Professional Documents
Culture Documents
Textbook Random Walks and Heat Kernels On Graphs 1St Edition Martin T Barlow Ebook All Chapter PDF
Textbook Random Walks and Heat Kernels On Graphs 1St Edition Martin T Barlow Ebook All Chapter PDF
Textbook Random Walks and Heat Kernels On Graphs 1St Edition Martin T Barlow Ebook All Chapter PDF
https://textbookfull.com/product/probability-on-graphs-random-
processes-on-graphs-and-lattices-geoffrey-grimmett/
https://textbookfull.com/product/probability-on-graphs-random-
processes-on-graphs-and-lattices-second-edition-grimmett/
https://textbookfull.com/product/generating-random-networks-and-
graphs-1st-edition-annibale/
https://textbookfull.com/product/random-graphs-geometry-and-
asymptotic-structure-1st-edition-michael-krivelevich/
From Random Walks to Random Matrices - Selected Topics
in Modern Theoretical Physics 1st Edition Jean Zinn-
Justin
https://textbookfull.com/product/from-random-walks-to-random-
matrices-selected-topics-in-modern-theoretical-physics-1st-
edition-jean-zinn-justin/
https://textbookfull.com/product/limit-theorems-for-some-long-
range-random-walks-on-torsion-free-nilpotent-groups-
springerbriefs-in-mathematics-chen/
https://textbookfull.com/product/heat-exchangers-first-edition-
holger-martin/
https://textbookfull.com/product/asl-grammar-the-workbook-1st-
edition-rochelle-barlow/
LONDON MATHEMATICAL SOCIETY LECTURE NOTE SERIES
The titles below are available from booksellers, or from Cambridge University Press at
http://www.cambridge.org/mathematics
M A RT I N T. BA R L OW
University of British Columbia, Canada
University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
4843/24, 2nd Floor, Ansari Road, Daryaganj, Delhi – 110002, India
79 Anson Road, #06–04/06, Singapore 079906
www.cambridge.org
Information on this title: www.cambridge.org/9781107674424
DOI: 10.1017/9781107415690
c Martin T. Barlow 2017
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2017
Printed in the United Kingdom by Clays, St Ives plc
A catalogue record for this publication is available from the British Library.
Library of Congress Cataloging-in-Publication Data
Names: Barlow, M. T.
Title: Random walks and heat kernels on graphs / Martin T. Barlow,
University of British Columbia, Canada.
Description: Cambridge : Cambridge University Press, [2017] |
Series: London Mathematical Society lecture note series ; 438 |
Includes bibliographical references and index.
Identifiers: LCCN 2016051295 | ISBN 9781107674424
Subjects: LCSH: Random walks (Mathematics) | Graph theory. |
Markov processes. | Heat equation.
Classification: LCC QA274.73 .B3735 2017 | DDC 511/.5–dc23
LC record available at https://lccn.loc.gov/2016051295
ISBN 978-1-107-67442-4 Paperback
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party Internet Web sites referred to in this publication
and does not guarantee that any content on such Web sites is, or will remain,
accurate or appropriate.
To my mother, Yvonne Barlow, and in memory of my father, Andrew Barlow.
Contents
Preface page ix
1 Introduction 1
1.1 Graphs and Weighted Graphs 1
1.2 Random Walks on a Weighted Graph 6
1.3 Transition Densities and the Laplacian 11
1.4 Dirichlet or Energy Form 15
1.5 Killed Process 21
1.6 Green’s Functions 22
1.7 Harmonic Functions, Harnack Inequalities, and the
Liouville Property 26
1.8 Strong Liouville Property for Rd 31
1.9 Interpretation of the Liouville Property 32
vii
viii Contents
The topic of random walks on graphs is a vast one, and has close connections
with many other areas of probability, as well as analysis, geometry, and alge-
bra. In the probabilistic direction, a random walk on a graph is just a reversible
or symmetric Markov chain, and many results on random walks on graphs also
hold for more general Markov chains. However, pursuing this generalisation
too far leads to a loss of concrete interest, and in this text the context will be
restricted to random walks on graphs where each vertex has a finite number of
neighbours.
Even with these restrictions, there are many topics which this book does not
cover – in particular the very active field of relaxation times for finite graphs.
