1: \def\versiondate{5 Jan. 2009}

2: \input math.macros

3: \input Ref.macros

4: 

5: %\proofmodetrue

6: %\leftsectionheadtrue

7: \checkdefinedreferencetrue

8: %\continuousnumberingtrue

9: \continuousfigurenumberingtrue

10: \theoremcountingtrue

11: \sectionnumberstrue

12: %\figuresectionnumberstrue

13: \forwardreferencetrue

14: %\lefteqnumberstrue

15: \tocgenerationtrue

16: \citationgenerationtrue

17: \nobracketcittrue

18: \hyperstrue

19: \initialeqmacro

20: 

21: \input\jobname.key

22: \bibsty{myapalike}

23: 

24: \def\cp{\tau}   % complexity

25: \def\gh{G}  % graph

26: \def\gp{\Gamma}  % group

27: \def\gpe{\gamma}  % group element

28: \def\verts{{\ss V}}

29: \def\VV{\verts}

30: \def\edges{{\ss E}}

31: \def\EE{\edges}

32: \def\tr{{\rm Tr}}  % trace

33: \def\bfo{{\bf 1}}

34: \def\bfz{{\bf 0}}

35: \def\detp{{\rm det}'}

36: \def\fo{{\frak F}}  %% forest

37: \def\tree{{\frak T}}  %% tree

38: \def\wsf{{\ss WSF}}

39: \def\fsf{{\ss FSF}}

40: \def\UST{{\ss UST}}

41: \def\ust{{\ss UST}}

42: \def\bdv{\partial_{\verts}}

43: \def\cpe{\cp_{\rm e}}

44: \def\Aut{{\rm Aut}} % automorphism group

45: \def\grn{{\cal G}}  % Green gen function = return series

46: \def\rad{{\bf r}}   % radius of cnv

47: \def\cat{{\rm G}}  % Catalan's constant

48: \def\asym{{\bf h}}   % asymptotics

49: \def\hloc{h_{\rm loc}}  % local entropy

50: \def\ip#1{(\changecomma #1)}

51: \def\bigip#1{\bigl(\changecomma #1\bigr)}

52: \def\Bigip#1{\Bigl(\changecomma #1\Bigr)}

53: \def\biggip#1{\biggl(\bigchangecomma #1\biggr)}

54: \def\leftip#1{\left(\leftchangecomma #1\right)}

55: \def\changecomma#1,{#1,\,}

56: \def\bigchangecomma#1,{#1,\;}

57: \def\leftchangecomma#1,{#1,\ }

58: \def\G{{\cal G}}

59: \def\GG{{\cal G}}

60: \def\PGW{{\ss PGW}}

61: \def\AGW{{\ss AGW}}

62: \def\UGW{{\ss UGW}}

63: \def\sgn{{\rm sgn}}

64: \def\rtd{\mu}  % the measure on rooted graphs

65: \def\cd{\Rightarrow}  % convergence in distribution

66: \def\sfrac#1#2{{#1 \over #2}}

67: \def\frac#1#2{{#1 \over #2}}

68: \def\nroot{\bp}  % network root; \root is already in TeX

69: \def\eps{\epsilon}

70: \def\mathrm#1{{\rm #1}}

71: \def\ed{\buildrel \cal D \over =}  %% equal in distribution

72: \def\A{{\cal A}}

73: \def\Ah{\widehat{\cal A}}

74: \def\marks{{\Xi}}  % space of marks

75: \def\mk{\xi}  % a mark

76: \def\mkmp{\psi}  % mark map

77: \def\etail#1{#1^-}   % tail of an edge

78: \def\ehead#1{#1^+}   % head of an edge

79: \def\bbar#1{{\overline #1}}

80: \def\fdom{{\ss FC}}  % finite components

81: \def\Ph{{\widehat{\rtd}}}

82: \def\Php{{\widehat{\rtd'}}}

83: \def\ev#1{{\cal #1}} % Events

84: \def\traj{\theta}    % Markov chain trajectory measure

85: \def\invar{{\cal I}}  %% invariant events

86: \def\SS{{\cal S}}   %% shift

87: \def\rln{{\cal R}}   %% relation

88: \def\rlnmap{\phi}

89: \def\assumptions{Let\/ $\P$ be a percolation on a

90: unimodular random network}

91: \def\genassumptions{Let\/ $\P$ be an insertion-tolerant percolation on a

92: unimodular random network}

93: \def\ct{n}  % count edges leaving

94: \def\paths{{\cal P}_*}   % paths in a network

95: \def\invarp{{\invar_\infty}}  % invariant $\sigma$-field for paths

96: \def\fin{{\ss FG_2}}  % finite graphs with 2 distinguished vertices

97: \def\gtwo{{{\cal G}_{**}}}  % networks with 2 distinguished vertices

98: \def\uni{{\cal U}}  % unimodular probability measures on $\GG_*$

99: \def\projection{{\ss prj}}  % projection on the first coordinate

100: \def\dmn{\GG_\#}  % distinct-mark networks

101: \def\itimes{\boxtimes}  % independent product

102: \def\dist{{\rm dist}}  % distance

103: \def\bn{{\rm B}}  % Bernoulli

104: \def\alg{{\ss Alg}}  % algebra

105: \def\affalg{\overline{\ss Alg}}  % affiliated algebra

106: \def\two{\{0, 1\}}

107: \def\expdeg{\overline{\rm deg}}  % expected degree

108: 

109: \def\ins{\Pi}   %% Insertion

110: \def\isoe{\iota_\edge}

111: \def\bde{\bd_\edge}

112: \def\bd{\partial}

113: \def\pc{p_{\rm c}}

114: \def\pcb{p_{\rm c}^{\rm bond}}

115: \def\pu{p_{\rm u}}

116: \def\USF{{\ss USF}}

117: \def\UST{{\ss UST}}

118: \def\MST{{\ss MST}}

119: \def\WUSF{{\ss WUSF}}

120: \def\WMSF{{\ss WMSF}}

121: \def\FUSF{{\ss FUSF}}

122: \def\FMSF{{\ss FMSF}}

123: \def\verts{{\ss V}}

124: \def\vertex{{\ss V}}

125: \def\vertices{{\ss V}}

126: \def\edges{{\ss E}}

127: \def\edge{{\ss E}}

128: \def\Stab{{\ss Stab}}

129: \def\bp{o}

130: \def\isol{\heartsuit}  % isolated vertex (type)

131: \def\lift{\lambda}  % lift probability measure from class to network

132: \def\dual#1{#1^\dagger}

133: \def\collec{{\cal C}}  % collection

134: \def\K{{\cal K}}  % compact set

135: \def\gve{{\cal G}_{*-}}  % network with vertex and edge

136: \def\type{\tau}  % type of a vertex

137: \def\HH{{\cal H}}   % a Hilbert space

138: \def\DD{{\scr D}}   % domain

139: \def\dmn{{\scr D}}   % domain

140: \def\spm{\sigma}  % trace of spectral measure

141: 

142: \def\BLPSusf{\ref b.BLPS:usf/, hereinafter referred to as %

143: \def\BLPSusf{BLPS (\refbyear {BLPS:usf})}\BLPSusf}

144: 

145: \def\BLPSgip{\ref b.BLPS:gip/, hereinafter referred to as %

146: \def\BLPSgip{BLPS (\refbyear {BLPS:gip})}\BLPSgip}

147: 

148: \def\LPSmsf{\ref b.LPS:msf/}

149: 

150: \ifproofmode \relax \else\head{{\it Electron. J. Probab.}, {\bf 12}, Paper

151: 54 (2007), 1454--1508.}

152: {Version of \versiondate}\fi

153: \vglue20pt

154: 

155: \def\firstheader{\eightpoint\ss\underbar{\raise2pt\line 

156: {{\it Electron.\ J.\ Probab.}, {\bf 12}, Paper 54 (2007), 1454--1508. 

157:      \hfil Version of \versiondate}}}

158: 

159: \beginniceheadline

160: 

161: \title{Processes on Unimodular Random Networks}

162: 

163: \author{David Aldous and Russell Lyons}

164: 

165: 

166: %\centerline{\tt NOT FOR DISTRIBUTION}

167: 

168: 

169: 

170: \abstract{

171: We investigate unimodular random networks.

172: Our motivations include 

173: their characterization via reversibility of an associated random walk

174: and their similarities to unimodular quasi-transitive graphs.

175: We extend various theorems concerning random walks, percolation, spanning

176: forests, and amenability from the known context of unimodular

177: quasi-transitive graphs to the more general context of unimodular random

178: networks.

179: We give properties of a trace associated to unimodular random networks

180: with applications to stochastic comparison of

181: continuous-time random walk.

182: }

183: 

184: \bottomII{Primary

185: 60C05. % Combinatorial probability

186: Secondary

187: 60K99,  % special processes

188:  05C80. %   Random graphs

189: }

190: {Amenability, equivalence relations, infinite graphs, percolation,

191: quasi-transitive, random walks, transitivity, weak convergence,

192: reversibility, trace, stochastic comparison, spanning forests,

193: sofic groups.}

194: {Research partially supported by NSF grants DMS-0203062,

195: DMS-0231224, DMS-0406017, and DMS-0705518.}

196: 

197: \articletoc

198: 

199: \bsection{Introduction}{s.intro}

200: 

201: In the setting of infinite discrete graphs,

202: the property of being a Cayley graph of a group is a strong form of 

203: ``spatial homogeneity":

204: many results not true for arbitrary graphs are true under this strong property.

205: As we shall soon explain, weaker regularity properties sufficient for many

206: results have been studied.  

207: In this paper, we turn to random graphs, investigating

208: a notion of

209: ``statistical homogeneity" or ``spatial stationarity"

210: that we call

211: a {\it unimodular random rooted network}.

212: The root is merely a distinguished vertex of the network and the

213: probability measure is on a certain space of rooted networks.

214: In a precise sense, the root is ``equally likely" to be any vertex of the

215: network, even though we consider infinite networks.

216: We shall show that many results known for

217: deterministic graphs under previously-studied regularity conditions 

218: do indeed extend to

219: unimodular random rooted networks.

220: 

221: Thus, a probabilistic motivation for our investigations is the study of

222: stochastic processes under unimodularity.

223: A second motivation is combinatorial: One often asks for asymptotics of

224: enumeration or optimization problems on finite networks as the size of the

225: networks tend to infinity.

226: One can sometimes answer such questions with the aid of a suitable limiting

227: infinite object. A survey of this approach is given by \ref b.AS:obj/.

228: We call ``random weak limit" the type of limit one considers; it is the

229: limiting ``view" from a uniformly chosen vertex of the finite networks.

230: What limiting objects can arise this way?

231: It has been observed before that

232: the probabilistic objects of interest,

233: unimodular random rooted networks, contain all

234: the combinatorial objects of interest, random

235: weak limits of finite networks.

236: One open question is whether these two classes in fact coincide.

237: An affirmative answer would have many powerful consequences, as we shall

238: explain.

239: 

240: To motivate this by

241: analogy, recall a simple fact about stationary sequences

242: $\Seq{Y_i}_{i \in \Z}$

243: of random variables.

244: For each $n \geq 1$, let 

245: $\Seq{Y_{n,i}}_{1 \leq i \leq n}$

246: be arbitrary. Center it at a uniform index 

247: $U_n \in \{1,2,\ldots,n\}$ 

248: to get a bi-infinite sequence

249: $\Seq{Y_{n,U_n+i}}_{i \in \Z}$,

250: interpreted arbitrarily outside its natural range.  

251: If there is a weak limit 

252: $\Seq{Y_i}_{i \in \Z}$ as $n \to \infty$ 

253: of these randomly centered sequences, then the limit is

254: stationary, and conversely any stationary sequence can be obtained

255: trivially as such a limit.

256: 

257: %Though the intrinsic definition

258: %(\ref s.notation/)

259: %is somewhat abstract, its motivation is easy to describe.

260: By analogy, then,

261: given a finite graph, take a uniform random vertex as root.

262: Such a randomly rooted graph automatically has a certain property

263: %(briefly, an associated measure on graphs with a distinguished oriented edge

264: %is invariant under edge-reversal)

265: (in short, if mass is redistributed in the graph, then the expected mass

266: that leaves the root is equal to the expected mass the arrives at the root)

267: and in \ref s.notation/,

268: we abstract this property as unimodularity.

269: It is then immediate that any infinite random rooted graph that is a limit

270: (in an appropriate sense that we call ``random weak limit")

271: of uniformly randomly rooted finite graphs will be unimodular, whereas

272: the above question asks

273: conversely whether any unimodular random rooted graph arises as a random

274: weak limit of some sequence of randomly rooted finite graphs.

275: 

276: Additional motivation for the definition arises from random walk considerations.

277: Given any random rooted graph, simple random walk induces a Markov chain on

278: rooted graphs.

279: Unimodularity of a probability measure $\rtd$ on rooted graphs

280: is equivalent to the property that a reversible stationary

281: distribution for this chain is given by the root-degree biasing of $\rtd$,

282: just as on finite graphs,

283: a stationary distribution for simple random walk is proportional to the

284: vertex degrees; see \ref s.extreme/.

285: 

286: Let us return now to the case of deterministic graphs.

287: An apparently minor relaxation of the Cayley graph property

288: is the ``transitive" property 

289: (that there is an automorphism taking any vertex to any other vertex).

290: By analogy with the shift-invariant interpretation of stationary sequences, one might expect every

291: transitive graph to fit into our set-up.

292: But this is false.

293: Substantial research over the last ten years has shown that

294: the most useful regularity condition is that of a

295: {\it unimodular transitive graph} (or, more generally, quasi-transitive).

296: %It turns out that this lets us make precise in a powerful way 

297: %the intuition that 

298: Intuitively, this is

299: an unrooted transitive graph that can be given 

300: a random root in such a way that each vertex is equally

301: likely to be the root.

302: This notion is, of course, precise in itself for a finite graph.

303: To understand how this is extended to infinite graphs, and then to

304: unimodular random rooted graphs, consider a finite graph $\gh$ and a

305: function $f(x, y)$ of ordered pairs of vertices of $\gh$.

306: Think of $f(x, y)$ as an amount of mass that is sent from $x$ to $y$.

307: Then the total mass on the graph $\gh$ before transport equals the total

308: after, since mass is merely redistributed on the graph.

309: We shall view this alternatively as saying that for a randomly

310: uniformly chosen vertex, the expected mass it receives is equal to the

311: expected mass it sends out.

312: This, of course, depends crucially on choosing the vertex uniformly and,

313: indeed, characterizes the uniform measure among all probability measures on

314: the vertices.

315: 

316: Consider now an infinite transitive graph, $\gh$.

317: Since all vertices ``look the same", we could just fix one, $\bp$,

318: rather than try to choose one uniformly.

319: However, a mass transport function $f$ will not conserve the mass at $\bp$

320: without some assumption on $f$ to make up for the fact that $\bp$ is fixed.

321: Although it seems special at first, it turns out that a very useful

322: assumption is that $f$ is invariant under the diagonal action of

323: the automorphism group of $\gh$.

324: (For a finite graph that happened to have no automorphisms other than the

325: identity, this would be no restriction at all.)

326: This is still not enough to guarantee ``conservation of mass", i.e., that

327: $$

328: \sum_x f(\bp, x) = \sum_x f(x, \bp)

329: \,,

330: \label e.mtp1

331: $$

332: but it turns out that \ref e.mtp1/ does hold when the

333: automorphism group of $\gh$ is unimodular. Here, ``unimodular" is used in

334: its original sense that the group

335: admits a non-trivial Borel measure that is invariant under both left and

336: right multiplication by group elements.

337: We call $\gh$ itself unimodular in that case; see Sections

338: \briefref s.notation/ and \briefref s.related/ for more

339: on this concept.

340: The statement that \ref e.mtp1/ holds under these assumptions is called the

341: Mass-Transport Principle for $\gh$.

342: If $\gh$ is quasi-transitive, rather than transitive, we still have a

343: version of \ref e.mtp1/, but we can no longer consider only one fixed

344: vertex $\bp$.

345: Instead, each orbit of the automorphism group must have a representative

346: vertex.

347: Furthermore, it must be weighted ``proportionally to its frequency" among

348: vertices; see \ref t.fixedgraph/. 

349: This principle was introduced to the study of percolation by \ref

350: b.Hag:deptree/, then developed and exploited heavily by \BLPSgip.

351: Another way of stating it is that \ref e.mtp1/ holds in expectation 

352: when $\bp$ is chosen randomly by an appropriate probability measure.

353: If we think of $\bp$ as the root, then we arrive at the notion of

354: random rooted graphs, and the corresponding statement that \ref e.mtp1/

355: holds in expectation is a general form of the Mass-Transport Principle.

356: This general form was called the ``Intrinsic Mass-Transport Principle" by

357: \ref b.BS:rdl/.

358: %It is, in fact, equivalent to the corresponding probability measure being

359: %unimodular by our definition.

360: We shall call a probability measure on rooted graphs unimodular precisely

361: when this general form of the Mass-Transport Principle holds.

362: We develop this in \ref s.notation/.

363: 

364: Thus, we can extend many results known

365: for unimodular quasi-transitive graphs to our new setting of

366: unimodular random rooted graphs, as noted by \ref b.BS:rdl/.

367: As a bonus, our set-up allows the treatment of

368: quasi-transitive graphs to be precisely parallel to that of

369: transitive graphs, with no additional notation or thought needed, which had

370: not always been the case previously.

371: 

372: 

373: 

374: To state results in their natural generality, as well as for technical

375: convenience, we shall work in the setting of networks,

376: which are just graphs with ``marks" (labels) on edges and vertices.

377: Mainly, this paper is organized to progress from the most general to the

378: most specific models.

379: An exception is made in \ref s.related/, where we discuss random networks

380: on fixed underlying graphs.

381: This will not only help to understand and motivate the general setting, but

382: also will be useful in deriving consequences of our general results.

383: 

384: \ref s.extreme/ elaborates the comment above about reversible stationary

385: distributions for random walk,  discussing extremality and invariant

386: $\sigma$-fields, speed of random walk, and continuous-time random walk and

387: their explosions.

388: \ref s.trace/ discusses a trace associated to unimodular random networks

389: and comparison of return probabilities of different continuous-time random

390: walks, which partially answers a question of Fontes and Mathieu.

391: We then write out the extensions to unimodular random rooted graphs of results

392: known for fixed graphs in the context of

393: percolation (\ref s.extend/),

394: spanning forests (\ref s.span/)

395: and amenability (\ref s.amen/).

396: These extensions are in most (though not all) cases straightforward.

397: Nevertheless, we think it is useful to list these extensions in the order

398: they need to be proved so that others need not check the entire (sometimes

399: long) proofs or chains of theorems from a variety of papers.

400: Furthermore, we were required to find several essentially new results along

401: the way.

402: %Moreover, occasionally our approach leads to a simpler and shorter proof

403: %than the original, albeit one based on our fundamental finite-approximation

404: %result, \ref t.TII/.

405: 

406: In order to appreciate the scope of our results, we list many examples of

407: unimodular probability measures in \ref s.ex/.

408: In particular, there is a significant and important overlap between our

409: theory and the theory of graphings of measure-preserving equivalence

410: relations. This overlap

411: is well known among a few specialists, but

412: deserves to be made more explicit. We do that here in \ref x.equivalence/.

413: 

414: Among several open problems, we spotlight a special case of

415: \ref q.coupling/: Suppose we

416: are given a partial order on the mark space and two unimodular probability

417: measures, one stochastically dominating the other. That is, there is

418: a monotone coupling of the two unimodular distributions that puts the

419: networks on the same graphs, but has higher marks for the second network

420: than for the first. Does this

421: imply the existence of a {\it unimodular\/} monotone coupling?

422: A positive answer would be of great benefit in a variety of ways.

423: 

424: Another especially important open question is \ref q.TII/, whether every

425: unimodular probability measure is a limit of uniformly rooted finite

426: networks.

427: For example, in the case that the random rooted infinite network is just a

428: Cayley graph (rooted, say, at the identity) with the edges marked by the

429: generators, a positive answer to this question on finite approximation would

430: answer a question of \ref b.Weiss:sofic/, by showing that all finitely

431: generated groups are ``sofic", although this is contrary to the belief

432: expressed by \ref b.Weiss:sofic/.

433: (Sofic groups were introduced, with a different definition, by \ref

434: b.Gromov:endo/; see \ref b.Elek:sofic/ for a proof that the definitions are

435: equivalent.)

436: This would establish several conjectures, since they are known to hold for

437: sofic groups: the direct finiteness

438: conjecture of \ref b.Kaplansky/ on group algebras

439: (see \ref b.Elek:sofic/), a

440: conjecture of \ref b.Gottschalk/ on ``surjunctive" groups in topological

441: dynamics

442: (see \ref b.Gromov:endo/), the Determinant

443: Conjecture on Fuglede-Kadison determinants

444: (see \ref b.Elek:hyper/), and Connes's (\refbyear{Connes})

445: Embedding Conjecture

446: for group von Neumann algebras (see \ref b.Elek:hyper/).

447: %The latter conjecture is referred to by saying that all countable groups are

448: %``hyperlinear".

449: %Since our work combined with that of \ref b.Elek:hyper/ shows that this is

450: %true, we ``refute the famous `theorem' of Gromov claiming that a

451: %proposition which holds for all countable discrete groups is either trivial

452: %or false", in the words of \ref b.Ozawa/.

453: The Determinant Conjecture in turn implies the Approximation Conjecture of

454: \ref b.Schick/ and the Conjecture of Homotopy Invariance of $L^2$-Torsion

455: due to \ref b.Luck:3man/; see Chap.~13 of \ref b.Luck:book/ for these

456: implications and more information.

457: \ref b.Weiss:sofic/ gave another proof

458: of Gottschalk's conjecture for sofic groups.

459: One may easily extend that proof to show a form of Gottschalk's conjecture 

460: for all quasi-transitive unimodular

461: graphs that are limits of finite graphs, but there are easy

462: counterexamples for general transitive graphs.

463: 

464: Further discussion of the

465: question on approximation by finite networks is given in

466: \ref s.finite/.

467: A positive answer would

468: provide solid support for the intuition that the root of a

469: unimodular random rooted network is equally likely to be any vertex.

470: \ref s.finite/ also contains some variations that would result from a

471: positive answer and some

472: additional consequences for deterministic graphs.

473: 

474: The notion of weak convergence of rooted locally finite graphs or networks  

475: (needed to make sense of convergence of randomly rooted

476: finite graphs to a limit infinite graph)

477: has arisen before in several different contexts.

478: Of course, the special case where the limit network is a Cayley diagram

479: was introduced by \ref b.Gromov:endo/ and \ref b.Weiss:sofic/.

480: In the other cases, the limits provide examples of unimodular

481: random rooted graphs.

482: \ref b.Aldous:fringe/

483: gives many examples of models of random finite trees

484: which have an infinite-tree limit 

485: (and one such example, the limit of uniform random labeled trees

486: being what is now called the 

487: Poisson-Galton-Watson tree,

488: ${\PGW}^\infty(1)$, goes back to 

489: \ref b.Grimmett:ltrees/).

490: The idea that random weak limits of finite planar graphs of

491: uniformly bounded degree provide an interesting class of infinite planar

492: graphs was developed by \ref b.BS:rdl/, who showed that random walk on

493: almost any such limit graph is recurrent.

494: (Thus, such graphs do not include regular trees or hyperbolic graphs,

495: other than trivial examples like $\Z$.)

496: A specialization to random weak limits of

497: plane triangulations was studied in more detail in

498: interesting recent work of \ref b.AngelSchramm/ and \ref b.Angel/.

499: %Another type of approximation is due to \ref b.VershikGordon/.

500: 

501: \ref x.PWIT/ describes an infinite-degree tree, arising as a limit

502: of weighted finite complete graphs. 

503: This example provides an interface between our setting and 

504: related ideas of 

505: ``local weak convergence"

506: and ``the objective method in the probabilistic analysis of algorithms".

507: A prototype is that the distribution of $n$ random points in a square of area $n$

508: converges in a natural sense as $n \to \infty$ to the distribution

509: of a Poisson point process on the plane of unit intensity.

510: One can ask whether solutions of combinatorial optimization

511: problems over the $n$ random points 

512: (minimum spanning tree, minimum matching, traveling salesman problem)

513: converge to limits that are the solutions of analogous

514: optimization problems over the Poisson point process in the whole plane.

515: \ref x.PWIT/ can be regarded as a mean-field analogue of random

516: points in the plane, and $n \to \infty$ limits of solutions of combinatorial optimization problems

517: within this model have been studied using the non-rigorous

518: cavity method from statistical physics.  

519: \ref b.APercus:SU/ illustrate what can be done by non-rigorous means,

520: while \ref b.AS:obj/ survey introductory rigorous theory. 

521: 

522: 

523: The reader may find it helpful to keep in mind one additional example,

524: a unimodular version of family trees of Galton-Watson branching

525: processes; see also \ref x.fdd/.

526: 

527: \procl x.AGW \procname{Unimodular Galton-Watson}

528: Let $\Seq{p_k \st k \ge 0}$ be a probability distribution on $\N$.

529: Take two independent Galton-Watson trees with offspring distribution

530: $\Seq{p_k}$, each starting with one particle, the root,

531: and join them by a new edge

532: whose endpoints are their roots. Root the new tree

533: at the root of the first Galton-Watson

534: tree. This is {\bf augmented Galton-Watson}

535: measure, \AGW. (If $p_0 \ne 0$, then we have the

536: additional options to condition on either

537: non-extinction or extinction of the joined trees.) Now bias by the

538: reciprocal of the degree of the root to get {\bf unimodular Galton-Watson}

539: measure, $\UGW$.

540: In different language, \ref b.LPP:GW/ proved that this measure,

541: $\UGW$, is unimodular.

542: Note that

543: the mean degree of the root is

544: $$

545: \expdeg(\UGW) = \left(\sum_{k \ge 0} {p_k \over k+1} \right)^{-1}

546: \,.

547: \label e.degAGW

548: $$

549: %We shall revisit this example in the contexts of limits of finite graphs

550: %and of random walks.

551: \endprocl

552: 

553: 

554: 

555: 

556: 

557: \bsection{Definitions and Basics}{s.notation}

558: 

559: We denote a (multi-)graph $\gh$ with vertex set $\verts$ and undirected

560: edge set $\edges$ by $\gh = (\verts, \edges)$.

561: When there is more than one graph under discussion, we write $\verts(\gh)$ or

562: $\edges(\gh)$ to avoid ambiguity.

563: We denote the degree of a vertex $x$ in a graph $\gh$ by $\deg_\gh(x)$.

564: {\bf Simple random walk} on $\gh$ is the Markov chain whose state space is

565: $\verts$ and whose transition probability from $x$ to $y$ equals the number of

566: edges joining $x$ to $y$ divided by $\deg_\gh(x)$.

567: 

568: A {\bf network} is a (multi-)graph $\gh = (\vertex, \edge)$ together with a

569: complete separable metric space $\marks$ called the {\bf mark space} and

570: maps from $\vertex$ and $\edge$ to $\marks$. Images in $\marks$ are called

571: {\bf marks}.

572: Each edge is given two marks, one associated to (``at") each of its endpoints.

573: The only assumption on degrees is that they are finite.

574: We shall usually assume that $\marks$ is Baire space $\N^\N$, since every

575: uncountable complete separable metric space is Borel isomorphic to Baire

576: space by Kuratowski's theorem (Theorem 15.10 of \ref b.Royden/).

577: We generally omit mention of the mark maps from our notation for

578: networks when we do not need them.

579: For convenience, we consider graphs as special cases of networks in which

580: all marks are equal to some fixed mark.

581: 

582: We now define ends in graphs.

583: In the special case of a tree, an infinite path that starts at any vertex

584: and does not backtrack is called a {\bf ray}. Two rays are {\bf equivalent}

585: if they have

586: infinitely many vertices in common. An equivalence class of rays is

587: called an {\bf end}.

588: In a general infinite graph, $\gh$, an {\bf end} of $\gh$ is an equivalence

589: class of infinite simple paths in $\gh$, where two paths

590: are equivalent if for every finite $K\subset \vertex(\gh)$,

591: there is a connected component of $\gh\setminus K$ that intersects

592: both paths.

593: 

594: Let $\gh$ be a graph.  For a subgraph $H$, let its {\bf (internal) vertex

595: boundary} $\bdv H$ be the set of vertices of $H$ that are adjacent to some

596: vertex not in $H$.

597: We say that $\gh$ is {\bf (vertex) amenable}

598: if there exists a sequence of subsets $H_n\subset \vertex(\gh)$

599: with

600: $$

601: \lim_{n \to\infty} {|\bdv H_n| \over |\verts(H_n)|} = 0\,,

602: $$

603: where $|\cbuldot|$ denotes cardinality.

604: Such a sequence is called a {\bf F\o{}lner sequence}.

605: A finitely generated group is {\bf amenable} if its Cayley graph is amenable.

606: For example, every finitely generated abelian group is amenable.

607: For more on amenability of graphs and groups, see \BLPSgip.

608: 

609: A {\bf homomorphism} $\varphi: \gh_1 \to \gh_2$ from one

610: graph $\gh_1=(\verts_1,\edges_1)$ to another $\gh_2=(\verts_2,\edges_2)$

611: is a pair of maps $\varphi_\verts:\verts_1\to\verts_2$ and

612: $\varphi_\edges:\edges_1\to\edges_2$

613: such that $\varphi_\verts$ maps the endpoints of $e$ to the endpoints

614: of $\varphi_\edges(e)$ for every edge $e \in \edges_1$.

615: When both maps $\varphi_\verts:\verts_1\to\verts_2$ and

616: $\varphi_\edges:\edges_1\to\edges_2$ are bijections, then $\varphi$ is

617: called an {\bf isomorphism}.

618: When $\gh_1 = \gh_2$, an isomorphism is called

619: an {\bf automorphism}.

620: The set of all automorphisms of $\gh$ forms a group under composition,

621: denoted by $\Aut(\gh)$.

622: The action of a group $\gp$ on a graph $\gh$ by automorphisms is said to be {\bf

623: transitive} if there is only one $\gp$-orbit in $\verts(\gh)$ and to be {\bf

624: quasi-transitive} if there are only finitely many orbits in $\verts(\gh)$.

625: A graph $\gh$ is {\bf transitive} or {\bf quasi-transitive} according as

626: whether the corresponding action of $\Aut(\gh)$ is.

627: %if for every $x,y\in\verts$, there is an

628: %automorphism of $\gh$ taking $x$ to $y$.

629: %In other words, the orbit

630: %space $\verts/\Aut(\gh)$ of $\verts$ under the group $\Aut(\gh)$ of

631: %automorphisms of $\gh$ is a singleton.

632: For example, every Cayley graph is transitive.

633: %A graph is {\bf quasi-transitive} if

634: %the orbit space $\verts/\Aut(\gh)$ is finite.

635: All the same terms are applied to networks when the maps in question preserve

636: the marks on vertices and edges.

637: 

638: A locally compact group is called {\bf unimodular} if its

639: left Haar measure is also right invariant.

640: In particular, every discrete countable group is unimodular.

641: We call a graph $\gh$ {\bf unimodular} if $\Aut(\gh)$ is unimodular, where

642: $\Aut(\gh)$ is given the weak topology generated by its action on $\gh$.

643: Every Cayley graph and, as \ref b.SoardiWoess/ and \ref b.Salvatori/ proved,

644: every quasi-transitive amenable graph is unimodular.

645: See \ref s.related/ and \BLPSgip\ for more details on unimodular graphs.

646: 

647: A {\bf rooted network} $(\gh, \bp)$ is a network $\gh$ with a distinguished

648: vertex $\bp$ of $\gh$, called the {\bf root}.

649: A {\bf rooted isomorphism} of rooted networks is an isomorphism of the

650: underlying networks that takes the root of one to the root of the other.

651: We generally do not distinguish between a rooted network and its

652: isomorphism class.

653: When needed, however, we use the following notation to make these distinctions:

654: $\gh$ will denote a graph, $\overline \gh$ will denote a network with underlying

655: graph $\gh$, and $[\overline \gh, \bp]$ will denote the class of rooted networks

656: that are rooted-isomorphic to $(\overline \gh, \bp)$.

657: We shall use the following notion introduced (in slightly different

658: generalities) by \ref b.BS:rdl/ and \ref b.AS:obj/.

659: %Given a positive integer $R$, a finite rooted network $H$, and a probability

660: %distribution $\rtd$ on rooted networks with a fixed mark space, let $p(R, H,

661: %\rtd)$ denote the probability that $H$ is rooted isomorphic to the ball of

662: %radius $R$ about the root of a network chosen with distribution $\rtd$.

663: %A sequence $\rtd_n$ is said to {\bf converge weakly} to $\rtd$, written $\rtd_n

664: %\cd \rtd$, if $\lim_{n \to\infty} p(R, H, \rtd_n) = p(R, H, \rtd)$

665: %for any positive integer $R$ and any finite network $H$.

666: %

667: %This agrees with the usual sense of weak convergence of probability

668: %measures on a metric space as follows.

669: Let $\GG_*$ denote the set of rooted isomorphism classes of rooted

670: {\it connected\/} locally finite networks.

671: Define a metric on $\GG_*$ by letting the distance between $(G_1, o_1)$ and

672: $(G_2, o_2)$ be $1/(1+\alpha)$, where $\alpha$ is

673: the supremum of those $r > 0$ such that

674: %(if we take

675: %$(G_i, o_i)$ to be representatives of their isomorphism classes) there is a

676: %network $(G, o)$ whose underlying graph is

677: %rooted isomorphic to subgraphs of the graphs

678: %underlying both $(G_i, o_i)$ with the properties

679: %\smallskip

680: %\item{(i)} every vertex of finite degree within distance $1/\epsilon$ of

681: %$o_i$ belongs to the image of $G$;

682: %\item{(ii)} every vertex of infinite degree within distance $1/\epsilon$ of

683: %$o_i$ that belongs to the image of $G$ has degree at least $1/\epsilon$

684: %in $G$; and

685: %\item{(iii)} every mark in $G$ is within distance $\epsilon$ of the

686: %corresponding marks in $G_i$.

687: %\smallskip

688: %\noindent

689: there is some rooted isomorphism of

690: the balls of (graph-distance) radius $\flr{r}$ around the roots of $G_i$ 

691: such that each pair of corresponding marks has distance less than $1/r$.

692: It is clear that $\GG_*$ is separable and complete in this metric.

693: For probability measures $\rtd$, $\rtd_n$ on $\GG_*$, we write $\rtd_n \cd

694: \rtd$ when $\rtd_n$ converges weakly with respect to this metric.

695: 

696: For a probability measure $\rtd$ on rooted networks, write $\expdeg(\rtd)$

697: for the {\it expectation\/} of the degree of the root with respect to

698: $\rtd$.

699: In the theory of measured equivalence relations (\ref x.equivalence/),

700: this is twice the {\bf

701: cost} of the graphing associated to $\rtd$.

702: Also, by the {\bf degree} of $\rtd$ we mean the {\it distribution\/} of

703: the degree of the root under $\rtd$.

704: 

705: For a locally finite connected rooted network, there is a canonical choice

706: of a rooted network in its rooted-isomorphism class.

707: More specifically, there is a continuous map $f$ from $\GG_*$ to the space of

708: networks on $\N$ rooted at 0 such that $f\big([\gh, \bp]\big) \in [\gh,

709: \bp]$ for all $[\gh, \bp] \in \GG_*$.

710: To specify this, consider the following total ordering on rooted networks

711: with vertex set $\N$ and root $0$.

