-theoretical solutions to the quantum YangBaxter equation, Duke Math. J. 100 (1999), 169–209.
[9] Tatiana Gateva-Ivanova, Set-theoretic solutions of the Yang-Baxter equation, Braces, and
Symmetric groups, arXiv:1507.02602 [math.QA].
[10] T. Gateva-Ivanova and M. Van den Bergh, Semigroups of I-type, J. Algebra 206 (1998),
97–112.
10
CEDÓ, GATEVA-IVANOVA, SMOKTUNOWICZ
[11] Peter Hubert, Nearrings and a Construction of Triply Factorized Groups, doctoral dissertation, Johannes Gutenberg-Universitat in Mainz, 2005.
[12] Eric Jespers and Jan Okninski Binomial semigroups, J. Algebra 202 (1998), 250–275.
[13] D.S. Passman, The Algebraic Structure of Group Rings, Wiley, New York, 1977.
[14] D. J. S. Robinson, A Course in the Theory of Groups, Springer-Verlag GTM Vol. 80 (1982).
[15] Wolfgang Rump, Braces, radical rings, and the quantum Yang-Baxter equation, Journal of
Algebra Volume 307, Issue 1, 1 January 2007, Pages 153–170.
[16] Ya. P. Sysak, Products of groups and local nearrings, Note. Mat. 28 (2008), suppl. n. 2.
177-214.
[17] M. Takeuchi, Survey on mached pairs of groups. An elementary approach to the ESS-LYZ
theory, Banach Center Publ. 61 (2003), 305–331.
Ferran Cedó, Departament de Matemàtiques, Universitat Autònoma de Barcelona,
08193 Bellaterra (Barcelona), Spain
Tatiana Gateva-Ivanova, American University in Bulgaria, 2700 Blagoevgrad, and
Institute of Mathematics and Informatics, Bulgarian Academy of Sciences, 1113 Sofia,
Bulgaria
Agata Smoktunowicz, School of Mathematics, The University of Edinburgh, James
Clerk Maxwell Building, The Kings Buildings, Mayfield Road EH9 3JZ, Edinburgh
E-mail: [email protected], [email protected], [email protected]
| 4math.GR
|
Constructing Top-k Routes with Personalized
Submodular Maximization of POI Features
Hongwei Liang,
Ke Wang
School of Computing Science, Simon Fraser University, Canada
Abstract—We consider a practical top-k route problem: given
a collection of points of interest (POIs) with rated features and
traveling costs between POIs, a user wants to find k routes from
a source to a destination, that maximally match her needs on
feature preferences and can be completed within a travel cost
budget. One challenge is dealing with the personalized diversity
requirement where each user has a different trade-off between
quantity (the number of POIs with a specified feature) and variety
(the coverage of specified features). Another challenge is the large
scale of the POI network and the great many alternative routes
to search. We model the personalized diversity requirement by
the whole class of submodular functions, and present an optimal
solution to the top-k route problem through an index structure
for retrieving relevant POIs in both feature and route spaces
and various strategies for pruning the search space using user
preferences and constraints. We also present heuristic solutions
and evaluate all the solutions on real life POI network data.
I. I NTRODUCTION
The advancement in location-acquisition on mobile devices
has triggered a revolution in location based services [1]. One
emerging thread is route planning, with applications such as
trip recommendation, ride sharing, and intra-city delivery. The
majority of current route planning systems focuses on shortest/least cost paths or explores points of interest (POIs) that
are popular or geographically close [2], [3]. Other works [4],
[5] recommend routes based on learnt user’s preferences from
historical data. The popular travel planner, Google Trips, only
suggests day plans traversing famous places or user selected
POIs, and it is not able to respond a user’s detailed requirement
such as a trip containing museum and park. A relatively
understudied problem that has many practical applications is
that, a user wants to be suggested a small number of routes that
best meet her requirements while satisfying certain constraints.
Let us consider several such applications.
Scenario 1: A new visitor to Rome wants to well spend
the remaining hours before taking the flight back home. She
wishes to be recommended a trip, starting from her hotel
and ending at the airport, that allows her to visit museums,
souvenir shops, and eat at some good Italian restaurants (not
necessarily in this order). She values the variety over the
number of places visited. For example, a route consisting of
one museum, one shop, and one Italian restaurant is preferred
to a route consisting of two museums and two shops. Scenario
2: A home seeker wants to plan a route to see open houses
that have many of these features: 3 bedrooms, close to schools,
and within the 500k budget. Scenario 3: A ride-sharing driver
wants to be suggested a sequence of pickups and the route
( 0, 1, 0 )
Feature vector: ( park, museum, food )
v1
( 0, 0.8, 0 )
v4
14
( 0, 0.4, 0.6 )
v6
3
arXiv:1710.03852v1 [cs.SI] 10 Oct 2017
[email protected], [email protected]
( 0, 0, 0.9 )
v5
( 0.7, 0, 0 )
( 1, 0, 0.2 )
v2
5
v3
Fig. 1: A sample POI map. Each nodes vi represents a POI.
Each POI has 3 features, (park, museum, food), with each
feature having a numeric rating in the range [0, 1], indicated
by the vector aside the POI. Each edge has a weight indicating
the cost of traveling the edge, e.g., traveling time.
from the previous pickups to the next, to optimize a certain
objective that integrates factors factors such as profit and
driving conditions and allow him to have dinner in time.
A. Top-k Route Problem
The above examples illustrate some common structures and
requirements of various emerging route planning applications.
First, the database is a POI map containing a collection of
POIs connected by edges weighted by traveling cost between
POIs, where each POI has a location and a set of features
(e.g., museum, 3 bedroom) with each feature having a numeric
or binary rating. The features and ratings on a POI are
ubiquitous and easily accessible in the real world. For instance,
Foursquare encourages its users to rate a venue with a numeric
score on features such as cleanliness and write text Tips.
Features and ratings can also be extracted from check-ins and
user provided text reviews [6].
Second, the user wants to find top-k routes through a
query Q with several parameters: a pair (x, y) of origin
and destination of a route, a weight vector w specifying
her personalized weight wh of each feature h, and a budget
constraint b on the total travel cost of a route. In addition, each
user has a personalized route diversity requirement in terms
of the “quantity vs. variety” trade-off for the POIs on a route
(i.e., Scenarios 1 and 2), which can be modeled by a carefully
chosen aggregation function Φh for each feature h over the
h’s scores of the POIs PV on a route P. The match degree of
a route P with Q can be defined by a gain function
X
Gain(PV ) =
wh Φh (PV ).
h
The top-k routes have the highest Gain(PV ) while satisfying
the origin and destination constraints with the cost budget b.
•
B. Challenges
One challenge of modeling the top-k route problem is
choosing a general aggregation function Φh that can model a
range of personalized route diversity requirement. Intuitively,
the sum aggregation Φh returns the sum of the scores on h
of all POIs on the input route, whereas the max aggregation
Φh returns the max score on h of all POIs on the input
route. These choices present the two extremes of the “quantity
vs variety” trade-off, they cannot model the other trade-offs.
Consider the POI map in Figure 1 and a query Q specified
by x = v1 , y = v5 , w = (0.4, 0.6, 0) for (park, museum,
food), and b = 18. The sum aggregation prefers the route
v1 → v6 → v4 → v5 , though this route fails to cover the
specified feature park. The max aggregation does not distinguish the two routes v1 → v3 → v5 and v1 → v2 → v3 → v5 ,
as both have v1 and v3 with the maximum score for museum
and park respectively, though the latter may be preferred by
some user owing to the additional v2 .
The second challenge is the large search space of the
top-k route problem. Even with a single feature, the top-k
route problem subsumes the NP-hard orienteering problem
[7]. While heuristic solutions could speed up the search, they
provide no guarantee on optimality. The complexity of the
top-k route problem is practically compounded by the feature
space and the route space, which makes it quite meaningful to
develop efficient exact algorithms that provide optimal results
in reasonable responding time. The key is to push the user’s
requirements into the search to prune unpromising routes and
to speed up the shortest distance computation through precomputed index structures. Furthermore, this approach should
work for a wide range of the aggregation function Φh to model
the personalized route diversity requirement.
C. Contributions
For a variety minded user, it is important to cover as
many specified features as possible, so a desirable Φh should
diminish the incremental value of an additional POI over a
repeating feature h. Such diminishing marginal utility properties can be naturally modeled by a submodular set function
Φh . While submodularity has been widely used in many real
world problems such as [8][9], to our knowledge it has not
been used to model the route diversity requirement as in our
problem. Our contributions are as follows.
• We formulate the top-k route problem with the personalized route diversity requirement by a submodular aggregation function Φh as a query parameter. We show that
the power-law function fulfils the submodularity of such
Φh , where the personalized route diversity requirement
can be specified by the law parameter. (Section III)
• We propose a novel structure to index POIs by both
feature scores and travel cost. This component eliminates
irrelevant POIs, the expensive I/O access, and travel cost
computation at query time. (Section IV)
•
•
We propose an optimal solution, PACER, to the topk route problem. The novelty is the design of various
strategies for reducing computation and search space that
work for any submodular function Φh , including a route
enumeration strategy that facilitates reuse of computed
results, a cost-based pruning strategy that eliminates dominated routes, and an utility-based upper bound pruning
strategy. (Section V)
We present two heuristic algorithms that have far smaller
search spaces with a good utility. (Section VI)
We evaluate our algorithms analytically and empirically,
which suggests that PACER is significantly faster than the
state-of-the-art algorithms. (Section VII)
II. R ELATED W ORKS
Orienteering Problem (OP). OP [7] finds a path, limited
in length, that visits some vertices and maximizes the total
points collected from the vertices on the path. Arc Orienteering
Problem (AOP) [10] associates the utility with edges instead
of nodes. Our top-k route problem is an generalization of OP
by modeling each node as POI that has features and ratings,
and modeling the personalized user preference and diversity
requirement. Our problem also generalizes AOP because edge
utility can be modeled by inserting a dummy POI on each
edge. However, it is not clear that a solution to OP or AOP
can solve our problem.
Next POI or Travel Package Recommendation. These
works either recommend the POI to be visited next [11] or
recommend a set of POIs [12] [13]. They are quite different
from our aim that find a route as a sequence of featured POIs.
Trajectory Search. The works in this category require that a
database of trajectories is collected, and either retrieve existing
(segments of) trajectories that match certain similarity query
[14] [15], or constructs/recommends a route based on segments
of trajectories [16] [17]. Our top-k problem assumes a POI
map, instead of a trajectory database, for route construction.
Sequential Location Recommendation/Planning. This
body of works suggests a POI sequence or a location path to
a user. [4] learns from historical travel behaviors and recommends travel routes by sequentially predicting the next location
using a Markov model. This approach works only for existing
users who have historical data. [18], [5] interactively plans
a route based on user feedback or selection. Our approach
specifies the desired routes through a user query, without
requiring historical data or interactive feedbacks from the user.
Several works recommend a route by maximizing certain
user satisfaction. [19] assumes that each POI has a single type
and searches for a route with POIs following a pre-determined
order of types. [20] allows the user to specify a minimum
number of POI types, instead of exact types, in a route.
[21] uses user’s historical photos to estimate personalized
ratings and stay time on POIs. [22] estimates temporal-based
user preferences. None of these works considers user-specific
feature preferences or queries like ours.
[23] finds a route that maximizes the number of POI
having the desired keywords given a distance threshold. [24]
constructs an optimal route covering user-specified categories
of locations. Since the coverage is through containing the
keyword associated with locations, visiting one park is considered as good as visiting two parks. Neither of them addresses
personalized diversity requirement of users.
[25], perhaps most related to our work, adopts a keyword
coverage function to measure the degree to which a set of
query keywords are covered by a route, similar to ours.
However, their pruning strategies are designed specifically for
their specific keyword coverage function. Thus their work
does not address the personalized route diversity requirement
where a different submodular function may be required for
each user. Our pruning strategies depend on the submodularity
of the function Φh , but not the specific form of Φh ; thus,
our approach supports the personalized route diversity requirement. Finally [25] produces a single route and its empirical
performance does not surpass the brute-force method by much.
III. P RELIMINARY
We formally define the problem studied in this paper. Table
I summarizes the notations frequently used throughout the
paper. The variables in bold-face are vectors or matrices.
A. Problem Statement
Definition 1: [A POI Map] A POI map G = (V, E) is
a directed/undirected and connected graph, where V is a set
of geo-tagged POI nodes and E ⊆ V × V is a set of edges
between nodes (i, j), i, j ∈ V. H is a set of features on POIs.
F ∈ R|V|×|H| denotes the POI-feature matrix, where Fi,h ∈
[0, β] is the rating on a feature h for the POI i. Each POI
i ∈ V is associated with a staying cost si . Each edge ei,j ∈ E
has a travel cost ti,j .
The choices of si and ti,j depend on applications and can
be time, expenses, or other cost.
Definition 2: [Routes] A route P is a path x → · · · i · · · →
y in G from the origin x to the destination y through a
sequence of non-repeating POIs i except possibly x = y. PV
denotes the set of POIs on P. Ti,j denotes the least traveling
cost from i to the next visited j, where i, j are not necessarily
adjacent in the POI map. The cost of P is defined as
X
X
cost(P) =
si +
Ti,j .
(1)
i∈PV
i→j∈P
A route P contains only the POIs i that the user actually
“visits” by consuming the staying time at i. Each i → j on a
route is a path from i to j with the least traveling cost Ti,j . Any
POI on such a path other than i and j serves an intermediate
node to go from i to j and will not be visited by the user.
The staying times at x and/or y can be either considered or
ignored depending on the user choice. The latter case can be
modeled by setting sx = sy = 0.
At the minimum, the user has an origin x and a destination y
for a route, not necessarily distinct, and a budget b on the cost
of the route. In addition, the user may want the POIs to have
certain features specified by a |H|-dimensional vector w with
wh being the weight of feature h, where 1 ≤ h ≤ |H|. wh ∈
[0, 1] and Σh wh = 1. The user can also specify a filtering
TABLE I: Nomenclature
Notation
G = (V, E )
H
si
F ∈ R|V|×|H|
Fi,h
ti,j
Ti,j
P, PV
Q=
(x, y, b, w, θ, Φ)
VQ
n
F̃i,h
Gain(PV , Q)
topK
Interpretation
POI map G with node set V and edge set E
feature set on POIs
staying cost on POI i ∈ V
POI-feature matrix
the rating on feature h ∈ H for POI i
the traveling cost on edge ei,j ∈ E
the least traveling cost from any POI i to any j
route P with the included POI set PV
user query with parameters:
x and y – source and destination location
b – travel cost budget
w ∈ R|H| – feature preference vector
θ ∈ R|H| – filtering vector on feature ratings
Φ – feature aggregation functions
POI candidates set retrieved by Q
size of VQ
Fi,h after filtered by θ
gain of a route P given query Q
the found top-k routes
vector θ so that Fi,h is set to 0 if it is less than θ h . F̃i,h
denotes Fi,h after this filtering. Finally, the user may specify
a route diversity requirement through a feature aggregation
function vector Φ, with Φh for each feature h. Φh (PV ) returns
the aggregated rating on feature h over the POIs in PV .
Definition 3: [Query and Gain] A user query Q is a 6tuple (x, y, b, w, θ, Φ). A route P is valid if it starts from x
and ends at y, and cost(P) ≤ b. The gain of P w.r.t. Q is
X
Gain(PV , Q) =
wh Φh (PV ).
(2)
h
Note that only the specification of x, y, b is required; the
specification of w, θ, Φ is optional, and if not provided by
a user, their default choices can be used. Gain(PV , Q) is
a set function and all routes P that differ only in the order
of POIs have the same Gain, and the order of POIs affects
only cost(P). Among all routes having the same POI set, we
consider only the route having the smallest cost(P).
Definition 4: [Top-k route problem] Given a query Q and
an integer k, find (construct) k valid routes P that have
different POI sets PV and the highest Gain(PV , Q) (if ties,
ranked by cost(P)). The k routes are denoted by topK.
In the rest of the paper, we use Gain(PV ) for
Gain(PV , Q).
B. Submodular Function Φh
To address the personalized route diversity requirement, we
consider a submodular Φh to model the diminishing marginal
utility as more POIs with feature h are added to a route. A set
function f : 2V → R is submodular if for every X ⊆ Y ⊆ V
and v ∈ V \ Y , f (X ∪{v})− f (X) ≥ f (Y ∪{v})− f (Y ), and
is monotone if for every X ⊆ Y ⊆ V , f (X) ≤ f (Y ). The
next theorem follows from [26] and the fact that Gain(PV )
is a nonnegative linear combination of Φh .
Theorem 1: If Φh (PV ) is nonnegative, monotone and submodular, so is Gain(PV ).
1
=0
= 0.5
=1
=2
= 100
Coeff
0.8
0.6
0.4
0.2
0
1
2
3
4
5
6
7
8
Rank
Fig. 2: Power law coefficient R−α vs. rank R for varied α
We aim to provide a general solution to the top-k route
problem for the whole class of nonnegative, monotone and
submodular Φh , which model various personalized route diversity requirement. To illustrate the modeling power of such Φh ,
for example, consider Φh defined by the power law function
X
Φh (PV ) =
Rh (i)−αh F̃i,h ,
(3)
i∈PV
where Rh (i) returns the rank of POI i on the rating of feature
h among all the POIs in PV (the largest value ranks the first),
and αh ∈ [0, +∞) is the power law exponent for feature
−α
h. Rh (i) h is non-increasing as Rh (i) increases. Note that
Rh (i) for all existing POIs i will change when a new POI
j is added to the route P, so it is incorrect to compute the
new Φh (PV ) by simply adding the marginal brought by j to
existing value of Φh (PV ). For ease of presentation, we drop
the subscript h and use α for αh in the rest of the paper.
−α
Figure 2 shows how Rh (i)
varies as the rank increases
for different α. Note that the sum aggregation (α = 0) and the
max aggregation (α = ∞) are the special cases. In general,
a larger α > 0 means that the lower ranked POIs have a
faster diminishing factor for their feature ratings F̃i,h , i.e., a
diminishing incremental value on h. Hence, Eqn. (3) supports
a spectrum of diversity requirement ranging from the sum
aggregation to the max aggregation through the setting of α.
Theorem 2: Φh (PV ) defined in Eqn. (3) is nonnegative,
monotone and submodular.
Proof: The nonnegativity and monotonicity of Φh (PV ) in
Eqn. (3) is straightforward. We show its submodularity below.
Let X and Y be the set of POIs contained in two routes, and
X ⊆ Y . The POIs in X and Y are arranged in descending
order of F̃i,h . Let v ∈ V \ Y so that X ′ = X ∪ {v} and
Y ′ = Y ∪ {v}. Let ∆X = Φh (X ′ ) − Φh (X) and ∆Y =
Φh (Y ′ ) − Φh (Y ). It suffices to show ∆X ≥ ∆Y .
If X = Y , the proof is straightforward. Thus, we assume
X ⊂ Y . We also assume that Y contains exactly one more
POI than X, say y. The general case of containing l > 1
POIs can be proved by repeating the argument for the assumed
case l times. Besides, we assume v has feature h; otherwise,
F̃v,h = 0, then ∆X ≥ ∆Y is always true.
Given v with feature h, the rank of an existing POI i is
reduced after inserting v if F̃i,h < F̃v,h . Thus, ∆X or ∆Y
consists of two parts, i.e., the increment brought in by v, and
the decrement of existing POIs caused by their reduced ranks.
As Y includes one extra POI y compared to X, v’s rank in
X ′ and Y ′ can only have the following two cases.
Case i: If the ranks of v in X ′ and Y ′ are the same, y
must be ranked lower than v in Y ′ . In this case, the increment
brought in by v is the same in X ′ and Y ′ but the decrement
in Y ′ is larger owing to the extra y. Hence, ∆X ≥ ∆Y .
Case ii: If the ranks of v are different in X ′ and Y ′ , y must
be ranked ahead of v in Y ′ , and the POIs ranked lower than v
in X ′ and Y ′ are the same. Suppose that the rank of v in X ′ is
p, then in Y ′ is p + 1. After the insertion of v, the ranks of the
POIs ranked higher than v remain unchanged, but the rank of
each POI ranked lower than v falls by one. Thus, let Xj (resp.
Xj′ ) denote the POI whose rank on feature h is j in X (resp.
′
X ′ ), we have Xp = Xp+1
. Then the increments on rank p is
−α
′
∆X(p) = p (Xp − Xp ). Similarly, for Y ′ the increments
′
on rank p + 1 is ∆Y (p + 1) = (p + 1)−α (Yp+1
− Yp+1 ). As
′
′
Xp = Yp+1 = v, Xp = Yp+1 , we get ∆X(p) > ∆Y (p + 1).
By this analogy, we get ∆X(p + i) > ∆Y (p + 1 + i) for the
subsequent ranks. Hence, ∆X > ∆Y .
In the rest of the paper, we assume that Φh is nonnegative,
monotone, and submodular, so is Gain(PV ) (Theorem 1). The
top-k route problem is NP-hard, as it subsumes two NP-hard
problems, i.e., the submodular maximization problem and the
orienteering problem. Thus, for a POI map, the feature space
and path space must be carefully indexed, and novel search
strategies are necessary to prune unpromising sub-spaces as
much as possible. We first consider an indexing strategy in
Section IV and then consider search strategies in Section V.
IV. OVERVIEW
AND I NDEXING
Our algorithm has the offline component and the online
component. Before processing any query, the offline component builds the indices based on the POI map for speeding
up POI selection and travel cost computation. The online
component responds to the user query Q with POI Candidates
Retrieval that retrieves the sub-indices relevant to Q, and
Routes Finding that searches for the top-k routes using the
sub-indices. In this section, we explain the offline component
and POI Candidates Retrieval of the online component. Routes
Finding will be presented in later sections.
A. Offline: Indexing
The POI map data is stored on disk. To answer user
queries rapidly with low I/O access and speed up travel cost
computation, we build two indices, FI and HI stored on disk.
FI is an inverted index mapping each feature h to a list of
POIs having non-zero rating on h. An entry (vi , Fi,h ) indicates
the feature rating Fi,h for POI vi , sorted in descending order
Sample POI map
2-Hop Index (HI)
v1: { (v1,0) (v2,12) (v6,12) (v4,14) }
v2: { (v2,0) (v6,4) (v3,5) }
Feature Index (FI)
v3: { (v3,0) (v4,4) (v6,5) }
park: { (v3,1) (v2,0.7) }
v4: { (v4,0) (v6,3) }
museum: { (v1,1) (v4,0.8) (v6,0.4) }
v5: { (v5,0) (v3,1) (v4,3) (v6,6) }
food: { (v5,0.9) (v6,0.6) (v3,0.2) }
v6: { (v6,0) }
Fig. 3: FI and HI built from the POI map in Figure 1.
of Fi,h . FI helps retrieving the POIs related to the features
specified by a query.
The least traveling cost Ti,j between arbitrary two POIs
i and j is frequently required in the online component. To
compute Ti,j efficiently, we employ the 2-hop labeling [27] for
point-to-point shortest distance querying on weighted graphs.
[27] shows scalable results for finding 2-hop Labels for both
unweighted and weighted graphs. Our HI index is built using
2-hop Labels generated by their algorithm.
HI. For an undirected graph, there is one list of pivot labels
for each node vi , where each label (u, d) contains a pivot node
u and the traveling cost d between vi and u. HI(vi ) denotes
the list of labels for vi , sorted in the ascending order of d.
According to [27], Ti,j between vi and vj is computed by
FI
HI
Q = (v6, v2, 13, ( 0.5, 0.5, 0 ), 0.6, 1)
Query specific 2-Hop Index (HIQ)
v1: { (v1,0) (v2,12) (v6,12) }
v2: { (v2,0) (v6,4) (v3,5) }
Query specific Feature Index (FIQ)
Q
v3: { (v3,0) (v4,4) (v6,5) }
park: { (v3,1) (v2,0.7) }
v4: { (v4,0) (v6,3) }
museum: { (v1,1) (v4,0.8) }
v6: { (v6,0) }
Fig. 4: Retrieve POI candidates VQ . Given a query Q, retrieve
the subindex FIQ from FI and the subindex HIQ from HI.
V. F INDING O PTIMAL ROUTES
We now present the second part of the online component,
i.e., the algorithm for finding top-k routes based on the
Ti,j =
min
(d1 + d2 ).
(4) extracted POI candidate set VQ and the sub-indices FIQ and
(u,d1 )∈HI(vi )∩(u,d2 )∈HI(vj )
HIQ . The algorithm design has two goals in mind: i. prune unThe FI and HI for the POI map in Figure 1 are shown in promising routes as aggressively as possible while preserving
Figure 3. For example, to compute T2,5 , we search for the the optimality of the top-k answers; ii. ensure that the pruning
common pivot nodes u from the pivot label lists of v2 and v5 strategies are applicable to the whole class of nonnegative,
and find that v3 is the pivot node that minimizes the traveling monotone and submodular aggregation functions Φh . We procost between v2 and v5 , so T2,5 = 5 + 1 = 6.
pose a novel algorithm, called Prefix bAsed Compact statEs
In the case of a directed graph, each POI vi will have two gRowth (PACER), to incorporate the idea of memoization
lists of labels in HI, HI(viin ) for vi as the ending node, and (a variation of dynamic programming) [28] and fuse a costHI(viout ) for vi as the starting node, and Eqn. (4) becomes
based pruning strategy and a gain-based pruning strategy in
Ti,j =
min
(d1 + d2 ).
(5) an unified way. First, we introduce several terminologies.
(u,d1 )∈HI(viout )∩(u,d2 )∈HI(vjin )
A route P is associated with several variables: PV ,
Gain(P
V ), the ending POI end(P), and cost(P). If x is not
B. Online: POI Candidates Retrieval
visited (not consume the stay time), sx and F̃x,h for every
Given a query Q, the first thing is to retrieve the POI h are set to 0; the same is applied to y. A POI sequence is
candidates VQ that are likely to be used in the routes finding an open route if it starts from x and visits several POIs other
part. In particular, the POIs that do not contain any feature than y; it is a closed route if it starts from x and ends at y.
in the preference vector w or do not pass any threshold in The initial open route includes only x. An open route P is
θ will never be used, nor the ones that cannot be visited on feasible if its closed form P → y satisfies cost(P → y) ≤ b.
the way from the source x to the destination y within the In the following discussion, P denotes either an open route
budget b. This is implemented by retrieving the query specific or a closed route. An open route P − with ending POI i can
sub-indices FIQ from FI and HIQ from HI.
be extended into a longer open route P = P − → j by a POI
Figure 4 illustrates how the retrieval works for a query
j 6∈ PV− ∪ {y}. The variables for P are updated by
Q = (x = v6 , y = v2 , b = 13, w = (0.5, 0.5, 0), θ = 0.6, α = 1),
PV = PV− ∪ {j}
P
Gain(PV ) = h wh Φh (PV )
where the weights in w are for (Park, Museum, Food), and α
(6)
end(P) = j
is the power low exponent in Eqn. (3). Here the elements in
cost(P) = cost(P − ) + Ti,j + sj .
each vector θ and α have the same value for all features.
FIQ , a sub-index of FI, is retrieved using w and θ. w
Next, we present our enumeration and pruning strategies,
directly locates the lists for the user preferred (with wh > 0) followed by the algorithm and the complexity analysis.
features, i.e., Park and Museum. θ is used to cut off lower
rated POIs on the sorted lists indicated by red scissors. A. Compact State Growth with Memoization
VQ = {v1 , v2 , v3 , v4 } contains the remaining POIs.
Compact states C. PV and Gain(PV ) depend on the POI
HIQ , a sub-index of HI, is then formed by retrieving the set of the route P but are independent of how the POIs are
lists for each POI in VQ and also those for x and y, and b is ordered. Therefore, we can group all open routes sharing the
used to cut off the sorted lists, indicated by red scissors. We same PV as a compact state, denoted as C, and let CL denote
also check whether a POI i in current VQ is actually reachable the list of open routes having C as the POI set. C is associated
by checking the single-point visit cost: if sx +Tx,i +si +Ti,y + with the following fields:
sy > b, we remove i from VQ and remove its list from HIQ ,
Gain(C) : the gain of routes grouped by C
as indicated by the blue shading. The final POI candidates in
(7)
CL : ∀P ∈ CL , end(P), cost(P).
VQ are typically much less than that in the original V.
FIQ and HIQ are retrieved only once and kept in memory. These information is cached in a hash map with C as the key.
0
1
{A}→B
2
{A}
{A}→C
Ø
B. Cost-based Pruning Strategy
4
{B}
{B}→C
8
{C}
{D}
{A}→D 10
{B}→D
{C}→D
{BD}{D}→B 12{CD}{D}→C
{AB}{B}→A 5{AC}{C}→A 6{BC}{C}→B 9{AD}{D}→A
3
{AB}→D
{AB}→C
7
{ABC}{AC}→B
{BC}→A
11
{ABD}{AD}→B
{AC}→D
13
{ACD}{AD}→C
{BD}→A
{BC}→D
14
{CD}→A
15
{BCD}{BD}→C
{CD}→B
{ABC}→D
{ABD}→C
{ABCD}{ACD}→B
{BCD}→A
Fig. 5: The compact state enumeration tree for PACER. The
number aside each node indicates the order of enumeration.
We assume that the POIs in VQ are arranged in the lexicographical order of POI IDs. The compact states are enumerated
as the subsets of VQ represented by the nodes of a tree. x is
included in every compact state, so we omit x.
Figure 5 shows a compact state enumeration tree for VQ =
{A, B, C, D}, excluding x and y. Each capital letter represents
a POI, each node represents a compact state. We define the set
of POIs that precede i, in the above order, in a POI set as the
prefix of a POI i. The compact states are generated in a specific
prefix-first depth-first manner so that longer open routes are
extended from earlier computed shorter ones. Initially, the root
is labeled by the empty set ∅. A child node C of the current
node C− is generated by appending a POI i that precedes
any POIs in C− to the front of C− , and all child nodes are
arranged by the order of i. For example, Node 7 {ABC} is
generated as a child node of Node 6 {BC} by appending A
to the front of {BC} because A precedes B and C.
The routes in CL are generated by extending the cached
routes in the compact states C−j = C \ {j} where j ∈ C.
There are |C| such C−j . At node C, we generate each open
route P = P − → j by selecting the routes P − from each
C−j
L and append j at the end, and compute the gain and cost
of each P based on the accessed information for C−j from
the hash map. P is kept in CL if it is feasible.
For example, to generate the open routes at the node
{ABC}, we access the cached open routes at nodes {AB},
{AC} and {BC} and append the missing POI as the ending
POI. {AB} → C represents all the open routes that have C
as the ending POI and A and B as the first two POIs in any
order, i.e., x → A → B → C and x → B → A → C. Note
that there is no need to materialize the entire tree in memory,
but materialize only the current expanded branch of the tree.
A closed route P → y for each P ∈ CL is used to update
the top-k routes topK. If CL is empty, this compact state
is not kept. If no compact state is expandable, we stops the
enumeration and yield the final topK. Section V-D will give
the detailed algorithm. Note that CL can include |C|! open
routes and enumerating all the routes can be very expensive.
We present two strategies to prune unpromising routes.
Pruning-1: cost dominance pruning. Consider two feasible open routes P and P ′ . We say P dominates P ′ if they
have the identical POI set and the same ending POI and P has
no more cost than P ′ . Thus, if we can produce an extended
route P ′ → P̂, we can also produce an extended route P → P̂
using the same extension P̂, and the latter has a cost no larger
than the former. Note that P and P ′ not only have the same
POI set, but also have the same ending POI. Hence, at the
compact state C, when generating P = P − → j for a given
j, we only select the open route P − from C−j
L such that P
is feasible and has the least cost, thus, dominates all other
routes P ′− → j with P ′− from C−j
L . This reduces |C|! open
routes to at most |C| dominating open routes at the compact
state C, one for each j in C, without affecting the optimality.
For example, {AB} → C on node 7 {ABC} now represents
only one open route. We call this strategy cost dominance
pruning. We emphasize that Pruning-1 is a subtree pruning.
For example, if A → B on node 3 is pruned, all the open
routes starting with A → B, such as A → B → C on node
7 and A → B → D on node 11, will never be considered.
Though Pruning-1 prunes all dominated open routes, many of
the remaining dominating open routes may not lead to the topk closed routes due to the too small gain. Our next strategy
further prunes unpromising dominating routes.
C. Gain-based Pruning Strategy
We can extend a dominating open route P step by step using
the remaining budget ∆b = b − cost(P) into a closed route
P → P̂ with the gain Gain(PV ∪ P̂V ). The POIs used for
extension at each step should be reachable from the current
end(P), therefore, chosen from the set
U = {i|Tend(P),i + si + Ti,y + sy ≤ ∆b},
(8)
where i is an unvisited POI other than y. Tend(P),i and Ti,y
can be computed with the help of HIQ . The marginal gain by
concatenating P̂ to the existing P is
∆Gain(P̂V |PV ) = Gain(PV ∪ P̂V ) − Gain(PV ).
(9)
Let P → P̂ ∗ denote the P → P̂ with the highest gain. If
P → P̂ ∗ ranks lower than the current k-th top routes topK[k],
P is not promising and all the open routes extended from P
can be pruned. This motivates the next pruning strategy.
Pruning-2: marginal gain upper bound pruning. Finding
P̂ ∗ can be as hard as finding an optimal route from scratch,
so we seek to estimate an upper bound U P of the marginal
gain ∆Gain(P̂V |PV ), such that if Gain(PV ) + U P is less
than the gain of the current topK[k], P is not promising, thus,
P and all its extensions can be pruned without affecting the
optimality. We call this marginal gain upper bound pruning.
As more routes are enumerated, the gain of the current
topK[k] increases and this pruning becomes more powerful.
The challenge is that it is hard to estimate the cost of the
extended part P̂ without knowing the order of the POIs. As
∆Gain(P̂V |PV ) is independent of the order of the POIs, we
can ignore the order and estimate the “route cost” by the “set
cost”, i.e., the sum of some cost c(i) of each POI i on the
remaining route, where c(i) is no larger than i’s actual cost if
it is included on the remaining route. We define c(i) as:
min(tj,i ) + min(ti,k )
,
(10)
2
where tj,i is the cost on an in-edge ej,i and ti,k is the cost
on an out-edge ei,k . Since the order of POIs is ignored, it is
easy to verify that min ensures the above property of c(i). For
end(P) and the destination y, their cost is “one-sided”, i.e.,
c(end(P)) = min(tend(P),k )/2 and c(y) = sy + min(tj,y )/2,
with c(end(P)) has no si . (This “set cost” can further approach the “route cost” by replacing tj,i and ti,k with Tj,i and
Ti,k , respectively, and choosing j and k from U, but finding
min(Tj,i ) and min(Ti,k ) will incur the computation cost.)
The remaining task of estimating U P becomes optimizing
the following problem:
X
max ∆Gain(S|PV ) s.t.
c(i) ≤ B,
(11)
i∈S
S⊆U ∪{y}
c(i) = si +
where U is defined in Eqn. (8) and B = ∆b − c(end(P)).
Note that S should include y because P̂ is ended with y. Let
∆Gain(S ∗ |PV ) denote the maximum of ∆Gain(S|PV ). As
c(i) and c(end(P)) are no larger than their actual costs on P̂,
∆Gain(S ∗ |PV ) ≥ ∆Gain(P̂V |PV ) for any P̂. Thus, using
∆Gain(S ∗ |PV ) as U P never loses the optimality. To solve
Eqn. (11), we first show the properties of ∆Gain(P̂V |PV ).
Theorem 3: Assume that Φh ’s are nonnegative, monotone,
and submodular. For any given open route P and the extended
part of route P̂, ∆Gain(P̂V |PV ) is a nonnegative, monotone,
submodular, and polynomially computable function on P̂V .
Proof: We only show that ∆Gain(P̂V |PV ) is submodular
on P̂V , as the proof of other properties is straightforward.
According to [26], if a set function g : 2V → R is submodular,
and X, Y ⊂ V are any disjoint sets, the residual function
f : 2Y → R defined as f (S) = g(S ∪ X) − g(X) is also
submodular. Since Gain is submodular (Theorem 1) and since
PV and U are disjoint, ∆Gain(P̂V |PV ) = Gain(PV ∪ P̂V )−
Gain(PV ) is residual on P̂V , thus, is submodular.
Apparently, Eqn. (11) is a submodular maximization problem, which unfortunately is also NP-hard [26] and there is no
polynomial time exact algorithm for it unless N P = P . Thus,
we switch to estimate the upper bound of ∆Gain(S ∗ |PV ).
[29] showed that there exists a polynomial time greedy algorithm, presented in [30], that obtain an approximate solution
∆Gain(S ′ |PV ) for the above problem with approximation
ratio of 1 − e−1 . Then the upper bound of ∆Gain(S ∗ |PV ) is
achieved by ∆Gain(S ′ |PV )/(1 − e−1 ). However, the greedy
algorithm runs in Ω(B|U|4 ), where B is defined in Eqn. (11).
A simplified version of this algorithm runs in O(B|U|) but
only achieve the approximation ratio of 12 (1 − e−1 ).
The above mentioned offline bounds, i.e., 1 − e−1 and
1
(1
− e−1 ), can be stated in advance before running the actual
2
algorithm. Compared to the offline bounds, the authors in
[31] showed that they can use the submodularity to acquire
a tight online bound for arbitrary given solution Ŝ (Ŝ is
obtained using any algorithm) to the problem of maximizing
a submodular set function subject to a cost constraint. They
empirically show that this bound is much tighter than the
bound 12 (1 − e−1 ). By applying [31] to the problem in Eqn.
(11), we have the following theorem.
Theorem 4: For a set Ŝ ⊆ U ∪ {y}, and each POI i ∈ U ∪
{y} \ Ŝ, let δi = ∆Gain(Ŝ ∪ {i}|PV ) − ∆Gain(Ŝ|PV ). Let
ri = δi /c(i), and let i1 , · · · , im be the sequence of P
these POIs
l−1
with ri in decreasing order. Let l be such that C = j=1
c(ij )
Pl
≤ B and j=1 c(ij ) > B. Let λ = (B − C)/c(il ). Then
Xl−1
δij + λδil . (12)
∆Gain(S ∗ |PV ) ≤ ∆Gain(Ŝ|PV ) +
j=1
Theorem 4 measures how far ∆Gain(Ŝ|PV ) for any given
set Ŝ is from the optimum ∆Gain(S ∗ |PV ). As our purpose is
estimating the upper bound of ∆Gain(S ∗ |PV ), we let Ŝ = ∅.
Then ∆Gain(Ŝ|PV ) = 0 and δi = ∆Gain({i}|PV ). Hence,
Xl−1
(13)
δij + λδil ≥ ∆Gain(S ∗ |PV ).
UP =
j=1
U P is computed without running a greedy algorithm. We also
empirically proved that this online bound in Eqn. (13) outperforms the offline bounds on both tightness and computational
cost. Thus, we finally choose the online bound.
D. Algorithm
Algorithm 1: PACER(C−, I) (Recursive funcion)
Required : Q = (x, y, b, w, θ, Φ), VQ , FIQ and HIQ to
compute Gain(C) and cost(P), and k
Parameters: compact state C− and the set of POIs I for
extending C−
Output
: a priority queue topK
1 forall POI i in set I in order do
2
C ← {i} ∪ C− ;
3
compute Gain(C);
4
forall POI j in C do
5
C−j ← C \ {j};
−j
6
P − ← the dominating route in CL
such that
−
cost(P → j) is minimum; // prune-1
7
P ← P − → j;
8
if cost(P → y) ≤ b then
9
Compute U P using Eqn. (13);
10
if Gain(C) + U P ≥ Gain(topK[k]) then
11
insert route P into CL ; // prune-2
12
13
UpdateTopK(CL , topK);
PACER(C, prefix of i in I);
PACER(C− , I) in Algorithm 1 incorporates the above
compact state enumeration and pruning strategies. The global
input is a query Q, VQ , FIQ and HIQ used to compute Gain(C)
and cost(P) as in Eqn. (6), and an integer k. The output
is a priority queue storing the topK results. PACER(C− , I)
recursively enumerates the subtree at the current compact state
C− with the POI set I available for extending C− . The initial
call is PACER(∅, VQ ), when only x is included.
As explained in Section V-A, Line 1 - 3 extends C− by each
i in the set I in order, creating the child node C and computing
Gain(C). Lines 4 - 11 generate the dominating and promising
open routes CL . Specifically, each j ∈ C is selected as the
ending POI and the rest POIs C−j is a previously computed
−j
compact state. It finds the dominating route P − in CL
(Line
6). This corresponds to Pruning-1. P is further considered only
if it is feasible. Pruning-2 is applied to check if the sum of
Gain(C) and the estimated upper bound of the marginal gain
is no less than the gain of the k-th route in topK (Lines 9
- 11) and if so, P is inserted into CL . After CL is finalized,
the closed route P → y for an open route P in CL such that
P → y has the least cost is chosen to update topK (Line 12).
The information of the new compact state C, as in Eqn. (7),
is added to the hash map. At last, C is extended recursively
with the POIs in the prefix of i in I (Line 13).
Let us summarize the good properties of PACER as follows.
Properties of PACER. First, PACER relies on Φh being
nonnegative, monotone, and submodular, but is independent
of any form of Φh . This is important for dealing with the
personalized diversity requirement. Second, PACER enumerates open routes as compact states in a prefix-first depthfirst order, which enables constructing the open routes at
C incrementally from those at C \ {i}, i.e., memoization.
Third, each compact state represents at most |C| dominating
feasible open routes, instead of |C|! routes, i.e., Pruning-1.
Fourth, only the dominating feasible open routes that have
an estimated maximum achieved gain no less than that of
the current topK[k] are kept, i.e., Pruning-2. This pruning
is tightened up as more closed routes are enumerated.
E. Complexity Analysis
One reasonable measure of computational complexity is the
number of routes examined. Two main factors having the
impact on the number of routes examined are the size of
the POI candidate set, i.e., |VQ |, and the maximum length of
routes examined (excluding x and y), i.e., the maximum |P|.
Let n and p denote these parameters, respectively. p ≪ n.
We analyze PACER relatively to the brute-force search and a
state-of-the-art approximation solution.
PACER. The compact states on the l-th level of the enumeration tree compute the routes containing
l POIs, and thanks to
Pruning-1, there are at most nl compact states on level l and
each represents at most l routes, each computed only once.
Thus, with p ≪ n and the Pascal’s rule [32], the number of
routes examined is at most
p
p
X
X
n
n−1
n−1
n−1
l
=n
≈ n(
+
)
l
l−1
p−1
p−2
l=1
l=1
n!
n
n
.
=n
=
(n − p + 1)(p − 1)! (n − p)!
p−1
(14)
Therefore, the computation cost of PACER is
1
n!
O( (p−1)!
(n−p)! ). If Pruning-2 is also enabled and prunes the
γ percent of the routes examined by PACER with Pruning-1,
n!
1
the computation cost of PACER is O((1 − γ) (p−1)!
(n−p)! ).
Brute-force algorithm (BF). The brute-force algorithm
n!
based on the breadth-first expansion examines O( (n−p)!
)
routes, which is (p − 1)! times of that for PACER with only
Pruning-1. In general, the next POI visited in a route does not
have to be an immediate neighbor of the previous one.
Approximation algorithm (AP). [33] proposed a quasipolynomial time approximation algorithm for the Orienteering
Problem. AP can be modified to solve our problem. It produces
a single route with the approximation ratio ⌈log p⌉ + 1 at
the computation cost O((2n · b)log p ). An improved version
reduces the cost to O((n · OP T · log b)log p ), where OP T
is an estimated upper bound on the optimal Gain. However,
both costs remain expensive if b or OP T has many discrete
values. For example, for b = 512 minutes, n = 50, p = 8
and OP T = 10.0 (100 discrete values with the single decimal
point precision), the computation cost for the improved version
is (50·100·log 512)log 8 = 9.11×1013. [34] noted that AP took
more than 104 seconds for a small graph with 22 nodes. In
comparison, the computation cost of PACER with Pruning-1
9
given by Eqn. (14) is only 50 × 50
7 = 4.99 × 10 . This cost is
further reduced by Pruning-2. PACER finds the optimal top-k
routes whereas AP only finds single approximate solution. We
will experimentally compare PACER with AP.
VI. H EURISTIC S OLUTIONS
Although PACER employs carefully designed enumeration
and pruning strategies, it remains expensive when the constraints are too loose, such as a large cost budget b and a
large POI candidate set VQ . We need either approximation or
heuristic algorithms to further limit the search space when such
extreme cases arise. The state-of-art approximation algorithm
[33] as mentioned above is shown practically not scalable.
Therefore, in this section, we design two heuristics.
State collapse heuristic. The cost dominance pruning in
PACER keeps at most l open routes for a compact state
representing a set of l POIs (excluding x and y). A more
aggressive pruning is to keep only a single open at each
compact state, i.e., the one with the least cost, with the
heuristic that this route likely visits more POIs. Clearly, this
heuristic no longer guarantees the optimality but can produce
k routes. We denote this heuristic algorithm by PACER-SC,
where SC stands for “State Collapsing”.
Let n and p be defined as in Section V-E. The number of
examined routes is no P
more than
the number of nodes in the
p
enumeration tree, i.e., l=1 nl . Still, with p ≪ n, we have
p
X
n
n
n
n+1
≈
+
=
.
(15)
l
p
p−1
p
l=1
The computation cost is O( np ), which is around 1/p of that
for PACER in Eqn. (14). With Pruning-2 enabled in
PACERSC, this computation cost reduces to O((1 − γ) np ).
Greedy algorithm. PACER-SC’s computation complexity
remains exponential in the route length p. Our next greedy
algorithm runs in polynomial time. It starts with the initial
route x → y and iteratively inserts an unvisited POI i to the
current route to maximize the marginal gain/cost ratio
Gain({i} ∪ C) − Gain(C)
,
si + Tx,i + Ti,y
(16)
where C denotes the set of POIs on the current route. It inserts
i between two adjacent POIs in the current route so that the
total cost of the resulting route is minimized. The term Tx,i +
Ti,y constrains the selected POIs i to be those not too far away
from the two end points. The expansion process is repeated
until the budget b is used up. The algorithm only produces a
single route and examines O(pn) routes because each insertion
will consider at most n unvisited POIs.
VII. E XPERIMENTAL E VALUATION
We evaluated the proposed algorithms on two real life data
on a Ubuntu 16.04.1 LTS machine with Intel (R) Core (TM)
i7-3770 CPU @ 3.40 GHz and 16G of RAM. All algorithms
were implemented in C++.
A. Experimental Setup
1) Datasets: Singapore denotes the Foursquare check-in
data collected in Singapore, and Austin denotes the Gowalla
check-in data collected in Austin, previously used in [25].
Singapore has 189,306 check-ins at 5,412 locations by 2,321
users, and Austin has 201,525 check-ins at 6,176 locations
by 4,630 users. Same as suggested in [24], [25], we built
an edge between two locations if they were visited on the
same date by the same user. The POI map for Singapore has
1,625 locations, and that for Austin has 2,609 locations. The
locations not connected by edges were ignored. We filled in the
edge costs ti,j by querying the traveling time in minute using
Google Maps API under driving mode. The average ti,j for
Singapore is 16.24 minutes and for Austin is 11.12 minutes.
The staying time si were generated following the Gaussian
distribution, si ∼ N (µ, σ 2 ), with mean µ = 1.5 hours and
standard deviation σ = 0.25 hours. All costs are in minute.
We extracted 202 and 252 unique features, respectively, for
Singapore and Austin based on the user mentioned features
at check-ins, as suggested in [25]. The rating of a feature h
on POI i is calculated by
Fi,h = min{
N Ch (i)
β
P
× , β},
1/|Sh| × j∈Sh N Ch (j)
2
(17)
where N Ch (i) is the number of check-ins at POI i containing
the feature h, Sh is the set of POIs containing h, β is the
maximumP
feature rating and is set to β = 5 for both data sets.
1/|Sh | × j∈Sh N Ch (j) is the average number of check-ins
on h. This calculation scales the middle value β2 by the ratio
of a POI’s check-in count to the average check-in count on h.
2) Algorithms: We compared the following algorithms.
BF is the brute-force method as mentioned in Section V-E.
PACER+1 is our optimal algorithm in Section V with only
Pruning-1 (cost dominance pruning) enabled. PACER+2 enables both Pruning-1 and Pruning-2 (marginal gain upper
bound pruning). PACER-SC is the state collapse algorithm
in Section VI. GR is the greedy algorithm in Section VI.
AP is the approximation algorithm proposed by [33] (see
Section V-E). A* is the A* algorithm proposed by [25]. Since
A* works only for its specific keyword coverage function,
it is not compared until Section VII-D where we adapt our
submodular aggregation function to their coverage function.
For a fair comparison, all algorithms use the extracted subindices in Section IV-B to reduce I/O access and the search
space. Note that BF, PACER+1, PACER+2 and A* produce
an exact solution and PACER-SC, GR, and AP produce an
greedy or approximate solution.
3) Queries: A query Q has the six parameters x, y, b, w,
θ, Φ. For concreteness, we choose Φh in Eqn. (3) with α
controlling the diversity of POIs on a desired route. We assume
θh and αh are the same for all features h. For Singapore,
we set x as Singapore Zoo and y as Nanyang Technological
University; and for Austin, we set x as UT Austin and y as
Four Seasons Hotel Austin.
For each data set, we generated 50 weight vectors w
to model the feature preferences of 50 users as follows.
First, we draw the number of features for w, m, from an
uniform distribution, i.e., m ∼ U(1, 4). Then, we draw m
features
h following the probability distributions: Pr(h) =
P
i∈Sh N Ch (i)
P
P
. N Ch (i) and Sh are defined in Eqn. (17).
h∈H
i∈Sh N Ch (i)
For each selected feature h, Let HQ P
be the set of selected
features. For each h ∈ HQ , wh =
P
i∈Sh
h∈HQ
P
N Ch (i)
N Ch (i) .
i∈Sh
Finally, we consider b ∈ {4, 5, 6, 7, 8, 9} in hours, θ ∈
{0, 1.25, 2.5, 3.75}, and α ∈ {0, 0.5, 1, 2} with the default
settings in bold face. For each setting of b, θ, α, we generated
50 queries Q = (x, y, b, w, θ, α) using the 50 vectors w
above. b here is in hours, therefore, b = 5 specifies the budget
of 300 mins.
We first study the effect of indexing (Section VII-B), and
then evaluate the performance of our proposed algorithms
(Section VII-C), finally we compare with the A* algorithm
(Section VII-D).
B. Effect of Indexing
The POI Candidate Retrieval (Section IV-B) extracts two
sub-indices and the POI candidate set VQ relevant to the query
Q. The search space for Q is largely determined by the size of
VQ , denoted by n below. Table II compares the average n for
the 50 queries Q for each setting of θ with α and b fixed at
their default settings. Compared to the number of POIs in the
original data, i.e., 1625 for Singapore and 2609 for Austin,
the significantly reduced POI candidate set paves the way for
subsequent efficient online route search.
TABLE II: Average size of the POI candidate set VQ vs. θ
Singapore
Austin
θ=0
215.24
207.90
θ = 1.25
98.66
89.58
θ = 2.5
72.20
60.34
θ = 3.75
50.70
41.70
PACER+1
PACER+2
OPTIMAL
PACER-SC
BF
PACER+1
PACER+2
OPTIMAL
PACER-SC
PACER-SC
GR
AP
GR
AP
PACER-SC
GR
AP
GR
AP
7
8
9
4
8
9
101
10-1
2.5
(d) Runtime vs. θ
10-1
10-2
10-3
0
0.5
1
2
(g) Runtime vs. α
7
8
9
8
10
4
6
102
4
1.25
2.5
3.75
(e) # of routes vs. θ
5
103
10
1
0
0
1.25
2.5
0.5
1
2
(h) # of routes vs. α
4
0
0.5
1
2
(i) Gain vs. α
Fig. 6: Experimental results for Singapore. Run time and
search space (# of routes) are in logarithmic scale. The labels
beside data points indicate the ratio of queries succesfully
responded by the algorithm under the parameter setting. No
label if no query fail. Data point or bar is not drawn if more
than half fail. AP can only respond queries with small b.
C. Performance Study
We evaluate all algorithms under various settings of
(b, θ, α) chosen from the ranges mentioned above. In particular, we evaluate Gain, CPU runtime, and search space in the
number of examined open routes for processing a query, and
report the average for the 50 queries (i.e., vectors w) under
each setting of (b, θ, α). GR and AP cannot produce top-k
routes for k > 1, thus, we first set k = 1 in order to compare
all algorithms. The impact of k will be presented in Section
VII-C4.
Figure 6 and 7 report the experiments for Singapore and
Austin, respectively. Each row corresponds to various settings
of one of b, θ, α while fixing the other two at the default settings. OPTIMAL denotes the same optimal gain of PACER+2,
PACER+1 and BF. We terminated an algorithm for a given
query after it runs for 1 hour or runs out of memory, and
used the label beside a data point to indicate the percentage
of finished queries. If more than an half of the queries were
terminated, no data point is shown.
1) Impact of budget b: b affects the length of routes (the
number of POIs included). Figure 6a - 6c and 7a - 7c shows
the runtime, search space, and Gain.
AP is the worst. This is consistent with the analysis in Section V-E, which shows that AP suffers from a high complexity
when b and OP T have many discrete values. For example,
6
7
8
10
Gain
4
10-1
5
6
7
8
4
9
(d) Runtime vs. θ
100
10-1
10-2
10-3
0
0.5
1
2
(g) Runtime vs. α
6
7
8
9
(c) Gain vs. b
108
106
10
8
6
4
3.75
0
1.25
2.5
4
3.75
(e) # of routes vs. θ
101
5
10
46/50
102
2.5
4
Time budget b (hour)
(b) # of routes vs. b
10-3
1.25
8
Time budget b (hour)
1010
101
10
6
9
46/50
3
0
(f) Gain vs. θ
8
5
(a) Runtime vs. b
3.75
12
10
12
4
Time budget b (hour)
16
107
10
4
(c) Gain vs. b
10
0
# of routes
100
6
10
6
3.75
101
5
108
10-3
1.25
4
106
102
Time budget b (hour)
1010
0
Runtime (sec)
7
Gain
10
6
(b) # of routes vs. b
3
# of routes
Runtime (sec)
(a) Runtime vs. b
5
Time budget b (hour)
10
-3
14
37/50 40/50
108
Gain
6
10-1
# of routes
5
Time budget b (hour)
4
1
105
103
1
0
1.25
2.5
3.75
(f) Gain vs. θ
107
10
0
16
Gain
6
102
4
8
10
1010
37/50 40/50
# of routes
4
10
103
# of routes
10
Runtime (sec)
12
106
-3
14
31/50
Runtime (sec)
10
42/50
8
Runtime (sec)
1
10-1
10
1010
31/50
Gain
10
42/50
Gain
103
# of routes
Runtime (sec)
BF
0.5
1
2
(h) # of routes vs. α
12
8
4
0
0.5
1
2
(i) Gain vs. α
Fig. 7: Experimental results for Austin
the setting of b = 6 has 360 discrete values in minute, a
majority of the queries cannot finish. BF is the second worst.
The efficiency of BF drops dramatically as b increases, since
the number of open routes becomes huger and processing them
is both time and memory consuming.
PACER+1’s search space is two orders of magnitude smaller
than that of BF, thanks to the compact state enumeration
strategy and the cost dominance pruning. PACER+2 is the
best among all the exact algorithms. In comparing PACER+2
against PACER+1, there is one order of magnitude speedup
in runtime and two orders of magnitude reduction in search
space. This clearly demonstrates the additional pruning power
of the Gain based upper bound pruning. PACER-SC trades
optimality for efficiency. Surprisingly, as shown in Figure 6c
and 7c, PACER-SC performs quite well with Gain being close
to that of OPTIMAL.
GR always finishes in less than 10−2 seconds. For Singapore, the achieved gain is worse than that of OPTIMAL,
whereas for Austin, the difference is small. This is because the
source x and destination y for Singapore are relatively remote
to the central city. GR will greedily select a POIs i not too
far away from x and y, thus, many POIs with possibly higher
feature ratings located in the central city are less likely to be
chosen, which results in the larger gap of Gain for Singapore.
In contrast, x and y for Austin are in the downtown area and
this situation is avoided in most cases.
2) Impact of of filtering threshold θ: Figure 6d - 6f and 7d 7f show the performance versus the threshold θ. When θ = 0,
there is no feature cut-off. As θ becomes larger, the size of
POI candidate set is reduced and all the algorithms run faster.
The majority of the experiments for AP cannot finish and its
D: Esplanade Park
F: Tiong Shian Ea!ng House
E: Peach Garden
(a) α = 0.5 (with diversity requirement)
E: Bukit Timah Nature Reserve
D: Tampines Eco Green
C: Bedok Reservoir Park
Hilton Singapore
A: Na!onal Museum
B: Fort Canning Park
Hilton Singapore →
A (M:5.0): National Museum →
B (P:4.1): Fort Canning Park →
C (P:4.0): Bedok Reservoir Park →
D (P:5.0): Tampines Eco Green →
E (P:4.8): BukitTimah Nature Reserve →
Hilton Singapore;
Gain = 8.66 ; Cost = 535 mins
(b) α = 0 (without diversity requirement)
Fig. 8: Two routes found from Singapore by PACER+2 for the
query Q = (x, y, b = 9, w = (P : 0.4, M : 0.3, R : 0.3), θ =
2.5, α), where x and y are Hilton Singapore, and P, M and R
represent Park, Museum, and Chinese Restaurant.
results are not shown. The study suggests that a reasonable
value of θ, e.g., 2.5, reduces the searching cost greatly while
having little loss on the quality of the found routes.
3) Impact of diversity parameter α: Figure 6g - 6i and 7g
- 7i show the results for various settings of α that represent
user’s diversity requirements on the route. PACER+2 and
PACER-SC are slightly affected. As α increases, the marginal
return diminishes faster and Φh behaves more towards the max
aggregation. In this case, Pruning-2 becomes less effective.
When α = 0, Eqn. (3) becomes the sum aggregation and both
the value of Gain and the difference between OPTIMAL and
GR reach the maximum. We emphasize that it does not make
sense to compare the Gain value under different settings of α;
in fact, a different setting of α values things differently.
Figure 8 evaluates the effectiveness of our power law
function in Eqn. (3) for modeling the personalized route
diversity requirement. We run two queries on Singapore, one
with α = 0.5, which specifies a diversity requirement, and one
with α = 0, which specifies the usual sum aggregation. The
other query parameters are the same. The figures show the best
routes found for each query, with the POIs on a route labeled
sequentially as A, B · · · . The red dots represent the source x
and destination y. The route for α = 0.5 covers all specified
features, i.e., two POIs for each feature, while maximizing the
total Gain. While the route for α = 0 has four parks out of
five POIs, due to the higher weight of Park in w. Although
the Gain value of the second route, i.e., 8.66, is higher than
that of the first route, i.e., 7.34, it is less preferred by a user
who values diversity. In fact, the second route’s Gain value
when evaluated using α = 0.5 is only 6.60.
4) Impact of k: We vary k in range [1, 100] while fixing
b, θ, α at the default values and run the algorithms, except GR
and AP, on both datasets. As k only influences the gain-based
pruning, the performance of BF and PACER+1 are unchanged.
109
# of routes
PACER+2
A*
103
101
10-1
4
5
6
7
8
PACER+2
A*
107
105
103
9
4
Time budget b (hour)
(a) Runtime - Singapore
103
101
10-1
4
5
6
7
8
6
7
8
9
(b) # of routes - Singapore
109
39/50
PACER+2
A*
5
Time budget b (hour)
# of routes
B: Fort Canning Park
C: Singapore Art Museum
Runtime (sec)
A: Na!onal Museum
Hilton Singapore →
A (M:5.0): National Museum →
B (P:4.1): Fort Canning Park →
C (M:5.0): Singapore Art Museum →
D (P:3.1): Esplanade Park →
E (R:3.6): Peach Garden →
F (R:5.0): Tiong Shian Eating House →
Hilton Singapore;
Gain = 7.34115 ; Cost = 540 mins
Runtime (sec)
Hilton Singapore
9
Time budget b (hour)
(c) Runtime - Austin
10
39/50
PACER+2
A*
7
105
103
4
5
6
7
8
9
Time budget b (hour)
(d) # of routes - Austin
Fig. 9: PACER+2 vs. A* (logarithmic scale).
For PACER+2 and PACER-SC, the change is limited (less than
10% slower for k = 10 and less than 25% slower for k = 100).
Because when k is small, the Gain of the k-th best route is
usually not far away to that of the best route, thus, the marginal
gain upper bound pruning is not seriously influenced. We omit
the figures about varying k due to limited space.
Discussion. As is shown, in fact our exact algorithm PACER
practically runs much faster than the state-of-art approximation
algorithm with a theoretical approximation guarantee. Therefore in real applications, when budget b and |VQ | are not large,
we prefer PACER that returns best solutions in reasonable
responding time. When b and |VQ | are relatively large, we
switch to the collapse heuristic version of PACER, which
sacrifices the optimality a little bit but runs much faster. And
when b and |VQ | are very large, we switch to the fastest greedy
algorithm, which can usually return an acceptable solution.
D. Comparison with A*
A* [25] only Q
works for their keyword coverage function:
Φh (PV ) = 1 − i∈PV [1 − F̃i,h ], and finds single route. In
[25], F̃i,h is in the range [0, 1] and it is set to 1 if the number
of check-ins on POI i for feature h is above average. In this
case, the single POI in P yields the maximum Φh (PV ) value;
the feature h of other POIs will not be considered. To compare
our algorithm with A*, we set β = 0.5 in Eqn. (17). To speed
up A* and for a fair comparison, we also equip A* with our
indices for POI Candidate retrieval. Note that the maximum
budget b in [25] is 15 kilometers in their efficiency study,
which is about 20 minutes by Google Maps under driving
mode, much smaller than our settings of 4 to 9 hours.
Figure 9 shows the comparison between PACER+2 and the
modified A* on both datasets. The report of Gain is omitted
as they are both exact algorithms. PACER+2 outperforms A*,
especially for a large b. Several queries of A* on Austin
even failed for b = 9 hours. Although A* has a pruning
strategy specifically for their keyword coverage function, the
search strategy itself is a bottleneck. Besides, their gainbased pruning is based on the greedy algorithm presented in
[30] with a bound looser than our online bound. In fact, the
experiments in [25] showed that A* is just 2-3 times faster than
the brute-force algorithm. In contrast, our algorithm enables
a dynamic programming based search method and multiple
pruning strategies, making the search space much smaller than
that of the brute-force algorithm, as shown in Section V-E.
VIII. C ONCLUSION AND E XTENSIONS
We considered a personalized top-k route search problem.
The large scale of POI maps and the combination of search in
feature space, spatial space, and path space make this problem
computationally hard. The personalized route diversity requirement further demands a solution that works for any reasonable
route diversity specification. We presented an exact solution to
address these challenges through multiple pruning strategies
that are applicable to the whole class of diversity requirements
corresponding to the class of submodular functions. The analytical and empirical evaluation suggested that our solutions
are significantly faster than the state-of-the-art algorithms. We
also presented high-performance heuristic solutions.
We introduce several possible extensions of this work.
Feature order constraints. Some user may have certain
order preference of the features in a route, e.g., at least one
POI with feature “Food” should be visited before the POIs
with “Shopping”. All this kind of feature partial orders can
be represented by a topological sorting. Once an open route
violates any order constraint, the violation cannot be removed
by appending more POIs to its end, thus, it is safe to prune
all the routes that are extension of the open route. That is, the
order constraints are anti-monotone. Our algorithms presented
in Sections V and VI can be easily adopted to prune the search
space for any additional anti-monotone constraint.
Feature combination requirements. The following (not
limited to) features combination requirements may be interesting in real application. i. Two POIs having the same feature
A cannot be visited consecutively; ii. Either both feature A
and B are visited, or none is visited; iii. Exactly only one of
A and B is visited, cannot both. With these constraints, less
routes are satisfied, thus, extra pruning is introduced.
R EFERENCES
[1] I. A. Junglas and R. T. Watson, “Location-based services,” Communications of the ACM, vol. 51, no. 3, pp. 65–69, 2008.
[2] Y. Zheng, L. Zhang, X. Xie, and W.-Y. Ma, “Mining interesting locations
and travel sequences from gps trajectories,” in WWW, 2009, pp. 791–
800.
[3] H. Yoon, Y. Zheng, X. Xie, and W. Woo, “Social itinerary recommendation from user-generated digital trails,” Personal and Ubiquitous
Computing, vol. 16, no. 5, pp. 469–484, 2012.
[4] T. Kurashima, T. Iwata, G. Irie, and K. Fujimura, “Travel route recommendation using geotags in photo sharing sites,” in CIKM. ACM,
2010, pp. 579–588.
[5] J. Sang, T. Mei, J.-T. Sun, C. Xu, and S. Li, “Probabilistic sequential
pois recommendation via check-in data,” in SIGSPATIAL. ACM, 2012,
pp. 402–405.
[6] Q. Diao, M. Qiu, C.-Y. Wu, A. J. Smola, J. Jiang, and C. Wang, “Jointly
modeling aspects, ratings and sentiments for movie recommendation
(jmars),” in the 20th ACM SIGKDD. ACM, 2014, pp. 193–202.
[7] C. Chekuri, N. Korula, and M. Pál, “Improved algorithms for orienteering and related problems,” ACM Transactions on Algorithms (TALG),
vol. 8, no. 3, p. 23, 2012.
[8] A. Krause and C. Guestrin, “Beyond convexity: Submodularity in
machine learning,” ICML Tutorials, 2008.
[9] D. Kempe, J. Kleinberg, and É. Tardos, “Maximizing the spread of
influence through a social network,” in Proceedings of the ninth ACM
SIGKDD. ACM, 2003, pp. 137–146.
[10] G. Jossé, Y. Lu, T. Emrich, M. Renz, C. Shahabi, U. Demiryurek, and
M. Schubert, “Scenic routes now: Efficiently solving the time-dependent
arc orienteering problem,” arXiv preprint arXiv:1609.08484, 2016.
[11] W. Zhang and J. Wang, “Location and time aware social collaborative
retrieval for new successive point-of-interest recommendation,” in Proceedings of the 24th ACM CIKM. ACM, 2015, pp. 1221–1230.
[12] Y. Ge, Q. Liu, H. Xiong, A. Tuzhilin, and J. Chen, “Cost-aware travel
tour recommendation,” in Proceedings of the 17th ACM SIGKDD.
ACM, 2011, pp. 983–991.
[13] Q. Liu, Y. Ge, Z. Li, E. Chen, and H. Xiong, “Personalized travel
package recommendation,” in ICDM. IEEE, 2011, pp. 407–416.
[14] K. Zheng, S. Shang, N. J. Yuan, and Y. Yang, “Towards efficient search
for activity trajectories,” in 29th ICDE. IEEE, 2013, pp. 230–241.
[15] B. Zheng, N. J. Yuan, K. Zheng, X. Xie, S. Sadiq, and X. Zhou,
“Approximate keyword search in semantic trajectory database,” in 31st
ICDE. IEEE, 2015, pp. 975–986.
[16] L.-Y. Wei, Y. Zheng, and W.-C. Peng, “Constructing popular routes
from uncertain trajectories,” in Proceedings of the 18th ACM SIGKDD.
ACM, 2012, pp. 195–203.
[17] J. Dai, B. Yang, C. Guo, and Z. Ding, “Personalized route recommendation using big trajectory data,” in ICDE. IEEE, 2015, pp. 543–554.
[18] S. Basu Roy, G. Das, S. Amer-Yahia, and C. Yu, “Interactive itinerary
planning,” in ICDE. IEEE, 2011, pp. 15–26.
[19] A. Gionis, T. Lappas, K. Pelechrinis, and E. Terzi, “Customized tour
recommendations in urban areas,” in ACM WSDM, 2014, pp. 313–322.
[20] C. Zhang, H. Liang, and K. Wang, “Trip recommendation meets realworld constraints: Poi availability, diversity, and traveling time uncertainty,” ACM TOIS, vol. 35, no. 1, p. 5, 2016.
[21] M. De Choudhury, M. Feldman, S. Amer-Yahia, N. Golbandi, R. Lempel, and C. Yu, “Automatic construction of travel itineraries using social
breadcrumbs,” in Proceedings of the 21st ACM conference on Hypertext
and hypermedia. ACM, 2010, pp. 35–44.
[22] E. H.-C. Lu, C.-Y. Chen, and V. S. Tseng, “Personalized trip recommendation with multiple constraints by mining user check-in behaviors,” in
SIGSPATIAL. ACM, 2012, pp. 209–218.
[23] W. Li, J. Cao, J. Guan, M. L. Yiu, and S. Zhou, “Retrieving routes of
interest over road networks,” in International Conference on Web-Age
Information Management. Springer, 2016, pp. 109–123.
[24] X. Cao, L. Chen, G. Cong, and X. Xiao, “Keyword-aware optimal route
search,” VLDB Endowment, vol. 5, no. 11, pp. 1136–1147, 2012.
[25] Y. Zeng, X. Chen, X. Cao, S. Qin, M. Cavazza, and Y. Xiang, “Optimal
route search with the coverage of users’ preferences,” in 24th IJCAI.
AAAI Press, 2015, pp. 2118–2124.
[26] A. Krause and D. Golovin, “Submodular function maximization,”
Tractability: Practical Approaches to Hard Problems, vol. 3, no. 19,
p. 8, 2012.
[27] M. Jiang, A. W.-C. Fu, R. C.-W. Wong, and Y. Xu, “Hop doubling label
indexing for point-to-point distance querying on scale-free networks,”
VLDB Endowment, vol. 7, no. 12, pp. 1203–1214, 2014.
[28] D. Michie, “Memo functions and machine learning,” Nature, vol. 218,
no. 5136, pp. 19–22, 1968.
[29] M. Sviridenko, “A note on maximizing a submodular set function subject
to a knapsack constraint,” Operations Research Letters, vol. 32, no. 1,
pp. 41–43, 2004.
[30] S. Khuller, A. Moss, and J. S. Naor, “The budgeted maximum coverage
problem,” Information Processing Letters, vol. 70, no. 1, pp. 39–45,
1999.
[31] J. Leskovec, A. Krause, C. Guestrin, C. Faloutsos, J. VanBriesen,
and N. Glance, “Cost-effective outbreak detection in networks,” in
Proceedings of the 13th ACM SIGKDD. ACM, 2007, pp. 420–429.
[32] D. M. Burton, Elementary number theory. Tata McGraw-Hill Education, 2006.
[33] C. Chekuri and M. Pal, “A recursive greedy algorithm for walks in
directed graphs,” in Foundations of Computer Science, 2005. FOCS
2005. 46th Annual IEEE Symposium on. IEEE, 2005, pp. 245–253.
[34] A. Singh, A. Krause, C. Guestrin, W. J. Kaiser, and M. A. Batalin,
“Efficient planning of informative paths for multiple robots.” in IJCAI,
vol. 7, 2007, pp. 2204–2211.
| 8cs.DS
|
SWISH: SWI-Prolog for Sharing
Jan Wielemaker1 and Torbjörn Lager2 and Fabrizio Riguzzi3
1
arXiv:1511.00915v1 [cs.PL] 3 Nov 2015
2
Web and Media group, VU University Amsterdam, The Netherlands,
[email protected]
Department of Philosophy, Linguistics and Theory of Science, University of Gothenburg,
Sweden
[email protected]
3
Dipartimento di Matematica e Informatica, University of Ferrara, Italy
[email protected]
Abstract. Recently, we see a new type of interfaces for programmers based on
web technology. For example, JSFiddle, IPython Notebook and R-studio. Web
technology enables cloud-based solutions, embedding in tutorial web pages, attractive rendering of results, web-scale cooperative development, etc. This article
describes SWISH, a web front-end for Prolog. A public website exposes SWIProlog using SWISH, which is used to run small Prolog programs for demonstration, experimentation and education. We connected SWISH to the ClioPatria
semantic web toolkit, where it allows for collaborative development of programs
and queries related to a dataset as well as performing maintenance tasks on the
running server and we embedded SWISH in the Learn Prolog Now! online Prolog
book.
1
Introduction
Web technology has emerged to a state where it becomes useable for implementing programming development environments. All major modern browsers now implement HTML5 and JavaScript and there are mature components available such as the
CodeMirror4 and ACE5 code editors, the Bootstrap6 framework for styling and UI widgets and vizualization libraries such as D3.js.7 Using web technology rather than traditional GUI based technology such as Eclipse,8 Microsoft Visual Studio,9 XEmacs,10
etc. has various advantages. Being network transparent, it allows for controlling cloud
hosted applications as well as Prolog processes running on headless devices. Web technology provides a great infrastructure for mashups, pages that integrate material from
several sources. For example, embedding Prolog in tutorial pages or embedding Prolog
queries that can be modified and re-evaluated in documents that describe data collections.
4
5
6
7
8
9
10
https://codemirror.net/
http://ace.c9.io
http://getbootstrap.com/
http://d3js.org/
https://eclipse.org/
https://www.visualstudio.com/
http://www.xemacs.org/
With SWISH (SWI-Prolog for Sharing), we provide this technology for (SWI)Prolog. SWISH consists of JavaScript client (browser) code and a number of Prolog
libraries that realise the server as a Prolog application. The client code consists of a
series of jQuery11 plugins that deal with editing source code, managing a shared source
repository, entering queries and rendering answers produced by Prolog. The server-side
libraries serve the overall web application, implement the source store and support the
editor with predicate documentation, templates, cross-reference results, etc. For executing Prolog queries, SWISH relies on Pengines (Prolog engines, [3]). A pengine is
a Prolog engine that can be controlled similarly to Prolog running in a terminal using
HTTP requests. The SWISH infrastructure was originally developed as a Prolog version of JSFiddle. It was later reimplemented as a modular jQuery based infrastructure
aiming at collaborative exploration of data hosted on a SQL or SPARQL server. This
use-case is described in section 5.2.
This article is organised as follows. Section 2 describes related work, which in our
case are the systems that have inspired us. Section 3 describes the architecture and
components of SWISH. In section 5 we describe four applications of the current system.
We conclude with future work and conclusions.
2
Related work
We are not aware of other initiatives that aim at developing a rich web-based development environment for Prolog. We do not compare SWISH with traditional editor or GUI
based development environments for Prolog because web-based environments provide
new opportunities and pose new challenges. Instead, we discuss three applications that
have served as inspiration for SWISH: JSFiddle,12 R-Studio13 and IPython Notebook.14
– As stated, the initial inspiration for SWISH was JSFiddle. Unlike JSFiddle though,
Prolog is executed on the server rather than in the browser.
– R-Studio [2] is an interface to the R statistical package. Although not a web application, it is based on the Qt webkit framework and uses web based technology in the
background. R-Studio came into the picture when the COMMIT/ project provided
a grant for developing SWISH as a toolkit for analysis of relational (SQL) data.
The R-studio interface has a similar layout as SWISH, providing a source window,
a console and an output plane that typically shows results in tables or charts.
– IPython Notebook [5] allows mixing markdown or HTML text with Python
sources. The rendered Notebook shows the text, sources and possible results in
the form of numbers, tables or charts.
SWISH embodies most of the ideas behind JSFiddle and R-Studio. Embedding of
SWISH in documents is demonstrated in section 5.3. Interactive editing of documents
that embed SWISH is discussed in future work (section 6).
11
12
13
14
https://jquery.com/
https://jsfiddle.net/
http://www.rstudio.com/
http://ipython.org/notebook.html
100
Both R-Studio and IPython Notebook work on the basis of authentication (either to
the OS or application), after which any command may be executed. SWISH can operate
both as a public service granting access to non-intrusive queries and as an authenticated
service to run arbitrary queries, for example for maintenance purposes. See section 5.2.
3
The SWISH application
SWISH consists of two parts. The client side, running in a browser, is implemented
as a series of jQuery plugins, using Bootstrap for styling and RequireJS15 for package
management. The server side is completely implemented in SWI-Prolog [7]. It builds on
top of the SWI-Prolog HTTP server libraries, the Pengines library and the IDE support
libraries that provide data for auto completion, documentation and highlighting.
In the following sections we describe SWISH in terms of interface components,
where we discuss the requirements, the user aspects, the client code and supporting
server functionality for each component. First, we provide a screendump that illustrates
the main components in figure 1.
Answer pane with
runner
Render as
chessboard
Hover shows
details
Query pane
Fig. 1. Screendump of SWISH. The left pane shows the source code, while the top-right pane
shows a query runner that exploits the current selected answer renderer and buttons on how to
continue after the first answer. The bottom-right pane provides the query editor with access to
example queries stored in the source, query history, apply solution modifiers, result presentation
and a Run! button to start the query.
15
http://requirejs.org/
101
3.1
The code editor
A proper editor is the most important component of a usable programming environment.
The editor must support the language, including syntax highlighting, auto indentation,
code completion based on templates and already existing code and highlighting of errors
and warning from the compiler. The editor is used both for editing the source code and
editing queries.
Prolog is a difficult language to support in code editors due to the lack of reserved
keywords, e.g., the word if in C starts an if-statement if not embedded in comment or
a string, but the word is in Prolog can refer to the built-in predicate is/2, but also some
predicate with a different arity, just a constant, etc. Another example is X-Y which can
both be an arithmetic expression or a pair as used with e.g., keysort/2. Next to the
lack of keywords the ability to extend the syntax using new operators complicates the
implementation of syntax support while editing. SWI-Prolog’s built-in Emacs oriented
editor resolves this problem by closely integrating Prolog with the editor. While typing,
the current term (clause or directive) is parsed and analysed in the context of the current
file and the file’s imports after each keystroke. If the term has valid syntax, all tokens
are coloured according to their syntactic role as well their relation to the remainder of
the program. For example, a call to a non-existing predicate is coloured red, a call to a
built-in or imported predicate is blue and a call to a locally defined predicate is black.
The libraries that implement this analysis have been decoupled from the built-in editor,
both to support source colouring for the SWI-Prolog documentation system PlDoc [6]
and ProDT16
There are two dominant open source and actively maintained in-browser code editors available: ACE and CodeMirror. When we started SWISH, ACE had a very basic
Prolog mode and CodeMirror had none. We nevertheless opted for CodeMirror because its highlighting is based on raw JavaScript code rather than a regular expression
based template language as used for ACE. The low level implementation allows for
a novel highlighting implementation. The highlighter consists of a JavaScript implemented Prolog tokeniser. Tokenizing Prolog is sufficient to colour comments, quoted
material (strings, quoted atoms), variables and constants (atoms and numbers). It is
also sufficient to support smart indentation. As discussed above, it is not sufficient for
highlighting the role played by atoms and compound terms.17
We provide semantic highlighting as illustrated in figure 2 by (1) forwarding the
changes to the content of the editor to the server server which maintains a mirror of the
editor content and (2) asking the server to produce a list of semantically enriched tokens
for the source. The tokens are returned as a list-of-lists, where each inner list represents
the tokens for a source term (clause or directive). Grouping the tokens per source term
allows for incremental update (not yet implemented) as well as re-synchronisation (see
below). For example, a fragment of an enriched token list may look like this: [ functor,
(undefined goal), variable (singleton), . . . ]. The JavaScript tokeniser matches its tokens
16
17
http://prodevtools.sourceforge.net, these libraries are not yet used by ProDT.
An additional complication is formed by CodeMirror’s token-based highlighting which does
not support look-ahead. As a consequence, we must decide on the colour of e.g., asserta(
while we do not know the arity of the term.
102
built-in
not called
auto-loaded
locally defined
recursive
Fig. 2. The semantic highlighter classifies, in addition to the syntactic category such as comment
or variable, terms that define or call predicates based on cross-referencing the source code.
with this list. If the basic type (e.g., ‘functor’ or ‘variable’) matches, it uses the enrichment information (e.g., ‘singleton’) to decide on the style. If the basic token type does
not match, it highlights the token using the basic syntactical category and schedules a
request to the server for a new list of enriched tokens. This request is sent if the user
pauses typing for 2 seconds. The request is accompanied by the full source if this is
small or the list of changes since the last request if the source is large. While waiting
for up-to-date enriched tokens, the JavaScript highlighting code heuristically tries to
re-synchronise and either uses the uncertain results or falls back to the plain tokens.
Re-synchronisation checks whether a single added, deleted or modified token gets the
token stream in-sync. If this fails it tries to re-synchronise on a full-stop with the next
clause or directive.
A CodeMirror hover plugin is used to show basic information about tokens if the
pointer hovers over it. For goals, this includes where the goal is defined (ISO, SWIProlog built-in, a library, locally) and the documentation summary information if available. This information is requested from the server.
A CodeMirror template plugin is configured from templates (e.g.,
atom length(+Atom, -Length)) extracted from the SWI-Prolog manual
and PlDoc documentation of imported libraries. This plugin shows a menu of
applicable predicates with their templates on Control-Space.
Finally, if the user uses the Run! button to execute a query, the program is compiled.
If the compiler generates errors or warnings, these are inserted as notes in the source
code.
3.2
Source code and query management
As JSFiddle formed the initial inspiration for SWISH, SWISH has a facility to save the
program. The current version of SWISH explicitly targets the cooperative development
103
of Prolog programs and queries related to a dataset (see section 5.2). This triggered the
implementation of a more organised storage facility. The server-side storage module is
implemented in Prolog and inspired by GIT. Each file is versioned independently rather
than maintaining the version of a hierarchy of files. Files can be referenced by content
using their GIT compatible SHA1 hash or by name. The name can be considered a
version head and refers to the latest version with that name. The file save and load
interface provides the following operations:
– Saving a file anonymously, which produces a randomly generated URL similar to
JSFiddle.
– Saving a file by name.
– Saving a new version. The interface shows the available versions and the modifications.
– Forking a file under a new name. The history remains linked to the original.
Prolog source files can include other sources on the same server
using :- include(filename)., including the latest version or
:- include(hash). to include a specific version.
Prolog source files can embed example queries using structured comments, where
each sequence from ?- to the matching full stop token is added to the Examples menu
of the query panel (see figure 6). The comment below illustrates a call to append/3
embedded in the source window.
/** <examples>
?- append([one], [two,three], List).
*/
3.3
The query editor
The query editor is based on the same jQuery plugin that realises the code editor and
thus profits from the syntax highlighting, template insertion and hovering plugins. In
addition, it provides three popup menus:
Examples This menu is filled from the structured comments described above. The examples menu is shown in figure 6.
History This menu provides earlier executed queries.
Solutions This menu embeds an existing query in a meta-call to alter the result. Currently provided operations are Aggregate (count all), Order by, Distinct, Limit, Time
and Debug (trace). Figure 3 shows how the menu is used to count the solutions of
a goal.
104
Fig. 3. The Solutions menu can be used to count results, order them, filter duplicates, etc. The
upper runner shows answers to the query as a table.
3.4
Running a query: runners in the answer pane
The answer pane is a placeholder for runners, where each runner represents a query. The
answer pane provides a menu for operations on all runners inside it. Provided actions are
Collapse all, Expand all, Stop all and Clear. The query may be completed, running
or waiting for user input. SWISH can manage multiple active queries at the same time,
up to an application defined maximum (default 3).
Each runner provides its own set of commands to control the specific query. During
execution a runner provides an Abort button. After query evaluation completes with an
answer and more answers may be available the runner allows for asking the next 1, 10,
100 or 1,000 results or to Stop the query. In addition, the runner shows a text input field
when the application wants to read input and may show debugger interaction buttons if
the tracer is being used (see section 3.4).
A runner can render answers in two modes, the classical Prolog mode or as a table, similar to what many database interfaces provide. The ‘table’ mode is particularly
appealing for querying datasets (see figure 3), while the former is more suitable for rendering small amounts of complex answers such as the chessboard position in figure 1.
By default, Prolog terms are rendered as structured HTML objects, where the rendered
text is the same as Prolog’s writeq/1 predicate.
The server can provide rendering libraries that render Prolog terms using
dedicated HTML. In figure 1, the ‘chess’ renderer is loaded due to the :use rendering(chess) directive. The ‘chess’ renderer translates a list of length
N holding integers in the range 1..N as a chessboard with queens. In addition to the
chess rendering library, SWISH provides rendering libraries for sudoku puzzles, parse
trees and tables. The ClioPatria version adds a renderer for RDF resources that renders
the resource more compactly and provides a hyperlink for obtaining details. If a term
105
can be rendered in multiple ways, the interface provides a hover menu to select between
the alternatives. Figure 4 illustrates this functionality. The render facility is similar to the
Prolog portray/1 hook that allows changing the result printed for terms with a specific
shape. However, it can exploit the full potential of HTML (or SVG) and the interface
allow for switching the selected rendering.
A rendering library is a module that must define a non-terminal (grammar rule)
term rendering//3, calling html//118 to produce HTML from the Prolog input
term, a list of variable bindings (Name = Variable) and user provided options. In the
current version, new rendering modules must be loaded into the SWISH server and
cannot be created by the SWISH user.
Fig. 4. With the ‘chess’ render library, a list of integers is interpreted as queens on a chessboard.
The user can select rendering as a ‘Prolog term’ to see the actual term.
Server side execution of the query Server-side execution of a query is supported
by the Pengines [3] library. The Pengines library allows for creating a Prolog engine
represented by a Prolog thread. Optionally, the pengine is handed a Prolog program that
is loaded into the pengine’s workspace (program space). The workspace is a temporary
module that is disposed of after the pengine terminates. The pengine may be asked
questions through HTTP queries, similar to a traditional Prolog user interacting with
Prolog running in a terminal.19
If the SWISH user hits the Run! button, the content of the source pane is used to
create a new pengine. Subsequently, the content of the query pane is sent as the one and
only query that will be executed by the pengine.20 Before execution, the query is verified
to be safe, unless sandboxing is disabled (see section 5.2). The sandbox component is
discussed below.
18
19
20
http://www.swi-prolog.org/pldoc/doc for?object=html/3
https://www.youtube.com/watch?v=G_eYTctGZw8
Pengines can execute multiple queries. We do not use this facility because a fresh pengine
starts in a predictable state (standard operators, empty dynamic database).
106
The pengine’s default working module may be pre-loaded with code. SWISH uses
this facility to redefine the Prolog I/O predicates such as read/1, write/1, format/1,2,3,
etc. The ClioPatria version (section 5.2) also preloads the RDF libraries, so users can
run queries on the RDF database without explicitly importing the required libraries.
Sandboxing queries A Prolog environment contains global state in the form of loaded
modules, defined operators, dynamic predicates, etc. Prolog exposes a rich and potentially dangerous interface to the operating system. For an anonymous services, we want
each query to start in a well defined state and we must ensure that execution of the query
does not make unwanted changes to the hosting computer or leaks sensitive information.
Both for education purposes and data analysis one can write meaningful programs
without making permanent changes to the server or the server’s filesystem. That is
where the sandbox library comes in. The sandbox library is active while loading the
source, where it refuses to add clauses to other modules than the pengine’s workspace
and where it only accepts a restricted set of directives, also aimed at keeping all changes
local to the workspace. Prior to execution, the sandbox unfolds the query and compares
all reachable goals with a whitelist. The whitelist contains all side-effect free built-in
Prolog predicates, safe meta-predicates (e.g., findall/3) and allows for using the dynamic database, provided that the head of the affected predicate is not module-qualified
(and thus the referenced predicate lives in the temporary program space of the Pengine)
and the body is safe. It does not allow for cross-module calls (Module:Goal) to private
predicates and does not provide access to object-enumeration predicates such as current atom/1, current predicate/1, etc., both to avoid leaking sensitive information.
The sandbox test fails under one of these conditions:
– It discovers a (meta-) goal for which it cannot deduce the called code. The traditional example is read(X), call(X). If such a goal is encountered, it signals
an instantiation error, together with a trace that explains how the insufficiently
instantiated goal can be reached. Note that it can deal with normal high-order predicates if the meta-argument is specified. For example, the following goal is accepted
as safe.
?- maplist(plus(1), [1,2,3])
– It discovers a goal that is not whitelisted. In this case it signals a permission error,
again accompanied with a trace that explains how the goal can be reached. Note that
pure Prolog predicates are unfolded, also if it concerns predicates from the libraries
or belonging to the set of built in predicates.
– It discovers a cross-module (M:Goal) call to a predicate that is not public. Normally, SWI-Prolog, in the tradition of Quintus Prolog, allows for this. Allowing it
in SWISH would imply that no data can be kept secret. With this limitation, libraries
can keep data in local dynamic predicates that remain invisible to non-authorised
users.
Debugging The SWISH debugger is based on the traditional 4-port debugging model
for Prolog. Figure 5 shows the tracer in action on sublist/2 from the Lists example
107
source. The debugger was triggered by a break-point on line 10 set by clicking on the
line-number in the code editor. The debugging interaction is deliberately kept simple
and similar to traditional programming environments. A retry button is added to the
commonly seen ‘step into’, ‘step over’ and ‘step out’ for highlighting the unique feature
of Prolog to re-evaluate a goal.
Fig. 5. Debugging applications in SWISH
4
Portability
The SWISH client libraries are based on mature and well maintained JavaScript libraries. The client runs all modern major browsers with HTML5, CSS and JavaScript
support. It is frequently tested on FireFox, Chrome, Safari and Internet Explorer 11.
The server code is basically non-portable. Many of the required libraries and features are shared with at least one other Prolog implementation, but none is capable to
support the full range. Below we summarise the main problems.
– The scale of the involved Prolog libraries demands for a closely compatible Prolog
module system. Probably only SICStus and YAP can be used without significant
rewrites.
108
– The HTTP server libraries are heavily based on C code that interacts with the SWIProlog foreign language interface to Prolog streams. YAP has copied the low-level
libraries and is capable to run (an old version of) these libraries.
– The Pengines library depends on the HTTP library and the multi-thread interface.
The SWI-Prolog thread API is also provided by YAP and XSB.
– The sandbox library (section 3.4) assumes that whitelisted predicates are indeed
safe. This requires robust handling of invalid calls and resource overflows. Few
Prolog systems can satisfy this requirement. SICStus Prolog would be a candidate,
but SICStus does not support multi-threading.
– The semantic syntax highlighting depends on detailed source layout information
provided by read term/3. SWI-Prolog’s support for term layout is an extended
version of the Quintus Prolog term layout functionality.
– Significant parts of the code rely on SWI-Prolog version 7 extensions, notably the
dict and string types that facilitate a natural mapping between Prolog and JSON
data.
From the above list it should be clear that a fully functional port of SWISH to
another Prolog system is not immediately feasible. YAP probably comes closest but
still requires a significant amount of work.
There is a more realistic scenario though. In this setup, SWI-Prolog provides the
web interface and most of the development tools and another language, not even necessarily Prolog, provides the query solving. The interface between the two can be based
on interprocess communication or, if the target system is robust, safe and capable of
supporting threads, by linking the target system into the process and using the C interface for communication.
5
Applications
In this section we describe and evaluate four publicly available SWISH applications.
All these services are regularly updated to run the latest version of SWISH and SWIProlog.
5.1
SWISH
SWISH21 runs a plain publicly accessible copy of SWISH that allows running sandboxed (see section 3.4) Prolog programs. The server has collected 10,800 programs
between September 29, 2014 and June 2, 2015. Over the month May 2015, it has executed 258,809 Prolog queries. The web site is regularly used by users of the ##prolog
IRC channel to discuss programming solutions and is in active use for education.22
21
22
http://swish.swi-prolog.org
Steve Matuszek, UMBC (via e-mail: “Thank you very much for this fantastic resource! I used
it while teaching Prolog this semester, and it really helped tighten the loop for my students.
We spent zero time on tool installation and other overhead, and all the time on understanding
the concepts. I even had them turn their assignments in via SWISH, with their test queries in
the examples block.”
109
5.2
ClioPatria
ClioPatria is a semantic web (RDF) framework for SWI-Prolog. It consists of an RDF
triple store, a SPARQL server and a web frontend to manage the server and explore
the data in the RDF store. ClioPatria can be extended using cpacks (ClioPatria pack or
plugin). SWISH is available as a ClioPatria cpack23 and makes the Prolog shell available for querying as well as maintenance tasks. Without login, user can run side-effect
free queries over the RDF data stored in ClioPatria’s RDF database. After login with
administrative rights, the sandbox limitations are lifted and the Prolog shell can be used
to perform maintenance tasks on the RDF data such as data transformation, cleanup,
etc., as well as program maintenance tasks such as reloading modified source files.
SWISH has been used in the Talk Of Europe creative camp24 to explore data on the
speeches in the European parliament.25 Although still immature, users appreciated the
ability to define more efficient and expressive queries than provided by the SPARQL
query interface. Above all, the ability to save and share programs that perform interesting tasks on the data was frequently used, in particular to seek help fixing queries.
5.3
Learn Prolog Now!
Learn Prolog Now!26 is an online version of a Prolog book by Patrick Blackburn, Johan
Bos, and Kristina Striegnitz [1]. We established a proof of concept that embeds SWISH
in the online course material.27 It is realised as a Prolog hosted proxy that fetches the
pages from the main site and serves the enhanced pages to the user. The proxy identifies
and classifies the code fragments in terms of ‘source code’ ‘queries’ and dependencies.
Next, it adds a button to the source fragments that, when pressed, replaces the HTML
<pre> element with in <iframe> running SWISH filled with the source while the
example queries are added to the Examples menu (figure 6). The queries are executed
by http://swish.swi-prolog.org. Program and queries are transferred using
the following HTTP parameters: code (the source code), background (source code
that is loaded into the pengine but not visible in the editor), examples (queries that appear in the Examples menu) and q (the initial query). The proxy server served 19,700
pages during May 2015.
5.4
cplint on SWISH
cplint on SWISH28 is a web application based on SWISH for reasoning with probabilistic logic programs under the distribution semantics. The Prolog source window
is used to write a logic program with annotated disjunction. A query in the form of a
ground atom is answered by returning its probability of being true in the program. The
computation of the probability is done with the cplint system [4] in the server using
23
24
25
26
27
28
http://cliopatria.swi-prolog.org/packs/swish
http://www.talkofeurope.eu/
http://purl.org/linkedpolitics
http://www.learnprolognow.org
http://lpn.swi-prolog.org
http://cplint.lamping.unife.it/
110
Fig. 6. Screendump of Learn Prolog Now with opened SWISH instance that shows the collected
source as well as example queries from the subsequent text that are classified as relating to this
source. The embedded SWISH provides all functionality available in the stand-alone SWISH. If
the user presses the close button, SWISH will be removed and the original code re-appears.
Pengines. The input program is translated into an internal representation using source
to source transformation.
SWISH was modified only in the JavaScript code for the runner. The source code
is prepended code for loading the cplint library and enabling the source to source
transformation while the query is wrapped into a call of the inference predicate. This
call has a variable argument Prob which will hold the probability and will be shown to
the user in the answer pane.
5.5
TRILL on SWISH
TRILL on SWISH29 is a probabilistic OWL reasoner that reuses SWISH. As SWISH
for ClioPatria, described in section 5.2, it is a ClioPatria cpack. The Prolog source window is replaced by an RDF/XML editor window that can be used to upload an OWL
ontology while the query editor can be used to pose Prolog queries against the OWL
29
http://trill.lamping.unife.it/
111
ontology. The probability of queries is computed using TRILL [8], a reasoner in Prolog
that adopts the distribution semantics for probabilistic description logics.
Also for TRILL on SWISH we had only to modify the JavaScript code for the
runner. The source code sent to the Pengine is obtained by adding Prolog code for
parsing an RDF/XML string, by calling the parsing predicate and by wrapping the query
in a meta-call that performs syntactic checks for misspellings.
6
Future work
Although definitely usable in its current state, SWISH is work in progress. We are confident that the basic component selection and organisation of the server and client code
are stable. More work is needed to improve the current system. Notably the semantic
highlighting is not yet perfect and often fails to degrade gradually if the server side annotation does not match the client tokens perfectly. The Pengine’s sandbox protection is
often too restrictive, while several security flaws have been reported and fixed already.
It is, and probably always will be, advised to run public SWISH-enabled services in
an operating system sandbox. The current server suffers from memory leaks and stability problems. Although the situation has improved significantly, the main demo server
needs to be restarted about once a week.30
We foresee several extensions to SWISH that will improve current applications and
enable new opportunities for deploying SWISH.
– ClioPatria’s authorised usage of SWISH shows some of the potential for controlling servers or embedded Prolog engines. In addition to small temporary Prolog
programs, we would like to be able to edit existing and create new Prolog modules as well as pages in other languages, such as JavaScript, HTML and CSS. Full
editing capabilities would allow for shared development of server software without
shell access to the server on which SWISH enabled software is running.
– Multi-document editing can enhance the sandboxed SWISH application by providing input and output documents. Compare this to TRILL (section 5.5) using an
RDF/XML document as input.
– We plan to provide a markdown-based format specifically for writing tutorials and
well as dataset analysis documents. The first will look like the Learn Prolog Now!
example discussed in section 5.3. For the second, we envision a document with
embedded code and query fragments, where the query fragments produce tables or
charts. This is similar to IPython Notebook.
– Turn http://swish.swi-prolog.org into a reliable and scalable resource.
Examine the possibility for schools to instantiate a private version that is preloaded
with course material and assignments.
30
A restart of the server has only small consequences to active users. Open queries are killed.
The source code mirror is lost, but automatically recovered if the client asks for a new set of
semantically enriched tokens.
112
7
Conclusion
This article presented SWISH, SWI-Prolog for Sharing. SWISH provides a webenabled interface to Prolog that is based on ideas from JSFiddle, R-Studio and IPython
Notebook. It consists of a JavaScript client side, while the server side is based on SWIProlog’s HTTP and Pengines (Prolog engines) libraries. SWISH can be deployed in
many settings, such as education, data analysis and server development and maintenance. SWISH as a whole is tied to SWI-Prolog, but other languages, not even limited
to Prolog, could be controlled from SWI-Prolog. SWISH is made available as open
source and can be downloaded from github.31
Acknowledgements
The development of SWISH was supported by the Dutch national program COMMIT/.
References
1. Patrick Blackburn, Johan Bos, and Kristina Striegnitz. Learn prolog now!, volume 7. College
Publications, 2006.
2. Christopher Gandrud. Reproducible Research with R and R Studio. CRC Press, 2013.
3. Torbjörn Lager and Jan Wielemaker. Pengines: Web logic programming made easy. TPLP,
14(4-5):539–552, 2014.
4. Fabrizio Riguzzi and Terrance Swift. Well-definedness and efficient inference for probabilistic
logic programming under the distribution semantics. Theory Pract. Log. Program., 13(Special
Issue 02 - 25th Annual GULP Conference):279–302, March 2013.
5. Cyrille Rossant. Learning IPython for interactive computing and data visualization. Packt
Publishing Ltd, 2013.
6. Jan Wielemaker and Anjo Anjewierden. PlDoc: Wiki style literate programming for Prolog.
In Patricia Hill and Wim Vanhoof, editors, Proceedings of the 17th Workshop on Logic-Based
methods in Programming Environments, pages 16–30, 2007.
7. Jan Wielemaker, Zhisheng Huang, and Lourens van der Meij. Swi-prolog and the web. TPLP,
8(3):363–392, 2008.
8. Riccardo Zese, Elena Bellodi, Evelina Lamma, Fabrizio Riguzzi, and Fabiano Aguiari. Semantics and inference for probabilistic description logics. In Uncertainty Reasoning for the
Semantic Web III, volume 8816 of LNCS, pages 79–99. Springer, 2014.
31
https://github.com/SWI-Prolog/swish
113
| 6cs.PL
|
Efficient Multi-Person Pose Estimation with Provable Guarantees
arXiv:1711.07794v1 [cs.CV] 21 Nov 2017
Shaofei Wang
Beijing A&E Technologies
Beijing, China
[email protected]
Konrad Paul Kording
University of Pennsylvania
Philadelphia, PA
[email protected]
Julian Yarkony
Experian Data Labs
San Diego, CA
[email protected]
Abstract
per/lower bounds.
Multi-person pose estimation (MPPE) in natural images
is key to the meaningful use of visual data in many fields
including movement science, security, and rehabilitation.
In this paper we tackle MPPE with a bottom-up approach,
starting with candidate detections of body parts from a convolutional neural network (CNN) and grouping them into
people. We formulate the grouping of body part detections into people as a minimum-weight set packing (MWSP)
problem where the set of potential people is the power set of
body part detections. We model the quality of a hypothesis
of a person which is a set in the MWSP by an augmented
tree-structured Markov random field where variables correspond to body-parts and their state-spaces correspond to
the power set of the detections for that part.
1. Introduction
We study the problem of multi-person pose estimation
(MPPE) [1] which we model as the problem of selecting
a subset of non-overlapping proposals of people supported
by image evidence and a prior model. This formulation
of MPPE corresponds to a minimum weight set packing
(MWSP) [16] problem where elements correspond to detections of body parts and sets (referred to as poses) correspond
to subsets of those body parts detections. Poses are associated with real valued costs based on occurrence and cooccurrence probabilities of detections in a pose as defined
by a deep neural network [20, 17, 25, 2] and (augmented)
tree structured part/spring model [11, 10, 26] respectively.
This fully specifies MPPE as an optimization problem.
Since the set of poses grows exponentially in the number of detections we employ an implicit column generation
(ICG) strategy [4, 13, 14] for inference in MWSP. We exploit the augmented tree structure of the cost of a pose to
frame pricing as a dynamic program [5] where variables
correspond to body parts and the state of a given variable
corresponds to a subset of the detections for that part. Since
the state-space of each variable is enormous (power set of
detections of a given part), we introduce a tool from operations research called the nested Benders decomposition
(NBD) [7, 8] which avoids considering the vector product of
the state-spaces of adjacent variables in the tree. NBD has
been used for a variety of applications including: agricultural planning, factory production planning, and stock portfolio optimization. Our NBD formulation is guaranteed to
achieve exact inference in the pricing problem and in practice is orders of magnitude faster than naive dynamic programming. NBD exploits the fact that pricing problems are
We describe a novel algorithm that combines efficiency
with provable bounds on this MWSP problem. We employ
an implicit column generation strategy where the pricing
problem is formulated as a dynamic program. To efficiently
solve this dynamic program we exploit the problem structure utilizing a nested Bender’s decomposition (NBD) exact
inference strategy which we speed up by recycling Bender’s
rows between calls to the pricing problem.
We test our approach on the MPII-Multiperson dataset,
showing that our approach obtains comparable results with
the state-of-the-art algorithm for joint node labeling and
grouping problems, and that NBD achieves considerable
speed-ups relative to a naive dynamic programming approach. Typical algorithms that solve joint node labeling and grouping problems use heuristics and thus can not
obtain proofs of optimality. Our approach, in contrast,
proves that for over 99 percent of problem instances we
find the globally optimal solution and otherwise provide up1
where Rd indicates the part associated with detection d.
Each detection is associated with a single part prior to
MWSP optimization. For short hand we use Dr to denote
the set of detections associated with part r. We use P
to denote the set of all possible poses which we index by
p. We associate the members of P with detections using
P ∈ {0, 1}|D|×|P| which we index by d, p where Pdp = 1
indicates that detection d is in pose p. This allows us to
formulate MPPE as a search for low cost, non-overlapping
sets.
We model human poses with an augmented tree structure over the set of parts described using matrix T ∈
{0, 1}|R|×|R| . We index T with r1 , r2 where Tr1 r2 = 1 if
and only if part r2 is a child of part r1 in the augmented tree.
We use fourteen body parts (head, neck, and the left/right of
shoulder, elbow, wrist, hip, knee, ankle), as standardized by
MPII dataset. We form an augmented tree over these parts
defined by a typical tree structure over all parts excluding
the neck, followed by connecting the neck to each of the
other thirteen parts; this model design is based on the observation that in real images people’s necks are rarely occluded, thus having connections from neck to all other body
parts can handle cases where other body parts are occluded
while keeping the model relatively simple. For short hand
we use T0 to denote the root of the tree which is a part other
than the neck selected arbitrarily.
The cost of a pose is defined using terms θ1 ∈ R|D| , and
2
θ ∈ R|D|×|D| which we index with d, and d1 , d2 respectively. We refer to the θ1 , θ2 terms as unary and pairwise
respectively. We use θd1 to indicate the cost of including detection d in a pose. Similarly we use θd21 d2 to indicate the
cost of including detections d1 , d2 in a common pose. The
augmented tree structure is respected with regards to these
costs; thus θd21 ,d2 can only be non-zero if Rd1 = Rd2 or if
Rd2 is a child of Rd1 . We model a prior on the number of
poses in the image using θ0 ∈ R which is a constant offset
added to the cost of each pose. We define the mapping of
poses to costs using Γ ∈ R|P| which we index with p where
Γp is the cost associated with pose p which is defined formally below.
X
X
Γp = θ 0 +
θd1 Pdp +
θd21 d2 Pd1 p Pd2 p (1)
similar across iterations of ICG by hot starting optimization
in a given iteration with Benders rows produced in previous
iterations. The combination of ICG with NBD promises
efficient and provably optimal solutions without having to
enumerate the vector product of state-spaces.
1.1. Related Work
Our work is closely related to the sub-graph multi-cut integer linear programming (ILP) formulation of [19, 15, 18]
which we refer to as MC for shorthand. MC models the
problem of MPPE as partitioning detections into fourteen
body parts (plus false positive) and clustering those detections into poses. Clustering is done according to the correlation clustering [3] criteria with costs parameterized by the
part associated with the detection. This formulation is notable as it models non-max suppression by allowing poses
to be associated with multiple detections of a given body
part. However, the optimization problem of MC is often too
hard to solve and is thus attacked with heuristic methods.
Motivated by the difficulty of inference in MC, the work
of [24] introduces an alternative model called the two tier
formulation (TTF). TTF truncates the model of MC in such
a way to achieve fast inference using ICG. The TTF model
with ICG inference outperforms MC with regards to finding
difficult-to-localize parts such as ankle and wrist, and also
provides a marginal improvement in overall accuracy. Furthermore TTF provides some additional capacities beyond
that of MC such as a prior on the number of people in an
image. In [22] inference in the TTF model is attacked using
a non-nested Benders decomposition [12, 7] though inference is demonstrated to not be as fast as in the ICG strategy.
While the TTF model works well in practice, it does not
optimize the full cost but privileges a single exemplar detection for each body part to model inter-part co-association
costs of a pose, ignoring all inter-part co-association costs
between non-exemplar detections. Furthermore TTF requires detections to be associated with parts in advance of
optimization.
1.2. Outline
Our paper is structured as follows. In Section 2 we introduce a novel MWSP formulation for MPPE and address
inference with ICG. In Section 3 we introduce our NBD approach to solving the pricing problem of ICG. In Section
4 we present experiments on the MPII-Multiperson validation set. Finally we conclude in Section 5. We provide
additional derivations in the appendix.
d∈D
(d1 ,d2 )∈D
We have thus fully defined the cost of a single pose. In
the next section we formulate the MWSP problem using the
costs of poses.
2.1. ILP/ LP Relaxation Formulation
2. Our MWSP Formulation of MPPE
We frame the search for the lowest cost set of nonoverlapping poses as an integer linear program (ILP). We
use γ ∈ {0, 1}|P| to define a selection of poses where
γp = 1 if and only if pose p is selected. We write the constraint that selected poses do not overlap as P γ ≤ 1. We
We now formulate MPPE as MWSP. We use R,D to denote the sets of human body parts and the sets of body part
detections, which we index with r,d respectively. We describe a surjection of detections to parts using R ∈ {R}|D|
2
express the ILP along with the corresponding primal/dual
linear program (LP) relaxations as follows using Lagrange
|D|
multipliers λ ∈ R0+ which we index by d.
min Γ> γ ≥ min Γ> γ =
γ∈{0,1}
P γ≤1
γ≥0
P γ≤1
max
λ≥0
Γ+P > λ≥0
−1> λ
Algorithm 1 Implicit Column Generation Algorithm
P̂ ← {}
repeat
did add← 0
λ ← max
(2)
λ≥0
P
Γp + d∈D Pdp λd ≥0
for D̆ ⊆ Dneck do
P
p ← min p∈P Γp + d∈D Pdp λd
p↔D̆
P
if Γp + d∈D Pdp λd < 0 then
P̂ ← [P̂ ∪ p]
did add← 1
end if
end for
until did add=0
In Appendix A we study dual-optimal bounds [6, 27] on λ
that when applied can accelerate inference without loosening the LP relaxation.
2.2. Inference via Implicit Column Generation
Given that P grows exponentially in the number of detections we can not explicitly consider it during MWSP optimization. We thus construct a sufficient subset P̂ ⊂ P
using ICG so as to solve exactly the dual problem in Eq 2.
ICG consists of two alternating optimizations referred to as
the restricted master problem (RMP) and the pricing problem respectively.
Figure 1. Implicit column generation procedure. We iteratively
solve the RMP followed by pricing. During pricing we compute
one pose associated with each member of the power set of neck
detections. We add the pose to nascent set P̂ if and only if it corresponds to a violated constraint in the dual. We terminate when
no pose in P corresponds to a violated dual constraint.
• RMP: We solve the dual optimization over the set P̂
providing dual variables λ. We write dual optimization
of the RMP below.
−1> λ
max
λ≥0
P
Γp + d∈D Pdp λd ≥0
(3)
We write ICG in Alg 1 and consider the pricing problem
which is a dynamic program in Section 2.4. At termination
of Alg 1 we solve MWSP over P̂ using an ILP solver. In
practice we find that the LP relaxation provides an integral
solution at termination for over 99% of cases. If needed
we could employ branch and price [4] or tighten the bound
with odd set inequalities while preserving the structure of
the pricing problem [23].
∀p∈P̂
• Pricing: Using λ we identify a subset of the most violated constraints corresponding to members of P and
add it to P̂. This subset includes the most violated constraint over all P. When no violated constraints exist
ICG terminates. The slack in the dual constraint corresponding to a given primal variable is referred to as its
reduced cost, thus pricing identifies the lowest reduced
cost terms in the primal.
2.3. Anytime Lower Bounds
We compute an anytime lower bound on our objective by
adding a term based on the columns produced to the objective of the RMP. Recall that each detection can be assigned
to at most one body part. Thus we can rely on the proof that
the LP for MWSP can be bounded by RMP objective plus
the lowest reduced cost times the cardinality of the set of elements [9, 23] (if a negative reduced cost term exists). We
compute this lower bound below given any non-negative λ
provided by the RMP as follows.
X
X
(
−λd ) + |D| min(0, min Γp +
Pdp λd ) (5)
During the pricing step we iterate through the power
set of neck detections and compute the lowest reduced
cost pose containing exactly those neck detections. We
index the power set of neck detections with D̆ and use
p ↔ D̆ to indicate that the neck detections in p are
exactly those in D̆. We write pricing below for an arbitrary subset of the neck detections D̆.
min Γp +
p∈P
p↔D̆
X
λd Pdp
−1> λ
∀p∈P̂
(4)
d∈D
d∈D
p∈P
d∈D
Observe that the minimization in Eq 5 is computed each
time we do pricing in Alg 1.
Note that when conditioned on a specific set of neck
detections, the pairwise costs from these neck detections to all other detections can be added to unary costs
of the other detections. Thus the augmented-tree structure becomes a typical tree structure, and exact inference can be done via dynamic programming.
2.4. Pricing Using a Naive Dynamic Program
Observe that Eq 4 corresponds to computing the maximum a posteriori probability (MAP) of a Markov random
3
field (MRF) where there is bijection between parts (except
the neck) and variables in the MRF. Similarly for a variable
in the MRF there is a bijection between the state space of
that variable and the power set of detections for the associated part. This MRF is tree structured and hence amenable
to exact inference via dynamic programming which we now
consider.
follows.
We use S r to denote the state space of variable r which
r
|D r |
we describe using S r ∈ {0, 1}|D |×2
. We index S r with
r
r
r
d, s where d ∈ D and s ∈ S respectively. We use Sds
=1
r
to indicate that detection d ∈ D is included in configuration s ∈ S r . We use µr̂ŝ←r to refer to the value of the
lowest cost solution to the sub-tree rooted at r conditioned
on its parent r̂ taking on state ŝ. The cost µr̂ŝ←r includes
the pairwise interaction terms between members of Dr and
Dr̂ . We define µ formally using helper terms Φ, ψ below.
The existence of such a decomposition is a known result
of the stochastic programming literature [8]. Notice that
the minimization in Eq 7 does not require either the configuration of the children of r nor “messages” from those
children. Hence the state of T0 can be determined independently of the other variables. Similarly each variable can be
determined independently of its children given the state of
its parent. If the sets Z r are of small cardinality then solving Eq 7 is easy thus we construct a sufficient subset of Z r
denoted Z̈ r for each r (other than the root) using row generation (cutting planes). We refer to the collection of the
nascent sets as Z̈. Given the nascent sets Z̈ we construct a
lower bound on µr̂ŝ←r denoted µ−
r̂ŝ←r defined below using
helper function µ∗−
rs .
µr̂ŝ←r = minr Φŝs + ψrs +
s∈S
X
µrs←r̄
µr̂ŝ←r = minr Φŝs + µ∗rs
s∈S
X
X
∗
r
maxr̄ (Ωr̄z0 +
µrs = ψrs +
Ωr̄zd Sds
)
r̄∈R
Trr̄ =1
(6)
r̄∈Trr̄=1
Φŝs =
X
θd21 d2 Sdr̂1 ŝ Sdr2 s
X
d∈D r
r
(θd1 + λd )Sds
+
X
θd21 d2 Sdr1 s Sdr2 s
r̄∈R
Trr̄ =1
d1 ∈D r
d2 ∈D r
+
X
d∈D r
µ−
minr Φŝs + µ∗−
rs
r̂ŝ←r = s∈S
X
X
r
µ∗−
max (Ωr̄z0 +
Ωr̄zd Sds
)
rs = ψrs +
d1 ∈D r̂
d2 ∈D r
ψrs =
z∈Z
z∈Z̈ r̄
(7)
(8)
d∈D r
The root variable T0 is not associated with µ−
r̂ŝ←r terms
since it has no parent but it is associated with µ∗−
rs terms.
θd21 d2 Sdr2 s
d1 ∈D̆
d2 ∈D r
3.1. Overview of Constructing Z̈
We now consider the construction of small sufficient sets
∗
Z̈ such that mins∈S T0 µ∗−
T0 s = mins∈S T0 µT0 s . We outline the remainder of this section as follows. In Section
3.2 we produce upper/lower bounds on the MAP of the
MRF, which are identical at termination of NBD. The upper
bound is accompanied by a configuration with cost equal to
the upper bound. In Section 3.3 we compute the gap between the upper and lower bounds introduced at each variable in the tree and select the variable r∗ associated with
the largest increase in the gap. Then in Section 3.4 we add
a Benders row to Z̈ r∗ . Next in Section 3.5 we increase
Ωr0z terms hence tightening the relaxation without generating new rows. Then in Section 3.6 we combine the steps
above to produce a complete NBD inference approach. Finally we provide implementation details in Section 3.7.
Computing µr̂ŝ←r for all ŝ ∈ S r̂ is expensive since we need
to search over all possible combinations of state spaces of
two adjacent variables, where the number of possible states
of each variable can be up to 50k in our experiments.
3. A Nested Benders Decomposition Alternative to Naive Dynamic Programming
We now consider NBD as an alternative to naive dynamic programming that avoids the expensive computation
of µr̂ŝ←r for all ŝ ∈ S r̂ . We express µ using the sum of
convex functions each constructed from the maximum of a
unique set of affine functions called Benders rows. Specifically for any part r ∈ R (other than the root T0 ) we denote the corresponding set of Benders rows as Z r which we
r
r̂
index by z. We describe Z r using Ωr ∈ R|Z |×(1+|D |)
where r̂ is the parent of r and index with (z, d) or (z, 0).
For a given z we use Ωrzd to indicate value associated with
d ∈ Dr̂ and Ωrz0 to be the offset. Using Z we provide an
alternative description of µ using helper term µ∗ defined as
3.2. Producing a Configuration and Corresponding
Bounds using Nested Benders Decomposition
In this section produce a configuration for the MRF by
proceeding from root to leaves and selecting the state for
a given variable r given the state of its parent only. We
describe the configuration produced using x where xr ∈
4
|D r |
S r . We use xr̂ to refer to the state of the parent of r. The
process of producing x is defined below.
R0+ .
min Φŝs + µ∗−
rs
s∈S r
xT0 ← arg min µ∗−
T0 s
(9)
s∈S T0
xr ← arg minr Φxr̂ s + µ∗−
rs
∀r ∈ R − T0
s∈S
(11)
=
X
min
P
s∈S r [xr =s]=1
−yd1 d2 +Sdr̂
The cost of the configuration is an upper bound on the
MAP and is associated with a lower bound on the MAP
mins∈S T0 µ∗−
T0 s .
1 ŝ
P
+ s∈S r [xr =s]Sdr
2s
≤1
∈R
δd1z d ≥0
δd2z d ≥0
Mr2 = max (Ωrz0 +
z∈Z̈ r
Mr3
=
(Mr1
1 2
1 2
s.t.
−
−
X
+ δ 0z +
(δd1z1 d2 − δd3z1 d2 )Sdr2 s ≥ 0
θd21 d2 − δd1z1 d2 + δd2z1 d2 + δd3z1 d2 ≥ 0
∀s ∈ S r
∀(d1 ∈ Dr̂ , d2 ∈ Dr )
After computing δ 0z , δ 1z , δ 2z , δ 3z we produce a new Benders row denoted z that is defined as follows.
X
Ωrz0 ← −δ 0z −
δd1z1 d2
(12)
d1 ∈D r̂
d2 ∈D r
(10)
Ωrzd1 ←
X
d2
r̂
Ωrzd Sdx
)
r̂
X
µ∗−
rs
d1 ∈D r̂
d2 ∈D r
δd1z1 d2 − δd2z1 d2
d1 ∈ Dr̂
∈D r
When solving optimization in the dual of Eq 11 we add
a tiny negative bias to objective corresponding to terms
δ 1z , δ 2z , δ 3z . This ensures that the corresponding terms do
not increase beyond what is needed to produce an optimal
dual solution, which stabilizes optimization. The additional
small
may be understood intuitively as ensuring that
P biases
1
2
δ
d2 ∈D r d1 d2 − δd1 d2 corresponds to the marginal cost for
using d1 in the solution. In Appendix B we solve the dual
LP in Eq 11 efficiently by reducing it to an equivalent LP
with |Dr | variables and far fewer constraints.
d∈D r̂
Mr2 )
δd1z1 d2 (Sdr̂1 ŝ − 1) − δd2z1 d2 Sdr̂1 ŝ
δd3z d ≥0
r̄∈Tr,r̄ =1
X
θd21 d2 yd1 d2
d1 ∈D r̂
d2 ∈D r
1 2
In this section we identify the variable in the tree associated with the largest increase in the gap between the upper and lower bounds given a configuration x produced in
Section 3.2. The gap between upper and lower bounds introduced at r is the difference between the upper and lower
bounds at variable r minus the corresponding gaps at its
children. We use Mr1 , Mr2 , Mr3 , which are defined below, to
denote the cost of the sub-tree rooted at r; the corresponding lower bound; and the gap introduced at r respectively.
= Φxr̂ xr + ψxr̂ xr +
X
= max
−δ 0z +
0z
δ
Mr̄1
X
d1 ∈D r̂
d2 ∈D r
3.3. Computing the Gap Introduced at each Variable in the Tree
X
s∈S r
yd1 d2 ≤Sdr̂ ŝ
1
P
yd1 d2 ≤ s∈S r [xr =s]Sdr s
2
x≥0
y≥0
+
Mr1
[xr = s]µ∗−
rs
(Mr̄1 − Mr̄2 )
r̄∈Tr,r̄ =1
3.4. Generating Benders Rows
In this section we identify the most violated Benders row
denoted z for a given r, given that its parent r̂ takes on ŝ
and the Z̈ sets. We formulate this as a small scale linear
program described below.
3.5. Re-using Rows: Rapidly Updating δ 0z while
leaving δ 1z , δ 2z , δ 3z fixed
In this sub-section we provide a complementary mechanism to generating new Benders rows. This mechanism
is motivated by the observation that µ∗−
rs may increase (but
never decrease) when Benders rows are added to the descendants of a given variable r. This mechanism takes in existing Benders rows and sets the corresponding δ 0z term to
We use decision variable [xr = s] = 1 to indicate that
s is the state associated
with variable r. We use decision
|D r̂ |×|D r |
variable y ∈ RP
which we index by d1 , d2 where
0+
yd1 d2 = Sdr̂1 ŝ ( s∈S r [xr = s]Sdr2 s ). We introduce dual
variables δ 0z ∈ R and δ 1z , δ 2z , δ 3z each of which lie in
5
the minimum feasible value thus tightening the corresponding constraint while leaving δ 1z , δ 2z , δ 3z fixed. This task is
faster than generating Benders rows via the method of Section 3.4.
Observe that given δ 1z , δ 2z , δ 3z satisfying θd21 d2 −
1z
δd1 d2 + δd2z1 d2 + δd3z1 d2 ≥ 0 there always exists a feasible
setting of δ 0z . Given fixed δ 1z , δ 2z , δ 3z we select the smallest feasible value for δ 0z as follows.
X
δ 0z ← −(minr µ∗−
(δd1z1 d2 − δd3z1 d2 )Sdr2 s ) (13)
rs +
s∈S
Algorithm 2 Overall Algorithm of Nested Benders Given
Initial Z̈
repeat
Step 1: Update δ 0 terms
for r ∈ R from leaves to children of the root do
for z ∈ Z̈ r do
P
1z
δ 0z ← − mins∈S r µ∗−
rs +
d1 ∈D r̂ (δd1 d2 −
d2 ∈D r
δd3z1 d2 )Sdr2 s
P
Ωrz0 ← −δ 0z − d∈Dr δd1z
end for
end for
STEP 2: Compute configuration and bounds
xT0 ← arg mins∈S T0 µ∗−
T0 s
for r ∈ R from children of T0 to leaves do
xr ← arg mins∈S r Φxr̂ s + µ∗−
rs
end for
p ← pose corresponding
to x
P
U B ← Γp + d∈D λd Pdp
LB ← mins∈S T0 µ∗−
T0 s
STEP 3: Select variable to add Benders row to
for r ∈ R from leaves to children
P of T0 do
Mr1 ← Φxr̂ xr + ψxr̂ xr + r̄∈Tr,r̄ =1 Mr̄1
P
r̂
Mr2 = maxz∈Z̈ r (Ωrz0 + d∈Dr̂ Ωrzd Sdx
)
r̂
P
2
1
1
3
Mr = (Mr − Mr ) − r̄∈Tr,r̄ =1 (Mr̄ − Mr̄2 )
end for
r∗ ← arg maxr∈R−T0 Mr3
STEP 4: Generate a new Benders row for r∗
if U B 6= LB then
generate δ 0z , δ 1z , δ 2z , δ 3z by solving dual in Eq
11 given parent taking on
Pstate xr̂∗ .
Ωrz0∗ ← −δ 0z − d1 ∈Dr̂ δd1z1 d2
d1 ∈D r̂
d2 ∈D r
3.6. Our Complete Nested Benders Decomposition
Algorithm
Our NBD approach iterates through the following steps.
1. Step 1: Proceeding from the leaves to the children of
the root: Set δ 0z terms to the minimum feasible value
given fixed δ 1z , δ 2z , δ 3z . This is done on the first iteration of NBD only if Z̈ r is not empty for each r.
2. Step Two: Proceed from root to leaves: select the state
for each variable conditioned on its parent (if it has
one) and the Benders rows associated with its children.
This produces upper and lower bounds associated with
each variable in the tree.
3. Step Three: Select the variable r∗ corresponding to
the largest increase in the gap between the upper and
lower bounds in the tree.
4. Step Four: Add a new Benders row to Z̈ r∗
We repeat this procedure until no additional Benders rows
need be added at which point the configuration produced in
step two is guaranteed to be the global optima. We formalize this procedure in Alg 2.
Ωrzd∗ 1
r∗
P
d2 ∈D r
1z
2z
d2 ∈D r δd1 d2 − δd1 d2
←
← Z̈ r∗ ∪ z
d1 ∈ Dr̂
Z̈
end if
until UB=LB
3.7. Implementation Details
In this section we provide implementation details with
regards Alg 2.
Accelerating Step One of Alg 2: Observe that δ 0z terms
associated with Z̈ r can only be decreased if a new Benders
row is added to one of the descendants of r. Thus when
executing Alg 2 we only update the δ 0 terms associated with
the ancestors of the variable r∗ which had its Benders row
set augmented. Observe that we need only update the δ 0z
terms associated with the leaves in the first iteration of NBD
in a given call from ICG.
Accelerating Step Two of Alg 2. Recall that the choice
of the state of a given variable r is a function of the Benders
rows associated with its children and the configuration of
the parent (if it has a parent). At any iteration of NBD other
than the first we consider the previous configuration of the
MRF produced in step two of Alg 2 and only update the
state of a variable if the state of its parent was changed or
if it is an ancestor of the variable r∗ which had its Benders
row set augmented.
Limiting the Number of Neck Detections in a Pose:
We found that our best results with regards to timing and
modeling occur when we require that each pose include exactly one neck detection.
Limiting State Space of a Variable: We limit the number of states of a given variable to a given user defined parameter value V (V =50,000). We construct this set as follows. We begin with the state corresponding to zero detec6
(a)
(b)
Figure 2. Timing comparison and speed-ups achieved by NBD. (a) Accumulated running time over problem instances for NBD and DP,
respectively. (b) Factor of speed-up of NBD relative to DP, as a function of computation time spent for DP pricing. Note that in general the
factor of speed-up grows as the problem gets harder for DP.
constant bias θ0 = 60 (set by hand) to regularize the number of people in the solution.
We compare solutions found by NBD and DP at each
step of ICG; for all problem instances and all optimization
steps, NBD obtains exactly the same solutions as DP (up
to a tie in costs). Comparing total time spent doing NBD
vs DP across problem instances we found that NBD is 44x
faster than DP, and can be up to 500x faster on extreme
problem instances. Comparison of accumulated running
time used by NBD and DP over all 418 instances are shown
in Fig. 2. We observe that the factor speed up provided by
NBD increases as a function of the computation time of DP.
With regards to cost we observe that the integer solution
produced over P̂ is identical to the LP value in over 99%
of problem instances thus certifying that the optimal integer solution is produced. For those instances on which LP
relaxation fails to produce integer results, the gaps between
the LP objectives and the integer solutions are all within
1.5% of the LP objectives.
For the sake of completeness, we also report MPPE accuracy in terms of average precisions (APs) and compare
it against a state-of-the-art solver [18] which uses primal
heuristics. Note that the cost formulation of [18] differs
from ours in that it allows a pose to be associated with multiple neck detections or none, while our model requires that
each pose must have exactly one neck detection and maps
detections to parts prior to ICG. Also our model includes a
prior on the number of poses as modeled by θ0 . As shown in
Table 1, we achieve equivalent results to [18]. We note that
although our algorithm does not run as fast as [18], our code
is implemented in pure MATLAB and can benefit further
from using commercial LP solvers and parallelizing pricing
routines. More importantly, our formulation provides up-
tions included, then add in the group of states corresponding
to one detection included; then add in the group of states
corresponding to two detections included etc. If adding a
group would have the state space exceed V states for the
variable we don’t add the group and terminate.
Caching Integrals: We accelerate Alg 2 by storing
the value of repeatedly used integrals that do not change
in value over the course of optimization.
time a new z is produced we store
P Thus each
r
r̂
Ω
S
for each ŝ ∈ S r̂ . Similarly we store
r̂
zd
dŝ
Pd∈D
1z
3z
r
r
d1 ∈D r̂ (δd1 d2 − δd1 d2 )Sd2 s for each s ∈ S .
d2 ∈D r
Initializing Z̈: We do not initialize Z̈ r with any Benders rows for the first round of pricing in ICG. Thus the initial state for a variable in NBD ignores the existence of its
children and the corresponding initial lower bound is −∞.
Timing Observation: Experimentally we observe that
the total time consumed by steps in NBD is ordered from
greatest to least as [1,2,4,3]. Note that the step solving the
LP is the second least time consuming step of NBD.
Selecting the Root: Observe that Alg 2 requires solving
LPs in step four for variables except the root. The number
of constraints in the LP for part r is exponential in the size
of |Dr |. We avoid solving the largest LP by selecting as the
root the part associated with the most detections. Alternatively we could select the root so as to have a more balanced
tree with the goal of leveraging parallel processing.
4. Experiments
We evaluate our approach against a naive dynamic programming based formulation on MPII-Multiperson validation set [1], which consists of 418 images. The unary θ1 and
pairwise θ2 costs are trained using the code of [15], with a
7
Part
Head
Shoulder
Elbow
Wrist
Hip
Knee
Ankle
mAP(UBody)
mAP
time (s/frame)
Ours
90.6
87.3
79.5
70.1
78.5
70.5
64.8
81.8
77.6
1.95
[18]
93.0
88.2
78.2
68.4
78.9
70.0
64.3
81.9
77.6
0.136
Table 1. We display average precision of our approach versus [18]. Running times are measured on an Intel i7-6700k quad-core CPU.
Figure 3. Example output of our system.
References
per/lower bounds and in over 99% of cases certificates of
optimality.
[1] M. Andriluka, L. Pishchulin, P. Gehler, and B. Schiele. 2d
human pose estimation: New benchmark and state of the art
analysis. In Proc. of CVPR, 2014.
[2] P. Baldi, P. Sadowski, and D. Whiteson. Searching for exotic
particles in high-energy physics with deep learning. Nature
communications, 5, 2014.
5. Conclusion
We have described multi-person pose estimation as a
minimum-weight set packing (MWSP) problem which we
address using implicit column generation. We solve the
corresponding pricing problem using a novel nested Benders decomposition (NBD) approach, which reuses Bender’s rows between calls to NBD. For over 99% of cases
we find provably optimal solutions, which is practically important in domains where knowledge of certainty matters,
such as interventions in rehabilitation. Our procedure for
solving the pricing problem vastly outperforms a baseline
dynamic programming approach. We expect that NBD will
find many applications in machine learning and computer
vision, especially for solving dynamic programs with large
state spaces for individual variables. For example we could
formulate sub-graph multi-cut tracking [21] as a MWSP
with pricing using NBD.
[3] N. Bansal, A. Blum, and S. Chawla. Correlation clustering.
In Journal of Machine Learning, pages 238–247, 2002.
[4] C. Barnhart, E. L. Johnson, G. L. Nemhauser, M. W. P.
Savelsbergh, and P. H. Vance. Branch-and-price: Column
generation for solving huge integer programs. Operations
Research, 46:316–329, 1996.
[5] R. E. Bellman. Dynamic Programming. Dover Publications,
Incorporated, 2003.
[6] H. Ben Amor, J. Desrosiers, and J. M. Valério de Carvalho.
Dual-optimal inequalities for stabilized column generation.
Operations Research, 54(3):454–463, 2006.
[7] J. F. Benders. Partitioning procedures for solving mixedvariables programming problems. Numerische mathematik,
4(1):238–252, 1962.
8
[26] Y. Yang and D. Ramanan. Articulated pose estimation with
flexible mixtures-of-parts. In Proc. of CVPR, 2011.
[8] J. R. Birge. Decomposition and partitioning methods for
multistage stochastic linear programs. Operations research,
33(5):989–1007, 1985.
[9] J. Desrosiers and M. E. Lübbecke. A primer in column generation. In Column generation, pages 1–32. Springer, 2005.
[10] P. Felzenszwalb, D. McAllester, and D. Ramanan. A discriminatively trained, multiscale, deformable part model. In
Proc. of CVPR, 2008.
[11] P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan. Object detection with discriminatively trained partbased models. IEEE transactions on pattern analysis and
machine intelligence, 32(9):1627–1645, 2010.
[12] A. M. Geoffrion and G. W. Graves. Multicommodity distribution system design by benders decomposition. Management science, 20(5):822–844, 1974.
[13] P. Gilmore and R. Gomory. A linear programming approach
to the cutting-stock problem. Operations Research (volume
9), 1961.
[14] P. Gilmore and R. E. Gomory. Multistage cutting stock problems of two and more dimensions. Operations research,
13(1):94–120, 1965.
[15] E. Insafutdinov, L. Pishchulin, B. Andres, M. Andriluka, and
B. Schiele. Deepercut: A deeper, stronger, and faster multiperson pose estimation model. CoRR, abs/1605.03170, 2016.
[16] R. M. Karp. Reducibility among combinatorial problems.
The IBM Research Symposia Series, pages 85–103. Plenum
Press, New York, 1972.
[17] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
Proc. of NIPS, 2012.
[18] E. Levinkov, J. Uhrig, S. Tang, M. Omran, E. Insafutdinov,
A. Kirillov, C. Rother, T. Brox, B. Schiele, and B. Andres.
Joint graph decomposition and node labeling: Problem, algorithms, applications. In Proc. of CVPR, 2017.
[19] L. Pishchulin, E. Insafutdinov, S. Tang, B. Andres, M. Andriluka, P. Gehler, and B. Schiele. DeepCut: Joint subset
partition and labeling for multi person pose estimation. In
Proc. of CVPR, 2016.
[20] D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Parallel distributed processing: Explorations in the microstructure of cognition, vol. 1. chapter Learning Internal Representations by Error Propagation, pages 318–362. MIT Press,
Cambridge, MA, USA, 1986.
[21] S. Tang, B. Andres, M. Andriluka, and B. Schiele. Subgraph
decomposition for multi-target tracking. In CVPR, 2015.
[22] S. Wang, K. Kording, and J. Yarkony. Exploiting skeletal
structure in computer vision annotation with benders decomposition. arXiv preprint arXiv:1709.04411, 2017.
[23] S. Wang, S. Wolf, C. Fowlkes, and J. Yarkony. Tracking objects with higher order interactions via delayed column generation. In Artificial Intelligence and Statistics, pages 1132–
1140, 2017.
[24] S. Wang, C. Zhang, M. A. Gonzalez-Ballester, A. Ihler, and
J. Yarkony. Multi-person pose estimation via column generation. arXiv preprint arXiv:1709.05982, 2017.
[25] R. Wu, S. Yan, Y. Shan, Q. Dang, and G. Sun. Deep
image: Scaling up image recognition. arXiv preprint
arXiv:1501.02876, 7(8), 2015.
[27] J. Yarkony and C. Fowlkes. Planar ultrametrics for image
segmentation. In Proc. of NIPS, 2015.
9
A. Appendix: Dual Optimal Inequalities on λ
dominates NBD computation. To achieve this we make the
following observations about the primal LP form in Eq 11.
In this section we provide upper bounds on the Lagrange multipliers λ called dual optimal inequalities (DOI)
[6] which are computed prior to ICG. The use of DOI decreases the search space that ICG needs to explore and thus
decreases the number of iterations of pricing required.
Observe that at any given iteration of ICG the optimal
solution to the primal LP relaxation need not lie in the span
of P̂. If limited to producing a primal solution over P̂ it
is useful to allow some values of P γ exceed one. We in|D|
troduce a slack vector ξ ∈ R0+ indexed by d that tracks
the presence of any detections included more than once and
prevents them from contributing to the objective when the
corresponding contribution is negative. To do this we offset
the cost for “over-including” a detection with a cost that at
least compensates and likely overcompensates.
Observe that the removal of a detection d from a pose
2
removes from the cost the associated θd1 , θdd
, θd21 d for any
1
d1 in the pose and if d is the only detection the θ0 term.
We define Ξd such that it is an upper bound on the increase
in the cost of a pose given that d is removed. To express
Ξd compactly we introduce the following terms θd2+
=
1 d2
2
0−
0
max(0, θd21 d2 ),θd2−
=
min(0,
θ
),θ
=
min(0,
θ
).
d1 d2
1 d2
We define Ξd as follows.
X
2−
Ξd = − min(0, θ0− + θd1 +
θdd
+ θd2−
)
(14)
1
1d
• Given θd21 d2 ≤ 0 or Sdr̂1 ŝ = 0: The constraint −yd1 d2 +
P
Sdr̂1 ŝ + s∈S r [xr = s]Sdr2 s ≤ 1 is inactive so δd1z1 d2 =
0.
• Given θd21 d2 ≥ 0 or Sdr̂1 ŝ = 1: The constraint yd1 d2 ≤
Sdr̂1 ŝ is inactive so δd2z1 d2 = 0.
P
• Given θd21 d2 ≥ 0 the constraint yd1 d2 ≤ s∈S r [xr =
s]Sdr2 s inactive so δd3z1 d2 = 0.
Observe that the following is true any pair d1 ∈ Dr̂ , d2 ∈
Dr such that θd21 d2 < 0.
θd21 d2 + δd2z1 d2 + δd3z1 d2 ≥ 0
Since δd2z1 d2 is non-negative and associated with a nonnegative term in the objective we observe the following.
θd21 d2 + δd2z1 d2 = −δd3z1 d2
−θd21 d2
Sdr̂
1
p∈P
p∈P
1 2
1 2
)θd2−d ≥δd2z d
1 2
1 2
(18)
s.t.
µ∗−
rs
+δ
0z
Ξ≥λ≥0
Γ+P > λ≥0
X
+
(δd1z1 d2
+
δd2z1 d2
+
θd2−
)Sdr2 s
1 d2
≥0
d1 ∈D r̂
d2 ∈D r
d∈D
max
d1 ∈D r̂
d2 ∈D r
1 ŝ
∀d∈D
=
δ ∈R
δ 1z ≥0
δ 2z ≥0
2+
1z
θ
ŝ d d ≥δd d
−(1−Sdr̂
The expanded MWSP objective and its dual LP relaxation
are given below:
X
X
min
Γp γp +
Ξd ξd
(15)
γp ∈{0,1}
ξ≥0
Pdp γp −ξd ≤1
≥
(17)
δd2z1 d2
Using these observations we rewrite optimization.
X
max
δd1z1 d2 (Sdr̂1 ŝ − 1) − δd2z1 d2 Sdr̂1 ŝ − δ 0z
0z
d1 ∈D
P
(16)
Now observe that δ 0z is the only term that is associated with
a non-zero objective and which is not bound to zero thus
it has the same value as the primal LP and therefor δ 0 =
mins∈S r Φŝs + µ∗−
rs .
Observe that for all δ 1z , δ 2z terms not bound to zero for a
given d2 that they co-occur in the objective with value zero
and in each constraint over S r with the common value. Using this we merge δ 1z , δ 2z terms across d2 as follows using
r
δ 4z ∈ R|D | which we index by d2 and helper term Qd2 .
X
δd4z2 =
δd1z1 d2 + δd2z1 d2
(19)
−1> λ
Observe that the dual relaxation bounds λ by Ξ from above.
These bounds are called DOI.
For cases where a pose is required to include a neck detection we can not use this bound for neck detections as the
removal of the neck makes the pose invalid. Therefore we
ignore the θ0− term when computing Ξ and for Rd = neck
we set Ξd = ∞.
To ensure that the DOI are not active at termination of
ICG we offset Ξ with a small positive constant.
d1 ∈D r̂
r̂
δd1z1 d2 = δd4z2 Sŝd
1
B. Appendix: Deriving a Compressed LP for
Benders Row Generation
r̂
δd2z1 d2 = δd4z2 (1 − Sŝd
)
1
In this Section we compress the dual form of the LP in
Eq 11 so as to accelerate inference. In fact by compressing
the LP we observe that optimization of this LP no longer
Qd2 =
X
d1
10
∈D r̂
θd2+
1 d2
Qd2
−θd2−
1 d2
Qd2
r̂
Sdr̂1 ŝ θd2+
− (1 − Sŝd
)θd2−
1
1 d2
1 d2
We add a tiny magnitude negative objective to δd4z2 for each
d2 so as to ensure that the smallest valued solution is produced. This ensures that that the corresponding terms do not
increase beyond what is needed to produce a dual feasible
solution which stabilizes optimization. We use = 10−10
to express this.
X
max
−
δd4z2
(20)
4z
Qd2 ≥δd ≥0
2
d2 ∈D r
0z
s.t. µ∗−
+
rs + δ
X
d2 ∈D
(δd4z2 +
r
X
θd2−
)Sdr2 s ≥ 0
1 d2
d1 ∈D r̂
We now remove constraints in the LP corresponding to
0z
members
of S rP
(constraints of the form µ∗−
+
rs + δ
P
2−
4z
r
+
θ
)S
≥
0)
without
altering
r (δ
r̂
d2
d2 s
d1 ∈D
d1 d2
d2 ∈D
the solution of the LP.
Observe that the slack in these constraints does not decrease as δ 4z increases. Thus to determine which constraints of the form above that we need not consider we set
δ 4z to be the zero vector. Any such constraints that are are
not violated will not be violated for any setting of δ 4z and
so can be ignored when solving the LP. In practice we find
that that the proportion of the constraints that we remove is
very large and thus we achieve considerable time savings
when solving the LP.
11
| 1cs.CV
|
Optimal Longest Paths by Dynamic Programming∗
Tomáš Balyo1 , Kai Fieger1 , and Christian Schulz1,2
1
Karlsruhe Institute of Technology, Karlsruhe, Germany
{tomas.balyo, christian.schulz}@kit.edu, [email protected]
University of Vienna, Vienna, Austria
[email protected]
2
arXiv:1702.04170v1 [cs.DS] 14 Feb 2017
Abstract
We propose an optimal algorithm for solving the longest path problem in undirected weighted
graphs. By using graph partitioning and dynamic programming, we obtain an algorithm that is
significantly faster than other state-of-the-art methods. This enables us to solve instances that
have previously been unsolved.
1998 ACM Subject Classification G.2.2 Graph Theory
Keywords and phrases Longest Path, Graph Partitioning
Digital Object Identifier 10.4230/LIPIcs.SEA.2017.
1
Introduction
The longest path problem (LP) is to find a simple path of maximum length between two given
vertices of a graph where length is defined as the number of edges or the total weight of the
edges in the path. The problem is known to be NP-complete [5] and has several applications
such as designing circuit boards [8, 7], project planning [2], information retrieval [15] or
patrolling algorithms for multiple robots in graphs [9]. For example, when designing circuit
boards where the length difference between wires has to be kept small [8, 7], the longest
path problem manifests itself when the length of shorter wires is supposed to be increased.
Another example application is project planning/scheduling where the problem can be used
to determine the least amount of time that a project could be completed in [2].
We organize the rest of paper as follows. After introducing basic concepts and related
work in Section 2, we present our main contribution in Section 3: an optimal algorithm
for the longest path problem in undirected graphs. The main ingredient of our algorithm
is a dynamic programming technique based on hierarchical partitions of the graph. A
summary of extensive experiments done to tune the algorithm and evaluate its performance
is presented in Section 4. This includes a study of the algorithm’s performance with respect
to different partitioning strategies in order to find a good balance between the time spent
for partitioning and the overall runtime of our algorithm. We also compare our algorithm
with optimal algorithms presented in recent literature [14]. Experiments show that our new
algorithm solves significantly more instances and is also faster than other optimal algorithms
on large/hard instances. Finally, we conclude with Section 5.
∗
This work was partially supported by DFG grants SA 933/11-1.
© Tomas Balyo, Kai Fieger, Christian Schulz;
licensed under Creative Commons License CC-BY
16th International Symposium on Experimental Algorithms (SEA 2017).
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
XX:2
Optimal Longest Paths by Dynamic Programming
2
2.1
Preliminaries
Definitions
In the following we consider an undirected graph G = (V, E, ω) with edge weights ω :
P
E → R≥0 , n = |V |, and m = |E|. We extend ω to sets, i.e., ω(E 0 ) :=
e∈E 0 ω(e).
N (v) := {u : {v, u} ∈ E} denotes the neighbors of v. A subgraph is a graph whose vertex
and edge set are subsets of another graph. We call a subgraph induced if it has every possible
edge. A subset of a graph’s vertices is called a clique if the graph contains an edge between
every two distinct vertex of the subset. A matching is a subset of the edges of a graph where
no two edges have vertices in common. A sequence of vertices s → · · · → t such that each
pair of consecutive vertices is connected by an edge, is called an s-t path. We say that s is
the source and t is the target. A path is called simple if it does not contain a vertex more
than once. The length of a path is defined by the sum of its edge weights. If the graph is
unweighted, then edge weights are assumed to be one.
Given a graph G = (V, E, ω) as well as two vertices s, t ∈ V , the longest path (LP) problem
is to find the longest simple path from s to t. Another version of the longest path problem is
to find the longest simple path between any two vertices. However, the problem can be solved
by introducing two vertices s, t, connecting them to all other vertices in the graph by edges of
weight zero and then running algorithms tackling the LP problem on the modified instance.
A k-way partition of a graph is a division of V into blocks of vertices V1 ,. . . ,Vk , i.e.
V1 ∪ · · · ∪ Vk = V and Vi ∩ Vj = ∅ for i =
6 j. A balancing constraint demands that
∀i ∈ {1..k} : |Vi | ≤ Lmax := (1 + )d |Vk | e for some imbalance parameter . The objective is
P
typically to minimize the total cut i<j w(Eij ) where Eij := {{u, v} ∈ E : u ∈ Vi , v ∈ Vj }.
2.2
Related Work
This paper is based on the bachelor thesis of Kai Fieger [4]. Previous work by Stern et al. [14]
mainly focuses on the possibility of applying algorithms that are usually used to solve the
shortest path problem (SP) to the longest path problem. Stern et al. [14] make clear why
LP is so difficult compared to SP. Several algorithms are presented that are frequently used
to solve SP or other minimization search problem. They are modified in order to be able
to solve LP. The search algorithms are put into three categories: heuristic, uninformed and
suboptimal. Each of the algorithms in the first two categories yields optimal solutions to the
problem. The most relevant category for this paper is heuristic searches. Here, a heuristic
can provide extra information about the graph or the type of graph. Heuristic searches
require a heuristic function that estimates the remaining length of a solution from a given
vertex of the graph. This can give important information helping to prune the search space.
Stern et al. [14] show that heuristic searches can be used efficiently for the longest path
problem. Some examples of algorithms in this category are Depth-First-Branch-and-Bound
(DFBnB) and A*. Another category represents “uninformed” searches, which do not require
any information other than what is already given in the definition of the problem. Examples
from this category are Dijksra’s algorithm or DFBnB without a heuristic. Modifying these
algorithms to fit LP basically leads to brute force algorithms, which means that they still
have to look at every possible path in the search space. Hence, these uninformed search
strategies are not very beneficial for LP. The last category are the suboptimal searches. The
authors looked at a large number of these algorithms that only find approximations of a
longest path. They are, however, not relevant for this paper since we present an optimal
algorithm. We are not aware of any recent work other than [14] related to optimal LP solving.
T. Balyo et. al.
XX:3
Karlsruhe High Quality Partitioning
Within this work, we use the open source multilevel graph partitioning framework KaHIP [10,
12] (Karlsruhe High Quality Partitioning). More precisely, we use the distributed evolutionary
algorithm KaFFPaE contained therein to create partitions of our graphs that are better
suited for our dynamic programming approach. The algorithms in KaHIP have been able to
improve the best known partitioning results in the Walshaw Benchmark [13] for many inputs
using a short amount of time to create the partitions. We refer the reader to [11] for more
details.
3
Longest Path by Dynamic Programming
We now introduce the main contribution of our paper which is a new algorithm to tackle the
longest path problem based on principles of dynamic programming. Hence, our algorithm is
called “Longest Path by Dynamic Programming” (LPDP). Our algorithm solves the longest
path problem (LP) for weighted undirected graphs. We start this section by introducing the
main approach which includes preprocessing the graph as well as combining paths. At the
end of the section, we show how to improve the approach by using hierarchical partitions of
the input instance and by reducing the number of auxiliary nodes which reduces the size of
the search space.
3.1
The Basic Approach
A simple way to solve the longest path problem is exhaustive depth-first search [14]. A regular
depth-first search (DFS) starts at a root vertex. By default, a vertex has two states: marked
and unmarked. Initially, all vertices are unmarked, except the root which is marked. DFS
calls itself recursively on each unmarked vertex reachable by an edge from the current vertex,
which is the parent of these vertices. The vertex is marked. Once it is done it backtracks to
its parent. The search is finished once DFS backtracks at the root vertex.
Exhaustive DFS is a DFS that unmarks a vertex upon backtracking. In that way every
simple path in the graph starting from the root vertex is explored. The LP problem can be
solved with exhaustive DFS by using the start vertex as root. During the search the length
of the current path is stored and compared to the previous best solution each time the target
vertex is reached. If the current length is greater than that of the best solution, it is updated
accordingly. When the search is done a path with maximum length from s to t has been
found. The main idea of LPDP is to partition a graph into multiple blocks and run a search
similar to exhaustive DFS on each block in a preprocessing step. Afterwards the results can
be combined into a single longest path for G.
Partitioning and Preprocessing
We now explain our preprocessing routine. First of all, we partition a graph into a predefined
number of blocks and modify the input instance. Partitioning can be done using a graph
partitioning algorithm, e.g. KaHIP. We then replace every cut edge e = {x, y}, i.e. an edge
running between two blocks, by introducing two new vertices ve ,ve0 and the edges {x, ve }
and {ve0 , y}. One of these edges retains the weight of the original edge e, the other edge
weight is set to zero. Additionally, we insert two new vertices. One is connected to the start
vertex and the other one to the target vertex. In both cases, we use an edge having weight
zero. All newly generated vertices are called auxiliary-vertices throughout the rest of the
section. Next, we compute subgraphs GA := (VA , EA ) where VA is the set vertices of block
SEA 2017
XX:4
Optimal Longest Paths by Dynamic Programming
A as well as all auxiliary-vertices connected to them, and EA are all edges that run between
those vertices. See Figure 1 for an example.
Now observe the following property: for a longest simple path in G the auxiliary-vertices
function as entry and exit points for their block of the partition. That means a longest simple
path from s to t can only enter and exit a block through the inserted auxiliary-vertices. For
a block that does not contain s or t, every time the path enters a block, it also has to leave it
again, connecting auxiliary-vertices in pairs of two. The sets of auxiliary-vertex-pairs PA that
we are interested in for any subset s of the auxiliary-vertex-pairs for any block are equivalent
to the matchings that exist for a clique-graph consisting of all auxiliary-vertices of that block.
In other words, each endpoint can only appear once in a subset of pairs. The pairs have to
be connected by non-intersecting simple paths. Our preprocessing algorithm computes the
longest of these connections for each block and set of auxiliary-vertex-pairs with a modified
version of exhaustive DFS executed on the respective subgraph, whose descriptions follows.
Our modified exhaustive DFS is executed multiple times each time using a different
auxiliary-vertex as a root. The search algorithm divides its current search path into different
path segments. It traverses the vertices of the graph as usual with the exception of the
auxiliary-vertices. The first segment starts from the root auxiliary-vertex. The segment is
completed once a different auxiliary-vertex is reached. When this happens, the algorithm
starts a new segment by jumping to an other auxiliary-vertex and continues traversing the
graph as before. This way each segment starts and ends in a auxiliary-vertex. The startand endpoints of all segments resemble the auxiliary-vertex pairs PA . The best current
result for each possible element in PA is stored and updated if necessary every time a
path segment is completed.
To avoid unnecessary traversal (due to symmetry) of the graph, a path segment is only
allowed to end in a vertex having an id higher than its start vertex. Additionally a path
segment can only start from an auxiliary-vertex, if this vertex is higher than all other starting
vertices in the current search path.
Combining Paths
After having performed preprocessing, we can now find the longest (simple) path of G in an
auxiliary graph that only contains auxiliary-vertices and edges between them. In this graph
two auxiliary-vertices are connected by an edge if they belong to the same graph GA . Note
that every block of the original partition is now represented by a clique of its auxiliary-vertices
in our auxiliary graph. Additionally, the edges {ve , ve0 } get introduced for all edges e that
were replaced in the previous graph, where ve and ve0 are the auxiliary-vertices that replaced
e. These edges represent the connections between the blocks.
In order to solve the longest path problem, we use another modified version of exhaustive
DFS. It starts from the vertex representing the start vertex. The algorithm creates a set
of auxiliary-vertex-pairs BA ∈ PA for every block A from its search path. An edge {v, w}
is part of a block A if both v and w are auxiliary-vertices of A. While this edge is part
of the search path the pair {v, w} is an element of BA . The pair {v, w} ∈ BA represents
a connection of the corresponding auxiliary-vertices of v and w in GA through a simple
path. The simple paths of all these pairs in BA cannot intersect with each other. The best
possibility to do this and the combined length of these paths has already been calculated in
the preprocessing phase. We do the following in order to only receive valid auxiliary-vertex
pairs BA when trying to append new edges to the current search path: first of all, for every
block A a solution for BA has to exist. This can be looked up, since the best possible solutions
have been calculated in the preprocessing step. Second, we have to search the graph in
T. Balyo et. al.
XX:5
Figure 1 An example illustrating the basic idea of the LPDP algorithm. In the upper left corner
we have a graph partitioned into three blocks indicated by the colors green, yellow and blue. The
starting node is in the green block and the target in the blue block. The border nodes are named
xi , yi and zi . In the upper right corner we see the three graphs created by removing the edges
connecting the blocks and adding the auxiliary nodes (indicated as small circles). Finally, in the
bottom we have the graphs used in the second stage of the algorithm – combining paths. On the
left side we see the simple version and right side contains the improved version with fever auxiliary
nodes and vertices.
an alternating pattern between edges that are part of a block and edges that connect two
blocks. Otherwise {a, b}, {b, c} ∈ BA would be possible, which means that the two paths in
GA would intersect. Now every time the vertex representing of the original target-vertex
has been found, the paths in GA for every BA are looked up and their weight summed up.
At the end the different BA with the highest combined weight are returned. This weight is
the weight/length of the longest simple path in G. The actual longest simple path can be
constructed by looking up all the pre-calculated paths in GA for the given connections of its
auxiliary-vertices BA .
Improvement through Hierarchical Partitioning
While preprocessing and computations can be done fast and even in parallel, the auxiliary
graph has to be searched with a variant of the naive brute-force approach. We can avoid this
by applying the same principles that we used to accelerate exhaustive DFS recursively.
The main idea here is to combine groups of blocks and the paths that have been calculated
for each of the blocks into a single new coarser block. After assigning each vertex to its
new/updated block we split the edges that connect different blocks as before. Note that the
cliques representing the previous blocks stay intact. The auxiliary-vertices representing the
start or target vertex is also connected to a new auxiliary-vertex as before. The computation
SEA 2017
XX:6
Optimal Longest Paths by Dynamic Programming
or combination of the paths is again done using the modified exhaustive DFS between the
new block’s auxiliary-vertices. In this way, we are actually doing a recursive call of our
algorithm. In a sense, the partition of the original graph G has to be hierarchical. Blocks
are subsequently combined step by step into larger ones, until only one is left and a longest
path of G is calculated.
Improvement through Reducing the Number of Auxiliary Vertices
Above we defined that every cut edge e = {x, y} is replaced with two auxiliary-vertices
and edges {x, ve } {ve0 , y}. Let Ax := {b1 , b2 , ...} be the set of all auxiliary-vertices that are
connected to a vertex x. If we ignore the possibility of two elements of Ax connecting to
each other (bi − x − bj ), we see that a path segment of the current block connecting b ∈ Ax
is also a valid path if we replace it with any other b0 ∈ Ax . This leads to an improved
algorithm that only has a single auxiliary-vertex bx for every vertex. Hence, in contrast to
the previous formulation, the weight of a cut edge {x, y} does not get divided among the
replacing edges {x, ve }, {ve0 , y} as before. Instead the edge {ve , ve0 } that gets introduced in
the auxiliary graph while combining the paths retains the weight of {x, y}. The weights of
{x, bx }, {by , y} are set to zero. Note that in this variant of the algorithm, bx is the same for
every other edge {x, z} that is replaced, meaning bx is the only auxiliary-vertex of x.
In turn, the auxiliary graph cannot be traversed as before, since all edges between different
blocks have weights associated with them. To obtain the correct result we sum up the weight
of all weighted edges in the search path and add it to the length of the currently induced
path. Additionally a single vertex can be connected to multiple different blocks. To still
search through all possible valid paths we have to allow multiple consecutive edges in the
search path that connects different blocks, compared to the purely alternating pattern from
before. Entering and leaving a vertex x through two such edges corresponds to a (bi − x − bj )
connection mentioned above, which represents the usage of a single vertex x in the block. It
follows that the search path not only induces the set of auxiliary-vertex-pairs BA for a block
A, but also the set of “excluded” vertices CA . Hence, we have to find the best possibility to
connect the pairs in BA without using any vertices in CA . During preprocessing, we calculate
all possibilities for each BA and choose the best one. By looking at the unused vertices of a
current solution, we get the best possibility for each BA and CA . Note that any vertex in
CA would simply correspond to an additional pair {bi , bj } ∈ BA in the previous formulation
of the algorithm. For an example of the improved auxiliary graph see Figure 1.
4
Experimental Evaluation
Methodology
We have implemented the algorithm described above using C++ and compiled it using gcc
4.9.4 with full optimizations turned on (-O3 flag). Our implementation is freely available in
the Karlsruhe Longest Paths package (KaLP) under the GNU GPL v2.0 license [1]. We use
multiple implementations provided by Stein et al. [14] for comparison: Exhaustive DFS is
the naive brute-force approach as well as the A* algorithm and the DFBnB solver. We run
each algorithm and input pair with a time limit of one hour. Experiments were run on a
machine that is equipped with two Intel® Xeon® Processors X5355 (2.66 GHz with 4 cores)
and 24 GB RAM.
We present multiple kinds of data: first and foremost, we use cactus plots in which
the number of problems is plotted against the running time. The plot shows the runtimes
T. Balyo et. al.
XX:7
achieved by the algorithm on each problem. The running times are sorted in ascending order
for each algorithm. The point (x, t) on a curve means that the xth fastest solved problem
was solved in t seconds. Problems that were not solved within the time limit are not shown.
In addition we utilize tables reporting the number of solved problems as well as scatter plots
to compare running times of two different solvers A, B by plotting points (tA , tB ) for each
instance.
Benchmark Problems
We mainly use instances similar to the ones that have been used in previous work by
Stein et al. [14], i.e. based on mazes in grids as well as the road network of New York.
Additionally we use subgraphs of a word association graph [6]. The graph describes the
results of an experiment of free word association performed by more than 6000 participants.
Vertices correspond to words and arcs represent a cue-target pair.
The first set of instances is generated by using mazes
s
in N × N grids of square fields with a given start and
target field. One can move to adjacent fields horizontally
or vertically but only if the field is not an obstacle. The
goal is to find the longest simple path from the start field
to the target field. We represent the grids as graphs: for
every free field we insert a vertex and we add an edge
with weight 1 between any two vertices, whose fields are
horizontally or vertically adjacent to each other. We
generate the grids as Stein et al. [14]: the top left and
bottom right field are the start and target fields. Random
t
fields of the grid are consecutively made into obstacles
Figure 2 An example maze with
until a certain percentage x ∈ {30%, 40%} of all fields is
obstacles and the longest s-t path.
filled. Afterwards a path between the start and target is
searched for to make sure that a solution of the longest path problem exists. An example
maze is shown in Figure 2. The sizes of the used mazes range from 10x10 up to 120x120.
The second and third set of instances are subgraphs of the road network of New York [3]
as well as subgraphs of the word association graph [6], respectively. A subgraph is extracted
as follows: we start a breadth-first search from a random vertex of the network and stop it
when a certain number of vertices is reached. The vertices touched by the breadth-first search
induce the instance. One of the touched vertices is randomly chosen as the target-vertex.
Experimental Results
We now compare A*, DFBnB, and exhDFS presented by Stein et al. [14] to our algorithm
LPDP using two configurations. Our configurations differ in the amount of time that we
spent on partitioning the input instance. We use either the eco/default configuration of
KaFFPa (LPDPe), which is a good trade off between solution quality and running time, or
the strong configuration of KaFFPaE which aims at better partitions while investing more
time for partitioning (LPDPs). In the latter case, we set the amount of block imbalance
to 10%. Note that LPDPs spends much more time in the graph partitioning phase of the
algorithm than LDPDe. All results reporting running time in this paper include the time
spent for partitioning.
Figures 3–5 and Table 1 summarize the results of our experiments. It is apparent from
the cactus plots in Figure 3 that both configurations of LPDP significantly outperform the
SEA 2017
Optimal Longest Paths by Dynamic Programming
10000
1000
Time in seconds
100
10
1
0.1
LPDPe
LPDPs
A*
DFBnB
exhDFS
0.01
0.001
0.0001
0
50
100
150
200
250
300
Random Grid Maze Problems
10000
1000
Time in seconds
100
10
1
0.1
0.01
LPDPe
LPDPs
A*
DFBnB
exhDFS
0.001
0.0001
1e-05
0
20
40
60
80
100
120
140
160
Road Network Problems
10000
1000
100
Time in seconds
XX:8
10
1
0.1
0.01
0.001
LPDPe
LPDPs
A*
DFBnB
exhDFS
0.0001
1e-05
1e-06
0
10
20
30
40
50
60
Word Association Problems
Figure 3 Cactus plots for the three kinds of benchmark problems comparing the previous
algorithms to LPDP which three different partitioning configurations. The running times include
time spent on partitioning for the LPDP variants.
T. Balyo et. al.
XX:9
10000
A* vs. LPDPe
DFBnB vs. LPDPe
exhDFS vs. LPDPe
A* vs. LPDPs
DFBnB vs. LPDPs
exhDFS vs. LPDPs
1000
Speedups
100
10
1
0.1
0.01
0.001
0.0001
0
20
40
60
80
100
120
140
Problems
Figure 4 Speedup of LPDPe and LPDPs in relation to previous LP algorithms on problems that
were solved within the time limit by each of the five tested algorithms.
Time of LPDPs in sec
1000
100
10
1
0.1
0.1
1
10
100
1000
Time of LPDPe in sec
Figure 5 A scatter plot comparing running times of LPDPe and LPDPs on the entire benchmark
set. Points above (below) the green line represent instances where LPDPe (LPDPs) was faster.
Points on the right (top) blue line represent instances that were not solved within the time limit of
one hour by LPDPe (LPDPs).
SEA 2017
XX:10
Optimal Longest Paths by Dynamic Programming
Solver
A*
DFBnB
Exhaustive DFS
LPDPe
LPDPs
Grid Mazes
34
36
31
288
296
Number of Solved Instances
Road Network Word Assoc.
71
39
83
46
71
44
141
55
143
55
Total
144
165
146
484
494
Table 1 The number of instances solved within the time limit of one hour by the tested solver
configurations for each collection of benchmark problems and in total.
previous algorithms for each kind of tested benchmark except for very easy problems. These
problems are typically solved under a few seconds by any of the algorithms. In these cases,
most of the time of our algorithm is spent in the partitioning phase. Moreover, our LPDP
algorithms can solve significantly more problems, which can be seen in the cactus plots as
well as in Table 1.
There are 140 problem instances that were solved by all five solvers within the time limit.
In Figure 4 we provide the speedup of LPDPe and LPDPs against the three original LP
algorithms. For most of these instances the speedup is actually below 1, but from our data
we know, that this happens only for easy problems (solvable within a couple of seconds
by each solver). The slowdown on these easy instances is due to the overhead caused by
partitioning. Nevertheless, the average speedups are still above 1: 17.45, 24.44 and 166.01
for LPDPe vs. A*, DFBnB, and exhDFS respectively and 3.24, 3.83 and 31.22 for LPDPs vs.
A*, DFBnB, and exhDFS respectively.
The differences in running time are highest for the grid maze instances and lowest
for word association graph problems. We believe this is due to the structure of these
graphs, in particular, how well they can be partitioned to loosely connected subgraphs. Our
algorithm excels on problems that can be successfully partitioned but is competitive on
all kinds of graphs.
As of evaluating our algorithm with different partitioning configurations, we see that
spending extra time on partitioning to get better solutions pays off. In particular, LPDPs is
able to solve more instances. Especially if the instance appears to be hard it is worth while
to invest more time in partitioning. Additionally, this depends on how well the graphs can
be partitioned (highest for grid mazes, smallest for word association).
Looking at the scatter plot in Figure 5, we can see that LPDPe is faster for most of
the instances but has significantly more unsolved instances. Nevertheless, there are three
instances that are solved by LDPDe and not by LPDPs. Each of these three instances come
from a different benchmark set. This shows that spending more effort on the partitioning
does not necessarily increase the number of solved instances.
5
Conclusion
We presented an optimal algorithm for the longest path problem in undirected graphs which
is based on dynamic programming. Experiments show that our new algorithm is faster for
nontrivial problems than the previous optimal algorithms and can solve significantly more
benchmark instances if a time limit per instance is given. Important future work includes
parallelization of our algorithm to improve the solver speed even further.
T. Balyo et. al.
XX:11
References
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
T. Balyo, K. Fieger, and C. Schulz. KaLP – Karlsruhe Longest Paths Homepage. http:
//algo2.iti.kit.edu/kalp/index.html.
P. Brucker. Scheduling Algorithms. Springer-Verlag New York, Inc., Secaucus, NJ, USA,
1995.
C. Demetrescu, A. V. Goldberg, and D. S. Johnson. The Shortest Path Problem: Ninth
DIMACS Implementation Challenge, volume 74. American Mathematical Soc., 2009. URL:
http://www.dis.uniroma1.it/challenge9/download.shtml.
K. Fieger. Using Graph Partitioning to Accelerate Longest Path Search. Bachelor’s Thesis,
Karlsruhe Institute of Technology, 2016.
M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of
NP-Completeness. W. H. Freeman & Co., New York, NY, USA, 1979.
University of Milano Laboratory of Web Algorithms. Datasets. URL: http://law.dsi.
unimi.it/datasets.php.
M. M. Ozdal and M. D. F. Wong. A Length-Matching Routing Algorithm for HighPerformance Printed Circuit Boards. IEEE Transactions on Computer-Aided Design of
Integrated Circuits and Systems, 25(12):2784–2794, 2006.
M. M. Ozdal and M. D. F. Wong. Algorithmic study of single-layer bus routing for highspeed boards. IEEE Transactions on Computer-Aided Design of Integrated Circuits and
Systems, 25(3):490–503, 2006.
D. Portugal and R. Rocha. MSP Algorithm: Multi-robot Patrolling Based on Territory
Allocation Using Balanced Graph Partitioning. In Proceedings of the 2010 ACM Symposium
on Applied Computing, SAC ’10, pages 1271–1276, New York, USA, 2010. ACM.
P. Sanders and C. Schulz. Engineering Multilevel Graph Partitioning Algorithms. In
Proceedings of the 19th European Symposium on Algorithms, volume 6942 of LNCS, pages
469–480. Springer, 2011.
P. Sanders and C. Schulz. Distributed Evolutionary Graph Partitioning. In Proceedings of
the 12th Workshop on Algorithm Engineering and Experimentation (ALENEX’12), pages
16–29, 2012.
P. Sanders and C. Schulz. Think Locally, Act Globally: Highly Balanced Graph Partitioning. In Proceedings of the 12th International Symposium on Experimental Algorithms
(SEA’13), volume 7933 of LNCS, pages 164–175. Springer, 2013.
A. J. Soper, C. Walshaw, and M. Cross. A Combined Evolutionary Search and Multilevel
Optimisation Approach to Graph-Partitioning. Journal of Global Optimization, 29(2):225–
241, 2004.
R. Stern, S. Kiesel, R. Puzis, A. Feller, and W. Ruml. Max is more than min: Solving maximization problems with heuristic search. In Proceedings of the Seventh Annual Symposium
on Combinatorial Search (SoCS 2014), 2014.
W. Y. Wong, T. P. Lau, and I. King. Information Retrieval in P2P Networks Using Genetic
Algorithm. In Special Interest Tracks and Posters of the 14th International Conference on
World Wide Web, WWW ’05, pages 922–923, New York, USA, 2005. ACM.
SEA 2017
| 8cs.DS
|
ANALYSIS OF BLOCK-PRECONDITIONERS FOR MODELS OF
COUPLED MAGMA/MANTLE DYNAMICS
arXiv:1311.6372v3 [math.NA] 29 May 2014
SANDER RHEBERGEN∗ , GARTH N. WELLS† , RICHARD F. KATZ‡ , AND ANDREW J.
WATHEN§
Abstract. This article considers the iterative solution of a finite element discretisation of
the magma dynamics equations. In simplified form, the magma dynamics equations share some
features of the Stokes equations. We therefore formulate, analyse and numerically test a Elman,
Silvester and Wathen-type block preconditioner for magma dynamics. We prove analytically and
demonstrate numerically the optimality of the preconditioner. The presented analysis highlights the
dependence of the preconditioner on parameters in the magma dynamics equations that can affect
convergence of iterative linear solvers. The analysis is verified through a range of two- and threedimensional numerical examples on unstructured grids, from simple illustrative problems through to
large problems on subduction zone-like geometries. The computer code to reproduce all numerical
examples is freely available as supporting material.
Key words. Magma dynamics, mantle dynamics, finite element method, preconditioners.
AMS subject classifications. 65F08, 76M10, 86A17, 86-08.
1. Introduction. The mantle of Earth extends from the bottom of the crust to
the top of the iron core, some 3000 km below. Mantle rock, composed of silicate minerals, behaves as an elastic solid on the time scale of seismic waves but over geological
time the mantle convects at high Rayleigh number as a creeping, viscous fluid [31].
This convective flow is the hidden engine for plate tectonics, giving rise to plate boundaries such as mid-ocean ridges (divergent) and subduction zones (convergent). Plate
boundaries host the vast majority of terrestrial volcanism; their volcanoes are fed by
magma extracted from below, where partial melting of mantle rock occurs (typically
at depths less than ∼100 km).
Partially molten regions of the mantle are of interest to geoscientists for their
role in tectonic volcanism and in the chemical evolution of the Earth. The depth
of these regions makes them inaccessible for direct observation, and hence studies of
their dynamics have typically involved numerical simulation. Simulations are often
based on a system of partial differential equations derived by McKenzie [27] and since
elaborated and generalised by other authors [e.g., 10, 33, 34]. The equations describe
two interpenetrating fluids of different density and vastly different viscosity: solid
and molten rock (i.e., mantle and magma). The grains of the rock form a viscously
deformable, permeable matrix through which magma can percolate. This is captured
in the theory by a coupling of the Stokes equations for the mantle with Darcy’s law
for the magma. Although each phase is independently incompressible, the two-phase
mixture allows for divergence or convergence of the solid matrix, locally increasing or
decreasing the volume fraction of magma. This process is modulated by a compaction
∗ Mathematical Institute, University of Oxford, Andrew Wiles Building, Radcliffe Observatory Quarter, Woodstock Road, Oxford OX2 6GG, United Kingdom and Department of
Earth Sciences, University of Oxford, South Parks Road, Oxford OX1 3AN, United Kingdom
([email protected]).
† Department of Engineering, University of Cambridge, Trumpington Street, Cambridge CB2 1PZ,
United Kingdom ([email protected]).
‡ Department of Earth Sciences, University of Oxford, South Parks Road, Oxford OX1 3AN,
United Kingdom ([email protected]).
§ Mathematical Institute, University of Oxford, Andrew Wiles Building, Radcliffe Observatory
Quarter, Woodstock Road, Oxford OX2 6GG, United Kingdom ([email protected]).
1
2
S. RHEBERGEN, G. N. WELLS, R. F. KATZ AND A.J. WATHEN
viscosity, and gives rise to much of the interesting behaviour associated with coupled
magma/mantle dynamics [35, 36, 20, 37].
The governing equations have been solved in a variety of contexts, from idealised
studies of localisation and wave behaviour [e.g. 1, 8] to applied studies of plate-tectonic
boundaries, especially mid-ocean ridges [e.g. 15, 18]. These studies have employed finite volume techniques on regular, Cartesian grids [e.g. 21]. Unlike mid-ocean ridges,
subduction zones have a plate geometry that is awkward for Cartesian grids; it is, however, conveniently meshed with triangles or tetrahedra, which can also focus resolution
where it is most needed [38]. Finite element simulations of pure mantle convection in
subduction zones are common in the literature, but it remains a challenge to model
two-phase, three-dimensional, magma/mantle dynamics of subduction, even though
this is an area of active research [22, 39]. Such models require highly refined computational meshes, resulting in very large systems of algebraic equations. To solve these
systems efficiently, iterative solvers together with effective preconditioning techniques
are necessary. Although the governing equations are similar to those of Stokes flow,
there has been no prior analysis of their discretisation and numerical solution by the
finite element method.
The most computationally expensive step in modelling the partially molten mantle is typically the solution of a Stokes-like problem for the velocity of the solid matrix.
To address this bottleneck in the context of large, unstructured grids for finite element discretisations, we describe, analyse, and test a preconditioner for the algebraic
system resulting from the simplified McKenzie equations. The system of equations
is similar to the Stokes problem, for which the Silvester–Wathen preconditioner [32]
has been proven to be optimal, i.e., the iteration count of the iterative method is
independent of the size of the algebraic system for a variety of discretisations of the
Stokes equations (see also [26]). The key lies in finding a suitable approximation to
the Schur complement of the block matrix resulting from the finite element discretisation. We follow this approach to prove and demonstrate numerically the optimality
of the preconditioner for coupled magma/mantle dynamics problems. The analysis
and numerical examples highlight some issues specific to magma/mantle dynamics
simulations regarding the impact of model parameters on the solver performance. To
the best of our knowledge, together with the work of Katz and Takei [19], we present
the first three dimensional computations of the (simplified) McKenzie equations, and
the first analysis of a preconditioner for this problem.
In this work we incorporate analysis, subduction zone-inspired examples, and software implementation. The analysis is confirmed by numerical examples that range
from illustrative cases to large, representative models of subduction zones solved using
parallel computers. The computer code to reproduce all presented examples is parallelised and is freely available under the Lesser GNU Public License (LGPL) as part of
the supporting material [30]. The proposed preconditioning strategies have been implemented using libraries from the FEniCS Project [2, 24, 25, 28] and PETSc [5, 6, 7].
The FEniCS framework provides a high degree of mathematical abstraction, which
permits the proposed methods to be implemented quickly, compactly and efficiently,
with a close correspondence between the mathematical presentation in this paper and
the computer implementation in the supporting material.
The outline of this article is as follows. In Section 2 we introduce the simplified
McKenzie equations for coupled magma/mantle dynamics, followed by a finite element
method for these equations in Section 3. A preconditioner analysis is conducted in
Section 4 and its construction is discussed in Section 5. Through numerical simulations
ANALYSIS OF PRECONDITIONERS FOR COUPLED MAGMA/MANTLE DYNAMICS
3
in Section 6 we verify the analysis; conclusions are drawn in Section 7.
2. Partially molten magma dynamics. Let Ω ⊂ Rd be a bounded domain
with 2 ≤ d ≤ 3. The McKenzie [27] model on Ω reads
(2.1)
∂t φ − ∇ · (1 − φ)u = 0,
(2.2)
−∇ · 2η(u) + ∇pf = ∇ ζ − 32 η ∇ · u − ρ̄ge3 ,
κ
∇ · u = ∇ · ∇ (pf + ρf gz) ,
(2.3)
µ
where φ is porosity, u is the matrix velocity, (u) = (∇u + (∇u)T )/2 is the strain
rate tensor, κ is permeability, µ is the melt viscosity, η and ζ are the shear and bulk
viscosity of the matrix, respectively, g is the constant acceleration due to gravity, e3
is the unit vector in the z-direction (i.e., e3 = (0, 1) when d = 2 and e3 = (0, 0, 1)
when d = 3), pf is the melt pressure, ρf and ρs are the constant melt and matrix
densities, respectively, and ρ̄ = ρf φ + ρs (1 − φ) is the phase-averaged density. Here
we assume that µ, η and ζ are constants and that κ is a function of φ. The magma
(fluid) velocity uf can be obtained from u, φ and pf through:
κ
∇ (pf + ρf gz) .
(2.4)
uf = u −
φµ
It will be useful to decompose the melt pressure as pf = p − ρs gz, where p is the
dynamic pressure and ρs gz the ‘lithostatic’ pressure. Equations (2.2), (2.3) and (2.4)
may then be written as
−∇ · 2η(u) + ∇p = ∇ ζ − 23 η ∇ · u + g∆ρφe3 ,
(2.5)
κ
(2.6)
∇ · u = ∇ · ∇ (p − ∆ρgz) ,
µ
κ
(2.7)
uf = u −
∇ (p − ∆ρgz) ,
φµ
where ∆ρ = ρs − ρf . Constitutive relations are given by
n
φ
(2.8)
κ = κ0
, ζ = rζ η,
φ0
where φ0 is the characteristic porosity, κ0 the characteristic permeability, n ≥ 1 is a
dimensionless constant and rζ is the ratio between matrix bulk and shear viscosity.
We non-dimensionalise (2.1), (2.5), (2.6) and (2.7) using
(2.9)
u = u0 u0 , x = Hx0 , t = (H/u0 )t0 , κ = κ0 κ0 , p = ∆ρgHp0 ,
where primed variables are non-dimensional, u0 is the velocity scaling, given by
(2.10)
u0 =
∆ρgH 2
,
2η
and H is a length scale. Dropping the prime notation, the McKenzie equations ((2.1),
(2.5) and (2.6)), in non-dimensional form are given by
(2.11)
∂t φ − ∇ · (1 − φ)u = 0,
(2.12)
−∇ · (u) + ∇p = ∇ 12 rζ − 23 ∇ · u + φe3 ,
!
n
2R2
φ
(2.13)
∇·u=
∇·
(∇p − e3 ) ,
rζ + 4/3
φ0
4
S. RHEBERGEN, G. N. WELLS, R. F. KATZ AND A.J. WATHEN
where R = δ/H with δ the compaction length defined as
s
(2.14)
δ=
(rζ + 4/3)κ0 η
,
µ
and (2.7) becomes
(2.15)
uf = u −
2R2 1
rζ + 4/3 φ
φ
φ0
n
(∇p − e3 ) .
When solving the McKenzie model numerically for time-dependent simulations,
(2.11) is usually decoupled from (2.12) and (2.13). Porosity is updated with (2.11)
after which the velocity and pressure are determined by solving (2.12) and (2.13);
iteration can be used to better capture the coupling. The most expensive part of this
procedure is solving (2.12) and (2.13). In this work we study an optimal solver for
equations (2.12) and (2.13) for a given porosity field. We remark that an alternative
to decoupling (2.11) from (2.12) and (2.13) is to use a composable linear solver for
the full system (2.11)-(2.13), see Brown et al. [12]. In this case, our optimal solver
may be used as a preconditioner for part of this composable linear solver.
For the rest of this paper we replace (rζ − 2/3)/2 by a constant α. Furthermore,
we replace
n
φ
R2
(2.16)
α + 1 φ0
by a spatially variable function k(x) (independent of α and φ) and we obtain the
problem
(2.17a)
−∇ · (u) + ∇p = ∇(α∇ · u) + φe3 ,
(2.17b)
∇ · u = ∇ · (k(∇p − e3 )).
For coupled magma/mantle dynamics problems, α may range from −1/3 to approximately 1000. For this reason we will assume in this paper that −1/3 ≤ α ≤ 1000. We
also bound k: 0 ≤ k∗ ≤ k(x) ≤ k ∗ for all x ∈ Ω. In the infinite-dimensional setting,
we note that if k(x) = 0 everywhere in Ω, the compaction stress ∇(α∇ · u) vanishes as
the velocity field is divergence free and (2.17) reduces to the Stokes equations. This
will not generally be the case for a finite element formulation, as will be discussed in
the following section.
On the boundary of the domain, ∂Ω, we impose
(2.18)
u = g,
(2.19)
−k(∇p − e3 ) · n = 0,
where g : ∂Ω → Rd is given boundary data satisfying the compatibility condition
Z
(2.20)
0=
g · n ds.
∂Ω
3. Finite element formulation. In this section we assume, without loss of
generality, homogeneous boundary conditions on u.
ANALYSIS OF PRECONDITIONERS FOR COUPLED MAGMA/MANTLE DYNAMICS
5
Let Th be a triangulation of Ω with associated finite element spaces Xh ⊂
and Mh ⊂ H 1 (Ω) ∩ L20 (Ω). The finite element weak formulation for (2.17)
and (2.18) is given by: find uh , ph ∈ Xh × Mh such that
Z
Z
(3.1)
B(uh ; ph , v; q) =
φe3 · v dx −
ke3 · ∇q dx
∀v, q ∈ Xh × Mh ,
d
H01 (Ω)
Ω
Ω
where
(3.2)
B(u; p, v; q) = a(u, v) + b(p, v) + b(q, u) − c(p, q),
and
Z
(u) : (v) + α(∇ · u)(∇ · v) dx,
Z
b(p, v) = −
p∇ · v dx,
Z Ω
c(p, q) =
k∇p · ∇q dx.
a(u, v) =
Ω
(3.3)
Ω
Proposition 3.1. For α > −1, there exists a cα > 0 such that
d
2
(3.4)
a(v, v) ≥ cα kvk1 ∀ v ∈ H01 (Ω) .
Proof. The proposition follows from
(3.5)
2
k∇ · vk ≤ (v)
2
≤ k∇vk
2
d
∀ v ∈ H01 (Ω) ,
(see Ref. [16, Eq. (3.4)]) and the application of Korn’s inequality.
We will consider finite elements that are inf-sup stable [11] in the degenerate limit
of k = 0, i.e., a(u, v) is coercive (see Proposition 3.1), c(p, p) ≥ 0 ∀p ∈ Mh and for
which there exists a constant c1 > 0 independent of h such that
(3.6)
max
vh ∈Xh
b(qh , vh )
≥ c1 kqh k
k∇vh k
∀qh ∈ Mh .
In particular, we will use Taylor–Hood (P 2 –P 1 ) finite elements on simplices. We note
that while in the infinite-dimensional setting the Stokes equations are recovered from
(2.17) when k = 0, this is not generally the case for the discrete weak formulation
in (3.1) when α 6= 0. Obtaining the Stokes limit in the finite element setting when
α 6= 0 requires the non-trivial property that the divergence of functions in Xh lie in
the pressure space Mh . This is not the case for Taylor–Hood finite elements.
The discrete system (3.1) can be written in block matrix form as
"
#
u
f
A BT
(3.7)
=
,
g
B −Ck p
where u ∈ Rnu and p ∈ N np = {q ∈ Rnp |q 6= 1} are, respectively, the vectors of
the discrete velocity and pressure variables with respect to appropriate bases for Xh
and Mh . The space N np satisfies the zero mean pressure condition.
6
S. RHEBERGEN, G. N. WELLS, R. F. KATZ AND A.J. WATHEN
For later convenience, we define the negative of the ‘pressure’ Schur complement S:
S = BA−1 B T + Ck ,
(3.8)
and the scalar pressure mass matrix Q such that
2
kqh k = hQq, qi,
(3.9)
for qh ∈ Mh and where q ∈ Rnp is the vector of the coefficients associated with the
pressure basis and h·, ·i denotes the standard Euclidean scalar product.
The differences between the matrix formulation of the magma/mantle equations (2.17) and the Stokes equations lie in the matrices A and Ck . In the case
of the magma/mantle dynamics, A includes the discretisation of compaction stresses:
a ‘grad-div’ term weighted by the factor α. Such ‘grad-div’ terms are known to be
problematic in the context of multigrid methods as the modes associated with lowest
eigenvalues are not well represented on a coarse grid [3]. There have been a number of investigations into this issue for H(div) finite element problems, e.g. [4, 23].
The second matrix which differs from the Stokes discretisation is Ck . For sufficiently
large k, this term provides Laplace-type pressure stabilisation for elements that would
otherwise be unstable for the Stokes problem.
4. Optimal block diagonal preconditioners. To model three-dimensional
magma/mantle dynamics of subduction, efficient iterative solvers together with preconditioning techniques are needed to solve the resulting algebraic systems of equations. The goal of this section is to introduce and prove optimality of a class of block
diagonal preconditioners for (3.7).
To prove optimality of a block preconditioner for the McKenzie problem, we first
present a number of supporting results.
Proposition 4.1. The bilinear form c in (3.3) satisfies
2
∀q ∈ M h .
c(q, q) ≥ k∗ k∇qk
(4.1)
Proof. This follows directly from
c(q, q) = k 1/2 ∇q
(4.2)
2
1/2
2
≥ k∗ ∇q
.
Lemma 4.2. For the matrices A, B and Ck given in (3.7), the pressure Schur
complement S in (3.8) and the pressure mass matrix Q in (3.9), for an inf-sup stable
formulation satisfying (3.6), the following bounds hold
(4.3)
0 < cq ≤
hSq, qi
≤ cq ,
h(Q + Ck )q, qi
∀q ∈ N np ,
where cq is given by
(
(4.4)
q
c =
1/(1 − |α|)
1
if − 1/3 ≤ α < 0,
if α ≥ 0,
ANALYSIS OF PRECONDITIONERS FOR COUPLED MAGMA/MANTLE DYNAMICS
7
and cq by
(4.5)
cq = min
!
c21 + cP k∗ (1 + |α|)
, 1 ,
(1 + |α|)(1 + cP k∗ )
where c1 is the inf-sup constant and cP the Poincaré constant.
Proof. Since A is symmetric and positive definite, and from the definition of S
hSq, qi = hA−1 B T q, B T qi + hCk q, qi
(4.6)
= sup
v∈Rnu
hv, B T qi2
+ hCk q, qi,
hAv, vi
for all q ∈ N np . From the definition of matrices A, B, Ck and Q it then follows that
(4.7)
(qh , ∇ · vh )2
hSq, qi = sup
vh ∈Xh
(vh )
2
+ αk∇ · vh k
+ (k∇qh , ∇qh ).
2
Using (3.5) and the Cauchy–Schwarz inequality,
2
(qh , ∇ · vh )2 ≤ kqh k
(4.8)
(vh )
2
.
For −1/3 ≤ α < 0,
(vh )
2
(4.9)
1
(vh )
1+α
1
(vh )
≤
1+α
=
2
2
+ α (vh )
2
2
+ αk∇ · vh k
,
and for α ≥ 0,
(4.10)
(vh )
2
2
≤ (vh )
2
+ αk∇ · vh k .
Hence,
(4.11)
(qh , ∇ · vh )2 ≤ cq kqh k
2
(vh )
2
+ αk∇ · vh k
2
,
where
(
1/(1 − |α|) if − 1/3 ≤ α < 0,
c =
1
if α ≥ 0.
q
(4.12)
Combining (4.7) and (4.11),
2
(4.13) hSq, qi ≤ cq kqh k + (k∇qh , ∇qh ) = cq hQq, qi + hCk q, qi ≤ cq h(Q + Ck )q, qi.
This proves the upper bound in (4.3).
Next we determine the lower bound. Using (3.5) and the inf-sup condition (3.6),
max
vh ∈Xh
(4.14)
(qh , ∇ · vh )2
(vh )
2
2
+ αk∇ · vh k
≥ max
vh ∈Xh
≥
(qh , ∇ · vh )2
(1 + |α|)k∇vh k
c21
2
kqh k ,
1 + |α|
2
8
S. RHEBERGEN, G. N. WELLS, R. F. KATZ AND A.J. WATHEN
which leads to
(4.15)
hSq, qi ≥
c21
hQq, qi + hCk q, qi.
1 + |α|
Using Proposition 4.1 and the Poincaré inequality,
2
hCk q, qi = (1 − ξ)c(qh , qh ) + ξ k 1/2 ∇qh
(4.16)
≥ (1 − ξ)c(qh , qh ) + ξcP k∗ kqh k
2
= (1 − ξ)hCk q, qi + ξcP k∗ hQq, qi,
for any ξ ∈ [0, 1]. Combining (4.15) and (4.16),
!
c21
+ ξcP k∗ hQq, qi + (1 − ξ)hCk q, qi,
(4.17)
hSq, qi ≥
1 + |α|
and setting ξ = (1 − c21 /(1 + |α|))/(1 + cP k∗ ) in the case that c21 /(1 + |α|) ≤ 1, and
otherwise setting ξ = 0,
!
c21 + cP k∗ (1 + |α|)
, 1 h(Q + Ck )q, qi,
(4.18)
hSq, qi ≥ min
(1 + |α|)(1 + cP k∗ )
from which cq is deduced.
For the discretisation of the Stokes equations, it was shown that the pressure
mass-matrix is spectrally equivalent to the Schur complement [32]. This is recovered
from Lemma 4.2 when k = 0 everywhere and α = 0.
Lemma 4.3. For the matrices A, B and Ck in (3.7), S in (3.8) and the pressure
mass matrix Q in (3.9), if the inf-sup condition in (3.6) is satisfied, then
(4.19)
h(B T (Q + Ck )−1 Bv, vi
≤ cq
hAv, vi
∀v ∈ Rnu ,
where cq is the constant from in (4.4).
Proof. From Lemma 4.2, symmetry of A and positive semi-definiteness of C,
q T BA−1 B T + Ck q
q T BA−1 B T q
(4.20)
≤
≤ cq ∀q ∈ N np .
q T (Q + Ck ) q
q T (Q + Ck ) q
Inserting q ← (Q + Ck )1/2 q,
(4.21)
q T (Q + Ck )−1/2 BA−1 B T (Q + Ck )−1/2 q
≤ cq
qT q
∀q ∈ N np .
Defining H = (Q + Ck )−1/2 BA−1 B T (Q + Ck )−1/2 and denoting the maximum eigenvalue of H by λmax and associated eigenvector x, since H is symmetric it follows that
λmax ≥ v T Hv/(v T v) ∀v ∈ Rn and λmax = xT Hx/(xT x). Hence, λmax ≤ cq , and
(4.22)
(Q + Ck )−1/2 BA−1 B T (Q + Ck )−1/2 x = λmax x,
and pre-multiplying both sides by A−1/2 B T (Q + Ck )−1/2 ,
(4.23) A−1/2 B T (Q + Ck )−1/2 (Q + Ck )−1/2 BA−1/2 A−1/2 B T (Q + Ck )−1/2 x
= λmax A−1/2 B T (Q + Ck )−1/2 x.
ANALYSIS OF PRECONDITIONERS FOR COUPLED MAGMA/MANTLE DYNAMICS
9
Letting v = A−1/2 B T (Q + Ck )−1/2 x, the above becomes
(4.24)
A−1/2 B T (Q + Ck )−1 BA−1/2 v = λmax v,
and it follows from λmax ≤ cq that
(4.25)
v T A−1/2 B T (Q + Ck )−1 BA−1/2 v
≤ cq
vT v
∀v ∈ Rnu ,
or, taking v ← A−1/2 v,
(4.26)
v T B T (Q + Ck )−1 Bv
≤ cq
v T Av
∀v ∈ Rnu ,
and the Lemma follows.
We now consider diagonal block preconditioners for (3.7) of the form
P 0
(4.27)
P=
,
P ∈ Rnu ×nu , T ∈ Rnp ×np .
0 T
We assume that P and T are symmetric and positive-definite, and that they satisfy
(4.28) δAP ≤
hAv, vi
≤ δ AP ∀v ∈ Rnu ,
hP v, vi
δQT ≤
h(Q + Ck )q, qi
≤ δ QT ∀q ∈ N np ,
hT q, qi
where δAP , δ AP , δQT and δ QT are independent of h, but may depend on model
parameters.
The discrete system in (3.7) is indefinite, and hence has both positive and negative eigenvalues. The speed of convergence of the MINRES Krylov method for the
preconditioned system
#
−1
−1 "
u
P 0
f
P 0
A BT
=
,
(4.29)
0 T
g
0 T
B −Ck p
depends on how tightly the positive and negative eigenvalues of the generalised eigenvalue problem
#
"
v
P 0 v
A BT
(4.30)
=λ
,
0 T q
B −Ck q
are clustered [13, Section 6.2]. Our aim now is to develop bounds on the eigenvalues
in (4.30) that are independent of the mesh parameter h.
Theorem 4.4. Let cq and cq be the constants in Lemma 4.2, and the matrices
A, B and Ck be those given in (3.7), S be the pressure Schur complement in (3.8)
and Q the pressure mass matrix in (3.9). If P and T satisfy (4.28), all eigenvalues
λ < 0 of (4.30) satisfy
q
q QT
1
2
(4.31)
−c δ
≤ λ ≤ 2 δAP − δAP + 4cq δQT δAP ,
and eigenvalues λ > 0 of (4.30) satisfy
(4.32)
δAP ≤ λ ≤ δ AP + cq δ QT .
10
S. RHEBERGEN, G. N. WELLS, R. F. KATZ AND A.J. WATHEN
Proof. Lemmas 4.2 and 4.3 provide the bounds
(4.33)
cq ≤
hSq, qi
≤ cq ,
h(Q + Ck )q, qi
h(B T (Q + Ck )−1 Bv, vi
≤ cq ,
hAv, vi
for all q ∈ N np and ∀v ∈ Rnu . Using these bounds together with the bounds given
in (4.28), the result follows directly by following the proof of Theorem 6.6 in Elman
et al. [13], or more generally Pestana and Wathen [29].
The main result of this section, Theorem 4.4, states that the eigenvalues of the
generalised eigenvalue problem (4.30) are independent of the problem size. From
Theorem 4.4 we see that
"
# [ h
q
i
1
2
(4.34) λ ∈ −cq δ QT ,
δAP − δAP
δAP , δ AP + cq δ QT ,
+ 4cq δQT δAP
2
in which all constants are independent of the problem size (independent of h). This
tells us that if we can find a P and T that are spectrally equivalent to A and Q +
Ck , respectively, then an iterative method with preconditioner (4.27) will be optimal
for (3.7).
The interval in (4.34) shows the dependence of the eigenvalues on α and k. The
upper and lower bounds on the positive eigenvalues are well behaved, as is the lower
bound on the negative eigenvalues, for all α and k. It is only when cq 1 that the
upper bound on the negative eigenvalues tends to zero. If this is the case, the rate of
convergence of the iterative method may slow. From (4.5), we see that cq 1 only if
α 1 and, at the same time, k∗ 1.
5. Preconditioner construction. Implementation of the proposed preconditioner requires the provision of symmetric, positive definite matrices P and T that
satisfy (4.28). Obvious candidates are P = A and T = Q + Ck , with a direct solver
used to compute the action of P −1 and T −1 . We will use this for small problems
in the following section to study the performance of the block preconditioning; the
application of a direct solver is not practical, however, when P and T are large, in
which case we advocate the use of multigrid approximations of the inverse.
To provide more general guidance, we first reproduce the following Lemma from
Elman et al. [13, Lemma 6.2].
Lemma 5.1. If u is the solution to the system Au = f and
(5.1)
ui+1 = (I − P −1 A)ui + P −1 f,
then if the iteration error satisfies hA(u − ui+1 ), u − ui+1 i ≤ ρhA(u − ui ), u − ui i, with
ρ < 1,
(5.2)
1−ρ≤
hAv, vi
≤1+ρ
hP v, vi
∀v.
Proof. See Elman et al. [13, proof of Lemma 6.2].
Lemma 5.1 implies that a solver that is optimal for Au = f will satisfy (4.28),
and is therefore a candidate for P , and likewise for T . The obvious candidates for P
and T are multigrid preconditioners applied to A and Q + Ck , respectively. However,
as we will show by example in Section 6, as α increases, and therefore the compaction
stresses (a ‘grad-div’ term) become more important, multigrid for P becomes less
effective as a preconditioner. More effective treatment of the large α case is the
subject of ongoing investigations.
ANALYSIS OF PRECONDITIONERS FOR COUPLED MAGMA/MANTLE DYNAMICS
11
6. Numerical simulations. In this section we verify the analysis results through
numerical examples. In all test cases we use P 2 –P 1 Taylor–Hood finite elements on
simplices. The numerical examples deliberately address points of practical interest
such as spatial variations in the parameter k, a wide range of values for α and large
problem sizes on unstructured grids of subduction zone-like geometries.
We consider two preconditioners. For the first, we take P = A and T = Q +
Ck in (4.27) and apply a direct solver to compute the action of the inverses. This
preconditioner will be referred to as the ‘LU’ preconditioner. For the second, we use
P −1 = AAMG and T −1 = (Q + Ck )AMG , where we use (·)AMG to denote the use
of algebraic multigrid to approximate the inverse of (·). This preconditioner will be
referred to as the ‘AMG’ preconditioner. The LU preconditioner is introduced as a
reference preconditioner to which the AMG preconditioner can be compared. The LU
preconditioner is not suitable for large scale problems. Note that we never construct
the inverse of P or T , but that we just use the action of the inverse.
All tests use the MINRES method, and the solver is terminated once a relative true
residual of 10−8 is reached. For multigrid approximations of P −1 , smoothed aggregation algebraic multigrid is used via the library ML [14]. For multigrid approximations
of T −1 , classical algebraic multigrid is used via the library BoomerAMG [17]. Unless
otherwise stated, we use multigrid V-cycles, with two applications of Chebyshev with
Jacobi smoothing on each level (pre and post) in the case of smoothed aggregation,
and symmetric Gauss–Seidel for the classical algebraic multigrid. The computer code
is developed using the finite element library DOLFIN [24], with block preconditioner
support from PETSc [12] to construct the preconditioners. The computer code to
reproduce all examples is freely available in the supporting material [30].
6.1. Verification of optimality. In this test case we verify optimality of the
block preconditioned MINRES scheme by observing the convergence of the solver
for varying h, α, k ∗ and k∗ . We solve (2.17) and (2.18) on the unit square domain
Ω = (0, 1)2 using a regular mesh of triangular cells. For the permeability, we consider
(6.1) k =
k ∗ − k∗
tanh(10x − 5) + tanh(10z − 5)
4 tanh(5)
2(k ∗ − k∗ ) − 2 tanh(5)(k∗ + k ∗ )
+
2
.
+
k∗ − k ∗
We ignore body forces but add a source term f to the right hand side of (2.17a). The
Dirichlet boundary condition g and the source term f are constructed such that the
exact solution pressure p and velocity u are:
(6.2)
(6.3)
(6.4)
p = − cos(4πx) cos(2πz),
ux = k∂x p + sin(πx) sin(2πz) + 2,
1
uz = k∂z p + cos(πx) cos(2πz) + 2.
2
Table 6.1 shows the number of iterations the MINRES method required to converge using the LU and AMG preconditioners with k∗ = 0.5 and k ∗ = 1.5, when
varying α from −1/3 to 1000. We clearly see that the LU preconditioner is optimal
(the iteration count is independent of the problem size), as predicted by the analysis
(see Theorem 4.4). Using the AMG preconditioner, there is a very slight dependence
on the problem size. The results in Table 6.1 indicate that the LU preconditioner
12
S. RHEBERGEN, G. N. WELLS, R. F. KATZ AND A.J. WATHEN
Table 6.1: Number of iterations for the LU and AMG preconditioned MINRES for
the unit square test with different levels of mesh refinement and for different values
of α. The number of degrees-of-freedom is denoted by N . For the α = 1000 case, four
applications of a Chebyshev smoother, with one symmetric Gauss-Seidel iteration for
each application, was used.
N
9,539
37,507
148,739
592,387
α = − 13
LU AMG
9
29
9
33
8
39
8
42
α=0
LU AMG
9
30
9
36
8
40
8
44
α=1
LU AMG
9
35
9
40
9
47
7
52
α = 10
LU AMG
8
67
8
80
7
96
7
106
α = 1000
LU AMG∗
7
202
6
283
6
366
6
432
z
Γ2
Ltx
Lz
Γ1
Lbx
(0, 0)
Γ3
Γ3
x
Fig. 6.1: Description of the wedge geometry for a two-dimensional subduction zone.
is uniform with respect to α. Theorem 4.4 indicates a possible dependence on α
through the constant cq . However, for α sufficiently small or sufficiently large, the
dependence of cq on α becomes negligible, and α has only a small impact on the iteration count. The AMG preconditioner, on the other hand, shows a strong dependence
on α. The issue with the ‘grad-div’ for multigrid solvers was discussed in Section 3,
and is manifest in Table 6.1. It has been observed in tests that the effectiveness of
a multigrid preconditioned solver for the operator A deteriorates with increasing α.
This is manifest in an increasing ρ in (5.2) for increasing α.
Results for the case of large spatial variations in permeability k are presented in
Tables 6.2 and 6.3 for the cases α = 1 and α = 100, respectively. A dependence of
the iteration count on the permeability is observed. The smaller k ∗ , the larger the
iteration counts for both the AMG and the LU preconditioners. We also observe that
for a given k ∗ there is little influence of k∗ on the iteration count. Comparing the
results in Tables 6.2 and 6.3 we see that the LU preconditioner shows no dependence
on α. For the AMG preconditioner the iteration count increases as α increases from
1 to 100.
6.2. A magma dynamics problem in two dimensions. In this test case we
solve (2.17) and (2.18) on a domain Ω, depicted in Figure 6.1, using unstructured
meshes with triangular cells. We take Ltx = 1.5,√Lbx = 0.5 and Lz = 1. We set the
permeability as k = 0.9(1 + tanh(−2r)) with r = x2 + z 2 and the porosity φ = 0.01.
ANALYSIS OF PRECONDITIONERS FOR COUPLED MAGMA/MANTLE DYNAMICS
13
Table 6.2: Number of iterations to reach a relative tolerance of 10−8 using preconditioned MINRES for the unit square test with varying levels of mesh refinement and
varying (k∗ , k ∗ ) pairs for α = 1. The number of degrees of freedom is denoted by N .
k ∗ = 10−4
N
9,539
37,507
148,739
592,387
k∗ = 0
LU AMG
32
88
35
108
38
130
36
143
k∗ = 10−8
LU AMG
32
88
35
108
37
130
36
143
k∗ = 10−6
LU AMG
32
88
35
108
38
127
35
135
k∗ = 5 · 10−5
LU AMG
32
80
35
97
33
111
33
122
k∗ = 1
N
9,539
37,507
148,739
592,387
k∗ = 0
LU AMG
27
67
28
78
28
93
27
101
k∗ = 0.1
LU AMG
10
37
10
44
10
50
10
54
k∗ = 0.5
LU AMG
9
36
9
42
9
48
9
52
k∗ = 0.9
LU AMG
9
36
9
42
7
47
7
52
k ∗ = 1000
N
9,539
37,507
148,739
592,387
k∗ = 0
LU AMG
3
24
3
27
3
34
3
37
k∗ = 1
LU AMG
3
26
3
27
3
33
3
37
k∗ = 10
LU AMG
3
24
3
27
3
34
3
37
k∗ = 100
LU AMG
3
24
3
30
3
33
3
40
k ∗ = 108
N
9,539
37,507
148,739
592,387
k∗ = 0
LU AMG
1
15
2
18
2
21
2
21
k∗ = 1
LU AMG
1
15
2
18
2
21
2
21
k∗ = 103
LU AMG
1
15
2
18
2
21
2
21
k∗ = 106
LU AMG
1
15
2
18
2
21
2
21
We consider two test cases for this geometry. The first test problem we denote as
the analytical corner flow test problem and the second as the traction-free test prob√
lem. In both problems we prescribe the following conditions: u = uslab = (1, −1)/ 2
on Γ1 , u = 0 on Γ2 and −k (∇p − e3 ) · n = 0 on ∂Ω.
6.2.1. Analytic corner flow. For the analytical corner flow problem we prescribe u = ucorner = (ux , uz ) on Γ3 , which is the analytic expression for corner-flow [9,
Section 4.8]. The corner-flow velocity components ux and uz are given by
(6.5)
ux = cos(θ)ur + sin(θ)uθ ,
uz = − sin(θ)ur + cos(θ)uθ ,
where θ = − arctan(z̃/x), z̃ = z − 1 and
(6.6) ur = Cθ sin(θ) + D(sin(θ) + θ cos(θ)),
uθ = C(sin(θ) − θ cos(θ)) + Dθ sin(θ),
with
(6.7)
C=
β sin(β)
,
− sin2 (β)
β2
D=
β cos(β) − sin(β)
.
β 2 − sin2 (β)
14
S. RHEBERGEN, G. N. WELLS, R. F. KATZ AND A.J. WATHEN
Table 6.3: Number of iterations to reach a relative tolerance of 10−8 using preconditioned MINRES for the unit square test with varying levels of mesh refinement and
varying (k∗ , k ∗ ) pairs for α = 100. The number of degrees of freedom is denoted by N .
=0
AMG
1605
1922
2179
2356
k∗ = 10−8
LU AMG
67
1598
75
1922
76
2177
73
2356
k∗ = 10−6
LU AMG
66
1557
71
1909
72
2146
68
2311
k∗ = 5 · 10−5
LU AMG
58
1385
62
1730
59
1972
59
2156
k∗ = 1
N
9,539
37,507
148,739
592,387
k∗ = 0
LU AMG
28
350
28
445
28
545
28
597
k∗ = 0.1
LU AMG
9
179
9
212
9
247
9
271
k∗ = 0.5
LU AMG
8
171
8
205
8
236
8
265
k∗ = 0.9
LU AMG
7
169
8
202
8
234
8
265
k ∗ = 1000
N
9,539
37,507
148,739
592,387
k∗ = 0
LU AMG
3
75
3
94
3
116
3
139
k∗ = 1
LU AMG
3
75
3
94
3
116
3
139
k∗ = 10
LU AMG
3
75
3
94
3
116
3
139
k∗ = 100
LU AMG
3
75
3
94
3
116
3
139
k ∗ = 108
N
9,539
37,507
148,739
592,387
k∗ = 0
LU AMG
1
11
1
13
1
20
1
23
k∗ = 1
LU AMG
1
11
1
13
1
20
1
23
k∗ = 103
LU AMG
1
11
1
13
1
20
1
23
k∗ = 106
LU AMG
1
11
1
13
1
20
1
23
k ∗ = 10−4
N
9,539
37,507
148,739
592,387
k∗
LU
67
75
76
73
Here β = π/4 is the angle between Γ1 and Γ2 . In Figure 6.2 we show the computed
streamlines of the magma and matrix velocity fields for this problem.
Table 6.4 presents the number of solver iterations for the LU and AMG preconditioners for different values of α. We observe very similar behaviour as we saw for the
test in Section 6.1. The LU preconditioner is optimal and uniform. The AMG preconditioner again shows slight dependence on the problem size, and as α is increased
the iteration count grows.
6.2.2. Traction-free problem. For the traction-free problem, instead of prescribing
ucorner , we prescribe the zero-traction boundary condition, ((u) − pI + α∇ ·
uI · n = 0 on Γ3 . Figure 6.3 shows the computed streamlines of the magma and
matrix velocity fields for this problem.
The solver iteration counts for this problem with different levels of mesh refinement and for different values of α are presented in Table 6.5. As for the analytic
corner flow problem of Section 6.2.1, the LU-based preconditioner is optimal and uniform. As expected, using the AMG-based preconditioner, the solver is not uniform
ANALYSIS OF PRECONDITIONERS FOR COUPLED MAGMA/MANTLE DYNAMICS
(a) α = 1
15
(b) α = 1000
Fig. 6.2: Streamlines of the magma (light) and matrix (dark) velocity fields in the
wedge of a two-dimensional subduction zone using the corner flow boundary condition
on Γ3 . The solution was computed on a mesh with 116176 elements.
Table 6.4: Number of iterations required for the corner flow problem using LU and
AMG preconditioned MINRES for different levels of mesh refinement and varying α.
For the α = 1000 case, four applications of a Chebyshev smoother, with one symmetric
Gauss-Seidel iteration for each application, was used.
N
34,138
133,777
526,719
α=1
LU AMG
26
69
26
75
24
81
α = 10
LU AMG
30
140
29
151
29
171
α = 100
LU AMG
30
367
27
390
26
446
α = 1000
LU AMG∗
28
572
27
669
27
758
with respect to α.
6.3. Magma dynamics problem in three dimensions. In the final case we
test the solver for a three-dimensional problem that is geometrically representative of a
subduction zone. We solve (2.17) and (2.18) on the domain Ω depicted in Figure 6.4.
We set Ltx = 1.5, Lbx = 0.5, Ly = 1 and Lz = 1, and use unstructured meshes of
tetrahedral
cells. Again we set the permeability as k = 0.9(1 + tanh(−2r)), with
√
r = x2 + z 2 , and the porosity φ = 0.01.
√
As boundary conditions, we prescribe u = uslab = (1, 0.1, −1)/ 2 on Γ1 , u = 0
on Γ2 , (u) − pI + α∇ · uI · n = 0 on Γ3 and −k (∇p − e3 ) · n = 0 on ∂Ω. In
Figure 6.5 we show computed vector plots of the matrix and magma velocities for
α = 1 and α = 1000.
Table 6.6 shows the number of iterations needed for the AMG preconditioned
MINRES method for the three-dimensional wedge problem. The LU preconditioned
solver is not practical for this problem when using reasonable mesh resolutions. All
cases have been computed in parallel using 16 processes. The computed examples
span a range of problem sizes, and only relatively small changes in the iteration count
are observed for changes in the number of degrees of freedom. Again, as α becomes
larger, so too does the iteration count.
7. Conclusions. In this work we introduced and analysed an optimal preconditioner for a finite element discretisation of the simplified McKenzie equations for
magma/mantle dynamics. Analysis of the preconditioner showed that the Schur com-
16
S. RHEBERGEN, G. N. WELLS, R. F. KATZ AND A.J. WATHEN
(a) α = 1
(b) α = 1000
Fig. 6.3: Streamlines of the magma (light) and matrix (dark) velocity fields in the
wedge of a 2D subduction zone using no stress boundary conditions on Γ3 . The
solution was computed on a mesh with 116176 elements.
Table 6.5: Number of iterations to reach a relative tolerance of 10−8 using LU and
AMG preconditioned MINRES for different values of α for the no-stress test. For
the α = 1000 case, four applications of a Chebyshev smoother, with one symmetric
Gauss-Seidel iteration for each application, was used.
N
34,138
133,777
526,719
α=1
LU AMG
24
65
23
73
23
80
α = 10
LU AMG
29
143
27
159
26
175
α = 100
LU AMG
27
375
27
424
27
475
α = 1000
LU AMG∗
25
626
24
718
24
798
plement of the block matrix arising from the finite element discretisation of the simplified McKenzie equations may be approximated by a pressure mass matrix plus a
permeability matrix. The analysis was verified through numerical simulations on a
unit square and two- and three-dimensional wedge flow problems inspired by subduction zones. For all computations we used P 2 –P 1 Taylor–Hood finite elements as they
are inf-sup stable in the degenerate limit of vanishing permeability. Numerical tests
demonstrated optimality of the solver. We observed that the multigrid version of the
preconditioner was not uniform with respect to the bulk-to-shear-viscosity ratio α.
As α is increased, the iteration count for the solver increases. We observe a similar
behaviour as k ∗ increases.
The analysis and testing of an optimal block preconditioning method for magma/mantle
dynamics presented in this work lays a basis for creating efficient and optimal simulation tools that will ultimately be put to use to study the genesis and transport
of magma in plate-tectonic subduction zones. Optimality has been demonstrated,
but some open questions remain regarding uniformity with respect to some model
parameters.
Acknowledgements. We thank L. Alisic and J. F. Rudge for the many discussions held related to this paper. We also thank the reviewers M. Knepley, M. Spiegelman, C. Wilson and one that remained anonymous, whose comments helped improve
this paper. The authors acknowledge the support of the Natural Environment Research Council under grants NE/I026995/1 and NE/I023929/1. Katz is furthermore
ANALYSIS OF PRECONDITIONERS FOR COUPLED MAGMA/MANTLE DYNAMICS
17
Ltx
z
Γ2
Γ3
Γ3
Lz
Γ3
y
Γ1
Γ3
(0, 0, 0)
Lbx
Ly
x
Fig. 6.4: Description of the wedge in a three-dimensional subduction zone.
(a) Matrix velocity, α = 1.
(b) Magma velocity, α = 1.
(c) Matrix velocity, α = 1000.
(d) Magma velocity, α = 1000.
Fig. 6.5: Vector plots of the magma and matrix velocities in the wedge of a threedimensional subduction zone for α = 1 and α = 1000 using the stress-free boundary
conditions on Γ3 .
grateful for the support of the Leverhulme Trust.
References.
[1] E. Aharonov, J. A. Whitehead, P. B. Kelemen, and M. Spiegelman. Channeling
instability of upwelling melt in the mantle. J. Geophys. Res., 100:20433–20450,
1995.
[2] M. S. Alnæs, A. Logg, K. B. Ølgaard, M. E. Rognes, and G. N. Wells. Unified
Form Language: A domain-specific language for weak formulations of partial
differential equations. ACM Trans Math Software, 40(2):9:1–9:37, 2014.
18
S. RHEBERGEN, G. N. WELLS, R. F. KATZ AND A.J. WATHEN
Table 6.6: Number of iterations required for AMG preconditioned MINRES for the
three-dimensional subduction model for different levels of mesh refinement and and
different values of α. The number of degrees of freedom is denoted by N . For the
α = 1000 case, four applications of a Chebyshev smoother, with one symmetric GaussSeidel iteration for each application, was used. All tests were run using 16 MPI
processes.
N
88,500
400,690
1,821,991
8,124,691
α=1
42
44
43
41
α = 10
127
122
122
120
α = 100
363
355
367
355
α = 1000
654
692
732
775
[3] D. N. Arnold, R. S. Falk, and R. Winther. Preconditioning in H(div) and applications. Math. Comp., 66(218):957–984, 1997.
[4] D. N. Arnold, R. S. Falk, and R. Winther. Multigrid in H(div) and H(curl).
Numer. Math., 85:197–217, 2000.
[5] S. Balay, W. D. Gropp, L. C. McInnes, and B. F. Smith. Efficient management
of parallelism in object oriented numerical software libraries. In E. Arge, A. M.
Bruaset, and H. P. Langtangen, editors, Modern Software Tools in Scientific
Computing, pages 163–202. Birkhäuser Press, 1997.
[6] S. Balay, J. Brown, K. Buschelman, V. Eijkhout, W. D. Gropp, D. Kaushik,
M. G. Knepley, L. C. McInnes, B. F. Smith, and H. Zhang. PETSc users manual.
Technical Report ANL-95/11 - Revision 3.4, Argonne National Laboratory, 2013.
[7] S. Balay, J. Brown, K. Buschelman, W. D. Gropp, D. Kaushik, M. G. Knepley,
L. C. McInnes, B. F. Smith, and H. Zhang. PETSc Web page, 2014. URL
http://www.mcs.anl.gov/petsc.
[8] V. Barcilon and F. Richter. Nonlinear waves in compacting media. J. Fluid
Mech., 164:429–448, 1986.
[9] G. K. Batchelor. An Introduction to Fluid Dynamics. Cambridge University
Press, New York, 1967.
[10] D. Bercovici and Y. Ricard. Energetics of a two-phase model of lithospheric
damage, shear localization and plate-boundary formation. Geophys. J. Int., 152:
581–596, 2003.
[11] F. Brezzi and M. Fortin. Mixed and Hybrid Finite Element Methods. SpringerVerlag, New York, 1991.
[12] J. Brown, M. G. Knepley, D. A. May, L. C. McInnes, and B. Smith. Composable
linear solvers for multiphysics. In 2012 11th International Symposium on Parallel
and Distributed Computing (ISPDC), pages 55–62, 2012. URL http://dx.doi.
org/10.1109/ISPDC.2012.16.
[13] H. C. Elman, D. J. Silvester, and A. J. Wathen. Finite Elements and Fast
Iterative Solvers. Numerical Mathematics and Scientific Computation. Oxford
University Press, 2005.
[14] M. W. Gee, C. M. Siefert, J. J. Hu, R. S. Tuminaro, and M. G. Sala. ML 5.0
smoothed aggregation users guide. Sandia National Laboratories, 2006. Tech.
Rep. SAND2006-2649.
[15] A. Ghods and J. Arkani-Hamed. Melt migration beneath mid-ocean ridges. Geophys. J. Int., 140:687–697, 2000.
ANALYSIS OF PRECONDITIONERS FOR COUPLED MAGMA/MANTLE DYNAMICS
19
[16] P. P. Grinevich and M. A. Olshanskii. An iterative method for the Stokes-type
problem with variable viscosity. SIAM J. Sci. Comput., 31:3939–3978, 2009.
[17] V. E. Henson and U. M. Yang. BoomerAMG: A parallel algebraic multigrid
solver and preconditioner. Appl. Numer. Math., 41(1):155–177, 2002.
[18] R. F. Katz. Magma dynamics with the enthalpy method: Benchmark solutions
and magmatic focusing at mid-ocean ridges. J. Petrol., 2008.
[19] R. F. Katz and Y. Takei. Consequences of viscous anisotropy in a deforming,
two-phase aggregate: Part 2. Numerical solutions of the full equations. J. Fluid
Mech., 734:456–485, 2013.
[20] R. F. Katz, M. Spiegelman, and B. Holtzman. The dynamics of melt and shear
localization in partially molten aggregates. Nature, 442:676–679, 2006.
[21] R. F. Katz, M. G. Knepley, B. Smith, M. Spiegelman, and E. T. Coon. Numerical simulation of geodynamic processes with the portable extensible toolkit for
scientific computation. Phys. Earth Planet. In., 163:52–68, 2007.
[22] T. Keller, D. A. May, and B. J. P. Kaus. Numerical modelling of magma dynamics
coupled to tectonic deformation of lithosphere and crust. Geophys. J. Int., 195:
1406–1442, 2013.
[23] T. V. Kolev and P. S. Vassilevski. Parallel auxiliary space amg solver for H(div)
problems. SIAM J. Sci. Comput., 34(6):A3079–A3098, 2012.
[24] A. Logg and G. N. Wells. DOLFIN: Automated finite element computing. ACM
Trans. Math. Software, 37(2):20:1–20:28, 2010.
[25] A. Logg, K.-A. Mardal, and G. N. Wells, editors. Automated Solution of Differential Equations by the Finite Element Method, volume 84 of Lecture Notes in
Computational Science and Engineering. Springer, 2012.
[26] D. A. May and L. Moresi. Preconditioned iterative methods for stokes flow
problems arising in computational geodynamics. Phys. Earth Planet. In., 171:
33–47, 2008.
[27] D. McKenzie. The generation and compaction of partially molten rock. J. Petrol.,
25:713–765, 1984.
[28] K. B. Ølgaard and G. N. Wells. Optimisations for quadrature representations
of finite element tensors through automated code generation. ACM Trans Math
Software, 37(1):8:1–8:23, 2010.
[29] J. Pestana and A. J. Wathen. Natural preconditioners for saddle point systems.
2013. URL http://eprints.maths.ox.ac.uk/1754.
[30] S. Rhebergen, R. F. Katz, G. N. Wells, and A. Wathen. Supporting material,
2013. URL https://bitbucket.org/magma-dynamics/preconditioning.
[31] G. Schubert, D. Turcotte, and P. Olson. Mantle Convection in the Earth and
Planets. Cambridge University Press, 2001.
[32] D. Silvester and A. Wathen. Fast iterative solution of stabilised Stokes systems.
Part II: Using general block preconditioners. SIAM J. Numer. Anal., 31:1352–
1367, 1994.
[33] G. Simpson, M. Spiegelman, and M. I. Weinstein. A multiscale model of partial
melts: 1. Effective equations. J. Geophys. Res., 115, 2010.
[34] G. Simpson, M. Spiegelman, and M. I. Weinstein. A multiscale model of partial
melts: 2. Numerical results. J. Geophys. Res., 115, 2010.
[35] M. Spiegelman. Flow in deformable porous-media. Part 1. Simple analysis. J.
Fluid Mech., 247:17–28, 1993.
[36] M. Spiegelman. Flow in deformable porous media. Part 2. Numerical analysis—
The relationship between shock waves and solitary waves. J. Fluid Mech., 247:
20
S. RHEBERGEN, G. N. WELLS, R. F. KATZ AND A.J. WATHEN
39–63, 1993.
[37] Y. Takei and R. F. Katz. Consequences of viscous anisotropy in a deforming,
two-phase aggregate: Part 1. Governing equations and linearised analysis. J.
Fluid Mech., 734:424–455, 2013.
[38] P. E. van Keken, C. Currie, S. D. King, M. D. Behn, A. Cagnioncle, J. He, R. F.
Katz, S.-C. Lin, E. M. Parmentier, M. Spiegelman, and K. Wang. A community
benchmark for subduction zone modeling. Phys. Earth Planet. In., 171:187–197,
2008.
[39] C. R. Wilson, M. Spiegelman, and P. E. van Keken. TerraFERMA: the Transparent Finite Element Rapid Model Assembler for multi-physics problems in
Earth sciences. 2014. URL https://bytebucket.org/tferma/tferma/wiki/
TFpaperGcubed.pdf.
| 5cs.CE
|
Bounded Degree Cosystolic Expanders of Every Dimension
Shai Evra
∗
Tali Kaufman
†
arXiv:1510.00839v3 [math.CO] 26 Jan 2017
January 27, 2017
Abstract
In this work we present a new local to global criterion for proving a form of high dimensional expansion, which we term cosystolic expansion. Applying this criterion on Ramanujan
complexes, yields for every dimension, an infinite family of bounded degree complexes with
the topological overlapping property. This answer affirmatively an open question raised by
Gromov.
1
Introduction
Expander graphs have been central objects of study in both computer science and mathematics
during the past few decades. Informally, expanders are sparse and highly connected graphs
and as such have numerous applications, see [HLW] and [Lub1] and the references therein. In
recent years, a new theory of high dimensional expanders has emerged, pioneered by the works
of Linial-Meshulam-Wallach [LinM],[MW], and Gromov [Gro] (for a recent survey see [Lub2]).
Linial-Meshulam-Wallach and Gromov suggested two, essentially equivalent, generalizations of
the notion of graph expansion to higher dimensions (even though both works were in completely
different research directions).
1.1
Gromov’s question
In [Gro], Gromov studied the complexity of embedding simplicial complexes into Euclidean
spaces. More specifically, he considered the following topological overlapping property:
Definition 1.1. A d-dimensional simplicial complex X has the c-topological overlapping property, for some c > 0, if for every continuous map f : X → d , there exists a point p ∈ d ,
whose preimage, f −1 (p), intersects at least c-fraction of the d-simplices of X, i.e.,
R
|{σ ∈ X(d) : p ∈ f (σ)}| ≥ c · |X(d)|.
R
(1.1)
A family of d-dimensional simplicial complexes is said to have the topological overlapping property, if there exists c > 0, such that each member of the family has the c-topological overlapping
property.
Gromov then proved the remarkable result that, for a fixed d ∈ N, the family of complete ddimensional complexes have the topological overlapping property. This result is a very striking
generalization of classical results from convex combinatorics due to Boros-Furedi [BF] (for d = 2)
∗
Hebrew University, ISRAEL. Email: [email protected]. Research supported in part by the ERC.
Bar-Ilan University, ISRAEL. Email: [email protected]. Research supported in part by the IRG, ERC and
BSF.
†
1
and Barany [Bar] (for d ≥ 3). Gromov then proceeded to give other examples of families
of complexes which posses the topological overlapping property (e.g. spherical buildings and
random complexes). However, all the examples he provided were of unbounded degree, i.e. the
number of faces incident to a single vertex grows with the number of vertices in the complex.
This naturally led Gromov to the following question:
Question 1.2 (Gromov). Is there an infinite family of bounded degree d-dimensional simplicial
complexes with the topological overlapping property, for arbitrary d ≥ 2?
In a recent work [KKL], Kaufman, Kazhdan and Lubotzky were able to give an affirmative
answer to Gromov’s question in dimension two, i.e., they showed that there exists an infinite
family of 2-dimensional, bounded degree complexes with the topological overlapping property,
see [KKL, Theorem 1.3]. However, their proof method is inherently suited for dimension two.
In this work we give a complete answer to Gromov’s question.
Theorem 1.3. For every d ∈ N, there exist an infinite family of d-dimensional bounded degree
complexes with the topological overlapping property.
An immediate application of Theorem 1.3, is an improvement of a result by Gromov and
Guth [GG, Theorem 2.2], which gives a high dimensional generalization to a classical result of
Kolmogorov and Barzdin [BK].
Corollary 1.4. For every D ≥ 2d + 1, there exists C = C(D) > 0 and an infinite family of ddimensional bounded degree complexes, {Xn }n , which satisfies the following: For any embedding
of Xn into RD , such that the images of any two non-adjoint simplices of Xn are of distance at
1
least 1, the volume of the 1-neighborhood of the image of Xn is at least C · |Xn | D−d .
1.2
A criterion for the topological overlapping property
Let us start by addressing the question: How can one prove that a certain complex posses a
topological overlapping property? In [Gro], Gromov showed that the topological overlapping
property is implied by a certain notion of high-dimensional expansion, which we now wish to
define. To do so, we first need some notations.
Definition 1.5. Let X be a d-dimensional simplicial complex. For any −1 ≤ k ≤ d, define:
• X(k) = {σ ∈ X | |σ| = k + 1} - the collection of k-dimensional faces of X.
• C k = C k (X; F2 ) = {f : X(k) → F2 } - the space of k-dimensional F2 -cochains of X.
P
• δk : C k → C k+1 , δf (τ ) = σ⊂τ f (σ) - the k-dimensional F2 -coboundary map of X.
• Z k = Z k (X; F2 ) = ker(δk ) - the space of k-dimensional F2 -cocycles of X.
• B k = B k (X; F2 ) = Im(δk−1 ) - the space of k-dimensional F2 -coboundaries of X.
P
| σ⊂τ }|
• k · k : C k → [0, 1], kf k = σ∈X(k) |{τ ∈X(d)
· f (σ) - a k-dimensional norm on C k .
d+1
(k+1
)·|X(d)|
We are now in a position to present our first notion of high dimensional expansion.
Definition 1.6 (Coboundary expanders). Let X be a d-dimensional simplicial complex, and
let ǫ > 0. We say that X is an ǫ-coboundary expander, if for every k = 0, 1, . . . , d − 1,
kδ(f )k
k
k
k
|f ∈C \B
≥ ǫ.
(1.2)
Expb (X) := min
minb∈B k kf + bk
2
The notion of coboundary expansion first appeared implicitly in the work [LinM], whose motivation was to generalize to higher dimensional complexes, the phase transition phenomenon of
connectivity of random graphs in the Erdos-Reyni model. Later on, Gromov [Gro], came across
essentially the same notion of expansion, while studying fiberwise complexetiy, and showing
that coboundary expansion implies the topological overlapping property.
Unfortunately, up to this date, no bounded degree coboundary expanders are known to exists, so one cannot use Gromov’s result, to answer his question. The problem is that coboundary
expansion requires vanishing of F2 -cohomology of the complex, a property which is so strong
that even certain Ramanujan complexes are known to not satisfy it (see [KKL]).
In order to bypass this problem, we define a weaker notion of expansion, which we call
cosystolic expansion. This notion is strictly weaker than coboundary expansion, since it allows
the existence of cocycles which are not coboundaries, as long as they are sufficiently large.
Definition 1.7 (Cosystolic expanders). Let X be a d-dimensional simplicial complex, and let
ǫ, µ > 0. We say that X is an (ǫ, µ)-cosystolic expander, if for every k = 0, 1, . . . , d − 1,
kδ(f )k
| f ∈ Ck \ Zk ≥ ǫ
(1.3)
Expkz (X) := min
minz∈Z k kf + zk
and
Systk (X) := min{kzk | z ∈ Z k \ B k } ≥ µ.
(1.4)
In [DKW], the authors strengthen the aforementioned result of Gromov, proving that cosystolic expansion implies the topological overlapping property.
Theorem 1.8 (Gromov’s Topological Overlapping Criterion [DKW]). For every d ∈ N and
ǫ, µ > 0, there exist c = c(d, ǫ, µ) > 0, such that if X is a d-dimensional (ǫ, µ)-cosystolic
expander, then X has the c-topological overlapping property.
Hence, by Gromov’s topological overlapping criterion, in order to prove Theorem 1.3, one
needs to prove the existence of an infinite family of bounded degree cosystolic expanders.
1.3
A criterion for cosystolic expansion
The first infinite family of bounded degree cosystolic expanders was constructed in [KKL],
where it was shown that the 2-dimensional skeletons (see below) of 3-dimensional Ramanujan
complexes of sufficiently large degree are cosystolic expanders.
In this work our intention was to generalize the results of [KKL] to all dimensions. However,
the proof method of [KKL] is specifically designed for the two-dimensional case. Indeed, in order
to apply their proof to dimension three or higher, one will need to assume that the graph, whose
vertex set is the set of edges in the complex and its edge set is the set of pairs of edges which
form a triangle in the complex, is an excellent expander graph. As it turns out, this condition
is too strong and does not hold in any Ramanujan complex.
The main novelty of our work is to present a new method for transforming expansion in
lower dimension to high dimension in the complex. This allow us to prove our main result,
which is a local to global criterion for cosystolic expansion, in all dimensions. To state this
criterion, we need to define the notions of skeletons, links and skeleton expander of a complex.
Definition 1.9. Let X be a d-dimensional simplicial complex.
• For k ≤ d, the k-dimensional skeleton of X, denoted X (k) , is the complex obtained by
deleting from X all faces of dimension greater then k.
3
• For σ ∈ X, the link of the face σ in X, denoted Xσ , is the complex obtained by picking
only the faces in X that contains σ, and removing σ from each of these faces. A link is
called a proper link of X, if 1 ≤ |σ| ≤ d − 1.
• A complex X, is said to be an α-skeleton expander, for some α > 0, if its 1-skeleton satisfy
the following graph expansion property 1 :
∀ A ⊂ X(0),
kE(A, A)k ≤ 4 · (kAk2 + α · kAk),
(1.5)
where E(A, A) ⊂ X(1) are the edges in X with both vertices in A.
We are now ready to state our local to global criterion for cosystolic expansion.
Theorem 1.10 (Main Theorem). For any d, Q ∈ N and β > 0, there exists ǫ = ǫ(d, β, Q) > 0,
µ = µ(d, β) > 0 and α = α(d, β) > 0, such that if X is a d-dimensional complex satisfying:
• X is Q-bounded degree, i.e. maxv∈X(0) |Xv | ≤ Q.
• All the proper links of X are β-coboundry expanders.
• All the proper links of X and X itself are α-skeleton expanders.
Then the (d − 1)-skeleton of X is an (ǫ, µ)-cosystolic expander.
It seems very natural to compare our local to global expansion result, with that of Garland
[Gar]. However, other then philosophically, the two results are unrelated, Garland’s work is
over R, while ours is over F2 . Moreover, the proof method of both results is quite different.
1.4
Ramanujan complexes
The well known Ramanujan complexes (see [Lub2]), constructed over a fixed residue field, are
bounded degree high dimensional complexes, which are excellent skeleton expanders (this follows
directly from the Ramanujan property). Their proper links are spherical buildings admitting a
strongly transitive action, and it follows from the work of Gromov [Gro] (see also [LMM]) that
they are coboundary expanders. Therefore, we are left to prove that their links, i.e., that the
spherical buildings, are also good skeleton expanders 2 .
Theorem 1.11. For any d ∈ N and α > 0, there exist q0 = q0 (d, α) > 0, such that for any
prime power q ≥ q0 , all d-dimensional, q-thick spherical buildings admitting a strongly transitive
action are α-skeleton expanders.
In conclusion, Ramanujan complexes of sufficiently large but fixed degree satisfies the requirements of Theorem 1.10, which together with Gromov’s topological overlapping criterion
(Theorem 1.8), implies the following corollary regarding the expansion of Ramanujan complexes.
Corollary 1.12. For any d ∈ N, there exists q0 = q0 (d), such that for any prime power q ≥ q0 ,
there exists µ = µ(d) > 0, ǫ = ǫ(d, q) > 0 and c = c(d, q) > 0, such that if X is the d-dimensional
skeleton of a (d + 1)-dimensional q-thick Ramanujan complex, then
1
The constant 4 in our definition of skeleton expansion is unnatural, as it implies a notion strictly weaker then
graph expansion. See [KM1] for a more accurate notion of skeleton expansion. However, our notion of skeleton
expansion will be sufficient for our use here.
2
After a completion of our paper, Izhar Oppenheim has pointed to us that a similar result to Theorem 1.11,
could potentially be deduced from the work of [Opp], as a special case of Theorem 8.12 there. Our proof that
exploits the geometric structure of the spherical buildings, allows us to get a proof which is significantly shorter.
4
• X is an (ǫ, µ)-cosystolic expander.
• X has the c-topological overlapping property.
Finally, by combining the explicit construction of Ramanujan complexes of [LSV2], with
Corollary 1.12, we get an explicit construction of bounded degree cosystolic expanders with the
topological overlapping property in every dimension (Theorem 1.3). It is interesting to note
that no such random construction, i.e., of bounded degree high dimensional complexes with the
topological overlapping property, is currently known.
1.5
Organization of the paper
In section 2 we review some basic definition related to simplicial complexes, norms, links,
skeletons and high dimensional expansions. In section 3 we prove the local to global criterion
for cosystolic expansion (Theorem 1.3). In section 4 we prove a one sided mixing lemma for the
1-skeletons of regular complexes. In section 5 we review the definition of spherical buildings and
show that they are good skeleton expanders. Finally, in section 6, we prove that Ramanujan
complexes satisfies the conditions of Theorem 1.3, hence their skeletons are therefore examples
of bounded degree complexes with the topological overlap property.
1.6
Acknowledgments
The authors are grateful to David Kazhdan and Alex Lubotzky for useful discussions and
advices, and for Ron Rosenthal for many valuable improvements for this paper. The first
author would like to especially thank Alex Lubotzky for introducing, teaching and encouraging
him to work on this problem, without which this work would not have been possible. We also
thank the ERC, and BSF for their support. This work is part of the Ph.D. thesis of the first
author at the Hebrew University of Jerusalem, Israel.
2
Preliminaries on complexes and expansions
In this section we present the basic definitions and properties of simplicial complexes with
norms, as well as notions of high-dimensional expansions.
2.1
Complexes
Throughout this paper we shall use the following notations regarding simplicial complexes:
A simplicial complex, X, over a set of vertices, V , is family of subsets of V , X ⊂ 2V , which
is closed under inclusions, i.e. if F ∈ X and E ⊂ F then E ∈ X (note that the empty-set is
always a face in any complex). Call the elements of X, faces or simplices. The dimension of
a simplex F ∈ X, is defined as dim(F ) = |F | − 1, and the dimension of the entire complex is
defined as the maximal dimension of a simplex in it, i.e. dim(X) = maxF ∈X dim(F ). A complex
is said to be pure if all its maximal faces are of the same dimension.
For convenience sake, by a k-face we mean a k-dimensional face of the complex, and by a
d-complex we will always mean a finite d-dimensional pure simplicial complex.
Let X be a d-complex. For any −1 ≤ k ≤ d, denote by X(k) the collection of k-faces in X,
and by C k = C k (X; F2 ) = {f : X(k) → F2 } the space of k-cochains. As usual, denote by
X
δ = δk : C k → C k+1 ,
(δf )(τ ) =
f (σ)
(2.1)
σ⊂τ
5
the k-coboundry map, and by B k = B k (X, F2 ) = Im(δk−1 ) and Z k = Z k (X, F2 ) = ker(δk ) the
spaces of k-coboundries and k-cocycles, respectively. Recall that δk ◦ δk−1 = 0, hence B k ⊂ Z k .
Since we are working over F2 , any k-cochain is an indicator function of a subset of k-faces,
and vice versa, any subset of k-faces defines a k-cochain, i.e. {f : X(k) → F2 } ∼
= {A ⊂ X(k)},
where f 7→ supp(f ) and A 7→ 1A . So, from now we will identify each subset of k-faces, A ⊂ X(k),
with the following k-cochain f = 1A ∈ C k .
Definition 2.1 (Norm). Define the following norm on the space of cochains:
k · k = k · kk : C k → [0, 1],
kAk =
X
w(σ)
where
w(σ) =
σ∈A
|{F ∈ X(d) | σ ⊂ F }|
d+1
|σ| · |X(d)|
(2.2)
Note that for any −1 ≤ k ≤ d, and any A, B ∈ C k (i.e. A, B ⊂ X(k)), then:
• kAk = 0 if and only if A = ∅, and kAk = 1 if and only if A = X(k).
• kA ∪ Bk ≤ kAk + kBk with equality if and only if A and B are disjoint.
Definition 2.2 (Container). Let X be a d-complex and let −1 ≤ k ≤ r ≤ d. For any A ∈ C k ,
define Γr (A) ∈ C r to be the following r-cochain,
Γr (A) = {F ∈ X(r) | ∃ σ ∈ A, σ ⊂ F }.
(2.3)
Lemma 2.3. Let X be a d-complex and let −1 ≤ k ≤ r ≤ d. Then for any A ∈ C k ,
r+1
r
kAk ≤ kΓ (A)k ≤
· kAk.
k+1
(2.4)
Proof. First note that for any σ ∈ X(k), then
kΓr ({σ})k =
X
w(τ ) =
σ⊂τ ∈X(r)
=
X
1
|X(d)|
1
|X(d)|
X
kΓ (A)k = k
[
σ∈A
r
Γ ({σ})k ≤
X
σ∈A
=
d+1
X
1
|X(d)|
X
τ ∈X(r) F ∈X(d) r+1
F ∈X(d) τ ∈X(r)
σ⊂τ
τ ⊂F
σ⊂F σ⊂τ ⊂F
d+1
d−k
X d−k
·
r+1
r−k
k+1
r−k
=
w(σ) =
d+1
d+1
k+1
r+1
F ∈X(d) r+1
σ⊂F
So, for any A ∈ C k , on the one hand,
r
1
r
kΓ ({σ})k =
X r + 1
σ∈A
k+1
· w(σ) =
1
d+1
r+1
· w(σ) (2.5)
r+1
· kAk.
k+1
(2.6)
r+1
On the other hand, since each τ ∈ Γr (A) contains at most k+1
k-faces, i.e. τ can belong to at
r+1
r
most k+1 of the Γ ({σ}), hence, from the previous calculation we get,
X
[
r+1
r+1
r+1
r
r
· kAk =
kΓ ({σ})k ≤
·k
Γ ({σ})k =
· kΓr (A)k,
k+1
k+1
k+1
σ∈A
σ∈A
which finishes the proof.
6
(2.7)
2.2
Links
Next we introduce the notion of a link of complex, which is a combinatorial analogue of a unit
sphere in the complex. Thus, the links will serve us as local views of the complex.
Definition 2.4 (Link). Let X be a d-complex and let σ ∈ X. The link of σ in X is defined as
the following (d − |σ|)-complex,
Xσ = {τ ∈ X|σ ⊔ τ ∈ X}.
(2.8)
For any −1 ≤ k ≤ d − |σ|, denote by Cσk = C k (Xσ , F2 ) the space of k-cochains of Xσ , by
k · kσ = k · kkσ : Cσk → [0, 1] the norm associated to the complex Xσ , and by δσ = δσk : Cσk → Cσk+1
the k-coboundry map.
Let us observe the following cases of degenerate links of a d-complex:
• The link of the unique (−1)-face, the empty set ∅ ∈ X(−1), is the entire complex, X∅ = X.
• The link of a (d − 1)-face is a 0-complex, i.e. a collection of isolated vertices.
So, in order to avoid these degenerate cases, by a proper link we mean a link of a face σ of
dimension 0 ≤ dim(σ) ≤ d − 2.
Definition 2.5 (Localization and lifting). Let X be a d-complex and let σ ∈ X. Define the
following maps between the original complex, X, and the link complex, Xσ :
• The first map, called the localization (w.r.t. σ), takes a cochain of the original complex,
and restrict only to the faces which contains σ, and then delete σ from each of them,
producing a cochain of the link. Concretely,
Iσ : C ∗ → Cσ∗−|σ| ,
Iσ (A) := {τ ∈ Xσ | τ ⊔ σ ∈ A}.
(2.9)
• The second map, called the lifting (w.r.t. σ), takes a cochain of the link complex, and adds
σ to each face in it, producing a cochain of the original complex. Concretely,
I σ : Cσ∗ → C ∗+|σ| ,
I σ (A) := {τ ⊔ σ ∈ X | τ ∈ A}.
(2.10)
Let us mention a few immediate observations about the localization and lifting operators.
k−|σ|
Let X be a d-complex, σ ∈ X, |σ| ≤ k ≤ d, A ∈ Cσ
and A′ ∈ C k . Then the following holds:
Iσ ◦ I σ (A) = A
(2.11)
I σ ◦ Iσ (A′ ) = {τ ∈ A′ | σ ⊂ τ }
(2.12)
δ ◦ I σ (A) = I σ ◦ δσ (A)
(2.13)
The following three Lemmas 2.6, 2.7 and 2.8, describes some relations between the localization and lifting operators and the global and local norms of the complex (i.e. the norms of the
original complex X and its links Xσ , for any σ ∈ X).
Lemma 2.6. Let X be a d-complex and let σ ∈ X. For any 0 ≤ k ≤ d − |σ|, and any
A ∈ Ck − |σ|σ , then,
|σ| + k + 1
σ
kI (A)k =
· w(σ) · kAkσ .
(2.14)
k+1
7
Proof. Denote by, w = wX , the weight function of the original complex, and by, wσ = wXσ ,
the weight function of the link. In the language of links, the weight norm is interpreted as
τ (d−|τ |)|
(and similarly for wσ ). Since the norm of a cochain is define by extending
w(τ ) = |X
·|X(d)|
(d+1
|τ | )
linearly the weight function, it is suffice to show the claim for A which is a singleton. So, for
any τ ∈ Xσ (k) and A = {τ } ∈ Cσk , we have
w(σ) · kAkσ = w(σ) · wσ (τ ) =
=
|Xσ (d − |σ|)|
·
d+1
|σ| · |X(d)|
which finishes the proof.
|(Xσ )τ (d − |σ| − |τ |)|
|Xσ (d − |σ|)|
· d−|σ|+1
d+1
· |Xσ (d − |σ|)|
|σ| · |X(d)|
|τ |
|Xσ∪τ (d − |σ ∪ τ |)|
d−|σ|+1
· |Xσ (d − |σ|)|
|τ |
d+1
|σ∪τ |
= d+1 d−|σ|+1
|σ| ·
|τ |
=
|Xσ∪τ (d − |σ ∪ τ |)|
d−|σ|+1
d+1
· |X(d)|
|σ| ·
|τ |
· w(σ ∪ τ ) =
1
|σ|+|τ |
|τ |
· kI σ (A)k, (2.15)
Lemma 2.7. Let X be a d-complex,σ ∈ X and let |σ| ≤ k ≤ d.
k−|σ|
1. For any A, B ∈ Cσ
k−|σ|
2. If A ∈ C k , B ∈ Cσ
Proof.
, then kAkσ ≤ kBkσ if and only if kI σ (A)k ≤ kI σ (B)k.
and kAk ≤ kA + I σ (B)k, then kIσ (A)kσ ≤ kIσ (A) + Bkσ .
1. Follows immediately from Lemma 2.6.
2. First note that, every face of I σ (B) contains σ, i.e. for any σ 6⊂ τ , τ ∈ A ⇔ τ ∈ A+I σ (B).
Combining this observation with equation (2.12), we get
X
X
w(τ ) =
kAk − kI σ Iσ (A)k =
w(τ ) = kA + I σ (B)k − kI σ Iσ (A + I σ (B))k.
τ ∈A+I σ (B)
σ6⊂τ
τ ∈A
σ6⊂τ
(2.16)
Using now the assumption that kAk ≤ kA + I σ (B)k, imply that
kI σ Iσ (A)k ≤ kI σ Iσ (A + I σ (B))k.
(2.17)
Finally applying fact 1. on the last inequality, together with equation (2.11), yields
kIσ (A)kσ ≤ kIσ (A + I σ (B))kσ = kIσ (A) + Bkσ
(2.18)
which finishes the proof.
Lemma 2.8. Let X be a d-complex, 0 ≤ j ≤ k ≤ d, and let A ∈ C k . Then
X
k+1
· kAk =
kI σ (Iσ (A))k.
j+1
(2.19)
σ∈X(j)
Proof. By the definition of the norm and Fubini’s Theorem,
X
σ∈X(j)
σ
kI (Iσ (A))k =
X
X
σ∈X(j) σ⊂τ ∈A
w(τ ) =
XX
τ ∈A σ⊂τ
8
w(τ ) =
X k + 1
τ ∈A
j+1
· w(τ ) =
k+1
· kAk.
j+1
(2.20)
2.3
Skeletons
Here we define the notions of skeletons of a complex, which are basically what we get when we
forget about the higher dimensional faces of the given complex.
Definition 2.9 (Skeleton). Let X be a d-complex and let 0 ≤ k ≤ d. Then the k-skeleton of X
is defined as the following k-complex,
X (k) = {τ ∈ X| dim(τ ) ≤ k}.
For example, the 1-skeleton of a complex is simply its underlying graph.
Note that for any d-complex X, and any 0 ≤ t ≤ k ≤ d, the spaces of t-cochains of the
complex and its k-skeleton, are the same, and for t < k, the t-coboundary operators of the
complex and its k-skeleton, are also the same. On the other hand, the norms of the complex
and its k-skeleton, might be quite different. However, if the complex is of bounded degree, then
the norms of X and X (k) are proportional.
Lemma 2.10. Let t ≤ k ≤ d, Q ∈ N and let X be a d-complex which is Q-bounded degree at
dimension k, i.e. maxσ∈X(k) |Xσ (d − |σ|)| ≤ Q. Then for any A ∈ C t ,
d+1
1
· kAkX ≤ kAkX (k) ≤ Q ·
· kAkX
(2.21)
d−t
k+1
Q · k−t
Proof. Denote by, w = wX , the weight function of the original complex, and by, wk = wX (k) ,
the weight function of the skeleton. Since the norm of a cochain is define by extending linearly
the weight function, it is suffice to show the claim for A which is a singleton, i.e. it suffice to
prove for any σ ∈ X(t),
d+1
1
· w(σ)
(2.22)
· w(σ) ≤ wk (σ) ≤ Q ·
k+1
Q · d−t
k−t
Next note that, since the complex is pure and of Q-bounded degree at dimension k, then
1
· |X(k)| ≤ |X(d)| ≤ Q · |X(k)|.
d+1
k+1
(2.23)
Similarly, since that for any σ ∈ X(t), the link, Xσ , is pure and of Q-bounded degree at
dimension k − t − 1, then
1
d−t
k−t
· |Xσ (k − t − 1)| ≤ |Xσ (d − t − 1)| ≤ Q · |Xσ (k − t − 1)|.
we get that for any σ ∈ X(t),
and
|Xσ (k − t − 1)|
wk (σ) = k+1
≤
t+1 · |X(k)|
d−t
k−t ·
Q·
|Xσ (d − t − 1)|
w(σ) = d+1
≤
t+1 · |X(d)|
which finishes the proof.
Q · |Xσ (d − t − 1)|
=
k+1
t+1 · |X(d)|
d+1
k+1 · |Xσ (k − t
d+1
t+1 · |X(k)|
9
− 1)|
=
(2.24)
d+1
· Q · w(σ),
k+1
(2.25)
d−t
· Q · wk (σ),
k−t
(2.26)
Next, we define a property of graph expansion for a complex, called skeleton expansion,
which says that the 1-skeleton of the complex satisfies a form of weak mixing behavior.
Definition 2.11 (Skeleton expander). Let X be a d-complex and let α > 0. Call X an αskeleton expander, if for any A ⊂ X(0), then
kE(A, A)k ≤ 4 · (kAk2 + α · kAk)
(2.27)
where E(A, A) ⊂ X(1) are the edges in X with both vertices in A.
In section §4, we prove a one sided mixing Lemma for the skeleton of a regular complex,
which gives a criterion for skeleton expansion in terms of non trivial eigenvalues.
2.4
High dimensional expansion
Here we present different notions concerning high dimensional expansion for simplicial complexes, which arose in the works of Linial-Meshulam-Wallach [LinM], [MW] and Gromov [Gro].
Definition 2.12 (Coboundary and cocycle expansion). Let X be a d-complex and 0 ≤ k < d.
Define the k-dimensional cobonudary expansion parameter of X to be:
Expkb (X) = min{
kδ(A)k
minb∈B k kA + bk
|
A ∈ C k \ B k }.
(2.28)
Define the k-dimensional cocycle expansion parameter of X to be:
Expkz (X) = min{
kδ(A)k
minz∈Z k kA + zk
|
A ∈ C k \ Z k }.
(2.29)
Let us spell out what both expansion parameters says in the special case of graphs. In the
graph case d = 1 and k = 0, the coboundaries are just V and ∅, and the cocycles are the unions
of connected components of the graph. The cobonudary expansion parameter is equal to the
Cheeger constant of the entire graph (up to some normalization), while the cocycle expansion
parameter is equal to the minimum of the Cheeger constants in each connected component of
the graph. So, a large cocycle expansion parameter imply that each connected component of the
graph is a good expander on its own, however the graph itself can be disconnected, in particular
not an expander.
Let us recall the definition of coboundary expanders, from the introduction.
Definition 2.13 (Coboundary expander). A d-complex, X, is said to be an ǫ-coboundry expander, for some ǫ > 0, if Expkb (X) ≥ ǫ, for any 0 ≤ k ≤ d − 1.
The notion of coboundary expansion was first originate in the work of [LinM], in connection
to vanishing of cohomology. (The actual term, coboundary expansion, later came from [DK]).
Recall that for any 0 ≤ k ≤ d, each k-coboundary is a k-cocycle, i.e. B k (X; F2 ) ⊂ Z k (X; F2 ),
and the k-th cohomology of X (in F2 -coefficients) is the quotient space
H k (X; F2 ) = Z k (X; F2 )/B k (X; F2 ).
(2.30)
It is a simple exercise to see that the following equivalences holds:
Expkb (X) > 0
⇔
Z k (X; F2 ) = B k (X; F2 )
10
⇔
H k (X; F2 ) = 0.
(2.31)
Furthermore, if the k-th cohomology is trivial, hence any k-cocycle is a k-coboundary, then the
k-coboundary and k-cocycle expansion parameters are equivalent:
H k (X; F2 ) = 0
⇒
Expkb (X) = Expkz (X).
(2.32)
All in all we get the following equivalent characterization for coboundary expansion:
Expkb (X) ≥ ǫ
⇔
Expkz (X) ≥ ǫ
and Z k = B k .
(2.33)
As noted by Gromov (see also [DKW]), this notion of vanishing of cohomology is too strong
for some application, since the existence of a cocycle which is not a coboundary, is acceptable
just as long as it is not too small. This is where the definition of cosystoles comes into play.
Definition 2.14 (Cosystoles). Let X be a d-complex and 0 ≤ k ≤ d − 1. Define the k-cosystole
of X to be the minimal size of a k-cococycle which is not a k-coboundary, i.e.
Systk (X) = min{kzk | z ∈ Z k (X, F2 ) \ B k (X, F2 )}.
(2.34)
Finally, we recall the definition of cosystolic expanders from the introduction.
Definition 2.15 (Cosystolic expander). A d-complex X is said to be an (ǫ, µ)-cosystolic expander, for some ǫ, µ > 0, if Expkz (X) ≥ ǫ and Systk (X) ≥ µ, for any 0 ≤ k ≤ d − 1.
Remark 2.16. In recent years, the notion of cosystoles (and its close cousin, systoles) over F2 ,
have found applications in the field of quantum error correcting codes (see [GL] and [EOT]),
where good lower bounds on the systoles and cosystoles give rise to quantum codes with good
parameters.
2.5
Minimal cochains
Let us now introduce the following notions of minimal and locally minimal cochains, which we
will need later on to prove our local to global cosystolic expansion criterion.
Definition 2.17 (Mininmal and locally minimal). A cochain A ∈ C k is said to be minimal if
kAk = min{kA + bk | b ∈ B k (X, F2 )}.
(2.35)
A cochain A ∈ C k is said to be locally minimal if for any ∅ =
6 σ ∈ X, the localization of A w.r.t.
σ, Iσ (A), is a minimal cochain in the link Xσ .
Note that a d-complex, X, is an ǫ-coboundary expander if and only if for any 0 ≤ k ≤ d − 1,
any minimal k-cochain, A ∈ C k , satisfies kδ(A)k ≥ ǫ · kAk.
Lemma 2.18. If A ∈ C k is a minimal cochain, then A is a locally minimal cochain.
k−1−|σ|
Proof. Let ∅ =
6 σ ∈ X and let γ ∈ Cσ
Lemma 2.7 part 1, we get that
. From the minimality of A, equation (2.13), and
kAk ≤ kA + δI σ (γ)k = kA + I σ δσ (γ)k ⇒ kIσ (A)kσ ≤ kIσ (A) + δσ (γ)kσ .
(2.36)
This proves that A is locally minimal.
Lemma 2.19. If A is a minimal cochain and A′ ⊂ A, then A′ is also a minimal cochain.
11
Proof. First note that, since A \ A′ and A′ are disjoint, kAk = kA \ A′ k + kA′ k. Next note that,
since the sum of two cochains is equal to their symmetric difference, for any cochain c ∈ C k ,
(A + c) \ (A′ + c) = ((A \ c) ∪ (c \ A)) \ ((A′ \ c) ∪ (c \ A′ )) ⊂
⊂ ((A \ c) \ (A′ \ c)) ∪ ((c \ A) \ (c \ A′ )) = ((A \ c) \ (A′ \ c)) ⊂ A \ A′ (2.37)
where in the second to last step, the equality follows from the fact that A′ ⊂ A. Hence,
kA + ck − kA′ + ck ≤ k(A + c) \ (A′ + c)k ≤ kA \ A′ k = kAk − kA′ k
(2.38)
Now, if c ∈ B k is a coboundry, then by the minimality of A we get,
kA′ k = kA′ k − kAk + kAk ≤ kA′ k − kAk + kA + ck ≤ kA′ + ck
(2.39)
where the last inequality is (2.38), which finishes the proof.
3
Cosystolic expansion criterion
In this section we prove the following local to global cosystolic expansion criterion. The following
result is slightly stronger Theorem 1.10 from the introduction.
Theorem 3.1. For any d ∈ N, β > 0 and Q ∈ N, there exists ǫ = ǫ(d, β, Q) > 0, µ = µ(d, β) > 0
and α = α(d, β) > 0, such that if X is a d-complex satisfying:
• The complex X is Q-bounded degree, i.e. maxv∈X(0) |Xv | ≤ Q.
• The link Xσ is a β-coboundry expander, for any σ ∈ X, 1 ≤ |σ| ≤ d − 1.
• The link Xσ is an α-skeleton expander, for any σ ∈ X, 0 ≤ |σ| ≤ d − 1.
Then for any 0 ≤ k ≤ d − 2 and any 0 ≤ r ≤ d − 1,
Expkz (X) ≥ ǫ
and
Systr (X) ≥ µ
(3.1)
In particular, the (d − 1)-skeleton of X is an (ǫ, µ)-cosystolic expander.
In order to prove Theorem 3.1 we follow [KKL] strategy, who noticed that the following
(co)isoperimetric inequality for small cochains imply cosystolic expansion.
Theorem 3.2. For any d ∈ N and β > 0, there exists ǭ = ǭ(d, β), µ̄ = µ̄(d, β) > 0 and
α = α(d, β) > 0, such that if X is a d-complex satisfying:
• The link Xσ is a β-coboundry expander, for any σ ∈ X, 1 ≤ |σ| ≤ d − 1.
• The link Xσ is an α-skeleton expander, for any σ ∈ X, 0 ≤ |σ| ≤ d − 1.
Then for any 0 ≤ k ≤ d − 1,
A ∈ C k is locally minimal and kAk ≤ µ̄
=⇒
kδ(A)k ≥ ǭ · kAk
(3.2)
Remark 3.3. Note that Theorem 3.2, unlike Theorem 3.1, does not require any bounded degree
assumption. This simple fact, makes Theorem 3.2 useful also for unbounded degree complexes,
as was shown in [LLR], who proved that for any d ∈ N, there exists d-dimensional coboundary
expanders, which are of bounded degree at dimension d − 1, i.e., every (d − 1)-face is contained
in a bounded number of d-faces. (The work [LLR] is a generalization of [LubM]).
12
Remark 3.4. The constants in Theorem 3.1, are
µ :=
1
·
3(d + 2)2d+3
β
(d + 2) · 2d+2
d !2d+1
,
ǫ := min{
1
, µ},
Q
α := µ
1+
1
2d+1
. (3.3)
The constants in Theorem 3.2, are
µ̄ :=
1
·
3(d + 2)2d+3
β
(d + 2) · 2d+2
d !2d+1
d
β
1
ǭ := ·
,
3 (d + 2) · 2d+2
,
α := µ
1+
1
2d+1
.
(3.4)
We note that we made no attempt to optimize the constants in this work.
3.1
Sketch of the proof of Theorem 3.2
Before diving into the proof of Theorem 3.2, which require us to define a few new technical
notions, we would like to provide a proof sketch.
Fix a locally minimal cochain, A ∈ C k . We define the notion of fat faces with respect to A
in the complex, which essentially says that a face is fat if the localization of A with respect to
that face, is large. Consider the coboundaries of the localization of A with respect to the fat
faces, and note that these ”local” coboundaries are large by the assumption that the links are
cobundary expanders. The main idea of the proof is that each such ”local” coboundary, is either
an actual coboundary in the original complex, or it contains a fat face of smaller dimension.
This last claim, is the content of Proposition 3.11, where the term Lη (A, i) essentially stands
for the sub-cochain of elements in A which contains a fat i-faces, and Υη (A) stands for the error
term. This error term is negligible due to the skeleton expansion assumption (Proposition 3.12).
So, iterating Proposition 3.11, we get that either δ(A) is large, or that there is a fat (−1)-face.
But the unique (−1)-face is fat precisely when A is large, which finishes the proof.
3.2
Fat machinery
In order to prove the (co)isoperimetric inequality (Theorem 3.2), we first construct a ”fatmachinery” which allow us to move calculations from higher to lower dimensions in the complex.
Definition 3.5 (Fat faces). Fix a cochain A ∈ C k and 0 < η < 1. Define inductively the
i-cochain of fat faces, w.r.t. A and η, i = 0, . . . , k, by
Sηk (A) = A
and
k−i
Sηi−1 (A) = {σ ∈ X(i − 1) | kIσ (Sηi (A))kσ ≥ η 2
}
(3.5)
For i = −1, 0, . . . , k, call the elements of Sηi (A) fat faces (w.r.t. A and the fatness constant η).
The following Lemma shows that the sizes of the cochains of fat faces is bounded by the size
of the original cochain (up to some constant).
Lemma 3.6. Let X be a d-complex, −1 ≤ i ≤ k ≤ d, A ∈ C k and 0 < η < 1. Then,
k−i
kSηi (A)k ≤ η −2
· kAk
(3.6)
Proof. For any, −1 ≤ j ≤ k − 1, and any fat j-face, σ ∈ Sηj (A), by applying Lemma 2.6 on the
cochain Iσ (Sηj+1 (A)) ∈ Cσ0 , we get
kI σ (Iσ (Sηj+1 (A)))k
k−j−1
η −2
≤
w(σ) =
j +2
(j + 2) · kIσ (Sηj+1 (A))kσ
13
· kI σ (Iσ (Sηj+1 (A)))k
(3.7)
Hence, combining this with Lemma 2.8,
kSηj (A)k =
X
σ∈Sηj (A)
w(σ) ≤
k−j−1
X
η −2
k−j−1
·
· k(Sηj+1 (A))k (3.8)
kI σ (Iσ (Sηj+1 (A)))k = η −2
j+2
j
σ∈Sη (A)
Hence, by iterating on equation (3.8) for j = i, . . . , k − 1, we get,
k−i−1
kSηi (A)k ≤ η −2
k−i−1 +2k−i−2 )
· k(Sηi+1 (A))k ≤ η −(2
... ≤ η −
Pk−1
j=i
2k−j
· k(Sηi+2 (A))k ≤ ...
k−i +1
kSηk (A)k = η −2
k−i
· kAk ≤ η −2
· kAk (3.9)
which finishes the proof.
From Lemma 3.6, we get the following consequence, which says that for a small cochain the
unique (−1)-face is a non-fat face. (This simple fact will serve as the finishing argument in the
proof of Theorem 3.2).
k+1
Corollary 3.7. Let X be a d-complex, k ≤ d, A ∈ C k and 0 < η < 1. If kAk < η 2
unique (−1)-face, the empty set is not a fat face, i.e. ∅ 6∈ Sη−1 (A).
then the
Proof. Note that the empty set has the following interesting property: ”A local view by ∅ is
everything”, i.e. for any Y ⊂ X then I∅ (Y ) = Y and k · k∅ = k · k. So, from Lemma 3.6 and the
assumption on the size of A, we get
k
k
kI∅ (Sη0 (A))k∅ = kSη0 (A)k ≤ η −2 · kAk < η 2
(3.10)
hence, by definition, the empty set is not a fat face.
Next, we define the cochain of degenerate faces, which intuitively one should think of as the
error-term when one is trying to move from higher dimension to lower dimension.
Definition 3.8 (Degenerate faces). Fix a cochain A ∈ C k and 0 < η < 1. A dead-end is a pair
of two equal sized fat faces, (σ, σ ′ ), whose intersection is a codimension-1 non-fat face, i.e.
|σ| = |σ ′ | = |σ ∩ σ ′ | + 1, σ, σ ′ ∈ Sη∗ (A), σ ∩ σ ′ 6∈ Sη∗−1 (A)
(3.11)
A face p ∈ X is said to be degenerate if it contains a dead-end in it, and define Υη (A) ∈ C k+1
to be the cochain of all (k + 1)-faces which are degenerate.
Next, we define the notion of a fat ladder.
Definition 3.9 (Fat ladders). Fix a cochain A ∈ C k and 0 < η < 1. For any fat i-face,
σ ∈ Sηi (A), define the k-cochain of fat-ladders of A, siting on σ, to be
Lη (A, σ) = {t ∈ A | ∃ σ = σi ⊂ . . . ⊂ σk = t, σj ∈ Sηj (A), ∀ j = i, . . . , k}
S
Define the k-cochain of i-fat-ladders of A, by Lη (A, i) = σ∈Sηi (A) Lη (A, σ).
(3.12)
The next lemma essentially says that if inside a (k + 1)-face you have a fat ladder and a
fat k-face, then you can either go down to a deeper ladder, or you contain a dead end, which
makes the ambient face degenerate.
Lemma 3.10. Let X be a d-complex, −1 ≤ i ≤ k ≤ d, A ∈ C k and 0 < η < 1. Let σ ∈ Sηi (A),
t ∈ Lη (A, σ), t′ ∈ A and t, t′ ⊂ p ∈ X(k + 1). Then, either t′ ∈ Lη (A, t′ ∩ σ) or p ∈ Υη (A).
14
Proof. By definition, there exists σ = σi ( . . . ( σk = t, where all σj are fat. Define σk′ = t′
′
′
and σj−1
= t′ ∩ σj = σj′ ∩ σj for any j = i, . . . , k. If all the σj′ are fat, and since t′ ∩ σ = σi−1
⊂
′
′
′
′
′
′
. . . ⊂ σk = t , so by removing repetitions if needed (σj = σj+1 ) we get that t ∈ Lη (A, t ∩ σ).
′
Otherwise, there is a non-fat σj′ , and w.l.o.g. we may assume j is maximal, i.e. σj+1
is fat, and
′
′
′
since that σj+1 6⊂ t (otherwise σj = σj+1 is fat), we get that |σj+1 | = |σj+1 | = |σj+1 ∩ σj+1 | − 1,
hence p ∈ Υη (A).
3.3
Proof of Theorem 3.2
The following Proposition is the heart of the ”fat machinery” which will allow us to move
calculations from higher to lower dimensions.
Proposition 3.11. Let k < d ∈ N, 0 < η < 1, 0 < β and let X be a d-complex such that for
any σ ∈ X, 1 ≤ |σ| ≤ d − 1, the link, Xσ , is a β-coboundary expander. Then for any locally
minimal k-cochain, A ∈ C k , and any 0 ≤ i ≤ k,
β
k+2
i+1
· kLη (A, i)k ≤ kδ(A)k + (k + 2) · kLη (A, i − 1)k + kΥη (A)k
Proof. Let us evaluate the following expression,
[
R=
I σ ◦ δσ ◦ Iσ (Lη (A, σ)) ⊂ X(k + 1).
(3.13)
(3.14)
σ∈S i (A)
On the one hand, since A is locally minimal, then Iσ (A) is minimal, and Iσ (Lη (A, σ)) ⊂ Iσ (A),
so by Lemma 2.19 , Iσ (Lη (A, σ)) is also minimal. So, by the assumption that all the proper
links of X are β-coboundry expanders, we get that for any σ ∈ Sηi (A) ⊂ X(i),
β · kIσ (Lη (A, σ))kσ ≤ kδσ ◦ Iσ (Lη (A, σ))kσ .
(3.15)
Note that Lη (A, σ) ⊂ {τ ∈ X(k)|σ ⊂ τ }, hence by equation (2.12), I σ ◦Iσ (Lη (A, σ)) = Lη (A, σ).
Combining this with equation
(3.15), Lemma 2.7 part 1, and the fact that each element of R
contains at most k+2
faces
from
Sηi (A) ⊂ X(i), we get
i+1
X
X
k+2
σ
β ·kLη (A, i)k ≤
β ·kLη (A, σ)k ≤
kI ◦δσ ◦Iσ (Lη (A, σ))k ≤
·kRk. (3.16)
i+1
i
i
σ∈Sη (A)
σ∈Sη (A)
On the other hand, for any σ ∈ Sηi (A) and any p ∈ I σ ◦ δσ ◦ Iσ (Lη (A, σ)), one of the following
three possibilities must occur:
1. All the k-faces in p which belongs to A, contains σ, and they are all from Lη (A, σ). In
this case p ∈ δ(A).
2. All the k-faces in p which belongs to A, contains σ, but not all of them are from Lη (A, σ).
In this case there is some t′ ∈ A \ Lη (A, σ) such that σ ⊂ t′ ⊂ p. Now, since p ∈
I σ ◦ δσ ◦ Iσ (Lη (A, σ)), there must be at least one t ∈ Lη (A, σ) such that t ⊂ p. So, by
Lemma 3.10 we get that p ∈ Υη (A).
3. There is a k-face in p which belongs to A, and it does not contain σ. In this case there
is some, t′ ∈ A, such that σ 6⊂ t′ ⊂ p. Like before, since p ∈ I σ ◦ δσ ◦ Iσ (Lη (A, σ)), there
must be at least one t ∈ Lη (A, σ) such that t ⊂ p. And again by Lemma 3.10 we get that
either p ∈ Υη (A), or t′ ∈ Lη (A, t′ ∩ σ) ⊂ Lη (A, i − 1), i.e. p is a (k + 1)-face which contains
a k-face from Lη (A, i − 1), hence by definition 2.2 p ∈ Γk+1 (Lη (A, i − 1)).
15
In conclusion we get that
R ⊂ δ(A) ∪ Γk+1 (Lη (A, i − 1)) ∪ Υη (A).
(3.17)
Combining equation (3.16) and (3.17), together with Lemma 2.3, yields
β
k+2
i+1
· kLη (A, i)k ≤ kRk ≤ kδ(A)k + (k + 2) · kLη (A, i − 1)k + kΥη (A)k
(3.18)
which finishes the proof.
The following Proposition gives an effective bound on the cochain of degenerate faces in
terms of the skeleton expansion and the fatness constant.
k+1
and let X be a d-complex such
Proposition 3.12. Let k < d ∈ N, 0 < η < 1, 0 < α ≤ η 2
that for any σ ∈ X, 0 ≤ |σ| ≤ d − 1, the link, Xσ , is an α-skeleton expander. Then for any
k-cochain A ∈ C k ,
kΥη (A)k ≤ (k + 2) · 2k+4 · η · kAk.
(3.19)
Proof. For any t ≤ d and any t-cochain Y , denote by E(Y, Y ) the (t + 1)-cochain of (t + 1)-faces
which contains at least two different t-faces from Y , and let Γk+1 (Y ) ∈ C k+1 be as in Lemma
2.3. Then, by the definition of the fat-degenerate faces, we get
Υη (A) =
k−1
[
[
Γk+1 (I σ (E(Iσ (Sηi+1 (A)), Iσ (Sηi+1 (A)))))
(3.20)
i=−1 σ∈X(i)\Sηi (A)
So, by Lemma 2.3, we get
kΥη (A)k ≤
k−1
X
X
i=−1 σ∈X(i)\Sηi (A)
k+2
· kI σ (E(Iσ (Sηi+1 (A)), Iσ (Sηi+1 (A))))k
i+2
(3.21)
From the skeleton expansion, we get for any σ ∈ X(i),
kE(Iσ (Sηi+1 (A)), Iσ (Sηi+1 (A)))kσ ≤ 4 · kIσ (Sηi+1 (A))kσ · (kIσ (Sηi+1 (A))kσ + α)
k
|σ|
Now, by the Lemma 2.6, we can multiply both sides by
(3.22)
· w(σ), and get
kI σ (E(Iσ (Sηi+1 (A)), Iσ (Sηi+1 (A))))k ≤ 4 · kI σ (Iσ (Sηi+1 (A)))k · (kIσ (Sηi+1 (A))kσ + α)
(3.23)
Next, if σ ∈ X(i) \ Sηi (A), then by the definition of fat faces,
k−i
kI σ (E(Iσ (Sηi+1 (A)), Iσ (Sηi+1 (A))))k ≤ 4 · kI σ (Iσ (Sηi+1 (A)))k · (η 2
+ α)
(3.24)
Summing this over all non-fat i-faces,
X
k−i
σ∈X(i)\Sηi (A)
kI σ (E(Iσ (Sηi+1 (A)), Iσ (Sηi+1 (A))))k ≤ 4(η 2
k−i
≤ 4(η 2
+ α) ·
X
σ∈X(i)
+α)·
σ∈X(i)\Sηi (A)
k−i
kI σ (Iσ (Sηi+1 (A)))k = 4(η 2
16
X
kI σ (Iσ (Sηi+1 (A)))k
+ α) · (i + 2) · kSηi+1 (A)k (3.25)
where the last equality follows from Lemma 2.8. Applying Lemma 3.6, we get
X
k−(i+1)
k−(i+1)
kI σ (E(Iσ (Sηi+1 (A)), Iσ (Sηi+1 (A))))k ≤ 4(i+2)·(η 2
+α·η −2
)·kAk (3.26)
σ∈X(i)\Sηi (A)
Combining equations (3.21) and (3.26) together, we get
k−1
X
k+2
k−(i+1)
k−(i+1)
kΥη (A)k ≤
) · kAk
+ α · η −2
· 4(i + 2) · (η 2
i+2
i=−1
!
!
k−1
k−1
X
X
k+2
k+1
k
−2k−0
2k−k
≤ 4·
)·kAk ≤ 4(k + 2) ·
+α·η
· (i + 2) ·(η
·(η+α·η −2 )·kAk
i+2
i+1
i=−1
i=−1
k+3
≤ (k + 2) · 2
· (η + α · η
−2k
) · kAk ≤ (k + 2) · 2k+3 · 2η · kAk (3.27)
which finishes the proof.
Finally, we are able to prove the isoperimetric inequality for small cochains (Theorem 3.2).
Proof of Theorem 3.2. Define the following constants,
c0 :=
β
,
(k + 2) · 2k+2
ǭ :=
ck0
,
2
η :=
ǭ
,
(k + 2)2k+4
k+1
µ̄ := η 2
and
k+1 +1
α := η 2
. (3.28)
k+1
Note that by definition Lη (A, k) = A, and if kAk ≤ µ̄ = η 2 , then by Corollary 3.7 the only
(−1)-face, the empty set, is non fat (w.r.t. A and η), and hence kLη (A, −1)k = 0. Therefore,
for any constant c ≤ 1 we get,
k
c · kAk =
k
X
i=0
i−1
c
k
X
(c · kLη (A, i)k − kLη (A, i − 1)k). (3.29)
(c · kLη (A, i)k − kLη (A, i − 1)k) ≤
i=0
β
, for any 0 ≤ i ≤ k. So, by applying Proposition
(k+2
i+1 )
3.11 on equation (3.29), with the constant c0 , we get
Note that c0 ≤ 1 and that (k + 2) · c0 ≤
k
ck0 · kAk ≤
X β
1
( k+2 · kLη (A, i)k − (k + 2) · kLη (A, i − 1)k) ≤ kδ(A)k + kΥη (A)k. (3.30)
·
k+2
i+1
i=0
Combining this with Proposition 3.12, we get
kδ(A)k ≥ (ck0 − (k + 2)2k+4 · η) · kAk = ǭ · kAk,
(3.31)
which finishes the proof.
3.4
Proof of Theorem 3.1
The fact that a coisoperimetric inequality for small cochains (Theorem 3.2), implies a cosystolic
expansion (Theorem 3.1), was first shown in [KKL, § 4], but for the sake of being self-contained
we add here their argument.
Proposition 3.13. [KKL, Proposition 2.5] Let X be a d-complex, and define Q = maxv∈X(0) |Xv |.
Let 0 ≤ k ≤ d. Then for any A ∈ C k , there exists γ ∈ C k−1 , which satisfies:
17
1. The cochain A + δ(γ) is locally minimal.
2. kA + δ(γ)k ≤ kAk.
3. kγk ≤ Q · kAk.
d+1
Proof. First note that N (A) := k+1
·|X(d)|·kAk is a non-negative integer. We prove the claim
by induction on N (A). In the base case N (A) = 0, then A = ∅ ∈ C k is the empty k-cochain,
and the claim holds for γ = ∅ ∈ C k−1 the empty (k − 1)-cochain. Assume the claim holds for
all cochains A′ ∈ C k such that N (A′ ) < N (A), i.e. such that kA′ k < kAk.
If A is locally minimal, then the claim holds for γ = ∅ ∈ C k−1 the empty (k − 1)-cochain.
k−1−|σ|
Otherwise, there exist ∅ =
6 σ ∈ X, and some c ∈ Cσ
, such that
kIσ (A) + δσ (c)kσ < kIσ (A)kσ
(3.32)
Denote c′ = I σ (c) ∈ C k−1 . Then by equation (2.13) and Lemma 2.7 part 2, we get
kA + δ(c′ )k < kAk.
(3.33)
So, N (A + δ(c′ )) < N (A), and since both are natural numbers, then N (A + δ(c′ )) ≤ N (A) − 1.
By the induction assumption there exist γ ′ ∈ C k−1 , such that:
1. (A + δ(c′ )) + δ(γ ′ ) = A + δ(c′ + γ ′ ) is a locally minimal cochain,
2. kA + δ(c′ + γ ′ )k ≤ kA + δ(c′ )k < kAk, and
3. kγ ′ k ≤ Q · kA + δ(c′ )k ≤ Q · (kAk −
1
d+1
(k+1
)·|X(d)|
).
Q
(since c ⊂ Xσ and |Xσ | ≤ Q), we
d+1
(k+1
)·|X(d)|
get that kγk ≤ kγ ′ k + kck ≤ Q · kAk, which finishes the proof.
Hence by taking γ = γ ′ + c, and noting that kck ≤
Now, using Proposition 3.13 and the isoperimetric inequality for small cochains (Theorem
3.2), we are able to prove the cosystolic expansion criterion (Theorem 3.1).
Proof of Theorem 3.1. Let µ̄ and ǭ be the constants from Theorem 3.2 and let Q = maxv∈X(0) |Xv |.
1
} and µ = µ̄.
Define ǫ = min{µ̄, Q
We begin by proving the cocycle expansion. Let A ∈ C k . First note that if kδ(A)k ≥ µ, and
since kAk ≤ 1 for any cochain, then
kδ(A)k ≥ µ ≥ ǫ ≥ ǫ · kAk
(3.34)
So, let us assume kδ(A)k ≤ µ, and let γ ∈ C k , be as in Proposition 3.13 apply on the cochain
δ(A). Then δ(A) + δ(γ) = δ(A + γ) is a locally minimal cochain and kδ(A + γ)k ≤ kδ(A)k ≤ µ.
By Theorem 3.2 and the fact that δ ◦ δ = 0 we get that
0 = kδ(δ(α + γ))k ≥ ǭ · kδ(α + γ)k
⇒
δ(α + γ) = 0.
(3.35)
So, A + γ ∈ Z k , and hence γ = A + (A + γ) ∈ {A + z | z ∈ Z k }. Now, by Proposition 3.13 part
(3), kγk ≤ Q · kδ(A)k, and we get
kδ(A)k ≥
1
· kγk ≥ ǫ · kγk = ǫ · kA + (A + γ)k ≥ ǫ · min kA + zk
Q
z∈Z k
18
(3.36)
which gives us the cocycle expansion Expkz (X) ≥ ǫ.
Next we prove the cosystolic bound. Let A ∈ Z k \ B k (if no such cocycle exists there is
nothing to prove). By Proposition 3.13, let A′ = A + δ(γ) be such that A′ is locally minimal
and kA′ k ≤ kAk. Note that since A ∈ Z k \ B k and δ(γ) ∈ B k then A′ ∈ Z k \ B k also. If
kA′ k ≤ µ = µ̄, then by Theorem 3.2 and the fact that A′ is a cocycle, we get
0 = kδ(A′ )k ≥ ǭ · kA′ k
⇒
A′ = 0
(3.37)
which is a contradiction since, 0 ∈ B k , and A′ is not in B k . So, kAk ≥ kA′ k ≥ µ, which gives
us the cosystolic bound Systk (X) ≥ µ.
4
Skeleton mixing lemma
The purpose of this section is to prove a one sided mixing lemma for the 1-skeleton of a regular
complex (see below), i.e. giving a spectral criterion for skeleton-expansion.
Remark 4.1. We note that mixing behavior, spectral gaps and random walks of high dimensional complexes, are subjects that have been already intensively studied in several works (see
[EGL], [GS], [GW], [KM1], [KM2], [KR], [Opp] [Par], [PRT], [PR] and [Ros]). The mixing
Lemma that we present here, is much simpler then the ones appearing in the above mentioned
works, and it is actually a result about partite-regular graph, rather then complexes.
Definition 4.2 (Regular complex). A d-complex X, is said to be regular if it satisfies:
F
Q
• There exist a partition of the vertices X(0) = di=0 Vi , such that X(d) ⊂ di=0 Vi .
Q
• For any I ⊂ J ⊂ [d] := {0, 1, . . . , d}, thereQexist kIJ ∈ N, such that each p ∈ X ∩ i∈I Vi ,
is contained in exactly kIJ faces from X ∩ j∈J Vj .
For example, in the simplest case where X is of dimension 1, i.e. a graph, then X is a
regular complex if and only if X is a bipartite biregular graph.
Remark 4.3. Note that if X is a regular complex, then so does all of its links and skeletons.
Next, we define what is the ”second eigenvalue” of the 1-skeleton of a regular complex.
Definition 4.4 (Non-trivial eigenvalue). Let X be a regular d-complex.
For any 0 ≤ i < j ≤ d,
F
define the (i, j)-type induced bipartite graph to be X(i,j) = (Vi Vj , X(1) ∩ Vi × Vj ). Denote
by λ(X(i,j) ) its normalized second largest eigenvalue. Define the normalized largest non-trivial
eigenvalue of (the 1-skeleton of ) X to be
λ(X) := max λ(X(i,j) ).
i6=j
Let us now state the one sided mixing lemma for the 1-skeleton of a regular complex.
Proposition 4.5. Let X be a regular complex, and let λ(X) be its normalized largest non-trivial
eigenvalue. Then, for any A, B ⊂ X(0),
kE(A, B)k ≤ 2(
p
d+1
) · (kAk · kBk + λ(X) · kAk · kBk)
d
where E(A, B) ⊂ X(1) are the edges in X with vertices from both A and B.
Note that in the 1-dimensional case, such a Mixing Lemma is already known.
19
(4.1)
F
Lemma 4.6. [EGL, Corollary 3.4] Let G = (V1 V2 , E) be a bipartite biregular graph, and let
λ(G) be its normalized second largest eigenvalue. Then, for any A ⊂ V1 and any B ⊂ V2 ,
s
|A| |B|
|E(A, B)|
|A| |B|
−
| ≤ λ(G) ·
(4.2)
|
|E(X)|
|V1 | |V2 |
|V1 | |V2 |
This bipartite mixing lemma will imply the general skeleton mixing lemma.
Proof of Proposition 4.5. First note that since X is Q
a regular complex, and let X(0) =
be the partition, then for any I ⊂ [d] and A ⊂ X ∩ i∈I Vi ,
[d]
X |{F ∈ X(d) | σ ∈ F }|
|A| · kI
|A|
d+1
Q
=
=
· kAk =
Q
[d]
|X(d)|
|X
∩
|I|
|X
∩
V
|
·
k
i∈I Vi |
σ∈A
i∈I i
I
Fd
i=0 Vi
(4.3)
In particular, for any i, j ∈ [d] and any A ⊂ Vi , B ⊂ Vj ,
kAk =
1
d+1
1
|A|
, kBk =
|Vi |
1
d+1
1
|B|
, and kE(A, B)k =
|Vj |
1
d+1
2
|E(A, B)|
|X(1) ∩ Vi × Vj |
(4.4)
So restating Lemma 4.6 in terms of the complex norm, we get for any A ⊂ Vi , B ⊂ Vj ,
kE(A, B)k ≤
d+12
1
kAk
d+1
2
· kBk + λ(X(i,j) )
d+1 p
1
kAk
d+1
2
≤ 2(
· kBk
λ(X) p
d+1
) · (kAk · kBk +
· kAk · kBk) (4.5)
d
d+1
Now, let A, B ⊂ X(0),
F and denote Ai = A ∩ Vi and Bi = B ∩ Vi for any 0 ≤ i ≤ d. Since X is
partite, E(A, B) = i6=j E(Ai , Bj ), hence
kE(A, B)k =
X
i6=j
Similarly, since A =
kE(Ai , Bj )k ≤ 2(
F
λ(X)
d+1 X
)·
(kAi k · kBj k +
d
d+1
i6=j
q
kAi k · kBj k)
F
Ai and B = j Bj , then
X
X
X
kAi k · kBj k ≤ (
kAi k) · (
kBj k) = kAk · kBk
(4.6)
i
i
i6=j
(4.7)
j
Next, note that for any N non-negative numbers x1 , . . . , xN ∈ R≥0 , one has
N
N
X
X
x2i
xi )2 ≤ N · max (x2i ) ≤ N ·
(
i=1
1≤i≤N
p
Applying this for N = (d + 1)2 and xi = kAi k · kBj k, we get
s
X
Xq
kAi k · kBj k ≤ (d + 1)2 ·
kAi k · kBj k ≤ (d + 1) · kAk · kBk
i,j
i,j
which finishes the proof.
In particular, since 2( d+1
d ) ≤ 4 for any d ∈ N, we get the following.
Corollary 4.7. Let X be a regular d-complex. Then X is an λ(X)-skeleton expander.
20
(4.8)
i=1
(4.9)
5
Spherical buildings
The object of this section is to introduce the notion of spherical buildings, and to show that
they are good skeleton expanders.
5.1
Definition of spherical buildings
Here we give a definition of spherical buildings, and list some of their properties which we shall
use. For more on buildings we refer to [AB].
Before defining a building, let us first define the notion of a chamber complex.
Definition 5.1 (Chamber complex). A d-dimensional simplicial complex, X, is said to be a
chamber complex, if it is pure (i.e. all maximal faces are d-dimensional), and for any two
maximal faces C, C ′ ∈ X(d), there is a sequence of d-faces, C = C1 , . . . , Cn = C ′ , such that for
each i = 1, . . . , n − 1, the intersection Ci ∩ Ci+1 is a (d − 1)-face.
For a d-dimensional chamber complexes, call a d-face a chamber, call a (d − 1)-face a panel,
and call a sequence as above, C = C1 , . . . , Cn = C ′ , a gallery from C to C ′ .
A chamber complex, X, is said to be thin if each panel is contained in exactly 2 chambers,
and it is said to be q-thick, q > 1, if each panel is contained in exactly q + 1 chambers. By a
thick chamber complex, we mean q-thick for some q > 1.
Then, one way to define a building is as follows (for the equivalence for the more common
definition, see [AB, Theorem 4.131]).
Definition 5.2 (Building). A building is a thick chamber complex together with a family of
subcomplexes, called apartments, which satisfy the following axioms:
• Each apartment is a thin chamber complex.
• Any two faces in the complex are contained in a common apartment.
• Any two apartments have an isomorphism which fixes their intersection.
A building is said to be spherical/affine/hyperbolic if it is finite/locally finite/otherwise, respectively.
Let us note that if B is a d-dimensional building (i.e. a d-dimensional chamber complex
which satisfy the axioms of the building), then each apartment of B is also d-dimensional.
Remark 5.3. Throughout this paper we only concerns ourselves with buildings which are simplicial complexes. However, it should be mentioned that buildings can also be poly-simplicial
complexes (just by allowing chamber complexes to be such).
Let us present an example of a spherical building.
Example 5.4. [AB, § 4.3] Let q be a prime power, d ∈ N, and denote V = Fdq . Consider the
simplicial complex, P(V ), whose vertices are the proper (i.e. not {0} or V ) subspaces of V , and
his faces are the flags of subspaces in V , i.e. {0} < W1 < . . . < Wt < V . Then P(V ) is a
(d − 2)-dimensional q-thick spherical building. Moreover, the group P GLd (Fq ) acts on P(V ) in
a strongly transitive way (see below).
Next, we wish to list some basic properties of spherical buildings, for the complete proofs
we refer to [AB].
21
Lemma 5.5. For any d ∈ N define θd := max{2d · (d + 1)!, 192 · 11!}. Then each apartment in
a d-dimensional spherical building is of size at most θd .
Proof. By [AB, Theorem 4.131] each apartment in a spherical building is a spherical Coxeter
complex, and by [AB, § 1.3,1.5.6] the spherical Coxeter complexes were completely classified,
and θd is taken to be the maximal size of all such possible complexes.
Corollary 5.6. For any d, q ∈ N define θd := max{2d · (d + 1)!, 192 · 11!} and Qd,q := ((d + 1) ·
(q + 1))θd . Then each d-dimensional q-thick spherical building is of size at most Qd,q .
Proof. Let C0 some chamber in the building. Since the building is d-dimensional and q-thick,
we get that C0 has at most (d + 1) · (q + 1) chambers of gallery distance at most 1 from C0 .
Iterating this fact we get that there are at most ((d+1)·(q +1))n chambers of gallery distance at
most n from C0 . Since each chamber in the building is contained in some apartment containing
C0 , and since the distance of two chambers inside the apartment is at most the size of the
apartment, which by Lemma 5.5 is at most θd , the claim is proven.
Remark 5.7. The bound ((d + 1) · (q + 1))θd on the size of the spherical building is by no mean
tight, since we are not trying to optimize the constants in this work.
Definition 5.8 (Type function). A d-dimensional complex, X, is said to admit a (d + 1)-type
function on the vertices if there is a function F
τX : X(0) → [d] := {0,
Qd1, . . . , d}, such that, setting
d
−1
Vi := τX ({i}) for i = 0, . . . , d, then X(0) = i=0 Vi and X(d) ⊂ i=0 Vi .
Lemma 5.9. [AB, Proposition 4.6] Let B be a d-dimensional building. Then B admits a
(d + 1)-type function on its vertices.
Lemma 5.10. [AB, Proposition 4.9] Let B be a building and let σ ∈ B be any face in it. Then
the link, Bσ , is also a building.
Lemma 5.11. [AB, Proposition 4.40] Let B be a building and let A be an apartment in it. Then
A is gallery convex, i.e. for any two chambers in A, C, C ′ ∈ A, and any gallery from C to C ′
in the building B, C = C0 , . . . , Cn = C ′ , which is of minimal length among all possible galleries
from C to C ′ , then the gallery sits completely inside the apartment A, i.e. C0 , . . . , Cn ∈ A.
Lemma 5.12. [AB, Proposition 5.122 (2)] Let B be a spherical building and let C be a chamber
op
in B. For any apartment containing C, A, there is a unique chamber in A, denoted CA
, which
is of maximal gallery distance from C.
Next we define a notion of a building which posses many symmetries.
Definition 5.13 (Strongly transitive action). A building, B, is said to posses a strongly transitive action, if there exist a group of automorphisms on the building, G ≤ Aut(B), such that:
• G preserves the (d + 1)-type function of the building as defined in Lemma 5.9.
• For any two pairs, (C1 , A1 ) and (C2 , A2 ), of a chamber, Ci , and an apartment containing
the chamber, Ai , i = 1, 2, there exists g ∈ G, such that g(C1 ) = C2 and g(A1 ) = A2 .
Lemma 5.14. Let B be a d-dimensional building and G a group that acts strongly transitively
on it. Then B is a regular complex (as defined in §4).
22
Proof. By Lemma 5.9, there exists a type-function τB : B(0) → [d], i.e. if Vi := τB−1 ({i}) ⊂ B(0),
F
Q
i = 0, . . . , d, then B(0)
= di=0 Vi and B(d) ⊂ di=0 Vi . Now, let I ⊂ J ⊂ [d] be two type sets,
Q
and let σ, σ ′ ∈ B ∩ i∈I Vi be two I-type faces. Choosing two chambers, C and C ′ , which
contains σ and σ ′ respectively, and by the second property of the strong transitivity there exists
g ∈ G such that g(C) = C ′ . Also by the first property of the strong transitivity, g preserves
the types of τB , hence g(σ) = σ ′ . Since g is an automorphism, the J-type faces containing σ
are mapped bijectively to the J-type faces containing σ ′ , in particular they are of the same
cardinality, proving that the building is regular.
Throughout this section we shall make use of the following notion from group theory.
Definition 5.15 (Stabilizer). Let X be any simplicial complex and G ≤ Aut(X) a group of
automorphisms on X. For any σ ∈ X, define the stabilizer of σ in G to be the following subgroup
of G:
Gσ = stabG (σ) = {g ∈ G | g(σ) = σ}.
Lemma 5.16. Let B be a building and G a group that acts strongly transitively on it. Then
for any face, σ ∈ B, and any apartment containing it, A, every Gσ -orbit in B passes through
the apartment A, i.e. for any τ ∈ B there exists g ∈ Gσ such that g(τ ) ∈ A.
Proof. Let C ∈ A be a maximal face containing σ, and let C ′ ∈ B be a maximal face containing
τ . By the second axiom of the building, there exists an apartment A′ which contains both C
and C ′ . By the strong-transitivity, there exists g ∈ G such that g(C) = C and g(A′ ) = A. On
the one hand, g preserves the type function, hence g(σ) = σ, i.e. g ∈ stabG (σ). On the other
hand, g(A′ ) = A, in particular g(τ ) ⊂ g(C ′ ) ∈ A, as needed.
Lemma 5.17. Let B be a building which posses a strongly transitive action, and let σ ∈ B.
Then the link Bσ also posses a strongly transitive action.
Proof. Let G ≤ Aut(B) be the group that acts strongly transitive on B. define Gσ = stabG (σ)
the stabilizer of σ in G, then Gσ admits an action on the link, i.e. Gσ ≤ Aut(Bσ ). The action
of Gσ is type-preserving since the action of G is. Any pair of a chamber and an apartment
containing it inside Bσ , (C, A), can be lifted to a pair of a chamber and an apartment containing
it inside B, (C̃, Ã). Let (C, A) and (C ′ , A′ ), be two pairs, where each pair contains a chamber
and an apartment containing that chamber, inside Bσ . Lifting them to such pairs in B, (C̃, Ã)
and (C̃ ′ , Ã′ ), then there is g ∈ G, such that (g(C̃), g(Ã)) = (C̃ ′ , Ã′ ), and since both pairs
contains σ, then g ∈ Gσ , and hence g(C) = C ′ and g(A) = A′ , which finishes the proof.
5.2
Expansion of spherical buildings
It was first observed by Gromov [Gro] that spherical buildings are coboundary expanders (see
[LMM] for a simplified proof).
Theorem 5.18. For any d ∈ N, there exist βd > 0, such that any d-dimensional spherical
building is a βd -coboundary expander.
The purpose of this subsection is to prove that the spherical buildings of sufficiently large
thickness are good skeleton expanders.
Theorem 5.19. Let X be a d-dimensional q-thick spherical building which posses a strongly
θd
d
transitively action. Then X is an √
q -skeleton expander, where θd := max{2 · (d + 1)!, 192 · 11!}.
23
The strategy for proving Theorem 5.19 is as follows: First, we define a property for bipartite
graphs (symmetric-convex), and show that such graphs have good bound on their second largest
eigenvalue. Second, we show that the type-induced graphs of a spherical building which posses
a strongly transitively action satisfy this property. Finally, by applying Corollary 4.7 we get
the skeleton expansion.
F
Definition 5.20 (Symmetric convex graph). Let X = (V1 V2 , E) be a bipartite graph, let
G ≤ Aut(X) be a group that acts by graph automorphisms on X, and let θ ∈ N. For any vertex
v in X, denote by, Gv = stabG (v) = {g ∈ G|g(v) = v}, its stabilizer.
Say that X is a θ-symmetric-convex graph, if for any v ∈ V1 , then:
1. The number of Gv -orbits in X is at most θ.
2. There is a unique Gv -orbit of vertices in V1 of maximal distance from v, and a unique
Gv -orbit of vertices in V2 of maximal distance from v.
3. For any vertex u, which is not of maximal distance from v, the number of neighbors w of
u such that dist(v, w) < dist(v, u), is at most θ.
Proposition 5.21. Let X be a bipartite (k, k ′ )-biregular θ-symmetric-convex connected graph,
then the normalized second largest eigenvalue of X is bounded by,
1
1
λ(X) ≤ θ · max{ √ , √ }
k
k′
(5.1)
Proof. Let A = AX ∈ End(CV (X) ) be the adjacency operator of X, and let Spec(A) =
{λn , . . . , λ2 , λ1 } be its set of eigenvalues. Let us recall some basic facts (see [EGL, § 3]): Since X
is an undirected graph the operator A is self-adjoint, hence Spec(A) ⊂ R. Since X is bipartite,
√
then λn−i+1 = −λi for any i = 1, . . . , n. Finally, since X is (k, k ′ )-biregular, then λ1 = k · k′ .
Note that any eigenvectors of the eigenvalues, λn , λ1 , are non-zero on every vertex, and if
f2 ∈ CV (X) is an eigenvector of λ2 , there most exist v ∈ V1 , such that f2 (v) 6= 0.
Fix such a vertex v ∈ V1 , let K = stabG (v) be its stabilizer in G, and define the following
directed-multi-graph X̄ = X/K as follows: The vertices of X̄ are the K-orbits of the vertices of
X, i.e. [u] = {k(u)|k ∈ K} for some u ∈ V (X), and the number of edges in X̄ from [u] to [w] is
equal to the number of edges in X between the vertex u and the set of vertices {k(w)|k ∈ K}
(note that this is independent of the choice of u). Finally, let Ā ∈ End(CV (X̄) ) be the adjacency
operator of X̄, and let Spec(Ā) ⊂ C be its set of eigenvalues. By the definition of X̄ and Ā, we
get that for any u ∈ V (X) and any [w] ∈ V (X̄),
X
Ā[u],[w] =
Au,w .
(5.2)
w∈[w]
Note that since X̄ is directed, a priori there is no reason for all the eigenvalues of Ā to be
real, however, if λ is an eigenvalue of Ā, and f ∈ CV (X̄) is an eigenvector of λ, then defining
f ′ ∈ CV (X) by f ′ (u) = f ([u]), we get that for any u ∈ V (X),
X
X
Ā[u],[w]f ([w]) = Āf ([u]) = λ · f ([u]) = λ · f ′ (u) (5.3)
Af ′ (u) =
Au,w f ′ (w) =
w∈V (X)
[w]∈V (X̄)
hence, f ′ is an eigenvector of A with eigenvalue λ. In particular,
Spec(Ā) ⊂ Spec(A) ⊂ R
24
(5.4)
On the other hand, if λ ∈ Spec(A),Pwith an eigenvector f ∈ CV (X) such that f (v) 6= 0, then
1
¯
defining f¯ ∈ CV (X̄) by f¯([u]) = |K|
k∈K f (k(u)), we get f 6≡ 0, such that for any [u] ∈ V (X̄),
Āf¯([u]) =
X
Ā[u],[w]f¯([w]) =
X
k∈K w∈V (X)
Ā[u],[w]
1 X
Af (k(w))
|K|
k∈K
[w]∈V (X̄)
[w]∈V (X̄)
1 X
=
|K|
X
1 X
1 X
Au,w f (k(w)) =
Af (k(u)) =
λ · f (k(u)) = λ · f¯([u]) (5.5)
|K|
|K|
k∈K
k∈K
hence f¯ is an eigenvector of Ā with eigenvalue λ. In particular,
λn , λ2 , λ1 ∈ Spec(Ā) ⊂ R
(5.6)
and by applying the trace formula for Ā2 , we get
λ22 + 2 · k · k′ = λ2n + λ22 + λ21 ≤
X
λ2 = tr(Ā2 )
(5.7)
λ∈Spec(Ā)
Finally, let us use the properties of the symmetric-convex graph: By property 1 the number
of vertices in X̄ is at most θ. By property 2 in each of the two parts of X̄, there is a unique
vertex, [vi ], i = 1, 2, of maximal distance from [v], and since X is (k, k′ )-biregular, so does
X̄, and hence [v1 ], [v2 ] has at most k · k′ directed 2-paths starting and ending with them. By
property 3 for any other vertex [u] in X̄, [u] 6= [v1 ], [v2 ], the number of directed 2-paths starting
and ending with [u] is at most θ · max{k, k′ }, since such a 2-path corresponds to a a following
2-path in X, u ∼ w ∼ u′ = k(u) for some vertex w and k ∈ K, so either dist(v, u) < dist(v, w)
in which case there are at most θ such possible u′ = k(u) (note dist(v, u) = dist(v, k(u))), or
dist(v, u) > dist(v, w) in which case there are at most θ such possible w. All in all we get that
X
tr(Ā2 ) =
#{directed 2-paths starting and ending with [u]} ≤ 2·k·k′ +(θ−2)·θ·max{k, k′ }
[u]∈X̄
(5.8)
Combining this with (5.2), we get
λ22 ≤ tr(Ā2 ) − 2 · k · k′ ≤ θ 2 · max{k, k′ }
and noting that λ(X) =
λ2
λ1
=
√λ2 ,
k·k ′
(5.9)
finishes the proof.
The following Proposition shows that the type-induced bipartite graphs of the spherical
buildings are symmetric-convex, with a constant θ that depends only on the dimension of the
building (and not on its thickness!).
Proposition 5.22. Let B be a d-dimensional q-thick spherical building which posses a strongly
transitive action. Let i 6= j ∈ [d] and let B(i,j) be the (i, j)-type induced bipartite biregular graph
of B. Then B(i,j) is an θd -symmetric-convex graph with both regularity degrees at least q + 1,
where θd := max{2d · (d + 1)!, 192 · 11!}.
Proof. Let v be a vertex in the building, Gv = stabG (v), and let A be some fixed apartment
that contains v.
First, let us prove the claim on the regularity degrees: By Lemma 5.14, B is a regular
complex, hence B(i,j) is a bipartite biregular graph. Assume v is of i-type, so it most be
contained in some panel of j-cotype, σ, (i.e. the σ contain a vertex of each type except j), and
25
by the q-thickness σ is contained in q + 1 chambers, C0 , . . . , Cq , each of which contains a unique
j-type vertex, u0 , . . . , uq , which is a neighbour of v in the graph B(i,j) (and of course the same
reasoning apply when replacing i and j).
1) By Lemma 5.16, the number of Gv -orbits is bounded by the size of an apartment in the
building, which, by Lemma 5.5, is bounded by θd .
2) Let C be some chamber in A which contains v. By Lemma 5.12 there is a unique chamber
op
∈ A of maximal gallery distance from C in A, and let e = {v1 , v2 } ⊂ C op be (the unique)
CA
edge of type {i, j} inside it. Now, since gallery distance is coarser then graph distance (any
gallery path contains in it a graph path), then v1 and v2 are the two farthest vertices from v
(w.r.t. the graph metric of B(i,j) ), of type i and j respectively, inside the apartment A. On the
other hand, since Gv is a collection of automorphisms, hence preserves distances, and by Lemma
op
5.12, there is a unique chamber, CA
′ , of maximal gallery distance from C, in each apartment
op
op
′
A , we get that for any two apartments A′ , A′′ , there is some k ∈ Gv such that k(CA
′ ) = CA′′ .
So, [v1 ] is the unique Gv -orbit in Vi of maximal distance vertices from v, and similarly [v2 ] is
the unique Gv -orbit in Vj of maximal distance vertices from v.
3) Since u is not of maximal distance from v, there is some neighbour w∗ of u such that
dist(v, u) < dist(v, w∗ ). Now by the axiom of the building let A be an apartment that contains
both v and the edge (u, w∗ ). Let w1 , . . . , wn be all the vertices in B(i,j) , which are neighbours
of u and satisfy dist(v, wi ) < dist(v, u) for any i = 1, . . . , n. Then, any minimal path from v to
w∗ which passes through u, most also pass through some wi . Hence, since gallery distance is
coarser then graph distance, any minimal gallery in the building from a maximal face containing
{v}, to a maximal face containing {u, w∗ }, most also pass through a maximal face containing
{u, wi } for some i. Now, if A is an apartment containing v and {u, w∗ }, then by Lemma 5.11 all
these minimal galleries lies in A, in particular all w1 , . . . , wn lies in A, i.e. n ≤ |A| ≤ θd (where
the last inequality is Lemma 5.5).
Finally, by combining Propositions 5.21, 5.22 and Corollary 4.7, we are able to prove that
spherical buildings are good skeleton expanders (Theorem 5.19).
Proof of Theorem 5.19. By Proposition 5.22, each type induced bipartite graph, B(i,j) , of the
building is θd -symmetric-convex. By Proposition 5.21, each B(i,j) has a normalized largest nonθd
trivial eigenvalue at most √
q . Finally applying Corollary 4.7 we get that the building is an
θd
√
q -skeleton
6
expander.
Ramanujan complexes
Ramanujan complexes were defined in [LSV1], and were explicitly constructed in [LSV2]. They
are finite quotients of Bruhat-Tits buildings of type Ã, which exhibits excellent spectral properties. For more on Ramanujan complexes, we refer the readers for the survey [Lub2].
The object of this section is to show that Ramanujan complexes of sufficiently large degree
satisfies the requirement of Theorem 3.1, namely:
Theorem 6.1. Let X be a d-dimensional q-thick Ramanujan complex. Then:
• The complex X is Qd,q -bounded degree, where Qd,q is as in Corollary 5.6.
• Each proper link of X is a βd -link coboundry expander, where βd is as in Theorem 5.18.
• Each proper link of X is a
θd
√
q -skeleton
expander, where θd is as in Theorem 5.19.
26
• X itself is a 2d · q −(d−1)/2 -skeleton expander.
Before proving the Theorem we shall need the following Lemma.
Lemma 6.2. Any proper link of a Ramanujan complex is a spherical building which posses a
strongly transitive action.
Proof of Lemma 6.2. First, since a Ramanujan complex is a quotient of an affine building of
type Ã, any link of a Ramanujan complex is a link of the covering affine building. Second, since
an Ã-type building posses a strongly transitive action, then by Lemmas 5.10 and 5.17, each link
of it is also a building which posses a strongly transitive action. Finally, since an affine building
is locally finite, then each proper link of it is finite, hence a spherical building.
Proof of Theorem 6.1. The first three claims follows from Lemma 6.2 together with Corollary
5.6, Theorem 5.18 and Theorem 5.19. We are left to show that X itself is an excellent skeleton
expander. Using the notation of [Lub2, Section 2.1], the adjacency operator of the type induced
graph X(i,j) is the Hecke operator Aj−i restricted to that graph, hence by [Lub2, Remark 2.1.5],
−(d−1)
d − i(d−i)
2
≤ 2d q 2
λ(X) ≤ max λ(Ai ) ≤ max
q
1≤i≤d
1≤i≤d i
(6.1)
Combining this with Corollary 4.5, proves that X is a (2d · q −(d−1)/2 )-skeleton expander.
Consequentially, combining the above Theorem 6.1, together with Theorems 3.2, 3.1 and
1.8, give us the following high dimensional expansion properties for Ramanujan complexes
of sufficiently large degree (this is Corollary 1.12 from the introduction, together with the
coisoperimetric inequality for small cochains).
Corollary 6.3. For any d, q ∈ N, there exists q0 = q0 (d) > 0, µ = µ(d) > 0, ǭ = ǭ(d) > 0,
ǫ = ǫ(d, q) > 0 and c = c(d, q) > 0, such that if X is the d-dimensional skeleton of a (d + 1)dimensional q-thick Ramanujan complex with q ≥ q0 , then:
• If A is a locally minimal cochain of X and kAk ≤ µ, then kδ(A)k ≥ ǭ · kAk.
• X is an (ǫ, µ)-cosystolic expander.
• X posses the c-topological overlapping property.
Finally, by [LSV2, Theorem 1.1], we get that for any 2 ≤ d ∈ N and any prime power
q, there exist an infinite family of d-dimensional q-thick Ramanujan complexes. Moreover,
these Ramanujan complexes are constructed explicitly. Hence, by applying Corollary 6.3 on
the Ramanujan complexes constructed in [LSV2], we get an affirmative answer to Gromov’s
question from the introduction (Theorem 1.3).
As mentioned in the introduction, as an immediate application of Theorem 1.3, we get:
Corollary 6.4 (Corollary 1.4). For any D ≥ 2d + 1, there exists C = C(D) > 0 and an
infinite family of d-dimensional bounded degree complexes, {Xn }n , which satisfies the following:
For any embedding of Xn into RD , such that the distance between the images of any two non
adjacent simplices of Xn is at least 1, then the volume of the 1-neighborhood of the image of
1
Xn , is at least C · |Xn | D−d .
Proof of Corollary 6.4. Combine Theorem 1.3 with [GG, Proposition 2.5].
27
The original result [GG, Theorem 2.2] gave the slightly weaker result that, for any ǫ > 0,
there exists an infinite family of d-dimensional complexes, {Xn }n , whose degree might depends
1
of ǫ (but not on |Xn |), and the lower bound on the volume of the 1-neighborhood is C ·|Xn | D−d −ǫ .
The reader is referred to [GG] to learn more on the work of Kolmogorov and Barzdin, and its
generalization to higher dimensions.
We close this work with the following concluding remarks:
Remark 6.5. Theorem 6.1 and Corollary 6.3 holds for any finite quotient of a Bruhat-Tits
building of large degree, not just for Ramanujan complexes. To prove this generalization, one
needs only to prove that the underlying graph of any finite quotient of a Bruhat-Tits building
of large degree, is a good expander graph. This can be done by using Oh’s explicit property (T)
[Oh], to get explicit bounds on the second eigenvalue of such a graph.
Remark 6.6. Note that for a d-dimensional Ramanujan complexes of sufficiently large degree,
we were only able to prove that their (d − 1)-skeletons are cosystolic expanders. Following
[Lub2, Conjecture 3.2.4], we suspect that the Ramanujan complexes themselves (and in fact
all finite quotients of Bruhat-Tits buildings) are cosystolic expanders. In contrast, in [KKL,
Proposition 1.5] it was shown that not all Ramanujan complexes of sufficiently large degree are
coboundary expanders, i.e. cosystolic expansion is the best one could hope for in general.
References
[AB]
P. Abramenko, K. S. Brown, Buildings Theory and Applications, Graduate Texts
in Mathematics 248, Springer.
[Bar]
I. Barany, A generalization of Carathodorys theorem, Discrete Mathematics
40(2): 141-152, (1982).
[BF]
E. Boros, Z. Furedi, The number of triangles covering the center of an n-set,
Geometriae Dedicata 17(1): 69-77, (1984).
[BK]
Y. M. Barzdin, A. N. Kolmogorov, On the realization of nets in 3-dimensional
space, Probl. Cybernet, 8: 261-268, (1967). See also Selected Works of A.N.
Kolmogorov, Kluwer Academic Publishers, 3: 194-202,245, (1993).
[DK]
D. Dotterrer, M. Kahle, Coboundary expanders, Journal of Topology and Analysis, 4(4): 499-514, (2012).
[DKW]
D. Dotterrer, T. Kaufman, U. Wagner, On Expansion and Topological Overlap, LIPIcs-Leibniz International Proceedings in Informatics. Vol. 51. Schloss
Dagstuhl-Leibniz-Zentrum fuer Informatik, (2016).
[EGL]
S. Evra, K. Golubev, A. Lubotzky, Mixing properties and the chromatic number of Ramanujan complexes, International Mathematics Research Notices,
2015(22): 11520-11548, (2015).
[EOT]
L. Eldar, M. Ozols, K. F. Thompson, Upper Bounds on Systoles of HighDimensional Expanders Using Quantum Codes, arXiv preprint arXiv:1610.07478,
(2016).
[FGLNP]
J. Fox, M. Gromov, V. Lafforgue, A. Naor, J. Pach, Overlap properties of geometric expanders, Journal fr die reine und angewandte Mathematik, 671: 49-83,
(2012).
28
[Gar]
H. Garland, p-Adic curvature and the cohomology of discrete subgroups of p-adic
groups, Annals of Mathematics, 375-423, (1973).
[Gro]
M. Gromov, Singularities, Expanders and Topology of Maps. Part 2: from Combinatorics to Topology Via Algebraic Isoperimetry, Geometric and Functional
Analysis, 20(2): 416-526, (2010).
[GG]
M. Gromov, L. Guth, Generalizations of the KolmogorovBarzdin embedding estimates, Duke Mathematical Journal, 161(13): 2549-2603, (2012).
[GL]
L. Guth, A. Lubotzky, Quantum error correcting codes and 4-dimensional arithmetic hyperbolic manifolds, Journal of Mathematical Physics 55.8: 082202,
(2014).
[GS]
A. Gundert, M. Szedlk, Higher dimensional discrete Cheeger inequalities, Proceedings of the thirtieth annual symposium on Computational geometry, ACM,
(2014).
[GW]
A. Gundert, U. Wagner, On Laplacians of random complexes, Proceedings of the
twenty-eighth annual symposium on Computational geometry, ACM, 151-160,
(2012).
[HLW]
S. Hoory, N. Linial, A. Wigderson, Expander graphs and their applications, Bulletin of the American Mathematical Society, 43(4): 439-561, (2006).
[KKL]
T. Kaufman, D. Kazhdan, A. Lubotzky, Isoperimetric Inequalities for Ramanujan Complexes and Topological Expanders, Geometric and Functional Analysis
26(1): 250-287, (2016).
[KM1]
T. Kaufman, D. Mass, High Dimensional Combinatorial Random Walks and Colorful Expansion, arXiv preprint arXiv:1604.02947, (2016).
[KM2]
T. Kaufman, D. Mass, Walking on the Edge and Cosystolic Expansion, arXiv
preprint arXiv:1606.01844, (2016).
[KR]
A. Knowles, R. Rosenthal, Eigenvalue confinement and spectral gap for random
simplicial complexes, arXiv preprint arXiv:1509.02034, (2015).
[Lub1]
A. Lubotzky, Expander graphs in pure and applied mathematics, Bulletin of the
American Mathematical Society, 49(1): 113-162, (2012).
[Lub2]
A. Lubotzky, Ramanujan complexes and high dimensional expanders, Japanese
Journal of Mathematics, 9(2): 137-169, (2014).
[LinM]
N. Linial, R. Meshulam, Homological connectivity of random 2-complexes, Combinatorica, 26(4): 475-487, (2006).
[LubM]
A. Lubotzky, R. Meshulam, Random Latin squares and 2-dimensional expanders,
Advances in Mathematics 272 : 743-760, (2015).
[LMM]
A. Lubotzky, R. Meshulam, S. Mozes, Expansion of building-like complexes,
Groups, Geometry, and Dynamics, 10(1): 155-175, (2016).
[LSV1]
ed , Israel
A. Lubotzky, B. Samuels, U. Vishne, Ramanujan complexses of type A
Journal of Mathematics, 149: 267-300, (2005).
29
[LSV2]
[LLR]
A. Lubotzky, B. Samuels, U. Vishne, Explicit construction of Ramanujan comed , European Journal of Combinatorics, 26(6): 965-993, (2005).
plexes of type A
A. Lubotzky, Z. Luria, R. Rosenthal, Random Steiner systems and bounded degree coboundary expanders of every dimension, arXiv preprint arXiv:1512.08331,
(2015).
[MW]
R. Meshulam, N. Wallach, Homological connectivity of random k-dimensional
complexes, Random Structures & Algorithms 34(3): 408-417, (2009).
[Oh]
H. Oh, Uniform pointwise bounds for matrix coefficients of unitary representations and applications to Kazhdan constants, Duke mathematical journal,
113(1): 133-192, (2002).
[Opp]
I. Oppenheim, Isoperimetric Inequalities for Local Spectral Expanders and Topological Expanders, arXiv preprint, arXiv:1501.04940, (2015).
[Par]
O. Parzanchevski, Mixing in high-dimensional expanders, arXiv preprint,
arXiv:1310.6477, (2013).
[PRT]
O. Parzanchevski, R. Rosenthal, R. J. Tessler, Isoperimetric inequalities in simplicial complexes, Combinatorica 1-33, (2012).
[PR]
O. Parzanchevski, R. Rosenthal, Simplicial complexes: spectrum, homology and
random walks, Random Structures & Algorithms, (2016).
[Pin]
M.S. Pinsker, On the complexity of a concentrator, 7th International Teletraffic
Conference, Stockholm, 4: 1-318, (1973).
[Ros]
R. Rosenthal, Simplicial branching random walks and their applications, arXiv
preprint, arXiv:1412.5406, (2014).
30
| 4math.GR
|
Condition number-free query and active learning of linear families
arXiv:1711.10051v1 [cs.LG] 27 Nov 2017
Xue Chen∗
[email protected]
The University of Texas at Austin
Eric Price
[email protected]
The University of Texas at Austin
November 29, 2017
Abstract
We consider the problem of learning a function from samples with ℓ2 -bounded noise. In the
simplest agnostic learning setting, the number of samples required for robust estimation depends
on a condition number that can be arbitrarily large. We show how to improve this dependence in
two natural extensions of the setting: a query access setting, where we can estimate the function
at arbitrary points, and an active learning setting, where we get a large number of unlabeled
points and choose a small subset to label. For linear spaces of functions, such as the family of
n-variate degree-d polynomials, this eliminates the dependence on the condition number. The
technique can also yield improvements for nonlinear spaces, as we demonstrate for the family of
k-Fourier-sparse signals with continuous frequencies.
∗
Supported by NSF Grant CCF-1526952 and a Simons Investigator Award (#409864, David Zuckerman), part of
this work was done while the author was visiting the Simons Institute for the Theory of Computing.
1
Introduction
A common task in many fields is to estimate a signal f (x) from noisy observations yi ≈ f (xi ). This
problem takes many forms, depending on the measurement model, the signal structure, and the
desired approximation norms. The classical solution to this problem is empirical risk minimization
(ERM), which for m observations (x1 , y1 ), . . . , (xm , ym ) and the ℓ2 loss function outputs
n
1 X
(yi − f ′ (xi ))2
fe = arg min
f ′ ∈F m
(1)
i=1
where F is the family of signals being learned. An important special case of this is when F is a
linear vector space, such as the set of polynomials. In this case, solving (1) is referred to as ordinary
least squares (OLS) regression.
Consider the mean-zero bounded-variance noise setting, where the xi are drawn from some
distribution D, the signal f actually lies in F, and yi = f (xi ) + ηi for independent mean zero
random variables ηi with E[ηi2 ] ≤ σ 2 . In this case, we would like to recover an fe that satisfies
kfe − f k2D := E (fe(x) − f (x))2 ≤ εσ 2
x∼D
(2)
with 3/4 probability. In the body of this paper we will also consider noise that is not independent
or mean zero, for which the desired guarantee is weaker.
The sample complexity m required to achieve the guarantee (2) depends on the function family
F and the sample distribution D. If functions in F have high variance at positions x that are
rarely sampled under D, then the sample complexity can be quite high. One way to bound this is
to measure the “condition number”
K := sup
sup
f (x)2
(3)
f ∈F x∈supp(D)
kf kD =1
where kf k2D = Ex∼D f (x)2 as in (2). A bounded condition number lets us apply the Chernoff
bound, so for any fixed function f ∈ F the empirical norm will approximate the actual norm with
1 − δ probability after O(K log(1/δ)) samples. When F is a dimension d vector space, classical
results (e.g., Chapter 11 of [GKKW06]) used this to show an O( Kd εlog d ) bound on the sample
complexity for OLS to guarantee (2). By applying matrix Chernoff, this has been improved to
O(K log d + Kε ) [CDL13].
The condition number K is always at least d, but can be much larger in natural settings. For
example, univariate degree d − 1 polynomials sampled uniformly over [−1, 1] have K = Θ(d2 ),
because Chebyshev polynomials contain most of their ℓ2 mass in a 1/d2 size region [CDL13]. And
K can be even worse; adversarial distributions D can drive K arbitrarily large. Unfortunately, as
we show in Section 7, the sample complexity required to achieve (2) under random samples really
is Θ(K log d + Kε ), for any K, d, and ε, even via algorithms other than OLS.
In this work, we show how to avoid the dependence on the condition number in two alternative
models of access to the signal f . The first model is that we may query f wherever we want. The
second model is active learning, where we get a fairly large number of unlabeled points xi and may
choose a small set of them to see the labels yi .
Query access. In the query access model, we may freely choose the points xi at which we see
yi = f (xi ) + ηi . We also know the distribution D, which is only used to measure distances under
1
k·kD ; one may think of D as uniform over an interval such as [−1, 1]. The query access model is the
one typically used in function estimation problems such as sparse Fourier transforms. Even though
the set of functions with sparse Fourier transform do not form a linear space, our techniques do
have implications for the problem, as we shall discuss.
An algorithm in the query access model can of course simulate one in the sampling model by
choosing x ∼ D, getting an O(K log d + Kε ) sample complexity. But by biasing the choices of x
towards points of high variance—that is, engaging in importance sampling—we can improve the
sample complexity. In particular, we sample our points from the distribution
DF (x) = D(x) · sup
f ∈F
kf kD =1
f (x)2
d
(4)
and output the fe that minimizes an appropriately weighted version of (1). In Corollary 5.1, we
show that this algorithm achieves (2) with only O(d log d+ dε ) samples. We also show in Theorem 7.1
that at least the dε term is necessary.
We then consider arbitrary non–independent non–mean-zero noise, where yi = f (xi ) + g(xi ) for
an arbitrary function g. In this case, since f + g may itself lie in F, one cannot hope to converge
to f as the number of samples grows to infinity. Instead of (2), we guarantee
kfe − f kD ≤ (1 + ε)kgkD
(5)
with the same O(d log d + dε ) sample complexity.
Active learning. Active learning is a model of semisupervised learning, appropriate in situations
where unlabeled data is common but labeling data requires expensive human annotation. Such
situations arise in a variety of domains, including speech recognition, information extraction, and
document classification [Set09].
In the active learning model we consider, the algorithm receives a set of unlabeled points xi
from the unknown distribution D. It then chooses a subset of those points, and receives the labels
yi = f (xi ) + ηi associated with them. One would like to minimize both the number of labeled and
unlabeled examples required to learn the function.
Let us start by trying to optimize each parameter in isolation. If we merely wanted to minimize
the number of unlabeled samples, we would label every element and revert to the standard sampling
model. Hence any active learning algorithm that achieves the guarantee (2) must take at least
Ω(K log d + Kε ) unlabeled samples.
On the other hand, if we merely wanted to minimize the number of labeled samples, we would
take an infinite number of unlabeled samples. These samples would let us learn the distribution D,
and let us label whatever points x ∈ supp(D) that we want. The resulting problem is identical to
the query access model, so we know that any active learning algorithm must label Ω( dε ) examples,
and we cannot hope for less than O(d log d + dε ) labels without improving the result in the query
access model.
We show that there is no tradeoff between the two criteria: we give an active learning algorithm
that takes O(K log d + Kε ) unlabeled points and requests O(d log d + dε ) labels, and achieves (2). As
with the query access model, if the noise is not independent or mean zero, the algorithm achieves (5)
with the same sample complexity.
2
Nonlinear function families. The above discussion is focused on function families F that are
linear vector spaces, such as n-variate polynomials of degree d. Some of the techniques have more
general application; we present one such example.
Consider estimating a function in the family of continuous k-Fourier-sparse signals
f (x) =
k
X
vj e2πifj x
(6)
j=1
with real-valued frequencies fj having bandlimit |fj | ≤ F , over the domain D = [−1, 1]. Because
the frequencies fj may be arbitrarily close to each other, this family is not very well conditioned.
It was shown in [CKPS16] how to achieve (2) for this family in the query access model using
poly(k log F ) samples and time. An important step was to bound the condition number
K=
sup f (x)2
sup
x∈supp(D)
f ∈F
kf kD =1
by O(k4 log3 k). We show that by querying points according to a nonuniform distribution biased towards high-variance positions x, analogous to DF from (4)—in fact, identical up to normalization—
the relevant condition number changes to
κ=
sup f (x)2 .
E
x∈supp(D)
f ∈F
kf kD =1
When F was a dimension d linear space, one could show that κ = d, which is why we could
replace K by d in the sample complexities. For k-Fourier-sparse signals, we can instead show
that κ = O(k2 log2 k)—which is still better than our bound on K. This in turn leads to a k2
factor improvement in the sample complexity. We use a simple net argument to show this for an
exponential time recovery algorithm, but the same thing should hold for fast algorithms by going
through the (rather long) [CKPS16] machinery.
The final sample complexity bound we obtain is presumably not optimal, both because the other
pieces of the proof remain inefficient and because κ might actually be smaller. Still, the argument
demonstrates how non-uniform sampling can improve the sample complexity for nonlinear function
spaces, and we expect that a similar sampling scheme will be necessary to achieve an optimal
sample complexity for continuous Fourier-sparse signals.
1.1
Related work
Many different query strategies have been considered in the active learning literature (see [Set09]
and references therein). Most of these, such as uncertainty sampling [LG94] and query-by-committee [SOS92],
are designed for the “serial” setting, where the algorithm iteratively requests labels and updates its
model estimate. Our algorithm only queries a single batch, which is still (given our lower bounds)
within a log factor of optimal. Variance reduction sampling techniques in the literature are also
serial [Coh94, GBD92], but one could consider our procedure as performing a form of worst-case
variance reduction. Our result has a similar “agnostic” form to the active learning considered
in [BBL06], although the problems are quite different.
Our proposed algorithm is more similar to strategies proposed in the importance sampling
literature [Owe13]. Importance sampling aims to find a sampling distribution D ′ , much like our
DF , that minimizes the variance in estimating a given function. One key difference is that the
3
alternative distribution D ′ in importance sampling is typically not defined with respect to a class
F of functions f , as done here.
There is an enormous literature on least squares regression, in a variety of settings, which we
cannot do justice to; see, for example, [SL12] for an overview. Many of these results are analyzed
in a “fixed design” setting, which considers a deterministic set of observation points xi , and only
wants to minimize the error at those points. One relevant work is [HKZ14], which analyzes ridge
regression—that is, linear regression with ℓ2 regularization—and aims for robust recovery given
samples from the same distribution D as error is measured in. The sample complexity then depends
on properties of D.
Another body of closely related work involves leverage score sampling [DMM08, MI10, Woo14],
the goal of which is to approximately solve a finite linear regression problem y = Xβ + η; this is
essentially the same as the fixed design linear regression problem. Our sample distribution DF can
be seen as a continuous limit of the leverage score sampling distribution, so the condition number
bound we show in Theorem 4.1 is very similar to the result in [MI10].
In the non-fixed-design setting, we previously observed that [GKKW06] gives an O( Kd εlog d )
bound to solve (2) for linear spaces. The bound shown is actually somewhat stronger: by truncating
∗
2
, where f ∗ is the
the output, the condition number K can be replaced by K ′ = supx∈supp(D) fkf(x)
k2D
actual function being learned. That is, the sample complexity depends on the conditioning of the
actual function rather than the worst-case function in the family. It is a good question whether
one can get a similar result without the factor d loss.
As discussed earlier, [CDL13] shows that O(K log d) samples suffice for (2) in linear function
spaces. They also show for univariate polynomials that sampling from the Chebyshev measure
√1
leads to an O(d log d) sample complexity, and that this can be improved to O(d) via a
π 1−x2
similarly explicit non-independent sampling. Our distribution DF behaves analogously to the
Chebyshev measure on univariate polynomials, and extends the result to arbitrary linear function
spaces.