This book is mainly concerned with infinite graphs, and in particular those
which have polynomial volume growth. The main topic is the relation between
geometric properties of the graph and asymptotic properties of the random
walk. A particular emphasis is on properties which are stable under minor per-
turbations of the graph – for example the addition of a number of diagonal
edges to the Euclidean lattice Z2 . The precise definition of ‘minor perturbation’
is given by the concept of a rough isometry, or quasi-isometry. One example
of a property which is stable under these perturbations is transience; this sta-
bility is proved in Chapter 2 using electrical networks. A considerably harder
theorem, which is one of the main results of this book, is that the property of
satisfying Gaussian heat kernel bounds is also stable under rough isometries.
The second main theme of this book is deriving bounds on the transition
density of the random walk, or the heat kernel, from geometric information on
the graph. Once one has these bounds, many properties of the random walk
can then be obtained in a straightforward fashion.
Chapter 1 gives the basic definition of graphs and random walks, as well
as that of rough isometry. Chapter 2 explores the close connections between
ix
x Preface
random walks and electrical networks. The key contribution of network the-
ory is to connect the hitting properties of the random walk with the effective
resistance between sets. The effective resistance can be bounded using varia-
tional principles, and this introduces tools which have no purely probabilistic
counterpart.
Chapter 3 introduces some geometric and analytic inequalities which will
play a key role in the remainder of the book – two kinds of isoperimetric
inequality, Nash inequalities, and Poincaré inequalities. It is shown how the
two analytic inequalities (i.e. Nash and Poincaré) can be derived from the
geometric information given by the isoperimetric inequalities.
Chapter 4 studies the transition density of the discrete time random walk,
or discrete time heat kernel. Two initial upper bounds on this are proved – an
‘on-diagonal’ upper bound from a Nash inequality, and a long range bound,
the Carne–Varopoulos bound, which holds in very great generality. Both these
proofs are quite short. To obtain full Gaussian or sub-Gaussian bounds much
more work is needed. The second half of Chapter 4 makes the initial steps, and
introduces conditions (bounds on exit times from balls, and a ‘near-diagonal
lower bound’) which if satisfied will lead to full Gaussian bounds.
Chapter 5 introduces the continuous time random walk, which is in many
respects an easier and more regular object to study than the discrete time ran-
dom walk. In this chapter it is defined from the discrete time walk and an
independent Poisson process – another construction, from the transition densi-
ties, is given in the Appendix. It is proved that Gaussian or sub-Gaussian heat
kernel bounds hold for the discrete time random walk if and only if they hold
for the continuous time walk.
Chapter 6 proves sub-Gaussian or Gaussian bounds for two classes of
graphs. The first are ‘strongly recurrent’ graphs; given the work done in
Chapter 4 the results for these come quite easily from volume and resistance
estimates. The second class are graphs which satisfy a Poincaré inequality –
these include Zd and graphs which are roughly isometric to Zd . The proof of
Gaussian bounds for these graphs is based on the tools developed by Nash to
study divergence form PDE; these methods also work well in the graph context.
Chapter 7 gives a brief introduction to potential theory. Using the method
of balayage, the heat kernel bounds of Chapter 6 lead to a Harnack inequality,
and this in turn is used to prove the strong Liouville property (i.e. all positive
harmonic functions are constant) for graphs which are roughly isometric to Zd .
This book is based on a course given at the first Pacific Institute for the
Mathematical Sciences (PIMS) summer school in Probability, which was held
at UBC in 2004. It is written for first year graduate students, though it is prob-
ably too long for a one semester course at most institutions. A much shorter
Preface xi
version of this course was also given in Kyoto in 2005, and I wish to thank
both PIMS and the Research Institute for the Mathematical Sciences (RIMS)
for their invitations to give these courses.
I circulated a version of these notes several years ago, and I would like to
thank those who have read them and given feedback, and in particular, Alain
Sznitman and Gerard Ben-Arous, for encouraging me to complete them. I also
wish to thank Alain for making available to me some lecture notes of his own
on this material.
I have learnt much from a number of previous works in this area – the
classical text of Spitzer [Spi], as well as the more recent works [AF, LL, LP,
Wo].
I am grateful to those who have sent me corrections, and in particular to
Yoshihiro Abe, Alma Sarai Hernandez Torres, Guillermo Martinez Dibene,
Jun Misumi, and Zichun Ye.
