Welcome to Scribd. Sign in or start your free trial to enjoy unlimited e-books, audiobooks & documents.Find out more
Download
Standard view
Full view
of .
Look up keyword
Like this
1Activity
0 of .
Results for:
No results containing your search query
P. 1
Ricci Curvature of Finite Markov Chains via Convexity of the Entropy

Ricci Curvature of Finite Markov Chains via Convexity of the Entropy

Ratings: (0)|Views: 13|Likes:
Published by arashbehboodi

More info:

Published by: arashbehboodi on Jul 02, 2012
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

01/17/2014

pdf

text

original

 
RICCI CURVATURE OF FINITE MARKOV CHAINS VIACONVEXITY OF THE ENTROPY
MATTHIAS ERBAR AND JAN MAAS
Abstract.
We study a new notion of Ricci curvature that applies toMarkov chains on discrete spaces. This notion relies on geodesic con-vexity of the entropy and is analogous to the one introduced by Lott,Sturm, and Villani for geodesic measure spaces. In order to apply tothe discrete setting, the role of the Wasserstein metric is taken over bya different metric, having the property that continuous time Markovchains are gradient flows of the entropy.Using this notion of Ricci curvature we prove discrete analogues of fundamental results by Bakry–´Emery and Otto–Villani. Furthermorewe show that Ricci curvature bounds are preserved under tensorisation.As a special case we obtain the sharp Ricci curvature lower bound forthe discrete hypercube.
Contents
1. Introduction12. The metric
73. Geodesics174. Ricci curvature195. Examples256. Basic constructions297. Functional inequalities32References371.
Introduction
In two independent contributions Sturm[41] and Lott and Villani[27] solved the long-standing open problem of defining a synthetic notion of Riccicurvature for a large class of metric measure spaces.The key observation, proved in[39], is that on a Riemannian manifold
M
,the Ricci curvature is bounded from below by some constant
κ
R
, if andonly if the Boltzmann-Shannon entropy
H
(
ρ
) =
 
ρ
log
ρ
dvol is
κ
-convexalong geodesics in the
L
2
-Wasserstein space of probability measures on
M
.The latter condition does not appeal to the geometric structure
M
, but onlyrequires a metric (to define the
L
2
-Wasserstein metric
2
) and a referencemeasure (to define the entropy
H
). Therefore this condition can be usedin order to define a notion of Ricci curvature lower boundedness on more
Date
: June 25, 2012.JM is supported by Rubicon subsidy 680-50-0901 of the Netherlands Organisation forScientific Research (NWO).
1
 
2 MATTHIAS ERBAR AND JAN MAAS
general metric measure spaces. This notion turns out to be stable underGromov-Hausdorff convergence and it implies a large number of functionalinequalities with sharp constants. The theory of metric measure spaces withRicci curvature bounds in the sense of Lott, Sturm, and Villani is still underactive development [2,3]. However, the condition of Lott-Sturm-Villani does not apply if the
L
2
-Wasserstein space over
does not contain geodesics. Unfortunately, thisis the case if the underlying space is discrete (even if the underlying spaceconsists of only two points). The aim of the present paper is to developa variant of the theory of Lott-Sturm-Villani, which does apply to discretespaces.In order to circumvent the nonexistence of Wasserstein geodesics, we re-place the
L
2
-Wasserstein metric by a different metric
, which has been in-troduced in[28]. There it has been shown that the heat flow associated with a Markov kernel on a finite set is the gradient flow of the entropy with re-spect to
(see also the independent work [15] containing related results forFokker-Planck equations on graphs, as well as [31], where this gradient flow structure has been discovered in the setting of reaction-diffusion systems). Inthis sense,
takes over the role of the Wasserstein metric, since it is knownsince the seminal work by Jordan, Kinderlehrer, and Otto that the heat flowon
R
n
is the gradient flow of the entropy[23] (see [2,18,19,20,32,36]for variations and generalisations). Convexity along
-geodesics may thus beregarded as a discrete analogue of McCann’s displacement convexity[29], which corresponds to convexity along
2
-geodesics in a continuous setting.Since every pair of probability densities on
can be joined by a
-geodesic, it is possible to define a notion of Ricci curvature in the spiritof Lott-Sturm-Villani by requiring geodesic convexity of the entropy withrespect to the metric
. This possibility has already been indicated in[28]. We shall show that this notion of Ricci curvature shares a number of  properties which make the LSV definition so powerful: in particular, it isstable under tensorisation and implies a number of functional inequalities,including a modified logarithmic Sobolev inequality, and a Talagrand-typeinequality involving the metric
.
Main results.
Let us now discuss the contents of this paper in more detail.We work with an irreducible Markov kernel
:
X ×X →
R
+
on a finite set
, i.e., we assume that
y
∈X 
(
x,y
) = 1for all
x
, and that for every
x,y
there exists a sequence
{
x
i
}
ni
=0
such that
x
0
=
x
,
x
n
=
y
and
(
x
i
1
,x
i
)
>
0 for all 1
i
n
.Basic Markov chain theory guarantees the existence of a unique stationaryprobability measure (also called steady state)
π
on
, i.e.,
x
∈X 
π
(
x
) = 1 and
π
(
y
) =
x
∈X 
π
(
x
)
(
x,y
)
 