It is an interesting question whether the O(d log d) term we get for general linear spaces can also
be improved to O(d). The situation is somewhat similar to that of spectral sparsifiers of graphs:
O(n log n) sparsifiers follow from appropriately weighted sampling using matrix Chernoff [SS11],
while O(n) sparsifiers exist via more clever constructions [BSS12]. Perhaps a more clever choice of
queries can also avoid the log d we get from matrix Chernoff in this setting.
Organization: We outline the proofs of our main results in Section 2. We introduce notation and
tools in Section 3. Then we consider the number of samples for empirical norms of linear families
in Section 4. Next we study the query access model in Section 5 and active learning in Section 6.
We demonstrate lower bounds for query complexity and sample complexity in Section 7. Finally,
we provide an application to nonlinear function spaces in Section 8.
2
Proof Overview
Consider observations of the form y(x) = f (x) + g(x) for f in a (not necessarily linear) family F
and g an arbitrary, possibly random function.
Improved conditioning by better sampling. We start with the noiseless case of g = 0 in the
query access model, and consider the problem of estimating kyk2D = kf k2D with high probability.
If we sample points xi ∼ D ′ for some distribution D ′ , then we can estimate kf k2D as the empirical
4
norm
m
1 X D(xi )
|f (xi )|2
m
D ′ (xi )
(7)
i=1
which has the correct expectation. To show that this concentrates, we would like to apply Chernoff
bounds, which depend on the maximum value of summand. In particular, the concentration depends
on the reweighted condition number
D(x) |f (x)|2
′
·
KD = sup sup
.
′
kf k2D
x
f ∈F D (x)
We define DF to minimize this quantity, by making the inner term the same for every x. Namely,
we pick
#
"
|f (x)|2
1
|f (x)|2
DF (x) = D(x) · sup
for κ = E sup
.
(8)
2
x∼D f ∈F kf k2
κ
f ∈F kf kD
D
This shows that by sampling from DF rather than D,
number of our estimate (7)
" the condition
#
2
2
to κ = E sup |fkf(x)|
.
improves from K = sup
sup |fkf(x)|
k2
k2
x∼D
D
x∈supp(D) f ∈F
κ·log 1δ
Chernoff bound, O( ε2 )
f ∈F
D
From the
samples from DF let us estimate kf k2D to within accuracy
1 ± ε with probability 1 − δ for any fixed function f ∈ F. To be able to estimate every f ∈ F, a
basic solution would be to apply a union bound over an ε-net of F.
Linear function families F let us improve the result in two ways. First, we show that κ = d
for any dimension d linear function space. Second, we can replace the union bound by a matrix
d log
d
concentration bound, showing that O( ε2 δ ) samples from DF suffice to estimate kf k2D to within
1 ± ε for every f ∈ F with probability 1 − δ. In particular, O(d log d) samples suffice to estimate
every kf k2D to within constant factors, with 99.9% probability.
The above results appear in Section 4.
The effect of noise. Now we consider our actual problem, which is to estimate f from y = f + g
for nonzero noise g of E[kgk2D ] = σ 2 . Given samples xi ∼ DF , the empirical risk minimizer is the
function fe ∈ F closest to y under the empirical norm (7) for D ′ = DF . When F is a linear family,
the solution fe is a linear projection, so it acts on f and g independently. The previous discussion
showed that with O(d log d) samples, the empirical norm is a good estimator for every function in
F, which indicates the projection of the samples xi on f into the linear subspace F equals f . Hence
the error fe − f is the projection of g onto F under the empirical norm.
First, suppose that g is orthogonal to F under the true norm k · kD —for instance, if g(x) is
an independent mean-zero random variable for each x. In this case, the expected value of the
projection of g is zero. At the same time, we can bound the variance of the projection of a single
random sample of g drawn from DF by the condition number d · σ 2 , which indicates the projection
of m = O( dε ) independent samples will have squared norm less than ε · σ 2 . This gives the following
theorem, proven in Section 5:
Theorem 2.1. Consider any dimension d linear space F of functions from a domain G to C. Let
(D, Y ) be a joint distribution over G × C and f = arg min
E
[|y − h(x)|2 ]. There exists an
h∈F
(x,y)∼(D,Y )
queries and outputs f˜ ∈ F satisfying
efficient algorithm that makes m = O(d log d +
|y − f (x)|2 with probability ≥ 0.99.
E |f˜(x) − f (x)|2 ≤ ε ·
E
d
ε)
x∼D
(x,y)∼(D,Y )
5
For a noise function g not orthogonal to F in expectation, let g ⊥ and gk denote the decomposition of g where g⊥ is the orthogonal part and g k = g − g ⊥ ∈ F. The above theorem indicates
√
kfe−f kD ≤ kg k kD + ε·kg⊥ kD , which provides the desired (1+ε)kgkD -closeness via the Pythagorean
theorem. This result appears in Corollary 5.1 of Section 5.
Active learning. Next we consider the active learning setting, where we don’t know the distribution D and only receive samples xi ∼ D, but can choose which xi receive labels yi . Suppose the
function family F is a dimension-d linear space, and let K be the condition number (3). We give
an algorithm that uses m1 = O(K log d + Kε ) unlabeled samples and m = O(d log d + dε ) labeled
samples, and achieves the same guarantees as in the query access model.
As an intermediate step, suppose that the algorithm knew the distribution D. To simulate our
query-access algorithm, we would just need to simulate samples from DF , which we could do using
2
. To
rejection sampling by labeling each sample x with probability proportional to supf ∈F |fkf(x)|
2
kD
make the probabilities at most 1, the actual probability of labeling x would be this value divided
by the maximum it takes—which is K. The chance that a random sample xi ∼ D receives a label
would become
KDF
1
d
|f (x)|2
=
E sup
= .
K x∼D f ∈F kf k2D
K
K
Hence, after O(K log d + Kε ) unlabeled samples, we will get the O(d log d + dε ) labeled samples we
need to run the query-access algorithm. The only problem is that we don’t know D.
Fortunately, the only way in which the above algorithm uses D is to estimate kf kD for functions
f ∈ F. After at least O(K
log d) unlabeled samples, matrix concentration inequalities show that
1 Pm1
the empirical estimate m1 i=1 f (xi )2 is a constant-factor approximation to kf k2D for every f ∈ F.
This suffices to get the desired bound, giving the following theorem proven in Section 6:
Theorem 2.2. Consider any dimension d linear space F of functions from a domain G to C. Let
(D, Y ) be a joint distribution over G × C and f = arg min
E
[|y − h(x)|2 ].
h∈F
Let K =
O(K log d +
sup
supx∈G |h(x)|2
.
khk2D
h∈F :h6=0
K
ε ) unlabeled
For any ε > 0, there exists an efficient algorithm that takes
samples from D and requests O(d log d + dε ) labels to output fe satisfying
E [|f˜(x) − f (x)|2 ] ≤ ε ·
x∼D
(x,y)∼(D,Y )
E
[|y − f (x)|2 ] with probability ≥ 0.99.
(x,y)∼(D,Y )
Lower bounds. We first prove a lower bound on the query complexity using information theory. The Shannon-Hartley Theorem indicates that under the i.i.d. Gaussian noise N (0, 1/ε), for a
function f with |f (x)| ≤ 1 at every point x, any observation y(x) = f (x) + N (0, 1/ε) obtains O(ε)
information about f . Because the dimension of F is d, this indicates Ω(d/ε) queries is necessary
to recover a function in F.
Next, for any K, d, and ε we construct a distribution D and dimension-d linear family F with
condition number K over D, such that the sample complexity of achieving (2) is Ω(K log d + K/ε).
The first term comes from the coupon collector problem, and the second from the above query
bound. We summarize these results in Table 1.
Query complexity
Sample complexity
Lower bound
Ω(d/ε) in Theorem 7.1
Ω(K log d + K/ε) in Theorem 7.3
Upper bound
O(d log d + d/ε)
O(K log d + K/ε)
Table 1: Lower bounds and upper bounds in different models
6
Signals with k-sparse Fourier transform. We now consider the nonlinear family F of functions with k-sparse Fourier transform defined in (6), over the distribution D = [−1, 1]. As discussed
at (8), even for nonlinear function families, sampling from DF improves the condition number from
K to
|f (x)|2
,
κ = E sup
x∈D f ∈F kf k2
D
e 4)
effectively replacing the supx by Ex . Before we describe how to bound κ, let us revisit the O(k
bound for K shown in [CKPS16]. The key step—Claim 5.2 in [CKPS16]—showed
that for any
∆ > 0 and f ∈ F, f (x) can be expressed as a linear combination of f (x + j∆) | j = 1, . . . , l}
with constant coefficients and l = Õ(k2 ). This gives an upper bound on |f (−1)|2 in terms of
|f (−1 + ∆)|2 + · · · + |f (−1 + l · ∆)|2 , which bounds |f (−1)|2 /kf k2D by integrating ∆ from 0 to 2/l.
To give a better bound for κ, we show that f (x) can be expressed as a constant-coefficient
linear combination of the k elements of an arithmetic sequence on both sides of x, i.e., {f (x − k ·
e 2 ) elements required by [CKPS16]
∆), . . . , f (x + k · ∆)} \ f (x). This is much shorter than the O(k
for the one-sided version, and lets us show that
2
k log k
kf k2D
|f (x)|2 = O
1 − |x|
for any x ∈ (−1, 1). This leads to κ = O(k2 log2 k), which appears in Theorem 8.1 of Section 8.
We note in passing that these two bounds are analogous to the Markov Brothers’ inequality
and Bernstein inequality for polynomials: when F is the space of univariate degree-d polynomials,
the Markov Brothers’ inequality shows K = O(d2 ) and the Bernstein inequality shows κ = O(d).
3
Preliminaries
We use [k] to denote the subset {1, 2, . . . , k} and 1E ∈ {0, 1} to denote the indicator function of an
event E.
We first state the Chernoff bound for real numbers.
Lemma 3.1 (Chernoff Bound [Che52, Tar09] ). Let X1 , X2 , · · · , Xn be independent random variables. Assume that 0 ≤ Xi ≤ 1 always, for each i ∈ [n]. Let X = X1 + X2 + · · · + Xn and
n
P
µ = E[X] =
E[Xi ]. Then for any ε > 0,
i=1
Pr[X ≥ (1 + ε)µ] ≤ exp(−
ε2
ε2
µ) and Pr[X ≥ (1 − ε)µ] ≤ exp(− µ).
2+ε
2
Corollary 3.2. Let X1P
, X2 , · · · , Xn be independent random variables in [0, R] with expectation 1.
n
Xi
with expectation 1 satisfies
For any ε < 1/2, X = i=1
n
Pr[|X − 1| ≥ ε] ≤ 2 exp(−
ε2 n
· ).
3 R
P
k 1/k .
For a vector ~v = v(1), . . . , v(m) ∈ Cm , let k~v kk denote the ℓk norm, i.e.,
i∈[m] |v(i)|
Given a sequence S = (t1 , . . . , tm ) (allowing
Prepetition in S)2 and corresponding weights (w1 , . . . , wm ),
let kf k2S,w denote the weighted ℓ2 norm m
j=1 wj · |f (tj )| . For convenience, we omit w if it is a
1/2
uniform distribution on S, i.e., kf kS = Ei∈[m] |f (ti )|2
.
7
v k2
Given a matrix A ∈ Cm×m , let kAk denote the operator norm kAk = max~v6=~0 kA~
k~
v k2 and λ(A)
denote all eigenvalues of A. Given a self-adjoint matrix A, let λmin (A) and λmax (A) denote the
smallest eigenvalue and the largest eigenvalue of A.
We state the matrix Chernoff inequality from [Tro12].
Theorem 3.3 (Theorem 1.1 of [Tro12]). Consider a finite sequence {Xk } of independent, random,
self-adjoint matrices of dimension d. Assume that each random matrix satisfies
Xk 0
and
λ(Xk ) ≤ R.
P
P
Define µmin = λmin ( k E[Xk ]) and µmax = λmax ( k E[Xk ]). Then
(
)
µmin /R
X
e−δ
for δ ∈ [0, 1], and
Pr λmin (
Xk ) ≤ (1 − δ)µmin ≤ d
(1 − δ)1−δ
k
)
(
µmax /R
X
e−δ
Pr λmax (
Xk ) ≥ (1 + δ)µmax ≤ d
for δ ≥ 0
(1 + δ)1+δ
(9)
(10)
k
We state the Shannon-Hartley theorem in information theory.
Theorem 3.4 (The Shannon-Hartley Theorem [Har28, Sha49]). Let S be a real-valued random
2
variable with E[S 2 ] = τ 2 and T ∼ N (0, σ 2 ). The mutual information I(S; S + T ) ≤ 12 log(1 + στ 2 ).
Given a function f with domain G and a distribution D over G, we use kf kD to denote the
1/2
expected ℓ2 norm of f (x) where x ∼ D, i.e., kf kD = E [|f (x)|2 ]
.
x∼D
Weights between different distributions. Given a distribution D, to estimate khk2D of a
function h through random samples from D ′ , we use the following notation to denote the reweighting of h between D ′ and D.
Definitionq
3.5. For any distribution D ′ over the domain G and any function h : G → C, we define
′
D(x)
h(D ) (x) = D
′ (x) · h(x).
For any function h ∈ F, from the definition, E
x∼D
h
i
h
i
D(x)
(D ′ ) (x)|2 = E
2 = E |h(x)|2 .
|h
|h(x)|
′ (x)
D
′
′
x∼D
x∼D
Let S = (t1 , . . . , tm ) be m random samples from the distribution D ′ over G. We always choose
wj =
D(tj )
for each tj
m · D ′ (tj )
(11)
to keep the expectation ES [khk2S,w ] the same as khk2D :
E
t1 ,...,tm ∼D ′
4
m
X
j=1
2
wj · |h(tj )|
m
|h(tj )|2 · D(tj )
1 X
=
E ′
= E |h(x)|2 .
′
x∼D
tj ∼D
m
D (tj )
j=1
Condition Number Under Reweighted Sampling
The main result in this section is a distribution for sample-efficient ℓ2 estimation of functions in a
given linear family F of finite dimension.
8
Theorem 4.1. Let F be any linear family of functions" with dimension #
d from the domain G
to C and D be any distribution over G. For κ = E
x∼D
sup
sup
|h(x)|2
and the distribution
h∈F :khkD =1
|h(x)|2
h∈F :khkD =1
, a sequence S of m = O( εd2 log dδ ) random samples (x1 , . . . , xm )
DF (x) = D(x) ·
κ
from DF with corresponding weights (w1 , . . . , wm ) guarantees, with probability at least 1 − δ,
∀h ∈ F,
khk2S,w = [1 − ε, 1 + ε] · khk2D .
(12)
For generality, we prove the sample complexity from an arbitrary distribution D ′ to estimate
the ℓ2 norm of functions in F under D. We bound the number of samples by the condition number
KD′ of D ′ , i.e.,
(
)
|h(x)|2 D(x)
.
KD′ = sup sup
·
khk2D D ′ (x)
x∈G h6=0
Lemma 4.2. Let F be any linear family of dimension d from the domain
G to C′ and D
be
(D ) (x)|2
|h
any distribution over G. Let D ′ be a distribution over G and KD′ = sup sup
. A
khk2
x∈G
K
O( εD2 ′
d
δ)
sequence S of m =
random samples (x1 , . . . , xm ) from
log
(w1 , . . . , wm ) guarantees that with probability at least 1 − δ,
∀h ∈ F,
D′
h6=0
D
with corresponding weights
khk2S,w = [1 − ε, 1 + ε] · khk2D .
We prove
κ= E
x∼D
"
|h(x)|2
sup
2
h∈F :h6=0 khkD
#
(13)
=d
in Section 4.1 such that Theorem 4.1 follows Lemma 4.2 with the condition number KDF = κ = d.
We focus on the proof of Lemma 4.2 in the rest of this section. The condition number KDF = d
d
indicates O(d2 log εδ
/ε2 ) random samples from DF with corresponding weights could guarantee
Property (12) from a union bound over an ε-net of F. We improve this by applying the matrix
Chernoff bound to get that O(d log dδ /ε2 ) samples suffice.
We first restate the guarantee of khk2S,w = (1 ± ε)khk2D in (13) as bounding the eigenvalues
of a matrix determined by S = (x1 , . . . , xm ) and corresponding weights. Let {v1 , . . . , vd } be any
orthonormal basis of F, where inner products are taken under the distribution D, i.e., E [vi (x) ·
x∼D
vj (x)] = 1i=j for any i, j ∈ [d]. In this work, we always use A to denote the following matrix.
Definition 4.3. Given the sample points S = (x1 , . . . , xm ) and their weights (w1 , . . . , wm ), we
define a matrix A ∈ Cm×d based on S and w as
√
(14)
A(i, j) = wi · vj (xi ) for i ∈ [m], j ∈ [d].
Let A∗ denote the conjugate transpose of A, i.e., A∗ (j, i) = A(i, j). We prove Property (13) is
equivalent to bounding the eigenvalues of A∗ · A.
Lemma 4.4. For any ε > 0, given S = (x1 , . . . , xm ) and their weights (w1 , . . . , wm ), let A be the
matrix defined in (14). Then
khk2S,w = (1 ± ε)khk2D
for every h ∈ F
if and only if the eigenvalues of A∗ A are in [1 − ε, 1 + ε].
9
Next we bound the eigenvalues of A∗ A by a matrix Chernoff bound.
Lemma 4.5. Let D ′ be an arbitrary distribution over G and
′
KD′
|h(D ) (x)|2
.
= sup sup
khk2D
h∈F :h6=0 x∈G
(15)
There exists an absolute constant C such that for any n ∈ N+ , d ∈ N+ , ε ∈ (0, 1) and δ ∈ (0, 1),
when we sample S = (x1 , . . . , xm ) independently from the distribution D ′ with m = εC2 · KD′ · log dδ
and choose wj =
D(xj )
m·D ′ (xj )
for each j ∈ [m], the matrix A defined in (14) satisfies
kA∗ A − Ik ≤ ε with probability at least 1 − δ.
We finish the proof of Lemma 4.2 here and defer the proofs of Lemma 4.4 and Lemma 4.5 to
Section 4.2.
Proof of Lemma 4.2.
We define A using (14) from the sample points in S and their weights
(w1 , . . . , wm ). We first apply Lemma 4.5 to bound the eigenvalues of A∗ A, i.e., λ(A∗ A) ∈ [1−ε, 1+ε]
with probability 1 − δ.
From Lemma 4.4, λ(A∗ A) ∈ [1 − ε, 1 + ε] indicates the property (13) for any h ∈ F.
4.1
The Condition Number of DF
We describe the distribution DF with KDF = d that provides an almost optimal sample complexity.
We first observe that the condition number of DF is always κ for any family F (not necessarily
linear).
Claim 4.6. For any family F and any distribution D on
its domain, let DF be the distribution
2
· |fkf(x)|
is at most κ.
defined in (8) with κ. The condition number KDF = sup sup DD(x)
k2
F (x)
x
f ∈F
D
Proof. For any g ∈ F and x in the domain G,
2
|g(x)|
· D(x)
|g(x)|2 D(x)
kgk2D
·
=
≤ κ.
2
kgk2D DF (x)
supf ∈F |fkf(x)|
· D(x)/κ
k2
D
Next we use the linearity of F to prove KDF = d. Let {v1 , . . . , vd } be any orthonormal basis of
F, where inner products are taken under the distribution D.
Lemma 4.7. For any linear family F of dimension d,
E
sup
x∼D h∈F :khkD =1
such that DF (x) = D(x) ·
sup
|h(x)|2 = d
|h(x)|2 /d has a condition number KDF = d. Moreover, there
h∈F :khkD =1
exists an efficient algorithm to sample x from DF and compute its weight
10
D(x)
DF (x) .
Algorithm 1 SampleDF
procedure GeneratingDF(F = span{v1 , . . . , vd }, D)
2:
Sample j ∈ [d] uniformly.
3:
Sample x from the distribution Wj (x) = |vj (x)|2 .
.
4:
Set the weight of x to be Pd d
2
1:
i=1
end procedure
5:
|vi (x)|
Proof. Given an orthonormal basis v1 , . . . , vd of F, for any h ∈ F with khkD = 1, there exists
α1 , . . . , αd such that h(x) = αi · vi (x). Then for any x in the domain, from the Cauchy-Schwartz
inequality,
P
P
P
X
| i∈[d] αi vi (x)|2
( i∈[d] |αi |2 ) · ( i∈[d] |vi (x)|2 )
|v(x)|2
P
P
sup
=
sup
=
=
|vi (x)|2 .
2
2
|α
|
|α
|
kvk2D
v
α1 ,...,αd
i
i
i∈[d]
i∈[d]
i∈[d]
P
This is tight because there always exist α1 = v1 (x), α2 = v2 (x), . . . , αd = vd (x) such that |
αi vi (x)|2 =
i∈[d]
P
P
(
|αi |2 ) · (
|vi (x)|2 ). Hence
i∈[d]
i∈[d]
X
|h(x)|2
|vi (x)|2 = d.
= E
2
x∼D
x∼D h∈F :h6=0 khkD
E
sup
i∈[d]
From the above calculation,
DF (x) =
Let
h(DF ) (x)
denote
and Claim 4.6,
q
D(x)
DF (x)
D(x) · sup |h(x)|2
· h(x) =
khkD =1
d
r
P d
|vi (x)|2
i∈[d]
=
D(x) ·
P
2
i∈[d] |vi (x)|
d
· h(x) for any h ∈ F. From the definition of DF
KDF = sup sup |h(DF ) (x)|2 = sup sup |h(x)|2 ·
x khkD =1
.
x khkD =1
D(x)
= d.
DF (x)
We present our sampling procedure in Algorithm 1.
4.2
Proofs of Lemma 4.4 and Lemma 4.5
We first prove Lemma 4.4 then prove Lemma 4.5.
P
Proof of Lemma 4.4.
For any function h = i∈[d] αi vi , let α(h) ∈ Cd denote the vector of the
coefficients (α1 , . . . , αd ) such that kα(h)k2 = khkD .
We first prove the ⇐ direction. When all eigenvalues of (A∗ · A) are in [1 − ε, 1 + ε],
kA · α(h)k22 = α(h)∗ · (A∗ · A) · α(h) ∈ [1 − ε, 1 + ε] · kα(h)k22 = [1 − ε, 1 + ε] · khk2D .
P
2
2
From our definitions of A and α(h), khk2S,w = m
j=1 wj |h(xj )| = kA · α(h)k2 .
11
Now we prove the other direction. If there exists an eigenvalue of (A∗ · A) not in [1 − ε, 1 + ε],
there exists a vector α(h) such that
α(h)∗ · (A∗ · A) · α(h) ∈
/ [1 − ε, 1 + ε]kα(h)k22 .
P
2
2
2
2
From the above discussion, α(h)∗ · (A∗ · A) · α(h) = m
j=1 wj |h(xj )| = khkS,w and kα(h)k2 = khkD ,
which indicates khk2S,w = (1 ± ε) E |h(x)|2 does not hold for some function in F.
x∼D
We apply the matrix Chernoff bound of Theorem 3.3 to bound the eigenvalues of A∗ A.
Proof of Lemma 4.5.
Let v1 , · · · , vd be the orthonormal basis of F in the definition of matrix
A. For any h ∈ F, let α(h) = (α1 , · · · , αd ) denote the coefficients of h under v1 , · · · , vd such
P
′
(D ′ )
that khk2D = kα(h)k22 . At the same time, for any fixed x, |h(D ) (x)|2 = | di=1 αi hi (x)|2 ≤
P
(D ′ )
kα(h)k22 · i∈[d] |vi (x)|2 . Thus
def
KD′ = sup
x∈G
′
|h(D ) (x)|2
khk2D
h∈F :h6=0
For each point xj in S with weight wj =
a vector in
Cd
indicates
sup
sup
x∈G
D(xj )
m·D ′ (xj ) ,
X
i∈[d]
(D ′ )
|vi
(x)|2 ≤ KD′ .
(16)
let Aj denote the jth row of the matrix A. It is
defined by
Aj (i) = A(j, i) =
√
(D ′ )
wj · vi (xj ) =
vi
√
(xj )
.
m
P
∗
So A∗ A = m
j=1 Aj · Aj .
For A∗j · Aj , it is always 0. Notice that the only non-zero eigenvalue of A∗j · Aj is
X
′
KD′
1
(D )
|vi (xj )|2 ≤
λ(A∗j · Aj ) = Aj · A∗j =
m
m
i∈[d]
from (16).
P
∗
At the same time, m
j=1 E[Aj ·Aj ] equals the identity matrix of size d×d because the expectation
of the entry (i, i′ ) in A∗j · Aj is
(D ′ )
′
E ′ [A(j, i) · A(j, i )] =
xj ∼D
E ′[
xj ∼D
vi
(D ′ )
(xj ) · vi′
m
(xj )
]
vi (xj ) · vi′ (xj )
D(x) · vi (xj ) · vi′ (xj )
]= E [
] = 1~i=~i′ /m.
xj ∼D
m · D ′ (xj )
m
P
∗
Now we apply Theorem 3.3 on A∗ A = m
j=1 (Aj · Aj ):
=
E [
xj ∼D ′
Pr [λ(A∗ A) ∈
/ [1 − ε, 1 + ε]] ≤ d
e−ε
(1 − ε)1−ε
−
≤ 2d · e
ε2 · m
K ′
D
3
≤δ
1/ KD′
m
given m ≥
12
+d
e−ε
(1 + ε)1+ε
6KD′ log dδ
.
ε2
1/ KD′
m
5
Results for the Query Access Model
We prove Theorem 2.1 and a corollary in this section.
Theorem 2.1. Consider any dimension d linear space F of functions from a domain G to C. Let
(D, Y ) be a joint distribution over G × C and f = arg min
E
[|y − h(x)|2 ]. There exists an
h∈F
efficient algorithm that makes m = O(d log d +
E |f˜(x) − f (x)|2 ≤ ε ·
x∼D
E
d
ε)
(x,y)∼(D,Y )
(x,y)∼(D,Y )
queries and outputs f˜ ∈ F satisfying
|y − f (x)|2 with probability ≥ 0.99.
We now give a corollary of Theorem 2.1 for specific kinds of noise. In the first case, we consider
noise functions representing independently mean-zero noise at each position x such as i.i.d. Gaussian
noise. Second, we consider arbitrary noise functions on G.
Corollary 5.1. Consider any dimension d linear space F of functions from a domain G to C
and distribution D over G. Let y(x) = f (x) + g(x) be our observed function, where f ∈ F and g
denotes a noise function. For any ε > 0, there exists an efficient algorithm that observes y(x) at
m = O(d log d + dε ) points and outputs fe such that with probability 0.99,
1. kf˜ − f k2D ≤ ε · E[kgk2D ], when g(x) is a random function from G to C where each g(x) is an
g
independent random variable with E[g(x)] = 0.
g
2. kf˜ − f kD ≤ (1 + ε) · kgkD when g is a deterministic function.
For generality, we prove a lemma that applies to resampling according to an arbitrary distributions D ′ on G. For any x ∈ G, let Y (x) denote the conditional distribution (D, Y |D = x) and
(D ′ , Y (D ′ )) denote the distribution that first generates x ∼ D ′ then generates y ∼ Y (x).
Lemma 5.2. Consider any dimension d linear space F of functions from a domain G to C. Let
(D, Y ) be a joint distribution over G × C, and f = arg min
E
[|y − h(x)|2 ].
h∈F
Let
D′
be any distribution on G, and define KD′ =
(x,y)∼(D,Y )
sup
(D ′ )
sup |h khk(x)|
2
h∈F :h6=0 x∈G
2
. There exists an
D
K
efficient algorithm that takes m = O(KD′ log d + εD′ ) samples from (D ′ , Y (D ′ )) and outputs f˜ ∈ F
satisfying
E |f˜(x) − f (x)|2 ≤ ε ·
|y − f (x)|2 with probability ≥ 0.99.
E
x∼D
(x,y)∼(D,Y )
Theorem 2.1 follows from Lemma 5.2 with the distribution DF in Lemma 4.7 that has a condition number KDF = d.
We introduce several notations in our algorithm. Let v1 , . . . , vd be an orthonormal basis of F
where the inner product is taken under the marginal distribution
P D. For any function h ∈ F, let
α(h) denote the coefficients (α(h)1 , . . . , α(h)d ) such that h = di=1 α(h)i · vi .
The coefficients (α(f )1 , . . . , α(f )d ) of the function f = arg min
E
[|y − h(x)|2 ] satisfy
h∈F
αi (f ) =
E
(x,y)∼(D,Y )
(x,y)∼(D,Y )
[y · vi (x)]. This indicates
E
[ y − f (x) · vi (x)] = 0,
(x,y)∼(D,Y )
13
∀i ∈ [d].
(17)
Given S = (x1 , . . . , xm ) and their weights (w1 , . . . , wm ), let A ∈ Cm×d denote the matrix in
√
Definition
4.3, i.e., A(i,
j) = wi ·vj (xi ) for i ∈ [m] and j ∈ [d]. Because the number of samples m =
O KD′ log d +
KD ′
ε
is large enough, we assume its eigenvalues satisfy λ(A∗ A) ∈ [1 − 1/4, 1 + 1/4],
which happens with probability 1 − 10−4 from Lemma 4.5.
√
√
Let ~yw denote ( w1 · y1 , . . . , wm · ym ). From the definition of A and α(h),
√
√
A · α(h) =
w1 · h(x1 ), . . . , wm · h(xm ) .
Hence the empirical distance kh − (y1 , . . . , ym )kS,w equals kA · α(h) − ~yw k2 for any h ∈ F. The
function fe minimizing kh − (y1 , . . . , ym )kS,w = min kA · α(h) − ~yw k2 is the pseudoinverse of A on
α(h)∈Cd
~yw , i.e., α(fe) = (A∗ · A)−1 · A∗ · ~yw .
We present our algorithm in Algorithm 2. We defer the proofs of Lemma 5.2 and Claim 5.4 to
Section 5.1 and the proof of Corollary 5.1 to Section 5.2.
Algorithm 2 Learning any linear subspace
procedure SubspaceLearning(ε, F, D, D ′ )
2:
Let {v1 , . . . , vd } be an orthogonal basis of F in D, i.e., E [vi (x) · vj (x)] = 1i=j .
1:
3:
4:
5:
6:
7:
8:
9:
10:
m ← C · KD′ · (log d + 1/ε) for some constant C.
Sample S = {x1 , . . . , xm } independently from D ′ .
D(xj )
for each j ∈ [m].
Set wj = m·D′ (x
j)
Sample yi = Y (xi ) independently for each xi .
√
A(i, j) ← wi · vj (xi ) for i ∈ [m] and j ∈ [d].
√
~yw (i) ← wi · yi for i ∈ [m].
α(f˜) ← (A∗ · P
A)−1 · A∗ · ~yw .
return f˜ =
α(f˜)i · vi .
x∼D
i∈[d]
11:
end procedure
Remark 5.3. The running time of our algorithm is the time to obtain an orthonormal basis in
F and compute the inverse (A∗ A)−1 . Given a basis e1 , . . . , ed of F with an procedure to compute
Ex∼D [ei (x) · ej (x)] in time T , we could obtain an orthonormal basis of F in time O(d3 T ) by GramSchmidt orthogonalization. The running time of computing (A∗ A)−1 is O(dω ) where ω is the matrix
multiplication constant.
5.1
Correctness
For any function h : G → C, let ~hS denote the vector (h(x1 ), . . . , h(xm )) and ~hS,w denote ~h under
the weights (w1 , . . . , wm ):
√
√
~hS,w def
w1 · h(x1 ), . . . , wm · h(xm ) .
=
(18)
We first prove the following lemma to bound k(A∗ · A)−1 · A∗ · ~yw − α(f )k2 = k(A∗ · A)−1 · A∗ · (~yw −
f~S,w )k2 .
Claim 5.4. Let f = arg min
h∈F
E
(x,y)∼(D,Y )
E
(x1 ,y1 ),...,(xm ,ym )
h
[|y − h(x)|2 ]. Then
A∗ · (~yw − f~S,w )
2
2
i
14
≤
KD′
·
E
[|y − f (x)|2 ].
m (x,y)∼(D,Y )
Proof. Recall that for any function f on G,
s
D(x)
′
f (D ) (x) =
f (x)
D ′ (x)
and
wj =
D(xj )
m · D ′ (xj )
for each xj ∈ S. For (x1 , y1 ), . . . , (xm , ym ) independently randomly generated from (D ′ , Y (D ′ )), we
first write
X X √
√
2
wj · vi (xj ) · wj · (f (xj ) − yj )
E[kA∗ · (f~S,w − ~
yw )k22 ] = E
i∈[d] j∈[m]
X
=
i∈[d]
E
X
j∈[m]
2
wj · vi (xj ) · (f (xj ) − yj ) =
X X
i∈[d] j∈[m]
i
h
2
E wj · vi (xj ) · (f (xj ) − yj ) ,
where the last step uses the property (17)
E
xj ∼D ′ ,yj ∼Y (xj )
1
wj · vi (xj ) · (f (xj ) − y) =
vi (xj ) · (f (xj ) − yj ) = 0
E
m (xj ,yj )∼(D,Y )
and the independence between different j ∈ [m] to eliminate the cross terms. Then we further
simplify it by swapping the summations of i and j:
i
h
X X
2
2
2
E
wj · |vi (xj )| · |f (xj ) − yj |
(19)
xj ∼D ′ ,yj ∼Y (D ′ )
i∈[d] j∈[m]
=
X
j∈[m]
≤
X
j∈[m]
E
xj ∼D ′ ,yj ∼Y (D ′ )
xj
X
max
∈[−1,1]n
Notice that
i
X
i∈[d]
wj · |vi (xj )|2 · wj · |f (xj ) − yj |2
wj · |vi (xj )|2
!
·
E
xj
∼D ′ ,y
j ∼Y
(D ′ )
wj · |f (xj ) − yj |2 .
(20)
(21)
1
wj · |f (xj ) − yj |2 =
|f (xj ) − yj |2 .
E
m (x,y)∼(D,Y )
xj
j ∼Y
P
wj · |vi (xj )|2 using the definition of wj :
On the other hand, we bound
E
∼D ′ ,y
(D ′ )
i∈[d]
X
i∈[d]
D(xj )
1 X (D′ )
|vi (xj )|2 ≤ KD′ /m
· |vi (xj )|2 =
′
m · D (xj )
m
i∈[d]
where we use (16) in the last step.
From all discussion above, we bound
∗
E[kA
S
· (fS,w − ~yw )k22 ]
≤
m
X
KD′
j=1
E
m2 (x,y)∼(D,Y )
KD′
|f (xj ) − yj |2 .
|f (xj ) − yj |2 ≤
E
m (x,y)∼(D,Y )
15
Next we prove Lemma 5.2.
Proof of Lemma 5.2.
Let m = C(KD′ log d + KD′ /ε) for a large constant C. From Lemma 4.5,
∗
λ(A A) ∈ [1 − 1/4, 1 + 1/4] with probability 0.999.
P
Because v1 , . . . , vd constitute an orthonormal basis of F, for fe = ki=1 α(fe)i vi ,
E [|fe(x) − f (x)|2 ] = kα(fe) − α(f )k22 = k(A∗ · A)−1 · A∗ · ~yw − α(f )k22 .
x∼D
Because (A∗ · A)−1 · A∗ · f~w = α(f ),
k(A∗ · A)−1 · A∗ · ~yw − α(f )k22 = k(A∗ · A)−1 · A∗ · (~yw − f~w )k22
≤ k(A∗ · A)−1 k2 · kA∗ · (~yw − f~w )k22 ≤ (
|f (xj ) − yj |2 . From the Markov inequality,
(x,y)∼(D,Y )
∗
with probability at least 0.991, kA · (~yw − f~w )k22 ≤ 2ε ·
|f (xj ) − yj |2 .
E
(x,y)∼(D,Y )
2
e
|f (xj ) − yj |2 with probability
From all discussion above, Ex∼D [|f (x) − f (x)| ] ≤ ε ·
E
From Claim 5.4, E[kA∗ · (~yw − f~w )k22 ] ≤
KD ′
m
E
1
)2 · kA∗ · (~yw − f~w )k22 .
1 − 1/4
(x,y)∼(D,Y )
0.99.
5.2
Proof of Corollary 5.1
In this section, we finish the proof of Corollary 5.1.
Proof of Corollary 5.1.
Let DF be the distribution of the linear subspace F under D with a
condition number KDF = d from Lemma 4.7. We apply Procedure SubspaceLearning with
ε, F, D, DF to obtain fe.
For the first part, let (D, Y ) = D, f (x) + g(x) be our joint distribution of (x, y). Because the
expectation E[g(x)] = 0 for every x ∈ G, arg min
E
[|y − v(x)|2 ] = f . From Lemma 5.2, for
(x,y)∼(D,Y )
v∈V
α(fe) = (A∗ · A)−1 · A∗ · ~yw and m = C(d log d + dε ) with a large constant C,
kfe − f k2D = kα(fe) − α(f )k22 ≤ ε ·
E
(x,y)∼(D,Y )
[|y − f (x)|2 ] = ε · E[kgk2D ], with probability 0.99.
For the second part, let g k be the projection of g(x) to F and g ⊥ = g − gk be orthogonal
to F. Let α(gk ) denote the coefficients of gk in the fixed orthonormal basis (v1 , . . . , vd ) so that
kα(gk )k2 = kg k kD . We decompose ~yw = f~w + ~gw = f~w + g~k w + g~⊥ w . Therefore
α(fe) = (A∗ A)−1 · A∗ · (f~w + g~k w + g~⊥ w ) = α(f ) + α(gk ) + (AA∗ )−1 A∗ · g~⊥ w .
The distance kfe − f kD = kα(fe) − α(f )k2 equals
k(A∗ A)−1 ·A∗ ·~yw −α(f )k2 = kα(f )+α(gk )+(A∗ A)−1 ·A∗ ·g~⊥ w −α(f )k2 = kα(gk )+(A∗ A)−1 ·A∗ ·g~⊥ w k2 .
√
From Lemma 5.2, with probability 0.99, k(A∗ A)−1 · A∗ · g~⊥ w k2 ≤ ε · kg ⊥ kD . Thus
k(A∗ A)−1 · A∗ · ~
yw − α(f )k2 = kα(gk ) + (A∗ A)−1 · A∗ · g~⊥ w k2
√
≤ kgk kD + ε · kg⊥ kD .
16
Let 1 − β denote kgk kD /kgkD such that kg ⊥ kD /kgkD =
p
2β − β 2 . We rewrite it as
r
p
√ p
√ p
ε 2 ε
2
) +
1 − β + ε · 2β − β kgkD ≤ (1 − β + ε · 2β)kgkD ≤ 1 − ( β −
kgkD .
2
2
From all discussion above, kfe− f kD = kα(fe) − α(f )k2 = k(A∗ A)−1 · A∗ · ~yw − α(f )k2 ≤ (1 + ε)kgkD .
6
Results for Active Learning
In this section, we investigate the case where we do not know the distribution D of x and only
receive random samples from D. We finish the proof of Theorem 2.2 for linear families that bounds
the number of unlabeled samples by the condition number of D and the number of labeled samples
by dim(F) to find the truth through D.
Theorem 2.2. Consider any dimension d linear space F of functions from a domain G to C. Let
(D, Y ) be a joint distribution over G × C and f = arg min
E
[|y − h(x)|2 ].
h∈F
Let K =
O(K log d +
sup
supx∈G |h(x)|2
.
khk2D
h∈F :h6=0
K
ε ) unlabeled
For any ε > 0, there exists an efficient algorithm that takes
samples from D and requests O(d log d + dε ) labels to output fe satisfying
E [|f˜(x) − f (x)|2 ] ≤ ε ·
x∼D
(x,y)∼(D,Y )
E
[|y − f (x)|2 ] with probability ≥ 0.99.
(x,y)∼(D,Y )
We present our algorithm in Algorithm 3. It first takes m0 = O(K log d + K/ε) unlabeled
samples and defines a distribution D0 to be the uniform distribution on these m0 samples. Then
we use D0 to simulate D such that we could apply Procedure SubspaceLearning in Algorithm 2
to require labels.
Algorithm 3 Regression over an unknown distribution D
1:
2:
3:
4:
5:
6:
7:
8:
procedure RegressionUnknownDistribution(ε, F, D)
Set C to be a large constant and m0 = C · (K log d + K/ε) .
Take m0 unlabeled samples x1 , . . . , xm0 from D.
Let D0 be the uniform distribution over (x1 , . . . , xm0 ).
Let v1 , . . . , vdPbe an orthonormal basis of F under D0 .
Set D1 (x) = di=1 |vi (x)|2 /d to be the sample efficient distribution of F on D0 with KD1 = d.
Apply Procedure SubspaceLearning in Algorithm 2 with parameters ε, F, D0 , D1 .
end procedure
Proof. We still use kf kD to denote
q
E [|f (x)|2 ] and kf kD0 to denote
x∼D
From Lemma 4.5, with probability at least 1 − 10−3 ,
khkD0 = (1 ± 1/4) · khkD for every h ∈ F.
q
E [|f (x)|2 ] separately.
x∼D0
(22)
Let yi denote a random label of xi from Y (xi ) for each i ∈ [m0 ] including the unlabeled samples in
the algorithm and the labeled samples in the 2nd stage of applying Procedure SubspaceLearning.
Let
|yi − h(xi )|2 .
(23)
f ′ = arg min
E
h∈F
xi ∼D0 ,yi ∼Y (xi )
17
From Claim 5.4,
[kf ′ − f k2D ] ≤
E
y1 ,...,ym0
K
m0
·
E
(x,y)∼(D,Y )
[|y − f (x)|2 ]. Let the constant in m0 be large
enough such that from the Markov inequality, with probability 1 − 4 · 10−3 ,
kf ′ − f k2D ≤ ε ·
[|y − f (x)|2 ].
E
(x,y)∼(D,Y )
For the uniform distribution of (D0 , Y ′ ) over (x1 , y1 ), . . . , (xm0 , ym0 ), let S denote the m0 unlabeled samples {x1 , . . . , xm0 }. Let D1 be the sample efficient distribution of F under D0 with condition number d and S ′ denote the m = O( dε ) labeled samples generated by D1 . From Lemma 5.2 and
the definition of f ′ in (23), when S ′ is generated by D1 , the empirical minimizer fe of D1 satisfies
m0
X
d
d
1
d
2
E′ [kfe − f ′ k2D0 ] ≤
· kfe − f ′ k2D0 =
·
· E |yi − f ′ (xi )|2 .
yi − f (xi ) =
S
m
m
m0
m xi ∼S
j=1
Notice that
E
E
(x1 ,y1 ),...,(xm0 ,ym0 ) xi ∼S
≤
E
(x1 ,y1 ),...,(xm0 ,ym0 )
≤2
Hence
E
E
(x,y)∼(D,Y )
2 E
|yi − f ′ (xi )|2
xi ∼S
2
|yi − f (xi )|
|y − f (x)|2 + 3
E′ [kfe − f ′ k2D0 ] ≤
(x1 ,y1 ),...,(xm0 ,ym0 ) S
+ 2kf −
E
(x1 ,y1 ),...,(xm0 ,ym0 )
d
m
·3
E
(x,y)∼(D,Y )
f ′ k2D0
kf − f ′ k2D
|y − f (x)|2 .
from (22)
Let the constant in m be large enough such that from the Markov inequality, with probability
1 − 4 · 10−3 over (x1 , y1 ), . . . , (xm0 , ym0 ) and S ′ ,
|y − f (x)|2 .
kfe − f ′ k2D0 ≤ ε ·
E
(x,y)∼(D,Y )
From all discussion above, we have
kfe−f k2D ≤ 2kfe−f ′ k2D +2kf ′ −f k2D ≤ 2ε·
E
(x,y)∼(D,Y )
[|y−f (x)|2 ]+3kf ′ −f k2D0 ≤ 5ε·
E
[|y−f (x)|2 ]
(x,y)∼(D,Y )
such that rescaling ε gives the desired result.
Corollary 6.1. Let F be a family of functions from a domain G to C with dimension d and D be a
supx∈G |h(x)|2
distribution over G with bounded condition number K = sup Ex∼D
. Let y(x) = f (x)+g(x)
[|h(x)|2 ]
h∈F :h6=0
be our observation with f ∈ F.
For any ε > 0, there exists an efficient algorithm that takes O(K log d + Kε ) unlabeled samples
from D and require O(d log d + dε ) labels to output fe such that with probability 0.99,
1. kfe − f k2D ≤ ε · E[kgk2D ], when g is a random function where each g(x) is an independent
random variable with E[g(x)] = 0.
2. kfe − f kD ≤ (1 + ε) · kgkD otherwise.
18
Proof. This proof is essentially identical to that of Corollary 5.1. When g is a random function where
each g(x) is an independent random variable with E[g(x)] = 0, let the joint distribution (D, Y )
be x ∼ D and y = f (x) + g(x) such that f = arg min
E
[|y − h(x)|2 ]. From Theorem 2.2,
h∈F
(x,y)∼(D,Y )
Algorithm 3 outputs fe satisfying kfe − f k2D ≤ ε · E[kgk2D ] with probability 0.99.
For the second part, let g k be the projection of E[g] onto F under D and g⊥ = g − gk be the
√
orthogonal part. From the first part, fe satisfies kfe − f − g k kD ≤ ε · kgkD . Thus
kfe − f kD ≤ kg k kD +
√
ε · kgkD .
e
Similar to the proof of Corollary 5.1, we prove
pkf − f kD ≤ (1 + ε) · kgkD as follows. Let 1 − β
k
⊥
denote kg kD /kgkD such that kg kD /kgkD = 2β − β 2 . We rewrite it as
r
p
p
√ p
ε 2 ε
2
1 − β + ε · 2β − β kgkD ≤ (1 − β + ε 2β)kgkD ≤ 1 − ( β −
) +
kgkD .
2
2
7
Lower Bounds
We present two lower bounds on the number of samples in this section. We first prove a lower
bound for query complexity based on the dimension d. Then we prove a lower bound for the
sample complexity based on the condition number of the sampling distribution.
1
Theorem 7.1. For any d and any ε < 10
, there exist a distribution D and a linear family F of
functions with dimension d such that for the i.i.d. Gaussian noise g(x) = N (0, 1ε ), any algorithm
which observes y(x) = f (x)+g(x) for f ∈ F with kf kD = 1 and outputs fe satisfying kf − fekD ≤ 0.1
with probability ≥ 34 , needs at least m ≥ 0.8d
ε queries.
Notice that this lower bound almost matches the upper bound in Theorem 2.1. In the rest of
this section, we focus on the proof of Theorem 7.1. Let F = {f : [d] → R} and D be the uniform
distribution over [d]. We first construct a packing set M of F.
Claim 7.2. There exists a subset M = {f1 , . . . , fn } ⊆ F with the following properties:
1. kfi kD = 1 for each fi ∈ M.
2. kfi k∞ ≤ 1 for each fi ∈ M.
3. kfi − fj kD > 0.2 for distinct fi , fj in M.
4. n ≥ 20.7d .
Proof. We construct M from U = f : [d] → {±1} in Procedure ConstructM. Notice that
|U | = 2d before the while loop. At the same time, Procedure ConstructM removes at most
d
0.3d functions every time because kg − hk < 0.2 indicates Pr[g(x) 6= h(x)] ≤ (0.2)2 /4 =
D
≤0.01d ≤ 2
0.01. Thus n ≥ 2d /20.3d ≥ 20.7d .
19
Algorithm 4 Construct M
1:
2:
3:
4:
5:
6:
7:
8:
procedure ConstructM(d)
Set n = 0 and U = f : [d] → {±1} .
while U 6= ∅ do
Choose any h ∈ U and remove all functions h′ ∈ U with kh − h′ kD < 0.2.
n = n + 1 and fn = h.
end while
Return M = {f1 , . . . , fn }.
end procedure
We finish the proof of Theorem 7.1 using the Shannon-Hartley theorem.
Proof of Theorem 7.1.
Because of Yao’s minimax principle, we assume A is a deterministic
algorithm given the i.i.d. Gaussian noise. Let I(fe; fj ) denote the mutual information of a random
function fj ∈ M and A’s output fe given m observations (x1 , y1 ), . . . , (xm , ym ) with yi = fj (xi ) +
N (0, 1ε ). When the output fe satisfies kfe− fj kD ≤ 0.1, fj is the closest function to fe in M from the
. This indicates
third property of M. From Fano’s inequality [Fan61], H(fj |f˜) ≤ H( 14 ) + log(|M|−1)
4
I(fj ; fe) = H(fj ) − H(fj |f˜) ≥ log |M| − 1 − log(|M| − 1)/4 ≥ 0.7 log |M| ≥ 0.4d.
At the same time, by the data processing inequality, the algorithm A makes m queries x1 , . . . , xm
and sees y1 , . . . , ym , which indicates
I(fe; fj ) ≤ I
y1 , . . . , ym ; fj
m
X
I yi ; fj (xi ) y1 , · · · , yi−1 .
=
(24)
i=1
For the query xi , let Di,j denote the distribution of fj ∈ M in the algorithm A given the first i − 1
observations x1 , y1 , . . . , xi−1 , yi−1 . We apply Theorem 3.4 on Di,j such that it bounds
1
I yi = fj (xi ) + N (0, ); fj (xi ) y1 , · · · , yi−1
ε
E [f (xi )2 ]
f ∼Di,j
1
≤ log 1 +
2
1/ǫ
max[f (xi )2 ]
1
f ∈M
≤ log 1 +
2
1/ε
ε
1
= log 1 + ε ≤ ,
2
2
where we apply
the second property of M in the second step to bound f (x)2 for any f ∈ M.
Pm
Hence we bound i=1 I(yi ; fj |y1 , · · · , yi−1 ) by m · 2ε . This implies
0.4d ≤ m ·
0.8d
ε
⇒m≥
.
2
ε
Next we consider the sample complexity of linear regression.
20
Theorem 7.3. For any K, d, and ε > 0, there exist
a distribution
D, a linear family of functions
2
F with dimension d whose condition number sup
equals K, and a noise function
sup |h(x)|
khk2
h∈F :h6=0
D
x∈G
g orthogonal to V such that any algorithm observing y(x) = f (x) + g(x) of f ∈ F needs at least
√
Ω(K log d + Kε ) samples from D to output fe satisfying kfe − f kD ≤ 0.1 ε · kgkD with probability 34 .
Proof. We fix K to be an integer and set the domain of functions in F to
be [K]. We choose D to
be the uniform distribution over [K]. Let F denote the family of functions f : [K] → C|f (d + 1) =
2
equals K. h(x) = 1x=1
f (d + 2) = · · · = f (K) = 0 . Its condition number sup sup |h(x)|
khk2
provides the lower bound ≥ K. At the same time,
D
h∈F :h6=0 x∈G
|h(x)|2
|h(x)|2
= PK
2
khk2D
i=1 |h(x)| /K
≤ K indicates the upper
bound ≤ K.
We first consider the case K log d ≥ Kε . Let g = 0 such that g is orthogonal to V . Notice that
√
kfe − f kD ≤ 0.1 ε · kgkD indicates fe(x) = f (x) for every x ∈ [d]. Hence the algorithm needs to
sample f (x) for every x ∈ [d] when sampling from D: the uniform distribution over [K]. From the
lower bound of the coupon collector problem, this takes at least Ω(K log d) samples from D.
Otherwise, we
prove that the algorithm needs Ω(K/ε) samples. Without loss of generality,
we assume E |f (x)|2 = 1 for the truth f in y. Let g(x) = N (0, 1/ε) for each x ∈ [d]. From
x∼[d]
Theorem 7.1, to find fe satisfying E |fe(x) − f (x)|2 ≤ 0.1 E |f (x)|2 , the algorithm needs at
x∼[d]
x∼[d]
least Ω(d/ε) queries of x ∈ [d]. Hence it needs Ω(K/ε) random samples from D, the uniform
distribution over [K].
8
Application to Continuous k-sparse Fourier Transforms
We consider the nonlinear function space containing signals with k-sparse Fourier transform in the
continuous setting. Let D be the uniform distribution over [−1, 1] and F be the bandlimit of the
frequencies. We fix the family F in this section to be
k
X
vj e2πi·fj x vj ∈ C, |fj | ≤ F .
F = f (x) =
j=1
The main result in this section is an estimation of κ =
2
] for signals with k-sparse
[sup |fkf(x)|
k2
E
x∈[−1,1] f ∈F
Fourier transform, which is better than the condition number
D
2
] used in [CKPS16].
sup [sup |fkf(x)|
k2
x∈[−1,1] f ∈F
D
Theorem 8.1. For signals with k-sparse Fourier transform,
#
"
|f (x)|2
= O(k2 log2 k).
E
sup
x∈[−1,1] f ∈F kf k2
D
Moreover, there exists c = O(1) such that a distribution
(
c
for |x| ≤ 1 −
log k ,
DF (x) = (1−|x|)
2
c · k log k, for |x| > 1 −
guarantees for any f (x) =
k
X
j=1
vj e2πifj x ,
|f (x)|2 ·
21
1
k 2 log2 k
1
k 2 log2 k
D(x)
= O(k2 log2 k) · kf k2D
DF (x)
∀x ∈ [−1, 1].
We first state the condition number result in [CKPS16].
P
Lemma 8.2 (Lemma 5.1 of [CKPS16]). For any f (x) = kj=1 vj e2πifj x ,
|f (x)|2
4
3
2 = O(k log k).
kf
k
x∈[−1,1]
D
sup
Our main technical lemma is an upper bound of |f (t)|2 in every position of t ∈ (−1, 1).
Lemma 8.3. For any t ∈ (−1, 1),
k2 log k
|f (t)|2
.
.
2
1 − |t|
f ∈F kf kD
sup
The main ingredient is an interpolation lemma of complex numbers.
Lemma 8.4. Given
Pany z1 , . . . , zk ∈ C with |z1 | = |z2 | = · · · = |zk | = 1, there exists a Laurent
polynomial P (z) = kj=−k cj z j such that
1. c0 = 1.
2. |cj | ≤ 1 for every j ∈ [−k, k].
3. P (zj ) = 0 for every j ∈ [k].
Q
P
Proof. Let Q(z) = kj=1 (z − zj ) whose coefficients are (Q0 , . . . , Qk ), i.e., Q(z) = kj=0 Qj · z j . Let
i0 ∈ {0, 1, . . . , k} denote the power z i0 in Q(z) with the largest coefficient, i.e., i0 = arg max Qi .
We set P (z) = Q(z)/(Qi0 ·
statement.
z i0 )
i=0,1,...,m
Pk−i0
Qj+i0
j=−i0 Qi0
=
·
zj ,
which satisfies the three conditions in the
Proof of Lemma 8.3. PGiven k frequencies f1 , . . . , fk and ∆ > 0, we set z1 = e2πif1 ·∆ , . . . , zk =
e2πifk ·∆ . Let P (z) = kj=−k cj · z j be the Laurent polynomial in Lemma 8.4 of z1 , . . . , zk . For
P
f (x) = kj=1 vj e2πifj x , we rewrite f (x + l∆) for l ∈ [−k, k] as a linear combination of z1 , . . . , zk :
f (x + l∆) =
k
X
vj e2πifj (x+l∆) =
j=1
j=1
Pk
k
X
vj · e2πifj ·x · e2πifj ·(l∆) =
k
X
j=1
vj · e2πifj ·x · zjl .
· zjl = 0 for all j ∈ [k] because z1 , . . . , zk are the roots of P (z). This provides a linear
relationship of f (x − k∆), . . . , f (x), . . . , f (x + k∆) with coefficients c−k , . . . , ck , i.e.,
!
k
k
k
k
k
X
X
X
X
X
vj e2πifj ·x ·
cl · zjl = 0.
vj · e2πifj ·x · zjl =
cl · f (x + l∆) =
cl
l=−k cl
l=−k
Hence c0 f (x) = −
|f (x)|2 ≤
P
X
l∈[−k,...,k]\{0} cl
l∈[−k,...,k]\{0}
j=1
j=1
l=−k
c2l ·
l=−k
· f (x + l∆), which indicates
X
l∈[−k,...,k]\{0}
|f (x + l∆)|2 ≤ 2k ·
22
X
l∈[−k,...,k]\{0}
|f (x + l∆)|2 .
Without loss of generality, we assume t = 1 − ε ∈ [0, 1) and integrate ∆ from 0 to ε/k:
ε/k · |f (t)|2 ≤
Z
ε/k
∆=0
2k ·
l∈[−k,...,k]\{0}
X
≤ 2k
X
Z
ε/k
l∈[−k,...,k]\{0} ∆=0
X
≤ 2k
l∈[1,...,k]
X
≤ 2k
l∈[1,...,k]
. k log k ·
Z
1
·
l
1
·
l
1
t=−1
Z
Z
|f (t + l∆)|2 d∆
|f (t + l∆)|2 d∆
ε·l/k
∆′ =0
′ 2
|f (t + ∆ )| d∆ + 2k
ε
∆′ =−ε
X
l∈[1,...,k]
by choosing ε =
the upper bound of
κ = O(k2 log2 k).
ε·l/k
∆′ =0
|f (t − ∆′ )|2 d∆′
k 2 log k
ε
·
E
[|f (x)|2 ].
x∈[−1,1]
We bound
[sup
E
1
.
k3
Z
|f (t)|2 dt.
|f (x)|2
]
x∈[−1,1] f ∈F kf k2
D
Z
|f (x)|2
1 1
sup
=
dx
2 x=−1 f ∈F kf k2D
Z 1−ε
|f (x)|2
4
3
sup
.
2 dx + ε · k log k
kf
k
f
∈F
x=−1+ε
D
Z 1−ε
k2 log k
.
dx + ε · k4 log3 k
1
−
|x|
x=−1+ε
1
2
. k log k · log + ε · k4 log3 k . k2 log2 k
ε
κ=
1
·
l
|f (t + ∆′ )|2 d∆′
From all discussion above, we have |f (1 − ε)|2 .
Proof of Theorem 8.1.
′
Next we define DF (x) = D(x) ·
sup
f ∈F ,f 6=0
|f (x)|2
kf k2D
from Lemma 8.2
from Lemma 8.3
2
sup |f (x)|
2
f ∈F ,f 6=0 kf kD
κ
. The description of DF (x) follows
in Lemmas 8.2 and 8.3. From Claim 4.6, its condition number is
P
Corollary 8.5. For any F > 0, T > 0, ε > 0, and observation y(x) = kj=1 vj e2πifj x + g(x) with
|fj | ≤ F for each j, there exists a non-adaptive algorithm that takes m = O(k 5 log3 k + k3 log2 k ·
P
e
log FεT ) random samples t1 , . . . , tm from DF and outputs fe(x) = kj=1 vej e2πifj x satisfying
i
h
|f (x)|2 with probability 0.9.
|g(x)|2 + ε
E
E
E
|fe(x) − f (x)|2 .
Proof of Corollary 8.5.
in [CKPS16], let Nf =
x∼[−T,T ]
x∼[−T,T ]
x∼[−T,T ]
We first state the main tool from the previous work. From Lemma 2.1
· Z ∩ [−F, F ] denote a net of frequencies for a constant C. For any
ε
T ·k Ck2
23
Algorithm 5 Recover k-sparse FT
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
procedure SparseFT(y, F, T, ε)
m ← O(k5 log3 k + k3 log2 k log FεT )
Sample t1 , . . . , tm from DF independently
Set the corresponding weights (w1 , . . . , wm ) and S = (t1 , . . . , tm )
Query y(t1 ), . . . , y(tm ) from the observation y
Nf ← εCk2 · Z ∩ [−F, F ] for a constant C
T ·k
for all possible k frequencies f1′ , . . . , fk′ in Nf do
′
′
Find h(x) in span{e2πi·f1 x , . . . , e2πi·fk x } minimizing kh − ykS,w
Update fe = h if kh − ykS,w ≤ kfe − ykS,w
end for
Return fe.
end procedure
signal f (x) =
Pk
2πifj (x) ,
j=1 vj e
f ′ (x) =
there exists a k-sparse signal
k
X
j=1
′
vj′ e2πifj (x) satisfying kf − f ′ kD ≤ εkf kD
whose frequencies f1′ , . . . , fk′ are in Nf . We rewrite y = f + g = f ′ + g ′ where g′ = g + f − f ′ with
kg′ kD ≤ kgkD + εkf kD . Our goal is to recover f ′ .
We construct a δ-net with δ = 0.05 for
2k
X
b
vj e2πi·hj x khkD = 1, b
hj ∈ Nf .
h(x) =
j=1
We first pick 2k frequencies b
h1 , . . . , b
h2k in Nf then construct a δ-net on the linear subspace
b
b
2πi
h
x
2πi
h
x
1 ,...,e
2k }. Hence the size of our δ net is
span{e
4F T ·kCk2
ε
2k
2
· (12/δ)2k ≤ (
4F T · kCk 3k
) .
ε·δ
Now we consider the number of random samples from DF to estimate signals in the δ-net. Based
on the condition number of DF in Theorem 8.1 and the Chernoff bound of Corollary 3.2, a union
bound over the δ-net indicates
2
k log2 k
k2 log2 k
FT
3
m = O(
· log |net|) = O
· (k log k + k log
)
δ2
δ2
εδ
random samples from DF would guarantee that for any signal h in the net, khk2S,w = (1 ± δ)khk2D .
From the property of the net,
for any h(x) =
2k
X
j=1
which is sufficient to recover f ′ .
vj e2πihj (x) with b
hj ∈ Nf ,
24
khk2S,w = (1 ± 2δ)khk2D ,
We present the algorithm in Algorithm 5 and bound kf − fekD as follows. The expectation of
kf − fekD over the random samples S = (t1 , . . . , tm ) is
kf − f ′ kD + kf ′ − fekD ≤ kf − f ′ kD + 1.1kf ′ − fekS,w
≤ kf − f ′ kD + 1.1(kf ′ − ykS,w + ky − fekS,w )
≤ kf − f ′ kD + 1.1(kg′ kS,w + ky − f ′ kS,w )
≤ εkf kD + 2.2(kgkD + εkf kD ).
From the Markov inequality, with probability 0.9, kf − fekD . εkf kD + kgkD .
Acknowledgements
The authors would like to thank Adam Klivans and David Zuckerman for many helpful comments
about this work.
References
[BBL06] Maria-Florina Balcan, Alina Beygelzimer, and John Langford. Agnostic active learning.
In Proceedings of the 23rd international conference on Machine learning, pages 65–72.
ACM, 2006.
[BSS12] Joshua Batson, Daniel A Spielman, and Nikhil Srivastava. Twice-ramanujan sparsifiers.
SIAM Journal on Computing, 41(6):1704–1721, 2012.
[CDL13] Albert Cohen, Mark A Davenport, and Dany Leviatan. On the stability and accuracy of
least squares approximations. Foundations of computational mathematics, 13(5):819–
834, 2013.
[Che52] Herman Chernoff. A measure of asymptotic efficiency for tests of a hypothesis based
on the sum of observations. The Annals of Mathematical Statistics, 23:493–507, 1952.
[CKPS16] Xue Chen, Daniel M. Kane, Eric Price, and Zhao Song. Fourier-sparse interpolation
without a frequency gap. In FOCS 2016, 2016.
[Coh94] David A Cohn. Neural network exploration using optimal experiment design. In
Advances in neural information processing systems, pages 679–686, 1994.
[DMM08] Petros Drineas, Michael W Mahoney, and S Muthukrishnan. Relative-error cur matrix
decompositions. SIAM Journal on Matrix Analysis and Applications, 30(2):844–881,
2008.
[Fan61] Robert Fano. Transmission of information; a statistical theory of communications.
Cambridge, Massachusetts, M.I.T. Press, 1961.
[GBD92] Stuart Geman, Elie Bienenstock, and René Doursat. Neural networks and the
bias/variance dilemma. Neural Computation, 4:1–58, 1992.
[GKKW06] László Györfi, Michael Kohler, Adam Krzyzak, and Harro Walk. A distribution-free
theory of nonparametric regression. Springer Science & Business Media, 2006.
25
[Har28] Ralph Hartley. Transmission of information. Bell System Technical Journal, 1928.
[HKZ14] Daniel Hsu, Sham M. Kakade, and Tong Zhang. Random design analysis of ridge
regression. Foundations of Computational Mathematics, 14(3):569–600, 2014.
[LG94] David D Lewis and William A Gale. A sequential algorithm for training text classifiers.
In Proceedings of the 17th annual international ACM SIGIR conference on Research
and development in information retrieval, pages 3–12. Springer-Verlag New York, Inc.,
1994.
[MI10] Malik Magdon-Ismail. Row sampling for matrix algorithms via a non-commutative
bernstein bound. arXiv preprint arXiv:1008.0587, 2010.
[Owe13] Art B. Owen. Monte Carlo theory, methods and examples. 2013.
[Set09] Burr Settles. Active learning literature survey. Computer Sciences Technical Report
1648, University of Wisconsin–Madison, 2009.
[Sha49] Claude Shannon. Communication in the presence of noise. Proc. Institute of Radio
Engineers, 37(1):10–21, 1949.
[SL12] George AF Seber and Alan J Lee. Linear regression analysis, volume 936. John Wiley
& Sons, 2012.
[SOS92] H Sebastian Seung, Manfred Opper, and Haim Sompolinsky. Query by committee.
In Proceedings of the fifth annual workshop on Computational learning theory, pages
287–294. ACM, 1992.
[SS11] Daniel A Spielman and Nikhil Srivastava. Graph sparsification by effective resistances.
SIAM Journal on Computing, 40(6):1913–1926, 2011.
[Tar09] Robert E. Tarjan. Lecture 10: More chernoff bounds, sampling, and the chernoff +
union bound. Princeton Class Notes, Probability and Computing, pages 1–9, 2009.
[Tro12] Joel A. Tropp. User-friendly tail bounds for sums of random matrices. Foundations of
Computational Mathematics, 12:389–434, 2012.
[Woo14] David P Woodruff. Sketching as a tool for numerical linear algebra. arXiv preprint
arXiv:1411.4357, 2014.
26
| 8cs.DS
|
1
Joint Sum Rate And Error Probability
Optimization: Finite Blocklength Analysis
Mahdi Haghifam, Mohammad Robat Mili, Behrooz Makki, Masoumeh
arXiv:1707.09735v2 [cs.IT] 1 Aug 2017
Nasiri-Kenari, Tommy Svensson.
Abstract
We study the tradeoff between the sum rate and the error probability in downlink of wireless
networks. Using the recent results on the achievable rates of finite-length codewords, the problem is
cast as a joint optimization of the network sum rate and the per-user error probability. Moreover, we
develop an efficient algorithm based on the divide-and-conquer technique to simultaneously maximize
the network sum rate and minimize the maximum users’ error probability and to evaluate the effect of
the codewords length on the system performance. The results show that, in delay-constrained scenarios,
optimizing the per-user error probability plays a key role in achieving high throughput.
I. I NTRODUCTION
The fifth generation of wireless communication (5G) must support novel traffic types for
which low latency, high data rate, and ultra reliability are of interest. Particularly, in many
applications such as vehicle-to-vehicle and vehicle-to-infrastructure communications for traffic
efficiency/safety or real-time video processing for augmented reality, the codewords are required
to be short (in the order of ∼ 100 channel uses) with stringent requirements on the latency and
reliability [1]. Therefore, it is interesting to optimize the performance of wireless networks in
the presence of finite-length codewords.
In 2010, [2] presented accurate information-theoretic approximations for the achievable rates
of finite blocklength codes. Using [2], the performance of wireless networks with short packets
Mahdi Haghifam, Mohammad Robat Mili and Masoumeh Nasiri-Kenari are with Electrical Engineering Department,
Sharif University of Technology, Tehran, Iran. Emails: haghifam [email protected], [email protected],
[email protected]. Behrooz Makki and Tommy Svensson are with Chalmers University of Technology, Gothenburg, Sweden.
Emails:{behrooz.makki, tommy.svensson}@chalmers.se.
2
has been studied in various papers, for the cases with cognitive radio [3], relay networks [4],
hybrid automatic repeat request technique [5].
In this letter, we consider a wireless network with an access point (AP) serving multiple users.
Using short packets, the AP transmits packets in the downlink to the users, which have different
target error probability requirements. Particularly, using the recent results of [2], we propose a
joint sum rate and per-user error probability optimization problem and investigate the effect of
the codeword length on the system performance. To solve the joint sum rate and per-user error
probability optimization problem, we develop a low-complexity two-level algorithm based on the
divide-and-conquer approach. Also, we derive a closed-form expression for the optimal per-user
error probability (Theorem 1). Finally, we find an efficient and close-to-optimal power allocation
algorithm, in terms of sum rate and error probability, based on the augmented Lagrange method
[6] (Algorithm 1).
The simulation and analytical results show that 1) our proposed algorithm can reach (almost) the same performance as in the exhaustive search-based approach with considerably less
implementation complexity (Figs. 1 and 2). Also, 2) the throughput is sensitive to the length
of short packets while its sensitivity to the packet length decreases for long packets (Fig. 2b).
Finally, 3) optimal error probability assignment with water-filling (WF) power allocation achieves
higher throughput, compared to using the optimal power allocation with equal error probability
assignment (Figs. 1, 2a).
II. S YSTEM M ODEL
We consider a downlink communication model with N single-antenna users which are served
by a single-antenna AP. It is assumed that each user is allocated an orthogonal channel to the
AP, for instance they could be separated in the frequency or time domain. Let us denote the
instantaneous channel gain between the AP and the i-th user, i = 1, . . . , N, by gi . The channel
gain gi can be expressed as gi = ḡi θ i , where θ i represents the small scale fading and ḡi is the
average channel gain, obtained by considering the path loss effects and shadowing. Thus, with
i-th user located at distance di from the AP, we have ḡi = κi di−δi where κi is the signal power gain
at distance 1 meter from the AP and δi is the path loss exponent. Moreover, the power allocated
by the AP to the signal of user i is denoted by pi . Thus, the instantaneous signal-to-noise ratio
(SNR) received by user i is γi =
pi gi
,
σ2
where σ 2 is the noise power density. We characterize
the network performance when the AP employs packets of short length. Specifically, the user
3
i’s message is encoded into the packets of length L and transmitted with power pi . In this way,
the maximum achievable information rate in nats per channel use (npcu) for user i which can
be decoded with block error probability no greater than ǫi is given by [7, Thm. 1]
s
log (L)
1
1
Q −1 (ǫi ) +
1−
.
ri = log (1 + γi pi ) −
2
L
L
(1 + γi pi )
(1)
In (1), the achievable rate increases unboundedly as the error probability tends towards one. On
the other hand, the rate decreases significantly in the cases with strict error probability requirements, i.e., small ǫi ’s. Also, the achievable rate increases with the signal length L monotonically
and letting L → ∞, the achievable rate (1) converges to Shannon’s capacity formula in the cases
with asymptotically long codewords.
Motivated by the tradeoff between the achievable rates and the error probability in (1), we
consider a joint sum rate maximization and per-user error probability minimization problem. Assuming perfect channel state information (CSI) at the AP, we study a multi-objective optimization
problem
maximize
ǫ,p
N
Õ
i=1
log (1 + γi pi ) −
minimize
max{ǫ1, . . ., ǫ N }
subject to
0 ≤ ǫi ≤ εmax,i
ǫ,p
0 ≤ pi
N
Õ
i=1
s
1
1
1−
Q −1 (ǫi )
2
L
(1 + γi pi )
(2b)
∀i ∈ {1, . . . , N},
∀i ∈ {1, . . ., N},
pi ≤ Pmax,
(2a)
(2c)
(2d)
(2e)
where ǫ = [ǫ1, . . . , ǫ N ] and p = [p1, . . ., p N ]. Also, in (2b), the goal is to minimize the maximum
error probability of the users. Then, in (2c), εmax,i is the maximum error probability constraint
of user i which indicates the supporting quality of service (QoS) requirement of user i. Also,
the total power constraint of the AP is denoted by Pmax . In this way, (2) is of interest in
emerging applications of 5G calling for heterogeneous QoS requirements on data rate and
reliability. For instance, massive machine-type communication and ultra-reliable and low latency
communication scenarios demand short-length packet exchange with stringent requirement on
reliability at moderately low rate [1]. The requirements of the aforementioned services illustrate
how the framework of (2) can be utilized to balance conflicting performance objectives, namely,
sum rate and maximum error probability. Moreover, as seen in the following, our discussions are
4
well applicable to the cases when optimizing the network sum throughput, which is defined as
the product of the rate and the successful message decoding probability. However, as opposed to
throughput optimization, (2) is flexible in optimizing the rates and error probabilities individually
based on the QoS requirements.
Depending on the number of users, there may be no closed-form solution for (2). Thus, we
follow the same method as in [8] to convert the problem into a single objective optimization
using the weighted sum method while normalizing the objectives. Also, as seen in Section IV,
our proposed sub-optimal approach can reach (almost) the same performance as in the optimal
exhaustive search-based scheme. With no loss of generality, we assume εmax,1 ≤ εmax,2 ≤ . . . ≤
εmax,N < 12 . Also, to guarantee a consistent comparison between the objectives in (2a) and (2b),
we normalize them as
U1 (p, ǫ) =
where SR∞
ÍN
i=1 log (1
+ γi pi ) −
r
1
L
1−
1
(1+γi pi )2
Q
−1
(ǫi )
(3)
SR∞
εmax,N − max{ǫ1, . . ., ǫ N }
,
(4)
U2 (ǫ) =
εmax,N
is a normalization factor that can be found by plugging the water-filling power
allocation expression into the Shannon’s capacity formula which provides an upper bound for
(2a). Then, we use the weighted sum method to rewrite (2) as the single-objective optimization
problem
maximize ωU1 (p, ǫ) + (1 − ω) U2 (ǫ)
(5a)
subject to (2c) − (2e).
(5b)
ǫ,p
Here, 0 ≤ ω ≤ 1 is the weighting parameter. Note that, with ω ranging from 0 to 1, scenarios
with strict rate requirements and relaxed error probability requirements to scenarios with low
rate requirements and ultra-reliability are addressed.
III. P ROPOSED A LGORITHM
The optimization problem (5) belongs to the class of non-convex problems which has a multimodal objective function, so finding its global optimal solution is computationally infeasible.
For this reason, we apply the primal decomposition approach [9] to optimize ǫ and p separately.
In this way, to solve (5), we use the following iterative approach
p[0] → ǫ[1] → p[1] → . . . ǫ[T] → p[T],
|{z}
| {z }
initialization
optimal solution
5
where ǫ[t] and p[t] are the optimal error probability and power allocation vectors at iteration t,
and T is the maximum number of iterations considered by the network designer. The details of
our proposed optimization approach are as follows.
A. Error Probability Optimization For A Given Power Allocation
Here, for a given power allocation p⋆ [t − 1] := p, we find the optimal error probabilities of
each user at iteration t denoted by ǫ⋆[t]. Setting z = max{ǫ1, . . ., ǫ N } and assuming a given
power allocation, (5) is rephrased as
minimize
ǫ,z
N
ω Õ
SR∞ i=1
s
subject to ǫi ≤ z
1−ω
1
1
Q −1 (ǫi ) +
1−
z
2
L
εmax,N
(1 + γi pi )
∀i ∈ {1, . . ., N}
0 ≤ z ≤ ǫmax,N
0 ≤ ǫi ≤ εmax,i
(6a)
(6b)
(6c)
∀i ∈ {1, . . ., N}
(6d)
Theorem 1 gives a closed-form expression for the optimal error probability assignment of each
user in terms of (6).
Theorem 1. The optimal error probabilities of the users are given by
h
i
εmax,1, . . . , εmax,k−1, βk, . . ., βk
βk ∈ Ik
| {z }
⋆
,
(7)
ǫ =
N−k+1 times
εmax,1, . . . , εmax,N
otherwise
for k = 1,
Ik = (εmax,k−1, εmax,k ] with I1 = (0, εmax,1 ], and
!
q. . ., N. Here, √we define the intervals
L(1−ω)SR∞
√22
βk = Q 2 log
.
√ ÍN
εmax, N ω 2π i=k
γ p +2γi pi
i i
1+γi pi
Proof. Since the constraints in (6b)-(6d) are affine functions in ǫ and z, it is sufficient to prove
that the objective function in (6a) is convex. The second derivative of Q −1 (x) is given by
2
d2 Q −1 (x)
−1
−1
= 2πQ (x) exp Q (x)
> 0 if x < 21 . Therefore, considering the fact that ǫi ≤
dx 2
ǫmax,i ≤ 21 , the objective function in (6a) is a sum of convex functions and an affine function,
i.e., z. Hence, (6) is a convex optimization problem, and the optimal solution can be found by
considering Karush-Kuhn-Tucker (KKT) conditions. Thus, we write the Lagrangian function of
(6) as
N
ω Õ
L (ǫ, z, λ, ν, η) =
SR∞ i=1
s
1
1
Q −1 (ǫi ) +
1−
2
L
(1 + γi pi )
6
N
N
Õ
Õ
1−ω
z−
λi (z − ǫi ) −
νi ǫmax,i − ǫi − η εmax,N − z ,
εmax,N
i=1
i=1
(8)
where λ = [λ1, . . . , λ N ] 0, η ≥ 0, and ν = [ν1, . . ., νN ] 0 are dual variables associated with
constraints in (6b), (6c), and (6d), respectively. According to the KKT conditions, the optimal
solution, which is denoted by ǫ⋆ and z⋆, should satisfy
s
√
1
ω
∂L
1
1
−
=
−
2π exp
SR∞ L
∂ǫi⋆
(1 + γi pi )2
+ λ⋆i + νi⋆ = 0,
(9b)
(9c)
(9d)
(9e)
(6b)-(6d).
ǫi⋆ is equal
2 !
(9a)
N
∂L
1−ω Õ ⋆
=
−
λ + η⋆ = 0,
∂z⋆ εmax,N i=1 i
λ⋆i z⋆ − ǫi⋆ = 0,
νi⋆ εmax,i − ǫi⋆ = 0,
η⋆ εmax,N − z⋆ = 0,
In (9a), we
Q −1 ǫi⋆
2
(9f)
√
−1
(Q (x))
. From (9c) and (9d), it can be verified that
have used dQdx(x) = − 2π exp
2
to either z⋆ or εmax,i ; otherwise, λ⋆i and νi⋆ must be equal to zero which contradict
2
−1
with (9a). Assume z⋆ < εmax,N and z⋆ ∈ Ik , so according to (9e), we have η⋆ = 0. Note that,
for 1 ≤ i ≤ k − 1, ǫi⋆ must be equal to εmax,i since we have z⋆ > εmax,i . Also, due to the fact
that z⋆ < εmax,i , it can be inferred that ǫi⋆ = z⋆ for k ≤ i ≤ N. Thus, in the cases with z⋆ ∈ Ik ,
we have ǫ⋆ = εmax,1, . . ., εmax,k−1, z⋆, . . . , z⋆ . Then, from (9c) and (9d), it can be concluded
| {z }
N
− k + 1 times
⋆
⋆
that λi = 0 for 1 ≤ i ≤ k − 1 and νi = 0 for k ≤ i ≤ N. In this way, (9a) can be expressed as
s
2 !
√
Q −1 z⋆
1
1
ω
1−
= λ⋆i ,
(10)
2π exp
2
SR∞ L
2
(1 + γi pi )
for k ≤ i ≤ N. Also, from (9b) and (10), we have
N
1−ω Õ ⋆
=
λ .
εmax,N i=k i
(11)
Plugging (10) into (11), the upper branch of (7) is found. In this way, depending on z⋆ being
in each region Ik , the closed-form solution for ǫ⋆ is provided. Then, given z⋆ = εmax,N , it is
straightforward to show that the objective function in (6a) is a decreasing function in each ǫi , so
7
the lower branch of (7) provides the optimal solution. Note that because of the strict convexity
of (6a), there is an optimal solution for ǫ which is found by searching in N + 1 branches of
(7).
B. Optimal Power Allocation For A Given Error Probability
Consider a given ǫ[t] := ǫ. Then, (5) is relaxed to
maximize ωU1 (p, ǫ)
(12a)
subject to
(12b)
p
(2d)-(2e).
Since the function in (12a) is non-concave in p, problem (12) belongs to the class of nonconvex optimization problems. In a non-convex problem, there is a nonzero duality gap between
primal and dual problems. Here, we use the augmented Lagrange approach [6, sec 4] to deal
with this non-convex optimization which reduces the duality gap by augmenting a penalty-like
quadratic term added to the Lagrangian function. In [6, Sec 4.2], it has been proved that the
augmented Lagrangian is locally convex when the penalty parameter is sufficiently large. In
contrast to the penalty functions approach, the augmented Lagrangian function largely preserves
smoothness and does not require an asymptotically large penalty parameter for the method to
converge, meaning that the penalization is exact. Augmented Lagrangian algorithms are based on
successive maximization of the augmented Lagrangian function in which the multiplier estimates
and penalty parameter are fixed in each iteration and then updated between iterations. Applying
the augmented Lagrangian method on (12), which eliminates the constraints and adds them to
the objective function, gives the augmented Lagrangian function
s
#
" N
1
1
ω Õ
Q −1 (ǫi )
1−
log (1 + γi pi ) −
L µ,ζ (p) =
2
SR∞ i=1
L
(1 + γi pi )
(13)
(
!)!
2
N
Õ
1
−
max 0, ζ − µ Pmax −
− ζ 2 ,
pi
2µ
i=1
where µ is a positive coefficient denoting the penalty parameter and ζ is the Lagrangian dual
variable associated with (2e). Then, at stage l of the power allocation problem, we solve
maximize
p
L µ(l),ζ (l) (p) ,
(14)
8
which approximates (12) to find the power allocation at iteration l+1 denoted by p(l+1) . Moreover,
variables ζ (l) and µ(l) are updated according to
(
ζ (l+1) = max 0, ζ (l) − µ(l) Pmax −
N
Õ
(l)
pi
i=1
!)
,
(15)
µ(l+1) = 2µ(l),
respectively. In this way, as µ(l) increases, the violations introduced by constraints are penalized
more severely so that the maximizer of the penalty function in (14) gives the results closer to
the feasible region. In [6, Sec 4.2], it has been shown that while the constraints are nonlinear,
the convergence rate of the augmented Lagrangian method is linear.
The iterative joint error probabilities and power allocation algorithm is summarized in Algorithm 1. In order to analyze the complexity order of the proposed algorithm, we note that the
optimal error probabilities can be found by (7) with the complexity O (N). Also, the complexity
of the power allocation at each iteration is O N 2 . Thus, the complexity of Algorithm 1 is
O N 2 + O (N) = O N 2 .
Algorithm 1 Error Probabilities Assignment and Power Allocation
1:
For every given ω, Pmax , {εmax,1, . . . , εmax,N }, µ(0) , and ζ (0) .
2:
Initialize: p[0] and t = 0.
3:
while {ǫ ⋆[t]} converges do
4:
Calculate ǫ⋆[t + 1] via (7) with given p⋆[t].
5:
Initialize: l = 0.
6:
while {p(l) } converges do
7:
Calculate p(l+1) via (14) with given ǫ⋆[t + 1], ζ (l) , and µ(l) .
8:
Update ζ (l+1) and µ(l+1) via (15).
9:
l = l + 1.
10:
end while
11:
p⋆[t + 1] = p(l) .
12:
t =t+1
13:
end while
9
IV. N UMERICAL R ESULTS
AND
C ONCLUSION
Here, we study the trade-off between the sum rate and maximum error probability. We set the
noise power σ 2 = 1, the number of users N = 4, and the error probability constraints εmax =
[10−5, 5×10−5, 10−4, 5×10−4 ] which are often assumed for vehicular-to-vehicular communications
[1]. Also, it is assumed that the users are equidistant from the AP. Also, we consider Rayleigh
fading with mean 1. Finally, we set µ(0) = 1 and ζ (0) = 0.15 in Algorithm 1. For the numerical
results, we consider the cases with L ≥ 100 channel uses, for which the approximation (1) is
tight enough [2]. Also, we compare our method with three baseline algorithms: 1) WF-based
power allocation with the error probabilities of all users set to the minimum of the required
error probabilities, called minmax error probability assignment, i.e., ǫ = [εmax,1, . . ., εmax,1 ], 2)
the proposed method for power allocation with the minmax error probability assignment, and 3)
equal power allocation with the proposed method for the error probabilities assignment. Finally,
the results are obtained by averaging over 104 different channel realizations.
Figure 1 shows the the tradeoff between the sum rate and the error probability for different
algorithms with Pmax = 6 dB and L = 200 channel uses. As a performance metric, we define
the sum throughput as
T =
N
Õ
i=1
ri (1 − ǫi ) ,
(16)
where the user i’s codeword rate and error probability are given by ri and ǫi , respectively. Then,
Fig. 2a demonstrates the sum throughput versus the AP’s total power constraint Pmax by setting
L = 100 channel uses and ω = 0.9. Finally, Fig. 2b evaluates the effect of the codeword length
on the sum throughput. The results lead to the following conclusions:
• The scheme with the WF power allocation and the error probability assignment based on
Theorem 1 achieves the tradeoff region close to the proposed method optimizing both the error
probability and the power allocation (Fig. 1).
• For short codewords, the throughput is remarkably affected by the length of the codeword.
However, the effect of increasing the codeword length decreases for long codewords (Fig. 2b).
Also, optimal error probability assignment with WF power allocation achieves higher throughput
compared to optimal power allocation with equal error probability assignment. Moreover, the
performance of WF with minmax error probability assignment is close to that of the scheme
with the proposed power allocation with minmax error probability assignment (Figs. 2a and 2b).
• For short codeword (say, L ≤ 1000 channel uses), the proposed algorithm leads to considerable
10
Sum Rate (npcu)
1.7
Exhaustive search
Proposed method
WF power allocation+proposed error
1.65
1.6
1.55
Pmax = 6 dB, L = 200 cu
1.5
10
−7
−6
10
max{ǫ⋆1 , ..., ǫ⋆N }
−5
10
Figure 1: Sum rate vs maximum error probability.
throughput improvement in comparison with other schemes. For instance, when L = 100 cu,
the performance of the proposed method has 66% of improvement. However, the performance
difference of the schemes decreases in the cases with long codewords.
• Finally, as observed in Figs. 1, 2a, and 2b, the gap between the developed algorithm and
the exhaustive search-based algorithm diminishes by increasing Pmax or L. Thus, our proposed
algorithm can be effectively applied to jointly optimize the sum rate and the error probability
of multi-user networks in delay-constrained applications.
R EFERENCES
[1] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski, “Five disruptive technology directions for 5G,” IEEE
Commun. Mag., vol. 52, no. 2, pp. 74–80, Feb. 2014.
[2] Y. Polyanskiy, H. V. Poor, and S. Verdu, “Channel coding rate in the finite blocklength regime,” IEEE Trans. Inf. Theory,
vol. 56, no. 5, pp. 2307 – 2359, May 2010.
[3] B. Makki, T. Svensson, and M. Zorzi, “Finite block-length analysis of spectrum sharing networks using rate adaptation,”
IEEE Trans. Commun., vol. 63, no. 8, pp. 2823–2835, Aug. 2015.
[4] Y. Li, M. C. Gursoy, and S. Velipasalar, “Throughput of two-hop wireless channels with queueing constraints and finite
blocklength codes,” in Proc. IEEE ISIT’2016, Barcelona, Spain, July 2016, pp. 2599–2603.
[5] B. Makki, T. Svensson, and M. Zorzi, “Finite block-length analysis of the incremental redundancy HARQ,” IEEE Wireless
Commun. Lett., vol. 3, no. 5, pp. 529–532, Oct. 2014.
[6] D. P. Bertsekas, Nonlinear programming.
Athena scientific Belmont, 1999.
[7] V. Y. F. Tan and M. Tomamichel, “The third-order term in the normal approximation for the awgn channel,” IEEE Trans.
Inf. Theory, vol. 61, no. 5, pp. 2430–2438, 2015.
11
Sum Throughput (npcu)
2
1.5
Exhaustive search
Proposed method
Proposed power allocation+minmax error
WF power allocation+proposed error
WF power allocation+minmax error
Subplot: (a)
1
L=300 cu, ω = 0.9
0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
AP’s total power 10 log(Pmax) (dB)
4.5
5
Sum Throughput (npcu)
1.8
1.5
Exhustive search
Proposed method
Proposed power allocation+minmax error
WF power allocation+proposed error
WF power allocation+minmax error
1
Pmax = 4 dB, ω = 0.9
0.5
100
500
Subplot: (b)
1000
1500
2000
2500
Codeword length L (channel uses)
3000
Figure 2: Sum throughput of the considered algorithms. Subplot (a): Sum throughput vs the total
power constraint of the AP. Subplot (b): Sum throughput vs the codeword length.
[8] W. Yu, L. Musavian, and Q. Ni, “Tradeoff analysis and joint optimization of link-layer energy efficiency and effective
capacity toward green communications,” IEEE Trans. Wireless Commun., vol. 15, no. 5, pp. 3339–3353, May 2016.
[9] D. P. Palomar and M. Chiang, “A tutorial on decomposition methods for network utility maximization,” IEEE J. Sel. Areas
Commun., vol. 24, no. 8, pp. 1439–1451, Aug. 2006.
| 7cs.IT
|
Using Table Valued Functions in SQL Server 2005 To
Implement a Spatial Data Library
Jim Gray
Microsoft Research
Alex Szalay, Gyorgy Fekete,
The Johns Hopkins University
August 2005
Technical Report
MSR-TR-2005-122
Microsoft Research
Advanced Technology Division
Microsoft Corporation
One Microsoft Way
Redmond, WA 98052
Using Table Valued Functions in SQL Server 2005
To Implement a Spatial Data Library
Jim Gray, Microsoft (contact author)
Alex Szalay, Johns Hopkins University
Gyorgy Fekete, Johns Hopkins University
August 2005
Abstract
This article explains how to add spatial search functions (point-near-point and point in
polygon) to Microsoft SQL Server™ 2005 using C# and table-valued functions. It is
possible to use this library to add spatial search to your application without writing any
special code. The library implements the public-domain C# Hierarchical Triangular Mesh
(HTM) algorithms from Johns Hopkins University. That C# library is connected to SQL
Server 2005 via a set of scalar-valued and table-valued functions. These functions act as a
spatial index.
Resources
The article is illustrated by examples that can be downloaded from
http://msdn.microsoft.com/sql/2005/. The sample package includes:
• An 11 MB sample spatial database of United States cities and river-flow gauges.
• The sample queries from the sql\testScript.sql article.
• A Visual Studio 2005 project, \htm.sln, with all the SQL and C# code.
• A paper, doc\Table_Valued_Functions.doc.
• An article, doc\HtmCsharp.doc, that provides a manual page for each routine.
• An article, doc\HTM.doc, that explains the Hierarchical Triangular Mesh
algorithms in detail.
• An article, doc\There_Goes_the_Neighborhood.doc, which explains how
the HTM algorithms are used in Astronomy. This article also explains two other
approaches: zones for batch-oriented point-to-point and point-area comparisons, and
regions for doing Boolean algebra on areas. Public domain implementations of those
approaches implemented for SQL Server are used in the SkyServer, a popular
Astronomy website for the Sloan Digital Sky Survey (http://SkyServer.SDSS.org/ and
by several other astronomy data servers.
Table of Contents
Abstract .......................................................................................................................................................... 1
Introduction .................................................................................................................................................... 3
Table Valued Functions: The Key Idea ......................................................................................................... 3
Using Table-Valued Functions to Add a Spatial Index .................................................................................. 4
The Datasets ................................................................................................................................................... 6
USGS Populated Places (23,000 cities) ..................................................................................................... 7
USGS Stream Gauges (17,000 instruments).............................................................................................. 7
The Spatial Index Table............................................................................................................................. 9
A Digression: Cartesian Coordinates...................................................................................................... 10
Typical Queries ............................................................................................................................................ 12
1. Find Points near point: find towns near a place. ............................................................................ 12
2. Find places inside a box. ................................................................................................................. 13
3. Find places inside a polygon. .......................................................................................................... 14
4. Advanced topics – complex regions. .............................................................................................. 15
References .................................................................................................................................................... 18
Appendix: The Basic HTM Routines ........................................................................................................... 19
HTM library version: fHtmVersion() returns versionString ............................................................... 19
Generating HTM keys: fHtmLatLon (lat, lon) returns HtmID............................................................ 19
LatLon to XYZ: fHtmLatLonToXyz (lat,lon) returns Point (x, y, z).................................................. 19
XYZ to LatLon: fHtmXyzToLatLon (x,y,z) returns Point (lat, lon)................................................... 19
Viewing HTM keys: fHtmToString (HtmID) returns HtmString ....................................................... 19
HTM trixel Centerpoint: fHtmToCenterpoint(HtmId) returns Point (x, y, z) .................................... 20
HTM trixel corner points: fHtmToCornerpoints(HtmId) returns Point (x, y, z) ................................ 20
Computing distances: fDistanceLatLon(lat1, lon1, lat2, lon2) returns distance ................................ 20
Finding nearby objects: fHtmNearbyLatLon(type, lat, lon, radius) returns SpatialIndexTable......... 20
Finding the nearest object: fHtmNearestLatLon(type, lat, lon) returns SpatialIndexTable ............... 21
Circular region HTM cover: fHtmCoverCircleLatLon(lat, lon, radius) returns trixelTable .............. 21
General region specification to HTM cover: fHtmCoverRegion(region) returns trixelTable............. 21
General region simplification: fHtmRegionToNormalFormString(region) returns regionString ....... 22
Cast RegionString as Table: fHtmRegionToTable(region) returns RegionTable .............................. 22
Find Points Inside a Region: fHtmRegionObjects(region, type) returns ObjectTable....................... 22
General region diagnostic: fHtmRegionError(region ) returns message............................................. 22
Introduction
Spatial data searches are common in both commercial and scientific applications. We
developed a spatial search system in conjunction with our effort to build the SkyServer
(http://skyserver.sdss.org/) for the astronomy community. The SkyServer is a multiterabyte database that catalogs about 300 million celestial objects. Many of the questions
astronomers want to ask of it involve spatial searches. Typical queries include, “What is
near this point?” “What objects are inside this area?” and “What areas overlap this area?”
For this article, we have added the latitude/longitude (lat/lon) terrestrial sphere (the earth)
grid to the astronomer’s right ascension/declination (ra/dec) celestial sphere (the sky)
grid. The two grids have a lot in common, but the correspondence is not exact; the
traditional order lat-lon corresponds to dec-ra. This order reversal forces us to be explicit
about the coordinate system. We call the Greenwich-Meridian-Equatorial terrestrial
coordinate system the LatLon coordinate system. The library supports three coordinate
systems:
• Greenwich Latitude-Longitude, called LatLon
• Astronomical right-ascension–declination called J2000
• Cartesian (x, y, z) called Cartesian.
Astronomers use arc minutes as their standard distance metric. A nautical mile is an arc
minute, so the distance translation is very natural. Many other concepts are quite similar.
To demonstrate these, this article will show you how to use the spatial library to build a
spatial index on two USGS datasets: US cities, and US stream-flow gauges. Using these
indexes and some spatial functions, the article provides examples of how to search for
cities near a point, how to find stream gauges near a city, and how to find stream gauges
or cities within a state (polygonal area).
We believe this approach is generic. The spatial data spine schema and spatial data
functions can be added to almost any application to allow spatial queries. The ideas also
apply to other multi-dimensional indexing schemes. For example, the techniques would
work for searching color space or any other low-dimension metric space.
Table Valued Functions: The Key Idea
The key concept of relational algebra is that every relational operator consumes one or
more relations and produces an output relation. SQL is syntactic sugar for this idea,
allowing you to define relations (data definition language) and to manipulate relations
with a select-insert-update-delete syntax.
Defining your own scalar functions lets you make some extensions to the relational
database – you can send mail messages, you can execute command scripts, and you can
compute non-standard scalars and aggregate values such as tax() or median().
However, if you can create tables, then you can become part of the relational engine –
both a producer and consumer of relational tables. This was the idea of OLEDB, which
allows any data source to produce a data stream. It is also the idea behind the SQL Server
2000 Table Valued Functions.
Implementing table valued functions in Transact-SQL is really easy:
create function t_sql_tvfPoints()
returns @points table (x float, y float)
as begin
insert @points values(1,2);
insert @points values(3,4);
return;
end
This is fine if your function can be done entirely in Transact-SQL. But implementing
OLEDB data sources or Table Valued Functions outside of Transact-SQL is a real
challenge in SQL Server 2000.
The common language runtime (CLR) integration of SQL Server 2005 makes it easy to
create a table-valued function. You create a list, an array, or any IEnumerable object
(anything you can do foreach on), and then you cast it as a table. That’s all there is to it.
[SqlFunction(
TableDefinition
= "x float, y float" ,
FillRowMethodName = "FillPair")]
public static IEnumerable csPoints( )
{
int[,] points = { { 1, 2 }, { 3, 4 } };
return (IEnumerable) points;
}
You compile this in Visual Studio and click Deploy. The table-valued function is
installed in the database.
Using Table-Valued Functions to Add a Spatial Index
There is a lot of confusion about indexes. Indexes are really simple – they are tables with
a few special properties:
• SQL Server has only one kind of associative (by value) index – a B-tree. The Btree can have multi-field keys, but the first field carries most of the selectivity.
• Conceptually, the B-Tree index is a table consisting of the B-Tree key fields, the
base table key fields, and any included fields that you want to add to the index.
• B-tree indexes are sorted according to the index key, such as.ZIP code or
customer ID, so that lookup or sequential scan by that key is fast.
• Indexes are often smaller than the base table, carrying only the most important
attributes, so that looking in the index involves many fewer bytes than examining
the whole table. Often, the index is so much smaller that it can fit in main
memory, thereby saving even more disk accesses.
• When you think you are doing an index lookup, you are either searching the index
alone (a vertical partition of the base table), or you are searching the index, and
then joining the qualifying index rows to rows in the base table via the base-table
primary key (a bookmark lookup).
The central idea is that the spatial index gives you a small subset of the data. The index
tells you where to look and often carries some helpful search information with it (called
All
Objects
C oarse
Subset
Caref ul
Test
Good
False
included columns or covering columns by the
experts.) The selectivity of an index tells how
big this initial reduction is (the coarse subset
of Figure 1). When the subset is located, a
careful test examines each member of the
subset and discards false positives. That
process is indicated by the diamond in Figure
1. A good index has few false positives. We
use the Figure 1 metaphor (the coarse subset
and the careful test) throughout this article.
Figure 1: The key idea: the spatial index gives
you a small subset of the data (at least 100x
smaller) and then a careful test discards all
false positives. An index is good if there are
relatively few false positives. We use this idea
(the coarse subset and the careful test)
throughout this article.
B-trees and table-valued functions can be
combined as follows to let you build your own
spatial index that produces coarse subsets:
1. Create a function that generates keys
that cluster related data together. For
example, if items A and B are related, then the keys for A and B should be nearby
in the B-tree key space.
2. Create a table-valued function that, given a description of the subset of interest,
returns a list of key ranges (a “cover”) containing all the pertinent values.
You cannot always get every key to be near all its relatives because keys are sorted in one
dimension and relatives are near in two-dimensional space or higher. However, you can
come close. The ratio of false-positives to correct
answers is a measure of how well you are doing.
nwnw nwne nenw nene
The standard approach is to find some space filling curve
and thread the key space along that curve. Using the
standard Mercator map, for example, you can assign
everyone in the Northwest to the Northwest key range,
and assign everyone in the Southeast to the Southeast key
range. Figure 2 shows the 2nd order space-filling curve
that traverses all these quadrants, assigning keys in
sequence. Everyone in the Northwest-Southwest quadrant
has the key prefix nwsw. If you have an area like the
circle shown in Figure 2, you can look in the key range
key between ‘nwsw’ and ‘nwse’
This search space is eight times smaller than the whole
table and has about 75 percent false positives (indicated
by the area outside the circle but inside the two boxes).
This is not a great improvement, but it conveys the idea.
A better index would use a finer cell division. With fine
enough cells, the converging area could have very few
false positives. A detailed review of space-filling curves
and space-partitioning trees can be found in the books of
Hanan Samet [Samet].
nwsw nwse nesw nese
Figure 2: The start of a space
filling Peano or Hilbert curve
(one recursively divides each
cell in a systematic way.) The
cells are labeled. All points in
cell ‘nwse’ have a key with that
prefix so you can find them all
in the ‘nwse’ section of the Btree, right before ‘nwne’ and
right after ‘nwsw’. The circle is
an area of interest that overlaps
two such cells.
Now we are going to define a space-filling curve – the Hierarchical Triangular Mesh
(HTM) that works particularly well on the sphere. The earth is round and the celestial
sphere is round, so this spherical system is very convenient for geographers and
astronomers. We could do similar things for any metric space. The space-filling curve
gives keys that are the basis of the spatial index. Then, when someone has a region of
interest, our table valued function will give them a good set of key-ranges to look at (the
coarse filter of Figure 1). These key ranges will cover the region with spherical triangles,
called trixels, much as the two boxes in Figure 2 cover the circle. The search function
need only look at all the objects in the key ranges of these trixels to see if they qualify
(the careful test in Figure 1).
To make this concrete, assume we have a table of Objects
create table Object (
objID bigint primary key,
lat
float, -- latitude
lon
float, -– longitude
HtmID bigint)-– the HTM key
and a distance function dbo.fDistanceLatLon(lat1, lon1, lat2, lon2) that gives
the distance in nautical miles (arcminutes) between two points. Further assume that the
following table-valued function gives us the list of key ranges for HtmID points that are
within a certain radius of a lat-lon point.
define function
fHtmCoverCircleLatLon(@lat float, @lon float, @radius float)
returns @TrixelTable table(HtmIdStart bigint, HtmIdEnd bigint)
Then the following query finds points within 40 nautical miles of San Francisco (lat,lon)
= (37.8,-122.4):
select O.ObjID, dbo.fDistanceLatLon(O.lat,O.lon, 37.8, -122.4)
from fHtmCoverCircleLatLon(37.8, -122.4, 40) as TrixelTable
join Object O
on O.HtmID between TrixelTable.HtmIdStart
-- coarse test
and TrixelTable.HtmIdEnd
where dbo.fDistanceLatLon(lat,lon,37.8, -122.4) < 40
-- careful test
We now must define the HTM key generation function, the distance function, and the
HTM cover function. That’s what we do next using two United States Geological spatial
datasets as an example. If you are skeptical that this scales to billions of objects, go to
http://skyserver.sdss.org/ and look around the site. That Web site uses this same code to
do its spatial lookup on a multi-terabyte astronomy database.
This article is about how you use SQL Table Valued Functions and a space-filling curve
like the HTM to build a spatial index. As such, we treat the HTM code itself as a black
box documented elsewhere [Szalay], and we focus on how to adapt it to our needs within
an SQL application.
The Datasets
The US Geological Survey gathers and publishes data about the United States. Figure 3
shows the locations of 18,000 USGS-maintained stream gauges that measure river water
flows and levels. The USGS also publishes a list of 23,000 place names and their
populations.
Figure 3: Graphical display of the latitude and longitude (lat/lon) of USGS stream gauges and of
USGS places. These two datasets are about 20,000 items each and are about 4 MB in all. We use
them to motivate the spatial search examples.
USGS Populated Places (23,000 cities)
The USGS published a list of place names and some of their attributes in 1993. There are
newer lists at the USGS website but they are fragmented by state, so it is difficult to get a
nationwide list. The old list will suffice to demonstrate spatial indicies. The data has the
following format:
create table Place(
PlaceName
varchar(100)
State
char(2)
Population int
Households int
LandArea
int
WaterArea
int
Lat
float
Lon
float
HtmID
bigint
)
not
not
not
not
not
not
not
not
not
null, -- City name
null, -- 2 char state code
null, -- Number of residents (1990)
null, -- Number of homes (1990)
null, -- Area in sqare KM
null, -- water area within land area
null, -- latitude in decimal degrees
null, -- longitude decimal degrees
null primary key --spatial index key
To speed name lookups, we add a name index, but the data is clustered by the spatial key.
Nearby objects are co-located in the clustering B-tree and thus on the same or nearby disk
pages.
create index Place_Name on Place(PlaceName)
All except the HtmID data can be downloaded from the USGS Web site. The SQL Server
2005 data import wizard can be used to import the data (we have already done that in the
sample database.) The HtmID field is computed from the Lat Lon by:
update Place set HtmID = dbo.fHtmLatLon(lat, lon)
USGS Stream Gauges (17,000 instruments)
The USGS has been maintaining records of river flows since 1854. As of 1 Jan 2000,
they had accumulated over 430 thousand years of measurement data. About six thousand
active stations were active, and about four thousand were online. The gauges are
described in detail at http://waterdata.usgs.gov/nwis/rt. A NOAA site shows the data
from a few hundred of the most popular stations in a very convenient way:
http://weather.gov/rivers_tab.php.
Our database has just the stations in the continental United States (see Figure 3). There
are also stations in Guam, Alaska, Hawaii, Puerto Rico, and the Virgin Islands that are
not included in this database. The stream gauge station table is:
create table Station (
StationName
varchar(100)
State
char(2)
Lat
float
Lon
float
DrainageArea float
FirstYear
int
YearsRecorded int
IsActive
bit
IsRealTime
bit
StationNumber int
HtmID
bigint
not
not
not
not
not
not
not
not
not
not
not
null,
null,
null,
null,
null,
null,
null,
null,
null,
null,
null,
-------------
USGS Station Name
State location
Latitude in Decimal
Longitude in Decimal
Drainage Area (km2)
First Year operation
Record years (at Y2k)
Was it active at Y2k?
On Internet at Y2K?
USGS Station Number
HTM spatial key
(based on lat/lon)
primary key(htmID, StationNumber) )
As before, the HtmID field is computed from the Lat Lon fields by:
update Station set HtmID = dbo.fHtmLatLon(lat, lon)
There are up to 18 stations at one location, so the primary key must include the station
number to make it unique. However, the HTM key clusters all the nearby stations
together in the B-tree. To speed lookups, we add a station number and a name index:
create index Station_Name
on Station(StationName)
create index Station_Number on Station(StationNumber)
The Spatial Index Table
Now we are ready to create our spatial index. We could have added the fields to the base
tables, but to make the stored procedures work for many different tables, we found it
convenient to just mix all the objects together in one spatial index. You could choose
(type,HtmID) as the key to segregate the different types of objects; but, we chose
(HtmID, key) as the key so that nearby objects of all types (cities and steam gagues) are
clustered together. The spatial index is:
create table SpatialIndex (
HtmID
bigint
not null , -- HTM spatial key (based on lat/lon)
Lat
float
not null , -- Latitude in Decimal
Lon
float
not null , -- Longitude in Decimal
x
float
not null , -- Cartesian coordinates,
y
float
not null , -- derived from lat-lon
z
float
not null , --,
Type
char(1) not null , -- place (P) or gauge (G)
ObjID
bigint
not null , -- object ID in table
primary key (HtmID, ObjID) )
The Cartesian coordinates will be explained later in this topic. For now, it is enough to
say that the function fHtmCenterPoint(HtmID) returns the Cartesian (x,y,z) unit vector
for the centerpoint of that HTM triangle. This is the limit point of the HTM, as the center
is subdivided to infinitely small trixels.
The SpatialIndex table is populated from the Place and Station tables as follows:
insert SpatialIndex
select
P.HtmID, Lat, Lon, XYZ.x, XYZ.y, XYZ.z,
'P' as type, P. HtmID as ObjID
from
Place P cross apply fHtmLatLonToXyz(P.lat, P.lon)XYZ
insert SpatialIndex
select S.HtmID, Lat, Lon, XYZ.x, XYZ.y, XYZ.z,
'S' as type, S.StationNumber as ObjID
from
Station S cross apply fHtmLatLonToXyz(S.lat, S.lon) XYZ
To clean up the database, we execute:
DBCC
DBCC
DBCC
DBCC
DBCC
DBCC
DBCC
INDEXDEFRAG
INDEXDEFRAG
INDEXDEFRAG
INDEXDEFRAG
INDEXDEFRAG
INDEXDEFRAG
SHRINKDATABASE
(
(
(
(
(
(
(
spatial
spatial
spatial
spatial
spatial
spatial
spatial
,
,
,
,
,
,
,
Station, 1)
Station, Station_Name)
Station, Station_Number)
Place,
1)
Place,
Place_Name)
SpatialIndex, 1)
1 ) -- 1% spare space
A Digression: Cartesian Coordinates
You can skip this if you like. It is not needed to use the library. The HTM code heavily
uses a trick to avoid spherical geometry: it moves from the 2D surface of the sphere to
3D. This allows very quick tests for “inside a polygon” and for “nearby a point” queries.
Every lat/lon point on the sphere can be
represented by a unit vector in threedimensional space v = (x,y,z). The north and
south poles (90° and -90°) are v = (0,0,1), and
v = (0,0,-1) respectively. Z represents the axis
of rotation, and the XZ plane represenst the
Prime (Greenwich) Meridian ,having
longitude 0° or longitude 180°. The formal
definitions are:
x = cos(lat)cos(lon)
y =cos(lat)sin(lon)
z = sin(lat)
z
Lat = 90
x y z = 0,0,1
x
Lat= 0
Lon = 0
x y z = 1,0,0
y
Lat = -90 = x y z = 0,0,-1
Lon = 90
Lat = 0
x y z = 0,1,0
Figure 4: Cartesian coordinates allow
quick tests for point-in-polygon and
point-near-point. Each lat/lon point has a
corresponding (x,y,z) unit vector.
These Cartesian coordiates are used as follows.
Given two points on the unit sphere, p1=(x1,y1,z1)
and p2 = (x2,y2,z2), then their dot product, p1•p2 = x1*x2+y1*y2+z1*z2, is the cosine of the
angle between these two points. It is a distance metric.
If we are looking for points within 45 nautical miles (arc minutes) of point p1, that is at
most 45/60 degrees away from p1. The dot product of such points with p1 will be less
than d=cos(radians(45/60). The “is nearby” test becomes { p2 | p2•p1 < d}, which is a
very quick test.
Cartesian coordinates also allow a
quick test for point-inside-polygon.
All our polygons have great-circle or
cos θ
small-circle edges. Such edges lie
θ
along a plane intersecting the sphere.
So the edges can be defined by the
unit vector, v, normal to the plane
Figure 5: Each great or small circle is the intersection
and by a shift along that vector. For
of a plane with the circle. A point is inside the circle if
its dot product with the plane’s normal vector is less
example, the equator is the vector v
= (0,0,1) and shift zero. Latitude 60° than cos(θ) where 2θ is the circle’s arc-angle diameter.
is defined by vector v = (0,0,1) with
a shift of 0.5, and a 60° circle around Baltimore is defined by vector v = (0.179195, 0.752798, 0.633392) with a shift of 0.5. A place, p2, is within 60° of Baltimore if p2•v
< 0.5. The same idea lets us decide if a point is inside or outside a HTM triangle by
evaluating three such dot products. That is one of the main reasons the HTM code is so
efficient and fast.
We have implemented several helper procedures to convert from LatLon to Cartesian
coordiantes:
fHtmXyz(HtmID) returns the xyz vector of the centerpoint of an HtmID
fHtmLatLonToXyz(lat,lon) returns an xyz vector
fHtmXyzToLatLon(x,y,z) returns a lat,lon vector.
They are used below and documented in the the API spec and Intellisense [Fekete].
The library here defaults to 21-deep HTM keys (the first level divides the sphere into 8
faces and each subsequent level divides the speherical triangle into 4-sub-triangles.) The
table below indicates that a 21-deep trixel is fairly small. The code can be modified to go
31-deep deep before the 64-bit representation runs out of bits.
Table 1: Each HTM level subdivdes the sphere. For each level, this table shows the area in
square degrees, arc minutes, arc seconds, and meters. The Trixel colum shows some charactic
sizes: the default 21-deep trixels is about .3 arc second2. The USGS data has about ½ object per
12-deep trixel.
HTM
depth
sphere
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
deg
2
41253
5157
1289
322
81
20
5
1
3E-1
8E-2
2E-2
5E-3
1E-3
3E-4
8E-5
2E-5
5E-6
1E-6
3E-7
8E-8
2E-8
5E-9
1E-9
3E-10
7E-11
2E-11
5E-12
1E-12
arc min
2
148,510,800
18,563,850
4,640,963
1,160,241
290,060
72,515
18,129
4,532
1,133
283
71
18
4
1
3E-1
7E-2
2E-2
4E-3
1E-3
3E-4
7E-5
2E-5
4E-6
1E-6
3E-7
7E-8
2E-8
4E-9
Area
arc sec2
534,638,880,000
66,829,860,000
16,707,465,000
4,176,866,250
1,044,216,563
261,054,141
65,263,535
16,315,884
4,078,971
1,019,743
254,936
63,734
15,933
3,983
996
249
62
16
4
1
2E-1
6E-2
2E-2
4E-3
9E-4
2E-4
6E-5
1E-5
2
earth m
5.E+14
6E+13
2E+13
4E+12
1E+12
2E+11
6E+10
2E+10
4E+9
1E+9
2E+8
6E+7
2E+7
4E+6
943816
235954
58989
14747
3687
922
230
58
14
4
1
2E-1
6E-2
1E-2
trixel
1 deg2
1 amin2
1 asec2
1 km2
1 m2
objects / trixel
SDSS USGS
3E+8
8E+7
2E+7
5E+6
1E+6
3E+5
73242
18311
4578
1144
286
72
18
4
1
0.3
.
30,000
7,500
1,875
468
117
29
7
2
0.5
0.1
Typical Queries
Assuming we can get the functions defined, we are ready to do a few queries.
1. Find Points near point: find towns near a place.
The most common query is to find all places nearby a certain place or point. Consider
the query, “Find all towns within 100 nautical miles of Baltimore Maryland.” The HTM
triangles covering a 100 nautical mile circle (100 arc minutes from) Baltimore are
obtained by
select *
-- find a HTM cover 100 NM around Baltimore
from fHtmCoverCircleLatLon(39.3, -76.6, 100)
This returns the Trixel Table at right. That is, the
The Baltimore circle HTM cover
fHtmCoverCircleLatLon() function returns a set of
HtmIdStart
HtmIdEnd
HTM triangles that “cover” the circle (in this case, a
14023068221440 14027363188735
single trixel). The HTM keys of all objects inside the
circle are also inside one of these triangles. Now we need to look in all those triangles
and discard the false positives (the careful test of Figure 1). We will order the answer set
by the distance from Baltimore, so that if we want the closest place, we can just select the
TOP 1 WHERE distance > 0 (we want to exclude Baltimore itself from being closest).
declare @lat float, @lon float
select @lat = lat, @lon = lon
from Place
where Place.PlaceName = 'Baltimore'
and State = 'MD'
select ObjID, dbo.fDistanceLatLon(@lat,@lon, lat, lon) as distance
from SpatialIndex join fHtmCoverCircleLatLon(@lat, @lon, 100)
On HtmID between HtmIdStart and HtmIdEnd
-- coarse test
and type = 'P'
and dbo.fDistanceLatLon(@lat,@lon, lat, lon) < 100 -- careful test
order by distance asc
The cover join returns 2,928 rows (the coarse test); 1,122 of them are within 100 air miles
(the careful test). This gives us 61% false positives – all within 9 milliseconds.
These are such common tasks that there are standard functions for them:
fHtmNearbyLatLon(type, lat, lon, radius)
fHtmNearestLatLon(type, lat, lon)
so the query above becomes:
select ObjID, distance
from fHtmNearestLatLon('P', 39.3, -76.61)
2. Find places inside a box.
Applications often want to find all the objects inside a square view-port when displaying
a square map or window. Colorado is almost exactly square with corner points (41N, 109°3’W) in the NW corner and (37°N-102° 3’E) in the SW corner. The state’s center
point is (39°N, -105°33’E) so one can cover that square with a circle centered at that
point.
declare @radius float
set @radius = dbo.fDistanceLatLon(41,-109.55,37,-102.05)/2
select *
from Station
where StationNumber in (
select ObjID
from fHtmCoverCircleLatLon(39, -105.55, @radius) join SpatialIndex
on HtmID between HtmIdStart and HtmIdEnd
and lat between 37 and 41
and lon between -109.05 and -102.048
and type = 'S')
OPTION(FORCE ORDER)
This example returns 1,030 stream gauges in about 46 milliseconds. Five other Colorado
gauges are right on the border that wanders south of 37° by up to 1 nautical mile. These
extra 5 stations appear when the southern latitude is adjusted from 37° to 36.98°1. The
cover circle returns 36 triangles. The join with the SpatialIndex table returns 1,975
gauges. That’s 47 percent false positives. The next section shows how to improve on
this by using the HTM regions to specify a polygon cover rather than a cover for a circle.
The FORCE ORDER clause is an embarrassment – if missing, the query runs ten times
longer because the optimizer does a nested-loops join with the spatial index as the outer
table. Perhaps if the tables were larger (millions of rows), the optimizer would pick a
different plan, but we cannot count on that. Paradoxically, the optimizer chose the
correct plan without any hints for all the queries in the previous section.
1
GIS systems and astronomical applications often want a buffer zone around a region. The HTM code
includes support for buffer zones, and they are much used in real applications, Look at reference [Szalay] to
see how this is done.
3. Find places inside a polygon.
The HTM code lets us specify the area as a circle, a rectangle, a convex hull, or a union
of these regions. In particular, the HTM library allows you to specify a region using the
following linear syntax:
circleSpec
rectSpec
hullSpec
convexSpec
areaSpec
regionSpec
:=
|
|
:=
|
|
:=
|
|
:=
:=
:=
‘CIRCLE LATLON ’
lat lon radius
‘CIRCLE J2000 ’
ra dec radius
‘CIRCLE [CARTESIAN ]’ x y z
radius
‘RECT LATLON ’
{ lat lon }2
‘RECT J2000 ’
{ ra dec }2
‘RECT [CARTESIAN ]’
{ x y z
}2
‘CHULL LATLON ’
{ lon lat }3+
‘CHULL J2000 ’
{ ra dec }3+
‘CHULL [CARTESIAN ]’ { x y z
}3+
‘CONVEX ’ [ ‘CARTESIAN ’] { x y z D }*
rectSpec | circleSpec | hullSpec | convexSpec
‘REGION ’ {areaSpec}* | areaSpec
To give examples of region specifications:
CIRCLE
A point specification and a 1.75 nautical mile (arc minute) radius.
'CIRCLE LATLON 39.3 -76.61 100'
'CIRCLE CARTESIAN 0.1792 -0.7528 0.6334 100'
RECT
Two corner points defining the minimum and maximum of the lat, lon. The
longitude coordinates are interpreted in the wrap-around sense, i.e.,
lonmin=358.0 and lonmax=2.0, is a 4 degree wide range. The latitudes must be
between the North and South Pole. The rectangle edges are constant latitude
and longitude lines, rather than the great-circle edges of CHULL and CONVEX.
'RECT LATLON 37 -109.55
CHULL
41 -102.05'
Three or more point specifications define a spherical convex hull with edges
of the convex hull connecting adjacent points by great circles. The points
must be in a single hemisphere, otherwise an error is returned. The order of
the points is irrelevant.
'CHULL LATLON 37 -109.55 41 -109.55 41 -102.051 37 -102.05'
CONVEX
Any number (including zero) of constraints in the form of a Cartesian vector
(x,y,z) and a fraction of the unit length of the vector.
'CONVEX
REGION
-0.17886 -0.63204 -0.75401 0.00000
-0.97797 0.20865 -0.00015 0.00000
0.16409 0.57987 0.79801 0.00000
0.94235 -0.33463 0.00000 0.00000'
A region is the union of zero or more circle, rect, chull, and convex areas.
‘REGION CONVEX 0.7 0.7 0.0 –0.5 CIRCLE LATLON 18.2 –22.4 1.75’
Any of these region descriptions can be fed to the fHtmCoverRegion() routine that
returns a trixel table describing a set of trixels (triangular areas) covering that region. The
simpler code for the Colorado query is:
select S.*
from ( select ObjID
from fHtmCoverRegion('RECT LATLON 37 -109.55
loop join SpatialIndex
on HtmID between HtmIdStart and HtmIdEnd
and lat between 37 and 41
and lon between -109.05 and -102.048
and type = 'S') as G
join Station S on G.objID = S.StationNumber
OPTION( FORCE ORDER)
41 -102.05')
This unusual query format is required to tell the optimizer exactly the order in which to
perform the join (to make the “force order” option work correctly). It is difficult to
modify the optimizer in this way, but until table-valued functions have statistics, they are
estimated to be very expensive. You have to force them into the inner loop join.
The query returns 1030 stream gauges and has 1,365 candidates from the cover, so there
are 25 percent false positives. Note that the rectangle cover is better than the circular
cover, which had 61% false positives. There is polygon syntax for non-rectangular states,
but this article is about table valued functions, not about the HTM algorithms. You can
see the HTM code in the project, and also in the documentation for the project.
A similar query can be cast as a convex hull as:
select S.*
from ( select ObjID
from fHtmCoverRegion(
'CHULL LATLON 37 -109.55 41 -109.55 41 -102.05 37 -102.05')
loop join SpatialIndex
on HtmID between HtmIdStart and HtmIdEnd
and lat between 37 and 41
and lon between -109.05 and -102.048
and type = 'S') as G
join Station S on G.objID = S.StationNumber
OPTION( FORCE ORDER)
The query returns 1030 stream gauges and has 1,193 candidates from the cover, so there
are 14 percent false positives. The convex hull cover is even better than the equivalent
rectangular cover in this case.
4. Advanced topics – complex regions.
The previous examples gave the syntax for regions and a discussion of point-near-point
and point-in-rectangle searches. Regions can get quite complex. They are Boolean
combinations of convex areas. We do not have the space here to explain regions in
detail, but the HTM library in the accompanying project has the logic to do Boolean
combinations of regions, simplify regions, compute region corner points, compute region
areas, and has many other features. Those ideas are described in [Fekete], [Gray], and
[Szalay].
To give a hint of these ideas, consider the state of Utah. Its boundaries are approximately
defined by the union of two rectangles:
declare @utahRegion varchar(max)
set @utahRegion = 'region '
+ 'rect latlon 37 -114.0475 41 -109.0475 ' -- main part
+ 'rect latlon 41 -114.0475 42 -111.01 ' -- Ogden & Salt Lake.
Now we can find all stream gauges in Utah with the query:
select S.*
from (
select ObjID
from fHtmCoverRegion(@utahRegion)
loop join SpatialIndex
on HtmID between HtmIdStart and HtmIdEnd
and (((
lat between 37
and
41)
and (lon between -114.0475 and -109.04))
or ((
lat between 41
and
42)
and (lon between -114.0475 and -111.01))
)
and type = 'S' ) as G
join Station S on G.objID = S.StationNumber
OPTION( FORCE ORDER)
-----
careful test
are we inside
one of the two
boxes?
The cover returns 38 trixels. The join returns 775 stations. The careful test finds 670
stations in Utah, and two Wyoming stations that are right on the border (14 percent false
positives).
Most states require much more complex regions. For example, a region string to
approximate California is:
declare @californiaRegion varchar(max)
set @californiaRegion = 'region '
+ 'rect latlon 39
-125 '
+ '42
-120 '
+ 'chull latlon 39
-124 '
+ '39
-120 '
+ '35
-114.6 '
+ '34.3 -114.1 '
+ '32.74 -114.5 '
+ '32.53 -117.1 '
+ '33.2 -119.5 '
+ '34
-120.5 '
+ '34.57 -120.65 '
+ '36.3 -121.9 '
+ '36.6 -122.0 '
+ '38
-123.03 '
select stationNumber
from fHtmCoverRegion(@californiaRegion)
loop join SpatialIndex
on HtmID between HtmIdStart and HtmIdEnd
/* and <careful test> */
and type = 'S'
join Station S on objID = S.StationNumber
OPTION( FORCE ORDER)
---------------
nortwest corner
center of Lake Tahoe
Pt. Arena
Lake tahoe.
start Colorado River
Lake Havasu
Yuma
San Diego
San Nicholas Is
San Miguel Is.
Pt. Arguelo
Pt. Sur
Monterey
Pt. Rayes
The cover returns 108 trixels, which cover 2,132 stations. Of these, 1,928 are inside
California, so the false positives are about 5 percent -- but the careful test is nontrivial.
That same query, done for places rather than stations, with the careful test included, looks
like this:
select *
from Place
where HtmID in
( select distinct SI.objID
from fHtmCoverRegion(@californiaRegion)
loop join SpatialIndex SI
on SI.HtmID between HtmIdStart and HtmIdEnd
and SI.type = 'P'
join place P on SI.objID = P.HtmID
cross join fHtmRegionToTable(@californiaRegion) Poly
group by SI.objID, Poly.convexID
having min(SI.x*Poly.x + SI.y*Poly.y + SI.z*Poly.z - Poly.d) >= 0
)
OPTION( FORCE ORDER)
This uses the convex-halfspace representation of California and the techniques described
in [Gray] to quickly test if a point is inside the California convex hull. It returns 885
places, seven of which are on the Arizona border with California (the polygon
approximates California). It runs in 0.249 seconds on a 1GHz processor. If you leave off
the “OPTION( FORCE ORDER)” clause it runs slower, taking 247 seconds.
Because this is such a common requirement, and because the code is so tricky, we added
a procedure fHtmRegionObjects(Region,Type) that returns object IDs from
SpatialIndex. This procedure encapsulates the tricky code above, so the two California
queries become:
select *
-- Get all the California River Stations
from Station
where stationNumber in -- that are inside the region
(select ObjID
from fHtmRegionObjects(@californiaRegion,'S'))
select *
-- Get all the California Cities
from Place
where HtmID in
-- that are inside the region
(select ObjID
from fHtmRegionObjects(@californiaRegion,'P'))
The Colorado and Utah queries are also simplified by using this routine.
4. Summary
The HTM spatial indexing library presented here is interesting and useful in its own right.
It is a convenient way to index data for point-in-polygon queries on the sphere. But, the
library is also a good example of how SQL Server and other database systems can be
extended by adding a class library that does substantial computation in a language like
C#, C++, Visual Basic, or Java. The ability to implement powerful table-valued
functions and scalar functions and integrate these queries and their persistent data into the
database is a very powerful extension mechanism that starts to deliver on the promise of
Object-Relational databases. This is just a first step. In the next decade, programming
languages and database query languages are likely to get even better data integration.
This will be a boon to application developers.
References
[Gray] “There Goes the Neighborhood: Relational Algebra for Spatial Data Search”, Jim
Gray, Alexander S. Szalay, Gyorgy Fekete, Wil O’Mullane, Maria A. NietoSantisteban, Aniruddha R. Thakar, Gerd Heber, Arnold H. Rots, MSR-TR-200432, April 2004
[Szalay] “Indexing the Sphere with the Hierarchical Triangular Mesh”, Alexander S.
Szalay, Jim Gray, George Fekete, Peter Z. Kunszt, Peter Kukol, Aniruddha R.
Thakar, To appear, included in this project.
[Fekete] “SQL SERVER 2005 HTM Interface Release 4” George Fekete, Jim Gray,
Alexander S. Szalay, May 15, 2005, included in this project.
[Samet1] Applications of Spatial Data Structures: Computer Graphics, Image
Processing, and GIS, Hanan Samet, Addison-Wesley, Reading, MA, 1990.
ISBN0-201-50300-0.
[Samet2] The Design and Analysis of Spatial Data Structures, Hanan Samet, AddisonWesley, Reading, MA, 1990. ISBN 0-201-50255-0.
Appendix: The Basic HTM Routines
This section describes the HTM routines. The companion document [Szalay] has a
manual page for each routine, and the routines themselves are annotated to support
Intellisense.
In what follows, lat and lon are in decimal degrees (southern and western latitudes are
negative), and distances are in nautical miles (arc minutes.)
HTM library version: fHtmVersion() returns versionString
The routine returns an nvarchar(max) string giving the HTM library version.
Example use:
print dbo.fHtmVersion()
Returns something like:
‘C# HTM.DLL V.1.0.0 1 August
2005 ’
Generating HTM keys: fHtmLatLon (lat, lon) returns HtmID
The routine returns the 21-deep HTM ID of that LatLon point.
Example use:
There
update Place set HtmID = dbo.fHtmLatLon(Lat,Lon)
are also fHtmXyz() and fHtmEq() functions for astronomers.
LatLon to XYZ: fHtmLatLonToXyz (lat,lon) returns Point (x, y, z)
The routine returns the Cartesian coordinates of that Lat Lon point.
Example use (this is the identity function):
Select LatLon.lat, LatLon.lon-360
from fHtmLatLonToXyz(37.4,-122.4) as XYZ cross apply
fHtmXyzToLatLon(XYZ.x, XYZ.y, XYZ.z) as LatLon
There is also an fHtmEqToXyz() functions for astronomers.
XYZ to LatLon: fHtmXyzToLatLon (x,y,z) returns Point (lat, lon)
The routine returns the Cartesian coordinates of that Lat Lon point.
Example use (this is the identity function):
Select LatLon.lat, LatLon.lon-360
from fHtmLatLonToXyz(37.4,-122.4) as XYZ cross apply
fHtmXyzToLatLon(XYZ.x, XYZ.y, XYZ.z) as LatLon
There is also an fHtmXyzToEq() functions for astronomers.
Viewing HTM keys: fHtmToString (HtmID) returns HtmString
Given an HtmID, the routine returns a nvarchar(32) in the form [N|S]t1t2t3…tn where each
triangle number ti is in {0,1,2,3} describing the HTM trixel at that depth of the triangular
mesh. .
Example use:
print 'SQL Server development is at: ' +
dbo.fHtmToString(dbo.fHtmLatLon(47.646,-122.123))
which returns: ‘N132130231002222332302’.
There are also fHtmXyz() and fHtmEq() functions for astronomers.
HTM trixel Centerpoint: fHtmToCenterpoint(HtmId) returns Point (x, y, z)
Returns the Cartesian center point of the HTM trixel specified by the HtmID.
Example use:
select * from fHtmToCenterPoint(dbo.fHtmLatLon(47.646,-122.123))
HTM trixel corner points: fHtmToCornerpoints(HtmId) returns Point (x, y, z)
Returns the three Cartesian corner points of the HTM trixel specified by the HtmID.
Example use:
select * from fHtmToCornerPoints(dbo.fHtmLatLon(47.646,-122.123))
Computing distances: fDistanceLatLon(lat1, lon1, lat2, lon2) returns distance
Computes the distance, in nautical miles (arc minutes) between two points.
Example use:
declare @lat float, @lon float
select @lat = lat, @lon = lon
from Place
where PlaceName = 'Baltimore' and State = 'MD'
select PlaceName,
dbo.fDistanceLatLon(@lat,@lon, lat, lon) as distance
from Place
There are also fDistanceXyz() and fDistanceEq() functions for astronomers.
The following routines return a table which serves as a spatial index. The returned spatial
index table has the data definition:
SpatialIndexTable table (
HtmID
bigint
not null , -- HTM spatial key (based on lat/lon)
Lat
float
not null , -- Latitude in Decimal
Lon
float
not null , -- Longitude in Decimal
x
float
not null , -- Cartesian coordinates,
y
float
not null , -- derived from lat-lon
z
float
not null , --,
Type
char(1) not null , -- place (P) or gauge (G)
ObjID
bigint
not null , -- object ID in table
distance float
not null , -- distance in arc minutes to object
primary key (HtmID, ObjID) )
Finding nearby objects: fHtmNearbyLatLon(type, lat, lon, radius) returns
SpatialIndexTable
Returns a list of objects within the radius distance of the given type and their distance
from the given point. The list is sorted by nearest object.
Example use:
select distance, Place.*
from fHtmNearbyLatLon('P', 39.3, -76.6, 10) I join Place
on I.objID = Place.HtmID
order by distance
There are also fHtmGetNearbyEq () and fHtmGetNearbyXYZ() functions
astronomers.
for
Finding the nearest object: fHtmNearestLatLon(type, lat, lon) returns SpatialIndexTable
Returns a list containing the nearest object of the given type to that point.
Example use:
select distance, Place.*
from fHtmNearestLatLon('P', 39.3, -76.6) I join Place
on I.objID = Place.HtmID
There are also fHtmGetNearestEq () and fHtmGetNearestXYZ() functions
astronomers.
for
The following routines return a table describing the HtmIdStart and HtmIdEnd of a set of
trixels (HTM triangles) covering the area of interest. The table definition is:
TrixelTable table (
HtmIdStart
HtmIdEnd
)
bigint not null , -- min HtmID in trixel
bigint not null
-- max HtmID in trixel
Circular region HTM cover: fHtmCoverCircleLatLon(lat, lon, radius) returns trixelTable
Returns a trixel table covering the designated circle.
Example use:
declare @answer nvarchar(max)
declare @lat float, @lon float
select @lat = lat, @lon = lon
from Place
where Place.PlaceName = 'Baltimore'
and State = 'MD'
set @answer = ' using fHtmCoverCircleLatLon() it finds:
'
select @answer = @answer
+ cast(P.placeName as varchar(max)) + ', '
+ str( dbo.fDistanceLatLon(@lat,@lon, I.lat, I.lon) ,4,2)
+ ' arcmintes distant.'
from SpatialIndex I join fHtmCoverCircleLatLon(@lat, @lon, 5)
On HtmID between HtmIdStart and HtmIdEnd -- coarse test
and type = 'P'
-- it is a place
and dbo.fDistanceLatLon(@lat,@lon, lat, lon)
between 0.1 and 5 -- careful test
join Place P on I.objID = P.HtmID
order by dbo.fDistanceLatLon(@lat,@lon, I.lat, I.lon) asc
print 'The city within 5 arcminutes of Baltimore is: '
+ 'Lansdowne-Baltimore Highlands, 4.37 arcminutes away'
There are also fHtmCoverCircleEq() for astronomers.
General region specification to HTM cover: fHtmCoverRegion(region) returns trixelTable
Returns a trixel table covering the designated region (regions are described earlier in this
topic).
select S.*
from ( select ObjID
from fHtmCoverRegion('RECT LATLON 37 -109.55
loop join SpatialIndex
on HtmID between HtmIdStart and HtmIdEnd
and lat between 37 and 41
and lon between -109.05 and -102.048
and type = 'S') as G
join Station S on G.objID = S.StationNumber
41 -102.05')
OPTION( FORCE ORDER)
General region simplification: fHtmRegionToNormalFormString(region) returns
regionString
Returns a string of the form REGION {CONVEX {x y z d}* }* where redundant
halfspaces have been removed from each convex; the convex has been simplified as
described in [Fekete]
print dbo.fHtmToNormalForm('RECT LATLON 37 -109.55
41 -102.05')
The following routine returns a table describing the HtmIdStart and HtmIdEnd of a set of
trixels (HTM triangles) covering the area of interest. The table definition is:
RegionTable ( convexID
bigint not null ,
halfSpaceID bigint not null
x
y
z
d
)
float
float
float
float
not
not
not
not
null
null
null
null
---------
ID of the convex, 0,1,…
ID of the halfspace
within convex, 0,1,2,
Cartesian coordinates of
unit-normal-vector of
halfspace plane
displacement of halfspace
along unit vector [-1..1]
Cast RegionString as Table: fHtmRegionToTable(region) returns RegionTable
Returns a table describing the region as a union of convexes, where each convex is the
intersection of the x,y,z,d halfspaces. The convexes have been simplified as described in
[Fekete]. Section 4 of this article describes the use of this function.
select *
from dbo.fHtmToNormalForm('RECT LATLON 37 -109.55
41 -102.05')
Find Points Inside a Region: fHtmRegionObjects(region, type) returns ObjectTable
Returns a table containing the objectIDs of objects in SpatialIndex that have the
designated type and are inside the region.
select *
-- find Colorado places.
from Places join
where HtmID in
select objID
from dbo. fHtmRegionObjects('RECT LATLON 37 -109.55
41 -102.05',‘P’)
General region diagnostic: fHtmRegionError(region ) returns message
Returns “OK” if region definition is valid; otherwise, returns a diagnostic saying what is
wrong with the region definition followed by a syntax definition of regions.
print dbo.fHtmRegionError ('RECT LATLON 37 -109.55
41 -102.05')
SQL SERVER 2005
HTM Interface Release 4
Alex Szalay, Gyorgy Fekete, Jim Gray, August 2005
This document describes the SQL Server 2005 interfaces to the C# HTM functions. It also explains how to
install and modify those procedures. A tutorial on the Hierarchical Triangular Mesh (HTM) is at
http://www.sdss.jhu.edu/htm/.
Table of Contents
HTM Concepts Reviewed .............................................................................................................................. 2
Find Version of Installed HTM Code............................................................................................................. 3
fHtmVersion()....................................................................................................................................... 3
Geometric Conversion Functions ................................................................................................................... 3
fHtmXyzToLatLon(x, y, z) ................................................................................................................... 3
fHtmXyzToEq(x, y, z) .......................................................................................................................... 3
fHtmLatLonToXyz(lat, lon).................................................................................................................. 3
fHtmEqToXyz(ra, dec) ......................................................................................................................... 4
Compute HtmID for a point............................................................................................................................ 4
fHtmLatLon (lat, lon)............................................................................................................................ 4
fHtmEq(ra, dec) .................................................................................................................................... 4
fHtmXyz(x,y,z) ..................................................................................................................................... 5
HTM Triangle Functions ................................................................................................................................ 6
fHtmToCenterPoint (HtmID)................................................................................................................ 6
fHtmToCornerPoints (HtmID).............................................................................................................. 6
Distances Between Points............................................................................................................................... 6
fDistanceLatLon( lat1, lon1, lat2, lon2)................................................................................................ 6
fDistanceEq( ra1, dec1, ra2, dec2) ........................................................................................................ 6
fDistanceXyz( x1,y1,z1, x2,y2,z2)........................................................................................................ 7
REGIONS....................................................................................................................................................... 8
Region Specifications ........................................................................................................................... 8
Examples of Region Specifications ...................................................................................................... 8
fHtmRegionToNormalFormString (regionSpec) ................................................................................ 10
fHtmRegionToTable (regionSpec)...................................................................................................... 10
fHtmRegionObjects (regionSpec, type) .............................................................................................. 10
fHtmRegionError(regionSpec)............................................................................................................ 11
HTM Covers: Compute HtmID Ranges For a Region................................................................................ 12
fHtmCoverRegion (regionSpec) ......................................................................................................... 12
fHtmCoverCircleLatLon(lat, lon, radiusArcMin) ............................................................................... 13
fHtmCoverCircleEq(ra, dec, radiusArcMin)....................................................................................... 13
fHtmCoverCircleXyz(x, y, z, radiusArcMin) ..................................................................................... 13
Installing the HTM Code.............................................................................................................................. 14
Compiling, Modifying, and Debugging the HTM Code .............................................................................. 14
Table of HTM Depths And Approximate Areas .......................................................................................... 15
HTM Concepts Reviewed
The Hierarchical Triangular Mesh is a multilevel, recursive decomposition of the sphere. At the top, depth
1, there are eight spherical triangles, four each for the Northern and Southern hemispheres. Four of these
triangles share a vertex at the pole. The sides opposite the pole form the equator. You can imagine these by
orienting a regular octahedron so that two of its vertices are at the poles, and the other four are equally
spaced on the equator. The spherical polygons are the projection of the edges of the octahedron onto the
circumscribing sphere. There are eight unique integers that represent these triangles.
Triangles in the mesh scheme are called trixels. Each trixel can be split into four smaller trixels by
introducing new vertices at the midpoints of each side, and adding a great circle arc segment to connect the
new vertices with the existing one. Trixel division repeats recursively and indefinitely to produce smaller
and smaller trixels. Each trixel has a level number that corresponds to the number of times an original
(octant) triangle had to be split. Points in this decomposition are represented by a leading 1 bit and then the
level 0 trixel number [0..7] and then the successive sub-trixel numbers [0..3]. This gives each trixel and its
center-point a unique 64 bit identifier, called an HTM ID (HtmID) that represents a particular trixel in the
HTM hierarchy. The smallest valid HtmID is 8 – being the level 0 HtmID of the triangle 0. HtmIDs are
numbered from level 0. The term depth tells how many levels are involved: depth = level+1.
Though the division process can continue indefinitely, the 64-bit representation runs out of bits at depth 31.
Depth 25 is good enough for most applications—about 0.6 meter on the surface of the Earth or 0.02 arc
seconds. The code here defaults to depth 21 (0.3 arc seconds). Note that this numbering scheme is not a
complete cover on the positive integers, and not all bit patterns form valid HtmID numbers.
Some of the functions described here return a region, or area, on the sphere. The return value of these
functions are referred to as trixel tables, which are tables of HtmID ranges for trixels that overlap or cover
the region. The structure of these tables is quite simple: they are rows of two 64-bit numbers (BIGINTs)
[Htm_start, Htm_stop] that are the starting and ending values of HtmIDs of trixels in the range. The
HtmIDs in this context are always for trixels at depth 21.
The library supports three coordinate systems:
1. LatLon is the Greenwich Meridian spherical coordinate system of latitude and longitude (lat,
lon) used by geographers.
2. J2000 or Equatorial is the celestial right ascension and declination (ra, dec) spherical
coordinate system used by astronomers (the vector pointing at the center of the Milky Way
defines the intersection of the J2000 prime meridian with the J2000 equator).
3. Cartesian is a the unit vector representation of a sphere; a point on the sphere (either LatLon
or J2000) has a corresponding unit vector. The North Pole is (0,0,1) and the prime meridian
intersection with the equator is (1,0,0).
This library is installed in the sample Spatial database, which uses the LatLon coordinates. Many of the
examples are geared to that. The astronomy examples are drawn from tables at http://skyserver.sdss.org/.
Find Version of Installed HTM Code
fHtmVersion()
Returns a string describing version of HTM code. A typical description has a version and timestamp:
“C# HTM.DLL V.1.0.0; 2005 July 30”.
Returns:
version varchar(max)
version string of installed code
Example use:
declare @version varchar(max)
select @version = dbo.fHtmVersion()
print 'Installed version is: ' + @version
Produces:
Installed version is: C# HTM.DLL V.1.0.0; 2005 July 30
Geometric Conversion Functions
fHtmXyzToLatLon(x, y, z)
Given the Cartesian coordinates (x, y, z), returns a table containing the corresponding LatLon coordinates.
Parameters:
x: float not null
x
y: float not null
y
z: float not null
z
Returns:
VertexTable(lat float, lon float)
The lat/lon equivalent of the given point.
Example use:
Select * from dbo.fHtmXyzToLatLon(1.0, 0.0, 0.0)
Errors: Empty table is returned if (x, y, z) is too close to (0, 0, 0) (within 1e9 of zero).
fHtmXyzToEq(x, y, z)
Given the Cartesian coordinates (x, y, z), returns a table containing the corresponding J2000 (ra, dec)
coordinates.
Parameters:
x: float not null
x
y: float not null
y
z: float not null
z
Returns:
VertexTable(ra float, dec float)
The RA/DEC equivalent of the given point.
Example use:
Select * from dbo.fHtmXyzToEq(1.0, 0.0, 0.0)
Errors: Empty table is returned, if (x, y, z) is too close to (0, 0, 0) (within 1e9 of zero).
fHtmLatLonToXyz(lat, lon)
Given a LatLon point, returns a table containing the corresponding Cartesian coordinates (x, y, z).
Parameters:
lat: float not null
latitude
lon:
float not null
longitude
Returns:
VertexTable(x float, y float, z float)
Example use:
Select * from dbo.fHtmLatLonToXyz(0.0, 0.0)
Errors: None. Extreme latitude values are truncated to [-90 … 90]
fHtmEqToXyz(ra, dec)
Given a J2000 (equatorial) point returns a table containing the corresponding Cartesian coordinate (x, y, z).
Parameters:
ra: float not null
right ascension
dec: float not null
declination
Returns:
VertexTable(x float, y float, z float)
Example use:
Select * from dbo.fHtmEqToXyz(0.0, 0.0)
Errors: None. Extreme declination values are truncated to [-90 … 90]
Compute HtmID for a point
fHtmLatLon (lat, lon)
Given a LatLon point, returns the 21-deep HtmID of that point on the earth.
Parameters:
lat: float not null
latitude
It is converted to the range [-90 ... 90]
lon: float not null
longitude
Returns:
htmID bigint not null
the 21-deep HtmID of that (ra,dec) point.
Example use:
Declare @htmID bigint
Select @htmID =dbo.fHtmLatLon(lat,lon)
From Place
where placeName = ‘Baltimore’ and state = ‘MD’
Errors: None.
fHtmEq(ra, dec)
Given a J2000 (equatorial) point, returns the 21-deep HtmID of that point on the celestial sphere.
Parameters:
Ra: float not null
right ascension in degrees.
It is converted to a [0 … 360) range
Dec: float not null
declination in degrees.
It is converted to the range [-90 ... 90]
Returns:
htmID bigint not null
the 21-deep HtmID of that (ra,dec) point.
Example use:
Declare @htmID bigint
Select @htmID =dbo.fHtmEq(ra,dec)
From Stars
Where id = 42
Errors: None.
fHtmXyz(x,y,z)
Given a Cartesian point, returns the 21-deep HtmID of that point on the sphere.
Parameters:
x: float not null
vector to galaxy center or prime meridian intersection with
equator
y: float not null
vector normal to x in galactic plane or normal to prime
meridian
z: float not null
vector normal to galactic plane or equator
xyz will be normalized to 1. ( (0, 0, 0) is converted to RA = 0, DEC = 0, i.e. (1, 0, 0)).
Returns:
htmID bigint not null
the 21-deep HtmID of that (x,y,z) point.
Example use:
Declare @htmID bigint
Select @htmID = dbo.fHtmXyz(1,0,0)
Errors: None.
HTM Triangle Functions
fHtmToCenterPoint (HtmID)
Given an HtmID, return its Cartesian x,y,z centerpoint as a vertex table.
Parameters:
HtmID: bigint not null
HtmID of a triangle.
Returns:
VertexTable(x float, y float, z float).
Example use:
Select x,y,z from fHtmToCenterPoint(dbo.fHtmLatLon(38,115))
Errors: None.
fHtmToCornerPoints (HtmID)
Given an HtmID, returns the three Cartesian x,y,z corner points of the triangle as a vertex table. If the
HtmID has less shallow depth, this will be a large triangle. For example, HtmID = 8 returns the corner
points of the entire octant.
Parameters:
HtmID: bigint not null
HtmID of a triangle.
Returns:
VertexTable(x float, y float, z float).
Example use:
Select x,y,z from fHtmToCornerPoints(dbo.fHtmLatLon(38,115))
Errors: None.
Distances Between Points
fDistanceLatLon( lat1, lon1, lat2, lon2)
Given two LatLon points, fDistanceLatLon() returns distance between them in arc minutes (nautical miles).
Parameters:
lat1: float not null
latitude in degrees truncated to [-90, 90].
lon1: float not null
longitude in degrees truncated to [0,360).
lat2: float not null
latitude in degrees truncated to [-90, 90].
lon2: float not null
longitude in degrees truncated to [0,360).
Returns:
Float not null
distance in arc minutes
Example use:
If (60 != dbo.fDistanceLatLon(0, 0, 1, 0) ) print 'error'
Errors: None.
fDistanceEq( ra1, dec1, ra2, dec2)
Given two J2000 (equatorial) points, fDistanceEq() returns distance between them in arc minutes.
Parameters:
Ra1: float not null
right ascension in degrees truncated to [0,360].
Dec1: float not null
declination in degrees truncated to [-90, 90].
Ra2: float not null
right ascension in degrees truncated to [0,360].
Dec2: float not null
declination in degrees truncated to [-90, 90]. .
Returns:
Float not null
distance in arc minutes
Example use:
If (60 != dbo.fDistanceEq(0, 0, 1, 0) ) print 'error'
Errors: None.
fDistanceXyz( x1,y1,z1, x2,y2,z2)
Given two Cartesian points, fHtmXyz returns distance between them in arc minutes.
Parameters:
x1, x2: float not null
vector to galaxy center or prime meridian intersection with
equator.
y1, y2: float not null
vector normal to x in galactic plane or prime meridian
z1, z2: float not null
vector normal to galactic plane or to north pole.
(x, y, z) will be normalized. (0,0,0) will be converted to (0, 0, 1)
Returns:
Float not null
distance in arc minutes
Example use:
If (5400 != dbo.fDistanceXyz(0, 1, 0, 0, 0, 1) ) print 'error'
Errors: None.
REGIONS
A region is an area of interest on the celestial sphere. You can specify a region as a polygon, a convex hull
of a polygon, a rectangle, or a circle. Inside the kernel of the HTM engine, all regions are represented as a
union of convexes, which are, in turn, intersections of halfspaces. For more information, see the article
(Htm.doc) in the Geospatial project.
Syntactically, a region is a list of convexes. Furthermore, a convex is a list of halfspaces, and a halfspace is
a 4-tuple {x, y, z, D}.
Region Specifications
This is the syntax for region (cover) specifications:
circleSpec
rectSpec
polySpec
hullSpec
convexSpec
areaSpec
regionSpec
:=
|
|
:=
|
|
:=
|
|
:=
|
|
:=
|
|
:=
|
|
|
|
:=
|
'CIRCLE J2000'
'CIRCLE LATLON'
'CIRCLE [CARTESIAN ]'
'RECT J2000'
'RECT LATLON'
'RECT [CARTESIAN ]'
'POLY J2000'
'POLY LATLON'
'POLY [CARTESIAN ]'
'CHULL J2000'
'CHULL LATLON'
'CHULL [CARTESIAN ]'
'CONVEX J2000'
'CONVEX LATLON'
'CONVEX [CARTESIAN ]’
circleSpec
rectSpec
polySpec
hullSpec
convexSpec
'REGION' {areaSpec}*
areaSpec
ra dec rad
lat lon rad
x y z
rad
{ra dec}2
{lat lon}2
{x y z }2
{ra dec}3+
{lat lon}3+
{x y z }3+
{ra dec}3+
{lat lon}3+
{x y z }3+
{ra dec D}*
{lat lon D}*
{x y z
D}*
Examples of Region Specifications
REGION
A number of convexes (including zero)
REGION CONVEX 1 0 0 0.7 0 1 0 0.7
REGION CONVEX J2000 0 0 0.99
5 3 0.99
REGION CONVEX J2000 0 0 0.99 CONVEX J2000 5 3 0.99
REGION CONVEX LATLON 90 0 0
REGION
CONVEX
Any number of (including zero) constraints
CONVEX CARTESIAN 0.7 0.7 0.0 -0.5 0.7 –0.7 0.0 –0.5
CONVEX
CIRCLE
A Point specification, like J2000 ra,dec, and an arc minutes radius. Angles are in
degrees. Represented as a CONVEX consisting of a single constraint.
CIRCLE J2000 182.25 –22.432 1.75
CIRCLE CARTESIAN 0.7 0.0 0.7 1.75
RECT
Followed by two angular point specs, defining the minimum and maximum of the ra,dec. The
latmin must be smaller than latmax. In a similar case for the longitudes, they are interpreted in
the wrap-around sense, i.e., ramin=358.0 and ramax=2.0, means a four-degree wide range.
RECT J2000 182.25 –1.432 184.75 1.44
POLY
Followed by an optional single coordinate specification and a number of corresponding point
specifications (two or three numbers each). The spherical polygon will be created by
connecting the points by great circle segments. Because it is restricted to a convex polygon,
the order does not matter, but must be consistent. If there is a bowtie pattern in the points, or
if the polygon is not convex, an error will result.
POLY J2000 -109.55 41 -102.05 41 -102.05 37 -109.55 37
CHULL
Followed by an optional single coordinate specification, and a number of corresponding point
specifications (two or three numbers each). The spherical convex hull will be created by
connecting the adjacent points by great circles. At least three points are needed. The points
should all be within a single hemisphere, otherwise an error is returned. The order of the
points is irrelevant.
CHULL J2000 180. -1. 190. -2. 185. 3. 182. 4. 185. 5.
fHtmRegionToNormalFormString (regionSpec)
Given a string describing a region, fHtmRegionToNormalFormString () returns the normalized
representation of that region as a union of non-empty convex hulls, with redundant constraints (halfspaces)
discarded from each convex.
Parameters:
regionSpec: nvarchar(max) not null see syntax for region specifications above.
Returns:
nvarchar(max) not null
returns the normalized region spec of the form
REGION {CONVEX {x y z d}* }*
(or null if error).
Example use:
Declare @regionSpec nvarchar(max)
Select @regionSpec = dbo.fHtmToNormalForm('CIRCLE J2000 195 0 1')
Errors:
regionSpec syntax error, returns empty string, see fHtmRegionError()
fHtmRegionToTable (regionSpec)
Given a string describing a region, fHtmRegionToTable() returns the tabular representation of the region as
a union of non-empty convex hulls, with redundant constraints (halfspaces) discarded from each convex.
The tabular representation has the schema described below:
Parameters:
regionSpec: nvarchar(max) not null see syntax for region specifications above.
Returns:
RegionTable (
convexID
bigint not null ,
halfSpaceID bigint not null
x
y
z
d
float
float
float
float
not
not
not
not
null
null
null
null
---------
ID of the convex, 0,1,…
ID of the halfspace
within convex, 0,1,2,
Cartesian coordinates of
unit-normal-vector of
halfspace plane
displacement of halfspace
along unit vector [-1..1]
)
Or empty table if error.
Example use:
select *
from fHtmToNormalForm('CIRCLE J2000 195 0 1')
Errors:
regionSpec syntax error, returns empty table., see fHtmRegionError()
fHtmRegionObjects (regionSpec, type)
This routine is particular to the SQL Server 2005 sample spatial database and library which has Place,
Station, and SpatialIndex tables, and has these functions installed. Given a string describing a region and a
type “P” for place or “S” for station, fHtmRegionObjects() returns the tabular list of all the SpatialIndex
objects of that type that are inside that region:
Parameters:
regionSpec: nvarchar(max) not null see syntax for region specifications above.
Type: char(1): “P” for Places in the Place table, “S” for Stations in the Station table
Returns:
ObjectTable (
objID bigint not null primary key , -- ID of the object,…
-- if type is “S”, it is the Station.stationNumber
-- if type is “T”, it is the Place.HtmID
Or empty table if error.
Example use:
select *
-- find Colorado places.
from Place
where HtmID in
(select objID
from fHtmRegionObjects('RECT LATLON 37 -109.55 41 -102.05','P'))
Errors:
regionSpec syntax error, returns empty table., see fHtmRegionError()
fHtmRegionError(regionSpec)
Returns “OK” if a valid regionSpec, else returns syntax error message.
Parameters:
regionSpec: nvarchar(max) not null see syntax for region specifications above.
Returns:
nvarchar(max) not null
diagnostic message
Example use:
Declare @diagnostic nvarchar(max)
Select @ diagnostic =
dbo.fHtmRegionError('CIRCLE J2000 195 0')
Errors: None.
HTM Covers: Compute HtmID Ranges for a Region
This suite of routines, given a region specification, returns a table of trixels. The trixels cover the specified
region. The trixels are described by a start-stop HTM pair. All points within the trixel are between the startstop of the 21-deep HTM pair; in fact, they are in the closed interval [Htm_Start, Htm_Stop].
The resulting table has the definition:
TrixelTable(Htm_Start bigint, Htm_Stop bigint)
Simple regions can be described as standard geometric shapes (circle, rectangle) giving the parameters. But
typically, regions are described by using the linear syntax described above. Because the enumeration of
HtmIDs tends to form locally connected intervals, the interface unifies these contiguous triangles as one
large trixel.
fHtmCoverRegion (regionSpec)
Given a string describing a region, fHtmRegionCover () returns the trixel table covering that region.
Parameters:
regionSpec: nvarchar(max) not null see syntax for region specifications above.
Returns:
TrixelTable(Htm_Start bigint, Htm_Stop bigint)
Example use:
Select * from fHtmCoverRegion('CIRCLE J2000 195 0 1')
Errors:
In case of error, returns the empty table. Use fHtmRegionError(RegionString) to get diagnostic.
fHtmCoverCircleLatLon(lat, lon, radiusArcMin)
Given a (latitude, longitude) point and a radius in arc minutes, fHtmCoverCircleLatLon () returns the trixel
table covering that circle.
Parameters:
Lat: float not null
latitude in degrees.
It is converted to the range [-90 ... 90]
Lon: float not null
longitude in degrees.
radiusArcMin: float not null circle’s radius in
arc minutes.
Radius should be positive and less than 180 degrees, i.e., 10800 minutes of arc.
Returns:
TrixelTable(Htm_Start bigint, Htm_Stop bigint)
Example use:
Select * from fHtmCoverCircleLatLon(195,0,1)
Errors: None:
fHtmCoverCircleEq(ra, dec, radiusArcMin)
Given a J2000 ra, dec point and a radius in arc minutes, fHtmCoverCircleEq () returns the trixel table
covering that circle.
Parameters:
Ra: float not null
right ascension in degrees.
It is converted to a [0 … 360)
Dec: float not null
declination in degrees.
It is converted to the range [-90 ... 90]
radiusArcMin: float not null circle’s radius in arcminutes.
Radius should be positive and less than 180 degrees, i.e., 10800 minutes of arc.
Returns:
TrixelTable(Htm_Start bigint, Htm_Stop bigint)
Example use:
Select * from fHtmCoverCircleEq(195,0,1)
Errors: None:
fHtmCoverCircleXyz(x, y, z, radiusArcMin)
Given a string describing a region, fHtmCoverCircleXyz() returns the trixel table covering that circle.
Parameters:
x: float not null
vector to galaxy center
y: float not null
vector normal to x in galactic plane
z: float not null
vector normal to galactic plane
radiusArcMin: float not null circle’s radius in arc minutes.
xyz will be normalized to 1. (0,0,0) is converted to the North Pole: (0,0,1).
Radius is range limited to [0…10800]
Returns:
TrixelTable(Htm_Start bigint, Htm_Stop bigint)
Example use:
Select * from fHtmCoverCircleXyz(1,0,0,1)
Errors: None:
Installing the HTM Code
1.
Install the SQL Server 2005 samples by following the instructions in the Installing Samples topic
in SQL Server Books Online. By default the sample will be installed in drive:\Program
Files\Microsoft SQL Server\90\Samples\Engine\Programmability\CLR\Spatial\, where drive is the
system drive.
2.
Compile the provided solution by using Visual Studio 2005 or the Microsoft .NET Framwork
SDK 2.0 using a command similar to the following in a .NET Framework SDK command prompt:
msbuild /property:configuration=debug CS\Spatial.sln
3. Attach the Spatial database in the data directory by using SQL Server Management Studio or by
executing the Scripts\AttachSpatialDatabase.bat command file in a command
prompt window if you have not done so already.
4.
Execute the spHtmCsharp.sql script in SQL Server Management Studio, or by executing a
command similar to the following in a command prompt window:
sqlcmd -S "(local)" -d Spatial -E -i "C:\Program Files\Microsoft
SQL
Server\90\Samples\Engine\Programmability\CLR\Spatial\Scripts\spHt
mCsharp.sql"
This script requires SQL Server 2005 be the default database server on the local system. This command is
part of the contents of Scripts\BuildSpatialDatabase.bat. It does not work with SQL Server 2000
or earlier versions. The spHtmCsharp.sql script enables the common language runtime (CLR), drops
any existing HTM assembly and replaces it with the current assembly
(CS\Spatial\bin\debug\Spatial.dll), and then defines all the HTM functions in that assembly.
Compiling, Modifying, and Debugging the HTM Code
The C:\HtmCsharp\htm.sln is a C# database project. You should be able to use the “deploy” and the
“debug” features and breakpoints will just work.
Table of HTM Depths And Approximate Areas
The library here defaults to 21-deep HTM keys (the first level divides the sphere into
eight faces, and each subsequent level divides the speherical triangle into four subtriangles.) The table below indicates that a 21-deep trixel is fairly small. The code can be
modified to go 31-deep deep before the 64-bit representation runs out of bits, but the
floating point representation and transcendental functions lose precision near level 25 .
Table 1: Each HTM level subdivdies the sphere. For each level, this table shows the area
in square degrees, arc minutes, arc seconds, and meters. The Trixel colum shows some
characteristic sizes: the default 21-deep trixels is about .3 arc second2. The USGS data has
about ½ object per 12-deep trixel.
HTM
Area
Objects/Trixel
2
2
2
2
depth degees
minute
area sec
earth meters trixel SDSS USGS
41253 148,510,800 534,638,880,000
5.E+14
sphere
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
5157
1289
322
81
20
5
1
3E-1
8E-2
2E-2
5E-3
1E-3
3E-4
8E-5
2E-5
5E-6
1E-6
3E-7
8E-8
2E-8
5E-9
1E-9
3E-10
7E-11
2E-11
5E-12
1E-12
18,563,850
4,640,963
1,160,241
290,060
72,515
18,129
4,532
1,133
283
71
18
4
1
3E-1
7E-2
2E-2
4E-3
1E-3
3E-4
7E-5
2E-5
4E-6
1E-6
3E-7
7E-8
2E-8
4E-9
66,829,860,000
16,707,465,000
4,176,866,250
1,044,216,563
261,054,141
65,263,535
16,315,884
4,078,971
1,019,743
254,936
63,734
15,933
3,983
996
249
62
16
4
1
2E-1
6E-2
2E-2
4E-3
9E-4
2E-4
6E-5
1E-5
6E+13
2E+13
4E+12
1E+12
2E+11
6E+10
2E+10
4E+9
1E+9
2E+8
6E+7
2E+7
4E+6
943816
235954
58989
14747
3687
922
230
58
14
4
1
2E-1
6E-2
1E-2
1 deg2
1 amin2
1 asec2
1 km2
1 m2
3E+8
8E+7
2E+7
5E+6
1E+6
3E+5
73242
18311
4578
1144
286
72
18
4
1
0.3
.
30,000
7,500
1,875
468
117
29
7
2
0.5
0.1
| 5cs.CE
|
THE CONGRUENCE TOPOLOGY,
GROTHENDIECK DUALITY AND THIN GROUPS
arXiv:1709.06179v1 [math.GR] 18 Sep 2017
ALEXANDER LUBOTZKY AND T.N. VENKATARAMANA
Abstract. This paper answers a question raised by Grothendieck
in 1970 on the “Grothendieck closure” of an integral linear group
and proves a conjecture of the first author made in 1980. This
is done by a detailed study of the congruence topology of arithmetic groups, obtaining along the way, an arithmetic analogue of
a classical result of Chevalley for complex algebraic groups. As an
application we also deduce a group theoretic characterization of
thin subgroups of arithmetic groups.
0. Introduction
If ϕ : G1 → G2 is a polynomial map between two complex varieties, then in general the image of a Zariski closed subset of G1 is not
necessarily closed in G2 . But here is a classical result:
Theorem (Chevalley). If ϕ is a polynomial homomorphism between
two complex algebraic groups then ϕ(H) is closed in G2 for every closed
subgroup H of G1 .
There is an arithmetic analogue of this issue: Let G be a Q-algebraic
group, let Af = Π∗p prime Qp be the ring of finite ádeles over Q. The
topology of G(Af ) induces the congruence topology on G(Q). If K
is compact open subgroup of G(Af ) then Γ = K ∩ G(Q) is called a
congruence subgroup of G(Q). This defines the congruence topology
on G(Q) and on all its subgroups. A subgroup of G(Q) which is closed
in this topology is called congruence closed. A subgroup ∆ of G commensurable to Γ is called an arithmetic group.
Now, if ϕ : G1 → G2 is a Q-morphism between two Q-groups, which
is a surjective homomorphism (as C-algebraic groups) then the image
of an arithmetic subgroup ∆ of G1 is an arithmetic subgroup of G2
([Pl-Ra, Theorem 4.1 p. 174]), but the image of a congruence subgroup is not necessarily a congruence subgroup. It is well known that
SLn (Z) has congruence subgroups whose images under the adjoint map
SLn (Z) → PSLn (Z) ֒→ Aut(Mn (Z)) are not congruence subgroups (see
[Ser] and Proposition 2.1 below for an exposition and explanation). So,
1
2
A. LUBOTZKY AND T.N. VENKATARAMANA
the direct analogue of Chevalley theorem does not hold. Still, in this
case, if Γ is a congruence subgroup of SLn (Z), then ϕ(Γ) is a normal
subgroup of ϕ(Γ), the (congruence) closure of ϕ(Γ) in PSLn (Z), and
the quotient is a finite abelian group. Our first technical result says
that the general case is similar. It is especially important for us that
when G2 is simply connected, the image of a congruence subgroup of
G1 is a congruence subgroup in G2 (see Proposition 0.1 (ii) below).
Before stating the result, we give the following definition and set
some notations for the rest of the paper:
Let G be a linear algebraic group over C, G0 - its connected component, and R = R(G) - its solvable radical, i.e. the largest connected
normal solvable subgroup of G. We say that G is essentially simply
connected if Gss := G0 /R is simply connected.
Given a subgroup Γ of GLn , we will throughout the paper denote by
0
Γ the intersection of Γ with G0 , where G0 is the connected component
of G - the Zariski closure of Γ. Therefore, Γ0 is always a finite index
normal subgroup of Γ.
The notion “essentially simply connected” will play an important role
in this paper due to the following proposition, which can be considered
as the arithmetic analogue of Chevalley’s result above:
Proposition 0.1. (i) If ϕ : G1 → G2 is a surjective (over C) algebraic homomorphism between two Q-defined algebraic groups,
then for every congruence closed subgroup Γ of G1 (Q), the image
ϕ(Γ0 ) is normal in its congruence closure ϕ(Γ0 ) and ϕ(Γ0 )/ϕ(Γ0 )
is a finite abelian group.
(ii) If G2 is essentially simply connected, and Γ a congruence subgroup
of G1 then ϕ(Γ) = ϕ(Γ), i.e., the image of a congruence subgroup
is congruence closed.
This analogue of Chevalley’s theorem, and a result of [Nori], [Weis]
enable us to prove:
Proposition 0.2. If Γ1 ≤ GLn (Z) is a congruence closed subgroup
(i.e. closed in the congruence topology) with Zariski closure G, then
there exists a congruence subgroup Γ of G, such that [Γ, Γ] ≤ Γ01 ≤ Γ.
If G is essentially simply connected then the image of Γ1 in G/R(G) is
actually a congruence subgroup.
We apply Proposition 0.1 (ii) in two directions:
(A) Grothendieck-Tannaka duality for discrete groups, and
CONGRUENCE TOPOLOGY
3
(B) A group theoretic characterization of thin subgroups of arithmetic
groups.
Grothendieck closure. In [Gro], Grothendieck was interested in the
following question:
Question 0.3. Assume ϕ : Γ1 → Γ2 is a homomorphism between
two finitely generated residually finite groups inducing an isomorphism
ϕ̂ : Γ̂1 → Γ̂2 between their profinite completions. Is ϕ already an
isomorphism?
To tackle Question 0.3, he introduced the following notion. Given a
finitely generated group Γ and a commutative ring A with identity, let
ClA (Γ) be the group of all automorphisms of the forgetful functor from
the category ModA (Γ) of all finitely generated A-modules with Γ action
to ModA ({1}), preserving tensor product. Grothendieck’s strategy was
the following: he showed that, under the conditions of Question 0.3, ϕ
induces an isomorphism from ModA (Γ2 ) to ModA (Γ1 ), and hence also
between ClA (Γ1 ) and ClA (Γ2 ). He then asked:
Question 0.4. Is the natural map Γ ֒→ ClZ (Γ) an isomorphism for a
finitely generated residually finite group?
An affirmative answer to Question 0.4 would imply an affirmative
answer to Question 0.3. Grothendieck then showed that arithmetic
groups with the (strict) congruence subgroup property do indeed satisfy
ClZ (Γ) ≃ Γ.
Question 0.4 basically asks whether Γ can be recovered from its category of representations. In [Lub], the first author phrased this question in the framework of Tannaka duality, which asks a similar question
for compact Lie groups. He also gave a more concrete description of
ClZ (Γ):
(0.1)
ClZ (Γ) = {g ∈ Γ̂|ρ̂(g)(V ) = V,
∀ (ρ, V ) ∈ ModZ (Γ)}.
Here ρ̂ is the continuous extension ρ̂ : Γ̂ → Aut(V̂ ) of the original
representation ρ : Γ → Aut(V ).
However, it is also shown in [Lub], that the answer to Question 0.4
is negative. The counterexamples provided there are the arithmetic
groups for which the weak congruence subgroup property holds but not
the strict one, i.e. the congruence kernel is finite but non-trivial. It
was conjectured in [Lub, Conj A, p. 184], that for an arithmetic group
4
A. LUBOTZKY AND T.N. VENKATARAMANA
Γ, ClZ (Γ) = Γ if and only if Γ has the (strict) congruence subgroup
property. The conjecture was left open even for Γ = SL2 (Z).
In the almost 40 years since [Lub] was written various counterexamples were given to question 0.3 ([Pl-Ta1], [Ba-Lu], [Br-Gr], [Py]) which
also give counterexamples to question 0.4, but it was not even settled
whether ClZ (F ) = F for finitely generated non-abelian free groups F .
We can now answer this and, in fact, prove the following surprising
result, which gives an essentially complete answer to Question 0.4.
Theorem 0.5. Let Γ be a finitely generated subgroup of GLn (Z). Then
Γ satisfies Grothendieck-Tannaka duality, i.e. ClZ (Γ) = Γ if and only
if Γ has the congruence subgroup property i.e., for some (and consequently for every) faithful representation Γ → GLm (Z) such that the
Zariski closure G of Γ is essentially simply connected, every finite index
subgroup of Γ is closed in the congruence topology of GLn (Z). In such
a case, the image of the group Γ in the semi-simple (simply connected)
quotient G/R is a congruence arithmetic group.
The Theorem is surprising as it shows that the cases proved by
Grothendieck himself (which motivated him to suggest that the duality holds in general) are essentially the only cases where this duality
holds.
Let us note that the assumption on G is not really restrictive. In
Lemma 3.6, we show that for every Γ ≤ GLn (Z) we can find an “over”
representation of Γ into GLm (Z) (for some m) whose Zariski closure is
essentially simply connected.
Theorem 0.5 implies Conjecture A of [Lub].
Corollary 0.6. If G is a simply connected semisimple Q-algebraic
group, and Γ a congruence subgroup of G(Q), then ClZ (Γ) = Γ if and
only if Γ satisfies the (strict) congruence subgroup property.
In particular:
Corollary 0.7. ClZ (F ) 6= F for every finitely generated free group on
at least two generators; furthermore, ClZ (SL2 (Z)) 6= SL2 (Z).
In fact, it will follow from our results that ClZ (F ) is uncountable.
Before moving on to the last application, let us say a few words
about how Proposition 0.1 helps to prove a result like Theorem 0.5.
CONGRUENCE TOPOLOGY
5
The description of ClZ (Γ) as in Equation 0.1 implies that
(0.2)
ClZ (Γ) = lim ρ(Γ)
←
ρ
when the limit is over all (ρ, V ) when V is a finitely generated abelian
group, ρ a representation ρ : Γ → Aut(V ) and ρ(Γ) = ρ̂(Γ̂) ∩ Aut(V ) ⊆
Aut(V̂ ). This is an inverse limit of countable discrete groups, so one
can not say much about it unless the connecting homomorphisms are
surjective, which is, in general, not the case. Now, ρ(Γ) is the congruence closure of ρ(Γ) in Aut(V ) and Proposition 0.1 shows that the
corresponding maps are “almost” onto, and are even surjective if the
modules V are what we call here “simply connected representations”,
namely those cases when V is torsion free (and hence isomorphic to Zn
for some n) and the Zariski closure of ρ(Γ) in Aut(C ⊗ V ) = GLn (C)
Z
is essentially simply connected. We show further that the category
ModZ (Γ) is “saturated” with such modules (see Lemma 3.6) and we
deduce that one can compute ClZ (Γ) as in Equation 0.1 by considering
only simply connected representations. We can then use Proposition
0.1(b), and get a fairly good understanding of ClZ (Γ). This enables us
to prove Theorem 0.5. In addition, we also deduce:
Corollary 0.8. If (ρ, V ) is a simply connected representation, then the
induced map ClZ (Γ) → Aut(V ) is onto Clρ (Γ) := ρ(Γ) - the congruence
closure of Γ.
From Corollary 0.8 we can deduce our last application.
Thin groups. In recent years, following [Sar], there has been a lot
of interest in the distinction between thin subgroups and arithmetic
subgroups of algebraic groups. Let us recall:
Definition 0.9. A subgroup Γ ≤ GLn (Z) is called thin if it is of
infinite index in G ∩ GLn (Z), when G is its Zariski closure in GLn .
For a general group Γ, we will say that it is a thin group (or it has
a thin representation) if for some n there exists a representation
ρ : Γ → GLn (Z) for which ρ(Γ) is thin.
During the last five decades a lot of attention was given to the study
of arithmetic groups, with many remarkable results, especially for those
of higher rank (cf. [Mar], [Pl-Ra] and the references therein). Much
less is known about thin groups. For example, it is not known if there
exists a thin group with property (T ). Also, given a subgroup of an
arithmetic group (say, given by a set of generators) it is difficult to
6
A. LUBOTZKY AND T.N. VENKATARAMANA
decide whether it is thin or arithmetic (i.e., of finite or infinite index
in its integral Zariski closure).
It is therefore of interest and perhaps even surprising that our results
enable us to give a purely group theoretical characterization of thin
groups Γ ⊂ GLn (Z). Before stating the precise result, we make the
topology on ClZ (Γ) explicit. If we take the class of simply connected
representations (ρ, V ) for computing the group ClZ (Γ), one can then
Q
show that ClZ (Γ)/Γ is a closed subspace of the product ρ (Clρ (Γ)/Γ),
where each Clρ (Γ)/Γ is given the discrete topology. This is the topology
on the quotient space ClZ (Γ)/Γ in the following theorem. We can now
state:
Theorem 0.10. Let Γ be finitely generated Z-linear group. Then Γ
is a thin group if and only if it satisfies (at least) one of the following
conditions:
(1) Γ is not F Ab (namely, it does have a finite index subgroup with
an infinite abelianization), or
(2) ClZ (Γ)/Γ is not compact.
Warning There are groups Γ which can be realized both as arithmetic groups as well as thin groups. For example, the free group is an
arithmetic subgroup of SL2 (Z), but at the same time a thin subgroup
of every semisimple group, by a well known result of Tits [Ti]. In our
terminology this is a thin group.
T.N.V. thanks the Math Department of the Hebrew University for
great hospitality while a major part of this work was done. He would
also like to thank the JC Bose fellowship (SR/S2/JCB-22/2017) for
support during the period 2013-2018.
The authors thank the Math Department of the University of Marseilles, and the conference at Oberwolfach, where the work was completed. We would especially like to thank Bertrand Remy for many
interesting discussions and for his warm hospitality.
A.L. is indebted to ERC, NSF and BSF for support.
1. Preliminaries on Algebraic Groups over Q
We recall the definition of an essentially simply connected group:
CONGRUENCE TOPOLOGY
7
Definition 1.1. Let G be a linear algebraic group over C with maximal
connected normal solvable subgroup R (i.e. the radical of G) and identity component G0 . We say that G is essentially simply connected
if the semi-simple part G0 /R = H is a simply connected.
Note that G is essentially simply connected if and only if, the quotient G0 /U of the group G0 by its unipotent radical U is a product
Hss × S with Hss simply connected and semi-simple, and S is a torus.
For example, a semi-simple connected group is essentially simply
connected if and only if it is simply connected. The group Gm × SLn
is essentially simply connected; however, the radical of the group GLn
is the group R of scalars and GLn /R = SLn /centre, so GLn is not
essentially simply connected. We will show later (Lemma 1.3(iii)) that
every group has a finite cover which is essentially simply connected.
Lemma 1.2. Suppose G ⊂ G1 × G2 is a subgroup of a product of
two essentially simply connected linear algebraic groups G1 , G2 over
C; suppose that the projection πi of G to Gi is surjective for i = 1, 2.
Then G is also essentially simply connected.
Proof. Assume, as we may, that G is connected. Let R be the radical
of G. The projection of R to Gi is normal in Gi since πi : G → Gi is
surjective. Moreover, Gi /πi (R) is the image of the semi-simple group
G/R; the latter has a Zariski dense compact subgroup, hence so does
Gi /πi (R); therefore, Gi /πi (R) is reductive and is its own commutator.
Hence Gi /πi (R) is semi-simple and hence πi (R) = Ri where Ri is the
radical of Gi . Let R∗ = G ∩ (R1 × R2 ). Since R1 × R2 is the radical
of G1 × G2 , it follows that R∗ is a solvable normal subgroup of G and
hence its connected component is contained in R. Since R ⊆ R1 × R2 ,
it follows that R is precisely the connected component of the identity
of R∗ . We then have the inclusion G/R∗ ⊂ G1 /R1 × G2 /R2 with
projections again being surjective.
By assumption, each Gi /Ri = Hi is semi-simple, simply connected.
Moreover G/R∗ = H where H is connected, semi-simple. Thus we
have the inclusion H ⊂ H1 × H2 . Now, H ⊂ H1 × H2 is such that the
projections of H to Hi are surjective, and each Hi is simply connected.
Let K be the kernel of the map H → H1 and K 0 its identity component.
Then H/K 0 → H1 is a surjective map of connected algebraic groups
with finite kernel. The simple connectedness of H1 then implies that
H/K 0 = H1 and hence that K = K 0 ⊂ {1} × H2 is normal in H2 .
8
A. LUBOTZKY AND T.N. VENKATARAMANA
Write H2 = F1 × · · · × Ft where each Fi is simple and simply connected. Now, K being a closed normal subgroup of H2 must be equal
Q
to i∈X Fi for some subset X of {1, · · · , t}, and is simply connected.
Therefore, K = K 0 is simply connected.
From the preceding two paragraphs, we have that both H/K and
K are simply connected, and hence so is H = G/R∗ . Since R is the
connected component of R∗ and G/R∗ is simply connected, it follows
that G/R = G/R∗ and hence G/R is simply connected. This completes
the proof of the lemma.
1.1. Arithmetic Groups and Congruence Subgroups. In the introduction, we defined the notion of arithmetic and congruence subgroup of G(Q) using the adelic language. One can define the notion of
arithmetic (res. congruence) group in more concrete terms as follows.
Given a linear algebraic group G ⊂ SLn defined over Q, we will say that
a subgroup Γ ⊂ G(Q) is an arithmetic group if is commensurable to
G ∩ SLn (Z) = G(Z); that is, the intersection Γ ∩ G(Z) has finite index
both in Γ and in G(Z). It is well known that the notion of an arithmetic
group does not depend on the specific linear embedding G ⊂ SLn . As
b of G(Q) as the
in [Ser], we may define the arithmetic completion G
completion of the group G(Q) with respect to the topology on G(Q)
as a topological group, obtained by designating arithmetic groups as a
fundamental systems of neighbourhoods of identity in G(Q).
Given G ⊂ SLn as in the preceding paragraph, we will say that an
arithmetic group Γ ⊂ G(Q) is a congruence subgroup if there exists
an integer m ≥ 2 such that Γ contains the “principal congruence subgroup” G(mZ) = SLn (mZ) ∩ G where SLn (mZ) is the kernel to the
residue class map SLn (Z) → SLn (Z/mZ). We then get the structure of
a topological group on the group G(Q) by designating congruence subgroups of G(Q) as a fundamental system of neighbourhoods of identity.
The completion of G(Q) with respect to this topology, is denoted G.
Again, the notion of a congruence subgroup does not depend on the
specific linear embedding G → SLn .
Since every congruence subgroup is an arithmetic group, there exists
b → G which is easily seen to be surjective, and
a map from π : G
b This is
the kernel C(G) of π is a compact profinite subgroup of G.
called the congruence subgroup kernel. One says that G(Q) has the
congruence subgroup property if C(G) is trivial. This is easily seen to
be equivalent to the statement that every arithmetic subgroup of G(Q)
is a congruence subgroup.
CONGRUENCE TOPOLOGY
9
It is known (see p. 108, last but one paragraph of [Ra2] or [Ch]) that
solvable groups G have the congruence subgroup property.
Moreover, every solvable subgroup of GLn (Z) is polycyclic. In such
a group, every subgroup is intersection of finite index subgroups. So
every solvable subgroup of an arithmetic group is congruence closed.
We will use these facts frequently in the sequel.
Another (equivalent) way of viewing the congruence completion is
(see [Ser], p. 276, Remarque) as follows: let Af be the ring of finite
adeles over Q, equipped with the standard adelic topology and let Zf ⊂
Af be the closure of Z. Then the group G(Af ) is also a locally compact
group and contains the group G(Q). The congruence completion G of
G(Q) may be viewed as the closure of G(Q) in G(Af ).
Lemma 1.3. Let H, H ∗ be linear algebraic groups defined over Q.
(i) Suppose H ∗ → H is a surjective Q-morphism. Let (ρ, WQ ) be a
representation of H defined over Q. Then there exists a faithful Q-representation (τ, VQ ) of H ∗ such that (ρ, W ) is a subrepresentation of (τ, V ).
(ii) If H ∗ → H is a surjective map defined over Q , then the image
of an arithmetic subgroup of H ∗ under the map H ∗ → H is an
arithmetic subgroup of H.
(iii) If H is connected, then there exists a connected essentially simply
connected algebraic group H ∗ with a surjective Q-defined homomorphism H ∗ → H with finite kernel.
(iv) If H ∗ → H is a surjective homomorphism of algebraic Q-groups
which are essentially simply connected, then the image of a congruence subgroup of H ∗ (Q) is a congruence subgroup of H(Q).
Proof. Let θ : H ∗ → GL(E) be a faithful representation of the linear
algebraic group H ∗ defined over Q and τ = ρ ⊕ θ as H ∗ -representation.
Clearly τ is faithful for H ∗ and contains ρ. This proves (i).
Part (ii) is the statement of Theorem (4.1) of [Pl-Ra].
We now prove (iii). Write H = RG as a product of its radical R and
∗
a semi-simple group G. Let Hss
→ G be the simply connected cover
∗
of G. Hence Hss acts on R through G, via this covering map. Define
∗
H ∗ = R ⋊ Hss
as a semi-direct product. Clearly, the map H ∗ → H has
finite kernel and satisfies the properties of (iii).
To prove (iv), we may assume that H and H ∗ are connected. If
U ∗ , U are the unipotent radicals of H ∗ and H, the assumptions of (iv)
do not change for the quotient groups H ∗ /U ∗ and H/U. Moreover,
since H ∗ is the semi-direct product of U ∗ and H ∗ /U ∗ (and similarly
10
A. LUBOTZKY AND T.N. VENKATARAMANA
for H, U) and the unipotent Q-algebraic group U has the congruence
subgroup property, it suffices to prove (iv) when both H ∗ and H are
reductive. By assumption, H ∗ and H are essentially simply connected;
∗
∗
i.e. H ∗ = Hss
× S ∗ and H = Hss × S where S, S ∗ are tori and Hss
, Hss
are simply connected semi-simple groups. Thus we have connected
reductive Q-groups H ∗ , H with a surjective map such that their derived
groups are simply connected (and semi-simple), and the abelianization
(H ∗ )ab is a torus (similarly for H).
∗
Now, [H ∗ , H ∗] = Hss
is a simply connected semi-simple group and
hence it is a product F1 ×· · ·×Fs of simply connected Q-simple algebraic
∗
groups Fi . Being a factor of [H ∗ , H ∗ ] = Hss
, the group [H, H] = Hss is a
product of a (smaller) number of these Fi ’s. After a renumbering of the
indices, we may assume that Hss is a product F1 ×· · ·×Fr for some r ≤ s
∗
and the map π on Hss
is the projection to the first r factors. Hence
∗
the image of a congruence subgroup of Hss
is a congruence subgroup
of Hss .
The tori S ∗ , S have the congruence subgroup property by a result
of Chevalley (as already stated at the beginning of this section, this is
true for all solvable algebraic groups). Hence the image of a congruence
subgroup of S ∗ is a congruence subgroup of S. We thus need only prove
that every subgroup of the reductive group H of the form Γ1 Γ2 , where
Γ1 ⊂ Hss and Γ2 ⊂ S are congruence subgroups, is itself a congruence
subgroup of H. We use the adelic form of the congruence topology.
Suppose K is a compact open subgroup of the H(Af ) where Af is the
ring of finite adeles. The image of H(Q) ∩ K under the quotient map
H → H ab = S is a congruence subgroup in the torus S and hence
H(Q) ∩ K ′ ⊂ (Hss (Q) ∩ K)(S(Q) ∩ K) for some possibly smaller open
subgroup K ′ ⊂ H(Af ). This proves (iv).
Note that part (iii) and (iv) prove Proposition 0.1(ii).
2. The Arithmetic Chevalley Theorem
In this section, we prove Proposition 0.1(i). Assume that ϕ : G1 →
G2 is a surjective morphism of Q-algebraic groups. We are to prove that
ϕ(Γ0 ) contains the commutator subgroup of a congruence subgroup of
G2 (Q) containing it.
Before starting on the proof, let us note that in general, the image
of a congruence subgroup of G1 (Z) under ϕ need not be a congruence
subgroup of G2 (Z). The following proposition gives a fairly general
situation when this happens.
CONGRUENCE TOPOLOGY
11
Proposition 2.1. Let π : G1 → G2 be a finite covering of semi-simple
algebraic groups defined over Q with G1 simply connected and G2 not.
Assume G1 (Q) is dense in G1 (Af ). Write K for the kernel of π and
Kf for the kernel of the map G1 (Af ) → G2 (Af ). Let Γ be a congruence
subgroup of G1 (Q) and H its closure in G1 (Af ). Then the image π(Γ) ⊂
G2 (Q) is a congruence subgroup if and only if KH ⊃ Kf .
Before proving the proposition, let us note that while K is finite, the
group Kf is a product of infinitely many finite abelian groups and that
Kf is central in G1 . This implies
Corollary 2.2. (i) There are infinitely many congruence subgroups
Γi with π(Γi ) non-congruence subgroups of unbounded finite index
in their congruence closures Γi .
(ii) For each of these Γ = Γi , the image π(Γ) contains the commutator
subgroup [Γ, Γ], and is normal in Γ (with abelian quotient).
We now prove Proposition 2.1.
Proof. Let G3 be the image of the rational points of G1 (Q):
G3 = π(G1 (Q)) ⊂ G2 (Q).
Define a subgroup ∆ of G3 to be a quasi-congruence subgroup if the
inverse image π −1 (∆) is a congruence subgroup of G1 (Q). Note that
the quasi-congruence subgroups of G3 are exactly the images of congruence subgroups of G1 (Q) by π. It is routine to check that by declaring
quasi-congruence subgroups to be open, we get the structure of a topological group on G3 . This topology is weaker or equal to the arithmetic
topology on G3 . However, it is strictly stronger than the congruence
topology on G3 . The last assertion follows from the fact that the completion of G3 = G1 (Q)/K(Q) is the quotient G1 /K where G1 is the
congruence completion of G1 (Q), whereas the completion of G3 with
respect to the congruence topology is G1 /Kf .
Now let Γ ⊂ G1 (Q) be a congruence subgroup and ∆1 = π(Γ); let
∆2 be its congruence closure in G3 . Then both ∆1 and ∆2 are open
in the quasi-congruence topology on G3 . Denote by G∗3 the completion
of G3 with respect to the quasi-congruence topology, so G∗3 = G1 /K
and denote by ∆∗1 , ∆∗2 the closures of ∆1 , ∆2 in G∗3 . We then have the
equalities
∆2 /∆1 = ∆∗2 /∆∗1 , ∆∗2 = ∆∗1 Kf /K.
Hence ∆∗1 = ∆∗2 if and only if K∆∗1 ⊃ Kf . This proves Proposition
2.1.
12
A. LUBOTZKY AND T.N. VENKATARAMANA
The proof shows that ∆∗1 is normal in ∆∗2 (since Kf is central) with
abelian quotient. The same is true for ∆1 in ∆2 and the corollary is
also proved.
To continue with the proof of Proposition 0.1, assume, as we may (by
replacing G1 with the Zariski closure of Γ), that G1 has no characters
defined over Q. For, suppose that G1 is the Zariski closure of Γ ⊂
G1 (Z). Let χ : G1 → Gm be a non-trivial (and therefore surjective)
homomorphism defined over Q; then the image of the arithmetic group
G1 (Z) in Gm (Q) is a Zariski dense arithmetic group. However, the only
arithmetic groups in Gm (Q) are finite and cannot be Zariski dense in
Gm . Therefore, χ cannot be non-trivial. We can also assume that G1
is connected.
We start by proving Proposition 0.1 for the case that Γ is a congruence subgroup.
If we write G1 = R1 H1 where H1 is semi-simple and R1 is the radical, we may assume that G1 is essentially simply connected (Lemma
1.3(iii)), without affecting the hypotheses or the conclusion of Proposition 0.1.
Hence G1 = R1 ⋊ H1 is a semi direct product. Then clearly, every
congruence subgroup of G1 contains a congruence subgroup of the form
∆⋊Φ where ∆ ⊂ R1 and Φ ⊂ H1 are congruence subgroups. Similarly,
write G2 = R2 H2 . Since ϕ is easily seen to map R1 onto R2 and H1
onto H2 , it is enough to prove the proposition for R1 and H1 separately.
We first recall that if G is a solvable linear algebraic group defined
over Q then the congruence subgroup property holds for G, i.e., every
arithmetic subgroup of G is a congruence subgroup (for a reference
see p. 108, last but one paragraph of [Ra2] or [Ch]). Consequently,
by Lemma 1.3 (ii), the image of a congruence subgroup in R1 is an
arithmetic group in R2 and hence a congruence subgroup. Thus we
dispose of the solvable case.
In the case of semi-simple groups, denote by H2∗ by the simply connected cover of H2 . The map ϕ : H1 → H2 lifts to a map from H1 to
H2∗ . For simply connected semi-simple groups, a surjective map from
H1 to H2∗ sends a congruence subgroup to a congruence subgroup by
Lemma 1.3 (iv).
We are thus reduced to the situation H1 = H2∗ and ϕ : H1 → H2 is
the simply connected cover of H2 .
By our assumptions, H1 is now connected, simply connected and
semisimple. We claim that for any non-trivial Q-simple factor L of H1 ,
CONGRUENCE TOPOLOGY
13
L(R) is not compact. Otherwise, the image of Γ, the arithmetic group,
there is finite and as Γ is Zariski dense, so H1 is not connected. The
strong approximation theorem ([Pl-Ra, Theorem 7.12]) gives now that
H1 (Q) is dense in H1 (Af ). So Proposition 2.1 can be applied to finish
the proof of Proposition 0.1 in the case Γ is a congruence subgroup.
We need to show that it is true also for the more general case when
Γ is only congruence closed. To this end let us formulate the following
Proposition which is of independent interest.
Proposition 2.3. Let Γ ⊆ GLn (Z), G its Zariski closure and Der =
[G0 , G0 ]. Then Γ is congruence closed if and only if Γ ∩ Der is a
congruence subgroup of Der.
Proof. If G0 has no toral factors, this is proved in [Ve], in fact, in this
case a congruence closed Zariski dense subgroup is a congruence subgroup. (Note that this is stated there for general G, but the assumption
that there is no toral factor was mistakenly omitted as the proof there
shows.)
Now, if there is a toral factor, we can assume G is connected, so
ab
G = V × S where V is unipotent and S a torus. Now Γ ∩ [G, G]
is Zariski dense and congruence closed, so it is a congruence subgroup
by [Ve] as before. For the other direction, note that the image of Γ is
U × S, being solvable, is always congruence closed, so the Proposition
follows.
Now, we can end the proof of Proposition 0.1 for congruence closed
subgroups by looking at ϕ on G3 = Γ the Zariski closure of Γ and apply
the proof above to Der(G03 ). It also proves Proposition 0.2.
Of course, Proposition 2.3 is the general form of the following result
from [Ve] (based on [Nori] and [Weis]), which is, in fact, the core of
Proposition 2.3.
Proposition 2.4. Suppose Γ ⊂ G(Z) is Zariski dense, G simply connected and Γ a subgroup of G(Z) which is closed in the congruence
topology. Then Γ is itself a congruence subgroup.
3. The Grothendieck closure
3.1. The Grothendieck Closure of a group Γ.
Definition 3.1. Let ρ : Γ → GL(V ) be a representation of Γ on a
lattice V in a Q-vector space V ⊗ Q. Then we get a continuous hob → GL(Vb ) (where, for a group ∆, ∆
b denotes its
momorphism ρb : Γ
14
A. LUBOTZKY AND T.N. VENKATARAMANA
profinite completion) which extends ρ .
Denote by Clρ (Γ) the subgroup of the profinite completion of Γ,
b : ρb(g)(V ) ⊂ V }. In
which preserves the lattice V : Clρ (Γ) = {g ∈ Γ
fact, since det(ρ̂(g)) = ±1 for every g ∈ Γ and hence also for every
g ∈ Γ̂, for g ∈ Clg (Γ), ρ̂(g)(V ) = V , and hence Clρ (Γ) is a subgroup of
Γ̂. We denote by Cl(Γ) the subgroup
(3.1)
b : ρb(g)(V ) ⊂ V
Cl(Γ) = {g ∈ Γ
∀ lattices V }.
Therefore, Cl(Γ) = ∩ρ Clρ (Γ) where ρ runs through all integral representations of the group Γ.
Suppose now that V is any finitely generated abelian group (not
necessarily a lattice i.e. not necessarily torsion-free) which is also a
Γ-module. Then the torsion in V is a (finite) subgroup with finite
exponent n say. Then nV is torsion free. Since Γ acts on the finite
b also acts on
group V /nV by a finite group via, say, ρ, it follows that Γ
b we have ρb(g)(V /nV ) =
the finite group V /nV via ρb. Thus, for g ∈ Γ
V /nV . Suppose now that g ∈ Cl(Γ). Then g(nV ) = nV by the
definition of Cl(Γ). Hence g(V )/nV = V /nV for g ∈ Cl(Γ). This
is an equality in the quotient group Vb /nV . This shows that g(V ) ⊂
V + nV = V which shows that Cl(Γ) preserves all finitely generated
abelian groups V which are Γ -modules.
By ClZ (Γ) we mean the Grothendieck closure of the (finitely generated) group Γ. It is essentially a result of [Lub] that the Grothendieck
closure ClZ (Γ) is the same as the group Cl(Γ) defined above (in [Lub],
the group considered was the closure with respect to all finitely generated Z modules which are also Γ modules, whereas we consider only
those finitely generated Z modules which are Γ modules and which are
torsion-free; the argument of the preceding paragraph shows that these
closures are the same). From now on, we identify the Grothendieck closure ClZ (Γ) with the foregoing group Cl(Γ).
Notation 3.2. Let Γ be a group, V a finitely generated torsion-free
abelian group which is a Γ-module and ρ : Γ → GL(V ) the corresponding Γ-action. Denote by Gρ the Zariski closure of the image ρ(Γ)
in GL(V ⊗ Q), and G0ρ its connected component of identity. Then
both Gρ , G0ρ are linear algebraic groups defined over Q, and so is
Derρ = [G0ρ , G0ρ ].
CONGRUENCE TOPOLOGY
15
b ∩ GL(V ). Since the profiLet B = Bρ (Γ) denote the subgroup ρb(Γ)
nite topology of GL(V̂ ) induces the congruence topology on GL(V ), Bρ (Γ)
is the congruence closure of ρ(Γ) in GL(V ).
We denote by D = Dρ (Γ) the intersection of B with the derived
subgroup Derρ = [G0 , G0 ]. We thus have an exact sequence
1 → D → B → A → 1,
where A = Aρ (Γ) is an extension of a finite group G/G0 by an abelian
group (the image of B∩G0 in the abelianization (G0 )ab of the connected
component G0 ).
3.2. Simply Connected Representations.
Definition 3.3. We will say that ρ is simply connected if the group
G = Gρ is essentially simply connected. That is, if U is the unipotent
radical of G, the quotient G0 /U is a product H × S where H is semisimple and simply connected and S is a torus.
An easy consequence of Lemma 1.2 is that simply connected representations are closed under direct sums.
Lemma 3.4. Let ρ1 , ρ2 be two simply connected representations of an
abstract group Γ. Then the direct sum ρ1 ⊕ρ2 is also simply connected.
We also have:
Lemma 3.5. Let ρ : Γ → GL(W ) be a sub-representation of a representation τ : Γ → GL(V ) such that both ρ, τ are simply connected.
Then the map r : Bτ (Γ) → Bρ (Γ) is surjective.
Proof. The image of Bτ (Γ) in Bρ (Γ) contains the image of Dτ . By
Proposition 2.3, Dτ is a congruence subgroup of the algebraic group
Derτ . The map Derτ → Derρ is a surjective map between simply
connected groups. Therefore, by part (iv) of Lemma 1.3, the image of
Dτ is a congruence subgroup F of Dρ . Now, by Proposition 2.3, Dρ ·ρ(Γ)
is congruence closed, hence equal to Bρ which is the congruence closure
of ρ(Γ) and Bτ → Bρ is surjective.
3.3. Simply-Connected to General.
Lemma 3.6. Every (integral) representation ρ : Γ → GL(W ) is a subrepresentation of a faithful representation τ : Γ → GL(V ) where τ is
simply connected.
16
A. LUBOTZKY AND T.N. VENKATARAMANA
Proof. Let ρ : Γ → GL(W ) be a representation. Let Der be the derived
subgroup of the identity component of the Zariski closure H = Gρ of
ρ(Γ). Then, by Lemma 1.3(iii), there exists a map H ∗ → H 0 with finite
∗
kernel such that H ∗ is connected and H ∗ /U ∗ = (H ∗ )ss × S ∗ where Hss
is a simply connected semi-simple group. Denote by WQ the Q-vector
space W ⊗ Q. By Lemma 1.3(i), ρ : H 0 → GL(WQ ) may be considered
as a sub-representation of a faithful representation (θ, EQ ) of the covering group H ∗ .
By (ii) of Lemma 1.3, the image of an arithmetic subgroup of H ∗ is
an arithmetic group of H. Moreover, as H(Z) is virtually torsion free,
one may choose a normal, torsion-free arithmetic subgroup ∆ ⊂ H(Z)
such that the map H ∗ → H 0 splits over ∆. In particular, the map
H ∗ → H 0 splits over a normal subgroup N of Γ of finite index. Thus,
θ may be considered as a representation of the group N.
Consider the induced representation IndΓN (WQ ). Since WQ is a representation of Γ, it follows that IndΓN (WQ ) = WQ ⊗IndΓN (trivN ) ⊃ WQ .
Since, by the first paragraph of this proof, WQ ⊂ EQ as H ∗ modules,
it follows that WQ |N ⊂ EQ and hence WQ ⊂ IndΓN (EQ ) =: VQ . Write
τ = IndΓN (EQ ) for the representation of Γ on VQ . The normality of N
in Γ implies that the restriction representation τ |N is contained in a
direct sum of the N-representations n → θ(γnγ −1 ) as γ varies over the
finite set Γ/N.
Write Gθ|N for the Zariski closure of the image θ(N). Since Gθ|N has
∗
∗
H as its Zariski closure and the group Hss
is simply connected, each θ
composed with conjugation by γ is a simply connected representation
of N. It follows from Lemma 3.4 that τ |N is simply connected. Since
simple connectedness of a representation is the same for subgroups
of finite index, it follows that τ , as a representation of Γ, is simply
connected.
We have now proved that there exists Γ-equivariant embedding of
the module (ρ, WQ ) into (τ, VQ ) where W, V are lattices in the Q-vector
spaces WQ , VQ . A basis of the lattice W is then a Q-linear combination
of a basis of V ; the finite generation of W then implies that there exists
an integer m such that mW ⊂ V , and this inclusion is an embedding
of Γ-modules. Clearly, the module (ρ, W ) is isomorphic to (ρ, mW ) the
isomorphism given by multiplication by m. Hence the lemma follows.
CONGRUENCE TOPOLOGY
17
The following is the main technical result of this section, from which
the main results of this paper are derived:
Proposition 3.7. The group Cl(Γ) is the inverse limit of the groups
Bρ (Γ) where ρ runs through simply connected representations and
Bρ (Γ) is the congruence closure of ρ(Γ). Moreover, if ρ : Γ → GL(W )
is simply connected, then the map Cl(Γ) → Bρ (Γ) is surjective.
b
Proof. Denote temporarily by Cl(Γ)sc the subgroup of elements of Γ
which stabilize the lattice V for all simply connected representations
(τ, V ). Let W be an arbitrary finitely generated torsion-free lattice
which is also a Γ-module; denote by ρ the action of Γ on W .
By Lemma 3.6, there exists a simply connected representation (τ, V )
which contains (ρ, W ). If g ∈ Cl(Γ)sc , then τb(g)(V ) ⊂ V ; since Γ is
b and stabilizes W , it follows that for all x ∈ Γ,
b τb(x)(W
c) ⊂ W
c;
dense in G
c ∩ V = W.
in particular, for g ∈ Cl(Γ)sc , ρb(g)(W ) = τb(g)(W ) ⊂ W
Thus, Cl(Γ)sc ⊂ Cl(Γ).
The group Cl(Γ) is, by definition, the set of all elements g of the
b which stabilize all Γ stable torsion free latprofinite completion Γ
tices. It follows in particular, that these elements g stabilize all Γstable lattices V associated to simply connected representations (τ, V );
hence Cl(Γ) ⊂ Cl(Γ)sc . The preceding paragraph now implies that
Cl(Γ) = Cl(Γ)sc . This proves the first part of the proposition (see
Equation 0.2).
We can enumerate all the simply connected integral representations
ρ, since Γ is finitely generated. Write ρ1 , ρ2 , · · · , ρn · · · , for the sequence of simply connected representations of Γ. Write τn for the
direct sum ρ1 ⊕ ρ2 ⊕ · · · ⊕ ρn . Then τn ⊂ τn+1 and by Lemma 3.4 each
τ is simply connected; moreover, the simply connected representation
ρn is contained in τn .
By Lemma 3.5, it follows that Cl(Γ) is the inverse limit of the totally ordered family Bτn (Γ); moreover, Bτn+1 (Γ) maps onto Bτn (Γ). By
taking inverse limits, it follows that Cl(Γ) maps onto the group Bτn (Γ)
for every n. It follows, again from Lemma 3.5, that every Bρn (Γ) is
a homomorphic image of Bτn (Γ) and hence of Cl(Γ). This proves the
second part of the proposition.
18
A. LUBOTZKY AND T.N. VENKATARAMANA
Definition 3.8. Let Γ be a finitely generated group. We say that Γ is
F Ab if the abelianization ∆ab is finite for every finite index subgroup
∆ ⊂ Γ.
Corollary 3.9. If Γ is F Ab then for every simply connected representation ρ, the congruence closure Bρ (Γ) of ρ(Γ) is a congruence subgroup
and Cl(Γ) is an inverse limit over a totally ordered set τn of simply
connected representations of Γ, of congruence groups Bn in groups
Gn = Gτn with G0n simply connected. Moreover, the maps Bn+1 → Bn
are surjective. Hence the maps Cl(Γ) → Bn are all surjective.
Proof. If ρ : Γ → GL(V ) is a simply connected representation, then
for a finite index subgroup Γ0 the image ρ(Γ0 ) has connected Zariski
closure, and by assumption, G0 /U = H × S where S is a torus and H
is simply connected semi-simple. Since the group Γ is FAb it follows
that S = 1 and hence G0 = Der(G0 ). Now Proposition 2.4 implies
that Bρ (Γ) is a congruence subgroup of Gρ (V ). The Corollary is now
immediate from the Proposition 3.7. We take Bn = Bτn in the proof
of the proposition.
We can now prove Theorem 0.5. Let us first prove the direction
claiming that the congruence subgroup property implies Cl(Γ) = Γ.
This was proved for arithmetic groups Γ by Grothendieck, and we
follow here the proof in [Lub] which works for general Γ. Indeed, if
ρ : Γ → GLn (Z) is a faithful simply connected representation such that
ρ(Γ) satisfies the congruence subgroup property, then it means that the
map ρ̂ : Γ̂ → GLn (Ẑ) is injective. Now ρ (Cl(Γ)) ⊆ GLn (Z) ∩ ρ̂(Γ̂), but
the last is exactly the congruence closure of ρ(Γ). By our assumption,
ρ(Γ) is congruence closed, so it is equal to ρ(Γ). So in summary ρ̂(Γ) ⊂
ρ̂ (Cl(Γ)) ⊆ ρ(Γ) = ρ̂(Γ). As ρ̂ is injective, Γ = Cl(Γ).
In the opposite direction: Assuming Cl(Γ) = Γ. By the description
of Cl(Γ) in (0.1) or in (3.1), it follows that for every finite index subgroup Γ′ of Γ, Cl(Γ′ ) = Γ′ (see [Lub, Proposition 4.4]). Now, if ρ is
a faithful simply connected representation of Γ, it is also such for Γ′
and by Proposition 3.6, ρ (Cl(Γ′ )) is congruence closed. In our case
it means that for every finite index subgroup Γ′ , ρ(Γ′ ) is congruence
closed, i.e. ρ(Γ) has the congruence subgroup property.
4. Thin Groups
Let Γ be a finitely generated Z-linear group, i.e. Γ ⊂ GLn (Z), for
some n. Let G be its Zariski closure in GLn (C) and ∆ = G ∩ GLn (Z).
CONGRUENCE TOPOLOGY
19
We say that Γ is a thin subgroup of G if [∆ : Γ] = ∞, otherwise Γ is an
arithmetic subgroup of G. In general, given Γ, (say, given by a set of
generators) it is a difficult question to determine if Γ is thin or arithmetic. Our next result gives, still, a group theoretic characterization
for the abstract group Γ to be thin. But first a warning: an abstract
group can sometimes appear as an arithmetic subgroup and sometimes
as a thin subgroup. For example, the free group on two generators
F = F2 is a finite index subgroup of SL2 (Z), and so, arithmetic. But
at the same time, by a well known result of Tits asserting that SLn (Z)
contains a copy of F which is Zariski dense in SLn [Ti]; it is also thin.
To be precise, let us define:
Definition 4.1. A finitely generated Z-linear group Γ is called a thin
group if it has a faithful representation ρ : Γ → GLn (Z) for some
Z
n ∈ Z, such that ρ(Γ) is of infinite index in ρ(Γ) ∩ GLn (Z) where
Z
ρ(Γ) is the Zariski closure of Γ in GLn . Such a ρ will be called a thin
representation of Γ.
We have assumed that i : Γ ⊂ GLn (Z). Assume also, as we may
(see Lemma 3.6) that the representation i is simply connected. By
b which preserves
Proposition 3.7, the group Cl(Γ) is the subgroup of Γ
the lattices Vn for a totally ordered set (with respect to the relation of
being a sub representation) of faithful simply connected integral representations (ρn , Vn ) of Γ with the maps Cl(Γ) → Bn being surjective,
where Bn is the congruence closure of ρn (Γ) in GL(Vn ). Hence, Cl(Γ)
is the inverse limit (as n varies) of the congruence closed subgroups Bn
and Γ is the inverse limit of the images ρn (Γ). Equip Bn /ρn (Γ) with
the discrete topology. Consequently, Cl(Γ)/Γ is a closed subspace of
Q
the Tychonov product n (Bn /ρn (Γ)). This is the topology on Cl(Γ)/Γ
considered in the following theorem.
Theorem 4.2. Let Γ be a finitely generated Z-linear group, i.e. Γ ⊂
GLm (Z) for some n. Then Γ is not a thin group if and only if Γ
satisfies both of the following two properties:
(a) Γ is an F Ab group (i.e. for every finite index subgroup Λ of Γ,
Λ/[Λ, Λ] is finite), and
(b) The group Cl(Γ)/Γ is compact
Proof. Assume first that Γ is a thin group. If Γ is not F Ab we are
done. So, assume Γ is F Ab. We must now prove that Cl(Γ)/Γ is
not compact. We know that Γ has a faithful thin representation ρ :
20
A. LUBOTZKY AND T.N. VENKATARAMANA
Γ → GLn (Z) which in addition, is simply connected. This induces
a surjective map (see Proposition 3.7) Cl(Γ) → Bρ (Γ) where Bρ (Γ)
is the congruence closure of ρ(Γ) in GLn (Z). As Γ is F Ab, Bρ (Γ) is
a congruence subgroup, by Corollary 3.9. But as ρ is thin, ρ(Γ) has
infinite index in Bρ (Γ). Thus, Cl(Γ)/Γ is mapped onto the discrete
infinite quotient space Bρ (Γ)/ρ(Γ). Hence Cl(Γ)/Γ is not compact.
Assume now Γ is not a thin group. This implies that for every faithful integral representation ρ(Γ) is of finite index in its integral Zariski
closure. We claim that Γ/[Γ, Γ] is finite. Otherwise, as Γ is finitely
generated, Γ is mapped on Z. The group Z has a Zariski dense integral
representation τ into Ga ×S where S is a torus; take any integral matrix
g ∈ SLn (Z) which is neither semi-simple nor unipotent, whose semisimple part has infinite order. Then both the unipotent and semisimple
part of the Zariski closure H of τ (Z) are non trivial and H(Z) cannot
contain τ (Z) as a subgroup of finite index since H(Z) is commensurable
to Ga (Z) × S(Z) and both factors are non trivial and infinite. The representation ρ × τ (where ρ is any faithful integral representation of Γ)
will give a thin representation of Γ. This proves that Γ/[Γ, Γ] is finite.
A similar argument (using an induced representation) works for every
finite index subgroup, hence Γ satisfies F Ab.
We now prove that Cl(Γ)/Γ is compact. We already know that Γ is
F Ab, so by Corollary 3.9, Cl(Γ) = limBρn (Γ) when Bn = Bρn (Γ) are
←
congruence groups with surjective homomorphisms Bn+1 → Bn . Note
that as Γ has a faithful integral representation, we can assume that all
the representations ρn in the sequence are faithful and
(4.1)
Γ = lim
ρn (Γ).
←−
n
This implies that Cl(Γ)/Γ = lim
Bn /ρn (Γ). Now, by our assumption,
←−
n
each ρn (Γ) is of finite index in Bn = Bρn (Γ). So Cl(Γ)/Γ is an inverse
limit of finite sets and hence compact.
5. Grothendieck closure and super-rigidity
Let Γ be a finitely generated group. We say that Γ is integral superrigid if there exists an algebraic group G ⊆ GLm (C) and an embedding
i : Γ0 7→ G of a finite index subgroup Γ0 of Γ, such that for every
integral representation ρ : Γ → GLn (Z), there exists an algebraic representation ρ̃ : G → GLn (C) such that ρ and ρ̃ agree on some finite
index subgroup of Γ0 . Note: Γ is integral super-rigid if and only if a
finite index subgroup of Γ is integral super-rigid.
CONGRUENCE TOPOLOGY
21
Example of such super-rigid groups are, first of all, the irreducible
(arithmetic) lattices in high rank semisimple Lie groups, but also the
(arithmetic) lattices in the rank one simple Lie groups Sp(n, 1) and
F−20
(see [Mar], [Cor], [Gr-Sc]). But [Ba-Lu] shows that there are such
4
groups which are thin groups.
Now, let Γ be a subgroup of GLm (Z), whose Zariski closure is essentially simply connected. We say that Γ satisfies the congruence
subgroup property (CSP) if the natural extension of i : Γ → GLm (Z)
to Γ̂, i.e. ĩ : Γ̂ → GLm (Ẑ) has finite kernel.
Theorem 5.1. Let Γ ⊆ GLm (Z) be a finitely generated subgroup satisfying (F Ab). Then
(a) Cl(Γ)/Γ is compact if and only if Γ is an arithmetic group which
is integral super-rigid.
(b) Cl(Γ)/Γ is finite if and only if Γ is an arithmetic group satisfying
the congruence subgroup property.
Remarks. (a) The finiteness of Cl(Γ)/Γ implies, in particular, its compactness , so Theorem 5.1 recovers the well known fact (see [BMS],
[Ra2]) that the congruence subgroup property implies super-rigidity.
(b) As explained in §2 (based on [Ser]) the simple connectedness is a
necessary condition for the CSP to hold. But by Lemma 3.6, if Γ
has any embedding into GLn (Z) for some n, it also has a simply
connected one.
We now prove Theorem 5.1.
Proof. : Assume first Cl(Γ)/Γ is compact in which case, by Theorem
4.2, Γ must be an arithmetic subgroup of some algebraic group G.
Without loss of generality (using Lemma 3.6) we can assume that G
is connected and simply connected, call this representation ρ : Γ → G.
Let θ be any other representation of Γ.
Let τ = ρ ⊕ θ be the direct sum. The group Gτ is a subgroup of Gρ ×
Gθ with surjective projections. Since both τ and ρ are embeddings of
the group Γ, and Γ does not have thin representations, it follows (from
Corollary 3.9) that the projection π : Gτ → Gρ yields an isomorphism
of the arithmetic groups τ (Γ) ⊂ Gτ (Z) and ρ(Γ) ⊂ Gρ (Z).
Assume, as we may, that Γ is torsion-free and Γ is an arithmetic
group. Every arithmetic group in Gτ (Z) is virtually a product of the
form Uτ (Z) ⋊ Hτ (Z) where Uτ and Hτ are the unipotent and semisimple parts of Gτ respectively (note that G0τ cannot have torus as
quotient since Γ is F Ab). Hence Γ ∩ Uτ (Z) may also be described as
22
A. LUBOTZKY AND T.N. VENKATARAMANA
the virtually maximal normal nilpotent subgroup of Γ. Similarly for
Γ ∩ Uρ (Z). This proves that the groups Uτ and Uρ have isomorphic
arithmetic groups which proves that π : Uτ → Uρ is an isomorphism.
Otherwise Ker(π), which is a Q-defined normal subgroup of Uτ , would
have an infinite intersection with the arithmetic group Γ ∩ Uτ .
Therefore, the arithmetic groups in Hτ and Hρ are isomorphic and
the isomorphism is induced by the projection Hτ → Hρ . Since Hρ
is simply connected by assumption, and is a factor of Hτ , it follows
that Hτ is a product Hρ H where H is a semi-simple group defined
over Q with H(Z) Zariski dense in H. But the isomorphism of the
arithmetic groups in Hτ and Hρ then shows that the group H(Z) is
finite which means that H is finite. Therefore, π : Hτ0 → Hρ is an
isomorphism and so the map G0τ → Gρ is also an isomorphism since it
is a surjective morphism between groups of the same dimension, and
since Gρ is simply connected.
This proves that Γ is a super-rigid group.
In [Lub], it was proved that if Γ satisfies super rigidity in some
simply connected group G, then (up to finite index) Cl(Γ)/Γ is in 1-1
correspondence with C(Γ) = Ker(Γ̂ → G(Ẑ)). This finishes the proof
of both parts (a) and (b).
Remark. In the situation of Theorem 5.1, Γ is an arithmetic group,
satisfying super-rigidity. The difference between parts (a) and (b), is
whether Γ also satisfies CSP. As of now, there is no known arithmetic
group (in a simply connected group) which satisfies super-rigidity without satisfying CSP. The conjecture of Serre about the congruence subgroup problem predicts that arithmetic lattices in rank one Lie groups
(−20)
fail to have CSP. These include Lie groups like Sp(n, 1) and F4
for
which super-rigidity was shown (after Serre had made his conjecture).
Potentially, the arithmetic subgroups of these groups can have Cl(Γ)/Γ
compact and not finite. But (some) experts seem to believe now that
these groups do satisfy CSP. Anyway as of now, we do not know any
subgroup Γ of GLn (Z) with Cl(Γ)/Γ compact and not finite.
References
[BMS] H. Bass, J. Milnor and J.-P. Serre, Solution of the congruence subgroup
problem for SLn (n ≥ 3) and Sp2n (n ≥ 2). Publ. Math. I.H.E.S. 33 (1967),
59–137.
[Ba-Lu] H. Bass and A. Lubotzky, Nonarithmetic superrigid groups: counterexamples to Platonov’s conjecture. Ann. of Math. 151 (2000), 1151–1173.
CONGRUENCE TOPOLOGY
[Bo]
23
A. Borel, Introduction Aux Groupes Arithmetiques. Actualite Scientifiques
et Industrielles 1341, Hermann, Paris, 1969.
[Br-Gr] M.R. Bridson and F.J. Grunewald, Grothendieck’s problems concerning
profinite completions and representations of groups. Ann. of Math. 160
(2004), 359–373.
[Ch]
J.S. Chahal, Solution of the congruence subgroup problem for solvable algebraic groups. Nagoya Math. J. 79 (1980), 141–144.
[Cor] K. Corlette, Archimedean superrigidity and hyperbolic geometry. Ann. of
Math. 135 (1992), 165–182.
[Go]
E.S. Golod, On nil-algebras and finitely approximable p-groups. (Russian)
Izv. Akad. Nauk SSSR Ser. Mat. 28 (1964), 273–276.
[Gri]
R.I. Grigorchuk, On Burnside’s problem on periodic groups. (Russian)
Funktsional. Anal. i Prilozhen. 14 (1980), 53–54.
[Gro] A. Grothendieck, Representations lineaires et compactification profinie des
groupes discrets. Manuscripta Math. 2 (1970), 375–396.
[Gr-Sc] M. Gromov and R. Schoen, Harmonic maps into singular spaces and p-adic
superrigidity for lattices in groups of rank one. Publ. Math. I.H.E.S. 76
(1992), 165–246.
[Lub] A. Lubotzky, Tannaka Duality for Discrete Groups. American J. of Math.
102 (1980), 663–689.
[Mar] G.A. Margulis, Discrete Subgroups of Semi-simple Lie Groups. Ergebnisse
der Mathematik und ihrer Grenzgebiete (3) 17. Springer-Verlag, Berlin,
1991.
[Nori] M.V. Nori, On Subgroups of GLn (Fp ). Invent Math 88 (1987), 257–275.
[Pl-Ra] V. Platonov and A. Rapinchuk, Algebraic Groups and Number Theory. In
series: Pure and Applied Mathematics, V. 139, Academic Press, 1994, 625
pps.
[Pl-Ta1] V. Platonov and O.I. Tavgan, Grothenkieck’s problem on profinite completions of groups. Soviet Math. Dokl. 33 (1986), 822–825.
[Pl-Ta2] V. Platonov and O.I. Tavgan, Grothendieck’s problem on profinite completions and representations of groups. K-Theory 4 (1990), 89–101.
[Py]
L. Pyber, Groups of intermediate subgroup growth and a problem of
Grothendieck. Duke Math. J. 121 (2004), 169–188.
[Ra1] M.S. Raghunathan, Discrete Subgroups of Lie Groups. Ergebnisse der
Mathematik und ihrer Grenzgebiete, Band 68. Springer-Verlag, New YorkHeidelberg, 1972. ix+227 pp.
[Ra2] M.S. Raghunathan, On the Congruence Subgroup Problem. Publ. Math.
I.H.E.S. 46 (1976), 107–161.
[Sar]
P. Sarnak, Notes on Thin Matrix Groups, Thin Groups and Superstrong
Approximation. Math. Sci. Res. Publ. 61 (2014), 343–362, Cambridge University Press, Cambridge.
[Ser]
J-P. Serre, Groupes de Congruence, (d’après H. Bass, H. Matsumoto, J.
Mennicke, J. Milnor, C. Moore). Séminaire Bourbaki, Vol. 10, Exp. No.
330, 275–291, Soc. Math. France, Paris, 1995.
[Ti]
J. Tits, Free subgroups in linear groups. J. Algebra 20 (1972), 250–270.
[Ve]
T.N. Venkataramana, A remark on Extended Congruence Subgroups.
IMRN (1999), no 15, 835–838.
24
A. LUBOTZKY AND T.N. VENKATARAMANA
[Weis] B. Weisfeiler, Strong Approximation for Zariski dense subgroups. Ann of
Math. 120 (1984), 271–315.
Institute of Mathematics, Hebrew University, Jerusalem 9190401,
Israel, [email protected]
Tata Institute of Fundamental Research, Homi Bhabha Road, Colaba, Mumbai 400005, India, [email protected]
| 4math.GR
|
A QUANTITATIVE GENERALIZATION OF PRODANOV-STOYANOV THEOREM
ON MINIMAL ABELIAN TOPOLOGICAL GROUPS
arXiv:1706.05411v2 [math.GN] 20 Jun 2017
TARAS BANAKH
To the memory of Ivan Prodanov (1935–1985)
Abstract. A topological group X is defined to have compact exponent if for some number n ∈ N the set
{xn : x ∈ X} has compact closure in X. Any such number n will be called a compact exponent of X. Our
principal result states that a complete Abelian topological group X has compact exponent (equal to n ∈ N)
if and only if for any injective continuous homomorphism f : X → Y to a topological group Y and every
y ∈ f (X) there exists a positive number k (equal to n) such that y k ∈ f (X). This result has many interesting
implications: (1) an Abelian topological group is compact if and only if it is complete in each weaker Hausdorff
group topology; (2) each minimal Abelian topological group is precompact (this is the famous ProdanovStoyanov Theorem); (3) a topological group X is complete and has compact exponent if and only if it is
closed in each Hausdorff paratopological group containing X as a topological subgroup (this confirms an old
conjecture of Banakh and Ravsky).
1. Introduction
In this paper we prove a theorem, which can be considered as a quantitative version of the fundamental
Prodanov-Stoyanov Theorem on the precompactness of minimal Abelian topological groups. All topological
groups considered in this paper are Hausdorff.
We recall that a topological group X is precompact if its completion in the two-sided uniformity is compact.
This happens if and only if X is totally bounded in the sense that for any neighborhood U ⊂ X of the unit
there exists a finite subset F ⊂ X such that X = F U . For a subset A of a topological space X by Ā we denote
the closure of A in X.
We shall say that a topological group X has compact exponent if for some number n ∈ N the set nX = {xn :
x ∈ X} has compact closure nX in X. In this case, the number n is called a compact exponent of X. Observe
that a topological group is compact if and only if 1 is its compact exponent. For the first time the concept of
compact exponent (without naming) has appeared in Banakh and Ravsky [2].
By a powertopological semigroup we understand a semigroup S endowed with a Hausdorff topology such
that for any a, b ∈ S and n ∈ N the map S → S, x 7→ axn b, is continuous. It is clear that each topological
semigroup is powertopological and each powertopological semigroup is semitopological which means that for
any a, b ∈ S the shift S → S, x 7→ axb, is continuous.
This paper contains a unique principal result:
Theorem 1.1. Let n ∈ N be any number. For a complete Abelian topological group X the following conditions
are equivalent:
(1) X has compact exponent (equal to n);
(2) for any injective continuous homomorphism f : X → Y to a topological group Y and every point
y ∈ f (X) ⊂ Y there exists a number k ∈ N (equal to n) such that y k ∈ f (X);
(3) for any continuous homomorphism f : X → Y to a powertopological semigroup Y and every point
y ∈ f (X) ⊂ Y there exists a number k ∈ N (equal to n) such that y k ∈ f (X).
This theorem has many interesting and non-trivial implications.
1991 Mathematics Subject Classification. 22A05, 22A15, 54D30, 54E15.
Key words and phrases. Abelian topological group, minimal topological group, paratopological group, topological semigroup,
topological group of compact exponent.
1
2
TARAS BANAKH
Corollary 1.2. An Abelian topological group X is compact if and only if X for any injective continuous
homomorphism f : X → Y to a topological group Y the image f (X) is closed in Y .
This corollary can be reformulated as follows.
Corollary 1.3. An Abelian topological group X is compact if and only if X is complete in each weaker
Hausdorff group topology.
This corollary implies the famous Prodanov-Stoyanov Theorem [14] on the precompactness of minimal
Abelian topological groups. We recall [3], [5] that a topological group X is minimal if X does not admit a
strictly weaker Hausdorff group topology.
Corollary 1.4 (Prodanov, Stoyanov). Each minimal Abelian topological group X is precompact.
Proof. The precompactness of X is equivalent to the compactness of the completion X̄ of X. By Corollary 1.2,
the compactness of X̄ will follow as soon as we check that for any continuous injective homomorphism f : X̄ →
Y to a topological group Y the image f (X̄) is closed in Y . We lose no generality assuming that the topological
group Y is complete and f (X̄) is dense in Y . By the minimality of X, the restriction f |X : X → f (X) is
a topological isomorphism. Then it uniquely extends to a topological isomorphism of the completions X̄ and
f (X) = Y of the topological groups X and f (X). So, f : X̄ → Y is a topological isomorphism and f (X̄) = Y
is closed in Y .
Another corollary of Theorem 1.1 confirms an old conjecture of Banakh and Ravsky [2], [16].
Corollary 1.5. For a complete Abelian topological group X the following conditions are equivalent:
(1) X has compact exponent;
(2) for any continuous homomorphism f : X → Y to a Hausdorff powertopological semigroup Y the image
f (X) is closed in Y ;
(3) for any injective continuous homomorphism f : X → Y to a topological group Y the quotient group
f (X)/f (X) is a torsion group;
(4) for any isomorphic topological embedding f : X → Y of X into a Hausdorff paratopological group Y
the image f (X) is closed in Y .
Proof. The equivalences (1) ⇔ (2) ⇔ (3) follow immediately from the corresponding equivalences in Theorem 1.1 and (3) ⇔ (4) was proved by Ravsky [16].
Remark 1.6. As was noticed by Dikranjan and Megrelishvili [5, §5], the Stoyanov-Prodanov Theorem 1.4 fails
for nilpotent groups: the Weyl-Heisenberg group H(w0 ) = H(R)/Z where
1 a b
1 0 b
H(R) = 0 1 c : a, b, c ∈ R
and Z = 0 1 0 : b ∈ Z
0 0 1
0 0 1
is nilpotent Lie group (of nilpotence degree 2), which is minimal but not compact, see also [4] and [12].
Nonetheless, Dikranjan and Uspenskij [8] proved the following two extensions of Corollary 1.2 to nilpotent and
solvable topological groups.
Theorem 1.7 (Dikranjan, Uspenskij). A nilpotent topological group X is compact if and only if for every
continuous homomorphism f : X → Y to a topological group Y the image f (X) is closed in Y .
Theorem 1.8 (Dikranjan, Uspenskij). A solvable topological group X is compact if and only if for every
continuous homomorphism f : Z → Y from a closed normal subgroup Z ⊂ X of X to a topological group Y
the image f (Z) is closed in Y .
Our proof of Theorem 1.1 follows the line of the proof of Stoyanov-Prodanov Theorem from [6], with an
additional care of (non)completeness and compact exponent. Because of that it is quite long and technical.
We separate the proof into several steps. In Section 2 we establish some preliminary results related to bounded
sets in topological groups and properties of topological groups of (pre)compact exponent. In Section 3, on each
Abelian topological group X we define a weaker group topology τ⋄ dependent on a sequence (xn )n∈ω of points
in X and prove two Key Lemmas. The first Key Lemma 3.1 produces a sequence (xn )n∈ω in X with some
A QUANTITATIVE GENERALIZATION OF PRODANOV-STOYANOV THEOREM
3
additional properties and the second Key Lemma 3.3 shows that these properties ensure that the topology τ⋄
is Hausdorff. We shall apply these Key Lemmas once in the proof of the “bounded” version of Theorem 1.1
(with k = n) and three times for the “unbounded” version (with k arbitrary). So, Theorem 1.1 follows from
its “bounded” and “unbounded” versions, proved in Theorem 4.4 and 7.2, respectively.
2. Preliminaries
In this section we fix some standard notations used in the paper. Also we recall some known results and
prove some simple lemmas.
2.1. Standard Notations. By N we denote the set of natural numbers, i.e., positive integer numbers and by
ω the set of non-negative integer numbers (= finite ordinals). By Z we denote the additive group of integer
numbers.
By P we denote the set of all prime numbers. For a number n ∈ N let Pn be the set of prime divisors of n.
For a subset A of a topological space X by Ā we denote the closure of A in X. A point x ∈ X is called an
accumulation point of a sequence (xn )n∈ω in a topological space X if each neighborhood Ox ⊂ X of x contains
infinitely many points xn , n ∈ ω.
For group X by e we denote its unit, i.e., a unique element e such that ex = x = xe. For two subsets A, B
of a group X we put AB = {ab : a ∈ A, b ∈ B}. The powers An are defined by induction: A0 = {e} and
−n
n −1
−1 n
An+1 = An A for n ∈ ω. Also A−1 = {a−1 : a ∈ A} and A
P = (A ) = (A ) for n ∈ ω.
For a sequence
(An )n∈ω of subsets of a group X by n∈ω An we denote the subgroup of X generated by
S
the union n∈ω An .
For a subset A of a group X and a natural number n ∈ N let nA = {an : a ∈ A} be the set of n-th powers
of the elements of A. Observe that nA ⊂ An and nA 6= An in general. If A is a subgroup of an Abelian group
X, then nA is a subgroup of X.
A topological group X is complete if it is complete in its two-sided uniformity (generated by the entourages
{(x, y) ∈ X × X : y ∈ xU ∩ U x} where U runs over neighborhoods of the unit e) of X).
It is well-known (see [1, §3.6] or [18]) that the completion X̄ of a topological group X by its two-sided
uniformity carries a natural structure of the topological group, which contains X as a dense subgroup.
2.2. Bounded sets in topological groups. In this section we establish some properties of bounded and
precompact sets in topological groups.
A subset B of a group X is called U -bounded for a set U ⊂ X if B ⊂ F U ∩ U F for some finite subset F ⊂ X.
A subset X of a topological group is called totally bounded if it is U -bounded for any neighborhood U of the
unit in X. It is well-known that a subset B of a complete topological group X has compact closure B̄ in X if
and only if B is totally bounded. Because of that totally bounded subsets in topological groups are also called
precompact sets, see [1, §3.7].
The following simple (but important) lemma shows that the failure of the total boundedness can be recognized by countable subgroups.
Lemma 2.1. If a subset A of a group X is not U U −1 -bounded for some set U ⊂ X, then for some countable
subgroup Z ⊂ X the set Z ∩ A is not U -bounded.
Proof. Since A is not U U −1 -bounded, for every finite set F = F −1 ⊂ X we have A 6⊂ F U U −1 ∩ U U −1 F and
hence A−1 6⊂ U U −1 F −1 ∪ F −1 U U −1 . By Zorn’s Lemma, there exists a maximal subset E ⊂ A ∪ A−1 which is
U U −1 -separated in the sense that x ∈
/ yU U −1 for any distinct points x, y ∈ E. The maximality of E guarantees
−1
that for any x ∈ A ∪ A there exists y ∈ E such that x ∈ yU U −1 or y ∈ xU U −1 (and hence x ∈ yU U −1 ).
Consequently, A ∪ A−1 = EU U −1 and A ⊂ EU U −1 ∩ U U −1 E −1 . The choice of U ensures that the set E is
infinite. Then we can choose any countable subgroup Z ⊂ X that has infinite intersection Z ∩ E with E.
We claim that the set Z ∩ A is not U -bounded. Assuming the opposite, we could find a finite subset F ⊂ X
such that Z ∩ A ⊂ U F . Since the set Z ∩ E is infinite, for some x ∈ F there are two distinct points y, z ∈ Z ∩ E
such that y n , z n ∈ U x. Then z n ∈ U x ⊂ U U −1 y n , which contradict the choice of the set E.
2.3. Localizations. A topological group X is called p-singular for a prime number p if for any n ∈ N, not
divisible by p the set nX is dense in X. For an Abelian topological group X the union Tp (X) of all p-singular
4
TARAS BANAKH
subgroups of X is the largest (closed) p-singular subgroup of X, see [6, §2.6]. For an Abelian discrete topological
n
group X the set Tp (X) coincides with the p-Sylow subgroup {x ∈ X : ∃n ∈ N xp = e}.
We shall need the following fact whose proof can be found in [6, 2.6.2].
P
Lemma 2.2. If an Abelian topological group is a union of compact subgroups, then the subgroup p∈P Tp (X)
is dense in X.
Lemma 2.3. If X is a compact Abelian topological group, then for any prime number p and any number n ∈ N
we have nTp (X) = pnp Tp (X), where np ∈ ω is the largest number such that pnp divides n.
Lemma 2.4. Let X be an Abelian topological group. For any number n ∈ N we have
X
pnp Tp (X) ⊂ nX,
p∈P
where np ∈ ω is the largest number such that pnp divides n. Moreover, if X is a union of compact subgroups,
P
then p∈P pnp Tp (X) is dense in nX.
Proof. For every p the p-singularity of Tp (X) implies that pnp Tp (X) = nTp (X) ⊂ nX. Taking into account
P
that nX is a group, we conclude that p∈P pnp Tp (X) ⊂ nX.
P
Now assume that X is a union of compact subgroups. To see that p∈P pnp Tp (X) in dense in nX, it suffices
P
for any open set U ⊂ X intersecting nX to find an element x ∈ U ∩ nX ∩ p∈P pnp Tp (X). Since U ∩ nX 6= ∅,
there exists a point z ∈ X such that z n ∈ U . Choose a neighborhood V ⊂ X of z such that V n ⊂ U . Since X
is a union of compact
subgroups, the closure Z̄ of the cyclic group Z := {z n }P
n∈Z is compact. By Lemma 2.2,
P
the subgroup p∈P Tp (Z̄) is dense in Z̄. So, we can find an element v ∈ V ∩ p∈P Tp (Z̄). Lemma 2.3 implies
P
P
P
that nTp (Z̄) = pnp Tp (Z̄). Then v n ∈ V n ∩ p∈P nTp (Z̄) = V n ∩ p∈P pnp Tp (Z̄) ⊂ U ∩ p∈P pnp Tp (X).
2.4. The Bohr topology and Følner Theorem. The Bohr topology on a semitopological group X is the
weakest topology on X in which all continuous homomorphisms into compact topological groups remain continuous. The Bohr topology is not necessarily Hausdorff. Topological groups with Hausdorff Bohr topology
and called maximally almost periodic.
In the proof of Theorem 1.1 we shall apply the following fundamental theorem of Følner [10], whose proof
can be found in [6, 1.4.3].
Theorem 2.5 (Følner, 1954). Let X be an Abelian semitopological group such that for every n ∈ Z the map
X → X, x 7→ xn , is continuous. If the group X is E-bounded for some subset E ⊂ X, then for any neighborhood
U ⊂ X the set U U −1 (EE −1 )8 is a neighborhood of the unit in the Bohr topology of X.
3. Two Key Lemmas
In this section we prove the two lemmas following the ideas of the proof of Lemma 2.7.3 from [6]. The first
lemma will help us to construct a nice sequence (xn )n∈ω in an Abelian topological group X. This sequence
determines a weaker group topology τ⋄ on X and in the second lemma we find conditions under which this
weaker topology is Hausdorff.
Lemma 3.1. Let X be an Abelian group, (zk )k∈ω be a sequence in X, (Wk )n∈ω be a decreasing sequence of
sets in X and (Gk )k∈ω be a sequence of subgroups of X satisfying the following conditions:
(1) for every k ∈ N the set Wk contains the unit e of X and Wk = Wk−1 ;
(2) for any n, k ∈ N either nG0 ⊂ Gk or nG0 is not Wk2 Gk -bounded.
Then there exists a sequence of points (xm )m∈ω in G0 satisfying the conditions:
/ Fm Wk Gk , where
(3) for any numbers m ∈ N, n ≤ 2m+1 and k ≤ m with nG0 6⊂ Gk we have xnm ∈
εm−1
(4) Fm = ziε0 xε11 · · · xm−1
: i ≤ m, ε0 , ε1 , . . . , εm−1 ∈ Z, max |εi | ≤ 2m+1 .
0≤i<m
Proof. The construction of the sequence (xk )k∈N is inductive. Let x0 = e and assume that for some m ∈ N
the points x0 , . . . , xm−1 have been constructed. Consider the finite set Fm defined by the condition (4) of
Lemma 3.1. Let Pm be the set of all pairs (n, k) ∈ N × N such that k ≤ n ≤ 2m+1 and nG0 6⊂ Gk . By the
condition (2), for every (n, k) ∈ Pm the set nG0 is not Wk2 Gk -bounded.
A QUANTITATIVE GENERALIZATION OF PRODANOV-STOYANOV THEOREM
5
The group G0 , being Abelian, is amenable and hence admits an invariant finitely additive invariant probability measure µ : P(G0 ) → [0, 1] defined on the Boolean algebra of all subsets of G0 , see [13, 0.15].
Claim 3.2. For any (n, k) ∈ Pm and any z ∈ Fm the set Sn,k,z = {x ∈ G0 : xn ∈ zWk Gk } has measure
µ(Sn,k,z ) = 0.
Proof. Using Zorn’s Lemma, choose a maximal subset M ⊂ G0 such that xn ∈
/ y n Wk2 Gk for any distinct points
x, y ∈ M . The maximality of M guarantees that for every x ∈ G0 there exists y ∈ M such that xn ∈ y n Wk2 Gk
or y n ∈ xn Wk2 Gk and hence xn ∈ y n Wk−2 Gk = y n Wk2 Gk . This means that nX ⊂ M Wk2 Gk . Since the set nG0
is not Wk2 Gk -bounded, the set M is infinite.
We claim that for any distinct points x, y ∈ M the sets xn Wk Gk and y n Wk Gk are disjoint. Assuming that the
−1 −1
n
intersection xn Wk Gk ∩y n Wk Gk contains some point z, we conclude that y n ∈ zWk−1 G−1
k ⊂ x Wk Gk Wk Gk =
xn Wk2 Gk , which contradicts the choice of the set M .
Therefore, the family (xn Wk Gk )x∈M is disjoint and so is the family (xn zWk Gk )x∈M
. Now consider the homo
morphism p : G0 → G0 , p : x 7→ xn and observe that the family p−1 (xn zWk Gk ) x∈M = xp−1 (zWk Gk ) x∈M
is disjoint. Now the additivity and invariantness of the measure µ ensures that the set p−1 (zWk Gk ) = {x ∈
X : xn ∈ zWk Gk } = Sn,k,z has measure µ(Sn,k,z ) = 0.
S
S
The additivity of the measure µ and Claim 3.2 imply that the set S = (n,k)∈Pm z∈Fm Sn,k,z has measure
zero. Consequently, we can find a point xm ∈ G0 \ S. It is clear that this point xm satisfies the condition (3)
of Lemma 3.1.
Let X be an Abelian topological group and (xn )n∈N be a sequence of points in X. For every positive real
number δ consider the set
n
m
o
n
X
X
|εi |
εi = 0 ⊂ X.
<
δ,
♦δ = xε11 · · · xεmm : m ∈ N, ε1 , . . . , εm ∈ Z,
i
2
i=1
i=1
Let T be the topology of X and B be the largest precompact group topology on the group X, i.e., B is the
Bohr topology of the group X, endowed with the discrete topology. Let Te = {T ∈ T : e ∈ T = T −1 } and
Be = {B ∈ B : e ∈ B = B −1 } be the families of open symmetric neighborhoods of the unit e in the topological
groups (X, T ) and (X, B), respectively.
It is easy to see that the family
τe := {T · (♦δ ∩ B) : T ∈ T , δ > 0, B ∈ B}
is a neighborhood base at the unit for some group topology τ⋄ on X. This topology τ⋄ will be called the
topology, ♦-generated by the sequence (xn )n∈ω , and this sequence will be called ♦-generating the topology τ⋄ .
If the topology τ⋄ is Hausdorff, then its ♦-generating sequence (xn )n∈ω will be called ♦-Hausdorff. In this
case X⋄ := (X, τ⋄ ) is a Hausdorff topological group and we can consider its completion X̄⋄ , which is a complete
topological group.
The following lemma detects ♦-Hausdorff sequences.
Lemma 3.3. Let X be an Abelian topological group, (zn )n∈ω be a sequence in X, (Wk )n∈ω be a decreasing
sequence of neighborhoods of the unit e in X and (Gk )k∈ω be a sequence of closed subgroups of X satisfying
the following conditions:
(1)
(2)
(3)
(4)
2
for every k ∈ N we have Wk = Wk−1 and Wk+1
⊂ Wk ;
2
for any n ∈ N and k ≤ n either
nG
⊂
G
or
nG
0
k
0 is not Wk Gk -bounded;
T
the closed subgroup Gω := k∈ω Gk is compact;
for every neighborhood U ⊂ X of the unit there exists k ∈ N such that Gk ⊂ Gω U .
If a sequence (xm )m∈ω of points of G0 satisfies the condition (3) of Lemma 3.1, then it is ♦-Hausdorff and
the set {xm }m∈ω is totally bounded in the topological group (X, τ⋄ ).
Proof. We separate the proof of this lemma into four claims.
Claim 3.4. For every k ∈ ω we have ♦1 ∩ Wk Gk ⊂ Gk .
6
TARAS BANAKH
Proof. To derive a contradiction, assume that the set (♦1 ∩ Wk Gk ) \ Gk contains some point x. Write x as
Pl
x = xε11 · · · xεl l for some number l ∈ N and some integer numbers ε1 , . . . , εl such that i=1 |ε2ii | < 1. Since
ε
/ Gk and thus
/ Gk . Let j ≤ l be the largest number such that xj j ∈
x ∈
/ Gk , for some i ≤ l, we have xεi i ∈
εi
εj G0 6⊂ Gk . The maximality of j guarantees that xi ∈ Gk for every i ∈ (j, l], and hence
ε
Wk Gk ∋ x = xε11 · · · xj j Gk
ε
−ε
and xj j ∈ x1−ε1 · · · xj−1j−1 Wk Gk G−1
k ⊂ Fj Wk Gk , which contradicts the property (3) of Lemma 3.1.
Claim 3.5. For any k ∈ N and any x ∈ ♦1 \ Gk there exists ε > 0 such that x ∈
/ ♦ε Wk .
|δi |
i=1 pi < 1.
δi
/ Gk .
Since x ∈
/ Gk , for some i ≤ n we have xi ∈
We claim that the real number ε = 21n has the required property. To derive a contradiction, assume that
x ∈ ♦ε Wk and find a point y ∈ P
♦ε such that x ∈ yWk . Write y as y = xε11 · · · xεmm for some integer numbers
m
m > n and ε1 , . . . , εm such that i=1 2εii < ε = 21n . The latter inequality implies that εi = 0 for all i ≤ n.
The inclusion x ∈ yWk implies that y −1 x ∈ Wk and y −1 x = xγ11 · · · xγmm where γi = δi for i ≤ n and γi = −εi
γ
/ Gk .
for n < i ≤ m. It follows that |γi | ≤ max{|δi |, |εi |} < 2i . Let j ≤ m be the largest number such that xj j ∈
/ Gk for some i ≤ n. Then γi G0 6⊂ Gk .
Such number j exists since xγi i = xδi i ∈
γj−1 γj
xj Gk
The maximality of j guarantees that xδi i ∈ Gk for all i ∈ (j, m]. It follows that Wk ∋ y −1 x ∈ xγ11 · · · xj−1
γ
−γj−1
1
W
G
⊂
F
W
G
,
which
contradicts
the
condition
(3)
of
Lemma
3.1.
and hence xj j ∈ x−γ
·
·
·
x
k k
j k k
1
j−1
Proof. Write x as x = xδ11 · · · xδnn for some n ∈ N and some integer numbers δ1 , . . . , δn such that
Pn
Claim 3.6. The topology τ⋄ is Hausdorff.
Proof. Given an element x ∈ X \ {e}, we should find T ∈ Te , ε > 0 and B ∈ Be such that x ∈
/ T · (♦ε ∩ B).
If x does not belong to the closure H̄ of the subgroup H, generated by the set {xn }n∈N , then we can find a
neighborhood T = T −1 ∈ Te of unit in X such that xT ∩ H = ∅ and conclude that T ♦1 ⊂ T H does not contain
x.
S
Next, we consider the case x ∈ H̄ \ Gω = k∈ω H̄ \ Gk and hence x ∈
/ Gk for some k ∈ N. Since the subgroup
Gk is closed in X, there exists a neighborhood T ∈ Te of the unit such that T = T −1 ⊂ Wk+1 and xT ∩ Gk = ∅.
If x ∈
/ T ♦1, then, we are done. So, assume that x ∈ T ♦1. In this case there exists z ∈ ♦1 such that x ∈ T z.
It follows that z ∈ xT −1 ∩ ♦1 = xT ∩ ♦1 ⊂ ♦1 \ Gk . Applying Claim 3.5, find ε > 0 such that z ∈
/ ♦ε Wk . We
2
claim that x ∈
/ ♦ε Wk+1 . In the opposite case, z ∈ xT −1 ⊂ xWk+1 ⊂ ♦ε Wk+1
⊂ ♦ε Wk , which contradicts the
choice of ε.
Finally, we consider the case x ∈ Gω . Choose a neighborhood U ∈ Te of the unit such that U = U −1 and
x∈
/ U 37 . By the condition (4) of Lemma 3.3, there exist k ∈ N such that Gk ⊂ Gω U . By the compactness of
Gω , there exists a finite set F ⊂ Gω ⊂ Gk such that Gω ⊂ F U . Then Gk ⊂ F U 2 . Consider the neighborhood
V = Gk ∩ U 2 of e in the group Gk and observe that Gk = F V . By Følner Theorem 2.5, the set (V V −1 )9 is a
neighborhood of the unit in the Bohr topology of the group Gk endowed with the discrete topology. The Baer
Theorem [17, 4.1.2] on extensions of homomorphisms into divisible groups implies that the Bohr topology of the
group Gk coincides with the subspace topology inherited from the Bohr topology on the group X. Consequently,
there exists a Bohr neighborhood B ∈ Be or e such that B ∩ Gk ⊂ (V V −1 )9 ⊂ U 36 . Let T = Wk ∩ U . We claim
that the neighborhood T · (♦1 ∩ B) ∈ τ⋄ of e does not contain the point x. Assuming that x ∈ T · (♦1 ∩ B),
we can find points t ∈ T and b ∈ ♦1 ∩ B with x = tb and conclude that b = xt−1 ∈ Gω T −1 ⊂ Gω Wk and
b ∈ ♦1 ∩ Gω Wk ⊂ Gk according to Claim 3.4. Then b ∈ B ∩ Gk ⊂ U 36 and x ∈ T b ⊂ U U 36 = U 37 , which
contradicts the choice of the neighborhood U .
Claim 3.7. The set {xn }n∈ω is totally bounded in the topological group (X, τ⋄ ).
Proof. Given a neighborhood U ∈ τ⋄ of e, we need to find a finite set F ⊂ X such that {xn }n∈ω ⊂ F U . We
can assume that the neighborhood U is of basic form U = T (♦ε ∩ B) for some T = T −1 ∈ Te , ε > 0 and
B = B −1 ∈ Be .
Let F ⊂ {xn }n∈ω be a maximal set such that x ∈
/ yU for any distinct elements x, y ∈ F . The maximality
of F guarantees that for any point x ∈ {xn }n∈ω there exists a point y ∈ F such that x ∈ yU or y ∈ xU (and
x ∈ yU −1 = yU ). So, {xn }n∈ω ⊂ F U . We claim that the set F is finite. To derive a contradiction, assume
that F is infinite.
A QUANTITATIVE GENERALIZATION OF PRODANOV-STOYANOV THEOREM
7
Choose a neighborhood D = D−1 ∈ Be of e such that D2 ⊂ B. The total boundedness of the topology B
yields a finite set E ⊂ X such that X = ED. By the Pigeonhole principle, there exists a point z ∈ E such
that the set zD contains two distinct points xn , xm ∈ E with 21n + 21m < ε. The last inequality ensures that
−1
−1
xn x−1
= DD−1 ⊂ B and hence xn x−1
m ∈ ♦ε . On the other hand, xn xm ∈ (zD)(zD)
m ∈ ♦ε ∩ B ⊂ U and
xn ∈ xm U , which contradicts the definition of the set F . This contradiction shows that the set F is finite and
the set {xk }k∈ω ⊂ F U is totally bounded.
4. Topological groups of compact exponent
We recall that a topological group X has compact exponent if for some n ∈ N the set nX = {xn : x ∈ X}
has compact closure in X.
Lemma 4.1. If a complete Abelian topological group has compact exponent, then X contains a ♦-Hausdorff
sequence (xm )m∈ω which has an accumulation point x∞ in the completion X̄⋄ of the topological group X⋄ :=
(X, τ⋄ ) such that for every n ∈ N the power xn∞ belongs to X⋄ if and only if the set nX is precompact.
Proof. Let s ∈ N be the smallest number such that sX is precompact. Using Lemma 2.1, find a countable
subgroup Z = {zk }k∈ω ⊂ X such that for every n < s the set nZ is not precompact.
Put G0 = Z̄ and Gk = sZ for all k ∈ N.
For every prime number p let sp ∈ ω be the largest number such that psp divides s. The
Pcompactness of
exponent implies that X is a union of compact subgroups. By Lemma 2.4, the subgroup p∈P psp Tp (Z̄) is
dense in sZ.
Claim 4.2. For any prime divisor q of s the subgroup psq −1 Tq (Z̄) is not precompact.
tp
Proof. Consider the number t = s/q and for every p ∈ P let tp be the largest number
P such tpthat p divides
t. Observe that tq = sq − 1 and tp = sp for p 6= q. By Lemma 2.4, the subgroup p∈P p Tp (Z̄) is dense
P
in tZ. Assuming that q sq −1 Tq (Z̄) is precompact, we would conclude that the subgroup p∈P ptp Tp (Z̄) =
P
q
q sq −1 Tq (Z̄) · p∈P psp Tp (Z̄) ⊂ q s −1 Tq (Z̄) · sX is precompact and so is its closure tZ. But this contradicts the
minimality of s.
Choose a neighborhood W0 ⊂ X of the unit such that for every divisor p of s the (non-precompact) subgroup
psp −1 Tp (Z̄) is not W02 sZ-bounded. Choose a decreasing sequence (Wk )k∈N of neighborhoods of the unit such
2
that Wk = Wk−1 and Wk+1
⊂ Wk .
It is easy to see that the sequences (Wk )k∈ω and (Gk )k∈ω satisfy the conditions (1)–(2) of Lemma 3.1 and
(1)–(4) of Lemma 3.3. Applying these lemmas, we obtain a ♦-Hausdorff sequence (xm )m∈ω in G0 satisfying the
conditions (3),(4) of Lemma 3.1 and such that the set {xn }n∈ω is totally bounded in the Hausdorff topological
group X⋄ = (X, τ⋄ ) and hence has an accumulation point x∞ ∈ X̄⋄ in the completion X̄⋄ of X⋄ = (X, τ⋄ ).
Claim 4.3. For every n ∈ N the power xn∞ belongs to X⋄ if and only if the subgroup nX is precompact.
Proof. If the set nX is precompact, then its closure nX ⊂ X is compact and hence closed in X̄⋄ . Then the
dense subset {x ∈ X̄ : xn ∈ nX} ⊃ X⋄ is closed in X̄ and hence contains the point x∞ . Consequently,
xn∞ ∈ X⋄ .
Next, we assume that the set nX is not precompact. In this case we should prove that xn∞ ∈
/ X⋄ . To derive
a contradiction, assume that xn∞ ∈ X⋄ . First we observe that the number s does not divide n. Otherwise
nX ⊂ sX would be precompact. Consequently, for some prime number q with sq > 0 the power q sq does not
divide n.
Since xn∞ ∈ X⋄ , for every neighborhood V ⊂ X of the unit the neighborhood xn∞ ♦1 V ⊂ xn∞ G0 V ∈ τ⋄
intersects the set {xnk }k∈ω ⊂ G0 , which implies that xn∞ ∈ Ḡ0 = Z̄. Then there exists a number l ∈ ω such
1
and observe that zl ♦ε Ws ∈ τ⋄ is a neighborhood of xn∞ in the topology τ⋄ .
that xn∞ ∈ zl Ws . Let ε = 2l+s
n
Since x∞ is an accumulation point of the sequence (xnk )k∈ω , there exists a number k > max{l, ns} such that
xnk ∈ zl ♦ε Ws and hence xnk ∈ zl yWs for some y ∈ ♦ε . Write the point y as y = xε11 · · · xελλ for some λ > k and
P
P
1
. The last inequality implies
some integer numbers ε1 , . . . , ελ such that λi=1 εi = 0 and λi=1 |ε2ii | < ε = 2l+s
that εi = 0 for all i ≤ l + s and |εi | < 2i−s for all i ≤ λ.
8
TARAS BANAKH
It follows that zl Ws ∋ y −1 xnk = xδ11 · · · xδλλ where δk = n − εk and δi = −εi for all i ∈ {1, . . . , λ} \ {k}. We
claim that |δi s| < 2i+1 for all i ≤ λ. If i 6= k, then |δi s| = |εi |s < 2i−s s < 2i . For i = k we have
|δi |s = |n − εk |s ≤ ns + |εk |s ≤ k + 2k−s s < 2k+1 = 2i+1 .
P
Since λi=1 δi = n and n is not divisible by q sq , for some i ≤ λ, the number δi is not divisible by q sq . Let
j ≤ λ be the largest number such that δj is not divisible by q sq . Since δi = −εi = 0 for i ≤ l, the number j is
greater than l.
The maximality of j guarantees that xδi i ∈ q sq G0 for all i ∈ (j, λ] and hence
−δ
−δ
δ
−δ
λ
1
1
⊂ x−δ
xj j ∈ x−δ
· · · xj−1j−1 zl Ws q sq Z̄.
· · · xj−1j−1 zl W1 xj+1j+1 · · · x−δ
1
1
λ
Let s′ = s/q sq and consider the power
δ s′
xj j
′
−δ
s′
′
′
1s
∈ x−δ
· · · xj−1j−1 zls Wss (s′ psp Z̄) ⊂ Fj W1 (sZ̄) ⊂ Fj W1 G1 ,
1
which contradicts the condition (3) of Lemma 3.1 as maxi≤j |δi s′ | ≤ maxi≤j 2i+1 = 2j+1 and δj s′ Z 6⊂ G1 = sZ.
To see the last non-inclusion, for every prime number p let tp be the largest number such that ptp divides δj s′ .
It follows that tq <
tp ≥ sp for p 6= q. By Lemma 2.4, the closure of the set δj s′ Z coincides with the
Psq and
tp
closure of the set p∈P p Tp (Z̄), which is not precompact because it contains the set q tq Tq (Z̄) ⊃ q sq −1 Tq (Z̄)
which is not precompact according to Claim 4.2. Then the set δj s′ Z is not precompact and hence cannot be
contained in the precompact set sZ = G1 .
Lemma 4.1 implies the “bounded” (or rather “quantitative”) part of Theorem 1.1.
Theorem 4.4. For a number n ∈ N and a complete Abelian topological group X of compact exponent the
following conditions are equivalent:
(1) the subgroup nX is totally bounded;
(2) for any closed subsemigroup Z ⊂ X, any continuous homomorphism f : Z → Y to a powertopological
semigroup Y and any y ∈ f (Z) ⊂ Y we have y n ∈ f (X);
(3) for any injective continuous homomorphism f : X → Y to a topological group Y and any y ∈ f (X) we
have y n ∈ f (X).
Proof. To prove that (1) ⇒ (2), assume that the subgroup nX is totally bounded and hence has compact closure
nX in the complete topological group X. Fix a closed subsemigroup Z ⊂ X and a continuous homomorphism
f : Z → Y to a powertopological semigroup Y . Being compact, the set Z ∩ nX has compact (and hence
closed) image f (Z ∩ nX) in the powertopological semigroup Y . The continuity of the power map π : Y → Y ,
π : y 7→ y n , implies that the set F = {y ∈ Y : y n ∈ f (Z ∩ nX)} is closed in Y . Since f (X) ⊂ F , the
closed set F contains the closure f (X) of f (X) in Y . Then any point y ∈ f (X) belongs to F and hence has
y n ∈ f (Z ∩ nX) ⊂ Y .
The implication (2) ⇒ (3) is trivial and (3) ⇒ (1) follows from Lemma 4.1.
Remark 4.5. Theorem 4.4(2) cannot be further extended to semitopological semigroups: just take any infinite
discrete topological group X of compact exponent and consider its one-point compactification Y = X ∪ {∞}.
Extend the group operation of X to a semigroup operation on Y letting y∞ = ∞ = ∞y for all y ∈ Y . It is
easy to check that Y is a compact Hausdorff semitopological semigroup containing X as a non-closed discrete
subgroup.
5. Topological groups of hypocompact exponent
We shall say that a topological group X has hypocompact exponent if for any neighborhood U ⊂ X of the
unit there exist n ∈ N and a finite subset F ⊂ X such that nX ⊂ F U . It is clear that each topological group
of precompact exponent has hypocompact exponent.
Lemma 5.1. If a topological group X has hypocompact exponent, then each finitely generated Abelian subgroup
A of X is precompact.
A QUANTITATIVE GENERALIZATION OF PRODANOV-STOYANOV THEOREM
9
Proof. Given any neighborhood U ⊂ X of the unit, we need to find a finite subset F ⊂ X such that A ⊂ F U .
Since X has hypocompact exponent, there exist n ∈ N and F1 ⊂ X such that nX ⊂ F1 U . Since the group A is
Abelian and finitely generated, the quotient group A/nA is finite. So, we can find a finite subset F2 ⊂ A such
that A = F2 · nA. Then A = F2 · nA ⊂ F2 F1 U and hence the finite set F = F1 F2 has the required property.
Lemma
5.2. If a complete Abelian topological group X has hypocompact exponent, then the subgroup ωX :=
T
nX
is compact and for every U ⊂ X there exists n ∈ N such that nX ⊂ ωX · U .
n∈N
Proof. To see that the closed set ωX is compact, it suffices to check that it is totally bounded in the complete
topological group X. Given any open neighborhood U ⊂ X of the unit, use the hypocompactness of the
exponent of X and find m ∈ N and a finite set F ⊂ X such that mX ⊂ F U and hence ωX ⊂ mX ⊂ F Ū ,
which means that ωX is precompact and hence compact.
We claim that nX ⊂ ωX · U for some n ∈ N. In the opposite case, for every n ∈ N we could choose a point
xn ∈ n!X \ ωXU . Taking into account that the topological group X has hypocompact exponent, we see that
the set {xn }n∈ω is precompact in the complete topological group X and hence the sequence (xn )n∈ω has an
accumulation pointTx∞ ∈ X \ ωXU . On the other hand, for every m ≥ n we have xm ∈ m!X ⊂ n!X, which
implies that x∞ ∈ n∈N nX and this is a desired contradiction.
We shall say that a subset B of a topological group X is precompact modulo a set A ⊂ X if for any open
neighborhood U ⊂ X of the unit, the set B is AU -bounded.
Lemma 5.3. Let A, B be two subsets of an Abelian topological group. If A is precompact modulo B, then for
any n ∈ N the set nA = {an : a ∈ A} is precompact modulo nB.
Proof. Given any neighborhood U ⊂ X of the unit, find a neighborhood V ⊂ X of the unit such that nV ⊂ U .
Since A is precompact modulo B, there exists a finite subset F ⊂ X such that A ⊂ F BV . It follows that for
every a ∈ A, there are points f ∈ F , b ∈ B and v ∈ V such that a = f bv. Then an = f n bn v n and hence
nA ⊂ (nF )(nB)(nV ) ⊂ (nF )(nB)U , which means that nA is precompact modulo nB.
A topological group X is ω-narrow if for any neighborhood U ⊂ X of the unit there exists a countable
subset C ⊂ X such that X = CU . It is clear that each separable topological group is ω-narrow.
Lemma 5.4. Assume that an ω-narrow complete Abelian topological group X has hypocompact exponent and
for some prime number p the p-component Tp (X) of X does not have compact exponent. Then X contains a
♦-Hausdorff sequence {xm }m∈ω ⊂ Tp (X) which has an accumulation point x∞ in X̄⋄ such that xn∞ ∈
/ X⋄ for
all n ∈ N.
Proof. First we prove the following statement.
Claim 5.5. For every k ∈ ω the set pk Tp (X) is not precompact modulo pk+1 Tp (X).
Proof. To derive a contradiction, assume that for some k ∈ ω the set pk Tp (X) is precompact modulo pk+1 Tp (X).
Lemma 5.3 implies that for every n ≥ k the set pn Tp (X) is precompact modulo pn+1 Tp (X) and hence pk Tp (X)
is precompact modulo pn Tp (X). We claim that pk Tp (X) is precompact. Indeed, for every neighborhood U
of the unit, by the hypocompactness of the exponent of X, there exist n ∈ N and finite subset F ⊂ X such
that nTp (X) ⊂ F U . Let d ∈ ω be the largest number such that pd divides n. Since the subgroup Tp (X) is
p-singular, the subgroup nTp (X) is dense in pd Tp (X). Consequently,
pd Tp (X) ⊂ pd Tp (X) = nTp (X) ⊂ nX ⊂ F Ū .
Since pk Tp (X) is precompact modulo pn Tp (X), there exists a finite subset E ⊂ X such that pk Tp (X) ⊂
(pn Tp (X))EU ⊂ F U EU = (F E)U 2 , which means that pk Tp (X) is precompact and hence the closed subgroup
Tp (X) of the complete group X has compact exponent. But this contradicts the assumption of Lemma 5.4.
Using Claim 5.5, choose a decreasing sequence (Wk )k∈ω of neighborhoods of the unit in X such that for
every k ∈ ω the following conditions are satisfied:
2
(1) Wk = Wk−1 and Wk+1
⊂ Wk ;
(2) pk Tp (X) is not Wk3 · pk+1 X-bounded.
10
TARAS BANAKH
The topological group X, being ω-narrow, contains a countable subset Z = {zk }k∈ω such that X = ZWk
for all k ∈ ω. For every k ∈ ω let Gk = pk Tp (X). Since the subgroup Tp (X) is p-singular, for every n ∈ N
we have nTp (X) = pd Tp (X) where d ∈ ω is the largest number such that pd divides n. This observation, the
choice of the sequence (Wk )k∈ω and Lemma 5.2 guarantee that the sequences (Wk )k∈ω and (Gk )k∈ω satisfy
the conditions (1),(2) of Lemma 3.1 and (1)–(4) of Lemma 3.3.
Applying Lemmas 3.1, construct a sequence (xn )n∈ω satisfying the conditions (3),(4) of this lemma. Applying
Lemma 3.3, we conclude that this sequence is ♦-Hausdorff and the set {xm }m∈ω is totally bounded in the
topological group X⋄ = (X, τ⋄ ). Choose any accumulation point x∞ ∈ X̄⋄ of the sequence (xn )n∈ω (which
exists since the set {xm }m∈ω is totally bounded in the complete topological group X̄⋄ ).
Claim 5.6. For every n ∈ N the power xn∞ ∈ X̄⋄ does not belong to X⋄ .
Proof. To derive a contradiction, assume that xn∞ ∈ X⋄ . Let d ∈ ω be the largest number such that pd divides
n. The p-singularity of the subgroup Tp (X) and Claim 5.5 guarantee that
nG0 = nTp (X) = pd Tp (X) 6⊂ pd+1 Tp (X) = Gd+1 .
Since X = ZWd+1 , there exists a number l ∈ ω such that xn∞ ∈ zl Wd+1 . Let ε = 21l . Since the sequence
accumulates at xn∞ in X⋄ , the neighborhood zl ♦ε Wd+1 ∈ τ⋄ of xn∞ contains a point xnλ with λ >
max{l, n}. Consequently, xnλ ∈ zl yWd+1 for some y ∈ ♦ε . Write y as y = xε11 · · · xεmm for some m ≥ λ and some
P
Pm |εi |
1
integer numbers ε1 , . . . , εm such that m
i=1 εi = 0 and
i=1 2i < ε = 2l . The last inequality implies that
εi = 0 for all i ≤ l and |εi | < 2i for all i ≤ m.
It follows that xnλ y −1 ∈ zl Wd+1 and xnλ y −1 = xδ11 · · · xδmn where δλ = n − ελ and δi = −εi for all i ∈
λ
λ+1
{1, . . . , m} \ {λ}. It follows that
and hence |δi | < 2i+1 for all i ≤ m.
Pm|δλ | = |n − ελ | ≤ n + 2 < 2
d+1
Taking into account that
, we conclude that for some j ≤ m the
i=1 δi = n is not divisible by p
number δj is not divisible by pd+1 . We can assume that j is the largest possible number with this property,
i.e. xδi i ∈ pd+1 Tp (X) ⊂ Gd+1 for all i ∈ (j, m]. It follows that j > max{l} (as δi = 0 for i ≤ l). Let c ∈ ω be
the largest number such that pc divides δj . Taking into account that δj is not divided by pd+1 , we conclude
that c ≤ d and Gd ⊂ Gc . By the p-singularity of Tp (X), G|δj | = δj Tp (X) = pc Tp (X) = Gc 6⊂ Gc+1 .
(xnm )m∈ω
δ
Then zl Wd+1 ∋ xnλ y −1 = xδ11 · · · xδmm ∈ xδ11 · · · xj j Gd+1 and hence
δ
−δ
1
xj j ∈ x−δ
· · · xj−1j−1 zl Wd+1 Gd+1 ⊂ Fj Wc+1 Gc+1 ,
1
which contradicts the condition (3) of Lemma 3.1.
Next we treat the case of topological groups with hypocompact exponent whose p-components have compact
exponent.
Lemma 5.7. Assume that an ω-narrow complete Abelian topological group X has hypocompact exponent and for
every prime number p the p-component Tp (X) has precompact exponent. If X fails to have compact exponent,
then X contains a ♦-Hausdorff sequence (xn )n∈ω which has an accumulation point x∞ ∈ X̄⋄ such that xn∞ ∈
/ X⋄
for all n ∈ N.
T
Proof. By Lemma 5.2, the subgroup ωX = n∈N nX is compact and for every neighborhood U ⊂ X of the
unit there exists a number n ∈ N such that nX ⊂ ωXU .
By our assumption, for every prime number p the p-component Tp (X) of X has precompact exponent.
Consequently, there exists sp ∈ ω such that psp Tp (X) is precompact. We can assume that the number sp is the
smallest possible. If sp = 0 then the subgroup Tp (X) is precompact. If sp > 0, then psp Tp (X) is precompact
but psp −1 Tp (X) is not.
P
Claim 5.8. The subgroup S = p∈P psp Tp (X) is precompact.
Proof. Given a closed neighborhood U ⊂ X of the unit, choose a neighborhood W ⊂ X of the unit such that
W 2 ⊂ U . By the hypocompactness of the exponent of X, there exist m ∈ N and a finite set F1 ⊂ X such that
mX ⊂ F1 W . For every p ∈
/ Pm the p-locality of the subgroup Tp (X) implies that Tp (X) = mTp (X) ⊂ mX.
P
Since mX is a subgroup of X, we also get p∈P\Pm Tp (X) ⊂ mX ⊂ F1 W.
A QUANTITATIVE GENERALIZATION OF PRODANOV-STOYANOV THEOREM
11
The precompactness of the groups psp Tp (X) for p ∈ Pm implies the precompactness of their (finite) sum
P
sp
sp
p∈Pm p Tp (X). So, we can find a finite subset F2 ⊂ X such that
p∈Pm p Tp (X) ⊂ F2 W . Then for the
finite set F = F1 F2 we have the desired inclusion
X
X
X
S=
psp Tp (X) ·
psp Tp (X) ⊂ F1 W · F2 W = F W 2 ⊂ F U,
psp Tp (X) =
P
p∈P
p∈Pn
p∈P\Pn
witnessing that the subgroup S is precompact.
Claim 5.9. The set P> = {p ∈ P : sp > 0} is infinite.
Proof. To derive
Q a contradiction, assume that the set P> is finite. In this case we shall prove that for the
number s = p∈P> psp the set sX is precompact, which will contradict the assumption of Lemma 5.7.
Indeed, for every p ∈PP the p-singularity of the group Tp (X) implies that sTp (X) = psp Tp (X). By Lemmas 5.1
and 2.2, the subgroup p∈P Tp (X) is dense in X. By the continuity of the map X → X, x 7→ xs , the subgroup
P
P
P
P
s( p∈P Tp (X)) = p∈P sTp (X) is dense in sX. Then the closed subgroup p∈P sTp (X) = p∈P psp Tp (X) = S̄
coincides with sX. By Claim 5.8, the subgroup S̄ = sX is compact, witnessing that X has precompact
exponent. But this contradicts the assumption of Lemma 5.7.
The compactness of the subgroups ωX and S̄ imply that the compactness of the subgroup K = ωX · S̄.
To shorten notations, for every prime number p ∈ P> denote the subgroup psp −1 Tp (X) by Sp (X) and observe
that it is not precompact but pSp (X) = psp Tp (X) is precompact.
Claim 5.10. There exists a strictly increasing sequence (pn )n∈ω of prime numbers and a decreasing sequence
(Wn )n∈ω of neighborhoods of the unit in X satisfying the following conditions for every n ∈ ω:
2
(1) Wn = Wn−1 and Wn+1
⊂ Wn ;
pn
(2) s > 0;
(3) Spn(X) is not KWn3 -bounded;
P
(4)
p≥pn+1 Tp (X) ⊂ ωX · Wn ⊂ KWn .
Proof. The construction of the sequences (pn )n∈ω and (Wn )n∈ω is inductive. To start the inductive construction, let p0 be the smallest prime number with sp0 > 0. Since the group K is compact and the subgroup Sp0(X)
is not precompact, it is not KW03 -bounded for some neighborhood W0 ⊂ X of the unit in X.
Assume that for some n ∈ N the prime number pn−1 and a neighborhood Wn−1 have been constructed.
By the hypocompactness of the exponent of X and Lemma 5.2, there exists a number m ∈ N such that
mX ⊂ ωX · Wn−1 . Choose any prime number pn ∈ {p ∈ P : sp > 0} such that pn > max(Pm ∪ {pn−1 }). Then
for any prime number p ≥ pn the number m is not divided by p and by the p-singularity of the group Tp (X)
its subgroup mTp (X) is dense in Tp (X). Then Tp (X) = mTp (X) ⊂ mX. Since mX is a subgroup, we also get
Q
p≥pn Tp (X) ⊂ mX ⊂ ωX · Wn−1 .
Since the group ωX is compact and the group Spn(X) is not precompact, it is not ωXWn3 -bounded for some
neighborhood Wn ⊂ X of the unit such that Wn = Wn−1 and Wn2 ⊂ Wn−1 . This completes the inductive
step.
Since the topological group X is ω-narrow, there exists a countable subset
P∞Z = {zk }k∈ω such that X = ZWk
for all k ∈ ω. For every k ∈ ω let Gk be the closure of the subgroup K · n=k Spn(X) in X.
Claim 5.11. For numbers n ∈ N and k ∈ ω the inclusion nG0 ⊂ Gk holds if and only if {pi : i < k} ⊂ Pn .
Proof. If {pi : i < k} ⊂ Pn , then
X
X
X
X
X
nSpi(X) ⊂
pi Spi(X) ·
Spi(K) ⊂ K ·
Spi(X) ⊂ Gk
Spi(X) =
n
i∈ω
P
i∈ω
i<k
i≥k
i≥k
and hence nG0 = n i∈ω Spi(X) ⊂ Ḡk = Gk .
Next, assume that for some i < k the number pi does not divide n. The pi -singularity of the pi -component
Tpi(X) implies that nSpi(X) = Spi(X) and then nG0 contains the subgroup Spi(X) which is not KWi3 -bounded
by condition (3) of Claim 5.10. On the other hand, by the condition (4) of Claim 5.10, the group Gk =
P
3
3
j≥k KSpj(X) is contained in KWk ⊂ KWi and hence is KWi -bounded. So, nG0 6⊂ Gk .
12
TARAS BANAKH
Let us show that the sequences (Wk )k∈ω and (Gk )k∈ω satisfy the conditions (1)–(4) of Lemma 3.3. The
condition (1) coincides with the condition (1) in Claim 5.10 and hence is satisfied. To check the condition
(2), fix two numbers n ∈ N, k ∈ ω, and assume that nG0 6⊂ Gk . By Claim 5.11, n is not divisible by some
prime number pi with i < k. Then Spi(X) = nSpi(X)) ⊂ nG0 . By the condition (3) of Claim 5.10, the set
Spi (X) is not KWi3 -bounded and hence nG0 is not KWi3 -bounded and nG0 is not KWk2 -bounded (as i < k
and Wi3 ⊂ Wi4 ⊂ Wk2 ).
Since K ⊂ Gω , the conditions (3), (4) of Lemma 3.3 will follow as soon as we check that for every neighborhood U ⊂ X of the unit there exists k ∈ ω such that Gk ⊂ KU . By Lemma 5.2, there exists n ∈ N such that
nX ⊂ ωX · U . Choose k ∈ N so large that for every i ≥ k the prime number pi does not divide n and hence
Spi(X) ⊂ Tpi(X) = nTpi(X) ⊂ nX. Then
X
Gk =
KSpi(X) ⊂ K · nX ⊂ K·ωX·U = KU.
i≥k
Therefore the conditions (1)–(4) of Lemma 3.3 are satisfies. Then the conditions (1),(2) of Lemma 3.1 are
satisfied, too. So, we can apply Lemmas 3.1 and 3.3 and construct a ♦-Hausdorff sequence {xk }k∈ω ⊂ G0
satisfying the condition (3),(4) of Lemma 3.1 and such that the set {xk }k∈ω is totally bounded in the topological
group X⋄ = (X, τ⋄ ) and hence has an accumulation point x∞ in the completion X̄⋄ of the topological group
X⋄ .
Claim 5.12. For every n ∈ N the power xn∞ ∈
/ X⋄ .
Proof. To derive a contradiction, assume that xn∞ ∈ X⋄ . Let d ∈ ω be the smallest number such that the prime
number pd does not divide n.
Since X = ZWd+1 , there exists a number l ∈ ω such that xn∞ ∈ zl Wd+1 . Let ε = 21l . Since the sequence
n
(xm )m∈ω accumulates at xn∞ in X⋄ , the neighborhood zl ♦ε Wd+1 ∈ τ⋄ of xn∞ contains a point xnλ with λ >
max{l, n}. Consequently, xnλ ∈ zl yWd+1 for some y ∈ ♦ε . Write y as y = xε11 · · · xεmm for some m ≥ λ and some
Pm
Pm
integer numbers ε1 , . . . , εm such that i=1 εi = 0 and i=1 |ε2ii | < ε = 21l . The last inequality implies that
εi = 0 for all i ≤ l and |εi | < 2i for all i ≤ m.
It follows that xnλ y −1 ∈ zl Wd+1 and xnλ y −1 = xδ11 · · · xδmn where δλ = n − ελ and δi = −εi for all i ∈
{1, . . . , m} \ {λ}. Also, |δλ | = |n − ελ | ≤ n + 2λ < 2λ+1 and hence |δi | < 2i+1 for all i ≤ m.
Let µ ∈ ω be the smallest number for which there exists a number i ≤ mPsuch that δi is not divisible by the
m
prime number pµ . Such number exists and is ≤ d since the number n = i=1 δi is not divisibleQby pd . The
minimality of µ guarantees that for every i ≤ m the number δi is divisible by the number q = j<µ pj and
hence xδi i ∈ qG0 ⊂ Gµ according to Claim 5.11.
Let j ≤ m be the largest number for which δj is not divisible by pµ . The maximality of j guarantees
that δi is divisible by pµ for all i ∈ (j, m]. It follows that j > l (as δi = 0 for i ≤ l). Then zl Wd+1 ∋
δ
−δ
δ
1
xnλ y −1 = xδ11 · · · xδmm ∈ xδ11 · · · xj j Gµ and hence xj j ∈ x−δ
· · · xj−1j−1 zl Wd+1 Gµ ⊂ Fj Wµ Gµ , which contradicts
1
the condition (3) of Lemma 3.1 as δj X 6⊂ Gµ (by Claim 5.11).
6. Treating topological groups which do not have hypocompact exponent
The following lemma was proved by Ravsky [16]. We shall give an alternative proof of this lemma deriving
it from Key Lemmas 3.1 and 3.3.
Lemma 6.1. If an Abelian topological group X is not of hypocompact exponent, then X contains a ♦-Hausdorff
sequence (xm )m∈ω , which has an accumulation point x∞ ∈ X̄⋄ such that xn∞ ∈
/ X⋄ for all n ∈ N.
Proof. Since X is not of hypocompact exponent, there exists a neighborhood W0 = W0−1 of the unit such that
for every n ∈ N the set nX is not W02 -bounded.
−1
Choose a sequence (Wk )∞
= Wk and Wk2 ⊂ Wk−1 for all k ∈ N.
k=1 of the unit such that Wk
Claim 6.2. There exists a countable subgroup Z = {zk }k∈ω of X such that for every n ∈ N the subgroup nZ
is not W0 -bounded.
A QUANTITATIVE GENERALIZATION OF PRODANOV-STOYANOV THEOREM
13
Proof. Using Zorn Lemma, for every n ∈ N choose a maximal subset Mn ⊂ X such that for any distinct
elements x, y ∈ Mn we have xn ∈
/ y n W02 . The maximality of Mn implies that for every x ∈ X there exists
n
n
2
y ∈ Mn such that x ∈ y W0 or y n ∈ xn W02 and hence xn ∈ y n W0−2 = y n W02 . Then xn ∈ y n W02 ⊂ nMn ·W02
and hence nX ⊂ nMn ·W02 . Since the set nX is not W02 -bounded, the set Mn is infinite. So, we can choose a
countable subgroup Z = {zk }k∈ω ⊂ X such that for every n ∈ N the intersection Z ∩ Mn is infinite.
We claim that for every n ∈ N the subgroup nZ is not W0 -bounded. Assuming the opposite, we could find
a finite subset F ⊂ X such that nZ ⊂ F W0 . Since the set Z ∩ Mn is infinite, for some x ∈ F there are two
distinct points y, z ∈ Z ∩ Mn such that y n , z n ∈ xW0 . Then z n ∈ xW0 ⊂ y n W0−1 W0 = y n W02 , which contradict
the choice of the set Mn .
Let G0 = Z̄ and Gk = {e} for all k > 0. It is clear that the sequences (Wk )k∈N and (Gk )k∈N satisfy
the conditions (1),(2) of Lemma 3.1 and (1)–(4) of Lemma 3.3. Applying these two lemmas, we obtain a ♦Hausdorff sequence (xn )n∈N satisfying the conditions (3),(4) of Lemma 3.1 such that the set {xn }n∈ω is totally
bounded in the topological group X⋄ := (X, τ⋄ ) and hence has an accumulation point x∞ in the completion
X̄⋄ of X⋄ .
Claim 6.3. For every n ∈ N the point xn∞ does not belong to X⋄ .
Proof. Assuming that xn∞ ∈ X⋄ , we would conclude that for every neighborhood V ⊂ X of the unit the
neighborhood xn∞ ♦1 V ⊂ G0 V ∈ τ⋄ intersects the set {xnk }k∈ω ⊂ G0 , which implies that xn∞ ∈ G0 = Z̄. Then
there exists a number λ ∈ ω such that xn∞ ∈ zλ W1 .
Let ε = 21λ . Since the sequence (xnm )m∈ω accumulates at xn∞ in the topological group X⋄ , the neighborhood
zl ♦ε W1 ∈ τ⋄ of xn∞ contains a point xnk with k > max{λ, n}. Consequently, xnk ∈ zλ yW1 for
Pmsome y ∈ ♦ε .
Write y as y = xε11 · · · xεmm for some m ≥ k and some integer numbers ε1 , . . . , εm such that i=1 εi = 0 and
Pm |εi |
i=1 2i < ε. The last inequality implies that εi = 0 for all i ≤ λ.
It follows that xnk y −1 ∈ zλ W1 and xnk y −1 = xδ11 · · · xδmn where δk = n − εk and δi = −εi for all i ∈
{1, . . . , m} \ {k}. We claim that |δi | ≤ 2i+1 for all i ≤ m. For i 6= k this follows from |δi | = |εi | < 2i . For i = k,
we have |δ
| = |n − εk | ≤ n + |εk | ≤ k + 2k < 2k+1 = 2i+1 .
Pkm
Since i=1 δi = n, for some i ≤ m the number δi is not equal zero. Let j ≤ m be the largest number such
δ
that δj 6= 0. It follows that j > λ (as δi = 0 for i ≤ λ). Then zλ W1 ∋ xnk y −1 = xδ11 · · · xδmm = xδ11 · · · xj j and
δ
−δ
1
hence xj j ∈ x−δ
· · · xj−1j−1 zλ W1 ⊂ Fj W1 G1 , which contradicts the condition (3) of Lemma 3.1 as maxi≤j |δj | ≤
1
maxi≤j 2i+1 ≤ 2j+1 and δj X 6⊂ G1 = {e}.
7. Joining pieces together
In this section we combine Lemmas 5.4, 5.7, 6.1 and prove the “unbounded” part of Theorem 1.1. First
observe that these three lemmas imply the following corollary.
Corollary 7.1. If an ω-narrow complete Abelian topological group X does not have compact exponent, then X
admits a weaker Hausdorff group topology τ⋄ such that the completion X̄⋄ of the topological group X⋄ = (X, τ⋄ )
contains an element y ∈ X̄⋄ such that y n ∈
/ X⋄ for all n ∈ N.
Now we can prove the promised “unbounded” part of Theorem 1.1.
Theorem 7.2. For a complete Abelian topological group X the following conditions are equivalent:
(1) X is complete and has compact exponent;
(2) for any continuous homomorphism f : X → Y to a powertopological semigroup Y and every point
y ∈ f (X) ⊂ Y there exists a number k ∈ N such that y k ∈ f (X);
(3) for any injective continuous homomorphism f : X → Y to a topological group Y and every point
y ∈ f (X) ⊂ Y there exists a number k ∈ N such that y k ∈ f (X).
Proof. The implication (1) ⇒ (2) follows from Theorem 4.4 and (2) ⇒ (3) is trivial. To prove that (3) ⇒ (1),
assume that the compact Abelian topological group X does not have compact exponent. Then for every n ∈ N
the set nX is not precompact. Using Lemma 2.1, we can find a closed separable subgroup Z ⊂ X such that for
every n ∈ N the subgroup nZ is not precompact, which means that Z does not have compact exponent. By
14
TARAS BANAKH
Corollary 7.1, Z admits a weaker Hausdorff group topology τ⋄ such that the completion Z̄⋄ of the topological
group Z⋄ = (Z, τ⋄ ) contains an element z ∈ Z̄⋄ such that z n ∈
/ Z⋄ for all n ∈ N.
Let Te be the family of all open neighborhoods of the unit in the topological group X. It is easy to se that
the family
τe = {U V : U ∈ Te , e ∈ V ∈ τ⋄ }
satisfies the Pontryagin Axioms [1, 1.3.12] and hence is a neighborhood base at the unit of some Hausdorff
group topology τ in which the subgroup Z remains closed and the subspace topology on Z inherited from
(X, τ ) coincides with the topology τ⋄ . Then the completion X̄τ of the topological group Xτ := (X, τ ) contains
the completion Z̄⋄ of the topological group Z⋄ and Z̄⋄ ∩ Xτ = Z⋄ . Consequently, the element z ∈ Z̄⋄ ⊂ X̄τ has
the required property: z n ∈
/ Xτ for all n ∈ N.
Acknowledgements
The author expresses his thanks to Alex Ravsky for fruitful discussions on the topic of this paper and
to Michael Megrelishvili for helpful remarks and comments on the text. Special thanks are due to Dikran
Dikranjan who told to the author the genuine story of the proof of Prodanov-Stoyanov Theorem from [6]
(which differs a bit from the original proof of this theorem in [14]) and suggested to dedicate this paper to the
memory of Ivan Prodanov who died of heart attack at the age of 49 — precisely the age of the author at the
moment of writing this paper.
References
[1] A. Arhangel’skii, M. Tkachenko, Topological groups and related structures, Atlantis Studies in Mathematics, 1. Atlantis Press,
Paris; World Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2008.
[2] T. Banakh, A. Ravsky, On H-closed paratopological groups, The Ukrainian Congress of Mathematics, Kyiv, 2001.
[3] D. Dikranjan, Recent advances in minimal topological groups, Topology Appl. 85:1-3 (1998), 53–91.
[4] D. Dikranjan, M. Megrelishvili, Relative minimality and co-minimality of subgroups in topological groups, Topology Appl.
157:1 (2010), 62–76.
[5] D. Dikranjan, M. Megrelishvili, Minimality conditions in topological groups, Recent Progress in General Topology III, Hart,
K.P., van Mill, Jan, Simon, P (Eds.) Springer Verlag (Atlantis Press), Berlin, (2014), 229–327.
[6] D. Dikranjan, I. Prodanov, L. Stoyanov, Topological Groups: Characters Dualities and Minimal Group Topologies, (2nd edn.),
Monographs and Textbooks in Pure and Applied Mathematics, Vol. 130, Marcel Dekker, New York (1989).
[7] D. Dikranjan, D. Shakhmatov, Selected topics from the structure theory of topological groups, in E. Pearl, ed., Open Problems
in Topology 2, Elsevier (2007), 389–406.
[8] D. Dikranjan, V.V. Uspenskij, Categorically compact topological groups, J. Pure Appl. Algebra, 126 (1998), 149–168.
[9] R. Engelking, General Topology, 2nd ed., Heldermann, Berlin, 1989.
[10] E. Følner, Generalization of a theorem of Bogoliouboff to topological abelian groups, Math. Scand. 2 (1954) 5–18.
[11] I.I. Guran, Topological groups similar to Lindelöf groups, Dokl. Akad. Nauk SSSR 256:6 (1981), 1305–1307.
[12] M. Megrelishvili, Generalized Heisenberg groups and Shtern’s question, Georgian Math. J. 11:4 (2004), 775–782.
[13] A. Paterson, Amenability, Amer. Math. Soc., Providence, RI, 1988.
[14] I. Prodanov, L.N. Stojanov, Every minimal abelian group is precompact, C. R. Acad. Bulgar. Sci., 37 (1), (1984), 23–26.
[15] D.A. Raikov, On a completion of topological groups, Izv. Akad. Nauk SSSR 10:6 (1946), 513–528 (in Russian).
[16] A. Ravsky, On H-closed paratopological groups, Visnyk Lviv Univ., Ser. Mekh.-Mat. 61, (2003), 172–179.
[17] D. Robinson, A course in the theory of groups, Springer-Verlag, New York, 1996.
[18] W. Roelcke, S. Dierolf, Uniform structures on topological groups and their quotients, McGrawHill, 1981.
T. Banakh: Ivan Franko National University of Lviv (Ukraine) and
Jan Kochanowski University in Kielce (Poland)
E-mail address: [email protected]
| 4math.GR
|
International Journal in Foundations of Computer Science & Technology (IJFCST) Vol.5, No.6, November 2015
THE NEW HYBRID COAW METHOD FOR SOLVING
MULTI-OBJECTIVE PROBLEMS
Zeinab Borhanifar and Elham Shadkam*
Department of Industrial Engineering, Faculty of Eng.; Khayyam University, Mashhad,
Iran
ABSTRACT
In this article using Cuckoo Optimization Algorithm and simple additive weighting method the hybrid
COAW algorithm is presented to solve multi-objective problems. Cuckoo algorithm is an efficient and
structured method for solving nonlinear continuous problems. The created Pareto frontiers of the COAW
proposed algorithm are exact and have good dispersion. This method has a high speed in finding the
Pareto frontiers and identifies the beginning and end points of Pareto frontiers properly. In order to
validation the proposed algorithm, several experimental problems were analyzed. The results of which
indicate the proper effectiveness of COAW algorithm for solving multi-objective problems.
KEYWORDS
Cuckoo Optimization Algorithm (COA), simple additive weighting (SAW), Pareto frontier, Multi-objective
optimization problem (MOP).
1. INTRODUCTION
There are many methods for solving nonlinear constrained programming problems such as
Newton, Genetic algorithm, the algorithm of birds and so on. In this paper using the emerging
Cuckoo Optimization Algorithm and simple additive weighting a method to solve multi-objective
problems is presented.
In single-objective optimization, it is assumed that the decision makers communicate only with
one goal like: profit maximization, cost minimization, waste minimization, share minimization
and so on. But in the real world it is not possible to consider single goals and usually more than
one goal are examined. For example, in the control of the projects if only the time factor is
considered, other objectives such as cost and quality are ignored and the results are not reliable.
So it is necessary to use multi-objective optimization problems.
Ehrgott and Gandibleux presented a detailed approximation method regarding the problems
related to combinatorial multi-objective optimization [1]. Klein and Hannan for multiple
objective integer linear programming problems (MOILP) presented and algorithm in which some
additional restrictions is used to remove the known dominant solutions [2]. Sylva and Crema
offered a method to find the set of dominant vectors in multiple objective integer linear
programming problems [3]. Arakawa et al. used combined general data envelopment analysis and
Genetic Algorithm to produce efficient frontier in multi-objective optimization problems [4].
Deb analyzed the solution of multi-objective problems by evolutionary algorithms [5]. Reyesseerra and Coello Coello analyzed the solution of multi-objective problems by particle swarm [6].
Cooper et al. have worked on the solution of multi-objective problems by the DEA and presenting
DOI:10.5121/ijfcst.2015.5602
15
International Journal in Foundations of Computer Science & Technology (IJFCST) Vol.5, No.6, November 2015
an application [7]. Pham and Ghanbarzadeh solved multi-objective problems by bee algorithm
[8]. Nebro et al. analyzed a new method based on particle swarm algorithm for solving multiobjective optimization problems [9]. Gorjestani et al. proposed a COA multi objective algorithm
using DEA method [10].
For multi-objective optimization problems usually it is not possible to obtain the optimal solution
that simultaneously optimizes all the targets in question. Therefore we should try to find good
solutions rather than the optimal ones known as Pareto frontier. Given that so far the Simple
Additive Weighting method is not used in meta-heuristic, especially cuckoo algorithms, this
paper presents a combined method.
The first section introduces Cuckoo optimization algorithm, then in the second section Simple
Additive Weighting (SAW) method is discussed as a combined method for solving multiobjective described. Finally, the fourth section provides the proposed implemented approach,
numerical results and a comparison which is made with other methods.
2. CUCKOO OPTIMIZATION ALGORITHM
Cuckoo optimization algorithm was developed by Xin-She Yang and Suash Deb in 2009. Thence
Cuckoo optimization algorithm was presented by Ramin Rajabioun in 2011 [11]. Cuckoo
algorithm flowchart is as figure 1. This algorithm applied in several researches such as production
planning problem [12], portfolio selection problem [13], evaluation of organization efficiency
[14], evaluation of COA [15] and so on. For more information about the algorithm refer to [11].
3. SIMPLE ADDITIVE WEIGHTING METHOD
SAW is one of the most practical methods designed for decision-making with multiple criteria
presented by Hong and Eun in 1981. In this method which is also known as weighted linear
combination after scaling the decision matrix by weighted coefficients of criteria, the free scale
weighted decision matrix id obtained and according to this scale the score of each option is
selected. The most important feature of this method is the simple application because of its
mathematical logic.
Assuming the multiple target model (1) and defining the parameters w1 and w2 which are the
weight of the objective functions and defined based on the importance of the functions by the
decision maker, the model can be converted to single-objective models (2):
Max F(x)=(
,
,….,
s.t.
≤
≥0
Max F(x)=
+
+⋯+
+
+ ⋯+
=1
In these models
x …
of the decision maker.
x are objective functions.
(1)
(2)
is the weight defined by the importance
16
International Journal in Foundations of Computer Science & Technology (IJFCST) Vol.5, No.6, November 2015
Figure 1: The Cuckoo optimization algorithm flowchart
4. PRESENTATION OF HYBRID COAW ALGORITHM
In this section we present the method COAW which is proposed in this paper. The steps of this
algorithm are as follows. Also the flowchart of COAW algorithm is as figure 2.
Step1 Different random w1 and w2 are generated subject to the summation of these two values
equals to one.
Step 2 The present locations of Cuckoos are determined randomly
Step 3 A number of eggs
gs are allocated to each Cuckoo
Step 4 The laying radius of each
ch Cucko
Cuckoo is determined
Step 5 The Cuckoos hatch in the nests of the hosts that are within their laying radius
Step 6 Eggs that are detected by the host birds are destroyed
Step 7 The eggs of the identified cuckoos are nurtured
Step 8 The habitats of the new cuckoos are evaluated by SAW method and determined weights
Step 9 the maximum number of cuckoos living at each location are determined and the ones
on in
wrong areas are destroyed
Step 10 The cuckoos are clustered by K-means
K
and the best cluster of cuckoos is determined as
the residence
Step 11 The new population of cuckoos m
moves toward the target location
Step 12 Stop condition is established
ablished otherwise go to step 2Step
2Step 13 the value of , are
determined for the best solutions and the Pareto frontier is gained based on ,
17
International Journal in Foundations of Computer Science & Technology (IJFCST) Vol.5, No.6, November 2015
SAW module
Determination of
weights problem
Evaluation of cost
function based on
determined weights
Figure 2. The flowchart of COAW algorithm
5. IMPLEMENTATION OF COAW ALGORITHM ON SOME TEST PROBLEMS
PROBLEM
In this section in order to validation
validat
the COAW algorithm some test problems are analyzed. Test
problems are presented
esented in Table 1.
Table 1. Test problems
Number of problem
Objectives
1
=
=
,
2 +
≥0
2
=2
=
,
1 +
≥0
3
Constraints
=
=
≥
3
1,
2
4≤0
≤0
≤0
≤2
18
International Journal in Foundations of Computer Science & Technology (IJFCST) Vol.5, No.6, November 2015
Given that determining input parameters is one of the effective problems in meta-heuristic
algorithms, so the parameters of the algorithm are presented as follows: the number of initial
population=5, minimum number of eggs for each cuckoo= 2, maximum number of eggs for each
cuckoo =4, maximum iterations of the Cuckoo Algorithm=50, number of clusters that we want to
make=1, Lambda variable in COA paper=5, accuracy in answer is needed=-3.75, maximum
number of cuckoos that can live at the same time=10, Control parameter of egg laying=5,
cuckooPopVariance = 1e-13.
6. THE SOLUTION OF TEST PROBLEMS
In this section the experimental problems of the previous section are solved by the proposed
algorithm and the results are compared and examined with the same algorithm.
6.1. The First Problem
2
1.8
1.6
1.4
f2
1.2
1
0.8
0.6
0.4
0.2
0
0
0.5
1
1.5
2
2.5
f1
Figure 3. Pareto frontiers created by COAW algorithm for first problem
(a)
(b)
(C)
Figure 4. Pareto frontiers created by: (a) Ranking method (b) DEA method (c) GDEA Method for first
problem
19
International Journal in Foundations of Computer Science & Technology (IJFCST) Vol.5, No.6, November 2015
6.2. The Second Problem
0
-0.2
f2
-0.4
-0.6
-0.8
-1
-1.2
-1
-0.5
0
0.5
1
1.5
2
2.5
f1
Figure 5. Pareto frontiers created by COAW for the first problem
(c)
(b)
(a)
Figure 6. Pareto frontiers created by: (a) Ranking method (b) DEA method (c) GDEA Method for second
problem
6.3. The Third Problem
2
1.5
1
f2
0.5
0
-0.5
-1
-1.5
-2
-1
-0.8
-0.6
-0.4
-0.2
0
f1
0.2
0.4
0.6
0.8
1
Figure 7. Pareto frontiers established by COAW for third problem
20
International Journal in Foundations of Computer Science & Technology (IJFCST) Vol.5, No.6, November 2015
(a)
(b)
(C)
Figure 8. Pareto frontiers created by: (a) Ranking method (b) DEA method (c) GDEA Method for third
problem
After the implementation of the proposed approach on test problems the Pareto frontiers are
obtained according to figures 3, 5 and 7 in order to compare the COAW method with other
methods, ranking method, DEA method and GDEA method are implemented on problems. The
results are show as figures 4, 5 and 8.
As figures indicate the created Pareto frontiers of the COAW proposed algorithm are exact and
have good dispersion. This method has a high speed in finding the Pareto frontiers and identifies
the beginning and end points of Pareto frontiers properly. The COAW algorithm not only solves
the problems with lower initial population 5 but also it presents better and more exact answers in
fewer repetitions than similar methods.
7. CONCLUSION
In this paper the hybrid COAW algorithm was presented to solve multi-objective problems. The
hybrid approach includes Cuckoo Algorithm and Simple Additive Weighting method. The
algorithm was analyzed for a number of experimental problems and compared with several
similar methods. The results indicate the accuracy in finding Pareto frontiers. Also the Pareto
frontier is better than similar methods and as a result COAW proposed method is reliable, fast and
simple to solve multi-objective optimization problems.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
Ehrgott, M., Gandibleux, X., Bound Sets for Bi objective Combinatorial Optimization Problems,
Computers & Operations Research, Vol. 34, Issue 9, 2674-2694, 2007.
Klein, D., Hannan, E., An Algorithm for the Multiple Objective Integer Linear Programming
Problem, European Journal of Operational Research, 1982, 9, 378–385.
Sylva, J., Crema, A., A Method for Finding the Set of Non-Dominated Vectors for Multiple Objective
Integer Linear Programs, European Journal of Operational Research, 158, pp. 46–55, 2004.
Arakawa, M., Nakayama, H., Hagiwara, I., Yamakawa, H., Multiobjective Optimization using
adaptive range genetic algorithms with data envelopment analysis, Vol.3, 1998.
Deb, K., Multi-Objective Optimization using Evolutionary Algorithms, John & Wiley Sons, Ltd.,
2001.
Reyes-Sierra, M., Coello Coello, CA., Multiple objective particle swarm optimizers: A survey of the
state-of-art. International Journal of Computational Intelligence Research 2(3), 287–308, 2006.
Cooper, W.W., Seiford, L.M., Tone, K., Data Envelopment Analysis: A Comprehensive Text with
Models, Applications, References and DEA Solver Software. Springer, New York, 2007.
21
International Journal in Foundations of Computer Science & Technology (IJFCST) Vol.5, No.6, November 2015
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
Pham, DT., Ghanbarzadeh, A., Multi-objective optimization using the bees algorithm. In: Third
international virtual conference on intelligent production machines and systems (IPROMS 2007):
Whittles, Dunbeath, Scotland, 2007.
Nebro, A.J., Durillo, J.J., Garc´ıa-Nieto, J., Coello Coello, CA., Luna F and Alba E (2009) SMPSO:
A new PSO-based metaheuristic for multi-objective optimization. 2009 IEEE Symposium on
Computational Intelligence in Multi criteria Decision-Making (MCDM 2009). IEEE Press, New
York, pp. 66–73.
Gorjestani, M., Shadkam, E., Parvizi, M., Aminzadegan, S., A HYBRID COA-DEA METHOD FOR
SOLVING MULTI-OBJECTIVE PROBLEMS, International Journal on Computational Science &
Applications, Vol.5, No.4, 2015.
Rajabioun, R., Cuckoo Optimization Algorithm, Applied Soft Computing, Vol. 1, pp 5508–5518,
2011.
Akbarzadeh, A., Shadkam, E., THE STUDY OF CUCKOO OPTIMIZATION ALGORITHM FOR
PRODUCTION PLANNING PROBLEM, International Journal of Computer-Aided technologies,
Vol.2, No.3, 2015.
Shadkam, E., Delavari, R., Memariani, F., Poursaleh, M., PORTFOLIO SELECTION BY THE
MEANS OF CUCKOO OPTIMIZATION ALGORITHM, International Journal on Computational
Sciences & Applications, Vol.5, No.3, 2015.
Shadkam E., Bijari M., The Optimization of Bank Branches Efficiency by Means of Response
Surface Method and Data Envelopment Analysis: A Case of Iran, Journal of Asian Finance,
Economics and Business Vol. 2 No. 2, 13-18, 2015.
Shadkam E., Bijari M., EVALUATION THE EFFICIENCY OF CUCKOO OPTIMIZATION
ALGORITHM, International Journal on Computational Sciences & Applications. Vol.4, No.2, pp. 3947, 2014.
22
| 9cs.NE
|
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
1
Contrast-Oriented Deep Neural Networks for
Salient Object Detection
arXiv:1803.11395v1 [cs.CV] 30 Mar 2018
Guanbin Li and Yizhou Yu
Abstract—Deep convolutional neural networks have become a
key element in the recent breakthrough of salient object detection.
However, existing CNN-based methods are based on either patchwise (region-wise) training and inference or fully convolutional
networks. Methods in the former category are generally timeconsuming due to severe storage and computational redundancies
among overlapping patches. To overcome this deficiency, methods
in the second category attempt to directly map a raw input
image to a predicted dense saliency map in a single network
forward pass. Though being very efficient, it is arduous for these
methods to detect salient objects of different scales or salient
regions with weak semantic information. In this paper, we develop
hybrid contrast-oriented deep neural networks to overcome
the aforementioned limitations. Each of our deep networks is
composed of two complementary components, including a fully
convolutional stream for dense prediction and a segment-level
spatial pooling stream for sparse saliency inference. We further
propose an attentional module that learns weight maps for fusing
the two saliency predictions from these two streams. A tailored
alternate scheme is designed to train these deep networks by finetuning pre-trained baseline models. Finally, a customized fully
connected CRF model incorporating a salient contour feature
embedding can be optionally applied as a post-processing step
to improve spatial coherence and contour positioning in the
fused result from these two streams. Extensive experiments on
six benchmark datasets demonstrate that our proposed model
can significantly outperform the state of the art in terms of all
popular evaluation metrics.
Index Terms—Deep Contrast Network, Salient Object Detection, Conditional Random Fields.
I. I NTRODUCTION
Visual saliency detection aims to locate the most conspicuous regions in images according to the human visual system
and has recently received increasing research interest. Image
saliency detection is traditionally approached in the form of
either eye-fixation prediction or salient object detection. The
former focuses on the natural mechanism of visual attention
and aims at accurately predicting human eye attended image
locations. However, previous research has pointed out that
salient object detection, which is more concerned with the
integrity of the predicted object regions, is more conducive to
a series of computer vision tasks including semantic segmentation [2], object localization and detection [3], [4], contentaware image editing [5], visual tracking [6] and person reidentification [7]. Although numerous valuable models have
This work was supported in part by the National Natural Science Foundation
of China under Grant 61702565 and was also sponsored by CCF-Tencent
Open Research Fund.
G. Li is with Sun Yat-sen University, Guangzhou 510006, China (e-mail:
[email protected]).
Y. Yu is with the Department of Computer Science, The University of Hong
Kong (e-mail: [email protected]).
A preliminary version of this paper appeared in CVPR 2016 [1].
been proposed, salient object detection remains challenging
due to a variety of complex factors in real-world scenarios.
Perceptual studies [8], [9] have shown that visual contrast
is the key factor that affects visual saliency. A series of
conventional salient object detection algorithms based on local
or global contrast modeling [10], [11], [12] have been successfully proposed. In previous research efforts, visual contrast
modeling is generally focused on the differences among various handcrafted low-level features and coupled with heuristic
saliency priors. Although handcrafted features tend to perform
well in simple cases, they are not robust enough for more
challenging scenarios. For example, it is hard for local contrast
models to accurately segment out large homogeneous regions
inside salient objects while global contrast information may
fail to handle images with cluttered background. Although
there exist machine learning based algorithms for salient object
detection [13], [14], [15], [16], they are basically focused
on integrating various handcrafted features [14] or merging
multiple saliency maps computed by different methods [16].
Recently, deep convolutional neural networks have been
widely used in salient object detection [17], [18], [19] because
of their powerful feature representations and have achieved
substantially better performance than traditional methods.
Methods based on deep convolutional neural networks can be
roughly divided into two categories. Methods in the first category generally perform patch-wise (or region-wise) training
and inference. Specifically, an image is first divided into a
set of regions or patches, and deep CNN based regression
or classification models are then trained to independently map
each image patch or region to a saliency score or a binary class
label (salient or non-salient). However, this results in serious
storage and computational redundancies, making training and
testing very time-consuming. For example, training a patchoriented CNN model takes over two GPU days while requiring hundreds of megabytes of storage to save deep features
extracted from one single image. Inspired by the latest trends
of developing fully convolutional neural networks for pixellevel image understanding problems [20], [21], [22], methods
in the second category train end-to-end models that directly
map an input image of arbitrary size to a saliency map with
the same size, performing dense feedforward computation and
backpropagation over the entire image. This type of methods
have rapidly become the cornerstone of this field as they not
only achieve very favorable performance but also are very
efficient. However, it is still arduous for these methods to
detect salient objects of different scales or salient regions with
weak semantic information. Moreover, pixel-level correlation
is typically not considered in such fully convolutional networks (FCNs), which usually give rise to incomplete salient
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
regions with blurry contours.
In this work, we develop hybrid contrast-oriented deep
neural networks to overcome the aforementioned limitations of
two types of contemporary CNN-based salient object detection
methods. Our deep networks are composed of a fully convolutional stream for dense prediction and a segment-level spatial
pooling stream for sparse saliency inference. We devise a
multi-scale fully convolutional network (MS-FCN) in the first
stream, which receives an entire image as input and directly
learns to map it to a dense saliency prediction with pixellevel accuracy. Our MS-FCN can not only learn multi-scale
feature representations, but also accurately judge the saliency
of every pixel by mining visual contrast information hidden in
multi-scale receptive fields. The segment-level spatial pooling
stream computes another sparse saliency map over superpixels by modeling the contrast between every superpixel and
its spatially adjacent regions. It extracts multi-scale regional
features very efficiently by performing feature masking in the
feature map of an intermediate layer of MS-FCN. At the end,
we produce our final saliency map by merging the saliency
maps from both streams with weight maps generated from a
proposed attentional module in our deep network. Our MSFCN can also be re-trained to generate a contour map for
salient objects. This contour map can be used to improve
contour localization in the fused saliency map via a fully
connected CRF.
In summary, this paper has the following contributions:
•
•
•
We propose end-to-end contrast-oriented deep neural
networks for localizing salient objects using multi-scale
contextual information. They incorporate a fully convolutional stream for dense prediction and a segment-wise
spatial pooling stream for sparse inference. A tailored
alternate scheme is designed to train these deep networks
by fine-tuning pre-trained baseline models.
A multi-scale VGG-16 or ResNet-101 network pretrained for image classification is re-purposed as the fully
convolutional stream to infer a dense saliency prediction
directly from the raw input image in a single forward
pass. This fully convolutional network can also be retrained to infer a salient object contour map, which can
be represented as a feature embedding and incorporated in
a fully connected CRF model to further improve contour
localization in the final result.
We have also devised a segment-wise spatial pooling
stream complementary to the fully convolutional stream
in our deep network. This stream efficiently masks out
segment-wise features from one designated feature map
of MS-FCN, and accurately models visual contrast among
superpixels and well captures saliency discontinuities
along region boundaries.
The rest of this paper is organized as follows. Section II
reviews related work on salient object detection. In Section III,
we introduce our proposed contrast-oriented deep neural networks. The complete algorithm is presented in Section IV.
Section V provides extensive performance evaluation as well
as comparisons against state-of-the-art models. Finally, we
conclude this paper in Section VI.
2
II. R ELATED W ORK
Traditional salient object detection can be categorized
into bottom-up approaches with handcrafted low-level features [23], [24], [15], [25], [26], [11], [14], [27], [10], [28] and
top-down approaches incorporating high-level knowledge [29],
[30], [31], [32], [33], [34], [35]. Bottom-up methods are
usually based on the center bias or background priors and
infer saliency maps from global or local contrast represented
as a combination of handcrafted low-level features (e.g. color,
texture and image gradient). Bottom-up computational models
are primarily based on a center-surround scheme and compute
saliency maps using a linear or non-linear combination of lowlevel features such as color, intensity, texture and orientation
of edges [24], [10], [36], [15]. Top-down methods are in
general task-dependent and require a machine learning scheme
to incorporate high-level knowledge into a process which was
originally limited to specified objects or assumptions [34],
[35], [33]. Graph based methods have also been widely used
to enhance spatial consistency and refine detected saliency
maps [37], [11], [1]. Recently, deep learning based methods
have been widely used for salient object detection and have
promoted its research into a new phase. Since the focus of
this paper is deep learning based salient object detection, we
highlight the most relevant previous work in the following
discussion.
In recent years, the successful application of deep convolutional neural networks has triggered a revolution in machine
learning and artificial intelligence, and has yielded significant
improvement in a variety of visual comprehension tasks,
including image classification [38], object detection [39] and
semantic segmentation [20], closing the gap to human-level
performance. Motivated by this, several attempts have also
been made to apply deep neural network models to salient
object detection [40], [1], [41], [42], [43]. Han et al. [44] first
attempted to develop stacked denoising autoencoders to learn
powerful representations for salient object detection in an unsupervised and bottom-up manner. In [45], a weighted sparse
coding framework is proposed for image saliency detection.
Recently, with the widespread application of convolutional
neural networks in image analysis and comprehension tasks,
it is not surprising to see a surging number of research papers
where very good results have been achieved on salient object
detection via the application of CNNs. Li et al. [17], [40]
trained a multi-layer fully connected network for deriving the
saliency value of every superpixel from its contextual CNN
features. Wang et al. [19] proposed two deep neural networks,
which take into account both low-level features and high-level
objectness, for salient object detection at the patch level. A
multi-context deep CNN framework incorporating both global
and local contexts is presented in [18]. However, all these
methods include fully connected layers and infer saliency
maps in an isolated patch-wise manner, the crucial spatial
information in the input image is ignored. However, since all
the image patches are treated as independent samples during
network training and inference, there is no shared computation
among overlapping image segments, which results in significant redundancies and excessive computational cost during
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
training and testing.
To address these issues, inspired by the seminal work
of developing end-to-end deep networks for semantic image
segmentation [20], [21], variant of fully convolutional neural
networks have been introduced to solve the problem of salient
object detection since the publication of our earlier conference
version [1]. Li et al. [41] proposed to explore the correlations
between saliency detection and semantic image segmentation
using a multi-task fully convolutional neural network. Liu et
al. [46] propose a hierarchical recurrent CNN to progressively
refine the details of saliency maps from a coarse prediction
result generated from the forward pass of a fully convolutional
VGG-16 network. Kuen et al. [47] proposed a recurrent attentional convolution-deconvolution network (RACDNN), which
consists of a recurrent neural network and a spatial transform
module, to recurrently attend to selected image sub-regions
for saliency refinement. In [48], Wang et al. introduced a
recurrent fully convolutional network (RFCN) to iteratively
refine the saliency map with incorporated prior knowledge.
These FCN based models have greatly improved both accuracy
and efficiency in saliency detection, there are still three aspects
of the flaws. First of all, these models are mostly based on the
topmost feature map of the network for saliency inference, the
over-reliance on the regional semantic feature may result in the
pool detection performance on the salient region with weak
semantic information. Second, all of these methods consider
feature modeling at a single scale and may not accurately
detect salient objects of very different sizes. And finally, as
the value at each position of a saliency map generated from
FCN-based models is derived from a context with a fixed
size (receptive field), the contours of salient objects can hardly
be well detected, and the generated saliency maps usually have
inadequate spatial consistency. Our proposed method instead
delves into the nature of saliency prediction, capturing the
key aspect in this problem, which is contrast learning. The
proposed method is not only able to infer a saliency probability
map from the contrast information in a multiscale deep CNN
but also from edge-preserving region-wise contrast information. In addition, it has been proven that fully connected
CRFs can be formulated as recurrent neural networks (RNNs).
However, experimental results show that RNNs can hardly be
trained to achieve comparable results as CRFs. Our proposed
method therefore exploits the effectiveness of a contour-aware
CRF. Our experimental results demonstrate the superiority of
our proposed method in comparison to all existing FCN based
salient object detection techniques.
Note that the initial deep contrast network reported in
CVPR 2016 [1] can be viewed as the first piece of work that
aims at designing an end-to-end fully convolutional network
for visual contrast modeling. To a certain extent, it inspired
the subsequent development of FCN-based models in this
field. Our updated contrast-oriented deep neural network for
salient object detection has several improvements over its
initial version. First, we adapt the state-of-the-art ResNet-101
network [49] for image classification to a fully convolutional
network and use it to replace the VGG-16 network in the original fully convolutional stream, achieving better performance.
Second, the fully convolutional stream is run on multiple
3
scaled versions of the original input image while the segmentwise spatial pooling stream is trained using segments from
multi-level image segmentation. These strategies make our
deep model more accurately detect salient objects at different
scales. Third, we propose to add an attentional module which
learns pixel-wise soft weights for fusing the two saliency
maps respectively generated from the two streams. Fourth,
we discover that the proposed multi-scale fully convolutional
stream in our deep network can be re-trained to detect salient
region contours, which can be integrated into a fully connected
CRF model to further improve contour localization in the
final saliency map. Finally, we present a more comprehensive
experimental comparison among multiple model variants and
report improved results on all benchmarks using all evaluation
metrics.
III. D EEP C ONTRAST N ETWORK
As illustrated in Fig. 1, our proposed contrast-oriented
deep neural network is composed of two complementary
components, a fully convolutional stream for dense saliency
prediction and a segment-wise spatial pooling stream for
sparse saliency inference. Specifically, the first component is
a multi-scale fully convolutional network (MS-FCN), which
receives an entire image as input and is trained to map the
input to a dense saliency map S1 in an end-to-end mode
by exploiting visual contrast across multiple levels of feature
maps. The segment-wise spatial pooling stream is trained to
infer the saliency map S2 at the segment level by discovering
the contrast among spatially adjacent regions on the basis of
features masked out from one designated feature map of the
first stream and a multi-layer perceptron. At the end, these two
intermediate saliency predictions from the above two network
streams are merged according to weight maps prescribed by a
trained attention module. The merged map becomes our final
saliency map S.
A. Multi-Scale Fully Convolutional Network
Inspired by the groundbreaking application of fully convolutional networks in pixel-level image comprehension, we focus
on constructing an end-to-end pixelwise regression network,
which can directly map a raw input image to a dense saliency
map. Considering the centrality of contrast modeling for
saliency detection, we have the following considerations when
designing the structure of this end-to-end network. First, the
network should be deep enough to accommodate features from
multiple levels since visual saliency relies on modeling the
contrast among both low-level appearance features as well as
high-level semantic features. Second, the network needs to be
able to explore the visual contrast across multiple feature maps
and detect salient objects of various scales. Finally, due to the
lack of training images with pixel-wise labeling, it is much
desired to fine-tune an existing pre-trained network instead of
training from scratch.
As VGG [50] and ResNet [49] are the two most representative and widely used deep classification networks with
publicly available pre-trained models, we choose them as
our pre-trained networks and adapt for our requirements.
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
4
...
...
... ...
...
...
...
...
...
...
SP
SFM
Output
NN_Layer2
NN_Layer1
Fea_s3 Fea_s2 Fea_s1
Fig. 1: The overall architecture of our proposed contrast-oriented deep neural network. It consists of a fully convolutional
stream (upper part), a segment-wise spatial pooling stream (lower part) and an attentional module to fuse the intermediate
saliency maps from the two streams. “SFM” refers to the segment feature masking layer while “SP” refers to the spatial pooling
operation.
Here we describe in detail the transformation of the VGG16 network, and ResNet-101 can be similarly transformed to
satisfy the requirements. To re-purpose the VGG-16 network
for dense saliency map generation, we first convert the two
fully connected layers of VGG-16 into 1 × 1 convolutional
ones as described in [20]. Moreover, as the original VGG16 network consists of 5 max pooling layers and each with
stride 2, the resulting network can only yield low-resolution
prediction maps with 1/32 the input resolution. To make the
resulting saliency map have a higher resolution, we remove
the downsampling operation in the last two max-pooling
layers by simply setting their “stride” to 1, which results in
downsampling by a factor of 8 instead of 32. At the same
time, to maintain the same size of the receptive fields of
the convolutional layers that follow, we refer to [21], [51]
and apply the dilation operation to the corresponding filter
kernels. The dilation algorithm (also called à trous algorithm),
which was originally proposed to improve the computational
efficiency of undecimated wavelet transforms [52], has recently been incorporated into the Caffe framework [21], [51]
as “dilated convolution” to efficiently control the resolution
of feature maps within deep CNNs without the need to learn
extra parameters. It works by inserting zeros between filter
weights. Specifically, consider applying the dilated version of a
convolutional filter w to an input feature map x, and generating
an output feature map y. The output value at position i is
calculated as
y[i] =
X
k
x[i + r · k]w[k],
(1)
where the dilation rate r corresponds to the stride with which
we sample the input feature map. This is equivalent to applying
convolution to the input feature map x with filters up-sampled
by inserting r − 1 zeros between any two originally adjacent
filter elements along each dimension. This dilated convolution
allows us to explicitly control the density of feature responses
in our customized fully convolutional networks. In our implementation, after setting the stride of the last two pooling
layers to 1, we replace all subsequent convolutional layers
with dilated convolutional layers with dilation rate r = 2 or
r = 4 (r = 2 for the three consecutive convolutional layers
after the penultimate max-pooling layer and r = 4 for the last
two newly converted 1 × 1 convolutional layers).
VGG-16 has five max pooling layers performing downsampling operations. If we start from the pooling layer closest
to the input image, these pooling layers have an increasingly
larger receptive field containing contextual information. To
design a deep convolutional network that is capable of mining
visual contrast information crucial in saliency inference, we
further develop a multiscale network from the above fully
convolutional version of VGG-16. As shown in the left part of
Fig. 2, we connect three extra convolution layers to each of the
first four max-pooling layers. The first extra layer uses 3 × 3
convolution kernels and has 128 channels while the second one
uses 1×1 convolution kernels and also has 128 channels. And
the third extra layer has one 1 × 1 kernel and a single channel,
which is used to produce the output saliency map. To make
the output feature maps of the four sets of extra convolutional
layers have the same size (8× downsampling resolution), the
stride of the first layer in these four sets are set to 4, 2, 1,
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
and 1, respectively. Although the four resulted feature maps
are of the same size, they are computed using receptive fields
with different sizes and hence represent contextual features at
4 different scales. We further stack these four feature maps
with the last output feature map of the above customized
fully convolutional conversion. The stacked feature maps (5
channels) are fed into a final convolution layer with a 1 × 1
kernel and a single output channel, which is modulated by the
sigmoid activation function to produce the saliency probability
map. Though the resulting saliency map of this network stream
has a downsampling factor of 8 in comparison to the input
image, it is smooth enough and allows us to use simple bilinear
interpolation to restore the resolution of the original input at
a negligible computational cost. We call this resized saliency
map S1 .
Note that the ResNet-101 network has no hidden fully
connected layers. To adapt ResNet-101 for dense saliency
prediction, we simply replace its 1000-way linear classification
layer with a linear convolutional layer with a 1 × 1 kernel and
a single output channel. Similar to VGG-16, the resolution
of the feature maps before the linear convolutional layer
is only 1/32 of that of the original input image because
the original ResNet-101 consists of one pooling layer and 4
convolutional layers, each of which has stride 2. We call these
five layers “down-sampling layers”. As described in [49], the
101 layers in ResNet-101 can be divided into five groups.
Feature maps computed by different layers in each group
share the same resolution. To increase the resolution of the
final saliency map, we replace the last two down-sampling
layers with dilated convolution layers, skip subsampling by
setting their stride to 1, and correspondingly increase the
dilation rate of subsequent convolution kernels to enlarge their
receptive fields. Therefore, all the features maps in the last
three groups have the same resolution, 1/8 original resolution,
after network transformation. To develop a multiscale version
of the above end-to-end extension of ResNet-101, as shown
in the right of Fig. 2, we connect an extra sub-network with
three convolutional layers to each of the final layers in the first
four groups. These additional layers have the same structure
as those added to VGG-16. Similar to the multiscale extension
of VGG-16, the four output feature maps from these four subnetworks are stacked together with the final output feature
map of the transformed ResNet-101, and fed into a final
convolutional layer with a 1 × 1 kernel and a single output
channel for final saliency map inference.
B. Segment-Level Saliency Inference
Salient objects in images are usually presented in a variety
of irregular shapes and the corresponding saliency map often
exhibits discontinuities along the object boundaries. Our multiscale fully convolutional network operates at a subsampled
pixel level and equally treats each pixel in the input image
without explicitly taking into account such saliency discontinuities. To better model visual contrast between regions and
visual saliency along region boundaries, we design a segmentwise spatial pooling stream in our network.
We first divide an input image into a set of superpixels,
and call each superpixel a segment. A mask is computed for
5
every segment in the feature map generated from one selected
convolutional layer of MS-FCN, which is named the feature
masking layer. We choose the convolutional layer Conv5 3 as
the feature masking layer in the MS-FCN based on VGG-16,
and the last convolutional layer in the fourth layer group as the
feature masking layer in the MS-FCN based on ResNet-101
as suggested in [49]. Since the activations at each location in
the feature masking layer is controlled by a receptive field in
the input image, we first project every location in the feature
masking layer to the center of its receptive field as in [53]. For
each segment in the input image, we first generate a binary
mask within its bounding box. In this mask, pixels inside the
segment are labeled ‘1’ while others are labeled ‘0’. Each
pixel labeled as ‘1’ in the binary mask is first assigned to
the closest receptive field center and then backprojected onto
the feature masking layer. Thus, each location in the feature
masking layer collects multiple ‘1’ labels backprojected from
its receptive field. The ratio between the number of collected
‘1’ labels at the location and the number of pixels in the input
image closest to its receptive field center is recorded. To yield
a binary mask for the segment on the feature masking layer,
the previously computed ratio at every location is thresholded
at 0.5, and the set of locations with nonzero values after
thresholding form the segment mask. In the event that the
ratio at all locations is below 0.5, the set of locations with
nonzero ratios before thresholding form the segment mask.
The resulting segment mask is then applied to the output
feature map of the feature masking layer by simply multiplying
this binary mask with each channel of the feature map. We call
the resulting features segment-masking features in our method.
Note that the feature map generated from the feature masking
layer has a downsampling factor of 8 instead of 32 in the
original VGG-16 network or 16 in the original ResNet-101
network since subsampling has been skipped in the last two
downsampling layers as described in Section III-A. Therefore,
the resolution of the feature map generated from the feature
masking layer is sufficient for segment masking.
Since segments have irregular shapes and variable sizes
when projected onto the feature masking layer, we further
perform a spatial pooling (SP) operation to produce a feature
vector of fixed length for each segment. It is a simplified
version of spatial pyramid pooling described in [54]. Specifically, we divide the bounding box of a projected segment
into h × w cells and perform Max- or mean-pooling over valid
positions (with mask label ‘1’) in each grid cell. This results
in h × w feature vectors of size C, which is the number of
convolutional filters in the feature masking layer. Afterwards,
we concatenate the feature vectors extracted from all grid cells
of the same segment to obtain the final feature vector with
h × w × C dimensions for that segment.
To discover segment-level visual contrast, we represent each
segment with a concatenation of three feature vectors respectively for three nested and increasingly larger regions masked
out from the designated feature map. These three regions
include the bounding box of the considered segment, the
bounding box of the immediate neighboring segments as well
as the entire feature map from the feature masking layer (with
the considered segment excluded to indicate the position of
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
6
CONV(3×3)
Stride:4
·
·
CONV1_1+RELU
CONV1_2+RELU
·
·
CONV2_1+RELU
CONV2_2+RELU
CONV1 (7×7, 64)
CONV(3×3)
Stride:4
POOLING_2
·
·
·
CONV3_1+RELU
CONV3_2+RELU
CONV3_3+RELU
·
·
·
CONV4_1+RELU
CONV4_2+RELU
CONV4_3+RELU
POOLING_3
POOLING_4
·
·
·
CONV5_1+RELU
CONV5_2+RELU
CONV5_3+RELU
128
POOLING (3×3, max)
POOLING_1
128
128
1
128
CONV(3×3)
Stride:1
128
1
CONV3_x
CONV(3×3)
Stride:1
CONV4_x
CONV(3×3)
Stride:1
128
CONV(3×3)
Stride:1
128
128
1
128
128
1
1
5
5
128
POOLING_5
CONV+RELU+DropOut
1
CONV2_x
CONV(3×3)
Stride:2
128
128
CONV(3×3)
Stride:1
128
1
CONV5_x
128
128
1
CONV+RELU+DropOut
VGG16
ResNet-101
Fig. 2: The architecture of VGG-16 based multi-scale fully convolutional network (left) and ResNet-101 based multi-scale
fully convolutional network (right). We connect three extra convolutional layers to each of the first four max-pooling layers of
VGG-16 and convert it to a multiscale version. For ResNet-101, we divide the 101 layers into five groups and connect an extra
sub-network with three convolutional layers to each of the final layers in the first four groups to form the multiscale version.
the segment). The above-mentioned feature representation of
each segment is further fed into two fully connected layers.
The output of the second fully connected layer is fed into
a “Sigmoid” layer which employs the sigmoid function to
perform logistic regression and produces a distribution over
binary saliency labels. We call the saliency map generated in
this way S2 .