712: First, total order $\N \times \N$ by the lexicographic order:

713: $(i_1, j_1) \prec (i_2, j_2)$ if

714: either $i_1 < i_2$ or $i_1 = i_2$ and $j_1 < j_2$.

715: Second, the lexicographic order on Baire space $\marks$ is also a total order.

716: We consider networks on $\N$ rooted at $0$.

717: Define a total order on such networks as follows.

718: Regard the edges as oriented for purposes of identifying the edges with $\N

719: \times \N$; the mark at $i$ of an edge between $i$ and $j$ will be considered

720: as the mark of the oriented edge $(i, j)$.

721: Suppose we are given a pair of networks on $\N$ rooted at $0$.

722: If they do not have the same edge sets, then the network that contains the

723: smallest edge in their symmetric difference is deemed to be the smaller

724: network.

725: If they do have the same edge sets, but not all the vertex marks are the same,

726: then the network that contains the vertex with the smaller mark on the

727: least vertex where they differ is deemed the smaller network.

728: If the networks have the same edge sets and the same vertex marks, but not all

729: the edge marks are the same, then

730: the network that contains the oriented edge with the smaller mark on the

731: least oriented edge where they differ is deemed the smaller network.

732: Otherwise, the networks are identical.

733: 

734: We claim that the rooted-isomorphism class of each locally finite connected

735: network contains a unique smallest rooted network on $\N$ in the above

736: ordering.

737: This is its {\bf canonical} representative.

738: To prove our claim, consider only the networks in the class that have

739: vertex set $\N$ and are rooted at $0$.

740: It is fairly easy to see that there is a smallest graph that supports a

741: network in the class:

742: the neighbors of 0 are $[1, n_1]$ for some $n_1 \ge 1$,

743: the neighbors of 1, besides 0, are $[n_1+ 1, n_2]$ for some $n_2 \ge n_1$, the

744: neighbors of 2, other than possibly 0 and 1, are $[n_2+1, n_3]$ for some

745: $n_3\ge n_2$, etc.

746: In general, the neighbors of $k \ge 1$ that are larger than $k$ are $[n_k+1,

747: n_{k+1}]$ for some $n_{k+1} \ge n_k$.

748: Write $n_0 := 0$.

749: Let $A_{-1}$ be the set of networks in the class that are on this smallest

750: graph.

751: All have the same mark at $0$.

752: Write $\mkmp(k)$ for the mark at a vertex $k$ and $\mkmp(j, k)$ for the

753: mark at the oriented edge $(j, k)$.

754: Define $A_k$ recursively as follows.

755: Given $A_{k-1}$, let $A_k$ be the subset of networks in $A_{k-1}$ such that

756: the marks $\mkmp(j)$ are increasing for $j \in [n_k+1, n_{k+1}]$.

757: Then let $B_{-1} := \bigcap_k A_k$.

758: %If the graph has no other vertices, then $A_0$ is a singleton and the

759: %canonical representative.

760: %Suppose there are other vertices, so that $n_1$ defined above is at least 1.

761: %Let $A_1$ be the subset of $A_0$ where the marks $\mkmp(k)$ are increasing

762: %in $k \in [1, n_1]$.

763: This is non-empty.

764: Now define $B_k$ recursively by letting $B_k$ be the subset of $B_{k-1}$

765: such that $j \mapsto \mkmp(k, j)$ is increasing on $[n_k+1, n_{k+1}]$.

766: The set $\bigcap_k B_k$ is then a singleton, the desired canonical

767: representative.

768: 

769: 

770: For a (possibly disconnected)

771: network $\gh$ and a vertex $x \in \verts(\gh)$, write $\gh_x$ for the

772: connected component of $x$ in $\gh$.

773: If $\gh$ is a network with probability distribution $\rtd$ on its

774: vertices, then $\rtd$ induces naturally a distribution on $\GG_*$,

775: which we also denote by $\rtd$; namely, the probability of $(\gh_x, x)$ is

776: $\rtd(x)$.

777: More precisely, $\rtd\big([\gh_x, x]\big) := \sum \big\{ \rtd(y) \st y \in

778: \verts(\gh), \ (\gh_y, y) \in [\gh_x, x] \big\}$.

779: For a finite network $\gh$, let $U(\gh)$ denote the distribution on $\GG_*$

780: obtained this way by choosing a uniform random vertex of $\gh$ as root.

781: Suppose that $\gh_n$ are finite networks and that $\rtd$ is a

782: probability measure on $\GG_*$.

783: We say the {\bf random weak limit} of $\gh_n$ is $\rtd$ if $U(\gh_n) \cd \rtd$.

784: %Note that only the component of the root matters for convergence to $\rtd$.

785: If $\rtd\Big(\big\{[\gh, \bp]\big\}\Big) = 1$ for a fixed transitive

786: network $\gh$ (and (any) $\bp \in \vertex(\gh)$), then we say that the {\bf

787: random weak limit} of $\gh_n$ is $\gh$.

788: 

789: As usual, call a collection $\collec$

790: of probability measures on $\GG_*$

791: {\bf tight} if for each $\epsilon > 0$, there is a

792: compact set $\K \subset \GG_*$ such that $\rtd(\K) > 1 - \epsilon$ for all

793: $\rtd \in \collec$.

794: Because $\GG_*$ is complete, any tight collection

795: has a subsequence that possesses a weak limit.

796: %If $\rtd_1$ and $\rtd_2$ are probability measures on $\verts(\gh)$ such that for

797: %any positive integer $R$ and any finite graph $H$, we have $p(R, H, \rtd_1) =

798: %p(R, H, \rtd_2)$, then we write $\rtd_1 \equiv \rtd_2$.

799: %If $\gh$ is a regular graph, write $d_\gh$ for its degree.

800: 

801: 

802: 

803: The class of probability measures

804: $\rtd$ that arise as random weak limits of finite networks is contained in

805: the class of unimodular $\rtd$, which we now define.

806: Similarly to the space $\GG_*$, we define the space $\gtwo$ of isomorphism

807: classes of locally

808: finite connected networks with an ordered pair of distinguished vertices

809: and the natural topology thereon.

810: We shall write a function $f$ on $\gtwo$ as $f(\gh, x, y)$.

811: 

812: \procl d.unimodular 

813: Let $\rtd$ be a probability measure on $\GG_*$.

814: We call $\rtd$ {\bf unimodular} if it obeys the {\bf Mass-Transport

815: Principle}:

816: For all Borel

817: %universally measurable\ftnote{*}{This means ``measurable with

818: %respect to every Borel measure".}

819: $f : \gtwo \to [0, \infty]$,

820: %and invariant in the sense that for any (non-rooted)

821: %network isomorphism $\gpe$ of $\gh$ and any $x, y \in \vertex(\gh)$, we have

822: %$f(\gpe \gh, \gpe x, \gpe y) = f(\gh, x, y)$.

823: we have

824: $$

825: \int \sum_{x \in \vertex(\gh)} f(\gh, \bp, x) \,d\rtd\big([\gh, \bp]\big)

826: =

827: \int \sum_{x \in \vertex(\gh)} f(\gh, x, \bp) \,d\rtd\big([\gh, \bp]\big)

828: \,.

829: \label e.mtpgen

830: $$

831: Let $\uni$ denote the set of unimodular Borel probability measures on $\GG_*$.

832: \endprocl

833: 

834: Note that to define the sums that occur here, we choose a specific network

835: from its rooted-isomorphism class, but which one we choose makes no

836: difference when the sums are computed.

837: We sometimes call $f(\gh, x, y)$ the amount of ``mass" sent from $x$ to $y$.

838: The motivation 

839: for the name ``unimodular" is two fold: One is the extension of the concept

840: of unimodular automorphism groups of networks. The second is

841: that the Mass-Transport

842: Principle expresses the equality of two measures on $\gtwo$

843: associated to $\rtd$, the ``left" measure $\rtd_{\rm L}$ defined by

844: $$

845: \int_\gtwo f \,d\rtd_{\rm L}

846: :=

847: \int_{\GG_*} \sum_{x \in \vertex(\gh)} f(\gh, \bp, x) \,d\rtd\big([\gh,

848: \bp]\big)

849: $$ 

850: and the ``right" measure $\rtd_{\rm R}$ defined by

851: $$

852: \int_\gtwo f \,d\rtd_{\rm R}

853: :=

854: \int_{\GG_*} \sum_{x \in \vertex(\gh)} f(\gh, x, \bp) \,d\rtd\big([\gh,

855: \bp]\big)

856: \,.

857: $$

858: Thus, $\rtd$ is unimodular iff $\rtd_{\rm L} = \rtd_{\rm R}$, which can

859: also be expressed by saying that the left measure is absolutely continuous

860: with respect to the right measure and has Radon-Nikod\'ym derivative 1.

861: 

862: It is easy to see that any $\rtd$ that is a random weak limit of finite

863: networks is unimodular, as observed by \ref b.BS:rdl/, who introduced this

864: general form of the Mass-Transport Principle under the name 

865: ``intrinsic Mass-Transport Principle".

866: The converse is open.

867: 

868: A special form of the Mass-Transport Principle was considered, in different

869: language, by \ref b.AS:obj/.

870: Namely, they defined $\rtd$ to be {\bf involution invariant} if \ref

871: e.mtpgen/ holds for those $f$ supported on $(\gh, x, y)$ with $x \sim y$.

872: In fact, the Mass-Transport Principle holds for general $f$ if it holds for

873: these special $f$:

874: 

875: \procl p.invinv

876: A measure is involution invariant iff it is unimodular.

877: \endprocl

878: 

879: \proof

880: Let $\rtd$ be involution invariant.

881: The idea is to send the mass from $x$ to $y$ by single steps, equally

882: spread among the shortest paths from $x$ to $y$.

883: For the proof, we may assume that $f(\gh, x, y) = 0$ unless $x$ and $y$ are

884: at a fixed distance, say $k$, from each other, since any $f$ is a sum of

885: such $f$.

886: Now write $L(\gh, x, y)$ for the set of paths of length $k$ from $x$ to $y$.

887: Let $n_j(\gh, x, y; z, w)$ be the number of paths in $L(\gh, x, y)$ such

888: that the $j$th edge goes from $z$ to $w$.

889: Define $f_j(\gh, z, w)$ for $1 \le j \le k$ and $z, w \in \verts(\gh)$

890: by

891: $$

892: f_j(\gh, z, w) :=

893: \sum_{x, y \in \verts(\gh)} {f(\gh, x, y) n_j(\gh, x, y; z, w) \over

894: |L(\gh, x, y)|}

895: \,.

896: $$

897: Then $f_j(\gh, z, w) = 0$ unless $z \sim w$. Furthermore,

898: $f_j(\gh, z, w) := f_j(\gh', z', w')$ if

899: $(\gh, z, w)$ is isomorphic to $(\gh', z', w')$.

900: Thus, $f_j$ is well defined and Borel on $\gtwo$,

901: whence involution invariance gives us

902: $$

903: \int \sum_{x \in \vertex(\gh)} f_j(\gh, \bp, x) \,d\rtd(\gh, \bp)

904: =

905: \int \sum_{x \in \vertex(\gh)} f_j(\gh, x, \bp) \,d\rtd(\gh, \bp)

906: \,.

907: $$

908: On the other hand,

909: $$

910: \sum_{x \in \vertex(\gh)} f(\gh, \bp, x)

911: =

912: \sum_{x \in \vertex(\gh)} f_1(\gh, \bp, x)

913: \,,

914: $$

915: $$

916: \sum_{x \in \vertex(\gh)} f(\gh, x, \bp)

917: =

918: \sum_{x \in \vertex(\gh)} f_k(\gh, x, \bp)

919: \,,

920: $$

921: and for $1 \le j < k$, we have

922: $$

923: \sum_{x \in \vertex(\gh)} f_j(\gh, x, \bp)

924: =

925: \sum_{x \in \vertex(\gh)} f_{j+1}(\gh, \bp, x)

926: \,.

927: $$

928: Combining this string of equalities yields the desired equation for

929: $f$.

930: \Qed

931: 

932: 

933: \comment{

934: If $\gh$ is a network, $x \in \verts(\gh)$, and $e \in \edges(\gh)$, then

935: the {\bf isomorphism class} of $(\gh, x, e)$ is the class $\{(\gpe \gh, \gpe

936: x, \gpe e) \st \gpe \hbox{ is an isomorphism of }\gh\}$.

937: Let $\gve$ be the set of such isomorphism classes with the natural

938: topology.

939: Given a rooted network $(\gh, x)$ and an edge $e$ incident to $x$, define

940: the involution $\iota(\gh, x, e) := (\gh, y, e)$, where $y$ is the other

941: endpoint of $e$.

942: Given a probability measure $\rtd$ on rooted networks, define the probability

943: measure $\widehat\rtd$ to be the law of the isomorphism class of

944: $(\gh, x, e)$, where $(\gh, x)$ is chosen

945: according to $\rtd$ and $e$ is then chosen uniformly among the edges

946: incident to $x$.

947: Thus,

948: $$

949: \widehat\rtd(\A) :=

950: \int_{\GG_*} {1 \over \deg_\gh x} |\{e \sim x \st (\gh, x, e) \in \A\}|

951: \,d\rtd\big([\gh, x]\big)

952: $$

953: for Borel sets $\A \subseteq \gve$.

954: (Note that we actually choose $[\gh, x]$ according to $\rtd$, then choose

955: $(\gh, x) \in [\gh, x]$ arbitrarily; this latter choice has no effect on

956: the number of $e \sim x$ such that $(\gh, x, e) \in \A$.)

957: Also, define $\widetilde \rtd$ to be the (non-probability) measure that is the

958: result of biasing $\widehat\rtd$ by the degree of the root; that is, the

959: Radon-Nikod\'ym derivative of $\widetilde \rtd$ with respect to $\widehat\rtd$

960: at the isomorphism class of $(\gh, x, e)$ is $\deg_\gh(x)$:

961: $$

962: \widetilde\rtd(\A) :=

963: \int_{\GG_*} |\{e \sim x \st (\gh, x, e) \in \A\}|

964: \,d\rtd\big([\gh, x]\big)

965: $$

966: for Borel sets $\A \subseteq \gve$.

967: (If the expected degree of the root is finite, one could obtain a

968: probability measure from $\widetilde \rtd$ by dividing by the expected degree;

969: but this is not always the case.)

970: The involution $\iota$ induces a pushforward map $\widetilde\rtd \mapsto

971: \widetilde\rtd \circ \iota^{-1}$.

972: We say that $\rtd$ is {\bf unimodular} or {\bf involution invariant} if

973: $\widetilde\rtd \circ \iota^{-1}= \widetilde\rtd$.

974: \comment{This definition works only for locally finite networks. We must resort

975: to the definition by \ref b.AS:obj/ in the general case. The latter seems much

976: harder for probabilists to comprehend.}

977: }%

978: 

979: %It turns out that the proof of our finite-approximation result, \ref

980: %t.TII/, will use only involution invariance.

981: %Thus, an alternative proof of \ref p.invinv/ can be given from the fact

982: %that unimodularity holds for random weak limits of finite networks, in

983: %combination with \ref t.TII/.

984: 

985: Occasionally one uses the Mass-Transport Principle for functions $f$

986: that are not nonnegative. It is easy to see that this use is justified

987: when

988: $$

989: \int \sum_{x \in \vertex(\gh)} |f(\gh, \bp, x)| \,d\rtd(\gh, \bp)

990: < \infty

991: \,.

992: $$

993: 

994: As noted by Oded Schramm (personal communication, 2004), unimodularity can

995: be defined for probability measures on other structures, such as

996: hypergraphs, while involution invariance is limited to graphs (or networks

997: on graphs).

998: 

999: \comment{

1000: Suppose that $\gh$ is an infinite quasi-transitive amenable connected network.

1001: Choose one element $o_i$ from each vertex orbit.

1002: It is shown in \BLPSgip, Proposition 3.6, that there is a probability

1003: measure $\rho$ on the set $ \{ o_i \}$ such that for any F\o{}lner sequence

1004: $H_n$, the relative frequency of vertices in $H_n$ that are in the same orbit

1005: as $o_i$ converges to $\rho(o_i)$. We call this measure $\rho$ the {\bf

1006: natural frequency distribution} of $\gh$.

1007: }%

1008: 

1009: We shall sometimes use the following property of marks.

1010: Intuitively, it says that each vertex has positive probability to be the

1011: root.

1012: 

1013: \procl l.unmark \procname{Everything Shows at the Root}

1014: Suppose that $\rtd$ is a unimodular probability measure on $\GG_*$.

1015: Let $\mk_0$ be a fixed mark and $\marks_0$ be a fixed 

1016: Borel set of marks.

1017: If the mark of the root is a.s.\ $\mk_0$, then the mark of every vertex is

1018: a.s.\ $\mk_0$.

1019: If every edge incident to the root a.s.\

1020: has its edge mark at the root in $\marks_0$,

1021: then all edge marks a.s.\ belong to $\marks_0$.

1022: \endprocl

1023: 

1024: \proof

1025: In the first case, each vertex

1026: sends unit mass to each vertex with a mark different from $\mk_0$.

1027: The expected mass received at the root is

1028: zero. Hence the expected mass sent is 0.

1029: The second case is a consequence of the first, where we put the mark $\mk_0$

1030: at a vertex when all the edge marks at that vertex lie in $\marks_0$.

1031: \Qed

1032: 

1033: When we discuss percolation in \ref s.extend/, we shall find it crucial

1034: that we have a unimodular coupling of the various measures (given by the

1035: standard coupling of Bernoulli percolation in this case).

1036: It would also be very useful to have unimodular couplings in more general

1037: settings.

1038: We now discuss what we mean.

1039: 

1040: Suppose that $\rln \subseteq \marks \times \marks$ is a closed set, which we

1041: think of as a binary relation such as the lexicographic order on Baire space.

1042: Given two measures $\rtd_1, \rtd_2 \in \uni$, say that

1043: $\rtd_1$ is {\bf $\rln$-related} to $\rtd_2$ if there is a probability measure

1044: $\nu$, called an {\bf $\rln$-coupling of $\rtd_1$ to $\rtd_2$},

1045: on rooted networks with mark space

1046: $\marks \times \marks$ such that $\nu$ is concentrated on networks all of

1047: whose marks lie in $\rln$ and whose marginal given by taking the $i$th

1048: coordinate of each mark is $\rtd_i$ for $i = 1, 2$.

1049: In particular, $\rtd_1$ and $\rtd_2$ can be coupled to have the same

1050: underlying rooted graphs.

1051: \comment{

1052: Call $\rln$ {\bf extra separable} if there is a sequence $\Seq{\mk_l}$ and two

1053: Borel maps $\rlnmap_i : \marks \to 2^{\Seq{\mk_l}}$ ($i = 1, 2$)

1054: such that for all $(\mk, \mk') \in \rln$, for all $\zeta_1 \in

1055: \rlnmap_1(\mk)$, and for all $\zeta_2 \in \rlnmap_2(\mk')$, we have $(\zeta_1,

1056: \zeta_2) \in \rln$.

1057: For example, this is the case with the relation $ \le $ on $\R$, where, e.g.,

1058: we may take $\Seq{\mk_l} := \Q$, $\rlnmap_1(\mk) := (-\infty, \mk) \cap \Q$,

1059: and $\rlnmap_2(\mk) := (\mk, \infty) \cap \Q$.

1060: }%

1061: 

1062: It would be very useful to have a positive answer to the following question.

1063: Some uses are apparent in \ref s.trace/ and in \ref s.finite/,

1064: while others appear in \ref b.Lyons:est/ and are hinted at elsewhere.

1065: 

1066: \procl q.coupling \procname{Unimodular Coupling}

1067: Let $\rln \subseteq \marks \times \marks$ be a closed set.

1068: %that is extra separable.

1069: If $\rtd_1, \rtd_2 \in \uni$ and $\rtd_1$ is

1070: $\rln$-related to $\rtd_2$, is there then a unimodular $\rln$-coupling of

1071: $\rtd_1$ to $\rtd_2$?

1072: \endprocl

1073: 

1074: The case where $\rtd_i$ are amenable is established affirmatively in \ref

1075: p.amen-average/.

1076: However, the case where $\rtd_i$ are supported by a fixed underlying

1077: non-amenable Cayley graph is open even when the marks take only two values.

1078: Here is a family of examples to illustrate what we do not know:

1079: 

1080: \procl q.end-couple

1081: Let $T$ be the Cayley graph of $\Z_2 * \Z_2 * \Z_2$ with respect to the

1082: generators $a$, $b$, $c$, which are all involutions.

1083: We label the edges with the generators.

1084: Fix three Borel symmetric

1085: functions $f_a$, $f_b$, $f_c$ from $[0, 1]^2$ to $[0, 1]$.

1086: Also, fix an end $\xi$ of $T$.

1087: Let $U(e)$ be i.i.d.\ Uniform$[0, 1]$ random variables indexed by the edges

1088: $e$ of $T$.

1089: For each edge $e$, let $I_e$ be the two edges adjacent to $e$ that lead

1090: farther from $\xi$ and let $J_e$ be the two other edges that are adjacent to

1091: $e$.

1092: Let $L(e)$ denote the Cayley label of $e$, i.e., $a$, $b$, or $c$.

1093: For an edge $e$ and a pair of edges $\{e_1, e_2\}$, write $f\big(e, \{e_1,

1094: e_2\}\big) := f_{L(e)}\big(U(e_1), U(e_2)\big)$.

1095: Define $X(e) := f(e, I_e)$ and $Y(e) := \max \big\{f(e, I_e), f(e, J_e)\big\}$.

1096: Let $\nu$ be the law of $(X, Y)$.

1097: Let $\rtd_1$ be the law of $X$ and $\rtd_2$ be the law of $Y$.

1098: We use the same notation for the measures in $\uni$ given by rooting $T$ at

1099: the vertex corresponding to the identity of the group.

1100: Let $\rln$ be $\le$ on $[0, 1] \times [0, 1]$.

1101: Since $X(e) \le Y(e)$ for all $e$, $\nu$ is an $\rln$-coupling of $\rtd_1$ to

1102: $\rtd_2$.

1103: In addition, $\rtd_2$ is clearly $\Aut(T)$-invariant (recall that the edges

1104: are labeled), while

1105: the same holds for $\rtd_1$ since it is an i.i.d.\ measure.

1106: Thus, $\rtd_i$ are both unimodular for $i = 1, 2$.

1107: On the other hand, $\nu$ is not $\Aut(T)$-invariant.

1108: Is there an invariant $\rln$-coupling of $\rtd_1$ to $\rtd_2$?

1109: In other words, is there a unimodular $\rln$-coupling of $\rtd_1$

1110: to $\rtd_2$?

1111: %Even much more elementary forms of this question have been open for several

1112: %years.

1113: \endprocl

1114: 

1115: Another example concerns monotone coupling of the wired and free uniform

1116: spanning forests (whose definitions are given below in \ref s.span/).

1117: This question was raised in \BLPSusf; a partial answer was given by \ref

1118: b.Bowen:coupling/.

1119: This is not the only interesting situation involving graph inclusion.

1120: To be more precise about this relation,

1121: for a map $\psi : \marks \to \marks$ and a network $G$, let $\psi(G)$

1122: denote the network obtained from $G$ by replacing each mark with its image

1123: under $\psi$.

1124: Given a Borel subset $\marks_0 \subseteq \marks$ and a network $\gh$,

1125: call the subnetwork consisting of those edges both of whose edge marks lie in

1126: $\marks_0$ the {\bf $\marks_0$-open subnetwork} of $\gh$.

1127: If $\rtd$ and $\rtd'$ are two probability measures on rooted networks, let

1128: us say that $\rtd$ is {\bf edge dominated} by $\rtd'$ if there exists a

1129: measure $\nu$ on $\GG_*$, a Borel subset $\marks_0 \subseteq \marks$, and

1130: Borel functions $\psi, \psi': \marks \to \marks$ such that if $(\gh', \bp)$

1131: denotes a network with law $\nu$ and $(\gh, \bp)$ the component of $\bp$ in

1132: the $\marks_0$-open subnetwork,

1133: then $\big(\psi(\gh), \bp\big)$ has law $\rtd$ and

1134: $\big(\psi'(\gh'), \bp\big)$ has law $\rtd'$.

1135: If the measure $\nu$ can be chosen to be unimodular, then we say that

1136: $\rtd$ is {\bf unimodularly edge dominated} by $\rtd'$.

1137: As a special case of \ref q.coupling/, we do not know whether the existence

1138: of such a measure $\nu$ that is not unimodular implies the existence of

1139: $\nu$ that is unimodular when $\rtd$ and $\rtd'$ are both unimodular

1140: themselves.

1141: 

1142: 

1143: \bsection{Fixed Underlying Graphs}{s.related}

1144: 

1145: Before we study general unimodular probability measures, it is useful to

1146: examine the relationship between unimodularity in the classical sense for

1147: graphs and unimodularity in the sense investigated here for random rooted

1148: network classes.

1149: 

1150: Given a graph $\gh$ and a vertex $x \in \verts(\gh)$, write $\Stab(x) := \{

1151: \gpe \in \Aut(\gh) \st \gpe x = x\}$ for

1152: the {\bf stabilizer subgroup} of $x$.

1153: Also, write $[x] := \Aut(\gh) x$ for the orbit of $x$.

1154: Recall the following principle from \BLPSgip:

1155: 

1156: \proclaim Mass-Transport Principle. If $\gh= (\vertex, \edge)$ is any graph,

1157: %$\gp \subseteq \Aut(\gh)$ is closed,

1158: $f : \vertex \times \vertex \to [0,

1159: \infty]$ is invariant under the diagonal action of $\Aut(\gh)$, and $o, o' \in

1160: \vertex$, then

1161: $$

1162: \sum_{z \in [o']} f(o, z) |\Stab(o')| = \sum_{y \in [o]} f(y, o') |\Stab(y)| \,.

1163: $$

1164: 

1165: Here, $|\cbuldot|$ denotes Haar measure on $\Aut(\gh)$, although we

1166: continue to use this notation for cardinality as well.

1167: Since $\Stab(x)$ is compact and open, $0 < |\Stab(x)| < \infty$.

1168: As shown in \ref b.Schlichting/ and \ref b.Trofimov/,

1169: $$

1170: |\Stab(x)y| / |\Stab(y)x| = |\Stab(x)| / |\Stab(y)|\,;

1171: \label e.schl-trof

1172: \,.

1173: $$

1174: It follows easily that $\gh$ is unimodular iff

1175: $$

1176: |\Stab(x)y| = |\Stab(y)x|

1177: \label e.unimod-schl-trof

1178: $$

1179: whenever $x$ and $y$ are in the same orbit.

1180: 

1181: \procl t.fixedgraph \procname{Unimodular Fixed Graphs}

1182: Let $G$ be a fixed connected graph. Then

1183: $G$ has a random root that gives a unimodular

1184: measure iff $G$ is a unimodular graph with

1185: $$

1186: c := \sum_i |\Stab(o_i)|^{-1} <\infty

1187: \,,

1188: \label e.alm-trn

1189: $$

1190: where $\{ o_i\} $ is a complete orbit

1191: section.

1192: In this case, there is only one such measure $\rtd$ on random rooted graphs

1193: from $G$ and it satisfies

1194: $$

1195: \rtd([G, x]) = c^{-1} |\Stab(x)|^{-1}

1196: \label e.mustab

1197: $$

1198: for every $x \in \verts(G)$.

1199: \endprocl

1200: 

1201: Of course, a similar statement holds for fixed networks.

1202: An example of a graph satisfying \ref e.alm-trn/, but that is not

1203: quasi-transitive, is obtained from the random weak limit of balls in a

1204: 3-regular tree.

1205: That is, let $\vertex := \N \times \N$.

1206: Join $(m, n)$ by edges to each of $(2m, n-1)$ and $(2m+1, n-1)$ for $n \ge

1207: 1$.

1208: The result is a tree with only one end and $\big|\Stab\big((m, n)\big)\big|

1209: = 2^n$.

1210: 

1211: 

1212: \proof

1213: Suppose first that $G$ is unimodular and that $c <\infty$.

1214: Define $\rtd$ by

1215: $$

1216: \all i \rtd([G, o_i]) := c^{-1} |\Stab(o_i)|^{-1}

1217: \,.

1218: $$

1219: To show that $\rtd$ is unimodular, let

1220: $f : \gtwo \to [0, \infty]$ be Borel.

1221: Since we are concerned only with the graph $\gh$, we shall write $f$

1222: instead as a function

1223: $f : \vertex \times \vertex \to [0,

1224: \infty]$ that is $\Aut(\gh)$-invariant.

1225: Then

1226: $$\eqaln{

1227: \int \sum_x f(o, x) \,d\rtd(G, o)

1228: &=

1229: c^{-1} \sum_i \sum_x f(o_i, x) |\Stab(o_i)|^{-1}

1230: \cr&=

1231: c^{-1} \sum_i |\Stab(o_i)|^{-1}

1232: \sum_j |\Stab(o_j)|^{-1} \sum_{x \in [o_j]} f(o_i, x) |\Stab(o_j)|

1233: \cr&=

1234: c^{-1} \sum_i |\Stab(o_i)|^{-1}

1235: \sum_j |\Stab(o_j)|^{-1} \sum_{y \in [o_i]} f(y, o_j) |\Stab(y)|

1236: \cr&\hskip1in\hbox{[by the Mass-Transport Principle for $\gh$]}

1237: \cr&=

1238: c^{-1} \sum_i |\Stab(o_i)|^{-1}

1239: \sum_j |\Stab(o_j)|^{-1} \sum_{y \in [o_i]} f(y, o_j) |\Stab(o_i)|

1240: \cr&\hskip1in\hbox{[by unimodularity of $\gh$]}

1241: \cr&=

1242: c^{-1} \sum_j \sum_y f(y, o_j) |\Stab(o_j)|^{-1}

1243: \cr&=

1244: \int \sum_y f(y, o) \,d\rtd(G, o)

1245: \,.

1246: }$$

1247: Since $\rtd$ satisfies the Mass-Transport Principle, it is unimodular.

1248: 

1249: Conversely, suppose that $\rtd$ is a unimodular probability measure on rooted

1250: versions of $\gh$. To see that $\gh$ is unimodular, consider any $u, v$ in the

1251: same orbit.

1252: Define

1253: $$

1254: \rtd\big([x]\big) :=

1255: \rtd\big([\gh, x]\big)

1256: \,.

1257: $$

1258: We first show that $\rtd\big([u]\big) > 0$.

1259: Every graph isomorphic to $\gh$ has a well-defined notion of vertices of

1260: type $[u]$.

1261: Let each vertex $x$ send mass 1 to each vertex of type $[u]$ that is

1262: nearest to $x$.

1263: This is a Borel function on $\gtwo$ if we transport no mass on graphs

1264: that are not isomorphic to $\gh$.

1265: The expected mass sent is positive, whence so is the expected mass

1266: received.

1267: Since only vertices of type $[u]$ receive mass, it follows that

1268: $\rtd\big([u]\big) > 0$, as desired.

1269: 

1270: Let $f(x, y) := \I{\gp_{u, x} v}(y)$, where $\gp_{u, x} := \{ \gpe

1271: \in\Aut(\gh) \st \gpe u = x \}$.

1272: Note that $y \in \gp_{u, x} v$ iff $x \in \gp_{v, y} u$.

1273: It is straightforward to check that $f$ is diagonally invariant under

1274: $\Aut(\gh)$.

1275: Note that

1276: $$

1277: |\Stab(x) y| \I{[x]}(o)

1278: =

1279: |\gp_{x, \bp} y|

1280: $$

1281: for all $x, y, \bp \in \verts(\gh)$.

1282: Therefore, we have

1283: $$\eqaln{

1284: |\Stab(u) v| \rtd([u])

1285: &=

1286: \int |\gp_{u, o} v| \,d\rtd(\gh, o)

1287: =

1288: \int \sum_x \I{\gp_{u, o} v}(x) \,d\rtd(\gh, o)

1289: \cr&=

1290: \int \sum_x f(o, x) \,d\rtd(\gh, o)

1291: =

1292: \int \sum_x f(x, o) \,d\rtd(\gh, o)

1293: \cr&\hskip1in\hbox{[by the Mass-Transport Principle for $\rtd$]}

1294: \cr&=

1295: \int \sum_x \I{\gp_{u, x} v}(o) \,d\rtd(\gh, o)

1296: =

1297: \int \sum_x \I{\gp_{v, o} u}(x) \,d\rtd(\gh, o)

1298: \cr&=

1299: \int |\gp_{v, o} u| \,d\rtd(\gh, o)

1300: =

1301: |\Stab(v) u| \rtd([v])

1302: \,.

1303: }$$

1304: That is,

1305: $$

1306: |\Stab(u) v| \rtd([u])

1307: =

1308: |\Stab(v) u| \rtd([v])

1309: \,.

1310: \label e.rtdstab

1311: $$

1312: Since $u$ and $v$ are in the same orbit, we have $[u] = [v]$, so $\rtd([u]) =

1313: \rtd([v])$. Since $\rtd([u]) > 0$, we obtain \ref e.unimod-schl-trof/.

1314: That is, $\gh$ is unimodular.

1315: Comparison of \ref e.rtdstab/ with \ref e.schl-trof/ shows \ref e.mustab/.

1316: \Qed

1317: 

1318: Automorphism invariance for random unrooted networks on fixed underlying

1319: graphs is also closely tied to unimodularity of random rooted networks.

1320: Here, we shall need to distinguish between graphs, networks, and

1321: isomorphism classes of rooted networks.

1322: Recall that

1323: $\overline \gh$ denotes a network whose

1324: underlying graph is $\gh$ and $[\overline \gh, \bp]$ denotes an

1325: equivalence class of networks $\overline \gh$ on $\gh$ with root $\bp$.

1326: %Note that $[\overline \gh, \bp]$ carries a natural probability measure

1327: %$\lift_{[\overline \gh, \bp]}$

1328: %induced by the Haar probability measure on $\Stab(\bp)$.

1329: 

1330: Let $\gh$ be a fixed connected unimodular graph satisfying \ref e.alm-trn/.

1331: Fix a complete orbit section $\{\bp_i\}$ of $\vertex(\gh)$.

1332: For a graph $\gh'$ and $x \in \vertex(\gh')$, $z \in

1333: \vertex(\gh)$, let $\Phi(x, z)$ be the set of rooted isomorphisms, if any,

1334: from $(\gh', x)$ to $(\gh, z)$.

1335: When non-empty, this set carries a natural probability measure,

1336: $\lift'_{(\gh', x; z)}$ arising from the Haar probability measure on

1337: $\Stab(z)$.

1338: When $\Phi(x, z) = \emptyset$, let $\lift'_{(\gh', x, z)} := 0$.

1339: Define

1340: $$

1341: \lift_{(\gh', x)} := \sum_{z \in \vertex(\gh)} \lift'_{(\gh', x; z)}

1342: \,.

1343: $$

1344: This is the analogue for isomorphisms from $\gh'$ to $\gh$

1345: of Haar measure on $\Aut(\gh)$.

1346: In particular, any $\gpe \in \Aut(\gh)$ pushes forward $\lift'_{(\gh', x,

1347: z)}$ to $\lift'_{(\gh', x, \gpe z)}$.

1348: 

1349: For a graph $\gh'$ isomorphic to $\gh$ and $x \in \vertex(\gh')$, let

1350: $\type(\gh', x) := \bp_i$ for the unique $\bp_i$ for which

1351: $\Phi(x, \bp_i) \ne \emptyset$.

1352: Note that $\lift_{(\gh', x)} = \lift_{(\gh', y)}$ when $\type(\gh', x) =

1353: \type(\gh', y)$.