RICCI CURVATURE OF FINITE MARKOV CHAINS 3
for all
y
∈ X 
. We assume that
π
is
reversible
for
, which means that thedetailed balance equations
(
x,y
)
π
(
x
) =
(
y,x
)
π
(
y
) (1.1)hold for
x,y
.Let
P
(
) :=
ρ
:
X
R
+
|
x
∈X 
π
(
x
)
ρ
(
x
) = 1
be the set of 
probability densities
on
. The subset consisting of thoseprobability densities that are strictly positive is denoted by
P
(
). Weconsider the metric
defined for
ρ
0
,ρ
1
P
(
) by
(
ρ
0
,ρ
1
)
2
:= inf 
ρ,ψ
12
 
10
x,y
∈X 
(
ψ
t
(
x
)
ψ
t
(
y
))
2
ˆ
ρ
t
(
x,y
)
(
x,y
)
π
(
x
) d
t
,
where the infimum runs over all sufficiently regular curves
ρ
: [0
,
1]
P
(
)and
ψ
: [0
,
1]
R
satisfying the ‘continuity equation’
dd
tρ
t
(
x
) +
y
∈X 
(
ψ
t
(
y
)
ψ
t
(
x
))ˆ
ρ
t
(
x,y
)
(
x,y
) = 0
x
,ρ
(0) =
ρ
0
, ρ
(1) =
ρ
1
.
(1.2)Here, given
ρ
P
(
), we write ˆ
ρ
(
x,y
) :=
 
10
ρ
(
x
)
1
 p
ρ
(
y
)
 p
d
 p
for the loga-rithmic mean of 
ρ
(
x
) and
ρ
(
y
). The relevance of the logarithm mean in thissetting is due to the identity
ρ
(
x
)
ρ
(
y
) = ˆ
ρ
(
x,y
)(log
ρ
(
x
)
log
ρ
(
y
))
,
which somehow compensates for the lack of a ‘discrete chain rule’. Thedefinition of 
can be regarded as a discrete analogue of the Benamou-Brenier formula [7]. Let us remark that if 
t
ρ
t
is differentiable at some
t
and
ρ
t
belongs to
P
(
), then the continuity equation (1.2) is satisfiedfor some
ψ
t
R
, which is unique up to an additive constant (see[28, Proposition 3.26]).Since the metric
is Riemannian in the interior
P
(
), it makes senseto consider gradient flows in (
P
(
)
,
) and it has been proved in[28] that the heat flow associated with the continuous time Markov semigroup
t
=
e
t
(
)
is the gradient flow of the entropy
H
(
ρ
) =
x
∈X 
π
(
x
)
ρ
(
x
)log
ρ
(
x
)
,
(1.3)with respect to the Riemannian structure determined by
.In this paper we shall show that every pair of densities
ρ
0
,ρ
1
P
(
) canbe joined by a constant speed geodesic. Therefore the following definitionin the spirit of Lott-Sturm-Villani seems natural.
Definition 1.1.
We say that 
has
non-local Ricci curvature bounded frombelow by
κ
R
if for any constant speed geodesic
{
ρ
t
}
t
[0
,
1]
in 
(
P
(
)
,
)
we have
H
(
ρ
t
)
(1
t
)
H
(
ρ
0
) +
t
H
(
ρ
1
)
κ
2
t
(1
t
)
(
ρ
0
,ρ
1
)
2
.

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->