In fact, this segment-wise spatial pooling stream of our network is an accelerated version of our previous work proposed
in [17]. Although they share the identical idea of inferring
saliency from contrast among multiscale contextual regions,
feature extraction and processing in the current method is
much more efficient as hundreds of segmental features for the
same image are instantaneously masked out from the feature
map generated by MS-FCN in a single forward pass. Moreover, our segment-wise spatial pooling stream also achieves
better results as segment features are extracted from our
multiscale fully convolutional network, which has been finetuned for salient object detection, instead of from the original
VGG-16 model for image classification.
C. Attentional Module for Saliency Map Fusion
To merge predicted saliency scores from the two different streams, there are three straightforward options: average
pooling, max-pooling and 1 × 1 convolution. However, all
these strategies are image content independent. As our two
network streams have complementary strengths in saliency
map prediction, inspired by [55], [56], we design a trainable
attentional module to generate content-dependent weight maps
for fusing the results from the two streams.
Let S1 and S2 be the probabilistic saliency maps from
the two network streams, respectively. The final saliency map
from our deep contrast network is calculated as a weighted
sum of these two maps. The spatially varying weights are
adaptively learned. Therefore, they are called weight maps. Let
S be the fused saliency map, W1 be the weight map for the
saliency map generated from the MS-FCN stream and W2 be
the weight map for the saliency map generated from the second
stream. The merged saliency map is calculated by summing the
element-wise product between each probability map (resized
to 1/8 the input image resolution) and its corresponding
weight map:
S = W1
S1 + W2
S2 .
(2)
We refer to [56] and call W1 and W2 attention weights as
they reflect how much attention should be paid to individual network streams as well as saliency scores at different
spatial locations. These two attention weights can also be
considered as feature maps that have the same size as the
predicted saliency maps, and thus can be jointly trained in
a fully convolutional network. In this work, we employ a
differentiable attention module to our deep network to infer
these attention weights. As illustrated in Fig. 1, the proposed
attention module receives as input the output feature map from
the feature masking layer, and it contains two convolutional
layers. The first layer has 512 filters with kernel size 3 × 3
while the second layer has two convolutional filters with kernel
size 1 × 1. The output feature map has two channels, further
fed into a SoftMax layer, which generates two score maps
corresponding to the aforementioned two attention weights.
D. Deep Contrast Network Training
We propose an alternate training scheme to train our network. Specifically, in the initialization phase, we pre-compute
the segments of all training images and train the segmentwise spatial pooling stream alone until convergence to obtain
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
its initial network parameters. Segment-wise saliency labeling
is performed by thresholding the average pixel-wise labeling
inside each segment, and the segment features are extracted
using the VGG-16 or ResNet-101 image classification model
pre-trained on the ImageNet dataset [57]. After initialization,
we alternately update the weights in the two network streams.
First, we fix the weights of the second stream and train the
MS-FCN as well as the attention module for one epoch.
Note that the weights in the attention module for adaptively
merging the predicted saliency maps from the two streams
are trained simultaneously with the MS-FCN stream in an
end-to-end mode. Next, we fix the weights in the MS-FCN
as well as the attention module, and fine-tune the parameters
in the second stream for one epoch using segment features
extracted with an updated VGG-16 or ResNet-101 network
embedded in the MS-FCN stream. We alternately train the
two streams 8 times (16 epochs in total) until the whole
training process converges. We define the following classbalanced cross-entropy as the loss function for training the
multi-scale fully convolutional steam and the attention module
of our network,
L = − βi
|I|
X
Gi log P (Si = 1|Ii , W )
i=1
− (1 − βi )
|I|
X
(3)
(1 − Gi ) log P (Si = 0|Ii , W ) ,
i=1
where βi represents the class balancing weight, denoted as
|I|+
βi = |I|
|I| and 1 − βi =
|I| , where |I|, |I|+ and |I|
respectively indicate the total number of pixels, salient pixels
and non-salient ones in image I. G represents the groundtruth
annotation and W represents the collection of all network
weights in the MS-FCN stream and the attention module.
When fine-tuning the segment-wise spatial pooling stream, we
use a batch of images as a unit and update parameters by
minimizing the summed squared errors accumulated over all
segments from the same batch of training images.
IV. T HE C OMPLETE A LGORITHM
A. Superpixel Segmentation
The segment-wise spatial pooling stream of our network requires the input image to be decomposed into non-overlapping
segments. In order to better avoid artificial boundaries in the
generated saliency map, each segment should be a perceptually homogeneous region while at the same time, strong
contours and edges should still be well preserved. In our
earlier version [1], we use a geodesic distance [58] based SLIC
algorithm for superpixel generation. In this work, we discover
that graph based image segmentation [59] produces segments
with better edge preservation than the SLIC algorithm, and
using segments generated from multiple levels of image segmentation can further improve the performance. Therefore, we
refer to [59] and employ the graph based image segmentation
algorithm therein to generate three levels of segments with
different parameter settings. We train a single segment-wise
spatial pooling stream for all the segments across three levels
of segmentation instead of learning different model parameters
7
for segments from different levels of segmentation. When
generating a saliency map from the segment-wise spatial
pooling stream, we apply the same stream to infer a saliency
map for each level of segmentation and then simply average
the three resulting saliency maps.
B. Salient Contour Detection
While in most cases our proposed deep contrast network
works well, it sometimes produces saliency maps where salient
region boundaries are not accurately localized, particularly for
images containing small salient regions. Meanwhile, we find
that our multi-scale fully convolutional network described in
Section III-A, when re-trained using annotated salient region
contours, is also capable of detecting the contours of salient
regions. The detected contours can be further encoded as
feature vectors and embedded into a CRF framework to enhance spatial coherence and the preservation of salient region
contours in saliency maps. To prepare training data for salient
region contour detection, boundary pixels of salient regions in
the groundtruth saliency maps are labeled ‘1’, and all other
pixels are labeled ‘0’. Such salient region contour maps are
taken as the groundtruth annotations when MS-FCN is trained
for salient region contour detection, and the class-balancing
weight is updated according to the fraction of pixels on salient
region contours.
Given a detected salient region contour map M , we apply
the normalized cut [60] algorithm to generate per-pixel feature
vectors, which are used in a fully connected CRF to improve
boundary localization in our final saliency map. First, we
construct a sparse graph where every pixel is connected to
other pixels in its 11 × 11 neighborhood. The affinity matrix
W of this graph is defined as follows,
M (p)2
,
(4)
Wij = exp − max
ρ
p∈ij
where Wij denotes the affinity between pixels i and j, p
represents pixels along the line segment (ij) connecting pixels
i and j, M (p) indicates the probability of pixel p being on a
salient region contour, and ρ is a constant scaling factor, which
is set to 0.1 in our experiments. The idea is that two pixels
should have a similar saliency value if there is no salient region
contour crossing the line segment connecting these two pixels.
Given an affinity matrix W , we further define Dii = Σi6=j Wij ,
and solve for generalized eigenvectors of the following system,
(D − W ) v = λDv. We use these eigenvectors as additional
features to improve spatial coherence. In our experiments, we
use eigenvectors corresponding to the 16 smallest eigenvalues.
C. Spatial Coherence
Since both streams of our deep contrast network independently infer the saliency score of each individual pixel
or segment without considering the impact of the correlation among pixels and segments on saliency prediction, the
resulting saliency maps contain more or less incomplete or
false positive salient objects. To mitigate this issue, we adopt
a fully connected conditional random field (CRF) [61] in a
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
post-processing step to enhance spatial coherence. The energy
function of the CRF model is formulated as
X
X
E (L) = −
log P (li ) +
θij (li , lj ) ,
(5)
i
i,j
where L is the binary label prediction for all pixels (salient or
not salient). P (li ) indicates the probability of pixel xi being
labeled li . As an initialization, P (li = 1) = Si and P (li =
0) = 1 − Si , where Si refers to the predicted probabilistic
saliency value at pixel xi of the saliency map S generated from
our deep contrast network. The pairwise potential θij (li , lj )
is defined as
"
2
2
kIi − Ij k
kpi − pj k
−
θij = µ (li , lj ) ω1 exp −
2σα2
2σβ2
!
! # (6)
2
2
kvi − vj k
kpi − pj k
−
+ ω2 exp −
,
2σγ2
2σ2
where µ (li , lj ) = 1 if li 6= lj , and zero, otherwise. It involves
a summation of two Gaussian kernels. The first kernel is
based on the observation that neighboring pixels should be
assigned similar saliency scores if they have similar colors but
do not have intervening salient region contours. It therefore
depends on pixel positions (p), pixel intensities (I) and the
contour feature embedding (v) discussed in Section IV-B. The
importance of color similarity, spatial closeness and salient
region contours are controlled by three parameters (σα , σβ
and σγ ), respectively. The second kernel is only dependent on
pixel positions with hyperparameter σ controlling the scale
of the Gaussian function. As pointed out in [62], it helps to
enhance label smoothness and remove small isolated regions.
As it has been proved in [61], this energy minimization
process can be modeled as efficient approximate probabilistic
inference by adopting a mean-field approximation to the
original CRF. High-dimensional filtering can be employed to
speed up the computation. We adapt the publicly available
implementation of [61] to minimize the above energy function.
The optimization process takes less than 0.5 second for an
image with 300 × 400 pixels. After CRF model optimization,
a saliency map Scrf can be generated from the pixelwise
posterior probabilities of saliency labels. We visualize the
effectiveness of our CRF in Fig. 3. As can be seen, the
(a)Source
(b)w/o CRF
(c)w/ CRF,
w/o contour
(d)saliency
contour
(e) w/ CRF,
w/ contour
(f)GT
Fig. 3: Examples of saliency maps generated with and without
a CRF (including CRFs with and without a contour feature
embedding).
8
original saliency maps from the proposed method without
CRF are rather coarse and the integrity (spatial coherence)
of detected salient regions can hardly be maintained. Though
saliency maps generated with a traditional CRF (without the
contour feature embedding) can enhance the spatial coherence
of detected salient regions to some extent, salient region
contours still may not be well positioned and there may be
false detections in the smooth background (e.g. the third row).
The fourth column of the figure demonstrates salient region
contours detected by our proposed method. As can be seen,
it is usually possible to accurately capture the boundaries
of salient regions and its corresponding embedded features
can further enhance the consistency of saliency prediction
across salient region contours and correct prediction errors.
A quantitative analysis of our CRF based saliency refinement
will be provided in Section V-C2.
V. E XPERIMENTAL R ESULTS
A. Experimental Setup
1) Datasets: We evaluate our proposed method on 6
widely used saliency detection benchmarks, including MSRAB [15], HKU-IS [17], PASCAL-S [35], DUT-OMRON [11],
ECSSD [63] and SOD [64]. MSRA-B includes 5,000 images, most of which holds a single salient object. HKUIS is proposed in our previous work [17], which has 4,447
images and most of the images include multiple separate
salient objects. PASCAL-S is based on the validation set of
PASCAL VOC2010 segmentation challenge [65] and contains
850 natural images. DUT-OMRON has 5,168 challenging images, which have relatively complex and diversified contents.
SOD has 300 images and was originally designed for image
segmentation. It is very challenging as most of the images
contain multiple objects and have low contrast or cluttered
background. We train the proposed contrast-oriented deep
neural networks based on the combination of both the training
sets of the MSRA-B (2500 images) and the HKU-IS (2500
images). The two validation sets are also combined as our
final validation, which contains a total of 1,000 images. We
test the model trained on this combined training set over all
other datasets to verity the model’s adaptability.
2) Evaluation Criteria: We employ precision-recall (PR)
curves, F-measure and mean absolute error (MAE) to quantitatively evaluate the performance of our method as well as
other salient object detection methods. Given a saliency map
with continuous values normalized to the range of 0 and 255,
we compute the binary masks by using every possible fixed
integer threshold. A pair of precision/recall values can be
computed by comparing each binary mask against the ground
truth. The precision is defined as the ratio between detected
groundtruth salient pixels and all predicted salient pixels in the
binary mask while the recall being the ratio between detected
groundtruth salient pixels and all groundtruth salient pixels.
Once the precision/recall pairs of all binary maps have been
computed, the PR curve can be plotted by averaging all pairs
of precision and recall values over all saliency maps of a
given dataset. F-measure is defined as the harmonic mean of
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
(a)Source
(b)HS
(c)DRFI
(d)PISA
(e)BSCA
(f)LEGS
(g)MC
(h)MDF
(i)DS
9
(j)DHSNET
(k)RFCN
(m)DCL+
(l)DCL
(n)GT
Fig. 4: Visual comparison between our methods (DCL and DCL+ ) and other state-of-the-art methods. Source: the input images;
GT: ground truth saliency maps; DCL+ : DCL with CRF refinement. DCL+ consistently achieves the best results in a variety
of complex scenarios.
Data Set
MSRA-B
ECSSD
HKU-IS
DUT-OMRON
PASCAL-S
SOD
Metric
maxF
MAE
maxF
MAE
maxF
MAE
maxF
MAE
maxF
MAE
maxF
MAE
SF
0.700
0.166
0.548
0.219
0.590
0.173
0.495
0.147
0.493
0.240
0.516
0.267
GC
0.719
0.159
0.597
0.233
0.588
0.211
0.495
0.218
0.539
0.266
0.526
0.284
HS
0.813
0.161
0.727
0.228
0.710
0.213
0.616
0.227
0.641
0.264
0.646
0.283
DRFI
0.845
0.112
0.782
0.170
0.776
0.167
0.664
0.150
0.690
0.210
0.699
0.223
PISA
0.837
0.102
0.764
0.150
0.753
0.127
0.630
0.141
0.660
0.196
0.660
0.223
BSCA
0.830
0.130
0.758
0.183
0.723
0.174
0.617
0.191
0.666
0.224
0.654
0.251
LEGS
0.870
0.081
0.827
0.118
0.770
0.118
0.669
0.133
0.752
0.157
0.732
0.195
MC
0.894
0.054
0.837
0.100
0.798
0.102
0.703
0.088
0.740
0.145
0.727
0.179
MDF
0.885
0.066
0.832
0.105
0.861
0.076
0.694
0.092
0.764
0.145
0.785
0.155
DS
0.898
0.067
0.900
0.079
0.866
0.079
0.773
0.084
0.834
0.108
0.829
0.127
RFCN
—
—
0.899
0.091
0.896
0.073
0.747
0.095
0.832
0.118
0.805
0.161
DHSNet
—
—
0.907
0.059
0.892
0.052
—
—
0.824
0.094
0.823
0.127
DCL
0.929
0.046
0.921
0.061
0.909
0.050
0.799
0.070
0.851
0.098
0.848
0.122
DCL+
0.931
0.042
0.925
0.058
0.913
0.041
0.811
0.064
0.857
0.092
0.857
0.120
TABLE I: Quantitative comparison in terms of maximum F-measure (larger is better) and MAE (smaller is better). The three
best performing algorithms are marked in red, blue, and green, respectively. As the testing set of the MSRA-B dataset is used
as part of the training set in the released model of DHSNet [46] and RFCN [48], and the part of the DUT-OMRON dataset
is also used in training the DHSNet model, we exclude the corresponding results here.
1.0
1.0
1.0
0.9
0.9
0.9
0.8
0.8
0.7
0.7
DR
0.5
GC
HS
0.4
0.3
0.2
LEGS
MC
0.6
HS
0.5
0.3
PISA
SF
0.6
0.7
BSCA
GC
HS
0.5
LEGS
MC
0.4
0.3
PISA
SF
MDF
MDF
SF
DS
DHSNet
0.2
0.2
DHSNet
MSRA-B
DS
0.1
DCL+
ECSSD
RFCN
0.1
0.0
0.2
0.4
0.6
Recall
0.8
1.0
0.0
0.3
0.2
0.4
0.6
0.8
1.0
Recall
0.0
MC
PISA
MDF
RFCN
0.1
HKU-IS
DS
DUT-OMRON
DS
DCL+
0.0
0.2
HS
LEGS
0.4
DCL+
0.0
DR
GC
0.5
RFCN
DCL+
0.0
0.6
SF
PISA
MDF
0.1
LEGS
MC
0.4
0.8
BSCA
DR
GC
Precision
BSCA
Precision
Precision
DR
0.6
0.9
0.8
BSCA
Precision
0.7
1.0
0.0
0.2
0.4
0.6
Recall
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
Recall
Fig. 5: Precision-recall curves of our method and 12 other state-of-the-art algorithms on 4 benchmark datasets. Our DCL+ (DCL
with CRF) consistently performs better than other methods across all the benchmarks.
the average precision and the average recall, which can be
calculated as
(1 + β 2 ) · P recision · Recall
Fβ =
,
(7)
β 2 · P recision + Recall
2
where β is set to 0.3 to place more emphasis on precision
than recall, as suggested in [24]. During evaluation, we report
the maximum F-measure (maxF) among all F-measure scores
computed from precision/recall pairs on the PR curve. We
also use twice the mean value of every saliency map as
the threshold to generate the corresponding binary map and
report the average precision, recall and F-measure of all binary
maps. As a complement, we also calculate the mean absolute
error (MAE) [26] as follows to quantitatively measure the
average absolute per-pixel difference between an estimated
saliency map S and the corresponding groundtruth saliency
map G.
M AE =
W X
H
X
1
|S(x, y) − G(x, y)|.
W × H x=1 y=1
(8)
3) Implementation: Our proposed model has been implemented on top of the open source code of DeepLab [21], which
is based on the Caffe platform [66]. It was trained with a GTX
Titan X GPU and Intel-i7 3.6GHz CPU.
During training, we resize all the images and their corresponding groundtruth saliency maps to 321×321, and perform
data augmentation by horizontal flipping. While training the
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
0.95
MSRA-B
0.90
pre
rec
0.85
fmea
0.80
0.75
0.70
0.65
0.60
0.55
0.50
0.45
DCL+
MC
MDF
DS
LEGS
PISA
DRFI
BSCA
HS
GC
SF
0.95
0.90
0.85
0.80
0.75
0.70
0.65
0.60
0.55
0.50
0.45
0.40
0.35
0.30
10
0.90
ECSSD
0.90
HKU-IS
0.85
pre
0.85
rec
0.80
rec
0.80
rec
fmea
0.75
fmea
0.75
fmea
0.70
0.70
0.65
0.65
0.60
0.60
0.55
0.55
0.50
0.50
0.45
MDF
MC
LEGS DRFI PISA BSCA
HS
GC
SF
0.35
pre
0.45
0.40
DCL+ RFCN DHSNet DS
DUT-OMRON
pre
0.40
DCL+ RFCN DHSNet MDF
DS
MC
LEGS PISA DRFI BSCA
HS
SF
GC
0.35
DCL+ RFCN MDF
DS
MC
LEGS DRFI
PISA
HS
BSCA
SF
GC
Fig. 6: Precision, recall and F-measure achieved using an adaptive threshold for every image. Our proposed method consistently
performs best among 13 different methods on 4 datasets.
methods in various challenging cases, e.g., salient regions
touching the image boundary (the first and fifth rows), low
contrast between salient objects and the background (the third
and sixth rows) and images with multiple separate salient
objects (the last three rows).
Our method significantly outperforms all other methods,
including those fully convolutional network based deep models published after our earlier conference version [1], by
a large margin on all public datasets in terms of the PR
curve (Fig. 5) as well as average precision, recall and Fmeasure (Fig. 6). Moreover, for the purpose of quantitative
evaluation, we report a comparison of maximum F-measure
and MAE in Table I. Our complete model (DCL+ ) clearly
outperforms the previous best-performing method by 3.67%,
1.98%, 1.90%, 10.64%, 2.76% and 3.38% in terms of maximum F-measure on MSRA-B (skipping RFCN and DHSNet
on this dataset), ECSSD, HKU-IS, DUT-OMRON (skipping
DHSNet), PASCAL-S and SOD, respectively. And at the same
time, it respectively lowers the MAE by 22.22%, 1.69%,
21.15%, 23.81%, 2.13% and 5.51%. It can also be observed
that the proposed method (DCL) without CRF-based postprocessing already outperforms all evaluated methods on all
considered datasets. We also compare run-time efficiency
among the considered algorithms. As shown in Table II,
our DCL model needs around 0.68 second to generate a
saliency map in the testing phase, which is comparable to
other fully convolutional methods (DS [41], RFCN [48] and
DHSNet [46]), and is much more efficient than other regionbased CNN models (LEGS [19], MC [18], MDF [17]).
1.0
0.91
0.9
0.90
0.8
Precision
MS-FCN stream, we set the learning rate for all newly added
layers to 10−3 and the learning rate for the rest of the layers to
10−4 . We employ a “poly” learning rate updating
powerpolicy [67]
iter
after each
(the learning rate is scaled by 1 − max
iter
iteration, and power = 0.9). We set the weight decay to 0.0005
and the momentum parameter to 0.9 during training. For the
segment-wise spatial pooling stream, we refer to [59] and
obtain 300 segments for each image from 3 levels of image
segmentation achieved with different parameter settings. We
set the grid size to 2×2 while performing spatial pooling over
each segment and the aggregated feature is of 6144 dimensions
in the VGG-16 based MS-FCN and 12288 dimensions in the
ResNet-101 based MS-FCN. This feature is further fed into
a sub-network consisting of two fully connected layers, each
of which contains 300 neurons. As in [61], we determine the
parameters of the fully connected CRF by performing cross
validation on the validation set. Finally, the actual value of
w1 , w2 , σα , σβ , σγ and σ are respectively set to 3.0, 5.0,
3.0, 50.0, 3.0 and 9.0 during evaluation.
We use DCL+ and DCL to respectively represent our best
saliency detectors with and without CRF-based refinement.
While it takes approximately 25 hours to train our model,
it only costs around 0.7 second for DCL to process an image
of size 400 × 300 on a PC with NVIDIA Titan X GPU and
Intel-i7 3.6GHz CPU. Note that this is far more efficient than
region-wise deep saliency detectors which independently treat
all image patches or superpixels during saliency estimation.
However, CRF-based post-processing is more expensive, and
requires additional 8 seconds since we need to compute
generalized eigenvectors used in the CRF model. Experimental
results reported in the following section show that DCL alone
without CRF refinement already performs better than most of
the existing state-of-the-art methods. A specific comparison of
the computational cost of different methods is summarized in
Table II.
0.89
0.7
0.88
0.6
DCL+
0.87
0.5
DCL
0.4
0.86
MSFCN
pre
Segment-Level
0.85
0.3
SC_MSFCN
0.2
0.5
0.6
0.7
0.8
0.9
1.0
0.84
rec
fmea
DCL+
DCL
MSFCN
Segment-Level SC_MSFCN
Recall
B. Comparison with the State of the Art
We compare our models (DCL and DCL+ ) with 9 other
state-of-the-art algorithms, including SF [26], GC [10],
HS [63], DRFI [14], PISA [68], BSCA [69], LEGS [19],
MC [18], MDF [17], DS [41], RFCN [48], DHSNet [46].
The last three are fully convolutional neural network based
methods, which were published after the publication of our
earlier conference version [1].
For qualitative evaluation, Figure 4 provides a visual comparison of saliency detection results, and the results from
our proposed method achieve much improvement over those
from other state-of-the-art algorithms. Specifically, our method
is capable of highlighting salient regions missed by other
Fig. 7: Component-wise validation of the proposed model and
the effectiveness of CRF based refinement.
C. Ablation Studies
1) Component-wise Effectiveness of Deep Contrast Network: To validate the necessity and effectiveness of the two
components contained in our deep contrast network, we take
the VGG-16 based version as a representative and compare
the saliency maps S1 inferred from the first stream (MSFCN), the saliency maps S2 from the second stream as well
as the fused ones based on S1 and S2 . As shown in Fig. 7,
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
SF
0.115
Time(s)
GC
0.25
HS
0.43
DRFI
47.08
PISA
0.65
BSCA
2.03
LEGS
2.00*
MC
2.38∗
11
MDF
8.00∗
DS
0.25∗
RFCN
4.60∗
DHSNet
0.24∗
DCL
0.68∗
TABLE II: Comparison of running time. *: GPU time.
VGG16
√
√
√
MSFCN
Attentional
ResNet-101
Module
√
√
√
√
√
√
√
√
√
√
√
Multi-Scale
Input
Segment-Level
SLIC
Multi-Scale
Superpixel
Segmentation
√
√
√
√
√
√
√
√
√
CRF
w/o contour
Metric
w/ contour
√
√
√
√
√
√
maxF
MAE
0.733
0.757
0.746
0.773
0.792
0.799
0.804
0.811
0.084
0.080
0.082
0.076
0.071
0.070
0.068
0.064
TABLE III: Performance evaluation of different model factors on DUT-OMRON Dataset.
(a)Source
(b)SC_MSFCN
(c)MSFCN
(d)Segment_Level
(e)DCL
(f)GT
Fig. 8: Sample visualizations demonstrating the componentwise efficacy of our deep contrast network.
the fused saliency map consistently performs best under all
evaluation metrics on the testing set of the MSRA-B dataset,
and the fully convolutional stream contributes to the merged
prediction far more than the segment-wise spatial pooling
stream. The two streams of our deep contrast network are
complementary and are capable of discovering global and local
contrast collaboratively through multiscale feature aggregation
in both streams. To validate the effectiveness of MS-FCN, we
have also generated saliency maps from the last scale of MSFCN for comparison. As illustrated in Fig. 7, a single scale
of MS-FCN (SC MSFCN) may lead to significantly inferior
performance when compared to the full version of MS-FCN in
terms of the PR curve as well as average precision, recall and
F-measure. Fig. 8 shows sample visualizations to demonstrate
the complementary nature of the two streams inside the
DCL network. As shown in the figure, although the fully
convolutional stream and the segment-wise spatial pooling
stream can produce promising saliency maps, they are far from
perfect. MS-FCN tends to generate very smooth saliency maps
but cannot well maintain the integrity of salient regions while
the segment-wise stream predicts saliency maps in the unit
of superpixels, it can hardly capture the global contrast and
cannot well handle images with a complex background. However, the fused DCL model exploits the advantages of both and
produces more accurate saliency predictions, which confirms
the complementarity of these two sub-networks. In particular,
there are examples (e.g. the second image in Fig. 8) where the
two streams have different mistakenly predicted regions, but
our proposed network still preferentially integrate respectively
predicted salient pixels and produce more accurate results.
This further demonstrates the robustness of our network and
the strong complementarity of the two network streams.
2) Effectiveness of Contour Guided CRF: As described in
Section IV-C, we incorporate a fully connected CRF with embedded contour features to further improve spatial coherence
and contour positioning in the saliency maps generated from
our deep contrast network. We compare the performance of
the generated saliency maps with and without CRF as postprocessing. As shown in Fig. 7, CRF significantly increases the
accuracy of the saliency maps generated for the testing images
of the MSRA-B dataset. We also show a visual comparison in
Figure 3 to illustrate the effectiveness of conventional CRF
post-processing and CRF incorporating salient region contours. As shown in this figure, conventional CRF improves the
spatial consistency of predicted results to a certain extent while
incorporating salient region contours enhances the confidence
of saliency predictions especially for pixels near detected
salient region boundaries.
D. Improvements after Conference Version
After the conference version of this work, we have made the
following five major modifications to our method: (1) adding
an attention module to infer spatially varying weights for
saliency map fusion, (2) employing the ResNet-101 network
in the fully convolutional stream, (3) running the fully convolutional stream on multiple scaled versions of the original
input image and fusing the results using max-pooling, (4)
training and testing the segment-wise spatial pooling stream
using segments from multi-level image segmentation, and (5)
performing salient region contour detection and incorporating
detected contours in the fully connected CRF during postprocessing. In Table III, we evaluate how each of these factors
affects the maximum F-measure and MAE on the DUTOMRON dataset. As shown in the table, these five factors
together contribute a 7.13% improvement in the maximum Fmeasure and a 20.0% decline in MAE in comparison to the
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
best reported results in the earlier conference version of this
paper.
1) Effectiveness of Attention Module: As described in Section III-C, instead of simply adding a 1×1 convolutional layer
on top of the saliency maps from the two network streams, we
design an attention module to infer spatially varying weight
maps. To validate its effectiveness, we conduct a performance
comparison between a deep contrast network with a trained
attention module and another deep contrast network with a
simple 1 × 1 convolutional layer. As shown in Table III,
adopting the attention module for saliency map fusion improves the maximum F-measure on the DUT-OMRON dataset
by 1.77% while lowering the MAE by 2.38%. Because of
the effectiveness of this mechanism, we always integrate this
module in our network in subsequent experiments.
2) Effectiveness of ResNet-101 in MS-FCN: As described
in Section III-A, we have attempted to replace the VGG-16
network with a transformed ResNet-101 network in the fully
convolutional stream of our deep network. To demonstrate its
effectiveness, we have trained a new deep contrast network
model for comparison. This new model is trained using the
same setting as Section V-D1 except that the transformed
VGG-16 network is replaced with a pre-trained and transformed ResNet-101. As shown in Table III, adopting ResNet101 instead of VGG-16 significantly improves the maximum
F-measure on the DUT-OMRON dataset by 3.62% while
lowering the MAE by 7.32%. We have also reached the same
conclusion as the VGG based DCL network that ResNet-101
in the single scale setting generates over-smoothed saliency
maps with prediction errors and performs much worse than
the multi-scale version with side branches. As shown in the
second and third columns of Fig 9, our proposed DCL network
with multi-scale ResNet-101 generates much more confident
and cleaner results than DCL with the original single-scale
ResNet-101.
3) Effectiveness of Multiple Scaled Inputs: Inspired by [56],
we adopt a multi-scale input strategy when generating a
saliency map from the fully convolutional stream. Specifically, we obtain three scaled versions of the original input
image with the scaling factor respectively set to 1, 0.75,
and 0.5, and independently feed these scaled images to the
fully convolutional stream. The three resulting saliency maps
are fused by taking the maximum response across scales
(a)Source
(b)DCL(Resnet)
(c)DCL
(MS-Resnet)
(d)DCL
(MS-Resnet +
Multiscale Input)
(e)GT
Fig. 9: Effectiveness of ResNet-101 in our DCL model.
12
for each position (i.e. max pooling). As shown in Table III,
multi-scale input brings an extra 2.46% improvement in the
maximum F-measure while lowering the MAE by 6.58%.
Sample visualizations are shown in the fourth column of Fig 9,
where fusing saliency predictions from multi-scale inputs gives
rise to more accurate saliency maps especially when there
exists multiple salient objects of different scales in the testing
image.
4) Effectiveness of Multi-Level Image Segmentation: As
described in Section IV-A, the final saliency map from the
revised segment-wise spatial pooling stream is the average
of three saliency maps, each of which is computed using all
superpixels from one of 3 levels of image segmentation. As
shown in Table III, multi-level image segmentation further
improves the maximum F-measure by 0.88% and lowers the
MAE by 1.40%.
5) Effectiveness of Salient Region Contours: As described
in Section IV, we revise the CRF-based post-processing step
in this version by integrating an additional feature vector
computed from detected salient region contours. Salient region
contours are detected using a separately trained contour detection model, which has the same network structure as the MSFCN stream. We compare saliency maps computed without
CRF, with CRF but without contour saliency features, and
with contour guided CRF, respectively. As shown in Table III,
post-processing our saliency maps with a dense CRF always
yields performance improvement. For the VGG16 based deep
contrast network, running CRF as a post-processing step
boosts the maximum F-measure by 3.27% and lowers the
MAE by 4.76%. For the ResNet-101 based deep contrast
network, which already achieves a much better performance
itself, adding a dense CRF still brings a 0.63% improvement in
the maximum F-measure and a 2.86% decrease in MAE. It is
worth noting that contour guided CRF results in more accurate
saliency maps with a 1.50% improvement in the maximum Fmeasure and a 8.57% decrease in MAE.
VI. C ONCLUSIONS
In this work, we have proposed end-to-end contrast-oriented
deep neural networks for salient object detection. Our deep
networks contain two complementary sub-networks and are
capable of extracting a wide variety of visual contrast information. The first sub-network is based on a multiscale fully
convolutional network, and is intended to infer pixel-wise
saliency by looking into contexts (receptive field) of multiple
scales around each pixel. The second sub-network is designed
to capture the contrast information among adjacent regions,
which can not only maintain the consistency of saliency
prediction within homogeneous regions but also better detect
discontinuities along salient region boundaries. An attentional
module with learnable weights is introduced to adaptively fuse
the two saliency maps from the two sub-networks. Finally, to
produce more accurate saliency predictions, we incorporate
a CRF with a contour feature embedding to further enhance
the spatial coherence and contour localization of the produced
saliency map. Experimental results show that the proposed
model achieves state-of-the-art performance on six public
benchmark datasets under various evaluation metrics.
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
R EFERENCES
[1] G. Li and Y. Yu, “Deep contrast learning for salient object detection,”
in Proc. IEEE Conf. CVPR, June 2016.
[2] Y. Wei, X. Liang, Y. Chen, X. Shen, M.-M. Cheng, J. Feng, Y. Zhao,
and S. Yan, “Stc: A simple to complex framework for weakly-supervised
semantic segmentation,” IEEE Trans. Pattern Anal. Mach. Intell., 2016.
[3] V. Navalpakkam and L. Itti, “An integrated model of top-down and
bottom-up attention for optimizing detection speed,” in Proc. IEEE Conf.
CVPR, vol. 2, 2006, pp. 2049–2056.
[4] Z. Wang, T. Chen, G. Li, R. Xu, and L. Lin, “Multi-label image
recognition by recurrently discovering attentional regions,” in Proc.
IEEE Conf. ICCV, 2017.
[5] S. Avidan and A. Shamir, “Seam carving for content-aware image
resizing,” in ACM Transactions on graphics (TOG), vol. 26, no. 3.
ACM, 2007, p. 10.
[6] H. Wu, G. Li, and X. Luo, “Weighted attentional blocks for probabilistic
object tracking,” The Visual Computer, vol. 30, no. 2, pp. 229–243, 2014.
[7] S. Bi, G. Li, and Y. Yu, “Person re-identification using multiple experts
with random subspaces,” Journal of Image and Graphics, vol. 2, no. 2,
2014.
[8] W. Einhäuser and P. KoÈnig, “Does luminance-contrast contribute
to a saliency map for overt visual attention?” European Journal of
Neuroscience, vol. 17, no. 5, pp. 1089–1097, 2003.
[9] D. Parkhurst, K. Law, and E. Niebur, “Modeling the role of salience in
the allocation of overt visual attention,” Vision research, vol. 42, no. 1,
pp. 107–123, 2002.
[10] M. Cheng, N. J. Mitra, X. Huang, P. H. Torr, and S. Hu, “Global contrast
based salient region detection,” IEEE Trans. Pattern Anal. Mach. Intell.,
vol. 37, no. 3, pp. 569–582, 2015.
[11] C. Yang, L. Zhang, H. Lu, X. Ruan, and M.-H. Yang, “Saliency detection
via graph-based manifold ranking,” in Proc. IEEE Conf. CVPR, 2013,
pp. 3166–3173.
[12] Q. Wang, Y. Yuan, and P. Yan, “Visual saliency by selective contrast,”
IEEE Transactions on Circuits and Systems for Video Technology,
vol. 23, no. 7, pp. 1150–1155, 2013.
[13] S. Lu, V. Mahadevan, and N. Vasconcelos, “Learning optimal seeds for
diffusion-based salient object detection,” in Proc. IEEE Conf. CVPR,
2014, pp. 2790–2797.
[14] P. Jiang, H. Ling, J. Yu, and J. Peng, “Salient region detection by ufo:
Uniqueness, focusness and objectness,” in Proc. IEEE Conf. ICCV, 2013,
pp. 1976–1983.
[15] T. Liu, Z. Yuan, J. Sun, J. Wang, N. Zheng, X. Tang, and H.-Y. Shum,
“Learning to detect a salient object,” IEEE Trans. Pattern Anal. Mach.
Intell., vol. 33, no. 2, pp. 353–367, 2011.
[16] L. Mai, Y. Niu, and F. Liu, “Saliency aggregation: a data-driven
approach,” in Proc. IEEE Conf. CVPR, 2013, pp. 1131–1138.
[17] G. Li and Y. Yu, “Visual saliency based on multiscale deep features,”
in Proc. IEEE Conf. CVPR, June 2015.
[18] R. Zhao, W. Ouyang, H. Li, and X. Wang, “Saliency detection by multicontext deep learning,” in Proc. IEEE Conf. CVPR, 2015, pp. 1265–
1274.
[19] L. Wang, H. Lu, X. Ruan, and M.-H. Yang, “Deep networks for saliency
detection via local estimation and global search,” in Proc. IEEE Conf.
CVPR, 2015, pp. 3183–3192.
[20] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks
for semantic segmentation,” Proc. IEEE Conf. CVPR, 2015.
[21] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille,
“Semantic image segmentation with deep convolutional nets and fully
connected crfs,” arXiv preprint arXiv:1412.7062, 2014.
[22] S. Xie and Z. Tu, “Holistically-nested edge detection,” Proc. IEEE Conf.
ICCV, 2015.
[23] D. Gao and N. Vasconcelos, “Bottom-up saliency is a discriminant
process,” in Proc. IEEE Conf. ICCV, 2007, pp. 1–6.
[24] R. Achanta, S. Hemami, F. Estrada, and S. Susstrunk, “Frequency-tuned
salient region detection,” in Proc. IEEE Conf. CVPR, 2009, pp. 1597–
1604.
[25] D. Klein, S. Frintrop et al., “Center-surround divergence of feature
statistics for salient object detection,” in Proc. IEEE Conf. ICCV. IEEE,
2011, pp. 2214–2219.
[26] F. Perazzi, P. Krähenbühl, Y. Pritch, and A. Hornung, “Saliency filters:
Contrast based filtering for salient region detection,” in Proc. IEEE Conf.
CVPR, 2012, pp. 733–740.
[27] W. Zhu, S. Liang, Y. Wei, and J. Sun, “Saliency optimization from
robust background detection,” in Proc. IEEE Conf. CVPR, 2014, pp.
2814–2821.
13
[28] Q. Wang, Y. Yuan, P. Yan, and X. Li, “Saliency detection by multipleinstance learning,” IEEE transactions on cybernetics, vol. 43, no. 2, pp.
660–672, 2013.
[29] T. Judd, K. Ehinger, F. Durand, and A. Torralba, “Learning to predict
where humans look,” in Proc. IEEE Conf. ICCV, 2009.
[30] K.-Y. Chang, T.-L. Liu, H.-T. Chen, and S.-H. Lai, “Fusing generic
objectness and visual saliency for salient object detection,” in Proc. IEEE
Conf. ICCV. IEEE, 2011, pp. 914–921.
[31] S. Goferman, L. Zelnik-Manor, and A. Tal, “Context-aware saliency
detection,” TPAMI, vol. 34, no. 10, pp. 1915–1926, 2012.
[32] X. Shen and Y. Wu, “A unified approach to salient object detection via
low rank matrix recovery,” in Proc. IEEE Conf. CVPR, 2012.
[33] R. Liu, J. Cao, Z. Lin, and S. Shan, “Adaptive partial differential
equation learning for visual saliency detection,” in Proc. IEEE Conf.
CVPR, 2014.
[34] Y. Jia and M. Han, “Category-independent object-level saliency detection,” in Proc. IEEE Conf. ICCV. IEEE, 2013, pp. 1761–1768.
[35] Y. Li, X. Hou, C. Koch, J. M. Rehg, and A. L. Yuille, “The secrets
of salient object segmentation,” in Proc. IEEE Conf. CVPR, 2014, pp.
280–287.
[36] X. Hou and L. Zhang, “Saliency detection: A spectral residual approach,”
in Proc. IEEE Conf. CVPR, 2007.
[37] J. Lei, B. Wang, Y. Fang, W. Lin, P. Le Callet, N. Ling, and C. Hou, “A
universal framework for salient object detection,” IEEE Transactions on
Multimedia, vol. 18, no. 9, pp. 1783–1795, 2016.
[38] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105.
[39] R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature
hierarchies for accurate object detection and semantic segmentation,”
in Proc. IEEE Conf. CVPR, 2014, pp. 580–587.
[40] G. Li and Y. Yu, “Visual saliency detection based on multiscale deep
cnn features,” IEEE Transactions on Image Processing, vol. 25, no. 11,
pp. 5012–5024, 2016.
[41] X. Li, L. Zhao, L. Wei, M. Yang, F. Wu, Y. Zhuang, H. Ling, and
J. Wang, “Deepsaliency: Multi-task deep neural network model for
salient object detection,” arXiv preprint arXiv:1510.05484, 2015.
[42] G. Li, Y. Xie, L. Lin, and Y. Yu, “Instance-level salient object segmentation,” Proc. IEEE Conf. CVPR, 2017.
[43] G. Li, Y. Xie, and L. Lin, “Weakly supervised salient object detection
using image labels,” in Proc. Conf. AAAI, Feb 2018.
[44] J. Han, D. Zhang, S. Wen, L. Guo, T. Liu, and X. Li, “Two-stage
learning to predict human eye fixations via sdaes,” IEEE transactions
on cybernetics, vol. 46, no. 2, pp. 487–498, 2016.
[45] N. Li, B. Sun, and J. Yu, “A weighted sparse coding framework for
saliency detection,” in Proc. IEEE Conf. CVPR, 2015, pp. 5216–5223.
[46] N. Liu and J. Han, “Dhsnet: Deep hierarchical saliency network for
salient object detection,” in Proc. IEEE Conf. CVPR, 2016, pp. 678–
686.
[47] J. Kuen, Z. Wang, and G. Wang, “Recurrent attentional networks for
saliency detection,” arXiv preprint arXiv:1604.03227, 2016.
[48] L. Wang, L. Wang, H. Lu, P. Zhang, and X. Ruan, “Saliency detection
with recurrent fully convolutional networks,” in Proc. Conf. ECCV.
Springer, 2016, pp. 825–841.
[49] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” arXiv preprint arXiv:1512.03385, 2015.
[50] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[51] H. Li, R. Zhao, and X. Wang, “Highly efficient forward and backward
propagation of convolutional neural networks for pixelwise classification,” arXiv preprint arXiv:1412.4526, 2014.
[52] S. Mallat, A wavelet tour of signal processing. Academic press, 1999.
[53] R. Girshick, “Fast r-cnn,” in International Conference on Computer
Vision (ICCV), 2015.
[54] K. He, X. Zhang, S. Ren, and J. Sun, “Spatial pyramid pooling in deep
convolutional networks for visual recognition,” in Proc. Conf. ECCV.
Springer, 2014, pp. 346–361.
[55] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by
jointly learning to align and translate,” Proc. Conf. ICLR, 2015.
[56] L.-C. Chen, Y. Yang, J. Wang, W. Xu, and A. L. Yuille, “Attention
to scale: Scale-aware semantic image segmentation,” arXiv preprint
arXiv:1511.03339, 2015.
[57] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet:
A large-scale hierarchical image database,” in Proc. IEEE Conf. CVPR,
2009, pp. 248–255.
[58] A. Criminisi, T. Sharp, C. Rother, and P. Pérez, “Geodesic image and
video editing.” ACM Transactions on graphics (TOG).
TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. XX, NO. X, MONTH 2018
[59] P. F. Felzenszwalb and D. P. Huttenlocher, “Efficient graph-based image
segmentation,” IJCV, vol. 59, no. 2, pp. 167–181, 2004.
[60] J. Shi and J. Malik, “Normalized cuts and image segmentation,” IEEE
Trans. Pattern Anal. Mach. Intell., vol. 22, no. 8, pp. 888–905, 2000.
[61] P. Krähenbühl and V. Koltun, “Efficient inference in fully connected crfs
with gaussian edge potentials,” arXiv preprint arXiv:1210.5644, 2012.
[62] J. Shotton, J. Winn, C. Rother, and A. Criminisi, “Textonboost for
image understanding: Multi-class object recognition and segmentation
by jointly modeling texture, layout, and context,” International Journal
of Computer Vision, vol. 81, no. 1, pp. 2–23, 2009.
[63] Q. Yan, L. Xu, J. Shi, and J. Jia, “Hierarchical saliency detection,” in
Proc. IEEE Conf. CVPR, 2013.
[64] D. Martin, C. Fowlkes, D. Tal, and J. Malik, “A database of human
segmented natural images and its application to evaluating segmentation
algorithms and measuring ecological statistics,” in Proc. IEEE Conf.
ICCV, vol. 2, 2001, pp. 416–423.
[65] M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman, “The pascal visual object classes (voc) challenge,” International
journal of computer vision, vol. 88, no. 2, pp. 303–338, 2010.
[66] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
S. Guadarrama, and T. Darrell, “Caffe: Convolutional architecture for
fast feature embedding,” in Proceedings of the ACM International
Conference on Multimedia. ACM, 2014, pp. 675–678.
[67] W. Liu, A. Rabinovich, and A. C. Berg, “Parsenet: Looking wider to
see better,” arXiv preprint arXiv:1506.04579, 2015.
[68] K. Wang, L. Lin, J. Lu, C. Li, and K. Shi, “Pisa: Pixelwise image
saliency by aggregating complementary appearance contrast measures
with edge-preserving coherence,” IEEE Transactions on Image Processing, vol. 24, no. 10, pp. 3019–3033, Oct 2015.
[69] Y. Qin, H. Lu, Y. Xu, and H. Wang, “Saliency detection via cellular
automata,” in Proc. IEEE Conf. CVPR, 2015, pp. 110–119.
Guanbin Li is currently a research associate professor in School of Data and Computer Science,
Sun Yat-sen University. He received his PhD degree
from the University of Hong Kong in 2016. He
was a recipient of Hong Kong Postgraduate Fellowship. His current research interests include computer
vision, image processing, and deep learning. He
has authorized and co-authorized on more than 20
papers in top-tier academic journals and conferences.
He serves as an area chair for the conference of
VISAPP. He has been serving as a reviewer for
numerous academic journals and conferences such as TPAMI, TIP, TMM,
TC, CVPR2018 and IJCAI2018.
Yizhou Yu received the PhD degree from University of California at Berkeley in 2000. He is
currently a professor at The University of Hong
Kong, and was a faculty member at University of
Illinois, Urbana-Champaign between 2000 and 2012.
He is a recipient of 2002 US National Science
Foundation CAREER Award, and 2007 NNSF China
Overseas Distinguished Young Investigator Award.
He has served on the editorial board of IET Computer Vision, IEEE Transactions on Visualization
and Computer Graphics, The Visual Computer, and
International Journal of Software and Informatics. He has also served on
the program committee of many leading international conferences, including
SIGGRAPH, SIGGRAPH Asia, and International Conference on Computer
Vision. His current research interests include deep learning methods for
computer vision, computational visual media, geometric computing, video
analytics and biomedical data analysis.
14
| 9cs.NE
|
Distributed estimation from relative measurements of heterogeneous and
uncertain quality
arXiv:1710.08632v1 [cs.SY] 24 Oct 2017
Chiara Ravazzi
Nelson P. K. Chan
Abstract— This paper studies the problem of estimation from
relative measurements in a graph, where a vector indexed
over the nodes has to be reconstructed from relative pairwise
measurements of differences between the values of nodes
connected by an edge. In order to model the heterogeneity and
uncertainty of the measurements, we assume that measurements
are affected by additive noise distributed according to a Gaussian mixture. In this original setup, we formulate the problem
of computing the Maximum-Likelihood (ML) estimates and we
propose two novel Expectation-Maximization (EM) algorithms
for its solution. The main difference between the two algorithm
is that one of them is distributed and the other is not. The
former algorithm is said to be distributed because it allows
each node to compute the estimate of its own value by using
only information that is directly available at the node itself or
from its immediate neighbors. We prove convergence of both
algorithms and present numerical simulations to evaluate and
compare their performance.
I. I NTRODUCTION
Whenever measurements are used to estimate a quantity
of interest, measurement errors must be properly taken into
account and the statistical properties of these errors should be
identified to enable an effective estimation. In this paper, we
look at a specific case of this broad issue, within the context
of network systems. Namely, we consider the problem of
distributed estimation from relative measurements, defined
as follows. We assume to have a real vector, indexed over
the nodes of a graph: the nodes are allowed to take pairwise
measurements of the differences between their vector entries
and those of their neighbors in the graph. The estimation
problem consists in reconstructing the original vector, up
to an additive constant. This prototypical problem can be
applied in a variety of contexts [1]. One example is relative
localization of mobile automated vehicles, where the vehicles
have to locate themselves by using only distance measurements [2]. Another example is statistical ranking, where a
set of items needs to be sorted according to their quality,
which can only be evaluated comparatively [3], [4]. In all
these scenarios, the noise affecting the measurements can be
The research leading to this paper has been partly developed while
C. Ravazzi was at Department of Electronics and Telecommunications
(DET), Politecnico di Torino, Italy, and N. P. K. Chan and P. Frasca
were with the Department of Applied Mathematics, University of Twente,
Enschede, the Netherlands.
Chiara Ravazzi is with the National Research Council (CNR), Institute
of Electronics, Computers and Telecommunication Engineering (IEIIT), c/o
Politecnico di Torino, Italy. E-mail: [email protected].
Nelson P. K. Chan is with the University of Groningen, Groningen, the
Netherlands. E-mail: [email protected].
Paolo Frasca is with Univ. Grenoble Alpes, CNRS, Inria, GIPSA-lab, F38000 Grenoble, France and Research Associate at the IEIIT-CNR, Torino,
Italy. E-mail:[email protected].
Paolo Frasca
drastically heterogeneous and, more importantly, its distribution may not be known a priori. For instance, in vehicle
localization, distances between vehicles may be measured
by more or less accurate sensors; in a ranking system, the
items upon evaluation can be compared by more or less
trustworthy entities. It is thus important to identify unreliable
measurements and weight them differently in the estimation.
In order to model this uncertainty, in this paper we assume
that the measurement noise is sampled from a mixture of
two Gaussian distributions with different variances, representing good and poor measurements, respectively. Our
solution to this problem builds on the classical ExpectationMaximization (EM) approach [5], [6], where the likelihood
is maximized by alternating operations of expectation and
maximization.
We are particularly interested in finding effective distributed algorithms to solve this problem. More precisely, we
say that an algorithm is distributed if it requires each node
to use information that is directly available at the node itself
or from its immediate neighbors. Actually, many distributed
algorithms for relative estimation are available [7], [1], [8],
[9], [10], [11], [12], [13], but to the best of our knowledge
all of them assume that the quality of the measurements is
known beforehand. At the same time, there is a large literature on robust estimation that also covers estimation from
relative measurements, but often provides algorithms that are
not distributed; see for instance [14] and references therein.
We fill this gap by proposing a distributed algorithm. In this
way, we contribute to the growing body of research on distributed algorithms for network-related estimation problems
with heterogeneous and unknown measurements [15], [16],
where distributed algorithms based on consensus and ranking
procedures have been proposed to approximate MaximumLikelihood (ML) estimates.
Other authors have used EM to estimate Gaussian mixtures’ parameters in other problems of distributed inference
in sensor networks [17], [18], [19]. In these works, a network
is given and each node independently performs the E-step
from local observations and this information is suitably
propagated to collaboratively perform the M-step. EM is also
a key instrument in learning from unreliable observations in
the context of social sensing, which has the goal of designing
reliable inference systems based on unreliable information
reported by users [20], [21].
Our contribution: In this paper, we define the problem
of robust estimation from relative measurements when measurement noise is drawn from a Gaussian mixture and we
design two iterative algorithms that solve it. Both algorithms
are based on combining classical Weighted Least Squares
(WLS) with Expectation-Maximization (EM), which is a
popular tool in statistical estimation problems involving
incomplete data [22], [23]. The first algorithm is centralized,
whereas the second algorithm is distributed but requires to
know (approximately) the two variances of the Gaussian
mixture. This knowledge is not necessary for the centralized
version. Both algorithms are proved to converge and their
performance is compared on synthetic data. We observe that
the centralized algorithm has better performance, achieving
smaller estimation errors. The centralized algorithm also requires less iterations to converge, but each iteration involves
more computations.
Organization of the paper: We formally present the problem of relative estimation in Section II. The centralised LSEM algorithm is described in Section III and the Distributed
LS-EM algorithm in Section IV. Section V contains some
numerical examples and Section VI our conclusions. The
details of most proofs are postponed to the Appendix.
II. E STIMATION FROM RELATIVE MEASUREMENTS
A set of nodes V = {1, . . . , N } of cardinality N is
considered, each of them endowed with an unknown scalar
quantity x
ev ∈ R for v ∈ V. The relative estimation problem
consists, for each node u ∈ V, in estimating the scalar value
x
eu , based on noisy measurements of differences x
eu − x
ev
with v and u in V. The set of available measurements can
be conveniently represented by an oriented graph G = (V, E),
where each edge represents a measurement. The orientation
of the edges is conventionally assumed to be such that
(v, u) ∈ E only if u < v.
We let A ∈ {0, ±1}E×V be the edge incidence matrix of
the graph G, defined as follows. The rows and the columns
of A are indexed by elements of E and V, respectively. The
(e, w)-entry of A is 0 if vertex w and edge e are not incident,
and otherwise it is 1 or −1 according as e originates or
terminates at w, respectively. More precisely,
+1 if e = (v, w)
Aew = −1 if e = (w, v)
0
otherwise
for every e ∈ E. We let b ∈ RE be the vector collecting the
measurements
b = Ae
x + η,
where ηe , e ∈ E are mutually independent random variables
distributed N(0, σe2 ), having
σe = (1 − ze )α + ze β,
with 0 < α < β, ze ∼ Ber(p) and p ∈ (0, 1/2). Provided
α β, the value ze = 1 is associated to a measurement
that is unreliable. With this formulation the random variables
{ηe }e∈E are Gaussian mixtures, whose model is completely
described by three parameters: p, α and β. For convenience,
from now on we consider p known: the knowledge of p does
not entail a significant restriction to our analysis.
Our main goal is to obtain a robust estimate of the state
vector x
e by suitably taking into account the different quality
of the measurements. We thus consider a joint Maximum
Likelihood estimation
x
bML =
argmax
L(x, α, β)
(1)
x∈RV ,α>0,β>0
where
L(x, α, β) := log fb|x,α,β (b|x, α, β)
and
fb|x,α,β (b|x, α, β) =
Y 1−p
(b − Ax)2e
√
exp −
2α2
2πα2
e∈E
#
p
(b − Ax)2e
+p
.
exp −
2β 2
2πβ 2
(2)
The computational complexity of optimization problem (1)
makes a brute force approach infeasible for large graphs.
A. Estimation via Weighted Least Squares
Problem (1) becomes much simpler if we assume to know
the distribution that has produced the noise term for each
measurement. Using the noise source information α, β, and
zee for all e ∈ E, where zee is the realization of ze , the MLestimation becomes
x
bML (b|e
z , α, β) = argmax log fb|x,ez,α,β (b|x, ze, α, β)
x∈RV
where
Y 1 − zee
(b − Ax)2e
√
exp −
2α2
2πα2
e∈E
#
zee
(b − Ax)2e
+p
exp −
.
2β 2
2πβ 2
Q
P
Noticing that log( e xe ) =
e is a
e log(xe ) and that z
binary vector, it is easy to see that ML is equivalent to solving
the Weighted Least Square (WLS) problem
fb|x,ez,α,β (b|x, ze, α, β) =
1
1
argmin kb−Axk2W = argmin (b−Ax)> W (b−Ax) (3)
2
2
V
V
x∈R
x∈R
with W = diag((1 − zee ) α−2 + zee β −2 ). The following
lemma describes the solutions of (3).
Lemma 1 (WLS estimator): Let the graph G be connected
and X be the set of solutions of (3), and let LW := A> W A
denote the weighted Laplacian of the graph. The following
facts hold true:
1) x ∈ X if and only if A> W Ax = A> W b;
2) there exists a unique x
bwls ∈ X such that kb
xwls k2 =
minx∈X kxk2 ;
3)
x
bwls = L†W A> W b,
(4)
where L† denotes the Moore-Penrose pseudo-inverse
of the weighted Laplacian LW .
Further useful properties are collected in the following result.
b(4,5)
x
e5
b(1,5)
x
e1
b(2,5)
b(3,4)
The gradient descent algorithm (5) converges to the WLS
solution, provided τ < 2kLW k−1
2 ; more details can be found
in [24].
b(1,2)
x
e4
III. C ENTRALIZED ALGORITHM
x
e2
b(2,3)
x
e3
Fig. 1.
The network of 5 nodes considered in Example 1.
Lemma 2 (Moments of WLS estimator): Provided G is
connected, it holds true that
1
E[b
xwls ] = I − 11> x
e
N
wls
wls
wls
wls >
E[(b
x − E[b
x ])(b
x − E[b
x ]) ] = L†W ,
In this section, we tackle the likelihood maximization
problem (1) in its full generality. Since there is no closed
form solution to it, we shall design an iterative algorithm
for its solution. Preliminarily to designing our algorithm,
we convert the Maximum Likelihood problem into a minimization problem by the following result, whose proof is
postponed to Appendix A.
Theorem 3: The following optimization problems have
the same solutions
(6)
− min min min V (x, π, α, β)
(7)
α,β
α,β
where 1 is a vector of length N whose entries are all 1.
It should be stressed that determining the state vector x
e
from relative measurements is only possible up to an additive
constant, being A1 = 0, and X = x
bwls + span(1). This
ambiguity can be avoided by assuming the centroid of the
nodes as the origin of the Cartesian coordinate system.
In view of this comment and of the results above, we shall
assume from now on that G is connected and 1> x
e = 0.
Next we provide a simple example for illustration.
Example 1: Consider the connected network in
Figure 1 with N = 5 nodes and set of edges
E
=
{(2, 1), (5, 1), (3, 2), (5, 2), (4, 3), (5, 4)}. Let
x
e
=
[1.9047, −0.1308, −0.0526, 0.2068, −1.9280]T ,
ze
=
[1, 1, 1, 1, 0, 0]T
and
η
=
[−0.0607, −0.0557, −0.0003, 0.0766, −0.0077, 0.0037]T .
Then, the incidence matrix A and the vector of measurements
can be easily constructed as
1 −1 0
0
0
1 0
0
0 −1
0 1 −1 0
0
A=
0
0 −1
0 1
0 0
1 −1 0
0 0
0
1 −1
T
and b = [1.9748, 3.7770, −0.0785, 1.8738, −0.2671, 2.1385]
Then the vector xwls can be computed as x
bwls =
T
[1.8636, −0.0982, −0.0524, 0.2134, −1.9264] . Instead, the
position estimated by an unweighted least squares is given
by x
bls = [1.8573 − 0.0988 − 0.03580.2157 − 1.9384]T .
We obtain that kb
xwls − x
ek2 /ke
xk22 = 3.7758 · 10−4 and
2
−4
ls
2
xk2 = 5.0452 · 10 .
kb
x −x
ek2 /ke
As shown in Lemma 1, the WLS solution is explicitly
known and can be easily computed solving a linear system.
Furthermore, the following distributed computation is also
possible, by using a gradient descent algorithm. Observe that
the gradient of the cost function in (3) is given by LW x −
A> W b. Set an initial condition x(0) = 0 and fix τ > 0 and
consider
x(t+1) = (I − τ LW )x(t) + τ A> W b.
max max L(x, α, β)
(5)
x
x
π∈[0,1]E
where
1 − πe
πe
1X
+
(8)
(b − Ax)2e
2
α2
β2
e∈E
X
1−p
p
− H(πe ) ,
+
−πe log − (1 − πe ) log
β
α
V (x, π, α, β) =
e∈E
H : [0, 1] → R is the natural entropy function
H(ξ) = −ξ log ξ − (1 − ξ) log(1 − ξ).
Note that, with respect to the original problem (6), problem (7) explicitly introduces the variable π ∈ [0, 1]E which
represents the estimated probabilities that the edges have
large variances. Actually, instead of solving problem (7), we
will solve a suitably modified problem, which we are going
to define next. This modification marks a key difference with
classical EM approaches. Namely, we shall solve
min min min Ṽ (x, π, α, β, ),
α,β
x
π∈ΣN −1
(9)
where Ṽ : RV × [0, 1]E × R>0 × R>0 × R≥0 → R is defined
as
Ṽ (x, π,α, β, )
(10)
1X
1
−
π
π
e
e
+ 2
=
(b − Ax)2e +
2
|E|
α2
β
e∈E
X
p
1−p
+
−πe log − (1 − πe ) log
− H(πe )
β
α
e∈E
and Σk = {π ∈ [0, 1]E : |supp(π)| ≤ |E| − k}. Compared
to (7), the optimization problem (9) introduces
• the positive variable , which has the goal to avoid possible singularities when one of the Gaussian components
of the mixture collapses to one point;
• the constraint set ΣN −1 , which forces the algorithm
to take into account at least N − 1 measurements as
reliable.
As will become clear in the proofs, these additional variables are instrumental to guarantee the convergence of the
algorithm.
The following lemma summarizes the main properties of
Ṽ in minimization problems that only involve one variable
at the time.
Proposition 4 (Partial minimizations):
x
b=x
b(π, α, β, ) = argmin Ṽ (x, π, α, β, )
x∈RV
π
b=π
b(x, α, β, ) = argmin Ṽ (x, π, α, β, )
π∈ΣN −1
α
b=α
b(x, π, β, ) = argmin Ṽ (x, π, α, β, )
α>0
b π, α, ) = argmin Ṽ (x, π, α, β, )
βb = β(x,
β>0
and denote
1−π
π
+ 2
W = diag
α2
β
LW = A> W A
π, we use them to re-estimate the mixture parameters α and
β.
The procedure is iterated until a suitable stopping criterion
is satisfied, e.g. a maximum number of iteration Tmax can be
fixed (SC ={t ≤ Tmax }) or the algorithm can be run until
the estimate stops changing (SC ={kx(t+1) −x(t) k/kx(t) k <
tol} for some tol > 0).
Algorithm 1 LS-EM
Require: Data: (b, A). Parameters: c1 , c2 > 0, p ∈ (0, 12 ).
1: Initialization: t ← 0, α(t) ← α0 , β (t) ← β0 , π (t) ←
0, (t) ← 1
2: while SC = 0 do
3:
Computation of weights: ∀e ∈ E
(t)
we(t+1) ←
ξe = f (ze = 1|x, α, β)
4:
WLS solution: W (t+1) ← diag(w(t+1) )
x(t+1) ← L†W (t+1) A> W (t+1) b,
Then, it holds true that
x
b = L†W A> W b
5:
Posterior distribution evaluation: ∀e ∈ E
π
b = PN −1 (ξ)
sP
α
b=
e (1
sP
βb =
(t)
πe
1 − πe
+ (t) 2
(α(t) )2
(β )
e
ξe(t+1) ← f (ze = 1|x(t+1) , α(t) , β (t) )
− πe )|be − (Ax)e |2 +
k1 − πk1
πe |be − (Ax)e |2 +
kπk1
where PN −1 : [0, 1]E → ΣN −1 is the projection that zeroes
the N − 1 smallest components of the given vector.
Proof: The statements can be directly verified by
differentiating Ṽ .
Using the insights obtained by Proposition 4, we propose
an alternating method for the minimization of (9). The
resulting method is a combination of Iterative Reweighted
Least Squares (IRLS) and Expectation Maximization. The
algorithm, which is detailed in Algorithm 1, is based on
the following four fundamental steps, which are iteratively
repeated until convergence.
WLS solution: Given the relative measurements b and
current parameters πe , α, β, a new estimation of the variable
x is obtained by solving the WLS problem with weights
we = (1 − πe )α−2 + πe β −2 ,
∀e ∈ E.
Expectation: The posterior distribution of the edges ξ is
evaluated, based on the current x, α, β.
Projection: The N − 1 smallest components of the posterior probabilities ξ are zeroed to ensure that at least N − 1
measurements are trusted and used in the weighted least
squares problem. The vector π = PN −1 (ξ) is the best
(|E| − N + 1)-approximation of the posterior probabilities
ξ.
Maximization: Given the projected posterior probabilities
6:
Best (|E| − N + 1)-approximation:
π (t+1) ← PN −1 (ξ (t+1) )
7:
Regularization parameter:
κ(t+1) ← dim(ker(LW (t+1) ))
θ(t+1) ←
8:
9:
10:
1
+c1 kx(t+1) −x(t) k+c2 (κ(t+1) −1)
log(t + 1)
(t+1) ← min (t) , θ(t+1)
Parameters estimation:
s
P
(t+1)
(t) + e (1 − πe
)|be − (Ax(t+1) )e |2
(t+1)
α
←
k1 − π (t+1) k1
s
P (t+1)
(t) + e πe
|be − (Ax(t+1) )e |2
β (t+1) ←
(t+1)
kπ
k1
t←t+1
end while
Although Algorithm 1 is a modified version of classical
EM, this fact is not sufficient to guarantee the convergence
of the proposed method. In fact, as observed in [22], a
generic EM algorithm is not guaranteed to converge to a
limit point but only to produce a sequence of points along
which the log-likelihood function does not decrease. Hence,
an explicit convergence proof is required in all specific cases.
Algorithm 1 also includes a regularization sequence (t) ,
which appears in the “Maximization” step and is designed
to be monotonic and to go to zero upon convergence of the
algorithm. The presence of such regularization is actually
instrumental to prove the convergence to a local maximum
of the log-likelihood function. In order to state it, denote
ζ (t) = (x(t) , π (t) , α(t) , β (t) , (t) ): then Algorithm 1 can be
seen as a map from RV × [0, 1]E × R>0 × R>0 × R≥0 to
itself that produces the sequence of iterates (ζ (t) ).
Theorem 5 (LS-EM convergence): For any b ∈ RE ,
the whole sequence ζ (t) converges to ζ ∞
=
(x∞ , π ∞ , α∞ , β ∞ , ∞ ) such that
x∞ = L†W ∞ A> W ∞ b, W ∞ = diag(w∞ )
(11a)
∞
∞
π
1
−
π
e
e
+ ∞
(11b)
we∞ =
|α∞ |2
|β |2
πe∞ = PN −1 (f (ze = 1|x∞ , α∞ , β ∞ ))
(11c)
sP
∞
∞
2
∞
e (1 − πe )|be − (Ax )e | +
(11d)
α∞ =
∞
k1 − π k1
sP
∞
∞
2
∞
e πe |be − (Ax )e | +
β∞ =
.
(11e)
kπ ∞ k1
∞
The converge point ζ is a fixed point of the algorithm and
a local minimum of Ṽ (·, ·, ·, ·, ε∞ ). If ε∞ = 0, then ζ ∞
locally maximizes the log-likelihood.
The proof of Theorem 5 is based on observing that function Ṽ in (9) is a Lyapunov function that is not increasing
along the sequence of iterates. Details are postponed to
Appendix B.
IV. D ISTRIBUTED ALGORITHM
In this section, we design and study a distributed algorithm
to solve problem (1), starting from the centralized one
proposed in the previous section. Preliminary, let us examine
steps 3–8 in Algorithm 1 in order to identify whether they
are amenable to a distributed computation. Steps 3 and 5
only require information depending on the edge e and are
thus inherently decentralized. Furthermore, we already know
that least squares problem in Step 4 can be computed by
a distributed procedure. Instead, steps 6–8 involve global
information and can not easily be distributed.
Based on this discussion, we propose the following simple
but effective variation of LS-EM algorithm, detailed in
Algorithm 2. The new algorithm is based on two design
choices. The first one is inspired by the distributed gradient
dynamics (5): instead of fully solving a WLS problem at
each iteration, we only perform one step of the corresponding
gradient iteration. In the second one, we assume α and β to
be known, thus removing the need for their estimation. A
further advantage is that keeping α and β fixed during the
evolution of the algorithm avoids certain difficulties in the
convergence analysis in Appendix B and namely removes
the need for regularization and projection steps. Even though
the knowledge of α and β can be a restrictive assumption,
we have observed that the algorithm is fairly robust to
uncertainties in these values: this quality is further discussed
in Remark 1 below.
Algorithm 2 Distributed LS-EM
Require: Data: (b, A). Parameters: p ∈ (0, 21 ), τ > 0, 0 <
α β.
1: Initialization: t ← 0, π (t) ← 0, x(t) ← 0
2: while SC = 0 do
3:
Computation of weights: ∀e ∈ E
(t)
we(t+1) ←
4:
(t)
πe
1 − πe
+ 2
2
α
β
Gradient step:
x(t+1) ← (I−τ LW )x(t) +τ A> W b, W ← diag(w(t+1) )
5:
Posterior distribution evaluation: ∀e ∈ E
πe(t+1) ← f (ze = 1|x(t+1) , α, β)
6:
7:
t←t+1
end while
The convergence of Algorithm 2 can be proved similarly to
Theorem 5, under the condition that the parameter τ belongs
to a certain range: details are postponed to Appendix C.
Theorem 6 (Distributed LS-EM convergence): If τ <
α/kAk2 , then for any b ∈ RE the sequence (x(t) , π (t) )
generated by Algorithm 2 converges to (x∞ , π ∞ ) such that
x∞ = L†W ∞ A> W ∞ b, W ∞ = diag(w∞ )
(12a)
∞
∞
1 − πe
π
we∞ =
+ e2
(12b)
α2
β
πe∞ = f (ze = 1|x∞ , α, β).
(12c)
The limit point (x∞ , π ∞ ) is a fixed point of the algorithm
and a local minimum of V .
V. N UMERICAL RESULTS
In this section, we provide simulations illustrating the
performance of the proposed algorithms: we are mainly
interested in comparing them in terms of convergence time
and final estimation error. At the same time, we would like to
examine how their performance depends on the parameters of
the problem, such as the parameters of the Gaussian mixture
(α, β, p) and the topology of the measurement graph G.
As a measure of performance, we consider the normalized
quadratic error (NQE), defined as
kb
x−x
ek22
∗ 100 [%] .
ke
xk22
(13)
where k · k2 denotes the 2-norm of a vector.
Let us begin by describing our baseline simulation setup
in details. First, we generate synthetic data to define the
estimation problem. The number of nodes is set to N =
50. The N components of the state vector are generated
randomly according to a uniform distribution in the interval
(0, 1): then, the mean is subtracted yielding a state vector
with mean 0. The topology is generated as Erdős-Rényi
random graphs with edge probability pedge ranging from 0.1
1
44
25
17
kx̂ − xk22
[%]
kxk22
0.4
0.2
0
0
5
10
15
20
25
30
35
40
45
50
Iteration
Fig. 3. NQE plotted against iteration count for 4 randomly chosen trials
obtained using LS-EM Algorithm (each color represents the result of a trial);
the parameter set is N = 50, pedge = 0.3, p = 0.1, α = 0.05, β/α = 5.
Note that the color of each chosen trial matches its counterpart in Fig. 4
20
18
16
14
12
10
8
6
2
7
11
10
34
29
41
30
42
45
9
13
19
16
0
0
5
10
15
20
25
30
35
40
45
50
Iteration
40
38
18
4820
24
22
14
Fig. 4. NQE plotted against iteration count for 4 randomly chosen trials
obtained using Distributed LS-EM Algorithm (each color represents the
result of a trial); the parameter set is N = 50, pedge = 0.3, p = 0.1, α =
0.05, β/α = 5. Note that the color of each chosen trial matches its
counterpart in Fig. 3
23
15
37
0.6
4
2
12
36
0.8
kx̂ − xk22
[%]
kxk22
up to 1 (i.e., an edge is created between two arbitrary nodes
with probability pedge ): a sample realization is in Fig. 2. In
the extreme case of pedge = 1, the graph generated is the
complete graph where all nodes are connected to all others.
We fix α = 0.05 and β/α in the range from 2 to 10. Also the
probability of getting a bad measurement p is taken between
0 and 1/2. The noise vector is then sampled from a normal
distribution using a combination of the above parameters.
Next, we simulate the iterative algorithms. We initialize
the state vector to be all zeros and also the vector π to
be all zeros, meaning that all measurements are initially
presumed to be good. For the LS-EM, an initial value for
α and β is specified: α is randomly chosen from the set
{0.1, 0.2, 0.3, 0.4, 0.5} and β = 2α in order to meet the
constraint β > α. They are then held fixed for the different
trials. For the Distributed LS-EM, the fixed values for α and
β are taken to be the true values. The stopping criterion
SC is chosen according to a tolerance tol = 10−4 , which
has been verified to be small enough to represent numerical
convergence. Then, we simulate different trials whereby for
each trial the vector ze is regenerated.
4
1
5
27
46
3
43
32
33
49
28
21
39
47
26
35
6
31
50
8
Fig. 2. An example of a generated random graph with the folowing details:
N = 50, pedge = 0.1, M = 110 with M the number of edges of the
realization.
In order to illustrate the evolution of the algorithms, we
plot the NQE against the iteration count for Algorithm 1 in
Fig. 3 and Algorithm 2 in Fig. 4. We have chosen four trials
out of a set of 250: note that the same trials (that is, the
same random graphs and measurements) are chosen for both
algorithms. We can observe from the plots that Algorithm 1
converges faster than Algorithm 2, but the two algorithms
achieve similar final errors in a majority of trials.
The comparison between Algorithms 1 and Algorithm 2
is further explored in Fig. 5, where the distributions of
the final NQEs for all the 250 trials mentioned above are
summarized via the boxplot command in MATLAB with
the default settings, showing the 25th (lower edge), 50th or
median (central mark) and 75th (upper edge) percentiles. In
order to make the comparison more complete and provide
benchmarks, we also include the weighted least squares as
per (4) (that uses the exact values for the parameters α, β
and p) and the “naive” unweighted least squares estimator
x
bls , in which we assume all measurements to be good. We
can observe that all our approaches have a median that is
clearly lower than the median of the LS approach. Actually,
the bulks of the error distributions are very similar to the
WLS benchmark, except for few trials of the Distributed LSEM that perform more poorly. A careful inspection of these
few trials shows that these large errors are due to incorrect
classification of the type of a small number of edges. This
phenomenon is not observed in LS-EM, possibly thanks
to a better diffusion of the information by the centralized
algorithm.
We also performed a parameter study in order to quantify
the behavior of the mean NQE with respect to pedge , p and
β. In Fig. 6, we can observe that the mean NQE decreases
with increasing pedge : as the graph becomes more connected,
there are more measurements available to estimate the state
variables. From the same figure, we can also observe that
starting from pedge = 0.4, the performance of the Distributed
LS-EM is similar to that of LS-EM. In Fig. 7, the mean
NQE increases with increasing p: this is due to the presence
of more bad measurements. A similar reasoning explains the
increase of NQE for increasing ratios β/α in Fig. 8. These
3
4
kx̂ − xk22
[%]
kxk22
kx̂ − xk22
[%]
kxk22
2.5
3
2
1.5
2
1
1
0.5
0
0
WLS
LS
LS-EM
0.50
LS-EM D
0.75
1.00
1.25
1.50
c
Fig. 5. Boxplot showing NQE for the different approaches; the parameter
set is N = 50, pedge = 0.3, p = 0.1, α = 0.05, β/α = 5. + are the NQE
considered as outliers by the boxplot command;
2.5
Fig. 9.
Boxplot of NQE for different initialization of α and β obtained using Distributed LS-EM; we consider β/α to be known and the
actual α and β values are set as c times their real values, with c ∈
{0.50, 0.75, 1.00, 1.25, 1.50}. The parameter set is N = 50, pedge =
0.3, p = 0.1, α = 0.05, β/α = 5.
2
1.75
kx̂ − xk22
kxk22
4
[%]
2.25
7
1.5
6
kx̂ − xk22
[%]
kxk22
Ave
3
1.25
1
5
0.75
4
0.5
3
0.25
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
pedge
2
1
0
Fig. 6. Mean NQE with respect to pedge ; the parameter set is N = 50, p =
0.1, α = 0.05, β/α = 5 and the number of trials is 1000. ◦ = WLS, + =
LS, = LS-EM, × = Distributed LS-EM.
2.25
[%]
4
6
8
10
β/α
Fig. 10. Boxplot of NQE for different initialization of β obtained using
Distributed LS-EM; we consider α to be known and let the actual ratio
β/α to be {2, 4, 6, 8, 10}. The parameter set is N = 50, pedge = 0.3, p =
0.1, α = 0.05, β/α = 5.
2.5
2
1.75
kx̂ − xk22
kxk22
4
2
1.5
Ave
3
1.25
1
0.75
0.5
0.25
0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
p
Fig. 7. Mean NQE with respect to p; the parameter set is N = 50, pedge =
0.3, α = 0.05, β/α = 5 and the number of trials is 1000. ◦ = WLS, + =
LS, = LS-EM, × = Distributed LS-EM.
2.5
2
1.75
kx̂ − xk22
kxk22
4
[%]
2.25
3
Remark 1 (Parameter uncertainty in Distributed LS-EM):
Crucially, in Algorithm 2 the values for α and β are assumed
to be known a priori. In practice one would usually not be
able to have this information: hence we want to explore the
sensitivity of the algorithm to incorrect choices of α and β.
1.5
1.25
Ave
dependencies on the parameters are consistent with intuition.
From these three figures, it is clear that the Distributed LSEM has larger average error than the centralised LS-EM
(which has in turn a larger error than the WLS estimate).
However, we should recall that the mean error of the Distributed LS-EM is driven up by the aforementioned sporadic
errors: a comparison of median values would show a smaller
gap from the centralized approach.
1
0.75
0.5
0.25
0
0
1
2
3
4
5
6
7
8
9
10
β/α
Fig. 8. Mean NQE with respect to β; the parameter set is N = 50, pedge =
0.3, p = 0.1, α = 0.05 and the number of trials is 1000. ◦ = WLS, + =
LS, = LS-EM, × = Distributed LS-EM.
In Fig. 9, we assume not to know the actual values of α
and β, but only their ratio β/α: we can observe that choosing
α and β to be smaller than their actual value yields a higher
median, while larger values than the real one yield similar
results. In Fig. 10, we assume to know α but not β: we can
observe that an incorrect and too large value of β increases
the presence of large errors, even though the bulk of the
error distribution remains similar. Overall, we conclude that
the algorithm is fairly robust to moderate uncertainties in the
knowledge of the parameters.
VI. C ONCLUDING REMARKS
In this paper, we have studied the problem of estimation
from relative measurements with heterogeneous quality. We
have introduced an novel formulation for the problem and we
have proposed two novel algorithms based on the method
of Expectation-Maximization. One of the algorithms has
the important feature of being fully distributed and thus
amenable to applications where communication is limited or
expensive. The other algorithm also distinguishes itself from
standard EM approaches, due to the presence of regularization variables and of a projection step, which help use coping
with the graph-dependent nature of the problem. Besides
designing the algorithms, we have proved their convergence
to a local maximum of the log-likelihood function (or to an
approximation when regularization is employed). We have
also presented a number of simulations that support the
good performance of the algorithms and their robustness
against uncertainties in the choice of the parameters. Despite
a generally good performance, the algorithms (an particularly the distributed one) may perform poorly on some
instances: this could be explained by the local nature of
the optimality results. Several interesting problems remain
open: we mention three of them here. First, one could further
investigate the role of the topology of the measurement
graphs in determining the performance of the algorithms:
namely, some topologies could be more effective for the
same number of measurements taken. Second, one could
look for distributed algorithms that need not to assume the
knowledge of the mixture parameters α and β. Third, one
could explore algorithms that perform a “hard” classification
of the measurements, as opposed to the “soft” classification
that is done in this paper, where the measurements are
assigned a probability of being of type α or β: some
preliminary results and designs with hard classification are
available in [25].
R EFERENCES
[1] P. Barooah and J. P. Hespanha, “Estimation from relative measurements: Electrical analogy and large graphs,” IEEE Transactions on
Signal Processing, vol. 56, no. 6, pp. 2181–2193, 2008.
[2] ——, “Estimation from relative measurements: Algorithms and scaling
laws,” IEEE Control Systems Magazine, vol. 27, no. 4, pp. 57–74,
2007.
[3] X. Jiang, L.-H. Lim, Y. Yao, and Y. Ye, “Statistical ranking and
combinatorial Hodge theory,” Mathematical Programming, vol. 127,
no. 1, pp. 203–244, 2011.
[4] B. Osting, C. Brune, and S. J. Osher, “Optimal data collection for
improved rankings expose well-connected graphs,” Journal of Machine
Learning Research, vol. 15, pp. 2981–3012, 2014.
[5] A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood
from incomplete data via the EM algorithm,” Journal of the Royal
Statistical Society. Series B, vol. 39, no. 1, pp. 1–38, 1977.
[6] T. K. Moon, “The expectation-maximization algorithm,” IEEE Signal
Processing Magazine, vol. 13, no. 6, pp. 47–60, 1996.
[7] A. Giridhar and P. R. Kumar, “Distributed clock synchronization over
wireless networks: Algorithms and analysis,” in IEEE Conference on
Decision and Control, San Diego, CA, USA, Dec. 2006, pp. 4915–
4920.
[8] S. Bolognani, S. Del Favero, L. Schenato, and D. Varagnolo,
“Consensus-based distributed sensor calibration and least-square parameter identification in WSNs,” International Journal of Robust and
Nonlinear Control, vol. 20, no. 2, pp. 176–193, 2010.
[9] W. S. Rossi, P. Frasca, and F. Fagnani, “Distributed estimation
from relative and absolute measurements,” IEEE Transactions on
Automatic Control, vol. PP, no. 99, pp. 1–1, 2017. [Online]. Available:
http://dx.doi.org/10.1109/TAC.2017.2661400
[10] A. Carron, M. Todescato, R. Carli, and L. Schenato, “An asynchronous
consensus-based algorithm for estimation from noisy relative measurements,” IEEE Transactions on Control of Network Systems, vol. 1,
no. 3, pp. 283–295, 2014.
[11] N. M. Freris and A. Zouzias, “Fast distributed smoothing of relative
measurements,” in IEEE Conference on Decision and Control, Maui,
HI, USA, Dec. 2012, pp. 1411–1416.
[12] C. Ravazzi, P. Frasca, R. Tempo, and H. Ishii, “Ergodic randomized algorithms and dynamics over networks,” IEEE Transactions on Control
of Network Systems, vol. 2, no. 1, pp. 78–87, 2015.
[13] P. Frasca, H. Ishii, C. Ravazzi, and R. Tempo, “Distributed randomized
algorithms for opinion formation, centrality computation and power
systems estimation,” European Journal of Control, vol. 24, pp. 2–13,
Jul. 2015.
[14] L. Carlone, A. Censi, and F. Dellaert, “Selecting good measurements
via `1 relaxation: A convex approach for robust estimation over
graphs,” in IEEE/RSJ Int. Conf. on Intelligent Robots & Systems, Sep.
2014, pp. 2667–2674.
[15] A. Chiuso, F. Fagnani, L. Schenato, and S. Zampieri, “Gossip algorithms for simultaneous distributed estimation and classification
in sensor networks,” IEEE Journal of Selected Topics in Signal
Processing, vol. 5, no. 4, pp. 691–706, 2011.
[16] F. Fagnani, S. M. Fosson, and C. Ravazzi, “A distributed classification/estimation algorithm for sensor networks,” SIAM Journal on
Control and Optimization, vol. 52, no. 1, pp. 189–218, 2014.
[17] R. D. Nowak, “Distributed EM algorithms for density estimation
and clustering in sensor networks,” IEEE Transactions on Signal
Processing, vol. 51, no. 8, pp. 2245–2253, 2003.
[18] M. Rabbat and R. Nowak, “Distributed optimization in sensor networks,” in Third International Symposium on Information Processing
in Sensor Networks, 2004. IPSN 2004, April 2004, pp. 20–27.
[19] D. Gu, “Distributed EM algorithm for Gaussian mixtures in sensor
networks,” IEEE Transactions on Neural Networks, vol. 19, no. 7, pp.
1154–1166, 2008.
[20] D. Wang, L. Kaplan, and T. F. Abdelzaher, “Maximum likelihood analysis of conflicting observations in social sensing,” ACM Transactions
on Sensor Networks, vol. 10, no. 2, p. 30, 2014.
[21] D. Wang, T. F. Abdelzaher, and L. Kaplan, Social Sensing: Building
Reliable Systems on Unreliable Data. Morgan Kaufmann, 2015.
[22] C. M. Bishop, Pattern Recognition and Machine Learning. Springer,
2006.
[23] M. R. Gupta and Y. Chen, “Theory and use of the EM algorithm,”
Foundations and Trends R in Signal Processing, vol. 4, no. 3, pp.
223–296, 2011.
[24] W. S. Rossi, P. Frasca, and F. Fagnani, “Transient and limit performance of distributed relative localization,” in IEEE Conference on
Decision and Control, Maui, HI, USA, Dec. 2012, pp. 2744–2748.
[25] N. P. Chan, “On analysis and design of algorithms for robust
estimation from relative measurements,” Master’s thesis, University
of Twente, Enschede, the Netherlands, Mar. 2016. [Online]. Available:
http://essay.utwente.nl/69425/
A PPENDIX
A. Properties of the likelihood
Theorem 3 converts the ML problem into a minimization
problem. Before proving it, we recall expression (2) and
introduce some useful notation:
fbe |x,α,β (be |x, α, β)
(b − Ax)2e
1−p
exp −
=√
2α2
2πα2
p
(b − Ax)2e
+p
exp −
2β 2
2πβ 2
(14)
fbe ,ze |x,α,β (be , ze |x, α, β)
1−p
(b − Ax)2e
= (1 − ze ) √
exp −
2α2
2πα2
#
p
(b − Ax)2e
+ze p
exp −
2β 2
2πβ 2
fbe |ze ,x,α,β (be |ze , x, α, β)
1 − ze
(b − Ax)2e
= √
exp −
2α2
2πα2
#
ze
(b − Ax)2e
+p
exp −
2β 2
2πβ 2
fze |be ,x,α,β (ze |be , x, α, β) =
in πe ∈ {0, 1}). Therefore
(15)
e∈E
Using the definition of function V in (8) we obtain
(16)
fbe ,ze |x,α,β (be , ze |x, α, β)
fbe |x,α,β (be |x, α, β)
e∈E
log fbe |x,α,β (be |x, α, β)
e∈E
=
=
X
log
X
fbe ,ze |x,α,β (be , ze |x, α, β)
e∈E
ze ∈{0,1}
X
X
e∈E
log
ze ∈{0,1}
fbe ,ze |x,α,β (be , ze |x, α, β)
q(ze )
q(ze )
for any q : {0, 1} → (0, 1) such that q(0) + q(1) = 1. From
Jensen’s inequality we get
X X
fb ,z |x,α,β (be , ze |x, α, β)
L(x, α, β) ≥
q(ze ) log e e
q(ze )
e∈E ze ∈{0,1}
|E|
log(2π) (17)
2
|E|
= − min V (x, π, α, β) −
log(2π)
2
π∈[0,1]E
L(x, α, β) ≥ max −V (x, π, α, β) −
π∈[0,1]E
L(x, α, β) = log fb|x,α,β (b|x, α, β)
Y
= log
fbe |x,α,β (be |x, α, β)
=
X
πe
1 − πe
(b − Ax)2e
+
−
H(π
)
e
2β 2
2α2
π∈[0,1]E
e∈E
X
p
1−p
|E|
+
πe log + (1 − πe ) log
−
log(2π)
β
α
2
≥ max −
Proof of Theorem 3: From (14) we have
X
L(x, α, β)
By differentiating V (x, π, α, β) with respect to πe we
write the optimality condition as
1 − πe
p
(b − Ax)2e
∂V
= − log
− log +
∂πe
πe
β
2β 2
2
1 − p (b − Ax)e
+ log
−
=0
α
2α2
from which we obtain
2
e
√ p 2 exp − (b−Ax)
2
2β
2πβ
π
be =
2
2
(b−Ax)
e
√ p 2 exp − 2β 2 e + √1−p 2 exp − (b−Ax)
2
2α
2πα
2πβ
= fze =1|be ,x,α,β (ze = 1|be , x, α, β).
Using this expression for π
b we get
X X
L(x, α, β) ≥
fze |be ,x,α,β (ze |be , x, α, β)
e∈E ze ∈{0,1}
fb ,z |x,α,β (be , ze |x, α, β)
× log e e
fz |b` ,x,α,β (ze |be , x, α, β)
X eX
=
fze |be ,x,α,β (ze |be , x, α, β)
e∈E ze ∈{0,1}
× log fbe |x,α,β (be |x, α, β) = L(x, α, β),
Let πe = q(ze = 1) = 1 − q(ze = 0). Therefore, we have
X
fb ,z =1|x,α,β (be , ze = 1|x, α, β)from which we conclude that the inequality is actually an
L(x, α, β) ≥
q(ze = 1) log e e
equality. Therefore
q(ze = 1)
e∈E
|E|
fb ,z =0|x,α,β (be , ze = 0|x, α, β)
log(2π).
L(x, α, β) = − min V (x, π, α, β) −
2
+ q(ze = 0) log e e
π∈[0,1]E
q(ze = 0)
where the last expression is obtained using the definition of
(b−Ax)2e
p
√
exp
−
2
X
2β
2
function V in (8). We conclude that
2πβ
=
πe log
πe
max max L(x, α, β) = − min min min V (x, π, α, β) + c
e∈E
x
α,β
α,β x π∈[0,1]E
(b−Ax)2e
1−p
√
exp
−
2α2
2πα2
with c = − |E|
+ (1 − πe ) log
2 log(2π).
1
−
π
e
X
πe
1 − πe
=−
(b − Ax)2e
+
− H(πe )
B. Proof of Theorem 5
2β 2
2α2
e∈E
X
In this section we describe the proof of Theorem 5,
p
1−p
|E|
+
πe log + (1 − πe ) log
−
log(2π).
showing
convergence of Algorithm 1.
β
α
2
e∈E
Lemma 7 (Monotonicity): The function Ṽ
defined
This inequality is true for all πe ∈ [0, 1] and e ∈ E (the in (10) is nonincreasing along the iterates ζ (t) =
function on the right-hand-side can be extended by continuity (x(t) , π (t) , α(t) , β (t) , (t) ).
Proof: By repeatedly applying Proposition 4 we see
that
A> W (t) A. From (19) we then have
Ṽ (x(t) , π (t) , α(t) , β (t) , (t) )
Ṽ (ζ
(t+1)
) = Ṽ (x
(t+1)
,π
(t+1)
≤ Ṽ (x
(t+1)
,π
(t+1)
≤ Ṽ (x
(t+1)
,α ,β
,π
(t)
(t)
(t)
(t)
≤ Ṽ (x , π
,α
(t+1)
,β
(t+1)
(t)
(t)
(t)
,α ,β
(t)
,α ,β
(t)
(t)
(t+1)
,
)
− Ṽ (x(t+1) , π (t+1) , α(t+1) , β (t+1) , (t+1) )
, )
≥Ṽ (x(t) , π (t) , α(t) , β (t) , (t) )
(t)
, )
(t)
, ) ≤ Ṽ (ζ
(t)
− Ṽ (x(t+1) , π (t) , α(t) , β (t) , (t) )
!
(t)
(t)
πe
(t)
1 − πe
1X
(t) 2
+ (t) 2
(b − Ax )e +
=
2
|E|
(α(t) )2
(β )
e∈E
!
(t)
(t)
(t)
1X
π
1
−
π
e
e
−
+ (t) 2
(b − Ax(t+1) )2e +
2
|E|
(α(t) )2
(β )
e∈E
)
for every time t, proving the result.
The following lemma implies that Algorithm 1 converges
numerically.
Lemma 8 (Asymptotic regularity): If (x(t) ) is the sequence generated by Algorithm 1, then x(t+1) − x(t) → 0 as
t → ∞.
Proof: If α(t) → 0 or β (t) → 0 as t → ∞, we have
→ 0 and, consequently, kx(t+1) − x(t) k2 → 0 and the
assertion is verified. If instead neither α nor β converge to
zero, then there exists a constant K > 0 and a sequence
of integers t` : N → N such that t` → ∞ as ` → ∞,
min{α(t` ) , β (t` ) } > K for all ` ∈ N. It holds true in general
that
X
X
πe(t) log β (t) +
(1 − πe(t) ) log α(t) − |E| log 2
e∈E
e∈E
(18)
≤ Ṽ (ζ (t) ) ≤ Ṽ (ζ (1) )
(t)
where the last inequality follows from Lemma 7. Then,
Ṽ (ζ (t` ) ) ≥ (|E| log K − |E| log 2).
Since x(t+1) = argmin Ṽ (x, π (t) , α(t) , β (t) , (t) ) we have
1
= (x(t) )> LW (t) x(t) − (x(t) )> A> W (t) b
2
1
− (x(t+1) )> LW (t) x(t+1) + (x(t+1) )> A> W (t) b
2
1
= (x(t) − x(t+1) )> LW (t) (x(t) + x(t+1) )
2
− (x(t) − x(t+1) )> A> W (t) b
1
= (x(t) − x(t+1) )> LW (t) x(t)
2
1
+ (x(t) − x(t+1) )> ( LW (t) x(t+1) − A> W (t) b)
2
From (20) we get
Ṽ (ζ (t) ) − Ṽ (ζ (t+1) )
1
1
≥ (x(t) − x(t+1) )> LW (t) x(t) − (x(t) − x(t+1) )> LW (t) x(t+1)
2
2
≥ ckx(t) − x(t+1) k2
for a suitable positive constant c. The last inequality is true
since LW (t) is positive semidefinite, the multiplicity of the
eigenvalue 0 is equal to 1 and 1> x(t) = 0 for all t ∈ N. In
fact, suppose by contradiction that there exists (tj ) such that
limj→∞ λ(tj ) = 0 with
λ(tj ) = min
v
Then, there needs to exist a subsequence t` such that α(t` )
or β (t` ) diverge. If β (t` ) → ∞, then (18) implies α(t` ) → 0.
From Step 8 in Algorithm 1 we obtain (t` ) → 0 and κ(t` ) →
0. We deduce that there exists `0 ∈ N such that κ(t` ) = 1 for
all ` > `0 . from which we get the contradiction λ(t` ) > c > 0
for all ` > `0 . The case α(t` ) → ∞ is analogous.
We now compute for ` ∈ N
x∈RV
Ṽ (x(t+1) , π (t) , α(t) , β (t) , (t) ) ≤ Ṽ (x(t) , π (t) , α(t) , β (t) , (t) )
(19)
= A> W (t) Ax(t+1) − A> W (t) b
(t+1)
(t)
>
t=1
tX
` −1
ckx(t) − x(t+1) k2
Ṽ (ζ (t) ) − Ṽ (ζ (t+1) )
= Ṽ (ζ (1) ) − Ṽ (ζ (t` ) )
(20)
(t)
= LW (t) x
−A W b=0
(t)
1−π
π (t)
= diag (α
+
and LW (t)
(t) )2
(β (t) )2
tX
` −1
t=1
∇x [Ṽ (x, π (t) , α(t) , β (t) , (t) )](x(t+1) )
where W
0≤
≤
and
v > (LW (tj ) − 11> )v
.
kvk2
≤ Ṽ (ζ (1) ) − (|E| log K − |E| log 2) = K 0 .
By letting ` → ∞, we obtain that kx(t) − x(t+1) k → 0.
=
Lemma 9: The sequence (x(t) )t∈N is bounded.
Proof: We first prove that (α(t) )t∈N and (β (t) )t∈N
are both upper bounded. By contradiction, there exists a
subsequence (t` )`∈N such that lim`→∞ β (t` ) = ∞. By
monotonicity of function Ṽ (see Lemma 7) and inequality (18) we have
X
X
πe(t` ) log β (t` ) +
(1 − πe(t` ) ) log α(t` ) − |E| log 2
e∈E
≤ Ṽ (x
(t` )
≤ Ṽ (x
(1)
,π
,π
e∈E
(t` )
(t` )
,α
(1)
,α
(1)
,β
,β
(t` )
(1)
(t` )
,
(1)
,
)
since the other conditions are immediate by continuity. In
order to verify (11c), we need to prove that for all i ∈
supp(π ] )
] 2
πi]
=
|x |
1−p
exp(− 2|α
] |2 ) α]
|x] |2
1−p
exp(− 2|α
] |2 )
α]
+
|x] |2
p
exp(− 2|β
] |2 )
β]
and for any i ∈
/ supp(π ] ) then
|x] |2
exp(− 2|αi] |2 ) 1−p
α]
).
|x] |2
1−p
exp(− 2|αi] |2 )
α]
From Step 8 and Step 7 in Algorithm 1 we get
+
|x] |
p
exp(− 2|βi] |2 )
β]
|x] |2
lim α(t` ) = 0,
`→∞
≤
lim (t` ) = 0,
`→∞
lim κ(t` ) − 1 = 0.
(21)
exp(− 2|αj] |2 ) 1−p
α]
|x] |
1−p
exp(− 2|αj] |2 )
α]
+
(22)
|x] |
p
exp(− |β ]j|2 )
β]
for all j ∈ supp(π ] ).
`→∞
We deduce that there exists `0 ∈ N such that κ(t` ) = 1 for
all ` > `0 . Since
v
uP
(t )
u
(1 − πe ` )|be − (Ax(t` ) )e |2
(t` )
,
α
= t e∈E P
(t` )
e∈E (1 − πe )
Since lim`→∞ π (t` ) = π ] , then there exists `0 such that,
(t )
∀` > `0 and ∀i ∈ supp(π ] ), πi ` 6= 0 and
we have that if there exist > 0 and {t`j }j∈N such that
|be − (Ax(t`j ) )| > then π (t`j ) → 1 as j → ∞. On the
other hand, if |be − (Ax(t` ) )e | → 0 then π (t` ) → 0. This
means that, defined the following set
and (21) is verified.
and so does x(t) by asymptotic regularity. Then β (t) must
be bounded, contradiction.
Since both (α(t) )t∈N and (β (t) )t∈N are upper bounded by
a constant χ > 0, then
X
X
0 ≤ kb − Ax(t) k22 =
(1 − πe )α(t) +
πe β (t) ≤ χ|E|,
We conclude that for all i ∈
/ supp(π ∞ )
|x] |2 1−p
exp(− 2|α
] |2 ) α]
]
.
πi = 0 = PN −1 1−p
|x] |2
|x] |2
p
exp(−
)
+
exp(−
)
]
]
2
]
]
2
α
2|α |
β
2|β |
(t )
πi `
=
|x] |2 1−p
exp(− 2|α
] |2 ) α]
(t` ) `→∞
ξi
→ 1−p
]
2
|x |
|x] |2
p
exp(− 2|α
] |2 ) + β ] exp(− 2|β ] |2 )
α]
= πi]
If i ∈
/ supp(π ] ), then we have to distinguish the following
(t )
(t )
two cases: either (a) πi ` is zero eventually or (b) πi `
converges to zero asymptotically. In case (a), there exists
(t )
(t`j )
that ∀` > `0 , πi ` = 0, from which
∆ = {e ∈ E : ∃ > 0 and (t`j )j for which |be −(Ax
)| > },`0(t )∈ N (tsuch
)
ξi ` < ξj ` , ∀j ∈ supp(π ] ) and (22) is satisfied. In case
(
1 if e ∈ ∆
(b), there exists a strictly positive sub-sequence (`q )q∈N such
lim π (t`j ) =
(t` )
(t` )
(t` )
j→∞
0 otherwise
that πi q = ξi q → πi] = 0. Since at the same time ξj q
]
]
), there exists q0 ∈ N
It should be noticed that |∆c | ≥ N − 1 and, together with converges to πi > 0 for all j ∈(tsupp(π
(t`q )
`q )
(t`j )
such
that
∀q
>
q
we
have
ξ
<
ξ
, and, by letting
0
κ
= 1, we conclude
i
j
q
→
∞,
lim A> W (t`j ) Ax(t`j ) = lim A> W (t`j ) b
|x] |2
j→∞
j→∞
exp(− 2|αi] |2 ) 1−p
α]
(t`j )
A>
= lim A>
] 2
∆c A∆c lim x
∆ c b∆ c
|
|x
|x] |
1−p
j→∞
j→∞
exp(− 2|αi] |2 ) + βp] exp(− 2|βi] |2 )
α]
(23)
and, consequently, (x(t`j ) )j∈N converges
|x] |2
exp(− 2|αj] |2 ) 1−p
]
α
† >
≤
lim x(t`j ) = (A>
|x]j |
|x] |
∆c A∆c ) A∆c b∆c .
1−p
p
j→∞
exp(− 2|α] |2 ) + β ] exp(− |β ]j|2 )
α]
e∈E
i
e∈E
which guarantees that x(t) is bounded as well.
Lemma 10: Any accumulation point of ζ (t) is a fixed
point of the algorithm and satisfies the equalities (11a)-(11e).
Proof: If (x] , π ] , α] , β ] , ] ) is an accumulation point of
the sequence (x(t) , π (t) , α(t) , β (t) , (t) )t∈N , then there exists
a subsequence (x(t` ) , π (t` ) , α(t` ) , β (t` ) , (t` ) )`∈N that converges to (x] , π ] , α] , β ] , ] ) as ` → ∞. We now show (11c),
Since the sequence (x(t) ) is bounded (see Lemma 9), there
exists a subsequence (xtj )j∈N such that limj→∞ x(tj ) = x∞ ,
limj→∞ α(tj ) = α∞ , limj→∞ β (tj ) = β ∞ , limj→∞ ξ (tj ) =
π ∞ , and limj→∞ π (tj ) = π ∞ . From Lemma 8, we get
limt→∞ x(tj +1) = limt→∞ x(tj ) = x∞ , proving convergence. Finally, Lemma 10 ensures that ζ (t) converges to a
fixed point.
Since
C. Proof of Theorem 6
In this section we prove Theorem 6, showing convergence
of Distributed LS-EM (Algorithm 2).
Let us consider the function V : RV ×[0, 1]E → R defined
from (8) by fixing the variables α and β, together with a
surrogate function V S : RV × RV × [0, 1]E → R
π (t+1) = argmin V (x(t+1) , π)
π
then we have V (x(t+1) , π (t+1) ) ≤ V (x(t+1) , π (t) ) and,
combining with (25) and (26), we get
1
(1 − τ µ)kx(t) − x(t+1) k2
2τ
T h
i
X
≤
V S (x(t+1) , x(t) , π (t) ) − V (x(t+1) , π (t+1) )
0≤
1
(x−z)> (I −τ LW )(x−z), (24)
2τ
π
1−π
+
W = diag
α2
β2
V S (x, z, π) = V (x, π)+
where
t=1
≤
We let the reader verify the following lemma.
Lemma 11 (Partial minimizations): If
x
b = argmin V S (x, z, π)
and
x∈RV
π
b
=
= argmin V (x, π),
π∈[0,1]E
t=1
|be −(Ax)e |2
∀e ∈ E
1
(1 − τ µ)kx(t) − x(t+1) k2
2τ
≤ V (x(1) , π (1) ) − V (x(T +1) , π (T +1) )
0≤
−
2α2
+ (1−p)
α e
Proposition 12: The function V defined in this section is
nonincreasing along the iterates ζ (t) = (x(t) , π (t) ).
Proof: From Lemma 11 we have
p
βe
V (ζ (t+1) ) = V (x(t+1) , π (t+1) )
≤ V (x(t+1) , π (t) )
= V S (x(t+1) , x(t+1) , π (t) )
≤ V S (x(t+1) , x(t) , π (t) )
≤ V S (x(t) , x(t) , π (t) )
= V (x(t) , π (t) ) = V (ζ (t) )
for every time t ∈ N, thereby proving the result.
The following lemma implies that the algorithm converges
numerically.
Lemma 13: If x(t) is the sequence generated by Algorithm 2, then x(t+1) − x(t) → 0 as t → ∞.
Proof: Define µ = maxt kA> W (t) Ak ≤ kAk2 /α.
Since from assumption τ < α/kAk2 < µ−1 we have
1
(1 − τ µ)kx(t) − x(t+1) k2
2τ
1
≤
(1 − τ kA> W (t) Ak)kx(t) − x(t+1) k2
2τ
1 (t)
≤
(x − x(t+1) )> (I − τ A> W (t) A)(x(t) − x(t+1) ).
2τ
(25)
0≤
If we take the sum until T
0≤
=
T
X
1 (t)
(x − x(t+1) )> (I − τ A> W (t) A)(x(t) − x(t+1) )
2τ
t=1
T h
X
i
cated series is telescopic, from which
|be −(Ax)e |2
2β 2
|b −(Ax) |2
− e 2β 2 e
V (x(t) , π (t) ) − V (x(t+1) , π (t+1) )
i
x
x
b = (I − τ LW )z + τ A> W b
π
be =
t=1
T h
X
V S (x(t) , x(t) , π (t) ) − V (x(t+1) , π (t+1) )
where the last inequality follows from the fact x(t+1) =
argmin V S (x, x(t) , π (t) ). Finally, we observe that the trun-
then
p −
βe
T h
X
V S (x(t+1) , x(t) , π (t) ) − V (x(t+1) , π (t) )
i
t=1
(26)
≤ V (x(1) , π (1) )
− λ(N log max{p/β, (1 − p)/α} − N log 2) = C 0
This last inequality holds for any T ∈ N, then by letting
T → ∞, we obtain that the series is convergent, from which
we deduce that as t → ∞
1 (t)
(x − x(t+1) )> (I − τ A> W (t) A)(x(t) − x(t+1) ) → 0
2τ
and by inequality (25) the claim is proved.
Lemma 14: The sequence (x(t) )t∈N is bounded.
Proof: Since 1> x(t) = 0 for all t ∈ N, from Lemma 11
we have
kx(t+1) k
1
11> x(t) + τ A> W (t) bk
N
1
>
(t)
≤ k(I − τ A W A) − 11> kkx(t) k + kτ A> W (t) bk
N
≤ (1 − τ µ2 )kx(t) k + τ γ
= k(I − τ A> W (t) A) I −
where µ2 = mint kA> W (t) A − N1 11> k > 0 and γ =
maxt kA> W (t) bk (notice that W (t) belongs to a finite set
of matrices). We conclude that
lim kx(t+1) k ≤ lim (1 − τ µ2 )t+1 kx(0) k +
t→∞
t→∞
∞
X
(1 − τ µ2 )s γτ
s=0
γ
≤
µ2
By Lemma 14 the sequence (x(t) ) is bounded and
then there exists a subsequence (xtj )j∈N such that
limj→∞ x(tj ) = x∞ , limj→∞ α(tj ) = α∞ , limj→∞ β (tj ) =
β ∞ , and limj→∞ π (tj ) = π ∞ . From Lemma 13, we get
limt→∞ x(tj +1) = limt→∞ x(tj ) = x∞ Since πe (x) =
f (ze = 1|x, α, β) is a continuous function of x, then also
π (t) → π ∞ and (x∞ , π ∞ ) is a fixed point.
| 3cs.SY
|
23 COMPUTATIONAL TOPOLOGY
OF GRAPHS ON SURFACES
Éric Colin de Verdière
arXiv:1702.05358v2 [cs.CG] 5 Sep 2017
INTRODUCTION
This chapter surveys computational topology results in the special, low-dimensional case
where the ambient space is a surface. Surface topology is very well-understood and comparably simpler than the higher-dimensional counterparts; many computational problems
that are undecidable in general (e.g., homotopy questions) can be solved efficiently on
surfaces. This leads to a distinct flavor of computational topology and to dedicated techniques for revisiting topological problems on surfaces from a computational viewpoint.
Topological surfaces and graphs drawn on them appear in various fields of mathematics and computer science, and these aspects are not surveyed here:
• in topology of three-dimensional manifolds, also in connection to the recent resolution of the Poincaré conjecture, combinatorial and algebraic structures defined
on surfaces are often relevant, e.g., via the study of mapping class groups and
Teichmüller spaces [FM11];
• in topological graph theory, a branch of structural graph theory, graphs on surfaces
are studied from a combinatorial point of view, also in relation to the theory of
Robertson and Seymour on graph minors; for example, colorability questions of
graphs on surfaces, generalizing the four-color theorem for planar graphs, are wellstudied [MT01];
• in enumerative combinatorics, a natural problem is to count (exactly or asymptotically) maps with given properties in the plane or on surfaces, with the help
of generating series; moreover, typical properties of random maps are investigated [Mie09, Bet12, LZ04];
• various applications involve surface meshes, in particular in geometry processing and computer graphics, for approximation [CDP04], topological simplification [GW01, WHDS04], compression [AG05], and parameterization [GY03]. Techniques for general surfaces apply also to subsets of the plane, and are thus relevant
in VLSI design [LM85] and map simplification [BKS98].
This chapter is organized as follows. We first review the basic concepts and properties
of topological surfaces and graphs embedded on them (Sections 23.1 and 23.2). Then we
consider three categories of topological problems, mostly from a computational perspective: drawing an abstract input graph on a surface (Section 23.3), homotopy questions
and variations (Section 23.4), and optimization of curves and graphs on surfaces, also
Author’s affiliation: CNRS, LIGM, Université Paris-Est Marne-la-Vallée, France.
Email:
[email protected]. Part of this work was done while the author was at CNRS,
Département d’Informatique, École normale supérieure, Paris, France.
2
É. Colin de Verdière
from a homological point of view (Section 23.5). Then we survey techniques that allow us
to solve general graph problems faster in the case where the input graph is embedded on a
fixed surface (Section 23.6). Finally, we collect other miscellaneous results (Section 23.7).
23.1 SURFACES
Surfaces are considered from a topological point of view: Two homeomorphic surfaces
are regarded as equivalent. Surfaces such as the sphere or the disk are topologically
uninteresting; our focus is on surfaces in which some closed curves are non-contractible
(they cannot be deformed to a point by a continuous motion on the surface).
GLOSSARY
Homeomorphism: Given two topological spaces X and X 0 , a map h : X → X 0
is a homeomorphism if h is bijective and both h and its inverse are continuous.
Surface (topological definition): In this chapter, a surface S is a compact
two-dimensional manifold possibly with boundary. Equivalently, S is a compact
topological space that is Hausdorff (any two distinct points have disjoint neighborhoods) and such that every point has a neighborhood homeomorphic to the plane
or the closed half-plane. The set of points of a surface S that have no neighborhood
homeomorphic to the plane is the boundary of S.
Surface (combinatorial definition): Equivalently, a surface S is a topological
space obtained from finitely many disjoint triangles by identifying some pairs of
edges of the triangles (by the quotient topology). The boundary of S is the union
of the edges that are not identified with any other edge.
Path: A path on S is a continuous map p : [0, 1] → S. Its two endpoints are
p(0) and p(1).
Connectedness: A surface is connected if any two points of the surface are the
endpoints of some path. The inclusionwise maximal connected subsets of a surface
form its connected components.
Orientability: A surface is non-orientable if some subset of it (with the induced topology) is homeomorphic to the Möbius strip (defined in Figure 23.1).
Otherwise, it is orientable.
PROPERTIES: CLASSIFICATION OF SURFACES
Every connected surface is homeomorphic to exactly one of the following surfaces:
• the orientable surface of genus g ≥ 0 with b ≥ 0 boundary components (or,
more concisely, boundaries), obtained from the sphere by removing g disjoint
open disks, attaching a handle (defined in Figure 23.1) to each of the resulting
g circles, and finally removing b open disks with disjoint closures;
Chapter 23: Computational topology of graphs on surfaces
disk (orientable, g =
0, b = 1)
sphere
entable,
b = 0)
S2
(orig = 0,
Möbius strip (nonorientable, g = 1,
b = 1)
3
handle (orientable,
g = 1, b = 1)
a
a
a
b
a
a
a
b
a
a
b
a
Figure 23.1: Examples of surfaces. Each surface (top row) comes with a polygonal
schema (bottom row), a polygon with some labeled and directed edges; the surface can be
obtained by identifying the pairs of edges with the same labels, respecting their direction.
The genus g and number of boundary components b are specified, as well as whether the
surface is orientable.
• the non-orientable surface of genus g ≥ 1 with b ≥ 0 boundary components
(or boundaries), obtained from the sphere by removing g disjoint open disks,
attaching a Möbius strip (defined in Figure 23.1) to each of the resulting g circles,
and finally removing b open disks with disjoint closures.
Every surface can be obtained by identifying pairs of edges of disjoint triangles. More
concisely, every surface can be defined by a polygonal schema, a polygon with labels and
directions on some of the edges specifying how they must be identified. In particular, one
can define a canonical polygonal schema for every connected surface without boundary:
• The canonical polygonal schema of the orientable surface of genus g ≥ 1 is a 4g-gon
whose successive edges are labeled a1 , b1 , ā1 , b̄1 , . . . , ag , bg , āg , b̄g , and where edge x
is directed clockwise, edge x̄ is directed counterclockwise. Identifying edge x with
x̄, as indicated by their directions, gives the orientable surface of genus g. See
Figure 23.2.
• Similarly, the canonical polygonal schema of the non-orientable surface of genus
g ≥ 1 is a 2g-gon whose successive edges are labeled a1 , a1 , . . . , ag , ag , and where
all edges are directed clockwise.
Figure 23.2: A double torus with a system of loops (left); the surface cut along the loops
(middle) is a disk, shown in the form of a (canonical) polygonal schema (right).
4
É. Colin de Verdière
Surface
Sphere
Disk
Annulus = cylinder
Pair of pants
Torus
Handle
Double torus
Projective plane
Möbius strip
Klein bottle
Orientable?
Yes
Yes
Yes
Yes
Yes
Yes
Yes
No
No
No
Genus
0
0
0
0
1
1
2
1
1
2
# of boundary components
0
1
2
3
0
1
0
0
1
0
Table 23.1: Some common surfaces.
EXAMPLES
Table 23.1 lists some common connected surfaces. See also Figures 23.1 and 23.2.
23.2 GRAPHS ON SURFACES
GLOSSARY
Let S be a surface.
Loop: A loop is a path whose two endpoints are equal to a single point, called
the basepoint of the loop.
Closed curve: A closed curve on S is a continuous map from the unit circle S1
to S. This is almost the same as a loop, except that a closed curve has no distinguished basepoint. A closed curve is sometimes called a cycle, although, contrary
to the standard terminology in graph theory, here a cycle may self-intersect.
Curve: A curve is either a path or a closed curve. For most purposes, the parameterization is unimportant; for example, a path p could be regarded as equivalent
to p ◦ ϕ, where ϕ : [0, 1] → [0, 1] is bijective and increasing.
Simplicity: A path or a closed curve is simple if it is injective. A loop ` : [0, 1] →
S is simple if its restriction to [0, 1) is injective.
Graph: In this chapter, unless specified otherwise, graphs are finite, undirected,
and may have loops and multiple edges.
Curve (in a graph): A curve in a graph G (also called walk in the terminology
of graph theory) is a sequence of directed edges e1 , . . . , ek of G where the target
of ei equals the source of ei+1 . Repetitions of vertices and edges are allowed. The
endpoints of the curve are the source of e1 and the target of ek . If they are equal,
the curve is closed.
Graph embedding (topological definition): A graph G naturally leads to a
topological space Ĝ, defined as follows: One considers a disjoint set of segments,
Chapter 23: Computational topology of graphs on surfaces
5
one per edge of G, and identifies the endpoints that correspond to the same vertex
of G. This gives a topological space, from which Ĝ is obtained by adding one
isolated point per isolated vertex of G. (As a special case, if G has no loop and
no multiple edge, then G is a one-dimensional simplicial complex, and Ĝ is the
associated topological space.) An embedding of G is a continuous map from Ĝ
into S that is a homeomorphism from Ĝ onto its image.
Graph embedding (concrete definition): Equivalently, an embedding of G
on S is a “crossing-free” drawing of G: It maps the vertices of G to distinct points
of S, and its edges to paths of S whose endpoints are the images of their incident
vertices; the image of an edge can self-intersect, or intersect the image of another
edge or vertex, only at its endpoints. When no confusion arises, we identify G with
its embedding on S, or with the image of that embedding.
Face: The faces of an embedded graph G are the connected components of the
complement of the image of G.
Degree: The degree of a vertex v is the number of edges incident to v, counted
with multiplicity (if an edge is a loop). The degree of a face f is the number of
edges incident to f , counted with multiplicity (if an edge has the same face on both
sides).
Cellular embedding:
morphic to open disks.
A graph embedding is cellular if its faces are homeo-
Triangulation: A graph embedding is a triangulation if it is cellular and its faces
have degree three. The triangulation may fail to be a simplicial complex: A triangle
is not necessarily incident to three distinct vertices, or even to three distinct edges.
Cutting: Given an embedded graph G on S without isolated vertex, the operation
of cutting S along G results in a (possibly disconnected) surface with boundary,
denoted S\\G (or sometimes SQG or similar); each connected component of S\\G
corresponds to a face of G on S, and by identifying pieces of the boundaries of
these components in the obvious way, one recovers the surface S. Similarly, one
can cut along a set of disjoint, simple closed curves. (Technically, if S has nonempty boundary, an additional condition is needed: The intersection of an edge
with the boundary of S can be either the entire edge, its two endpoints, or one of
its endpoints.)
Planarity: A graph is planar if it has an embedding to the plane (or equivalently
the sphere).
Dual graph: A dual graph of a cellularly embedded graph G on S (assumed
without boundary) is a graph G∗ embedded on S with one vertex f ∗ inside each
face f of S, and with an edge e∗ for each edge e of G, such that e∗ crosses e and
no other edge of G. A dual graph is cellularly embedded. Its combinatorial map
(see below) is uniquely determined by the combinatorial map of G.
Euler genus: The Euler genus ḡ of a connected surface S with genus g equals 2g
if S is orientable, and g if S is non-orientable.
Euler characteristic: The Euler characteristic of a cellularly embedded graph G
equals χ(G) := v − e + f , where v, e, and f are its number of vertices, edges, and
faces, respectively.
6
É. Colin de Verdière
PROPERTIES: EULER’S FORMULA AND CONSEQUENCES
1. Euler’s formula: If G is cellularly embedded on a connected surface S of Euler
genus ḡ with b boundary components, then χ(G) = 2−ḡ−b. In particular, χ(G) does
not depend on G, only on S, and is consequently called the Euler characteristic
of S.
2. The number of vertices and faces of a graph G cellularly embedded on a connected
surface is at most linear in the number of its edges. In particular, the combinatorial
complexity of G is linear in its number of edges.
3. Conversely, let G be a (not necessarily cellular) graph embedding on a connected
surface with Euler genus ḡ and b boundaries. Assume that G has no face of degree
one or two that is an open disk. Then the numbers e of edges and v of vertices
of G satisfy e = O(v + ḡ + b).
DATA STRUCTURES
In all the problems we shall consider, the exact embedding of a graph on a surface is
irrelevant; only the actual combinatorial data associated to the embedding is meaningful.
If G is a graph cellularly embedded on a surface S without boundary, we only need the
information of G together with the facial walks, namely, the closed walks in G encountered
when walking along the boundary of the faces of G. This information is called the
combinatorial map of G, and allows us to reconstruct the surface, by attaching disks to
every facial walk. (Some conditions on the walks are needed to ensure that the resulting
space is indeed a surface.) If S has boundaries, one can specify the corresponding faces
of G. If S is orientable and G has no loop edge, instead of the facial walks one could as
well specify the cyclic ordering of the edges incident to each vertex.
However, more complicated data structures are needed to perform basic operations
efficiently. For example, one should be able to compute the degree of a face in time linear
in the degree; to count the number of faces of G in linear time; to determine whether the
surface is orientable in linear time; etc. (The last two operations, together with counting
the number of vertices and edges, allow us to identify the topology of the surface in linear
time using Euler’s formula; this can also be done in logarithmic space [BEKT16].)
Figure 23.3: The graph-encoded map data structure. Each edge bears four flags, drawn
parallel to it. Three operations allow us to move from a flag to a nearby flag.
Chapter 23: Computational topology of graphs on surfaces
7
One such data structure, the graph-encoded map or gem representation [Lin82], uses
flags (quarter-edges, or, equivalently, incidences between a vertex, an edge, and a face)
of G, see Figure 23.3; three involutive operations can be applied to a flag to move to an
incident flag. Alternative data structures have been designed for more general situations
(e.g., to allow surfaces with boundaries) or to take advantage of special situations (e.g.,
in the case where G is a triangulation, or where S is orientable); see the survey [Ket99].
However, the choice of the data structure is irrelevant for the theoretical design and
asymptotic analysis of the algorithms.
CONVENTIONS FOR THIS CHAPTER
Henceforth, we assume all surfaces to be connected.
In several works mentioned in the following, only orientable surfaces are considered.
In some cases, non-orientable surfaces are just as easy to handle, but sometimes they lead
to additional difficulties. We refer to the original articles to determine whether
the results hold on non-orientable surfaces.
Also, in most problems studied in this chapter, surfaces with boundaries are no harder
to handle than surfaces without boundary: Any algorithm for surfaces without boundary
immediately implies an algorithm for surfaces with boundary (with the same running
time, or by replacing g by g + b in the complexity, where g and b are the genus and the
number of boundary components). For this reason, we mostly focus on computational problems for surfaces without boundary.
Finally, when we consider cellularly embedded graphs for algorithmic problems, we
implicitly assume that they are specified in the form of a data structure as
described above (e.g., a graph-encoded map).
23.3 EMBEDDING AND DRAWING GRAPHS ON SURFACES
Being able to build embeddings of a graph on a surface with small genus is important;
almost all algorithms for graphs embeddable on a fixed surface require an embedding of
the input graph (there are a few exceptions [ES14, Kel06, MS12]). We discuss algorithmic
results related to the problem of embedding a graph on a surface, and then consider more
general drawings where crossings are allowed.
EMBEDDING GRAPHS ON SURFACES
Let G be an abstract graph (not embedded on any surface), given, e.g., by the (unordered)
list of the edges incident to every vertex. We assume that G is connected. Let n denote
the combinatorial complexity of G, that is, the total number of vertices and edges of G.
1. General facts: An embedding on an orientable surface with minimum possible genus
is cellular. If G is embeddable on an orientable (resp., non-orientable) surface of
genus g, then it is embeddable on an orientable (resp., non-orientable) surface of
genus g 0 , for every g 0 ≥ g.
8
É. Colin de Verdière
2. General bound: G can be cellularly embedded on some orientable surface with
genus O(n).
3. Planar case: There is an O(n)-time algorithm for deciding embeddability in the
sphere (equivalently, in the plane) [HT74]; also in O(n) time, the graph can be
embedded with straight-line segments in the plane [Sch90a] (see also [NR04, Ch. 4]),
if it has no loop or multiple edge. See Chapter 55 for more results on graph drawing.
4. Time complexity: Given a graph G and a surface S, specified by its Euler genus ḡ
and by whether it is orientable, determining whether G embeds on S is NPhard [Tho89], but can be done in 2poly(ḡ) · n time [KMR08, Moh99] (where poly(ḡ)
is a polynomial in ḡ), which is linear if ḡ is fixed. Such an embedding can be
computed in the same amount of time if it exists.
5. Space complexity: For every fixed ḡ, determining whether an input graph G embeds
on some surface (orientable or not) of Euler genus at most ḡ can be done in space
logarithmic in the input size [EK14].
6. Approximation: Given as input a graph G and an integer ḡ, one can in polynomial
time either correctly report that G embeds on no surface of Euler genus ḡ, or
compute an embedding on some surface of Euler genus ḡ O(1) [KS15].
Except for the planar case, these algorithms are rather complicated, and implementing
them is a real challenge. For example, there seems to be no available implementation of a
polynomial-time algorithm for testing embeddability in the torus, and no publicly available implementation of any algorithm to decide whether a graph embeds on the double
torus; attempts of implementing some known embedding algorithms, even in the simplest
cases, have unveiled some difficulties [MK11]. On the other hand, a recent approach is
promising in practice for graphs of moderate size, using integer linear programming or
Boolean satisfiability reformulations [BCHK16].
In contrast, determining the maximum genus of an orientable surface without boundary on which a graph can be cellularly embedded can be done in polynomial time [FGM88].
There are also results on the embeddability of two-dimensional simplicial complexes on
surfaces [Moh97].
On a less algorithmic side, in the field of topological graph theory, a lot more is known
about the embeddability of some classes of graphs on some surfaces; see, e.g., [Arc96,
Sect. 4.2] and references therein.
GLOSSARY ON DRAWINGS
Let G be a graph and S be a surface.
Drawing: Drawings are more general than embeddings in that they allow a finite
set of crossing points, where exactly two pieces of edges intersect and actually cross.
Formally, recall that G has an associated topological space Ĝ. A (topological)
drawing of G on S is a continuous map from Ĝ into S such that the preimage
of every point in S has cardinality zero or one, except for a finite set of points
(“crossings”), whose preimages have cardinality two; moreover, each such crossing
point has a disk neighborhood that contains exactly the images of two pieces of
edges of Ĝ, which form, up to homeomorphism, two crossing straight lines.
Chapter 23: Computational topology of graphs on surfaces
9
Arrangement: Let D be a drawing of G on S. The arrangement of D on S is
the graph G0 embedded on S that has the same image as D and is obtained from D
by inserting a vertex of degree four at each crossing in D and subdividing the edges
of G accordingly. Similarly, one can consider the arrangement of a set of curves
drawn on S.
Crossing number: The crossing number of G with respect to S is the minimum
number of crossings that G has in any drawing of G on S.
Pair crossing number:
The pair crossing number of G with respect to S is
the minimum number of pairs of edges of G that cross, over all drawings of G on S.
Odd crossing number: The odd crossing number of G with respect to S is the
minimum number of pairs of edges of G that cross an odd number of times, over
all drawings of G on S.
DRAWING GRAPHS ON SURFACES WITH FEW CROSSINGS
1. Crossing numbers: Computing the planar crossing number of a graph is NP-hard
(even in very special cases, such as that of a planar graph with a single additional
edge [CM13]), and there exists no polynomial-time algorithm with approximation
guarantee better than a certain constant [Cab13]. However, for every fixed k, one
can, in linear time, determine whether an input graph has planar crossing number
at most k [KR07], although the problem admits no polynomial kernel [HD16]. Some
approximation algorithms for the planar crossing number are known in restricted
cases, such as bounded maximum degree [HC10, Chu11].
2. Variations on crossing numbers: The relations between the various notions of crossing numbers are not fully understood. Let c, p, and o denote the planar crossing
number, planar pair crossing number, and planar odd crossing number, respectively,
of some graph G. It is clear that o ≤ p ≤ c, and it is known that the left inequality
can be strict [PSŠ08]. It is widely believed that p = c, but the best bound known
so far is c = O(p3/2 log2 p) (this follows essentially from [Tót12]). See, e.g., [Mat14]
for more details, and [Sch13a] for a wide survey on the various notions of crossing
numbers.
3. Hanani–Tutte theorem: The (weak) Hanani–Tutte theorem [Han34, Tut70], however, states that if o = 0 then c = 0. Furthermore it holds not only for the plane,
but for arbitrary surfaces [CN00, PSŠ09b]: If a graph G can be drawn on a surface S in a way that every pair of edges crosses an even number of times, then
G can be embedded on S. In the planar case, it actually suffices to assume that
every pair of independent edges (which do not share any endpoints) crosses an even
number of times, but whether this generalizes to arbitrary surfaces is open, except
for the projective plane [PSS09a, CVK+ 16]. We refer to surveys [Sch13b, Sch14]
for more details.
23.4 HOMOTOPY AND ISOTOPY
Most works in computational topology for surfaces do not take as input a given abstract
graph, as in the previous section; instead, they consider an already embedded graph,
10 É. Colin de Verdière
given by its combinatorial map.
GLOSSARY
Let S be a surface.
Reversal: The reversal of a path p : [0, 1] → S is the path p−1 : [0, 1] → S defined
by p−1 (t) = p(1 − t).
Concatenation: The concatenation of two paths p, q : [0, 1] → S with p(1) = q(0)
is the path p · q defined by (p · q)(t) = p(2t) if t ≤ 1/2 and (p · q)(t) = q(2t − 1) if
t ≥ 1/2.
Homotopy for paths: Given two paths p, q : [0, 1] → S, a homotopy between p
and q is a continuous deformation between p and q that keeps the endpoints fixed.
More formally, it is a continuous map h : [0, 1] × [0, 1] → S such that h(0, ·) = p,
h(1, ·) = q, and both h(·, 0) and h(·, 1) are constant maps (equal, respectively,
to p(0) = q(0) and to p(1) = q(1)). The paths p and q are homotopic. Being
homotopic is an equivalence relation, partitioning the paths with given endpoints
into homotopy classes.
Fundamental group: The homotopy classes of loops with a given basepoint form
a group, where concatenation of loops accounts for the multiplication and reversal
accounts for the inverse operation: if [p] denotes the homotopy class of path p, then
we have [p · q] = [p] · [q] and [p−1 ] = [p]−1 .
Homotopy for closed curves (also called free homotopy): Given two closed
curves γ, δ : S1 → S, a homotopy between γ and δ is a continuous deformation
between them, namely, a continuous map h : [0, 1] × S1 → S such that h(0, ·) = γ
and h(1, ·) = δ.
Contractibility: A loop or closed curve is contractible if it is homotopic to a
constant loop or closed curve.
Isotopy: An isotopy between two simple paths, loops, or closed curves is a homotopy h that does not create self-intersections: for each t, h(t, ·) is a simple path,
loop, or closed curve. An isotopy of a graph G is a continuous family of embeddings
of G (the vertices and edges move continuously).
Ambient isotopy: An ambient isotopy of a surface S is a continuous map i :
[0, 1] × S → S such that for each t ∈ [0, 1], i(t, ·) is a homeomorphism.
Minimally crossing: A family of closed curves Γ = (γ1 , . . . , γk ) is minimally
crossing if for every family of closed curves Γ0 = (γ10 , . . . , γk0 ) with γi and γi0 homotopic for each i, the number of intersections and self-intersections in Γ is no larger
than in Γ0 .
Covering space: Let S̃ be a possibly non-compact connected surface. A continuous map π : S̃ → S is a covering map if every point x ∈ S has a connected
neighborhood U such that π −1 (U ) is a disjoint union of open sets (Ui )i∈I and
π|Ui : Ui → U is a homeomorphism for each i. We say that (S̃, π) is a covering
space of S. A lift of a path p is a path p̃ on S̃ such that π ◦ p̃ = p. Finally, if each
Chapter 23: Computational topology of graphs on surfaces
11
loop in S̃ is contractible, then (S̃, π) is a universal covering space of S, which
is essentially unique (precisely: if (S̃, π) and (S̃ 0 , π 0 ) are universal covering spaces,
then there is a homeomorphism τ : S̃ → S̃ 0 such that π = π 0 ◦ τ ).
BASIC PROPERTIES
1. Two paths p and q are homotopic if and only if p · q −1 is a (well-defined and)
contractible loop.
2. Two loops p and q with the same basepoint are freely homotopic (viewed as closed
curves without basepoint) if the homotopy classes of the loops p and q are conjugates
in the fundamental group.
3. The fundamental group of a surface S without boundary of genus g is best understood by looking at a canonical polygonal schema of the surface: If S is orientable,
it is the group generated by 2g generators a1 , b1 , . . . , ag , bg and with a single re−1
−1 −1
lation, a1 b1 a−1
1 b1 . . . ag bg ag bg , corresponding to the boundary of the polygonal
schema. Similarly, if S is non-orientable, it is the group generated by g generators
a1 , . . . , ag and with a single relation, a1 a1 . . . ag ag .
4. The fundamental group of a surface with at least one boundary component is a free
group (because such a surface has the homotopy type of a graph).
5. Let (S̃, π) be a covering space of S. Every path p on S admits lifts on S̃; moreover,
if x̃ is a lift of p(0), then p has a unique lift p̃ such that p̃(0) = x̃. Two paths are
homotopic on S if and only if they have homotopic lifts on S̃. In particular, two
paths are homotopic if they admit lifts with the same endpoints in the universal
covering space.
DECIDING HOMOTOPY AND ISOTOPY
1. Homotopy: One of the first and most studied problems regarding curves on surfaces
is concerned with homotopy tests: (1) The contractibility problem: Is a given closed
curve (or, equivalently here, loop) contractible? (2) The free homotopy problem:
Are two given closed curves (freely) homotopic? These problems translate to central
problems from group theory, in the special case of fundamental groups of surfaces:
Given a finitely generated group, presented in the form of generators and relations,
(1) does a given word in the generators represent the trivial element of the group
(the word problem)? Do two given words in the generators represent conjugate
elements in the group (the conjugacy problem)?
In computational geometry, these problems are studied in the following context:
The input is a cellularly embedded graph G and one or two closed curves in G,
represented as closed walks in G. There exist linear-time (and thus optimal) algorithms for both the contractibility and the free homotopy problems [LR12, EW13].
(An earlier article [DG99] claims the same results, but it is reported [LR12] that
the algorithm for free homotopy in that article has a subtle flaw.) The approaches
rely on the construction of a part of the universal covering space, or on results from
small cancellation theory in group theory [GS90]. We remark that Dehn’s algorithm [Deh12] can be implemented in linear time, but assuming that the surface is
12 É. Colin de Verdière
fixed and that the graph has a single face, which the other algorithms mentioned
above do not require.
2. Isotopy: Deciding whether two simple closed curves are isotopic can also be done in
linear time, because this equivalence relation is a simple refinement of homotopy for
simple closed curves [Eps66]. Deciding isotopy of graph embeddings is more complicated, but can also be done efficiently, since it essentially reduces to homotopy
tests for closed curves [CVM14].
3. Minimum-cost homotopies: Often, when it is known that two curves are homotopic, one would like to compute a “reasonable” homotopy. Relevant questions include finding a homotopy that sweeps the minimum possible area (in a discretized
sense) [CW13], or has the minimum possible number of “steps”; a homotopy in
which the maximum length of the intermediate curves is minimal (“height” of the
homotopy) [CL09]; a homotopy in which the maximum distance traveled by a point
from the first to the second curve is minimal (“width” of the homotopy—this is related to the homotopic Fréchet distance) [HPN+ 16]; etc. Several of these questions
have been studied only in the case of the plane, and extensions to surfaces are still
open.
ELEMENTARY MOVES AND UNCROSSING
Figure 23.4: The four Reidemeister moves, up to ambient isotopy. The pictures represent
the intersection of the union of the curves with a small disk on S; in particular, in these
pictures, the regions bounded by the curves are homeomorphic to disks, and no other
parts of curves intersect the parts of the curves shown.
1. Elementary moves: Every family of closed curves in general position can be made
minimally crossing by a finite sequence of Reidemeister moves, described in Figure 23.4. If a closed curve has k self-crossings, Ω(k 2 ) Reidemeister moves can be
needed; this is tight if the curve is homotopic to a simple curve, but in general no
subexponential upper bound seems to be known [CE16]. Actually, one can deform a
family of curves continuously to make it minimally crossing without increasing the
total number of crossings at any step, and moreover, in a minimally crossing family, each curve is itself minimally self-crossing, and each pair of curves is minimally
crossing [GS97] (see also [HS94a]). There are other characterizations of curves not
in minimally crossing position [HS85].
2. Making curves simple: Let G be a graph cellularly embedded on a surface S. One
can decide whether an input curve, represented by a closed walk in G, is homotopic
to a simple closed curve in S in near-linear time. More generally, one can compute
the minimum number of self-intersections of a curve in S homotopic to an input
closed walk in G, and the minimum number of intersections between two curves in S
respectively homotopic to two input closed walks in G, in quadratic time [DL17].
Chapter 23: Computational topology of graphs on surfaces
13
3. Untangling curves by a homeomorphism: Given two families of disjoint, simple
curves, one can try to minimize the number of crossings between them by changing
one of them by a homeomorphism of the surface; some bounds are known on the
number of crossings that one can achieve [MSTW16].
4. Simultaneous graph drawing: This also relates to the problem of embedding two
input graphs on the same surface in a way that the embeddings cross each other
few times. Here also some results are known [Neg01, RS05, HKMT16]; one can also
require both combinatorial maps to be fixed.
5. Number of homotopy classes: How many simple closed curves in different homotopy
classes can one draw such that they pairwise cross at most k times, for a given
integer k? On orientable surfaces of genus g ≥ 2 without boundary and k = 0, the
answer is 3g − 2 (a pants decomposition, see below, together with a contractible
closed curve). The problem is more interesting for larger values of k; it was recently
proved that, for fixed k, the number of curves one can draw is polynomial in the
genus [Prz15].
23.5 OPTIMIZATION: SHORTEST CURVES AND GRAPHS
The problem of computing shortest curves and graphs satisfying certain topological properties on surfaces has been widely considered. This leads to problems with a flavor of
combinatorial optimization.
For these problems to be meaningful, a metric must be provided. In computational
geometry, one could naturally consider piecewise linear surfaces in some Euclidean space
(perhaps R3 ); however, efficient algorithms for computing shortest paths in such surfaces [MMP87, CH96] need additional assumptions because distances involve square
roots, which leads to deep and unrelated questions on the complexity of comparing sums
of square roots [Blö91]. Furthermore, in the context of graph problems in the specific
case of surface-embedded graphs (Section 23.7 below), that model would be insufficient.
The notions of combinatorial and cross-metric surfaces, defined below, have been developed to avoid these technical distractions, and are suitable in various settings. On the
other hand, with an oracle for shortest path computations, several of the results in this
section extend to more geometric settings, for example piecewise linear surfaces in some
Euclidean space (see, e.g., [EW05, Sect. 3.6]).
GLOSSARY
Discrete metrics on surfaces
Combinatorial surface: A combinatorial surface is the data of a cellular graph
embedding G, with positive weights on the edges. The only allowed curves are walks
in G; the length of a curve is the sum of the weights of the edges of G traversed by
the curve, counted with multiplicity. Algorithmically, curves are stored as closed
walks in G. The complexity of the combinatorial surface is the complexity of the
embedding G (asymptotically, its number of edges).
14 É. Colin de Verdière
Cross-metric surface: A cross-metric surface [CVE10] is also the data of a
cellular graph embedding G on some surface S, with positive weights on the edges.
However, in contrast to the combinatorial surface model, here the curves are drawn
on the surface S in general position with respect to G; the length of a curve is the
sum of the weights of the edges of G crossed by the curve, counted with multiplicity.
Algorithmically, a family of curves (or a graph) on a cross-metric surface is stored
by the combinatorial map of the arrangement of that family of curves (or graph)
together with G. The complexity of the cross-metric surface is the complexity of
the embedding G (asymptotically, its number of edges).
Without loss of generality, one could draw the curves in a neighborhood of the dual
graph G∗ of G. Pushing them completely onto G∗ would transform them into curves
on the combinatorial surface defined by G∗ . However, the cross-metric surface
defined by G retains more information than the combinatorial surface defined by G∗ :
In the latter case, when curves share edges of G∗ , they automatically overlap; the
cross-metric model allows us to make them disjoint except at some well-defined
crossing points. (We should point out that it is still possible to define the notion of
crossing between two curves in a combinatorial surface, but this is still insufficient
for some of the algorithms described below.)
n
d
n
s
s
Figure 23.5: Left: Some closed curves on surfaces, (d) disk-bounding, (n) non-separating,
(s) splitting. Right: A pants decomposition of a surface.
Types of simple closed curves
Let γ be a simple closed curve in the interior of a surface S. See Figure 23.5.
Disk-bounding curve: γ is disk-bounding if the surface S cut along γ (denoted
by S\\γ) has two connected components, one of which is homeomorphic to the disk.
Separating curve: γ is separating if S\\γ has two connected components.
Splitting curve:
Essential curve:
γ is splitting if γ is separating but not disk-bounding.
γ is essential if no component of S\\γ is a disk or an annulus.
Topological decompositions
Cut graph: A cut graph is a graph G embedded on a surface S such that S\\G
is homeomorphic to a closed disk.
System of loops: A system of loops on a surface without boundary is a cut
graph with a single vertex. See Figure 23.2.
Chapter 23: Computational topology of graphs on surfaces
15
Canonical system of loops: A system of loops G on a surface without boundary S is canonical if the edges of the polygon S\\G appear in the same order as in
a canonical polygonal schema (see Section 23.1)
Pants decomposition: A pants decomposition of an orientable surface S is a
family Γ of simple, disjoint closed curves on S such that S\\Γ is a disjoint union of
pairs of pants. See Figure 23.5.
Octagonal decomposition: An octagonal decomposition of an orientable surface S without boundary is a family Γ of closed curves on S such that each (self)intersection point in Γ is a crossing between exactly two closed curves, and each
face of the arrangement of Γ on S is an octagon (a disk with eight sides).
Homology
In the context of graphs on surfaces, one-dimensional homology on surfaces over the field
Z/2Z is used; it can be described somewhat more concisely than more general homology
theories. Let S be a surface. Here we assume graph embeddings to be piecewise linear
(with respect to a fixed triangulation of S).
Homological sum: By the previous assumption, the closure of the symmetric
difference of the images of two graph embeddings G and G0 is the image of some
graph embedding G00 , called the homological sum of G and G0 . (G00 is defined up to
subdivision of edges with degree-two vertices, insertion of isolated vertices, and the
reverse operations; here, graph embeddings are considered up to such operations.)
Homology cycle: A graph G embedded on S is a homology cycle if every vertex
of G has even degree. The set of homology cycles forms a vector space over the
field Z/2Z: The empty graph is the trivial element and addition is the homological
sum.
Homology boundary: A graph G embedded on S is a homology boundary if the
faces of G can be colored in two colors, say black and white, such that G is the
“boundary” between the two colors: Exactly one side of each edge of G is incident
to a black face. The set of homology boundaries forms a vector space over Z/2Z.
Every homology boundary is a homology cycle.
Homology group: It is the Z/2Z-vector space, denoted by H1 (S), that is the
quotient of the homology cycles by the homology boundaries. A graph embedding
is homologically trivial if it is a homology boundary.
The homology of sets of loops or closed curves can be defined similarly, because these loops
and closed curves are the images of some graph embedding. Using the more advanced
theory of singular homology one can remove the restriction of dealing with piecewiselinear graph embeddings.
BASIC PROPERTIES
1. A simple closed curve is disk-bounding if and only if it is contractible.
2. A simple closed curve is separating if and only if it is homologically trivial.
16 É. Colin de Verdière
3. The homology group of a surface S without boundary has dimension ḡ, the Euler
genus of S, and is generated by the loops appearing on the boundary of a canonical
polygonal schema.
SHORTEST CURVES
U
N
W
EI
G
H
T
ED
W
EI
G
H
T
ED
Deciding whether a simple closed curve in a cross-metric (or combinatorial) surface is
separating or disk-bounding can be done in time linear in the size of the data structure
used to store the cellular graph and the curve; this boils down to determining whether
some graph is connected, or whether some surface is a disk (which is easy using Euler’s
formula). Here we consider the optimization version, by looking for shortest curves with
a given topological type in a combinatorial or cross-metric surface. Non-disk-bounding
or non-separating curves are of particular interest, because cutting along such a curve
simplifies the topology of a surface. Below we use non-trivial as a shorthand for either
non-disk-bounding or non-separating.
UNDIRECTED
O(n2 log n) [EHP04]
O(g 3/2 n3/2 log n) non-sep
[CM07]
g O(g) n3/2
non-db
g O(g) n log n [Kut06]
O(g 3 n log n) [CC07]
O(g 2 n log n) [CCE13]
g O(g) n log log n [INS+ 11]
2O(g) n log log n [Fox13]
O(gn log n) for 2-approx. [EHP04]
DIRECTED
O(n2 log n) [CCVL16]
O(g 1/2 n3/2 log n) [CCVL16]
2O(g) n log n non-sep [EN11b]
O(g 2 n log n) non-sep
[Eri11]
g O(g) n log n non-db
O(g 3 n log n) non-db [Fox13]
O(n3 ) [Tho90] (see [MT01])
O(n2 ) [CCVL12]
O(gnk) [CCVL12]
O(gn/ε) for (1 + ε)-approx. [CCVL12]
O(n2 ) [CCVL16]
O(gnk) [CCVL16]
Table 23.2: Algorithms for shortest non-trivial closed curves on surfaces without boundary, depending on whether the graph is weighted and whether it is directed. “Non-sep”
and “non-db” mean non-separating and non-disk-bounding, respectively; k is the size of
the output. The best complexities known to date are in bold (there can be several of
them in each category due to the tradeoff between g, n, and k). Of course, the undirected
case reduces to the directed case, and the unweighted case reduces to the weighted case;
in each cell, we do not repeat the algorithms that are available for more general scenarios.
1. Structural properties: In a combinatorial surface, a shortest noncontractible or nonnull-homologous loop based at a vertex x is made of two shortest paths from x and
of a single edge (this is the so-called 3-path condition [Tho90]). It follows that
the globally shortest non-contractible and non-null-homologous closed curves do
not repeat vertices and edges, and are also shortest non-disk-bounding and nonseparating closed curves. More generally, in the algorithms mentioned below, a
typical tool is to prove a bound on the number of crossings between the (unknown)
shortest curve and any shortest path.
Chapter 23: Computational topology of graphs on surfaces
17
2. Different scenarios for shortest non-trivial curves: Table 23.2 summarizes the running times of the known algorithms. In such problems, it is relevant to look for
more efficient algorithms in the case where the genus g is smaller compared to the
complexity n of the graph defining the surface. The standard scenario, which is
the only one considered elsewhere in this chapter, is that of a combinatorial (or
equivalently, cross-metric) surface (the undirected, weighted case, in the upper left
corner in Table 23.5.1). One can also aim for faster algorithms in the unweighted
case (unit weights). Finally, one can extend the techniques to the case of directed
graphs, where the edges of the combinatorial surface are directed and can only be
used in a specified direction (equivalently, the edges of the cross-metric surface can
only be crossed in a specific direction).
3. Other topological types: Shortest simple closed curves of other topological types
have been investigated as well (in the following, n denotes the complexity of
the cross-metric surface): shortest splitting curves [CCV+ 08] (NP-hard, but computable in O(n log n) time for fixed genus); shortest essential curves [EW10] (O(n2 log n)
time, or O(n log n) for fixed genus and number of boundaries—in this case, surfaces
with boundary require more sophisticated techniques); and non-separating curves
which are shortest in their (unspecified) homotopy class [CDEM10] (O(n log n)).
4. Shortest homotopic curves: A slightly different problem is that of computing a
shortest curve homotopic to a given curve (either a path or a closed curve); this is
also doable in small polynomial time, using octagonal decompositions to build a part
of the universal covering space [CVE10] (earlier algorithms dealt with simple curves
only, with an iterated shortening process that leads to a global optimum [CVL05,
CVL07]).
5. Shortest paths: All these algorithms rely on shortest path computations on combinatorial (or cross-metric) surfaces, which can be done in O(n log n) time using
Dijkstra’s algorithm [Dij59] classically speeded up with Fibonacci heaps [FT87] in
the primal (or dual) graph. This actually computes the shortest paths from a single source to all other vertices of the combinatorial surface. Other algorithms are
available for computing multiple shortest paths quickly under some conditions on
the locations of the endpoints [CCE13].
SHORTEST DECOMPOSITIONS
Decompositions of surfaces are central in topology; for example, the standard proof of the
classification theorem transforms an arbitrary cut graph into a canonical system of loops.
Many algorithms described in the previous subsection rely on topological decompositions
and their properties.
1. Shortest cut graph: The problem of computing a shortest cut graph on a crossmetric surface has been extensively studied. Computing the shortest cut graph is
NP-hard, but there is an O(log2 g)-approximation algorithm that runs in O(g 2 n log n)
time [EHP04]. Moreover, for every ε > 0 one can compute a (1 + ε)-approximation
in f (ε, g) · n3 time, for some function f [CAM15]. If one is looking for a shortest cut graph with a specified vertex set P (for example, a shortest system of
loops with given basepoint [EW05]), then there is an algorithm with running time
O(n log n + gn + |P |) [CV10]. At the root of several of these articles lies the treecotree property [Epp03]: If G is a cellular graph embedding, there exists a partition
18 É. Colin de Verdière
(T, C, X) of the edges of G such that T is a spanning tree of G and the edges dual
to C form a spanning tree of the dual graph G∗ . Contracting T and deleting C
transforms G into a system of loops, each loop corresponding to an element of X.
2. Other topological decompositions: Some canonical system of loops (for orientable
surfaces without boundary) can be computed in O(gn) time [LPVV01]. An octagonal decomposition or a pants decomposition made of closed curves which are
as short as possible in their respective homotopy classes can be computed in
O(gn log n) time [CVE10]. But in general the complexity of computing shortest
such decompositions is open. On the other hand, there are bounds on the maximum length of some decompositions, assuming that the combinatorial surface is
an unweighted triangulation, or, dually, that the cross-metric surface is unweighted
and each vertex has degree three [CVHM15].
3. Stretch: Let S be a cross-metric surface, and let G be the associated embedded
graph. The stretch of S is the minimum of the product of the lengths of γ and δ,
over all closed curves γ and δ crossing exactly once. This quantity is related to the
planar crossing number and the size of a largest toroidal grid minor of G∗ [HC10],
and can be computed in small polynomial time [CCH14].
HOMOLOGY AND ITS RELATION TO CUTS AND FLOWS
As hinted above, homology is useful because a simple closed curve is separating if and
only if it is null-homologous; the algorithms for computing shortest non-separating closed
curves actually compute shortest non-null-homologous closed curves, which turn out to
be simple.
Homology is a natural concept; in particular, it is interesting to look for a family
of closed curves, of minimum total length, the homology classes of which generate the
homology group. Some efficient algorithms have been given for this purpose [EW05], also
in connection with an algorithm to compute a minimum cycle basis of a surface-embedded
graph [BCFN16].
Another reason for the importance of homology is its relation to cuts: Given a graph G
cellularly embedded on a surface S without boundary, the (s, t)-cuts in G are dual to
the subgraphs of G∗ in some fixed homology class on the surface obtained from S by
removing the faces of G∗ containing s and t. Thus, computing minimum cuts amounts
to computing shortest homologous subgraphs. This property has been exploited to study
general graph problems, where better algorithms can be designed in the specific case of
graphs embedded on a fixed surface, to:
1. compute minimum (s, t)-cuts in near-linear time [CEN09, EN11b]. The best algorithm runs in 2O(g) n log n time, where g is the genus [EN11b], and relies on the
homology cover , a particular type of covering space;
2. compute maximum (s, t)-flows faster, by exploiting further the duality between
flows and cuts [CEN12, BEN+ 16];
3. count and sample minimum (s, t)-cuts efficiently [CFN14];
4. compute global minimum cuts efficiently (without fixing s and t) [EFN12];
Chapter 23: Computational topology of graphs on surfaces
19
5. deal with other problems, e.g., to compute the edge expansion and other connectivity measures [Pat13] or to bound the space complexity of bipartite matching [DGKT12].
23.6 ALGORITHMS FOR GRAPHS EMBEDDED ON A
FIXED SURFACE
Some general graph problems can be solved faster in the special case of graphs embedded
on a fixed surface. Examples include cut and flow problems (see previous section), multicommodity problems, domination and independence problems, connectivity problems
(Steiner tree, traveling salesman problem, etc.), disjoint paths problems, shortest paths
problems, subgraph problems, and more.
Sometimes the problems are solvable in polynomial-time on arbitrary graphs, and the
goal is to obtain faster algorithms for surface-embedded graphs. But in many cases, the
problems considered are NP-hard on arbitrary graphs, and polynomial-time algorithms
are obtained for graphs embeddable on a fixed surface (occasionally by fixing some other
parameters of the problem). Typically, optimization problems are considered, in which
case it is relevant to look for approximation algorithms.
The methods involved usually combine topological aspects (as described above) with
techniques from structural and algorithmic graph theory.
GLOSSARY
Minor: A graph H is a minor of another graph G if H can be obtained from G
by removing edges and isolated vertices, and contracting edges.
Minor-closed family: A family F of graphs is minor-closed if every minor of a
graph in F is also in F .
Tree decomposition: A tree decomposition of a graph G = (V, E) is a tree T in
which each node is labeled by a subset of V , such that:
– for each v ∈ V , the set of nodes in T whose labels contain v induces a
non-empty connected subtree of T , and
– if G has an edge connecting vertices u and v, then the label of at least one
node of T contains both u and v.
Width: The width of a tree decomposition is the maximum cardinality of the
labels minus one.
Treewidth: The treewidth of a graph G is the minimum width of a tree decomposition of G.
SURVEY OF TECHNIQUES
Central to algorithmic and structural graph theory is the study of minor-closed families
of graphs; by a deep result of Robertson and Seymour [RS04], for each such family F ,
20 É. Colin de Verdière
there is a finite set XF of graphs such that G ∈ F if and only if no graph in XF is a
minor of G. We refer to [KM07] for a survey on these structural aspects.
The graphs embeddable on a fixed surface form a minor-closed family, and have the
benefit that they can be studied using topological techniques. Robertson and Seymour
provide a decomposition theorem for minor-closed families of graphs involving graphs
embeddable on a fixed surface [RS03]; efficient algorithms for surface-embedded graphs
are sometimes extended to minor-closed families of graphs (different from the family of
all graphs).
It is impossible to list all results in algorithms for surface-embedded graphs here, so
we focus on general methods. Several algorithms are based on topological techniques
described in the previous sections (in particular, shortest non-trivial curves or shortest
decompositions), in several cases with advanced algorithmic techniques [EN11a, KKS11,
ES14, PPSL14]. Sometimes the same techniques have led to new results for planar
graphs [Eri10, EN11c, CV17b]. Methods applicable to several algorithmic problems have
also emerged, in many cases extending previous ones invented for planar graphs:
1. Graph separators and treewidth: Let G be a graph with n vertices embedded on
a surface with genus g. In linear time, one can compute a balanced separator
√
√
of size O( gn), namely, a set of O( gn) vertices whose removal leaves a graph
without connected component of more than 2n/3 vertices [GHT84, Epp03]. Also,
√
the treewidth of G is O( gn).
2. Dynamic programming: Small treewidth implies efficient algorithms using dynamic
programming in arbitrary graphs. When the graph is embedded, one can exploit
this fact to obtain algorithms with smaller dependence on the treewidth for some
problems [Bon12, RST13, RST14].
3. Irrelevant vertex technique: Several graph problems enjoy the following property [Thi12]:
If the input graph has large treewidth, there exists an irrelevant vertex, whose removal creates an equivalent instance of the problem (e.g., a vertex at the center of
a large grid minor). This property is widely used in structural graph theory and
has been exploited several times in the context of algorithms for surface-embedded
graphs [KR07, KT12, RS12].
4. Polynomial-time approximation schemes (PTASs): Baker [Bak94] has introduced
a technique for designing approximation schemes for some optimization problems
with local constraints in planar graphs: She has showed that one can delete a small
part of the input graph without changing too much the value of the solution and
such that the resulting graph has small treewidth. The technique has been extended
to graphs embeddable on a fixed surface [Epp00], to graphs that can be drawn on
a fixed surface with a bounded number of crossings per edge [GB07], and to more
general contraction-closed problems where contraction instead of deletion must be
used [Kle05, DHM10]. A crucial step in making the latter technique effective is
the construction of a spanner : In the case of a minimization problem, this is a
subgraph of the input graph containing a near-optimal solution and whose weight
is linear in that of the optimal solution. Brick decomposition is a technique that
builds spanners for some problems, originally in planar graphs, but also sometimes
in graphs on surfaces [BDT14].
5. Bidimensionality: This theory [Thi15, DFHT05, DHT06, DH08] applies to minimization problems on unweighted graphs where contracting an edge of the graph
Chapter 23: Computational topology of graphs on surfaces
21
does not increase the value of the solution, and where the value of the solution
in grid graphs (and generalizations) is large. It leads to output-sensitive algorithms
for graphs embeddable on a fixed surface with running time of the form
√
O(
k)
2
· nO(1) , where k is the value of the solution and n is the input size. This also
provides PTASs in some cases [DH05]. For the problems where bidimensionality
applies, PTASs can sometimes also be obtained in weighted graphs using a different
framework [CACV+ 16].
6. Stochastic embeddings: Let G = (V, E) and G0 = (V 0 , E 0 ) be positively edgeweighted graphs. A non-contracting metric embedding f from G to G0 is a mapping
from V to V 0 such that d0 (f (x), f (y)) ≥ d(x, y) for each x, y ∈ V , where d and d0
represent shortest path distances in G and G0 , respectively. The distortion of f is
the maximum of d0 (f (x), f (y))/d(x, y) over all x 6= y ∈ V (see Chapter 8). Every
graph G embeddable on an orientable surface S of genus g admits a probability
distribution of non-contracting metric embeddings into planar graphs such that for
each x, y ∈ V , one has E[d0 (f (x), f (y))] ≤ O(log g) · d(x, y), where the expectation
is over all f in the distribution [Sid10]. This reduces several optimization problems
on graphs on S to the same problem in planar graphs, up to the loss of an O(log g)
factor. Actually, such a distribution can be computed in polynomial time even if
no embedding of G on S is known [MS12].
23.7 OTHER MODELS
A rather large number of results relate to the concepts described in this chapter, and it
would be impossible to cover them all. Below, we provide a selection of miscellaneous
results that consider other models for representing graphs on surfaces.
COMPUTATIONAL TOPOLOGY IN THE PLANE WITH OBSTACLES
The plane minus finitely many points or polygons (“obstacles”) forms a (non-compact)
surface S. Taking any cellular graph embedding on S makes S a combinatorial (or crossmetric) surface, so most of the topological algorithms above apply. However, it is much
more natural to consider arbitrary piecewise-linear curves in S, whose length is defined
by the Euclidean metric. In this model, S is defined by the obstacles (a finite set of
disjoint simple polygons, for simplicity of exposition); curves are arbitrary polygonal
lines avoiding the interior of the obstacles. Some of the problems defined in the previous
sections and related problems have been studied in this model:
1. Homotopy and isotopy tests: There are efficient algorithms to test whether two
curves are (freely) homotopic [CLMS04], or whether two graphs are isotopic [CVM14].
2. Shortest homotopic paths can be computed efficiently as well [HS94b, Bes03, EKL06];
see also Section 31.2. A variant where several simple and disjoint paths must be
shortened while preserving their homotopy class and keeping their neighborhoods
simple and disjoint (i.e., the paths are “thick”) has also been investigated [GJK+ 88].
3. Shortest disjoint paths: Here the goal is to compute disjoint paths with minimum
total length (or, more precisely, non-crossing paths, since in the limit case, the
22 É. Colin de Verdière
solution may consist of overlapping paths). If the endpoints lie on the boundary of
a bounded number of obstacles, the problem is solvable in polynomial time [EN11c].
4. Other results include a constant-factor approximation algorithm for the shortest
pants decomposition in the case where the obstacles are points [Epp09] and an
algorithm for computing the homotopic Fréchet distance, a measure of similarity between curves that takes the obstacles into account topologically [CCV+ 10,
HPN+ 16].
SIMPLE AND DISJOINT CURVES IN GRAPHS
In the cross-metric model, defined by a cellularly embedded graph G, one can think of
curves as being drawn in a neighborhood of G∗ . So, intuitively, curves are drawn in G∗ ,
but they can share vertices and edges of G∗ while being simple and pairwise disjoint.
It is very natural, especially in topological graph theory, to forbid such overlaps:
A set of disjoint simple curves cannot repeat any vertex or edge of G∗ . Many of the
problems mentioned in the previous sections make sense in this setup, which turns out
to be generally more difficult to handle than the cross-metric model. In this model, the
following results are known (here by circuit we mean a closed curve in the graph without
repeated vertex, and containing at least one edge):
1. Determining whether there exists a separating (resp., splitting) circuit is NPcomplete [CCVL11].
2. Determining some contractible (resp., non-contractible, resp., non-separating) circuit, if such a circuit exists, is possible in linear time, even if one requires the circuit
to pass through a given vertex [CCVL11].
3. Computing a shortest contractible circuit is possible in polynomial time, but if
one requires the circuit to pass through a given vertex, the problem becomes NPhard [Cab10].
4. Computing a shortest separating circuit is NP-hard [Cab10].
5. There is a combinatorial characterization on whether curves can be made simple and
disjoint in the graph by a homotopy on the surface [Sch91]. In the case of a planar
surface with boundaries, this leads to a polynomial-time algorithm [Sch90b, Th. 31],
which in turn has some algorithmic consequences on the problem of computing
vertex-disjoint paths in planar graphs [Sch90b, Th. 34]. See also [Sch03, Ch. 76].
NORMAL CURVES ON SURFACES
Let Γ be a family of disjoint simple closed curves on a surface S in general position with
respect to a triangulation T of S. A natural way to represent Γ, as described in the
previous sections, is by its arrangement with T . Normal curves are a more economical
representation, at the price of a mild condition: For every triangle t of T , the intersection of the image of Γ with t must be a set of (disjoint simple) paths, called normal arcs,
connecting different sides of t. For such a Γ, and for each triangle t of T , one stores three
integers recording the number of normal arcs connecting each of the three pairs of sides
Chapter 23: Computational topology of graphs on surfaces
23
of t. Overall, Γ is described by 3n non-negative integers, where n is the number of triangles in T . Conversely, given a vector of 3n non-negative integers, one can unambiguously
reconstruct Γ up to normal isotopy, that is, up to an ambient isotopy that leaves the
edges of T globally unchanged.
To store the vector of normal coordinates, O(n log(X/n)) bits are needed, where X
is the number of crossing points of Γ with T . In contrast, representing these curves on a
cross-metric surface requires at least to store a constant amount of information per vertex
of the arrangement, which is Θ(n+X) in total. So the normal curve representation can be
exponentially compressed compared to the “naı̈ve” one. Despite this, in time polynomial
in the input size one can:
1. count the number of connected components of a normal curve (note that a “normal
curve” does not have to be connected), and partition these components according
to their (normal or not) isotopy classes, given by their multiplicities and the normal
coordinates of a representative [SSŠ02, EN13];
2. decide whether two normal curves are isotopic [SSŠ02, EN13];
3. compute the algebraic [SSŠ02, EN13] or the geometric [SSŠ08] intersection number
of two normal curves. (The algebraic intersection number of γ and δ is the sum,
over all crossings between γ and δ, of the sign of the crossing, which is +1 if γ
crosses δ from left to right at that crossing point and −1 otherwise; this is welldefined if the surface is orientable, since it is invariant by isotopy. The geometric
intersection number of γ and δ is the minimum number of crossings between curves
γ 0 and δ 0 isotopic to γ and δ.)
These problems have been initially studied using straight-line programs, a concise
encoding of words over a finite alphabet; many algorithms on words can be solved efficiently using the straight-line program representation, in particular because straight-line
programs can represent exponentially long words; this leads to efficient algorithms for
normal curves [SSŠ02, SSŠ08]. The same and other problems have been revisited using
more topological techniques [EN13]. Normal curves are the lower-dimensional analog of
normal surfaces, widely used in three-dimensional topology.
23.8 OTHER RESOURCES
Books. Graphs on surfaces from a combinatorial viewpoint are treated in detail in [MT01];
see also [GT87]. For basic surface topology, we recommend [Arm83, Sti93, Hen94].
Survey. [Eri12] surveys optimization problems for surface-embedded graphs, providing
more details on a large fraction of Section 23.5.
Course notes and unpublished material. [Eri13] provides some notes in computational topology with a strong emphasis on graphs on surfaces. [CV12, CV17a] survey
some algorithms for optimization of graphs and curves on surfaces. [DMST11] emphasizes
graph algorithms for surface-embedded graphs.
24 É. Colin de Verdière
RELATED CHAPTERS
Chapter 20: Polyhedral maps
Chapter 31: Shortest paths and networks
Chapter 55: Graph drawing
ACKNOWLEDGMENTS
Many thanks to Sergio Cabello, Vincent Cohen-Addad, Jeff Erickson, Francis Lazarus,
Arnaud de Mesmay, and Dimitrios Thilikos for their careful reading of preliminary versions, and for their numerous comments that greatly improved this chapter.
REFERENCES
[AG05]
P. Alliez and C. Gotsman. Recent advances in compression of 3D meshes. In N.A.
Dodgson, M.S. Floater, and M.A. Sabin, editors, Advances in Multiresolution for
Geometric Modelling, pages 3–26, Springer-Verlag, Berlin, 2005.
[Arc96]
D. Archdeacon. Topological graph theory. A survey. Congr. Numer., 115:5–54,
1996.
[Arm83]
M.A. Armstrong. Basic Topology. Undergraduate Texts in Mathematics, SpringerVerlag, Berlin, 1983.
[Bak94]
B.S. Baker. Approximation algorithms for NP-complete problems on planar
graphs. J. ACM, 41:153–180, 1994.
[BCFN16]
G. Borradaile, E.W. Chambers, K. Fox, and A. Nayyeri. Minimum cycle and
homology bases of surface embedded graphs. In Proc. 32nd Sympos. Comput.
Geom., vol 51 of LIPIcs, article 23, Schloss Dagstuhl, 2016.
[BCHK16]
S. Beyer, M. Chimani, I. Hedtke, and M. Kotrbčı́k. A practical method for the
minimum genus of a graph: Models and experiments. In Proc. 15th Sympos.
Experimental Algorithms, vol. 9685 of LNCS, pages 75–88, Springer, Cham, 2016.
[BDT14]
G. Borradaile, E.D. Demaine, and S. Tazari. Polynomial-time approximation
schemes for subset-connectivity problems in bounded-genus graphs. Algorithmica,
68:287–311, 2014.
[BEKT16]
B.A. Burton, M. Elder, A. Kalka, and S. Tillmann. 2-manifold recognition is in
logspace. J. Comput. Geom., 7:70–85, 2016.
[BEN+ 16]
G. Borradaile, D. Eppstein, A. Nayyeri, and C. Wulff-Nilsen. All-pairs minimum
cuts in near-linear time for surface-embedded graphs. In Proc. 32nd Sympos.
Comput. Geom., vol. 51 of LIPIcs, article 22, Schloss Dagstuhl, 2016.
[Bes03]
S. Bespamyatnikh. Computing homotopic shortest paths in the plane. J. Algorithms, 49:284–303, 2003.
[Bet12]
J. Bettinelli. The topology of scaling limits of positive genus random quadrangulations. Ann. Probab., 40:1897–1944, 2012.
[BKS98]
M. de Berg, M. van Kreveld, and S. Schirra. Topologically correct subdivision
simplification using the bandwidth criterion. Cartography and GIS, 25:243–257,
1998.
[Blö91]
J. Blömer. Computing sums of radicals in polynomial time. In Proc. 32nd IEEE
Sympos. Found. Comp. Sci., pages 670–677, 1991.
Chapter 23: Computational topology of graphs on surfaces
25
[Bon12]
P. Bonsma. Surface split decompositions and subgraph isomorphism in graphs on
surfaces. In Proc. 29th Sympos. Theoret. Aspects Comp. Sci., vol. 14 of LIPIcs,
pages 531–542, Schloss Dagstuhl, 2012.
[Cab10]
S. Cabello. Finding shortest contractible and shortest separating cycles in embedded graphs. ACM Trans. Algorithms, 6:24, 2010.
[Cab13]
S. Cabello. Hardness of approximation for crossing number. Discrete Comput.
Geom., 49:348–358, 2013.
[CACV+ 16]
V. Cohen-Addad, É. Colin de Verdière, P.N. Klein, C. Mathieu, and D. Meierfrankenfeld. Approximating connectivity domination in weighted bounded-genus
graphs. In Proc. 48th ACM Sympos. Theory Comput., pages 584–597, 2016.
[CAM15]
V. Cohen-Addad and A. de Mesmay. A fixed parameter tractable approximation
scheme for the optimal cut graph of a surface. In Proc. 23rd European Sympos.
Algorithms, vol. 9294 of LNCS, pages 386–398, Springer, Berlin, 2015.
[CC07]
S. Cabello and E.W. Chambers. Multiple source shortest paths in a genus g graph.
In Proc. 18th ACM-SIAM Sympos. Discrete Algorithms, pages 89–97, 2007.
[CCE13]
S. Cabello, E.W. Chambers, and J. Erickson. Multiple-source shortest paths in
embedded graphs. SIAM J. Comput., 42:1542–1571, 2013.
[CCH14]
S. Cabello, M. Chimani, and P. Hliněný. Computing the stretch of an embedded
graph. SIAM J. Discrete Math., 28:1391–1401, 2014.
[CCV+ 08]
E.W. Chambers, É. Colin de Verdière, J. Erickson, F. Lazarus, and K. Whittlesey.
Splitting (complicated) surfaces is hard. Comput. Geom., 41:94–110, 2008.
[CCV+ 10]
E.W. Chambers, É. Colin de Verdière, J. Erickson, S. Lazard, F. Lazarus, and
S. Thite. Homotopic Fréchet distance between curves — or, walking your dog in
the woods in polynomial time. Comput. Geom., 43:295–311, 2010.
[CCVL11]
S. Cabello, É. Colin de Verdière, and F. Lazarus. Finding cycles with topological
properties in embedded graphs. SIAM J. Discete Math., 25:1600–1614, 2011.
[CCVL12]
S. Cabello, É. Colin de Verdière, and F. Lazarus. Algorithms for the edge-width
of an embedded graph. Comput. Geom., 45:215–224, 2012.
[CCVL16]
S. Cabello, É. Colin de Verdière, and F. Lazarus. Finding shortest non-trivial
cycles in directed graphs on surfaces. J. Comput. Geom., 7:123–148, 2016.
[CDEM10]
S. Cabello, M. DeVos, J. Erickson, and B. Mohar. Finding one tight cycle. ACM
Trans. Algorithms, 6:61, 2010.
[CDP04]
S.-W. Cheng, T.K. Dey, and S.-H. Poon. Hierarchy of surface models and irreducible triangulations. Comput. Geom., 27:135–150, 2004.
[CE16]
H.-C. Chang and J. Erickson. Untangling planar curves. In Proc. 32nd Sympos.
Comput. Geom., vol. 51 of LIPIcs, article 29, Schloss Dagstuhl, 2016.
[CEN09]
E.W. Chambers, J. Erickson, and A. Nayyeri. Minimum cuts and shortest homologous cycles. In Proc. 25th Sympos. Comput. Geom., pages 377–385, ACM Press,
2009.
[CEN12]
E.W. Chambers, J. Erickson, and A. Nayyeri. Homology flows, cohomology cuts.
SIAM J. Comput., 41:1605–1634, 2012.
[CFN14]
E.W. Chambers, K. Fox, and A. Nayyeri. Counting and sampling minimum cuts
in genus g graphs. Discrete Comput. Geom., 52:450–475, 2014.
[CH96]
J. Chen and Y. Han. Shortest paths on a polyhedron. Internat. J. Comput. Geom.
Appl., 6:127–144, 1996.
[Chu11]
J. Chuzhoy. An algorithm for the graph crossing number problem. In Proc. 43rd
ACM Sympos. Theory Comput., pages 303–312, 2011.
26 É. Colin de Verdière
[CL09]
E.W. Chambers and D. Letscher. On the height of a homotopy. In Proc. 21st
Canad. Conf. Comput. Geom., pages 103–106, 2009. See also the erratum at
mathcs.slu.edu/~chambers/papers/hherratum.pdf.
[CLMS04]
S. Cabello, Y. Liu, A. Mantler, and J. Snoeyink. Testing homotopy for paths in
the plane. Discrete Comput. Geom., 31:61–81, 2004.
[CM07]
S. Cabello and B. Mohar. Finding shortest non-separating and non-contractible
cycles for topologically embedded graphs. Discrete Comput. Geom., 37:213–235,
2007.
[CM13]
S. Cabello and B. Mohar. Adding one edge to planar graphs makes crossing number
hard. SIAM J. Comput., 42:1803–1829, 2013.
[CN00]
G. Cairns and Y. Nicolayevsky. Bounds for generalized thrackles. Discrete Comput.
Geom., 23:191–206, 2000.
[CV10]
É. Colin de Verdière. Shortest cut graph of a surface with prescribed vertex set.
In Proc. 18th European Sympos. Algorithms, part 2, vol. 6347 of LNCS, pages
100–111, Springer, Berlin, 2010.
[CV12]
É. Colin de Verdière. Topological algorithms for graphs on surfaces. Habilitation thesis, École normale supérieure, Paris, 2012. Available at http://monge.
univ-mlv.fr/~colinde/pub/12hdr.pdf.
[CV17a]
É. Colin de Verdière. Algorithms for embedded graphs. Course notes, http://
monge.univ-mlv.fr/~colinde/cours/all-algo-embedded-graphs.pdf, 2017.
[CV17b]
É. Colin de Verdière. Multicuts in planar and bounded-genus graphs with bounded
number of terminals. Algorithmica, 78:1206–1224, 2017.
[CVE10]
É. Colin de Verdière and J. Erickson. Tightening nonsimple paths and cycles on
surfaces. SIAM J. Comput., 39:3784–3813, 2010.
[CVHM15]
É. Colin de Verdière, A. Hubard, and A. de Mesmay. Discrete systolic inequalities
and decompositions of triangulated surfaces. Discrete Comput. Geom., 53:587–620,
2015.
[CVK+ 16]
É. Colin de Verdière, V. Kaluža, P. Paták, Z. Patáková, and M. Tancer. A direct
proof of the strong Hanani–Tutte theorem on the projective plane. In Proc. 24th
Sympos. Graph Drawing Network Visualization, pages 454–467, vol. 9801 of LNCS,
Springer, Cham, 2016.
[CVL05]
É. Colin de Verdière and F. Lazarus. Optimal system of loops on an orientable
surface. Discrete Comput. Geom., 33:507–534, 2005.
[CVL07]
É. Colin de Verdière and F. Lazarus. Optimal pants decompositions and shortest
homotopic cycles on an orientable surface. J. ACM, 54:18, 2007.
[CVM14]
É. Colin de Verdière and A. de Mesmay. Testing graph isotopy on surfaces. Discrete
Comput. Geom., 51:171–206, 2014.
[CW13]
E.W. Chambers and Y. Wang. Measuring similarity between curves on 2-manifolds
via homotopy area. In Proc. 29th Sympos. Comput. Geom., pages 425–434, ACM
Press, 2013.
[Deh12]
M. Dehn. Transformation der Kurven auf zweiseitigen Flächen. Math. Ann.,
72:413–421, 1912.
[DFHT05]
E.D. Demaine, F.V. Fomin, M.T. Hajiaghayi, and D.M. Thilikos. Subexponential parameterized algorithms on bounded-genus graphs and H-minor-free graphs.
J. ACM, 52:866–893, 2005.
[DG99]
T.K. Dey and S. Guha. Transforming curves on surfaces. J. Comput. System Sci.,
58:297–325, 1999.
Chapter 23: Computational topology of graphs on surfaces
27
[DGKT12]
S. Datta, A. Gopalan, R. Kulkarni, and R. Tewari. Improved bounds for bipartite
matching on surfaces. In Proc. Sympos. Theoret. Aspects Comp. Sci., vol. 14 of
LIPIcs, pages 254–265, Schloss Dagstuhl, 2012.
[DH05]
E.D. Demaine and M. Hajiaghayi. Bidimensionality: new connections between
FPT algorithms and PTASs. In Proc. 16th ACM-SIAM Sympos. Discrete Algorithms, pages 590–601, 2005.
[DH08]
E.D. Demaine and M. Hajiaghayi. The bidimensionality theory and its algorithmic
applications. The Computer Journal, 51:292–302, 2008.
[DHM10]
E.D. Demaine, M. Hajiaghayi, and B. Mohar. Approximation algorithms via contraction decomposition. Combinatorica, 30:533–552, 2010.
[DHT06]
E.D. Demaine, M. Hajiaghayi, and D.M. Thilikos. The bidimensional theory of
bounded-genus graphs. SIAM J. Discrete Math., 20:357–371, 2006.
[Dij59]
E.W. Dijkstra. A note on two problems in connexion with graphs. NumerṀath.,
1:269–271, 1959.
[DL17]
V. Despré and F. Lazarus. Computing the geometric intersection number of
curves. In Proc. 33rd Sympos. Comput. Geom., vol. 77 of LIPIcs, article 35, Schloss
Dagstuhl, 2017.
[DMST11]
E. Demaine, S. Mozes, C. Sommer, and S. Tazari. Algorithms for Planar Graphs
and Beyond. Course notes, http://courses.csail.mit.edu/6.889/fall11/
lectures/, 2011.
[EFN12]
J. Erickson, K. Fox, and A. Nayyeri. Global minimum cuts in surface embedded
graphs. In Proc. 23rd ACM-SIAM Sympos. Discrete Algorithms, pages 1309–1318,
2012.
[EHP04]
J. Erickson and S. Har-Peled. Optimally cutting a surface into a disk. Discrete
Comput. Geom., 31:37–59, 2004.
[EK14]
M. Elberfeld and K. Kawarabayashi. Embedding and canonizing graphs of bounded
genus in logspace. In Proc. 46th ACM Sympos. Theory of Computing, pages 383–
392, 2014.
[EKL06]
A. Efrat, S.G. Kobourov, and A. Lubiw. Computing homotopic shortest paths
efficiently. Comput. Geom., 35:162–172, 2006.
[EN11a]
J. Erickson and A. Nayyeri. Computing replacement paths in surface-embedded
graphs. In Proc. 22nd ACM-SIAM Sympos. Discrete Algorithms, pages 1347–1354,
2011.
[EN11b]
J. Erickson and A. Nayyeri. Minimum cuts and shortest non-separating cycles via
homology covers. In Proc. 22nd ACM-SIAM Sympos. Discrete Algorithms, pages
1166–1176, 2011.
[EN11c]
J. Erickson and A. Nayyeri. Shortest non-crossing walks in the plane. In Proc.
22nd ACM-SIAM Sympos. Discrete Algorithms, pages 297–308, 2011.
[EN13]
J. Erickson and A. Nayyeri. Tracing compressed curves in triangulated surfaces.
Discrete Comput. Geom., 49:823–863, 2013.
[Epp00]
D. Eppstein. Diameter and treewidth in minor-closed graph families. Algorithmica,
27:275–291, 2000.
[Epp03]
D. Eppstein. Dynamic generators of topologically embedded graphs. In Proc. 14th
ACM-SIAM Sympos. Discrete Algorithms, pages 599–608, 2003.
[Epp09]
D. Eppstein. Squarepants in a tree: sum of subtree clustering and hyperbolic pants
decomposition. ACM Trans. Algorithms, 5, 2009.
28 É. Colin de Verdière
[Eps66]
D.B.A. Epstein. Curves on 2-manifolds and isotopies. Acta Math., 115:83–107,
1966.
[Eri10]
J. Erickson. Maximum flows and parametric shortest paths in planar graphs. In
Proc. 21st ACM-SIAM Sympos. Discrete Algorithms, pages 794–804, 2010.
[Eri11]
J. Erickson. Shortest non-trivial cycles in directed surface graphs. In Proc. 27th
Sympos. Comput. Geom., pages 236–243, ACM Press, 2011.
[Eri12]
J. Erickson. Combinatorial optimization of cycles and bases. In A. Zomorodian,
editor, Advances in Applied and Computational Topology, vol. 70 of Proc. Sympos.
Appl. Math., pages 195–228, AMS, Providence, 2012.
[Eri13]
J. Erickson. Computational Topology. Course notes, http://compgeom.cs.uiuc.
edu/~jeffe/teaching/comptop/, 2013.
[ES14]
J. Erickson and A. Sidiropoulos. A near-optimal approximation algorithm for
asymmetric TSP on embedded graphs. In Proc. 13th Sympos. Comput. Geom.,
pages 130–135, ACM Press, 2014.
[EW05]
J. Erickson and K. Whittlesey. Greedy optimal homotopy and homology generators. In Proc. 16th ACM-SIAM Sympos. Discrete Algorithms, pages 1038–1046,
2005.
[EW10]
J. Erickson and P. Worah. Computing the shortest essential cycle. Discrete Comput. Geom., 44:912–930, 2010.
[EW13]
J. Erickson and K. Whittlesey. Transforming curves on surfaces redux. In Proc.
24th ACM-SIAM Sympos. Discrete Algorithms, pages 1646–1655, 2013.
[FGM88]
M.L. Furst, J.L. Gross, and L.A. McGeoch. Finding a maximum-genus graph
imbedding. J. ACM, 35:523–534, 1988.
[FM11]
B. Farb and D. Margalit. A Primer on Mapping Class Groups. Princeton University Press, 2011.
[Fox13]
K. Fox. Shortest non-trivial cycles in directed and undirected surface graphs. In
Proc. 24th ACM-SIAM Sympos. Discrete Algorithms, pages 352–364, 2013.
[FT87]
M.L. Fredman and R.E. Tarjan. Fibonacci heaps and their uses in improved
network optimization algorithms. J. ACM, 34:596–615, 1987.
[GB07]
A. Grigoriev and H.L. Bodlaender. Algorithms for graphs embeddable with few
crossings per edge. Algorithmica, 49:1–11, 2007.
[GHT84]
J.R. Gilbert, J.P. Hutchinson, and R.E. Tarjan. A separator theorem for graphs
of bounded genus. J. Algorithms, 5:391–407, 1984.
[GJK+ 88]
S. Gao, M. Jerrum, M. Kaufmann, K. Mehlhorn, W. Rülling, and C. Storb. On
continuous homotopic one layer routing. In Proc. 4th Sympos. Comput. Geom.,
pages 392–402, ACM Press, 1988.
[GS90]
S.M. Gersten and H.B. Short. Small cancellation theory and automatic groups.
Invent. Math., 102:305–334, 1990.
[GS97]
M. de Graaf and A. Schrijver. Making curves minimally crossing by Reidemeister
moves. J. Combin. Theory Ser. B, 70:134–156, 1997.
[GT87]
J.L. Gross and T.W. Tucker. Topological Graph Theory. Wiley, New York, 1987.
[GW01]
I. Guskov and Z.J. Wood. Topological noise removal. In Proc. Graphics Interface,
pages 19–26, Canad. Inf. Process. Soc., Toronto, 2001.
[GY03]
X. Gu and S.-T. Yau. Global conformal surface parameterization. In Proc. Eurographics/ACM Sympos. Geom. Processing, pages 127–137, 2003.
Chapter 23: Computational topology of graphs on surfaces
29
[Han34]
C. Chojnacki (H. Hanani). Über wesentlich unplättbare Kurven im dreidimensionalen Raume. Fund. Math., 23:135–142, 1934.
[HC10]
P. Hliněný and M. Chimani. Approximating the crossing number of graphs embeddable in any orientable surface. In Proc. ACM-SIAM Sympos. Discrete Algorithms,
pages 918–927, 2010.
[HD16]
P. Hliněný and M. Derňár. Crossing number is hard for kernelization. In Proc.
32nd Sympos. Comput. Geom., vol. 51 of LIPIcs, article 42, Schloss Dagstuhl,
2016.
[Hen94]
M. Henle. A Combinatorial Introduction to Topology. Dover Publications, Mineola,
1994.
[HKMT16]
A. Hubard, V. Kaluža, A. de Mesmay, and M. Tancer. Shortest path embeddings
of graphs on surfaces. In Proc. 32nd Sympos. Comput. Geom., vol. 51 of LIPIcs,
article 43, Schloss Dagstuhl, 2016.
[HPN+ 16]
S. Har-Peled, A. Nayyeri, M. Salavatipour, and A. Sidiropoulos. How to walk your
dog in the mountains with no magic leash. Discrete Comput. Geom., 55:39–73,
2016.
[HS85]
J. Hass and P. Scott. Intersections of curves on surfaces. Israel J. Math., 51:90–120,
1985.
[HS94a]
J. Hass and P. Scott. Shortening curves on surfaces. Topology, 33:25–43, 1994.
[HS94b]
J. Hershberger and J. Snoeyink. Computing minimum length paths of a given
homotopy class. Comput. Geom., 4:63–98, 1994.
[HT74]
J. Hopcroft and R. Tarjan. Efficient planarity testing. J. ACM, 21:549–568, 1974.
[INS+ 11]
G.F. Italiano, Y. Nussbaum, P. Sankowski, and C. Wulff-Nilsen. Improved algorithms for min cut and max flow in undirected planar graphs. In Proc. 43rd ACM
Sympos. Theory of Computing, pages 313–322, 2011.
[Kel06]
J.A. Kelner. Spectral partitioning, eigenvalue bounds, and circle packings for
graphs of bounded genus. SIAM J. Comput., 35:882–902, 2006.
[Ket99]
L. Kettner. Using generic programming for designing a data structure for polyhedral surfaces. Comput. Geom., 13:65–90, 1999.
[KKS11]
K. Kawarabayashi, P.N. Klein, and C. Sommer. Linear-space approximate distance
oracles for planar, bounded-genus and minor-free graphs. In Proc. Int. Coll. Automata, Languages and Progr., part 1, vol. 6755 of LNCS, pages 135–146, Springer,
Berlin, 2011.
[Kle05]
P.N. Klein. A linear-time approximation scheme for planar weighted TSP. In Proc.
46th IEEE Sympos. Found. Comp. Sci., pages 647–657, 2005.
[KM07]
K. Kawarabayashi and B. Mohar. Some recent progress and applications in graph
minor theory. Graphs Combin., 23:1–46, 2007.
[KMR08]
K. Kawarabayashi, B. Mohar, and B. Reed. A simpler linear time algorithm for
embedding graphs into an arbitrary surface and the genus of graphs of bounded
tree-width. In Proc. 49th IEEE Sympos. Found. Comp. Sci., pages 771–780, 2008.
[KR07]
K. Kawarabayashi and B. Reed. Computing crossing number in linear time. In
Proc. 39th ACM Sympos. Theory of Computing, pages 382–390, 2007.
[KS15]
K. Kawarabayashi and A. Sidiropoulos. Beyond the Euler characteristic: approximating the genus of general graphs. In Proc. 47th ACM Sympos. Theory of Computing, pages 675–682, 2015.
30 É. Colin de Verdière
[KT12]
M. Kamiński and D.M. Thilikos. Contraction checking in graphs on surfaces. In
Proc. 29th Sympos. Theoret. Aspects Comp. Sci., vol. 14 of LIPIcs, pages 182–193,
Schloss Dagstuhl, 2012.
[Kut06]
M. Kutz. Computing shortest non-trivial cycles on orientable surfaces of bounded
genus in almost linear time. In Proc. 22nd Sympos. Comput. Geom., pages 430–438,
ACM Press, 2006.
[Lin82]
S. Lins. Graph-encoded maps. J. Combin. Theory Ser. B, 32:171–181, 1982.
[LM85]
C.E. Leiserson and F.M. Maley. Algorithms for routing and testing routability of
planar VLSI layouts. In Proc. 17th ACM Sympos. Theory of Computing, pages
69–78, 1985.
[LPVV01]
F. Lazarus, M. Pocchiola, G. Vegter, and A. Verroust. Computing a canonical
polygonal schema of an orientable triangulated surface. In Proc. 17th Sympos.
Comput. Geom., pages 80–89, ACM Press, 2001.
[LR12]
F. Lazarus and J. Rivaud. On the homotopy test on surfaces. In Proc. 53rd IEEE
Sympos. Found. Comp. Sci., pages 440–449, 2012.
[LZ04]
S.K. Lando and A.K. Zvonkin.
Springer-Verlag, Berlin, 2004.
[Mat14]
J. Matoušek. String graphs and separators. In J. Nešetřil and M. Pellegrini, Geometry, Structure and Randomness in Combinatorics, pages 61–97, Scuola Normale
Superiore, Pisa, 2014.
[Mie09]
G. Miermont. Tessellations of random maps of arbitrary genus. Annales Scientifiques de l’École normale supérieure, Quatrième série, 42:725–781, 2009.
[MK11]
W. Myrvold and W. Kocay. Errors in graph embedding algorithms. J. Comput.
System Sci., 77:430–438, 2011.
[MMP87]
J.S.B. Mitchell, D.M. Mount, and C.H. Papadimitriou. The discrete geodesic
problem. SIAM J. Comput., 16:647–668, 1987.
[Moh97]
B. Mohar. On the minimal genus of 2-complexes. J. Graph Theory, 24:281–290,
1997.
[Moh99]
B. Mohar. A linear time algorithm for embedding graphs in an arbitrary surface.
SIAM J. Discete Math., 12:6–26, 1999.
[MS12]
Y. Makarychev and A. Sidiropoulos. Planarizing an unknown surface. In Proc.
15th Workshop on Approximation, vol. 7408 of LNCS, pages 266–275, Springer,
Berlin, 2012.
[MSTW16]
J. Matoušek, E. Sedgwick, M. Tancer, and U. Wagner. Untangling two systems of
noncrossing curves. Israel J. Math., 212:37–79, 2016.
[MT01]
B. Mohar and C. Thomassen. Graphs on Surfaces. Johns Hopkins University
Press, 2001.
[Neg01]
S. Negami. Crossing numbers of graph embedding pairs on closed surfaces.
J. Graph Theory, 36:8–23, 2001.
[NR04]
T. Nishizeki and M.S. Rahman. Planar Graph Drawing. World Scientific, Singapore, 2004.
[Pat13]
V. Patel. Determining edge expansion and other connectivity measures of graphs
of bounded genus. SIAM J. Comput., 42:1113–1131, 2013.
[PPSL14]
Ma. Pilipczuk, Mi. Pilipczuk, P. Sankowski, and E.J. van Leeuwen. Network
sparsification for Steiner problems on planar and bounded-genus graphs. In Proc.
55th IEEE Sympos. Found. Comp. Sci., pages 186–195, 2014.
Graphs on Surfaces and Their Applications.
Chapter 23: Computational topology of graphs on surfaces
31
[Prz15]
P. Przytycki. Arcs intersecting at most once. Geom. Funct. Anal., 25:658–670,
2015.
[PSŠ08]
M.J. Pelsmajer, M. Schaefer, and D. Štefankovič. Odd crossing number and crossing number are not the same. Discrete Comput. Geom., 39:442–454, 2008.
[PSS09a]
M.J. Pelsmajer, M. Schaefer, and D. Stasi. Strong Hanani–Tutte on the projective
plane. SIAM J. Discrete Math., 23:1317–1323, 2009.
[PSŠ09b]
M.J. Pelsmajer, M. Schaefer, and D. Štefankovič. Removing even crossings on
surfaces. European J. Combin., 30:1704–1717, 2009.
[RS03]
N. Robertson and P.D. Seymour. Graph minors. XVI. Excluding a non-planar
graph. J. Combin. Theory Ser. B, 89(1):43–76, 2003.
[RS04]
N. Robertson and P.D. Seymour. Graph minors. XX. Wagner’s conjecture. J. Combin. Theory Ser. B, 92:325–357, 2004.
[RS05]
R.B. Richter and G. Salazar. Two maps with large representativity on one surface.
J. Graph Theory, 50:234–245, 2005.
[RS12]
N. Robertson and P.D. Seymour. Graph minors. XXII. Irrelevant vertices in linkage
problems. J. Combin. Theory Ser. B, 102:530–563, 2012.
[RST13]
J. Rué, I. Sau, and D.M. Thilikos. Asymptotic enumeration of non-crossing partitions on surfaces. Discrete Math., 313:635–649, 2013.
[RST14]
J. Rué, I. Sau, and D.M. Thilikos. Dynamic programming for graphs on surfaces.
ACM Trans. Algorithms, 10:8, 2014.
[Sch90a]
W. Schnyder. Embedding planar graphs on the grid. In Proc. 1st ACM-SIAM
Sympos. Discrete Algorithms, pages 138–148, 1990.
[Sch90b]
A. Schrijver. Homotopic routing methods. In B. Korte, L. Lovász, H.J. Prömel,
and A. Schrijver, editors, Paths, Flows, and VLSI-layout, pages 329–371. SpringerVerlag, Berlin, 1990.
[Sch91]
A. Schrijver. Disjoint circuits of prescribed homotopies in a graph on a compact
surface. J. Combin. Theory Ser. B, 51:127–159, 1991.
[Sch03]
A. Schrijver. Combinatorial optimization: Polyhedra and efficiency. Vol. 24 of
Algorithms and Combinatorics, Springer-Verlag, Berlin, 2003.
[Sch13a]
M. Schaefer. The graph crossing number and its variants: A survey. Electron.
J. Combin., Dynamic Surveys, article 21, 2013. Updated in 2014.
[Sch13b]
M. Schaefer. Toward a theory of planarity: Hanani–Tutte and planarity variants.
J. Graph. Alg. Appl., 17:367–440, 2013.
[Sch14]
M. Schaefer. Hanani–Tutte and related results. In I. Bárány, K.J. Böröczky, and
L. Fejes Tóth, editors, Geometry—Intuitive, Discrete, and Convex: A tribute to
László Fejes Tóth, vol. 24 of Bolyai Society Math. Studies, pages 259–299, Springer,
Berlin, 2014.
[Sid10]
A. Sidiropoulos. Optimal stochastic planarization. In Proc. 51st IEEE Sympos.
Found. Comp. Sci., pages 163–170, 2010.
[SSŠ02]
M. Schaefer, E. Sedgwick, and D. Štefankovič. Algorithms for normal curves and
surfaces. In Proc. 8th Conf. Computing and Combinatorics, vol. 2387 of LNCS,
pages 370–380, Springer, Berlin, 2002.
[SSŠ08]
M. Schaefer, E. Sedgwick, and D. Štefankovič. Computing Dehn twists and geometric intersection numbers in polynomial time. In Proc. 20th Canad. Conf.
Comput. Geom., pages 111–114, 2008.
[Sti93]
J. Stillwell. Classical Topology and Combinatorial Group Theory, 2nd edition.
Springer-Verlag, New York, 1993.
32 É. Colin de Verdière
[Thi12]
D.M. Thilikos. Graph minors and parameterized algorithm design. In H.L. Bodlaender, R.G. Downey, F.V. Fomin, and D. Marx, editors, The Multivariate Algorithmic Revolution and Beyond, vol. 7370 of LNCS, pages 228–256, Springer,
Berlin, 2012.
[Thi15]
D.M. Thilikos. Bidimensionality and parameterized algorithms. In Proc. 10th
Sympos. Parameterized and Exact Computation, vol. 43 of LIPIcs, pages 1–16,
Schloss Dagstuhl, 2015.
[Tho89]
C. Thomassen. The graph genus problem is NP-complete. J. Algorithms, 10:568–
576, 1989.
[Tho90]
C. Thomassen. Embeddings of graphs with no short noncontractible cycles.
J. Combin. Theory Ser. B, 48:155–177, 1990.
[Tót12]
G. Tóth. A better bound for pair-crossing number. In J. Pach, editor, Thirty
Essays on Geometric Graph Theory, pages 563–567, Springer, New York, 2012.
[Tut70]
W.T. Tutte. Toward a theory of crossing numbers. J. Combin. Theory, 8:45–53,
1970.
[WHDS04]
Z.J. Wood, H. Hoppe, M. Desbrun, and P. Schröder. Removing excess topology
from isosurfaces. ACM Trans. Graph., 23:190–208, 2004.
| 8cs.DS
|
Eye-Movement behavior identification for
AD diagnosis
Juan Biondi
∗1,2
, Gerardo Fernandez1 , Silvia Castro2,1 , and Osvaldo Agamennoni1,3
1
arXiv:1702.00837v3 [cs.NE] 15 Jan 2018
Laboratorio de Desarrollo en Neurociencia Cognitiva, Instituto de Investigaciones en Ingenierı́a Eléctrica
(IIIE), Departamento de Ingenierı́a Eléctrica y de Computadoras (DIEC), Universidad Nacional del Sur (UNS)
- CONICET, Bahı́a Blanca, Argentina
2
Laboratorio de Visualización y Computación Gráfica (VyGLab), Departamento de Ciencias e Ingenierı́a de la
Computación (DCIC), Universidad Nacional del Sur (UNS), Bahı́a Blanca, Argentina
3
Comisión de Investigaciones Cientı́ficas de la Provincia de Buenos Aires (CIC), Argentina
Abstract
In the present work, we develop a deep-learning approach to differentiate between the
eye-movement’s behavior of people with neurodegenerative diseases over healthy control subjects, from reading. The subjects with and without Alzheimer’s disease read well-defined
and previously validated sentences including high-, low-predictable sentences, and proverbs.
From these eye-tracking data we derive trial-wise information consisting of descriptors that
capture the reading behavior of the subjects. With this information we train a set of denoising sparse-autoencoders and build a deep neural network using the trained autoencoders
and a softmax classifier that allows identifying patients with Alzheimers disease with 89.78%
of accuracy. Our results are very encouraging and show that these models promise to be
helpful to understand the dynamics of the eye movement behavior and its relation with the
underlying neuropsychological processes.
Keywords: Eye-tracking, Deep-learning, Alzheimer’s Disease
1
Introduction
Alzheimer’s disease (AD) is a nonreversible neurodegenerative disease characterized by progressive impairment of cognitive and memory functions that develops over a period of years being the most prevalent
cause of dementia in elderly subjects. Initially, people experience memory loss and confusion, which may
be mistaken for the kinds of memory changes that are sometimes associated with normal aging Waldemar et al. (2007). The subtle changes in behavior and response of the early manifestation of this disease
make it difficult to diagnose by using the classical neuropsychological tests such as the Mini-Mental State
Examination. The use of more advanced diagnosis tools such as MRI and PET results is critical for its
early diagnosis. Since AD is nonreversible, its early treatment can improve the patient’s life delaying the
full manifestation of the disease. In the last years, the study of the eye movement, known as eye-tracking,
during reading, has proved to help performing this task (Fernández et al., 2015b, 2016b, 2015a, 2013).
Reading is a cognitive activity that has received considerable attention of researchers to evaluate
human cognitive performance. This requires the integration of several central cognitive subsystems, from
attention and oculomotor control to word identification and language comprehension. Eye movements
show a reproducible pattern during normal reading. Each eye movement ends up in a fixation point, which
allows the brain to process the incoming information and to program the following saccade. Different
neuropsychiatric pathologies produce abnormalities in eye movements and disturbances in reading, having
each of them a particular pattern that can be registered and measured (Fernández et al., 2016c,a; Holzman
et al., 1974; Iacono et al., 1992; Riby and Hancock, 2009; Kellough et al., 2008). Eye movements can be
classified into three groups:
∗
Corresponding Author.
Email: [email protected]
1
• Movement for maintaining the image on the fovea (area of the retina with acuity vision), compensating head or object movements;
• Movements for shifting the eyes, when the attention changes from one object to another. There are
subtypes of shifting movements: saccades (looking for a new center of visual attention), monitoring
and vergence (slower than saccades and are responsible for carrying the image of interest to both
foveae, allowing stereoscopic vision);
• Movements of binocular fixation that also prevent fading of the image. These movements have
three variations: tremor, drift, and microsaccade.
Saccades are rapid big eye movements particularly important from the cognitive point of view since
cognitive processes have a direct influence on such movements. Each saccade has its direction. People,
depending on language, read from left to right and most of the saccadic eye movements are oriented
accordingly. These normal reading movements are called forward saccades. Reading movements going
from right to left are called regressions. The saccade movement alternates with a fixation made when
the eyes are directed to a particular target (See (Rayner, 1998) for a review). As shown in (Fernández
et al., 2013), patients with early Alzheimer disease show alterations during the execution of tasks, such
as reading, and these alterations can be related to an impairment in their working memory (Fernández
et al., 2014a, 2016b). In fact, it has been shown that through this differentiation in the eye-movements,
it is possible to infer a diagnosis (Fernández et al., 2015a).
The use of computer-aided diagnosis is a key challenge since the growth of computational power
permits the creation of more complex models. These models can be used to create biomarkers that
help in disease identification. Since the popularization of deep-learning neural networks Schmidhuber
(2015); Deng and Yu (2014), many efforts have been made in their use in the field of medicine. This
technique is commonly used in conjunction with imaging diagnoses such as PET or MRI mainly because
the feature representation that this technology provides may help even when data is incomplete Li et al.
(2014). Specifically, there were advances in the detection and pattern differentiation of the physical brain
alterations that neurodegenerative diseases produce, such as AD and mild cognitive impairment (MCI)
Suk and Shen (2013); Suk et al. (2014). Even there were advances in its early diagnostic Liu et al. (2014).
The problem is that, when a brain physical alteration is observable, the damages made to the brain are
irreversible (even though the disease is in an early stage) and may cause deterioration in the quality of
life of the patient. The eye-tracking technique allows us to find subtler changes that were made by the
brain to alleviate small memory deficits in the patient. These changes are not noticed by the patient
but small changes in the way they read our set of sentences can be found with the technique presented
in this paper.
In this work, we use a deep-learning neural network trained on reading information extracted from
controls and patients with probable AD in order to identify the patterns made by them in the reading
process and later cluster them in their respective groups. Throughout this work, we may use AD patients
and patients with probable AD in an interchangeable manner because of the nature of the AD diagnosis.
The hypothesis was that using deep-learning in the feature identification of the key characteristics of the
patient’s eye behavior during reading sentences may lead to a correct classification that can be used to
infer a diagnosis. Using this type of technology may improve the results obtained in (Fernández et al.,
2015a) since it provides a smaller granularity in the detection of the disease and consequently, a better
performance. Additionally, this technology allows us to improve the effectiveness of the classification as
we collect more ground truth subjects.
2
2.1
Methods
Ethics Statement
The investigation adhered to the principles of the Declaration of Helsinki and was approved by the
Institutional Bioethics Committee of the Hospital Municipal de Agudos (Bahı́a Blanca, Buenos Aires,
Argentina). All patients and their caregivers, and all control subjects signed an informed consent prior
to their inclusion in the study.
2
2.2
Participants
Twenty six patients (mean age 69 years, SD = 7.3 years) with the diagnosis of probable AD were
recruited at Hospital Municipal of Bahı́a Blanca (Buenos Aires, Argentina). The clinical criteria to
diagnose AD at its early stages remains under debate McKhann et al. (1984). In the present work,
the diagnosis was based on the criteria for dementia outlined in the Diagnostic and Statistical Manual
of Mental Disorders (DSM-IV). All AD patients underwent a detailed clinical history revision, physical/neurological examination and thyroid function test. They all presented an APO E3E4 Genotype.
Magnetic resonance images were obtained from twelve patients and computerized tomography scans
from the other. All the patients underwent biochemical analysis to discard other common pathologies
(hemoglobin, full blood count, erythrocyte sedimentation rate, urea and electrolytes, blood glucose). All
this data provided a more precise diagnosis of AD. Patients were excluded if: (1) they suffered from any
medical conditions that could account for, or interfere with, their cognitive decline; (2) had evidence of
vascular lesions in computed tomography or FMRI; (3) had evidence for an Axis I diagnosis (e.g. major
depression or drug abuse) as defined by the DSM-IV. To be eligible for the study, patients had to have at
least one caregiver providing regular care and support. Patients taking cholinesterase inhibitors (ChE-I)
were not included. None of the subjects were taking hypnotics, sedative drugs or major tranquilizers.
The control group consisted of 43 elderly adults (mean age 71 years, SD = 6.1 years), with no known
neurological or psychiatric disease according to their medical records, and no evidence of cognitive decline
or impairment in daily activities. A one-way ANOVA showed no significant differences between the ages
of AD and Control individuals. Those participants diagnosed of suffering from Ophthalmologic disease
such as glaucoma, visually significant cataract or macular degeneration as well as visual acuity less than
20/20 were excluded from the study.
The mean scores of Controls and AD patients in the Mini-Mental State Examination (MMSE)
Folstein et al. (1975) were 27.8 (SD = 1.0) and 24.2 (SD = 0.8), respectively, the latter suggesting early
mental impairment. A one-way ANOVA evidenced significant differences between MMSE in AD patients
and Controls (p < 0.001). The mean score of AD patients in the Adenbrook’s Cognitive Examination Revised (ACE-R) Mioshi et al. (2006) was 84.4, (SD = 1.1), the cut-off being of 86. The mean school
education trajectories in AD patients and Controls were 15.2 (SD = 1.3) years and 15.1 (SD = 1.0)
years, respectively. A one-way ANOVA showed no significant differences between the education of AD
and Control individuals.
2.3
Apparatus and eye movement data
Single sentences were presented on the center line of a 20-inch LCD Monitor (1024x 768 pixels resolution;
font: regular; New Courier; 12 points, 0.2◦ in height). Participants sat at a distance of 60 cm from
the monitor. Head movements were minimized using a chin rest. Correction for the 60 cm viewing
distance was performed by using the Eyelink1000 corneal reflection system, which assessed changes in
gaze position by measuring both the reflection of an infrared illuminator on the cornea and the pupil
size, by means of a video camera sensitive to light in the infrared spectrum.
Eye movements were recorded with an EyeLink1000 Desktop Mount (SR Research) eye tracker, with
a sampling rate of 1000 Hz and an eye position resolution of a 20-s arc. All recordings and calibration
were binocular. Eye movement data from 69 participants reading 184 sentences resulted in a total of
48716 fixations - 13002 for Control and 35714 for the people with AD. This data was cleaned from blinks
and track losses. Prior to removing the analysis fixations shorter than 51ms and longer than 750ms,
and fixations on the first and last word of each sentence (see (Kliegl et al., 2006) for a description of the
analytic procedure), we measured, for each patient, the elapsed time between the instant when sentences
were first presented, and the instant when participants looked at the final spot: mean reading time in
high-predictable sentences was 3495ms vs. 5635ms (Controls and AD) and 4828ms vs. 6881ms (controls
and AD) in low predictable-sentences.
2.4
Procedures
Participant’s gaze was calibrated with a standard 13-point grid for both eyes. After validating the
calibration, a trial began with the appearance of a fixation point on the position where the first letter of
the sentence was to be presented. As soon as both eyes were detected within a radius equal to 1◦ from
3
the fixation spot, the sentence was presented. After reading it, participants looked at a dot in the lower
right corner of the screen; when the gaze was detected on the final spot, the trial ended. Occasionally,
external factors such as minor movements and slippages of the head gear could cause small drifts. To
avoid them, we performed a drift correction before the presentation of each spot.
To assess whether subjects comprehended the texts, they were presented with a three alternative
multiple-choice question about the sentence in progress in 20% of the sentence trials. Participants
answered the questions by moving a mouse and choosing the response with a mouse click. Overall mean
accuracy was 95% (SD = 3.2%) in Control and 91% (SD = 5.4%) in AD. A one-way ANOVA showed no
significant differences between comprehension of the answers in Controls and in AD patients. The latter
were only marginally less accurate than Control subjects, probably because they were in an early stage
of the pathology, as indicated by the MMSE and ACE-R values. Once the comprehension test ended,
the next trial started with the presentation of the fixation spot. An extra calibration was done after 15
sentences or if the eye tracker did not detect the eye at the initial fixation point within 2 s.
2.5
Sentence corpus
The sentence corpus was composed of short sentences of a line with 75 low predictable sentences, 45
high-predictable sentences and 64 proverbs (e.g., “Maria is always laughing and in a good mood”, “It is
worthwhile to think before talking” and “A bird in the hand is worth two in the bush”)Fernández et al.
(2014b). All the sentences comprised a well-balanced number of content and function words and had
similar grammatical structure.
2.5.1
Word and Sentence Lengths
Sentences ranged from a minimum of 5 words to a maximum of 14 words. Mean sentence length was 8.1
(SD = 1.4) words for low predictability sentences, 7.6 words (SD = 1.5) for high predictability sentences
and 7.3 words (SD = 1.9) for proverbs. Words ranged from 1 to 14 letters. Mean word length was 4.6,
4.1 and 4.0 (SD = 2.5, SD = 2.3 and SD = 2.0) for low-, high-predictable sentences and proverbs,
respectively.
2.5.2
Word Frequencies
We used the Spanish Lexical Lexesp corpus Sebastián-Gallés (2000) for assigning a frequency to each
word of the sentence corpus. Word frequencies ranged from 1 to 264721 per million, so we transformed
it to log10 (f requency). Mean log10 (f requency) was 3.4 (SD = 1.3) for low predictability sentences, 3.4
(SD = 1.5) for high predictability sentences and 3.47 (SD = 1.36) for proverbs.
2.5.3
Word Predictability
It was measured in an independent experiment with 18 researchers of the Electrical Engineering and
Computer Science Department of Universidad Nacional del Sur. We used an incremental cloze task
procedure in which participants had to guess the next word given only the prior words of the sentence.
Participants guessed the first word of the unknown sentence and entered it via the keyboard. In return, the computer presented the first word of the original sentence on the screen. Responding to this,
participants entered their guess for the second word and so on, until a period indicated the end of the
sentence. Correct words stayed on the screen. Participants were between 31 and 62 years old and
did not participate in the reading experiment. Academic background of the reading experiment group
and the cloze task group was similar. Word predictabilities ranged from 0 to 1 with a mean of 0.38
(SD = 0.36). The average predictability measured from the cloze task was transformed using a logit
function 0.5 ∗ ln(pred/(1 − pred)); predictabilities of zero were replaced with 1/(2 ∗ 18) = −2.55 and
those among the five perfectly predicted words with (2 ∗ 18 − 1)/(2 ∗ 18) = +2.55, where 18 represents
the number of complete predictability protocols. Mean logit predictability was −0.9 (SD = 0.9) for
low predictability sentences, 0.0 (SD = 1.29) for high predictability sentences and 0.08 (SD = 1.23) for
proverbs.
As in other languages, we find strong correlations in Spanish between word length, word frequency,
and word predictability. Long words are of low frequency (r = −0.80 and r = −0.75 in low and in high
4
predictability sentences, respectively). Frequent words are highly predictable (r = 0.47 and r = 0.37 in
low and in high predictability sentences, respectively), and highly predictable words tend to be short
words (r = −0.47 and r = −0.38 in low and in high predictability sentences, respectively).
2.6
Used information
The information used for this work was a trial-wise compaction of the original data where we keep
descriptors of the mean reading behavior of the subjects in each read sentence. We measured the
saccade amplitude, fixation duration and duration of the fixation on a single word of the subject during
the reading of each sentence but only kept the mean and the standard deviation. In addition, we measured
the total number of fixations and classified them by first pass fixations, refixations, unique fixations and
total fixations:
• First pass fixations: The first fixation on a specific word of the sentence.
• Unique fixations: Fixations that occur once in a word that was skipped in the first pass.
• Multiple fixations: Multiple fixations on a word in the first pass.
• Refixations: Fixations that take place once a word that already has a first pass fixation or a unique
fixation implying a regression.
Categorical data as the diagnosis (used as training labels) was replaced by numerical values in order
to unify the data types and improve during the classification process. An integer with two possible values
was used for the diagnosis information construction: 0 for “Control” and 1 for “AD”. The identification
and the diagnosis information of the subject were kept apart from the data. A detail of the variables used
as input for the model construction is shown in table 1. Since the tag (AD or Control ) is associated to
the patient and not to each sentence, and, since we use a per-trial classification approach, the subject’s
tag was applied to all the sentences read by him/her. Following this approach may introduce noise in
the classifying stage because, as we use a per-trial classification approach, a Control subject could, for
example, be distracted during the reading of a specific sentence thus making him read as a non-healthy
person. Anyway, the system should be able to detect and ignore these artifacts because many samples
are used in the training stage.
Table 1: Used variables for the model construction.
Name
nw
gaze
sd gaze
as
sd as
ntf
ntm
dfp
sd dfp
fpp
rf
nfu
dfu
sd dfu
Description
Number of words in the sentence.
Global (sentence) mean of the sum of fixation durations on the same word.
Standard deviation of gaze.
Mean saccade amplitude in the sentence.
Standard deviation of as
Count of the total number of fixations on the sentence.
Count of the number of multifixations on the sentence.
Mean duration of the first pass fixations on the sentence.
Standard deviation of dfp.
Count of the number of first pass fixations on the sentence.
Count of refixations on the sentence.
Count of unique fixations on the sentence.
Mean duration of unique fixations on the sentence.
Standard deviation of dfu.
All the data was previously outlier-checked by establishing a dropout policy in order to use a cleaner
dataset. The outlier check policy consisted of finding the mean and the SD of each condition group and
checking the two groups separately. All the trials where the standard deviation was bigger than two
5
times the standard deviation of the group were considered as outliers and was dropped out, resulting
in a 10% samples lost. The sentence identication, order, and type were kept separate from the training
information. This is because the standard deviations in the information for the AD patients for proverbs
and high predictability sentences, appear to be particularly high after the data is outlier-checked, causing
a highly unbalanced dataset.
The resulting dataset consisted on 3235 trials with mean 46.88 (SD = 10.76) trials per subject. The
dataset was splitted in two groups: one for the training of the network with data of 61 subjects and other
for testing with data of 8 subjects randomly picked. Finally, the training dataset consisted on 2922 trials
of 61 subjects - 39 Control and 22 AD - with 47.9 (SD = 10.47) mean trials each; the testing dataset
consisted on 313 trials of 8 subjects - 4 Control and 4 AD - with 39.12 (SD = 10.37) mean trials each.
Splitting the data in this way ensures that the network can’t infer the condition in other way, avoids
over-fitting and ensures that the testing data is totally unknown by the network.
2.7
Deep learning with denoising sparse-autoencoders
In this work we used sparse-autoencoders for the codification stage. The sparse-autoencoders work just
as regular autoencoders, i.e. there are neural networks under supervised learning with the targets set
equal to the input (the identity) but, in the case of sparse-autoencoders, an average number of activations
per neuron restriction was applied in the hidden layer by penalizing the average number of activations
different from the desired (known as sparsity proportion) adding a penalty term to the cost function.
This restriction is introduced so that each neuron specializes on a particular feature. The lower the
sparsity proportion, the more specific the feature. The resulting trained neural network can be thought
of as: an encoder, involving the input and the hidden layer, and a decoder, involving the hidden and the
output layer. In this case, we set an activation restriction equal to 10%.
In a denoising-autoencoder, the idea is to force the hidden layer to discover more robust features
and to prevent it from simply learning the identity, by training the autoencoder to reconstruct the input
from a corrupt version. The altered version of the input was generated by introducing noise, which was
obtained by clamping some of the fields to zero. The corrupt data was used as the sparse-autoencoder
input, and the clean (unaltered) data as the target. Using this type of data corruption mechanism forces
the network to learn a way of reconstructing a field based on others. This, combined with the sparsity
restriction, results in more robust features.
The deep-learning neural network was built using two stages of these denoising sparse-autoencoders.
In each stage, we train the autoencoder by corrupting the encoded clean data obtained from the previous
stage, and providing it to the next as its input. At the end of the two stages, we set a softmax layer as
a classifier, training it with uncorrupted data and the corresponding tag. As we used a per-trial classification approach, the patient diagnosis was extended to all the sentences read by him and the classifier
was trained with this data as the target. The softmax layer is a non-linear, multiclass generalization of
the binary Logistic Regression, and its output is the “probability” for each class (we quoted the word
“probability” because it’s shape depends on of the regularization used in the training stage, it can be
more diffuse or peaky).
3
Results
Several configurations were generated by varying the sparsity proportion, the number of units and layers
and the shape of the network (same vs. decreasing number of units between layers). We adopted the
one that produced the best results which consisted of two layers of denoising sparse-autoencoders with
16 and 4 hidden units in the hidden layer each, using a sparsity proportion of 10%. After the training
of the network, a series of tests were performed with data not included in the training dataset. This
dataset, as mentioned, was composed of 313 sentences from 8 subjects - 4 Control and 4 AD - with
39.12 (SD = 10.37) mean trials each. We used a softmax layer for the classification trained using the
condition translation with 0 for Control subjects and 1 for AD subjects. This means that we have a
single class “AD” and, since the output of the classifier is a real number between 0 and 1, the read
sentences classified by the network with values close to 0 have a small “probability” of being read by
an AD patient (i.e. high probability of being read by a Control patient) and vice-versa. The “ground
truth” values are known, so we can split the output into groups and observe the number of sentences
6
misclassified by the network. Based on this, we show the output of the network where values below 0.5
are considered as classified as Control, and higher values are considered as classified as AD (see Figure
1). As can be seen, the output of the network was consistent with the expected values.
Figure 1: Classification results histogram representing number of sentences, split by “ground
truth” values. Values below 0.5 are considered as classified as Control, and higher values are
considered as classified as AD.
Figure 2: Classification results. Values below 0.5 are considered as classified as Control (class
0), and higher values are considered as classified as AD (class 1).
Now, we can round the values so we can plot a confusion matrix and approximate the number of
misclassified sentences in order to measure the performance of the network. In the figure 2 we can see
7
a confusion matrix of the output. The “X” axis represents the expected output values and the “Y”
axis the rounded output of the network. As can be seen, the overall performance of the network was
good giving a 89.8% of well-classified sentences. The performance of the network using sentences read
by Control patients (88.7% correctness) approximates to the performance observed using sentences read
by AD patients (91.0% correctness).
Figure 3: Number of misclassified sentences by type, split by “ground truth” label.
On the other hand, misclassified sentences were not concentrated on a particular type of sentence
as can be seen in figure 3. Here we can see the original concentration of sentences types in the testing
dataset and the correctness of the classification following the mentioned method.
Figure 4: Parallel coordinates plot with two subsets (one composed of AD patients and the
other of Control subjects) of trials that have similar values for the input in each field and its
codification during the different stages of the network. As expected, similar values encoded
“together”. Control subjects encoded closer than AD subjects; This may be due the high
“within group” variability of the AD group.
This result, added to the fact that neither was a concentration of misclassication in a particular
sentence, may suggest that most of the misclassications occurred due to presumably stochastic processes.
In addition, the trained networks were evaluated using a spread result test in order to determine the
8
softness of the model. These tests checked if similar information is encoded in a similar way in the
subsequent stages of the network. A significant differentiation in later stages of encoding may show
over-fitting in the network (and/or in the different stages).
Two subsets (one composed of AD patients and the other of Control subjects) of trials that have
similar values for the input in each field are shown in Figure 4. As shown, similar input values map
to similar encoded values on each stage of the autoencoders. This is because the modeled function is
smooth. Furthermore, the data through the subsequent stages of codification tends to group. These
results have shown that the output information such as the encoding in the different stages are reliable.
On the other hand, they show that certain neurons in later stages tend to specialize on the detection of
specific AD or Control input features.
4
Conclussions and future work
The results showed that using a deep-learning architecture for identifying the characteristic eye movements patterns of neurodegenerative diseases like Alzheimer’s disease is a good approach since this
technology is focused on pattern finding and is suitable for this work. Moreover, the high performance
in a per-trial classification approach, leads us to conclude that, since a single patient reads many sentences, the assertion rate per patient is higher than the 89.8% accuracy reported in this work. Using the
policy that network outputs higher than 0.5 are classified as AD and below as Control, if we tag each
patient using “majority voting” over all of his/her read sentences, the network gets a 100% classification
accuracy for the testing set - 8 well classified subjects from 8 total -. This was expected since, on this
test set, the total number of missclassified sentences is 32 and every patient, after the dataset cleaning,
have 39.12 (SD = 10.37) mean sentences.
Table 2: Comparison of mean value given by the network and “severity of the disease” score
given by psychiatrists.
IDPat
58
57
66
60
56
55
63
64
70
69
65
59
71
62
67
68
53
Mean
0,97
0,95
0,49
0,95
0,96
0,94
0,87
0,51
0,90
0,84
0,91
0,58
0,75
0,76
0,47
0,40
0,84
SD
0,17
0,17
0,32
0,16
0,16
0,17
0,25
0,34
0,24
0,25
0,22
0,36
0,33
0,32
0,35
0,31
0,31
Score
0,9
0,5
0,5
0,6
0,8
0,7
0,8
0,5
0,5
0,5
0,6
0,6
0,6
0,6
0,5
0,8
0,8
Mean
SD
Difference
0,07
0,45
0,01
0,35
0,16
0,24
0,07
0,01
0,40
0,34
0,31
0,02
0,15
0,16
0,03
0,40
0,04
0,19
0,15
Additionally, we asked the head psychiatrists of other AD patients (that were not included in training
process) to elaborate a score of the overall severity of the disease of each patient on the traditional tests
using a scale from 0 to 1, without knowing the results given by our network. The process of creating the
9
score required that the physicians have a deep knowledge of the psychiatric history of each patient, the
recompilation of the results of every neuropsychological tests made by the patient and its comprehension.
Table 2 shows the scores given by the psychiatrists compared with the mean value obtained in our
network for all the sentences read by the subjects and its standard deviation (S D). As seen, for most of
the patients, the values obtained are very similar to the scores given by the psychiatrists with a mean
value of 0.19 (SD = 0.15). The obtained results show that the created marker is reasonably close to the
score but involves a much simpler process.
Finding a better way to interpret the output of the classifier is left as future work. An improvement is required because values near 0.5 are not determined to be identified as AD or Control (equal
“probabilities”). The policy chosen in this work is a first rough approach that doesn’t reflect the actual
power of the network. Using a fuzzy-logic encoder to obtain the overall diagnosis of a patient might
lead to a more accurate result. Determining whether the number given by the classier is related to the
severity of the disease is left as a future improvement. This task is particularly difficult since there are
no ground truth measurements to corroborate the information due to the current psychological testing
methods. Anyway, although we adopted a per-trial classification approach, it’s easy to think that the
overall diagnosis may be related to a measurement extracted from the entire test and not from a single
trial. As shown before, even with the policy used on this work and simply using the mean of the scores
or the “controlling” label, this network behaved as expected.
References
Li Deng and Dong Yu. Deep learning: Methods and applications. Foundations and Trends in Signal
Processing, 7(3–4):197–387, 2014.
Gerardo Fernández, Pablo Mandolesi, Nora P Rotstein, Oscar Colombo, Osvaldo Agamennoni, and
Luis E Politi. Eye movement alterations during reading in patients with early alzheimer disease, eye
movement behavior in alzheimer disease patients. Investigative ophthalmology & visual science, 54
(13):8345–8352, 2013.
Gerardo Fernández, Jochen Laubrock, Pablo Mandolesi, Oscar Colombo, and Osvaldo Agamennoni.
Registering eye movements during reading in alzheimer’s disease: Difficulties in predicting upcoming
words. Journal of clinical and experimental neuropsychology, 36(3):302–316, 2014a.
Gerardo Fernández, Diego E Shalom, Reinhold Kliegl, and Mariano Sigman. Eye movements during
reading proverbs and regular sentences: The incoming word predictability effect. Language, Cognition
and Neuroscience, 29(3):260–273, 2014b.
Gerardo Fernández, Liliana R Castro, Marcela Schumacher, and Osvaldo E Agamennoni. Diagnosis of
mild alzheimer disease through the analysis of eye movements during reading. Journal of integrative
neuroscience, 14(01):121–133, 2015a.
Gerardo Fernández, Marcela Schumacher, Liliana Castro, David Orozco, and Osvaldo Agamennoni.
Patients with mild alzheimer’s disease produced shorter outgoing saccades when reading sentences.
Psychiatry research, 229(1):470–478, 2015b.
Gerardo Fernández, Salvador Guinjoan, Marcelo Sapognikoff, David Orozco, and Osvaldo Agamennoni.
Contextual predictability enhances reading performance in patients with schizophrenia. Psychiatry
research, 241:333–339, 2016a.
Gerardo Fernández, Facundo Manes, Luis E Politi, David Orozco, Marcela Schumacher, Liliana Castro,
Osvaldo Agamennoni, and Nora P Rotstein. Patients with mild alzheimer’s disease fail when using
their working memory: Evidence from the eye tracking technique. Journal of Alzheimer’s Disease, 50
(3):827–838, 2016b.
Gerardo Fernández, Marcelo Sapognikoff, Salvador Guinjoan, David Orozco, and Osvaldo Agamennoni.
Word processing during reading sentences in patients with schizophrenia: evidences from the eyetracking technique. Comprehensive psychiatry, 68:193–200, 2016c.
10
Marshal F Folstein, Susan E Folstein, and Paul R McHugh. mini-mental state: a practical method for
grading the cognitive state of patients for the clinician. Journal of psychiatric research, 12(3):189–198,
1975.
Philip S Holzman, Leonard R Proctor, Deborah L Levy, Nicholas J Yasillo, Herbert Y Meltzer, and
Stephen W Hurt. Eye-tracking dysfunctions in schizophrenic patients and their relatives. Archives of
general psychiatry, 31(2):143–151, 1974.
William G Iacono, Margaret Moreau, Morton Beiser, Jonathan AE Fleming, and Tsung-Yi Lin. Smoothpursuit eye tracking in first-episode psychotic patients and their relatives. Journal of Abnormal Psychology, 101(1):104, 1992.
Jennifer L Kellough, Christopher G Beevers, Alissa J Ellis, and Tony T Wells. Time course of selective
attention in clinically depressed young adults: An eye tracking study. Behaviour research and therapy,
46(11):1238–1243, 2008.
Reinhold Kliegl, Antje Nuthmann, and Ralf Engbert. Tracking the mind during reading: the influence
of past, present, and future words on fixation durations. Journal of experimental psychology: General,
135(1):12, 2006.
Rongjian Li, Wenlu Zhang, Heung-Il Suk, Li Wang, Jiang Li, Dinggang Shen, and Shuiwang Ji. Deep
learning based imaging data completion for improved brain disease diagnosis. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2014, pages 305–312. Springer, 2014.
Siqi Liu, Sidong Liu, Weidong Cai, Sonia Pujol, Ron Kikinis, and Dagan Feng. Early diagnosis of
alzheimer’s disease with deep learning. In Biomedical Imaging (ISBI), 2014 IEEE 11th International
Symposium on, pages 1015–1018. IEEE, 2014.
Guy McKhann, David Drachman, Marshall Folstein, Robert Katzman, Donald Price, and Emanuel M
Stadlan. Clinical diagnosis of alzheimer’s disease report of the nincds-adrda work group* under the
auspices of department of health and human services task force on alzheimer’s disease. Neurology, 34
(7):939–939, 1984.
Eneida Mioshi, Kate Dawson, Joanna Mitchell, Robert Arnold, and John R Hodges. The addenbrooke’s
cognitive examination revised (ace-r): a brief cognitive test battery for dementia screening. International journal of geriatric psychiatry, 21(11):1078–1085, 2006.
Keith Rayner. Eye movements in reading and information processing: 20 years of research. Psychological
bulletin, 124(3):372, 1998.
Deborah Riby and Peter JB Hancock. Looking at movies and cartoons: eye-tracking evidence from
williams syndrome and autism. Journal of Intellectual Disability Research, 53(2):169–181, 2009.
Jürgen Schmidhuber. Deep learning in neural networks: An overview. Neural Networks, 61:85–117, 2015.
Núria Sebastián-Gallés. LEXESP: Léxico informatizado del español. Edicions Universitat Barcelona,
2000.
Heung-Il Suk and Dinggang Shen. Deep learning-based feature representation for ad/mci classification. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2013, pages 583–590.
Springer, 2013.
Heung-Il Suk, Seong-Whan Lee, Dinggang Shen, Alzheimer’s Disease Neuroimaging Initiative, et al.
Hierarchical feature representation and multimodal fusion with deep learning for ad/mci diagnosis.
NeuroImage, 101:569–582, 2014.
G Waldemar, B Dubois, M Emre, J Georges, IG McKeith, M Rossor, P Scheltens, P Tariska, and
B Winblad. Recommendations for the diagnosis and management of alzheimer’s disease and other
disorders associated with dementia: Efns guideline. European Journal of Neurology, 14(1):e1–e26,
2007.
11
| 9cs.NE
|
THE MUTUAL INFORMATION OF LDGM CODES
arXiv:1707.04413v2 [cs.IT] 26 Jul 2017
JAN VAN DEN BRAND AND NOR JAAFARI
A BSTRACT. We provide matching upper and lower bounds on the mutual information in noisy reconstruction of parity check
codes and thereby prove a long-standing conjecture by Montanari [IEEE Transactions on Information Theory 2005]. Besides
extending a prior concentration result of Abbe and Montanari [Theory of Computing 2015] to the case of odd check degrees,
we precisely determine the conjectured formula for code ensembles of arbitrary degree distribution, thus capturing a broad
class of capacity approaching codes.
1. I NTRODUCTION
Sparse random binary matrices provide a natural way of encoding messages without exhausting the transmission rate.
Let m be a number larger than the blocklength of a message ξ ∈ Fn2 . By choosing a random generator matrix A over
the field Fm×n
, we obtain a codeword x by simple matrix multiplication x = Aξ. Given A and a noisy observation x̃
2
obtained from the binary memoryless symmetric (BMS) channel, we can most likely recover ξ by solving the system
of linear equations. That is, provided m is sufficiently large and the matrix A imposes adequate redundancy averaged
over the bits of ξ. Properly structured sparse random matrices A induce the class of Low Density Generator Matrix
(LDGM) codes. LDGM codes have been known for many decades, and although we can describe them in very few
lines, to date the status of research has hardly advanced, partly due to the fact that while simplifying the encoding
and decoding process, sparsification also severely exacerbates the analysis. This is unsurprising but remarkable, as we
know that at least for some code constructions these codes perform remarkably well. This paper aims to break the first
ground by proving a precise formula for the mutual information of LDGM codes which was previously conjectured
by Montanari [34].
1.1. Optimal codes on graphs. As may be expected, structured sparse code ensembles are readily constructed from
bipartite graphs. Also known as tanner graphs (factor graphs), they consist of variable nodes representing bits of a
signal on the left-hand side and check nodes (factor nodes) representing parity check equations on the right-hand side.
In a more general context factor graphs are used to model constraint satisfaction problems (CSPs) by having factor
nodes impose constraints on participating variable nodes. In LDGM codes each variable node participates in at least
one parity equation via its neighboring check nodes: If a check node constrained to xj ∈ F2 is adjacent to k bits
ξ1 , . . . , ξk , then codewords need to satisfy ξ1 + . . .+ ξk = xj in F2 . The intersection of codewords satisfying all parity
check equations modeled by the graph form a linear code.
The crux is that while local interactions are constructed in a simple fashion, the global structure of the code has to
perform in a complex interplay, as to ensure efficient coding and decoding. Importantly, the placement and number of
check nodes can be chosen in a sophisticated way as to keep encoding and decoding complexity low, while maintaining
a sufficient amount of redundancy. In a nutshell these so-called standard code ensembles achieve their performance
by imposing a specific degree distribution on the nodes of the factor graph. For example, we constrain check nodes to
perform a parity check on a constant number k of variable nodes with degree distribution D. The mutual information
of (D, k)-code ensembles has been rigorously studied in previous work. In a pioneering paper [34] Montanari derives
an upper bound on the mutual information for even k and conjectures the bound to be tight. In subsequent work [2]
Abbe and Montanari were able to prove the existence of a limit for the mutual information in the same scenario. Our
result comprehensively determines this limit for the mutual information of (D, k)-code ensembles for both even and
odd k.
This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may
no longer be accessible.
1
1.2. The mutual information. In the last decade there has been a critical endeavor to analyze such standard code
ensembles with respect to error-free decodeability. The mutual information I(X, Y ) captures how much information
the output of the channel Y contains about the input of the channel X and thus provides the essential measure to
quantify information-theoretic limits of decodeability. As it artlessly entails bounds on error probabilities the mutual
information proves to be key in many related areas of decoding noisy signals [8, 9, 10, 26], most importantly in the
analyses of random linear codes [38]. In the setting of BMS channels with noise η ∈ (0, 1/2) each bit is independently
correctly transmitted with probability 1 − η and flipped with probability η. Computing the exact mutual information in
this configuration is a highly non-trivial task and despite a substantial amount of research [18, 23, 24, 34], one is usually
merely able to provide bounds and occasionally tailor formulas to individual scenarios. In a general case analysis for
LDGM codes with given variable degrees Montanari [34] derives an upper bound on the mutual information, subject
to the condition that check degrees satisfy a convexity assumption. He conjectures the bound to be sharp, as it matches
explicit formulas from auspicious but non-rigorous calculations in the spin-glass theory of statistical mechanics.
In this paper, building upon the indispensable groundwork by Montanari, we establish the Aizenman-Sims-Starr [5]
cavity computation for standard code ensembles to derive a matching lower bound and with that prove the conjectured bound to be tight. Furthermore, we introduce a new technique to drop the assumptions on the variable degree
distribution and extend the results a comprehensive class of standard code ensembles.
1.3. Results. The following theorem proves Montanaris long-standing conjecture from [34]. Particularly, it is the
first result to precisely determine the mutual information in random LDGM codes with given variable degrees without
imposing restrictions on the degree distribution or the magnitude of noise.
The predicted formula comes in the form of a stochastic fixed-point equation. To state our Theorem denote by
P0 ([−1, 1]) the set of probability distributions on [−1, 1] with mean zero and let π ∈ P0 ([−1, 1]). Fix a degree
distribution D, a number k > 0 and let (θi,j )i≥0,1≤j≤k be independently identically distributed samples from π, and
γ be chosen from D. Write Be(η) for the outcome of a random Bernoulli experiment with parameter η. Further, let
Λ(x) = x ln x for x ∈ (0, ∞) and Λ(0) = 0. With a sequence J 1 , J 2 , . . . of independent copies of J = 1 − 2Be(η),
we let
i
i
h P
h
Qk
Qk−1
Qγ
L(k, D, η) = 12 E Λ
.
− k−1
j=1 θ 0,j
j=1 θ a,j
a=1 1 + J a σ
σ∈{±1}
k E[γ] E Λ 1 + J
Theorem 1.1. Let Cn be a random LDGM-code with blocklength n, variable degree distribution D and check degree
k. Let X be a message chosen uniformly at random from Cn . If Y is the message obtained by passing X through a
memoryless binary symmetric channel with error probability η > 0 then
E[γ]
1
((1 − η) ln(1 − η) + ln 2 + η ln(η)) + ln 2.
lim I(X, Y ) = − sup L(k, D, η) +
n→∞ n
2k
P0 ([−1,1])
The argument we develop to prove the mutual information in standard graph ensembles is reasonably general. We
expect it to extend to LDPC code ensembles and similar problems of conditional random fields. Possibly our approach
may facilitate simpler proofs to the analyses of spatially coupled codes. It is important to mention that while spatial
coupling was invented to engineer error-correcting codes, the technique is now applied beyond the context of coding
theory [4, 20].
In Section 3, we will therefore actually prove a general version of the main theorem to include a broader class of factor
graph ensembles encompassing many problems related to conditional random fields. Theorem 1.1 deals with its most
natural application being the problem of noisy reconstruction in standard LDGM code ensembles.
1.4. Background and related work. Following Shannon’s work, early codes based on algebraic constructions were
analyzed, only to realize that these codes do not saturate the capacity limit by a far margin. Progress stalled in the
subsequent decades until the early 90s. The introduction of Turbo codes [12] reignited interest, as Turbo codes were
able to deliver performance close to the Shannon limit. Ensuing generalizations uncovered the power of codes based
on graphs as Low Density Parity Check (LDPC) codes, originally put forward in Robert Gallagher’s PhD thesis in
1962, were rediscovered. Having been neglected for quite some time, LDPC codes reemerged to a broad audience,
when a series of papers by Luby, Mitzenmacher, Shokrollahi, Spielman and Steman [28, 29, 30, 31] followed up by
work of Richardson, Shokrollahi and Urbanke [36, 37] proved that Gallagher’s parity check codes perform at rates
close to Shannon capacity and can be designed such that efficient decoding is possible. Crucially, in constructing
codes that approach capacity one has to keep in mind the performance of the decoder. Recently, parity check codes
have become widely used and successfully implemented in the context of satellite communication, WiFi transmission
2
and data protocols. At the same time significant progress has been made with regard to the design of optimal codes,
that not only achieve capacity but also allow for efficient coding and decoding. This is usually achieved by prescribing
a degree distribution on the variable nodes. Driven by the success of spatial coupling, the analysis of standard code
ensembles has gathered tremendous momentum. Nonetheless, the construction and analysis of spatially coupled codes
in particular is notoriously complicated.
In the analysis of linear codes, Sourlas’ work [40, 41, 42] dating back to the early 90s provided a crucial link to the
physics theory of spin systems that would betoken the path to numerous fruitful results. Many methods therefore base
on developments in the rigorous theory of mean field spin glasses. So far the most promising analyses of standard
LDGM code ensembles utilize Guerra and Toninelli’s interpolation method [19] to provide general bounds on the
inference threshold for graphical models. Subject to the condition that the generating function of the left degree distribution is convex, Abbe and Montanari [2] show that the entropy of the transmitted message conditional to the received
one concentrates around a well defined deterministic limit. In a previous work by Montanari [34] the interpolation
method was employed to lower bound the entropy by an asymptotic expression derived in [16], which was obtained
using heuristic statistical mechanics calculations. In a recent paper by Coja-Oghlan, Krkazala, Perkins and Zdeborová
[14] the entropy was derived for Erdős-Rényi type LDGM graph ensembles. However, these are merely fragmentarily
applicable for error-correcting codes, because the corresponding adjacency matrices are unstructured and may exhibit
empty columns.
2. P ROOF O UTLINE
In this section we outline the pillars of our proof. To derive the mutual information in noisy observations for a broad
class of random factor graph models, we make slightly more general assumptions than needed for codes. As such our
result includes a number of problems related to conditional random fields. The teacher-student model can be viewed
as a natural generalization of the retrieval problem in BMS-distorted LDGM codes. Instead of {±1} we consider an
arbitrary finite set Ω of possible bit values and generalize the parity check constraints to a finite set Ψ of positive weight
functions ψ : Ω k → (0, 2) for some fixed k > 1. At the basis of Ψ is a probability space (Ψ, p) with a prior distribution
p on the weight functions. We write ψ for a random choice from p. Further, we specify unweighted factor graphs
G = (V, F, (∂a)a∈F ) by their bipartition V, F into variable nodes V and check nodes F as well as their neighborhood
structure (∂a)a∈F , ∂a ∈ V k . A (weighted) factor graph additionally carries weight functions ψa : Ω ∂a → (0, 2)
on each factor node a ∈ F that locally evaluate signals (assignments) σ ∈ Ω V on the variable nodes of the graph
G = (V, F, (∂a)a∈F , (ψa )a∈F ). For a set of variable nodes of size n we write V = Vn and omit the index if it is
apparent from the context. Commonly, if ℓ is an integer, we write [ℓ] = {1, . . . , ℓ} and identify Vn with the set [n].
If a is a check node we denote by ∂a its neighborhood in V k and by (∂1 a, . . . , ∂k a) the vector of its k neighbors in
ascending order. To exert supplementary methods and results from [14], throughout the paper we require p to satisfy
two assumptions SYM and POS that are on the one hand easily verified for the class of LDGM codes and a number
of related applications and on the other hand directly imply the assumptions from [14]. To this end let P(Ω) denote
the set of probability distributions on Ω. Further, let P∗2 (Ω) denote the set of all probability measures on P(Ω) whose
mean corresponds to the uniform distribution.
P
SYM: Let ξ = |Ω|−k τ ∈Ω k E[ψ(τ )]. For all σ ∈ Ω k we have E[ψ(σ)] = ξ.
POS: For all π, π ′ ∈ P∗2 (Ω) the following is true. With µ1 , µ2 , . . . chosen from π, µ′1 , µ′2 , . . . chosen from π ′
and ψ ∈ Ψ chosen from p, all mutually independent, we have
P
h P
Qk
Qk
′
µ
(τ
)
ψ(τ
)
µ
(τ
)
+
(k
−
1)Λ
ψ(τ
)
E Λ
k
k
i=1 i i
τ ∈Ω
i=1 i i
τ ∈Ω
P
i
Qk
′
−kΛ
≥ 0.
τ ∈Ω k ψ(τ )µ1 (τ1 )
i=2 µi (τi )
Linear codes satisfy both SYM and POS as we will see in Section 3. The general story now goes as follows. The
teacher chooses a ground truth σ ∗ ∈ Ω V uniformly at random that he finds himself unable to directly convey to his
students. He uses the teacher-student model in which he may set up a random (D, k)-graph G∗ = G∗ (n, p, σ ∗ )
on n variable nodes and M = M (D, k, n) check nodes, where parity checks are chosen proportionally to the local
evaluation, that is
P[ψa = ψ] ∝ p(ψ)ψ(σ ∗ (x1 ), . . . , σ ∗ (xk ))
3
for x1 , . . . , xk ∈ ∂a ⊂ V, ψ ∈ Ψ.
The students get to see the random graph G∗ but not the ground truth and stand before the task of deciphering as much
information about σ ∗ as possible. Thus the limit on the amount of deductable information is quantified by the mutual
information I(σ ∗ , G∗ ).
(π)
(π)
Theorem 2.1. With γ chosen from D, ψ 1 , ψ 2 , . . . ∈ Ψ chosen from p, µ1 , µ2 , . . . chosen from π ∈ P∗2 (Ω) and
h1 , h2 , . . . ∈ [k] chosen uniformly at random, all mutually independent let
i
h
P
Qγ P
Q
(π)
1
B(D, π) = |Ω|
E ξ −γ Λ
σ∈Ω
j6=hi µki+j (τj )
i=1
τ ∈Ω k 1{τhi = σ}ψ i (τ )
i
h P
Qk
(π)
(τ
)
.
E[γ]
E
Λ
ψ(τ
)
µ
− k−1
k
j
τ ∈Ω
j=1 j
kξ
Let (σ ∗ , G∗D ) be an assignment/factor graph pair from the teacher-student model. Then
1
E[γ] X
lim I(σ ∗ , G∗D ) = − sup B(D, π) + ln |Ω| +
E[Λ(ψ(τ ))].
n→∞ n
kξ|Ω|k
π∈P∗2 (Ω)
k
τ ∈Ω
The mutual information in the teacher-student model has been previously derived for Erdős-Rényi type LDGM graph
ensembles in [14]. We include standard (D, k)-graph ensembles G∗D by a delicate extension of their techniques to a
modified Poisson approximation inspired by Montanaris Multi-Poisson ensembles [34]. The technical proofs involved
will be carried out for arbitrarily precise approximations G∗α,β,D of random (D, k)-graphs G∗D and then bridged to the
original model by way of a coupling. The following Proposition establishes the precision of our approximative model.
Proposition 2.2. Let Gn = G∗α,β,D and G′n = G∗D . There is a bounded function f : [0, 1) × (0, ∞) → R and a
coupling of Gn and G′n , such that if we add the missing O(αn) check nodes in Gn the expected number of check
nodes that have a different neighborhood in G′n is O(f (α, β)).
In other words, we can obtain a random (D, k)-graph G∗D by first generating a Multi-Poisson approximation G∗α,β,D
and then rewiring a small number of edges. We prove Proposition 2.2 in Section 3 by coupling G∗α,β,D to the exact
instance G∗D chosen from the configuration model. We track the incremental influence of small perturbations during
the construction process and prove that the approximation error is negligible. The coupling is similar to the procedure
from [34], yet by introducing the additional approximation parameter β we achieve an error that is asymptotically
independent of n.
To outline the proof of Theorem 2.1 suppose that (σ ∗ , G∗D ) is a jointly distributed tuple from the teacher-student
model, where the underlying factor graph is a random (D, k)-graph. As the mutual information can be expressed by
the marginal entropies, we can write
I(σ ∗ , G∗D ) = H(σ ∗ ) − H(σ ∗ |G∗D ).
While H(σ ∗ ) is easily calculated the derivation of H = H(σ ∗ |G∗D ) requires a more sophisticated approach, making
use of several methods from the physicist’s toolbox for sparse mean-field spin-glasses. Apart from inherent properties
of the teacher-student model we will prove and make use of the so called Nishimori property from physics. This
property allows us to identify the reweighed a posteriori measure induced by the graph, commonly named the Gibbs
measure µG∗D , with samples from the the ground truth. Results from [14] imply that given the Nishimori property we
can perform slight perturbations on µG∗D that without significant impact on the mutual information allow us to factorize
its marginals. This novel pinning technique is the crucial new ingredient that facilitates the proof of Montanaris
conjecture.
Therewith we have layed a foundation to derive an upper bound on the conditional entropy, i.e., a lower bound on the
mutual information, by means of the so-called Aizenman-Sims-Starr scheme. The proof is carried out in Section 4. In
essence the procedure boils down to estimating the change in entropy if we go from a model with n variable nodes to
one with n + 1 variable nodes (as indicated in Figure 1).
Denote by ∆H (n) the change in conditional entropy when going from n to n + 1.The representation n1 Hn =
Pn−1
Pn−1
1
1
j=0 (Hn+1 − Hn ) = n
j=0 ∆H (n) clearly implies that if ∆H (n) converges, then its limit is also the limit
n
of Hn . Unfortunately, we will not be able to compute the limit directly, but we can settle by using the representation
to obtain an upper bound on the entropy in the limit
lim sup Hn ≤ lim sup ∆H (n).
n→∞
n→∞
4
1
2
...
...
n
n+1
F IGURE 1. A change of entropy is induced by introducing another variable into the random factor
graph model.
This is achieved by generating our graphs from a random graph process and rigorously coupling the underlying models. The result of an explicit calculation carried out in Section 4 is that ∆H (n) can itself be upper bounded by the
conjectured formula. This translates into the following lower bound on the mutual information.
Proposition 2.3. We have
E[γ] X
1
E[Λ(ψ(τ ))].
lim inf I(σ ∗ , G∗D ) ≥ − sup B(D, π) + ln |Ω| +
n→∞ n
kξ|Ω|k
π∈P∗2 (Ω)
k
τ ∈Ω
In Section 5, using the Guerra-Toninelli interpolation method we derive a matching lower bound on the conditional
entropy. We interpolate between the original model and a much simpler graph model, where the Gibbs measure is more
easily understood and we can effortlessly verify that the entropy coincides with the conjectured formula. The random
graph approximation is assembled in a sequence of layers, such that depending on the approximation parameter β > 0
each layer contributes Po(β) many check nodes. We refine the interpolation argument by segmentation into layers, as
portrayed by Figure 2.
s
s+1
F IGURE 2. The interpolation scheme for k = 3 during t ∈ (0, 1) in layer s, originally consisting of
Po(β) = 2 factor nodes (dashed neighborhoods). One of two k-ary factor nodes in layer s is split
into k unary nodes.
Beginning with the surface layer of G1 = GD we split its k-ary factor nodes one at a time, replacing them by unary
nodes with weight functions simulating the complex underlying structure. We thus peel apart the intricate composition
of GD into a forest G0 of unary variable components. By independently splitting each k-ary factor node of layer
s with probability t ∈ [0, 1], sequentially for each layer s, we ensure a continuous interpolation between G1 and
G0 . Ultimately, we aim to show that the entropy in the latter model does indeed upper bound the original conditional
∂
entropy. This amounts to controlling ∂t
Hn (Gt ) within the interpolation interval and involves yet another clever
coupling argument, carried out in full detail in section 5. Taken as a whole, because the derivative with respect to t is
positive during the entire interpolation, we obtain the desired lower bound on the conditional entropy and as such the
following upper bound on the mutual information.
Proposition 2.4.
lim sup
n→∞
E[γ] X
1
I(σ ∗ , G∗D ) ≤ − sup B(D, π) + ln |Ω| +
E[Λ(ψ(τ ))].
n
kξ|Ω|k
π∈P∗2 (Ω)
k
τ ∈Ω
5
Together, Propositions 2.3 and 2.4 immediately imply the assertion for any planted CSP ensemble with given variable
degree distributions. Theorem 1.1 follows by explicitly calculating the quantities in the case of LDGM codes, and
verifying that SYM and POS hold. These simple but technical computations are put off to Section 3.
3. T HE T EACHER -S TUDENT
MODEL AND
S YMMETRY
In this section we set up the tools utilized in our proof. We formalize the notion of retrieval by constructively defining
the teacher-student model for (D, k)-graph ensembles as well as their Poissonian approximation. To this end we verify
a number of expedient properties brought with the model in our setting. We analyze the approximative random graph
model in terms of its free energy. This quantity is closely related to the mutual information and the main object of
interest in many physics models of spin-glasses. Moreover, to derive tight bounds from the expressions we obtain for
the free energy, we employ the Pinning Lemma from [14]. Finally we show how bounds on the free energy imply our
general main result. Following this section, the lower and upper bound are derived in Sections 4 and 5 respectively.
3.1. Preliminaries
P and Notation. With a factorQgraph G = (V, F, (∂a)a∈F , (ψa )a∈F ) we associate its partition function Z(G) = σ∈Ω n ψG (σ), where ψG (σ) = a∈F ψa (σ(∂1 a, . . . , ∂k a)). This gives rise to the Gibbs distribution
P
µG (σ) = Z(G)−1 ψG (σ). If X : Ω n → R is a functional, we write hX(σ)iG = σ∈Ω n µG (σ)X(σ), when referring
to its average with regard to the Gibbs distribution.
Additionally, we define the empirical distribution on the graph as
P
the average Gibbs marginal πG = |V |−1 x∈V δµG,x . If µ is a probability measure, we write µ⊗ℓ to denote its ℓ-fold
product measure. Moreover, we write µ(π) for a random sample from π ∈ P∗2 (Ω).
The standard graph ensembles follow a predefined left-degree distribution. Considering degree sequences chosen from
those distributions will be an expedient tool for the construction of random graphs from a given ensemble. We call
a degree sequence d finite, if supx d(x) < ∞. If not explicitly stated otherwise, we assume the number of variable
nodes in a given graph to be n and degree sequences to be finite with maximal degree ∆. For a real number x we let
(x)+ = max{0, x}.
Before we formally specify the random (D, k)-graph construction process, let us define a process that samples graphs
with specific left-degree sequences. This is commonly known as the configuration model.
Definition 3.1. Let d : [n] → N be a degree sequence. By Ḡn,d we denote the random unweighted factor graph
obtained from the following process.
EX1: Let F = ∅, s = 1. Initiate a vector δ s by setting δ s (x) = d(x) for x = 1, . . . , n.
EX2: While δ s 6= 0: Choose a random xs ∈ [n] from the measure ν s defined by
ν s (x) = P
δ s (x)
.
y∈[n] δ s (y)
Update δ s+1 = δ s − (1{xs })x∈[n] and set s = s + 1. If s − 1 is a multiple of k, then add a new check node
a to F and set its neighborhood to be {xs−k−1 , . . . , xs−1 }.
P
The resulting graph is Ḡn,d = ([n], F, (∂a)a∈F ) and satisfies the degree sequence as long as n divides x∈[n] d(x).
We also write Ḡd if the number of variable nodes n is apparent from the context.
Definition 3.2. If [n] allows for a partition (Vl )l with |Vl | = nD(l), a random unweighted factor graph from the
(D, k)-ensemble can be obtained by choosing such a partition (V l )l uniformly at random, setting d(x) = l for
x ∈ V l , l = 1, 2, . . .. We then write GD for a random choice from Ḡd .
At times it is beneficial to generate the random graph in a sequence of batches of factor nodes according to a specified
structure, rather than one node-socket at a time. This is facilitated by means of the following model that we will then
use to lay down the Poissonian approximation.
Definition 3.3. Given a degree sequence d as well as a vector m = (m1 , . . . , msmax ) with non-negative integer
entries, we define Gn,m,d to be the random graph obtained from the following experiment. For the sake of clarity we
omit the subscript n if it is unambiguous.
AP1: Let F = ∅, s = 1. Initiate a vector δ s by setting δ s (x) = d(x) for x = 1, . . . , n.
AP2: In layers s = 1, . . . , smax choose ms random neighborhoods ∂as,i from ν ⊗k
and add factor nodes
s
as,1 , . . . , as,ms with respective neighborhoods to the graph. Denote by ∇s (x) the number of times the variable node x has been chosen as a neighbor in round s and update δ s+1 = (δ s − ∇s )+ , s = s + 1.
6
Given a degree distribution D we say that a degree sequence d is a random D-partition of [n] if for a uniformly random
variable x ∈ [n] and any ℓ ≥ 0 the degrees satisfy P[d(x) = ℓ] = D(ℓ).
Definition 3.4. Given a degree distribution D and a vector m = (m1 , . . . , msmax ) with non-negative integer entries,
we let d be a random D-partition of [n] and thereby define Gm,D = Gm,d .
3.2. The approximation scheme. To obtain an approximation of GD we introduce two additional parameters α ∈
[0, 1), β > 0 that aid in laying down an appropriate approximation-step vector m.
P
Definition 3.5. When α, β and a degree sequence d are fixed let smax = ⌊(1 − α)β −1 k −1 x∈[n] d(x)⌋. Further,
let m1 , m2 , . . . be a sequence of independent Po(β) distributed numbers. We then obtain an (α, β)-approximation of
Ḡd by setting m = (m1 , . . . , msmax ) and letting Gα,β,d = Gm,d . Analogously we define Gα,β,D = Gm,D , where
we first choose a random D-partition d and generate m thereafter.
P
Note that the graph-generating procedure as stated in Definition 3.3 will fail if x∈V δ s (x) = 0 for some 1 ≤ s ≤
−1
smax
P. However, it will succeed w.h.p. as n tends to infinity because the probability of adding more than k (1 −
α) x∈V d(x) factor nodes during the construction tends towards zero. Particularly, as α, β tend to 0 this becomes
an arbitrarily close approximation of Ḡd , in the sense that we obtain Ḡd from Gα,β,d by rewiring O(1) node sockets.
This is enabled by the following observation.
Fact 3.6. With α, β fixed and s ∈ [smax ] let (δ s (v))v∈[n] be the degree sequence from Definitions 3.1 and 3.3. Let
c : [n] → Z be a bounded integer valued function on the variable nodes and let δ ′s (v) = (δ s (v) − c(v))+ . Consider
distributions ν s ∝ δ s and ν ′s ∝ δ ′s . Then
P
x∈V [c(x)]+
′
P
kν s − ν s kT V =
.
O(αn) − x∈V c(x)
P
P
Proof. We simply write out the total variation distance and bound x δ s (x) ≥ x δ smax (x) = Ω(αn) to obtain
!
X
X
δ
(v)
−
c(v)
δ
(v)
s
s
P
P
−P
(ν s (v) − ν ′s (v))+ =
kν s − ν ′s kT V =
δ s (x)
δ s (x) − x∈[n] δ c (x)
x∈[n]
x∈[n]
v∈[n]
v∈[n]
+
!
!
P
P
X c(v) x∈[n] δ s (x) − δ s (v) x∈[n] c(x)
X
c(v)
P
P
P
=
≤
.
c(x)
δ
(x)
−
(δ
x∈[n]
x∈[n] s
x∈[n] s (x) − c(x))
v∈[n]
+
v∈[n]
+
In the light of Lemma 3.6 consider generating an approximative graph Gα,β,D . If we independently generate the
missing O(αn) factor nodes in the same fashion as the exact procedure from Definition 3.1, we can verify that the
distribution of the resulting factor graph matches the distribution of GD after we rewire a finite number of factor
nodes. Importantly, this number is asymptotically independent of n.
Proof of Proposition 2.2. If β is sufficiently small, most layers will contain at most one check node, and very few will
contain a constant number of check nodes. For the sake of simplicity suppose that each layer consists of at most one
check node. The argument extends to any constant number of check nodes.
First observe that during the layer-wise creation of Gn in any step of layer s, if the approximative graph matches
the original graph except for C > 0 factor nodes, there are at most Ck different entries in the degree sequences.
That is kν s − ν ′s kT V = O( Ck
αn ) by Fact 3.6. As the degree sequence in the exact model updates immediately, the
approximative degree sequences adds up to k − 1 incremental deviations during the creation of one k-ary check node.
C
This results in a total variation distance of kO( (C+1)k
αn−k ) = O( αn ) for the choice of the next neighbor. We will now
couple the configuration model process in both graphs inductively. Suppose that during the process, we have an
optimal coupling such that both graphs differ in exactly i − 1 ≥ 0 check nodes. By the coupling Lemma both graphs
i
coincide with probability O( αn
), thus we can choose the same neighborhood for the next X i check nodes, where
i
). Hence, once the process is complete
X i is a geometrically distributed random variable with parameter pi = O( αn
the total number C F of different check nodes is the minimum number of independent random variables X 1 , X 2 , . . .
PC F
i
X i ≥ Ω((1 − α)n). As n tends
) and i=1
such that X i is chosen from the geometric distribution with pi = O( αn
Xi
to infinity, the random variables αn can be arbitrarily well approximated by random variables Y i chosen from the
7
PC F
exponential distribution with parameter p̂i = O(i), such that i=1
Y i ≥ Ω( 1−α
α ). As a consequence C F can be
modeled as a function of α and is asymptotically independent of n.
We have assumed that each layer s = 1, . . . , smax consists of at most one check node. The argument generalizes to
any sequence of layers with independent Po(β) many check nodes. Because the tails of the Poisson distribution show
sub-exponential decay, we can guarantee that kνs − νs′ k = O( C+β
αn ), if in layer s − 1 the coupling differs in at most C
check nodes. By induction we can write C F as a function of α and β.
3.3. The Teacher-Student Model. To derive the mutual information between the transmitted message and the received one, we analyze the free energy density within the (D, k)-teacher-student model. To this end, we generalize the
scheme from [14] to arbitrary degree distributions. We define a random factor graph G∗ via the following construction.
TCH1: Choose σ ∗ : [n] → Ω uniformly at random.
TCH2: Choose a random graph G = ([n], F, (∂a)a∈F ) from Gα,β,D .
TCH3: For each check node a ∈ F choose a random ψ a in Ψ from the distribution
p(ψ)ψ(σ ∗ (∂1 a), . . . , σ∗ (∂k a))
P[ψ a = ψ] = P
′
∗
∗
ψ ′ ∈Ψ ψ (σ (∂1 a), . . . , σ (∂k a))
We let G∗ = G∗ (σ ∗ ) = ([n], F, (∂a)a∈F , (ψ a )a∈F ) denote the resulting (weighted) random factor graph
and omit the parameters α, β and n for the sake of readability. We write G∗D for the limit α, β → 0.
3.4. Pinning and Symmetry. Another indispensable tool that we make use of is pinning. As we derive bounds
on the free energy E[ln Z(G∗ )] we are confronted with drawing k-tuples of bit-assignments chosen from the Gibbs
measure. Due to the confinement by weight functions certain configurations of tuples are favored within the Gibbs
measure. Hence, in most cases variable node assignments sampled from the Gibbs measure, where the nodes are
connected through a common check node will be highly correlated by means of the weight function and thus far from
independent. This correlation ought to persist for any nodes within a finite distance in the graph. On the other hand
if we choose two variable nodes at random they are typically far apart, which is why we can hope that their Gibbs
measure is almost a product measure. This is captured by the notion of ε-symmetry, or more generally (δ, ℓ)-symmetry.
Definition 3.7. Let µG be a probability measure on Ω n . For ℓ ≥ 2 and {x1 , . . . , xℓ } ⊂ [n] let µG,x1 ,...xℓ denote the
marginal distribution of an ℓ-tuple chosen from µG . We say that µG is (δ, ℓ)-symmetric, if
X
1
kµG,x1,...,xℓ − µG,x1 ⊗ · · · ⊗ µG,xℓ kT V < δ.
ℓ
n
x1 ,...,xl ℓ∈[n]
If µG is (ε, 2)-symmetric, we simply speak of ε-symmetry.
Fortunately, the following Pinning Lemma from [14] guarantees a degree of ε-symmetry in exchange for small modifications in the original measure.
Lemma 3.8. [14] For any ε > 0 there is T = T (ε, Ω) > 0 such that for every n > T and every probability measure
µ ∈ P(Ω n ) the following is true. Obtain a random probability measure µ̌ ∈ P(Ω n ) as follows: Draw a sample
σ̌ ∈ Ω n from µ, independently choose a number θ ∈ (0, T ) uniformly at random and obtain a random set U ⊂ [n]
by including each i ∈ [n] with probability θ/n independently. The measure µ̌ defined by
µ̌(σ) =
µ(σ)1{∀i ∈ U : σi = σ̌ i }
,
µ({τ ∈ Ω n : ∀i ∈ U : τi = σ̌ i })
σ ∈ Ωn
is ε-symmetric with probability at least 1 − ε.
When applying the procedure from Lemma 3.8 to the Gibbs measure of a graph the perturbation equates to pinning
the color of each variable node i in U to its coloring under σ̌, i.e. adding a constraint ψ i (σ) = 1{σ = σ̌ i } to those
variable nodes.
Definition 3.9. Given a factor graph G, a subset U ⊂ [n] of its variable nodes and an assignment σ̌ ∈ Ω n obtain
GU,σ̌ from G by adding unary check nodes ax with weights ψax (τ ) = 1{σ̌(x) = τ } to each variable node x ∈ U .
Moreover, for T ≥ 0 let U = U (T ) ⊂ V be a random subset of vertices generated by first choosing θ ∈ [0, T ]
uniformly at random and then including each variable node v ∈ [n] in U with probability θ/n. If σ ∈ Ω n is chosen
independently and uniformly at random from Ω for each variable node, we let GT = GU ,σ .
8
The following Lemma establishes, that the perturbations performed in the pinning procedure yield an ε-symmetric
measure with probability at least 1 − ε.
Lemma 3.10 ([14, Lemmata 2.8 and 3.5]). Let µG be a Gibbs measure on Ω n .
1. For any ε > 0 there is a T0 = T0 (Ω, ε) such that for all T > T0 and sufficiently large n the meausre µGT is
ε-symmetric with probability at least 1 − ε.
2. Moreover, for any l ≥ 3, δ > 0 there is ε > 0 such that if µGT is ε-symmetric, then µ is (δ, l)-symmetric.
Also note that the pinning procedure has merely constant additive impact on the free energy. In the factor graph model
all weight functions on check nodes are positive and all variable degrees are at most ∆. Such being the case, pinning
a single variable node changes the free energy of the graph by at most O(1). Because the expected number of pinned
variable nodes is at most T , we have the following Lemma.
Lemma 3.11. If G is a random factor graph from the (D, k)-ensemble, then
E[ln Z(GT )] = E[ln Z(G)] + O(1).
Lemma 3.10 together with the so-called Nishimori property will be key tools utilized in the proof. The Nishimori
property stands testament to the fact that if the prior p satisfies a certain symmetry condition, the Gibbs measure in
the teacher student model µG∗ does indeed resemble the a posteriori distribution of our planted assignment σ ∗ given
the graph outcome G∗ . We will use this argument in the same way as done in the proofs of [14]. Nevertheless, the
symmetry condition from [14] imposes a much weaker restriction on the prior p. While being a direct implication
of SYM, their condition merely yields mutual contiguity of both measures. In the class (Ψ, p) of models that covers
error-correcting codes transmitted over binary memoryless channels, we obtain the following stronger formulation.
Lemma 3.12. Suppose that p satisfies SYM. Then
P[σ ∗ = σ, G∗ (σ ∗ ) = G] = P[G∗ = G]µG (σ).
(1)
∗
In particular, the distribution of σ coincides with the Gibbs measure µG∗ .
Proof. If G is a random graph model and the event |F (G)| = M has positive probability, we write GM to be the
conditional random graph on M factor nodes. Observe that for any choice of M, σ and any event A
P[G∗M (σ ∗ ) ∈ A|σ ∗ = σ] =
E[ψGM (σ)1{GM ∈ A}]
.
E[ψGM (σ)]
(2)
To see (2) it suffices to write out the weights for any event
A = {(V, F, (∂a)a∈F , (ψa )a∈F : (ψa )a∈F ∈ W ⊂ Ψ F }.
Doing so, we get
P[G∗M (σ ∗ ) ∈ A|σ ∗ = σ] = P[GM = (V, F, (∂a)a∈F )] P[(ψ a )a∈F ∈ W |σ ∗ = σ]
= P[GM = (V, F, (∂a)a∈F )]ξ −M E[ψ(V,F,(∂a)a∈F ) (σ)1{(ψ a )a∈F ∈ W }]
=ξ −M E[ψGM (σ)1{GM ∈ A}]
and are left to verify that summing over all graphs G = (V, F, (∂a)a∈F ) on M factor nodes gives
X
E[ψGM (σ)] =
P[GM = G] E[ψG(σ) |G = G]
G
=
X
P[GM = (V, F, (∂a)a∈F )]
G
Y
E[ψ(σ(∂a)] = ξ M .
a∈F
Having established (2) let us now derive (1). If we write M for the distribution of F (G∗ ), then G∗M and G∗ have the
same distribution. Moreover, for any M we have
X
X ξ 1 E[ψG (σ)1{GM ∈ A}]
M
P[G∗M ∈ A] =
P[σ ∗ = σ] P[G∗M (σ ∗ ) ∈ A ∈ A|σ ∗ = σ] =
n
ξ
|Ω|
E[ψ GM (σ)]
n
n
σ∈Ω
σ∈Ω
X E[ψG (σ)] E[ψG (σ)1{GM ∈ A}]
E[Z(GM )1{GM ∈ A}]
M
M
=
.
=
E[Z(GM )]
E[ψGM (σ)]
E[Z(GM )]
n
σ∈Ω
9
Consequently, if we write σ for a sample from the Gibbs measure µG∗ , the joint distribution (σ, G∗ ) satisfies
1
ψGM (σ)
∗
P[σ = σ, GM (σ) = A] =
E
Z(GM )1{GM ∈ A} = P[σ ∗ = σ] P[G∗M |σ ∗ = σ]
E[Z(GM )]
Z(GM )
by multiplying the denominator and numerator with E[ψGM (σ)] and applying SYM in the final step.
The following lemma shows that applying the pinning process in any order preserves the Nishimori property.
Lemma 3.13. For any set of vertices U ⊂ [n] the following distributions on pairs of factor graphs and assignments
are identical.
(1) Choose σ (1) = σ ∗ , then choose G(1) = G∗U,σ(1) σ (1) and output σ (1) , G(1) .
(2) Choose G = G∗ . Choose σ (2) from µG , then G(2) = GU,σ(2) and output σ (2) , G(2) .
(3) Choose G(3) = G∗ (σ)U,σ . Then choose σ (3) from µG(3) and output σ (3) , G(3) .
(4) Choose G′ = G∗ and σ ′ from µG′ . Pin G(4) = GU,σ′ and choose σ (4) from µG(4) . Output σ (4) , G(4) .
Proof. By Lemma 3.12 the pairs (1) and (2) as well as (3) and (4) are identical. To see that (2) and (4) coincide, it
suffices to prove that
P[σ (2) = σ|G(2) ∈ A] = P[σ (4) = σ|G(4) ∈ A]
holds for any event A = {(G, σ̌) : G ∈ G}, where G is an arbitrary set of unpinned factor graphs and σ̌ ∈ Ω U is a
pinning of the variables in U . Denote by σ G a sample from the Gibbs measure µG and write σU for the restriction of
σ to U . Then
(2)
P[σ (2) = σ|G(2) ∈ A] = P[σ (2) = σ|G ∈ G, σ U = σ̌] = P[σ G = σ|G ∈ G, σ G,U = σ̌]
1
ψG (σ)1{σU = σ̌}
E P
=
1{σ G,U = σ̌}1{G ∈ G}
E[1{σG,U = σ̌}1{G ∈ G}]
τ ∈Ω n ψG (τ )1{τU = σ̌}
ψ (2) (σ)
(3)
= E[1{G(2) ∈ G}]−1 E G (2) 1{G(2) ∈ G} .
Z(G )
By pinning the same neighborhood
the evaluation in the teacher-student model remains the same and thus (3) equals
E[1{G(4) ∈ G}]−1 E Z(G(4) )−1 ψG(4) (σ)1{G(4) ∈ G} = P[σ (4) = σ|G(4) ∈ A].
3.5. Proof of the main result. We will now prove the conjectured formula by showing that the mutual information
per bit n1 I(σ ∗ , G∗D ) converges to the solution of a stochastic fixed-point equation. To state our result in a more general
setting we will write the convergence in terms of the free energy n1 E[ln Z(G∗D )]
Proposition 3.14. If SYM and POS hold, then lim supn→∞ − n1 E[ln Z(G∗D )] ≤ − supπ∈P∗2 (Ω) B(D, π).
We prove Proposition 3.14 in Section 5 by performing the Guerra-Toninelli interpolation between G1 = G∗D and a
forest of isolated variable nodes G0 . Taking account of the details that come along, this can be understood as simply
splitting each k-ary factor node into k unary factors with probability 1 − t, t ∈ [0, 1]. As − E[ln Z(G0 )] is easy
to compute, we can show that the resulting expression is also an upper bound for − E[ln Z(G1 )] by verifying that
∂/∂t E[ln Z(G1 )] is positive on (0, t).
Proposition 3.15. If SYM holds, then lim supn→∞ − n1 E[ln Z(G∗D )] ≥ − supπ∈P∗2 (Ω) B(D, π).
In Section 4 we prove the lower bound in Proposition 3.15 via the previously described Aizenman-Sims-Starr scheme.
That is we bound the difference between the free energy of Gn and Gn+1 . By coupling the joint distribution on a
large common subgraph G0 , we can compute the expected change in the free energy given by the additional constraints
going from G0 to the graphs Gn and Gn+1 . This is possible due to Lemma 3.8 as these constraints are generated
from the empirical Gibbs marginals on G0 which does not significantly differ from πGn or πGn+1 .
By writing out the mutual information, Propositions 3.14 and 3.15 translate into Propositions 2.3 and 2.4, thus revealing the main theorem.
Proof of Theorem 2.1. Theorem 2.1 is a corollary of Propositions 3.14 and 3.15. Writing out the mutual information
with Lemma 3.12 we get
X
X
µG (σ)
= H(σ ∗ ) − E[H(µG∗ )].
I(σ ∗ , G∗D ) =
P[G∗D = G]
µG (σ) ln
∗ = σ]
P[σ
σ
G
10
As σ chosen from µG∗ coincides with σ ∗ given G∗ , we obtain
ψG∗ (σ) ψG∗ (σ)
E[H(µG∗ )] = − E
= E ln Z(G∗ ) − E[hψG∗ (σ)iG∗ ]
ln
Z(G∗ )
Z(G∗ )
n E[γ] X
= E ln Z(G∗ ) − E ln ψG∗ (σ ∗ ) = E ln Z(G∗ ) −
E [Λ(ψ(τ ))] .
k|Ω|k
k
τ ∈Ω
∗
n
Because σ is chosen from the uniform distribution on Ω the assertion follows.
Theorem 2.1 readily implies Theorem 1.1 by verifying SYM and POS for the case of η-noisy (D, k)-LDGM code.
Proof of Theorem 1.1. Let Ω = {±1},
Q η ∈ (0, 1/2) and (Ψ, p) = ({ψ1 , ψ−1 }, p), where p is the uniform distribution
on Ψ and ψs (σ) = 1 + s(1 − 2η) i=1 σi for s ∈ Ω, ω ∈ Ω k . Clearly, SYM holds as E[ψ(σ)] = ψ1 (σ)/2 +
ψ−1 (σ)/2 = (1 ± (1 − 2η))/2 + (1 ∓ (1 − 2η))/2 = 1. To see POS write ψ = ψs for a uniform choice of
s ∈ Ω. Writing µ1 , µ2 , . . . for independent samples from π ∈ P∗2 (Ω) and µ′1 , µ′2 , . . . for independent samples from
π ′ ∈ P∗2 (Ω), for every l ≥ 1 and i ∈ [k] we obtain
!l
k
l
X
Y
Q
P
l
σµj (σ) ,
1 − σ∈Ω k ψ(σ) j=1 µj (σj ) = ((1 − 2η)s)
l
Q
P
1 − σ∈Ω k ψ(σ) j=1 µ′j (σj ) = ((1 − 2η)s)l
j=1
σ∈Ω
k
Y
X
j=1
l
P
Q
1 − σ∈Ω k ψ(σ)µi (σI ) j6=i µ′j (σj ) = ((1 − 2η)s)l
!l
σµ′j (σ)
σ∈Ω
X
!l
σµi (σ)
σ∈Ω
Y
j6=i
,
X
!l
σµ′j (σ)
σ∈Ω
.
P
P
Setting Xl = E[( σ σµ1 (σ))l ], Yl = E[( σ σµ′1 (σ))l ] and writing Λ in a logarithmic series expansion, it is clearly
sufficient to show that for any l ≥ 1 we have E[((1 − 2η)s)l ](Xlk + (k − 1)Ylk − kXl Ylk−1 ) ≥ 0. The case of odd l
is immediate. For even l, X and Y are non-negative and thus Xlk + (k − 1)Ylk − kXl Ylk−1 is non-negative.
4. T HE LOWER BOUND
In this section we perform the technical computations for the Aizenman-Sims-Starr scheme. For the remainder of
this section we let α ∈ (0, 1), β > 0 and D be arbitrary but fixed. Let G∗n be the random graph obtained from the
experiment TCH on n vertices and let G∗n+1 be chosen from TCH on n + 1 vertices. Moreover, we let G∗T,n and
G∗T,n+1 signify the respective graphs obtained after performing the pinning procedure from Definition 3.9.
Establishing the following proposition, by way of a generalization of an argument from [14], we will pave the way to
proving our lower bound.
Proposition 4.1. Let ∆T (n) = E[ln Z(G∗T,n+1 )] − E[ln Z(G∗T,n )]. Then
lim sup lim sup ∆T (n) ≤
T →∞
n→∞
sup
B(D, π).
(4)
π∈P∗2 (Ω)
Let us write f (n, T ) = oT (1) if limT →∞ lim supn→∞ |f (n, T )| = 0. Proposition 4.1 then immediately implies
n−1
1
1 X
1
∗
∗
E[ln Z(GT,n )] = E[ln Z(GT,1 )] +
∆T (N ) ≤ oT (1) + sup B(D, π).
n
n
n
π∈P∗2 (Ω)
N =1
Taking the lim sup over T → ∞ and subsequently over α, β → 0, we obtain Proposition 3.15.
4.1. The coupling. To prove Proposition 4.1 we construct a coupling of the two graphs in play by first sampling maximal common subgraph G0 from which we then obtain graphs G′ and G′′ to mirror G∗T,n and G∗T,n+1 respectively.
For T > 0 let θ be chosen uniformly from [0, T ]. Choose a random D-partition d of [n + 1] and set δ 1 = d on the
subdomain [n + 1] and δ 1 (n + 1) = 0. Denote by D a random sample from the distribution D.
11
Let
1−α
1−α
1−α
(n + 1) E D −
D,
nED
min
,
k
k
k
+
1−α
1−α
λ′ =
n E D − λ0 ,
λ′′ =
(n + 1) E D − λ0 ,
k
k
smax = ⌊β −1 λ0 ⌋, s′max = ⌊β −1 λ′ ⌋, s′′max = ⌊β −1 λ′′ ⌋, sb = ⌊β −1 D⌋.
λ0 =
Moreover, let γ 1 , γ 2 , . . . , γ smax , γ ′1 , γ ′2 , . . . , γ ′s′max , γ ′′1 , γ ′′2 , . . . , γ ′′s′′max and γ b,1 , . . . , γ b,sb be independently chosen
from Po(β).
CPL1: On n variables choose σ ∗n ∈ Ω n uniformly and obtain a weighted factor graph by performing the
following. Setting s = 1 and beginning with the empty graph consisting only of variable nodes
(1) for i = 1, . . . , γ s add a factor node as,i with neighborhood chosen from
Qk
j=1 δ s (xj )
P[∂as,i = (x1 , . . . , xk )] = P
(5)
Qk
y1 ,...,yk
j=1 δ s (yj )
and weight ψ as,i chosen from
P[ψ as,i = ψ] =
p(ψ)ψ(σ ∗n (∂as,i ))
.
Eψ [ψ(σ ∗n (∂as,i ))]
(6)
(2) Set δ s+1 = (δ s − ∇s )+ , where ∇s (x) counts the number of times x was drawn as a neighbor in round
s. Increase s = s + 1 and abort
s > smax .
Pwhen
smax
Po(β) check nodes
We obtain a graph that consists of s=1
CPL2: With probability θ/(n + 1) independently pin each variable node x to σ ∗n , i.e., to each vertex x ∈ [n]
attach a unary constraint node ψax with ψax (τ ) = 1{σ∗n (x) = τ }.
We identify the random graph resulting from CPL1 and CPL2 with G0 .
CPL1’: With the pair (σ ∗n , G0 ) from the former experiment perform another s′max rounds, that is for s =
smax + 1, smax + 2, . . . , smax + s′max and i = 1, . . . , γ ′s add new weighted factor nodes a′s,i from (5) and (6)
while updating δ s with each increment of s as in CPL1 (2).
CPL2’: Then independently pin each x ∈ [n] that is not yet pinned with probability θ/(n(n + 1 − θ)).
Identify the resulting graph with G′ .
CPL1”: With (σ ∗n , G0 ) and δ smax from the first experiment CPL1, CPL2 extend σ ∗n to σ ∗n+1 by choosing σ ∗n+1 (xn+1 ) independently and uniformly at random. Then — independently from the experiment
CPL1’,CPL2’ — from G0 create a graph G′′ by adding factor nodes as follows. Let s = smax .
(1) For i = 1, . . . , γ ′′s add a factor node cs,i with neighborhoods and weights from
P[ψ cs,i = ψ, ∂cs,i = xi1 , . . . , xik ] ∝ p(ψ)ψ(σ ∗n (xi1 , . . . , xik ))
k
Y
δ s (xij )
(7)
j=1
(2) Set δ s+1 = (δ s − ∇s )+ , where ∇s (x) denotes the number of times x was drawn as a neighbor in round
s. Halt when s > smax + s′′max − β −1 D.
Consequently set δ 1 = δ s and reset s = 1. Then
(a) For i = 1, . . . , γ b,s add a factor node bs,i with neighborhoods and weights from
P[ψ bs,i = ψ, ∂bs,i = xi1 , . . . , xik ]
∝1{n + 1 ∈ {i1 , . . . , ik }}p(ψ)ψ(σ∗n+1 (xi1 , . . . , xik ))
Y
δ s (xij )
(8)
ij 6=n+1
(b) Set δ s+1 = (δ s − ∇s )+ , where ∇s (x) denotes the number of times x was drawn as a neighbor in round
s. Halt when s > sb .
CPL2”: Finally, pin xn+1 to σ ∗n+1 (xn+1 ) with probability θ/(n + 1).
Identify the resulting graph with G′′ .
Note that by Lemma 3.12 we can replace (7) and (8) by first drawing the neighborhoods from ν ⊗k
s and ν s respectively,
then subsequently adding weights proportional to its evaluation (6).
12
Lemma 4.2. For sufficiently large n we have
d
d
G′ = G∗T,n and G′′ = G∗T,n+1 .
Proof. Clearly, for large enough n we have
1−α
1−α
1−α
λ0 = min
(n + 1) E D −
D,
nED .
k
k
k
The choice of neighborhoods is independent from the pinning process. Hence, we might as well switch the order
and perform CPL1, CPL2, CPL1’, CPL2’. In each of the experiments we add Po(β) many factor nodes from the
distribution (5). The neighborhoods are then chosen independently from everything but the planted coloring σ ∗ and the
neighborhood. Consequently, it suffices to compare the processes that generate these random neighborhoods. Because
the number of rounds that we perform in both CPL1, CPL2 and the model TCH on n vertices are ⌊k −1 (1 − α)n E D⌋
we can perfectly couple the occurences of each round. Moreover, in the process of performing CPL1’ and then CPL2’
each variable gets pinned independently with probability θ/n and therefore the procedure results in the graph G∗T,n .
It remains to prove the second distributional equality. On the event that D = 0 we see in the same fashion as with G′
that the procedure CPL1, CPL1” yields G∗n+1 . Suppose that D > 0. Even though we first consider xn+1 in the last
step, our choice of sb and (7) ensure that its degree in expectation over the outcome of our Poisson random variables
γ b,1 , . . . , γ b,sb is as if chosen from (5) on n+1 vertices. By (5), (7) and (8) so is the neighborhood distribution. Finally,
as with G′ , due to CPL2, CPL2” each variable node in G′′ is independently pinned with probability θ/(n + 1) and
therefore G′′ is distributed as GT,n+1 .
By Lemma 4.2 we now have a characterization of the left-hand side in (4)
Z(G′′ )
Z(G′ )
Z(G′′ )
= E ln
− E ln
∆ = E ln
Z(G′ )
Z(G0 )
Z(G0 )
(9)
which is approachable by calculating the contributions that are added to the partition function when going from G0 to
either G′ or G′′ . To assure that (9) mediates the assertion in (4) we establish the following fact.
Lemma 4.3. We have kG0 − G∗T k = o(1).
Proof. T is fixed and θ is chosen uniformly in [0, T ]. If U 0 denotes the random set that is pinned in G0 then each
variable in U 0 was chosen with probability θ. As T /n−T /(n+1) = o(1) the sets U and U 0 can be coupled such that
they coincide with probability 1 − o(1). Moreover, E[|E(G∗T )| − |m(G0 )|] = O(1) whereas the variance of |E(G∗T )|,
|E(G0 )| is at least Ω(n). Hence, all but O(1) factor nodes can be optimally coupled, such that the neighborhoods are
at most o(1) apart in total variation distance.
Before we proceed to calculate (9) we want to capture that our coupling is typically well-behaved.
Claim 4.4. Let U be the event that given G0 the process CPL2’ does not pin any additional vertices. Then
E ln(Z(G′ )/Z(G0 ))|G0 = oT (1) + E 1U ln(Z(G′ )/Z(G0 ))|G0 .
(10)
Let Y be the set of vertices belonging to the neighborhood chosen in CPL1’. Let m0 denote the number of factor
nodes added during CPL1’. There is ε = o(1) such that given T is sufficiently large and given G0 the event
Y = {kµG0 ,Y − ⊗y∈Y µG0 ,y kT V ≤ ε and |Y | = km0 }
occurs with probability at least 1 − ε.
Proof. To validate the first part, observe that for any graph G0 during CPL2’ each variable node is independently
pinned with probability θ/(n(n − 1 + θ)) ≤ T /(n(n − 1)). Thus P[U|G0 ] ≥ 1 − T /(n − 1) ≥ 1 − 2T /n. Equation
(10) is immediate as all weights added during the process are strictly positive.
Each neighborhood (x1 , . . . , xk ) in CPL1’ is chosen from a product measure proportional to a slight perturbation of
Qk
i=1 δ smax (xi ), where the neighbors are chosen from a set of size O(αn) w.h.p. The expected number of new factor
nodes is E[m0 ] = O(β), whence P[|Y | = km0 |G0 ] = 1 − o(1). Moreover, by the first part of Lemma 3.10 there is
ε = o(1) such that for sufficiently large T the graph G∗T is ε-symmetric with probability at least 1 − ε. Combining this
with the second part of Lemma 3.10 we find ε′ = o(1) such that given G0 the event Y = {kµG0 ,Y −⊗y∈Y µG0 ,y kT V ≤
ε and |Y | = km0 } occurs with probability at least 1 − ε′ as long as T > T0 (ε′ ) is sufficiently large.
13
Lemma 4.5. Let πG0 be the empirical distribution of the Gibbs marginals of G0 . We have
k
Y
X
k
−
1
(π
)
0
E ln(Z(G′ )/Z(G0 ))|G0 = oT (1) + (1 − α)
E DΛ
ψ(τ )
µh G (τj )
kξ
k
τ ∈Ω
h=1
0
with probability at least 1 − oT (1) over the choice of G .
Proof. On the event U ∩Y the graph G′ is obtained from G0 by simply adding m′ = P o(λ′ ) check nodes b1 , . . . , bm′
with neighborhoods and weights chosen from (5) and (6) respectively.
* ′
+
m
Y
E ln(Z(G′ )/Z(G0 ))|G0 , σ ∗n = oT (1) + E ln
G0
ψ bi (σ(∂1 bi ), . . . σ(∂k bi ))
i=1
= oT (1) + E ln
G0
′
X
µG0 ,Y (τ )
τ ∈Ω Y
m
Y
i=1
ψ bi (τ (∂1 bi ), . . . , τ (∂k bi )) G0 .
(11)
As the factor nodes b1 , . . . , bm′ are independently chosen from the same distribution w.h.p., with Claim 4.4 the expression (11) simplifies to
′
k
m
Y
X
X
µG0 ,∂h bi (τh ) G0
ψ bi (τ )
oT (1) + E
ln
i=1
h=1
τ ∈Ω k
k
Y
X
(1 − α)(k − 1)
E D ln
µG0 ,∂h b1 (τh ) G0 .
ψ b1 (τ )
=oT (1) +
k
k
τ ∈Ω
(12)
h=1
Hence, with i1 , . . . , ik chosen independently and uniformly at random from [n], using SYM2 in the distribution (5),(6)
of b1 and writing (12) with the empirical distribution of G0 we obtain
k
Y
X
(1
−
α)(k
−
1)
(π )
E ln(Z(G′ )/Z(G0 ))|G0 = oT (1) +
E DΛ
ψ(τ )
µh G0 (τj ) G0 .
kξ
k
τ ∈Ω
h=1
Claim 4.6. Let U ′ be the event that given G0 the process CPL1” yields m′ = 0 factor nodes and CPL2” does not pin
xn+1 . If β is sufficiently small, then
E ln(Z(G′′ )/Z(G0 ))|G0 = oT (1) + E 1U ′ ln(Z(G′′ )/Z(G0 ))|G0 .
For j = 1, . . . , γ b,sb let hj be independently uniformly chosen indices in [k] and y j = (y j1, , . . . , y j,k ) uniformly
random choices of neighborhoods subject to the condition that y j,hj = xn+1 6= y j,i , i 6= hj . Let Y ′ be the set of
neighbors of b1 , . . . , bsb chosen in CPL1” without xn+1 . Let Y = {yj,h : j ≤ γ b,sb , h ∈ [k]} \ {xn+1 } and D be a
sample from D. Then Y and Y ′ are mutually contiguous and there is ε = o(1) such that given G0 the event
Y ′ = {kµG0 ,Y ′ − ⊗y∈Y ′ µG0 ,y kT V ≤ ε and |Y ′ | = (k − 1)D}
occurs with probability at least 1 − ε.
Proof. The probability of xn+1 not being pinned is at least 1 − T /(n + 1) = o(1). As D has finite support, if β is
sufficiently small we have E[m′ ] = 0 and thus the first assertion follows.
Given G0 , for each j = 1, . . . , γ b,sb the probability of drawing a neighbor x with δ s (x) = 0 is upper bounded by
n−1 α−1 k, while E[γ b,sb ] = (1 − α)k −1 E[D] = O(1). Therefore Y ′ is of size (k − 1)D asymptotically almost
surely. Moreover, because δ s (y j,i ) is bounded by a small constant for any 1 ≤ j ≤ D, i ≤ k, we have mutually
contiguity of Y and Y ′ . As G∗T is ε-symmetric for sufficiently large T , we may again apply Lemma 3.10 to obtain a
sequence ε′ = o(1) such that Y ′ = {kµG0 ,Y ′ − ⊗y∈Y ′ µG0 ,y kT V ≤ ε and |Y ′ | = (k − 1)D} occurs with probability
at least ε′ .
14
Lemma 4.7. Let πG0 be the empirical distribution of the Gibbs marginals of G0 . We have
D X
−D
Y
X
Y
ξ
(πG0 )
E ln(Z(G′′ )/Z(G0 ))|G0 = oT (1) + E
Λ
µki+j
(τj )
1{τhi = σ}ψ i (τ )
|Ω|
k
i=1
σ∈Ω
j6=h i
τ ∈Ω
with probability at least 1 − oT (1) over the choice of G0 .
Proof. As in the proof of the previous lemma, given that U ′ ∩ Y ′ occurs, Claim 4.6 implies that G′′ is obtained from
G0 by simply adding D weighted check nodes from (8) and therefore
D
Y
X
Z(G′′ ) 0
E ln
(13)
ψ bi (τ (∂1 bi ), . . . , τ (∂k bi )) G0 .
G = oT (1) + E ln
µG0 ,Y (τ |Y )
Z(G0 )
Y ∪{x
}
i=1
τ ∈Ω
n+1
Hence, with the natural extension of σ ∗n to σ ∗n+1 and h1 , h2 , . . . uniformly chosen from [k], for i = 1, . . . , D let
(ω i,1 , . . . , ω i,k ) ∈ Ω k and ψ ′i be chosen from
P[(ω i,1 , . . . , ωi,k ) = (ω1 , . . . , ωk ), ψ ′ = ψ] ∝ 1{ωi,hi = σ ∗n+1 (xn+1 )}ξ −1 p(ψ)ψ(ω1 , . . . , ωk ).
Together with the ε-symmetry statement of Claim 4.6 equation (13) becomes
D X
Y
XY
oT (1) + E ln
µG0 ,yi,h (τh )|G0 .
1{τhi = σ}ψ ′i (τ )
σ∈Ω i=1 τ ∈Ω k
h6=hi
Again, writing out the probabilities for the independently chosen check nodes, with SYM we get
D X
−D
Y (π 0 )
XY
ξ
G
µki+h
(τh ) .
Λ
1{τhi = σ}ψ i (τ )
oT (1) + E
|Ω|
k
i=1
σ∈Ω
h6=hi
τ ∈Ω
Finally, to prove Proposition 4.1 we are going to make use of the following fact that is immediate from 3.12.
Fact 4.8 ([14, Corollary 3.13]). For all T ≥ 0 and all ω ∈ Ω we have
Eh||σ −1 (ω)| − n/|Ω||iG∗T = o(1).
Moreover, note that by performing a continuous transformation of the functional utilized in [14] we preserve the
following property.
Fact 4.9 ([14, Lemma 2.9]). The functional π ∈ P 2 (Ω) → B(D, π) is weakly continuous.
Proof of Proposition 4.1. By Lemmata 4.5 and 4.7 we have established that
∆T (n) ≤ oT (1) + E[B(D, πG0 )].
(14)
To bridge the gap to π ∈ P∗2 (Ω) we have to carve out that πG0 is arbitrarily close to the set P∗2 (Ω) as n → ∞. Fact
4.9 together with (14) then yields the assertion when taking the limits in the specified order.
Because σ ∗ is the uniform distribution on Ω n , Fact 4.8 gives Eh||σ −1 (ω)| − n/|Ω||iG∗T = o(1). Therefore, using
Lemma 4.3 we obtain
Z
1 X
h1{σ(x) = ω}iG0 − |Ω|−1
E
µ(ω)dπG0 (µ) − |Ω|−1 = E
n
x∈[n]
−1
|σ (ω)|
1
≤E
−
|
= o(1),
n
|Ω| G0
R
whence in expectation over G0 the measure µdπG0 (µ) converges to the uniform distribution in total variation
distance. That is w.h.p. there is α(G0 ) ≥ 0, E[α(G0 )] = o(1) and a measure ν(G0 ) ∈ P(Ω) such that the convex
combination (1 − α(G0 ))πG0 + α(G0 )δν(G0 ) ∈ P∗2 (Ω) closes the gap.
15
5. T HE
UPPER BOUND
In this section we carry out the calculations for the Guerra-Toninelli interpolation. For any given π ∈ P∗2 (Ω) we set
up a family of graphs (Gt )t∈[0,1] , by which we can interpolate between the original graph G1 = G∗ and a graph G0
with free energy −nB(D, π) + o(n). By proving that the derivative ∂/∂t E[Z(Gt )] is positive on the entire interval,
we obtain
1
(15)
− E[ln Z(G∗ )] ≤ −B(D, π) + o(1).
n
Throughout the section we assume that α ∈ (0, 1), β > 0, the degree distribution D and π ∈ P∗2 (Ω) are arbitrary but
fixed.
5.1. The interpolation. In the interpolation we will utilize the fact, that our graph G = Gα,β,D consists of smax
layers of Poissonian degree. This being the case, for any s ∈ {1, . . . , smax } we define an interpolation as follows.
A codeword in the interpolation model consists of s − 1 layers of parity checks, followed by a layer, where with
probability 1 − t′ , t′ ∈ [0, 1] each parity check is replaced by a repetition of the codebits it contains, and a final
smax − s layers of simple blocks of repetition code altogether satisfying the degree distribution. To make this precise,
we define a random factor graph model Gs,t = Gs,t (m, γ) as follows.
I1: Draw a random D-partition d of [n]. Let X ′ be chosen from Po(βt), X ′′ be chosen from Po(β(1 − t)) and
let X 1 , X 2 , . . . be a sequence of Po(β) random variables all mutually independent. Define vectors m and γ
by letting
ℓ<s
0,
Xℓ , ℓ < s
′′
′
for ℓ = 1, . . . , smax .
γs = X , ℓ = s
mℓ = X , ℓ = s ,
Xℓ , ℓ > s
0,
ℓ>s
I2: For i = 1, . . . , ms add a k-ary factor node as,i with neighborhood from
Qk
j=1 δ s (xj )
.
P[∂as,i = (x1 , . . . , xk )] = P
Qk
y1 ,...,yk
j=1 δ s (yj )
For j = 1, . . . , kγ s add a unary factor bs,j with neighbors chosen from
δ s (x)
.
P[∂bs,j = x] = P
y δ s (y)
(16)
Set δ s+1 = (δ s − ∇s )+ , where ∇s (x) denotes the number of times x was drawn as a neighbor in round s.
Increase s and abort when s > smax .
I3: To each k-ary factor node a in the graph independently assign a weight function ψ a chosen from p.
I4: To each unary factor node b in the graph independently assign a unary weight function ψ b as follows. Choose
ψ from p, indenpendently choose i from the uniform distribution on [k] and choose µ1 , . . . , µk iid from π.
Let ψ b be the map
X
Y
σ 7→
ψ(τ1 , . . . , τi−1 , σ, τi+1 , . . . , τk )
µh (τh ).
τ1 ,...,τk
h6=i
Having established the interpolation null-model, we can now define our original interpolation in the teacher-student
model by reweighing. With a signal σ : [n] → Ω we define the distribution G∗s,t (σ) = G∗s,t (σ, m, γ) by letting
P[G∗s,t (σ) ∈ A] =
E[ψGs,t (σ)1{Gs,t ∈ A}]
.
E[ψGs,t (σ)]
If the signal σ is chosen uniformly at random, we write G∗s,t = G∗s,t (m, γ) for G∗s,t (σ, m, γ). Notice that s =
smax , t′ = 1 yields the original graph model G∗ and s = 1, t′ = 0 corresponds to the graph of a simple repetition
code. All we do by layer-wise interpolation is to split the interpolation interval into smax intervals of equal length.
Finally, to ensure symmetry, we apply the pinning procedure. For this purpose fix T > 0.
IP1: Choose σ̌ from the uniform distribution on Ω n .
IP2: Choose a random graph G = ([n], F , (∂a)a∈F , (ψ a )a∈F ) from G∗s,t (σ̌, m, γ).
IP3: Choose θ in [0, T ] uniformly at random and let U be a random θ/n-subset of [n].
16
IP4: To each x ∈ U connect a unary check node ax with weight function ψax (τ ) = 1{σ̌(x) = τ }.
Let us write G∗T,s,t for the resulting graph. Note that by Lemma 3.10 there is T0 that only depends on ε and Ω such
that for T ≥ T0 the Gibbs measure of G∗T,s,t is ε-symmetric with probability at least 1 − ε. Thus, we can fix a
sufficiently large T > 0 before performing the interpolation and guarantee that throughout the process the family of
graphs remains ε-symmetric.
The tally of total factor nodes in the partition function during interpolation is accounted for in a correction term. To
this end, with independent samples µ1 , µ2 , . . . from π, let
Γs,t
k
Y
(s + t − 1)β(k − 1) X
(π)
µj (τj ) .
E Λ
=
ψ(τ )
ξ
k
j=1
τ ∈Ω
The following lower bound is the main ingredient to our interpolation argument as it tethers the derivative arbitrarily
closely to zero.
Proposition 5.1. Let
ΦT,s : t ∈ [0, 1] 7→ E[ln Z(G∗T,s,t )] + Γs,t /n.
Then for all t ∈ [0, 1] and all s ∈ [smax ] we have Φ′T,s (t) > oT (1)n−1 .
We will prove Proposition 5.1 in section 5.3. Let us first see how Proposition 5.1 implies (15) and as such Proposition
3.14.
5.2. Proof of Proposition 3.14. With the fundamental theorem of calculus we write
1
1
1
E[ln Z(G∗T,n )] + Γsmax ,1 = (E[ln Z(G∗T,smax ,1 )] + Γsmax ,1 )
n
n
n
sX
max Z t
1
Φ′T,s (t)dt
= (E[ln Z(G∗T,1,0 )] + Γ1,0 ) +
n
s=1 0
sX
max Z t
1
1
≥ E[ln Z(G∗T,1,0 )] −
oT (1)dt
n
n s=1 0
(17)
1
E[ln Z(G∗T,1,0 )] + oT (1).
(18)
n
If (ψ i )i≥1 is a sequence with entries independently chosen from p, (µi,j )i,j≥1 has entries independently chosen from
π, h1 , h2 , . . . are independently uniform choices from [k] and γ is a random sample from D, a simple calculation
unfolds that
γ X
X
Y
Y
1
1
E[ln Z(G∗T,1,0 )] =
E ξ −γ Λ
µb,j (τj ) .
(19)
1{τhb = σ}ψ b (τ )
n
|Ω|
k
=
σ∈Ω b=1 τ ∈Ω
j∈[k]\{h b }
Also, by definition of Γs,t and smax ,
Γsmax ,1
k
Y
(1 − α)n E[D] (k − 1) X
(π)
≤
µj
E Λ
ψ(τ )
k
ξ
k
j=1
(20)
τ ∈Ω
Plugging (19) and (20) into (17) and consequently taking the lim inf of T → ∞, we have
1
lim inf E[Z(G∗α,β,D )]
n→∞ n
γ X
XY
Y
1
≥ lim inf
E ξ −γ Λ
µb,j (τj )
1{τhb = σ}ψ b (τ )
n→∞
|Ω|
k
σ∈Ω b=1 τ ∈Ω
j∈[k]\{h b }
k
Y
X
(1 − α)(k − 1)
(π)
E[D] E Λ
µj .
ψ(τ )
−
kξ
k
j=1
τ ∈Ω
17
(21)
The weights in Ψ are strictly positive. Hence, Proposition 2.2 allows us to compare the free energy of our exact model
Z(G∗D ) with the approximative free energy as
E[ln Z(G∗D )] = E[ln Z(G∗α,β,D )] + O(αn).
Thus, (21) extends to the exact model when ultimately taking our approximation to the limit
lim sup −
n→∞
1
E[Z(G∗D )]
n
γ X
XY
Y
1
≤ − sup lim sup lim sup
E ξ −γ Λ
µb,j (τj )
1{τhb = σ}ψ b (τ )
|Ω|
π∈P∗2 (Ω) α,β→0 n→∞
k
σ∈Ω b=1 τ ∈Ω
j∈[k]\{h b }
k
Y
X
(1 − α)(k − 1)
(π)
−
µj
E[D] E Λ
ψ(τ )
kξ
k
j=1
τ ∈Ω
=−
sup
B(D, π).
π∈P∗2 (Ω)
5.3. Proof of Proposition 5.1. To prove Proposition 5.1 we derive a more practical expression of the derivative Φ′T,s
which is comparable to the expression from POS.
Proposition 5.2. Let (σ̌, G∗T,s,t ) be chosen from the IP experiment. Let ψ be chosen from p, µ1 , µ2 , . . . , µk be chosen
from π, all mutually independent. For s ∈ [smax ] let δ s be chosen from AP2’ in Definition 3.5 and set
δs
νs = P
y∈[n]
δ s (y)
.
With y, y 1 , . . . , y k chosen uniformly from the set of variable nodes, we let
Ξs,t,l = E
"
k
Y
ν s (y i )h1 −
ψ(σ(y 1 ), . . . , σ(y k )ilG∗T ,s,t
i=1
− E ν s (y)h1 −
X
ψ(τ )1{τi = σ(y i )}
Y
j6=i
τ ∈Ω k
#
µj (τj )ilG∗T ,s,t
l
k
Y
X
µj (τj ) .
ψ(τ )
+ (k − 1) E 1 −
j=1
τ ∈Ω k
Then uniformly for all t ∈ (0, 1), s ∈ [smax ] and T ≥ 0 we have
d
β X Ξs,t,l
ΦT,s (t) = oT (1)n−1 +
.
dt
nξ
l(l − 1)
l≥2
To prove Proposition 5.2 we begin by rewriting
distribution with parameters λ > 0 satisfies
∂
∂t
E[ln Z(G∗T,s,t )] into a similar expression. Note that the Poisson
∂ λm −λ
λm−1 −λ λm −λ
∂
Po(λ)({m}) =
e =
e −
e = Po(λ)({m − 1}) − Po(λ)({m}),
∂λ
∂λ m!
(m − 1)!
m!
18
m ≥ 1.
Hence, as m and γ are independent but add Po(β) new neighbors to each layer when put together, differentiating
E[ln Z(G∗T,s,t )] with respect to t yields
∂
E[ln Z(G∗T,s,t )]
∂t
X
∂
=
E[ln Z(G∗T,s,t )|ms = m, γ s = γ] Po(tβ)({m})Po((1 − t)β)({γ})
∂t
m,γ
X
E[ln Z(G∗T,s,t )|ms = m + 1] − E[ln Z(G∗T,s,t )|ms = m] Po(tβ)({m})
=β
m
−β
X
γ
E[ln Z(G∗T,s,t )|γ s = γ + 1] − E[ln Z(G∗T,s,t )|γ s = γ] Po((1 − t)β)({γ})
=β E[ln Z(G∗T,s,t (m + 1s , γ))] − E[ln Z(G∗T,s,t (m, γ))]
− β E[ln Z(G∗T,s,t (m, γ + 1s ))] − E[ln Z(G∗T,s,t (m, γ))] .
The term β E[ln Z(G∗T,s,t (m, γ))] cancels and we can write
∂
E[ln Z(G∗T,s,t )] = β E[ln Z(G∗T,s,t (m + 1s , γ))] − β E[ln Z(G∗T,s,t (m, γ + 1s ))]
∂t
(22)
Now consider a graph model G∗,1
T,s,t , where we slightly alter the procedure I2 as follows.
I2’: For any ℓ 6= s we construct the graph as described by I2. If ℓ = s, instead of increasing s and moving onto
(1)
(1)
I3, we add another k unary check nodes b1 , . . . , bk with neighborhoods chosen from (16), update δ s+1
accordingly and equip each of the new check nodes with constant weight functions ψb(h) = 1, h = 1, . . . , k.
Afterwards we increase s and continue.
∗
In short G∗,1
T,s,t differs from GT,s,t by having k additional unary check nodes within layer s, which have neutral weight.
By letting
∆T,s,t = E[ln Z(G∗T,s,t (m + 1s , γ))] − E[ln Z(G∗,1
T,s,t (m, γ))]
∆′T,s,t
=k
−1
E[ln Z(G∗T,s,t (m, γ
+ 1s ))] − k
−1
and
E[ln Z(G∗,1
T,s,t (m, γ))]
we can write (22) as
∂
E[ln Z(G∗T,s,t )] = β∆T,s,t − kβ∆′T,s,t .
(23)
∂t
The Nishimori property naturally extends to the case of the interpolation model. By choosing neutral weights in I2’
this includes G∗,1
T,s,t .
∗
Lemma 5.3. Lemma 3.13 remains true if we replace G∗ by either G∗,1
T,s,t or GT,s,t .
The proof is analogous to Lemma 3.12
Lemma 5.4. Let (σ̌, G∗T,s,t ) be chosen from the IP experiment. Let a be a check node with ∂a chosen from ν ⊗k
s and
ψ a chosen from
P[ψ = ψ] =
p(ψ)
ψ(σ(∂a)).
ξ
Moreover, let b be a unary check node with ∂b chosen from ν s and assign to it a weight function ψ b defined by
Y
X
ψ ′ (τ )1{τi = σ}
µ′h (τh ),
ψ b (σ) =
h6=i
τ ∈Ω k
′
µ′1 , . . . , µ′k
where the index i ∈ [k], the weight function ψ ∈ Ψ and
are chosen from
Z Y
X
1{τi = σ̌(∂b)}ψ(τ )
µ′j (τj )dπ ⊗k (µ′1 , . . . , µ′k ).
P[i = i, (µ′1 , . . . , µ′k ) ∈ A, ψ ′ = ψ] ∝ p(ψ)
A j6=i
τ ∈Ω k
19
(24)
Then
∆T,s,t = E[lnhψ a (σ(∂a))iG∗T ,s,t ],
(25)
∆′T,s,t
(26)
= E[lnhψ b (σ(∂b))iG∗T ,s,t ] + oT (1).
∗
Proof. We begin by showing (25). Note that G∗,1
T,s,t (m, γ) differs from GT,s,t (m, γ) by having k additional unary
check nodes with neutral weight assigned during layer s. Of course the sockets chosen as neighbors of these additional
∗
check nodes induce a perturbation on the distributions ν ℓ , ℓ ≥ s. However, G∗,1
T,s,t (m, γ) and GT,s,t (m + 1s , γ) can
be coupled such that afore-said neighborhood is assigned to the the additional k-ary check node a, while all remaining
choices of neighborhoods and weight functions are chosen from the same distribution within both graphs. Given this
coupling, we have
Z(G∗T,s,t (m + 1s , γ))
Z(G∗,1
T,s,t (m, γ))
=
=
P
σ∈Ω k
ψ a (σ(∂a))
Q
c∈F \{a}
ψc (σ)
∗,1
Z(GT,s,t
(m, γ))
X
ψa (σ(∂a))µG∗,1
T ,s,t (m,γ)
=
(σ)
X ψG∗,1
T ,s,t (m,γ)
σ∈Ω k
Z(G∗,1
T,s,t (m, γ))
ψa (σ(∂a))
(σ) = hψ a (σ(∂a))iG∗,1 .
T ,s,t
σ∈Ω k
Taking the logarithm and integrating immediately gives (25). To show (26) we couple G∗T,s,t (m, γ + 1s ) and
(1)
(1)
∗,1
G∗,1
T,s,t (m, γ) in a similar fashion. That is in layer s, the neighborhood chosen by b1 , . . . , bk of GT,s,t (m, γ)
is equally assigned to the k additional unary check nodes of G∗T,s,t (m, γ + 1s ) that result from the positive entry in
1s . These k unary nodes are then independently assigned weight functions chosen from (24). We couple all remaining
random variables trivially by copying the choice of G∗,1
T,s,t (m, γ). With this joint distribution, we can write
Z(G∗T,s,t (m, γ + 1s ))
=
∗,1
Z(GT,s,t
(m, γ))
Qk
P
(1) Q
σ∈Ω k
i=1 ψ b(1) (σ(∂bi ))
c∈F \{∪k
(1)
i=1 bi }
i
ψc (σ)
Z(G∗,1
T,s,t (m, γ))
=
*
k
Y
i=1
+
(1)
ψ b(1) (σ(∂bi ))
i
.
(27)
G∗,1
T ,s,t
By taking the logarithm and integrating (27) gives
*
k∆′T,s,t = E[ln Z(G∗T,s,t (m, γ + 1s ))] − E[ln Z(G∗,1
T,s,t (m, γ))] = E
k
Y
(1)
+
ψ b(1) (σ(∂bi ))
i=1
i
G∗,1
T ,s,t
.
(1)
∗,1
While the initial Gibbs measure of Gs,t
does not necessarily factorize over the marginals of ∂bi , i = 1, . . . , k,
Lemma 3.10 guarantees that our pinned measure G∗,1
T,s,t is in fact (oT (1), k)-symmetric with probability 1 − oT (1).
(1)
Because the random set Y = {∪ki=1 bi } is contiguous with respect to a uniformly random choice Y ′ from [n]k this
implies
*
E
k
Y
(1)
+
ψ b(1) (σ(∂bi ))
i=1
i
G∗,1
T ,s,t
i
h
= k E hψ (1) (σ(∂b(1)
∗,1
+ oT (1)
))i
1
G
b
1
and thus proves the assertion.
T ,s,t
Before we can verify Proposition 5.2 we have to write out (25) and (26). This is a simple, but technical computation.
20
Claim 5.5. With the assumptions of Proposition 5.2 we have
* l
+
k
X
Y
Y
1−ξ
1
1
∆T,s,t = −
+
E ν s (y i )
1 − ψ(σ h (y 1 ), . . . , σ h (y k ))
ξ
ξ
l(l − 1)
i=1
l≥2
∆′T,s,t =oT (1) −
+
1
E ν s (y)
l(l − 1)ξ
X
G∗
T ,s,t
1−ξ
ξ
l≥2
The definition ∆′′s,t =
h=1
∂
1
β(k−1) ∂t Γs,t
∆′′s,t = −
*
l
Y
X
1−
h=1
Y
ψ(τ )1{τi = σ̌(y)}
+
G∗
T ,s,t
gives
1−ξ
+
ξ
ξ
1X
l≥2
1
l(l − 1)
(28)
µj (τj )
j6=i
τ ∈Ω k
.
(29)
l
k
Y
X
µj (τj ) .
E 1 −
ψ(τ )
(30)
j=1
τ ∈Ω k
Proof. We begin by showing (28). In the interpolation model (σ̌, G∗T,s,t ), each k-ary check node a chooses its
neighborhood from P[∂a = (x1 , . . . , xk )] = ν ⊗k
s (x1 , . . . , xk ) and then obtains a weight function from P[ψ a |σ̌ =
σ] = ξ −1 p(ψ)ψσ(∂a). Because all weight functions take values in (0, 2) writing σ 1 , σ 2 , . . . for independent samples
from µG∗,1 and expanding the logarithm
T ,s,t
lnhψ(σ(y1 ), . . . , σ(yk ))iG∗,1 = −
T ,s,t
X1
l≥1
=−
l
X1
l≥1
l
h1 − ψ(σ(y1 ), . . . , σ(yk ))ilG∗,1
T ,s,t
*
l
Y
+
1 − ψ(σ h (y1 ), . . . , σ h (yk ))
h=1
,
G∗,1
T ,s,t
we obtain
i
h
E ln hψ a (σ(∂a))iG∗,1
T ,s,t
!
* l
+
k
X
Y
Y
X 1
ν s (yi ) ψ(σ̌(y1 ), . . . , σ̌(yk )))
E
1 − ψ(σ h (y1 ), . . . , σ h (yk ))
=−
lξnk y ,...,y
i=1
l≥1
=
X 1
lξnk
l≥1
1
h=1
k
X
E
y1 ,...,yk
"
k
Y
!
ν s (yi ) (1 − ψ(σ̌(y1 ), . . . , σ̌(yk )))
i=1
·
X 1
−
lξnk
l≥1
G∗,1
T ,s,t
X
y1 ,...,yk
E
"
k
Y
i=1
*
l
Y
1 − ψ(σ h (y1 ), . . . , σ h (yk ))
h=1
ν s (yi ) h1 −
G∗,1
T ,s,t
ψ(σ(y1 ), . . . , σ(yk ))ilG∗,1
T ,s,t
#
.
(32)
∗,1
By Lemma 5.3, the pairs (σ̌, GT,s,t
) and (σ h , G∗,1
T,s,t ), h ≥ 1 are identically distributed. Thus, we can write
!
* l
+
k
Y
Y
ν s (yi ) (1 − ψ(σ̌(y1 ), . . . , σ̌(yk )))
E
1 − ψ(σ h (y1 ), . . . , σ h (yk ))
i=1
i=1
G∗,1
T ,s,t
h=1
! * l+1
+
k
Y
Y
ν s (yi )
=E
1 − ψ(σ h (y1 ), . . . , σ h (yk ))
h=1
21
G∗,1
T ,s,t
(31)
+
!
.
Moreover, SYM gives
1
ξnk
X
E
y1 ,...,yk
"
k
Y
!
ν s (yi ) h1 − ψ(σ(y1 ), . . . , σ(yk ))iG∗,1
i=1
T ,s,t
#
=
1−ξ
ξ
and thus (32) simplifies to
−
1−ξ
+
ξ
X X
l≥2 y1 ,...,yk
k
Y
1
E
l(l − 1)ξnk
i=1
! * l+1
+
Y
ν s (yi )
1 − ψ(σ h (y1 ), . . . , σ h (yk ))
h=1
G∗,1
T ,s,t
.
Lemma 5.4 then implies (30).
To calculate (29), recall that in layer s a unary check node b is added by chosing a neighbor y from νs and equipping b
with a weight function chosen from (24). For y ∈ [n] write by for a check node chosen in the same way but conditioned
on the event that ∂b = y. Then
i
i
h
X h
E ν s (y) lnhψ by (σ(y))iG∗,1 .
E lnhψ b (σ(∂b))iG∗,1 =
(33)
T ,s,t
T ,s,t
y∈[n]
By SYM the normalization in (24) is
in equation (33) we get
1
nkξ
X
y∈[n],i∈[k]
E ν s (y)
P
τ ∈Ω k
X
Pk
1{τi = σ̌(∂b)} E[ψ(τ )
i=1
Q
j6=i
µ′j (τj )] = kξ. By writing out (24)
1{τi = σ̌(x)}ψ(τ )
(34)
τ ∈Ω k
·
Y
µ′j (τj ) ln
j6=i
*
X
1{σi = σ(y)}ψ(σ)
Y
+
µ′j (σj )
j6=i
σ∈Ω k
G∗,1
T ,s,t
.
(35)
Again, we utilize the fact that our weight functions take values within (0, 2) and expand the logarithm to write (35) as
−
k X
X X
y∈[n] i=1 l≥1
Y
X
1
E ν s (y)
1{τi = σ̌(x)}ψ(τ )
µ′j (τj )
nklξ
k
j6=i
τ ∈Ω
·
=−
k X
X X
y∈[n] i=1 l≥1
*
l
Y
1−
h=1
X
1{σi = σ h (y)}ψ(σ)
+
µ′j (σj )
j6=i
σ∈Ω k
Y
G∗,1
T ,s,t
Y
X
1
E ν s (y) 1 −
1{τi = σ̌(x)}ψ(τ )
µ′j (τj )
nklξ
j6=i
τ ∈Ω k
* l
+
Y
Y
X
′
·
1{σi = σ h (y)}ψ(σ)
1−
µj (σj )
h=1
− ν s (y)
*
σ∈Ω k
l
Y
1−
h=1
22
X
σ∈Ω k
j6=i
1{σi = σ h (y)}ψ(σ)
Y
j6=i
G∗,1
T ,s,t
+
µ′j (σj )
G∗,1
T ,s,t
Once again we simplify the expression by using that the distributions σ̌ and σ h , h ≥ 1 coincide
Y
X
E ν s (y) 1 −
1{τi = σ̌(x)}ψ(τ )
µ′j (τj )
j6=i
τ ∈Ω k
·
= E ν s (y)
and obtain
* l+1
Y
1−
h=1
*
l
Y
1−
h=1
X
X
1{σi = σ h (y)}ψ(σ)
Y
µ′j (σj )
j6=i
σ∈Ω k
1{σi = σ h (y)}ψ(σ)
Y
+
G∗,1
T ,s,t
i
h
E lnhψ b (σ(∂b))iG∗,1
T ,s,t
*
+
Y
X
1 X X
′
=−
E ν s (y) 1 −
1{σi = σ 1 (y)}ψ(σ)
µj (σj )
nkξ
k
+
X
l≥2
X X
y∈[n] i∈[k]
1−ξ
ξ
=−
+
1
nkξl(l − 1)
X
l≥2
j6=i
σ∈Ω
E ν s (y)
X X
1
E ν s (y)
nkξl(l − 1)
y∈[n] i∈[k]
*
*
l
Y
1−
h=1
l
Y
X
h=1
X
G∗,1
T ,s,t
1{σi = σ 1 (y)}ψ(σ)
Y
1{σi = σ 1 (y)}ψ(σ)
Y
+
G∗,1
T ,s,t
µ′j (σj )
j6=i
σ∈Ω k
+
µ′j (σj )
j6=i
σ∈Ω k
1−
G∗,1
T ,s,t
µ′j (σj )
j6=i
σ∈Ω k
y∈[n] i∈[k]
+
G∗,1
T ,s,t
by employing SYM in the final equation.
Qk
P
(π)
Finally let us derive (30). By definition Γs,t = ξ −1 (s + t − 1)β(k − 1) E[Λ( τ ∈Ω k ψ(τ ) j=1 µj (τj ))], where
(π)
(π)
we write µ1 , µ2 , . . . for independent samples from π. As in the previous cases, we perform the same procedure of
expanding the logarithm and simplifying the telescopic sum to obtain
k
Y
X
1
∂
(π)
µj (τj )
Γs,t = ξ −1 E Λ
ψ(τ )
β(k − 1) ∂t
k
j=1
τ ∈Ω
l
k
k
X1
Y
X
Y
X
(π)
(π)
µj (τj )
=ξ −1
µj (τj ) 1 −
ψ(τ )
E 1 −
ψ(τ )
l
k
k
j=1
j=1
l≥1
τ ∈Ω
τ ∈Ω
= − ξ −1 E 1 −
=−
X
ψ(τ )
τ ∈Ω k
1X
1
1−ξ
+
ξ
ξ
l(l − 1)
l≥2
k
Y
j=1
(π)
− 1 −
µj (τj ) + ξ −1
X
l≥2
X
ψ(τ )
τ ∈Ω k
1
l(l − 1)
k
Y
j=1
l
k
Y
X
(π)
µj (τj )
E 1 −
ψ(τ )
l
k
Y (π)
X
µj (τj ) .
ψ(τ )
E 1 −
τ ∈Ω k
l
(π)
µj (τj )
τ ∈Ω k
j=1
j=1
23
Proof of Proposition 5.2. The assertion is now immediate as (23) and Claim 5.5 yield
∂
β X Ξs,t,l
ΦT,s (t) = n−1 oT (1) + β∆s,t − kβ∆′s,t + β(k − 1)∆′′s,t = oT (1)n−1 +
.
∂t
n
l(l − 1)
l≥2
We complete the proof of Proposition 5.1 by comparing the expressions Ξs,t,l to a non-negative value given by POS.
To this end let ρ1 , ρ2 , . . . be a sequence of independently samples from πG∗T ,s,t , let µ1 , µ2 , . . . be independently
chosen from π. Define
′
Ξs,t,l
l
l
k
k
Y
X
Y
X
µi (τi )
ρi (τi ) + (k − 1) 1 −
ψ(τ )
ψ(σ)
=E 1 −
i=1
σ∈Ω k
τ ∈Ω k
i=1
l
k
Y
X
X
1 −
ψ(τ )ρ1 (τ1 )
µj (τj ) .
−
i=1
(36)
j6=i
τ ∈Ω k
Let σ be a chosen from µG∗T ,s,t . By Lemma 5.3 the pairs (σ ∗ , G∗T,s,t ) and (σ, G∗T,s,t ) are identically distributed. Thus,
for any choice of s, t the mean of the empirical marginal distribution πG∗T ,s,t is given by the uniform distribution. As
POS holds, expanding Λ shows that the expression (36) is non-negative for any l ≥ 0.
Proof of Proposition 5.1. With this consideration the assertion follows from Proposition 5.2 by verifying
′
|Ξs,t,l − Ξs,t,l
| = oT (1)
(37)
for any choice of s, t and l.
For the first and third summand in (36) we will use the Pinning Lemma to sufficiently regularize the underlying graph
models, such that exchangeability with the corresponding term from Ξs,t,l is possible. For the time being, consider
the first suammnd. Observe that for independently uniform choices y 1 , y 2 , . . . among the variable nodes [n], we can
write
E 1 −
=E
k
Y
i=1
X
ψ(σ)
σ∈Ω k
ν s (y i ) 1 −
k
Y
i=1
l
ρi (σi ) G∗T,s,t
X
ψ(σ)
k
Y
i=1
σ∈Ω k
24
l
µG∗T ,s,t (σi ) G∗T,s,t .
Hence, for any ψ ∈ Ψ the triangle inequality and the Cauchy-Schwarz inequality yield
l
k
k
Y
X
Y
X
1
ρi (σi ) G∗T,s,t
ν s (yi )h1 − ψ(σ(y1 ), . . . , σ(yk )ilG∗T ,s,t − E 1 −
ψ(σ)
nk y ,...,y i=1
k
i=1
1
≤
1
nk
σ∈Ω
k
k
X Y
ν s (yi ) h1 − ψ(σ(y1 ), . . . , σ(yk )ilG∗T ,s,t
y1 ,...,yk i=1
l
k
Y
X
µG∗T ,s,t (σi ) G∗T,s,t
ψ(σ)
− E 1 −
i=1
σ∈Ω k
≤n−k
k
X Y
y1 ,...,yk i=1
X
·
y1 ,...,yk
ν s (yi )2
!1/2
(38)
X
h1 − ψ(σ(y1 ), . . . , σ(yk )ilG∗T ,s,t − E 1 −
ψ(σ)
σ∈Ω k
k
Y
i=1
l
2 1/2
µG∗T ,s,t (σi ) G∗T,s,t
.
We can rewrite (38) as the ratio between the 2- andP
1-norm of the vector (δ s (v))v∈n . The expression is maximized by
sparse vectors, where each vector contains at most v δ 1 (v)/(α maxv δ 1 (v)) non-zero entries. Hence, we can bound
the factor (38) by
!1/2
!k
pP
k
X Y
δ s (v)2
O(1)
2
−k
−k
v
P
ν s (yi )
≤n
n
·
≤ k/2 .
n
v δ s (v)
y ,...,y
i=1
1
k
Moreover, as each ψ evaluates in (0, 2), Lemma 5.3 implies that for any C > 0, ε > 0, l ≥ 1 there is δ > 0 such that
if G∗T,s,t is δ-symmetric, we have
l
k
Y
X
X
µG∗T ,s,t (σi ) G∗T,s,t < ε.
h1 − ψ(σ(y1 ), . . . , σ(yk )ilG∗T ,s,t − E 1 −
ψ(σ)
Cn−k
y1 ,...,yk
i=1
σ∈Ω k
In the same way, for any ψ ∈ Ψ and i ∈ [k] we can bound
k
X
Y
1 XY
ν s (yi )h1 −
ψ(τ )1{τi = σ i (y)}
µj (τj )ilG∗T ,s,t
n y i=1
j6=i
τ ∈Ω k
l
X
Y
− E 1 −
ψ(τ )ρ1 (τi )
µj (τj ) G∗T,s,t < ε.
τ ∈Ω k
j6=i
By Lemma 3.8 we can guarantee that G∗T,s,t is oT (1)-symmetric with probability 1 − oT (1), which implies (37) with
probability 1 − oT (1).
Acknowledgement. We thank Amin Coja-Oghlan for helpful discussions.
25
R EFERENCES
[1] E. Abbe: Community detection and stochastic block models: recent developments. arXiv preprint, arXiv:1703.10146 (2017).
[2] E. Abbe, A. Montanari: Conditional random fields, planted constraint satisfaction and entropy concentration. Theory of Computing 11 (2015),
413–443.
[3] D. Achlioptas, H. Jia, C. Moore: Hiding satisfying assignments: Two are Better than One. Journal of Artificial Intelligence Research 24,
(2005), 623-639.
[4] D. Achlioptas, H. Hassani, N. Macris, R. Urbanke: Bounds for random constraint satisfaction problems via spatial coupling. Proceedings of
the 27th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA (2016), 469–479.
[5] M. Aizenman, R. Sims, S. Starr: An extended variational principle for the SK spin-glass model. Physical Review B 68(21), (2003).
[6] J. Banks, C. Moore, J. Neeman, P. Netrapalli: Information-theoretic thresholds for community detection in sparse networks. Proceedings of
the 29th COLT, (2016), 383–416.
[7] J. Barbier, M. Dia, N. Macris: Threshold saturation of spatially coupled sparse superposition codes for all memoryless channels. Information
Theory Workshop (ITW), IEEE, (2016).
[8] J. Barbier, M. Dia, N. Macris, F. Krzakala, T. Lesieur, L. Zdeborová: The mutual information in random linear estimation. Proccedings of the
54th Annual Allerton Conference, (2016), 625–632.
[9] J. Barbier, M. Dia, N. Macris, F. Krzakala, T. Lesieur, L. Zdeborová: Mutual information for symmetric rank-one matrix estimation: A proof
of the replica formula. In Advances in Neural Information Processing Systems (2016), 424–432.
[10] J. Barbier, N. Macris: I-MMSE relations in random linear estimation and a sub-extensive interpolation method. arXiv preprint,
arXiv:1704.04158, (2017).
[11] E. Berlekamp, R.J. McEliece, H. C A Van Tilborg: On the inherent intractability of certain coding problems. IEEE Transactions on Information
Theory 24(3), (1978), 384—386.
[12] C. Berrou, A. Glavieux, P. Thitimajshima: Near Shannon limit error-correcting coding and decoding: Turbo-codes. 1. Proceedings of the IEEE
International Conference on Commununication Geneva (1993), 1064-–1070.
[13] C. Bordenave, M. Lelarge, L. Massoulié: Non-backtracking spectrum of random graphs: community detection and non-regular Ramanujan
graphs. Procedings of the 56th Annual Symposium on Foundations of Computer Science, FOCS (2015), 1347–1357.
[14] A. Coja-Oghlan, F. Krzakala, W. Perkins, L. Zdeborova: Information-Theoretic Thresholds from the Cavity Method. Proceedings of the 49th
Annual ACM SIGACT Symposium on Theory of Computing STOC (2017), 146–157.
[15] V. Feldman, W. Perkins, S. Vempala: On the complexity of random satisfiability problems with planted solutions. Proceedings of the 47th
Annual ACM on Symposium on Theory of Computing STOC, (2015), 77–86.
[16] S. Franz, M. Leone, A. Montanari, F. Ricci-Tersenghi: Dynamic phase transition for decoding algorithms. Physical Review E 66(4), (2002).
[17] A. Clauset, A. Ghasemian, C. Moore, L. Peel, P. Zhang: Detectability thresholds and optimal algorithms for community structure in dynamic
networks. Physical Review X, 6(3), (2016).