1354: 

1355: Every probability measure $\rtd$ on $\GG_*$ that is

1356: concentrated on network classes whose underlying graph is $\gh$ induces a

1357: probability measure $\lift_\rtd$ on unrooted networks on $\gh$:

1358: $$

1359: \lift_\rtd(\A)

1360: :=

1361: \int \int_{\textstyle \Phi\big(\bp, \type(\gh', \bp)\big)}

1362: \I{\A}(\phi \overline \gh')  \,d\lift_{(\gh', \bp)}(\phi)

1363: \,d\rtd([\overline \gh', \bp])

1364: $$

1365: for Borel sets $\A$ of networks on $\gh$.

1366: It is easy to see that this is well defined (the choice of $(\overline

1367: \gh', \bp)$ in its equivalence class not mattering).

1368: 

1369: \procl t.uni-vs-invar \procname{Invariance and Unimodularity}

1370: Let $\gh$ be a fixed connected unimodular graph satisfying \ref e.alm-trn/.

1371: Let $\nu$ be an $\Aut(\gh)$-invariant probability measure on unrooted

1372: networks whose underlying graph is $\gh$.

1373: Then randomly rooting the network as in \ref e.mustab/ gives a measure $\rtd

1374: \in \uni$.

1375: Conversely, let $\rtd \in \uni$ be supported on networks whose underlying

1376: graph is $\gh$.

1377: %Fix $z \in \vertex(\gh)$.

1378: %Let $\rtd_z$ be $\rtd$ conditioned on the root being in the orbit of $z$.

1379: %Define $\nu$ to be the measure induced by $\rtd_z$

1380: %on networks on $\gh$ by identifying $z$

1381: %with $\bp$ under a random element of $\Aut(\gh)$ that takes $z$ to the root.

1382: %\msnote{Phrase better. The point is that $\rtd$ is defined on

1383: %rooted-isomorphism classes.} Then $\nu$ is $\Aut(\gh)$-invariant.

1384: Then $\lift_\rtd$ is $\Aut(\gh)$-invariant.

1385: \endprocl

1386: 

1387: \proof

1388: The first part of the theorem is proved just as is the first part of \ref

1389: t.fixedgraph/,

1390: so we turn to the second part.

1391: Let $\gpe_0 \in \Aut(\gh)$

1392: and $F$ be a bounded Borel-measurable function of networks on $\gh$.

1393: Invariance of $\lift_\rtd$ means that $\int F(\overline

1394: \gh)\,d\lift_\rtd(\overline \gh) = \int F(\gpe_0 \overline

1395: \gh)\,d\lift_\rtd(\overline \gh)$.

1396: To prove that this holds, let

1397: $$

1398: f(\overline \gh', x, y) :=

1399: \int_{\textstyle \Phi\big(x, \type(\gh', y)\big)

1400: \atop \textstyle \cap \Phi\big(y, \gpe_0 \type(\gh, y)\big)}

1401: F(\phi \overline \gh')\,d\lift_{(\gh', y)}(\phi)

1402: \,.

1403: $$

1404: It is straightforward to check that

1405: $f$ is well defined and Borel on $\gtwo$.

1406: %if we define $f$ to be 0

1407: %for networks that are not defined on graphs isomorphic to $\gh$.

1408: Therefore, unimodularity of $\rtd$ gives

1409: $$\eqaln{

1410: \int F(\gpe_0 \overline \gh)\,d\lift_\rtd(\overline\gh)

1411: &=

1412: \int \int_{\textstyle \Phi\big(\bp, \type(\gh', \bp)\big)}

1413: F(\gpe_0 \phi \overline \gh') \,d\lift_{(\gh', \bp)}(\phi)

1414: \,d\rtd([\overline \gh', \bp])

1415: \cr&=

1416: \int \int_{\textstyle \Phi\big(\bp, \gpe_0 \type(\gh', \bp)\big)}

1417: F(\phi \overline \gh') \,d\lift_{(\gh', \bp)}(\phi)

1418: \,d\rtd([\overline \gh', \bp])

1419: \cr&=

1420: \int \sum_{x \in \vertex(\gh')}

1421: \int_{\textstyle \Phi\big(x, \type(\gh', \bp)\big)

1422: \atop \textstyle \cap \Phi\big(\bp, \gpe_0 \type(\gh', \bp)\big)}

1423: F(\phi \overline \gh') \,d\lift_{(\gh', \bp)}(\phi)

1424: \,d\rtd([\overline \gh', \bp])

1425: \cr&=

1426: \int \sum_{x \in \vertex(\gh')}

1427: f(\overline \gh', x, \bp)

1428: \,d\rtd([\overline \gh', \bp])

1429: \cr&=

1430: \int \sum_{x \in \vertex(\gh')}

1431: f(\overline \gh', \bp, x)

1432: \,d\rtd([\overline \gh', \bp])

1433: \cr&=

1434: \int \sum_{x \in \vertex(\gh')}

1435: \int_{\textstyle \Phi\big(\bp, \type(\gh', x)\big)

1436: \atop \textstyle \cap \Phi\big(x, \gpe_0 \type(\gh', x)\big)}

1437: F(\phi \overline \gh') \,d\lift_{(\gh', x)}(\phi)

1438: \,d\rtd([\overline \gh', \bp])

1439: \cr&=

1440: \int \sum_{x \st \type(\gh', x) = \type(\gh', \bp)}

1441: \int_{\textstyle \Phi\big(\bp, \type(\gh', x)\big)

1442: \atop \textstyle \cap \Phi\big(x, \gpe_0 \type(\gh', x)\big)}

1443: F(\phi \overline \gh') \,d\lift_{(\gh', x)}(\phi)

1444: \,d\rtd([\overline \gh', \bp])

1445: \cr&=

1446: \int \sum_{x \st \type(\gh', x) = \type(\gh', \bp)}

1447: \int_{\textstyle \Phi\big(\bp, \type(\gh', \bp)\big)

1448: \atop \textstyle \cap \Phi\big(x, \gpe_0 \type(\gh', \bp)\big)}

1449: F(\phi \overline \gh') \,d\lift_{(\gh', \bp)}(\phi)

1450: \,d\rtd([\overline \gh', \bp])

1451: \cr&=

1452: \int \sum_{x \in \vertex(\gh')}

1453: \int_{\textstyle \Phi\big(\bp, \type(\gh', \bp)\big)

1454: \atop \textstyle \cap \Phi\big(x, \gpe_0 \type(\gh', \bp)\big)}

1455: F(\phi \overline \gh') \,d\lift_{(\gh', \bp)}(\phi)

1456: \,d\rtd([\overline \gh', \bp])

1457: \cr&=

1458: \int

1459: \int_{\textstyle \Phi\big(\bp, \type(\gh', \bp)\big)}

1460: F(\phi \overline \gh') \,d\lift_{(\gh', \bp)}(\phi)

1461: \,d\rtd([\overline \gh', \bp])

1462: \cr&=

1463: \int F(\overline \gh)\,d\lift_\rtd(\gh)

1464: \,.

1465: \Qed

1466: }$$

1467: 

1468: 

1469: \procl r.anygroup

1470: As this section shows, unimodular quasi-transitive graphs are special cases

1471: of unimodular rooted networks.

1472: However, sometimes one is interested in random networks on a graph $\gh$

1473: that are not necessarily invariant under the full group $\Aut(\gh)$, but

1474: only under some subgroup, $\gp \subset \Aut(\gh)$.

1475: This is common when $\gh$ is a Cayley graph of $\gp$.

1476: In this case, we could mark the edges by the generators they represent;

1477: that is, if $x, y \in \gp$ and $y = x a$ with $a$ one of the generators

1478: used to form $\gh$, then we can mark the edge $[x, y]$ at $x$ by $a$.

1479: This makes the full automorphism group of the network

1480: $\overline \gh$ equal to $\gp$, rather than to $\Aut(\gh)$.

1481: The theory here then goes through with only a complication of notation.

1482: However, given any graph $\gh$ and any closed subgroup $\gp \subset

1483: \Aut(\gh)$ that acts quasi-transitively on $\gh$, we do not know whether it

1484: is possible to mark the edges and vertices of $\gh$ to get a network whose

1485: automorphism group is equal to $\gp$.

1486: Yet, the theory for quasi-transitive subgroups is the same; see \BLPSgip.

1487: \endprocl

1488: 

1489: 

1490: 

1491: 

1492: 

1493: 

1494: \bsection{Random Walks and Extremality}{s.extreme}

1495: 

1496: Random walks on networks,

1497: besides being of intrinsic interest,

1498: form an important tool for studying networks.

1499: A random walk is most useful when it has a stationary measure, in other

1500: words, when the distribution of $(\gh, w_0)$ is the same as the

1501: distribution of $(\gh, w_1)$, where $w_0$ is the initial location of the

1502: random walk and $w_1$ is the next location of the random walk.

1503: 

1504: Consider simple random walk on a random graph chosen by a unimodular

1505: probability measure $\rtd$ on rooted graphs, where we start the random walk

1506: at the root.

1507: Just as for finite graphs, we do not expect $\rtd$ to be stationary for the

1508: random walk; rather, we get a stationary measure by biasing $\rtd$ by the

1509: degree of the root.

1510: The fact that this measure is stationary follows from the definition of

1511: involution invariance;

1512: in fact, the definition is precisely that simple random walk is

1513: reversible, i.e., that the distribution of $\big((\gh, w_0), (\gh,

1514: w_1)\big)$ is the same as the distribution of $\big((\gh, w_1), (\gh,

1515: w_0)\big)$, where $(\gh, w_0)$ has distribution $\rtd$ biased by the degree

1516: of the root and $w_1$ is a uniform random neighbor of the 

1517: root.\ftnote{*}{Note that the degree times counting measure is

1518: reversible on every graph, regardless of

1519: unimodularity of the measure on rooted graphs.}

1520: If $\expdeg(\rtd) < \infty$, then we can normalize the biased measure

1521: to obtain a probability measure.

1522: 

1523: In particular, recall from

1524: \ref x.AGW/ the definition of the augmented Galton-Watson measure \AGW.

1525: In \ref b.LPP:GW/, it was remarked in reference to the stationarity of $\AGW$

1526: for simple random walk that ``unlike the situation for finite

1527: graphs, there is no biasing in favor of vertices of large degree".

1528: However, we now see that contrary to this remark,

1529: the situations of finite graphs and $\AGW$ are, in fact, parallel.

1530: That is because the biasing by the degree has already been made part of the

1531: probability measure $\AGW$. The correct comparison of the uniform measure

1532: on vertices of finite graphs is to the

1533: unimodular Galton-Watson probability measure on trees, $\UGW$, because it

1534: is for this measure that ``all vertices are equally likely to be the root".

1535: 

1536: More generally, we can consider

1537: stationarity of random walk in a random environment with random scenery.

1538: Here, if the graph underlies a network, the marks are not restricted to play a

1539: passive role, but may, in fact, determine the transition probabilities (as

1540: in \ref s.trace/) and provide a scenery for the random walk.

1541: %That is, let $\mkmp(x)$ be the vertex mark at $x$.

1542: That is, a Borel function $p : \gtwo \to [0, 1]$, written as

1543: $p : (\gh, x, y) \mapsto p_\gh(x, y)$, such that $\sum_{y \in \verts}

1544: p_\gh(x, y) = 1$ for all vertices $x$

1545: %and such that

1546: %$$

1547: %p_{(\phi(\gh), \phi(\bp))}(\phi(x), \phi(y))

1548: %= p_{(\gh, \bp)}(x, y)

1549: %$$

1550: %for every non-rooted isomorphism $\phi$

1551: is called an {\bf environment}.

1552: A Borel map $\nu : \GG_* \to (0, \infty)$, written

1553: $\nu : (\gh, x) \mapsto \nu_{\gh}(x)$,

1554: %such that

1555: %$$

1556: %\nu_{(\phi(\gh), \phi(\bp))}(\phi(x)) = \nu_{(\gh, \bp)}(x)

1557: %$$

1558: %for every non-rooted isomorphism $\phi$

1559: is called an {\bf initial bias}.

1560: It is called {\bf $p$-stationary} if for all $\gh$, the measure

1561: $\nu_\gh$ is stationary for

1562: the random walk on $\gh$ given by the environment $p_\gh$.

1563: Write $\paths$ for the set of (equivalence classes of)

1564: pairs $\big((\gh, w_0), \Seq{w_n \st n \ge

1565: 0}\big)$ with $(\gh, w_0) \in \GG_*$ and $w_n \in \vertex(\gh)$.

1566: Let $\Ph$ denote the distribution on $\paths$ of

1567: the trajectory of the Markov chain

1568: determined by the environment starting at $\bp$ with initial distribution

1569: equal to $\rtd$ biased by $\nu_{\gh}(\bp)$.

1570: That is, if $\traj_{(\gh, \bp)}$ denotes the probability measure on $\paths$

1571: determined by the environment on $\gh$ with initial vertex

1572: $w_0 = \bp$, then for all events $\ev B$, we have

1573: $$

1574: \Ph(\ev B)

1575: :=

1576: \int_{\GG_*} \traj_{(\gh, \bp)}(\ev B) \nu_{\gh}(\bp)

1577: \,d\rtd(\gh, \bp)

1578: \,.

1579: $$

1580: Let $\invar$ denote the $\sigma$-field of events (in the Borel

1581: $\sigma$-field of $\GG_*$) that are invariant under non-rooted

1582: isomorphisms.

1583: To avoid possible later confusion,

1584: note that this does not depend on the measure $\rtd$, so that even if there

1585: are no non-trivial non-rooted isomorphisms $\rtd$-a.s., the $\sigma$-field

1586: $\invar$ is still not equal (mod 0) to the $\sigma$-field of

1587: $\rtd$-measurable sets.

1588: It is easy to see that

1589: for any $\rtd \in \uni$ and $\ev A \in \invar$ with $\rtd(\ev A) > 0$, the

1590: probability measure $\rtd(\;\cbuldot \mid \ev A)$ is also unimodular.

1591: %Let $\invarp$ be the collection of Borel subsets of $\paths$ that are

1592: %invariant under non-rooted isomorphisms of the underlying network with the

1593: %diagonal action, i.e., under maps $\big((\gh, \bp), \Seq{w_n}\big) \mapsto

1594: %\big((\phi(\gh), \phi(\bp)), \Seq{\phi(w_n)}\big)$ for isomorphisms $\phi:\gh

1595: %\to\gh$.

1596: Define the {\bf shift} $\SS : \paths \to \paths$ by

1597: $$

1598: \SS \big((\gh, w_0), \Seq{w_n}\big) := \big((\gh, w_1), \Seq{w_{n+1}}\big)

1599: \,.

1600: $$

1601: %and $\SS \big((\gh, \bp), \Seq{w_n}\big) := \big((\gh, \bp),

1602: %\Seq{w_n}\big)$ when $\bp \ne w_0$.

1603: The following extends Theorem 3.1 of \ref b.LS:rwrers/;

1604: the proof is essentially the same.

1605: 

1606: %\comment{Here and later, we require local finiteness if this turns out not to

1607: %be in our definition of $\GG_*$.}

1608: 

1609: \procl t.rwrers \procname{Random Walk in a Random Environment and Random

1610: Scenery}

1611: Let $\rtd$ be a unimodular probability measure on $\GG_*$.

1612: Let $p_{\cbuldot}(\cbuldot)$ be an environment and

1613: $\nu_{\cbuldot}(\cbuldot)$ be an initial bias that is $p$-stationary.

1614: Let $\Ph$ be the corresponding measure on trajectories.

1615: Then $\Ph$ is stationary for the shift.

1616: If $p$ is also reversible with respect to $\nu_{\cbuldot}(\cbuldot)$, then

1617: $\Ph$ is reversible, in other words, for all events $\ev A$, $\ev B$, we

1618: have 

1619: $$

1620: \Ph[(\gh, w_0) \in \ev A, (\gh, w_1) \in \ev B]

1621: =

1622: \Ph[(\gh, w_1) \in \ev A, (\gh, w_0) \in \ev B]

1623: \,.

1624: $$

1625: If

1626: $$

1627: \int \nu_{\gh}(\bp) \,d\rtd(\gh, \bp) = 1

1628: \,,

1629: \label e.proba

1630: $$

1631: then $\Ph$ is a probability measure.

1632: \endprocl

1633: 

1634: \proof

1635: The reversibility was not mentioned in prior work, so we give that proof

1636: here.

1637: Assuming that $p$ is $\nu$-reversible, we have 

1638: $$\eqaln{

1639: \Ph[(\gh, w_0) \in \ev A, (\gh, w_1) \in \ev B]

1640: &=

1641: \Ebig{\sum_{x \in \vertex(\gh)} \I{\ev A}(\gh, \bp) \nu_\gh(\bp) p_\gh(\bp,

1642: x) \I{\ev B}(\gh, x)}

1643: \cr&=

1644: \Ebig{\sum_{x \in \vertex(\gh)} \I{\ev A}(\gh, \bp) \nu_\gh(x) p_\gh(x,

1645: \bp) \I{\ev B}(\gh, x)}

1646: \,.

1647: \cr

1648: }$$

1649: The Mass-Transport Principle now gives that this 

1650: $$

1651: =

1652: \Ebig{\sum_{x \in \vertex(\gh)} \I{\ev A}(\gh, x) \nu_\gh(\bp) p_\gh(\bp,

1653: x) \I{\ev B}(\gh, \bp)}

1654: =

1655: \Ph[(\gh, w_1) \in \ev A, (\gh, w_0) \in \ev B]

1656: \,. \Qed

1657: $$

1658: 

1659: \procl r.choice

1660: This theorem is made more useful by noticing that for any $\rtd \in \uni$,

1661: there is a choice of $p_{\cbuldot}(\cbuldot)$ and

1662: $\nu_\cbuldot(\cbuldot)$ that satisfies all the hypotheses, including

1663: \ref e.proba/.

1664: For example, if $F_\gh(x)$

1665: denotes $\sum_{y \sim x} 1/\deg_\gh(y)$,

1666: then let $p_\gh(x, y) := 1/[F_\gh(x) \deg_\gh(y)]$ and

1667: $\nu_\gh(x) := Z^{-1} F_\gh(x)/\deg_\gh(x)$,

1668: where 

1669: $$

1670: Z := \int F_\gh(\bp)/\deg_\gh(\bp) \,d\rtd(\gh, \bp)

1671: \,.

1672: $$

1673: It is clear that $p$ is an environment.

1674: Since $F_\gh(\bp) \le \sum_{y \sim x} 1 = \deg_\gh(\bp)$, we also have that

1675: $Z < \infty$, so that $\nu$ is a $p$-stationary initial bias and $p$ is

1676: $\nu$-reversible.

1677: \endprocl

1678: 

1679: Given a network with positive edge

1680: weights and a time $t > 0$, form the {\bf transition operator} $P_t$ for

1681: continuous-time random walk whose rates are the edge weights; in the case

1682: of unbounded weights (or degrees), we take the minimal process, which dies

1683: after an explosion.

1684: That is, if the entries of a matrix $A$ indexed by the vertices are equal

1685: off the diagonal to the negative of the edge weights and the diagonal

1686: entries are chosen to make the row sums zero, then $P_t := e^{-A t}$; in

1687: the case of unbounded weights, we take the self-adjoint extension of $A$

1688: corresponding to the minimal process.

1689: The matrix $A$ is called the {\bf infinitesimal generator} or the {\bf

1690: Laplacian} of the network.

1691: %the infinitesimal generator is an extension of the Laplacian and it is true

1692: %that the transition probabilities are given by exponentiation (Theorem

1693: %13.37 of Rudin's F.A., 1st ed.). This should be in some probability text,

1694: %too.

1695: 

1696: 

1697: \procl c.continuous-time-stationary

1698: Suppose that $\rtd \in \uni$ is carried by networks with non-negative edge

1699: weights such that the corresponding continuous-time Markov chain has no

1700: explosions a.s.

1701: Then $\rtd$ is stationary and reversible.

1702: \endprocl

1703: 

1704: \proof

1705: Fix $t > 0$ and let $p_\gh(x, y) := P_t(x, y)$.

1706: It is well known that $p$ is reversible with respect to

1707: the uniform measure $\nu_\gh \equiv 1$.

1708: Thus, \ref t.rwrers/ applies.

1709: \Qed

1710: 

1711: We can also obtain a sufficient condition for lack of explosions:

1712: 

1713: \procl c.no-explode

1714: Suppose that $\rtd \in \uni$ is carried by networks with non-negative edge

1715: weights $c_\gh(e)$ such that $Z := \E[\sum_{x \sim \bp} c_\gh(\bp, x)] <

1716: \infty$.

1717: Then the corresponding continuous-time Markov chain has no explosions.

1718: \endprocl

1719: 

1720: \proof

1721: In this case, consider the discrete-time Markov chain corresponding to

1722: these weights.

1723: It has a stationary probability measure arising from the choice

1724: $\nu_\gh(x) := \sum_{y \sim \bp} c_\gh(x, y)/Z$.

1725: It is well known that explosions occur iff 

1726: $$

1727: \sum_{n \ge 0} \nu_\gh(w_n)^{-1} < \infty

1728: $$

1729: with positive probability.

1730: However, stationarity guarantees that this sum is infinite a.s.\

1731: (by the Poincar\'e recurrence theorem).

1732: \Qed

1733: 

1734: \procl r.poss-explode

1735: It is possible for explosions to occur:

1736: For example, consider the uniform spanning tree $T$ in $\Z^2$ (see

1737: \BLPSusf). The only fact we use about $T$ is that it has one end a.s.\

1738: and has an invariant distribution.

1739: Let $c_\gh(e) := 0$ for $e \notin T$ and

1740: $c_\gh(e) := 2^{f(e)}$ when $e \in T$

1741: and $f(e)$ is the number of vertices in the finite component of $T

1742: \setminus e$.

1743: Then it is easy to verify that the corresponding continuous-time Markov

1744: chain explodes a.s.

1745: 

1746: Furthermore, explosions may occur

1747: on a fixed transitive graph that is not unimodular, 

1748: even if the condition in \ref c.no-explode/ is satisfied.

1749: To see this, let

1750: $\xi$ be a fixed end of a regular tree $T$ of degree 3. 

1751: Thus, for every vertex $x$ in $T$, there is a unique ray $x_\xi := 

1752: \Seq{x_0=x, x_1, x_2, \ldots}$ starting at $x$ such that $x_\xi$ and

1753: $y_\xi$ differ by only finitely many vertices for any pair $x$, $y$.

1754: Call $x_1$ the {\bf $\xi$-parent} of $x$, call $x$ a {\bf $\xi$-child} of

1755: $x_1$, and call

1756: $x_2$ the {\bf $\xi$-grandparent} of $x$. Let $\gh$ be the graph obtained

1757: from $T$ by adding the edges $(x, x_2)$ between each $x$ and its

1758: $\xi$-grandparent.

1759: Then $\gh$ is a transitive graph, first mentioned by \ref b.Trofimov/.

1760: In fact, every automorphism of $\gh$ fixes $\xi$.

1761: Now consider the following random weights on $\gh$.

1762: Put weight 0 on every edge in $\gh$ that is not in $T$.

1763: For each vertex of $\gh$, declare open the edge to precisely

1764: one of its two $\xi$-children, chosen uniformly and

1765: independently for different vertices.

1766: The open components are rays.

1767: Let the weight of every edge that is not open also be 0.

1768: If an edge $(x, y)$ between a vertex $x$ and its $\xi$-parent $y$ is open and

1769: $y$ is at distance $n$ from the beginning of the open ray containing $(x, y)$,

1770: then let the weight of the edge be $(3/2)^n$.

1771: Since this event has probability $1/2^{n+1}$, the condition of \ref

1772: c.no-explode/ is clearly satisfied.

1773: It is also clear that the Markov chain explodes a.s.

1774: \endprocl

1775: 

1776: The class $\uni$ of unimodular probability measures on $\GG_*$ is clearly

1777: convex.

1778: An element of $\uni$ is called {\bf extremal} if it cannot be written as a

1779: convex combination of other elements of $\uni$.

1780: We shall show that the extremal measures are those for which $\invar$

1781: contains only sets of measure 0 or 1.

1782: Intuitively,

1783: they are the extremal measures for unrooted networks since

1784: the distribution of the root is forced given the distribution of the

1785: unrooted network.

1786: For example, one may show that $\UGW$ is extremal when conditioned on

1787: non-extinction.

1788: %To do this, use ergodicity of simple random walk, proved in \ref b.LPP/,

1789: %and then appeal to \ref t.erg/, in particular, \ref e.shiftev/, which

1790: %shows that if all such $\ev B$ are trivial, then so are all $\ev A$,

1791: %there, i.e., all of $\invar$.

1792: %Is there a more direct way?

1793: First, we show the following ergodicity property, analogous to Theorem 5.1

1794: of \ref b.LS:indist/.

1795: Recall that a $\sigma$-field is called {\bf $\rtd$-trivial} if all its

1796: elements have measure 0 or 1 with respect to $\rtd$.

1797: 

1798: \procl t.erg \procname{Ergodicity}

1799: Let $\rtd$ be a unimodular probability measure on $\GG_*$.

1800: Let $p_{\cbuldot}(\cbuldot)$ be an environment that satisfies

1801: $$

1802: \all{\gh} \all{x, y \in \vertex(\gh)} \quad x \sim y \Longrightarrow

1803: p_\gh(x, y) > 0

1804: \label e.getsall

1805: $$

1806: and

1807: $\nu_{\cbuldot}(\cbuldot)$ be an initial bias that is $p$-stationary and

1808: satisfies \ref e.proba/.

1809: Let $\Ph$ be the corresponding probability measure on trajectories.

1810: If $\invar$ is $\rtd$-trivial, then

1811: every event that is shift invariant is $\Ph$-trivial.

1812: More generally, the events $\ev B$ in the $\Ph$-completion of the

1813: shift-invariant $\sigma$-field are those of the form 

1814: $$

1815: \ev B

1816: =

1817: \big\{ \big((\gh, \bp), w\big) \in \paths \st (\gh, \bp) \in \ev A \big\}

1818: \xor \ev C

1819: \label e.shiftev

1820: $$

1821: for some $\ev A \in \invar$ and some event $\ev C$ with $\Ph(\ev C) = 0$.

1822: \endprocl

1823: 

1824: \proof

1825: Let $\ev B$ be a shift-invariant event.

1826: As in the proof of Theorem 5.1 of \ref b.LS:indist/, we have $\traj_{(\gh,

1827: \bp)}(\ev B) \in \{ 0, 1 \}$ $\rtd$-a.s.

1828: The set $\ev A$ of $(\gh, \bp)$ where this probability equals 1 is in

1829: $\invar$ by \ref e.getsall/, and a little thought reveals that

1830: \ref e.shiftev/ holds for some $\ev C$ with $\Ph(\ev C) = 0$.

1831: If $\invar$ is $\rtd$-trivial, then

1832: $\rtd(\ev A) \in \{0, 1\}$, whence $\Ph(\ev B) \in \{0, 1\}$ as desired.

1833: Conversely, every event $\ev B$ of the form \ref e.shiftev/ is clearly 

1834: in the $\Ph$-completion of the shift-invariant $\sigma$-field.

1835: \Qed

1836: 

1837: We may regard the space $\paths$ as the space of sequences of rooted

1838: networks, where all roots belong to the same network.

1839: Thus, $\paths$ is the natural trajectory space for the Markov chain with

1840: the transition probability from $(\gh, x)$ to $(\gh, y)$ given by $p_\gh(x,

1841: y)$.

1842: With this interpretation, \ref t.erg/ says that this Markov chain is

1843: ergodic when $\invar$ is $\rtd$-trivial.

1844: The next theorem says that this latter condition is, in turn, equivalent to

1845: extremality of $\rtd$.

1846: 

1847: \procl t.findroot \procname{Extremality}

1848: %Let $(\gh, \bp, \bp')$ be a random connected network with two distinguished

1849: %vertices such that the laws $\rtd$ of $(\gh, \bp)$ and $\rtd'$ of $(\gh,

1850: %\bp')$ are both unimodular.

1851: %Then $\rtd = \rtd'$.

1852: %Equivalently, a unimodular probability measure $\rtd$ on $\GG_*$ is

1853: A unimodular probability measure $\rtd$ on $\GG_*$ is

1854: extremal iff $\invar$ is $\rtd$-trivial.

1855: \endprocl

1856: 

1857: \comment{

1858: For the entire proof, choose an environment and stationary initial bias

1859: that satisfy \ref e.proba/ and \ref e.getsall/.

1860: Also, let $\kappa$ be the law of $(\gh, \bp, \bp')$ on $\gtwo$ that couples

1861: $\rtd$ and $\rtd'$.

1862: 

1863: Let $\A$ be an event of $\GG_*$.

1864: Let $\alpha$ be the function on $\paths$ that gives

1865: the asymptotic frequency of visits to $\A$:

1866: $$

1867: \alpha\big(((\gh, w_0), \Seq{w_n})\big) :=

1868: \liminf_{N \to\infty} {1 \over N} |\{n \le N \st (\gh, w_n) \in \A\}|

1869: \,.

1870: $$

1871: Then $\alpha$ is a shift-invariant function.

1872: %that is measurable with respect to $\invarp$.

1873: For any $t \in [0, 1]$, let $\ev B_t := \{(\gh, \bp) \st \traj_{(\gh,

1874: \bp)}[\alpha \le t] \in \{0, 1\}\}$.

1875: Then $\rtd(\ev B_t), \rtd'(\ev B_t) \in \{ 0, 1 \}$,

1876: as we saw in the proof of \ref t.erg/.

1877: Therefore, there exists a Borel function $f : \GG_* \to [0, 1]$ such

1878: that $\rtd\big[\traj_{(\gh, \bp)}[\alpha = f(\gh, \bp)] = 1 \big] = 1$.

1879: Since a random walk has positive probability of going from $\bp$ to $\bp'$,

1880: it follows that

1881: $$

1882: \kappa[\big[\traj_{(\gh, \bp)}[\alpha = f(\gh, \bp) = f(\gh, \bp')] = 1 \big]

1883: = 1

1884: \,.

1885: \label e.same

1886: $$

1887: Let $\nu \rtd$ stand for the measure $d(\nu \rtd)(\gh, \bp) = \nu_\gh(\bp)

1888: d\rtd(\gh, \bp)$, and likewise for $\nu\rtd'$.

1889: By \ref t.rwrers/, these are stationary measures for simple random walk,

1890: whence the ergodic theorem gives

1891: $$\eqaln{

1892: \int \I\A \nu_\gh(\bp) \,d\rtd(\gh, \bp)

1893: &=

1894: \int f(\gh, \bp) \nu_\gh(\bp) \,d\rtd(\gh, \bp)

1895: }$$

1896: }%

1897: 

1898: \proof

1899: Let $\A \in \invar$. If $\A$ is

1900: not $\rtd$-trivial, then we may write $\rtd$ as a convex combination of $\rtd$

1901: conditioned on $\A$ and $\rtd$ conditioned on the complement of $\A$. 

1902: Each of these two new probability measures is unimodular,

1903: yet distinct, so $\rtd$ is not extremal.

1904: 

1905: Conversely, suppose that $\invar$ is $\rtd$-trivial.

1906: Choose an environment and stationary initial bias

1907: that satisfy \ref e.proba/ and \ref e.getsall/, as in \ref r.choice/.

1908: Let $\A$ be an event of $\GG_*$.

1909: Let $\alpha$ be the function on $\paths$ that gives

1910: the frequency of visits to $\A$:

1911: $$

1912: \alpha\big((\gh, w_0), \Seq{w_n})\big) :=

1913: \liminf_{N \to\infty} {1 \over N} |\{n \le N \st (\gh, w_n) \in \A\}|

1914: \,.

1915: $$

1916: \ref t.rwrers/ allows us to apply the ergodic theorem to deduce that

1917: $\int \alpha \,d\Ph = (\nu\rtd)(\A)$,

1918: where $\nu \rtd$ stands for the measure $d(\nu \rtd)(\gh, \bp) = \nu_\gh(\bp)

1919: d\rtd(\gh, \bp)$.

1920: On the other hand, $\alpha$ is a shift-invariant function, which,

1921: according to \ref t.erg/, means that $\alpha$ is a constant $\Ph$-a.s.

1922: Thus, we conclude that

1923: $\alpha = (\nu\rtd)(\A)$ $\Ph$-a.s.

1924: %\ref t.rwrers/ allows us to apply the ergodic theorem to deduce that

1925: %$\alpha = \int \alpha \,d\Ph = (\nu\rtd)(\A)$ $\Ph$-a.s.,

1926: %where $\nu \rtd$ stands for the measure $d(\nu \rtd)(\gh, \bp) = \nu_\gh(\bp)

1927: %d\rtd(\gh, \bp)$.

1928: Consider any non-trivial

1929: convex combination of two unimodular probability

1930: measures, $\rtd_1$ and $\rtd_2$, that gives $\rtd$.

1931: Then $\Ph$ is a (possibly different) convex combination of $\Ph_1$ and $\Ph_2$.

1932: The above applies to each of $\Ph_i$ ($i=1, 2$) and the associated

1933: probability measures $a_i \nu \rtd_i$, where

1934: $a_i := \left(\int \nu_\gh(\bp) \,d\rtd_i(\gh, \bp)\right)^{-1}$.

1935: Therefore, we obtain that $a_1(\nu\rtd_1)(\A) = (\nu\rtd)(\A) =

1936: a_2(\nu\rtd_2)(\A)$.

1937: Since this holds for all $\A$, we obtain $a_1 (\nu\rtd_1) = a_2

1938: (\nu\rtd_2)$.

1939: Since $\rtd_1$ and $\rtd_2$ are probability measures, this is the same as

1940: $\rtd_1 = \rtd_2$, whence $\rtd$ is extremal.

1941: \Qed

1942: 

1943: We define the {\bf speed} of a path $\Seq{w_n}$ in a graph $\gh$ to be

1944: $\lim_{n \to\infty} \dist_\gh(w_0, w_n)/n$ when this limit exists, where

1945: $\dist_\gh$ indicates the distance in the graph $\gh$.

1946: 

1947: The following extends Lemma 4.2 of \ref b.BLS:pert/.

1948: 

1949: \procl p.speedexists \procname{Speed Exists}

1950: Let $\rtd$ be a unimodular probability measure on $\GG_*$ with an

1951: environment and stationary initial distribution $\nu_{\cbuldot}(\cbuldot)$ with

1952: $\int \nu_{\gh}(\bp) \,d\rtd(\gh, \bp) = 1$, so that the

1953: associated random walk distribution $\Ph$ is a probability measure.

1954: Then the speed of random walk exists $\Ph$-a.s.\ and is equal $\Ph$-a.s.\

1955: to an $\invar$-measurable function.

1956: The same holds for simple random walk when $\expdeg( \rtd) < \infty$.

1957: \endprocl

1958: 

1959: \proof

1960: Let $f_n\big((\gh, \bp), w\big) := \dist_\gh\big(w(0), w(n)\big)$. Clearly

1961: $$

1962: f_{n+m}\big((\gh, \bp), w\big)

1963: \le f_n\big((\gh, \bp), w\big) + f_m\big(\SS^n \big((\gh, \bp), w\big)\big)

1964: \,,

1965: $$

1966: so that the Subadditive Ergodic Theorem ensures

1967: that the speed $\lim_{n \to\infty} f_n\big( (\gh, \bp), w\big)/n$ exists

1968: $\Ph$-a.s.

1969: Since the speed is shift invariant, \ref t.erg/ shows that the speed is

1970: equal $\Ph$-a.s.\ to an $\invar$-measurable function.

1971: The same holds for simple random walk since it has an equivalent stationary

1972: probability measure (degree biasing) when $\expdeg(\rtd) < \infty$.