1
Introduction
1
2 Introduction
Note that while the metric d(x, ·) is integer valued, we allow r here to be
real. Of course B(x, r ) = B(x, r ).
(7) We define the exterior boundary of A by
∂ A = {y ∈ Ac : there exists x ∈ A with x ∼ y}.
Set also
We write 1 for the function f which is identically 1, and 1 A for the indicator
function (or characteristic function) of the set A. We write 1x for 1{x} .
(i) μx y = μ yx ,
(ii) μx y ≥ 0 for all x, y ∈ V,
(iii) if x = y then μx y > 0 if and only if x ∼ y.
Definition 1.2 We say (, μ) has bounded weights if there exists C1 < ∞
such that C 1−1 ≤ μe ≤ C 1 for all e ∈ E, and μx x ≤ C1 for all x ∈ V. We say
(, μ) has controlled weights if there exists C 2 < ∞ such that
μx y 1
≥ whenever x ∼ y.
μx C2
Lemma 1.3 Suppose (, μ) satisfies (H5) with constant C2 . Then (H3) holds
and, for n ≥ 0, |B(x, n)| ≤ 2C2n , μ(B(x, n)) ≤ 2C2n μx .
μxHy = μx y , x, y ∈ H.
We define μxH = y∈H μxHy ; note that μxH ≤ μx . Clearly (H, E H ) need
not be connected, even if is.
(2) Collapse of a subset to a point. Let A ⊂ V. The graph obtained by col-
lapsing A to a point a (where a ∈ V) is the graph = (V , E ) given
by
V = (V − A) ∪ {a},
E = {x, y} ∈ E : x, y ∈ V − A ∪ {x, a} : x ∈ ∂ A .
Note that μxa ≤ μx < ∞. In general this graph need not be locally finite,
but will be if |∂ A| < ∞. It is easy to check that if is connected then so
is .
6 Introduction
(3) Join of two graphs. Let i = (Vi , E i ), i = 1, 2 be two graphs. (We regard
V1 , V2 as being distinct sets.) If xi ∈ Vi , the join of 1 and 2 at x1 , x2 is
the graph (V, E) given by
V = V1 ∪ V2 , E = E 1 ∪ E 2 ∪ {x 1 , x2 } .
μ (2)
(x,y1 ),(x,y2 ) = μ y1 ,y2 , μ (1)
(x 1 ,y),(x 2 ,y) = μx 1 ,x 2 ,
μ× (1) (2)
(x1 ,y1 ),(x2 ,y2 ) = μx 1 ,x 2 μ y1 ,y2 .
We call X the simple random walk (SRW) on (, μ) and (P(x, y)) the tran-
sition matrix of X . At each time step X moves from a vertex x to a neighbour
y with probability proportional to μx y . Since we can have μx x > 0 we do
in general allow X to remain at a vertex x with positive probability. If ν
1.2 Random Walks on a Weighted Graph 7
Set
Pn (x, y) = Px (X n = y).
μx0 Px0 (X 0 = x0 , X 1 = x1 , . . . , X n = xn )
= μxn Pxn (X 0 = xn , X 1 = x n−1 , . . . , X n = x0 ).
Proof Using the μ-symmetry of P it is easy to check that both sides equal
n
i=1 μxi−1 ,xi .
T A = min{n ≥ 0 : X n ∈ A},
T A+ = min{n ≥ 1 : X n ∈ A}.
Here we adopt the convention that min ∅ = +∞, so that T A = ∞ if and only
if X never hits A. Note also that if X 0 ∈ A then T A = T A+ . Write Tx = T{x}
and
τ A = T Ac = min{n ≥ 0 : X n ∈ A} (1.4)
for the time of the first exit from A. If A, B ⊂ V we will write {T A < TB }
for the event that could more precisely be denoted {T A < TB , T A < ∞}.
If A, B are disjoint then the sample space can be written as the disjoint
union
(d) For all x, y ∈ V with x = y either Px (Ty < ∞) < 1 or P y (Tx < ∞) < 1.
(e) For all x, y ∈ V,
∞
Px (X hits y only finitely often) = Px 1(X n =y) < ∞ = 1.
n=0
∞
Px (X hits y infinitely often) = Px 1(X n =y) = ∞ = 1.
n=0
Let us start with the Euclidean lattices. Polya [Pol] proved the following in
1921.