[18] A. Giurgiu, N. Macris, R. Urbanke: Spatial coupling as a proof technique and three applications. IEEE Transactions on Information Theory
62(10), (2016), 5281–5295.
[19] F. Guerra, F. L. Toninelli: The thermodynamic limit in mean field spin glasses. Communications in Mathematical Physics 230 (2002), 71—79.
[20] S.H. Hassani, N. Macris, R. Urbanke: Threshold saturation in spatially coupled constraint satisfaction problems. Journal of Statistical Physics
150(5), (2013), 807–850.
[21] H. Jia, C. Moore, D. Strain: Generating hard satisfiable formulas by hiding solutions deceptively. Journal of Artifical Intelligence Research
28, (2007), 107 – 118.
[22] Y. Kabashima, F. Krzakala, M. Mézard, A. Sakata, L. Zdeborová: Phase transitions and sample complexity in bayes-optimal matrix factorization. IEEE Transactions on Information Theory, 62(7), (2016), 4228–4265.
[23] S. Kudekar, N. Macris: Proof of replica formulas in the high noise regime for communication using LDGM codes. Information Theory
Workshop, (2008), 416–420.
[24] S. Kudekar, N. Macris: Sharp bounds for optimal decoding of low-density parity-check codes. IEEE Transactions on Information Theory
55(10) (2009), 4635–4650.
[25] S. Kumar, A. J. Young, N. Macris and H. D. Pfister: Threshold Saturation for Spatially Coupled LDPC and LDGM Codes on BMS Channels.
IEEE Transactions on Information Theory 60, (2014), 7389–7415
[26] M. Lelarge, L. Miolane: Fundamental limits of symmetric low-rank matrix estimation. Proceedings of the 2017 Conference on Learning
Theory, PMLR 65, (2017), 1297–1301.
[27] M. Lelarge, L. Massoulié, J. Xu: Reconstruction in the Labeled Stochastic Block Model. IEEE Transactions on Network Science and Engineering. 2(4), (2015), 152-163.
[28] M. Luby, M. Mitzenmacher, A. Shokrollahi, D.A. Spielman: Analysis of low density codes and improved designs using irregular graphs.
Proceedings of the 30th Annual ACM Symposium on Theory of Computing, STOC (1998), 249-–258.
[29] M. Luby, M. Mitzenmacher, A. Shokrollahi, D.A. Spielman: Efficient erasure correcting codes. IEEE Transactions on Information Theory, 47
(2001), 569-–584.
[30] M. Luby, M. Mitzenmacher, A. Shokrollahi, D.A. Spielman: Improved low-density parity-check codes using irregular graphs. IEEE Transactions on Information Theory, 47 (2001), 585—598.
[31] M. Luby, M. Mitzenmacher, A. Shokrollahi, D.A. Spielman, V. Stemann: Practical loss-resilient codes. Proceedings of the 29th annual ACM
Symposium on Theory of Computing STOC (1997), 150—159.
[32] A. Manoel, F. Krzakala, M. Mézard, L. Zdeborová: Multi-Layer Generalized Linear Estimation. arXiv preprint, arXiv:1701.06981 (2017).
[33] L. Massoulié: Community detection thresholds and the weak Ramanujan property. Proceedings of the 46th Annual ACM Symposium on
Theory of Computing STOC, (2014), 694-–703.
26
[34] A. Montanari: Tight bounds for LDPC and LDGM codes under MAP decoding. IEEE Transactions on Information Theory 51, (2005), 3221–
3246.
[35] C. Moore: The computer science and physics of community detection: landscapes, phase transitions, and hardness. arXiv preprint,
arXiv:1702.00467 (2017).
[36] T. Richardson, A. Shokrollahi, R. Urbanke: Design of capacity-approaching irregular low-density parity-check codes. IEEE Transactions on
Information Theory 47, (2001), 619-–637.
[37] T. Richardson, R. Urbanke: The capacity of low-density parity check codes undermessage-passing decoding. IEEE Transactions on Information Theory 47, (2001), 599—618.
[38] T. Richardson, R. Urbanke: Modern Coding Theory. Cambridge University Press 2008.
[39] C.E. Shannon: A Mathematical Theory of Communication. Bell System Technical Journal 27, (1948), 379-423, 623-656.
[40] N. Sourlas: Spin-glass models as error-correcting codes. Nature 339, (1989), 693-694.
[41] N. Sourlas: Statistical Mechanics and Error-Correcting Codes. Statistical Mechanics of Neural Networks, Lecture Notes in Physics 368, edited
by L. Garrido, (1990), Springer New York.
[42] N. Sourlas: Statistical Mechanics and Error-Correcting Codes. From Statistical Physics to to Statistical Inference and Back, edited by P.
Grassberger and J.-P. Nadal, (1994), Kluwer Academic, Dordrecht.
JAN VAN DEN B RAND , [email protected], KTH ROYAL I NSTITUTE OF T ECHNOLOGY, S TOCKHOLM , S WEDEN .
N OR JAAFARI , [email protected], G OETHE U NIVERSITY, M ATHEMATICS I NSTITUTE , F RANKFURT, G ERMANY.
27
| 7cs.IT
|
arXiv:1508.03362v1 [math.AG] 9 Aug 2015
RAMIFICATION OF VALUATIONS AND LOCAL RINGS IN POSITIVE
CHARACTERISTIC
STEVEN DALE CUTKOSKY
1. Introduction
In this paper we consider birational properties of ramification of excellent local rings.
Suppose that K ∗ /K is a finite separable field extension, S is an excellent local ring of
∗
K (S has quotient field QF(S) = K ∗ ) and R is an excellent local ring of K such that
dim S = dim R, S dominates R (R ⊂ S and the maximal ideals mS of S and mR of R
satisfy mS ∩ R = mR ) and ν ∗ is a valuation of K ∗ which dominates S (the valuation ring
Vν ∗ of ν ∗ dominates S). Let ν be the restriction of ν ∗ to K.
The notation that we use in this paper is explained in more detail in Section 2.
1.1. Local Monomialization.
Definition 1.1. R → S is monomial if R and S are regular local rings of the same
dimension n and there exist regular systems of parameters x1 , . . . , xn in R and y1 , . . . , yn
in S, units δ1 , . . . , δn in S and an n × n matrix A = (aij ) of natural numbers with nonzero
determinant such that
n
Y
a
(1)
xi = δi
yj ij for 1 ≤ i ≤ n.
j=1
If R and S have equicharacteristic zero and algebraically closed residue fields, then
within the extension R̂ → Ŝ there are regular parameters giving a form (1) with all δi = 1.
More generally, we ask if a given extension R → S has a local monomialization along
the valuation ν ∗ .
Definition 1.2. A local monomialization of R → S is a commutative diagram
R1 → S1 ⊂ Vν ∗
↑
↑
R → S
such that the vertical arrows are products of monoidal transforms (local rings of blowups
of regular primes) and R1 → S1 is monomial.
It is proven in Theorem 1.1 [10] that a local monomialization always exists when K ∗ /K
are algebraic function fields over a (not necessarily algebraically closed) field k of characteristic 0, and R → S are algebraic local rings of K and K ∗ respectively. (An algebraic
local ring is essentially of finite type over k.)
We can also define the weaker notion of a weak local monomialization by only requiring
that the conclusions of Definition 1.2 hold with the vertical arrows being required to be
birational (and not necessarily factorizable by products of monoidal transforms).
Partially supported by NSF.
1
This leads to the following question for extensions R → S as defined in the beginning
of this paper: When does there exist a local monomialization (or at least a weak local
monomialization) of an extension R → S of excellent local rings dominated by a valuation
ν∗ ?
As commented above, the question has a positive answer within algebraic function fields
over an arbitrary field of characteristic zero by Theorem 1.1 [10].
For the question to have a positive answer in positive characteristic or mixed characteristic it is of course necessary that some form of resolution of singularities be true. This
is certainly true in equicharacteristic zero, and is known to be true very generally in dimension ≤ 2 ([2], [29], [7]) and in positive characteristic and dimension 3 ([4], [11], [8] and
[9]). A few recent papers going beyond dimension three are [18], [22], [6], [5], [23], [35],
[36], [24] and [37].
The case of two dimensional algebraic function fields over an algebraically closed field
of positive characteristic is considered in [14], where it is shown that monomialization is
true if R → S is a defectless extension of two dimensional algebraic local rings over an
algebraically closed field k of characteristic p > 0 (Theorem 7.3 and Theorem 7.35 [14]).
We will discuss the important concept of defect later on in this introduction.
In [12], we give an example showing that weak monomialization (and hence monomialization) does not exist in general for extensions of algebraic local rings of dimension ≥ 2
over a field k of char p > 0. We prove the following theorem in [12]:
Theorem 1.3. (Theorem 1.4 [12], Counterexample to local and weak local monomialization) Let k be a field of characteristic p > 0 with at least 3 elements and let n ≥ 2. Then
there exists a finite separable extension K ∗ /K of n dimensional function fields over k,
a valuation ν ∗ of K ∗ with restriction ν to K and algebraic regular local rings A and B
of K and K ∗ respectively, such that B dominates A, ν ∗ dominates B and there do not
exist regular algebraic local rings A′ of K and B ′ of K ∗ such that ν ∗ dominates B ′ , B ′
dominates A′ , A′ dominates A, B ′ dominates B and A′ → B ′ is monomial.
We have that the defect δ(ν ∗ /ν) = 2 in the example of Theorem 1.3 (with ν = ν ∗ |K).
In [10] and [14], a very strong form of local monomialization is established within
characteristic zero algebraic function fields which we call strong local monomialization
(Theorem 5.1 [10] and Theorem 48 [14]). This form is stable under appropriate sequences
of monoidal transforms and encodes the classical invariants of the extension of valuation
rings. In [14], we show that strong local monomialization is true for defectless extensions
of two dimensional algebraic function fields (Theorem 7.3 and Theorem 7.35 [14]). We
give an example in [14] (Theorem 7.38 [14]) showing that strong local monomialization is
not generally true for defect extensions of two dimensional algebraic function fields (over
a field of positive characteristic).
In this paper, we establish that local monomialization (and strong local monomialization) hold for defectless extensions of two dimensional excellent local rings. We will first
state our theorem (which is proven in Section 3), and then we will define and discuss the
defect.
Theorem 1.4. Suppose that R is a 2 dimensional excellent local domain with quotient
field QF(R) = K. Further suppose that K ∗ is a finite separable extension of K and S
is a two dimensional excellent local domain with quotient field QF(S) = K ∗ such that S
dominates R. Let ν ∗ be a valuation of K ∗ which dominates S and let ν be the restriction
2
of ν ∗ to K. Suppose that the defect δ(ν ∗ /ν) = 0. Then there exists a commutative diagram
(2)
R1 → S1 ⊂ Vν ∗
↑
↑
R → S
such that the vertical arrows are products of quadratic transforms along ν ∗ and R1 → S1
is monomial.
The proof of the theorem actually produces stable strong monomialization.
We now define the defect of an extension of valuations. The role of this concept in local
uniformization was observed by Kuhlmann [26] and [27]. A good introduction to the role
of defect in valuation theory is given in [26]. A brief survey which is well suited to our
purposes is given in Section 7.1 of [14]. Suppose that K ∗ /K is a finite Galois extension
of fields of characteristic p > 0. The splitting field K s (ν ∗ /ν) of ν is the smallest field
between K and K ∗ with the property that ν ∗ is the only extension to K ∗ of ν ∗ |L. The
defect δ(ν ∗ /ν) is defined by the identity
[K ∗ : K s (ν ∗ /ν)] = f (ν ∗ /ν)e(ν ∗ /ν)pδ(ν
∗ /ν)
(Corollary to Theorem 25 , Section 12, Chapter VI [40]). In the case when K ∗ /K is only
finite separable, we define the defect by
δ(ν ∗ /ν) = δ(ν ′ /ν) − δ(ν ′ /ν ∗ )
where ν ′ is an extension of ν ∗ to a Galois closure K ′ of K ∗ over K.
The defect is equal to zero if the residue field Vν /mν has characteristic zero (Theorem
24, Section 12, Chapter VI [40]) or if Vν is a DVR (Corollary to Theorem 21, Section 9,
Chapter V [39]).
1.2. Associated graded rings of valuations. The semigroup of R with respect to the
valuation ν is
S R (ν) = {ν(f ) | f ∈ R \ {0}}.
The group generated by S R (ν) is the valuation group Γν of ν which is well understood
([30], [31], [40], [27]); the semigroup can however be extremely complicated and perverse
([17], [15]).
The associated graded ring of ν on R, as defined in [35] and [36], is
M
grν (R) =
Pγ (R)/Pγ+ (R).
γ∈S R (ν)
Here Pγ (R) is the ideal in R of elements of value ≥ γ and Pγ+ (R) is the ideal in R
of elements of value > γ. This ring plays an important role in Teissier’s approach to
resolution of singularity (it is completely realized for Abhyankar valuations in arbitrary
characteristic in [36]).
We always have that QF(grν (R)) = QF(grν (Vν ) and
[QF(grν ∗ (S)) : QF(grν (R))] = f (ν ∗ /ν)e(ν ∗ /ν).
In [19] and [20] it is proven that there exists a strong local monomialization R1 → S1
for defectless extensions R → S of two dimensional algebraic local rings in a two algebraic
function field over an algebraically closed field, which has the property that the induced
extension of associated graded rings along the valuation
(3)
grν (R) → grν ∗ (S)
3
is of finite type and is even a “toric extension”. This result is extended in [16] to the
case of two dimensional algebraic function fields over an arbitrary field of characteristic
zero. These proofs all make use of the technique of generating sequences of a valuation
on a local ring, which is developed by Spivakovsky in [34] for two dimensional regular
local rings with algebraically closed residue fields, and is extended in [15] to arbitrary
regular local rings of dimension two. Unfortunately, this technique is special to dimension
two, and does not extend well to higher dimension local rings, or even to normal local
rings of dimension two (the examples of strange semigroups in [17] and [15] show this).
An interesting construction of generating sequences within a valuation ring which exhibits
the defect of an extension of valuations is given in [38], and a different general construction
of generating sequences is given in [32].
In general, the extension (3) is not of finite type, even for equicharacteristic zero algebraic regular local rings of dimension two (Example 9.4 [15]), so blowing up to reach a
good stable form is required to obtain that (3) has a good form.
It is not difficult to show that the extension (3) is of finite type and is toric when ν ∗ is
an Abhyankar valuation (equality holds in Abhyankar’s inequality (Theorem 1 [1])
trdegS/mS Vν ∗ /mν ∗ + dimQ Γν ∗ ⊗ Q ≤ dim S.
From a special case of Theorem 5.1 [10] (recalled in Theorem 4.1 of this paper) we
give the precise statement of the stable strongly monomial forms R1 → S1 obtained by a
rational rank 1 valuation (dimQ Γν ∗ ⊗ Q = 1) in the case when R → S is an extension of
algebraic local rings in an extension of algebraic function fields over an arbitrary field of
characteristic zero. In this case, there are regular parameters x1 , . . . , xn in R1 and regular
parameters y1 , . . . , yn in S and a unit δ in S1 such that
x1 = δy1e , x2 = y2 , . . . , xn = yn
where e = e(ν ∗ /ν) = |Γν ∗ /Γν |. In this paper, we give a simple proof (Theorem 4.2) that in
the case when R → S is an extension of algebraic local rings in an extension of algebraic
function fields over an arbitrary field of characteristic zero, and ν has rational rank 1 a
strongly monomial extension R1 → S1 has the property that
(4)
grν ∗ (S1 ) ∼
= grν (R1 ) ⊗R1 /mR1 S1 /mS1 [Z]/(Z e − [δ]−1 [x1 ]),
where [γ1 ], [x1 ] are the respective classes in grν (R1 ) ⊗R1 /mR1 S1 /mS1 . The degree of the
extension of quotient fields of grν (R) → grν ∗ (S) is e(ν ∗ /ν)f (ν ∗ /ν), where
f (ν ∗ /ν) = [Vν ∗ /mν ∗ : Vν /mν ].
In particular, the extension of associated graded rings along the valuation is finite and
“toric”.
We show in Theorem 4.3 of this paper that the stable strongly monomial forms found
in Theorem 1.4 of defectless extensions of two dimensional excellent local rings dominated
by a valuation ν ∗ of rational rank 1 have an extension of associated graded rings along the
valuation of the form (4). Since stable forms of Abhyankar valuations have a finite type
“toric extension” as commented above, we conclude that stable strongly monomial forms
of a defectless extension of two dimensional excellent local rings always has a finite type
“toric” extension (3).
In contrast, we do not have such a nice stable form of the extension of associated
graded rings along a valuation which has positive defect, as is shown by the example of
Theorem 38 [14], analyzed in Section 5 of this paper. Using the notation of Section 7.4
4
[14], explained in the following subsection on invariants of stable forms along a valuation,
it follows from Remark 7.44 [14] that the graded domains grν ∗ (Sn ) are integral but not
finite over grν (Rn ) for all n, in contrast to the situation when the defect δ(ν ∗ /ν) is zero in
Theorem 4.3. The quotient fields of grν (Rn ) and grν ∗ (Sn ) are equal under the extension
of Theorem 38 [14], so the degree is 1 = e(ν ∗ /ν)f (ν ∗ /ν) as in the conclusions of Theorem
4.3.
1.3. Invariants of stable forms along a valuation. Suppose that R → S is an inclusion of regular two dimensional algebraic local rings within function fields K and K ∗
respectively, over an algebraically closed field k of positive characteristic p > 0, such that
K ∗ /K is finite separable, S dominates R, and there is a valuation ν ∗ of K ∗ with restriction
ν to K which dominates R and S, such that
1) ν ∗ dominates S.
2) The residue field Vν ∗ /mν ∗ of Vν ∗ is algebraic over S/mS .
3) The value group Γν ∗ of ν ∗ has rational rank 1 (so it is isomorphic as an ordered
group to a subgroup of Q).
It is shown in Corollary 7.30 and Theorem 7.33 [14] that there are sequences of quadratic
transforms along ν ∗ (each vertical arrow is a product of quadratic transforms) constructed
by the algorithm of Section 7.4 [14], where we have simplified notation, writing Rn for Rrn
and Sn for Ssn .
Vν
↑
..
.
(5)
→ Vν ∗
↑
..
.
↑
Rn →
↑
..
.
↑
Sn
↑
..
.
↑
R2 →
↑
R →
↑
S2
↑
S
where each Rn is an algebraic regular local ring of K and Sn is an algebraic regular local
ring of K ∗ such that Sn dominates Rn , and no quadratic transform of Rn factors through
Sn . For n ≫ 0, Rn has regular parameters xn , yn and Sn has regular parameters un , vn
such that there are “stable forms”
(6)
αn
un = γn xnap , vn = xbnn fn
where γn is a unit in Sn , dn = ν n (fn mod xn ) = pβn , where ν n is the natural valuation
of the DVR Sn /xn Sn , with bn , αn , βn ≥ 0 and αn + βn does not depend on n, a does not
depend on n.
In Theorem 7.38 [14], an example is given where K ∗ /K is a tower of two Artin Schreier
extensions with δ(ν ∗ /ν) = 2, a = 1, αn = 1 for all n, βn = 1 for all n and bn = 0
for all n. In particular, this shows that “strong local monomialization” fails for this
extension. However, it is also shown in the example that local monomialization is true for
this extension (by considering different sequences of quadratic transforms above R and S).
5
In Corollary 7.30 and Theorem 7.33 of [14], it is shown that αn + βn is a constant for
n ≫ 0, where αn and βn are the integers defined above which are associated to the stable
forms (5) of an extension of valued two dimensional algebraic function fields. If Γν is
not p-divisible, it is further shown that αn and βn are both constant for n ≫ 0, and pβn
is the defect of the extension. However, if Γν is p-divisible, then it is only shown that
the sum αn + βn is constant for n ≫ 0, and that pαn +βn is the defect of the extension.
In Remark 7.34 [14] it is asked if αn and βn (and some other numbers computed from
the stable forms) are eventually constant in the case when Γν is p-divisible. We give
examples in Section 5, equations (31) - (34), showing that this is not the case, even within
(defect) Artin Schreier extensions. The examples are found by considering a factorization
of the example of Theorem 3.8 [14] into a product of two Artin Schreier extensions, and
computing generating sequences on the intermediary rings.
2. Notation and Preliminaries
2.1. Local algebra. All rings will be commutative with identity. A ring S is essentially
of finite type over R if S is a local ring of a finitely generated R-algebra. We will denote
the maximal ideal of a local ring R by mR , and the quotient field of a domain R by QF(R).
(We do not require that a local ring be Noetherian). Suppose that R ⊂ S is an inclusion
of local rings. We will say that S dominates R if mS ∩ R = mR . If the local ring R is a
domain with QF(R) = K then we will say that R is a local ring of K. If K is an algebraic
function field over a field k (which we do not assume to be algebraically closed) and a
local ring R of K is essentially of finite type over k, then we say that R is an algebraic
local ring of k.
Suppose that K → K ∗ is a finite field extension, R is a local ring of K and S is a local
ring of K ∗ . We will say that S lies over R if S is a localization of the integral closure T of
R in K ∗ . If R is a local ring, R̂ will denote the completion of R by its maximal ideal mR .
Suppose that R is a regular local ring. A monoidal transform R → R1 of R is a local
ring of the form R[ Px ]m where P is a regular prime ideal in R (R/P is a regular local ring)
and m is a prime ideal of R[ Px ] such that m ∩ R = mR . R1 is called a quadratic transform
if P = mR .
2.2. Valuation Theory. Suppose that ν is a valuation on a field K. We will denote by
Vν the valuation ring of ν:
Vν = {f ∈ K | ν(f ) ≥ 0}.
We will denote the value group of ν by Γν . Good treatments of valuation theory are
Chapter VI of [40] and [3], which contain references to the original papers. If ν is a
valuation ring of an algebraic function field over a field k, we insist that ν vanishes on
k \ {0}, and say that ν is a k valuation.
If ν is a valuation of a field K and R is a local ring of K we will say that ν dominates R
if the valuation ring Vν dominates R. Suppose that ν dominates R. A monoidal transform
R → R1 is called a monoidal transform along ν if ν dominates R1 .
Suppose that K ∗ /K is a finite separable extension, ν ∗ is a valuation of K ∗ and ν is the
restriction of ν to K. The ramification index is
e(ν ∗ /ν) = |Γν ∗ /Γν |
and reduced degree is
f (ν ∗ /ν) = [Vν ∗ /mν ∗ : Vν /mν ].
6
The defect δ(ν ∗ /ν) is defined in the introduction to this paper. Its basic properties are
developed in Section 11, Chapter VI [40], [26] and Section 7.1 of [14].
We will call a ring a DVR if it is a valuation ring with value group Z.
2.3. Galois theory of local rings. Suppose that K ∗ /K is a finite Galois extension, R
is a local ring of K and S is a local ring of K ∗ which lies over R. The splitting group
s
Gs (S/R), splitting field K s (S/R) = (K ∗ )G (S/R) and inertia group Gi (S/R) are defined
and their basic properties developed in Section 7 of [3].
2.4. Galois theory of valuations. The Galois theory of valuation rings is developed in
Section 12 of Chapter VI of [40] and in Section 7 of [3]. Some of the basic results we need
are surveyed in Section 7.1 [14]. If we take S = Vν ∗ and R = Vν where ν ∗ is a valuation
of K ∗ and ν is the restriction of ν to K, then we obtain the splitting group Gs (ν ∗ /ν), the
splitting field K s (ν ∗ /ν) and the inertia group Gi (ν ∗ /ν). In Section 12 of Chapter VI of
[40], Gs (ν ∗ /ν) is written as GZ and called the decomposition group. Gi (ν ∗ /ν) is written
as GT . The ramification group GV of ν ∗ /ν is defined in Section 12 of Chapter VI of [40]
and is surveyed in Section 7.1 [14]. We will denote this group by Gr (ν ∗ /ν).
2.5. Semigroups and associated graded rings of a local ring with respect to a
valuation. Suppose that ν is a valuation of field K which dominates a local ring R of K.
We will denote the semigroup of values of ν on S by
S R (ν) = {ν(f ) | f ∈ R \ {0}}.
Suppose that γ ∈ Γν . We define ideals in R
and
Pγ (R) = {f ∈ R | ν(f ) ≥ 0}
Pγ+ (R) = {f ∈ R | ν(f ) > 0}
and define (as in [35]) the associated graded ring of R with respect to ν by
M
grν (R) =
Pγ (R)/Pγ+ .
γ∈S R (ν)
2.6. Birational geometry of two dimensional regular local rings. We recall some
basic theorems which we will make frequent use of.
Theorem 2.1. (Theorem 3 [1]) Suppose that K is a field, and R is a regular local ring
of dimension two of K. Suppose that S is another 2 dimensional regular local ring of K
which dominates R. Then there exists a unique sequence of quadratic transforms of R
of R which factor R → S.
R → R1 → · · · → R n = S
Lemma 2.2. (Lemma 12 [1]) Suppose that A is a two dimensional regular local ring of a
field K and ν is a valuation of K which dominates ν. Let
R → R1 → R2 → · · ·
be the infinite sequence of quadratic transforms along ν. Then
Vν ∗ = ∪ ∞
i=1 Ri .
We also make use of the fact that “embedded resolution of singularities ” is true within
a regular local ring of dimension 2 (Theorem 2 [1]), and the fact that resolution of singularities is true for two dimensional excellent local rings ([29], [7]).
7
3. Local monomialization of two dimensional defectless extensions
In this section we prove Theorem 1.4, establishing local monomialization for defectless
extensions of two dimensional excellent local domains. This extends the result for extensions of two dimensional algebraic local rings in two dimensional algebraic function fields
over an algebraically closed field in Theorem 7.3 and 7.35 [14].
We indicate the differences between the proof in [14] of the analogue of Theorem 1.4
for algebraic local rings over an algebraically closed field, and the proof of Theorem 1.4
in this paper. The essential case is of rational rank 1 valuations (Theorem 3.3). An
essential ingredient in the proof is the computation of complexity in Proposition 3.1,
which generalizes Proposition 7.2 [14]. The steps of the proof are the same, but some of
the individual calculations require different methods, as we do not have coefficient fields
in general in the situation of this paper, and the completions of local rings are no longer
extensions of power series rings over a field. In [14], the analogue of Theorem 3.3 is deduced
as a consequence of a detailed analysis of stable forms (Theorem 7.33 [14]) which makes
essential use of the assumption that there is no residue field extension (the ground field is
algebraically closed). In this paper, we give a different, more direct argument to deduce
Theorem 3.3.
The proof of Theorem 1.4 actually produces stable strong monomialization. We first
establish strong monomialization in the two essential cases of the theorem, and give the
proof of Theorem 1.4 at the end of this section. We will make use of the list of good
properties of excellent rings given in Scholie 7.8.3 [21].
3.1. Degree formulas. In this subsection we generalize the formulas of Proposition 7.2
[14].
Proposition 3.1. Suppose that R and S are two dimensional regular excellent local rings
such that S dominates R and K ∗ = QF(S) is a finite separable extension of K = QF(R),
R has a regular system of parameters u, v and S has a regular system of parameters x, y
such that there is an expression
u = γxa , v = xb f
where a > 0, b ≥ 0, γ is a unit in S, x 6 | f in S and f is not a unit in S. Then there exist
inclusions
R → R0 → S
of local rings with the following properties: R0 is a two dimensional normal local ring of
K which is essentially of finite type over R such that S lies above R0 . Let ν be the natural
valuation of the DVR S/xS. Then we have
[QF(Ŝ) : QF(R̂0 )] = ad[S/mS : R/mR ]
where d = ν(f mod x).
Proof. First suppose that b > 0. Let s = gcd(a, b), I be the ideal which is the integral
a
b
closure of (u s , v s ) in R. Let
a
vs
ϕ= b
us
and
1
.
I
R0 = R
b
u s mS ∩R 1b I
us
8
R0 is normal since I is integrally
Lclosed. (Powers of I are integrally closed by Theorem
2’, Appendix 5 [40], so the ring n≥0 I n is integrally closed and R0 is a local ring of the
L
normal scheme Proj( n≥0 I n )). The elements xa , ϕ ∈ mR0 S so mR0 S is mS -primary.
By Lemma 9 [1], there exists a rank 2 valuation ν of K ∗ which dominates S. Let A
be the integral closure of R0 in K ∗ and let T = AmS ∩A . By Theorem 1 [1], Vν /mν is
finite over T /mT so S/mS is finite over T /mT . T is normal of dimension two (since T is
excellent and normal). Since mT S is mS -primary, we have that T = S by the version of
Zariski’s main theorem in (10.7) [4]. Thus S lies over R0 .
Let W ∗ be the DVR W ∗ = S(x) . ϕ ∈ R0 is neither a unit nor divisable by x in S. Thus
the prime ideal p = xS ∩ R0 has height one in R0 . Thus
W = (R0 )p = W ∗ ∩ K
is a DVR.
Let t be a regular parameter in W . W is the valuation ring of the valuation ordt . Since
I is a monomial ideal, the value group Z of ordt is generated by ordt u and ordt v. Thus
gcd(ordt u, ordt v) = 1. Since ordt ϕ = 0, we have that ordt u = as and ordt v = sb .
We have that
b
a b
IW = u s W = (t s s ).
Since
ab
IW ∗ = (x s ),
we have that tW ∗ = (xs ), so the ramification index of W ∗ /W is
e(W ∗ /W ) = s.
The integrally closed ideal I is generated by all monomials um v n such that
n
m
+
≥ 1.
b/s a/s
Since
a b
gcd( , ) = 1,
s s
we have that
ordx (um v n ) = ma + nb >
(7)
ab
s
for such a monomial provided
b
a
(m, n) 6∈ {( , 0), (0, )}.
s
s
Thus for um v n ∈ I,
um v n
b
us
unless
∈p
b
a
um v n = u s or um v n = v s .
Since ν(ϕ) =
(8)
ad
s
> 0, we have that
R0 /p ∼
= (R/mR [ϕ])(ϕ)
where ϕ is the residue of ϕ in R0 /p, and
(9)
R0 /mR0 = R/mR .
9
By Corollary 1 of Section 2, Chapter II of Local Fields [33] and (ii) of Theorem 1,
[
Section 3 of Chapter II [33], and by (9) and the fact that ϕ is a regular parameter in R
0 /p
by (8), we have that the inclusion
[
\
R
0 /p → S/xS
is a finite extension of complete DVRs, and
\ : QF(R
[
\ [
\ [
[QF(S/xS)
0 /p)] = e(S/xS/R0 /p)f (S/xS/R0 /p)
= e(S/xS/R0 /p)f (S/xS/R0 /p) =
ad
s [S/mS
: R/mR ].
We have that mR S ⊂ pS so xS is the only prime ideal of S lying over p. Further, pR̂0
and xŜ are prime ideals since R0 /p and S/xS are regular local rings. R̂0 and Ŝ are normal
since R0 and S are normal and excellent. Also, Ŝ is a finite extension of R̂0 by (10.13) and
(10.2) of [4], since S lies over R0 . Thus QF(Ŝ) is a finite field extension of QF(R̂0 ) and
Ŝ is the integral closure of R̂0 in QF(Ŝ). We obtain that Ŵ ∗ = ŜxŜ is the unique DVR of
QF(Ŝ) which dominates Ŵ = (R̂0 )pR̂0 .
Thus, by Theorem 20, page 60 [40],
[QF(Ŝ) : QF(R̂0 )] = e(Ŵ ∗ /Ŵ )f (Ŵ ∗ /Ŵ ).
We have that
e(Ŵ ∗ /Ŵ ) = e(W ∗ /W ) = s
since R0 and S are analytically unramified (as they are excellent), and
f (Ŵ ∗ /Ŵ ) = [QF(Ŝ/xŜ) : QF(R̂0 /pR̂0 )] =
ad
[S/mS : R/mR ].
s
Thus
[QF(Ŝ) : QF(R̂0 )] = ad[S/mS : R/mR ].
Now suppose that b = 0. Then taking R0 = R, W = RuR and W ∗ = SxS , a simpler
variant of the above proof shows that e(Ŵ ∗ /Ŵ ) = a, f (Ŵ ∗ /Ŵ ) = d[S/mS : R/mR ] and
[QF(Ŝ) : QF(R̂)] = ad[S/mS : R/mR ].
Proposition 3.2. Suppose that R and S are two dimensional excellent regular local rings,
such that S dominates R and K ∗ = QF(S) is a finite separable extension of K = QF(R),
R has a regular system of parameters u, v and S has a regular system of parameters x, y
such that there is an expression
u = γ1 xa y c , v = γ2 xb y d
where γ1 , γ2 are units in S and ad − bc 6= 0. Then there exist inclusions
R → R0 → S
of local rings with the following properties: R0 is a two dimensional normal local ring of
K which is essentially of finite type over R such that S lies above R0 and
[QF(Ŝ) : QF(R̂0 )] = |ad − bc|[S/mS : R/mR ].
10
Proof. The conclusions of this proposition follow from Proposition 3.1 if one of a, b, c, d is
zero, so we may assume that a, b, c, d are all positive. After possibly interchanging x and
y, it can be assumed that ad − bc > 0. The proof of this proposition is a generalization of
the proof of Proposition 3.1, and we give an outline of the proof, indicating the essential
differences.
Let s = gcd(a, b) and s′ = gcd(c, d). Let I ⊂ R (respectively I ′ ⊂ R) be the integral
d
c
a
b
closure of the the ideal (u s , v s ) (respectively (u s′ , v s′ )). We have that
II ′
R0 = R
c
b
u s v s′ mS ∩R bII ′ c
u s v s′
is a normal local ring (products of integrally closed ideals in a 2-dimensional regular local
ring are integrally closed by Theorem 2’, Appendix 5 [40]).
Let p = xS ∩ R0 . Let W ∗ be the DVR W ∗ = S(x) and let W be the DVR W = (R0 )p =
∗
W ∩ K. Let
I
′
p = (xS) ∩ R
b
us
and let
a
vs
ϕ= b.
us
h i
ϕ 6∈ p′ , so p′ is a height one prime in R Ib . Thus
us
′′
W =R
I
b
us
p′
is a DVR dominated by W ∗ . Thus W = W ′′ . As in the proof of Proposition 3.1, we
conclude that
e(W ∗ /W ) = s
and that
R0 /p ∼
= (R/mR [ϕ])(ϕ)
where ϕ is the residue of ϕ in R0 /p. We have that
ν(ϕ) =
ad − bc
s
where ν is the natural valuation of S/xS, and R0 /mR0 ∼
= R/mR . We calculate (as in the
proof of Proposition 3.1) that
ad − bc
\ : QF(R
[
f (Ŵ ∗ /Ŵ ) = [QF(S/xS)
[S/mS : R/mR ]
0 /p)] =
s
and that
[QF(Ŝ) : QF(R̂0 )] = e(Ŵ ∗ /Ŵ )f (Ŵ ∗ /Ŵ ) = (ad − bc)[S/mS : R/mR ].
11
3.2. Rational rank 1. Throughout this subsection we will have the following assumptions. Suppose that R is a 2 dimensional excellent local domain with quotient field
QF(R) = K. Further suppose that K ∗ is a finite separable extension of K and S is a
2 dimensional local domain with quotient field QF(S) = K ∗ such that S dominates R.
Suppose that ν ∗ is a valuation of K ∗ such that
1) ν ∗ dominates S.
2) The residue field Vν ∗ /mν ∗ of Vν ∗ is algebraic over S/mS .
3) The value group Γν ∗ of ν ∗ has rational rank 1 (so it is isomorphic as an ordered
group to a subgroup of Q).
Let ν be the restriction of ν ∗ to K.
Theorem 3.3. Suppose that the defect δ(ν ∗ /ν) = 0. Then there exists a commutative
diagram
(10)
R1 → S1 ⊂ Vν ∗
↑
↑
R → S
such that the vertical arrows are products of quadratic transforms along ν ∗ and
1) R1 and S1 are two dimensional regular local rings.
2) R1 has a regular system of parameters u1 , v1 and S1 has a regular system of parameters x1 , y1 and there exists a unit γ1 ∈ S1 such that
u1 = γ1 xe1 and v1 = y1
where e = e(ν ∗ /ν) = |Γν ∗ /Γν |, and the class of ν ∗ (x1 ) is a generator of the group
Γν ∗ /Γν ∼
= Ze .
3) Vν ∗ /mν ∗ is the join Vν ∗ /mν ∗ = (Vν /mν )(S1 /mS1 ) and [S1 /mS1 : R1 /mR1 ] =
f (ν ∗ /ν) = [Vν ∗ /mν ∗ : Vν /mν ].
It will follow from our proof that there exists a diagram (10) such that the conditions
1), 2) and 3) of Theorem 3.3 are stable under further appropriate sequences of quadratic
transforms above R and S.
Proposition 3.4. There exists a local ring R′ of K which is essentially of finite type over
R, is dominated by ν and dominates R such that if we have a commutative diagram
(11)
Vν → Vν ∗
↑
↑
R1 → S 1
↑
R′
↑
↑
R → S
where R1 is a regular local ring of K which is essentially of finite type over R and dominates
R, S1 is a regular local ring of K ∗ which is essentially of finite type over S and dominates
S, R1 has a regular system of parameters u, v and S1 has a regular system of parameters
x, y such that there is an expression
u = γxa , v = xb f
12
where a > 0, b ≥ 0, γ is a unit in S, x 6 | f in S1 and f is not a unit in S1 , then
ad[S1 /mS1 : R1 /mR1 ] = e(ν ∗ /ν)f (ν ∗ /ν)pδ(ν
(12)
∗ /ν)
.
where d = ν(f mod x) with ν being the natural valuation of the DVR S/xS.
Proof. We first prove the proposition with the assumption that K ∗ /K is Galois with Galois
group G.
Let g be the number of extensions of ν to K ∗ . Writing f (ν ∗ /ν) = f0 ps where gcd(f0 , p) =
1 and e(ν ∗ /ν) = e0 pt where gcd(e0 , p) = 1, we have by the Corollary to Theorem 25, page
78 [40] that [G : Gs (ν ∗ /ν)] = g, [Gs (ν ∗ /ν) : Gi (ν ∗ /ν)] = f0 , [Gi (ν ∗ /ν) : Gr (ν ∗ /ν)] = e0
∗
and |Gr (ν ∗ /ν)| = ps+t+δ(ν /ν) , so that
[K ∗ : K] = ge(ν ∗ /ν)f (ν ∗ /ν)pδ(ν
Since
(13)
[K s (ν ∗ /ν)
∗ /ν)
: K] = g, we have that
[K ∗ : K s (ν ∗ /ν)] = e(ν ∗ /ν)f (ν ∗ /ν)pδ(ν
By the argument at the top of page 86 in [3], there exists
diagram
(14)
.
∗ /ν)
R′
.
as above such that for any
Vν → Vν ∗
↑
↑
R∗ → S ∗
↑
↑
′
R
↑
↑
R → S
where R∗ and S ∗ are normal local rings of K and K ∗ respectively such that R∗ is essentially
of finite type over R′ and dominates R′ and S ∗ lies over R∗ , we have that
(15)
Gs (S ∗ /R∗ ) = Gs (ν ∗ /ν).
Now applying Proposition 3.1 to an extension R1 → S1 satisfying (11) (so that R1
dominates R′ ), we have a commutative diagram
Vν ∗
↑
R∗
↑
R1
→ Vν
↑
→ S1
ր
such that R∗ and S1 are normal local rings such that S1 lies over R∗ and
(16)
[QF(Ŝ1 ) : QF(R̂∗ )] = ad[S1 /mS1 : R1 /mR1 ].
Let S ′ = K s (ν ∗ /ν ∗ ) ∩ S1 . Then R∗ → S ′ is unramified with S ′ /mS ′ = R∗ /mR∗ by
Theorem 1.47 [3], since K s (ν ∗ /ν ∗ ) = K s (S1 /R∗ ) by (15). Thus R̂∗ = Ŝ ′ by (10.1) [4]. We
have that
(17)
[K ∗ : K s (ν ∗ /ν)] = [QF(Ŝ ′ ) : QF(R̂∗ )]
by II of Proposition 1 (page 498) [2], since there is a unique local ring in K ∗ lying over
S ′ . Now combining (13), (17) and (16), we obtain formula (12). (The proof in [2] is valid
in our more general situation since S ′ is excellent.)
We will now establish the proposition in the general case, when K ∗ /K is only assumed
to be finite and separable. Let K ′ be a Galois closure of K ∗ /K, and let ν ′ be an extension
13
of ν ∗ to K ′ . By the Galois case, there exists a normal local ring R′ of K giving the property
of Proposition 3.4 within the extension K ′ /K, and a normal local ring S ′ of K ∗ giving
the property of Proposition 3.4 within the extension K ′ /K ∗ . We can choose R′ and S ′ so
that S ′ lies over R′ .
Now suppose that we are given a diagram (11). We then have that S ′ ⊂ S1 . We have
by Proposition 3.1 local rings R0 of K and S1 of K ∗ such that S1 lies over R0 and
(18)
[QF(Ŝ1 ) : QF(R̂0 )] = ad[S1 /mS1 : R1 /mR1 ].
Let T be the integral closure of S1 in K ′ and let T ′ = Tmν ′ ∩T . By (13) and (17), we have
that
′
[QF(T̂ ′ ) : QF(R̂0 )] = e(ν ′ /ν)f (ν ′ /ν)pδ(ν /ν)
and
′
∗
[QF(T̂ ′ ) : QF(Ŝ1 )] = e(ν ′ /ν ∗ )f (ν ′ /ν ∗ )pδ(ν /ν ) .
Thus
[QF(T̂ ′ ) : QF(R̂0 )]
∗
[QF(Ŝ1 ) : QF(R̂0 )] =
= e(ν ∗ /ν)f (ν ∗ /ν)pδ(ν /ν)
′
[QF(T̂ ) : QF(Ŝ1 )]
δ
since e, f and p are multiplicative. Now formula (12) follows from (18).
Proposition 3.5. There exists a local ring R′′ of K which is essentially of finite type over
R, is dominated by ν and dominates R such that if we have a commutative diagram
(19)
Vν → Vν ∗
↑
↑
R1 → S 1
↑
↑
R′′
↑
↑
R → S
where R1 is a regular local ring of K which is essentially of finite type over R′′ , S1 is a
regular local ring of K ∗ which is essentially of finite type over S and dominates S, R1 has
a regular system of parameters u, v and S1 has a regular system of parameters x, y such
that there is an expression
u = γxa , v = xb f
where a > 0, b ≥ 0, γ is a unit in S, x 6 | f in S1 and f is not a unit in S1 and there exists
a unit τ ∈ S1 and n ∈ Z+ such that h = τ xn , then
1) ν ∗ (x) is a generator of Γν ∗ /Γν ∼
= Ze(ν ∗ /ν) ,
2) Vν ∗ /mν ∗ = (Vν /mν )(S1 /mS1 ),
3)
∗
ad[S1 /mS1 : R1 /mR1 ] = e(ν ∗ /ν)f (ν ∗ /ν)pδ(ν /ν)
where d = (f mod x) with ν being the natural valuation of the DVR S/xS.
Proof. Let R′ be the local ring of the conclusions of Proposition 3.4 and let g1 , . . . , gs ∈ Vν ∗
be such that the classes of g1 , . . . , gs in Vν ∗ /mν ∗ are a Vν /mν basis. Let R′′ be a regular
local ring of K which dominates R′ and is essentially of finite type over R′ such that
g1 , . . . , gs are in the integral closure of R′′ in K ∗ .
Suppose that we have a diagram (19). Then g1 , . . . , gs ∈ S1 so conclusion 2) holds.
Since ν ∗ (h) = nν ∗ (x) is a generator of Γν ∗ /Γν we have that ν ∗ (x) is a generator of
Γν ∗ /Γν .
14
Finally, 3) holds by Proposition 3.4, since R′′ dominates R′ .
Corollary 3.6. Let assumptions be as in Proposition 3.5, and further assume that δ(ν ∗ /ν) =
0. Let R′′ be the local ring of the conclusions of Proposition 3.5. Suppose that we have a
commutative diagram (19). Then
a = e(ν ∗ /ν), d = 1 and [S1 /mS1 : R1 /mR1 ] = f (ν ∗ /ν).
Proof. We have that aν ∗ (x) ∈ Γν and ν ∗ (x) is a generator of Γν ∗ /Γν ∼
= Ze(ν ∗ /ν) by 1) of
Proposition 3.5 so e(ν ∗ /ν) divides a, and thus
(20)
Further,
(21)
a ≥ e(ν ∗ /ν).
[S1 /mS1 : R1 /mR1 ] ≥ f (ν ∗ /ν)
since Vν ∗ /mν ∗ = (Vν /mν )(S1 /mS1 ) by 2) of Proposition 3.5, so S1 /mS1 contains a basis
of Vν ∗ /mν ∗ over Vν /mν . Further, we have that
(22)
ad[S1 /mS1 : R1 /mR1 ] = e(ν ∗ /ν)f (ν ∗ /ν)
by 3) of Proposition 3.5 since δ(ν ∗ /ν) = 0. Thus by equations (20), (21) and (22), we have
a[S1 /mS1 : R1 /mR1 ] ≥ e(ν ∗ /ν)f (ν ∗ /ν) = ad[S1 /mS1 : R1 /mR1 ]
giving the conclusions of the corollary.
We now give the proof of Theorem 3.3. Let R′′ be the ring of the conclusions of
Proposition 3.5. We may assume, after replacing R and S with appropriate sequences of
quadratic transforms of R and S, that R and S are regular, R dominates R′′ and R has
regular parameters u, v and S has regular parameters x, y such that
(23)
u = γxa , v = xb y
and h = τ xn where h ∈ Vν ∗ is such that ν ∗ (h) is a generator of Γν ∗ /Γν and τ is a unit in S.
(We have that d = 1 by Corollary
√ 3.6). Let S → S1 be the smallest sequence of quadratic
transforms along ν ∗ such that xyS1 is a prime ideal (this is possible since ν ∗ has rational
rnak 1 so ν ∗ (x) and ν ∗ (y) are rationally dependent),
√ and let R → R1 be the smallest
sequence of quadratic transforms along ν such that uvR1 is a prime ideal. We will show
that S1 dominates R1 and we have regular parameters u1 , Q in R1 and x1 , Q in S1 and
e(ν ∗ /ν)
a unit γ ′ in S1 such that u1 = γ ′ x1
. Since these conclusions will then hold under
further sequences of quadratic transforms, we will then have established the conclusions
of the theorem, and the remark on stability following the statement of Theorem 3.3.
We have that
S1 = S[x1 , y1 ]mν ∗ ∩S[x1,y1 ]
where
(24)
m′
n′
n1 1
1
1
x = xm
1 y 1 , y = x1 y 1
with m1 n′1 − n1 m′1 = ±1, ν ∗ (x1 ) > 0 and ν ∗ (y1 ) = 0.
S[x1 , y1 ]/x1 S[x1 , y1 ] ∼
= S/mS [y1 ]
is a polynomial ring over S/mS . Let f ∈ S/mS [y1 ] be the monic generator of
(mν ∗ ∩ S[x1 , y1 ])/x1 S[x1 , y1 ].
15
There exists P ∈ S[x1 , y1 ] such that the residue of P in S[x1 , y1 ]/x1 S[x1 , y1 ] is f . Then
x1 , P are regular parameters in S1 .
Now substitute (24) into (23), to obtain
m′ a
1a
u = γxm
y1 1
1
m′ b+n′
v = x1m1 b+n1 y1 1 1 .
Let s = gcd(m1 a, m1 b + n1 ). By a sequence of substitutions û0 = u, v̂0 = v, and
ûi = ûi+1 , v̂i = ûi+1 v̂i+1
or
ûi = ûi+1 v̂i+1 , v̂i = v̂i+1
for 0 ≤ i ≤ λ we obtain an expression
ûλ = γ c1 xs1 y1t1
v̂λ = γ c2 xs1 y1t2 .
We have that
s|t1 − t2 | = Det
s t1
s t2
=
m1 a
m′1 a
′
(m1 b + n1 ) (m1 b + n′1 )
We thus have that t2 − t1 6= 0. Set
u1 = ûλ , v1 =
=a
m1 m′1
n1 n′1
= a.
v̂λ
if t2 − t1 > 0,
ûλ
ûλ
, v1 = v̂λ if t2 − t1 < 0.
v̂λ
Without loss of generality, we may assume that t2 − t1 > 0. We then have an expression
u1 =
m′
n′
n2 2
2
2
u = um
1 v1 , v = u1 v1
with m2 n′2 − n2 m′2 = ±1, ν(u1 ) > 0 and ν(v1 ) = 0. Now
R1 = R[u1 , v1 ]mν ∩R[u1 ,v1 ] .
We have that R[u1 , v1 ]/u1 R[u1 , v1 ] ∼
= R/mR [v1 ] is a polynomial ring over R/mR . Let
g ∈ R/mR [v1 ] be the monic generator of
(mν ∩ R[u1 , v1 ])/u1 R[u1 , v1 ].
There exists Q ∈ R[u1 , v1 ] such that the residue of Q in R[u1 , v1 ]/u1 R[u1 , v1 ] is g. u1 , Q
are regular parameters in R1 . We have an expression
u1 = (γ c1 y1t1 )xs1 , v1 = γ c2 −c1 y1t2 −t1 .
The inclusion R → S induces a natural homomorphism
R/mR [v1 ] ∼
= R[u1 , v1 ]/u1 R[u1 , v1 ] → S[x1 , y1 ]/x1 S[x1 , y1 ] ∼
= S/mS [y1 ].
Let 0 6= γ0 be the residue of γ in S/mS . Then γ0 is the residue of γ in S/mS [y1 ] and
γ0c2 −c1 y1t2 −t1 is the residue of v1 in S/mS [y1 ]. The residue of Q in S/mS [y1 ] is g(γ0 y1t2 −t1 )
which is nonzero. Thus b(S1 /R1 ) = 0. We further have that a(S1 /R1 ) = e(ν ∗ /ν),
d(S1 /R1 ) = 1 and [S1 /mS1 : R1 /mR1 ] = f (ν ∗ /ν) by Proposition 3.5. This completes
the proof of Theorem 3.3.
16
3.3. Rational rank 2. We have the following Theorem 3.7 for rational rank 2 valuations
dominating a two dimensional excellent local domain, most of whose simple proof is as on
page 40 of [14]. The statement that the extension is defectless follows from the argument
of Proposition 3.4 which establishes (12), replacing references to Proposition 3.1 with
Proposition 3.2. We obtain the formula
|ad − bc|[S1 /mS1 : R1 /mR1 ] = e(ν ∗ /ν)f (ν ∗ /ν)pδ(ν
∗ /ν)
instead of (12). Since e(ν ∗ /ν) = |ad − bc|, we obtain that δ(ν ∗ /ν) = 0. Related results are
proven for valuations of maximal rational rank in algebraic function fields in Theorem 3.1
[25]
Theorem 3.7. Suppose that R is a 2 dimensional excellent local domain with quotient
field QF(R) = K. Further suppose that K ∗ is a finite separable extension of K and S is
a 2-dimensional local domain with quotient field QF(S) = K ∗ such that S dominates R.
Suppose that ν ∗ is a valuation of K ∗ such that
1) ν ∗ dominates S.
2) The residue field Vν ∗ /mν ∗ of Vν ∗ is algebraic over S/mS .
3) The value group Γν ∗ of ν ∗ has rational rank 2.
Let ν be the restriction of ν ∗ to K. Then the defect δ(ν ∗ /ν) = 0, and there exists a
commutative diagram
R1 → S1 ⊂ Vν ∗
↑
↑
R → S
(25)
such that the vertical arrows are products of quadratic transforms along ν ∗ and
1) R1 and S1 are two dimensional regular local rings.
2) R1 has a regular system of parameters u1 , v1 and S1 has a regular system of parameters x1 , y1 and there exist units γ1 , τ1 ∈ S1 such that
u1 = γ1 xa1 y1b and v1 = τ1 xc1 y1d
where
∗
e = e(ν /ν) = |Γν ∗ /Γν | = Det
a b
c d
and the classes of ν ∗ (x1 ), ν ∗ (y1 ) are generators of the group Γν ∗ /Γν ∼
= Z2 /AZ2 .
3) Vν ∗ /mν ∗ is the join Vν ∗ /mν ∗ = (Vν /mν )(S1 /mS1 ) and [S1 /mS1 : R1 /mR1 ] =
f (ν ∗ /ν) = [Vν ∗ /mν ∗ : Vν /mν ].
There exists a diagram (25) such that the conditions 1), 2) and 3) of Theorem 3.7 are
stable under further appropriate sequences of quadratic transforms above R and S.
3.4. Proof of Theorem 1.4. We now give the proof of Theorem 1.4.
trdegS/mS Vν ∗ /mν ∗ + dimQ Γν ∗ ⊗ Q ≤ 2
by Abhyankar’s inequality (Theorem 1 [1]).
If trdegR/mR S/mS = 1, then Vν ∗ and Vν are DVRs and algebraic local rings of K ∗
and K respectively, so that Vν → Vν ∗ is a monomial mapping, which is obtained from R,
respectively S by a sequence of quadratic transforms along ν ∗ (there exists a sequence of
quadratic transforms R → R∗ along ν such that Vν /mν is algebraic over R/mR . Thus
17
R∗ = Vν be Zariski’s Main Theorem, (10.7) [4]). (For this type of valuation we must have
that δ(ν ∗ /ν) = 0 by Theorem 20, Section 11, Chapter VI [40]).
Suppose that dimQ Γν ∗ ⊗ Q = 2. Then Vν ∗ /mν ∗ is algebraic over S/mS by Abhyankar’s
inequality. In this case there exists a monomialization by Theorem 3.7 (we also always
have that δ(ν ∗ /ν) = 0 for this type of valuation by Theorem 3.7).
The remaining case is when Vν ∗ /mν ∗ is algebraic over S/mS , dimQ Γν ∗ ⊗ Q = 1 and
δ(ν ∗ /ν) = 0. The existence of a monomialization in this case follows from Theorem 3.3.
4. Extensions of Associated Graded Rings of Valuations
In this section, we extend the results of [19] and [20] calculating the extension of associated graded rings of a valuation for defectless extensions of 2 dimensional algebraic
function fields over an algebraically closed field, and of [15] for 2 dimensional algebraic
function fields over a (not necessarily closed) characteristic zero field. We refer to the
introduction of this paper for a discussion of this problem.
We recall the following theorem on strong monomialization of rational rank 1 valuations
in an extension of characteristic zero function fields from [10].
Theorem 4.1. (The rational rank 1 case of Theorem 5.1 [10], Theorem 6.1 [14] and
Theorem 6.5 [13]) Let K an algebraic function field over a field k of characteristic zero,
K ∗ a finite algebraic extension of K and ν ∗ a rational rank 1 k valuation of K ∗ . Suppose
that S ∗ is an algebraic local ring with quotient field K ∗ which is dominated by ν ∗ and R∗
is an algebraic local ring with quotient field K which is dominated by S ∗ . Let ν be the
restriction of ν ∗ to K. Then there exists a commutative diagram
R → S ⊂ Vν ∗
↑
↑
∗
R → S∗
(26)
such that the vertical arrow are product of monoidal transforms along ν ∗ and
1) R and S are regular local rings of dimension equal to n = trdegk Vν ∗ /mν ∗ .
2) R has a regular system of parameters x1 , . . . , xn and S has a regular system of
parameters y1 , . . . , yn and there exists a unit δ ∈ S such that
x1 = δy1e and xi = yi for i ≥ 2
3)
4)
5)
6)
(27)
where e = e(ν ∗ /ν) = |Γν ∗ /Γν |.
The class of ν(y1 ) is a generator of the group Γν ∗ /Γν .
Vν ∗ /mν ∗ = (Vν /mν )(S/mS ).
[S/mS : R/mR ] = f = f (ν ∗ /ν) = [Vν ∗ /mν ∗ : Vν /mν ]
The conclusions of 1) and 3) - 5) of the theorem continue to hold for R1 → S1
whenever there exists a commutative diagram
R1 → S1 ⊂ Vν ∗
↑
↑
R → S
such that the vertical arrow are product of monoidal transforms along ν ∗ and 2)
holds for R1 → S1 .
We now show that with the assumptions of the conclusions of Theorem 4.1, we further
have a very simple description of the extension of associated graded rings of the valuations.
18
Theorem 4.2. Let assumptions be as in Theorem 4.1, and let R → S be as in the
conclusions of Theorem 4.1. Then we have a natural isomorphism of graded rings
grν ∗ (S) ∼
= grν (R) ⊗R/mR S/mS [Z]/(Z e − [δ]−1 [x1 ]),
where [γ1 ], [x1 ] are the respective classes in grν (R1 ) ⊗R1 /mR1 S1 /mS1 . The degree of the
extension of quotient fields of grν (R) → grν ∗ (S) is e(ν ∗ /ν)f (ν ∗ /ν).
Proof. Let γ1 , . . . , γf ∈ S be such that their residues γ 1 , . . . , γ f in S/mS are a basis of
S/mS over R/mR . We now establish the following formula:
P
(28)
Suppose that h1 , . . . , hf ∈ R. Then ν ∗ ( fi=1 γi hi ) = min{ν(hi )}.
Without loss of generality, we may suppose that ν(h1 ) is this minimum. Then
all i so we have classes
hi
[ ] ∈ Vν /mν ⊂ Vν ∗ /mν ∗ .
h1
P
If ν ∗ ( fi=1 γi hi ) > ν ∗ (h1 ), then
X hi
γ i [ ] = 0 in Vν ∗ /mν ∗ ,
h1
hi
h1
∈ Vν for
which is impossible since γ 1 , . . . , γ f are a basis of Vν ∗ /mν ∗ over Vν /mν , by the assumptions
of the theorem. We have thus established formula (28).
Suppose that z ∈ S. Let t be a positive integer such that tν ∗ (mS ) > ν ∗ (z). We have
an expression
z=
f
e−1 X
X
j=0 i=1
mtS .
gij γi y1j + w
where gij ∈ R for all i, j and w ∈
We will next establish the following formula:
(29)
ν ∗ (z) = min{ν(gij ) + jν ∗ (y1 )}.
With our assumption on t,
f
e−1 X
X
gij γi y1j .
ν ∗ (z) = ν ∗
j=0 i=1
Set
hj =
f
X
i=1
gij γi for 0 ≤ j ≤ e − 1.
By (28), ν ∗ (hj ) ∈ Γν for all j. For 0 ≤ j ≤ e − 1, we have
ν ∗ (hk y1k ) − ν ∗ (hj y1j ) = (k − j)ν ∗ (y1 ) + ν ∗ (hk ) − ν ∗ (hj ).
Since the class of ν ∗ (y1 ) has order e in Γν ∗ /Γν , we have that
(30)
ν ∗ (hk y1k ) 6= ν ∗ (hj y1j ) for j 6= k.
Formula (29) now follows from (30) and (28).
For γ ∈ Γν , we have a homomorphism
Pγ (R)/Pγ+ (R) ⊗R/mR S/mS → Pγ (S)/Pγ+ (S)
19
defined by gi ⊗ γi 7→ gi γi . This map is 1-1 by the proof of (28). Thus
M
Pγ (R)/Pγ+ (R) ⊗R/mR S/mS
γ∈Γν
is a graded subalgebra of
M
τ ∈Γν ∗
Pτ (S)/Pτ+ (S).
Suppose that h ∈ S. Let τ = ν ∗ (h). By (30), there exists a unique i with 0 ≤ i ≤ e − 1
and γ ∈ Γν such that τ = γ + iν ∗ (y1 ). Further,
f
X
γj gj y1i + h2
h=
j=1
where gj ∈ R with ν(gj ) = γ for 1 ≤ j ≤ f and h2 ∈ S satisfies ν ∗ (h2 ) > τ . Thus
Pτ (S)/Pτ+ (S) = Pγ (R)/Pγ+ (R) ⊗R/mR S/mS [y 1 ]
where y 1 is the class of y1 . We have that
(y 1 )e − [δ]−1 [x1 ] = 0
where [δ], [x1 ] are the classes of δ and x1 . All other relations on y 1 are divisible by this
relation by (29).
Theorem 4.3. Let assumptions be as in Theorem 3.3, and let R1 → S1 be as in the
conclusions of Theorem 3.3. Then we have a natural isomorphism of graded rings
grν ∗ (S1 ) ∼
= grν (R1 ) ⊗R1 /mR1 S1 /mS1 [Z]/(Z e − [γ1 ]−1 [u1 ]),
where [γ1 ], [u1 ] are the respective classes in grν (R1 ) ⊗R1 /mR1 S1 /mS1 . The degree of the
extension of quotient fields of grν (R1 ) → grν ∗ (S1 ) is e(ν ∗ /ν)f (ν ∗ /ν).
Proof. The proof is exactly the same as the proof of Theorem 4.2.
We also obtain from Theorem 3.7 the following result, showing that even in positive
and mixed characteristic, the associated graded rings of Abhyankar valuations dominating
a stable extension of two dimensional excellent regular local rings have a nice form.
Theorem 4.4. Let assumptions be as in Theorem 3.7, and let R1 → S1 be as in the
conclusions of Theorem 3.7. Then we have a natural isomorphism of graded rings
a b
−1
c d
−1
gr
(R
)
⊗
S
/m
grν ∗ (S1 ) ∼
=
1
S1 [X, Y ]/(X Y − [γ1 ] [u1 ], X Y − [τ1 ] [v1 ]),
R1 /mR1 1
ν
where [γ1 ], [τ1 ], [u1 ], [v1 ] are the respective classes in grν (R1 ) ⊗R1 /mR1 S1 /mS1 . The degree
of the extension of quotient fields of grν (R1 ) → grν ∗ (S1 ) is e(ν ∗ /ν)f (ν ∗ /ν).
20
5. Non constancy of αn and βn
In Corollary 7.30 and Theorem 7.33 of [14], it is shown that αn + βn is a constant for
n ≫ 0, where αn and βn are the integers defined in (6) which are associated to the stable
forms (5) of an extension of valued two dimensional algebraic function fields. If Γν is not
p-divisible, it is further shown that αn and βn are both constant for n ≫ 0, and pβn is the
defect of the extension. However, if Γν is p-divisible, then it is only shown that the sum
αn + βn is constant for n ≫ 0, and that pαn +βn is the defect of the extension. In Remark
7.34 [14] it is asked if αn and βn (and some other numbers computed from the stable
forms) are eventually constant in the case when Γν is p-divisible. We give examples (31) (34) here showing that this is not the case, even within defect Artin Schreier extensions.
The example in Theorem 7.38 [14] is a tower K ∗ /K of two Artin Schreier extensions,
K → K1 → K ∗ ,
where K = k(u, v), K1 = k(x, v) and K ∗ = k(x, y), over an algebraically closed field k of
characteristic p > 0, with
xp
, v = y p − xc y
u=
1 − xp−1
where p − 1 divides c. R and S are defined to be R = k[u, v](u,v) and S = k[x, y](x,y) .
The valuation ν ∗ on K ∗ is defined by the generating sequence (41) in S and the valuation
ν = ν ∗ |K is defined by the generating sequence (46) in R. In Theorem 7.38 [14], it is shown
that Γν ∗ = Γν = p1∞ Z, δ(ν ∗ /ν) = 2, and in the stable forms Rn → Sn above R → S, we
have αn = 1 and βn = 1 for all n.
Let ν1 = ν ∗ |K1 . We define A = k[x, v](x,v) , a local ring of K1 . We will show in this
section that δ(ν ∗ /ν1 ) = δ(ν1 /ν) = 1 and we have stable forms Rj → Aj of K1 /K (Theorem
5.7) and Aj → Sj of K ∗ /K1 (Theorem 5.6) such that by the conclusions of Theorem 5.6,
1 if j is even
(31)
αj (Sj /Aj ) =
0 if j is odd
and
0
1
if j is even
if j is odd
for Aj → Sj , and by Theorem 5.7, we have
0
(33)
αj (Aj /Rj ) =
1
if j is even
if j is odd
(32)
βj (Sj /Aj ) =
and
(34)
for Rj → Aj .
βj (Aj /Rj ) =
1
0
if j is even
if j is odd
We now prove these statements. We make use of the notation introduced in Section 7.11
[14] in the construction of the example of Theorem 7.38 [14]. The proof makes essential
use of the theory of generating sequences of a valuation dominating a two dimensional
regular local ring, as developed in [34] and extended in [15].
We define a k-valuation ν1 of K1 by prescribing a generating sequence in A, starting
with
(35)
U0 = x, U1 = v, U2 = v p − x
21
and for j ≥ 2,
(36)
Uj+1 = Ujp − xp
Uj+1 =
3
Ujp
−
2j−2
Uj−1 if j is odd
2j−1
p
Uj−1 if j is even.
x
We now establish that {Ui } determines a unique valuation ν1 on K1 such that ν1 (x) = 1.
Define γ j by
1
γ 0 = 1, γ 1 =
p
and for j ≥ 2,
(
1 2j−2
if j is odd
+ γ j−1 )
p (p
γj =
(37)
1
2j−1
+ γ j−1 ) if j is even
p3 (p
By induction on j ≥ 0, we have that
( 2j−2 Pj
p
( j ′ =0
P
(38)
γ j+1 =
2j−1
p
( jj ′ =0
Let
ni =
p
p3
1
)
p4j ′
1
)
p4j ′
if j is odd
if j is even
if j is odd
if j is even
Let Γj be the group generated by γ 0 , . . . , γ j . By Remark 7.171 [14] or Theorem 1.1 [15]
and its proof, {Ui } is a generating sequence of a unique valuation ν1 on K1 such that
ν1 (Ui ) = γ i for all i if
(39)
ni = [Γi : Γi−1 ]
for i ≥ 1 and
γ i+1 > ni γ i
(40)
for i ≥ 1.
By (37), ni γ i ∈ Γi−1 , so [Γi : Γi−1 ] = ni if γ i has order precisely ni in Γi /Γi−1 .
By (38), we have that
(
1
Z if i is odd
p2i−2
Γi−1 =
1
Z if i is even
p2i−3
for i ≥ 1, so that by (38), γi has order ni in Γi /Γi−1 . Since (40) holds by (38), we have
that {Ui } is a generating sequence in A which determines a valuation ν1 of K1 .
The following Propositions 5.1 and 5.2 are a little stronger than Proposition 7.40 [14].
Proposition 5.1. Let
S1 = S → S2 → · · ·
be the sequence of quadratic transforms such that Si = Sri′ in the notation of Definition
7.11 [14] so that Vν ∗ = ∪Si . Let Qi be the generating sequence of S of (71) of [14]
determining ν ∗ (Proposition 7.40 [14])
(41)
Q0
Q1
Q2
Qj+1
=
=
=
=
x
y
2
yp − x
2
2j−2
Qj−1 for j ≥ 2.
Qpj = xp
22
Then there exist generating sequences {Q(k)i } in Sk such that
Q2
Q2
=
xS2 = Q(2)0 = Q1 = y, yS2 = Q(2)1 =
Q0
x
are regular parameters in S2 such that Q(2)0 = 0 is a local equation of the exceptional
locus of Spec(S2 ) → Spec(S) and for k ≥ 2,
Qk+1
xSk+1 = Q(k + 1)0 = Q(k)1 , ySk+1 = Q(k + 1)1 = p2k−2
x
Qk−1
are regular parameters in Sk+1 , such that Q(k+1)0 = 0 is a local equation of the exceptional
locus of Spec(Sk+1 ) → Spec(S).
In S2 , the generating sequence {Q(2)i } is defined by
Qi+1
(42)
Q(2)0 = Q1 , Q(2)i = 2(i−1)
xp
for i ≥ 1, and for k ≥ 3, the generating sequence {Q(k)j } in Sk is defined by
(43)
Q(k)0 = Q(k − 1)1 =
Qk−1
2(k−3)
p
x
Q
k−3
and
(44)
Q(k)i =
Qi+k−1
x
p2(i+k−3)
2(i−1)
p
Qk−2
for i ≥ 1.
It follows from Proposition 5.1 that ν ∗ (Q(k)0 ) = p2 ν ∗ (Q(k)1 ), so we have regular parameters xSk+1 , y Sk+1 in Sk+1 defined by
(45)
2
Q(k)0 = xpSk+1 (y Sk+1 + 1), Q(k)1 = xSk+1 .
Proposition 5.2. Let
R1 = R → R2 → · · ·
be the sequence of quadratic transforms such that Ri = Rri′ in the notation of Definition 7.11 [14] so that Vν = ∪Ri . Let Pi be the generating sequence of R of (76) of [14]
determining ν (Corollary 7.41 [14])
(46)
P0
P1
P2
Pi+1
=
=
=
=
u
v
2
vp − u
2
2i−2
Pi−1 for i ≥ 2.
Pip − up
Then there exist generating sequences {P (k)i } in Rk determining ν such that
P2
P2
uR2 = P (2)0 = P1 = v, vR2 = P (2)1 =
=
P0
u
are regular parameters in R2 such that uR2 = 0 is a local equation of the exceptional locus
of Spec(S2 ) → Spec(S) and for k ≥ 2,
Pk+1
uRk+1 = P (k + 1)0 = P (k)1 , vRk+1 = P (k + 1)1 = p2k−2
u
Pk−1
are regular parameters in Rk+1 , such that P (k+1)0 = 0 is a local equation of the exceptional
locus of Spec(Rk+1 ) → Spec(R).
23
In R2 , the generating sequence {P (2)j } is defined by
Pi+1
2(i−1)
up
for i ≥ 1, and for k ≥ 3, the generating sequence {P (k)j } in Rk is defined by
(47)
P (2)0 = P1 , P (2)i =
(48)
P (k)0 = P (k − 1)1 =
Pk−1
2(k−3)
p
u
Pk−3
and
(49)
P (k)i =
Pi+k−1
u
p2(i+k−3)
2(i−1)
p
Pk−2
for i ≥ 1.
It follows from Proposition 5.2 that we have regular parameters uRk+1 , v Rk+1 in Rk+1
defined by
2
P (k)0 = upRk+1 (v Rk+1 + 1), P (k)1 = uRk+1 .
(50)
Proposition 5.3. The sequence {Uj } is a generating sequence in A of a unique k-valuation
ν1 of K1 such that ν1 (x) = 1. Let
A1 = A → A2 → · · ·
be the sequence where Ai = Aa′i in the notation of Definition 7.11 [14], so that Vν1 = ∪Ai .
Then there exist generating sequences {U (k)i } in Ak determining ν1 such that
xA2 = U (2)0 = U1 = v, vA2 = U (2)1 =
U2
U2
=
U0
x
are regular parameters in A2 such that U (2)0 = 0 is a local equation of the exceptional
locus of Spec(A2 ) → Spec(A). For j ≥ 2,
Uj+1
p2j−2
if j is odd
x
Uj−1
(51)
xAj+1 = U (j + 1)0 = U (j)1 , vAj+1 = U (j + 1)1 =
Uj+1
p2j−1
if j is even
x
Uj−1
are regular parameters in Aj+1 , such that U (j+1)0 = 0 is a local equation of the exceptional
locus of Spec(Aj+1 ) → Spec(A).
In A2 , the generating sequence {U (2)j } is defined by
(52)
U (2)0 = U1
and for j ≥ 1,
(53)
U (2)j =
(
Uj+1
2j−2
xp
Uj+1
2j−1
xp
if j is odd
if j is even
and for k ≥ 3, the generating sequence {U (k)j } in Ak is defined by
Uk−1
p2k−6
if k is odd
x
Uk−3
(54)
U (k)0 = U (k − 1)1 =
Uk−1
p2k−5
if k is even
x
and for k odd, j ≥ 1,
24
Uk−3
(55)
U (k)j =
and for k even, j ≥ 1,
(56)
U (k)j =
Uj+k−1
2(j+k)−5 p2j−2
Uk−2
Uj+k−1
2(j+k)−6 p2j−3
Uk−2
xp
if j is odd
Uj+k−1
2(j+k)−6 p2j−2
Uk−2
Uj+k−1
2(j+k)−5 p2j−1
Uk−2
xp
if j is odd
xp
if j is even
xp
if j is even
Further, there exist units δ(k)j in Ak with δ(k)j ≡ 1 mod mAk such that if k is odd,
then
U (k)2 = U (k)p1 − δ(k)0 U (k)0
(57)
and for j ≥ 2,
(58)
U (k)j+1 =
(
U (k)pj − δ(k)j+1 U (k)p0
3
U (k)pj
−
2j−2
U (k)j−1
p2j−1
δ(k)j+1 U (k)0
U (k)j−1
if j is odd
if j is even
and if k is even, then
3
U (k)2 = U (k)p1 − δ(k)0 U (k)0
(59)
and for j ≥ 2,
(60)
U (k)j+1 =
(
3
U (k)pj − δ(k)j+1 U (k)p0
U (k)pj
−
2j−2
U (k)j−1
p2j−3
δ(k)j+1 U (k)0
U (k)j−1
if j is odd
if j is even
Proof. The fact that the sequence {Uj } is a generating sequence of a unique k-valuation
ν1 of K1 was shown before Proposition 5.1.
The remainder of the proposition is proved by induction on k. By (35) and (36) if
k = 1 and (57) - (60) if k > 1 and by Theorem 7.1 [15], there exists a generating sequence
{Ũ (k + 1)} in Ak+1 defined if k is odd by
(61)
Ũ (k + 1)0 = U (k)1 , Ũ (k + 1)1 =
and for j ≥ 2,
(62)
and if k is even, then
(63)
Ũ (k + 1)j =
U (k)j+1
2j−1
U (k)1p
U (k)j+1
U (k)p1
2j
if j is odd
if j is even
Ũ (k + 1)0 = U (k)1 , Ũ (k + 1)1 =
25
U (k)2
U (k)p1
U (k)2
U (k)p1
3
and for j ≥ 2,
(64)
Ũ (k + 1)j =
U (k)j+1
if j is odd
2j+1
U (k)1p
U (k)j+1
U (k)p1
2j
if j is even
To prove that the conclusions of the proposition hold for {U (k + 1)i }, we use the
induction assumption and the appropriate equations (66), (68) or (69) which are stated
below to first verify that there are units λ(k + 1)j in Ak+1 with λ(k + 1)j ≡ 1 mod mAk+1
such that
(65)
U (k + 1)j = λ(k + 1)j Ũ (k + 1)j
for all j, verifying that {U (k + 1)j } is a generating sequence in Ak+1 for ν1 , and then that
the appropriate equations (57) - (60) hold and finally that the appropriate equations (52)
and (53), or (54) - (56) hold.
We now state and prove the equations (66), (68) and (69).
There exists a unit ε(2) in A2 with ε(2) ≡ 1 mod mA2 such that
U1p = ε(2)x.
(66)
Equation (66) follows from (35) and the fact (Theorem 7.1 [15]) that A2 has regular
parameters xA2 and v A2 defined by
(67)
U (1)0 = x = xpA2 (v A2 + 1), U (1)1 = v = xA2
There exists a unit ε(k + 1) ∈ Ak+1 with ε(k + 1) ≡ 1 mod mAk+1 such that
U (k)p1 = ε(k + 1)
(68)
Uk−1
p
Uk−2
if k ≥ 2 is odd and
3
U (k)p1 = ε(k + 1)
(69)
Uk−1
3
p
Uk−2
if k ≥ 2 is even.
We now simultaneously verify the equations (68) and (69). We first verify that Ak+1
has regular parameters xAk+1 and v Ak+1 defined if k is odd by
(70)
U (k)0 = xpAk+1 (v Ak+1 + 1), U (k)1 = xAk+1
and if k is even by
(71)
3
U (k)0 = xpAk+1 (v Ak+1 + 1), U (k)1 = xAk+1 .
If k is odd, we have that pν1 (U (k)1 ) = ν1 (U (k)0 ) by (57) and if k is even we have that
p3 ν1 (U (k)1 ) = ν1 (U (k)0 ) by (59), so (70) and (71) follow from (57), (59) and Theorem
7.1 [15]. Thus for j ≤ k, Uj is a power of U (k + 1)0 times a unit in Ak+1 which is
equivalent to 1 mod mAk+1 . Thus by (36), there exists a unit ε(k + 1) in Ak+1 with
ε(k + 1) ≡ 1 mod mAk+1 such that
Ukp = ε(k + 1)xp
and
3
Ukp = ε(k + 1)xp
2k−2
2k−1
26
Uk−1 if k is odd
Uk−1 if k is even.
Thus if k is odd,
U (k)p1
=
Uk
xp2k−3 Uk−2
2k−2
p
ε(k+1)xp
Uk−1
p
xp2k−2 Uk−2
=
Ukp
2k−2 p
p
x
Uk−2
=
U
= ε(k + 1) Uk−1
.
p
k−2
if k is even,
3
U (k)p1
=
Uk
xp2k−4 Uk−2
ε(k+1)xp
=
x
2k−1
p2k−1
p 3
Ukp
=
Uk−1
3
3
p
xp2k−1 Uk−2
= ε(k + 1)
3
p
Uk−2
Uk−1
3
p
Uk−2
.
Lemma 5.4. We have that
U0 = x = Q0 , U1 = Qp1 − xc y,
(72)
and for j ≥ 1, if j is odd, then
(73)
Uj+1 = Qj+1 + x
P j−1
p2j−2 ( j ′2=0
1
′
p4j
)
fj+1 (x, y)
where fj+1 ∈ k[[x]][y], x divides fj+1 in k[[x]][y], and degy fj+1 = p2j−1 .
If j is even, then
(74)
Uj+1 =
Qpj+1
+x
P j −1
p2j−1 ( j2′ =0
1
′
p4j
)
fj+1 (x, y)
where fj+1 ∈ k[[x]][y], x divides fj+1 in k[[x]][y], and degy fj+1 = p2j .
Proof. We have that
U1 = v = y p − xc y = Qp1 − xc y,
U2 = v p − x = Q2 − xcp y p ,
verifying (73) for j = 1 and
3
3
4
U3 = U2p − xp U1 = Qp3 − xcp y p + xp
3 +c
y
verifying (74) for j = 2.
We prove the equations (73) and (74) for j ≥ 3 by induction. First assume that j is
odd. We have that
Uj+1 = Ujp − xp
2j−2
= Qj+1 + x
Uj−1
P j−3
p2j−2 ( j ′2=0
1
′
p4j
) p
fj
−x
P j−3
p2j−2 +p2j−6 ( j ′2=0
The formula (73) then follows since
j−3
j−1
2
2
X
X
1
1
2j−2
)
=
p
(
p2j−2 (
′
′) + 1
4j
4j
p
p
′
′
j =0
j =0
and
j−1
j−3
p
2j−2
+p
2j−6
2
2
X
X
1
1
2j−2
)
=
p
(
).
(
4j ′
4j ′
p
p
′
′
j =0
j =0
27
1
′
p4j
)
fj−1 .
Now assume that j ≥ 3 is even. We have that
3
Uj+1 = Ujp − xp
=
Qpj+1
2j−1
+x
Uj−1
P 2j −1
1
j ′ =0 p4j ′
p2j−1 (
) p3
fj
−x
P j −2
p2j−1 +p2j−5 ( j2′ =0
1
′
p4j
)
fj−1 .
The formula (74) then follows since
p
2j−1
+p
2j−5
j
j
j =0
j =0
−2
−1
2
2
X
X
1
1
2j−1
(
)=p
(
).
4j ′
4j ′
p
p
′
′
Proposition 5.5. We have Vν1 = Vν ∗ ∩ K1 and ν ∗ |K1 = ν1 . Further, ν ∗ is the unique
extension of ν1 to K ∗ .
Proof. Since ν1 (x) = ν ∗ (x) = 1, it suffices to show that Vν1 ⊂ Vν ∗ . From (51) and the fact
that Vν1 = ∪Aj , we need only show that vAj+1 ∈ Vν ∗ for all j. From equation (72) of [14],
we obtain that
j−1
(75)
∗
ν (Qj+1 ) = β j+1 < p
2j−2
2
X
1
)+1
(
4j ′
p
′
j =0
if j is odd, and
j
(76)
ν
∗
(Qpj+1 )
= pβ j+1 < p
2j−1
−1
2
X
1
)+1
(
4j ′
p
′
j =0
if j is even. Thus by Lemma 5.4,
(77)
∗
ν (Uj+1 ) =
ν ∗ (Qj+1 )
pν ∗ (Qj+1 )
Thus by (51) and Proposition 7.40 [14],
∗
ν (ySj+1 )
∗
ν (vAj+1 ) =
pν ∗ (ySj+1 )
if j is odd
if j is even
if j odd
if j even
The fact that ν ∗ is the unique extension of ν to K ∗ follows since ν1 |K = ν ∗ |K = ν by
Proposition 7.40 [14], and ν ∗ is the unique extension of ν to K ∗ by (1) of Theorem 7.38
[14].
Theorem 5.6. For all j, Sj is a finite extension of Aj . If j is odd, we have expressions
(78)
xAj+1 = τj+1 xpSj+1 , vAj+1 = γj+1 ysj+1 + xSj+1 Ωj+1
where τj+1 , γj+1 are units in Sj+1 and Ωj+1 ∈ Sj+1 .
If j is even, we have expressions
(79)
xAj+1 = τj+1 xSj+1 , vAj+1 = γj+1 ySp j+1 + xSj+1 Ωj+1
where τj+1 , γj+1 are units in Sj+1 and Ωj+1 ∈ Sj+1 .
28
Proof. For i ≤ j, Qi are units in Sj+1 times a power of xSj+1 . By (73), (74) and (75) and
(76), we then have that Ui are units in Sj+1 times a power of xSj+1 and for i ≤ j,
αQpi where α ∈ Sj+1 is a unit if i is odd
(80)
Ui =
αQi where α ∈ Sj+1 is a unit if i is even.
We have from (75) that
(81)
j−1
ν ∗ (Qj−1 ) = β j−1
j−1
−1
j−2
2
2
X
X
X
1
1
1
2j−6
2j−2
2j−2
2j−6
)
<
p
(
)
+
1
=
p
(
=p
(
′
′
′) +1 − p
4j
4j
4j
p
p
p
′
′
′
j =0
j =0
j =0
if j is odd and by (76),
(82)
pν ∗ (Qj−1 ) = pβ j−1
j
j
j =0
j =0
−2
−1
j−2
2
2
X
X
X
1
1
1
2j−5
2j−1
2j−1
2j−5
(
(
=p
(
′) < p
′)+ 1 = p
′)+ 1− p
4j
4j
4j
p
p
p
′
′
′
j =0
if j is even.
If j is odd, by equations (51), (73), (80) and (81) and Proposition 5.1, we have an
expression
vAj+1
P j−1
p2j−2 ( j ′2=0
1
′
)
p4j f
Qj+1 + x
Uj+1
j+1
=
= γySj+1 + xSj+1 Ω
= p2j−2
2j−2
p
x
Uj−1
αx
Qj−1
for some unit γ ∈ Sj+1 and Ω ∈ Sj+1 .
If j is even, by equations (51), (74), (80) and (82) and Proposition 5.1, we have an
expression
vAj+1
Uj+1
=
= p2j−1
x
Uj−1
Qpj+1
+x
P j −1
p2j−1 ( j2′ =0
αx
p2j−1
1
′
p4j
)
fj+1
Qpj−1
= γySp j+1 + xSj+1 Ω
for some unit γ ∈ Sj+1 and Ω ∈ Sj+1 .
By (80), we have that
xAj+1
αQpj
Uj
= p2j−3
= p2j−3 p = αxpSj+1
x
Uj−2
x
Qj−2
for some unit α ∈ Sj+1 if j is odd and
xAj+1 =
Uj
2j−4
p
x
U
=
j−2
αQj
2j−4
p
x
Q
j−2
= αxSj+1
for some unit α ∈ Sj+1 if j is even.
The extension Aj → Sj is finite for all j since each Aj → Sj is quasi finite and ν ∗ is the
unique extension of ν to K ∗ .
Theorem 5.7. For all j, Aj is a finite extension of Rj . If j is odd, we have expressions
(83)
p
uRj+1 = σj+1 xAj+1 , vRj+1 = λj+1 vA
+ xAj+1 Λj+1
j+1
where σj+1 , λj+1 are units in Aj+1 and Λj+1 ∈ Aj+1 .
29
If j is even, we have expressions
(84)
uRj+1 = σj+1 xpAj+1 , vRj+1 = λj+1 vAj+1 + xAj+1 Λj+1
where σj+1 , λj+1 are units in Aj+1 and Λj+1 ∈ Aj+1 .
Proof. We have that Rj+1 ⊂ Aj+1 and Rj+1 → Aj+1 is finite since Rj+1 → Sj+1 and
Aj+1 → Sj+1 are finite, Aj+1 is normal and QF(Rj+1 ) ⊂ QF(Aj+1 ). The expressions (83)
and (84) follow from (2) of Theorem 7.38 [14] and Theorem 5.6.
References
[1] S. Abhyankar, On the valuations centered in a local domain, Amer. J. Math. 78 (1956), 321
- 348.
[2] S. Abhyankar, Local uniformization of algebraic surfaces over ground fields of characteristic
p 6= 0, Annals of Math. 63 (1956), 491 -526.
[3] S. Abhyankar, Ramification theoretic methods in algebraic geometry, Princeton Univ Press,
1959.
[4] S. Abhyankar, Resolution of singularities of embedded algebraic surfaces, second edition,
Springer Verlag, New York, Berlin, Heidelberg, 1998.
[5] A. Benito, O. Villamayor U., Techniques for the study of singularities with application to
resolution of 2-dim schemes, Math. Ann. 353 (2012), 1937 - 1068.
[6] A. Bravo and O. Villamayor U., Singularities in positive characteristic, Stratification and
simplification of the singular locus, Advances in Math. 224, (2010), 1349 - 1418.
[7] V. Cossart, U. Jannsen and S. Saito, Canonical embedded and non-embedded resolution of
singularities for excellent two-dimensional schemes, arXiv:0905.2191
[8] V. Cossart and O. Piltant, Resolution of singularities of threefolds in positive characteristic
I, Reduction to local uniformization on Artin-Schreier and purely inseparable coverings, J.
Algebra 320 (2008), 1051 - 1082.
[9] V. Cossart, and O. Piltant, Resolution of singularities of threefolds in positive characteristic
II, J. Algebra 321 (2009), 1836 - 1976.
[10] S.D. Cutkosky, Local factorization and monomialization of morphisms, Astérisque 260, 1999.
[11] S.D. Cutkosky, Resolution of Singularities for 3-folds in positive characteristic, Amer. J.
Math. 131 (2009), 59 - 127.
[12] S.D. Cutkosky, Counterexamples to local monomialization in positive characteristic, Math.
Ann. 362 (2015), 321 - 334..
[13] S.D. Cutkosky and L. Ghezzi, Completions of valuation rings, Contemp. math. 386 (2005),
13 - 34.
[14] S.D. Cutkosky and O. Piltant, Ramification of Valuations, Advances in Math. 183 (2004),
1-79.
[15] S.D. Cutkosky and Pham An Vinh, Valuation semigroups of two dimensional local rings,
Proceedings of the London Mathematical Society 108 (2014), 350 - 384.
[16] S.D. Cutkosky and Pham An Vinh, Ramification of local rings along valuations, to appear
in Journal of pure and applied algebra.
[17] S.D. Cutkosky and B. Teissier, Semigroups of valuations on local rings, Mich. Math. J. 57
(2008), 173 - 193.
[18] A.J. de Jong, Smoothness, semi-stablility and alterations, Inst. Hautes Etudes Sci. Publ.
math. 83 (1996) 51 -93.
[19] L. Ghezzi, Huy Tài Hà and O. Kashcheyeva, Toroidalization of generating sequences in
dimension two function fields, J. Algebra 301 (2006) 838-866.
[20] L. Ghezzi and O. Kashcheyeva, Toroidalization of generating sequences in dimension two
function fields of positive characteristic, J. Pure Appl. Algebra 209 (2007), 631-649.
[21] A. Grothendieck, and A. Dieudonné, Eléments de géométrie algébrique IV, vol. 2, Publ.
Math. IHES 24 (1965).
[22] H. Hauser, On the problem of resolution of singularities in positive characteristic (or: a proof
we are waiting for) Bull. Amer. Math. Soc. 47 (2010), 1-30.
[23] H. Hironaka, Three key theorems on infinitely near singularities, Singularités FrancoJaponaises, 87 - 126, Sémin. Congr. 10 Soc. Math. France, Paris 2005.
30
[24] H. Knaf and F.-V. Kuhlmann, Every place admits local uniformization in a finite extension
of the function field, Adv. Math. 221 (2009), 428 - 453.
[25] H. Knaf and F-V. Kuhlmann, Abhyankar places admit local uniformization in any characteristic, Ann. Sci. École Norm. Sup. 38 (2005), 833 - 846.
[26] F.-V. Kuhlmann, Valuation theoretic and model theoretic aspects of local uniformization, in
Resolution of Singularities - A Research Textbook in Tribute to Oscar Zariski, H. Hauser, J.
Lipman, F. Oort, A. Quiros (es.), Progress in Math. 181, Birkhäuser (2000), 4559 - 4600.
[27] F.-V. Kuhlmann, Value groups, residue fields, and bad places of algebraic function fields,
Trans. Amer. Math. Soc. 356 (2004), 363 - 395.
[28] F.-V. Kuhlmann, A classification of Artin Schreier defect extensions and a characterization
of defectless fields, Illinois J. Math. 54 (2010), 397 - 448.
[29] J. Lipman, Desingularization of 2-dimensional schemes, Annals of Math. 107 (1978), 115 –
207.
[30] S. MacLane, A construction for absolute values in polynomial rings, Trans. Amer. Math.
Soc. 40 (1936), 363 - 395.
[31] S. MacLane and O. Schilling, Zero-dimensional branches of rank 1 on algebraic varieties,
Annals of Math. 40 (1939), 507 - 520.
[32] M. Moghaddam, A construction for a class of valuations of the field K(X1 , . . . , Xd , Y ) with
large value group, J. Algebra 319, 7 (2008), 2803-2829.
[33] J.P. Serre, Corps Locaux, Hermann, 1962.
[34] M. Spivakovsky, Valuations in function fields of surfaces, Amer. J. Math. 112 (1990), 107 156.
[35] B. Teissier, Valuations, deformations and toric geometry, Valuation theory and its applications II, F.V. Kuhlmann, S. Kuhlmann and M. Marshall, editors, Fields Institute Communications 33 (2003), Amer. Math. Soc., Providence, RI, 361 – 459.
[36] B. Teissier, Overweight deformations of affine toric varieties and local uniformization,
preprint.
[37] M. Temkin, Inseparable local uniformization, J. Algebra 373 (2013), 65 - 119.
[38] M. Vaquié, Famille admissible de valuations et défaut d’une extension, J. algebra 311 (2007),
859 - 876.
[39] O. Zariski and P. Samuel, Commutative Algebra Volume I, Van Nostrand, 1958.
[40] O. Zariski and P. Samuel, Commutative Algebra Volume II, Van Nostrand, 1960.
Steven Dale Cutkosky, Department of Mathematics, University of Missouri, Columbia,
MO 65211, USA
E-mail address: [email protected]
31
| 0math.AC
|
1
Channel Hardening and Favorable Propagation in
Cell-Free Massive MIMO with Stochastic Geometry
arXiv:1710.00395v1 [cs.IT] 1 Oct 2017
Zheng Chen and Emil Björnson
Abstract—Cell-Free (CF) Massive MIMO is an alternative
topology for Massive MIMO networks, where a large number
of single-antenna access points (APs) are distributed over the
coverage area. There are no cells but all users are jointly served
by the APs using network MIMO methods. In prior work, it
has been claimed that CF Massive MIMO inherits the basic
properties of cellular Massive MIMO, namely channel hardening
and favorable propagation. In this paper, we evaluate if one can
rely on these properties when having a realistic stochastic AP
deployment. Our results show that the level of channel hardening
depends strongly on the propagation environment and there is
generally little hardening, except when the pathloss exponent
is small. We further show that using 5–10 antennas per AP,
instead of one, we can greatly improve the hardening. The level of
favorable propagation is affected by the propagation environment
and by the distance between the users, where only spatially well
separated users exhibit favorable propagation. The conclusion
is that we cannot rely on channel hardening and favorable
propagation when analyzing and designing CF Massive MIMO
networks, but we need to use achievable rate expressions and
resource allocation schemes that work well in the absence of
these properties. Some options are described in this paper.
Index terms— Cell-Free Massive MIMO, channel hardening, favorable propagation, achievable rates, stochastic geometry.
I. I NTRODUCTION
The throughput of conventional cellular networks is limited
by the uncoordinated inter-cell interference. To mitigate this
interference, Shamai and Zaidel introduced the co-processing
concept in 2001 [2], which is more commonly known as
network multiple-input multiple-output (MIMO) [3]. The key
idea is to let all access points (APs) in the network jointly
serve all users, in downlink as well as uplink, thereby turning
interference into useful signals [4]. Despite the great theoretical potential, the 3GPP LTE standardization of this technology
failed to provide any remarkable gains [5].
Two practical issues with network MIMO are to achieve
scalable channel acquisition and sharing of data between APs.
The former can be solved by utilizing only local channel state
information (CSI) at each AP [6], which refers to knowledge
of the channels between the AP and the users. These channels
can be estimated by exploiting uplink pilot transmission and
channel reciprocity in time-division duplex (TDD) systems,
Z. Chen and E. Björnson are with the Department of Electrical Engineering
(ISY), Linköping University, Linköping, Sweden (email: [email protected],
[email protected]). A part of this work will appear in Globecom Workshops 2017 [1].
This work was supported in part by ELLIIT, CENIIT, and the Swedish
Foundation for Strategic Research.
thus making TDD a key enabler for network MIMO. Usercentric clustering, where all APs reasonably close to a user
transmit signals to it, is key to reduce the data sharing
overhead [7]. These concepts are not easily incorporated into
LTE, which relies on codebook-based channel acquisition and
network-centric AP clustering.
The network MIMO concept has recently reappeared under
the name Cell-Free (CF) Massive MIMO [8], [9], which
refers to a network with a massive number of geographically
distributed single-antenna APs that jointly serve a smaller
number of users. In particular, it has been presented as a
better option for providing coverage than using uncoordinated
small cells [10]. The CF concept is fundamentally the same
as in the network MIMO paper [6], where the APs perform
joint transmission with access to data to every user but only
local CSI. The main novelty introduced by CF Massive MIMO
is the capacity analysis that takes practical pilot allocation
and imperfect CSI into account [10], [11], using similar
methodology as in the cellular Massive MIMO literature [12].
Conventional cellular Massive MIMO systems consist of
non-cooperating APs equipped with a massive number of colocated antennas. Such systems deliver high spectral efficiency
by utilizing the channel hardening and favorable propagation phenomena [12]. Channel hardening means that the
beamforming turns the fading multi-antenna channel into an
nearly deterministic scalar channel [13]. Favorable propagation
means that the users’ channel vectors are almost orthogonal
[14]. These are both consequences of the law of large numbers.
CF Massive MIMO is essentially a single-cell Massive
MIMO system with antennas distributed over a wide geographical area, which makes the joint channel from the APs
to a user strongly spatially correlated—some APs are closer
to the user than others. The capacity of Massive MIMO with
spatial correlation and imperfect CSI has been analyzed in
[15]–[17], among others, but with channel models that provide
channel hardening and favorable propagation. It is claimed in
[10] that the outstanding aspect of CF Massive MIMO is that
it can also utilize these phenomena, but this has not been fully
demonstrated so far. Hence, it is not clear if the known Massive
MIMO capacity lower bounds are useful in the CF context or if
they underestimate the achievable performance. For example,
[18] derived a new downlink capacity lower bound for the
case when the precoded channels are estimated using downlink
pilots. That bound provides larger values than the common
bound that estimates the precoded channels by relying on
channel hardening, but it is unclear whether this indicates the
need for downlink pilots and/or the lack of channel hardening
in the CF Massive MIMO setup.
2
This paper aims at answering the following open questions:
• Can we observe channel hardening and favorable propagation in CF Massive MIMO with single-antenna APs?
• Is it more beneficial to deploy more antennas on few
APs or more APs with few antennas, in order to achieve
a reasonable degree of channel hardening and favorable
propagation?
• Are there any other important factors that affect the
conditions of these two properties?
• Which capacity bounds in conventional cellular Massive
MIMO are appropriate to use in CF Massive MIMO?
In order to answer these questions, we model the AP distribution by a homogeneous Poisson Point Process (PPP),
where each AP is equipped with N ≥ 1 of antennas. Unlike
the conventional regular grid model for the base station
deployment, the stochastic point process model considered
in this work can capture the irregular and semi-random AP
deployment in real networks [19], [20]. First, conditioning
on a specific network realization with APs located at fixed
locations and a reference user point at the origin, we define
the channel hardening and favorable propagation criteria as
functions of the AP-user distances. Then, we examine the
spatially averaged percentage/probability of randomly located
users that satisfy these criteria. The separation of the randomness caused by small-scale fading and the spatial locations of
APs allows us to study the statistical performance of largescale CF Massive MIMO network with time-scale separation.
It is similar to the concept of meta distribution proposed in
[21], where the difference mainly lies in the definition of the
studied performance metrics.
Our analysis is carried out by considering different number
of antennas per AP and different non-singular pathloss models: the single-slope model with different pathloss exponents
[22], [23] and the multi-slope model [10]. Compared to the
conference paper [1], which focuses on the channel hardening
aspect of CF Massive MIMO, in this paper, we provide thorough investigation for both channel hardening and favorable
propagation, based on which we give insights into the selection
of achievable rate expressions in CF Massive MIMO.
The remainder of this paper is organized as follows. In
Section II we describe the CF Massive MIMO network model,
including the AP distribution and the channel models. Next,
Section III analyzes the channel hardening and Section IV
analyzes the favorable propagation in CF Massive MIMO.
Section V considers different capacity lower bounds from
cellular Massive MIMO and demonstrates which ones are
useful in CF systems. Section VI concludes this paper.
II. S YSTEM M ODEL
We consider a CF Massive MIMO system in a finitesized network region A. The APs are distributed on the twodimensional Euclidean plane according to a homogeneous PPP
Φ A with intensity λ A, measured by per m2 . [24]. Each AP is
equipped with N ≥ 1 antennas, which is a generalization of the
N = 1 CF Massive MIMO considered in prior works [8]–[11].
All the APs are connected to a central processing unit (CPU)
through backhaul, and the CPU codes and decodes the data
CPU
UE k
gm,k
AP m
Fig. 1. Cell-Free Massive System. Here, each AP can be equipped with either
single or multiple antennas.
signals; see Fig. 1 for an illustration. Different to a smallcell network, all the APs are coordinated to serve all users
simultaneously using the same frequency-time resources. The
number of users and their locations are generated by another
independent point process. Denote by L the number of APs
in a specific realization of the PPP Φ A, we have that L is a
Poisson random variable (RV) with mean value
E[L] = λ A S(A),
(1)
where S(A) denotes the area of the network region A. Let M
denote the total number of antennas existing in A, then we
have M = LN and E[M] = Nλ A S(A).
Given the user distribution, we assume that there are K
users in a specific network realization, where K M. When
the AP density is much larger than the user density, the
boundary effect caused by the finite-size network region is
weak, i.e., users located at the network boundary are still
likely to have nearby dominant APs that makes their received
signal distribution similar to network-center users. Consider
a typical user at the origin, the spatially averaged network
statistics seen at this typical user can represent the average
network performance seen by randomly located users in the
network. Denote by gk the M × 1 channel vector between all
the antennas and the typical user (labeled as user k), the m-th
element gm,k is modeled by
gm,k =
p
l(dm,k )hm,k ,
(2)
where hm,k represents the small-scale fading and l(dm,k )
represents the distance-dependent pathloss and it is a function
of the distance dm,k between the m-th antenna and the user
k. Since every N antennas are co-located at the same AP, we
have d(i−1)·N +1,k = d(i−1)·N +2,k = . . . = di ·N,k , for i = 1, . . . , L.
We assume independent Rayleigh fading from each antenna
to the typical user, which means that {hm,k } are independently
and identically distributed (i.i.d.) CN (0, 1) RVs. In the first
part of this paper, we consider a non-singular pathloss model
l(r) = min(1, r −α ), where r is the antenna-user distance and
3
α > 1 is the pathloss exponent.1 A three-slope pathloss model
will also be studied in Section III-D.
Note that we do not include shadow fading in our analysis.
With the commonly used log-normal shadowing model, the
random shadowing coefficients from randomly located APs
do not have any fundamental impact on the channel gain
distribution. Therefore, the inclusion of shadowing coefficients
would not change the general trends observed in this paper.
A. Main Advantage of CF Massive MIMO
When α > 2, kgk k 2 is guaranteed to have finite mean even
with infinite network size, i.e., ρ → ∞. When 1 < α < 2,
kgk k 2 increases unboundedly when the network size grows.
It is particularly interesting to study the case when the
antenna density µ = Nλ A is fixed. We then observe that
the mean channel gain in (5) is the same, irrespective of
whether there is a high density of single-antenna APs or
a smaller density of multi-antenna APs. The variance is,
however, proportional to (N 2 + N)λ A = (N + 1)µ and thus
grows with N.
Similar to other distributed antenna systems, the main
advantage of CF Massive MIMO is the macro-diversity; that is,
reduced distance between a user and its nearest APs. This can
be demonstrated by analyzing the distribution of the squared
norm of the channel vector,
(3)
i ∈Φ A
|hm,k
m=(i−1)·N +1
. . . = di ·N,k
0.5
0.4
which we refer to as the channel gain. Here, M depends on
each realization of the PPP Φ A.
Each AP is equipped with N antennas and therefore the
sum of the small-scale fading coefficients of its N co-located
antennas is a RV following a Gamma(N, 1) distribution, which
has mean N and variance N. We define the distance vector
r = [r1, . . . , rL ]T , where each element ri denotes the distance
from the i-th AP to the typical user at the origin. Thus, the
squared norm in (3) can be written as
Õ
kgk k 2 =
Hi l(ri ),
(4)
where Hi =
0.7
0.6
|hm,k | 2 l(dm,k ),
m=1
iÍ
·N
0.8
CDF
kgk k 2 =
M
Õ
1
0.9
|2
∼ Gamma(N, 1) and ri =
d(i−1)·N +1,k =
for i = 1, . . . , L.
Note that there are two sources of randomness in (4):
{Hi } and Φ A. When studying the channel distribution for a
randomly located user, it is natural to consider the distribution
of kgk k 2 with respect to both sources of randomness. From
prior studies on the application of stochastic geometry in
wireless networks, it is well known that the sum of the received
power from randomly distributed nodes is described by a shot
noise process. The mean and variance of kgk k 2 averaging
over the spatial distribution of the antennas are known for
the unbounded and bounded pathloss models [23]. For our
considered pathloss model l(r) = min(1, r −α ), in a finite
network region with radius ρ centered around the typical user,
we have
(
2(1−ρ2−α )
if α , 2
Nλ A π 1 + α−2
2
(5)
E kgk k =
Nλ A π (1 + 2 ln(ρ))
if α = 2
1 − ρ2−2α
Var kgk k 2 = (N 2 + N)λ A π 1 +
.
(6)
α−1
Proof: See Appendix A.
1 Note that the unbounded pathloss model l(r) = r −α is not appropriate
when analyzing CF Massive MIMO with stochastic geometry, because the
antennas can then be arbitrarily close to the user, which might result in
unrealistically high power gain when using the unbounded pathloss model.
0.3
N=1
N=10
N=100
0.2
0.1
0
-70
-60
-50
-40
-30
-20
-10
0
10
20
2
||g|| [dB]
Fig. 2. The CDF of the squared norm of the channel vector kg k k 2 to user
k with respect to small-scale fading and PPP realizations. The number of
antennas per AP is N = {1, 10, 100}. The antenna density is fixed µ =
N λ A = 0.001/m2 (103 /km2 ).
Fig. 2 shows the cumulative distribution function (CDF)
of kgk k 2 , with respective to random spatial locations and
small-scale fading realizations. We consider µ = 103 /km2 and
different numbers of antennas per AP: N ∈ {1, 10, 100}. Note
that the horizontal axis is shown in decibel, thus a few users
are very close to an AP and have large values of kgk k 2 while
the majority have substantially smaller values. The long-tailed
exponential distribution of the small-scale fading |hm,k | 2 has
a strong impact on the CDF, but also the AP density makes a
difference. The larger N is, the longer tail the distribution has;
at the 95%-likely point, N = 1 achieves a 12 dB higher value
than N = 100. The reason behind the increasing tail with N is
that Var[kgk k 2 ] is proportional to (N + 1), as described above.
A practical interpretation is that having higher AP density
reduces the average distance between the APs and this macrodiversity reduces the risk that a randomly located user has large
distances to all of its closest APs. This observation evinces a
key motivation of CF Massive MIMO with N = 1 and high
AP density, in terms of providing more uniform coverage to
users at random locations than a conventional cellular Massive
MIMO deployment with N = 100 and a low AP density.
From [23] and [25], it is known that a Gamma distribution
provides a good approximation of the interference distribution
in a Poisson random field with non-singular pathloss. Here, the
expression of kgk k 2 coincides with the definition of interference power in [23] and [25]. Thus, the Gamma distribution can
be used to approximate the distribution of kgk k 2 . The details
are omitted since it is outside the scope of this work.
4
III. M EASURE OF C HANNEL H ARDENING
B. Conditional Channel Distribution at Fixed Location
The previous analysis characterized the channel gain distribution that a user will observe when moving around in a
large network. Once the APs are deployed, for a user at a
fixed location (e.g., located in a room), the small-scale fading
varies over time but the large-scale fading from the APs to the
user remains the same. Conditioning on a specific network
realization of Φ A, assuming that there are L APs in the
network, the distances between the APs and the typical user
are basically fixed. The conditional distribution of the channel
statistics with respect to the small-scale fading distribution
is essential for performance evaluation (e.g., computing the
ergodic capacity) of CF Massive MIMO networks with a fixed
topology and users at fixed but random locations.
With single-antenna APs, i.e., N = 1, the total number of
antennas M is equal to the number of APs L. As a result
of the exponentially distributed small-scale fading coefficient
|hm,k | 2 , the conditional distribution of the channel gain kgk k 2
in (3) follows a Hypoexponential distribution, denoted by
Hypo(l(r1 )−1, . . . , l(rL )−1 ), which is usually a long-tailed distribution when the coefficients l(ri ) are distinct [26].
With N > 1 antennas per AP, the channel gain kgk k 2
is given by (4) where Hi ∼ Gamma(N, 1). As a result, the
conditional distribution of the channel gain from the i-th AP
is Hi l(ri ) ∼ Gamma(N, l(ri )) for i = 1, . . . , L. Due to the sum
of independent Gamma RVs with different scale parameters,
the mean and variance of kgk k 2 conditioning on the distance
vector r are
h
i
E kgk k 2 r = N
h
i
Var kgk k 2 r = N
L
Õ
i=1
L
Õ
l(ri )
(7)
l 2 (ri ).
(8)
i=1
The exact conditional probability density function (PDF) of
kgk k 2 can be computed using the approach in [27] and the
exact expression is available in [28, Eq. (6)]. By looking at
(7) and (8), it is unclear how the channel gain behaves; for
example, if it is the mean value or the channel variations that
grow faster.
In cellular Massive MIMO with M co-located antennas,
conditioning on a specific location of the user, denote by β =
E[kgk k 2 |r]/M the pathloss from the M co-located antennas
to the user. The squared norm of the channel gain kgk k 2 then
follows a Gamma(M,√β)-distribution, with mean value βM and
standard deviation β M. When M increases, the distribution
approaches a normal distribution and is (relatively speaking)
concentrated around the mean since it grows faster than the
standard deviation. The different channel gain distributions in
CF and cellular Massive MIMO highlight the fundamental
difference between the channel statistics of these two types
of networks. In the remainder of this paper, we will proceed
to investigate if in a CF Massive MIMO network, we could
observe the classical Massive MIMO phenomena, namely
channel hardening and favorable propagation.
In cellular Massive MIMO, when the number of antennas
grows, the channel between the AP and the user behaves
as almost deterministic. This property of is called channel
hardening. Conditioning on a specific network realization with
distance vector r = [r1, . . . , rL ]T , channel hardening appears
in CF Massive MIMO when the following condition holds:
kgk k 2
→ 1 as M → ∞,
(9)
E kgk k 2 |r
L
Í
where kgk k 2 =
Hi l(ri ) is the channel gain from the L APs
i=1
to the typical user. One way to prove channel hardening (with
convergence2 in (9) in mean square sense) is to show that the
channel gain variation
"
#
Var kgk k 2 |r
kgk k 2
r =
Var
2 → 0 as M → ∞.
E kgk k 2 |r
E kgk k 2 |r
(10)
For a large wireless network, studying the channel statistics
at a specific location is of limited interest and the results cannot
be generalized to users at other arbitrary locations. To quantify
the channel gain variation for users at arbitrary locations, we
define the following channel hardening measure:
i
h
Var kgk k 2 r
(11)
≤
θ
pθ = P h
.
i 2
E kgk k 2 r
Var[ kg k k 2 |r]
2 given a certain threshold θ.
(E[ kgk k 2 |r])
Here, the probability is obtained over different network realizations that generate different distance vector r. As mentioned
in Section II, the spatially averaged probability pθ provides
the percentage of randomly located users that experience
Var[ kg k k 2 |r]
2 smaller or equal to θ. Notice that pθ = 1 implies
(E[ kgk k 2 |r])
that all users have channel gain variations that are smaller than
θ. The ideal case is p0 = 1 where the variance is zero for all
users. When the threshold θ is small enough, the larger pθ
is, with higher possibility we observe channel hardening for
users at arbitrary locations.
This is the CDF of
A. Necessary Conditions for Channel Hardening
With N ≥ 1 antennas per AP, from (7) and (8), the channel
hardening measure in (11) can be written as
Í
L
N Í L l 2 (r )
2
i
i=1 l (ri )
i=1
pθ = P
≤
θ
≤
θ
=
P
.
2
2
ÍL
Í L
N i=1 l(ri )
N
l(r
)
i
i=1
(12)
Since N appears in the denominator, for a given θ, pθ always
increases with N. This implies that regardless of the AP
density, having more antennas per AP always helps the channel
to harden. In the following, we fix the number of antennas N
per AP and study the impact that the AP density λ A has on
the channel hardening criterion.
2 Note
that convergence in mean square implies convergence in probability.
5
For a given network realization with L APs, by defining
ÍL
ÍL 2
Y1 = i=1
l(ri ), Y2 = i=1
l (ri ), and
Xch =
Y2
=
NY12
ÍL
i=1
N
Í
l 2 (ri )
L
i=1
l(ri )
2 ,
(13)
we can write the channel hardening measure as
1) α > 2: As the network radius ρ → ∞, we have ρ2−2α →
0 and ρ2−α → 0, which implies
α 2
2
α
E Y1
→ λ Aπ
+ λ Aπ
, (19)
α−1
α−2
α
,
(20)
E[Y2 ] → λ A π
α−1
1/N
E[Y2 ]
2 →
.
(21)
NE Y1
1 + λ A π α(α−1)2
(α−2)
pθ = P [Xch ≤ θ] .
(14)
The exact of distribution of Xch is difficult to analyze even with
the joint PDF of ri , i = 1, . . . , L. One objective of this work is
to provide intuitive insights into the relation between channel
hardening and the AP density. Specifically, if pθ should approach 1 when the AP density λ A increases,
Í L we need
Í Xch → 0
when λ A → ∞. Since Y12 = Y2 + i=1
l(ri ) Lj=1, j,i l(r j ),
it follows that Y2 and Y12 are highly correlated. Though the
distributions of Y2 and Y12 are not trivial to obtain, their mean
and variance can be obtained by Campbell’s theorem as in
Section II-A. For the non-singular pathloss model l(r) =
min(1, r −α ), in a network region with radius ρ, we have
(
2(1−ρ2−α )
α−2
λ Aπ 1 +
λ A π(1 + 2 ln(ρ))
E [Y1 ] =
if α , 2
if α = 2
1 − ρ2−2α
Var [Y1 ] = λ A π 1 +
.
α−1
(15)
(16)
Then, using E[Y12 ] = Var[Y1 ] + (E[Y1 ])2 , we obtain
E Y12 =
For Y2 =
α−ρ2−2α
α−1
α−2ρ2−α
α−2
2
λ Aπ
+ λ Aπ
2−2α
α−ρ
2
+
λ
π
+
2
ln(ρ))
(1
λ Aπ
A
α−1
ÍL
i=1
if α , 2
if α = 2.
(17)
l 2 (ri ), using again Campbell’s theorem, we have
α − ρ2−2α
= Var[Y1 ].
(18)
α−1
From the above results, we make the following observations:
E[Y2 ] = λ A π
•
•
•
Y2 scales proportionally to λ A;
The higher order element of Y12 scales proportionally to
λ2A;
When the pathloss is bounded, both Y2 and Y12 have finite
mean, which increase with λ A.
Given these observations, one intuitive conclusion is that when
E[Y2 ]
Y2
λ A increases, NY
2 → 0, which implies N E[Y 2 ] → 0. In other
1
E[Y2 ]
approaching 0, the AP
N E[Y12 ]
1. Since the AP density is
density should satisfy λ A απ(α−1)
(α−2)2
2
measured in APs per m , the condition for channel hardening
With small N, in order to have
1
2]
words, if NE[Y
does not converge to zero when λ A increases,
E[Y12 ]
adding more APs will not help the channel to harden. We
continue to investigate this necessary condition for channel
hardening below.
When the network region grows infinity large, i.e., ρ → ∞,
depending on the pathloss exponent, we have the following
cases:
is only satisfied if λ A ∼ 1 AP/m2 , which is a rather unrealistic
condition in practice.
2) α = 2: This case behaves as in a free-space propagation
environment. As ρ → ∞, we have ln(ρ) → ∞ and ρ2−2α → 0,
which implies
E[Y2 ] → λ A π
α
,
α−1
(22)
E[Y2 ]
1/N
→ 0,
NE Y12
1 + λ A π(1 + 2 ln(ρ))2 α−1
α
(23)
where the operator means that the difference between the
expressions vanishes asymptotically. From (23), we observe
that channel hardening is achieved as the network radius
increases.
3) 1 < α < 2: One example of this case is the indoor
near field propagation. With ρ → ∞, we have ρ2−α → ∞ and
ρ2−2α → 0, which implies
E[Y2 ]
1/N
2
4−2α
NE Y1
1 + λ A π 4ρ
(2−α)2
α−1
α
→ 0.
(24)
2]
From the above equations, we see that NE[Y
decreases
E[Y12 ]
rapidly with λ A and ρ when α ≤ 2. When the network region
2]
will eventually approach 0. This
grows infinitely large, NE[Y
E[Y12 ]
suggests that with smaller pathloss exponents, e.g., free-space
propagation and indoor near field propagation, it is more
likely to observe channel hardening in CF Massive MIMO.
With the two-ray ground-reflection pathloss model and α = 4
[29], the convergence to channel hardening only happens with
impractically high antenna density.
In order to validate our analytical predictions, we present
in Fig. 3 the simulated pθ (i.e., the CDF of Xch ) for different
AP densities, obtained with pathloss exponents α ∈ {3.76, 2}.
Note that in this figure we only consider N = 1. We have
chosen large values of λ A in order to see the behavior of
pθ when λ A → ∞. Fig. 3 shows that with α = 3.76, for a
given threshold θ, the channel hardening measure pθ does
not change much with the AP intensity, unless we reach
λ A = 105 /km2 (0.1/m2 ). However, having λ A > 103 /km2 is
probably practically unreasonable. With α = 2, the convergence of the channel hardening measure pθ to one becomes
more obvious when the AP density grows, which indicates
that the probability to observe channel hardening at random
locations is fairly large.
6
an array of 5-10 antennas. With a smaller pathloss exponent,
the required number of antennas per AP to achieve reasonably
strong channel hardening is also smaller.
1
0.9
0.8
0.7
1
5
=10 /km , =2
0.9
=103 /km 2 , =2
0.8
=102 /km 2 , =2
A
0.7
A
0.5
A
0.4
0.3
2
A
0.2
A
0.1
A
5
2
0.6
3
2
0.5
N=50,
2
2
0.4
N=20,
=10 /km , =3.76
p
p
0.6
=10 /km , =3.76
=10 /km , =3.76
0.3
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.2
N=5,
0.1
Fig. 3. The CDF of Xch , with pathloss exponent α ∈ {3.76, 2}. The
network radius is ρ = 0.5 km and N = 1. The AP density is λ A ∈
{102, 103, 105 }/km2 , which is equivalent to {10−4, 10−3, 0.1}/m2 .
Property 1. Increasing the number of antennas per AP in CF
Massive MIMO always helps the channel to harden. With one
antenna per AP, increasing the AP density does not lead to
channel hardening when using typical pathloss exponents and
AP densities. In a propagation environment with a very small
pathloss exponent, α ≤ 2, the channel hardening criterion has
higher chance to be satisfied as the AP density increases.
B. More Antennas on Few APs or More APs with Few
Antennas?
When the antenna density µ = Nλ A is fixed, whether to
choose larger N with smaller AP density λ A or vice versa to
achieve a high level of channel hardening can be inferred from
(21) for α > 2. We can rewrite (21) as
E[Y2 ]
1
1
=
=
.
α(α−1)
NE Y12
N + Nλ A π
N + µπ α(α−1)2
2
(α−2)
(25)
(α−2)
Since the denominator contains N plus a constant term for
fixed µ, we will clearly obtain more channel hardening by
having more antennas on fewer APs if we have a limited
number of antennas to deploy.3 For α ≤ 2, we can get the
same observations from (23) and (24). Note that the stronger
hardening comes at the price of less macro diversity.
In Fig. 4, we present pθ (i.e., the CDF of Xch ) for different
λ A and N while keeping the overall antenna density fixed at
µ = Nλ A = 103 /km2 (10−3 /m2 ). This figure confirms our
prediction from (25) that having multiple antennas per AP
will substantially help the channel to harden, and the level of
channel hardening clearly increases with N. The curve N = 50
can be interpreted as a cellular Massive MIMO system, due
to the massive number of antennas per AP. The largest gains
occur when going from N = 1 to N = 5 (or to N = 10), thus
we can achieve reasonable strong channel hardening within
the scope of CF Massive MIMO if each AP is equipped with
3 This
result was obtained with uncorrelated fading between the user and
the antennas on an AP. If there instead is spatially correlated fading, due to
insufficient scattering around the AP, this will slightly reduce the hardening,
but more antennas will still be beneficial.
A
N=10,
1
Threshold
A
N=1,
A
A
A
=20/km 2
=50/km 2
=100/km 2
=200/km 2
=1000/km 2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Threshold
Fig. 4. The CDF of Xch with pathloss exponent α = 3.76 and network radius
ρ = 0.5 km. The antenna density is µ = N λ A = 1000/km2 (10−3 /m2 ).
C. Distance vs. Number of Antennas
With higher AP density, the average number of antennas in
a certain area is larger, thus the distances from the nearest
APs to the user are generally smaller—this is the macro
diversity effect. The effect of larger AP density on the channel
hardening criteria can be related to the reduced distance and/or
the increased antenna number. To understand whether the
distance or the number of antennas plays the dominant role,
we now examine the impact of the AP density while assuming
that the typical user is only served by the Mc nearest antennas,
where Mc is usually much smaller than the total antenna
number M in the network region. Here, we focus only on
the case with a single antenna per AP. Thus, the Mc nearest
antennas represent also the Mc nearest APs.
From existing results on the distance distribution in
Poisson networks, the joint PDF of the distances r =
[r1, r2, . . . , r Mc ]T from the Mc nearest antennas to the typical
2
user is fr (x1, . . . , x Mc ) = e−πλ A x M c (2πλ A) Mc , for 0 < x1 <
. . . < x Mc < ∞ [30]. To avoid the dependence between
the distribution of the Mc distance variables, we consider
an approximately equivalent case where a fixed number Mc
antennas are uniformly and independently distributed in a disk
B(0, R) centered at the typical user at the origin. Here, R is an
average radius
determined by the equivalent antenna density,
q
Mc
i.e., R = πλ A .
As a result of the i.i.d. uniform distribution, the joint PDF
of the Mc points within B(0, R) is
Mc
Ö
2xi
fr (x1, . . . , x Mc ) =
(26)
R2
i=1
for x1, . . . , x Mc ∈ [0, R]. Then we have
∫
E [Xch ] ≈
0
R
∫
···
0
R
M
Íc
l 2 (rm ) Ö
Mc
2xi
dx1 · · · dx Mc . (27)
M
2
R2
Íc
i=1
l(rm )
m=1
m=1
Here, the approximation comes from two parts:
7
•
•
Mc M, i.e., we consider that only the Mc nearest APs
contribute to the majority of the channel gain and Mc is
usually much smaller than M;
Our assumption q
on having Mc antennas uniformly disMc
tributed on B(0, πλ
) is an approximation of the real
A
joint distance distribution of the Mc nearest antennas.
more APs will affect the network statistics observed from
the user.
Property 2. When the AP density grows, the shortened distances from nearby APs to the user have little impact on the
channel hardening criterion. Increasing the total number of
antennas in an environment with small pathloss exponent have
stronger influence on the channel hardening than in a system
with a larger pathloss exponent.
1
=3.76, M c=10, Approximation
0.9
In this section, we extend our analysis to the scenario with
a multi-slope pathloss model, which models the fact that the
pathloss exponent generally increases with the propagation
distance. Similar to [10], we consider the three-slope pathloss
model
Cr −3.5 if r > d1
−2
Cr d1−1.5 if d0 ≤ r ≤ d1
l(r) =
(28)
Cd −2 d −1.5 if r < d0,
0 1
=3.76, M c=M, Simulation
0.8
Mean Value of Xch
D. Multi-Slope Pathloss Model
=3.76, M c=10, Simulation
=2, M c=10, Approximation
0.7
=2, M c=10, Simulation
=2, M c=M, Simulation
0.6
0.5
0.4
0.3
0.2
0.1
0
500
1000
1500
2000
2500
AP Density
3000
3500
4000
4500
5000
A
Fig. 5. The mean value of Xch . The approximation results are obtained from
(27) with Mc = 10. The simulation results are obtained for the case with
Mc = 10 nearest APs transmitting to the user and the case with all the M
antennas inside the entire network area of radius ρ = 0.5 km transmitting to
the user. λ A is measured in AP/km2 .
In Fig. 5, we present the approximation and simulation results
of E [Xch ] for different antenna densities. We compare the
approximation obtained from (27) with Mc = 10, and the mean
value of E [Xch ] obtained in simulations when assuming only
the nearest Mc = 10 antennas are transmitting to the user. For
comparison, we present also simulation results of E [Xch ] with
all the M antennas inside the entire network area of radius
ρ = 0.5 km transmitting to the user at the origin. From this
figure we make the following observations:
•
•
•
Both analytical and simulation results show that the mean
value of Xch is barely affected by the AP density λ A when
λ A > 103 /km2 , but it depends on the pathloss exponent
α. The smaller α is, the smaller mean of Xch we have,
the more likely we can observe channel hardening, which
evinces our previous observations;
The approximation in (27) provides an accurate measure
of E [Xch ] when only a few nearest APs affect the channel
gain variation, i.e., when the pathloss exponent α is large;
With smaller pathloss exponent, e.g., α = 2, the mean
value of Xch when the typical user is served by M
Mc antennas is much lower than by Mc = 10 nearest
antennas. This implies that the number of APs that affect
the channel gain variation is much higher than 10. For
example, when λ A = 103 /km2 , based on Monte Carlo
simulations, approximately the nearest 200 APs can affect
the distribution of Xch . The pathloss exponent determines
the speed of signal attenuation, therefore, with smaller α,
where d0 and d1 are fixed distances at which the slope starts to
change, C is a constant that depends on the carrier frequency
and antenna height. Since the constant factor C does not affect
the channel hardening measure, for simplicity, we consider
C = 1 in the remainder of this section.
2]
→ 0 as the
As previously in this paper, we consider NE[Y
E[Y12 ]
necessary condition of channel hardening. When ρ → ∞,
∫
7
−1.5
E[Y1 ] = λ A2π l(r)rdr = 2λ A πd1
ln d1 − ln d0 +
;
6
R
(29)
∫
3
Var[Y1 ] = λ A2π l 2 (r)rdr = 2λ A πd1−3 d0−2 − d1−5 . (30)
10
R
E Y12 = Var[Y1 ] + (E[Y1 ])2
"
#
2
7
3 −5
−3
−2
= 2λ A πd1 2πλ A ln d1 −ln d0 +
+d0 − d1 .
6
10
(31)
Since E[Y2 ] = Var[Y1 ], we have
E[Y2 ]
=
NE Y12
≈
1/N
1 + 2λ A π
(ln d1 −ln d0 + 76 )
2
3 −5
d0−2 − 10
d1
1/N
1+
2λ A πd02
ln d1 − ln d0 +
7 2
6
,
(32)
where the approximation holds when d1 1 m. Based on
2]
this, using larger d0 and d1 will make NE[Y
approach 0 with
E[Y12 ]
higher speed when λ A increases. This is intuitive, given the
previous observation that a smaller pathloss exponent improves
the hardening, because as d0 and d1 increase, the number of
APs with small pathloss exponents increases. Adding more
antennas to the APs will also improve the channel hardening,
both for a fixed λ A and when the total antenna density µ =
Nλ A is fixed.
Fig. 6 shows pθ (i.e., the CDF of Xch ) obtained with the
three-slope pathloss model, with d0 = 10 m and d1 = 50 m.
In this figure, we compare pθ obtained with different values
8
1
0.9
0.8
0.7
p
0.6
A
0.5
A
0.4
A
0.3
A
0.2
A
0.1
A
=200/km 2, N=10
=100/km 2, N=10
=50/km 2, N=10
=2000/km 2, N=1
=1000/km 2, N=1
=500/km 2,N=1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Threshold
Fig. 6. The CDF of Xch with the three-slope pathloss model. d0 = 10 m and
d1 = 50 m. The total antenna density is µ = N λ A = {500, 1000, 2000}/km2 .
of the AP density λ A for N = 1 and N = 10 antennas per
AP. First, with N = 1, when the antenna density increases,
the increase of pθ is substantial, and more influential than the
results in Fig. 3 and Fig. 4 with the one-slope model. Second,
comparing the results obtained with N = 1 and N = 10, we see
that the number of antennas per AP plays a more important
role than increasing the AP density in helping the channel
to harden, in terms of achieving a small Xch with practically
reasonable AP density values.
Property 3. With the three-slope pathloss model, due to
the small pathloss exponent of the propagation environment
nearby the user, the channel gain variance declines rather fast
with the AP density compared to the mean value. Furthermore,
having large number of antennas per AP can guarantee small
channel variation, which makes the channel hardening easier
to achieve.
IV. FAVORABLE P ROPAGATION
In this section, we define and analyze the favorable propagation conditions in CF Massive MIMO networks. Similar to
the previous section, we will consider both conventional CF
networks with single-antenna APs and a generalization with
multiple antennas per AP.
Recall that the channel vector from the M antennas to the
user k is p
gk = [g1,k , . . . , g M,k ]T , where the m-th element
is gm,k = l(dm,k )hm,k . To have favorable propagation, the
channel vectors between the BS and the user terminals should
be orthogonal, which means
gkH g j =
0
kgk k 2 , 0
if
if
k,j
k = j.
(33)
When this condition is satisfied, each user can get the same
communication performance as if it is alone in the network
[31]. In practice, this condition is not fully satisfied, but can be
approximately achieved when the number of antennas grows
to infinity, in which case the channels are said to provide
asymptotically favorable propagation. To be more specific, in
CF Massive MIMO, the asymptotically favorable propagation
condition can be defined as follows:
gkH g j
q
−→ 0, when M → ∞, k , j.
E kgk k 2 |dk E kg j k 2 |d j
(34)
Here, we have conditioned on a specific network realization with distance vectors dk = [d1,k , . . . , d M,k ]T and d j =
[d1, j , . . . , d M, j ]T , and each element dm,k represents the distance from the m-th antenna to the user k. Recall that
every N antennas are co-located at the same AP, we have
d(i−1)·N +1,k = . . . = di ·N,k for i = 1, . . . , L.
Different from cellular Massive MIMO with co-located
antennas, the large-scale fading coefficients from each antenna
in CF Massive MIMO to a user are different, which can be
viewed as a type of spatial channel correlation. Thus, we have
M q
Õ
∗
l(dm,k )l(dm, j )hm,k
hm, j
(35)
gkH g j =
m=1
and
v
u
tM
M
q
Õ
Õ
2
2
E kgk k |dk E kg j k |d j =
l(dm,k )
l(dm, j ).
m=1
m=1
(36)
Since {hm,k } are i.i.d. CN (0, 1) RVs, we have
h
i 0 if k , j
∗
E hm,k
hm, j =
1 if k = j
and it follows that
gkH g j
E q
d
,
d
=
k
j
2
2
E kgk k |dk E kg j k |d j
(37)
0
1
if k , j
if k = j.
(38)
With this mean value, the convergence in (34) holds (in mean
square sense and in probability) if the variance of the left-hand
side goes asymptotically to zero. Using (35) and (36), we have
gkH g j
Var q
d
,
d
(39)
k j
E kgk k 2 |dk E kg j k 2 |d j
L
Í
N
l(di ·N,k )l(di ·N, j )
i=1
=
(40)
L
L
Í
Í
N 2 l(di ·N,k ) l(di ·N, j )
i=1
i=1
L
≤
N L2
1
L
L
Í
i=1
l(di ·N,k )
1
L
L
Í
.
(41)
l(di ·N, j )
i=1
As the AP density λ A grows, L increases.
1
L
L
Í
i=1
l(di ·N,k ) will
approach E[l(di ·N,k )] > 0, which is a positive value that only
depends on the network size and the pathloss model. Thus,
(40) is upper-bounded by a positive value that decreases as
1/L when L increases. When L → ∞, the variance of the
channel orthogonality will approach 0. Combined with (38),
we can prove that the asymptotically favorable propagation
condition defined in (34) holds for CF Massive MIMO.
9
i=1
1
0.9
0.8
0.7
0.6
p
For finite L, we use (40) to define the channel orthogonality
metric
L
Í
l(di ·N,k )l(di ·N, j )
i=1
Xfp =
(42)
L
L
Í
Í
N
l(di ·N,k ) l(di ·N, j )
i=1
0.5
A
0.4
A
0.3
and consider the probability that two users at random locations
have Xfp no larger than a threshold γ:
pγ = P[Xfp ≤ γ].
A
0.2
A
0.1
A
=25/km 2, N=20
=50/km 2, N=10
=100/km 2, N=5
=500/km 2, N=1
=100/km 2, N=1
0
(43)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Threshold
A. Impact of Antenna Density on Channel Orthogonality
The impact of the antenna density µ = Nλ A will be analyzed
in two cases: fixed λ A with different N or fixed N with
different λ A. As mentioned above, Xfp is inversely proportional
to N when L is fixed, so increasing N always helps the
channels to become more orthogonal. In the other case, when
L increases, the denominator of Xfp grows almost as L 2 ,
while the numerator increases almost linearly with L. Knowing
that L is a Poisson RV with mean value proportional to λ A,
consequently, Xfp should scale roughly inversely proportional
to λ A, which evinces that for a given γ, pγ will grow with the
AP density λ A. Combining the two cases, we can see that both
larger N and larger AP density λ A can help the channel to offer
more favorable propagation. Fig. 7 presents the CDF of the
channel orthogonality metric Xfp with different λ A and N. By
comparing the results obtained with λ A = {500, 100}/km2 and
N = {1, 5} (marked with circle, left triangle and plus sign), we
validate that both increasing λ A and increasing N can improve
the channel orthogonality.
Fig. 7. The CDF of Xfp . Pathloss exponent α = 3.76. Network radius ρ = 0.5
km. For the first four curves (marked with upward triangle, cross, circle and
left triangle), the antenna density is µ = N λ A = 500/km2 (5 × 10−4 /m2 ). The
distance between user j and user k is 70 m.
C. Impact of Inter-User Distance on Channel Orthogonality
From (34), it is obvious that the distance
between two users
p
∗ h
affects the variance of each term l(dm,k )l(dm, j )hm,k
m, j .
When two users are far apart, their channel vectors are more
likely to be orthogonal with smaller variance. This result
comes from the fact that l(dm,k )l(dm, j ) for all m = 1, . . . , M
will become much smaller when the distance lk, j between user
ÍM
ÍM
k and user j is large. In addition, m=1
l(dm,k ) m=1
l(dm, j )
will not vary much with the inter-user distance lk, j when M is
fixed. Therefore, Xfp becomes smaller when the distance lk, j
increases.
1
0.9
0.8
0.7
0.6
p
Clearly, in order to have asymptotically favorable propagation,
when the antenna density grows, pγ should approach one for
any γ ≥ 0. For practical purposes, it is desirable that pγ is
large for values of the threshold γ that are close to zero. In the
following, we will analyze how the antenna density, the interuser distance, and the pathloss exponent affects the channel
orthogonality.
A
=200/km 2, lk,j =212m
A
=100/km 2,l k,j =212m
A
=50/km 2,l k,j =212m
A
=200/km 2, lk,j =70m
A
=100/km 2, l k,j =70m
A
=50/km 2, l k,j =70m
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Threshold
B. More Antennas on Few APs or More APs with Few
Antennas?
From (42), we see that the value of Xfp is always upperbounded by N1 . When the antenna density µ = Nλ A is fixed,
increasing N means smaller λ A. As the result, the average
number of APs within close distance to the user will be less.
Thus, it is hard to predict whether it is more beneficial to have
more antennas on few APs or more APs with few antennas.
In Fig. 7, we present the CDF of Xfp when fixing the total
antenna density Nλ A = 500/km2 . We see that increasing N
does not necessarily lead to higher or lower pγ for a given
value of γ. We also observe that when choosing sufficiently
large N, e.g., N ≥ 20, the channel orthogonality metric
Xfp becomes very small. In other words, sufficiently large N
will help the channels to different users to be asymptotically
orthogonal.
Fig. 8. The CDF of Xfp with different λ A and different inter-user distances
ld, j , N = 1, and pathloss exponent α = 3.76.
In Fig. 8, we present pγ for different AP densities λ A ∈
{50, 100, 200} APs/km2 , N = 1, and inter-user distances lk, j ∈
{70, 212} m. First, it is shown that with larger λ A, the variance
of the orthogonality metric Xfp is smaller. Second, when the
distance between two users is larger, they are more likely to
have nearly orthogonal channels. This observation showcases
the importance of serving spatially separated users in order to
ensure near channel orthogonality.
D. Impact of Pathloss Exponent on Channel Orthogonality
When the number of antennas M and their locations are
fixed, we consider two extreme cases: user k and user j are
10
very close or extremely far from each other. Since Xch coincides with Xfp in the special case when di ·N,k ' di ·N, j for all
m = 1, . . . , M, we infer from Section III that smaller pathloss
exponent will also lead to smaller Xfp . In the other extreme
case,Íwhen the two users are far apart, in the
Í Ldenominator of
L
Xfp , i=1
l(di ·N, j ) is almost independent of i=1
l(di ·N,k ), and
both terms increase much faster than the numerator, especially
when α is small. Combining these two extreme cases, we
expect that smaller pathloss exponent would help the channels
to become asymptotically orthogonal when M is sufficiently
large.
1
0.9
0.8
Consider a CF Massive MIMO system with M singleantenna APs and K users, which are assigned mutually orthogonal pilot sequences. The transmission is divided into
coherence intervals of τc samples, whereof τp are used for
uplink pilot signaling and K ≤ τp ≤ τc . The channels are
modeled as in previous sections: gk ∼ CN (0, Bk ), where
Bk = diag(β1,k , . . . , β M,k ) and βm,k = l(rm,k ). Each user transmits its orthogonal pilot sequence. User k uses the transmit
powers ρk and pk for pilot and data, respectively. Since the
elements in gk are independent, it is optimal to estimate them
separately at the receiving antenna. The MMSE estimate of
gm,k is
p
τp ρk βm,k √
τρk gm,k + wm,k ,
(44)
ĝm,k =
τp ρk βm,k + 1
0.7
where wm,k ∼ CN (0, 1) is i.i.d. additive noise. If we denote
4
by γm,k = E | ĝm,k | 2 the mean square of the MMSE estimate
of gm,k , then it follows from (44) that
p
0.6
0.5
0.4
0.3
=2
=3
=4
0.2
0.1
γm,k =
2
τρk βm,k
τρk βm,k + 1
.
(45)
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Threshold
Fig. 9. The CDF of Xfp with different pathloss exponent α ∈ {2, 3, 4},
N = 1, and the inter-user distance lk, j = 70 m.
Fig. 9 shows pγ obtained with pathloss exponents α ∈
{2, 3, 4}, N = 1, and an inter-user distance of lk, j = 70 m. The
figure shows that with smaller α the channels become more
orthogonal, which is line with our prediction from above. If
we would instead use the three-slope pathloss model in (28),
APs close to the user (distance smaller than d1 ) will have
pathloss exponent α ≤ 2, and users at larger distances have
pathloss exponent 3.5. Therefore, compared to the single-slope
non-singular pathloss model l(r) = min(r −3.76, 1), the channels
between two users will in the average be more orthogonal.
Summarizing the above analysis and observations, we have
the following conclusions.
Property 4. Increasing the antenna density by increasing either the AP density or the number of antennas per AP can both
help the user channels to offer favorable propagation. Smaller
pathloss also helps the channels to become asymptotically
orthogonal. The larger the distance between two users, the
more likely their channels will be nearly orthogonal.
V. C APACITY B OUNDS FOR C ELL -F REE M ASSIVE MIMO
A key conclusion from the previous sections is that CF
Massive MIMO systems exhibit little channel hardening, as
compared to cellular Massive MIMO. Hence, although CF
Massive MIMO is equivalent to a single-cell Massive MIMO
system with strong spatial channel correlation, we must be
careful when reusing results from the Massive MIMO literature. In particular, capacity bounds that were derived by
relying on channel hardening can potentially be very loose
when applied to CF systems. In this section, we explain which
capacity lower bounds are suitable for CF systems.
We will now compare different achievable rate expressions for
uplink and downlink when using maximum ratio (MR) processing, which is commonly assumed in CF Massive MIMO
since it can be implemented distributively. The expressions are
lower bounds on the capacity, thus we should use the one that
gives the largest value the accurately predict the achievable
performance.
A. Uplink Achievable Rate
During the uplink data transmission, all K users simultaneously transmit to the M APs. When using MR, an achievable
rate (i.e., a lower bound on the capacity) of user k is
2
M
Í
©
ª
pk
γm,k
®
m=1
UatF
®,
Ru,k
= log2 1 +
(46)
®
K
M
M
Í
Í
Í
®
pj
γm,k βm, j +
γm,k
j=1
m=1
m=1
«
¬
which was used for CF Massive MIMO in [10]. This bound
is derived based on the use and then forget (UatF) principle
[12], where the channel estimates are used for MR but then
“forgotten” and channel hardening is utilized to obtain a
UatF is a special
simple closed-form expression. Note that Ru,k
case of the general expression in [32] for correlated singlecell Massive MIMO systems that apply MR processing. Alternatively, the achievable rate expression in [16] for spatially
correlated channels can be used:
Ru,k
©
ª
H
2
®
pk |ak ĝk |
®
= E log2
Í
® ,
K
K
Í
H
H
p j |ak ĝ j | 2 + ak
p j (B j − Γ j ) + I M ak ®
j=1
j,k
«
¬
(47)
11
constant factor C (dB) is given by
1
0.9
0.8
C =105 + 46.3 + 33.9 log 10( f ) − 13.82 log 10(h AP )
0.7
− (1.1 log 10( f ) − 0.7)hu + (1.56 log 10( f ) − 0.8),
CDF
0.6
0.5
0.4
0.3
Perfect CSI
General rate
UatF rate
0.2
0.1
0
0
1
2
3
4
5
6
7
8
Per-user rate [bit/s/Hz]
Fig. 10. The CDF of the uplink achievable rates obtained with the UatF
bound, the general bound, and the rate with perfect CSI. M = 100, N = 1,
K = 20, τ p = 20.
1
0.9
0.8
0.7
0.6
CDF
(48)
0.5
0.4
0.3
Perfect CSI
General rate
UatF rate
0.2
0.1
0
0
1
2
3
4
5
6
7
8
9
10
Per-user rate [bit/s/Hz]
Fig. 11. Same as Fig. 10. M = 100, N = 5, K = 20, τ p = 20.
where f is the carrier frequency, h AP and hu are the AP
and user antenna height, respectively [10]. The simulation
parameters are summarized in Table I. Fig. 10 and Fig. 11
show the CDFs of the user rates for random distances between
the L APs and the K users. The general rate expression in
(47) provides almost identical rates to the perfect CSI case,
which indicates that the estimation errors to the closest APs
are negligible. In contrast, the UatF rate in (46) is a much
looser bound on the capacity, particularly for the users that
support the highest data rates. Some users get almost twice
the rate when using the general expression. This is due to the
lack of channel hardening for users that are close to only a
few APs. Hence, a general guideline is to only use (47) for
achievable rates in CF Massive MIMO with single-antenna
APs. An important observation from Fig. 10 and Fig. 11 is that,
with multiple antennas per AP, e.g., N = 5, the differences
between the UatF bound and the perfect CSI case becomes
smaller, particularly for the users that support the smallest
rates. This comes from the fact that having multiple antennas
per AP gives better channel hardening, as we have seen in
Section III. However, the average rate reduces when going
from many single-antenna APs to fewer multi-antenna APs,
so it is not obvious what is the better choice in practice.
TABLE I
S IMULATION S ETUP
Parameters
Values
M, K, τp , τc
d0 , d1
Carrier frequency f
Antenna height h AP , hu
Uplink pilot power ρk , data power pk
Downlink power per UE q
100, 20, 20, 500
10 m, 50 m
1.9 GHz
1.5 m, 1.65 m
100 mW, 100 mW
100 mW
where ĝk = [ĝ1,k , . . . , ĝ M,k ]T , Γ j = diag(γ1, j , . . . , γ M, j ), and
ak is the combining vector, which is ak = ĝk for MR. This
general bound does not rely on channel hardening.
In Fig. 10 and Fig. 11, we compare the uplink achievable
rates obtained with (46) and (47), and also the rate with perfect
CSI (obtained from (47) by letting ρk → ∞). The simulation
is performed in a network area of 1 km ×1 km. K = 20 users
are randomly and uniformly distributed in the network region,
and we have τp = K. The total number of antennas is M =
100. The results in Fig. 10 are obtained with L = 100 singleantenna APs and those in Fig. 11 are obtained with L = 20 APs
with N = 5 antennas per AP. All the APs are independently
and uniformly distributed in the network region. The length
of the coherence block is τc = 500. The large-scale fading
coefficients between the antennas and the users are generated
from 300 different PPP realizations. The three-slope pathloss
model in (28) is used with d0 = 10 m and d1 = 50 m. The
B. Downlink Achievable Rate
In the downlink of Massive MIMO, it is popular to use rate
expressions where the only CSI available at the user is the
channel statistics. When there is channel hardening, one can
achieve rates similar to the uplink. If MR precoding is used,
then one can use the rate expression for spatially correlated
fading from [16] and obtain the achievable rate
UatF
Rd,k
2
M √
Í
©
ª
pm,k γm,k
®
m=1
®
= log2 1 +
®
K
M
Í Í
pm,k 0 βm,k + 1 ®
k 0 =1 m=1
«
¬
(49)
for user k, where pm,k is the average transmit power allocated
to user k by AP m. This type of expression was used for CF
Massive MIMO in [10], using a slightly different notation. We
call this the UatF rate since we use the received signals for
detection, but “forget” to use them for blind estimation of the
instantaneous channel realizations.
Suppose ak is the precoding vector (including power allocation) assigned to user k. To avoid relying on channel
hardening, user k can estimate its instantaneous channel after
precoding, akH gk , from the collection of τd received downlink
12
signals in the current coherence block. By following a similar
approach as in [33, Lemma 3], we obtain the achievable rate
Rd,k =
©
ª
K
H
2
Õ
®
|ak gk |
H
® − 1
E log2
g
]
.
log
1
+
τ
Var[a
k
d
2
j
® τd
K
Í
®
H
2
j=1
|a j gk | + 1
« j,k
¬
(50)
The first term in (50) is the rate with perfect CSI and the second term is a penalty term from imperfect channel estimation
at the user. Note that the latter term vanishes as τd → ∞,
thus this general rate expression is a good lower bound when
the channels change slowly. The
is
q precoding
q rate with MR
p1, k
p M, k T
obtained by setting ak = ĝ1,k γ1, k . . . ĝ M,k γ M, k .
1
0.9
0.8
0.7
CDF
0.6
0.5
0.4
0.3
Perfect CSI
General rate
UatF rate
0.2
0.1
0
0
1
2
3
4
5
VI. C ONCLUSION
In this work, we provided thorough investigation of the
channel hardening and favorable propagation phenomena in
CF Massive MIMO systems from a stochastic geometry perspective. By studying the channel distribution from stochastically distributed APs with either single or multiple antennas
per AP, we characterized the channel gain distribution, based
on which we examined the conditions of channel hardening
and favorable propagation. Our results showed that whether the
channel hardens with increasing number of antennas strongly
depends on the propagation environment and pathloss model,
but one should generally not expect much hardening. However,
one can improve the hardening by deploying multiple antennas
per AP. There are several factors that can help the channel to
provide favorable propagation, such as smaller pathloss exponent, higher antenna density, and spatially separated users with
larger distances. Well separated users will generally exhibit
favorable propagation since they are mainly communicating
with different subsets of the APs.
A main implication of this work is that one should not
rely on channel hardening and favorable propagation when
computing the achievable rates in CF Massive MIMO, because
this could lead to a great underestimation of the achievable
performance. There is a good uplink rate expression for
spatially correlated Massive MIMO systems that can be used.
Further development of downlink rate expressions is needed
to fully understand the achievable downlink performance in
CF Massive MIMO.
6
Per-user rate [bit/s/Hz]
A PPENDIX
Fig. 12. The CDF of the downlink achievable rates obtained with the UatF
bound, the general bound, and the rate with perfect CSI. M = 100, N = 1,
K = 20, τ p = 20.
In Fig. 12, we compare the downlink achievable rates
obtained with (49) and (50), and also the rate with perfect
CSI at user (obtained from (50) by letting τd → ∞). The
simulation parameters are summarized in Table I, which are
the same as in Section V-A for the uplink. The downlink power
pm,k is given by
q · γm,k
γm,k
pm,k =
,
(51)
=q
2
M
Í
E[k ĝk k ]
0
γm ,k
m0 =1
where q is the downlink power allocated to each user k, which
is chosen as 100 mW in the simulations. By doing so, we have
E[kak k 2 ] = q, which is the same for all users. Fig. 12 shows
the CDFs of the user rates for random antenna-user distances.
As in the uplink, there is a substantial gap between the rates
achieved by perfect CSI curve and the UatF rate. The curve
for the general rate in (50) is in the middle, which implies
that the users should estimate their downlink channels and not
rely on channel hardening in CF Massive MIMO. It is unclear
whether the gap from the perfect CSI curve is due to limited
downlink estimation quality or is an artifact from the capacity
bounding technique that lead to (50). In any case, there is
a need to further study the achievable downlink rates in CF
Massive MIMO.
A. Proof of (5) and (6)
Recall Campbell’s theorem as follows [23]. If f (x) :
Rd → [0, +∞] is a measurable function and Φ is a stationary/homogeneous PPP with density λ, then
"
#
∫
Õ
E
f (x) = λ
f (x)dx.
(52)
Rd
x ∈Φ
Since Φ A is a two-dimensional homogeneous PPP, for a finite
network region with radius ρ, we have
"
#
Õ
2
E kgk k = E
Hi l(ri )
i ∈Φ A
= E [Hi ] · λ A
∫
l(k xk)dx.
B(0,ρ)
∫ ρ
= λ A · E [Hi ] 2π
l(r)rdr.
(53)
0
Note that l(r) = min(1, r −α ) and E [Hi ] = N as a result of the
Gamma distribution, thus
∫ 1
∫ ρ
E kgk k 2 = Nλ A2π
rdr +
r 1−α dr .
(54)
0
Depending on the value of α, we have
(
2(1−ρ2−α )
Nλ A π 1 + α−2
2
E kgk k =
Nλ A π (1 + 2 ln ρ)
1
if α , 2
if α = 2
(55)
13
From [23], we have the expression for the variance
"
#
∫
Õ
Var
f (x) = λ
f (x)2 dx.
x ∈Φ
(56)
Rd
2
= (E [Hi ])2 +
Since
i ∼ Gamma(N, 1), we have E Hi
H
2
2
Var Hi = N + N, thus
∫ ρ
2
2
Var kgk k = λ A · E Hi 2π
l 2 (r)rdr
0
∫ 1
∫ ∞
2
1−2α
= (N + N)λ A2π
rdr +
r
dr
1
0
1
2
2−2α
= (N + N)λ A π 1 +
. (57)
1−ρ
α−1
Note that (57) will have a different form when α = 1, which
is unlikely to happen in a real propagation environment. Thus,
this case is not discussed here.
R EFERENCES
[1] Z. Chen and E. Björnson, “Can we rely on channel hardening in cell-free
Massive MIMO?” accepted in IEEE GLOBECOM Workshops, 2017.
[2] S. Shamai and B. M. Zaidel, “Enhancing the cellular downlink capacity
via co-processing at the transmitting end,” in Proc. IEEE VTC-Spring,
vol. 3, 2001, pp. 1745–1749.
[3] S. Venkatesan, A. Lozano, and R. Valenzuela, “Network MIMO: Overcoming intercell interference in indoor wireless systems,” in Proc. IEEE
ACSSC, 2007, pp. 83–87.
[4] D. Gesbert, S. Hanly, H. Huang, S. Shamai, O. Simeone, and W. Yu,
“Multi-cell MIMO cooperative networks: A new look at interference,”
IEEE J. Sel. Areas Commun., vol. 28, no. 9, pp. 1380–1408, 2010.
[5] A. Osseiran, J. F. Monserrat, and P. Marsch, 5G Mobile and Wireless
Communications Technology. Cambridge University Press, 2016, ch.
9, Coordinated Multi-Point Transmission in 5G.
[6] E. Björnson, R. Zakhour, D. Gesbert, and B. Ottersten, “Cooperative
multicell precoding: Rate region characterization and distributed strategies with instantaneous and statistical CSI,” IEEE Trans. Signal Process.,
vol. 58, no. 8, pp. 4298–4310, 2010.
[7] E. Björnson, N. Jaldén, M. Bengtsson, and B. Ottersten, “Optimality
properties, distributed strategies, and measurement-based evaluation
of coordinated multicell OFDMA transmission,” IEEE Trans. Signal
Process., vol. 59, no. 12, pp. 6086–6101, 2011.
[8] H. Q. Ngo, A. E. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta,
“Cell-free Massive MIMO: Uniformly great service for everyone,” in
Proc. IEEE SPAWC, 2015.
[9] E. Nayebi, A. Ashikhmin, T. L. Marzetta, and H. Yang, “Cell-free
Massive MIMO systems,” in Proc. Asilomar, 2015.
[10] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta,
“Cell-free Massive MIMO versus small cells,” IEEE Trans. Wireless
Commun., vol. 16, no. 3, pp. 1834–1850, 2017.
[11] E. Nayebi, A. Ashikhmin, T. L. Marzetta, H. Yang, and B. D. Rao,
“Precoding and power optimization in cell-free Massive MIMO systems,” IEEE Trans. Wireless Commun., vol. 16, no. 7, pp. 4445–4459,
2017.
[12] T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo, Fundamentals
of Massive MIMO. Cambridge University Press, 2016.
[13] H. Q. Ngo and E. Larsson, “No downlink pilots are needed in TDD
Massive MIMO,” IEEE Transactions on Wireless Communications,
vol. 16, no. 5, pp. 2921–2935, May 2017.
[14] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Aspects of favorable
propagation in Massive MIMO,” in European Signal Processing Conference (EUSIPCO), September 2014, pp. 76–80.
[15] H. Huh, G. Caire, H. Papadopoulos, and S. Ramprashad, “Achieving
“massive MIMO” spectral efficiency with a not-so-large number of
antennas,” IEEE Trans. Wireless Commun., vol. 11, no. 9, pp. 3226–
3239, 2012.
[16] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need?” IEEE J. Sel.
Areas Commun., vol. 31, no. 2, pp. 160–171, 2013.
[17] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach
to channel estimation in large-scale multiple-antenna systems,” IEEE J.
Sel. Areas Commun., vol. 31, no. 2, pp. 264–273, 2013.
[18] G. Interdonato, H. Q. Ngo, E. G. Larsson, and P. Frenger, “How much
do downlink pilots improve cell-free Massive MIMO?” in Proc. IEEE
GLOBECOM, 2016.
[19] W. Lu and M. Di Renzo, “Stochastic geometry modeling of cellular
networks: Analysis, simulation and experimental validation,” in ACM
International Conference on Modeling, Analysis and Simulation of
Wireless and Mobile Systems, November 2015, pp. 179–188.
[20] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach
to coverage and rate in cellular networks,” IEEE Transactions on
Communications, vol. 59, no. 11, pp. 3122–3134, November 2011.
[21] M. Haenggi, “The meta distribution of the SIR in Poisson bipolar and
cellular networks,” IEEE Transactions on Wireless Communications,
vol. 15, no. 4, pp. 2577–2589, April 2016.
[22] R. K. Ganti and M. Haenggi, “Interference in ad hoc networks with
general motion-invariant node distributions,” in IEEE International
Symposium on Information Theory, July 2008, pp. 1–5.
[23] M. Haenggi, R. K. Ganti et al., “Interference in large wireless networks,”
Foundations and Trends R in Networking, vol. 3, no. 2, pp. 127–248,
2009.
[24] M. Haenggi, Stochastic geometry for wireless networks. Cambridge
University Press, 2012.
[25] M. Kountouris and N. Pappas, “Approximating the interference distribution in large wireless networks,” in International Symposium on Wireless
Communications Systems (ISWCS), August 2014, pp. 80–84.
[26] K. Smaili, T. Kadri, and S. Kadry, “Hypoexponential distribution with
different parameters,” Applied mathematics, vol. 4, no. 04, p. 624, 2013.
[27] S. V. Amari and R. B. Misra, “Closed-from expressions for distribution
of sum of exponential random variables,” IEEE Trans. Rel., vol. 46,
no. 4, pp. 519–522, 1997.
[28] E. Björnson, D. Hammarwall, and B. Ottersten, “Exploiting quantized
channel norm feedback through conditional statistics in arbitrarily correlated MIMO systems,” IEEE Trans. Signal Process., vol. 57, no. 10,
pp. 4027–4041, 2009.
[29] A. Goldsmith, Wireless Communications. Cambridge University Press,
2005.
[30] D. Moltchanov, “Distance distributions in random networks,” Ad Hoc
Networks, vol. 10, no. 6, pp. 1146–1166, 2012.
[31] F. Rusek, D. Persson, B. Lau, E. Larsson, T. Marzetta, O. Edfors, and
F. Tufvesson, “Scaling up MIMO: Opportunities and challenges with
very large arrays,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–
60, 2013.
[32] E. Björnson, L. Sanguinetti, and M. Debbah, “Massive MIMO with
imperfect channel covariance information,” in Proc. ASILOMAR, 2016.
[33] G. Caire, “On the ergodic rate lower bounds with applications to Massive
MIMO,” CoRR, vol. abs/1705.03577, 2017.
| 7cs.IT
|
A case for multiple and parallel RRAMs as synaptic
model for training SNNs
Aditya Shukla
Dept. of Electrical Engineering
Indian Institute of Technology
Bombay
Mumbai, India
[email protected]
Sidharth Prasad
Dept. of Electrical Engineering
Indian Institute of Technology
Bombay
Mumbai, India
[email protected]
Abstract— To enable a dense integration of model synapses in
a spiking neural networks hardware, various nano-scale devices
are being considered. Such devices, besides exhibiting spike-time
dependent plasticity (STDP), need to be highly scalable, have a
large endurance and require low energy for transitioning
between states. In this work, we first introduce and empirically
determine two new specifications for a synapse in SNNs: number
of conductance levels per synapse and maximum learning-rate.
To the best of our knowledge, there are no RRAMs that meet the
latter specification. As a solution, we propose the use of multiple
RRAMs in parallel within a synapse. While synaptic reading, all
PCMO-RRAMs are simultaneously read and for each synaptic
conductance-change event, the mechanism for conductance
STDP is initiated on only one RRAM, randomly picked from the
set. Second, to validate our solution, we experimentally
demonstrate STDP of conductance of a PCMO-RRAM and then
show that due to a large learning-rate, a single PCMO-RRAM
fails to model a synapse in the training of an SNN. As anticipated,
network training improved as more PCMO-RRAMs are added to
the synapse. Fourth, we discuss the circuit-requirements for
implementing such a scheme, to conclude that the requirements
are within bounds. Thus, our work presents specifications for
synaptic devices in trainable SNNs, indicates the shortcomings of
state-of-art synaptic contenders, and provides a solution to
extrinsically meet the specifications and discusses the peripheral
circuitry that implements the solution.
Keywords—Fisher’s Iris dataset, memristor, PrCaMnO
(PCMO), resistive random-access memory (RRAM), spike-time
dependent plasticity, spiking neural network, synapse
I. INTRODUCTION
While decoding a brain‟s functioning, mimicking biology
using models of brain has been a significant approach taken by
neuro-scientific community. Also, to perform cognitive
computing tasks while maintaining energy efficiency of a
human brain has been the grand challenge in the post-scaling
and data-driven era. Executing these tasks using spiking (or,
third generation) neural networks in a bio-inspired hardware
that is functionally and structurally similar to a brain will
greatly help in achieving these goals. Such bio-mimicry has led
to adoption of a bottom-up approach that distributes and
XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
Sandip Lashkare
Dept. of Electrical Engineering
Indian Institute of Technology
Bombay
Mumbai, India
[email protected]
Udayan Ganguly
Dept. of Electrical Engineering
Indian Institute of Technology
Bombay
Mumbai, India
[email protected]
integrates the processing and memory – electronic neurons as
computing units and electronic synapses as memory units
connected as required in network. Though various VLSIamenable circuits have been proposed that can mimic a
synapse [1], [2], to integrate as any many as 1014 synapses
within a volume of a human brain requires the model synapse
to have dimensions in the order of the thickness of a synaptic
cleft (~10nm) [3]–[5]. This has led to an appreciable research
and development of plethora of novel nano-scaled devices that
can faithfully mimic a synapse [4], [6]–[17]. Of all the novel
material based nano-scaled devices for modeling synapses in
hardware, memristors/RRAMs (Resistive Random-AccessMemory) have gained remarkable research interest. These are
non-volatile resistive memory-elements whose state/resistance
can altered by applying sufficiently strong voltage-pulses and
are strong candidates as weights in electronic in-situ trainable
neural networks.
Spike-time dependent plasticity (STDP) rule, a type of
Hebbian „local‟ learning-rule, is considered to be the essential
property of any SNN based synapse that their models must
exhibit. By changing the strength of voltage-pulses being
applied as a function of time-since-last-spike, STDP rule has
been demonstrated on various RRAMs [4], [10]–[17]. But, for
an extensive and large-scale utilization of these devices in
cross-bars as a synaptic array, it needs to (1) be highly scalable,
(2) have excellent endurance, (3) have low-energy switchability and (4) be compatible with CMOS [4], [18], [19].
However, other important but relatively unexplored requisites
from the synaptic RRAM that are discussed in this work
include: (1) analog range of conductance or ample number of
memory states/bits and (2) a low value of maximum STDP
|
|
based weight-change (mathematically, low
) at each weight-change event, or learning-rate. Requisite
(1) is based on the fact that in nature, most STDP rules
observed in biology are analog [20], [21] and so are the datasets (e.g. data-sets of images in color/grey-scale, Fisher‟s Iris,
Wisconsin‟s breast-cancer, chemical assays like wine [22]–
[24].) With only additional costs can either be synthetically
transformed into the binary domain, thus often necessitating
analog or multi-level synapse. Requisite (2) comes from the
fact that all STDP learning-rules have a point of maximum
weight-change (at point(s) of highest/least time or ratecorrelation; [20], [21]), and this value should be kept small for
a stable weight-evolution in a network while training.
In this paper, we first empirically show that for softwareequivalent training, (1) the learning-rate (maximum
at
each weight-change event) must be less than 2% and (2) a
resolution of at least 256 levels (8 bits) per synapse are needed.
Second, we show that STDP demonstrations on RRAMs up todate depict a learning-rate of 20%-400%, thus, these devices do
not meet our specifications. To tackle this problem we propose
the use of multiple ( ) and parallel RRAMs in a synapse.
Within such a synapse, reading requires all RRAMs but the
conductance-change, as dictated by the STDP rule, is brought
about in only one randomly picked RRAM from the set of
RRAMs of the synapse. This way, the learning-rate is lowered
from
to
, enabling a software-equivalent
learning. Second, we validate our proposal using an
interpolation-model of STDP measurements of a standalone
PrCaMnO-RRAM (accompanying all non-idealities) for
training an SNN with multiple and parallel RRAMs in all its
synapse. Next, the learning performance with is evaluated to
show that
produces excellent peak learning performance
but with significant fluctuations from epoch to epoch, while
is necessary for software-equivalent learning. In
comparison, 256 binary synapses are needed for equivalent
programming (4x improvement). Fourth, architectural
consideration of circuit implementation is then discussed and a
simple circuit to implement the random programming scheme
is presented. Thus, our work presents the specification for
synaptic devices for analog datasets, demonstrates that
challenge for synaptic candidates in literature, and presents an
architectural solution to enable learning and provides a circuit
implementation.
This work is organized as follows: in section 2, we an
overview of the STDP rule, followed by the basis for our claim
of 2% learning-rate as a necessary condition for acceptable
SNN training performance. In section 3 we report the
procedure for STDP demonstration of the PCMO-RRAM and
the results. In continuation, using the STDP-data we show that
the this device fails as synapse. In section 4, we validate our
proposal of using multiple and parallel RRAMs in a synapse.
Lastly, in section 5, we discuss the hardware requirements and
other consideration for adopting the proposed approach.
II. IDEAL LEARNING-RATE FOR SNNS
A. STDP-Overview
Spike-time dependent plasticity (STDP) is the most
common learning-rule used in Spiking Neural Networks
(SNNs). It gives a relation between
(the time gap between
the pre-synaptic and the post-synaptic spikes) and
, the
weight change of the synapses, as illustrated in Figure 1.
Figure 1: Biological STDP rule (after [20])
Figure 2: STDP rule of Eq. 1 with
in Eq. 3
{
Several other time-dependences of weight-change are
possible [21].
model the biological saturation-effects
[20] and ensures that doesn‟t increase/decrease indefinitely.
In Eq. 1,
is positive while
is negative (to strengthen
causality and weaken anti-causality between spikes). As an
example, we have plotted in Figure 2: (1)
a function of
for various (or ) and (2)
as a function of (or ) for
various
with
50ms,
,
and
and:
B. Ideal learning-rate for training
Since
in Eq. 1 (or any STDP equation for that matter)
set the maximum conductance-change, they determine any
network‟s learning-rate. While training an artificial neural
network, the learning-rate needs to be carefully chosen [25].
Since no work exists that studies learning-rate for spiking
neural networks trained using STDP, we empirically determine
the learning-rate using the SNN given in [26]. This single
layered feed-forward SNN can be trained via an exponential
STDP rule to classify data-points of Fisher‟s Iris, Wisconsin‟s
breast-cancer and wine data-sets. For training, following STDP
rule has been used:
For its illustration, we use an exponential weight-update
rule comprising of exponential- dependence term, saturation
factor
and scaling factors ( ), as given in Eq. 2.
|
{
|
|
(
|
)(
(
)
)(
)
Figure 4: Classification accuracy (CA, %) versus learning-rate parameters
and
per cent of
Figure 3: Distribution of classification errors (CE) versus number of
conductance levels per synapse for (a) Iris dataset (b) Wisconsin‟s breastcancer dataset, both trained with 4 linear sensors at input
For large
but small
(~2%), the classification
accuracy settles, but is not the maximum, which
means training is partial
3. For large
but small , the network is unable to
learn
4. For small
, the learning is the best and the
classification accuracy settles
Further, several SNNs trained using STDP in literature rely
on learning-rates less than 1% [27]–[31]. Though no work
focuses on how learning-rate affects the overall training, it is
very likely that the performances may suffer from similar
degradation as learning-rates are increased.
2.
Figure 5: Classification accuracies (CA) for four values of
.
Where,
represents parameters for conductance
potentiation,
represents parameters for conductance
depression,
set the maximum conductance change,
set
the STDP time-constant,
is the maximum conductance
(with 0 as minimum) and is conductance‟s saturation-factor.
To quantitatively study the effect of
on training, we
simulate the training of the network to classify Iris dataset with
various values of the pair { , } and observe the evolution
of classification accuracy as training proceeds.
From Figure 4, which plots the mean-accuracy for last five
epochs of training, it is observed that
1. Lower the learning-rates, better the training in terms of
both maximum accuracy and stability
2. Low synaptic weight‟s depression-rate (~2%) is a
necessary condition for high (more than ~90%) posttraining accuracy
3. Synaptic potentiation rate affects the training
performance, but to a lower extent than depression rate
Figure 4 shows four classification accuracy (CA) plots for
all possible pairs of
, where the percentages
represent fraction of the maximum weight of the synapse,
. Following observations can be made:
1.
For large
(~35%), the training never gets
completed and the classification accuracy doesn‟t
settle
C. Number of bits per digital synapses
To empirically determine the number of bits needed in a
digital synapse for software-equivalent training, we simulated
the training of the network used above to classify Fisher‟s iris
and Wisconsin‟s breast-cancer datasets. The total conductance
range was divided uniformly into
levels. During a
conductance-update event, the originally continuous change in
conductance is discretized to the nearest conductance level.
Figure 3 plots the mean post-training per-cent classification
error (CE) for 10 training experiments over 10 training-epochs.
It is observed that at least 8 bits/256 discrete levels are needed
to ensure lowest CE.
III. PCMO-RRAM AS SYNAPSE IN SNNS
A. STDP of RRAMs using neural write-pulses
RRAM can be approximately modeled as resistor whose
resistance can be changed with voltage-pulses, if the pulsestrength exceeds writing threshold (Figure 6a) [32], [33]. This
change increases, as strength above its threshold is increased.
Given this approximation, to realize STDP on an RRAM,
carefully shaped write-pulses are applied at the two ends of the
synaptic RRAM by the two spiking neurons in context (Figure
6b) [19], [34]. These pulses are so shaped that when two such
pulses, corresponding to each of the neurons attached to a
synapse, relatively displaced in time, are subtracted,
1.
There exists a portion in the net that always just
reaches (or, goes above) the RRAM‟s threshold
2. the height of the net above the RRAM‟s threshold
(or compactly, the overdrive) is a function of the
relative displacement
Each time a neuron spikes, such pulses are applied
immediately in response at the terminal of the synaptic RRAM.
This way, the RRAM sees a net-voltage equal to the
(
N
Postsynaptic
neuron
Presynaptic
neuron
(0,0)
(a)
)
(
)
Here,
(b)
Vw
Δt
t
t
1.
and
are the positive and negative
amplitudes.
and
are set such that they are
less than
and
, respectively
2.
and
are the decay constants.
sets the
write-time of the pulse and sets the exponential
STDP time-constant
(0,0)
3.
(c)
)
(
{
ΔG
)
(
N
(d)
Figure 6: (a) RRAM conductance-change (ΔG) modeled as a function of
rectangular pulse of height VW, (b) Pre-synaptic and post-synaptic neuron
connected through a synapse. Right after each of their spiking instants,
each produce and apply a fresh write-pulse at the RRAM‟s terminal (c) A
possible synaptic write-pulse for exponential STDP, (d) Subtraction of
two relatively displace (
) write-pulses or voltage across RRAM.
The circled portion of the subtraction, crossing the threshold, changes the
conductance of the RRAM
and
define the spike lengths in the time axis
B. Demonstration of exponential STDP of PCMO
PCMO-RRAMs have been experimentally demonstrated as
endurable, fast and highly scalable non-volatile analog
memristive contender for synaptic applications [8], [35]–[37].
PCMO-RRAM with cross-section shown in Figure 7 and an
area of
, originally reported in reported in [37]
was used for the demonstrating STDP. The device was
initialized to its low-resistance state by applying a long-lasting
and constant negative voltage-pulse of –2.5 V and compliance
set to 10 mA. For writing, pulses in Eq. 2 were used and the
values for ‟s, ‟s and A‟s are given in Table 1. The procedure
followed to demonstrate the STDP of the device is:
1) the conductance was read using a small
rectangular voltage pulse (0.5V), yielding the
initial conductance value
Figure 7: Measured PrCaMnO-RRAM‟s structure
3) the final conductance value
using a 0.5 V voltage pulse
TABLE I
WRITE-PULSE PARAMETERS USED FOR STDP
Writepulse
Parameter
Value
2) a was randomly chosen from [-100ms, 100ms]
and the subtraction of relatively displaced writepulses corresponding to
was directly applied to
terminal 1 and with terminal 2 grounded
Parameter
Value
1V
-1V
1 s
100ms
0.5 s
50ms
1V
-1V
1 s
100ms
0.5 s
50ms
was recorded by
These three steps were repeated 1000 times for the new
state
that served as
for the new iteration. Of the pulses
applied, the ones leading to increase in conductance have been
plotted in green/Δ in Figure 8and those leading to decrease, in
blue/∇. The conductance values have been normalized by
dividing each value with the maximum conductance observed.
For a better visual, uniformly spaced iso-initial-conductance
(iso- ) points have been plotted in Figure 9 and iso-time-
subtraction of the displaced pulses applied by the pre- and the
post-synaptic neurons (Figure 6d). For implementing an
exponential STDP rule, the write-pulse may be given the
following shape:
Figure 8: LTP (green/Δ) and LTD (blue/∇ sc tt r lots
Figure 9: (a) Observed STDP ( G) of the PCMO-RRAM‟s conductance
with various iso-initial-conductance ( ) in distinct colors. (b) Interpolation
model of measurements.
in the legend represents the initial conductance
Figure 11: Training behaviour with single RRAM based synapse. (a)
Regardless of the initial condition post-training classification accuracies
(CA) settle to low values right after first training iteration (epoch); (b)
Conductance‟s evolution with initial condition (IC)-1; (c) Conductance‟s
evolution with IC-2.
Figure 12: Learning-rate ( or ) offered by various analog RRAMs in
literature. None of them meet our specifications
Figure 10: (a) Observed STDP ( G) of the PCMO-RRAM‟s conductance
with various iso-spike-time-difference (iso-Δt) in distinct colors. (b)
Interpolation model of measurements in (a)
difference (iso) in Figure 10. To use this data for
simulations, we used its interpolation model. Iso- and isoSTDP curves obtained using such a model are plotted in Figure
9b and Figure 10b. Though the STDP demonstrated
experimentally has a time-constant of 50ms, it can be altered
by scaling the time-keeping portion of the write-pulses.
C. Training with Single PCMO-RRAM as a synapse
We replaced the mathematical synaptic model in the
network mentioned in Sec. II with the interpolation model of
PCMO-RRAM. To do so, we modified the:
1. Read equation: Each read-instance of
, the
mathematical conductance, was replaced with
the RRAM‟s conductance. Since the latter was
normalized, an additional factor of
was added,
leading to the following replacement:
2. STDP rule equation: The STDP rule specified in Eq. 3
was replaced with following equation:
(
Where,
is determined
interpolation model mentioned above.
)
from
the
Next, the network‟s training was simulated. Figure 10 plots
the evolution of classification accuracy as training proceeds.
Clearly, the performance is worse in comparison to that
obtained with a mathematical synaptic model.
Our hypothesis, based on observation made from Figure 4
and Figure 11, is that such marked reduction in performance is
observed because of large percentage change in conductance of
PCMO-RRAM. This is equivalent to having large learningrates in a mathematical model.
STDP has been demonstrated on several analog RRAMs
[4], [10]–[17]. As observed above, it is necessary for the
maximum conductance change to be less than 2% of the
synaptic conductance-range to get software-equivalent
evolution of conductance. However, all analog conductance‟s
plasticity demonstrations up until now have a maximum
| |
conductance change (
) of more than 20% and
can go up to 400% in some devices (Figure 12). Thus, all
currently existing analog RRAMs will fail to produce softwareequivalent post-training classification accuracy for our
network.
Our second hypothesis, based on observation made from
Figure 6 and 7, is that training performance can improved
without changing the network, by reducing the maximum
| or
change in conductance change (i.e., lower |
|
| in Eq. 3). The validity of this hypothesis can
be tested by using an STDP-data like we used for an RRAM
with a lower maximum conductance. Since in our knowledge,
all RRAMs in the literature do not meet this constraint, this is
not possible for us at the moment. To test our hypothesis and as
a step towards better memristor based synaptic models, we
propose using a set of n PCMO-RRAMs in parallel in a way
that
1. they function aggregately as synapse
2. the conductance-change pulses are applied to only one
of n-RRAMs
Next, we validate our proposal.
IV. PARALLEL AND MULTIPLE PCMOS AS SYNAPSE
To test the ability of multiple (n) PCMO-RRAMs to
collectively act as a synapse, we continued with same network
and trained it to classify Iris data-set. The mapping from a
mathematical synapse to an n-RRAM based synapse is similar
to the one described in Sec. 3C, with a slight difference,
described as follows:
1. Each read-instance of
replaced with:
, the mathematical weight, was
Figure 13: Classification accuracies for 2, 16 and 100 PCMO-RRAMs in
the synapse, with same initial conductance configuration
∑
2. STDP rule equation: For every spiking instant, STDPrule based weight-update is replaced with the following
two equations:
(
)
Note that RRAM‟s conductance increase corresponds to
increase in the conductance of the synapse of which it‟s a part.
We let take all values from the set {2, 4, 16, 36, 64, 100}.
Figure 13 plots the CA, as training proceeds for n=2, 16 and
100. It is observed that learning is more stable for synapses
with more RRAMs. Figure 13 plots the CA‟s quantiles for all
n‟s, as the training progresses with number of training-epochs
in Figure 14. The number of training-epochs was determined
empirically. For n=100, approximately 50 epochs were needed
to stabilize learning. Since learning-rate is roughly inversely
proportional to the number of RRAM, the number of training
iterations needed is proportionally increased/decreased. For
very large learning-rates (n=2, 4 and 16), a lower limit of 20
was set. The evolution conductance of each synapse for various
‟s has been plotted in Figure 15. It is observed that:
1. starting from n=4, the network‟s CA reaches softwaremaximum of 97.3% at least once, while training (Figure
13)
2. despite being trained for more than adequate epochs,
network with synapses having very low number of
RRAMs are unable learn stably (Figure 13 and Figure
14a)
3. as n is increased, the CA distribution in Figure 14a
follows a trend similar to the one exhibited by Figure 5
showing the increase in CA as LTP and LTD rates are
decreased simultaneously – higher mean and lower
variation
4. conductance evolve more smoothly as
~~
increases
V. DISCUSSION
The scheme discussed does not escalate the circuit
requirements. During reading, all RRAMs in a synapse need to
be simultaneously read from. This is done by applying same
reading pulse-voltage to all row-bars of the synapse. The
current from each of the branch associated to a synapse is then
summed up, to get a current proportional to the synaptic
weight.
Figure 14: (a) Classification-accuracy (CA) quantiles as the number of
PCMO-RRAMs in the synapse is varied. (b) Number of epochs for training
Figure 15: Synaptic weight evolution smoothens as more PCMO-RRAMs
are added within the synapse
For performing the writing operation, one RRAM is
randomly chosen from
RRAMs within a synapse, with a
uniform probability for each. This is done by applying the presynaptic write-pulse to a random row among
row-bars and
the post-synaptic write-pulse to a random column among
column-bars (for a synapse with
RRAM rows and
RRAM columns). This way, the RRAM selection is uniformly
random among all
RRAMs. Though there can be other
schemes for uniformly selecting RRAMs, this in our opinion,
shouldn‟t require complicated and/or large area peripheral
circuits.
The reading and writing phases may be multiplexed in time
using a global control signal
(reading indicated by
and writing by
). During the writing phase, to
allow random selection of a row/column, a set of global onehot selection-lines is laid out along the periphery of the array.
The active line in this set is changed, periodically. Whenever a
neuron ( ) spikes (assumed to be random in time), the content
of the selection line set is copied onto the RRAM selection
register with output vector ̅ (Figure 17). This way, the same
row/column of the synapse remains selected until the next
spike, as the active line of the synapse‟s selection- register
reading pulse is applied on pre-synaptic terminals and postsynaptic terminals grounded.
W
R
W
R
Σ
W
One-hot
select
Σ
(a)
W
(b)
Figure 16: (a) Read process consists of applying read pulses (R, in
green) to all rows of a synapse and summing up current at the array‟s
virtually grounded output; (b) Weight update process consists of application
of write-pulses (generated by W, blue) to only one randomly chosen row
and column of the synapse. Choice of the row/column is determined by
global one-hot select lines that are read at the spiking instant
Clock
Global
one-hot
vector
Ni
Selection
latch – i
Si
Use of multiple RRAMs within a synapse clearly requires a
bigger cross-bar array. However, a larger cross-bar implies a
larger attenuation in the voltage applied across RRAMs at
cross-points far away from the either input and output sides of
the array. Thus, for a constant wire-resistance, the cross-bar
array size, or more fundamentally, the number of RRAMs in
the synapses cannot be made arbitrarily large. Not only their
number, but also their arrangement needs to be carefully
designed to maintain a certain minimum level of read-write
fidelity. For simplicity, consider an SNN with just one nRRAM synapse and all RRAMs within it in same conductance
state. If the potential-drop due to wire-resistance is assumed to
be a linear function of cross-point index within a synapse, then
the read current-error of a cross-point
within the array
can be expressed as:
For a synapse with
RRAMs, the maximum error in
current will be for the cross-point at the corner furthest from
the inputs and outputs, i.e., the one with indices
. For
fixed number of RRAMs in a synapse, say
, the maximum
error in read current will be
Pr
R(1)/W(0)
Figure 17: At every rising edge of neural spike, selection-latch copies
content of the global one-hot vector. For each neuron, there is a unique
selection latch
Si
R
1
W
0
Sj
R(1)/W(0)
1
0
W
Figure 18: During writing phase (RW=0), MOSFET-based switches
select the RRAM being written onto. During reading phase (RW=1), all
RRAMs are accessed by turning on all switches.
remains same. During the reading phase, the selection register
is over-written using the pre-set control input so that ̅
.
Write-pulses may be applied to a single row (column) out
of M1 (M2) via MOSFET based switches as shown in Figure
18. Each of the MOSFETs‟ gate is connected to the one-hot
selection vector . Thus, only one MOSFET out of M1 (M2)
is conducting and will allow write-pulses to be applied to the
corresponding row (column). However, during the reading
phase (
), all MOSFETs are turned on by setting
̅
. Thus, for all RRAMs within a synapse, same
This happens when the synapse has an arrangement of
cross-points in configuration closest to that of a square. Thus,
within a synapse, the RRAMs should be arranged in a squarelike configuration.
VI. CONCLUSION
In this work, we introduce and empirically determine two
new specifications for an SNN based synapse: number of
conductance levels per synapse and maximum learning-rate.
To the best of our knowledge, there are no RRAMs that meet
the latter specification. As a solution, we proposed the use of
multiple RRAMs in parallel within a synapse. While synaptic
reading, all RRAMs are simultaneously read and for each
synaptic conductance-change event, the writing pulses for
STDP are applied on only one RRAM, randomly picked from
the set. To validate our solution, we experimentally
demonstrated STDP of conductance of a PCMO-RRAM and
showed that due to large learning-rate, a single device fails to
model a synapse in the training of an SNN. As anticipated,
network training improved as more RRAMs are added to the
synapse. Lastly, we discuss the circuit-requirements for
implementing such a scheme, to conclude that the requirements
are within bounds.
REFERENCES
[1] G. Indiveri, F. Corradi, and N. Qiao, “Neuromorphic Architectures for
Spiking Deep Neural Networks,” pp. 68–71, 2015.
[2] G. Indiveri, E. Chicca, and R. Douglas, “A VLSI array of low-power
spiking neurons and bistable synapses with spike-timing dependent plasticity,”
IEEE Trans. Neural Networks, vol. 17, no. 1, pp. 211–221, 2006.
[3]
Wikipedia,
“Chemical
synapse.”
https://en.wikipedia.org/wiki/Chemical_synapse.
[Online].
Available:
[4] S. Yu, D. Kuzum, and H. S. P. Wong, “Design considerations of synaptic
device for neuromorphic computing,” Proc. - IEEE Int. Symp. Circuits Syst.,
pp. 1062–1065, 2014.
type.,” J. Neurosci., vol. 18, no. 24, pp. 10464–10472, 1998.
[5] S. B. Eryilmaz, D. Kuzum, S. Yu, and H.-S. P. Wong, “Device and system
level design considerations for analog-non-volatile-memory based
neuromorphic architectures,” in 2015 IEEE International Electron Devices
Meeting (IEDM), 2015, p. 4.1.1-4.1.4.
[22] U. E. Paulista, M. Filho, N. Prof, F. Breve, and I. Esta, “Iris Data Set,”
2013. [Online]. Available: https://archive.ics.uci.edu/ml/datasets/iris.
[6] B. L. Jackson et al., “Nanoscale electronic synapses using phase change
devices,” J. Emerg. Technol. Comput. Syst., vol. 9, no. 2, p. 12:1–12:20, 2013.
[7] M. Suri et al., “Phase change memory as synapse for ultra-dense
neuromorphic systems: Application to complex visual pattern extraction,” 2011
Int. Electron Devices Meet., p. 4.4.1-4.4.4, 2011.
[8] S. Park et al., “RRAM-based synapse for neuromorphic system with pattern
recognition function,” Tech. Dig. - Int. Electron Devices Meet. IEDM, pp. 231–
234, 2012.
[9] D. Garbin et al., “HfO2-Based OxRAM Devices as Synapses for
Convolutional Neural Networks,” IEEE Trans. Electron Devices, vol. 62, no. 8,
pp. 2494–2501, Aug. 2015.
[10] M. Prezioso, F. Merrikh Bayat, B. Hoskins, K. Likharev, and D. Strukov,
“Self-Adaptive Spike-Time-Dependent Plasticity of Metal-Oxide Memristors,”
Sci. Rep., vol. 6, no. 1, p. 21331, Aug. 2016.
[11] Z. Wang et al., “Memristors with diffusive dynamics as synaptic
emulators for neuromorphic computing,” Nat. Mater., vol. 16, no. 1, pp. 101–
108, 2017.
[12] Y.-F. Wang, Y.-C. Lin, I.-T. Wang, T.-P. Lin, and T.-H. Hou,
“Characterization and Modeling of Nonfilamentary Ta/TaOx/TiO2/Ti Analog
Synaptic Device,” Sci. Rep., vol. 5, no. 1, p. 10150, Sep. 2015.
[13] Z. Wang, S. Ambrogio, S. Balatti, and D. Ielmini, “A 2-transistor/1resistor artificial synapse capable of communication and stochastic learning in
neuromorphic systems,” Front. Neurosci., vol. 9, no. JAN, pp. 1–11, 2015.
[14] S. H. Jo, T. Chang, I. Ebong, B. B. Bhadviya, P. Mazumder, and W. Lu,
“Nanoscale memristor device as synapse in neuromorphic systems,” Nano
Lett., vol. 10, no. 4, pp. 1297–1301, 2010.
[15] N. Panwar, D. Kumar, N. K. Upadhyay, P. Arya, U. Ganguly, and B.
Rajendran, “Memristive synaptic plasticity in Pr0.7Ca0.3MnO 3 RRAM by
bio-mimetic programming,” Device Res. Conf. - Conf. Dig. DRC, no. June
2012, pp. 135–136, Jun. 2014.
[16] Y. Li et al., “Activity-Dependent Synaptic Plasticity of a Chalcogenide
Electronic Synapse for Neuromorphic Systems,” Sci. Rep., vol. 4, no. 1, p.
4906, May 2015.
[17] Y. Li et al., “Ultrafast Synaptic Events in a Chalcogenide Memristor,” Sci.
Rep., vol. 3, no. 1, p. 1619, Dec. 2013.
[18] G. W. Burr et al., “Neuromorphic computing using non-volatile memory,”
Adv. Phys. X, vol. 2, no. 1, pp. 89–124, Jan. 2017.
[19] B. Rajendran et al., “Specifications of Nanoscale Devices and Circuits for
Neuromorphic Computational Systems,” IEEE Trans. Electron Devices, vol.
60, no. 1, pp. 246–253, Jan. 2013.
[20] G. Q. Bi and M. M. Poo, “Synaptic modifications in cultured hippocampal
neurons: dependence on spike timing, synaptic strength, and postsynaptic cell
[21] L. F. Abbott and S. B. Nelson, “Synaptic plasticity: taming the beast,” Nat.
Neurosci., vol. 3 Suppl, no. Supp, pp. 1178–83, Nov. 2000.
[23] D. W. et al H, “Breast Cancer Wisconsin Dataset.”
[24]
P.
Forina,
M.
et
al,
“Wine
Data
Set,”
https://archive.ics.uci.edu/ml/datasets/Wine, 2009. [Online]. Available:
https://archive.ics.uci.edu/ml/datasets/wine.
[25] D. R. Wilson and T. R. Martinez, “The need for small learning rates on
large problems,” in IJCNN’01. International Joint Conference on Neural
Networks. Proceedings (Cat. No.01CH37222), 2001, vol. 1, no. 1, pp. 115–
119.
[26] A. Biswas, S. Prasad, U. Ganguly, S. Lashkare, and U. Ganguly, “A
simple and efficient SNN and its performance & robustness evaluation method
to enable hardware implementation,” Dec. 2016.
[27] P. U. Diehl and M. Cook, “Unsupervised learning of digit recognition
using spike-timing-dependent plasticity,” Front. Comput. Neurosci., vol. 9, no.
August, pp. 1–9, 2015.
[28] S. Song and L. F. Abbott, “Cortical Development and Remapping through
Spike Timing-Dependent Plasticity,” Neuron, vol. 32, no. 2, pp. 339–350, Oct.
2001.
[29] S. Song, K. D. Miller, and L. F. Abbott, “Competitive Hebbian learning
through spike-timing-dependent synapticplasticity,” Nat. Neurosci., vol. 3, no.
9, pp. 919–926, Sep. 2000.
[30] Q. Yu, H. Tang, K. C. Tan, and H. Yu, “A brain-inspired spiking neural
network model with temporal encoding and learning,” Neurocomputing, vol.
138, pp. 3–13, 2014.
[31] S. R. Kheradpisheh, M. Ganjtabesh, S. J. Thorpe, and T. Masquelier,
“STDP-based spiking deep convolutional neural networks for object
recognition,” 2016.
[32] Y. V. Pershin and M. Di Ventra, “SPICE model of memristive devices
with threshold,” vol. 22, no. 2, pp. 485–489, Apr. 2012.
[33] T. Serrano-Gotarredona, T. Masquelier, T. Prodromakis, G. Indiveri, and
B. Linares-Barranco, “STDP and sTDP variations with memristors for spiking
neuromorphic learning systems,” Front. Neurosci., vol. 7, no. 7 FEB, pp. 1–15,
2013.
[34] B. L. Jackson, D. S. Modha, and B. Rajendran, “Producing spike-timing
dependent plasticity in an ultra-dense synapse cross-bar array,” 23-Jun-2011.
[35] Dong-Jun Seong et al., “Resistive-Switching Characteristics of Al/PCMO
for Nonvolatile Memory Applications,” IEEE Electron Device Lett., vol. 30,
no. 9, pp. 919–921, Sep. 2009.
[36] S. Jung et al., “Thermally-assisted Ti/PCMO ReRAM with excellent
switching speed and retention characteristics,” in 2011 International Electron
Devices Meeting, 2011, p. 3.6.1-3.6.4.
[37] P. Kumbhare, I. Chakraborty, A. K. Singh, S. Chouhan, N. Panwar, and U.
Ganguly, “A selectorless RRAM with record memory window and nonlinearity
based on trap filled limit mechanism,” 2015 15th Non-Volatile Mem. Technol.
Symp. NVMTS 2015, vol. 1, pp. 15–17, 2016.
| 9cs.NE
|
Classification for Dynamical Systems:
Model-based Approach and Support Vector Machines
arXiv:1803.10552v1 [cs.SY] 28 Mar 2018
Giorgio Battistelli and Pietro Tesi
Abstract— We consider the problem of classifying trajectories
generated by dynamical systems. We investigate a model-based
approach, the common approach in control engineering, and
a data-driven approach based on Support Vector Machines, a
popular method in the area of machine learning. The analysis
points out connections between the two approaches and their
relative merits.
I. I NTRODUCTION
Assume that we are given two dynamical systems, whose
underlying dynamics might be unknown. We are interested
in designing a classifier, that is a machine which, given an
observed trajectory generated by either the systems, correctly
identifies which of the candidate systems has generated such
a trajectory. We will refer to this problem as the problem of
classification for dynamical systems.
Problems of this type are ubiquitous within the community
of systems and control. For instance, in fault detection one
system may represent the behavior in nominal conditions and
another a prescribed faulty behavior. Similarly, in networked
control one system may represent the closed-loop behavior
when data are successfully transmitted, while another system
may represent the open-loop behavior in the presence of
packet dropouts. Problems of this type naturally arise also
when dealing with systems that naturally exhibit multiple
operating modes, for example switched circuits. On the
other hand, in computer science, classification is the core of
machine learning for pattern recognition [1], [2], which has
found several applications in different fields of engineering,
including applications that are intensively studied in control
engineering like fault detection and diagnosis [3]. Yet, with
regard to this problem, the interaction between these two
communities has been low.
Within the systems and control community, classification
for dynamical systems has been studied in connection with
the analysis of switched / multi-mode systems [4]-[9], often
under the term mode-identification, which is defined as the
problem of reconstructing the active mode of a switched
system from its output trajectories. The common approach
is a model-based approach: assuming a correct model of
the system for each operating mode, one can check whether
or not mode-identification is feasible via dynamic-dependent
conditions, and mode-identifiers (in fact, classifiers) can be
obtained in terms of rank tests, least-square functions or
dynamical systems. While the theory also generalizes to
noisy observations [9], [10] and to some types of nonlinear
dynamics [11], [12], little is known on how to approach
The authors are with DINFO, University of Florence, 50139 Florence,
Italy {giorgio.battistelli,pietro.tesi}@unifi.it.
model-based classification if one departs from the hypothesis
that the dynamics of the system are known with perfect
accuracy. Even in the simplest case where the dynamics are
associated with parametric uncertainty, building a classifier
is a non-trivial task. The difficulty is similar to the one
encountered in adaptive control based on multiple models,
where a main issue is indeed to guarantee that modelling
inaccuracies do not destroy the learning capability of the
control scheme [13].
In computer science, the main paradigm to classification
is instead the data-driven paradigm. Classifiers are designed
by choosing a function with adjustable parameters selected
using a number of training data, called the examples. The
resulting function (the classifier) is then evaluated according
to its capability to generalize from the training dataset, that
is to correctly map new examples. Popular methods are
Neural Networks (NN) and Support Vector Machines (SVM),
whose capability to generalize from a training dataset can
be quantified via suitable loss functions such as the risk
function [1]. Data-driven methods have the intrinsic potential
to overcome issues related to model uncertainty, and have
already proven their effectiveness in challenging applications
such as the prediction of epileptic seizures from recording
of EEG signals [14]. However, it is not obvious how to
tailor the analysis and design of data-driven methods to the
specific context where data come from dynamical systems.
It is worth pointing out that classification of data generated
from dynamical systems is not new in computer science. In
fact, it can be regarded as classification of time series once
we assume the existence of an underlying data-generating
system. Yet, approaches which take this standpoint still try
to incorporate models into the learning task, either to extract
from data informative features [15] or to construct suitable
kernel functions [16], [17]. While incorporating models is a
natural step to take, it leads to the previous question of how
to handle model uncertainty, and does not help to understand
the performance achievable by model-free schemes. Similar
issues related to model-based approaches have been pointed
out also in the context of clustering [18].
In this paper, we consider autonomous linear systems and
approach the classification problem from both model-based
and data-driven perspectives, pointing out relative merits and
establishing connections between the two. We first consider
a model-based approach and derive a classifier assuming the
knowledge of the system dynamics. This approach has two
fundamental merits: i) to highlight necessary conditions for
the existence of a correct classifier (problem feasibility); ii)
to guide the design and analysis of a data-driven solution.
In connection with ii), the model-based approach shows that
under problem feasibility one can design a correct classifier
which can be interpreted in terms of polynomial kernels [2].
Building on this result, we consider a data-driven approach
based on SVM. By using properties stemming from the
model-based solution, we provide bounds on the margin of
the classifier and quantify its generalization performance [19]
as a function of the systems one wishes to classify.
The rest of this paper is as follows. Sections II and III
formalize the problem of interest, and recall basic concepts
regarding SVM. In Sections IV and V, we present the main
results. Section VI discusses the results and open problems.
Numerical simulations are reported in Section VII, while
Section VIII provides concluding remarks.
II. F RAMEWORK
Consider two linear dynamical systems
(
xi (t + 1) = Ai xi (t)
Σi ∼
, i = 1, 2
yi (t) = Ci xi (t)
A. Limitations of the classification problem
(1)
where t ∈ Z+ := {0, 1, . . .} denotes time; xi ∈ Rni is
the state; yi ∈ Rm is the output; Ai and Ci are state and
output transition matrices. We will assume that each Σi is
observable (in a control-theoretic sense). This entails no loss
of generality in that if Σi is not observable, all subsequent
developments apply to the observable subsystem obtained via
a Kalman observability decomposition. Assume now that we
are given a sequence
Y := col(y(0), y(1), . . . , y(N − 1))
(2)
of N measurements generated by one of the two systems, that
is Y = col(yi (0), yi (1), . . . , yi (N − 1)) with i ∈ {1, 2}, but
we have no direct information on which of the two systems
has generated Y . We are interested in determining which of
the two systems has generated Y , referring to this problem
as the problem of classification.
To make the problem definition precise, let
Ci
Ci Ai
(3)
Oi :=
..
.
Ci (Ai )N −1
be the observability matrix of order N of the pair (Ci , Ai ),
and let
Li := {Oi x, x ∈ Rni : Oi x 6= 0}
The problem of interest is to construct correct classifiers.
We will investigate two approaches:
(i) Model-based classification: The classifier depends on
the knowledge of the matrices Ai and Ci , i = 1, 2.
(ii) Data-driven (model-free) classification: The classifier
does not depend on the knowledge of the matrices Ai
and Ci , i = 1, 2, and has to be determined on the basis
of a given number of sample trajectories, that is points
in the sets L1 and L2 .
Case (i) reflects the situation where the dynamics of the
systems are known and this information is exploited in the
design of the classifier. On the contrary, case (ii) reflects the
situation where the dynamics of the systems are unknown or
this information is not directly exploited in the design of the
classifier.
(4)
be the set of all possible nonzero trajectories of N samples
that can be generated by the i-th system.
Definition 1 (Classifiers and correctness): A classifier is
any function f : Y → R, where Y is the space of the input
data. A classifier for the dynamical systems in (1) is said to
be correct if it satisfies:
(
>0
if Y ∈ L1
f (Y )
(5)
<0
if Y ∈ L2
Clearly, the knowledge of the system dynamics provides
an extra degree of information that can be used to properly
design a classifier. Yet, there are certain limitations which
cannot be overcome even in the ideal situation where one
has perfect knowledge of the dynamics. In particular, the
following result holds true.
Theorem 1 (Limitations of the classification problem):
A correct classifier for the dynamical systems in (1) exists
only if rank [ O1 O2 ] = n1 + n2 .
Proof of Theorem 1. By definition, there exists no correct
classifier whenever L1 ∩ L2 6= ∅, because this implies the
existence of trajectories compatible with both the systems.
This is equivalent to the fact that the dynamical system
resulting from the parallel interconnection of Σ1 and Σ2 is
observable, that is that the observability matrix of order N
of the pair (C, A) with
A1
0
C1 C2
(6)
, C =
A =
0
A2
has column rank n1 + n2 . This gives the result.
The condition in Theorem 1 can be satisfied only if the two
systems do not share common eigenvalue-eigenvector pairs.
This condition also requires N m ≥ n1 + n2 , which means
that a large enough observation window must be chosen to
render classification feasible.
We will take this condition as a standing assumption.
Assumption 1. rank [ O1 O2 ] = n1 + n2 .
III. S UPPORT V ECTOR M ACHINES
In this section, we briefly recall some concepts on SVM
focusing on the case of separable data. This material of this
section is adapted from [2].
Assume we have L observations (Yk , ℓk ), k = 1, 2, . . . , L,
each one consisting of a vector Yk ∈ Rd plus a label
ℓk ∈ {−1, 1} specifying the class to which Yk belongs. In
connection with the problem introduced in Section II, one
can think of (Yk , ℓk ) as an observation collected from one of
the two candidate systems Σi , where Yk is the measurement
and ℓk specifies which of the two systems has generated Yk .
Consider the problem of classifying the vectors Yk using
hyperplanes H(Y, α) = {Y ∈ Rd : w⊤ Y + b = 0}, where
α = (w, b) is a vector of adjustable weights. If there exists
a vector α satisfying
w⊤ Yk + b > 0 if ℓk = 1
(7)
w⊤ Yk + b < 0 if ℓk = −1
for k = 1, 2, . . . , L, then the vectors Yk are called linearly
separable, and the function
f (Y, α) = w⊤ Y + b,
α = (w, b)
(8)
defines a linear classifier which is correct with respect to the
data (Yk , ℓk ), k = 1, 2, . . . , L.
For the linearly separable case, an SVM searches for
the separating hyperplane with largest margin ρ, that is it
searches for the value of α which maximizes
|w⊤ Yk + b|
(9)
ρ := min
k=1,2,...,L
kwk
This can be cast as a convex program:
1
kwk2
2
w⊤ Yk + b ≥ 1
subject to
w⊤ Yk + b ≤ −1
min
α
if ℓk = 1
(10)
if ℓk = −1
The reason to search for the hyperplane with largest margin
is related to the fact that f (Y, α) is obtained from a finite set
of observations, the so-called training set. On the other hand,
one would like f (Y, α) to be able to correctly classify also
data which are not present in the training set. This property is
usually called the generalization performance [2], and SVM
can guarantee a good generalization performance. We will
discuss this point in more detail in Section V.
Problem (10) involves L constraints and d + 1 unknowns.
When d > L it can be more convenient to resort to a dual
formulation of the problem, called Wolfe dual:
1
max µ⊤ 1 − µ⊤ Zµ
µ
2
X
(11)
subject to µ 0,
µk ℓ k = 0
k=1,2,...,L
where µ := col(µ1 , µ2 , . . . , µL ) is the vector of Lagrange
multipliers, Z = [Zkj ] is a symmetric L × L matrix such
that Zkj = ℓk ℓj Yk⊤ Yj , k, j = 1, 2, . . . , L, and where 1 is
the vector of ones. Problem (11) involves L constraints and
unknowns. The solution has the form
X
w=
µk ℓk Yk
(12)
k=1,2,...,L
and each vector Yk associated to a positive multiplier µk
is a support vector. This means that for the optimal linear
classifier resulting from (10) the parameter w is given by a
linear combination of the support vectors, which can be then
interpreted as the most representative points in the training
dataset.
A. Kernel functions
Finding a separating surface which is linear with respect
to the space Y of the input data is not always possible. One
of the most important results about SVM is related to the
possibility of finding non-linear separating surfaces in a very
straightforward manner.
By looking at the optimization problem (11), one sees that
the data appears only through the products Yk⊤ Yj . One can
think of mapping the input space Y into a higher-dimensional
space H through a function Φ : Y → H, and search for a
function κ : Y × Y → R such that
κ(Y, Z) = hΦ(Y ), Φ(Z)i,
∀ X, Z ∈ Y
(13)
The space H is called the feature space, while Φ(Y ) is called
the feature vector. Any function κ satisfying (13) is called a
kernel function. Kernel functions define separating surfaces
which are linear with respect to H. The remarkable feature
of kernel functions is that there is no need to use or know
the function Φ in order to compute or use w. In fact, in order
to compute the solution of (11) with respect to Φ one can
simply use Zkj = ℓk ℓj κ(Yk , Yj ). Moreover,
X
w⊤ Φ(Y ) =
µk ℓk κ(Yk , Y )
(14)
k=1,2,...,L
Thus one can use κ instead of Φ also for the classification
task. Kernel functions are also advantageous from the point
of view of computations since κ operates in the input space
Y, which has usually lower dimension than H. Common
kernel functions are polynomial, Gaussian and hyperbolic
tangent functions [2].
In the sequel, we will show that for classifying dynamical
systems polynomial kernels are good candidates.
IV. M ODEL - BASED C LASSIFICATION
We now consider a model-based approach to classification.
The following example shows that no correct classifier exists
which is linear in the input space.
Example 1. Consider two systems as in (1), where A1 = 1,
A2 = −1 and C1 = C2 = 1. Assumption 1 clearly holds true
for N ≥ 2. However, as depicted in Figure 1, there exists
no linear classifier for the two candidate systems, and this is
independent of the particular choice of N .
Example 1 indicates that for dynamical systems there is no
correct classifier which is linear with respect to Y . However,
Figure 1 suggests that a correct classifier for Example 1 exists
and is given by f (Y ) = y(0)y(1), which can be rewritten as
f (Y ) = w⊤ Φ, where
0
1
1
, Φ = Y ⊗Y
(15)
w⊤ =
2 1
0
where ⊗ stands for Kronecker product. We will see that Φ
defines a polynomial kernel. Before doing this, we show that
the choice Φ = Y ⊗ Y is general in the sense that it applies
to any linear dynamical system.
y(1)
A. Form of the model-based classifier: Kernel function and
support vectors interpretation
y(0)
0
f (Y )
Fig. 1: Left: Pictorial representation of the possible observation points for
Example 1 when N = 2. The trajectories that can be generated by the
first system correspond to points (red circles) which always falls in the first
or third quadrant of the Cartesian plane, while the trajectories that can be
generated by the second system correspond to points (blue circles) which
always falls in the second or fourth quadrant of the Cartesian plane. Right:
Pictorial representation of f (Y ) := y(0)y(1).
Let Gi := Oi⊤ Oi , i = 1, 2, be the observability Gramian
corresponding to the i-th system. Note that Gi is nonsingular
under Assumption 1. Also, let
Q := Q1 − Q2
(16)
where
Qi := Oi Gi−1 Oi⊤ ,
i = 1, 2
(17)
The following result holds true.
Theorem 2 (Model-based classifier): Let Φ = Y ⊗ Y and
let wM = vec(Q), where vec(·) is the vectorization operator.
⊤
Under Assumption 1, f (Y ) = wM
Φ is a correct classifier
for the dynamical systems in (1).
Proof of Theorem 2. The idea is to show that computing
⊤
wM
Φ is equivalent to determining which of the sets Li the
vector Y belongs to. Consider the point-set distance
πi (Y ) := minn kOi x − Y k2 ,
x∈R
i = 1, 2
(18)
Notice that if Y ∈ L1 then π1 (Y ) = 0 and π2 (Y ) > 0 in
view of Assumption 1. Likewise, if Y ∈ L2 then π1 (Y ) > 0
and π2 (Y ) = 0. Hence, the function
g(Y ) := π2 (Y ) − π1 (Y )
(19)
Theorem 2 could have been stated directly in terms of
f (Y ) = Y ⊤ QY . Yet, the form f (Y ) = w⊤ Φ turns out to be
useful because it provides guidelines for the formulation of
the data-driven approach. In fact, it guarantees the existence
of a solution for the SVM formulation if we use Φ = Y ⊗ Y
as input to the training algorithm. Moreover, it is immediate
to verify that
Φ⊤ Φ = (Y ⊤ Y )2
that is Φ defines a homogeneous polynomial kernel. This
means that in the SVM formulation one can work directly in
the space of Y by emplyoing the kernel κ(Y, Z) = (Y ⊤ Z)2 .
This option is possible also for the model-based solution if
we write wM in terms of support vectors. This interpretation
is simple and worth mentioning.
A (reduced) singular value decomposition of the matrix Oi
yields Oi = Ui Si Vi⊤ , where Ui ∈ RN m×ni has orthonormal
columns, Si ∈ Rni ×ni is a diagonal matrix with positive
entries (due to Assumption 1), and Vi ∈ Rni ×ni is unitary.
Thus we have Qi = Ui Ui⊤ . Let Yi,k be the k-th column of
Ui and let Φi,k = Yi,k ⊗ Yi,k be the corresponding feature
vector. Hence,
ni
X
Φi,k =
k=1
=
g(Y ) = Y ⊤ (I − Q2 )Y − Y ⊤ (I − Q1 )Y
⊤
k=1
ni
X
Yi,k ⊗ Yi,k
⊤
vec(Yi,k Yi,k
)
= vec(Ui Ui⊤ )
= vec(Qi )
(20)
where the first equality follows because Qi is idempotent.
Thus g(·) = f (·), which concludes the proof.
Remark 1 (Invariance to coordinate transformations):
Notice that wM is independent of the particular state-space
realization adopted for the dynamical systems since Q1 and
Q2 are projection matrices.
(22)
where the second equality follows from the vectorization rule
vec(ABC) = (C ⊤ ⊗ A)vec(B) for matrices A, B and C of
appropriate dimension. Thus,
wM = vec(Q)
n1
n2
X
X
=
Φ1,k −
Φ2,k
(23)
k=1
and the support vectors (cf. (12)) are the left singular vectors
of the observability matrices O1 and O2 .
V. DATA - DRIVEN CLASSIFICATION BASED
S UPPORT V ECTOR M ACHINES
⊤
= Y QY = vec(Q) (Y ⊗ Y )
⊤
= wM
Φ
ni
X
k=1
k=1
defines a correct classifier for the dynamical systems in (1).
2
Notice now that πi (Y ) = k(I − Q⊤
i )Y k for all Y . Hence,
we get
(21)
ON
The model-based approach suggests an SVM formulation
for the problem of classifying data generated by dynamical
systems. We will first describe the SVM formulation and we
will then make some considerations on the generalization
performance of the solution. An interesting result is that the
generalization performance of the data-driven classifier can
be quantified as a function of the dynamics of the systems
which generate the training dataset.
A. Data-driven classification based on SVM
Ltr
i
Let
⊂ Li , i = 1, 2, be a finite nonempty subset of Li
consisting of all the nonzero trajectories recorded from Σi .
tr
tr
Thus Yk ∈ Ltr := (Ltr
1 ∪ L2 ), k = 1, 2, . . . , L, L := |L |,
tr
is the k-th training vector. Let ℓk = 1 if Yk ∈ L1 , and
ℓk = −1 if Yk ∈ Ltr
2 . Finally, let Φk = Yk ⊗ Yk be the
feature vector associated to Yk . Following Section III and
Theorem 2, we formulate the data-driven approach as the
problem of finding the hyperplane that contains the origin
and separates the training datasets with maximum margin,
that is:
1
kwk2
2
w ⊤ Φk ≥ 1
subject to
w⊤ Φk ≤ −1
min
w
if ℓk = 1
(24)
if ℓk = −1
The following result holds true.
Theorem 3 (Data-driven classifier): Let Assumption 1 be
satisfied, and consider an arbitrary training dataset Ltr . Then,
the solution wD to the optimization problem (24) exists and
⊤
is unique. Hence, f (Y ) = wD
Φ is a correct classifier with
tr
respect to L .
Proof of Theorem 3. The proof follows from Theorem 2.
In fact, the model-based solution wM guarantees that a1 :=
⊤
⊤
minℓk =1 wM
Φk > 0 and a2 := maxℓk =−1 wM
Φk < 0.
Thus w M := wM /a with a := min{a1 , −a2 } guarantees
the feasibility of the set of constraints. Uniqueness follows
as the optimization problem is a convex program.
The constraint that the solution must contain the origin is
simply to mimic the model-based solution. This constraint is
actually not needed, and a standard formulation (10) would
still guarantee existence and uniqueness of the solution. If
we constrain the solution to contain the origin the Wolfe dual
becomes
1
max µ⊤ 1 − µ⊤ Zµ
µ
2
(25)
subject to µ 0
P
which does not involve the constraint k=1,2,...,L µk ℓk = 0.
We notice that in this case the dimension of the feature
space is (N m)2 . Nonetheless, by considering a kernel-based
implementation one can remain in the N m-dimensional
space of the sequences Y .
Theorem 3 indicates that one can find a surface separating
the training dataset without information about the underlying
systems except for their linearity, which suggests the feature
space of choice. In the remaining part of this section, we will
discuss on the capability of this SVM classifier to generalize
to observations outside the training set.
B. Expected risk
Ideally, one would like to establish the correctness of the
data-driven classifier in the same sense as Definition 1. This
is a non-trivial problem which, to the best of our knowledge,
has not yet been solved. In the sequel, we consider another
way to characterize the generalization performance of the
SVM classifier, based on the notion of expected risk. While
this notion does not provide deterministic bounds, it has the
merit to capture the situation where the training dataset is
randomly chosen. Hence, it has the merit to describe cases
in which one cannot perform dedicated experiments on the
systems.
Consider a training dataset of L random i.i.d. observations
drawn according to a probability distribution P (Y, ℓ). Given
a classifier f (Y ), its expected risk can be defined as [1]:
Z
1
|ℓ − sgn(f (Y ))| dP (Y, ℓ)
(26)
R :=
2
where sgn is the sign function. The expected risk quantifies
the capability of a classifier to generalize from the training
dataset. Several studies have been devoted to provide upper
bounds on the expected risk for a given family of classifiers.
An interesting bound for linear classifiers which serves our
discussion is reported hereafter.
Theorem 4 ([19]): Let Y ∈ Rd belong to the sphere of
radius R, and consider the class F of real-valued functions
defined as F := {Y 7→ w⊤ Y : kwk ≤ 1, kY k ≤ R}. There
is a constant c such that, for all probability distributions, with
probability at least 1 − η over L randomly i.d.d. vectors, if
a classifier has margin at least ρ on all the examples then its
expected error is not larger than
c R2
2
log
L
+
log
(1/η)
(27)
L ρ2
The constant c is related to the so-called fat-shattering
dimension of linear classifiers, and its explicit expression can
be found in [19]. Theorem 4 shows that one can quantify
the generalization performance of a linear classifier as a
function of the margin ρ obtained for the training dataset.
We now show that the margin of the SVM classifier can
be related to the margin of the model-based solution. This
permits to quantify the generalization performance of the
SVM classifier in terms of the dynamics of the systems that
one wishes to classify. The analysis which follows holds
for normalized data. We will briefly comment later on the
general case.
Consider normalized training data
Y k :=
Yk
kYk k
(28)
with feature vector Φk := Y k ⊗ Y k . It holds that kΦk k = 1.
Consider now the optimal solution wD to (24) computed
with respect to Φk , whose existence and uniqueness is again
ensured by the model-based solution. We can assume without
loss of generality that kwD k ≤ 1. Let now ρM and ρD
represent the margin corresponding to the model-based and
the data-driven solutions, respectively,
ρi :=
min
k=1,2,...,L
|wi⊤ Φk |
,
kwi k
i ∈ {M, D}
(29)
It holds that
VI. D ISCUSSION
ρD ≥ ρM
(30)
irrespective of the training dataset, because the data-driven
solution is the margin maximizer. The next result shows that,
using normalized data, ρM is bounded from below by a
positive quantity that depends solely on the dynamics of the
systems one wishes to classify. We refer the reader to [20]
for a definition of principal angles.
Theorem 5 (Bound on the data-driven classifier margin):
Let Assumption 1 be satisfied. Consider an arbitrary training
dataset of vectors Yk , and let Φk := Y k ⊗ Y k where Y k is
as in (28). Let wD be the unique solution to the optimization
problem (24) computed with respect to Φk . Then, it holds
that
β
ρD ≥ p
(31)
2 (n1 + n2 )
where n1 and n2 are the orders of the dynamical systems
in (1), and β is the squared sine of the smallest principal
angle between the subspaces spanned by the columns of the
observability matrices O1 and O2 .
Proof of Theorem 5. Since ρD ≥ ρM it is sufficient to
bound ρM . The term kwM k satisfies
kwM k2 = vec(Q)⊤ vec(Q)
= kQk2F
≤ 2 (n1 + n2 )
(32)
where k · kF denotes Frobenius norm. The inequality follows
from kQk2F ≤ 2kQ1 k2F + 2kQ2 k2F and kQi k2F = ni because
the Qi ’s are projection matrices. Consider now the term
⊤
|wM
Φk |. Assume without loss of generality that its minimum
is attained for some Y∗ ∈ Ltr
1 . It holds that
min
k=1,2,...,L
⊤
⊤
|wM
Φk | = |wM
Φ∗ |
= |π2 (Y ∗ ) − π1 (Y ∗ )|
= π2 (Y ∗ )
(33)
The third equality comes from the fact that Y∗ ∈ Ltr
1 implies
π2 (Y ∗ ) > 0 and π1 (Y ∗ ) = 0 in view of Assumption 1. As
shown in [9, Theorem 1], π2 (Y ∗ ) ≥ βkY ∗ k2 . Hence, the
proof follows from kY ∗ k = 1.
Theorem 5 permits to bound the risk of the data-driven
classifier based on the dynamics of the systems one wishes
to classify, and formalizes the intuition that the risk bound
becomes smaller as the dynamics of the systems to classify
are more distant from one another. In fact, the higher β the
larger the coefficient of inclination between the subspaces
spanned by the columns of O1 and O2 , which is maximal
when the two spaces are orthogonal.
Data normalization ensures that ρM is bounded away from
zero. This property does not hold in general since trajectories
of dynamical systems can be arbitrarily close to the origin.
Nonetheless, one can obtain a very similar bound by adding
to (27) an extra term which accounts for training data below
the margin ρ [19, Theorem 1.7].
It is intuitive that incorporating models can be beneficial
to the classification task. This fact is obvious also from the
analysis shown in this paper since under Assumption 1 no
classifier can outperform the model-based classifier when the
models are exact and the data are noise-free. However, as
mentioned before, the model-based approach introduces the
non-trivial issue of how to quantify the effect of modelling
inaccuracies. The data-driven bypasses the intermediate step
of identification, and thus it has the potential to be applicable
also when accurate models are difficult to obtain. Hereafter,
we briefly elaborate on this point also in connection with a
number of open problems.
A. Linear systems with noisy observations
When observations are corrupted by noise, identification
may be difficult and require, even for linear systems, many
careful provisions [21]. In contrast, an SVM formulation
can address the problem in a rather straightforward manner.
Consider a soft-margin SVM [2]:
X
1
kwk2 +
C ξk
(w,ξ) 2
k=1,2,...,L
w⊤ Φk ≥ 1 − ξk
subject to
w⊤ Φk ≤ −1 + ξk
min
if ℓk = 1
(34)
if ℓk = −1
where C is a parameter and ξ := col (ξ1 , ξ2 , . . . , ξL ) is the
vector of slack variables, which account for the fact that
noise may render the data non-separable. On one hand, there
exist many studies aimed at quantifying the generalization
performance of SVM also for soft-margin formulations [22].
On the other hand, even with noisy data one can still give a
separation measure between linear systems (the margin ρM )
as a function of their dynamics and the signal-to-noise ratio
[9, Theorem 2]. This means that even with noisy data one can
quantify the generalization performance of an SVM classifier
along the same lines as in Section V-B.
We point out that while this reinforces the idea that for
linear systems polynomial kernels are good candidates, it
remains unclear if better performance can be obtained with
different kernel functions.
B. Nonlinear systems
Classification for nonlinear systems is another situation in
which an SVM formulation can bypass difficulties related to
system identification. This is related to the capability of SVM
to find non-linear separating surfaces in a straightforward
manner through the kernel trick. Interestingly, even in the
nonlinear case one can define a separation measure between
dynamics [12, Theorem 3]. However, in contrast with the
linear case where this measure involves principal angles
between observability subspaces, for nonlinear systems this
measure involves K-functions, which are often difficult to
relate to the underlying dynamics. Like for linear systems,
a deeper understanding of this point would be beneficial to
figure out which types of kernel functions are most suitable
for a given class of nonlinear systems.
In fact, theoretical studies on classification for nonlinear
systems are recent also within computer science, and the
approaches appear largely diversified; for example, see [23]
for an interesting recent account. Yet, also in this context,
the question of which kernel functions are most suitable for
a given class of dynamics is unresolved.
C. Classifiers in-the-loop
Thanks to their simple form, classifiers have the potential
to be used in real-time applications, thus for control purposes.
This fact has been noted in [24], where the authors introduce
the term classifier in-the-loop to describe a framework in
which a classifier can modify online the control action by
looking at the process data. A notion of generalization is
considered, which characterizes the capability of a classifier
to work under small perturbations of the system vector fields,
which is a sensitivity-type analysis. While the results are
promising, it remains unexplored how to handle more general
forms of uncertainty. Ideally, one should provide bounds on
the risk function of a classifier that hold for all the possible
system trajectories, and relate such bounds with closed-loop
stability properties. A non-trivial difficulty is that much of
the theory on the generalization properties of classifiers have
been developed in a probabilistic setting, while for robust
stability it is desirable to guarantee worst-case deterministic
bounds.
Interestingly, the architecture considered in [24] can be
regarded as a supervisory control system [13]. In supervisory
control, the supervisor selects based on process data which
candidate control law (hypothesis) is most appropriate at any
given time. This is done by assigning to each candidate law a
score function (cost function) that quantifies the performance
level achievable by the control law given the process data, In
supervisory control, one often uses the term cost detectability
[25], [26] to measure the capability of a supervisor to learn
from data an appropriate control law even when the process
does not match the models used to design the control laws.
In fact, the supervisor is a classifier and cost detectability
is a measure of its generalization performance. The idea
of adaptive control as learning-from-data is indeed not new
[27], but a firm theoretical link with the realm of machine
learning has not yet been established.
VII. A N UMERICAL E XAMPLE
Consider a system with transfer function
G(s) =
s+1
(s + 10)(s2 + s + 1)
(35)
where s is the Laplace variable. To improve performance, the
system is controlled with a proportional controller K = 30
under negative feedback. The goal is to design a classifier
which can detect the loss of control effectiveness. We denote
by Σ1 the open-loop system and by Σ2 the closed-loop
system. Hence, Σ1 and Σ2 have transfer functions G(s)
and W (s) := KG(s)/(1 + KG(s)), respectively. Finally,
we denote by Σ1 and Σ2 the corresponding sampled-data
systems under sampling time Ts . The systems are as in (1)
with n1 = n2 = 3 and m = 1. Using the previous notation,
we let N be the length of the observation sequences, and L
the number of training data. We let Q be the number of data
used for validation. In order for Assumption 1 to be satisfied
one needs N ≥ 6. Under such condition, Assumption 1 holds
for a generic choice of Ts .
We focus on the SVM classifier because the model-based
classifier is always correct under Assumption 1. We note that
classifying the two systems is non-trivial, as one can observe
from Figure 2. For instance, for Ts = 0.1, in the ideal case
of N = ∞ one has β = 0.004 in Theorem 5, and a cepstral
distance [20] equal to 1.045.
We report in Table I simulations results for various choices
of N , L and Ts , with Q = 1000 validation data. The SVM
classifier is computed as in Theorem 5. For the training and
the validation test, trajectories are generated from random
initial conditions with zero mean and variance σ 2 = 100.
The error in the validation test is defined as:
X
1
⊤
Φk )
(36)
ℓk − sgn(wD
Rtest :=
2Q
k=1,2,...,Q
Variation of the parameter N (Ts = 0.1, L = 50)
N =2
N =5
N = 10
N = 50
N = 100
Rtest
0.4720
0.0760
0.0685
0.0150
0.0150
Variation of the parameter L (Ts = 0.1, N = 10)
L=3
L=5
L = 10
L = 50
L = 100
Rtest
0.1480
0.1480
0.0720
0.0685
0.0620
Variation of the parameter Ts (N = 10, L = 50)
Ts = 0.01
Ts = 0.05
Ts = 0.1
Ts = 0.5 Ts = 1
Rtest
0.4850
0.0730
0.0685
0.0385
0.0835
TABLE I: Numerical results for the SVM classifier with Q = 1000.
One sees that the classifier performs well for reasonable
choices of the parameters. In particular:
(i) Dependence on N . As N goes to zero, the performance
is clearly that of a random guess. One the other hand,
remarkably, classification becomes accurate exactly as
soon as one approaches the theoretical bound N ≥ 6.
The performance saturates after N = 50. To further
decrease the error we need to increase L (with L > 300
one can achieve an error below 1%).
(ii) Dependence on L. The performance variations are less
evident in this case. This suggest that L is less critical
than N . The intuition is that random initial conditions
generically ensure the excitation of all system dynamics
so that even few examples may suffice.
(iii) Dependence on Ts . The sampling time does not play a
major role as long as we avoid over-sampling, in which
case both A1 and A2 tend to the identity matrix, or
under-sampling, in which case both A1 and A2 tend to
the zero matrix.
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0
0
-0.2
-0.2
-0.4
-0.4
-0.6
-0.6
-0.8
-0.8
rag replacements
-1
1
2
3
4
5
6
7
PSfrag
replacements
8
9
10
-1
1
2
3
4
5
6
7
8
9
10
y(1)
y(1)
sampled trajectories
sampled trajectories
Fig. 2: Output trajectories of the two systems (Left: open-loop; Right: closed-loop) with N = 10 and Ts = 0.1. The figures report 1000 normalized
trajectories for each system generated from random initial conditions with zero mean and variance σ2 = 100.
VIII. C ONCLUDING R EMARKS
We have considered the problem of classifying trajectories
generated by dynamical systems, looking at a model-based
approach, the common approach in control engineering,
as well as at a data-driven approach based on Support
Vector Machines, a popular method in computer science.
The present discussion suggests that both the approaches
have distinct merits. A deeper understanding of the interplay
between these two approaches would help to establish a
sound theory for dynamical systems more general than those
considered in this paper.
R EFERENCES
[1] V. Vapnik, “The nature of statistical learning theory.” Springer-Verlag,
New York, 1995.
[2] C. Burges, “A tutorial on support vector machines for pattern recognition,” Data Mining and Knowledge Discovery, vol. 2, pp. 121–167,
1998.
[3] S. Mahadevan and S. Shah, “Fault detection and diagnosis in process
data using one-class support vector machines,” Journal of Process
Control, vol. 19, pp. 1627–1639, 2009.
[4] R. Vidal, A. Chiuso, S. Soatto, and S. Sastry, “Observability of linear
hybrid systems,” in Hybrid Systems: Computation and Control, Lecture
Notes in Computer Science, O. Maler and A. Pnueli, Eds. Berlin
Heidelberg: Springer-Verlag, 2003, vol. 2623, pp. 523–539.
[5] M. Babaali and M. Egerstedt, “Observability of switched linear systems,” in Hybrid Systems: Computation and Control, Lecture Notes in
Computer Science, M. Morari and L. Thiele, Eds. Berlin Heidelberg:
Springer-Verlag, 2004, vol. 3414, pp. 48–63.
[6] L. Vu and D. Liberzon, “Invertibility of switched linear systems,”
Automatica, vol. 44, no. 4, pp. 949–958, 2008.
[7] H. Lou and R. Yang, “Conditions for distinguishability and observability of switched linear systems,” Nonlinear Analysis: Hybrid Systems,
vol. 5, pp. 427–445, 2011.
[8] G. Battistelli, “On stabilization of switching linear systems,” Automatica, vol. 49, pp. 1162–1173, 2013.
[9] M. Baglietto, G. Battistelli, and P. Tesi, “Mode-observability degree
in discrete-time switching linear systems,” Systems & Control Letters,
vol. 70, pp. 69–76, 2014.
[10] A. Tanwani and D. Liberzon, “Robust invertibility of switched linear
systems,” in Proceedings of the 50st IEEE Conference on Decision
and Control, Orlando, FL, USA, 2011, pp. 441–446.
[11] ——, “Invertibility of switched nonlinear systems,” Automatica,
vol. 46, no. 12, pp. 1962–1973, 2010.
[12] M. Baglietto, G. Battistelli, and P. Tesi, “Distinguishability of discretetime nonlinear systems,” IEEE Transactions on Automatic Control,
vol. 59, pp. 1014–1020, 2014.
[13] D.Liberzon, Switching in Systems and Control. Birkhäuser, 2003.
[14] L. Chisci, A. Mavino, G. Perferi, M. Sciandrone, C. Anile, G. Colicchio, and F. Fuggetta, “Real-time epileptic seizure prediction using
AR models and support vector machines,” IEEE Transactions on
Biomedical Engineering, vol. 57, pp. 1124–1132, 2010.
[15] K. Brodersen, T. M. Schofield, A. Leff, C. Soon Ong, E. Lomakina,
J. Buhmann, and K. Stephan, “Generative embedding for model-based
classification of fMRI data,” PLoS Computational Biology, vol. 7,
e1002079, 2011.
[16] T. Jebara, R. Kondor, and A. Howard, “Probability product kernels,”
Journal of Machine Learning Research, vol. 5, pp. 819–844, 2004.
[17] S. Vishwanathan, A. Smola, and R. Vidal, “Binet-Cauchy kernels on
dynamical systems and its application to the analysis of dynamic
scenes,” International Journal of Computer Vision, vol. 73, pp. 95–
119, 2007.
[18] O. Lauwers and B. De Moor, “A time series distance measure
for efficient clustering of input/output signals by their underlying
dynamics,” in Proceedings of the 56th IEEE Conference on Decision
and Control, 2017.
[19] P. Bartlett and J. Shawe-Taylor, “Generalization performance of support vector machines and other pattern classifiers,” in Advances in
Kernel Methods - Support Vector Learning, B. Schölkopf, C. Burges,
and A. Smola, Eds. Cambridge, MA, MIT Press, 1999, pp. 43–54.
[20] K. De Cock and B. De Moor, “Subspace angles between ARMA
models,” Systems & Control Letters, vol. 46, pp. 265–270, 2002.
[21] G. Pillonetto, F. Dinuzzo, T. Chen, G. De Nicolao, and L. Ljung, “Kernel methods in system identification, machine learning and function
estimation: A survey,” Automatica, vol. 50, pp. 657–682, 2014.
[22] J. Shawe-Taylor and N. Cristianini, “On the generalization of soft margin algorithms,” IEEE Transactions on Information Theory, vol. 48,
pp. 2721–2735, 2002.
[23] Y. Shen, P. Tino, and K. Tsaneva-Atanasova, “Classification framework for partially observed dynamical systems,” PHYSICAL REVIEW
E, vol. 95, 043303, 2017.
[24] H. Poonawala and U. Topcu, “Robustness of classifier-in-the-loop
control systems: A hybrid-systems approach,” in Proceedings of the
56th IEEE Conference on Decision and Control, 2017.
[25] M. Stefanovic and M. Safonov, “Safe adaptive switching control:
Stability and convergence,” IEEE Transactions on Automatic Control,
vol. 53, pp. 2012–2021, 2008.
[26] G. Battistelli, J. Hespanha, E. Mosca, and P. Tesi, “Model-free adaptive
switching control of uncertain time-varying plants,” IEEE Transactions
on Automatic Control, vol. 8, pp. 1208–1220, 2013.
[27] M. Safonov and T. Tsao, “The unfalsified control concept and learning,” IEEE Transactions on Automatic Control, vol. 42, pp. 843–847,
1997.
| 3cs.SY
|
Sentence Object Notation:
Multilingual sentence notation based on Wordnet
Abdelkrime Aries
Djamel eddine Zegour
Walid Khaled Hidouci
Ecole nationale Supérieure d’Informatique (ESI, ex. INI), Algiers, Algeria
Emails: {ab aries, d zegour, w hidouci}@esi.dz
arXiv:1801.00984v2 [cs.CL] 10 Jan 2018
Abstract
The representation of sentences is a very important task. It can be used as a way to exchange data interapplications. One main characteristic, that a notation must have, is a minimal size and a representative form. This
can reduce the transfer time, and hopefully the processing time as well.
Usually, sentence representation is associated to the processed language. The grammar of this language
affects how we represent the sentence. To avoid language-dependent notations, we have to come up with a new
representation which don’t use words, but their meanings. This can be done using a lexicon like wordnet, instead
of words we use their synsets. As for syntactic relations, they have to be universal as much as possible.
Our new notation is called STON ”SenTences Object Notation”, which somehow has similarities to JSON. It
is meant to be minimal, representative and language-independent syntactic representation. Also, we want it to be
readable and easy to be created. This simplifies developing simple automatic generators and creating test banks
manually. Its benefit is to be used as a medium between different parts of applications like: text summarization,
language translation, etc. The notation is based on 4 languages: Arabic, English, Franch and Japanese; and there
are some cases where these languages don’t agree on one representation. Also, given the diversity of grammatical
structure of different world languages, this annotation may fail for some languages which allows more future
improvements.
Keywords: Sentence annotation, Sentence structure, Multilingual languages, Data exchange languages, Knowldge
representation, Natural language processing
1
Introduction
Tagging sentences is a very important task in natural language processing. One of the famously known methods is
syntactic tagging. The main idea is to detect the different parts (structure) of a sentence, such as nominal phrases,
verbal phrases, nouns, verbs, etc. This structure can be expressed using some languages like XML, JSON, etc.
[Bertran et al., 2008, Recasens and Martı́, 2009, Lopatková et al., 2011]. The problem with syntactic tagging is its
dependency to the processed language. Indeed, it is a good way if our system is destined for a specific language.
But, when it comes to multilingual or cross-lingual systems, it is better to come up with a way to represent the
sentence structure independently from languages.
Semantic representation of sentences is one solution to this problem. The meaning of a sentence is something
beyond languages; it is related to the different concepts of its words and the different relations between them.
Words are just a way to describe a concept in a given language. For example, the words ” èQ m. /shajarah/ ” in
Arabic, ”tree” in English, ”arbre” in French and ”木 /ki/ ” in Japanese refer to the same concept. This concept in
Wordnet 3.01 [Miller, 1995] is defined as ”a tall perennial woody plant having a main trunk and branches forming
1 WordNet is a large lexical database of English. Nouns, verbs, adjectives and adverbs are grouped into sets of cognitive synonyms (synsets),
each expressing a distinct concept. URL: https://wordnet.princeton.edu
1
a distinct elevated crown; includes both gymnosperms and angiosperms”. The idea of semantic representation is
to represent the words as concepts and each language can link its own words to these concepts. Then, the semantic
relations between these concepts are extracted from the sentence [Uchida et al., 1999, Banarescu et al., 2013]. It
will be great if we can gather all concepts of different languages, then make links to each one of them. Also, to
detect the semantic relations between the concepts in a sentence automatically, that will need a large amount of
knowledge and processing power.
Our idea is to propose a simple language which will help us transfer the information in sentences between
applications. For example, we can use it to send sentences from a summarization system to a translation system.
That allows us to create what is known as cross-lingual summarization. The language has to be simple and unambiguous so the developers can, easily, create tools to encode a specific natural language to this one. It must
have a minimal size to minimize the transfer time and also the processing one. One of the main characteristics on
which we insist is the readability, which helps us to create STON representations manually. To this end, we want to
represent the syntactic relations in sentences taking in mind the multilingualism aspect. We must insist on the term
multilingual which means ”using several languages”, so it doesn’t have to represent all world languages. In our
study, we are interested in four languages: Arabic, English, French and Japanese. For non Latin scripts, namely
Arabic and Japanese, we will provide the ALA-LC romanization2 .
The rest of this paper is organized as follows. Section 2 presents some related words in the subject of text
annotation. Section 3 describes our main proposition and different parts of STON. Section 4 addresses the cases
where we have adpositional phrases, relative clauses and comparison. In section 5, some solutions are proposed
concerning the coordination between references, proper names, verbs with two objects, complementizer and passive voice. Section 8 is reserved to discuss our work, its grounds, its contribution in comparison with other works,
its benefits and its limits. Finally, section 9 is reserved for conclusion and future improvements.
2
Related works
Sentence structure can be represented using generic-purpose languages such as XML or JSON. Lets take XML as
an example, to represent a sentence we have to specify some roles in a DTD file. It would contain different structures of a sentence, such as subject, object, verb, tense of the verb, etc. For example, Recasens and Martı́ [2009]
uses XML to represent sentences in order to annotate corpora for Spanish and Catalan. They used AnCoraPipe
[Bertran et al., 2008] to create the corpus. Figure 1 represents an example of XML annotation of the sentence ”La
Comisión Europea anunció . . . ” which means ”The European Commission announced . . . ”. They use XML tags
to express nouns, verbs, nominal phrases, etc. where each tag has some properties.
<sn arg="arg0" entity="entity1" entityref="ne"
func="suj" ne="organization" tem="agt">
<spec gen="f" num="s">
<d gen="f" lem="el" num="s" posttype="article" wd="La" />
</spec>
<grup.nom gen="f" num="s">
<n entityref="ne" gen="c" lem="Comisión_Europea"
ne="organization" num="c" postype="proper"
sense="16:cs1" wd="Comisión_Europea" />
</grup.nom>
</sn>
<grup.verb>
<v els="a2" lem="aunciar" mood="indicative" num="s"
person="3" postype="main" tense="past" wd="anunció" />
<grup.verb>
Figure 1: Example of XML sentence annotation [Recasens and Martı́, 2009]
Knowledge-based interlingual machine translation uses a representation of sentences as medium between the
source language and the destination one. KANT [Mitamura et al., 1991] is an example system that uses an interlingua to represent sentences before translation. KANT interlingua is a list-based structural representation scheme
2 ALA-LC (American Library Association - Library of Congress) is a set of standards for romanization to represent texts in other writing
systems using the Latin script. URL: https://www.loc.gov/catdir/cpso/roman.html
2
(*A-REMAIN ; action rep for ’remain’
(FORM FINITE)
(TENSE PAST)
(MOOD DECLARATIVE)
(PUNCTUATION PERIOD)
(IMPERSONAL -) ; passive + expletive subject
(ARGUMENT-CLASS THEME+PREDICATE) ; predicate argument structure
(Q-MODIFIER ; PP semrole (generic)
(*K-DURING ; PP interlingua
(POSITION FINAL) ; clue for translation
(OBJECT ; PP object semrole
(*O-TIME ; object rep for ’time’
(UNIT -)
(NUMBER SINGULAR)
(REFERENCE DEFINITE)
(DISTANCE NEAR)
(PERSON THIRD)))))
(THEME ; object semrole
(*O-DEFAULT-RATE ; object rep for ’default rate’
(PERSON THIRD)
(UNIT -)
(NUMBER SINGULAR)
(REFERENCE DEFINITE)))
(PREDICATE ; adjective phrase semrole
(*P-CLOSE ; property rep for ’closer’
(DEGREE POSITIVE)
(Q-MODIFIER
(*K-TO
(OBJECT
(*O-ZERO
(UNIT -)
(NUMBER SINGULAR)
(REFERENCE NO-REFERENCE)
(PERSON THIRD))))))))
Figure 2: Representing the sentence ”The default rate remained close to zero during this time.” in KANT interlingua [Czuba et al., 1998]
using nested frames. Each interlingua frame contains a head concept, series of feature-value pairs, and semantic
slots containing nested interlingua frames. Concepts are symbols which begins with an asterisk (*) followed by a
concept prefix defining its category (e.g. *A-DRIVE which is the action drive). KANT interlingua distinguishes
many categories such as action, object, manner, proper name, etc. It uses certain features of the input text, such as
modality, aspect, discourse markers, etc. in order to generate grammatically accurate output texts [Mitamura et al.,
1991]. Semantic roles are relations between frames, such as agent, theme, etc. Figure 2 represents the KANT
interlingua’s representation of the sentence ”The default rate remained close to zero during this time.”. It contains
concepts such as *A-REMAIN, *K-DURING, etc.; features such as FORM, TENSE, etc. and semantic roles such
as Q-MODIFIER, THEME, etc.
Despite being an interlingua, KANT shows some limitations when representing sentences [Czuba et al., 1998].
It is more close to English semantics than being multilingual, because it is intended for English to other languages
translation. Also, It is designed for technical domains, thus the vocabulary is limited to a subset of meanings.
Universal Networking Language (UNL) is a knowledge representation language to represent the meaning of
texts without ambiguity. It was developed in 1996, as un intermediate multilingual language to be used through
the Internet [Uchida et al., 1999, Uchida and Zhu, 2005]. The major commitments of the UNL are the following:
• It must represent information: represent ”what was meant” and not ”what was said” or ”how it was said”.
• It must be a language for computers: like HTML, SGML, XML, etc.
• It must be self-sufficient: The UNL representation must not depend on any implicit knowledge and should
explicitly codify all information.
3
{unl}
agt(affect(icl>do).@present.@entry:01,
human(icl>animal).@pl)
obj(affect(icl>do).@present.@entry:01,
environment(icl>abstract thing).@pl)
{/unl}
Figure 3: Representing sentence ”Human affect the environment” in UNL
• It must be general-purpose: Its primary objective is to serve as an infrastructure for handling knowledge. It
can be used for different tasks such as: translation, text mining, multilingual document generation, summarization, etc.
• It must be independent from any particular natural language.
UNL defines some tags for the structure of the text: document ”[D]”, paragraph ”[P]” and sentence ”[S]”. Concepts
are represented as character-strings called ”Universal Words (UWs)”, where each natural language (En, Fr, etc.)
has its own word dictionary. UNL expressions are based on binary relations, where each binary relation has two
UWs as parameters. Also, UNL specifies some attributes to represent information conveyed by natural language
grammatical categories (such as tense, mood, aspect, number, etc). Figure 3 represents UNL formulation of
the sentence ”Human affect the environment”. The sentence starts with the tag ”{unl}” followed by two binary
relations. The relation ”agt” defines a thing which initiates an action. In our example, ”Human” which is a plural
noun is the one who do (in present tense) the action of ”affecting”. The relation ”obj” defines a thing in focus
which is directly affected by an event or state. In our example, ”environment” is the direct object of the action of
”affecting”.
Indeed, UNL has a minimal size, with a multilingual background and covers a large number of languages.
However, it shows some limitations when it comes to which relations we must choose. Boguslavsky [2013] claims
that the selection of relations differs from team to team as it is sometimes ambiguous which one to choose. An
example of that is the phrase ”freedom for all” which was described with the purpose relation ”pur” and with
the beneficiary relation ”ben” by another team. Martins [2013] raises some other issues concerning UNL. One of
these issues is the proper nouns, are they treated as permanent UWs or just temporary? Also, a concept can be
represented by simple UW or a compound UW. For example ”the physiological need for food” can be represented
using the UW ”hunger”, as it can be represented with the compound UW ”hungry.@ness”. Technically speaking,
we can say it is hard to design an encoder from natural language to UNL. After parsing a sentence, we have to find
the relations between each part of this sentence. It is, sometimes, too ambiguous to select between two relations
manually, let alone selecting them automatically.
A most recent representation language is AMR (Abstract Meaning Representation) proposed by Banarescu
et al. [2013]. It is a semantic representation language designed to represent the meaning of English sentences.
The text is represented as a graph, where the leaves are labeled with concepts such as ”(b / boy)” which means an
instance called ”b” of the concept ”boy” (See Figure 4). The concepts are English words (”boy”), PropBank framesets [Kingsbury and Palmer, 2002] (”want-01”), or special keywords: special entity types (”date-entity”, ”worldregion”, etc.), quantities (”monetary-quantity”, ”distance-quantity”,etc.) and logical conjunctions (”and”, etc).
The relations between concepts are:
• Frame arguments, following PropBank conventions.
• General semantic relations.
• Relations for quantities.
• Relations for date-entities.
• Relations for lists.
AMR is light weight annotation, but it is heavily based on English. It has some limitations where it comes
to inflectional morphology for tense and number. It does not deeply capture many noun-noun or noun-adjective
relations. Also, because it relies on Propbank framesets, it is subject to the Propbank constraints.
4
∃ w, b, g:
instance(w, want-01)
∧ instance(g, go-01)
∧ instance(b, boy)
∧ arg0(w, b) ∧
arg1(w, g) ∧ arg0(g, b)
(w / want-01
:ARG0 (b / boy)
:ARG1 (g / go-01
:ARG0 b))
LOGIC format
AMR format
GRAPH format
Figure 4: AMR representation of the sentence ”The boy wants to go” [Banarescu et al., 2013].
Ar
En
Fr
Ja
Ar2STON
En2STON
Fr2STON
Ja2STON
...
...
STON-Generator
Ar
Wornet
En
Wornet
STON
Fr
Wornet
Ja
Wornet
Language-independent applications
STON2Ar
Ar
STON2En
STON
STON2Fr
STON2Ja
...
En
STON-Parser
Fr
Ja
...
Figure 5: Example of STON’s purpose
3
STON representation
STON is intended to represent sentences’ syntactic structures in a multilingual context. Figure 5 represents our
vision on how STON may be used. We can use lexical parsers of several languages to get the different parts of
speech. Then, we can use them to create STON representation using a STON generator and Wordnet lexicon
of each language. The representation can be used in multiple multilingual applications, like the example we
mentioned earlier. Then having a STON parser, we can extend it to handle each language apart. Using a realizer
or a language generator, we can reproduce a readable text for the destination language.
In our representation, we look at different parts of sentences as either actions or roles. Actions are the dynamic
part of the sentence. Each action contains the verb and its morphological specifications like tense, negation, etc.
While roles are, generally, nominal phrases that have a purpose in the action. They can be agents, themes, places,
times, etc.
3.1
Roles
Nominal phrases play roles in an action; They can be agents, themes, places, times, etc. In our case, agents are
those who do the action (either in active or passive voice), and themes are those who undergo or experience the
action. For example, in table 1, ”the fat man” and ”delicious food” are roles played in the action of ”eating” which
happened in the present. The first phrase plays the role of an agent, while the second plays the role of a theme.
In our representation (STON), we list all the roles that appear in the sentences. Each role contains the following
attributes (see Figure 6):
• id: a name for the role in order to reference it.
5
Table 1: Roles and Actions in Arabic, English, French and Japanese.
Language Sentence
Arabic
English
French
Japanese
Ég. QË @
. YK YÊË@ ÑjÊË@ É» AK á ÒË@
/al-Rajulu al-Samı̄nu ya’kulu al-Lah.ma al-Ladhidh./
The fat man eats the delicious meat.
Le gros homme mange la viande déliciouse.
太った男性は美味しい肉を食べます。
/futotta dansei wa oishii niku o tabemasu./
@r:[
r:{
id: <role-ID>;
syn: <synset>;
qnt: <[O]number/PL>;
def: <Y/N>;
@adj:[
adj:{
syn: <synset>;
adv: [synset, ...];
adj:}
adj:{
...
adj:}
adj:];
r:}
r:{
...
r:}
r:]
Figure 6: Roles representation.
• syn: the synset number of the noun in the lexicon (Wordnet in our case).
• A set of adjectives blocks that modify the noun. Each adjective block contains its synset and a set of adverbs
synsets.
• qnt: the quantity; it describes the amount of the noun. For example ”10 apples”, the quantity is 10. By
default, it equals 1; it is a number but it can have the term ”PL” for plural. Some languages, like Arabic, has
dual numbers ” úæ J Ó /muthanna/ ” which can simply represented by the number 2. Then, the language
generator can handle this. As for ordinal numbers, we add ”O” before the number; example, ”O2” means
second.
• def: defined; many languages, such as Arabic, English and French, have the ability to identify a noun from
many using the definite articles.
3.2
Actions
The verb in a sentence, or in a clause, represents an action. In languages, like Arabic, there are some nominal
sentences, which doesn’t include a verb. Nevertheless, we can add a verb to express the action, like the examples
in Table 2. The first sentence is composed of a subject ” @ Y JJ. Ó /mubtada’/ ” (which is a Noun) and a predicate ”
Q .g /khabar/ ” (which is an Adjective). In the second one, the subject is a noun and the predicate is a prepositional
phrase. In both sentences, we can use the copula (to be) as if it is the action, the subject will be considered as an
agent and the predicate as a theme. In case of the third sentence, where we have an active participle ” É« A ¯ Õæ @
/ism fā‘il/ ”, and the origin verb is a movement verb like ”to go” we can consider it as present continuous [Haak,
1997].
To represent the actions, we define a set of actions blocks, where each block contains some attributes. The
attributes are generally related to the verb, since the action is all about the verb. But, it contains also some links to
agents, themes, etc. These are the main attributes contained in each action, where ”id” and ”syn” are compulsory
attributes (See Figure 7):
• id: a name for the action in order to reference it.
• syn: the synset number of the verb in the lexicon (Wordnet).
6
Table 2: Example of nominal sentences in Arabic.
Arabic
English
Noun +
Adjective
Noun + Prep.
+ Noun
Noun + Active
participle + PP
. á ÖÞ ( àñºK) Ég. QË @
The man (is) fat.
ú¯ Ég. QË @
. ñË@
The man (is) in the market.
úÍ@ I. ë @ X Ég. QË @
. ñË@
The man is going to the market.
/al-Rajulu (yakūnu) samı̄nun./
/al-Rajulu fı̄ al-Sūqi./
/al-Rajulu dhahibun ilá al-Sūqi./
@act:[
act:{
id: <action_ID>;
syn: <synset>;
tns: <PA/PR/FU>;
prg: <Y/N>;
neg: <Y/N>;
mod: <CAN/MAY/MUST>;
agt: [<id>,...];
thm: [<id>,...];
@adv:[
adv:{
syn: <synset>;
adv: [synset, ...];
adv:}
adv:]
act:}
act:{
...
act:}
act:]
Figure 7: Actions representation.
• agt: a set of role IDs; those who did the action.
• thm: a set of role IDs; those who receive the action.
• A set of adverbs blocks that modify the verb. Each adverb block contains its synset and a set of adverbs
synsets.
• Some verb specifications: tense (”tns”), progression (”prg”), perfect aspect (prf), negation (”neg”) and
modality (”mod”).
The tense can be: past (”PA”), present (”PR”) or future (”FU”). The absence of tense means the action is
tense-free; e.g. ”To do so, you must try”. There are languages which doesn’t have future tense, such as Arabic and
Japanese. In this case, we can detect the tense using adverbs (”Tomorrow”) or temporal prepositional phrases (”in
the next year”). Other languages define (far/near) past and (far/near) future. For example, in Arabic, there is no
tense called ”future” but it can be expressed using auxiliaries. For near future, we use ” /sa/ ” attached to the verb
in present tense (” I
. ë X A /sa’dhhabu/ ”, I will go <soon>). For far future, we use ” ¬ñ /sawfa/ ” detached from
the verb in present tense (” I
. ë X @ ¬ñ /sawfa adhhabu/ ”, I will go <later>). But, since this can be detected using
adverbs such as ”soon” and ”later”, we can ignore it.
There are two aspects which are mostly used in occidental languages: progressive and perfect. The perfect
aspect refers to some actions prior to the time under consideration, which is viewed as already completed. In
STON representation, the tense is imperfect unless we add (”prf: Y;”). As for progressive aspect, it is a situation
7
Table 3: Example of Consecutive actions.
Language
Sentence
Arabic
úÍ@ Õç' Q» I.ë X
.PA®ÊJË@ YëA Õç' È Q ÜÏ @ úÍ@ XA« Õç' ñË@
English
French
Japanese
/dhahaba Karı̄m ilá al-Sūqi thumma ‘āda ilá al-manzili thumma shāhada al-Tilfāza./
Karim went to the market, came back home, then watched T.V.
Karim est allé au marché, il est revenu à la maison, ensuite il a regardé la télé.
カリムさんは市場に行って、家に戻って、テレビを目ました。
/Karimu-san wa itchiba ni itte, ie ni modotte, terebi o memashita./
@st:[
st:{
typ: <AFF/EXCL/QUES>;
act: [<act_id>,...|...]
st:}
st:]
Figure 8: Sentences representation in STON.
where a verb is (was) in motion for an interval of time. Likely, the action is not progressive unless we add (prg:
Y;).
Modality, in our case, can express possibility (”MAY”), admissibility (”CAN”) or obligation (”MUST”). The
modal verb ”will” is used to express the future, which is a tense in our case. Advice (”You should see a doctor”),
prohibition (”You mustn’t smoke here”), certainty (”He must be rich, since he lives there”), permission (”You can
leave now”) and lack of necessity (”You don’t have to do anything”) can be represented using these three modal
verbs.
3.3
Sentences
A Role-Action representation is not sufficient, since there are sentences which contains consecutive actions. In
Table 3, we can observe three actions: going to the market, coming back and watching T.V. If we represent
this sentence as three sentences instead, we will loose the information that these actions are consecutive. Not to
mention, we have to specify which actions are the main ones (there are actions which are relatives of roles and
actions).
The sentence part lists some sentences blocks. Each block contains the type of the sentence: affirmation
(”AFF”), exclamation (”EXC”), question (”QST”) and imperative (”IMP”). It have an attribute ”act” to list the
references of actions in this sentence. The annotation of sentences in STON is illustrated in Figure 8.
4
Relations
There are many relations between the clauses and phrases of a sentence. Attributes such as agents and themes in an
action can just represent simple sentences. A more complicated sentence can contain adpositional phrases, relative
clauses, etc. Figure 9 shows our view about the 4 relations between the roles and the actions.
A Role-Role relation can be expressed by adpositional relations as the example ”the man in the car”. Adpositional relations can, also, be used to express Action-Role relation (”he is in the car”). Another relation is the
relation Role-Action found in relative clauses (”the man who is driving”). Like adjectives, a relative clause can
modify (describe) a nominal phrase. A noun can be described by a relative clause, such in the sentence ”the man
who was strong”. The late example is different from ”the strong man” because the first sentence adds the information of a past quality of strength. The relation Action-Action can be found in adverbial clauses (he is where I can
see him). Comparison between two roles is another issue; Are they equal or one is more or less than the other?
They can share a verb such in ”I do more work than you” or an adjective ”I am stronger than you”. More structure
have to be added in order to allow these types of relations.
8
[(the man) IN (the car)]
[(he is) WHERE (I can see him)]
Adverbial
Adpositional
[(he is) IN (the car)]
Adpositional
Role
Action
[(the man) WHO (is driving)]
Relative
Figure 9: Relations between roles and actions
Language
Arabic
English
French
Japanese
4.1
Table 4: Example of relative clauses.
Sentence
. ZAÜÏ @ H. Qå ÑjÊË@ É¿ @ ø YË@ Ég. QË @
/al-Rajulu al-Ladhı̄ akala al-Lah.ma yashrabu al-Mā’./
The man who ate the meat is drinking water.
L’homme qui a mangé de la viande boit de l’eau.
肉を食べた男性が水を飲んでいます。
/niku o tabeta dansei ga mizu o nonde imasu./
Relative clauses
Relative clauses are a little challenging to be represented using our Role-Action notation. Our notation starts with
roles then actions. Since relative clauses are actions which describe a role, they have to be represented in the role
itself. In Table 4, the man drinking water, is the one who ate the meat earlier.
One proposition is to use ulterior referencing in roles to reference an action as a relative. Then, the relationship
between the role and the relative clause must be specified. For example, the text in Table 4 can be represented as
in Figure 10.
Looking at relative pronouns (See Table 5), we can consider 4 main types of relative phrases: subject, possessive, direct object and indirect object. In this categorization, we considered a person as same as a thing, because
the difference between them can be taken in consideration in text generation task.
• The subject type (”SBJ”) indicates that the main clause is a subject of the relative one. For example, in the
phrase ”the man who ate the meat”, the main clause ”the man” is the subject of the relative one ”ate the
meat”.
• The possessive type (”POS”) indicates that the noun in the relative clause is possessed by the main one. For
example, the phrase ”the man whose car is so expensive” describes a man who possess an expensive car.
• The object type (”OBJ”) indicates that the main clause is a direct object of the relative one. For example,
the phrase ”the man whom I saw yesterday” describes a man who is an object of me seeing him. We can
deduce the following information from it ”I saw the man yesterday”.
• The indirect object, or the clauses that starts with prepositions, are various. The meaning of these clauses
follows the meaning of their prepositions. In this case, we can form their types using the keyword (”IO ”)
followed by the Adpositional relation (next subsection). For example, ”The town from which I came” describes a town which is an indirect object of me coming. We can deduce the following information from
Table 5: Relative pronouns.
Person
thing
Place
Subject
Object
Possessive
who/that
who/whom/that
whose
which /that
which/that
whose
9
where
Time
Reason
when
why
@r:[
r:{
id: man;
syn: 10287213;
@rel:{
typ: SBJ;
ref: [ate]
rel:}
r:}
r:{
id: meat;
syn: 7649854;
r:}
r:{
id: water;
syn: 14845743;
r:}
r:]
@act:[
act:{
id: ate;
syn: 1168468;
tns: PA;
obj: [meat];
act:}
act:{
id: drink;
syn: 1168468;
tns: PR;
prg: Y;
agt: [man];
thm: [water];
act:}
act:]
Figure 10: Example of STON representation in case of relative clauses.
it ”I came from the town”. As for the relative adverbs ”where” and ”when”, we can replace them with a
preposition. For example, ”The town where I met him” is equivalent to ”The town in which I met him”.
• One other type is the reason (”RSN”), expressed by the relative adverb ”why”.
4.2
Adpositional phrases
An adpositional phrase includes prepositional phrases, postpositional phrases, and circumpositional phrases. Usually, they are used to describe the time or the place of an action. Arabic, English and French use prepositional
phrases, while Japanese uses postpositional phrases. They can be related to an action: ”He works at 8”, or a role:
”The mother of the boy”.
Our objective is to define as few relations as possible which can represent the meaning of most adpositions.
For that matter, we define these relations, which can be extended in the future:
• AGO: An amount of time back in the past. e.g. ”I bought it 2 years ago”.
• FRM: An origin or change of state; e.g. ”I came from Algeria”.
• IN: The existence of a particular time, place or situation. e.g. ”I wake up at 7 am.”, ”I was born in 1986”, ”I
am in town.”, ”I am at home.”
• SNC: A period from past till a particular time. e.g. ”I have worked since 2013”.
• TO: A destination; It can be a location, a person, etc. e.g. ”I am going to the market.”, ”I will give it to
him.”,”We waited till noon”.
• FOR: An amount of time or an objective. e.g. ”I will sleep for 2 hours”, ”the foundation for the support of
the cinema.”, ”during the war”.
• BEF: A time earlier than another (before) or a place in front of something. e.g. ”She’s always up before
down”, ”I sit in front the TV.”.
• AFT: A time further than another (after) or a place at the back of something (behind). e.g. ”He always sleep
after 10 pm”, ”the towel is behind the door”.
• BY: Not later than a specified time, or a place besides something or an agent; Here, we consider these
prepositions as the same: ”by”, ”next to”, ”besides” and ”near”. e.g. ”I will finish by 5pm.”, ”He walks by
the river.”, ”diffraction by crystals”.
10
• INS: Something inside something else. e.g. ”The present is inside the box.”.
• OUT: Something outside something else. e.g. ”I am outside the home”.
• BLW: A place under something; Here, we treat ”Below” and ”Under” as the same. e.g. ”He swims under
the bridge.”.
• ABV: A place above something; Here, we treat ”Above” and ”over” as the same. e.g. ”we walked over the
bridge.”.
• BTW: A place between two or more things. e.g. ”The museum is between two flats.”.
• THR: From one side to another, or surrounded; Here, we don’t make a difference between ”through” and
”across”. e.g. ”She walked through the forest.”.
• ON: A subject or something connected to something else. e.g. ”She borrowed a book about mathematics”.
• WTH: Being together or being involved. e.g. ”I am with him”.
• OF: Possession or belonging. e.g. ”The leafs of the tree”. This relation is used to represent compound
nouns: ”printer cartridge” will be ”cartridge of printer”.
• AS: A role. e.g. ”He works as a teacher.”.
• UND: A situation. e.g. ”He did the exam under a lot of pressure.”.
In English, ”at” and ”in” are used to express ”exact” and ”wide” locations or times respectively. In Japanese,
there is no difference between exact or wide range locations and times. But, there is another aspect which is used
to distinguish the prepositions ”に /ni/ ” and ”で /de/ ”. The first one (/ni/) is used to express the existence of
something or someone in a place. e.g. ”犬は公園にいます。 /inu wa kōen ni imasu./ ” (”The dog is in the
park.”). While the second one (/de/) is used to express the place where an action takes place. e.g. ”犬が公園で
吠えます。 ” /inu ga kōen de hoemasu./ (The dog barks in the park.). Unfortunately, our representation doesn’t
take these aspects into consideration. Nevertheless, we believe this can be solved in text generation task. In case of
”at-in” differentiation, we can test the noun to generate the right preposition. Likewise, for Japanese postpositions
”ni-de”, we can use the verb.
4.3
Adverbial clauses
An adverbial clause is a dependent clause which functions as an adverb. in our representation, it represents the
relation between an action and another. These are the relations used to express the adverbial clauses:
• WHN: A specific time. e.g. ”He listens when you talk.”.
• WHL: A period of time. e.g. ”He listens while you talk.”.
• WHR: A place. e.g. ”He started where he stopped.”.
• IF: A condition. e.g. ”He will come if he can.”.
• SO: A purpose. e.g. ”He tries so he can succeed.”.
• BCS: A reason. e.g. ”He can’t be angry with her because he likes her.”.
• THG: A concession. e.g. ”I will come although I don’t like traveling.”.
• LIK: A manner. e.g. ”He did the job as you asked.”.
• FTR: After (time). e.g. ”He start the job after he wake up.”.
• BFR: Before (time). e.g. ”Before we came here, the door was shut.”.
11
Table 6: Example of sentences with comparative.
Language
Arabic
English
French
Japanese
Sentence
È ñ£ @ Õç' Q» /Karı̄m atwalu min akhı̄h./
. éJk @ áÓ
.
É ¯ @ Y« A Õç' Q» /Karı̄m yusā‘idu aqalla min akhı̄h./
. éJk @ áÓ
Karim is taller than his brother.
Karim helps less than his brother.
Karim est plus grand que son frère.
Karim aide moins que son frère.
カリムさんは兄弟より背が高いです。 /Karimu-san wa kyōdai yori se ga takai desu./
カリムさんは兄弟より手伝うのが少ないです。 /Karimu-san wa kyōdai yori tetsudau no ga
sukunai desu./
@act:[
act:{
...
@cmp:[
cmp:{
typ: <L/M/LT/MT/EQ>;
adj: [<synset>, ...];
ref: [<id>,...|...];
cmp:}
cmp:]
act:}
act:]
Figure 11: Comparison block.
4.4
Comparison
Comparison is a little bit tricky; Mostly it includes an adjective which is shared between two roles, but also to the
action. For instance, in the sentence ”Karim is taller than his brother”, the adjective ”tall” is a property of ”Karim”
as it is of ”his bother”. We can represent the adjectives in the two roles, but the relation between them (superlative)
is better to be represented in the action ”To be”. Sure, the relation of comparison can be represented in one of
the roles ”Karim” and we can reference the other ”his bother”. But, there are cases where the comparison doesn’t
include adjectives; it is about the verb instead. For example, the sentence ”Karim helps less than his brother”
contains a comparison over the action and not the adjective. In the end, the comparison must be represented in the
action rather than the role. Table 6 contains the two examples quoted previously. We must point out that the Arabic
example for the sentence ”Karim helps less than his brother” is not so fluent. In order to be fluent, the sentence
must be ”. éJk @ áÓ è Y«AÓ É¯ @ Õç' Q» /Karı̄m aqallu musā‘adatan min akhı̄h./ ”. This can be literally translated as
”Karim is less helpful than his brother”. In Japanese example, the direct translation is ”Karim is less in help than
his brother”. Also, we can use ”カリムさんは兄弟よりも少ないながらも手伝いします。 /Karimu-san wa
kyōdai yorimo sukunai nagara tetsudaishimasu./ ”, which means literally ”Karim helps despite this is less than his
brother do.”; A form which is more polite. What matters for us is the comparison itself, when it means ”less than”
we represent it in the ”cmp” block. Then, when it comes to text generation task, each language handles how it has
to be generated fluently.
The idea is to use a block ”cmp” in the action block (see Figure 11). There are three types of comparison:
comparative, superlative and equality. If we add less comparison, we will have five types: less (”L”) and more
(”M”) for comparative, least (”LT”) and most (”MT”) for superlative and equal (”EQ”). The first parameter of
the comparison would be the agent itself, and the second is a reference in the comparison block. The STON
representations of Table 6 examples are illustrated in Figure 12.
12
@act:[
act:{
id: be_tall;
syn: 2604760;
tns: PR;
@cmp:[
cmp:{
typ: M;
adj: [2385102];
ref: [brother];
cmp:}
cmp:];
agt: [karim];
act:}
act:]
@act:[
act:{
id: help;
syn: 2547586;
tns: PR;
@cmp:[
cmp:{
typ: L;
ref: [brother];
cmp:}
cmp:];
agt: [karim];
act:}
act:]
Karim is taller than his brother.
Karim helps less than his brother.
Figure 12: Example of STON representation of comparison.
5
Some issues and their solutions
With the Role-Action-Sentence notation, we can represent a lot of sentences. But, there are some ambiguities
when we want to represent some others. To handle this, we will show some problems and their solutions in context
of STON.
5.1
Coordination between references
One problem is, how to represent the coordination between references. There are, principally, two main coordinations which are disjunction (”or”) and conjunction (”and”). To limit ambiguity when representing a sentence, we
want to use either disjunctions of conjunctions or the inverse. For example:
• Mother and son ate food.
• Father and son ate food.
These two sentences can be aggregated in two ways:
1. Mother and son or father and son ate food. (disjunctions of conjunctions);
2. Mother or father and son ate food. (conjunctions of disjunctions).
The first one sounds more appropriate than the second. Moreover, the second sounds like either the mother alone
or the father and the son ate the food. So, the notation of agents and themes must be changed. Taking the late
example, the agents in the sentence can be represented as ”agt: [mother, child | father, child];”
5.2
Proper names
A proper name is a phrase that identifies one unique entity from a class of entities. For example, ”London” is
distinguished from the common noun ”city”; It is more specific. Proper names can be people (Eg. ”Abdelkrime
Aries”), locations (Eg. ”Algeria”), organizations (Eg. ”ESI”), etc.
The problem with proper names is: how to represent them inside the role? STON is heavily based on Wordnet’s
synsets (or any other lexicon), each word must have one synset to be represented. Some proper names already have
synsets in Wordnet, such as U.S. cities. But, a city such as ”Jijel” for example, which is an Algerian city, has no
synset number in Wordnet. There are many other proper names, such as persons’ names, which doesn’t exist in
Wordnet. One solution is to add an attribute ”nam” to the role section, which contains the named entity.
In STON, a role must always have a synset or a pronoun. To afford much information about the role (Here, the
proper name), we can specify its hypernym’s synset. For instance, the sentence ”Karim lives at Jijel” (See Figure
13) has two proper names: ”Karim” and ”Jijel”. For each, we afford the synset of ”person” and ”city” respectively.
13
@r:[
r:{
id: karim;
syn: 7846;
nam: Karim;
r:}
r:{
id: jijel;
syn: 8524735;
nam: Jijel;
r:}
r:]
@act:[
act:{
id: lives;
syn: 2649830;
tns: PR;
agt: [karim];
@rel:[
rel:{
typ: IN;
ref: [jijel];
rel:}
rel:]
act:}
act:]
@st:[
st:{
typ: AFF;
act: [lives];
st:}
st:]
Figure 13: STON representation of the sentence ”Karim lives at Jijel”.
Table 7: Example of sentences with two objects.
Language Sentence
Arabic
English
French
Japanese
5.3
ù¢« @
. éK Yë É®¢Ë@
Ég. QË@
/a‘t.á al-Rajulu al-T.ifla hadiyyatan./
The man gave the boy a gift.
L’homme a donné un cadeau à l’enfant.
男性は少
少年に 贈り物があげました。
/dansei wa shōnen ni okurimono ga agemashita./
Verbs with two objects
In languages like Arabic and English, a verb can have two objects without using any preposition. Table 7 represents
an example of the sentence ”the man gave the boy a gift”. The bold phrase in each language represents the indirect
object. We can use the prepositions ”to” and ” Ë /li/ ” with the indirect object. The English sentence will be ”The
ù¢« @ /a‘t.á al-Rajulu
man gave a gift to the boy.”. Similarly, the Arabic sentence will be ”. É®¢ÊË éK Yë Ég. QË@
hadiyyatan li-al-T.ifli./ ”. So, we can use adpositional phrases to represent the relation with the indirect object.
5.4
Pronouns
In a first time, we thought to use references to the original role instead of personal pronouns; In generation phase,
we can generate them when the role is referred many times. This is can be applied in normal situations when
we have all the information including the anaphoric relations. In situations when we want to represent sentences
generated from extractive summarization for example, we may have a lot of trouble recovering these types of
relations. Also, to boost the analysis (from natural language to STON) and generation (from STON to natural
language) tasks, the pronouns are a need.
The pronouns can be classified using many features; Table 8 represents the eight features used for pronouns
classification according to Seah and Bond [2014]. Based on these features, we present the pronouns using two
attributes:
• typ: the type of the pronoun which is encoded on 6 characters:
– 1st: ”D” for demonstrative pronoun (this, that, etc.); ”S” for subjective personal pronoun (I, he, etc.);
”O” for objective personal pronoun (me, him, etc.) and ”P” for possessive pronoun (my, his, etc.).
– 2nd: ”F” for first person; ”S” for second person and ”T” for third person.
– 3rd: ”S” for singular; ”D” for dual; ”P” for plural and ”N” for not defined number.
14
Head
Table 8: Pronouns features accoding to [Seah and Bond, 2014].
Demonstratives
Entity
Time
Manner
Person
Place
Reason
Thing
Personal (1e, 1i, 2, 3)
Quantifier
Number
Gender
Case
Type
Formality
Politeness
Proximity
Dual
Plural
Singular
Feminine
Masculine
Neuter
Objective
Possessive
Subjective
Assertive
Elective
Negative
Other
Reciprocal
Universal
Interrogative
Reflexive
Formal
Informal
Polite
Distal
Medial
Proximal
r:{
id: his_novel;
syn: 3833065;<novel>
typ: PTSMFN;<his>
qnt: O1;<first>
r:}
Figure 14: STON representation of the role ”his first novel”.
– 4th: ”F” for female; ”M” for male and ”N” for neuter. Even if the source language attribute sex to
objects such as Arabic and French, we consider them as neuter. For example, in Arabic, the chair is
masculine while it is feminine in French.
– 5th: This is reserved for formality and politeness. ”R” for rude; ”C” for casual; ”F” for formal and
”P” for polite. In languages where there is no formality level in pronouns, we choose to use formal by
default.
– 6th: The proximity can be: ”D” for distal; ”M” for medial; ”P” for proximal and ”N” for not defined.
• ref: the reference(s) to the role(s) related to this pronoun.
The pronouns attribute can be accompanied with a synset to afford a more compact format. For instance,
Figure 14 represents a case where a pronoun and a noun are packed together. The pronoun ”his” is represented
as ”PTSMFN” which means: possessive, third person, singular, masculine, formal and without proximity. This
representation is better than transforming the clause to ”first novel of him” before being represented.
5.5
Passive voice
The passive voice hasn’t a specific representation in STON. A sentence like ”Karim ate the apple” (See Table 9)
has the same representation as ”The apple was eaten by Karim”. The choice of using active or passive voice is
decided by the text generation task. As for sentences which doesn’t have an agent such as ”An apple was eaten”,
we simply don’t define an agent in the representation. in this case, when we generate the sentence, we have to use
passive voice.
Table 9: Example of sentences in passive voice.
Language Sentence
Arabic
I Ê¿ @ ékA® JË @
. Õç' Q» ¬ Q£ áÓ
English
French
Japanese
/al-tuffah.ah ukilat min t.arafi Karı̄m./
The apple was eaten by Karim.
La pomme a été mangée par Karim.
林檎がカ
カリ ムさ ん に 食べられました。
/Ringo ga Karimu-san ni taberaremashita./
15
@act:[
act:{
id: represent;
syn: 988028;
tns: PR;
agt: [ston];
thm: [sentence];
act:}
act:{
id: hope;
syn: 1811441;
tns: PR;
agt: [karim];
thm: [represent];
act:}
act:]
Figure 15: Actions of the sentence ”Karim hopes that STON represents sentences”.
5.6
Complementizer phrase
A complementizer phrase can be a subject or an object. In our representation, we rather consider it as an agent or a
theme. Figure 15 represents an example of STON annotation in case of complementizer phrases. Here, the phrase
”STON represents sentences” is a theme and ”Karim” is the agent of the action ”Hoping”.
The structure ”VERB + TO + INFINITIVE” will be solved as a complimentizer. A sentence like ”I want to
go there” can be seen as if ”to go there” is the theme of ”wanting”.
6
STON tools and corpora
STON notation is intended to be used in NLP applications where the concern is about the sentence syntax in a
multilingual context. It can be used in language generation as an intermediate language for machines. This implies
that we can use it as a mean for text translation, text summarization, etc. Parsing STON language would be fast
since its grammar is well defined (blocks, references, etc.). The most challenging task is to create tools to parse a
text into STON and to generate a text from STON. Corpora for testing is a must, this is why we started to annotate
some materials already annotated by UNL.
6.1
Annotation process
Annotating some materials can be helpful in the future, especially when we intend to use STON for other applications. Since some grammatical structures are not able to be represented before transformation, like the apposition
(which can be replaced by ”which is”). Here, some steps to follow in order to have a good annotation in STON:
• Begin to represent roles (Nominal phrases) such as the dependent role must be at last. For example, in the
sentence ”the statue of liberty” we represent the role ”liberty” then the role ”statue” which has a relation
”OF” to the first one.
• Adjectives can’t be represented alone. For predicative adjectives, create a role with the noun in the subject
and the adjective. For example, ”The man is friendly” would be represented as ”The man is friendly man”.
• As for proper names, if they exist in Wordnet just put their synset (e.g. ”Cairo”). If they don’t, put them as
value of the attribute ”nam” (spaces must be transformed to underscores) and afford the synset of their type:
city, person, animal, dog, cat, etc. (the more specific the type, the better).
• Transform enumerations of a role to a relation ”OF”. For example, the sentence ”He did many jobs: web
designer, engineer and teacher” will be ”He did many jobs of web designer, engineer and teacher”.
• When you find some prepositions we didn’t talk about, try to find a similar in the relations. For example,
The preposition ”As well as” can be considered as a simple ”and”.
• Omit the chronological order indicators (first, then, finally) and represent it as consecutive actions.
• Try to transform expressions in order to fit the STON representation. For example: ”much of it experimental”
will be ”which is much experimental”.
• Comparison doesn’t exist in roles, it just exists in actions. To solve this problem, we add the expression ”an
amount which is ”; the comparison must not have an adjective. For example, ”the author of more than 200
articles” will be ”the author of an amount which is more than 200 articles”.
16
Table 10: Statistics on the annotated texts.
Naguib Mahfouz Bio
Louis de Broglie Bio
Original texts statistics:
# Sentences
# Words
# Words/Sentence
10
287
29
35
1055
30
UNL statistics:
# UWs with redundancy
# Relations
# Attributes
264
172
229
# Synsets with redundancy
# Roles
# Actions
# Relations
132
85
23
61
803
507
703
STON statistics:
373
237
82
183
We started to annotate some biographies3 already annotated by UNL. You can check the annotated texts on
Nolporas project4 which aims to create corpora mostly for STON. Table 10 represents some statistics on the two
biographies we annotated. The biographies contains long sentences with an average of 29 words. The synsets in
STON are the equivalent of UWs in UNL; here the number is less because we use references to roles. Roles sometimes are repeated in other sentences, so we don’t need to repeat their descriptions. Relations in UNL corresponds
to the relations (adpositions, relatives and adverbials) and the references of themes and agents. Concerning the
manual annotation, each sentence took us from a half to one hour since we have to choose between different senses
and the right relations.
The annotation process must be automatic which will be more interesting. We started working on another
project called NaLanPar5 (Natural language parser) which aims to generate sentence representations, including
STON, from texts. It uses open source text parsers such as Stanford Parser [Chen and Manning, 2014], which is a
syntactic parser for many languages (English, Arabic, etc.) licensed under GPL license. Currently, we are working
on transforming English text to STON, and more languages are to be added after finishing. At the mean time, it
can handle the sentences of form ”Subject Verb Object Preposition Noun”.
6.2
STON parsing
We created a parser for STON in a project called SentRep6 (sentence representation). Its aim is to implement
parsers for different sentence representation languages, taking STON as primary focus. The parser is an abstract
class which can be extended to implement the cases where the parser finds a role, adjective, action, etc.
To test the speed of parsing, we used a PC with i7 processor (2 GHz 4 cores) and openJDK 7. The parser
do nothing when it finds the different parts (just calls the functions which will treat them). For each biography
(NaguibMahfouz and LouisdeBroglie) we executed the parser 10 times and calculated the time of parsing in milliseconds. NaguibMahfouz biography parsing took between 60ms and 71ms with an average of 67ms, and almost
8ms per sentence. As for LouisdeBroglie biography parsing, it took between 132ms to 176ms with an average of
163ms, and almost 5ms per sentence.
6.3
Generation
To generate sentences from STON representation, we are currently working on another project called NaLenGen7
(Natural language generator). NaLenGen aims to generate text from different sentence representations (It uses
SentRep for that). It uses open source text realizers such as SimpleNLG [Gatt and Reiter, 2009], which is a
3 http://www.undl.org/unldoc/bb.htm
4 Nolporas
project: https://github.com/kariminf/NaLanPar
source: https://github.com/kariminf/NaLanPar
6 SentRep source: https://github.com/kariminf/SentRep
7 NaLenGen source: https://github.com/kariminf/NaLanGen
5 NaLenPar
17
Source
Table 11: Example of sentences generation to English and French.
Born in Cairo in 1911, Naguib Mahfouz began writing when he was seventeen. His first
novel was published in 1939 and ten more were written before the Egyptian Revolution of
July 1952, when he stopped writing for several years.
English generated text
Naguib Mahfouz which was given birth in Cairo in 1911 began writing when he was 17
years. First, his novel was published in 1939 and 10, more novels were written before the
revolution of Egyptians of July 1952 in which he discontinued writing for several years.
French generated text
Naguib Mahfouz que a été accouché à un Le Caire à 1911 a débuté un œuvre quand lui a été
de 17 années. Son premier nouveau a été publié à 1939 et de 10 nouveaux plus ont été écrits
avant le tour de des égyptiens de July 1952 à lequel lui a cessé un œuvre pour des années
es.
realization engine for English licensed under MPL license. It has been adapted to many other languages: German
[Bollmann, 2011], French [Vaudry and Lapalme, 2013] and Brazilian Portuguese [de Oliveira and Sripada, 2014].
To map Wordnet synsets to other languages, we use Open Multilingual Wordnet [Bond and Foster, 2013]. The
mapping is not complete for many languages, as a result the generation may fail when the synset is not found.
We tried to generate some English and French text from STON annotation. Examine Table 11 which is an
example of the two first sentences of Naguib Mahfouz biography. The resulted text was fair for English and a little
bad for French. This is why more work has to be done to address some issues we have found:
• Sometimes the mapping to other languages is done automatically which leads to some errors in translating
concepts. For example, we found that the concept ”Cairo” was mapped to ”OSS 117 : Le Caire” in French.
As we can see in the example, many other concepts are wrong.
• The synsets contain many words, and some words are more adequate than others. This is why we have to
propose a method in order to choose these words based on their frequency of use for example.
• We can’t generate directly from STON, and some changes must be done to have more fluent texts. For
example, predicative adjectives are expressed as a role with a noun. The sentence ”The child is happy” is
represented as ”The child is happy child”.
7
Limitations and challenges
Although STON can represent a wide range of sentences, it shows some limitations. Sometimes, the variation
between languages can prevent us from representing the sentences properly. In some languages, we don’t always
find the same syntactic representation of a meaning. For instance, considering Table 12, the adjective ”hungry” has
no exact translation in Japanese. In fact, the Japanese sentence literally means ”my stomach emptied”. The same
adjective is translated to a noun in French, where the sentence literally means ”I have a hunger”. In Arabic, we
ª k. A K @ /anā ju‘t/ ”, which means the same thing. This sentence is composed of the
can use the sentence ” . I
pronoun ”I” and a verb ”to be hungry” conjugated in the past. Also, since STON is based on Wordnet synsets, it
is limited to the concepts extracted from English language. So, if we want to represent a sentence from Arabic for
example, we have to find Wornet concepts that are close to the meaning of the sentence’s words.
STON can’t deal with text structures like paragraphs, lists, tables, etc. Many languages have complex predicates
which can represent the same meaning as a singleton verb. For instance, ”I gave the baby a bath” and ”I bathed the
baby” have the same meaning. Unfortunately, STON is not a fully semantic representation which means it handles
those as two different forms. A sentence like ”the purpose is to teach mathematics and develop physics” can be
represented without problem. But, in case of this sentence: ”the purpose is to teach and develop physics”, we have
to use redundancy and represent it as ”the purpose is to teach physics and develop physics”.
8
Discussion
STON is all about representing the different parts of a sentence independently from its structure in natural languages. It is meant to transport sentences information between different applications (programs). It represents the
syntactic relations between the different parts of sentences in a multi-lingual way. When the syntax fails to keep
the multilingual aspect, it uses semantic relations instead. To understand STON better, we have to know what this
language is not about.
18
Table 12: Example of sentences which doesn’t have the same POS in different languages.
Language Sentence
Romanization
Arabic
English
French
Japanese
Criteria
/anā jaw‘ān./
. àA«ñk. AK @
I am hungry.
J’ai faim.
お腹が空いた。
/onaka ga suita./
Table 13: Comparaison between STON and other well-known annotation formats.
KANT
UNL
AMR
STON
Objective
Interlingua
for
machine
translation from English
technical manuals.
Represent the meaning of texts without ambiguity, to be
used as a language
of the web.
Write down the
meanings of English
sentences.
Represent sentences
in a multilingual
way without basing so much on
semantics, to be
used as interchange
format
between
applications.
Aspects
Semantics,
with
morphological
aspects (tense, etc.)
Morphology,
Semantics, Pragmatics
Semantics, no morphological aspects
(tense, etc.)
Syntax, with morphological aspects
(tense, etc.)
and
semantics
when
syntax fails to be
multilingual
Concepts
Own
UWs (UNL Ontology)
PropBank frames
WordNet synsets
Dependency
Domain dependent
Language independent
English
language
dependent
Language independent
Relations
Own
Own
PropBank relations
Own
Readability
Readable
Difficult to follow
Less readable
Less readable, difficult when we have a
big text
• It does not represent the relations between the parts of speech semantically, even if there are some relations
in relative clauses. For example, it does not allow us to represent relations like UNL does, such as beneficiary
relation and purpose relation.
• It is not a format for storing texts, such as Open document format (ODF) and Microsoft office format which
are based on XML.
There are similarities and differences between STON and other representation languages. A concise comparison
between KANT, UNL, AMR and STON is given in Table 13.
When we check the meaning, UNL is the representation that mostly represents the meaning. AMR, in the
other hand, is used to represent the meaning but it lacks some morphological aspects such as verb tense. KANT
and STON annotations are less depending on semantics than the two previous ones. Our objective is to represent
sentences with a minimum cost (time and processing effort). This is why we represent relations like ”from” as
they are, even if they can mean many things: ”from 2 am”, ”from London”, etc.
Semantic representation is a powerful tool because it allows us to represent what could be understood from a
sentence. It can eliminate redundancy in sentences; For example [Banarescu et al., 2013]:
• The soldier was afraid of battle.
• The soldier feared battle.
• The soldier had a fear of battle.
19
act:{
id: was;
syn: 2604760;
tns: PA;
agt: [soldier];
thm: [afraid-soldier];
@rel:[
rel:{
typ: OF;
ref: [battle];
rel:}
rel:]
act:}
act:{
id: feared;
syn: 1780729;
tns: PA;
agt: [soldier];
thm: [battle];
act:}
act:{
id: had;
syn: 121046;
tns: PA;
agt: [soldier];
thm: [fear];
@rel:[
rel:{
typ: OF;
ref: [battle];
rel:}
rel:]
act:}
Figure 16: STON action representation of 3 sentences with same meaning.
These sentences have the same meaning, therefore their representation must be the same. In contrast with UNL
and AMR, STON doesn’t go deep into the semantic relations, or to comprehend the sentence as a whole. Figure
16 shows STON representation of these 3 sentences. It is clear that the representation makes difference between
the verbs, nouns and adjectives.
The four languages are based on different ontologies and lexicons to represent their concepts. KANT annotation is based on concepts defined especially for the KANT system, generally limited by technical reports domain.
Likewise, UNL defines its own concepts base called UNL ontology, and each concept is referred to as a universal
word (UW). AMR and STON use PropBank and Wordnet respectively, this is why they are limited to these two
bases.
AMR and KANT are more based on English, while UNL and STON seeks to be multilingual. When it comes
to multilingual aspect, the UWs of UNL are so powerful. For the mean time, we use Wordnet to represent the
different concepts. Unfortunately, till nowadays, the mapping to other languages is not complete. For this reason,
we had some problems in generating French text from STON annotation when the synset can’t be found.
Readability is important when we want to create a sentence representation manually or to check it after automatic generation. KANT annotation is more readable than the other three representations. STON is developed as a
machine language such as UNL, even so, we want to allow some space for readability. This can be helpful in case
we want to test a system that uses STON, because it will be easier to create test banks.
9
Conclusion
In this work, we proposed an inter-application language (STON) aimed to represent sentences structures in a
multilingual context. STON is based somehow on the JSON representation with some adjustments to speed up
the parsing. It is based on the assumption that anything in the sentence is either a role or an action with relations
between them. The representation uses both syntactic structure (noun definition, verb tense, subjects, objects,
etc) and semantic relations (time and place relations, etc.). To support multilingualism, we use concepts instead
of words (in our case, we use Wordnet’s synsets). Our intention is to use STON as a mean of communication
between different applications. More specifically, the language is intended to be used in cross-lingual automatic
text summarization.
STON is far beyond being complete or being perfect. There still are some improvements to be made in the
future, such as the problem of words’ syntactic alignment inter-languages. Because it is based on Wordnet, the
concepts are limited to English. There are a lot of concepts that doesn’t exist in English but exist in other languages.
Exploiting a larger semantic network with a more knowledge base like BabelNet8 [Navigli and Ponzetto, 2012]
may improve inter-languages representation.
8 http://babelnet.org
20
Acknowledgement
Special thanks to Hisham Omar for his valuable feedback concerning Japanese examples.
References
Manuel Bertran, Oriol Borrega, Marta Recasens, and Bàrbara Soriano. Ancorapipe: A tool for multilevel annotation. 41:291–292, 2008. ISSN 1135-5948.
Marta Recasens and M. Antònia Martı́. Ancora-co: Coreferentially annotated corpora for spanish and catalan. Language Resources and Evaluation, 44(4):315–345, 2009. ISSN 1574-0218. doi: 10.1007/s10579-009-9108-x.
URL http://dx.doi.org/10.1007/s10579-009-9108-x.
Markéta Lopatková, Petr Homola, and Natalia Klyueva. Annotation of sentence structure. Language Resources
and Evaluation, 46(1):25–36, 2011. doi: 10.1007/s10579-011-9162-z. URL http://dx.doi.org/10.1007/
s10579-011-9162-z.
George A. Miller. Wordnet: A lexical database for english. Commun. ACM, 38(11):39–41, November 1995. ISSN
0001-0782. doi: 10.1145/219717.219748. URL http://doi.acm.org/10.1145/219717.219748.
Hiroshi Uchida, Meiying Zhu, and Tarcisio Della Senta. The unl, a gift for a millennium. 1999. URL http:
//www.undl.org/publications/gm/index.htm.
Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffitt, Ulf Hermjakob, Kevin Knight, Philipp
Koehn, Martha Palmer, and Nathan Schneider. Abstract meaning representation for sembanking. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse, pages 178–186, Sofia, Bulgaria, August 2013. Association for Computational Linguistics. URL http://www.aclweb.org/anthology/
W13-2322.
Teruko Mitamura, Eric H Nyberg, and Jaime G Carbonell. An efficient interlingua translation system for multilingual document production. In Proceedings of the Third Machine Translation Summit, 1991.
Krzysztof Czuba, Teruko Mitamura, and Eric H Nyberg. Can practical interlinguas be used for difficult analysis
problems? In Proceedings of AMTA-98 Workshop on Interlinguas, 1998.
Hiroshi Uchida and Meiying Zhu. Unl2005 from language infrastructure toward knowledge infrastructure. Special
Speech, Pacific Association for Computational Linguistics (PCLING 2005), 2005.
Igor Boguslavsky. Some lexical issues of unl. Universal Networking Language: Advances in Theory and Applications, pages 101–108, 2013.
R. Martins. Lexical Issues of UNL: Universal Networking Language 2012 Panel. EBSCO ebook academic collection. Cambridge Scholars Publishing, 2013. ISBN 9781443852814. URL https://books.google.dz/
books?id=tdcwBwAAQBAJ.
Paul Kingsbury and Martha Palmer. From treebank to propbank. In Language Resources and Evaluation, 2002.
M. Haak. The Verb in Literary and Colloquial Arabic. Functional grammar series. Mouton de Gruyter, 1997.
ISBN 9783110154016. URL https://books.google.dz/books?id=I21G6qVQibkC.
Yu Jie Seah and Francis Bond. Annotation of pronouns in a multilingual corpus of mandarin chinese, english and
japanese. In 10th Joint ACL - ISO Workshop on Interoperable Semantic Annotation, pages 82–87, Reykjavik,
Iceland, 2014.
Danqi Chen and Christopher Manning. A fast and accurate dependency parser using neural networks. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 740–
750, Doha, Qatar, October 2014. Association for Computational Linguistics. URL http://www.aclweb.org/
anthology/D14-1082.
Albert Gatt and Ehud Reiter. SimpleNLG: A realisation engine for practical applications. In Proceedings of the
12th European Workshop on Natural Language Generation (ENLG 2009), pages 90–93, Athens, Greece, March
2009. Association for Computational Linguistics. URL http://www.aclweb.org/anthology/W09-0613.
21
Marcel Bollmann. Adapting simplenlg to german. In Proceedings of the 13th European Workshop on Natural Language Generation, pages 133–138, Nancy, France, September 2011. Association for Computational Linguistics.
URL http://www.aclweb.org/anthology/W11-2817.
Pierre-Luc Vaudry and Guy Lapalme. Adapting simplenlg for bilingual english-french realisation. In Proceedings
of the 14th European Workshop on Natural Language Generation, pages 183–187, Sofia, Bulgaria, August 2013.
Association for Computational Linguistics. URL http://www.aclweb.org/anthology/W13-2125.
Rodrigo de Oliveira and Somayajulu Sripada. Adapting simplenlg for brazilian portuguese realisation. In Proceedings of the 8th International Natural Language Generation Conference (INLG), pages 93–94, Philadelphia,
Pennsylvania, U.S.A., June 2014. Association for Computational Linguistics. URL http://www.aclweb.
org/anthology/W14-4412.
Francis Bond and Ryan Foster. Linking and extending an open multilingual wordnet. In Proceedings of the 51st
Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1352–1362,
Sofia, Bulgaria, August 2013. Association for Computational Linguistics. URL http://www.aclweb.org/
anthology/P13-1133.
Roberto Navigli and Simone Paolo Ponzetto. BabelNet: The automatic construction, evaluation and application of
a wide-coverage multilingual semantic network. Artificial Intelligence, 193:217–250, 2012.
22
| 2cs.AI
|
arXiv:1604.01551v1 [math.AC] 6 Apr 2016
Stably co-tame polynomial automorphisms
over commutative rings
Shigeru Kuroda∗
Abstract
We say that a polynomial automorphism φ in n variables is stably
co-tame if the tame subgroup in n variables is contained in the subgroup generated by φ and affine automorphisms in n + 1 variables. In
this paper, we give conditions for stably co-tameness of polynomial
automorphisms.
1
Introduction
Let R be a commutative ring of characteristic p 6= 1, R[x] := R[x1 , . . . , xn ]
the polynomial ring in n variables over R, and GAn (R) := AutR R[x] the
automorphism group of the R-algebra R[x]. We identify each φ ∈ GAn (R)
with the n-tuple (φ(x1 ), . . . , φ(xn )) of elements of R[x]. The composition is
defined by
φ ◦ ψ = (g1 (f1 , . . . , fn ), . . . , gn (f1 , . . . , fn ))
(1.1)
for φ = (f1 , . . . , fn ), ψ = (g1 , . . . , gn ) ∈ R[x]n . For each r ≥ 1, we regard
GAn (R) as a subgroup of GAn+r (R) by identifying each φ ∈ GAn (R) with
the unique extension φ̃ ∈ GAn+r (R) of φ defined by φ̃(xn+i ) = xn+i for
i = 1, . . . , r.
We say that φ ∈ GAn (R) is affine if φ = (x1 , . . . , xn )A + b for some
A ∈ GLn (R) and b ∈ Rn , and set Aff n (R) := {φ ∈ GAn (R) | φ is affine}. We
define
ǫ(f ) := (x1 + f, x2 , . . . , xn ) ∈ GAn (R) for each f ∈ R[x̂] := R[x2 , . . . , xn ],
and set En (R) := {ǫ(f ) | f ∈ R[x̂]}. We call Tn (R) := hAff n (R), En (R)i
the tame subgroup, and elements of Tn (R) are said to be tame. Here,
∗
Partly supported by JSPS KAKENHI Grant Number 15K04826
1
for subsets S1 , . . . , Sr and elements g1 , . . . , gs of a group
S G, we denote by
hS1 , . . . , Sr , g1 , . . . , gs i the subgroup of G generated by ri=1 Si ∪ {g1 , . . . , gs }.
If n ≥ 3 and R contains Q, then Tn (R) = hAff n (R), σi holds for σ = ǫ(x22 )
by Derksen (cf. [Es, Thm. 5.2.1]). We remark that Derksen’s theorem requires
that R is generated by R∗ as a Q-vector space, but this assumption is in
fact unnecessary (cf. §3). When R is a field and p = 0, Bodnarchuk [B]
proved a similar result for more general σ’s. The situation is different if p is
a prime. Maubach-Willems [MW] showed that T3 (F2 ) 6= hAff 3 (F2 ), ǫ(x22 )i,
and conjectured that, if n ≥ 3 and R is a finite field, then no finite subset E
of Tn (R) satisfies Tn (R) = hAff n (R), Ei.
Edo [Ed] found a class of φ ∈ GAn (R) for which hAff n (R), φi contains
(R).
Such φ is said to be co-tame. If R is a field, no element of GA2 (R)
Tn
is co-tame thanks to Jung [J] and van der Kulk [K]. For n ≥ 3, it is easy to
find elements of GAn (R) \ Aff n (R) which are not co-tame if R is not a field
(cf. (4.1)), or if p is a prime (cf. [EK, §1 (d)]). In the case where n ≥ 3, R is
a field and p = 0, the first example of such an automorphism was found by
Edo-Lewis [EL] for n = 3.
Recall that φ ∈ GAn (R) is said to be stably tame if φ belongs to Tn+1 (R).
It is known that some non-tame automorphisms are stably tame (cf. [BEW],
[N], [SU], [S]). The following is an analogue to the stably tame automorphisms.
Definition 1.1. We say that φ ∈ GAn (R) is stably co-tame if hAff n+1 (R), φi
contains Tn (R), or equivalently hAff n+1 (R), φi contains En (R).
Clearly, co-tame automorphisms are stably co-tame. When R is a field,
there exist elements of T3 (R) which are not co-tame but stably co-tame in
both cases p = 0 and p > 0 (cf. §6.1). The purpose of this paper is to
study when elements of GAn (R) are stably co-tame or not. If R contains
an infinite field, we have a necessary and sufficient condition for stably cotameness (Corollary 2.4).
This paper is organized as follows. The main results are stated in Section 2, and three key results are proved in Sections 3, 4 and 5. In Section 6,
we study stably co-tameness of the example of Edo-Lewis. We also discuss
a technique which is useful when R does not contain an infinite field.
2
Main results
Since T1 (R) = Aff 1 (R), we always assume that n ≥ 2 unless otherwise
stated. Take any φ ∈ GAn (R). We define Mφ to be the R-submodule of
2
R[x] generated by 1, xi and η(φ(xi )) for i = 1, . . . , n and η ∈ Aff n (R). By
definition, we have
η(Mφ ) ⊂ Mφ
for each η ∈ Aff n (R).
(2.1)
The following theorem holds for any commutative ring R of characteristic
p 6= 1.
Theorem 2.1. φ ∈ GAn (R) is stably co-tame in the following four cases:
(a) Mφ contains xi xj for some 1 ≤ i < j ≤ n.
(b) 2 is a unit of R, and Mφ contains x2i for some 1 ≤ i ≤ n.
(c) n = p = 2 and Mφ contains x21 x2 or x1 x22 .
(d) n = p = 2, Mφ contains x31 or x32 , and there exists ξ ∈ R∗ satisfying
ξ + 1 ∈ R∗ .
Next, assume that p is zero or a prime. For a ∈ R and t1 , . . . , tn ≥ 0, we
call axt11 · · · xtnn a good monomial in the following five cases:
(I) p = 0 and t1 + · · · + tn ≥ 2.
(II) p ≥ 2 and there exist 1 ≤ i < j ≤ n such that ti , tj ≡ 1 (mod p).
(III) p ≥ 3 and there exists 1 ≤ i ≤ n such that ti 6≡ 0, 1 (mod p).
(IV) n = p = 2 and there exist i, j ∈ {1, 2} such that ti ≡ 1 and tj ≡ 2
(mod 4).
(V) n = p = 2 and there exists 1 ≤ i ≤ 2 such that ti ≡ 3 (mod 4).
For each f ∈ R[x], let Cf denote the set of the coefficients of good
monomials
appearing in f . We define Iφ to be the ideal of R generated
Sn
by i=1 Cφ(xi ) .
Theorem 2.2. Assume that p is zero or a prime. If φ ∈ GAn (R) satisfies
Iφ 6= R, then φ is not stably co-tame.
Throughout this paper, let k be a field. When R is a commutative kalgebra, we say that f ∈ R[x] satisfies the degree condition if
degsxi f ≤ #k − 2 for i = 1, . . . , n.
(2.2)
Here, degsxi f denotes the separable degree of f if p > 0, and the standard
degree degxi f of f if p = 0, as a polynomial in xi . The separable degree of
a polynomial g(x) ∈ R[x] in one variable is defined as the degree of h(x) ∈
e
R[x] \ R[xp ] such that h(xp ) = g(x) for some e ≥ 0. We say that φ ∈
GAn (R) satisfies the degree condition if φ(xi ) satisfies the degree condition
for i = 1, . . . , n.
Now,
φ to be the ideal of R generated by the union of Cf for
P we define JP
f ∈ ni=1 Rφ(xi ) + ni=1 Rxi satisfying the degree condition. Since no good
3
monomial is linear, Jφ is contained in Iφ . If φ satisfies the degree condition,
then Jφ is equal to Iφ . Hence, (ii) of the following theorem is a consequence
of (i).
Theorem 2.3. Let k be a field, R a commutative k-algebra, and φ ∈ GAn (R).
(i) If Jφ = R, then φ is stably co-tame.
(ii) If φ satisfies the degree condition and if Iφ = R, then φ is stably co-tame.
From Theorems 2.2 and 2.3 (ii), we obtain the following corollary.
Corollary 2.4. Let k be an infinite field, and R a commutative k-algebra.
Then, φ ∈ GAn (R) is stably co-tame if and only if Iφ = R.
In particular, if R = k is an infinite field, then φ ∈ GAn (k) is stably
co-tame if and only if a good monomial appears in φ(xi ) for some i. When
p = 0, this is the same as φ is non-affine.
In the following three sections, we prove Theorems 2.1, 2.2 and 2.3 (i).
3
Proof of Theorem 2.1
Let R be any commutative ring, and Γ a subgroup of GAn+1 (R) containing
Aff n+1 (R), where n ≥ 2. First, we study properties of Γ. Define
ǫ̂(f ) := (x1 , . . . , xn , xn+1 + f ) ∈ GAn+1 (R) for each f ∈ R[x].
We identify each permutation σ ∈ Sn+1 with (xσ(1) , . . . , xσ(n+1) ) ∈ Aff n+1 (R),
and write φσ := σ −1 ◦ φ ◦ σ for φ ∈ GAn+1 (R). Then, we have the following:
(A) If φ = (f1 , . . . , fn ) ∈ GAn (R) belongs to Γ, then ǫ̂(fi ) = φ ◦ ǫ̂(xi ) ◦ φ−1
belongs to Γ for each 1 ≤ i ≤ n.
(B) If f, g ∈ R[x̂] satisfy ǫ(f ), ǫ(gxn+1 ) ∈ Γ, then we have ǫ̂(f ) = ǫ(f )(1,n+1) ∈
Γ, and so
ǫ(f g) = ǫ̂(f ) ◦ ǫ(gxn+1 ) ◦ ǫ̂(f )−1 ◦ ǫ(gxn+1 )−1 ∈ Γ.
In particular, ǫ(f ) ∈ Γ implies ǫ(af ) ∈ Γ for each f ∈ R[x̂] and a ∈ R, since
ǫ(axn+1 ) belongs to Aff n+1 (R), and hence to Γ.
When Q ⊂ R and n ≥ 3, Derksen showed that Tn (R) is generated by
Aff n (R) and {ǫ(ax22 ) | a ∈ R}, and Tn (R) = hAff n (R), ǫ(x22 )i holds if R is
generated by R∗ as a Q-vector space (cf. [Es, Thm. 5.2.1]). We remark that,
since n ≥ 3, the first statement and (B) imply that Tn (R) = hAff n (R), ǫ(x22 )i
whenever Q ⊂ R.
4
Lemma 3.1. If Γ contains ǫ̂(x1 x2 ), or if n = 2 and Γ contains ǫ̂(x1 x22 ), then
Γ contains Tn (R).
Proof. We show that Γ contains ǫ(f ) for each f ∈ R[x̂]. Since ǫ(g + h) =
ǫ(g)◦ǫ(h) holds for each g, h ∈ R[x̂], we may assume that f is a monomial. In
the former case, Γ contains ǫ̂(x1 x2 )(1,n+1)(2,i) = ǫ(xi xn+1 ) for each 2 ≤ i ≤ n.
By (B), it follows that ǫ(f ) ∈ Γ implies ǫ(xi f ) ∈ Γ. Since Γ contains ǫ(a) for
each a ∈ R, the assertion follows by induction on deg f . In the latter case,
Γ contains ǫ̂(x1 x22 )(1,3) = ǫ(x22 x3 ). Hence, ǫ(axl2 ) ∈ Γ implies ǫ(axl+2
2 ) ∈ Γ for
each a ∈ R and l ≥ 0 similarly. Since Γ contains ǫ(a) and ǫ(ax2 ), it follows
that Γ contains ǫ(axl2 ) for all l ≥ 0.
The following two implications hold for the conditions listed in Theorem 2.1.
Lemma 3.2. (b) implies (a), and (d) implies (c).
Proof. By (2.1), (b) implies (x1 + x2 )(x1 − x2 ) = x21 − x22 ∈ Mφ . Since 2 is a
unit of R, we have (x1 +x2 , x1 −x2 , x3 , . . . , xn ) ∈ Aff n (R). Hence, Mφ contains
x1 x2 by (2.1). Similarly, (d) implies x21 x2 + x1 x22 = x31 + x32 + (x1 + x2 )3 ∈ Mφ .
Hence, Mφ contains ((ξx1 )2 x2 + ξx1 x22 ) + ξ 2 (x21 x2 + x1 x22 ) = ξ(ξ + 1)x1 x22 .
Since ξ(ξ + 1) is a unit of R, this implies that Mφ contains x1 x22 .
Now, we prove Theorem 2.1. Thanks to Lemmas 3.1 and 3.2 and (2.1),
it suffices to show that Γ := hAff n+1 (R), φi contains ǫ̂(f ) for each f ∈ Mφ .
Write φ = (f1 , . . . , fn ) and
f = h+
n X
r
X
i=1 j=1
ai,j ηj (fi ), where h ∈
n
X
Rxi + R, ai,j ∈ R and ηj ∈ Aff n (R).
i=1
Then, ǫ̂(f ) is the product of ǫ̂(h) and ǫ̂(ηj (ai,j fi )) for i = 1, . . . , n and j =
1, . . . , r. Since Γ contains ǫ̂(fi ) for each i by (A), Γ contains ǫ̂(ai,j fi ) for each
i, j by (B). Hence, Γ contains ηj ◦ ǫ̂(ai,j fi ) ◦ ηj−1 = ǫ̂(ηj (ai,j fi )) for each i, j.
Since ǫ̂(h) is affine, it follows that Γ contains ǫ̂(f ). This completes the proof
of Theorem 2.1.
4
Proof of Theorem 2.2
Assume that p is zero or a prime. We define NGn (R) to be the set of φ ∈
GAn (R) such that no good monomial appears in φ(x1 ), . . . , φ(xn ). If p = 0,
then we have NGn (R) = Aff n (R). In this case, the following theorem is
obvious.
5
Theorem 4.1. NGn (R) is a subgroup of GAn (R), and no element of NGn (R)
is stably co-tame. In fact, En (R) 6⊂ hAff n+r (R), NGn (R)i holds for any r ≥ 1.
To prove Theorem 4.1, we need a lemma, where
L p need not be zero or a
prime. Consider the standard grading R[x]
L = l>0 R[x]l . An R-submodule
V of R[x] is said to be graded if V =
l>0 (V ∩ R[x]l ). If V is generated
by monomials, then V is graded. Recall that each (f1 , . . . , fn ) ∈ R[x]n is
identified with the substitution map R[x] → R[x] defined by xi 7→ fi for
i = 1, . . . , n, and R[x]n forms a monoid for the composition defined in (1.1).
Note that V n is closed under this operation if and only if φ(V ) ⊂ V holds
for each φ ∈ V n .
Lemma 4.2. Let V be a graded R-submodule of R[x] such that V n is closed
under composition. If x1 , . . . , xn ∈ V , then V n ∩ GAn (R) is a subgroup of
GAn (R).
Proof. Since V n ∩ GAn (R) contains (x1 , . . . , xn ), and is closed under composition, we show that φ−1 belongs to V n for each φ ∈ V n ∩ GAn (R).
L There
exists η ∈ V n ∩ Aff n (R) for which ψ := φ ◦ η satisfies ψ(xi ) ∈ xi + l>2 R[x]l
for i = 1, . . . , n. Since φ−1 = η ◦ ψ −1 , and V n is closed under composition,
it suffices to verify that ψ −1 belongs to V n . Suppose to theP
contrary that
f := ψ −1 (xi ) does not belong to V for some i. Write f = l>0 fl , where
fl ∈ R[x]
holds for some d ≥ 0. Take the minimal d, and set
P l . Then, fd 6∈ V P
f ′ := l<d fl and f ′′ := l>d fl . Let h, h′ and h′′ be the homogeneous components of ψ(fd ), ψ(f ′ ) and ψ(f ′′ ) of degree d, respectively. Then, h + h′ + h′′
belongs to V , since ψ(fd ) + ψ(f ′ ) + ψ(f ′′ ) = ψ(f ) = xi belongs to V , and
V is graded by assumption. By the minimality of d, we have f ′ ∈ V . Since
ψ = φ ◦ η belongs to V n , it follows that ψ(f ′ ) belongs to V . This implies that
h′ belongs to L
V as before. As for h and h′′ , we have h = fd and h′′ = 0, since
ψ(xi ) ∈ xi + l>2 R[x]l holds for i = 1, . . . , n. Thus, h + h′ + h′′ = fd + h′
does not belong to V , a contradiction. Therefore, ψ −1 belongs to V n .
P
We remark that, if V = s∈Σ As for some Σ ⊂ R[x] and an R-subalgebra
A of R[x], and if φ(A) ⊂ A and φ(Σ) ⊂ V hold for each φ ∈ V n , then
V n is closed under composition. For example, assume that R has prime
l
l
l
characteristic p, and define R[xp ] := R[xp1 , . . . , xpn ] for each l ≥ 0. Set
Zd := {a ∈ Z | a ≥ d} for each d ∈ Z. Let d, e ∈ Z0 and ∅ =
6 N ⊂ Z0 be
such that d ≤ e, and each u, v ∈ N satisfy u + v ∈ N ∪ Zd . Then, we define
a graded R-submodule of R[x] by
d
V := V (d, e, N) := R[xp ] +
n X
X
i=1 u∈N
6
e
u
R[xp ]xpi .
Let us prove that V n is closed under composition using the remark for A :=
u
e
d
R[xp ] and Σ := R[xp ] ∪ {xpi | 1 ≤ i ≤ n, u ∈ N}. First, note that f ∈ V
w
u
v
v
implies f p ∈ V for any v ∈ N, since each u, v ∈ N satisfy (xpi )p ∈ {xpi |
u
d
d
e
w ∈ N} ∪ R[xpi ], and R[xp ] contains R[xp ]. Hence, φ(xpi ) belongs to V for
l
l
each φ ∈ V n , 1 ≤ i ≤ n and u ∈ N. Clearly, φ(R[xp ]) ⊂ R[xp ] holds for any
φ ∈ R[x]n and l ≥ 0. Therefore, we have φ(A) ⊂ A and φ(Σ) ⊂ V for each
φ ∈ V n.
Now, let us prove Theorem 4.1 when p is a prime. Clearly, u, v ∈ {0}
implies u + v ∈ {0} ∪ Z1 . So, we define Vn := V (1, 1, {0}) and Wn :=
V (1, 2, {0}), i.e.,
Vn = R[xp ] + R[xp ]x1 + · · · + R[xp ]xn
2
2
Wn = R[xp ] + R[xp ]x1 + · · · + R[xp ]xn .
If n = p = 2, then t1 , t2 ≥ 0 satisfy xt11 xt22 ∈ W2 if and only if t1 , t2 ≡ 0
(mod 2), or ti ≡ 1, tj ≡ 0 (mod 4) for some i, j ∈ {1, 2}. Hence, we have
xt11 xt22 6∈ W2 if and only if xt11 xt22 is a good monomial. Similarly, when (n, p) 6=
(2, 2), a nonzero monomial m is good if and only if m does not belong to Vn .
Thus, NGn (R) is equal to W22 ∩ GA2 (R) if n = p = 2, and to Vnn ∩ GAn (R)
otherwise. Therefore, NGn (R) is a subgroup of GAn (R) by Lemma 4.2. Note
n+r
n+r
that hAff n+r (R), NGn (R)i is contained in Wn+r
if n = p = 2, and in Vn+r
n+r
otherwise. Since En (R) \ Vn+r
contains ǫ(x2 x3 ) if n ≥ 3, and ǫ(x22 ) if p ≥ 3,
n+r
while En (R) \ Wn+r contains ǫ(xp+1
2 ) for any n, p ≥ 2, we get the last part of
Theorem 4.1. This completes the proof.
If I is a proper ideal of R, then each φ ∈ GAn+r (R) induces an element
φI of GAn+r (R/I). Since En (R) ∋ τ 7→ τI ∈ En (R/I) is surjective, (σ ◦ τ )I =
σI ◦τI for each σ, τ ∈ GAn+r (R), and τ ∈ Aff n+r (R) implies τI ∈ Aff n+r (R/I),
we see that
En (R) ⊂ hAff n+r (R), φi implies En (R/I) ⊂ hAff n+r (R/I), φI i
(4.1)
for any r ≥ 0 and φ ∈ GAn (R). Hence, Theorem 4.1 implies the following
corollary.
Corollary 4.3. Let I be a proper ideal of R such that char(R/I) is zero
or a prime. If φ ∈ GAn (R) satisfies φI ∈ NGn (R/I), then En (R) 6⊂
hAff n+r (R), φi holds for any r ≥ 1. Here, char(R/I) is the characteristic
of R/I.
Now, we can prove Theorem 2.2. By assumption, I := Iφ is a proper ideal
of R. If p is a prime, then R contains Fp , and so char(R/I) = p. Hence, φI
belongs to NGn (R/I) by the definition of I. Thus, φ is not stably co-tame
7
by Corollary 4.3. If p = 0, then φI is affine. By (4.1), this implies that φ is
not stably co-tame.
Finally, we remark that, if p = 2, then R[x2 ] is contained in Vn . Hence,
we have E2 (R) ⊂ V22 , and so T2 (R) ⊂ V22 . Thus, for any φ ∈ T2 (R), no
monomial xt11 xt22 with t1 and t2 odd appears in φ(x1 ) and φ(x2 ). If R is a
domain, the same holds for any φ ∈ GA2 (R), since T2 (K) = GA2 (K) for any
field K (cf. [J], [K]).
Lemma 4.4. Assume that p = 2, φ ∈ GA2 (R), 1 ≤ i ≤ 2 and t1 , t2 ≥ 0
are odd. If a ∈ R is the coefficient of xt11 xt22 in φ(xi ), then a belongs to the
nilradical of R.
Proof. Let p be any prime ideal of R. Since R/p is a domain of characteristic
two, xt11 xt22 does not appear in φp (xi ) as mentioned. Hence, a belongs to p.
5
Proof of Theorem 2.3 (i)
Assume that R is a commutative k-algebra. To prove Theorem 2.3 (i), we
verify that one of (a), (b) and (c) of Theorem 2.1 holds when Jφ = R. If
Jφ = R, a good monomial appears in a polynomial satisfying the degree
condition. This implies that #k > 2, and so there always exists ξ ∈ R∗
satisfying ξ + 1 ∈ R∗ .
P
e
Take any f (x) ∈ R[x] and write f (x) = di=0 ui xip with d, e ≥ 0 and
ui ∈ R, where we regard pe = 1 if p = 0. If k contains d + 1 distinct
e
elements ξ0 , . . . , ξd , then ui xip can be written as a k-linear combination of
f (ξ0 x), . . . , f (ξd x) for each 0 ≤ i ≤ d by linear algebra. This remark is used
to prove the following lemma.
Lemma 5.1. If f ∈ R[x] satisfies the degree condition, each monomial appearing in f is written as a k-linear combination of f (ξ1 x1 , . . . , ξn xn ) for
ξ1 , . . . , ξn ∈ k ∗ .
Proof. We prove the lemma by P
induction on n. The case n = 0 is clear. Ase
sume that n ≥ 1, and write f = di=0 fi xip
1 , where d := #k−2, fi ∈ R[x̂] and
e ≥ 0. Take any monomial m appearing in f . Then, there exists i such that
m appears in fi xi1 . Since #k = d + 2, we may find distinct α0 , . . . , αd ∈ k ∗ .
Then, fi xi1 is written as a k-linear combination of f (αl x1 , x2 , . . . , xn ) for
l = 0, . . . , d as remarked. By the choice of i, we have m = m′ xi1 for some
monomial m′ appearing in fi . Note that degsxj fi ≤ degsxj f ≤ d for all
j. Hence, by induction assumption, m′ is written as a k-linear combination of fi (β2 x2 , . . . , βn xn ) for β2 , . . . , βn ∈ k ∗ . Thus, m = m′ xi1 is written
as a k-linear combination of fi (β2 x2 , . . . , βn xn )xi1 ’s, and therefore that of
f (αl x1 , β2 x2 , . . . , βn xn ) for l = 0, . . . , d and β2 , . . . , βn ∈ k ∗ .
8
Next, we define λ ∈ Aff n (R) by λ(xi ) = xi + 1 for i = 1, . . . , n.
Lemma 5.2. If xt11 · · · xtnn is a good monomial, there appears in λ(xt11 · · · xtnn )
the monomial uxi xj with u ∈ k ∗ in the cases (I), (II) and (III), xi x2j in the
case (IV), and x3i in the case (V), where i, j ∈ {1, . . . , n}, and where i 6= j if
(II) or (IV).
′
Proof. In the case (IV), we have λ(xt11 xt22 ) = (xi +1)(x2j +1)(x4i +1)ti (x4j +1)tj
for some i, j ∈ {1, 2} with i 6= j, and t′1 , t′2 ≥ 0. Then, it is easy to see that the
monomial xi x2j appears in λ(xt11 xt22 ). Other cases are verified similarly.
′
Now, let us prove Theorem 2.3 (i).
by assumption, we may
PSince Jφ = RP
find a1 , . . . , ar ∈ R and g1 , . . . , gr ∈ ni=1 Rφ(xi ) + ni=1 Rxi with r ≥ 1 as
follows:
(1) For
Pr each 1 ≤ l ≤ r, a good monomial ml appears in gl with coefficient al .
(2) l=1 al = 1, and g1 , . . . , gr satisfy the degree condition.
Since gl satisfies the degree condition for each l, so does ml , and hence
so does λ(ml ). Clearly, gl belongs to Mφ . Hence, ml belongs to Mφ by
Lemma 5.1 and (2.1). Then, λ(ml ) belongs to Mφ , so each monomial appearing in λ(ml ) belongs to Mφ similarly. By Lemma 5.2, there appears in
λ(ml ) the monomial nl := ul al xil xjl with ul ∈ k ∗ in the cases (I), (II) and
(III), nl := al xil x2jl in the case (IV), and nl := al x3il in the case (V), where
il , jl ∈ {1, . . . , n}, and where il 6= jl if (II) or (IV). Since nl belongs to Mφ ,
we have al x1 x2 ∈ Mφ in the case (II). The proof of Lemma 3.2 shows that
al x1 x22 ∈ Mφ in the case (V) as well as in the case (IV).
If n ≥P3 and p = 2, then ml must be of type (II) for all l. Hence, Mφ
contains rl=1 al x1 x2 = x1 x2 . Therefore, (a) of Theorem 2.1 holds.
If n = p = 2, then ml is of type (II) or (IV)
Pror (V) for each l. In the
case (II),
l=1 al = 1, it follows that
P al is nilpotent by Lemma 4.4. Since
P
u := ′ al is a unit of R, where the sum ′ isPtaken over l such that ml is
of type (IV) or (V). Hence, Mφ contains u−1 ′ al x1 x22 = x1 x22 . Therefore,
(c) of Theorem 2.1 holds.
If p 6= 2, then each ml is of type (I) or (II) or (III), so Mφ contains
ul al xil xjl . By (2.1), Mφ contains al x21 if il = jl . If il 6= jl , then Mφ contains
2
a
l x1 x2 , and hence contains al x1 (x2 + x1 ) − al x1 x2 = al x1 . Thus, Mφ contains
P
r
2
2
l=1 al x1 = x1 . Therefore, (b) of Theorem 2.1 holds. This completes the
proof of Theorem 2.3 (i).
9
6
Remarks
6.1. When R = k, Edo-Lewis [EL] showed that θN := (β ◦ π)−N ◦ π ◦ (β ◦ π)N
is neither affine nor co-tame for each N ≥ 3, where
β := (x1 + x22 (x2 + x23 )2 , x2 + x23 , x3 ) and π := (x2 , x1 , x3 ).
Here, we mention that the expression of θN above is slightly different from the
original one, due to the difference in the definitions of composition (cf. §2.2
of [EL] and (1.1)). Since R = k and θN is not affine, θN is stably co-tame
if p = 0 by the remark after Corollary 2.4. If p = 2, then θN belongs to
NG3 (k), and hence is not stably co-tame by Theorem 4.1. When p ≥ 3, we
have the following theorem.
Theorem 6.1. If p ≥ 3, N ≥ 1 and #k ≥ 42N −1 + 2, then θN is stably
co-tame.
Proof. Observe that θN is written as σ1 ◦ τ1 ◦ · · · ◦ σN ◦ τN ◦ σ, where
σi ∈ {π ◦ β ◦ π, π ◦ β −1 ◦ π},
τi ∈ {β, β −1 } and σ ∈ {id, π}
with σ1 = π ◦ β −1 ◦ π =: (h1 , h2 , x3 ). For each w = (w1 , w2 , w3 ) ∈ Z3 , we
define the w-weighted grading on R[x] by degw xi := wi for i = 1, 2, 3. For
f ∈ R[x] \ {0}, we denote by f w the highest w-homogeneous part of f . Now,
′
set θN
:= θN ◦ σ. We claim that, if degw h2 is greater than (1/4) degw h1 and
2w3 , then
2N−i
′
4
θN
(xi )w = (σ1 ◦ τ1 ◦ · · · ◦ σN ◦ τN )(xi )w = ±(hw
2 )
for i = 1, 2.
In fact, the case N = 1 is checked directly, and the case N ≥ 2 follows by
induction on N, since (σi ◦ τi )(x2 ) = x2 ± x21 (x1 ± x23 )2 ± x23 and
(σi ◦τi )(x1 ) = x1 ±x23 ±(x2 ±x21 (x1 ±x23 )2 )2 (x2 ±x21 (x1 ±x23 )2 ±x23 )2
for each i.
Since h1 = x1 − x23 and h2 = x2 − x21 h21 , the assumption of the claim holds
for w = (1, 0, 0), (0, 1, 0), (0, 0, 1), (2, 0, 1), for which
4
2 4
2
2 2
hw
2 = −x1 , x2 , −x1 x3 , −x1 (x1 − x3 ) ,
′
respectively. The first three of these imply degxi θN
(x2 ) ≤ 4 · 42N −2 ≤
′
#k − 2 for i = 1, 2, 3. Hence, θN
(x2 ) satisfies the degree condition. Since
′
(2,0,1)
2 t
θN (x2 )
= ±(x1 (x1 − x3 )) with t := 2 · 42N −2 similarly, the monomial
′
±(x1 x23 )t appears in θN
(x2 ). Since p ≥ 3, we see that xt1 x2t
3 is a good mono′
mial of type (III). Thus, we get JθN = JθN = R. Therefore, θN is stably
co-tame by Theorem 2.3 (i).
10
6.2. When k is a finite field, Theorem 2.3 might not be useful due to the
degree condition. In such a case, the following technique may be effective.
Take any η ∈ Aff n (R), and set δ := η − id. Then, δ(Mφ ) ⊂ Mφ holds for
each φ ∈ GAn (R) by (2.1). Since φ(xj ) belongs to Mφ for each j, we get the
following lemma.
Lemma 6.2. If φ ∈ GAn (R) and h ∈ R[x] satisfy the following condition,
then h belongs to Mφ : There exist 1 ≤ j ≤ n and η1 , . . . , ηl ∈ Aff n (R) with
l ≥ 1 such that (δ1 ◦ · · · ◦ δl ◦ φ)(xj ) belongs to R∗ h + Mφ , where δi := ηi − id
for each i.
Assume that p is a prime. For i = 1, . . . , n, take any ai ∈ R∗ , define
ηi ∈ Aff n (R) by ηi (xi ) = xi + ai and ηi (xj ) = xj for j 6= i, and put δi :=
ηi − id. Then, we have δip = 0 since ηip = id, degxi δi (xdi ) < d for d ≥ 0, and
P
Rxji if p ∤ d − l for all 0 ≤ l < t. Note that ker δi
δit (xdi ) ∈ R∗ xd−t
+ d−t−1
i
j=0
is an R-subalgebra of R[x], and δi is ker δi -linear. Also, ker δi contains the
R-subalgebra Ai generated by qi := xpi − ap−1
xi and xj for j 6= i. We claim
i
that ker δi = Ai . In fact, each f ∈ ker δi \ {0} satisfies p | degxi f =: lp,
hence degxi (f − hqil ) < lp holds for some h ∈ R[{xj | j 6= i}]. Thus, we get
f − hqil ∈ Ai by induction on l, and so f ∈ Ai .
Now, fix l = (l1 , . . . , ln ) ∈ {0, . . . , p − 1}n , and set δ l := δ1l1 ◦ · · · ◦ δnln and
Ml :=
n
X
i=0
l
Rxi x +
li −1
n X
X
Ai xji , where x0 := 1 and xl := xl11 · · · xlnn .
i=1 j=0
Then, since δi ’s commute
P with each other, and are Ai -linear, we see that
δ l (Ml ) is contained in ni=0 Rxi . Hence, δ l (Ml ) ⊂ Mφ holds for every φ ∈
GAn (R). Using this, we can give sufficient conditions for stably co-tameness
of φ. For P
example, assume that l1 , l2 ≤ p − 2. Then, δ l (x1 x2 xl ) belongs to
R∗ x1 x2 + 2i=0 Rxi . Hence, if φ satisfies φ(xj ) ∈ R∗ x1 x2 xl + Ml for some
1 ≤ j ≤ n, then (δ l ◦ φ)(xj ) belongs to R∗ x1 x2 + Mφ . This implies that x1 x2
belongs to Mφ by Lemma 6.2. Therefore, φ is stably co-tame by Theorem 2.1.
Similarly, if p ≥ 3, l1 ≤ p − 3, and φ(xj ) belongs to R∗ x21 xl + Ml for some
1 ≤ j ≤ n, then φ is stably co-tame.
6.3. When n ≥ 3, Motoki Kuroda [MK] showed that ǫ(x22 ) is stably cotame for R = Fp with p ≥ 3, and found it hard to decide whether ǫ(x52 ) is
stably co-tame for R = F3 . Note that ǫ(x52 ) is stably co-tame for R = F9 by
Theorem 2.3 (ii).
6.4. The Tame Generators Problem asks if GAn (R) = Tn (R). In the context of stably co-tame automorphisms, NGn (R) is a natural generalization of
Aff n (R). So, the following problem is of interest, as a generalization of the
11
Tame Generators Problem in prime characteristic (see also [EK] for other
generalizations).
Problem. When p is a prime, does it hold that GAn (R) = hNGn (R), Tn (R)i?
References
[BEW] J. Berson, A. van den Essen and D. Wright, Stable tameness of
two-dimensional polynomial automorphisms over a regular ring, Adv.
Math. 230 (2012), no. 4-6, 2176–2197.
[B]
Y. Bodnarchuk, On generators of the tame invertible polynomial maps
group, Internat. J. Algebra Comput. 15 (2005), no. 5-6, 851–867.
[Ed]
E. Edo, Coordinates of R[x, y]: constructions and classifications,
Comm. Algebra 41 (2013), no. 12, 4694–4710.
[EK]
E. Edo and S. Kuroda, Generalisations of the tame automorphisms
over a domain of positive characteristic, Transform. Groups 20
(2015), no. 1, 65–81.
[EL]
E. Edo and D. Lewis, The affine automorphism group of A3 is not a
maximal subgroup of the tame automorphism group, Michigan Math.
J. 64 (2015), no. 3, 555–568.
[Es]
A. van den Essen, Polynomial automorphisms and the Jacobian
conjecture, Progress in Mathematics, Vol. 190, Birkhäuser, Basel,
Boston, Berlin, 2000.
[J]
H. Jung, Über ganze birationale Transformationen der Ebene, J.
Reine Angew. Math. 184 (1942), 161–174.
[K]
W. van der Kulk, On polynomial rings in two variables, Nieuw Arch.
Wisk. (3) 1 (1953), 33–41.
[MK]
M. Kuroda, Stably Derksen polynomial automorphisms over finite
fields (Japanese), Master’s Thesis, Tokyo Metropolitan University,
January 2016.
[MW] S. Maubach and R. Willems, Polynomial automorphisms over finite
fields: mimicking tame maps by the Derksen group, Serdica Math. J.
37 (2011), no. 4, 305–322 (2012).
12
[N]
M. Nagata, On Automorphism Group of k[x, y], Lectures in Mathematics, Department of Mathematics, Kyoto University, Vol. 5, Kinokuniya Book-Store Co. Ltd., Tokyo, 1972.
[SU]
I. Shestakov and U. Umirbaev, The tame and the wild automorphisms
of polynomial rings in three variables, J. Amer. Math. Soc. 17 (2004),
197–227.
[S]
M. K. Smith, Stably tame automorphisms, J. Pure Appl. Algebra 58
(1989), 209–212.