1973: \Qed

1974: 

1975: In the case of simple random walk on trees, we can actually calculate the

1976: speed\ftnote{*}{The publisher inadvertently changed the following proposition

1977: to a theorem in the published version. Also, the published version had an

1978: incorrect formula for \ref e.genspeed/.}:

1979: 

1980: \procl p.treespeed \procname{Speed on Trees}

1981: Let $\rtd \in \uni$ be concentrated on infinite trees.

1982: If $\rtd$ is extremal and $\expdeg(\rtd) < \infty$, then the speed of

1983: simple random walk is $\Ph$-a.s.\ 

1984: $$

1985: 1 - {2 \over \overline{\deg}(\mu)}

1986: \,.

1987: \label e.genspeed

1988: $$

1989: \endprocl

1990: 

1991: \proof Given a rooted tree $(\gh, \bp)$ and $x \in \vertex(\gh)$, write

1992: $|x|$ for the distance in $\gh$ between $\bp$ and $x$.

1993: The speed of a path $\Seq{w_n}$ is the limit

1994: $$

1995: \lim_{n \to \infty} {1 \over n} |w_n|

1996: = \lim_{n \to \infty} {1 \over n} \sum_{k = 0}^{n-1} \bigl(|w_{k+1}|

1997:        - |w_{k}|\bigr)\,.

1998: $$

1999: Now the strong law of large numbers for martingale differences (\ref

2000: b.Feller:book/, p.~243) gives

2001: $$

2002: \lim_{n \to \infty} {1 \over n} \sum_{k = 0}^{n-1} \bigl(|w_{k+1}|

2003:        - |w_{k}|\bigr)

2004: = \lim_{n \to \infty} {1 \over n} \sum_{k = 0}^{n-1} \E\bigl[|w_{k+1}|

2005:        - |w_{k}| \bigm| \Seq{w_i \st i \le k}\bigr] \quad\hbox{a.s.} %\label e.1

2006: $$

2007: Provided $w_k \ne o$, the $k$th term on the right equals

2008: $$

2009: {\deg_\gh w_k - 2 \over \deg_\gh w_k }\,. %\label e.2

2010: $$

2011: Since $\gh$ is a.s.\ infinite,

2012: $w_k = \bp$ for only a set of $k$ of density 0 a.s., whence the speed equals

2013: $$

2014: \lim_{n \to \infty} {1 \over n} \sum_{k = 0}^{n-1} {\deg_\gh w_k - 2 \over

2015: \deg_\gh w_k }\,. %\label e.sumdrift

2016: $$

2017: Since this is the limit of averages of an ergodic stationary sequence for

2018: the measure $d\sigma(\gh, \bp) = \deg_\gh (\bp)\, d\rtd(\gh,

2019: \bp)/\overline{\deg}(\rtd)$,

2020: the ergodic theorem

2021: tells us that it converges a.s.\ to the $\sigma$-mean of an element of the

2022: sequence,

2023: $$

2024: \int {\deg_\gh(\bp) - 2 \over \deg_\gh(\bp)} d\sigma(\gh, \bp)

2025: \,,

2026: $$

2027: which is the same as \ref e.genspeed/.

2028: \Qed

2029: %Also, see \ref t.deg-infinite/.

2030: 

2031: When we study percolation, the following consequence will be useful.

2032: 

2033: \procl p.transtrees \procname{Comparison of Transience on Trees}

2034: Suppose $\rtd \in \uni$ is concentrated on networks whose underlying graphs

2035: are trees that are transient for simple random walk.

2036: Suppose that the mark space is $(0, \infty)$, that marks $\mkmp(\cbuldot,

2037: \cbuldot)$ on edges are the same at both endpoints, that the environment is

2038: $p_\gh(x, y) := \mkmp(x, y)/\nu_\gh(x)$, where $\nu_\gh(x) := \sum_{y \sim

2039: x} \mkmp(x, y)$, and that $\int \nu_{\gh}(\bp) \,d\rtd(\gh, \bp) = 1$, so

2040: that the associated random walk distribution $\Ph$ is a probability

2041: measure.

2042: Then random walk is also transient with respect to the environment

2043: $p_\cbuldot(\cbuldot)$ $\rtd$-a.s.

2044: \endprocl

2045: 

2046: \proof

2047: Let $\ev A$ be the set of $p_\cbuldot(\cbuldot)$-recurrent networks.

2048: Suppose that $\rtd(\ev A) > 0$.

2049: By conditioning on $\ev A$, we may assume without loss of

2050: generality that $\rtd(\ev A) = 1$.

2051: By \ref t.deg2/ and the recurrence of simple random walk on trees with at

2052: most two ends, we have $\expdeg(\rtd) > 2$.

2053: Let $\epsilon > 0$ be sufficiently small that 

2054: $$

2055: \int |\{x \sim \bp \st \mkmp(\bp, x) \ge \epsilon\}| \,d\rtd(\gh, \bp) > 2

2056: \,.

2057: $$

2058: Since finite trees have average degree strictly less than 2, is follows

2059: that the subnetwork $(\gh_\epsilon, \bp)$, defined to be the connected

2060: component of $\bp$ formed by the edges with marks at least $\epsilon$,

2061: is infinite with positive probability.

2062: Let $\rtd'$ be the law of $(\gh_\epsilon, \bp)$ when $(\gh, \bp)$ has the

2063: law $\rtd$, and conditioned on the event $\ev B$ that 

2064: $(\gh_\epsilon, \bp)$ is infinite.

2065: Then $\rtd' \in \uni$ and $\expdeg(\rtd') > 2$.

2066: %Let $\Ph'$ be the associated measure on random walks.

2067: By \ref p.treespeed/, simple random walk has positive speed $\rtd'$-a.s.,

2068: so, in particular, is transient a.s.

2069: Now simple random walk is the walk corresponding to all edge weights in

2070: $\gh_\epsilon$ being, say, $1/\epsilon$.

2071: Rayleigh's monotonicity principle (\ref b.DoyleSnell/ or \ref b.LP:book/)

2072: now implies that random walk is transient $\Ph$-a.s.\ on $\ev B$.

2073: Thus, $\ev A \cap \ev B = \emptyset$.

2074: This contradicts our initial assumption that $\rtd(\ev A) = 1$.

2075: \Qed

2076: 

2077: The converse of \ref p.transtrees/ is not true, as there are transient

2078: reversible random walks on 1-ended trees (see \ref x.halfplane/ for an

2079: example of such graphs; weights can be defined appropriately).

2080: Also, \ref p.transtrees/ does not extend to arbitrary networks, as one may

2081: construct an invariant network on $\Z^3$ that gives a recurrent random

2082: walk.

2083: %Put very small conductances on very large periodically spaced cube

2084: %boundaries to effectively cut off the origin.

2085: 

2086: 

2087: 

2088: Given two probability measures $\rtd$ and $\rtd'$

2089: on rooted networks and one of the standard

2090: notions of product networks, one can define the {\bf independent product}

2091: $\rtd \itimes \rtd'$ of

2092: the two measures by choosing a network from each measure independently and

2093: taking their product, rooted at the ordered pair of the original roots.

2094: 

2095: \procl p.product \procname{Product Networks}

2096: Let $\rtd$ and $\rtd'$ be two unimodular probability measures on $\GG_*$.

2097: Then their independent product $\rtd \itimes \rtd'$ is also unimodular.

2098: If $\rtd$ and $\rtd'$ are both extremal, then so is $\rtd \itimes \rtd'$.

2099: \endprocl

2100: 

2101: \proof

2102: Let $\gh_n$ and $\gh'_n$ be finite connected networks whose random weak

2103: limits are $\rtd$ and $\rtd'$, respectively.

2104: Then $\gh_n \times \gh'_n$

2105: clearly has random weak limit $\rtd \itimes \rtd'$, whence the product

2106: is unimodular.

2107: Now suppose that both $\rtd$ and $\rtd'$ are extremal.

2108: Let $\ev A \in \invar$.

2109: Then $\ev A_{(\gh', \bp')} :=

2110: \{(\gh, \bp) \st (\gh, \bp) \times (\gh', \bp') \in \ev A \} \in

2111: \invar$ since $\Aut(\gh) \times \Aut(\gh') \subseteq \Aut(\gh \times

2112: \gh')$.

2113: Therefore, $\rtd \ev A_{(\gh', \bp')} \in \{0, 1\}$.

2114: On the other hand, $\ev A_{(\gh', \bp')} = \ev A_{(\gh', \bp'')}$ for all

2115: $\bp'' \in \verts(\gh')$ because $\ev A \in \invar$.

2116: Therefore,

2117: $\ev B := \{(\gh', \bp') \st \rtd \ev A_{(\gh', \bp')}  = 1\} \in \invar$,

2118: whence $\rtd' \ev B \in \{0, 1\}$.

2119: Hence Fubini's theorem tells us that $(\rtd \itimes \rtd')(\ev A) \in

2120: \{0, 1\}$, as desired.

2121: \Qed

2122: 

2123: 

2124: \procl r.weakmixing

2125: Another type of product that can sometimes be defined does not always

2126: produce an extremal measure from two extremal measures.

2127: That is, suppose that

2128: $\rtd$ and $\rtd'$ are two extremal unimodular probability measures on

2129: $\GG_*$ that, for simplicity, we assume are concentrated on networks with a

2130: fixed underlying transitive graph, $\gh$.

2131: Let $\rtd''$ be the measure on networks given by taking a fixed root,

2132: $\bp$, and choosing the marks as $(\mkmp, \mkmp')$, where $\mkmp$ gives a

2133: network with law $\rtd$, $\mkmp'$ gives a network with law $\rtd'$, and

2134: $\mkmp$, $\mkmp'$ are independent.

2135: Then it may be that $\rtd''$ is not extremal.

2136: For an example, consider the following.

2137: Fix an irrational number, $\alpha$.

2138: Given $x \in [0, 1]$, form the network $\gh_x$ on the

2139: integer lattice graph by marking each integer $n$ with the indicator

2140: that the fractional part of $n\alpha + x$ lies in $[0, 1/2]$.

2141: Let $\rtd$ be the law of $(\gh_x, 0)$ when $x$ is chosen uniformly.

2142: Then $\rtd$ is unimodular and extremal (by ergodicity of Lebesgue measure

2143: with respect to rotation by $\alpha$), but if $\rtd' = \rtd$ and $\rtd''$

2144: is the associated measure above, then $\rtd''$ is not

2145: extremal since when the marks come from $x, y \in [0, 1]$,

2146: the fractional part of $x - y$ is $\invar$-measurable.

2147: \endprocl

2148: 

2149: It may be useful to keep in mind the vast difference between stationarity

2150: and reversibility in this context.

2151: For example, let $T$ be a 3-regular tree and $\zeta$ be an end of $T$.

2152: Mark each edge by two independent random variables, one that is uniform on

2153: $[0, 1]$ and the other uniform on $[1, 2]$, with the latter one

2154: at its endpoint closer to $\zeta$ and with all these random variables

2155: mutually independent for different edges.

2156: Then simple random walk is stationary in this scenery, but not

2157: reversible, even though $T$ is a Cayley graph.

2158: 

2159: 

2160: 

2161: \bsection{Trace and Stochastic Comparison}{s.trace}

2162: 

2163: There is a natural trace associated to every measure in $\uni$.

2164: This trace is useful for making various comparisons.

2165: We illustrate this by extending results of \ref b.PittSC/ and

2166: \ref b.FontesMathieu/ on return

2167: probabilities of continuous-time random walks.

2168: %We also give some results that are used by \ref b.Lyons:est/ for

2169: %comparative enumeration.

2170: 

2171: Suppose that $\rtd$ is a unimodular probability measure on $\GG_*$.

2172: %Assume that the vertex marks are a.s.\ distinct.

2173: Consider the Hilbert space $H := \int^\oplus \ell^2\big(\vertex(\gh)\big)

2174: \,d\rtd(\gh, \bp)$, a direct integral (see, e.g., \ref b.Nielsen/ or \ref

2175: b.KR/, Chapter 14).

2176: Here, we always choose the canonical representative for each network,

2177: which, recall, is a network on the vertex set $\N$.

2178: The space $H$ is defined as the set

2179: of ($\rtd$-equivalence classes of) $\rtd$-measurable functions $f$ defined on

2180: canonical rooted networks $(\gh, \bp)$ that satisfy $f(\gh, \bp) \in

2181: \ell^2(\N)$ and $\int \|f(\gh, \bp)\|^2 \,d\rtd(\gh, \bp) < \infty$.

2182: We write $f = \int^\oplus f(\gh, \bp) \,d\rtd(\gh, \bp)$.

2183: The inner product is given by $(f, g) := \int \big(f(\gh, \bp), g(\gh,

2184: \bp)\big) \,d\rtd(\gh, \bp)$.

2185: Let $T : (\gh, \bp) \mapsto T_{\gh, \bp}$ be a measurable assignment of

2186: bounded linear operators on $\ell^2\big(\vertex(\gh)\big) = \ell^2(\N)$

2187: with finite

2188: supremum of the norms $\|T_{\gh, \bp}\|$.

2189: Then $T$ induces a bounded linear operator $T := T^\rtd := \int^\oplus T_{\gh,

2190: \bp} \,d\rtd(\gh, \bp)$ on $H$ via

2191: $$

2192: T^\rtd : \int^\oplus f(\gh, \bp)

2193: \,d\rtd(\gh, \bp) \mapsto \int^\oplus T_{\gh, \bp} f(\gh, \bp)

2194: \,d\rtd(\gh, \bp)

2195: \,.

2196: $$

2197: The norm $\|T^\rtd\|$ of $T^\rtd$ is the $\rtd$-essential supremum of

2198: $\|T_{\gh, \bp}\|$.

2199: Identify each $x \in \vertex(\gh)$ with the vector $\II x \in

2200: \ell^2\big(\vertex(\gh)\big)$.

2201: Let $\alg = \alg(\rtd)$ be the von Neumann algebra of ($\mu$-equivalence

2202: classes of) such maps $T$ that are equivariant in

2203: the sense that for all network isomorphisms $\phi : \gh_1 \to \gh_2$, all

2204: $\bp_1, x, y \in \verts(\gh_1)$ and all $\bp_2 \in \verts(\gh_2)$,

2205: we have $(T_{\gh_1, \bp_1} x, y) = (T_{\gh_2, \bp_2} \phi x, \phi y)$.

2206: For $T \in \alg$, we have in particular that $T_{\gh, \bp}$ depends on

2207: $\gh$ but not on the root $\bp$, so we shall simplify our notation and

2208: write $T_\gh$ in place of $T_{\gh, \bp}$.

2209: For simplicity, we shall even write $T$ for $T_\gh$ when no confusion can

2210: arise.

2211: Recall that if $S$ and $T$ are self-adjoint operators on a Hilbert space

2212: $H$, we write $S \le T$ if $\ip{S u, u} \le \ip{T u, u}$ for all $u \in H$.

2213: We claim that

2214: $$

2215: \tr(T) := \tr_\rtd(T)

2216: := \Ebig{(T_\gh \bp, \bp)} := \int (T_\gh \bp, \bp)

2217: \,d\rtd(\gh, \bp)

2218: $$

2219: is a {\bf trace} on $\alg$, i.e., $\tr(\cbuldot)$ is linear,

2220: $\tr(T) \ge 0$ for $T \ge 0$, and $\tr(S T) = \tr(T

2221: S)$ for $S, T \in \alg$.

2222: Linearity of $\tr$ is obvious.

2223: Also, the second property is obvious since the integrand is nonnegative for $T

2224: \ge 0$. The third property follows from the Mass-Transport Principle:

2225: We have

2226: $$\eqaln{

2227: \Ebig{(S T \bp, \bp)}

2228: &=

2229: \Ebig{(T \bp, S^* \bp)}

2230: =

2231: \Ebigg{\sum_{x \in \vertex(\gh)} (T \bp, x) (x, S^* \bp)}

2232: \cr&=

2233: \Ebigg{\sum_{x \in \vertex(\gh)} (T \bp, x) (S x, \bp)}

2234: =

2235: \Ebigg{\sum_{x \in \vertex(\gh)} (T x, \bp) (S \bp, x)}

2236: \cr&=

2237: \Ebig{(T S \bp, \bp)}

2238: \,.

2239: }$$

2240: In order to justify this use of the Mass-Transport Principle, we check

2241: absolute integrability:

2242: $$\eqaln{

2243: \Ebigg{\sum_{x \in \vertex(\gh)} |(T \bp, x) (x, S^* \bp)|}

2244: &\le

2245: \bigg(\Ebigg{\sum_{x \in \vertex(\gh)} |(T \bp, x)|^2}

2246: \Ebigg{\sum_{x \in \vertex(\gh)} |(x, S^* \bp)|^2} \bigg)^{1/2}

2247: \cr&=

2248: \bigg(\Ebig{\|T \bp\|^2} \Ebig{\|S^* \bp\|^2} \bigg)^{1/2}

2249: %\le

2250: %\|T_{\gh, \bp}\| \cdot \|S_{\gh, \bp}\|

2251: %\,.

2252: %Furthermore, we have

2253: %\Ebig{ \|T_{\gh, \bp}\| \cdot \|S_{\gh, \bp}\|}

2254: \le

2255: %\bigg(\Ebig{ \|T_{\gh, \bp}\|^2} \Ebig{\|S_{\gh, \bp}\|^2} \bigg)^{1/2}

2256: %=

2257: \|T\| \cdot \|S\|

2258: <\infty

2259: \,.

2260: }$$

2261: 

2262: A general property of traces that are finite and normal, as ours is,

2263: is that if $S \le T$,

2264: then $\tr f(S) \le \tr f(T)$ for any increasing function $f : \R \to \R$.

2265: %$\tr(e^S) \le

2266: %\tr(e^T)$. (Proof: for a positive integer $n$, we have $\tr(T^n) - \tr(S^n) =

2267: %\tr(T^n - S^n) = \tr((T-S)U) = \tr((T-S)^{1/2} U (T-S)^{1/2}) \ge 0$ for a

2268: %positive operator $U$.)

2269: One proof is as follows.

2270: First, if $f \ge 0$ and $T$ is self-adjoint, then $f(T) \ge 0$.

2271: Second, if $S, T \ge 0$, then $\tr (S T) = \tr \big(S^{1/2} T S^{1/2}\big)

2272: \ge 0$ since $S^{1/2} T S^{1/2} = (T^{1/2} S^{1/2})^* (T^{1/2}S^{1/2})  \ge 0$.

2273: Third, if $f$ is an increasing polynomial and $S \le T$, then

2274: $$

2275: {d \over d z} \Big(\tr f\big(S + z(T-S)\big)\Big)

2276: =

2277: \tr \Big(f'\big(S + z(T-S)\big)(T-S)\Big)

2278: \ge 0

2279: $$

2280: for $z \ge 0$

2281: since $f' \ge 0$, $S + z(T-S)$ is self-adjoint, and $0 \le T-S$.

2282: This shows that with these restrictive hypotheses, $\tr f(S) \le \tr f(T)$.

2283: Fourth, any monotone increasing function can be approximated by 

2284: an increasing polynomial.

2285: This shows the result in general.

2286: See \ref b.BrownKosaki/, pp.~6--7 for another proof. (They stated the result

2287: only for continuous $f$ with $f(0) = 0$ because they dealt with more

2288: general traces and operators, but such restrictions are not needed in our

2289: situation.)

2290: %They also assume the trace is faithful, which ours is, but that is not used

2291: %in their proof.

2292: 

2293: Recall from \ref s.extreme/ that given

2294: a network with positive edge

2295: weights and a time $t > 0$, we form the transition operator $P_t$ for

2296: continuous-time random walk whose rates are the edge weights; in the case

2297: of unbounded weights (or degrees), we take the minimal process, which dies

2298: after an explosion.

2299: If $A$ is the Laplacian of the network,

2300: then $P_t := e^{-A t}$.

2301: 

2302: \procl t.rtnpr \procname{Return Probabilities}

2303: Let the mark space be $\R^+$ and let $\rln$ be $\le$.

2304: Let $\rtd_i \in \uni$ have edge weights that are the same at both ends of each

2305: edge ($i = 1, 2$).

2306: Suppose that there is a unimodular $\rln$-coupling $\nu$ of

2307: $\rtd_1$ to $\rtd_2$.

2308: Let $P^{(i)}_t$ be the transition operators corresponding to the edge

2309: weights ($i = 1, 2$).

2310: Then 

2311: $$

2312: \int P^{(1)}_t(\bp, \bp) \,d\rtd_1(\gh, \bp) \ge \int P^{(2)}_t(\bp,

2313: \bp) \,d\rtd_2(\gh, \bp)

2314: $$

2315: for all $t > 0$.

2316: \endprocl

2317: 

2318: \proof

2319: Suppose first that the weights are bounded a.s.

2320: The assumptions tell us that the infinitesimal generators $A^{(i)} \in

2321: \alg$ satisfy $A^{(1)} \le A^{(2)}$, so that for all $t > 0$, we have

2322: $-A^{(1)} t \ge -A^{(2)} t$.

2323: Therefore

2324: $\int P^{(1)}_t(\bp, \bp) \,d\rtd_1(\gh, \bp) = \tr_{\nu}(e^{-A^{(1)} t})

2325: \ge \tr_{\nu}(e^{-A^{(2)} t}) = \int P^{(2)}_t(\bp, \bp)

2326: \,d\rtd_2(\gh, \bp)$.

2327: 

2328: In the case of unbounded weights, note that $P^{(i)}_t(\bp, \bp) = \lim_{n

2329: \to\infty} P^{(i, n)}_t(\bp, \bp)$ for $i = 1, 2$, where the superscript $n$

2330: refers to truncating all weights larger than $n$. (This is because we are

2331: looking at the minimal process, which is reversible.)

2332: %Given t >0, there is some distance R

2333: %such that the probability of having gone more than R while returning to o

2334: %at time t is less than epsilon. Thus, truncation will eventually give the

2335: %same rates within distance R of o, so we get the correct limiting return

2336: %probability at time t. Now use the BCT (we are taking expectation over (G,

2337: %o).)

2338: Thus, the result follows from the bounded case, which we have proved.

2339: \Qed

2340: 

2341: \ref t.rtnpr/ extends a result of \ref b.FontesMathieu/, who proved it in the

2342: case of $\Z^d$ for processes without explosions.

2343: \ref b.PittSC/, Lemma 3.1, prove an analogous

2344: comparison result for Cayley graphs, but with different assumptions on the

2345: pairs of rates (which are deterministic for them).

2346: The case of \ref t.rtnpr/ specialized to

2347: finite networks was proved earlier by Benjamini and Schramm;

2348: see Theorem 3.1 of \ref b.HeicklenHoffman/.

2349: 

2350: This theorem also gives a partial answer to a

2351: question of Fontes and Mathieu, who asked whether the same holds when

2352: $\rtd_i$ are supported on a single Cayley graph, are invariant under the

2353: group action, and are $\rln$-related.

2354: For example, the theorem shows that

2355: it holds when the networks (which are the environments for the

2356: random walks) are given by i.i.d.\ edge marks, since in such a case, it is

2357: trivial that being $\rln$-related implies the existence of a unimodular

2358: $\rln$-coupling.

2359: Also, in the amenable case, the existence of a unimodular $\rln$-coupling

2360: follows from the existence of an $\rln$-coupling, as is proved in \ref

2361: p.amen-average/; in the case of fixed amenable transitive graphs, this is a

2362: well-known averaging principle.

2363: 

2364: \procl q.rtnpr

2365: Does \ref t.rtnpr/ hold without the assumption of a unimodular

2366: $\rln$-coupling, but just an $\rln$-coupling?

2367: \endprocl

2368: 

2369: This question asks whether we can compare the traces from two different von

2370: Neumann algebras.

2371: One situation where we can do this is as follows.

2372: 

2373: 

2374: 

2375: If $\rtd_1$ and $\rtd_2$ are probability measures on $\GG_*$, then a

2376: probability measure $\nu$

2377: on $\GG_* \times \GG_*$ whose coordinate

2378: marginals are $\rtd_1$ and $\rtd_2$ is called a {\bf monotone graph

2379: coupling} of $\rtd_1$ and $\rtd_2$ if

2380: $\nu$ is concentrated on pairs of rooted

2381: networks $\big( (\gh_1, \bp), (\gh_2, \bp) \big)$

2382: that share the same roots and satisfy $\vertex(\gh_1) \subseteq

2383: \vertex(\gh_2)$.

2384: %In this case, decompose $\ell^2\big(\vertex(\gh_2)\big) =

2385: %\ell^2\big(\vertex(\gh_1)\big) \oplus \ell^2\big(\vertex(\gh_2) \setminus

2386: %\vertex(\gh_1)\big)$ and let $I = I_1 \oplus I_2$ be the corresponding

2387: %decomposition of the identity operator.

2388: In this instance,

2389: let $V_1$ be the inclusion of $\ell^2\big(\vertex(\gh_1)\big)$ in

2390: $\ell^2\big(\vertex(\gh_2)\big)$.

2391: %and $\bfz_2$ denote the 0 operator on

2392: %$\ell^2\big(\vertex(\gh_2) \setminus \vertex(\gh_1)\big)$.

2393: %Thus, if $T$ is an operator on $\ell^2\big(\vertex(\gh_1)\big)$, then $T

2394: %\oplus \bfz_2$ is its natural extension to

2395: %an operator on $\ell^2\big(\vertex(\gh_2)\big)$.

2396: %Conversely, if $T$ is an operator on $\ell^2\big(\vertex(\gh_2)\big)$, then

2397: %$V_1^* T V_1$ is its restriction to $\ell^2\big(\vertex(\gh_1)\big)$.

2398: %These give two natural ways to compare an operator on

2399: %$\ell^2\big(\vertex(\gh_1)\big)$ to an operator on

2400: %$\ell^2\big(\vertex(\gh_2)\big)$, which accounts for the two parts of our

2401: %next proposition.

2402: 

2403: When there is a unimodular coupling (as in \ref t.rtnpr/),

2404: the following result is easy.

2405: The fact that it holds more generally is useful.

2406: 

2407: 

2408: \comment{

2409: \procl p.equality \procname{Trace Comparison}

2410: Let $\nu$ be a monotone graph coupling of two unimodular probability measures

2411: $\rtd_1$ and $\rtd_2$.

2412: Let $T^{(i)} \in \alg$  ($i = 1, 2$).

2413: Suppose that 

2414: $$

2415: T^{(1)}_{\gh_1} \oplus \bfz_2 \le T^{(2)}_{\gh_2}

2416: \label e.cond1

2417: $$

2418: for $\nu$-almost all pairs $\big( (\gh_1, \bp), (\gh_2, \bp) \big)$.

2419: Then $\tr_{\rtd_1}(T^{(1)}) \le \tr_{\rtd_2}(T^{(2)})$ with equality iff

2420: $$

2421: T^{(1)}_{\gh_1} \oplus \bfz_2 = T^{(2)}_{\gh_2} \quad \nu\hbox{-a.s.}

2422: \label e.defeq1

2423: $$

2424: Similarly, if

2425: $$

2426: T^{(1)}_{\gh_1} \le V_1^* T^{(2)}_{\gh_2} V_1

2427: \label e.cond2

2428: $$

2429: for $\nu$-almost all pairs $\big( (\gh_1, \bp), (\gh_2, \bp) \big)$,

2430: then $\tr_{\rtd_1}(T^{(1)}) \le \tr_{\rtd_2}(T^{(2)})$ with equality iff

2431: $$

2432: T^{(1)}_{\gh_1} = V_1^* T^{(2)}_{\gh_2} V_1 \quad \nu\hbox{-a.s.}

2433: \label e.defeq2

2434: $$

2435: \endprocl

2436: }

2437: 

2438: \procl p.equality \procname{Trace Comparison}

2439: Let $\nu$ be a monotone graph coupling of two unimodular probability measures

2440: $\rtd_1$ and $\rtd_2$.

2441: Let $T^{(i)} \in \alg(\rtd_i)$ be self-adjoint with

2442: $$

2443: T^{(1)}_{\gh_1} \le V_1^* T^{(2)}_{\gh_2} V_1

2444: \label e.cond1

2445: $$

2446: for $\nu$-almost all pairs $\big( (\gh_1, \bp), (\gh_2, \bp) \big)$.

2447: Then $\tr_{\rtd_1}(T^{(1)}) \le \tr_{\rtd_2}(T^{(2)})$.

2448: If in addition

2449: for $\nu$-almost all pairs $\big( (\gh_1, \bp), (\gh_2, \bp) \big)$ and for

2450: all $x \in \vertex(\gh_1)$ we have

2451: $$

2452: \deg_{\gh_1}(x) < \deg_{\gh_2}(x)

2453: \implies

2454: \bigip{T^{(1)}_{\gh_1} x, x} < \bigip{T^{(2)}_{\gh_2} x, x}

2455: \,,

2456: \label e.cond2

2457: $$

2458: then either

2459: $$

2460: \tr_{\rtd_1}(T^{(1)}) < \tr_{\rtd_2}(T^{(2)})

2461: \label e.concl1

2462: $$

2463: or 

2464: $$

2465: \vertex(\gh_1) = \vertex(\gh_2)\,,\ \edges(\gh_1) = \edges(\gh_2)\,,

2466: \ \hbox{ and } \ T^{(1)}_{\gh_1} = T^{(2)}_{\gh_2} \quad \nu\hbox{-a.s.}

2467: \label e.concl2

2468: $$

2469: \endprocl

2470: 

2471: \proof

2472: Suppose that \ref e.cond1/ holds.

2473: The fact that $\tr_{\rtd_1}(T^{(1)}) \le \tr_{\rtd_2}(T^{(2)})$ is an

2474: immediate consequence of the definition of trace and of the hypothesis:

2475: $$\eqaln{

2476: \tr_{\rtd_1}(T^{(1)}) 

2477: &=

2478: \int \bigip{T^{(1)}_{\gh} \bp, \bp} \,d\rtd_1(\gh, \bp)

2479: =

2480: \int \bigip{T^{(1)}_{\gh_1} \bp, \bp} \,d\nu\big( (\gh_1, \bp), (\gh_2,

2481: \bp) \big)

2482: \cr&\le

2483: \int

2484: \bigip{V_1^* T^{(2)}_{\gh_2} V_1 \bp, \bp} \,d\nu\big( (\gh_1, \bp), (\gh_2,

2485: \bp) \big)

2486: \cr&=

2487: \int \bigip{T^{(2)}_{\gh_2} \bp, \bp} \,d\nu\big( (\gh_1, \bp), (\gh_2,

2488: \bp) \big)

2489: =

2490: \tr_{\rtd_2}(T^{(2)})

2491: \,.

2492: }$$

2493: 

2494: Suppose that equality holds in this inequality, i.e., \ref e.concl1/ fails.

2495: Then

2496: $$

2497: \bigip{T^{(1)}_{\gh_1} \bp, \bp}

2498: =

2499: \bigip{T^{(2)}_{\gh_2} \bp, \bp}

2500: \quad \nu\hbox{-a.s.}

2501: $$

2502: \comment{

2503: Let $B_n(\bp)$ be the vertex ball of radius $n$ about $\bp$.

2504: By the Mass-Transport Principle and the preceding equation, we have

2505: $$\eqaln{

2506: \int \sum_{x \in B_n(\bp)} \bigip{T^{(1)}_\gh x, x} \,d\rtd_1(\gh, \bp)

2507: &=

2508: \int |B_n(\bp)| \bigip{T^{(1)}_\gh \bp, \bp} \,d\rtd_1(\gh, \bp)

2509: \cr&=

2510: \int |B_n(\bp)| \bigip{T^{(2)}_\gh \bp, \bp} \,d\rtd_2(\gh, \bp)

2511: \cr&=

2512: \int \sum_{x \in B_n(\bp)} \bigip{T^{(2)}_\gh x, x} \,d\rtd_2(\gh, \bp)

2513: \,.

2514: }$$

2515: Comparing the first and last quantities in this chain of equalities and using

2516: the hypothesis again yields that

2517: }

2518: %

2519: Of course, we also have by hypothesis that $\nu$-a.s.,

2520: $$

2521: \bigip{T^{(1)}_{\gh_1} x, x}

2522: \le

2523: \bigip{T^{(2)}_{\gh_2} x, x}

2524: \label e.ineqatx

2525: $$

2526: for all $x \in \vertex(\gh_1)$.

2527: Assume now that \ref e.cond2/ holds.

2528: We shall prove that \ref e.concl2/ holds.

2529: First,

2530: we claim that $\nu$-a.s.,

2531: $$

2532: \vertex(\gh_1) = \vertex(\gh_2),\ 

2533: \edges(\gh_1) = \edges(\gh_2)\,,

2534: \label e.eqgraphs

2535: $$

2536: and

2537: $$

2538: \bigip{T^{(1)}_{\gh_1} x, x}

2539: =

2540: \bigip{T^{(2)}_{\gh_2} x, x}

2541: %\quad \hbox{for all } x \in B_n(\bp)\;\ \nu\hbox{-a.s.}

2542: \label e.atx

2543: $$

2544: %Hence the same holds for all $x \in \vertex(\gh_1)$.

2545: for all $x \in \vertex(\gh_2)$.

2546: If not, let $k$ be the smallest integer such that with positive

2547: $\nu$-probability, there is a vertex $x$ at $\gh_1$-distance $k$ from $\bp$

2548: where \ref e.atx/ does not hold.

2549: Such a $k$ exists by virtue of \ref e.cond2/.

2550: Consider $\bigip{T^{(i)}_{\gh_i} x, x}$ as part of the mark at $x$.

2551: According to \ref t.rwrers/, the random walk on $\gh_i$ given in \ref

2552: r.choice/ yields a shift-stationary measure $\Ph_i$ on trajectories

2553: $\big((\gh_i, w_0), \Seq{w_n \st n \ge 0}\big)$ for each $i$.

2554: In particular, the distribution of $\bigip{T^{(i)}_{\gh_i} w_k, w_k}$ is

2555: the same as that of $\bigip{T^{(i)}_{\gh_i} w_0, w_0}$.

2556: Now the latter is the same for $i=1$ as for $i=2$ (since $w_0 = \bp$).

2557: Note that for all $x$ at distance less than $k$ from $\bp$, we have

2558: $\deg_{\gh_1}(x) = \deg_{\gh_2}(x)$ by \ref e.cond2/.

2559: Thus, the walks may be coupled together up to time $k$, whence

2560: the distribution of $\bigip{T^{(i)}_{\gh_i} w_k, w_k}$ is

2561: not the same for $i=1$ as for $i=2$ in light of \ref

2562: e.ineqatx/ and choice of $k$.

2563: This is a contradiction.

2564: It follows that $\deg_{\gh_1}(x) = \deg_{\gh_2}(x)$ for all $x \in

2565: \vertex(\gh_1)$, whence \ref e.eqgraphs/ and \ref e.atx/ hold.

2566: 

2567: Now $T := T^{(2)}_{\gh_2} - T^{(1)}_{\gh_1} \ge 0$ is self-adjoint.

2568: It follows that

2569: for any $x, y \in \vertex(\gh_2)$

2570: and any complex number $\alpha$ of modulus 1,

2571: $$

2572: 0

2573: \le

2574: \bigip{T (\alpha x + y), \alpha x + y}

2575: =

2576: 2 \Re \big\{ \alpha \ip{T x, y} \big\}

2577: \,,

2578: $$

2579: whence $\bigip{T x, y} = 0$.