Polya used the fact that X n is a sum of independent random variables, and
that therefore the Fourier transform of Pn (0, x) is a product. Inverting the
Fourier transform, he deduced that P2n (0, 0) ∼ cd n −d/2 , and hence proved
his theorem.
1.2 Random Walks on a Weighted Graph 9
after a bit more work. (Or we can use the trick of writing Z n = ( 12 (X n +
Yn ), 12 (X n − Yn )), where X and Y are independent SRW on Z.) In general we
obtain cd n −d/2 , so the series converges when d ≥ 3. This expression is to be
expected: by the local limit theorem (a variant of the central limit theorem),
we have that the distribution of X n is approximately the multivariate Gaussian
N (0, n 1/2 d −1/2 Id ). (Of course the fact that, after rescaling, the SRW X on Zd
with d ≥ 3 converges weakly to a transient process does not prove that X is
transient, since transience/recurrence is not preserved by weak convergence.)
The advantage of the combinatorial argument is that it is elementary. On
the other hand, the details are a bit messy for d ≥ 3. More significantly, the
technique is not robust. Consider the following three examples:
(1) The SRW on the hexagonal lattice in R2 .
(2) The SRW on a graph derived from a Penrose tiling.
(3) The graph (, μ) where = Zd , and the weights μ satisfy μx y ∈ [c−1 , c].
Example (1) could be handled by the combinatorial method, but the details
would be awkward, since one now has to count six kinds of steps. Also it
is plainly a nuisance to have to give a new argument for each new lattice.
Polya’s method does work for this example, but both methods look hopeless
for (2) or (3), since they rely on having an exact expression for Pn (x, x) or
its transform. (For the convergence of SRW on a Penrose tiling to Brownian
motion see [T3, BT].)
Definition 1.11 Let P be some property of a weighted graph (, μ), or the
SRW X on it. P is stable under bounded perturbation of weights (weight
stable) if whenever (, μ) satisfies P, and μ are weights on such that
c−1 μx y ≤ μx y ≤ cμx y , x, y ∈ V,
then (, μ ) satisfies P. (We say the weights μ and μ are equivalent.)
10 Introduction
If there exists a rough isometry between two spaces they are said to be roughly
isometric. (One can check that this is an equivalence relation.)
Example Consider the graphs (Z+ , μ(α) ) defined in Examples 1.4, and let
ϕ : Z+ → Z+ be the identity map. Then ϕ is (of course) a rough isometry
between the graphs Z+ and Z+ , but if α = β it is not a rough isometry between
the weighted graphs (Z+ , μ(α) ) and (Z+ , μ(β) ). We will see in Chapter 2 that
(Z+ , μ(α) ) is recurrent if and only if α ≤ 1, and that the type of a graph is
stable under rough isometries (of weighted graphs).
we also have
A p→ p = sup{g A f 1 : f p ≤ 1, gq ≤ 1}. (1.10)
Proof (a) and (b) are easy. For (c), let first p ∈ [1, ∞) and q be the conjugate
index. Let f ∈ L p (V, μ). Then, using Hölder’s inequality,
p
p
||P f || p = |P f (x)| p μx = p(x, y) f (y)μ y μx
x x y
1/ p 1/q p
≤ p(x, y)| f (y)| μ y p
p(x, y)1q μ y μx
x y y
p
= p(x, y)| f (y)| p μx μ y = | f (y)| p μ y = || f || p .
x y y
Remark Note that this proof did not actually use the symmetry of p(x, y),
but just the fact that x μx p(x, y) = 1, that is, that μ is an invariant measure
for P.
≤ μx A x y | f (y)| p M p/q ≤ M 1+ p/q | f (y)| p μ y
x y y
p
=M p
|| f || p ,
|| Com || p→ p ≤ 2M.
One can then define the ‘delayed RW’ on (, μ) by considering the random
walk associated with the operator
1
P=I− Com .
M
This is the SRW on the weighted graph (V, E, μ ), where μx y = M −1 μx y if
x = y and μx x = 1 − y=x μx y .
If M = ∞ then the operator Com is not naturally associated with any
discrete time random walk. However it can be associated in a natural way with
a continuous time random walk – see Remark 5.7.
Remark 1.19 See Example 1.35 in Section 1.6 for a graph and functions
u, v such that both u and v have finite support, but
v(x) u(x)μx = u(x) v(x)μx .
x x
whenever
this sum converges absolutely. This is the discrete analogue of
∇ f · ∇g. Let
H 2 (V, μ) = H 2 (V) = H 2 = { f ∈ C(V) : E ( f, f ) < ∞}.
Choose a ‘base point’ o ∈ V and define
|| f ||2H 2 = E ( f, f ) + f (o)2 .
The sum in (1.15) is well defined for f, g ∈ H 2 (V). If A ⊂ V we will
sometimes use the notation
E A ( f, g) = 12 ( f (x) − f (y))(g(x) − g(y))μx y . (1.16)
x∈A y∈A
PLATE XLIV
When attention is to be drawn to the fact that the birds are laying,
or sitting on eggs, a number of small circles are drawn about the
track (Fig. 39).
A lizard track is indicated by drawing a number of dots, equally
spaced along a straight line, and on alternate sides of it. The dots
are the claw impressions, the line the trail left by the tail (Fig. 40).
A single wavy line, or a number of parallel wavy lines, represents a
snake or a snake track (Fig. 41).
The human footprint is either correctly shown in detail, or is
reduced to a short, straight line at one end of which five
(occasionally only three or four) dots are drawn in a sloping line to
indicate the toes. When walking is to be implied, these footprints are
either shown one behind the other in a straight line, at equal
distances apart, or they stand alternately right and left by an
imaginary central line. The common way of showing the last-named
system conventionally is to connect the alternate footprints, whether
they be actually shown or not, by a zig-zag line (Fig. 42).
In all cases mentioned above, the track may stand equally well for
the object itself.
Fig. 54. Camps consisting of a man and his wife (left) and of eight men.
There are not many places left in the world where the man of the
Stone Age can still be seen roaming the wilds of his inherited
possessions. Even in Australia, although there remain one or two
areas where comparatively little havoc has been wrought among the
primitive institutions of the indigenous man, yet the influence of
civilization is slowly, but very surely, encroaching indirectly upon his
ancient cults by such means as inter-tribal barter, if not actually by
the hand of the white intruder. Especially do these remarks apply to
the manufacture and utilization of stone implements; it is, of course,
only to be expected that the superiority of the metal blades of the
white man’s implements would appeal to the native who formerly had
to spend hours making a crude cutting edge which only too often
broke when applied to the test for the first time. We shall, however,
treat the subject regardless of the alterations which have been
brought about by our appearance upon the scene, and without
attempting to draw up a hard and fast scheme of classification.
At the present time, whilst there are not only some of the primitive
men alive still, but also a limited number of observers who have had
the good fortune of seeing them at work, it is of vastly greater
importance to record the living facts than to write exhaustively, nay,
even speculatively, upon the comparative shapes and embodied
techniques of artefacts whose stony composition will ensure their
keeping, even fossilized, long after the men who made them, and the
scientists who lived among them, have passed into oblivion.
The Australian aboriginal makes adequate use of any suitably
shaped pieces of stone he happens to find whilst in pursuit of game;
both in the Musgrave Ranges and the northern Kimberleys stones
are used in their natural shape for hurling into a flying flock of birds,
for shying at a bounding wallaby, for bringing down nuts of the
boabab, and for precipitating fledgelings out of a nest.
Other stones, usually oblong and rounded pebbles gathered in a
river bed, are used for pounding and cracking purposes. At any
camping ground these pebbles can be picked up in great numbers,
showing one or two places, usually the points, at which the
percussions have worn the stone away; pounding stones and
hammers of this description are equally plentiful in the sandhills on
the plains of Adelaide, all over central Australia, and along the north
coast. They are used for pounding seeds and foliage (the latter of
which is to serve for corrobboree decorations), for pulverizing ochre,
for cracking nuts and hollow bones containing marrow. Vide Plate
LIV, 1.
The underlying surface consists either of a level portion of an
outcrop or another, but larger, stone, which takes the place of an
anvil. Some of the coastal tribes of eastern Australia used to shape
their heavier pounding stones by chipping away material at one side
until a stout, cylindrical handle was formed, the whole resembling a
pestle; dumb-bell shaped pounders were also made, but were rare.
We have already learnt that natural pebbles or rock fragments are
also used, together with a wooden rod, for knocking out teeth during
initiation ceremonies.
PLATE XLVII