Department of Mathematics and Information Sciences
Tokyo Metropolitan University
1-1 Minami-Osawa, Hachioji
Tokyo 192-0397, Japan
[email protected]
13
| 0math.AC
|
Continuous monitoring of ℓp norms in data streams
Jaroslaw Blasiok∗
Jian Ding†
Jelani Nelson‡
arXiv:1704.06710v3 [cs.DS] 9 Nov 2017
November 10, 2017
Abstract
In insertion-only streaming, one sees a sequence of indices a1 , a2 , . . . , am ∈ [n]. The stream defines
def
a sequence of m frequency vectors x(1) , . . . , x(m) ∈ Rn with (x(t) )i = |{j : j ∈ [t], aj = i}|. That is,
(t)
x is the frequency vector after seeing the first t items in the stream. Much work in the streaming
literature focuses on estimating some function f (x(m) ). Many applications though require obtaining
estimates at time t of f (x(t) ), for every t ∈ [m]. Naively this guarantee is obtained by devising an
algorithm with failure probability ≪ 1/m, then performing a union bound over all stream updates to
guarantee that all m estimates are simultaneously accurate with good probability. When f (x) is some
ℓp norm of x, recent works have shown that this union bound is wasteful and better space complexity
is possible for the continuous monitoring problem, with the strongest known results being for p = 2
[HTY14, BCIW16, BCI+ 17]. In this work, we improve the state of the art for all 0 < p < 2, which we
obtain via a novel analysis of Indyk’s p-stable sketch [Ind06].
1
Introduction
Estimating statistics of frequency vectors implicitly defined by insertion-only update streams, as defined
in the abstract, was first studied by Flajolet and Martin in [FM85]. They studied the so-called distinct
elements problem, in which f (x) is the support size of x. In the insertion-only model, the support size of x is
equivalent to the number of distinct ai appearing in the stream. One goal in such streaming algorithms, both
for this particular distinct elements problem as well as for many others function estimation problems studied
in subsequent works, is to minimize the space consumption of the stream-processing algorithm, ideally using
o(n) words of memory (note there is always a trivial n space algorithm by storing x explicitly in memory).
For over two decades, work on estimating statistics of frequency
P vectors of streams remained dormant,
until the work of [AMS99] on estimating the p-norm kxkp = ( i xpi )1/p in streams for integer p ≥ 1.
Since then several works have studied these and several other problems, from the perspective of both upper
and lower bounds, including estimating kxkp for all 0 < p ≤ 2 (not necessarily integral) [AMS99, Ind06,
IW03, Woo04, Li08, Li09, KNW10a, NW10, KNPW11, JW13], kxkp for p > 2 [AMS99, BYJKS04, CKS03,
IW05, BGKS06, Gro09, Jay09, AKO11, BO13, BKSV14, Gan15], empirical entropy [CBM06, CCM10, BG06,
HNO08] and other information-theoretic quantities [IM08, GIM08, BO10a], cascaded norms [CM05, JW09,
Jay13], and several others. There have also been general theorems classifying which statistics of frequency
vectors admit space-efficient streaming estimation algorithms [BO10b, BC15, BOR15, BCWY16, BCKY17].
Taking a dynamic data structural viewpoint, “streaming algorithms” is simply a synonym for “dynamic
data structures” but with an implied focus on minimizing memory consumption (typically striving for an
algorithm using sublinear memory). Elements in the stream can be viewed as updates to the frequency
∗ Harvard
University. [email protected]. Supported by ONR grant N00014-15-1-2388.
of Chicago. [email protected]. Partially supported by NSF grant DMS-1455049 and an Alfred P.
Sloan Research Fellowship.
‡ Harvard University. [email protected]. Supported by NSF grant IIS-1447471 and CAREER award CCF-1350670,
ONR Young Investigator award N00014-15-1-2388, and a Google Faculty Research Award.
† University
1
vector x (seeing a ∈ [n] in the stream can be seen as update(a, 1), causing the change xa → xa + 1), and
the request for an estimate of some statistic of x is a query. In this data structural language, all the works
cited in the previous paragraph provide Monte-Carlo guarantees of the following form for queries: starting
from any fixed frequency vector and after executing any fixed sequence of updates, the probability that the
output of a subsequent query then fails is at most δ. Here we say a query fails if, say, the output is not a
good approximation to some particular f (x) (this will be made more formal later). In many applications
however, one does not simply want the answer to one query at the end of some large number of updates,
but rather one wants to continuously monitor the data stream. That is, the sequence of data structural
operations is an intermingling of updates and queries. For example, one may have a threshold T in mind,
and if f (x) ever increases beyond T some data analyst should be alerted. Such a goal could be achieved
(approximately) by querying after every update to determine whether the updated frequency vector satisfies
this property. Indeed, the importance of supporting continuous queries in append-only databases (analogous
to the insertion-only model of streaming) was recognized 25 years ago in [TGNO92], with several later works
focused on continuous stream monitoring with application areas in mind such as trend detection, anomaly
detection, financial data analysis, and (bio)sensor data analysis [BW01, CÇC+ 02, OJW03].
If one assumes that a query is being issued after every update, then in a stream of m updates the failure
probability should be set to δ ≪ 1/m so that, by a union bound, all queries succeed. Most Monte-Carlo
streaming algorithms achieve some space S to achieve failure probability 1/3, at which point one can achieve
failure probability δ by running Θ(lg(1/δ)) instantiations of the algorithm in parallel and returning the
median estimate (see for example [AMS99]). This method increases the space from S to Θ(S lg(1/δ)), and
for many problems (such as ℓp -norm estimation) it is known that at least in the so-called strict turnstile
model (i.e. update(a, ∆) is allowed for both positive and negative ∆, but we are promised xi ≥ 0 for all i
at all times) this form of space blow-up is necessary [JW13]. Nevertheless, although improved space lower
bounds have been given when desiring that the answer to a single query fails with probability at most δ,
now such blow-up has been shown necessary for the continuous monitoring problem in which one wants,
with failure probability 1/3, to provide simultaneously correct answers for m queries intermingled with m
updates. In fact to the contrary, in certain scenarios such as estimating distinct elements or the ℓ2 -norm in
insertion-only streams, improved upper bounds have been given!
Definition 1. We say a Monte-Carlo randomized streaming algorithm A provides strong tracking for f
in a stream of length m with failure probability η if at each time t ∈ [m], A outputs an estimate f˜t such that
P(∃t ∈ [m] : |f˜t − f (x(t) )| > εf (x(t) )) < η.
We say that A provides weak tracking for f if
′
P(∃t ∈ [m] : |f˜t − f (x(t) )| > ε sup f (x(t ) )) < η.
t′ ∈[m]
′
Note if f is monotonically increasing, then for insertion-only streams supt′ ∈[m] f (x(t ) ) is simply f (x(m) ).
The first non-trivial tracking result we are aware of which outperformed the median trick for insertion-only
streaming was the RoughEstimator algorithm given in [KNW10b] for estimating the number of distinct
elements in a stream. RoughEstimator provided a strong tracking guarantee for f (x) = | support(x)| (the
distinct elements problem) for constant ε, η, using the same space as what is what is required to answer only
a single query. This strong tracking algorithm was used as a subroutine in the main non-tracking algorithm
of that work for approximating the number of distinct elements in a data stream up to 1 + ε.
For ℓp -estimation for p ∈ (0, 2], without tracking, it is known that O(ε−2 lg(1/δ)) words of memory is
achievable to return a (1 + ε)-approximate value of f (x) = kxkp with failure probability δ [AMS99, Ind06,
KNW10a]1 . This upper bound thus implies a strong tracking algorithm with space complexity O(ε−2 lg m)
for tracking failure probability η = 1/3, by setting δ < 1/(3m) and performing a union bound. The work
[HTY14] considered the strong tracking variant of ℓp -estimation in insertion-only streams for for any p in the
more restricted interval (1, 2]. They showed that the same algorithms of [AMS99, Ind06], unchanged, provide
1 For
constant δ and p = 2, [AMS99] shows that space O(ε−2 (lg n + lg lg m)) bits is achievable in insertion-only streams.
2
strong tracking with η = 1/3 with space O(ε−2 (lg n + lg lg m + lg(1/ε))) words2 . This is an improvement
over the standard median trick and union bound when the stream length is very long (m > nω(1) ) and ε
is not too small (ε > 1/mo(1)). They also showed that in an update model which allows deletions of items
(“turnstile streaming”), any algorithm which only maintains a linear sketch Πx of x must use Ω(lg m) words
of memory for constant ε, showing that the median trick is optimal for this restricted class of algorithms.
A different algorithm was given in [BCIW16] for strong tracking for ℓ2 using space O(ε−2 (lg(1/ε) +
lg lg m)). It was then most recently shown in [BCI+ 17] that the AMS sketch itself of [AMS99] (though
with 8-wise independent hash functions instead of the original 4-wise independence proposed in [AMS99])
provides strong tracking in space O(ε−2 lg lg m), and weak tracking in space O(1/ε2 ). That is, the AMS
sketch provides weak tracking without any asymptotic increase in space complexity over the requirement to
correctly answer only a single query.
Despite the progress in upper bounds for tracking ℓ2 , the only non-trivial improvement for tracking ℓp is
the O(ε−2 (lg n + lg lg m + lg(1/ε))) upper bound of [HTY14]. Although this bound provides an improvement
for very long streams (m super-polynomial in n), it does not provide any improvement over the standard
median trick for the case most commonly studied case in the literature of m, n being polynomially related.
Our contribution. We show that Indyk’s p-stable sketch [Ind06] for 0 < p ≤ 2, derandomized using
bounded independence as in [KNW10a], provides weak tracking while using O(lg(1/ε)/ε2 ) words of space. It
also provides strong tracking using O(ε−2 (lg lg m + lg(1/ε)) words of space. Our bounds thus both improve
the space complexity achieved in [HTY14] for ℓp -tracking, and well as the range of p supported from p ∈ (1, 2]
to all p ∈ (0, 2] (note for p > 2, it is known that any algorithm requires polynomial space even to obtain a
2-approximation for a single query, i.e. the non-tracking variant of the problem [BYJKS04]).
2
Notation
We use [n] for integer n to denote {1, . . . , n}. We measure space in words unless stated otherwise, where
a single word is at least lg(nm) bits. For p ∈ (0, 2], we let Dp denote the symmetric p-stable distribution,
scaled so that for Z ∼ Dp , P(|Z| > 1) = 12 . The distribution Dp has theP
property that it is supported on the
n
reals, and for any fixed vector v ∈ Rn and Z1 , . . . , Zn , Z i.i.d. from Dp , i=1 Zi xi is equal in distribution to
kxkp · Z. See [Nol17] for further reading on these distributions.
For two vectors u, v ∈ Rn we write u v to denote coordinatewise comparison, i.e. u v iff ∀i ui ≤ vi .
For a finite set S, we write #S to denote cardinality of this set.
3
Preliminaries
The following lemma is standard. A proof with explicit constants can be found in [Nel11, Theorem 42].
Lemma 2. If Z ∼ Dp , then P(Z > λ) ≤
Cp
λp
for some explicit constant Cp depending only on p.
We also state some other results we will need.
Lemma 3 (Paley-Zygmund). If Z ≥ 0 is a random variable with finite variance, then
P(Z > θ E Z) ≥ (1 − θ)2
(E Z)2
.
E(Z 2 )
Corollary 4. For fixed vector v ∈ Rn , if σ ∈ {±1}n is a vector of 4-wise independent random signs, then
P(hσ, vi2 ≥
2
1
kvk22 ) ≥
3
27
2 For p = 2 their space is as written including the space required to store all hash functions, but for 1 < p < 2 this space
bound assumes that the storage of hash functions is for free.
3
Proof. This follows from Ehσ, vi4 < 3(Ehσ, vi2 )2 and the Paley-Zygmund inequality.
Theorem 5 ([BCIW16, BCI+ 17, Theorem 15]). Let v (1) , v (2) , . . . v (m) ∈ Rn , be a sequence of vectors such
that 0 v (1) v (2) . . . v (m) . Let σ ∈ {±1}n be a vector of 4-wise independent random signs. Then
C
P sup |hσ, v (i) i| > λkv (n) k2 < 2
λ
i≤m
for some universal constant C.
Theorem 6. [KNW10a, DKN10] If Zi ∼ Dp for i ∈ [n] are k-wise independent random variables, then for
every vector x ∈ Rn and every pair a, b ∈ R ∪ {±∞} we have
P(hZ, xi ∈ (a, b)) = P(kxkp Z1 ∈ (a, b)) ± O(k −1/p )
Theorem 7. [BR94, P
Lemma 2.3] Let X1 , . . . Xn ∈ {0, 1} be a sequence of k-wise independent random
variables, and let µ = E Xi . Then
X
∀λ > 0, P(
Xi ≥ (1 + λ)µ) ≤ exp(−Ω(min{λ, λ2 }µ)) + exp(−Ω(k))
4
Overview of approach
Indyk’s p-stable sketch picks a random matrix Π ∈ Rd×n such that each entry is drawn according to the
distribution Dp . It then maintains the sketch Πx(t) of the current frequency vector. This sketch can be
easily updated as the frequency vector changes, i.e. after observing an index aj ∈ [n] we update the sketch
by Πx(t+1) := Πx(t) + Πeaj . An kxkp -estimate query is answered by returning the median of |Πx(i) |j over
j ∈ [d]. Since storing Π in memory explicitly is prohibitively expensive, we generate it so that the entries in
each row are k-wise independent for k = O(1/εp ) (as done in [KNW10a]), and the d seeds used to generate
the rows of Π are O(lg(1/(εδ)))-wise independent. We also work with discretized p-stable random variables
to take bounded memory. All together, the bounded independence and discretization, also performed in
[KNW10a], allow us to store Π using low memory.
We then show that instantiating Indyk’s algorithm with d = O(ε−2 lg(1/(εδ))) provides the weak tracking
guarantee with failure probability δ. The analysis of the correctness of this algorithm is as follows. Let πi
denote the ith row of Π. We first show a result resembling the Doob’s martingale inequality — namely,
in Section 5 we show that for a fixed i, if we look at the evolution of hπi , x(t) i as t increases, the largest
attained value (supt≤m hπi , x(t) i) is with good probability not much larger than the median of the distribution
|hπi , x(m) i|, which is the typical magnitude of the counter at the end of the stream. This fact resembles similar
facts shown in [BCIW16, BCI+ 17] for when the πi have independent Rademachers as entries, though our
situation is complicated by the fact that p-stable random variables have much heavier tails.
We then, discussed in Section 5.1, show how the previous paragraph implies a weak tracking algorithm
with d = O(ε−2 lg(1/(εδ))): we split the sequence of updates into poly(1/ε) intervals such that the ℓp -norm
of the frequency vector of updates in each of those intervals, i.e. kx(t+1) − x(t) kp , is of the order εΘ(1) kx(m) kp .
We then union bound over the poly(1/ε) intervals to argue that the algorithm’s estimate is good at each of
the interval endpoints. This is the source of the extra factor of lg(1/ε) in our space bound: to obtain ε−Ω(1)
failure probability to union bound over these intervals. On the other hand, within each of the intervals most
of the counters do not change too rapidly by the argument developed in Section 5.
Finally, in Section 5.2 we show how given an algorithm satisfying a weak tracking guarantee, one can
use it to get a strong-tracking algorithm with slightly larger space complexity. This argument was already
present in [BCI+ 17]. One first identifies q points in the input stream at which the ℓp norm roughly doubles
when compared to the previously marked point. There are only O(lg m) such intervals. It is then enough to
ensure that our algorithm satisfies weak tracking for all those O(lg m) prefixes simultaneously, in order to
deduce that the algorithm in fact satisfies strong tracking. This is done by union bound over O(lg m) bad
events (as opposed to standard union bound over O(m) bad events), which introduces an extra lg lg m factor
in the space complexity as when compared to weak tracking.
4
5
Analysis
We first show two lemmas that play a crucial role in our weak tracking analysis.
Lemma 8. Let x ∈ Rn be a fixed vector, and Z ∈ Rn be a random vector with k-wise independent entries
drawn according to Dp . Then
X
C
P(
x2i Zi2 ≥ λ2 kxk2p ) ≤ p + O(k −1/p )
λ
for some universal constant C.
P 2 2
Proof. Let E0 be the event
xi Zi ≥ λ2 kxk2p . Note that E0 depends only on |Zi |, and does not depend on
the signs of the Zi . We write Zi = |Zi |σi , where σi are k-wise independent random signs. Conditioning on
|Zi |,
X
X
2
E
x2i Zi2
xi |Zi |σi |Z1 |, . . . |Zn | =
σ
and therefore for any |Z1 |, . . . , |Zm | for which E0 holds, by Corollary 4
X
X
2
2
2 2
1
2X 2 2
2
P
xi |Zi |σi ≥ λ kxkp |Z1 |, . . . , |Zm | ≥ P
xi Zi |Z1 |, . . . , |Zm | ≥
xi |Zi |σi ≥
σ
σ
3
3
27
and thus
P
σ
X
xi |Zi |σi
2
2
≥ λ2 kxk2p |Z1 |, . . . |Zn |
3
≥
1E0
,
27
where 1E0 is an indicator random variable for event E0 . Integrating over |Zi |,
X
2
2
1
P
xi |Zi |σi ≥ λ2 kxk2p ≥
P(E0 ).
σ,Z
3
27 Z
(1)
On the other hand |Zi |σi has the same distribution as Zi , and moreover
!
r
X
2
2
2 2
2
P
λkvkp
xi Zi ≥ λ kxkp = P |hx, Zi| ≥
Z
Z
3
3
!
r
2
≤ P kxkp Z̃ ≥
λkxkp + O(k −1/p )
Z
3
≤
C
+ O(k −1/p )
λp
(2)
where Z̃ ∼ Dp . The inequalities are obtained via Theorem 6 and Lemma 2. Combining (1), (2) yields
P(E0 ) ≤
Z
27C
+ O(k −1/p ).
λp
Lemma 9. Let x(1) , x(2) , . . . x(m) ∈ Rn satisfy 0 x(1) x(2) . . . x(m) . Let Z ∈ Rn have k-wise
independent entries marginally distributed according to Dp . Then for some Cp depending only on p,
1
−1/p
P sup |hZ, x(k) i| ≥ λkx(m) kp ≤ Cp
.
+
k
λ2p/(2+p)
k≤m
5
Proof. Observe that for any β we have
P
sup |hZ, x(k) i| ≥ λkx(m) kp
k≤m
≤P
X
Zi2 (x(m) )2i ≥ β 2 kx(m) k2p
i
+P
sup |hZ, x(k) i| ≥ λkx(m) kp
k≤m
!
X
Zi2 (x(m) )2i < β 2 kx(m) k2p .
Lemma 8 directly implies that
P
X
C
C
Zi2 (x(m) )2i ≥ β 2 kx(m) k2p ≤ p + 1/p .
β
k
(3)
On the other hand we can write Zi = |Zi |σi , where σi are k-wise independent Rademacher random
variables, independent from |Zi |. Let us define w(k) ∈ Rn for k ∈ [m] to be the vector with coordinates
(w(k) )i := (x(k) )i |Zi |, so that hx(k) , Zi = hw(k) , σi, and in particular
sup hZ, x(i) i = sup hσ, w(i) i .
k≤m
k≤m
Now, if we condition on |Z1 |, . . . |Zn |, then the sequence w(1) , . . . w(k) of vectors satisfies the assumptions
of Theorem 5, and we can conclude that
Cβ 2
λ (m)
(k)
P sup hσ, w i > kw k2 ≤ 2 .
β
λ
k≤m
Moreover if |Zi | are such that
P
Zi2 (x(m) )2i ≤ β 2 kx(m) k2p , or equivalently kw(m) k22 ≤ β 2 kx(m) k2p , we have
P
sup hσ, w
(k)
(m)
i > λkx
k≤m
kp
≤
Cβ 2
,
λ2
which implies
P
sup |hZ, x(k) i| ≥ λkx(m) kp
k≤m
X
(m) 2
(Zi xi
) < βkx(m) k2p
≤
Cβ 2
.
λ2
This together with Eq. (3) yields
1
Cβ 2
C
(k)
(m)
P sup |hZ, x i| ≥ λkx kp ≤ p + 2 + 1/p
β
λ
k
k≤m
2
We can take β := Θ(λ 2+p ), to have
5.1
1
βp
+
Cβ 2
λ2
2p
= O(λ− 2+p ).
Weak tracking of kxkp
In this section we upper bound the number of rows needed in Indyk’s p-stable sketch with boundedly
independent entries to achieve weak tracking.
Lemma 10. Let x(1) , . . . x(m) ∈ Rn be any sequence satisfying 0 x(1) x(2) . . . x(m) . Take Π ∈ Rd×n
to be a random matrix with entries drawn according to Dp , and such that the rows are r-wise independent,
and all entries within a row are s-wise independent.
For every k ∈ [m], define αk to be median |(Πx(k) )1 |, . . . , |(Πx(k) )d | . If d = Ω(ε−2 (lg 1ε + lg 1δ )), r =
Ω(lg 1ε + lg δ1 ) and s = Ω(ε−p ), then with probability at least 1 − δ we have
∀k ∈ [m], kx(k) kp − εkx(m) kp ≤ αk ≤ kx(k) kp + εkx(m) kp
6
Proof. Consider a sequence of indices 1 < t1 < t2 < . . . < tq+1 = m, constructed inductively in the following
way. We take t1 to be the smallest index with kx(t1 ) kp ≥ ε4/p kx(m) kp . Given tk , we take tk+1 to be the
smallest index such that kx(tk+1 ) − x(tk ) kp ≥ ε4/p kx(m) kp if there exists one, and tk+1 = m otherwise.
Observe that q ≤ ε−8/p . Indeed, for p ≥ 1 we have
X
X
kx(m) kpp = kx(t1 ) +
(x(ti+1 ) − x(ti ) )kpp ≥ kx(t1 ) kpp +
kx(ti+1 ) − x(ti ) kpp ≥ qε4 kx(m) kpp
1≤i<q
1≤i<q
P
where the inequality kx(t1 ) + i≥1 (x(ti+1 ) − x(ti ) )kpp ≥ kx(t1 ) kpp + kx(ti+1 ) − x(ti ) kpp holds because all vectors
x(1) and x(ti+1 ) − x(ti ) for every i have non-negative entries — we can consider
each
separately,
P
P coordinate
and use the fact that for p ≥ 1 and nonnegative numbers ai we have ( ai )p ≥
api — or equivalently,
kakp1 ≥ kakpp . After rearranging this yields q ≤ ε−4 ≤ ε−8/p .
P
P
Similarly, for p ≤ 1, we have that for non-negative numbers ai , ( i≤q ai )p ≥ q p−1 api (this is true
P
P p
because for fixed
ai , the sum
ai is maximized when all ai are equal), and therefore
X
X
kx(ti+1 ) − x(ti ) kpp ≥ q p ε4 kx(m) kpp
kx(m) kpp = kx(t1 ) +
(x(ti+1 ) − x(ti ) )kpp ≥ q p−1 kx(t1 ) kpp +
P
1≤i<q
1≤i<q
which implies q ≤ ε−4/p .
For j ∈ [m], let us define
lj := #{i : |hπi , x(j) i| < (1 − ε)kx(j) kp }
uj := #{i : |hπi , x(j) i| > (1 + ε)kx(j) kp }.
Let π̃i be a vector of i.i.d. random variables drawn according to Dp . We know that hπ̃i , x(j) i ∼ kx(j) kp Dp .
Hence P(|hπ̃i , x(j) i| > kx(j) kp ) = 12 , and P(|hπ̃i , x(j) i| > (1 + ε)kx(j) kp ) ≤ 21 − 2Cε for some universal constant
C. Similarly P(|hπ̃i , x(j) i| < (1 − ε)kx(j) kp ) ≤ 21 − 2Cε.
Entries of πi are s-wise independent, for s ≥ C2 ε−p with some large constant C2 depending on C. Thus by
Theorem 6, P(|hπi , x(j) i| < (1 − ε)kx(j) kp ) ≤ P(|hπ̃i , x(j) i| < (1 − ε)kx(j) kp ) + Cε ≤ 12 − Cε, and analogously
for P(|hπi , x(j) i| > (1 + ε)kx(j) kp ) < 21 − Cε.
Hence
1
E lj ≤ d
− Cε
2
1
− Cε .
E uj ≤ d
2
For j ∈ [q], let Sj be the event
d Cd
d Cd
ltj ≤ −
ε ∧ utj ≤ −
ε
2
2
2
2
Note that for fixed j and varying i, indicator random variables for the events “|hπi , x(j) i| < (1−ε)kx(j) kp ”
are r-wise independent. Thus by Theorem 7, P(Sj ) ≥ 1 − C ′ exp(−Ω(dε2 )) − exp(−Ω(r)). Taking d =
8/p
1
Ωp (ε−2 (lg 1ε + lg δ1 )) and r = Ωp (lg εδ
) we can get P(Sj ) ≥ 1 − δε2 , and hence by a union bound all Sj hold
simultaneously except with probability at most δ2 since the number of events Sj is q ≤ ε−8/p .
For i ∈ [d] and j ∈ [q], let Ei,j be the event
∃t ∈ [tj , tj+1 − 1], |hx(t) − x(tj ) , πi i| > εkx(m) kp .
7
By construction of the sequence tj , all x(t) − x(tj ) above have ℓp norm at most ε4 kx(m) kp , we can invoke
2p
4/p 2+p
Lemma 9 to deduce that P(Eij ) ≤ C3 ε ε
+ C3 s−1/p ≤ C3 ε + C3 s−1/p . Again if we pick s ≥ C4 ε−p
for sufficiently large C4 and small enough ε we have P(Eij ) ≤
E
X
1Eij ≤
i
C
4 ε.
Therefore for any fixed j, we have
C
dε
4
And finally again by Theorem 7, for each j
X
C
1Eij ≥ dε) . exp(−C ′ dε) + exp(−C ′ r)
P(
2
i
1
δ
We have d ≥ C3 ε−2 lg δε
, and q ≤ ε−8/p , hence for sufficiently small ε, we have exp(−C ′ dε) ≤ 2q
. On
1
δ
′
the other hand if r = Ωp (lg δε ) is sufficiently large, we have exp(−C r) ≤ 2q . We invoke the union bound
over all j to deduce that with probability at least 1 − 2δ the following event V holds:
∀j,
X
1Eij ≤
i
C
dε.
2
We know that with probability at least 1 − δ simultaneously V and all the events Sj hold. We will
show now that, when these events all hold, then ∀k kx(k) kp − Kεkx(m) kp ≤ αk ≤ kx(k) kp + Kεkx(m) kp for
some universal constant K. Indeed, consider some k, and let us assume that tj ≤ k ≤ tj+1 . With event Sj
satisfied, we know that #{i : |hπi , x(tj ) i| ≤ kx(tj) kp + εkx(m) kp } ≥ d 12 + Cε
2 , and with event V satisfied,
(k)
(tj )
(m)
we know that for all but Cε
d
of
indices
i
we
have
|hπ
,
x
−
x
i|
≤
εkx
k.
i
2
(k)
(tj )
(k)
(tj )
By the triangle inequality |hπi , x i| ≤ |hπi , x i| + |hπi , x − x i|, yielding
#{i : |hπi , vk i| ≤ kvtj kp + 2εkvm kp } ≥
d
.
2
With similar reasoning we can deduce that
#{i : |hπi , x(k) i| ≥ kx(tj ) kp − 2εkx(m) kp } ≥
d
,
2
which implies the median of |hπi , x(k) i| over i ∈ [d] is in the range kx(tj ) kp ± 2εkx(m) kp . In other words
kx(tj ) kp − 2εkx(m) kp ≤ αk ≤ kx(ti ) kp + 2εkx(m) kp .
Finally we also have kx(k) kp − kx(tj ) kp ≤ εkx(m) kp by construction of the sequence {tj }qj=1 , so the claim
follows up to rescaling ε by a constant factor.
Lemma 11. The above algorithm can be implemented using O(ε−2 lg(1/(εδ)) lg m) bits of memory to store
fixed precision approximations of all counters (Πx(k) )i , and O(ε−p lg(1/(εδ)) lg(nm)) bits to store Π.
Proof. Consider a sketch matrix Π as in Lemma 10 — i.e. Π ∈ Rd×n with random Dp entries, such that
all rows are r-wise independent and all entries within a row are s-wise independent. Moreover let us pick
some γ = Θ(εm−1 ) and consider discretization Π̃ of Π, namely each entry Π̃ij is equal to Πij rounded
to the nearest integer multiple of γ. The analysis identical to the one in [KNW10a, A.6] shows that this
discretization have no significant effect on the accuracy of the algorithm, and moreover that one can sample
from a nearby distribution using only τ = O(lg mε−1 ) uniformly random bits. Therefore we can store such
a matrix succinctly using O (rs(lg n + τ ) + r lg d) bits of memory, by storing a seed for a random r-wise
independent hash function h : [d] → {0, 1}O(s(lg n+τ )) and interpreting each h(i) as a seed for an s-wise
8
independent hash function describing the i-th row of Π̃ [Vad12, Corollary
Hence the
−1 3.34].
total space
lg δ +lg ε−1
complexity of storing the sketch matrix Π̃ in a succinct manner is O
(lg n + lg m) bits.
εp
Additionally we have to store the sketch of the current frequency vector itself, i.e. for all i ∈[d] we need to
store hπ̃i , x(k) i; for every such counter we need O(lg mε−1 ) = O(lg m) bits, and there are d = O
counters.
lg ε−1 +lg δ−1
ε−2
We thus have the following main theorem of this section.
Theorem 12. For any p ∈ (0, 2] there is an insertion-only streaming
algorithm that provides
the weak
n
−1
−1
(lg
ε
+
lg
δ
)
bits
of
memory.
tracking guarantees for f (x) = kxkp with probability 1−δ using O lg m+lg
ε2
5.2
Strong tracking of kxkp
In this section we discuss achieving a strong tracking guarantee. The same argument for ℓ2 -tracking appeared
in [BCI+ 17]. The reduction is in fact general, and shows that for any monotone function f the strong tracking
problem for f reduces to the weak tracking version of the same problem with smaller failure probability.
Lemma 13. Let f : Rn → R+ be any monotone functon of Rn (i.e. x y =⇒ f (x) ≤ f (y)), such
that mini f (ei ) = 1 (where ei are standard basis vectors). Let A be an insertion-only streaming algorithm
satisfying weak tracking for any sequence of updates with probability 1 − δ and accuracy ε. Then for a
sequence of frequency vectors 0 x(1) . . . x(m) algorithm A satisfies strong tracking with probability
1 − δ lg f (x(m) ) and accuracy 2ε.
Proof. Define t1 < t2 < · · · < tq so that ti is the smallest index in [m] larger than ti−1 with f (x(ti ) ) ≥ 2i (if
no such index exists, define q = i and tq = m). Note that q ≤ lg f (x(m) ).
The algorithm will fail with probability at most δ to satisfy the conclusion of Theorem 12 for a particular
sequence of vectors x(1) , x(2) , . . . x(tj ) . That is, for every j, with probability 1 − δ, we have that
∀i ≤ tj , f (x(i) ) − εf (x(tj ) ) ≤ f˜i ≤ f (x(i) ) + εf (x(tj ) ),
where f˜t is the estimate output by the algorithm at time t.
We can union bound over all j ∈ [q] to deduce that except with probability qδ ≤ δ lg f (x(m) ),
∀i ≤ tj , f (x(i) ) − εf (x(tj ) ) ≤ f˜i ≤ f (x(i) ) + εf (x(tj ) ).
By construction of the sequence of tj , we know that for every i, if we take tj to be smallest such that
i ≤ tj , then f (x(tj ) ) ≤ 2f (x(i) ), and the claim follows.
Theorem 14. For any p ∈ (0, 2] there is an insertion-only streaming algorithm that provides strong tracking
guarantees for estimating the ℓp -normof the frequency vector with probability
1 − δ and multiplicative error
1 + ε, with space usage bounded by O
lg m+lg n
(lg ε−1
ε2
+ lg δ −1 + lg lg m) bits.
Proof. This follows from Lemma 11 and Lemma 13 by observing that after a sequence of m insertions, the
ℓp norm of the frequency vector is bounded by m2 , i.e. lg(kx(m) kp ) = O(lg m).
References
[AKO11]
Alexandr Andoni, Robert Krauthgamer, and Krzysztof Onak. Streaming algorithms via precision
sampling. In Proceedings of the 52nd IEEE Annual Symposium on Foundations of Computer
Science (FOCS), pages 363–372, 2011.
9
[AMS99]
Noga Alon, Yossi Matias, and Mario Szegedy. The space complexity of approximating the
frequency moments. J. Comput. Syst. Sci., 58(1):137–147, 1999.
[BC15]
Vladimir Braverman and Stephen R. Chestnut. Universal sketches for the frequency negative moments and other decreasing streaming sums. In Proceedings of the 18th International
Workshop on Approximation, Randomization, and Combinatorial Optimization: Algorithms and
Techniques (APPROX), pages 591–605, 2015.
[BCI+ 17]
Vladimir Braverman, Stephen R. Chestnut, Nikita Ivkin, Jelani Nelson, Zhengyu Wang, and
David P. Woodruff. BPTree: an ℓ2 heavy hitters algorithm using constant memory. In Proceedings of the 36th SIGMOD-SIGACT-SIGART Symposium on Principles of Database Systems
(PODS), 2017.
[BCIW16] Vladimir Braverman, Stephen R. Chestnut, Nikita Ivkin, and David P. Woodruff. Beating
countsketch for heavy hitters in insertion streams. In Proceedings of the 48th Annual ACM
SIGACT Symposium on Theory of Computing (STOC), pages 740–753, 2016.
[BCKY17] Vladimir Braverman, Stephen R. Chestnut, Robert Krauthgamer, and Lin F. Yang. Streaming
symmetric norms via measure concentration. In Proceedings of the 49th Annual ACM Symposium
on Theory of Computing (STOC), to appear, 2017.
[BCWY16] Vladimir Braverman, Stephen R. Chestnut, David P. Woodruff, and Lin F. Yang. Streaming
space complexity of nearly all functions of one variable on frequency vectors. In Proceedings of the
35th ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems (PODS),
pages 261–276, 2016.
[BG06]
Lakshminath Bhuvanagiri and Sumit Ganguly. Estimating entropy over data streams. In Proceedings of the 14th Annual European Symposium on Algorithms (ESA), pages 148–159, 2006.
[BGKS06] Lakshminath Bhuvanagiri, Sumit Ganguly, Deepanjan Kesh, and Chandan Saha. Simpler algorithm for estimating frequency moments of data streams. In Proceedings of the 17th Annual
ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 708–713, 2006.
[BKSV14] Vladimir Braverman, Jonathan Katzman, Charles Seidell, and Gregory Vorsanger. An optimal algorithm for large frequency moments using o(n1−2/k ) bits. In Proceedings of the 17th
International Workshop on Approximation, Randomization, and Combinatorial Optimization:
Algorithms and Techniques (APPROX), pages 531–544, 2014.
[BO10a]
Vladimir Braverman and Rafail Ostrovsky. Measuring independence of datasets. In Proceedings
of the 42nd ACM Symposium on Theory of Computing (STOC), pages 271–280, 2010.
[BO10b]
Vladimir Braverman and Rafail Ostrovsky. Zero-one frequency laws. In Proceedings of the 42nd
ACM Symposium on Theory of Computing (STOC), pages 281–290, 2010.
[BO13]
Vladimir Braverman and Rafail Ostrovsky. Approximating large frequency moments with pickand-drop sampling. In Proceedings of the 16th International Workshop on Approximation, Randomization, and Combinatorial Optimization: Algorithms and Techniques (APPROX), pages
42–57, 2013.
[BOR15]
Vladimir Braverman, Rafail Ostrovsky, and Alan Roytman. Zero-one laws for sliding windows and universal sketches. In Proceedings of the 18th International Workshop on Approximation, Randomization, and Combinatorial Optimization: Algorithms and Techniques (APPROX),
pages 573–590, 2015.
[BR94]
Mihir Bellare and John Rompel. Randomness-efficient oblivious sampling. In Proceedings of the
35th Annual IEEE Symposium on Foundations of Computer Science (FOCS), pages 276–287,
1994.
10
[BW01]
Shivnath Babu and Jennifer Widom. Continuous queries over data streams. SIGMOD Record,
30(3):109–120, 2001.
[BYJKS04] Ziv Bar-Yossef, T. S. Jayram, Ravi Kumar, and D. Sivakumar. An information statistics approach to data stream and communication complexity. Journal of Computer and System Sciences, 68(4):702–732, 2004.
[CBM06]
Amit Chakrabarti, Khanh Do Ba, and S. Muthukrishnan. Estimating entropy and entropy norm
on data streams. Internet Mathematics, 3(1):63–78, 2006.
[CÇC+ 02] Donald Carney, Ugur Çetintemel, Mitch Cherniack, Christian Convey, Sangdon Lee, Greg Seidman, Michael Stonebraker, Nesime Tatbul, and Stanley B. Zdonik. Monitoring streams - A new
class of data management applications. In Proceedings of the 28th International Conference on
Very Large Data Bases (VLDB), pages 215–226, 2002.
[CCM10]
Amit Chakrabarti, Graham Cormode, and Andrew McGregor. A near-optimal algorithm for
estimating the entropy of a stream. ACM Trans. Algorithms, 6(3):51:1–51:21, 2010.
[CKS03]
Amit Chakrabarti, Subhash Khot, and Xiaodong Sun. Near-optimal lower bounds on the multiparty communication complexity of set disjointness. In Proceedings of the 18th Annual IEEE
Conference on Computational Complexity (CCC), pages 107–117, 2003.
[CM05]
Graham Cormode and S. Muthukrishnan. Space efficient mining of multigraph streams. In
Proceedings of the 24th ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database
Systems (PODS), pages 271–282, 2005.
[DKN10]
Ilias Diakonikolas, Daniel M. Kane, and Jelani Nelson. Bounded independence fools degree-2
threshold functions. In Proceedings of the 51st Annual IEEE Symposium on Foundations of
Computer Science (FOCS), pages 11–20, 2010.
[FM85]
Philippe Flajolet and G. Nigel Martin. Probabilistic counting algorithms for data base applications. J. Comput. Syst. Sci., 31(2):182–209, 1985.
[Gan15]
Sumit Ganguly. Taylor polynomial estimator for estimating frequency moments. In Proceedings
of the 42nd International Colloquium on Automata, Languages, and Programming (ICALP),
pages 542–553, 2015.
[GIM08]
Sudipto Guha, Piotr Indyk, and Andrew McGregor. Sketching information divergences. Machine
Learning, 72(1-2):5–19, 2008.
[Gro09]
André Gronemeier. Asymptotically optimal lower bounds on the nih-multi-party information
complexity of the and-function and disjointness. In Proceedings of the 26th International Symposium on Theoretical Aspects of Computer Science (STACS), pages 505–516, 2009.
[HNO08]
Nicholas J. A. Harvey, Jelani Nelson, and Krzysztof Onak. Sketching and streaming entropy via
approximation theory. In Proceedings of the 49th Annual IEEE Symposium on Foundations of
Computer Science (FOCS), pages 489–498, 2008.
[HTY14]
Zengfeng Huang, Wai Ming Tai, and Ke Yi. Tracking the frequency moments at all times. CoRR,
abs/1412.1763, 2014.
[IM08]
Piotr Indyk and Andrew McGregor. Declaring independence via the sketching of sketches. In
Proceedings of the 19th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages
737–745, 2008.
[Ind06]
Piotr Indyk. Stable distributions, pseudorandom generators, embeddings, and data stream
computation. J. ACM, 53(3):307–323, May 2006.
11
[IW03]
Piotr Indyk and David P. Woodruff. Tight lower bounds for the distinct elements problem. In
Proceedings of the 44th Annual IEEE Symposium on Foundations of Computer Science (FOCS),
pages 283–, 2003.
[IW05]
Piotr Indyk and David P. Woodruff. Optimal approximations of the frequency moments of data
streams. In Proceedings of the 37th Annual ACM Symposium on Theory of Computing (STOC),
pages 202–208, 2005.
[Jay09]
T. S. Jayram. Hellinger strikes back: A note on the multi-party information complexity of
AND. In Proceedings of the 12th International Workshop on Randomization and Approximation
Techniques (RANDOM), pages 562–573, 2009.
[Jay13]
T. S. Jayram. On the information complexity of cascaded norms with small domains. In IEEE
Information Theory Workshop (ITW), pages 1–5, 2013.
[JW09]
T. S. Jayram and David P. Woodruff. The data stream space complexity of cascaded norms. In
Proceedings of the 50th Annual IEEE Symposium on Foundations of Computer Science (FOCS),
pages 765–774, 2009.
[JW13]
T. S. Jayram and David P. Woodruff. Optimal bounds for johnson-lindenstrauss transforms and
streaming problems with subconstant error. ACM Trans. Algorithms, 9(3):26:1–26:17, 2013.
[KNPW11] Daniel M. Kane, Jelani Nelson, Ely Porat, and David P. Woodruff. Fast moment estimation
in data streams in optimal space. In Proceedings of the 43rd ACM Symposium on Theory of
Computing (STOC), pages 745–754, 2011.
[KNW10a] Daniel M. Kane, Jelani Nelson, and David P. Woodruff. On the exact space complexity of
sketching and streaming small norms. In Proceedings of the 21st Annual ACM-SIAM Symposium
on Discrete Algorithms (SODA), pages 1161–1178, 2010.
[KNW10b] Daniel M. Kane, Jelani Nelson, and David P. Woodruff. An optimal algorithm for the distinct elements problem. In Proceedings of the 29th SIGMOD-SIGACT-SIGART Symposium on
Principles of Database Systems (PODS), pages 41–52, 2010.
[Li08]
Ping Li. Estimators and tail bounds for dimension reduction in ℓα (0 < α ≤ 2) using stable
random projections. In Proceedings of the 19th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 10–19, 2008.
[Li09]
Ping Li. Compressed counting. In Proceedings of the 20th Annual ACM-SIAM Symposium on
Discrete Algorithms (SODA), pages 412–421, 2009.
[Nel11]
Jelani Nelson. Sketching and streaming high-dimensional vectors. PhD thesis, Massachusetts
Institute of Technology, 2011.
[Nol17]
J. P. Nolan. Stable Distributions - Models for Heavy Tailed Data. Birkhauser, Boston, 2017. In
progress, Chapter 1 online at http://fs2.american.edu/jpnolan/www/stable/stable.html.
[NW10]
Jelani Nelson and David P. Woodruff. Fast manhattan sketches in data streams. In Proceedings
of the 29th ACM SIGMOD-SIGACT-SIGART Symposium on Principles of Database Systems
(PODS), pages 99–110, 2010.
[OJW03]
Chris Olston, Jing Jiang, and Jennifer Widom. Adaptive filters for continuous queries over
distributed data streams. In Proceedings of the ACM SIGMOD International Conference on
Management of Data (SIGMOD), pages 563–574, 2003.
[TGNO92] Douglas B. Terry, David Goldberg, David A. Nichols, and Brian M. Oki. Continuous queries
over append-only databases. In Proceedings of the ACM SIGMOD International Conference on
Management of Data (SIGMOD), pages 321–330, 1992.
12
[Vad12]
Salil P. Vadhan. Pseudorandomness. Foundations and Trends in Theoretical Computer Science,
7(1-3):1–336, 2012.
[Woo04]
David Woodruff. Optimal space lower bounds for all frequency moments. In Proceedings of the
15th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 167–175, 2004.
13
| 8cs.DS
|
arXiv:1705.10212v1 [math.GR] 29 May 2017
Effective Separability of Finitely Generated Nilpotent Groups
1
Effective Twisted Conjugacy
Separability of Nilpotent Groups
Jonas Deré∗ and Mark Pengitore†
May 30, 2017
Abstract
This paper initiates the study of effective twisted conjugacy separability for finitely generated groups, which measures the complexity of separating distinct twisted conjugacy classes
via finite quotients. The focus is on nilpotent groups, and our main result shows that there
is a polynomial upper bound for twisted conjugacy separability. That allows us to study regular conjugacy separability in the case of virtually nilpotent groups, where we compute a
polynomial upper bound as well. As another application, we improve the work of the second
author by giving a precise calculation of conjugacy separability for finitely generated nilpotent
groups of nilpotency class 2.
1
Introduction
Let G be a finitely generated group with a fixed automorphism ϕ . We say that x, y ∈ G are ϕ twisted conjugate if there exists z ∈ G such that z x ϕ (z)−1 = y. That is an equivalence relation
and the equivalence classes of x ∈ G is called the ϕ -twisted conjugacy class which is denoted as
[x]ϕ . We say G is ϕ -twisted conjugacy separable if for each non-ϕ -twisted conjugate pair x, y ∈ G,
there exists a surjective group morphism to a finite group π : G → Q such that π (x) ∈
/ π ([y]ϕ ), or
/ π ([x]ϕ ). When G is ϕ -twisted conjugacy separable for all ϕ ∈ Aut(G), we
equivalently, π (y) ∈
say that G is twisted conjugacy separable.
The interest in twisted conjugacy classes arises in many different areas of mathematics such as
Reidemeister fixed point theory [17, 18, 26], Selberg theory [1, 39], and algebraic geometry [23].
Twisted conjugacy separability was originally introduced formally in [19] and has been further
studied in [16].
When ϕ = id, the ϕ -twisted conjugacy classes are equal to the usual notion of conjugacy classes.
Similarly, ϕ -twisted conjugacy separable is equivalent to the usual notion of conjugacy separable
∗ KU Leuven Kulak, Kortrijk, Belgium.
Email: [email protected].
fellowship of the Research Foundation -- Flanders (FWO).
† Purdue University, West Lafayette, IN. E-mail: [email protected]
Supported by a postdoctoral
Effective Separability of Finitely Generated Nilpotent Groups
2
when ϕ = id . Thus, the notion of ϕ -twisted conjugacy separability is a natural generalization of
conjugacy separability where we allow our conjugacy classes to be twisted by ϕ . In particular,
if G is a finitely generated group that contains a characteristic, finite index, twisted conjugacy
separable subgroup, then G is twisted conjugacy separable, see [16, Thm 5.2]. That is in contrast to conjugacy separability which is not closed with respect to finite extensions or finite index
subgroups [21, 31]. Thus, twisted conjugacy separability gives us an important tool in studying
the more classical notion of conjugacy separability since twisted conjugacy separability implies
conjugacy separability.
Conjugacy separability, along with residually finiteness, subgroup separability, and other residual
properties, have been extensively studied and used in resolving important conjectures in geometry
such as Agol’s work on the Virtual Haken conjecture. Previous work in the literature has been
to understand what groups satisfy these properties. For instance, polycyclic groups, free groups,
residually free groups, limit groups, Bianchi groups, surface groups, fundamental groups of compact, orientable 3-manifolds, and virtually compact special hyperbolic groups have been shown to
be conjugacy separable and subsequently, residually finite [2, 12, 13, 14, 20, 25, 32, 40]. Recently,
there has been considerable activity in establishing effective versions of the above separability
properties (see [3, 4, 5, 6, 7, 8, 9, 10, 11, 27, 28, 29, 34, 35, 37, 41]). The main purpose of this
article is to improve on the effective conjugacy separability results of [36] for nilpotent groups and
to establish effective twisted conjugacy separability for the class of virtual nilpotent groups.
For a finitely generated group G with a finite generating subset S and an automorphism ϕ , we
ϕ
introduce a function ConjG,S (n) on the natural numbers that quantifies ϕ -twisted conjugacy separability. To be specific, the value on a natural number n is the maximum order of the minimal
finite quotient needed to distinguish pairs of non-ϕ -conjugate elements as one varies over the nball. We also quantify the more general property of twisted conjugacy separability via the function
TconjG,S (n). We start by establishing a norm kϕ kS for automorphisms and then define TconjG,S (n)
ϕ
on a natural number n to be the maximum value of ConjG,S (n) as one varies over automorphisms
ϕ satisfying kϕ kS ≤ n.
ϕ
When the automorphism ϕ is the identity map, the function ConjG,S is equal to the function that
quantifies conjugacy separability introduced by Lawton-Louder-McReynolds [30], which is in
this case denoted as ConjG,S (n). As a natural consequence, ConjG,S (n) is always bounded by
TconjG,S (n). Additionally, we will see that if H is a characteristic, twisted conjugacy separable
subgroup of G of index r, then there exist automorphisms { fi }ri=1 of H such that ConjG,S (n) is
fi
r
′
bounded by ∏ri=1 (ConjH,S
′ (n)) where S and S are finite generating subsets of G and H, respectively (see Theorem 7.7). Hence, if we are given a class of conjugacy separable groups that is
closed under finite index subgroups, then we may study conjugacy separability of any finite extension of these groups by studying effective twisted conjugacy separability of the original group.
So far, previous papers have only studied the asymptotic behavior of ConjG,S (n). For instance, if
G is a finitely generated nilpotent group, then the second author [36] demonstrated ConjG,S (n)
nd1 for some d1 ∈ N, and if, in addition, G is a finitely generated nilpotent group that is not
virtually abelian, then there exists d2 ∈ N such that nd2 ConjG,S (n). Lawton-Louder-McReynolds
2
demonstrated that ConjG,S (n) nn when G is a finite rank free group or a surface group. This
Effective Separability of Finitely Generated Nilpotent Groups
3
article is the first to study effective twisted conjugacy separability for any class of groups. As of
now, there are no effective proofs of conjugacy separability for classes of groups such as polycyclic
groups, fundamental groups of compact, orientable 3-manifolds, Bianchi groups, etc. Even in the
context of surface groups or free groups, there is no good asymptotic lower bound for ConjG,S (n)
other than the one provided by effective residual finiteness.
To state our results, we require some notation. For two non-decreasing functions f , g : N → N, we
write f (n) g(n) if there exists a C ∈ N such that f (n) ≤ C g(C n) for all n. We write f ≈ g when
f g and g f . If N is a nilpotent group, we write c(N) for its nilpotent class, as we will recall
in Section 2.3.
Our first result is the precise computation of ConjN,S (n) when N is a 2-step, torsion free, finitely
generated nilpotent group and S is a finite generating subset.
Theorem 5.4. Let N be a torsion free, finite generated nilpotent group with a finite generating
subset S such that c(N) = 2. Then there exists Φ(N) ∈ N such that ConjN,S (n) ≈ nΦ(N) .
An explicit expression for Φ(N) given in Section 2.3.
The lower bound is provided by [36, Thm 1.8]. For the upper bound, we improve on the techniques
of [36] in the context of two step nilpotent groups to obtain precise asymptotic upper bounds. The
purpose of this result is to both improve the results of [36] in the context of two step nilpotent
groups and to demonstrate the techniques used in the proof of the main theorem.
The next result is the main theorem of the article. In the following theorem, we give the first effective upper bound for ϕ -twisted conjugacy separability for nilpotent groups and the first effective
upper bound for twisted conjugacy separability for any finitely generated group.
Theorem 6.3. Let N be a torsion free, finitely generated nilpotent group with a finite generating
subset S. Let x ∈ N and ϕ ∈ Aut(N). Then there exist natural numbers k1 , k2 , k3 such that
ϕ
ConjN,x,S (n) (kϕ kS )k1 (kxkS )k2 nk3 .
ϕ
In particular, ConjN,S (n) (kϕ kS )k1 nk2 +k3 and TconjN,S (n) nk1 +k2 +k3 .
The proof of Theorem 6.3 generalizes the techniques of [2] and [36] to the context of ϕ -conjugacy
classes by introducing the notion of the i-th twisted centralizer corresponding to an automorphism
ϕ . When ϕ = Inn(x), then the i-th twisted centralizer is equal to the group of elements that
centralize x modulo the i-th term of the lower central series. Using these twisted centralizers, we
may proceed by induction on step length.
More generally, we have a similar result as Theorem 6.3 for virtually nilpotent groups.
Theorem 7.8. Suppose that G is a virtually nilpotent group, and suppose that S is a finite generating subset of G. For ϕ ∈ Aut(G) and x ∈ G, there exist natural numbers k1 , k2 , k3 such that
ConjG,x,S (n) (kϕ kS )k1 (kxkS )k2 nk3 .
ϕ
In particular, ConjG,S (n) (kϕ kS )k1 nk2 +k3 and TconjG,S (n) nk1 +k2 +ks3 .
Effective Separability of Finitely Generated Nilpotent Groups
4
For this result, we write the ϕ -twisted conjugacy class of an element of G as a finite union of right
translates of twisted conjugacy classes of a finite index, characteristic finitely generated nilpotent
subgroup. We then apply Theorem 6.3.
The last result of this article extends the work of the second author in [36] to the context of finite
extensions of nilpotent groups.
Theorem 7.11. Let G be a virtually nilpotent group with a finite generating subset S, and let x ∈ G.
There exist k1 , k2 ∈ N such that ConjG,x,S (n) (kxkS )k1 nk2 . In particular, ConjG,S (n) nk1 +k2 . If
G is not virtually abelian, then
n(c(N)−1) Φ(N) ConjG,S (n) nk1 +k2
where N is any infinite finitely generated nilpotent subgroup of finite index in G.
In order to prove the upper bound of Theorem 7.11, we apply Theorem 7.8. For the lower bound,
we follow [36, Thm 1.8] using the fact that every conjugacy class in G is the union of at most
[G : N] conjugacy classes in N.
ϕ
We finish by working out asymptotic upper bounds for ConjH3 (Z) (n) for any ϕ ∈ Aut(H3 (Z))
where H3 (Z) is the 3-dimensional integral Heisenberg group. We also work out a 5-dimension
examples with a fixed automorphism.
Acknowledgements
We would like to thank Karel Dekimpe for his suggestion to study twisted conjugacy classes in the
context of finitely generated nilpotent groups. The second author would like to thank his advisor
Ben McReynolds for his continued support.
2
Background
In this section, we introduce necessary definitions for this paper and start by fixing some notation.
Let G be a group with finite generating subset S. The order of a finite group G is denoted as |G|.
We write kxkS to be the word length of x with respect to S and denote the identity element of G as
1. We denote the order of g as an element of the group G as ordG (g). We define the commutator
of x, y ∈ G as [x, y] = x y x−1 y−1 . For any subset X ⊂ G, we let hX i be the subgroup generated by
the set X . For a normal subgroup H E G, we set πH : G → G/H to be the natural projection and
sometimes write x̄ = πH (x) if the normal subgroup H is clear from the context.
For any x ∈ G, we let Inn(x) ∈ Aut(G) be the associated inner automorphism i.e. Inn(x)(y) =
x y x−1 . For an integer k and prime p, we define v p (k) to be the largest natural number such that
pv p (k) divides k.
We define Gm to be the subgroup generated by m-th powers of elements in G which is a characteristic subgroup. We define the associated projection as rm = πGm . We also define the abelianization
Effective Separability of Finitely Generated Nilpotent Groups
5
of G as Gab = G/[G, G] with the associated projection πab = π[G,G] . We define the center of G as
Z(G) and the centralizer of x in G as CG (x).
2.1
Norms of subgroups and automorphisms
We associate a norm to finitely generated subgroups in the following definition.
Definition 2.1. Let G be a finitely generated group with a finite generating subset S. For any finite
subset X ⊆ G, we define kX kS = max{kxkS | x ∈ X }. For a finitely generated subgroup H ≤ G, we
define
kHkS = min{kX kS | X is a finite generating subset for H}.
Let S1 and S2 be two generating subsets of a group G such thath |s|S2 ≤ K for all s ∈ S1 , or
equivalently, such that kxkS2 ≤ KkxkS1 for all x ∈ G. Then kHkS2 ≤ KkHkS1 for all finitely generated subgroups H ≤ G. Indeed, take generators hi ∈ H for H such that |hi |S1 ≤ kHkS1 , then
|hi |S2 ≤ K|hi |S1 ≤ KkHkS1 . Since these elements generate H, the statement follows. In particular,
we get the following relation between norms of subgroups for different generating subsets.
Lemma 2.2. Let G be a finitely generated group with finite generating subsets S1 and S2 , and let
H ≤ G be a finitely generated subgroup. Then C1 kHkS1 ≤ kHkS2 ≤ C kHkS1 for some C > 0.
Similarly, we define a norm for morphisms of finitely generated groups, and subsequently, define
a norm for automorphisms of a finitely generated group.
Definition 2.3. Let G and G′ be finitely generated groups with finite generating subsets S and S′ .
Let ϕ : G → G′ be a group morphism. We define
kϕ kS,S′ = max{kϕ (s)kS′ | s ∈ S}.
If ϕ is an automorphism of G, then we assume S = S′ and write kϕ kS .
Equivalently, kϕ kS,S′ is the smallest natural number such that kϕ (x)kS′ ≤ kϕ kS,S′ kxkS for all x ∈ G.
Let S1 and S2 be two generating subsets of G such that kskS2 ≤ K for all s ∈ S1 . We show that in
this case, kϕ kS1 ,S′ ≤ Kkϕ kS2 ,S′ . Note that
kϕ (x)kS′ ≤ kϕ kS2 ,S′ kxkS2 ≤ Kkϕ kS2 ,S′ kxkS1
for all x ∈ G, and thus, the conclusion follows.
Similarly, one can show for S1′ and S2′ generating subsets for G′ with kskS2′ ≤ K for all s ∈ S1′ , that
kϕ kS,S1′ ≤ Kkϕ kS,S2′ .
In particular, we have the following lemma.
Lemma 2.4. Let G be a group with two finite generating subsets S, S′ . There exists a constant
C > 0 such that for every automorphism ϕ ∈ Aut(G), we have C1 kϕ kS′ ≤ kϕ kS ≤ Ckϕ kS′ .
Effective Separability of Finitely Generated Nilpotent Groups
2.2
6
Finitely generated groups and separability
Let G be a finitely generated group with a finite generating subset S, and let X ⊆ G be an arbitrary
subset. Following Bou-Rabee in [5], we define the relative depth function
DG (X , · ) : G \ X → N ∪ {∞}
as
DG (X , y) = min{|Q| | π : G ։ Q, |Q| < ∞, and π (y) ∈
/ π (X )}
with the understanding that DG (X , y) = ∞ if no such Q exists.
Definition 2.5. We say that a non-empty subset X ≤ G is separable if D(X , y) < ∞ for all y ∈ G \X .
/ πH (X ).
We say that a finite quotient G/H separates X and y if πH (y) ∈
Recall the twisted conjugacy class [x]ϕ for x ∈ G and ϕ ∈ Aut(G) from Section 1.
Definition 2.6. Let G be a finitely generated group with and let ϕ : G → G be an automorphism.
We say that G is ϕ -twisted conjugacy separable if for all x ∈ G the twisted conjugacy class [x]ϕ is
separable. We say that G is twisted conjugacy separable if G is ϕ -twisted conjugacy separable for
all ϕ ∈ Aut(G).
Let G be a finitely generated group with a finite generating subset S, and let ϕ : G → G be an
automorphism. To quantify ϕ -twisted conjugacy separability relative to x for some fixed x ∈ N,
ϕ
we define the following function ConjG,x,S (n) : N → N given by
ϕ
/ [x]ϕ }.
ConjG,x,S (n) = max{D([x]ϕ , y) | kykS ≤ n, y ∈
By allowing x to vary, we are able to quantify ϕ -twisted conjugacy separability for any given
ϕ
group G and automorphism ϕ . We define ConjG,S : N → N to be given by
ϕ
ϕ
ConjG,S (n) = max{ConjG,x,S (n) | kxkS ≤ n}.
Finally, we obtain a method to quantify twisted conjugacy separability by taking automorphisms
with norm at most n. We now define Tconj : N → N to be given by
ϕ
TconjG,S (n) = max{ConjG,S (n) | ϕ ∈ Aut(G) and kϕ kS ≤ n}.
ϕ
Lemma 2.7. Let S1 and S2 be two finite generating subsets of G. If ϕ ∈ Aut(G), then ConjG,S1 (n) ≈
ϕ
ConjG,S2 (n). Similarly, TconjG,S1 (n) ≈ TconjG,S2 (n).
The proof is similar to [5, Lem 1.1] (see also [6, Lem 1.2] and [30, Lem 2.1]).
ϕ
Observe that if ϕ = id, then ConjG,S is equal to the conjugacy separability ConjG,S introduced
by Lawton, Louder, and McReynolds [30]; subsequently, ConjG,S (n) TconjG,S (n). If, for some
ϕ ∈ Aut(G), we have that ConjϕG,S (n) < ∞ for all n ∈ N, then G is ϕ -twisted conjugacy separable.
Similarly, if TconjG,S (n) < ∞ for all n, then G is twisted conjugacy separable, and subsequently, G
is ϕ -conjugacy separable for all ϕ ∈ Aut(G). In particular, G is conjugacy separable in that case.
Effective Separability of Finitely Generated Nilpotent Groups
2.3
7
Nilpotent Groups
Most of the groups we work on in this paper are nilpotent groups, and we recall their basic properties. See [24, 38] for a more thorough account of the theory of nilpotent groups.
A central series for a group N is a sequence of subgroups 1 = N0 ≤ N1 ≤ · · · ≤ Nc = N such that
[N, Ni+1 ] ≤ Ni . There are two special central series which play an important role when studying
def
nilpotent groups. The i-th term of the lower central series is defined by γ1 (N) = G and inductively
def
def
γi (N) = [G, γi−1 (N)] for i > 1. The i-th term of the upper central series is defined by ζ1 (N) = Z(N)
def
(Z(G/ζi−1 (N))) for i > 1. We denote their associated projections
and inductively ζi (N) = πζ−1
i−1 (N)
i
as πi = πγi (N) and π = πζi (N) when N is understood from context.
Definition 2.8. We say that N is nilpotent of step size c if c is the minimal natural number such
that γc+1 (N) = 1, or equivalently, ζc (N) = N. We write the nilpotency class of N as c(N). We
define the Hirsch length of N as
c(N)
h(N) =
∑
i=1
.
i+1 (N)
rankZ γi (N) γ
We define T (N) to be the finite characteristic subgroup of finite order elements. If N is a torsion
free, finitely generated nilpotent group, we say that N is a F -group.
For nilpotent groups N, the subgroup N m is always of finite index in N.
Lemma 2.9. Let N be a F -group, and let H ≤ N be a subgroup of index pα where p is a prime.
α
Then N p ≤ H.
Proof. For normal subgroups H ≤ N, this result is trivial. Using the fact that in nilpotent groups
every subgroup H in N is subnormal, see [38], meaning that there exists a sequence H0 = H ≤
H1 ≤ . . . ≤ Hk = N such that Hi is normal in Hi+1 , the result follows for every subgroup H.
The following subgroup will be useful in assigning numerical invariants to subgroups and group
morphisms.
Definition 2.10.
Let N be a F -group, and let H ≤ N be a subgroup. We define the isolator of H
√
N
in N, denoted H, as the set
√
N
H = {x ∈ N | there exists k ∈ N such that xk ∈ H}.
√
N
H is a subgroup for all H ≤ N when N is a F -group. Additionally,
From [38], it follows
that
√
N
we have that | H : H| < ∞.
We now define the notion of a determinant of a subgroup of a finite generated nilpotent group.
Definition 2.11. Let N be a√nilpotent group, and let H ≤ N be a subgroup. We define the determinant det(H) as det(H) = | N H : H|. If ϕ : N1 → N2 is a morphism of nilpotent groups, we write
det(ϕ ) = det(ϕ (N1 )).
Effective Separability of Finitely Generated Nilpotent Groups
8
If ϕ is an injective map of torsion-free abelian groups of the same rank, then det(ϕ ) is equal to the
usual determinant of the matrix representative of ϕ with respect to a fixed choice of basis. Thus,
we have a generalization of the usual notion of determinant to a more general class of groups and
group morphisms. If ϕ : A → B is a map of abelian groups where ϕ (A) has rank equal to B, then
det(ϕ )B ⊆ ϕ (A).
We need one more invariant for nilpotent groups.
Definition 2.12. Let N be a F -group, and let z ∈ N be a primitive central element of N. An one
dimensional
quotient of N associated to z is a quotient N H such that N/H is a F -group
central
where Z N H = hπH (z)i.
For any primitive central element, the existence of an associated one dimensional central quotient of N is guaranteed by [36, Prop 3.1]; however, uniqueness is not guaranteed motivating the
following definition.
Definition 2.13. Let N be a F -group. We define Φ(N) to be the smallest integer such that for
N
every primitive
z ∈ Z(N), there exists an one dimensional central quotient H associated to z
such that h N H ≤ Φ(N).
The value Φ(N) for a F -group is the value found in the statement of [36, Thm 1.1] and is originally
defined in [36, Defn 3.4].
3
Twisted centralizers and twisted determinants
In this section, we introduce twisted centralizers and study the projections of these subgroups to
finite quotients. The key concept for understanding these finite projections is the twisted determinant which we introduce in Definition 3.7. From now on, N is a nilpotent group of nilpotency
class c and with a fixed automorphism ϕ : N → N.
Definition 3.1. Let N be a nilpotent group, and let ϕ : N → N be an automorphism. For every
i ≥ 1, we define the subgroups
ϕ
Ni = {x | x ϕ (x)−1 ∈ γi (N)}
and the corresponding subsets
ϕ
ϕ
Xi = {x ϕ (x)−1 | x ∈ Ni } ⊆ γi (N).
ϕ
We call Ni the i-th twisted centralizer corresponding to the automorphism ϕ .
ϕ
ϕ
Note that N1 = N and Nc+1 = {x ∈ N | ϕ (x) = x} by definition. If ϕ = Inn(y) with y ∈ N, then
ϕ
Nc+1 is the centralizer of the element y, hence the name.
ϕ
The subsets Xi play an important role in studying twisted conjugacy classes as they determine if
two elements are twisted conjugate when they differ by an element of γi (N)
Effective Separability of Finitely Generated Nilpotent Groups
9
Lemma 3.2. Let N be a nilpotent group with an automorphism ϕ . For every x, y ∈ N with y ∈
γi (N), it holds that
ϕ
yx ∼ϕ x ⇐⇒ y ∈ Xi x .
Proof. Note that yx ∼ϕ x if and only if there exists z ∈ N such that z x ϕ (z)−1 = y x. This last
ϕ
statement is the same as z x ϕ (z)−1 x−1 = y. Equivalently, y ∈ Xi x .
ϕ
The twisted centralizers Ni are used to define the maps ψϕ ,i .
Definition 3.3. Let N be a nilpotent group, and let ϕ ∈ Aut(N). For each i, we define a map
ψϕ ,i : Niϕ → γi (N) γ (N) as
i+1
ψϕ ,i (x) = x ϕ (x)−1
Lemma 3.4. The map ψϕ ,i : Ni → γi (N) γ
ϕ
i+1 (N)
mod γi+1 (N).
is a group morphism for all i.
ϕ
Proof. Let x, y ∈ Ni . By computation,
ψϕ ,i (xy) =
=
=
=
=
ϕ
xy ϕ (xy)−1
mod γi+1 (N)
−1
ϕ (x)−1 mod γi+1 (N)
x ψϕ ,i (y) ϕ (x)−1 mod γi+1 (N)
ψϕ ,i (x) ψϕ ,i (y) [ψϕ ,i (y)−1 , ϕ (x)] mod γi+1 (N)
ψϕ ,i (x) ψϕ ,i (y).
xy ϕ (y)
ϕ
In order to understand how Ni relates to Ni+1 for each i, we have the following lemma.
ϕ
Lemma 3.5. With notations as above, we have ker(ψϕ ,i ) = Ni+1 .
ϕ
Proof. Take any x ∈ Ni , then x ∈ ker(ψϕ ,i ) if and only if
xϕ (x)−1 γi+1 (N) = ψϕ ,i (x) = 1 mod γi+1 (N)
ϕ
which is equivalent to x ∈ Ni+1 .
ϕ
Each ψϕ ,i induces an injective map ψ̄ϕ ,i : Ni N ϕ → γi (N) γ
i+1
i+1 (N)
of abelian groups.
With the above lemma, we can define a subgroup which will be of importance in the effective
upper bound of twisted conjugacy separability.
ϕ
Definition 3.6. Let N be a nilpotent group, and let ϕ ∈ Aut(N). The set Xc is a subgroup since it
ϕ
is the image of ψϕ ,c . We define the central subgroup Nϕ = Xc ≤ γc+1 (N) ≤ Z(N).
Effective Separability of Finitely Generated Nilpotent Groups
10
One can see that x ∈ Nϕ are elements of γc (N) that take the form x = y ϕ (y)−1 for some y ∈ N.
Definition 3.7. Let Γ be a F -group, and let ϕ ∈ Aut(N). For each i, we define Dϕ ,i = det(ψ̄ϕ ,i ).
Additionally, we define the twisted determinant Dϕ as
c
Dϕ = ∏ Dϕ ,i .
i=1
The twisted determinant is the main ingredient for the following proposition.
Proposition 3.8. Let N be a F -group and p be a prime. There exists a natural number k∗ (p, c)
k+k∗ (p,c)+v p (Dϕ )
such that for every automorphism ϕ , natural number k, and x ∈ N with x ϕ (x)−1 ∈ N p
k
there exists y ∈ N p such that y ϕ (y)−1 = x ϕ (x)−1 . The number k∗ (p, c) can be chosen such that
∗
there exists a constant C > 0 with pk (p,c)+v p (Dϕ ) ≤ C|Dϕ | for all primes p.
Observe that k∗ (p, c) is independent of the choice of automorphism. The above proposition is a
twisted generalization of [2, Lem 2] and is an effective version of [15, Thm 4.1.].
The proof of Proposition 3.8 relies on the following two lemmas.
Lemma 3.9. Let p be a prime,
and let ϕ : A → B be a group morphism of abelian groups. If
b ∈ ϕ (A) ∩ pk+v p (det(ϕ )) · B for some k ≥ 0, then there exists an a ∈ pk · A such that ϕ (a) = b.
ϕ )) · g for some g ∈ B. Since b ∈ ϕ (A), we know that ord(g · ϕ (A)) as
Proof. Write b = pk+v p (det(
p
p
an element of the group B ϕ (A)/ϕ (A) divides pk+v p (det(ϕ )) . Since det(ϕ ) = | B ϕ (A)/ϕ (A)|, it
follows that g · ϕ (A) has order at most pv p (det(ϕ )) . Hence, pv p (det(ϕ )) · g ∈ ϕ (A). Let ã ∈ A with
ϕ (ã) = g, then the element a = pk ã is our desired element.
We reproduce the proof of [2, Lem 2] in order to estimate the associated value that is constructed.
Lemma 3.10. Let N be a F -group of nilpotency class c, and let p be a prime. There exists an
k
k+k(p,c)
, then there exists y ∈ N such that x = y p . Additionally,
integer k(p, c) ≥ 0 such that if x ∈ N p
p(k, c) can be chosen such that pk(p,c) ≤ c! for all primes p.
Proof. For a natural number m, we let d p (m) be the largest integer such that pd p (m) ≤ m. We show
that the lemma holds for the value k(p, c) = ∑ci=1 d p (i). We proceed by induction on nilpotency
class length, and since the statement is evident for abelian groups, we may assume that N has
k+k(p,c)
k+k(p,c)
where xi ∈ N for all i. Then
, we write x = ∏ri=1 xip
nilpotency class c > 1. For x ∈ N p
[24, Thm 6.3] implies that we may write
! pk+k(p,c)
k+k(c,p)
k+k(c,p)
c
c
r
r
(p i )
(p i )
pk+k(p,c)
y
y
=
= ∏ xi
x = ∏ xi
∏i
∏i
i=1
i=1
i=2
i=1
where yi ∈ γi (N) for each i and y1 = ∏ri=1 xi . We may write the binomial coefficient
k+k(p,c) k+k(p,c)
p
p
=
pℓi ui
i
Effective Separability of Finitely Generated Nilpotent Groups
11
where gcd(ui , p) = 1. Thus, this value is divisible by pk+k(p,c)−ℓi ; therefore, we write
pk+k(p,c)
i
= mi pk+k(p,c)−ℓi .
Since N has nilpotent class c and yi ∈ γi (N), we have g pℓi ui = 1 for pℓi ui > c. If pℓi ui ≤ c, then
ℓi ≤ d p (c). Letting si = d p (c) − ℓi , we write
pk+k(p,c)
i
= mi pk+k(p,c)−ℓi = mi pk+k(p,c)+si −d p (c)
for i ≤ c. Subsequently, for i ≤ c we have
(p
yi
k+k(p,c)
i
)
k+k(p,c)+si −d p (c)
ip
= ym
i
si
k+k(p,c)−d p (c)
ip
= (ym
)p
i
.
si
ip
Letting zi = ym
, we may write
i
r
k+k(p,c)
x = ∏ xip
i=1
c
k+k(p,c)−d p (c)
= ∏ zip
i=1
c
k+k(p,c−1)
= ∏ zip
.
i=1
where zi ∈ [N, N] for 2 ≤ i ≤ c. Then [24, Lem 1.3] implies that hz1 , [N, N]i has nilpotency
class strictly less than c. Since {z1 , z2 , · · · , zc } ∈ hz1 , [N, N]i, the inductive hypothesis gives us
k
a y ∈ hz1 , [N, N]i such that x = y p . By construction, k(p, c) = ∑ci=1 d p (i) which implies pk(p,c) ≤
∏ci=1 pd p (i) ≤ c!.
Proof of Proposition 3.8. Let k(p, c) be the natural number from Lemma 3.10. We proceed by
ϕ
induction on i where x ∈ Ni . If i = c + 1, or equivalently, x ϕ (x)−1 ∈ γc+1 (N) = 1, then we take
y = 1.
ϕ
Now assume that x ϕ (x)−1 ∈ γi (N) ∩ N p for k sufficiently big. Letting A = Ni N ϕ
k
i+1
and B =
γi (N)
γi+1 (N), we consider the injective morphism of abelian groups ψ̄ϕ ,i : A → B induced by
k−k(p,c)
k−k(p,c)
which implies ψ̄ϕ ,i (x) ∈ B p
ψϕ ,i . Lemma 3.10 gives us a z ∈ N such that x ϕ (x)−1 = z p
.
k−k(p,c)−v
(det(
ϕ̄
))
p
ϕ p
ϕ
ϕ
Lemma 3.9 shows that there exists xi ∈ (Ni )
such that ψ̄ϕ ,i (x Ni+1 ) = ϕ̄i (xi Ni+1 ).
k−k(p,c)−v p (Di )
k−k(p,c)−v p (Di )
ϕ
−1
p
Hence, xi x ∈ Ni+1 , and since xi , x ∈ N
, it follows that x̃ ∈ N p
. Thus,
k−k(p,c)−v p (Di )
−1
p
x̃ ϕ (x̃) ∈ γi+1 (N) ∩ N
. Induction implies the theorem follows with the constant given
∗
by k (p, c) = (c − 1) k(p, c).
We finish by noting Lemma 3.8 implies that k(p, c) ≤ c !, and thus,
pk
∗ (p,c)+v
p (Dϕ )
= p(c−1)·k(p,c)+v p (Dϕ ) ≤ (pk(p,c) )c−1 pv p (Dϕ ) ≤ (c)c−1 |Dϕ |.
We denote by ϕ̄ pk the automorphism induced by ϕ on the quotient N pk .
N
Effective Separability of Finitely Generated Nilpotent Groups
12
Corollary 3.11. Let N be a F -group of nilpotency class c, and let ϕ : N → N be an automorphism.
Let p be a prime and k be some natural number. Take k∗ (p, c − 1) as in Proposition 3.8, and define
pk+k0 → N/N pk as the natural projection such that
k0 = v p (Dϕ ) + k∗ (p, c − 1). Define ρ : N/N
!
N pk+k0
π pk = ρ ◦ π pk+k0 . Then ρ
= π pk (Nϕ ).
N
ϕ̄ k+k0
p
Proof. First assume that x ∈ Nϕ , or equivalently, x = y ϕ (y)−1 ∈ γc (N) for some y ∈ N. Note that
k+k
π pk+k0 (x) ∈ γc (N/N p 0 ) since π pk+k0 is a group morphism. Now
π pk+k0 (x) = π pk+k0 (yϕ (y)−1 ) = π pk+k0 (y) ϕ pk+k0 (π pk+k0 (y))−1 ∈ N pk+k0
;
N
ϕ̄ k+k0
p
hence, ρ (π pk+k0 (x)) = π pk (x) ∈ ρ
N pk+k0
N
ϕ̄
pk+k0
!
N pk+k0
. Hence, π pk (Nϕ ) ⊆ ρ
N
N pk+k0
For the other inclusion, let x ∈ γc (N) with π pk (x) ∈ ρ
N
ϕ̄
ϕ̄
pk+k0
!
pk+k0
!
.
. Thus, there exists
k+k
y ∈ N such that y ϕ (y)−1 ∈ x N p 0 . Applying Proposition 3.8 to N/γc (N) and the map induced by
k
ϕ , there exists z ∈ N p such that z ϕ (z)−1 ∈ y ϕ (y)−1 γc (N). Hence, y−1 z ϕ (y−1 z)−1 ∈ Nϕ . Since
π pk (z) = π pk (ϕ (z)) = 1, we get
π pk (x) = π pk (y ϕ (y)−1 ) = π pk (z−1 y ϕ (z−1 y)−1 ) ∈ π pk (Nϕ ).
!
N pk+k0
= π pk (Nϕ ).
We conclude that ρ
N
ϕ̄ k+k0
p
4
Bounding the Twisted Determinant
In the previous section, we identified the twisted determinant as the crucial factor for studying the
twisted centralizers in finite quotients. The goal of this section is to bound the twisted determinant
of an automorphism in terms of the norm of that automorphism.
Before we start, we provide some propositions and examples that relate the determinates of subgroups with the norms of those subgroups.
Example 4.1. Consider the group G = Z with standard generating subset S = {±1}. The norm of
the subgroup nZ with n ∈ N satisfies
knZkS = n = [Z : nZ] = det(nZ).
We generalize the previous example to all finitely generated abelian groups.
Proposition 4.2. Let A be a finitely generated abelian group with a finitely generating subset S.
There exists a constant C > 0 such that if H ≤ A is a subgroup, then det(H) ≤ C (kHkS )rankZ (H) .
Effective Separability of Finitely Generated Nilpotent Groups
13
√
Proof. Since T (A) is always a finite normal subgroup with T (A) ≤ A H, we may assume that
A = Zk . Moreover, we may assume that S is the standard generating subset since the statement is
invariant under changing the generating subset.
Let ℓ be the rank of H. There exist h1 , . . . , hl ∈ H such that H ′ = hh1 , · · · , hℓ i has rank l and khi kS ≤
kHkS . Thus, we may restrict our attention to subgroups of the form H ′ since det(H) ≤ det(H ′ )
and kH ′ kS ≤ kHkS .
Let π : Zk → Zℓ be the projection to the first ℓ coordinates. By relabeling coordinates as necessary, we may assume that π |H is injective. Since π |H is injective on H, it follows that π | √
A
H is
p
√
√
√
A
A
A
Zℓ
ℓ
π (H) = Z , it follows that
injective. Thus, [π ( H) : π (H)] = [ H : H], and since π ( H) ≤
det(π (H)) ≥ det(H). Thus, without loss of generality, we may assume that H has full rank in A.
Let X = [h1 , h2 , . . . , hℓ ]T . Since H is full rank, it follows that X has non-zero determinant. Observing that det(X ) = [Zk : H] = det(H) and that each coefficient of X is bounded by kHkS , the
formula for the determinant gives the desired bound.
Applying the above proposition to group morphisms, we have the following.
Corollary 4.3. Let A be a finitely generated abelian group with a finite generating subset S. There
exists a constant C > 0 such that for every morphism ϕ : B → A of abelian groups, it follows that
det(ϕ ) ≤ C(kϕ (B)kS )rank(ϕ (B)) ≤ C(kϕ kS )rank(ϕ (B)).
Proof. Our result follows from Proposition 4.2 since the value only depends on the image.
The following example shows that one cannot bound the norm of ϕ (B) by the determinant of ϕ .
Example 4.4. Consider the group Z2 with the standard generating subset S and subgroup H =
h(1, n)i. A computation shows that
kHkS = n + 1,
but det(H) = 1.
The following lemma will be useful to estimate the norm of the kernel of group morphisms.
Lemma 4.5. Let S be a generating subset for Zk . There exists a constant C > 0 such that for every
injective group morphism ϕ : Zk → Zk and every element x ∈ Zk , it holds that
kxkS ≤ C
(max{kϕ kS , kϕ (x)kS })k
.
| det(ϕ )|
Proof. The statement is invariant under changing the generating subset, so we assume that S is the
standard generating subset. Writing x = (xi )ki=1 , Cramers rule implies
xi =
det(ϕi )
det(ϕ )
Effective Separability of Finitely Generated Nilpotent Groups
14
where ϕi is the morphism given by replacing the i-th column of the matrix representative of ϕ by
ϕ (x). Moreover, each entry of the matrix representative of ϕi is bounded by max{kϕ kS , kϕ (x)kS }.
Thus, the explicit formula for the determinant of a (k × k)−matrix gives our result.
Lemma 4.6. Let A be a finitely generated abelian group of rank k > 0 with a finite generating
subset S′ and m a fixed natural number. There exists a constant C > 0 such that for every group
morphism ϕ : Zℓ → A and generating subset S with |S| ≤ m for Zℓ it holds that
k ker(ϕ )kS ≤ C(kϕ kS,S′ )k .
Moreover, there exists m′ , not depending on ϕ , such that this bound can be achieved with a generating subset with ≤ m′ elements.
Proof. Without loss of generality, we may assume that A is torsion-free, so A = Zk . The statement
is invariant under changing the generating subset of A, so we may assume that S′ is the standard
generating subset. Additionally, we assume that ϕ (Zℓ ) has rank k by taking a quotient of A if
necessary.
First, we prove the lemma for the standard generating subset S = {e1 , . . . , el } for Zl . We assume
that ϕ (ei ) are linearly independent for 1 ≤ i ≤ k, and let D = det(ϕ (e1 ), ϕ (e2 ), . . . , ϕ (ek )). Note
that there exists a constant C0 such that |D| ≤ C0 kϕ kkS,S′ .
For each 1 ≤ j ≤ l − k, we construct a vector v j ∈ ker(ϕ ). Apply Lemma 4.5 gives a x j ∈ Zk such
that ϕ (x j ) = D ϕ (ek+ j ) and |x j |S ≤ C1 (kϕ kS,S′ )k . Letting ι : Zk → Zl be the natural inclusion,
the vector v j = ι (x j ) − D ek+ j is in ker(ϕ ) by construction. The vectors v j generate a finite index
subgroup of ker(ϕ ). Since S is assumed to be the standard generating subset, an easy computation
shows that k ker(ϕ )kS ≤ m max{kv j k}. Therefore, k ker(ϕ )kS ≤ C(kϕ kS,S′ )k and the bound m′
also follows from this construction.
Now assume that S = {s1 , . . . , sm } is an arbitrary generating subset for Zℓ . Consider the morphism
f : Zm → Zℓ which maps ei to the generator si . For the composition ϕ ◦ f , we have
k ker(ϕ ◦ f )kS ≤ C(kϕ kS,S′ )k .
Now f maps the generators of this kernel to generators of the kernel of ϕ , and moreover, k f (x)kS ≤
kxk for every x ∈ Zm . Thus, the lemma follows.
Example 4.7. Fix n ∈ N and take n distinct primes p1 < . . . < pn . Consider the matrix given by
p1j
0
Aj = .
..
0
0 ...
p2j . . .
.. . .
.
.
0 ...
0
0
..
.
1
1
..
.
pnj 1
and the corresponding linear map ϕ j : Zn+1 → Zn . Let S and S′ be the standard generating sets
for Zn+1 and Zn respectively. If we write k = p1 . . . pn , then the kernel of ϕ j is generated by the
Effective Separability of Finitely Generated Nilpotent Groups
element
j
kj
, . . . , k j , −k j
p1j
pn
15
. A computation shows that
k ker(ϕ j )kS = k
kj
p1j
,...,
kj
pnj
!
, −k j kS ≥ k j ≥ pnjn .
j
On the other hand, kϕ j kS′ = max{pn , n}. This example implies that the bound in Proposition 4.6
is optimal.
Proposition 4.8. For every m ≥ 0, there exists a constant C such that for every nilpotent group N
and any generating subset S with |S| ≤ m, it holds that
p
k N γ2 (N)kS ≤ C
Proof. If s1 , . . . , sk are the generators of N, then consider the group morphism ϕ : Zk → N p
N
γ2 (N)
p
N
with ϕ (ei ) = si γ2 (N). From Lemma 4.6, it follows that there exists a C such that k ker(ϕ )k ≤ C.
Forpevery generator x j ∈pZk for ker(ϕ ), fix an element y j ∈ N such that ky j kS ≤ kx j k and ϕ (x j ) =
y j N γ2 (N). The group N γ2 (N) is now generated by the y j and γ2 (N), where kγ2 (N)kS ≤ 4.
Proposition 4.9. Let A be a finitely generated abelian group of rank k > 0 with a finite generating
subset S′ . For every m > 0, there exists a constant C > 0 such that if N is a F -group with a finite
generating subset S with |S| ≤ m and ϕ : N → A is a group morphism, then
k ker(ϕ )kS ≤ C(kϕ kS,S′ )rankZ (A) .
Moreover, there exists m′ such that this bound can be achieved with a generating subset with ≤ m′
elements.
p
Proof. We can assume that A is torsion-free. First note that N γ2 (N) is a subgroup of ker ϕ whose
norm is bounded as described in Proposition 4.8. So it suffices to bound the norm of the kernel of
ϕ̄ : N p
N
γ (N) → A, and thus, Lemma 4.6 gives our proposition. The statement about the number
2
of generators also follows from the same lemma.
Theorem 4.10. Let N be a F -group with a finite generating subset S, and let i > 0. Then there
ϕ
exists a natural number k and constant C > 0 such that kNi kS ≤ C (kϕ kS )k for any ϕ ∈ Aut(N).
ϕ
Proof. We proceed by induction on i that every Ni satisfies the condition of the theorem, with
the additional assumption that the number of elements in such a generating subset is uniformly
ϕ
bounded by a constant mi . For i = 1, we have that γ1 (N) = N, and thus, N1 = N. Our theorem is
now evident for this case with m1 = |S| and k = 1.
ϕ
Now assume that the result holds for the group Ni with the constant C1 , integer k1 and number
ϕ
of generators mi . By assumption, there exists a finite generating subset S′ for Ni with |s|S ≤
C1 (kϕ kS )k1 for all s ∈ S′ and |S′ | ≤ mi .
Effective Separability of Finitely Generated Nilpotent Groups
16
For the quotient γi (N) γ
, we fix the generating subset Si given by the projections of commutators of length i. Denote ri to be the rank of γi (N) γ (N) and fix the constant C2 as in
i+1
Proposition 4.9 for the nilpotency class c with the generating subset Si , which is independent of
the group morphism ϕ .
i+1 (N)
From the computation after Definition 2.3, it follows that kϕ kS′ ,Si ≤ C1 (kϕ kS )k1 kϕ kS,Si . Take C3
as any constant such that kskS ≤ C3 for all s ∈ Si . (It is easy to give an explicit form for the constant
C3 .) Hence, it follows that kϕ kS,Si ≤ C3 kϕ kS , and consequently, kϕ kS′ ,Si ≤ C1 C3 (kϕ kS )k1 +1 .
ϕ
ϕ
From the computation after Definition 2.1, we have kNi+1 kS ≤ C1 (kϕ kS )k1 kNi+1 kS′ . By using
Proposition 4.9, we get that
ϕ
ϕ
kNi+1 kS ≤ C1 (kϕ kS )k1 kNi+1 kS′
≤ C1 C2 (kϕ kS )k1 kϕ kS′ ,Si
ri
≤ C1ri +1 C2 C3ri (kϕ kS )k1 +k1 ri +ri .
It follows that there exists a bound mi+1 on the number of generators by Proposition 4.9.
An important application of Theorem 4.10 is to bound the twisted determinant of an automorphism
in terms of its norm.
Corollary 4.11. Let N be a F -group with generating subset S. There exists a constant C > 0 and
k ∈ N such that for every automorphism ϕ : N → N the twisted determinant Dϕ satisfies
|Dϕ | ≤ C (kϕ kS )k .
Proof. It suffices to give such a bound for every determinant Di of ψi . For this bound, we use
ϕ
Theorem 4.10 on the group Ni to find a generating subset S′ whose word length kS′ kS ≤ C(kϕ kS )k
ϕ
for some C, k ∈ N. In particular, the norm kψi (Ni )kS ≤ C(kϕ kS )k+1 . We next apply Corollary 4.3
to the group morphism ψi to find a bound on the determinant Di .
As a final application of the above bounds, we have the following estimate which is essential for
Theorem 6.3.
Corollary 4.12. Let N be a F -group with a finite generating subset S. Let p be prime, and let
k∗ (p, c) be the constant from Corollary 3.11. Then there exists some constant C > 0 and an integer
∗
k such that pk (p,c)+v p (Dϕ ) ≤ C (kϕ kS )k for every automorphism ϕ : N → N.
Proof. This follows immediately from the bounds in Proposition 3.8 and Corollary 4.11.
5
Precise Conjugacy Separability of Two Step Nilpotent Groups
Before we give a proof of the main theorem, we apply the techniques that we have developed
so far in a specific setting, namely conjugacy separability of nilpotent groups of nilpotency class
Effective Separability of Finitely Generated Nilpotent Groups
17
2. The main goal of this section is to give a precise computation of the asymptotic behavior of
ConjN,S (n) where N is a F -group (see Theorem 5.4). We start with some preliminary results and
observations.
Let N be a F -group of nilpotency class 2 with x ∈ N. From Definition 3.1, we have
Inn(x)
N2
= {y ∈ N| y x y−1 x−1 ∈ γ2 (N)} = N
since N is a nilpotent group of nilpotency class 2. We also observe that the map
ψInn(x),2 : N → γ2 (N)
is given by ψInn(x),2 (y) = [y, x]. Note that for every x, y, z ∈ N, it holds that
[z, xy] = zxyz−1 y−1 x−1 = zxz−1 x−1 xzyz−1 y−1 x−1 = [z, x][z, y].
Inn(x)
This calculation can also be observed from the fact that N3
= 1 for all x ∈ N.
In the case of inner automorphisms, we have a stronger version of Theorem 4.10.
Proposition 5.1. Let N be a F -group with a finite generating
subset S such that c(N) = 2. Then
p
there exists a C > O such that k Im(ψInn(x),2 )kS ≤ C kxkS for every x ∈ N.
Proof. Write S′ = {[si , s j ] | si , s j ∈ S} as a finite generating subset for γ2 (N). From [33], it follows
that there exists a constant C > 0 such that
p
kxkS ≤ C kxkS′
for all x ∈ γ2 (N). The subgroup Im(ψInn(x),2 ) is generated by elements of the form [s, x] with s ∈ S.
p
It follows immediately that k[s, x]kS′ ≤ kxkS , and hence, k Im(ψInn(x),2 )kS ≤ C kxkS .
Another crucial ingredient for the main results is separability of central subgroups. We start with
an easy example which we will use in the proof of the next proposition.
Example 5.2. Take the group Z with standard generating subset S = {±1}. Consider any nontrivial subgroup H = hhi 6= 0. We can seperate any element x ∈ Z, x ∈
/ H in the finite quotient Z H .
The order of this finite quotient is |h| = kHkS . Note that Z H is a direct sum of subgroups with
order a prime power. In particular we can separate every element x ∈
/ H from H in a quotient of
the form Z pk Z with pk ≤ kHkS .
For a general generating subset S′ for Z, we find that there exists a C > 0 such that every element
can be separated in a finite quotient Z pk Z with pk ≤ CkHkS′ .
The following result generalizes this example to nilpotent groups.
Proposition 5.3. Let N be a F -group of nilpotency class c with finite generating subset S. There
exists a constant C such that for all central subgroups H ≤ Z(N) and every x ∈ N, x ∈
/ H, we can
separate x and H in a finite quotient Q with
|Q| ≤ C max{kHkS , log (kxkS )}c Φ(N) .
Effective Separability of Finitely Generated Nilpotent Groups
18
Proof. Assume that H is a central subgroup and x ∈ N \ H. We construct a finite√quotient which
separates x from H in two different cases, according to whether or not x ∈ H0 = N H.
First assume that x ∈ H0 . By taking a quotient M of N if necessary, we can assume that Z(N) has
rank 1 and that this quotient satisfies h(M) ≤ Φ(N). Let z be a generator of Z(M). From Example
5.2, it follows that there exists a prime power pk such that x and H are separated in Z(M) pk Z(M)
with pk ≤ CkHkS′ for S′ a generating subset for Z(M). There exists a constant C′ such that kHkS′ ≤
C′ kHkcS by [33]. Now take k(p, c) as in Lemma 3.10, and consider the quotient M pk+k(p,c) . Note
M
k+k(p,c)
p
k
that M
∩ Z(M) ≤ p · Z(M) by Lemma 3.10, and thus, x and H are seperated in this finite
quotient. The order of this quotient is bounded by ph(M)k(p,c) ph(M)k ≤ (c!)Φ(N) CkHkc Φ(N) .
Now assume that x ∈
/ H0 . In this case, consider the quotient group N H0 which is a torsion-free
nilpotent group with πH0 (x) 6= 1. From [36, Thm 1.1], where ΨRF (N) = Φ(N), it follows there
exists a constant C such that x is separated from H in a quotient of order ≤ C log (kxkS )c Φ(N) .
Since we have a bound in both cases, the proof is finished.
Now that we have all of the necessary tools, we can give the precise calculation of the conjugacy
separability of F -groups of nilpotency class 2.
Theorem 5.4. Let N be a F -group with a finite generating subset S such that c(N) = 2. Then
ConjN,S (n) ≈ nΦ(N) .
Proof. The lower bound is given in [36, Theorem 1.8(ii)], where ΨLower (N) = Φ(N) in this case.
Thus, we need only to demonstrate that ConjN,S (n) nΦ(N) . To that end, suppose x, y ∈ N such
that kxkS , kykS ≤ n and x ≁id y.
If πab (x) ≁id πab (y), or equivalently, if πab (x) 6= πab (y), then [36, Thm 1.1.] implies that there
exists a surjective group morphism π : N/[N, N] → Q such that π (πab (x y−1 )) 6= 1 and where
|Q| ≤ C log(kx y−1 k). Since π (πab (x)) and π (πab (y)) are non-equal central elements, they are not
conjugate. Thus, DN ([x]id , y) ≤ 2C log(2n).
Now we may assume that y = x z where kzkS ≤ n. Consider the group morphism ψInn(x),2 as before.
√
Proposition 5.1 implies k Im(ψInn(x),2 )kS ≤ C1 n for some C1 ∈ N. Proposition 5.3 implies there
exists a surjective group morphism π : N → Q such that π (z) ∈
/ π (Im(ψInn(x),2 )) and where |Q| ≤
2Φ(N)
C2 (k Im(ψInn(x),2 )kS )2 Φ(N) ≤ C1
C2 nΦ(N) . We claim that π (y) ≁id π (x z). Suppose for a contradiction there exists an a ∈ N such that π (a x a−1 ) = π (x z). That implies π (a x a−1 x−1 ) = π (z).
2Φ(N)
C2 nΦ(N) ,
Thus, π (z) ∈ π (Im(ψInn(x),2 )) which is a contradiction. Hence, DN ([x]id , x z) ≤ C1
and subsequently, ConjN,S (n) nΦ(N) .
6
Effective Twisted Conjugacy Separability
In this section we prove the main results of this paper. We start with the abelian case.
Effective Separability of Finitely Generated Nilpotent Groups
19
Proposition 6.1. Let A be a finitely generated abelian group with a finite generating subset S. For
every x ∈ A, it holds that
ϕ
ConjA,x,S (n) kϕ kS log (kxkS + n)
ϕ
for all ϕ ∈ Aut(A). Subsequently, ConjA,S (n) kϕ kS log(n) and TconjA,S (n) n log(n).
Proof. Let ϕ ∈ Aut(A), and take x ∈ A. Suppose y ∈ A satisfies kykS ≤ n and x ≁ϕ y. We may
write
[x]ϕ = {z + x − ϕ (z) | z ∈ A} = {x + (IA − ϕ )(z) | z ∈ A}
/ Im(IA − ϕ ).
where IA : A → A is the identity map. It thus holds that y ∈
/ [x]ϕ if and only if y − x ∈
Observe that Im(ψϕ ,1 ) = Im(IA − ϕ ) if we use the notations from Section 3. For s ∈ S, we observe
that {(IA − ϕ )(s) | s ∈ S} is a generating subset for Im(IA − ϕ ). Thus, for each s ∈ S we may write
k(IA − ϕ )(s)kS ≤ kskS + kϕ (s)kS ≤ 1 + kϕ kS .
Subsequently, k Im(IA − ϕ )kS ≤ 2kϕ kS . Proposition 5.3 implies that there exists a surjective group
/ π (Im(I − ϕ )) and where
morphism π : A → Q such that π (y − x) ∈
|Q| ≤ C max {log (ky − xkS ) , kϕ kS } ≤ C max {log (kxkS + n) , kϕ kS }
for some C ∈ N. By construction, it holds that π (y) ∈
/ π ([x]ϕ ). The statements of the theorem then
follow immediately.
We still need the following technical result, which is a generalization of Lemma 4.3.
Lemma 6.2. Let N be a F -group with finite generating subset S and an automorphism ϕ : N → N.
ϕ
There exists a constant C > 0 and an integer k > 0 such that for every y ∈ N with y ∈ X1 , it holds
that y = x ϕ (x)−1 for some x ∈ N with
kxk ≤ C max{kϕ kS , kykS }k .
ϕ
ϕ
Proof. We proceed by induction on j such that y ∈ Xc+1− j . If j = 0 or thus y ∈ Xc+1 = 1, we
can take x = 1. Now assume that j > 0 and write c + 1 − j = i. Consider the induced map
ϕ
ψ̄ϕ ,i : Ni N ϕ → γi (N) γ (N). Use Theorem 4.10 to find constants C1 , k1 > 0 and a generating
i+1
i+1
ϕ
subset S′ for Ni such that kskS ≤ C1 kϕ kkS1 for all s ∈ S′ . We apply Lemma 4.5 to find C2 , k2 > 0
ϕ
such that there exists x ∈ Ni with
kxkS′ ≤ C2 max{kϕ kS , kykS }k2
ϕ
and ψ̄ϕ ,i (x Ni+1 ) = x ϕ (x)−1 γi+1 (N) = y γi+1 (N). Write z = x ϕ (x)−1 . In particular, we get that
kxkS ≤ C1C2 kϕ kkS1 max{kϕ kS , kykS }k2 ≤ C1C2 max{kϕ kS , kykS }k1 +k2 .
ϕ
By construction, z−1 y ∈ Xi+1 with
kz−1 ykS ≤ kykS + kxkS + kϕ (x)kS
≤ kykS +C1C2 max{kϕ kS , kykS }k1 +k2 +C1C2 max{kϕ kS , kykS }k1 +k2 +1 ,
and so, we can use the inductive hypothesis to finish the proof.
Effective Separability of Finitely Generated Nilpotent Groups
20
For the rest of this section, we fix an automorphism ϕ : N → N and work with the automorphisms
ϕx : N → N given by ϕx (n) = x ϕ (n) x−1 . Additionally, we denote the induced automorphism on
N m where m ∈ N as ϕ̄x,m The subgroup Nϕ will be denoted as Nx to simplify notation. Similarly,
x
N
N
we denote N m
via Nx,m .
ϕ̄x,m
Theorem 6.3. Let N be a F -group with a finite generating subset S. Let x ∈ N and ϕ ∈ Aut(N).
There exist natural numbers k1 , k2 , and k3 such that
ϕ
ConjN,x,S (n) (kϕ kS )k1 (kxkS )k2 nk3
ϕ
for any ϕ ∈ Aut(N). In particular, ConjN,S (n) (kϕ kS )k1 nk2 +k3 and TconjN,S (n) nk1 +k2 +k3 .
Proof. We proceed by induction on nilpotency class, and observe that the base case is given by
Proposition 6.1. Thus, we may assume that c(N) > 1.
Let ϕ ∈ Aut(N) and fix x ∈ N. Suppose that y ∈ N satisfies x ≁ϕ y and kykS ≤ n. Our goal is to
construct a finite quotient N/K such that πK (y) ∈
/ πK ([x]ϕ̃ ) and then bound |N/K| in terms of kxkS ,
kϕ kS , and n.
Denote by ϕ̄ the automorphism of N γ (N) induced by ϕ . If x γc(N) (N) ≁ϕ̄ y γc(N) (N), then
c(N)
the inductive hypothesis implies there exist integers k1 , k2 , k3 and C1 > 0 satisfying the following. There exists a surjective group morphism to a finite group π1 : N γ (N) → Q1 such that
c(N)
/ [π1 (x γc(N) (N))]ϕ̄ and where
π1 (y γc(N) (N)) ∈
|Q1 | ≤ C1 (kϕ̄ kS̄ )k1 (kx̄kS )k2 nk3 .
Thus, we can assume that x γc(N) (N) ≁ϕ̄ y γc(N) (N). Writing y = yx−1 x, Lemma 3.2 implies that
ϕ
yx−1 ∈ X1 x γc (N). By Lemma 6.2, there exists y0 with the norm bounded as described in the lemma
and yx−1 = y0 ϕx (y0 )−1 z for some z ∈ γc (N). Solving for z, we get that there exist constants C2 , k4
such that
kzkS ≤ C2 (kϕx kS kxkS n)k4 .
Now we see that y−1
0 yϕ (y0 ) = xz or thus y ∼ϕ xz.
Theorem 4.10 implies that kNx kS ≤ C3 (kϕx kS )k5 where C3 > 0 is some constant and k5 ∈ N. For
each s ∈ S, we may write kϕx (s)kS = kx ϕ (s) x−1 kS ≤ 2 kxkS + kϕ (s)kS . Subsequently, kϕx kS ≤
2 kxkS kϕ kS , and thus, kNx kS ≤ C4 (kϕ kS )k5 (kxkS )k5 where C4 = C3 2k5 .
Now fix k6 ∈ N and constant C5 > 0 from Proposition 5.3 for N. Lemma 3.2 implies that z ∈
/ Nx .
/ π2 (Nx ) and
Therefore, there exists a surjective group morphism π2 : N → Q2 such that π2 (z) ∈
where
|Q2 | ≤ C5 (max{kNx kS , log (kzkS )})k6 ≤ C5 (kNx kS · kzkS )k6
(k +k5 )k6
≤ 2k4 (C2C4 )k6 C5 kϕ kS 4
(2k4 +k5 )k6 k4 k6
kxkS
n
.
α
Moreover, we may assume that |Q2 | = pα where p is a prime. Lemma 2.9 implies that N p ≤
/ π pα (Nx ).
ker(π2 ), and subsequently, π pα (z) ∈
Effective Separability of Finitely Generated Nilpotent Groups
21
Using the natural number k∗ (p, c(N)−1) and notation from Corollary 3.11, it follows that ρ1 (Nx,m ) =
∗
/ πm ([x]ϕ ). Indeed otherwise,
πm (Nx ) where m = pα +v p (Dϕx )+k (p,c−1) . We claim that πm (x z) ∈
πm (z) ∈ Nx,m . Thus, π pα (z) = ρ1 (πm (z)) ∈ π pα (Nx ) which is a contradiction.
For the bound on the order, we combine the previous inequalities. Corollary 4.12 implies there
exists a constant C6 > 0 and an integer k7 such that pk0 +v p (Dϕ ) ≤ C6 (kϕ kS )k7 . Therefore,
h(N)
|N/N m | ≤ (pα +k0 +v p (Dϕ ) )h(N) = pα h(N) C6 (kϕ kS )k7
,
and thus, the first statement holds because of the bounds on pα . The last two statements of the
theorem follow immediately.
7
Virtually Nilpotent Groups
This section is broken up into two parts. The first part is a technical detour, whereas the second
subsection contains the main results of the section.
7.1
Separability of Extensions of Twisted Conjugacy Separable Groupss
We start this subsection with the following definition.
Definition 7.1. Let X ⊂ G be a subset (not necessarily finite). Let FG,X,S : N → N be defined as
FG,X,S (n) = max{DG (X , g) | g ∈ (G \ X ) ∩ BG,S (n)},
where we take max 0/ = 0.
Recall that the function DG (X , ·) was introduced on page 6. The function FG,X,S measures the
complexity to separate elements of G from the set X in finite quotients. The function does not
depend on the choice of generating subset.
Lemma 7.2. Let G be a finitely generated group, and let X ⊂ G be a subset. Let S1 and S2 is a
two finite generating subsets for G, then FG,X,S1 (n) ≈ FG,X,S2 (n).
The proof is similar to [5, Lem 1.1] (see also [6, Lem 1.2] and [30, Lem 2.1]). We first give some
lemmas about the function FG,X,S before coming to our main result.
Let H be a finite index normal subgroup of G, and let X ⊂ H be a separable subset. For any x ∈
H \ X , the following lemma relates the complexity of separating x from X in G to the complexity
of separating x from X in H.
Lemma 7.3. Let G be a finitely generated group, and let H be a finite index normal subgroup.
Suppose that S1 and S2 are finite generating subsets for G and H respectively. Suppose that X ⊆ H
is a separable subset. Then FG,X,S1 (n) (FH,X,S2 (n))[G:H] .
Effective Separability of Finitely Generated Nilpotent Groups
22
Proof. Since H is a finite index subgroup, H is an undistorted subgroup. Subsequently, there
exists C > 0 such that kxkS2 ≤ C kxkS1 for all x ∈ H. Let x ∈ G such that x ∈ (G \ X ) ∩ BG,S1 (n).
Suppose that x is not an element of H. Since X ⊆ H, we may pass to the quotient G H which is a
finite group by assumption. Thus, we may assume that x ∈ H \ X .
/
Note that kxkS2 ≤ Cn. Hence, there exists a surjective group morphism π : H → Q such that π (x) ∈
π (X ) where |Q| ≤ FH,X,S2 (Cn). Since finite groups are linear, we have a finite group Q̃ ⊃ Q and an
induced morphism π̃ : G → Q̃ such that π̃ restricted to H is equal to our original group morphism
π , and moreover, |Q̃| ≤ |Q||G:H| . Thus, |Q̃| ≤ FH,X,S2 (Cn))|G:H| , and subsequently, FG,X,S1 (n)
(FH,X,S2 (n))|G:H| .
For separable subsets {Xi }ki=1 and x ∈ G \ ∪ki=1 Xi, this next lemma relate the complexity of separating x from ∪ki=1 Xi to the complexity of separating x from each Xi individually.
Lemma 7.4. Let G be a finitely generated group with a finite generating subset S, and let {Xi }ki=1
be a finite collection of proper separable subsets. If Y = ∪ki=1 Xi , then
k
FG,Y,S (n) ∏ FG,Xi ,S (n).
i=1
Proof. Let x ∈ G such that x ∈ (G \Y ) ∩ BG,S (n). It follows that x ∈ (G \ Xi) ∩ BG,S (n). Thus,
there exists a surjective group morphism πi : G → Qi such that π (x) ∈
/ π (Xi ) and |Q| ≤ FG,Xi ,S (n).
k
Let K = ∩i=1 ker(πi ). By selection, πK (x) ∈
/ πK (Xi ) for each i, and hence, πK (x) ∈
/ πK (Y ). Now
|G K | ≤ ∏ki=1 |Qi | ≤ ∏ki=1 FG,Xi ,S (n). We conclude that FG,Y,S (n) ∏ki=1 FG,Xi ,S (n).
This last lemma relates the complexity of separating an element x ∈ G/X from X to the complexity
of separating x y from X · y.
Lemma 7.5. Let G be a finitely generated subgroup with a finite generating subset S. Suppose
that X ⊂ G is a separable subset, and let y ∈ G. Then FG,X·y,S (n) ≈ FG,X,S (n).
Proof. We need only show that FG,X,S (n) FG,X·y,S (n). Suppose that x ∈ G such that kxkS ≤ n
and x ∈
/ X · y. That implies that x y−1 ∈
/ X . Therefore, there exists a surjective group morphism
−1
π : G → Q such that π (x y ) ∈
/ π (X ) and |Q| ≤ FG,X,S (kykS + n) ≤ FG,X,S (kykS · n). It follows that
π (x) ∈
/ π (X · y) since right translation is a bijection of G. Therefore FG,X·y,S (n) FG,X,S (n).
Suppose G is a finite extension of H. The following proposition shows that the conjugacy class of
any element of G can be written as a finite union of right translates of twisted conjugacy classes
of elements in H. The following proof follows [16, Thm 5.2].
Proposition 7.6. Suppose that G is a finite generated group that contains a finite index charac|G:H|
teristic subgroup H, and let {si }i=1 be a set of representatives for the right cosets of H. Fix
Effective Separability of Finitely Generated Nilpotent Groups
23
an automorphism ϕ ∈ Aut(G) and let fx,i be the automorphism of H induced by conjugation by
xi = si x ϕ (si )−1 . If ϕ̄ is the automorphism of H induced by ϕ , then
[G:H]
[
[x]ϕ =
i=1
[1H ] fx,i ◦ϕ̄ · xi .
Proof. Let ϕ̄ be the automorphism of H induced by ϕ . We may write
[G:H]
−1
[x]ϕ = {y x ϕ (y)
| y ∈ G} =
[
i=1
[G:H]
=
{z si x ϕ (si )−1 ϕ (z)−1 | z ∈ H}
[G:H]
[
{z xi ϕ (z)−1 x−1
i xi | z ∈ H} =
[
{z fx,i (ϕ (z))−1 | z ∈ H} · xi =
i=1
[G:H]
=
i=1
[
{z fx,i (ϕ (z))−1 xi | z ∈ H}
[
[1H ] fx,i ◦ϕ̄ · xi .
i=1
[G:H]
i=1
Suppose that H is a twisted conjugacy separable group and that G is a finite extension. Additionally, assume that ϕ ∈ Aut(G) and x ∈ G. The following theorem relates the quantification of the
ϕ -twisted conjugacy class of x in G with the quantification of ψi -twisted conjugacy separability of
H where ψi are a finite fixed collection of automorphisms of H, depending both on ϕ and x.
Theorem 7.7. Suppose that G is a finite generated group that contains a finite index characteristic
|G:H|
subgroup H, and let {si }i=1 be a set of representatives for the right cosets of H. Fix an automorphism ϕ ∈ Aut(G) and a x ∈ G. Let fx,i be the automorphism of H induced by conjugation by
xi = si x ϕ (si )−1 . If SG and SH are finite generating subsets of G and H respectively, then
ϕ
ConjG,x,SG (n)
[G:H]
∏
i=1
[G:H]
fx,i ◦ϕ
ConjH,1,S
(n)
.
H
Proof. For simplicity in the following arguments, let Xi = [1H ] fx,i ◦ϕ · xi and Xi′ = [1H ] fx,i ◦ϕ .
Lemma 7.5 implies that FG,Xi ,SG (n) ≈ FG,Xi′ ,SG (n), and since Xi′ ⊆ H, Lemma 7.3 implies that
f ◦ϕ
x,i
(n). Given
FG,Xi′ ,SG (n) (FH,Xi′ ,SH (n))[G:H] . Since Xi′ = [1H ] fxi ◦ϕ , we have FH,Xi′ ,SH (n) = ConjH,1,S
H
[G:H]
that [x]ϕ = ∪i=1 Xi , Lemma 7.4 implies that
[G:H]
ϕ
ConjG,x,SG (n)
∏ FH,X ,S
i=1
′
i
H
(n).
Taking everything together, we have
ϕ
ConjG,x,SG (n)
[G:H]
∏
i=1
[G:H]
fx,i ◦ϕ
ConjH,1,S
(n)
.
H
Effective Separability of Finitely Generated Nilpotent Groups
7.2
24
Effective Twisted Conjugacy Separability of Virtually Nilpotent Groups
We now apply Theorem 7.7 to the context of virtually nilpotent groups to get the first of the two
main results of this section.
Theorem 7.8. Suppose that G is a virtually nilpotent group, and suppose S is a finite generating
subset of G. For ϕ ∈ Aut(G) and x ∈ G, there exists a natural numbers k1 , k2 , k3 such that
ϕ
ConjG,x,S (n) (kϕ kS )k1 (kxkS )k2 nk3 .
ϕ
In particular, ConjG,S (n) (kϕ kS )k1 nk2 +k3 and TconjG,S (n) nk1 +k2 +k3 .
Proof. If G is finite, then the theorem is clear. Thus, we may assume that G is infinite. We can
assume that N is a characteristic F -subgroup of G. Let S′ be a finite generating subset for N and
|G:N|
{si }i=1 be a set of right coset representatives of N in G.
Consider the automorphism fx,i of N induced by conjugation by si x ϕ (si )−1 . Theorem 7.7 implies
ϕ
ConjG,x,S (n)
[G:N]
∏
i=1
[G:N]
fx,i ◦ϕ̄
ConjN,1,S
.
′ (n)
Theorem 6.3 implies there exist natural numbers k1 , k2 , k3 such that
f ◦ϕ
x,i
(n) (k fx,i ◦ ϕ̄ kS′ )k1 (k1kS′ )k2 nk3 = (k fx,i ◦ ϕ̄ kS′ )k1 nk3 .
ConjH,1,S
H
Thus, to finish, we give a bound for k fx,i ◦ ϕ̄ kS′ .
Since the subgroup N is undistorted, it suffices to find a bound on k fx,i ◦ ϕ kS ≤ k fx,i kS · kϕ kS . Note
that fx,i is conjugation by si x ϕ (si )−1 , and thus,
k fx,i kS ≤ 2ksi x ϕ (si )−1 kS + 1 ≤ 2(ksi kS )2 kxkS kϕ kS + 1 ≤ CkxkS kϕ kS
for some C ∈ N since there are only finitely many si .
That implies k fx,i ◦ ϕ̄ k ≤ C kxkS (kϕ kS )2 . Therefore,
f ◦ϕ
x,i
(n) Ck1 (kxkS )k1 (kϕ kS )2 k1 nk3 .
ConjH,1,S
H
By taking everything together, we may write
ϕ
2
2
2
2
ConjG,x,S (n) ≤ Ck1 [G:N] (kxkS )k1 |G:N| (kϕ kS )2 k1 |G:N| nk3 |G:N| .
The last two inequalities follow immediately.
For this section’s last result, we need the following proposition which is similar to [36, Cor 10.5]).
Proposition 7.9. Let H3 (Z) be the 3-dimensional integral Heisenberg group with presentation
given by hx, y, z| [x, y] = z, z centrali, and let p be any prime. Suppose π : H3 (Z) → Q is a surjective
group morphism where Q is a q-group for some prime q distinct from p, then π (x p ) ∼id π (x p zm )
where m is any natural number.
Effective Separability of Finitely Generated Nilpotent Groups
25
Proof. Write the conjugacy class [x p ] = {x p zt p | t ∈ Z}. Let qk be the order of the element π (z) in
Q. Since gcd(p, qk ) = 1, there exists integers a, b such that a p + b qk = 1. We see that
k
π (x p ) ∼id π (x p zm a p ) = π (x p zm a p+m b q ) = π (x p zm ).
We reproduce the proof of [36, Prop 13.1].
Proposition 7.10. Let G and H be conjugacy separable, finitely generated groups, and suppose
that H is a subgroup of G. If x, y ∈ H are two non-conjugate elements of G, then DN ([x]id , y) ≤
DG ([x]id , y).
Proof. Take π : G → Q be surjective morphism with |Q| = DG ([x]id , y) and π (y) ∈
/ π ([x]id ). The
restriction of π to the subgroup H separates the conjugacy classes of x and y. Moreover, |π (H)| ≤
|Q| = DG ([x]id , y).
We now have the following theorem which gives the asymptotic behavior for the conjugacy separability quantification function for the class of virtually nilpotent groups.
Theorem 7.11. Let G be a virtually nilpotent group with a finite generating subset S, and let x ∈ G.
There exist k1 , k2 ∈ N such that ConjG,x,S (n) (kxkS )k1 nk2 . In particular, ConjG,S (n) nk1 +k2 . If
G is not virtually abelian, then
n(c(N)−1) Φ(N) ConjG,S (n) nk1 +k2
where N is any infinite finitely generated nilpotent subgroup of finite index in G.
Proof. Note that if G is finite, then the upper bound clearly holds. Therefore, we may assume that
G contains a F -group N as a characteristic finite index subgroup. Thus, Theorem 7.8 implies that
k1 k2
id
there exist integers k1 , k2 such that Conjid
G,x,S (n) (kxkS ) n . Since ConjG,x,S (n) = ConjG,x,S (n)
and Conjid
G,S (n) = ConjG,S (n), the first two statements are evident.
For the lower bound, we follow the proof of [36, Thm 1.10(ii)]. Assuming that G is not virtually abelian, there exists a F -group N that is a finite index characteristic subgroup of G where
c(N) ≥ 2. We construct an infinite sequence of non-conjugate elements at , bt in N that are also not
(c(N)−1) Φ(N)
where S′ is a finite
conjugate in G such that kat kS′ , kbt kS′ ≈ nt and DN ([at ]id , bt ) ≈ nt
generating subset of N. Proposition 7.10 implies that DN ([at ]id , bt ) ≤ DG ([at ]id , bt ). Since N is a
finite index subgroup of G, it follows that N is undistorted in G. In particular, kat kS , kbt kS ≈ nt .
At that point, we have the last statement of the theorem.
Take elements x ∈ γc−1 (N), y ∈ N and z ∈ γc (N) ≤ Z(N) such that [x, y] = zk1 and z is primitive.
By using the techniques from [36, Prop3.6.], we can assume that every one dimensional central
quotient N H associated to z satisfies h N H ≥ Φ(N). Via [36, Defn 4.9] and [36, Prop 4.10],
there exists k2 ∈ N such that for every prime q > k2 and every surjective morphism π : N → Q to
a finite q-group with π (z) 6= 1, it holds that |Q| ≥ qΦ(N) . Take k = k1 k2 . By replacing x by xk2 ,
Effective Separability of Finitely Generated Nilpotent Groups
26
we find a subgroup x, y, zk of G which is isomorphic to the 3-dimensional integral Heisenberg
group.
∞ be an enumeration of primes greater than max{k, [G : N]}, we consider the eleLetting {pt }t=1
ments αt,i = x pt zi k for 1 ≤ i ≤ [G : N] + 1. Since αt,i are pairwise non-equal central elements of
N
H · N pt , they are pairwise non-conjugate as elements of N. We claim that there exist i0 such that
|G:N|
αt,1 ≁id αt,i0 as elements of G. Letting {si }i=1 be a set of right coset representatives of N in G,
we may write the conjugacy class of any element x ∈ N as
[G:N]
[x]id = {gxg−1 | g ∈ G} =
[
i=1
[G:N]
{n si x si−1 n−1 | n ∈ N} =
[
i=1
{n xi n−1 | n ∈ N},
so as the union of [G : N] conjugacy classes of elements xi = si x s−1
i in N. Since the [G : N] + 1
elements αt,i all lie in different conjugacy classes of N, the claim follows. Now take at = αt,1 and
bt = αt,i0 .
1/(c(N)−1)
. We claim that DN ([at ]id , bt ) ≥
From [22, Lem 3.B], it follows that kat kS , kbt kS ≈ pt
Φ(N)
Φ(N)
pt
. So it suffices to show that for all surjective group morphisms π : N → Q where |Q| < pt
that π (at ) ∼id π (bt ). If π (zk ) = 1, then π (at ) ∼id π (bt ). Thus, we may assume that π (zk ) 6= 1.
By [24, Thm 2.7], we may assume that Q is a finite q-group where q is a prime. From our
Φ(N)
assumptions on p j and π , we know that if q = p j , then |Q| ≥ pt
. Hence, we can also assume that
q 6= p j . Since x, y, zk is isomorphic to the 3-dimensional integral Heisenberg group, Proposition
7.9 implies that there exists g ∈ x, y, zk such that π (g at g−1 ) = π (bt ). Therefore, π (at ) ∼id π (bt )
and as explained above this argument ends the proof.
8
Some Examples
In this last section, we work some explicit examples.
8.1
Heisenberg group
In this section, we work out the twisted conjugacy seperability function for the discrete Heisenberg
group H3 (Z) = {(x, y, z) | x, y, z ∈ Z} with group law given by
(x1 , y1 , z1 )(x2 , y2 , z2 ) = (x1 + x2 , y1 + y2 , z1 + z2 + x1 y2 ).
Fix the generating subset S = {(1, 0, 0), (0, 1, 0)}.
Any automorphism ϕ : H3 (Z) → H3 (Z) can be written as
ϕ (x, y, z) = (ax + by, cx + dy, ex + f y + Dz)
where D = ad − bc = ±1. We let
a b
A=
∈ GL(2, Z).
c d
Effective Separability of Finitely Generated Nilpotent Groups
27
Note that if D = −1, then (0, 0, 2) ∈ (H3 (Z))ϕX for every X ∈ H3 (Z). So in this case, k(H3 (Z))ϕX kS
is uniformly bounded over all X .
ϕ
We compute the function ConjH3 (Z),S (n) for a general automorphism ϕ . Let X = (x1 , y1 , z1 ) and
Y = (x2 , y2 , z2 ) be two elements in H3 (Z) with kX kS , kY kS ≤ n such that X ≁ϕ Y . If πγ2 (H3 (Z)) (X ) ≁ϕ
πγ2 (H3 (Z)) (Y ), then Proposition 6.1 (with the automorphism ϕ fixed here) implies that we can seperate the twisted conjugacy classes in a finite quotient of norm log(n). So we will always assume
ϕ
that πγ2 (H3 (Z)) (X ) ∼ϕ πγ2 (H3 (Z)) (Y ). In this way, it also follows that log(n) ConjH3 (Z),S (n).
ϕ
ϕ
ϕ
Note that (H3 (Z))2 = (H3 (Z))2 X for all X ∈ N. The group (H3 (Z))2 can have rank 1, 2 or 3,
corresponding to dimension 0, 1 or 2 for the eigenspace of A corresponding to eigenvalue 1.
ϕ
Case 1: rank((H3 (Z))2 ) = 1
ϕ
In this case, (H3 (Z))2 = γ2 (H3 (Z)), and thus, NϕX = 1 That implies that x1 = x2 and y1 = y2 . If
ϕ
D = −1, we get that ConjH3 (Z),S (n) is bounded (since there are only a finite number of twisted
ϕ
conjugacy classes). If D = 1, then ConjH3 (Z),S (n) (log(n))3 since it corresponds to separating
central elements in the group H3 (Z). By using the elements (0, 0, 1) and (0, 0, p) for increasing
ϕ
primes, it is easy to show that indeed ConjH3 (Z),S (n) ≈ (log(n))3 .
ϕ
Case 2: rank(N2 ) = 2
Note that the second eigenvalue of A must be −1 and thus D = −1. The norm k(H3 (Z))ϕX k
is uniformly bounded and hence there is a fixed quotient in which we can seperate the twisted
ϕ
conjugacy classes of X and Y . We conclude that in this case ConjH3 (Z),S (n) ≈ log(n).
ϕ
Case 3: rank((H3 (Z))2 ) = 3
ϕ
In this case, A = I2 and (H3 (Z))2 = H3 (Z). In this case, we can copy the proof of [36, Thm 1.6]
ϕ
3
to find that ConjH3 (Z),S (n) ≈ Conjid
H3 (Z),S (n) ≈ n .
ϕ
Note that in all three cases, we have ConjH3 (Z),S (n) n3 ≈ ConjH3 (Z),S (n).
8.2
A 5-dimensional Example
Let N be the nilpotent group given by
N = ha1 , a2 , a3 , b1 , b2 | [a1 , a2 ] = b1 , [a2 , a3 ] = b2 , and b1 , b2 central i
with generating set S = {a1 , a2 , a3 , b1 , b2 }. Let ϕ : N → N be the automorphism given by
a1 7→ a1
a2 7→ a2
ϕ : a3 7→ a1 a3
b1 7→ b1
b2 7→ b1 b2 .
Effective Separability of Finitely Generated Nilpotent Groups
28
Every element x ∈ N can be uniquely expressed as
β
β
x = a1α1 a2α2 a3α3 b1 1 b2 2
for some αi , β j ∈ Z.
ϕ
ϕ
Note that the subgroup N2 x = N2 for every x ∈ N. It follows immediately that
ϕ
N2 x = ha1 , a2 , b1 , b2 i ;
ϕ
hence, kN2 x kS = 1. These facts allow us to prove the following proposition.
Proposition 8.1. There exists a constant C ∈ N such that for every x ∈ N with kxkS ≤ n, it holds
√
that k Im ψϕx ,2 ) kS ≤ C n.
Proof. The subgroup Im ψϕx ,2 ) is generated by the elements a1 x ϕ (a1 )−1 x−1 , a2 x ϕ (a2 )−1 x−1 ,
b1 x ϕ (b1 )−1 x−1 = 1 and b2 x ϕ (b2 )−1 x−1 = b−1
1 . The statement now follows similarly as in Proposition 5.1.
ϕ
We now compute the asymptotic upper bound for ConjN,S (n).
Proposition 8.2. For the group N and ϕ : N → N as above, it holds
ϕ
ConjN,S (n) n3 .
Proof. Let x, y ∈ N such that kxkS , kykS ≤ n and where x ≁ϕ y. First assume that π[N,N] (x) ≁ϕ
π[N,N] (y). Proposition 6.1 (with the automorphism ϕ fixed here) implies there exists a constant
C1 > 0 such that
DN/[N,N] ([π[N,N] (x)]ϕ̄ , π[N,N] (y)) ≤ C1 log kπ[N,N] (x)kS + kπ[N,N] (y)kS .
/
In particular, there exists a surjective group morphism π1 : N/[N, N] → Q1 such that π1 (y) ∈
π1 ([x]ϕ ) and where
|Q1 | ≤ C1 kπ[N,N] (x)kS n.
Since kπ[N,N] (x)kS̄ ≤ n, it follows that DN ([x]ϕ , y) ≤ C1 log(2n).
Thus, we may assume that y = x z where z ∈ γ2 (N) = hb1 , b2 i and kzkS ≤ C2 n. Proposition 5.3
/ π2 (Im(ψϕx ,2 ))
implies that there exists a surjective group morphism π2 : N → Q2 such that π2 (z) ∈
and where
2 Φ(N)
|Q2 | ≤ C2 max k Im(ψϕx ,2 )kS , log (kzkS )
for some C2 ∈ N. We claim that π2 (x z) ∈
/ π2 ([x]ϕ ), and for a contradiction suppose otherwise.
Thus, there exists y ∈ N such that π2 (x z) = π2 (y x ϕ (y)−1 ). That implies we may write π2 (z) =
π2 (y x ϕ (y)−1 x). In particular, π2 (z) ∈ Im(ψϕx ,2 ) which is a contradiction.
By calculation, one can see that Φ(N) = 3. We conclude that for some C3 > 0, we have DN ([x]ϕ , y) ≤
ϕ
C3 max{n3 , log(n)6 } and thus ConjN,S n3 .
Effective Separability of Finitely Generated Nilpotent Groups
9
29
Future Questions
We list some of the open problems which remain after the main results of this paper.
In Theorem 5.4 we computed the precise conjugacy separability function for nilpotency class 2.
We conjecture that for a general nilpotent group, this function is of the following form.
Conjecture 1. Let N be a F -group of nilpotency class c > 1, then ConjN (n) ≈ nΦ(N)(c−1) .
The lower bound for this conjecture was already given in [36].
In Proposition 5.3, we computed an upper bound for the effective seperability function in the case
of central subgroups. There is not yet a description for the subgroup separability function for
general subgroups. That leads to the following question.
Question 1. Compute the effective subgroup seperability function for finitely generated nilpotent
groups.
ϕ
In Section 8 we gave several explicit examples of the function ConjN,S for nilpotent groups N. In
all these examples, an upper bound was given by ConjN,S , so in the case of the identity map. We
conjecture that that is always the case.
Conjecture 2. Let N be a nilpotent group with generating subset S. For every automorphism
ϕ ∈ Aut(N), it holds that
ϕ
ConjN,S (n) ConjN,S (n).
For all known examples, the conjugacy function only depends on the rational or even real Mal’cev
completion of the F -group N. It is an open question whether that is true in general.
Question 2. Let N1 and N2 be two (abstractly) commensurable finitely generated virtually nilpotent groups. Is it true that
ConjN1 (n) ≈ ConjN2 (n)?
References
[1] James Arthur and Laurent Clozel. Simple algebras, base change, and the advanced theory
of the trace formula, volume 120 of Annals of Mathematics Studies. Princeton University
Press, Princeton, NJ, 1989.
[2] Norman Blackburn. Conjugacy in nilpotent groups. Proc. Amer. Math. Soc., 16:143–148,
1965.
[3] K. Bou-Rabee and D. B. McReynolds. Bertrand’s postulate and subgroup growth. J. Algebra,
324(4):793–819, 2010.
[4] K. Bou-Rabee and D. B. McReynolds. Asymptotic growth and least common multiples in
groups. Bull. Lond. Math. Soc., 43(6):1059–1068, 2011.
Effective Separability of Finitely Generated Nilpotent Groups
30
[5] Khalid Bou-Rabee. Quantifying residual finiteness. J. Algebra, 323(3):729–737, 2010.
[6] Khalid Bou-Rabee. Approximating a group by its solvable quotients. New York J. Math.,
17:699–712, 2011.
[7] Khalid Bou-Rabee, Mark F. Hagen, and Priyam Patel. Residual finiteness growths of virtually special groups. Math. Z., 279(1-2):297–310, 2015.
[8] Khalid Bou-Rabee and Tasho Kaletha. Quantifying residual finiteness of arithmetic groups.
Compos. Math., 148(3):907–920, 2012.
[9] Khalid Bou-Rabee and D. B. McReynolds. Extremal behavior of divisibility functions.
Geom. Dedicata, 175:407–415, 2015.
[10] Khalid Bou-Rabee and Brandon Seward. Arbitrarily large residual finiteness growth. J.
Reine Angew. Math., 710:199–204, 2016.
[11] N. V. Buskin. Efficient separability in free groups. Sibirsk. Mat. Zh., 50(4):765–771, 2009.
[12] S. C. Chagas and P. A. Zalesskii. Bianchi groups are conjugacy separable. J. Pure Appl.
Algebra, 214(9):1696–1700, 2010.
[13] Sheila C. Chagas and Pavel A. Zalesskii. Finite index subgroups of conjugacy separable
groups. Forum Math., 21(2):347–353, 2009.
[14] Sheila C. Chagas and Pavel A. Zalesskii. Limit groups are subgroup conjugacy separable. J.
Algebra, 461:121–128, 2016.
[15] Jonas Deré. Periodic and eventually periodic points of affine infra-nilmanifold endomorphisms. Discrete Contin. Dyn. Syst., 36(10):5347–5368, 2016.
[16] A. Fel’shtyn and E. Troitsky. Twisted conjugacy separable groups. Preprint, 2012.
[17] Alexander Fel’shtyn. Dynamical zeta functions, Nielsen theory and Reidemeister torsion.
Mem. Amer. Math. Soc., 147(699):xii+146, 2000.
[18] Alexander Fel’shtyn. New directions in Nielsen-Reidemeister theory.
157(10-11):1724–1735, 2010.
Topology Appl.,
[19] Alexander Fel’shtyn and Evgenij Troitsky. Twisted Burnside-Frobenius theory for discrete
groups. J. Reine Angew. Math., 613:193–210, 2007.
[20] Edward Formanek. Conjugate separability in polycyclic groups. J. Algebra, 42(1):1–10,
1976.
[21] A. V. Goryaga. Example of a finite extension of an FAC-group that is not an FAC-group.
Sibirsk. Mat. Zh., 27(3):203–205, 225, 1986.
[22] M. Gromov. Asymptotic invariants of infinite groups. In Geometric group theory, Vol. 2 (Sussex, 1991), volume 182 of London Math. Soc. Lecture Note Ser., pages 1–295. Cambridge
Univ. Press, Cambridge, 1993.
Effective Separability of Finitely Generated Nilpotent Groups
31
[23] A. Grothendieck. Formules de nielse-wecken et de lefschetz en géométrie algébrique. In
Séminaire de Géométrie Algébrique du Bois-Marie 1965-66., volume 569 of SGA 5. Lecture
Notes in Math., pages 407–441. Springer-Verlag, Berlin, 1977.
[24] Philip Hall. The Edmonton notes on nilpotent groups. Queen Mary College Mathematics
Notes. Mathematics Department, Queen Mary College, London, 1969.
[25] Emily Hamilton, Henry Wilton, and Pavel A. Zalesskii. Separability of double cosets and
conjugacy classes in 3-manifold groups. J. Lond. Math. Soc. (2), 87(1):269–288, 2013.
[26] Bo Ju Jiang. Lectures on Nielsen fixed point theory, volume 14 of Contemporary Mathematics. American Mathematical Society, Providence, R.I., 1983.
[27] Martin Kassabov and Francesco Matucci. Bounding the residual finiteness of free groups.
Proc. Amer. Math. Soc., 139(7):2281–2286, 2011.
[28] O. Kharlampovich, A. Myasnikov, and M. Sapir. Algorithmically complex residually finite
groups. Preprint, 2012.
[29] Gady Kozma and Andreas Thom. Divisibility and laws in finite simple groups. Math. Ann.,
364(1-2):79–95, 2016.
[30] Sean Lawton, Larsen Louder, and D. B. McReynolds. Decision problems, complexity, traces,
and representations. Groups Geom. Dyn., 11(1):165–188, 2017.
[31] Armando Martino and Ashot Minasyan. Conjugacy in normal subgroups of hyperbolic
groups. Forum Math., 24(5):889–910, 2012.
[32] Ashot Minasyan and Pavel Zalesskii. Virtually compact special hyperbolic groups are conjugacy separable. Comment. Math. Helv., 91(4):609–627, 2016.
[33] D. V. Osin. Subgroup distortions in nilpotent groups. Comm. Algebra, 29(12):5439–5463,
2001.
[34] Priyam Patel. On a theorem of Peter Scott. Proc. Amer. Math. Soc., 142(8):2891–2906, 2014.
[35] Priyam Patel. On the residual finiteness growths of particular hyperbolic manifold groups.
Geom. Dedicata, 185:87–103, 2016.
[36] M. Pengitore. Effective separability of finitely generated nilpotent groups. Preprint, 2016.
[37] Igor Rivin. Geodesics with one self-intersection, and other stories. Adv. Math., 231(5):2391–
2412, 2012.
[38] Daniel Segal. Polycyclic Groups. Cambridge University Press, 1983.
[39] Salahoddin Shokranian. The Selberg-Arthur trace formula, volume 1503 of Lecture Notes in
Mathematics. Springer-Verlag, Berlin, 1992. Based on lectures by James Arthur.
Effective Separability of Finitely Generated Nilpotent Groups
32
[40] Peter F. Stebe. Conjugacy separability of groups of integer matrices. Proc. Amer. Math. Soc.,
32:1–7, 1972.
[41] Andreas Thom. About the length of laws for finite groups. Israel J. Math., 219(1):469–478,
2017.
| 4math.GR
|
Predicting Lane Keeping Behavior of Visually Distracted Drivers Using
Inverse Suboptimal Control*
arXiv:1604.03984v3 [cs.SY] 29 Apr 2016
Felix Schmitt1 , Hans-Joachim Bieg1 , Dietrich Manstetten1 , Michael Herman1 and Rainer Stiefelhagen2
Abstract— Driver distraction strongly contributes to crashrisk. Therefore, assistance systems that warn the driver if her
distraction poses a hazard to road safety, promise a great
safety benefit. Current approaches either seek to detect critical
situations using environmental sensors or estimate a driver’s
attention state solely from her behavior. However, this neglects
that driving situation, driver deficiencies and compensation
strategies altogether determine the risk of an accident. This
work proposes to use inverse suboptimal control to predict these
aspects in visually distracted lane keeping. In contrast to other
approaches, this allows a situation-dependent assessment of the
risk posed by distraction. Real traffic data of seven drivers are
used for evaluation of the predictive power of our approach.
For comparison, a baseline was built using established behavior
models. In the evaluation our method achieves a consistently
lower prediction error over speed and track-topology variations.
Additionally, our approach generalizes better to driving speeds
unseen in training phase.
I. INTRODUCTION
A. Motivation
Driver distraction is a psychological concept that can be
defined as "diversification of attention away from activities
critical for safe driving towards a competing activity" [1].
Common examples are looking at an in-vehicle display
instead of the road ahead of the vehicle (visual distraction), dialing a telephone number which requires removing
one hand from the steering-wheel (manual distraction) or
conversation (cognitive distraction). According to the U.S.
National Highway Traffic Safety Administration 25% of
police-reported crashes are caused by inattention in form of
distracted or fatigued (e.g. drowsy) drivers [2]. An analysis
of the 100-Car Naturalistic Driving Study even concludes
that approximately 80% of all crashes and 65% of all near
crashes involve inattentive drivers as a contributing factor
[3].
Safe driving requires predominately the acquisition and
processing of visual information for lateral and longitudinal vehicle control [4]. Accordingly, several studies found
distinct negative effects of visual distraction, i.e. too long
off-road glances, on driving performance, e.g. slower and
impaired response to lead vehicle braking [5] or decreased
lane-keeping performance [6]. Although conventional driver
*This work is part of the public project UR:BAN which is co-funded
by the Federal Ministry for Economic Affairs and Energy on basis of a
decision by the German Bundestag.
1 F. Schmitt, H.-J. Bieg, D. Manstetten and M. Herman are with Robert
Bosch GmbH, Corporate Research, 70465 Stuttgart, Germany
2 R. Stiefelhagen is with the Institute for Anthropomatics and Robotics,
Karlsruhe Institute of Technology, Germany.
accepted for IEEE Intelligent Vehicles Symposium (IV 2016)
assistance systems, e.g. forward collision and lane departure
warning, can already partially mitigate these decrements,
optimal assistance has to consider the driver’s distraction
state. Ultimately, the driver needs only support if her abilities
do not suffice to resolve a hazard. Hence, the knowledge of
whether or whether not the driver has currently perceptual
insufficiencies can spare her of false alerts or interventions
by the assistance system, as demonstrated in [7].
B. Related work
Motivated by its importance, several algorithms for detection and assessment of inattention have been proposed
(reviewed by [8]). Due to progress in real-time estimation of
head orientation [9] and eye tracking [10], current approaches
that address visual distraction often process the driver’s gazedirection (or the head-orientation as an approximation) for
prediction. In this work we, similar as other authors ([3],
reviewed by [11]), build on the binary variable x̂ot that
indicates whether the driver’s gaze intersects a so-called
Region Of Interest (ROI) (x̂ot = 1 if and x̂ot = 0 if not).
This region is defined as a plane, through which the driver
has to look to gather necessary visual information for lane
keeping and headway control in the forward road scenery
(Fig. 1).
ROI
Fig. 1.
Intersection of driver’s gaze with a potential ROI
An open question in driver distraction research is how to
algorithmically decide that the driver is too distracted based
on his behavior in a specific situation. In visual distraction
assessment difficulties arise from the fact that safe driving
sometimes even requires to look off-road, e.g. for checking
the speedometer. However, an assistance system would warn
or even intervene on the algorithms decision. Consequently,
if the driver was falsely detected as being critically distracted
too frequently the system would become a nuisance and
cause the driver to turn it off. Nevertheless, any dangerous
situation has to be classified correctly.
Several authors, e.g. [12] and [13], proposed algorithms for
decision on critical distraction that in fact detect the periods
the driver is engaging into a secondary task during the
primary task of safe driving. In most cases such an additional
task of subjective utility and requiring visual attention is the
c 2016 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media,
including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers
or lists, or reuse of any copyrighted component of this work in other works. DOI: <DOI No.>
reason for long off-road glances. However, this approach
has an issue: Particular visually distracting activities, e.g.
selecting a radio station, are commonly socially accepted.
These familiar tasks were for example performed in up to
25% of the driving time in one naturalistic driving study
[14]. In many situations, they pose little risk, as drivers
have developed compensatory mechanisms [15]. If those
engagements indeed have utility to the driver, then she should
rather not receive interventions from the system.
As an alternative to detection of secondary task engagement
the subjectively perceived distraction can be predicted [16].
To achieve this, a-posterior distraction rating based on 10s of
video of the driver and the outside scenery was used. While
this method captures the subjective risk of distraction, it does
not reliably relate to objective risk (e.g. probability a crash)
which can rapidly change over seconds. Additionally, human
ratings are often ambiguous.
Prediction of the objective relative crash risk by different
ROI-based algorithms was investigated by [11] using data
of a naturalistic driving study [3]. Here, similar to the
other approaches, the dependence of "safe" gaze behavior
on the specific situation is neglected. Furthermore, it is
hardly tractable to conduct a study with prototype vehicles
containing a similar amount of crash-data for evaluation.
We follow a different idea: Instead of estimating a driver
distraction state, we predict both the future in-ROI-state x̂ot0
of the driver and the trajectory of the vehicle as driven by the
potentially distracted driver. This allows to assess the current
state x̂ot wrt. lane keeping by estimating the probability of
critical situations caused by driver impairment.
From a methodological point of view, similar inverse optimal
control concepts have been applied to other problems of
assisted [17] and autonomous driving [18]. However, none
of these is suitable for the stochastic and partially observable
control problem considered in this work.
C. Contributions
The main contribution of this work is the proposal of an
inverse suboptimal control approach to predict behavior in
distracted lane keeping. We consider this as a step towards
situation specific visual distraction assessment. Therefore, we
introduce the class of partially observable Markov decision
processes in II-A and show as a first result in II-B how
distracted lane keeping can be modeled as a strategy therein.
Next, we explain how the driver’s objective inferred by
inverse optimal control can be used to predict her behavior
in new situations II-C. Maximum causal entropy inverse
optimal control [19] is presented in II-D as an approach
to account for suboptimal driver behavior. Following, we
report the setting of the driving study, we conducted for
evaluation in III. As a second contribution we finally present
in IV a numerical comparison in terms of predictive power of
our approach against a baseline built from Salvucci’s wellknown two-point steering model [20] and Johnson’s model
for gaze-allocation [21]. Thereby, we investigate both the
overall prediction error and the performance when applying
the models to speeds unseen in the training phase.
II. PROBLEM FORMULATION
A. Partially Observable Markov Decision Processes
Most optimal control problems belong to the class
of Markov Decision Processes (MDPs). An MDP consists of a state x, a control u, a stochastic process model P(xt+1 |xt , ut ) and a reward-function r(x, u).
The objective of an MDP is to find a stochastic policy/controller
π(u|x) that maximizes the expected reward
PT
E[ t=0 r(xt , ut )|π, P, p0 ] over a time horizon T , conditioned on the policy π, the dynamics P and an initial-state
distribution p0 .
The problem addressed in this work belong to the class
of partially observable MDPs (POMDP)s [22], that extend
MDPs to states that cannot directly be observed. Here, first
the a-posterior distribution p(xt |o1,...,t ) of xt has to be
inferred from observations o1,...,t distributed according to
an Observation Model (OM) po (ot |xt ). If the a-posterior
distribution, the so-called belief bt , is known, the POMDP
can be transformed to an equivalent MDP in the belief-state
x0 := b (see [22] for details).
We motivate the application of POMDPs by the idea that
distracted driving is an optimal policy for a combination of
driving performance and utility of an arbitrary secondary
task. As both tasks are in conflict, the primary task by itself is
no longer solved optimally and critical situations can occur.
This is the case, especially, if the driver’s behavior is only
suboptimal as considered in II-D.
B. Model for Distracted Lane Keeping
Distracted lane keeping can be modeled as an optimal
policy π in a POMDP in system- and OM-states (xs , xo )> ,
observations o, system- and OM-controls (us , uo )> and parametric reward r = θs> ϕs (xs , us ) + θo> ϕo (xo , uo ) consisting
of:
1) A time-varying linear-affine model of the dynamics of
the vehicle in xs , us (system-state dynamics)
xst+1 = At xst + Bt ust + at + s ,
(1)
with normal-distributed disturbance N (s |0, Σs ).
Specifically, we used a discretization of the linear
kinematic model given by the ordinary differential
equations
ẏ
0 vt
0
y
0
0
φ̇ = 0 0 c vt φ + 0 α̇ + −vt κt ,
0 0
0
α
1
0
α̇
{z
} |{z} |{z} |{z} | {z }
|{z} |
ẋst
Ât
xst
B̂t
ust
at
with the variables explained in Table I and Fig. 3.
2) A reward θs> ϕs on xst and ust , that encodes the driving
objectives of lane keeping. As a driver seeks to keep
the vehicle in the center of the road by smooth and
minimal steering, the squares of the lane position, the
steering angle and the steering velocity are penalized.
Also ẏ 2 is considered, as the lateral velocity relates to
the impact severity in case of a crash with an object
change of driver’s gaze position
driver’s decision on gaze position
(ROI / not ROI)
driver’s hypothesis
about vehicle states
Kalman-Filter
xot
Observation Model
ot = H(xot )xst
µK
t−1
ΣK
t−1
ΣK
t
Belief Update
H(xo )
ΣK
←−t ΣK
t
t−1
µK
t
H(xo
t ),ot
←−
xot
Policy
K
π(ust , uot |xot , µK
t , Σt )
µK
t
µK
t−1
uot
Observation-Model-State Dynamics
xot+1 = xot + uot mod 2
ust
System-State Dynamics
xst+1 = At xst + Bt ust + at + s
xst
driver’s steering based on
expected vehicle states
Fig. 2.
Control-flow in belief-MDP II-B. Elements of the driver model are indicated in red, while the vehicle model is denoted in blue
tangent
on lc
vt ∆t
dynamics
t
yt φt
curve
radius
1/κ
xot+1 = (xot + uot )
t+1
Fig. 3. Change of the vehicle’s (denoted by 4) position yt → yt+1 for
steering wheel angle αt = 0 and φt , vt , ∆t, κ
TABLE I
VARIABLES OF K INEMATIC V EHICLE M ODEL
α
v
κ
c
Definition
lateral position wrt. lane center lc
angle between tangent
of lane l and vehicle’s longitudinal axis
steering-angle
vehicle’s absolute velocity
curvature of lane
steering-wheel transmission ratio
mod 2.
(4)
4) A reward θo> ϕo (xot , uot ) to account for the secondary
task utility and costs for switching the gaze. We assume
a constant utility for every time-step of engagement by
progress in the task completion and a constant penalty
for gaze-switches. This leads to the reward function
yt+1
lc
y
φ
vehicle’s response to steering input
Unit
m
rad
rad
m/s
1/m
on the neighboring lane. Altogether, this results in a
reward model
θs> ϕs (xst , ust ) = −θ1 yt2 − θ2 ẏt2 − θ3 αt2 − θ4 α̇t2 . (2)
3) A model of the driver’s impairment during visual
distraction. Thereby, the OM-state xot is nothing but
the in-ROI indicator x̂ot from I-B. For the purpose of
this work the attentive driver is assumed to perfectly
perceive all vehicle states, while the driver engaged in
the secondary task is assumed to not perceive y, φ but
sense α with one hand on the steering wheel (see Fig.
4). This can be mathematically implemented by the
xot -dependent observation model
(
e3 e>
xot = 0
3
o s
o
ot = H(xt )xt , H(xt ) =
(3)
I
xot = 1
where I is the identity matrix and e3 = (0, 0, 1)> .
The driver’s ability to switch his gaze between the
secondary task and the road at will, is incorporated
by the OM-control uot ∈ {0, 1} and the OM-state
θo> ϕo (xot , uot ) = θ5 Ixot =1 + θ6 Iuot =1
(5)
where θ5 ≥ 0, θ6 ≤ 0 and I(.) is the indicator (Iχ = 1
for a true expression χ, Iχ = 0 otherwise).
Because of the linearity in (1) and (3) this POMDP can easily
be transformed into its belief-MDP: The belief bt in xst is
fully specified by the a-posterior expectation µK
t and the aposterior covariance ΣK
t , resulting from the corresponding
Kalman-filter. Fig. 2 visualizes the final belief-MDP.
C. Inverse Optimal Control for Behavior Prediction
To predict future system states, the presented POMDP
process model has to be combined with a model of the
driver’s policy π. In this work we indirectly infer π from data
D = {(x, u)i }, by estimating the parameters θ of the driver’s
reward r(x, u) = θ> ϕ(xs , xo , us , uo ) under the assumption
of optimal behavior. This approach is referred to as Inverse
Optimal Control (IOC). Once the reward is known, the
driver’s behavior can be predicted on arbitrary vt , κt by
computation of the policy based on an optimality criterion.
Compared to direct estimation of π, IOC has the advantage
that the parameters θ are less dependent on the external
model variables vt , κt , as they model driver preferences,
potentially applying globally. We empirically investigate this
hypothesis in Sec. IV.
D. Maximum Causal Entropy Inverse Optimal Control
In our case of a linear combination of reward features
r(x, u) = θ> ϕ(x, u), matching the empirical feature expectation
T
hX
i
E[ϕ(x, u)|D] = E
ϕ(xt , ut ) π† , P, p0
(6)
t=0
by the optimal policy π† of reward θ†> ϕ(x, u) results in:
T
hX
i
E[θ†> ϕ(x, u)|D] = E
θ†> ϕ(xt , ut ) π† , P, p0 ,
III. COLLECTED DATA IN DRIVING STUDY
A. Experiment
(7)
t=0
hence θ†> ϕ(x, u) is a reward for which the observed behavior
is optimal [23]. Although many approaches to IOC in MDPs
have been proposed [24] not all are suitable for modeling
real-world driving behavior. The reason is that humans often
deviate from the optimal behavior [25] what needs to be
taken into account in modeling distracted driving. Therefore,
we apply the Maximum Causal Entropy (MCE) method [19].
Here, a stochastic policy π(u|x) of maximum causal entropy,
that fullfills condition (6), is computed by the following
optimization problem:
T
hX
i
max − E
log π(ut |xt ) π, P, p0
π(u|x)
(8)
t=0
T
hX
i
s.t. E[ϕ(x, u)|D] = E
ϕ(xt , ut ) π, P, p0 .
(9)
t=0
The rationale behind the maximization of the casual entropy
of π is that the resulting policy effectively deviates least from
the true behavior π ? if a worst-case π ? is assumed (details
in [19]). Hence, the estimation is very robust.
1) Suboptimal policies from MCE: Although reward parameters θ are not explicitly estimated in MCE, it is considered an inverse optimal control approach due to the following
property: The likelihood of the estimated π to act like any
deterministic policy u = π̄(x) is a monotonic function of
T
hX
i
E
θ̂> ϕ(xt , ut ) π̄, P, p0 ,
To evaluate our modeling approach, we conducted a study
on a public highway. We decided against a study in a driving
simulator because of possible influences on the participants’
behavior by absence of real risk. In addition to that, evaluation of prediction robustness on realistic sensor-input, i.e.
noisy signals, is important.
We recruited seven drivers (6 male, 1 female). All drivers
took part in a driving safety training prior to the experiment.
The experiment consisted of four fixed driving speed conditions {80, 90, 100, 110} km/h. Vehicle speed was controlled
by the vehicle’s Adaptive Cruise Control (ACC) to prevent
drivers from adjusting their speed as a compensatory action
while being engaged in a secondary task. A conservative time
gap was employed to ensure that the distance to preceding
vehicles did have the least possible influence on the drivers’
behavior. When the vehicle traveled at the required speed,
the measurement periods were started. Such a period was
either a reference, where the driver had to keep the lane fully
attentive or involved a visually distracting secondary task. At
each speed three secondary tasks and three reference periods
per participant were triggered by the investigator.
The task we used in this experiment consisted of incrementally reading a series of 30 random numbers ni ∈ {1, 2}
from a display and typing them in a number pad (see Fig.
4). Thereby, the participants were instructed to "perform the
(10)
t=0
where θ̂ is the Lagrangian multiplier of constraint (9)
[26](c.6). That means the higher the expected value of
θ̂> ϕ(x, u) the likelier is the behavior according to π̄. Consequently, π is interpreted as a suboptimal, yet goal-oriented
policy, with respect to the reward model θ̂> ϕ(x, u).
2) Parameter learning in MCE-IOC: The optimization
problem (8)-(9) is solved for optimal π? by gradient descent
on its Lagrangian-dual problem in θ [19]:
1) Calculate π for current θ using the recursion
Q̃(uT , xT ) = θ> r(uT , xT ),
(11)
Z
V (xt+1 ) = log exp Q̃(ut+1 , xt+1 ) dut+1 ,
Q̃(ut , xt ) = θ> r(ut , xt ) + E V (xt+1 ) P ,
Z
π(ut |xt ) = exp(Q̃(ut , xt ))/ exp(Q̃(ũt , xt )) dũt .
2) Simulate π in the MDP starting from p0 to compute
T
hX
i
E=E
ϕ(xt , ut ) π, P, p0
t=0
(12)
3) Update θ ← θ − η E[ϕ(x, u)|D] − E using a suitable
step-size η.
Fig. 4.
Artificial secondary task used in the experiment
secondary task as quickly and correctly as possible while
not endangering the driving safety". This artificial task was
chosen, as it resembles the principle of a variety of real
visual-manual tasks performed while driving, and possesses
several advantages. First, the task state is fully measurable
and can be modeled easily, in contrast to the vehicle’s
interaction system. Second, the participants needed only little
practice to reach maximum execution performance, resulting
in no significant learning effects during the experiment.
B. Recorded Data
We used a series mono-camera system for tracking the
lane boundaries and recorded the position of the lane, the
angle between tangent of the lane boundary and the vehicle’s
longitudinal axis and the curvature via the vehicles Controller
Area Network (CAN). A commercial three-camera infra red
eye tracking system with active illumination was used for
estimation of the driver’s gaze direction, while an own ROIbased algorithm detected whether the driver’s gaze was off
or on the road. Steering wheel position and velocity as
well as absolute velocity were also recorded via the CAN.
Hence, beside the eye-tracking systems we relied solely
on signals that are already accessible in today’s seriesproduction vehicles.
y
φ
β1
w/2
~v
β2
1/κ
C. Preprocessing
In order to ensure a quality dataset for the numerical
evaluation, the following preprocessing and filtering steps
were performed on the collected raw data:
1) Selection of valid trials according to protocol: We
automatically excluded lane changes and their preparation
phases. As a different driving maneuver than lane keeping it
requires a different driving and gaze policy. Also situations
where the ACC controller reduced the vehicle speed by more
than 5% were left out due to possible influence on the
drivers’ behavior. The final dataset consisted of 136 valid
segments comprising reference and secondary task periods
with an average duration of ≈ 50 s.
2) Sub-sampling and filtering of signals: We first subsampled the variables of the vehicle model to 25 Hz. Thereafter, the Rauch-Tung-Striebel-smoother [27] was employed
for filtering of the signals using the kinematic vehicle model
(1). The system parameters c, Σs as well as the sensor
noise-covariance Σm were estimated by an expectationmaximization approach as proposed for vehicle models by
[28].
IV. EVALUATION
To demonstrate the effectiveness of our IOC-based approach we conducted two numerical evaluations. Here we
compared against a baseline comprising of established behavior models.
A. Baseline
[21] presented a model for gaze-allocation in visual dualtasking, where the probability of a glance to a task is a
logistic function of the uncertainty in its states. In our case
uncertainty is only present in the vehicle states - the random
number ni is either known to the driver if she has seen it
on the display or unknown else. Therefore, we applied the
following variant of the original approach
Iuo =0 exp(λ1 ) + Iuot =1
(13)
= 1) = t
exp(λ1 ) + 1
Iuo =0 exp(λ2 + tr(ΛΣK
t )) + Iuo
t =1
p(uot |xot = 0) = t
. (14)
K
exp(λ2 + tr(ΛΣt )) + 1
p(uot |xot
Here λ1 , λ2 , Λ are parameters and tr(.) is the sum of the
matrix diagonal elements.
For modeling human predictive steering we adopted the twopoint-steering model of [20]. In this model the driver steers
by means of a visual near-angle β1 and a visual far-angle β2 .
In a lane l of width w, β1 is defined as the angle between
the straight through the vehicle’s center and a near point
on the lane center lc and the vehicle’s longitudinal axis. β2
is defined as the angle of minimal magnitude between the
tangents from the vehicle’s center to both boundaries lb and
lb
Fig. 5.
lc
Illustration of the variables of the two-point steering model.
the vehicle’s longitudinal axis (see Fig. 5).
We computed β1 , β2 using the approximations
β1 = − arctan y (2 m)−1 − φ
(
−1
− arccos 1 − κ (w/2 + y)
−φ
−1
β2 =
+ arccos 1 + κ (w/2 − y)
−φ
(15)
κ<0
κ≥0
(16)
and fed both angles and the steering-wheel angle into the
stochastic policy α̇ = k1 β1 + k2 β2 + k3 α + k with parameters k1 , k2 , k3 and a normally-distributed random-variable k
([20] used a PD-controller that was, however, less stable in
combination with our approximation due to jumps in β2 ).
Both models were linked by replacing y, φ in (15),(16) with
their expectations if xot = 0, i.e. when the modeled driver
looks off the road, similar to [29].
B. Evaluation protocol
For the numerical evaluation we further subdivided all
valid segments into snippets of ≈ 5 s (overlap ≈ 2.5 s)
to account for a realistic prediction horizon in a real-time
system.
In the training-phase of the MCE-IOC approach we relaxed
the feature matching condition (6) to
i
hP
T
E[ϕ(x, u)|D]i − E
t=0 ϕ(x, u) π, P
i
≤ 1%, (17)
E[ϕ(x, u)|D]i
(equivalent to a L1 regularization of θ) to prevent over-fitting.
The parameters of the baseline were inferred using L1regularized maximum likelihood estimation for generalized
linear models as implemented in MATLAB’s lassoglm
function [30].
In every interval 100 trajectories xs , xo were sampled for
each model, starting in the first state of the segment. Thereby,
we incrementally simulated the driver’s behavior π and
the responses of the POMDP model (II-B) using the preestimated parameters c, Σs from III-C. On the obtained data
the following metrics were computed:
•
Expected squared error of lane-position
T
1 X
SE(y) = E
(yt − ytD )2 |π, P, p0
T t=0
(18)
•
Kullback-Leiber divergence between empirical off-road
gaze duration d and prediction of the model
X
KL(pD ||p) =
pD (d) log pD (d) − log p(d)
d
SE
0.4
0.2
(19)
1) Overall prediction performance: We first evaluated
the overall prediction performance by splitting the dataset
into a training set and a test set of equal size randomly
and independently of driver, velocity and track-topology.
Afterwards the roles of the datasets were swapped. For better
estimation of the error statistics this 2-fold cross-validation
procedure was repeated 10 times.
2) Transfer performance: To investigate the generalization quality on unseen velocities we conducted a second
evaluation. Here we trained on a random selection of half of
the data of one single speed condition. We thereafter tested
on the remaining half with the same and on all data of other
velocities. In this evaluation we performed 5 repetitions of
2-fold cross-validation.
0
KL
0.6
0.4
0.2
0
ours
Train
ours
Test
ours
Same
ours
Trans
base
Train
base
Test
base
Same
base
Trans
Fig. 6.
Box indicates the [25, 75]-% interval I with the median as
horizontal line. Data whose distance to the median exceeds 1.5× distance
median to corresponding quantile, is considered outliers (··). Train and Test
denote the conditions of IV-B.1, while Same denotes the test error on the
same speed and Transfer the test error on the other unseen speeds of IV-B.2.
C. Results
The results of IV-B.1 are summarized in Table II. Due to
the skew-shaped error distributions (see Fig.6) we report the
total median (instead of the mean) in training and test of
both methods in both metrics.
TABLE II
OVERALL P REDICTION PERFORMANCE
Metric
SE
KL
Baseline
Train
0.0482
0.0998
Test
0.0484
0.1006
Ours
Train
0.0154
0.0731
Test
0.0157
0.0727
Table III shows the errors of IV-B.2. We present the total
median errors of the 10 evaluations. Thereby, the results of
our approach are in bold letters.
TABLE III
T RANSFER PERFORMANCE
Train
80 km/h
SE
KL
90 km/h
SE
KL
100 km/h
SE
KL
110 km/h
SE
KL
80 km/h
0.0154
0.0293
0.0851
0.1280
0.0152
0.0310
0.0833
0.1307
0.0147
0.0311
0.0838
0.13743
0.0157
0.0401
0.0825
0.1249
90 km/h
0.0143
0.0572
0.0744
0.1180
0.0148
0.0365
0.0748
0.1146
0.0160
0.0390
0.0763
0.1124
0.0152
0.0408
0.0724
0.1076
Test
100 km/h
0.0164
0.0725
0.0738
0.1025
0.0159
0.0477
0.0730
0.1018
0.0155
0.0450
0.0731
0.1100
0.0179
0.0479
0.0755
0.0901
110 km/h
0.0154
0.0836
0.0701
0.1216
0.0163
0.0627
0.0712
0.1200
0.0161
0.0584
0.0693
0.1173
0.0177
0.0603
0.0721
0.1102
Fig. 6 depicts the error distributions of both evaluations.
D. Discussion
IV-B.1 shows significantly lower prediction error of our
approach based on MCE-IOC compared to the baseline
(Wilcoxon signed-rank ptest ≤ 10−12 ). As training and test
error for both methods and in both metrics differed only
slightly (Wilcoxon ranksum ptest ≥ 0.01), adequate regularization was chosen for both methods. A possible explanation
for the better SE of our approach can be that the explicit
incorporation of the vehicle’s velocity and the future track
curvature allows more precise trajectory prediction. Although
Johnson’s barrier model already incorporates the growing
uncertainty over time it might have worse KL than the MCEIOC approach because it lacks a direct link to lane keeping
controller.
IV-B.2 gives further insight into the performance differences:
MCE-IOC performed significantly better in both metrics
(ptest ≤ 10−12 ) over all conditions except for KL Train
110 km/h Test 100 km/h (ptest = 0.01). Although a test on
between-condition differences revealed significant variations
(Kruskal-Wallis ptest ≤ 10−5 ), these were more pronounced
in the baseline (ptest ≤ 10−12 vs. ptest ≥ 10−6 in MCE-IOC)
what could be validated by significantly lower variance
(Ansari-Bradley ptest ≤ 10−12 ).
The hypothesis that the driver preferences θ estimated by
MCE-IOC are better transferable than a directly estimated
policy could be confirmed in this experiment by IV-B.2:
The difference between the error on unseen speeds and the
median test error on seen speeds was significantly higher for
the baseline approach in both metrics (Wilcoxon ranksum
SE ptest = 10−11 , KL ptest = 10−9 ). However, the general
hypothesis has to be further investigated on significantly
lower speeds in future work.
V. C ONCLUSION
We presented a inverse optimal control based method
to predict the behavior of potentially distracted driver in
lane keeping. Here, the key elements are a model of goaloriented driver behavior wrt. a reward, an explicit model
of the driver’s impairment and a dynamics model of the
vehicle responses. Once the reward is inferred from observed
behavior, it can be transferred to new situations for estimation
of likely driver behavior what was empirically validated for
unseen driving speeds. In the evaluation we also showed
lower prediction error of our method compared to a baseline
comprising of domain-specific models [21],[20].
Our approach can be used for situation-dependent assessment
of visual driver distraction wrt. lane keeping, by prediction
of the probability of critical incidents, i.e. lane-departure.
However, for deployment in a final distraction mitigation
system it has to be completed with prediction of lateral driving behavior. Hence, in future work we will address inverse
optimal control approaches for modeling speed adjustment
and headway control of distracted drivers. Additionally, we
also want to incorporate estimation of observation models
II-B, eq. (3) into the inference procedure. This is necessary
for behavior prediction in secondary task, e.g. reading the
speedometer, where the driver can to some extent perceive
the road by peripheral vision. Here, the knowledge of suboptimal behavior of the driver can also be exploited [31].
R EFERENCES
[1] J. D. Lee, K. L. Young, and M. A. Regan, “Defining driver distraction,”
Driver distraction: Theory, effects, and mitigation, p. 31, 2008.
[2] T. A. Ranney, W. Garrott, and M. Goodman, “Nhtsa driver distraction
research: past, present, and future,” in Proceedings of the International
Technical Conference on the Enhanced Safety of Vehicles, vol. 2001.
National Highway Traffic Safety Administration, 2001, pp. 9–p.
[3] S. G. Klauer, T. A. Dingus, V. L. Neale, J. D. Sudweeks, and D. J.
Ramsey, “The impact of driver inattention on near-crash/crash risk:
An analysis using the 100-car naturalistic driving study data,” National
Highway Traffic Safety Administration, Washington D.C., Tech. Rep.
DOT HS 810 594, 2006.
[4] M. Sivak, “The information that drivers use: is it indeed 90% visual?”
Perception, vol. 25, pp. 1081–1089, 1996.
[5] T. A. Dingus, M. C. Hulse, J. F. Antin, and W. W. Wierwille, “Attentional demand requirements of an automobile moving-map navigation
system,” Transportation Research Part A: General, vol. 23, no. 4, pp.
301–315, 1989.
[6] Y. Peng, L. N. Boyle, and S. L. Hallmark, “Driver’s lane keeping
ability with eyes off road: Insights from a naturalistic study,” Accident
Analysis & Prevention, vol. 50, pp. 628–634, 2013.
[7] C. Blaschke, F. Breyer, B. Färber, J. Freyer, and R. Limbacher, “Driver
distraction based lane-keeping assistance,” Transportation research
part F: traffic psychology and behaviour, vol. 12, no. 4, pp. 288–299,
2009.
[8] Y. Dong, Z. Hu, K. Uchimura, and N. Murayama, “Driver inattention
monitoring system for intelligent vehicles: A review,” IEEE Transactions on Intelligent Transportation Systems., vol. 12, no. 2, pp. 596–
614, 2011.
[9] A. Tawari, S. Martin, and M. M. Trivedi, “Continuous head movement estimator for driver assistance: Issues, algorithms, and on-road
evaluations,” IEEE Transactions on Intelligent Transportation Systems,
vol. 15, no. 2, pp. 818–830, 2014.
[10] D. W. Hansen and Q. Ji, “In the eye of the beholder: A survey of
models for eyes and gaze,” IEEE Transactions on Pattern Analysis
and Machine Intelligence., vol. 32, no. 3, pp. 478–500, 2010.
[11] Y. Liang, J. D. Lee, and L. Yekhshatyan, “How dangerous is looking
away from the road? algorithms predict crash risk from glance patterns
in naturalistic driving,” Human Factors: The Journal of the Human
Factors and Ergonomics Society, vol. 54, no. 6, pp. 1104–1116, 2012.
[12] M. Wollmer, C. Blaschke, T. Schindl, B. Schuller, B. Farber, S. Mayer,
and B. Trefflich, “Online driver distraction detection using long
short-term memory,” IEEE Transactions on Intelligent Transportation
Systems, vol. 12, no. 2, pp. 574–582, 2011.
[13] J. D. Lee, J. Moeckli, T. L. Brown, S. C. Roberts, C. Schwarz,
L. Yekhshatyan, E. Nadler, Y. Liang, T. Victor, D. Marshall et al.,
“Distraction detection and mitigation through driver feedback,” National Highway Traffic Safety Administration, Washington D.C., Tech.
Rep. DOT HS 811 547A, 2013.
[14] B. Metz, S. Schoch, M. Just, and F. Kuhn, “How do drivers interact
with navigation systems in real life conditions?: Results of a fieldoperational-test on navigation systems,” Transportation Research Part
F: Traffic Psychology and Behaviour, vol. 24, pp. 146–157, 2014.
[15] N. Schömig, B. Metz, and H.-P. Krüger, “Anticipatory and control
processes in the interaction with secondary tasks while driving,”
Transportation research part F: traffic psychology and behaviour,
vol. 14, no. 6, pp. 525–538, 2011.
[16] N. Li and C. Busso, “Predicting perceived visual and cognitive
distractions of drivers with multimodal features,” IEEE Transactions
on Intelligent Transportation Systems,, vol. PP, no. 99, pp. 1–15, 2014.
[17] M. Shimosaka, T. Kaneko, and K. Nishi, “Modeling risk anticipation
and defensive driving on residential roads with inverse reinforcement
learning,” in Proceedings of the IEEE 17th International Conference
on Intelligent Transport Systems (ITSC), 2014.
[18] M. Kuderer, S. Gulati, and W. Burgard, “Learning driving styles for
autonomous vehicles from demonstration,” in Proceedings of the IEEE
International Conference on Robotics & Automation (ICRA), vol. 134,
2015.
[19] B. D. Ziebart, J. A. Bagnell, and A. K. Dey, “The principle of
maximum causal entropy for estimating interacting processes,” IEEE
Transactions on Information Theory., vol. 59, no. 4, pp. 1966–1980,
2013.
[20] D. D. Salvucci and R. Gray, “A two-point visual control model of
steering,” Perception-London, vol. 33, no. 10, pp. 1233–1248, 2004.
[21] L. Johnson, B. Sullivan, M. Hayhoe, and D. Ballard, “Predicting
human visuomotor behaviour in a driving task,” Philosophical Transactions of the Royal Society B: Biological Sciences, vol. 369, no. 1636,
p. 20130044, 2014.
[22] R. D. Smallwood and E. J. Sondik, “The optimal control of partially
observable markov processes over a finite horizon,” Operations Research, vol. 21, no. 5, pp. 1071–1088, 1973.
[23] P. Abbeel and A. Y. Ng, “Apprenticeship learning via inverse reinforcement learning,” in Proceedings of the 21th International conference
on Machine learning (ICML). ACM, 2004, p. 1.
[24] S. Zhifei and E. M. Joo, “A review of inverse reinforcement learning theory and recent advances,” in Proceedings of the 2012 IEEE
Congress on Evolutionary Computation (CEC)., 2012, pp. 1–8.
[25] H. A. Simon, “Theories of decision-making in economics and behavioral science,” The American economic review, pp. 253–283, 1959.
[26] B. D. Ziebart, J. Bagnell, and A. K. Dey, “Modeling interaction via
the principle of maximum causal entropy,” in Proceedings of the 27th
International Conference on Machine Learning (ICML), 2010, pp.
1255–1262.
[27] H. E. Rauch, C. Striebel, and F. Tung, “Maximum likelihood estimates
of linear dynamic systems,” AIAA journal, vol. 3, no. 8, pp. 1445–
1450, 1965.
[28] J. E. Stellet, F. Straub, J. Schumacher, W. Branz, and J. M. Zöllner,
“Estimating the process noise variance for vehicle motion models,” in
Proceedings of the IEEE 18th International Conference on Intelligent
Transportation Systems (ITSC), 2015, pp. 1512–1519.
[29] L. Johnson, B. Sullivan, M. Hayhoe, and D. Ballard, “A soft barrier
model for predicting human visuomotor behavior in a driving task,” in
Proceedings of the 35th Annual Conference of the Cognitive Science
Society. Citeseer, 2013, pp. 689–691.
[30] MATLAB Statistics ToolboxTM User’s Guide, R2014b ed., The MathWorks, Inc., 3 Apple Hill Drive Natick, MA 01760-2098, 2014.
[31] M. Golub, S. Chase, and M. Y. Byron, “Learning an internal dynamics model from control demonstration,” in Proceedings of the 30th
International Conference on Machine Learning (ICML)., 2013, pp.
606–614.
| 3cs.SY
|
Type Classes for Lightweight Substructural Types
Edward Gan
Jesse A. Tov
Greg Morrisett
Facebook, Menlo Park
Northeastern University, Boston
Harvard University, Cambridge
[email protected]
[email protected]
[email protected]
Linear and substructural types are powerful tools, but adding them to standard functional programming languages often means introducing extra annotations and typing machinery. We propose a
lightweight substructural type system design that recasts the structural rules of weakening and contraction as type classes; we demonstrate this design in a prototype language, Clamp.
Clamp supports polymorphic substructural types as well as an expressive system of mutable
references. At the same time, it adds little additional overhead to a standard Damas–Hindley–Milner
type system enriched with type classes. We have established type safety for the core model and
implemented a type checker with type inference in Haskell.
1
Introduction
Type classes [12, 19] provide a way to constrain types by the operations they support. If the type class
predicate Dup α indicates when assumptions of type α are subject to contraction (duplication), and
Drop α indicates whether they are subject to weakening (dropping), then linear, relevant, affine, and
unlimited typing disciplines are all enforced by some subset of these classes. Linear types, then, are
types that satisfy neither Dup nor Drop. This idea, suggested in one author’s dissertation [16], forms the
basis of our prototype substructural programming language Clamp.
Clamp programs are written in a Haskell-like external language in which weakening and contraction
are implicit. This is easier for programmers to work with, but to specify the type system and semantics
the external language is elaborated into an internal language that is linear (i.e. variables are used exactly
once.) The internal language provides explicit dup and drop operations, which impose the corresponding
type class constraints on their arguments. Thus, in the internal language one might think of dup and drop
as functions with these qualified types:
dup : ∀α. Dup α ⇒ α → α × α
drop : ∀αβ . Drop α ⇒ α → β → β
In the internal language all nonlinear usage is mediated by the dup and drop operations. For example,
the internal language term λ x. x + x is ill formed because it uses variable x twice, but the term
λ x. let (x1 , x2 ) = dup x in x1 + x2
is well typed. Because elaboration into the internal language ensures that the resulting program is linear,
it can then be checked using nearly-standard Damas–Hindley–Milner type reconstruction [8] with type
classes [12, 19]; improper duplication and dropping is indicated by unsatisfiable type class constraints.
S. Alves and I. Cervesato (Eds): Third International Workshop
on Linearity 2014 (LINEARITY’14)
EPTCS 176, 2015, pp. 34–48, doi:10.4204/EPTCS.176.4
Gan, Tov & Morrisett
fst
fst
35
:: Drop b => (a , b ) -U> a
= \(x , y ) -U> x
constU :: (Dup a , Drop a , Drop b ) => a -U> b -U> a
constU = \x -U> \y -U> x
constL :: Drop b => a -U> b -L> a
constL = \x -U> \y -L> x
Figure 1: Prelude functions with inferred signatures
Contributions. We believe that Clamp offers substructural types with less fuss than many prior approaches to programmer-facing substructural type systems. Throughout the design, we leverage standard
type class machinery to deal with most of the constraints imposed by substructural types. Implementing
type inference for Clamp (§3) is straightforward and it is also easy to extend the system with custom
resource aware structures (§2.4). The specific contributions in this paper include:
• a type system design with polymorphic substructural types and a type safety theorem (§2);
• a flexible system for managing weak and strong references (§2.4);
• a type checker with type inference derived from a type checker for Haskell (§3); and
• a dup-and-drop–insertion algorithm that is in some sense optimal (§3.1).
1.1
Clamp Basics
In this section, we introduce the Clamp external language, in which dup and drop operations are implicit.
The concrete syntax is borrowed from Haskell, but one prominent difference in Clamp is that each
function type and term must be annotated with one of four substructural qualifiers: U for unlimited, R for
relevant, A for affine, or L for linear.
Three examples of Clamp functions, translated from the Haskell standard prelude, appear in figure 1.
Their types need not be written explicitly, and are inferred by Clamp’s type checker.
Consider the fst function, which projects the first component of a pair. Because we would like be
able to use library functions any number of times or not at all, we annotate the arrow in the lambda expression with qualifier U. This annotation determines the function type’s structural properties—meaning,
in this case, that fst satisfies both Dup and Drop. (Note that this is a property of the function itself, not
of how it treats its argument.) Because fst does not use the second component of the pair, this induces
the Drop b constraint on type variable b . In particular, elaboration into the internal language inserts
a drop operation for y to make the term linear: \(x , y ) -U> drop y x . The presence of drop ,
which disposes of its first argument and returns its second, causes the Drop type class constraint to be
inferred.
Function constU imposes a similar constraint on its second argument, but it also requires the type of
its first argument be unlimited. This is because constU returns an unlimited closure containing the first
argument in its environment. The argument is effectively duplicated or discarded along with the closure,
so it inherits the same structural restrictions. Alternatively, we can lift this restriction with constL ,
which returns a linear closure and thus allows the first argument to be linear.
36
Type Classes for Lightweight Substructural Types
e ::=
|
|
v ::=
x | v | e1 e2 | e [τi ] | (e1 , e2 ) | letp (x1 , x2 ) = e in e2 | Inl e | Inr e
case e of Inl x1 → e1 ; Inr x2 → e2 | newrq e | releaserq e | swaprq e1 with e2
dup e1 as x1 , x2 in e2 | drop e1 in e2
λ aq x:τ. e | Λαi [P] . v | (v1 , v2 ) | Inl v | Inr v | ` | ()
aq
rq
τ ::= α | τ1 −→ τ2 | τ1 × τ2 | τ1 + τ2 | Unit | Ref τ | ∀αi .P ⇒ τ
P
rq
aq
K
::=
::=
::=
::=
(K 1 τ1 , . . . , K n τn )
s (strong) | w (weak)
U (unlimited) | R (relevant) | A (affine) | L (linear)
Dup | Drop
(terms)
(values)
(types)
(constraints)
(reference qualifiers)
(arrow qualifiers)
(predicate constructors)
Figure 2: Syntax of λcl
2
Formalizing λcl
To validate the soundness of our approach, we have developed λcl , a core model of the Clamp internal
language. λcl is based on System F [11] with a few modifications: variable bindings are treated linearly,
arrows are annotated with qualifiers, and type class constraints [12, 19] are added under universal quantifiers. As an example of how one can define custom usage-aware datatypes in Clamp, λcl also includes
a variety of operations for working with mutable references.
The λcl type system shares many similarities with Tov’s core Alms calculus: a λms [17]. Unlike the
external Clamp language prototype (§3), λcl provides first-class polymorphism and does not support type
inference.
2.1
Syntax of λcl
The syntax of λcl appears in figure 2. Most of the language is standard, but notably arrow types and
λ terms in Clamp are annotated with an arrow qualifier (aq). These annotations determine which structural operations a function supports, as well as the corresponding constraints imposed on the types in
aq
its closure environment. Unlike some presentations of linear logic, −→ here constrains usage of the
U
function itself, not usage of the function’s argument. Thus one can call dup on a −→ arrow but not on an
L
−→ arrow. Type abstractions specify the type class constraints that they abstract over; their bodies are
restricted to values, so unlike λ terms, type abstractions do not need an arrow qualifier.
The newrq e and releaserq e forms introduce and eliminate mutable references. Each comes in two
flavors depending on its reference qualifier (rq), which records whether the reference supports strong
or merely weak updates. Weak (conventional) updates must preserve a reference cell’s type, but strong
updates can modify both the value and type of a cell.
Form swaprq e1 with e2 provides linear access to a reference by exchanging its contents for a different
value. The release operator deallocates a cell and returns its contents if it is not aliased. Store locations
(`) appear at run time but are not written by the programmer.
To incorporate type classes, universal types may include constraints on their type variables. A constraint P denotes a set of atomic predicate constraints Kτ, each of which is a predicate constructor K
applied to a type. For the sake of our current analysis, K is either Dup or Drop.
Gan, Tov & Morrisett
E ::=
|
|
µ ::=
37
[·] | E e | v E | E [τi ] | (E, e) | (v, E) | Inl E | Inr E
(evaluation contexts)
rq
case E of Inl x1 → e1 ; Inr x2 → e2 | letp (x1 , x2 ) = E in e | new E | releaserq E
swaprq E with e | swaprq v with E | dup E as x1 , x2 in e | drop E in e
` 7→i v, µ | ·
(stores)
Figure 3: Runtime structures
(µ ; (λ x:τ. e) v) 7−→ (µ ; {v/x} e)
(µ ; (Λαi [P] . v) [τi ]) 7−→ (µ ; {τi /αi }v)
(µ ; newrq v) 7−→ (µ, ` 7→1 v ; `)
` fresh
(µ, ` 7→i v1 ; swaprq ` with v2 ) 7−→ (µ, ` 7→i v2 ; (`, v1 ))
(µ, ` 7→1 v ; releasew `) 7−→ (µ ; Inl v)
(µ, ` 7→i v ; releasew `) 7−→ (µ, ` 7→i−1 v ; Inr ())
1
when i > 1
s
(µ, ` 7→ v ; release `) 7−→ (µ ; v)
(µ ; dup v as x1 , x2 in e) 7−→ (incr(locs (v) ; µ) ; {v/x2 } {v/x1 } e)
(µ ; drop v in e) 7−→ (decr(locs (v) ; µ) ; e)
(µ1 ; E [e1 ]) 7−→ (µ2 ; E [e2 ])
when (µ1 ; e1 ) 7−→ (µ2 ; e2 )
Figure 4: Small-step relation
2.2
Semantics
The execution of λcl terms can be defined by a call-by-value small-step semantics with evaluation contexts and a global, reference-counted store µ. The run-time structures needed to define this small step
relation are given in figure 3. Reference counts are used to track when reference cells can be safely
deallocated in the presence of aliasing. Exchange properties for the store are implicitly assumed. A
selection of the small step relation rules are given in figure 4, focusing on the rules for reference cells
and substructural operations. Most of the complexity here comes from the reference counts.
The swap operator exchanges the contents of a cell in the heap with a different value. The release
operator deallocates a cell and returns its contents if it is not aliased. However, if a cell has been aliased
it decrements the reference count and returns a unit. The dup and drop operators manipulate reference
counts as expected.
A few metafunctions given in figure 5 are necessary to maintain reference counts, similar to those
in [7]. The functions incr(`; µ) and decr(`; µ) allow us to increment and decrement reference counts in
the heap. Incrementing a location is straightforward, but a decrement must be defined recursively since
deallocating the last pointer to a reference cell involves decrementing the reference counts of all cells the
deallocated contents originally pointed to.
The locs meta-function is a convenient way of extracting the multiset of locations that a value uses.
Note that the + (or ]) operator is a multiset operator which additively combines occurrences, and is used
again in section 3.1. The locs function is also designed to operate on well-typed terms, so it only looks
at one branch of a case expression, assuming that the other branch must share the same location typing
context.
38
Type Classes for Lightweight Substructural Types
incr(`; ` 7→ j v, µ) = ` 7→ j+1 v, µ
(
decr(locs (v) ; µ)
j
decr(`; ` 7→ v, µ) =
` 7→ j−1 v, µ
locs (`) = {`}
aq
locs (λ x:τ. e) = locs (e)
j=1
j>1
incr({`1 , . . . , `k }; µ) = incr(`1 ; · · · incr(`k ; µ))
decr({`1 , . . . , `k }; µ) = decr(`1 ; · · · decr(`k ; µ))
locs (e1 e2 ) = locs (e1 ) + locs (e2 )
case e of
locs Inl x1 → e1 ; = locs (e) + locs (e1 )
Inr x2 → e2
···
Figure 5: Reference count management
VAR
TA PP
TA BS
P1 , P2 ; Γ; Σ ` v : τ
P; x:τ; · ` x : τ
P1 ; Γ; Σ ` e : ∀αi .P2 ⇒ τ
dom P2 ⊆ αi
P; Γ, x:τ1 ; Σ ` e : τ2
P
Constrainaq (Γ; Σ)
A PP
aq
P; Γ1 ; Σ1 ` e1 : τ2 −→ τ
aq
aq
{τi /αi }P2
P1 ; Γ; Σ ` e [τi ] : {τi /αi }τ
P1 ; Γ; Σ ` Λαi [P2 ] . v : ∀αi .P2 ⇒ τ
A BS
P1
P; Γ; Σ ` λ x : τ1 . e : τ1 −→ τ2
P; Γ2 ; Σ2 ` e2 : τ2
P; Γ1 + Γ2 ; Σ1 + Σ2 ` e1 e2 : τ
C ASE
P; Γ1 ; Σ1 ` e1 : τ11 + τ12
P; Γ2 , x21 :τ11 ; Σ2 ` e21 : τ2
P; Γ2 , x22 :τ12 ; Σ2 ` e22 : τ2
P; Γ1 + Γ2 ; Σ1 + Σ2 ` case e1 of Inl x21 → e21 ; Inr x22 → e22 : τ2
D UP
D ROP
P; Γ1 ; Σ1 ` e1 : τ1
P; Γ2 , x1 :τ1 , x2 :τ1 ; Σ2 ` e2 : τ2
P
Dup τ1
P; Γ1 + Γ2 ; Σ1 + Σ2 ` dup e1 as x1 , x2 in e2 : τ2
P; Γ1 ; Σ1 ` e1 : τ1
P; Γ2 ; Σ2 ` e2 : τ2
P Drop τ1
P; Γ1 + Γ2 ; Σ1 + Σ2 ` drop e1 in e2 : τ2
Figure 6: Selected λcl term typing rules
2.3
Term Typing
Variable contexts Γ ::= x1 :τ1 , . . . , xn :τn associate variables with types, where each variable appears at
most once. Location contexts (store typings) Σ ::= `s 7→s τs , . . . , `w 7→kww τw , . . . associate locations (`)
with their reference types Ref τ, and distinguish between strong and weak locations; weak locations
carry a reference count k to track aliasing.
Linearity is enforced in λcl via standard context-splitting. Because Γ and Σ are linear environments,
we need operations to join them. The join operation + is defined only on pairs of compatible environments, written Γ1 ^ Γ2 and Σ1 ^ Σ2 . Two variable contexts are compatible so long as they are disjoint.
Two location contexts are compatible if the strong locations are disjoint and the weak locations in their
intersection agree on their types. Joining variable contexts appends the two sets of bindings together,
while joining location contexts also involves adding the reference counts of any shared weak locations.
Contexts are identified up to permutation.
The term typing judgment (P; Γ; Σ ` e : τ) assigns term e type τ under constraint, variable, and
location contexts P, Γ, and Σ. Selected typing rules for the core language appear in figure 6, and the
Gan, Tov & Morrisett
39
N EW
L OC W
L OC S
P; ·; ` 7→1w τ ` ` : Ref w τ
P; ·; ` 7→s τ ` ` : Ref s τ
P; Γ; Σ ` e : τ
P; Γ; Σ ` newrq e : Ref rq τ
R ELEASE W
R ELEASE S
P; Γ; Σ ` releasew e : Unit + τ
P; Γ; Σ ` releases e : τ
P; Γ; Σ ` e : Ref s τ
P; Γ; Σ ` e : Ref rq τ
S WAP W
P; Γ1 ; Σ1 ` e1 : Ref rq τ
P; Γ2 ; Σ2 ` e2 : τ
w
P; Γ1 + Γ2 ; Σ1 + Σ2 ` swap e1 with e2 : Ref rq τ × τ
S WAP S
P; Γ1 ; Σ1 ` e1 : Ref s τ1
P; Γ2 ; Σ2 ` e2 : τ2
s
P; Γ1 + Γ2 ; Σ1 + Σ2 ` swap e1 with e2 : Ref s τ2 × τ1
Figure 7: λcl term reference cell typing rules
typing rules for reference cells are given in figure 7. Consistency conditions Σ1 ^ Σ2 and Γ1 ^ Γ2 are
assumed whenever contexts are combined. The core language typing rules split and share the linear
contexts as needed, but are otherwise a natural extension of System F to support type class constraints.
We impose a syntactic restriction, similar to Haskell 98’s context reduction restrictions [15], on the
form of constraints in type schemes introduced by the TA BS rule: type abstractions may only constrain
the type variables that they bind, and not compound or unrelated types. This simplifies induction over
typing derivations for TA BS since it means that no constraints on external type variables can be introduced by a type abstraction. Additionally, in rule A BS, the variable and location contexts are constrained
by the function’s arrow qualifier, to ensure that values captured by the closure support any structural
operations that might be applied to the closure itself; this constraint must be entailed ( ) by the constraint context. Here Constrainaq (Γ; Σ) is shorthand for the appropriate set of Dup and Drop constraints
applied to every type mapped in Γ and Σ, so that for instance ConstrainL imposes no constraints, while
ConstrainR imposes only Dup constraints.
The dup and drop forms constrain the types of their parameters in the expected way, by requiring
their types to be members of the Dup or Drop type classes, respectively (again entailed by the constraint
context).
Since λcl supports both strong and weak references with different substructural properties, there are
a variety of typing rules governing their usage. The swap operation needs to return both an updated
reference and the old contents, so it packages those in a pair. Weak and strong forms of reference cell
operations are provided, and it is safe to apply the weak operations to both strong and weak references.
The releaserq e forms are used to deallocate a reference cell and possibly retrieve its contents. Notably, in
the case of a weak reference, since the contents could be linear, we preserve its linearity while allowing
for aliasing by returning the contents of the reference only when the last alias to the cell is released, and
unit otherwise.
40
Type Classes for Lightweight Substructural Types
(Dup α1 , Dup α2 ) ⇒ Dup (α1 × α2 )
(Drop α1 , Drop α2 ) ⇒ Drop (α1 × α2 )
(Dup α1 , Dup α2 ) ⇒ Dup (α1 + α2 )
(Drop α1 , Drop α2 ) ⇒ Drop (α1 + α2 )
U
U
R
A
() ⇒ Dup (α1 −→ α2 ) () ⇒ Drop (α1 −→ α2 ) () ⇒ Dup Unit
() ⇒ Drop Unit
() ⇒ Dup (α1 −→ α2 ) () ⇒ Drop (α1 −→ α2 ) () ⇒ Dup (Ref w α) (Drop α) ⇒ Drop(Ref rq α)
Figure 8: Dup and Drop instances
2.4
Type Class Instances
Throughout the type system, type class constraints are propagated via entailment, P1 P2 , which specifies
when one set of type class predicates (P2 ) is implied by another (P1 ) in the context of the fixed background
instance environment Γis . For example, entailment allows our type system to derive that Unit × Unit is
duplicable because Unit is. Rules for entailment are given by Jones [12] and adapt naturally to this
setting. The substructural essence of the type class system in Clamp is the set of base Dup and Drop
instances Γis , which appears in figure 8.
Since pairs and sums contain values that might be copied or ignored along with the pair or sum value,
their instance rules require instances for their components. Functions impose constraints on their closure
environments when they are assigned a qualifier during term typing, so the instance rules for arrows
depend only on the arrow qualifier.
Dealing correctly with references is more subtle, as seen in λ refURAL [2]. In Clamp, some references
support strong updates, which can change not only the value but the type of a mutable reference. However
it is unsafe to alias a reference cell whose type might change.
In λ refURAL , the restrictions on reference types are given in a sizable table, but Dup and Drop instances make it easy to express these restrictions in Clamp. Clamp classifies references by the kind of
updates they support: strong or weak. This is specified by the rq qualifier in the Ref rq type.
Qualitatively, the constraints we impose are that:
• Strong references may not be duplicated.
• Only references with droppable contents may be dropped.
• Only strong references support direct deallocation.
• Weak references can be deallocated, but only return their contents when unaliased.
The above four rules capture the same restrictions as λ refURAL references. They also increase the
expressiveness of the system by explicitly distinguishing weak and strong references and allowing for
the deallocation of weak references. They are expressed in λcl with two type class instances and the
typing judgments R ELEASE W and R ELEASE S.
As an example of the kinds of structures we can build using these rules, consider the type
Ref w (fhandle)
for a linear file handle fhandle. This weak reference can be aliased to provide shared access to the
file handle, but cannot be dropped based on the type class instances, since fhandle cannot be dropped.
Anyone that uses this reference must release the reference and close the file if necessary.
Gan, Tov & Morrisett
2.5
41
Type Safety
Here we sketch part of the type safety proof; more details may be found in Gan’s thesis [9].
The bulk of the work goes into proving preservation, and the key lemma in proving preservation
relates constraints to bindings. Intuitively, this lemma says that structural constraints on a value’s type
respect the structural constraints of everything the value contains or points to, via the variable and location contexts. Syntactic forms like Dup Γ are used to denote the set of Dup constraints on all types in Γ,
and similarly Dup Σ applies Dup to all of the Ref rq τ types mapped by Σ. Note that the lemma does not
hold for arbitrary expressions.
Lemma 1 (Constraints capture bindings). Suppose that P; Γ; Σ ` v : τ. If P
if P Drop τ then P (Drop Σ, Drop Γ).
Dup τ then P
(Dup Σ, Dup Γ);
Proof. By induction on the typing derivation for v.
Lemma 1 is essential to proving the substitution lemma (Lemma 2).
Lemma 2 (Substitution). If
• P; Γ, x:τx ; Σ1 ` e : τ ,
• P; ·; Σ2 ` v : τx , and
• Σ1 ^ Σ2 ,
then P; Γ; Σ1 + Σ2 ` {v/x} e : τ
Proof. By induction on the typing derivation for e, making use of lemma 1 in the λ case.
In proving Preservation, it is also useful to separate out a Replacement Lemma which specifies exactly how substitution interacts with evaluation contexts.
Lemma 3 (Replacement). If P; Γ; Σ ` E [M] : τ then ∃τ 0 , Σ1 , Σ2 , Γ1 , Γ2 such that
• Σ = Σ1 + Σ2 and Γ = Γ1 + Γ2 and P; Γ1 ; Σ1 ` M : τ 0 and furthermore
0
0
0
0
0
0
• If P; Γ1 ; Σ1 ` M 0 : τ 0 with Γ1 ^ Γ2 and Σ1 ^ Σ2 , then P; Γ1 + Γ2 ; Σ1 + Σ2 ` E [M 0 ] : τ for any
0
0
M 0 , Γ1 , Σ1
Proof. By induction on E.
Another key lemma for proving preservation relates the locs function used to maintain dynamic
reference counts with the store context that a value requires. To state this lemma, we overload the locs
function to also return the multiset of occurrences (multiple for reference counted weak location stores)
of locations in the domain of a store context.
Lemma 4 (Store Contexts map Free Locations). If P; Γ; Σ ` e : τ then locs (e) = locs (Σ).
Finally, to prove preservation and type soundness we need to introduce store and configuration typings which are given in Figure 9. The remainder of the type soundness proof is then mostly standard.
Lemma 5 (Preservation). If `c (µ1 ; e1 ) : τ and (µ1 ; e1 ) 7−→ (µ2 ; e2 ) then `c (µ2 ; e2 ) : τ
Theorem 1 (Type soundness). If `c (· ; e) : τ then either it diverges or it reduces to a value configuration
(µ ; v) such that `c (µ ; v) : τ.
42
Type Classes for Lightweight Substructural Types
S T-N IL
S T-C ONS S
S T-C ONS W
Σ1 `s µ : Σ2
Σ `s · : ·
Σ1 `s µ : Σ2
·; ·; Σv ` v : τ
i
Σ1 + Σv `s µ, ` 7→ v : Σ2 , `
7 iw
→
·; ·; Σv ` v : τ
1
Σ1 + Σv `s µ, ` 7→ v : Σ2 , ` 7→s τ
τ
C ONF
Σ1 `s µ : Σ1 + Σ2
·; ·; Σ2 ` e : τ
`c (µ ; e) : τ
Figure 9: Store and configuration typing
3
Implementing the Clamp Type Checker
We have implemented a type checker that infers Damas–Hindley–Milner style type schemes for Clamp
terms. The type checker is an extension of Jones’s “Typing Haskell in Haskell” type checker [13]. Its
source code may be found at https://github.com/edgan8/clampcheck.
The process of modifying a Haskell type checker to support Clamp was straightforward and illustrates
one of the strengths of Clamp’s design: It requires only small and orthogonal additions to a language like
Haskell. Besides adding qualifiers to arrow types, we made three additions to a Haskell type checker:
1. an elaboration pass that inserts dups and drops,
2. Dup and Drop type classes and instances, and
3. substructural qualifiers and constraints on arrow types.
3.1
Inferring dups and drops.
The elaboration pass is the bridge between a concise user-facing language and leveraging conventional,
nonlinear type checking techniques. The pass takes as input a term with arbitrary variable usages; it
inserts the appropriate dup and drop operations and renames the duplicated copies so that in the resulting
term all variable usage is strictly linear. Structural properties are then enforced by the constraints imposed
by dup and drop.
Since different elaborations can lead to different static and dynamic semantics, we have proven that
our algorithm generates an optimal elaboration in two senses:
• It minimizes the program’s live variables.
• It imposes minimal type class constraints.
In what follows we define a core linear language to formalize and estabilsh these two points.
An Abstract Linear Language To focus on the essential problems, we can work with an abstraction
of the linear λ calculus, λ̂lin . By modeling only usage and binding, λ̂lin allows us to focus on inserting
dup and drop operations independently of the particular types and term forms of a language. Its syntax
is given in figure 10. Extending the results to cover other term forms is straightforward.
The product expression e1 ⊗e2 abstracts multiplicative forms such as pairs and function applications—
that is, pairs of expressions where both will be evaluated. The sum expression e1 & e2 abstracts additive
forms such as linear logic’s additive conjuction, and the relationship between the branches of a case—
that is, pairs of expressions where exactly one will be evaluated.
Gan, Tov & Morrisett
43
e ::= x | λ x.e | e1 ⊗ e2 | e1 & e2
ae ::= x | λ x.ae | ae1 ⊗ ae2 | ae1 & ae2 | dup Γ in ae | drop Γ in ae
(unannotated terms)
(annotated terms)
Figure 10: λ̂lin syntax
L-VAR
L-A BS
{x} ` x
L-C HOICE
L-PAIR
Γ + {x} ` ae
x∈
/Γ
Γ ` λ x.ae
Γ1 ` ae1
Γ2 ` ae2
Γ1 + Γ2 ` ae1 ⊗ ae2
L-D UP
Γ ` ae1
Γ ` ae2
Γ ` ae1 & ae2
L-D ROP
Γ1 + Γ2 + Γ2 ` ae
Γ1 ` ae
Γ1 + Γ2 ` dup Γ2 in ae
Γ1 + Γ2 ` drop Γ2 in ae
Figure 11: Annotated expression well-formedness
Expressions, e, are unannotated and don’t explicitly satisfy linear usage constraints. Annotated expression, ae, use dup and drop operations to explicitly specify nonlinear usage of variables. The dup and
drop operations work over contexts, Γ, which are multisets of variables
The contexts Γ in λ̂lin manage scope and binding by restricting contraction and weakening to explicit
dup and drop annotations, but do not track the types of variables. In order to avoid the messy but
straightforward process of generating names and renaming variables when inserting a dup, we think
of contexts as multisets (e.g., {x, x, y, z, . . .}) of in-scope variables, or equivalently as functions from
variables to natural numbers. Thus Γ(x) will be used to denote the number of times x appears in Γ.
We use these multiset contexts to define a notion of well-formedness in figure 11, which describes
when an annotated term ae in λ̂lin properly accounts for all nonlinear usage of its variables through
explicit dup and drop operations.
Inference Algorithm An inference algorithm for annotating terms is given in figure 12. The strategy
is to recursively transform a term bottom-up based on the free variables fv in each recursively transformed sub-term. Note that the fv function always returns a set and that ∩ and \ denote the standard set
intersection and difference operators.
Dup operations are inserted where the free variables of two sub-terms of a multiplicative form (e.g.,
application, but not branching) are discovered to intersect; drops are added under binders when the bound
variable is not free in its scope, and when a variable used in one branch (say, of an if-then-else) is not
free in the other.
We outline the key steps in our optimality argument here; additional details may be found in [9].
Lemma 6 states that the algorithm is sound.
Lemma 6 (Soundness). fv(e) ` infer(e)
If ae is an annotation of e, then fv(e) v fv(ae), so lemma 7 shows that our algorithm in fact generates
an annotation which requires a minimal context Γ for well-formedness.
Lemma 7 (Minimal contexts). If Γ ` ae then fv(ae) v Γ.
With some technical lemmas, we can then prove that the algorithm introduces no unnecessary dups
or drops on variables. In order to compare different potential annotations of the same term, we de-
44
Type Classes for Lightweight Substructural Types
infer (x) = x
(
λ x. infer (e)
infer (λ x.e) =
λ x.drop x in infer (e)
if x ∈ fv(e);
otherwise
infer (e1 ⊗ e2 ) = dup fv(e1 ) ∩ fv(e2 ) in infer(e1 ) ⊗ infer(e2 )
infer (e1 & e2 ) = (drop fv(e2 ) \ fv(e1 ) in infer(e1 )) & (drop fv(e1 ) \ fv(e2 ) in infer(e2 ))
Figure 12: Inference algorithm
fine a function erase : ae → e, which removes dup and drop annotations from an annotated term in the
straightforward way, yielding an unannotated term. It should be evident that erase(infer(e)) = e.
Lemma 8 (Forced drop). If Γ ` ae, Γ (x) ≥ 1, and x ∈
/ fv (erase (ae)), then ae contains a subterm
drop Γ0 in ae0 such that x ∈ Γ0 .
Lemma 9 (Forced dup). If Γ ` ae, Γ (x) ≤ 1, and there exists a subderivation Γs ` aes of Γ ` ae with
Γs (x) ≥ 2, then ae contains a subterm dup Γ0 in ae0 such that x ∈ Γ0
Lemma 10 (No Unnecessary Drops). Let ae = infer (e). If ae contains a subterm drop Γd in aes with
0
0
0
x ∈ Γd then any other well-formed ae0 with erase (ae0 ) = e contains a subterm drop Γd in aes with x ∈ Γd .
Lemma 11 (No Unnecessary Dups). Let ae = infer (e). If ae contains a subterm dup Γd in aes with
0
0
0
0
x ∈ Γd then any other ae0 with erase (ae0 ) = e and Γ ` ae0 for Γ (x) ≤ 1 contains a subterm dup Γd in aes
0
with x ∈ Γd .
3.2
Constraint processing.
After the dup and drop insertion pass, type inference can proceed without needing to count variable
usages or split contexts, since the insertion pass has made every dup and drop operation explicit. With
the exception of the extra constraints imposed on closure environments, inferring types for the Clamp
internal language is like inferring types for Haskell. For the constraint solver, the type classes Dup and
Drop and their instances are no different than any other type class.
Type checking in this system is thus separated into two self-contained steps: first, usage analysis
as performed by elaboration, and second, checking substructural constraints in the same manner as any
other type class system. A similar division was used by de Vries et al. to integrate a uniqueness typing
system into Damas–Hindley–Milner [18].
In Table 1 we present the sizes of the components of our Clamp implementation. It compares favorably to the implementation of languages such as Alms [17], whose type inference engine is 15,000 lines
of Haskell. The dup/drop insertion sits on top of the stack and is the main addition we have had to make
to a Haskell type checker design. Besides that, we have included the base set of type class instances and
altered arrow kinds throughout.
4
Related Work
Of the existing work in linear type systems, we will focus here on those which develop general purpose
polymorphic linear types. Research on the mathematical expressiveness of linearity [3] or more tailored
Gan, Tov & Morrisett
45
Component
Lines of code
Dup/drop insertion
Type class instances
160
60
Syntax and Types
Parser/Lexer
Unification Engine
703
319
373
Table 1: Type checker code breakdown
use cases [10] for instance often do not aim at broad usability and polymorphism.
The first linear type systems derive directly from intuitionistic linear logic, and use the exponential “!” to indicate types that support structural operations [1, 6]. Some later type systems, in order to
support parametric polymorphism over linearity, replace “!” with types composed of a qualifier and a
pretype [2, 20], so that all types in these languages have a form like q τ. Similarly, the Clean programming
language makes use of qualifier variables and inequalities to capture a range of substructural polymorphism [5, 18]. Though Clean uses uniqueness rather than linear types, many of its design decisions can
be applied in a linear settings as well.
More recent languages such as Alms [17] and F ◦ [14] eliminate the notational overhead of annotating
every type with substructural qualifiers by using distinct kinds to separate substructural types. Thus,
rather than working with types like A file, a file type in Alms can be defined to have kind A. Like Clean,
Alms is highly polymorphic, but it makes use of compound qualifier expressions on function types, as
well as dependent kinds and sub-kinding.
Compared to type systems like those of Clean and Alms, we believe Clamp offers advantages in
simplicity and extensibility. Like Alms and F ◦ , Clamp avoids the burden in Clean of annotating every
type with a qualifier. Type classes themselves are a general and powerful feature; for a language that is
going to have type classes anyway, the Clamp approach allows adding the full spectrum of URAL types
with little additional complexity for programmers and type checkers. Programmers already familiar with
type classes will be well prepared to understand Clamp-style substructural types.
Further, type classes provide a clean formalism for constraining state-aware datatypes such as the
system of weak and strong mutable references found in Clamp (§2.4). Finally, we anticipate that userdefined Dup and Drop instances, not yet supported by Clamp, will allow defining custom destructors and
copy constructors, which should enable a variety of resource management strategies.
However, compared to Alms and Clean, Clamp does not provide as much polymorphism because
each arrow is assigned a concrete qualifier. Consider, for instance, a curry function in Clamp. Unlike in
Alms or Clean, Clamp requires different versions for different desired structural properties. For instance,
two possible type schemes for a curry function are
(Dup a , Drop a ) => ((a , b ) -U> c ) -U> a -U> b -U> c
and
((a , b ) -L> c ) -U> a -L> b -L> c.
We believe that extending Clamp with qualifier variables and type class implications could increase its
expressiveness to the point where curry has a principal typing.
46
5
5.1
Type Classes for Lightweight Substructural Types
Future Work
Custom dup and drop
In Clamp’s current design, the semantics of dup and drop are fixed. Allowing programmers instead to
define their own implementations of dup and drop on user-defined types would enable scenarios similar
to those possible in C++ via copy constructors and destructors. Programmers could then define data types
that automatically manage resources in ways that meet particular needs; but unlike in C++, we believe
this could be done without unsafe operations.
Programmers could define types and instances to manage their memory in whatever way is most
appropriate, for instance by choosing between deep-copy and shallow-copy dup operations (or perhaps
some hybrid approach), or between eager and lazy drop operations. For such a system to be practical, it
is important that the dup/drop-insertion algorithm be easy to understand, since the insertion of dups and
drops can affect the dynamic semantics of the language.
5.2
Polymorphic Arrows
In most cases, Clamp allows programmers to define functions that are inherently polymorphic over the
substructural properties of their arguments. In these cases, a function that uses one of its arguments
linearly can accept any type, whether it satisfies Dup or Drop, for that argument. However, this is not the
case for function types, which are annotated with fixed qualifiers determined at the function definition
point.
Alms is able to accommodate more polymorphic arrow types by introducing a subtyping relation on
qualified arrows [17], and it refines the types of arrows further by introducing usage qualifiers that depend
on the substructural properties of type variables. This allows one to write a function whose qualifier is
inferred from the closure environment and inherits any polymorphism present in that environment.
There are many options for increasing the polymorphic expressiveness of Clamp without resorting to
the complexities of subtyping. One could make qualifiers into first class types and allow quantification
over qualifiers in arrow types. This is implemented in an ad-hoc way in our current type checker.
The idea of expanding the language of qualifiers could also be profitable in Clamp. This often means
annotating arrow types with the types of their closure environments, and in a sense assigning them a
closure-converted type. The dup and drop instances for arrows could then use the closure environment
types to determine the arrow type’s substructural properties.
5.3
Implementation
The current implementation of the Clamp type checker reflects λcl , and could benefit from the addition of
some standard language features found in full Haskell. In particular, the addition of algebraic datatypes,
user-defined instance rules, and a module system would allow programmers to define libraries that expose
custom types with varying substructural properties.
It would also be interesting to implement a compiler for Clamp that takes advantage of substructural
properties for reference-counted memory management [7]. Eagerly reusing the storage occupied by
linear and affine values may offer particular performance advantages, and substructural analyses can
enable a variety of other optimizations as well [4, 21].
Gan, Tov & Morrisett
6
47
Conclusion
Clamp introduces techniques that make it easier and more desirable to add substructural types to functional programming languages. The external / internal language distinction gives us both a programmer
friendly syntax and a direct path to type inference. The Dup and Drop classes also support polymorphism over the URAL lattice and can represent state aware types such as strong and weak references.
Type classes are an expressive and well-established language feature, and Clamp shows that they can
serve as a base for substructural types.
References
[1] Samson Abramsky (1993): Computational Interpretations of Linear Logic. Theor. Comput. Sci. 111(1-2),
doi:10.1016/0304-3975(93)90181-R.
[2] Amal Ahmed, Matthew Fluet & Greg Morrisett (2005): A Step-Indexed Model of Substructural
State. In: Proc. 10th ACM SIGPLAN International Conference on Functional Programming (ICFP’05),
doi:10.1145/1086365.1086376.
[3] Sandra Alves, Maribel Fernández, Mário Florido & Ian Mackie (2011): Linearity and Recursion in a
Typed Lambda-calculus. In: Proceedings of the 13th International ACM SIGPLAN Symposium on Principles and Practices of Declarative Programming, PPDP ’11, ACM, New York, NY, USA, pp. 173–182,
doi:10.1145/2003476.2003500.
[4] Henry G. Baker (1994):
doi:10.1145/181748.181750.
A “linear logic” Quicksort.
SIGPLAN Not. 29(2), pp. 13–18,
[5] Erik Barendsen & Sjaak Smetsers (1996): Uniqueness Typing for Functional Languages with Graph Rewriting Semantics. In: Math. Struct. Comp. Sci.
[6] Gavin M. Bierman (1993): On Intuitionistic Linear Logic. Ph.D. thesis, University of Cambridge.
[7] Jawahar Chirimar, Carl A. Gunter & Jon G. Riecke (1996): Reference Counting as a Computational Interpretation of Linear Logic. Journal of Functional Programming 6, pp. 6–2, doi:10.1017/S0956796800001660.
[8] Luis Damas & Robin Milner (1982): Principal Type-Schemes for Functional Programs. In: Proc. 9th Annual
ACM Symposium on Principles of Programming Languages (POPL’82), doi:10.1145/582153.582176.
[9] Edward Gan (2013): Clamp: Type Classes for Substructural Types. Senior Thesis, Harvard University.
Available at http://edgan8.github.io/.
[10] Simon J. Gay & Vasco T. Vasconcelos (2010): Linear type theory for asynchronous session types. Journal of
Functional Programming 20, pp. 19–50, doi:10.1017/S0956796809990268.
[11] Jean-Yves Girard (1972): Interprétation fonctionnelle et Elimination des coupures de l’arithmétique d’ordre
supérieur. These d’état, Université de Paris 7, doi:10.1016/0304-3975(87)90045-4.
[12] Mark P. Jones (1995): Qualified Types: Theory and Practice. Cambridge University Press, New York.
[13] Mark P. Jones (1999): Typing Haskell in Haskell. In: Proc. 1999 Haskell Workshop.
[14] Karl Mazurak, Jianzhou Zhao & Steve Zdancewic (2010): Lightweight Linear Types in System F◦ .
In: Proc. 5th ACM SIGPLAN Workshop on Types in Language Design and Implementation,
doi:10.1145/1708016.1708027.
[15] Simon Peyton Jones & John Hughes, ed. (1999): Haskell 98: A Non-Strict, Purely Functional Language.
[16] Jesse A. Tov (2012): Practical Programming with Substructural Types. Ph.D. thesis, Northeastern University.
[17] Jesse A. Tov & Riccardo Pucella (2011): Practical Affine Types. In: Proc. 38th Annual ACM SIGPLANSIGACT Symposium on Principles of Programming Languages (POPL’11), doi:10.1145/1926385.1926436.
48
Type Classes for Lightweight Substructural Types
[18] Edsko Vries, Rinus Plasmeijer & David M. Abrahamson (2008): Implementation and Application of Functional Languages. chapter Uniqueness Typing Simplified, Springer-Verlag, Berlin, Heidelberg, pp. 201–218,
doi:10.1007/978-3-540-85373-2 12.
[19] Philip Wadler & Stephen Blott (1989): How to Make Ad-Hoc Polymorphism Less Ad Hoc. In:
Proc. 16th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL’89),
doi:10.1145/75277.75283.
[20] David Walker (2005): Advanced Topics in Types and Programming Languages, chapter 1. MIT Press, Cambridge, Mass., U.S.A.
[21] Keith Wansbrough & Simon Peyton Jones (1999): Once upon a Polymorphic Type. In: Proc. 26th ACM
SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL’99), ACM, pp. 15–28,
doi:10.1145/292540.292545.
| 6cs.PL
|
SYMBOLIC MODELS FOR RETARDED JUMP-DIFFUSION SYSTEMS
arXiv:1704.03690v3 [math.OC] 25 Jan 2018
PUSHPAK JAGTAP AND MAJID ZAMANI
Abstract. In this paper, we provide for the first time an automated, correct-by-construction, controller synthesis scheme for a class of infinite dimensional stochastic systems, namely, retarded jump-diffusion systems.
First, we construct finite dimensional abstractions approximately bisimilar to original retarded jump-diffusion
systems having some stability property, namely, incremental input-to-state stability. Second, we construct
finite abstractions approximately bisimilar to constructed finite dimensional abstractions. Both types of abstractions are derived without any state-space discretization. By using the transitivity property of approximate
bisimulation relations, we establish that the constructed finite abstractions are also approximately bisimilar
to original retarded jump-diffusion systems with a precision that can be chosen a-priori. Given those finite
abstractions, one can synthesize controllers for original systems satisfying high-level logic requirements in a
systematic way. Moreover, we provide sufficient conditions for the proposed notion of incremental stability in
terms of the existence of incremental Lyapunov functions which reduce to linear matrix inequalities (LMI) for
the linear systems. Finally, the effectiveness of the results is illustrated by synthesizing a controller regulating
the temperatures in a ten-room building modeled as a delayed jump-diffusion system.
1. Introduction
Finite (a.k.a. symbolic) abstraction techniques have gained significant attention in the last few years since
they provide tools for automated, correct-by-construction, controller synthesis for several classes of control
systems. In particular, such abstractions provide approximate models that are related to concrete systems by
aggregating concrete states and inputs to the symbolic ones. Having such finite abstractions, one can make use
of the existing automata-theoretic techniques [MPS95] to synthesize hybrid controllers enforcing rich complex
specifications (usually expressed as linear temporal logic formulae or automata on infinite strings) over the
original systems.
In the past few years, there have been several results providing bisimilar finite abstractions for various
continuous-time non-probabilistic as well as stochastic systems. The results include construction of approximately bisimilar abstractions for incrementally stable control systems [PGT08], switched systems [GPT10],
stochastic control systems [ZMEM+ 14], and randomly switched stochastic systems [ZA14]. However, the abstractions obtained in these results are based on state-space quantization which suffer severely from the curse
of dimensionality, i.e., the computational complexity increases exponentially with respect to the state-space
dimension of the concrete system.
To alleviate this issue, authors in [CGG13] proposed an alternative approach for constructing approximately
bisimilar abstractions for incrementally stable non-probabilistic switched systems without discretizing the
state-space. The concept is further extended to provide finite abstractions for incrementally stable stochastic switched systems [ZAG15], stochastic control systems [ZTA17], and infinite dimensional non-probabilistic
control systems [Gir14]. For a comparison between state-space discretization based and free approaches, we
refer the interested readers to the discussion in Section 5.4 in [ZTA17].
On the other hand, retarded stochastic systems are widely used to model various processes in finance, ecology,
medical, and engineering (see, e.g., [Sha13, BSP08, KM13]). However, the construction of symbolic models for
such classes of systems are still unaddressed due to underlying challenges such as infinite-dimensional functional
state-space and dependency on state history. The authors in [PPDBT10, PPD15] provide the construction of
abstractions for incrementally stable non-probabilistic time-delayed systems by spline-based approximation of
functional spaces. However, the proposed results are complex from the implementation point of view and also
suffer from the curse of dimensionality with respect to the state-space dimension of the concrete system. This
1
2
P. JAGTAP AND M. ZAMANI
motivates our work in this paper to provide a scheme for the construction of approximately bisimilar finite
abstractions for a class of infinite dimensional stochastic systems, namely, retarded jump-diffusion systems
without discretizing the state-space.
The main contribution of this paper is twofold. First, we introduce a notion of incremental input-to-state
stability for retarded jump-diffusion systems and provide sufficient conditions for it in terms of the existence
of a notion of incremental Lyapunov functions. In the linear case, we show that the sufficient conditions reduce to a linear matrix inequality (LMI). Second, under the assumption of incremental stability property, we
provide finite dimensional abstractions which are approximately bisimilar to the original infinite dimensional
stochastic systems. Then, we provide approximately bisimilar finite abstractions for the constructed finite dimensional abstractions. Further, with the help of transitivity property of approximate bisimulation relations,
we show that the obtained finite abstractions are approximately bisimilar to the concrete infinite dimensional
systems with a precision defined a-priori. Further, we demonstrate the effectiveness of the proposed results by
synthesizing a controller keeping temperatures in a comfort zone in a ten-room building modeled as a linear
delayed jump-diffusion system.
2. Retarded Jump-Diffusion Systems
2.1. Notations. Let the triplet (Ω, F , P) denote a probability space with a sample space Ω, filtration F , and
the probability measure P. The filtration F = (Fs )s≥0 satisfies the usual conditions of right continuity and
completeness [ØS05]. The symbols N, N0 , Z, R, R+ , and R+
0 denote the set of natural, nonnegative integer,
integer, real, positive, and nonnegative real numbers, respectively. We use Rn×m to denote a vector space of
real matrices with n rows and m columns. Symbol ei ∈ Rn denotes a vector whose all elements are zero, except
the ith one, which is one. Let the family of continuous functions ζ : [−τ, 0] → Rn be denoted by C([−τ, 0]; Rn )
with a norm kζk[−τ,0] := sup−τ ≤θ≤0 kζ(θ)k, where k · k denotes the Euclidean norm in Rn . We denote by ζ ≡ c
k
n
a constant function ζ ∈ C([−τ, 0]; Rn ) with value c ∈ Rn . For k > 0 and t ∈ R+
0 , LFt ([−τ, 0]; R ) denotes
n
the family of all Ft -measurable C([−τ, 0]; R )-valued random processes φ := {φ(θ) | −τ ≤ θ ≤ 0} such that
b
sup−τ ≤θ≤0 E[kφ(θ)kk ] < ∞. The notation CF
([−τ, 0]; Rn ) denotes the family of all bounded F0 -measurable
0
n
C([−τ, 0]; R )-valued random variables. For a matrix A ∈ Rn×m , kAk represents the Euclidean norm of A.
We use λmin (A) and λmax (A) to denote the minimum and maximum eigenvalue of a symmetric matrix A,
respectively. The diagonal set △ ⊂ R2n is defined as △ = {(x, x)|x ∈ Rn }.
The closed ball centered at x ∈
Rm with radius R is defined by BR (x) = {y ∈ Rm | kx − yk ≤ R}. A set
Qm
m
B ⊆ R is called a box if B = i=1 [ci , di ], where ci , di ∈ R with ci < di for each i ∈ {1, . . . , m}. The span of
a box B is defined as span(B) = min{|di − ci | | i = 1, . . . , m}, where |a| represents the absolute value of a ∈ R.
S
ki η
, ki ∈ Z}, the set b∈[Rm ]η BR (b) is a countable covering of Rm for any
By defining [Rm ]η := {z ∈ Rm | zi = √
m
η ∈ R+ and R ≥ η/2. For a box B ⊆ Rm and η ≤ span(B), define the η-approximation [B]η := [Rm ]η ∩B. Note
that [B]η 6= ∅ for any η ≤ span(B). We extend the notions of span and of approximation to finite unions of
SM
boxes as follows. Let A = j=1 Aj , where each Aj is a box. Define span(A) = min{span(Aj ) | j = 1, . . . , M },
SM
and for any η ≤ span(A), define [A]η = j=1 [Aj ]η .
+
A continuous function γ : R+
0 → R0 belongs to class K if it is strictly increasing and γ(0) = 0; it belongs to
+
+
class K∞ if γ ∈ K and γ(r) → ∞ as r → ∞. A continuous function β : R+
0 × R0 → R0 belongs to class KL if
for each fixed s, the map β(r, s) belongs to class K with respect to r and, for each fixed r 6= 0, the map β(r, s)
n
is decreasing with respect to s and β(r, s) → 0 as s → ∞. Given a measurable function f : R+
0 → R , the
(essential) supremum of f is denoted by kf k∞ ; we recall that kf k∞ := (ess)sup{kf (t)k, t ≥ 0}. We identify a
relation R ⊆ A × B with the map R : A → 2B defined by b ∈ R(a) if and only if (a, b) ∈ R.
2.2. Retarded Jump-Diffusion Systems. Let (Ws )s≥0 be an r-dimensional F-Brownian motion and (Ps )s≥0
be an r̃-dimensional F-Poisson process. We assume that the Poisson process and the Brownian motion are
independent of each other. The Poisson process Ps := [Ps1 ; . . . ; Psr̃ ] models r̃ kinds of events whose occurrences
are assumed to be independent of each other.
SYMBOLIC MODELS FOR RETARDED JUMP-DIFFUSION SYSTEMS
3
Definition 2.1. A retarded jump-diffusion system (RJDS) is a tuple ΣR = (Rn , X , U, U, f, g, r), where:
Rn is the Euclidean space;
X is a subset of C([−τ, 0]; Rn ), for some τ ∈ R+
0;
U ⊆ Rm is the bounded input set which is finite unions of boxes;
U is a subset of the set of all measurable, locally essentially bounded functions of time from R+
0 to U;
f : X × U → Rn , satisfies the following Lipschitz assumption: there exist constants Lf , Lu ∈ R+ , such
that kf (xt , u) − f (x̂t , û)k ≤ Lf kxt − x̂t k[−τ,0] + Lu ku − ûk for all xt , x̂t ∈ X and all u, û ∈ U;
• g : X → Rn×r satisfies the following Lipschitz assumption: there exists a constant Lg ∈ R+
0 such that
kg(xt ) − g(x̂t )k ≤ Lg kxt − x̂t k[−τ,0] for all xt , x̂t ∈ X ;
• r : X → Rn×r̃ satisfies the following Lipschitz assumption: there exists a constant Lr ∈ R+
0 such that
kr(xt ) − r(x̂t )k ≤ Lr kxt − x̂t k[−τ,0] for all xt , x̂t ∈ X .
•
•
•
•
•
An Rn -valued continuous-time process ξ is said to be a solution process for ΣR if there exists υ ∈ U satisfying
(2.1)
d ξ(t) = f (ξt , υ(t)) d t + g(ξt ) d Wt + r(ξt ) d Pt ,
P-almost surely (P-a.s.), where f , g, and r are the drift, diffusion, and reset terms, respectively, and ξt := {ξ(t+
θ)|−τ ≤ θ ≤ 0}. We emphasize that postulated assumptions on f , g, and r ensure the existence and uniqueness
of the solution process ξ on t ≥ −τ [ØS05, Theorem 1.19]. Throughout the paper we use the notation ξζ,υ (t) to
b
denote the value of a solution process starting from initial condition ζ = {ξ(θ)| − τ ≤ θ ≤ 0} ∈ CF
([−τ, 0]; Rn )
0
P-a.s. and under the input signal υ at time t. We also use the notation ξt,ζ,υ to denote the solution process
b
starting from initial condition ζ = {ξ(θ)| − τ ≤ θ ≤ 0} ∈ CF
([−τ, 0]; Rn ) P-a.s. and under the input signal
0
+
υ. Note that for any t ∈ R0 , ξζ,υ (t) is a random variable taking values in Rn and ξt,ζ,υ is a random variable
taking values in C([−τ, 0]; Rn ). Here, we assume that the Poisson processes Psi , for any i ∈ {1, 2, . . . , r̃}, have
the rates of λi . Now we will introduce delayed jump-diffusion system (ΣD )(DJDS) as a special case of retarded
jump-diffusion system which is given by
(2.2)
d ξ(t)=F (ξ(t), ξ(t − τ1 ), υ(t)) d t+G(ξ(t), ξ(t − τ2 )) d Wt + R(ξ(t), ξ(t − τ3 )) d Pt ,
where F : Rn × Rn × U → Rn , G : Rn × Rn → Rn×r , and R : Rn × Rn → Rn×r̃ are the drift, diffusion, and
reset terms, respectively. The constants τ1 , τ2 , and τ3 are the state delay in the drift, diffusion, and reset
terms, respectively.
2.3. Incremental Stability for RJDS and DJDS. Here, we introduce a notion of incremental stability
for RJDS (resp. DJDS).
Definition 2.2. An RJDS ΣR (resp. DJDS ΣD ) is incrementally input-to-state stable in the k th moment,
where k ≥ 1, denoted by (δ-ISS-Mk ), if there exist a KL function β and a K∞ function γ such that for any
b
n
t ∈ R+
0 , any two initial conditions ζ, ζ̂ ∈ CF0 ([−τ, 0]; R ), and any υ, υ̂ ∈ U the following condition is satisfied:
(2.3)
E[kξζ,υ (t)−ξζ̂,υ̂ (t)kk ] ≤β(E[kζ − ζ̂kk[−τ,0]], t)+ γ(kυ − υ̂k∞ ).
One can readily verify that in the absence of delay, Definition 2.2 reduces to Definition 3.1 in [ZMEM+ 14].
Although the left hand side of condition (2.3) is based on the Euclidean norm in Rn , one can also derive a
similar property using the functional norm over C([−τ, 0]; Rn ) as shown in the following technical lemma.
Lemma 2.3. Consider a δ-ISS-Mk retarded jump-diffusion system ΣR (resp. DJDS ΣD ). Then for any
b
n
t ∈ R+
0 , any two initial conditions ζ, ζ̂ ∈ CF0 ([−τ, 0]; R ), and any υ, υ̂ ∈ U, there exist a KL function β̃ and
a K∞ function γ such that the following inequality holds:
(2.4)
E[kξt,ζ,υ −ξt,ζ̂,υ̂ kk[−τ,0] ]≤ β̃(E[kζ − ζ̂kk[−τ,0]], t)+γ(kυ − υ̂k∞ ),
where β̃(s, t) = e−(t−τ ) s + β(s, max{0, t − τ }) and β and γ are the functions appearing in (2.3).
4
P. JAGTAP AND M. ZAMANI
Proof. From inequality (2.3), we obtain the following inequalities:
(2.5)
E[kξt,ζ,υ − ξt,ζ̂,υ̂ kk[−τ,0] ] ≤ β(E[kζ − ζ̂kk[−τ,0] ], t − τ ) + γ(kυ − υ̂k∞ ), ∀t ≥ τ,
(2.6)
E[kξt,ζ,υ − ξt,ζ̂,υ̂ kk[−τ,0] ] ≤ E[kζ − ζ̂kk[−τ,0] ]+β(E[kζ − ζ̂kk[−τ,0] ], 0) + γ(kυ − υ̂k∞ ), ∀t ∈ [0, τ ).
Moreover, we also have
e−(t−τ ) kζ − ζ̂kk[−τ,0] ≥ kζ − ζ̂kk[−τ,0] ,
(2.7)
∀t ∈ [0, τ ).
The inequalities (2.5) and (2.6) along with (2.7) yield
E[kξt,ζ,υ − ξt,ζ̂,υ̂ kk[−τ,0] ] ≤ e−(t−τ ) E[kζ − ζ̂kk[−τ,0] ] + β(E[kζ − ζ̂kk[−τ,0] ], max{0, t − τ })+γ(kυ − υ̂k∞ ), ∀t ≥ 0.
One can further represent this inequality as
E[kξt,ζ,υ − ξt,ζ̂,υ̂ kk[−τ,0] ] ≤ β̃(E[kζ − ζ̂kk[−τ,0] ], t)+γ(kυ − υ̂k∞ ),
for all t ≥ 0, where β̃(s, t) := e−(t−τ ) s + β(s, max{0, t − τ }) is a KL function.
For later use, we provide the infinitesimal generators (denoted by operator L) for an RJDS ΣR and a DJDS
ΣD using Itô’s differentiation [JP09, equation (23)]. Let function V : Rn × Rn → R+
0 be twice differentiable on
Rn × Rn \ △. The infinitesimal generator of V associated with an RJDS ΣR in (2.1) is an operator, denoted
n
by LV , from C([−τ, 0]; Rn ) × C([−τ, 0]; Rn) to R, and ∀t ∈ R+
0 , ∀xt , x̂t ∈ C([−τ, 0]; R ) and ∀u, û ∈ U it is given
by
f (xt , u)
∂ V ∂x,x̂ V
1
g(xt ) T
g (xt ) g T (x̂t ) x,x
LV (xt , x̂t ) := ∂x V ∂x̂ V
+ Tr
f (x̂t , û)
g(x̂t )
∂x̂,x V ∂x̂,x̂ V
2
(2.8)
+
r̃
X
i=1
λi V (xt (0) + r(xt )ei , xˆt (0) + r(x̂t )ei ) − V (xt (0), xˆt (0)) .
The infinitesimal generator of V associated with a DJDS ΣD in (2.2) is an operator, denoted by LV , from
R8n to R and ∀x, x̂, y, ŷ, z, ẑ, p, p̂ ∈ Rn , and ∀u, û ∈ U it is given by
F (x, y, u)
∂ V ∂x,x̂ V
1
G(x, z) T
G (x, z) GT (x̂, ẑ) x,x
+ Tr
LV (x, x̂, y, ŷ, z, ẑ, p, p̂) := ∂x V ∂x̂ V
F (x̂, ŷ, û)
G(x̂, ẑ)
∂x̂,x V ∂x̂,x̂ V
2
(2.9)
+
r̃
X
i=1
λi V (x + R(x, p)ei , x̂ + R(x̂, p̂)ei ) − V (x, x̂) .
The symbols ∂x and ∂x,x̂ in (2.8) and (2.9) represent first and second-order partial derivatives with respect
to x (1st argument) and x̂ (2nd argument), respectively. Note that we dropped the arguments of ∂x V , ∂x̂ V ,
∂x,x V , ∂x̂,x V , ∂x,x̂ V , and ∂x̂,x̂ V in (2.8) and (2.9) for the sake of simplicity.
Now we describe δ-ISS-Mk in terms of existence of so-called δ-ISS-Mk Lyapunov functions for RJDS and DJDS
using Razumikhin-type condition as defined next.
Definition 2.4. Consider an RJDS ΣR and a continuous function V : Rn × Rn → R+
0 that is twice differentiable on Rn × Rn \ △. The function V is called a δ-ISS-Mk Lyapunov function for ΣR for k ≥ 1, if there
exist K∞ functions α, α, and ϕ, such that:
(i) α(resp. α) is a convex (resp. concave) function;
(ii) ∀x, x̂ ∈ Rn , α(kx − x̂kk ) ≤ V (x, x̂) ≤ α(kx − x̂kk );
(iii) ∀u, û ∈ U and ∀t ≥ 0,
(2.10)
E[LV (φ, φ̂)] ≤ −E[ω(φ(0), φ̂(0))] + ϕ(ku − ûk),
for all φ, φ̂ ∈ LkFt ([−τ, 0]; Rn ) satisfying
(2.11)
E[V (φ(θ), φ̂(θ))] ≤ E[q(φ(0), φ̂(0))], ∀θ ∈ [−τ, 0];
SYMBOLIC MODELS FOR RETARDED JUMP-DIFFUSION SYSTEMS
5
where ω : Rn × Rn → R+ is a nonnegative function such that there exists a K∞ function ω̃ satisfying ω(x, x̂) ≥
k
)
> 0; q : Rn × Rn → R is a function such that q(x, x̂) − V (x, x̂) ≥ q(kx − x̂k),
ω̃(kx − x̂kk ) and limksk→∞ ω̃(ksk
α(kskk )
where q is a K∞ function satisfying limksk→∞
q(ksk)
α(kskk )
> 0.
Definition 2.5. Consider a DJDS ΣD and a continuous function V : Rn ×Rn → R+
0 that is twice differentiable
on Rn × Rn \ △. Function V is called a δ-ISS-Mk Lyapunov function for ΣD for k ≥ 1, if there exist constants
P
κ0 , κ1 , κ2 , κ3 such that κ0 ≥ 3i=1 κi ≥ 0, a nonnegative function ψ : Rn × Rn → R+ , K∞ functions α, α, ϕ,
and K function ω̂ such that: conditions (i) and (ii) in Definition 2.4 hold and ∀x, x̂, y, ŷ, z, ẑ, p, p̂ ∈ Rn and
∀u, û ∈ U,
LV (x, x̂, y, ŷ, z, ẑ, p, p̂) ≤ − κ0 V (x, x̂) − ψ(x, x̂) + κ1 V (y, ŷ) + κ2 V (z, ẑ) + κ3 V (p, p̂) + ϕ(ku − ûk),
and ψ(x, x̂) ≥ ω̂(kx − x̂kk ) and limksk→∞
ω̂(kskk )
α(kskk )
> 0.
Now we provide the description of δ-ISS-Mk for an RJDS ΣR in terms of existence of δ-ISS-Mk Lyapunov
functions in the following theorem.
Theorem 2.6. An RJDS ΣR is δ-ISS-Mk if it admits a δ-ISS-Mk Lyapunov function as in Definition 2.4.
Proof. The proof is inspired by the proof of Theorem 3.1 in [HM09]. Denote ϕ = ϕ(kυ − υ̂k∞ ) and V 0 =
b
([−τ, 0]; Rn ). By using Lemma 3.2 and 3.3 in [HM09], there exist a
α(E[kζ − ζ̂kk[−τ,0]]), ∀υ, υ̂ ∈ U and ∀ζ, ζ̂ ∈ CF
0
constant aq > 0 and a K∞ function µω such that ∀t ≥ 0, E[ω(ξζ,υ (t), ξζ̂,υ̂ (t))] ≥ 2ϕ and E[q(ξζ,υ (t), ξζ̂,υ̂ (t))] −
E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≥ aq , whenever E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≥ µ−1
ω (2ϕ). Without loss of generality, assume
k
µ−1
ω (2ϕ) < α(sup−τ ≤θ≤0 E[kζ(θ) − ζ̂(θ)k )] ≤ V 0 . Let J be the minimal nonnegative integer such that
−1
M 0 = µω
(2ϕ) + Jaq > V 0 . Let τ̂ = max{τ, M0 /ϕ} and tj = j τ̂ for j ∈ {0, 1, . . . , J}. In order to prove the
theorem, we need to show
(2.12)
E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≤ min{V 0 , Mj }, ∀t ≥ tj ,
where Mj = µ−1
ω (2ϕ) + (J − j)aq and j ∈ {0, 1, . . . , J}. First we show that E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≤ V 0 , ∀t ≥ t0 .
Suppose that ta := inf{t > t0 | E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] > V 0 } < ∞. Since E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] is continuous in
time t ≥ 0, there exist a pair of constants tb and tc such that t0 ≤ tb ≤ ta < tc and
(2.13)
t = tb ;
E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] = V0 ,
V0 < E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] < V0 + aq , tb < t ≤ tc .
However, by generalized Itô’s formula [Sko09] and condition (2.10) in Definition 2.4, we have
Z t
E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] = E[V (ξζ,υ (tb ), ξζ̂,υ̂ (tb ))] +
E[LV (ξs,ζ,υ , ξs,ζ̂,υ̂ )] d s ≤ V0 − ϕ(t − tb ) ≤ V0
tb
fot all t ∈ (tb , tc ], which contradicts (2.13). Thus the inequality E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≤ V 0 must be true for all
t ≥ t0 . Now we show that E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≤ M1 , ∀t ≥ t1 . Let tm := inf{t ≥ t0 | E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≤
M1 } < ∞. If tm > t1 , then ∀t ∈ [t0 , t1 ], we have
(2.14)
E[q(ξζ,υ (t), ξζ̂,υ̂ (t))]≥E[V (ξζ,υ (t), ξζ̂,υ̂ (t))]+aq >M1 +aq >V 0 ≥E[V (ξζ,υ (t + θ), ξζ̂,υ̂ (t + θ))], ∀θ ∈ [−τ, 0].
Using condition (2.10) in Definition 2.4, inequality (2.14) implies
E[LV (ξt,ζ,υ , ξt,ζ̂,υ̂ )] ≤ −ϕ, ∀t ∈ [t0 , t1 ].
Consequently, by generalized Itô’s formula, we have E[V (ξζ,υ (t),ξζ̂,υ̂ (t))] ≤ V 0 − ϕτ̂ < 0, which contradicts
the property of E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≥ 0, ∀t ≥ 0. Hence, we must have tm ≤ t1 . Let
t̄a := inf{t > tm | E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] > M1 } < ∞.
6
P. JAGTAP AND M. ZAMANI
Again as E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] is continuous in t ≥ 0, there exists constants t̄b and t̄c such that t1 ≤ t̄b ≤ t̄a < t̄c
and
E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] = M1 ,
t = t̄b ;
M1 < E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] < M1 + aq , t̄b < t ≤ t̄c .
By using similar reasoning as before, generalized Itô’s formula [Sko09] and condition (2.10) in Definition
2.4, the assumption results in contradiction, thus we have (2.12) for j = 1. Now define tj := inf{t ≥ tj−1 |
E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≤ Mj } < ∞ for j = 2, 3, . . . , J. By similar type of reasoning, we get E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≤
Mj , ∀t ≥ tj . Particularly, E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≤ MJ = µ−1
ω (2ϕ), ∀t ≥ tJ . By following Jensen’s inequality,
one obtains
E[kξζ,υ (t) − ξζ̂,υ̂ (t)kk ] ≤ γ(kυ − υ̂k∞ ), ∀t ≥ tJ ,
(2.15)
+
where γ(s) = α−1 (µ−1
ω (2ϕ(s))) for all s ∈ R0 . Now choose a KL function β such that β(V 0 , t) ≥ 2V 0 −
∀t ∈ [0, tJ ]. So we have E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≤ β(V 0 , t), ∀t ∈ [0, tJ ] which implies
V0
tJ t,
E[kξζ,υ (t) − ξζ̂,υ̂ (t)kk ] ≤ β(E[kζ − ζ̂kk[−τ,0] ], t), ∀t ∈ [0, tJ ],
(2.16)
where β(s, t) = α−1 (β(α(s), t)) for any s, t ∈ R+
0 . From (2.15) and (2.16), one can readily verify inequality
(2.3) which implies that ΣR is δ-ISS-Mk .
The next corollary proposes similar results as in the previous theorem but for DJDS.
Corollary 2.7. A DJDS ΣD is δ-ISS-Mk if it admits a δ-ISS-Mk Lyapunov function as in Definition 2.5.
Proof. Let ω(x, x̂) =
1
1+κ0 ψ(x, x̂)
for all x, x̂ ∈ Rn . Now by considering Definition 2.5, we have
E[LV (φ, φ̂)]=E[LV (φ(0), φ̂(0), φ(−τ1 ), φ̂(−τ1 ), φ(−τ2 ), φ̂(−τ2 ), φ(−τ3 ), φ̂(−τ3 ))]
≤ − κ0 E[V (φ(0), φ̂(0))] − E[ψ(φ(0), φ̂(0))] + κ1 E[V (φ(−τ1 ), φ̂(−τ1 ))] + κ2 E[V (φ(−τ2 ), φ̂(−τ2 ))]
+ κ3 E[V (φ(−τ3 ), φ̂(−τ3 ))] + ϕ(ku − ûk)
≤ − κ0 E[V (φ(0), φ̂(0))] + E[ω(φ(0), φ̂(0))] + κ1 E[V (φ(−τ1 ), φ̂(−τ1 ))] + κ2 E[V (φ(−τ2 ), φ̂(−τ2 ))]
+ κ3 E[V (φ(−τ3 ), φ̂(−τ3 ))] − E[ω(φ(0), φ̂(0))] + ϕ(ku − ûk)
3
X
κi ) E[V (φ(0), φ̂(0))] + E[ω(φ(0), φ̂(0))] − E[ω(φ(0), φ̂(0))] + ϕ(ku − ûk)
≤ − (κ0 −
i=1
≤ − E[ω(φ(0), φ̂(0))] + ϕ(ku − ûk)
for all t ≥ 0 and φ, φ̂ ∈ LkFt ([−τ, 0]; Rn ) satisfying condition (2.11) in Definition 2.4 with function q(φ(0), φ̂(0)) :=
ω̂(s)
V (φ(0), φ̂(0)) + ω(φ(0), φ̂(0)). Moreover, functions ω̃(s) = q(s) = 1+κ
, ∀s ∈ R+
0 , satisfy properties required in
0
condition (iii) in Definition 2.4. Therefore, V satisfies all the conditions in Definition 2.4. Thus by following
Theorem 2.6, we obtain that ΣD is δ-ISS-Mk .
In the following lemma, we provide a similar result as in Corollary 2.7 but tailored to linear delayed jumpdiffusion systems in which sufficient conditions boil down to an LMI.
Lemma 2.8. Consider a DJDS ΣD as given in (2.2), where for all x, y, z, p ∈ Rn and u ∈ U, F (x, y, u) :=
A1 x + A2 y + Bu, for some A1 , A2 ∈ Rn×n and B ∈ Rn×m , G(x, z) := [G1 x G2 x · · · Gr x] + [G1 z G2 z · · ·
SYMBOLIC MODELS FOR RETARDED JUMP-DIFFUSION SYSTEMS
7
Gr z] and R(x, p) := [R1 x R2 x · · · Rr̃ x] + [R1 p R2 p · · · Rr̃ p], for some Gi , Gi , Ri , Ri ∈ Rn×n . Then, system
P4
ΣD is δ-ISS-M2 if there exist constants c1 , c2 , c3 , c4 , c5 ∈ R+ satisfying c1 > i=2 ci and
(2.17)
∆
P A2
T
A2 P
0
r̄
P
T
Gi P Gi
0
i=1
r̃
P
λi (RT P +RT P Ri )
0
i
i
i=1
BT P
r̄
P
i=1
r̄
P
i=1
r̃
P
GTi P Gi
i=1
0
λi (P Ri +RiT P Ri )
0
T
Gi P Gi
0
r̃
P
0
i=1
0
0
T
λi Ri P Ri
0
P B
−c1 P
0
0
0
0
0
0
c2 P
0
0
0
0 0
0 c3 P
0
0
0
0
0 c4 P
0
0
0
0
0 c5 Im
0
0
where P is a symmetric positive definite matrix and ∆ = P A1 + AT1 P +
RiT P Ri ).
r̄
P
i=1
GTi P Gi +
r̃
P
i=1
λi (P Ri + RiT P +
Proof. Consider a function V : Rn × Rn → R+
0 given by
V (x, x̂) :=
(2.18)
1
(x − x̂)T P (x − x̂),
2
∀x, x̂ ∈ Rn ,
where P is a symmetric positive definite matrix. One can readily verify that the function V in (2.18) satisfies
properties (i) and (ii) in Definition 2.4 with functions α(s) := 12 λmin (P )s and α(s) := 21 λmax (P )s for all s ∈ R+
0
and k = 2. By considering the infinitesimal generator in (2.9) associated with the considered linear delayed
jump-diffusion system, Lipschitz assumptions, Young’s inequality, consistency of norms, and (2.17) one can
obtain the following chains of inequalities
LV (x, x̂, y, ŷ, z, ẑ, p, p̂) = (x − x̂)T P (A1 (x − x̂) + A2 (y − ŷ) + B(u − û))
r
r̃
T
1X
T
h
1 X
Gi (x − x̂) + Gi (z − ẑ) P Gi (x − x̂) + Gi (z − ẑ) +
λi (x − x̂) + Ri (x − x̂) + Ri (p − p̂)
2 i=1
2 i=1
i
P (x − x̂) + Ri (x − x̂) + Ri (p − p̂) − (x − x̂)T P (x − x̂)
+
r
≤(x − x̂)T P (A1 (x − x̂) + A2 (y − ŷ) + B(u − û)) +
1 Xh
(x − x̂)T GTi P Gi (x − x̂) + (x − x̂)T GTi P Gi (z − ẑ)
2 i=1
r̃
h
i 1X
λi (x − x̂)T P (Ri (x − x̂) + Ri (p − p̂))
+ (z − ẑ)
− x̂) + (z − ẑ)
− ẑ) +
2 i=1
i
+ (Ri (x − x̂) + Ri (p − p̂))T P (x − x̂) + (Ri (x − x̂) + Ri (p − p̂))T P (Ri (x − x̂) + Ri (p − p̂))
T
T
Gi P Gi (x
T
T
Gi P Gi (z
r
≤(x − x̂)T P (A1 (x − x̂) + A2 (y − ŷ) + B(u − û)) +
T
+ (z − ẑ)
T
Gi P Gi (x
T
− x̂) + (z − ẑ)
T
T
Gi P Gi (z
1 Xh
(x − x̂)T GTi P Gi (x − x̂) + (x − x̂)T GTi P Gi (z − ẑ)
2 i=1
r̃
h
i 1X
λi (x − x̂)T P (Ri (x − x̂) + Ri (p − p̂))
− ẑ) +
2 i=1
+ ((x − x̂)T RiT + (p − p̂)T Ri )P (x − x̂) + (x − x̂)T RiT P Ri (x − x̂) + (x − x̂)T RiT P Ri (p − p̂)
i
T
T
+ (p − p̂)T Ri P Ri (x − x̂) + (p − p̂)T Ri P Ri (p − p̂))
8
P. JAGTAP AND M. ZAMANI
T
x − x̂
y − ŷ
1
z − ẑ
≤
2 p − p̂
u − û
∆
P A2
T
A2 P
0
r̄
P
T
Gi P Gi
0
i=1
r̃
P
λi (RT P +RT P Ri )
0
i
i
i=1
BT P
r̄
P
i=1
r̄
P
i=1
0
GTi P Gi
0
r̃
P
i=1
λi (P Ri +RiT P Ri )
0
T
Gi P Gi
0
0
r̃
P
i=1
0
T
λi Ri P Ri
P B
x − x̂
0
y − ŷ
0
z − ẑ
p − p̂
0
u − û
0
0
1
≤ − c1 (x − x̂)T P (x − x̂) + c2 (y − ŷ)T P (y − ŷ) + c3 (z − ẑ)T P (z − ẑ) + c4 (p − p̂)T P (p − p̂) + c5 ku − ûk2
2
c5
≤ − c1 V (x, x̂) + c2 V (y, ŷ) + c3 V (z, ẑ) + c4 V (p, p̂) + ku − ûk2 .
2
P4
Thus by following the proof of Corollary 2.7 with κ0 =
i=2 ci , ψ(x, x̂) = (c1 − κ0 )V (x, x̂), ω(x, x̂) =
c1 −κ0
1
b
ψ(x,
x̂)
=
κV
(x,
x̂),
where
κ
=
,
∀t
≥
0,
∀ζ,
ζ̂
∈
CF
([−τ, 0]; Rn ), and ∀φ, φ̂ ∈ LkFt ([−τ, 0]; Rn )
1+κ0
1+κ0
0
satisfying (2.11), one obtains
c5
(2.19)
E[LV (φ, φ̂)] ≤ −κE[V (ξζ,υ (t), ξζ̂,υ̂ (t))] + ku − ûk2 .
2
By using generalized Ito’s formula, (2.19), condition (ii) in Definition 2.4, and Gronwall’s inequality, we have
Z t
Z t
E[LV (φ, φ̂)] d s
E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] = E[V (ξζ,υ (0), ξζ̂,υ̂ (0)) +
LV (φ, φ̂) d s] = E[V (ξζ,υ (0), ξζ̂,υ̂ (0))] +
0
0
Z t
λmax (P )
2
≤
E[LV (φ, φ̂)] d s
E[kξζ,υ (0) − ξζ̂,υ̂ (0)k ] +
2
0
Z t
c5
λmax (P )
− κE[V (ξζ,υ (s), ξζ̂,υ̂ (s))] + kυ(s) − υ̂(s)k2 d s
E[kζ − ζ̂k2[−τ,0]] +
≤
2
2
0
λmax (P )
c5
≤
E[kζ − ζ̂k2[−τ,0]]e−κt +
kυ − υ̂k2∞ .
2
2κ
Now by using condition (ii) in Definition 2.4, one obtains
c5
λmax (P )
λmin (P )
E[kξζ,υ (t) − ξζ̂,υ̂ (t)k2 ] ≤ E[V (ξζ,υ (t), ξζ̂,υ̂ (t))] ≤
E[kζ − ζ̂k2[−τ,0] ]e−κt +
kυ − υ̂k2∞ ,
2
2
2κ
and, hence,
E[kξζ,υ (t) − ξζ̂,υ̂ (t)k2 ] ≤
λmax (P )
c5
E[kζ − ζ̂k2[−τ,0] ]e−κt +
kυ − υ̂k2∞ .
λmin (P )
λmin (P )κ
Therefore, by introducing functions β and γ as
(2.20)
β(s, t) =
c5
λmax (P ) −κt
e s, γ(s) =
s2 ,
λmin (P )
λmin (P )κ
for any s, t ∈ R+
0 , inequality (2.3) is satisfied.
2.4. Noisy and Noiseless Trajectory. In this subsection, we provide an important technical lemma, which is
used later to construct finite dimensional or finite abstractions of retarded jump-diffusion systems. The lemma
provides an upper bound on the distance between solution of ΣR taking values randomly in C([−τ, 0]; Rn ) and
those of corresponding non-probabilistic retarded systems (denoted by ΣR ) obtained by removing diffusion
and reset terms (that is, g and r in (2.1)). From now onwards, we use notation ξ ζ,υ (t) to denote the value of
a solution process of ΣR in Rn and ξ t,ζ,υ to denote the solution of ΣR in C([−τ, 0]; Rn ) at time t ∈ R+
0 started
b
n
from non-probabilistic initial condition ζ ∈ CF
([−τ,
0];
R
),
where
F
is
the
trivial
sigma-algebra,
and
under
0
0
input signal υ. From now on, we raise a supplementary assumption on δ-ISS-Mk Lyapunov functions V which
is used to show later results.
SYMBOLIC MODELS FOR RETARDED JUMP-DIFFUSION SYSTEMS
9
Assumption 2.9. For a δ-ISS-Mk Lyapunov function V defined in Definition 2.4, one has
(i) kV (x, y) − V (x, z)k ≤ γ̃(ky − zk);
(ii) ω(x, y) ≥ κV (x, y)
for some concave K∞ function γ̃, for some constant κ ∈ R+ , and any x, y, z ∈ Rn .
The condition (i) in Assumption 2.9 is not restrictive as long as one is only interested in working in a compact
subset of Rn × Rn . For all x, y, z ∈ D, where D is a compact subset of Rn , by applying mean value theorem to
(x,y)
ks.
the function y → V (x, y), one obtains, kV (x, y) − V (x, z)k ≤ γ̃(ky − zk), where γ̃(s) = max k ∂V∂y
(x,y)∈D\△
Now we provide the main result of this subsection.
Lemma 2.10. Consider an RJDS ΣR such that f (o, 0m ) = 0n , g(o) = 0n×r , and r(o) = 0n×r̃ , where o ≡ 0.
Suppose that k ≥ 2 and there exists a δ-ISS-Mk Lyapunov function V for ΣR with Assumption 2.9 such that its
Hessian is a positive semidefinite matrix in R2n×2n and ∂x,x V (x, x̂) ≤ P , for any x, x̂ ∈ Rn , and some positive
b
semidefinite matrix P ∈ Rn×n . Then for any ζ ∈ CF
([−τ, 0]; Rn ), where F0 is the trivial sigma-algebra, and
0
υ ∈ U, we have
E[kξt,ζ,υ − ξ t,ζ,υ kk[−τ,0] ] ≤ σζ (g, r, t),
(2.21)
where the nonnegative valued function σζ tends to zero as t → 0, t → +∞, or as Lg → 0 and Lr → 0, where
Lg and Lr are the Lipschitz constants, introduced in Definition 2.1.
Proof. In the proof, we use notation H(V )(x, x̂) to denote the Hessian matrix of V at (x, x̂) ∈ R2n . For the
sake of simplicity, we drop the arguments of ∂x,x V, ∂x V , ∂x̂ V , and H(V ). By using Jensen’s inequality and
properties of V in Definition 2.4, we have
α(E[kξζ,υ (t) − ξ ζ,υ (t)kk ]) ≤ E[α(kξζ,υ (t) − ξ ζ,υ (t)kk )] ≤ E[V (ξζ,υ (t), ξ ζ,υ (t))] =
Z t h
f (ξs,ζ,υ , υ(s)) 1
+ Tr(g(ξs,ζ,υ )g T(ξs,ζ,υ )∂x,x V )
= E ∂x V ∂x̂ V
,
υ(s))
f
(ξ
2
s,ζ,υ
0
r̃
i
X
+ λi V (ξζ,υ (s)+r(ξs,ζ,υ )ei , ξ ζ,υ (s))−V (ξζ,υ (s), ξ ζ,υ (s)) d s
=
Z
i=1
t h
E
0
+
∂x V
∂x̂ V
f (ξs,ζ,υ , υ(s))
1
g(ξs,ζ,υ ) T
+ Tr
g (ξs,ζ,υ )
g(ξ s,ζ,υ )
f (ξ s,ζ,υ , υ(s))
2
Z
t
0
E[LV (ξs,ζ,υ , ξ s,ζ,υ )] d s
g T (ξ s,ζ,υ ) H(V )
r̃
X
λi V (ξζ,υ (s)+r(ξs,ζ,υ )ei , ξ ζ,υ (s)+r(ξ s,ζ,υ )ei ) − V (ξζ,υ (s), ξ ζ,υ (s))
i=1
r̃
X
1
+ Tr(g(ξs,ζ,υ )g T(ξs,ζ,υ )∂x,x V ) +
λi V (ξζ,υ (s) + r(ξs,ζ,υ )ei , ξ ζ,υ (s))
2
i=1
X
r̃
i
1
g(ξs,ζ,υ ) T
T
λi V (ξζ,υ (s) + r(ξs,ζ,υ )ei , ξ ζ,υ (s) + r(ξ s,ζ,υ )ei ) d s.
− Tr
g (ξs,ζ,υ ) g (ξ s,ζ,υ ) H(V ) −
g(ξ s,ζ,υ )
2
i=1
10
P. JAGTAP AND M. ZAMANI
With the help of Assumption 2.9 and Gronwall’s inequality, one obtains
Z t
Z t
1
−κE[V (ξζ,υ (s), ξ ζ,υ (s))] d s +
E[ Tr(g(ξs,ζ,υ )g T (ξs,ζ,υ )∂x,x V )
α(E[kξζ,υ (t) − ξ ζ,υ (t)kk ]) ≤
2
0
0
X
r̃
1
g(ξs,ζ,υ ) T
λi γ̃(kr(ξ s,ζ,υ )ei k)] d s
− Tr
g (ξs,ζ,υ ) g T (ξ s,ζ,υ ) H(V ) +
)
g(ξ
2
s,ζ,υ
i=1
Z t
Z t
r̃
X
1
≤
−κE[V (ξζ,υ (s), ξ ζ,υ (s))] d s +
E[ Tr(g(ξs,ζ,υ )g T (ξs,ζ,υ )∂x,x V ) +
λi γ̃(kr(ξ s,ζ,υ )ei k)] d s
2
0
0
i=1
≤σ ζ (g, r, t)e−κt ,
where the function σ ζ can be computed as
Z t
r̃
X
1 p
2
σ ζ (g, r, t):=
E[ k ∂x,x V g(ξs,ζ,υ )kF +
λi γ̃(kr(ξ s,ζ,υ )ei k)] d s.
2
0
i=1
Now by using Lipschitz assumptions on the diffusion and reset term one can obtain
Z t
Z t
r̃
X
1 p
λi
γ̃(Lr kξ s,ζ,υ k[−τ,0] ) d s.
σ ζ (g, r, t) ≤ k P k2 n min{n, r}L2g
E[kξs,ζ,υ k2[−τ,0]] d s +
2
0
0
i=1
Since V is a δ-ISS-Mk Lyapunov function for k ≥ 2, f (o, 0m ) = 0n , g(o) = 0n×r , r(o) = 0n×r̃ , and using
Lemma 2.3, we have
k
k
(E[kξt,ζ,υ k2[−τ,0]]) 2 ≤ E[(kξt,ζ,υ k2[−τ,0] ) 2 ] = E[kξt,ζ,υ kk[−τ,0] ]
≤ β̃(kζkk[−τ,0] , t) + γ(kυk∞ ) ≤ β̃(kζkk[−τ,0] , t) +γ(sup{kuk}).
u∈U
Thus, we have
2
E[kξt,ζ,υ k2[−τ,0] ] ≤ (β̃(kζkk[−τ,0] , t) + γ(sup{kuk})) k
u∈U
and
1
E[kξt,ζ,υ k[−τ,0]] ≤ (β̃(kζkk[−τ,0] , t) + γ(sup{kuk})) k .
u∈U
By using these inequalities, one has
Z t
1 p
2
σ ζ (g, r, t) ≤ k P k2 n min{n, r}L2g (β̃(kζkk[−τ,0] , s) + γ(sup{kuk})) k d s
2
u∈U
0
Z
r̃
t
X
1
λi
γ̃(Lr ((β̃(kζkk[−τ,0] , s) + γ(sup{kuk})) k )) d s.
+
u∈U
0
i=1
By defining
Z t
1 p
2
2
2 −κt
(β̃(kζkk[−τ,0] , s) + γ(sup{kuk})) k d s
σ̂ζ (g, r, t) =α
k P k n min{n, r}Lg e
2
u∈U
0
Z t
r̃
X
1
λi e−κt
γ̃(Lr ((β̃(kζkk[−τ,0] , s) + γ(sup{kuk})) k )) d s ,
+
−1
(2.22)
i=1
u∈U
0
we obtain E[kξζ,υ (t) − ξ ζ,υ (t)kk ] ≤ σ̂ζ (g, r, t). Now by definition of ξt = {ξ(t + θ)|θ ∈ [−τ, 0]} one can obtain
(2.23)
E[kξt,ζ,υ −ξ t,ζ,υ kk[−τ,0] ]≤ sup E[kξζ,υ (t+θ)−ξ ζ,υ (t+θ)kk ] ≤ σζ (g, r, t),
θ∈[−τ,0]
where σζ (g, r, t) :=
sup {σ̂ζ (g, r, t + θ)}, which completes the proof. Note that σζ (g, r, t) satisfies all the
θ∈[−τ,0]
conditions proposed in the lemma.
SYMBOLIC MODELS FOR RETARDED JUMP-DIFFUSION SYSTEMS
11
3. Systems and Approximate Equivalence Relations
We recall the notion of system introduced in [Tab09] which later serves as a unified modeling framework for
both retarded jump-diffusion systems and their finite dimensional abstractions and symbolic models.
Definition 3.1. A systems is a tuple S = (X, X0 , U, −→, Y, H) where X is a set of states (possibly infinite),
X0 ⊆ X is a set of initial states, U is a set of inputs (possibly infinite), −→⊆ X × U × X is a transition
relation, Y is a set of outputs, and H : X → Y is an output map.
u
We denote x −→ x′ as an alternative representation for a transition (x, u, x′ ) ∈−→, where state x′ is called a
u-successor (or simply successor) of state x, for some input u ∈ U . Moreover, a system S is said to be
•
•
•
•
metric, if the output set Y is equipped with a metric d : Y × Y → R+
0.
finite (or symbolic), if X and U are finite.
deterministic, if there exists at most a u-successor of x, for any x ∈ X and u ∈ U .
nonblocking, if for any x ∈ X, there exists some u-successor of x, for some u ∈ U .
For a system S, the finite state-run generated from initial state x0 ∈ X0 is a finite sequence of transitions:
u
un−2
u
un−1
1
0
· · · −→ xn−1 −→ xn ,
x1 −→
x0 −→
(3.1)
u
i
xi+1 , for i ∈ {0, 1, . . . , n − 1}. The associated finite output-run is given by yi = H(xi ), for
such that xi −→
i ∈ {0, 1, . . . , n − 1}. These finite runs can be directly extended to infinite runs as well.
Now, we provide the notion of approximate (bi)simulation relation between two systems, introduced in [GP07],
which is later used for analyzing and synthesizing controllers for retarded jump-diffusion systems.
Definition 3.2. Let S1 = (X1 , X10 , U1 , −→, Y1 , H1 ) and S2 = (X2 , X20 , U2 , −→, Y2 , H2 ) be two metric systems
1
2
having the same output sets Y1 = Y2 and metric d. For ε ∈ R+
0 , a relation R ⊆ X1 × X2 is said to be an
ε-approximate bisimulation relation between S1 and S2 if it satisfies the following conditions:
(i) ∀(x1 , x2 ) ∈ R, we have d(H1 (x1 ), H2 (x2 )) ≤ ε;
u2
u1
x′2 in S2 satisfying (x′1 , x′2 ) ∈ R;
x′1 in S1 implies x2 −→
(ii) ∀(x1 , x2 ) ∈ R, x1 −→
1
u
2
u
2
1
1
2
x′1 in S1 satisfying (x′1 , x′2 ) ∈ R.
x′2 in S2 implies x1 −→
(iii) ∀(x1 , x2 ) ∈ R, x2 −→
If we remove condition (iii), then R ⊆ X1 × X2 is said to be an ε-approximate simulation relation from S1 to
S2 .
The system S1 is ε-approximate bisimilar to S2 , denoted by S1 ∼
=εS S2 , if there exists an ε-approximate
bisimulation relation R between S1 and S2 such that: ∀x10 ∈ X10 , ∃x20 ∈ X20 with (x10 , x20 ) ∈ R and ∀x20 ∈
X20 , ∃x10 ∈ X10 with (x10 , x20 ) ∈ R. In order to present the main results of the paper, we need to employ the
notion of system as an abstract representation of a retarded jump-diffusion system. First, we define a metric
system associated with the retarded jump-diffusion system ΣR , denoted by S(ΣR ) = (X, X0 , U, −→, Y, H),
where
X is the set of all C([−τ, 0]; Rn )-valued random variables defined on the probability space (Ω, F , P);
b
X0 is a subset of CF
([−τ, 0]; Rn );
0
U = U;
υ
+
ζ −→ ζ ′ if ζ and ζ ′ are measurable in Ft and Ft+h , respectively, for some t ∈ R+
0 and h ∈ R , and
k
n
′
there exists a solution ξt ∈ LFt ([−τ, 0]; R ) of ΣR satisfying ξt = ζ and ξh,ζ,υ = ζ P-a.s.;
• Y = X;
• H(ζ) = ζ.
•
•
•
•
1
We assume that the output Y is equipped with the metric d(y, y ′ ) = (E[ky − y ′ kk[−τ,0] ]) k , for any y, y ′ ∈ Y and
some k ≥ 1. Also, note that the system S(ΣR ) is deterministic and nonblocking. From now on, we restrict
12
P. JAGTAP AND M. ZAMANI
our attention to the sampled-data system, where control signals (in ΣR ) are piecewise-constant over intervals
of length h ∈ R+ , i.e.
Uh = {υ ∈ U | υ(t) = υ(ih), t ∈ [ih, (i + 1)h), i ∈ N0 }.
The metric systems associated with the sampled-data retarded jump-diffusion systems can be defined as
υh
Sh (ΣR ) = (Xh , Xh0 , Uh , −→, Yh , Hh ), where Xh = X, Xh0 = X0 , Uh = Uh , Yh = Y , Hh = H, and ζh −→
ζh′
h
h
if ζh and ζh′ are measurable in Fih and F(i+1)h , respectively, for some i ∈ N0 , and there exists a solution
ξt ∈ LkFt ([−τ, 0]; Rn ) of ΣR satisfying ξt = ζh and ξh,ζh ,υh = ζh′ P-a.s. In other words, a finite stateυ
υ
υN −1
h
h
h
1
0
· · · −→ ζN , where υi ∈ Uh and ζi+1 = ξh,ζi ,υi P-a.s. for
ζ1 −→
run of Sh (ΣR ), represented by ζ0 −→
i ∈ {0, 1, . . . , N − 1}, captures solutions of RJDS ΣR at the sampling times t = 0, h, . . . , N h, started from
ζ0 ∈ X0 and resulting from control input υ obtained by the concatenation of input signals υi ∈ Uh . Moreover,
the corresponding finite output-run is {y0 , y1 , . . . , yN }.
4. Finite Dimensional Abstractions for RJDS
In this section, we introduce a finite dimensional abstraction for Sh (ΣR ). Consider metric systems associated
with the sampled-data retarded jump-diffusion systems Sh (ΣR ) and consider triple ρ = (h, N, ζs ) of parameters, where h ∈ R+ is the sampling time, N ∈ N is a temporal horizon, and ζs ∈ C([−τ, 0]; Rn ) is a source
state. Let us define two metric systems as
Sρ (ΣR ) = (Xρ , Xρ0 , Uρ , −→, Yρ , Hρ ),
ρ
S ρ (ΣR ) = (Xρ , Xρ0 , Uρ , −→, Yρ , H ρ ),
ρ
where
• Xρ = UN , Xρ0 = Xρ , Uρ = U, Yρ = Yh ;
uρ
• xρ −→ x′ρ , where xρ = (u1 , u2 , . . . , uN ) ∈ Xρ , if and only if x′ρ = (u2 , . . . , uN , uρ );
ρ
• Hρ (xρ ) = ξN h,ζs ,xρ (H ρ (xρ ) = ξ N h,ζs ,xρ ).
Here, we abuse notation by identifying xρ = (u1 , u2 , . . ., uN ) ∈ UN as an input curve υ : [0, N h) → U such
that υ(t) = uk for any t ∈ [(k − 1)h, kh) for k ∈ {1, . . . , N } in ξN h,ζs ,xρ and ξ N h,ζs ,xρ . We use similar notations
in the rest of the paper as well. Notice that system Sρ (ΣR ) (resp. S ρ (ΣR )) is deterministic, non-blocking,
and finite dimensional (but not necessarily symbolic unless U is a finite set). Note that Hρ and H ρ are the
output maps from non-probabilistic state xρ ∈ Xρ to a C([−τ, 0]; Rn )-valued random variable ξN h,ζs ,xρ and to
a non-probabilistic C([−τ, 0]; Rn )-valued variable ξ N h,ζs ,xρ , respectively.
The next theorem provides one of the main results of this section on the construction of finite dimensional
abstractions which are approximately bisimilar to RJDS ΣR .
Theorem 4.1. Consider an RJDS ΣR , admitting a δ-ISS-Mk Lyapunov function of the form explained in
Lemma 2.10. Given any ε > 0, let the sampling time h, temporal horizon N , and source state ζs be such that
1
1
1
(β̃(εk , h))) k + (σζs (g, r, (N + 1)h)) k + (β̃(Z(ζs ), N h)) k ≤ ε,
(4.1)
where Z(ζs ) = sup kξ h,ζs ,u1 − ζs kk[−τ,0] . Then, the relation
u1 ∈U
1
R1 = {(ζ, xρ ) ∈ Xh × Xρ | (E[kH(ζ) − H ρ (xρ )kk[−τ,0] ]) k ≤ ε}
is an ε-approximate bisimulation relation between Sh (ΣR ) and S ρ (ΣR ).
Proof. Consider any (ζ, xρ ) ∈ R1 , where ζ ∈ Xh and xρ = (u1 , u2 , . . . , uN ) ∈ Xρ . Then we have (E[kH(ζ) −
1
H ρ (xρ )kk[−τ,0] ]) k ≤ ε. Thus condition (i) in Definition 3.2 holds. Now we show that condition (ii) in Definition
SYMBOLIC MODELS FOR RETARDED JUMP-DIFFUSION SYSTEMS
13
3.2 holds. Consider any υh : [0, h[→ uh for some uh ∈ U and ζ ′ = ξh,ζ,υh . Consider uρ = uh and x′ρ =
(u2 , . . . , uN , uρ ) and let xρ = (u1 , u2 , . . . , uN , uρ ) denote input sequence in UN +1 . With the help of Jensen
inequality, Lemma 2.10, triangle inequality, and (2.4) one obtains the following chains of inequalities:
1
1
(E[kH(ζ ′ ) − H ρ (x′ρ )kk[−τ,0] ]) k = (E[kH(ζ ′ ) − Hρ (xρ ) + Hρ (xρ ) − H ρ (x′ρ )kk[−τ,0] ]) k
1
= (E[kξh,ζ,υh − ξ(N +1)h,ζs ,xρ + ξ(N +1)h,ζs ,xρ − ξ N h,ζs ,x′ρ kk[−τ,0] ]) k
1
= (E[kξh,ζ,υh − ξ(N +1)h,ζs ,xρ + ξ(N +1)h,ζs ,xρ − ξ (N +1)h,ζs ,xρ + ξ (N +1)h,ζs ,xρ − ξ N h,ζs ,x′ρ kk[−τ,0] ]) k
1
1
≤ (E[kξh,ζ,υh − ξ(N +1)h,ζs ,xρ kk[−τ,0] ]) k + (E[kξ(N +1)h,ζs ,xρ − ξ (N +1)h,ζs ,xρ kk[−τ,0]]) k
1
+ (E[kξ (N +1)h,xs ,xρ − ξ N h,ζs ,x′ρ kk[−τ,0] ]) k
1
1
≤ (E[kξh,ζ,υh − ξh,ξN h,ζs ,xρ ,uρ kk[−τ,0]]) k + (E[kξ(N +1)h,ζs ,xρ − ξ (N +1)h,ζs ,xρ kk[−τ,0] ]) k
+ (kξ N h,ξh,ζ
1
s ,u1
,x′ρ
− ξ N h,ζs ,x′ρ kk[−τ,0] ) k .
1
1
1
≤ (β̃(E[kζ − ξN h,ζs ,xρ kk[−τ,0] ], h)) k + (σ(g, r, (N + 1)h)) k + (β̃(kξ h,ζs ,υ1 − ζs kk[−τ,0], N h)) k
1
1
1
≤ (β̃(εk , h)) k + (σζs (g, r, (N + 1)h)) k + (β̃(Z(ζs ), N h)) k ≤ ε.
Hence, (ζ ′ , x′ρ ) ∈ R1 . Thus condition (ii) in Definition 3.2 holds. In a similar way, one can show that condition
(iii) in Definition 3.2 holds which completes the proof.
Note that in the above theorem, given any ε > 0, one can select temporal horizon N to be sufficiently large to
1
enforce terms σζs (g, r, (N +1)h) and β̃(Z(ζs ), N h) to be sufficiently small. This results in (α−1 (e−κh α(εk ))) k <
ε which enforces a lower bound for the sampling time h.
Now we establish the results on the existence of non-probabilistic finite dimensional abstraction S ρ (ΣR ) such
that Sh (ΣR ) ∼
=εS S ρ (ΣR ) given the result in Theorem 4.1.
Theorem 4.2. Consider the results in Theorem 4.1. If we select
1
b
([−τ, 0]; Rn )|(E[kH(ζ)−H ρ (xρ0)kk[−τ,0] ]) k ≤ε, ∃xρ0 ∈Xρ0 },
Xh0 ⊆{ζ ∈ CF
0
then we have Sh (ΣR ) ∼
=εS S ρ (ΣR ).
1
Proof. For every ζ ∈ Xh0 , there always exists xρ0 ∈ Xρ0 such that (E[kH(ζ) − H ρ (xρ0 )kk[−τ,0] ]) k ≤ ε. Hence
(ζ, xρ0 ) ∈ R1 . In a similar way, we can show that for every xρ0 ∈ Xρ0 there exists ζ ∈ Xh0 such that
(ζ, xρ0 ) ∈ R1 , which completes the proof.
The next theorems provide results that are similar to those of Theorems 4.1 and 4.2, but by using finite
dimensional abstraction Sρ (ΣR ) rather than S ρ (ΣR ). Note that Sρ (ΣR ) is a stochastic finite dimensional
abstraction for ΣR .
Theorem 4.3. Consider a δ-ISS-Mk RJDS ΣR . Given any ε > 0, let the sampling time h, temporal horizon
N , and source state ζs be such that
1
1
(β̃(εk , h))) k + (β̃(Z̃(ζs ), N h)) k ≤ ε,
(4.2)
where Z̃(ζs)= sup E[kξh,ζs ,u1− ζs kk[−τ,0]]. Then, the relation
u1∈U
1
R1 = (ζ, xρ )∈Xh × Xρ |(E[kH(ζ) − Hρ (xρ )kk[−τ,0] ]) k ≤ε
is an ε-approximate bisimulation relation between Sh (ΣR ) and Sρ (ΣR ).
Proof. The proof is similar to the one of Theorem 4.1.
14
P. JAGTAP AND M. ZAMANI
Theorem 4.4. Consider the results in Theorem 4.3. If we select
1
b
([−τ, 0]; Rn )|(E[kH(ζ)−Hρ (xρ0 )kk[−τ,0] ]) k ≤ε, ∃xρ0 ∈Xρ0 },
Xh0 ⊆{ζ ∈ CF
0
then we have Sh (ΣR ) ∼
=εS Sρ (ΣR ).
Proof. The proof is similar to the one of Theorem 4.2.
Now we provide results on the construction of finite abstractions.
5. Finite Abstractions for RJDS
In this section, we provide a finite (a.k.a. symbolic) abstraction for Sh (ΣR ) by quantizing input set U. Let
us consider tuple ρ = (h, N, ζs , η), where η > 0 is a quantization parameter and the quantized input set is
denoted by [U]η (cf. Notation in Subsection 2.1). Now, we can define corresponding finite systems as
Sρ (ΣR ) = (Xρ , Xρ0 , Uρ , −→, Yρ, Hρ ),
ρ
S ρ (ΣR ) = (Xρ , Xρ0 , Uρ , −→, Yρ, H ρ ),
ρ
where
• Xρ = [U]N
η ,Xρ0 = Xρ , Uρ = [U]η , Yρ = Yh ;
uρ
• xρ −→ xρ′ , where xρ = (u1 , u2 , . . . , uN ) ∈ Xρ , if and only if x′ρ = (u2 , . . . , uN , uρ );
ρ
• Hρ (xρ ) = ξN h,ζs ,xρ (H ρ (xρ ) = ξ N h,ζs ,xρ ).
In order to provide approximate bisimulation relation between sampled retarded jump-diffusion systems and
symbolic models, we need following technical lemmas.
Lemma 5.1. Consider a δ-ISS-Mk RJDS ΣR and a quantization parameter η such that 0 < η ≤ span(U).
Then the relation R2 given by
(5.1)
R2 = {(xρ , xρ ) ∈ Xρ × Xρ | xρ = (u1 , u2 , . . . , uN ), xρ = ([u1 ]η , [u2 ]η , . . . , [uN ]η )}
1
1
(γ(η)) k
S ρ (ΣR ).
is a (γ(η)) k -approximate bisimulation relation between S ρ (ΣR ) and S ρ (ΣR ), and S ρ (ΣR ) ∼
=S
Proof. Let (xρ , xρ ) ∈ R2 , then kui − [ui ]η k ≤ η for i ∈ {1, 2, . . . , N } implies that kxρ − xρ k∞ ≤ η. By using
the definition of δ-ISS-Mk , one obtains
1
1
1
1
(kH ρ (xρ )−H ρ (xρ )kk ) k =(kξ N h,ζs ,xρ −ξ N h,ζs ,xρ kk[−τ,0] ) k ≤ (γ(kxρ − xρ k∞ )) k ≤ (γ(η)) k .
Then, the first condition in Definition 3.2 holds. Now, consider any (xρ , xρ ) ∈ R2 , where xρ = (u1 , u2 , . . . , uN )
u
and xρ = (u1 , u2 , . . . , uN ). Let u ∈ Uρ and consider xρ −→ x′ρ := (u2 , . . . , uN , u) in S ρ (ΣR ). Choose u = [u]η
ρ
u
and consider xρ −→
ρ
x′ρ
:= (u2 , . . . , uN , u) in S ρ (ΣR ). It is obvious that (x′ρ , x′ρ ) ∈ R2 and, hence, condition
1
(ii) in Definition 3.2 holds. Similarly, condition (iii) in Definition 3.2 holds which shows R2 is a (γ(η)) k approximate bisimulation relation between S ρ (ΣR ) and S ρ (ΣR ). For any xρ0 := (u1 , u2 , . . . , uN ) ∈ Xρ0 , there
always exists xρ0 := ([u1 ]η , [u2 ]η , . . . , [uN ]η ) ∈ Xρ0 and, hence, (xρ0 , xρ0 ) ∈ R2 . Note that the existence of
such xρ0 is guaranteed by U being a finite union of boxes and by the inequality η ≤ span(U). Moreover, for
1
(γ(η)) k
any xρ0 ∈ Xρ0 and by choosing xρ0 = xρ0 , one readily gets (xρ0 , xρ0 ) ∈ R2 and, hence, S ρ (ΣR ) ∼
=
S ρ (ΣR ).
S
The next lemma provides results that are similar to those of Lemma 5.1 but by using symbolic model Sρ (ΣR ).
SYMBOLIC MODELS FOR RETARDED JUMP-DIFFUSION SYSTEMS
15
Lemma 5.2. Consider a δ-ISS-Mk RJDS ΣR and a quantization parameter 0 < η ≤ span(U). Then the
1
relation R2 as given in Lemma 5.1 is a (γ(η)) k -approximate bisimulation relation between Sρ (ΣR ) and Sρ (ΣR ),
1
(γ(η)) k
and Sρ (ΣR ) ∼
Sρ (ΣR ).
=
S
Proof. The proof is similar to the one of Lemma 5.1.
Now we provide the main results of this section on establishing an approximate bisimulation relation between Sh (ΣR ) and S ρ (ΣR ), which is an immediate consequence of the transitivity property of approximate
bisimulation relations [GP07, Proposition 4] as recalled next.
Proposition 5.3. Consider metric systems S1 , S2 , and S3 such that S1 ∼
=δS S2 and S2 ∼
=δ̂S S3 , for some
+
δ+
δ̂
δ, δ̂ ∈ R0 . Then, we have S1 ∼
=S S3 .
Now we provide the first main result of this section.
Theorem 5.4. Consider an RJDS ΣR . Given any ε > 0 and a quantization parameter 0 < η ≤ span(U),
consider the results in Theorem 4.1 and Lemma 5.1. Then, the relation R given by
R = (xh , xρ ) ∈ Xh × Xρ | ∃xρ ∈ Xρ , (xh , xρ ) ∈ R1 and (xρ , xρ ) ∈ R2
1
is a (ε + (γ(η)) k )-approximate bisimulation relation between Sh (ΣR ) and S ρ (ΣR ).
Note that relations R1 and R2 in Theorem 5.4 have been defined in Theorem 4.1, and Lemma 5.1, respectively.
Theorem 5.5. Under the assumptions of Theorem 5.4, let us choose Xh0 as in Theorem 4.2. Then, we have
1
ε+(γ(η)) k
S ρ (ΣR ).
Sh (ΣR ) ∼
=
S
The proofs of Theorems 5.4 and 5.5 are a simple consequence of the results in Theorem 4.4, Lemma 5.1, and
Proposition 5.3.
Note that by referring to the discussion after Theorem 4.1 and by allowing a lower bound on h, one can achieve
any abstraction precision in Theorem 5.5 by choosing sufficiently large temporal horizon N and sufficiently
small input set quantization η.
In a similar way, with the help of transitivity property of bisimulation relations as in Proposition 5.3, we can
provide an approximate bisimulation relation between sampled retarded jump-diffusion system Sh (ΣR ) and
the symbolic model Sρ (ΣR ) with probabilistic output values as the following.
Theorem 5.6. Consider an RJDS ΣR . Given any ε > 0 and quantization parameter 0 < η ≤ span(U),
consider the results in Theorem 4.3 and Lemma 5.2. Then, the relation R given by
R = (xh , xρ ) ∈ Xh × Xρ | ∃xρ ∈ Xρ , (xh , xρ ) ∈ R1 and (xρ , xρ ) ∈ R2
1
is a (ε + (γ(η)) k )-approximate bisimulation relation between Sh (ΣR ) and Sρ (ΣR ).
Theorem 5.7. Under the assumptions of Theorem 5.6, let us choose Xh0 as in Theorem 4.4. Then we have
1
∼ε+(γ(η)) k Sρ (ΣR ).
Sh (ΣR ) =
S
The proofs of Theorems 5.6 and 5.7 are a simple consequence of the results in Theorem 4.3, Lemma 5.2, and
Proposition 5.3.
16
P. JAGTAP AND M. ZAMANI
Room 7
Room1
Room 2
Room 8
Room 3
Room 4
Room 5
Heater 1
Heater 2
Room 9
Room 10
Room 6
Figure 1. A schematic of ten-room building.
6. An Example
To show the effectiveness of the proposed results, we consider a simple thermal model of ten-room building
as shown schematically in Figure 1. The model used here is similar to that used in [ZAG15]. In addition,
we modified arrangement of the rooms to increase state-space dimensions and considered that the dynamic is
affected by delays in states and by jumps (modeling door and window opening). The dynamic of the considered
delayed jump-diffusion system ΣD is given by the following delayed stochastic differential equations:
d ξ1 (t) =(α21 (ξ2 (t) − ξ1 (t)) + αe1 (Te − ξ1 (t)) − α1τ1 ξ1 (t − τ1 )) d t + (g1 ξ1 (t) + g1τ2 ξ1 (t − τ2 )) d Wt1 + r1 ξ1 (t) d Pt1 ,
d ξ2 (t) =(α12 (ξ1 (t) − ξ2 (t)) + α72 (ξ7 (t) − ξ2 (t)) + α92 (ξ9 (t) − ξ2 (t)) + α32 (ξ3 (t) − ξ2 (t)) + αe2 (Te − ξ2 (t))
− α2τ1 ξ2 (t − τ1 ) + αH1 (Th − ξ2 (t))υ1 ) d t + (g2 ξ2 (t) + g2τ2 ξ2 (t − τ2 )) d Wt2 + r1 ξ2 (t) d Pt2 ,
d ξ3 (t) =(α23 (ξ2 (t) − ξ3 (t)) + α43 (ξ4 (t) − ξ3 (t)) + αe3 (Te − ξ3 (t)) − α3τ1 ξ3 (t − τ1 )) d t
+ (g3 ξ3 (t) + g3τ2 ξ3 (t − τ2 )) d Wt3 + r3 ξ3 (t) d Pt3 ,
d ξ4 (t) =(α34 (ξ3 (t) − ξ4 (t)) + α54 (ξ5 (t) − ξ4 (t)) + αe4 (Te − ξ4 (t)) − α4τ1 ξ4 (t − τ1 )) d t
+ (g4 ξ4 (t) + g4τ2 ξ4 (t − τ2 )) d Wt4 + r4 ξ4 (t) d Pt4 ,
d ξ5 (t) =(α45 (ξ4 (t) − ξ5 (t)) + α85 (ξ8 (t) − ξ5 (t)) + α105 (ξ10 (t) − ξ5 (t)) + α65 (ξ6 (t) − ξ5 (t)) + αe5 (Te − ξ5 (t))
− α5τ1 ξ5 (t − τ1 ) + αH2 (Th − ξ5 (t))υ2 ) d t + (g5 ξ5 (t) + g5τ2 ξ5 (t − τ2 )) d Wt5 + r5 ξ5 (t) d Pt5 ,
d ξ6 (t) =(α56 (ξ5 (t) − ξ6 (t)) + αe6 (Te − ξ6 (t)) − α6τ1 ξ6 (t − τ1 )) d t + (g6 ξ6 (t) + g6τ2 ξ6 (t − τ2 )) d Wt6 + r6 ξ6 (t) d Pt6 ,
d ξ7 (t) =(α27 (ξ2 (t) − ξ7 (t)) + αe7 (Te − ξ7 (t)) − α7τ1 ξ7 (t − τ1 )) d t + (g7 ξ7 (t) + g7τ2 ξ7 (t − τ2 )) d Wt7 + r7 ξ7 (t) d Pt7 ,
d ξ8 (t) =(α58 (ξ5 (t) − ξ8 (t)) + αe8 (Te − ξ8 (t)) − α8τ1 ξ8 (t − τ1 )) d t + (g8 ξ8 (t) + g8τ2 ξ8 (t − τ2 )) d Wt8 + r8 ξ8 (t) d Pt8 ,
d ξ9 (t) =(α29 (ξ2 (t) − ξ9 (t)) + αe9 (Te − ξ9 (t)) − α9τ1 ξ9 (t − τ1 )) d t + (g9 ξ9 (t) + g9τ2 ξ9 (t − τ2 )) d Wt9 + r9 ξ9 (t) d Pt9 ,
d ξ10 (t) =(α510 (ξ5 (t) − ξ10 (t)) + αe10 (Te − ξ10 (t)) − α10τ1 ξ10 (t − τ1 )) d t + (g10 ξ10 (t) + g10τ2 ξ10 (t − τ2 )) d Wt10
+ r10 ξ10 (t) d Pt10 ,
where the terms Wti and Pti , i ∈ {1, 2, . . . , 10}, denote the standard Brownian motion and Poisson process with
rate λi = 0.1, respectively; ξi , i ∈ {1, 2, . . . , 10}, denote the temperature in each room; Te = 15 (degree Celsius)
is the external temperature; TH1 = TH2 = 100 are the temperatures of two heaters; τ1 = 15 time units and
τ2 = 10 time units are state delays in drift and diffusion terms, respectively; and the control inputs υ1 , υ2 are
amounted to 1 if corresponding heaters are on and to 0 if corresponding heaters are off. Here, we assume that
at most one heater is on at each time instance which results in the finite input set U = {(1, 0), (0, 1), (0, 0)}.
The system parameters are chosen as αij = 5×10−2 for all i 6= j and i, j ∈ {1, 2, . . . , 10}, αe2 = αe5 = 8×10−3,
αe1 = αe3 = αe4 = αe6 = αe7 = αe8 = αe9 = αe10 = 5 × 10−3 , αH1 = αH2 = 3.6 × 10−3 , αiτ1 = 1 × 10−4 ,
gi = 2 × 10−3 , giτ2 = 1 × 10−4 , and ri = 1 × 10−3 , for i ∈ {1, 2, . . . , 10}. Using the result in Lemma 2.8 with
P = I10 , one can readily compute function β(s, t) = e−κt s, ∀s ∈ R+
0 , with κ = 0.0129.
By considering the 2nd moment, i.e. k = 2, a constant function ζs ≡ [17, 17, 17, 17, 17, 17, 17, 17, 17, 17]T ∈
C([−τ, 0]; Rn ),where τ = 15, a precision ε = 0.31 and by fixing sampling time h = 30 time units, one can
SYMBOLIC MODELS FOR RETARDED JUMP-DIFFUSION SYSTEMS
21
20
20
20
4
6
21
2
21
19
19
19
18
18
18
18
19
20
17
21
18
1
19
20
18
21
19
20
21
5
3
21
20
20
8
10
21
19
19
18
18
18
19
20
18
21
19
20
21
9
7
Figure 2. A few realizations of the solution process ξζ,υ with initial condition ζ ≡
[19, 19, 19, 19, 19, 19, 19, 19, 19, 19]T .
1
1
0.5
0
0
200
400
0
200
400
600
800
1000
1200
600
800
1000
1200
2
1
0.5
0
Time
Figure 3. The evolution of input signals υ1 and υ2 .
obtain temporal horizon N = 14 for S ρ (ΣD ), satisfying inequality (4.1) in Theorem 4.1. Therefore, the
resulting cardinality of the set of states of S ρ (ΣD ) is 314 = 4782969. Note that using the results in Theorem
4.1 one obtains N = 18 which results in a bigger abstraction. Remark that since the input set is finite, the
finite dimensional abstraction S ρ (ΣD ) is also symbolic.
18
P. JAGTAP AND M. ZAMANI
(E[|| , (t)|| 2W])1/2
0.1
0.08
0.06
0.04
0.02
0
0
200
400
600
800
1000
1200
Time
Figure 4. The square root of average values (over 10000 realizations) of the squared distance
of the solution process ξζ,υ to the set W.
Now consider the objective to design a controller enforcing the trajectories of ΣD to stay within comfort zone
W = [18, 21]10 . This corresponds to the LTL specification W. The computation of the symbolic model
S ρ (ΣD ) and the controller synthesis have been performed using tool QUEST [JZst] on a computer with CPU
3.5GHz Intel Core i7. Figure 2 shows a few realizations of the closed-loop solution process ξζ,υ starting form
initial condition ζ ≡ [19, 19, 19, 19, 19, 19, 19, 19, 19, 19]T ∈ Xh0 . The synthesized control signals υ1 and υ2
are shown in Figure 3. In Figure 4, we show the square root of the average value (over 10000 realizations)
of the squared distance in time of the solution process ξζ,υ to the set W, namely kξζ,υ (t)k2W , where the point
to set distance is defined as kxk2W = inf w∈W kx − wk2 . One can readily observe that the empirical average is
much lower than the precision ε = 0.31. In Table 1, we report sizes of finite abstractions (given by number of
transitions), sizes of controllers, lower bounds for precision ε, and computation time required for constructing
finite abstractions and controllers for different values of N .
Table 1. Performance comparison for different values of temporal horizon N .
N
15
14
13
11
9
Number of transitions in S ρ (ΣD )
43046721 14348907 4782969 531441 59049
Number of transitions in the controller 695971
222862
70757
6819
982
Abstraction computation time (Sec)
93.37
24.204
6.84
0.413 0.015
Controller computation time (Sec)
3650.57
1150.98
354.59
2.56
2.07
precision ε
0.26
0.31
0.38
0.55
0.8
References
[BSP08]
[CGG13]
[Gir14]
[GP07]
M. Bandyopadhyay, T. Saha, and R. Pal. Deterministic and stochastic analysis of a delayed allelopathic phytoplankton model within fluctuating environment. Nonlinear Analysis: Hybrid systems, 2(3):958–970, 2008.
E. Le Corronc, A. Girard, and G. Goessler. Mode sequences as symbolic states in abstractions of incrementally
stable switched systems. In 52nd IEEE Conference on Decision and Control, pages 3225–3230, Dec 2013.
A. Girard. Approximately bisimilar abstractions of incrementally stable finite or infinite dimensional systems. In
53rd IEEE Conference on Decision and Control, pages 824–829, Dec 2014.
A. Girard and G. J. Pappas. Approximation metrics for discrete and continuous systems. IEEE Transactions on
Automatic Control, 52(5):782–798, May 2007.
SYMBOLIC MODELS FOR RETARDED JUMP-DIFFUSION SYSTEMS
19
[GPT10]
A. Girard, G. Pola, and P. Tabuada. Approximately bisimilar symbolic models for incrementally stable switched
systems. IEEE Transactions on Automatic Control, 55(1):116–126, January 2010.
[HM09]
L. Huang and X. Mao. On input-to-state stability of stochastic retarded systems with Markovian switching. IEEE
Transactions on Automatic Control, 54(8):1898–1902, August 2009.
[JP09]
A. A. Julius and G. J. Pappas. Approximations of stochastic hybrid systems. IEEE Transaction on Automatic
Control, 54(6):1193–1203, June 2009.
[JZst]
P. Jagtap and M. Zamani. Quest: A tool for state-space quantization-free synthesis of symbolic controllers. In
International Conference on Quantitative Evaluation of Systems, pages 309–313. Springer International Publishing,
2017, https://www.hcs.ei.tum.de/en/software/quest/.
[KM13]
I. Kolmanovsky and T. Maizenberg. Stochastic optimal control of jump diffusion excited energy harvesters. In 2013
American Control Conference, pages 5049–5055, June 2013.
[MPS95]
O. Maler, A. Pnueli, and J. Sifakis. On the synthesis of discrete controllers for timed systems. In Annual Symposium
on Theoretical Aspects of Computer Science, pages 229–242. Springer, 1995.
[ØS05]
B. Øksendal and A. Sulem. Applied Stochastic Control of Jump Diffusions. Universitext. Springer-Verlag, Berlin,
2005.
[PGT08]
G. Pola, A. Girard, and P. Tabuada. Approximately bisimilar symbolic models for nonlinear control systems. Automatica, 44(10):2508–2516, October 2008.
[PPD15]
G. Pola, P. Pepe, and M. D. Di Benedetto. Symbolic models for timevarying timedelay systems via alternating
approximate bisimulation. International Journal of Robust and Nonlinear Control, 25(14):1099–1239, 2015.
[PPDBT10] G. Pola, P. Pepe, M. D. Di Benedetto, and P. Tabuada. Symbolic models for nonlinear time-delay systems using
approximate bisimulations. Systems & Control Letters, 59(6):365–373, June 2010.
[Sha13]
L. Shaikhet. Lyapunov functionals and stability of stochastic functional differential equations. Springer Science &
Business Media, 2013.
[Sko09]
A. Skorokhod. Asymptotic methods in the theory of stochastic differential equations. American Mathematical Soc.,
2009.
[Tab09]
P. Tabuada. Verification and control of hybrid systems: a symbolic approach. Springer Science & Business Media,
2009.
[ZA14]
M. Zamani and A. Abate. Approximately bisimilar symbolic models for randomly switched stochastic systems.
Systems & Control Letters, 69:38–46, July 2014.
[ZAG15]
M. Zamani, A. Abate, and A. Girard. Symbolic models for stochastic switched systems: A discretization and a
discretization-free approach. Automatica, 55:183–196, May 2015.
[ZMEM+ 14] M. Zamani, P. M. Esfahani, R. Majumdar, A. Abate, and J. Lygeros. Symbolic control of stochastic systems via
approximately bisimilar finite abstractions. IEEE Transactions on Automatic Control, 59(12):3135–3150, December
2014.
[ZTA17]
M. Zamani, I. Tkachev, and A. Abate. Towards scalable synthesis of stochastic control systems. Discrete Event
Dynamic Systems, 27(2):341–369, June 2017.
1 Department
of Electrical and Computer Engineering, Technical University of Munich, D-80290 Munich, Germany.
E-mail address: {pushpak.jagtap,zamani}@tum.de
URL: http://www.hcs.ei.tum.de
| 3cs.SY
|
On privacy amplification, lossy compression, and their duality to channel coding
Joseph M. Renes
arXiv:1708.05685v2 [cs.IT] 24 Aug 2017
Institute for Theoretical Physics, ETH Zürich, Switzerland
We examine the task of privacy amplification from information-theoretic and coding-theoretic points of
view. In the former, we give a one-shot characterization of the optimal rate of privacy amplification against
classical adversaries in terms of the optimal type-II error in asymmetric hypothesis testing. The converse
significantly improves on previous bounds based on smooth min-entropy by Watanabe and Hayashi [ISIT
2013] and turns out to be equivalent to a recent formulation in terms of the Eγ divergence by Yang, Schaefer,
and Poor [arXiv:1706.03866 [cs.IT]]. In the latter, we show that protocols for privacy amplification based
on linear codes can be easily repurposed for channel simulation. Combined with known relations between
channel simulation and lossy source coding, this implies that privacy amplification can be understood as
a basic primitive for both channel simulation and lossy compression. Applied to symmetric channels or
lossy compression settings, our construction leads to protocols of optimal rate in the asymptotic i.i.d. limit.
Finally, appealing to the notion of channel duality recently detailed by us in [arXiv:1701.05583 [quant-ph]],
we show that linear error-correcting codes for symmetric channels with quantum output can be transformed
into linear lossy source coding schemes for classical variables arising from the dual channel. This explains
a “curious duality” in these problems for the (self-dual) erasure channel observed by Martinian and Yedidia
[Allerton 2003; arXiv:cs/0408008] and partly anticipates recent results on optimal lossy compression by
polar and low-density generator matrix codes.
1
Introduction
Packing and covering are at the core of most simple information processing primitives. In noisy channel
coding, for instance, where inputs lead to probability distributions over output symbols, one would like to
pack as many of these distributions into the space of all possible output distributions such that no two of them
overlap significantly. This gives an error-correcting code, as the associated inputs can be reliably inferred from
the channel output. Covering is in some sense dual to packing, as the goal is to find a set of distributions
whose empirical average approximates (“covers”) a target distribution. In channel simulation, for instance,
we would like to approximate the channel output for a given input with the minimal possible amount of
additional randomness.
In this paper we examine the simple covering task of privacy amplification, also known as randomness
extraction, both from an information-theoretic as well as a coding-theoretic point of view. The goal of privacy
amplification, originally introduced in [1], is to deterministically transform a given random variable Y , which
may be correlated with Z, into the largest possible new random variable V which is uniformly-distributed and
independent of Z. Regarding Z as information held by an adversary or eavesdropper Eve and Y as the variable
each held by Alice and Bob, privacy amplification can be understood as a means of extracting a random secret
key from information partially correlated with the adversary.
The natural information-theoretic question is how much randomness can be extracted, and the answer
depends on the setting. In cryptography, one is interested in making as few assumptions on the correlations
to the eavesdropper as possible and usually considers constraints formulated in terms of the min-entropy the
adversary has about Y , which is related to the maximal probability of guessing Y [2, 3]. One can also consider
adversaries holding quantum information, as opposed to classical, information, but this will not be our focus.
Instead, we will consider the setting where the complete distribution PY Z is known and Z is classical. In
§3 we give upper and lower bounds on the optimal rate of privacy amplification in a one-shot setting that are
formulated in terms of asymmetric hypothesis testing. In particular, the minimal type-II error of discriminating
between the actual distribution PY Z and an uncorrelated distribution R Y ×Q Z plays an important role (see
Theorem 1), where R Y is the uniform distribution and Q Z is arbitrary. Previous work in [1, 4–7] is based
on the min-entropy or collision entropy, though see also [8] and the very recent [9]. Our converse bound
is reminiscient of the metaconverse in channel coding [10, 11][12, Lemma 4.7] not only in appearance, but
also in that it has an elegant proof and leads to tight, computationally-tractable bounds at finite blocklengths.
The converse makes a substantial improvement to a bound based on the min-entropy [7, Theorem 3], and
turns out to be equivalent to bound recently formulated by Yang, Schaefer, and Poor [9, Lemma 5].
Turning to coding theory, in §4 we show that privacy amplification can be used as a primitive to construct
protocols for channel simulation and lossy compression. First we show that privacy amplification based on
linear functions can be used for simulating the action of a given channel W : X → Y on a given input X , such
that the simulation error in the latter is precisely equal to to the security parameter in the former. The idea
behind the construction, stated in detail in Proposition 1, is to consider privacy amplification of the channel
output Y relative to the input X . If we extend the function from Y to V to be reversible, say g : Y → (T, V ),
1
then only T needs to be transmitted from the encoder to the decoder in order to reconstruct Y by applying
g −1 to T and common randomness V . By considering linear functions, we can immediately infer the size
of T to be |Y |/|V |. For symmetric channels, this is sufficient to achieve the optimal rate of communication
required for the simulation task, provided the amount of common randomness available to the encoder and
decoder is large enough.
As shown in [13, 14], simulating the optimal channel in the rate-distortion function gives a means of
turning channel simulation into lossy compression. Hence, privacy amplification can also be used to perform
lossy compression, the precise details of which are stated in Corollary 2. For sources and distortion functions
symmetric in a certain sense, such as the canonical example of compressing a uniformly-random input and
considering Hamming distortion, our construction achieves the rate-distortion bound.
Finally, §5 shows that lossy compression can be accomplished by repurposing a good error-correcting
code for an appropriate “dual channel” as recently investigated by us in [15]. In particular, suppose that X
is a random variable to be compressed and reconstructed as X 0 according to a given distortion measure, and
W = PX |X 0 is the optimal channel in the associated rate-distortion function. Then we show that if a code C
is good for the dual channel W ⊥ , then there exists a similarly good lossy compression scheme for X , where
the reconstructed X 0 are based on codewords of C ⊥ (see Corollary 3 for precise details). While it happens
that the encoder of the channel code and the decompressor are related, the lossy encoder is unrelated to the
channel decoder. Hence, no guarantees can be made on the efficiency of the lossy compressor even if efficient
channel decoding is known to be possible.
The dual channel usually has a quantum output, but one important exception is the erasure channel.
In this case, for a channel with erasure probability q, the associated lossy compression problem is precisely
the binary erasure quantization considered by Martinian and Yedidia [16], who established that in this case
channel codes can be converted to lossy source codes and vice versa. Thus, we can understand the forward
implication as resulting from the deeper structure of duality of codes and channels. Establishing this more
general relation precisely is one of the main goals of this paper.
2
Mathematical setup
We shall only consider random variables with a finite alphabet and will treat their associated probability
distributions (probability mass functions) as vectors. For a random variable X with alphabet X , we denote
the probability mass function as PX and consider it to be an element of R|X | . Joint distributions are labelled by
all the relevant random variables, and R denotes the uniform distribution. Product distributions are denoted
by ×, e.g. P×Q, which corresponds to the tensor product at the level of the vector representation.
Events and observables can also be treated as elements of R|X | , and in particular the set of tests will be
important for our purposes. These are simply vectors whose entries lie in the interval [0, 1]. For a test Λ and
probability PX , the probability of the test itself will be denoted 〈Λ, PX 〉, which denotes the Euclidean inner
product.
We shall also have occasion to consider quantum states and tests, and this notation can also be employed
in the quantum setting. Probability distributions P are replaced by density operators ρ, positive operators of
unit trace on C|X | , tests by positive operators on the same space whose eigenvalues do not exceed unity, and
the inner product by the Hilbert-Schmidt inner product 〈Λ, ρ〉 = Tr[Λρ].
The variational distance of two distributions P and Q is defined by δ(P, Q) := max0≤Λ≤1 〈Λ, P − Q〉, where
1 denotes the vector of all ones. From this definition it is immediate that δ(P, Q) satisfies the triangle and
data processing inequalities, and it is easy to show that δ(P, Q) = 21 kP − Qk1 .
Given two distributions, the set of all pairs of probabilities achievable by all possible tests forms the testing
region R(P, Q) in the unit square. We shall make use of the lower boundary of this region, given by the function
βα (P, Q) := min{〈Λ, Q〉 : 〈Λ, P〉 ≥ α, 0 ≤ Λ ≤ 1} .
Λ
(1)
This can be interpreted as the minimal type-II error in an asymmetric hypothesis test between P and Q, when
the type-I error is constrained to be smaller than 1−α. From the definition it is immediate that βα satisfies the
data processing inequality. That is, for any stochastic map (channel) W , βα (P, Q) ≤ βα (W (P), W (Q)), since
the optimal test Λ? for βα (W (P), W (Q)) induces a feasible test Λ0 for βα (P, Q) by 〈Λ0 , P〉 = 〈Λ? , W (P)〉.
The optimization in 1 is a linear program (see, e.g. [17]), whose dual formulation is
βα (P, Q) = max{µα − 〈1, S〉 : µP − S ≤ Q, µ ≥ 0, S ≥ 0} .
µ,S
(2)
Complementary slackness conditions for the primal and dual programs lead to the well-known NeymanPearson lemma [18] that the optimal test Λ satisfies Λ(x) = 1 for µP(x) > Q(x), Λ(x) = 0 for µP(x) < Q(x),
2
and the values of Λ(x) for x with µP(x) = Q(x) are chosen so that the type-I error is 1 − α. Here µ, the
optimal value in the dual, is the cutoff (inverse) likelihood ratio for deciding between P and Q; clearly the
optimal S is just S(x) = max{µP(x) − Q(x), 0}. Thus, the region R(P, Q) is the convex hull of the points
(αk , βαk ) obtained by tests of the form Λ(x) = 1 for x such that P(x) ≥ γQ(x) for some γ ≥ 0, and zero
otherwise.
P(x)
P(x)
The quantity βα (P, Q) is equivalent to the divergence Eγ (P, Q) := P[ Q(x) ≥ γ] − γQ[ Q(x) ≥ γ] in that
Eγ (P, Q) = α(γ) − γβα(γ) (P, Q) ,
P(x)
(3)
P(x)
where α(γ) = P[ Q(x) ≥ γ] [19, Theorem 21]. To see this, note that (1) implies βα(γ) (P, Q) ≤ Q[ Q(x) ≥ γ], and
therefore Eγ is upper-bounded
by the righthand side of (3). On the other hand, with µ = 1/γ in (2), it follows
P
that γβα(γ) ≥ α(γ) − x:P(x)≥γQ(x) P(x) − γQ(x) = α(γ) − Eγ (P, Q). Rerunning the argument but leaving the
optimization over S in the dual implies that Eγ (P, Q) = maxΛ 〈Λ, P〉 − γ〈Λ, Q〉. In this context we also mention
the bound
α
(4)
βα (P, Q) ≤ ,
γ
P(x)
P(x)
which holds for γ such that P[ Q(x) ≥ γ] ≥ α (cf. [19, Equation 2.68]). To derive it, note that Λ = 1[ Q(x) ≥ γ] is
P
P
feasible for βα (P, Q) by assumption, and therefore βα (P, Q) ≤ 〈Λ, Q〉 = x:Q(x)≤P(x)/γ Q(x) ≤ γ1 x:Q(x)≤P(x)/γ P(x) =
α/γ.
The mean and variance of the log-likelihood ratio log P(x)/Q(x) are also the important quantities in
the large-n asymptotics of testing between P ×n and Q×n . For any distributions P and Q, let D(P, Q) =
P
P(x)
x P(x) log Q(x) be the average of the log-liklihood ratio under P, i.e. the relative entropy, and V (P, Q) =
P
P(x) 2
P(x)
log
− D(P, Q)2 its variance. Stein’s lemma [20, Theorem 2]1 gives the first-order behavior for
x
Q(x)
large n, and following [11, Lemma 58] and [21, §VI] we have the following second-order refinement.
P
P(x)
Lemma 1. Suppose P and Q are any two distributions with finite D(P, Q), V (P, Q), and t(P, Q) = x P(x)(log Q(x) −
D(P, Q))3 . Then, for α ∈ (0, 1) and η ∝
of the unit normal Gaussian,
p1
n
such that α ± η ∈ (0, 1) and Φ the cumulative distribution function
− log βα±η (P ×n , Q×n ) = nD(P, Q) +
Æ
nV (P, Q) Φ−1 (α) + O(log n) .
(5)
Proof. Start with Lemma 58 of [11], which is based on the Berry-Esseen refinement to the central limit theorem. For b = 6V (P, Q)−3/2 t(P, Q), it establishes the bounds
Æ
b+"
log βα (P ×n , Q×n ) ≥ −nD(P, Q) − nV (P, Q) Φ−1 α − p
and
(6)
+ log " − 21 log n
n
Æ
b
,
(7)
log βα (P ×n , Q×n ) ≤ −nD(P, Q) − nV (P, Q) Φ−1 α + p
n
p
for any " > 0, provided the argument of Φ−1 is in (0, 1). Now set η = c/ n for some c > 0 and make the
p
p −1
p
replacement α → α + c/ n. By Taylor’s theorem, we have nΦ (α + pan ) = nΦ−1 (α) + a(Φ−1 )0 (ξ) for some
ξ ∈ [α, α + pan ], so the two bounds are equal to O(log n).
It is not difficult to see that the variational distance of P and Q is the length of the longest vertical line segment one can place inside above (or below) the diagonal inside the testing region. The following proposition
gives bounds on βα (P, Q) just in terms of α and δ(P, Q).
Lemma 2. For any distributions P and Q and α ∈ [0, 1],
α − δ(P, Q) ≤ βα (P, Q) ≤ α (1 − (1 − α)δ(P, Q)) .
(8)
Proof. Supposing Λ is the optimal test in βα (P, Q), it follows immediately that δ(P, Q) ≥ 〈Λ, P − Q〉 = α −
βα (P, Q). For the upper bound, let Γ be the optimal test in δ(P, Q) and set α? = 〈Γ , P〉. By this definition we
have δ(P, Q) = 〈Γ , P − Q〉. Next, set Λ = a1 + bΓ for a and b to be determined later, but such that 〈Λ, P〉 = α,
i.e. a + bα? = α. It then follows that βα (P, Q) ≤ 〈Λ, Q〉 = α − bδ(P, Q). Now there are two cases to consider, α
smaller or larger than α? . For the former, the choice a = 0 and b = α? /α ensures that Λ is a valid test. Then
βα (P, Q) ≤ α(1 − α1? δ(P, Q)), which implies the desired bound since 1/α? ≥ 1 ≥ 1 − α. For the latter, choosing
1−α
b = 1 − a and a = (α − α? )/(1 − α? ) again leads to a valid test. Here we have βα (P, Q) ≤ α(1 − 1−α
? δ(P, Q))
1
which implies the desired bound since 1−α? ≥ 1 ≥ α.
1
Chernoff attributes the result to Stein in [20].
3
Note that α? = P[P(x) ≥ Q(x)] appearing in the proof is precisely the value of α for which the vertical
distance between the diagonal and the lower boundary of R(P, Q) is the variational distance.
The lower bound corresponds to the upper bound in [22, Proposition 3.2], while the upper bound is due
to Frédéric Dupuis. We mention in passing that the lower bound claimed in Proposition 3.2 is in error. In
particular, for P and Q deterministic and completely uniform distributions of a binary-valued random variable,
respectively, we have δ(P, Q) = 21 and βα (P, Q) = α2 . This violates the ostensible bound for α < 1/2. This
1
1
example also shows the claimed Pinsker-like inequality 1−α
α δ(P, Q) ≤ − log α βα (P, Q) is violated for α < /3.
3
Bounds on extractable randomness
Given a joint distribution PY Z , the task of randomness extraction, or privacy amplification, of Y relative to
Z is to apply a function f : Y → V such that the resulting distribution PV Z is essentially the same as R V ×PZ .
This setup is depicted in Figure 1. We sometimes refer to f as the extractor or extractor function, though note
that in the cryptography community an extractor refers to a set of functions useful for generating randomness
from a source characterized only in terms of min-entropy. We measure closeness by the variational distance,
and say that f is a protocol for (k, ") privacy amplification for PY Z when δ(PV Z , R V ×PZ ) = " and log |V | = k.
PY Z
Y
f
RV
V
V
≈"
PZ
Z
Z
Figure 1: Schematic representation of privacy amplification (randomness extraction) of Y relative to
Z. The function f should produce a random variable V which is "-close to being uniformly random
and independent of Z, as measured by the variational distance.
Letting K" (Y |Z) P be the largest K ∈ N such that there exists a (log K, ") privacy amplification protocol for
PY Z , we can show the following result.
Theorem 1. For any joint distribution PY Z ,
K" (Y |Z) P ≤
min
K" (Y |Z) P ≥ max
η∈[0,"]
1
η β"+η (PY Z ,
1Y ×PZ )
and
4η2
"−η max β"−η (PY Z , 1Y × Q Z ) .
η∈[0,1−"]
QZ
(9)
(10)
We will show the second inequality, the achievability statement, by employing two-universal hashing [23].
Thus, linear functions are capable of achieving the stated bounds. The argument is a combination of the argument given for channel resolvability by Hayashi for channels with classical [24, Lemma 2] and quantum [12,
Lemma 9.2] output with that of the leftover hashing lemma of [25], adapted to yield a bound involving βα .
The first inequality, the converse, is an adaptation of the converse involving min-entropy reported in [21,
§III.A], itself based on [26, §8.2.2]. Both the underlying achievability and converse arguments also apply
when the adversary holds quantum information (i.e. Z is quantum), but this is decidedly not the case for the
hypothesis testing versions. We shall remark on the steps that fail for quantum side information.
3.1
Proof of the converse
Let us begin by showing the converse. It relies on the following lemma, the hypothesis-testing version of the
statement that min-entropy of Y conditioned on Z cannot increase by the application of a function to Y .
Lemma 3. For any function f : Y → V , βα (PV Z , 1V ×PZ ) ≤ βα (PY Z , 1Y ×PZ ) for all α ∈ [0, 1].
Proof. Using the conditional distributions PY |Z=z we can construct a stochastic map W from PV Z back to PY Z :
W ( y, z|v, z 0 ) = δz,z 0 P
4
δ f ( y),v PY |Z=z ( y)
y0
δ f ( y 0 ),v PY |Z=z ( y 0 )
.
(11)
Clearly W is stochastic and W (PV Z ) = PY Z . By the data processing inequality we have βα (PV Z , 1V ×PZ ) ≤
βα (PY Z , W (1V ×PZ )). Now observe that
X
[W (1V ×PZ )]( y, z) =
W ( y, z|v, z 0 )PZ (z 0 )
(12a)
v,z 0
= PZ (z) P
≤ PZ (z) .
PY |Z=z ( y)
y0
δ f ( y), f ( y 0 ) PY |Z=z ( y 0 )
(12b)
(12c)
Hence βα (PY Z , W (1V ×PZ )) ≤ βα (PY Z , 1Y ×PZ ), completing the proof.
Now suppose f : Y → V is the extractor function of a (log K" (Y |Z) P , ") privacy amplification protocol. By
Lemma 2, we have β"+η (PV Z , R V ×PZ ) ≥ η for any 0 ≤ η ≤ 1 − ". This is equivalent to β"+η (PV Z , 1V ×PZ ) ≥
η|V | = ηK" (Y |Z) P , whence Lemma 3 and a minimization over η gives (9).
Lemma 3 does not hold for quantum Z, as one can find counterexamples. That this might be the case
can be anticipated by noticing that in the proof we make use of the distribution of Y conditioned on the
value of Z, for which there is no quantum analog. To state a specific counterexample, consider the qubit
density operators ϕ j = | j〉〈 j| for j = 0, 1 and {| j〉} an orthonormal basis, along with ϕ2 = |+〉〈+|, where
|+〉 = p12 (|0〉 + |1〉). Suppose Y is uniformly-distributed in {0, 1, 2} and take the state of Z given Y = y to
be ϕ y ; the marginal state of Z averaged over Y is ϕ̄ = 31 (1 + |+〉〈+|) For the function f mapping 0 and 1
to 1 and 2 to itself, then V has distribution (2/3, 1/3) and corresponding conditional states 12 1 and ϕ2 . In the
lower bound of βα (PV Z , 1V ×PZ ) we can choose µ = 2 and S = 31 |−〉〈−| so that βα (PV Z , 1V ×PZ ) ≥ 2α − 31 . In
the upper bound of βα (PY Z , 1Y ×PZ ) we can take the test for Y = y to be a scaled projection onto the zero
9 −3
1 −3
eigenspace of 49 ϕ y − ϕ̄ for Y = 0, 1 and zero for Y = 2. In particular, Λ0 = α6 −3
and Λ0 = α6 −3
9 , and
1
4
4
1
Λ3 = 0. Then 〈Λ, ρY Z 〉 = α and 〈Λ, 1Y ⊗ ρ Z 〉 = 3 α, meaning βα (PY Z , 1Y ×PZ ) ≤ 3 α. For α > /2 this is smaller
than 2α − 31 , so this example shows that the lemma cannot hold for arbitrary α. It is an open question if it
holds for α < 1/2.
3.2
Proof of achievability
Now we move to the proof of the direct part. Let f : Y → V be an arbitrary function with |V | = K" (Y |Z) P , and
set ∆ := δ(PV Z , R V × PZ ). For arbitrary test ΛY ZPon Y Z, define the rescaled probability distributions P̂V Z and
P̄V Z such that P̂V Z + P̄V Z = PV Z and P̂V Z (v, z) = y: f ( y)=v PY Z ( y, z)ΛY Z ( y, z). Using the triangle inequality we
have
∆ ≤ +δ(PV Z , R V × P̄Z ) + δ(R V × P̄Z , R V ×PZ )
= δ( P̂V Z + P̄V Z , R V × P̄Z ) + δ( P̄Z , PZ )
≤ δ( P̄V Z , R V × P̄Z ) + δ( P̄Z , PZ ) +
1
2 k P̂V Z k1
(13a)
(13b)
(13c)
Due to the form of P̄V Z and P̂V Z , the latter two terms combine to give 〈ΛY Z , PY Z 〉, so that
∆ ≤ 21 k P̄V Z − R V × P̄Z k1 + 〈ΛY Z , PY Z 〉 .
(14)
P
P
To bound
fact that ( i |x i |)2 ≤ i si−1 x i2 for any set of si > 0
P the first term in this expression, we use the
p
x
such that i si = 1. This follows by writing kxk1 = k ps sk1 and applying the Hölder inequality with p =
q = 2. Choosing s vz = Q Z (z)/|V | for some normalized distribution Q Z with strictly positive probabilities to be
determined later, we have
X
k P̄V Z − R V × P̄Z k21 ≤ |V |
Q Z (z)−1 ( P̄V Z (v, z) − |V1 | P̄Z (z))2
(15a)
= |V |
vz
X
vz
Q Z (z)−1 P̄V Z (v, z)2 −
X
z
Q Z (z)−1 P̄Z (z)2 .
(15b)
Now let us deal with the summation over v in the first term. With Λ0 ( y, z) = 1 − Λ( y, z) and omitting the Y Z
random variable subscripts, we have
X
X X
X
P̄V Z (v, z)2 =
P( y, z)Λ0 ( y, z)
P( y, z)Λ0 ( y, z)
(16a)
v
v
=
X
y
y: f ( y)=v
P( y, z)2 Λ0 ( y, z)2 +
y 0 : f ( y 0 )=v
X
y6= y 0
5
δ f ( y)= f ( y 0 ) P( y, z)P( y 0 , z)Λ0 ( y, z)Λ0 ( y 0 , z) .
(16b)
Taking the expectation over f chosen uniformly at random from a family of two-universal hash functions, we
then obtain
X
X
X
P̄V Z (v, z)2 =
P( y, z)2 Λ0 ( y, z)2 + |V1 |
P( y, z)P( y 0 , z)Λ0 ( y, z)Λ0 ( y 0 , z)
(17a)
Ef
v
y
≤
y6= y 0
X
0
P( y, z) Λ ( y, z) +
2
y
2
1
2
|V | P̄Z (z) .
(17b)
Using this in (15) gives
E f k P̄V Z − R Z × P̄Z k21 ≤ |V |
X
yz
Q Z (z)−1 PY Z ( y, z)2 Λ0Y Z ( y, z)2 .
(18)
Finally, let ΛY Z be the optimal test in βη (PY Z , 1Y × Q Z ), so that 〈ΛY Z , PY Z 〉 = η and βη (PY Z , 1Y × Q Z ) ≤
µη for the optimal µ in the dual formulation (2). By the properties of the optimal test it follows that
Λ0Y Z ( y, z)(µPY Z ( y, z) − Q Z (z)) ≤ 0. Therefore, Λ0Y Z ( y, z)PY Z ( y, z)Q Z (z)−1 ≤ µ1 Λ0Y Z ( y, z). From (18) we then
obtain
E f k P̄V Z − R V × P̄Z k21 ≤
≤
≤
|V | X
PY Z ( y, z)Λ0Y Z ( y, z)2
µ yz
(19a)
|V |
µ
(19b)
η|V |
.
βη (PX Y , 1X × Q Y )
(19c)
By Jensen’s inequality, E f k P̄V Z − R V × P̄Z k21 ≥ (E f k P̄V Z − R V × P̄Z k1 )2 . Returning to (14), we have
v
u
η|V |
1t
Ef ∆ ≤ η +
.
2 βη (PY Z , 1Y × Q Z )
(20)
4δ
maxQ Z β"−δ (PY Z , 1Y ×Q Z )c ensures there exists an f such that ∆ ≤ ". Since
Choosing η = "−δ and |V | = b "−δ
βα (P, Q) is continuous in Q, we may maximize over all Q Z , not just those with strictly positive probability. This
completes the proof.
Were Z quantum rather than classical, the test ΛY Z would take the form of a set of positive operators Λ y ,
and we would immediately be confronted with the possibility that Λ y may not commute with the conditional
quantum states of Z given Y = y, i.e. the quantum versions of PZ|Y = y . In the proof we assume these objects
commute, e.g. in (16). One method of dealing with this issue, as done in [12], is to “pinch” the quantum
states (remove the off-diagonal elements) to restore commutation, and appeal to bounds between the pinched
and unpinched states. It is unclear if this can be done in combination with the steps taken here to end up
with a bound in terms of βα .
2
3.3
Analysis of the bounds
The first- and second-order asymptotics of log K" (Y n |Z n ) now follow immediately from Lemma 1. The firstorder behavior, i.e. the optimal rate, follows from [27, Theorem 1] by setting X = Y . The second-order
behavior was originally shown in [7, Theorem 3], starting from a one-shot bound in terms of the smooth
min-entropy (see also [6]).
Corollary 1. For any PY Z and " ∈ (0, 1),
log K" (Y n |Z n ) P ×n = nH(Y |Z) +
Proof. In the upper bound, set η =
1
η
p1 .
n
Æ
nV (Y |Z) Φ−1 (") + O(log n) .
(21)
Lemma 1 applies for n large enough so that " + η < 1, and the factor
disappears when taking the limit n → ∞. Likewise, in the lower bound we can choose Q Z = PZ and η =
so that Lemma 1 again applies for n large enough such that η ≤ ".
p1
n
After publication of the initial version of this manuscript, Wei Yang pointed out that the bounds contained
herein are related to those recently derived in [9]. The achievability bound can be derived from [9, Lemma
6
2], which states that there exists an (n, K, ") privacy amplification scheme such that, for all distributions Q Z
and γ > 0,
v
t γK
" ≤ Eγ (PY Z , R Y ×Q Z ) +
E P [exp(−|i(Y :Z) − log γ|)] ,
(22)
4|Y | Y Z
where i( y, z) = log PY Z ( y, z) − log R Y ( y)Q Z (z). To get back to (10), observe that the expectation term is
P Z ( y,z)
P Z ( y,z)
necessarily smaller than 1 and Eγ ≤ PY Z [ RY (Y y)Q
≥ γ]. Now set α = PY Z [ RY (Y y)Q
≥ γ] and use (4) to get
Z (z)
Z (z)
p
" ≤ α + 12 αK/βα (PY Z , 1Y ×Q Z ), which is (20).
The converse in (9) is equivalent to that of [9, Lemma 5], which states that every (n, K, ") privacy amplification protocol for PY Z must satisfy
" ≥ E|Y |/K (PY Z , R Y ×PZ ) .
(23)
To obtain (9) from this expression, use the variational form of Eγ to write " ≥ maxΛ 〈Λ, PY Z 〉 −
P(x)
|Y |
K 〈Λ, R Y ×PZ 〉.
Now consider tests that yield the vertices of R(PY Z , R Y ×PZ ), specifically Λk = 1[ Q(x) ≥ γk ] such that αk =
〈Λk , PY Z 〉 is larger than ". Using βαk = 〈Λk , R Y ×PZ 〉 and rearranging the inequality, we have |Y |βαk (PY Z , R Y ×PZ ) ≥
(αk − ")K. The function α 7→ βα interpolates linearly between vertices, and therefore the relation holds for
arbitrary α. Setting α − " = η recovers (9).
To show that (9) is equivalent, we use (23) to choose a suitable η in the optimization. Specifically, suppose
K ? is the optimizer in (23) so that " = E|Y |/K ? (PY Z , R Y ×PZ ), and let γ? = |Y |/K ? . Now define η = α − " for
P
( y,z)
α = PY Z [ RY Y( Zy)PZ (z) ≥ γ? ]. Clearly η ≤ 1 − " since α must be positive. On the other hand, η must be positive
P
( y,z)
since we have " = α − γ? (R Y ×PZ )[ RY Y( Zy)PZ (z) ≥ γ? ], the second term of which is positive. Indeed, by (3), we
must have " = " + η − γ? β"+η (PY Z , R Y ×PZ ). Rearranging the expression gives |Y |β"+η (PY Z , 1Y ×PZ ) = ηK ? ,
and therefore K" (Y |Z) P ≤ K ? . Since we have K" (Y |Z) P ≤ K ? from the previous argument, this implies that
the choice of η is optimal.
log K" (Y n |Z n )
0.45
0.4
extraction rate
1
n
0.35
0.3
upper bound [7, (10)]
upper bound (9) & [9, (154)]
normal approximation (21)
0.25
lower bound [9, (150)]
lower bound (10) & [7, (9)]
0.2
2000
4000
6000
8000
10000
blocklength n
Figure 2: Comparison of finite blocklength bounds on randomness extraction from Y n relative to Z n
for the i.i.d. case of Z obtained from uniform Y through a BSC of crossover probability 0.11 and a
target security parameter of " = 10−10 . The asymptotic rate for this example is 1/2. The shaded region
indicates the improvement of (9) over [7, (10)].
To compare the bounds for a fixed example, consider Z obtained from uniform Y through a binary symmetric channel with crossover probability 0.11 and a target security parameter of " = 10−10 . We follow [11,
Theorem 35] in computing βα (PY×nZ , 1Y n ×R Z n ). Assuming η = "/2 for the auxiliary variables and Q Z uniform
in (10), it turns out that the lower bound of (10) and that of [7, Equation 9] are essentially identical. Our
(10) always has the advantage, though it is miniscule: an increase in |V | by no more than 10 for the range
100 ≤ n ≤ 10000. As shown in Figure 2, Theorem 18 in [9]2 is considerably better. On the converse side,
2
Equations 150 and 153 contain a small error: The term g n (γ) inside the square root should be g n (γ)2 (Wei Yang, private communication).
7
(9) substantially improves over [7, Equation 10]. It is satisfying to see the normal approximation obtained
by disregarding the O(log n) term in (21) lies midway between the upper and lower bounds. Note that in this
example the O(log n) term is provably absent from the asymptotic expansion, as shown in [9, Theorem 19].
4 Repurposing randomness extraction
4.1 Channel simulation
Given a channel W : X → Y , the task of channel simulation is to reproduce the joint input and output
statistics of W applied to X by making use of an ideal (noiseless) channel and common randomness at the
encoder and decoder. Randomness is of course necessary to simulate the stochastic nature of W , and the main
information-theoretic question is to characterize the amounts of communication and randomness required for
a given channel W and input X . A stronger variant is universal channel simulation, in which the simulation
works for any input X , i.e. the encoder is not constructed using the knowledge of the particular input X ; in
this paper we are only interested in the non-universal case.