2580: %$$

2581: %\bigip{T^{(1)}_{\gh_1} x, y}

2582: %=

2583: %\bigip{T^{(2)}_{\gh_2} x, y}

2584: %\quad \hbox{for all } x , y \in \vertex(\gh_2)\;\ \nu\hbox{-a.s.}

2585: %$$

2586: That is, $T = 0$ and \ref e.concl2/ holds.

2587: \Qed

2588: 

2589: \comment{

2590: Since we also want to handle unbounded operators, sometimes arising as

2591: unbounded functions of bounded operators, we recall that

2592: a closed densely defined operator is {\bf affiliated}

2593: with $\alg$ if it commutes with all unitary operators

2594: that commute with $\alg$; see, e.g., \ref b.KadRing1/, p.~342.

2595: Write $\affalg$ for the set of closed densely defined operators affiliated

2596: with $\alg$.

2597: Also note that our trace on $\alg$ is obviously finite.

2598: Let $T \in \affalg$ be a self-adjoint operator 

2599: with spectral resolution $E_T$. We define the Borel measure $\spm_{\rtd, T}$

2600: by

2601: $$

2602: \spm_{\rtd, T}(B) :=

2603: \tr_\rtd\big(E_T(B)\big)

2604: \label e.specmsr

2605: $$

2606: for Borel subsets $B \subseteq \R$.

2607: We extend the trace by defining 

2608: $$

2609: \tr_\rtd(T) := 

2610: \int_0^\infty \lambda \,d\spm_{\rtd, T}(\lambda) 

2611: $$

2612: for positive operators $T \in \affalg$ and then by linearity to all of

2613: $\affalg$ when it makes sense.

2614: If $T \ge 0$ and $\tr(T) < \infty$, then $\bp \in \dmn(\sqrt T)$ since

2615: $$\eqaln{

2616: \lim_{n \to\infty} \|E_T[0, n] \sqrt T E_T[0, n] \bp \|^2 

2617: &=

2618: \lim_{n \to\infty} \bigip{E_T[0, n] T E_T[0, n] \bp, \bp}

2619: \cr&=

2620: \lim_{n \to\infty} \int_0^n \lambda \,d\spm_{\rtd, T}(\lambda) 

2621: =

2622: \tr_\rtd(T) < \infty

2623: \,.

2624: }$$

2625: In particular, 

2626: $$

2627: \|\sqrt T \bp\|^2 = \tr_\rtd(T)

2628: \,.

2629: $$

2630: Similarly, if $T$ is self-adjoint with polar decomposition $T = U |T|$ and

2631: $\tr_\rtd\big(|T|\big) < \infty$,

2632: then 

2633: $$

2634: \bigip{U \sqrt{|T|} \bp, \sqrt{|T|} \bp} = \tr_\rtd(T)

2635: \,.

2636: $$

2637: 

2638: % $S \le T$ means that the closure of $T - S$ is positive and self-adjoint.

2639: 

2640: % $V_1^* T^{(2)} V_1$ is affiliated with $\alg(\rtd_1)$ and its traces on

2641: % both algebras agree. We have the inequality of the $\rtd_1$-trace,

2642: % therefore.

2643: 

2644: We now extend \ref p.equality/.

2645: %For the definition of the measure topology, see \ref b.FackKos/, \S 1.5.

2646: 

2647: \procl p.equalityunbdd \procname{Trace Comparison of Unbounded Operators}

2648: Let $\nu$ be a monotone graph coupling of two unimodular probability measures

2649: $\rtd_1$ and $\rtd_2$.

2650: Let $T^{(i)} \in \affalg(\rtd_i)$ be self-adjoint with

2651: $\tr_{\rtd_i} (|T^{(i)}|) < \infty$ ($i = 1, 2$).

2652: %Let $e$ be the spectral resolution of $T^{(2)}$ and $e_n := e(-n, n)$ for

2653: %$n \ge 1$.

2654: %Suppose that $T^{(2, n)} \in \alg(\rtd_2)$ satisfy 

2655: %$$

2656: %|T^{(2, n)}| \le |T^{(2)}|

2657: %\label e.approxs

2658: %$$

2659: %and 

2660: %Then $V_1^* e_n T^{(2)} e_n V_1$ converges in the measure topology to an

2661: %operator $S \in \affalg(\rtd_1)$.

2662: %Suppose that the closure of $S - T^{(1)}$ is a positive operator.

2663: If $\|\sqrt{|T^{(1)}_{\gh_1}|} \bp\| \le \|\sqrt{|T^{(2)}_{\gh_2}|} \bp\|$

2664: for $\nu$-almost all pairs $\big( (\gh_1, \bp), (\gh_2, \bp) \big)$,

2665: then $x \in \dmn(\sqrt{|T^{(i)}|})$ for all $x \in \vertex(\gh_i)$ ($i = 1,

2666: 2$) and $\tr_{\rtd_1}(T^{(1)}) \le \tr_{\rtd_2}(T^{(2)})$.

2667: If in addition

2668: for $\nu$-almost all pairs $\big( (\gh_1, \bp), (\gh_2, \bp) \big)$ and for

2669: all $x \in \vertex(\gh_1)$ we have

2670: $$

2671: \deg_{\gh_1}(x) < \deg_{\gh_2}(x)

2672: \implies

2673: \|\sqrt{|T^{(1)}_{\gh_1}|} x\| < \|\sqrt{|T^{(2)}_{\gh_2}|} x\|

2674: \,,

2675: \label e.cond3

2676: $$

2677: then either

2678: \ref e.concl1/ holds or

2679: \ref e.concl2/ holds.

2680: \endprocl

2681: 

2682: \proof

2683: Note first that $e_n |T^{(2)}| e_n$ is a bounded operator, which is easily

2684: seen to be in $\alg(\rtd_1)$.

2685: Furthermore, this is an increasing sequence of operators that converges

2686: since $L^1(\rtd_1)$ is complete by Theorem 13 of \ref b.Segal/.

2687: \msnote{Or we might want to not use $\le$ in the hypo and use instead

2688: sqrts. Maybe then we don't even need the limits.}

2689: }

2690: 

2691: 

2692: \comment{

2693: \procl r.eq-ext

2694: An extension of the previous proposition is proved similarly, where we

2695: consider marks in $\marks^2$ and assume that $\nu$ is concentrated on pairs

2696: of rooted networks such that when the second coordinate is forgotten, we

2697: obtain identical rooted networks.

2698: \msnote{Explain what this means}

2699: \endprocl

2700: }

2701: 

2702: \comment{

2703: The following extension is useful as well (see \ref b.Lyons:est/).

2704: A continuous function $f : (0, \infty) \to \R$ is called {\bf operator

2705: monotone on $(0, \infty)$} if for any bounded self-adjoint operators $A, B$

2706: with spectrum in $(0, \infty)$ and $A \le B$, we have $f(A) \le f(B)$.

2707: According to L\"owner's theorem,

2708: such $f$ are precisely those that can be represented as

2709: $$

2710: f(t) =

2711: \alpha + \beta t + \int \left[{s \over 1+s^2} - {1 \over t +s}\right]

2712: d\kappa(s)

2713: \label e.pick

2714: $$

2715: for some $\alpha \in \R$, $\beta \ge 0$, and

2716: some positive Borel measure $\kappa$ on $\CO{0, \infty}$ with $\int

2717: d\kappa(s)/(1+s^2) < \infty$; see \ref b.Bhatia/, pp.~138--144 and \ref

2718: b.Donoghue/.

2719: Given $f$ as above, $0 \le T \in \alg$ and $\rtd \in \uni$, say that the

2720: pair $(f, T)$ is {\bf mild} with respect to $\rtd$ if 

2721: $$

2722: \int 

2723: \left[

2724: \int_0^1 \bigip{(T+s)^{-1} \bp, \bp}

2725: \,d\kappa(s)

2726: +

2727: \int_1^\infty \Big( {1 \over s} - \bigip{(T+s)^{-1} \bp, \bp} \Big)

2728: \,d\kappa(s)

2729: \right]

2730: \,d\rtd(\gh, \bp)

2731: < \infty

2732: \,.

2733: $$

2734: We are interested in extending the equality condition of \ref p.equality/

2735: to unbounded functions of unbounded operators.

2736: For simplicity, we extend $\alg$ to the set of equivariant measurable maps

2737: $T$ such that $T_{\gh}$ is a linear operator 

2738: defined on the linear span of the vectors $x \in \vertex(\gh)$.

2739: %Note that if the range of $T_{\gh} \ge I$

2740: %is dense in $\ell^2\big(\vertex(\gh)\big)$,

2741: %then the operator $T_\gh^{-1}$ has norm bounded by

2742: %$1$ on the range of $T_\gh$,

2743: %and therefore it extends to a bounded operator on all of

2744: %$\ell^2(\vertex(\gh)$.

2745: 

2746: 

2747: \procl l.incrlimit

2748: Let $T_n$ be bounded self-adjoint operators on a Hilbert space $\HH$ and

2749: $T$ an unbounded self-adjoint operator on $\HH$ with domain $\DD(T)$.

2750: Suppose that $I \le T_n \le T_{n+1}$ for all $n$ and that for all $x \in

2751: \DD(T)$, we have $\bigip{T_n x, x} \to \bigip{T x, x}$ as $n \to\infty$.

2752: Then $\bigip{T_n^{-1} x, x} \to \bigip{T^{-1} x, x}$ as $n \to\infty$ for

2753: all $x \in \HH$.

2754: \endprocl

2755: 

2756: \proof

2757: Note that $T$ is invertible by

2758: Theorems 13.13 and 13.31 of \ref b.Rudin:FA/.

2759: Since the function $t \mapsto -1/t$ is operator monotone on $(0, \infty)$,

2760: we have that the pointwise limit $A$ of $T_n^{-1}$ exists and satisfies

2761: $T_n^{-1} \ge A \ge T^{-1}$.

2762: Therefore

2763: $T_n \le A^{-1} \le T$.

2764: (The proof of operator monotonicity on p.~114 of \ref b.Bhatia/ applies as

2765: well to unbounded self-adjoint operators.)

2766: On the other hand, the pointwise limit of $T_n$ is $T$ on $\DD(T)$, whence

2767: $A = T$ on $\DD(T)$.

2768: Since $A$ is also self-adjoint, it follows that its domain is the same as

2769: $\DD(T)$ and so $A = T$.

2770: \Qed

2771: 

2772: \msnote{Perhaps we want to do the previous theorem for unbounded operators

2773: instead.}

2774: 

2775: \procl c.unbdd \procname{Traces of Unbounded Operators}

2776: Let $\nu$ be a monotone graph coupling of two unimodular probability measures

2777: $\rtd_1$ and $\rtd_2$.

2778: Let $T^{(i, M)} \in \alg$ for $i = 1, 2$ and $M \in \N$ be such that

2779: $$

2780: T^{(1, M)}_{\gh_1} \oplus \bfz_2 \le T^{(2, M)}_{\gh_2} 

2781: %\,.

2782: \label e.mono

2783: $$

2784: and

2785: $$

2786: 0 \le T^{(i, M)}_{\gh_i} \le T^{(i, M+1)}_{\gh_i}

2787: %\qquad \nu\hbox{-a.s.}

2788: \label e.increasing

2789: $$

2790: for $i = 1, 2$, all $M \in \N$ and for

2791: $\nu$-almost all pairs $\big( (\gh_1, \bp), (\gh_2, \bp) \big)$, where

2792: $\bfz_2$ denotes the 0 operator on

2793: $\ell^2\big(\vertex(\gh_2) \setminus \vertex(\gh_1)\big)$.

2794: Assume that for all $\gh$ and all $x, y \in \vertex(\gh)$, we have

2795: $\lim_{M \to\infty} \bigip{T^{(i, M)}_{\gh} x, y} = \bigip{T^{(i)}_{\gh} x,

2796: y}$ for some $T^{(i)} \in \alg$ that has a self-adjoint

2797: extension ($i = 1, 2$).

2798: %Suppose that for all $s > 0$ and $i = 1, 2$, the range of $T^{(i)}_{\gh_i} +

2799: %s$ is dense in $\ell^2\big(\vertex(\gh_i)\big)$.

2800: Let $f$ be an operator monotone function on $(0, \infty)$ with

2801: $$

2802: \lim_{t \downarrow 0} t f(t) = 0

2803: \label e.nomass

2804: $$

2805: and such that $(f, T^{(i, 1)})$ is mild with respect to $\rtd_i$ for $i = 1,

2806: 2$.

2807: If

2808: $$

2809: \lim_{M \to\infty} \lim_{\epsilon \downarrow 0}

2810: \left[

2811: \tr_{\rtd_2}\big(f(T^{(2, M)} + \epsilon)\big) -

2812: \tr_{\rtd_1}\big(f(T^{(1, M)} + \epsilon)\big)

2813: \right]

2814: =

2815: 0

2816: \label e.limit

2817: $$

2818: and

2819: for $\nu$-almost all pairs $\big( (\gh_1, \bp), (\gh_2, \bp) \big)$ and for

2820: all $x \in \vertex(\gh_1)$ we have

2821: $$

2822: \deg_{\gh_1}(x) < \deg_{\gh_2}(x)

2823: \implies

2824: \forall s > 0 \quad

2825: \Bigip{\big(T^{(1)}_{\gh_1} + s\big)^{-1} x, x} <

2826: \Bigip{\big(T^{(2)}_{\gh_2} + s\big)^{-1} x, x}

2827: \,,

2828: \label e.invs

2829: $$

2830: then $T^{(1)}_{\gh_1} = T^{(2)}_{\gh_2}$ $\nu$-a.s.

2831: \endprocl

2832: 

2833: \proof

2834: Write $f$ as in \ref e.pick/.

2835: If $\kappa = 0$, then the result is trivial, so assume $\kappa \ne 0$.

2836: The assumption \ref e.nomass/ ensures that $\kappa\big(\{0\}\big) = 0$.

2837: For $\epsilon > 0$, $M \in \N$ and 

2838: $\nu$-almost each pair $\big( (\gh_1, \bp), (\gh_2, \bp) \big)$, 

2839: we have

2840: $$\eqaln{

2841: \bigip{f(T^{(2, M)}_{\gh_2} + \epsilon) \bp, \bp}

2842: &-

2843: \bigip{f(T^{(1, M)}_{\gh_1} + \epsilon) \bp, \bp}

2844: \cr&\hskip-1in=

2845: \beta \bigg[\bigip{T^{(2, M)}_{\gh_2} \bp, \bp}

2846: -

2847: \bigip{T^{(1, M)}_{\gh_1} \bp, \bp} \bigg]

2848: \cr&\hskip-1in\qquad+

2849: \int

2850: \left[

2851: \Bigip{ \big(T^{(1, M)}_{\gh_1} + \epsilon +s \big)^{-1} \bp, \bp} -

2852: \Bigip{ \big(T^{(2, M)}_{\gh_2} + \epsilon +s \big)^{-1} \bp, \bp}

2853: \right]

2854: d\kappa(s)

2855: \cr&\hskip-1in\ge

2856: \int

2857: \left[

2858: \Bigip{ \big(T^{(1, M)}_{\gh_1} + \epsilon +s \big)^{-1} \bp, \bp} -

2859: \Bigip{ \big(T^{(2, M)}_{\gh_2} + \epsilon +s \big)^{-1} \bp, \bp}

2860: \right]

2861: d\kappa(s)

2862: %\cr&\hskip-1in\ge

2863: %\int

2864: %\left[

2865: %\Bigip{ \big(T^{(1)}_{\gh_1} + \epsilon +s \big)^{-1} \bp, \bp} -

2866: %\Bigip{ \big(T^{(2)}_{\gh_2} + \epsilon +s \big)^{-1} \bp, \bp}

2867: %\right]

2868: %d\kappa(s)

2869: \,,

2870: \label e.upb

2871: }

2872: $$

2873: where the inequality is from

2874: \ref e.mono/ and the fact that $\beta \ge 0$.

2875: %, while the second inequality

2876: %is from \ref e.increasing/.

2877: By \ref e.increasing/,

2878: %and the fact that

2879: %the function $t \mapsto -1/t$ is operator monotone on $(0, \infty)$,

2880: we have for every $M \ge 1$, $\epsilon > 0$, and $s > 0$,

2881: $$

2882: \bigip{(T^{(i, M)}_\gh+\epsilon + s)^{-1} \bp, \bp}

2883: \le

2884: \bigip{(T^{(i, 1)}_\gh+s)^{-1} \bp, \bp}

2885: \,,

2886: $$

2887: and so for $\epsilon < 1 < s$,

2888: $$

2889: {1 \over s} - \bigip{(T^{(i, M)}_\gh+\epsilon + s)^{-1} \bp, \bp} 

2890: \le

2891: {1 \over 1+s^2} + {1 \over 1+s} - \bigip{(T^{(i, 1)}_\gh+1+s)^{-1} \bp, \bp} 

2892: \,.

2893: $$

2894: Because $(f, T^{(i, 1)})$ is mild with respect to $\rtd_i$, we may

2895: therefore apply the Lebesgue

2896: dominated convergence theorem and \ref l.incrlimit/ to obtain that

2897: $$\eqaln{

2898: \lim_{M \to\infty}

2899: \lim_{\epsilon \downarrow 0}

2900: &\int

2901: \int_0^\infty

2902: \left[

2903: {s \over 1 + s^2} -

2904: \Bigip{ \big(T^{(i, M)}_{\gh} + \epsilon +s \big)^{-1} \bp, \bp}

2905: \right]

2906: d\kappa(s)

2907: \,d\rtd_i(\gh, \bp)

2908: \cr&=

2909: \int

2910: \int_0^\infty

2911: \left[

2912: {s \over 1 + s^2} -

2913: \Bigip{ \big(T^{(i)}_{\gh} + \epsilon +s \big)^{-1} \bp, \bp}

2914: \right]

2915: d\kappa(s)

2916: \,d\rtd_i(\gh, \bp)

2917: \,.

2918: }$$

2919: Combining this with \ref e.limit/ and \ref e.upb/ yields

2920: $$

2921: \int

2922: \int_0^\infty

2923: \left[

2924: \Bigip{ \big(T^{(1)}_{\gh_1} +s \big)^{-1} \bp, \bp} -

2925: \Bigip{ \big(T^{(2)}_{\gh_2} +s \big)^{-1} \bp, \bp}

2926: \right]

2927: d\kappa(s)

2928: \,d\nu\big( (\gh_1, \bp), (\gh_2, \bp) \big)

2929: =

2930: 0

2931: \,.

2932: $$

2933: Now the integrand is non-negative, whence

2934: for $\kappa$-almost all $s > 0$ and

2935: for $\nu$-almost all pairs $\big( (\gh_1, \bp), (\gh_2, \bp) \big)$,

2936: $$

2937: \Bigip{ \big(T^{(1)}_{\gh_1} +s \big)^{-1} \bp, \bp}

2938: =

2939: \Bigip{ \big(T^{(2)}_{\gh_2} +s \big)^{-1} \bp, \bp}

2940: \,.

2941: $$

2942: Note that the operators $\big(T^{(i)}_{\gh_i} +s \big)^{-1}$ are

2943: bounded with norm at most $s^{-1}$.

2944: Therefore, we may conclude from \ref p.equality/ that these operators are

2945: equal $\nu$-a.s.

2946: Now Theorems 13.13 and 13.31 of \ref b.Rudin:FA/ tell us that

2947: $T^{(i)}_{\gh} + s$ is a bijection of its domain with

2948: $\ell^2\big(\vertex(\gh)\big)$.

2949: Hence $T^{(1)}_{\gh_1} + s = T^{(2)}_{\gh_2} + s$ $\nu$-a.s.,

2950: which gives us the desired result.

2951: \Qed

2952: }

2953: 

2954: 

2955: \bsection{Percolation}{s.extend}

2956: 

2957: We now begin our collection of

2958: extensions of results that are known for unimodular fixed graphs.

2959: For most of the remainder of the paper, we consider graphs without marks, or,

2960: equivalently, with constant marks, except that marks are used as explained

2961: below to perform percolation on the given graphs.

2962: We begin this section on percolation with some preliminary results on

2963: expected degree.

2964: 

2965: \procl t.deg-infinite \procname{Minimal Expected Degree}

2966: If $\rtd$ is a unimodular probability measure on $\GG_*$ concentrated on

2967: infinite graphs, then $\expdeg(\rtd) \ge 2$.

2968: \endprocl

2969: 

2970: This is proved exactly like Theorem 6.1 of \BLPSgip\ is proved.

2971: In the context of equivalence relations, this is well known and was perhaps

2972: first proved by \ref b.Levitt/.

2973: 

2974: \procl t.deg2 \procname{Degree Two}

2975: If $\rtd$ is a unimodular probability measure on $\GG_*$ concentrated on

2976: infinite graphs,

2977: then $\expdeg (\rtd) = 2$ iff $\rtd$-a.s.\ $\gh$ is a tree with at most 2 ends.

2978: \endprocl

2979: 

2980: The proof is like that of Theorem 7.2 of \BLPSgip.

2981: 

2982: 

2983: \procl p.limittree \procname{Limits of Trees}

2984: If $G_n$ are finite trees with random weak limit $\rtd$, then

2985: $\expdeg(\rtd) \le 2$ and $\rtd$ is concentrated on trees with at most 2

2986: ends.

2987: \endprocl

2988: 

2989: \proof

2990: Since $\expdeg\big(U(G_n)\big) < 2$, we have $\expdeg(\rtd) \le 2$.

2991: The remainder follows from \ref t.deg2/.

2992: \Qed

2993: 

2994: We now discuss what we mean by percolation on a random rooted network.

2995: Given a probability measure $\rtd \in \uni$, we may wish to randomly

2996: designate some of the edges of the random network ``open".

2997: For example, in Bernoulli($p$) bond percolation, each edge is independently

2998: open with probability $p$.

2999: %Formally, fix a homeomorphism $\phi_2 : \marks \times [0, 1] \to \marks$.

3000: Recall that in \ref s.finite/ 

3001: we have added a second independent uniform $[0, 1]$ coordinate to

3002: the edge marks to form $\rtd^\bn \in \uni$, the standard coupling of

3003: Bernoulli percolation on $\rtd$.

3004: For $p \in [0, 1]$, one can define {\bf Bernoulli($p$) bond

3005: percolation on} $\rtd$

3006: as the measure $\rtd^\bn_p$ that replaces the second coordinate of each edge

3007: mark by ``open" if it is at most $p$ and by ``closed" otherwise.

3008: 

3009: Note that if $\phi : \marks \times [0, 1] \to \marks$ is the projection

3010: onto the first coordinate, then $\rtd = \rtd^\bn_p \circ \phi^{-1}$.

3011: We have changed the mark space, but a fixed homeomorphism would bring it back

3012: to $\marks$.

3013: Thus, more generally, if $\psi : \marks \to \marks$ is Borel, then we call

3014: $\rtd$ a {\bf percolation} on $\rtd \circ \psi^{-1}$.

3015: 

3016: 

3017: \procl d.ins

3018: Let $G=\big(\verts(G),\edges(G)\big)$ be a graph.

3019: Given a configuration $A\in \two^{\edges(G)}$ and an edge

3020: $e\in\edges(G)$, denote $\ins_e A$ the element of $\two^{\edges(G)}$ that

3021: agrees with $A$ off of $e$ and is $1$ on $e$.

3022: For $\A\subset \two^{\edges(G)}$, we write $\ins_e\ev A := \{\ins_e A \st A \in

3023: \A\}$.

3024: For bond percolation, call an edge ``closed" if it is marked ``0" and ``open"

3025: if it is marked ``1".

3026: A bond percolation process $\P$ on $G$ is {\bf insertion tolerant} if

3027: $\P(\ins_e\ev A)>0$

3028: for every $e\in\edges(G)$ and every Borel $\A\subset \two^{\edges(G)}$

3029: satisfying $\P(\A)>0$.

3030: The primary subtlety in extending this notion to percolation on unimodular

3031: random networks is that it may not be possible to pick an edge measurably from

3032: a rooted-automorphism-invariant set of edges.

3033: Thus, we shall make an extra assumption of distinguishability with marks.

3034: That is, a percolation process $\P$ on a unimodular probability measure on

3035: $\GG_*$ is {\bf insertion tolerant} if $\P$-a.s.\ there is no nontrivial

3036: rooted isomorphism of the marked network and

3037: for any event $\A \subseteq \{ (A,\gh)

3038: \st A \subseteq \edges(\gh),\; \gh \in \GG_* \}$ with $\P(\A) > 0$ and any

3039: Borel function $e : \gh \mapsto e(\gh) \in \edges(\gh)$ defined on

3040: $\GG_*$, we have $\P(\ins_e \A) > 0$.

3041: \endprocl

3042: 

3043: For example, Bernoulli($p$) bond percolation

3044: is insertion tolerant when $p\in \OC{0, 1}$.

3045: 

3046: We call a connected component of open edges (and their endpoints) a {\bf

3047: cluster}.

3048: Given a rooted graph $(\gh, \bp)$, define

3049: $$

3050: \pc(\gh, \bp) := \sup \{ p \st \hbox{Bernoulli}(p)\hbox{ percolation on }

3051: \gh \hbox{ has no infinite clusters a.s.}\}

3052: \,.

3053: $$

3054: Clearly $\pc$ is $\invar$-measurable, so if $\rtd \in \uni$ is extremal,

3055: then there is a constant $\pc(\rtd)$ such that $\pc(\gh, \bp) = \pc(\rtd)$

3056: for $\rtd$-a.e.\ $(\gh, \bp)$.

3057: 

3058: 

3059: \procl x.pclower

3060: Even if $\rtd \in \uni$ satisfies $\expdeg(\rtd) < \infty$, it does

3061: not necessarily follow that $\pc(\gh) > 0$ for $\rtd$-a.e.\ $(\gh, \bp)$.

3062: For example, let $p_k := 1/[k(k+1)]$ for $k \ge 1$ and $p_0 := 0$.

3063: Let $\UGW$ be the corresponding unimodular Galton-Watson measure (see

3064: \ref x.AGW/).

3065: Then $\expdeg(\UGW) = 6/(12-\pi^2)$ by \ref e.degAGW/,

3066: but since $\sum k p_k = \infty$,

3067: we have $\pc(\gh) = 0$ a.s.\ by \ref b.Lyons:rwpt/.

3068: \endprocl

3069: 

3070: A more elaborate example

3071: shows that no stochastic bound on the

3072: degree of the root,

3073: other than uniform boundedness,

3074: implies $\pc(\rtd) > 0$:

3075: 

3076: \procl x.pclower2

3077: Given $a_n > 0$, we shall construct $\rtd \in \uni$ such that

3078: $\rtd[\deg_\gh(\bp) \ge n] < a_n$ for all large $n$

3079: and $\pc(\gh) = 0$ for $\rtd$-a.e.\ $(\gh, \bp)$.

3080: We may assume that $\sum a_n < \infty$.

3081: Consider the infinite tree $T$ of degree $3$ and $\bp \in \vertex(T)$.

3082: Let $j \geq 2$ and set $p_j := \frac{1}{2} + \frac{1}{j}$. 

3083: Consider supercritical Bernoulli($p_j$)

3084: bond percolation on $T$,

3085: so that

3086: $$

3087: \theta_j := \P[\bp

3088: \hbox{ belongs to an infinite component}] > 0

3089: \,.

3090: $$

3091: Define $q_j < 1$ by

3092: $p_j q_j = \frac{1}{2} + \frac{1}{2j}$,

3093: so that the fragmentation of an infinite $p_j$-component by an independent

3094: $q_j$ percolation process will still contain infinite components.

3095: Let $N_j$ be the smallest integer with

3096: $(1 - \frac{1}{j^2})^{N_j} < 1 - q_j$.

3097: Finally, choose some sequence $1 > r_j \downarrow 0$ sufficiently fast.

3098: We label the edges of $T$ by the following

3099: operations, performed independently for each $j \geq 2$.

3100: 

3101: Take the infinite components of Bernoulli($p_j$) bond percolation on $T$.

3102: ``Thin" by retaining each component independently with probability $r_j$ and

3103: deleting other components.

3104: For each edge $e$ in the remaining components, let $L_j(e) := N_j$, while

3105: $L_j(e) := 1$ for deleted edges $e$.

3106: 

3107: Now let $L(e) := \sup_j L_j(e)$.

3108: Since $r_j \to 0$ fast, $L(e) < \infty$ for all $e$ a.s.

3109: Consider the graph $G$ obtained by replacing each edge $e$ of $T$ by $L(e)$

3110: parallel chains of length 2.

3111: To estimate $\deg_G(\bp)$, note that the chance that $\bp$ is incident

3112: to an edge $e$ with $L_j(e) = N_j$ (thus contributing at most $3N_j$ to the

3113: degree) equals $r_j\theta_j$.  

3114: Thus, by choice of $\Seq{r_j}$, we can make the root-degree distribution

3115: have tail probabilities eventually less than $a_n$.

3116: Now consider Bernoulli($1/j$) bond percolation on $G$.

3117: For an edge $e$ of $T$ which is replaced by $N_j$ chains of $G$,

3118: the chance of percolating across at least one of these $N_j$

3119: chains is (by definition of $N_j$) larger than $q_j$.

3120: Thus the percolation clusters on $G$ dominate the $q_j$-percolation

3121: clusters on the retained components of the original infinite

3122: $p_j$-percolation clusters on $T$, and as observed above must therefore

3123: contain infinite components.

3124: 

3125: To see how to make this into a probability measure in $\uni$, note that

3126: $L$ is an invariant random network on $T$.

3127: Therefore, we obtain a probability measure in

3128: $\uni$ by \ref t.uni-vs-invar/.

3129: We may now use the edge labels to replace an edge labeled $n$ by $n$ parallel

3130: chains of length 2, followed by

3131: a suitable re-rooting

3132: as in \ref x.stretched/ (below).

3133: This gives a new probability measure in $\uni$ that has

3134: the property desired, since 

3135: the expected degree of the root is finite by the hypothesis $\sum_n a_n <

3136: \infty$.

3137: Also, the re-rooting stochastically decreases the degree since it

3138: introduces roots of degree 2.

3139: \endprocl

3140: 

3141: 

3142: The following extends a well-known result of \ref b.HP:uniq/.

3143: The proof is the same.

3144: 

3145: \procl t.uniq \procname{Uniqueness Monotonicity and Merging Clusters}

3146: Let $\rtd$ be a unimodular probability measure on $\GG_*$.

3147: Let $p_1 < p_2$ and $\P_i$ ($i=1,2$) be the corresponding Bernoulli($p_i$) bond

3148: percolation processes on $\rtd$.

3149: If there is a unique infinite cluster $\P_1$-a.s., then

3150: there is a unique infinite cluster $\P_2$-a.s.  Furthermore, in the

3151: standard coupling of Bernoulli percolation processes, if $\rtd$ is extremal,

3152: then $\rtd$-a.s.\ for all $p_1, p_2$ satisfying $\pc(\rtd) < p_1 < p_2 \le 1$,

3153: every infinite $p_2$-cluster contains an infinite $p_1$-cluster.

3154: \endprocl

3155: 

3156: As a consequence, for extremal $\rtd \in \uni$,

3157: there is a constant $\pu(\rtd)$ such that for any

3158: $p > \pu(\rtd)$, we have $\P_{p}$-a.s., there is a

3159: unique infinite cluster, while for any $p < \pu(\rtd)$,

3160: we have $\P_{p}$-a.s., there is not a unique infinite cluster.

3161: 

3162: Every unimodular probability measure on $\GG_*$ can be written as a Choquet

3163: integral of extremal measures.

3164: In the following, we refer to these extremal measures as ``extremal

3165: components".

3166: 

3167: \procl l.ergcomp

3168: If\/ $\P$ is an insertion-tolerant percolation on a

3169: unimodular random network that is concentrated on infinite graphs, then

3170: almost every extremal component of\/ $\P$ is insertion tolerant.

3171: \endprocl

3172: 

3173: \proof

3174: The proof can be done precisely as that of Lemma 1 of \ref b.GKN/, but

3175: by using the present

3176: Theorems \briefref t.rwrers/, \briefref t.erg/, and \briefref

3177: t.findroot/, as well as \ref r.choice/ to replace the use of a

3178: measure-preserving transformation in \ref b.GKN/ by the shift on

3179: trajectories of a stationary Markov chain.

3180: The fact that $\P$ is concentrated on infinite graphs is used to deduce

3181: that the corresponding Markov chain is not positive recurrent, whence the

3182: asymptotic frequency of visits to any given neighborhood of the root is 0.

3183: \Qed

3184: 

3185: \procl c.ninf \procname{Number of Infinite Clusters}

3186: If\/ $\P$ is an insertion-tolerant percolation on a

3187: unimodular random network, then

3188: $\P$-almost surely, the number of infinite clusters is $0,1$ or $\infty$.

3189: \endprocl

3190: 

3191: The proof is standard for the extremal components; cf.\ \ref b.NS:number/.

3192: 

3193: The following extends \ref b.HP:uniq/ and Proposition 3.9 of \ref

3194: b.LS:indist/.

3195: The proof is parallel to that of the latter.

3196: 

3197: \procl p.niso

3198: \assumptions.

3199: Then $\P$-a.s.\ each infinite cluster that has at least $3$ ends has

3200: no isolated ends.

3201: \endprocl

3202: 

3203: The following corollary is proved just like Proposition 3.10 of \ref

3204: b.LS:indist/.

3205: There is some overlap with Theorem 3.1 of \ref b.Paulin/.

3206: 

3207: \procl c.isolated \procname{Many Ends}

3208: \genassumptions.

3209: If there are infinitely many infinite clusters $\P$-a.s., then $\P$-a.s.\ every

3210: infinite cluster has continuum many ends and no isolated end.

3211: \endprocl

3212: 

3213: The following extends Lemma 7.4 and Remark 7.3

3214: of \BLPSgip\ and is proved similarly.

3215: 

3216: \procl l.trim \procname{Subforests}

3217: \assumptions.

3218: If\/ $\P$-a.s.\ there is a component of the open subgraph

3219: $\omega$ with at least three ends, then

3220: there is a percolation $\fo$ on $\omega$ whose components are trees

3221: such that

3222: a.s.\ whenever a component $K$ of $\omega$ has at least three ends,

3223: there is a component of $K \cap \fo$ that has infinitely many ends and has

3224: $\pc < 1$.

3225: \endprocl

3226: 

3227: The following extends Proposition 3.11 of \ref b.LS:indist/ and is proved

3228: similarly (using the preceding \ref l.trim/).

3229: 

3230: \procl p.trans \procname{Transient Subtrees}

3231: \genassumptions.

3232: If there are $\P$-almost surely infinitely many

3233: infinite clusters,  then $\P$-a.s.\ each infinite cluster is transient and,

3234: in fact, contains a transient tree.

3235: \endprocl

3236: 

3237: In order to use this, we shall use the comparison of simple to network

3238: random walks given in \ref p.transtrees/.

3239: 

3240: \procl d.inds

3241: A percolation process $\P$ on a unimodular probability measure on

3242: $\GG_*$ has {\bf indistinguishable infinite clusters} if

3243: for any event $\A \subseteq \{ \big(A,(\gh, \bp)\big)

3244: \st A \in \two^{\vertices(\gh)} \times \two^{\edges(\gh)},\;

3245: (\gh, \bp) \in \GG_* \}$ that is invariant under non-rooted isomorphisms,

3246: almost surely, for all infinite clusters $C$ of the open subgraph

3247: $\omega$, we have $(C,\omega)\in\ev A$, or for all

3248: infinite clusters $C$, we have $(C,\omega)\notin\ev A$.

3249: \endprocl

3250: 

3251: 

3252: The following extends Theorem 3.3 of \ref b.LS:indist/ and is proved

3253: similarly using the preceding results:

3254: For example, instead of the use of delayed simple random walk by \ref

3255: b.LS:indist/, we use the network random walk in \ref r.choice/. 

3256: This is a reversible random walk corresponding to edge weights $(x, y)

3257: \mapsto 1/[(\deg x)(\deg y)]$.

3258: It is transient by Propositions \briefref p.trans/ and \briefref

3259: p.transtrees/, combined with Rayleigh's monotonicity principle.

3260: 

3261: \procl t.cerg \procname{Indistinguishable Clusters}

3262: If\/ $\P$ is an insertion-tolerant percolation on a

3263: unimodular random network, then

3264: $\P$ has indistinguishable infinite clusters.

3265: \endprocl

3266: 

3267: Among the several consequences of this result is the following extension of

3268: Theorem 4.1 of \ref b.LS:indist/, proved similarly.

3269: 

3270: \procl t.tau \procname{Uniqueness and Long-Range Order}

3271: \genassumptions, $\rtd$.

3272: If\/ $\P$ is extremal and there is more than one infinite cluster $\P$-a.s.,

3273: then $\rtd$-a.s.,

3274: $$

3275: \inf \big \{ \P[\hbox{there is an open path from $x$ to } y \mid \gh]

3276: \st x, y \in \vertex(\gh) \big \} = 0

3277: \,.

3278: $$

3279: \endprocl

3280: 

3281: The following extends Theorem 6.12 of \ref b.LS:indist/ and is proved

3282: similarly.

3283: 

3284: \procl t.product \procname{Uniqueness in Products}

3285: Suppose that $\rtd$, $\rtd_1$, and $\rtd_2$ are extremal unimodular probability

3286: measures on $\GG_*$, with $\rtd$ supported on infinite graphs and $\rtd_1$

3287: a percolation on $\rtd_2$. Then $\pu(\rtd \itimes \rtd_1) \ge

3288: \pu(\rtd \itimes \rtd_2)$. In particular, $\pu(\rtd) \ge

3289: \pu(\rtd \itimes \rtd_2)$.

3290: \endprocl

3291: 

3292: More results on percolation will be presented in \ref s.amen/.

3293: 

3294: 

3295: 

3296: \bsection{Spanning Forests}{s.span}

3297: 

3298: An interesting type of percolation other than Bernoulli is given by certain

3299: random forests.

3300: There are two classes of such random forests that have been widely studied,

3301: the uniform ones and the minimal ones.

3302: 

3303: We first discuss the uniform case.

3304: Given a finite connected graph, $\gh$, let $\ust(\gh)$ denote the uniform

3305: measure on spanning trees on $\gh$.

3306: \ref b.Pemantle:ust/

3307: proved a conjecture of Lyons, namely,

3308: that if an infinite connected graph $\gh$ is exhausted by a sequence of finite

3309: connected subgraphs $\gh_n$, then the weak limit of

3310: $\Seq{\ust(\gh_n)}$ exists.

3311: However, it may happen that the limit measure is not supported on trees,

3312: but on forests.

3313: This limit measure is now called the {\bf free (uniform) spanning forest} on

3314: $\gh$, denoted $\fsf$ or $\FUSF$.

3315: If $\gh$ is itself a tree, then this measure is

3316: trivial, namely, it is concentrated on $\{\gh\}$. Therefore, \ref

3317: b.Hag:umesft/

3318: introduced another limit that had been considered on $\Z^d$

3319: more implicitly by \ref b.Pemantle:ust/ and explicitly by

3320: \ref b.Hag:rcust/, namely, the weak limit of the uniform

3321: spanning tree measures on $\gh_n^*$, where $\gh_n^*$ is the graph $\gh_n$ with

3322: its boundary identified (``wired")

3323: to a single vertex. As \ref b.Pemantle:ust/ showed,

3324: this limit also always exists

3325: on any graph and is now called the {\bf wired (uniform) spanning forest},

3326: denoted $\wsf$ or $\WUSF$.

3327: It is clear that both $\fsf$ and $\wsf$ are concentrated on the set of

3328: spanning forests\ftnote{*}{By a

3329: ``spanning forest'', we mean a subgraph without cycles that contains every

3330: vertex.} of $\gh$ that are {\bf essential}, meaning that all their

3331: trees are infinite.

3332: %Furthermore, since the limits exist regardless of the exhaustion chosen, both

3333: %$\fsf_\gh$ and $\wsf_\gh$ are invariant under any automorphisms that $\gh$ may

3334: %have.

3335: Both $\fsf$ and $\wsf$ are important in their own right; see \ref

3336: b.Lyons:bird/ for a survey and \BLPSusf\ for a comprehensive

3337: treatment.

3338: 

3339: In all the above, one may work more generally with a weighted graph, where

3340: the graph has positive weights on its edges.

3341: In that case, $\ust$ stands for the measure such that the probability of a

3342: spanning tree is proportional to the product of the weights of its edges.

3343: The above theorems continue to hold and we use the same notation for the

3344: limiting measures.

3345: 

3346: 

3347: Most results known about the uniform spanning forest measures hold for

3348: general graphs. Some, however, require extra hypotheses such as

3349: transitivity and unimodularity.

3350: We extend some of these latter results here.

3351: 

3352: Given $\rtd$, taking the wired uniform spanning forest on each graph

3353: gives a percolation that we denote $\WUSF(\rtd)$.

3354: Our first result shows, among other things,

3355: that the kind of limit considered in this paper,

3356: i.e., random weak convergence, gives another natural way to define $\wsf$.

3357: It might be quite useful to have an explicit description of measures on

3358: finite graphs whose random weak limit is the {\it free\/} spanning forest.

3359: 

3360: \procl p.ust-limit \procname{UST Limits}

3361: If $\rtd$ is a unimodular probability measure on infinite networks in

3362: $\GG_*$, then $\expdeg\big(\WUSF(\rtd)\big) = 2$.

3363: If $\gh_n$ are finite connected networks whose random weak limit is $\rtd$,

3364: then $\UST(\gh_n) \cd \WUSF(\rtd)$.

3365: More generally, if $\rtd_n$ are unimodular probability measures on $\GG_*$

3366: with $\rtd_n \cd \rtd$, then $\WUSF(\rtd_n) \cd \WUSF(\rtd)$.

3367: \endprocl

3368: 

3369: \proof

3370: Suppose first that $\rtd$ is concentrated on recurrent networks.

3371: Then Wilson's algorithm gives both $\UST(\gh_n)$ and $\WUSF(\rtd)$.

3372: Hence it is clear that $\UST(\gh_n) \cd \WUSF(\rtd)$.

3373: Since $\expdeg(\UST(\gh_n)) = 2 - 2/|\vertex(\gh_n)| \le 2$,

3374: this limit relation in combination with

3375: Fatou's lemma shows that $\expdeg(\WUSF(\rtd)) \le 2$,

3376: whence equality results from \ref t.deg-infinite/.

3377: %We could also have used, say, \ref p.treespeed/.

3378: 

3379: Suppose next that $\rtd$ is concentrated on transient networks.

3380: Then the proof of Theorem 6.5 of \BLPSusf\ gives the same result.

3381: 

3382: Finally, if $\rtd$ is concentrated on neither recurrent nor transient

3383: networks, then we may write $\rtd$ as a mixture of two unimodular measures

3384: that are concentrated on recurrent or on transient networks and apply the

3385: preceding.

3386: 

3387: This proves the first sentence. The second sentence is a special case of

3388: the third, so we prove the third.

3389: 

3390: \comment{

3391: Suppose first that

3392: $\gh_n$ are finite connected networks whose random weak limit is $\rtd$.

3393: Given a positive integer $R$, let $\UST_R(\rtd)$ be the uniform spanning

3394: tree on the wired ball of radius $R$ about the root, and likewise for

3395: $\UST_R(\gh_n)$.

3396: By definition, we have $\UST_R(\rtd) \cd \WUSF(\rtd)$ as $R \to\infty$.

3397: Clearly, $\UST_R(\gh_n) \cd \UST_R(\rtd)$ as $n \to\infty$.

3398: Furthermore, the intersection of $\UST(\gh_n)$ with the ball of radius $R$

3399: stochastically dominates $\UST_R(\gh_n)$ by the theorem of \ref b.FedMih/.

3400: Therefore, every weak limit point of $\Seq{\UST(\gh_n)}$ stochastically

3401: dominates $\UST_R(\rtd)$ and therefore also $\WUSF(\rtd)$.

3402: On the other hand, from \ref p.limittree/,

3403: every weak limit point of $\Seq{\UST(\gh_n)}$

3404: has expected degree at most 2, while by \ref t.deg-infinite/, we have

3405: $\expdeg\big(\WUSF(\rtd)\big) \ge 2$.

3406: It follows that $\UST(\gh_n) \cd \WUSF(\rtd)$ as $n \to\infty$ and that

3407: $\expdeg\big(\WUSF(\rtd)\big) = 2$.

3408: 

3409: By \ref t.TII/, it follows that

3410: $\expdeg\big(\WUSF(\rtd)\big) = 2$ for all $\rtd \in \uni$.

3411: }%

3412: %

3413: %We may now establish the last assertion of the proposition.

3414: If $\rtd_n \cd \rtd$, then

3415: $\UST_R(\rtd_n) \cd \UST_R(\rtd)$ as $n \to\infty$ and the

3416: intersection of $\WUSF(\rtd_n)$ with the ball of radius $R$

3417: stochastically dominates $\UST_R(\rtd_n)$, so that every weak limit point of

3418: $\Seq{\WUSF(\rtd_n)}$ stochastically dominates $\WUSF(\rtd)$.

3419: However, since they all have expected degree 2 by what we have just shown,

3420: they are equal.

3421: \Qed

3422: 

3423: We next show that the trees of the $\wsf$ have only one end a.s.

3424: The first theorem of this type was proved by \ref b.Pem:ust/.

3425: His result was completed and extended in

3426: Theorem 10.1 of \BLPSusf, which

3427: dealt with the transitive unimodular case.

3428: The minor modifications needed for the quasi-transitive unimodular case

3429: were explained by \ref b.Lyons:est/.

3430: Another extension is given by \ref b.LMS:wsf/, who showed that for graphs with

3431: a ``reasonable" isoperimetric profile, each tree has only one end

3432: $\wsf$-a.s.

3433: 

3434: \procl t.1endUSF \procname{One End}

3435: If $\rtd$ is a unimodular probability measure on $\GG_*$ that is

3436: concentrated on transient networks with bounded degree,

3437: then $\WUSF(\rtd)$-a.s., each tree has exactly one end.

3438: \endprocl

3439: 

3440: \proof

3441: The proof is essentially the same as in \BLPSusf, with the following

3442: modifications.

3443: In the proof of Theorem 10.3 of \BLPSusf, which is the case where there is

3444: only one tree a.s., we replace $x$ and $y$ there by $\bp$ and $X_n$, where

3445: $\Seq{X_n}$ is the simple random walk starting from the root; to use

3446: stationarity, we bias the underlying network by the degree of the root.

3447: This gives a measure equivalent to $\rtd$, so that almost sure conclusions for

3448: it hold for $\rtd$ as well.

3449: The stationarity and reversibility give that the probability that the

3450: random walk from $\bp$ ever visits $X_n$ is equal to the probability that a

3451: random walk from $X_n$ ever visits $\bp$.

3452: By transience, this tends to 0 as $n \to\infty$, which allows the proof of

3453: \BLPSusf\ to go through.

3454: [Here, we needed finite

3455: expected degree to talk about probability since we used the equivalent

3456: probability measure of biasing by the degree.]

3457: 

3458: In the proof of Theorem 10.4 of \BLPSusf, the case when there is more than

3459: one tree a.s.,

3460: we need the degrees to be bounded for the displayed equality on p.~36 of

3461: \BLPSusf\ to hold up to a constant factor.

3462: \Qed

3463: 

3464: For our proof, we had to assume transience;

3465: there is presumably an extension to the recurrent case, which would say

3466: that the number of ends $\WUSF(\rtd)$-a.s.\ is the same as the number of

3467: ends $\rtd$-a.s.\ when $\rtd$ is concentrated on recurrent networks.

3468: Also, presumably the assumption that the degrees are bounded is not needed.

3469: In any case, our result here goes beyond what has been done before and

3470: gives a partial answer to Question 15.4 of \BLPSusf; removing the

3471: assumption of bounded degrees would completely answer that question.

3472: It also applies, e.g., to transient clusters of Bernoulli percolation; see

3473: \ref b.GKZ/ and \ref b.BLS:pert/ for sufficient conditions for transience.

3474: 

3475: We now prove analogous results for the other model of spanning trees, the

3476: minimal ones.

3477: Given a finite connected graph, $\gh$, and independent uniform $[0, 1]$

3478: random variables on its edges, the spanning tree that minimizes the sum of

3479: the labels of its edges has a distribution denoted $\MST(\gh)$, the {\bf

3480: minimal spanning tree} measure on $\gh$.

3481: If $\gh$ is infinite, there are two analogous measures, as in the uniform

3482: case. They can be defined by weak limits, but also directly (and by

3483: pointwise limits).

3484: Namely, given independent uniform $[0, 1]$ edge labels, remove all edges

3485: whose label is the largest in some cycle containing that edge.

3486: The remaining edges form the {\bf free minimal spanning forest},

3487: $\FMSF$.

3488: If one also removes the edges $e$ both of whose endpoints belong to infinite

3489: paths of edges that are all labeled smaller than $e$ is, then the resulting

3490: forest is called the {\bf wired minimal spanning forest}, $\WMSF$.

3491: 

3492: The following is analogous to \ref p.ust-limit/ above and is proved

3493: similarly to it and part of Theorem 3.12 of \LPSmsf, using \ref t.death/

3494: below. Parts of it were also proved by \ref b.AS:obj/.

3495: %It extends results of \LPSmsf, as well.

3496: %As noted in \ref b.AS:obj/, it holds without the usual assumption that the

3497: %edge labels are independent uniform random variables, but with the

3498: %assumption merely that the labels are distinct.

3499: 

3500: \procl p.mst-limit \procname{MST Limits}

3501: If $\rtd$ is a unimodular probability measure on infinite networks in

3502: $\GG_*$, then $\expdeg\big(\WMSF(\rtd)\big) = 2$.

3503: If $\gh_n$ are finite connected networks whose random weak limit is $\rtd$,

3504: then $\MST(\gh_n) \cd \WMSF(\rtd)$.

3505: More generally, if $\rtd_n$ are unimodular probability measures on $\GG_*$

3506: with $\rtd_n \cd \rtd$, then $\WMSF(\rtd_n) \cd \WMSF(\rtd)$.

3507: \endprocl

3508: 

3509: The following extends a result of \LPSmsf, which in turn extends a result

3510: of \ref b.Alexander:MSF/, who proved this in fixed Euclidean lattices.

3511: Our proof follows slightly different lines.

3512: For information on when the hypothesis is satisfied, see \ref t.death/

3513: below.

3514: 

3515: \procl t.msf1end \procname{One End}

3516: If $\rtd$ is an extremal

3517: unimodular probability measure on infinite networks in

3518: $\GG_*$ and there is $\P_{\pc(\rtd)}$-a.s.\ no infinite cluster,

3519: then $\WMSF(\rtd)$-a.s., each tree has exactly one end.

3520: \endprocl

3521: 

3522: \proof

3523: By the first part of \ref p.mst-limit/ and by \ref t.deg2/, each tree has at

3524: most 2 ends $\WMSF(\rtd)$-a.s.

3525: Suppose that some tree has 2 ends with positive probability.

3526: A tree with precisely two ends has a {\bf trunk}, the unique bi-infinite

3527: simple path in the tree.

3528: By the definition of $\WMSF$, the labels on a trunk cannot have a maximum.

3529: By the Mass-Transport Principle, the limsup in one direction must equal the

3530: limsup in the other direction, since otherwise we could identify the one

3531: edge that has label larger than the average of the two limsups and is the

3532: last such edge in the direction from the larger limsup to the smaller

3533: limsup.

3534: Let $p$ be this common limsup.

3535: By the preceding, all the labels on the trunk are strictly less than $p$.

3536: The root belongs to the trunk with positive probability.

3537: Assume this happens.

3538: Then the root belongs to an infinite $p$-cluster (the one containing the

3539: trunk).

3540: Now invasion from the root will fill (the vertices of)

3541: this $p$-cluster and is part of the

3542: tree containing the root (see \LPSmsf), whence the tree contains (the

3543: vertices of) the entire $p$-cluster of the root.

3544: By \ref t.uniq/, the tree therefore also contains an infinite $p'$-cluster

3545: for every $p' \in (\pc(\rtd), p)$ if $p > \pc(\rtd)$.

3546: Let $x$ be a vertex in the tree that is in an infinite $p'$-cluster $C$ for

3547: $p' := (\pc(\rtd) + p)/2$.

3548: Now invasion from $x$ has a finite symmetric difference with

3549: invasion from $o$ (see \LPSmsf),

3550: invasion from $x$ does not leave $C$, and invasion from $\bp$ fills the

3551: trunk.

3552: It follows from the definition of $p$ that $p' \ge p$.

3553: That is, $p = \pc(\rtd)$.

3554: Therefore, there is $\P_{\pc(\rtd)}$-a.s.\ an infinite cluster.

3555: \Qed

3556: 

3557: 

3558: 

3559: \bsection{Amenability and Nonamenability}{s.amen}

3560: 

3561: Recall that a graph $\gh$ is (vertex) amenable

3562: iff there is a sequence of subsets $H_n\subset \vertex(\gh)$

3563: with

3564: $$

3565: \lim_{n \to\infty} {|\bdv H_n| \over |\verts(H_n)|} = 0\,,

3566: $$

3567: where $|\cbuldot|$ denotes cardinality.

3568: 

3569: Amenability, originally defined for groups, now appears in several areas of

3570: mathematics, including probability theory and ergodic theory.

3571: Its presence provides many tools one is used to from $\Z$ actions, yet its

3572: absence also provides a powerful threshold principle.

3573: There are many equivalent definitions of amenability.

3574: We choose one that is not standard, but is useful for our probabilistic

3575: purposes.

3576: We show that it is equivalent to other definitions.

3577: Then we shall illustrate its uses.

3578: 

3579: \procl d.amen

3580: Let $\projection : \marks \to \marks$ be the composition of a homeomorphism of

3581: $\marks$ with $\marks^2$ followed by the projection onto the first coordinate.

3582: If a rooted network $(\gh, \bp)$ is understood,

3583: then for a subset $\marks_0 \subseteq \marks$ and vertex $x$, the {\bf

3584: $\marks_0$-component of $x$} is the set of vertices that can be reached

3585: from $x$ by edges both of whose marks lie in $\marks_0$.

3586: Write $K(\marks_0)$ for the $\marks_0$-component of the root.

3587: For a probability measure $\rtd$ on rooted graphs, denote by $\fdom(\rtd)$ the

3588: class of percolations on $\rtd$ that have only finite components.

3589: That is, $\fdom(\rtd)$ consists of pairs $(\nu, \marks_0)$ such that $\nu$ is

3590: a unimodular probability measure on $\GG_*$, $\marks_0 \subseteq \marks$ is

3591: Borel, $\rtd = \nu \circ \projection^{-1}$, and

3592: $K(\marks_0)$ is finite $\nu$-a.s.

3593: (By \ref l.unmark/, all $\marks_0$-components are then finite $\nu$-a.s.)

3594: For $\marks_0 \subseteq \marks$ and $x \in \vertex(\gh)$, write

3595: $$

3596: \ct(x, \marks_0) := | \{y \in \vertex(\gh) \st (x, y) \in \edges(\gh),

3597: \hbox{ some edge mark of } (x, y) \hbox{ is }\notin \marks_0 \} |

3598: \,.

3599: $$

3600: For $K \subset \vertex(\gh)$, define

3601: $$

3602: \ct(K, \marks_0) :=

3603: \sum_{x \in K} \ct(x, \marks_0)

3604: \,.

3605: $$

3606: Define

3607: $$

3608: \isoe(\rtd)

3609: :=

3610: \inf \left\{ \int {\ct\big(K(\marks_0), \marks_0\big) \over |K(\marks_0)|}

3611: d\rtd'(\gh, \bp)

3612: \st (\rtd', \marks_0) \in \fdom(\rtd) \right\}

3613: \,.

3614: $$

3615: Call $\rtd$ {\bf amenable} if $\isoe(\rtd) = 0$.

3616: Define 

3617: $$

3618: \expdeg(\rtd', \marks_0)

3619: :=

3620: %\int |\{x \in \vertex(\gh) \st (\bp, x) \in \edges(\gh) \hbox{ and both

3621: %edge marks of $(\bp, x)$ lie in } \marks_0\}| \,d\rtd'(\gh, \bp)

3622: \int \big[\deg_\gh(\bp) - n(\bp, \marks_0)\big] \,d\rtd'(\gh, \bp)

3623: \,,

3624: $$

3625: the expected degree in the $\marks_0$-component of the root,

3626: and

3627: $$

3628: \alpha(\rtd)

3629: :=

3630: \sup \big\{ \expdeg(\rtd', \marks_0) \st (\rtd', \marks_0) \in \fdom(\rtd)

3631: \big\}

3632: \,.

3633: $$

3634: Of course, neither $\isoe(\rtd)$ nor $\alpha(\rtd)$ depends on the choice

3635: of homeomorphism in $\projection$.

3636: Furthermore, these quantities depend only on the probability measure on

3637: the underlying graphs of the networks, not on the marks.

3638: \endprocl

3639: 

3640: This definition of amenability is justified in three ways: It agrees with

3641: the usual definition of amenability for fixed unimodular quasi-transitive

3642: graphs by Theorems 5.1 and 5.3 of \BLPSgip; it agrees with the usual notion

3643: of amenability for equivalence relations by \ref t.many/ below; and it

3644: allows us to extend to non-amenable unimodular random rooted graphs many

3645: theorems that are known for non-amenable unimodular fixed graphs, as we

3646: shall see.

3647: 

3648: We say that a graph $\gh$ is {\bf anchored amenable}

3649: if there is a sequence of subsets $H_n\subset \vertex(\gh)$

3650: such that $\bigcap_n H_n \ne \emptyset$, each $H_n$ induces a connected

3651: subgraph of $\gh$, and

3652: $$

3653: \lim_{n \to\infty} {|\bdv H_n| \over |\verts(H_n)|} = 0\,.

3654: $$

3655: The relationship between amenability of $\rtd$ and amenability or anchored

3656: amenability of $\rtd$-a.e.\ graph is as follows.

3657: The first clearly implies the third, which implies the second, but the third

3658: does not imply the first.

3659: Indeed, take a 3-regular tree and

3660: randomly subdivide its edges by a number of vertices whose

3661: distribution does not have a finite

3662: exponential tail. \ref b.ChenPeres/ show that the result is anchored

3663: amenable a.s.

3664: However, there is an appropriate unimodular version if the subdividing

3665: distribution has finite mean (see Example 2.4.4 of \ref

3666: b.Kaim:harmonic/ or \ref x.stretched/ below),

3667: and it is non-amenable by \ref c.amentrees/ below.

3668: 

3669: In order to work with this definition, we shall need some easy facts.

3670: 

3671: \procl l.bdryo

3672: If $\rtd$ is a unimodular probability measure on $\GG_*$ and $(\rtd',

3673: \marks_0) \in \fdom(\rtd)$, then

3674: $$

3675: \int {\ct\big(K(\marks_0), \marks_0\big) \over |K(\marks_0)|} d\rtd'

3676: =

3677: \int \ct(\bp, \marks_0) \, d\rtd'

3678: \,.

3679: $$

3680: \endprocl

3681: 

3682: \proof

3683: Let $K_x$ be the $\marks_0$-component of $x$.

3684: Let each vertex $x$ send mass $\ct(y, \marks_0)/|K_x|$ to each $y \in K_x$.

3685: Then the left-hand side is the expected mass sent from the root and the

3686: right-hand side is the expected mass received by the root.

3687: \Qed

3688: 

3689: 

3690: %\procl l.finiteavg

3691: %If $\rtd$ is a unimodular probability measure on $\GG_*$ and $\rtd' \in

3692: %\fdom(\rtd)$, then

3693: %$$

3694: %\expdeg(\rtd)

3695: %=

3696: %\int {\sum_{x \in K(\bp)} \deg(x) \over |K(\bp)|} d\rtd'(\gh, \bp)

3697: %\,,

3698: %\label e.alpha-alt

3699: %$$

3700: %where $K(\bp)$ is the component of the root in $(\gh, \bp)$ and degree is

3701: %measured in $\rtd$. \msnote{Fix wording.}

3702: %\endprocl

3703: %

3704: %\proof

3705: %This is a consequence of the Mass-Transport Principle:

3706: %each vertex $x$ distributes mass $\deg(x)$ equally among the vertices in its

3707: %$\rtd'$-component $K$.

3708: %The expected total mass sent out from the root is $\expdeg(\rtd)$, while the

3709: %expected total mass received at the root is the right-hand side of

3710: %\ref e.alpha-alt/.

3711: %\Qed

3712: 

3713: 

3714: \procl p.alpha

3715: If $\rtd$ is a unimodular probability measure on $\GG_*$, then

3716: $$

3717: \isoe(\rtd) + \alpha(\rtd) = \expdeg(\rtd)

3718: \,.

3719: \label e.alpha

3720: $$

3721: Therefore, if\/ $\expdeg(\rtd) < \infty$, then $\rtd$ is amenable iff

3722: $\alpha(\rtd) = \expdeg(\rtd)$.

3723: \endprocl

3724: 

3725: \proof

3726: %In the definition of $\alpha(\rtd)$, replace $\expdeg(\rtd')$ with its equivalent

3727: %expression from \ref e.alpha-alt/ (applied to $\rtd'$).

3728: %Adding this to $\isoe(\rtd)$,

3729: %the result follows by another application of

3730: %\ref e.alpha-alt/ (applied to $\rtd$).

3731: This is obvious from \ref l.bdryo/.

3732: \Qed

3733: 

3734: 

3735: \procl l.disint

3736: Let $\rtd, \nu \in \uni$ with $\nu$ a percolation on $\rtd$, that is, there

3737: is some Borel $\psi : \marks \to \marks$ such that $\rtd = \nu \circ

3738: \psi^{-1}$.

3739: Let $\kappa$ be a regular conditional probability measure of $\rtd$

3740: with respect to the $\sigma$-field generated by $\psi$, i.e., a

3741: disintegration of $\nu$ with respect to $\psi$, with $\kappa_{(\gh, \bp)}$

3742: being the probability measure on the fiber over $(\gh, \bp)$.

3743: Let $h : \gtwo \to [0, 1]$ be Borel and symmetric: $h(G, x, y) = h(G, y,

3744: x)$.

3745: Define $k(G, x, y) := \int h(G, x, y) \,d\kappa_{(\gh, x)}$.

3746: Then there is a symmetric Borel $\lambda$ such that for $\rtd$-a.e.\ $(\gh,

3747: \bp)$ and all $x \in \vertex(\gh)$, we have $k(\gh, \bp, x) = \lambda(\gh,

3748: \bp, x)$.

3749: \endprocl

3750: 

3751: %This could also be proved by showing it on finite networks and using \ref

3752: %t.TII/, as suggested by B\'alint Virag.

3753: %It is false without unimodularity.

3754: 

3755: \proof

3756: It suffices to show that for all $f : \gtwo \to [0, 1]$, we have 

3757: $$

3758: \int \sum_{x \in \vertex(\gh)} k(\gh, \bp, x) f(\gh, \bp, x) \,d\rtd(\gh,

3759: \bp)

3760: =

3761: \int \sum_{x \in \vertex(\gh)} k(\gh, x, \bp) f(\gh, \bp, x) \,d\rtd(\gh,

3762: \bp)

3763: \,,

3764: $$

3765: since this shows symmetry of $k$ a.e.\ with respect to the left measure

3766: $\rtd_{\rm L}$.

3767: To see that this equation holds, observe that 

3768: $$\eqaln{

3769: \int \sum_{x \in \vertex(\gh)} k(\gh, \bp, x) f(\gh, \bp, x) &\,d\rtd(\gh,

3770: \bp)

3771: =

3772: \int \int \sum_{x \in \vertex(\gh)} h(\gh, \bp, x) f(\gh, \bp, x)

3773: \,d\kappa_{(\gh, \bp)} \,d\rtd(\gh, \bp)

3774: \cr&=

3775: \int \int \sum_{x \in \vertex(\gh)} h(\gh, x, \bp) f(\gh, x, \bp)

3776: \,d\kappa_{(\gh, \bp)} \,d\rtd(\gh, \bp)

3777: \cr&\hskip1in\hbox{[by the Mass-Transport Principle for $\nu$]}

3778: \cr&=

3779: \int \int \sum_{x \in \vertex(\gh)} h(\gh, \bp, x) f(\gh, \bp, x)

3780: \,d\kappa_{(\gh, x)} \,d\rtd(\gh, \bp)

3781: \cr&\hskip1in\hbox{[by the Mass-Transport Principle for $\rtd$]}

3782: \cr&=

3783: \int \int \sum_{x \in \vertex(\gh)} h(\gh, x, \bp) f(\gh, \bp, x)

3784: \,d\kappa_{(\gh, x)} \,d\rtd(\gh, \bp)

3785: \cr&\hskip1in\hbox{[by symmetry of $h$]}

3786: \cr&=

3787: \int \sum_{x \in \vertex(\gh)} k(\gh, x, \bp) f(\gh, \bp, x) \,d\rtd(\gh,

3788: \bp)

3789: \,.

3790: \Qed

3791: }$$

3792: 

3793: 

3794: We now prove some properties that are equivalent to amenability.

3795: Most of these are standard in the context of equivalence relations.

3796: With appropriate modifications, these equivalences hold with a weakening of

3797: the assumption of unimodularity.

3798: They are essentially due to \ref b.CFW/ and \ref b.Kaim:amen/, although

3799: (ii) seems to be new.

3800: 

3801: \procl t.many \procname{Amenability Criteria}

3802: Let $\rtd$ be a unimodular probability measure on $\GG_*$ with $\expdeg(\rtd)

3803: < \infty$.

3804: The following are equivalent:

3805: \beginitems

3806: \itemrm{(i)} $\rtd$ is amenable;

3807: \itemrm{(ii)} there is

3808: a sequence of Borel functions $\lambda_n : \gtwo \to [0, 1]$

3809: such that for all $(\gh, x, y) \in \gtwo$ and all $n$,

3810: we have

3811: $$

3812: \lambda_n(\gh, x, y) =

3813: \lambda_n(\gh, y, x)

3814: \label e.symm

3815: $$

3816: and for $\rtd$-a.e.\ $(\gh, \bp)$, we have

3817: $$

3818: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x) = 1

3819: \label e.prmsrae

3820: $$

3821: and

3822: $$

3823: \lim_{n \to\infty}

3824: \sum_{x \in \vertex(\gh)}\sum_{y \sim x}

3825: \left|\lambda_n(\gh, \bp, x) -

3826: \lambda_n(\gh, \bp, y)\right|

3827: = 0

3828: \,;

3829: \label e.asymsame

3830: $$

3831: \itemrm{(iii)} there is

3832: a sequence of Borel functions $\lambda_n : \gtwo \to [0, 1]$

3833: such that for $\rtd$-a.e.\ $(\gh, \bp)$,

3834: $$

3835: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x) = 1

3836: $$

3837: and

3838: $$

3839: \lim_{n \to\infty}

3840: \sum_{y \sim \bp} \sum_{x \in \vertex(\gh)}

3841: \left|\lambda_n(\gh, \bp, x) -

3842: \lambda_n(\gh, y, x)\right|

3843: = 0

3844: \,;

3845: $$

3846: \itemrm{(iv)}

3847: $\rtd$ is hyperfinite, meaning that there

3848: is a unimodular measure $\nu$ on $\GG_*$, an increasing sequence of 

3849: Borel subsets $\marks_n \subseteq \marks$, and a Borel

3850: function $\psi: \marks \to \marks$ such that if $G$ denotes a

3851: network with law $\nu$ and $G_n$ the subnetwork consisting of those edges

3852: both of whose edge marks lie in $\marks_n$, then $\psi(G)$ has law $\rtd$,

3853: all components of $G_n$ are finite, and $\bigcup_n \marks_n = \marks$.

3854: \enditems

3855: \endprocl

3856: 

3857: \proof

3858: Assume from now on that $\rtd$ is carried by networks with distinct marks.

3859: We shall use the following construction.

3860: Suppose that $(\nu, \marks_0) \in \fdom(\rtd)$.

3861: Let $\kappa$ be a regular conditional probability measure of

3862: $\rtd$ with respect to the $\sigma$-field generated by $\projection$.

3863: By \ref l.disint/, there is a Borel symmetric 

3864: $\lambda : \gtwo \to [0, 1]$ such that

3865: for $\rtd$-a.e.\ $(\gh, \bp)$ and $x \in \vertex(\gh)$, we have

3866: $$

3867: \lambda(\gh, \bp, x)

3868: =

3869: \int \II{x \in K(\marks_0)}/|K(\marks_0)|

3870: \,d\kappa_{(\gh, \bp)}

3871: \,.

3872: \label e.connect

3873: $$

3874: Clearly,

3875: $$

3876: \sum_{x \in \vertex(\gh)} \lambda(\gh, \bp, x) = 1

3877: \label e.prmsr

3878: $$

3879: for $\rtd$-a.e.\ $(\gh, \bp)$.

3880: For $\rtd$-a.e.\ $(\gh, \bp)$, we have

3881: $$\eqaln{

3882: \sum_{x \in \vertex(\gh)} \sum_{y \sim x}

3883: |\lambda(\gh, \bp, x) &- \lambda(\gh, \bp, y)|

3884: \cr&\le

3885: \sum_{x \in \vertex(\gh)} \sum_{y \sim x}

3886: \int |\II{x \in K(\marks_0)} -

3887: \II{y \in K(\marks_0)}|/|K(\marks_0)|

3888: \,d\kappa_{(\gh, \bp)}

3889: \cr&\le

3890: \int {2\ct\big(K(\marks_0), \marks_0\big) \over |K(\marks_0)|}

3891: \,d\kappa_{(\gh, \bp)}

3892: \,.

3893: \label e.isobd

3894: }$$

3895: 

3896: Now if (i) holds, then we may choose $(\rtd_n, \marks_n) \in \fdom(\rtd)$

3897: such that

3898: $$

3899: \int

3900: \sum_{n}

3901: {\ct\big(K(\marks_n), \marks_n\big) \over |K(\marks_n)|}

3902: \,d\rtd_{n} < \infty

3903: \,.

3904: $$

3905: Let $\kappa^{(n)}$ and

3906: $\lambda_n$ be as above (but for $\rtd_n$).

3907: Then by \ref e.isobd/, we have

3908: $$\eqaln{

3909: \sum_{n} \int

3910: \sum_{x \in \vertex(\gh)}\sum_{y \sim x}

3911: \big|\lambda_n(\gh, \bp, x) &-

3912: \lambda_n(\gh, \bp, y)\big|

3913: \,d\rtd(\gh, \bp)

3914: \cr&\le

3915: \sum_{n} \int

3916: \int {2\ct\big(K(\marks_n), \marks_n\big) \over |K(\marks_n)|}

3917: \,d\kappa^{(n)}_{(\gh, \bp)}

3918: \,d\rtd(\gh, \bp)

3919: \cr&=

3920: \sum_{n}

3921: \int {2\ct\big(K(\marks_n), \marks_n\big) \over |K(\marks_n)|}

3922: \,d\rtd_{n} < \infty

3923: \,,

3924: }$$

3925: which shows that (ii) holds.

3926: 

3927: Next, suppose that (ii) holds.