Let us first specify the setup more concretely. Given an input distribution PX , the channel W leads to a
joint distribution PX Y . An (m, r, ") protocol for simulation of W applied to X consists of an encoding map
E : (X , V ) → T and a decoding map D : (T, V ) → Y 0 , such that log |T | = m, log |V | = r, and δ(PX Y , PX Y 0 ) ≤ ",
where Y 0 = D(E(X , V ), V ) and V is a uniformly-distributed random variable.
PX
X
E
D
T
V
RV
≈"
Y0
PX
X
W
Y
V
Figure 3: Simulation of the channel W acting on input random variable X by means of an ideal
channel and common randomness.
The following shows that privacy amplification can be repurposed for channel simulation (see also [28]).
Proposition 1. Let W : X → Y be an arbitrary channel and PX an arbitrary input distribution. Suppose the
linear function f is a protocol for (k, ") privacy amplification of Y relative to X . Then a (log |Y |− k, k, ") protocol
for simulating the action of W on X can be constructed from f and PX Y .
Proof. The linear function f can always be extended to a reversible linear function g : Y → (T, V ), and given
the joint distribution PX Y , g induces a conditional distribution PT |X V . Interpreted as a channel, PT |X V defines
the encoder E. It requires log |V | bits of random input and produces a message of log |T | = log |Y |/|V | bits.
Meanwhile, the decoder D is just g −1 , calling the output Y 0 instead of Y .
By design, E(V, X ) = T , meaning the combined action of D and E on PX V gives back PX Y . In the protocol,
this combined action is instead applied to PX × R V , producing PX Y 0 . By the data processing inequality we
therefore have δ(PX Y , PX Y 0 ) ≤ δ(PX V , PX × R V ). But the premise is that the latter quantity is bounded from
above by ", so the proof is complete.
Since we are considering linear f , it makes sense to associate the function to a linear code. Suppose we
consider f to output the syndromes of some code C encoding n − k bits into n, i.e. f ( y n ) = H y n , where H
is the k × n parity check matrix of C. Then, if g is to be reversible, the output in T must correspond to the
message encoded in C, and the action of g −1 is therefore to produce a codeword of C determined by t, offset
to a coset determined by v. To see this more concretely, extend H to an invertible n × n matrix M which
defines g, say M = M̄
with M̂ = H. The reconstruction operation will apply M −1 , so that y n = M −1 (t ⊕ v).
M̂
0
0
−1 T
Now let M = (M ) and set M 0 = M̄
for M̄ 0 an (n − k) × n matrix. In terms of M 0 , the action of g −1 is
M̂ 0
given by ( y n ) T = t T ⊕ v T )M 0 = t T M̄ 0 ⊕ v T M̂ 0 . Since M M 0T = 1, H M̄ 0T = 0, meaning M̄ 0 is the generator
matrix of C, and the action of g −1 is as claimed.
For channels W whose input X induces a uniformly-random output Y , the above construction directly
leads to protocols which achieve the optimal communication rate of channel simulation in the asymptotic
i.i.d. scenario, assuming unlimited common randomness. The amount of communication required is |T n | =
|Y n |/K" (Y n |X n ), and so in the i.i.d. case with uniform Y we have limn→∞ 1n log |T n | = 1n (log |Y n |−H(Y n |X n )) =
I(X :Y ) while " → 0 (cf. [29, Theorem 2]). To achieve the mutual information rate in the general case of a
non-uniform Y , one option is to concatenate the above protocol with data compression of Y n , though we will
not pursue this further here (for a similar approach applied to noisy channel communication, see [30]).
8
PX
X
E
C
D
PX
X0
Λ
Z
0
X
≈"
X
Λ
Z
X
Figure 4: Lossy compression of the random variable X . Λ is a binary-valued test that indicates 1 if
d(x, x 0 ) ≥ d ? and zero otherwise.
4.2
Lossy source coding
The task of lossy source coding is to compress a random variable X to a smaller alphabet such that the reconstructed X 0 is close to X as measured by a distortion function d : X ×X 0 → R+ . There are two main approaches
to specifying the distortion constraint, by requiring either a fixed average value or a fixed probability of exceeding a specified target value. In the latter case one defines a (k, d ? , ") protocol for lossy compression of
X ∼ PX to consist of an encoder E : X → C and a decoder D : C → X 0 such that P[d(X , D ◦ E(X )) > d ? ] ≤ "
and log |C| = k. This excess distortion probability can be expressed as the expectation under PX X 0 of a test
function Λ which is 1 whenever d(x, x 0 ) > d ? and zero otherwise, i.e. P[d(X , D ◦ E(X )) > d ? ] = 〈Λ, PX X 0 〉.
This setup is depicted in Figure 4. In the former case, a (k, d̄) protocol consists of an encoder and decoder
such that E PX X 0 d(X , X 0 ) = d̄. The latter is a stronger criterion, since any "-good scheme with target value d ?
can be converted into an average scheme with average distortion less than d ? + " max x,x 0 d(x, x 0 ) (see, e.g.
[31, Lemma 6]).
Channel simulation is one way to construct a lossy compression protocol, as shown in [13, 14] (and later
in the quantum setting in [31–33]). For the case of excess distortion probability, we formalize the statement
as follows.
Proposition 2. Given PX , a distortion function d and a target distortion d ? , suppose that W : X → X 0 is a
channel such that the target distortion for X and X 0 is exceeded with probability no larger than " 0 . Then a (k, ")
protocol for channel simulation of W applied to X can be used to construct a (k, d ? , " + " 0 ) lossy compression
protocol for X .
Proof. Let X 00 be the output of the channel simulation protocol, for which δ(PX X 0 , PX X 00 ) ≤ ". By the properties
of the variational distance, δ(PX X 00 , PX X 0 ) ≥ 〈Λ, PX X 00 〉 − 〈Λ, PX X 0 〉, and therefore the probability 〈Λ, PX X 00 〉 that
X and X 00 exceed the target distortion is no larger than " + " 0 .
As opposed to channel simulation, in lossy compression there is no need for randomized encoding or
decoding operations. Simulation protocols adapted to lossy compression can in principle be derandomized:
Since the output distribution PX X 00 is a mixture over the values of V and the distortion probability is linear
in PX X 00 , one may as well pick that value of V leading to the smallest excess distortion probability. This still
leaves the randomness in the mapping PT |X =x,V =v , but this can be derandomized in the same manner.
In the i.i.d. setting it is perhaps more natural to consider the case of average distortion
directly, since
P
here we are interested in symbolwise distortion functions of the form d(x n , x 0n ) = 1n i dsym (x i , x i0 ) for some
function dsym , as well as constructing protocols by simulating single symbol channels W : X i → X i0 (rather
than W : X n → X 0n ). In this case we have
Proposition 3. Given PX×n and a symbolwise distortion function d with 0 ≤ dsym (x i , x i0 ) ≤ 1 for all x i , x i0 , suppose
that W : X → X 0 is a channel such that E PX dsym (X , X 0 ) = d̄ ≤ 1. Then a (k, ") simulation protocol for W ×n acting
on X n can be used to construct a (k, d̄ + ") lossy compression scheme for X n .
Proof. Let X 00n be the output of the simulation protocol. Since the protocol is "-good, δ(PX i X i0 , PX i X i00 ) ≤ " for
all i ∈ [n]. But by the definition of the variational distance,
X
δ(PX i X i0 , PX i X i00 ) = max
f (x, y)(PX i X i00 (x, y) − PX i X i0 (x, y))
(24a)
0≤ f ≤1
≥
X
x, y
x, y
dsym (x, y)(PX i X i00 (x, y) − PX i X i0 (x, y))
= E PX X 00 dsym (X i , X i00 ) − d̄ .
i i
Thus E PX X 00 dsym (X i , X i00 ) ≤ d̄ + ", and averaging over i gives the desired result.
i i
9
(24b)
(24c)
In the i.i.d. setting the optimal rate is given by the rate-distortion function R(d̄) = infW :d(X ,W (X ))≤d̄ I(X :
W (X )). When the optimal channel W in this expression gives a uniformly-random X 0 , then we can employ an
optimal rate channel simulation protocol to construct an optimal rate lossy compression procedure. Hence,
in this case we can ultimately rely on privacy amplification of X 0 relative to X by linear functions to perform
lossy compression of X at the optimal rate. For fixed blocklength we have the following corollary:
Corollary 2. In the setting of Proposition 3, suppose PX X 0 is a joint distribution such that E PX X 0 dsym (X , X 0 ) ≤ d̄.
Then any linear (k, ") privacy amplification protocol for X 0n relative to X n can be used to construct a (n− k, d̄ +")
lossy compression scheme for X n .
The standard example of a uniformly-random source X and Hamming distortion dsym (x, x 0 ) = δ x,x 0 has
a uniform X 0 , since the optimal channel is just a binary symmetric channel with crossover probability d̄ [34,
§10.3.1]. So, too, does the binary erasure quantization example of Martinian and Yedidia [16]. Here the
input X has alphabet {0, 1, ?}, with probabilities (1 − e)/2, (1 − e)/2 and e for some 0 ≤ e ≤ 1, respectively,
and the symbol distortion function dsym (x, x 0 ) is 0 if x =? or x = x 0 , and 1 otherwise. As reported by Kostina
d̄
))
and Verdú in [35, Equation 202], the rate distortion function for this case is R(d̄) = (1 − e)(1 − h2 ( 1−e
(after adapting the notation to the present setting). The optimal channel is a concatenation of the map which
randomly assigns ? inputs to 0 and 1, but leaves those input values untouched, and a binary symmetric channel
with crossover probability d̄/(1 − e). It therefore has uniform output over {0,1} for any value of e.
5
Lossy compression from channel coding
By making use of duality relations for channels and codes, we can show that linear error-correcting codes
can be used to build lossy source codes. Suppose X is the random variable to be lossily compressed, and
PX X 0 is the optimal joint distribution in the rate distortion function. This induces a channel W = PX |X 0 given
the marginal PX 0 ; note that this channel is defined in the opposite sense to §4.1. Corollary 2 establishes
that lossy compression can be constructed from privacy amplification of the input of W relative to its output.
But, following [15, 36], this task is dual to channel coding (or lossless compression) for the dual channel
W ⊥ : X 0 → B, where now W ⊥ is a channel whose output is quantum-mechanical (see [15, §3.2] for a precise
definition). Here we restrict attention to symmetric channels W .
Specifically, Corollary 8 of [15] ensures that a (k, ") code C for W ⊥ (where " is the average error probability
p
under the optimal decoder) leads to a linear (k, 2 ") privacy amplification protocol for X 0 relative to X . The
extractor function f is given by the generator matrix G of C acting to the right, i.e.p
the parity check matrix of
C ⊥ . In fact, using Theorem 5.1 of [36] we can improve the security parameter to 2" (the difference stems
from the use of the max-entropy in Corollary 8, which involves an optimization over the marginal of Z rather
than using actual marginal PZ directly). Combining this with Corollary 2, we obtain
Corollary 3. For PX |X 0 the optimal conditional distribution appearing in the rate-distortion function, let W ⊥ be
p
the dual channel according to [15]. Then a (k, ") code C for W ⊥ can be used to construct an (n − k, 2") lossy
compression scheme for X n .
The reconstruction operation outputs codewords of C ⊥ , shifted to a coset determined by the common
randomness V . Meanwhile, the quantizer or compressor is stochastic, based on the conditional distribution
distribution PT |X V as in Proposition 1.
The dual channels for the examples mentioned above can be explicitly given. For the case of Hamming
distortion, the optimal channel from X 0 to X is also a BSC with crossover
probability
δ = d̄, which means the
p
p
dual channel takes the classical input z to the pure state |θz 〉 = δ |0〉 + (−1)z 1 − δ |1〉. For binary erasure
quantization, the optimal channel is a concatenation of a BSC with crossover probability δ = d̄/(1 − e) with
an erasure channel with erasure probability e. The dual of this channel is computed in Example 3.9 of [37].
Its output consists of two independent parts, one classical and one quantum-mechanical. The classical part
is the just the output of the erasure channel with erasure probability 1 − e, while the quantum part is exactly
the output of the dual of the BSC. Thus, with probability e the input to the channel shows up unchanged in
the classical part of the output, but even when the classical part is useless, the quantum part contains some
information about the input.
Note that when d̄ = 0 in the latter case, the two quantum states |θz 〉 are identical, so the quantum part
of the dual output is useless. Then the dual is effectively just the erasure channel with erasure probability
1 − e. The above relation between error-correcting code C for W ⊥ and the use of the dual code C ⊥ for privacy
amplification of W partly explains the “curious duality between erased/known symbols in source coding and
known/erased symbols in channel coding” observed by Martinian and Yedidia [16]. By more direct analysis
of error-correction and lossy compression for the erasure channel, they show an equivalence between the
10
two tasks. Here we have made use of the more general theory of duality and shown one direction of the
equivalence, namely that error-correction implies lossy compression.
This also makes sense of recent results on the optimality of polar codes and spatially-coupled low-density
generator matrix codes for lossy compression, shown in [38] and [39], respectively. Polar codes are their own
duals in the sense that the dual of a polar code is given by using the frozen bits instead of the information bits,
which is again a code constructed by the properties of synthesized channels. Thus, the fact that polar codes
achieve the capacity for classical-quantum channels described above, which follows from the general result
of [40], implies that polar codes achieve the rate-distortion bound for the associated sources. One simply has
to base the scheme on the synthesized inputs with high entropy, exactly as done in [38]. Similarly, it is not
unreasonable to suspect that spatially-coupled low-density parity check (LDPC) codes achieve the capacity
of these classical-quantum channels under optimal decoding. This would imply that their duals, low-density
generator matrix (LDGM) codes, achieve the rate-distortion bound for the associated sources, precisely as
shown in [39]. Duality also helps explain that LDPC codes are themselves not useful for lossy compression, as
observed in [16], as otherwise their duals, LDGM codes would be good for channel coding and this is known
not to be the case [41]. Note that the implications based on duality say nothing about complexity of encoding
or decoding operations for either channel coding or lossy compression; indeed, establishing low complexity
bounds is the better part of the results of [38] and [39].
6
Discussion
We have given new bounds on the optimal rate of privacy amplification in a one-shot setting and seen that
the converse bound improves substantially on previously-known results. While the achievability bound is not
the tightest known in the literature, the formulation of both in terms of hypothesis testing has advantages of
its own. One is the clear connection of information theory to statistics, which in particular immediately gives
the asymptotic expansion of the optimal rate to second order. More conceptually, using a common quantity in
optimal rate bounds allow us to see the concrete relationship of covering and packing problems more plainly.
Namely, for approximation parameter ", rate bounds for covering problems involve β" , while those for packing
problems involve β1−" . This supports the notion that packing is dual to covering. An open question is whether
the hypothesis-testing approach can be extended to covering problems involving quantum information.
We have also shown that privacy amplification is a primitive for constructing channel simulation and
lossy compression protocols. Doing so enables us to extend the known duality of codes for packing (lossless
compression) and covering (privacy amplification) to the covering problem of lossy compression. Specifically,
coding duality implies that duals of good channel codes lead to good lossy source codes, at least for symmetric
channels and lossy compression setups. An immediate open question in this context is whether one can go in
the other direction, from lossy source coding back to channel coding. This was observed to be the case for the
binary erasure channel and binary erasure quantization in [16]. Perhaps the most straightforward approach
to demonstrating this would be to show that lossy source codes can be used for privacy amplification, since
duality ensures that a good privacy amplification protocol implies the existence of a good channel code for the
dual (cf [15, Corollary 8], [36, Theorem 5.2]). One could also investigate whether duality leads to improved
finite blocklength bounds on privacy amplification. However, since the dual setup involves a quantum output,
this seems doubtful as the bounds available in this case are not as tight as for classical output (see [21]).
A much bigger and more tantalizing open question is whether duality also provides a link between algorithms for channel decoding and source quantization, e.g. using belief propagation (BP). Such a link was
shown in Theorem 3 of [16] for the binary erasure coding and quantization problems, and progress on the
general case could help in finding new bounds on the performance of BP. A first step in this direction would be
to extend the notion of BP decoding to the duals of classical channels. This is presumably possible by extending our construction in [42], which dealt with the dual of the BSC, since any symmetric binary-input classical
channel can be regarded as a mixture of BSCs. Though duality was not used in the construction therein, in
retrospect its role is evident. In particular, the unitaries for combining quantum information at check and
variable nodes could have been determined by appealing to the convolution rules for the BSC and Theorem 1
of [15], which states that the dual of a check convolution is the variable convolution of the duals and similarly
for the dual of a variable convolution.3 All of which hints concretely to the possibility that duality can shed
light on BP, but the details remain to be seen.
Acknowledgments. I thank Renato Renner and Marco Tomamichel for helpful discussions. Particular
thanks to Frédéric Dupuis for providing the upper bound of Lemma 2, to Henry Pfister for pointing me to
[16], and Wei Yang for pointing out the relationship between Theorem 1 and the results of [9]. This work
3
Despite my best efforts, an error in the check node convolution in an initial draft of [15] persists in the final published version (though
not the most recent arXiv version): The check node convolution should read Uf = CNOT1→2 . Thanks to Narayanan Rengaswamy for
pointing this out, twice!
11
was supported by the Swiss National Science Foundation (SNSF) via the National Centre of Competence in
Research “QSIT”, as well as the Air Force Office of Scientific Research (AFOSR) via grant FA9550-16-1-0245.
References
[1] C. H. Bennett, G. Brassard, and J.-M. Robert, “Privacy Amplification by Public Discussion”, SIAM Journal on Computing 17, 210–229 (1988).
[2] R. Impagliazzo, L. A. Levin, and M. Luby, “Pseudo-random generation from one-way functions”, in Proceedings of
the twenty-first annual ACM symposium on Theory of computing (1989), pp. 12–24.
[3] N. Nisan and D. Zuckerman, “Randomness is Linear in Space”, Journal of Computer and System Sciences 52, 43–52
(1996).
[4] C. H. Bennett, G. Brassard, C. Crépeau, and U. M. Maurer, “Generalized privacy amplification”, IEEE Transactions
on Information Theory 41, 1915–1923 (1995).
[5] R. Renner and S. Wolf, “Simple and Tight Bounds for Information Reconciliation and Privacy Amplification”, in
Advances in Cryptology - ASIACRYPT 2005, Lecture Notes in Computer Science (2005), pp. 199–216.
[6] M. Hayashi, “Tight Exponential Analysis of Universally Composable Privacy Amplification and Its Applications”, IEEE
Transactions on Information Theory 59, 7728–7746 (2013), arXiv:1010.1358 [cs.IT].
[7] S. Watanabe and M. Hayashi, “Non-asymptotic analysis of privacy amplification via Rényi entropy and inf-spectral
entropy”, in Proceedings of the 2013 IEEE International Symposium on Information Theory (2013), pp. 2715–2719,
arXiv:1211.5252 [cs.IT].
[8] M. H. Yassaee, M. R. Aref, and A. Gohari, “Non-Asymptotic Output Statistics of Random Binning and Its Applications”,
(2013), arXiv:1303.0695 [cs.IT].
[9] W. Yang, R. F. Schaefer, and H. V. Poor, “Wiretap Channels: Nonasymptotic Fundamental Limits”, (2017), arXiv:1706.03866
[cs.IT].
[10] H. Nagaoka, “Strong converse theorems in quantum information theory”, in Proceedings of the ERATO Conference
on Quantum Information Science (EQIS), Vol. 33 (2001).
[11] Y. Polyanskiy, H. Poor, and S. Verdú, “Channel Coding Rate in the Finite Blocklength Regime”, IEEE Transactions on
Information Theory 56, 2307–2359 (2010).
[12] M. Hayashi, Quantum Information Theory, Graduate Texts in Physics (Springer, Berlin, Heidelberg, 2017).
[13] Y. Steinberg and S. Verdú, “Simulation of random processes and rate-distortion theory”, IEEE Transactions on Information Theory 42, 63–86 (1996).
[14] A. Winter, “Compression of sources of probability distributions and density operators”, (2002), arXiv:quant-ph/0208131.
[15] J. M. Renes, “Duality of channels and codes”, (2017), arXiv:1701.05583 [quant-ph].
[16] E. Martinian and J. S. Yedidia, “Iterative Quantization Using Codes On Graphs”, in Proceedings of the 41st Annual
Allerton Conference on Communication, Control, and Computing (2003), p. 1317, arXiv:cs/0408008.
[17] R. J. Vanderbei, Linear programming: foundations and extensions (Springer, New York, 2013).
[18] J. Neyman and E. S. Pearson, “On the Problem of the Most Efficient Tests of Statistical Hypotheses”, Philosophical
Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 231, 289–337 (1933).
[19] Y. Polyanskiy, “Channel coding: non-asymptotic fundamental limits”, PhD (Princeton University, Nov. 2010).
[20] H. Chernoff, “Large-Sample Theory: Parametric Case”, The Annals of Mathematical Statistics 27, 1–22 (1956).
[21] M. Tomamichel and M. Hayashi, “A Hierarchy of Information Quantities for Finite Block Length Analysis of Quantum
Tasks”, IEEE Transactions on Information Theory 59, 7693–7710 (2013), arXiv:1208.1478 [quant-ph].
[22] F. Dupuis, L. Krämer, P. Faist, J. M. Renes, and R. Renner, “Generalized Entropies”, in XVIIth International Congress
on Mathematical Physics, edited by A. Jensen (World Scientific, 2013), pp. 134–153, arXiv:1211.3141 [quant-ph].
[23] J. L. Carter and M. N. Wegman, “Universal classes of hash functions”, Journal of Computer and System Sciences 18,
143–154 (1979).
[24] M. Hayashi, “General nonasymptotic and asymptotic formulas in channel resolvability and identification capacity
and their application to the wiretap channel”, IEEE Transactions on Information Theory 52, 1562–1575 (2006).
[25] M. Tomamichel, C. Schaffner, A. Smith, and R. Renner, “Leftover Hashing Against Quantum Side Information”, IEEE
Transactions on Information Theory 57, 5524–5535 (2011), arXiv:1002.2436 [quant-ph].
[26] M. Tomamichel, “A Framework for Non-Asymptotic Quantum Information Theory”, PhD Thesis (ETH Zürich, 2012),
arXiv:1203.2142 [quant-ph].
[27] R. Ahlswede and I. Csiszàr, “Common randomness in information theory and cryptography. I. Secret sharing”, IEEE
Transactions on Information Theory 39, 1121–1132 (1993).
[28] J. M. Renes, “A Theory of Information (?)”, in Beyond i.i.d. in information theory (Jan. 2013).
[29] C. Bennett, P. Shor, J. Smolin, and A. Thapliyal, “Entanglement-assisted capacity of a quantum channel and the reverse Shannon theorem”, Information Theory, IEEE Transactions on 48, 2637–2655 (2002), arXiv:quant-ph/0106052.
[30] J. M. Renes and R. Renner, “Noisy Channel Coding via Privacy Amplification and Information Reconciliation”, IEEE
Transactions on Information Theory 57, 7377–7385 (2011), arXiv:1012.4814 [quant-ph].
12
[31] N. Datta, J. Renes, R. Renner, and M. Wilde, “One-Shot Lossy Quantum Data Compression”, IEEE Transactions on
Information Theory 59, 8057–8076 (2013), arXiv:1304.2336 [quant-ph].
[32] Z. Luo and I. Devetak, “Channel Simulation With Quantum Side Information”, IEEE Transactions on Information
Theory 55, 1331–1342 (2009), arXiv:quant-ph/0611008.
[33] N. Datta, M. H. Hsieh, and M. M. Wilde, “Quantum Rate Distortion, Reverse Shannon Theorems, and Source-Channel
Separation”, IEEE Transactions on Information Theory 59, 615–630 (2013), arXiv:1108.4940 [quant-ph].
[34] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. (Wiley-Interscience, 2006).
[35] V. Kostina and S. Verdu, “Fixed-Length Lossy Compression in the Finite Blocklength Regime”, IEEE Transactions on
Information Theory 58, 3309–3338 (2012), arXiv:1102.3944 [cs.IT].
[36] J. M. Renes, “Duality of privacy amplification against quantum adversaries and data compression with quantum side
information”, Proceedings of the Royal Society A 467, 1604–1623 (2011), arXiv:1003.0703 [quant-ph].
[37] J. M. Renes, D. Sutter, and S. H. Hassani, “Alignment of Polarized Sets”, IEEE Journal on Selected Areas in Communications 34, 224–238 (2016), arXiv:1411.7925 [quant-ph].
[38] S. B. Korada and R. L. Urbanke, “Polar Codes are Optimal for Lossy Source Coding”, IEEE Transactions on Information
Theory 56, 1751–1768 (2010), arXiv:0903.0307 [cs.IT].
[39] V. Aref, N. Macris, and M. Vuffray, “Approaching the Rate-Distortion Limit With Spatial Coupling, Belief Propagation,
and Decimation”, IEEE Transactions on Information Theory 61, 3954–3979 (2015), arXiv:1307.5210 [cs.IT].
[40] M. M. Wilde and S. Guha, “Polar codes for classical-quantum channels”, IEEE Transactions on Information Theory
59, 1175–1187 (2013), arXiv:1109.2591 [quant-ph].
[41] D. MacKay, “Good error-correcting codes based on very sparse matrices”, IEEE Transactions on Information Theory
45, 399–431 (1999).
[42] J. M. Renes, “Belief propagation decoding of quantum channels by passing quantum messages”, New Journal of
Physics 19, 072001 (2017), arXiv:1607.04833 [quant-ph].
13
| 7cs.IT
|
arXiv:1601.01894v1 [math.GR] 8 Jan 2016
On The UNRECOGNIZABILITY BY PRIME
GRAPH FOR THE ALMOST SIMPLE GROUP
PGL(2, 9)
Ali Mahmoudifar
Department of Mathematics, Tehran North Branch,
Islamic Azad University, Tehran, Iran
e-mail: [email protected]
Abstract
The prime graph of a finite group G is denoted by Γ(G). Also G is called recognizable by
prime graph if and only if each finite group H with Γ(H) = Γ(G), is isomorphic to G. In this
paper, we classify all finite groups with the same prime graph as PGL(2, 9). In particular,
we present some solvable groups with the same prime graph as PGL(2, 9).
2000 AMS Subject Classification: 20D05, 20D60, 20D08.
Keywords : Projective general linear group, prime graph, recognition.
1
Introduction
Let n be a natural number. We denote by π(n), the set of all prime divisors of n. Also Let G be
a finite group. The set π(|G|) is denoted by π(G). The set of element orders of G is denoted by
πe (G). We denote by µ(S), the maximal numbers of πe (G) under the divisibility relation. The
prime graph of G is a graph whose vertex set is π(G) and two distinct primes p and q are joined
by an edge (and we write p ∼ q), whenever G contains an element of order pq. The prime graph
of G is denoted by Γ(G). A finite group G is called recognizable by prime graph if for every finite
group H such that Γ(G) = Γ(H), then we have G ∼
= H. So G is recognizable by prime graph
whenever there exists a fin finite group K such that Γ(K) = Γ(G) in while K is not isomorphic
to G.
For the almost simple group PGL(2, q), there are a lot of results about the recognition by
prime graph. In [8], it is proved that if p is a prime number which is not a Mersenne or Fermat
prime and p 6= 11, 19 and Γ(G) = Γ(PGL(2, p)), then G has a unique nonabelian composition factor which is isomorphic to PSL(2, p) and if p = 13, then G has a unique nonabelian composition
factor which is isomorphic to PSL(2, 13) or PSL(2, 27). We know that PGL(2, 2α ) ∼
= PSL(2, 2α ).
For the characterization of such simple groups we refer to [9, 10]. In [1], it is proved that if
q = pα , where p is an odd prime and α is an odd natural number, then PGL(2, q) is uniquely
determined by its prime graph.
By the above description, we get that the characterization by prime graph of PGL(2, pk ),
where p is an odd prime number and k is even, is an open problem. In this paper as the
main result we consider the recognition by prime graph of the almost simple groups PGL(2, 32 ).
Moreover, we construct some solvable group with the same prime graph as PGL(2, 32 ).
1
2
Preliminary Results
Lemma 2.1. Let G be a finite group and N E G such that G/N is a Frobenius group with kernel
F and cyclic complement C. If (|F |, |N |) = 1 and F is not contained in N CG (N )/N , then
p|C| ∈ πe (G) for some prime divisor p of |N |.
Lemma 2.2. Let G be a Frobenius group with kernel F and complement C. Then the following
assertions hold:
(a) F is a nilpotent group.
(b) |F | ≡ 1 (mod |C|).
(c) Every subgroup of C of order pq, with p, q (not necessarily distinct) primes, is cyclic.
In particular, every Sylow subgroup of C of odd order is cyclic and a Sylow 2-subgroup of C
is either cyclic or generalized quaternion group. If C is a non-solvable group, then C has a
subgroup of index at most 2 isomorphic to SL(2, 5) × M , where M has cyclic Sylow p-subgroups
and (|M |, 30) = 1.
By using [13, Theorem A] we have the following result:
Lemma 2.3. Let G be a finite group with t(G) ≥ 2. Then one of the following holds:
(a) G is a Frobenius or 2-Frobenius group;
(b) there exists a nonabelian simple group S such that S ≤ G:= G/N ≤ Aut(S) for some
nilpotent normal π1 -subgroup N of G and G/S is a π1 -group.
3
Main Results
Lemma 3.1. There exists a Frobenius group G = K : C, where K is an abelian 3-group and
π(C) = {2, 5}, such that Γ(G) = Γ(PGL(2, 9)).
Proof. Let F be a finite field with 34 elements. Also let V be the additive group of F and H be
the multiplicative group F \ {0}. We know that H acts on V by right product. So G := V ⋊ H
is a finite group such that π(G) = {2, 3, 5}, since |V | = 34 and |H| = 80. On the other hand H
acts fixed point freely on V , so G is a Frobenius group with kernel V and complement H. Since
the multiplicative group F \ {0} is cyclic, H is cyclic too. Therefore, G has an element of order
10, and so the prime graph of G consists just one edge, which is the edge between 2 and 5. This
implies that Γ(G) = Γ(PGL(2, 9)), as desired.
Lemma 3.2. There exists a Frobenius group G = K : C, where π(K) = {2, 5} and C is a cyclic
3-group such that Γ(G) = Γ(PGL(2, 9)).
Proof. Let F1 and F2 be two fields with 22 and 52 elements, respectively. Let V := F1 × F2
be the direct product of the additive groups F1 and F2 . Also let H := H1 × H2 , be the direct
product of H1 and H2 , which are the multiplicative groups F1 \ {0} and F2 \ {0}, respectively.
We know that Hi acts fixed point freely on Fi , where 1 ≤ i ≤ 2. So we define an action of H
′
on V as follows: for each (h, h′ ) ∈ H and (g, g′ ) ∈ V , we define (g, g′ )(h,h ) := (hg, h′ g′ ). It is
clear that this definition is well defined. So we may construct a finite group G = V ⋊ H. On
the other hand H acts fixed point freely on V . So G is a Frobenius group with kernel V and
complement H such that π(V ) = {2, 5} and π(H) = {3}. Finally, since V is nilpotent, we get
that Γ(G) contains an edge between 2 and 5, so Γ(G) = Γ(PGL(2, 9)).
Lemma 3.3. There exists a 2-Frobenius group G with normal series 1 ⊳ H ⊳ K ⊳ G, such that
π(H) = {5}, π(G/K) = {2} and π(K/H) = {3}, such that Γ(G) = Γ(PGL(2, 9)).
2
Proof. Let F be a field with 52 elements and V be its additive group. We know that F =
{0, 1, α, α2 , . . . , α23 }, where α is a generator of the multiplicative group F \ {0}. Hence |α| = 24,
and so β := α8 has order 3. Also hβi ∼
= Z3 and Aut(Z3 ) ∼
= Z2 . This argument implies that we
may construct a Frobenius group T := hβi : hγi, where γ is an involution.
x y
Now we define an action of T on V as follows: for each β x γ y ∈ T and v ∈ V , v β γ := β x v,
where 1 ≤ x ≤ 3 and 1 ≤ y ≤ 2. Therefore G := V : T is a 2 Frobenius group with desired
properties.
Theorem 3.4. Let G be a finite group. Then Γ(G) = Γ(PGL(2, 32 )) if and only if G is isomorphic to one of the following groups:
(1) A Frobenius group K : C, where K is an abelian 3-group and π(C) = {2, 5},
(2) A Frobenius group K : C, where π(K) = {2, 5} and C is a cyclic 3-group,
(3) A 2-Frobenius group with normal series 1 ⊳ H ⊳ K ⊳ G, such that π(H) ⊆ {2, 5},
π(G/K) = {2} and π(K/H) = {3},
(4) Almost simple group PGL(2, 32 ).
Proof. Throughout the proof, we assume that G is a finite group with the same prime graph as
the almost simple group PGL(2, 32 ). First we note that by [16, Lemma 7], we have:
µ(PGL(2, 9)) = {3, 8, 10}.
Hence in Γ(PGL(2, 32 )) (and so in Γ(G)), there is only one edge which is the edge between
2 and 5 and so 3 is an isolated vertex. This implies that Γ(G) has two connected components
{3} and {2, 5}. Thus by Lemma 2.3, we get that G is a Frobenius group or 2-Frobenius group
or there exists a nonabelian simple group S such that S ≤ G/F it(G) ≤ Aut(S). We consider
each possibility for G.
Let G be a Frobenius group with kernel K and complement C. We know that K is nilpotent
and C is a connected component of the prime graph of G. Also by the above description, 3
is not adjacent to 2 and 5 in Γ(G). This shows that either π(K) = {3} or π(C) = {3}. We
consider these cases, separately.
Case 1. Let π(K) = {3}. Hence the order of complement C, is even and so K is an
abelain subgroup of G. Also by the above description, π(C) = {2, 5}. Since C is a connected
component of Γ(G), there is and edge between 2 and 5. So it follows that Γ(G) = Γ(PGL(2, 9)),
which implies groups satisfying in (1).
Case 2. Let π(C) = {3}. Hence π(K) = {2, 5}. Since K is nilpotent, we get that 2 and 5
are adjacent in Γ(G). This means Γ(G) = Γ(PGL(2, 9)) and so we get (2).
Case 3. Let G be a 2-Frobenius group with normal series 1 ⊳ H ⊳ K ⊳ G. Since π(K/H)
and π(H) ∪ π(G/K) are the connected components of Γ(G), we get that π(K/H) = {3} and
π(H) ∪ π(G/K) = {2, 5}. This implies (3).
Case 4. Let there exist a nonabelian simple group S, such that S ≤ Ḡ := G/Fit(G) ≤
Aut(S). Since π(S) ⊆ π(G), π(S) = {2, 3, 5}. The finite simple groups with this property
are classified in [12, Table 8]. So we get that S is isomorphic to one of the simple groups A5 ,
PSU(4, 2) and PSL(2, 9)(∼
= A6 ).
Subcase 4.1. Let S ∼
= A5 . We know that Aut(A5 ) = S5 . So Ḡ is isomorphic to the
alternating group A5 or the symmetric group S5 . Since in the prime graph of S5 , there is an
edge between 2 and 3, hence we get that Ḡ is not isomorphic to S5 . Thus, G/Fit(G) = A5 . In
the prime graph of A5 , 2 and 5 are nonadjacent. Then at least one of the prime numbers 2 or
5, belongs to π(Fit(G)).
3
Let 5 ∈ π(Fit(G)). Let F5 be a Sylow 5-subgroup of Fit(G). Since F5 is a characteristic
subgroup of Fit(G) and Fit(G) is a normal subgroup of G, F5 E G. On the other hand in alternating group A5 , the subgroup h(12)(34), (13)(24)i : h(123)i is a Frobenius subgroup isomorphic
to 22 : 3. We recall that by the previous argument, F5 E G, and so G has a subgroup isomorphic
to 5α : (22 : 3). So by Lemma 2.1, we get that 3 is adjacent to 5, a contradiction.
Subcase 4.2. Let S ∼
= PSU(4, 2). By [4], there is an edge between 3 and 2 which is a
contradiction.
Subcase 4.3. Let S ∼
= PSL(2, 9). Then Ḡ is isomorphic to PSL(2, 9) or PSL(2, 9) : hθi,
where θ is a diagonal, field or diagonal-field automorphism of PSL(2, 9). In particular θ is an
involution. If θ is a field or diagonal-field automorphism of PSL(2, 9), then the semidirect product
PSL(2, 9) : hθi contains an element of order 6. Hence θ is neither a field automorphism nor a
diagonal-field automorphism. Therefore θ is a diagonal automorphism and so Ḡ ∼
= PGL(2, 9).
∼
By the above discussion, G/Fit(G) = PGL(2, 9). It is enough to prove that Fit(G) = 1. On
the contrary, let r ∈ π(Fit(G)). Also let Fr be the Sylow r-subgroup of Fit(G). Since Fit(G) is
nilpotent, we can write Fit(G) = Or′ (Fit(G)) × Fr . So if we put G̃ = G/Or′ (Fit(G)), then we
get that:
G ∼ G̃ ∼ G̃/Φ(Fr )
.
PGL(2, 9) ∼
=
=
=
Fit(G)
Fr
Fr /Φ(Fr )
Since Fr /Φ(Fr ) is an elementary abelian group, without loose of generality we may assume that
F := Fit(G) is an elementary abelian r-group and G/F ∼
= PGL(2, 9).
If r = 3, then by , we conclude that in Γ(G), 2 and 3 are adjacent, which is a contradiction.
So let r 6= 3. Also let S3 be a Sylow 3-subgroup of PGL(2, 9). We know that S3 is not cyclic. On
the other hand F ⋊ S3 is a Frobenius group since 3 is an isolated vertex of Γ(G). This follows
that S3 is cyclic which is impossible. Therefore F = 1 and so G ∼
= PGL(2, 9), which completes
the proof.
References
[1] Z. Akhlaghi, M. Khatami and B. Khosravi, Characterization by prime graph of PGL(2, pk )
where p and k are odd, International Journal of Algebra and Computation 20 (7) (2010)
847-873.
[2] S. Shariati Beynekalae, A. Iranmanesh and M. Foroudi Ghasemabadi, Quasirecognition by
prime graph of the simple group, Southeast Asian Bull. Math., Bn (2), 39 (2015) 181–193.
[3] L. Wang, Quasirecognition of L22 (2) by Its prime graph, Southeast Asian Bull. Math., 35
(2015) 883–889.
[4] A. A. Buturlakin, Spectra of finite linear and unitary groups, Algebra and Logic 47(2)
(2008) 91–99.
[5] J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker and R. A. Wilson, Atlas of Finite
Groups (Oxford University Press, Oxford, 1985).
[6] K. W. Gruenberg and K. W. Roggenkamp, Decomposition of the augmentation ideal and
of the relation modules of a finite group, Proc. London Math. Soc. 31(2) (1975) 149–166.
[7] M. Hagie, The prime graph of a sporadic simple group, Comm. Algebra 31(9) (2003) 4405–
4424.
4
[8] M. Khatami, B. Khosravi and Z. Akhlaghi, NCF-distinguishability by prime graph of
P GL(2, p), where p is a prime, Rocky Mountain J. Math., to appear.
[9] B. Khosravi, n-Recognition by prime graph of the simple group P SL(2, q), J. Algebra Appl.
7(6) (2008) 735–748.
[10] B. Khosravi, B. Khosravi and B. Khosravi, 2-Recognizability of P SL(2, p2 ) by the prime
graph, Siberian Math. J. 49(4) (2008) 749.757.
[11] B. Khosravi, B. Khosravi and B. Khosravi, On the prime graph of P SL(2, p) where p > 3
is a prime number, Acta. Math. Hungarica 116(4) (2007) 295–307.
[12] R. Kogani-Moghadam and A. R. Moghaddamfar, Groups with the same order and degree
pattern, Sci. China Math., 55 (4) (2012), 701–720.
[13] A. S. Kondrat’ev, Prime graph components of finite simple groups, Math. USSR-SB. 67(1)
(1990) 235–247.
[14] V. D. Mazurov, Characterizations of groups by arithmetic properties, Proceedings of the
International Conference on Algebra, Algebra Colloq. 11(1) (2004) 129–140.
[15] V. D. Mazurov, Characterizations of finite groups by sets of their element orders, Algebra
Logic 36(1) (1997) 23–32.
[16] A. R. Moghaddamfar and W. J. Shi, The number of finite groups whose element orders is
given., Beitrage Algebra Geom., 47(2) (2006) 463–479.
[17] J. S. Williams, Prime graph components of finite groups, J. Algebra 69(2) (1981) 487–513.
[18] A. V. Zavarnitsin, Recognition of finite groups by the prime graph, Algebra Logic 43(4)
(2006) 220–231.
[19] K. Zsigmondy, Zur theorie der potenzreste, Monatsh. Math. Phys. 3 (1892) 265–284.
5
| 4math.GR
|
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
1
Deep convolutional neural networks for predominant
instrument recognition in polyphonic music
arXiv:1605.09507v3 [cs.SD] 26 Dec 2016
Yoonchang Han, Jaehun Kim, and Kyogu Lee, Senior Member, IEEE
Abstract—Identifying musical instruments in polyphonic music recordings is a challenging but important problem in the
field of music information retrieval. It enables music search
by instrument, helps recognize musical genres, or can make
music transcription easier and more accurate. In this paper,
we present a convolutional neural network framework for predominant instrument recognition in real-world polyphonic music.
We train our network from fixed-length music excerpts with a
single-labeled predominant instrument and estimate an arbitrary
number of predominant instruments from an audio signal with
a variable length. To obtain the audio-excerpt-wise result, we
aggregate multiple outputs from sliding windows over the test
audio. In doing so, we investigated two different aggregation
methods: one takes the average for each instrument and the
other takes the instrument-wise sum followed by normalization.
In addition, we conducted extensive experiments on several
important factors that affect the performance, including analysis
window size, identification threshold, and activation functions
for neural networks to find the optimal set of parameters.
Using a dataset of 10k audio excerpts from 11 instruments for
evaluation, we found that convolutional neural networks are
more robust than conventional methods that exploit spectral
features and source separation with support vector machines.
Experimental results showed that the proposed convolutional
network architecture obtained an F1 measure of 0.602 for micro
and 0.503 for macro, respectively, achieving 19.6% and 16.4% in
performance improvement compared with other state-of-the-art
algorithms.
Index Terms—Instrument recognition, convolutional neural
networks, deep learning, multi-layer neural network, music
information retrieval
I. I NTRODUCTION
M
USIC can be said to be built by the interplay of
various instruments. A human can easily identify what
instruments are used in a music, but it is still a difficult
task for a computer to automatically recognize them. This is
mainly because music in the real world is mostly polyphonic,
which makes extraction of information from an audio highly
challenging. Furthermore, instrument sounds in the real world
vary in many ways such as for timbre, quality, and playing
style, which makes identification of the musical instrument
even harder.
In the music information retrieval (MIR) field, it is highly
desirable to know what instruments are used in an audio
sample. First of all, instrument information per se is an
Y. Han, J. Kim, and K. Lee are with the Music and Audio Research
Group, Graduate School of Convergence Science and Technology, Seoul
National University, Seoul 08826, Republic of Korea, e-mail: ([email protected], [email protected], [email protected]).
K. Lee is also with the Advanced Institutes of Convergence Technology,
Suwon, Republic of Korea
Manuscript received December 28, 2016; revised December 28, 2016.
important and useful information for users, and it can be
included in the audio tags. There is a huge demand for
music search owing to the increasing number of music files
in digital format. Unlike text search, it is difficult to search
for music because input queries are usually in text format. If
an instrument information is included in the tags, it allows
people to search for music with the specific instrument they
want. In addition, the obtained instrument information can
be used for various audio/music applications. For instance,
more instrument-specific and tailored audio equalization can
be applied to the music; moreover, a music recommendation
system can reflect the preference of users for musical instruments. Furthermore, it can also be used to enhance the
performance of other MIR tasks. For example, knowing the
number and type of the instrument would significantly improve
the performance of source separation and automatic music
transcription; it would also be helpful for identifying the genre
of the music.
Instrument recognition can be performed in various forms.
Hence, the term “instrument recognition” or “instrument identification” might indicate several different research topics.
For instance, many of the related works focus on studiorecorded isolated notes. To name a few, Eronen used cepstral
coefficients and temporal features to classify 30 orchestral
instruments with several articulation styles and achieved a
classification accuracy of 95% for instrument family level
and about 81% for individual instruments [1]. Diment et al.
used a modified group delay feature that incorporates phase
information together with mel-frequency cepstral coefficients
(MFCCs) and achieved a classification accuracy of about
71% for 22 instruments [2]. Yu et al. applied sparse coding
on cepstrum with temporal sum-pooling and achieved an Fmeasure of about 96% for classifying 50 instruments [3].
They also reported their classification result on a multi-source
database, which was about 66%.
Some previous works such as Krishna and Sreenivas [4]
experimented with a classification for solo phrases rather than
for isolated notes. They proposed line spectral frequencies
(LSF) with a Gaussian mixture model (GMM) and achieved
an accuracy of about 77% for instrument family and 84% for
14 individual instruments. Moreover, Essid et al. [5] reported
that a classification system with MFCCs and GMM along
with principal components analysis (PCA) achieved an overall
recognition accuracy of about 67% on solo phrases with five
instruments.
More recent works deal with polyphonic sound, which is
closer to real-world music than to monophonic sound. In
the case of polyphonic sound, a number of research stud-
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
ies used synthesized polyphonic audio from studio-recorded
single tones. Heittola et al. [6] used a non-negative matrix
factorization (NMF)-based source-filter model with MFCCs
and GMM for synthesized polyphonic sound and achieved a
recognition rate of 59% for six polyphonic notes randomly
generated from 19 instruments. Kitahara et al. [7] used various
spectral, temporal, and modulation features with PCA and
linear discriminant analysis (LDA) for classification. They
reported that, using feature weighting and musical context,
recognition rates were about 84% for a duo, 78% for a trio,
and 72% for a quartet. Duan et al. [8] proposed the uniform
discrete cepstrum (UDC) and mel-scale UDC (MUDC) as
a spectral representation with a radial basis function (RBF)
kernel support vector machine (SVM) to classify 13 types of
Western instruments. The classification accuracy of randomly
mixed chords of two and six polyphonic notes, generated using
isolated note samples from the RWC musical instrument sound
database [9], was around 37% for two polyphony notes and
25% for six polyphony notes.
As shown above, most of the previous works focused on the
identification of the instrument sounds in clean solo tones or
phrases. More recent research studies on polyphonic sounds
are closer to the real-world situation, but artificially produced
polyphonic music is still far from professionally produced
music. Real-world music has many other factors that affect
the recognition performance. For instance, it might have a
highly different timbre, depending on the genre and style of the
performance. In addition, an audio file might differ in quality
to a great extent, depending on the recording and production
environments.
In this paper, we investigate a method for predominant
instrument recognition in professionally produced Western
music recordings. We utilize convolutional neural networks
(ConvNets) to learn the spectral characteristics of the music
recordings with 11 musical instruments and perform instrument identification on polyphonic music excerpts. The major
contributions of the work presented in this paper are as
follows.
1. We present the ConvNet architecture for predominant
musical instrument identification where the training data
are single labeled and the target data are multi-labeled
with an unknown number of classes existing in the data.
2. We introduce a new method to aggregate the outputs
of ConvNets from short-time sliding windows to find the
predominant instruments in a music excerpt with variable
length, where the conventional method of majority vote
often fails.
3. We conduct an extensive experiment on activation function for the neurons used in ConvNets, which can cause
a huge impact on the identification result.
The remainder of the paper is organized as follows. In section II, we introduce emerging deep neural network techniques
in the MIR field. Next, the system architecture section includes
audio preprocessing, the proposed network architecture with
2
detailed training configuration, and an explanation of various
activation functions used for the experiment. Section IV, the
evaluation section, contains information about the dataset,
testing configuration including aggregation strategy, and our
evaluation scheme. Then, we illustrate the performance of the
proposed ConvNet in section V, the Results section, with an
analysis of the effects of activation function, analysis window
size, aggregation strategy, and identification threshold, and
with an instrument-wise analysis. Moreover, we present a
qualitative analysis based on the visualization of the ConvNet’s
intermediate outputs to understand how the network captured
the pattern from the input data. Finally, we conclude the paper
in section VI.
II. P ROLIFERATION OF D EEP N EURAL N ETWORKS IN
M USIC I NFORMATION R ETRIEVAL
The ability of traditional machine learning approaches was
limited in terms of processing input data in their raw form.
Hence, usually the input for the learning system, typically
a classifier, has to be a hand-crafted feature representation,
which requires extensive domain knowledge and a careful
engineering process. However, it is getting more common
to design the system to automatically discover the higherlevel representation from the raw data by stacking several
layers of nonlinear modules, which is called deep learning
[10]. Recently, deep learning techniques have been widely
used across a number of domains owing to their superior
performance. A basic architecture of deep learning is called
deep neural network (DNN), which is a feedforward network
with multiple hidden layers of artificial neurons. DNN-based
approaches have outperformed previous state-of-the-art methods in speech applications such as phone recognition, largevocabulary speech recognition, multi-lingual speech recognition, and noise-robust speech recognition [11].
There are many variants and modified architectures of deep
learning, depending on the target task. Especially, recurrent
neural networks (RNNs) and ConvNets have recently shown
remarkable results for various multimedia information retrieval
tasks. RNNs are highly powerful approaches for sequential
inputs as their recurrent architecture enables their hidden units
to implicitly maintain the information about the past elements
of the sequence. Since languages natively contain sequential
information, it is widely applied to handle text characters or
spoken language. It has been reported that RNNs have shown
a successful result on language modeling [12] and spoken
language understanding [13], [14].
On the other hand, ConvNet is useful for data with local
groups of values that are highly correlated, forming distinctive
local characteristics that might appear at different parts of the
array [10]. Hence, it is one of the most popular approaches
recently in the image processing area such as handwritten
digit recognition [15], [16], [17] for the MNIST dataset and
image tagging [18], [19] for the CIFAR-10 dataset. In addition,
it has been reported that it has outperformed state-of-the-art
approaches for several computer vision benchmark tasks such
as object detection, semantic segmentation, and category-level
object recognition [11], and also for speech-recognition tasks
[20].
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
3
Fig. 1. Schematic of the proposed ConvNet containing 4 times repeated double convolution layers followed by max-pooling. The last max-pooling layer
performs global max-pooling, then it is fed to a fully connected layer followed by 11 sigmoid outputs.
The time-frequency representation of a music signal is
composed of harmonics from various musical instruments and
a human voice. Each musical instrument produces a unique
timbre with different playing styles, and this type of spectral
characteristics in music signal might appear in a different
location in time and frequency as in the image. ConvNets
are usually composed of many convolutional layers, and
inserting a pooling layer between convolutional layers allows
the network to work at different time scales and introduces
translation invariance with robustness against local distortions.
These hierarchical network structures of ConvNets are highly
suitable for representing music audio, because music tends to
present a hierarchical structure in time and different features
of the music might be more salient at different time scales
[21].
Hence, although ConvNets have been a more commonly
used technique in image processing, there are an increasing
number of attempts to apply ConvNets for music signal. It
has been reported that ConvNet has outperformed previous
state-of-the-art approaches for various MIR tasks such as onset
detection [22], automatic chord recognition [23], [24], and
music structure/boundary analysis [25], [26].
An attempt to apply ConvNets for musical instrument
identification can be found in the recent report from Park
et al. [27] and Li et al. [28], although it is still an ongoing
work and is not a predominant instrument recognition method;
hence, there are no other instruments but only target instrument
sounds exist. Our research differs from [27] because we deal
with polyphonic music, while their work is based on the
studio recording of single tones. In addition, our research
also differs from [28] because we use single-label data for
training and estimate multi-label data, while they used multilabel data from the training phase. Moreover, they focused
on an end-to-end approach, which is promising in that using
raw audio signals makes the system rely less on domain
knowledge and preprocessing, but usually it shows a slightly
lower performance than using spectral input such as melspectrogram in recent papers [29], [30].
III. S YSTEM ARCHITECTURE
A. Audio Preprocessing
The convolutional neural network is one of the representation learning methods that allow a machine to be fed with raw
data and to automatically discover the representations needed
for classification or detection [10]. However, appropriate preprocessing of input data is still an important issue to improve
the performance of the system.
In the first preprocessing step, the stereo input audio is
converted to mono by taking the mean of the left and right
channels, and then it is downsampled to 22,050 Hz from the
original 44,100 Hz of sampling frequency. This allows us to
use frequencies up to 11,025 Hz, the Nyquist frequency, and
it is sufficient to cover most of the harmonics generated by
musical instruments while removing noises possibly included
in the frequencies above this range. Moreover, all audios are
normalized by dividing the time-domain signal with its maximum value. Then, this downsampled time-domain waveform is
converted to a time-frequency representation using short-time
Fourier transform (STFT) with 1024 samples for the window
size (approx. 46 ms) and 512 samples of the hop size (approx.
23 ms).
Next, the linear frequency scale-obtained spectrogram is
converted to a mel-scale. We use 128 for the number of melfrequency bins, following the representation learning papers
on music annotation by Nam et al. [31] and Hamel et al. [21],
which is a reasonable setting that sufficiently preserves the
harmonic characteristics of the music while greatly reducing
the dimensionality of the input data. Finally, the magnitude of
the obtained mel-frequency spectrogram is compressed with a
natural logarithm.
B. Network Architecture
ConvNets can be seen as a combination of feature extractor and the classifier. Our ConvNet architecture generally
follows a popular AlexNet [18] and VGGNet [32] structure,
which contains very deep architecture using repeated several
convolution layers followed by max-pooling, as shown in
Figure 1. This method of using smaller receptive window size
and smaller stride for ConvNet is becoming highly common
especially in the computer vision field such as in the study
from Zeiler and Fergus [33] and Sermanet et al. [34], which
has shown superior performance in ILSVRC-2013.
Although the general architecture style is similar to that of
other successful ConvNets in the image processing area, the
proposed ConvNet is designed according to our input data. We
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
TABLE I
P ROPOSED C ONV N ET S TRUCTURE . T HE I NPUT S IZE D EMONSTRATED IN
T HIS TABLE I S FOR AN A NALYSIS W INDOW S IZE OF 1 S ECOND (N UMBER
OF FILTERS × TIME × FREQUENCY ). T HE ACTIVATION F UNCTION I S
F OLLOWED BY E ACH C ONVOLUTIONAL L AYER AND A F ULLY
C ONNECTED L AYER . T HE I NPUT OF E ACH C ONVOLUTION L AYER I S
Z ERO -PADDED WITH 1 × 1, B UT I S N OT S HOWN FOR B REVITY.
Input size
Description
1 × 43 × 128
32 × 45 × 130
32 × 47 × 132
32 × 15 × 44
32 × 15 × 44
64 × 17 × 46
64 × 19 × 48
64 × 6 × 16
64 × 6 × 16
128 × 8 × 18
128 × 10 × 20
128 × 3 × 6
128 × 3 × 6
256 × 5 × 8
256 × 7 × 10
256 × 1 × 1
1024
1024
11
mel-spectrogram
3 × 3 convolution, 32 filters
3 × 3 convolution, 32 filters
3 × 3 max-pooling
dropout (0.25)
3 × 3 convolution, 64 filters
3 × 3 convolution, 64 filters
3 × 3 max-pooling
dropout (0.25)
3 × 3 convolution, 128 filters
3 × 3 convolution, 128 filters
3 × 3 max-pooling
dropout (0.25)
3 × 3 convolution, 256 filters
3 × 3 convolution, 256 filters
global max-pooling
flattened and fully connected
dropout (0.50)
sigmoid
use filters with a very small 3 × 3 receptive field, with a fixed
stride size of 1, and spatial abstraction is done by max-pooling
with a size of 3 × 3 and a stride size of 1.
In Table I, we illustrate the detailed ConvNet architecture
with the input size in each layer with parameter values except
the zero-padding process. The input for each convolution layer
is zero-padded with 1 × 1 to preserve the spatial resolution
regardless of input window size, and we increase the number
of channels for the convolution layer by a factor of 2 after
every two convolution layers, starting from 32 up to 256.
In the last max-pooling layer after the eight convolutional
layers, we perform global max-pooling followed by one fully
connected layer. Recently, it has been reported that the use of
global average pooling without a fully connected layer before
a classifier layer is less prone to overfitting and shows better
performance for image processing datasets such as CIFAR-10
and MNIST [35]. However, our empirical experiment found
that global average pooling slightly decreases the performance
and that global max-pooling followed by a fully connected
layer works better for our task.
Finally, the last classifier layer is the sigmoid layer. It
is common to use a softmax layer when there is only one
target label, but our system must be able to handle multiple
instruments present at the same time, and, thus, a sigmoid
output is used.
C. Training Configuration
The training was done by optimizing the categorical crossentropy between predictions and targets. We used Adam [36]
4
as an optimizer with a learning rate of 0.001, and the minibatch size was set to 128. To accelerate the learning process
with parallelization, we used a GTX 970 GPU, which has 4GB
of memory.
The training was regularized using dropout with a rate of
0.25 after each max-pooling layer. Dropout is a technique
that prevents the overfitting of units to the training data by
randomly dropping some units from the neural network during
the training phase [37]. Furthermore, we added dropout after
a fully connected layer as well with a rate of 0.5 since a fully
connected layer easily suffers from overfitting.
In addition, we conducted an experiment with various time
resolutions to find the optimal analysis size. As our training
data were a fixed 3-s audio, we performed the training with
3.0, 1.5, 1.0, and 0.5 s by dividing the training audio and used
the same label for each divided chunk. The audio was divided
without overlap for training as it affects the validation loss
used for the early stopping. Fifteen percent of the training data
were randomly selected and used as a validation set, and the
training was stopped when the validation loss did not decrease
for more than two epochs.
The initialization of the network weights is another important issue as it can lead to an unstable learning process,
especially for a very deep network. We used a uniform
distribution with zero biases for both convolutional and fully
connected layers following Glorot and Bengio [38].
D. Activation Function
The activation function is followed by each convolutional
layer and fully connected layer. In this section, we introduce
several activation functions used in the experiment for the
comparison.
The traditional way to model the activation of a neuron is by
using a hyperbolic tangent (tanh) or sigmoid function. However, non-saturating nonlinearities such as the rectified linear
unit (ReLU) allow much faster learning than these saturating
nonlinearities, particularly for models that are trained on large
datasets [18]. Moreover, a number of works have shown that
the performance of ReLU is better than that of sigmoid and
tanh activation [39]. Thus, most of the modern studies on
ConvNets use ReLU to model the output of the neurons [28],
[32], [33], [34].
ReLU was first introduced by Nair and Hinton in their work
on restricted Boltzmann machines [40]. The ReLU activation
function is defined as
yi = max(0, zi )
(1)
where zi is the input of the ith channel. ReLU simply
suppresses the whole negative part to zero while retaining
the positive part. Recently, there have been several modified
versions of ReLU introduced to improve the performance
further. First, leaky-ReLU (LReLU), introduced by Mass et
al. [41], compresses the negative part rather than make it all
zero, which might cause some initially inactive units to remain
inactive. It is defined as
zi
zi ≥ 0
yi =
(2)
αzi zi < 0
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
where α is a parameter between 0 and 1 to give a small gradient in the negative part. Second, parametric ReLU (PReLU),
introduced by He et al. [42], is basically similar to LReLU
in that it compresses the negative part. However, PReLU
automatically learns the parameter for the negative gradient,
unlike LReLU. It is defined as
zi
zi ≥ 0
yi =
(3)
αi zi zi < 0
where αi is the learned parameters for the ith channel.
The choice of activation function considerably influences
the identification performance. It is difficult to say which
specific activation function always performs the best because
it highly depends on the parameter setting and the input data.
For instance, an empirical evaluation of the ConvNet activation
functions from Xu et al. [43] reported that the performance
of LReLU is better than those of ReLU and PReLU, but
sometimes it is worse than that of basic ReLU, depending on
the dataset and the value for α. Moreover, most of the works
regarding activation function are on the image classification
task, not on the audio processing domain.
Hence, we empirically evaluated several activation functions
explained above such as tanh, ReLU, LReLU, and PReLU to
find the most suitable activation function for our task. For
LReLU, very leaky ReLU (α = 0.33) and normal leaky ReLU
(α = 0.01) were used, because it has been reported that the
performance of LReLU considerably differs depending on the
value and that very leaky ReLU works better [43].
We used separate test audio data from the IRMAS dataset,
which were not used for the training. First, a sliding window
was used to analyze the input test audio, which was of the
same size as the analysis window in the training phase. The
hop size of the sliding window was set to half of the window
size. Then, we aggregated the sigmoid outputs from the sliding
windows by summing all outputs class-wise to obtain a total
amount of activation for each instrument. These 11 summed
sigmoid activations were then normalized to be in a range
between 0 and 1 by dividing all with the maximum activation.
IV. E VALUATION
A. IRMAS Dataset
The IRMAS dataset includes musical audio excerpts with
annotations of the predominant instruments present and is
intended to be used for the automatic identification of the
predominant instruments in the music. This dataset was used
in the paper on predominant instrument classification by Bosch
et al. [44] and includes music from various decades from the
past century, hence differing in audio quality to a great extent.
In addition, the dataset covers a wide variability in musical instrument types, articulations, recording and production styles,
and performers.
The dataset is divided into training and testing data, and
all audio files are in 16-bit stereo wave with 44,100 Hz of
sampling rate. The training data consisted of 6705 audio files
with excerpts of 3 s from more than 2000 distinct recordings.
Two subjects were paid to obtain the data for 11 pitched
instruments, as shown in Table II from selected music tracks,
5
TABLE II
L IST OF M USICAL I NSTRUMENTS U SED IN THE E XPERIMENT WITH T HEIR
A BBREVIATIONS , AND THE N UMBER OF L ABELS OF THE T RAINING AND
T ESTING AUDIO .
Instruments
Abbreviations
Training (n)
Testing (n)
Cello
Clarinet
Flute
Acoustic guitar
Electric guitar
Organ
Piano
Saxophone
Trumpet
Violin
Voice
cel
cla
flu
acg
elg
org
pia
sax
tru
vio
voi
388
505
451
637
760
682
721
626
577
580
778
111
62
163
535
942
361
995
326
167
211
1044
with the objective of extracting music excerpts that contain a
continuous presence of a single predominant instrument.
On the other hand, the testing data consisted of 2874 audio
files with lengths between 5 s and 20 s, and no tracks from
the training data were included. Unlike the training data,
the testing data contained one or more predominant target
instruments. Hence, the total number of training labels was
identical to the number of audio files, but the number of testing
labels was more than the number of testing audio files as the
latter are multi-label. For both the training and the testing
dataset, other musical instruments such as percussion and bass
were not included in the annotation even if they exist in the
music excerpts.
B. Testing Configuration
In the training phase, we used a fixed length window
because the input data for ConvNet should be in a specific
fixed shape. However, our testing audios had variable lengths
between 5 s and 20 s, which were much longer than those of
the training audio. Developing a system that can handle variable length of input data is valuable because music in real life
varies in its length. We performed short-time analysis using
overlapping windows to obtain local instrument information in
the audio excerpts. Since an annotation exists per audio clip,
we observed multiple sigmoid outputs and aggregated them to
make a clip-wise decision. We tried two different strategies
for the aggregation, which are the average and the normalized
sum, which are referred as S1 and S2 throughout the paper,
respectively.
For S1, we simply took an average of the sigmoid outputs
class-wise (i.e., instrument-wise) over the whole audio clip and
thresholded it without normalization. This method is intended
to capture the existence of each instrument with its mean
probability such that it might return the result without any
detected instrument. For S2, we first summed all sigmoid outputs class-wise over the whole audio excerpt and normalized
the values by dividing them with the maximum value among
classes such that the values were scaled to be placed between
zero and one, followed by thresholding. This method is based
on the assumption that humans perceive the “predominant”
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
6
TABLE III
E XPERIMENT VARIABLES FOR THE ACTIVATION F UNCTION , S IZE OF THE
A NALYSIS W INDOW, AGGREGATION S TRATEGY, AND I DENTIFICATION
T HRESHOLD . D EFAULT S ETTINGS ARE I NDICATED IN B OLD .
Variables
Activation func.
analysis win. size
Agg. strategy
θ (S1)
θ (S2)
tanh, ReLU, PReLU, LReLU (0.01), LReLU (0.33)
0.5 s, 1.0 s, 1.5 s, 3.0 s
S1 (mean), S2 (sum and normalized)
0.02, 0.04, 0.06, 0.08, 0.10, 0.12, 0.14, 0.16, 0.18
0.20, 0.25, 0.30, 0.35, 0.40, 0.45, 0.50, 0.55, 0.60
R=
Fig. 2. Schematic of obtaining a multi-label output from a test audio signal.
Input audio was analyzed with sliding window, and these multiple sigmoid
outputs were aggregated using two different strategies, S1 and S2, to estimate
the predominant instrument for the testing audio excerpt.
instrument in a more relatively scaled sense such that the
strongest instrument is always detected and the existence of
other instruments is judged by their relative strength compared
to the most activate instrument.
Majority vote, one of the most common choices for a
number of classification tasks, is not used in our system.
Majority vote first predicts the classes for each analysis frame
and the one with more vote wins. However, using this method
for our task would result in disregarding accompaniment
instruments, piano for example, because a music signal is
composed of various musical instruments and usually the
sounds are overlapped in time domain, and a presence of
accompaniments are usually much weaker than that of voice
or lead instruments.
As our target is to identify an arbitrary number of predominant instruments in testing data, instruments with aggregated
value over the threshold were all considered as predominant
instruments. Using a higher value for the identification threshold will lead to better precision, but it will obviously decrease
the recall. On the other hand, a lower threshold will increase
the recall, but will lower the precision. Hence, we tried a
range of values for the threshold to find the optimal value for
the F 1 measure, which is explained in the next Performance
Evaluation section.
For S1, values between 0.02 and 0.18 were used, and for
S2, values between 0.2 and 0.6 were used as a threshold θ.
These threshold values were empirically chosen but set to be a
wide enough range to find the best performance (i.e., highest
F 1 measure). The schematic of this aggregation process is
illustrated in Figure 2.
C. Performance Evaluation
Following the evaluation method widely used in the instrument recognition task, we computed the precision and recall,
which are defined as
P =
tp
tp + f p
(4)
tp
tp + f n
(5)
where tp is true positive, f p is false positive, and f n is false
negative. In addition, we used the F 1 measure to calculate
the overall performance of the system, which is the harmonic
mean between precision and recall:
F1 =
2P R
P +R
(6)
Since the number of annotations for each class (i.e., 11
musical instruments) was not equal, we computed the precision, recall, and F 1 measure for both the micro and the
macro averages. For the micro averages, we calculated the
metrics globally regardless of classes, thus giving more weight
to the instrument with a higher number of appearances. On
the other hand, we calculated the metrics for each label and
found their unweighted average for the macro averages; hence,
it is not related to the number of instances, but represents the
overall performance over all classes. Finally, we repeated each
experiment three times and calculated the mean and standard
deviation of the output.
V. R ESULTS
We used LReLU (α = 0.33) for the activation function, 1
s for the analysis window, S2 for the aggregation strategy,
and 0.50 for the identification threshold as default settings
of the experiment where possible, which showed the best
performance. The experiment variables are listed in Table III.
First, we compared the performance of the proposed ConvNet with that of the existing algorithm on the IRMAS dataset.
The effect of activation function, analysis window, aggregation
strategy, and identification threshold on the recognition performance was analyzed separately in the following subsections.
A. Comparison to Existing Algorithms
For the result, our network achieved 0.602 for the micro
F 1 measure and 0.503 for the macro F 1 measure. The
existing algorithm from Fuhrmann and Herrera [45] used
typical hand-made timbral audio features with their framewise mean and variance statistics to train SVMs, and Bosch et
al. [44] improved this algorithm with source separation called
FASST (Flexible Audio Source Separation Framework) [46]
in a preprocessing step.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
Fuhrmann
Bosch
Proposed
0.7
0.6
0.5
micro F1-measure
0.62
F1-measure
0.8
7
1.0s
1.5s
3.0s
0.60
0.58
0.56
0.54
0.4
0.2
0.1
Prec.
Rec.
Micro
F1
Prec.
Rec.
Macro
F1
Fig. 3. Performance comparison of the predominant instrument recognition
algorithm from Fuhrmann and Herrra [45], Bosch et al. [44], and our proposed
ConvNet.
In terms of precision, Fuhrmann and Herrera’s algorithm
showed the best performance for both the micro and the
macro measure. However, its recall was very low, around
0.25, which resulted in a low F 1 measure. Our proposed
ConvNet architecture outperformed existing algorithms on the
IRMAS dataset for both the micro and the macro F 1 measure,
as shown in Figure 3. From this result, it can be observed
that the learned feature from the input data that is classified
through ConvNet works better than the conventional handcrafted features with SVMs.
B. Effect of Activation Function
In the case of using rectified units as an activation function,
it was possible to observe a significant performance improvement compared to the tanh baseline as expected, as shown
in Table IV. Unlike the result presented in the ImageNet
classification work from He et al. [42], PReLU did not show
any performance improvement, but just showed a matching
performance with ReLU in our task. On the other hand, using
LReLU showed better performance than using normal ReLU
and PReLU. While using LReLU with a small gradient (α
= 0.01) showed similar performance to ReLU as expected,
LReLU with a very leaky alpha setting (α = 0.33) showed the
best identification performance, which matched the result of
the empirical evaluation work on ConvNet activation function
from Xu et al. [43].
This result shows that suppressing the negative part of the
activation rather than making it all zero certainly improves the
performance compared to normal ReLU because making the
whole negative part zero might cause some initially inactive
units to be never active as mentioned above. Moreover, this
result shows that using leaky ReLU, which has been proved
to work well in the image classification task, also benefits the
musical instrument identification.
F1-measure
0.3
0.0
0.5s
0.52
0.2
0.25
0.3
0.51
0.50
0.49
0.48
0.47
0.46
0.45
0.44
0.43
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0.55
0.6
0.35 0.4 0.45 0.5
identification threshold
0.55
0.6
macro F1-measure
Fig. 4. Micro and macro F 1 measure of an analysis window size of 0.5, 1.0,
1.5, and 3.0 s according to the identification threshold.
TABLE IV
I NSTRUMENT R ECOGNITION P ERFORMANCE OF THE P ROPOSED
C ONV N ET WITH VARIOUS ACTIVATION F UNCTIONS .
Activation func.
tanh
ReLU
PReLU
LReLU (α=0.01)
LReLU (α=0.33)
P
Micro
R
F1
P
Macro
R
F1
0.416
0.640
0.612
0.640
0.655
0.625
0.550
0.565
0.552
0.557
0.499
0.591
0.588
0.593
0.602
0.348
0.521
0.502
0.530
0.541
0.537
0.508
0.516
0.507
0.508
0.399
0.486
0.490
0.492
0.503
C. Effect of Analysis Window Size
As mentioned above, we conducted an experiment with
diverse analysis window sizes such as 3.0, 1.5, 1.0, and 0.5
s to find the optimal analysis resolution. Figure 4 shows the
micro and macro F 1 measure with various analysis frame sizes
according to identification threshold, and it can be observed
that the use of the longest 3.0-s window clearly performed
poorer than the use of shorter window sizes regardless of identification threshold. However, shortening the analysis frame
down to 0.5 s decreased the overall performance again.
From this result, it can be seen that 1.0 s is the optimal
analysis window size for our task. Using a shorter analysis
frame helped to increase the temporal resolution, but 0.5 s
was found to be too short a window size for identifying the
instrument.
D. Effect of Identification Threshold
Using a higher value for the identification threshold leads
to better precision, but it decreases the recall. On the contrary,
a lower threshold results in better recall with lower precision.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
1.0
F1-measure
0.8
0.6
8
Optimal setting
Agg. strategy S2 to S1
Analysis win. 1s to 3s
Identification thr. 0.50 to 0.20
0.4
0.2
0.0
cel
cla
flu
acg
elg
org
pia
sax
tru
vio
voi
Fig. 5. Class-wise performance of the instrument identification. To analyze the effect of each parameter on each instrument, we compared the optimal setting
(i.e., default setting) with a different aggregation strategy, analysis window size, and LReLU identification threshold.
TABLE V
I NSTRUMENT R ECOGNITION P ERFORMANCE OF THE P ROPOSED
C ONV N ET U SING T WO D IFFERENT AGGREGATION S TRATEGIES WITH
THE T HRESHOLD θ THAT R ETURNED THE H IGHEST F 1 M EASURE FOR
E ACH S TRATEGY FOR C OMPARISON .
Agg. strategy
S1 (θ = 0.16)
S2 (θ = 0.50)
P
Micro
R
F1
0.627
0.655
0.572
0.557
0.595
0.602
P
Macro
R
F1
0.525
0.541
0.502
0.508
0.486
0.503
Hence, we used the F 1 measure, which is the harmonic mean
of precision and recall to evaluate the overall performance. In
terms of F 1 measure, we found that 0.5 is the most appropriate
threshold as it showed the best performance for the macro F 1
measure, as shown in Figure 4.
The current system uses a certain identification threshold
for all instruments. However, we think that there might be
a room for improvement by using different thresholds for
each instrument because there are various types of instruments
included in the experiment. For example, the amplitude of the
piano sound was relatively small in a number of music excerpts
because it is usually used as an accompanying instrument. On
the other hand, the flute sound in the music was mostly louder
than others because it is usually used as a lead instrument.
E. Effect of Aggregation Strategy
We conducted an experiment with two different strategies,
S1 and S2, for the aggregation of ConvNet outputs as explained
in the Testing Configuration section. The performance of S1
and S2 is demonstrated in Table V with a threshold θ that
returned the highest F 1 measure for each strategy. As a
result, S2 showed better identification performance than S1
overall. There was only a slight performance gap between
S1 and S2 for the micro F 1 measure, but the difference
was notable for the macro F 1 measure. This result shows
that performing a class-wise sum followed by normalization
is a better aggregation method for predominant instrument
identification than taking class-wise mean values. It is likely
due to the training and testing audios differing in quality to a
great extent, depending on the recording and production time,
and the audio-excerpt-wise normalization helped to minimize
the effect of quality differences between audio excerpts, which
would result in a more generalized output.
F. Analysis of Instrument-Wise Identification Performance
The results demonstrated above were focused on the overall
identification performance. In this section, we analyze and
discuss the result instrument-wise (i.e., class-wise) to observe
the system performance in detail. As shown in Figure 5,
identification performance varies to a great extent, depending
on the instruments. Regardless of parameter setting, it can be
observed that the system recognizes the voice in the music
very well, showing an F 1 measure of about 0.90. On the other
hand, cello and clarinet showed relatively poor performance
compared to other instruments, showing an F 1 measure of
around 0.20.
These results were highly likely affected by the insufficient
number of training audio samples. For deep learning, the
number of training examples is critical for the performance
compared to the case of using hand-crafted features because
it aims to learn a feature from the low-level input data. As
illustrated in Table II, the number of training audio samples
for voice is 778, which is the largest number of training audio.
On the contrary, 338 and 505 audio excerpts were used for
cello and clarinet, respectively, which were the least and third
least number of training audio. We believe that increasing
the number of training data for cello and clarinet would be
helpful to increase the identification performance for these
instruments.
In addition, the number of test audio samples for cello and
clarinet was much less than those for other instruments too.
The dataset only has 111 and 62 test audio samples for cello
and clarinet, respectively, which are the first and second least
number of test audio, while it has 1044 audio samples for the
human voice. Evaluating the system on a small number of test
data would make the result less reliable and less stable than
other identification results.
Apart from the issue related to the number of audio,
high identification performance of the voice class is highly
likely owing to its spectral characteristic that is distinct from
other musical instruments. The other instruments used in the
experiment usually produce relatively clear harmonic patterns;
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
cel
cla
flu
gac
9
gel
org
pia
sax
tru
vio
voi
Fig. 6. Visualization of the t-SNE clustering result. It represents the clustering result for each intermediate state of the proposed model. From left to right,
the first four plots are the clustering result of the activations at the end of each convolutional block, and the last two plots are the clustering result of the
activations of the hidden dense layer and the final sigmoid output, respectably. The upper plots are drawn from the sample used in the training, and the lower
plots are from the validation data samples.
however, the human voice produces highly unique spectral
characteristics that contain much more inharmonic spectrum
with a natural vibrato.
Regarding aggregation strategy, using S1 instead of S2
decreased the identification performance for organ and piano.
This result indicates that S1 showed a slight advantage on
instruments that are usually used as an accompaniment instrument, while using S1 for aggregation was better for most of
the cases. On the other hand, using a 3-s analysis window
instead of the default 1-s window considerably decreased
the performance, especially for flute, acoustic guitar, electric
guitar, and violin. This result shows that using a longer
analysis window is a disadvantage for most of the cases.
Finally, using a very low identification threshold, 0.20, caused
considerable performance loss especially for flute, saxophone,
trumpet, and violin, while it showed a slight improvement for
electric guitar, organ, and piano.
This result can be understood to mean that using a lower
threshold for identification performance helps to detect instruments that are usually in the background, while using a
higher threshold is suitable for instruments that are frequently
used as a lead instrument or for wind instruments that usually
show relatively strong presence in the music. As mentioned
in the Results section, this result indicates that there can be
a potential performance improvement by using a different
identification threshold for each instrument.
G. Qualitative Analysis with Visualization Methods
To understand the internal mechanism of the proposed
model, we conducted a visual analysis with various visualization methods. First, we tried clustering for each layer’s
intermediate hidden states from a given input data sample
to verify how the encoding behavior of each layer contributes to the clustering of input samples. We selected the
t-distributed stochastic neighbor embedding (t-SNE) [47] algorithm, which is a technique for dimensionality reduction of
high-dimensional data. Second, we exploited the deconvolution [33], [48] method to identify the functionality of each
unit in the proposed ConvNet model by visual analysis. Our
system basically repeats two convolutional layers followed by
one pooling layer, and we grouped these three components
and call it “convolutional block” throughout this section for
simplicity.
The t-SNE algorithm is based on the stochastic neighbor
embedding (SNE) algorithm, which converts the similarities between given data points to joint probability and then
embeds high-dimensional data points to lower-dimensional
space by minimizing the Kuller-Leibler divergence between
the joint probability of low-dimensional embedding and the
high-dimensional data points. This method is highly effective
especially in a dataset where its dimension is very high
[47]. This advantage of the algorithm accorded well with
our condition, where the target observations were necessarily
in a high dimension since we reshaped each layer’s filter
activations to a single vector respectively.
With the visualization exploiting t-SNE, we could observe
how each layer contributed to the classification of the dataset.
Reflecting a gradually changing inter-distance of data points at
each stage of the proposed model, four intermediate activations
were extracted at the end of each convolutional block and one
from the hidden fully connected layer, and another one from
the final output layer. For the compression of dimensionality
and computational efficiency, we pooled the maximum values
for activation matrices of each unit. By this process, the
dimensionality of each layer’s output could be diminished
to each layer’s unit size. We visualized on both randomly
selected training and validation data samples from the entire
dataset to verify both how the model exactly works and how it
generalizes its classification capability. In Figure 6, it is clearly
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
shown that data samples under the same class of instrument
are well grouped and each group is separated farther, with the
level of encoding being higher, particularly on the training set.
While the clustering was not clearer than the former case, the
tendency of clustering on the validation set was also found to
be similar to the training set condition.
Another visualization method, deconvolution, has recently
been introduced as a useful analysis tool to qualitatively
evaluate each node of a ConvNet. The main principle of
this method is to inverse every stage of operations reaching
to the target unit, to generate a visually inspectable image
that has been, as a consequence, filtered by the trained subfunctionality of the target unit [33]. With this method, it is
possible to reveal intuitively how each internal sub-function
works within the entire deep convolutional network, which
tends to be thought of as a “black box”.
By this process, the functionality of a sub-part of the
proposed model is explored. We generated deconvoluted images like those in Figure 7 from the arbitrary input melspectrogram, for each unit in the entire model. From the visual
analysis of the resulting images, we could see several aspects
of the sub-functionalities of the proposed model: (1) Most
units in the first layer tend to extract vertical, horizontal, and
diagonal edges from the input spectrogram, just like the lower
layers of ConvNets do in the usual image object recognition
task. (2) From the second layer through the fourth layer, each
deconvoluted image indicates that each unit of the mid-layers
has a functionality that searches for particular combinations of
the edges extracted from the first layer. (3) It was found that
it is difficult to strongly declare each sub-part of the proposed
model that detects a specific musical articulation or expression.
However, in an inductive manner, we could see that some units
indicate that they can be understood as a sub-function of such
musical expression detector.
We conducted a visual analysis of the deconvoluted image
of two independent music signals, which have the same kind
of sound sources, but differently labeled.1 For both cases, the
most activated units of the first layer strongly suggested that
their primary functionality is to detect a harmonic component
on the input mel-spectrogram by finding horizontal edges in
it, as shown in the top figures in Figure 7. However, from
the second layer to higher layers, the highly activated units’
behavior appeared to be quite different for each respective
input signal. For instance, the most activated unit of signal
(A)’s second layer showed a functionality similar to onset
detection, by detecting a combination of vertical and horizontal
edges. Compared to this unit, the most activated units of
the third layer showed a different functionality that seems to
activate unstable components such as the vibrato articulation
or the “slur” of the singing voice part, by detecting a particular
combination of diagonal and horizontal edges. On the other
hand, the model’s behavior in signal (B) was very different.
As is clearly shown in the second and the third layers’ output
in Figure 7, the highly activated sub-functions were trying
to detect a dense field of stable, horizontal edges which are
1 Both signals were composed of a “voice” and an “acoustic guitar”
instrument, but the predominant instrument of signal (A) was labeled as the
“voice,” while (B) was labeled as the “acoustic guitar”.
10
often found in harmonic instruments like guitar. Each field
detected from those units corresponded to the region where
the strumming acoustic guitar sound is.
VI. C ONCLUSION
In this paper, we described how to apply ConvNet to identify
predominant instrument in the real-world music. We trained
the network using fixed-length single-labeled data, and identify
an arbitrary number of the predominant instrument in a music
clip with a variable length.
Our results showed that very deep ConvNet is capable of
achieving good performance by learning the appropriate feature automatically from the input data. Our proposed ConvNet
architecture outperformed previous state-of-the-art approaches
in a predominant instrument identification task on the IRMAS
dataset. Mel-spectrogram was used as an input to the ConvNet,
and we did not use any source separation in the preprocessing
unlike in existing works.
We conducted several experiments with various activation
functions for ConvNet. Tanh and ReLU were used as a baseline, and the recently introduced LReLU and PReLU were also
evaluated. Results confirmed that ReLU worked reasonably
well, which is a de facto standard in recent ConvNet studies.
Furthermore, we obtained the better results with LReLU than
with normal ReLU, especially with the very leaky setting (α
= 0.33). The performance of Tanh was worse than those of
other rectifier functions as expected, and PReLU just showed
a matching performance with ReLU for our task.
This paper also investigated different aggregation methods
for ConvNet outputs that can be applied to music excerpts
with various lengths. We experimented with two different aggregation methods, which are the class-wise mean probability
S2 and the class-wise sum followed by normalization S2. The
experimental results showed that S2 is a better aggregation
method because it effectively deals with the quality difference
between audios through the audio-excerpt-wise normalization
process. In addition, we conducted an extensive experiment
with various analysis window sizes and identification thresholds. For the analysis window size, using a shorter window
improved the performance by increasing the temporal resolution. However, 0.5 s was too short to obtain an accurate
identification performance, and 1.0 s was found to be the
optimal window size. There was a trade-off between precision
and recall, depending on the identification threshold; hence, we
used an F 1 measure, which is the harmonic mean of precision
and recall. For the result, a threshold value of 0.5 showed the
best performance.
Visualization of the intermediate outputs using t-SNE
showed that the feature representation became clearer each
time the input data were passed through the convolutional
blocks. Moreover, visualization using deconvolution showed
that the lower layer tended to capture the horizontal and
vertical edges, and that the higher layer tended to seek the
combination of these edges to describe the spectral characteristics of the instruments.
Our study shows that many recent advances in a neural
network on the image processing area are transferable to the
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
11
LAYER 1
LAYER 1
LAYER 2
LAYER 2
LAYER 3
LAYER 3
LAYER 4
LAYER 4
(A)
(B)
Fig. 7. Mel-spectrogram of two input signals and their respective deconvoluted results. The left two columns and right two columns of the image, denoted as
(A) and (B), respectively, were calculated from two independent music signals. Both signals were a 3-s polyphonic music segment that was randomly cropped
from the original music. Moreover, both signals (A) and (B) consist mainly of the voice and the acoustic guitar sound. However, the dominant instrument of
(A) is labeled as the voice, while (B) is labeled as the acoustic guitar. Each row of images represents a deconvoluted signal overlaid on the original signal. We
extracted these results from four intermediate stages of the proposed model. Deconvolution outputs were extracted from the end of each convolutional block.
For both target signals, the two highest activated units of each point were chosen and deconvoluted to be visualized. From left to right, images are arranged
in order of decreasing absolute unit activation. The red region and the green region of each deconvoluted image indicate the positive value and negative value
of the result, respectably. The remaining area is where the magnitude of activation is relatively lower than in those regions. The range of activation result is
normalized for the purpose of clear visualization.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
audio processing domain. However, audio signal processing,
especially music signal processing, has many different aspects
compared to the image processing area where ConvNets are
most extensively used. For example, spectral characteristics
are usually overlapped in both time and frequency unlike the
objects in an image, which makes the detection difficult. Moreover, music signals are much more repetitive and continuous
compared to natural images and are present in various lengths.
We believe that applying more musical knowledge on the aggregation part with adaptive thresholding for each instrument
can improve the performance further, which warrants deeper
investigation.
ACKNOWLEDGMENT
This research was supported partly by the MSIP (Ministry
of Science, ICT and Future Planning), Korea, under the ITRC
(Information Technology Research Center) support program
(IITP-2016-H8501-16-1016) supervised by the IITP (Institute
for Information & communications Technology Promotion),
and partly by a National Research Foundation of Korea (NRF)
grant funded by the MSIP (NRF-2014R1A2A2A04002619).
R EFERENCES
[1] A. Eronen and A. Klapuri, “Musical instrument recognition using cepstral coefficients and temporal features,” in Acoustics, Speech, and Signal
Processing, 2000. ICASSP’00. Proceedings. 2000 IEEE International
Conference on, vol. 2. IEEE, 2000, pp. II753–II756.
[2] A. Diment, P. Rajan, T. Heittola, and T. Virtanen, “Modified group
delay feature for musical instrument recognition,” in 10th International
Symposium on Computer Music Multidisciplinary Research (CMMR).
Marseille, France, 2013.
[3] L.-F. Yu, L. Su, and Y.-H. Yang, “Sparse cepstral codes and power
scale for instrument identification,” in Acoustics, Speech and Signal
Processing (ICASSP), 2014 IEEE International Conference on. IEEE,
2014, pp. 7460–7464.
[4] A. Krishna and T. V. Sreenivas, “Music instrument recognition: from
isolated notes to solo phrases,” in Acoustics, Speech, and Signal Processing, 2004. Proceedings.(ICASSP’04). IEEE International Conference on,
vol. 4. IEEE, 2004, pp. iv–265.
[5] S. Essid, G. Richard, and B. David, “Musical instrument recognition
on solo performances,” in Signal Processing Conference, 2004 12th
European. IEEE, 2004, pp. 1289–1292.
[6] T. Heittola, A. Klapuri, and T. Virtanen, “Musical instrument recognition
in polyphonic audio using source-filter model for sound separation.” in
ISMIR, 2009, pp. 327–332.
[7] T. Kitahara, M. Goto, K. Komatani, T. Ogata, and H. G. Okuno,
“Instrument identification in polyphonic music: Feature weighting to
minimize influence of sound overlaps,” EURASIP Journal on Applied
Signal Processing, vol. 2007, no. 1, pp. 155–155, 2007.
[8] Z. Duan, B. Pardo, and L. Daudet, “A novel cepstral representation for
timbre modeling of sound sources in polyphonic mixtures,” in Acoustics,
Speech and Signal Processing (ICASSP), 2014 IEEE International
Conference on. IEEE, 2014, pp. 7495–7499.
[9] M. Goto, H. Hashiguchi, T. Nishimura, and R. Oka, “Rwc music
database: Music genre database and musical instrument sound database.”
in ISMIR, vol. 3, 2003, pp. 229–230.
[10] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, pp. 436–444, 2015.
[11] L. Deng and D. Yu, “Deep learning: Methods and applications,” Foundations and Trends in Signal Processing, vol. 7, no. 3–4, pp. 197–387,
2014.
[12] T. Mikolov, M. Karafiát, L. Burget, J. Cernockỳ, and S. Khudanpur,
“Recurrent neural network based language model.” in INTERSPEECH,
vol. 2, 2010, p. 3.
[13] G. Mesnil, X. He, L. Deng, and Y. Bengio, “Investigation of recurrentneural-network architectures and learning methods for spoken language
understanding.” in INTERSPEECH, 2013, pp. 3771–3775.
12
[14] K. Yao, G. Zweig, M.-Y. Hwang, Y. Shi, and D. Yu, “Recurrent neural
networks for language understanding.” in INTERSPEECH, 2013, pp.
2524–2528.
[15] Y. LeCun, L. Jackel, L. Bottou, C. Cortes, J. S. Denker, H. Drucker,
I. Guyon, U. Muller, E. Sackinger, P. Simard et al., “Learning algorithms
for classification: A comparison on handwritten digit recognition,”
Neural networks: the statistical mechanics perspective, vol. 261, p. 276,
1995.
[16] A. Calderón, S. Roa, and J. Victorino, “Handwritten digit recognition
using convolutional neural networks and gabor filters,” Proc. Int. Congr.
Comput. Intell, 2003.
[17] X.-X. Niu and C. Y. Suen, “A novel hybrid cnn–svm classifier for
recognizing handwritten digits,” Pattern Recognition, vol. 45, no. 4, pp.
1318–1325, 2012.
[18] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105.
[19] J. Ngiam, Z. Chen, D. Chia, P. W. Koh, Q. V. Le, and A. Y. Ng, “Tiled
convolutional neural networks,” in Advances in Neural Information
Processing Systems, 2010, pp. 1279–1287.
[20] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly,
A. Senior, V. Vanhoucke, P. Nguyen, T. N. Sainath et al., “Deep neural
networks for acoustic modeling in speech recognition: The shared views
of four research groups,” Signal Processing Magazine, IEEE, vol. 29,
no. 6, pp. 82–97, 2012.
[21] P. Hamel, S. Lemieux, Y. Bengio, and D. Eck, “Temporal pooling
and multiscale learning for automatic annotation and ranking of music
audio.” in ISMIR, 2011, pp. 729–734.
[22] J. Schluter and S. Bock, “Improved musical onset detection with convolutional neural networks,” in Acoustics, Speech and Signal Processing
(ICASSP), 2014 IEEE International Conference on. IEEE, 2014, pp.
6979–6983.
[23] E. J. Humphrey and J. P. Bello, “Rethinking automatic chord recognition with convolutional neural networks,” in Machine Learning and
Applications (ICMLA), 2012 11th International Conference on, vol. 2.
IEEE, 2012, pp. 357–362.
[24] N. Boulanger-Lewandowski, Y. Bengio, and P. Vincent, “Audio chord
recognition with recurrent neural networks.” in ISMIR, 2013, pp. 335–
340.
[25] K. Ullrich, J. Schlüter, and T. Grill, “Boundary detection in music
structure analysis using convolutional neural networks.” in ISMIR, 2014,
pp. 417–422.
[26] T. Grill and J. Schlüter, “Music boundary detection using neural networks on combined features and two-level annotations,” in Proceedings
of the 16th International Society for Music Information Retrieval Conference (ISMIR 2015), Malaga, Spain, 2015.
[27] T. Park and T. Lee, “Musical instrument sound classification with deep
convolutional neural network using feature fusion approach,” arXiv
preprint arXiv:1512.07370, 2015.
[28] P. Li, J. Qian, and T. Wang, “Automatic instrument recognition in
polyphonic music using convolutional neural networks,” arXiv preprint
arXiv:1511.05520, 2015.
[29] Y. Hoshen, R. J. Weiss, and K. W. Wilson, “Speech acoustic modeling
from raw multichannel waveforms,” in Acoustics, Speech and Signal
Processing (ICASSP), 2015 IEEE International Conference on. IEEE,
2015, pp. 4624–4628.
[30] D. Palaz, R. Collobert, and M. M. Doss, “Estimating phoneme class
conditional probabilities from raw speech signal using convolutional
neural networks,” arXiv preprint arXiv:1304.1018, 2013.
[31] J. Nam, J. Herrera, M. Slaney, and J. O. Smith, “Learning sparse feature
representations for music annotation and retrieval.” in ISMIR, 2012, pp.
565–570.
[32] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[33] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” in Computer vision–ECCV 2014. Springer, 2014, pp.
818–833.
[34] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun, “Overfeat: Integrated recognition, localization and detection using
convolutional networks,” arXiv preprint arXiv:1312.6229, 2013.
[35] M. Lin, Q. Chen, and S. Yan, “Network in network,” arXiv preprint
arXiv:1312.4400, 2013.
[36] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
arXiv preprint arXiv:1412.6980, 2014.
[37] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: A simple way to prevent neural networks from over-
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, MAY 2016
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
fitting,” The Journal of Machine Learning Research, vol. 15, no. 1, pp.
1929–1958, 2014.
X. Glorot and Y. Bengio, “Understanding the difficulty of training deep
feedforward neural networks,” in International conference on artificial
intelligence and statistics, 2010, pp. 249–256.
J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liu,
X. Wang, and G. Wang, “Recent advances in convolutional neural
networks,” arXiv preprint arXiv:1512.07108, 2015.
V. Nair and G. E. Hinton, “Rectified linear units improve restricted boltzmann machines,” in Proceedings of the 27th International Conference
on Machine Learning (ICML-10), 2010, pp. 807–814.
A. L. Maas, A. Y. Hannun, and A. Y. Ng, “Rectifier nonlinearities
improve neural network acoustic models,” in Proc. ICML, vol. 30, 2013,
p. 1.
K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers:
Surpassing human-level performance on imagenet classification,” in
Proceedings of the IEEE International Conference on Computer Vision,
2015, pp. 1026–1034.
B. Xu, N. Wang, T. Chen, and M. Li, “Empirical evaluation of rectified
activations in convolutional network,” arXiv preprint arXiv:1505.00853,
2015.
J. J. Bosch, J. Janer, F. Fuhrmann, and P. Herrera, “A comparison of
sound segregation techniques for predominant instrument recognition in
musical audio signals.” in ISMIR, 2012, pp. 559–564.
F. Fuhrmann and P. Herrera, “Polyphonic instrument recognition for exploring semantic similarities in music,” in Proc. of 13th Int. Conference
on Digital Audio Effects DAFx10, 2010, pp. 1–8.
N. Ono, K. Miyamoto, H. Kameoka, J. Le Roux, Y. Uchiyama,
E. Tsunoo, T. Nishimoto, and S. Sagayama, “Harmonic and percussive
sound separation and its application to mir-related tasks,” in Advances
in music information retrieval. Springer, 2010, pp. 213–236.
L. Van der Maaten and G. Hinton, “Visualizing data using t-sne,” Journal
of Machine Learning Research, vol. 9, no. 2579-2605, p. 85, 2008.
J. Yosinski, J. Clune, A. Nguyen, T. Fuchs, and H. Lipson, “Understanding neural networks through deep visualization,” arXiv preprint
arXiv:1506.06579, 2015.
Yoonchang Han was born in Seoul, Republic of
Korea, in 1986. He studied electronic engineering
systems at King’s College London, UK, from 2006
to 2009, and then moved to Queen Mary University
of London, UK, and received an MEng (Hons)
degree in digital audio and music system engineering
with First Class Honours in 2011. He is currently a
PhD candidate in digital contents and information
studies at the Music and Audio Research Group
(MARG), Seoul National University, Republic of
Korea. His main research interest lies within developing deep learning techniques for automatic musical instrument recognition.
Jaehun Kim was born in Seoul, Republic of Korea,
in 1986. He is currently a researcher at the Music
and Audio Research Group. His research interests
include signal processing and machine learning techniques applied to music and audio. He received a
BA in English literature and linguistics from Seoul
National University and received an MS degree in
the digital contents and information studies at the
Music and Audio Research Group (MARG), Seoul
National University, Republic of Korea.
13
Kyogu Lee is an associate professor at Seoul National University and leads the Music and Audio
Research Group. His research focuses on signal
processing and machine learning techniques applied
to music and audio. Lee received a PhD in computerbased music theory and acoustics from Stanford
University.
| 9cs.NE
|
The Spin Group in Superspace
Hennie De Schepper, Alí Guzmán Adán, Frank Sommen
Clifford Research Group, Department of Mathematical Analysis, Faculty of Engineering and
Architecture, Ghent University, Krijgslaan 281, 9000 Gent, Belgium.
arXiv:1804.00963v1 [math.GR] 3 Apr 2018
Abstract
There are two well-known ways of describing elements of the rotation group SO(m). First,
according to the Cartan-Dieudonné theorem, every rotation matrix can be written as an even
number of reflections. And second, they can also be expressed as the exponential of some
anti-symmetric matrix.
In this paper, we study similar descriptions of the corresponding extension of SO(m) to
superspace. The setting is natural to describe the behavior of bosonic and fermionic particles.
This group of super-rotations SO0 is also an extension of the symplectic group. While still
being connected, it is thus no longer compact. As a consequence, it cannot be fully described
by just one action of the exponential map on its Lie algebra. Instead, we obtain an Iwasawatype decomposition for this group in terms of three exponentials acting on three direct
summands of the corresponding Lie algebra of supermatrices.
At the same time, SO0 strictly contains the group generated by super-vector reflections.
Therefore, its Lie algebra is isomorphic to a certain extension of the algebra of superbivectors.
This means that the Spin group in superspace has to be seen as the group generated by the
exponentials of the so-called extended superbivectors in order to cover SO0 . We also study
the actions of this Spin group on supervectors and provide a proper subset of it that is a
double cover of SO0 . Finally, we show that every fractional Fourier transform in n bosonic
dimensions can be seen as an element of the spin group in superspace.
Keywords. Spin groups, symplectic groups, Clifford analysis, superspace
Mathematics Subject Classification (2010). 30G35 (22E60)
1
Introduction
Mathematical analysis on superspace (super analysis) is a very important tool in the study of several
branches of contemporary theoretical physics such as supergravity or superstring theories. It was introduced by Berezin, see [1], and possesses a huge mathematical impact since it breaks several traditional
patterns of classical analysis, such as commutation among coordinate variables and positive dimensions.
Indeed, elements in superspace are defined by means of bosonic (commuting) variables xj and fermionic
(anti-commuting) variables x`j , where, in a natural way, the bosonic variables describe positive dimensions
while the fermionic ones describe negative dimensions.
Generalizations of some objects from classical analysis appear in superspace with a wider nature.
For example, the notion of a supervector variable is introduced as a variable that takes values in some
Grassmann envelope and, through the corresponding definitions of bosonic and fermionic partial derivatives ∂xj and ∂x`j , the supervector derivative (supergradient or super Dirac operator) can be introduced
as well. These analogies lead to the introduction of Clifford analysis in superspace as it was done in the
papers [2, 3, 13, 14, 16].
A proper extension of Clifford analysis to this setting needs a set of rules (or axioms) that guarantee
the preservation of the Clifford nature in the new objects, although they can have a very different behavior
compared to the classical ones. This set of rules is provided by the notion of a radial algebra, which is an
abstract setting that algebraically describes all main properties which any Clifford environment should
1
satisfy, see [3, 5, 12, 15]. This approach does not depend on an a priori defined vector space or signature.
In particular, the main exigence is that the anticommutator {x, y} of every pair of vector variables is a
commuting element, leading to the definition of an abstract inner product by the formula
1
hx, yi = − {x, y}.
2
(1)
However this abstract description leaves some issues unclarified. In particular, it is not possible to describe
the whole set of linear transformations leaving the above bilinear form invariant and in consequence, there
is no proper definition for the spin group in this general abstract setting. We can, of course, define vector
reflections by ψ(w)[x] = wxw where w, x are abstract vector variables. Yet this doesn’t enable us to
prove that all linear transformations leaving (1) invariant can be written as the composition of vector
reflections. In order to find the set of those transformations it is necessary to work with a concrete
representation of the radial algebra, i.e., an underlying vector space V where the vector variables are
defined endowed with a fixed bilinear form and signature.
For example, in the Clifford polynomial representation of the radial algebra (see [5, 12]), we have the
Euclidean space V = Rm where the most important group leaving the inner product (1) invariant is the
set of rotations SO(m). The spin group appears in this setting as a double cover of SO(m) given by
2k
Y
Spin(m) :=
wj : k ∈ N, wj ∈ Sm−1 ,
j=1
where Sm−1 = {w ∈ Rm : w2 = −1} denotes the unit sphere in Rm . The relation between Spin(m) and
SO(m) is easily seen through the Lie group representation h : Spin(m) → SO(m)
h(s)[x] = sxs,
s ∈ Spin(m), x ∈ Rm ,
which describes the action of every element of SO(m) in terms of Clifford multiplication. Such a description of the spin group follows from the Cartan-Dieudonné theorem which states that every orthogonal
transformation in an m-dimensional symmetric bilinear space can be written as the composition of at
most m reflections.
The situation may however not be the same in another representation of the radial algebra where
the Cartan-Dieudonné theorem is no longer valid. In this paper we deal with one of those cases: the
representation in superspace. Our main goal is to properly define the (super) spin group as a set of
elements describing every superrotation through Clifford multiplication in superspace. To that end, we
consider vector variables taking values in the Grassmann envelope Rm|2n (ΛN ). This makes it possible
to study the group of supermatrices leaving (1) invariant and to define in a proper way the spin group
in superspace. It is worth noticing that the superstructures are absorbed by the Grassmann algebras
leading to classical Lie groups and Lie algebras instead of supergroups or superalgebras.
The paper is organized as follows. We start with some preliminaries on Grassmann algebras, Grassmann envelopes and supermatrices in Section 2. In particular, we carefully recall the notion of an
exponential map for Grassmann numbers and supermatrices as elements of finite dimensional associative
algebras. In Section 3 we briefly describe the Clifford setting in superspace leading to the introduction of
the Lie algebra of superbivectors. An extension of this algebra plays an important rôle in the description
of the super spin group. The use of the exponential map in such an extension (which takes us out of the
radial algebra) necessitates the introduction of the corresponding tensor algebra. Section 4 is devoted to
the study of the invariance of the “inner product” (1) in this setting. There, we study several groups of
supermatrices and in particular, the group of superrotations SO0 and its Lie algebra so0 , which combine
both orthogonal and symplectic structures. We prove that every superrotation can be uniquely decomposed as the product of three exponentials acting in some special subspaces of so0 . Finally, in section 5
we study the problem of defining the spin group in this setting and its differences with the classical case.
We show that the compositions of even numbers of vector reflections are not enough to fully describe
SO0 since they only show an orthogonal structure and don’t include the symplectic part of SO0 . Next
we propose an alternative, by defining the spin group through the exponential of extended superbivectors
and show that they indeed cover the whole set of superrotations. In particular, we explicitly describe a
subset S which is a double covering of SO0 and contains in particular every fractional Fourier transform.
2
2
Preliminaries
2.1
Grassmann algebras and Grassmann envelopes
Let ΛN be the Grassmann algebra of order N ∈ N over the field K (K = R or C) with canonical generators
f1 , . . . , fN which are subject to the multiplication rules
fj fk + fk fj = 0,
(2)
implying in particular that fj2 = 0. Every element of ΛN can be written in the form
X
a=
aA fA ,
(3)
A⊂{1,...,N }
where aA ∈ K and fA = fj1 · · · fjk , for A = {j1 , . . . , jk }, with 1 ≤ j1 ≤ . . . ≤ jk ≤ N ; here we put f∅ = 1.
(k)
We define the space of homogeneous elements of degree k by ΛN = spanK {fA : |A| = k}, where in
(k)
particular ΛN = {0} for k > N . It then easily follows from (2)-(3) that
ΛN =
N
M
(k)
(k)
(`)
(k+`)
ΛN and ΛN ΛN ⊂ ΛN
k=0
P
(k)
The projection of ΛN on its k-homogeneous part is denoted by [·]k : ΛN → ΛN , i.e. [a]k = |A|=k aA fA .
In particular we denote [a]0 = a∅ =: a0 . It is well known that ΛN shows a natural Z2 -grading. In
L
L
(2k)
(2k+1)
as the spaces of homogeneous even and odd
fact, defining ΛN,0 = k≥0 ΛN and ΛN,1 = k≥0 ΛN
elements respectively, we obtain the superalgebra structure ΛN = ΛN,0 ⊕ ΛN,1 where ΛN,j ΛN,k ⊂ ΛN,j+k
for j, k ∈ Z2 . We recall that ΛN is graded commutative in the sense that
vw = wv,
vw`= w`v,
v`w`= −w`v`,
v, w ∈ ΛN,0 ,
v`, w`∈ ΛN,1 .
(0)
Any element a ∈ ΛN is the sum of a number a0 ∈ ΛN = K and a nilpotent element. In fact, every
LN
(k)
N +1
a ∈ Λ+
= 0. It is easily seen that the projection [·]0 : ΛN → K is an algebra
N :=
k=1 ΛN satisfies a
homomorphism, i.e. [ab]0 = a0 b0 for a, b ∈ ΛN . In particular the following property holds.
Lemma 1. Let a ∈ ΛN such that a2 ∈ K \ {0}. Then a ∈ K.
It sometimes is useful to consider ΛN embedded in a Grassmann algebra of higher order; in particular,
we consider ΛN ⊂ ΛN +1 := AlgK {f1 , . . . , fN , fN +1 } being fN +1 a new canonical Grassmann generator.
(k)
(k)
This embedding preserves all grading structures since ΛN ⊂ ΛN +1 and ΛN,j ⊂ ΛN +1,j for j ∈ Z2 .
The algebra ΛN is a K-vector space of dimension 2N . As every finite dimensional K-vector space,
ΛN becomes a Banach space with the introduction of an arbitrary norm, all norms being equivalent. In
particular, the following result is obtained through straightforward computation.
P
Lemma 2. The norm k · k1 defined on ΛN by kak1 = A⊂{1,...,N } |aA | satisfies
kabk1 ≤ kak1 kbk1 ,
for every a, b ∈ ΛN .
(4)
The exponential of a ∈ ΛN , denoted by ea or exp(a), is defined by the power series
ea =
∞
X
aj
j=0
(5)
j!
Proposition 1. The series (5) converges for every a ∈ ΛN and ea is a continuous function in ΛN .
Proof.
From (4) it follows that
∞
X
kaj k1
j=0
j!
≤
∞
X
kakj
1
j=0
3
j!
= ekak1
whence (5) (absolutely) converges in the Banach space ΛN . Now consider BR := {a ∈ ΛN : kak1 ≤ R}
for some R > 0, where it holds that
kaj k1
kakj1
Rj
≤
≤
j!
j!
j!
P∞ j
P∞ Rj
and j=0 j! converges, whence from the Weierstrass M -criterium we have that j=0 aj! uniformly cona
a
verges in BR and e thus is continuous in BR . Then, e is continuous in ΛN .
Now consider the graded vector space Kp|q with standard homogeneous basis e1 , . . . , ep , e`1 , . . . , e`q , i.e.
K = Kp|0 ⊕ K0|q where {e1 , . . . , ep } is a basis for Kp|0 and {e`1 , . . . , e`q } is a basis for K0|q . Elements in
Kp|0 are called even homogeneous elements while elements in K0|q are called odd homogeneous elements.
The Grassmann envelope Kp|q (ΛN ) is defined as the set of formal linear combinations of the form
p|q
x = x + x`=
p
X
xj ej +
j=1
q
X
where xj ∈ ΛN,0 ,
x`j e`j ,
x`j ∈ ΛN,1 .
(6)
j=1
Then Kp|q (ΛN ) is a K-vector space of dimension 2N −1 (p + q), inheriting the Z2 -grading of Kp|q , i.e.
Kp|q (ΛN ) = Kp|0 (ΛN ) ⊕ K0|q (ΛN ), where Kp|0 (ΛN ) denotes the subspace of vectors of the form (6) with
x`j = 0, and K0|q (ΛN ) denotes the subspace of vectors of the form (6) with xj = 0. The subspaces
Kp|0 (ΛN ) and K0|q (ΛN ) are called the Grassmann envelopes of Kp|0 and K0|q , respectively.
In Kp|q (ΛN ),Pthere exists a subspace which is naturally isomorphic to Kp|0 . It consists of
(6)
Pvectors
m
p
of the form x = j=1 xj ej with xj ∈ K. The map [·]0 : Kp|q (ΛN ) → Kp|0 defined by [x]0 = j=1 [xj ]0 ej
will be useful.
T
The standard basis of Kp|q can be represented by the columns ej = (0, . . . , 1, . . . , 0) (1 on the j-th
T
place from the left) and e`j = (0, . . . , 1, . . . , 0) (1 on the (p + j)-th place from the left). In this basis,
T
elements of Kp|q (ΛN ) take the form x = (x1 , . . . , xp , x`1 , . . . , x`q ) .
2.2
Supermatrices
The Z2 -grading of Kp|q yields the Z2 -grading of the space End Kp|q of endomorphisms on Kp|q . This
space is isomorphic to the space Mat(p|q) of block matrices of the form
A B`
A 0
0 B`
M=
=
+
(7)
0 D
C` D
C` 0
where1 A ∈ Kp×p , B` ∈ Kp×q , C` ∈ Kq×p and D ∈ Kq×q . The first term in (7) is the even part of M and
the second term is the odd one. The Grassmann envelope of Mat(p|q) is denoted by Mat(p|q)(ΛN ) and
consists of matrices of the form (7) but with entries in ΛN (namely, A, D with even entries and B`, C`
with odd entries). Elements in Mat(p|q)(ΛN ) are called supermatrices.
The Z2 -grading of Mat(p|q)(ΛN ), inherited from Mat(p|q), together with the usual matrix multiplication, provides a superalgebra structure on this Grassmann envelope. More precisely, for any k ∈ N let
(2k)
(k)
Mat(p|q)(ΛN ) be the space of all homogeneous supermatrices of degree k. This is, Mat(p|q)(ΛN ) con(2k)
(2k+1)
sists of all diagonal block matrices with entries in ΛN , and Mat(p|q)(ΛN
) consists of all off-diagonal
(2k+1)
block matrices with entries in ΛN
. These subspaces define a grading in Mat(p|q)(ΛN ) by
Mat(p|q)(ΛN ) =
N
M
(k)
Mat(p|q)(ΛN ),
and
(k)
(`)
(k+`)
Mat(p|q)(ΛN ) Mat(p|q)(ΛN ) ⊂ Mat(p|q)(ΛN
).
k=0
Then, it is clear that every supermatrix M can be written as the sum of a numeric matrix M0 ∈
LN
(0)
(k)
Mat(p|q)(ΛN ) and a nilpotent supermatrix M ∈ Mat(p|q)(Λ+
N ) :=
k=1 Mat(p|q)(ΛN ). In accordance
with the general ideas valid for Grassmann algebras and Grassmann envelopes we define the algebra
(0)
homomorphism [·]0 : Mat(p|q)(ΛN ) → Mat(p|q)(ΛN ) as the projection:
A B`
A0 0
M=
−→
= M0 = [M ]0
0 D0
C` D
1
Given a set S, we use the notation Sp×q to refer to the set of matrices of order p × q with entries in S.
4
where A0 and D0 are the numeric projections of A and D on Kp×p and Kq×q respectively. We recall that
(0)
Mat(p|q)(ΛN ) is equal to the even subalgebra of Mat(p|q). Given a set of supermatrices S we define
[S]0 = {[M ]0 : M ∈ S}.
Every supermatrix M defines a linear operator over Kp|q (ΛN ) which acts on a vector x = x + x`,
x ∈ Kp|0 (ΛN ), x`∈ K0|q (ΛN ) as
A B`
x
Ax + B`x`
Mx =
=
∈ Kp|q (ΛN ).
x`
C` D
C`x + Dx`
However, the supermatrix M ∈ Mat(p|q)(ΛN ) defining an endomorphism of Kp|q (ΛN ) is not unique. In
fact, the zero endomorphism is defined by every supermatrix on the ideal HN ⊂ Mat(p|q)(ΛN ) given by
0 0
0 B`
(N )
(N )
HN = L =
: L ∈ Mat(p|q)(ΛN ) or HN = L =
: L ∈ Mat(p|q)(ΛN ) ,
0 D
0 0
for N odd or even respectively. Hence, all supermatrices in M + HN define the same endomorphism over
Kp|q (ΛN ) as M . But this situation changes if we consider endomorphisms over Kp|q (ΛN +1 ) defined by
supermatrices with entries in ΛN .
Lemma 3. Let E be an endomorphism over Kp|q (ΛN +1 ). If E admits a supermatrix representation in
Mat(p|q)(ΛN ), then this representation is unique in Mat(p|q)(ΛN ).
This result is easily proved using the introduction of a new canonical Grassmann generator fN +1 to
show that an element a ∈ ΛN satisfies afN +1 = 0 if and only if a = 0.
To study group structures in Mat(p|q)(ΛN ) we start from the Lie group GL(p|q)(ΛN ) of all invertible
elements of Mat(p|q)(ΛN ). The following is a well-known characterization of this group, see [1].
A B`
Theorem 1. Let M =
∈ Mat(p|q)(ΛN ). Then the following statements are equivalent.
C` D
(i) M ∈ GL(p|q)(ΛN ).
(ii) A, D are invertible.
(iii) A0 , D0 are invertible.
In addition, for every M ∈ GL(p|q)(ΛN ) its inverse is
M
−1
=
−1
A − B`D−1 C`
−1
−D−1 C` A − B`D−1 C`
−1 !
−A−1 B` D − C`A−1 B`
−1
.
D − C`A−1 B`
The usual definitions of transpose, trace and determinant of a matrix are not appropriate in the
graded case. For example, although the transpose
T
A
C`T
MT =
B`T DT
of a supermatrix M is a well defined element of Mat(p|q)(ΛN ), the usual property (M L)T 6= LT M T will
not hold in general. This problem is fixed by introducing the supertranspose by
AT
C`T
M ST =
.
−B`T DT
The transpose and supertranspose operations satisfy the following relations, see [1].
p×q
and C` ∈ Λq×p
. Then,
Proposition 2. Let M, L ∈ Mat(p|q)(ΛN ), x ∈ Kp|q (ΛN ), B` ∈ ΛN,1
N,1
(i)
T
B`C` = −C`T B`T ,
ST
(ii) (M L)
= LST M ST ,
5
T
(iii) (M x) = xT M ST ,
A −B`
Ip
ST ST
2
(iv) M
=
= SM S, where S =
0
−C` D
ST
−1
(v) M −1
= M ST
for every M ∈ GL(p|q)(ΛN ).
0
−Iq
,
The situation for the trace is similar. The usual trace tr(M ) of an element M ∈ Mat(p|q)(ΛN ) is well
defined, but in general, tr(M L) 6= tr(LM ) for M, L ∈ Mat(p|q)(ΛN ). The notion of supertrace provides
a solution to this problem; it is defined as the map str : Mat(p|q)(ΛN ) → ΛN,0 given by
str
A B`
C` D
= tr(A) − tr(D).
The following properties easily follow from the above definition, see [1].
q×p
Proposition 3. Let M, L ∈ Mat(p|q)(ΛN ), B` ∈ Λp×q
and C` ∈ ΛN,1
. Then
N,1
(i) tr B`C` = − tr C`B` ,
(ii) str(M L) = str(LM ),
(iii) str M ST = str(M ).
The superdeterminant or Berezinian is a function from GL(p|q)(ΛN ) to ΛN,0 defined by:
sdet(M ) =
det(A)
det(A − B`D−1 C`)
=
.
det(D)
det(D − C`A−1 B`)
Some of its most basic properties are given in the following proposition, see [1].
Proposition 4. Let M, L ∈ GL(p|q)(ΛN ), then
i) sdet(M L) = sdet(M ) sdet(L),
ii) sdet M ST = sdet(M ).
Since Mat(p|q)(ΛN ) is a finite dimensional vector space, every two norms in this space are equivalent,
whence, without loss of generality, we can define for M = {mj,k }j,k=1,...,p+q ∈ Mat(p|q)(ΛN ) the norm
Pp+q
kM k = j,k=1 kmj,k k1 , satisfying kM Lk ≤ kM kkLk for every pair M, L ∈ Mat(p|q)(ΛN ).
Similarly to Proposition 1 the following result can be proven.
Proposition 5. For every M ∈ Mat(p|q)(ΛN ) the exponential defined by
eM =
∞
X
Mj
j=0
j!
absolutely converges. In consequence, eM is a continuous function in Mat(p|q)(ΛN ).
Also the supertranspose, the supertrace and the superdeterminant are continuous functions.
Proposition 6. Let M, L ∈ Mat(p|q)(ΛN ). Then
(i) e0 = Ip+q ;
ST
ST
(ii) eM
= eM ;
(iii) If M L = LM then eM +L = eM eL ;
−1
(iv) eM ∈ GL(p|q)(ΛN ) and eM
= e−M ;
2
Ik denotes the identity matrix in Rk×k .
6
(v) e(a+b)M = eaM ebM for every pair a, b ∈ ΛN,0 ;
(vi) eCM C
−1
= CeM C −1 for every C ∈ GL(p|q)(ΛN );
d tM
dt e
(vii) etM (t ∈ R) is a smooth curve in Mat(p|q)(ΛN ),
(viii) sdet eM = estr(M ) .
= M etM = etM M and
d tM
dt e
t=0
= M;
Remark 2.1. The proofs of (i)-(vii) are straightforward computations. A detailed proof for (viii) can be
found in [1]. Similar properties to (i) and (iii)-(vii) can be obtained for the exponential map in ΛN .
We also can define the notion of logarithm for a supermatrix M ∈ Mat(p|q)(ΛN ) by
ln(M ) =
∞
X
(−1)j+1
j=1
(M − Ip+q )j
,
j
(8)
wherever it converges.
Proposition 7.
(i) The series (8) converges and yields a continuous function near Ip+q .
(ii) In Mat(p|q)(ΛN ), let U be a neighbourhood of Ip+q on which ln is defined and let V be a neighbourhood of 0 such that exp(V ) := {eM : M ∈ V } ⊂ U . Then eln(M ) = M , ∀M ∈ U , and ln(eL ) = L,
∀L ∈ V .
Proof.
(i) Observe that
∞
X
(M − Ip+q )j
j
j=1
≤
∞
j
X
kM − Ip+q k
.
j
j=1
Whence, since the radius of convergence of the last series is 1, (8) absolutely converges and defines
a continuous function in the ball kM − Ip+q k < 1.
(ii) The statement immediately follows from the absolutely convergence of the series for exp and ln,
and from the formal identities eln x = ln(ex ) = x in the indeterminate x.
It is worth mentioning that the same procedure can be repeated in ΛN . With the above definitions
of the exponential and logarithmic maps, it is possible to obtain all classical results known for Lie groups
and Lie algebras of real and complex matrices .
The exponential of a nilpotent matrix M ∈ Mat(p|q)(Λ+
N ) clearly is given by a finite sum, which
yields the bijective mapping:
+
exp : Mat(p|q)(Λ+
N ) → Ip+q + Mat(p|q)(ΛN )
with inverse
+
ln : Ip+q + Mat(p|q)(Λ+
N ) → Mat(p|q)(ΛN ),
since both expansions have a finite number of non-zero terms, whence problems of convergence do not
arise.
We recall that a supermatrix M belongs to GL(p|q)(ΛN ) if and only if its numeric projection M0 has
an inverse. Then, M = M0 (Ip+q + M0−1 M) = M0 exp(L), for some unique L ∈ Mat(p|q)(Λ+
N ).
7
3
3.1
The superspace framework with Grassmann coefficients
The Clifford setting
In order to set up the Clifford analysis framework in superspace, take p = m, q = 2n (m, n ∈ N)
and K = R. The canonical homogeneous basis e1 , . . . , em , e`1 , . . . , e`2n of Rm|2n can be endowed with an
orthogonal and a symplectic structure by the multiplication rules
ej ek + ek ej = −2δj,k ,
ej e`k + e`k ej = 0,
e`j e`k − e`k e`j = gj,k ,
where the symplectic form gj,k is defined by
g2j,2k = g2j−1,2k−1 = 0,
g2j−1,2k = −g2k,2j−1 = δj,k ,
j, k = 1, . . . , n.
Following these relations, elements in Rm|2n generate an infinite dimensional algebra denoted by Cm,2n .
A similar approach in the Grassmann envelope Rm|2n (ΛN ) leads to the definition of the algebra
Am,2n (ΛN ) = AlgR (f1 , . . . , fN , e1 , . . . , em , e`1 , . . . , e`2n ) = ΛN ⊗ Cm,2n
where it is assumed that elements of ΛN commute with elements of Cm,2n .
α2n
Every element in Am,2n (ΛN ) can be written as a finite sum of terms of the form aej1 · · · ejk e`1α1 · · · e`2n
2n
where a ∈ ΛN , 1 ≤ j1 ≤ . . . ≤ jk ≤ m and (α1 , . . . , α2n ) ∈ (N ∪ {0}) is a multi-index. In this algebra
we consider the corresponding generalization of the projection [·]0 which now goes from Am,2n (ΛN ) to
α2n
α2n
.
]0 = [a]0 ej1 · · · ejk e`1α1 · · · e`2n
Cm,2n and is defined by [aej1 · · · ejk e`1α1 · · · e`2n
Developing a proper Clifford analysis in this framework requires a suitable realization of the radial
algebra in Am,2n (ΛN ). For a detailed treatment of the radial algebra setting we refer the reader to
[4, 5, 12, 15]. This realization in Am,2n (ΛN ) is seen through a set of vector variables taking values in the
Grassmann envelope Rm|2n (ΛN ), i.e., supervectors of the form:
x = x + x`=
m
X
xj ej +
j=1
2n
X
x`j e`j ,
xj ∈ ΛN,0 ,
x`j ∈ ΛN,1 ,
j=1
which satisfy the fundamental axiom of radial algebra, i.e. the anti-commutator of every pair of them is
a commuting element. Indeed, for every pair x, y ∈ Rm|2n (ΛN )
{x, y} = −2
m
X
j=1
xj yj +
n
X
(x`2j−1 y`2j − x`2j y`2j−1 ) ∈ ΛN,0 .
(9)
j=1
The algebra generated by all elements in Rm|2n (ΛN ) is called the radial algebra embedded in Am,2n (ΛN )
and will be denoted by Rm|2n (ΛN ). This algebra is generated by the set of elements
{fA ej : A ⊂ {1, . . . , N }, |A| even , j = 1, . . . , m} ∪ {fA e`j : A ⊂ {1, . . . , N }, |A| odd , j = 1, . . . , 2n} ,
which turns Rm|2n (ΛN ) into a finite dimensional real space. For more details about this kind of realization
of the radial algebra we refer the reader to [3, 4, 13, 14].
3.2
Superbivectors
Superbivectors play a very important rôle in this work. Following the radial algebra approach, the space
of bivectors can be defined as the space of all ΛN,0 -linear combinations of wedge products of supervectors,
i.e.
X
X
X
1
x ∧ y = [x, y] =
(xj yk − xk yj )ej ek +
(xj y`k − x`k yj )ej e`k +
(x`j y`k + x`k y`j ) e`j e`k ,
2
1≤j≤m
1≤j<k≤m
where e`j
1≤j≤k≤2n
1≤k≤2n
(2)
e`k = 12 {e`j , e`k }. Hence, the space Rm|2n (ΛN ) of superbivectors consists of elements of the form
X
X
X
B=
bj,k ej ek +
b`j,k ej e`k +
Bj,k e`j e`k ,
(10)
1≤j<k≤m
1≤j≤k≤2n
1≤j≤m
1≤k≤2n
8
where bj,k ∈ ΛN,0 , b`j,k ∈ ΛN,1 and Bj,k ∈ ΛN,0 ∩ Λ+
N, . Observe that the coefficients Bj,k are commuting
but nilpotent since they are generated by elements of the form x`j y`k + x`k y`j that belong to Λ+
N, . This
(2)
constitutes an important limitation for the space of superbivectors because it means that Rm|2n (ΛN )
doesn’t allow for any other structure than the orthogonal one. In
P fact, the real projection [B]0 of
every superbivector B is just the classical Clifford bivector: [B]0 = 1≤j<k≤m [bj,k ]0 ej ek . Thence it is
(2)E
(2)
necessary to introduce an extension Rm|2n (ΛN ) of Rm|2n (ΛN ) containing elements B of the form (10)
but with Bj,k ∈ ΛN,0 . This extension allows to consider two different structures in the same element B:
the orthogonal and the symplectic one. In fact, in this case we have
X
X
[B]0 =
[bj,k ]0 ej ek +
[Bj,k ]0 e`j e`k .
1≤j<k≤m
1≤j≤k≤2n
(2)
(2)E
Remark 3.1. Observe that Rm|2n (ΛN ) and Rm|2n (ΛN ) are finite dimensional real vector spaces with
m(m − 1)
+ 2N −1 2mn + 2N −1 − 1 n(2n + 1),
2
m(m
− 1)
(2)E
+ 2N −1 2mn + 2N −1 n(2n + 1).
dim Rm|2n (ΛN ) = 2N −1
2
(2)
dim Rm|2n (ΛN ) = 2N −1
(2)E
The extension Rm|2n (ΛN ) of the superbivector space clearly lies out of the radial algebra Rm|2n (ΛN )
and generates an infinite dimensional algebra using the multiplication rules defined in Am,2n (ΛN ). Ele(2)E
ments in Rm|2n (ΛN ) are called extended superbivectors. Both superbivectors and extended superbivectors
preserve several properties of classical bivectors.
(2)E
(2)
Proposition 8. The space Rm|2n (ΛN ) is a Lie algebra. In addition, Rm|2n (ΛN ) is a Lie subalgebra of
(2)E
Rm|2n (ΛN ).
Proof.
We only need to check that the Lie bracket defined by the commutator in the associative algebra
(2)E
(2)
Am,2n (ΛN ) is an internal binary operation in Rm|2n (ΛN ) and Rm|2n (ΛN ). Direct computation shows
that for a, b ∈ ΛN,0 and a`, b`∈ ΛN,1 we get:
[aej ek , ber es ] = ab (2δj,s er ek − 2δs,k er ej + 2δr,j ek es − 2δr,k ej es ),
[aej ek , b`er e`s ] = ab`(2δr,j ek e`s − 2δr,k ej e`s ),
[aej ek , be`r
e`s ] = 0,
e`s + (1 − δj,r )gs,k ej er ),
[a`ej e`k , b`er e`s ] = a`b`(2δr,j e`k
[a`ej e`k , be`r
[ae`j
e`k , be`r
e`s ] = a`b (gk,s ej e`r + gk,r ej e`s ),
e`s ] = ab (gj,s e`r
e`k + gk,s e`r
e`j + gj,r e`k
e`s + gk,r e`j
e`s ).
It is well known from the radial algebra framework that the commutator of a bivector with a vector
always is a linear combination of vectors with coefficients in the scalar subalgebra. Indeed, for vector
variables x, y, z we obtain
1
1
[x ∧ y, z] = [[x, y], z] = [2xy − {x, y}, z] = [xy, z] = {y, z}x − {x, z}y.
2
2
(2)E
This property can be easily generalized to Rm|2n (ΛN ) by straightforward computation. In particular, the
following result holds.
Proposition 9. Let x ∈ Rm|2n (ΛN ), let {b1 , . . . , b2N −1 } be a basis for ΛN,0 and let {b`1 , . . . , b`2N −1 } be a
basis for ΛN,1 . Then,
[br ej ek , x] = 2br (xj ek − xk ej ),
[b`r ej e`2k−1 , x] = b`r (2xj e`2k−1 + x`2k ej ),
[b`r ej e`2k , x] = b`r (2xj e`2k − x`2k−1 ej ),
[br e`2j
[br e`2j−1
[br e`2j−1
e`2k , x] = −br (x`2j−1 e`2k + x`2k−1 e`2j ),
e`2k−1 , x] = br (x`2j e`2k−1 + x`2k e`2j−1 ),
e`2k , x] = br (x`2j e`2k − x`2k−1 e`2j−1 ).
The above computations also are valid for supervectors x in any extension of Rm|2n (ΛN ) such as Rm|2n (ΛN +1 ).
9
3.3
Tensor algebra and exponential map
Since Am,2n (ΛN ) is infinite dimensional, the definition of the exponential map by means of the power
P∞ j
series ex = j=0 xj! is not as straightforward as it was for the algebras ΛN or Mat(p|q)(ΛN ). A correct
definition of the exponential map in Am,2n (ΛN ) requires the introduction of the tensor algebra. More
details about the general theory of tensor algebras can be found several basic references, see e.g. [6, 11, 17].
Let T (V )L
be the tensor algebra of the vector space V with BV = {f1 , . . . , fN , e1 , . . . , em , e`1 , . . . , e`2n },
∞
i.e., T (V ) = j=0 T j (V ) where T j (V ) = spanR {v1 ⊗ · · · ⊗ vj : v` ∈ BV } is the j-fold tensor product of
V with itself. Then Am,2n (ΛN ) can be seen as a subalgebra of T (V )/I where I ⊂ T (V ) is the twosided
ideal generated by the elements:
fj ⊗ fk + fk ⊗ fj ,
fj ⊗ ek − ek ⊗ fj ,
fj ⊗ e`k − e`k ⊗ fj ,
ej ⊗ ek + ek ⊗ ej + 2δj,k ,
ej ⊗ e`k + e`k ⊗ ej ,
e`j ⊗ e`k − e`k ⊗ e`j − gj,k .
Indeed, T (V )/I is isomorphic to the extension of Am,2n (ΛN ) which also contains infinite sums of
α2n
where a ∈ ΛN , 1 ≤ j1 ≤ . . . ≤ jk ≤ m and
arbitrary terms of the form aej1 · · · ejk e`1α1 · · · e`2n
2n
(α1 , . . . , α2n ) ∈ (N ∪ {0}) is a multi-index.
P∞ j
The exponential map exp(x) = ex = j=0 xj! is known to be well defined in the tensor algebra T (V ),
see e.g. [6], whence it also is well defined in T (V )/I. It has the following mapping properties:
exp : Am,2n (ΛN ) → T (V )/I,
exp : Rm|2n (ΛN ) → Rm|2n (ΛN ).
The first statement directly follows from the definition of T (V )/I, while the second one can be obtained following the standard procedure established for ΛN and Mat(p|q)(ΛN ), since Rm|2n (ΛN ) is finite
dimensional.
The ortho symplectic structure in Rm|2n (ΛN )
4
4.1
Invariance of the inner product
The ΛN,0 -valued function given by the anti-commutator of supervectors (9) leads to the definition of the
symmetric bilinear form h·, ·i : Rm|2n (ΛN ) × Rm|2n (ΛN ) −→ ΛN,0 given by
m
n
X
1X
1
xj yj −
(x`2j−1 y`2j − x`2j y`2j−1 ) ∈ ΛN,0 .
hx, yi = − {x, y} =
2
2 j=1
j=1
which we will use as a generalized inner product. It can be easily written in terms of supermatrices, since
Im
0
0 1
T
hx, yi = x Qy
where Q =
, and J2n = diag
.
−1 0
0 − 21 J2n
Pm
When N = 1, the inner product hx, yi = j=1 xj yj coincides with the Euclidean inner product in Rm .
So from now on, we assume N > 1.
In order to find all the supermatrices M ∈ Mat(m|2n)(ΛN ) the corresponding linear operators of
which leave the inner product h·, ·i invariant, observe that
T
hM x, M yi = hx, yi ⇐⇒ (M x) QM y = xT Qy ⇐⇒ xT M ST QM − Q y = 0.
We thus have to identify those supermatrices L ∈ Mat(m|2n)(ΛN ) for which
xT L y = 0
∀x, y ∈ Rm|2n (ΛN ).
(11)
Lemma 4. A supermatrix L ∈ Mat(m|2n)(ΛN ) satisfies (11) if and only if L ∈ WN where WN is a
twosided ideal of Mat(m|2n)(ΛN ) defined as follows:
• if N = 2` + 1, W2`+1 is generated by the set of supermatrices
0 B`
0 0
(N )
(N −1)
L=
:
L
∈
Mat(m|2n)(Λ
)
∪
L
=
:
L
∈
Mat(m|2n)(Λ
)
;
N
N
0 D
C` 0
10
• if N = 2`, W2` is generated by the set of supermatrices
0 0
(N )
L=
: L ∈ Mat(m|2n)(ΛN ) .
0 D
Proof.
We can easily check by straightforward computations that the two subspaces defined above are two-sided
ideals of Mat(m|2n)(ΛN ) i.e., M WN ⊂ WN and WN M ⊂ WN for every M ∈ Mat(m|2n)(ΛN ). Let
A B`
L=
C` D
2n×2n
, B` = {b`j,k } ∈ Λm×2n
, C` = {c`j,k } ∈ Λ2n×m
and D = {dj,k } ∈ ΛN,0
.
with A = {aj,k } ∈ Λm×m
N,1
N,1
N,0
Choosing for x, y all possible coordinate supervectors we obtain that L satisfies (11) if and only if its
entries are such that
xj aj,k yk = 0
∀xj , yk ∈ ΛN,0 ,
xj b`j,k y`k = 0
∀xj ∈ ΛN,0 , y`k ∈ ΛN,1 ,
x`j c`j,k yk = 0
∀x`j ∈ ΛN,1 , yk ∈ ΛN,0 ,
x`j dj,k y`k = 0
∀x`j , y`k ∈ ΛN,1 .
Taking in the first equation xj = yk = 1 we obtain aj,k = 0 for every j, k = 1, . . . , m. For the remaining
coefficients we need to distinguish between two cases.
(N )
Case: N = 2` + 1. We first prove that b`j,k , c`j,k ∈ spanR {f1 · · · f2`+1 }. If b`j,k ∈ ΛN , it is clear that
b`j,k y`k = 0, for every y`k ∈ ΛN,1 . It then easily follows that xj b`j,k y`k = 0 for every pair xj ∈ ΛN,0 , y`k ∈
(N )
ΛN,1 . If on the contrary, b`j,k ∈
/ ΛN , there would be at least one of the generators fr missing in one of
the canonical terms of b`j,k . Then, taking xj = 1 and y`k = fr we will have xj b`j,k y`k 6= 0. Hence,
xj b`j,k y`k = 0
∀xj ∈ ΛN,0 , y`k ∈ ΛN,1
if and only if b`j,k ∈ spanR {f1 · · · f2`+1 }.
The same holds for c`j,k . Next we prove that dj,k ∈ spanR {f1 · · · fr−1 fr+1 · · · f2`+1 : 1 ≤ r ≤ 2` + 1}. In
(N −1)
fact, if dj,k ∈ ΛN
it is clear that x`j dj,k y`k = 0, since x`j dj,k y`k is the sum of homogeneous terms of
(N −1)
degree at least N + 1. If on the contrary, dj,k ∈
/ ΛN
, there would be at least two different Grassmann
generators fr , fs missing in one of the canonical terms of dj,k . Then, taking x`j = fr and y`k = fs we have
that x`j dj,k y`k 6= 0. Hence,
x`j dj,k y`k = 0
∀x`j , y`k ∈ ΛN,1
if and only if dj,k ∈ spanR {f1 · · · fr−1 fr+1 · · · f2`+1 : 1 ≤ r ≤ 2` + 1}.
Case: N = 2`. We first prove that b`j,k = c`j,k = 0. If b`j,k 6= 0, there will be always at least one
Grassmann generator fr missing in one of the terms of b`j,k since N is an even number. Then, taking
xj = 1 and y`k = fr we will have xj b`j,k y`k 6= 0. Hence,
xj b`j,k y`k = 0
∀xj ∈ ΛN,0 , y`k ∈ ΛN,1
if and only if b`j,k = 0.
(N )
The same holds for c`j,k . We now prove that dj,k ∈ spanR {f1 · · · f2` }. Indeed, if dj,k ∈ ΛN it is clear that
(N )
x`j dj,k y`k = 0. If on the contrary, dj,k ∈
/ ΛN , there would be at least two different Grassmann generators
fr , fs missing in one of the canonical terms of dj,k . Then, taking x`j = fr and y`k = fs we have that
x`j dj,k y`k 6= 0. Hence,
x`j dj,k y`k = 0
∀x`j , y`k ∈ ΛN,1
if and only if dj,k ∈ spanR {f1 · · · f2` }.
As a direct consequence of Lemma 4 we have the following result.
Corollary 1. The set O(m|2n)(ΛN ) of supermatrices in Mat(m|2n)(ΛN ) leaving the inner product h·, ·i
invariant is characterized by
O(m|2n)(ΛN ) = {M ∈ Mat(m|2n)(ΛN ) : M ST QM − Q ∈ WN }.
11
Remark 4.1. The form L = M ST QM − Q suggests that we do not need the whole ideal WN to describe
O(m|2n)(ΛN ). In fact, supermatrices of the previous form satisfy
LST = M ST QST M ST
ST
− QST = M ST QS (SM S) − QS = M ST QM − Q S = LS,
whence the subspace WN∗ = {L ∈ WN : LST = LS} can be considered in the above definition. However
the use of WN∗ is not important for the purposes of this paper, whence we will continue working with WN .
We will now study the algebraic structure of O(m|2n)(ΛN ).
Theorem 2. The following statements hold:
(i) O(m|2n)(ΛN ) ⊂ GL(m|2n)(ΛN ).
(ii) O(m|2n)(ΛN ) is a group under the usual matrix multiplication.
(iii) O(m|2n)(ΛN ) is a closed subgroup of GL(m|2n)(ΛN ).
Proof.
(i) To prove that every supermatrix in O(m|2n)(ΛN ) is invertible, first note that the real projection
of a supermatrix in WN (N > 1) is zero. Then for every M ∈ O(m|2n)(ΛN ) we have
ST
A0 0
T
M
Q[M
]
−
Q
=
[M
]
Q[M
]
−
Q
=
0,
where
[M
]
=
.
0
0
0
0
0
0 D0
This can be rewritten in terms of the real blocks A0 and D0 as AT0 A0 = Im and D0T J2n D0 = J2n ,
implying that A0 ∈ O(m) and D0 ∈ Sp(2n). On account of Theorem 1, M thus is invertible.
(ii) Here we only need to prove that matrix inversion and matrix multiplication are internal operations in O(m|2n)(ΛN ). For the inversion, the condition M ST QM − Q = L ∈ WN implies that
ST
ST
M −1
QM −1 − Q = − M −1
LM −1 ∈ WN since WN is an ideal. Hence, M ∈ O(m|2n)(ΛN )
−1
implies M
∈ O(m|2n)(ΛN ). For the multiplication, let M1 , M2 ∈ O(m|2n)(ΛN ) and L1 , L2 ∈
WN such that MjST QMj − Q = Lj , j = 1, 2. Then:
(M1 M2 )
ST
Q (M1 M2 ) − Q = M2ST M1ST QM1 M2 − Q = M2ST (Q + L1 ) M2 − Q
= L2 + M2ST L1 M2 .
Since WN is an ideal, L2 + M2ST L1 M2 ∈ WN . Hence, M1 M2 ∈ O(m|2n)(ΛN ).
(iii) Let {Mj }j∈N ⊂ O(m|2n)(ΛN ) be a sequence that converges to a supermatrix M ∈ Mat(m|2n)(ΛN ).
Since algebraic operations are continuous in Mat(m|2n)(ΛN ) we have that
lim MjST QMj − Q = M ST QM − Q.
j→∞
But MjST QMj − Q ∈ WN for every j ∈ N and, since WN is a finite dimensional subspace of
Mat(m|2n)(ΛN ), it is closed. Whence the limit M ST QM − Q belongs to WN .
Remark 4.2. The above theorem states that O(m|2n)(ΛN ) is a Lie group.
The group O(m|2n)(ΛN ) can be partitioned in a natural way into the classes
OL = OL (m|2n)(ΛN ) := {M ∈ O(m|2n)(ΛN ) : M ST QM − Q = L},
L ∈ WN .
Not for every L ∈ WN the set OL is non empty. In particular, OL = ∅ for every L ∈ WN \ WN∗ .
Proposition 10. The following statements hold:
(i) Let L1 , L2 ∈ WN and M1 ∈ OL1 , M2 ∈ OL2 . Then
M1 M2 ∈ OM2ST L1 M2 +L2
and
M1−1 ∈ O−(M −1 )ST L
1
12
−1
1 M1
.
(ii) OL is a subgroup of O(m|2n)(ΛN ) if and only if L = 0.
(iii) The binary relation:
R = {(M1 , M2 ) ∈ O(m|2n)(ΛN )2 : M1 , M2 ∈ OL ; L ∈ WN }
= {(M1 , M2 ) ∈ O(m|2n)(ΛN )2 : M1 M2−1 ∈ O0 }.
is an equivalence relation.
Proof.
(i) This was already proven in the proof of Theorem 2 (ii).
(ii) If OL is a subgroup then Im+2n ∈ OL and L = Im+2n QIm+2n − Q = 0. To prove now that O0
is a subgroup we just have to consider L1 = L2 = 0 in (i). Then M1 , M2 ∈ O0 directly implies
M1 M2 ∈ O0 and M1−1 ∈ O0 .
ST
(iii) Let M1 ∈ OL1 and M2 ∈ OL2 . From (i) we get that M1 M2−1 ∈ OL where L = M2−1
(L1 − L2 ) M2−1 .
−1
Thence, M1 , M2 ∈ OL for some L ∈ WN if and only if M1 M2 ∈ O0 . Since O0 is a subgroup of
O(m|2n)(ΛN ) it now easily follows that R is an equivalence relation.
Remark 4.3. The subgroup O0 is a closed subgroup of GL(m|2n)(ΛN ), whence it is a Lie group. It
plays a crucial rôle in the study of O(m|2n)(ΛN ) since given any representative element M of OL we can
describe the whole set OL by means of the relation OL = O0 M . For that reason, we will, from now on,
focus our attention to O0 .
Proposition 11. The following statements hold:
(i) A supermatrix M ∈ Mat(m|2n)(ΛN ) belongs to O0 if and only if
1 `T
`
T
A A − 2 C J2n C = Im ,
1 `T
T `
A B − 2 C J2n D = 0,
`T ` 1 T
B B + 2 D J2n D = 12 J2n .
(12)
(ii) sdet(M ) = ±1 for every M ∈ O0 .
(iii) [O(m|2n)(ΛN )]0 = [O0 ]0 = O(m) × Sp(2n).
Proof.
i) The relation M ST QM = Q can be written in terms of A, B`, C`, D as:
Im
AT A − 12 C`T J2n C`
AT B` − 21 C`T J2n D
=
0
−B`T A − 12 DT J2n C` −B`T B` − 21 DT J2n D
0
− 12 J2n
.
(ii) On account of Proposition 4, M ST QM = Q implies that sdet(M )2 sdet(Q) = sdet(Q), whence
sdet(M )2 = 1. The statement then follows from Lemma 1.
(iii) See the proof of Theorem 2 i).
4.2
Group of superrotations SO0 .
As in the classical way, we can introduce now the set of superrotations by
SO0 = SO0 (m|2n)(ΛN ) = {M ∈ O0 : sdet(M ) = 1}.
This is easily seen to be a Lie subgroup of O0 with real projection equal to SO(m) × Sp(2n). In fact,
the conditions M ST QM = Q and sdet(M ) = 1 imply that M0T QM0 = Q and sdet(M0 ) = 1. This means
that
A0 0
M0 =
0 D0
with AT0 A0 = Im , D0T J2n D0 = J2n and det(A0 ) = det(D0 ). But D0 ∈ Sp(2n) implies det(D0 ) = 1. Then
det(A0 ) = 1 and A0 ∈ SO(m).
The following proposition states that, as in the classical case, SO0 is connected and in consequence,
it is the identity component of O0 .
13
Proposition 12. SO0 is a connected Lie group.
Proof.
Since the real projection SO(m) × Sp(2n) of SO0 is a connected group, it suffices to prove that for
every M ∈ SO0 there exist a continuous path inside SO0 connecting M with its real projection M0 . To
PN
(j)
that end, let us write M = j=0 [M ]j , where [M ]j is the projection of M on Mat(m|2n)(ΛN ) for each
j = 0, 1, . . . , N . Then, observe that
N
N
N
k
X
X
X
X
M ST QM = Q ⇐⇒
M ST j Q [M ]j = Q ⇐⇒
M ST j Q[M ]k−j = Q
j=0
j=0
⇐⇒ M0T QM0 = Q,
and
j=0
k=0
k
X
ST
M
Q[M ]k−j = 0,
j
k = 1, . . . , N.
j=0
PN
Let us now take the path M (t) = j=0 tj [M ]j . For t ∈ [0, 1] this is a continuous path with M (0) = M0
and M (1) = M . In addition, M (t)T0 QM (t)0 = M0T QM0 = Q and for every k = 1, . . . , N we have,
k
k
X
X
ST
Q[M ]k−j = 0.
M (t)ST j Q[M (t)]k−j = tk
M
j
j=0
j=0
Hence, M (t)ST QM (t) = Q, t ∈ [0, 1]. Finally, observe that sdet(M (t)) = 1 for every t ∈ [0, 1], since
sdet(M (t)0 ) = sdet(M0 ) = 1.
We will now investigate the corresponding Lie algebras of O(m|2n)(ΛN ), O0 and SO0 .
Theorem 3.
(i) The Lie algebra so(m|2n)(ΛN ) of O(m|2n)(ΛN ) is given by
so(m|2n)(ΛN ) = {X ∈ Mat(m|2n)(ΛN ) : X ST Q + QX ∈ WN }.
(ii) The Lie algebra so0 = so0 (m|2n)(ΛN ) of O0 coincides with the Lie algebra of SO0 and is given by
the space of all "super anti-symmetric" supermatrices
so0 = {X ∈ Mat(m|2n)(ΛN ) : X ST Q + QX = 0}.
A B`
(iii) A supermatrix X =
∈ Mat(m|2n)(ΛN ) belongs to so0 if and only if
C` D
AT + A = 0,
1
B` − C`T J2n = 0
2
and DT J2n + J2n D = 0.
(iv) [so(m|2n)(ΛN )]0 = [so0 ]0 = so(m) ⊕ sp(2n).
Proof.
(i) If X ∈ Mat(p|q)(ΛN ) is in the Lie algebra of O(m|2n)(ΛN ) then etX ∈ O(m|2n)(ΛN ) for every
ST
t ∈ R, i.e. etX QetX − Q = L(t) ∈ WN . Differentiating at t = 0 we obtain X ST Q + QX = L0 (0),
0
where L (0) ∈ WN since WN is a closed subspace. On the other hand, if X ∈ Mat(p|q)(ΛN ) satisfies
X ST Q + QX = L ∈ WN , then X ST = L1 − QXQ−1 where L1 = LQ−1 ∈ WN . Computing the
exponential of tX ST we obtain
j
j
∞
∞
X
X
Q(−tX)Q−1 + tL1
Q(−tX)Q−1
ST
=
+ L2 (t) = Qe−tX Q−1 + L2 (t),
etX =
j!
j!
j=0
j=0
where L2 (t) is a convergent infinite sum of products that contain the factor tL1 at least one time.
Then, since WN is a two-sided ideal, L2 (t) ∈ WN for every t ∈ R, whence
etX
ST
QetX − Q = L2 (t)QetX ∈ WN
14
∀t ∈ R.
(ii) To prove that so0 is the Lie algebra of O0 it suffices to repeat the above reasoning with L = L1 =
L2 = 0. From Proposition 6 it easily follows that the Lie algebra of SO0 is
{X ∈ Mat(p|q)(ΛN ) : X ST Q + QX = 0, str(X) = 0}.
But X ST Q + QX = 0 implies str(X) = 0. In fact, the condition X ST = −QXQ−1 implies that
str(X ST ) = − str(QXQ−1 ) = − str(X), yielding str(X) = str(X ST ) = − str(X) and str(X) = 0.
Hence, the Lie algebra of SO0 is so0 .
(iii) Observe that the relation X ST Q + QX = 0 can be written in terms of A, B`, C`, D as follows:
AT + A
− 12 C`T J2n + B`
= 0.
−B`T − 12 J2n C` − 12 DT J2n − 12 J2n D
(iv) As the real
we have [so(m|2n)(ΛN )]0 = [so0 ]0 .
projection
of every element in WN (N > 1) is zero,
A B`
A0 0
Let X =
∈ so0 , then X0 = [X]0 =
satisfies X0ST Q+QX0 = 0. Using (iii)
0 D0
C` D
we obtain AT0 + A0 = 0 and D0T J2n + J2n D0 = 0 which implies that A0 ∈ so(m) and D0 ∈ sp(2n).
Remark 4.4. As in the Remark 4.1, the supermatrices of the form L = X ST Q + QX satisfy LST =
QXS + X ST QS = LS. Then, the subspace WN∗ can replace WN in the above definition of so(m|2n)(ΛN ).
The connectedness of SO0 allows to write any of its elements as a finite product of exponentials of
supermatrices in so0 , see [8]. In the classical case, a single exponential suffices for such a description since
SO(m) is compact and in consequence exp : so(m) → SO(m) is surjective. This property, however, does
not hold in the group of superrotations SO0 , since the exponential map from sp(2n) to the non-compact
Lie group Sp(2n) ∼
= {Im } × Sp(2n) ⊂ SO0 is not surjective, whence not every element in SO0 can be
written as a single exponential of a supermatrix in so0 . Nevertheless, it is possible to find a decomposition
for elements of SO0 in terms of a fixed number of exponentials of so0 elements.
Every supermatrix M ∈ SO0 has a unique decomposition M = M0 + M = M0 (Im+2n + L) where
−1
M0 is its real projection, M ∈ Mat(m|2n)(Λ+
N ) its nilpotent projection and L = M0 M. We will now
separately study the decompositions for M0 ∈ SO(m) × Sp(2n) and Im+2n + L ∈ SO0 .
First consider M0 ∈ SO(m)×Sp(2n). We already mentioned that exp : so(m) → SO(m) is surjective,
while exp : sp(2n) → Sp(2n) is not. However, it can be proved that Sp(2n) = exp(sp(2n)) · exp(sp(2n)),
invoking the following polar decomposition for real algebraic Lie groups, see Proposition 4.3.3 in [9].
Proposition 13. Let G ⊂ GL(p) be an algebraic Lie group such that G = GT and g its Lie algebra.
Then every A ∈ G can be uniquely written as A = ReX , R ∈ G ∩ O(p), X ∈ g ∩ Sym(p), where Sym(p)
is the subspace of all symmetric matrices in Rp×p .
Taking p = 2n and G = Sp(2n) in the above proposition we get that every symplectic matrix D0 can
be uniquely written as D0 = R0 eZ0 with R0 ∈ Sp(2n) ∩ O(2n) and Z0 ∈ sp(2n) ∩ Sym(2n). But the group
Sp(2n) ∩ O(2n) is isomorphic to U (n) which is connected and compact. Then the exponential map from
the Lie algebra sp(2n) ∩ so(2n) ∼
= u(n) is surjective on Sp(2n) ∩ O(2n). This means that D0 ∈ Sp(2n) can
be written as D0 = eY0 eZ0 with Y0 ∈ sp(2n) ∩ so(2n) and Z0 ∈ sp(2n) ∩ Sym(2n). Hence, the supermatrix
M0 ∈ SO(m) × Sp(2n) can be decomposed as
X
X
Im
0
e 0
0
e 0
0
M0 =
=
= eX eY ,
0 eZ0
0
e Y0 e Z 0
0
e Y0
X0 0
0 0
where X =
∈ so(m) × [sp(2n) ∩ so(2n)] and Y =
∈ {0m } × [sp(2n) ∩ Sym(2n)].
0 Y0
0 Z0
Now consider the element Im+2n + L ∈ SO0 . As shown at the end of Section 2, the function
+
exp : Mat(m|2n)(Λ+
N ) → Im+2n + Mat(m|2n)(ΛN ) is a bijection with the logarithmic function defined in
(8) as its inverse. Then the supermatrix Z = ln(Im+2n + L) satisfies eZ = Im+2n + L and is nilpotent.
Those properties suffice for proving that Z ∈ so0 . From now on we will denote the set so0 ∩Mat(m|2n)(Λ+
N)
by so0 (m|2n)(Λ+
).
N
15
Z
Proposition 14. Let Z ∈ Mat(m|2n)(Λ+
N ) such that e ∈ SO0 . Then Z ∈ so0 .
Proof.
ST
It suffices to prove that etZ ∈ SO0 for every t ∈ R. The expression etZ QetZ − Q can be written as the
following polynomial in the real variable t.
"
#
N
N
k k
j
ST
j
X
X
t
Z
t
(Z
)
tZST
tZ
Q
−Q
P (t) = e
Qe − Q =
j!
k!
j=0
k=0
N X
k
k
N
N
j
ST j
k−j k−j
k X
X
X
X
t (Z )
t Z
k
tk
t
=
Q
=
(ZST )j QZk−j =
Pk (Z),
j!
(k − j)!
k! j=0 j
k!
j=0
k=1
k=1
k=1
Pk
where Pk (Z) = j=0 kj (ZST )j QZk−j . If P (t) is not identically zero, i.e. not all the Pk (Z) are 0, we can
take k0 ∈ {1, 2, . . . , N } to be the largest subindex for which Pk0 (Z) 6= 0. Then,
lim
1
t→∞ tk0
P (t) =
Pk0 (Z)
6= 0,
k0 !
contradicting that P (Z) = {0}. So P (t) identically vanishes, yielding etZ ∈ SO0 for every t ∈ R.
This way, we have proven the following result.
Theorem 4. Every supermatrix in SO0 can be written as M = eX eY eZ with X ∈ so(m) × [sp(2n) ∩
so(2n)], Y ∈ {0m } × [sp(2n) ∩ Sym(2n)] and Z ∈ so0 (m|2n)(Λ+
N ). Moreover, the elements Y and Z are
unique.
4.3
Relation with superbivectors.
Theorem 3 allows to compute the dimension of so0 as a real vector space.
+
2mn
+
n(2n
+
1)
.
Corollary 2. The dimension of the real Lie algebra so0 is 2N −1 m(m−1)
2
Proof.
Since so0 is the direct sum of the corresponding subspaces of block components A, B`, C` and D respectively,
it suffices to compute the dimension of each one of them. According to Theorem 3 (iii) we have:
A
0
0
C`
0
0
V1 =
V2 =
V3 =
0
0
m×m
: AT = −A, A ∈ ΛN,0
1 `T
C J2n
2
∼
= ΛN,0 ⊗ so(m),
2n×m
∼
,
= ΛN,1 ⊗ R
0
0
∼
: DT J2n + J2n D = 0, D ∈ Λ2n×2n
= ΛN,0 ⊗ sp(2n),
N,0
D
2n×m
: C` ∈ ΛN,1
then dim V1 = 2N −1
m(m − 1)
,
2
then dim V2 = 2N −1 m2n,
then dim V3 = 2N −1 n(2n + 1).
(2)E
Comparing this result with the one in Remark 3.1 we obtain that dim Rm|2n (ΛN ) = dim so0 . This
means that both vector spaces are isomorphic. This isomorphism also holds on the Lie algebra level.
Following the classical Clifford approach, the commutator
[B, x]
(2)E
∀ B ∈ Rm|2n (ΛN ), x ∈ Rm|2n (ΛN ),
(13)
(2)E
should be the key for the Lie algebra isomorphism. Proposition 9 shows that for every B ∈ Rm|2n (ΛN )
the commutator (13) defines an endomorphism over Rm|2n (ΛN ) that can be represented by a supermatrix
in Mat(m|2n)(ΛN ). But as it was explained in Section 2, that supermatrix is not unique. This issue can
be solved with the natural extension of the linear operator defined in (13) to Rm|2n (ΛN +1 ).
16
(2)E
Lemma 5. The map φ : Rm|2n (ΛN ) → Mat(m|2n)(ΛN ) defined by
φ(B)x = [B, x]
(2)E
∀ B ∈ Rm|2n (ΛN ), x ∈ Rm|2n (ΛN +1 ),
(14)
takes values in so0 . In particular, if we consider {b1 , . . . , b2N −1 } and {b`1 , . . . , b`2N −1 } to be the canonical
basis of ΛN,0 and ΛN,1 respectively, we obtain the following basis for so0 .
Ek,j − Ej,k 0
,
φ(br ej ek ) = 2br
0
0
0
Ej,2k
,
φ(b`r ej e`2k−1 ) = b`r
2E2k−1,j
0
0
−Ej,2k−1
φ(b`r ej e`2k ) = b`r
,
2E2k,j
0
0
0
φ(br e`2j e`2k ) = −br
,
0 E2j,2k−1 + E2k,2j−1
0
0
φ(br e`2j−1 e`2k−1 ) = br
,
0 E2j−1,2k + E2k−1,2j
0
0
,
φ(br e`2j−1 e`2k ) = br
0 E2k,2j − E2j−1,2k−1
0
0
,
φ(br e`2j e`2k−1 ) = br
0 E2j,2k − E2k−1,2j−1
1 ≤ r ≤ 2N −1 , 1 ≤ j < k ≤ m,
1 ≤ r ≤ 2N −1 , 1 ≤ j ≤ m, 1 ≤ k ≤ n,
1 ≤ r ≤ 2N −1 , 1 ≤ j ≤ m, 1 ≤ k ≤ n,
1 ≤ r ≤ 2N −1 , 1 ≤ j ≤ k ≤ n,
1 ≤ r ≤ 2N −1 , 1 ≤ j ≤ k ≤ n,
1 ≤ r ≤ 2N −1 , 1 ≤ j ≤ k ≤ n,
1 ≤ r ≤ 2N −1 , 1 ≤ j < k ≤ n,
where Ej,k denotes the matrix with all entries equal 0, except the one in the j-th row and k-th column
which is equal to 1, and the order of Ej,k should be deduced from the context.
Proof.
The above equalities can be directly obtained from Proposition 9, whence we should only check that all
supermatrices obtained above form a basis for so0 . The matrices Ej,k satisfy the relations
T
Ej,k
= Ek,j , Ej,2k−1 J2n = Ej,2k , Ej,2k J2n = −Ej,2k−1 , J2n E2j,k = E2j−1,k , J2n E2j−1,k = −E2j,k .
Then
• for φ(br ej ek ) we have A = 2br (Ek,j − Ej,k ), B` = 0, C` = 0 and D = 0, whence
AT = 2br (Ej,k − Ek,j ) = −A;
• for φ(b`r ej e`2k−1 ) we have A = 0, B` = b`r Ej,2k , C` = 2b`r E2k−1,j and D = 0, whence
1 `T
C J2n = b`r Ej,2k−1 J2n = b`r Ej,2k = B`;
2
• for φ(b`r ej e`2k ) we have A = 0, B` = −b`r Ej,2k−1 , C` = 2b`r E2k,j and D = 0, whence
1 `T
C J2n = b`r Ej,2k J2n = −b`r Ej,2k−1 = B`;
2
• for φ(br e`2j
e`2k ) we have A = 0, B` = 0, C` = 0 and D = −br (E2j,2k−1 + E2k,2j−1 ), whence
DT J2n + J2n D = −br (E2k−1,2j J2n + E2j−1,2k J2n + J2n E2j,2k−1 + J2n E2k,2j−1 )
= −br (−E2k−1,2j−1 − E2j−1,2k−1 + E2j−1,2k−1 + E2k−1,2j−1 ) = 0;
• for φ(br e`2j−1
e`2k−1 ) we have A = 0, B` = 0, C` = 0 and D = br (E2j−1,2k + E2k−1,2j ), whence
DT J2n + J2n D = br (E2k,2j−1 J2n + E2j,2k−1 J2n + J2n E2j−1,2k + J2n E2k−1,2j )
= br (E2k,2j + E2j,2k − E2j,2k − E2k,2j ) = 0;
17
• for φ(br e`2j−1
e`2k ) we have A = 0, B` = 0, C` = 0 and D = br (E2k,2j − E2j−1,2k−1 ), whence
DT J2n + J2n D = br (E2j,2k J2n − E2k−1,2j−1 J2n + J2n E2k,2j − J2n E2j−1,2k−1 )
= br (−E2j,2k−1 − E2k−1,2j + E2k−1,2j + E2j,2k−1 ) = 0.
The above computations show that all supermatrices obtained belong to so0 . Direct verification shows
that they form a set of 2N −1 m(m−1)
+ 2N −1 2mn + 2N −1 n(2n + 1) linear independent elements, i.e. a
2
basis of so0 .
(2)E
Theorem 5. The map φ : Rm|2n (ΛN ) → so0 defined in (14) is a Lie algebra isomorphism.
Proof.
From Lemma 5 follows that φ is a vector space isomorphism. In addition, due to the Jacoby identity in
(2)E
the associative algebra Am,2n (ΛN ) we have for every B1 , B2 ∈ Rm|2n (ΛN ) and x ∈ Rm|2n (ΛN +1 ) that
[φ(B1 ), φ(B2 )] x = φ(B1 )φ(B2 )x − φ(B2 )φ(B1 )x
= [B1 , [B2 , x]] + [B2 , [x, B1 ]] = [[B1 , B2 ] , x] = φ ([B1 , B2 ]) x.
implying, by Lemma 3, that [φ(B1 ), φ(B2 )] = φ ([B1 , B2 ]), i.e., φ is a Lie algebra isomorphism.
5
The Spin group in Superspace
So far we have seen that the Lie algebra so0 of the Lie group of superrotations SO0 has a realization
in Am,2n (ΛN ) as the Lie algebra of extended superbivectors. In this section, we discuss the proper way
of defining the corresponding realization of SO0 in T (V )/I, i.e., the analogue of the Spin group in the
Clifford superspace framework.
5.1
Supervector reflections
The group generated by the supervector reflections was briefly introduced in [14] using the notion of
the unit super-sphere defined as the super-surface S(m|2n)(ΛN ) = {w ∈ Rm|2n (ΛN ) : w2 = −1}. The
reflection associated to the supervector w ∈ S(m|2n)(ΛN ) is defined by ψ(w)[x] = wxw, x ∈ Rm|2n (ΛN ).
It is known from the radial algebra setting that ψ(w) maps vectors into vectors. Indeed,
wxw = {x, w}w − w2 x = {x, w}w + x = x − 2hx, wiw.
Every supervector reflection can be identified with a unique matrix in O(m|2n)(ΛN ). On account of
the results in Section 2, an option for doing this is extending the operator ψ(w) to Rm|2n (ΛN +1 ) in the
natural way, i.e.
ψ(w)[x] = wxw
x ∈ Rm|2n (ΛN +1 ).
(15)
Pm
P2n
Lemma 6. Let w = w + w`= j=1 wj ej + j=1 w`j e`j ∈ S(m|2n)(ΛN ). Then, the endomorphism (15)
can be represented by an unique supermatrix
A(w) B`(w)
ψ(w) =
∈ O(m|2n)(ΛN )
C`(w) D(w)
with A(w) = −2Dw Em×m Dw + Im , B`(w) = Dw Em×2n Dw`J2n , C`(w) = −2Dw`E2n×m Dw , and finally
D(w) = Dw`E2n×2n Dw`J2n + I2n , where
w1
w`1
1 1 ... 1
. .
.
p×q
..
..
Dw =
, Dw` =
and Ep×q = .. .. . . . .. ∈ R .
.
.
wm
w`2n
18
1
1
...
1
Proof.
Pm
P2n
Observe that ψ(w)[x] = wxw = {x, w}w + x = k=1 yk ek + k=1 y`j e`k , where
m
n
X
X
yk = −2
wj wk xj + xk +
w`2j−1 wk x`2j − w`2j wk x`2j−1 ,
j=1
j=1
m
n
X
X
y`k = −2
wj w`k xj + x`k +
−w`2j−1 w`k x`2j + w`2j w`k x`2j−1 .
j=1
Then, ψ(w)x =
A(w) B`(w)
C`(w) D(w)
A(w) = −2
w12
w1 w2
..
.
w1 wm
w2 w1
w22
..
.
w2 wm
j=1
x
x`
...
...
..
.
...
where,
wm w1
wm w2
..
.
2
wm
+ Im = −2Dw Em×m Dw + Im ,
−w`2 w1 w`1 w1 . . . −w`2n w1
−w`2 w2 w`1 w2 . . . −w`2n w2
B`(w) =
..
..
..
..
.
.
.
.
−w`2 wm w`1 wm . . . −w`2n wm
w1 w`1
w2 w`1 . . . wm w`1
w1 w`2
w2 w`2 . . . wm w`2
C`(w) = −2
..
..
..
..
.
.
.
.
w1 w`2n w2 w`2n . . . wm w`2n
w`2 w`1
−w`1 w`1 . . . w`2n w`1
w`2 w`2
−w`1 w`2 . . . w`2n w`2
D(w) =
..
..
..
..
.
.
.
.
w`2 w`2n −w`1 w`2n . . . w`2n w`2n
w`2n−1 w1
w`2n−1 w2
..
.
w`2n−1 wm
= Dw Em×2n Dw`J2n ,
= −2Dw`E2n×m Dw ,
−w`2n−1 w`1
−w`2n−1 w`2
..
.
−w`2n−1 w`2n
+ I2n = Dw`E2n×2n Dw`J2n + I2n .
The uniqueness of such a supermatrix in Mat(m|2n)(ΛN ) is guaranteed by Lemma 3 and it is easily seen
that ψ(w) belongs to O(m|2n)(ΛN ) since
1
1
1
hψ(w)[x], ψ(w)[y]i = − {wxw, wyw} = w{x, y}w = − {x, y} = hx, yi.
2
2
2
Algebraic operations with the matrices A(w), B`(w), C`(w), D(w) are easy since
m
n
X
X
2
Ep×m Dw
Em×q =
wj2 Ep×q ,
Ep×2n Dw`J2n Dw`E2n×q = 2
w`2j−1 w`2j Ep×q .
j=1
j=1
We can now define the bosonic Pin group in superspace as
Pinb (m|2n)(ΛN ) = {w1 · · · wk : wj ∈ S(m|2n)(ΛN ), k ∈ N},
and extend the map ψ to a Lie group homomorphism ψ : Pinb (m|2n)(ΛN ) → O(m|2n)(ΛN ) by
ψ(w1 · · · wk )[x] = w1 · · · wk x wk · · · w1 = ψ(w1 ) ◦ · · · ◦ ψ(wk )[x].
Proposition 15. Let w ∈ S(m|2n)(ΛN ). Then ψ(w) ∈ O0 and sdet (ψ(w)) = −1.
Proof.
19
(16)
`
To prove that ψ(w) ∈ O0 it suffices to prove thatPA(w), B`(w),
PnC (w), D(w) satisfy (12). This can be easily
m
2
2
done using (16) and the identity −1 = w = − j=1 wj + j=1 w`2j−1 w`2j . In fact, we have
2
Em×m Dw − 4Dw Em×m Dw + Im
A(w)T A(w) = 4Dw Em×m Dw
m
X
= 4
wj2 Dw Em×m Dw − 4Dw Em×m Dw + Im ,
j=1
n
X
C`(w)T J2n C`(w) = 4Dw Em×2n Dw`J2n Dw`E2n×m Dw = 8
w`2j−1 w`2j Dw Em×m Dw .
j=1
Then, A(w)T A(w) − 21 C`(w)T J2n C`(w) = 4
Also,
hP
m
2
j=1 wj − 1 −
i
w
`
w
`
j=1 2j−1 2j Dw Em×m Dw + Im = Im .
Pn
2
Em×2n Dw`J2n + Dw Em×2n Dw`J2n
A(w)T B`(w) = −2Dw Em×m Dw
m
X
= −2
wj2 Dw Em×2n Dw`J2n + Dw Em×2n Dw`J2n ,
j=1
C`(w)T J2n D(w) = −2Dw Em×2n Dw`J2n Dw`E2n×2n Dw`J2n − 2Dw Em×2n Dw`J2n
n
X
= −4
w`2j−1 w`2j Dw Em×2n Dw`J2n − 2Dw Em×2n Dw`J2n .
j=1
h P
i
Pn
m
Hence, A(w)T B`(w)− 12 C`(w)T J2n D(w) = 2 − j=1 wj2 + 1 + j=1 w`2j−1 w`2j Dw Em×2n Dw`J2n = 0. In
the same way we have
m
X
2
Em×2n Dw`J2n = −
wj2 J2n Dw`E2n×2n Dw`J2n ,
B`(w)T B`(w) = −J2n Dw`E2n×m Dw
j=1
T
D(w) J2n D(w) = J2n Dw`E2n×2n Dw`J2n Dw`E2n×2n Dw`J2n + 2J2n Dw`E2n×2n Dw`J2n + J2n
n
X
= 2
w`2j−1 w`2j + 1 J2n Dw`E2n×2n Dw`J2n + J2n ,
j=1
whence
m
n
X
X
1
1
1
B`(w)T B`(w) + DT (w)J2n D(w) = −
wj2 + 1 +
w`2j−1 w`2j J2n Dw`E2n×2n Dw`J2n + J2n = J2n .
2
2
2
j=1
j=1
Then, A(w), B`(w), C`(w), D(w) satisfy (12) and in consequence, ψ(w) ∈ O0 . To prove that sdet(ψ(w)) =
−1, first observe that ψ(w) = ψ(w)−1 since ψ(w) ◦ ψ(w)[x] = wwxww = x. Hence, due to Theorem 1 we
−1
obtain A(w) = A(w) − B`(w)D(w)−1 C`(w)
, yielding
det A(w) − B`(w)D(w)−1 C`(w)
1
=
.
sdet(ψ(w)) =
det[D(w)]
det[A(w)] det[D(w)]
We will compute det[D(w)] using the formula det[D(w)] = exp(tr ln D(w)) and the fact that D(w) − I2n
is a nilpotent matrix. Observe that
∞
∞
j
X
X
Dw`E2n×2n Dw`J2n
j+1 (D(w) − I2n )
ln D(w) =
(−1)
=
(−1)j+1
j
j
j=1
j=1
20
j
.
It follows from (16) that
Dw`E2n×2n Dw`J2n
j
j−1
n
X
= 2j−1
w`2j−1 w`2j
Dw`E2n×2n Dw`J2n .
j=1
Then,
P
j−1
n
j−1
∞
2
w
`
w
`
j=1 2j−1 2j
X
ln D(w) = (−1)j+1
Dw`E2n×2n Dw`J2n
j
j=1
and in consequence,
tr ln D(w) = −
∞
X
(−1)j+1
2j
P
n
j=1
w`2j−1 w`2j
j
= − ln 1 + 2
j
j=1
n
X
w`2j−1 w`2j .
j=1
Hence
1
det(D(w)) =
1+2
n
P
w`2j−1 w`2j
j=1
Similar computations yield det(A(w)) = 1 − 2
m
P
j=1
wj2 which shows that sdet(ψ(w)) = −1.
The above proposition states that the Lie group homomorphism ψ takes values in O0 , and its restriction to the bosonic spin group, defined as
Spinb (m|2n)(ΛN ) = {w1 · · · w2k : wj ∈ S(m|2n)(ΛN ), k ∈ N},
takes values in the subgroup SO0 .
In the classical case, the Pin and the Spin groups are a double covering of the groups O(m) and
SO(m) respectively. A natural question in this setting is whether Pinb (m|2n)(ΛN ) and Spinb (m|2n)(ΛN )
cover the groups O0 and SO0 . The answer to this question is negative and the main reason is that the
real projection of every vector w ∈ S(m|2n)(ΛN ) is in the unitary sphere Sm−1 of Rm , i.e.,
[w]0 =
m
X
[wj ]0 ej
and
[w]20 = −1.
j=1
Then, the real projection of ψ (Pinb (m|2n)(ΛN )) is just O(m), while [O0 ]0 = O(m) × Sp(2n). This means
that these bosonic versions of Pin and Spin do not describe the symplectic parts of O0 and SO0 . This
phenomenon is due to the natural structure of supervectors: their real projections belong to a space
with an orthogonal structure while the symplectic structure plays no rôle. Up to a nilpotent vector, they
are classical Clifford vectors, whence it is impossible to generate by this approach the real symplectic
geometry that is also present in the structure of O0 and SO0 . That is why we have chosen the name
of "bosonic" Pin and "bosonic" Spin groups. This also explains why we had to extend the space of
superbivectors in section 3. The ordinary superbivectors are generated over ΛN,0 by the wedge product
of supervectors. Then, they can only describe so(m) and not sp(2n) and in consequence, they do not
cover so0 .
As in the classical setting (see [7]), it is possible to obtain the following result that shows, from
another point of view, that Pinb (m|2n)(ΛN ) cannot completely describe O0 .
(2)
Proposition 16. The Lie algebra of Pinb (m|2n)(ΛN ) is included in Rm|2n (ΛN ).
Proof.
Let γ(t) = w1 (t) · · · wk (t) be a path in Pinb (m|2n)(ΛN ) with wj (t) ∈ S(m|2n)(ΛN ) for every t ∈ R and
Pk
0
γ(0) = 1. The tangent to γ at t = 0 is dγ
j=1 w1 (0) · · · wj (0) · · · wk (0). We will show that each
dt t=0 =
summand of
dγ
dt t=0
(2)
belongs to Rm|2n (ΛN ).
21
For j = 1 we have w10 (0)w2 (0) · · · wk (0) = −w10 (0)w1 (0). But w1 (t)w1 (t) ≡ −1 implies
w10 (0)w1 (0) + w1 (0)w10 (0) = 0
⇒
hw10 (0), w1 (0)i = 0.
(2)
Then w10 (0)w1 (0) = −hw10 (0), w1 (0)i + w10 (0) ∧ w1 (0) = w10 (0) ∧ w1 (0) ∈ Rm|2n (ΛN ). For j = 2,
w1 (0)w20 (0) · · · wk (0) = w1 (0)w20 (0)w2 (0)w1 (0) = − [w1 (0)w20 (0)w1 (0)] · [w1 (0)w2 (0)w1 (0)]
= −ψ(w1 (0))[w20 (0)] · ψ(w1 (0))[w2 (0)].
But ψ(w1 (0)) preserves the inner product, so w1 (0)w20 (0) · · · wk (0) = ψ(w1 (0))[w20 (0)] ∧ ψ(w1 (0))[w2 (0)]
(2)
∈ Rm|2n (ΛN ). We can proceed similarly for every j = 3, . . . , k.
5.2
A proper definition for the group Spin(m|2n)(ΛN )
The above approach shows that the radial algebra setting does not contain a suitable realization of SO0
(2)E
in the Clifford superspace framework. Observe that the Clifford realization of so0 given by Rm|2n (ΛN )
lies outside of the radial algebra Rm|2n (ΛN ), which suggests that something similar should happen with
the Clifford realization of the Lie group SO0 . In this case, a proper definition for the Spin group would
(2)E
be generated by the exponentials (in general contained in T (V )/I) of all the elements in Rm|2n (ΛN ), i.e.
n
o
(2)E
Spin(m|2n)(ΛN ) := eB1 · · · eBk : B1 , . . . , Bk ∈ Rm|2n (ΛN ), k ∈ N ,
and the action of this group on Rm|2n (ΛN ) is given by the group homomorphism h : Spin(m|2n)(ΛN ) →
SO0 defined by the restriction to Rm|2n (ΛN ) of
h(eB )[x] = eB xe−B ,
(2)E
B ∈ Rm|2n (ΛN ), x ∈ Rm|2n (ΛN +1 ).
In fact, for every extended superbivector B, h(eB ) maps supervectors into supervectors and admits a
supermatrix representation in Mat(m|2n)(ΛN ) belonging to SO0 . This is summarized below.
(2)E
Proposition 17. Let B ∈ Rm|2n (ΛN ). Then, h(eB )[x] = eφ(B) x for every x ∈ Rm|2n (ΛN +1 ).
Proof.
k
X
k
B j x(−B)k−j holds. Then,
In every associative algebra, the identity [B, [B . . . [B, x] . . .]] =
{z
}
|
j
j=0
k
h(eB )[x] = eB xe−B =
∞
X
Bk
k=0
k!
!
x
∞
X
(−B)k
k=0
!
=
k!
∞
X
k=0
k
j
k−j
X
B
(−B)
x
j!
(k − j)!
j=0
∞
k
∞
∞
X
X
X
1
φ(B)k x
1 X k
B j x(−B)k−j =
[B, [B . . . [B, x] . . .]] =
= eφ(B) x.
=
{z
}
k! j=0 j
k! |
k!
k=0
k=0
k
k=0
(2)E
Remark 5.1. The above proposition means that the Lie algebra isomorphism φ : Rm|2n (ΛN ) → so0 is
the derivative at the origin of the Lie group homomorphism h : Spin(m|2n)(ΛN ) → SO0 , i.e.,
etφ(B) = h(etB )
(2)E
∀t ∈ R, B ∈ Rm|2n (ΛN ).
On account of the connectedness of SO0 it can be shown that the group Spin(m|2n)(ΛN ) is a realization of SO0 in T (V )/I through the representation h.
Theorem 6. For every M ∈ SO0 there exist s ∈ Spin(m|2n)(ΛN ) such that h(s) = M .
22
Proof.
Since SO0 is a connected Lie group (Proposition 12), for every supermatrix M ∈ SO0 there exist
(2)E
X1 , . . . , Xk ∈ so0 such that eX1 · · · eXk = M , see Corollary 3.47 in [8]. Taking B1 , . . . , Bk ∈ Rm|2n (ΛN )
such that φ(Bj ) = Xj , j = 1, . . . , k, we obtain
M x = eX1 · · · eXk x = eφ(B1 ) · · · eφ(Bk ) x = h(eB1 ) ◦ · · · ◦ h(eBk )[x] = h(eB1 · · · eBk )[x].
The above equality is valid for every x ∈ Rm|2n (ΛN +1 ). Then, since h(eB1 · · · eBk ) and M belong to
Mat(m|2n)(ΛN ), Lemma 3 guarantees that s = eB1 · · · eBk ∈ Spin(m|2n)(ΛN ) satisfies h(s) = M .
The decomposition of SO0 given in Theorem 4 provides the exact number of exponentials of extended
superbivectors to be considered in Spin(m|2n)(ΛN ) in order to cover the whole group SO0 . If we consider
(2)E
the subspaces S1 , S2 , S3 of Rm|2n (ΛN ) given by
m(m − 1)
+ n2 ,
2
dim S2 = n2 + n,
S1 = φ−1 so(m) × [sp(2n) ∩ so(2n)] ,
S2 = φ−1 {0m } × [sp(2n) ∩ Sym(2n)] ,
S3 = φ−1 so0 (m|2n)(Λ+
N)
dim S1 =
dim S3 = dim so0 −
m(m − 1)
− n(2n + 1),
2
(2)E
we get the decomposition Rm|2n (ΛN ) = S1 ⊕S2 ⊕S3 , leading to the subset S = exp(S1 ) exp(S2 ) exp(S3 ) ⊂
Spin(m|2n)(ΛN ) which suffices for describing SO0 . Indeed, from Theorem 4 it follows that the restriction
h : S → SO0 is surjective. We now investigate the explicit form of the superbivectors in each one of the
subspaces S1 , S2 and S3 .
Proposition 18. The following statements hold.
1 ≤ j < k ≤ m,
ej ek ,
(i) A basis for S1 is e`2j−1 e`2k−1 + e`2j e`2k , 1 ≤ j ≤ k ≤ n,
e`2j−1 e`2k − e`2j e`2k−1 , 1 ≤ j < k ≤ n.
1 ≤ j ≤ n,
e`2j−1 e`2j ,
(ii) A basis for S2 is: e`2j−1 e`2k−1 − e`2j e`2k , 1 ≤ j ≤ k ≤ n,
e`2j−1 e`2k + e`2j e`2k−1 , 1 ≤ j < k ≤ n.
(iii) S3 consists of all elements of the form (10) with bj,k , Bj,k ∈ ΛN,0 ∩ Λ+
N, and b`j,k ∈ ΛN,1 .
Proof.
We first recall that a basis for the Lie algebra sp(2n) is given by the elements
Aj,k := E2j,2k−1 + E2k,2j−1 , 1 ≤ j ≤ k ≤ n
Bj,k := E2j−1,2k + E2k−1,2j , 1 ≤ j ≤ k ≤ n,
Cj,k := E2k,2j − E2j−1,2k−1 , 1 ≤ j ≤ k ≤ n,
Dj,k := E2j,2k − E2k−1,2j−1 , 1 ≤ j < k ≤ n,
where the matrices Ej,k ∈ Rn×n are defined as in Lemma 5. It holds that ATj,k = Bj,k for 1 ≤ j ≤ k ≤ n,
T
T
Cj,k
= Dj,k for 1 ≤ j < k ≤ n and Cj,j
= Cj,j for 1 ≤ j ≤ n. Hence, for every matrix D0 ∈ sp(2n) we
have
X
X
D0 =
(aj,k Aj,k + bj,k Bj,k + cj,k Cj,k ) +
dj,k Dj,k ,
1≤j≤k≤n
D0T =
X
1≤j<k≤n
(aj,k Bj,k + bj,k Aj,k ) +
1≤j≤k≤n
X
1≤j<k≤n
where aj,k , bj,k , cj,k , dj,k ∈ R.
23
(cj,k Dj,k + dj,k Cj,k ) +
n
X
j=1
cj,j Cj,j ,
(i) From the previous equalities we get that D0T = −D0 if and only if
X
X
D0 =
aj,k (Aj,k − Bj,k ) +
cj,k (Cj,k − Dj,k ) .
1≤j≤k≤n
1≤j<k≤n
Then, a basis for sp(2n) ∩ so(2n) is {Aj,k − Bj,k : 1 ≤ j ≤ k ≤ n} ∪ {Cj,k − Dj,k : 1 ≤ j < k ≤ n}.
The rest of the proof directly follows from Lemma 5.
(ii) In this case we have that D0T = D0 if and only if
X
D0 =
X
aj,k (Aj,k + Bj,k ) +
1≤j≤k≤n
cj,k (Cj,k + Dj,k ) +
n
X
cj,j Cj,j ,
j=1
1≤j<k≤n
whence a basis for sp(2n) ∩ Sym(2n) is
{Aj,k + Bj,k : 1 ≤ j ≤ k ≤ n} ∪ {Cj,j : 1 ≤ j ≤ n} ∪ {Cj,k + Dj,k : 1 ≤ j < k ≤ n}.
The rest of the proof directly follows from Lemma 5.
iii) This trivially follows from Lemma 5.
5.3
Spin covering of the group SO0
It is a natural question in this setting whether the spin group still is a double covering of the group of
rotations, as it is in classical Clifford analysis. In order to answer this question, we investigate how many
times S ⊂ Spin(m|2n)(ΛN ) covers SO0 , or more precisely, the cardinal of the set {s ∈ S : h(s) = M }
given a certain fixed element M ∈ SO0 .
From Proposition 17 we have that the representation h of an element s = eB1 eB2 eB3 ∈ S, Bj ∈ Sj ,
has the form h(s) = eφ(B1 ) eφ(B2 ) eφ(B3 ) . Following the decomposition M = eX eY eZ given in Theorem 4
for M ∈ SO0 , we get that h(s) = M if and only if eφ(B1 ) = eX , B2 = φ−1 (Y ) and B3 = φ−1 (Z). Then,
the cardinal of {s ∈ S : h(s) = M } only depends on the number of extended superbivectors B1 ∈ S1 that
satisfy eφ(B1 ) = eX . It reduces our analysis to finding the kernel of the restriction h|exp(S1 ) : exp(S1 ) →
SO(m) × [Sp(2n) ∩ SO(2n)] of the Lie group homomorphism h to exp(S1 ). This kernel is given by
ker h|exp(S1 ) = {eB : eφ(B) = Im+2n , B ∈ S1 }.
(2)
We recall, from Proposition 18, that B ∈ S1 may be written as B = Bo + Bs where Bo ∈ R0,m is a
classical real bivector and Bs ∈ φ−1 ({0m } × [sp(2n) ∩ so(2n)]). The components Bo , Bs commute and in
consequence, eB = eBo eBs . Consider the projections φo and φs of φ over the algebra of classical bivectors
(2)
R0,m and φ−1 ({0m } × [sp(2n) ∩ so(2n)]) respectively, i.e.
(2)
φs : φ−1 ({0m } × [sp(2n) ∩ so(2n)]) → sp(2n) ∩ so(2n),
φo : R0,m → so(m),
where
φ(B) =
φo (Bo )
0
0
φs (Bs )
or equivalently: φo (Bo )[x] = [Bo , x] for x ∈ Rm|0 (ΛN +1 ), and φs (Bs )[x] = [Bs , x], for x ∈ R0|2n (ΛN +1 ).
Hence eφ(B) = Im+2n if and only if eφo (Bo ) = Im and eφs (Bs ) = I2n . For the first condition, we know
(2)
from classical Clifford analysis that Spin(m) = {eB : B ∈ R0,m } is a double covering of SO(m) and in
φo (Bo )
B0
consequence, e
= Im implies e = ±1. Let us now compute all possible values for eBs for which
φs (Bs )
e
= I2n . To that end, we need the following linear algebra result.
Proposition 19. Every matrix D0 ∈ so(2n) ∩ sp(2n) can be written in the form D0 = RΣRT where
R ∈ SO(2n) ∩ Sp(2n) and
Σ=
0
−θ1
θ1
0
..
.
0
−θn
θn
0
24
,
θj ∈ R,
j = 1, . . . , n.
(17)
Proof.
c
The map Ψ(D0 ) = 12 QD0 QT , where
Q=
1
0
..
.
i
0
..
.
0
1
..
.
0
i
..
.
...
...
..
.
0
0
..
.
0
0
..
.
0
0
0
0
...
1
i
∈ Cn×2n ,
is a Lie group isomorphism between SO(2n) ∩ Sp(2n) and U (n). In addition, Ψ is its own infinitesimal representation on the Lie algebra level, and in consequence, a Lie algebra isomorphism
between
c
so(2n) ∩ sp(2n) and u(n). The inverse of Ψ is given by Ψ−1 (L) = 21 QT L Q + QT Lc Qc . For every
D0 ∈ so(2n) ∩ sp(2n), let us consider the skew-Hermitian matrix L = Ψ(D0 ) ∈ u(n). It is know that
every skew-Hermitian matrix is unitarily diagonalizable and all its eigenvalues are purely imaginary, see
c
[10]. Hence, L = Ψ(D0 ) can be written as L = U Ξ U T where U ∈ U (n) and Ξ = diag(−iθ1 , . . . , −iθn ),
−1
T
−1
θj ∈ R. Then, D0 = Ψ (L) = RΣR where R = Ψ (U ) ∈ SO(2n) ∩ Sp(2n) and Σ = Ψ−1 (Ξ) has the
form (17).
e
Since φs (Bs ) ∈ so(2n) ∩ sp(2n), we have φs (Bs ) = RΣRT as in the previous proposition. Hence,
= ReΣ RT where eΣ is the block-diagonal matrix
φs (Bs )
eΣ = diag(eθ1 J2 , . . . , eθn J2 ) with eθj J2 = cos θj I2 + sin θj J2 .
Hence eφs (Bs ) = I2n if and only if eΣ = I2n , which is seen to be equivalent to θj = 2kj π, kj ∈ Z
(j = 1, . . . , n), or to
Σ=
n
X
2kj π (E2j−1,2j − E2j,2j−1 ) ,
kj ∈ Z (j = 1, . . . , n).
j=1
Now, SO(2n) ∩ Sp(2n) being connected and compact, there exists BR ∈ φ−1 (so(2n) ∩ sp(2n)) such
φ(BR )
that R
. We recall that the h-action leaves any multivector structure invariant, in particular,
=e
(2)E
(2)E
(2)E
B
h[e ] Rm|2n (ΛN ) ⊂ Rm|2n (ΛN ) for every B ∈ Rm|2n (ΛN ). Then, using that φ is the derivative at the
origin of h, we get that the extended superbivector h(eBR )[φ−1 (Σ)] = eBR φ−1 (Σ)e−BR is such that
φ eBR φ−1 (Σ)e−BR = eφ(BR ) Σe−φ(BR ) = RΣRT = φ(Bs ),
−1
implying that Bs = eBR φ−1 (Σ)e−BR . Then, in order to compute eBs = eBR eφ
−1
to compute eφ (Σ) . Following the correspondences given in Lemma 5 we get
φ−1 (Σ) =
n
X
2kj π φ−1 (E2j−1,2j − E2j,2j−1 ) =
j=1
n
X
(Σ) −BR
e
, we first have
2
2
kj π e`2j−1
+ e`2j
.
j=1
and, in consequence
−1
eφ
(Σ)
= exp
n
X
n
Y
2
2
2
2
kj π e`2j−1
+ e`2j
=
exp kj π e`2j−1
+ e`2j
.
j=1
(18)
j=1
2
2
Let us compute exp π e`2j−1
+ e`2j
, j ∈ {1, . . . , n}. Consider x = e`2j−1 −ie`2j , y = e`2j−1 +ie`2j ; where i is
2
2
2
2
the usual imaginary unit in C. It is clear that xy = e`2j−1
+e`2j
+i(e`2j−1 e`2j −e`2j e`2j−1 ) = e`2j−1
+e`2j
+i and
2
2
[x, y] = 2i which is a commuting element. Then, exp π e`2j−1 + e`2j = exp (π xy − iπ) = − exp (π xy).
In order to compute exp (π xy) we first prove the following results.
Lemma 7. For every k ∈ N the following relations hold.
(i) yk x , xyk = −2ik yk−1 ,
(ii) xk yk xy = xk+1 yk+1 − 2ik xk yk .
25
Proof.
(i) We proceed by induction. For k = 1 we get [y, x] = −2i which obviously is true. Now assume that
(i) is true for k ≥ 1, then for k + 1 we get
yk+1 x = y yk x = yxyk − 2ik yk = (xy − 2i)yk − 2ik yk = xyk+1 − 2i(k + 1)yk .
(ii) From (i) we get xk yk xy = xk xyk − 2ik yk−1 y = xk+1 yk+1 − 2ik xk yk .
Pk
k
k−j
Lemma 8. For every k ∈ N it holds that (xy) =
S(k, j) xj yj , where S(n, j) is the
j=1 (−2i)
Stirling number of the second kind corresponding to k and j.
Remark 5.2. The Stirling number of the second kind S(k, j) is the number of ways of partitioning a
set of k elements into j non empty subsets. Among the properties of the Stirling numbers we recall the
following ones:
S(k, 1) = S(k, k) = 1,
∞
X
S(k + 1, j + 1) = S(k, j) + (j + 1)S(k, j + 1),
k=j
j
S(k, j)
(ex − 1)
xk
=
.
k!
j!
Proof of Lemma 8.
We proceed by induction. For k = 1 the statement clearly is true. Now assume it to be true for k ≥ 1.
Using Lemma 7, we have for k + 1 that
(xy)k+1 =
k
k
X
X
(−2i)k−j S(k, j) xj yj xy =
(−2i)k−j S(k, j)xj+1 yj+1 + (−2i)k+1−j j S(k, j) xj yj
j=1
j=1
k
= (−2i) xy +
k−1
X
!
(−2i)
k−j
[S(k, j) + (j + 1)S(k, j + 1)] x
j+1
y
j+1
+ xk+1 yk+1
j=1
=
k+1
X
(−2i)k+1−j S(k + 1, j) xj yj .
j=1
Then we obtain
eπxy =
∞
X
πk
k=0
=1+
k!
(xy)k = 1 +
k=1 j=1
k!
(−2i)k−j S(k, j) xj yj
∞
k
X
(−2πi)
(−2i)−j
(−2i)k−j S(k, j) xj yj = 1 +
S(k, j) xj yj
k!
k!
j=1
∞
X
∞ X
∞
X
πk
j=1 k=j
=1+
∞ X
k
X
πk
∞
X
k=j
j
−1
xj yj = 1.
j!
−2πi
(−2i)−j
e
j=1
2
2
from which we conclude that exp π e`2j−1
+ e`2j
= − exp (π xy) = −1.
Remark 5.3. Within the algebra AlgR {e`1 , . . . , e`2n } the elements e`2j−1 , e`2j may be identified with the opeπ
π
rators e 4 i ∂aj , e− 4 i aj respectively, the aj ’s being real variables. Indeed, these identifications immediately
lead to the Weyl algebra defining relations
π
π
π
π
e 4 i ∂aj e− 4 i ak − e− 4 i ak e 4 i ∂aj = ∂aj ak − ak ∂aj = δj,k .
2
2
Hence e`2j−1
+ e`2j
may be identified with the harmonic oscillator i ∂a2j − a2j and in consequence, the
h
i
2
2
element exp π e`2j−1
+ e`2j
corresponds to exp πi ∂a2j − a2j . We recall that the classical Fourier
transform in one variable can be written as an operator exponential
π
π
F[f ] = exp
i exp
i ∂a2j − a2j [f ].
4
4
h
i
Hence, exp πi ∂a2j − a2j = −F 4 = −id, where id denotes the identity operator.
26
P
kj
Qn
−1
2
2
= (−1) kj , whence eBs = ±1. Then,
Going back to (18) we have eφ (Σ) = j=1 exp π e`2j−1
+ e`2j
for B = Bo + Bs ∈ S1 such that eφ(B) = Im+2n , we have eB = eBo eBs = ±1, i.e. ker h|exp(S1 ) = {−1, 1}.
Theorem 7. The set S = exp(S1 ) exp(S2 ) exp(S3 ) is a double covering of SO0 .
Pn θ
2
2
,
Remark 5.4. As shown before, every extended superbivector of the form B = j=1 2j π e`2j−1
+ e`2j
θj ∈ R, belongs to S1 . Then, though the identifications made in remark 5.3 we can see all the operators
n
n
n
i Y
h π
X
Y
θ
π
j
exp
πi(∂a2j − a2j ) =
exp θj i(∂a2j − a2j ) =
exp −θj i Fa2θj j ,
2
2
2
j=1
j=1
j=1
2θ
as elements of the Spin group in superspace. Here, Faj j denotes the one-dimensional fractional Fourier
transform of order 2θj in the variable aj .
6
Conclusions and future work
In this paper we have shown that vector reflections in superspace are not enough to describe the set of
linear transformations leaving the inner product invariant. This constitutes a very important difference
with the classical case in which the algebra of bivectors x ∧ y is isomorphic to the special orthogonal
algebra so(m). Such a property is no longer fulfilled in this setting. The real projection of the algebra of
(2)
superbivectors Rm|2n (ΛN ) does not include the symplectic algebra structure which is present in the Lie
algebra of supermatrices so0 , corresponding to the group of super rotations.
That fact has an major impact on the definition of the Spin group in this setting. The set of
elements defined through the multiplication of an even number of unit vectors in superspace does not
suffice for describing Spin(m|2n)(ΛN ). A suitable alternative, in this case, is to define the (super) spin
elements as products of exponentials of extended superbivectors. Such an extension of the Lie algebra
of superbivectors contains, through the corresponding identifications, harmonic oscillators. This way, we
obtain the Spin group as a cover of the set of superrotations SO0 through the usual representation h. In
addition, every fractional Fourier transform can be identified with a spin element.
In forthcoming work, we will prove the invariance of the (super) Dirac operator ∂x under the corresponding actions of this (super) Spin group. We will also study the invariance of the Hermitian system
under the action of the corresponding Spin subgroup in superspace.
Acknowledgements
Alí Guzmán Adán is supported by a BOF-doctoral grant from Ghent University with grant number
01D06014.
References
[1] F. A. Berezin. Introduction to Super Analysis. D. Reidel Publishing Co., Inc., New York, NY, USA,
1987.
[2] H. De Bie and F. Sommen.
322(12):2978–2993, 2007.
A clifford analysis approach to superspace.
Annals of Physics,
[3] H. De Bie and F. Sommen. Correct rules for clifford calculus on superspace. Advances in Applied
Clifford Algebras, 17(3):357–382, 2007.
[4] H. De Schepper, A. Guzman Adan, and F. Sommen. Hermitian clifford analysis on superspace.
Submitted for publication, 2016.
[5] Hennie De Schepper, Alí Guzmán Adán, and Frank Sommen. The radial algebra as an abstract
framework for orthogonal and hermitian clifford analysis. Complex Analysis and Operator Theory,
pages 1–34, 2016.
27
[6] P. DIENES. The exponential function in linear algebras. The Quarterly Journal of Mathematics,
os-1(1):300–309, 1930.
[7] Thomas Friedrich. Dirac operators in Riemannian geometry, volume 25 of Graduate Studies in
Mathematics. American Mathematical Society, Providence, RI, 2000. Translated from the 1997
German original by Andreas Nestke.
[8] Brian Hall. Lie groups, Lie algebras, and representations: An elementary introduction, volume 222
of Graduate Texts in Mathematics. Springer, Cham, second edition, 2015.
[9] Joachim Hilgert and Karl-Hermann Neeb. Structure and geometry of Lie groups. Springer Monographs in Mathematics. Springer, New York, 2012.
[10] Roger A. Horn and Charles R. Johnson. Matrix analysis. Cambridge University Press, Cambridge,
second edition, 2013.
[11] Anthony W. Knapp. Lie groups beyond an introduction, volume 140 of Progress in Mathematics.
Birkhäuser Boston, Inc., Boston, MA, second edition, 2002.
[12] F. Sommen. An algebra of abstract vector variables. In Portugaliae Mathematica 54 (3) (1997),
pages 287–310, 1997.
[13] F. Sommen. An extension of clifford analysis towards super-symmetry. In Clifford algebras and their
applications in mathematical physics, pages 199–224. Springer, 2000.
[14] F. Sommen. Clifford analysis on super-space. Advances in Applied Clifford Algebras, 11(1):291–304,
2001.
[15] F. Sommen. Analysis Using Abstract Vector Variables, pages 119–128. Birkhäuser Boston, Boston,
MA, 2002.
[16] F. Sommen. Clifford analysis on super-space. ii. Progress in analysis, 1:383–405, 2003.
[17] Takeo Yokonuma. Tensor spaces and exterior algebra, volume 108 of Translations of Mathematical
Monographs. American Mathematical Society, Providence, RI, 1992. Translated from the 1977
Japanese edition by the author.
Hennie de Schepper
Clifford Research Group, Department of Mathematical Analysis, Ghent University,
Krijgslaan 281, 9000 Gent, Belgium.
e-mail: [email protected]
Alí Guzmán Adán
Clifford Research Group, Department of Mathematical Analysis, Ghent University,
Krijgslaan 281, 9000 Gent, Belgium.
e-mail: [email protected]
Frank Sommen
Clifford Research Group, Department of Mathematical Analysis, Ghent University,
Krijgslaan 281, 9000 Gent, Belgium.
e-mail: [email protected]
28
| 4math.GR
|
Uniform Asymptotic Inference and the Bootstrap After Model
Selection
Ryan Tibshirani, Alessandro Rinaldo, Rob Tibshirani, and Larry Wasserman
arXiv:1506.06266v3 [math.ST] 9 Aug 2017
Carnegie Mellon University and Stanford University
Abstract
Recently, Tibshirani et al. (2016) proposed a method for making inferences about parameters
defined by model selection, in a typical regression setting with normally distributed errors. Here,
we study the large sample properties of this method, without assuming normality. We prove that
the test statistic of Tibshirani et al. (2016) is asymptotically valid, as the number of samples n
grows and the dimension d of the regression problem stays fixed. Our asymptotic result holds uniformly over a wide class of nonnormal error distributions. We also propose an efficient bootstrap
version of this test that is provably (asymptotically) conservative, and in practice, often delivers
shorter intervals than those from the original normality-based approach. Finally, we prove that
the test statistic of Tibshirani et al. (2016) does not enjoy uniform validity in a high-dimensional
setting, when the dimension d is allowed grow.
1
Introduction
There has been a recent surge of work on conducting formally valid inference in a regression setting
after a model selection event has occurred, see Berk et al. (2013), Lockhart et al. (2014), Tibshirani
et al. (2016), Lee et al. (2016), Fithian et al. (2014), Bachoc et al. (2014), just to name a few. Our
interest in this paper stems in particular from the work of Tibshirani et al. (2016), who presented a
method to produce valid p-values and confidence intervals for adaptively fitted coefficients from any
given step of a sequential regression procedure like forward stepwise regression (FS), least angle
regression (LAR), or the lasso (the lasso is meant to be thought of as tracing out a sequence of models
along its solution path, as the penalty parameter descends from λ = ∞ to λ = 0). These authors use
a statistic that is carefully crafted to be pivotal after conditioning on the model selection event. This
idea is not specific to the sequential regression setting, and is an example of a broader framework
that we might call selective pivotal inference, applicable in many other settings, as in, e.g., Taylor
et al. (2016), Lee et al. (2016), Lee & Taylor (2014), Loftus & Taylor (2014), Reid et al. (2017), Choi
et al. (2014), Fithian et al. (2014), Hyun et al. (2016).
A key to the methodology in Tibshirani et al. (2016) (and much of the work in selective pivotal
inference) is to the assumption of normality of the errors. To fix notation, consider the regression
of a response Y ∈ Rn on predictor variables X 1 , . . . , X d ∈ Rn , stacked together as columns of a matrix
X ∈ Rn×d . We will treat the predictors X are fixed (nonrandom), and assume the model
Yi = θ i + ² i , i = 1, . . . , n,
(1)
where θ ∈ Rn is an unknown mean parameter of interest. Tibshirani et al. (2016) assume that the
errors ²1 , . . . , ²n are i.i.d. N(0, σ2 ), where the error variance σ2 > 0 is known. An advantage of their
1
approach is that it does not require θ to be an exact linear combination of the predictors X 1 , . . . , X d ,
and makes no assumptions about the correlations among these predictors. But as far as the finitesample guarantees are concerned, normality of the errors is crucial. In this work, we examine the
properties of the test statistic proposed in Tibshirani et al. (2016)—hereafter, the truncated Gaussian
(TG) statistic—without using an assumption about normal errors. We only assume that ²1 , . . . , ²n are
i.i.d. from a distribution with mean zero and essentially no other restrictions.
A high-level description of the selective pivotal inference framework for sequential regression is
as follows (details are provided in Section 2). FS, LAR, or the lasso is run for some number of steps
k, and a model is selected, call it M. For FS and LAR, this model will always have k active variables,
and for the lasso, it will have at most k, as variables can be added to or deleted from the active set
at each step. We specify a linear contrast of the mean vT θ of interest, e.g., one giving the coefficient
of a variable of interest in the model M at step k, in the regression of θ onto the active variables. By
assuming normal errors in (1), and examining the distribution of vT Y conditional on having selected
c ) = M, we can construct a confidence interval C α satisfying
model M, which we denote by M(Y
¯
³
´
¯c
P vT θ ∈ C α ¯ M(Y
) = M = 1 − α,
for a given α ∈ [0, 1]. The interpretation: if we were to repeatedly draw Y from (1) and run FS, LAR,
or the lasso for k steps, and only pay attention to cases in which we selected model M, then among
these cases, the constructed intervals C α = C α (Y ; M) contain vT θ with frequency tending to 1 − α.
The above is a conditional perspective of the selective pivotal inference framework for FS, LAR,
and lasso. An unconditional or marginal point of view is also possible, which we now describe. For
each possible selected model M, a constrast vector v M is specified, and the contrast vT
θ is considered
M
c ) = M. To be concrete, we can again think of a setup such that vT θ
when model M is selected, M(Y
M
gives the coefficient of a variable in the model M at step k, in the projection of θ onto the active set.
Confidence intervals are then constructed in exactly the same manner as above (without change),
and conditional coverage over all models M implies the following unconditional property for C α ,
³
´
P vTc θ ∈ C α = 1 − α.
M (Y )
The interpretation is different: if we were to repeatedly draw Y from (1) and run FS, LAR, or lasso
c )), then these intervals contain their
for k steps, and construct confidence intervals C α = C α (Y ; M(Y
T
respective targets v c θ with frequency approaching 1 − α. Notice that, by construction, the target
M (Y )
itself may change each time we draw Y , though it is the same for all Y that give rise to the same
selected model. In terms of the setting for regression contrasts described above, each time we draw Y
and carry out the inferential procedure, the interval C α covers the coefficient of a possibly different
variable in the active model, in the projection of θ onto the active variables. Figure 1 demonstrates
this point.
1.1
Uniform convergence
When making asymptotic inferential guarantees, as we do in this paper, it is important to be clear
about the type of guarantee. Here we review the concepts of uniform convergence and validity. Let
ξ1 , . . . , ξn ∈ Rs be random vectors with joint distribution (ξ1 , . . . , ξn ) ∼ F n , where F n ∈ P n , and P n is a
class of distributions. For example, we could have ξ1 , . . . , ξn ∈ Rs i.i.d. from F, and the class P n could
contain product distributions of the form F n = F × . . . × F (n times); our notation allows for a more
2
6
4
−2
0
2
●●
●
●
●●●●●●
●
●
●● ●
●●
●
●
●●●●
●
●
●●●●●
●
●
●
●● ●● ●
●
●
●
● ●●● ●●●● ●●●
●
●
●
●
●
●
●
● ●● ●●●●●●●●●●●●●● ●●●● ●●●● ●●●● ●●●● ●●● ●● ●
−6
Confidence interval
● ●
0
20
40
60
80
100
Repetition
Figure 1: An example of conditional and unconditional coverage for one step of FS (the variables are normalized, and this is equivalent to one step of LAR, or lasso). Here n = 20 and d = 3, and a response Y was drawn
100 times from a model as in (1) with i.i.d. N (0, σ2 ) errors. The different colors denote different active models
that were selected after one step, where an active model is a variable-sign pair, namely, the variable achieving
the largest absolute inner product with Y , and the sign of this inner product. Across the 100 repetitions, the
circles denote a target to be covered, and the segments are 90% confidence intervals. E.g, the color green corresponds to the model + X 2 , so in repetitions 1, 3, 11, 12, etc., X 2T Y was largest among all absolute inner products
of variables with Y , and the green segments denote 90% confidence intervals designed to cover the contrast X 2T θ .
Similarly, red corresponds to the model − X 1 , and blue to + X 3 . Dotted segments indicate that the given interval
does not cover its target. The empirical coverage among green intervals: 21/21, among red intervals: 61/70,
and among blue intervals: 8/9. Hence in each case, the empirical coverage is close to the nominal 90% level.
Further, in total, i.e., unconditionally, the empirical coverage is 90/100, right at the nominal 90% level.
general setup than this one. Let Wn = T n (ξ1 , . . . , ξn ) for a statistic T n , and W ∼ G, where Wn ,W ∈ R q .
We will say that Wn , converges uniformly in distribution to W, over P n , provided that
¯
¯
(2)
lim sup sup ¯PFn (Wn ≤ x) − P(W ≤ x)¯ = 0.
n→∞ F ∈P x∈R q
n
n
(The above inequalities, as in Wn ≤ x and W ≤ x, are meant to be interpreted componentwise; we are
also implicitly assuming that the limiting distribution G is continuous, otherwise the above inner
supremum should be restricted to continuity points x of G.) This is much stronger than the notion of
pointwise convergence in distribution, which only requires that
¯
¯
lim sup ¯PFn (Wn ≤ x) − P(W ≤ x)¯ = 0,
(3)
n→∞ x∈R q
for a particular sequence of distributions F n , n = 1, 2, 3, . . ..
A recent article by Kasy (2015) emphasizes the importance of uniformity in asymptotic approximations. This authors points out that a uniform version of the continuous mapping theorem follows
directly from a standard proof of the continuous mapping theorem (e.g., see Theorem 2.3 in van der
Vaart (1998)).
Lemma 1. Suppose that Wn converges uniformly in distribution to W, with respect to the class P n .
Let ψ : R q → R be a map that is continuous on a set D, such that P(W ∈ D) = 1. Then ψ(Wn ) converges
uniformly in distribution to ψ(W) with respect to P n .
3
Kasy (2015) also remarks that the central limit theorem for triangular arrays, specifically the
Lindeberg-Feller central limit theorem (e.g., Proposition 2.27 in van der Vaart (1998)) naturally extends to the uniform case. The logic is, roughly speaking: uniform convergence in (2) is equivalent to
pointwise convergence over all sequences of distributions F n , n = 1, 2, 3, . . ., and triangular arrays, by
design, can have a different distribution assigned to each row. Therefore if the Lindeberg condition
holds for any possible sequence, then so does the convergence to normality.
Lemma 2. Let ξ1 , . . . , ξn ∈ R q be a triangular array of independent random vectors, with joint distribution F n . Assume ξ1 , . . . , ξn have mean zero and finite variance. Also assume that for any sequence
F n ∈ P n , n = 1, 2, 3, . . ., we have
lim
n→∞
n
X
³
´
EFn kξ i k22 · 1{kξ i k2 ≥ ²} = 0, for all ² > 0,
i =1
and
lim
n→∞
n
X
CovFn (ξ i ) = Σ,
i =1
where Σ does not depend on the sequence F n , n = 1, 2, 3, . . .. Then Wn =
to W ∼ N(0, Σ), uniformly with respect to P n .
Pn
i =1 ξ i
converges in distribution
In our work, a motivating reason for the study of uniform convergence is the associated property
of uniform validity of asymptotic confidence intervals. That is, if Wn = Wn (µ) depends on a parameter
µ = µ(F n ) of the distribution F n , but W does not, then we can consider any (1 − α) confidence set C n,α
built from a (1 − α) probability rectangle R α of W,
C n,α = {µ : Wn (µ) ∈ R α },
and the uniform convergence of Wn to W, really just by rearranging its definition in (2), implies
¯
¯
³
´
¯
¯
lim sup sup ¯PFn µ(F n ) ∈ C n,α − (1 − α)¯ = 0.
(4)
Meanwhile, pointwise convergence as in (3) only implies
¯
¯
³
´
¯
¯
lim sup ¯PFn µ(F n ) ∈ C n,α − (1 − α)¯ = 0,
(5)
n→∞ F ∈P α∈[0,1]
n
n
n→∞ α∈[0,1]
for a particular sequence F n , n = 1, 2, 3, . . .. For a confidence set satisfying (4), and a given tolerance
² > 0, there exists a sample size n(²) such that the coverage is guaranteed to be at least 1 − α − ², for
n ≥ n(²), no matter the underlying distribution (over the class of distributions in question). Note that
this is not necessarily true for a pointwise confidence set as in (5), as the required sample size here
could depend on the particular distribution under consideration.
1.2
Summary of main results
An overview of our main contributions is as follows.
1. We establish that TG statistics for typical inferences along the FS, LAR, and lasso paths only
depend on the data (X , Y ) through n1 X T X and p1n X T Y (Lemmas 3, 4, and 5 in Section 3), which
is important since these two quantities have asymptotic limits in a standard low-dimensional
asymptotic setup.
4
2. Placing mild constraints on the mean and error distribution in (1), and treating the dimension
d as fixed, we prove that the TG test statistic is asymptotically pivotal, converging to U(0, 1)
(the standard uniform distribution), when evaluated at the true population value for its pivot
argument. We show that this holds uniformly over a wide class of distributions for the errors,
without any real restrictions on the predictors X (first part of Theorem 7 in Section 4).
3. The resulting confidence intervals are therefore asymptotically uniformly valid, over the same
class of distributions (second part of Theorem 7 in Section 4).
4. The above asymptotic results assume that the error variance σ2 is known, so for σ2 unknown,
we propose a plug-in approach that replaces σ2 in the TG statistic with a simple estimate, and
alternatively, an efficient bootstrap approach. Both allow for conservative asymptotic inference
(Theorem 11 in Section 5).
5. We present detailed numerical experiments that support the asymptotic validity of the TG pvalues and confidence intervals for inference in low-dimensional regression problems that have
nonnormal errors (Section 6). Our experiments reveal that the plug-in and bootstrap versions
also show good performance, and the bootstrap method can often deliver substantially shorter
intervals than those based directly on the TG statistic.
6. Our experiments also also suggest that the TG test statistic (and plug-in, bootstrap variants)
may be asymptotically valid in even broader settings not covered by our theory, e.g., problems
with heteroskedastic errors and (some) high-dimensional problems.
7. We prove that TG statistic does not exhibit a general uniform convergence to U(0, 1) when the
dimension d is allowed to increase (Theorem 12 in Section 7).
1.3
Related work
A recent paper by Tian & Taylor (2017) is very related to our work here. These authors examine the
asymptotic distribution of the TG statistic under nonnormal errors. Their main result proves that
the TG statistic is asymptotically pivotal, under some restrictions on the model selection events in
question. We view their work as providing a complementary perspective to our own: they consider a
setting where the dimension d grows, but place strong regularity conditions on the selected models;
we adopt a more basic setting with d fixed, and prove more broad uniformly valid convergence results
for the TG pivot, free of regularity conditions.
In a sequence of papers, Leeb & Potscher (2003, 2006, 2008) prove that in a classical regression
setting, it is impossible to find the distribution of a post-selection estimator of the underlying coefficients, even asymptotically. Specifically, they prove for an estimate βb of some underlying coefficient
p
vector β0 , any quantity of the form Q n = nA(βb − β0 ), for a linear transform A, cannot be used for
inference after model selection. Though Q n can be made to be pivotal or at least asymptotically pivotal (once A is chosen once appropriately), this is no longer true in the presence of selection, even if
the dimension d is fixed and the sample size n approaches ∞. Furthermore, they show that there is
no uniformly consistent estimate of the distribution of Q n (either conditionally or unconditionally),
which makes Q n unsuitable for inference. This fact is essentially a manifestation of the well-known
Hodges phenomenon. The selective pivotal inference framework, and hence our paper, circumvents
this problem as we do not claim (nor attempt) to estimate the distribution of Q n , and instead make
inferences using an entirely different pivot that is constructed via a careful conditioning scheme.
5
1.4
Notation
As our paper considers an asymptotic regime, with the number of samples n growing, we will often
use a subscript n to mark the dependence of various quantities on the sample size. An exception is
our notation for the predictors, response, and mean, which we will always denote by X , Y , θ , respectively. Though these quantities will (of course) vary with n, our notation hides this dependence for
simplicity.
When it comes to probability statements involving Y , drawn from (1), we will write P f (θ)=µ ( · ) to
denote the probability operator under a mean vector θ such that f (θ ) = µ. With a subscript omitted,
as in P( · ), it is implicit that the probability is taken under θ . Also, we will generally write y (lowercase) for an arbitrary response vector, and Y (uppercase) for a random response vector drawn from
(1). This is intended to distinguish statements that hold for an arbitrary y, and statements that hold
c the model selection procedure
for a random Y with a certain distribution. Lastly, we will denote M
associated with the regression algorithm under consideration (FS, LAR, or lasso), and we will treat
c
this as a mapping from Rn to the space of models, so that M(y)
is a fixed quantity, representing the
c ) is a random variable, representmodel selected when the response is the fixed vector y, and M(Y
ing the model selected when the response is the random vector Y . Similar notation will be used for
related quantities.
2
Selective inference
In this section, we review the selective pivotal inference framework for sequential regression procedures. We present interpretations for the inferences from both conditional and unconditional perpsectives, in Sections 2.2 and 2.5, respectively. The other subsections provide the necessary details for
understanding the framework, beginning with the selection events encountered along the FS, LAR,
and lasso paths.
2.1
Model selection
Consider forward stepwise regression (FS), least angle regression (LAR), or the lasso, run for a number of steps k, where k is arbitrary (but treated as fixed throughout this paper). Such a procedure
S
defines a partition of the sample space, Rn = M ∈M Π M , with elements
c = M }, M ∈ M.
Π M = { y : M(y)
(6)
c denotes the selected model from the given k-step procedure, run on y, and M is the space
Here M(y)
c ) a selected model may be bit of an abuse of common nomenclature,
of possible models. Calling M(Y
c
because, as we will see, M(y) will describe more than just a set of selected variables at the point y.
c as a representation of the decisions made by the algorithm across its k
In fact, one can think of M(y)
b` (y), sb` (y)) : ` = 1, . . . , k}, comprised of two things:
c = {( A
steps. For FS, we define M(y)
b` (y), ` = 1, . . . , k, denoting the variables that are given nonzero
1. a sequence of active sets A
coefficients, at each of the k steps;
2. a sequence of sign vectors sb` (y), ` = 1, . . . , k, denoting the signs of nonzero coefficients, at each
of the k steps.
6
b1 (y) ⊆ A
b2 (y) ⊆ A
b3 (y) ⊆ . . ., as FS selects one variable to add
The active sets are nested across steps, A
to the active set at each step. However, the sign vectors sb1 (y), sb2 (y), sb3 (y), . . . are not, since these are
determined by least squares on the active variables at each step. Hence, as defined, the number of
c
possible models M(y)
after k steps of FS is
2
|M| = d · (d − 1) · · · (d − k + 1) · 2 · 22 · · · 2k = O(d k 2k ).
Moreover, the corresponding partition elements Π M , M ∈ M in (6) are all convex cones. The proof of
this fact is not difficult, and requires only a slight modification of the arguments in Tibshirani et al.
(2016), given in Appendix A.1 for completeness. The result is easily seen for k = 1: after one step of
FS, assuming without a loss of generality that X 1 , . . . , X d have unit norm, we can express, e.g.,
© ¡
¢
ª ©
ª
b1 (y), sb1 (y) = (1, 1) = y : X T y ≥ ± X T y, j = 2, . . . , d
y: A
1
j
=
d ©
\
ª ©
ª
y : (X 1 − X j )T y ≥ 0 ∩ y : (X 1 + X j )T y ≥ 0 ,
j =2
2
3
the right-hand side above being an intersection of half-spaces passing through zero, and therefore a
b1 (y), sb1 (y)), these cones form a partition of
convex cone. As we enumerate the possible choices for ( A
n
R . Figure 2 shows an illustration.
1
− X3
+ X2
0
− X1
+ X1
+ X3
−3
−2
−1
− X2
−3
−2
−1
0
1
2
3
Figure 2: An example of the model selection partition from one step of FS (the variables are normalized, and
this is equivalent to one step of LAR, or lasso). Here n = 2 and d = 3. The colors indicate the regions of the
sample space R2 for which different models—pairs of active variables and signs—are selected, so that, e.g., the
red region contains points in R2 that are maximally aligned with X 1 .
c
For LAR and the lasso, we need to modify the definition of the selected model M(y)
in order for
the resulting partition elements in (6) to be convex cones. We add an “extra” bit of model information
7
b
c = {( A(y),
and define M(y)
sb(y), Ib` (y)) : ` = 1, . . . , k}, where Ib` (y) is a list of variables that play a special
role in the construction of the LAR or lasso active set at the `th step, but that a user would not
typically pay attention to. In truth, the latter quantity is only a detail that is included so that Π M ,
M ∈ M are convex cones (without it, the partition elements would each be a union of cones), and so
we do not describe it here. Furthermore, it does not affect our treatment of inference in what follows,
and for this reason, we will largely ignore the minor differences in model selection events between
FS, LAR, and lasso hereafter.
The description of Ib` (y), ` = 1, . . . , k, and the proof that the partition elements Π M , M ∈ M are
cones for LAR and lasso, mirrors that in Tibshirani et al. (2016), and is again given in Appendix A.1.
b1 (y) ⊆ A
b2 (y) ⊆ A
b3 (y) ⊆ . . ., since one variable is added
Like FS, the active sets from LAR are nested, A
to the active set at each step. But for the lasso, this is not necessarily true, as in this case variables
can be either added or deleted at each step.
2.2
Inference after selection
We review the selective pivotal inference approach for hypothesis testing after model selection with
FS, LAR, or the lasso. The technical details of the TG statistic are deferred to the next two subsections, as they are not needed to understand how the method is used. The null hypotheses we consider
are of the form H0 : vT θ = 0. An important special case occurs when the linear contrast vT θ gives
a normalized coefficient in the regression of θ onto a subset of the variables in X . To be specific, in
T
T
this case v = X A (X A
X A )−1 e j /(e Tj (X A
X A )−1 e j )1/2 , for a subset A ⊆ {1, . . . , d }, where we let X A ∈ Rn×| A |
T
denote the submatrix of X whose columns correspond to elements of A (with X A
X A assumed to be
invertible for the chosen subset), and we write e j for the jth standard basis vector. This gives
T
T
e Tj (X A
X A )−1 X A
θ
vT θ = q
:= β j (A),
T X )−1 e
e Tj (X A
j
A
(7)
and therefore H0 : vT θ = 0 is a test for the significance of the jth normalized coefficient in the linear
projection of θ onto X A , written as β j (A) for short. (Though the normalization in the denominator is
irrelevant for this significance test, it acts as a key scaling factor for the asymptotics in Section 4.)
The idea of using a projection parameter for inference, β j (A), has also appeared in, e.g., Berk et al.
(2013), Wasserman (2014), Lee et al. (2016). Here is now a summary of the testing framework.
• For each possible model M ∈ M, and any v ∈ Rn and µ ∈ R, a TG statistic T( · ; M, v, µ) is defined
(see (10), in the next subsection), whose domain is the partition element Π M . This can be used
as follows: if Y is drawn from (1), and lands in the partition element Π M for model M, then the
statistic T(Y ; M, v, µ) provides us with a test for the hypothesis H0 : vT θ = µ.
• A concrete case to keep in mind, denoting M = {(A ` , s ` ) : ` = 1, . . . , k}, is a choice of v such that
vT θ = β j (A ` ), in the notation of (7). This is the jth normalized coefficient in the regression of θ
onto the active variables X A ` , for an active set A ` at some step ` = 1, . . . , k.
• Assume i.i.d. N(0, σ2 ) errors in (1). Under the null hypothesis, the TG statistic has a standard
uniform distribution, over draws of Y that land in Π M . Mathematically, this is the property
¯
³
´
¯c
PvT θ=µ T(Y ; M, v, µ) ≤ t ¯ M(Y
) = M = t,
(8)
8
for all t ∈ [0, 1]. The probability above is taken over an arbitrary mean parameter θ for which
c ) = M only
vT θ = µ (in fact, the TG statistic is constructed so that the law of T(Y ; M, v, µ) | M(Y
T
depends on θ through v θ , so this is unambiguous). In order for (8) to hold, of course, v and µ
cannot be random, i.e., they cannot depend on Y , though they can be functions of M.
• Thus T(Y ; M, v, µ) serves as a valid p-value (with exact finite sample size) for testing the null
c ) = M.
hypothesis H0 : vT θ = µ, conditional on M(Y
• A confidence interval is obtained by inverting the test in (8). Given a desired confidence level
1 − α, we define C α to be the set of all values µ such that α/2 ≤ T(Y ; M, v, µ) ≤ 1 − α/2. Then, by
construction, the property in (8) (which we reiterate, assumes i.i.d. N(0, σ2 ) errors) translates
into
¯
³
´
¯c
P vT θ ∈ C α ¯ M(Y
) = M = 1 − α.
(9)
The interpretation of the above statement is straightforward: the random interval C α contains
c ) = M.
the fixed parameter vT θ with probability 1 − α, conditional on M(Y
2.3
The truncated Gaussian pivot
We now describe the truncated Gaussian (TG) pivotal quantity in detail. As defined in Section 2.1, if
c for the selected model from the given algorithm (FS, LAR, or lasso), run for k steps on
we write M(y)
c = M } is a convex cone, for any fixed achieveable model M. Hence
y, then Π M = { y : M(y)
c = M } = { y : Q M y ≥ 0},
Π M = { y : M(y)
for a fixed matrix Q M (here the inequality is meant to be interpreted componentwise). Now to define
the pivot T( · ; M, v, µ) for testing H0 : vT θ = µ, several preliminary quantities must be introduced:
w=
QM v
kvk22
, a(y; M, v) = vT y − min
i :w i >0
(Q M y) i
(Q M y) i
, and b(y; M, v) = vT y − max
.
i :w i <0
wi
wi
The TG pivot is then defined by
¶
µ T
¶
µ
v y−µ
b(y; M, v) − µ
−Φ
Φ
σkvk2
σkvk2
T(y; M, v, µ) = µ
¶
µ
¶.
a(y; M, v) − µ
b(y; M, v) − µ
Φ
−Φ
σkvk2
σkvk2
(10)
This has the following property, as stated in (8): when Y is drawn from (1) with i.i.d. N(0, σ2 ) errors,
c ) = M. See Lemmas 1
and vT θ = µ, the pivot T(Y ; M, v, µ) is uniformly distributed conditional on M(Y
and 2 in Tibshirani et al. (2016) for a proof of this result.
2.4
P-values and confidence intervals
For the null hypothesis H0 : vT θ = 0, we have seen from (8) that T(Y ; M, v, 0) acts as a proper (conditional) p-value. But as defined in (10), the statistic T(Y ; M, v, 0) is implicitly aligned to have power
against the one-sided alternative hypothesis H1 : vT θ > 0. Therefore, when seeking to test the significance of, say, the jth coefficient in the projected linear model of θ on X A ` , we will actually choose v
so that
vT θ = (s ` ) j β j (A ` ),
(11)
9
T
T
where recall (s ` ) j = sign(e Tj (X A
X A ` )−1 X A
y) is the sign of the jth coefficient in the regression of y
`
`
onto the set A ` of active variables, for y ∈ Π M . This orients the test in a meaningful direction: vT θ > 0
is now the hypothesis that the jth coefficient in the projection of θ onto X A ` is nonzero, and shares the
same sign as the jth coefficient in the projection of y onto X A ` , over y ∈ Π M ; that is, with the above
choice of v, the p-value T(Y ; M, v, 0) is designed to be small when the jth coefficient in the projection
of Y on X A ` corresponds to a projected population effect that is both large and of the same sign as
this computed coefficient. Beyond the current subsection, we will not be explicit about the sign factor
in (11) when discussing such contrasts (i.e., those giving regression coefficients in a projected linear
model for θ ), but it is implicitly understood when computing one-sided p-values.
A statistic aligned to have power against the two-sided alternative H1 : vT θ 6= 0 is simply given
by 2 min{T(Y ; M, v, 0), 1 − T(Y ; M, v, 0)}. For purely testing purposes, we find the one-sided p-values
discussed above to be more natural, and hence these will serve as our default. On the other hand,
for constructing confidence intervals, we prefer to invert the two-sided statistics, since these lead to
two-sided intervals. As
2 min{T(Y ; M, v, µ), 1 − T(Y ; M, v, µ)} ≥ α ⇐⇒ α/2 ≤ T(Y ; M, v, µ) ≤ 1 − α/2,
the previously described confidence interval in (9) is just given by inverting the two-sided pivot.
To summarize: the default in this work, as with Tibshirani et al. (2016), is to consider one-sided
hypothesis tests, but two-sided intervals. These are just two slightly different uses of the same pivot.
2.5
Inference after selection, revisited
We have portrayed selective pivotal inference, in sequential regression procedures, as a method for
producing conditional p-values and intervals. An unconditional interpretation of this framework is
also possible, which we describe here.
• For each model M ∈ M, a contrast vector v M ∈ Rn and pivot value µ M ∈ R are identified, so that
c = M. A
the hypothesis H0,M : vT
θ = µ M is to be tested whenever y ∈ Π M , i.e., whenever M(y)
M
TG statistic T ( · ; V ,U) is then defined, whose domain is the entire sample space Rn . Here we
write V = {v M : M ∈ M} and U = {µ M : M ∈ M} to denote the collection of contrast vectors and
pivot values, respectively, across partition elements—we will also refer to these as catalogs.
This unconditional TG statistic is defined by
X
T( · ; M, v M , µ M ) 1ΠM ( · ),
T ( · ; V ,U) =
M ∈M
where 1ΠM ( · ) denotes the indicator function for the partition element Π M (and T( · ; M, v M , µ M )
is as before, defined in (10)). The unconditional statistic can be used as follows: if a response Y
c ), v c , µ c ) to test the hypothesis
is drawn from (1), then we can form T (Y ; V ,U) = T(Y ; M(Y
M ( y) M ( y)
T
H0 : v c θ = µ M
c(Y ) .
M (Y )
• A concrete case to keep in mind is when V assigns a contrast vector v M to each model M, such
that vT
θ = β j M (A `M ), in the notation of (7), where M = {(A ` , s ` ) : ` = 1, . . . , k} as usual. This is
M
the j M th normalized coefficient from projecting θ onto X A `M , the active variables at step ` M .
• Assume that the errors in (1) are i.i.d. N(0, σ2 ). Then under the proper hypothesis, by summing
up the conditional property in (8) across partition elements, we have
³
´
PV T θ=U T (Y ; V ,U) ≤ t = t,
(12)
10
for all t ∈ [0, 1]. The assertion above holds for a parameter θ such that V T θ = U, which we use
as shorthand for vT
θ = µ M for all M ∈ M. Note that this full specification, across all M ∈ M,
M
is critical in order to apply the relevant null probability within each partition element (giving
rise to the equality in (12)).
• Therefore T (Y ; V ,U) serves as a valid p-value (with exact finite sample size)—but for testing
what null hypothesis? Formally, it is attached to H0 : V T θ = U, an exhaustive specification of
vT
θ = µ M , over all M ∈ M, but in truth, T (Y ; V ,U) carries no information about models other
M
c ). For this reason, we actually consider T (Y ; V ,U) to be a p-value for
than the selected one, M(Y
the random null hypothesis H0 : vTc θ = µ M
c(Y ) . This is made more precise through confidence
M (Y )
intervals.
• A confidence interval is obtained by inverting the test in (12). But the TG statistic at Y ,
X
¡
¢
c ), v c , µ c
T (Y ; V ,U) =
T(Y ; M, v M , µ M )1ΠM (Y ) = T Y ; M(Y
M (Y ) M (Y ) ,
M ∈M
only depends on U through µ M
c(Y ) . Thus, given a desired confidence level 1 − α, let us define D α
to be the set of U such that α/2 ≤ T (Y ; V ,U) ≤ 1 − α/2, and C α to be the set of µ M
c(Y ) such that
c ), v c , µ c ) ≤ 1 − α/2. Then we can see that
α/2 ≤ T(Y ; M(Y
M (Y ) M (Y )
U ∈ D α ⇐⇒ µ M
c(Y ) ∈ C α ,
c ), and
so the confidence interval is effectively infinite with respect to the values µ M , M 6= M(Y
inverting the test in (12) yields
³
´
P vTc θ ∈ C α = 1 − α.
(13)
M (Y )
The above expression says that the random interval C α traps the random parameter vTc θ
M (Y )
with probability 1 − α, and thus, this supports the interpretation of H0 : vTc θ = µ M
c(Y ) as the
M (Y )
null hypothesis underlying the unconditional TG statistic.
Remark 1. The pivotal property in (12) is derived under the distributional assumption that V T θ =
U, i.e., vT
θ = µ M for all M ∈ M, which may seem unnatural, as the catalog U of pivot value can be
M
large (e.g., on the order of d k after k steps of FS), and so this is condition on possibly many contrasts
of θ . However, it is worth emphasizing that the unconditional testing property in (12) is really only
useful in that it allows us to formulate the unconditional confidence interval property in (13), which
is a more natural statement about coverage of a single (random) parameter. When viewing selective
inference from an unconditional perpsective, we find it more natural to place the focus on confidence
intervals rather than hypothesis testing; in many ways, we find the former the more natural of the
two perspectives, unconditionally. Tibshirani et al. (2016) in fact suggest separate nomenclature for
the unconditional case, referring to the property in (13) as that of a selection interval (rather than
confidence interval), to emphasize that this interval covers a moving target.
3
The master statistic
c
Given a response y and predictors X , our description thus far of the selected model M(y),
statistics
T(y; M, v, µ) and T (y; V ,U), etc., has ignored the role of X . This was done for simplicity. The theory
to come in Section 4 will consider X to be nonrandom, but asymptotically X must (of course) grow
11
with n, and so it will help to be precise about the dependence of the selected model and statistics on
c , y), T(X , y; M, v, µ), and T (X , y; V ,U) to emphasize this
X . We will denote these quantities by M(X
dependence. We define
³1
´
1
Ωn =
XT X, p XT y ,
n
n
a d(d + 3)/2-dimensional quantity that we will call the master statistic. As its name might suggest,
this plays an important role: all normalized coefficients from regressing y onto subsets of the variables X can be written in terms of Ωn . That is, for an arbitrary set A ⊆ {1, . . . , p}, the jth normalized
coefficient from the regression of y onto X A is
T
T
T
T
y
X A )−1 p1n X A
(e Tj X A
X A )−1 X A
y e Tj n(X A
= q
,
q
T X )−1 e
T X )−1 e
e Tj (X A
e Tj n(X A
j
j
A
A
which only depends on (X , y) through Ωn . The same dependence is true, it turns out, for the selected
models from FS, LAR, and the lasso. We defer the proof of the next lemma, as with all proofs in this
paper, until the appendix.
Lemma 3. For each of the FS, LAR, and lasso procedures, run for k steps on data (X , y), the selected
c , y) only depends on (X , y) through Ωn = ( 1 X T X , p1 X T y), the master statistic.
model M(X
n
n
c , y) = M ⇐⇒ Q M (X ) y ≥ 0
In more detail, for any fixed M ∈ M, the matrix Q M (X ) such that M(X
1 T
1
T
p
can be written as Q M (X ) = P M ( n X X ) n X , where P M depends only on n1 X T X . Hence
³1
´ 1
c , y) = M ⇐⇒ P M X T X p X T y ≥ 0.
M(X
n
n
This lemma asserts that the master statistic governs model selection, as performed by FS, LAR,
c , y), the statistic
and the lasso. It is also central to TG pivot for these procedures. Denoting M = M(X
T(X , y; M, v, µ) in (10) only depends on (X , y) through three quantities:
vT y Q M (X ) v
,
, and Q M (X ) y.
kvk2
kvk2
The third quantity is always a function of Ωn , by Lemma 3. When v is chosen so that vT y is a normalized coefficient in the regression of y onto a subset of the variables in X , the first two quantities
are also functions of Ωn . Thus, in this case, the TG pivot only depends on (X , y) through the master
statistic Ωn ; in fact, it is continuous at any point such that n1 X T X is nonsingular and y does not lie
on the boundary of a model selection event.
Lemma 4. Fix any model M ∈ M, and suppose that v is chosen so that vT y is a normalized coefficient
from projecting y onto a subset of the variables in X . Then the TG statistic only depends on (X , y) by
means of Ωn , so that we may write
T(X , y; M, v, µ) = ψ M
³1
´
1
XT X, p XT y .
n
n
Further, the function ψ M is continous at any point (S, z) such that S is nonsingular and P M (S) z > 0.
Finally, we show that the conditional pivotal property of the TG statistic in (8) can be phrased
entirely in terms of the master statistic.
12
Lemma 5. Assume the conditions of Lemma 4, and additionally that Y is drawn from (1). Construct
the master statistic Ωn = ( n1 X T X , p1n X T Y ). Then there is a function g such that
vT θ = g(E(Ωn )).
Thus if the errors in (1) are i.i.d. N(0, σ2 ), then the conditional pivotal property (8) of the TG statistic
can be reexpressed as
¯
³
´
¯c
P g(E(Ωn ))=µ ψ M (Ωn ) ≤ t ¯ M(X
, Y ) = M = t,
for all t ∈ [0, 1].
Equipped with the last two lemmas, asymptotic theory for the TG test, when d is fixed, is not far
off. Under weak conditions on the data model in (1), the central limit theorem tells us that p1n X T Y
converges weakly to a normal random variable. With n1 X T X converging to a deterministic matrix,
the continuous mapping theorem will then provide the appropriate asymptotic limit for the statistic
T(X , y; M, v, µ) = ψ M ( n1 X T X , p1n X T Y ). This is made more precise next.
4
Asymptotic theory
Here we treat the dimension d as fixed, and consider the limiting distribution of the TG statistic as
n → ∞. (See Section 7 for the case when d grows.) Throughout, the matrix X ∈ Rn×d will be treated
as nonrandom, and we consider a sequence of predictor matrices satisfying two conditions:
lim
n→∞
1 T
X X = Σ,
n
(14)
for a nonsingular matrix Σ ∈ Rd ×d , and
lim max
n→∞ i =1,...,n
k x i k2
p = 0,
n
(15)
where x i ∈ Rd , i = 1, . . . , n denote the rows of X . These are not strong conditions.
4.1
A nonparametric family of distributions
We specify the class of distributions that we will be working with for Y in (1). Let σ2 > 0 be a fixed,
known constant. First we define a set of error distributions
½
¾
Z
Z
2
2
E = F : xdF(x) = 0, x dF(x) = σ .
The first moment condition in the above definition is needed to make the model identifiable, and the
second condition is used for simplicity. Aside from these moment conditions, the class E contains a
small neighborhood (say, as measured in the total variation metric) around essentially every element.
Thus, modulo the moment assumptions, E is strongly nonparametric in the sense of Donoho (1988).
Given µ ∈ R, let Fµ denote the distribution of µ + δ, where δ ∼ F, and given θ = (θ1 , . . . , θn ) ∈ Rn , let
F n (θ ) = Fθ1 × . . . , ×Fθn . Now we define a class of distributions
½
¾
P n (θ) = F n (θ) = Fθ1 × . . . × Fθn : F ∈ E .
(16)
13
In words, assigning a distribution Y ∼ F n (θ ) means that Y in drawn from the model (1), with mean
θ ∈ Rn , and errors ²1 , . . . , ²n i.i.d. from an arbitrary centered distribution F with variance σ2 .
As n grows, we allow the underlying mean θ to change, but we place a restriction on this parameter so that it has an appropriate asymptotic limit. Specifically, we consider a class Θ of sequences of
mean parameters such that p1n X T θ has an asymptotic limit lying in some compact set, with uniform
convergence to this limit. Formally, write (in a slight abuse of notation) θ ∈ Θ to denote a sequence of
mean parameters in Θ, and let E(Θ) denote the set of limit points of { p1n X T θ : θ ∈ Θ}. Then, for some
constant B > 0, we require of the class Θ,
¯
¯
¯
¯ 1 T
d
¯
E(Θ) ⊆ [−B, B] , and lim sup
sup ¯ p X θ − η¯¯ = 0.
(17)
n→∞ η∈E (Θ) 1 T
n
p X θ →η
n
We emphasize once again that θ ∈ Rn and X ∈ Rn× p will both vary with n, i.e., we can think of θ and
the columns of X as triangular arrays, but our notation suppresses this dependence for simplicity.
4.2
Uniform convergence results
We begin with a result on the uniform convergence of (the random part of) the master statistic to a
normal distribution, both marginally and conditionally.
Lemma 6. Assume that X has asymptotic covariance matrix Σ, as in (14), and satisfies the normalization condition in (15). Let Y ∼ F n (θ ) ∈ P n (θ ), this class as defined in (16), for a sequence of mean
parameters θ ∈ Θ, as defined in (17). Denote p1n X T θ → η as n → ∞. Then Z n = p1n X T Y converges in
distribution to Z ∼ N(η, σ2 Σ), uniformly over P n (θ ), and uniformly over all θ ∈ Θ. That is,
¯
¯
lim sup
sup
sup ¯P(Z n ≤ x) − P(Z ≤ x)¯ = 0.
n→∞ θ ∈Θ F (θ )∈P (θ )
x∈Rd
n
n
Further, given a sequence of matrices A n ∈ R q×d , n = 1, 2, 3, . . . with A n → A as n → ∞, such that the
set { z : Az ≥ 0} has nonempty interior, Z n | A n Z n ≥ 0 converges in distribution to Z | AZ ≥ 0, uniformly
over P n (θ ), and uniformly over all θ ∈ Θ.
This lemma, combined with Lemmas 4 and 5 of the last section, leads us to uniform asymptotic
theory for the TG test. We remind the reader that k, the number of steps, is to be considered fixed in
the next result (as it is throughout the paper).
Theorem 7. Assume the conditions of Lemma 6. Suppose FS, LAR, or the lasso is run for k steps on
(X , Y ). Below we describe the conditional and unconditional asymptotic results separately.
(a, Markovic) Fix any model M ∈ M. Let v be a vector such that vT θ gives a normalized coefficient in
the projection of θ onto some subset of the variables in X , and let µ be an arbitrary pivot value. Then
c , Y ) = M converges in distribution to
under vT θ = µ, the conditional TG statistic T(X , Y ; M, v, µ) | M(X
W ∼ U(0, 1), uniformly over P n (θ ), and over θ ∈ Θ. That is,
¯
¯
³
´ ¯¯
¯
¯c
lim sup
sup
, Y ) = M − t¯¯ = 0.
sup ¯¯PvT θ=µ T(X , Y ; M, v, µ) ≤ t ¯ M(X
n→∞
θ ∈Θ F n (θ )∈P n (θ ) t∈[0,1]
Moreover, if we define C n,α to be the set of µ such that α/2 ≤ T(X , Y ; M, v, µ) ≤ 1 − α/2, then C n,α is an
asymptotically uniformly valid confidence interval for vT θ . That is,
¯
¯
¯
³
´
¯
¯
¯c
lim sup
sup
sup ¯¯PvT θ=µ vT θ ∈ C n,α ¯ M(X
, Y ) = M − (1 − α)¯¯ = 0.
n→∞
θ ∈Θ F n (θ )∈P n (θ ) α∈[0,1]
14
(b) Let V = {v M : M ∈ M} be a catalog of vectors such that each vT
θ yields a normalized coefficient in
M
the projection of θ onto a subset of the variables in X , for M ∈ M, and U = {µ M : M ∈ M} be a catalog
of pivot values. Then under V T θ = U, the same results as in part (a) hold marginally. That is,
¯
³
´ ¯
¯
¯
lim sup
sup
sup ¯PV T θ=U T (X , Y ; V ,U) ≤ t − t¯ = 0.
n→∞ θ ∈Θ F (θ )∈P (θ ) t∈[0,1]
n
n
c , Y ), v c
and for C n,α defined to be the set of µ such that α/2 ≤ T(X , Y ; M(X
M ( X ,Y ) , µ) ≤ 1 − α/2,
¯ ³
¯
´
¯
¯
lim sup
sup
sup ¯P vTc
θ ∈ C n,α − (1 − α)¯ = 0.
M ( X ,Y )
n→∞ θ ∈Θ F (θ )∈P (θ ) α∈[0,1]
n
n
Remark 2. An initial version of this work contained only the unconditional result in part (b) of the
theorem. Jelena Markovic pointed out that the conditional result in part (a) should also be possible,
and thus this conditional result should also be attributed to her. Between the initial and the current
version of this paper, in addition to revising Theorem 7, we have also revised Theorems 11 and 12 to
include the appropriate conditional results.
5
Unknown σ2 and the bootstrap
The results of the previous section assumed that the error variance σ2 in the model (1) was known.
Here we consider two strategies when σ2 is unknown. The first plugs a (rather naive) estimate of σ2
into the usual TG statistic. The second is a computationally efficient bootstrap method. Both, as we
will show, yield asymptotically conservative p-values. (In practice, the bootstrap often gives shorter
confidence intervals than those based on the TG pivot; see Section 6.)
5.1
A simple plug-in approach
Given a model M ∈ M, contrast vector v, and pivot value µ, consider the TG statistic T(X , Y ; M, v, µ).
Let us abbreviate
b M = b(X , Y ; M, v),
abM = a(X , Y ; M, v), and b
where the latter two functions are as defined in Section 2.3. In this notation, we can succintly write
the TG statistic as
¶
µ T
¶
µb
v Y −µ
bM − µ
−Φ
Φ
σkvk2
σkvk2
(18)
T(X , Y ; M, v, µ) = µ
¶
µ
¶ .
bM − µ
b
abM − µ
Φ
−Φ
σkvk2
σkvk2
When σ2 is unknown, we propose a simple plug-in approach that replaces σ with cs Y , where
s2Y =
n
1X
| Y i − Y |2 ,
n i=1
P
the sample variance of Y (here Y = ni=1 Yi /n denotes the sample mean), and c > 1 is a fixed constant.
To be explicit, we consider the modified TG statistic
µ b
¶
µ T
¶
bM − µ
v Y −µ
Φ
−Φ
cs Y kvk2
cs Y kvk2
e
T(X , Y ; M, v, µ) = µ
(19)
¶
µ
¶.
b
bM − µ
abM − µ
Φ
−Φ
cs Y kvk2
cs Y kvk2
15
The scaling factor c facilitates our theoretical study of the above plug-in statistic, and practically, we
have found that ignoring it (i.e., setting c = 1) works perfectly well, though a choice of, say, c = 1.0001
seems to have a minor effect anyway.
When the mean θ of Y is nonzero, the sample variance s2Y is generally too large as an estimate of
σ2 . As we will show, the modified statistic in (19) thus yields asymptotically conservative p-values.
Residual based estimates of σ2 are not as useful in our setting because they depend more heavily on
the linearity of the underlying regression model, and they suffer practically when d is close to n (see
also the discussion at the start of Section 6).
5.2
An efficient bootstrap approach
As an alternative to the plug-in method of the last subsection, we investigate a highly efficient bootstrap scheme that does not rely on knowledge of σ2 . Our general framework so far treats X as fixed,
and for our bootstrap strategy to respect this assumption, we cannot use, say, the pairs bootstrap,
and must perform sampling with respect to Y only. The residual bootstrap is ruled out since we do
not assume that the mean θ follows a linear model in X . This leaves us to consider simple bootstrap
sampling of the components of Y . This is somewhat nonstandard, as the components of Y in (1) are
not i.i.d., but it provides a mechanism for provably conservative asymptotic inference, and it is what
makes our approach so computationally efficient.
Given Y = (Y1 , . . . , Yn ) drawn from the model in (1), let Y ∗ = (Y1∗ , . . . , Yn∗ ) denote a bootstrap sample
of Y . We will denote by P∗ the conditional distribution of Y ∗ on Y , and E∗ the associated expectation
operator. That is, P∗ (Y ∗ ∈ A) is shorthand for P(Y ∗ ∈ A |Y ), and similarly for E∗ . Using the notation
b M ), and assuming without a loss of generality that kvk2 = 1,
of the last subsection (notation for abM , b
let us motivate our bootstrap proposal by expressing the TG statistic as
¯
³
´
¯
bM , Y ,
T(X , Y ; M, v, µ) = P Zµ,σ2 ≥ vT Y ¯ abM ≤ Zµ,σ2 ≤ b
b M ) treated as fixed, and
where the probability on the right-hand side is taken with Y (and thus abM , b
2
with Zµ,σ2 denoting a N(µ, σ ) random variable. The main idea is now to approximate the truncated
normal distribution underlying the TG statistic with an appropriate one from bootstrap samples,
¯
¯
³
´
³
´
¯
b M , Y ≈ P∗ vT (Y ∗ − Y 1) + µ ≥ vT Y ¯¯ abM ≤ vT (Y ∗ − Y 1) + µ ≤ b
bM .
P Zµ,σ2 ≥ vT Y ¯ abM ≤ Zµ,σ2 ≤ b
P
Recall Y = ni=1 Yi /n is the sample mean of Y , so E∗ (vT Y ∗ ) = vT (Y 1) (with 1 ∈ Rn denoting the vector
of all 1s), and we have shifted vT Y ∗ so that the resulting quantity vT (Y ∗ − Y 1) + µ mimics a normal
variable with mean µ. The right-hand side above very nearly defines our bootstrap version of the TG
statistic, except that for technical reasons, we must make two small modifications. In particular, we
define the bootstrap TG statistic as
¢
¡
b M + δn
P∗ vT Y ≤ cvT (Y ∗ − Y 1) + µ ≤ b
∗
T (X , Y ; M, v, µ) =
,
(20)
¡
¢
b M + δn
P∗ abM ≤ cvT (Y ∗ − Y 1) + µ ≤ b
where c > 1 is a constant as before, and δn = γ n−1/4 for a small constant γ > 0. Again, we have found
that ignoring the scaling factor c (i.e., setting c = 1) works just fine in practice, though a choice like
c = 1.0001 does not cause major differences anyway. On the contrary, a nonzero choice of the padding
factor like δn = 10−4 n−1/4 does play an important practical role, since the bootstrap probabilities in
the numerator and denominator in (20) can sometimes be zero.
16
Lastly, it is worth emphasizing that practical estimation of the bootstrap probabilities appearing
in (20) is quite an easy computational task, because the regression procedure in question, be it FS,
LAR, or the lasso, need not be rerun beyond its initial run on the observed Y . After this initial run,
b M , and then draw, say, B = 1000 bootstrap samples Y ∗
we can just save the realized quantities abM , b
in order to estimate the probabilities in (20). This is not at all computationally expensive. Moreover,
to estimate (20) over multiple trial values of µ (so that we can invert these bootstrap p-values for a
bootstrap confidence interval), only a single common set of bootstrap samples is needed, since we can
just shift vT Y ∗ appropriately for each bootstrap sample Y ∗ .
5.3
Asymptotic theory for unknown σ2
Treating the dimension d as fixed, we will assume the previous limiting conditions (14), (15) on the
matrix X , and additionally, that
n
1X
k x i k32 = O(1).
(21)
n i=1
P
Note that (14) already implies that n1 ni=1 k x i k22 → tr(Σ), and the above is a little stronger, though it
is still not a strong condition by any means. For example, it is satisfied when max i=1,...,n k x i k2 = O(1).
These conditions on X imply important scaling properties for our usual choices of contrast vectors.
Lemma 8. Assume that X satisfies (14), (15), (21). If v is any vector such that vT θ gives a normalized
regression coefficient from projecting θ onto some subset of the variables in X , then
³ 1 ´
kvk33 = O p .
n
We specify assumptions on the distribution of Y in (1) that are similar to (but slightly stronger
than) those in Section 4.1. For constants σ2 , τ, κ > 0, we define a set of error distributions
½
¾
Z
Z
Z
Z
0
2
2
3
4
E = F : xdF(x) = 0, x dF(x) = σ , x dF(x) ≤ τ, x dF(x) ≤ κ .
We also define a class of distributions
P n0 (θ) =
½
F n (θ ) = F θ 1 × . . . × F θ n
¾
: F ∈E .
0
(22)
where as before, Fµ denotes the distribution of µ + δ, for δ ∼ F. We define a class Θ0 of sequences of
mean parameters that satisfies, as before,
¯
¯
¯ 1 T
¯
0
d
¯
E(Θ ) ⊆ [−B, B] , and lim sup
sup ¯ p X θ − η¯¯ = 0,
(23)
n→∞ η∈E (Θ0 ) 1 T
n
p X θ →η
n
for a constant B > 0, where recall E(Θ0 ) denotes the set of limit points in Θ0 ; also, for each θ ∈ Θ0 , at
each n, we require
n
n
1X
1X
s2θ =
|θ i − θ |2 ≤ S, and r 3θ =
|θ i − θ |3 ,
(24)
n i=1
n i=1
P
for constants S, R > 0, where θ = ni=1 θ i /n. Note that the assumptions Y ∼ F n (θ ), with F n (θ ) ∈ P n0 (θ )
and θ ∈ Θ0 , are not much stronger than our assumptions in Section 4.1: we require the existence of
two more moments for the error distribution, and place an additional weak condition on the growth
of (components of) θ . These conditions are sufficient to prove the following helpful lemma.
17
Lemma 9. Assume that X satisfies (14), (15). Let Y ∼ F n (θ ) ∈ P n0 (θ ), where this class is as defined in
(22), and let θ ∈ Θ0 , where this class is as in (23), (24). Then for any fixed M ∈ M, and c > 1,
¯
³
´
¯c
lim sup
sup
P cs Y ≥ σ ¯ M(X
, Y ) = M = 1.
n→∞ θ ∈Θ0 F (θ )∈P 0 (θ )
n
n
c , Y ) = M, uniformly over
In words, the event { cs Y ≥ σ} has probability tending to 1 conditional on M(X
0
0
P n (θ), and over θ ∈ Θ . Furthermore, denoting the sample third moment of Y as
r 3Y =
n
1X
| Y i − Y |3 ,
n i=1
we have that for any δ > 0, there exists C > 0 such that for sufficiently large n,
¯
µ 3
¶
¯
rY
¯
c
sup
sup
P 3 ≥ C ¯ M(X , Y ) = M ≤ δ,
sY
θ ∈Θ0 F n (θ )∈P n0 (θ )
c , Y ) = M, uniformly over P n0 (θ ), and over θ ∈ Θ0 .
In words, r 3Y /s3Y = OP (1) conditional on M(X
The last two lemmas allow us to tie the distribution function of our bootstrap contrast to that of
a normal random variable.
Lemma 10. Assume that X satisfies (14), (15), (21). Let Y ∼ F n (θ ) ∈ P n0 (θ ), as defined in (22), and let
θ ∈ Θ0 , as defined in (23), (24). Let M ∈ M, and let v be such that vT θ gives a normalized regression
coefficient from projecting θ onto a subset of the variables in X . Then for any δ > 0, there exists C > 0
such that sufficiently large n,
¯
¶
µ
¯ ¢¯
¯ ¡ T ∗
¢
¡
C ¯¯ c
¯
¯
¯
sup
sup
P sup P∗ v (Y − Y 1) ≤ t − P s Y Z ≤ t Y ≥ p ¯ M(X , Y ) = M ≤ δ,
n
t∈R
θ ∈Θ0 F n (θ )∈P n0 (θ )
where we use Z ∼ N(0, 1) for a standard normal random variate. In words, sup t∈R |P∗ (vT (Y ∗ − Y 1) ≤ t) −
p
c , Y ) = M, uniformly over P n0 (θ ), and over θ ∈ Θ0 .
P(s Y Z ≤ t | Y )| = OP (1/ n) conditional on M(X
We are now ready to present uniform asymptotic results for the plug-in and bootstrap TG statistics. We remind the reader the number of steps k is treated as fixed below (as it is throughout).
Theorem 11. Assume the conditions of Lemma 10. Suppose FS, LAR, or the lasso is run for k steps
e , Y ; M, v, 0) | M(X
c , Y ) = M and
on (X , Y ). Then under vT θ = 0, the conditional plug-in TG statistic T(X
∗
c
conditional bootstrap TG statistic T (X , Y ; M, v, 0) | M(X , Y ) = M are each asymptotically larger than
U(0, 1) in distribution, uniformly over P n0 (θ ), and over θ ∈ Θ0 . That is,
lim sup
sup
sup
¯
´ i
h
³
e , Y ; M, v, 0) ≤ t ¯¯ M(X
c , Y ) = M − t = 0,
PvT θ=0 T(X
lim sup
sup
sup
h
n→∞ θ ∈Θ0 F (θ )∈P 0 (θ ) t∈[0,1]
n
n
and
n→∞ θ ∈Θ0 F (θ )∈P 0 (θ ) t∈[0,1]
n
n
+
¯
³
´ i
¯c
, Y ) = M − t = 0,
PvT θ=0 T ∗ (X , Y ; M, v, 0) ≤ t ¯ M(X
+
where x+ = max{ x, 0} denotes the positive part of x. Further, given any catalog V = {µ M : M ∈ M} of
vectors such that each vT
θ yields a normalized coefficient in the projection of θ onto a subset of the
M
variables in X , for M ∈ M, the same results hold marginally under V T θ = 0.
18
Remark 3. For simplicity, we analyzed the plug-in and bootstrap statistics simultaneously. Consequently, the conditions assumed to prove asymptotic properties of the plug-in approach are stronger
than what we would need if we were to study this method on its own, but there are not major differences in these conditions.
Theorem 11 establishes that the plug-in and bootstrap versions of the TG statistic are asymptotically conservative when viewed as p-values under vT θ = 0. If we look more broadly at the distribution of these test statistics under vT θ = µ, for an arbitrary value of µ, then a technical barrier arises.
For each statistic, our proof of its asymptotic conservativeness leverages the fact that the truncated
Gaussian survival function decreases (in a pointwise sense), as its underlying variance parameter
decreases. To extend these results to the case of an arbitrary pivot value µ, we would need the analogous fact to hold when we replace the survival function of the Gaussian variate cs Y Z + µ truncated to
b M ], with that of σ Z + µ tuncated to [abM , b
b M ], on the event { cs Y ≥ σ}. Yet, without the guarantee
[abM , b
that abM ≥ µ (which clearly cannot always be true, for an arbitrary value of µ), it is no longer the case
that decreasing the variance from c2 s2Y to σ2 always decreases the survival functions of these two
truncated Gaussians; see Appendix A.11. This means that confidence intervals given by directly inverting either the plug-in or bootstrap TG statistic do not have provably correct asymptotic coverage
properties, under the current analysis.
From the arguments in the proof of Theorem 11, we can construct one-sided confidence intervals
with conversative asymptotic coverage, by forcing them to include abM . We do not pursue the details
here, as we have found that these one-sided intervals are practically too wide to be of interest.
Importantly, the plug-in and bootstrap TG statistics often display excellent empirical properties,
as we will show in the next section. A more refined analysis is needed to establish asymptotic uniformity for the distribution of these statistics under vT θ = µ. Such asymptotic uniformity, for arbitrary
µ, would lead to asymptotic coverage guarantees for confidence intervals produced by inverting these
statistics, and we leave this extension to future work.
6
Examples
We present empirical examples that support the theory developed in the previous sections, and also
suggest that there is much room to refine and expand our current set of results. The first two subsections examine a low-dimensional problem setting that is covered by our theory. The last two look at
substantial departures from this theoretical framework, the heteroskedastic and high-dimensional
settings, respectively. In all examples, the LAR algorithm was used for variable selection and associated inferences; results with the FS and lasso paths were roughly similar. Also, in all examples,
where not explicitly stated otherwise, the computed p-values are a test of whether the target population value is 0.
It may be worth discussing two potentially common reactions to our experimental setups, especially for the low-dimensional problems described in the next subsections. First, our plug-in statistic
uses s2Y as an estimate for σ2 ; why not use an estimate from the full least squares model of Y on
X , since this would be less conservative? While experiments (not shown) confirm that this works in
low-dimensional regression problems, such an estimate becomes anti-conservative as the number of
variables grows (particularly, irrelevant ones), and is obviously not applicable in high-dimensional
problems. Therefore, we stick with the simple estimate s2Y , as this is always applicable and always
conservative.
19
Second, to determine variable significance in a low-dimensional problem, one could of course fit a
full regression model and inspect the resulting p-values and confidence intervals. These p-values and
intervals could even be Bonferonni-adjusted to account for selection. Of course, this strategy would
not be possible for a high-dimensional problem, but if the number of predictors is small enough, then
it may work perfectly fine. So when should one use more complex tools for post-selection inference?
This is an important question, deserving of study, but it is not the topic of this paper. The examples
that follow are intended to portray the robustness of the selective pivotal inference method against
nonnormal error distributions; they are not meant to represent the ideal statistical practice in any
given scenario.
6.1
P-value examples
We begin by studying a low-dimensional setting with n = 50 and d = 10. We defined predictors X ∈
R50×10 , by drawing the columns independently according to the following mixture distribution: with
equal probability, a column was filled with i.i.d. entries from N(0, 1), Bern(0.5), or SN(0, 1, 5), where
SN(0, 1, 5) denotes the skew normal distribution (O’Hagan & Leonard 1976) with shape parameter
equal to 5. We then scaled the columns of X to have unit norm. The underlying mean was defined
as θ = X β0 , where β0 ∈ R10 has its first 2 components equal to −4 and 4, and the rest set to 0. Over
500 repetitions, we drew a response Y ∈ R50 from (1), with i.i.d. errors, and 4 different choices for the
error distribution: normal, Laplace, uniform, and skew normal. In each case, we centered the error
distribution, and we scaled it to have variance σ2 = 1 (for the skew normal distribution, we used a
shape parameter 5). Every 10 repetitions, the predictor matrix X was regenerated according to the
prescription described above.
Figure 3a displays QQ plots of p-values for testing the significance of the variable entered into
the active model, across 3 steps of LAR. (The QQ plots compare the p-values to a standard uniform
distribution.) The p-values were computed using the TG statistic with σ2 = 1, the plug-in TG statistic
with s2Y as its estimate for σ2 , and the bootstrap TG statistic with 50,000 bootstrap samples used
to approximate the probabilities in the numerator and denominator of (20), and padding factor δn =
10−4 n−1/4 . (The scaling factor was ignored, i.e., set to c = 1, for the plug-in and bootstrap statistics.)
In steps 1 and 2, the p-values are restricted to repetitions in which a correct variable selection was
made—i.e., variable 1 or 2 was entered into the active LAR model. In step 3, the p-values are from
repetitions in which an incorrect variable selection was made—i.e., one of variables 3 through 10
was entered into the active model. Since the underlying signal was fairly strong and the predictors
uncorrelated, such selections happened the majority of the time; specifically, the p-values displayed
for steps 1, 2, and 3 comprise approximately 95%, 85%, and 87% of the 500 repetitions, respectively.
The p-values in steps 1 and 2 show reasonable power, for all 3 statistics (TG, plug-in, and bootstrap
types), and all 4 error distributions. Also, the p-values in step 3 are uniform, as desired, again for
all statistics and all error distributions. Though the guarantees (for uniform null p-values) are only
asymptotic for the Laplace, uniform, and skew normal error distributions, such asymptotic behavior
appears to kick in quite early for these distributions, as the sample size here is only n = 50. Further,
the QQ plots reveal that the p-values for the nonnormal error distributions are not really any farther
from uniform than they are in the normal case. This is somewhat remarkable, recalling that the pvalues are, by construction, exactly uniform under normal errors.
Figure 3b inspects the TG statistic and plug-in and boostrap variants, when the pivot value µ is
set to the true population value. That is, we set µ = vT θ in computing the statistics in (18), (19), and
(20), in each data instance and each step of LAR. The figure collects the p-values across all 3 steps
20
Step 1, p-values
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
Observed
Normal errors
Laplace errors
1.0
TG
TG plug−in
Bootstrap
0.0
0.2
0.4
0.6
0.8
0.6
0.4
0.2
Expected
0.8
1.0
Skewed normal errors
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●●●●●
●●
●●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●●
●
●●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●●●
●
●
●
●●
●
●
●
●●●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●●●
●●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
1.0
0.8
0.6
Expected
TG
TG plug−in
Bootstrap
Observed
0.4
0.2
TG
TG plug−in
Bootstrap
Uniform errors
●
●●
●●
●
●
●
●●●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●●●●
●
●●
●
●
●
●
● ●●●
●
●
●
●●
●●●
●
●●●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●●
●●●
●●●
●
●
●●
●●
●●●
●
●
●
●●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●●
●
●●
●
●
●
●●●●
●
●●
●●
●
●
●●
●
●● ●
●●
●●●
●
●●●
●
●
●
●
●●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.2
0.4
Expected
0.6
0.8
1.0
Laplace errors
●
●
●
●
●●●
●
●
●
●●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●●
●●●
●●
●
●
●
●●●
●
●●●
●●
●●
●●
●●
●●
●
●●
●●●●
●●
●●●
●
●●●
●
●
●
●
●●●
●●
●
●
●●
●
●●●
●●
●
●●● ●
●●
●
●
●
●
●
●●
●
●●
●●
●●●
●●●●●
●
●
●
●●
●●
●●
●●●●● ●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.6
0.4
0.0
0.2
Expected
0.8
1.0
Normal errors
1.0
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●●
● ●●
●
●
●
●●
●
●
●●●
●
●
●●
●
●●
●●
●
●
●
●●●●
●
●
●
●●
●
●
● ●
●
●
●
●●
●
●
●●
●●
●
●
●●● ●
●
●●●
●
●●●
●
●
●
●
●●●● ●
●●
●
●●
●●
●●●
●
●●
●
●
●●●
●
●●
●
●
●●
●
●
●●●
● ●●●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
TG
TG plug−in
Bootstrap
0.0
0.2
0.4
0.6
0.8
Observed
Observed
Uniform errors
Skewed normal errors
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.8
0.6
Expected
TG
TG plug−in
Bootstrap
Observed
0.4
0.2
●
●
●●●
●
●●
●
●
●
●
●
●
●●●
●●●
●●
●
●
●●
●
●
●
●
●●
●●●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●●
●
●●
●●
●
●
●
●
●●
●
●
●●●
●
●●
● ●
●
●●
●
●● ●
●
●
●
●●
●●
●
●●●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
1.0
0.8
0.6
Expected
TG
TG plug−in
Bootstrap
Observed
0.4
0.2
TG
TG plug−in
Bootstrap
●
●
●
●
●●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●●
●
●●
●
●●●●
●
●
●
●●
●
●
●
●●●
●●●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●●
●●
●●●●
●
●
●
●
●●●●●
●
●
●
●●
●
●
●
●●
●●●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●●● ●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
1.0
0.8
0.6
Expected
0.4
0.2
●●
●●
●
●
●
●
●
●
●●
●
●●
●
●●
●● ●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●●
●●
●●
●
●
●
●
●
● ●●
●●
●●
●●
●●●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.6
0.4
0.0
0.2
Expected
0.8
1.0
Step 2, p-values
1.0
●
●
● ●●
●●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●●●
●
●●●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●●
●●
●
●
●●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
TG
TG plug−in
Bootstrap
0.0
0.2
0.4
0.6
0.8
Observed
Observed
Uniform errors
Skewed normal errors
1.0
0.2
0.4
0.6
0.8
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
0.8
0.6
Expected
TG
TG plug−in
Bootstrap
Observed
0.4
0.2
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
1.0
0.8
0.6
Expected
TG
TG plug−in
Bootstrap
Observed
0.4
0.2
0.6
1.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
TG
TG plug−in
Bootstrap
0.0
0.4
0.2
Expected
0.8
1.0
Laplace errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.6
0.4
0.0
0.2
Expected
0.8
1.0
Normal errors
1.0
Step 3, p-values
1.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
TG
TG plug−in
Bootstrap
0.0
0.2
Observed
0.4
0.6
0.8
1.0
Observed
(a) P-values are shown, after each of 3 steps of LAR.
All steps, pivotal statistics
0.0
0.2
0.4
0.6
Observed
0.8
1.0
0.0
0.2
0.4
0.6
Observed
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.8
0.6
Expected
0.4
0.2
TG
TG plug−in
Bootstrap
1.0
Skewed normal errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.8
0.6
Expected
0.4
0.2
TG
TG plug−in
Bootstrap
1.0
Uniform errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.8
0.6
Expected
0.4
0.2
TG
TG plug−in
Bootstrap
1.0
Laplace errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.6
0.4
0.0
0.2
Expected
0.8
1.0
Normal errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
TG
TG plug−in
Bootstrap
0.0
Observed
0.2
0.4
0.6
0.8
1.0
Observed
(b) Pivotal statistics are shown, aggregated over all 3 steps of LAR.
Figure 3: A simulation setup with n = 50 and d = 10, and a mean θ = X β0 , where β0 has 2 nonzero components.
of LAR, for each of the 4 error distribution types. According to our theory, the distribution of the TG
pivotal statistics here should be asymptotically uniform. This is clearly supported by the QQ plots.
Interestingly, both plug-in and bootstrap pivotal statistics also appear uniform in the QQ plots, and
yet, this is not a case handled by our asymptotic theory: recall, Theorem 11 fixes the pivot value µ to
be 0 (as, otherwise, technical difficulties are encountered in its proof). This gives empirical evidence
to the idea that a more refined analysis could extend Theorem 11 to the broader setting (of arbitrary
pivot values) handled by Theorem 7. Moreover, it suggests that inverting the plug-in and bootstrap
TG statistics should yield intervals with proper coverage, which is verified in the next subsection.
Lastly, we repeated all experiments in this subsection with the predictors X ∈ R50×10 generated
in such a way to induce a (population) correlation of 0.5 between all pairs of predictor variables. The
results are quite similar to those shown in Figure 3, and are hence deferred to Appendix A.12.
6.2
Confidence interval examples
We stay in same setting as the last subsection, so that n = 50, d = 10, and θ = X β0 for a coefficient
vector β0 with its first 2 components equal to −4 and 4, and the rest equal to 0. We invert the TG,
plug-in TG, and bootstrap TG statistics to obtain 90% confidence intervals at each LAR step. See Table 1 for a numerical summary. “Coverage” refers to the average fraction of intervals that contained
their respective targets over the 500 repetitions, “power” is the average fraction of intervals that
excluded zero, and “width” is the median interval width. These are all recorded in an unconditional
sense, i.e., no screening of repetitions was performed based on the variables that were selected across
the 3 steps of LAR (the conditional coverages however, were quite similar). From the table, we can
see that all 3 methods lead to accurate coverage (around 90%) in all cases. We can further see that
the intervals from the bootstrap TG statistic are shorter than those from the plug-in TG statistic in
all cases, and considerably shorter than both the plug-in and original TG statistics in steps 2 and 3.
The power from the bootstrap TG intervals is generally better than that from the plug-in TG intervals; also, it is on par with the power from the original TG statistic in step 1, but somewhat worse
in step 2. Recall that the original TG statistic uses knowledge of the error variance (σ2 = 1) but the
bootstrap and plug-in variants do not.
N
L
U
S
TG
Plug-in
Boot
TG
Plug-in
Boot
TG
Plug-in
Boot
TG
Plug-in
Boot
Coverage
0.914
0.928
0.932
0.904
0.944
0.944
0.912
0.928
0.924
0.892
0.940
0.936
Step 1
Power
0.508
0.378
0.528
0.568
0.410
0.566
0.538
0.396
0.540
0.540
0.402
0.520
Width
5.622
7.561
5.477
5.193
7.271
5.429
5.153
7.284
5.453
5.346
7.210
5.477
Coverage
0.890
0.914
0.916
0.926
0.930
0.944
0.902
0.910
0.910
0.878
0.896
0.912
Step 2
Power
0.520
0.404
0.424
0.536
0.440
0.454
0.504
0.390
0.422
0.504
0.380
0.394
Width
10.309
15.774
7.856
11.153
14.859
7.892
12.347
17.497
7.808
10.876
15.687
8.060
Coverage
0.910
0.918
0.930
0.912
0.904
0.924
0.894
0.886
0.892
0.906
0.910
0.918
Step 3
Power
0.114
0.100
0.090
0.118
0.120
0.108
0.128
0.126
0.118
0.116
0.106
0.102
Width
25.155
34.642
9.141
26.393
36.206
9.273
26.451
39.299
8.913
26.592
38.965
9.057
Table 1: Summary statistics for 90% confidence intervals constructed in the problem setting of Figure 3. The 4
blocks of rows correspond to the 4 types of noise: normal, Laplace, uniform, and skew normal, respectively. The
standard errors are about 0.01, 0.02, and 0.42 for the coverage, power, and width statistics, respectively.
22
It is a bit surprising that the bootstrap intervals can be shorter but still have worse power than
the original TG intervals. This is easier to understand once the intervals are visualized, as done in
Figure 4. The figure shows 100 sample intervals from the first LAR step, under normally distributed
errors. Sample intervals from the other error models are shown in Appendix A.13. We see that the
bootstrap TG intervals are indeed shorter, but compared to the original TG intervals, they are more
symmetric around the target population values. The original TG intervals, being more asymmetric,
are often shorter on the side (of the target value) facing 0, and this results in better power.
Again, we repeated the experiments here with the predictors X ∈ R50×10 generated to have pairwise correlation 0.5. Comparisons can be drawn between the results in a manner that roughly parallels the discussions following Table 1; however, on an absolute scale, all methods display a decrease
in power across the board (as correlated predictors clearly make the problem more difficult). Details
are provided in Appendix A.14.
6.3
Heteroskedastic errors
In the same setup as in Sections 6.1 and 6.2, with n = 50, d = 10, and the predictors X and mean θ
generated in the same manner, we consider a heteroskedastic model for Y by drawing ²0i , i = 1, . . . , n
i.i.d. from the given distribution—normal, Laplace, uniform, or skew normal—and then taking the
errors to be ² i = σ i ²0i , i = 1, . . . , n, where σ2i = 10k x i k22 , i = 1, . . . , n (and where x i ∈ Rd , i = 1, . . . , n denote
the rows of X .) The spread of error variances ended up being fairly substantial, from about 0.3 to
P
5.5. The original TG statistic was computed with σ2 = n1 ni=1 σ2i as a surrogate for the common error
variance; the plug-in and bootstrap variants were computed as usual. For brevity, we only plot the
pivotal statistics, aggregated over 3 steps of LAR, in Figure 5. (This is analogous to what is shown in
Figure 3b for the homoskedastic case. P-values at steps 1, 2, and 3, not shown, end up being similar
to those in Figure 3a, but the power from all methods is generally lower, due to the heteroskedastic
errors.) As we can see, the pivotal statistics in the figure look very close to uniformly distributed, as
desired. This is especially encouraging because the current problem setup lies outside of the scope
of our asymptotic theory (which assumes a constant error variance), and it suggests that our theory
could possibly be extended to accomodate errors with an (unknown) nonconstant variance structure.
6.4
High-dimensional examples
Finally, we consider a high-dimensional regime with n = 50 and d = 1000 predictors. The matrix X ∈
R50×1000 was generated according to the same recipe as before: each column, with equal probability,
was assigned i.i.d. entries from N(0, 1), Bern(0.5), or SN(0, 1, 5), and then scaled to have unit norm.
The mean was defined as θ = X β0 , where β0 ∈ R1000 has its first 2 components equal to -4 and 4,
and the rest 0. Over 500 repetitions, a response Y ∈ R50 was generated by adding normal, Laplace,
uniform, or skew normal noise to θ , with an error variance of σ2 = 1 (and every 10 repetitions, the
predictor matrix X was regenerated). Figure 6 plots the pivotal statistics aggregated over the first
3 steps of LAR. (This is as in Figure 3b for the low-dimensional case. P-values from the first 3 LAR
steps are omitted for brevity, and are roughly similar to those in Figure 3a, except that they display
less power, due to the high-dimensionality.) The pivotal statistics here look quite close to uniform, as
desired, and this is again encouraging, especially given that the current high-dimensional case lies
outside of the scope of our theory (which assumes that d is fixed). Further work on high-dimensional
asymptotic theory should be pursued (see also Tian & Taylor (2017)), though, as we show in the next
section, there is no hope for a uniform convergence result in high dimensions that holds as generally
23
5
●
●
●
●
●
●● ●
●● ●
●
●● ●
●●●
●●
●●● ● ●●
● ●●
●
0
●
●
●●
−5
●●● ●●
●
●
●
●
●●● ●●
● ●
●
● ●●
●●●●●
●
●
●● ●
●●
● ●
●
●● ●
●●●
● ●●●
●
●●●●● ●●
●●●
●● ●● ●● ●
● ●●
−10
Confidence interval
10
Normal errors, TG
0
20
40
60
80
100
Repetition
5
●
●
●
●
●
●● ●
●● ●
●
●● ●
●●●
●●
●●● ● ●●
● ●●
●
0
●
●
●●
−5
●●● ●●
●
●
●
●
●●● ●●
● ●
●
● ●●
●●●●●
●
●
●● ●
●●
● ●
●
●● ●
●●●
● ●●●
●
●●●●● ●●
●●●
●● ●● ●● ●
● ●●
−10
Confidence interval
10
Normal errors, TG plug−in
0
20
40
60
80
100
Repetition
5
●
●
●
●
●
●● ●
●● ●
●
●● ●
●●●
●●
●●● ● ●●
● ●●
●
0
●
●
●●
−5
●●● ●●
●
●
●
●
●●● ●●
● ●
●
● ●●
●●●●●
●
●
●● ●
●●
● ●
●
●● ●
●●●
● ●●●
●
●●●●● ●●
●●●
●● ●● ●● ●
● ●●
−10
Confidence interval
10
Normal errors, bootstrap
0
20
40
60
80
100
Repetition
Figure 4: Confidence intervals from 100 draws of Y from the same model as that in Figure 3. These intervals
are constructed from the first step of LAR, under a uniform distribution for noise. The colors are simply a visual
aid to mark the selection of different variables at step 1. The open circles denote the true population quantity to
be covered (here, the coefficient from projecting θ onto the first selected variable). Intervals that do not contain
their targets are drawn as dotted segments.
0.2
0.4
0.6
0.8
0.0
0.2
0.4
0.6
0.8
1.0
TG
TG plug−in
Bootstrap
0.0
0.2
Observed
0.4
0.6
0.8
0.6
0.4
0.2
Expected
0.8
1.0
Skewed normal errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
1.0
0.8
0.6
Expected
TG
TG plug−in
Bootstrap
Observed
0.4
0.2
0.6
1.0
Uniform errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
TG
TG plug−in
Bootstrap
0.0
0.4
0.2
Expected
0.8
1.0
Laplace errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.6
0.4
0.0
0.2
Expected
0.8
1.0
Normal errors
1.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
TG
TG plug−in
Bootstrap
0.0
0.2
Observed
0.4
0.6
0.8
1.0
Observed
Figure 5: A simulation setup with n = 50 and d = 10, but with heteroskedastic errors. Shown are the pivotal
statistics aggregated over 3 LAR steps.
0.0
0.2
0.4
0.6
Observed
0.8
1.0
0.2
0.4
0.6
0.8
1.0
TG
TG plug−in
Bootstrap
0.0
0.2
Observed
0.4
0.6
0.8
Observed
1.0
1.0
0.8
0.6
Expected
0.4
0.2
0.6
Skewed normal errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
TG
TG plug−in
Bootstrap
0.0
0.4
0.2
Expected
0.8
1.0
Uniform errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.8
0.6
Expected
0.4
0.2
TG
TG plug−in
Bootstrap
1.0
Laplace errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.6
0.4
0.0
0.2
Expected
0.8
1.0
Normal errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
TG
TG plug−in
Bootstrap
0.0
0.2
0.4
0.6
0.8
1.0
Observed
Figure 6: A simulation setup with n = 50 and d = 1000. Shown are the pivotal statistics over 3 LAR steps.
as the one we established in Theorem 7 for low dimensions.
7
A negative result in high dimensions
We prove that the TG statistic fails to converge to a uniform distribution, under the null hypothesis,
in a data model that has nonnormal errors and is high-dimensional, but otherwise represents a fairly
standard setting: the “many means” setting. We write the observation model as
Yi j = µ j + ² i j , i = 1, . . . , m, j = 1, . . . , d,
(25)
where we interpret i = 1, . . . , m as replications, and j = 1, . . . , d as dimensions. In total there are hence
n = md observations. Denote
m
1 X
Yj=
Yi j , j = 1, . . . , d.
m i=1
We will analyze the TG statistic, when selection is performed based on the largest of |Y j |, j = 1, . . . , d,
and inference is then performed on the corresponding mean parameter. A straightforward change of
notation will translate the above into a regression problem, with an orthogonal design X ∈ Rn×d , but
we stick with the many means formulation of the problem for simplicity.
We assume that the errors ² i j , i = 1, . . . , m, j = 1, . . . , d in (25) are i.i.d. from the following mixture:
π · N(−B, 1) + (1 − 2π) · N(0, 1) + π · N(B, 1).
25
(26)
The mixing proportion π and mean shift B will both scale with d. Moreover, they will be chosen so
that (for each d) the error variance is
σ2 = 1 + 2πB2 = 2.
As mentioned, we will consider model selection events of the form
c ) = ( j, s) ⇐⇒ sY j ≥ max |Y ` |.
M(Y
`6= j
We note that this is exactly the same selection event as that from the first step of FS, LAR, or lasso
paths, when run on the regression version of this problem with orthogonal design X . It is not hard
c ) = ( j, s), is
to check that the TG statistic for conditionally testing µ j = 0, given that M(Y
µp
¶
msY j
1−Φ
p
2
T(Y ; j, s, 0) =
p
µ
¶.
max`6= j m|Y ` |
1−Φ
p
2
(27)
As per the spirit of our paper, we can also view this statistic unconditionally; for this it is helpful to
define W1 = |Y 1 |, . . . ,Wd = |Y d |, and denote by W(1) ≥ . . . ≥ W(d ) the order statistics. Then from (27), we
can see that the unconditional TG statistic for testing the selected mean being 0 is
¶
µp
mW(1)
1−Φ
p
2
(28)
T (Y ; 0) =
µp
¶.
mW(2)
1−Φ
p
2
The framework underlying the TG statistic tells us that if the errors in (25) are i.i.d. N(0, 2), then for
c ) = ( j, s). Furany fixed model ( j, s), the pivot T(Y ; j, s, 0) is uniformly distributed conditional on M(Y
ther, if W(1) and W(2) are the largest and second largest absolute values of centered normal random
variables (each with variance 2/m), then the unconditional pivot T (Y ; 0) is again uniform. But when
W(1) ,W(2) are large, and are defined by the order statistics of nonnormal random variates, the statistic
T (Y ; 0)—which in this case is defined by the extreme tail behavior of the normal distribution—could
be nonuniform. The next theorem asserts that such nonuniformity does indeed happen asymptotically if we choose the mixture distribution in (26) appropriately.
Theorem 12. Assume the observation model (25), where the errors are all drawn i.i.d. from (26). Let
d and m scale in such a manner that (log d)/m → ∞. Further, let
s
µ ¶1/m
1
d 1/m
π=
, B=
,
d
2
so that the error variance is fixed at σ2 = 2. Then under the global null hypothesis, µ = 0, the unconditional TG statistic T (Y ; 0) in (28) does not converge in distribution to U(0, 1). In particular, on an
event whose limiting probability is at least 1/e, the statistic T (Y ; 0) converges to 0.
Further, the same results hold conditionally on any selected model. That is, for any fixed ( j, s), the
c ) = ( j, s) does not converge in distribution to U(0, 1), and on
conditional TG statistic T(Y ; j, s, 0) | M(Y
c ) = ( j, s)) at least 1/e, it converges to 0.
an event with limiting probability (conditional on M(Y
26
0.0
0.2
0.4
0.6
0.8
0.0
0.2
0.4
0.6
0.8
1.0
m=50000, d=2
●
●
●● ●●● ●●
●
●●● ●●
●●●●
●●
● ●●
●
●●
●●● ●
●●●
●
●
●
●
●
●● ●
●●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Expected
0.6
0.4
0.0
0.2
Expected
0.8
1.0
m=2, d=50000
1.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
Observed
0.2
0.4
0.6
0.8
1.0
Observed
Figure 7: The left plot shows a QQ plot of TG p-values, computed over 500 repetitions from the many means
setup exactly as described in Theorem 12, with d = 50, 000 and m = 2. We can see that the p-values are clearly
nonuniform, and 34% of the p-values are 0 (up to computer precision), close to the theoretically predicted proportion of 1/ e. The right plot shows p-values from the same model, but having reversed the roles of d and m (we
also had to cap π at 1/2); we can see that the p-values are essentially uniform.
Remark 4. The assumed condition (log d)/m → ∞ requires the dimension d to diverge to ∞, but not
necessarily the number of replications m, though it clearly allows m to diverge at a sufficiently slow
rate. On the other hand, if d were fixed and m diverged to ∞, then the result of the theorem would no
longer be true, and the limiting distribution of the TG p-value would revert to U(0, 1). (To be careful,
here we would have cap the mixing probability π at 1/2 in order for the mixture to make sense, since
the current definition of π diverges with d fixed and m tending to ∞.) In fact, this is ensured by our
low-dimensional result in Theorem 7: after reformulating the many means problem in appropriate
regression notation, all of the conditions of Theorem 7 are met by our current setup when d is fixed.
This is supported by the simulation in Figure 7.
Remark 5. The precise scaling (log d)/m → ∞ is chosen since this implies π = (1/d)1/m → 0, i.e., the
extreme mixture components N(−B, 1) and N(B, 1) each have probability tending to 0, an intuitively
reasonable property for the error distribution. But we note that this scaling is not important for any
other reason, and the proof would still remain correct if d/m → ∞.
Remark 6. In Theorem 3 of Tian & Taylor (2017), the authors show that the TG statistic converges
in distribution to a standard uniform random variable, in a high-dimensional problem setting, with
some restrictions on the sequences of selection events that are allowed. One might ask what part of
our high-dimensional setup here violates their conditions, because both results obviously cannot be
true simultaneously. As far as we can tell, the issue lies in the role of δn in Assumption 1 of Tian &
Taylor (2017). Namely, as we have defined the error distribution in (26), the value of δn needed to
certify the third condition Assumption 1 of their work is too small for the main assumption in their
Theorem 3 to hold. Hence Theorem 3 of Tian & Taylor (2017) does not apply to our current setup.
27
8
Discussion
We have studied the selective pivotal inference framework, with a focus on forward stepwise regression (FS), least angle regression (LAR), and the lasso, in regression problems with nonnormal errors.
We have shown that the truncated Gaussian (TG) pivot is asymptotically robust in low-dimensional
settings to departures from normality, in that it converges to a U(0, 1) distribution (its pivotal distribution under normality), and does so uniformly over a broad class of nonnormal error distributions.
When the error variance σ2 is unknown, we have proposed plug-in and bootstrap versions of the TG
statistic, both of which yield provably conservative asymptotic p-values.
Our numerical experiments revealed that the statistics under theoretical investigation generally
display excellent finite-sample performance, for highly nonnormal error distributions. These experiments also revealed findings not predicted by our theory: (i) the bootstrap TG statistic often produces
shorter confidence intervals than those based on the plug-in TG statistic, and even the TG statistic
that relies on the error variance σ2 ; and (ii) all three TG statistics show strong empirical properties
well-outside of the classic homoskedastic, fixed d regression setting that we presumed theoretically.
However, as we have clearly demonstrated, one should not hope for a convergence result in high
dimensions that is as general as the result obtained in low dimensions. In a relatively simple many
means problem, we showed the nonconvergence of the TG statistic to U(0, 1) as d → ∞, whereas in
the same problem but with d fixed, the TG statistic converges to its usual U(0, 1) limit.
There is still much left to do in terms of understanding the behavior of selective pivotal inference
tools that are constructed to have exact finite-sample guarantees under normality, like the TG statistic of Tibshirani et al. (2016), when applied in high-dimensional regression settings with nonnormal
data. When the pivot, the central cog of this framework, is constructed under the assumption of normality, this creates robustness issues that are especially worrisome in high dimensions. Appendix
A.16 provides a high-level discussion of some of these issues; a more detailed study will be the subject
of future research.
Acknolwedgements
We thank Jelena Markovic and Jonathan Taylor for many helpful discussions, and for their overall
generosity. An initial version of our work contained only unconditional (i.e., marginal) results in the
main theorems (Theorems 7, 11, and 12); Jelena Markovic pointed out that Theorem 7 should also
hold conditionally, and the current version of this work has been revised accordingly.
A
A.1
Appendix
Convex cones for FS, LAR, lasso
We describe a modification of the conic conditioning set in Tibshirani et al. (2016) for FS. Our version
is different in that we additionally condition on the sign of every active coefficient at every step, rather
than just the coefficient of the variable to enter the model at each step. The modifications needed for
the LAR and lasso conditioning sets, made on top of the sets for LAR and lasso given in Tibshirani
et al. (2016), will follow similarly to that described for FS, and hence we omit the details.
b` (y), ` = 1, . . . , k by a sorted
After k FS steps, we can always represent a sequence of active sets A
list of variables [ bj 1 (y), . . . , bj k (y)] that were chosen to enter the model at each step. Unfortunately, the
same cannot be done for a sequence of active signs sb` (y), ` = 1, . . . , k, because these do not obey such
28
a nested structure. We will write sb` (y) = [ sb`,1 (y), . . . sb`,` (y)] for the signs of coefficients corresponding
to the variables [ bj 1 (y), . . . , bj ` (y)], at the `th step.
b` (y) = A ` , sb` (y) = s ` , for ` = 1, . . . , k, using induction. At
Now we characterize the event that A
step ` = 1, we have that bj 1 (y) = j 1 and sb1 (y) = s 1 if and only if
s 1 X Tj1 y/k X j 1 k22 ≥ ± X Tj y/k X j k22 for all j 6= j 1 ,
or, rearranged,
¡
s 1 X j 1 /k X j 1 k22 ± X j /k X j k22
¢T
y ≥ 0 for all j 6= j 1 ,
b` (y) = A `
a set of 2(d − 1) linear inequalities in y. Assume that we have represented the event that A
and sb` (y) = s ` , for ` = 1, . . . , k −1, by a collection of linear inequalities in y. Then to represent bj k (y) = j k
and sbk (y) = [s k,1 , . . . , s k,k ], we must only append to this collection of inequalities. The former subevent
bj k (y) = j k is characterized by
¡
¢
e j /k X
e j k2 ± X
e j /k X
e j k2 T r ≥ 0 for all j 6= j 1 , . . . , j k ,
s k,k X
k
k 2
2
e j is the residual from regressing X j onto X A , and r is the residual from regression y onto
where X
k−1
e j = P ⊥ X j and r = P ⊥ X j , where P ⊥ projects onto the orthocomplement
X A k−1 . By expressing X
A k−1
A k−1
A k−1
of the column space of X A k−1 , we can rewrite the above constraints as
¡
¢
⊥
2
⊥
⊥
2 T
s k,k P ⊥
y ≥ 0 for all j 6= j 1 , . . . , j k ,
A k−1 X j k /kP A k−1 X j k k2 ± P A k−1 X j /kP A k−1 X j k2
a set of 2(d − k) linear inequalities in y. Meanwhile, the subevent sbk (y) = [s k,1 , . . . , s k,k ] can be characterized by k inequalities expressed in block form,
T
T
y ≥ 0.
diag(s 1,1 , . . . , s k,k ) (X A
X A k )−1 X A
k
k
This completes the proof.
A.2
Proof of Lemma 3
We prove the result for FS; the results for the LAR and lasso paths follows similarly, by inpsecting
the form of the linear inequalities that determine their selection events.
p
Consider the first FS step as described in Appendix A.1. Multiplying through by n, we see that
an equivalent set of inequalities that characterize the selection event bj 1 (y) = j 1 , sb1 (y) = s 1 is
s1
n
X Tj X j 1
1
T
X Tj y
n Xj y
1
p ± T
p ≥ 0 for all j 6= j 1 .
n
Xj Xj n
This is clearly of the desired form P1 ( n1 X T X ) p1n X T y ≥ 0, for a matrix P1 ( n1 X T X ) dependent only on
1 T
n X X . At the kth step of FS, there are two sets of inequalities to be examined: one that describes
the variable to enter bj k (y) = j k , and the second that describes the active signs sbk (y) = [s k,1 , . . . , s k,k ].
p
The first set, multiplying through by n, is
s k,k
nX Tj X A k−1 (X A T X A k−1 )−1
k
k−1
T
X Tj X A k−1 (X A T X A k−1 )−1 X A
k
k−1
k−1
X jk
T
T
nX Tj X A k−1 (X A T X A k−1 )−1
XA
y
XA
y
k−1
≥0
p ± T
p
T
−
1
n
X j X A k−1 (X A T X A k−1 ) X A X j n
k−1
k−1
for all j 6= j 1 , . . . , j k ,
29
while the second set, again multiplying through by
p
n, is
XT y
T
−1 A k
diag(s 1,1 , . . . , s k,k ) n(X A
X
)
p ≥ 0.
Ak
k
n
These inequalities are clearly all summarized by P k ( n1 X T X ) p1n X T y ≥ 0, where P k ( n1 X T X ) is a matrix that depends only on n1 X T X . This completes the proof.
A.3
Proof of Lemma 4
Under the conditions of the lemma, the TG pivot for fixed M in (8) depends only on X , y through the
master statistic, because, as explained above the lemma, the only dependence in the pivot on X , y is
through the quantities vT y/kvk2 , (Q M (X ) v)/kvk2 , Q M (X ) y, and each of these is in turn a function of
the master statistic Ωn . Moreover, we may reexpress the TG statistic in (10) as
¡
¢
¡
¢
Φ f 1 ( n1 X T X , p1n X T y) − Φ f 2 ( n1 X T X , p1n X T y)
T(X , y; M, v, µ) = ¡ 1
¢
¡
¢,
Φ f 1 ( n X T X , p1n X T y) − Φ f 3 ( n1 X T X , p1n X T y)
for some functions f 1 , f 2 , f 3 , or more succinctly, as T(X , y; M, v, µ) = ψ M ( n1 X T X , p1n X T y), where
¡
¢
¡
¢
Φ f 1 (S, z) − Φ f 2 (S, z)
¢
¡
¢.
ψ M (S, z) = ¡
Φ f 1 (S, z) − Φ f 3 (S, z)
Note that the quantities vT y/kvk2 , (Q M (X ) v)/kvk2 , Q M (X ) y depend smoothly on the master statistic
Ωn = ( n1 X T X , p1n X T y) at any point such that n1 X T X is nonsingular. This implies f 1 , f 2 , f 3 are smooth
functions of (S, z) at any point such that S is nonsingular. Lastly, for all S, z such that P M (S) z > 0,
we have f 1 (S, z) > f 3 (S, z), and thus the denominator of φ M (S, z) is positive. This proves the desired
continuity result on ψ M .
A.4
Proof of Lemma 5
For the master statistic Ωn = ( n1 X T X , p1n X T Y ), note that E(Ωn ) = ( n1 X T X , p1n X T θ ). As v is assumed
to be chosen such that vT θ is a normalized regression coefficient from the projection of θ onto some
subset of the columns in X , we may assume without a loss of generality that vT θ is as in (7) for some
A, j. Then, we see that we must only define
e Tj (S A,A )−1 z A
g(S, z) = q
,
e Tj (S A,A )−1 e j
where we use S A,A to denote the submatrix of S with rows in A and columns in A, and z A to denote
the subvector of z with entries in A.
A.5
Proof of Lemma 6
P
Define Z0,n = ni=1 ξ i , where ξ i = p1n x i ² i , x i is the ith row of X , and ² i = Yi − θ i , for i = 1, . . . , n. Note
that (ξ1 , . . . , ξn ) ∼ F n (0), with independent, mean zero components. We compute
n
X
i =1
Cov(ξ i ) =
n
σ2 X
n
i =1
30
x i xT
i =
σ2
n
XT X,
which converges to σ2 Σ as n → ∞, by assumption. Further, for any δ > 0, consider
µ
½
¾¶
´ 1X
n ³
n
X
k x i k2
2
2
2
E kξ i k2 · 1{kξ i k2 ≥ δ} =
k x i k2 E ² i · 1 p |² i | ≥ δ .
n i=1
n
i =1
P
We seek to show that this converges to 0 as n → ∞. As n1 ni=1 k x i k22 → tr(Σ), it suffices to show that
the maximum of the above expectations (in the summands) converges to 0, which is implied by the
p
assumption that max i=1,...,n k x i k2 / n → 0. As the above arguments did not depend on the sequence
F n (0), n = 1, 2, 3, . . ., we have verified the Lindeberg-Feller conditions uniformly, and hence the uniform Lindeberg-Feller central limit theorem, Lemma 2, implies that Z0,n converges in distribution
to Z0 ∼ N(0, σ2 Σ), uniformly over P n (0).
Now consider Z n = p1n X T Y = Z0,n + p1n X T θ . Writing Φ and φ for the standard normal CDF and
density,
sup
sup
¯
¯
sup ¯P(Z n ≤ x) − P(Z ≤ x)¯
θ ∈Θ F n (θ )∈P n (θ ) x∈Rd
= sup
sup
≤ sup
sup
≤ sup
sup
θ ∈Θ F n (θ )∈P n (θ )
θ ∈Θ F n (θ )∈P n (θ )
θ ∈Θ F n (θ )∈P n (θ )
|
¯
¯ µ
¶
¯
¯
1
sup ¯¯P Z0,n ≤ x − p X T θ − P(Z ≤ x)¯¯
n
x∈Rd
¯
µ
¶¯
¯
¯
¯
¯
1
¯
¯
sup P(Z0,n ≤ x) − P(Z0 ≤ x) + sup ¯¯Φ(x − η) − Φ x − p X T θ ¯¯
n
x∈Rd
x∈Rd
¯
¯
¯
¯
¯ ¯ 1
sup ¯P(Z0,n ≤ x) − P(Z0 ≤ x)¯ + ¯¯ p X T θ − η¯¯φ(0),
n
x∈Rd
{z
}
{z
} |
a
b
where the second line is due to the triangle inequality, and the third line is due to the simple bound
R x− t
|Φ(x − t) − Φ(x − s)| = | x−s φ(u) du| ≤ | t − s|φ(0), for any x, s, t. Note that a → 0 by the argument at the
start of this proof, and b → 0 by assumption in (17). This shows that Z n converges in distribution to
Z ∼ N(η, σ2 Σ), uniformly over P n (θ ), and over θ ∈ Θ.
Lastly, we establish the conditional result. By repeating the same arguments as above, the uniform Lindeberg-Feller central limit theorem and condition (17) imply that (Z n , A n Z n ) converges to
(Z, AZ), uniformly over P n (θ ), and over θ ∈ Θ. Thus, along sequence F n (θ ) ∈ P n (θ ), n = 1, 2, 3, . . . with
θ ∈ Θ, observe
P(Z n ≤ x, A n Z n ≥ 0)
P(Z ≤ x, AZ ≥ 0)
P(Z n ≤ x | A n Z n ≥ 0) =
→
,
P(A n Z n ≥ 0)
P(AZ ≥ 0)
at a rate that does not depend on the sequence in question. This is true because the numerator and
denominator each converge to their normal probability counterparts, and the denominator remains
bounded away from zero since { z : Az ≥ 0} has nonempty interior, and the set of limits of p1n X T θ was
assumed compact, in (17). Since x was arbitrary, and the distribution of Z | AZ ≥ 0 is continuous, we
have (e.g., Lemma 2.11 in van der Vaart (1998))
¯
¯
¯
¯
sup ¯P(Z n ≤ x | A n Z n ≥ 0) − P(Z ≤ x | AZ ≥ 0)¯ → 0.
x∈Rd
And as the sequence F n (θ ) ∈ P n (θ ), n = 1, 2, 3, . . . with θ ∈ Θ was arbitrary, we have shown the desired
uniform convergence.
31
A.6
Proof of Theorem 7
We begin with the proof of part (a). Let Z n = p1n X T Y and Z ∼ N(η, σ2 Σ). Also, let A n = P M ( n1 X T X )
c , Y ) = M, by Lemma 3. Also, Z n | A n Z n ≥ 0 converges
and A = P M (Σ). Recall that A n Z n ≥ 0 ⇐⇒ M(X
weakly to Z | AZ ≥ 0, uniformly over P n (θ ) and over θ ∈ Θ, by Lemma 6. As n1 X T X → Σ deterministically, we also have that Ωn = ( n1 X T X , Z n ) converges uniformly in distribution to Ω = (Σ, Z).
The choice of v as specified in the theorem is now important for two reasons. First, by Lemma 4,
we can express
T(X , Y ; M, v, µ) = ψ M (Ωn ),
for a function ψ M . Second, by Lemma 5, we can express vT θ = g(E(Ωn )) for a function g. Neither ψ M
nor g depend on n, and the distribution in question is that of ψ M (Ωn ) | A n Z n ≥ 0 under g(E(Ωn )) = µ.
The function ψ M is continuous at any point (S, z) such that S is nonsingular and Az > 0; recalling the
assumed nonsingularity of Σ, it is therefore continuous on a set of full probability under the limiting
distribution L(Ω | AZ ≥ 0). By the uniform continuous mapping theorem, Lemma 1, ψ M (Ωn ) | A n Z n ≥
0 converges uniformly to ψ(Ω) | AZ ≥ 0, which is distributed as U(0, 1) when g(E(Ω)) = µ by the pivotal
property of the TG statistic under normality, as in (8). The proof of uniform validity of TG confidence
intervals is just a rearrangement of the uniform asymptotic pivotal statement.
The proof of part (b) follows from the expansion
X
c , Y ) = M }.
T (X , Y ; V ,U) =
T(X , Y ; M, v M , µ M ) 1{ M(X
M ∈M
As the number possible models |M| is finite, we can simply apply the asymptotic pivotal result from
part (a) to each M ∈ M to establish the asymptotic pivotal property of T (X , Y ; V ,U). The confidence
interval result is again just a rearrangement of this pivotal property.
A.7
Proof of Lemma 8
By assumption, the vector v can be written as
T
X A (X A
X A )−1 e j
v= q
,
T X )−1 e
e Tj (X A
j
A
for some A, j. We compute
kvk33
Pn
=
=
T
−1
3
i =1 | X i,A (X A X A ) e j |
T X )−1 e |3/2
| e Tj (X A
j
A
1 Pn
T
| X i,A n(X A
X A )−1 e j |3
n3/2 i =1
.
T X )−1 e |3/2
| e Tj n(X A
j
A
The denominator converges to | e Tj (Σ A,A )−1 e j |3/2 by (14). The numerator satisfies
n
1 X
n3/2
n
1X
1
T
T
| X i,A n(X A
X A )−1 e j |3 ≤ p ·
k x i k32 · k n(X A
X A )−1 e j k32 ,
n
n
i =1
i =1
|
{z } |
{z
}
a
b
where a is bounded by (21) and b converges to k(Σ A,A )−1 e j k32 by (14). This completes the proof.
32
A.8
Proof of Lemma 9
We start by proving the result about the event { cs Y ≥ σ}. First let us study its asymptotic probability
marginally. Consider
n
1X
E| ² i + θ i − ² − θ | 2
n i=1
n
n
n
1X
1X
2X
E| ² i − ² | 2 +
|θ i − θ | 2 +
E(² i − ²)(θ i − θ )
=
n i=1
n i=1
n i=1
n−1 2
=
σ + s2θ ,
n
E(s2Y ) =
where ² =
Pn
i =1 ² i /n.
Hence
¡
¢
¡
¢
P cs Y ≤ σ = P c2 s2Y − c2 E(s2Y ) ≤ σ2 − c2 E(s2Y )
≤
c4 Var(s2Y )
(c2 E(s2Y ) − σ2 )2
,
where in the last line we used Chebyshev’s inequality. Recalling that c2 > 1, we have the lower bound
(c2 E(s2Y ) − σ2 )2 ≥ 0.999(c2 − 1)2 σ4 , for n large enough. Therefore, to show P(cs Y ≥ σ) → 1, it is enough
to show that Var(s2Y ) → 0 as n → ∞, uniformly. For this, we will use the simple inequality
Var(W1 + . . . + Wm ) ≤ m
m
X
Var(Wi ),
(29)
i =1
which follows from the fact that 2Cov(Wi ,W j ) ≤ Var(Wi ) + Var(W j ). We will also invoke Rosenthal’s
inequality (Rosenthal 1970), which for independent W1 , . . . ,Wm , having mean zero and E|Wi | t < ∞ for
i = 1, . . . , m, states that
¯
¯ m
½ m
µ m
¶ t/2 ¾
X
X
¯ X ¯t
t
2
¯
¯
E|Wi | ,
EWi
,
(30)
E¯ Wi ¯ ≤ C t max
i =1
i =1
i =1
for a constant C t > 0 only depending on t. Hence, observe that
¶
n
1X
2
|² i + θ i − ² − θ |
n i=1
µ n
¶
n
2X
1X
2
2
|² i + θ i − θ | + ² −
(² i + θ i − θ )²
= Var
n i=1
n i=1
µ n
¶
1X
2
2
|² i + θ i − θ | − ²
= Var
n i=1
µ n
¶
1X
2
≤ 2Var
|² i + θ i − θ | + 2Var(²2 ),
n i=1
|
{z
} | {z }
µ
Var(s2Y ) = Var
a
b
33
where in the last line we used (29). We consider a, b individually. We have
µ n
¶
n
n
X 2 X
X
2
a = 2 Var
²i +
² i (θ i − θ )
|θ i − θ |2 + 2
n
i =1
i =1
i =1
µ n ¶
µ n
¶
X 2
X
6
12
≤ 2 Var
² i + 2 Var
² i (θ i − θ )
n
n
i =1
i =1
6
12 2 2
≤ κ + σ s θ → 0,
n
n
where the second line again used (29), and the third used our assumptions on the error distribution
in (22), and on θ in (24). We also have
¯ n ¯4
¯
2 ¯¯ X
b = 4 E¯ ² i ¯¯
n
i =1
2
≤ 4 C 4 max{ nκ, n2 σ4 } → 0,
n
where the second line used Rosenthal’s inequality (30). This implies Var(s2Y ) ≤ a + b → 0, uniformly
over P n0 (θ ), and over θ ∈ Θ0 .
We have therefore shown P(cs Y ≥ σ) → 1, uniformly over P n0 (θ ), and over θ ∈ Θ0 . To see that the
c , Y ) = M, take any sequence F n (θ ) ∈ P n (θ ), n = 1, 2, 3, . . . where
same result holds conditional on M(X
0
θ ∈ Θ , and note that
¯
³
´ P(cs ≥ σ , A Z ≥ 0)
¯c
n n
Y
P cs Y ≥ σ ¯ M(X
,Y ) = M =
P(A n Z n ≥ 0)
P(A n Z n ≥ 0) − P(cs Y < σ)
≥
P(A n Z n ≥ 0)
P(AZ ≥ 0) − 0
→
= 1,
P(AZ ≥ 0)
where we have borrowed the notation and the normal convergence result P(A n Z n ≥ 0) → P(AZ ≥ 0)
from the proof of Lemma 5. The rate of convergence in the last line does not depend on the sequence
in consideration, because of the uniform convergence of A n Z n to AZ, and the fact the denominator
is bounded away from zero, since the set of limits of p1n X T θ is assumed to be compact, in (23). And
as F n (θ ) ∈ P n (θ ), n = 1, 2, 3, . . . with θ ∈ Θ0 was arbitrary, this completes the proof of the first part of
the lemma.
For the second part, on the boundedness of r 3Y /s3Y , consider that for any C > 0 we have
¯
µ 3
¶
¯
³
´
¯
rY
¯
c , Y ) = M ≥ P r 3 < σ3 C/c3 , s3 ≥ σ3 /c3 ¯¯ M(X
c ,Y ) = M
P 3 < C ¯ M(X
Y
Y
sY
¯
¯
´
³
´
³
¯c
¯c
≥ 1 − P r 3Y ≥ σ3 C/c3 ¯ M(X
, Y ) = M − P s3Y < σ3 /c3 ¯ M(X
,Y ) = M .
c , Y ) = M) → 0, uniformly, by what we showed above. It
The last term here satisfies P(s3Y < σ3 /c3 | M(X
c , Y ) = M) ≤ δ for
suffices to prove that, for any δ > 0, there exists C > 0 such that P(r 3Y > c3 C/σ3 | M(X
c , Y ) = M) is
large enough n, uniformly. By Markov’s inequality, this will be true as long as E(r 3Y | M(X
uniformly bounded. To this end, we will use the simple inequality,
| a + b | t ≤ 2 t | a| t + 2 t | b | t ,
34
(31)
and compute
E
³
¯
¯c
r 3Y ¯ M(X
,Y ) = M
´
¯
µ n
¶
¯
1 X
3¯ c
|² i + θ i − ² − θ | ¯ M(X , Y ) = M
= E
n i=1
¯
µ n
¶
¯
23 X
3¯ c
|² i − ²| ¯ M(X , Y ) = M + 23 r 3θ
≤ E
n i=1
¯
µ n
¶
¯
³
´
¯
26 X
¯c
3¯ c
|² i | ¯ M(X , Y ) = M + 26 E |²|3 ¯ M(X
≤ E
, Y ) = M + 23 r 3θ
n
i =1
≤ 26 τ M +
26
C 3 max{ nτ M , n2/3 σ3M } + 23 r 3θ ,
n3
(32)
where the second and third lines used (31), and the last line used Rosenthal’s inequality (30), along
with the abbreviations
¯
¯
µ n
µ n
¶
¶
¯
¯
1 X
1 X
3¯ c
2¯ c
2
τM = E
|² i | ¯ M(X , Y ) = M , and σ M = E
|² i | ¯ M(X , Y ) = M .
n i=1
n i=1
c , Y ) = M and the uniform convergence P(A n Z n ≥ 0) → P(AZ ≥ 0)
Once again using A n Z n ≥ 0 ⇐⇒ M(X
from Lemma 5, we have for large enough n,
τM ≤
τ
τ
E| ² 1 | 3
≤
≤
,
P(A n Z n ≥ 0) P(AZ ≥ 0)/2 ρ /2
where we have used the upper bound on the third moment of the error distribution in (22), and we
have used a lower bound P(AZ ≥ 0) ≥ ρ > 0 that holds uniformly over all θ ∈ Θ0 , due to the assumed
compactness of the set of limits of p1n X T θ , in (23). Thus we have shown that τ M is uniformly upper
bounded. Similar arguments show that σ M is uniformly upper bounded. As r 3θ ≤ R by assumption in
c , Y ) = M) is uniformly upper bounded. This completes the proof
(24), we see from (32) that E(r 3Y | M(X
of the second part, and the lemma.
A.9
Proof of Lemma 10
Let us write
n
vT (Y ∗ − Y 1) X
=
ξi ,
sY
i =1
P
where ξ1 , . . . , ξn are independent with mean zero and ni=1 Var∗ (ξ i ) = 1. By Theorem 3.7 of Chen et al.
(2011),
n
n
X
¯ ¡X
¯ ¢¯
¢
¡
sup ¯P∗
ξ i ≤ t − P Z ≤ t ¯ Y ¯ ≤ 10
E∗ | ξ i | 3 .
t∈R
i =1
i =1
But the right-hand side is precisely
10
n
X
E∗ |ξ i |3 = 10
i =1
r 3Y
s3Y
kvk33 .
p
c , Y ) = M, uniformly over P n0 (θ ), and
Lemmas 8 and 9 imply that this is OP (1/ n) conditional on M(X
0
over θ ∈ Θ , giving the result.
35
A.10
Proof of Theorem 11
First, we prove the result for the plug-in statistic. Denoting Z ∼ N(0, 1), we have
¯
³
´
bM , Y .
e , Y ; M, v, 0) = P cs Y Z ≥ vT Y ¯¯ abM ≤ cs Y Z ≤ b
T(X
c , Y ) = M, uniConsider the event { cs Y ≥ σ}, which has probability approaching 1 conditional on M(X
0
0
formly over P n (θ ), and over θ ∈ Θ , by Lemma 9. On this event, by the monotonicity of the truncated
Gaussian survival function in its variance parameter, shown in Appendix A.11, we can replace cs Y
by σ, and this cannot increase the value of the statistic. (To verify that the result in Appendix A.11
can indeed be applied, notice that abM ≥ 0, i.e., the left endpoint of the interval is at least the mean of
the truncated Gaussian, which follows from the fact that vT Y ≥ 0 by design.) Thus we can write
¯
³
´
bM , Y + E n,
e , Y ; M, v, 0) = P σ Z ≥ vT Y ¯¯ abM ≤ σ Z ≤ b
T(X
c , Y ) = M) → 0, uniformly over P n0 (θ ), and over θ ∈ Θ0 . Hence, for any t ∈ [0, 1],
where P(E n < 0 | M(X
¯
¯
³
´
³
´
e , Y ; M, v, 0) ≤ t ¯¯ M(X
c , Y ) = M ≤ PvT θ=0 T(X , Y ; M, v, 0) ≤ t ¯¯ M(X
c , Y ) = M + o(1),
PvT θ=0 T(X
where the o(1) remainder term above is uniform over t ∈ [0, 1], over P n0 (θ ), and over θ ∈ Θ0 . Applying
part (a) of Theorem 7 proves the conditional result for the plug-in statistic.
Next, we turn to the bootstrap result, whose proof is a little more involved. Define a function
¢
¡
b M + δn
©
ª
P∗ z ≤ cvT (Y ∗ − Y 1) ≤ b
∗
bM .
· 1 abM ≤ z ≤ b
G (z) = ¡
¢
b M + δn
P∗ abM ≤ cvT (Y ∗ − Y 1) ≤ b
Lemma 10 implies that we can write
¯ ¢
¡
b M ¯ Y + E n + δn
©
ª
P
z
≤
cs
Z
≤
b
Y
bM ,
G ∗ (z) = ¡
· 1 abM ≤ z ≤ b
¯ ¢
0
¯
b
P abM ≤ cs Y Z ≤ b M Y + E n + δn
p
c , Y ) = M, uniformly over z ∈ R, over P n0 (θ ), and over
where |E n |, |E 0n | = OP (1/ n) conditional on M(X
0
θ ∈ Θ . (Note that c in the above can be absorbed into the role of t in the lemma.) Dividing through
b M | Y ) + δn , we have
by the quantity P(abM ≤ cs Y Z ≤ b
¯ ¢
¡
b M ¯ Y + δn
P z ≤ cs Y Z ≤ b
En
+ ¡
¯ ¢
¯ ¢
¡
b M ¯ Y + δn P abM ≤ cs Y Z ≤ b
b M ¯ Y + δn ©
ª
P abM ≤ cs Y Z ≤ b
∗
bM
· 1 abM ≤ z ≤ b
G (z) =
0
En
1+ ¡
¯ ¢
b M ¯ Y + δn
P abM ≤ cs Y Z ≤ b
¯ ¢
¡
b M ¯ Y + δn
©
ª
P z ≤ cs Y Z ≤ b
bM + e n
= ¡
· 1 abM ≤ z ≤ b
¯ ¢
¯
b
P abM ≤ cs Y Z ≤ b M Y + δn
¯
³
´ ©
ª
¯
b M , Y · 1 abM ≤ z ≤ b
bM + e n
≥ P cs Y Z ≥ z ¯ abM ≤ cs Y Z ≤ b
¯
³
´ ©
ª
¯
b M , Y · 1 abM ≤ z ≤ b
bM + e n,
≥ P σ Z ≥ z ¯ abM ≤ σ Z ≤ b
c , Y ) = M, uniformly over z ∈ R, over P n0 (θ ), and over θ ∈ Θ0 , but
where | e n | = o P (1) conditional on M(X
the precise value of e n may differ from line to line. Above, in the second line, we used E n /δn = o P (1)
36
c , Y ) = M, uniformly, and similarly for E 0n ; in the third line, we used the fact that
conditional on M(X
(p + δ)/(q + δ) ≥ p/q for 0 < p ≤ q and δ ≥ 0; in the last line, we have used, as before, the monotonicity
of the truncated Gaussian survival function in its underlying variance parameter, and the fact that
c , Y ) = M) → 1, uniformly.
P(cs Y ≥ σ | M(X
Rewriting the result in the last display, we have
·
¸
¯
³
´ ©
ª
¯
∗
b
b
≤ | e n |,
sup G (z) − P σ Z ≥ z ¯ abM ≤ σ Z ≤ b M , Y · 1 abM ≤ z ≤ b M
z∈R
−
T
where x− = max{0, − x} denotes the negative part of x. In particular, at z = v Y , this implies
·
¸
T ∗ (X , Y ; M, v, 0) − T(X , Y ; M, v, 0) ≤ | e n |.
−
Finally, this means that we can write, at an arbitrary level t ∈ [0, 1],
¯
¯
³
´
³
´
¯c
¯c
PvT θ=0 T ∗ (X , Y ; M, v, 0) ≤ t ¯ M(X
, Y ) = M = PvT θ=0 T(X , Y ; M, v, 0) ≤ t − E 00n ¯ M(X
,Y ) = M ,
c , Y ) = M, uniformly over t ∈ [0, 1], over P n0 (θ ), and over θ ∈ Θ0 .
where (E 00n )− = o P (1) conditional on M(X
Therefore
¯
¯
³
´
³
´
¯c
¯c
PvT θ=0 T ∗ (X , Y ; M, v, 0) ≤ t ¯ M(X
, Y ) = M ≤ PvT θ=0 T(X , Y ; M, v, 0) ≤ t ¯ M(X
, Y ) = M + o(1),
where the o(1) term above is uniform over t ∈ [0, 1], over P n0 (θ ), and over θ ∈ Θ0 . Applying part (a) of
Theorem 7 proves the conditional result for bootstrap statistic.
The unconditional results for two modified TG statistics hold simply by marginalization.
A.11
Monotonicity of the truncated Gaussian distribution in σ2
Define
[a,b]
F 0,σ2 (x) =
Φ(b/σ) − Φ(x/σ)
,
Φ(b/σ) − Φ(a/σ)
the survival function for a normal random variable Z ∼ N(0, σ2 ), truncated to lie in an interval [a, b],
where a ≥ 0. We will show, following the proof of a similar monotonicity result in Lemma A.1 of Lee
et al. (2016), that for any 0 < σ21 < σ22 ,
[a,b]
[a,b]
F 0,σ2 (x) < F 0,σ2 (x) for all x ∈ [a, b].
1
2
To emphasize, the above property is only true when the interval [a, b] lies to the right of 0. Without
this restriction, the survival function will not be monotone increasing in σ2 (if [a, b] contains 0, then
it will generally be nonmonotone, and if [a, b] lies to the left of 0, then it will actually be monotone
decreasing).
[a,b]
Over σ2 > 0, the family of distributions F 0,σ2 forms an exponential family with natural parameter
1/σ2 , as it is just a family of Gaussian distributions with the carrier measure changed. Therefore, it
[a,b]
has a monotone likelihood ratio in its sufficient statistic − x2 , i.e., if we denote by f 0,σ2 the truncated
Gaussian density function, and we fix σ21 < σ22 , and a ≤ x1 < x2 ≤ b, then
f
[a,b]
(x2 )
0,σ21
f
[a,b]
(x2 )
0,σ22
<
37
f
[a,b]
(x1 )
0,σ21
f
[a,b]
(x1 )
0,σ22
.
Hence
f
[a,b]
[a,b]
[a,b]
[a,b]
(x2 ) f 2 (x1 ) < f 2 (x1 ) f 2 (x2 ),
0,σ21
0, σ 2
0, σ 1
0, σ 2
Integrating with respect to x1 over [a, x), for some x < x2 , we obtain
³
´ ³
´
[a,b]
[a,b]
[a,b]
[a,b]
f 2 (x2 ) 1 − F 0,σ2 (x) < 1 − F 0,σ2 (x) f 2 (x2 ).
0,σ1
2
1
0, σ 2
Now integrating with respect to x2 , over (x, b], we obtain
³
´ ³
´ [a,b]
[a,b]
[a,b]
[a,b]
F 0,σ2 (x) 1 − F 0,σ2 (x) < 1 − F 0,σ2 (x) F 0,σ2 (x).
1
2
1
2
Rearranging gives the result.
A.12
P-value examples for correlated predictors
Here we investigate the consequences of using correlated predictors in the simulation setup of Section 6.1. We constructed a preliminary matrix X ∈ R50×10 as before: each column was drawn independently to have either i.i.d. N(0, 1), Bern(0.5), or SN(0, 1, 5) entries, with equal probability. We then
took as our predictor matrix X 0 = X Σ1/2 , where Σ ∈ R10×10 has all diagonal entries equal to 1 and all
off-diagonal entries equal to 0.5 (and Σ1/2 is its symmetric square root). We scaled the columns of X 0
to have unit norm. The rest of the setup is then just as in Section 6.1.
Figure 8 shows the results, in the same format as Figure 3: p-values for LAR steps 1, 2, and 3,
and pivotal statistics aggregated over LAR steps, from 500 repetitions. The p-values at steps 1 and
2 were restricted to repetitions in which either variable 1 or 2 were selected (now comprising about
70% and 60% of the repetitions, respectively); the p-values at step 3 were restricted to repetitions in
which one of variables 3 through 10 was selected (comprising about 80% of the repetitions). Similar
to the display in Figure 3, we see power in the p-values from steps 1 and 2, albeit less power than in
the uncorrelated case, and uniform p-values in step 3, as well as uniform pivotal statistics.
A.13
Confidence intervals for uniform, Laplace, and skew normal noise
Figures 9 through 11 show sample confidence intervals for the problem setting of Section 6.2, when
the error distribution is uniform, Laplace, and skew normal, respectively.
A.14
Confidence interval summary statistics for correlated predictors
Table 2 gives summary statistics of confidence intervals obtained by inverting the original TG, plugin TG, and bootstrap TG statistics, as in Table 1 of Section 6.2, but for the correlated predictors setup
described in Section A.12.
A.15
Proof of Theorem 12
Let us denote by N j the number of observations in the jth column of the data array Yi j , i = 1, . . . , m,
j = 1, . . . , d that are drawn from the N(B, 1) mixture component. Similarly, let N 0j denote the number
of observations in the jth column drawn from the N(0, 1) mixture component. Then we will define E
to be the event
n
o
E = For some j = 1, . . . , d, we have N j = m and N`0 ≥ m − 2π md for all ` 6= j .
38
Step 1, p-values
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
Observed
Normal errors
Laplace errors
1.0
TG
TG plug−in
Bootstrap
0.0
0.2
0.4
0.6
0.8
0.6
0.4
0.2
Expected
0.8
1.0
Skewed normal errors
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●
●
●
●●
●
●●●●
●
●
●
●●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
1.0
0.8
0.6
Expected
TG
TG plug−in
Bootstrap
Observed
0.4
0.2
TG
TG plug−in
Bootstrap
Uniform errors
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.2
0.4
Expected
0.6
0.8
1.0
Laplace errors
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●●
●
●
●
●
●
●●● ●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.6
0.4
0.0
0.2
Expected
0.8
1.0
Normal errors
1.0
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
TG
TG plug−in
Bootstrap
0.0
0.2
0.4
0.6
0.8
Observed
Observed
Uniform errors
Skewed normal errors
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.8
0.6
Expected
TG
TG plug−in
Bootstrap
Observed
0.4
0.2
●
●
●
●
●
●
●
●●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●●
●
●●
●●
●
●●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
1.0
0.8
0.6
Expected
TG
TG plug−in
Bootstrap
Observed
0.4
0.2
TG
TG plug−in
Bootstrap
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
1.0
0.8
0.6
Expected
0.4
0.2
●●
●●
●
●●
●
●●
●●
●
●
●●●
●●
●
●
●●
●●
●
●
●
●●
●●
●
●
●
●●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.6
0.4
0.0
0.2
Expected
0.8
1.0
Step 2, p-values
1.0
●
●
●●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●●●
●
●●
●
●●
●
●
●
●
●●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
TG
TG plug−in
Bootstrap
0.0
0.2
0.4
0.6
0.8
Observed
Observed
Uniform errors
Skewed normal errors
1.0
0.2
0.4
0.6
0.8
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
0.8
0.6
Expected
TG
TG plug−in
Bootstrap
Observed
0.4
0.2
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
1.0
0.8
0.6
Expected
TG
TG plug−in
Bootstrap
Observed
0.4
0.2
0.6
1.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
TG
TG plug−in
Bootstrap
0.0
0.4
0.2
Expected
0.8
1.0
Laplace errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.6
0.4
0.0
0.2
Expected
0.8
1.0
Normal errors
1.0
Step 3, p-values
1.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
TG
TG plug−in
Bootstrap
0.0
0.2
Observed
0.4
0.6
0.8
1.0
Observed
(a) P-values are shown, after each of 3 steps of LAR.
All steps, pivotal statistics
0.0
0.2
0.4
0.6
Observed
0.8
1.0
0.0
0.2
0.4
0.6
Observed
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
0.8
0.6
Expected
0.4
0.2
TG
TG plug−in
Bootstrap
1.0
Skewed normal errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.8
0.6
Expected
0.4
0.2
TG
TG plug−in
Bootstrap
1.0
Uniform errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.8
0.6
Expected
0.4
0.2
TG
TG plug−in
Bootstrap
1.0
Laplace errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.6
0.4
0.0
0.2
Expected
0.8
1.0
Normal errors
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
TG
TG plug−in
Bootstrap
0.0
Observed
0.2
0.4
0.6
0.8
1.0
Observed
(b) Pivotal statistics are shown, aggregated over all 3 steps of LAR.
Figure 8: QQ plots as in Figure 3, but in a setup where the predictor variables have pairwise correlation 0.5.
5
● ●●
●
●●●●
●●●●
●●
●
●●●●●●●●●● ●
● ●● ●●
●● ●●
●
●●●
● ●●●●
●
●
●● ●●● ●
−5
●● ●
●●●
●
●●●
●●●●●
●
●
●
0
●●
●
●
●
● ●
●●
●
●●
●●●
●●●
●●
●●
●
●●
●
●●
●●
−10
Confidence interval
10
Laplace errors, TG
0
20
40
60
80
100
Repetition
5
● ●●
●
●●●●
●●●●
●●
●
●●●●●●●●●● ●
● ●● ●●
●● ●●
●
●●●
● ●●●●
●
●
●● ●●● ●
−5
●● ●
●●●
●
●●●
●●●●●
●
●
●
0
●●
●
●
●
● ●
●●
●
●●
●●●
●●●
●●
●●
●
●●
●
●●
●●
−10
Confidence interval
10
Laplace errors, TG plug−in
0
20
40
60
80
100
Repetition
5
● ●●
●
●●●●
●●●●
●●
●
●●●●●●●●●● ●
● ●● ●●
●● ●●
●
●●●
● ●●●●
●
●
●● ●●● ●
−5
●● ●
●●●
●
●●●
●●●●●
●
●
●
0
●●
●
● ●
●
●
●●
●
●●
●●●
●●●
●●
●
●●
●●
●
●●
●●
−10
Confidence interval
10
Laplace errors, bootstrap
0
20
40
60
80
100
Repetition
Figure 9: Confidence intervals from 100 draws of Y , similar to those in Figure 4, but under a uniform noise
distribution.
5
●● ●
●●
●
● ● ●
● ●
● ●
●
●
●●
●●●●●
●
●●
● ●
●●●
●
●●●●
●●●
0
●●●
●
●●●●●●
−5
●
● ●● ●●
●
●
● ●●●
● ● ●●
●●● ●●
●
●
●●●
●
●● ● ●●
●●●●
●● ●●
●
● ● ● ●●●●
●● ● ●
−10
Confidence interval
10
Uniform errors, TG
0
20
40
60
80
100
Repetition
5
●● ●
●●
●
● ● ●
● ●
● ●
●
●
●●
●●●●●
●
●●
● ●
●●●
●
●●●●
●●●
0
●●●
●
●●●●●●
−5
●
● ●● ●●
●
●
● ●●●
● ● ●●
●●● ●●
●
●
●●●
●
●● ● ●●
●●●●
●● ●●
●
● ● ● ●●●●
●● ● ●
−10
Confidence interval
10
Uniform errors, TG plug−in
0
20
40
60
80
100
Repetition
5
●● ●
●●
●
● ● ●
● ●
● ●
●
●
●●
●●●●●
●
●●
● ●
●●●
●
●●●●
●●●
0
●●●
●
●●●●●●
−5
●
● ●● ●●
●
●
● ●●●
●●● ●●
● ● ●●
●
●●●
●
●● ● ●●
●
●●●●
●● ●●
●
● ● ● ●●●●
●● ● ●
−10
Confidence interval
10
Uniform errors, bootstrap
0
20
40
60
80
100
Repetition
Figure 10: Confidence intervals from 100 draws of Y , similar to those in Figure 4, but under a Laplace noise
distribution.
5
●
●
●● ●
●
●●●
●●●●●
●
●●
●●●●●
●
● ●●
●
●
●
●
●
●●●●
●● ●
● ●●
●
0
●
●
●●
−5
● ●● ● ●●
●●
●●● ●● ●●
●●
●● ●
●
●●
● ● ●●
●● ●● ●●●
●●
●●●● ●●
● ●●
●● ●
● ●●●●
●
−10
Confidence interval
10
Skewed normal errors, TG
0
20
40
60
80
100
Repetition
5
●
●
●● ●
●
●●●
●●●●●
●
●●
●●●●●
●
● ●●
●
●
●
●
●
●●●●
●● ●
● ●●
●
0
●
●
●●
−5
● ●● ● ●●
●●
●●● ●● ●●
●●
●● ●
●
●●
● ● ●●
●● ●● ●●●
●●
●●●● ●●
● ●●
●● ●
● ●●●●
●
−10
Confidence interval
10
Skewed normal errors, TG plug−in
0
20
40
60
80
100
Repetition
5
●
●
●● ●
●
●●●
●●●●●
●
●●
●●●●●
●
● ●●
●
●
●
●
●
●●●●
●● ●
● ●●
●
0
●
●
●●
−5
● ●● ● ●●
●●
●●● ●● ●●
●●
●● ●
●
●●
● ● ●●
●● ●● ●●●
●●
●●●● ●●
● ●●
●● ●
● ●●●●
●
−10
Confidence interval
10
Skewed normal errors, bootstrap
0
20
40
60
80
100
Repetition
Figure 11: Confidence intervals from 100 draws of Y , similar to those in Figure 4, but under a skew normal
noise distribution.
N
L
U
S
TG
Plug-in
Boot
TG
Plug-in
Boot
TG
Plug-in
Boot
TG
Plug-in
Boot
Coverage
0.908
0.926
0.924
0.912
0.928
0.934
0.910
0.926
0.918
0.904
0.912
0.908
Step 1
Power
0.220
0.186
0.192
0.264
0.182
0.176
0.226
0.154
0.172
0.240
0.174
0.192
Width
6.907
8.186
4.973
6.510
7.341
5.117
6.826
8.192
4.949
6.479
7.717
4.973
Coverage
0.920
0.922
0.916
0.886
0.894
0.916
0.898
0.906
0.886
0.910
0.920
0.904
Step 2
Power
0.244
0.210
0.254
0.290
0.264
0.294
0.262
0.200
0.280
0.262
0.218
0.254
Width
25.960
30.113
8.745
23.668
26.841
8.769
25.371
29.211
8.817
24.502
28.979
8.697
Coverage
0.904
0.908
0.914
0.894
0.894
0.884
0.920
0.922
0.910
0.892
0.894
0.896
Step 3
Power
0.110
0.106
0.116
0.126
0.130
0.148
0.106
0.098
0.122
0.136
0.120
0.122
Width
55.614
66.083
10.667
54.351
60.831
10.583
52.786
63.915
10.474
56.700
68.143
10.486
Table 2: Summary statistics for 90% confidence intervals, as in Table 1, but in a modified problem setting such
that the predictor variables have pairwise correlation 0.5. The standard errors are roughly 0.01, 0.02, and 0.87
for the coverage, power, and width statistics, respectively.
In words, E is the event that exactly one column has all of its observations drawn from N(B, 1), and
each of the rest of the d − 1 columns have at least m − 2π md observations from N(0, 1). We calculate
¡
¢d −1
P(E) = d πm P N10 ≥ m − 2π md
³
¡
¢´d −1
e 1 ≥ 2π md
= 1 − P N10 + N
µ
¶d −1
1
≥ 1−
d
→ 1/e,
e j for
where in the second line we used that d πm = 1 by construction, and introduced the notation N
the number of observations in column j that are drawn from the N(−B, 1) mixture component; in the
third line we used Markov’s inequality.
On the event E, intersected with an event whose probability tends to one, we have W(1) ,W(2) → ∞,
and furthermore
p
p
mW(1) ≥
p
mB + Z0 ≥
mW(2) ≤ 2π m3/2 dB +
p
mB/2,
max
j =1,...,d −1
Z j ≤ 4π m3/2 dB,
where Z0 , Z1 , . . . , Z d −1 denote standard normals. We note that the ultimate bounds on the right-hand
sides in the two lines above are extremely loose, but will suffice for our purposes. Hence using Mills’
ratio, we can bound the TG statistic on the event in consideration by
µ ¡ mW 2 − mW 2 ¢ ¶
µ
¶
2
(1)
(2) W(2)
T (Y ; 0) ≤ exp −
1+
2
4
W(1)
mW(2)
µ ¡ mW 2 − mW 2 ¢ ¶
(1)
(2)
≤ 2 exp −
,
4
43
0.3
0.4
0.2
0.3
0.1
0.2
0.0
0.1
0.0
−10
−5
0
5
10
2
4
6
8
10
Figure 12: The left plot shows two densities p, q, in black and red; the right shows their tail functions H p , H q
(in corresponding colors).
for sufficiently large d. But on this same event we have that
¶
µ
1
2
2
mW(1)
− mW(2)
≥ mB2 − 16π2 m2 d 2 ,
4
and it is straightforward to check that the right-hand side of the bound above diverges to ∞, given
our assumptions on m, d, π, B. Therefore, we have shown that on an event whose probability tends
to at least 1/e, the TG statistic converges to 0.
As for the conditional result, notice that for any model ( j, s), we have by symmetry (under µ = 0)
c ) = ( j, s)) = P(T (Y ; 0) ≤ t), as well as P(E | M(Y
c ) = ( j, s)) = P(E). Hence the condiP(T(Y ; j, s, 0) ≤ t | M(Y
c ) = ( j, s) itself cannot be asymptotically uniform, and converges to
tional TG statistic T(Y ; j, s, 0) | M(Y
c ) = ( j, s).
0 on a event whose limiting probability is at least 1/e, conditional on M(Y
A.16
Some thoughts on instability in high dimensions
The TG statistic is defined by the ratio of normal tail probabilities. If the dimension d is large (in
which case we are searching through a large space of models), or there are some large effects, then
we often find ourselves evaluating the pivot far into the tails. The point of evaluation is given by a
linear function of the data, which should itself converge to a Gaussian distribution (at least when d
is finite). But even a small amount of non-Gaussianity is magnified when we are in the tails. To see
this, consider the function
R∞
p(z) dz
H p (t) = Rt+∞1
.
t p(z) dz
The left plot in Figure 12 shows two densities p and q which are nearly indistinguishable. The right
plot shows their corresponding tail functions H p and H q . Even though p and q are close, we see that
H p and H q are quite different. The message is that any inferential method that depends heavily on
extreme tail behavior could be unreliable.
Perhaps more visually striking is a plot of the TG statistic, when viewed as a function of y (for
X fixed). This is shown in Figure 13, where the statistic is used to test µ = 0, and we used the same
44
setup—thus the same model selection partition elements, and even matching colors—as in Figure 2.
Here n = 2, so it is possible to fully visualize the TG statistic as a function of y ∈ R2 . This function
is not well-behaved at the boundaries between partition elements corresponding to different model
selection events. Technically, this function is continuous on the interior of each partition element,
which permits an application of the (uniform) continuous mapping theorem when d is fixed. But the
derivatives at the boundaries are infinite and, especially in high-dimensional problem settings, there
is a nonnegligible probability of being near a boundary. Thus a small perturbation to the data could
have a dramatic effect on the value of the pivot.
References
Bachoc, F., Leeb, H. & Potscher, B. (2014), Valid confidence intervals for post-model-selection predictors. arXiv: 1412.4605.
Berk, R., Brown, L., Buja, A., Zhang, K. & Zhao, L. (2013), ‘Valid post-selection inference’, Annals of
Statistics 41(2), 802–837.
Chen, L., Goldstein, L. & Shao, Q.-M. (2011), Normal Approximation by Stein’s Method, Springer.
Choi, Y., Taylor, J. & Tibshirani, R. (2014), Selecting the number of principal components: estimation
of the true rank of a noisy matrix. arXiv: 1410.8260.
Donoho, D. (1988), ‘One-sided inference about functionals of a density’, Annals of Statistics
16(4), 1390–1420.
Fithian, W., Sun, D. & Taylor, J. (2014), Optimal inference after model selection. arXv: 1410.2597.
Hyun, S., G’Sell, M. & Tibshirani, R. J. (2016), Exact post-selection inference for changepoint detection and other generalized lasso problems. arXv: 1606.03552.
Kasy, M. (2015), Uniformity and the delta method. Unpublished manuscript.
Lee, J., Sun, D., Sun, Y. & Taylor, J. (2016), ‘Exact post-selection inference, with application to the
lasso’, Annals of Statistics 44(3), 907–927.
Lee, J. & Taylor, J. (2014), ‘Exact post model selection inference for marginal screening’, Advances in
Neural Information Processing Systems 27, 136–144.
Leeb, H. & Potscher, B. (2003), ‘The finite-sample distribution of post-model-selection estimators and
uniform versus nonuniform approximations’, Econometric Theory 19(1), 100–142.
Leeb, H. & Potscher, B. (2006), ‘Can one estimate the conditional distribution of post-model-selection
estimators?’, Annals of Statistics 34(5), 2554–2591.
Leeb, H. & Potscher, B. (2008), ‘Can one estimate the unconditional distribution of post-modelselection estimators?’, Econometric Theory 24(2), 338–376.
Lockhart, R., Taylor, J., Tibshirani, R. J. & Tibshirani, R. (2014), ‘A significance test for the lasso’,
Annals of Statistics 42(2), 413–468.
45
3
2
1
0
−1
−2
0.2 −3 −2 −1
0.4
0.6
0.8
0
1
2
−3
3
3
2
1
0
0.8
0.6
0.4
0.2
−1
−2
−3 −2 −1
0
1
2
−3
3
Figure 13: Two 3d views of the TG statistic, with the pivot value set at µ = 0, in same setup as in Figure 2.
Here n = 2, and the statistic is plotted as a function of y ∈ R2 .
Loftus, J. & Taylor, J. (2014), A significance test for forward stepwise model selection.
1405.3920.
arXiv:
O’Hagan, A. & Leonard, T. (1976), ‘Bayes estimation subject to uncertainty about parameter constraints’, Biometrika 63(1), 201–203.
Reid, S., Taylor, J. & Tibshirani, R. (2017), ‘Post-selection point and interval estimation of signal
sizes in Gaussian samples’, Canadian Journal of Statisitcs 45(2), 128–148.
Rosenthal, H. (1970), ‘On the subspaces of l p (p > 2) spanned by sequences of independent random
variables’, Israel Journal of Mathematics 8(3), 273–303.
Taylor, J., Loftus, J. & Tibshirani, R. J. (2016), ‘Inference in adaptive regression via the kac-rice
formula’, Annals of Statistics 44(2), 743–770.
Tian, X. & Taylor, J. (2017), ‘Asymptotics of selective inference’, Scandinavian Journal of Statistics
44(2), 480–499.
Tibshirani, R. J., Taylor, J., Lockhart, R., & Tibshirani, R. (2016), ‘Exact post-selection inference for
sequential regression procedures’, Journal of the American Statistical Association 111(514), 600–
620.
van der Vaart, A. (1998), Asymptotic Statistics, Cambridge University Press.
Wasserman, L. (2014), ‘Discussion: A significance test for the lasso’, Annals of Statistics 42(2), 501–
508.
47
| 10math.ST
|
39
Demand-Driven Pointer Analysis with Strong Updates via Value-Flow
Refinement
arXiv:1701.05650v1 [cs.PL] 20 Jan 2017
Yulei Sui, School of Computer Science and Engineering, UNSW Australia
Jingling Xue, School of Computer Science and Engineering, UNSW Australia
We present a new demand-driven flow- and context-sensitive pointer analysis with strong updates for C
programs, called S UPA, that enables computing points-to information via value-flow refinement, in environments with small time and memory budgets such as IDEs. We formulate S UPA by solving a graphreachability problem on an inter-procedural value-flow graph representing a program’s def-use chains, which
are pre-computed efficiently but over-approximately. To answer a client query (a request for a variable’s
points-to set), S UPA reasons about the flow of values along the pre-computed def-use chains sparsely (rather
than across all program points), by performing only the work necessary for the query (rather than analyzing the whole program). In particular, strong updates are performed to filter out spurious def-use chains
through value-flow refinement as long as the total budget is not exhausted. S UPA facilitates efficiency and
precision tradeoffs by applying different pointer analyses in a hybrid multi-stage analysis framework.
We have implemented S UPA in LLVM (3.5.0) and evaluate it by choosing uninitialized pointer detection
as a major client on 18 open-source C programs. As the analysis budget increases, S UPA achieves improved
precision, with its single-stage flow-sensitive analysis reaching 97.4% of that achieved by whole-program
flow-sensitive analysis by consuming about 0.18 seconds and 65KB of memory per query, on average (with a
budget of at most 10000 value-flow edges per query). With context-sensitivity also considered, S UPA’s twostage analysis becomes more precise for some programs but also incurs more analysis times. S UPA is also
amenable to parallelization. A parallel implementation of its single-stage flow-sensitive analysis achieves a
speedup of up to 6.9x with an average of 3.05x a 8-core machine with respect its sequential version.
CCS Concepts: •Software and its engineering Ñ Software verification and validation; Software defect
analysis; •Theory of computation Ñ Program analysis;
Additional Key Words and Phrases: strong updates, value flow, pointer analysis, flow sensitivity
1. INTRODUCTION
Pointer analysis is one of the most fundamental static program analyses, on which
virtually all others are built. The goal of pointer analysis is to compute an approximation of the set of abstract objects that a pointer can refer to. A pointer analysis is (1)
flow-sensitive if it respects control flow and flow-insensitive otherwise and (2) contextsensitive if it distinguishes different calling contexts and context-insensitive otherwise.
Strong updates, where stores overwrite, i.e., kill the previous contents of their abstract destination objects with new values, is an important factor in the precision of
pointer analysis [Hardekopf and Lin 2009; Lhoták and Chung 2011]. In the case of
weak updates, these objects are assumed conservatively to also retain their old contents. Strong updates are possible only if flow-sensitivity is maintained. In addition,
a flow-sensitive analysis can strongly update an abstract object written at a store if
and only if that object has exactly one concrete memory address, known as a singleton.
By applying strong updates where needed, a pointer analysis can improve precision,
thereby providing significant benefits to many clients, such as change impact analysis [Acharya and Robinson 2011], bug detection [Yan et al. 2016; Ye et al. 2014a],
security analysis [Arzt et al. 2014], type state verification [Fink et al. 2008], compiler
optimization [Sui et al. 2016b, 2013, 2014b], and symbolic execution [Blackshear et al.
2013].
In this paper, we introduce a demand-driven pointer analysis for C by investigating
how to perform strong updates effectively in a flow- and context-sensitive framework.
For C programs, flow-sensitivity is important in achieving the precision required by
the afore-mentioned client applications due to strong updates performed. If contextsensitivity is also considered, some more strong updates are possible for some pro-
39:2
Yulei Sui and Jingling Xue
grams at the expense of more analysis times. For object-oriented languages like Java,
context-sensitivity (without strong updates) is widely used in achieving useful precision [Lhoták and Hendren 2003; Li et al. 2014; Milanova et al. 2002, 2005; Smaragdakis et al. 2011; Sun et al. 2011; Xiao and Zhang 2011].
Ideally, strong updates at stores should be performed by analyzing all paths independently by solving a meet-over-all-paths (MOP) problem. However, even with branch
conditions being ignored, this problem is intractable due to potentially unbounded
number of paths that must be analyzed [Landi 1992; Ramalingam 1994].
Instead, traditional flow-sensitive pointer analysis (FS) for C [Hind and Pioli 1998;
Kam and Ullman 1977] computes the maximal-fixed-point solution (MFP) as an overapproximation of MOP by solving an iterative data-flow problem. Thus, the data-flow
facts that reach a confluence point along different paths are merged. Improving on
this, sparse flow-sensitive pointer analysis (SFS) [Hardekopf and Lin 2011; Li et al.
2011; Oh et al. 2012; Ye et al. 2014b; Yu et al. 2010] boosts the performance of FS in
analyzing large C programs while maintaining the same strong updates done by FS.
The basic idea is to first conduct a pre-analysis on the program to over-approximate
its def-use chains and then perform FS by propagating the data-flow facts, i.e., pointsto information sparsely along only the pre-computed def-use chains (aka value-flows)
instead of all program points in the program’s control-flow graph (CFG).
Recently, an approach [Lhoták and Chung 2011] for performing strong updates in C
programs is introduced. It sacrifices the precision of FS to gain efficiency by applying
strong updates at stores where flow-sensitive singleton points-to sets are available but
falls back to the flow-insensitive points-to information otherwise.
By nature, the challenge of pointer analysis is to make judicious tradeoffs between
efficiency and precision. Virtually all of the prior analyses for C that consider some degree of flow-sensitivity are whole-program analyses. Precise ones are unscalable since
they must typically consider both flow- and context-sensitivity (FSCS) in order to maximize the number of strong updates performed. In contrast, faster ones like [Lhoták
and Chung 2011] are less precise, due to both missing strong updates and propagating
the points-to information flow-insensitively across the weakly-updated locations.
In practice, a client application of a pointer analysis may require only parts of the
program to be analyzed. In addition, some points-to queries may demand precise answers while others can be answered as precisely as possible with small time and memory budgets. In all these cases, performing strong updates blindly across the entire
program is cost-ineffective in achieving precision.
For C programs, how do we develop precise and efficient pointer analyses that are
focused and partial, paying closer attention to the parts of the programs relevant
to on-demand queries? Demand-driven analyses for C [Heintze and Tardieu 2001;
Zhang et al. 2014a; Zheng and Rugina 2008] and Java [Lu et al. 2013; Shang et al.
2012; Sridharan and Bodı́k 2006; Su et al. 2016; Yan et al. 2011] are flow-insensitive
and thus cannot perform strong updates to produce the precision needed by some
clients. B OOMERANG [Späth et al. 2016] represents a recent flow- and context-sensitive
demand-driven pointer analysis for Java. However, its access-path-based approach
performs strong updates at a store a.f “ . . . only partially, by updating a.f strongly
and the aliases of a.f.˚ weakly. Elsewhere, advances in whole-program flow-sensitive
analysis for C have exploited some form of sparsity to improve performance [Hardekopf
and Lin 2011; Li et al. 2011; Oh et al. 2012; Ye et al. 2014b; Yu et al. 2010]. However,
how to replicate this success for demand-driven flow-sensitive analysis for C is unclear.
Finally, it remains open as to whether sparse strong update analysis can be performed
both flow- and context-sensitively on-demand to avoid under- or over-analyzing.
In this paper, we introduce S UPA, the first demand-driven pointer analysis with
strong updates for C, designed to support flexible yet effective tradeoffs between effi-
Demand-Driven Flow-Sensitive Pointer Analysis
Program
Preanalysis
39:3
Stages
Value-Flows
Efficiency
Refine
Queries
Precision
Stage[0] .... Stage[N-1]
On-Demand
Reachability Solver
Stage[i]
Select i
Budgets
Out-Of-Budget[i]?
No
Yes
i++
Fig. 1: Overview of S UPA
ciency and precision in answering client queries, in environments with small time and
memory budgets such as IDEs. As shown in Figure 1, the novelty behind S UPA lies
in performing Strong UPdate Analysis precisely by refining imprecisely pre-computed
value-flows away in a hybrid multi-stage analysis framework. Given a points-to query,
strong updates are performed by solving a graph-reachability problem on an interprocedural value-flow graph that captures the def-use chains of the program obtained
conservatively by a pre-analysis. Such over-approximated value-flows can be obtained
by applying Andersen’s analysis [Andersen 1994] (flow- and context-insensitively).
S UPA conducts its reachability analysis on-demand sparsely along only the precomputed value-flows rather than control-flows. In addition, S UPA filters out imprecise
value-flows by performing strong updates flow- and context-sensitively where needed
with no loss of precision as long as the total analysis budget is sufficient. The precision
of S UPA depends on the degree of value-flow refinement performed under a budget.
The more spurious value-flows S UPA removes, the more precise the points-to facts are.
S UPA handles large C programs by staging analyses in increasing efficiency but
decreasing precision in a hybrid manner. Currently, S UPA proceeds in two stages by
applying FSCS and FS in that order with a configurable budget for each analysis.
When failing to answer a query in a stage within its alloted budget, S UPA downgrades
itself to a more scalable but less precise analysis in the next stage and will eventually
fall back to the pre-computed flow-insensitive information. At each stage, S UPA will
re-answer the query by reusing the points-to information found from processing the
current and earlier queries. By increasing the budgets used in the earlier stages (e.g.,
FSCS), S UPA will obtain improved precision via improved value-flow refinement.
In summary, this paper makes the following contributions:
— We present the first demand-driven flow- and context-sensitive pointer analysis
with strong updates for C that enables computing precise points-to information by
refining away imprecisely precomputed value-flows, subject to analysis budgets.
— We introduce a hybrid multi-stage analysis framework that facilitates efficiency and
precision tradeoffs by staging different analyses in answering client queries.
— We have produced an implementation of S UPA in LLVM (3.5.0) [SUPA 2016]. We
evaluate S UPA with uninitialized pointer detection as a practical client by using a
total of 18 open-source C programs. As the analysis budget increases, S UPA achieves
improved precision, with its single-stage flow-sensitive analysis reaching 97.4% of
that achieved by whole-program flow-sensitive analysis, by consuming about 0.18
39:4
Yulei Sui and Jingling Xue
seconds and 65KB of memory per query, on average (with a per-query budget of at
most 10000 value-flow edges traversed). With context-sensitivity also being considered, more strong updates are also possible. S UPA’s two-stage analysis then becomes
more precise for some programs at the expense of more analysis times.
— We present four case studies to demonstrate that S UPA is effective in checking
whether variables are initialized or not in real-world applications.
— We show that S UPA is amenable to parallelization. To demonstrate this, we have
developed a parallel implementation of S UPA’s single-stage flow-sensitive analysis
based on Intel Threading Building Blocks (TBB), achieving a speedup of up to 6.9x
with an average of 3.05x a 8-core machine over its sequential version.
The rest of this paper is organized as follows. Section 2 provides the background
information. Section 3 presents a motivating example. Section 4 introduces our formalism for S UPA. Section 5 discusses and analyzes our experimental results. Section 6
contains four case studies. Section 7 describes a parallel implementation of S UPA. Section 8 describes the related work. Finally, Section 9 concludes the paper.
2. BACKGROUND
We describe how to represent a C program by an interprocedural sparse value-flow
graph to enable demand-driven pointer analysis via value-flow refinement. Section 2.1
introduces the part of LLVM-IR relevant to pointer analysis. Section 2.2 describes how
to put top-level variables in SSA form. Section 2.3 describes how to put address-taken
variables in SSA form. Section 2.4 constructs a sparse value-flow graph that represents
the def-use chains for both top-level and address-taken variables in the program.
2.1. LLVM-IR
We perform pointer analysis in the LLVM-IR of a program, as in [Balatsouras and
Smaragdakis 2016; Hardekopf and Lin 2011; Lhoták and Chung 2011; Li et al. 2011;
Sui et al. 2012; Ye et al. 2014b]. The domains and the LLVM instructions relevant to
pointer analysis are given in Table I. The set of all variables V are separated into two
subsets, O that contains all possible abstract objects, i.e., address-taken variables of a
pointer and P that contains all top-level variables.
In LLVM-IR, top-level variables in P “ S Y G, including stack virtual registers (symbols starting with ”%”) and global variables (symbols starting with ”@”) are explicit, i.e.,
directly accessed. Address-taken variables in O are implicit, i.e., accessed indirectly at
LLVM’s load or store instructions via top-level variables.
Only a subset of the complete LLVM instruction set that is relevant to pointer analysis are modeled. As in Table I, every function f of a program contains nine types of
instructions (statements), including seven types of instructions used in the function
body of f , and one F UN E NTRY instruction f pr1 , . . . , rn q with the declarations of the
parameters of f , and one F UN E XIT instruction retf p as the unique return of f . Note
that the LLVM pass UnifyFunctionExitNodes is executed before pointer analysis in order
to ensure that every function has only one F UN E XIT instruction.
Let us go through the seven types of instructions used inside a function. For an
A DDR O F instruction p“&o, known as an allocation site, o is one of the following objects:
(1) a stack object, o` , where ` is its allocation site (via an LLVM alloca instruction), (2) a
global object, i.e., a global object o` , where ` is its allocation site or a program function
of , where f is its name, and (3) a dynamically created heap object oh` , where ` is its
heap allocation site (e.g., via a malloc() call). For each object o (except for a function), we
write of ld to represent the sub-object that corresponds to its field f ld. For flow-sensitive
pointer analysis, the initializations for global objects take place at the entry of main().
Demand-Driven Flow-Sensitive Pointer Analysis
39:5
Table I: Domains and LLVM instructions used by pointer analysis.
Analysis Domains
`
PL
instruction labels
f ld
PC
constants (field accesses)
s
PS
stack virtual registers
g
PG
global variables
f
PF ĎG
program functions
p, q, r, x, y P P “ S Y G top-level variables
o, a, b, c, d P O
address-taken variables
v
P V “ P Y O program variables
LLVM Instruction Set
A DDR O F
p = &o
C OPY
p =q
P HI
p =φpq, rq
F IELD
p = &q Ñ f ld
L OAD
p = ˚q
S TORE
˚p = q
C ALL
p = qpr1 , . . . , rn q
F UN E NTRY f pr1 , . . . , rn q
F UN E XIT
retf p
C OPY denotes a casting instruction (e.g., bitcast) in LLVM. P HI is a standard SSA
instruction introduced at a confluence point in the CFG to select the value of a variable
from different control-flow branches. L OAD (S TORE) is a memory accessing instruction
that reads (write) a value from (into) an address-taken object.
Our handling of field-sensitivity is ANSI-compliant. Given a pointer to an aggregate
(e.g., a struct or an array), pointer arithmetic used for accessing anything other than
the aggregate itself has undefined behavior [ISO90 1990; Pearce et al. 2007] and thus
not modeled. To model the field accesses of a struct object, F IELD represents a getelementptr instruction with its field offset f ld as a constant value. A getelementptr instruction that operates on a non-constant field of a struct is modeled as C OPY instructions,
one for every field of the struct conservatively. Arrays are treated monolithically.
C ALL, p “ qpr1 , . . . , rn q, denotes a call instruction, where q can be either a global
variable (for a direct call) or a stack virtual register (for an indirect call).
2.2. SSA Form for Top-Level Variables
LLVM-IR is known as a partial SSA form since only top-level variables are in SSA
form. In LLVM-IR, top-level variables are explicit, i.e., directly accessed and can thus
be put in SSA form by using a standard SSA construction algorithm [Cytron et al.
1991] (with P HI instructions inserted at confluence points).
p = &a;
q = &c;
p = &a;
q = &c;
a = &b;
c = &d;
x = &b;
y = &d;
*p = x;
*q = y;
t1 = *p;
swap *p = *q;
*q = t1;
(a) C code
t1 = *p;
t2 = *q;
swap
*p = t2;
*q = t1;
(b) Partial SSA
Points-to relations for p and q
observed at runtime
p
q
p
q
a
c
a
c
b
d
b
d
(c) Before swap
(d) After swap
Fig. 2: A swap example and its partial SSA form.
Let us illustrate LLVM’s partial SSA form by using an example in Figure 2. Figure 2(a) shows a swap program in C and Figure 2(b) gives its corresponding partial
39:6
Yulei Sui and Jingling Xue
SSA form. Figures 2(c) and (d) depict some (runtime) points-to relations before and
after the swap operation. In this example, we have p, q, x, y, t1, t2 P P and a, b, c, d P O.
Note that x, y, t1 and t2 are new temporary registers introduced in order to put the program given in Figure 2(a) into the partial SSA form given in Figure 2(b). In particular,
˚p “ ˚q is decomposed into t2 “ ˚q and ˚p “ t2, where t2 is a top-level pointer.
In LLVM-IR, all top-level variables are in SSA form. In this example, all top-level
variables are trivially in SSA form, as each has exactly one definition only. As a result,
the def-use chains for top-level variables are readily available.
However, address-taken variables are accessed indirectly at loads and stores via toplevel variables and thus not in SSA form. For example, the address-taken variable a
is defined implicitly twice, once at ˚p “ x and once at ˚p “ t2, and the address-taken
variable c is also defined implicitly twice, once at ˚q “ y and once at ˚q “ t1. As a
result, the def-use chains for address-taken variables are not immediately available.
2.3. SSA Form for Address-Taken Variables
Starting with LLVM’s partial SSA form, we first perform a pre-analysis by using Andersen’s algorithm flow- and context-insensitively [Andersen 1994], implemented in
SVF [Sui and Xue 2016]. We then put address-taken variables in memory SSA form,
by using the SSA construction algorithm [Cytron et al. 1991]. Imprecise points-to information computed this way will be refined by our demand-driven pointer analysis.
Given a variable v, AnderPtspvq represents its points-to set computed by Andersen’s
algorithm. There are two steps [Sui et al. 2014a], illustrated in Figures 3(a) and (b)
intraprocedurally and in Figures 4(a) and (b) interprocedurally.
Step 1: Computing Modification and Reference Side-Effects. As shown in Figure 3(a),
every load, e.g., t1 “ ˚q is annotated with a µpaq operator for each object a pointed
by q, i.e., a P AnderPtspqq to represent a potential use of a at the load. Similarly,
every store, e.g., ˚p “ x is annotated with a a “ χpaq operator for each object a P
AnderPtsppq to represent a potential def and use of a at the store. If a can be strongly
updated, then a receives whatever x points to and the old contents in a are killed.
Otherwise, a must also incorporate its old contents, resulting in a weak update to a.
We compute the side-effects of a function call by applying a lightweight interprocedural mod-ref analysis [Sui et al. 2014a, §4.2.1]. For a given callsite `, it is annotated
with µpaq (a “ χpaq) if a may be read (modified) inside the callees of ` (discovered by
Andersen’s pointer analysis). In addition, appropriate χ and µ operators are also
added for the F UN E NTRY and F UN E XIT instructions of these callees in order to
mimic passing parameters and returning results for address-taken variables.
Figure 4(a) gives an example modified from Figure 3(a) by moving the four swap instructions into a function, swap. For read side-effects, µpaq and µpcq are added before
callsite `7 to represent the potential uses of a and c in swap. Correspondingly, swap’s
F UN E NTRY instruction `8 is annotated with a “ χpaq and c “ χpcq to receive the values of a and c passed from `7 . For modification side-effects, a “ χpaq and c “ χpcq are
added after `7 to receive the potentially modified values of a and c returned from
swap’s F UN E XIT instruction `13 , which are annotated with µpaq and µpcq.
Step 2: Memory SSA Renaming. All the address-taken variables are converted into
SSA form as suggested in [Chow et al. 1996]. Every µpaq is treated as a use of a.
Every a “ χpaq is treated as both a def and use of a, as a may admit only a weak
update. Then the SSA form for address-taken variables is obtained by applying a
standard SSA construction algorithm [Cytron et al. 1991].
For the program annotated with µ’s and χ’s in Figure 3(a), Figure 3(b) gives its memory SSA form. Similarly, Figure 4(b) gives the memory SSA form for Figure 4(a).
Demand-Driven Flow-Sensitive Pointer Analysis
ℓ1:
ℓ2:
ℓ3:
ℓ4:
ℓ5:
39:7
p = &a;
q = &c;
p = &a;
q = &c;
p = &a;
q = &c;
x = &b;
y = &d;
x = &b;
y = &d;
x = &b;
y = &d;
*p = x;
a1 = !(a0)
a1 = !(a0)
ℓ6: *q = y;
[a]
*q = y;
c = !(c)
"(a)
"(a1)
"(a1)
t1 = *p;
"(c)
[a]
"(c1)
ℓ8: t2 = *q;
ℓ9: *p = t2;
t2 = *q;
swap
a = !(a)
ℓ10: *q = t1;
c = !(c)
t1 = *p;
[c]
"(c1)
[c]
t2 = *q;
swap
*p = t2;
*p = t2;
a2 = !(a1)
a2 = !(a1)
*q = t1;
*q = t1;
c2 = !(c1)
c2 = !(c1)
(b) Step 2: renaming
(a) Step 1: adding "s$and$!s
*q = y;
c1 = !(c0)
c1 = !(c0)
ℓ7: t1 = *p;
swap
*p = x;
*p = x;
a = !(a)
(c) Sparse value-flows of a and c
Fig. 3: Memory SSA form and sparse value-flows constructed intraprocedurally for
Figure 2, obtained with Andersen’s analysis: AnderPtsppq “ tau and AnderPtspqq “ tcu.
foo(){
p = &a;
q = &c;
ℓ1 :
ℓ2 :
ℓ3 :
ℓ4 :
x = &b;
y = &d;
ℓ5: *p = x;
a = !(a)
ℓ6: *q = y;
c = !(c)
"(a)
"(c)
ℓ8: swap(p,q){
a = !(a)
c = !(c)
}
swap(p,q){
ℓ9: t1 = *p;
"(c)
ℓ10: t2 = *q;
ℓ11: *p = t2;
t1 = *p;
"(c1)
*q = y;
t2 = *q;
*p = t2;
a1 = !(a0)
ℓ12: *q = t1;
"(a1)
"(c1)
c = !(c)
*q = t1;
}
(a) Step 1: adding "s$and$!s
}
[a]
}
(b) Step 2: renaming
a1 = !(a0)
c1 = !(c0)
[a]
[a]
"(a1)
"(c1)
swap(p,q);
a2 = !(a1)
c2 = !(c1)
}
[a]
[c]
"(c1)
[c]
t2 = *q;
*p = t2;
c1 = !(c0)
[c]
"(a1)
t1 = *p;
*q = y;
"(a2)
"(c2)
a2 = !(a1)
c2 = !(c1)
"(a)
"(c)
a1 = !(a0)
c2 = !(c1)
swap(p,q);
ℓ13:
*p = x;
a2 = !(a1)
c1 = !(c0)
a = !(a)
x = &b;
y = &d;
"(a1)
*p = x;
swap(p,q){
foo(){
p = &a;
q = &c;
a1 = !(a0)
c1 = !(c0)
x = &b;
y = &d;
"(a)
ℓ7: swap(p,q);
a = !(a)
c = !(c)
foo(){
p = &a;
q = &c;
[c]
a2 = !(a1)
[a]
*q = t1;
c2 = !(c1)
[c]
"(a2)
"(c2)
[a]
[c]
}
(c) Sparse value-flows of a and c
Fig. 4: Memory SSA form and sparse value-flows constructed interprocedurally for an
example modified from Figure 2 with its four swap instructions moved into a separate
function, called swap. `8 and `13 correspond to the F UN E NTRY and F UN E XIT of swap.
2.4. Sparse Value-Flow Graph
Once both top-level and address-taken variables are in SSA form, their def-use chains
are immediately available, as shown in Table II. We discussed top-level variables earlier. For the two address-taken variables a and c in Figure 2, Figure 3(c) depicts their
def-use chains, i.e., sparse value-flows for the memory SSA form in Figure 3(b). Similarly, Figure 4(c) gives their sparse value-flows for the memory SSA form in Figure 4(b).
Given a program, a sparse value-flow graph (SVFG), Gvfg “ pN, Eq, is a multi-edged
directed graph that captures its def-use chains for both top-level and address-taken
39:8
Yulei Sui and Jingling Xue
Table II: Def-use information of both top-level and address-taken variables. Defv
(U sev ) denotes the set of definition (use) instructions for a variable v P V.
Instruction
`
p “ &o
p“q
p “ φpq, rq
p “ &q Ñ f ld
p “ ˚q
µpai q
˚p “ q
ai`1 “ χpai q
p “ qpr1 , . . . , rn q
µpai q aj`1 “ χpaj q
f pr1 , . . . , rn q ai`1 “ χpai q
retf p
µpai q
Defs and Uses of Variables in Memory SSA Form
t`u “ Defp
t`u “ Defp ` P U seq
t`u “ Defp ` P U seq ` P U ser
t`u “ Defp ` P U seq
t`u “ Defp ` P U seq ` P U seai
` P U sep ` P U seq ` P Defai`1 ` P U seai
t`u “ Defp ` P U seq @ i P 1, . . . , n : ` P U seri
` P U seai ` P Defaj`1 ` P U seaj
@ i P 1, . . . , n : ` P Defri ` P Defai`1 ` P U seai
` P U sep ` P U seai
` P Defp `1 P U sep
` P Defai `1 P U seai
[INTRA-ADDR]
p
a
`Ý
Ñ `1
`Ý
Ñ `1
` : p “ qpr1 , . . . , rn q of P AnderPtspqq `1 : f pr11 , . . . , rn1 q
[INTER-CALL-TOP]
ri
@i P 1, . . . , n : ` ÝÑ
`1
1
` : p “ qp. . . q af P AnderPtspqq ` : retf p1
[INTER-RET-TOP]
p
`1 Ý
Ñ`
` : p “ qp. . . q µpai q af P AnderPtspqq `1 : f p. . . q aj`1 “ χpaj q
[INTER-CALL-ADDR]
a
`Ý
Ñ `1
µpai q
` : “ qp. . . q aj`1 “ χpaj q af P AnderPtspqq `1 : retf
[INTER-RET-ADDR]
1 a
` Ý
Ñ`
[INTRA-TOP]
Fig. 5: Value-flow construction in Memory SSA form.
variables. N is the set of nodes representing all instructions and E is the set of edges
v
representing all potential def-use chains. In particular, an edge `1 Ý
Ñ `2 , where v P V,
from statement `1 to statement `2 signifies a potential def-use chain for v with its def at
v
`1 and use at `2 . We refer to `1 Ý
Ñ `2 a direct value-flow if v P P and an indirect valueflow if v P O. This representation is sparse since the intermediate program points
between `1 and `2 are omitted, thereby enabling the underlying points-to information
to be gradually refined by applying a sparse demand-driven pointer analysis.
Figure 5 gives the rules for connecting value-flows between two instructions
based on the defs and uses computed in Table II. For intraprocedural value-flows,
[INTRA-TOP] and [INTRA-ADDR] handle top-level and address-taken variables, respectively. In SSA form, every use of a variable only has a unique definition. For a use of a
identified as ai (with its i-th version) at `1 annotated with µpai q, its unique definition
a
in SSA form is ai at an ` annotated with ai “ χpai´1 q. Then, ` Ý
Ñ `1 is generated to rep1
resent potentially the value-flow of a from ` to ` . Thus, the PHI functions introduced
a
for address-taken variables will be ignored, as the value a in ` Ý
Ñ `1 is not versioned.
Let us consider interprocedural value-flows. The def-use information in Table II
is only intraprocedural. According to Figure 5, interprocedural value-flows are constructed to represent parameter passing for top-level variables ([INTER-CALL-TOP]
and [INTER-RET-TOP]), and the µ{χ operators annotated at F UN E NTRY, F UN E XIT
and C ALL for address-taken variables ([INTER-CALL-ADDR] and [INTER-RET-ADDR]).
Demand-Driven Flow-Sensitive Pointer Analysis
39:9
[INTER-CALL-TOP] connects the value-flow from an actual argument ri at a call instruction ` to its corresponding formal parameter ri1 at the F UN E NTRY `1 of every callee
f invoked at the call. Conversely, [INTER-RET-TOP] models the value-flow from the
F UN E XIT instruction of f to every callsite where f is invoked. Just like for top-level
variables, [INTER-CALL-ADDR] and [INTER-RET-ADDR] build the value-flows of addresstaken variables across the functions according to the annotated µ’s and χ’s. Note that
the versions i and j of an SSA variable a in different functions may be different. For
a
c
example, Figure 4(c) illustrates the four inter-procedural value-flows `7 Ý
Ñ `8 , `7 Ñ
Ý `8 ,
a
c
`13 Ý
Ñ `7 and `13 Ñ
Ý `7 obtained by applying the two rules to Figure 4(b).
a
The SVFG obtained this way may contain spurious def-use chains, such as `5 Ý
Ñ `9
in Figure 3, as Andersen’s flow- and context-insensitive pointer analysis is fast but
imprecise. However, this representation allows imprecise points-to information to be
refined by performing sparse whole-program flow-sensitive pointer analysis as in prior
work [Hardekopf and Lin 2011; Nagaraj and Govindarajan 2013; Sui et al. 2016a; Ye
et al. 2014b]. In this paper, we introduce a demand-driven flow- and context-sensitive
pointer analysis with strong updates that can answer points-to queries efficiently and
precisely on-demand, by removing spurious def-use chains in the SVFG iteratively.
3. A MOTIVATING EXAMPLE
Our demand-driven pointer analysis, S UPA, operates on the SVFG of a program. It
computes points-to queries flow- and context-sensitively on-demand by performing
strong updates, whenever possible, to refine away imprecise value-flows in the SVFG.
Our example program, shown in Figure 6(a), is simple (even with 16 lines). The
program consists of a straight-line sequence of code, with `1 – `10 taken directly from
Figure 2(b) and the six new statements `11 – `16 added to enable us to highlight some
key properties of S UPA. We assume that u at `11 is uninitialized but i at `12 is initialized. The SVFG embedded in Figure 6(a) will be referred to shortly below. We describe
how S UPA can be used to prove that z at `16 points only to the initialized object i, by
computing flow-sensitively on-demand the points-to query ptpx`16 , zyq, i.e., the points-to
set of z at the program point after `16 , which is defined in (1) in Section 4.
Figure 6(b) depicts the points-to relations for the six address-taken variables and
some top-level ones found at the end of the code sequence by a whole-program flowsensitive analysis (with strong updates) like SFS [Hardekopf and Lin 2011]. Due to
flow-sensitivity, multiple solutions for a pointer are maintained. In this example, these
are the true relations observed at the end of program execution. Note that SFS gives
rise to Figure 2(c) by analyzing `1 – `6 , Figure 2(d) by analyzing also `7 – `10 , and
finally, Figure 6(b) by analyzing `11 – `16 further. As z points to i but not u, no warning
is issued for z, implying that z is regarded as being properly initialized.
Figure 6(c) shows how the points-to relations in Figure 6(b) are over-approximated
flow-insensitively by applying Andersen’s analysis [Andersen 1994]. In this case, a
single solution is computed conservatively for the entire program. Due to the lack of
strong updates in analyzing the two stores performed by swap, the points-to relations
in Figures 2(c) and 2(d) are merged, causing ˚a and ˚c to become spurious aliases.
When `11 – `16 are analyzed, the seven spurious points-to relations (shown in dashed
arrows in Figure 6(c)) are introduced. Since z points to i (correctly) and u (spuriously), a
false alarm for z will be issued. Failing to consider flow-sensitivity, Andersen’s analysis
is not precise for this uninitialization pointer detection client.
Let us now explain how S UPA, shown in Figure 1, works. S UPA will first perform
a pre-analysis to the example program to build the SVFG given in Figure 6(a), as
discussed in Section 2. For its top-level variables, their direct value-flows, i.e., defuse chains are explicit and thus omitted to avoid cluttering. For example, q has three
39:10
Yulei Sui and Jingling Xue
Spurious Points-to
Points-to
Direct Value-flow
ℓ1: p = &a;
Query
.
pt(⟨ℓ
16 ,z⟩) =?
ℓ1: p = &a;
ℓ2: q = &c;
ℓ2: q = &c;
ℓ3: x = &b;
ℓ4: y = &d;
ℓ4: y = &d;
ℓ5: *p = x;
ℓ5: *p = x;
SU for c
ℓ6: *q = y;
[a]
swap
[q]
[c]
ℓ8: t2 = *q;
[a]
3
ℓ10: *q = t1;
x
ℓ12: v = &i;
[t2]
4
[p]
ℓ12: v = &i;
[a]
9 [v]
2
ℓ14: *t3 = w;
x
[p]
ℓ14: *t3 = w;
[d]
ℓ15: *t3 = v;
SU for d
ℓ16: z = *t3;
x
Spurious Value-Flows
q
a
c
t3
b
8
ℓ16: z = *t3;
1
[t3]
(d) The SUPA analysis for resolving pt(⟨ℓ16 ,z⟩) = {i} by
traversing from ⟨ℓ16 ,z⟩ backwards against the value-flows
(a) A program and its SVFG (with
only indirect value-flows shown)
p
[t3]
[d]
[b]
[d]
x
[d]
[b]
ℓ15: *t3 = v;
[b]
6
ℓ13: t3 = *p;
ℓ13: t3 = *p;
[b]
[c]
ℓ9: *p = t2;
ℓ11: w = &u;
[a]
7
ℓ8: t2 = *q;
SU for a
[c]
ℓ9: *p = t2;
[y]
ℓ6: *q = y;
[q]
5
ℓ7: t1 = *p;
[a]
Indirect Value-flow
p
q
a
c
t3
d
z
b
d
i
u
z
i
u
(b) Flow-sensitive points-to relations found
to hold at the end of the program
(with some for top-level pointers omitted)
(c) Flow-insensitive points-to relations
(with some for top-level pointers omitted)
Fig. 6: A motivating example for illustrating S UPA (SU stands for “Strong Update”).
q
q
q
def-use chains `2 Ý
Ñ `6 , `2 Ý
Ñ `8 and `2 Ý
Ñ `10 . For its address-taken variables, there
are nine indirect value-flows, i.e., def-use chains depicted in Figure 6(a). Let us see
how the two def-use chains for b are created. As t3 points to b, `14 , `15 and `16 will be
annotated with b “ χpbq, b “ χpbq and µpbq, respectively. By putting b in SSA form,
b
these three functions become b2 “ χpb1q, b3 “ χpb2q and µpb3q. Hence, we have `14 Ñ
Ý `15
b
and `15 Ñ
Ý `16 , indicating b at `16 has two potential definitions, with the one at `15
overwriting the one at `14 . The def-use chains for d and a are built similarly.
Demand-Driven Flow-Sensitive Pointer Analysis
39:11
Let us consider a single-stage analysis with Stage[N-1] “ Stage[0] “ F S in Figure 1. Figure 6(d) shows how S UPA computes ptpx`16 , zyq on-demand, starting from `16 ,
by performing a backward reachability analysis on the SVFG, with the visiting order
of def-use chains marked as 1 – 9 . Formally, this is done as illustrated in Figure 8.
The def-use chains for only the relevant top-level variables are shown. By filtering out
the four spurious value-flows (marked by ), S UPA finds that only i at `12 is backward
reachable from z at `16 . Thus, ptpx`16 , zyq “ tiu. So no warning for z will be issued.
S UPA differs from prior work in the following three major aspects:
— On-Demand Strong Updates
A whole-program flow-sensitive analysis like SFS [Hardekopf and Lin 2011] can
answer ptpx`16 , zyq precisely but must accomplish this task by analyzing all the 16
statements, resulting in a total of six strong updates performed at the six stores,
with some strong updates performed unnecessarily for this query. Unfortunately,
existing whole-program FSCS or even just FS algorithms do not scale well for large
C programs [Acharya and Robinson 2011].
In contrast, S UPA computes ptpx`16 , zyq precisely by performing only three strong
updates at `6 , `9 and `15 . The earlier a strong update is performed by S UPA during
its reachability analysis, the fewer the number of statements traversed. After 1 –
8 have been performed, S UPA finds that t3 points to d only. With a strong update
performed at `15 : ˚t3 “ v ( 9 ), S UPA concludes that ptpx`16 , zyq “ tiu.
— Value-Flow Refinement
Demand-driven pointer analyses [Shang et al. 2012; Sridharan and Bodı́k 2006;
Yan et al. 2011; Zhang et al. 2014a; Zheng and Rugina 2008] are flow-insensitive
and thus suffer from the same imprecision as their flow-insensitive whole-program
counterparts. In the absence of strong updates, many spurious aliases (such as ˚a
and ˚c) result, causing z to point to both i and u. As a result, a false alarm for z is
issued, as discussed earlier.
However, S UPA performs strong updates flow-sensitively by filtering out the four
b
spurious pre-computed value-flows marked by . As t3 points to d only, `15 Ñ
Ý `16 is
spurious and not traversed. In addition, a strong update is enabled at `15 : ˚t3 “ v,
b
d
a
rendering `14 Ñ
Ý `15 and `14 Ý
Ñ `15 spurious. Finally, `5 Ý
Ñ `9 is refined away due
to another strong update performed at `9 . Thus, S UPA has avoided many spurious
aliases (e.g., ˚a and ˚c) introduced flow-insensitively by pre-analysis, resulting in
ptpx`16 , zyq “ tiu precisely. Thus, no warning for z is issued.
— Query-based Precision Control
To balance efficiency and precision, S UPA operates in a hybrid multi-stage analysis framework. When asked to answer the query ptpx`16 , zyq under a budget, say, a
maximum sequence of three steps traversed, S UPA will stop its traversal from `9
to `8 (at 4 ) in Figure 6(d) and fall back to the pre-computed results by returning
ptpx`16 , zyq “ tu, iu. In this case, a false positive for z will end up being reported.
4. DEMAND-DRIVEN STRONG UPDATES
We introduce our demand-driven pointer analysis with strong updates, as illustrated
in Figure 1. We first describe our inference rules in a flow-sensitive setting (Section 4.1). We then discuss our context-sensitive extension (Section 4.2). Finally, we
present our hybrid multi-stage analysis framework (Section 4.3). All our analyses are
field-sensitive, thereby enabling more strong updates to be performed to struct objects.
39:12
Yulei Sui and Jingling Xue
q
` : p “ &o
x`, py Ðâ op
[ADDR]
` : p “ q `1 Ý
Ñ`
x`, py Ðâ x`1 , qy
[COPY]
q
r
` : p “ φpq, rq `1 Ý
Ñ ` `2 Ý
Ñ`
1
x`, py Ðâ x` , qy x`, py Ðâ x`2 , ry
[PHI]
q
[FIELD]
Ñ`
` : p “ &q Ñ f ld `1 Ý
x`, py Ðâ oy
f ld
x`1 , qy Ðâ op
q
o
` : p “ ˚q `2 Ý
Ñ ` x`2 , qy Ðâ op `1 Ý
Ñ`
x`, py Ðâ x`1 , oy
[LOAD]
q
p
[STORE]
[SU/WU]
Ñ`
Ñ ` x`2 , py Ðâ op `1 Ý
` : ˚p “ q `2 Ý
x`, oy Ðâ x`1 , qy
` : ˚p “
o
`1 Ý
Ñ`
o P Az killp`, pq
x`, oy Ðâ x`1 , oy
` : “ qp. . . , r, . . . q µpoj q
q
`2 Ý
Ñ ` x`2 , qy Ðâ ox
f
x`1 , r1 y Ðâ x`, ry
[CALL]
` : p “ qp. . . q oj`1 “ χpoj q
2 q
`1 : retf p1
1
1 p
µpoi q
1 o
` Ý
Ñ ` x` , qy Ðâ ox
` ÝÑ ` ` Ý
Ñ`
f
x`, py Ðâ xp1 , `1 y x`, oy Ðâ x`1 , oy
[RET]
[COMPO]
2
`1 : f p. . . , r1 , . . . q oi`1 “ χpoi q
r
o
`Ý
Ñ `1 ` Ý
Ñ `1
1
x` , oy Ðâ x`, oy
lv Ðâ lv 1
lv 1 Ðâ lv 2
lv Ðâ lv 2
$ 1
&to u
killp`, pq “ A
%
∅
if ptpx`, pyq “ to1 u ^ o1 P singletons
else if ptpx`, pyq “ ∅
otherwise
Fig. 7: Single-stage flow-sensitive S UPA analysis with demand-driven strong updates.
4.1. Formalism: Flow-Sensitivity
We present a formalization of a single-stage S UPA consisting of only a flow-sensitive
(FS) analysis. Given a program, S UPA will operate on its SVFG representation Gvfg
constructed by applying Andersen’s analysis [Andersen 1994] as a pre-analysis, as
discussed in Section 2.4 and illustrated in Section 3.
Let V “ L ˆ V be the set of labeled variables lv, where L is the set of statement
labels and V “ P Y O as defined in Table I. S UPA conducts a backward reachability
analysis flow-sensitively on Gvfg by computing a reachability relation, Ðâ Ď V ˆ V. In
our formalism, x`, vy Ðâ x`1 , v 1 y signifies a value-flow from a def of v 1 at `1 to a use of v at `
through one or multiple value-flow paths in Gvfg . For an object o created at an A DDR O F
Demand-Driven Flow-Sensitive Pointer Analysis
39:13
statement, i.e., an allocation site at `1 , identified as x`1 , oy, we must distinguish it from
x`, oy accessed elsewhere at ` in our inference rules. Our abbreviation for x`1 , oy is op.
Given a points-to query x`, vy, S UPA computes ptpx`, vyq, i.e., the points-to set of x`, vy
by finding all reachable target objects op, defined as follows:
ptpx`, vyq “ to | x`, vy Ðâ opu
(1)
Despite flow-sensitivity, our formalization in Figure 7 makes no explicit references
to program points. As S UPA operates on the def-use chains in Gvfg , each variable x`, vy
mentioned in a rule appears at the point just after `, where v is defined.
Let us examine our rules in detail. By [ADDR], an object op created at an allocation site
` is backward reachable from p at ` (or precisely at the point after `). The pre-computed
direct value-flows across the top-level variables in Gvfg are always precise ([COPY] and
[PHI]). In partial SSA form, [PHI] exists only for top-level variables (Section 2.4).
However, the indirect value-flows across the address-taken variables in Gvfg can
be imprecise; they need to be refined on the fly to remove the spurious aliases thus
introduced. When handling a load p “ ˚q in [LOAD], we can traverse backwards from
p at ` to the def of o at `1 only if o is actually used by, i.e., aliased with ˚q at `, which
requires the reachability relation x`2 , qy Ðâ op to be computed recursively. A store ˚p “ q
is handled similarly ([STORE]): q defined at `1 can be reached backwards by o at ` only
if o is aliased with ˚p at `.
If ˚q in a load ¨ ¨ ¨ “ ˚q is aliased with ˚p in a store ˚p “ ¨ ¨ ¨ executed earlier, then
p and q must be both backward reachable from op. Otherwise, any alias relation established between ˚p and ˚q in Gvfg by pre-analysis must be spurious and will thus be
filtered out by value-flow refinement.
[SU/WU] models strong and weak updates at a store ` : p “ . Defining its kill
set killp`, pq involves three cases. In Case (1), p points to one singleton object o1 in
singletons, which contains all objects in A except the local variables in recursion, arrays (treated monolithically) or heap objects [Lhoták and Chung 2011]. In Section 4.2,
we discuss how to apply strong updates to heap objects context-sensitively. A strong
update is then possible to o. By killing its old contents at `1 , no further backward travero
sal along the def-use chain `1 Ý
Ñ ` is needed. Thus, x`, oy Ðâ x`1 , oy is falsified. In Case
(2), the points-to set of p is empty. Again, further traversal to x`1 , oy must be prevented
to avoid dereferencing a null pointer as is standard [Hardekopf and Lin 2009, 2011;
Lhoták and Chung 2011]. In Case (3), a weak update is performed to o so that its old
contents at `1 are preserved. Thus, x`, oy Ðâ x`1 , oy is established, which implies that the
o
backward traversal along `1 Ý
Ñ ` must continue.
[FIELD] handles field-sensitivity. For a field access (e.g., p “ &q Ñ f ld), pointer p
points to the field object of ld of object o pointed to by q.
[CALL] and [RET] handle the reachability traversal interprocedurally by computing
the call graph for the program on the fly instead of relying on the imprecisely precomputed call graph built by the pre-analysis as in [Hardekopf and Lin 2011]. In the
SVFG, the interprocedural value-flows sinking into a callee function f may come from
a spurious indirect callsite `. To avoid this, both rules ensure that the function pointer
q at ` actually points to f ([CALL] and [RET]). Essentially, given a points-to query z
at an indirect callsite ` : z “ p˚f pqpq. Instead of analyzing all the callees found by the
pre-analysis, S UPA recursively computes the points-to set of f p to discover new callees
at the callsite and then continues refining ptpx`, zyq using the new callees.
Finally, Ðâ is transitive, stated by [COMPO].
Let us try all our rules, by first revisiting our motivating example where strong updates are performed (Example 4.1) and then examining weak updates (Example 4.2).
Yulei Sui and Jingling Xue
39:14
p
`1 : p “ &a
a
x`1 , py Ðâ p
q
`8 : t2 “ ˚q `2 Ý
Ñ `8
[ADDR]
`2 : q “ &c
x`2 , qy Ðâ p
c
x`8 , t2y Ðâ x`6 , cy
x`13 , t3y Ðâ x`6 , cy
d
a
`9 Ý
Ñ `13
[LOAD]
c
[ADDR]`6 Ñ
Ý `8
[LOAD]
[COMPO]
p
t2
a `8 ÝÑ `9
`9 : ˚p “ t2 `1 Ý
Ñ `9 x`1 , py Ðâ p
x`9 , ay Ðâ x`8 , t2y
x`6 ,cy Ðâ x`4 ,yy
v
y
[COMPO]
[STORE]
[COMPO]
[STORE]
[COMPO]
`4 : y “ &d
[ADDR]
[COMPO]
[ADDR]
[COMPO]
x`12 , vy Ðâ pi
`12 : v “ &i
x`4 ,yy Ðâ dp
[STORE]
c `4 Ñ
`6 : ˚q “ y `2 Ý
Ñ `6 x`2 , qy Ðâ p
Ý `6
t3
x`13 ,t3y Ðâ dp
(b) Deriving ptpx`13 , t3yq (corresponding to 5 – 7 in Figure 6(d))
x`13 ,t3y Ðâ x`4 ,yy
q
x`13 , t3y Ðâ x`8 , t2y
(a) Deriving ptpx`13 , t3yq (corresponding to 1 – 4 in Figure 6(d))
x`13 , t3y Ðâ x`9 , ay
`13 : t3 “ ˚p `1 Ý
Ñ `13
x`13 ,t3y Ðâ x`8 ,t2y
t3
x`15 , dy Ðâ x`12 , vy
`15 : ˚t3 “ v `13 ÝÑ `15 x`13 , t3y Ðâ dp `12 Ý
Ñ `15
[LOAD]
`16 : z “ ˚t3 `13 ÝÑ `16 x`13 , t3y Ðâ dp `15 Ý
Ñ `16
x`16 , zy Ðâ x`15 , dy
x`16 , zy Ðâ x`12 , vy
x`16 , zy Ðâ pi
(c) Deriving ptpx`16 , zyq (corresponding to 8 – 9 in Figure 6(d))
Fig. 8: Reachability derivations for ptpx`16 , zyq shown in Figure 6(d) (with reuse of cached points-to results inside each box).
Demand-Driven Flow-Sensitive Pointer Analysis
39:15
Example 4.1. Figure 8 shows how we apply the rules of S UPA to answer ptpx`16 , zyq
illustrated in Figure 6(d). [SU/WU] (implicit in these derivations) is applied to `6 , `9
and `15 to cause a strong update at each store. At `6 , ptpx`6 , qyq “ tcu, the old contents
a
in c are killed. At `9 , `5 Ý
Ñ `9 becomes spurious since x`9 , ay Ðâ x`5 , ay is falsified. At `15 ,
b
d
`14 Ñ
Ý `15 and `14 Ý
Ñ `15 are filtered out since x`15 , byÐâx`14 , by and x`15 , dy Ðâ x`14 , dy are
b
falsified. Finally, `15 Ñ
Ý `16 is ignored since t3 points to d only ([LOAD]).
l
S UPA improves performance by caching points-to results to reduce redundant traversal, with reuse happening in the marked boxes in Figure 8. For example, in Figure 8(c),
p computed in [LOAD] is reused in [STORE].
ptpx`13 , t3yq “ tdu
ℓ1 :
ℓ2 :
ℓ3 :
ℓ4 :
ℓ5 :
ℓ6 :
ℓ7 :
ℓ8 :
[a]
Indirect value-flow
p1 = &a;
x = &b;
y = &c;
r = &d;
*p1 = x;
Andersen's
Points-to:
*p1 = y;
SU for a
q = p1
[a]
ℓ9 :
if(*) p2 = &e;
p3 = 𝞥(p1,p2);
ℓ10:
ℓ11:
Direct value-flow
*p3 = r;
[a]
z = *q;
WU for a
pt(p1) = {a}
pt(p2) = {e}
pt(p3) = {a,e}
pt(q) = {a}
pt(x) = {b}
pt(y) = {c}
pt(r) = {d}
pt(a) = {b,c,d}
pt(z) = {b,c,d}
Query
pt(⟨ℓ11 ,z⟩) =?
Fig. 9: Resolving ptpx`11 , zyq “ tc, du with a weak update.
Example 4.2. Let us consider a weak update example in Figure 9 by computing
ptpx`11 , zyq on-demand. At the confluence point `9 , p3 receives the points-to information
from both p1 and p2 in its two branches: x`9 , p3 y Ðâ p
a and x`9 , p3 y Ðâ ep. Thus, a weak update is performed to the two locations a and e at `10 . Let us focus on p
a only. By applying
p By applying [SU/WU], x`10 , ay Ðâ x`6 , ay Ðâ x`3 , yy Ðâ p
[STORE], x`10 , ay Ðâ x`4 , ry Ðâ d.
c.
Thus, ptpx`11 , ayq “ tc, du, which excludes b due to a strong update performed at `6 . As
ptpx`7 , qyq “ tau, we obtain ptpx`11 , zyq “ tc, du.
l
Unlike [Lhoták and Chung 2011], which falls back to the flow-insensitive points-to
information for all weakly updated objects, S UPA handles them as precisely as (wholeprogram) flow-sensitive analysis subject to a sufficient budget. In Figure 9, due to
a weak update performed to a at `10 , ptpx`10 , ayq “ tc, du is obtained, forcing their
approach to adopt ptpx`10 , ayq “ tb, c, du thereafter, causing ptpx`11 , zyq “ tb, c, du. By
maintaining flow-sensitivity with a strong update applied to `6 to kill b, S UPA obtains
ptpx`11 , zyq “ tc, du precisely.
4.1.1. Handling Value-Flow Cycles. To compute soundly and precisely the points-to information in a value-flow cycle in the SVFG, S UPA retraverses it whenever new points-to
information is found until a fix point is reached.
x
z
Example 4.3. Figure 10 shows a value-flow cycle formed by `5 Ý
Ñ `6 and `6 Ý
Ñ `5 . To
compute ptpx`6 , zyq, we must compute ptpx`5 , xyq, which requires the aliases of ˚z at the
load `5 : x “ ˚z to be found by using ptpx`6 , zyq. S UPA computes ptpx`6 , zyq by analyzing
39:16
Yulei Sui and Jingling Xue
Direct Value-flow
ℓ1 :
ℓ2 :
ℓ3 :
ℓ4 :
Query
pt(⟨ℓ5 ,z⟩) =?
*q = p;
[b]
ℓ5 :
ℓ6 :
Indirect Value-flow
p = &a;
q = &b;
y = &b;
x = *z;
[z]
[x]
z = 𝞥(x,y);
.
Fig. 10: Resolving ptpx`5 , zyq “ ta, bu in a value-flow cycle.
this value-flow cycle in two iterations. In the first iteration, a pointed-to target pb is
found since x`6 , zy Ðâ x`4 , yy Ðâ pb. Due to x`2 , qy Ðâ pb, ˚z and ˚q are found to be aliases.
a is found since x`6 , zy Ðâ x`5 , xy Ðâ x`3 , by Ðâ
In the second iteration, another target p
a. Thus, ptpx`6 , zyq “ ta, bu is obtained.
x`1 , py Ðâ p
l
4.1.2. Field-Sensitivity. Field-insensitive pointer analysis does not distinguish different
fields of a struct object, and consequently, gives up opportunities for performing strong
updates to a struct object, as a struct object may actually represent its distinct fields.
In contrast, S UPA is truly field-sensitive, by avoiding the two limitations altogether.
Indirect Value-flow
bar() {
x = &a;
y = &b;
z = &c;
p = &(x->f);
q = &(x->g);
*p = y;
ℓ1:
ℓ2:
ℓ3:
ℓ4:
ℓ5:
ℓ6:
ℓ7:
[a]
ℓ9: foo(v) {
a1 = 𝝌(a0)
....
[a]
ℓ10: w = &(v->g);
𝜇(a1)
ℓ11: r = *w;
}
𝜇(a2)
foo(x);
[a]
ℓ1 :
ℓ2 :
ℓ3 :
ℓ4 :
ℓ5 :
ℓ6 :
ℓ9: foo(v) {
bar() {
a.g1 = 𝝌(a.g0)
x = &a;
y = &b;
....
[a.g]
z = &c;
[a.g]
ℓ10: w = &(v->g);
p = &(x->f);
q = &(x->g);
𝜇(a.g1)
ℓ11: r = *w;
*p = y;
}
a.f1 = 𝝌(a.f0)
ℓ7 :
*q = z;
a2 = 𝝌(a1)
[a]
ℓ8:
a1 = 𝝌(a0)
Indirect Value-flow
Query
pt(⟨ℓ11 ,r⟩) =?
}
(a) Field-insensitive value-flows (pt(<ℓ11 ,r>) = {b,c})
*q = z;
[a.g]
ℓ8 :
a.g1 = 𝝌(a.g0)
𝜇(a.g1)
foo(x);
Query
pt(⟨ℓ11 ,r⟩) =?
}
(b) Field-sensitive value-flows (pt(<ℓ11 ,r>) = {c})
Fig. 11: Resolving ptpx`11 , ryq “ tcu with field-sensitivity.
Example 4.4. Figure 11 illustrates the effects of field-sensitivity on computing
the points-to information for r at `11 . Without field-sensitivity, as illustrated in Figure 11(a), the two statements at `4 and `5 are analyzed as if they were `4 : p “ &x
and `5 : q “ &x. As a result, no strong update is possible at `6 and `7 , since x, which
represents possibly multiple fields, is not a singleton. Thus, ptpx`11 , ryq “ tb, cu.
S UPA is field-sensitive. To answer the points-to query for r at `11 , we compute first
a. By applying
x`11 , ry Ðâ x`10 , wy and then x`10 , vy Ðâ x`9 , vy Ðâ x`8 , xy Ðâ x`1 , xy Ðâ p
Demand-Driven Flow-Sensitive Pointer Analysis
39:17
[FIELD] at `10 and [LOAD] at `11 , we obtain x`11 , ry Ðâ x`11 , a.gy. By traversing the
b.g
a.g
a.g
three indirect def-use chains for a.g, `7 ÝÝÑ `8 , `8 ÝÝÑ `9 and `9 ÝÝÑ `11 , backwards
from `11 , we obtain ptpx`11 , ryq Ðâ x`9 , a.gy Ðâ x`8 , a.gy Ðâ x`7 , a.gy Ðâ x`3 , zy Ðâ p
c.
l
4.1.3. Properties
T HEOREM 4.5 pS OUNDNESSq. S UPA is sound in analyzing a program as long as its
pre-analysis (for computing the SVFG of the program) is sound.
P ROOF. When building the SVFG for a program, the def-use chains for its top-level
variables are identified explicitly in its partial SSA form. If the pre-analysis (for computing the sparse value-flow graph of the program) is sound, then the def-use chains
built for all the address-taken variables are over-approximate. According to its inference
rules in Figure 4, S UPA performs essentially a flow-sensitive analysis on-demand, by restricting the propagation of points-to information along the precomputed def-use chains,
and falls back to the sound points-to information computed by the pre-analysis when
running out of its given budgets. Thus, S UPA is sound if the pre-analysis is sound.
T HEOREM 4.6 pP RECISIONq. Given a points-to query x`, vy, ptpx`, vyq computed by
S UPA is the same as that computed by (whole-program) FS if S UPA can successfully
resolve the points-to query within a given budget.
P ROOF. Let ptS UPA px`, vyq and ptFS px`, vyq be the points-to sets computed by S UPA and
FS, respectively. By Theorem 1, ptS UPA px`, vyq Ě ptFS px`, vyq, since S UPA is a demanddriven version of FS and thus cannot be more precise. To show that ptS UPA px`, vyq Ď
ptFS px`, vyq, we note that S UPA operates on the SVFG of the program to improve its efficiency, by also filtering out value-flows imprecisely pre-computed by the pre-analysis.
For the top-level variables, their direct value-flows are precise. So S UPA proceeds exactly the same as FS ([ADDR], [COPY], [PHI], [FIELD], [CALL], [RET] and [COMPO]).
For the address-taken variables, S UPA establishes the same indirect value-flows flowsensitively as FS does but in a demand-driven manner, by refining away imprecisely
pre-computed value-flows ([LOAD], [STORE], [SU/WU], [CALL], [RET] and [COMPO]). If
S UPA can complete its query within the given budget, then ptS UPA px`, vyq Ď ptFS px`, vyq.
Thus, ptS UPA px`, vyq “ ptFS px`, vyq.
4.2. Formalism: Flow- and Context-Sensitivity
We extend our flow-sensitive formalization by considering also context-sensitivity to
enable more strong updates (especially now for heap objects). We solve a balancedparentheses problem by matching calls and returns to filter out unrealizable interprocedural paths [Lu et al. 2013; Reps et al. 1995; Shang et al. 2012; Sridharan and
Bodı́k 2006; Yan et al. 2011]. A context stack c is encoded as a sequence of callsites,
[κ1 . . . κm ], where κi is a call instruction `. c ‘ κ denotes an operation for pushing a
callsite κ into c. c a κ pops κ from c if c contains κ as its top value or is empty since a
realizable path may start and end in different functions.
With context-sensitivity, a statement is parameterized additionally by a context c,
e.g., c, ` : p “ &o, to represent its instance when its containing function is analyzed
under c. A labeled variable lv has the form xc, `, vy, representing variable v accessed
at statement ` under context c. An object op that is created at an A DDR O F statement
under context c is also context-sensitive, identified as pc, opq.
Given a points-to query xc, `, vy, S UPA computes its points-to set both flow- and
context-sensitively by applying the rules given in Figure 12:
ptpxc, `, vyq “ tpc1 , oq | xc, `, vy Ðâ pc1 , opqu
(2)
39:18
Yulei Sui and Jingling Xue
[C-PHI]
q
c, ` : p “ &o
xc, `, py Ðâ pc, opq
[C-ADDR]
Ñ`
c, ` : p “ q `1 Ý
xc, `, py Ðâ xc, `1 , qy
[C-COPY]
q
r
Ñ ` `2 Ý
Ñ`
`1 Ý
xc, `, py Ðâ xc, `2 , ry
c, ` : p “ φpq, rq
xc, `, py Ðâ xc, `1 , qy
q
[C-FIELD]
c, ` : p “ &q Ñ f ld `1 Ý
Ñ ` xc, `1 , qy Ðâ pc1 , opq
z
xc, `, py Ðâ pc1 , o.f
ldq
q
o
c, ` : p “ ˚q `2 Ý
Ñ ` xc, `2 , qy Ðâ pc1 , opq `1 Ý
Ñ`
xc, `, py Ðâ xc1 , `1 , oy
[C-LOAD]
q
p
[C-STORE]
[C-SU/WU]
[C-COMPO]
Ñ`
Ñ ` xc, `2 , py Ðâ pc1 , opq `1 Ý
c, ` : ˚p “ q `2 Ý
xc1 , `, oy Ðâ xc, `1 , qy
c, ` : ˚p “
o
`1 Ý
Ñ`
pc1 , oq P Az killpc, `, pq
1
xc , `, oy Ðâ xc1 , `1 , oy
lv Ðâ lv 1 lv 1 Ðâ lv 2
lv Ðâ lv 2
q
[C-CALL]
q
[C-RET]
r
o
c, ` : “ qp. . . , r, . . . q µpoj q
`2 Ý
Ñ ` xc, `2 , qy Ðâ p , ox
Ñ `1 ` Ý
Ñ `1
fq ` Ý
c1 , `1 : f p. . . , r1 , . . . q oi`1 “ χpoi q
c “ c1 a `
xc1 , `1 , r1 y Ðâ xc, `, ry xc1 , `1 , oy Ðâ xc, `, oy
p1
o
1
c, ` : p “ qp. . . q oj`1 “ χpoj q
`2 Ý
Ñ ` xc, `2 , qy Ðâ p , ox
Ñ ` `1 Ý
Ñ`
fq ` Ý
1 1
1
1
c , ` : retf p µpoi q
c “c‘`
xc, `, py Ðâ xc1 , p1 , `1 y xc, `, oy Ðâ xc1 , `1 , oy
$ 1 1
&tpc , o qu
killpc, `, pq “ A
%
∅
if ptpxc, `, pyq “ tpc1 , o1 qu ^ pc1 , o1 q P cxtSingletons
else if ptpxc, `, pyq “ ∅
otherwise
Fig. 12: Single-stage flow- and context-sensitive S UPA analysis with demand-driven
strong updates.
where the reachability relation Ðâ is now also context-sensitive.
Passing parameters to and returning results from a callee invoked at a callsite are
handled by [C-CALL] and [C-RET]. [C-CALL] deals with the direct and indirect valueflows backwards from the entry instruction of a callee function to each of its callsites
based on the call graph computed on the fly similarly as [CALL] in Figure 7, except that
[C-CALL] is context-sensitive. Likewise, [C-RET] deals with the direct and indirect
value-flows backwards from a callsite to the return instruction of every callee function.
With context-sensitivity, S UPA will filter out more spurious value-flows generated
by Andersen’s analysis, thereby producing more precise points-to information to enable more strong updates ([C-SU/WU]). At a store c, ` : ˚p “ , its kill set is contextsensitive. A strong update is applied if p points to a context-sensitive singleton pc1 , o1 q P
Demand-Driven Flow-Sensitive Pointer Analysis
ℓ1:
ℓ2:
bar() {
p = malloc(...); // a
q = malloc(...); // b
ℓ6:
[a]
𝜇(a0)
ℓ3:
foo(p)
[b]
a1 = 𝝌(a0)
ℓ4:
𝜇(b1)
foo(q);
[b]
𝜇(b2)
z = *q;
Indirect Value-flow
foo(x) {
a1 = 𝝌(a0)
b1 = 𝝌(b0)
[b]
ℓ7: y = malloc(..); // c
ℓ8:
[a]
[b] b2 = 𝝌(b1)
ℓ5:
39:19
*x = y;
b2 = 𝝌(b1)
a2 = 𝝌(a1)
ℓ9:
}
𝜇(a2)
𝜇(b2)
[a]
[a]
SU for heap
object b under
context [ℓ4]
[b]
Query
pt(⟨ℓ5 ,z⟩) =?
}
Fig. 13: Resolving ptpr s, `5 , zq “ tr`4 s, cu) with context-sensitive strong updates.
cxtSingletons, where o1 is a (non-heap) singleton defined in Section 4.1 or a heap object
with c1 being a concrete context, i.e., one not involved in recursion or loops.
Example 4.7. Let us use an example given in Figure 13 to illustrate the effects
of context-sensitive strong updates on computing the points-to information for z at
`5 . This example is adapted from a real application, milc-v6, given in Figure 17(c).
Without context-sensitivity, S UPA will only perform a weak update at `8 : ˚x “ y, since
x points to both a and b passed into foo() from the two callsites at `3 and `4 . As a
result, z at `5 is found to point to not only what y points to, i.e., c but also what b points
to previously (not shown to avoid cluttering). With context-sensitivity, S UPA finds that
xr s, `5 , zy Ðâ xr s, `5 , by Ðâ xr s, `4 , by Ðâ xr`4 s, `9 , by Ðâ xr`4 s, `8 , by Ðâ xr`4 s, `7 , yy Ðâ pr`4 s, p
cq.
Since xr`4 s, `8 , xy points to a context-sensitive singleton p`4 , bq at `8 , a strong update is
performed to b at `8 , causing the old contents in b to be killed.
l
Given a program, the SCCs (strongly connected components) in its call graph are
constructed on the fly. S UPA handles the SCCs in the program context-sensitively but
the function calls inside a SCC context-insensitively as in [Sridharan and Bodı́k 2006].
4.3. SUPA: Hybrid Multi-Stage Analysis
To facilitate efficiency and precision tradeoffs in answering on-demand queries, S UPA,
as illustrated in Figure 1, organizes its analyses in multiple stages sorted in increasing
efficiency but decreasing precision. Let there be M queries issued successively. Let the
N stages of S UPA, Stage[0], ¨ ¨ ¨ , Stage[N-1], be configured with budgets η0 , ¨ ¨ ¨ , ηN ´1 ,
respectively. In our current implementation, each budget is specified as the maximum
number of def-use chains traversed in the SVFG of the program.
S UPA answers a query on-demand by applying its N analyses successively, starting
from Stage[0]. If the query is not answered after budget ηi has been exhausted at
stage i, S UPA re-issues the query at stage i ` 1, and eventually falls back to the results
that are pre-computed by pre-analysis.
S UPA caches fully computed points-to information in a query and reuses it in subsequent queries, as illustrated in Figure 8. Let Q be the set of queried variables issued
from a program. Let I Ě Q be the set of variables reached from Q during the analysis.
Let p`, vq P Q be a queried variable. We write ptiηi px∆i , `, vyq to represent the points-to
39:20
Yulei Sui and Jingling Xue
set of a variable x`, vy computed at stage i under budget ηi , where ∆i is a contextual
qualifier at stage i (e.g., c in FSCS). By convention, ptN
ηN px∆N , `, vyq denotes the pointsto set obtained by pre-analysis, at Stage[N] (conceptually).
When resolving ptiηi px∆i , `, vyq at stage i, suppose S UPA has reached a variable
x`1 , v 1 y P I and needs to compute pti˚ px∆i , `1 , v 1 yq, where ˚pď ηi ) represents an unknown
budget remaining, with p`1 , v 1 q being p`, vq possibly (in a cycle).
Presently, S UPA exploits two types of reuse to improve efficiency with no loss of
precision in a hybrid manner:
Backward Reuse: p`1 , v 1 q P I . If ptj˚ px∆j , `1 , v 1 yq, where j ď i, was previously cached,
then pti˚ px∆i , `1 , v 1 yq “ ptj˚ px∆j , `1 , v 1 yq, provided that ptj˚ px∆j , `1 , v 1 yq is a sound representation of pti˚ px∆i , `1 , v 1 yq. For example, if Stage[i] “ F S and Stage[j] “ F SCS,
SCS
S
1 1
1
then ptF
pxc1 , `1 , v 1 yq can be reused for ptF
˚
˚ px` , v yq if c is true, representing a
context-free points-to set.
Forward Reuse: p`1 , v 1 q P Q . If ptjηj px∆j , `1 , v 1 yq, where j ą i, was previously computed
and cached but ptkηk px∆k , `1 , v 1 yq was not, where 0 ď k ă j, then S UPA will also fail
for ptk˚ px∆k , `1 , v 1 yq, where i ď k ă j, since ˚ ď ηk . Therefore, S UPA will exploit the
second type of reuse by setting pti˚ px∆i , `1 , v 1 yq “ ptjηj px∆j , `1 , v 1 yq.
Of course, many other schemes are possible with or without precision loss.
5. EVALUATION
We evaluate S UPA by choosing detection of uninitialized pointers as a major client.
The objective is to show that S UPA is effective in answering client queries, in environments with small time and memory budgets such as IDEs, by facilitating efficiency and
precision tradeoffs in a hybrid multi-stage analysis framework. We provide evidence
to demonstrate a good correlation between the number of strong updates performed
on-demand and the degree of precision achieved during the analysis.
5.1. Implementation
We have implemented S UPA in LLVM (3.5.0). The source files of a program are compiled under “-O0” (to facilitate detection of undefined values [Zhao et al. 2012]) into
bit-code by clang and then merged using the LLVM Gold Plugin at link time to
produce a whole program bc file. The compiler option mem2reg is applied to promote
memory into registers. Otherwise, SUPA will perform more strong updates on memory
locations that would otherwise be promoted to registers, favoring SUPA undesirably.
All the analyses evaluated are field-sensitive.
Positive weight cycles that arise from processing fields of struct objects are collapsed [Pearce et al. 2007]. Arrays are considered monolithic so that the elements in
an array are not distinguished. Distinct allocation sites (i.e., A DDR O F statements) are
modeled by distinct abstract objects.
We build the SVFG for a program based on our open-source software, SVF [Sui
and Xue 2016]. The def-use chains are pre-computed by Andersen’s algorithm flow
and context-insensitively. In order to compute soundly and precisely the points-to information in a value-flow cycle, S UPA retraverses the cycle whenever new points-to
information is discovered until a fix point is reached.
To compare S UPA with whole-program analysis, we have implemented a sparse flowsensitive (SFS) analysis described in [Hardekopf and Lin 2011] also in LLVM, as SFS
is a recent solution yielding exactly the flow-sensitive precision with good scalability.
However, there are some differences. In [Hardekopf and Lin 2011], SFS was imple-
Demand-Driven Flow-Sensitive Pointer Analysis
39:21
mented in LLVM (2.5.0), by using imprecisely pre-computed call graphs and representing points-to sets with binary decision diagrams (BDDs). In this paper, just like
S UPA, SFS is implemented in LLVM (3.5.0), by building a program’s call graph on the
fly (Section 4.1) and representing points-to sets with sparse bit vectors.
We have not implemented a whole-program FSCS pointer analysis in LLVM. There
is no open-source implementation either in LLVM. According to [Acharya and Robinson 2011], existing FSCS algorithms for C “do not scale even for an order of magnitude
smaller size programs than those analyzed” by Andersen’s algorithm. As shown here,
SFS can already spend hours on analyzing some programs under 500 KLOC.
5.2. Methodology
We choose uninitialized pointer detection as a major client, named Uninit, which requires strong update analysis to be effective. As a common type of bugs in C programs,
uninitialized pointers are dangerous, as dereferencing them can cause system crashes
and security vulnerabilities. For Uninit, flow-sensitivity is crucial. Otherwise, strong
updates are impossible, making Uninit checks futile.
We will show that S UPA can answer Uninit’s on-demand queries efficiently while
achieving nearly the same precision as SFS. For C, global and static variables are
default initialized, but local variables are not. In order to mimic the default uninitialization at a stack or heap allocation site ` : p “ &a for an uninitialized pointer a, we add
a special store ˚p “ u immediately after `, where u points to an unknown abstract object
(UAO), ua . Given a load x “ ˚y, we can issue a points-to query for x to detect its potential uninitialization. If x points to a ua (for some a), then x may be uninitialized. By
performing strong updates more often, a flow-sensitive analysis can find more UAO’s
that do not reach any pointer and thus prove more pointers to be initialized. Note that
SFS can yield false positives since, for example, path correlations are not modeled.
We do not introduce UAO’s for the local variables involved in recursion and array
objects since they cannot be strongly updated. We also ignore all the default-initialized
stack or heap objects (e.g., those created by calloc()).
We generate meaningful points-to queries, one query for the top-level variable x at
each load x “ ˚y. However, we ignore this query if x is found not to point to any UAO by
pre-analysis. This happens only when x points to either default-initialized objects or
unmodeled local variables in recursion cycles or arrays. The number of queries issued
in each program is listed in the last column in Table III.
5.3. Experimental Setup
We use a machine with a 3.7GHz Intel Xeon 8-core CPU and 64 GB memory. As
shown in Table III, we have selected a total of 18 open-source programs from a variety of domains: spell-1.1 (a spelling checker), bc-1.06 (a numeric processing language), milc-v6 (quantum chromodynamics), less-451 (a terminal pager), sed-4.2
(a stream editor), milc-v6 (quantum chromodynamics), hmmer-2.3 (sequence similarity searching), make-4.1 (a build automation tool), a2ps-4.14 (a postScript filter),
bison-3.04 (a parser), grep-2.2.1 (string searching), tar-1.28 (tar archiving),
wget-1.16 (a file downloading tool), bash-4.3 (a unix shell and command language),
gnugo-3.4 (a Go game), sendmail-8.15.1 (an email server and client), vim74 (a
text editor), and emacs-24.4 (a text editor).
For each program, Table III lists its number of lines of code, statements which
are LLVM instructions relevant to our pointer analysis, pointers, allocation sites (or
AddrOf statements), and queries issued (as discussed in Section 5.2).
39:22
Yulei Sui and Jingling Xue
Table III: Program characteristics.
Program
spell-1.1
bc-1.06
milc-v6
less-451
sed-4.2
hmmer-2.3
make-4.1
gzip-1.6
a2ps-4.14
bison-3.0.4
grep-2.21
tar-1.28
wget-1.16
bash-4.3
gnugo-3.4
sendmail-8.15
vim-7.4
emacs-24.4
Total
KLOC
0.8
14.4
15
27.1
38.6
36
40.4
64.4
64.6
113.3
118.4
132
140.0
155.9
197.2
259.9
413.1
431.9
2263.0
Statements
1011
17018
11713
6766
25835
27924
14926
22028
49172
36815
10199
30504
51556
59442
369741
86653
147550
189097
1157950
Pointers
1274
15212
29584
22835
34226
74689
36707
25646
116129
90049
33931
85727
63199
191413
286986
256074
466493
754746
2584920
Allocation Sites
42
654
865
1135
395
1472
1563
1180
3625
1976
1108
3350
726
6359
27511
7549
8960
12037
80507
Queries
17
689
3
100
1191
2043
1133
551
5065
4408
562
909
1142
5103
1970
2715
6753
4438
38792
5.4. Results and Analysis
We evaluate S UPA with two configurations, S UPA -FS and S UPA -FSCS. S UPA -FS is a
one-stage FS analysis by considering flow-sensitivity only. S UPA -FSCS is a two-stage
analysis consisting of FSCS and FS applied in that order.
5.4.1. Evaluating S UPA -FS. When assessing S UPA -FS, we consider two different criteria: efficiency (its analysis time and memory usage per query) and precision (its competitiveness against SFS). For each query, its analysis budget, denoted B, represents
the maximum number of def-use chains that can be traversed. We consider a wide
range of budgets with B falling into r10, 200000s.
S UPA -FS is highly effectively. With B “ 10000, S UPA -FS is nearly as precise as SFS,
by consuming about 0.18 seconds and 65KB of memory per query, on average.
Efficiency. Figure 14(a) shows the average analysis time per query for all the programs under a given budget, with about 0.18 seconds when B “ 10000 and about 2.76
seconds when B “ 200000. Both axes are logarithmic. The longest-running queries
can take an order of magnitude as long as the average cases. However, most queries
(around 80% across the programs) take much less than the average cases. Take emacs
for example. SFS takes over two hours (8047.55 seconds) to finish. In contrast, S UPA FS spends less than ten minutes (502.10 seconds) when B “ 2000, with an average
per-query time (memory usage) of 0.18 seconds (0.12KB), and produces the same answers for all the queries as SFS (shown in Figure 15 and explained below).
For S UPA, its pre-analysis is lightweight, as shown in Table IV, with vim taking the
longest at 531.57 seconds. The same pre-analysis is also shared by SFS in order to
enable its own sparse whole-program analysis. The additional time taken by SFS for
analyzing each program entirely is given in the last column.
Figure 14(b) shows the average memory usage per query under different budgets. Following the common practice, we measure the real-time memory usage
by reading the virtual memory information (VmSize) from the linux kernel file
Memory usage (KB)
Time per query (secs)
Demand-Driven Flow-Sensitive Pointer Analysis
39:23
10
2.7592
0.8375
0.4138
1
0.0754
0.0334
0.1
0.0074
0.0035
0.01
0.001
0.0007
0.0004
0.0001
0.0002
1
10
0.1836
0.0189
0.0017
102
103
104
(a) Analysis Time
1,000
128.19
100
16.28
6.21
10
1
0.1
0.01
0.44 0.48
0.42
101
2.1592
0.72 0.87
105
237.36
Budget
436.44
364.38
64.80
2.02
0.64
102
103
104
(b) Memory Usage
105
Budget
Fig. 14: Average analysis time and memory usage per query consumed by S UPA -FS
under different analysis budgets (with both axes being logarithmic).
(/proc/self/status). The memory monitor starts after the pre-analysis to measure
the memory usage for answering queries only. The average amount of memory consumed per query is small, with about 65KB when B “ 10000 and about 436KB when
B “ 200000. Even under the largest budget B “ 200000 evaluated, S UPA -FS never uses
more than 3MB for any single query processed.
100 %
80 %
60 %
40 %
20 %
sp
ell
b
mi c
lc
hm less
me
r
s
maed
k
gz e
a2 ip
bis ps
gron
ep
t
wgar
e
b t
gnash
sen ug
dm o
ail
emvim
ac
s
0%
B=200k
B=100k
B=40k
B=20k
B=10k
B=4k
B=2k
B=1k
B=400
B=200
B=100
B=40
B=20
B=10
Fig. 15: Percentage of queried variables proved to be initialized by S UPA -FS over SFS
under different budgets.
39:24
Yulei Sui and Jingling Xue
Table IV: Pre-processing times taken by pre-analysis shared by S UPA and SFS and
analysis times of SFS (in seconds).
Program
spell
bc
milc
less
sed
hmmer
make
gzip
a2ps
bison
grep
tar
wget
bash
gnugo
sendmail
vim
emacs
Pre-Analysis Times
Shared by S UPA and SFS
Andersen’s Analysis SVFG
Total
0.01
0.01
0.01
0.35
0.21
0.56
0.42
0.1
0.52
0.42
0.37
0.79
1.38
0.34
1.73
1.57
0.46
2.03
1.74
1.17
2.91
0.27
0.10
0.37
7.34
1.31
8.65
8.18
3.66
11.84
1.44
0.17
1.61
2.73
1.71
4.44
1.86
0.90
2.76
53.48
44.07
97.55
5.68
2.75
8.44
24.05
23.43
47.48
445.88
85.69 531.57
135.93 146.94 282.87
Analysis Time of SFS
0.01
0.98
0.16
1.94
5.46
1.07
13.94
0.20
60.61
44.16
2.39
12.27
3.47
2590.69
9.86
348.63
13823
8047.55
Precision. Given a query ptpx`, py), p is initialized if no UAO is pointed by p and potentially uninitialized otherwise. We measure the precision of S UPA -FS in terms of the
percentage of queried variables proved to be initialized by comparing with SFS, which
yields the best precision achievable as a whole-program flow-sensitive analysis.
Figure 15 reports our results. As B increases, the precision of S UPA -FS generally
improves. With B “ 10000, S UPA -FS can answer correctly 97.4% of all the queries
from the 18 programs. These results indicate that our analysis is highly accurate,
even under tight budgets. For the 18 programs except a2ps, bison and bash, S UPA FS produces the same answers for all the queries when B “ 100000 as SFS. When
B “ 200000 for these three programs, S UPA becomes as precise as SFS, by taking an
average of 0.02 seconds (88.5KB) for a2ps, 0.25 seconds (194.7KB) for bison, and 3.18
seconds (1139.3KB) for bash, per query.
Understanding On-Demand Strong Updates. Let us examine the benefits achieved
by S UPA -FS in answering client queries by applying on-demand strong updates. For
each program, Figure 16 shows a good correlation between the number of strong updates performed (#SU on the left y-axis) in a blue curve and the number of UAO’s
reaching some uninitialized pointers (#UAO on the right y-axis) in a red curve under
varying budgets (on the logarithmic x-axis). The number of such UAO’s reported by
SFS is shown as the lower bound for S UPA -FS in a dashed line.
In most programs, S UPA -FS performs increasingly more strong updates to block
increasingly more UAO’s to reach the queried variables as the analysis budget B increases, because S UPA -FS falls back increasingly less frequently from FS to the precomputed points-to information. When B increases, S UPA -FS can filter out more spurious value-flows in the SVFG to obtain more precise points-to information, thereby
enabling more strong updates to kill the UAO’s.
Demand-Driven Flow-Sensitive Pointer Analysis
Number of strong updates
spell
#UAO
#SU
6
10
4
5
39:25
Number of UAO by Supa
bc
#SU
80
20
60
15
100
40
10
50
20
5
0
0
200
150
2
0
0
0
101
102
103
104
0
0
105
101
102
Budget
#UAO
#SU
60
15
300
200
40
100
0
0
101
102
103
103
104
4
2
0
101
102
104
5
sed
0
0
0
105
101
102
Budget
103
#UAO
#SU
40
200
104
30
150
20
100
10
50
104
0
0
hmmer
105
100
50
0
0
105
#UAO
150
101
Budget
make
103
Budget
10
20
#UAO
6
0
105
milc
Budget
less
#SU
Number of UAO by SFS
#UAO
#SU
102
103
104
105
Budget
gzip
a2ps
#UAO
#SU
#UAO
#SU
200
40
60
15
1,500
60
150
30
40
10
1,000
40
100
20
50
10
20
5
500
20
0
0
#SU
0
0
0
101
102
103
104
0
0
105
101
102
Budget
104
0
0
105
101
102
Budget
bison
#SU
103
#UAO
1,500
150
1,000
100
103
104
105
Budget
grep
#SU
#UAO
#UAO
#SU
30
600
20
400
10
200
200
tar
#UAO
100
80
150
60
100
500
50
0
0
0
101
102
103
104
50
20
0
0
0
105
101
102
Budget
103
104
#UAO
300
100
200
#UAO
#SU
30
200
300
20
200
0
0
103
104
0
0
0
105
101
102
Budget
#UAO
800
600
103
104
gnugo
105
#UAO
40
150
103
104
20
50
0
0
0
105
101
Budget
sendmail
#SU
102
100
10
100
102
101
Budget
bash
#SU
50
100
101
0
0
105
400
0
0
Budget
wget
#SU
40
103
104
105
Budget
vim
#SU
102
#UAO
emacs
#SU
150
3,000
300
1,500
100
2,000
200
1,000
50
1,000
100
500
#UAO
80
60
400
40
200
0
0
0
101
102
103
Budget
104
105
0
0
0
101
102
103
Budget
104
105
20
0
0
0
101
102
103
104
105
Budget
Fig. 16: Correlating the number of strong updates with the number of UAO’s under
S UPA -FS with different analysis budgets.
When B “ 200000, S UPA -FS gives the same answers as SFS in all the 18 programs
except bison and vim, which causes S UPA -FS to report 16 and 35 more, respectively.
39:26
Yulei Sui and Jingling Xue
For some programs such as spell, bc, milc, hmmer and grep, most of their strong
updates happen under small budgets (e.g., B “ 1000). In hmmer, for example, 192
strong updates are performed when B “ 10000. Of the 5126 queries issued, S UPA -FS
runs out-of-budget for only three queries, which are all fully resolved when B “ 200000,
but with no further strong updates being observed.
For programs like bison, bash, gnugo and emacs, quite a few strong updates take
place when B ą 1000. There are two main reasons. First, these programs have many
indirect call edges (with 8709 in bison, 1286 in bash, 23150 in gnugo and 4708 in
emacs), making their on-the-fly call graph construction costly (Section 4.1.2). Second,
there are many value-flow cycles (with over 50% def-use chains occurring in cycles
in bison), making their constraint resolution costly (to reach a fixed point). Therefore,
relatively large budgets are needed to enable more strong updates to be performed.
Interestingly, in programs such as a2ps, gnugo and vim, fewer strong updates are
observed when larger budgets are used. In vim, the number of strong updates performed is 1492 when B “ 2000 but drops to 1204 when B “ 4000. This is due to the
forward reuse described in Section 4.3. When answering a query ptpx`, vyq under two
budgets B1 and B2 , where B1 ă B2 , S UPA -FS has reached x`1 , v 1 y and needs to compute
ptpx`1 , v 1 yq in each case. S UPA -FS may fall back to the flow-insensitive points-to set of
v 1 under B1 but not B2 , resulting in more strong updates performed under B1 in the
part of the program that is not explored under B2 .
5.4.2. Evaluating S UPA -FSCS. For C programs, flow-sensitivity is regarded as being important for achieving useful high precision. However, context-sensitivity can be important for some C programs, in terms of both obtaining more precise points-to information and enabling more strong updates. Unfortunately, whole-program analysis does
not scale well to large programs when both are considered (Section 5.1).
Table V: Average analysis times consumed and UAO’s reported by S UPA -FSCS (with a
budget of 10000 in each stage) and S UPA -FS (with a budget of 10000 in total).
Program
spell
bc
milc
less
sed
hmmer
make
gzip
a2ps
bison
grep
tar
wget
bash
gnugo
sendmail
vim
emacs
S UPA -FS
Time (ms) #UAO
0.01
0
18.35
69
0.02
3
15.15
37
355.60
32
11.41
86
124.40
26
0.64
5
126.01
34
465.54
94
124.46
14
26.31
70
24.51
104
188.69
17
72.73
28
200.32
94
168.67
218
159.22
45
S UPA -FSCS
Time (ms) #UAO
0.01
0
287.23
69
14.52
0
92.41
37
4725.42
32
135.05
71
229.44
26
4.28
5
448.26
32
529.20
86
197.66
14
83.10
68
84.90
104
327.16
17
80.08
27
250.19
85
473.25
218
222.65
45
Demand-Driven Flow-Sensitive Pointer Analysis
39:27
In this section, we demonstrate that S UPA can exploit both flow- and contextsensitivity effectively on-demand in a hybrid multi-stage analysis framework, providing improved precision needed by some programs. Table V compares S UPA -FSCS (with
a budget of 20000 divided evenly in its FSCS and FS stages) with S UPA -FS (with a
budget of 10000 in its single FS stage). The maximal depth of a context stack allowed
is 3. By allocating the budgets this way, we can investigate some additional precision
benefits achieved by considering both flow- and context-sensitivity.
In general, S UPA -FSCS has longer query response times than S UPA -FS due to the
larger budgets used in our setting and the times taken in handling context-sensitivity.
In milc, hmmer, a2ps, bison, tar , gnugo and sendmail, S UPA -FSCS reports fewer
UAO’s than S UPA -FS, for two reasons. First, S UPA -FSCS can perform strong updates
context-sensitively for stack and global objects, resulting in 0 UAO’s reported by S UPA FSCS for milc. Second, S UPA -FSCS can perform strong updates to context-sensitive
singleton heap objects defined in Section 4.2, by eliminating 8 UAO’s in bison, 1 in
tar and 1 in sendmail, which have been reported by S UPA -FS.
6. CASE STUDIES
114
115
119
120
// symtab.c
static
void symbols_sort(symbol **first, symbol **second) {
...
symbol* tmp = *first;
*first = *second;
121
*second
...
=
tmp;
SU for nd
123
}
623
624
628
static void
user_token_number_redeclaration(...) {
...
Query
symbols_sort (&st, &nd);
pt(⟨ℓ628 ,nd->location⟩)
...
complain_indent (&nd->location, ...);
635
}
627
93
98
104
113
114
115
117
119
120
121
123
125
128
129
130
131
133
135
(a) Code snippet from bison-3.0.4
//io_lat4.c
int qcdhdr_get_str(char *s, QCDheader *hdr, char **q) {
*q = (*hdr).value[i];
SU for q
}
int qcdhdr_get_int(char *s,QCDheader *hdr,int *q) {
char *p;
qcdhdr_get_str(s,hdr,&p);
Query
sscanf(p,"%d",...);
pt(⟨ℓ117 ,p⟩)
}
int qcdhdr_get_int32x(char *s,QCDheader *hdr,...) {
char *p;
qcdhdr_get_str(s,hdr,&p);
sscanf(p,"%x",...);
}
int qcdhdr_get_double(char *s, QCDheader *hdr, ...) {
char *p;
qcdhdr_get_str(s,hdr,&p);
sscanf(p,"%lf",...);
}
(c) Code snippet from milc-v6
// mark.c
68 static struct mark* getmark(int c){
72
register struct mark *m; static struct mark sm;
75
switch (c) {
77
case ’^’:
81
m = &sm;
...
84
m->m_ifile = curr_ifile; SU for sm.m_ifile
85
break;
108
case ’\’’:
112
m = &marks[LASTMARK];
113
break;
127
}
128
return (m);
Query
129 }
pt(⟨ℓ208 ,m->m_ifile⟩)
179 public void gomark(int c) {
186
m = getmark(c);
208
if (m->m_ifile){ ...}
218 }
(b) Code snippet from less-4.5.1
//argp-help.c
434 static struct hol * make_hol (...) {
442
struct hol *hol = malloc (sizeof (struct hol)); // Obj
501
return hol;
SU for Obj.short_options
502 }
849 static void hol_append (struct hol *hol, ...) {
934
hol->short_options = short_options;
939 }
1386 static struct hol * argp_hol (...) {
1390
struct hol *hol = make_hol (argp, cluster);
1401
hol_append(hol, ... );
1405 }
1588 static void _help (...)
1617
hol = argp_hol (argp, 0);
Query
1664
hol_usage (hol, fs);
pt(⟨ℓ1353 ,hol->short_options⟩)
1727 }
1346 static void hol_usage (struct hol *hol, ...) {
1353
strlen(hol->short_options);
1382 }
(d) Code snippet from tar-1.28
Fig. 17: Selected code snippets.
We examine some real code to see how client queries are answered precisely with
on-demand strong updates under four different scenarios.
Figure 17(a). There is a swap from bison. In line 121, second points to a singleton
stack object nd passed from line 627. So a strong update is applied. When querying
nd->location in line 628, S UPA knows that nd points to what st pointed to before.
39:28
Yulei Sui and Jingling Xue
Fig. 18: Speedups of S UPA -FS when parallelized over its sequential version with two,
four and eight threads (B “ 10000q.
Figure 17(b). In the code fragment from less, m->m ifile is initialized in two different branches, one recognized due to a strong update performed at the store in
line 84 and one due to the default initialization in line 112. According to S UPA,
m->m ifile in line 208 is initialized.
Figure 17(c). In the code fragment from milc, q in line 98 can point to several stack
variables that are all named p in lines 115, 123 and 131. With context-sensitivity,
S UPA finds that q points to one singleton under each context. Thus, a strong update
is performed so that each stack variable becomes properly initialized when queried
at each call to sscanf().
Figure 17(d). In the code fragment from tar, hol in line 1390 points to a heap object
o allocated in line 442. With o treated as a context-sensitive singleton (requiring a
context stack of at least depth 1), a strong update can be performed in line 934 to
initialize its field short options properly.
7. PARALLELIZING SUPA
To demonstrate that S UPA is amenable to parallelization as a demand-driven analysis, we have parallelized S UPA -FS by using Intel Threading Building Blocks (TBB).
A concurrent queue is used to store all the queries issued from a program. We use a
task group to allocate tasks for computing the queries from concurrent queue in parallel. The cached points-to information is shared with a concurrent hash map.
Figure 18 shows the speedups achieved by parallelization over the sequential setting with B “ 10000. With eight threads, the average speedup for the 18 programs
is 3.05x and the maximum speedup observed is 6.9x at a2ps. The time for each setting excludes the pre-analysis time. Some programs enjoy better speedups than others.
There are three main reasons. First, some programs, such as spell, less and milc,
have relatively few queries issued. Therefore, the performance benefits achieved from
query parallelization can be small. Second, different queries take different times to
answer, resulting in different degrees of workload imbalance in different programs.
Third, different programs suffer from different synchronization overheads in accessing the cached points-to information in concurrent hash map.
8. RELATED WORK
Demand-driven and whole-program approaches represent two important solutions
to long-standing pointer analysis problems. While a whole-program pointer analysis
aims to resolve all the pointers in the program, a demand-driven pointer analysis is
designed to resolve only a (typically small) subset of the set of these pointers in a
Demand-Driven Flow-Sensitive Pointer Analysis
39:29
client-specific manner. This work is not concerned with developing an ultra-fast wholeprogram pointer analysis. Rather, our objective is to design a staged demand-driven
strong update analysis framework that facilitates efficiency and precision tradeoffs
flow- and context-sensitively according to the needs of a client (e.g., user-specified budgets). Below we limit our discussion to the work that is most relevant to S UPA.
8.1. Flow-Sensitive Pointer Analysis
Strong updates require pointers to be analyzed flow-sensitively with respect to program execution order. Whole-program flow-sensitive pointer analysis has been studied extensively in the literature. Choi et al. [1993] and Emami and Hendren [1994]
gave some formulations in an iterative data-flow framework [Kam and Ullman 1977].
Wilson and Lam [1995] considered both flow- and context-sensitivity by representing
procedure summaries with partial transfer functions, but restricted strong updates to
top-level variables only. To eliminate unnecessary propagation of points-to information
during the iterative data-flow analysis [Hardekopf and Lin 2009, 2011; Oh et al. 2012;
Yu et al. 2010], some form of sparsity has been exploited. The sparse value-flows, i.e.,
def-use chains in a program are captured by sparse evaluation graphs (SEG) [Choi
et al. 1991; Ramalingam 2002] as in [Hind and Pioli 1998] and various SSA representations such as HSSA [Chow et al. 1996], partial SSA [Lattner and Adve 2004]
and SSI [Ananian 1999; Tavares et al. 2014]. The def-use chains for top-level pointers, once put in SSA, can be explicitly and precisely identified, giving rise to a so-called
semi-sparse flow-sensitive analysis [Hardekopf and Lin 2009]. Later, the idea of staged
analysis [Fink et al. 2008] has been leveraged to make pointer analysis full-sparse for
both top-level and address-taken variables by using fast Andersen’s analysis as precise analysis [Hardekopf and Lin 2011; Sui et al. 2016a; Ye et al. 2014b]. This paper is
the first to exploit sparsity to improve the performance of a flow- and context-sensitive
demand-driven analysis with strong updates being performed for C programs.
Recently, Balatsouras and Smaragdakis [Balatsouras and Smaragdakis 2016] propose a fine-grained field-sensitive modeling technique for performing Andersen’s analysis by inferring lazily the types of heap objects in order to filter out redundant field
derivations. This technique can be exploited to obtain a more precise pre-analysis to
improve the precision and/or efficiency of sparse flow-sensitive analysis.
8.2. Demand-Driven Pointer Analysis
Demand-driven pointer analyses for C [Heintze and Tardieu 2001; Zhang et al. 2014a;
Zheng and Rugina 2008] and Java [Lu et al. 2013; Shang et al. 2012; Sridharan
and Bodı́k 2006; Su et al. 2016; Yan et al. 2011] are flow-insensitive, formulated
in terms of CFL (Context-Free-Language) reachability [Reps et al. 1995]. Heintze
and Tardieu [2001] introduced the first on-demand Andersen-style pointer analysis for C. Later, Zheng and Rugina [2008] performed alias analysis for C in terms
of CFL-reachability flow- and context-insensitively with indirect function calls handled conservatively. Sridharan et al. gave two CFL-reachability-based formulations
for Java, initially without considering context-sensitivity [Sridharan et al. 2005] and
later with context-sensitivity [Sridharan and Bodı́k 2006]. Shang et al. [2012] and
Yan et al. [2011] investigated how to summarize points-to information discovered
during the CFL-reachability analysis to improve performance for Java programs.
Lu et al. [2013] introduced an incremental pointer analysis with a CFL-reachability
formulation for Java. Su et al. [2014] demonstrated that the CFL-reachability formulation is highly amenable to parallelization on multi-core CPUs. Recently, Feng
et al. [2015] focused on answering demand queries for Java programs in a contextsensitive analysis framework (without performing strong updates). Unlike these flow-
39:30
Yulei Sui and Jingling Xue
insensitive analyses, which are not effective for many clients like Uninit, S UPA can
perform strong updates on-demand flow and context-sensitively.
B OOMERANG [Späth et al. 2016] represents a recent flow- and context-sensitive
demand-driven pointer analysis for Java. However, its access-path-based analysis performs only strong updates partially at a store a.f “ . . . , by updating a.f strongly but
the aliases of a.f.˚ weakly, where a and b are different top-level variables. Let us explain this by using the following straight-line Java code and its corresponding C code.
`1 :
`2 :
`3 :
`4 :
`5 :
Java Code
q = new A()
p=q
p.f = new A()
q.f = new A()
x = p.f
// o1
// o2
// o3
`1 :
`2 :
`3 :
`4 :
`5 :
C Code
q = malloc()
p=q
*p = malloc()
*q = malloc()
x = *p
// o1
// o2
// o3
Let us consider B OOMERANG first. At `3 , a strong update is performed to p.f to make
it point to o2 only. At `4 , a strong update is performed to q.f to make it point to o3 but
a weak update is performed to all its aliases so that p.f now points to not only o2 as
before but also o3, As a result, x points-to both o2 and o3 at `5 . Let us consider now
S UPA. With both flow- and context-sensitivity enforced, a strong update is performed
to o1 pointed p and q at both `3 and `4 , respectively. Thus, x points to o3 only at `5 .
8.3. Hybrid Pointer Analysis
The basic idea is to find a right balance between efficiency and precision. For C programs, the one-level approach [Das 2000] achieves a precision between Steensgaard’s
and Andersen’s analyses by applying a unification process to address-taken variables
only. In the case of Java programs, context-sensitivity can be made more effective
by considering both call-site-sensitivity and object-sensitivity together than either
alone [Kastrinis and Smaragdakis 2013]. In [Guyer and Lin 2003], how to adjust the
analysis precision according to a client’s needs is discussed. Zhang et al. [2014b] focus
on finding effective abstractions for whole-program analyses written in Datalog via
abstraction refinement. Lhoták and Chung [Lhoták and Chung 2011] trades precision
for efficiency by performing strong updates only on flow-sensitive singleton objects but
falls back to the flow-insensitive points-to information otherwise. In this paper, we
propose to carry out our on-demand strong update analysis in a hybrid multi-stage
analysis framework. Unlike [Lhoták and Chung 2011], S UPA can achieve the same
precision as whole-program flow-sensitive analysis, subject to a given budget.
8.4. Parallel Pointer Analysis
Méndez-Lojo et al. [2010] introduced a parallel implementation of Andersen’s analysis
for C based on graph rewriting. Their parallel analysis is flow- and context-insensitive,
achieving a speedup of up to 3X on an 8-core CPU. Su et al. [2016] introduces an
improvement of this parallel implementation on GPUs. The whole-program sparse
flow-sensitive pointer analysis [Hardekopf and Lin 2009] has also been parallelized
on multi-core CPUs [Nagaraj and Govindarajan 2013] and GPUs [Nasre 2013]. The
speedups are up to 2.6X on a 8-core CPU. This paper presents the first parallel implementation of demand-driven pointer analysis with strong updates for C programs,
achieving an average speedup of 3.05X on a 8-core CPU.
9. CONCLUSION
We have introduced, S UPA, a demand-driven pointer analysis that enables computing
precise points-to information for C programs flow- and context-sensitively with strong
Demand-Driven Flow-Sensitive Pointer Analysis
39:31
updates by refining away imprecisely pre-computed value-flows, subject to some analysis budgets. S UPA handles large C programs effectively by allowing pointer analyses
with different efficiency and precision tradeoffs to be applied in a hybrid multi-stage
analysis framework. S UPA is particularly suitable for environments with small time
and memory budgets such as IDEs. We have evaluated S UPA by choosing uninitialized
pointer detection as a major client on 18 C programs. S UPA can achieve nearly the
same precision as whole-program flow-sensitive analysis under small budgets.
One interesting future work is to investigate how to allocate budgets in S UPA to its
stages to improve the precision achieved in answering some time-consuming queries
for a particular client. Another direction is to add more stages to its analysis, by considering, for example, path correlations.
REFERENCES
Acharya, M. and Robinson, B. (2011). Practical change impact analysis based on static
program slicing for industrial software systems. In ICSE ’11, pages 746–755.
Ananian, C. S. (1999). The static single information form. PhD thesis, Master’s Thesis,
MIT.
Andersen, L. (1994). Program analysis and specialization for the C programming language. PhD thesis, DIKU, University of Copenhagen.
Arzt, S., Rasthofer, S., Fritz, C., Bodden, E., Bartel, A., Klein, J., Le Traon, Y., Octeau,
D., and McDaniel, P. (2014). Flowdroid: Precise context, flow, field, object-sensitive
and lifecycle-aware taint analysis for android apps. In PLDI ’14, pages 259–269.
Balatsouras, G. and Smaragdakis, Y. (2016). Structure-sensitive points-to analysis for
c and c++. In SAS ’16.
Blackshear, S., Chang, B.-Y. E., and Sridharan, M. (2013). Thresher: Precise refutations for heap reachability. In PLDI ’13, pages 275–286.
Choi, J.-D., Burke, M., and Carini, P. (1993). Efficient flow-sensitive interprocedural
computation of pointer-induced aliases and side effects. In POPL ’93, pages 232–245.
Choi, J.-D., Cytron, R., and Ferrante, J. (1991). Automatic construction of sparse data
flow evaluation graphs. In POPL ’91, pages 55–66.
Chow, F., Chan, S., Liu, S., Lo, R., and Streich, M. (1996). Effective representation of
aliases and indirect memory operations in SSA form. In CC ’96, pages 253–267.
Cytron, R., Ferrante, J., Rosen, B., Wegman, M., and Zadeck, F. (1991). Efficiently computing static single assignment form and the control dependence graph.
TOPLAS’91, 13(4):490.
Das, M. (2000). Unification-based pointer analysis with directional assignments. In
PLDI ’00, pages 35–46.
Emami, M. Ghiya, R. and Hendren, J. (1994). Context-sensitive interprocedural
points-to analysis in presence of function pointers. In PLDI ’94, pages 242–256.
Feng, Y., Wang, X., Dillig, I., and Lin, C. (2015). EXPLORER: query- and demanddriven exploration of interprocedural control flow properties. In OOPSLA ’15, pages
520–534.
Fink, S. J., Yahav, E., Dor, N., Ramalingam, G., and Geay, E. (2008). Effective typestate
verification in the presence of aliasing. ACM TOSEM, 17(2):9.
Guyer, S. Z. and Lin, C. (2003). Client-driven pointer analysis. In SAS ’03, pages
1073–1073.
Hardekopf, B. and Lin, C. (2009). Semi-sparse flow-sensitive pointer analysis. In POPL
’09, pages 226–238.
Hardekopf, B. and Lin, C. (2011). Flow-Sensitive Pointer Analysis for Millions of Lines
of Code. In CGO ’11, pages 289–298.
39:32
Yulei Sui and Jingling Xue
Heintze, N. and Tardieu, O. (2001). Demand-driven pointer analysis. In PLDI ’01,
pages 24–34.
Hind, M. and Pioli, A. (1998). Assessing the effects of flow-sensitivity on pointer alias
analyses. In SAS ’98, pages 57–81.
ISO90 (1990). ISO/IEC. international standard ISO/IEC 9899, programming languages - C.
Kam, J. B. and Ullman, J. D. (1977). Monotone data flow analysis frameworks. Acta
Informatica, 7(3):305–317.
Kastrinis, G. and Smaragdakis, Y. (2013). Hybrid context-sensitivity for points-to analysis. In PLDI ’13, pages 423–434.
Landi, W. (1992). Undecidability of static analysis. ACM Letters on Programming
Languages and Systems (LOPLAS), 1(4):323–337.
Lattner, C. and Adve, V. (2004). LLVM: A compilation framework for lifelong program
analysis & transformation. In CGO ’04, pages 75–86.
Lhoták, O. and Chung, K.-C. A. (2011). Points-to analysis with efficient strong updates.
In POPL ’11, pages 3–16.
Lhoták, O. and Hendren, L. (2003). Scaling Java points-to analysis using Spark. CC
’03, pages 153 – 169.
Li, L., Cifuentes, C., and Keynes, N. (2011). Boosting the performance of flow-sensitive
points-to analysis using value flow. In FSE ’11, pages 343–353.
Li, Y., Tan, T., Sui, Y., and Xue, J. (2014). Self-inferencing reflection resolution for Java.
In ECOOP ’14, pages 27–53.
Lu, Y., Shang, L., Xie, X., and Xue, J. (2013). An incremental points-to analysis with
CFL-reachability. In CC’13.
Méndez-Lojo, M., Mathew, A., and Pingali, K. (2010). Parallel inclusion-based points-to
analysis. In OOPSLA ’10, pages 428–443.
Milanova, A., Rountev, A., and Ryder, B. G. (2002). Parameterized object sensitivity
for points-to and side-effect analyses for Java. ISSTA ’02.
Milanova, A., Rountev, A., and Ryder, B. G. (2005). Parameterized object sensitivity
for points-to analysis for Java. ACM Trans. Softw. Eng. Methodol., 14(1):1–41.
Nagaraj, V. and Govindarajan, R. (2013). Parallel flow-sensitive pointer analysis by
graph-rewriting. In PACT ’13, pages 19–28.
Nasre, R. (2013). Time- and space-efficient flow-sensitive points-to analysis. TACO
’13, 10(4):39:1–39:27.
Oh, H., Heo, K., Lee, W., Lee, W., and Yi, K. (2012). Design and implementation of
sparse global analyses for C-like languages. In PLDI ’12, pages 229–238.
Pearce, D., Kelly, P., and Hankin, C. (2007). Efficient field-sensitive pointer analysis of
C. ACM TOPLAS, 30(1):4–es.
Ramalingam, G. (1994). The undecidability of aliasing. ACM TOPLAS, 16(5):1467–
1471.
Ramalingam, G. (2002). On sparse evaluation representations. Theoretical Computer
Science, 277(1):119–147.
Reps, T., Horwitz, S., and Sagiv, M. (1995). Precise interprocedural dataflow analysis
via graph reachability. In POPL ’95, pages 49–61.
Shang, L., Xie, X., and Xue, J. (2012). On-demand dynamic summary-based points-to
analysis. In CGO ’12, pages 264–274.
Smaragdakis, Y., Bravenboer, M., and Lhoták, O. (2011). Pick your contexts well: understanding object-sensitivity. In POPL’11, pages 17–30.
Späth, J., Do, L. N. Q., Ali, K., and Bodden, E. (2016). Boomerang: Demand-driven
flow-and context-sensitive pointer analysis for java. ECOOP.
Sridharan, M. and Bodı́k, R. (2006). Refinement-based context-sensitive points-to
analysis for Java. PLDI ’06, pages 387–400.
Demand-Driven Flow-Sensitive Pointer Analysis
39:33
Sridharan, M., Gopan, D., Shan, L., and Bodı́k, R. (2005). Demand-driven points-to
analysis for Java. In OOPSLA ’05, pages 59–76.
Su, Y., Ye, D., and Xue, J. (2014). Parallel pointer analysis with cfl-reachability. In
ICPP ’14, pages 451–460.
Su, Y., Ye, D., Xue, J., and Liao, X. (2016). An efficient GPU implementation of
inclusion-based pointer analysis. IEEE Trans. Parallel Distrib. Syst., 27(2):353–366.
Sui, Y., Di, P., and Xue, J. (2016a). Sparse flow-sensitive pointer analysis for multithreaded programs. In CGO ’16, pages 160–170. ACM.
Sui, Y., Fan, X., Zhou, H., and Xue, J. (2016b). Loop-oriented array-and field-sensitive
pointer analysis for automatic simd vectorization. In LCTES ’16, pages 41–51. ACM.
Sui, Y., Li, Y., and Xue, J. (2013). Query-directed adaptive heap cloning for optimizing
compilers. In CGO ’13, pages 1–11.
Sui, Y. and Xue, J. (2016). SVF: Interprocedural static value-flow analysis in LLVM.
In CC ’16, pages 265–266.
Sui, Y., Ye, D., and Xue, J. (2012). Static memory leak detection using full-sparse
value-flow analysis. In ISSTA ’12, pages 254–264.
Sui, Y., Ye, D., and Xue, J. (2014a). Detecting memory leaks statically with full-sparse
value-flow analysis. TSE ’14, 40(2):107–122.
Sui, Y., Ye, S., Xue, J., and Zhang, J. (2014b). Making context-sensitive inclusion-based
pointer analysis practical for compilers using parameterised summarisation. SPE
’14, 44(12):1485–1510.
Sun, Q., Zhao, J., and Chen, Y. (2011). Probabilistic points-to analysis for Java. In CC
’11, pages 62–81.
SUPA (2016). SUPA. http://www.cse.unsw.edu.au/„corg/supa.
Tavares, A., Boissinot, B., Pereira, F., and Rastello, F. (2014). Parameterized construction of program representations for sparse dataflow analyses. In CC ’14, pages 18–
39. Springer.
Wilson, R. and Lam, M. (1995). Efficient context-sensitive pointer analysis for C programs. PLDI ’95, pages 1–12.
Xiao, X. and Zhang, C. (2011). Geometric encoding: forging the high performance context sensitive points-to analysis for Java. In ISSTA ’11, pages 188–198.
Yan, D., Xu, G., and Rountev, A. (2011). Demand-driven context-sensitive alias analysis
for Java. In ISSTA ’11, pages 155–165.
Yan, H., Sui, Y., Chen, S., and Xue, J. (2016). Automated memory leak fixing on valueflow slices for c programs. In Proceedings of the 31st Annual ACM Symposium on
Applied Computing, SAC ’16, pages 1386–1393, New York, NY, USA. ACM.
Ye, D., Sui, Y., and Xue, J. (2014a). Accelerating dynamic detection of uses of undefined
variables with static value-flow analysis. In CGO ’14.
Ye, S., Sui, Y., and Xue, J. (2014b). Region-based selective flow-sensitive pointer analysis. In SAS ’14, pages 319–336.
Yu, H., Xue, J., Huo, W., Feng, X., and Zhang, Z. (2010). Level by level: making flowand context-sensitive pointer analysis scalable for millions of lines of code. In CGO
’10, pages 218–229.
Zhang, Q., Xiao, X., Zhang, C., Yuan, H., and Su, Z. (2014a). Efficient subcubic alias
analysis for C. In PLDI ’14, pages 829–845.
Zhang, X., Mangal, R., Grigore, R., Naik, M., and Yang, H. (2014b). On abstraction
refinement for program analyses in Datalog. In PLDI ’14, pages 239–248.
Zhao, J., Nagarakatte, S., Martin, M. M., and Zdancewic, S. (2012). Formalizing the
LLVM intermediate representation for verified program transformations. In POPL
’12, pages 427–440.
Zheng, X. and Rugina, R. (2008). Demand-driven alias analysis for C. In POPL ’08,
pages 197–208.
| 6cs.PL
|
cuDNN: Efficient Primitives for Deep Learning
arXiv:1410.0759v3 [cs.NE] 18 Dec 2014
Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran
NVIDIA
Santa Clara, CA 95050
{schetlur, jwoolley, philippev, jocohen, johntran}@nvidia.com
Bryan Catanzaro
Baidu Research
Sunnyvale, CA 94089
[email protected]
Evan Shelhamer
UC Berkeley
Berkeley, CA 94720
[email protected]
Abstract
We present a library of efficient implementations of deep learning primitives.
Deep learning workloads are computationally intensive, and optimizing their kernels is difficult and time-consuming. As parallel architectures evolve, kernels must
be reoptimized, which makes maintaining codebases difficult over time. Similar
issues have long been addressed in the HPC community by libraries such as the
Basic Linear Algebra Subroutines (BLAS) [2]. However, there is no analogous
library for deep learning. Without such a library, researchers implementing deep
learning workloads on parallel processors must create and optimize their own implementations of the main computational kernels, and this work must be repeated
as new parallel processors emerge. To address this problem, we have created
a library similar in intent to BLAS, with optimized routines for deep learning
workloads. Our implementation contains routines for GPUs, although similarly
to the BLAS library, these routines could be implemented for other platforms.
The library is easy to integrate into existing frameworks, and provides optimized
performance and memory usage. For example, integrating cuDNN into Caffe, a
popular framework for convolutional networks, improves performance by 36% on
a standard model while also reducing memory consumption.
1
Introduction
Deep neural networks have been successful at solving many kinds of tasks [4]. Parallel processors
such as GPUs have played a significant role in the practical implementation of deep neural networks. The computations that arise when training and using deep neural networks lend themselves
naturally to efficient parallel implementations. The efficiency provided by these implementations allows researchers to explore significantly higher capacity networks, training them on larger datasets
[7]. This has led to greatly improved accuracy on tasks such as speech recognition and image classification, among others. For example, the majority of entries in the 2014 ILSVRC challenge [17]
use GPUs to implement deep neural networks, as pioneered by [13], and improved by many others,
including [18] [19]. Deep neural networks for speech recognition have also benefited from parallel
implementations on GPUs [10] [9] [15].
Convolutional Neural Networks (CNNs) [14] are an important and successful class of deep networks. Convolutional neural networks are computed using dense kernels that differ from traditional
dense linear algebra routines. Accordingly, modern deep learning frameworks, such as Torch7 [8],
Theano [5], and Caffe [11] feature suites of custom kernels that implement basic operations such
as tensor convolutions, activation functions and pooling. These routines represent the bulk of the
1
computation when training a CNN, and thus account for the majority of its execution time. The deep
learning community has been successful in finding optimized implementations of these kernels, but
as the underlying architectures evolve, these kernels must be re-optimized, which is a significant
investment. Optimizing these kernels requires a deep understanding of the underlying processor architecture, with careful scheduling of data movement, on-chip memory placement, register blocking,
and other optimizations in order to get acceptable performance.
We believe that providing a library of optimized routines for these computations will provide several important benefits. Firstly, deep learning frameworks can focus on higher-level issues rather
than close optimization of parallel kernels to specific hardware platforms. Secondly, as parallel
architectures evolve, library providers can provide performance portability, in much the same way
as the BLAS routines provide performance portability to diverse applications on diverse hardware.
Thirdly, a clearer separation of concerns allows specialization: library providers can take advantage
of their deep understanding of parallel architectures to provide optimal efficiency. Our goal is to
make it much easier for deep learning frameworks to take advantage of parallel hardware.
Our library addresses this goal by providing a flexible, easy-to-use C-language API for deep learning
workloads that integrates neatly into existing frameworks. It can provide immediate efficiency gains,
and it is rigorously tested and maintained in order to be reliable and performant across a range of
different processor architectures. Importantly, our library is designed to use the minimum possible
amount of auxiliary memory, which frees up scarce memory for larger models and datasets. We also
optimize performance across a wide range of potential use cases, including small mini-batch sizes.
2
Library
One of the primary goals of cuDNN is to enable the community of neural network frameworks to
benefit equally from its APIs. Accordingly, users of cuDNN are not required to adopt any particular
software framework, or even data layout.
Rather than providing a layer abstraction, we provide lower-level computational primitives, in order
to simplify integration with existing deep learning frameworks, each with their own abstractions.
The bulk of the API is dedicated to functions that perform primitive operations on data stored in
user-controlled buffers. By keeping the API low-level, the library integrates simply into other frameworks.
cuDNN supports forward and backward propagation variants of all its routines in single and double
precision floating-point arithmetic. These include convolution, pooling and activation functions.
The library allows variable data layout and strides, as well as indexing of sub-sections of input
images. It also includes a set of auxiliary tensor transformation routines that allow for the easy
manipulation of 4d-tensors.
2.1
Overview and Handles
The library exposes a host-callable C language API, but requires that input and output data be resident on the GPU, analogously to cuBLAS. The library is thread-safe and its routines can be called
from different host threads. Convolutional routines for the forward and backward passes use a common descriptor that encapsulates the attributes of the layer. Tensors and Filters are represented in
opaque descriptors, with the flexibility to specify the tensor layout using arbitrary strides along each
dimension for tensors.
2.1.1
Spatial Convolutions
The most important computational primitive in convolutional neural networks is a special form of
batched convolution. The parameters governing this convolution are listed in table 1. In this section,
we describe the forward form of this convolution - the other forms necessary for backpropagation
are closely related.
There are two inputs to the convolution: D ∈ RN CHW , which forms the input data, and F ∈
RKCRS , which forms the convolutional filters. The input data ranges over N images in a minibatch, C input feature maps, H rows per image, and W columns per image. The filters range over
2
Parameter
N
C
H
W
K
R
S
u
v
pad h
pad w
Meaning
Number of images in mini-batch
Number of input feature maps
Height of input image
Width of input image
Number of output feature maps
Height of filter kernel
Width of filter kernel
Vertical stride
Horizontal stride
Height of zero-padding
Width of zero-padding
Table 1: Convolutional parameters
K output feature maps, C input feature maps, R rows per filter, and S columns per filter. The output
is also a four-dimensional tensor O ∈ RN KP Q , where N and K were as defined previously, and
P = f (H, R, u, pad h), Q = f (W, S, v, pad w), meaning that the height and width of the output
images depends on the image and filter height and width, along with padding and striding choices.
The striding parameters u and v allow the user to reduce the computational load by computing
only a subset of the output pixels. The padding parameters allow users to specify how many rows or
columns of 0 entries are appended to each image. MATLAB’s “valid” convolution mode corresponds
to setting pad h = 0, pad w = 0, while MATLAB’s “same” convolution mode corresponds to
pad h = R2 , pad w = S2 , and MATLAB’s “full” convolution mode corresponds to pad h =
R − 1, pad w = S − 1.
More specifically,
H − R + 1 + 2pad h
f (H, R, u, pad h) =
u
(1)
Define an accessing function to account for striding, padding, and inverting for the convolution:
g(p, u, R, r, pad h) = p · u + R − r − 1 − pad h
(2)
Then, the forward convolution evaluates O[n, k, p, q] ∀n ∈ [0, N ), ∀k ∈ [0, K), ∀p ∈ [0, P ), ∀q ∈
[0, Q). For convenience, define D0 as a zero-extended version of D.
O[n, k, p, q] =
C−1
X R−1
X S−1
X
F [k, c, r, s] · D0 [n, c, g(p, u, R, r, pad h), g(q, v, S, s, pad w)]
(3)
c=0 r=0 s=0
As can be seen from Equation 3, computing the convolution involves a seven-way nested loop, with
four independent loops and three accumulation loops. There are many ways of implementing this
computation, some of which we will discuss in the next section.
cuDNN’s convolutional routines incorporate implementations of both the convolution as well as the
cross-correlation variants of these functions. These functions support user-defined strides along each
dimension of the input and output tensors. This is important because different frameworks store tensors using different memory layouts; for example, some frameworks interleave feature maps, while
others keep them separate. cuDNN allows the user to specify memory layout, which makes it much
simpler to integrate into existing frameworks. cuDNN’s routines also have a mode to either return
the raw gradients or to accumulate them in a buffer as needed for models with shared parameters or
a directed acyclic graph structure.
2.2
Other functions
cuDNN also provides other commonly used functions for deep learning. For example, it provides
three commonly used neuron activation functions; Sigmoid, Rectified Linear and Hyperbolic Tangent. It provides a softmax routine, which by default uses the numerically stable approach of scaling
each element to avoid overflow in intermediate results. Softmax may be computed per image across
3
the feature map, height and width dimensions or per spatial location, per image across the feature
map dimension. cuDNN provides average and max pooling operations, as well as a set of tensor
transformation routines, such as those that add tensors, with optional broadcasting. The goal of
providing these functions is to reduce the amount of parallel code that is required for deep learning
frameworks, by providing flexible, well-optimized versions of these commonly used functions. With
cuDNN, it is possible to write programs that train standard convolutional neural networks without
writing any parallel code, but simply using cuDNN and cuBLAS.
3
Implementation
The majority of functions that cuDNN provides have straightforward implementations. The convolution implementation is not as obvious, so we will outline the motivation and reasoning behind our
design choices.
There are several ways to implement convolutions efficiently. Our goal is to provide performance
as close as possible to matrix multiplication, while using no auxiliary memory. GPU memory is
high bandwidth, but low capacity, and is therefore a scarce resource. When training deep networks,
ideally the GPU memory should be filled with data, parameters, and neuron responses, not auxiliary
data structures needed by the convolution algorithm. Several approaches to computing convolutions require large auxiliary data structures, and therefore we do not consider these approaches for
cuDNN.
One approach is to lower the convolutions into a matrix multiplication, following [6]. This can be
done by reshaping the filter tensor F into a matrix Fm with dimensions K × CRS, and gathering a
data matrix by duplicating the original input data into a matrix Dm with dimensions CRS × N P Q.
The computation can then be performed with a single matrix multiply to form an output matrix Om
with dimension K × N P Q.
Image
data
D0
D1
D2
D0
D1
D2
D0
D1
D2
D4
D5
D7
D8
D3
D4
D5
D3
D4
D5
D3
D4
D5
D3
D4
D6
D7
D6
D7
D8
D6
D7
D8
D6
D7
D8
D1
D2
D4
D5
D0
D1
D3
D4
D4
D5
D7
D8
D3
D4
D6
D7
D1
D2
D4
D5
D0
D1
D3
D4
D4
D5
D7
D8
D3
D4
D6
D7
D1
D2
D4
D5
D0
D1
D3
D4
𝑫[𝟎,𝟎,:,:]
𝑫[𝟎,𝟏,:,:]
Filter
data
F0
F1
F0
F1
F0
F1
F2
F3
F2
F3
F2
F3
𝑭[𝟎,:,:,:
]
G0
G1
G0
G1
G0
G1
G2
G3
G2
G3
G2
G3
𝑭[𝟏,:,:,:]
𝑫[𝟎,𝟐,:,:]
N
=
1
C
=
3
H
=
3
W
=
3
K
=
2
R
=
2
S
=
2
u=v
=
1
pad_h
=
0
pad_w
=
0
F0
F1
F2
F3
F0
F1
F2
F3
F0
F1
F2
F3
G0
G1
G2
G3
G0
G1
G2
G3
G0
G1
G2
G3
Fm
Om
Figure 1: Convolution lowering
Figure 1 illustrates how a simple convolution can be lowered to a matrix multiplication. The colors
in this illustration represent the input feature maps, and elements of D and F are uniquely labeled in
the illustration so as to show how each participates in forming Dm and Fm . The filter matrix Fm has
dimensions K × CRS = 2 × 12, while the data matrix Dm has dimensions CRS × N P Q = 12 × 4.
Note that each element of D is duplicated up to RS = 4 times in Dm . The output matrix Om has
dimensions K × N P Q = 2 × 4.
4
Lowering convolutions to matrix multiplication can be efficient, since matrix multiplication is highly
optimized. Matrix multiplication is fast because it has a high ratio of floating-point operations
per byte of data transferred. This ratio increases as the matrices get larger, meaning that matrix
multiplication is less efficient on small matrices. Accordingly, this approach to convolution is most
effective when it creates large matrices for multiplication. As we mentioned earlier, the sizes of Dm
and Fm depend on products of the parameters to the convolution, not the parameters themselves.
This means that performance using this approach can be very consistent, since the algorithm does
not care if one of the parameters is small, as long as the product is large enough. For example,
it is often true that in early layers of a convolutional network, C is small, but R and S are large,
while at the end of the network, C is large, but R and S are small. However, the product CRS
is usually fairly large for all layers, so performance can be consistently good. The disadvantage of
this approach is that forming Dm involves duplicating the input data up to RS times, which can
require a prohibitively large temporary allocation. To work around this, implementations sometimes
materialize Dm piece by piece, for example, by calling matrix multiplication iteratively for each
element of the mini-batch. However, this limits the parallelism in the implementation, and can
lead to cases where the matrix multiplications are too small to effectively utilize the GPU. This
approach also lowers the computational intensity of the convolutions, because Dm must be written
and read, in addition to reading D itself, requiring significantly more memory traffic as a more direct
approach. Accordingly, we opt not to use this implementation directly, although as we will explain,
our implementation is related.
Another approach is to use the Fast Fourier Transform to compute the convolution. The FFT can
significantly lower the work complexity of the convolutions, and with clever engineering can be
used effectively for deep neural networks [16]. However, the FFT based approach uses a significant
amount of temporary memory, since the filters must be padded to be the same size as the inputs.
This is especially costly when the filters are small compared to the images, which often happens in
the first few layers of a convolutional network. Additionally, the FFT based approach does not perform efficiently when the striding parameters u and v are greater than 1, which is common in many
state-of-the art networks, such as the early layers in [18] and [19]. Striding reduces the computational work of the convolutions by a factor of uv, by computing only a sparse subset of the output.
However, the nature of the FFT algorithm is such that computing pruned FFTs is a non-trivial task,
and often is slower than computing the dense FFT, followed by an additional subsampling step. Due
to these drawbacks, we opted to forgo the FFT approach, although we agree that in some cases it is
useful.
Another common approach is to compute the convolutions directly. This can be very efficient, but
requires a large number of specialized implementations to handle the many corner cases implicit in
the 11-dimensional parameter space of the convolutions. Implementations following this approach
are often well-optimized for convolutions in certain parts of the parameter space, but perform poorly
for others. For example, cuda-convnet2 [12] performs well when batch sizes are large, but poorly
once batch size falls to 64 or below. Optimizing and maintaining all these specializations is a difficult
task. As we envision this library being maintained for some time, and being ported to yet-to-beconceived future architectures, we searched for something simpler that would perform more robustly
across the parameter space and be easier to port to new architectures.
3.1
Our approach
NVIDIA provides a matrix multiplication routine that achieves a substantial fraction of floatingpoint throughput on GPUs. The algorithm for this routine is similar to the algorithm described in
[20]. Fixed sized submatrices of the input matrices A and B are successively read into on-chip
memory and are then used to compute a submatrix of the output matrix C. We compute on tiles of A
and B while fetching the next tiles of A and B from off-chip memory into on-chip caches and other
memories. This technique hides the memory latency associated with the data transfer, allowing the
matrix multiplication computation to be limited only by the time it takes to perform the arithmetic.
As we discussed earlier, convolutions can be lowered onto matrix multiplication. This approach
provides simplicity of implementation as well as consistency of performance across the parameter
space, although materializing the lowered matrix in memory can be costly. Our solution follows
this approach, but we avoid the problems with materializing the lowered matrix in memory by lazily
materializing Dm in on-chip memory only, rather than by materializing it in off-chip memory before
5
1.80
1.60
1.40
TFlops
1.20
1.00
cuDNN
0.80
CAFFE
0.60
cuda-‐convnet2
0.40
0.20
-‐
0
32
64
96
128
minibatch
size
Figure 2: Comparative Performance
Layer 1
Layer 2
Layer 3
Layer 4
Layer 5
N
128
128
128
128
128
C
3
96
128
128
128
H
128
64
32
16
13
W
128
64
32
16
13
K
96
128
128
128
384
R
11
9
9
7
3
S
11
9
9
7
3
u
1
1
1
1
1
v
1
1
1
1
1
pad h
0
0
0
0
0
pad w
0
0
0
0
0
Table 2: Convolutional layer collection
calling a matrix multiplication routine. Since the tiling required for the matrix multiplication routine
is independent of any parameters from the convolution, the mapping between the tile boundaries of
Dm and the convolution problem is non-trivial. Accordingly, our approach entails computing this
mapping and using it to load the correct elements of A and B into on-chip memories. This happens
dynamically as the computation proceeds, which allows our convolution algorithm to exploit optimized infrastructure for matrix multiplication. We require additional indexing arithmetic compared
to a matrix multiplication, but fully leverage the computational engine of matrix multiplication to
perform the work. After the computation is complete, we perform the required tensor transposition
to store the result in the user’s desired data layout.
Computing the additional indexing requires repeated calculations of integer division and modulus
operations by launch-time constant divisors. We make use of the algorithm for integer division and
modulus presented in [21] to transform these costly operations into integer multiplies and shifts, and
thus reduce the indexing overhead required by our approach.
3.2
Performance
The convolution routines in cuDNN provide competitive performance with zero auxiliary memory
required.
Figure 2 shows the performance on an NVIDIA Tesla K40 of three convolution implementations:
cuDNN, Caffe, and cuda-convnet2. We evaluated these implementations using the layer configurations shown in table 2, which are commonly used for benchmarking convolution performance [1],
and quote average throughput for all these layers. cuDNN performance ranges from 0.8× to 2.25×
that of cuda-convnet2, with an advantage at smaller batch sizes. Compared to Caffe, cuDNN performance ranges from 1.0× to 1.41×. Importantly, even with a small mini-batch size of only 16,
cuDNN performance is still 86% of maximum performance, which shows that our implementation
performs well across the convolution parameter space.
Table 3 illustrates cuDNN’s performance portability across GPU architectures. The Tesla K40 is
built using the Kepler architecture, and has a peak single-precision throughput of 4.29 TFlops. The
6
Layer 1
Layer 2
Layer 3
Layer 4
Layer 5
K40 (TFlops)
0.99
1.52
1.51
1.38
1.01
GTX980 (TFlops)
1.46
2.49
2.42
2.19
1.87
K40 (%)
23%
35%
35%
32%
24%
GTX980 (%)
30%
51%
49%
45%
38%
Table 3: Performance portability
Geforce GTX 980 is built using the newer Maxwell architecture, and has a peak single-precision
throughput of 4.95 TFlops. cuDNN performance ranges from 23-35% of peak on the Tesla K40,
and from 30-51% of peak on the GTX 980. This data illustrates how cuDNN provides performance
portability across GPU architectures, with no need for users to retune their code as GPU architectures
evolve.
4
Caffe Integration
Caffe is a deep learning framework developed with expression, speed, and modularity in mind. Its
architecture mirrors the natural modularity of deep networks as compositional models made from a
collection of inter-connected layers. A deep network is defined layer-by-layer in a plaintext schema.
Each layer type is implemented according to a simple protocol of setup, forward, and backward
steps to encapsulate engineering details. Data and derivatives flow through layers and between
the host and device according to the framework’s unified memory interface that handles allocation
and communication. cuDNN integration raises the speed and memory efficiency of the framework
without sacrificing expression or modularity.
4.1
Development
Integration is made simple by the self-contained design of the cuDNN handles, descriptors, and
function calls together with the modularity of the framework. The core Caffe framework is unaltered, preserving the network, layer, and memory interfaces. Changes were isolated to new layer
definitions and implementations, helper functions for descriptors, and the corresponding tests. The
patch is almost purely additive.
In Caffe each type of model operation is encapsulated in a layer. Layer development comprises
declaring and implementing a layer class, defining the layer in the protocol buffer model schema,
extending the layer factory, and including tests. Computations are carried out through the layer
protocol of setup, forward, and backward steps. cuDNN layers fit this same scheme. Library handles
and descriptors are configured in setup while forward and backward calls are made in the respective
layer methods. The cuDNN primitives yield concise layer implementations. The cuDNN layers are
drop-in replacements to their standard Caffe counterparts.
Caffe has a standard array and memory interface, called a blob, for storing and communicating
data on the host and device. Blobs hold data, gradients, and parameters. In particular, layer inputs
and outputs are held in N × C × H × W dimensional blobs. cuDNN tensor and filter descriptors
are trivially constructed from blob arrays through its flexible support for dimension and stride. By
coordinating memory solely at the descriptor, Caffe retains control over memory and communication
for efficiency.
The Caffe+cuDNN convolution layer exploits the reduced memory consumption of cuDNN to speed
up execution. The forward pass parallelizes the computation of group convolution (filtering with sets
of filters with restricted channel connectivity). The backward pass parallelizes the computation of
the gradients with respect to the bias, filter weights, and bottom data.
Table 4 illustrates the performance improvements gained from integrating cuDNN into Caffe. Overall, training time for 200 iterations improved by 36%, when training the bvlc reference caffenet
model, using cuDNN R1 on an NVIDIA Tesla K40.
7
Backward Propagation
Forward Propagation
Testing
Update
Overall
Caffe (seconds)
156270
109310
21600
2926
290106
Caffe+cuDNN (seconds)
120651
75330
14529
2920
213431
Speedup
1.30
1.45
1.49
1.00
1.36
Table 4: Caffe Performance
4.2
User Experience
cuDNN computation is transparent to the user through drop-in integration. The model schema and
framework interfaces are completely unchanged. Setting a single compilation flag during installation
equips Caffe with cuDNN layer implementations and sets cuDNN as the default computation engine.
Layers automatically fall back to the standard Caffe functionality in all cases outside of the current
scope of cuDNN. Model execution can be tuned through per-layer ENGINE parameters that select
the implementation. The workflow for the user is virtually identical.
Future cuDNN releases will be integrated in like fashion.
5
Baidu Integration
Several deep learning projects at Baidu have integrated cuDNN. For example, it has been integrated
into PADDLE, Baidu’s internal deep learning framework. On a Tesla K10 GPU, we observed that
performance on a benchmark set of convolutional layers improves by 30% on average over an implementation that lowers convolution to matrix multiply.
We are also using cuDNN in other domains besides image processing, such as speech and language.
cuDNN’s ability to convolve non-square inputs with asymmetric padding is particularly useful for
these other domains. In our experience, cuDNN has reduced memory consumption compared with
matrix multiplication, which has enabled us to use larger models and larger mini batches. cuDNN
was simple to integrate into our code because we didn’t need to change our data structure layout,
thanks to cuDNN’s flexible interface.
6
Future Work
We are considering several avenues for expanding the performance and functionality of cuDNN.
Firstly, although our convolution routines are competitive with other available implementations,
more work remains to bring performance up to that attained by matrix multiplication. Over time,
we hope to shrink this gap. Secondly, we envision adding support for other primitives; for example:
1D and 3D convolutions would be useful for speech, language processing, and video applications,
among others. Local Receptive Field computations, which are similar to convolutions, but with
untied weights, are also useful and could be added to cuDNN. Finally, we would like this library to
help people use multiple GPUs to accelerate training.
7
Conclusion
This paper presents cuDNN, a library for deep learning primitives. We presented a novel implementation of convolutions that provides reliable performance across a wide range of input sizes, and takes
advantage of highly-optimized matrix multiplication routines to provide high performance, without
requiring any auxiliary memory. We also provide a set of routines that allow users to train and evaluate complete deep neural networks without the need to write parallel code manually. As parallel
architectures continue to evolve, such libraries will provide increasing value to the machine learning
community. The library is available at [3], and we welcome feedback at [email protected].
8
References
[1] convnet-benchmarks.
https://github.com/soumith/convnet-benchmarks,
2014.
[2] Netlib BLAS. http://www.netlib.org/blas/, 2014.
[3] NVIDIA cuDNN - GPU accelerated deep learning. https://developer.nvidia.
com/cuDNN, 2014.
[4] Yoshua Bengio, Aaron C. Courville, and Pascal Vincent. Unsupervised feature learning and
deep learning: A review and new perspectives. CoRR, abs/1206.5538, 2012.
[5] James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume Desjardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: a cpu and
gpu math expression compiler. In SciPy, volume 4, page 3, 2010.
[6] Kumar Chellapilla, Sidd Puri, Patrice Simard, et al. High performance convolutional neural
networks for document processing. In Workshop on Frontiers in Handwriting Recognition,
2006.
[7] Adam Coates, Brody Huval, Tao Wang, David Wu, Bryan Catanzaro, and Andrew Ng. Deep
learning with COTS HPC systems. In ICML, pages 1337–1345, 2013.
[8] Ronan Collobert, Koray Kavukcuoglu, and Clément Farabet. Torch7: A matlab-like environment for machine learning. In BigLearn, NIPS Workshop, 2011.
[9] George E Dahl, Dong Yu, Li Deng, and Alex Acero. Context-dependent pre-trained deep
neural networks for large-vocabulary speech recognition. Audio, Speech, and Language Processing, IEEE Transactions on, 20(1):30–42, 2012.
[10] Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly,
Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups.
Signal Processing Magazine, IEEE, 29(6):82–97, 2012.
[11] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick,
Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature
embedding. arXiv preprint arXiv:1408.5093, 2014.
[12] Alex Krizhevsky. cudaconvnet2. https://code.google.com/p/cuda-convnet2/,
2014.
[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep
convolutional neural networks. In NIPS, pages 1097–1105, 2012.
[14] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning
applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
[15] Andrew L Maas, Awni Y Hannun, Daniel Jurafsky, and Andrew Y Ng. First-pass large vocabulary continuous speech recognition using bi-directional recurrent dnns. arXiv preprint
arXiv:1408.2873, 2014.
[16] Michael Mathieu, Mikael Henaff, and Yann LeCun. Fast training of convolutional networks
through ffts. arXiv preprint arXiv:1312.5851, 2013.
[17] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng
Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li FeiFei. Imagenet large scale visual recognition challenge, 2014.
[18] Pierre Sermanet, David Eigen, Xiang Zhang, Michaël Mathieu, Rob Fergus, and Yann LeCun. Overfeat: Integrated recognition, localization and detection using convolutional networks. arXiv preprint arXiv:1312.6229, 2013.
[19] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov,
Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions.
arXiv preprint arXiv:1409.4842, 2014.
[20] Guangming Tan, Linchuan Li, Sean Treichler, Everett Phillips, Yungang Bao, and Ninghui
Sun. Fast implementation of DGEMM on Fermi GPU. In Supercomputing 2011, SC ’11,
pages 35:1–35:11, New York, NY, USA, 2011. ACM.
[21] Henry S. Warren. Hacker’s Delight. Addison-Wesley Professional, 2002.
9
| 9cs.NE
|
Decentralised L1 Adaptive Primary Controllers &
Distributed Consensus-Based Secondary Control
for DC Microgrids with Constant-Power Loads
Daniel O’Keeffe∗†1 , Stefano Riverso‡2 , Laura Albiol-Tendillo§2 , and Gordon
Lightbody¶1,3
arXiv:1804.00251v2 [cs.SY] 3 Apr 2018
1
Control & Intelligent Systems Group, School of Engineering, University College of
Cork, Ireland
2
United Technologies Research Centre Ireland Ltd, 4th Floor Penrose Business
Centre, Cork, Ireland
3
MaREI-SFI Research Centre, University College Cork, Ireland
Technical Report
March, 2018
Abstract
Constant-power loads are notoriously known to destabilise power systems, such as DC microgrids, due to their negative incremental impedance. This paper equips distributed generation
units with decentralised L1 adaptive controllers at the primary level of the microgrid control hierarchy. Necessary and sufficient conditions are provided to local controllers for overall microgrid
stability when constant-power loads are connected. The advantages of the architecture over conventional heuristic approaches are: (i) scalable design, (ii) plug-and-play functionality, (iii) well
defined performance and robustness guarantees in a heterogeneous and uncertain system, and
(iv) avoids the need for online measurements to obtain non-a priori system impedance information. The proposed primary control architecture is evaluated with distributed consensus-based
secondary level controls using a bus-connected DC microgrid, which consists of DC-DC buck and
boost converters, linear and non-linear loads. Stability of the overall hierarchical control system
is proven using a unit-gain approximation of the primary level.
Keywords: Consensus Algorithms, Constant-Power Loads, Decentralised & Distributed Control, Low-Voltage DC Islanded Microgrid, Robust-Adaptive Control, Scalable Design, Voltage Stability
∗ Research is supported by the Irish Research Council enterprise partnership scheme (Award No. EPSPG/2015/106)
in collaboration with University College Cork, Ireland and United Technologies Research Centre Ireland Ltd.
† Email: [email protected]; Corresponding author
‡ Email: [email protected]
§ Email: [email protected]
¶ Email: [email protected]
1
1
Introduction
The increasing complexities associated with large-scale interconnected systems (LSiS) has led to
the mass-adoption of decentralised control architectures over conventional centralised approaches [1,2].
In fact, decentralised controllers have become the control standard for distributed autonomous power
systems, also known as microgrids (mGs). The IEEE 2030.7 and 2030.8 standards outline guidelines
for providing flexibility and plug-and-play (PnP) features in LSiS, which have recently become more
imperative [3].
A key challenge for decentralised controllers in LSiS is to stipulate necessary and sufficient stability conditions. The stability of a mG can be compromised by the connection of constant-power loads
(CPLs). In a small-signal sense, CPLs reduce system damping between individually stable systems
by introducing negative-incremental impedance. Typically, mGs experience the effects of CPLs when
tightly regulated motor drives are interfaced with inadequately damped power converters, or when
multiple mGs are clustered together [4]. Conventional criteria for stabilising mGs with CPLs have
been reviewed in [5, 6]. Such approaches employ classical frequency domain analysis which is predicated on determining the impedance ratio between the interconnected power converters, known as the
minor-loop gain, and are limited to unidirectional power flows, unlike state-space approaches. This
requires a priori knowledge of the total load number and the effective impedance value. Moreover,
stability conditions are only sufficient, i.e. an unstable criteria can lead to a stable or an unstable
system. The passivity-based criterion was proposed in [7, 8] to overcome these restrictions by considering the overall system and measuring the effective impedance. Though the criterion is simple
and practical, the tuning guidelines for local controllers are not provided. Ultimately, these designs
are heuristic and not scalable, i.e. N controllers may require retuning. As complex LSiS proliferate
and become more heterogeneous, features such as flexibility, robustness to uncertainty and PnP operations will become increasingly more vital. Consequently, such restrictive approaches may become
prohibitive.
Decentralised PnP control architectures have been proposed at the primary and secondary control
levels of distributed generation units (DGUs) to guarantee overall voltage and current stability when
power converters are plugged-in/out, irrespective of system topology. [9–12, 14, 15]. To facilitate
system scalability and reconfiguration, a control design orientated approach is adopted by using a loadconnected model which treats loads as an exogenous disturbance. Kron reduction methods [16] are
subsequently used to map interconnections to more general network topologies such as bus-connected
[17]. Among these works, only [14] has considered CPL stabilisation within the PnP framework.
Here, each local model uses a low-frequency approximation of the CPL and neglects high-frequency
content. However, local stability conditions are only provided for DC-DC buck converters, while tests
do not consider CPLs at all. Moreover, accurate knowledge of load power-ratings are required, which
thus restricts compliance with flexible systems of different owners and stakeholders [18].
Due to the increasing uncertainty within mGs, adaptive controllers have been implemented at the
primary and secondary control levels [19,21,22]. However, these strategies are based on primary droop
controllers, which are the conventional heuristic method for load-power sharing [23–25]. These control
schemes depend on specific mG models and topologies, and do not provide transient or steady-state
robustness guarantees. Moreover, PnP operations are not considered. Recently, decentralised robustadaptive control architectures have become attractive due to their guaranteed robustness during fast
adaptation when compared to conventional model reference adaptive control (MRAC) architectures
[26, 27].
In [28], we developed a scalable decentralised L1 adaptive control (L1 AC) architecture to augment the primary voltage controllers of DGUs. Steady-state robustness guarantees are provided in
the presence of arbitrary topology, PnP operations, uncertain couplings and unknown load changes.
However, global asymptotic stability (GAS) is achieved in a conservative fashion, while its computation is non-scalable. Subsequently, we formulated a distributed L1 AC architecture to guarantee
GAS in a scalable, PnP manner [29]. Ultimately, these approaches are predicated on the ability to
access hardware or software in order to augment or retrofit existing baseline controllers. In some
applications this might not be available.
This paper extends our previous work by designing a purely adaptive architecture, and incorporating CPL stabilisation conditions for local controller tuning. Unlike the techniques of [4–6, 30],
which implement stabilisers at the load-side converter, this paper follows [14,31] whereby the DGU is
equipped with each stabiliser. The rationale for this is that increasing damping by reducing the loadside control bandwidth impairs load-power quality. Moreover, it is the generators of the traditional
2
utility grid which act to stabilise the overall system during load changes.
The proposed architecture is implemented in a bus-connected DC mG, consisting of DC-DC
boost and buck converters together with closed-loop speed controlled DC motors. Though the lowfrequency CPL approximation of [14] is also used, the low-pass filtering feature inherent in L1 AC
architectures, means that the high-frequency content of the CPL can be neglected without adversely
affecting transient stability. Furthermore, due to the adaptive nature of the architecture, mapping
bus-connected coupling parameters into the load-connected design model via Kron reduction can be
avoided as long as the parameters are contained within the uncertainty subset.
Finally, the proposed primary controllers are fitted with distributed consensus-based secondary
controllers to achieve voltage restoration and load-power sharing objectives. Consensus algorithms
have become popular in distributed systems as they can help decentralised controllers achieve centralisedlike performance. Consensus allows nodes in a sparse communications network to construct a vision
of the global system with limited information in a fault-tolerant manner. From the perspective of
the secondary controllers, the primary level can be modelled as a unit-gain approximation [13]. Lyapunov functions are again used to demonstrate asymptotic stability using both approximations. The
system’s PnP capabilities and resiliency to communication faults is tested.
A version of this paper has been submitted to the 12th UKACC International Conference on
Control.
2
Constant-power load model
A bus-connected mG is a typical topology in automotive, marine and aircraft applications [32] is
shown in Fig. 1.
Figure 1: Typical DC microgrid configuration with n generation-side and m load-side converters.
Fig. 1 shows a mG bus formed via n generation-side/DGUs and m load-side converters. Each
converter is locally fitted with voltage and current primary controllers. The primary control layer of
the typical mG control hierarchy is required to provide fast and stable voltage and current performance
in response to load changes and reconfiguration.
This section derives the negative-incremental impedance model of load-side converters acting as
CPLs, by decomposing each converter’s closed-loop impedance into low and high frequency components using the Extra Element Theorem [5]. At frequencies less than the closed-loop bandwidth i.e.
where the controller loop-gain, TkL (s) is large, and the controller works well, the closed-loop impedance
of each load-side converter, Zkin (s), approximates ZkN (s), the forced impedance response. At frequencies greater than the closed-loop bandwidth, Zkin (s) follows the open-loop impedance, ZkD (s). This
3
can be represented in terms of local admittance,
1
TkL (s)
1
1
1
=
+ D
in
L
N
Zk (s)
1 + Tk (s) Zk (s) Zk (s) 1 + TkL (s)
(1)
where TkL (s) = CkL (s)GLk (s) assuming unity sensor and actuator gain, and k ∈ M = {1, ..., m} denotes
the set of load-converters, or number of CPLs. The overall admittance can be written as,
m
X
1
Zkin (s)
(2)
PkL = vbus (t)ibus
k (t)
(3)
Yin (s) =
i=1
A CPL can be expressed as,
where PkL , vbus and ibus
represent the fixed power, and variable bus voltage and current. Within
k
the closed-loop bandwidth, the voltage controller works to maintain a fixed load-power by adjusting
its duty cycle to maintain a constant vkbus (t), regardless of any input voltage disturbances from a
mG bus. Typically, to exhibit good power quality, load-side converters are tuned to have very fast
control-bandwidth. Therefore, if the bus voltage at the power converter’s input terminals reduces,
then ibus
k (t) will increase. This is the manifestation of negative-incremental impedance, which is
analytically shown as,
L
∂
Pk
PL
∂vkbus
N
CPL
|(Vbus ,Ikbus ) = − k 2
Zk (s) = Rk
= bus |(Vbus ,Ikbus ) = bus
(4)
bus
∂ik
∂ik
Ik
Ikbus
For completion, the transfer function for the open-loop input impedance ZkD (s) can be found in [6].
3
Power converter models for DC microgrids
This section defines the models for DC-DC source-side and load-side power converters. Boost
converters, which increase input voltages, and buck converters, which decrease input voltages, are
considered. Following the control design approach of [11], loads are treated as exogenous inputs in
a load-connected topology. Fig.2 shows each converter circuit topology with RL power lines that
connect converters to each other.
Figure 2: DC-DC power converter circuit topologies with power line connections to neighbours.
4
Each converter can be modelled in small-signal state-space form as,
(
ẋ[i] (t) = Aii x[i] (t) + Bi u[i] (t) + Ei w[i] (t) + ζ[i] (t)
DGU
Σ[i] :
y[i] (t) = Ci x[i] (t)
(5)
where x[i] (t) = [ĩdci , ṽdci ]T , is the small-signal state vector, u[i] (t) is the small-signal control input, w[i] (t) = ĩLi is the small-signal load disturbance, represented as an exogenous input, ζ[i] (t) =
P
DGU
j∈Ni Aij x[j] (t) represents coupling with Σ[j] , and (i, j) ∈ N = {1, ..., n} denotes the set of DGUs,
and j ∈ Ni = {1, ..., n − 1} denotes the neighbour set of DGU i. From [28], the matrices of (5) for
the boost converter are defined as,
"
" Vdc #
#
R
i)
i
− (1−D
− Ltti
0
0
0
Lti
Lti
i
Aii = (1−Di )
Bi = −Idci Ei =
Ci = 0 1
Aij = 0
P
1
−1
1
−
Rij Ct
Ct
Cti
j∈Mi Rij Cti
Cti
i
i
where Aii ∈ R2×2 is the state matrix; Aij ∈ R2×2 is the coupling matrix, Bi ∈ R2 is the input vector,
V
Vini
, and Idci = (1−Diin)i2 RL , where RLi
and Ci ∈ R1×2 is the output vector. Furthermore, Vdci = (1−D
i)
i
is the effective local load resistance at each converter’s terminals. However, since the lad-connected
topology treats loads as disturbance RLi is unknown.
From [11], the matrices of (5) for the buck converter are defined as,
" Rt
#
1
− Lt i
− L1t
0
0
0
L
t
i
i
i
P
Ei = − 1
Aii =
Aij = 0
Bi =
Ci = 0 1
1
1
1
−
0
Rij Ct
Ct
Cti
j∈Mi Rij Cti
i
i
Comparing both converter models, it is clear that the dynamics of the boost are dependent on its
duty cycle operating point and is non-minimum phase (NMP) for output voltage control. The NMP
action makes controller tuning more difficult, particularly when coupled to unknown power lines as
addressed in [33].
+ili , where
To introduce the CPL model of (4), the exogenous input is altered such that iLi = iCPL
i
ĩli (t) represents the current disturbance due to non-CPL loads i.e. neighbouring DGUs or battery
banks. The large-signal equivalent line resistances and currents at the load-connected terminals of
each DGU must be mapped from the original bus-connected topology using Kirchoff’s voltage and
current laws, as in [10];
m
X
1
, ∀j 6= i
Rq
q=1
Pm
Pm bus
1
PL
ik
vbus
k=1
= P k=1 k
= P
m
m
Rk q=1 R1q
Rk q=1 R1q
Rij = Ri Rj
iCPL
i
(6)
where Ri and Rj are the line resistances that connect DGUs to the mG bus as in Fig. 1. Furthermore,
as the mG bus voltage varies depending on the line resistances and the number of grid-forming DGUs
present, vbus is represented in terms of the output voltage of each DGU.
Pn vdci
Using (4) and substituting
Pm
vbus =
1
k=1 RCPL
k
vdci
i=1 Ri
Pn
i=1 Ri
Pn
1
+ i=1 R1i
k=1 RCPL
k
vbus = Pm
= − v21
bus
±
r
P
Pm
k=1
PkL into (7) yields a quadratic solution for vbus ,
vdci
n
i=1 Ri
2
(7)
Pn
2
+4
1
i=1 Ri
Pm
k=1
PkL
Pn
1
i=1 Ri
(8)
Alternatively, if the upper-bound RkCPL is known, i.e. within the bounds of adaptation, the effective
CPL rating can be written as,
Pn 1 Pm
1
L
+
i=1 Ri
k=1 Pk
RCPL
CPL
k
P
(9)
Pi
=
m
1 P
1
Rk q=1 R1q
j∈Ni vdcj
Ri 1 + vdc
i
5
PiCPL
vdci
=
Finally, iLi must be linearised about the operating point as iCPL
i
ĩCPL
=−
i
is non-linear. This yields,
PiCPL
2 ṽdci
Vdc
i
(10)
As a result, the state matrix of the boost converter model becomes,
R
i)
− Ltti
− (1−D
L
ti
i
Aii = (1−Di )
P
PiCPL
1
1
− Ct
j∈Ni Rij − V 2
Ct
i
i
and the state matrix of the buck converter model becomes,
R
− ti
− L1t
i
Lti
Aii = 1
P
1
1
− Ct
j∈Ni Rij −
Ct
i
(11)
dci
i
PiCPL
2
Vdc
i
(12)
Ultimately, the low frequency approximation of the CPL can no longer be modelled as an exogenous input as it now directly influences the eigenvalues of both converters. This introduces greater
uncertainty into the model as loads are unknown when considering flexible and heterogeneous DC
mGs.
4
Decentralised L1 adaptive primary voltage control
Over the last decade, efforts to improve the transient performance and robustness guarantees
of conventional MRAC architectures with fast adaptation, led to the formulation of L1 AC theory
[26]. This robust-adaptive control architecture achieves transient performance and bounded state
and control signal guarantees by inserting a low-pass filter (LPF) at the input to both the plant
and state-predictor. Through the appropriate design of the LPF, the typical trade-off between fast
adaptation and robustness is decoupled. The theory has been developed for time-varying uncertainty
and disturbances, unmodelled dynamics, time-delays and unknown input gains and non-linearities
[26].
Centralised L1 AC architectures have been notably successful in safety-critical applications including; sub-scale NASA aircraft auto-pilots [34], manned aircraft [35], and unmanned water/aerial
vehicles [36, 37]. In addition to our decentralised [28] and distributed [29] L1 AC architectures, decentralised schemes were formulated in [38, 39], and were implemented to augment aircraft baseline
controllers [40]. Unlike [34–40], we consider unmatched interconnections and provide robustness to
constant disturbances via integral action.
Recent criticisms of L1 AC theory, found in [41–46], have been reviewed and rebuked in [29,47–49].
4.1
Plant structure
The plant has a known structure, but with unknown parameter values. The control objective
is to design a bounded control input u[i] (t), such that y[i] (t) = Ci x[t] (t) tracks a reference voltage
with convergent state and bounded parametric errors in the presence of matched uncertainty and
unmatched coupling and disturbances. Defining the new state,
ref
ξ˙[i] (t) = y[i]
(t) − y[i] (t)
Z t
Z t
ref
ref
(t) − Ci x[i] (t) dt
y[i]
(t) − y[i] (t) dt =
y[i]
ξ[i] (t) =
(13)
0
0
ref
ref
where y[i]
(t) = Vdc
− Vdci .
i
Remark 1. Initially, the unmatched coupling term ζ[i] is neglected to enable local decoupled design.
Subsequently, the term is reintroduced when GAS conditions are provided. Moreover, as in [28], the
integrator provides adequate robustness to unmatched constant disturbances, and therefore ĩl[i] (t) is
neglected. The augmented open-loop model, where a matched uncertainty term is introduced to
6
represent parametric uncertainty in the dynamics of ΣDGU
can be represented as,
[i]
Σ̄DGU
[i]
:
(
ref
x̄˙ [i] (t) = Āii x̄[i] (t) + B̄i u[i] (t) + θ̄[i] (t)x̂[i] (t) + F y[i]
(t)
ȳ[i] (t) = C̄i x̄[i] (t)
(14)
1
where x̄[i] (t) ∈ R3 , is the system measurable state vector; uL
[i] (t) ∈ R is the control signal; θ̄[i] (t) ∈ Θ ⊂
3
R is the unknown parametric uncertainty vector, which belongs to the known uniformly bounded
convex set Θ. The matrices of (14) are defined as,
0
Bi
Aii 02x1
F = 0 C̄i = Ci 0
B̄i =
Āii =
0
−Ci
0
1
4.2
Control law
The small-signal control input u[i] (t) for Σ̄DGU
is a fusion of a state-feedback controller and low[i]
frequency banded uncertainty compensation signal, as defined in the Laplace domain,
L1
L1
T
x̄
](t)
(15)
(t)
+
u
(t)
=
−
K
x̄
(t)
+
C(p)[
θ̂
C[i]
: uSF
i
[i]
[i]
[i]
[i]
[i]
where Ki = [Kii , Kiv , Kiξ ] ∈ R1×3 is the state-feedback control gain vector, C(p) represents a second
d
-order Butterworth LPF1 , and p , dt
represents the differential operator. The robustness of the
L1 AC is dependent on the LPF bandwidth ωc , as subsequently designed.
4.3
State-predictor
The state-predictor generates an estimate of the system states. Thereafter, the adaptive law is
used to asymptotically drive the uncertain plant dynamics to converge to the desired closed-loop
dynamics of the state-predictor. Without loss of generality, the state-predictor formulation and the
desired closed-loop dynamics are equal for all DGUs,
(
T
ref
1
x̂˙ [i] (t) = Âm x̂[i] (t) + B̂m (uL
[i] (t) + θ̂[i] (t)x̂[i] (t)) + F y[i] (t)
E[i] :
(16)
ŷ[i] (t) = Ĉi x̂[i] (t)
− B̄i Ki ∈ R3×3 Hurwitz, Âm is the design matrix that specifies the
where Ki renders Âm , Ānom
ii
desired closed-loop dynamics, and Ānom
is the state matrix in which the designer estimates as the
ii
nominal dynamics without uncertainty in order to design the nominal control gains.
4.4
Adaptive law
The adaptive law generates an estimate of the plant uncertainties. Defining the state-error and
parametric estimation error vectors as, x̃[i] (t) = x̄[i] (t) − x̂[i] (t) and θ̃[i] (t) = θ̄[i] (t) − θ̂[i] (t), the
state-error dynamics, used to drive the adaptive law, can be defined as,
x̃˙ [i] (t) = Âm x̃[i] (t) + B̂m θ̃[i] (t)x̂[i] (t)
(17)
The adaptive law is determined from Lyapunov’s second stability method. A quadratic Lyapunov
candidate is defined as a function in terms of x̃[i] (t) and θ̃[i] (t).
V[i] (x̃[i] (t), θ̃[i] (t)) = x̃[i] (t)T Pi x̃[i] (t) + θ̃[i] (t)T Γ−1
i θ̃[i] (t)
(18)
where, Pi ∈ R3×3 is a symmetric matrix, such that Pi = PiT > 0 is the solution to the algebraic
Lyapunov linear inequality ATii Pi + Pi Aii ≤ −Qi , for arbitrary Qi = QTi > 0, and Γi ∈ R+ is the
adaptive gain. If the time-derivative of (18) is at least negative semi-definite, then (14) is locally
1 A second-order Butterworth LPF is chosen as it has a maximally flat pass-band, and the order must be at least
equal to the order of the closed-loop plant to ensure a bounded control input - see section 4.6
7
stable since the energy along the trajectories of state and estimation errors decreases. By defining
the adaptive law as,
˙
θ̂[i] (t) , Γi Proj(θ̂[i] (t), −x̄[i] (t)x̃T[i] (t)Pi B̂m )
(19)
where the projection operator, defined in [50], bounds the parametric uncertainty estimate, then
V̇[i] (x̃[i] (t), θ̃[i] (t)) ≤ −x̃[i] (t)T Qi x̃[i] (t) ≤ 0
(20)
By invoking Barbalat’s Lemma it follows that lim x̃[i] (t) = 0. Hence, asymptotic convergence of
t→∞
x̃[i] (t) and boundedness of θ̃[i] (t) is proven.
4.5
Filter design
The phase-lag introduced by LPFs can reduce phase-margin and lead to instability. This section
defines the conditions for which local stability is ensured when the LPF is inserted. The closed-loop
dynamics are presented in the Laplace domain as,
−1
x̄[i] (s) = (sI − Âm )
−1
B̄i (1 − C(s)) η[i] (s) + (sI − Âm )
ref
(s) + (sI − Âm )
F y[i]
−1 0
x̄[i]
(21)
where x̄0[i] = x̄[i] (0) is the initial state and,
T
(t) x̄[i] (t) + x̃[i] (t)
η[i] (t) = θ̂[i]
(22)
From (19) and (20), the signals θ̂[i] (t) and x̃[i] (t) are bounded.
However, Σ̄DGU
is non-linear due to the term in (22). To determine stability conditions upon
[i]
insertion of the LPF and to prescribe performance specifications such as rise-time, settling-time etc.,
a linear time-invariant (LTI) reference model is defined where the uncertainty term is known. The
available reference system can be given as,
T
ref
∗
ref
ref
x̄˙ ref
[i] (t) = Âm x̄[i] (t) + B̄i (1 − C(p)) θ̄[i] x̄[i] (t) + F y[i] (t)
(23)
∗
where, θ̄[i] (t) = θ̄[i]
is known. From (23), estimation is decoupled from control as the identification
of the local state vector is independent of the control input. The performance and robustness specifications can now be set independent of the estimation process. Subsequently, section 4.6 shows that
the Σ̄DGU
converges to the LTI reference model with uniform and decoupled performance bounds as
[i]
the adaptation increases.
Using the LTI reference model, the conditions for local stability due to insertion of the LPF are
determined by the following Lemma:
Lemma 1. Following the small-gain theorem and Lemma 2.1.2 in [26], if ||(C(s)−1)(sI−Âm )−1 B̄i ||L1 θmax <
1, then the reference system (23) is bounded-input-bounded-state stable with respect to initial conditions and reference output.
Proof. From the definition of the closed-loop reference system (23), it follows that,
−1
−1
−1
∗T ref
ref
x̄ref
(s)
=
(sI
−
Â
)
B̄
(1
−
C(s))
θ̄
x̄
(s)
+ (sI − Âm ) F y[i]
(s) + (sI − Âm ) x̄0[i]
m
i
[i]
[i] [i]
(24)
As, (sI − Âm )−1 and C(s) are stable transfer functions, the following bound holds,
||x̄ref
[i] ||L∞
≤
−1 0
x̄[i] ||L∞
||(sI − Âm )
−1
+ ||(sI − Âm )
−1
1 − ||(C(s) − 1)(sI − Âm )
ref
F y[i]
||L∞
T
∗ ||
B̄i θ̄[i]
L1
(25)
ref
Since x̄i.c.
[i] (t) and y[i] (t) are uniformly bounded, the reference states are bounded as long as the
denominator of (25) does not equal zero. As defined in [26], the worst case adaptation bound is,
∗ 2
||1
θmax = 4 max ||θ̄[i]
θ∈Θ
8
(26)
∗
where θ̄[i]
is determined as the required gain such that the eigenvalues of Āii and Âm are the same.
Here, Āii is at its most uncertain, i.e. maximal deviation between plant and desired eigenvalues. θmax
represents the boundary of projection for estimating the parameters when using the adaptation law
(19). Finally, for the reference states to remain bounded, the following L1 -norm condition must be
satisfied,
λ = ||(C(s) − 1)(sI − Âm )−1 B̄i ||L1 θmax < 1
(27)
where the degree-of-freedom is ωc .
Consequently, Barbalat’s lemma can be used again to show lim x̃[i] (t) = 0, i.e. asymptotic
t→∞
tracking is maintained when inserting the LPF. By designing the LPF bandwidth to minimise (27),
the L1 AC guarantees uniform transient and steady-state performance bounds.
Remark 2. In order to neglect the high-frequency component of the CPL, the bandwidth of C(s)
should be smaller than the closed-loop natural frequency of the overall impedance Zin (s).
4.6
Design considerations
This section derives: local stability conditions when CPLs augment the dynamics of each DGU;
uniform and decoupled performance bounds of input and output signals between the non-linear system
(14) and LTI reference model (23); and global stability conditions.
The closed-loop desired dynamics matrix for boost converters is defined as,
(1−Di +Vdci Kiv )
Vdci Kiξ
R +V
Ki
−
−
− ti Ltdci i
Lti
Lti
i
1−Di +It K i
CPL
Boost
P
I
Kiξ
P
t
1
1
i
v
Âm
=
(28)
i
i
i
−
−
K
−
I
2
ti i
j∈Ni Rij
Cti
Cti
Cti
Vdc
i
0
−1
0
The closed-loop desired dynamics matrix for buck converters is defined as,
R i +Kii
(1+Kiv )
− tL
−
0
L
ti
ti
P
Kiξ
PiCPL
1
1
1
ÂBuck
=
m
− Ct
2
j∈Ni Rij − Vdc
Cti
Cti
i
i
0
−1
0
(29)
Necessary and sufficient conditions for the stability of the local closed-loop subsystems are:
Boost/Buck
Boost/Buck
) > 0. As a result, the control gains must satisfy the
) < 0 and det(Âm
trace(Âm
following conditions in order to preserve the local stability of the DC mG under CPLs. The conditions
for boost converters are,
Iti Rti + Vdci (1 − Di )
i
Ki < −
2Iti Vdci
CPL
X 1
1
P
C
t
(30)
Kiv <
− i 2 − i (Rti + Kii Vdci )
Iti
Rij
Vdci
Lti
j∈Ni
Kiξ > 0
The conditions for buck converters are,
CPL
X 1
P
Rt
1
− i2 − i
Kii <
Cti
Rij
Vdci
Lti
j∈Ni
X 1
P CPL
− i2 −1
Kiv < −(Rti + Kii )
Rij
Vdci
j∈Ni
Kiξ > 0
9
(31)
The convergence of the non-linear closed-loop adaptive system to the uncertainty free linear reference model in (23) can be shown by subtracting (23) from the uncertain plant in . The bound on
this is represented as
||x̄[i] − x̂ref
[i] ||L∞ ≤
||C(s)||L1 ||ρ[i] (s)||L1
1 − ||(C(s) − 1)(sI − Âm )−1 B̄i ||L1 θmax
(32)
where, from Lemma 2.2.5 in [26], ρ[i] (s) is the Laplace transform of the upper-bound on the non-zero
state-error initialisation,
s
−αi t
(V[i] (0) − θmax
λmin (Qi )
θmax
Γi )e
;
αi ,
+
(33)
ρ[i] (t) ,
λmin (Pi )
λmin (Pi )Γi
λmax (Pi )
(32) is rewritten as,
γ1
||x̄[i] − x̂ref
[i] ||L∞ ≤ √
Γi
(34)
where,
γ1 =
||C(s)||L1 ||ρ[i] (s)||L1
1 − ||(C(s) − 1)(sI − Âm )−1 B̄i ||L1 θmax
s
(Γi V[i] (0) − θmax )e−αi t
θmax
+
λmin (Pi )
λmin (Pi )
For bound between real and LTI reference control inputs, we define,
T
ref( s)
∗
(s)
=
−C(s)
η
(s)
+
η̃
(s)
+
C(s)
u[i] (s) − uref
x̂
θ̄
[i]
[i]
[i]
[i] [i]
T
∗
x̄[i] (s) − x̂ref
−C(s)η̃[i] (s)θ̄[i]
[i] (s)
(35)
where, following (22), η̃[i] (s) is the Laplace transform of η̃[i] (t) = θ̂[i]T (t x̃[i] (t). Using Lemma A.12.1
of [26], which implies that there exists an arbitrary c0 , we define the strictly proper and boundedinput-bounded-output stable transfer function,
H1 (s) = C(s)
1
cT0 (sI
− Âm )−1 B̄i
cT0
(36)
The error-dynamics in (17) can be written in the Laplace domain as,
x̃[i] (s) = (sI − Âm )−1 B̄i η̃[i] (s)
(37)
Using (36) and (37), C(s)η̃[i] (s) = H1 (s)x̃[i] (s), (35) becomes,
T
∗
u[i] (s) − uref
[i] (s) = −H1 (s)x̃[i] (s) + C(s)θ̄[i]
The performance bound of (38) is written as,
(s)
x̄[i] (s) − x̂ref
[i]
(38)
||u[i] − uref
[i] || ≤ ||H1 (s)||L1 ||x̃[i] ||L∞ + ||C(s)||L1 θmax γ1
(39)
γ2
||u[i] − uref
[i] || ≤ √
Γi
(40)
(Γi V[i] (0) − θmax )e−αi t
θmax
+
+ ||C(s)||L1 θmax γ1
λmin (Pi )
λmin (Pi )
(41)
or,
where,
γ2 = ||C(s)
1
cT0 (sI − Âm )−1 B̄i
cT0 ||L1
s
Remark 3. It is clear that if the relative degree of C(s) is not at least equal to the relative degree of
the plant, (sI− Âm )−1 B̄i , then the first term in (41) is not strictly proper, and will result in unbounded
10
control inputs, and therefore no transient guarantees can be provided. This is also why in standard
MRAC architectures, i.e. C(s) = 1, such guarantees cannot be provided.
Remark 4. The inversion of (sI−Âm )−1 B̄i results in unstable poles for non-minimum-phase systems.
Boost converters are non-minimum-phase for duty-cycle to output voltage control. Therefore, the
desired dynamics must be carefully designed to be minimum-phased. This can be done by transforming
local systems to control-canonical form, as performed in [29]. Finally, lim ||x̄[i] − x̂ref
[i] ||L∞ = 0 and
Γi →∞
lim ||u[i] −uref
[i] ||L∞ = 0 shows that the non-linear local plant converges to the linear reference model,
Γi →∞
and the performance bounds of both state and control inputs decrease as the adaptive gain increases.
Conservative global asymptotic stability conditions are derived offline using collective Lyapunov
functions. The overall Lyapunov function candidate that describes the global system can be written
as,
V(t) =
N
X
i=0
T
x̃T[i] (t)Pi x̃[i] (t) + θ̃[i]
(t)Γ−1
θ̃
(t)
[i]
i
(42)
Assumption 1. We assume local controllers exploit (19), and plant dynamics have converged to
desired dynamics.
The derivative of (42) is,
V̇ = −
N
X
x̃T[i] (t)Qi x̃[i] (t)
(43)
i=0
if and only if matrix P satisfies the Lyapunov inequality equation,
T
T
 P + PÂm + ÂC P + PÂC < 0
} |
{z
}
| m {z
(a)
(44)
(b)
where P = diag(Pi ) ∈ R3M×3M ; Âm = diag(Âm ) ∈ R3M×3M represents the overall desired dynamics;
ÂC = Â − Âm ∈ R3M×3M represents the coupling dynamics only. As each DGU is designed to be
locally asymptotically stable, the matrices of (a) are negative definite. Therefore, the design of K
is performed iteratively offline to ensure ||(a)|| > ||(b)||. This typically results in detuned controller
gains, as expected due to the conservative requirements of decentralised systems. Furthermore, this
method can suffer when system size expands as the retuning of K becomes more difficult, despite the
advantage of designing the desired dynamics as the same for all DGUs. The controller gains can be
tuned optimally using LQR guidelines, such as in [51], provided conditions (30) or (31) depending on
converter topology, and (44) are satisfied.
The decentralised L1 adaptive primary control architecture can be seen in Fig.3.
Figure 3: Purely adaptive architecture: decentralised L1 adaptive controller.
11
5
Coordinated secondary control
Coordination among multiple DGUs within any mG is imperative for dynamic operation and
system management. Two key control objectives include: (i) voltage restoration, and (ii) load-power
sharing [6, 23]. In reference to (i), as the bus voltage of a bus-connected mG cannot be directly
controlled and since unknown voltage drops occur across unknown power-lines, the secondary control
level is required to compute appropriate voltage references for the primary control level in order to
maintain each DGU voltage within a prescribed range. In reference to (ii), secondary controllers
are required to also provide a voltage reference that ensures DGUs can share the power delivery
between each other independent of both topology and power-lines. The mathematical definitions are
as follows.
Definition 1. Equal-current sharing is achieved if the output current of each DGU equals the average
load-current, i.e.
iout
= hIL i
i
(45)
= idci for buck converters, IL = [iL1 , iL2 , ..., iLN ]T
where,
= (1 − Di )idci for boost converters,
is the local unknown load current vector, and the operator h.i denotes the average value of a vector.
iout
i
iout
i
ref
ref
Assumption 2. Secondary control voltage references are identical Vbus
.
= Vbus
i
Definition 2. Under assumption 2, voltage restoration is achieved if the average voltage of all DGU
ref
, i.e.
output voltages equals Vbus
ref
hVi = Vbus
T
(46)
where, V = [vdc1 , vdc2 , ..., vdcN ] .
Remark 5. We consider a network of communication links consisting of a set of nodes T =
{T1 , T2 , ..., TN }, connected through edges E = T × T . Such a network is described by a graph
G = (T , E). Each node represents a DGU in the network and edges represent communication links
for information exchange. The Laplacian matrix, L(G) describes the graph in matrix form, and is
defined as,
if i 6= j and Ti is adjacent to Tj
−1,
L(G)ij = deg(Ti ), if i = j
(47)
0,
otherwise
where deg(Ti ) represents the number of adjacent nodes to connected to Ti .
The secondary control level is typically implemented at a slower bandwidth to the primary level,
and utilises a communication network in order to coordinate and manage voltage and current/power
levels within the overall mG.
5.1
Distributed consensus-based controller design
In a distributed system, the bus voltage of a bus-connected mG cannot be directly controlled.
Instead, each primary level voltage reference is controlled such that the average of all the local
voltages equals an estimate of the bus. Secondary controllers in [2, 10, 23, 33] have utilised a fullyconnected low-bandwidth communications (LBC) network, whereby the average voltage estimate
is determined by measuring every DGU. However, this clearly places restrictions on the network
i.e. fault-tolerance. As a result, a sparse-connected, peer-to-peer, LBC network has been proposed,
whereby dynamic consensus algorithms are implemented to estimate global information from a limited
number of nodes [19, 20].
The dynamic consensus algorithm used to estimate the mG bus voltage is denoted as,
v̂ibus (t)
, vdci (t) +
Zt X
0 j∈Ni
Zt
vdci (t) −
aij v̂jbus (τ ) − v̂ibus (τ ) dτ =
X
0 j∈Ni
12
aij v̂ibus (τ ) − v̂jbus (τ ) dτ
(48a)
(48b)
where aij = 1 if Σ̄DGU
is connected to Σ̄DGU
, otherwise aij = 0. Subsequently, the estimate of the
[j]
[i]
mG bus voltage is used to drive the average of all the voltages to equal a bus voltage reference.
To achieve (ii), the voltage reference for the primary level is controlled to ensure equal current
sharing, i.e. the current injected equals the average total current injected within the mG. A consensus
algorithm similar to above is also used for this.
iout
refi (t)
,
iout
i (t)
+
Zt X
0 j∈Ni
Zt
iout
i (t) −
out
aij îout
j (τ ) − îi (τ ) dτ =
(49a)
out
(τ
)
−
î
(τ
)
dτ
aij îout
i
j
(49b)
X
0 j∈Ni
Using remark 5, the global consensus algorithm can be represented by vector form as,
Z
V̂bus , V − L(G) V̂dτ =
Z
V + L(G) V̂dτ
depending on how the vector V̂ is constructed. Similarly,
Z
Iref = I − L(G) Îdτ =
Z
I + L(G) Îdτ
(50a)
(50b)
(51a)
(51b)
depending on how the vector Î is constructed.
ref
Finally, the following controlled correction terms can be added to Vbus
,
∆vdci (t) =
∆iout
i (t)
=
v
kP
i
i
kP
i
ref
vbus
−
v̂ibus (t)
+
kIvi
Zt
0
ref
− v̂ibus (τ ) dτ
vbus
Zt
iout
iout
i (t)
out
i (τ )
out
i
irefi −
irefi −
+ kIi
dτ
mi
mi
(52)
(53)
0
where mi is the load-sharing co-efficient i.e. for equal current-sharing mi = mj , and ∆vdci and ∆iout
i
are generated from PI controllers, represented in the Laplace domain as,
v
+
Cv (s) = kP
i
kIvi
s
(54)
i
+
Ci (s) = kP
i
kIii
s
(55)
and,
The PI controllers can be tuned by representing the primary level as a unit-gain approximation [12,15].
Finally, the voltage reference sent to the proposed primary control level to achieve (i) and (ii) is given
as,
ref
ref
= vbus
+ ∆vdci (t) + ∆iout
Vdc
i (t)
i
(56)
The overall distributed hierarchical control structure can be seen Fig.4,
5.2
Stability analysis
This section uses Lyapunov stability analysis to show that under the effect of secondary control,
stability of the overall system is maintained and all DGU output currents and voltages converge such
that (45) and (46) are satisfied. Following [13], unit-gain approximations of the primary L1 AC level
are made and subsequently, stability is proven.
13
Figure 4: Overall hierarchical control architecture with decentralised L1 adaptive primary controller
and distributed secondary PI controllers.
5.2.1
Unit-gain approximation
Approximating the primary L1 adaptive control loops with ideal unitary gains is a reasonable one
at low-frequencies since the loop can be shaped by the LPF of L1 AC. The global relationship between
the primary L1 AC level and the distributed-consensus based secondary control level is represented
as,
V = Vref
bus + ∆V + ∆I
(57)
out
out T
where ∆V = [∆vdc1 , ..., ∆vdc2 , ..., ∆vdcN ]T and ∆I = [∆iout
1 , ..., ∆i2 , ..., ∆iN ] . The overall adjustment signals from the PI controllers are defined in vector form as,
Z
V
∆V = KV
e
+
K
eV dτ
(58)
I
P V
and,
∆I =
KIP eI
+
KII
Z
eI dτ
(59)
V
I
I
v
v
i
i
where KV
P = diag(kPi ), KI = diag(kIi ), KP = diag(kPi ), and KI = diag(kIi ). The global secondary
control error dynamics are defined as,
Z
eV = Vref
V̂dτ
(60)
−
V
−
L(G)
bus
bus T
where V̂ = [v̂1bus , v̂2bus , ..., v̂N
] , and,
eI = Iref − I − L(G)
out
out T
where Î = [îout
1 , î2 , ..., îN ] .
Z
Îdτ
(61)
Lemma 2. The symmetric Laplacian matrix L(G) is positive definite.
Proof. An arbitrary vector, z ∈ Rn , can always be written as z = ẑ + z̄, where ẑ ∈ H1 and z̄ ∈ H1⊥ .
H1 ∈ R is a subspace composed of vectors with zero average, i.e. H1 = {ẑ ∈ Rn : hẑi = 0} and
H1⊥ ∈ R is a subspace orthogonal to H1 , i.e. H1⊥ = {α1n : α ∈ R}. Then,
z T L(G)z
is equivalent to the following cases,
T
If z > 0, then z L(G)z > 0
If z < 0, then z T L(G)z > 0
If z̄ = 0, then ẑ T L(G)z̄ > 0
Therefore, L(G) is positive definite.
14
(62)
(63)
For convenience, we prove convergence of the output current loop first by isolating the currentsharing adjustment term in (57), which is expressed as,
I = Iref
pri + ∆I
(64)
where Iref
pri = diag(0N ). The following Lyapunov function candidate is considered,
VI =
1 T
e PI eI
2 I
(65)
The derivative of which equates to,
V̇I ≤ eTI PI ėI
(66)
ėI = − İ + L(G)Î
(67)
˙ = −(I + KI )−1 KI L(G)Î − KI eI
İ = ∆I
P
I
P
(68)
V̇I ≤ −eTI PI (I + KIP )−1 (−KIP L(G)Î + KII eI + L(G)Î)
(69)
where the derivative of (61) is,
and the derivative of (64) yields,
Therefore, (66) can be written as,
or,
V̇I ≤ −
KII T
KI
eI PI A + AT PI eI − P eTI PI A + AT PI L(G)Î
2
2
(70)
where A = (I + KIP )−1 . To prove convergence, Barbalat’s lemma is invoked. The derivative of (70)
is,
V̈I ≤
KII T
eI PI A + AT PI AKIP L(G)Î + AKII eI
2
(71)
Since (65) and (71) are bounded, (70) is uniformly continuous and lim V̇I = 0. As both individual
t→∞
terms in (70) are positive definite, both must converge to 0 if lim V̇I = 0. As a result, the output
t→∞
currents converge to their reference.
Remark 6. A more trivial solution to above can be given if RIref = I, which is the case when the loadsharing co-efficients mi = mj = 1. As a result eI = −L(G) Îdτ , and therefore, V̇I = −eTI PI L(G)Î.
Next, the following Lyapunov function is considered to prove voltage stability and convergence,
VV =
1 T
e PV eV
2 V
(72)
The derivative of which equates to,
where,
V̇V ≤ −eTV PV V̇ + LV̂
(73)
˙ + ∆I
˙
V̇ = ∆V
(74)
Using (51) and similar analysis from above,
1
1
T
T
T
T
V̇V ≤ − KV
I eV PV A + A PV eV − L(G)eV PV A + A PV V̂−
2
2
1
1 I
T
T
KP L(G)eV PV A + A PV Î − eTV PV A + AT PV eI
2
2
15
(75)
To prove asymptotic stability, we again invoke Barbalat’s lemma where a bounded second derivative,
2
2
1 V T
1
T
T
V̂+
KI eV PV A + AT PV eV + KV
I eV PV A + A PV
2
2
1
1 I
2
2
K L(G)eTV PV A + AT PV Î + eTV PV A + AT PV eI
2 P
2
(76)
V̈V <
(77)
leads to a uniformly continuous (75) and lim V̇V = 0. Similar to before, the average of each output
t→∞
voltage converges to the bus voltage reference.
6
Results
Simulations are performed in Matlab/Simulink. Firstly, we show instability induced by negativeincremental impedance between two DC-DC power converters using non-adaptive state-feedback controllers. Subsequently, the proposed architecture is evaluated using a bus-connected mG consisting
of 6 DGUs and 2 power electronic loads.
6.1
Constant-power load instability
A lightly damped boost converter, stepping 100 V to 382 V is interfaced with a tightly regulated
buck converter which powers a resistive load at 48 V. Fig. 5(a) highlights the family of unstable and
stable closed-loop eigenvalues as the effective incremental impedance of the boost converter’s load
is varied from -10 Ω to 10 Ω. Fig.5(b) shows a poorly damped bus voltage in response to a step
800
386
600
Source DGU Voltage (V)
384
Imaginary Axis
400
200
Load = 10 Ω
0
Load = -10 Ω
-200
382
380
378
376
374
372
370
-400
368
-600
366
-800
2.5
-6
-4
-2
0
2
Real Axis
4
2.6
6
2.7
2.8
2.9
3
Time (s)
×10 5
(b) Effect of CPL instability on mG bus voltage:
(a) Family of closed-loop eigenvalues as RiCPL varies from
(red) tightly regulated load-side converter; (blue)
-10 Ω to 10 Ω.
closed-loop bandwidth detuned.
Figure 5: Constant-power load instability.
from 382 V - 375 V. Fig.5(b) also shows that by detuning the load-side converter controller, the bus
voltage response improves. However, as previously mentioned this would be at the cost of decreased
load performance.
6.2
Stable voltage restoration and equal-current sharing
L1
,
Next, the 6 DGUs of a bus-connected mG, similar to that in Fig.2, are equipped with C[i]
DGU
DGU
DGU
DGU
DGU
DGU
i = {1, ..., 6}, where Σ̄[1] , Σ̄[2] , Σ̄[3] , Σ̄[4] , and Σ̄[5] are boost converters while Σ̄[6] is a
buck converter. Each DGU provides power to a 380 V DC bus which is loaded with a 5 kW linear
resistive load and a 3.8 kW non-linear closed-loop controlled DC motor. Each load is interfaced to
the bus via buck converters, where the bus voltage is stepped down to 48 V. System parameters are
detailed in Table I of [28].
As shown in section 4.6, C(s) must at least match the smallest relative degree of the plant.
Therefore a second-order Butterworth LPF is designed according to (27), where a frequency sweep
16
0.48285
0.4828
0.48275
λ
0.4827
0.48265
0.4826
0.48255
0.4825
0.48245
0
1
2
3
4
5
6
7
8
9
ω c (rads -1)
10
×10 4
Figure 6: Calculation of appropriate LPF bandwidth ωc such that λ < 1.
from 100 − 100krads−1 is performed. From Fig.6, the bandwidth can be selected arbitrarily. However,
from remark 2, in order to allow Zin (s) to be approximated as a negative-incremental resistance, an
upper-bound on the bandwidth of C(s) should be provided. Fig.5(a) compares the Bode plot of
the overall closed-loop load impedance Zin (s) to the overall open-loop load impedance ZD (s), i.e.
overall high frequency content of CPL loads. In Fig.5(a), Zin (s) asymptotically approximates ZD (s)
at 3 krads−1 for small CPL levels i.e. each load equal to 100 W. In Fig.5(b), Zin (s) asymptotically
approximates ZD (s) at 50 krads−1 for large CPL levels i.e. each load equal to 10 kW.
Bode Diagram
Bode Diagram
60
40
20
0
Z in (s)
-20
Z D (s)
-40
90
45
0
-45
-90
-135
10 0
10 1
10 2
10 3
10 4
40
20
0
Z in (s)
Z D (s)
-20
90
Phase (deg)
Phase (deg)
Magnitude (dB)
Magnitude (dB)
60
0
-90
-180
-270
10 0
10 5
10 1
Frequency (rad/s)
10 2
10 3
10 4
10 5
10 6
Frequency (rad/s)
(a) PiL = 100 W.
(b) PiL = 10 kW.
Figure 7: Bode plots of Zin (s) under varying PiL .
As a result, the upper-bound is conservatively chosen as 3 krads−1 . In general, as it is assumed
that power levels are unknown, the upper-bound on C(s) should be selected using an a priori estimate
of the smallest expected load power within the overall mG. Therefore,
C(s) =
s2
9 × 106
+ 4.243 × 103 s + 9 × 106
(78)
DGU
DGU
DGU
DGU
The bus-connected mG is powered initially by DGUs Σ̄DGU
[1] , Σ̄[2] , Σ̄[3] , Σ̄[4] , and Σ̄[5] , while
ref
Σ̄DGU
powers a local load on its own. The bus voltage reference vbus
is set as 380 V.
[6]
17
6.2.1
Plug-and-play operation
At t = 8s, Σ̄DGU
[6] , which steps an input voltage of 700 V down to 380.5 V, plugs-in. The topology
change is represented by the following Laplacian matrices,
3 −1 −1 0
0 −1
2 −1 −1 0
0
−1 2
0 −1 0
0
−1 2
0 −1 0
−1 0
2 −1 0
0
; L(G)t=8s
−1
0
2
−1
0
L(G)t=0s
(79)
0 −1 −1 3
0 −1
0 −1 −1 3 −1
0
0
0 −1 2 −1
0
0
0 −1 1
−1 0
0
0 −1 2
Fig.8(a) shows the response of all the DGUs when a fully-connected communications network is
used. Fig.8(b) shows the response when a sparse connected communications network with optimal
link redundancy is used.
390
X: 3.189
Y: 384.1
390
X: 3.872
Y: 383.7
385
380
X: 3.573
Y: 381.2
X: 3.284
Y: 373.5
375
DGU output voltage (V)
DGU Output Voltages (V)
385
X: 3.173
Y: 384.1
X: 3.662
Y: 373.3
370
DGU 1
DGU 2
DGU 3
DGU 4
DGU 5
DGU 6
365
360
355
1.8
2
2.2
380
375
370
DGU 1
DGU 2
DGU 3
DGU 4
DGU 5
DGU 6
365
360
355
2.4
2.6
2.8
3
3.2
3.4
3.6
3.8
7.5
8
8.5
Time (s)
9
9.5
Time (s)
(a) Fully-connected communications network.
(b) Sparse communications network.
Figure 8: Voltage restoration of the microgrid bus voltage when Σ̄DGU
plugs-in.
[6]
Here, it is clear that all DGUs restore their output voltages such that their average equals 380 V.
The secondary control response is adequately fast with 0.7 s settling times and the sparse network
compares well with the fully-connected network.
Fig.10 shows the maintenance of equal-current sharing during the plug-in test. Again, a fullyconnected network and the sparse network are compared.
34
DGU 1
DGU 2
DGU 3
DGU 4
DGU 5
DGU 6
50
DGU output current (A)
DGU Output Currents (A)
32
30
X: 2.983
Y: 26.78
28
DGU 1
DGU 2
DGU 3
DGU 4
DGU 5
DGU 6
55
26
45
40
35
30
25
20
24
15
10
22
1.6
1.8
2
2.2
2.4
2.6
2.8
3
7.4
Time (s)
7.6
7.8
8
8.2
8.4
8.6
8.8
9
9.2
Time (s)
(a) Fully-connected communications network.
(b) Sparse communications network.
Figure 9: Equal-current sharing maintained when Σ̄DGU
plugs-in.
[6]
18
9.4
30
DGU 1
DGU 2
DGU 3
DGU 4
DGU 5
29
28
DGU 1
DGU 2
DGU 3
DGU 4
DGU 5
DGU 6
390
Average voltage estimate (V)
Total load current versus number of DGUs ratio (A)
When Σ̄DGU
joins the network, all DGU output current injections quickly converge to a current
[6]
of 26.78 A. Though steady-state performance is similar for both networks with a settling time of
in particular has a large 55% overshoot. As a
approximately 0.6 s, the transient response of Σ̄DGU
[5]
DGU
result Σ̄[6] undershoots by the same amount in order to satisfy (45).
The convergence rates at start-up, i.e. t = 0s, and at t = 8s when the topology reconfigures i.e.
L(G) changes, are plotted below.
27
26
25
24
385
380
375
370
23
365
22
0
1
2
3
4
5
6
7.5
7
8
8.5
9
9.5
10
10.5
11
Time (s)
Time (s)
(a) Convergence of DGU current estimates at start-up.
(b) Sparse communications network.
Figure 10: Convergence of DGU current and voltage estimates.
6.2.2
Link-failure resiliency
Finally, the system’s resiliency to communication link failures is investigated. At t = 14s, though
the topology of the physical system remains the same when Σ̄DGU
plugs-in at t = 8s, the topology of
[6]
the communication network changes. Edges ε12 and ε13 fail. Σ̄DGU
can only exchange information
[1]
DGU
with Σ̄[6] . The Laplace changes from L(G)t=8s in (79) to,
L(G)t=14s
1
0
0
0
0
0
0 −1
0
0
−1 0
0
0
0 −1
1 −1 0
0
1 −1 0
0
−1 2 −1 0
0 −1 2 −1
0
0 −1 2
(80)
The response to the change in communication topology is shown below. Fig.11 shows the system’s
fault tolerant capability. As long as each node in the communication network has a neighbour such
that information can be exchanged throughout the system either directly or indirectly then system
irrespective of communication faults. Ultimately, the topology change, reflected by the Laplacian
matrices, only affects the transient response. In general, the less nodes connected the slower the
response and the larger over/under shoots can be, as seen in Fig.11(a).
7
Conclusion
This paper develops a novel scalable and purely adaptive L1 AC architecture to stabilise DGUs in
a bus-connected DC mG with CPLs. The architecture draws on a control orientated load-connected
model which enables scalability and treats unknown loads and exogenous inputs. Due to the adaptive
nature of the architecture, a bus-connected topology does not need to be mapped into this loadconnected topology, as long as the bus-connected parameters exist within the uncertainty bound.
The destabilising current disturbances induced by the CPL at each DGU terminal is approximated
by its low-frequency content due to the LPF of the L1 AC architecture. Necessary and sufficient
stability conditions are provided for DC-DC boost and buck converters using a nominally expected
19
DGU 1
DGU 2
DGU 3
DGU 4
DGU 5
DGU 6
DGU output current (A)
50
40
DGU 1
DGU 2
DGU 3
DGU 4
DGU 5
DGU 6
388
386
DGU output voltage (V)
60
30
20
10
0
-10
384
382
380
378
376
374
-20
372
-30
13.95
14
14.05
14.1
14.15
13.9
Time (s)
14
14.1 14.2 14.3 14.4 14.5 14.6 14.7 14.8 14.9
Time (s)
(a) DGU output current response.
(b) DGU output voltage response.
Figure 11: Output current and voltage responses when communications links ε12 and ε13 fail.
load power. Subsequently, the smallest a priori expected load total is used to determine an upperbound for designing the bandwidth of the architecture’s LPF. As a result, very fast asymptotic
convergence of local state and estimation errors is theoretically proven and demonstrated with a busconnected mG. For overall system stability of the primary level, a conservative, offline approach is
used.
The proposed adaptive primary controllers are cascaded with distributed consensus-based secondary controllers where global knowledge of the average bus voltage and load current are estimated
using limited local information. Asymptotic stability of the hierarchical control system is shown using
Lyapunov functions. Voltage restoration and equal-current sharing responses are shown to be fast
and stable during PnP and communication link failure tests.
References
[1] M. A. Anuradha and A. Massoud, IEEE Vision for Smart Grid Controls: 2030 and Beyond:
Roadmap. IEEE CSS, 2013.
[2] M. Andreasson, D. V. Dimarogonas, H. Sandberg, and K. H. Johansson, “Distributed controllers
for multiterminal HVDC transmission systems,” IEEE Transactions on Control of Network Systems, vol. 4, no. 3, pp. 564–574, 2017.
[3] J. Ã. Stoustrup, “Plug & Play Control : Control Technology Towards New Challenges,” European
Journal of Control, vol. 15, no. 3-4, pp. 311–330, 2009.
[4] Q. Shafiee, T. Dragicevic, J. C. Vasquez, and J. M. Guerrero, “Modeling, stability analysis and
active stabilization of multiple DC-microgrid clusters,” in ENERGYCON 2014 - IEEE International Energy Conference, 2014, pp. 1284–1290.
[5] A. Riccobono and E. Santi, “Comprehensive review of stability criteria for dc power distribution
systems,” IEEE Transactions on Industry Applications, vol. 50, no. 5, pp. 3525–3535, 2014.
[6] T. Dragicevic, X. Lu, J. C. Vasquez, and J. M. Guerrero, “DC Microgrids - Part I: A Review
of Control Strategies and Stabilization Techniques,” IEEE Transactions on Power Electronics,
vol. 31, no. 7, pp. 4876–4891, 2016.
[7] A. Riccobono and E. Santi, “A novel passivity-based stability criterion (pbsc) for switching
converter dc distribution systems,” in Applied Power Electronics Conference and Exposition
(APEC), 2012 Twenty-Seventh Annual IEEE. IEEE, 2012, pp. 2560–2567.
[8] A. Riccobono and E. Santi, “Stability analysis of an all-electric ship mvdc power distribution
system using a novel passivity-based stability criterion,” in Electric Ship Technologies Symposium
(ESTS), 2013 IEEE. IEEE, 2013, pp. 411–419.
20
[9] S. Riverso, F. Sarzo, and G. Ferrari-Trecate, “Plug-and-Play Voltage and Frequency Control of
Islanded Microgrids with Meshed Topology,” IEEE Transactions on Smart Grid, vol. 6, no. 3,
pp. 1176–1184, 2015.
[10] S. Riverso, M. Tucci, J. C. Vasquez, J. M. Guerrero, and G. Ferrari-Trecate, “Stabilizing plugand-play regulators and secondary coordinated control for AC islanded microgrids with busconnected topology,” Applied Energy, no. August, pp. 1–21, 2017.
[11] M. Tucci, S. Riverso, J. C. Vasquez, J. M. Guerrero, and G. Ferrari-Trecate, “A Decentralized
Scalable Approach to Voltage Control of DC Islanded Microgrids,” IEEE Transactions on Control
Systems Technology, vol. 24, no. 6, pp. 1965–1979, 2016.
[12] M. Tucci, L. Meng, J. M. Guerrero, and G. Ferrari-Trecate, “Plug-and-play control and consensus
algorithms for current sharing in DC microgrids,” in IFAC-PapersOnLine, vol. 50, 2016, pp. 1–23.
[13] M. Tucci, L. Meng, J. M. Guerrero, and G. Ferrari-Trecate, “Consensus algorithms and
plug-and-play control for current sharing in DC microgrids,” arXiv, pp. 1–23, 2016. [Online].
Available: http://arxiv.org/abs/1603.03624
[14] M. S. Sadabadi, Q. Shafiee, and A. Karimi, “Plug-and-Play Robust Voltage Control of DC
Microgrids,” IEEE Transactions on Smart Grid, pp. 1–1, 2017.
[15] R. Han, M. Tucci, R. Soloperto, A. Martinelli, G. Ferrari-Trecate, and J. M. Guerrero, “Hierarchical Plug-and-Play Voltage/Current Controller of DC microgrid with Grid-Forming/Feeding
modules: Line-independent Primary Stabilization and Leader-based Distributed Secondary
Regulation,” no. July, 2017. [Online]. Available: http://arxiv.org/abs/1707.07259
[16] F. Dorfler and F. Bullo, “Kron reduction of graphs with applications to electrical networks,”
IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 60, no. 1, pp. 150–163, 2013.
[17] M. Tucci, A. Floriduz, S. Riverso, and G. Ferrari-Trecate, “Plug-and-play control of ac islanded
microgrids with general topology,” in Control Conference (ECC), 2016 European. IEEE, 2016,
pp. 1493–1500.
[18] F. Katiraei, A. Zamani, and R. Masiello, “Microgrid Control Systems,” IEEE Power and Energy
Magazine, vol. 15, no. 4, pp. 116–112, 2017.
[19] V. Nasirian, A. Davoudi, F. L. Lewis, and J. M. Guerrero, “Distributed adaptive droop control
for DC distribution systems,” IEEE Transactions on Energy Conversion, vol. 29, no. 4, pp.
944–956, 2014.
[20] Q. Shafiee, T. Dragicevic, F. Andrade, J. C. Vasquez, and J. M. Guerrero, “Distributed
consensus-based control of multiple dc-microgrids clusters,” in Industrial Electronics Society,
IECON 2014-40th Annual Conference of the IEEE. IEEE, 2014, pp. 2056–2062.
[21] X. Lu, K. Sun, J. M. Guerrero, J. C. Vasquez, and L. Huang, “State-of-Charge Balance Using
Adaptive Droop Control for Distributed Energy Storage Systems in DC Microgrid Applications,”
IEEE Trans. Ind. Electron., vol. 61, pp. 2804–2815, 2014.
[22] T. V. Vu, D. Perkins, F. Diaz, D. Gonsoulin, C. S. Edrington, and T. El-Mezyani, “Robust
adaptive droop control for DC microgrids,” Electric Power Systems Research, vol. 146, pp. 95–
106, 2017.
[23] X. Lu, J. M. Guerrero, K. Sun, and J. C. Vasquez, “An improved droop control method for dc
microgrids based on low bandwidth communication with dc bus voltage restoration and enhanced
current sharing accuracy,” IEEE Transactions on Power Electronics, vol. 29, no. 4, pp. 1800–
1812, 2014.
[24] J. W. Simpson-Porco, F. Dörfler, and F. Bullo, “Synchronization and power sharing for droopcontrolled inverters in islanded microgrids,” Automatica, vol. 49, no. 9, pp. 2603–2611, 2013.
[25] J. Zhao and F. Dörfler, “Distributed control and optimization in dc microgrids,” Automatica,
vol. 61, pp. 18–26, 2015.
21
[26] C. Cao and Hovakimyan, L1 Adaptive Control Theory: Guaranteed Robustness with Fast Adaptation. Society for Industrial and Applied Mathematics, 2010.
[27] T. E. Gibson, A. M. Annaswamy, and E. Lavretsky, “Improved Transient Response in Adaptive
Control Using Projection Algorithms and Closed Loop Reference Models,” AIAA Guidance,
Navigation, and Control Conference, no. August, pp. 1–13, 2012.
[28] D. O’Keeffe, S. Riverso, L. Albiol-Tendillo, and G. Lightbody, “Voltage Control of DC
Islanded Microgrids: Scalable Decentralised L1 Adaptive Controllers,” 2018. [Online]. Available:
arXivpreprintarXiv:1801.04508
[29] D. O’Keeffe, S. Riverso, L. Albiol-Tendillo, and G. Lightbody, “A Distributed Scalable
Architecture using L1 Adaptive Controllers for Primary Voltage Control of DC Microgrids,”
2018. [Online]. Available: arXivpreprintarXiv:1801.06484
[30] X. Lu, K. Sun, L. Huang, J. M. Guerrero, J. C. Vasquez, and Y. Xing, “Virtual impedance
based stability improvement for DC microgrids with constant power loads,” 2014 IEEE Energy
Conversion Congress and Exposition, ECCE 2014, vol. 6, no. 6, pp. 2670–2675, 2014.
[31] V. Arcidiacono, A. Monti, and G. Sulligoi, “Generation control system for improving design and
stability of medium-voltage dc power systems on ships,” IET Electrical Systems in Transportation, vol. 2, no. 3, pp. 158–167, 2012.
[32] L. Herrera and J. Wang, “Stability analysis and controller design of dc microgrids with constant
power loads,” in Applied Power Electronics Conference and Exposition (APEC), 2015 IEEE.
IEEE, 2015, pp. 691–696.
[33] D. O’Keeffe, S. Riverso, L. Albiol-Tendillo, and G. Lightbodyt, “Distributed Hierarchical Droop
Control of Boost Converters in DC Microgrids,” 28th IEEE Irish Signals and Systems Conference,
pp. 1–6, 2017.
[34] I. Gregory, E. Xargay, C. Cao, and N. Hovakimyan, “Flight Test of an L1 Adaptive Controller on
the NASA AirSTAR Flight Test Vehicle,” AIAA Guidance, Navigation, and Control Conference,
pp. 1–31, 2010.
[35] K. A. Ackerman, E. Xargay, R. Choe, N. Hovakimyan, C. M. Cotting, R. B. Jeffrey, M. P.
Blackstun, P. T. Fulkerson, T. R. Lau, and S. S. Stephens, “Evaluation of an L1 Adaptive
Flight Control Law on Calspans Variable-Stability Learjet,” Journal of Guidance, Control, and
Dynamics, pp. 1–10, 2017.
[36] C. H. Svendsen, N. O. Holck, R. Galeazzi, and M. Blanke, “L1 adaptive manoeuvring control
of unmanned high-speed water craft,” IFAC Proceedings Volumes (IFAC-PapersOnline), vol. 9,
no. PART 1, pp. 144–151, 2012.
[37] B. Michini and J. P. How, “L1 adaptive control for indoor autonomous vehicles: Design process
and flight testing,” Proceeding of AIAA Guidance, Navigation, and Control Conference, no.
August, pp. 1–15, 2009.
[38] S. Yoo, C. Cao, and N. Hovakimyan, “Decentralised L1 adaptive control for large-scale non-linear
systems with interconnected unmodelled dynamics,” IET Control Theory & Applications, vol. 4,
no. 10, pp. 1972–1988, 2010.
[39] S. Yoo, N. Hovakimyan, and C. Cao, Decentralized L1 adaptive control for large-scale systems
with unknown time-varying interaction parameters, 10 2010, pp. 5590–5595.
[40] G. Kumaresan and A. Kale, “Application of L1 adaptive controller for the design of a novel
decentralized leader follower formation algorithm,” IFAC-PapersOnLine, vol. 49, no. 1, pp. 706–
711, 2016.
[41] P. A. Ioannou, A. M. Annaswamy, K. S. Narendra, S. Jafari, L. Rudd, R. Ortega, and J. Boskovic,
“L1-Adaptive Control: Stability, Robustness, and Interpretations,” IEEE Transactions on Automatic Control, vol. 59, no. 11, pp. 3075–3080, 2014.
22
[42] J. D. Boskovic and R. K. Mehra, “Performance Analysis of a Simple L1-Adaptive Controller,”
in American Control Conference (ACC). IEEE, 2013, pp. 3376–3381.
[43] R. Ortega and E. Panteley, “Comments on-adaptive control: stabilisation mechanism, existing
conditions for stability and performance limitations,” International Journal of Control, vol. 87,
no. 3, pp. 581–588, 2014.
[44] R. Ortega and E. Panteley, “L1-”Adaptive” control always converges to a linear PI control and
does not perform better than the PI,” vol. 19, no. 3, pp. 6926–6928, 2014.
[45] R. Ortega and E. Panteley, “When is a parameterized controller suitable for adaptive control?”
European Journal of Control, vol. 22, pp. 13–16, 2015.
[46] R. Ortega, E. Panteley, and A. Bobtsov, “Comments on ’Comparison of Architectures and Robustness of Model Reference Adaptive Controllers and L1-Adaptive Controllers’,” International
Journal of Adaptive Control and Signal Processing, vol. 30, pp. 125–127, 2016.
[47] N. Hovakimyan, “L1 Adaptive Control,” Department of Mechanical Science and Engineering,
University of Illinois at Urbana-Champaign, Tech. Rep, pp. 1–15, 2014.
[48] T. Souanef and W. Fichter, “Comments on L1 Stability Condition,” International Journal of
Control, no. August, pp. 1 – 18, 2014.
[49] O. B. Altin, “Relaxing fundamental assumptions in iterative learning control,” Ph.D. dissertation, University of Michigan, 2016.
[50] E. Lavretsky and T. E. Gibson, “Projection Operator in Adaptive Systems,” arXiv preprint,
arXiv:1112.4232, 2011. [Online]. Available: http://arxiv.org/abs/1112.4232
[51] B. Kurucs, “State Space Control of Quadratic Boost Converter using LQR and LQG approaches,”
2015 Intl Conference on Optimization of Electrical & Electronic Equipment, no. 2, pp. 642–648,
2015.
23
| 3cs.SY
|
Differentiable Learning of Logical Rules for
Knowledge Base Reasoning
arXiv:1702.08367v3 [cs.AI] 27 Nov 2017
Fan Yang
Zhilin Yang
William W. Cohen
School of Computer Science
Carnegie Mellon University
{fanyang1,zhiliny,wcohen}@cs.cmu.edu
Abstract
We study the problem of learning probabilistic first-order logical rules for knowledge base reasoning. This learning problem is difficult because it requires learning
the parameters in a continuous space as well as the structure in a discrete space.
We propose a framework, Neural Logic Programming, that combines the parameter
and structure learning of first-order logical rules in an end-to-end differentiable
model. This approach is inspired by a recently-developed differentiable logic called
TensorLog [5], where inference tasks can be compiled into sequences of differentiable operations. We design a neural controller system that learns to compose
these operations. Empirically, our method outperforms prior work on multiple
knowledge base benchmark datasets, including Freebase and WikiMovies.
1
Introduction
A large body of work in AI and machine learning has considered the problem of learning models
composed of sets of first-order logical rules. An example of such rules is shown in Figure 1. Logical
rules are useful representations for knowledge base reasoning tasks because they are interpretable,
which can provide insight to inference results. In many cases this interpretability leads to robustness
in transfer tasks. For example, consider the scenario in Figure 1. If new facts about more companies
or locations are added to the knowledge base, the rule about HasOfficeInCountry will still be
usefully accurate without retraining. The same might not be true for methods that learn embeddings
for specific knowledge base entities, as is done in TransE [3].
HasOfficeInCity(New York, Uber)
CityInCountry(USA, New York)
In which country Y
does X have office?
HasOfficeInCountry(Y, X) ?
X = Uber
Y = USA
HasOfficeInCountry(Y, X) ß HasOfficeInCity(Z, X), CityInCountry(Y, Z)
X = Ly*
HasOfficeInCity(Paris, Ly*)
CityInCountry(France, Paris)
Y = France
Figure 1: Using logical rules (shown in the box) for knowledge base reasoning.
Learning collections of relational rules is a type of statistical relational learning [7], and when the
learning involves proposing new logical rules, it is often called inductive logic programming [18]
. Often the underlying logic is a probabilistic logic, such as Markov Logic Networks [22] or
ProPPR [26]. The advantage of using a probabilistic logic is that by equipping logical rules with
probability, one can better model statistically complex and noisy data. Unfortunately, this learning
problem is quite difficult — it requires learning both the structure (i.e. the particular sets of rules
included in a model) and the parameters (i.e. confidence associated with each rule). Determining
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
the structure is a discrete optimization problem, and one that involves search over a potentially large
problem space. Many past learning systems have thus used optimization methods that interleave
moves in a discrete structure space with moves in parameter space [12, 13, 14, 27].
In this paper, we explore an alternative approach: a completely differentiable system for learning
models defined by sets of first-order rules. This allows one to use modern gradient-based programming
frameworks and optimization methods for the inductive logic programming task. Our approach
is inspired by a differentiable probabilistic logic called TensorLog [5]. TensorLog establishes a
connection between inference using first-order rules and sparse matrix multiplication, which enables
certain types of logical inference tasks to be compiled into sequences of differentiable numerical
operations on matrices. However, TensorLog is limited as a learning system because it only learns
parameters, not rules. In order to learn parameters and structure simultaneously in a differentiable
framework, we design a neural controller system with an attention mechanism and memory to learn
to sequentially compose the primitive differentiable operations used by TensorLog. At each stage of
the computation, the controller uses attention to “softly” choose a subset of TensorLog’s operations,
and then performs the operations with contents selected from the memory. We call our approach
neural logic programming, or Neural LP.
Experimentally, we show that Neural LP performs well on a number of tasks. It improves the
performance in knowledge base completion on several benchmark datasets, such as WordNet18
and Freebase15K [3]. And it obtains state-of-the-art performance on Freebase15KSelected [25],
a recent and more challenging variant of Freebase15K. Neural LP also performs well on standard
benchmark datasets for statistical relational learning, including datasets about biomedicine and
kinship relationships [12]. Since good performance on many of these datasets can be obtained using
short rules, we also evaluate Neural LP on a synthetic task which requires longer rules. Finally, we
show that Neural LP can perform well in answering partially structured queries, where the query is
posed partially in natural language. In particular, Neural LP also obtains state-of-the-art results on the
KB version of the W IKI M OVIES dataset [16] for question-answering against a knowledge base. In
addition, we show that logical rules can be recovered by executing the learned controller on examples
and tracking the attention.
To summarize, the contributions of this paper include the following. First, we describe Neural LP,
which is, to our knowledge, the first end-to-end differentiable approach to learning not only the
parameters but also the structure of logical rules. Second, we experimentally evaluate Neural LP on
several types of knowledge base reasoning tasks, illustrating that this new approach to inductive logic
programming outperforms prior work. Third, we illustrate techniques for visualizing a Neural LP
model as logical rules.
2
Related work
Structure embedding [3, 24, 29] has been a popular approach to reasoning with a knowledge base.
This approach usually learns a embedding that maps knowledge base relations (e.g CityInCountry)
and entities (e.g. USA) to tensors or vectors in latent feature spaces. Though our Neural LP system can
be used for similar tasks as structure embedding, the methods are quite different. Structure embedding
focuses on learning representations of relations and entities, while Neural LP learns logical rules.
In addition, logical rules learned by Neural LP can be applied to entities not seen at training time.
This is not achievable by structure embedding, since its reasoning ability relies on entity-dependent
representations.
Neural LP differs from prior work on logical rule learning in that the system is end-to-end differentiable, thus enabling gradient based optimization, while most prior work involves discrete search
in the problem space. For instance, Kok and Domingos [12] interleave beam search, using discrete
operators to alter a rule set, with parameter learning via numeric methods for rule confidences. Lao
and Cohen [13] introduce all rules from a restricted set, then use lasso-style regression to select a
subset of predictive rules. Wang et al. [27] use an Iterative Structural Gradient algorithm that alternate
gradient-based search for parameters of a probabilistic logic ProPPR [26], with structural additions
suggested by the parameter gradients.
Recent work on neural program induction [21, 20, 1, 8] have used attention mechanism to “softly
choose” differentiable operators, where the attentions are simply approximations to binary choices.
The main difference in our work is that attentions are treated as confidences of the logical rules and
2
have semantic meanings. In other words, Neural LP learns a distribution over logical rules, instead of
an approximation to a particular rule. Therefore, we do not use hardmax to replace softmax during
inference time.
3
Framework
3.1
Knowledge base reasoning
Knowledge bases are collections of relational data of the format Relation (head, tail), where
head and tail are entities and Relation is a binary relation between entities. Examples of such
data tuple are HasOfficeInCity (New York, Uber) and CityInCountry (USA, New York).
The knowledge base reasoning task we consider here consists of a query1 , an entity tail that the
query is about, and an entity head that is the answer to the query. The goal is to retrieve a ranked list
of entities based on the query such that the desired answer (i.e. head) is ranked as high as possible.
To reason over knowledge base, for each query we are interested in learning weighted chain-like
logical rules of the following form, similar to stochastic logic programs [19],
α query (Y, X)←Rn (Y, Zn ) ∧ · · · ∧ R1 (Z1 , X)
(1)
where α ∈ [0, 1] is the confidence associated with this rule, and R1 , . . . , Rn are relations in the
knowledge base. During inference, given an entity x, the score of each y is defined as sum of the
confidence of rules that imply query (y, x), and we will return a ranked list of entities where higher
the score implies higher the ranking.
3.2
TensorLog for KB reasoning
We next introduce TensorLog operators and then describe how they can be used for KB reasoning.
Given a knowledge base, let E be the set of all entities and R be the set of all binary relations. We map
all entities to integers, and each entity i is associated with a one-hot encoded vector vi ∈ {0, 1}|E|
such that only the i-th entry is 1. TensorLog defines an operator MR for each relation R. Concretely,
MR is a matrix in {0, 1}|E|×|E| such that its (i, j) entry is 1 if and only if R (i, j) is in the knowledge
base, where i is the i-th entity and similarly for j.
We now draw the connection between TensorLog operations and a restricted case of logical rule
inference. Using the operators described above, we can imitate logical rule inference R (Y, X)← P (Y,
.
Z) ∧ Q (Z, X) for any entity X = x by performing matrix multiplications MP · MQ · vx = s. In other
words, the non-zero entries of the vector s equals the set of y such that there exists z that P (y, z) and
Q (z, x) are in the KB. Though we describe the case where rule length is two, it is straightforward to
generalize this connection to rules of any length.
Using TensorLog operations, what we want to learn for each query is shown in Equation 2,
X
αl Πk∈βl MRk
(2)
l
where l indexes over all possible rules, αl is the confidence associated with rule l and βl is an ordered
list of all relations in this particular rule. During inference, given an entity vx , the score of each
retrieved entity is then equivalent to the entries in the vector s, as shown in Equation 3.
X
s=
(αl (Πk∈βl MRk vx )) , score(y | x) = vyT s
(3)
l
To summarize, we are interested in the following learning problem for each query.
!
X
X
X
T
max
score(y | x) = max
vy
(αl (Πk∈βl MRk vx ))
{αl ,βl }
{x,y}
{αl ,βl }
{x,y}
(4)
l
where {x, y} are entity pairs that satisfy the query, and {αl , βl } are to be learned.
1
In this work, the notion of query refers to relations, which differs from conventional notion, where query
usually contains relation and entity.
3
Figure 2: The neural controller system.
3.3
Learning the logical rules
We will now describe the differentiable rule learning process, including learnable parameters and
the model architecture. As shown in Equation 2, for each query, we need to learn the set of rules
that imply it and the confidences associated with these rules. However, it is difficult to formulate a
differentiable process to directly learn the parameters and the structure {αl , βl }. This is because each
parameter is associated with a particular rule, and enumerating rules is an inherently discrete task. To
overcome this difficulty, we observe that a different way to write Equation 2 is to interchange the
summation and product, resulting the following formula with a different parameterization,
|R|
T X
Y
t=1
akt MRk
(5)
k
where T is the max length of rules and |R| is the number of relations in the knowledge base. The key
parameterization difference between Equation 2 and Equation 5 is that in the latter we associate each
relation in the rule with a weight. This combines the rule enumeration and confidence assignment.
However, the parameterization in Equation 5 is not sufficiently expressive, as it assumes that all rules
are of the same length. We address this limitation in Equation 6-8, where we introduce a recurrent
formulation similar to Equation 3.
In the recurrent formulation, we use auxiliary memory vectors ut . Initially the memory vector is set
to the given entity vx . At each step as described in Equation 7, the model first computes a weighted
average of previous memory vectors using the memory attention vector bt . Then the model “softly”
applies the TensorLog operators using the operator attention vector at . This formulation allows the
model to apply the TensorLog operators on all previous partial inference results, instead of just the
last step’s.
u0 = vx
(6)
|R|
ut =
X
akt MRk
!
bτt uτ
for 1 ≤ t ≤ T
(7)
τ =0
k
uT+1 =
t−1
X
T
X
bτT +1 uτ
(8)
τ =0
Finally, the model computes a weighted average of all memory vectors, thus using attention to select
the proper rule length. Given the above recurrent formulation, the learnable parameters for each
query are {at | 1 ≤ t ≤ T } and {bt | 1 ≤ t ≤ T + 1}.
We now describe a neural controller system to learn the operator and memory attention vectors.
We use recurrent neural networks not only because they fit with our recurrent formulation, but also
because it is likely that current step’s attentions are dependent on previous steps’. At every step
t ∈ [1, T + 1], the network predicts operator and memory attention vectors using Equation 9, 10,
4
and 11. The input is the query for 1 ≤ t ≤ T and a special END token when t = T + 1.
ht = update (ht−1 , input)
(9)
at = softmax (W ht + b)
(10)
bt = softmax [h0 , . . . , ht−1 ]T ht
(11)
The system then performs the computation in Equation 7 and stores ut into the memory. The memory
holds each step’s partial inference results, i.e. {u0 , . . . , ut , . . . , uT+1 }. Figure 2 shows an overview
of the system. The final inference result u is just the last vector in memory, i.e. uT+1 . As discussed
in Equation 4, the objective is to maximize vyT u. In particular, we maximize log vyT u because the
nonlinearity empirically improves the optimization performance. We also observe that normalizing
the memory vectors (i.e. ut ) to have unit length sometimes improves the optimization.
To recover logical rules from the neural controller system, for each query we can write rules and their
confidences {αl , βl } in terms of the attention vectors {at , bt }. Based on the relationship between
Equation 3 and Equation 6-8, we can recover rules by following Equation 7 and keep track of the
coefficients in front of each matrix MRk . The detailed procedure is presented in Algorithm 1.
Algorithm 1 Recover logical rules from attention vectors
Input: attention vectors {at | t = 1, . . . , T } and {bt | t = 1, . . . , T + 1}
Notation: Let Rt = {r1 , . . . , rl } be the set of partial rules at step t. Each rule rl is represented by
a pair of (α, β) as described in Equation 1, where α is the confidence and β is an ordered list of
relation indexes.
Initialize: R0 = {r0 } where r0 = (1, ( )).
for t ← 1 to T + 1 do
ct = ∅, a placeholder for storing intermediate results.
Initialize: R
for τ ← 0 to t − 1 do
for rule (α, β) in Rτ do
ct .
Update α0 ← α · bτt . Store the updated rule (α0 , β) in R
if t ≤ T then
Initialize: Rt = ∅
ct do
for rule (α, β) in R
for k ← 1 to |R| do
Update α0 ← α · akt , β 0 ← β append k. Add the updated rule (α0 , β 0 ) to Rt .
else
ct
Rt = R
return RT +1
4
Experiments
To test the reasoning ability of Neural LP, we conduct experiments on statistical relation learning, grid
path finding, knowledge base completion, and question answering against a knowledge base. For all
the tasks, the data used in the experiment are divided into three files: facts, train, and test. The facts
file is used as the knowledge base to construct TensorLog operators {MRk | Rk ∈ R}. The train and
test files contain query examples query (head, tail). Unlike in the case of learning embeddings,
we do not require the entities in train and test to overlap, since our system learns rules that are entity
independent.
Our system is implemented in TensorFlow and can be trained end-to-end using gradient methods.
The recurrent neural network used in the neural controller is long short-term memory [9], and the
hidden state dimension is 128. The optimization algorithm we use is mini-batch ADAM [11] with
batch size 64 and learning rate initially set to 0.001. The maximum number of training epochs is 10,
and validation sets are used for early stopping.
4.1
Statistical relation learning
We conduct experiments on two benchmark datasets [12] in statistical relation learning. The first
dataset, Unified Medical Language System (UMLS), is from biomedicine. The entities are biomedical
5
concepts (e.g. disease, antibiotic) and relations are like treats and diagnoses. The second
dataset, Kinship, contains kinship relationships among members of the Alyawarra tribe from Central
Australia [6]. Datasets statistics are shown in Table 1. We randomly split the datasets into facts, train,
test files as described above with ratio 6:2:1. The evaluation metric is Hits@10. Experiment results
are shown in Table 2. Comparing with Iterative Structural Gradient (ISG) [27], Neural LP achieves
better performance on both datasets. 2 We conjecture that this is mainly because of the optimization
strategy used in Neural LP, which is end-to-end gradient-based, while ISG’s optimization alternates
between structure and parameter search.
Table 1: Datasets statistics.
UMLS
Kinship
# Data
# Relation
# Entity
5960
9587
46
25
135
104
Table 2: Experiment results. T indicates the maximum rule length.
ISG
UMLS
Kinship
Figure 3: Accuracy on grid path finding.
4.2
Neural LP
T =2
T =3
T =2
T =3
43.5
59.2
43.3
59.0
92.0
90.2
93.2
90.1
Grid path finding
Since in the previous tasks the rules learned are of length at most three, we design a synthetic task
to test if Neural LP can learn longer rules. The experiment setup includes a knowledge base that
contains location information about a 16 by 16 grid, such as North ((1,2), (1,1)) and SouthEast
((0,2), (1,1)) The query is randomly generated by combining a series of directions, such as
North_SouthWest. The train and test examples are pairs of start and end locations, which are
generated by randomly choosing a location on the grid and then following the queries. We classify
the queries into four classes based on the path length (i.e. Hamming distance between start and
end), ranging from two to ten. Figure 3 shows inference accuracy of this task for learning logical
rules using ISG [27] and Neural LP. As the path length and learning difficulty increase, the results
show that Neural LP can accurately learn rules of length 6-8 for this task, and is more robust than
ISG in terms of handling longer rules.
4.3
Knowledge base completion
We also conduct experiments on the canonical knowledge base completion task as described in [3].
In this task, the query and tail are part of a missing data tuple, and the goal is to retrieve the
related head. For example, if HasOfficeInCountry (USA, Uber) is missing from the knowledge
base, then the goal is to reason over existing data tuples and retrieve USA when presented with
query HasOfficeInCountry and Uber. To represent the query as a continuous input to the neural
controller, we jointly learn an embedding lookup table for each query. The embedding has dimension
128 and is randomly initialized to unit norm vectors.
The knowledge bases in our experiments are from WordNet [17, 10] and Freebase [2]. We use the
datasets WN18 and FB15K, which are introduced in [3]. We also considered a more challenging
dataset, FB15KSelected [25], which is constructed by removing near-duplicate and inverse relations
from FB15K. We use the same train/validation/test split as in prior work and augment data files with
reversed data tuples, i.e. for each relation, we add its inverse inv_relation. In order to create a
2
We use the implementation of ISG available at https://github.com/TeamCohen/ProPPR. In Wang
et al. [27], ISG is compared with other statistical relational learning methods in a different experiment setup, and
ISG is superior to several methods including Markov Logic Networks [12].
6
facts file which will be used as the knowledge base, we further split the original train file into facts
and train with ratio 3:1. 3 The dataset statistics are summarized in Table 3.
Table 3: Knowledge base completion datasets statistics.
Dataset
# Facts
# Train
# Test
# Relation
# Entity
WN18
FB15K
FB15KSelected
106,088
362,538
204,087
35,354
120,604
68,028
5,000
59,071
20,466
18
1,345
237
40,943
14,951
14,541
The attention vector at each step is by default applied to all relations in the knowledge base. Sometimes
this creates an unnecessarily large search space. In our experiment on FB15K, we use a subset of
operators for each query. The subsets are chosen by including the top 128 relations that share common
entities with the query. For all datasets, the max rule length T is 2.
The evaluation metrics we use are Mean Reciprocal Rank (MRR) and Hits@10. MRR computes an
average of the reciprocal rank of the desired entities. Hits@10 computes the percentage of how many
desired entities are ranked among top ten. Following the protocol in Bordes et al. [3], we also use
filtered rankings. We compare the performance of Neural LP with several models, summarized in
Table 4.
Table 4: Knowledge base completion performance comparison. TransE [4] and Neural Tensor
Network [24] results are extracted from [29]. Results on FB15KSelected are from [25].
WN18
Neural Tensor Network
TransE
D IST M ULT [29]
Node+LinkFeat [25]
Implicit ReasoNets [23]
Neural LP
FB15K
FB15KSelected
MRR
Hits@10
MRR
Hits@10
MRR
Hits@10
0.53
0.38
0.83
0.94
0.94
66.1
90.9
94.2
94.3
95.3
94.5
0.25
0.32
0.35
0.82
0.76
41.4
53.9
57.7
87.0
92.7
83.7
0.25
0.23
0.24
40.8
34.7
36.2
Neural LP gives state-of-the-art results on WN18, and results that are close to the state-of-the-art on
FB15K. It has been noted [25] that many relations in WN18 and FB15K have inverse also defined,
which makes them easy to learn. FB15KSelected is a more challenging dataset, and on it, Neural LP
substantially improves the performance over Node+LinkFeat [25] and achieves similar performance
as D IST M ULT [29] in terms of MRR. We note that in FB15KSelected, since the test entities are rarely
directly linked in the knowledge base, the models need to reason explicitly about compositions of
relations. The logical rules learned by Neural LP can very naturally capture such compositions.
Examples of rules learned by Neural LP are shown in Table 5. The number in front each rule is the
normalized confidence, which is computed by dividing by the maximum confidence of rules for each
relation. From the examples we can see that Neural LP successfully combines structure learning
and parameter learning. It not only induce multiple logical rules to capture the complex structure in
the knowledge base, but also learn to distribute confidences on rules.
To demonstrate the inductive learning advantage of Neural LP, we conduct experiments where training
and testing use disjoint sets of entities. To create such setting, we first randomly select a subset of
the test tuples to be the test set. Secondly, we filter the train set by excluding any tuples that share
entities with selected test tuples. Table 6 shows the experiment results in this inductive setting.
3
We also make minimal adjustment to ensure that all query relations in test appear at least once in train and
all entities in train and test are also in facts. For FB15KSelected, we also ensure that entities in train are not
directly linked in facts.
7
Table 5: Examples of logical rules learned by Neural LP on FB15KSelected. The letters A,B,C are
ungrounded logic variables.
1.00 partially_contains(C, A) ← contains (B, A) ∧ contains (B, C)
0.45 partially_contains(C, A) ← contains (A, B) ∧ contains (B, C)
0.35 partially_contains(C, A) ← contains (C, B) ∧ contains (B, A)
1.00 marriage_location (C, A) ← nationality (C, B) ∧ contains (B, A)
0.35 marriage_location (B, A) ← nationality (B, A)
0.24 marriage_location (C, A) ← place_lived (C, B) ∧ contains (B, A)
1.00 film_edited_by (B, A)←nominated_for (A, B)
0.20 film_edited_by (C, A)←award_nominee (B, A) ∧ nominated_for (B, C)
Table 6: Inductive knowledge base completion. The metric is Hits@10.
TransE
Neural LP
WN18
FB15K
FB15KSelected
0.01
94.49
0.48
73.28
0.53
27.97
As expected, the inductive setting results in a huge decrease in performance for the TransE model4 ,
which uses a transductive learning approach; for all three datasets, Hits@10 drops to near zero, as
one could expect. In contrast, Neural LP is much less affected by the amount of unseen entities and
achieves performance at the same scale as the non-inductive setting. This emphasizes that our Neural
LP model has the advantage of being able to transfer to unseen entities.
4.4
Question answering against knowledge base
We also conduct experiments on a knowledge reasoning task where the query is “partially structured”,
as the query is posed partially in natural language. An example of a partially structured query would
be “in which country does x has an office” for a given entity x, instead of HasOfficeInCountry (Y,
x). Neural LP handles queries of this sort very naturally, since the input to the neural controller is a
vector which can encode either a structured query or natural language text.
We use the W IKI M OVIES dataset from Miller et al. [16]. The dataset contains a knowledge base and
question-answer pairs. Each question (i.e. the query) is about an entity and the answers are sets of
entities in the knowledge base. There are 196,453 train examples and 10,000 test examples. The
knowledge base has 43,230 movie related entities and nine relations. A subset of the dataset is shown
in Table 7.
Table 7: A subset of the W IKI M OVIES dataset.
Knowledge base
directed_by (Blade Runner, Ridley Scott)
written_by (Blade Runner, Philip K. Dick)
starred_actors (Blade Runner, Harrison Ford)
starred_actors (Blade Runner, Sean Young)
Questions
What year was the movie Blade Runner released?
Who is the writer of the film Blade Runner?
We process the dataset to match the input format of Neural LP. For each question, we identity the
tail entity by checking which words match entities in the knowledge base. We also filter the
words in the question, keeping only the top 100 frequent words. The length of each question is
limited to six words. To represent the query in natural language as a continuous input for the neural
controller, we jointly learn a embedding lookup table for all words appearing in the query. The
query representation is computed as the arithmetic mean of the embeddings of the words in it.
4
We use the implementation of TransE available at https://github.com/thunlp/KB2E.
8
We compare Neural LP with several embedding based QA models. The main difference between
these methods and ours is that Neural LP does not embed the knowledge base, but instead learns
to compose operators defined on the knowledge base. The comparison is summarized in Table 8.
Experiment results are extracted from Miller et al. [16].
Table 8: Performance comparison. Memory Network is from [28]. QA system is from [4].
Model
Figure 4: Visualization of learned logical rules.
Accuracy
Memory Network
QA system
Key-Value Memory Network [16]
Neural LP
78.5
93.5
93.9
94.6
To visualize the learned model, we randomly sample 650 questions from the test dataset and compute
the embeddings of each question. We use tSNE [15] to reduce the embeddings to the two dimensional
space and plot them in Figure 4. Most learned logical rules consist of one relation from the knowledge
base, and we use different colors to indicate the different relations and label some clusters by relation.
The experiment results show that Neural LP can successfully handle queries that are posed in natural
language by jointly learning word representations as well as the logical rules.
5
Conclusions
We present an end-to-end differentiable method for learning the parameters as well as the structure
of logical rules for knowledge base reasoning. Our method, Neural LP, is inspired by a recent
probabilistic differentiable logic TensorLog [5]. Empirically Neural LP improves performance on
several knowledge base reasoning datasets. In the future, we plan to work on more problems where
logical rules are essential and complementary to pattern recognition.
Acknowledgments
This work was funded by NSF under IIS1250956 and by Google Research.
References
[1] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural
networks for question answering. In Proceedings of NAACL-HLT, pages 1545–1554, 2016.
[2] Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. Freebase: a
collaboratively created graph database for structuring human knowledge. In Proceedings of
the 2008 ACM SIGMOD international conference on Management of data, pages 1247–1250.
ACM, 2008.
[3] Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko.
Translating embeddings for modeling multi-relational data. In Advances in neural information
processing systems, pages 2787–2795, 2013.
[4] Antoine Bordes, Sumit Chopra, and Jason Weston. Question answering with subgraph embeddings. arXiv preprint arXiv:1406.3676, 2014.
[5] William W Cohen. Tensorlog: A differentiable deductive database.
arXiv:1605.06523, 2016.
arXiv preprint
[6] Woodrow W Denham. The detection of patterns in Alyawara nonverbal behavior. PhD thesis,
University of Washington, Seattle., 1973.
[7] Lise Getoor. Introduction to statistical relational learning. MIT press, 2007.
[8] Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka GrabskaBarwińska, Sergio Gómez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou,
et al. Hybrid computing using a neural network with dynamic external memory. Nature, 538
(7626):471–476, 2016.
9
[9] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):
1735–1780, 1997.
[10] Adam Kilgarriff and Christiane Fellbaum. Wordnet: An electronic lexical database, 2000.
[11] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
[12] Stanley Kok and Pedro Domingos. Statistical predicate invention. In Proceedings of the 24th
international conference on Machine learning, pages 433–440. ACM, 2007.
[13] Ni Lao and William W Cohen. Relational retrieval using a combination of path-constrained
random walks. Machine learning, 81(1):53–67, 2010.
[14] Ni Lao, Tom Mitchell, and William W Cohen. Random walk inference and learning in a large
scale knowledge base. In Proceedings of the Conference on Empirical Methods in Natural
Language Processing, pages 529–539. Association for Computational Linguistics, 2011.
[15] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine
Learning Research, 9(Nov):2579–2605, 2008.
[16] Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and
Jason Weston. Key-value memory networks for directly reading documents. arXiv preprint
arXiv:1606.03126, 2016.
[17] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):
39–41, 1995.
[18] Stephen Muggleton, Ramon Otero, and Alireza Tamaddoni-Nezhad. Inductive logic programming, volume 38. Springer, 1992.
[19] Stephen Muggleton et al. Stochastic logic programs. Advances in inductive logic programming,
32:254–264, 1996.
[20] Arvind Neelakantan, Quoc V Le, and Ilya Sutskever. Neural programmer: Inducing latent
programs with gradient descent. arXiv preprint arXiv:1511.04834, 2015.
[21] Arvind Neelakantan, Quoc V Le, Martin Abadi, Andrew McCallum, and Dario Amodei.
Learning a natural language interface with neural programmer. arXiv preprint arXiv:1611.08945,
2016.
[22] Matthew Richardson and Pedro Domingos. Markov logic networks. Machine learning, 62(1-2):
107–136, 2006.
[23] Yelong Shen, Po-Sen Huang, Ming-Wei Chang, and Jianfeng Gao. Implicit reasonet: Modeling
large-scale structured relationships with shared memory. arXiv preprint arXiv:1611.04642,
2016.
[24] Richard Socher, Danqi Chen, Christopher D Manning, and Andrew Ng. Reasoning with neural
tensor networks for knowledge base completion. In Advances in neural information processing
systems, pages 926–934, 2013.
[25] Kristina Toutanova and Danqi Chen. Observed versus latent features for knowledge base and
text inference. In Proceedings of the 3rd Workshop on Continuous Vector Space Models and
their Compositionality, pages 57–66, 2015.
[26] William Yang Wang, Kathryn Mazaitis, and William W Cohen. Programming with personalized
pagerank: a locally groundable first-order probabilistic logic. In Proceedings of the 22nd ACM
international conference on Information & Knowledge Management, pages 2129–2138. ACM,
2013.
[27] William Yang Wang, Kathryn Mazaitis, and William W Cohen. Structure learning via parameter
learning. In CIKM 2014, 2014.
[28] Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. arXiv preprint
arXiv:1410.3916, 2014.
[29] Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Embedding entities and
relations for learning and inference in knowledge bases. In ICLR, 2015.
10
| 2cs.AI
|
A Restaurant Process Mixture Model for
Connectivity Based Parcellation of the Cortex
arXiv:1703.00981v1 [q-bio.NC] 2 Mar 2017
Daniel Moyer, Boris A. Gutman, Neda Jahanshad, and
Paul M. Thompson
Imaging Genetics Center, University of Southern California
[email protected]
Abstract. One of the primary objectives of human brain mapping is
the division of the cortical surface into functionally distinct regions, i.e.
parcellation. While it is generally agreed that at macro-scale different
regions of the cortex have different functions, the exact number and configuration of these regions is not known. Methods for the discovery of
these regions are thus important, particularly as the volume of available
information grows. Towards this end, we present a parcellation method
based on a Bayesian non-parametric mixture model of cortical connectivity.
Keywords: Human Connectome, Cortical Parcellation, Bayesian NonParametrics
1
Introduction
Historically, researchers proposed and investigated regional brain parcellations
through manual dissection and qualitative description [25]. The rise of noninvasive neuro-imaging coupled with advances in computing and computer vision allowed for the exploration of automated parcellation methods, both for
fitting existing atlases to data and for data-driven discovery of functionally and
structurally cohesive parcels [23]. The success of the former propelled the rise
in interest and analyses of brain connectomics in the last decade [20]. Connectomics is a topic of interest within all scales of neuroscience; at the macro-scale,
it is often defined by discrete networks of cortical gray matter regions as nodes
with weighted or binary edges connecting them. In structural connectomics–the
focus of this paper–the edge weights are usually based on counts of estimated
structural connections recovered using diffusion MRI tractography, sometimes
weighted by a microstructural measure.
A number of papers have focused on the converse, using the connection profile
of either voxels or vertices, or in the case of functional MRI, the pairwise signal
correlation for each vertex pair to define the parcellation of the cortex (see [4]
and the references therein). For connectivity based parcellation (CBP) methods
using structural connectivity, two modeling choices are required: 1) a spatial
resolution of the grid on which connections are defined, and 2) the criteria on
which clusters should be formed. In almost every existing method, connectivity
measures are first estimated for a high resolution grid of atomic units (choice
one), e.g. voxels or vertices, and the atomic units then combined under spatial
constraints optimizing a desiderata (choice two).
The first decision is essentially a division of connectivity into two scales: the
macro-region level and higher resolution voxel/vertex level, where our task is
to learn from the former the regions of the latter. The second modeling choice
is the criteria on which these atoms are clustered. Many popular choices come
from more general clustering literature, e.g. within-group sum of squares (explained variance), within-group statistical distances, and mixture model likelihoods [4,15,24]. These criteria use the connectivity profiles of each meso-scale
atom without regard to the network structure they induce at the macro-scale.
In other words, they treat each vertex or voxel as a data point with an associated vector (its row in the meso-scale adjacency matrix), and then cluster based
on this vector space. If one vertex is changed from one group to another, these
methods generally do not re-evaluate the quality of all the groups, though on
the macro-scale each of their connective profiles would have changed.
In the present work, we will address both these choices. We present a method
framed in the context of generative models, specifically Bayesian non-parametric
mixture models which place priors over all possible partitions of the higher resolution grid, and do not require the number of clusters to be predefined. One
large classes of such priors are the so-called “restaurant processes”, used here.
We implement a continuum form of connectivity for our mixture components,
and further leverage a conjugate likelihood-prior relationship to produce closed
form marginal likelihoods for network interactions, allowing efficient sampling.
Our paper is organized as follows: we first define terminology, rigorously
define the parcellation task, and describe the model as a whole. We then describe
each of its components in closer detail. We then present results on two datasets,
and discuss the model in relation to existing models and methods.
2
Model
Let Ω be the white matter/gray matter interface (the inner cortical surface),
with the acknowledgment that Ω is in general composed of two disjoint sheets,
each with a boundary at the medial wall. Fix a coordinate system
S over Ω, and
define
a
parcellation
P
as
any
set
of
regions
{E
}
where
E
⊂
Ω,
Ei = Ω, and
i
i
T
| Ei | = 0 (i.e. the regions Ei are almost disjoint). We assume there exists a
latent parcellation P ∗ that accurately describes the cortical surface with respect
to its underlying neuroanatomical structure. Our objective is the recovery of
P ∗ , specifically using structural connectivity information, and without specifying
the exact number of regions. In order to accomplish this, we construct a joint
generative model of parcellations and connectivity.
We start by choosing a model of partitions. We use the distance dependent Chinese Restaurant Process (ddCRP) [2], a variant of the popular Chinese
Restaurant Process (CRP) non-parametric Bayesian models. CRP models are
most commonly used in mixture models, providing a prior over all possible la-
bel assignments for any number of label-parameter pairs. A main assumption of
the CRP is the exchangeability of the data; the ddCRP removes this exchangeability assumption, allowing for non-trivial topologies of dependence between
data points. This is discussed in Section 2.1 in detail. Briefly, ddCRP allows us
to use non-parametric style mixture models on mesh grids, where we assume a
priori that neighboring patches are dependent. For example, we assume there
is spatial auto-correlation over the discrete manifold of the mesh. Practically
speaking, the ddCRP is the component responsible for merging or splitting the
parcels (clusters), and in general for their configuration.
We next choose a mixture component model; the distribution chosen here will
generate the observed network between estimated regions from the ddCRP. We
choose to follow the style of the Infinite Relational Model [12,1], where we model
interactions between clusters instead of the profiles of the clusters themselves.
Thus, we need a separate parameter for each pair of regions. Before diving into
this however, it is important to consider the form of our connectivity data.
Structural connectivity is estimated using streamlines (tractography), usually
via identifying tracts which intersect the cortical surface at two locations. Thus,
the evidence of connectivity is a set of endpoint pairs on Ω.
In traditional connectivity analysis, these endpoints are counted by region
pair, and a graph is formed from the resulting count statistics. These representations abstract away both knowledge of region geometry such as surface area,
curvature etc, as well as topological information, i.e. region adjacency; this is the
information we will be using in the ddCRP model. While it is possible to ignore
these conflicting motives and directly kluge a graph to a spatial patch model, we
instead attempt to retain spatial intuition in our connectivity representation.
Consider Ω×Ω, the set of all possible tract endpoints intersecting the cortical
surface. We model the observation of these pairs of endpoints as a spatial point
process on Ω×Ω. Assuming that each tract is independently recovered1 , this process is the Poisson point process. That is, for any region pair Ei ×Ej ⊂ Ω×Ω, the
number of tract Rendpoint
pairs observed in that region pair is Poisson distributed
R
with parameter Ei Ej λ(x, y)dydx. Here λ : Ω × Ω → R+ is a non-negative rate
function assumed to be integrable over all Ω × Ω. For Poisson processes, λ completely characterizes the process. While we discuss further the Poisson point
process in Section 2.2, in the view of the overall model it is important to note
one convenient property: disjoint regions have independent counts.
Moving back to the mixture components, we make the following simplifying
assumption on the form of the tract endpoint process: each region pair interacts
in a homogenous manner. That is, we assume λ is constant over any pair of
parcels. Thus, for any finite configuration of K parcels we have on order K 2
non-negative scalar parameters to estimate, and these parameters are the rate
parameters for Poisson spatial processes generating the evidence of connectivity.
We choose to use the Gamma distribution to model these parameters (i.e., each
pair of parcels (gi , gj ) draws a rate λij from a Gamma prior). As shown in
1
It is important to make the distinction between physical fascicles and recovered
tracts. Here, we define the latter to be the reconstructed tractography.
Section 2.2, the conjugacy of the Gamma distribution with the homogenous
Poisson process allows for closed form marginal distributions, and thus efficient
collapsed sampling methods. We also choose to use the mesh faces {fm }M
m=1
as the elements of our ddCRP-mixture. This is because connectivity is usually
defined over intersections of tracts with areal units, and both the ddCRP as well
as the Poisson process naturally operate over such regions.
Putting this all together, and leaving the meaning of cm for the next section,
the model is as follows:
cm , gi ∼ ddCRP(α, Adj)
λij ∼ Gamma(a, b)
Dij ∼ Poisson Point Process(λij )
2.1
The distance dependent Chinese Restaurant Process (ddCRP)
As suggested by its name, distance directed Chinese Restaurant Process is a
variant of the Chinese Restaurant Process (CRP), often used in non-parametric
Bayesian mixture models as a prior over possible mixture components (i.e. a
distribution over distributions). Let α be some positive constant concentration
parameter, and let G0 be a prior distribution over mixture component parameters (for us, a gamma distribution). The original CRP mixture model [17] describes an endless stream of customers (data) entering a restaurant with an infinite number of tables (clusters). Each customer either chooses (with prescribed
probability dependent on α) to sit at an existing table (which has a particular
component distribution) or sit an unoccupied table (draw a new component distribution from the prior). Up to the indexing of the tables, for any finite number
of observations any number of clusters and configuration of cluster associations
is possible.
In the original CRP, the data are assumed to be exchangeable; that is, the
joint likelihood of any observations is invariant under permutations of observation indices. However, in our spatial context we have a topology of face adjacencies. Permutations of the face indexes are non-trivial, and thus the faces are
not exchangeable. To model this, the ddCRP allows each customer to choose
another customer (possibly itself) to sit with based on its dependencies. This
forms a directed graph of seating choices; table assignments are then made to
each group of customers who have chosen to sit with each other, i.e. each connected component of the seating choice graph. Mixture components are drawn
for each table from G0 , and only then are the actual data drawn from each
mixture component. Clearly this is a two stage procedure. In our context, this
means each face will choose to be in a cluster with one of its neighbors or itself.
M
As above, let {fm }M
m=1 be the set of mesh faces, and let {cm }m=1 be the corresponding assignments, where each cm ∈ {1, . . . , M }. We draw cm conditioned
adjacency information Adj as follows:
p(cm
1 if j is adjacent to i
if m = j
= j|Adj) ∝ α
0
otherwise
We denote each cluster of faces as gk for k ∈ {1, . . . , K}, where K is the
number of clusters. Due to our restriction of cm to the indices of faces adjacent to
fm , each gk is a contiguous region, and the set of groups forms a valid parcellation
of Ω. We note that the original ddCRP is defined for more general distance
functions.
2.2
Mixture components: Poisson-Gamma
The evidence of pairwise interaction between regions in structural connectivity
is the set of tract endpoints D = {(xt , yt )}Tt=1 . Since the regional clusters are
defined over discrete grids of areal atoms (mesh faces), these are naturally aggregated to count measures over each pair of sub-regions. For any pair of regions
(gi , gj ) ⊂ Ω × Ω, define Dij = {(xt , yt ) ∈ gi × gj }. We model the counts |Dij |
using the Poisson process with fixed intensity λij , where the area gi ×gj contains
a random count |Dij | distributed
Z
|Dij | ∼ P oisson(
λdxdy)
gi ×gj
Using the independence assumption of the tract endpoints, the likelihood of any
configuration of tract endpoints can then be written
( Z
)
Y
|D |
L(D) =
exp −
λij dxdy λij ij
gi ,gj
gi ×gj
We use a Gamma prior for the λij parameters, the conjugate prior of the Poisson
distribution. Using the Gamma distribution allows us derive a simple closed form
marginal distribution for Dij that “integrates out” the λij ’s, leaving a likelihood
in terms of prior parameters a, b. It is as follows:
Z
Z
P (Dij |a, b) = P (Dij , µ|a, b)dµ = P (Dij |µ)P (µ|a, b)dµ
Z
=
( Z Z
) |Dij |
Y
ba
exp(−bλ)λa−1 dλ
exp −
λdA
λ×
Γ
(a)
gi gj
t=1
{z
}
{z
} |
|
Homogeneous Point Process
Z
= Z(a, b)
=
Here, Z(a, b) =
exp {−(|gi × gj | + b)λ} λ|Dij |+a−1 dλ
|
{z
}
Z(a, b)
=
Z(a0 , b0 )
ba
Γ (a)
Gamma Prior
Un-normalized Gamma Posterior
a
β
|gi × gj | + b
1
|gi × gj | + b
|Dij |
Γ (a + |Dij |)
Γ (a)
2.3
Combined Model and Collapsed Sampling Scheme
We will estimate the model via Collapsed Gibbs Sampling, specifically using
the closed form integral over λij to avoid sampling the interaction parameters.
Starting at iteration ` = 0, we update each c`m by the following conditional
likelihood:
Y
`+1
`
P (c`+1
P (Dij |a, b, c`+1
m = k|D) ∝ P (cm = k)
m = k, {cr }r<m , {cs }s>m )
i,j
Since we assume cm is restricted by our mesh topology, we have only a small
number of options to evaluate. We denote the seating graph edge c`m as a “critical
edge” if for any other node fm0 such that cm0 = m there exists a path to the face
with index c`m . Let gold be fm ’s previous component, and gcrit be the component
of fm without its own edge (its critical component). Without loss of generality
we may further order each neighbor of fm as fn for n = 1, . . . , N , and their
groups as g(n). Using these definitions, for triangle meshes we can write out all
possible scenarios:
1. If cm is not critical, and all neighbors are of the same component before
the update, then we can simply choose cm via P (cm = k), as there is no
difference with respect to the induced components.
2. If cm is not critical, but not surrounded by the same component, then we
are asking essentially “Should cm ’s previously induced component join one
of its as of yet independent neighbors”. Thus,
P (c`+1
m
= n|D) ∝ P (cm = n)
K
Y
P (D∗,k |gnew = gold ∪ gn , gk )
k
×
N
K
Y
Y
P (Dg(n),k |gn̂ , gk )
gn̂ :n̂6=n k
n̂6=old
for each neighbor n. Here D∗,k = {(xt , yt ) ∈ gold ∪ gn × gk }
3. If cm is critical, then for each neighboring component (including the component gold \gcrit in the neighbors) we have
P (c`+1
m = n|D) ∝ P (cm = n)
K
Y
P (D∗,k |gnew = gcrit ∪ gn , gk )
k
×
K
YY
P (Dn̂,k |gn̂ , gk ).
n̂6=n k
We iteratively update the face associates using P (c`+1
= k|D), collecting
m
samples after every pass. While this generates a posterior distribution over cm ,
we simply take the maximum a posteriori (MAP) estimate as our selected parcellation.
In general, updates made in Gibbs sampling algorithms are done sequentially;
this is because strong dependencies between concurrent updates will destabilize
some samplers. However, in cases of low dependence approximate asynchronous
parallel updates have been used with empirically strong results (so called “Hogwild” updates [11]). In our case, most updates are either within components, or
between small components (with correspondingly small interdependencies), so a
small degree of parallelism is possible. In practice we use a compromise between
the serial algorithm and the parallel version: we use a shared memory parallel
sampler for calculating the likelihoods of a small batch ci , then make a serial
updates based on these likelihoods. This allows a roughly linear speed-up in the
number of threads used, though there is a slowly scaling cost of the serial update.
2.4
Implementation notes
In fixing a coordinate system, it is common to split the white matter/gray matter
interface into two spheres, each with a null region where the corpus callosum
bridges the longitudinal fissure. Thus, an easy system can be constructed using
spherical coordinates and a marker for hemisphere.
The symmetry of each tract’s endpoints requires careful consideration to
avoid double counting; while the intuition of the model can be understood without thinking about the symmetry of the data, when evaluating joint probabilities
it is important to only include each data point once. This can be achieved by
only evaluating P (Dij |a, b) for i ≤ j. When computing the parallel updates, in
our experience it is much more efficient to keep the threads active but idle, and
simply have a single thread do the serial update. This avoids the overhead of
repeated thread spawns, which for some implementations/architectures can be
costly.
3
Procedure and Results
In order to test our proposed model, we use two open datasets, one composed
of 20 subjects each scanned twice from the Institute of Psychology, Chinese
Academy of Sciences (IPCAS) subset of the Consortium for Reliability and Reproducibility (CoRR) dataset [26], and the other composed of 30 subjects from
the Human Connectome Project (HCP) S900 release [22]. The pre-processing
differs slightly between the datasets, to account for the different imaging parameters. In general the HCP dataset has higher resolution (both in voxel size
and angular resolution) leading to different tractographies. On each dataset we
compare the performance of the proposed method against two recommended
alternatives: Ward’s method, a greedy hierarchical clustering method [4], and
Spectral Clustering [15].
3.1
Preprocessing and Tractography
IPCAS: T1-weighted (T1w) and diffusion weighted (DWI) images were obtained
on 3T Siemens TrioTim by the original investigators [26] using an 8-channel head
coil and 60 directions. Each subject was scanned twice, roughly two weeks apart.
T1w images were processed with Freesufer’s [5] recon-all pipeline to obtain a triangle mesh of the grey-white matter boundary registered to a shared spherical
space [6]. We resample this space to a geodesic grid (where each face has approximately equal area) with 10,000 total faces, doing so only after computing
tract intersections with the surface. Probabilistic streamline tractography was
conducted using the DWI in 2mm isotropic MNI 152 space, using Dipy’s [7] implementation of constrained spherical deconvolution (CSD) [21] with a harmonic
order of 6. Tractography streamlines were seeded at 2 random locations in each
white matter voxel labeled by FSL’s FAST. Streamline tracking followed directions randomly in proportion to the orientation function at each sample point
at 0.5mm steps, starting bidirectionaly from each seed point with 8 restarts
per seed. As per Dipy’s Anatomically Constrained Tractography (ACT) [19], we
retained only tracts longer than 5mm with endpoints in likely gray matter.
HCP: We used the minimally preprocessed T1-weighted (T1w) and diffusion
weighted (DWI) images rigidly aligned to MNI space. Briefly, the preprocessing
of these images included motion correction and eddy current correction (DWI),
and linear and nonlinear alignment (betweek T1w and DWI). We used the HCP
Pipeline (version 3.13.1) FreeSurfer protocol to run an optimized version of the
recon-all pipeline that computes surface meshes in a higher resolution (0.7mm
isotropic) space. We again resample this space to an geodesic grid after computing tract intersections with the surface. Tractography was conducted using
the DWI in the native 1.25mm isotropic voxel size in MNI space. Probabilistic
streamline tractography was performed as in IPCAS above.
3.2
Fitting and Results
We fit the proposed method using our parallel sampling scheme, using 60 passes
of the sampler with 8 parallel threads (approximately 600,000 updates per subject), using a = 1, b = 1, and α = 0.01. We use the MAP estimate as our
results. We fitted Ward clustering by maximizing Explained Variance over a
naı̈ve search of every possible merge. For the Spectral Clustering method we use
an exponential kernel, using the normalized cosine distance as a metric. We use
a number of eigenvectors equal to the number of clusters. For both baselines we
take the vector of connections as our feature vector. In both clustering schemes,
we specify the number of clusters to be equal to that of the proposed method.
We assess cluster quality using a KL-divergence based measure. We take the
number of tracts from each face fm to each region gi as the objective distribution, and measure how well this is approximated by the average number of
tracts from g(fm ) to gi . These form two matrices of dimension M × K. We then
normalize these matrices to sum to one and measure their KL divergence as in
[15]. If a cluster is well represented by its average connectivity profile, then this
divergence will be low. For the IPCAS dataset we have an additional measure
of Test-Retest reproducibility. This is measured by Normalized Mutual Information (NMI)[4,1], which measures cluster similarity without requiring similar
numbers of clusters. Let Z be a binary matrix of cluster assignments, where, for
Fig. 1. Plots of the KL Divergence based goodness of fit measure, for the three methods,
on both datasets. Here, lower is better.
each row i, each entry
p Zij is 1 if fi is in cluster j and zero otherwise. NMI is
defined as I(Z1 , Z2 )/ (H(Z1 )H(Z2 ), where I(·, ·) is mutual information, H(·)
is entropy, and Z1 and Z2 are the cluster assignments for the first and second
scan respectively. (This uses the convention customary in information theory
that 0 log 0 = 0). NMI is also invariant under permutations of labels. As can be
seen in Figure 1 and Figure 2, the proposed method is performing well compared
to the baseline methods. HCP dataset uses more clusters (around 250 per hemisphere) than the IPCAS dataset (around 175-200 per hemisphere). The difference
here may be due in part to the higher resolution of the HCP dataset, leading to
greater resolving power with respect to the regional connections. These averages
are at the upper range of the number suggested by Van Essen et al. [23].
4
Discussion
This model draws on the wide range of previously proposed methods in connectivity based parcellation. Several non-parametric Bayesian methods have been
proposed, in particular two excellent works Jbabdi et al. [10] and Baldassano et
al. [1], both of whom use Normal-inverse-Wilshart conjugations as their mixture
components (Baldassano et al. use a special case, the Normal-inverse-χ2 ). These
models also enjoy closed form marginal distributions, but do not have infinite
divisibility (the distributions they model are not spatial processes). Jbabdi et
al., whose work predates the ddCRP, use a Dirichlet Process with spatial priors
as their partition prior. They then further define a hierarchical process on top
of this that links multiple subjects. Baldassano et al. use the ddCRP directly,
but model voxel connections, again without the aid of a spatial process. Instead,
they model the aggregate connectivity as coming from a normal distribution.
Fig. 2. Left: Normalized Mutual Information between Test-Retest scans. Here, higher
is better. Right: histograms of the number of clusters selected for each subject.
The ddCRP is similar to a Markov Random Field model with a very strong
spatial prior. These models have been successful in obtaining parcellations from
functional connectivity [9,18], though few if any have used Bayesian non-parametrics.
This frame of reference leads us toward more traditional computer vision tasks
such as pixel labeling, where as in many cases surface parcellation has been
framed as vertex parcellation [3,15,24]. This is a small but relatively important
conceptual difference; the pixel and mesh-face models have areal units, but vertex parcellations are graphs of infinitesimal points. The intuition of the former
leads us toward the use of spatial processes.
A similar spatial process viewpoint of connectivity is proposed in Moyer et
al. [14], but the discovery of new parcellations is not discussed. Poisson count
processes for network interactions have also been explored in the literature [13],
as have infinite relational variants [8] though usually in the context of network
clustering via the stochastic blockmodel (i.e. clustering the regions themselves).
These usually ignore spatial constraints.
Alternative methods to Bayesian models usually specify the number of clusters. Of note is Parisot et al. [15] and a subsequent work by the same authors
[16], which propose spectral methods for the parcellation task, augmented with
a pre-processing local agglomeration. These papers note the propensity for Spectral Clustering to form equi-areal clusters; as can be seen in Figure 3, our method
does not form equi-areal groups. Thus, it may be the case that a lower number
of clusters for spectral clustering may perform better.
As there is a rich body of functional and anatomical knowledge regarding
the cortex, parcellations based on connectivity information alone would need
proper neuroanatomical, histological and functional validation, and more information from these sources would ideally be used to optimize parcellations. The
model presented here uses only spatial constraints and connectivity to estimate
Fig. 3. An exemplar parcellation from an HCP subject. Region colors are random.
feasible parcellations based on recoverable structural connections from imaging.
However, we believe that the modeling techniques explored here can easily be
imputed into larger, multi-modal models, and in general the improvements made
may increase the accuracy and reproducibility of studies of connectivity patterns.
These are critical to furthering our understanding of the living human brain.
Acknowledgements
This work was supported by NIH Grant U54 EB020403, as well as the NSF
Graduate Research Fellowship Program. The authors would like to thank the
reviewers as well as Greg Ver Steeg for multiple helpful conversations.
References
1. Baldassano, C., Beck, D.M., Fei-Fei, L.: Parcellating connectivity in spatial maps.
PeerJ 3, e784 (2015)
2. Blei, D.M., Frazier, P.I.: Distance Dependent Chinese Restaurant Processes. Journal of Machine Learning Research 12(Aug), 2461–2488 (2011)
3. Clarkson, M.J., Malone, I.B., Modat, M., Leung, K.K., Ryan, N., Alexander, D.C.,
Fox, N.C., Ourselin, S.: A framework for using diffusion weighted imaging to improve cortical parcellation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 534–541. Springer (2010)
4. Eickhoff, S.B., Thirion, B., Varoquaux, G., Bzdok, D.: Connectivity-based parcellation: Critique and implications. Human Brain Mapping 36(12), 4771–4792 (2015)
5. Fischl, B.: Freesurfer. NeuroImage 2(62), 774–781 (2012)
6. Fischl, B., et al.: High-resolution intersubject averaging and a coordinate system
for the cortical surface. Human Brain Mapping 8(4), 272–284 (1999)
7. Garyfallidis, E., et al.: Dipy, a library for the analysis of diffusion MRI data. Front.
Neuroinform 8(8) (2014)
8. Hinne, M., et al.: Probabilistic clustering of the human connectome identifies communities and hubs. PloS one 10(1), e0117179 (2015)
9. Honnorat, N., et al.: GraSP: geodesic graph-based segmentation with shape priors
for the functional parcellation of the cortex. NeuroImage 106, 207–221 (2015)
10. Jbabdi, S., Woolrich, M.W., Behrens, T.E.J.: Multiple-subjects connectivity-based
parcellation using hierarchical dirichlet process mixture models. NeuroImage 44(2),
373–384 (2009)
11. Johnson, M., et al.: Analyzing hogwild parallel Gaussian Gibbs sampling. In: Advances in Neural Information Processing Systems. pp. 2715–2723 (2013)
12. Kemp, C., et al.: Learning systems of concepts with an infinite relational model
(2006)
13. Moyer, D., et al.: Mixed membership stochastic blockmodels for the human connectome. MICCAI–Workshop on Bayesian and Graphical Models for Biomedical
Imaging 5, 6
14. Moyer, D., et al.: A continuous model of cortical connectivity. In: International
Conference on Medical Image Computing and Computer-Assisted Intervention.
pp. 157–165. Springer (2016)
15. Parisot, S., et al.: Tractography-driven groupwise multi-scale parcellation of the
cortex. In: International Conference on Information Processing in Medical Imaging.
pp. 600–612. Springer (2015)
16. Parisot, S., et al.: Group-wise parcellation of the cortex through multi-scale spectral
clustering. NeuroImage (2016)
17. Pitman, J., et al.: Combinatorial stochastic processes (2002)
18. Ryali, S., et al.: A parcellation scheme based on von Mises-Fisher distributions
and Markov random fields for segmenting brain regions using resting-state fMRI.
NeuroImage 65, 83–96 (2013)
19. Smith, R.E., et al.: Anatomically-constrained tractography: improved diffusion
MRI streamlines tractography through effective use of anatomical information.
NeuroImage 62(3), 1924–1938 (2012)
20. Sporns, O., Tononi, G., Kötter, R.: The human connectome: a structural description of the human brain. PLoS Comput Biol 1(4), e42 (2005)
21. Tournier, J.D., et al.: Resolving crossing fibres using constrained spherical deconvolution: validation using diffusion-weighted imaging phantom data. NeuroImage
42(2), 617–625 (2008)
22. Van Essen, D.C., Consortium, W., et al.: The WU-Minn human connectome
project: an overview. NeuroImage 80, 62–79 (2013)
23. Van Essen, D.C., Glasser, M.F., Dierker, D.L., Harwell, J., Coalson, T.: Parcellations and hemispheric asymmetries of human cerebral cortex analyzed on surfacebased atlases. Cerebral Cortex 22(10), 2241–2262 (2012)
24. Yeo, B.T., et al.: The organization of the human cerebral cortex estimated by
intrinsic functional connectivity. Journal of Neurophysiology 106(3) (2011)
25. Zilles, K., Amunts, K.: Centenary of Brodmann’s map—conception and fate. Nature Reviews Neuroscience 11(2), 139–145 (2010)
26. Zuo, X.N., et al.: An open science resource for establishing reliability and reproducibility in functional connectomics. Scientific Data 1 (2014)
| 5cs.CE
|
On the Limited Communication Analysis and Design
for Decentralized Estimation
arXiv:1801.05849v1 [cs.SY] 17 Jan 2018
Andreea B. Alexandru
†
Sérgio Pequito
Abstract— This paper pertains to the analysis and design
of decentralized estimation schemes that make use of limited
communication. Briefly, these schemes equip the sensors with
scalar states that iteratively merge the measurements and the
state of other sensors to be used for state estimation. Contrarily
to commonly used distributed estimation schemes, the only
information being exchanged are scalars, there is only one
common time-scale for communication and estimation, and the
retrieval of the state of the system and sensors is achieved in
finite-time. We extend previous work to a more general setup
and provide necessary and sufficient conditions required for
the communication between the sensors that enable the use of
limited communication decentralized estimation schemes. Additionally, we discuss the cases where the sensors are memoryless,
and where the sensors might not have the capacity to discern the
contributions of other sensors. Based on these conditions and
the fact that communication channels incur a cost, we cast the
problem of finding the minimum cost communication graph
that enables limited communication decentralized estimation
schemes as an integer programming problem.
I. I NTRODUCTION
Sensors are often geographically deployed to collect measurements over large-scale networked dynamical systems,
which are used by state estimators (implementing state
observers) to retrieve an estimate of the overall state of the
system. Then, the estimate is provided to the actuator that
implements a controller to steer the dynamical system to the
desired state [1], [2], [3], [4]. Estimators can be full-state,
reduced and extended state observers that implicitly explore
the trade-offs between communication and estimation [5],
[6]. In the context of distributed estimation, additional information used by the state estimators is shared to improve the
quality of the estimate. For example, they can share either
the estimate, the error between predicted state observation
and the measurement, or the innovation used as part of the
state estimation process [7], [8], [4], [9]. The information
is shared by resorting to communication between different sensors’ computational units, and the communication
capabilities impact the ability to retrieve the estimate of
the state. Therefore, it is fundamental to understand which
communication is required to ensure a successful recovery
of the system state [10], [11].
Most of the observers implemented in large-scale systems
require a large amount of information being exchanged
This work was supported in part by the TerraSwarm Research Center,
one of six centers supported by the STARnet phase of the Focus Center
Research Program (FCRP) a Semiconductor Research Corporation program
sponsored by MARCO and DARPA.
† Department of Electrical and Systems Engineering, School of Engineering and Applied Science, University of Pennsylvania
‡ Department of Civil and Environmental Engineering, and Institute for
Data, Systems, and Society, Massachusetts Institute of Technology
†
Ali Jadbabaie
‡
George J. Pappas
†
through communication (i.e., the state, the error between
measurement and predicted state observation). Furthermore,
the estimators are shown to be asymptotically stable (not
always with an arbitrary error decay) which might restrict
the actuation performance in the context of large-scale networked dynamical systems. To overcome such limitations,
in [12], we proposed an approach that equips the sensors
with scalar states which are exchanged with other sensors,
and together with sensors measurements, suffice to retrieve
in finite-time the state of the networked dynamical system
and those of the sensors, which we refer to as limited communication decentralized estimation scheme. Subsequently,
the information being exchanged between sensors is reduced
to the bare minimum, and the communication topologies analyzed are designed to ensure sensor-network state recovery.
In this paper, we seek to better understand the restrictions
and trade-offs of the limited communication decentralized
estimation scheme. Specifically, the main contributions of
this paper are: (i) we waive some implicit assumptions made
in [12] about the communication scheme performed by the
sensors (which are in general only sufficient, as we emphasize in Remark 2); (ii) we explore the implications in two
different setups: (a) the sensors are memoryless (i.e., they do
not keep track of their previous state); and (b) sensors might
not have the capacity to discern the contributions and/or state
of other sensors (e.g., those relying on radio technology); and
(iii) we leverage these new conditions to cast the problem
of determining the minimum communication cost required
to deploy a limited communication decentralized estimation
scheme as an integer programming problem.
II. P ROBLEM S TATEMENT
Let the evolution of a (possibly) large-scale networked
dynamical system be captured by
x[k + 1] = Ax[k],
k = 0, 1, . . .
(1)
n×1
where x[k] ∈ R
is the state of the system. Consider
m sensors with measurements yi ∈ R described as follows:
yi [k] = c⊺i x[k],
i = 1, . . . , m,
(2)
where ci ∈ Rn×1 is the output vector describing the
contributions of the different observed state variables. We
assume that (A, C = [c⊺1 c⊺2 . . . c⊺m ]⊺ ) is observable, but not
necessarily observable from a specific sensor i, i.e., (A, c⊺i )
is not necessarily observable.
In the limited communication decentralized estimation
scheme, we consider that the sensors possess a scalar state
and can communicate with each other. During this process,
they share their states, which enables the retrieval of the state
of both the networked dynamical system and the sensors.
The communication capabilities are captured by a directed
communication graph G = (V, E), where the set of vertices
V = {1, . . . , m} labels the m sensors, and an edge (i, j) ∈
E translates in the capability of sensor j to receive data
from sensor i. Besides, each sensor i computes a linear
combination of the (scalar) measurement and the scalar data
zj ∈ R received from the neighboring sensors, i.e., j ∈ N−
i ,
which can be described as follows:
X
wij zj [k], i ∈ V,
(3)
zi [k + 1] = yi [k] +
j∈N−
i
N−
i
where
= {j ∈ V : (i, j) ∈ E} are the indices of
the in-neighbors of sensor i given by the communication
graph G. Subsequently, we can write (1)-(3) using the following compact representation:
A 0n×m
x̃[k] =: Ã(G)x̃[k],
(4)
x̃[k + 1] =
C W(G)
where x̃ = [x1 . . . xn z1 . . . zm ]⊺ is the augmented system’s
state and W(G) the dynamics between sensors induced
by the communication graph, i.e., [W(G)]ij = wij when
(i, j) ∈ E and zero otherwise. It is worth noticing that some
of the weights wij may be set to zero, and, in particular,
if wii = 0 for i ∈ {1, . . . , m} then we are dealing with
memoryless sensors that work as relays – which cannot be
addressed by the setup explored in [12]. Additionally, the
augmented system’s output is described as follows:
− c⊺i − 01×m
ỹi [k] =
x̃[k] =: C̃i x̃[k], i ∈ V, (5)
i
0|Ji |×n
IJ
m
i
where IJ
m is the sub-matrix containing the rows of the
m × m identity matrix with indices in Ji ⊂ {1, . . . , m}.
In particular, Ji = N−
i when the linear combination of
incoming sensor’s states is performed (locally) at sensor
i, or Ji = {i} when sensors do not have the capacity to
discern the contributions and/or state of other sensors (e.g.,
those relying on radio technology). The latter case cannot be
addressed by the setup explored in [12].
In this paper, we seek solutions to the following problems:
Problem 1: Characterize the necessary and sufficient conditions that must be satisfied by G (and, subsequently, by
W(G)) ensuring that (Ã(G), C̃i ) is observable.
In particular, we provide the characterization required in
the memoryless sensor scenario, and in the case where a
sensor only has access to its own state. Next, we propose to
determine communication topologies that ensure the necessary and sufficient conditions required to solve the previous
problem, while minimizing the communication cost between
the different sensors:
Problem 2: Let Ωe ∈ R+
≥0 ∪ {∞} be the communication cost incurred by establishing a communication link
e = (i, j) ∈ E between the sensors i, j ∈ V to obtain a
communication graph G = (V, E). We aim to determine E
that solves the following optimization problem:
X
Ωe s.t. G satisfies conditions from Problem 1.
min
E
e∈E
III. T ERMINOLOGY
AND
P REVIOUS R ESULTS
In what follows, we rely on structural systems theory [13]
to assess system theoretical properties by considering only
the inter-dependencies between states and sensors. One such
system property is that of structural observability that considers the sparsity binary patterns (Ā, C̄), where an entry in
these matrices is zero if there is no direct dependency between two (state or sensor) variables and one otherwise [14],
[15]. A pair (Ā, C̄) is structurally observable if there exists
an observable pair (A, C) such that the zero entries in
(Ā, C̄) are also zero in (A, C). Subsequently, it can be
proved that if such an observable pair exists, then almost
all possible pairs satisfying the sparsity pattern are also
observable. Furthermore, structural properties (e.g., structural
observability) are necessary to ensure non-structural properties (observability). Therefore, in Section V, we rely on
structural systems to ensure first the necessary conditions,
and then we show that in fact these are also sufficient.
One of the key features of structural systems theory is that
we can interpret the sparsity patterns (Ā, C̄) as a directed
state graph D(Ā) ≡ (X , EX ,X ) and state-output graph
D(Ā, C̄) ≡ (X ∪ Y, EX ,X ∪ EX ,Y ), where the vertices are
labeled by the states and sensors and the edges capture the
inter-dependencies between state and sensor variables as follows: EX ,X = {(xi , xj ) : [Ā]ji 6= 0} and EX ,Y = {(xi , yj ) :
[C̄]ji 6= 0}. We will use EX ,X ∪Y := EX ,X ∪ EX ,Y for
brevity. Additionally, we can use graph-theoretical notions,
e.g., paths and cycles, to address the structural properties.
In particular, to characterize structural observability, we
introduce the notion of bipartite graph associated with the
state graph and state-output graph. The state bipartite graph
B(Ā) ≡ B(X , X , EX ,X ) (resp., the state-output bipartite
graph B(Ā, C̄) ≡ B(X , X ∪ Y, EX ,X ∪Y )), consists of two
sets X (resp., X and X ∪ Y) that can be graphically
interpreted and to which we refer to as left and right set
of vertices. Edges between the left and right set of vertices
encode the dependencies described by the edge-set of the
directed state graph (resp., directed state-output graph). Also,
due to the correspondence between these edges, paths and
cycles in the state-output graph can be captured by subsets
of vertex-disjoint edges in the state and state-output bipartite
graph, which are referred to as matchings, and the subset
with the largest number of edges referred to as maximum
matching. Consequently, those left (resp., right) vertices in
the state and state-output bipartite graph that do not belong
to any edge in the matching are referred to as left-unmatched
(resp., right-unmatched) vertices. Accordingly, we have the
following result:
Lemma 1 ([16]): Consider the digraph D(Ā, C̄) ≡ (X ∪
Y, EX ,X ∪Y ) and let M ∗ be a maximum matching associated
to the state-output bipartite graph B(Ā, C̄) ≡ B(X , X ∪
Y, EX ,X ∪Y ). Then, the digraph D ≡ (X ∪ Y, M ∗ ) comprises
a disjoint union of cycles and elementary paths, from the
left-unmatched vertices to the right-unmatched vertices of
M ∗ , that span D(Ā, C̄). Moreover, such a decomposition
is minimal, in the sense that no other spanning subgraph
decomposition of D(Ā, C̄) into elementary paths and cycles
contains strictly fewer elementary paths.
⋄
The different graph-theoretic concepts can come together
to assess structural observability of (Ā, C̄) as follows.
Theorem 1 ([12]): Let D(Ā, C̄) ≡ (X ∪ Y, EX ,X ∪Y )
denote the state-output digraph and B(Ā, C̄) the state-output
bipartite representation. The pair (Ā, C̄) is structurally observable if and only if the following two conditions hold:
(i) there is a path from every state vertex to an output
vertex in D(Ā, C̄); and
(ii) there exists a maximum matching M ∗ associated
to B(Ā, C̄) such that the left-unmatched vertices
UL (M ∗ ) = ∅.
⋄
Therefore, as previously mentioned, we can build upon
these results to analyze and design the limited communication decentralized estimation scheme. In [12], we have
provided necessary and sufficient conditions that G needs
to satisfy to ensure observability of (Ã(G), C̃i ) under the
following simplifying implicit assumption.
Implicit Assumption [12]: Each sensor retains its previous
state that is always weighted in the sensor dynamics (3), i.e.,
wii 6= 0, which implies that in (4) we have [diag(W(G))]ii 6=
0 for all i = 1, . . . , m.
◦
In other words, [12] excludes the case of memoryless sensors, which we address in this paper. We also explore other
setups, e.g., when the sensors are not able to differentiate
the individual contributions of other sensors due to the technology used. Next, we state two of the main results in [12]
for ease of comparison with the main results attained in this
work, where we waive the implicit assumption stated above.
Theorem 2 ([12]): Let D(Ã(G)) ≡ (V ≡ (X ∪ Z), EV,V )
be the state digraph, where X corresponds to the labels of
the state vertices and Z to the labels of the sensors’ states.
In addition, let Ni− = {v ∈ V : (v, zi ) ∈ E} be the set of
in-neighbors of a vertex zi representing a sensor in D(Ã(G)),
i = 1, . . . , m. The following two conditions are necessary
and sufficient to ensure that (Ã(G), C̃i ), for i = 1, . . . , m,
is generically observable:
(i) for every z ∈ Z there must exist a directed path from
any v ∈ V;
(ii) for every z ∈ Z there must exist a set of
left-unmatched vertices UL , associated with a maximum matching of the bipartite representation of
D(Ã(G)), such that UL ⊂ Ni− and UL ∩ X = ∅. ⋄
Hence, Theorem 2 can be used to obtain the next result
to Problem 1 under the implicit assumption stated above.
Theorem 3 ([12]): If (A, C) is observable and
(Ã(G), C̃i ) is structurally observable ∀i = 1, . . . , m, then
almost all realizations of W(G) ensure that (Ã(G), C̃i ) is
observable.
⋄
For brevity’s sake, we will use the shortened notation à =
Ã(G) in the rest of the paper.
IV. L IMITED C OMMUNICATION A NALYSIS
AND
D ESIGN
In this section, we introduce the main results of this paper.
Lemma 2 shows that the sensor capabilities impose strong
constraints on the network’s structure required to ensure
structural observability. This technical result plays a key role
in understanding Theorem 4, which states the necessary and
sufficient conditions required to address Problem 1. Specifically, it provides the conditions for the communication graph
G such that (Ã, C̃i ) is observable, i = 1, . . . , m. Next, we
consider the design of communication graphs that attain the
former conditions, while minimizing the total cost incurred
by the communication between the sensors (Problem 2). In
particular, we cast the problem as an integer programming
problem that can be solved with off-the-shelf solvers.
We start by showing that the structural observability of a
pair (Ā, C̄), that is often assessed through the state-output
graph properties (as captured in Theorem 1), can enforce a
particular structure of Ā under certain sensing capabilities.
Lemma 2: Let ei ∈ Rp×1 be the canonical column-vector
with one in the i’th position and the remaining entries equal
to zero. Given a structured adjacency matrix of a graph,
M̄ ∈ {0, 1}p×p , the pairs (M̄, e⊺i ), for i = 1, . . . , p, are
structurally observable if and only if the associated state
digraph D(M̄) is strongly connected and spanned by a
disjoint union of cycles.
⋄
Proof: (Necessity) Assume (M̄, e⊺i ) is structurally
observable, for all i = 1, . . . , p. Suppose by contradiction
that D(M̄) is not strongly connected. If D(M̄) is not strongly
connected, then the state-output digraph D(M̄, e⊺i ) is also
not strongly connected and its directed acyclic representation
contains a number of strongly connected components. Since
the output vertex in D(M̄, e⊺i ) is one vertex yi connected
only to vertex xi , then it readily follows that condition (i) of
Theorem 1 cannot hold.
Next, we prove that D(M̄) is spanned by a disjoint union
of cycles. Consider condition (ii) of Theorem 1: there exists
a maximum matching in B(M̄, e⊺i ) such that there is no
left-unmatched vertex in X . There are two possibilities for
the maximum matchings in B(M̄): (a) UL = ∅ and (b)
UL 6= ∅. Case (a) means that there is a perfect matching in
B(M̄), i.e., from Lemma 1, D(M̄) is spanned by a disjoint
union of cycles.
We want to prove now that case (b) cannot happen when
(M̄, e⊺i ) is structurally observable. Let us first address the
case when |UL | = 1. Let X represent the set of vertices in the
graph described by M̄. Consider a maximum matching M 1
in B(M̄) that has UL (M 1 ) = {xj } and UR (M 1 ) = {xk },
for some vertices xj , xk ∈ X . Since D(M̄) is strongly
connected, then, there exists a neighbor xl of xk such that
(xl , xk ) ∈ EX ,X . Let M 2 be another maximum matching
in B(M̄) such that UL = {xl } and UR = {xm }, for
some vertex xm ∈ X . By Lemma 4 in [16], there exists
a maximum matching M ∆ such that UL (M ∆ ) = {xl } and
UR (M ∆ ) = {xk }. However, M ∗ = M ∆ ∪ {(xl , xk )} is also
a maximum matching, in fact, a perfect matching, which
leads to a contradiction of the fact that M ∆ is a maximum
matching. Therefore, the set of left-unmatched vertices has
to be empty, meaning a maximum matching is also a perfect
matching, leading to the fact that D(M̄) is spanned by a
disjoint union of cycles. Now, for the case when |UL | > 1,
we can iteratively find augmented paths [17] and construct
larger cardinality maximum matchings, while thus reducing
the cardinality of the set of left-unmatched vertices with
respect to those matchings, until |UL | = 1.
(Sufficiency) Assume D(M̄) is strongly connected and
spanned by a disjoint union of cycles. It follows that also
D(M̄, e⊺i ) is strongly connected and condition (i) from
Theorem 1 is satisfied. Since D(M̄) is spanned by a disjoint
union of cycles, by Lemma 1, there exists a perfect matching
M ∗ (which is also a maximum matching) in B(M̄, e⊺i ). This
implies condition (ii), i.e., UL (M ∗ ) = ∅.
Remark 1: In the proof of Lemma 2, the technical challenge is to show that the state vertices in the state-output
bipartite graph cannot be always matched by an edge whose
right-vertex is a sensor vertex (when the state graph is
strongly connected), which implies that those states need
to be always matched by edges whose end-points are state
vertices. Therefore, by leveraging Lemma 1, it follows that
the state graph has to be spanned by cycles.
⋄
V. M AIN R ESULTS
Before we present the solution to the former problem, we
need to review the notion of linking (see, for instance, [13])
from the vertices in the state digraph to the vertices in the
communication digraph in the state-output graph associated
with the augmented system. Specifically, a linking P is a set
of vertex-disjoint and simple paths in D(Ã), from the vertices in D(Ā) to the vertices in D(W(G)). Additionally, for
each sensor i, we denote by Pi the communication-linking
(a linking where both the starting and ending vertices in the
vertex-disjoint simple paths belong to the communication
graph) from the set of sensor vertices that belong to the
edges in a maximum matching of B(Ā, C̄) to a subset
of in-neighbors of sensor i (Ji ) with equal cardinality. In
particular, there are as many of those sensor vertices as
left-unmatched vertices in a maximum matching associated
to B(Ā) due to the observability of (A, C), and its structural
observability, as prescribed by Theorem 1. Consequently, the
solution to Problem 1 can be formally stated as follows.
Theorem 4: Consider the system (4)-(5). For (A, C) observable, the pair (Ã, C̃i ) is observable for all i = 1, . . . , m,
if and only if D(W(G)) is strongly connected and there
exists a linking Pi such that D(W(G \ Pi )) is spanned by a
disjoint union of cycles, for all i = 1, . . . , m.
⋄
Proof: (Necessity) It is enough to show that one
condition from Theorem 2 or Theorem 3 does not hold when
D(W(G \Pi )) is not spanned by a disjoint union of cycles or
that D(W(G)) is not strongly connected. Therefore, assume
that (Ã, C̃i ) is structurally observable ∀i = 1, . . . , m and
D(W(G)) is not strongly connected. The proof follows
along the same lines as the proof of necessity in Lemma 2
since condition (i) of Theorem 2 fails if D(W(G)) is not
strongly connected. Now, assume that (Ã, C̃i ) is structurally
observable ∀i = 1, . . . , m and D(W(G \ Pi )) is not spanned
by a disjoint union of cycles. Using the second part of the
proof of Lemma 2 for M̄ corresponding to each of the
strongly connected components of D(W(G \ Pi )) proves the
contradiction to condition (ii) of Theorem 2 if D(W(G \Pi ))
is not spanned by a disjoint union of cycles.
(Sufficiency) Assume D(W(G \ Pi )) is spanned by a
disjoint union of cycles for all i = 1, . . . , m, and D(W(G))
is strongly connected. In order to prove sufficiency, we follow
similar steps to those in the proof in [12] and show that the
same conditions are satisfied by a more general W(G).
The necessity and sufficiency of condition (i) and sufficiency of condition (ii) in Theorem 2 follow as in [12],
since D(W(G)) is assumed to be strongly connected. The
original system (1)-(2) is also structurally observable, and,
from Theorem 1, we know there exists a maximum matching
M associated to B(Ā) such that, for every left-unmatched
vertex x ∈ UL (M ), there is a distinct sensor associated to
it. Expanding the maximum matching M to the augmented
system’s bipartite state graph B(Ã), we can match all the
vertices x ∈ UL (M ) with a distinct sensor measuring it.
This yields that the only possible left-unmatched vertices in
B(Ã) are z ∈ Z, hence, UL ∩ X = ∅. All sensors zj that
are not right-matched by a path from a previously unmatched
state vertex are spanned by disjoint cycles, by the assumption
on D(W(G)). Then, either these left-unmatched vertices are
already in-neighbors of sensor zi , or, following a similar
procedure as in the proof of Lemma 2, we can find another
maximum matching such that UL ⊂ Ni− .
Next, to show that there exists a realization of W(G)
that ensures observability of (Ã, C̃i ), we leverage the proof
of Theorem 3 in [12]. Specifically, we invoke the PopovBelevitch-Hautus criterion to assess the observability of the
system (Ã, C̃i ). As a result, W(G) must be such that the
following equalities hold for λ ∈ C, i = 1, . . . , m:
A−λIn
0n×m
C
W(G)−λIm
à − λIn+m
⊺
= rank − ci − 01×m = n + m.
rank
−
C̃i
N
0|N− |×n
Imi
{z
}
|
i
M
The structure of W(G) does not allow arbitrary placing
of the eigenvalues, as opposed to the proof of Theorem 3
in [12]. However, we are able to prove that the eigenvalues
of W(G) that we cannot place do not affect the rank of M.
The eigenvalues associated with the cycles in W(G \ Pi )
can be arbitrarily placed: for each cycle, composed of edges
with weights w1 , . . . , wrj , where 1 ≤ j ≤ |C| and |C| is
the number of cycles, the associated eigenvalues will have
√
2πi k−1
r
rj
, k = 1, . . . , rj .
the values: λjk = j w1 · . . . · wrj e
The eigenvalues that cannot be placed are associated to the
paths Pi and will be zero [18], [19]. The same analysis
holds for A, i.e., the zero eigenvalues are associated with
the paths that are not spanned by cycles. More specifically,
the vertices on these paths are exactly the left-unmatched
vertices with respect to a maximum matching in B(Ã, C̃i ).
In the state-output digraph D(Ã, C̃i ), in order to match
these left-unmatched vertices, the paths are extended through
the links described by C to the sensors’ states, in the
communication graph D(W(G)). Let p be the number of
vertices in the minimum length paths in D(Ā) and D(W(G))
that are not spanned by cycles, i.e., corresponding to the zero
eigenvalues in A and W(G). By suitable permutations, we
can separate the blocks in A and W(G) associated to the
disjoint cycles, denoted symbolically by AC , WC and the
blocks associated to the linkings Pi , respectively AP , WP :
" AP 0 ∗ 0 #
A 0n×m
= 0∗ W0P A0C 0∗ .
M :=
C W(G)
0
∗
0 W
C
The eigenvalues in WP can be chosen to be different than
the eigenvalues of AC , which are non-zero, and different than
zero, hence M has n+m−p non-zero eigenvalues. Moreover,
the end-vertices of the linkings Pi are measured by the
N−
N−
outputs given by Imi . Therefore, the pair (M(1:p,1:p) , Imi )
is observable and, by Popov’s criterion,
#
"
M(1:p,1:p) − λIp
= p.
rank
N−
Imi
This completes the proof that rank(M) = n + m, i.e., the
conditions of Theorem 4 are sufficient.
Remark 2: Theorem 4 accounts for scenarios where the
sensors are memoryless, i.e., they do not retain their previous
state to integrate it in the overall dynamics. This extends
the results in [12], revisited in Section III. Specifically, the
case where sensors are not readily memoryless leads to the
case where the communication graph is strongly connected
and has a subgraph spanned by a disjoint union of cycles,
since the access of a sensor to its state and incorporation
in the overall dynamics corresponds to a self-loop in the
communication graph, which is an elementary cycle.
⋄
Remark 3: In the context of limited communication decentralized estimation schemes that employ sensors which
cannot discern between the contributions coming from their
neighbors due to the technology used, i.e., when Ji = {i}
in (5), it follows that A is at most rank n − 1. More specifically, there will be only one possible communication-linking
ending at the i’th sensor vertex, implying that the state
bipartite graph’s maximum matching can have at most one
left-unmatched vertex.
⋄
Finally, given the necessary and sufficient conditions for
the communication graph provided in Theorem 4, we aim
to formulate the problem of designing a minimum cost
communication graph, as stated in Problem 2, as an integer
programming problem. To this end, we leverage problems
such as the minimum cost maximum matching problem and
minimum cost spanning trees [20]. We also need to formulate
the conditions on the communication graph, which require
to encode the minimum cardinality linkings Pi .
To better visualize the results, we write the communication
graph as G = (Z, EZ,Z ), where Z represents the set of sensor
states, and EZ,Z the set of communication links between the
sensors. Let Ωe ∈ R+
≥0 ∪ {∞} be the communication cost
incurred by establishing a link e = (i, j) ∈ EZ,Z from sensor j to sensor i. If we want to obtain a communication graph
dealing with memoryless sensors, then we prescribe Ωii =
∞, and obtain a finite cost graph as a feasible solution.
Briefly, the constraints that have to be satisfied by G can
be described in the following algorithm where steps are
addressed simultaneously: for every sensor i = 1, . . . , m,
1. Find the number of left-unmatched vertices in B(Ā);
2. Find the minimum cost linking Pi ;
3. Run the minimum cost maximum matching algorithm on
G \ Pi and select the edges that compose it; and
4. Add the minimum cost edges such that the graph G is
strongly connected.
We can leverage some insights provided by the heuristic
algorithm provided in [12] to obtain an integer programming
problem formulation without explicitly computing Pi . For a
sensor i, add a ‘virtual output’ (i.e., not part of the sensing
technology but with the same role under this intermediate
N−
step) to each of its in-neighbors, according to Imi . Denote
this set of vertices by S i . Next, let B(Ã, C̃i , S i ) = (X ∪
Z, X ∪ Z ∪ S i , EX ∪Z,X ∪Z∪S i ) be the state-output bipartite
graph of the system with virtual outputs and D(Ã, C̃i , S i )
the associated state-output digraph. We then expand the cost
structure as follows: assign weights Ωi w.r.t. sensor i to all
the edges that are not in the communication graph G and
Ωie = 0 for e ∈ EX ∪Z,X ∪Z ∪ EZ,S i . This setup ensures
that the minimum cost maximum matching algorithm in the
former state-output bipartite graph will return a matching that
partitions the state-output digraph in vertex-disjoint paths and
cycles, while incurring the minimum cost. Specifically, the
paths will contain those described by Pi , and the rest of
the digraph will be spanned by disjoint cycles. Furthermore,
notice that since there are no edges from the communication
digraph to the state digraph, there can be no cycle spanning
both vertices in X and vertices in Z.
To state the integer programming formulation, let us
denote by ue the binary variable associated to the existence
of edge e ∈ D(Ã, C̃i , S i ), i = 1, . . . , m. For a set S ⊂ V,
the cutset δ − (S) ⊂ E represents a subset of edges with the
start vertex in S and the end vertex in V \ S, for a given set
of edges E and set of vertices V. Hence, the constraints are
given by the matching problem on the state-output digraph
and by the strong connectivity of the communication graph,
which is imposed via rooted minimum spanning tree for each
vertex. For brevity, denote E i := EZ,Z∪S i . Thus, we obtain
the following formulation of Problem 2:
X
min
Ωe u e
ue
s.t
e∈∪m
i=1 EX ∪Z,X ∪Z∪S i
X
e∈δ − (v)
X
e∈δ − (S)
ue ≤ 1, v ∈ Z,
ue ≥ 1, ∀ ∅ ( S ( Z, v ∈
/ S, ∀ v ∈ Z,
ue = 1, ∀e ∈ EX ,X ∪ EX ,Z and ue = 0 otherwise,
ue ∈ {0, 1}, ∀e ∈ E i , i = 1, . . . , m.
where the edges in the minimum cost communication
graph G ∗ can be retrieved from U := {e ∈ EZ,Z |ue = 1}.
The design problem proposed in Problem 2 is NP-hard,
since it contains as a particular instance the design problem
addressed in [12]. Hence, a straightforward greedy algorithm
can be implemented by sequentially performing the steps
described in the pseudo-algorithm above. Nonetheless, the
solution will depend on the initial point since, at each
iteration, the previous selected edges in E will be set to one,
which does not guarantee that the final configuration of G has
indeed minimum cost. Consequently, we leverage the integer
y1
y2
y3
y4
y5
x1
x2
x3
x4
x5
Fig. 1. Plant with 5 state nodes (black) and 5 sensors deployed (blue).
The interconnections between the state nodes are depicted in black and the
measurement terminals between state nodes and sensors are depicted in red.
programming formulation which is also known to be NP-hard
in general, but which we can solve by resorting to highly
optimized off-the-shelf software toolboxes (e.g., YALMIP)
that have been efficiently deployed in practice when dealing
with large-scale complex problems [20], [21].
VI. I LLUSTRATIVE EXAMPLE
Consider the example in Figure 1 and associate the following cost matrix for the communication links:
Ω(G) =
"∞
1 1 1 1
1 ∞ 1 1 1
1 1 ∞ 1 1
1 1 1 ∞ 1
1 1 1 1 ∞
#
.
The minimum topology for the communication graph G
such that decentralized observability from all sensors is
ensured is depicted in Figure 2. Since there are two leftunmatched vertices in B(Ā), each sensor should have
at least two in-neighbors. Here, we illustrate how decentralized observability from sensor 1 is achieved. We
need to find a linking P1 such that G \ P1 is spanned
by a disjoint union of cycles. In this case, pick a
maximum matching in B(Ā, C̄) composed of the edges
(x1 , y1 ), (x2 , y2 ), (x3 , x2 ), (x4 , x3 ), (x5 , x4 ). The linking P1
from the matched sensors to the in-neighbors of sensor 1 can
be chosen as (y1 , y4 , y5 )∪(y2 , y3 ). We obtain that G\P1 = ∅,
which is trivially spanned by a union of disjoint cycles.
VII. C ONCLUSIONS
In this paper, we have extended the limited communication
decentralized estimation schemes to cope with general scenarios and provided necessary and sufficient conditions for
the communication graph such that retrieval of the state of
the system and sensors is possible. In particular, the present
extension enables the deployment of limited communication
decentralized estimation schemes in the scenarios where the
sensors are memoryless, and where the sensors do not have
the capacity to discern the contributions and/or state of other
sensors. Furthermore, we cast the design problem under
communication costs, i.e., the problem of determining the
minimum cost communication graph required to implement
a limited communication decentralized estimation scheme, as
an integer programming problem. This formulation enables
the use of off-the-shelf software toolboxes that are reliable
in practice when dealing with large-scale complex problems.
y1
y2
y3
y4
y5
x1
x2
x3
x4
x5
Fig. 2.
Minimum communication topology that ensure decentralized
observability for the plant in Figure 1.
Future research will focus on proposing energy-efficient
communication protocols between the sensors when subject
to constrained energy budgets. Towards this goal, it is important to understand the trade-offs between communication and
the information contained in the sensors states. In particular,
we aim to quantify and classify the role of the sensors’ state
dimension in the estimation process and accuracy, which
can be key when adding communication infrastructure is
prohibitive. Our findings suggest that as the number of
dimensions of the exchanged states increases, fewer communication links are required to guarantee decentralized
observability. Moreover, one can design the dimension of the
sensors’s memory such that no additional links are necessary.
R EFERENCES
[1] V. Gupta, “Distributed estimation and control in networked systems,”
Ph.D. dissertation, California Institute of Technology, 2007.
[2] Y.-F. Huang, S. Werner, J. Huang, N. Kashyap, and V. Gupta, “State
estimation in electric power grids: Meeting new challenges presented
by the requirements of the future grid,” IEEE Signal Processing
Magazine, vol. 29, no. 5, pp. 33–43, 2012.
[3] S. Kar, “Large scale networked dynamical systems: Distributed inference,” Ph.D. dissertation, Carnegie Mellon University, 2010.
[4] U. A. Khan, “High-dimensional consensus in large-scale networks:
Theory and applications,” Ph.D. dissertation, Carnegie Mellon University, 2009.
[5] L. Wang and A. Morse, “A distributed observer for a time-invariant
linear system,” in American Control Conference (ACC), 2017. IEEE,
2017, pp. 2020–2025.
[6] A. Mitra and S. Sundaram, “Distributed observers for lti systems,”
arXiv preprint arXiv:1608.01429, 2016.
[7] F. Garin and L. Schenato, “A survey on distributed estimation and
control applications using linear consensus algorithms,” in Networked
Control Systems. Springer, 2010, pp. 75–107.
[8] S. Das and J. M. Moura, “Distributed Kalman filtering with dynamic
observations consensus,” IEEE Transactions on Signal Processing,
vol. 63, no. 17, pp. 4458–4473, 2015.
[9] M. V. Subbotin, “Distributed decentralized estimation,” Ph.D. dissertation, University of California at Santa Barbara, 2008.
[10] U. A. Khan and A. Jadbabaie, “Coordinated networked estimation
strategies using structured systems theory,” in 50th Conference on
Decision and Control and European Control Conference. IEEE, 2011,
pp. 2112–2117.
[11] U. A. Khan, S. Kar, A. Jadbabaie, and J. M. Moura, “On connectivity,
observability, and stability in distributed estimation,” in 49th IEEE
Conference on Decision and Control. IEEE, 2010, pp. 6639–6644.
[12] A. B. Alexandru, S. Pequito, A. Jadbabaie, and G. J. Pappas, “Decentralized observability with limited communication between sensors,” in
55th Conference on Decision and Control. IEEE, 2016, pp. 885–890.
[13] J.-M. Dion, C. Commault, and J. Van Der Woude, “Generic properties
and control of linear structured systems: a survey,” Automatica, vol. 39,
no. 7, pp. 1125–1144, 2003.
[14] C.-T. Lin, “Structural controllability,” IEEE Transactions on Automatic
Control, vol. 19, no. 3, pp. 201–208, 1974.
[15] R. Shields and J. Pearson, “Structural controllability of Multi-Input
linear systems,” IEEE Transactions on Automatic control, vol. 21,
no. 2, pp. 203–212, 1976.
[16] S. Pequito, S. Kar, and A. P. Aguiar, “A framework for structural input/output and control configuration selection in large-scale systems,”
IEEE Transactions on Automatic Control, vol. 61, no. 2, pp. 303–318,
2016.
[17] R. Diestel, Graph theory, ser. Graduate Texts in Mathematics.
Springer, 2005.
[18] C. Coates, “Flow-graph solutions of linear algebraic equations,” IRE
Transactions on circuit theory, vol. 6, no. 2, pp. 170–187, 1959.
[19] K. J. Reinschke, Multivariable control: a graph theoretic approach.
Springer-Verlag, 1988.
[20] G. L. Nemhauser and L. A. Wolsey, Integer programming and combinatorial optimization. Springer, 1988, vol. 20.
[21] J. Lofberg, “YALMIP: A toolbox for modeling and optimization in
MATLAB,” in International Symposium on Computer Aided Control
Systems Design. IEEE, 2004, pp. 284–289.
| 3cs.SY
|
INTERPLAY BETWEEN HOMOLOGICAL DIMENSIONS OF A
COMPLEX AND ITS RIGHT DERIVED SECTION
arXiv:1404.3982v2 [math.AC] 28 Jul 2016
CYRUS JALALI
Abstract. Let (R, m) be a commutative Noetherian local ring, a be a proper ideal
of R and M be an R-complex in D(R). We prove that if M ∈ Df❁ (R) (respectively,
M ∈ Df❂ (R)), then idR RΓa (M ) = idR M (respectively, fdR RΓa (M ) = fdR M ). Next,
it is proved that the right derived section functor of a complex M ∈ D❁ (R) (R is not
necessarily local) can be computed via a genuine left-bounded complex G ≃ M of
Gorenstein injective modules. We show that if R has a dualizing complex and M is an
R-complex in Df (R), then GfdR RΓa (M ) = GfdR M and GidR RΓa (M ) = GidR M .
Also, we show that if M is a relative Cohen-Macaulay R-module with respect to a
(respectively, Cohen-Macaulay R-module of dimension n), then GfdR HahtM a (M ) =
GfdR M + n (respectively, GidR Hn
m (M ) = GidR M − n). The above results generalize
some known results and provide characterizations of Gorenstein rings.
1. Introduction
Throughout this paper, R is a commutative Noetherian ring, a is a proper ideal of
R and M is an R-complex. The category of R-complexes is denoted C(R), and we use
subscripts ❁, ❂ and to denote genuine boundedness conditions. So, C❂ (R) is the
full subcategory of C(R) of bounded below complexes (see [2, Definition 2.1.1]). Also,
the derived category is denoted D(R), and we use subscripts ❁, ❂ and to denote
homological boundedness conditions (see [2, Definition 4.1.15]). The symbol ≃ is the
sign for isomorphism in D(R) and quasiisomorphisms in C(R). We also use superscript f to signify that the homology modules are degreewise finitely generated. An
R-complex I is semiinjective if the functor HomR (−, I) converts injective quasiisomorphisms into surjective quasiisomorphisms. A semiinjective resolution of an R-complex
≃
M is a semiinjective complex I and a quasiisomorphism M −→ I. For an R-complex
M the injective dimension idR M is defined as
o
n
R−complex I such that
.
idR M = inf ℓ ∈ Z | ∃Msemiinjective
≃I in D(R) and Iv =0 f or all v<−ℓ
Several of the main results of this paper involve the hypothesis that R has a dualizing
complex. A complex D ∈ Df (R) is dualizing for R if it has finite injective dimension and
the canonical morphism χR
M : R → RHomR (D, D) is an isomorphism in D(R). If R has
a dualizing complex D, we may consider the functor −† = RHomR (−, D). The notion
of Gorenstein injective module was introduced by E.E. Enochs and O.M.G. Jenda in
[6]. An R-module M is said to be Gorenstein injective, if there exists a HomR (I, −)
exact acyclic complex E of injective R-modules such that M = Ker(E0 → E−1 ). The
Gorenstein injective dimension, GidR M , of M ∈ D❁ (R) is defined to be the infimum
of the set of integers n such that there exists a complex G ∈ C❁ (R) consisting of
Gorenstein injective modules satisfying M ≃ G and Gn = 0 for n < −ℓ. Also, an
2000 Mathematics Subject Classification. 13D05, 18E30.
Key words and phrases. Flat dimension, Injective dimension, Gorenstein injective dimension, Derived local cohomology.
1
2
C. JALALI
R-complex F is semiflat if the functor − ⊗R F preserves injective quasiisomorphisms.
For an R-complex M the flat dimension fdR M is defined as
o
n
semif lat R−complex F such that
fdR M = inf n ∈ Z | F∃≃M
in D(R) and Fv =0 f or all v>n .
An R-module M is said to be Gorenstein flat, if there exists an I ⊗R − exact acyclic
complex F of of flat R-modules such that M = Ker(F0 → F−1 ). The Gorenstein flat
dimension, GfdR M , of M ∈ D❂ (R) is defined to be the infimum of the set of integers
n such that there exists a complex F ∈ C❂ (R) consisting of Gorenstein flat modules
satisfying M ≃ F and Gn = 0 for n < ℓ. Let M be an R-complex in D(R). The right
derived section functor of the complex M is defined as RΓa (M ) = Γa (E), where E is
a semiinjective resolution of M (see [8] and [16]). If x = x1 , · · · , xr is a generating set
for the ideal a and Čx the corresponding Čech complex, then RΓa (M ) ≃ M ⊗L
R Čx (see
[15, Theorem 1.1(iv)]).
It has been shown in [7, Theorem 6.5] that the right derived section functor (with
support in any ideal a) sends complexes of finite flat dimension (respectively, finite
injective dimension) to complexes of finite flat dimension (respectively, finite injective
dimension). In section 2 we prove that if (R, m) is a local ring and M ∈ Df❁ (R), then
idR RΓa (M ) = idR M (see Theorem 2.2). It shows that the following statements are
equivalent:
(i) R is Gorenstein;
(ii) idR RΓa (R) = dim(R) for any ideal a of R;
(iii) idR RΓa (R) < ∞ for some ideal a of R.
This provides a characterization of Gorenstein rings, which recovers [20, Corollary
2.7]. Next, in 2.4, we prove that if (R, m) is a local ring and M is an R-complex in
−infRΓa (M )
Df❁ (R) with ampRΓa (M ) = 0, then idR Ha
(M ) = idR M + infRΓa (M ). Notice
that, this result is a generalization of [20, Theorem 2.5]. Also, a flat version of 2.2 is
demonstrated. Indeed, it is shown, in Theorem 2.7, that if (R, m) is a local ring and
M ∈ Df❂ (R), then fdR RΓa (M ) = fdR M .
It has been proved in [4, Theorem 5.9] that if M ∈ D (R), then
GfdR M < ∞ ⇒ GfdR RΓa (M ) < ∞.
Moreover, if R has a dualizing complex, the above implication may be reversed if a is
in the Jacobson radical of R and M ∈ Df (R). We show that if (R, m) is a local ring
and M ∈ Df (R), then
GfdR M < ∞ ⇒ GfdR RΓa (M ) = GfdR M, and
R admits a dualizing complex ⇒ GfdR RΓa (M ) = GfdR M.
Then, as a corollary, we prove that if M is a relative Cohen-Macaulay R-module with
respect to a, where is defined as in [20], and that n = grade(a, M ), then GfdR Hna (M ) =
GfdR M + n.
In section 3, first we prove that if C is an Γa -acyclic R-complex in C❁ (R), then
RΓa (C) ≃ Γa (C) (see Theorem 3.5). It implies that the right derived section functor of
a complex M ∈ D❁ (R) can be computed via a genuine left-bounded complex G ≃ M
of Gorenstein injective modules.
Also, as a main result, we show that if (R, m) is a local ring admitting a dualizing
complex and M is an R-complex in Df (R), then GidR RΓa (M ) = GidR M (see Theorem
3.9). It shows that the following statements are equivalent:
(i) R is Gorenstein;
(GORENSTEIN) INJECTIVE AND (GORENSTEIN) FLAT DIMENSIONS
3
(ii) GidR RΓa (R) = dim(R) for any ideal a of R;
(iii) GidR RΓa (R) < ∞ for some ideal a of R.
This provides a characterization of Gorenstein rings, which improves [20, Corollary 3.10]
and [19, Theorem 2.6], that is, we may prove them without assuming that R is CohenMacaulay. Next, in Theorem 3.10, we prove a complex version of 2.2, which improves
[20, Theorem 3.8]. As a corollary, in 3.11, we deduce that GidR Hnm (M ) = GidR M − n,
wherever (R, m) is a local ring and M is a Cohen-Macaulay R-module with dimR M = n.
2. right derived section functor, injective dimension and (Gorenstein)
flat dimension
The following lemma, which is an immediate consequence of [10, Corollary 3.4.4] and
[10, Proposition 3.2.2], determines the i -th Bass number µiRp ((RΓa (M ))p ) of (RΓa (M ))p
(see [2, Definition 6.1.18]).
Lemma 2.1. Let (R, m, k) be a local ring, and let M be an R-complex in D(R). Then
µiR (RΓa (M )) = µiR (M ) for all i ∈ Z; In particular, for every p ∈ V(a) there is an
equality µiRp ((RΓa (M ))p ) = µiRp (Mp ) for all i ∈ Z.
The following theorem, which is one of the main results of this section, provides a
comparison between the injective dimensions of a complex and its right derived section
functor.
Theorem 2.2. Let (R, m, k) be a local ring, and let M be an R-complex in Df❁ (R).
Then idR RΓa (M ) = idR M .
Proof. Let s := idR M < ∞. Then, in view of [2, Lemma 6.1.19] and Lemma 2.1,
µi+s
Rp ((RΓa (M ))p ) = 0 for all p ∈ Spec(R) and for all i > 0. Therefore, it follows from
[2, Lemma 6.1.19] that idR RΓa (M ) ≤ s.
For the opposite inequality, let t := idR RΓa (M ) < ∞. Then, by [2, Theorem
5.1.6], infRHomR (T, RΓa (M )) ≥ −t for all cyclic R-modules T . Hence, in view of [10,
Proposition 3.2.2], there are isomorphisms
H−t−i (RHomR (k, M )) ∼
= H−t−i (RHomR (k, RΓa (M ))) ∼
=0
for all i > 0. Therefore −infRHomR (k, M ) ≤ t, and so idR M ≤ t by [2, Theorem
6.1.13].
The following corollary, which recovers [20, Corollary 2.7], is an immediate consequence of the previous Theorem.
Corollary 2.3. Let (R, m) be a local ring. Then the following statements are equivalent:
(i) R is Gorenstein;
(ii) idR RΓa (R) = dim(R) for any ideal a of R;
(iii) idR RΓa (R) < ∞ for some ideal a of R.
The following theorem, which is an immediate consequence of Theorem 2.2, is a
generalization of [20, Theorem 2.5].
Theorem 2.4. Let (R, m) be a local ring. Suppose that M is an R-complex in Df (R)
such that ampRΓa (M ) = 0. Then
−infRΓa (M )
idR Ha
(M ) = idR M + infRΓa (M ).
4
C. JALALI
Proof. Let n := −infRΓa (M ). Since RΓa (M ) ≃ H−n (RΓa (M )), there is an equality
idR RΓa (M ) = idR Σn H−n (RΓa (M )) + n.
But Σn RΓa (M ) is equivalent to the module Σn H−n (RΓa (M )) in the category of Rmodules. So, we may identify Σn H−n (RΓa (M )) with Hna (M ). Hence idR Hna (M ) =
idR RΓa (M ) − n. The desired equality now follows from Theorem 2.2.
In the following we use the notion of a semifree resolution. A semifree resolution of an
R-complex M is a semifree complex F (see [2, Definition 3.1.1]) and a quasiisomorphism
≃
F −→ M .
Lemma 2.5. Let (R, m, k) be a local ring, and let M be an R-complex in D(R). Then
L
k ⊗L
R RΓa (M ) ≃ k ⊗R M.
Proof. Let F be a semifree resolution of the residue field k, and let x = x1 , · · · , xr be
a generating set for the ideal a and Čx be the Čech complex with respect to x. But,
as in the proof of [11, Lemma 2.4], there exists a quasiisomorphism F ⊗R Čx ≃ F in
C(R). The result now follows, since k ⊗L
R Čx ≃ k.
The following lemma, which is an immediate consequence of [10, Corollary 3.4.4] and
R
2.5, determines the i -th Betti number βi p ((RΓa (M ))p ) of (RΓa (M ))p (see [2, Definition
6.1.14]).
Lemma 2.6. Let (R, m, k) be a local ring, and let M be an R-complex in D(R). Then
βiR (RΓa (M )) = βiR (M ) for all i ∈ Z; In particular, for every p ∈ V(a) there is an
R
R
equality βi p ((RΓa (M ))p ) = βi p (Mp ) for all i ∈ Z.
Theorem 2.7. Let (R, m, k) be a local ring, and let a be a proper ideal of R. Suppose
that M is an R-complex in Df❂ (R). Then fdR RΓa (M ) = fdR M .
Proof. Let s := fdR M < ∞. Then, in view of [2, Lemma 6.1.15] and Lemma 2.6,
Rp
βi+s
((RΓa (M ))p ) = 0 for all p ∈ Spec(R) and for all i > 0. Therefore, it follows from
[2, Lemma 6.1.15] that fdR RΓa (M ) ≤ s.
For the opposite inequality, let t := fdR RΓa (M ) < ∞. Then, by [2, Theorem 5.1.9],
supT ⊗L
R RΓa (M )) ≥ t for all cyclic R-modules T . Hence, in view of Lemma 2.5, there
are isomorphisms
Ht+i (k ⊗L M ) ∼
= Ht+i (k ⊗L RΓa (M )) ∼
=0
for all i > 0. Therefore supk
R
L
⊗R M
R
≤ t, and so fdR M ≤ t by [2, Theorem 5.2.13].
The following theorem, which is an immediate consequence of Theorem 2.7, is a flat
version of 2.4.
Theorem 2.8. Let (R, m) be a local ring. Suppose that M is an R-complex in Df (R)
such that ampRΓa (M ) = 0. Then
−infRΓa (M )
fdR Ha
(M ) = fdR M − infRΓa (M ).
Proof. Straightforward verification similar to the proof of Theorem 2.4.
In the rest of this section, we make a comparison between the Gorenstein flat dimensions of a complex and its right derived section functor.
Proposition 2.9. Suppose that M is an R-complex in D (R). Then
GfdR RΓa (M ) ≤ GfdR M.
(GORENSTEIN) INJECTIVE AND (GORENSTEIN) FLAT DIMENSIONS
5
Proof. Notice that if GfdR M = ∞, then there is nothing to prove. So, we may assume
that GfdR M < ∞. Hence, it follows from [4, Theorem 5.9] that GfdR RΓa (M ) < ∞.
Now, by [9, Theorem 8.8], there exists p ∈ V(a) such that
GfdR RΓa (M ) = depthRp − depthRp (RΓa (M ))p .
But depthRp (RΓa (M ))p = depthRp Mp . It follows, again by [9, Theorem 8.8], that
GfdR RΓa (M ) = depthRp − depthRp Mp ≤ GfdR M
as desired.
Proposition 2.10. Let (R, m, k) be a local ring, and let M be an R-complex in Df (R)
such that GfdR M < ∞. Then
GfdR M ≤ GfdR RΓa (M ).
Proof. By [4, Theorem 5.9], GfdR RΓa (M ) < ∞. Hence, by [9, Theorem 8.7], there are
equalities
sup(E(k) ⊗L
R RΓa (M )) = depthR − depthR RΓa (M )
= depthR − depthR M = sup(E(k) ⊗L
R M ).
Since M ∈ Df (R), sup(E(k) ⊗L
R M ) = GfdR M . The result now follows from the fact
that GfdR RΓa (M ) ≥ sup(E(k) ⊗L
R RΓa (M )) (see [4, Corollary 3.6]).
The following theorem is a Gorenstein flat version of Theorem 2.7.
Theorem 2.11. Let (R, m) be a local ring, and let M be an R-complex in Df (R).
(i) If GfdR M < ∞, then GfdR RΓa (M ) = GfdR M .
(ii) If R admits a dualizing complex, then GfdR RΓa (M ) = GfdR M .