3928: Then by the bounded convergence theorem, we have

3929: $$

3930: \lim_{n \to\infty} \int

3931: \sum_{x \in \vertex(\gh)}\sum_{y \sim x}

3932: \left|\lambda_n(\gh, \bp, x) -

3933: \lambda_n(\gh, \bp, y)\right|

3934: \,d\rtd(\gh, \bp)

3935: = 0

3936: \,.

3937: $$

3938: On the other hand, the Mass-Transport Principle and \ref e.symm/

3939: show that this integral is the same as

3940: $$

3941: \int

3942: \sum_{y \sim \bp} \sum_{x \in \vertex(\gh)}

3943: \left|\lambda_n(\gh, \bp, x) -

3944: \lambda_n(\gh, y, x)\right|

3945: \,d\rtd(\gh, \bp)

3946: \,.

3947: $$

3948: Therefore, by taking a subsequence if necessary, we achieve (iii).

3949: 

3950: Next, suppose that (iii) holds.

3951: Then we may define a sequence similar to $\lambda_n$ on

3952: the corresponding equivalence relation (see \ref x.equivalence/),

3953: which implies that the equivalence

3954: relation is amenable by \ref b.Kaim:amen/, and hence hyperfinite by

3955: a theorem of \ref b.CFW/. (Another proof of the latter theorem was sketched

3956: by \ref b.Kaim:amen/, with more details

3957: given by \ref b.KechrisMiller/).

3958: Translating the definition of hyperfinite equivalence relation to rooted

3959: networks gives (iv).

3960: 

3961: Finally, that (iv) implies (i) is an immediate consequence of Lebesgue's

3962: Dominated Convergence Theorem, our assumption that $\expdeg(\rtd) <

3963: \infty$, and \ref l.bdryo/.

3964: \Qed

3965: 

3966: Now we show how to produce unimodular networks from non-unimodular ones on

3967: amenable measures, just as we can produce invariant measures from

3968: non-invariant ones on amenable groups.

3969: We illustrate in the context of couplings.

3970: 

3971: \procl p.amen-average \procname{Coupling From Amenability}

3972: Let $\rln \subseteq \marks \times \marks$ be a closed set.

3973: If $\rtd_1, \rtd_2 \in \uni$ are amenable and $\rtd_1$ is

3974: $\rln$-related to $\rtd_2$, then there is a unimodular $\rln$-coupling of

3975: $\rtd_1$ to $\rtd_2$.

3976: \endprocl

3977: 

3978: \proof

3979: Let $\nu$ be an $\rln$-coupling of $\rtd_1$ to $\rtd_2$.

3980: Let $\lambda_n$ be as in \ref t.many/(ii) for $(\rtd_1 + \rtd_2)/2$.

3981: Define the measures $\nu_n$ by

3982: $$

3983: \nu_n(\ev B)

3984: :=

3985: \int

3986: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x) \II{(\gh, x) \in \ev B}

3987: \,d\nu(\gh, \bp)

3988: $$

3989: for Borel $\ev B \subseteq \GG_*$ (with mark space $\marks \times \marks$).

3990: Then $\nu_n$ is a probability measure by \ref e.prmsrae/.

3991: Since $\nu$ is carried by networks all of whose marks are in $\rln$, so is

3992: $\nu_n$.

3993: If $\ev B$ is an event that

3994: specifies only the first coordinates of the marks, i.e., $\ev B = \ev

3995: B_1 \times 2^\marks$ for some $\ev B_1$, then

3996: $$\eqaln{

3997: \nu_n(\ev B)

3998: &=

3999: \int

4000: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x) \II{(\gh, x) \in \ev B}

4001: \,d\nu(\gh, \bp)

4002: \cr&=

4003: \int

4004: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x) \II{(\gh, x) \in \ev B_1}

4005: \,d\mu_1(\gh, \bp)

4006: \cr&=

4007: \int

4008: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, x, \bp) \II{(\gh, \bp) \in \ev B_1}

4009: \,d\mu_1(\gh, \bp)

4010: \cr&=

4011: \int

4012: \II{(\gh, \bp) \in \ev B_1}

4013: \,d\mu_1(\gh, \bp)

4014: \cr&=

4015: \mu_1(\ev B)

4016: }$$

4017: by the Mass-Transport Principle, \ref e.symm/, and \ref e.prmsrae/.

4018: Likewise, if $\ev B$ is an event that

4019: specifies only the second coordinates of the marks, then $\nu_n(\ev B) =

4020: \mu_2(\ev B)$.

4021: Thus, $\nu_n$ is an

4022: $\rln$-coupling of $\rtd_1$ to $\rtd_2$.

4023: 

4024: Now by definition of $\nu_n$, we have

4025: $$

4026: \int f(\gh, \bp)\,d\nu_n(\gh, \bp)

4027: =

4028: \int

4029: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x) f(\gh, x)

4030: \,d\nu(\gh, \bp)

4031: $$

4032: for every Borel $f : \GG_* \to [0, \infty]$.

4033: Therefore, for every Borel $h : \gtwo \to [0, 1]$ with $h(\gh, x, y) = 0$

4034: unless $x \sim y$, we have

4035: $$\eqaln{

4036: \int \sum_{y \in \vertex(\gh)} h(\gh, \bp, y) \,d\nu_n(\gh, \bp)

4037: &=

4038: \int \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x)

4039: \sum_{y \in \vertex(\gh)} h(\gh, x, y) \,d\nu_n(\gh, \bp)

4040: \cr&=

4041: \int \sum_{y \in \vertex(\gh)}

4042: \sum_{x \in \vertex(\gh)} 

4043: h(\gh, x, y) 

4044: \lambda_n(\gh, \bp, x)

4045: \,d\nu(\gh, \bp)

4046: }$$

4047: and

4048: $$\eqaln{

4049: \int \sum_{y \in \vertex(\gh)} h(\gh, y, \bp) \,d\nu_n(\gh, \bp)

4050: &=

4051: \int \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x)

4052: \sum_{y \in \vertex(\gh)} h(\gh, y, x) \,d\nu_n(\gh, \bp)

4053: \cr&=

4054: \int \sum_{y \in \vertex(\gh)}

4055: \sum_{x \in \vertex(\gh)} 

4056: h(\gh, x, y) 

4057: \lambda_n(\gh, \bp, y)

4058: \,d\nu(\gh, \bp)

4059: \,,

4060: }$$

4061: where, in the last step, we have interchanged $x$ and $y$.

4062: Therefore,

4063: $$\eqaln{

4064: \bigg| 

4065: \int \sum_{y \in \vertex(\gh)} h(\gh, \bp, y) \,d\nu_n(\gh, \bp)

4066: &- 

4067: \int \sum_{y \in \vertex(\gh)} h(\gh, y, \bp) \,d\nu_n(\gh, \bp)

4068: \bigg|

4069: \cr&\le

4070: \int \sum_{y \in \vertex(\gh)}

4071: \sum_{x \sim y}

4072: \left|

4073: \lambda_n(\gh, \bp, x)

4074: -

4075: \lambda_n(\gh, \bp, y)

4076: \right|

4077: \,d\nu(\gh, \bp)

4078: \cr&=

4079: \int \sum_{y \in \vertex(\gh)}

4080: \sum_{x \sim y}

4081: \left|

4082: \lambda_n(\gh, \bp, x)

4083: -

4084: \lambda_n(\gh, \bp, y)

4085: \right|

4086: \,d\rtd_1(\gh, \bp)

4087: \,,

4088: }$$

4089: which tends to 0 by \ref e.asymsame/.

4090: Thus, any limit point of $\nu_n$ is involution invariant and, since $\rln$ is

4091: closed, is an $\rln$-coupling of $\rtd_1$ to $\rtd_2$.

4092: \Qed

4093: 

4094: \comment{

4095: \proof

4096: Let $\nu$ be an $\rln$-coupling of $\rtd_1$ to $\rtd_2$.

4097: Let $\lambda_n$ be as in \ref t.many/ for, say, $(\mu_1+\mu_2)/2$.

4098: Write

4099: $$

4100: k(\gh, \bp) :=

4101: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, x, \bp)

4102: \,.

4103: $$

4104: Define the measures $\nu_n$ by

4105: $$

4106: \nu_n(\ev B)

4107: :=

4108: \int

4109: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x) \II{(\gh, x) \in \ev B}

4110: k(\gh, x)^{-1} \,d\nu(\gh, \bp)

4111: $$

4112: for Borel $\ev B \subseteq \GG_*$ (with mark space $\marks \times \marks$).

4113: Then $\nu_n$ is a probability measure since

4114: $$\eqaln{

4115: \int

4116: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x)

4117: k(\gh, x)^{-1} \,d\nu(\gh, \bp)

4118: &=

4119: \int

4120: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, x, \bp)

4121: k(\gh, \bp)^{-1} \,d\nu(\gh, \bp)

4122: \cr&=

4123: \int d\nu(\gh, \bp)

4124: = 1

4125: }$$

4126: by the Mass-Transport Principle.

4127: %Note that

4128: %$$

4129: %\int

4130: %\sum_{x \in \vertex(\gh)} \lambda_n(\gh, x, \bp) \,d\mu_1(\gh, \bp)

4131: %=

4132: %\int

4133: %\sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x) \,d\mu_1(\gh, \bp)

4134: %=

4135: %1

4136: %$$

4137: %by the Mass-Transport Principle.

4138: Since $\nu$ is carried by networks all of whose marks are in $\rln$, so is

4139: $\nu_n$.

4140: If $\ev B$ is an event that

4141: specifies only the first coordinates of the marks, i.e., $\ev B = \ev

4142: B_1 \times 2^\marks$ for some $\ev B_1$, then

4143: $$\eqaln{

4144: \nu_n(\ev B)

4145: &=

4146: \int

4147: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x) \II{(\gh, x) \in \ev B}

4148: k(\gh, x)^{-1} \,d\kappa(\gh, \bp)

4149: \cr&=

4150: \int

4151: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x) \II{(\gh, x) \in \ev B_1}

4152: k(\gh, x)^{-1} \,d\mu_1(\gh, \bp)

4153: \cr&=

4154: \int

4155: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, x, \bp) \II{(\gh, \bp) \in \ev B_1}

4156: k(\gh, \bp)^{-1} \,d\mu_1(\gh, \bp)

4157: \cr&=

4158: \int

4159: \II{(\gh, \bp) \in \ev B_1}

4160: \,d\mu_1(\gh, \bp)

4161: \cr&=

4162: \mu_1(\ev B)

4163: }$$

4164: by the Mass-Transport Principle.

4165: Likewise, if $\ev B$ is an event that

4166: specifies only the second coordinates of the marks, then $\nu_n(\ev B) =

4167: \mu_2(\ev B)$.

4168: Thus, $\nu_n$ is an

4169: $\rln$-coupling of $\rtd_1$ to $\rtd_2$.

4170: Now by definition of $\nu_n$, we have

4171: $$

4172: \int f(\gh, \bp)\,d\nu_n(\gh, \bp)

4173: =

4174: \int

4175: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x) f(\gh, x)

4176: k(\gh, x)^{-1} \,d\nu(\gh, \bp)

4177: $$

4178: for every Borel $f : \GG_* \to [0, \infty]$.

4179: Therefore,

4180: \msnote{The below does not calculate in sufficient generality.}

4181: $$\eqaln{

4182: \bigg| \int \sum_{y \sim \bp} \II{(\gh, \bp) \in \ev B} \,d\nu_n(\gh, \bp)

4183: &- \int \sum_{y \sim \bp} \II{(\gh, y) \in \ev B} \,d\nu_n(\gh, \bp)\bigg|

4184: \cr&\quad=

4185: \bigg|

4186: \int \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x)

4187: \sum_{y \sim x} \II{(\gh, x) \in \ev B} k(\gh, x)^{-1} \,d\nu(\gh, \bp)

4188: \cr&\qquad-

4189: \sum_{x \in \vertex(\gh)} \lambda_n(\gh, \bp, x)

4190: \sum_{y \sim x} \II{(\gh, y) \in \ev B} k(\gh, x)^{-1} \,d\nu(\gh, \bp)

4191: \bigg|

4192: \cr&\quad\le

4193: ...

4194: }$$

4195: \msnote{Now apply CBS and use that the square of a number in

4196: $[0, 1]$ is at most that number.}

4197: Thus, any limit point of $\nu_n$ is unimodular and, since $\rln$ is

4198: closed, is an $\rln$-coupling of $\rtd_1$ to $\rtd_2$.

4199: \Qed

4200: }%

4201: 

4202: 

4203: 

4204: \procl p.recurrent \procname{Recurrence Implies Amenability}

4205: If $\rtd \in \uni$ and simple random walk is $\rtd$-a.s.\ recurrent,

4206: then $\rtd$ is amenable.

4207: \endprocl

4208: 

4209: \proof

4210: Consider the ``lazy" simple random walk that moves nowhere with probability

4211: 1/2 and otherwise moves to a random neighbor, like simple random walk.

4212: It is recurrent by hypothesis and aperiodic by construction.

4213: Let $\lambda_n(\gh, \bp, x)$ be the probability that lazy simple random

4214: walk on $\gh$ starting from $\bp$ will be at $x$ at time $n$.

4215: By \ref b.Orey/ and recurrence, the functions $\lambda_n$

4216: satisfy property (iii) of \ref t.many/, whence $\rtd$ is amenable.

4217: \Qed

4218: 

4219: 

4220: The following is proved similarly to Remark 6.2 of \BLPSusf.

4221: 

4222: \procl p.amenforest \procname{Forests in Amenable Networks}

4223: If $\rtd \in \uni$ is amenable and $\marks_0 \subseteq \marks$ is such that

4224: the $\marks_0$-open subgraph $\fo$ of $(\gh, \bp)$ is a forest $\rtd$-a.s.,

4225: then the expected degree of $\bp$ in $\fo$ is at most 2.

4226: \endprocl

4227: 

4228: The following extends Theorem 5.3 of \BLPSgip.

4229: In the following, we say that $\P$ is a {\bf percolation} on $\rtd$ that

4230: {\bf gives} subgraphs $\A$ a.s.\ if there is a Borel

4231: function $\psi: \marks \to \marks$ such that $\rtd = \P \circ \psi^{-1}$

4232: and there is a Borel subset $\marks_0 \subseteq \marks$

4233: such that if $G(\marks_0)$ denotes the $\marks_0$-open subnetwork of $G$,

4234: then $\psi\big(G(\marks_0)\big) \in \A$ for $\P$-a.e.\ $G$.

4235: 

4236: \procl t.treechar

4237: Let $\rtd \in \uni$ with $\expdeg(\rtd) < \infty$. The following are equivalent:

4238: \beginitems

4239: \itemrm{(i)} $\rtd$ is amenable;

4240: \itemrm{(ii)} there is a percolation $\P$ on $\rtd$ that gives

4241: spanning trees with at most 2 ends a.s.;

4242: \itemrm{(iii)} there is a percolation $\P$ on $\rtd$ that gives

4243: non-empty connected subgraphs $\omega$ that satisfy $\pc(\omega) = 1$ a.s.

4244: \enditems

4245: \endprocl

4246: 

4247: \proof

4248: The proof that (i) implies (ii) is done as for Theorem 5.3 of \BLPSgip, but

4249: uses Propositions \briefref p.amenforest/ and \briefref p.limittree/.

4250: That (ii) implies (iii) is obvious.

4251: The proof that (iii) implies (i) follows the first part of the proof of

4252: Theorem 1.1 in \ref b.BLPS:crit/.

4253: \Qed

4254: 

4255: \procl c.amentrees \procname{Amenable Trees}

4256: A unimodular probability measure $\rtd$ on infinite

4257: rooted trees is amenable iff $\expdeg(\rtd) = 2$ iff $\rtd$-a.s.\ $\gh$ has

4258: 1 or 2 ends.

4259: \endprocl

4260: 

4261: \proof

4262: Combine \ref t.treechar/ with \ref t.deg2/.

4263: \comment{

4264: Since the average degree in any finite tree is less than 2, we have

4265: $\expdeg(\rtd') < 2$ for all $\rtd' \in \fdom(\rtd)$, and so $\alpha(\rtd) \le 2$.

4266: Therefore, $\expdeg(\rtd) > 2$ implies non-amenability by \ref p.alpha/.

4267: On the other hand, if $\expdeg(\rtd) = 2$, then by \ref t.deg2/, $\rtd$ is

4268: concentrated on trees with at most 2 ends, whence on trees with $\pc = 1$.

4269: Therefore, Bernoulli($p$) percolation on $\rtd$ produces only finite

4270: components and for $p$ close to 1, produces a $\rtd' \in \fdom(\rtd)$ with

4271: $\expdeg(\rtd')$ close to 2.

4272: Thus, $\alpha(\rtd) = 2$ and we have amenability by \ref p.alpha/.

4273: The last equivalence is from \ref t.deg2/.

4274: }%

4275: \Qed

4276: 

4277: 

4278: 

4279: The next result was proved for non-amenable unimodular transitive graphs in

4280: \BLPSgip\ with a more direct proof in \ref b.BLPS:crit/. This extension is

4281: proved similarly.

4282: Presumably, the hypothesis that $\rtd$ is non-amenable can be replaced by the

4283: assumption that $\pc(\rtd) < 1$. (This is a major open conjecture for

4284: quasi-transitive graphs.)

4285: 

4286: \procl t.death \procname{Critical Percolation}

4287: Let $\rtd$ be an extremal unimodular non-amen\-a\-ble

4288: probability measure on $\GG_*$

4289: with $\expdeg(\rtd) < \infty$.

4290: There is $\P_{\pc(\rtd)}$-a.s.\ no infinite cluster.

4291: \endprocl

4292: 

4293: %The finiteness condition is needed for ruling out 1 component by knowing

4294: %that P(edge open) \to 1 implies degree of percolation \to degree of graph.

4295: % I initially required $\rtd$ to be concentrated on non-amenable graphs,

4296: % but I don't see why this restriction is needed.

4297: 

4298: This can be interpreted for finite graphs as follows.

4299: Suppose that $\gh_n$ are finite connected graphs with bounded average

4300: degree whose random weak limit is

4301: extremal and non-amenable with critical value $\pc$.

4302: Consider Bernoulli($\pc$) percolation on $\gh_n$.

4303: Let $\alpha_n(\ell)$ be the random variable giving

4304: the proportion of vertices of $\gh_n$ that belong to simple open paths of

4305: length at least $\ell$.

4306: Then $\lim_{\ell \to\infty} \lim_{n \to\infty} \alpha_n(\ell) = 0$ in

4307: probability.

4308: 

4309: 

4310: The following theorem is proved similarly to \ref b.BS:hp/, as extended by

4311: \ref b.LP:book/, and by using \ref x.dual/.

4312: 

4313: \procl t.planarperc \procname{Planar Percolation}

4314: Let $\rtd \in \uni$ be extremal, non-amenable,

4315: and carried by plane graphs with one end and bounded degree.

4316: Then $0 < \pc(\rtd) < \pu(\rtd) < 1$ and Bernoulli($\pu(\rtd)$) percolation

4317: on $\rtd$ has a unique infinite cluster a.s.

4318: \endprocl

4319: 

4320: 

4321: 

4322: The following extends Theorems 3.1, 3.2, 3.5, 3.6, and 3.10 of \ref

4323: b.BLS:pert/ and is proved similarly.

4324: If $\P$ is a percolation on $\rtd$ with $\P \circ \psi^{-1} = \rtd$ and

4325: with $\marks_0 \subseteq \marks$

4326: defining the open subgraphs, we call $\P'$ a {\bf subpercolation} on $\P$

4327: that {\bf gives} subgraphs in $\A'$ {\bf with positive probability} if there

4328: is a Borel

4329: function $\psi': \marks \to \marks$ such that $\P = \P' \circ \psi'^{-1}$

4330: and there is a Borel subset $\marks_1 \subseteq \marks$ such that if

4331: $G(\marks_1)$ denotes the $\marks_1$-open subnetwork of $G$, then

4332: $\P'\left[\psi\Big(\psi'\big(G(\marks_1)\big)\big(\marks_0\big)\Big) \in

4333: \A'\right] > 0$.

4334: 

4335: \procl t.sub-iso \procname{Non-Amenable Subgraphs}

4336: Let $\rtd$ be a unimodular probability measure on $\GG_*$ with finite

4337: expected degree and $\P$ be a

4338: percolation on $\rtd$ with open subgraph $\omega$.

4339: \beginitems

4340: \itemrm{(i)} If $h > 0$ and $\E[\deg_\omega \bp \mid \bp \in \omega] \ge

4341: \alpha(\rtd) + 2h$, then there is a subpercolation $\P'$ on $\P$ that gives

4342: a non-empty subgraph $\omega'$ with $\isoe(\omega') \ge h$ with positive

4343: $\P'$-probability.

4344: \itemrm{(ii)} If $\omega$ is a forest a.s., $h > 0$ and $\E[\deg_\omega \bp

4345: \mid \bp \in \omega] \ge 2 + 2h$, then there is a subpercolation $\P'$ on

4346: $\P$ that gives a non-empty subgraph with $\isoe \ge h$ with positive

4347: probability.

4348: \itemrm{(iii)} If $\rtd$ is non-amenable and $\omega$ has exactly one

4349: infinite cluster $\P$-a.s., then there is a subpercolation $\P'$ on $\P$

4350: that gives a non-empty subgraph $\omega'$ with $\isoe(\omega') > 0$ $\P'$-a.s.

4351: \itemrm{(iv)} If $\omega$ has components with at least three ends

4352: $\P$-a.s., then there is a subpercolation $\P'$ on $\P$ that gives a

4353: non-empty forest $\fo$ with $\isoe(\fo) > 0$ $\P'$-a.s.

4354: \itemrm{(v)} If $\rtd$ is concentrated on subgraphs with spectral radius

4355: less than 1 and $\omega$ has exactly one infinite cluster $\P$-a.s., then

4356: there is a subpercolation $\P'$ on $\P$ that gives a non-empty forest $\fo$

4357: with $\isoe(\fo) > 0$ $\P'$-a.s.

4358: \enditems

4359: \endprocl

4360: 

4361: The following extends a result of \ref b.Hag:rcust/ and is proved

4362: similarly to Corollary 6.3 of \BLPSusf.

4363: 

4364: \procl p.amen-for \procname{Amenability and Boundary Conditions}

4365: Let $\rtd$ be an amenable unimodular probability measure on $\GG_*$.

4366: Then $\FUSF(\rtd) = \WUSF(\rtd)$ and $\FMSF(\rtd) = \WMSF(\rtd)$.

4367: \endprocl

4368: 

4369: 

4370: We may now strengthen \ref p.recurrent/, despite the fact that not every

4371: graph is necessarily non-amenable $\rtd$-a.s.

4372: It extends Theorem 4.3 of \ref b.BLS:pert/ and is proved similarly, using

4373: \ref t.sub-iso/(iii) with $\P := \rtd$.

4374: 

4375: 

4376: \procl t.posspeed \procname{Positive Speed on Non-Amenable Graphs}

4377: If $\rtd \in \uni$ is non-amenable and concentrated on graphs with bounded

4378: degree,

4379: then the speed of simple random walk is positive $\rtd$-a.s.

4380: \endprocl

4381: 

4382: 

4383: 

4384: \bsection{Examples}{s.ex}

4385: 

4386: 

4387: We present here a variety of interesting examples of unimodular

4388: measures.

4389: 

4390: \procl x.renew \procname{Renewal Processes}

4391: Given a stationary (delayed) renewal process on $\Z$, let $\rtd$ be the law

4392: of $(\Z, 0)$ with the graph $\Z$, some fixed mark at renewals, and some

4393: other fixed mark elsewhere. Then $\rtd \in \uni$.

4394: \endprocl

4395: 

4396: \procl x.halfplane \procname{Half-Plane}

4397: Fix $d \ge 3$ and let $T$ be the $d$-regular tree.

4398: Let $\rtd_1$ be the random weak limit of balls of growing radii in $T$.

4399: Note that $\rtd_1$ is carried by trees with only one end.

4400: Let $\rtd_2$ be concentrated on the fixed graph $(\Z, 0)$.

4401: Now let $\rtd := \rtd_1 \itimes \rtd_2$.

4402: This is a unimodular version of the half-plane $\N \times \Z$.

4403: \endprocl

4404: 

4405: \procl x.cover

4406: For a rooted network $(\gh, \bp)$, its {\bf universal cover} is the rooted

4407: tree $(T, \bp) = T(\gh, \bp)$ formed as follows.

4408: The vertices of $T$ are the finite paths in $\gh$ that start at $\bp$ and

4409: do not backtrack.

4410: Two such vertices are joined by an edge in $T$ if one is an extension of

4411: the other by exactly one edge in $\gh$.

4412: The path with no edges consisting of just the vertex $\bp$ in $\gh$ is the

4413: root $\bp$ of $T$.

4414: There is a natural rooted graph homomorphism $\pi : (T, \bp) \to (\gh,

4415: \bp)$ (the cover map) that maps paths to their last point.

4416: Marks on $T$ are defined by lifting the marks on $\gh$ via $\pi$.

4417: It is clear that

4418: if $\rtd \in \uni$ and $\nu$ is the law of $T(\gh, \bp)$ when $(\gh, \bp)$

4419: has the law $\rtd$, then $\nu \in \uni$

4420: and $\expdeg(\rtd) = \expdeg(\nu)$.

4421: \endprocl

4422: 

4423: 

4424: \procl x.cluster

4425: Let $\P$ be a unimodular percolation on $\rtd$ that labels edges either

4426: open or closed.

4427: Let $\nu$ be the law of the open cluster of the root when the network is

4428: chosen according to $\P$ conditional on the root belonging to an infinite

4429: open cluster.

4430: Then $\nu$ is unimodular, as a direct verification of the definition shows.

4431: When $\rtd$ is concentrated on a fixed unimodular graph, this fact has been

4432: widely used in the study of percolation.

4433: \endprocl

4434: 

4435: 

4436: \procl x.Voronoi \procname{Tilings}

4437: Let $X$ be a Euclidean space or hyperbolic space (of constant curvature).

4438: Write $\gp$ for its isometry group.

4439: There is a Mass-Transport Principle for $X$ that says the following; see

4440: \ref b.BS:hp/ for a proof.

4441: Let $\rho$ be a positive Borel measure on $X \times X$ that is invariant

4442: under the diagonal action of $\gp$.

4443: Then there is a constant $c$ such that for all Borel $A \subset X$ of

4444: volume $|A| > 0$, we have $\rho(A \times X) = \rho(X \times A) = c |A|$.

4445: Suppose that $P$ is a (countable) point process in $X$ whose law is

4446: $\gp$-invariant.

4447: For example, Poisson point processes are $\gp$-invariant.

4448: One often considers graphs $\gh$ that are functions $\gh = \beta(P)$, where

4449: $\beta$ commutes with the action of $\gp$.

4450: For a few recent examples, see

4451: \ref b.BS:hp/, \ref b.HolPer:PP/, or \ref b.Timar:PP/.

4452: For instance, the 1-skeleton of the Voronoi tessellation corresponding to

4453: $P$ is such a graph.

4454: In general, we call such measures on graphs {\bf $\gp$-equivariant factors}

4455: of $P$.

4456: They are necessarily $\gp$-invariant.

4457: \par

4458: Another way that invariant measures on graphs embedded in $X$ occur is

4459: through (aperiodic) tilings of $X$.

4460: Again, one can take the 1-skeleton.

4461: An important tool for studying aperiodic tilings is a limit measure

4462: obtained from translates of a given tiling (in the Euclidean case), or,

4463: more generally, invariant measures on tilings with special properties; see,

4464: e.g., \ref b.Rob/, \ref b.Radin:survey/, \ref b.Solomyak/,

4465: \ref b.Radin:book/, or

4466: \ref b.BowRad:densest/ for some examples.

4467: \par

4468: Let $\nu$ be any $\gp$-invariant probability measure on graphs embedded in

4469: $X$.

4470: Fix a Borel set $A \subset X$ of positive finite volume.

4471: If $v(A) := \int |\vertex(\gh) \cap A| \,d\nu(\gh) < \infty$,

4472: then define $\rtd$ as follows.

4473: Choose $\gh$ with the law $\nu$ biased by $|\vertex(\gh) \cap A|$.

4474: Then choose the root $\bp$ of $\gh$

4475: uniformly among all vertices that belong to $A$.

4476: The law of the resulting graph $(\gh, \bp)$ is $\rtd$.

4477: We claim that $\rtd$ is unimodular and does not depend on $A$.

4478: In fact, $\rtd$ is the same as the Palm measure of $(\gh, \bp)$, except

4479: that $\rtd$ is a measure on isomorphism classes of graphs that

4480: does not involve any geometric embedding.

4481: To prove our claims, we first write $\rtd$ in symbols:

4482: $$

4483: \rtd(\A) :=

4484: v(A)^{-1} \int \sum_{\bp \in A} \II{(\gh, \bp) \in \A} \,d\nu(\gh)

4485: $$

4486: for Borel $\A \subset \GG_*$.

4487: Let $f : \gtwo \to [0, \infty]$ be Borel.

4488: Define

4489: $$

4490: \rho(B \times C) :=

4491: \int \sum_{x \in \vertex(\gh) \cap B} \sum_{y \in \vertex(\gh) \cap C}

4492: f(\gh, x, y)\,d\nu(\gh)

4493: $$

4494: for Borel $B, C \subseteq X$.  Since $\nu$ is invariant, $\rho$ is

4495: diagonally invariant.

4496: Therefore,

4497: $$\eqaln{

4498: \int \sum_{x \in \vertex(\gh)} f(\gh, \bp, x) \,d\rtd(\gh, \bp)

4499: &=

4500: v(A)^{-1} \rho(A \times X)

4501: =

4502: v(A)^{-1} \rho(X \times A)

4503: \cr&=

4504: \int \sum_{x \in \vertex(\gh)} f(\gh, x, \bp) \,d\rtd(\gh, \bp)

4505: \,,

4506: }$$

4507: which means that $\rtd$

4508: satisfies the Mass-Transport Principle, i.e., is unimodular.

4509: Furthermore, if we take $f(\gh, x, y) := \II{x = y}$, then we see that

4510: $\rho(A \times X) = v(A)$, so that

4511: there is a constant $c$ such that $v(A) = c |A|$.

4512: Likewise, if $f(\gh, x, y) := \II{x = y, (\gh, x) \in \ev A}$, then we see

4513: that for each $\A$, there is another constant $c_\A$ such that $v(A)

4514: \rtd(\A) = c_\A |A|$.

4515: It follows that $\rtd$ does not depend on $A$.

4516: \endprocl

4517: 

4518: 

4519: %\procl x.aperiodic

4520: %Similarly, given any amenable graph $\gh$ and any network on it, if we take

4521: %a limit point of the sequence of probability measures gotten by taking a

4522: %uniform root in a F\o lner sequence, we get a unimodular measure.

4523: %\endprocl

4524: 

4525: 

4526: 

4527: \procl x.dual \procname{Planar Duals}

4528: Let $\rtd$ be a unimodular probability measure on plane graphs all of

4529: whose faces have finitely many sides.

4530: We are assuming that to each graph, there is a measurably associated

4531: plane embedding.

4532: Thus, each graph $\gh$ has a plane dual $\dual \gh$ with respect to its

4533: embedding.

4534: In fact, to be technically accurate in what follows, we replace the

4535: embedding by an assignment (possibly random) of marks to the edges that

4536: indicate the cyclic order in which they appear around a vertex

4537: in a fixed orientation of the plane.

4538: (For example, if a vertex $x$ has $d$ edges incident to it, then one can

4539: let the $d$ edge marks associated to $x$ be $\{1, 2, \ldots, d\}$ in cyclic

4540: order, with the one marked 1 chosen at random, independently of marks

4541: elsewhere.)

4542: Then the plane dual graph is defined entirely with respect to the

4543: resulting network in an automorphism-equivariant way, needing no reference

4544: to the plane.

4545: \par

4546: Provided a certain finiteness condition is satisfied,

4547: there is a natural unimodular probability measure on the dual graphs,

4548: constructed as follows.

4549: For a face $f$, let $\deg f$ denote the

4550: number of sides of $f$. For a vertex $x$, let $F(x) := \sum_{f \sim x}

4551: 1/\deg f$.

4552: Assume that $Z := \int F(\bp) \,d\rtd(\gh, \bp) < \infty$.

4553: To create a unimodular probability measure $\dual \rtd$ on the duals,

4554: first choose $(\gh, \bp)$ with law $\rtd$ biased by $F(\bp)/Z$.

4555: Then choose a face $f_0$ incident to $\bp$ with probability proportional to

4556: $1/\deg f_0$.

4557: The law of the resulting rooted graph $(\dual \gh, f_0)$ is $\dual \rtd$:

4558: $$

4559: \dual \rtd(\A) :=

4560: Z^{-1} \int \sum_{f_0 \sim \bp} {1 \over \deg f_0} \II{(\dual \gh, f_0) \in

4561: \A}

4562: \,d\rtd(\gh, \bp)

4563: $$

4564: for Borel $\A \subseteq \GG_*$.

4565: To prove that $\dual \rtd$ is indeed unimodular, let $k : \gtwo \to [0, \infty]$

4566: be Borel. Then

4567: $$\eqaln{

4568: Z \int \sum_{f \in \vertex(\dual \gh)} k(\dual \gh, f_0, f)

4569: \,d\dual \rtd(\dual \gh, f_0)

4570: &=

4571: \int

4572: \sum_{f_0 \sim \bp} {1 \over \deg f_0}

4573: \sum_{f \in \vertex(\dual \gh)} k(\dual \gh, f_0, f)

4574: \,d\rtd(\gh, \bp)

4575: \cr&=

4576: \int

4577: \sum_{f_0 \sim \bp}

4578: \sum_{f \in \vertex(\dual \gh)}

4579: {1 \over \deg f_0}

4580: k(\dual \gh, f_0, f)

4581: \sum_{x \sim f} {1 \over \deg f}

4582: \,d\rtd(\gh, \bp)

4583: \cr&=

4584: \int

4585: \sum_{x \in \vertex(\gh)}

4586: \sum_{f_0 \sim \bp}

4587: \sum_{f \sim x}

4588: {1 \over \deg f_0}

4589: {1 \over \deg f}

4590: k(\dual \gh, f_0, f)

4591: \,d\rtd(\gh, \bp)

4592: \cr&=

4593: \int

4594: \sum_{x \in \vertex(\gh)}

4595: \sum_{f_0 \sim x}

4596: \sum_{f \sim \bp}

4597: {1 \over \deg f_0}

4598: {1 \over \deg f}

4599: k(\dual \gh, f_0, f)

4600: \,d\rtd(\gh, \bp)

4601: \cr&\hskip1in\hbox{[by the Mass-Transport Principle for $\rtd$]}

4602: \cr&=

4603: \int

4604: \sum_{x \in \vertex(\gh)}

4605: \sum_{f \sim x}

4606: \sum_{f_0 \sim \bp}

4607: {1 \over \deg f_0}

4608: {1 \over \deg f}

4609: k(\dual \gh, f, f_0)

4610: \,d\rtd(\gh, \bp)

4611: \cr&=

4612: Z \int \sum_{f \in \vertex(\dual \gh)} k(\dual \gh, f, f_0)

4613: \,d\dual \rtd(\dual \gh, f_0)

4614: \,.

4615: }$$

4616: Thus, $\dual \rtd$ satisfies the Mass-Transport Principle, so is unimodular.

4617: A similar argument shows that $\dual{(\dual \rtd)} = \rtd$.

4618: \par

4619: Another important construction comes from combining the primal and dual

4620: graphs into a new plane graph by adding a vertex where each edge crosses

4621: its dual.