Proof. (i) A straightforward application of Proposition 2.9 and Proposition 2.10.
(ii) In view of part (i), we may assume that GfdR RΓa (M ) < ∞. Hence, by [4,
Theorem 5.9], GfdR M < ∞. The desired equality now follows from Proposition 2.9
and Proposition 2.10.
Corollary 2.12. Let (R, m) be a local ring. Suppose that M is relative Cohen-Macaulay
with respect to a and that n = grade(a, M ). Then GfdR Hna (M ) = GfdR M + n.
b m)
b is a
b is a local ring admitting a dualizing complex and M ⊗R R
Proof. Notice that (R,
b
b and that grade(aR,
b M ⊗R R)
b =
relative Cohen-Macaulay R-module
with respect to aR
n. Hence, in view of [3, Theorem 4.27], we may assume that R is complete; and so it
has a dualizing complex. The result therefore follows from Theorem 2.11.
3. right derived section functor and Gorenstein injective dimension
In this section the category of R-modules is denoted C(R). Recall from [1, Exercise
4.1.2] that the local cohomology modules of R-module M with respect to a can be
calculated by an Γa -acyclic resolution of M . First, we prove the complex version of it.
Definition 3.1. (see [17, 5.7.9]) Let F : C(R) → C(R) be a left exact functor, and
assume that M is an R-complex in C❁ (R). If 0 → M → C∗,0 → C∗,1 → · · · → C∗,q →
is a Cartan-Eilenberg injective resolution of M , where is defined as in [13, §10.5], define
Ri (FM ) to be Hi (Tot(FC)).
6
C. JALALI
Lemma 3.2. Let M and Ḿ be two R-complexes in C❁ (R), and let ζ : M → Ḿ be a
morphism of R-complexes. Suppose that 0 → M → C∗,0 → C∗,1 → · · · → C∗,q → and
0 → Ḿ → Ć∗,0 → Ć∗,1 → · · · → Ć∗,q → are Cartan-Eilenberg injective resolutions of
M and Ḿ , respectively. Then there exists a sequence {ζ∗,q }q∈N0 of morphisms ζ∗,q :
C∗,q → Ć∗,q of R-complexes over ζ.
Proof. A straightforward application of [12, Theorem 19].
Lemma 3.3. Let F : C(R) → C(R) be a left exact functor, and let M and Ḿ be two
R-complexes. Then
(i) Any quasiisomorphism ζ : M → Ḿ induces isomorphism
Ri (FM ) ∼
= Ri (FḾ )
for all i ∈ Z; and
(ii) If M is F-acyclic R-complex in C❁ (R), that is, Mi is F-acyclic for all i ∈ Z, then
Ri (FM ) = Hi (FM )
for all i ∈ Z.
Proof. Straightforward verification similar to the proof of [17, Corollary 5.7.7].
The following theorem, which is one of the main results of this section, enables us to
prove some interesting results.
Theorem 3.4. Let F : C(R) → C(R) be a left exact functor, and let M be an F-acyclic
R-complex in C❁ (R). Assume that I is F-acyclic and F(I) is injective for every injective
R-module I. Then F(M ) ≃ F(E), for every semiinjective resolution E ∈ C❁ (R) of M .
Proof. Let E be a semiinjective resolution of M with Ev = 0 for v > supM , and let
≃
ζ : M −→ E be an quasiisomorphism. By [13, Theorem 10.45], there exist CartanEilenberg injective resolutions 0 → M → C∗,0 → C∗,1 → · · · → C∗,q → and 0 →
E → Ć∗,0 → Ć∗,1 → · · · → Ć∗,q →. Hence, in view of Lemma 3.2, there is a sequence
{ζ∗,q }q∈N0 of morphisms of R-complexes such that the diagram
0 −−−−→ E −−−−→ Ć∗,0 −−−−→ Ć∗,1 −−−−→ · · · −−−−→ Ć∗,q −−−−→
x
x
x
x
ζ
ζ
ζ
ζ
∗,q
∗,1
∗,0
0 −−−−→ M −−−−→ C∗,0 −−−−→ C∗,1 −−−−→ · · · −−−−→ C∗,q −−−−→
commutes in C(R). By Lemma 3.3(ii), Rp (F(M )) = Hp (F(M )) for all p ∈ Z, so
that the natural morphism F(M ) → Tot(F(C)) is a quasiisomorphism. Similarly,
Rp (F(E)) = Hp (F(E)) for all p ∈ Z, so that the natural morphism F(E) → Tot(F(Ć))
is a quasiisomorphism. Thus, by [2, Proposition 3.3.5(a)], there exists a quasiisomorphism Tot(F(Ć)) → F(E), since F(E) is injective.
But, by Lemma 3.3(i), there are isomorphisms Rp (F(M )) ∼
= Rp (F(E)) for all p ∈ Z.
Hence the morphism ζ∗ : Tot(F(C)) −→ Tot(F(Ć)) is a quasiisomorphism, where
X
ζn =
F(ζp,q ) : Tot(F(C))n → Tot(F(Ć))n
p+q=n
for all n ∈ Z. Therefore, there are quasiisomorphisms
≃
≃
≃
F(M ) −→ Tot(F(C)) −→ Tot(F(Ć)) −→ F(E).
Now F(M ) ≃ F(E) as desired.
(GORENSTEIN) INJECTIVE AND (GORENSTEIN) FLAT DIMENSIONS
7
The next theorem, which offers an application of the previous theorem, is a complex
version of [1, Exercise 4.1.2].
Theorem 3.5. Let C be an Γa -acyclic R-complex in C❁ (R). Then RΓa (C) ≃ Γa (C).
The next corollary shows that if M ∈ D❁ (R), then RΓa (M ) can be computed via
a genuine left-bounded complex G ≃ M of Gorenstein injective modules. Also, notice
that [14, Theorem 3.4] is an immediate consequence of this fact.
Corollary 3.6. Let M be an R-complex in D❁ (R), and let G ∈ C❁ (R) be an R-complex
of Gorenstein injective modules such that M ≃ G. Then RΓa (M ) ≃ Γa (G).
Proof. Since by [18, Lemma 1.1] every Gorenstein injective module is Γa -acyclic, the
result follows from Theorem 3.5.
It has been proved in [14, Corollary 3.3] that if R admits a dualizing complex and
M ∈ D❁ (R), then GidR RΓa (M ) ≤ GidR M . The following proposition together with
[4, Theorem 5.9] recover this result.
Proposition 3.7. Suppose that M is an R-complex in D❁ (R) such that GidR RΓa (M ) <
∞. Then
GidR RΓa (M ) ≤ GidR M.
Proof. Notice that if GidR M = ∞, then there is nothing to prove. So, we may assume
that GidR M < ∞. By [5, Theorem 2.2], there exists p ∈ V(a) such that
GidR RΓa (M ) = depthRp − widthRp (RΓa (M ))p .
But widthRp (RΓa (M ))p = widthRp Mp . It follows, again by [5, Theorem 2.2], that
GidR RΓa (M ) = depthRp − widthRp Mp ≤ GidR M
as desired.
Proposition 3.8. Let (R, m) be a local ring, and let M be an R-complex in
such that GidR M < ∞. Then
Df (R)
GidR M ≤ GidR RΓa (M ).
Proof. By [4, Proposition 6.3], there is an inequality
GidR RΓa (M ) ≥ depthR − widthR RΓa (M ).
But widthR RΓa (M ) = widthR M . Thus, we have
GidR RΓa (M ) ≥ depthR − widthR M
= depthR − inf M.
The result therefore follows from [5, Corollary 2.3].
The following theorem, which is a Gorenstein injective version of Theorem 2.2, is
one of the main results of this section.
Theorem 3.9. Let (R, m) be a local ring admitting a dualizing complex, and let M be
an R-complex in Df (R). Then GidR RΓa (M ) = GidR M .
Proof. A straightforward application of [4, Theorem 5.9], Proposition 3.7 and Proposi
tion 3.8.
The next theorem, which is a Gorenstein injective version of Theorem 2.4, recovers
[20, Theorem 3.8].
8
C. JALALI
Theorem 3.10. Let (R, m) be a local ring admitting a dualizing complex. Suppose that
M is an R-complex in Df (R) such that ampRΓa (M ) = 0. Then
−infRΓa (M )
GidR Ha
(M ) = GidR M + infRΓa (M ).
Proof. Follows from Theorem 3.9 (similar to the proof of Theorem 2.4).
The following corollary, which improves [20, Corollary 3.9], is a consequence of the
previous Theorem.
Corollary 3.11. Let (R, m) be a local ring, and let M be a Cohen-Macaulay R-module
with dimR M = n. Then the following statements hold.
b = GidR M .
(i) GidRb RΓaRb (M ⊗R R)
n
(ii) GidR Hm (M ) = GidR M − n.
b m)
b is
b is a local ring admitting a dualizing complex and M ⊗R R
Proof. Notice that (R,
b
a Cohen-Macaulay R-module of dimension n.
(i) A straightforward application of Theorem 3.9 and [3, Theorem 3.24].
(ii) There is an inequality
b + n.
b = Gid b Hn (M ⊗R R)
GidRb RΓmb (M ⊗R R)
b
R m
b = GidR Hn (M ). The result now
But, in view of [14, Lemma 3.6], GidRb Hnmb (M ⊗R R)
m
follows from part (i).
The next corollary provides a characterization of Gorenstein rings, which together
with Corollary 2.3 show that [19, Theorem 2.6] and [20, Corollary 3.10] hold without
assuming that R is Cohen-Macaulay.
Corollary 3.12. Let (R, m) be a local ring admitting a dualizing complex. Then the
following statements are equivalent:
(i) R is Gorenstein;
(ii) GidR RΓa (R) = dim(R) for any ideal a of R;
(iii) GidR RΓa (R) < ∞ for some ideal a of R.
Proof. A straightforward application of Theorem 3.9 and [3, Proposition 3.11].
References
[1] M. P. Brodmann and R. Y. Sharp, Local Cohomology: An Algebraic Introduction with Geometric
Applications, (Cambridge University Press, Cambridge, 1998).
[2] L.W. Christensen and H-B. Foxby, Hyperhomological Algebra with Applications to Commutative
Rings 12 Dec 2006, (2006).
[3] L.W. Christensen, H-B. Foxby and H. Holm, Beyond totally reflexive modules and back, In:
Noetherian and Non-Noetherian Perspectives, edited by M. Fontana, S-E. Kabbaj, B. Olberding
and I. Swanson, (Springer Science+Business Media, LLC, New York, 2011) 101-143.
[4] L.W. Christensen, A. Frankild and H. Holm, On Gorenstein projective, injective and flat
dimensions- A functorial descripotion with aplications, J. Algebra, 302 (2006) 231–279.
[5] L. W. Christensen and S. Sather-Wagstaff, Transfer of Gorenstein dimensions along ring homomorphisms, J. Pure Appl. Algebra, 214 (2010), no. 6, 982-989.
[6] E.E. Enochs and O.M.G. Jenda, Relative Homological Algebra, De Gruyter Expositions in Mathematics, No. 30 (Walter De Gruyter, New York, 2000).
[7] H-B. Foxby, Bounded complexes of flat modules, J. Pure Appl. Algebra 15 (1979), no. 2, 149-172.
[8] S. I. Gelfand and Yu. I. Manin, Homological Algebra, (Springer-Verlag, Berlin, 1999).
[9] S. Iyengar and S. Sather-Wagstaff, G-dimension over local homomorphisms. Applications to the
Frobenius endomorphism, Illinois J. Math., 48(1) (2004) 241–272. MR2048224.
(GORENSTEIN) INJECTIVE AND (GORENSTEIN) FLAT DIMENSIONS
9
[10] J. Lipman, Lectures on local cohomology and duality, in: Local cohomology and its applications,
volume 226 of Lecture Notes in Pure and Appl. Math., (Dekker, New York, 2002) pages 39–89.
[11] W. Mahmood and P. Schenzel, On invariants and endomorphism rings of certain local cohomology
modules, J. Algebra 372 (2012) 56-67.
[12] D. G. Northcott, An Introduction to Homological Algebra, (Cambridge University Press, 1960).
[13] J.J. Rotman, An Introduction to Homological Algebra, (Springer Science+Business Media LLC,
2009).
[14] R. Sazeedeh, Gorenstein injective of the section functor, Forum Mathematicum, 22 (2010) 11171127.
[15] P. Schenzel, Proregular sequences, local cohomology, and completions, Math. Scand., 92 (2003)
271–289.
[16] N. Spaltenstein, Resolutions of unbounded complexes, Compositio Math., 65 (1988).
[17] C. A. Weibel, An Introduction to Homological Algebra, (Cambridge University Press, New York,
1994).
[18] S. Yassemi, A generalization of a theorem of Bass, Comm. Algebra, 35 (2007) 249-251.
[19] T. Yoshizawa, On Gorenstein injective of top local cohomology modules, Proc. Amer. Math. Soc.,
140 (2012) 1897-1907.
[20] M. R. Zargar and H. Zakeri, On injective and Gorenstein injective dimensions of local cohomology
modules, to appear in Algebra Colloquium.
C. Jalali, Faculty of mathematical sciences and computer, Kharazmi University, 599
Taleghani Avenue, Tehran 15618, Iran
E-mail address: [email protected]
| 0math.AC
|
Flexible Design for α-Duplex Communications
in Multi-Tier Cellular Networks
arXiv:1511.07903v3 [cs.IT] 20 Apr 2016
Ahmad AlAmmouri, Hesham ElSawy, and Mohamed-Slim Alouini
Abstract
Backward compatibility is an essential ingredient for the success of new technologies. In the context
of in-band full-duplex (FD) communication, FD base stations (BSs) should support half-duplex (HD)
users’ equipment (UEs) without sacrificing the foreseen FD gains. This paper presents flexible and
tractable modeling framework for multi-tier cellular networks with FD BSs and FD/HD UEs. The
presented model is based on stochastic geometry and accounts for the intrinsic vulnerability of uplink
transmissions. The results show that FD UEs are not necessarily required to harvest rate gains from FD
BSs. In particular, the results show that adding FD UEs to FD BSs offers a maximum of 5% rate gain
over FD BSs and HD UEs case if multi-user diversity is exploited, which is a marginal gain compared
to the burden required to implement FD transceivers at the UEs’ side. To this end, we shed light on
practical scenarios where HD UEs operation with FD BSs outperforms the operation when both the
BSs and UEs are FD and we find a closed form expression for the critical value of the self-interference
attenuation power required for the FD UEs to outperform HD UEs.
Index Terms
Full duplex, half duplex, stochastic geometry, network interference, network rate, network topology.
I. I NTRODUCTION
Time division duplexing (TDD) and frequency division duplexing (FDD) are the commonly
used techniques to protect receivers from their overwhelming self-interference (SI). This implies
The authors are with Computer, Electrical, and Mathematical Sciences and Engineering (CEMSE) Divison, King Abdullah
University of Science and Technology (KAUST), Thuwal, Makkah Province, Saudi Arabia. (Email: {ahmad.alammouri,
hesham.elsawy, slim.alouini}@kaust.edu.sa)
Part of this work will be presented in IEEE International Conference on Communications (ICC’16) [1].
that the resources (i.e., time or frequency) are divided between forward and reverse links, which
creates a performance trade-off between them. SI cancellation (SIC) eliminates such trade-off
via in-band full-duplex (FD) communication, which gives the forward and reverse links the
opportunity to simultaneously utilize the complete set of resources [2]–[6]. FD transceivers
are capable of sufficiently attenuating (up to -110 dB [7]) their own interference (i.e., SI)
and simultaneously transmit and receive on the same channel, which offers higher bandwidth
(BW) for FDD systems and longer transmission time for TDD systems. Consequently, FD
communication improves the performance of both the forward and reverse links, in which the
improvement depends on the efficiency of SIC.
Leveraging FD communication to large-scale networks, SI is not the only bottleneck due to the
increased mutual interference when compared to the half-duplex (HD) case. This is because each
FD link contains two active transmitters while each HD link contains one active transmitter and
one passive receiver. Therefore, rigorous studies that capture the effect of the network interference
on FD communication are required to draw legitimate conclusions about its operation in largescale setup. In this context, stochastic geometry can be used to model FD operation in large scale
networks and understand its behavior [8]. Stochastic geometry succeeded to provide a systematic
mathematical framework for modeling both ad-hoc and cellular networks [8]–[11].
Despite the higher interference injected into the network, recent studies have shown that FD
communications outperform HD communications in large scale setup if sufficient SIC is achieved.
For instance, the asymptotic study in [12] shows a maximum improvement of 80% rate gain,
which monotonically decreases in the link distance, for FD communication over the HD case.
A more realistic ad-hoc network setup in [13] shows that FD offers an average of 33% rate
gain when compared to the HD operation. In the case of cellular networks, [14] shows around
30% improvement in the total rate for FD when compared to the HD case. The authors in [15]
show that the increase of aggregate interference in FD networks creates a trade-off between the
average spectral efficiency and the coverage probability. However, [7] reveals that the FD gains
in cellular networks are mainly confined to the DL due to the high disparity between uplink
(UL) and downlink (DL) transmission powers. Furthermore, the authors in [16]–[19] show that
when a constrained power control is employed in the UL, the FD communication gains in the
DL may come at the expense of high degradation in the UL. The authors in [16] advise to use
FD communications in small cell tiers such that the users’ equipment (UEs) and base stations
(BSs) have comparable transmit powers. For FD operation in macro tiers with high disparity
between UL and DL transmit powers, the authors in [17] advocate using pulse shaping along
with partial overlap between UL and DL spectrum to neutralize DL to UL interference and avoid
deteriorating UL rate. With pulse shaping and partial UL/DL overlap, [17] shows a simultaneous
improvement of 33% and 28% in the UL and DL, respectively. It ought to be mentioned that,
in addition to the UL/DL transmit power disparity, the asymmetric UL/DL traffic that naturally
exists in practical cellular networks imposes another challenge to the FD operation [20].
To harvest the aforementioned gains, FD transceivers are required on both sides of each link.
However, cellular networks operators can only upgrade their BSs and do not have direct access
to upgrade UEs. Furthermore, the high cost of FD transceivers, in terms of complexity, power
consumption and price, may impedes their penetration to the UEs’ domain. Therefore, techniques
to achieve FD gains in cellular networks with FD BSs and HD UEs are required. In this context,
3-nodes topology (3NT) is proposed in [15], [21]–[25] to harvest FD gains by serving two HD
UEs within each FD BS. In 3NT, the BSs have SIC capabilities and can simultaneously serve HD
UL and HD DL users on the same channels. That is, each BS can merge each UL/DL channel
pair into a larger channel and reuse that channel to serve an UL and a DL users simultaneously.
The studies in [21]–[24] show the potential of 3NT to harvest HD gains. However, the results
in [21] are based on simulations, and the results in [22]–[25] are based on a simplistic system
models.
In this paper, we present a unified mathematical framework, based on stochastic geometry,
to model 3NT (i.e., FD BSs and HD users) and 2-nodes topology (2NT) (i.e., FD BSs and
FD UEs) in multi-tier cellular networks. The proposed mathematical framework is then used to
conduct rigorous comparison between 3NT and 2NT. Different from [23]–[25], the presented
system model accounts for the explicit performance of UL and DL for cell center users (CCUs)
and cell edge users (CEUs) in a multi-tier cellular network. It also captures more realistic system
parameters than [23]–[25] by accounting for pulse-shaping, matched filtering, UL power control,
maximum power constraint for UEs, UEs scheduling, and the different BSs’ characteristics
in each network tier. When compared to [17], the proposed framework considers a multi-tier
network with different FD topologies (i.e., 2NT and 3NT), flexible association, different pathloss exponents between different network elements, and incorporate uncertainties in the SIC.
However, we exploit the fine-grained duplexing strategy proposed in [17] that allows partial
overlap between the UL and DL channels, which is denoted as α-duplex (αD) scheme. The
parameter α ∈ [0, 1] controls the amount of overlap between UL and DL channels and captures
the HD (at α = 0) and FD (at α = 1) as special cases. Beside being used to optimize the spectrum
allocation to the UL and DL, the parameter α shows the gradual effect of the interference induced
via FD communication on the system performance, and to optimize the amount of the overlap
between UL and DL channels. The results show that 3NT can achieve close performance (within
5%) when compared to the 2NT with FD UEs that have efficient SIC if multi-user diversity and
UEs scheduling are exploited. On the other hand, if the FD UEs in the 2NT have poor SIC, the
3NT achieves a better performance. In both cases, it is evident that network operators do not
need to bear the burden of implementing SIC in the UEs to harvest FD gains.
The rest of the paper is organized as follows: in Section II, we present the system model and
methodology of the analysis. In Section III, we analyze the performance of the α-duplex system.
Numerical and simulation results with discussion are presented in Section IV before presenting
the conclusion in Section V.
Notations: E[.] denotes the expectation over all the random variables (RVs) inside [.], Ex [.]
denotes the expectation with respect to (w.r.t.) the RV x, 1{.} denotes the indicator function which
takes the value 1 if the statement {.} is true and 0 otherwise, .∗ denotes the convolution operator
and S ∗ denotes the complex conjugate of S, Lx (.) denotes the Laplace transform (LT) of the
probability density function (PDF) of RV x and Italic letters are used to distinguish variables
from constants.
II. S YSTEM M ODEL
A. Network Model
A K-tier cellular network is considered, in which the BSs1 in each tier are modeled via an
(k)
independent homogeneous 2-D Poisson point processes (PPPs) [9] Φd , where k ∈ {1, , 2, ..., K},
with intensity λk . The location of the ith BS in the k th tier is denoted by xk,i ∈ R2 . Beside
simplifying the analysis, the PPP assumption for abstracting cellular BSs is verified by several
experimental studies [9], [10], [28]. UEs are distributed according to a PPP Φu , which is
1
In this work we assume that both the BSs and the UEs are equipped with a single antenna. Combining FD with multiple-input
and multiple-output (MIMO) transmitters is covered in [26], [27].
independent from the BSs locations, with intensity λu , where λu ≫
(k)
(k)
BSs transmit with a constant power Pd , however, the value of Pd
K
P
λk . Within each tier, all
k=1
varies across different tiers.
In contrast, UEs employ a truncated channel inversion power control with maximum transmit
power constraint of Pu [29]. That is, each UE compensates for its path-loss to maintain a tierspecific target average power level of ρ(k) at the serving BS. UEs that cannot maintain the
threshold ρ(k) transmit with their maximum power Pu . UEs who can keep the threshold ρ(k) , are
denoted as cell center users (CCUs), while UEs who transmit with their maximum power are
denoted as cell edge users (CEUs) [30].
The power of all transmitted signals experiences a power law path loss attenuation with
exponent η > 2. Due to the different relative antenna heights and propagation environments,
we discriminate between the path loss exponent for the paths between two BSs (DL to UL
interference), two UEs (UL to DL interference), and a BS and a UE (UL to UL interference),
which are respectively denoted by ηdu , ηud , and ηuu , as shown in Fig. 1. Assuming channel
reciprocity, the path loss exponent between a BS and a UE (i.e., DL to DL interference), denoted
by ηdd , is equivalent to the one between a UE and a BS (i.e., UL to UL interference) ηuu , and
hence, both symbols are used interchangeably2 . Also, Rayleigh fading channels are assumed
such that the channels power gains are independent and identically distributed (i.i.d) exponential
RVs with unit means3 .
B. Operation Modes and Spectrum Allocation
We consider a fine grained αD scheme that allows partial overlap between UL and DL channels
and captures the FD and HD as special cases. We denote the BWs used in the HD case in the
UL and DL, respectively, as BuHD and BdHD , in which BuHD and BdHD are not necessarily equal.
To avoid adjacent channel interference, the BSs utilize a guard band of ǫB between each UL-DL
pair of bands, where B = min(BdHD , BuHD )4 . As shown in Fig. 2a, the BW used in the αD DL
is Bd (α) = BdHD + α(ǫ + 1)B, and in the αD UL is Bu (α) = BuHD + α(ǫ + 1)B. Note that the
2
We assume that the path-loss exponents in each direction is different but equivalent in all tiers. Assuming equal path-loss
exponents in all tiers is a common simplifying assumption in the literature. [29], [31], [32].
3
4
Extending the results to capture other fading models can be done following [33], [34].
The scheme proposed in [17] is captured by setting ǫ to zero, since no guard bands are assumed there.
Fig. 1: Channel allocation, interference types, and path-loss exponents for a) 2NT and b) 3NT.
(a) Frequency bands allocation.
(b) UEs’ scheduling in 3NT.
Fig. 2: Frequncy Bands allocation and UEs’ scheduling.
parameter α controls the partial overlap between the UL and DL frequency bands. Also, the HD
and FD modes are captured as special cases by setting α to 0 and 1, respectively. It is assumed
that each tier has its own duplexing parameter αk , which is used by all BSs within that tier.
Without loss of generality, we assume that each BS has only two pairs of UL-DL channels
that are universally reused across the network. For simplicity, we assume that the two channel
HD
HD
HD
HD
pairs are sufficiently separated in the frequency domain (i.e., fu1
< fd1
≪ fu2
< fd2
) to
avoid adjacent channel interference between different UL-DL pairs. It is worth noting that the
idealized rectangular frequency domain pulse shapes shown in Fig. 2a are used for illustration
only. However, as discussed later, we use time-limited pulse shapes that impose adjacent channel
interference due to the out of band ripples in the frequency domain.
In the 2NT network, UEs have FD transceivers and can use the UL and DL belonging to
the same UL-DL pair for their αD operation. In contrast, 3NT UEs have HD transceivers and
cannot transmit and receive on overlapping channels. Hence, each HD user is assigned his UL
and DL channels from two different UL-DL pairs as shown in Fig. 1 and Fig. 2a. Consequently,
3NT UEs can benefit from the larger BW channels without SI. Note that the FD BSs in all cases
as well as the FD UEs in the 2NT would experience SI as shown in Fig. 1. In contrast, 3NT
experience intra-cell interference on the DL direction due to the partial overlap between the UL
channel of the one UE and the DL channel of the other UE.
To this end, we assume that the BSs exploit multi-user diversity to control intra-cell interference in 3NT by imposing a minimum separation angle constraint between users scheduled on
the same channel as shown in Fig. 2b5 . In sectored BSs, the value of δo can be estimated to a
certain accuracy depending on the number of sectors. If the BSs’ cannot estimate the angeles
between users, then δo is set to zero and we refer to this case as random scheduling.
For the FD BSs and 2NT UEs, we denote the SI attenuation power as βu hs and βd hs ,
respectively, where βu , βd are positive constants representing the mean SIC power values in
the UL and DL, respectively, and hs follows a general unit mean distribution with PDF given
by fHs (·) which represents the uncertainty in SIC. Three special cases of interest for fHs (·) are
considered, namely, constant attenuation where fHs (·) is a degenerate distribution as in [14],
[17], [35] and random attenuation where fHs (·) is an exponential distribution as in [24] and
Rician fading as in [26] which captures the previous two cases as special cases. It is shown in
[1] that all distributions leads to the same performance trends.
C. UEs to BSs Association
We consider a biased and coupled6 BS-UE association scheme. Biasing is used to enable
flexible load balancing between tiers by encouraging UEs to connect to lower power BSs to
balance the average load served by the tiers across the network [37], [38]. We define a distance
dependent biasing factor τ and assume that all BSs within the same tier have the same biasing
factor. Hence, a UE connects to k th tier if {τk rk < τi ri ∀ i ∈ {1, .., K}, k 6= i}.
The used association scheme captures different association strategies as special cases. For
example, if τ is set to the same value for all tiers, then closest BS association is considered, if
5
More advanced and sophisticated scheduling and multi-user diversity techniques are postponed for future work.
6
Decoupled UL/DL association is analyzed using stochastic geometry in [36] for traditional HD multi-tier network, extending
this analysis to decoupled association is postponed to future work.
(a) τj = 1.
(j)
(b) τj = (Pd )
−1
η
.
Fig. 3: A realization of the associations areas assuming different association factors, where the
green squares, diamonds, and circles represent macro, micro, and pico BSs, respectively.
(k)
−1
τk = (Pd ) ηdd , then the UE connects to the BS providing the highest received signal strength
(RSS). Note that different association schemes changes the relative BSs’ association areas across
the tiers as shown in Fig. 3, where three tiers network is shown with 10W macro BSs, 5W micro
BSs, and 1W, pico BSs7 . In Fig. 3a, nearest BS association is considered, and hence, association
areas are represented by Voronoi tessellation [40]. In Fig. 3b, UE connects according to the
RSS, in this case the association areas construct multiplicative weighted Voronoi tessellation
(also denoted as circular tessellation) [40].
D. Pulse Shaping
FT
We employ time-limited pulse shapes8 , denoted as s(t, BW, bv ) ←→ S(f, BW, bv ) with unit
(k)
(k)
energy, where BW is the pulse null-to-null bandwidth, bd and bu indicate the pulse types used
by the k th tier in the DL and UL, respectively. We assume a flexible pulse shaping scheme,
where each tier has its own pulse shapes in the DL and UL, however, all BSs within the same
tier use the same pulse shapes. To have a unified effective BW for all values of αi in the αD
mode, the null-to-null BW of the pulse-shapes is kept equal to the channel BW. Hence, the pulse
shapes are also functions of the parameter αk .
7
The values of the transmit powers are based on [39].
8
In this work, we focus on time-limited pulse shapes to avoid inter-symbol-interference (ISI), since including the effect of ISI
will complicate the analysis much more. However, frequency-limited Nyquist pulses (e.g. root raised cosine) can be also used,
since it protect the nodes from ISI. For more information on the effect of different pulse shapes on the α-duplex scheme, refer
to [17].
E. Base-band Signal Representation
(k)
(k)
For the sake of simple presentation, we use αk , bd and bu to denote the duplexing factor,
the UL pulse shape, and the DL pulse shape, respectively in the k th tier. Also, we use v, v̄,
and w to indicate the desired transmission, where v, v̄, w ∈ {d, u}, v 6= v̄, for DL and UL,
respectively, and i, k as BSs’ tier index, where i, k ∈ {1, ..., K}. Exploiting this notation, the
received baseband signal at the input of the matched filter of a test transceiver in the ith tier
(BS or UE) can be expressed as
yv(i) (t) =Γo
q
K
K
X
X
X
X
(k)
(i)
(k)
(i)
(i)
(t) + n(t).
Iju →v (t) + Is(i)
Idj →v (t) +
Pvo ro−ηvv ho s(t, Bv (αi ), bv(i) ) +
v
k=1 j∈Ψ̃(k)
k=1 j∈Ψ̃(k)
u
d
(1)
(i)
where Γo , Pvo , ro , and ho denote the intended symbol, transmit power, link distance, and channel
(k)
(k)
(k)
(i)
power gains, respectively. Ψ̃d ⊆ Ψd is the set of interfering BSs in the k th tier, Ijd →v (t) is
(k)
(k)
the DL interference from the j th BS in k th tier, Ψ̃u ⊆ Ψu is the set of interfering UEs in the
(k)
(i)
(i)
k th tier, Iju →v (t) is the UL interference from j th UE connected to the k th tier, Isv (t) is the
SI term affecting the v direction, and n(t) is a white complex Gaussian noise with zero mean
and two-sided power spectral density No /2. The downlink and uplink interference are given by
(k)
(i)
Ijd →v (t)
Iju
(k)
→v (i)
=
r
(k)
(k)
Γdj s(t, Bd (αi ), bd )
r
(k)
(t) = Γ(k)
uj s(t, Bu (αi ), bu )
(k)
(k)
(k) (k)
Pd hdj
(k) (k)
Puj huj
−ηdv
(k)
(k)
exp j2π fd − fv(i) t ,
rdj
(2)
−ηuv
(k)
exp j2π fu(k) − fv(i) t .
ruj
(3)
where Γdj and Γuj denote the interfering symbol from the DL j th BS and interfering symbol
from the UL j th UE in the k th tier. Following the same interpretation of the subscripts and
(k)
(k)
superscripts defined for the interfering symbols, hdj and huj denote the DL and UL interfering
(k)
(k)
(k)
(k)
channel gains, Pd , and Puj denote the DL and UL interfering transmit powers, rdj , and ruj
(k)
(k)
denote the DL and UL interfering link distances, and fd and fu denote the center frequencies
of the DL and UL interfering frequency bands (see Fig. 1). Note that the BS index is removed
from the DL transmit power because all BSs in the same tier transmit with the same power.
(k)
(k)
Similarly, the BS and UE indices are removed from the center frequencies fd and fu
because
all elements in the same tier employ the same overlap parameter α(k) . The SI term in (1) is
given by
I(i)
su (t) = Γs
q
(i)
(i)
βu hs Pd s(t, Bd (αi ), bd ) exp j2π∆f (i) t .
(4)
(i)
Γ pβ h P s(t, B (α ), b(i)
t . 2NT
u ) exp −j2π∆f
u
i
s
d s uo
(i)
Isd (t) =
0.
3NT
(5)
(i)
where, βd represents the average attenuation power of the SI in the DL, βu is the average
attenuation power of the SI affecting a BS in the ith tier in the UL, hence, each tier can have a
different SIC capability depending on the BSs’ sizes and receivers complexity. Puo is the transmit
power of the tagged UE and
(i)
∆f (i) = fu(i) − fd ,
(6)
which represents the difference between the UL and DL center frequencies in the ith tier. Note
that this difference also depends on the chosen tier, since each tier can have a different duplexing
factor αi which leads to different UL/DL BWs and center frequencies.
F. Methodology of Analysis
The analysis is conducted on a test transceiver, which is a BS for the UL and a UE for the
DL, located at the origin and operating on a test channel pair. According to Slivnyak’s theorem
[9], there is no loss of generality in this assumption. Also, there is no loss of generality to focus
on a test channel pair as interferences on different bands are statistically equivalent. We asses the
impact of FD communication via the outage probability and the transmission rate. The outage
probability is defined as
O(θ) = P {SINR < θ} .
(7)
For the transmission rate [41], we assume that the nodes transmit with a fixed rate regardless
of the state of the channel (BW log2 (1 + θ)), hence, the transmission rate is defined as
R = BW log2 (1 + θ) P {SINR ≥ θ} .
(8)
In (8), the degraded SINR is compensated by the increased linear BW term. Hence, (8) can
be used to fairly assess the performance of FD operation. As shown (7) and (8), both the outage
probability and transmission rate are independent from the symbol structure and only depend on
(k)
(k)
the SINR. Consequently, all transmitted symbols Γo , Γdj and Γuj for all {k, j} are abstracted
to independent zero-mean unit-variance complex Gaussian random variables. Abstracting the
symbols via Gaussian random variables have negligible effect on the signal-to-interference-plusnoise-ratio (SINR) distribution as shown in [42], [43].
In the analysis, we start by modeling the effect of the matched and low-pass filtering on the
baseband signal. Then, based on the base-band signal format after filtering, the expressions for
the SINR in different cases (i.e., CCU-UL, CCU-DL, CEU-UL, and CEU-DL for 3NT, and 2NT)
are obtained. The performance metrics in (7) and (8) are then expressed in terms of the LT of
the PDF of the interference, which is obtained later to evaluate (7) and (8)9 .
III. P ERFORMANCE A NALYSIS
The received signal is first convolved with the conjugated time-reversed pulse shape template,
passed through a low-pass filter, and sampled at t = to . The baseband signal after filtering and
sampling at the input of the decoder is given by:
yv(i) (to ) =yv(i) (t). ∗ h(i)
v (t − t0 )|t=to
r
q
K
−ηvv
X
X
(k)
(i) −ηvv
(k) (k)
(k)
Iv (αi , αk )+
Γvj Pvj hvj rvj
ho Iv (αi , αi ) +
=A Pvo ro
k=1 j∈Ψ̃(k)
v
K
X
X
(k)
Γv̄j
k=1 j∈Ψ̃(k)
v̄
r
(k) (k)
Pv̄j hv̄j
(k)
rv̄j
−ηv̄v
(i)
Cv (αi , αk ) + I(i)
sv (t). ∗ hv (t − t0 )|t=to +
p
No |Iv (αi , αi )|2 .
(9)
(i)
where hv (t) is the combined matched and low-pass filter impulse response for a transceiver in
(i)
the ith tier. The frequency domain representation hv (t) is given by
S ∗ (f, B (α ), b(i)
v )
v
i
Hv(i) (f ) =
0
−
Bv (αi )
2
≤f ≤
Bv (αi )
.
2
(10)
elsewhere.
(i)
where S(f, Bv (αi ), bv ) represents the used pulse shape as discussed in section II-D.
The factors I(·, ·) and C(·, ·) in (9) represent the intra-mode (i.e., from UL-UL or DL-DL )
and cross-mode (i.e., from UL-DL or vice versa) effective received energy factors, respectively.
9
Expressions for the bit error probability can be derived by using the obtained SINR in the next section and following [17],
[44].
From (2), (3), (10), and expressing the convolution operation in the frequency domain, the pulse
shaping and filtering factors are obtained as,
Iv (αi , αk ) =
BvZ
(αi )/2
(k)
(i)
(k)
S ∗ (f, Bv (αi ), b(i)
v )S(f − fv + fv , Bv (αk ), bv )df,
BvZ
(αi )/2
S ∗ (f, Bv (αi ), b(i)
v )S(f − fv̄
(11)
−Bv (αi )/2
Cv (αi , αk ) =
(k)
(k)
+ fv(i) , Bv̄ (αk ), bv̄ )df,
(12)
−Bv (αi )/2
It should be noted that although same mode links use similar pulse shapes in the same tier,
the effective energy received from intra-mode intra-tier transmitters is not unity as shown in
(11). This is because (10) includes the combined impulse response of the matched and low-pass
filters, which extracts the desired frequency range from the received signal. Consequently, the
energy outside the desired BW is discarded and the energy contained within the pulse shape
is no longer unity. Also, the cross-mode interference factor in (12) is strictly less than unity
due to low-pass filtering, the possibly of different pulse shapes, and the partial overlap between
cross-mode channels.
o
n
(i)
(i)
(i)
(i)
Let Ξ = ro , rvj , ho , hvj , Pvo , Pvj , hs ; ∀i = {1, ..., K}, v ∈ {u, d} , then conditioning on Ξ
the SINR is given by
SINR(i)
v (Ξ) =
(i)
K
P
Pvo ro−ηvv ho
,
−ηv̄v
−ηvv
K
P
P
P
(k)
(k)
(k) (k)
(k) (k)
|C˜v (αi , αk )|2 + σ̃s2v (αi ) + No
|Ĩv (αi , αk )|2 +
Pvj hvj rvj
Pv̄j hv̄j rv̄j
k=1 j∈Ψ̃(k)
k=1 j∈Ψ̃(k)
v
v̄
(13)
where,
|Ĩv (αi , αk )|2 =
|C˜v (αi , αk )|2 =
|Iv (αi , αk )|2
,
|Iv (αi , αi )|2
|Cv (αi , αk )|2
,
|Iv (αi , αi )|2
(14)
(15)
and σ̃s2v (·) is the residual SI power normalized by |Iv (αi , αi )|2 . From (4) and (5), σ̃s2v can be
expressed for the UL and DL as
(i)
σ̃s2u (αi ) = βu(i) hs Pd |C˜u (αi , αi )|2 .
(16)
β h P |C˜ (α , α )|2 .
i
i
d s uo d
σ̃s2d (αi ) =
0.
2NT
(17)
3NT
The SINR in (13) is used in the next section to evaluate the outage provability and rate as
discussed in Section II-F.
A. Performance Metrics
From (7) and (13), the outage probability in the link v ∈ {u, d} in the ith tier can be written
as,
Ov(i) (θ) = P
(i)
Pvo ro−ηvv ho
<θ .
K
K
P
P
(k,i) ˜
(k,i)
2 + σ̃ 2 (α ) + N
2+
|
C
(α
,
α
)|
I
I
|
Ĩ
(α
,
α
)|
v→v v
i
o
i
k
i
k
v̄→v v
sv
k=1
k=1
where in general,
(k,i)
Iv→w
=
P
(k) (k)
Pvj hvj
(k)
of ho , it can be written as
Ov(i) (θ) = 1−
E e
η
−No ro vv θ
(i)
Pvo
e
j∈Ψ̃v
2 (α )rηvv θ
−σ̃s
i o
v
(i)
Pvo
K
Y
k=1
LI(k,i)
v→v
(18)
−ηvw
(k)
rvj
. By exploiting the exponential distribution
roηvv θ|Ĩv (αi , αk )|2
(i)
Pvo
!
!
roηvv θ|C˜v (αi , αk )|2
.
(i)
Pvo
LI(k,i)
(i)
v̄→v
(i)
(19)
where the expectation is over {ro , Pvo , σ̃s2v }. Since the {ro , Pvo } depends on the UEs type (CCU
or CEU) as discussed in section II-A, we present an explicit study for each type. The serving
distances ro for CCUs and CEU s are characterized via the following lemma.
Lemma 1. The serving distance distribution for a randomly selected CCU or CEU given that
it is connected to the ith tier denoted by fR(i)
(.) and fR(i)
(.), respectively, are given by the
c
e
following equations,
2π λ̄i r exp −π λ̄i r2
(
1 ) (r),
fR(i) (r) =
η2 1
ηdd
c
u
dd
0≤r≤ P(i)
Pu
¯
ρ
1 − exp −π λi ρ(i)
fR(i) (r) = 2π λ̄i r exp −π λ̄i r + π λ¯i
2
e
where λ̄i =
K
P
k=1
τk2
λ .
τk2 k
Pu
ρ(i)
η2 !
dd
1( P
u
ρ(i)
1
ηdd
(20)
) (r).
<r<∞
(21)
Proof: Refer to Appendix A.
From Lemma 1, it is straightforward to find the probability that a randomly selected UE from
the ith tier is a CCU or a CEU, which are given by,
2 !
Pu ηdd
P {CCU} = 1 − exp −π λ̄i
,
ρ(i)
2 !
Pu ηdd
P {CEU} = exp −π λ̄i
.
ρ(i)
(22)
(23)
By the law of total probability, the average outage probability can be expressed via the CCUs’
(i)
(i)
outage probability and the CEUs’ outage probability, denoted by Ovc and Ove , respectively, as
(i)
(i)
Ōv(i) (θ) = Ovc (θ)P {CCU} + Ove (θ)P {CEU} ,
(i)
(24)
(i)
where each of Ovc (θ) and Ove (θ) is represented as in (19), but with the specific parameters
related to the CCUs and the CEUs. From (19), it is clear that the LT of the aggregate interference
(k,i)
(i)
(i)
from each tier Iw→v is required to evaluate Ovc (θ) and Ove (θ). The aggregate interference, and
hence its LT, depends on the spatial distribution of the set of interfering BSs and UEs in the
(k)
(k)
(k)
tier, Ψ̃d and Ψ̃u , respectively. The set of interfering BSs Ψ̃d in the k th tier is the same as
(k)
the original set of BSs Ψd excluding the transmitting BS itself in the DL and the serving BS
P
(k)
in the UL. Hence, Ψ̃d is a PPP with intensity λk . From UEs associations and λu ≫ K
k=1 λk ,
(k)
the intensity of the interfering UEs Ψ̃u on a certain channel in the k th tier is also λk . However,
(k)
Ψ̃u is not a PPP because only one UE can use a channel in each Voronoi-cell, which impose
correlations among the positions of the interfering UEs on each channel and violates the PPP
(k)
assumption. Furthermore, the employed association makes the set of interfering UEs Ψ̃u and
(k)
the set of interfering BSs Ψ̃d
correlated. The inter-correlations between the interfering UEs
and the cross-correlations between the UEs and BSs impede the model tractability. Hence, to
maintain the tractability, we ignore these correlations. The used assumptions to keep the model
tractability are formally stated below.
(k)
Assumption 1. The set of interfering UEs Ψ̃u in the k th tier is a PPP with intensity λk .
(k)
(k)
Assumption 2. The point process Ψ̃d for the interfering BSs and the point process Ψ̃u for
the interfering UEs both in the k th tier are independent.
(k)
Assumption 3. The point processes Ψ̃u ’s which represent the interfering UEs connected to
different tiers are independent from each other.
Remark 1. The previous assumptions are necessary to maintain the model tractability. Assumption 1 has been used and validated in [14], [17], [29], [30], Assumption 2 in [15], [17],
and Assumption 3 in [29]. It is important to mention that these assumptions ignore the mutual
correlations between the interfering sources, however, the correlations between the interfering
sources and the test receiver are captured through the proper calculation for the interference
exclusion region enforced by association and/or UL power control. The accuracy of the developed
model with Assumptions 1-3 is validated via independent Monte Carlo simulation in Section IV.
Based on Assumptions 1-3, the LT of the aggregated interference is always generated from a
PPP Φ, but with different parameters such as interference exclusion regions, interferers intensity,
and transmit power distribution. For brevity, we present the following unified lemma for the LT
of the aggregate interference generated from a homogeneous PPP with general parameters and
use it to obtain all LTs in (19).
Lemma 2. Let LI (s) be the LT of the aggregate interference I generated from a PPP network
with intensity λ, i.i.d transmit powers Pj , unit means i.i.d exponentially distributed channel power
gains hj , and per-interferer protection region of B(o, aj ), where B(o, aj ) is a ball centered at
the origin (o) and has a radius aj . Then, LI (s) is given by,
LI (s) = exp
2
2
−2πλ
2−η
−η
sP 2 F1 1, 1 − ; 2 − ; −a P s
EP a
,
η−2
η
η
(25)
where 2 F1 (·, ·; ·; ·) is the hyper-geometric function [45], EP [·] is the expectation over the transmitted power of the sources, and η > 2 is a general path loss exponent. For the special case of
a = 0, equation (25) reduces to,
h
i
2
2
2π 2 λ
.
EP (sP ) η csc
LI (s) = exp −
η
η
(26)
Proof: Refer to Appendix B.
Due to the expectation over power distribution, the LT expression in (25) has an integral over
hyper-geometric function. If the interference exclusion distance a around the test receiver is
independent of the transmit powers, then the expression given by (25) can be lower-bounded by
the simplified closed-form expression given in the following lemma.
Lemma 3. Let LI (s) be the LT of the aggregate interference generated from a PPP network with
intensity λ, i.i.d transmit powers Pj , unit means i.i.d exponentially distributed channel power
gains, and interference protection region of B(o, a), where B(o, a) is a ball centered at the
origin (o) and has a radius a. Assuming that a is independent from Pj , ∀j, then LI (s) can be
lower-bounded by,
LI (s) ≥ exp
2
2
−2πλ 2−η
−η
a
sE [P ] 2 F1 1, 1 − ; 2 − ; −a E [P ] s .
η−2
η
η
(27)
Proof: Refer to Appendix C.
Lemma 3 precludes the necessity to integrate over the PDF of the transmit power of the
interfering sources and give the LT is a closed-form containing the first moment of the transmit
power, which reduces the computational complexity of the LTs. For the sake of simplified
expressions, we always use the bound in (27) whenever applicable and is verified in the section(k,i)
IV. Using Lemma 2 and Lemma 3 , the LTs of the aggregated interference Iv→v for the UL and
DL for CCUs and CEUs are given by the following lemma.
(e)
(c)
(e)
(c)
(e)
(c)
(e)
(c)
Lemma 4. Let L (k,i) L (k,i) , L (k,i) L (k,i) , L (k,i) L (k,i) and L (k,i) L (k,i) represent
Iu→u
Iu→u
Id→u
Id→u
Id→d
Id→d
Iu→d
Iu→d
the LTs of the UL to UL, DL to UL, DL to DL, and UL to DL aggregate interference generated
from the k th tier affecting a CCU (CEU) and its serving BS given that both of them are in the
ith tier, then these LTs are given by
1− η 2
η2
uu
−2πλk ρ(k)
2
2
uu
s 2 F1 1, 1 −
,2 −
, −ρ(k) s ,
E Pu(k)
L (k,i) (s) = exp
Iu→u
ηuu − 2
ηuu
ηuu
(c)
(c)
L
(e)
(c)
(k,i)
Id→d
(29)
−2πλk
2
2
,
,2 −
, −Pu(k) sro−ηuu
EP (k) Pu(k) s 2 F1 1, 1 −
ηuu − 2 u
ηuu
ηuu
h
i
2
2
−2πλk h (k) i
(k)
−ηuu
,
,2 −
, −E Pu sro
E Pu s 2 F1 1, 1 −
& exp
ηuu − 2
ηuu
ηuu
(k,i) (s|ro ) ≈ exp
Iu→u
L
2
2π 2 λk (k) ηdu
2π
,
sP
(s)
=
exp
−
csc
(k,i)
d
Id→u
ηdu
ηdu
(e)
(k,i) (s) = L
Id→u
L
(28)
(e)
(s|ro ) = L
(k,i)
Id→d
(s) = exp
−2πλk
ηdd − 2
ro τi
τj
2−ηdd
(k)
sPd 2 F1
"
2
2
1, 1 −
,2−
,−
ηdd
ηdd
(30)
(31)
ro τi
τj
−ηdd
(k)
Pd s
#!
(32)
,
1− η 2
η2
ud
−2πλk ρ(k)
2
2
(k,i)
ud
s 2 F1 1, 1 −
L (k,i) (s) ≈ exp
,2 −
, −ρ(k) s U1 (ro , s),
E Pu(k)
Iu→d
ηud − 2
ηud
ηud
(c)
(e)
L
(k,i) (s) ≈ exp
Iu→d
where,
h
i
2
2
−2πλk h (k) i
(k,i)
,2 −
, −E Pu(k) sro−ηud
U1 (ro , s),
E Pu s 2 F1 1, 1 −
ηud − 2
ηud
ηud
ρ
ζ
E Pu(k)
=
(k,i)
and U1
(k,i)
U1
(k) ζ
γ
ζηdd
2
+ 1, π λ̄k
π λ̄k
ζη2dd
Pu
ρ(k)
η2
dd
ζ
+ (Pu ) exp π λ̄k
Pu
ρ(k)
η2 !
dd
(33)
(34)
,
(35)
(ro , s) is the LT of intra-cell interference in the 3NT case, which is expressed as
(ro , s) =
Pu
ρ(i)
R
0
1
ηud
Rπ
δo
(i) (r)
Rc
P{CCU}f
1+sρ(i) (1+( rro )2 −2 rro cos(δ))
−ηud
2
dδdr
π−δo
+
R∞
Pu
ρ(i)
1
ηud
Rπ
δo
(i) (r)
Re
P{CEU}f
1+sPu (r 2 +ro2 −2ro r cos(δ))
−ηud
2
dδdr
π−δo .
1.
3NT
i=k
O.W
(36)
and γ(·, ·) is the lower incomplete gamma function [45].
Proof: Refer to Appendix D.
(k,i)
Note that U1
(·, ·) in equations (33) and (34) represents the intra-cell interference, in which
it has an effect on the 3NT case from tier that tagged transceiver belongs to. For the sake
(k,i)
of simplified expressions, we also present a closed-form approximation for U1
(·, ·) in the
following lemma.
Lemma 5. The LT of the intra-cell interference given in equation (36) can be approximated by
(k,i)
U1
(ro , s) =
P{CCU}
sin(δo )
1+sρ(i) (1+( rr̄oc )2 + 2 π−δ
o
1.
ro
r̄c
)
−ηud
2
+
P{CEU}
sin(δo )
ro r̄e )
1+sPu (r̄e2 +ro2 + 2 π−δ
o
−ηud
2
.
3NT
i=k
(37)
O.W
where
erf
r̄c =
p
η1
η1
η2
p
ud
ud
ud
u
u
u
− 2 λ̄i ρP(i)
π λ¯i ρP(i)
exp −π λ̄i ρP(i)
.
η 2
p
ud
u
2 λ̄i 1 − exp −π λ̄i ρP(i)
(38)
p
η2
η1
η1
p
ud
ud
ud
u
u
u
erfc
− 2 λ¯i ρP(i)
π λ̄i ρP(i)
exp π λ̄i ρP(i)
p
.
r̄e =
2 λ¯i
(39)
where erf(·) and erfc(·) are the error function and the complementary error function, respectively
[45].
Proof: By substituting r and cos(δ) by their average values.
Using the results in Lemmas 1-5 along with (19), the outage probabilities for all types of
connections in the depicted system model are characterized via the following theorem.
Theorem 1. The outage probabilities in the UL and the DL in the ith tier for CCUs and CEUs
are given by,
(i)
Ouc (θ)
(i)
Oue (θ)
=1−e
−No θ
ρ(i)
Z∞
=1−
(i)
USIu
Pu
ρ(i)
e
(i)
Odc (θ)
=1−
Pu
ρ(i)
Z
Y
K
k=1
η
−No ro uu θ
Pu
L
K
Y
(i)
k=1
(e)
L
(k,i)
Iu→u
e
Z∞
Pu
ρ(i)
θroηuu
Pu
θ|C˜u (αi , αk )|2
ρ(i)
!
,
(40)
roηuu θ|C˜u (αi , αk )|2
Pu
!
dro ,
(41)
roηdd θ|C˜d (αi , αk )|2
!
dro ,
(42)
roηdd θ|C˜d (αi , αk )|2
!
dro ,
(43)
(c)
L
(k,i)
Id→u
fR(i) (ro )
e
roηuu θ|Ĩu (αi , αk )|2
Pu
η
−No ro dd θ
(i)
P
d
(i)
USId
0
=1−
(k,i)
Iu→u
USIu
!
θ|Ĩu (αi , αk )|2
ρ(i)
(c)
!
(e)
L
(k,i)
Id→u
1
ηdd
×
(i)
Ode (θ)
θ
ρ(i)
1
ηuu
×
K
Y
k=1
e
(c)
L
(k,i)
Id→d
η
−No ro dd θ
(i)
P
d
×
K
Y
k=1
(e)
L
(k,i)
Id→d
(i)
Pd
!
fR(i) (ro )
c
roηdd θ|Ĩd (αi , αk )|2
(i)
Pd
(i)
USId
1
ηdd
θρro2ηdd
θPu roηdd
(i)
Pd
!
(i)
(c)
L
(k,i)
Iu→d
(i)
Pd
fR(i) (ro )
e
roηdd θ|Ĩd (αi , αk )|2
Pd
!
!
(e)
L
(k,i)
Iu→d
(i)
Pd
where,
(i)
USIu (x)
=
Z∞
0
(i)
exp −xβu(i) hPd |C˜u (αi , αi )|2 fHs (h)dh,
(44)
(i)
USId (x) =
∞
R
exp −xβd h|C˜d (αi , αi )|2 fHs (h)dh.
2NT
(45)
0
1.
3NT
(·) and fR(i)
(·) are given in equations (20)
and fHs (·) is the distribution of the SIC power. fR(i)
c
e
and (21), respectively, and the LTs are given in Lemma 4.
Proof: Refer to Appendix E.
A special case of interest that leads to simple forms of the outage probability is presented in
the following corollary.
Corollary 1. In an interference limited dense single tier cellular network with unbinding UL
transmit power, the outage probability in the DL and UL, assuming ηdd = ηuu = ηud = 4,
ηdu = 3, δo = 90o , and hs ∼ exp(1), are given by
Ou (θ) ≈ 1 −
Od (θ) ≈ 1 −
√
√
exp − θ arctan
θ −
4π√2 λ
3 3
θ|C̃u (α,α)|2 Pd
ρ
1 + βu Pd |C˜u (α, α)|2 ρθ
Z∞
32
(46)
2πλro UNT (θ, ro )
0
√
√
θ −
× exp −πλro2 − πλro2 θ arctan
s
ρro4 θ|C˜d (α, α)|2
Pd
where,
UNT (θ, ro ) =
≈
Pd
.
Pd +βd |C̃d (α,α)|2 ρro8 θ
∞
π
R R
Pd λr exp(−πλr 2 )
0 δo
Pd +ro4 θ|C̃d (α,α)|2 ρ(1+( rro )2 −2 rro cos(δ))
Pd
.
Pd +βd |C̃d (α,α)|2 ρro8 θ
Pd
√
−2 dr.
8
Pd +ro4 θ|C̃d (α,α)|2 ρ(1+4λro2 + π
λro )
s
arctan
ρro4 θ|C˜d (α, α)|2
Pd
dro
(47)
2NT
−2
dδdr.
3NT
(48)
2NT
3NT
(49)
Proof: The expressions follow from Theorem 1 and Lemmas 4-5 by considering a single
tier network and setting Pu → ∞.
Following (8), the rate can be expressed in terms of the outage probability as follows
R(θ) = BW log2 (1 + θ) (1 − O(θ)) .
(50)
Hence, general expressions for the α-duplex rate in a multi-tier network can be obtained by
directly substituting the outage probability expressions from Theorem 1 in equation (8). For the
sake of brevity, we only list the rate expressions for a special case of interest in the following
Corollary.
Corollary 2. In an interference limited dense single tier cellular network with unbinding UL
transmit power, the average rate in the DL and UL, assuming ηdd = ηuu = ηud = 4, ηdu = 3,
δo = 90o, and hs ∼ exp(1), are given by
√ 4π 2 λ
√
(Bu + (ǫ + 1)αB) log2 (1 + θ)
− θ arctan
exp
Ru (θ) ≈
θ − √
3 3
1 + βu Pd |C˜u (α, α)|2 ρθ
θ|C˜u (α, α)|2 Pd
ρ
! 23
.
(51)
Rd (θ) ≈ 2πλ(Bd + (ǫ + 1)αB) log2 (1 + θ)
s
s
Z∞
√
4
2
4
2
˜
˜
√
ρro θ|Cd (α, α)|
ρro θ|Cd (α, α)|
arctan
UNT (θ, ro )dro
θ −
× ro exp −πλro2 − πλro2 θ arctan
Pd
Pd
0
(52)
where,
UNT (θ, ro ) =
≈
Pd
.
Pd +βd |C̃d (α,α)|2 ρro8 θ
R∞ Rπ
Pd λr exp(−πλr 2 )
0 δo
1+( rro )2 −2 rro
(
Pd +ro4 θ|C̃d (α,α)|2 ρ
2NT
cos(δ))
Pd
.
Pd +βd |C̃d (α,α)|2 ρro8 θ
Pd
√
−2 dr.
8
Pd +ro4 θ|C̃d (α,α)|2 ρ(1+4λro2 + π
λro )
−2
dδdr.
3NT
(53)
2NT
3NT
(54)
Proof: Follows from Corollary 1 and equation (50).
From the last corollary, we can find the critical SIC βd at which the 2NT outperforms the
3NT as a function of the serving distance (ro ). This value is given by the following corollary.
Corollary 3. In an interference limited dense single tier cellular network with unbinding UL
transmit power, the approximate minimum value of βd required in the 2NT to outperform 3NT
as a function of the serving distance (ro ), assuming ηud = 4, δo = 90o , fully-overlapped channels
(α = 1 and |C˜d (α, α)|2 = 1), and hs ∼ exp(1), is given by
−2
8√ 3
2
4
λro + ro
βd ≈ 4λro +
π
(55)
TABLE I: Parameters Values.
Parameter
Value
Parameter
Value
Pu
3W
Pd
5W
1 BSs/km
No
0
1 MHz
θ
1
βd
−75 dB
βu
−110 dB
ρ
-60 dBm
ǫ
0.03134
2
λ
BuHD ,
BdHD
2
bd /bu
Sinc/Sinc
δo
90o
ηuu , ηdd , ηud
4
ηdu
3
Proof: Follows from equation (54).
Equation (55) expresses the critical value of βd as a function of ro and λ. To get more insights
on the critical value of βd with respect to the BSs’ intensity λ , we assume that the tagged UE
is located at the average serving distance, this assumption reduces (55) to
βd ≈
16λ2
9
(56)
In the next section, Theorem 1, Lemmas 1-5, and Corollaries 1-3 are used to analyze the
performance of cellular network under 2NT and 3NT operations.
IV. S IMULATIONS
AND
N UMERICAL R ESULTS
Throughout this section, we verify the developed mathematical paradigm via independent
system level simulations, where the BSs are realized via a PPP over an area of 600 km2 . Then,
the UEs are distributed uniformly over the area such that each BS has at least two UEs within its
association area. Each BS randomly selects two UEs to serve such that the δo angle separation
as illustrated in Fig. 2b is satisfied. The SINR is calculated by summing the interference powers
from all the UEs and the BSs after multiplying by the effective interference factors. In the UL,
the transmit powers of the UEs are set according to the power control discussed in Section II.
The results are taken for the UE and the BSs that are closest to the origin to avoid the edge
effect. Unless otherwise stated, the parameters values in Table 1 are used. Note that for the
average SIC power, the maximum reported value according to [7] is -110 dB, and hence, we set
βu to -110 and consider that βd ≥ βu because the BSs are more likely to have more powerful
SIC capabilities.
1.6
1.2
βd = −110 (dBm)
0.9
0.6
1.2
Rdc (α)
Rd (α)
Rde (α)
1.8
Rate (Mbps)
1.5
Rate (Mbps)
2
Rdc (α)
Rd (α)
Rde (α)
1.8
Ruc (α)
Rue (α)
Ru (α)
1
Rate (Mbps)
2.1
1.4
1.2
1
0.8
0.6
0.4
0.8
βd = −75 (dBm)
0.3
0
0
0.2
0.4
0.2
0.6
0.6
0.8
1
0.4
0
0.2
0.4
0.6
0.8
1
0
0
0.2
0.4
α
α
(a) 2NT DL throughput.
(b) 3NT DL throughput.
0.6
0.8
1
α
(c) UL throughput.
Fig. 4: Rates vs α for the 3NT and 2NT.
For the pulse shaping, we consider two basic pulse shapes, namely, Sinc2 and Sinc pulse
shape10 , which have the FTs given in (57),
S(f, BW, b) =
s
2f
)
SINC( BW
∞
R
2f
2
SINC ( BW )df
b = Sinc.
−∞
s
2f
SINC2 ( BW
)
∞
R
2f
4
SINC ( BW )df
b = Sinc2 .
(57)
−∞
where b = Sinc when the Sinc pulse is considered and b = Sinc2 when Sinc2 pulse is considered.
Unless otherwise stated, the SIC power distribution fHs (·) is assumed to be exponentially
distributed with unit mean.
Fig. 4 shows the rate variation for UL and DL versus α for the 2NT and 3NT, where α = 0 and
α = 1 represent the FD and the HD cases, respectively, the solid lines represent the analytical
results obtained from Theorem 1 with the exact LTs in Lemma 4, diamonds represent the results
obtained by simulations, the squares in Fig. 4b are found by using the approximation for intracell interference given in Lemma 5, and the squares in Fig. 4c by using the bounds for the
LTs given in Lemma 4. The close match between the analysis, approximations, and simulation
results validates the developed mathematical model and verifies the accuracy of the assumptions
in Section III-A, as well as the bound presented in Lemmas 4-5.
Several insights can be obtained from Fig .4. For instance, the figure shows that the CCUs
have better performance compared to the CEUs in all cases, which is intuitive due to the larger
service distances that lead to higher path-loss attenuation for CEUs compared to the CCUs. Note
10
Employing and designing more sophisticated pulse shapes for specific purposes is left for future work.
1
2
0.8
1.75
0.4
0.2
0
-150
1.5
α = 0.288
α = 0 (HD)
Rd (α) 2NT
Rd (α) 3NT
Rd (α) 3NT Approx.
-130
-110
1.25
0.75
0.5
-70
-50
α = 0.288
1
0.25
-90
0.6
Rde (α) (Mbps)
0.6
α = 1 (FD)
Rdc (α) (Mbps)
Rd (α) (Mbps)
0.8
0
-150
βd (dB)
(a) Average DL rate.
α = 1 (FD)
0.7
α = 1 (FD)
α = 0 (HD)
-110
0.4
α = 0 (HD)
0.3
0.2
Rdc (α) 2NT
Rdc (α) 3NT
Rdc (α) 3NT Approx.
-130
α = 0.288
0.5
0.1
-90
-70
-50
0
-150
βd (dB)
(b) CCU DL rate.
Rde (α) 2NT
Rde (α) 3NT
Rde (α) 3NT Approx.
-130
-110
-90
-70
βd (dB)
(c) CEU DL rate.
Fig. 5: DL Rates vs βd under different network topology, where FD denotes α = 1 and αD
denotes α ≈ 0.2886.
that the CEUs do not have sufficient power to invert their path-loss in the UL direction, and
hence, the received power at the serving BS is less than ρ, which leads to the deteriorated UL
CEU performance when compared to the CCU case. The figure also shows that there exist an
optimal value of partial overlap 0 < α < 1 that maximizes the UL transmission rate11 . Hence,
despite the efficient SIC (-110 dB), neither HD nor FD are optimal in the UL case due to the
prominent DL interference. On the other hand, the DL performance is mainly affected by the
SIC rather than the UL interference12 . Particularly, for UE with efficient SIC, the full overlap
(i.e., FD) is the best strategy for the DL. On the other hand, partial overlap is better for UE
with inefficient SIC. It is worth mentioning that the SI has more prominent effect on the CEU
than CCU, in which the SI nearly nullifies the DL rate for high value of α and efficient SIC.
This is because CEUs transmit with their maximum power, which makes the residual SI power
more prominent compared to CCUs.
Comparing Fig. 4.a and Fig. 4.b, we can see that the 3NT achieves close performance to the
2NT with sufficient SIC, and outperforms the 2NT with poor SIC. Note that 3NT UEs operates
in HD mode and hence are not affected by SIC as shown in Fig. 4b.
Fig. 5 plots the DL transmission rate vs SI attenuation power for the 2NT and 3NT. The figure
11
The UL performance is maximized at α = 0.28859 due to the orthogonality between the used pulse shapes at this particular
value, for more details refer to [17].
12
In the case of perfect SI or very low values of βd , e.g. βd = −110 dBm, the degradation in the SINR is only due to
the UL-to-DL interference, and since this is negligible compared to the DL-to-DL interference for a realistic set of network
parameters, the increase in BW (linearly) overcome the decrease in the SINR which results in approximately linear curve.
-50
shows that the DL FD rate outperforms the DL rate of both the HD and the α = 0.288 in all the
cases. Fig. 5 also shows that there is a critical value, for βd , at which the 3NT outperforms 2NT.
This critical value can be interpreted as the point at which the SI experienced by DL UEs in the
2NT becomes more significant than the intra-cell interference experienced by DL-UEs in 3NT
and we found an closed form approximation for it in Corollary 3 and (56). Interestingly, the gain
offered by the 2NT at low values of βd is not significant when compared to the 3NT. Hence, the
intra-cell interference is not a limiting parameter for the 3NT. In other words, network operators
can harvest FD gains by HD UEs almost similar to the gains harvested by FD UEs with efficient
SIC capabilities. The figure also shows that, in case of poor SIC at the UEs, the 3NT can offer
significant gains, specially for CEUs, when compared to the 2NT case.
To study the effect of the serving distance on the 2NT/3NT performance, we plot Fig. 6a
for λ = 20 BSs/km2 . The figure plots the minimum βd required in 2NT to outperform 3NT vs.
the serving distance along with the pdf of the serving distance, where sold (dashed) lines are
obtained from the exact (approximate) expression of the intra-cell interference given in equation
(53). The close match between the exact and the approximate results validates the approximation
given in (53). As the figure shows, the 3NT is more appealing for farther UEs because they
have a tighter constraint for the SIC β required for the 2NT to outperform the 3NT, which may
require more sophisticated and expensive FD transceivers. There are two reasons for this result;
first, large serving distance implies that the intra-cell interferer in 3NT is further on average
due to the enforced scheduling technique, which reduces the negative effect of the intra-cell
interference. Second, longer service distance implies larger transmit power due to the employed
power control, hence more powerful SIC is required. A useful design insight for Fig. 6 is that the
BSs should select the mode of operation (i.e., 3NT or 2NT) for the UEs based on their distances
along with their SIC. To get more insights on the network operation for different intensities we
plot Fig. 6b based on equation (56). As expected, 2NT becomes more appealing in dense cellular
networks because the intra-cell (self) interference is more(less) prominent in smaller cell area13 .
Finally, we study the rate gains of the 2NT/3NT with δo in Fig. 6c. As expected, by increasing
δo the distance between the two scheduled UEs increases, hence the intra-cell power in 3NT
13
Fig. 6b focuses on the comparing 2NT/3NT with different intensities. The effect of intensity on the FD gain for 2NT in
cellular network is covered in [17] and for ad-hoc network with CSMA-based transmitters is covered in [46].
0.0086
0
0.0071
-20
0.0057
-40
0.0043
-60
0.0029
-80
0.0014
-100
0
50
100
150
200
-50
2
-55
0
250
Rd (α) (Mbps)
0.01
20
fR (ro )
βd (dBm)
βd (dBm)
40
-60
-65
-70
10
20
ro (Meters)
30
40
λ (km2 )
50
60
Rcd (α)
Rcd (α)
Rd (α)
Rd (α)
Red (α)
Red (α)
1.5
1
0.5
o
0
45
o
90
o
135
o
2NT
3NT
2NT
3NT
2NT
3NT
180
o
δo
(a) Critical values of βd vs. the serv- (b) Critical values of βd vs. the BSs’
(c) DL rate vs δo as illustrated in Fig.
ing distance.
2b.
intesnity.
Fig. 6: (a), (b) Critical values of βd vs. the serving distance and BSs’ intesnity using equations
(56) and (57), and (c) DL rate vs δo as illustrated in Fig. 2b.
operation decreases. Moreover, the figure shows the necessity of UEs scheduling and multi-user
diversity in 3NT, otherwise the rate loss in 3NT compared to 2NT can go up to 20% in the case
of random scheduling (δo = 0).
V. C ONCLUSION
This paper presents a mathematical paradigm for multi-tier cellular networks with FD BSs and
HD/FD UEs. The presented model captures detailed system parameters including pulse shaping,
filtering, imperfect self-interference cancellation, partial uplink/downlink overlap, uplink power
control, limited users’ transmit powers, UE-BS association, and UEs’ scheduling. To this end,
unified rate expressions for 2-nodes topology (2NT) with FD users and 3-nodes topology (3NT)
with HD users are presented and used to compare their performance. The results show that there
exist a critical value for the self-interference cancellation, at which the performance of 3NT
outperforms the 2NT. Moreover, closed form approximations for this critical value as a function
of the serving distance and the BSs’ intensity are obtained. The results also show that even when
SI is efficiently canceled, the 2NT does not offer significant gains when compared to the 3NT
operation if multi-user diveristy and users’ scheduling are exploited. This implies that network
operators can harvest FD gains by implementing FD transceivers at their BSs regardless of the
state of the users (i.e., FD or HD).
A PPENDIX A
P ROOF
OF
L EMMA 1
Exploiting independence between the network tiers and using the null probability of the PPP,
the cumulative distribution function (CDF) of the ith -tier service distance is given by,
τ
FR(i) (r) = P{R(i) ≤ r} = P{Ri ≤ r|Ri ≤
P{Ri ≤ r ∩ Ri ≤ τji rj ∀j 6= i}
τj
rj ∀j ∈ {1, ..., K}, j 6= i} =
τ
τi
P{Ri ≤ τji rj ∀j 6= i}
(58)
The denominator is given by,
P{Ri ≤
τj
rj ∀j 6= i} =
τi
Z∞
0
fRi (ri )
Z∞
M
Y
j6=i
τi
j=1 (ri τj
and the nominator is given by,
τj
P{ri ≤ r ∩ ri ≤ rj ∀j 6= i} =
τi
Z∞
0
)
λi
fRj (rj )drj
dri = P
M
τi2
j=1
τj2
(59)
λj
λi
M Z∞
Y
fRj (rj )drj dri =
fRi (ri )
j6=i τi
j=1ri τ
j
1 − exp −πr2
M
P
j=1
τi2
τj2
M
P
j=1
τi2
λ
τj2 j
!!
λj
(60)
By substituting equations (59) and (60) in (58) it results in,
FR(i) (r) = 1 − exp −πr2
where λ̄i =
M
P
j=1
τi2
λ
τj2 j
M
X
τ2
i
2 λj
τ
j=1 j
= 1 − exp −π λ̄i r2
r ≥ 0.
(61)
and the PDF is given by,
fR(i) (r) = 2π λ̄i r exp −π λ̄i r2
r ≥ 0.
(62)
Given that the UE is a CCU, the PDF in (62) should be truncated according to channel
(i)
inversion power control. Let Rc denote the serving distance for a test CCU connected to the
ith tier, then its PDF is given by,
fR(i) (r) =
c
2π λ̄i r exp −π λ̄i r2
Pu
ρ(i)
R
0
1
ηdd
2π λ̄i r exp −π λ̄i r2 dr
2π λ̄i r exp −π λ̄i r2
(
1 ) (r).
=
η2 1
ηdd
u
dd
0≤r≤ P(i)
Pu
ρ
1 − exp −π λ̄i ρ(i)
(63)
Similarly, for the PDF of the service distance for a CEUs,
fR(i) (r) =
R∞
e
Pu
ρ(i)
2π λ̄i r exp −π λ̄i r2
1
ηdd
2π λ̄i r exp −π λ̄i r2 dr
= 2π λ̄i r exp −π λ¯i r2 + π λ̄i
Pu
ρ(i)
η2 !
dd
1(
Pu
ρ(i)
1
ηdd
) (r).
<r<∞
(64)
A PPENDIX B
P ROOF
OF
L EMMA 2
The proof is as follows,
LI (s) = E exp
(i)
= EΦ
Y
X
j∈Φ
−sPj hj rj−η 1 (rj > aj ) ,
−η
Ehj ,Pj exp −sPj hj rj 1 (rj > aj ) ,
rj ∈Φ
Z
(ii)
= exp −2πλEP
∞
a
(iii)
= exp
Eh
1 − exp −sP hr−η
rdr
,
2
−2πλ
2
2−η
−η
sP 2 F1 1, 1 − , 2 − , −a P s
EP a
,
η−2
η
η
(65)
where, (i) follows from the independence between Ψ̃ and hj , (ii) by using the probability
generation functional (PGFL) of PPP and (iii) by using the LT of h and by evaluating the
integral.
A PPENDIX C
P ROOF
OF
L EMMA 3
The lemma is proved by showing that the second derivative of the function which appears
inside the expectation of the exponent in (27) is positive w.r.t P. Hence, the function of interest
is convex in P and the result in Lemma 3 follows from Jensen’s inequality [47, Section 3.1.8].
Let y = a−η P s, the function of interest, denoted here as G(y), can be expressed as
2
2
G(y) = −y 2 F1 1, 1 − , 2 − , −y
η
η
(66)
The second derivative of G(y) is given by
2
d G(y)
=
dy 2
2 F1
1, 1 − η2 ; 2 − η2 ; −y −
y
1
y+1
2
1
1− 2 .
−
+
η
(y + 1)2
η
(67)
where (67) is found by using [45, Eqs (15.2.2),(15.2.10),(15.2.27)]
and some mathematical
2
2
1
simplifications. Owing to the fact that (1+y)2 , 1 − η , η , and y are positive for η > 2, the
1
is positive.
prove is completed by proving that G2 (y) = 2 F1 1, 1 − η2 ; 2 − η2 ; −y − y+1
Using the integral definition of the hypergeometric function [45, Eq. (15.3.1)] and projecting it
on our case, we have
Γ 2 − 2 Z1 −2
Z1 −2
Z1 −2
(ii)
(iii)
η
tη
tη
tη
2
2
1
2
(i)
dt = 1 −
dt >
dt >
> 0.
2 F1 1, 1 − ; 2 − ; −y =
2
η
η
η
1 + ty
1 + ty
1+y
Γ 1 − η 0 1 + ty
0
0
(68)
where (i) follows by [45, Eq. (6.1.15)], (ii) follows from the fact that η > 2, and (iii) is proved
as in the sequel. Taking the first derivative of the integrand in (68) as
t−2/η
ty + 1
′
=
t−
η+2
η
(−(η + 2)ty − 2)
,
η(ty + 1)2
(69)
shows that it is a decreasing function in t, and hence, the minimum occurs at the boundary 1.
Then (iii) in (68) follows by lower-bounding the integral by the minimum value of the integrand
multiplied by the integration region. Hence, the second derivative of G(y) in (66) is positive,
which completes the prove.
A PPENDIX D
P ROOF
OF
L EMMA 4
Based on Lemma 2 and 3, we only need to determine the interference exclusion region (IER)
for each tier (aj ) in each case.
•
LI (k,i) (s): Due to the association rule in Section II-C, ro τi ≤ rj τk is always satisfied. Hence,
d→d
the IER is defined by B(o, ro ττki ), by substituting a in Lemma 2 by ro ττki , the final expression
is found.
(c)
•
L
(k,i)
Id→d
•
L
Id→d
(e)
(k,i)
1
u
)η .
(s): Based on the power inversion for CCUs and following [29], a = ( ρP(k)
(s): Based on the power inversion for CEUs and following [30], a = ro , then by using
Lemma 3, the final expression is found.
•
LI (k,i) (s): The PPP assumption of the BSs location implies that there is no IER for both
d→u
•
cases (CCUs and CCUs), and hence a = 0.
LI (k,i) (s): We assume that the tagged BS is collocated its associated UE, hence LI (k,i) (s) =
u→d
u→d
LIu→u
(k,i) (s) in 2NT. In 3NT, the effect of intra-cell interference should be considered also. Let
(k,i)
U1 (s) denote the LT of the intra-cell interference and let Pu1 , h1−o , r1−o , and r1 denote
the transmitted power of the interfering user, the channel gain between the two users, the
distance between them and the distance between the interfering UE and the serving BS,
respectively, then the LT of the interfering power can be expressed as
(i,i)
U1
i
h
−η
(s) = E e−sPu1 h1−o r1−o ,
i
h
−η/2
2
2
(i)
,
= E e−sPu1 h1−o (ro +r1 −2ro r1 cos(δ))
(70)
where (i) follows by using the cosine rule (cf. Fig. 2b), where δ is the uniformly distributed
between δo and π. When the other UE is a CCU, which has a probability P{CCU}, then
Pu1 = ρ(i) r1η , and when it is CEU, which has a probability P{CEU}, then Pu1 = Pu by
(k,i)
substituting these values and by averaging over h1−o , the expression for U1
(s) is found.
A PPENDIX E
P ROOF
OF
T HEOREM 1
Starting by the UL outage probability, for CCUs the transmitted power is equal to ρroη , and
for the CEUs the transmitted power is set to the maximum Pu , by substituting these values in
(i)
(19) we get equations (40) and (41) except USIu which is found by substituting σ̃s2 by its value
given in (16) and then by averaging over hs while conditioning on ro . Similar steps are followed
to find the outage in the DL direction.
R EFERENCES
[1] A. AlAmmouri, H. ElSawy, and M.-S. Alouini, “Harvesting full-duplex rate gains in cellular networks with half-duplex
user terminals,” in Proc. 2016 IEEE International Conference on Communications (ICC), Accepted., 2016.
[2] D. Bharadia, E. McMilin, and S. Katti, “Full duplex radios,” in Proc. of the ACM SIGCOMM 2013 Conf.
on SIGCOMM, ser. SIGCOMM ’13.
New York, NY, USA: ACM, 2013, pp. 375–386. [Online]. Available:
http://doi.acm.org/10.1145/2486001.2486033
[3] S. Hong, J. Brand, J. Choi, M. Jain, J. Mehlman, S. Katti, and P. Levis, “Applications of self-interference cancellation in
5G and beyond,” IEEE Wireless Commun. Mag., vol. 52, no. 2, pp. 114–121, Feb. 2014.
[4] A. Sabharwal, P. Schniter, D. Guo, D. Bliss, S. Rangarajan, and R. Wichman, “In-band full-duplex wireless: Challenges
and opportunities,” IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. 1637–1652, Sep. 2014.
[5] D. Kim, H. Lee, and D. Hong, “A survey of in-band full-duplex transmission: From the perspective of phy and mac layers,”
IEEE Commun. Surveys Tuts., vol. 17, no. 4, pp. 2017–2046, Fourthquarter 2015.
[6] H. Alves, C. de Lima, P. Nardelli, R. Demo Souza, and M. Latva-aho, “On the average spectral efficiency of
interference-limited full-duplex networks,” in Proc. 9th International Conf. Cognitive Radio Oriented Wireless Networks
and Communications (CROWNCOM), Jun. 2014, pp. 550–554.
[7] S. Goyal, P. Liu, S. Panwar, R. Difazio, R. Yang, and E. Bala, “Full duplex cellular systems: will doubling interference
prevent doubling capacity?” IEEE Commun. Mag., vol. 53, no. 5, pp. 121–127, May 2015.
[8] H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic geometry for modeling, analysis, and design of multi-tier and cognitive
cellular wireless networks: A survey,” IEEE Commun. Surveys Tuts., vol. 15, no. 3, pp. 996–1019, Third quarter 2013.
[9] M. Haenggi, Stochastic Geometry for Wireless Networks.
Cambridge University Press, 2012, cambridge Books Online.
[Online]. Available: http://dx.doi.org/10.1017/CBO9781139043816
[10] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to coverage and rate in cellular networks,” IEEE Trans.
Wireless Commun., vol. 59, no. 11, pp. 3122–3134, Nov. 2011.
[11] W. Lu and M. D. Renzo, “Stochastic geometry modeling of cellular networks: Analysis, simulation and experimental
validation,” CoRR, vol. abs/1506.03857, 2015. [Online]. Available: http://arxiv.org/abs/1506.03857
[12] X. Xie and X. Zhang, “Does full-duplex double the capacity of wireless networks?” in Proc. IEEE INFOCOM, Apr. 2014,
pp. 253–261.
[13] Z. Tong and M. Haenggi, “Throughput analysis for full-duplex wireless networks with imperfect self-interference
cancellation,” IEEE Trans. Commun., vol. 63, no. 11, pp. 4490–4500, Nov 2015.
[14] J. Lee and T. Quek, “Hybrid full-/half-duplex system analysis in heterogeneous wireless networks,” IEEE Trans. on Wireless
Commun., vol. 14, no. 5, pp. 2883–2895, May 2015.
[15] S. Goyal, C. Galiotto, N. Marchetti, and S. S. Panwar, “Throughput and coverage for a mixed full and half duplex small
cell network,” CoRR, 2016. [Online]. Available: http://arxiv.org/abs/1602.09115
[16] I. Randrianantenaina, H. ElSawy, and M. S. Alouini, “Limits on the capacity of in-band full duplex communication in
uplink cellular networks,” in IEEE Global Commun. Conf. (GLOBECOM) Workshops, San Diego, USA, Dec. 2015.
[17] A. AlAmmouri, H. ElSawy, O. Amin, and M. S. Alouini, “In-band α-duplex scheme for cellular networks: A stochastic
geometry approach,” IEEE Trans. Wireless Commun., submitted, 2015. [Online]. Available: http://arxiv.org/abs/1509.00976
[18] I. Randrianantenaina, H. Elsawy, H. Dahrouj, and M.-S. Alouini, “Interference management with partial uplink/downlink
spectrum overlap,” in Proc. 2016 IEEE Int. Conf. on Commun. (ICC), accepted, 2016.
[19] H. ElSawy, A. AlAmmouri, O. Amin, and M.-S. Alouini, “Can uplink transmissions survive in full-duplex cellular
environments?” in Proc. 22nd European Wireless Conference (EW), Accepted., 2016.
[20] N. H. Mahmood, G. Berardinelli, P. E. Mogensen, and F. Frederiksen, “Throughput analysis of full duplex communication
with asymmetric traffic in small cell systems,” in Proc. The Eleventh Int. Conf. on Wireless and Mobile Commun. (ICWMC),
2015, pp. 57–60.
[21] K. Sundaresan, M. Khojastepour, E. Chai, and S. Rangarajan, “Full-duplex without strings: Enabling full-duplex with halfduplex clients,” in Proc. of the 20th Annu. Int. Conf. on Mobile Computing and Networking, ser. MobiCom ’14.
ACM,
2014, pp. 55–66.
[22] C. H. M. de Lima, H. Alves, P. H. J. Nardelli, and M. Latva-aho, “Hybrid half- and full-duplex communications under
correlated lognormal shadowing,” in Proc. IEEE 81st Vehicular Technology Conf. (VTC Spring), May 2015, pp. 1–5.
[23] S. Goyal, P. Liu, S. Hua, and S. Panwar, “Analyzing a full-duplex cellular system,” in Proc. 47th Annu. Conf. on Information
Sciences and Systems (CISS), Mar. 2013, pp. 1–6.
[24] M. Mohammadi, H. A. Suraweera, I. Krikidis, and C. Tellambura, “Full-duplex radio for uplink/downlink transmission
with spatial randomness,” in Proc. IEEE Int. Conf. Commun. (ICC), Accepted., London, UK, Jun. 2015.
[25] C. Psomas and I. Krikidis, “Outage analysis of full-duplex architectures in cellular networks,” in Vehicular Technology
Conf. (VTC Spring), 2015 IEEE 81st, May 2015, pp. 1–5.
[26] I. Atzeni and M. Kountouris, “Full-duplex mimo small-cell networks: Performance analysis,” in Proc. IEEE Global
Commun. Conf. (GLOBECOM), Dec 2015, pp. 1–6.
[27] C. Psomas, M. Mohammadi, I. Krikidis, and H. A. Suraweera, “Directional antennas for interference management in
full-duplex cellular networks,” CoRR, 2016. [Online]. Available: http://arxiv.org/abs/1602.02718
[28] A. Guo and M. Haenggi, “Spatial stochastic models and metrics for the structure of base stations in cellular networks,”
IEEE Trans. Wireless Commun., vol. 12, no. 11, pp. 5800–5812, Nov. 2013.
[29] H. ElSawy and E. Hossain, “On stochastic geometry modeling of cellular uplink transmission with truncated channel
inversion power control,” IEEE Trans. Wireless Commun., vol. 13, no. 8, pp. 4454–4469, Aug. 2014.
[30] A. AlAmmouri, H. ElSawy, and M.-S. Alouini, “Load-Aware modeling for uplink cellular networks in a multi-channel
environment,” in Proc. IEEE 25th PIMRC, 2014, Washington, DC, USA, Sep. 2014, pp. 1591–1596.
[31] H. S. Dhillon, R. K. Ganti, F. Baccelli, and J. G. Andrews, “Modeling and analysis of k-tier downlink heterogeneous
cellular networks,” IEEE J. Sel. Areas Commun., vol. 30, no. 3, pp. 550–560, April 2012.
[32] V. Chandrasekhar and J. G. Andrews, “Uplink capacity and interference avoidance for two-tier femtocell networks,” IEEE
Trans. Wireless Commun., vol. 8, no. 7, pp. 3498–3509, July 2009.
[33] M. D. Renzo, A. Guidotti, and G. E. Corazza, “Average rate of downlink heterogeneous cellular networks over generalized
fading channels: A stochastic geometry approach,” IEEE Tans. Commun., vol. 61, no. 7, pp. 3050–3071, Jul. 2013.
[34] A. AlAmmouri, H. ElSawy, A. Sultan-Salem, M. D. Renzo, and M.-S. Alouini, “Modeling cellular networks in fading
environments with dominant specular components,” in Proc. 2016 IEEE Int. Conf. on Commun. (ICC), Accepted., 2016.
[Online]. Available: http://arxiv.org/abs/1602.03676
[35] J. Yun, “Intra and inter-cell resource management in full-duplex heterogeneous cellular networks,” IEEE Trans. Mobile
Comput., 2015.
[36] S. Singh, X. Zhang, and J. Andrews, “Joint rate and SINR coverage analysis for decoupled uplink-downlink biased cell
associations in hetnets,” IEEE Trans. Wireless Commun., 2015.
[37] J. Andrews, S. Singh, Q. Ye, X. Lin, and H. Dhillon, “An overview of load balancing in hetnets: old myths and open
problems,” IEEE Trans. Wireless Commun., vol. 21, no. 2, pp. 18–25, Apr. 2014.
[38] H. Elsawy, E. Hossain, and D. I. Kim, “Hetnets with cognitive small cells: user offloading and distributed channel access
techniques,” IEEE Commun. Mag., vol. 51, no. 6, pp. 28–36, Jun. 2013.
[39] R. Jain, S. Katiyar, and N. Agrawal, “Hierarchical cellular structures in highcapacity cellular communication systems,”
(IJACSA) International Journal of Advanced Computer Science and Applications., vol. 2, no. 9, pp. 52–57, 2011.
[40] P. Ash and E. Bolker, “Generalized dirichlet tessellations,” Geometriae Dedicata, vol. 20, no. 2, pp. 209–243, 1986.
[41] C. Li, J. Zhang, and K. Letaief, “Throughput and energy efficiency analysis of small cell networks with multi-antenna
base stations,” IEEE Trans. Wireless Commun., vol. 13, no. 5, pp. 2505–2517, May 2014.
[42] L. H. Afify, H. ElSawy, T. Y. Al-Naffouri, and M.-S. Alouini, “The influence of Gaussian signaling approximation on
error performance in cellular networks,” IEEE Commun. Lett, vol. 19, no. 12, pp. 2202–2205, Dec 2015.
[43] A. Giorgetti and M. Chiani, “Influence of fading on the Gaussian approximation for BPSK and QPSK with asynchronous
cochannel interference,” IEEE Trans. Wireless Commun., vol. 4, no. 4, pp. 384–389, Aug. 2005.
[44] M. Di Renzo and W. Lu, “The equivalent-in-distribution (EiD)-based approach: On the analysis of cellular networks using
stochastic geometry,” IEEE Commun. Lett, vol. 18, no. 5, pp. 761–764, May 2014.
[45] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables.
New York: Dover, 1964.
[46] S. Wang, V. Venkateswaran, and X. Zhang, “Exploring full-duplex gains in multi-cell wireless networks: A spatial stochastic
framework,” in IEEE Conf. on Computer Commun. (INFOCOM), Apr. 2015, pp. 855–863.
[47] S. Boyd and L. Vandenberghe, Convex Optimization.
New York, NY, USA: Cambridge University Press, 2004.
| 10math.ST
|
arXiv:1509.04612v2 [cs.NE] 16 Sep 2015
Adapting Resilient Propagation for Deep Learning
Alan Mosca
George D. Magoulas
Department of Computer Science and
Information Systems
Birkbeck, University of London
Malet Street, London WC1E 7HX - United Kingdom
Email: [email protected]
Department of Computer Science and
Information Systems
Birkbeck, University of London
Malet Street, London WC1E 7HX - United Kingdom
Email: [email protected]
Abstract—The Resilient Propagation (Rprop) algorithm has
been very popular for backpropagation training of multilayer
feed-forward neural networks in various applications. The standard Rprop however encounters difficulties in the context of
deep neural networks as typically happens with gradient-based
learning algorithms. In this paper, we propose a modification of
the Rprop that combines standard Rprop steps with a special
drop out technique. We apply the method for training Deep
Neural Networks as standalone components and in ensemble formulations. Results on the MNIST dataset show that the proposed
modification alleviates standard Rprop’s problems demonstrating
improved learning speed and accuracy.
I. I NTRODUCTION
Deep Learning techniques have generated many of the stateof-the-art models [1], [2], [3] that reached impressive results
on benchmark datasets like MNIST [4]. Such models are usually trained with variations of the standard Backpropagation
method, with stochastic gradient descent (SGD). In the field of
shallow neural networks, there have been several developments
to training algorithms that have sped up convergence [5],
[6]. This paper aims to bridge the gap between the field
of Deep Learning and these advanced training methods, by
combining Resilient Propagation (Rprop) [5], Dropout [7] and
Deep Neural Networks Ensembles.
A. Rprop
The Resilient Propagation [5] weight update rule was
initially introduced as a possible solution to the “vanishing
gradients” problem: as the depth and complexity of an artificial
neural network increase, the gradient propagated backwards
by the standard SGD backpropagation becomes increasingly
smaller, leading to negligible weight updates, which slow
down training considerably. Rprop solves this problem by
using a fixed update value δij , which is increased or decreased
multiplicatively at each iteration by an asymmetric factor η+
and η− respectively, depending on whether the gradient with
respect to wij has changed sign between two iterations or not.
This “backtracking” allows Rprop to still converge to a local
minima, but the acceleration provided by the multiplicative
factor η+ helps it skip over flat regions much more quickly.
To avoid double punishment when in the backtracking phase,
Rprop artificially forces the gradient product to be 0, so that
the following iteration is skipped. An illustration of Rprop can
be found in Algorithm 1.
Algorithm 1 Rprop
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
η+ = 1.2, η− = 0.5, ∆max = 50, ∆min = 10−6
pick ∆ij (0)
∆wij (0) = − sgn ∂E(0)
∂wij · ∆ij (0)
for all t ∈ [1..T ] do
∂E(t−1)
if ∂E(t)
> 0 then
∂wij · ∂wij
∆ij (t) = min{∆ij (t − 1) · η+ , ∆max }
∆wij (t) = − sgn ∂E(t)
∂wij · ∆ij (t)
wij (t + 1) = wij (t) + ∆wij (t)
∂E(t−1)
= ∂E(t)
∂wij
∂wij
∂E(t−1)
< 0 then
else if ∂E(t)
∂wij · ∂wij
∆ij (t) = max{∆ij (t − 1) · η− , ∆min }
∂E(t−1)
=0
∂wij
else
∆wij (t) = − sgn ∂E(t)
∂wij · ∆ij (t)
wij (t + 1) = wij (t) + ∆wij (t)
∂E(t−1)
= ∂E(t)
∂wij
∂wij
end if
end for
B. Dropout
Dropout [7] is a regularisation method by which only a
random selection of nodes in the network is updated during each training iteration, but at the final evaluation stage
the whole network is used. The selection is performed by
sampling a dropout mask Dm from a Bernoulli distribution
with P (mutedi ) = Dr , where P (mutedi ) is the probability
of node i being muted during the weight update step of
backpropagation, and Dr is the dropout rate, which is usually
0.5 for the middle layers, 0.2 or 0 for the input layers, and
0 for the output layer. For convenience this dropout mask
is represented as a weight binary matrix D ∈ {0, 1}M×N ,
covering all the weights in the network that can be used to
multiply the weight-space of the network to obtain what is
called a thinned network, for the current training iteration,
where each weight wij is zeroed out based on the probability
of its parent node i being muted.
The remainder of this paper is structured as follows:
• In section II we explain why using Dropout causes an
incompatibility with Rprop, and propose a modification
•
•
to solve the issue.
In section III we show experimental results using the
MNIST dataset, first to highlight how Rprop is able to
converge much more quickly during the initial epochs,
and then use this to speed up the training of a Stacked
Ensemble.
Finally in section IV, we look at how this work can be
extended with further evaluation and development.
II. R PROP AND D ROPOUT
In this section we explain the zero gradient problem, and
propose a solution by adapting the Rprop algorithm to be
aware of Dropout.
A. The zero-gradient problem
In order to avoid double punishment when there is a change
of sign in the gradient, Rprop artificially sets the gradient
product associated with weight ij for the next iteration to
∂Et+1
∂Et
= 0. This condition is checked during the
∂wij · ∂wij
following iteration, and if true no updates to the weights wij
or the learning rate ∆ij are performed.
Using the zero-valued gradient product as an indication
to skip an iteration is acceptable in normal gradient descent
because the only other occurrence of this would be when
learning has terminated. When Dropout is introduced, an
additional number of events can produce these zero values:
• When neuron i is skipped, the dropout mask for all
weights wij going to the layer above has a value of 0
• When neuron j in the layer above is skipped, the gradient
propagated back to all the weights wij is also 0
These additional zero-gradient events force additional skipped
training iterations and missed learning rate adaptations that
slow down the training unnecessarily.
B. Adaptations to Rprop
By making Rprop aware of the dropout mask Dm, we are
able to distinguish whether a zero-gradient event occurs as
a signal to skip the next weight update or whether it occurs
for a different reason, and therefore w and ∆ updates should
be allowed. The new version of the Rprop update rule for
each weight ij is shown in Algorithm 2. We use t to indicate
the current training example, t − 1 for the previous training
example, t + 1 for the next training example, and where a
value with (0) appears, it is intended to be the initial value.
All other notation is the same as used in the original Rprop:
• E(t) is the error function (in this case negative log
likelihood)
• ∆ij (t) is the current update value for weight at index ij
• ∆wij (t) is the current weight update value for index ij
In particular, the conditions at line 5 and line 18 are
providing the necessary protection from the additional zerogradients, and implementing correctly the recipe prescribed
by Dropout, by completely skipping every weight for which
Dmij = 0 (which means that neuron j was dropped out and
therefore the gradient will necessarily be 0. We expect that
this methodolgy can be extended to other variants of Rprop,
such as, but not limited to, iRprop+ [8] and JRprop [6].
Algorithm 2 Rprop adapted for Dropout
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
η+ = 1.2, η− = 0.5, ∆max = 50, ∆min = 10−6
pick ∆ij (0)
∆wij (0) = − sgn ∂E(0)
∂wij · ∆ij (0)
for all t ∈ [1..T ] do
if Dmij = 0 then
∆ij (t) = ∆ij (t − 1)
∆wij (t) = 0
else
∂E(t−1)
> 0 then
if ∂E(t)
∂wij · ∂wij
∆ij (t) = min{∆ij (t − 1) · η+ , ∆max }
∆wij (t) = − sgn ∂E(t)
∂wij · ∆ij (t)
wij (t + 1) = wij (t) + ∆wij (t)
∂E(t−1)
= ∂E(t)
∂wij
∂wij
∂E(t−1)
< 0 then
else if ∂E(t)
∂wij · ∂wij
∆ij (t) = max{∆ij (t − 1) · η− , ∆min }
∂E(t−1)
=0
∂wij
else
if ∂E(t−1)
= 0 then
∂wij
∆wij (t) = − sgn ∂E(t)
∂wij · ∆ij (t)
wij (t + 1) = wij (t) + ∆wij (t)
else
∆ij (t) = ∆ij (t − 1)
∆wij (t) = 0
end if
end if
end if
end for
III. E VALUATING ON MNIST
In this section we describe an initial evaluation of performance on the MNIST dataset. For all experiments we
use a Deep Neural Network (DNN) with five middle layers,
of 2500, 2000, 1500, 1000, 500 neurons respectively, and a
dropout rate Drmid = 0.5 for the middle layers and no
Dropout on the inputs. The dropout rate has been shown
to be an optimal choice for the MNIST dataset in [9]. A
similar architecture has been used to produce state-of-the-art
results [3], however the authors used the entire training set
for validation, and graphical transformations of said set for
training. These added transformations have led to a “virtually
infinite” training set size, whereby at every epoch, a new
training set is generated, and a much larger validation set of
the original 60000 images. The test set remains the original
10000 image test set. An explanation of these transformations
is provided in [10], which also confirms that:
“The most important practice is getting a training
set as large as possible: we expand the training set
by adding a new form of distorted data”
We therefore attribute these big improvements to the transformations applied, and have not found it a primary goal to replicate these additional transformations to obtain the state-of-theart results and instead focused on utilising the untransformed
A. Compared to SGD
From the results in Table I we see that the modified version
of Rprop is able to start-up much quicker and reaches an
error value that is close to the minimum much more quickly.
SGD reaches a higher error value, and after a much longer
time. Although the overall error improvement is significant,
the speed gain from using Rprop is more appealing because it
allows to save a large number of iterations that could be used
for improving the model in different ways. Rprop obtains its
best validation error after only 35 epochs, whilst SGD reached
the minimum after 473. An illustration of the first 200 epochs
can be seen in Figure 1.
stays below it consistently until it reaches its minimum. Also,
the unmodified version does not reach the same final error as
the modified version, and starts overtraining much sooner, and
does not reach a better error than SGD. Table I shows with
more detail how the performance of the two methods compares
over the first 200 epochs.
Unmodified Rprop vs Modified Rprop
14
Modified
Unmodified
12
Validation Error (%)
dataset, using 50000 images for training, 10000 for validation
and 10000 for testing. Subsequently, we performed a search
using the validation set as an indicator to find the optimal
hyperparameters of the modified version of Rprop. We found
that the best results were reached with η+ = 0.01, η− = 0.1,
∆max = 5 and ∆min = 10−3 . We trained all models to the
maximum of 2000 allowed epochs, and measured the error on
the validation set at every epoch, so that it could be used to
select the model to be applied to the test set. We also measured
the time it took to reach the best validation error, and report its
approximate magnitude, to use as a comparison of orders of
magnitude. The results presented are an average of 5 repeated
runs, limited to a maximum of 2000 training epochs.
10
8
6
4
2
0
20
40
60
80
100
120
140
160
180
200
Training Epoch
Fig. 2: Validation Error - Unmod. vs Mod. Rprop
C. Using Modified Rprop to speed up training of Deep Learning Ensembles
B. Compared to unmodified Rprop
The increase in speed of convergence can make it practical
to produce Ensembles of Deep Neural Networks, as the
time to train each member DNN is considerably reduced
without undertraining the network. We have been able to train
these Ensembles in less than 12 hours in total on a singleGPU, single-CPU desktop system 1 . We have trained different
Ensemble types, and we report the final results in Table II. The
methods used are Bagging [11] and Stacking [12], with 3 and
10 member DNNs. Each member was trained for a maximum
of 50 epochs.
• Bagging is an ensemble method by which several different training sets are created by random resampling of
the original training set, and each of these are used to
train a new classifier. The entire set of trained classifiers
is usually then aggregated by taking an average or a
majority vote to reach a single classification decision.
• Stacking is an ensemble method by which the different
classifiers are aggregated using an additional learning
algorithm that uses the inputs of these first-space classifiers to learn information about how to reach a better
classification result. This additional learning algorithm is
called a second-space classifier.
In the case of Stacking the final second-space classifier
was another DNN with two middle layers, respectively of
size (200N, 100N ), where N is the number of DNNs in the
Ensemble, trained for a maximum of 200 epochs with the
We can see from Figure 2 that the modified version of
Rprop has a faster start-up than the unmodified version, and
1 We used a Nvidia GTX-770 graphics card on a core i5 processor,
programmed with Theano in python
SGD vs Modified Rprop
90
Modified Rprop
SGD
80
Validation Error (%)
70
60
50
40
30
20
10
0
0
20
40
60
80
100
120
Training Epoch
140
160
180
200
Fig. 1: Validation Error - SGD vs Mod. Rprop
Method
SGD
Rprop
Mod Rprop
Min Val Err
2.85%
3.03%
2.57%
Epochs
1763
105
35
Time
320 min
25 min
10 min
Test Err
3.50%
3.53%
3.49%
1st Epoch
88.65%
12.81%
13.54%
TABLE I: Simulation results
modified Rprop. We used the same original train, validation
and test sets for this, and collected the average over 5 repeated
runs. The results are still not comparable to what is presented
in [3], which is consistent with the observations about the
importance of the dataset transformations, however we note
that we are able to improve the error in less time it took to
train a single network with SGD. A Wilcoxon signed ranks test
shows that the increase in performance obtained from using
the ensembles of size 10 compared to the ensemble of size 3
is significant, at the 98% confidence level.
Method
Bagging
Bagging
Stacking
Stacking
Size
3
10
3
10
Test Err
2.56%
2.13%
2.48%
2.19%
Time
35 min
128 min
39 min
145 min
TABLE II: Ensemble performance
IV. C ONCLUSIONS
AND
F UTURE W ORK
We have highlighted that many training methods that have
been used in shallow learning may be adapted for use in Deep
Learning. We have looked at Rprop and how the appearance of
zero-gradients during the training as a side effect of Dropout
poses a challenge to learning, and proposed a solution which
allows Rprop to train DNNs to a better error, and still be much
faster than standard SGD backpropagation.
We then showed that this increase in training speed can
be used to train effectively an Ensemble of DNNs on a
commodity desktop system, and reap the added benefits of
Ensemble methods in less time than it would take to train a
Deep Neural Network with SGD.
It remains to be assessed in further work whether this improved methodology would lead to a new state-of-the-art error
when applying the pre-training and dataset enhancements that
have been used in other methods, and how the improvements
to Rprop can be ported to its numerous variants.
ACKNOWLEDGEMENT
The authors would like to thank the School of Business,
Economics and Informatics, Birkbeck College, University of
London, for the grant received to support this research.
R EFERENCES
[1] L. Wan, M. Zeiler, S. Zhang, Y. L. Cun, and R. Fergus, “Regularization
of neural networks using dropconnect,” in Proceedings of the 30th
International Conference on Machine Learning (ICML-13), 2013, pp.
1058–1066.
[2] D. Ciresan, U. Meier, and J. Schmidhuber, “Multi-column deep neural
networks for image classification,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE
Press, 2012, pp. 3642–3649.
[3] D. C. Ciresan, U. Meier, L. M. Gambardella, and J. Schmidhuber,
“Deep, big, simple neural nets for handwritten digit recognition,” Neural
computation, vol. 22, no. 12, pp. 3207–3220, 2010.
[4] Y. Lecun and C. Cortes, “The MNIST database of handwritten digits.”
[Online]. Available: http://yann.lecun.com/exdb/mnist/
[5] M. Riedmiller and H. Braun, “A direct adaptive method for faster
backpropagation learning: The rprop algorithm,” in proceeding of the
IEEE International Conference on Neural Networks. IEEE, 1993, pp.
586–591.
[6] A. D. Anastasiadis, G. D. Magoulas, and M. N. Vrahatis, “New
globally convergent training scheme based on the resilient propagation
algorithm,” Neurocomputing, vol. 64, pp. 253–270, 2005.
[7] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov, “Improving neural networks by preventing coadaptation of feature detectors,” CoRR, vol. abs/1207.0580, 2012.
[8] C. Igel and M. Hüsken, “Improving the Rprop learning algorithm,” in
Proceedings of the second international ICSC symposium on neural
computation (NC 2000), vol. 2000. Citeseer, 2000, pp. 115–121.
[9] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: A simple way to prevent neural networks from overfitting,” The Journal of Machine Learning Research, vol. 15, no. 1, pp.
1929–1958, 2014.
[10] P.
Y.
Simard,
D.
Steinkraus,
and
J.
C.
Platt,
“Best practices for convolutional neural networks applied
to visual document analysis,” 2003. [Online]. Available:
http://research.microsoft.com/apps/pubs/default.aspx?id=68920
[11] L. Breiman, “Bagging predictors,” Machine Learning, vol. 24, no. 2, pp.
123–140, 1996.
[12] W. D, “Stacked generalization,” Neural Networks, vol. 5, pp. 241–259,
1992.
| 9cs.NE
|
Fine-grained reductions
from approximate counting to decision∗
Holger Dell1 and John Lapinskas2
arXiv:1707.04609v1 [cs.DS] 14 Jul 2017
1
Saarland University and Cluster of Excellence (MMCI), [email protected]
2
Oxford University, [email protected]
July 18, 2017
Abstract
The main problems in fine-grained complexity are CNF-SAT, the Orthogonal Vectors problem,
3SUM, and the Negative-Weight Triangle problem (which is closely related to All-Pairs Shortest
Path). In this paper, we consider the approximate counting version of each problem; thus instead
of simply deciding whether a witness exists, we attempt to (multiplicatively) approximate the
number of witnesses. In each case, we provide a fine-grained reduction from the approximate
counting form to the usual decision form. For example, if there is an O(cn )-time algorithm that
solves k-SAT for all k, then we prove there is an O((c + o(1))n )-time algorithm to approximately
count the satisfying assignments. Similarly, we get that the exponential time hypothesis (ETH)
is equivalent to an approximate counting version. This mirrors a result of Sipser (STOC 1983)
and Stockmeyer (SICOMP 1985), who proved such a result in the classical polynomial-time
setting, and a similar result due to Müller (IWPEC 2006) in the FPT setting.
Our algorithm for polynomial-time problems applies in a general setting in which we approximately count edges of a bipartite graph to which we have limited access. In particular, this means
it can be applied to problem variants in which significant improvements over the conjectured
running time bounds are already known. For example, the Orthogonal Vectors problem over
GF(m)d for constant m can be solved in time n · poly(d) by a result of Williams and Yu (SODA
2014); our result implies that we can approximately count the number of orthogonal pairs with
essentially the same running time. Moreover, our overhead is only
√ polylogarithmic, so it can be
applied to subpolynomial improvements such as the n3 / exp(Θ( log n)) time algorithm for the
Negative-Weight Triangle problem due to Williams (STOC 2014).
1 Introduction
1.1 Approximate counting and decision in coarse-grained settings
It is clearly at least as hard to count objects as it is to decide their existence, and very often it is
harder. For example, Valiant [21] defined the class #P as the natural counting variant of NP, and
∗
Part of this work was done while the authors were visiting the Simons Institute for the Theory of Computing. The
research leading to these results has received funding from the European Research Council under the European
Union’s Seventh Framework Programme (FP7/2007-2013) ERC grant agreement no. 334828. The paper reflects
only the authors’ views and not the views of the ERC or the European Commission. The European Union is not
liable for any use that may be made of the information contained therein.
1
Toda [19] proved that P#P contains the entire polynomial hierarchy. The decision counterparts of
many #P-complete problems are in P, for example, counting perfect matchings is #P-complete and
detecting one is in P.
However, the situation changes substantially if we consider approximate counting rather than
exact counting. For all real ε with 0 < ε < 1, we say that x ∈ R is an ε-approximation to N ∈ R if
|x − N | ≤ εN holds. Clearly, computing an ε-approximation to N is at least as hard as deciding
whether N > 0 holds, but surprisingly, in many settings it is no harder. Indeed, Sipser [16] and
Stockmeyer [17] proved implicitly that every problem in #P has a polynomial-time randomised
ε-approximation algorithm using an NP-oracle; the result is later proved explicitly in Valiant and
Vazirani [22]. This is a foundational result in the wider complexity theory of polynomial approximate
counting initiated by Dyer, Goldberg, Greenhill and Jerrum [5].
Another example arises in parameterised complexity. Here, the usual goal is to determine whether
an instance of size n with parameter k can be solved in “FPT time” f (k) · poly(n) for some
computable function f : N → N. Hardness results are normally presented using the W -hierarchy
(see for example [6]). Müller [14] has proved that for any problem in #W[1], there is a randomised
ε-approximation algorithm using a W[1]-oracle which runs on size-n parameter-k instances in time
f (k) · poly(n, ε−1 ) for some computable f : N → N. He also proves analogous results for the rest of
the hierarchy.
We consider the subexponential-time setting. The popular (randomised) exponential time
hypothesis (ETH) introduced by Impagliazzo and Paturi [9] asserts that there exists δ > 0 such
that no randomised algorithm can solve an n-variable instance of 3-SAT in time O(2δn ).1 We prove
that ETH is equivalent to a seemingly weaker approximate counting version.
Theorem 1. ETH is true if and only if there exists δ > 0 such that no randomised ε-approximation
algorithm can run on n-variable instances of #3-SAT in time ε−2 · O(2δn ).
Note that the usual argument of Valiant and Vazirani [22] does not apply in this setting without
modification, as it adds clauses of linear width to the instance. Our proof will take a similar
approach, but making use of a sparse hashing technique due to Calabro, Impagliazzo, Kabanets and
Paturi [2]. (We give more detail in Section 1.4.)
1.2 Fine-grained results in the exponential setting
In fine-grained complexity, we are concerned not only with the classification of algorithms into broad
categories such as polynomial, FPT, or subexponential, but with their more precise running times.
A more fine-grained analogue of ETH is known as the strong exponential time hypothesis (SETH,
see Impagliazzo, Paturi and Zane [10]), which asserts that for all δ > 0, there exists k ∈ N such that
no randomised algorithm can solve n-variable instances of k-SAT in time O(2(1−δ)n ).
An analogue of Theorem 1 for SETH is implicit in Thurley [18], who provides a randomised
approximate counting algorithm that makes use of a decision oracle: SETH is true if and only if
for all δ > 0, there exists k ∈ N such that no randomised ε-approximation algorithm can run on
n-variable instances of k-SAT in time O(2(1−δ)n ). However, this result is not ideal from a fine-grained
perspective, as it does not guarantee that solving k-SAT and approximating #k-SAT have similar
time complexities in the limit as k → ∞. Indeed, given an algorithm for k-SAT with running
time Θ(2(1−δ)n ), Thurley’s approximation algorithm has worst-case running time ε−2 · Θ∗ (2(1−δ/2)n ),
that is, the exponential savings over exhaustive search goes down from δ for decision to δ/2 using
1
To streamline our discussion, we ignore the detail that some papers only allow for deterministic algorithms.
Throughout the paper, we require randomised algorithms to have success probability at least 2/3 unless otherwise
specified.
2
Thurley’s algorithm. Traxler [20] proved that if we allow super-constant clause width Ω(log n)
instead of considering k-SAT, then the same savings can be achieved for approximate counting and
decision. We strengthen Traxler’s result so that it applies in the setting of SETH.
Theorem 2. Let δ > 0. Suppose that for all k ∈ N, there is a randomised algorithm which runs
on n-variable instances of k-SAT in time O(2(1−δ)n ). Then for all δ 0 > 0 and all k ∈ N, there is
a randomised ε-approximation algorithm which runs on n-variable instances of #k-SAT in time
0
ε−2 · O(2(1−δ+δ )n ).
In particular, if SETH is false, then Theorem 2 yields an efficient algorithm for approximating
#k-SAT for sufficiently large k. Note that there is no particular reason to believe that an efficient
decision algorithm would yield an efficient counting algorithm directly. Indeed, when k = 3, the most
efficient known algorithms run in time O(1.308n ) for decision (due to Hertli [7]), in time O(1.537n )
for 12 -approximate counting (due to Thurley [18]), and in time O(1.642n ) for exact counting (due to
Kutzkov [12]).
It remains an open and interesting question whether a result analogous to Theorem 2 holds for
fixed k, that is, whether deciding k-SAT and approximating #k-SAT have the same time complexity
up to a subexponential factor. For (large) fixed k, the best-known decision, 12 -approximate counting,
and exact counting algorithms (due to Paturi, Pudlák, Saks, and Zane [15], Thurley [18], and
Impagliazzo, Matthews, and Paturi [8], respectively) all have running time 2(1−Θ(1/k))n , but with
2
progressively worse constants in the exponent. Our reduction has an overhead of 2Ω((log k) /k)n , so
we do not get improved approximate counting algorithms for fixed k.
1.3 Fine-grained results in the polynomial-time setting
Alongside SAT, perhaps the most important problems in fine-grained complexity are 3SUM, Orthogonal Vectors (OV), and All-Pairs Shortest Paths (APSP). All three problems admit well-studied
notions of hardness, in the sense that many problems reduce to them or are equivalent to them
under fine-grained reductions, and they are not known to reduce to one another. See Williams [26]
for a recent survey. We prove analogues of Theorem 2 for 3SUM and OV. While it is not clear
what a “canonical” counting version of APSP should be, we nevertheless prove an analogue of
Theorem 2 for the Negative-Weight Triangle problem (NWT) that is equivalent to APSP under
subcubic reductions. Our results, together with previous decision algorithms, immediately imply
three new approximate counting algorithms. (However, we believe that two of these, Theorems 4
and 8, may also be derived more directly by modifying their known decision algorithms.)
3SUM asks, given three integer lists A, B and C of total length n, whether there exists a tuple
(a, b, c) ∈ A × B × C with a + b = c. (Frequently the input is taken to be a single list rather than
three; it is well-known that the two versions are equivalent.) It is easy to see that 3SUM can be
solved in Õ(n2 ) operations by sorting C and iterating over all pairs in A × B, and it is conjectured
that for all δ > 0, 3SUM admits no O(n2−δ )-time randomised algorithm. We obtain an analogue
of Theorem 2 for the natural counting version of 3SUM, in which we approximate the number of
tuples (a, b, c) with a + b = c. (See Section 2 for details on our model of computation.)
Theorem 3. If 3SUM with n integers from [−N, N ] has a randomised T (n, N )-time algorithm, then
there is an (ε−2 T (n, N ) · O((log n)6 ) + ε−4 log N · Õ(n))-time randomised ε-approximation algorithm
for #3SUM.
Note that normally N is assumed to be at most n3 , in which case our algorithm has only
polylogarithmic overhead over decision. Thus independently of whether or not the 3SUM conjecture
3
is true, we can conclude that 3SUM and, say, 12 -approximating #3SUM have essentially the same
time complexity.
Chan and Lewenstein [3] have proved that the 3SUM conjecture fails when the problem is restricted
to instances in which elements of one list are somewhat clustered. This is an interesting special
case with several applications, including monotone multi-dimensional 3SUM with linearly-bounded
coordinates — see the introduction of [3] for an overview. By Chan and Lewenstein’s algorithm
combined with an analogue of Theorem 3, we obtain the following result.
Theorem 4. For all δ > 0, there is a randomised ε-approximation algorithm with running time
ε−4 · Õ(n2−δ/7 ) for instances of #3SUM with n integers in [−n3 , n3 ] such that at least one of A, B,
or C may be covered by n1−δ intervals of length n.
We next consider OV, which asks, given two lists A and B of total length n of zero-one vectors
over Rd , whether there exists an orthogonal pair (a, b) ∈ A × B. It is easy to see that OV can be
solved in O(n2 d) operations by iterating over all pairs, and it is conjectured that for all δ > 0, when
d = ω(log n), OV admits no O(n2−δ )-time randomised algorithm. This conjecture is implied by
SETH [23], and Abboud, Williams and Yu [1] proved that it fails when d = O(log n). We obtain
an analogue of Theorem 2 for the natural counting version of OV, in which we approximate the
number of orthogonal pairs.
Theorem 5. If OV with n vectors in d dimensions has a randomised T (n, d)-time algorithm,
then there is a randomised (ε−2 T (n, d) · O((log n)6 ) + ε−4 d · Õ(n))-time ε-approximation algorithm
for #OV.
Note that it is impossible to decide OV in time o(n), so when d is polylogarithmic (as is the usual
assumption), our algorithm has only polylogarithmic overhead over decision. Thus our result is able
to turn the n2−O(log n/d) -time algorithm of [1] into an approximate counting algorithm, but Chan
and Williams [4] already gave a deterministic exact counting algorithm of similar complexity.
A version of OV in which the real zero-one vectors are replaced by arbitrary vectors over finite fields
or rings is also studied, and there are efficient randomised algorithms due to Williams and Yu [25].
Their algorithms do not immediately generalise to counting, but by an analogue of Theorem 5, we
nevertheless obtain efficient approximate counting algorithms.
Theorem 6. Let m = pk be a constant prime power. Then there is a randomised ε-approximation
algorithm for n-vector instances of #OV over GF(m)d (resp. (Z/mZ)d ) in time ε−4 d(p−1)k · Õ(n)
(resp. ε−4 dm−1 · Õ(n)).
Note that the dependence on d may be close to best possible; under SETH, for all δ > 0 and
f : N → N with f (x) = o(x/ log x), OV over GF(m)d (resp. (Z/mZ)d ) cannot be solved in time
n2−δ df ((p−1)k) (resp. n2−δ df (m) ) for all but finitely many values of m = pk . [25]
Finally, we study the Negative-Weight Triangle problem (NWT) of deciding whether an edgeweighted tripartite graph contains a triangle of negative total weight, which Williams and Williams [27]
have shown is equivalent to APSP under fine-grained reductions. It is easy to see that NWT can be
solved in Õ(n3 ) operations by checking every possible triangle, and it is conjectured that for all δ > 0,
NWT admits no O(n3−δ )-time randomised algorithm. We obtain an analogue of Theorem 2 for the
natural counting version of NWT, in which we approximate the number of negative-weight triangles.
Theorem 7. If NWT for n-vertex graphs with weights from [−W, W ] has a randomised T (n, W )time algorithm, then there is a randomised (ε−2 T (n, W ) · O((log n)6 ) + ε−4 log W · Õ(n2 ))-time
ε-approximation algorithm for #NWT.
4
Note that it is impossible to decide NWT in time o(n2 ), so when W is polynomially bounded, our
algorithm has only polylogarithmic overhead over decision. Note also that [27] provides a subcubic
reduction from listing negative-weight triangles to NWT, although it has polynomial overhead and
so does not imply our result. Together with an algorithm of Williams [24], Theorem 7 implies the
following.
Theorem 8. There is a randomised ε-approximation√which runs on n-vertex instances of #NWT
with polynomially bounded weights in time ε−4 n3 /eΩ( log n) .
1.4 Techniques
We first discuss Theorems 1 and 2, which we prove in Section 3. In the polynomial setting,
the standard reduction from approximating #k-SAT to deciding k-SAT is due to Valiant and
Vazirani [22], and runs as follows. If a k-CNF formula F has at most 2δn solutions for some δ > 0,
then using a standard self-reducibility argument, one can count the number of solutions with O∗ (2δn )
calls to a k-SAT-oracle. Otherwise, for any m ∈ N, one may form a new formula Fm by conjoining F
with m uniformly random XOR clauses. It is relatively easy to see that as long as the number SAT(F )
of satisfying assignments of F is substantially greater than 2m , then SAT(Fm ) is concentrated
around 2−m SAT(F ). Thus by choosing m appropriately, one can count SAT(Fm ) exactly, then
multiply it by 2m to obtain an estimate for SAT(F ).
Unfortunately, this argument requires modification in the exponential setting. If F has n
variables, then each uniformly random XOR has length Θ(n) and therefore cannot be expressed
as a width-k CNF without introducing Ω(n) new variables. It follows that (for example) Fbn/2c
will contain Θ(n2 ) variables. This blowup is acceptable in a polynomial setting, but not an
2/3
exponential one — for example, given a Θ(2n )-time algorithm for k-SAT, it would yield a useless
4/3
Θ(2n )-time randomised approximate counting algorithm for #k-SAT. We can afford to add only
constant-length XORs, which do not in general result in concentration in the number of solutions.
We therefore make use of a hashing scheme developed by Calabro, Impagliazzo, Kabanets, and
Paturi [2] for a related problem, that of reducing k-SAT to Unique-k-SAT. They choose a 2s-sized
subset of [n] uniformly at random, where s is a large constant, then choose variables binomially at
random within that set. This still does not yield concentration in the number of solutions of Fm ,
but it turns out that the variance is sufficiently low that we can remedy this by summing over many
slightly stronger independently-chosen hashes.
Our results in Section 1.3 follow from a more general theorem, in which we consider the problem
of approximately counting edges in an arbitrary bipartite graph to which we have only limited
access. In particular, we only allow adjacency queries and independence queries: An adjacency
query checks whether an edge exists between two given vertices of the graph, and an independence
query checks whether a given set of vertices is an independent set in the graph. The standard
approach (as used by Thurley [18]) would be to use random adjacency queries to handle instances
with many edges, and independence queries and self-reducibility to handle instances with few edges.
This approach requires polynomially many independence queries, which is too many to establish the
tight relationship between approximate counting and decision required for the results of Section 1.3.
In contrast, our main algorithm (Theorem 10) approximates the number of edges in such a graph in
quasi-linear time and makes only poly-logarithmically many independence queries.
Using this algorithm, we obtain the results for polynomial-time problems in a straightforward
way. For example, in the proof of Theorem 5 for OV, the vertices of G are the input vectors and the
edges correspond to orthogonal pairs. An adjacency query corresponds to an orthogonality check,
which can be done in time Õ(d) in d dimensions, and an independence query corresponds to a call
5
to the decision oracle for OV on a sub-instance, which takes time T (n, d) by assumption. Since only
poly-logarithmically independence queries occur, Theorem 5 follows.
Our algorithm for Theorem 10 works roughly as follows. Let G be the bipartite graph whose
edges we are trying to count, and let U and V be the vertex classes of G. Using binary search
together with our independence oracle, we can quickly find non-isolated vertices of G. If G contains
few such vertices, then by the standard self-reducibility argument used in Theorems 1 and 2, we can
quickly determine |E(G)| exactly, so suppose G contains many such vertices. If every vertex of G
is contained in only a small proportion of its edges, then we can approximately halve |E(G)| by
passing to a uniformly random subset of V , and proceed similarly to Valiant and Vazirani. However,
in general this will not be the case and the number of edges in the resulting graph will not be
concentrated. We must therefore detect and remove problematic vertices as we go. The procedure
we use for this is quite technical, and forms the bulk of the proof, so we defer further explanation to
Section 4.
2 Preliminaries
We write N for the set of all positive integers. For a positive integer n, we use [n] to denote the set
{1, . . . , n}. We use log to denote the base-e logarithm, and lg to denote the base-2 logarithm.
2.1 Notation
We write N for the set of all positive integers. For a positive integer n, we use [n] to denote the set
{1, . . . , n}. We use log to denote the base-e logarithm, and lg to denote the base-2 logarithm.
We consider graphs G to be undirected, and write e(G) = |E(G)|. For all v ∈ V (G), we use
N (v) to denote the neighbourhood {w ∈ V (G) : {v, w} ∈ E(G)} of v. For convenience, we shall
generally present bipartite graphs G as a triple (U, V, E) in which (U, V ) is a partition of V (G) and
E ⊆U ×V.
When stating quantitative bounds on running times of algorithms, we assume the standard
word-RAM machine model with logarithmic-sized words. We assume that lists and functions in
the problem input are presented in the natural way, that is, as an array using at least one word
per entry. In general, we shall not be overly concerned with logarithmic factors in running times.
We shall write f (x) = Õ(g(x)) when for some constant c ∈ R, f (x) = O((log x)c g(x)) as x → ∞.
Similarly, we write f (x) = O∗ (g(x)) when for some constant c ∈ R, f (x) = O(xc g(x)) as x → ∞.
We require our problem inputs to be given as finite binary strings, and write Σ∗ for the set of
all such strings. A randomised approximation scheme for a function f : Σ∗ → N is a randomised
algorithm that takes as input an instance x ∈ Σ∗ and a rational error tolerance 0 < ε < 1, and
outputs a rational number z (a random variable depending on the “coin tosses” made by the
algorithm) such that, for every instance x, P((1 − ε)f (x) ≤ z ≤ (1 + ε)f (x)) ≥ 2/3. All of our
approximate counting algorithms will be randomised approximation schemes.
2.2 Probability theory
We will require some well-known results from probability theory, which we collate here for reference.
First, we state Chebyshev’s inequality.
Lemma 1. Let X be a real-valued random variable with mean µ and let t > 0. Then
P |X − µ| ≥ t ≤
6
Var(X)
.
t2
We also use the following concentration result due to McDiarmid [13].
Lemma 2. Suppose f is a real function of independent random variables X1 , . . . , Xm , and let
µ = E(f (X1 , . . . , Xm )). Suppose there exist c1 , . . . , cm ≥ 0 such that for all i ∈ [m] and all pairs
(x, x0 ) differing only in the ith coordinate, |f (x) − f (x0 )| ≤ ci . Then for all t > 0,
2/
P(|f (X1 , . . . , Xm ) − µ| ≥ t) ≤ 2e−2t
Pm
c2
i=1 i
.
Finally, we use the following Chernoff bounds, proved in (for example) Corollaries 2.3-2.4 and
Remark 2.11 of Janson, Łuczak and Rucinski [11].
Lemma 3. Suppose X is a binomial or hypergeometric random variable with mean µ. Then:
2 µ/3
(i) for all 0 < ε ≤ 3/2, P(|X − µ| ≥ εµ) ≤ 2e−ε
;
(ii) for all t ≥ 7µ, P(X ≥ t) ≤ e−t .
3 From decision to approximate counting CNF-SAT
In this section we prove our results for the satisfiability of CNF formulae, formally defined as follows.
Problem: k-SAT
Input: A k-CNF formula F .
Task: Decide if F is satisfiable.
Problem: #k-SAT.
Input: A k-CNF formula F .
Task: Compute the number SAT(F ) of satisfying assignments of F .
We also define a technical intermediate problem. For all s ∈ N, we say that a matrix A is s-sparse
if every row of A contains at most s non-zero entries. In the following definition, k ∈ N and s ∈ N
are constants.
Problem: Πk,s .
Input: An n-variable Boolean formula F of the form F (x) = F 0 (x) ∧ (Ax = b). Here F 0 is a
k-CNF formula, A is an s-sparse m×n matrix over GF(2) for some 0 ≤ m ≤ n, and b ∈ GF(2)m .
Task: Decide if F is satisfiable.
We define the growth rate πk,s of Πk,s as the infimum over all β > 0 such that Πk,s has a
randomised algorithm that runs in time O∗ (2βn ) and outputs the correct answer with probability at
least 2/3. Our main reduction is encapsulated in the following theorem.
Theorem 9. Let k ∈ N with k ≥ 2, let 0 < δ < 1, and suppose s ≥ 120 lg(6/δ)2 /δ. Then there is a
randomised approximation scheme for #k-SAT which, when given
an n-variable formula F and
approximation error parameter ε, runs in time ε−2 · O 2(πk,s +δ)n .
Before we prove this theorem, let us derive Theorems 1 and 2 as immediate corollaries.
Theorem 1 (restated). ETH is true if and only if there exists δ > 0 such that no randomised
ε-approximation algorithm can run on n-variable instances of #3-SAT in time ε−2 · O(2δn ).
Proof. First note that we may use any randomised approximation scheme for #3-SAT to decide
3-SAT with success probability at least 2/3 by taking ε = 1/2 and outputting ‘yes’ if and only
if the result is non-zero. Thus the backward implication of Theorem 1 is immediate. Conversely,
7
suppose ETH is false. A well-known result of Impagliazzo, Paturi and Zane [10, Lemma 10] then
implies that for all constant k ∈ N and δ > 0, there is a randomised algorithm which can decide
k-SAT in time O(2δn ) with success probability at least 2/3. Hence for all constant k, s ∈ N, by the
natural reduction from Πk,s to (max{k, s})-SAT, we obtain πk,s = 0. The result therefore follows
by Theorem 9.
Theorem 2 (restated). Let δ > 0. Suppose that for all k ∈ N, there is a randomised algorithm
which runs on n-variable instances of k-SAT in time O(2(1−δ)n ). Then for all δ 0 > 0 and all k ∈ N,
there is a randomised ε-approximation algorithm which runs on n-variable instances of #k-SAT in
0
time ε−2 · O(2(1−δ+δ )n ).
Proof. Suppose that δ > 0 is as specified in the theorem statement. Then for all constant k, s ∈ N,
by the natural reduction from Πk,s to (max{k, s})-SAT, we have πk,s ≤ 1 − δ. Thus the result again
follows by Theorem 9.
3.1 Proof of Theorem 9
Given access to an oracle that decides satisfiability queries, we can compute the exact number of
solutions of a formula with few solutions using a standard self-reducibility argument given below
(see also [18, Lemma 3.2]).
Algorithm Sparse(F, a). Given an instance F of Πk,s on n variables, a ∈ N, and access to an
oracle for Πk,s , this algorithm computes SAT(F ) if SAT(F ) ≤ a; otherwise it outputs FAIL.
(S1) (Query the oracle) If F is unsatisfiable, return 0.
(S2) (No variables left) If F contains no variables, return 1.
(S3) (Branch and recurse) Let F0 and F1 be the formulae obtained from F by setting the first
free variable in F to 0 and 1, respectively. If Sparse(F0 , a) + Sparse(F1 , a) is at most a, then
return this sum; otherwise abort the entire computation and return FAIL.
Lemma 4. Sparse is correct and runs in time at most (min{a, SAT(F )} + 1) · Õ(|F |) with at most
8n · (min{a, SAT(F )} + 1) calls to the oracle. Moreover, each oracle query is a formula with at most
n variables.
Proof. Consider the recursion tree of Sparse on inputs F and a. At each vertex, the algorithm
takes time at most Õ(|F |) to compute F0 and F1 , and it issues a single oracle call. For convenience,
we call the leaves of the tree at which Sparse returns 0 (in (S1)) or 1 (in (S2)) the 0-leaves and
1-leaves, respectively.
Let x be the number of 1-leaves. Each non-leaf is on the path from some 1-leaf to the root,
otherwise it would be a 0-leaf. There are at most x such paths, so there are at most nx non-leaf
vertices in total. Finally, every 0-leaf has a sibling which is not a 0-leaf, or its parent would be a
0-leaf, so there are at most (n + 1)x 0-leaves in total. Overall, the tree has at most 4nx vertices. An
easy induction using (S3) implies that x ≤ 2a, and certainly x ≤ SAT(F ), so the running time and
oracle access bounds are satisfied. Correctness likewise follows by a straightforward induction.
When our input formula F has too many solutions to apply Sparse efficiently, we first reduce
the number of solutions by hashing. In particular, we use the same hash functions as Calabro et
al. [2]; they are based on random sparse matrices over GF(2) and formally defined as follows:
8
Definition 5. Let s, m, n ∈ N. An (s, m, n)-hash is a random m × n matrix A over GF(2) defined
as follows. For each row i ∈ [m], let Ri be a uniformly random size-s subset of [n]. Then for all
i ∈ [m] and all j ∈ Ri , we choose values Ai,j ∈ GF(2) independently and uniformly at random, and
set all other entries of A to zero.
For intuition, suppose F is an n-variable k-CNF formula and S is the set of satisfying assignments
of F , and that |S| > 2δn holds for some small δ > 0. It is easy to see that for all m, s ∈ N
and uniformly random b ∈ GF(2)m , if A is an (s, m, n)-hash, then the number X of satisfying
assignments of F (x) ∧ (Ax = b) has expected value |S|/2m . (See Lemma 6.) If X were concentrated
around its expectation, then by choosing an appropriate value of m, we could reduce the number of
solutions to at most 2δn , apply Sparse to count them exactly, then multiply the result by 2m to
obtain an approximation to |S|. This is the usual approach pioneered by Valiant and Vazirani [22].
In the exponential setting, however, we can only afford to take s = O(1), which means that
X is not in general concentrated around its expectation. In [2], only very limited concentration
was needed, but we require strong concentration. To achieve this, rather than counting satisfying
assignments of a single formula F (x) ∧ (Ax = b), we will sum over many such formulae. We first
bound the variance of an individual (s, m, n)-hash when s and S are suitably large. Our analysis
here is similar to that of Calabro et al. [2], although they are concerned with lower-bounding the
probability that at least one solution remains after hashing and do not give bounds on variance.
Lemma 6. Let 0 < δ < 1/6 and let s, m, n ∈ N. Suppose m ≤ n and s ≥ 20 lg(1/δ)2 /δ. Let
S ⊆ GF(2)n and suppose |S| ≥ 2m+δn . Let A be an (s, m, n)-hash, and let b ∈ GF(2)m be
uniformly random and independent of A. Let S 0 = {x ∈ S : Ax = b}. Then E(|S 0 |) = 2−m |S| and
Var(|S 0 |) ≤ |S|2 2δn/16−2m .
Proof. For each x ∈ GF(2), let Ix be the indicator variable of the event that Ax = b. Exposing A
implies that P(Ix ) = 2−m for all x ∈ GF(2)n , and hence
E(|S 0 |) =
X
P(Ix ) = 2−m |S|.
x∈S
We now bound the second moment. We have
E(|S 0 |2 ) =
X
E(Ix Iy ) =
(x,y)∈S 2
=
X
X
P(Ax = Ay = b)
(x,y)∈S 2
m
Y
P((Ax)i = (Ay)i = bi ).
(1)
(x,y)∈S 2 i=1
It will be convenient to partition the terms of this sum according to Hamming distance, which we
denote by d. Write h : [0, 1] → [0, 1] for the binary entropy function h(δ) = −δ lg δ − (1 − δ) lg(1 − δ),
write h−1 : [0, 1] → [0, 1/2] for its left inverse, and let α = h−1 (δ). Then it follows immediately
from (1) that
E(|S 0 |2 ) =
X
m
Y
X
P((Ax)i = (Ay)i = bi ) +
(x,y)∈S 2 i=1
d(x,y)≤αn
m
Y
P((Ax)i = (Ay)i = bi ).
(2)
(x,y)∈S 2 i=1
d(x,y)>αn
Denote the projection of any vector x ∈ GF(2)n onto GF(2)Ri by xRi . For any x, y ∈ S and any
i ∈ [m] we have
P((Ax)i = (Ay)i = bi ) = P((Ax)i = (Ay)i ) · P((Ay)i = bi | (Ax)i = (Ay)i )
= P((Ax)i = (Ay)i )/2.
9
Since (Ax)i = (Ay)i whenever xRi = yRi , it follows that
P((Ax)i = (Ay)i = bi ) = 1 − P((A(x − y))i 6= 0 | xRi 6= yRi ) · P(xRi 6= yRi ) /2.
Since {j ∈ Ri : (x − y)j = 1} has an equal number of odd- and even-sized subsets, on exposing Ri
it follows that
P(xRi 6= yRi )
1
1−
P((Ax)i = (Ay)i = bi ) =
2
2
=
1 + P(xRi = yRi )
for all x, y ∈ S.
4
(3)
In particular, this implies P((Ax)i = (Ay)i = bi ) ≤ 1/2. Since a ball of Hamming radius αn in
GF(2)n contains at most 2h(α)n = 2δn points, it follows that
X
m
Y
P((Ax)i = (Ay)i = bi ) ≤ |S|2δn · 2−m = E(|S 0 |)2
(x,y)∈S 2 i=1
d(x,y)≤αn
2m+δn
≤ E(|S 0 |)2 .
|S|
(4)
Now suppose d(x, y) ≥ αn. Since |Ri | = s by definition,
P(xRi = yRi ) ≤
n−dαne
s
n
s
≤ (1 − dαne /n)s ≤ (1 − α)s ≤ e−αs .
Hence by (3), we have P((Ax)i = (Ay)i = bi ) ≤ (1 + e−αs )/4. It follows that
X
m
Y
(x,y)∈S 2
i=1
P((Ax)i = (Ay)i = bi ) ≤
|S|2 me−αs
|S|2
−αs m
(1
+
e
)
≤
e
.
22m
22m
(5)
d(x,y)>αn
Combining (2), (4) and (5), we obtain
Var(|S 0 |) = E(|S 0 |2 ) − E(|S 0 |)2 ≤
|S|2 me−αs
e
.
22m
Since δ < 1/6, we have α = h−1 (δ) ≥ δ/(2 lg(6/δ)) ≥ δ/(4 lg(1/δ)). It follows that αs ≥ 5 lg(1/δ) ≥
2
2 ln(4/δ), and so Var(|S 0 |) ≤ |S|2 eδ m/16 /22m . Since m ≤ n and δ < 1/ lg(e), the result follows.
We now state the algorithm we will use to prove Theorem 9, then use the lemmas above to prove
correctness. In the following definition, δ is a rational constant with 0 < δ < 1/3.
Algorithm SATSolveδ (#k-SAT). Given an n-variable instance F of #k-SAT, a rational number ε ∈ (0, 1), and access to an oracle for Πk,s for some s ≥ 40 lg(2/δ)2 /δ, this algorithm computes
a rational number z such that with probability at least 3/4, (1 − ε)SAT(F ) ≤ z ≤ (1 + ε)SAT(F ).
(A1) (Brute-force on constant-size instances) If n/ lg n ≤ 8/δ, solve the problem by brute force
and return the result.
(A2) (If there are few satisfying assignments, count them exactly) Let t = dδn/2 + 2 lg(1/ε)e,
and apply Sparse to F and a = 2t+δn/2 . Return the result if it is not equal to FAIL.
(A3) (Try larger and larger equation systems) For each m ∈ {0, . . . , n − t}:
(A3a) (Set maximum number of solutions to find explicitly) Let a = 2t+δn/2+2 .
(A3b) For each i ∈ {1, . . . , 2t }:
10
• (Prepare query) Independently sample an (s, m + t, n)-hash Am,i and a uniformly
random vector bm,i ∈ GF(2)m+t . Let Fm,i = F (x) ∧ (Am,i x = bm,i ).
• (Ask oracle using subroutine) Let zm,i be the output of Sparse(Fm,i , a).
• (Bad hash or m too small) If zm,i = FAIL, go to next m in the outer for-loop.
• Otherwise, set a := a − zm,i .
(A3c) (Return our estimate) Return 2m
P2t
i=1 zm,i .
(A4) (We failed, return garbage) Return −1.
Lemma 7. SATSolveδ is correct for all δ ∈ (0, 13 ) and runs in time at most ε−2 ·O∗ (2δn ). Moreover,
the oracle is only called on instances with at most n variables.
Proof. Let F be an instance of #k-SAT on n variables, and let ε ∈ (0, 1). The running time of
the algorithm is dominated by (A2) and (A3b). Clearly (A2) takes time at most O∗ (2t+δn/2 ) by
Lemma 4. In the inner for-loop, the number a controls the maximum running time we are willing to
spend. In particular, again by Lemma 4, the running time for one iteration of the inner for-loop is
O∗ (zm,i ) if zm,i 6= FAIL and otherwise it is bounded by O∗ (a + 1) but the remaining iterations of
P
the inner loop are then skipped. It is easy to see that ij=1 zm,j ≤ 2t+δn/2+2 holds at any point of
the inner loop, and hence the overall running time is O∗ (2t+δn/2 ) as required. Likewise, the oracle
access requirements are satisfied, so it remains to prove the correctness of SATSolveδ .
If SATSolveδ terminates at (A1) or (A2), then correctness is immediate. Suppose not, so that
n/ lg n > 8/δ holds, and the set S of solutions of F satisfies |S| ≥ 2t+δn/2 . Let M = max{m ∈ Z :
|S| ≥ 2m+t+δn/2 }, and note that 0 ≤ M ≤ n − t and |S| ≤ 2M +t+δn/2+1 . The formulas Fm,i are
oblivious to the execution of the algorithm, so for the analysis we may view them as being sampled
in advance. Let Sm,i be the set of solutions to Fm,i . For each m with 0 ≤ m ≤ M , let Em be the
following event:
t
2
X
|Sm,i | − 2−m |S| ≤ 2−m−(t−δn/2)/2 · |S| .
i=1
Thus Em implies 2
m P2t
i=1 |Sm,i | − |S| ≤ ε|S|. By
≤ i ≤ 2t we have E(|Sm,i |)
Lemma 6 applied to δ/2, s, m + t, n and S, for
all 0 ≤ m ≤ M and 1
= 2−m−t |S| and Var(|Sm,i |) ≤ |S|2 2δn/32−2m−2t .
Since the Sm,i ’s are independent, it follows by Lemma 1 that
P(Em ) ≥ 1 −
2t · |S|2 2δn/32−2m−2t
≥ 1 − 2−δn/4 ≥ 1 − 1/n2 .
2−2m−t+δn/2 |S|2
Thus a union bound implies that, with probability at least 3/4, the event Em occurs for all m
with 0 ≤ m ≤ M simultaneously. Suppose now that this happens. Then in particular, we
P t
have 2i=1 |SM,i | ≤ (1 + ε)2−M |S| ≤ 2t+δn/2+2 . But then, if SATSolveδ reaches iteration m = M ,
none of the calls to Sparse fail in this iteration and we have zM,i = |SM,i | for all i ∈ {1, . . . , 2t }.
Thus SATSolveδ returns some estimate z in (A3c) while m ≤ M . Moreover, since Em occurs, this
estimate satisfies (1 − ε)|S| ≤ z ≤ (1 + ε)|S| as required. Thus SATSolveδ behaves correctly with
probability at least 3/4, and the result follows.
Theorem 9 (restated). Let k ∈ N with k ≥ 2, let 0 < δ < 1, and suppose s ≥ 120 lg(6/δ)2 /δ.
Then there is a randomised approximation scheme for #k-SAT which, when given
an n-variable
formula F and approximation error parameter ε, runs in time ε−2 · O 2(πk,s +δ)n .
11
Proof. If ε < 2−n , then we solve the #k-SAT instance exactly by brute force in time O∗ (ε−1 ), so
suppose ε ≥ 2−n . By the definition of πk,s , there exists a randomised algorithm for Πk,s with failure
probability at most 1/3 and running time at most O∗ (2(πk,s +δ/3)n ). By Lemma 3(i), for any constant
C, by applying this algorithm lg(1/ε) · O(n) = O(n2 ) times and outputting the most common
result, we may reduce the failure probability to at most ε2 /Cn2δn/3 . We apply SATSolveδ/3 to
F and ε, using this procedure in place of the Πk,s -oracle. If we take C sufficiently large, then by
Lemma 7 and a union bound, the overall failure probability is at most 1/3, and the running time is
ε−2 · O∗ (2(πk,s +2δ/3)n ) = ε−2 · O(2(πk,s +δ)n ) as required.
4 General fine-grained result
We first define the setting of our result.
Definition 8. Let G = (U, V, E) be a bipartite graph. We define the independence oracle of G to
be the function indG : 2U ∪V → {0, 1} such that indG (S) = 1 if and only if S is an independent
set in G. We define the adjacency oracle of G to be the function adjG : U × V → {0, 1} such that
adjG (u, v) = 1 if and only if (u, v) ∈ E.
We think of edges of G as corresponding to witnesses of a decision problem. For example, in OV,
they will correspond to pairs of orthogonal vectors. Thus calling adjG will correspond to verifying a
potential witness, and calling indG will correspond to solving the decision problem on a sub-instance.
Our main result will be the following.
Theorem 10. There is a randomised algorithm A with the following properties:
(i) A is given as input two disjoint sets U and V and a rational number 0 < ε < 1;
(ii) for some bipartite graph G = (U, V, E) with |V (G)| = n, A has access to the independence and
adjacency oracles of G;
(iii) A returns a rational number x such that (1 − ε) · e(G) ≤ x ≤ (1 + ε) · e(G) holds with probability
at least 23 ;
(iv) A runs in time at most ε−4 · Õ(n);
(v) A makes at most ε−2 · O((log n)6 ) calls to the independence oracle.
Throughout the rest of the section, we take G = (U, V, E) to be the bipartite graph of the theorem
statement and n = |U ∪ V |. Moreover, for all X ⊆ V , we define ∂(X) = |E ∩ (U × X)| and
UX = {u ∈ U : (u, v) ∈ E for some v ∈ X}.
We briefly compare the performance of the algorithm of Theorem 10 with that of the standard
approach of sampling to deal with dense instances combined with brute force counting to deal
with sparse instances (as used in Thurley [18]). Suppose ε is constant, that we can evaluate indG
in time O(n2−α ) for some α > 0, that we can evaluate adjG in time O(1), and that the input
graph contains nβ edges for some β > 0. Then sampling requires Ω(n2−β ) time, and brute force
enumeration of the sort used in Sparse (p. 8) requires Ω(n2−α+β ) time. The worst case arises
when β = α/2, in which case the algorithm requires Ω(n2−α/2 ) time. However, the algorithm of
Theorem 10 requires only Õ(n2−α ) time in all cases. Thus it has only polylogarithmic overhead
over deciding whether the graph contains edges at all.
Similarly to Section 3, we shall obtain our approximation by repeatedly (approximately) halving
the number of edges in the graph until few remain, then counting the remaining edges exactly. For
12
the halving step, rather than hashing, if our current graph is induced by U ∪ X for some X ⊆ V
then we shall simply delete half the vertices in X chosen uniformly at random. However, if any
single vertex in X is incident to a large proportion of the remaining edges, then the edge count of
the resulting graph will not be well-concentrated around its expectation and so this approach will
fail. We now prove that this is essentially the only obstacle.
Definition 9. Given 0 < ξ < 1, we say a non-empty set X ⊆ V is ξ-balanced if every vertex in X
has degree at most ξ∂(X).
Lemma 10. Let 0 < ξ < 1, suppose X ⊆ V is ξ-balanced, and suppose |X| ≥ 24 log n. Let X 0 ⊆ X
be a random subset formed by including each vertex of X independently with probability 12 . Then
with probability at least 1 − 4/n, we have |X 0 | ≤ 3|X|/4 and
p
1
∂(X 0 ) − ∂(X) ≤ ξ log n · ∂(X).
2
Proof. First note that since E(|X 0 |) = |X|/2, by Lemma 3(i) we have
P |X 0 | ≥
3|X|
4
≤P
|X 0 | −
|X|
|X|
≥
2
4
≤ 2e−|X|/24 ≤
2
.
n
(6)
√
Now let R = ξ log n. For each vertex v ∈ X, let iv be the indicator random variable of the event
that v ∈ X 0 . Note that ∂(X 0 ) is a function of {iv : v ∈ X}, and that changing a single indicator
variable iv alters ∂(X 0 ) by exactly d(v). Moreover, E(∂(X 0 )) = ∂(X)/2. It therefore follows by
Lemma 2 that
!
∂(X)
−2R2 ∂(X)2
0
P ∂(X ) −
≥ R∂(X) ≤ 2 exp P
.
(7)
2
2
v∈X d(v)
Let t = |X|, and define a function f : Rt → R by
f (x1 , . . . , xt ) =
t
X
x2i .
i=1
Suppose that (x1 , . . . , xt ) maximises f subject to the constraints 0 ≤ x1 , . . . , xt ≤ ξ∂(X) and
Pt
2
2
2
2
i=1 xi = ∂(X). For all 0 < c ≤ x ≤ y, we have ((x − c) + (y + c) ) − (x + y ) = 2c(−x + y + c) > 0,
so we must have xi ∈ {0, ξ∂(X)} for all but at most one value of i ∈ [t]. (Otherwise, we could increase
the value of f by perturbing x1 , . . . , xt .) Thus up to reordering, we have x1 = · · · = xb1/ξc = ξ∂(X),
xb1/ξc+1 = (1 − ξ b1/ξc)∂(X), and xb1/ξc+2 , . . . , xt = 0. Since X is ξ-balanced, it follows that
X
d(v)2 ≤
v∈X
2
1 2
1
ξ ∂(X)2 + 1 − ξ
ξ
ξ
∂(X)2
≤ (ξ + ξ 2 )∂(X)2 ≤ 2ξ∂(X)2 .
By (7), it follows that
P
∂(X 0 ) −
∂(X)
2
2
≥ R∂(X) ≤ 2e−R /ξ = .
2
n
(8)
The result therefore follows from (6), (8), and a union bound.
To make X balanced, we will need to find a reasonable approximation of the set of high-degree
vertices in X, as captured by the following definition.
13
Definition 11. Given 0 < ξ < 1 and ∅ ( X ⊆ V , we say S ⊆ X is a ξ-core of X if it satisfies the
following properties:
(W1) every vertex in X with degree at least ξ|UX | is contained in S;
(W2) every vertex in S has degree at least ξ|UX |/24.
If 1 ≤ |S| < 24/ξ 2 , then we call S a ξ-unbalancer; otherwise, if S = ∅ or |S| ≥ 24/ξ 2 , we call S a
ξ-witness.
To find a ξ-core of X, we use the following procedure. Note that if there are few non-isolated
vertices remaining, it will instead return ∂(X) exactly.
Algorithm FindCore. The algorithm takes as input a set ∅ ( X ⊆ V and 0 < ξ < 1. If
|X| < 24 log n or |UX | < 24 log n/ξ, it returns ∂(X). Otherwise, it returns a set S which with
probability at least 1 − 3/n is a ξ-core of X.
(C1) If |X| < 24 log n, then use adjG to enumerate the edges incident to X and return the total
number.
(C2) Let u1 , . . . , ut be a uniformly random ordering of U , and let x = d24 log n/ξe.
(C3) Recursively define a sequence k1 , . . . , kx by
n
o
ki = max k ∈ {0, . . . , t} : indG (V ∪ {u1 , . . . , uk } \ {uk1 +1 , . . . , uki−1 +1 }) = 1 .
Using binary search, find k1 , . . . , kx , and let Y = {uki +1 : i ∈ [x], ki < t}.
(C4) If kx = t, then use adjG to enumerate the edges incident to Y and return the total number.
(C5) Otherwise, use adjG to find the set S ⊆ X of vertices which are adjacent to at least ξx/2
vertices in Y . Return S.
Lemma 12. When n ≥ 2, FindCore is correct. Moreover, it runs in ξ −1 · Õ(n) time, makes at
most ξ −1 · O((log n)2 ) calls to the independence oracle, and makes at most ξ −1 · Õ(n) calls to the
adjacency oracle.
Proof. Let x = d24 log n/ξe. First note that (C1) requires O(n log n) time and at most 24n log n
calls to adjG ; (C2) requires O(n log n) time; (C3) requires x · O(log n) time and at most x lg n
calls to indG ; (C4) and (C5) together require x · O(n) time and at most xn calls to adjG . Thus
the running time guarantees in the statement are correct. It is also immediate that the algorithm
behaves correctly if |X| < 24 log n.
Note that when ki < t, we have
ki + 1 = min{k ∈ [t] : indG (V ∪ {u1 , . . . , uk } \ {uk1 +1 , . . . , uki−1 +1 }) = 0}
= min{k ∈ [t] \ {k1 + 1, . . . , ki−1 + 1} : uk ∈ UX } .
Thus Y consists of the least x elements of UX in the random ordering u1 , . . . , ut , or all of UX if
|UX | < x. Thus if |UX | < x then the algorithm behaves correctly, and if |UX | ≥ x then Y is a
uniformly random size-x subset of UX . Suppose |UX | ≥ x. Then it remains to prove that S is a
ξ-core with probability at least 1 − 3/n.
Let Z be the set of vertices in X with degree at least ξ|UX |. For all v ∈ Z, |N (v) ∩ Y | follows a
hypergeometric distribution with mean (d(v)/|UX |) · x ≥ ξx ≥ 24 log n. By Lemma 3(i), it follows
that
for all v ∈ Z, P(v ∈
/ S) = P(|N (v) ∩ Y | < ξx/2) ≤ 2e−ξx/12 ≤ 2/n2 .
(9)
14
Conversely, let Z 0 be the set of vertices in X with degree less than ξ|UX |/24. For all v ∈ Z 0 ,
|N (v) ∩ Y | follows a hypergeometric distribution with mean (d(v)/|UX |) · x < ξx/24. Lemma 3(ii)
then yields
for all v ∈ Z 0 , we have P(v ∈ S) = P |N (v) ∩ Y | ≥ ξx/2 ≤ e−ξx/2 ≤ e−12 log n < 1/n2 .
(10)
We obtain the required upper bound on the failure probability of the algorithm by (9), (10), and a
union bound over all v ∈ Z ∪ Z 0 .
The following lemma implies that finding a ξ-core S of X ⊆ V with ξ small will allow us to either
apply Lemma 10 to halve ∂(X) (possibly after removing S) or halve the size of |UX | by removing S.
Lemma 13. Let ξ ∈ (0, 1), let X be a set with ∅ ( X ⊆ V , and let S be a ξ-core of X.
(i) If S is a ξ-witness of X, then X is ξ-balanced.
(ii) If S is a ξ-unbalancer of X, then either X \ S contains no 48ξ-unbalancer or |UX\S | ≤ |UX |/2.
Proof. Suppose S is a ξ-witness of X, so that either S = ∅ or |S| ≥ 24/ξ 2 . If S = ∅, then by (W1)
every vertex in X has degree at most ξ|UX |. Since every vertex in UX is incident to an edge to X,
we have |UX | ≤ ∂(X) and so X is ξ-balanced. Suppose instead |S| ≥ 24/ξ 2 . Then by (W2), at
least 24/ξ 2 vertices in X have degree at least ξ|UX |/24. Hence ∂(X) ≥ |UX |/ξ. Since every vertex
v ∈ X satisfies d(v) ≤ |UX |, it follows that d(v) ≤ ξ∂(X) and so once again X is ξ-balanced. Thus (i)
holds.
Now suppose S is a ξ-unbalancer of X, suppose S 0 is a 48ξ-unbalancer of X \ S, and let v ∈ S 0 be
arbitrary. By (W2), we have d(v) ≥ 2ξ|UX\S |. Moreover, since v ∈
/ S, by (W1) we have d(v) ≤ ξ|UX |.
It follows that |UX | ≥ 2|UX\S |, so (ii) holds.
We now prove the main result.
Algorithm EdgeCount. Given sets U and V with |U ∪ V | = n and a rational ε > 0, return a
rational number x such that (1 − ε) · e(G) ≤ x ≤ (1 + ε) · e(G) holds with probability at least 23 .
(E1) Let ζ = ε2 /362 (log n)3 , let X = V , and let t = N = 0. If n < 3000, count e(G) exactly
using adjG and output it.
(E2) Apply FindCore to X and ζ/48. If FindCore returns ∂(X), output 2t ∂(X) + N . Otherwise, it returns a subset S of X.
(E3) If S = ∅ or |S| > 24 · (48/ζ)2 , then remove each element of X independently with probability
1/2, increment t, and go to (E2).
(E4) Using adjG , find ∂(S).
(E5) Apply FindCore to X \ S with argument ζ. If FindCore returns ∂(X \ S), output
2t ∂(X) + N . Otherwise, it returns a subset S 0 of X \ S.
(E6) If S 0 = ∅ or |S 0 | > 24/ζ 2 , then remove S from X, remove each element of X independently
with probability 1/2, add 2t ∂(S) to N , increment t, and go to (E2).
(E7) Otherwise, remove S from X, add 2t ∂(S) to N , and go to (E2).
Theorem 10. There is a randomised algorithm A with the following properties:
(i) A is given as input two disjoint sets U and V and a rational number 0 < ε < 1;
15
(ii) for some bipartite graph G = (U, V, E) with |V (G)| = n, A has access to the independence and
adjacency oracles of G;
(iii) A returns a rational number x such that (1 − ε) · e(G) ≤ x ≤ (1 + ε) · e(G) holds with probability
at least 23 ;
(iv) A runs in time at most ε−4 · Õ(n);
(v) A makes at most ε−2 · O((log n)6 ) calls to the independence oracle.
Proof. We take A to be EdgeCount, modified to return an arbitrary value if the running time
or oracle access exceeds that allowed by (iv)–(vi). Thus every desired property of A is immediate
except (iii).
Suppose without loss of generality that n ≥ 3000. For all integer i ≥ 1, let Xi be the value of X
at the start of the ith iteration of (E2)–(E7), or Xi = Xi−1 if the algorithm terminates before
this point. Define ti and Ni in the same fashion. Let Si be the output of FindCore at the ith
iteration of (E2), or ∅ if the algorithm terminates before this point. Let Ai be the event that
|Xi ||UXi | ≤ (3/4)i−1 n2 and
p
1 − 2 ζ log n
ti
e(G) ≤ 2ti ∂(Xi ) + Ni ≤ 1 + 2 ζ log n
p
ti
e(G),
(11)
and let Bi be the event that EdgeCount terminates before the start of the ith iteration of
(E2)–(E7). Let Ei = Ai ∪ Bi , and note that P(A1 ) = 1.
Let m = d7 log ne + 1, and suppose that E1 , . . . , Em all occur. We first claim Bm+1 occurs. If Bm
occurs this is immediate, so suppose instead that Am \ Bm occurs. Then we have |Xm ||UXm | ≤
(3/4)7 log n n2 < 1, so either Xm = ∅ or UXm = ∅ and EdgeCount terminates at (E2). Thus Bm+1
must occur in all cases.
By Lemma 12: (E1) requires O(1) time and O(1) calls to the adjacency oracle; each iteration
of (E2) or (E5) requires ε−2 · Õ(n) time, ε−2 · O((log n)5 ) calls to the independence oracle,
and ε−2 · Õ(n) calls to the adjacency oracle; each iteration of (E3) or (E6) requires O(n) time; each
iteration of (E4) requires ε−4 · Õ(n) time and ε−4 · Õ(n) calls to the adjacency oracle; and each
iteration of (E7) requires O(n) time. Since m = O(log n), it follows that when E1 , . . . , Em occur,
EdgeCount satisfies the time and oracle access restrictions of (iv)–(vi), so A does not terminate
early.
Let j = max{i ∈ [m] : Bi does not occur}, so that A1 , . . . , Aj and Bj+1 occur and EdgeCount
outputs 2tj ∂(X
p for all i we have ti+1 ≤ ti + 1, it follows that tj ≤ m ≤ 9 log n.
√ j ) + Nj . Since
Moreover, 2tj ζ log n ≤ 18 ζ(log n)3 = ε/2 ≤ 1. Hence
p
1 − 2 ζ log n
tj
p
≥ 1 − 2tj ζ log n ≥ 1 − ε,
√
p
p
tj
1 + 2 ζ log n
≤ e2tj ζ log n ≤ 1 + 4tj ζ log n ≤ 1 + ε.
It therefore follows from (11) that when E1 , . . . , Em occur, (1 − ε)e(G) ≤ 2tj ∂(Xj ) + Nj ≤ (1 + ε)e(G),
as required by (iii).
It remains to prove that P(E1 , . . . , Em ) ≥ 2/3. Let Fi be the event that in the ith iteration of
(E2)–(E7), one of the following events occurs:
Fi1 : the invocation of FindCore at (E2) fails.
Fi2 : EdgeCount loops at (E3) and either |Xi+1 | > 3|Xi |/4 or
p
1
∂(Xi+1 ) − ∂(Xi ) > ζ log n · ∂(Xi ).
2
16
Fi3 : the invocation of FindCore at (E5) fails.
Fi4 : EdgeCount loops at (E6) and either |Xi+1 | > 3|Xi |/4 or
p
1
∂(Xi+1 ) − ∂(Xi \ Si ) > ζ log n · ∂(Xi \ Si ).
2
We claim that for all i ∈ [m − 1], conditioned on Ei , we have Ei+1 ⊆ Fi .
Suppose that Ei ∩ Bi+1 ∩ Fi occurs, and note that this implies that Ai ∩ Bi occurs. Then we must
prove that Ai+1 occurs. We split into cases depending on the behaviour of EdgeCount in the ith
iteration.
Case 1: EdgeCount loops at (E6). In this case, since Ai ∩ Fi4 occurs, we have |Xi+1 ||UXi+1 | ≤
(3/4)|Xi ||UXi | ≤ (3/4)i n2 . Moreover,
1 p
− ζ log n 2ti+1 ∂(Xi \ Si ) + Ni+1
2
p
= (1 − 2 ζ log n)2ti ∂(Xi \ Si ) + 2ti ∂(Si ) + Ni
2ti+1 ∂(Xi+1 ) + Ni+1 ≥
≥ (1 − 2 ζ log n)(2ti ∂(Xi ) + Ni )
p
p
≥ 1 − 2 ζ log n
ti+1
e(G),
t
√
and (likewise) 2ti+1 ∂(Xi+1 ) + Ni+1 ≤ 1 + 2 ζ log n i+1 e(G). Thus Ai+1 occurs.
Case 2: EdgeCount loops at (E3). This case is very similar to Case 1, with Fi2 in place of Fi4 ,
and so we omit it.
Case 3: EdgeCount loops at (E7). Since Fi1 does not occur, Si must be a (ζ/48)-core of Xi .
Since EdgeCount does not loop at (E3), it follows that Si is a (ζ/48)-unbalancer of Xi . Likewise,
since Fi3 does not occur and EdgeCount does not loop at (E6), the output Si0 of FindCore at
(E5) must be a ζ-unbalancer of Xi \ Si . It follows by Lemma 13(ii) that |UXi+1 | ≤ |UXi |/2. Since Ai
occurs, it follows that |Xi+1 ||UXi+1 | ≤ (3/4)i n2 . Moreover, since 2ti+1 ∂(Xi+1 )+Ni+1 = 2ti ∂(Xi )+Ni
and Ai occurs, (11) also holds and Ai+1 occurs. We have therefore shown that Ei+1 ⊆ Fi conditioned
on Ei , as claimed.
Now, it follows from the correctness of FindCore (Lemma 12) that for all i ∈ [m − 1],
P(Fi1 | E1 , . . . , Ei ) ≤ 3/n,
(12)
P(Fi3 | E1 , . . . , Ei , Fi1 , Fi2 ) ≤ 3/n.
(13)
If E1 , . . . , Ei and Fi1 occur, and EdgeCount loops at (E3), then Si must be a (ζ/48)-witness of Xi
(since Fi1 does not occur), so by Lemma 13(i), Xi is (ζ/48)-balanced. Moreover, since EdgeCount
does not halt at (E2), |Xi | ≥ 24 log n. Thus by Lemma 10,
P(Fi2 | E1 , . . . , Ei , F11 ) ≤ 4/n.
(14)
Likewise, if EdgeCount loops at (E6) during the ith iteration, then |Xi \ Si | ≥ 24 log n and the
output Si0 of FindCore at (E5) must be a ζ-witness of Xi \ Si , so
P(Fi4 | E1 , . . . , Ei , F11 , F12 , F13 ) ≤ 4/n.
(15)
It follows by (12)–(15) and a union bound that P(Ei+1 | E1 , . . . , Ei ) ≥ 1−14/n. Hence P(E1 , . . . , Em ) ≥
1 − 14m/n ≥ 1 − 112 log n/n ≥ 2/3, where the last inequality follows since n ≥ 3000.
17
5 Applications of Theorem 10
5.1 3SUM
We formally define the problems as follows.
Problem: 3SUM
Input: Three lists A, B and C of integers.
Task: Decide whether there exists a tuple (a, b, c) ∈ A × B × C such that a + b = c.
Problem: #3SUM
Input: Three lists A, B and C of integers.
Task: Count the number of tuples (a, b, c) ∈ A × B × C such that a + b = c.
Theorem 3 (restated). If 3SUM with n integers from [−N, N ] has a randomised T (n, N )time algorithm, then there is an (ε−2 T (n, N ) · O((log n)6 ) + ε−4 log N · Õ(n))-time randomised
ε-approximation algorithm for #3SUM.
Proof. Let (A, B, C) be an instance of #3SUM and let 0 < ε < 1. If ε ≤ n−3 then we solve the
problem exactly in time O(n3 log N ) = O(ε−1 log N ), so suppose ε > n−3 . Let E = {(a, b) ∈ A × B :
a + b ∈ C}, and let G = (A, B, E). Note that if C has been sorted, then adjG can be evaluated in
time (log N ) · O(log n) using binary search. Moreover, for all X ⊆ A ∪ B, indG (X) = 1 if and only
if (X ∩ A, X ∩ B, C) is a ‘no’ instance of 3SUM, so indG can be evaluated with failure probability
at most 1/3 in time O(n) + T (n, N ) = O(T (n, N )). (Recall that T (n, N ) = Ω(n).) Otherwise, as
in the proof of Theorem 9, for all constant K > 0, indG can be evaluated with failure probability
at most ε2 /(K(log n)6 ) in time Õ(T (n, N )). We therefore sort C in (log N ) · Õ(n) time, apply the
algorithm of Theorem 10 to A, B and ε, and return the result.
Theorem 4 (restated). For all δ > 0, there is a randomised ε-approximation algorithm with
running time ε−4 · Õ(n2−δ/7 ) for instances of #3SUM with n integers in [−n3 , n3 ] such that at least
one of A, B, or C may be covered by n1−δ intervals of length n.
Proof. Say a set S ⊆ Z is (n, δ)-clustered if it can be covered by at most n1−δ intervals of length n;
note that it can be checked in quasilinear time whether a set is (n, δ)-clustered. Let (A, B, C) be
an instance of #3SUM in which at least one of A, B or C is (n, δ)-clustered. By negating and
permuting sets if necessary, we may assume that C is (n, δ)-clustered. Exactly as in the proof
of Theorem 3, any randomised T (n, N )-time algorithm for 3SUM on such instances yields an
(ε−2 T (n, N ) · O((log n)6 ) + ε−4 log N · Õ(n))-time randomised approximation scheme. (In particular,
note that (X ∩ A, X ∩ B, C) remains an instance of the restricted problem.) Chan and Lewenstein [3,
Corollary 4.3] provide a randomised O(n1−δ/7 )-time algorithm for such instances, so the result
follows.
5.2 Orthogonal Vectors
We formally define the problems as follows.
Problem: OV
Input: Two lists A and B of zero-one vectors in Rd .
P
Task: Decide whether there exists a pair (u, v) ∈ A × B such that di=1 ui vi = 0.
18
Problem: #OV
Input: Two lists A and B of zero-one vectors in Rd .
P
Task: Count the number of pairs (u, v) ∈ A × B such that di=1 ui vi = 0.
Theorem 5 (restated). If OV with n vectors in d dimensions has a randomised T (n, d)-time
algorithm, then there is a randomised (ε−2 T (n, d) · O((log n)6 ) + ε−4 d · Õ(n))-time ε-approximation
algorithm for #OV.
Proof. Let (A, B) be an instance of #OV, and let 0 < ε < 1. If ε ≤ n−2 then we solve the problem
exactly in O(ε−1 d) time, so suppose ε > n−2 . Let E = {(a, b) ∈ A × B : ha, bi = 0}, and let
G = (A, B, E) be a bipartite graph. Note that adjG can be evaluated naïvely in O(d) time. Moreover,
for all X ⊆ A ∪ B, indG (X) = 1 if and only if (A ∩ X, B ∩ X) is a ‘no’ instance of OV, so indG can
be evaluated with failure probability at most 1/3 in time O(n) + T (n, d) = O(T (n, d)). As in the
proof of Theorem 3, it follows that indG can be evaluated with sufficiently low failure probability in
time Õ(T (n, d)). We therefore apply the algorithm of Theorem 10 to A, B and ε and return the
result.
In the following definitions, R is a constant finite ring.
Problem: OV(R)
Input: Two lists A and B of vectors in Rd .
P
Task: Decide whether there exists a pair (u, v) ∈ A × B such that di=1 ui vi = 0R .
Problem: #OV(R)
Input: Two lists A and B of vectors in Rd .
P
Task: Count the number of pairs (u, v) ∈ A × B such that di=1 ui vi = 0R .
Theorem 6 (restated). Let m = pk be a constant prime power. Then there is a randomised
ε-approximation algorithm for n-vector instances of #OV over GF(m)d (resp. (Z/mZ)d ) in time
ε−4 d(p−1)k · Õ(n) (resp. ε−4 dm−1 · Õ(n)).
Proof. Exactly as in the proof of Theorem 5, any randomised T (n, d)-time algorithm for OV(R)
yields an (ε−2 T (n, d)·O((log n)6 )+ε−4 d· Õ(n))-time randomised approximation scheme for #OV(R).
The result therefore follows from Theorems 1.6 and 1.3 (respectively) of Williams and Yu [25].
5.3 Negative-Weight Triangles
We formally define the problems as follows.
Problem: NWT
Input: A tripartite graph G and a symmetric function w : V (G)2 → Z.
Task: Decide whether there exists a triangle abc in G such that w(a, b) + w(b, c) + w(c, a) < 0.
Problem: #NWT
Input: A tripartite graph G and a symmetric function w : V (G)2 → Z.
Task: Count the number of triangles abc in G such that w(a, b) + w(b, c) + w(c, a) < 0.
Theorem 7 (restated). If NWT for n-vertex graphs with weights from [−W, W ] has a randomised
T (n, W )-time algorithm, then there is a randomised (ε−2 T (n, W ) · O((log n)6 ) + ε−4 log W · Õ(n2 ))time ε-approximation algorithm for #NWT.
19
Proof. Let (G, w) be an instance of #NWT, let A, B and C be the vertex classes of G, and let
0 < ε < 1. If ε < n−3 then we solve the problem exactly in O(ε−1 log W ) time, so suppose ε ≥ n−3 .
Let U = A, let V = {e ∈ E(G) : e ⊆ B ∪ C}, and let
n
o
E = (a, {b, c}) ∈ U × V : {a, b}, {a, c} ∈ E(G) and w(a, b) + w(b, c) + w(c, a) < 0 .
Let G0 = (U, V, E), so that G0 is a bipartite graph. Note that adjG0 can be evaluated naïvely in
O(log W ) time. Moreover, for all X ⊆ U ∪ V , define a graph GX by V (GX ) = (X ∩ A) ∪ B ∪ C and
n
o
E(GX ) = e ∈ E(G) : e ∩ X ∩ A 6= ∅ or e ∈ X ∩ V .
Let wX = w|V (GX )2 . Then for all X ⊆ U ∩ V , indG0 (X) = 1 if and only if (GX , wX ) is a
‘no’ instance of NWT, so indG0 can be evaluated with failure probability at most 1/3 in time
O(n2 ) + T (n, W ) = O(T (n, W )). As in the proof of Theorem 3, it follows that indG0 can be
evaluated with sufficiently low failure probability in time Õ(T (n, W )). We therefore apply the
algorithm of Theorem 10 to U , V and ε and return the result.
Problem: APSP
Input: A directed graph G and a function w : E(G) → Z.
Task: Output the matrix A such that for all u, v ∈ V (G), Au,v is the minimum weight of any
path from u to v in G.
Theorem 8 (restated). There is a randomised ε-approximation√which runs on n-vertex instances
of #NWT with polynomially bounded weights in time ε−4 n3 /eΩ( log n) .
instance of APSP with polynomially bounded
Proof. By Williams [24, Theorem 1.1], an n-vertex
√
edge weights can be solved in time n3 /eΩ( log n) . There is a well-known reduction from NWT
to APSP with only constant overhead, which we give explicitly in the following paragraph. The
result therefore follows√immediately from Theorem 7, noting that the polylogarithmic overhead is
subsumed into the eΩ( log n) term.
Let (G, w) be an instance of NWT, writing G = (V, E). Form an instance (G0 , w0 ) of APSP as
follows. Let V (G0 ) = (V × [3]), and let
E(G0 ) =
[
[
{{(u, i), (v, i + 1)}, {(v, i), (u, i + 1)}}.
i∈{1,2} {u,v}∈E
Let w0 ({(u, i), (v, i + 1)}) = w(u, v) for all {(u, i), (v, i + 1)} ∈ E(G0 ). Thus for all {u, v} ∈ E, the
paths from (u, 1) to (v, 3) in G0 correspond exactly to the triangles containing {u, v} in G. Let A
be the output of APSP on G0 . Then (G, w) is a ‘yes’ instance of NWT if and only if for some
{u, v} ∈ E(G), we have Au,v + w(u, v) < 0. This can be checked in O(n2 ) time.
References
[1] Amir Abboud, Richard Ryan Williams, and Huacheng Yu. More applications of the polynomial
method to algorithm design. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015, pages 218–230,
2015.
20
[2] Chris Calabro, Russell Impagliazzo, Valentine Kabanets, and Ramamohan Paturi. The complexity of unique k-SAT: An isolation lemma for k-CNFs. J. Comput. Syst. Sci., 74(3):386–393,
2008.
[3] Timothy M. Chan and Moshe Lewenstein. Clustered integer 3SUM via additive combinatorics.
In Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing,
STOC 2015, Portland, OR, USA, June 14-17, 2015, pages 31–40, 2015.
[4] Timothy M. Chan and Ryan Williams. Deterministic APSP, Orthogonal Vectors, and more:
Quickly derandomizing Razborov-Smolensky. In Proceedings of the Twenty-Seventh Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 2016, Arlington, VA, USA, January
10-12, 2016, pages 1246–1255, 2016.
[5] Martin Dyer, Leslie Ann Goldberg, Catherine Greenhill, and Mark Jerrum. The relative
complexity of approximate counting problems. Algorithmica, 38(3):471–500, 2004.
[6] Jörg Flum and Martin Grohe. The parameterized complexity of counting problems. SIAM J.
Comput., 33(4):892–922, April 2004.
[7] Timon Hertli. 3-SAT faster and simpler - unique-SAT bounds for PPSZ hold in general. SIAM
J. Comput., 43(2):718–729, 2014.
[8] Russell Impagliazzo, William Matthews, and Ramamohan Paturi. A satisfiability algorithm
for AC0 . In Proceedings of the Twenty-Third Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2012, Kyoto, Japan, January 17-19, 2012, pages 961–972, 2012.
[9] Russell Impagliazzo and Ramamohan Paturi. On the complexity of k-SAT. J. Comput. Syst.
Sci., 62(2):367–375, 2001.
[10] Russell Impagliazzo, Ramamohan Paturi, and Francis Zane. Which problems have strongly
exponential complexity? J. Comput. Syst. Sci., 63(4):512–530, 2001.
[11] Svante Janson, Tomasz Łuczak, and Andrzej Rucinski. Random Graphs. John Wiley & Sons,
2000.
[12] Konstantin Kutzkov. New upper bound for the #3-SAT problem. Inf. Process. Lett., 105(1):1–5,
2007.
[13] Colin McDiarmid. On the method of bounded differences. In Surveys in Combinatorics, 1989:
Invited Papers at the Twelfth British Combinatorial Conference, pages 148–188, 1989.
[14] Moritz Müller. Randomized approximations of parameterized counting problems. In Parameterized and Exact Computation: Second International Workshop, IWPEC 2006, Zürich,
Switzerland, September 13-15, 2006. Proceedings, pages 50–59, 2006.
[15] Ramamohan Paturi, Pavel Pudlák, Michael E. Saks, and Francis Zane. An improved exponentialtime algorithm for k-SAT. J. ACM, 52(3):337–364, May 2005.
[16] Michael Sipser. A complexity theoretic approach to randomness. In Proceedings of the Fifteenth
Annual ACM Symposium on Theory of Computing, STOC ’83, pages 330–335, New York, NY,
USA, 1983.
21
[17] Larry Stockmeyer. On approximation algorithms for #P. SIAM J. Comput., 14(4):849–861,
1985.
[18] Marc Thurley. An approximation algorithm for #k-SAT. In 29th International Symposium
on Theoretical Aspects of Computer Science, STACS 2012, February 29th - March 3rd, 2012,
Paris, France, volume 14 of LIPIcs, pages 78–87, 2012.
[19] Seinosuke Toda. PP is as hard as the polynomial-time hierarchy. SIAM J. Comput., 20(5):865–
877, October 1991.
[20] Patrick Traxler. The relative exponential time complexity of approximate counting satisfying
assignments. In Parameterized and Exact Computation: 9th International Symposium, IPEC
2014, Wroclaw, Poland, September 10-12, 2014. Revised Selected Papers, pages 332–341, 2014.
[21] Leslie G. Valiant. The complexity of computing the permanent. Theor. Comput. Sci., 8(2):189–
201, 1979.
[22] Leslie G. Valiant and Vijay V. Vazirani. NP is as easy as detecting unique solutions. Theor.
Comput. Sci., 47:85–93, 1986.
[23] Ryan Williams. A new algorithm for optimal 2-constraint satisfaction and its implications.
Theor. Comput. Sci., 348(2-3):357–365, 2005.
[24] Ryan Williams. Faster all-pairs shortest paths via circuit complexity. In Symposium on Theory
of Computing, STOC 2014, New York, NY, USA, May 31 - June 03, 2014, pages 664–673,
2014.
[25] Ryan Williams and Huacheng Yu. Finding orthogonal vectors in discrete structures. In
Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2014, Portland, Oregon, USA, January 5-7, 2014, pages 1867–1877, 2014.
[26] Virginia V. Williams. Hardness of easy problems: Basing hardness on popular conjectures such
as the Strong Exponential Time Hypothesis. In 10th International Symposium on Parameterized
and Exact Computation, IPEC 2015, September 16-18, 2015, Patras, Greece, pages 17–29,
2015.
[27] Virginia V. Williams and Ryan Williams. Subcubic equivalences between path, matrix and
triangle problems. In 51th Annual IEEE Symposium on Foundations of Computer Science,
FOCS 2010, October 23-26, 2010, Las Vegas, Nevada, USA, pages 645–654, 2010.
22
| 8cs.DS
|
arXiv:1801.08113v1 [cs.NE] 24 Jan 2018
Free Energy Minimization Using the
2-D Cluster Variation Method: Initial
Code Verification and Validation
THM TR2018-001(ajm)
Alianna J. Maren
Northwestern University School of Professional Studies
Master of Science in Data Science Program
Themasis Associates
Date: 2018-01-23
Version 1.0
1
Abstract
A new approach for general artificial intelligence (GAI), building
on neural network deep learning architectures, can make use of one
or more hidden layers that have the ability to continuously reach a
free energy minimum even after input stimulus is removed, allowing
for a variety of possible behaviors. One reason that this approach
has not been developed until now has been the lack of a suitable free
energy equation; one that would avoid some of the difficulties known in
Hopfield-style neural networks. The Cluster Variation Method (CVM)
offers a means for characterizing 2-D local pattern distributions, or
configuration variables, and provides a free energy formalism in terms
of these configuration variables. The equilibrium distribution of these
configuration variables is defined in terms of a single interaction enthalpy parameter, h, for the case of equiprobable distribution of bistate
(neural/neural ensemble) units. For non-equiprobable distributions,
the equilibrium distribution can be characterized by providing a fixed
value for the fraction of units in the active state (x1 ), corresponding
to the influence of a per-unit activation enthalpy, together with the
pairwise interaction enthalpy parameter h.
This paper provides verification and validation (V&V) for code
that computes the configuration variable and thermodynamic values
for 2-D CVM grids characterized by different interaction enthalpy
parameters, or h-values. This means that there is now a working
foundation for experimenting with a 2-D CVM-based hidden layer
that can, as an alternative to responding strictly to inputs, also now
independently come to its own free energy minimum. Such a system
can also return to a free energy-minimized state after it has been
perturbed, which will enable a range of input-independent behaviors
that have not been hitherto available. A further use of this 2-D CVM
grid is that by characterizing different kinds of patterns in terms of
their corresponding h-values (together with their respective fraction of
active-state units), we have a means for quantitatively characterizing
different kinds of neural topographies. This further allows us to connect
topographic descriptions (in terms of local patterns) with free energy
minimization, allowing a first-principles approach to characterizing
topographies and building new computational engines.
Keywords: artificial intelligence; neural networks; deep learning;
statistical thermodynamics; free energy; free energy minimization;
Cluster Variation Method; entropy; brain networks; neural connectivity
2
1
Introduction and Overview
This article documents the verification and validation (V&V) results for
the first two stages of code development for free energy minimization
within a 2-D Cluster Variation Method (CVM) system.
The intention is that this 2-D CVM system can have its free energy
minimized independent of its use in any other process. Ultimately,
the 2-D CVM system will be inserted as a single layer into a neural
network, creating a new form of computational engine, which I call
the CORTECON, standing for a COntent-Retentive, TEmporallyCONnected neural network, first described in [1] and [2], both of which
presented early results using a 1-D CVM.
This work described here focuses on a 2-D CVM grid, which can
operate as both a hidden layer and as a independent functional unit,
with the ability to achieve a free energy minimum when there is no
extraneous signal coming into the layer, is shown in Figure 1.
This notion of using a 2-D CVM as a computational engine’s hidden
layer advances ideas originally proposed in [1] and [2], along with [3],
and further incorporates (and makes practical) ideas put forth by
Karl Friston, whose notation was adopted for Figure 1. Specifically,
this figure illustrates the computational engine using Friston’s notion
of a set of computational (representational) units separated from an
external system by a Markov blanket. It also allows for the variational
Bayes (free energy minimization) approach described by Friston [4],
[5], and [6], and earlier by Beal [7].
In brief, Friston (building on work by Beal [7]) proposes a computational system in which a Markov blanket separates the computational
(representational) elements of the engine from external events, as shown
in Figure 1. The communication between the external system elements
(denoted ψ̃) with those of the representational system (denoted λ or
r̃) are mediated by two distinct layers or components of the Markov
blanket; the sensing (s̃) elements and the action (ã) ones.
This article provides V&V for the first two code development stages:
1. Computing values for the configuration variables in a
2-D system – for various values of the interaction enthalpy
parameter h, and
2. Computing the thermodynamic quantities associated with
the 2-D system – given the set of configuration variables, it is
possible to then compute enthalpy, entropy, and free energy.
3
Figure 1: Illustration of a Cluster Variation Method (CVM)-based computational engine, for which the Markov blanket of sensing and active units
corresponds to input and output layers (see Friston [4]). Unique to the approach advanced here, the computational layer is composed as a 2-D CVM,
for which the free energy equation can be explicitly written, and the free
energy minimum can be found either analytically or computationally, depending on the parameters used. The CVM layer comprises the internal or
representational units (r), and cannot communicate with the external field
(shown in two parts for visualization purposes only). However, units within
the representational layer can receive inputs from the sensory units (s) and
send signals to the active (a) units. The sensory units can receive inputs from
external stimulus, and send signals to the representational units. The active
units can receive inputs from the representational units, and send signals to
the external output units. (In the notional view advanced by Friston (op.
cit.), a broader set of interactions is allowed; for simplicity in this engine, the
interaction pathways have been streamlined.)
4
Most crucially, the code incorporates a free energy minimization
process, so that once an initial (randomly-generated pattern) has been
created, it is adjusted in a two-stage process:
1. Achieve desired x1 specification – this allows us to implicitly
enfold a nominal per-unit activation energy (where the relationship between this parameter ε0 and x1 cannot be explicitly stated
at this time), and
2. Achieve free energy minimization for the given set of
x1 , h-values – typically, the 2-D CVM grid needs to have state
changes in its various units to achieve a free energy minimum.
2
The Configuration Variables
The first V&V aspect of the task documented here was to ensure that
the configuration variables for the 2-D grid were counted correctly.
1. Configuration variable definitions – including how they are
counted in the 2-D CVM grid,
2. 2-D CVM grid specifications – size and wrap-arounds, and
3. V&V results – configuration variable counts for select examples.
2.1
Introducing the Configuration Variables
The Cluster Variation Method, introduced by Kikuchi [8], uses an entropy term that includes not only the distribution into simple “on” and
“off” states, but also distribution into local patterns, or configurations,
as illustrated in the following figures.
A 2-D CVM is characterized by a set of configuration variables,
which collectively represent single unit, pairwise combination, and
triplet values. The configuration variables are denoted as:
• xi - Single units,
• yi - Nearest-neighbor pairs,
• wi - Next-nearest-neighbor pairs, and
• zi - Triplets.
These configuration variables are illustrated for a single zigzag
chain in Figure 2.
5
xi
wi
yi
zi
Figure 2: The 1-D single zigzag chain is created by arranging two staggered
sets of M units each. The configuration variables shown are xi (single units),
yi (nearest-neighbors), wi (next-nearest-neighbors), and zi (triplets).
Table 1: Configuration Variables for the Cluster Variation Method
Name
Variable Instances
Unit
Nearest-neighbor
Next-nearest-neighbor
Triplet
xi
yi
wi
zi
2
3
3
6
For a bistate system (one in which the units can be in either
state A or state B), there are six different ways in which the triplet
configuration variables (zi ) can be constructed, as shown in Figure 3,
and also in Table 1.
Notice that within Figure 3, the triplets z2 and z5 have two possible
configurations each: A-A-B and B-A-A for z2 , and B-B-A and A-BB for z5 . This means that there is a degeneracy factor of 2 for
each of the z2 and z5 triplets.
The degeneracy factors βi and γi (number of ways of constructing a
given configuration variable) are shown in Figure 4; β2 = 2, as y2 and
w2 can be constructed as either A-B or as B-A for y2 , or as B- -A
or as A- -B for w2 . Similarly, γ2 = γ5 = 2 (for the triplets), as there
are two ways each for constructing the triplets z2 and z5 . All other
degeneracy factors are set to 1.
2.2
Counting the Configuration Variables
To experiment with the 2-D CVM system, I constructed various grids
of 256 (16 x 16) units each, as illustrated in Figure 5.
6
Figure 3: The six ways in which the configurations zi can be constructed.
I decided to use a 16 x 16 grid for several reasons:
1. Sufficient variety in local patterns – I was able to construct
grids that illustrated several distinct kinds of topographies (each
corresponding to different h-values),
2. Sufficient nodes – so that triplet-configuration extrema could
be explored in some detail, and
3. Countability – I needed to be able to manually count all the
configuration values for a given 2-D grid configuration, and match
them against the results from the program, as a crucial V&V
step.
One final advantage of the 16 x 16 grid layout was that the different
grid configurations were both large enough to show diversity, but small
enough so that I could create a figure illustrating the activation state
7
Figure 4: Illustration of the configuration variables for the Cluster Variation
Method, showing the ways in which the configurations variables yi , wi , and
zi can be constructed, together with their degeneracy factors βi and γi .
(A or B) of each node, thus illustrating the detailed particulars of each
configuration design.
I began with manually-designed grid configurations, such as the two
shown in Figure 5. These two configurations correspond (somewhat)
to the notions of “scale-free” and “rich-club” topographies, as observed
in various neural communities. (For references, please consult [3].)
These two different grid configurations are early attempts to characterize how the h-values can be identified for grids with different
total configuration variable values. The following Section 3 will discuss
h-values in the context of the free energy equation.
Both of these systems were created with the constraint of equiprobable occurrence of units in states (A or B; that is, x1 = x2 = 0.5. This
was done to facilitate the next V&V step, which will be discussed in
Section 3. Thus, for the configurations shown in Figure 5, both the (a)
and (b) grids have 128 nodes each of units in state A and in state B.
8
Figure 5: Illustration of the two different grids for experiments with the 2-D
CVM system.
The configuration on the left of Figure 5 is an effort to build a
“scale-free-like” system. The notion of a “scale-free” system is that
the same kind of pattern replicates itself throughout various scales of
observation in a system. Thus, for the “scale-free-like” configuration
shown in Figure 5 (a), I created a design that was originally intended
to be 180-degree symmetrical around a central axis (dihedral group-2
symmetry). Specifically, the left and right sides were to be identical in
a rotated-180-degree sense.
For ease in design of the “scale-free-like” system, I focused on
creating a pattern on one side and duplicating it on the other. I
used a paisley-like base pattern in order to create greater dispersion
of values across the zi triplets; that is, I wanted to minimize having
tightly-clustered islands that would yield little in the way of A-B-A
and B-A-B triplets (z2 and z5 , respectively).
The practical limitation of attempting to fit various “islands” of
A nodes (black) into a surrounding “sea” of B nodes (white) meant
that there were not quite enough B nodes to act as borders around
the more compact sets of A nodes. Thus, the pattern in the right half
of grid (a) is a bit more compressed than originally planned.
9
The original plan was that out of 256 nodes in the grid, half would
be on the right, and half on the left; 128 nodes on each side. Of
these, for each side, 64 nodes were to be in state A (black). Of these
nodes (per side), sixteen (16 nodes) would be used to create a large,
paisley-shaped island. The remaining 64 - 16 = 48 nodes would be
used for smaller-sized islands; two islands of eight nodes each, etc. The
plan is shown in Figure 6. The notation of “center” and “off-center“
refers to the placement of the various islands; the largest (16-node)
islands were to be placed more-or-less in the center of each of their
respective (left or right) sides of the grid, and the remaining islands
were to be “off-center”; situated around their primary respective large
islands.
The resulting patterns were close to the original plan, although not
exactly the same. (Again, for details, see Figure 6.)
Even though some changes had to be made to the original design
plan, the original constraint, that the number of units in states A and
B would be identical (128 nodes in each), was kept. The details are
shown in Figure 6.
The validation step for this stage of code development was to
manually count all the configuration variables for several different
configuration grids, such as the ones shown in Figure 5.
The counts for the “scale-free-like” grid shown in Figure 6 are shown
in Figure 7. It suffices to say that the results from the manual counting
(of all configuration variables) and those created by the computer code
were identical. These held true across several different grids with
different node configurations.
Note: To achieve the fractional variables shown in Figure 3, and
also in Table 1, the following relations are used:
• xi = Xi ,
• yi = Yi /2, for i = 1, 3 and y2 = Y2 /4, accounting for the degeneracy with which y2 occurs,
• wi = Wi /2, for i = 1, 3 and w2 = W2 /4, accounting for the
degeneracy with which w2 occurs, and
• zi = Zi /2, for i = 1, 3, 4, 6 and z2 = Z2 /4, z5 = Z5 /4, accounting
for the degeneracy with which z2 and z5 occur.
Note: The exact details of the row counts are difficult to read in
Figures 7 and 8; the original diagrams are in a corresponding slidedeck
10
Figure 6: A 2-D CVM “scale-free-like” system with an equal number state A
and state B nodes (128 nodes each).
that will be available in the associated GitHub repository; see details
at the end of this document.
The second configuration, for an “extreme-rich-club-like” configuration, is shown in Figure 8.
As a contrast to the grid configuration used in Figures 6 and 7, I
created a second configuration that had only one large compact region
of nodes in state A, which was wrapped-around the grid envelope,
as shown in Figure 8. This configuration was designed to maximize
the number of pairwise and triplet configurations that put “like-nearlike.” The previous configuration, shown in Figure 6, was more in the
direction of “like-near-unlike.”
The purpose of having configurations with such different dispersions among the configuration variable values was that they would
11
Figure 7: A 2-D CVM “scale-free-like” system with an equal number state A
and state B nodes (128 nodes each).
putatively yield different h-values, or correspond to different points
on an equilibrium curve for the free energy equation (in the case of
equiprobable units in states A and B). As I have analytic results for
that free energy minimum curve (the equilibrium point for the free
energy at different h-values, or interaction enthalpy values), it would
serve as both a useful experiment and V&V test. These results are
discussed in the following Section 3.
The V&V for the initial stage of code development; ascertaining
that the configuration variable counts were as they should be, was
complete.
There is an accompanying slidedeck that documents the code block
structure and provides other important elements of code documentation
(other than V&V); this will also be available on GitHub; see the end
12
Figure 8: A 2-D CVM “extreme-rich-club-like” system with an equal number
state A and state B nodes (128 nodes each).
of this document for details.
By far, the most complex element of the “configuration variable
counting” code was in counting the triplets. The V&V step ensured
that the counts wrapping around from right to left, and from top to
bottom (creating a completely-wrapped envelope of the initial 2-D
grid) performed as desired and expected.
3 Verification and Validation of Computing the Thermodynamic Variables
There were two primary means for obtaining validation that the code
computing the thermodynamic variables was correct:
13
1. Comparison with analytic for the equiprobable case –
for the case equiprobable distribution among the xi variables
(x1 = x2 = 0.5), I have developed an analytic solution, which
gives a means for comparing the code-generated results against
the expected (analytic) results, and
2. Comparison with analytic for the case where the interaction enthalpy is zero – the second means to check the codegenerated results is for the case where the distributuion of x
values is not equiprobable, however the interaction enthalpy is
set to zero (h = 1), and thus the exact distribution of other
configuration values can be precisely computed, allowing further
for exact analytic computation of thermodynamic variables.
The previous section described the patterns generated for the
validation of configuration variable counting. It was interesting to see
how the thermodynamic variables emerged for the systems described
there, however (as will be illustrated here), certain of those system were
not at equilibrium, even though they had equiprobable distribution of
xi values. As these results are more in the realm of theory and less
V&V, they will be discussed elsewhere.
The realization that manually-generated patterns would not necessarily be at equilibrium meant that I needed to have test cases where
the patterns would indeed be at equilibrium; this required not only
random generation of patterns, but also that they be modified so that
their associated free energy values achieved minimum. This generationand-modification process is described more thoroughly in the following
Section 4.
3.1
Validation support: analytic solution
The analytic solution for the case where x1 = x2 = 0.5 can be found
when we are using the full interaction enthalpy term of ε1 ∗(2y2 −y1−y3).
This solution is similar to the more limited enthalpy equation, used in
[3] as well as in the predecessor work [9], where ε1 ∗ (2y2 ).
The free energy equation for a 2-D CVM system, including configuration variables in the entropy term, is
14
Ḡ2−D = G2−D /N =
ε(z2 + z3 + z4 + z5 )
"
− 2
3
X
βi Lf (yi )) +
i=1
+µ(1 −
3
X
βi Lf (wi )) −
i=1
6
X
2
X
βi Lf (xi )) − 2
i=1
i=1
γi zi ) + 4λ(z3 + z5 − z2 − z4 )
i=1
(1)
where µ and λ are Lagrange multipliers, and we have set kβ T = 1.
Note: the full derivation of the 2-D CVM free energy is presented
in [9], and the preceding equation corresponds to Equations (2)–(14)
of that reference. Also, the single enthalpy parameter here is ε, with
the enthalpy parameter for unit activation implicitly set to zero, as
the earlier intention was to solve the above equation for an analytic
solution, which was possible only in the case where x1 = x2 = 0.5,
meaning that the per-unit enthalpy activation parameter ε0 = 0.
The enthalpy term used previously, in [3] and in [9], was
H̄2−D = H2−D /N = ε1 (2y2 ) = ε1 (z2 + z3 + z4 + z5 ).
(2)
The approach that I am using currently is to take the same enthalpy
equation as originally advocated by Kikuchi [8], [10], which gives
H̄2−D = H2−D /N = ε1 (2y2 − y1 − y3 ) = ε1 (z4 + z3 − z1 − z6 ). (3)
Both of these equations are found using equivalence relations, specifically
y2 = z 2 + z 4 = z 3 + z 5
(4)
2y2 = z2 + z4 + z3 + z5 .
(5)
I have previously found analytic solution for this equation, for the
condition where x1 = x2 = 0.5, it is
z3 =
(h2 − 3)(h2 + 1)
.
8[h4 − 6h2 + 1]
15
6
X
(6)
#
γi Lf (zi )
When the more complete enthalpy expression is used, viz. ε1 (2y2 −
y1 − y3 ) = ε1 (z4 + z3 − z1 − z6 ), the analytic solution becomes
z3 =
(h4 − 3)(h4 + 1)
.
8[h8 − 6h4 + 1]
(7)
(Note: the full derivation of these results will be published separately.)
The experimentally-generated results from probabilistically-generated
data sets correspond to the analytic results in the neighborhood of
h = 0. The reason that the range is so limited is that the analytic
solution makes use of equivalence relations as expressed above. The
resulting solution has divergences at h = 0.172 and h = 5.828.
The comparison is shown in the following Figure 9. In this figure,
the column in the table marked as z3Analyt1 corresponds to results
from Eqns. 3 and 7 (the current approach) and in the next column,
z3Analyt2 corresponds to results from Eqns. 2 and 6 (the previous
approach).
The graph is shown in the following Figure 10.
The divergent behavior in the analytic solution is likely due to the
use of equivalence relationships, as identified in Eqn. 4.
3.2 Validation support: basic thermodynamic
results
The following Figure 11 shows the results when x1 = 0.5, which is the
case where all of the results should conform with the analytic solution.
The corresponding Figure 12 presents the data table supporting
Figure 11.
4 Verification and Validation of Free
Energy Minimization
It is not enough to simply compute the thermodynamic variables for a
given 2-D grid configuration; it is important to have a mechanism by
16
Figure 9: Data table giving the probabilistically-generated z3 results (after
reaching free energy minimum) vs. the analytic results for two different
formulations of the enthalpy expression, all for the case where x1 = x2 = 0.5,
and where h = 0.8..1.8.
which the pattern of node activations on the grid can adjust in order
to reach a free energy minimum.
I accomplished this by writing the code for two stages:
1. Bring x1 close to desired value, and
2. Adjust configuration variables to achieve free energy minimum.
Adjusting total number of nodes to achieve desired x1 :
The code has an initial specification for the desired x1 value (in
**main**), and randomly generates a 2-D CVM grid according to a
probabilistic assignment of “1” (state A) or “0” (state B) to the units
in the grid. However, just because the probability (of the random
number generation) is set to a specified value (say, 0.35) does not mean
that the resulting total of state A nodes will be precisely 0.35 of the
total number of nodes (e.g., 0.35 * 256, or 90 nodes); thus, a few nodes
will have to be “flipped” in order to bring the actual number of nodes
17
Figure 10: Graph giving the probabilistically-generated z3 results (after
reaching free energy minimum) vs. the analytic results for two different
formulations of the enthalpy expression, all for the case where x1 = x2 = 0.5,
and where h = 0.8..1.8. As discussed in the body of this work, the analytic
solution diverges at h = 0.172, where the denominator becomes zero.
in state A closer to the desired value.
The code specifies a tolerance value for how close the actual x1
value needs to be to the desired value. It runs a function to randomly
select and flip unit values (as needed, going in the right direction), and
continues this until the resulting actual x1 is within tolerance of the
desired value.
Validation: Printing out the actual values for x1 , ensuring
that they are within tolerance of the desired value.
Adjusting configuration variables to achieve free energy
minimum:
18
Figure 11: Configuration variable and thermodynamic values for the case
where x1 = x2 = 0.5, and where the interaction enthalpy parameter h ranges
as h = 0.8..1.8. See detailed explanation of results in the following Section 4,
as their nature is similar to these results.
There is no guarantee (in the current version of the code) that the
free energy minimum is actually met; instead, the code will run this
entire process (generating a new grid, adjusting for x1 within tolerance, and then adjusting the units so that free energy is progressively
decreased) for a specified number of trials. During the debug phase,
the number of trials was between 1 - 3, so that I could closely monitor
the process. During actual runs, the trials were typically 10 - 20; there
was not much variability in the results.
The goal of this process is to keep adjusting the grid units so that
free energy is decreased. For each run, there is a constant value for x1 .
That means, before any nodes are flipped, the program will (randomly)
find a node in state A, and another node in state B. It will flip the two
(from state A to state B, and vice versa). It will then compute the new
free energy; this requires recomputing the entire set of configuration
variable values. While x1 is held constant with this process, it is likely
19
Figure 12: Data table for configuration variable and thermodynamic values for
the case where x1 = x2 = 0.5, and where the interaction enthalpy parameter
h ranges as h = 0.8..1.8.
that all other configuration variables (yi , wi , and zi ) will change.
The program computes the new free energy value (using the new
configuration variable values as well as the h-value that is being tested
for the run). If the free energy is lower, the change in the units is kept.
If not, both units are reverted back to their original values.
The trials are strictly probabilistic for this generation of code
development; there is no attempt to find nodes whose topographic
position (i.e., sets of neighbors, nearest-neighbors, and triplets) would
be most likely to produce a free energy decrease if the node were to
change.
One version of the code is designed less to run multiple trials, and
more to collect, print, and plot the thermodynamic variables over a
series of attempts to flip nodes and test the resulting free energy.
One validation step is visual observation of the thermodynamic
variables over the course of any one of these trials; noting that the free
energy does, in fact, decrease.
20
Another validation step is that when h = 1 (ε1 = 0), there is
no interaction energy. In this case, the final configuration variable
values should be very close to their probabilistic likelihoods. Thus, for
example, when h = 1 and x1 = 0.35, we expect that y1 = 0.35 ∗ 0.35 =
0.1225, etc. Thus, it is possible to compare the actual resultant
configuration variable values with the probabilistic expectancies.
A final validation step is to compare the resulting behaviors against
the theoretical expectations. This is discussed more fully in the following subsection.
4.1
Validation Support: Exemplar Code Run
An example is shown in the following Figure 13.
This data is actually from a perturbation run, where the 2-D grid
is established as previously described, and then perturbed by a given
amount (in this case, a fraction of 0.1 of the existing nodes are flipped),
and then taken to free energy minimum a second time.
Figure 13: Configuration variable and thermodynamic values for the case
where x1 = 0.35 and x2 = 0.65, and where the interaction enthalpy parameter
h ranges as h = 0.8..1.8. See detailed explanation of results within Section 4.
21
These results were obtained from the program 2D-CVM-perturbexpt-1-2b-2018-01-12.py, run on Friday, Jan. 12, 2018.
The parameter settings were for x1 = 0.35 and h = 0.8..1.8, with
a total of twenty trials (numT rials = 20) for each h-value. A data
table from this run is shown in Figure 14. All reported results for
configuration variable and thermodynamic values are averages over
numT rials runs, where numT rials = 20.
4.1.1
Validation support: y2 results
The values observed for y2 conform to expectations. In Figure 13, y2
is shown in green, as y2 − 0.8 (in order to bring the y2 values within
the same visual range as other results).
When h = 1.0, y2 = 0.2278, which is the expected result. (The
true expected results is y2 = 0.35 ∗ 0.65 = 0.2275; the observed value
of 0.2278 is an average over twenty trials. The deviance from the
theoretical expectation is acceptable. )
When h < 1.0, the y2 values are greater, and when h > 1.0, the y2
values are smaller. In fact, y2 ranges from y2 = 0.301 (when h = 0.8)
down to y2 = 0.151 (when h = 1.8). These again are expected results.
A separate document will address the theoretical expectations in more
detail. In brief, when h < 1.0, then ε1 < 0.0, meaning that the
interaction enthalpy parameter ε1 is negative. When ε1 is negative,
the enthalpy is decreased by increasing y2 , as the interaction enthalpy
ε1 multiplies the term (2y2 − y1 − y3 ). Thus, maximizing y2 is expected
when h < 1.0.
There is a limit as to how far y2 can be increased; presumably
it can approach 0.5, however, that would mean that the units were
arranged in a strict checkerboard manner; that there were no instances
of like-near-like at all. This is rather difficult to achieve; both in
creation of highly-ordered systems, and in this particular code, which
uses a simplistic find-and-flip strategy.
As previously noted, when h > 1.0, the y2 values are smaller. This
is the case where ε1 > 0.0, and system enthalpy is decreased when y2
is made smaller. Thus, the system moves more towards a like-with-like
configuration (increasing y1 and y3 ); maximizing the size of the various
“islands,” and decreasing the size of their borders (minimizing y2 ).
There is a practical limit as to how far y2 can be decreased; there
will always be a border area between the state A islands (or even a
single, massive state A continent) and the surrounding sea of state
22
Figure 14: Data table containing configuration variable and thermodynamic
values for the case where x1 = 0.35 and x2 = 0.65, and where the interaction
enthalpy parameter h ranges as h = 0.8..1.8. See detailed explanation of
results within Section 4.
B units. This means that y2 will not get close to zero. The actual
practical limit for y2 will actually depend on the total system size
(total number of nodes), because the border area will progressively
decrease (although not disappear) as more and more islands join to
become continents. Thus, the value of y2 < 0.157, which occurs when
h ≥ 1.2, is not surprising.
Once y2 is pushed to a suitably small value, it becomes increasingly
difficult for the simple find-and-flip strategy to (randomly) find nodes
where the flip will accomplish a free energy reduction. This is likely
why there is general stability in the y2 values beyond h ≥ 1.2; there are
simply not that many nodes where the flip will do much good, keeping
in mind that two nodes (each in a different state) have to be flipped in
order to maintain the x1 value.
Thus, the preliminary conclusion is that free energy minimization is
being accomplished, and that the y2 values are behaving as expected.
23
4.1.2
Validation support: delta results
Again referencing Figure 13, we examine the curve for delta (shown
in cyan), defined as (2y2 − y1 − y3 ), which is the actual term that is
multiplied by ε1 to achieve the interaction enthalpy term. (The delta
curve is shown in dark green in this figure.) This curve behaves as
expected.
In particular, we note that there is a nearly linear behavior in the
region between h = 0.8 and h = 1.3. When h = 0.8, delta = 0.2035
(according to the data table shown in Figure 14. When h = 1.2, delta
= -0.3730. When h = 1.0, we would expect that there would be purely
probabilistic distribution of units into their configurations, and thus
expect that y1 = 0.35 ∗ 0.35 = 0.1225, y3 = 0.65 ∗ 0.65 = 0.4225,
and y2 = .2275 (as mentioned earlier). We would have then that
2y2 − y1 − y3 = 2 ∗ .2275 − 0.1225 − 0.4225 = −0.090. The actual value
is delta = -0.0887, which is acceptably close.
Similar arguments hold for the expected and observed values of
delta as did for y2 in the preceding discussion.
We again note that the values for delta level out as h increases
beyond 1.2; this is because there are not that many units that the
simple find-and-flip strategy can easily find. In particular, we note
that the z3 value at h ≥ 1.3 is typically around z3 = 0.04, which is
very small.
In particular, we observe that this z3 value indicates that we have
pushed the system to its limit for minimizing z3 , which is the A-A-B
configuration. This z3 value indicates a border of a rather large island
of state A units in a sea of B units. Specifically, for the 256-unit
system that is the subject for this investigation, when z3 = 0.04, then
N = 0.4 ∗ 256/2 = 102.4/2 = 51 triplets involve border units around
islands / continents of state A. This is approximately 1/5th of the
total number of units available. This suggests that we have pushed
the system about as far as it can go. Of course, a visual inspection
of the resulting grid would be enormously useful in confirming these
assessments. This will be included in a subsequent document.
4.1.3
Validation support: thermodynamic results
The enthalpy is maximum when h = 0, which is to be expected. As we
minimize free energy, we minimize enthalpy. As soon as we introduce
some non-zero interaction enthalpy, we have an opportunity to adjust
the configuration values (specifically the yi , as just discussed) to lower
24
the enthalpy.
The entropy is similarly at a maximum (neg-entropy is at a minimum) when h = 0. The negative entropy increases for non-zero values
of h, as expected.
We particularly note that in the vicinity of h = 0, or more generally,
in the range of 0.8 ≤ h ≤ 1.3, the variances in the entropy and enthalpy
are approximately on the same scale; one does not appreciably dwarf
the other.
When we move beyond h ≥ 1.3, we find that the enthalpy term
strongly dominates the entropy, and thus dominates the free energy. It
does this because we are increasing the value of the interaction enthalpy
coefficient, ε1 , and not because we are gaining any appreciable difference
in the configuration values. As noted in the previous discussions, these
values have more-or-less stabilized in this range.
Thus, increasing h beyond h = 1.3 does not serve any useful value,
suggesting a practical bound on h-values for this kind of system.
Our actual and practical choices for the h-values should be based
on the kind of behavior that we want to see in the configuration values.
For modeling brain-like systems, we will most likely want h ≥ 0, as that
induces like-with-like clustering, which seems to characterize certain
neural collectives.
5 Code included in this V&V description
• 2D-CVM-perturb-expt-1-2b-2018-01-12.py - perturbation
analysis with user-specifiable (in **main**) values for x1 , h,
numT rials, and many other parameters.
Code Availability: All code referenced here will be made available on a public GitHub repository (https://github.com/ajmaren/2DCluster-Variation-Method) after a short delay from initial publication
of this V&V document. This and related code will be supported by
extensive documentation, which will also be placed in the GitHub
repository. Anyone desiring access to the original code prior to its
placement on a public GitHub repository should contact A.J. Maren
at: [email protected].
Copyright: All code referenced here has been independently develeoped by A.J. Maren. A.J. Maren holds the copyright to both the
25
code and this V&V document itself. arXiv is granted non-exclusive
and irrevocable license to distribute this article.
References
[1] A. Maren, “Free energy as driving function in neural networks,” in
Symposium on Nonlinear Theory and Its Applications, (Hawaii),
pp. 89–93, December 5-10 1993. doi:10.13140/2.1.1621.1529.
[2] A. Maren, E. Schwartz, and J. Seyfried, “Configurational entropy
stabilizes pattern formation in a hetero-associative neural network,”
in IEEE Int’l Conf. SMC, (Chicago, IL), pp. 89–93, October 1992.
doi:10.1109/ICSMC.1992.271796.
[3] A. Maren, “The cluster variation method: a primer for neuroscientists,” Brain Sciences, vol. 6, no. 4, p. 44, 2016.
[4] K. Friston, M. Levin, B. Sengupta, and G. Pezzulo,
“Knowing one’s place: a free-energy approach to pattern
regulation,” J. R. Soc. Interface, vol. 12, p. 20141383,
2015.
doi:10.1098/rsif.2014.1383; available online at:
http://dx.doi.org/10.1098/rsif.2014.1383.
[5] K. Friston, “The free-energy principle: a unified brain
theory?,” Nat. Rev. Neurosci., vol. 11, p. 127–138, 2010.
doi:10.1038/nrn2787.
[6] K. Friston, “Life as we know it,” Journal of The Royal Society
Interface, vol. 10, no. 86, 2013.
[7] M. J. Beal, Variational algorithms for approximate Bayesian
inference. PhD thesis, University College London, 2003. PDF:
http://www.cse.buffalo.edu/faculty/mbeal/papers/beal03.pdf.
[8] R. Kikuchi, “A theory of cooperative phenomena,” Phys. Rev.,
vol. 81, no. 81, p. 988, 1951.
[9] A. Maren, “The Cluster Variation Method II: 2-D grid of zigzag
chains: Basic theory, analytic solution and free energy variable
distributions at midpoint (x1 = x2 = 0.5),” Tech. Rep. THM
TR2014-003 (ajm), Themasis, 2014. doi:10.13140/2.1.4112.5446.
[10] R. Kikuchi and S. Brush, “Improvement of the Cluster Variation
Method,” J. Chem. Phys., vol. 47, p. 195, 1967.
26
| 9cs.NE
|
arXiv:1711.10789v1 [cs.LG] 29 Nov 2017
Efficient exploration with
Double Uncertain Value Networks
Thomas M. Moerland, Joost Broekens and Catholijn M. Jonker
Department of Computer Science
Delft University of Technology, The Netherlands
{T.M.Moerland,D.J.Broekens,C.M.Jonker}@tudelft.nl
Abstract
This paper studies directed exploration for reinforcement learning agents by tracking uncertainty about the value of each available action. We identify two sources of
uncertainty that are relevant for exploration. The first originates from limited data
(parametric uncertainty), while the second originates from the distribution of the
returns (return uncertainty). We identify methods to learn these distributions with
deep neural networks, where we estimate parametric uncertainty with Bayesian
drop-out, while return uncertainty is propagated through the Bellman equation as a
Gaussian distribution. Then, we identify that both can be jointly estimated in one
network, which we call the Double Uncertain Value Network. The policy is directly
derived from the learned distributions based on Thompson sampling. Experimental
results show that both types of uncertainty may vastly improve learning in domains
with a strong exploration challenge.
1
Introduction
Reinforcement learning (RL) is the dominant class of algorithms to learn sequential decision-making
from data. In RL we start with zero prior knowledge and need to actively collect our own data.
Therefore, we should not settle on a policy too early, instead of trying out actions we have not
properly explored yet. However, we neither want to continue exploring sub-optimal actions, when we
already know what is best. This challenge is known as the exploration/exploitation trade-off.
Most state-of-the-art reinforcement learning implementations use undirected forms of exploration,
such as -greedy or Boltzmann exploration. These methods act on point estimates of the mean actionvalue, usually applying some random perturbation to avoid only selecting the currently optimal action.
However, these undirected methods are known to be highly inefficient (Osband et al., 2014). By only
tracking point estimates of the mean state-action value, these algorithms lack the information to, for
example, discriminate between an action that has never been tried before (and requires exploration)
and an action that has been tried extensively and deemed sub-optimal (and can be avoided).
A natural solution to this problem originates from tracking uncertainties/distributions. The intuition
is that with limited data and large uncertainty there is reason to explore, while narrow distributions
naturally transfer to exploitation (see Appendix C for a detailed illustration). For this work we identify
two types of uncertainties/distributions that are interesting for exploration:
• Parametric uncertainty: This is the classical statistical uncertainty which is a function of the
number of available data points. The cardinal example is the posterior distribution of the
mean (action-value).
• Return uncertainty: This is the distribution over returns from a state-action pair given the
policy. For this work we focus on deterministic domains, which makes the return distribution
entirely induced by the (exploratory) stochastic policy.
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
We argue that - for deterministic environments - we can explore by acting probabilistically optimal
with respect to both distributions (see Section 3). We identify neural network methods to estimate
each of them separately, and subsequently show that both can be combined in one network, which we
call the Double Uncertain Value Network (DUVN). To the best of our knowledge, we are the first to
1) distinguish between uncertainty due to limited data (parametric) and uncertainty over the return
distribution, 2) propagate both through the Bellman equation, 3) track both with neural networks (i.e.,
high-capacity function approximators), and 4) use both to improve exploration.1
The remainder of this paper is organized as follows. In Section 2 we provide a general introduction
to Bayesian deep learning and distributional reinforcement learning. In Section 3, we discuss
parametric and return uncertainty, and identify their potential for exploration. Section 4 discusses
their implementations for policy evaluation with neural networks, and also discusses how to derive
a policy from the learned distributions based on Thompson sampling. Sections 4, 5 and 6 show
experimental results, discuss future work, and draw conclusions, respectively.
2
Pre-liminaries
2.1
Bayesian deep learning
Bayesian neural networks (MacKay, 2003) represent the uncertainty in the model through posterior
distributions over the model parameters. Assume we observe some random variables X and Y and
are interested in the conditional distribution p(Y |X). We introduced a neural network pφ (Y |X) with
parameters φ ∈ Φ to estimate this conditional distribution. In the Bayesian setting, we treat the model
parameters φ as random variables themselves. Given an observed dataset H, we may use the posterior
distribution over model parameters p(φ|H) to obtain the posterior predictive distribution
p(y ? |x? , H) =
Z
p(y ? |x? , φ)p(φ|H)dφ
(1)
for a new observed datapoint x? . In the non-linear neural networks of practical interest, the posterior
distribution p(φ|H) is analytically intractable. Gal and Ghahramani (2016) showed that the wellknown empirical procedure drop-out actually produces a Monte-Carlo approximation to Eq. 1,
providing samples from the posterior predictive distribution by simply retaining drop-out during test
time (prediction). We use this technique in this paper, and discuss alternative methods for Bayesian
inference in neural networks in the Future work section.
2.2
Distributional reinforcement learning
In reinforcement learning (RL) (Sutton and Barto, 1998) agents are studied that interact with an unknown environment with the goal to optimize some long-term performance measure. The framework
adopts a Markov Decision Process (MDP) given by the tuple {S, A, T , R, γ}. At every time-step
t we observe a state st ∈ S and pick an action at ∈ A = {1...NA }, for NA available discrete
actions. The MDP follows the transition dynamics st+1 = T (·|st , at ) ∈ S and returns rewards
rt = R(st , at ) ∈ R. For this work, we assume a discrete action space and deterministic transition
and reward functions.
We act in the MDP according to a stochastic policy π, i.e. a ∼ π(·|s) ∈ P(A). The discounted return
Z π (s, a) from a state-action pair (s, a) is a random process given by
Z π (s, a) =
∞
X
γ t rt ,
st+1 = T (·|st , at ), at+1 ∼ π(·|st+1 ), s0 = s, a0 = a
(2)
t=0
for discount factor γ ∈ [0, 1]. We emphasize that the return Z π is a random variable, where the
distribution of Z π is induced by the stochastic policy (as we assume a deterministic environment). We
1
As a side contribution, we introduce the Initial Return Entropy (IRE) as a measure of task exploration
difficulty. See Appendix B.
2
may rewrite equation 2 into a recursive form, known as the distributional Bellman equation (omitting
the π superscript from now on):
Z(s, a) = rt + γZ(s0 , a0 ),
s0 = T (·|s, a), a0 ∼ π(·|s0 ).
(3)
Note that the equality sign represents distributional equality here (Engel et al., 2005). We are now
ready to define the action-value function. Denote by Eπ the expectation over all traces induced by
the policy π. Applying this operator to Z(s, a) defines the state-action value Q(s, a) = Eπ [Z(s, a)].
Applying this operator to Eq. 2 gives
Q(s, a) = rt + γEs0 =T (·|s,a),a0 ∼π(·|s0 ) [Q(s0 , a0 )]
(4)
which is known as the Bellman equation (Sutton and Barto, 1998). Most RL papers actually start-off
from Eq. 4. We present the current introduction to emphasize that the mean action value Q(s, a) is a
quantity that we estimate by sampling from an underlying return distribution p(Z|s, a).2
We approximate the action-value (distribution) with a deep neural network. We write Qφ (s, a) for a
network predicting a (point estimate) action-value, and pφ (Z|s, a) for a network approximating the
entire return distribution. To learn the state-action value RL algorithms follow variants of a scheme
known as generalized policy iteration (GPI) (Sutton and Barto, 1998). GPI iterates between policy
evaluation, in which we calculate a new estimates Ψ(s, a) of the state-action value based on (new)
sample data (e.g., for one-step SARSA Ψ(s, a) = r(s, a) + γQ(s0 , a0 )), and policy improvement, in
which we use the estimate Ψ(s, a) to improve the policy (whether with a value-based, actor-critic or
policy gradient algorithm).
3
Distributional perspective on exploration
We will now argue for a probabilistic perspective on value functions and exploration. There are two
distributions that might be useful from an exploration point of view: 1) the statistical parametric
uncertainty of the mean action value, and 2) the distribution of the return.
Parametric uncertainty of the mean Given a policy the state-action value Q(s, a) is a scalar
number by definition, as it is an expectation over all possible future traces. However, from a statistical
point of view it makes sense to treat our estimate of Q(s, a) as a random variable, as we need to
approximate it from a finite number of samples. We call this the parametric uncertainty.
Parametric exploration, i.e. acting optimistic with respect to the uncertainty of the mean action-value,
has been very successful in the bandit setting. However, it has been sparsely applied to RL (see
Appendix A for related work). We believe this is due to a fundamental complication regarding
uncertainties in RL, which has only been identified by Dearden et al. (1998) before. Bandits are
one-step decision problems with pay-offs originating from a stationary distribution, which makes the
value approximation an ordinary supervised learning problem. However, in RL the target distribution
is highly non-stationary. A standard target like r + γQ(s0 , a0 ) falsely assumes that Q(s0 , a0 ) is known,
while it is actually uncertain itself. Therefore, repeatedly visiting a state-action pair should not
makes us certain about its value if we are still uncertain about what to do next. In other words: the
state-action value certainty depends on the future policy certainty. Standard parametric uncertainty
cannot account for this problem (the ‘local’ parametric uncertainty will converge as if it is supervised
learning), and we somehow need to propagate the uncertainty of future state-action pairs’ value (the
‘global’ uncertainty) back through the Bellman equation. An illustration is seen in Fig. 1b right,
where the uncertainty in φ influences both the current and future value estimates (ignoring the Z
distributions in that graph for now, as the need to propagate the parametric uncertainty p(φ) over
timesteps already applies when we learn mean action-values).
2
We empirically observe that the shape of this return distribution strongly differs between domains. This
matters because the shape of the return distribution also influences how easily we can estimate its expectation, or
some other quantity like an upper confidence bound, from samples. For example, a long, thin right tail in the
return distribution - as frequently the case in RL with only a few ‘good’ traces - may give our mean estimate high
variance (it would actually need importance sampling). In Appendix B we visualize return distributions for some
well-known domains, and also introduce the Initial Return Entropy as a measure of task exploration difficulty.
3
Figure 1: a) Three types of neural networks with different uncertanties/probabilitiy distributions. Circles are
probabilistic nodes. Left: parametric uncertainty over the mean action-value. Middle: propagating (return)
distributions for point estimate parameters. Right: parametric uncertainty over propagating distribution (=
Double Uncertain Value Network). b) Illustration of propagating distributions. Subscripts identify unique
state-action pairs. We initialize all state-action pairs with a prior parametric uncertainty p(φ) and prior output
distribution pφ (Z). Then, for a new observed transition, we want to update our estimates of pφ (Z) at the current
state action pair by propagating the distribution of the next node p(Z 0 ) through the Bellman operator T (instead
of just propagating the mean). For this work, we consider two quantities to propagate: i) the return distribution
0
at the next node
R pφ (Z0 ) (for point estimate φ), or ii) the parametric uncertain return distribution at the next
0
node p(Z ) = pφ (Z )p(φ|H)dφ. Arrows point backwards because we focus on the direction of uncertainty
propagation/back-up (which runs in the different direction than our exploration).
Return distribution Standard RL, and also the parametric uncertainty introduced above, usually
deal with the mean action-value Q(s, a). However, from an exploration point of view, it makes
more sense to learn the full return distribution p(Z|s, a). Note that we still focus on deterministic
environments. Therefore, the distribution over returns is solely induced by our own policy. As we
may modify our own policy, it makes sense to act optimistically with respect to the return distribution
we observe. As an illustration, consider a state-action pair with particular mean action value estimate
Q(s, a). It really matters whether this average originates from a highly varying return, or from
consistently the same return. It matters because our policy may influence the shape of this distribution,
i.e. for the highly varying returns we may actively transform the distribution towards the good returns.
In other words, what we really care about in deterministic domains is the best return, or the upper end
of the return distribution.3
It turns out that both challenges focus around propagating either parametric uncertainty and/or return
distributions through the Bellman equation (Fig 1b). The overall idea is to memorize the propagating
global MDP uncertainties in a neural network, which makes them locally available at action selection
time. We thereby avoid the need for any forward planning (to get global information), and our
approach is entirely model-free.
4
Double Uncertain Value Networks
4.1
Policy evaluation
We now discuss three probabilistic policy evaluation approaches that incorporate the uncertainties
introduced in the previous section: 1) (local) parametric uncertainty only, 2) return distribution only,
and 3) both combined. The respective network structures are illustrated in Fig. 1a. Implementation
details are provided in Appendix E.
Parametric uncertainty only To estimate our parametric uncertainty we may use any type of
Bayesian inference method suitable for neural networks. For this paper we consider the Bayesian
dropout (Gal et al., 2016), as it has a very simple practical implementation (see Sec. 2.1). This
gives
us a sample from the posterior predictive distribution of the mean action value: p(Q|s, a, H) =
R
Qφ (s, a)p(φ|H)dφ. The associated network structure is visualized in Figure 1a, left.
3
For stochastic domains the return distribution has additional noise for which we do want to act on the
expectation.
4
Return distribution only We next consider the problem of learning return distributions instead
of mean action-values. For this work we will assume that the return distribution p(Z|s, a) can be
approximated by a Gaussian. Therefore, we modify our neural network to output the distribution
parameters µZ (s, a) and σ Z (s, a), where clearly µZ (s, a) = Q(s, a). Note that the network parameters φ are point estimates now. The associated network structure is visualized in Figure 1a, middle.
During policy evaluation we need to estimate distributional targets instead of point estimate targets. We will construct bootstrap estimators based on the distributional Bellman equation (Eq. 3).
The derivation for the mean µZ (s, a) = Q(s, a) is well-known from standard RL, so we focus on
propagating the return standard deviation through the distributional Bellman equation:
h
i
h
i
h
i
X
X
Sd p(Z|s, a) = Sd r(s, a) + γ
π(a0 |s0 )p(Z|s0 , a0 ) = γ
π(a0 |s0 )Sd p(Z|s0 , a0 ) (5)
a0 ∈A
a0 ∈A
because γ ≥ 0, π(a|s) ≥ 0, and we assume the next state distributions are independent so we may
ignore the covariances.4 We see that the standard deviation of p(Z|s, a) is a linear combination of
the standard deviations p(Z|s0 , a0 ) (one timestep ahead), reweighted by the policy probabilities and
shrunken by γ. We approximate the sum over the policy probabilities π(a0 |s0 ) by sampling from our
policy (as is the usual solution in RL, which will be right in expectation over multiple traces). The
network may then be trained to move the current predictions closer to these targets, for example with
a squared loss
2
2
0 0
Z
Z 0 0
Z
L(φ) = r(s, a) + γµZ
(s
,
a
)
−
µ
(s,
a)
+
γσ
(s
,
a
)
−
σ
(s,
a)
φ
φ
φ
φ
(6)
where we as usual fix the bootstrap predictions at (s0 , a0 ), i.e. the training gradients w.r.t. φ are
blocked there. This approach can be seen as a form of analytic approximate return propagation with a
(heuristic) distributional loss (see Appendix A.2 for other distributional losses). Similar ideas with
approximate return propagation were recently explored with discrete network output distributions
(Bellemare et al., 2017), which may also accommodate for propagating multimodality.
A second, more simple propagation method which we also experimented with is sampling-based
propagation. In that setting we sample M values zm (s0 , a0 ) ∼ pφ (Z|s0 , a0 ), push these through
the Bellman operator to construct Ψm (s, a) = r(s, a) + γzm (s0 , a0 ), and train our network on this
collection of samples with, e.g., a maximum likelihood loss. This may require more samples and be
less accurate, but it will also work for complicated network output distributions (like deep generative
models) for which analytic propagation and projection is infeasible. Results of this approach are not
shown, but were comparable to the results with approximate return propagation shown in Section 5.
Parametric uncertainty over return distributions We finish with the observation that both ideas
may actually naturally be combined in one function approximator (Fig 1a, right). Note that we can
now propagate both the return distribution and its parametric uncertainty at the next timestep, i.e.
we are effectively propagating uncertain return distributions (the parametric uncertainty over the
network output distribution). Starting from a sampled transition now, we want to propagate the return
distribution weighted over the parametric uncertainty at the next timestep:
Z(s, a) =
Z h
Z
i
r + γZφ (s0 , a0 ) p(φ|H)dφ = r + γ Zφ (s0 , a0 )p(φ|H)dφ
(7)
Besides that, the same distributional Bellman propagating machinery applies as above.5 We refer to
the general mechanism of uncertainty propagation (parametric, return or both) as Bellman uncertainty.
The appearance of the network, with both uncertainty over the network parameters φ and over the
output distribution to track the propagating (uncertain) return distributions, makes us refer to it as the
For random variables X, Y and scalar constants a, b, c we have: Var[a + bX + cY ] = b2 Var[X] +
c Var[Y ] + 2bc Cov[X, Y ].
5
0
0
Z 0
0
Sample from p(φ|H) at the next time-step, make network predictions µZ
φ (s , a ) and σφ (s , a ), and do
the Bellman propagation. Repeated sampling of φ does Monte Carlo integration over p(φ|H), as a numerical
integration like in Dearden et al. (1998) is infeasible for the neural network setting.
4
2
5
Double Uncertain Value Network (DUVN) (Fig 1a, right). The intuition is that during early learning
we will mostly be propagating uncertainty, while with converging distributions we will eventually
start propagating true return distributions.
In summary, we identified three types of probabilistic policy evaluation algorithms (with the three
associated network structures visualized in Fig. 1a):6
R
1. The (local) parametric uncertainty of the mean value: p(Q|s, a, H) = Qφ (s, a)p(φ|H)dφ.
2. The (propagating) distribution of the return: pφ (Z|s, a, H) (with point parameters φ).
R
3. Both, (propagating) uncertain return distr.: p(Z|s, a, H) = pφ (Z|s, a, H)p(φ|H)dφ.
4.2
Policy improvement
We now describe how to use any of these distributions to naturally balance exploration versus
exploitation, based on Thompson sampling (Thompson, 1933) (see Appendix C as well). To
generalize notation, we introduce a new random variable Θ with distribution p(Θ|s, a) to capture any ofQ
the three policy evaluation distributions introduced in the previous section. We write
p(Θ|s) = a? ∈A p(Θ|s, a? ) for the joint action-value distribution in a state s, where we assume
the posterior distributions per action are independent. Thompson sampling selects action a with
probability equal to:
Z
π(a|s) =
p(Θa > Θa? 6=a )p(Θ|s)dΘ
(8)
where Θa = Θ(s, a) and Θa? 6=a notational convention for Θ(s, a? )∀a? ∈ A, a? 6= a. In words, we
choose action a with probability equal to the probability that the specific action is the optimal one
when averaging over all uncertainty in the joint distribution p(Θ|s). The practical implementation of
Thompson sampling is very simple, as we may just sample from p(Θ|s, a) for every a and argmax
over these values:
1. Sample φ ∼ p(φ) (or equivalently a dropout mask).
2. Sample Z(s, a? ) ∼ pφ (Z|s, a? ) ∀ a? ∈ A.
3. Select a = arg maxa? ∈A Z(s, a? ).
If we do not consider parametric uncertainty, then we ignore the first sampling step and just use the
current parameter point estimates. If we do not consider the Bellman uncertainty, then we replace the
second sampling step with a deterministic prediction Qφ (s, a).
Thompson sampling is not the only possible choice to make decisions under uncertainty, but it has
shown good empirical performance in the bandit literature (Chapelle and Li, 2011). It naturally
performs policy improvement, as it gradually starts to prefer the better actions when the distributions start narrowing/converging. We thereby hope to improve on the instability of greedy policy
improvement (see also Bellemare et al. (2017)) or undirected exploration. Ideally, the uncertain return
distribution would gradually narrow and for a deterministic environment eventually converge to a
Dirac distribution on the optimal value function.
5
Experiments
We now evaluate the different types of probabilistic policy evaluation in combination with Thompson
sampling exploration. We refer to Thompson sampling on the three types of discussed policy evaluation as parametric exploration, return exploration, and uncertain return exploration. Experimental
details are provided in Appendix E.
We first consider the Chain domain (Appendix D, Figure 6). The domain consists of a chain of states
of length N , with two available actions at each state. The only trace giving a positive, non-zero
6
We could think of another algorithm that does propagate (i.e., has a probabilistic network
output), but only
R
propagates the parametric uncertainty of the mean at the next time-step p(Q0 |H) = Q0φ p(φ|H)dφ (and not
the entire return distribution). We did not come up with such an algorithm, but concurrently with our work,
O’Donoghue et al. (2017) did focus on this problem. See Related Work.
6
Figure 2: Learning curves on Chain domain for Thompson sampling on parametric uncertainty, return distribution and uncertain return distribution versus -greedy exploration ( = 0.05). Plots progress row-wise for
increased depth of the Chain, i.e. increased exploration difficulty. Note that the correct action at each state in the
chain is initialized at random (i.e. not always action 2, as in the visualization in Fig. 6). Results averaged over 5
repetitions.
reward is to select the ‘correct’ action at every step. The correct action per step is determined at
domain initialization by sampling from a uniform Bernoulli. The domain has a strong exploration
challenge, which grows exponentially with the length of the chain (see Appendix D).
Learning curves for the Chain domain are shown in Fig. 2, for different lengths of the chain. First
of all, we note that the -greedy strategy does not learn in this domain at all (not even for the short
length). The three probabilistic approaches do explore, with best performance for the uncertain return
exploration. In the longest chain, of length 100, all probabilistic exploration methods also get trouble
solving the domain. However, they do see some rewards, which makes us hypothesize this could be
an issue of stabilization (more than that the exploration does not work). See Appendix D.1 for results
when the correct action is always the same, as in the original variants of this problem (Osband et al.,
2016).
We next test our method on a set of tasks from the OpenAI Gym repository (Fig. 3). We see that
our exploration methods manage to learn on all domains. The achieved end policies all reflect
good policies on each problem. -greedy exploration is a bit unstable on some domains (CartPole,
LunarLander), but generally performs reasonable as well. We note that the uncertainty exploration
methods, which have a completely different exploration mechanism compared to -greedy exploration,
never really perform worse on these domains.
We hypothesize these domains have too much structure and are not challenging enough to show
the same exploration difference as seen for the Chain domain. Future work should address more
challenging (high-dimensional) exploration problems. We also want to stress that probabilistic
exploration will not always outperform undirected methods, especially not on domains with relatively
simple exploration. Uncertainty methods will generally create a cautious agent, that first wants to
properly verify all parts of the domain. In contrast, undirected exploration agents may exploit sooner,
which can be beneficial in domains with non-deep exploration (i.e., with quick rewards).
6
Future work
We identify several directions for future work:
1. Other types of Bayesian inference in neural networks (for parametric uncertainty): we
hypothesize that the Bayesian drop-out may be too unstable and tedious to tune, as we
sometimes observed in our experiments as well. Potentially, different methods to approximate the posterior over the network parameters (e.g., Welling and Teh (2011)) may improve
estimation of parametric uncertainty.
2. More expressive output distributions (for Bellman uncertainty propagation): for this work
we only experimented with Gaussian distributions for propagation. Recently, Bellemare et al.
(2017) studied return distribution propagation with categorical distributions, which more
naturally accommodate for multi-modality (see Related Work as well). Another extension
could involve more expressive continuous network distributions, e.g. based on conditional
variational inference (Moerland et al., 2017).
7
Figure 3: Learning curves for parametric exploration, return exploration and uncertain return exploration on
different OpenAI Gym environments. Results averaged over 5 repetitions.
3. Continuous action-spaces: the current implementations only focussed on discrete action
spaces, where Thompson sampling can easily be applied by maintaining a distribution per
action and enumerating all actions for action selection. Extension to the continuous setting
would require either directly propagating policy uncertainty, or learning a parametric policy
whose distribution mimics the uncertainty in the value function.
4. Stochastic environments: this paper entirely focussed on domains with deterministic
reward and transition functions, which makes the return distribution only induced by the
stochastic policy. In stochastic domains the return distribution will have additional noise for
which we do want to act on the expectation, to prevent continuing to act optimistically with
respect to something we can’t influence.
Finally, we want to stress that the RL algorithms in this paper are entirely model-free. The uncertainty
theme also appears in model-based RL, where it is useful/necessary to track the uncertainty on an
estimated transition and/or reward function (Deisenroth and Rasmussen, 2011; Depeweg et al., 2017).
This parametric model uncertainty is different from the parametric value/policy uncertainty studied
in this work, but our ideas may be extended to the model-based setting as well (which would add
another source of uncertainty).
7
Conclusion
This paper introduced Double Uncertain Value Networks (DUVN), which, to the best of our knowledge, is the first algorithm that 1) distinguishes between uncertainty due to limited data (parametric)
and uncertainty over the return distribution, 2) propagates both through the Bellman equation, 3)
tracks both with neural networks (i.e., high-capacity function approximators), and 4) uses both to
improve exploration. We implemented the DUVN algorithm with Bayesian dropout for the parametric uncertainty and a Gaussian distribution for the Bellman uncertainty propagation. The main
appeal of this implementation is its simplicity: any deep Q-network implementation can be easily
extended as in this work by adding drop-out to the neural network layers and specifying a Gaussian
output distribution instead of a mean-only prediction. This should take no more then a few lines
of code in most automatic differentiation software packages. We showed that, even for the vanilla
implementation, we at least match or improve undirected exploration performance on a variety of
problems, and drastically improve performance on an exploration heavy domain (Chain). We believe
further improvements in the distributional approach to RL, e.g. with more expressive network output
distributions that capture multi-modality, is a promising direction for RL exploration research.
8
References
Auer, P., Cesa-Bianchi, N., and Fischer, P. (2002). Finite-time analysis of the multiarmed bandit problem.
Machine learning, 47(2-3):235–256.
Bellemare, M., Srinivasan, S., Ostrovski, G., Schaul, T., Saxton, D., and Munos, R. (2016). Unifying countbased exploration and intrinsic motivation. In Advances in Neural Information Processing Systems, pages
1471–1479.
Bellemare, M. G., Dabney, W., and Munos, R. (2017). A distributional perspective on reinforcement learning.
arXiv preprint arXiv:1707.06887.
Blundell, C., Cornebise, J., Kavukcuoglu, K., and Wierstra, D. (2015). Weight uncertainty in neural networks.
arXiv preprint arXiv:1505.05424.
Brafman, R. I. and Tennenholtz, M. (2002). R-max-a general polynomial time algorithm for near-optimal
reinforcement learning. Journal of Machine Learning Research, 3(Oct):213–231.
Chapelle, O. and Li, L. (2011). An empirical evaluation of thompson sampling. In Advances in neural
information processing systems, pages 2249–2257.
Dearden, R., Friedman, N., and Andre, D. (1999). Model based Bayesian exploration. In Proceedings of the
Fifteenth conference on Uncertainty in artificial intelligence, pages 150–159. Morgan Kaufmann Publishers
Inc.
Dearden, R., Friedman, N., and Russell, S. (1998). Bayesian Q-learning. In AAAI/IAAI, pages 761–768.
Deisenroth, M. and Rasmussen, C. E. (2011). PILCO: A model-based and data-efficient approach to policy
search. In Proceedings of the 28th International Conference on machine learning (ICML-11), pages 465–472.
Depeweg, S., Hernández-Lobato, J. M., Doshi-Velez, F., and Udluft, S. (2017). Uncertainty Decomposition in
Bayesian Neural Networks with Latent Variables. arXiv preprint arXiv:1706.08495.
Engel, Y., Mannor, S., and Meir, R. (2003). Bayes meets Bellman: The Gaussian process approach to temporal
difference learning. In Proceedings of the 20th International Conference on Machine Learning (ICML-03),
pages 154–161.
Engel, Y., Mannor, S., and Meir, R. (2005). Reinforcement learning with Gaussian processes. In Proceedings of
the 22nd international conference on Machine learning, pages 201–208. ACM.
Gal, Y. (2016). Uncertainty in deep learning. PhD thesis, PhD thesis, University of Cambridge.
Gal, Y. and Ghahramani, Z. (2016). Dropout as a Bayesian approximation: Representing model uncertainty in
deep learning. In international conference on machine learning, pages 1050–1059.
Gal, Y., McAllister, R. T., and Rasmussen, C. E. (2016). Improving PILCO with bayesian neural network
dynamics models. In Data-Efficient Machine Learning workshop, volume 951, page 2016.
Ghavamzadeh, M. and Engel, Y. (2007a). Bayesian actor-critic algorithms. In Proceedings of the 24th
international conference on Machine learning, pages 297–304. ACM.
Ghavamzadeh, M. and Engel, Y. (2007b). Bayesian policy gradient algorithms. In Advances in neural information
processing systems, pages 457–464.
Guez, A., Silver, D., and Dayan, P. (2012). Efficient Bayes-adaptive reinforcement learning using sample-based
search. In Advances in Neural Information Processing Systems, pages 1025–1033.
Houthooft, R., Chen, X., Duan, Y., Schulman, J., De Turck, F., and Abbeel, P. (2016). Vime: Variational
information maximizing exploration. In Advances in Neural Information Processing Systems, pages 1109–
1117.
Kearns, M. and Singh, S. (2002). Near-optimal reinforcement learning in polynomial time. Machine Learning,
49(2-3):209–232.
Kocsis, L. and Szepesvári, C. (2006). Bandit based monte-carlo planning. In ECML, volume 6, pages 282–293.
Springer.
MacKay, D. J. (2003). Information theory, inference and learning algorithms. Cambridge university press.
9
Mannor, S., Simester, D., Sun, P., and Tsitsiklis, J. N. (2004). Bias and variance in value function estimation. In
Proceedings of the twenty-first international conference on Machine learning, page 72. ACM.
Mannor, S., Simester, D., Sun, P., and Tsitsiklis, J. N. (2007). Bias and variance approximation in value function
estimates. Management Science, 53(2):308–322.
Mannor, S. and Tsitsiklis, J. (2011). Mean-variance optimization in Markov decision processes. arXiv preprint
arXiv:1104.5601.
Matthias, P., Rein, H., Prafulla, D., Szymon, S., Richard Y., C., Xi, C., Tamim, A., Pieter, A., and Marcin, A.
(2017). Parameter Space Noise for Exploration. arXiv preprint arXiv:1706.01905.
Moerland, T. M., Broekens, J., and Jonker, C. M. (2017). Learning Multimodal Transition Dynamics for
Model-Based Reinforcement Learning. arXiv preprint arXiv:1705.00470.
Morimura, T., Sugiyama, M., Kashima, H., Hachiya, H., and Tanaka, T. (2012). Parametric return density
estimation for reinforcement learning. arXiv preprint arXiv:1203.3497.
O’Donoghue, B., Osband, I., Munos, R., and Mnih, V. (2017). The Uncertainty Bellman Equation and
Exploration. arXiv preprint arXiv:1709.05380.
Osband, I., Blundell, C., Pritzel, A., and Van Roy, B. (2016). Deep exploration via bootstrapped DQN. In
Advances in Neural Information Processing Systems, pages 4026–4034.
Osband, I., Van Roy, B., and Wen, Z. (2014). Generalization and exploration via randomized value functions.
arXiv preprint arXiv:1402.0635.
Rasmussen, C. E., Kuss, M., et al. (2003). Gaussian Processes in Reinforcement Learning. In NIPS, volume 4,
page 1.
Sobel, M. J. (1982). The variance of discounted Markov decision processes. Journal of Applied Probability,
19(4):794–802.
Stadie, B. C., Levine, S., and Abbeel, P. (2015). Incentivizing exploration in reinforcement learning with deep
predictive models. arXiv preprint arXiv:1507.00814.
Sutton, R. S. and Barto, A. G. (1998). Reinforcement learning: An introduction, volume 1. MIT press Cambridge.
Tamar, A., Di Castro, D., and Mannor, S. (2016). Learning the variance of the reward-to-go. Journal of Machine
Learning Research, 17(13):1–36.
Thompson, W. R. (1933). On the likelihood that one unknown probability exceeds another in view of the
evidence of two samples. Biometrika, 25(3/4):285–294.
van Hoof, H., Tanneberg, D., and Peters, J. (2017). Generalized exploration in policy search. Machine Learning,
106(9-10):1705–1724.
Welling, M. and Teh, Y. W. (2011). Bayesian learning via stochastic gradient Langevin dynamics. In Proceedings
of the 28th International Conference on Machine Learning (ICML-11), pages 681–688.
White, D. (1988). Mean, variance, and probabilistic criteria in finite Markov decision processes: a review.
Journal of Optimization Theory and Applications, 56(1):1–29.
10
A
Related work
Exploration is a widely studied topic in reinforcement learning. We will discuss work based on
parametric (value/policy) uncertainty, return distributions/uncertainty, and add some context on
other exploration approaches (count-based/intrinsic motivation) and other uncertainty methods in RL
(uncertainty in model-based RL).
A.1
Parametric uncertainty of the mean
This research direction uses the uncertainty of the mean action value, either from a frequentist
or Bayesian direction, to direct exploration. Exploration is usually based on ’optimism under
uncertainty’. Parametric uncertainty has been extensively studied in the bandit setting, which are
environments with a single state, multiple actions and unknown, stochastic rewards. Succesful
approaches are UCB Auer et al. (2002), which acts on the upper confidence bound of a frequentist
confidence interval, and Thompson sampling Thompson (1933), which is also studied in this paper.
There are a few extensions of these ideas to the RL/MDP setting. The first occurrence of parametric
uncertainty in RL seems to be Bayesian Q-learning by Dearden et al. (1998). The authors use
tabular Q-learning with normal distributions and conjugate updating. They are also the only ones
that explicitly identify the necessity to propagate parametric uncertainty from future states. Their
exploration is based on either Thompson sampling (which they call Q-value sampling), while they
also consider myopic value of perfect information (VPI) as another exploration strategy.
Osband et al. (2014) extended these ideas to the linear function approximation setting with randomized
least-squares value iteration (RLSVI). In the neural network context, parametric uncertainty based
on variational inference was studied for bandits by Blundell et al. (2015). Gal and Ghahramani
(2016) studied the use of dropout uncertainty for parametric value uncertainty similar to our work,
but did not consider any propagation, nor the distribution over returns. Osband et al. (2016) also
studied parametric exploration in RL with neural networks, using the non-parametric bootstrap
(i.e., a frequentist approach to uncertainty estimation, not to be confused with the use of the term
‘bootstrapping’ in RL).
Concurrently with the present paper, O’Donoghue et al. (2017) also identified the need to propagate
parametric uncertainty over timesteps. Their approach is based on a variance estimate, which has a
similar role as the σ in our Gaussian uncertainty propagation. Their neural network implementation
derives the local parametric uncertainty estimates from the linearity of their last network layer and
frequentist uncertainty estimates known from linear regression. This contrasts to our Bayesian
approach to parametric uncertainty. Moreover, they still propagate uncertainty about the mean action
value, and do not consider the returns as in our paper.
There is more work that does track uncertainty for policy evaluation, but does not use these for policy
improvement / exploration. Most of these have focussed on Gaussian Process regression (Engel
et al., 2003). Rasmussen et al. (2003) uses two Gaussian Processes: one to track the parametric
uncertainty in the value, and a second one to model the uncertainty in the transition model. However,
the paper still uses a greedy policy improvement. The Gaussian Process approach was also extended
to continuous action spaces. There are actor-critic (Ghavamzadeh and Engel, 2007a) and policy
search (Ghavamzadeh and Engel, 2007b) algorithms that track the uncertainty in the gradients, but
again only to stabalize the update, not to direct exploration.
The idea of exploration based on parametric uncertainty also connects to the difference between
action space and parameter space / episode-based exploration (Matthias et al., 2017; van Hoof et al.,
2017). Most (undirected) exploration methods, like -greedy and Boltzmann, inject exploration noise
at the action space level. However, it can be beneficial to inject the noise at parameter level instead,
usually because it allows you to retain a particular noise setting over multiple steps (e.g. an entire
episode). The risk of action-space exploration noise is that the agent has to redecide at every timestep
and therefore cannot stick with an exploration decision. The effect might be jittering behaviour
between exploration and exploitation steps. This has also been identified as the challenge of ensuring
‘deep’ exploration Osband et al. (2016). We have not considered this problem in this paper, but it
could for example be implemented by fixing the dropout mask over an entire episode.
We also want to note that the exact same exploration problem occurs in classical (tree) search.
Succesful Monte Carlo Tree Search (MCTS) algorithms, like Upper Confidence Bounds for Trees
11
(UCT) Kocsis and Szepesvári (2006), act on the upper confidence bound of a frequentist confidence
interval of the value at each state-action pair. The overlap between reinforcement learning and (modelbased) search has been identified for long Sutton and Barto (1998), where RL i) does not assume
an a-priori known environment model, and ii) usually includes a parametric function approximator
to represent the value/policy, while search stores these in a tree structure (which is technically an
effective sparse form of tabular representation). But besides that, the same exploration themes appear
in both fields.
A.2
Return uncertainty
While the distributional Bellman equation (Eq. 3) is certainly not new Sobel (1982); White (1988),
nearly all RL research has focussed on the mean action-value. Most papers that do study the
underlying return distribution study the ’variance of the return’. Engel et al. (2005) learned the
distribution of the return with Gaussian Processes, but did not use it for exploration. Tamar et al.
(2016) studied the variance of the return with linear function approximation. Mannor and Tsitsiklis
(2011) theoretically studies policies that bound the variance of the return.
The variance of the return does not need to be used with ’optimism under uncertainty’, and actually
has more frequently been considered for risk-sensitive RL. In several scenarios we may want to
avoid incidental large negative pay-offs, which can e.g. be desastrous for a real-world robot, or in a
financial portfolio. Morimura et al. (2012) studied parametric return distribution propagation as well.
They do risk-sensitive exploration by softmax exploration over quantile Q-functions (also known as
the Value-at-Risk (VaR) in financial management literature). Their distribution losses are based on
KL-divergences (including Normal, Laplace and skewed Laplace distributions), which could be a
better distributional loss than the heuristic loss in Eq. 6. However, their implementations do remain
in the tabular setting.
Recently, Bellemare et al. (2017) theoretically studied the distributional Bellman operator. The authors
show that the operator is still a contraction in the policy evaluation setting, but not a contraction
in any distribution metric for the control setting. They hypothesize this is due to the ‘inherent
instability of greedy updates’ in the Bellman optimality operator. Their algorithm (called C51) uses a
categorical distribution to propagate returns distributions, which may more easily accommodate for
multimodality compared to the Gaussian distribution used in this work. C51 backs-up the complete
Bellman distributions, but they do not use these for exploration. Their methods nevertheless improves
over all other previous deep Q-networks on Atari games.
A.3
Count-based Exploration & Intrinsic Motivation
Count-based exploration uses a slightly different incentive for exploration, focussing or rewarding
regions of state-space that have not been visited (often). These ideas were extensively studied in the
tabula rasa setting, e.g. R-max Brafman and Tennenholtz (2002) and Explicit-Explore or Exploit
(E 3 ) Kearns and Singh (2002). Guez et al. (2012) explicitly plans ahead using Monte Carlo Tree
Search over uncertain transition dynamics models. Applications in high-dimensional domains include
Stadie et al. (2015) and Bellemare et al. (2016).
Intrinsic motivation generalizes this notion of novelty to any internal reward for domain-independent
characteristics, i.e. next to the domain-dependent external reward function. An example is rewarding
actions that decreases the parametric uncertainty in the transition model (Houthooft et al., 2016).
Alltogether, this class of exploration methods usually depends on the ability to learn good transition
models (from limited data), a problem which is not trivial itself (Deisenroth and Rasmussen, 2011;
Depeweg et al., 2017; Moerland et al., 2017).
A theoretical problem with count-based / intrinsic motivation approaches can be that they change the
RL objective itself. For example, bonuses on novelty might make an agent continue to visit a region
of state-space where the value functions are already very certain, yet not all states are frequently
visited yet (like continuously walking around a room to view it from all angles, which gives a new
visual state each time). Nevertheless, intrinsic motivation-based approaches hold the state-of-the-art
on challenging exploration problems like Montezuma’s Revenge (Bellemare et al., 2016).
12
A.4
Uncertainty in model-based RL
All work in this paper only considered model-free RL. However, similar issues with uncertainty
appear when learning the transition and/or reward function, known as ‘model-based RL’. Dearden
et al. (1999) was again the first to address this problem for the tabular setting. Mannor et al. (2004,
2007) studied two environment sources of variance that may influence the return distribution: the
‘internal variance’ due to a stochastic environment, and the ‘parametric (model) variance’ due to bias
in the environment model. Neither of these were considered in this work, but both may be added.
These ideas were studied in neural networks by Depeweg et al. (2017), who instead uses the terms
empistemic uncertainty for the model bias and aleatoric uncertainty for the inherent environment
noise/stochasticity. Their approach, which infers distributions on neural network parameters to
capture model bias, and uses expressive output distributions to capture true environment stochasticity,
actually has a similar structure as our Double Uncertain Value Network (in some sense, they learn
a ‘Double Uncertain Transition Network’). Of course, transition model learning does not involve
any uncertainty propagation (it is a well-defined supervised learning problem). Finally, Gal et al.
(2016) used Bayesian drop-out, as considered in this work for parametric value uncertainty, to track
parametric model uncertainty.
13
B
Initial Return Entropy (IRE) as a measure of initial exploration difficulty
Define the initial return distribution (IRD) pinit (Z) as the distribution over trace returns Z ∈ R when
sampling an initial state sinit from some initial state distribution and following a uniform random
policy from there on. For undirected exploration, the uniform policy is the best policy we can specify
until we start encountering varying returns. Define the initial return entropy (IRE) as the entropy of
this distribution:
Z
H(Z) =
pinit (Z) · log pinit (Z) dZ
(9)
We propose that the IRE is an interesting measure of the domain exploration difficulty, where
lower values indicate a higher exploration challenge. Figure 4 shows the IRD and IRE for various
domains from the OpenAI Gym repository. We see quite large differences between the shape of
these distributions. Importantly, some domains with hard exploration, for example the Atari game
Montezuma’s Revenge, show a very spiked IRD and therefore low IRE. The challenge of such
domains is that nearly all initial traces give the same return, which makes it hard to ever find a first
indication of where to go. In many of such domains nearly all traces then give 0 reward, but for
example Mountain Car shows all traces giving -200 reward (there is a -1 penalty per timestep, and
Gym caps MountainCar episodes at length 200 by default). The entropy of the return distribution
is of course robust against such reward function translations, making it a stable measure of initial
exploration difficulty.
We do not propose this is the only measure of domain exploration difficulty. For example, a wellknown exploration challenge is choosing between a small suboptimal pay-off and exploring further
to obtain a potential higher reward. This type of exploration challenge is not accurately reflected in
the IRE, as simple early rewards spread out the initial return distribution and may falsely suggest the
domain is easy. There are of course many more dimensions that influence the RL task difficulty, like
the state and action space cardinality, but the IRE nicely illustrates why a low-dimensional task like
the Chain (Appendix D) can actually be quite challenging.
Figure 4: Return distributions from the initial state in different environments for a uniform random policy.
Histogram produced over 50.000 traces of maximum 500 steps. The first 8 domains are directly taken from the
OpenAI Gym. The Chain domain is introduced in Appendix D. Orange line is a kernel density estimate, with the
vertical dashed line its empirical mean (a Monte Carlo estimate of Q(s, a) under a uniform random policy. The
top-right display the initial return entropy (IRE) estimate for the domain.
14
C
Illustration of undirected versus directed exploration
We will quickly elaborate on the difference between undirected exploration methods, like -greedy
and Boltzmann exploration, and directed methods, like Thompson sampling, in a theoretical example.
Consider two available actions which, given some observed data H, both have some posterior actionvalue distribution p(Q|H). Figure 5 shows two scenario’s, I (left) and II (right). The only difference
between both scenario’s is our uncertainty about the value of action a1 : in the second scenario we are
much more uncertain about its true value. We now compare how -greedy, Boltzmann and Thompson
sampling will act in both scenario’s. The main point will be that undirected methods cannot leverage
the uncertainty information.
Figure 5: Example posterior value distributions for two available actions. Scenario I (left): Action 1 (blue solid
line) has µ1 = 0, σ1 = 1, Action 2 (green dashed line) has µ2 = 2, σ2 = 1. Scenario II (right): The same
except for σ1 = 5.
1. -greedy exploration only uses the distribution means and will act the same in both scenarios,
preferring action 2 and selecting action 1 with (small) probability .
2. Boltzmann (soft-max) exploration is usually seen as more subtle, gradually preferring actions
with a higher pay-off. Boltzmann does consider the numerical scale of the action means, including
their difference (something -greedy ignores):
πBoltzmann (a|s) = P
eµa
a? ∈A
eµa?
However, it still acts the same in scenario I and II, because Boltzmann approximates a distribution
over both actions by still only considering their means. Although the softmax returns a probability distribution over actions, this should can not be interpreted as their uncertainty.7 Another problem with
Boltzmann action selection is that it is non-robust against translation of the reward function, making
it tedious to tune. Therefore, many undirected implementations still prefer -greedy exploration.
3. Thompson sampling, a directed exploration method, uses π(a1 |s) = p(Q1 > Q2 ). For the
example with normal random variables Q1 ∼ N (·|µ1 , σ1 ) and Q2 ∼ N (·|µ2 , σ2 ), we can analytically
calculate P (Q1 > Q2 ) = P (Q1 − Q2 > 0). Define X = Q1 − Q2 , then X will still
p have a normal
distribution, and by standard laws of probability, E[X] = µ1 − µ2 , and Sd[X] = (σ1 )2 + (σ2 )2 .
Applying
this to the example, gives us for Scenario I (left) π(a1 ) = N (Q1 −Q2 > 0|µX √
= −1, σX =
√
2) ≈ 0.08, and for Scenario II (right) π(a1 ) = N (Q1 − Q2 > 0|µX = −1, σX = 26) ≈ 0.35.
Note how Thompson sampling naturally assigns extra probability mass to action a1 in Scenario II,
where we are much more uncertain about its potential value.
7
A similar phenomenon happens with the softmax and cross-entropy loss in classification tasks. The outputs
of this softmax are also frequently falsely interpreted as a measure of uncertainty over classes (Gal, 2016).
However, when we extrapolate (far) away from our observed data, one of the classes usually gets a high
probability. It thereby appears as we are very certain, but since we have not observed any data in this region of
input space, we should actually be very uncertain. This illustrates how point estimates over a discrete set cannot
be transformed to uncertainties (what we need is an entire uncertainty/distribution per class output).
15
D
Illustration of exploration challenge: Chain domain
We now study the Chain domain, an example MDP (Fig. 6) that illustrates the difficulty of exploration
with sparse rewards. This domain is also empirically studied in the results section of this paper. The
MDP consists of a chain of states S = {1, 2..., N }. At each time step the agent has two available
actions: a1 (‘left’) and a2 (‘right’). At every step, one of both actions is the ‘correct’ one, which
deterministically moves the agent one step further in the chain. The wrong action terminates the
episode. All states have zero reward except the final chain state N , which has r = 1. Variants
of these problem have been studied more frequently in RL (Osband et al., 2014). In the ‘ordered’
implementation, the correct action is always the same (e.g. a2 ), and the optimal policy is to always
walk right. This is the variant illustrated in Fig. 6 as well.
Figure 6: Chain domain. Example MDP where undirected exploration is highly inefficient. Based on (Osband
et al., 2014).
Osband et al. (2014) studied the expected regret for a variant of this scenario. We here present a
different illustration, where we show the expected time until the first visit to the terminal state (i.e.
the first non-zero trace in this domain).
Example 1. Let l denote the number of episodes before we first reach state N . Clearly, before
we reach N for the first time, we have seen no reward information, and undirected exploration will
follow a uniform random policy. The probability of a trace reaching state N under the uniform policy
is p = 2−(N −1) . Therefore, the number of episodes until we first reach N follows a negative binomial
(N −1)
distribution with success probability p, i.e. l ∼ N B(1, p). It follows that E[l] = 1−p
− 1.
p =2
Example 1 shows that, for undirected exploration on point estimates, the required number of exploratory episodes scales exponentially with the exploration depth N . Although this is clearly a
simplified domain, it is important to note that this setting is actually very representative of the
exploration problem in sparse reward domains. This is well visible in Fig. 4, where we can see the
Chain domain having similar initial return distributions as for example Montezuma’s Revenge, a
game notorious for its challenging exploration.
D.1
Additional results for ordered Chain
The experiments section in the paper discusses the unordered Chain, where the correct action at
every step is randomized. We here compare to the ‘ordered’ Chain, where the correct action at every
step is always a2 . Although this is the standard implementation in literature, we believe there is a
systematic bias in this domain that makes them not really exponentially challenging for exploration.
The problem is that the optimal policy has a network predicting a2 at every step. This will happen
too easily in a function approximator (like a neural network) due to its natural tendency to generalize.
We show the results on the ordered Chain in Fig. 7. First of all, compared to the results in Fig. 2, we
see that exploration is indeed much easier in the ordered problem. For example, -greedy now also
solves the problem, something which we would not expect from the exponential exploration time
discussed above. Nevertheless, we see that the probabilistic exploration methods still outperform
-greedy, especially when the length of the chain increases.
16
Figure 7: Learning curves on the ordered Chain domain.
E
Implementation Details
Network architecture consists of a 3 layer network for each discrete action with 128 nodes in each
hidden layer and ReLu activations. For parametric uncertainty, each hidden layer has drop-out
applied to its output, with pkeep probability to keep a node. We use separate subnetworks per action
to explicitly separate their uncertainty. For larger problems, the initial representation layers may
be shared. Learning rates are fixed at 0.001 on all experiments. Optimization is performed with
stochastic gradient descent using Adam updates in Tensorflow. For the experiments with parametric
exploration (parametric uncertainty only) we train on a standard squared loss between new target
and predicted mean action value, i.e. the first half of Eq. 6. We use a target network and replay
database, where we replay 10% of times in a prioritized way (by maintaining a separate prioritized
replay queue based on the total temporal difference error in the previous time this trace was trained
on). All domains (except for the Chain) are taken from the OpenAI Gym repository available at
https://github.com/openai/gym.
All -greedy experiments have fixed at 0.05 throughout learning. Drop-out rates were either
pkeep = 0.75 (for parametric uncertainty only) or pkeep = 0.90 (for uncertain returns). All experiments used one-step SARSA (ie., on-policy updates with eligibility traces parameter λ = 0 (Sutton
and Barto, 1998)), except for the MountainCar experiments, which use λ = 0.9. Note that the ideas
about eligibility traces and cutting traces equally apply to the propagation of distributions, i.e. they
allow for quicker propagation over multiple timesteps (always at the risk of propagating on-policy,
exploratory results too quickly/far).8
Thompson sampling uses the same policy for exploration and evaluation. In some sense, proper uncertainty policies somewhat blur the line between on- and off-policy (where the behavioural/exploration
policy differs from the target/evaluation policy), as there is just on reasonable probabilistic policy
incorporating all uncertainty. Nevertheless, we could consider Thompson sampling exploration while
evaluating with a policy that does act on some mean value again.
8
We do believe that the uncertainty-based policies, like Thompson sampling, may also benefit work on
cutting traces. Trace cutting is usually based on importance sampling ratios between the exploratory policy
and the target policy. Thompson sampling may provide more realistic probabilities for exploratory actions,
which may allow for more natural trace cutting. For example, -greedy always strongly cuts a trace for every
exploratory step, no matter whether the exploratory action is very close to the best one, or known to be very bad.
In contrast, probabilistic policies will cut traces when other possible actions in the state have much uncertainty
left, which should indeed stop the speed of our back-ups. A challenge is that our neural network implementation
naturally samples a next action, but the associated probability of each action is not directly available. Of course,
for a small discrete action space we could approximate it by repeatedly sampling from our policy.
17
| 2cs.AI
|
Structured low-rank matrix completion for forecasting in time series
analysis
Jonathan Gillarda , Konstantin Usevichb,c,˚
arXiv:1802.08242v1 [stat.ME] 22 Feb 2018
b
a
Cardiff School of Mathematics, Cardiff University, Senghennydd Road, Cardiff, UK, CF24 4AG.
Université de Lorraine, CRAN, UMR 7039, Campus Sciences, BP 70239, 54506 Vandœuvre-lès-Nancy, France
c
CNRS, CRAN, UMR 7039, France
Abstract
In this paper we consider the low-rank matrix completion problem with specific application to
forecasting in time series analysis. Briefly, the low-rank matrix completion problem is the problem
of imputing missing values of a matrix under a rank constraint. We consider a matrix completion
problem for Hankel matrices and a convex relaxation based on the nuclear norm. Based on new
theoretical results and a number of numerical and real examples, we investigate the cases when the
proposed approach can work. Our results highlight the importance of choosing a proper weighting
scheme for the known observations.
Keywords: Hankel matrices; Low-rank matrix completion; Forecasting; Nuclear norm
1. Introduction
Low-rank representations and approximations have been shown to be a very useful tool in
time series forecasting [1]. One of the popular approaches is singular spectrum analysis (SSA)
forecasting [2] that embeds the time series into a Hankel matrix and uses a low-rank approximation
and continuation to compute the next values of a time series. SSA uses the fact that many time
series can be well approximated by a class of so-called time series of finite rank. Despite many
successful examples [3, 4, 5], SSA forecasting has a number of disadvantages.
In this paper we develop a method based on Hankel matrix completion. We follow the approach
of [6], where it was proposed to embed a time series into a Hankel matrix and the missing data
(to be forecasted) are stored in the bottom right-hand corner of this matrix. The method of [6] is
based on minimising the nuclear norm, which provides a convex relaxation to a low-rank matrix
completion problem which is non-convex and NP-hard in general (see for example [7] and [8]).
The nuclear norm (the sum of singular values) is a popular convex surrogate for the rank [7],
and is similar to using the ℓ1 -norm for sparse approximation [9]. It was shown to be a successful
tool in imputing the missing values of a matrix (see for example [7], [10], [11] and [12], [13]).
Nuclear norm relaxation has been a very popular tool for spectral estimation [13], recommender
systems [7], and system identification [14, 15, 16]. An advantage of the nuclear norm relaxation
˚
Corresponding author
Email addresses: [email protected] (Jonathan Gillard), [email protected]
(Konstantin Usevich)
Preprint submitted to Elsevier
February 22, 2018
considered in this paper is the ability to build more complex models to represent potentially more
complex behavior in observed time series.
An important question is when the convex relaxation solves the original low-rank matrix completion problem. Much famous research has been conducted on this topic, but most available
research [7, 10, 11, 12, 13] assumes that the position of the missing entries in the matrix is random,
and often that the known entries are also random; mostly unstructured matrices are considered.
Therefore these results are not applicable due to a special arrangement of missing data and due to
the Hankel structure in our problem. Moreover, as noted in [17] and [18], the case of structured
matrices is much more challenging.
There are few available results for completion of Hankel matrices with a fixed pattern of missing
values. In [19], a special case was analysed: square real-valued Hankel matrices with nearly half
of its values missing. It was shown that the nuclear norm relaxation gives the correct rank-one
completion only when the embedded time series can be written as a sum of decreasing exponentials.
In [20], this analysis was extended to the rank-r case for the same pattern of missing values.
This paper has several contributions. First, as in [6], we consider the general case of rectangular
Hankel matrices with potentially fewer missing values. We show that for few missing values, the
convex relaxation of the low-rank matrix completion using the nuclear norm will give identical
solutions without using the convex relaxation for time series with undamped or exponentially
increasing periodic components, and establish bounds on the number of missing values. We also
study the question of choosing the optimal shape of the Hankel matrix (parameterized by the socalled window length). Second, we suggest a new (relative to [6]) formulation of the low-rank matrix
completion problem for Hankel matrices, which allows the possibility to allocate past observations
different weights. In particular, exponential weighting is designed to overcome the problems related
to the performance of the nuclear norm for time series that can be expressed as a sum of increasing
exponentials.
Empirical comparisons show that, with the proper choice of weights, our novel formulation
performs well relative to a number of classical techniques. For the numerical examples in this
paper we use CVX, a MATLAB package for specifying and solving convex programs [21, 22].
This paper has the following structure. In Section 2 we formally define the problems to be
considered. We first define exact matrix completion, before considering an approximate version.
This section also describes the settings used throughout the paper. Some known theoretical results
necessary to be stated are reviewed in Section 3. First, the time series of finite rank are recalled,
and the solution of the exact minimal rank completion is summarized. Next, known results on time
series of finite rank are recalled. Section 4 contains the main results of the paper. First, we give
theoretical bounds for matrix completion in the case of arbitrary shape of the matrix and number of
missing values. We check the tightness of our bounds based on numerical experiments. Second, we
establish the connection between exponential weighting and preprocessing of time series. Finally,
the examples for forecasting of real and model time series demonstrating the advantages of the
proposed methodology are provided in Section 5.
2
2. Problem Statement
2.1. Hankel matrices
For a vector f “ pf1 , . . . , fn q with n ą 1 and a so-called window length L, the L ˆ pn ´ L ` 1q
Hankel matrix is defined as
˛
¨
f1
f2
¨ ¨ ¨ fn´L`1
˚
‹
.
˚ f2
. . fn´L`2 ‹
f3
˚
‹.
HL pfq “ ˚ .
‹
..
.
.
.
.
.
˝ .
‚
.
.
.
fL fL`1 ¨ ¨ ¨
fn
In what follows, we are going to pose the problem of forecasting a given time series as the low-rank
matrix completion of a Hankel matrix. Formally, let
p “ pp1 , p2 , . . . , pn`m q,
(1)
be a vector of length pn ` mq, with m ě 0. In what follows m will be the number of observations
forecasted, and n will be the length of the given time series that we wish to forecast. We use the
notation pp1:nq “ pp1 , p2 , . . . , pn q for the first n elements of p. Next, let L, K be the integers such
that L ` K ´ 1 “ m ` n. Then the matrix structure Sppq (parameterized by p) we consider is
¨
˚
˚
˚
˚
˚
Sppq “ HL ppq “ ˚
˚
˚
˚
˚
˝
¨¨¨
¨¨¨
.
..
p1
p2
p2
..
.
..
.
..
.
¨¨¨
.
.
..
¨ ¨ ¨ ..
.
.
.
.. ..
..
.
.
.
.. ..
..
¨ ¨ ¨ pn pn`1
pL
p3
¨¨¨
pK
..
.
¨¨¨
pn
¨¨¨
¨¨¨
.
..
pn`1
..
.
¨ ¨ ¨ pn`m
˛
‹
‹
‹
‹
‹
‹.
‹
‹
‹
‹
‚
(2)
In (2), the grey-shaded values are “known” and others are “missing”.
The Hankel matrix structure belongs to the class of affine matrix structures [18, §3.3] having
the form:
pn`mq
ÿ
pi Si
(3)
Sppq “ S0 `
i“1
where Si , i P t0, 1, . . . , pn ` mqu are given linearly independent basis matrices, and in particular,
for the Hankel matrix structure (2), the basis matrices matrices in (3) are given as S0 “ 0,
¨ 0 0 ¨¨¨ 0 0 ˛
¨ 0 0 ¨¨¨ 0 0 ˛
¨ 0 1 ¨¨¨ 0 0 ˛
¨ 1 0 ¨¨¨ 0 0 ˛
.
..
.
..
.
0 0 ..
0 0 ¨¨¨
˚0
.
S1 “ ˚
˝ ..
0
.
..
0 0‹
. .‹
. . .. ‚, S2
0 0
0 0
.
..
.
..
.
0 0 ..
0 0 ¨¨¨
˚1
.
“˚
˝ ..
0
.
..
0 0‹
. .‹
. . .. ‚, . . . , Sn`m´1
0 0
0 0
.
..
.
..
.
0 0 ..
0 0 ¨¨¨
˚0
.
“˚
˝ ..
0
.
..
0 0‹
. .‹
. . .. ‚, Sn`m
0 1
1 0
.
..
.
..
.
0 0 ..
0 0 ¨¨¨
˚0
.
“˚
˝ ..
0
.
..
0 0‹
. .‹
. . .. ‚.
0 0
0 1
In the following subsections, we describe the formal statement of the matrix completion problem,
and the nuclear norm relaxation approach that we propose.
3
2.2. Exact low-rank matrix completion
Let p0 “ pp0,1 , p0,2 , . . . , p0,n q be a given vector of observations (a time series). For a given
matrix structure (3), the Structured Low-Rank Matrix Completion (SLRMC) problem is posed as
p̃ “ arg min rank Sppq subject to pp1:nq “ p0 .
pPRpn`mq
(4)
The implicit low-rank assumption of the Hankel matrix corresponds to the class of time series of
so-called finite rank, which are described in Section 3.
The matrix completion problem (4) for general matrix structures is NP-hard (see [7] and [8]). A
convex relaxation of this problem based on the nuclear norm became increasingly popular recently.
Formally, for a matrix X P CLˆK its nuclear norm is defined as
}X}˚ “
minpL,Kq
ÿ
k“1
|σk pXq|,
where σk pXq are the singular values of X. A convex relaxation of (4) is obtained by replacing the
rank with the nuclear norm:
p̂ “ arg min }Sppq}˚ subject to pp1:nq “ p0 .
pPRpn`mq
(5)
The intuition behind this relaxation is the same as for using the ℓ1 norm in compressed sensing:
the nuclear norm is expected to force all but a few singular values to be zero (a low-rank solution).
Remark 2.1. For the Hankel matrix case, the solution of (4) is known and is given in Section 3.
Still, the performance of the nuclear norm relaxation is important for understanding the behavior
of forecasting in the approximate case to be introduced in the next subsection.
2.3. Approximate matrix completion
Let x “ px1 , x2 , . . . , xn q be a vector of length n. We denote
g
fn
fÿ
}x}W “ e wi x2i
(6)
i“1
where W “ pw1 , w2 , . . . , wn q, is a vector of weights, wi ą 0 for i “ 1, . . . , n.
The approximate rank minimization can be posed as follows. Given p0 , W , m ě 0 and τ ě 0
find
p̃ “ arg min rank Sppq subject to }pp1:nq ´ p0 }W ď τ .
(7)
pPRpn`mq
The parameter τ controls the precision of approximation. Two extreme cases can be distinguished:
• if τ “ 0, then (7) is equivalent to the exact matrix completion problem (4), i.e. there is no
approximation;
• if m “ 0 in (7), then p̃ is an approximation to the given vector p0 with no forecast.
4
Unlike (4), the problem (7) does not have a known solution. In fact, it is a dual problem to
structured low-rank approximation [18] which is known to be a difficult optimization problem [23].
In order to circumvent the complexity of the problem, we consider the following relaxation of
(7), using the nuclear norm:
p˚ “ arg min }Sppq}˚ subject to }pp1:nq ´ p0 }W ď τ .
pPRpn`mq
(8)
Remark 2.2. There are alternative ways to extend the problem (5) to approximate versions. For
example, consider the following equivalent formulations:
min
}pp1:nq ´ p0 }W subject to }Sppq}˚ ď δ,
min
}pp1:nq ´ p0 }W ` γ}Sppq}˚ ,
pPRpn`mq
pPRpn`mq
(9)
(10)
where δ and γ are regularisation parameters for each of the formulations.
In fact, it can be shown that (similarly to the result for equivalence of LASSO formulations
[24]) the problems (8), (9) and (10) are equivalent in the following sense: for any value of τ , there
exist δ and γ such that the solutions to (8), (9) and (10) coincide. However, the relation between
“equivalent” τ , δ and γ is not known a priori. See for example [24, §1.3].
2.4. Choice of weights
There are several natural choices of weights W defining (6) in the approximation problem (8):
1. Trapezoid weighting. Take W1 to be the vector of weights such that }pp1:nq ´ p0 }W1 “
}HL ppp1:nq q ´ HL pp0 q}F . In this case, the vector W1 “ pw1,1 , w1,2 , . . . , w1,n q is given by
w1,i
$
for i “ 1, . . . , L ´ 1,
& i,
L,
for i “ L, . . . , n ´ L ` 1,
“
%
n ´ i ` 1, for i “ n ´ L ` 2, . . . , n .
(11)
2. Uniform weighting. Take W2 to be the vector of weights such that W2 “ p1, 1, . . . , 1q.
3. Exponential weighting. take W3 to be the vector of weights W3 “ pw3,1 , w3,2 , . . . , w3,n q such
that w3,i “ exppα iq, i “ 1, . . . , n for some number α.
We make the following comments regarding each of the possible choices of weights given above.
W1 is the vector of weights ‘enforced’ upon the observations if the Frobenius norm of a Hankel matrix is used. The Frobenius norm for matrix optimization problems is commonly and traditionally
used due to classical results of the optimality of low rank approximations achieved by truncating
the singular value decomposition. However an unintended consequence of using the Frobenius norm
is that the vector of observations (when placed into a Hankel matrix) receive the weights as given in
W1 . Such weights may be unnatural for forecasting, for example, as the more recent observations
are given declining weight in time. W2 is the natural ‘correction’ to the weights offered in W1 by
giving each observation equal weight. As mentioned earlier, for the purposes of forecasting there
may be some advantages in adopting the weights given in W3 , where the most recent observations
receive larger weight.
5
3. Time series of finite rank
In this section, we recall the class of so-called time series of finite rank [2] and the solution of
the exact rank minimization problem (4) for the Hankel matrix structure (2).
3.1. Time series of finite rank and linear recurrent formulae
First, we recall basic properties of time series of finite rank1 [2]. Informally speaking, time
series of finite rank are the time series for which the Hankel matrix has low rank. It is known [2,
Chapter 5] that the class of such real-valued time series is given by sums of products of cosines,
exponential and polynomial functions, i.e.
s2
s1
ÿ
ÿ
(12)
Qj pkqρkj , k “ 1, 2, . . . ,
Qj pkqρkj cosp2πωj k ` φj q `
pk “
j“s1 `1
j“1
where Qj pkq are real polynomials of degrees µj ´ 1 (µj are positive integers), ωj P p0, 12 q, φj P R,
2
1
ρj P R are such that tρj usj“s
are distinct and the pairs tpρj , ωj qusj“1
are also distinct. As shown
1 `1
in [2], time series of the form (12) are particularly suitable to model trends, periodicities and
modulated periodicities in time series analysis. The number
s2
s1
ÿ
ÿ
(13)
µj
2µj `
r“
j“s1 `1
j“1
is called the finite difference dimension (or rank) of the time series and is equal to the rank of the
Hankel matrix Sppq given in (2) if L, K ě r (see [2, Chapter 5]).
Time series of finite rank can be more compactly represented in the complex-valued case. Some
of the results of this paper will be formulated for complex-valued time series. Next, we recall a
summary from [25], which is based on the results of [26]. Consider a complex-valued infinite time
series pp1 , p2 , . . . , pn , . . .q with
s
ÿ
Pj pkqλkj , k “ 1, 2, . . . ,
(14)
pk “
j“1
where Pj pkq are complex polynomials of degrees νj ´ 1, (νj are positive integers), and λj P Czt0u.
Note that time series (12) are special cases of time series (14). Indeed, take s “ s1 ` 2s2 ,
pλ1 , . . . , λs q “ pρ1 e2πω1 , ρ1 e´2πω1 , . . . , ρs1 e2πωs1 , ρs1 e´2πωs1 , ρs1 `1 , . . . , ρs1 `s2 q
φ
´φ
´φs
φs
and pP1 , . . . , Ps q “ p e 2 1 Q1 , e 2 1 Q1 , . . . , e 2 1 Qs1 , e 2 1 Qs1 , Qs1 `1 , . . . , Qs1 `s2 q.
It is known [25, Corollary 2.1] that time series (14) satisfy the minimal linear recurrent formula
pk`r “ ´
r´1
ÿ
qj pk`j ,
(15)
j“0
where the (complex) coefficients qj are the coefficients of the so-called characteristic polynomial
(16)
qpzq “ pz ´ λ1 qν1 ¨ ¨ ¨ pz ´ λs qνs “ z r ` qr´1 z r´1 ` ¨ ¨ ¨ ` q1 z ` q0 .
s
ř
This is why the number r “
νj is called the finite difference dimension of the time series (14).
j“1
1
In fact, we consider the class of time series of finite difference dimension; however, it coincides with the time
series of finite rank if the time series is infinite [2, Corollary 5.1]. This is why we use the term “time series of finite
rank” in this paper.
6
3.2. Solution of the rank minimization problem
For time series of finite rank, the solution of the rank minimization problem (4) for the Hankel
structure (2) is known. It is equivalent to the problem of minimal rank extension of Hankel matrices,
solved in [27] for square matrices and in [26] for rectangular matrices. We summarize the results
of [27, 26] in the form of a theorem, and the proof is given in Appendix B.
Theorem 3.1. Let r ď minpL, K, n2 q, and consider the Hankel matrix structure as given in (2).
Then for any complex-valued time series pp0,1 , p0,2 , . . . , p0,n , . . .q of finite rank r (i.e. of the
form (14)), and vector p0 defined as
p0 “ pp0,1 , p0,2 , . . . , p0,n q,
(17)
the solution of the rank minimization problem
r “ arg min rank Sppq subject to pp1:nq “ p0 .
p
(18)
r “ pp0,1 , p0,2 , . . . , p0,n`m q ,
p
(19)
pPCpn`mq
is unique and given by
where p0,n`1 , p0,n`2 . . . p0,n`m are computed using the linear recurrent formula (15).
Remark 3.2. Theorem 3.1 implies that the same holds for the real-valued case: for a real-valued
time series pp0,1 , p0,2 , . . . , p0,n , . . .q of finite rank r (i.e., time series of the form (12)) and the vector
p0 defined as (17), the solution of the rank minimization problem (4) is unique and given by (19).
Variants of Theorem 3.1 in the real-valued case are also used in the framework of SSA [2, 25].
In fact, the continuation by the linear recurrent formula (15) is at the core of forecasting methods
in SSA.
Remark 3.3. If the parametric form of the time series (i.e. (12) or (14)) is known, then the
minimal rank completion is given by the same formula (15). However, the advantage of the recursion (15) is that the parametric form does not have to be derived. Due to that reason such matrix
completion approaches are referred to as data-driven in [28]. The nuclear norm forecasting method
proposed in this paper has the same advantage.
3.3. Performance of the nuclear norm: known results
The performance of the nuclear norm (i.e. when the solution of (5) coincides with the solution
of (4)), was studied recently for a special case of the structure (2):
˛
¨
p1
p2 ¨ ¨ ¨
pn
‹
˚
.
˚ p2
. . pn`1 ‹
p3
‹
˚
(20)
Sppq “ ˚ .
.. ‹ ,
.
.
.
.
˝ ..
.
.
. ‚
pn pn`1 ¨ ¨ ¨ pn`m
i.e. when the matrix is square (L “ K), and all the values below the main antidiagonal of the
Hankel matrix Sppq are missing (i.e. L “ n “ m ` 1).
The first result, appeared in [19] and refined in [20], treats the rank-one case and is given below.
7
Theorem 3.4 ([20, Theorem 6]). Let p0 “ pp0,1 , p0,2 , . . . , p0,n q be a complex-valued vector given as
p0,k “ cλk ,
k “ 1, . . . , n,
where λ P C.
• If |λ| ď 1, the solution of (4), i.e.
p0,k “ cλk ,
k “ n ` 1, . . . , n ` m
is also a solution of
p “ arg min }Sppq}˚ subject to pp1:nq “ p0 ;
p
(21)
pPCpn`mq
in particular, if |λ| ă 1, then the solution of (21) is unique.
• If |λ| ą 1, then the unique solution of (21) is given by
p0,n`k “ c
λn
k
,
λ
k “ 1, . . . , m.
Theorem 3.4 is illustrated in Fig. 1. Note that the exponential needs to be decreasing (damped),
in order for the nuclear norm (21) to give the same solution for when there is no convex relaxation.
A similar result was proved for the rank-r case (where the exponentials should be decreasing
sufficiently fast).
0
1
2
3
4
5
6
7
8
9
10
n
0
(a) Decreasing exponential (ρ ă 1)
1
2
3
4
5
6
7
8
9
10
n
(b) Increasing exponential (ρ ą 1)
Figure 1: Forecasts (in red) given by the nuclear norm for the time series pk “ cρk , n “ 6.
Theorem 3.5 ([20, Theorem 7]). Fix the structure (20) (i.e. L “ K “ n “ m ` 1 in (2)) and
r ď n2 . Then there exists a number 0 ă ρmax,r,m ă 1 such that for any complex-valued time series
of finite rank r (14) with
|λj | ă ρmax,r,m ,
the solution of (21) is unique and coincides with the solution of (18) given in Theorem 3.1.
Theorem 3.5 implies that the same result holds for the real-valued problems (4) and (5). Note
that Theorem 3.5 only requires that the roots λj of the characteristic polynomial (16) should be all
within a disk of radius ρmax,r,m in the complex plane. Theorem 3.5 does not impose any conditions
on separation of the roots λj or their multiplicities νj .
Remark 3.6. Theorem 3.4 proves that (in the case of rank-one matrices) the radius ρmax,1,m is
equal to 1.
In the general case (r ą 1), Theorem 3.5 does not give a good estimate of the radius ρmax,1,m .
8
4. Theoretical results on nuclear norm forecasting
4.1. Square matrices: fewer missing values
In this subsection, we show that the radius obtained from Theorems 3.4 or 3.5 is also directly
applicable to the case of fewer missing values, which is typical in the case of forecasting.
Corollary 4.1. Let L “ K, n ě L, m “ 2L ´ 1 ´ n, and r ă L`1
2 . In this case we have fewer
missing values than in Theorem 3.5, i.e. the matrix Sppq is of the form
¨
˛
p1 p2 ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨
pL
.. ‹
˚
.
˚ p2 p3 ¨ ¨ ¨ . .
. ‹
¨¨¨
˚
‹
˚ ..
‹
˚ . ¨ ¨ ¨ ... ... ¨ ¨ ¨
‹
p
n
˚
‹.
Sppq “ ˚ .
(22)
‹
.
.
.
.
˚ .. . . . .
‹
.
¨
¨
¨
p
n`1 ‹
˚
˚ .
.. ‹
˝ .. . . . . . . . . . . . .
. ‚
pL ¨ ¨ ¨ pn pn`1 ¨ ¨ ¨ pn`m
Let ρmax,r,L be the radius from Theorem 3.5. Then for any complex-valued (or real-valued) time
series of the form (14) where
|λj | ă ρmax,r,L
the solution of the nuclear norm minimization (21) for the structure (22) coincides with the solution
of the rank minimization problem (18).
4.2. Rectangular matrices: the rank-one case
In this section, we aim to improve the results of the previous subsection. We give an explicit
bound for a single (complex) exponential for rectangular matrices.
Proposition 4.2. Let L, K be arbitrary, m ă minpL, Kq (so that n “ L`K ´1´m ě maxpL, Kq).
Let p0 “ pp0,1 , p0,2 , . . . , p0,n q be a complex-valued time series given as
p0,k “ cλk ,
where λ P C. If ρ “ |λ| satisfies Cpρq ă 1, where
$
L`K
&|ρm`1 ´ ρ´pm`1q | ¨ ? |ρ|
,
2L
|ρ ´1||ρ2K ´1|
Cpρq “ Cpρ, L, K, mq “
m`1
%?
,
LK
(23)
|ρ| ‰ 1,
|ρ| “ 1.
then the solution of the nuclear norm minimization (21) for the matrix (2) coincides with the
solution of of the rank minimization problem (18).
As an illustration of Proposition 4.2, Figure 2 contains plots of Cpρq against ρ for different m
with L “ 20, K “ 50 and n “ L ` K ´ 1 ´ m. Figure 3 contains plots of Cpρq against m for
different ρ with L “ 20, K “ 50 and n “ L ` K ´ 1 ´ m. In Fig. 3, we can see that for small
number of missing values the nuclear norm minimization gives the correct solution if the complex
exponential is undamped or increasing in magnitude. Fig. 3 shows that there is for fixed L, K, ρ
there is a limiting number of missing data until which the nuclear norm gives the correct solution.
In fact, we can find this number explicitly and this is given in Corollary 4.3.
9
(a) m “ 1
(b) m “ 5
(c) m “ 10
Figure 2: Plot of Cpρq against ρ for different m with L “ 20, K “ 50 and n “ L ` K ´ 1 ´ m.
(a) ρ “ 0.85
(b) ρ “ 0.95
(c) ρ “ 1.05
Figure 3: Plot of Cpρq against m for different ρ with L “ 20, K “ 50 and n “ L ` K ´ 1 ´ m.
Corollary 4.3. Let L, K and ρ be fixed, and the time series is given as in (23). Then the solution
of the nuclear norm minimization (21) for the matrix (2) coincides with the solution of (18) if
ˆ?
˙
$
y 2 `4´y
’
’
log
, |ρ| ă 1,
’ |ρ|
2
’
&?
m`1ă
LK,ˆ
˙ |ρ| “ 1,
’
?
’
2
’
y `4`y
’
%log|ρ|
, |ρ| ą 1,
2
where for |ρ| ‰ 1 we define
dˇ
ˇˇ
ˇ
ˇ
1 ˇˇ ˇˇ
1 ˇˇ
ˇ
y “ ˇ1 ´ 2L ˇ ˇ1 ´ 2K ˇ.
ρ
ρ
Next we look at the question of choosing the optimal window length. The next corollary shows
that the performance of the nuclear norm is maximised when the Hankel matrix is square or almost
square.
10
Corollary 4.4. If m and n are fixed, then Cpρq is minimised if L “ K (in the case m ` n odd)
or |L ´ K| “ 1 (in the case m ` n even).
In Figure 4 we plot Cpρq against m for different ρ and L, which confirms the conclusions of
Corollary 4.4. Indeed, the curves for the approximately square shape of the Hankel matrix are
lower than the others in Figure 4.
(a) ρ “ 0.99
(b) ρ “ 1.01
(c) ρ “ 1.05
Figure 4: Plot of Cpρq against m for different ρ and L. Here K “ 70 ´ L and n “ L ` K ´ 1 ´ m.
Next, we would like to see whether the bound given in Corollary 4.3 is tight. We take L “ K “
13, the time series pk “ ρk and test the performance of the nuclear norm for different values of ρ
and m. If the Frobenius norm between the minimal rank completion and minimal nuclear norm
completion is ď 10´4 , the nuclear norm completion is declared successful. The results are plotted
in Figure 5. From Figure 5 we see that the bound given by Corollary 4.3 is not optimal for small
1.2
1.1
rho
1.0
0.9
0.8
0.7
1
2
3
4
5
6
7
8
9
10 11 12
m
Figure 5: Success of the nuclear norm against m and ρ, L “ K “ 13. White: success, black: failure. Red curve: the
bound given by Corollary 4.3.
m, so the nuclear norm works for higher values of ρ.
11
4.3. Rectangular matrices: the rank-r case
First, we provide an improved result from Corollary 4.1 that is valid for rectangular matrices
and is stronger so that it may reduce the radius ρ.
Proposition 4.5. Let L, K, m, n be as in Theorem 3.1, and consider the matrix structure (2).
Fix the number r ď minpL,Kq`1
, and define m2 “ maxp2r ´ 1, mq. Let ρ1max,r,m2 be the radius from
2
Theorem 3.5 for matrices of size pm2 ` 1q ˆ pm2 ` 1q. Then for any time series of the form (14)
where
|λj | ă ρ1max,r,m2
the solution of the nuclear norm minimization (21) for the full matrix Sppq coincides with the
solution of the rank minimization problem (18) (the same holds for the real-valued problem statements).
Next, we verify numerically what is the maximum number of missing entries when the nuclear
norm gives the correct solution in the rank-r case. For different values of ρ, and r, we consider the
time series
r
ÿ
ρj eiωj k ,
pk “
j“1
where ωj are drawn uniformly in r0, 2πq. For M “ 20 realisations, we calculate the empirical
probability of success, based on the same criterion as in the previous example. The results are
plotted in Fig. 6.
5
4
4
4
r
6
5
r
6
5
r
6
3
3
3
2
2
2
1
1
1
2
3
4
5
6
7
8
9
10 11 12
1
1
2
3
4
5
6
m
7
8
9
10 11 12
m
(a) ρ “ 0.85
5
4
4
3
4
5
6
7
8
9
10 11 12
(c) ρ “ 1.00
r
5
r
6
2
m
(b) ρ “ 0.95
6
1
3
3
2
2
1
1
1
2
3
4
5
6
7
8
m
(d) ρ “ 1.05
9
10 11 12
1
2
3
4
5
6
7
8
9
10 11 12
m
(e) ρ “ 1.15
Figure 6: Probability of success of the nuclear norm against m, ρ and r, L “ K “ 13. Greyscale, white: 1, black: 0.
In Figure 6 we see that the higher the rank is, the smaller number of missing values can be
correctly imputed by the nuclear norm. For decaying exponentials, the nuclear norm recovers the
12
correct solution for a large number of missing values and ranks, but for the undamped or increasing
exponentials special care has to be taken.
4.4. Results and discussion on the approximate matrix completion
The previous subsections suggest that the nuclear norm heuristic works for sufficiently damped
sinusoids. Hence, it may be beneficial to preprocess the time series (by introducing an artificial
damping) before solving (8). Formally, for a given α ą 0 we define the scaled matrix structure:
Ssc ppq “ Sppsc q,
where
psc “
ˆ
(24)
ˆ
˙
ˆ
˙˙
´ α¯
2α
pm ` nqα
p1 ¨ exp ´
, p2 ¨ exp ´
, . . . , pm`n ¨ exp ´
2
2
2
(25)
p˚˚˚ “ arg min }Ssc ppq}˚ subject to }pp1:nq ´ p0 }W2 ď τ.
(26)
is the scaled time series (multiplied by a decreasing exponential). For simplicity, consider the case
of the uniform weights W2 “ p1, 1, . . . , 1q, and solve
pPRpn`mq
In fact, we will show in the next proposition that the problem (26) is equivalent to an exponentially
weighted problem (8) for the scaled time series.
Proposition 4.6. The solution of problem (26) is given by
1. Construct the scaled initial data vector psc,0 as in (25):
ˆ
ˆ
˙
´ α¯
´ nα ¯˙
2α
, p0,2 ¨ exp ´
.
psc,0 “ p0,1 ¨ exp ´
, . . . , p0,n ¨ exp ´
2
2
2
2. Solve the problem
psc,˚ “ arg min }Sppsc q}˚ subject to }psc,p1:nq ´ psc,0 }W3 ď τ,
pPRpn`mq
where W3 “ pw3,1 , w3,2 , . . . , w3,n q such that w3,i “ exppαiq, i “ 1, . . . , n.
3. Scale back the weighted approximation
ˆ
ˆ ˙
ˆ
˙˙
´α¯
2α
pm ` nqα
p˚˚˚ “ psc,˚,1 ¨ exp
, psc,˚,2 ¨ exp
, . . . , psc,˚,m`n ¨ exp
.
2
2
2
(27)
(28)
Hence, by Proposition 4.6 the exponential weighting may help to overcome the potential problems with the increasing exponentials in the time series.
5. Examples
5.1. Fortified wine
In this section we consider the classical time series ‘Fortified Wine’, where 120 observations
depict the monthly volumes of wine sales in Australia (thousands of litres) in the period from
January 1980 until December 1989. Denote by p0 the vector of these 120 observations, that is,
n “ 120. We do not consider a forecast and thus take m “ 0.
13
Figure 7 contains a plot of the vector p0 with three approximations. The approximations are
obtained from (8) with L “ 60 and weighting scheme (11), so that }p ´ p0 }W1 “ }Sppq ´ Spp0 q}F .
Define
τ prq “
min
}Spp0 q ´ X}F .
(29)
XPRLˆK , rankpXq“r
For given r, the value τ prq in (29) can be found using the singular value decomposition, see for
example [29, Sect. 2.4]. The motivation for taking τ prq as in (29) is to ensure that our approximation
p˚ is as ‘least as good’ as the what would be obtained from the unstructured low rank approximation
of Spp0 q. We consider three values of τ prq , with r “ 1, 3 and 10.
6000
5000
4000
3000
2000
1000
0
20
40
60
80
100
120
Figure 7: Fortified wine (solid line) with three approximations obtained from (8), τ “ τ p1q (long dash), τ “ τ p3q
(short dash) and τ “ τ p10q (dot-dash).
Figure 8 contains plots of the square root of the first fifteen singular values of the matrix Spp˚ q
where p˚ is the solution to (8), for three values of τ as described above. Minimizing the sum of
singular values results in many individual singular values going close to zero. The parameter τ can
be used to control the complexity of the approximation, the smaller the value of τ , the closer the
approximation will be to the given vector p0 .
500
500
500
400
400
400
300
300
300
200
200
200
100
100
100
0
0
5
(a) τ “ τ
10
p1q
15
0
0
5
10
(b) τ “ τ
p3q
15
0
0
5
(c) τ “ τ
10
15
p10q
Figure 8: Plot of the square root of the first 15 singular values of the matrix Spp˚ q where p˚ is the solution to (8).
14
5.2. Forecasting deaths
In this section we consider forecasting the famous ‘death’ series recording the monthly accidental
deaths in the USA between 1973 and 1978. This data has been studied by many authors (such as
[2]) and can be found in a number of time series data libraries. We wish to replicate the exercise
given in [30] which aimed to forecast the final six values of this series. The time series contains a
total of 78 observations. We truncate the series to the first 72 observations and will forecast the
remaining six observations. Denote these series of 72 observations by p0 . We consider (8) with
n “ 72 and m “ 6.
Table 1 contains forecasts of the final six data points of the data series by several methods
along with the square root of the mean square error. These results are taken from [30] and full
details of the fitted models can be found within. In summary Model I and Model II are examples
of SARIMA models as described by [31]. Model I is given by
∇12 pi “ 28.831 ` p1 ´ 0.478Bqp1 ´ 0.588B 12 qZi
and Model II is given by
∇12 pi “ 28.831 ` Zi ´ 0.596Zi´1 ´ 0.407Zi´6 ´ 0.685Zi´12 ` 0.460Zi´13
where Zi is a realisation of white noise with zero mean and variance 0.9439 and B is the backward
shift operator defined as: B j Zi “ Zi´j . HWS represents the model as fitted by the Holt-Winter
seasonal algorithm. ARAR represents the model as fitted by transforming the data prior to fitting
an autoregressive model.
?
M SE
1
2
3
4
5
6
Original data 7798 7406 8363 8460 9217
9316
Model I 8441 7704 8549 8885 9843 10279 582.626
Model II 8345 7619 8356 8742 9795 10179 500.500
HWS 8039 7077 7750 7941 8824
9329 401.263
ARAR 8168 7196 7982 8284 9144
9465 253.202
Table 1: Forecasted data using four different models, along with the square root of the mean square error.
We now consider the following exercise. We take L “ 24 and consider three forms of W defined
in Section 2.3. For the weighting scheme W3 we take α “ 0.05. We will solve (8) under these
three weighting schemes and select τ such that the solutions obtained are as least as close to the
solutions obtained by unstructured rank 3, 6 and 12 approximations to HL pp0 q. More precicely,
for a weight vector W and rank r, we take τ to be the solution of
q“
X
arg min
XPRLˆK , rankpXq“r
}HL pp0 q ´ X}F ,
q F,
q “ arg min }HL ppq ´ X}
p
pPRn
τ
prq
“ }q
p ´ p0 } W .
In particular, for the weighting scheme W1 it is equivalent to taking τ “ τ p3q , τ p6q and τ p12q as
q can be obtained by orthogonal projection on the space of the Hankel
in (29). Also, note that p
15
Original data
W1 , rank = 3
W1 , rank = 6
W1 , rank = 12
W2 , rank = 3
W2 , rank = 6
W2 , rank = 12
W3 , rank = 3
W3 , rank = 6
W3 , rank = 12
1
7798
7829
7860
7873
8371
8269
8096
8233
8029
7953
2
7406
7730
7722
7648
8219
8007
7704
7947
7444
7490
3
8363
7775
7779
7808
8233
8022
7916
7863
7874
7906
4
8460
8045
8030
8076
8501
8262
8123
8236
8349
8118
5
9217
8485
8714
8868
8945
9081
9040
9485
9433
9122
6
9316
8815
9006
9123
9356
9401
9293
10012
9782
9356
?
M SE
485.06
404.00
336.27
424.87
356.75
295.23
472.29
308.28
247.38
Table 2: Forecasted data using nine different models, along with the square root of the mean square error.
matrices. Hence, it can be computed by diagonal averaging [2, Sec. 6.2], and is, in fact, the SSA
approximation of p0 . The results are given in Table 2.
Figure 9 contains plots of p0 along with approximations p0 obtained from (8) with W and τ
as defined above. The effects of the different weighting schemes can clearly be seen, also the effect
of decreasing τ is visible.
Figure 10 contains a plot of the logarithm of the square root of the mean square error of the
forecast against α and τ for weighting scheme W3 . The smallest square root of the mean square
error is 219.91 obtained at α “ 0.01 and τ “ 8000.
We now make some remarks:
• As the rank increases, or more formally, as τ increases, then the quality of the forecast
improves.
• W1 appears to be the worst weighting scheme. Recall that in this case the weights are given in
(11). It can be seen that this weighting has the unfortunate characteristic of ‘down-weighting’
observations towards the end of the vector p0 , contradictory to the argument that the more
recent observations are more important for forecasting.
• Weighting scheme W3 gives increased weight to the more recent observations. It can be seen
that this weighting scheme gives the best forecast, in the sense that it provides a forecast
with the smallest mean square error (and smaller than the best mean square error found in
[30]).
5.3. Simulation study
We now consider a simulated example where we consider a time series p0 of n “ 100 observations
of the form pi “ si ` εi where si “ cosp2πi{10q (denoted Case 1) or si “ expp0.02iq cosp2πi{10q
(denoted Case 2) and εi is a white noise (Gaussian) error term with zero mean and standard
deviation σ “ 0.1. We truncate the series to the first n ´ m terms and the aim of this study is to
forecast the remaining m observations for m “ 1, 2, . . . , 15. We take L “ 30 in this example.
Figures 11 and 12 contain the square root of the mean square errors obtained from forecasting
the m remaining observations, using (8) under two possible weighting schemes, for both Case 1
and Case 2. We consider three weighting schemes (W1 , W2 , W3 ) defined in Section 2.3. We select
16
11000
11000
11000
10000
10000
10000
9000
9000
9000
8000
8000
8000
7000
7000
0
20
40
60
7000
0
(a) W1 , rank 3
20
40
60
0
(b) W1 , rank 6
11000
11000
10000
10000
10000
9000
9000
9000
8000
8000
8000
7000
0
20
40
60
20
40
60
0
(e) W2 , rank 6
11000
11000
10000
10000
10000
9000
9000
9000
8000
8000
8000
7000
20
40
(g) W3 , rank 3
60
20
40
60
(f) W2 , rank 12
11000
0
60
7000
0
(d) W2 , rank 3
7000
40
(c) W1 , rank 12
11000
7000
20
7000
0
20
40
(h) W3 , rank 6
60
0
20
40
60
(i) W3 , rank 12
Figure 9: Plot of the data (solid line) with approximation and forecasts (dashed line).
τ such that the solutions obtained are as least as close to the unstructured rank 2 approximation
Spp0 q in each of the weighting schemes given above, in the same way as in subsection 5.2.
We remark that weighting scheme W3 provides the better forecasts, especially for Case 2. Given
the form of the time series in Case 2, this result is to be expected. The time series in this case
grows exponentially in time. This gives an indication as to the merits of alternative weighting
schemes, as opposed to the ‘traditional’ weighting scheme given by W1 .
6. Conclusion
In this paper we have considered matrix completion as a tool for forecasting in time series
analysis. We have formulated a nuclear norm relaxation of structured low-rank matrix completion
suitable for this purpose, and have demonstrated its practical potential towards the end of the
paper. We have shown that the time series should be sufficiently damped (i.e. the exponentials
should not be increasing very fast) for the nuclear norm approach to work. This becomes par17
log10(rootMSE)
3
2.5
15000
2
0.2
10000
0.15
0.1
5000
0.05
tau
alpha
Figure 10: Logarithm of the square root of the mean square error of the forecast against α and τ for weighting
scheme W3 .
0.4
0.4
0.4
0.35
0.35
0.35
0.3
0.3
0.3
0.25
0.25
0.25
0.2
0.2
0.2
0.15
0.15
0.15
0.1
0.1
0.1
0.05
0.05
0.05
0
0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15
0
1
2
3
4
(a) W “ W1
5
6
7
8
9 10 11 12 13 14 15
1
2
3
4
(b) W “ W2
5
6
7
8
9 10 11 12 13 14 15
(c) W “ W3
Figure 11: Case 1; Square root of the mean square errors against the number of forecasted observations m for three
weighting schemes. Taken over 50 simulations.
0.6
0.6
0.6
0.5
0.5
0.5
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0
0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15
(a) W “ W1
0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15
(b) W “ W2
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15
(c) W “ W3
Figure 12: Case 2; Square root of the mean square errors against the number of forecasted observations m for three
weighting schemes. Taken over 50 simulations.
ticularly important when the rank of the time series or the number of values to be forecasted is
18
large. An approach based on exponential weighting is proposed, which is shown to be equivalent to
preprocessing of time series. Numerical experiments indicate that the use of exponential weighting
improves the performance of the nuclear norm forecasting.
Appendix A. Structured matrices and optimality conditions
The proofs of Theorems 3.5 and 3.4 and are based on the following necessary and sufficient
condition for a global optimum of (5) or (21).
p ˚ P Cn`m and the
Lemma Appendix A.1 (A generalization of [20, Proposition 14]). Let p
compact SVD of Spp
p˚ q be given by
UΣVH “ Spp
p˚ q.
Further, let Q1 “ UUH and Q2 “ VVH and B “ UVH . Then the following statements hold true.
p ˚ is a global minimizer of (21) (and a minimizer of (5) is Spp
1. p
p˚ q is real) if and only if there
exists a matrix M P Rmˆn such that }M}2 ď 1 and
xQ1 MQ2 ` B, Sk y “ 0,
(A.1)
holds for all k P tn ` 1, . . . , n ` mu.
p ˚ is the unique minimizer of (21) (and the minimizer (5)
2. If, in addition, }M}2 ă 1, then p
is Spp
p˚ q is real).
The matrix M in Lemma Appendix A.1 is called the optimality certificate. In fact, the proofs
of Theorems 3.5 and 3.4 proceed by constructing such an optimality certificate.
Appendix B. Proofs
Proof of Theorem 3.1. The proof follows from [26, Theorem 5.13], applied to a submatrix of Sppq.
More precisely, under the conditions of the theorem, the matrix Sppq contains an r ˆ pr ` 1q matrix
of rank r with only known entries (no missing values). By [26, §5] and [26, Theorem 5.13] in
particular, there exists only one extension to an L ˆ K matrix that preserves the rank r (any other
extension would increase the rank). This is exactly the completion given in (19).
Proof of Corollary 4.1. Let p̃ be the solution given in (19). By Theorem 3.5 the solution of (18)
is the unique minimiser of (21) for the matrix (20). Moreover, from the proof of Theorem 3.5 (see
also Lemma Appendix A.1 and the remark after it), there exists a matrix M P RLˆL , }M}2 ă 1
such equation (A.1) holds for all k P tL ` 1, . . . , n ` mu. Hence, in particular, equation (A.1)
holds for all k P tn ` 1, . . . , n ` mu, and by Lemma Appendix A.1, the point p̃ is also the unique
minimiser of (21) for the structure as given in (22).
Proof of Proposition 4.2. The solution of (4) is given by continuing the formula (23), i.e. p̃k “ cλk .
Let us show that the point p̃ is optimal for (18).
Denote the matrix X “ Spp̃q. Since the matrix is rank-one, its compact SVD has the form
X “ σuvH , and therefore, we have that B “ X
σ.
19
matrix X is rank-one, so X “ cabH ,
˛
Let us find explicitly the terms of the compact SVD. The
where
¨
¨
˛
1
1
˚ λ
˚ λ ‹
˚
˚
‹
a “ ˚ . ‹, b “ ˚
˚ ...
.
˝ . ‚
˝
λL´1
λ
K´1
‹
‹
‹.
‹
‚
Assume that ρ “ |λ|, λ “ ρ exppωq, is the imaginary unit, and c “ |c| exppφq. Then the singular
vectors u, v of a compact SVD can be found as
u“
a exppφq
,
}a}2
,v “
b
.
}b}2
hence the singular value σ can be expressed as
$?
& |pρ2L ´1qpρ2K ´1q|
, |ρ| ‰ 1,
|ρ2 ´1|
σ “ |c|}a}2 }b}2 “ |c| ¨ ?
% LK,
|ρ| “ 1.
In addition, the matrices B, Q1 and Q2 defined in Lemma Appendix A.1 are given by
ˆ
˙
ˆ
˙
aaH
bbH
Q1 “ I ´
, Q2 “ I ´
.
}a}22
}b}22
In what follows, we use the optimality condition given in Lemma Appendix A.1. Let us find a
matrix M that satisfies (A.1). We will do that by truncating the matrix X and using [20, Theorem
6] for the truncated matrix.
Consider the lower right pm ` 1q ˆ pm ` 1q submatrix of the matrix in (22)
˛
¨
pn´m pn´m`1 ¨ ¨ ¨
pn
‹
˚
˚ pn´m`1 pn´m`2 . . . pn`1 ‹
1
‹
˚
(B.1)
S ppq “ ˚
.. ‹ .
..
.
.
.
.
˝
.
.
.
. ‚
pn
pn`1
Then the matrix X1 “ S 1 pp̃q has the form
X1 “ cλn´m yyT ,
¨ ¨ ¨ pn`m
¨
˚
˚
y“˚
˝
1
λ
..
.
λm
˛
‹
‹
‹,
‚
and the singular vectors in its compact SVD X1 “ σ 1 u1 v1H can be chosen as
u1 “
y expppφ ` ωpm ´ nqqq
,
}y}2
v“
y
.
}y}2
Hence, the singular value is equal to
1
n´m
σ “ |c||ρ|
}y}22
n´m
“ |c||ρ|
20
¨
#
|ρ2pm`1q ´1|
|ρ2 ´1| ,
m ` 1,
|ρ| ‰ 1,
|ρ| “ 1.
1
Again, we denote B1 “ u1 v1H , Q1 “ pI ´ u1 u1H q, where the matrix B1 can be found as B1 “ X
σ1 .
Let us construct the optimality certificate. By Theorem 3.4 ([20, Theorem 6]), there exists a
matrix M P Cpm`1qˆpm`1q such that
xB1 ` Q1 M1 Q1T , S1k y “ 0,
@k P tn ` 1, . . . , pn ` mqu.
and also }M1 }2 “ |ρ|m , Q1 M1 Q1T “ M1 .
Next, we define the matrix M P CLˆK as follows:
˙
ˆ
σ1 0 0
.
M“
σ 0 M1
Then since QMQT “ M we have that
xB ` QMQT , Sk y “ xB, Sk y ` xM, Sk y,
where the first term can be found as
xB, Sk y “
1
1
σ1
xX, Sk y “ xX1 , S1k y “ xB1 , S1k y
σ
σ
σ
Hence we have that
xB ` QMQT , Sk y “
σ1 1 1
σ1
xB , Sk y ` xM1 , S1k y “ 0,
σ
σ
and M satisfies (A.1). Since, by construction, }M}2 “
Lemma Appendix A.1.
σ1
m
σ |ρ|
“ Cpρq, the proof is complete by
Proof of Corollary 4.4. For fixed m, n, the value of Cpρq is minimized if
$?
&cLK,
ˇ
% ˇˇ1 ´
ˇˇ
1 ˇˇ
1´
2L
ρ ˇˇ
|ρ| “ 1,
ˇ
1 ˇ
, |ρ| ‰ 1,
ρ2K ˇ
is maximized. It is clear that this quantity is a geometric mean of two quantities that monotonically
depend on L and K, which is maximised if L and K are as close as possible to each other.
Before proving Proposition 4.5, we will need a technical lemma about the bases of subspaces.
ˆ p1q ˙
A
Lemma Appendix B.1. Let A “
P Cnˆr , where Ap1q P Cn1 ˆr , Ap2q P Cn2 ˆr such that
Ap2q
rank A “ rank Ap2q “ r. Let Up2q P Cn2 ˆr and U P Cnˆr be the orthonormal bases for column
spaces of Ap2q and A respectively.
p2q
Further, let m ď n2 and UÓm , UÓm be the matrices containing last m rows of U, Up2q respectively:
p2q
p2q
UÓm “ Un´m`1:n,: , UÓm “ Un2 ´m`1:n2 ,: .
Then it holds that
p2q
}UÓm }F ď }UÓm }F .
21
(B.2)
Proof. First, we only need to prove (B.2) for n1 “ 1, so that Ap2q is A with the first row removed.
The case n1 ą 1 can be proved by recursively removing the first row n1 times. Second, the quantities
in (B.2) do not depend on particular bases of subspaces due to invariance of the Frobenius norm
under the unitary transformations. Hence, we can assume, for example, that U and Up2q are
obtained from the thin QR factorizations, i.e. A “ UR and Ap2q “ Up2q Rp2q .
Consider the case n1 “ 1, n2 “ n ´ 1, and the full QR factorization of A:
ˆ ˙
`
˘
R
A“Q
, where Q “ U W , W P Cnˆpn´rq .
0
Without loss of generality, we can assume that W has the form
ˆ
˙
β 0
W“
,
w ˚
where w P Cn2 and β P C, i.e. only one element in the first row of W is nonzero. Next, we use the
fact that Ap2q is obtained from A by removing the first row. From [29, Sec. 12.5.3] and from the
structure of W we have that
˛
¨
¨ T ˛
˙ vT
ˆ
ˆ p1q ˙
v
˘1
0
A
˝Rp2q ‚
(B.3)
A“
“ pQGr ¨ ¨ ¨ G1 q ˝Rp2q ‚ “
0 Qp2q
Ap2q
0
0
where v is a vector and Gk P Cnˆn , k P t1, . . . , ru are appropriate Givens rotations
in
˙
ˆ rotating
p2q
˘
`
R
is the
the plane of the indices pk, k ` 1q, and Qp2q “ Up2q Wp2q , so that Ap2q “ Qp2q
0
full QR factorization of Ap2q . Finally, since the rotation matrices Gr , . . . , G1 affect only first r ` 1
columns of Q, we have that
˙
ˆ
ˆ ˙˙
ˆ
˘1
0
β
r
r
U
Gr ¨ ¨ ¨ G1 “
w
0 Up2q
r r P Cpr`1qˆpr`1q are the reduced Givens rotation matrices. Hence, if wÓm is the vector of
where G
the last m elements of w, then
p2q
}UÓm }2F “ }UÓm }2F ` }wÓm }22 ď }UÓm }2F ,
which completes the proof.
Proof of Proposition 4.5. We are going to prove that there exists such a (certificate) matrix M P
CKˆL that satisfies (A.1). Similarly to the proof for Proposition 4.2, we are going to construct the
certificate in the form
ˆ
˙
0 0
, M1 P Cpm2 `1qˆpm2 `1q .
M“
0 M1
Then (A.1) can be rewritten as where
xBred ` Q1 M1 Q1T , S1k y “ 0,
22
(B.4)
where Bred P Cpm`1qˆpm`1q is the lower right submatrix of B and Q1 “ Im2 `1 ´P1 P Cpm2 `1qˆpm2 `1q
and P1 the projector on the column space of the lower right pm2 ` 1q ˆ pm2 ` 1q submatrix of Sppq
defined as
¨
˛
pn`m´2m2 pn`m´m2 `1 ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ ¨ pn`m´m2
..
˚
‹
.
˚ pn`m´m2 `1 pn`m´m2 `2 ¨ ¨ ¨ . .
‹
.
¨¨¨
˚
‹
˚
‹
..
.
.
.
.
˚
‹
.
.
.
¨¨¨
¨¨¨
pn
‹.
S 1 ppq “ ˚
(B.5)
˚
‹
.
..
.. ... ¨ ¨ ¨
..
˚
‹
.
.
pn`1 ‹
˚
˚
‹
..
..
.
.
.
.
.
.
.
.
˝
‚
.
.
.
.
.
.
pn`m´m2
M1
¨¨¨
pn
pn`1 ¨ ¨ ¨
pn`m
We are going to look for the matrix
that satisfies (B.4) for k P tn ` m ´ m2 ` 1, . . . , n ` mu (
a few extra constraints are added). As in [20, eqn. (23)], we rewrite the constraints as
ApP1 q vecpM1 q “ ´pS1 qT vecpBred q,
where
and
(B.6)
`
˘
S1 “ vecpSn`m´m2 `1 q ¨ ¨ ¨ vecpSn`m q ,
ApP1 q “ pS1 qT ppIm2 `1 ´ P1 q b pIm2 `1 ´ P1 qq “ pS1 qT pQ1 b Q1 q.
are the same matrices as in [20, eqn. (21)–(22)].
Next, as in the proof of [20, Lemma 21] we are going to find M1˚ such that it has the minimal
Frobenius norm and satisfies (B.4). Such a matrix is given by
vecpM1˚ q “ ´ApP1 q: pS1 qT vecpBred q “ ´ApP1 q: pS1 qT vecpBred ´ P10 Bred pP10 qT q,
where, as in [20, eqn. (27)]
M“
ˆ
˙
Ir 0
P Cpm2 `1qˆpm2 `1q
0 0
is the projector on the subspace spanned by the first r unit vectors.
As in [20, eqn. (31)], we note that
}M1˚ }2 ď } vecpM1˚ q}2 ď }ApP1 q: }2 }pS1 qT }2 }Bred ´ P10 Bred pP10 qT }F .
Now if prove that, as in [20, Lemma 20],
}Bred ´ P10 Bred pP10 qT }F ď }P1 ´ P10 }F .
then }M1˚ }2 will be bounded by the right hand side of [20, eqn. (31)], and the result will hold true
analogously to [20, Theorem 7] (which only relies on [20, eqn. (31)]).
As in the proof of [20, Lemma 20], we have that
}Bred ´ P10 Bred pP10 qT }2F ď }pI ´ P10 qBred }2F ` }P10 Bred pI ´ P10 q}2F
H
}2 ,
ď }UÓm2 `1´r }2F ` }VÓm
2 `1´r F
(B.7)
where the UÓm2 `1´r , VÓm2 `1´r P Cpm2 `1´rqˆr contain the last m2 ` 1 ´ r rows of the SVD factors
U and V of Sppq respectively.
23
Finally, by Lemma Appendix B.1, we have that the right hand side in (B.7) is bounded by
H
}UÓm2 `1´r }2F ` }VÓm
}2 ď 2}U1Óm2 `1´r }2F “ }P1 ´ P0 }2F
2 `1´r F
where U1 P Cpm2 `1qˆr is the left factor of the SVD of S 1 ppq defined in (B.5), and the last equation
is from [20, eqn. (29)] and the fact P1 “ U1 pU1 qH . Hence, the proof is complete.
Proof of Corollary 4.3. The case |ρ| “ 1 is easy. Consider the case |ρ| ‰ 1. It is evident that
Cpρq ă 1 if and only if
|ρm`1 ´ ρ´pm`1q | ă y.
Denote x “ |ρ|m`1 ą 0. If |ρ| ą 1, then the condition becomes
1
x ´ ă y ðñ x2 ´ yx ´ 1 ă 0 ðñ x ă
x
If |ρ| ă 1, the condition becomes
1
´ x ă y ðñ x2 ` yx ´ 1 ą 0 ðñ x ą
x
a
a
y 2 ` 4`y
.
2
y2 ` 4 ´ y
.
2
The statement follows from the properties of the logarithm.
Proof of Proposition 4.6. Consider the change of variables (25). Then it is obvious that (26) is
equivalent to (27). Indeed, it follows from (24) and from the fact that
}psc,p1:nq ´ psc,0 }W3 “ }pp1:nq ´ p0 }W2 .
The scaling (28) is needed to perform the change of variables that is inverse to (25).
References
[1] A.Dokumentov, R.Hyndman, Low-dimensional decomposition, smoothing and forecasting of sparse functional
data, Tech. Rep. Working Paper 16/14, Monash University, Department of Econometrics and Business Statistics
(2014).
[2] N. Golyandina, V. Nekrutkin, A. A. Zhigljavsky, Analysis of time series structure: SSA and related techniques,
CRC press, 2001.
[3] M. A. R. Khan, D. Poskitt, Forecasting stochastic processes using singular spectrum analysis: Aspects of the theory and application, International Journal of Forecasting 33 (1) (2017) 199–213.
doi:https://doi.org/10.1016/j.ijforecast.2016.01.003.
[4] F. Papailias, D. Thomakos, Exssa:
Ssa-based reconstruction of time series via exponential
smoothing of covariance eigenvalues, International Journal of Forecasting 33 (1) (2017) 214–229.
doi:https://doi.org/10.1016/j.ijforecast.2016.08.004.
[5] H. Hassani,
S. Heravi,
A. Zhigljavsky,
Forecasting european industrial production with
singular spectrum analysis,
International Journal of Forecasting 25 (1) (2009) 103–118.
doi:https://doi.org/10.1016/j.ijforecast.2008.09.007.
[6] H. Butcher, J. Gillard, Simple nuclear norm based algorithms for imputing missing data and forecasting in time
series, Statistics and Its Interface 10 (3) (2017) 19–25. doi:10.4310/SII.2017.v10.n1.a2.
[7] M. Fazel, Matrix rank minimization with applications, Ph.D. thesis, PhD thesis, Stanford University (2002).
[8] N. Gillis, F. Glineur, Low-rank matrix approximation with weights or missing data is NP-hard, SIAM Journal
on Matrix Analysis and Applications 32 (4) (2011) 1149–1165.
[9] I. Rish, G. Grabarnik, Sparse modeling: theory, algorithms, and applications, CRC press, 2014.
[10] E. J. Candès, B. Recht, Exact Matrix Completion via Convex Optimization, Foundations of Computational
Mathematics 9 (6) (2009) 717–772. doi:10.1007/s10208-009-9045-5.
24
[11] E. J. Candes, Y. Plan, Matrix completion with noise, Proceedings of the IEEE 98 (6) (2010) 925–936.
[12] B. Recht, M. Fazel, P. A. Parrilo, Guaranteed minimum-rank solutions of linear matrix equations via nuclear
norm minimization, SIAM review 52 (3) (2010) 471–501.
[13] Y. Chen, Y. Chi, Robust spectral compressed sensing via structured matrix completion, IEEE Transactions on
Information Theory 60 (10) (2014) 6576–6601. doi:10.1109/TIT.2014.2343623.
[14] Z. Liu, L. Vandenberghe, Interior-point method for nuclear norm approximation with application to system
identification, SIAM J. Matrix Ana. Appl. 31 (3) (2009) 1235–1256.
[15] M. Verhaegen, A. Hansson, N2sid: Nuclear norm subspace identification of innovation models, Automatica
72 (Supplement C) (2016) 57–63. doi:https://doi.org/10.1016/j.automatica.2016.05.021.
[16] N. Blomberg, On nuclear norm minimization in system identification, Ph.D. thesis, KTH Royal Institute of
Technology, licenciate thesis (2016).
[17] I. Markovsky, How effective is the nuclear norm heuristic in solving data approximation problems?,
in: Proc. of the 16th IFAC Symposium on System Identification, Brussels, 2012, pp. 316–321.
doi:10.3182/20120711-3-BE-2027.00125.
[18] I. Markovsky, Low Rank Approximation: Algorithms, Implementation, Applications, Communications and
Control Engineering, Springer, 2012.
[19] L. Dai, K. Pelckmans, On the nuclear norm heuristic for a Hankel matrix completion problem, Automatica 51
(2015) 268–272.
[20] K. Usevich, P. Comon, Hankel low-rank matrix completion: Performance of the nuclear norm relaxation, IEEE
Journal of Selected Topics in Signal Processing 10 (4) (2016) 637–646.
[21] M. Grant, S. Boyd, CVX: Matlab software for disciplined convex programming, version 2.1,
http://cvxr.com/cvx (Mar. 2014).
[22] M. Grant, S. Boyd, Graph implementations for nonsmooth convex programs, in: V. Blondel, S. Boyd, H. Kimura
(Eds.), Recent Advances in Learning and Control, Lecture Notes in Control and Information Sciences, SpringerVerlag Limited, 2008, pp. 95–110, http://stanford.edu/~ boyd/graph_dcp.html.
[23] J. Gillard, A. Zhigljavsky, Optimization challenges in the structured low rank approximation problem, Journal
of Global Optimization 57 (3) (2013) 733–751.
[24] J. Mairal, F. Bach, J. Ponce, Sparse modeling for image and vision processing, Foundations and Trends in
Computer Graphics and Vision 8 (2-3) (2014) 85–283. doi:10.1561/0600000058.
[25] K. Usevich, On signal and extraneous roots in Singular Spectrum Analysis, Statistics and Its Interface 3 (3)
(2010) 281–295. doi:10.4310/SII.2010.v3.n3.a3.
[26] G. Heinig, K. Rost, Algebraic methods for Toeplitz-like matrices and operators, Birkhäuser, Boston, 1984.
[27] I. Iohvidov, Toeplitz and Hankel matrices and forms. Algebraic theory, Birkhäuser, Basel, 1982.
[28] I. Markovsky, A missing data approach to data-driven filtering and control, IEEE Trans. Automat. Contr. 62
(2017) 1972–1978.
[29] G. H. Golub, C. F. Van Loan, Matrix computations, Vol. 3, JHU Press, 2012.
[30] P. J. Brockwell, R. A. Davis, Time series: theory and methods, Springer Science & Business Media, 2009.
[31] G. E. Box, G. M. Jenkins, G. C. Reinsel, Time series analysis: forecasting and control, John Wiley & Sons,
2013.
25
| 3cs.SY
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.