4622: That is, if $\gh$ is a plane graph and $\dual \gh$ its dual, then every

4623: edge $e \in \edge(\gh)$ intersects $\dual e \in \edge(\dual \gh)$ in one

4624: point, $v_e$.

4625: (These are the only intersections of $\gh$ and $\dual \gh$.)

4626: For $e \in \edge(\gh)$, write $\widehat e$ for the pair of edges that

4627: result from the subdivision of $e$ by $v_e$, and likewise for

4628: $\widehat{\dual e}$.

4629: This defines a new graph $\widehat \gh$, whose vertices are

4630: $\verts(\gh)\cup\verts(\dual \gh)\cup\Big\{v_e\st e\in\edges(\gh)\Big\}$

4631: and whose edges are $\bigcup_{e \in\edges(\gh)} (\widehat e \cup

4632: \widehat{\dual e})$.

4633: If $\expdeg(\rtd) < \infty$, then we may define a unimodular probability

4634: measure $\widehat \rtd$ on the graphs $\widehat \gh$ from $\rtd$ as

4635: follows.

4636: Let $\widehat Z := (3/2) \expdeg(\rtd) + Z \le (5/2) \expdeg(\rtd) <

4637: \infty$.

4638: For $x \in \vertex(\gh)$, let $\widehat N(x)$ be the set consisting of $x$

4639: itself plus the vertices of $\widehat \gh$ that correspond to edges or

4640: faces of $\gh$ that are incident to $x$.

4641: For $w \in \vertex(\widehat \gh)$, define

4642: $$

4643: \delta(w) :=

4644: |\{x \in \vertex(\gh) \st w \in \widehat N(x)\}|^{-1}

4645: =

4646: \cases{

4647: 1 &if $w \in \vertex(\gh)$,\cr

4648: 1/2 &if $w = v_e$ for some $e \in \edge(\gh)$,\cr

4649: 1/\deg w &if $w \in \vertex(\dual \gh)$.\cr

4650: }

4651: $$

4652: Define

4653: $$

4654: \widehat \rtd(\A) :=

4655: \widehat Z^{-1} \int \sum_{w_0 \in \widehat N(\bp)} \delta(w_0)

4656: \II{(\widehat \gh, w_0) \in \A} \,d\rtd(\gh, \bp)

4657: \,.

4658: $$

4659: Note that $\widehat\rtd$ is a probability measure.

4660: To prove that $\widehat\rtd$ is unimodular, let $k : \gtwo \to [0, \infty]$

4661: be Borel. Then

4662: $$\eqaln{

4663: \widehat Z \int \sum_{w \in \vertex(\widehat \gh)} k(\widehat \gh, w_0, w)

4664: &\,d\widehat\rtd(\widehat \gh, w_0)

4665: =

4666: \int

4667: \sum_{w_0 \in \widehat N(\bp)} \delta(w_0)

4668: \sum_{w \in \vertex(\widehat \gh)} k(\widehat \gh, w_0, w)

4669: \,d\rtd(\gh, \bp)

4670: \cr&=

4671: \int

4672: \sum_{w_0 \in \widehat N(\bp)}

4673: \sum_{w \in \vertex(\widehat \gh)}

4674: \delta(w_0)

4675: k(\widehat \gh, w_0, w)

4676: \sum_{x \st w \in \widehat N(x)} \delta(w)

4677: \,d\rtd(\gh, \bp)

4678: \cr&=

4679: \int

4680: \sum_{x \in \vertex(\gh)}

4681: \sum_{w_0 \in \widehat N(\bp)}

4682: \sum_{w \in \widehat N(x)}

4683: \delta(w_0)

4684: \delta(w)

4685: k(\widehat \gh, w_0, w)

4686: \,d\rtd(\gh, \bp)

4687: \cr&=

4688: \int

4689: \sum_{x \in \vertex(\gh)}

4690: \sum_{w_0 \in \widehat N(x)}

4691: \sum_{w \in \widehat N(\bp)}

4692: \delta(w_0)

4693: \delta(w)

4694: k(\widehat \gh, w_0, w)

4695: \,d\rtd(\gh, \bp)

4696: \cr&\hskip1in\hbox{[by the Mass-Transport Principle for $\rtd$]}

4697: \cr&=

4698: \int

4699: \sum_{x \in \vertex(\gh)}

4700: \sum_{w \in \widehat N(x)}

4701: \sum_{w_0 \in \widehat N(\bp)}

4702: \delta(w_0)

4703: \delta(w)

4704: k(\widehat \gh, w, w_0)

4705: \,d\rtd(\gh, \bp)

4706: \cr&=

4707: \widehat Z \int \sum_{w \in \vertex(\widehat \gh)} k(\widehat \gh, w, w_0)

4708: \,d\widehat \rtd(\widehat \gh, w_0)

4709: \,.

4710: }$$

4711: Thus, $\widehat\rtd$ satisfies the Mass-Transport Principle, so is unimodular.

4712: 

4713: \endprocl

4714: 

4715: 

4716: 

4717: \procl x.PWIT \procname{Poisson Weighted Infinite Tree}

4718: Our definition

4719: (\ref s.notation/)

4720: of the metric on the space $\GG_*$ of rooted graphs

4721: refers to ``balls of radius $r$" in which distance is

4722: graph distance, i.e., edges implicitly have length $1$.

4723: \ref b.AS:obj/ work in the setting of graphs whose edges

4724: have positive real lengths, so that distance becomes minimum

4725: path length.  This setting permits one to consider 

4726: graphs which may have infinite degree, but which are still

4727: ``locally finite" in the sense that only finitely vertices fall within

4728: any finite radius ball.

4729: Of course, edge lengths are a special (symmetric) case of edge marks.

4730: %We have restricted ourselves to locally finite graphs. 

4731: %However, if we allowed infinite degree with an $\N$-ordering of the edges

4732: %incident to any vertex, then our theory would extend.

4733: An important example is the following.

4734: Consider a regular rooted tree $T$ of infinite degree.  % marked as follows.

4735: Fix a continuous increasing function $\Lambda$ on $\CO{0, \infty}$ with

4736: $\Lambda(0) = 0$ and $\lim_{t \to\infty} \Lambda(t) = \infty$.

4737: Order the children of each vertex of $T$ via a bijection with $\Z^+$.

4738: For each vertex $x$, consider an independent Poisson process on $\R^+$ with

4739: mean function $\Lambda$.

4740: Define the length of the edge joining $x$ to its $n$th child to be the

4741: $n$th point of the Poisson process associated to $x$.

4742: This is a unimodular random network in the extended sense of \ref b.AS:obj/.

4743: It can be derived by taking the random weak limit of the complete graph on $n$

4744: vertices whose edge lengths are independent with cdf 

4745: %vertices with each edge marked independently by a random variable having cdf

4746: $t \mapsto 1 - e^{-\Lambda(t)/n}$ ($t \ge 0$)

4747: and then deleting all edges in the limit whose length

4748: is $\infty$.

4749: (We are working here with the mark space $[0, \infty]$.)

4750: %Triangles will not disappear in the weak limit. This is why

4751: %we need to delete infinitely long edges.

4752: See \ref b.Aldous:random-assign/.

4753: \endprocl

4754: 

4755: 

4756: %\procl x.planar

4757: %Many examples of random weak limits of finite random trees are described

4758: %in \ref b.Aldous:fringe/.

4759: %Interesting recent work of

4760: %\ref b.AngelSchramm/ and \ref b.Angel/ describes random weak limits of uniform plane

4761: %triangulations.

4762: %\endprocl

4763: 

4764: \procl x.stretched \procname{Edge Replacement}

4765: Here is a general way to create unimodular random rooted graphs from existing

4766: unimodular fixed graphs.

4767: This is an extension of the random subdivision (or stretching) introduced

4768: by \ref b.AdamsLyons/ and studied further in

4769: Example 2.4.4 of \ref b.Kaim:harmonic/ and \ref b.ChenPeres/.

4770: Let $\fin$ be

4771: the set of isomorphism classes of finite graphs with

4772: an ordered pair of distinct distinguished vertices.

4773: For our construction, we may start with a fixed

4774: unimodular quasi-transitive connected graph, $\gh$, or, more generally,

4775: with any unimodular probability measure $\rtd$ on $\GG_*$.

4776: In the former case, fix an orientation of the edges of $\gh$ and let $L$ be

4777: a random field on the oriented edges of $\gh$ that is invariant under the

4778: automorphism group of $\gh$ and takes values in $\fin$ and such that

4779: $\big|\vertex\big(L(e)\big)\big|$ has finite mean for each edge $e$.

4780: Replace each edge $e$ with the graph $L(e)$, where the first of the

4781: distinguished vertices of $L(e)$ is identified with the tail of $e$ and the

4782: second of its distinguished vertices is identified with the head of $e$.

4783: Call the resulting random graph $H$.

4784: There is a unimodular probability measure that is equivalent to this

4785: measure on random graphs.

4786: Namely, let $\mu_i$ be the law of $(H, o_i)$, where $\{o_i\}$ is a complete

4787: section of the vertex orbits of $\gh$.

4788: Given $H$, write

4789: $$

4790: A(x) := 2 + \sum_{e \sim x} \Big(\big|\vertex\big(L(e)\big)\big|-2\Big)

4791: $$

4792: and

4793: $$

4794: c := \sum_i \E\big[A(o_i)\big] \big|\Stab(o_i)\big|^{-1}

4795: \,.

4796: $$

4797: Choose $o_i$ with probability $c^{-1} \E\big[A(o_i)\big]

4798: \big|\Stab(o_i)\big|^{-1}$.

4799: Given $o_i$, choose $(H, o_i)$ with distribution $\mu_i$.

4800: Given this, list the non-distinguished vertices of all $L(e)$ for $e$

4801: incident to $o_i$ as $z_1, z_2, \ldots, z_{A(o_i) - 2}$ and set $z_{A(o_i) - 1}

4802: := z_{A(o_i)} := o_i$.

4803: Let $U$ be a uniform integer in $\big[1, A(o_i)\big]$.

4804: Then $(H, z_U)$ is unimodular and, clearly, has law with respect to which

4805: $\sum_i \mu_i$ is absolutely continuous.

4806: 

4807: Indeed, we state and prove this more generally.

4808: Suppose that $\rtd$ is a unimodular probability measure on $\GG_*$.

4809: Orient the edges of the rooted networks arbitrarily.

4810: Let $\mkmp(e)$ denote the ordered pair of the

4811: marks of the edge $e$ (ordered by the orientation of $e$).

4812: Suppose $L : \marks^2 \to \fin$ is Borel with the property

4813: that whenever $L(\mk_1, \mk_2) = (G, x, y)$, we also have $L(\mk_2, \mk_1) =

4814: (G, y, x)$.

4815: (This will ensure that the orientation of the edges will not affect the

4816: result.)

4817: If

4818: $$

4819: \int \sum_{e \sim \bp} \bigg[\Big|\vertex\Big(L\big(\mkmp(e)\big)\Big)\Big|

4820: - 2\bigg]

4821: \,d\rtd(\gh, \bp)

4822: < \infty

4823: \,,

4824: $$

4825: then let $\rtd'$ be the following measure.

4826: Define

4827: $$

4828: A(\gh, \bp) := 2 + \sum_{e \sim \bp}

4829: \bigg[\Big|\vertex\Big(L\big(\mkmp(e)\big)\Big)\Big| - 2\bigg]

4830: \,.

4831: $$

4832: Choose $(\gh, \bp)$ with probability distribution $\rtd$ biased by $A(\gh,

4833: \bp)$ and replace each edge $e$ by the graph $L\big(\mkmp(e)\big)$, where

4834: the tail and head of $e$ are identified with the first and second

4835: distinguished vertices of $L\big(\mkmp(e)\big)$, respectively; call the

4836: resulting graph $H$.

4837: Write $A := A(\gh, \bp)$ and

4838: list the non-distinguished vertices of all $L\big(\mkmp(e)\big)$ for $e$

4839: incident to $\bp$ as $z_1, z_2, \ldots, z_{A - 2}$ and set $z_{A - 1}

4840: := z_{A} := \bp$.

4841: Let $U$ be a uniform integer in $[1, A]$.

4842: Finally, let $\rtd'$ be the distribution of $(H, z_U)$.

4843: 

4844: This is unimodular by the following calculation.

4845: Write $H(\gh)$ for the graph $H$ formed as above from the network $\gh$.

4846: Let $\vertex_0(\mk_1, \mk_2)$ be the set of non-distinguished vertices of

4847: the graph $L(\mk_1, \mk_2)$.

4848: Write $z_i(\gh, \bp)$ ($1 \le i \le A(\gh, \bp)-2$) for the vertices of the

4849: neighborhood $B(\gh, \bp) := \bigcup_{e \sim \bp} \vertex_0\big(\mkmp(e)\big)$.

4850: Write $z_i(\gh, \bp) := \bp$ for $i = A(\gh, \bp)-1, A(\gh, \bp)$.

4851: Put $c := \int A(\gh, \bp) \,d\rtd(\gh, \bp)$.

4852: In order to show that $\rtd'$ is unimodular, let $f : \gtwo \to [0,

4853: \infty]$ be Borel.

4854: Define $\overline f : \gtwo \to [0, \infty]$ by

4855: $$\eqaln{

4856: \overline f(\gh, x, y) 

4857: &:=

4858: {1 \over c}

4859: \sum_{z \in B(\gh, x)}

4860: \sum_{z' \in B(\gh, y)}

4861: f\big(H(\gh), z, z'\big)

4862: +

4863: {2 \over c}

4864: \sum_{z \in B(\gh, x)}

4865: f\big(H(\gh), z, y\big)

4866: \cr&\qquad+

4867: {2 \over c}

4868: \sum_{z' \in B(\gh, y)}

4869: f\big(H(\gh), x, z'\big)

4870: +

4871: {2 \over c}

4872: f\big(H(\gh), x, y\big)

4873: \,.

4874: \cr

4875: }$$

4876: Then 

4877: $$\eqaln{

4878: \int \sum_{z \in \vertex(H)} &f(H, \bp, z) \,d\rtd'(H, \bp)

4879: \cr&=

4880: {1 \over c} \int {1 \over A(\gh, \bp)} \sum_{i=1}^{A(\gh, \bp)}

4881: \sum_{z \in \vertex(H(\gh))} f\big(H(\gh), z_i(\gh, \bp), z\big)

4882: A(\gh, \bp)\,d\rtd(\gh, \bp)

4883: \cr&=

4884: \int \sum_{x \in \vertex(\gh)} \overline f(\gh, \bp, x) \,d\rtd(\gh, \bp)

4885: \cr&=

4886: \int \sum_{x \in \vertex(\gh)} \overline f(\gh, x, \bp) \,d\rtd(\gh, \bp)

4887: \cr&=

4888: \int \sum_{z \in \vertex(H)} f(H, z, \bp) \,d\rtd'(H, \bp)

4889: \,.

4890: }$$

4891: %If $\rtd = U(G)$ for some finite graph $G$, then it is easy to see that

4892: %$\rtd' = U(H)$ for the graph $H$ constructed from $G$.

4893: %In the general case, we use \ref t.TII/ to see that $\rtd'$ is a random

4894: %weak limit of finite networks, whence unimodular.

4895: %Furthermore, the law of $H$ rooted at the root of $\gh$

4896: %equals $\rtd'$ conditioned on the root being one of the

4897: %distinguished vertices in the replacement graphs.

4898: \endprocl

4899: 

4900: Our final example details the correspondence between random rooted graphs

4901: and graphings of equivalence relations.

4902: 

4903: 

4904: \procl x.equivalence

4905: Let $\mu$ be a Borel probability measure on a topological space

4906: $X$ and $R$ be a Borel subset of

4907: $X^2$ that is an equivalence relation with finite or countable equivalence

4908: classes. We call the triple $(X, \mu, R)$ a {\bf measured equivalence

4909: relation}.

4910: For $x \in X$, denote its $R$-equivalence class by $[x]$.

4911: We call $R$ {\bf measure preserving} if

4912: $$

4913: \int_{x \in X} \sum_{y \in [x]} f(x, y) \,d\mu(x)

4914: =

4915: \int_{x \in X} \sum_{y \in [x]} f(y, x) \,d\mu(x)

4916: $$

4917: for all Borel $f : X^2 \to [0, \infty]$.

4918: A {\bf graphing} $\Phi$ of $R$ is a Borel subset of $X^2$ such that the

4919: smallest equivalence relation containing $\Phi$ is $R$.

4920: A graphing $\Phi$ induces the structure of a graph on the vertex set $X$

4921: by defining an edge between $x$ and $y$ if $(x, y) \in \Phi$ or $(y, x)

4922: \in \Phi$.

4923: Denote the subgraph induced on $[x]$ and rooted at $x$ by $\Phi(x)$.

4924: Given Borel maps $\psi : X \to \marks$ and $\phi : X^2 \to \marks$,

4925: we regard $\psi(x)$ as the mark at $x$ and $\phi(x, y)$ as the mark at $x$

4926: of the edge from $x$ to $y$.

4927: Thus, $\Phi(x)$ is a random rooted network. Its law (or, rather, the law of

4928: its rooted isomorphism class) is unimodular iff $R$ is measure preserving.

4929: 

4930: Conversely, suppose that $\rtd$ is a probability measure on $\GG_*$.

4931: Add independent uniform marks as second coordinates to the existing marks

4932: and call the resulting measure $\nu$.

4933: Write $\dmn \subset \GG_*$ for the set of (isomorphism classes of) rooted

4934: networks with distinct marks.

4935: Thus, $\nu$ is concentrated on $\dmn$.

4936: Define $R \subset \dmn^2$ to be the set of pairs of (isomorphism classes

4937: of) rooted networks that are

4938: non-rooted isomorphic.

4939: Define $\Phi \subset R$ to be the set of pairs of isomorphic rooted

4940: networks whose roots are neighbors in the unique (non-rooted) isomorphism.

4941: Then $(\dmn, \nu, R)$ is a measured equivalence relation with graphing

4942: $\Phi$.

4943: We have that $R$ is measure preserving iff $\rtd$ is unimodular.

4944: If we define the mark map $\projection$ that forgets the second coordinate,

4945: then $\nu$ pushes forward to $\rtd$, i.e., $\rtd = \nu \circ \projection^{-1}$.

4946: 

4947: Thus, the theory of unimodular random rooted networks has substantial

4948: overlap with the

4949: theory of graphed measure-preserving equivalence relations.

4950: The largest difference between the two theories lies in the foci of

4951: attention: We focus on probabilistic aspects of the graphing, while the

4952: other theory focuses on ergodic aspects of the equivalence relation (and,

4953: thus, considers all graphings of a given equivalence relation).

4954: The origins of our work lie in two distinct areas: one is group-invariant

4955: percolation on graphs, while the other is asymptotic analysis of finite

4956: graphs. The origin of the study of measured equivalence relations lies in

4957: the ergodic theory of group actions.

4958: Some references for the latter work, showing relations to von Neumann

4959: algebras and logic, among other things, are

4960: \refbmulti{FM1,FM2}, \ref b.FHM/,

4961: \ref b.CFW/,

4962: \ref b.Zimmer:book/,

4963: \ref b.KechrisMiller/,

4964: \ref b.BeckerKech:book/, \ref b.Kechris:turing/,

4965: \ref b.AdamsLyons/,

4966: \refbmulti {Kaim:amen,Kaim:harmonic},

4967: \ref b.Paulin/,

4968: \refbmulti{Gaboriau:cost,Gaboriau:betti},

4969: and \refbmulti{Furman1,Furman2}.

4970: \endprocl

4971: 

4972: 

4973: 

4974: 

4975: 

4976: \bsection{Finite Approximation}{s.finite}

4977: 

4978: Although we do not present any theorems in this section, because of its

4979: potential importance, we have devoted the whole section to the question of

4980: whether

4981: finite networks are weakly dense in

4982: $\uni$.

4983: Let us call random weak limits of finite networks {\bf sofic}.

4984: 

4985: \procl q.TII \procname{Finite Approximation}

4986: Is every probability measure in $\uni$ sofic? In other words, if

4987: $\rtd$ is a unimodular probability measure on $\GG_*$,

4988: do there exist finite networks $G_n$ such that

4989: $U(G_n) \cd \rtd$?

4990: \endprocl

4991: 

4992: To appreciate why the answer is not obvious, consider the

4993: special case of the (non-random) graph consisting of the

4994: infinite rooted 3-regular tree.

4995: It is true that there exist finite graphs $G_n$ that

4996: approximate the infinite 3-regular tree in the sense of random

4997: weak convergence; a moment's thought shows these {\it cannot}

4998: be finite trees.

4999: This special case is of course known

5000: (one can use finite quotient groups of $\Z_2 * \Z_2 * \Z_2$,

5001: random $3$-regular graphs, or expanders (\ref B.lub94/)),

5002: but the constructions in this special case do not readily extend to

5003: the general case.

5004: 

5005: Another known case of sofic measures is more difficult to establish.

5006: Namely, \ref b.Bowen:periodic/ showed

5007: that all unimodular networks on regular trees are sofic. (To deduce this from

5008: his result, one must use the fact, easily established, that networks with

5009: marks from a finite set are dense in $\GG_*$.)

5010: 

5011: \procl x.fdd

5012: The general unimodular Galton-Watson

5013: measure $\UGW$ (\ref x.AGW/) is also sofic.

5014: To see this, consider the following random networks,

5015: sometimes called ``fixed-degree

5016: distribution networks" and first studied by \ref b.MolloyReed/.

5017: Given

5018: $\Seq{r_k}$ and $n$ vertices, give each vertex $k$ balls with probability

5019: $r_k$, independently. Then pair the balls at random and place an edge for

5020: each pair between the corresponding vertices. There may be one ball left

5021: over; if so, ignore it. Let $m_0 := \sum k r_k$, which we assume is

5022: finite.  In the limit, we get a tree where the root has degree $k$ with

5023: probability $r_k$, each neighbor of the root, if any, has degree $k$ with

5024: probability $k r_k/m_0$, etc. In fact, we get $\UGW$ for the

5025: offspring distribution $k \mapsto (k+1) r_{k+1}/m_0$.  Thus, if we want

5026: the offspring distribution $\Seq{p_k}$, we need merely start with $r_k :=

5027: c^{-1} p_{k-1}/k$ for $k \ge 1$ and $r_0 := 0$, where $c := \sum_{k \ge 0}

5028: p_k/(k+1)$.

5029: \endprocl

5030: 

5031: Let us compare the

5032: intuitions behind amenability and unimodularity.

5033: One can define an average of any bounded function on the vertices of, say,

5034: an amenable Cayley graph; the average will be the same for any translate of

5035: the function.

5036: This can also be regarded as an average of the function with respect to a

5037: probability measure that chooses a group element uniformly at random;

5038: however, the precise justification of this requires a measure that, though

5039: group invariant, is only finitely additive.

5040: Nevertheless, this invariant measure is approximated by uniform measures on

5041: finite sets, namely, F\o{}lner sets.

5042: By contrast, the justification that a unimodular random rooted graph

5043: provides a uniform distribution on the vertices is via the Mass-Transport

5044: Principle. The measure itself is, of course, countably additive; if it is

5045: sofic, then it,

5046: too, is approximated by uniform measures on finite sets.

5047: The two intuitions concerning uniform measures that come from amenability

5048: and from unimodularity agree insofar as every amenable quasi-transitive graph

5049: is unimodular, as shown by \ref b.SoardiWoess/ and \ref b.Salvatori/.

5050: 

5051: One might think that if a sequence $\Seq{\gh_n}$ of finite graphs has a

5052: fixed transitive graph $\gh$ as its random weak limit, then any unimodular

5053: probability measure on networks supported by $\gh$ would be a random weak

5054: limit of some choice of networks on the same sequence $\Seq{\gh_n}$.

5055: This is false, however; e.g., if $\gh$ is a 3-regular tree, then almost any

5056: choice of a sequence of growing 3-regular graphs has $\gh$ as its random weak

5057: limit. However, there is a random independent set\ftnote{*}{This means that

5058: no two vertices of the set are adjacent.} of density 1/2 on $\gh$ whose

5059: law is automorphism invariant,

5060: while the density of independent sets in random 3-regular graphs is bounded

5061: away from 1/2 (see \ref b.FS/).

5062: Nevertheless, if \ref q.TII/ has a positive answer, then

5063: it is not hard to show that

5064: there is some sequence of finite graphs that has this

5065: property of carrying arbitrary networks.

5066: 

5067: Recall that the {\bf Cayley diagram} of a group $\gp$ generated by a finite

5068: subset $S$ is the network $(\gh, \bp)$ with vertex set $\gp$, edge set

5069: $\big\{(x, x s) \st x \in \gp,\ s \in S\big\}$, root $\bp$ the identity

5070: element of $\gp$, and edge marks $s$ at the endpoint $x$ of $(x, x s)$ and

5071: $s^{-1}$ at the endpoint $x s$ of $(x, x s)$, as in \ref r.anygroup/.

5072: We do not mark the vertices (or mark them all the same).

5073: \ref b.Weiss:sofic/ defined $\gp$ to be {\bf sofic} if its Cayley diagram 

5074: is a random weak limit of finite networks with edge marks from

5075: $S \cup S^{-1}$.

5076: It is easy to check that this property does not depend on the generating

5077: set $S$ chosen.

5078: By embedding $S \cup S^{-1}$ into $\marks$, we can use a positive answer to

5079: \ref q.TII/ to give every

5080: Cayley diagram as a random weak limit of {\it some\/} finite networks.

5081: Changing the marks on the finite networks to their nearest points in $S

5082: \cup S^{-1}$ gives the kind of approximating networks desired. That is, we

5083: would have that

5084: every finitely generated group is sofic.

5085: As we mentioned in the introduction, this would have plenty of

5086: consequences.

5087: 

5088: To illustrate additional consequences of

5089: a positive answer to

5090: \ref q.TII/, we establish

5091: the existence of various probability measures on sofic networks.

5092: The idea is that if a class of probability measures is specified by a

5093: sequence of local ``closed" conditions in such a way that there is a measure in

5094: that class on any finite graph, then there is an automorphism-invariant

5095: measure in that class on any sofic quasi-transitive graph.

5096: Rather than state a general theorem to that effect, we shall

5097: illustrate the principle by two examples.

5098: 

5099: 

5100: \procl x.gibbs \procname{Invariant Markov Random Fields}

5101: Consider networks with vertex marks $\pm 1$ and no (or constant) edge marks.

5102: %Let $H$ be a local potential function, i.e., for any finite network

5103: %$\gh$, $H(\gh) \in \OC{-\infty, + \infty}$.

5104: %For example, suppose $H$ is $\infty$ unless all marks are $\pm 1$.

5105: %Let us ignore edge marks.

5106: Given a finite graph $\gh$, $h \in \R$, and $\beta > 0$,

5107: the probability measure $\nu_\gh$ on mark maps $\mkmp : \verts(\gh) \to \{-1,

5108: +1\}$ given by

5109: $$

5110: \nu_\gh(\mkmp) := Z^{-1} \exp\Big\{\sum_{x \in \verts(\gh)} \beta h \mkmp(x) -

5111: \sum_{x \sim y} \beta \mkmp(x) \mkmp(y)\Big\}

5112: \,,

5113: $$

5114: where $Z$ is the normalizing constant required to make these probabilities

5115: add to 1, is known as the

5116: anti-ferromagnetic Ising model at inverse

5117: temperature $\beta$ and with external field $h$ on $\gh$.

5118: %is given by $H(\gh) =

5119: %- \beta h \epsilon$ when $\gh$ is an isolated vertex with mark $\epsilon$,

5120: %$H(\gh) = + \beta \epsilon_1 \epsilon_2$ when $\gh$ is a single edge whose

5121: %endpoints are marked $\epsilon_1$ and $\epsilon_2$, and $H(\gh) = 0$

5122: %otherwise.

5123: Let $\gh$ now be an infinite sofic transitive graph.

5124: Let $\Seq{\gh_n}$ be an approximating sequence of finite graphs

5125: and let $\nu_n := \nu_{\gh_n}$ be the corresponding probability measures.

5126: Write $\overline \gh_n$ for the corresponding random network on $\gh_n$

5127: and, as usual, $U(\overline \gh_n)$ for the uniformly rooted network.

5128: Since $U(\overline \gh_n)$ is unimodular, so is any weak limit point,

5129: $\rtd$.

5130: By tightness, there is such a weak limit point, and it is concentrated on

5131: networks with underlying graph $\gh$.

5132: By \ref t.uni-vs-invar/, we may lift $\rtd$ to an

5133: automorphism-invariant

5134: measure $\nu = \lift_\rtd$ on networks on $\gh$.

5135: This measure $\nu$ is a Markov random field with the required Gibbs

5136: specification, meaning that for any finite subgraph $H$ of $\gh$, the

5137: conditional $\nu$-distribution of $\mkmp \restrict \vertex(H)$ given $\mkmp

5138: \restrict \bdv H$ is equal to the conditional $\nu_H$-distribution of the

5139: same thing.

5140: One is often interested in invariant random fields, not just any random

5141: fields with the given Gibbs specification.

5142: One can easily get a Markov random field with the required Gibbs

5143: specification by taking a limit over subgraphs of $\gh$, but this will not

5144: necessarily produce an invariant measure. In case $\gh$ is amenable, one

5145: could take a limit of averages of the resulting measure to obtain an

5146: invariant measure, but this will not work in the non-amenable case.

5147: That is the point of the present construction.

5148: A variation on this is spin glasses:

5149: Here, for finite graphs $\gh$, the measure $\nu_\gh$ is

5150: $$

5151: \nu_\gh(\mkmp) := Z^{-1} \exp\Big\{\sum_{x \in \verts(\gh)} \beta h \mkmp(x) +

5152: \sum_{x \sim y} \beta J_{x, y} \mkmp(x) \mkmp(y)\Big\}

5153: \,,

5154: $$

5155: where $J_{x, y}$ are, say, independent $\pm 1$-valued random variables.

5156: Again, one can find an invariant spin glass (coupled to the independent

5157: interactions $J_{x, y}$) with the same parameters $h$ and $\beta$ on any

5158: transitive sofic graph by the above method.

5159: \endprocl

5160: 

5161: 

5162: \procl x.sandpile \procname{Invariant Sandpiles}

5163: Consider now networks with vertex marks in $\N$ and no edge marks.

5164: Given a finite rooted 

5165: graph $(\gh, \bp)$, a mark map $\mkmp : \vertex(\gh)

5166: \to \N$ is called {\bf critical} (or {\bf stable and recurrent}) if for all

5167: $x \in \vertex(\gh)$, we have $\mkmp(x) < \deg(x)$ and for all

5168: subgraphs $W$ of $\gh \setminus \{\bp\}$, there is some $x \in \vertex(W)$

5169: such that $\mkmp(x) \ge \deg_W(x)$; we may take $\mkmp(\bp) \equiv 0$.

5170: In this context, one usually calls the root ``the sink".

5171: It turns out that the set of such mark

5172: maps form a very interesting group, called the {\bf sandpile

5173: group} or {\bf chip-firing group} of $\gh$; see \ref b.btw/, \ref

5174: b.Dhar:survey/, \ref b.Biggs/, and \ref b.MRZ/.

5175: Let $\nu_{(\gh, \bp)}$ be the uniform measure on critical mark maps.

5176: Given a transitive sofic graph $\gh$ and an approximating sequence

5177: $\Seq{\gh_n}$ of finite graphs, let $\nu_n :=

5178: \nu_{(\gh_n, \bp_n)}$ be the corresponding probability measures for any

5179: fixed choice of roots $\bp_n$.

5180: Write $(\overline \gh_n, \bp_n)$ for the corresponding random network on

5181: $(\gh_n, \bp_n)$ and, as usual, $U(\overline \gh_n)$ for the uniformly rooted

5182: network. (This root is unrelated to $\bp_n$.)

5183: Since $U(\overline \gh_n)$ is unimodular, so is any weak limit point,

5184: $\rtd$.

5185: By tightness, there is such a weak limit point, and it is concentrated on

5186: networks with underlying graph $\gh$.

5187: (Probably the entire sequence $\Seq{U(\overline \gh_n)}$ in fact converges

5188: to $\rtd$.)

5189: By \ref t.uni-vs-invar/, we may lift $\rtd$ to an

5190: automorphism-invariant

5191: measure $\nu$ on networks on $\gh$.

5192: This measure $\nu$ is supported by networks with only critical mark maps in

5193: the sense that for all $x \in \vertex(\gh)$, we have $\mkmp(x) < \deg(x)$

5194: and for all subgraphs $W$ of $\gh$, there is some $x \in \vertex(W)$ such

5195: that $\mkmp(x) \ge \deg_W(x)$.

5196: \endprocl

5197: 

5198: However, we do not know how to answer the following question.

5199: 

5200: \procl q.color \procname{Invariant Coloring}

5201: Given a quasi-transitive infinite graph $\gh$ and a number $c$

5202: that is 

5203: at least the chromatic number of $\gh$, is there an $\Aut(\gh)$-invariant

5204: probability measure on proper $c$-colorings of the vertices of $\gh$?

5205: %The following does not work: Let $U(\gh_n) \cd \gh$ (say in the

5206: %transitive case). Consider a coloring of $\gh_n$ that maximizes the number

5207: %of vertices colored $\{0, 1, \ldots, c-1\}$. Will the resulting networks

5208: %have a limit that uses only these colors? In other words, if every large

5209: %neighborhood of a finite graph is $c$-colorable, then can a small

5210: %proportion of vertices be deleted from the graph to give a graph that is

5211: %$c$-colorable? NO; we need to be more careful. Can we be?

5212: Let $D := \max_{x \in \vertex(\gh)} \deg_\gh x $.

5213: A positive answer for $c \ge D + 1$ is due to Schramm (personal

5214: communication, 1997).

5215: If $\gh$ is sofic, then we can also obtain such a measure for $c = D$ by

5216: using a well-known result of Brooks (see, e.g., \ref b.Bollobas/, p.~148,

5217: Theorem V.3.)

5218: The question is particularly interesting when $\gh$ is planar and $c = 4$.

5219: In fact, it is then also of great interest to know whether there is a

5220: quasi-transitive proper 4-coloring of $\gh$.

5221: \endprocl

5222: 

5223: 

5224: 

5225: 

5226: 

5227: \medbreak

5228: \noindent {\bf Acknowledgements.}\enspace

5229: We are grateful to 

5230: %Hari Bercovici for helpful discussions related to \ref c.unbdd/ and to

5231: Oded Schramm for discussions of Theorems \briefref

5232: t.uni-vs-invar/ and \briefref t.msf1end/.

5233: We also thank Oded Schramm and \'Ad\'am Tim\'ar for various suggestions.

5234: We are indebted to

5235: G\'abor Elek for telling us about the connection of our work to

5236: various conjectures about groups.

5237: 

5238: 

5239: 

5240: \def\noop#1{\relax}

5241: \input \jobname.bbl

5242: 

5243: \filbreak

5244: \begingroup

5245: \eightpoint\sc

5246: \parindent=0pt\baselineskip=10pt

5247: 

5248: Department of Statistics,

5249: University of California Berkeley,

5250: Berkeley, CA, 94720-3860

5251: \emailwww{aldous@stat.berkeley.edu}

5252: {http://www.stat.berkeley.edu/users/aldous}

5253: 

5254: Department of Mathematics,

5255: Indiana University,

5256: Bloomington, IN 47405-5701

5257: \emailwww{rdlyons@indiana.edu}

5258: {http://mypage.iu.edu/\string~rdlyons/}

5259: 

5260: \endgroup

5261: 

5262: 

5263: \bye

5264: 

5265: 

5266: