1: %% LyX 1.3 created this file. For more info, see http://www.lyx.org/.
2: %% Do not edit unless you really know what you are doing.
3: \documentclass[oneside,german,english,reqno]{amsart}
4: \usepackage{palatino}
5: \usepackage[T1]{fontenc}
6: \usepackage[latin1]{inputenc}
7: \setcounter{tocdepth}{2}
8: \usepackage{graphicx}
9: \usepackage{amssymb}
10: \IfFileExists{url.sty}{\usepackage{url}}
11: {\newcommand{\url}{\texttt}}
12:
13: \makeatletter
14:
15: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% LyX specific LaTeX commands.
16: %% Because html converters don't know tabularnewline
17: \providecommand{\tabularnewline}{\\}
18:
19: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Textclass specific LaTeX commands.
20:
21: \theoremstyle{plain}
22: \theoremstyle{plain}
23: \newtheorem{thm}{Theorem}
24: \theoremstyle{plain}
25: \newtheorem{lem}{Lemma}
26: \theoremstyle{definition}
27: \newtheorem*{defn*}{Definition}
28: \theoremstyle{remark}
29: \newtheorem*{rem*}{Remark}
30: \theoremstyle{plain}
31: \newtheorem*{cor*}{Corollary}
32: \theoremstyle{plain}
33: \newtheorem*{conjecture*}{Conjecture}
34: \newenvironment{lyxlist}[1]
35: {\begin{list}{}
36: {\settowidth{\labelwidth}{#1}
37: \setlength{\leftmargin}{\labelwidth}
38: \addtolength{\leftmargin}{\labelsep}
39: \renewcommand{\makelabel}[1]{##1\hfil}}}
40: {\end{list}}
41: \theoremstyle{remark}
42: \newtheorem*{conclusion*}{Conclusion}
43:
44: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% User specified LaTeX commands.
45: \DeclareMathOperator{\diam}{diam}
46: \DeclareMathOperator{\supp}{supp}
47: \DeclareMathOperator{\len}{len}
48: \DeclareMathOperator{\real}{Re}
49: \DeclareMathOperator{\LE}{LE}
50: \DeclareMathOperator{\QL}{QL}
51: \DeclareMathOperator{\cut}{cut}
52: \DeclareMathOperator{\Haus}{Haus}
53: \DeclareMathOperator{\sh}{\subset_{\Haus}}
54: \DeclareMathOperator{\capa}{Cap}
55: \renewcommand{\theenumi}{(\roman{enumi})}
56: \renewcommand{\labelenumi}{\theenumi}
57: \numberwithin{lem}{section}
58: \usepackage{graphicx}
59: \usepackage{color}
60: \usepackage{caption2}
61: %\usepackage{hyperref}
62: \allowdisplaybreaks[1]
63:
64: \ifx\hyperlink\undefined
65: \newcommand{\refs}[1]{\ref{#1}}
66: \newcommand{\toolong}[2]{\url{#1}}
67: \newcommand{\hypertarget}[1]{}
68: \newcommand{\hrlb}{}
69: \else
70: \newcommand{\refs}[1]{\ref*{#1}}
71: \newcommand{\toolong}[2]{\href{#1}{\nolinkurl{#2} (too long)}}
72: \newcommand{\hrlb}{\\}
73: \fi
74:
75: \usepackage{babel}
76: \makeatother
77: \begin{document}
78: \theoremstyle{remark}
79: \newtheorem*{discussion}{Discussion}
80: \theoremstyle{plain}
81: \newtheorem{sublem}{Sublemma}[lem]
82: \newtheorem*{dsh}{Theorem (Delmotte; Holopainen and Soardi)}
83: \newtheorem*{dmt}{Theorem (Delmotte)}
84: \newtheorem*{hsc}{Theorem (Hebisch and Saloff-Coste)}
85: \newtheorem*{varop}{Theorem (Varopoulos)}
86: \newtheorem*{bpp}{Theorem (Benjamini, Pemantle and Peres)}
87: %\newtheorem*{mainlem}{Main lemma}
88: %\renewcommand{\labelenumi}{\theenumi .}
89: \newcounter{const}
90: \newcounter{Const}
91: \def\newc#1{
92: \refstepcounter{const}
93: \label{#1}
94: }
95: \def\newC#1{
96: \refstepcounter{Const}
97: \label{#1}
98: }
99:
100:
101: \title{The scaling limit of loop-erased random walk in three dimensions}
102:
103:
104: \author{Gady Kozma}
105:
106: \begin{abstract}
107: We show that the scaling limit exists and is invariant to dilations
108: and rotations. We give some tools that might be useful to show universality.
109: \end{abstract}
110: \maketitle
111: \tableofcontents{}
112:
113:
114: \section{Introduction}
115:
116: Loop-erased random walk (LERW) is a model for a random simple path,
117: created by taking a simple random walk and, whenever the random walk
118: hits its path, removing the resulting loop and continuing. See section
119: \ref{sub:Loop-erasure} for a precise definition. It is strongly related
120: to the uniform spanning tree (UST), a random spanning tree of a graph
121: $G$ selected uniformly between all spanning trees of $G$: the path
122: in the UST between two points is distributed like a LERW between them
123: \cite{P91}, and further, the entire UST can be generated using repeated
124: use of LERW by Wilson's algorithm \cite{W96}. Both models, and the
125: connections between them are interesting on a general graph, but we
126: shall be most interested in lattices on $\mathbb{R}^{d}$ and open
127: subsets thereof, in which case these models arise naturally in statistical
128: mechanics in conjunction with the Potts model.
129:
130: Of all the non-Gaussian models in statistical mechanics, LERW is probably
131: the most tractable. Above five dimensions, it can be analyzed using
132: the non-intersections of simple random walk directly \cite[chapter 7]{L91}
133: giving an easy proof that the scaling limit is Brownian motion. In
134: 4 dimensions a logarithmic correction is required \cite{L95}, and
135: that too has been proved with no use of the difficult technique of
136: lace expansion (see \cite{BS85,HvdHS03} for lace expansion). Borrowing
137: a term from physics we might say that the \emph{upper critical dimension}
138: for this model is $4$. In $2$ dimensions, LERW is conformally invariant
139: in the limit as the lattice becomes finer and finer. This allowed
140: physicists to make precise conjectures about fractal dimensions, critical
141: exponents and winding numbers \cite{D92,M92}. Rigorously, $3$ different
142: approaches proved fruitful: the connection to random domino tilings
143: \cite{K00a,K00b}, the connection to SLE \cite{LSW04}, and the approach
144: we will pursue in this paper, \cite{K}. In fact, SLE was discovered
145: \cite{S00} in the context of LERW.
146:
147: Attempts to understand LERW in dimension $3$ focused mainly on the
148: number of steps it takes to reach the distance $r$. Physicists conjecture
149: that it is $\approx r^{\xi}$ and did numerical experiments to show
150: that $\xi=1.62\pm0.01$ \cite{GB90}. Rigorously the existence of
151: $\xi$ is not proved (so we must talk about an upper and lower exponents
152: $\underline{\xi}\leq\overline{\xi}$), and the best estimates known
153: are $1<\underline{\xi}\leq\overline{\xi}\leq5/3$ \cite{L99}. LERW
154: has no natural continuum equivalent in dimensions smaller than $4$
155: --- Brownian motion has a dense set of loops and therefore it is not
156: clear how to remove them in chronological order. In two dimensions
157: the scaling limit is radial SLE $2$, but it is not clear if this
158: can be interpreted as a {}``Brownian motion with loops removed''.
159: For example, take a coupling of Brownian motion and SLE 2 which is
160: the scaling limit of the couple $(R,\LE(R))$ --- it is not proved
161: that this limit exists, but for the purpose of the discussion we may
162: assume it does or alternatively take a subsequential limit. It is
163: not known whether in that coupling the SLE $2$ path is a function
164: of the Brownian path (I was informed of this question by O. Schramm).
165:
166: In this paper we shall show that LERW has a scaling limit in three
167: dimensions. More precisely we shall show the following theorem:
168:
169: \begin{thm}
170: \label{thm:Z3scal}Let $\mathcal{D}\subset\mathbb{R}^{3}$ be a polyhedron
171: and let $a\in\mathcal{D}$. Let $\mathbb{P}_{n}$ be the distribution
172: of the loop-erasure of a random walk on $\mathcal{D}\cap2^{-n}\mathbb{Z}^{3}$
173: starting from $a$ and stopped when hitting $\partial\mathcal{D}$.
174: Then $\mathbb{P}_{n}$ converge in the space $\mathcal{M}(\mathcal{H}(\overline{\mathcal{D}}))$.
175: \end{thm}
176: Here $\mathcal{H}(\mathcal{X})$ is the space of compact subsets of
177: $\mathcal{X}$ with the Hausdorff metric, and $\mathcal{M}(\mathcal{X})$
178: is the space of measures on $\mathcal{X}$ with the topology of weak
179: convergence (these, and a couple of other notations are explained
180: in section \ref{sub:Preliminaries}). In general, the choice of topologies
181: above is not canonical. For example, \cite{LSW04} shows the existence
182: of a scaling limit for LERW in two dimensions replacing $\mathcal{H}$
183: above with the somewhat stronger topology of {}``minimal distance
184: after optimal change of variables''. However, for our techniques
185: the Hausdorff metric is the natural choice. I believe that the tools
186: that will be developed here can be used for a number of convergence
187: questions for LERW (e.g.~the existence of $\xi$, the existence of
188: the scaling limit on more general domains, and in stronger topologies,
189: universality and so on). However, as this paper is long as it is,
190: I chose to show only the simplest consequences: that the limit exists
191: and is invariant to dilations and rotations.%
192: \begin{figure}
193: \includegraphics[%
194: scale=0.6]{lerw3_color.eps}
195:
196:
197: \caption{A loop-erased random walk in three dimensions. Color indicates position
198: on the curve (red to blue to green and back to red).}
199: \end{figure}
200:
201:
202: Since we are interested in scaling limits it might be useful to review
203: quickly known results of this type. The archetypical example is of
204: course the Donsker invariance principle \cite[page 16]{RW94} stating
205: that the scaling limit of simple random walk is Brownian motion in
206: any dimension. As already remarked, in two dimensions the scaling
207: limit of LERW is radial SLE $2$, and a good deal of other discrete
208: models have been shown to converge to SLE: critical percolation on
209: the triangular lattice converges to chordal SLE $6$ \cite{S01,SW01},
210: the Peano curve of the UST converges to chordal SLE $8$ \cite{LSW04},
211: and the harmonic explorer converges to SLE $4$ \cite{SS}. The case
212: of the self-avoiding walk demonstrates the difficulties involved nicely:
213: it has been proved that if the limit exists and is conformally invariant,
214: it would be chordal SLE $8/3$ \cite{LSW}, but the existence of the
215: limit is still open. In high dimensions lace expansion has been used
216: to show that the scaling limit of the self avoiding walk is Brownian
217: motion \cite{BS85,HS92}, and that the scaling limit of percolation,
218: oriented percolation and lattice trees is integrated super-Brownian
219: excursion \cite{NY95,DS98,S99,vdHS03}. In intermediate dimensions
220: much less is known. The discrete Gaussian free field converges to
221: the continuum Gaussian free field and the Richardson model was shown
222: to have a limit shape from subadditivity arguments in any dimension
223: \cite{R73}, and I cannot resist citing the beautiful work on branched
224: polymers in dimension $3$ \cite{BI03a}. But these examples are the
225: exception, not the rule.
226:
227:
228: \subsection{Sketch of the proof}
229:
230: The core of the argument is very similar to that of \cite{K}, so
231: let us recall the argumentation there. Let $R$ be a random walk on
232: a {}``three dimensional graph'' starting from $a$ and stopped on
233: the boundary of some domain $\mathcal{D}$. Let $B\subset\mathcal{D}$
234: be some (small) ball. We write \[
235: \LE(R)=\gamma_{1}\cup\gamma_{2}\cup\gamma_{3}\]
236: where $\gamma_{1}$ is the portion of $\LE(R)$ until the first time
237: when $\LE(R)$ hits $B$. Notice that this is \textbf{not} the same
238: as the loop-erasure of a random walk stopped on $\partial B$! $\gamma_{2}$
239: is the portion of $\LE(R)$ until the last time when $\LE(R)$ is
240: inside $B$, and $\gamma_{3}$ is the reminder (the precise form of
241: this division is in the proof of lemma \ref{lem:0_to_Y}, page \pageref{lem:0_to_Y}).
242: Tracing the process of loop-erasure in $\mathcal{D}$ one sees that
243: $\gamma_{1}$ does not depend on anything that happens inside $B$:
244: when one knows all entry and exit points of $R$ from $B$, and all
245: the trajectories that $R$ does outside $B$, one can calculate $\gamma_{1}$.
246: In particular, if we compare random walks $R^{1}$ and $R^{2}$ on
247: graphs $G^{1}$ and $G^{2}$, where $G^{1}\setminus B=G^{2}\setminus B$
248: and inside $B$ we have some estimate of the sort\begin{equation}
249: p^{1}(v)\simeq p^{2}(v)\label{eq:pisketch}\end{equation}
250: where $p^{i}(v)$ is the probability of a random walk on $G^{i}$
251: to exit $B$ in a particular vertex $v$, then we should have that
252: \[
253: \gamma_{1}^{1}\simeq\gamma_{1}^{2}.\]
254: This argument and the precise meaning of {}``$\simeq$'' are contained
255: in lemma \ref{lem:xizLE}. To make this argument work for $\gamma_{3}$,
256: we have to use the symmetry of loop-erased random walk.
257:
258: Now, if $\gamma_{2}$ is large then we are in the situation that was
259: coined in \cite{S00} a {}``quasi-loop'', namely two close points
260: on the path with a long way between them. In two dimensions it is
261: possible to show that $\LE(R)$ has few quasi-loops by tracing the
262: process that creates them and seeing that it necessitates that a random
263: walk that starts quite close to a loop-erased random walk will avoid
264: hitting it for a long while. This, however, contradicts the discrete
265: Beurling projection principle (see \cite{K87}) that states that a
266: random walk starting near any path has a high probability to intersect
267: it%
268: \footnote{Kesten's theorem is stronger, and claims that the minimum probability
269: is achieved, up to a constant, for a straight line, in which case
270: it can be estimated directly. However we will not need this level
271: of accuracy.%
272: }. See \cite[lemma 2.1]{S00} or \cite[lemma 18]{K}. Unfortunately
273: this argument no longer holds in three dimensions. Random walk starting,
274: say, in a distance $n$ from a straight line, has a reasonable probability
275: to intersect it only after extending to a distance of $n^{2}$, and
276: even after $n^{2}$ the probability to not intersect the line only
277: decreases logarithmically. In other words, in three dimensions not
278: all paths are {}``hittable'', and we have to show specifically that
279: loop-erased random walk is, using facts about its structure. This
280: will be done in chapter \ref{sec:Quasi-loops} and we shall show that
281: the probability that a random walk starting near a loop-erased random
282: walk will avoid hitting it decreases like a power law. This will allow
283: to repeat the above argumentation in three dimensions, show that there
284: are no quasi-loops and hence that LERW is similar on our $G^{1}$
285: and $G^{2}$.
286:
287: The proof that loop-erased random walk is hittable is based on searching
288: for (local) cut times. A cut time for a random walk $R$ is a time
289: $t$ such that $R[0,t]\cap R[t+1,\infty[\;=\emptyset$. The number
290: of cut times is connected to the non-intersection exponent: by considering
291: the parts of $R$ up to $t$ and from $t$ on as two random walks,
292: and reversing the first part, we see that it is important to estimate
293: the probability that two random walks of length $t$ will not intersect.
294: This is $\approx t^{-2\xi}$ \cite{L96b} where $\xi$ is the famous
295: non-intersection exponent of Brownian motion. See section \ref{sub:Background-exponent}
296: for a description of this topic. Heuristically speaking, a set is
297: {}``hittable'' if its Hausdorff dimension is $>1$, and the set
298: of cut times has dimension $2-\xi$ so the argument terminates by
299: the well known fact that $\xi<1$ in three dimensions \cite{BL90b}.
300:
301: Once the fact that $\LE(R)$ has no quasi-loops on either $G^{1}$
302: or $G^{2}$ is established, we get that it is similar on these two
303: graphs. Hence we can show that $\LE(R)$ is similar on $\mathbb{Z}^{3}$
304: and $2\mathbb{Z}^{3}$ by interpolating between them: dissecting into
305: a grid of cubes of intermediate size, and at each step change one
306: cube from $\mathbb{Z}^{3}$ to $2\mathbb{Z}^{3}$. Hence all of the
307: above discussion actually referred to graphs formed by cutting and
308: gluing together cubes of these two graphs.
309:
310: Here are some corresponding reading recommendations:
311:
312: \begin{itemize}
313: \item If you are familiar with \cite{K}, the part most interesting for
314: you would probably be the proof that there are no quasi loops. Read
315: the definition of a Euclidean net (section \ref{sub:DefinitionNet},
316: page \pageref{sub:DefinitionNet}), the definition of an isotropic
317: graph (section \ref{sub:DefinitionIsotropic}, page \pageref{sub:DefinitionIsotropic})
318: and the statements of theorems \ref{lem:escape} and \ref{thm:kof}
319: (pages \pageref{lem:escape} and \pageref{thm:kof} respectively)
320: and jump directly to chapter \ref{sec:Quasi-loops} (page \pageref{sec:Quasi-loops}).
321: \item If you are unfamiliar with \cite{K}, the part most interesting for
322: you would probably be the proof core sketched above. Read the definitions
323: of a Euclidean net and an isotropic graph as above; and the definition
324: of a quasi loop and the statement of theorem \ref{thm:QL} (page \pageref{thm:QL}).
325: Then jump directly to chapter \ref{sec:Isotropic-gluing} (page \pageref{sec:Isotropic-gluing})
326: or even to section \ref{sub:Definition-II} (page \pageref{sub:Definition-II}).
327: \item If you are the kind of person who prefers explicit examples to generalizations,
328: start with chapter \ref{sec:Examples} (page \pageref{sec:Examples})
329: and read a few examples of isotropic graphs and isotropic interpolations.
330: Then you can read the rest of the paper keeping in mind that an {}``isotropic
331: graph'' is really cubes of $\mathbb{Z}^{3}$ and $2\mathbb{Z}^{3}$
332: (and other variations) cut and sewn together so that random walk would
333: behave like Brownian motion. Section \ref{sub:Invariance} also contains
334: the proof of the invariance of the scaling limit to dilations and
335: rotations.
336: \item Chapters \ref{sec:euclidean-nets} and \ref{sec:Brownian-graphs}
337: are mostly recommended for students and non-special\-ists. Chapter
338: \ref{sec:euclidean-nets} consists mostly of citing well known connections
339: between rough isometries, heat kernel decay, Harnack inequality and
340: similar topics. In chapter \ref{sec:Brownian-graphs} we are forced
341: to replicate the results of Lawler \cite{L96b} in our settings. Roughly
342: we show that a relatively simple estimate of hitting probabilities
343: allow to couple random walk and Brownian motion and then Lawler's
344: argument goes through almost unchanged, giving that on the graphs
345: that interest us random walk has many cut times.
346: \end{itemize}
347: Finally I wish to point out how this paper improves over \cite{K}
348: in the two dimensional case. The use of a computer to calculate precise
349: estimates for the harmonic potential on $\mathbb{Z}^{2}$ \cite{KS04}
350: and on {}``hybrid graphs'' has been made completely unnecessary
351: by the use of electrical conductance techniques (see lemma \ref{lem:a}).
352: The use of {}``nice rectangles'' to ensure that hitting probabilities
353: are comparable was replaced by a multi-scale application of Harnack's
354: inequality together with a coupling argument (and in particular we
355: use spheres throughout rather than rectangles). See lemmas \ref{lem:sball}--\ref{lem:striag}
356: and \ref{lem:nodir}--\ref{lem:iiexit2}. Here the representation
357: is of comparable length, but is possibly less cumbersome. Finally,
358: the proof here of the final limit process is much shorter and simpler.
359:
360:
361: \subsection{About the settings}
362:
363: The usual settings for these problems is that of a lattice in $\mathbb{R}^{d}$.
364: However, as explained above, the proof has to cut and saw together
365: different graphs, and even if these graphs were to be grids, the \emph{intermediate
366: objects} we must handle would not be. Hence we need to understand
367: random walks on graphs which are {}``similar'' to $\mathbb{Z}^{d}$.
368: It turns out that there are two important levels of similarity, which
369: correspond to {}``metric'' and {}``conformal'' properties.
370:
371: Much effort has gone into understanding what properties of random
372: walk are related to the metric structure only, or, more formally,
373: are satisfied by any graph \emph{roughly isometric} to $\mathbb{Z}^{d}$
374: (see definition and background on page \pageref{sub:Background-on-RI}).
375: However, one cannot expect LERW on a graph roughly isometric to $\mathbb{Z}^{d}$
376: to converge to a limit independent of the graph. Indeed, the scaling
377: limit of the random walk on the graph $\mathbb{Z}^{2}\times2\mathbb{Z}$
378: is not Brownian motion but a stretched version of it. These are not
379: identical --- indeed, even their hitting distribution on, say, a sphere,
380: differ, which implies that the scaling limit of LERW on $\mathbb{Z}^{3}$
381: and $\mathbb{Z}^{2}\times2\mathbb{Z}$ also differ. Hence we need
382: some condition to ensure that locally the graph is not stretched in
383: any direction. In other words, we need to preserve the conformal structure
384: of $\mathbb{R}^{d}$.
385:
386: Properties related to the conformal structure are less well understood.
387: The {}``invariance principle'', that is the fact that random walk
388: converges to Brownian motion, which is a conformal property, has been
389: researched intensively, but it seems in different contexts than here.
390: Hence we will use a definition of \emph{isotropic graph} which is,
391: to the best of my knowledge, new. These graphs will satisfy the invariance
392: principle (this is more or less a tautology) and they preserve many
393: properties which are not preserved by the metric structure alone,
394: such as escape probabilities from a line, the non-intersection exponent,
395: and so on.
396:
397: Our definition of an isotropic graph (see chapter \ref{sec:Brownian-graphs})
398: is definitely not the most general imaginable. There are at least
399: two important examples which fall out of its scope. The first is a
400: conformal map of a grid --- for example the graph $\mathbb{Z}^{2}/\{(a,b)\sim(-a,-b)\}$
401: embedded into $\mathbb{C}$ via the map $(a,b)\mapsto(a+ib)^{2}$.
402: The second is random graphs, such as the Delaunay triangulation of
403: a Poisson process or the infinite cluster of super-critical percolation.
404: These graphs are not even roughly isometric to $\mathbb{R}^{2}$ and
405: yet are {}``isotropic'' in some heuristic sense. For example, the
406: percolation cluster is isotropic in the sense that it satisfies the
407: invariance principle, see \cite{DFGW89,SS04,BB,MP}. I conjecture
408: that the results here extend to these graphs, but will not complicate
409: the paper by considering them.
410:
411: We shall prove theorem \ref{thm:Z3scal} (and other results) in both
412: the two and three dimensional cases. While the three dimensional case
413: is the more interesting one, the two dimensional proof is not quite
414: a subset of known results: it is proved for multiply connected domains
415: and for graphs more general than grids. However, at points the presentation
416: of specifically two dimensional issues will be sketchy.
417:
418:
419: \subsection{Acknowledgements}
420:
421: Enormous thanks go to Itai Benjamini for many useful discussions,
422: encouragements, and for pointing out to me the relevance of rough
423: isometries and of the non-intersection exponent to this project. Many
424: thanks go to Gidi Amir and Omer Angel for useful discussions, in particular
425: with respect to counterexamples around lemmas \ref{lem:roughxi} and
426: \ref{lem:plane}. Lemma \ref{lem:omer} was discovered together with
427: Omer Angel.
428:
429: This project was carried out while I was enjoying the hospitality
430: of, in chronological order, Université Bordeaux I, The Weizmann Institute
431: of Science (Charles Clore fund), Tel Aviv University and the Institute
432: of Advanced Study in Princeton (Oswald Veblen fund). I wish to thank
433: all these institutions, and especially A. Olevski\u\i{} from Tel
434: Aviv University who went to great efforts for me at unusual times.
435:
436:
437: \subsection{\label{sub:Preliminaries}Preliminaries}
438:
439: A weighted graph is a couple $(G,\omega)$ where $G$ is a set and
440: $\omega:G\times G\to\left[0,\infty\right[$ such that $\omega(v,w)=\omega(w,v)$.
441: We shall often call $(G,\omega)$ simply $G$ and use $\omega$ only
442: in the places it is needed. For $v\in G$ the neighbors of $v$ are
443: the vertices $w$ such that $\omega(v,w)>0$. We denote by $v\sim w$
444: the neighborhood relation. We shall assume always that the number
445: of neighbors of every vertex is bounded and that the graph has \emph{bounded
446: weights} i.e.\[
447: \sup_{v,w}\omega(v,w)<\infty\quad\inf_{v\sim w}\omega(v,w)>0.\]
448: We do not assume $\omega(v,v)=0$ i.e.~we allow self loops.
449:
450: A directed graph is a graph where $\omega(v,w)$ might be different
451: from $\omega(w,v)$. We will only use directed graphs once, in section
452: \ref{sub:Beurling}. Unless specifically marked {}``directed'' everything
453: below should be assumed to hold for undirected graphs only.
454:
455: For a subset $X\subset G$ we denote by $\partial X$ the external
456: boundary, namely all vertices of $G\setminus X$ with a neighbor in
457: $X$. When this is not clear from the context, we shall denote $\partial_{G}X$
458: for the graph boundary and $\partial_{\textrm{cont}}$ for the boundary
459: of subsets of $\mathbb{R}^{d}$ in the usual sense. We write $\overline{X}=X\cup\partial X$.
460:
461: A path in a graph is a function $\gamma$ from $\{1,\dotsc,n\}$ to
462: $G$ such that $\gamma(n)$ and $\gamma(n+1)$ are neighbors. $n$
463: is the length of the path, denoted by $\len\gamma$. If $\gamma_{1}$
464: and $\gamma_{2}$ are two paths and $\gamma_{1}(\len\gamma_{1})$
465: is a neighbor of $\gamma_{2}(1)$ (in which case we call $\gamma_{1}$
466: and $\gamma_{2}$ {}``concatenatable'') we shall define $\gamma_{1}\cup\gamma_{2}$
467: to be the path of length $\len\gamma_{1}+\len\gamma_{2}$ obtained
468: by concatenating them. It will be convenient to regard $\emptyset$
469: as a path of length $0$ and define $\gamma\cup\emptyset=\emptyset\cup\gamma=\gamma$.
470: The notation $\gamma[a,b]$ will be a short for the path of length
471: $b-a+1$ defined by $\gamma'(i)=\gamma(a+i-1)$, and also for the
472: set $\{\gamma(t):t\in[a,b]\}$ (there will rarely be a need to differentiate
473: between a path and its image). The same holds for other types of segments
474: (open, half-open). We say that $\gamma$ is {}``between'' $\gamma(1)$
475: and $\gamma(\len\gamma)$ and call $G$ connected if there exists
476: a path between any two vertices.
477:
478: A random walk on a weighted graph is a process $R$ in discrete time
479: such that $R(t+1)$ depends only on $R(t)$ and \begin{equation}
480: \mathbb{P}(R(t+1)=w\,|\, R(t)=v)=\frac{\omega(v,w)}{\omega(v)}\quad\omega(v):=\sum_{w}\omega(v,w).\label{eq:omega}\end{equation}
481: $\mathbb{P}$ denoting the probability. We shall denote by $\mathbb{E}$
482: the expectation. When we shall need to specify the starting point,
483: we shall do so using $\mathbb{P}^{v}$ for the probability when $R(0)=v$,
484: and similarly $\mathbb{E}^{v}$. When we shall need to specify the
485: graph, we shall do so using $\mathbb{P}_{G}^{v}$ etc. Occasionally
486: (as in the statement of theorem \ref{thm:Z3scal}) we will have a
487: graph with an embedding in $\mathbb{R}^{d}$ and the {}``starting
488: point'' would be an $a\in\mathbb{R}^{d}$. In this case we mean by
489: $\mathbb{P}_{G}^{a}$ a random walk on $G$ starting from the point
490: of $G$ closest to $a$ (if more than one such point exist, choose
491: one, say by lexicographic order).
492:
493: For a subset $X\subset G$ and a random walk $R$ we denote by $T(X)$
494: the hitting time of $X$ i.e.\[
495: T(X):=\min\{ t\geq1:R(t)\in X\}\]
496: or $\infty$ if the set is never hit. If $X=\{ x\}$ we shall write
497: $T(x)$ as short for $T(\{ x\})$. If $d$ is some metric on $G$
498: and if $X=\partial B(v,r)$ where $B(v,r)$ is a ball around $v$
499: of radius $r$ in the metric $d$, we will denote for short $T_{v,r}:=T(\partial B(v,r))$
500: (and assume the metric is clear from the context). Note that even
501: if we start from $v$, $T(v)$ is non-trivial since hitting is defined
502: only for $t\geq1$.
503:
504: Sometimes we will have a few independent walks denoted by $R^{1},R^{2},\dotsc$.
505: In this case the corresponding stopping times will be denoted by $T^{i}(X)$
506: and $T_{v,r}^{i}$. Similarly we shall denote $\mathbb{P}^{1,v,2,w}$
507: when we want to denote that $R^{1}$ started from $v$ while $R^{2}$
508: started from $w$.
509:
510: The strong Markov property says that for any stopping time $T$ the
511: random walk after $T$ behaves like a regular random walk. We shall
512: often use it, say for an event $E$ that depends only on what happened
513: after $T$, in the form $\mathbb{P}(E)=\mathbb{E}\mathbb{P}^{R(T)}(E)$.
514: Here $\mathbb{E}$ denotes expectation over the value of $R(T)$.
515:
516: Random walk is symmetric in the sense that the probabilities to traverse
517: a given path in one direction and in the opposite direction are equal
518: up to the ratio of $\omega$ at the beginning and end. In particular
519: we can sum over all paths of length $t$ and get\begin{equation}
520: \omega(v)\mathbb{P}^{v}(R(t)=w)=\omega(w)\mathbb{P}^{w}(R(t)=v)\quad\forall t,v,w.\label{eq:sym}\end{equation}
521: A similar argument shows that if $v,w\in A\subset G$ then\begin{equation}
522: \omega(v)\mathbb{P}^{v}(T(A)<\infty,R(T(A))=w)=\omega(w)\mathbb{P}^{w}(T(A)<\infty,R(T(A))=v).\label{eq:symT}\end{equation}
523:
524:
525: For a function $f:G\to\mathbb{R}$ (or to any linear space over $\mathbb{R}$)
526: we define the (discrete) Laplacian of $f$, $\Delta f$ by \[
527: (\Delta f)(v)=-f(v)+\sum_{w\sim v}\frac{\omega(v,w)}{\omega(v)}f(w)\]
528: A function $f$ such that $\Delta f$ is zero will be called (discretely)
529: harmonic. If $\Delta f$ is zero on a set $A\subset G$ we shall call
530: $f$ {}``harmonic on $A$''. Harmonic functions satisfy the maximum
531: principle, i.e.~a function harmonic on $A$ attains its maximum in
532: $\overline{A}$ on the boundary $\partial A$. Harmonic functions
533: are related to random walks by the following simple and well known
534: fact: if $f$ is harmonic on $A$ and $v\in A$ then\begin{equation}
535: f(v)=\mathbb{E}^{v}(f(R(T(\partial A)))).\label{eq:fEfRT}\end{equation}
536:
537:
538: For two sets $A$ and $B$ in a metric space $(X,d)$, we define their
539: distance by \[
540: d(A,B):=\inf_{a\in A,\, b\in B}d(a,b).\]
541: If $x\in X$ we write $d(x,A)$ as a short for $d(\{ x\},A)$. The
542: Hausdorff distance between $A$ and $B$ is defined by \[
543: d_{\Haus}(A,B):=\max(\sup_{a\in A}d(a,B),\sup_{b\in B}d(b,A)).\]
544: The diameter of a set is defined by\[
545: \diam A:=\sup_{a,b\in A}d(a,b).\]
546: If the metric space has an addition structure, we will use the notation
547: $A+B$ for the Minkowski sum of the sets $A$ and $B$ i.e.\[
548: A+B:=\{ a+b:a\in A,b\in B\}.\]
549: In particular, if $B$ is a ball centered at $0$ then \[
550: A+B(0,r)=\{ x:d(x,A)<r\}.\]
551: We will sometimes abuse notations by denoting the right hand side
552: by $A+B(r)$ even when the metric space has no addition structure.
553:
554: A domain is a non-empty bounded open connected subset of $\mathbb{R}^{d}$.
555: A polyhedron is a domain whose boundary is composed of a collection
556: of non-degenerate linear polyhedra of dimension $d-1$. In particular,
557: we do not require that the boundary of the polyhedron is connected.
558: For simplicity, however, we will not allow slits.
559:
560: By $C$ and $c$ we denote absolute constants which may be different
561: from place to place. $C$ will usually denote constants which are
562: {}``large enough'' and $c$ {}``small enough''. We shall number
563: ($c_{1},c_{2},\dotsc$) only constants to which we will need to refer
564: to again. Sometimes we shall also write $C(\cdot)$ and $c(\cdot)$
565: for a constant which is not properly absolute --- it depends on some
566: parameters --- but is best thought of as absolute. This notation implicitly
567: means I cannot think of any applications where the parameters are
568: not themselves constants. Again, $C(G)$ could change from place to
569: place, and we shall number only those that we shall need to refer
570: to in the future. If, say, $C_{187}$ depends on some parameters we
571: shall only note this once and from that place on refer to it as simply
572: $C_{187}$, not $C_{187}(\alpha,\tau,\mathcal{H})$.
573:
574: As usual we denote by $\left\lfloor x\right\rfloor $ the largest
575: integer $\leq x$ and by $\left\lceil x\right\rceil $ the smallest
576: integer $\geq x$.
577:
578:
579: \subsection{\label{sub:Loop-erasure}Loop erasure}
580:
581: For a finite path $\gamma:\{1,\dotsc,n\}\rightarrow G$ we define
582: its loop erasure, $\LE(\gamma)$, which is a simple path in $G$,
583: by the consecutive removal of loops from $\gamma$. Formally, \begin{eqnarray}
584: \LE(\gamma)(1) & := & \gamma(1)\nonumber \\
585: \LE(\gamma)(i+1) & := & \gamma(j_{i}+1)\quad j_{i}:=\max\{ j:\gamma(j)=\LE(\gamma)(i)\}\quad.\label{eq:defLE}\end{eqnarray}
586: Which is defined for all $i$ such that $j_{i-1}<n$.
587:
588: \begin{lem}
589: \label{lem:condLE_sym}Let $b^{0},b^{1}\in B\subset G$. Let $R^{i}$
590: be a random walk starting at $b^{i}$, stopped at $B$ and conditioned
591: to hit $b^{1-i}$. Then $\LE(R^{0})$ has the same distribution as
592: the reversal of $\LE(R^{1})$.
593: \end{lem}
594: \noindent This is well known. See e.g.~\cite[lemma 2]{K}.
595:
596: The following lemma was discovered with Omer Angel. To the best of
597: our knowledge, it has never been published before.
598:
599: \begin{lem}
600: \label{lem:omer}Let $G$ be a weighted graph and let $v,w\in G$.
601: Let $R$ be a random walk on $G$ starting from $v$ and let $T_{n}$
602: be the $n$-th time $R$ is at $w$. Then\[
603: \LE(R[0,T_{1}])\sim\LE(R[0,T_{n}])\quad\forall n=2,3,\dotsc\]
604:
605: \end{lem}
606: The notation $\sim$ here stands, as usual, for {}``having the same
607: distribution''.
608:
609: \begin{proof}
610: Let $\gamma$ be any path starting from $v$ not containing $w$ and
611: let $k=1,\dotsc,n$. Denote $l=\len\gamma$. Define $X_{\gamma,k}$
612: to be the event that $\gamma=\LE(R[0,T_{n}])[1,l]$ and that $j_{l}\in[T_{n-k},T_{n-k+1}]$
613: where $j_{l}$ is from the definition of $\LE$ above and where we
614: consider $T_{0}$ to be $0$. Let $x$ be any neighbor of $\gamma(l)$.
615: We have that $x$ is the next element of $\LE(R[0,T_{n}])$ if and
616: only if $R(j_{l}+1)=x$. Denote this event by $N_{x}$.
617:
618: Conditioning by $X_{\gamma,k}$ we get that $R[j_{l},T_{n}]$ is a
619: random walk on $G$ starting from $\gamma(l)$ and conditioned not
620: to hit $\gamma$ before hitting $w$ for $k$ times. Therefore \[
621: \mathbb{P}(N_{x}\,|\, X_{\gamma,k})=\mathbb{P}^{\gamma(l)}(R(1)=x\,|\, T_{k}<T(\gamma)).\]
622: The point of the lemma is that the right hand side is independent
623: of $k$ --- after $T_{1}$ it is no longer possible to know anything
624: about the value of $R(1)$. Therefore\[
625: \mathbb{P}(N_{x}\,|\, X_{\gamma,k})=\mathbb{P}(N_{x}\,|\, X_{\gamma,1})\]
626: and summing over $k$ we get (denoting $X_{\gamma}=\bigcup X_{\gamma,k}$)\[
627: \mathbb{P}(N_{x}\,|\, X_{\gamma})=\sum_{k=1}^{n}\frac{\mathbb{P}(N_{x}\cap X_{\gamma,k})}{\mathbb{P}(X_{\gamma})}=\sum_{k=1}^{n}\frac{\mathbb{P}(N_{x}|X_{\gamma,1})\mathbb{P}(X_{\gamma,k})}{\mathbb{P}(X_{\gamma})}=\mathbb{P}(N_{x}\,|\, X_{\gamma,1})\]
628: and this last term is equal to the probability that $\LE(R[0,T_{1}])$
629: will, if conditioned to start from $\gamma$, will have as its next
630: vertex $x$. Indeed, this is the well known {}``Laplacian random
631: walk'' representation of loop-erased random walk, see \cite{L87}.
632: \end{proof}
633: \begin{lem}
634: \label{lem:omerB}Lemma \ref{lem:omer} holds also when the random
635: walk is conditioned not to hit a given $B\subset G$, $w\not\in B$.
636: In a formula,\[
637: \LE(R[0,T_{1}])\,|\, R[1,T_{1}]\cap B=\emptyset\sim\LE(R[0,T_{n}])\,|\, R[1,T_{n}]\cap B=\emptyset\quad\forall n=2,3,\dotsc\]
638:
639: \end{lem}
640: \noindent The proof is identical to that of the previous lemma, except
641: the random walk is conditioned to not hit $B\cup\gamma$ instead of
642: just $\gamma$.
643:
644:
645: \section{\label{sec:euclidean-nets}Euclidean nets}
646:
647:
648: \subsection{\label{sub:Background-on-RI}Background on rough isometries}
649:
650: Let $X$ and $Y$ be two metric spaces. A function $f:X\to Y$ is
651: called a \textbf{rough morphism} if \[
652: d(f(x),f(y))\leq Cd(x,y)+C\]
653: for some $C$ which depends on $f$. A \textbf{rough identity} is
654: a function $f:X\to X$ satisfying \[
655: d(f(x),x)\leq C\]
656: for some $C$ which depends on $f$. Notice that $f$ need be neither
657: one-to-one nor onto! $X$ and $Y$ would be called \textbf{roughly
658: isometric} if there exist rough morphisms $f:X\to Y$ and $g:Y\to X$
659: such that both $f\circ g$ and $g\circ f$ are rough identities. In
660: this case we call both $f$ and $g$ \textbf{rough isometries}. The
661: term was introduced by Kanai in \cite{K85}, though in more restricted
662: settings it already appeared in \cite{G81}. There are various equivalent
663: definitions in the literature, but I prefer the above {}``categorical''
664: one. A rough isometry completely ignores all local structure, and
665: in fact $\mathbb{R}^{d}$ is roughly isometric to $\mathbb{Z}^{d}$
666: and more generally, any manifold is roughly isometric to any net inside
667: it.
668:
669: To talk about rough isometry of graphs, we need to introduce a metric.
670: Let therefore $G$ be a weighted connected graph. Define \[
671: \delta(v,w):=\min_{\gamma:v\to w}\len\gamma\]
672: where the minimum is taken over all paths $\gamma$ from $v$ to $w$.
673: Clearly this makes $G$ a metric space.
674:
675: The {}``Euclidean nets'' we are going to define in the next section
676: are graphs roughly isometric to $\mathbb{R}^{d}$. Whether properties
677: of random walks are preserved under rough isometries is in general
678: not obvious. In some cases (e.g.~transience) this requires an equivalent
679: representation as a geometric property. In others (e.g.~Harnack's
680: inequality) it is actually unknown. Let us therefore state here some
681: of connections between random walks and the geometry of the graph
682: that we will use.
683:
684: \begin{defn*}
685: We say that a graph $G$ satisfies the volume doubling property if
686: there exists a constant $C$ such that for any $v\in G$ and any $r\geq1$,
687: $\omega(B(v,2r))\leq C\omega(B(v,\linebreak[1]r))$ where $\omega(A):=\sum_{x\in A}\omega(x)$,
688: $\omega$ from (\ref{eq:omega}).
689: \end{defn*}
690:
691:
692: \begin{defn*}
693: We say that a graph $G$ satisfies the weak Poincaré inequality if
694: there exists a constant $C$ such that for any function $f:G\to\mathbb{R}$,
695: any $v\in G$ and any integer $r$,\begin{gather*}
696: \sum_{w\in B(v,r)}\omega(w)\left|f(w)-\overline{f}\right|^{2}\leq Cr^{2}\sum_{w\sim x\in B(v,2r)}\omega(w,x)|f(w)-f(x)|^{2}\\
697: \overline{f}:=\frac{1}{\omega(B(v,r))}\sum_{x\in B(v,r)}\omega(x)f(x).\end{gather*}
698:
699: \end{defn*}
700: The inequality is called {}``weak'' because the sum on the right
701: hand side is over a ball of radius $2r$. The regular Poincaré inequality
702: is defined with the sum over the ball of radius $r$. However, under
703: the assumption of the volume doubling property, these properties are
704: equivalent, see \cite[\S 5]{J86} (the settings there are a little
705: different but the proof carries through literally the same). In fact,
706: the equivalence (under volume doubling) of the weak Poincaré inequality
707: under different constants $>1$ replacing the {}``$2$'' in the
708: radius of the ball is much easier and the only thing we will use:
709: this easily implies that the combination of volume doubling and weak
710: Poincaré inequality is invariant to rough isometries.
711:
712: Another common variation on this inequality is an $L^{1}$ version
713: i.e.~$\sum\omega|f(w)-\overline{f}|\leq Cr\sum\omega|f(w)-f(x)|.$
714: The $L^{1}$ version is stronger --- indeed, since the $L^{\infty}$
715: version $|f(w)-\overline{f}|\leq2r\max|f(w)-f(x)|$ is obviously always
716: true, the $L^{2}$ version follows from the $L^{1}$ version by interpolation.
717:
718:
719: \begin{defn*}
720: We say that a graph $G$ satisfies the elliptic Harnack inequality
721: if there exists a constant $C$ such that for any $v\in G$ and $r\geq1$
722: and any function $f$ harmonic and positive on $B(v,2r)$ one has\begin{equation}
723: \max\{ f(x):x\in B(v,r)\}\leq C\min\{ f(x):x\in B(v,r)\}.\label{eq:ellipt-Harnack}\end{equation}
724:
725: \end{defn*}
726:
727:
728: \begin{defn*}
729: We say that a graph $G$ satisfies the parabolic Harnack inequality
730: if there exists a constant $C$ such that for any $v\in G$ and $r\geq1$
731: and any positive function $f$ on $B(v,2r)\times[0,4r^{2}]$ satisfying
732: \begin{equation}
733: f(\cdot,t+1)-f(\cdot,t)=\Delta f(\cdot,t)\label{eq:parab-Harnack}\end{equation}
734: one has\begin{multline}
735: \max\{ f(x,t):(x,t)\in B(v,r)\times[r^{2},2r^{2}]\}\leq\\
736: C\min\{ f(x,t):(x,t)\in B(v,r)\times[3r^{2},4r^{2}]\}.\label{eq:parab-res}\end{multline}
737:
738: \end{defn*}
739: Clearly, the parabolic Harnack inequality is stronger than the elliptic
740: one. A difficulty in applying this fact is as follows: if the graph
741: is bipartite (say $\mathbb{Z}^{d}$) then the parabolic Harnack inequality
742: cannot hold --- for example, $f(x,t)=\mathbb{P}(R(t)=x)$ satisfies
743: (\ref{eq:parab-Harnack}) but the right hand side of (\ref{eq:parab-res})
744: is $0$ for any $r>1$, since $f(x,t)=0$ whenever $t+\sum x_{i}$
745: is odd. However, adding self loops will allow the graph to satisfy
746: the parabolic Harnack inequality without changing the set of harmonic
747: functions at all. After adding self-loops it is not at all easy to
748: construct examples of graphs satisfying the elliptic Harnack inequality
749: without satisfying the parabolic one. See \cite{BB99,GSC05} for some
750: constructions (see also \cite{HS01}).
751:
752: \begin{dmt}Let $G$ be an infinite connected graph and assume $\omega(v,v)>c$
753: for all $v\in G$. Then the following are equivalent
754:
755: \begin{enumerate}
756: \item $G$ satisfies the volume doubling property and the weak Poincaré
757: inequality.
758: \item $G$ satisfies the parabolic Harnack inequality.
759: \item The random walk on $G$ satisfies upper and lower Gaussian estimates,
760: namely\[
761: \frac{c}{\omega(B(v,\sqrt{t}))}e^{-C\delta(v,w)^{2}/t}\leq\mathbb{P}^{v}(R(t)=w)\leq\frac{C}{\omega(B(v,\sqrt{t}))}e^{-c\delta(v,w)^{2}/t}\]
762: for all $\delta(v,w)\leq t$.
763: \end{enumerate}
764: Further, for any two clauses, all constants in the first depend only
765: on the constants in the second.\end{dmt}
766:
767: See \cite{D99}. One of the important consequences of this theorem
768: is that the parabolic Harnack inequality is invariant to rough isometries:
769: as already remarked, the combination of the volume doubling property
770: and the Poincaré inequality is invariant to rough isometries. For
771: the elliptic Harnack inequality, the question of its invariance is
772: still open.
773:
774:
775: \subsection{\label{sub:DefinitionNet}Definition}
776:
777: A $d$-dimensional \textbf{Euclidean net} is a graph $(G,\omega)$
778: such that $G\subset\mathbb{R}^{d}$ and
779:
780: \begin{enumerate}
781: \item $G$ has bounded weight;
782: \item The inclusion $i:G\to\mathbb{R}^{d}$ is a rough isomorphism between
783: $(G,\delta)$ and $\mathbb{R}^{d}$;
784: \item $\inf\{|v-w|:v\neq w\in G\}>0$.
785: \end{enumerate}
786: We shall mostly be interested in the $\mathbb{R}^{d}$ distance on
787: $G$, which we will denote by $d(v,w)$ or $|v-w|$. Likewise, the
788: notation $B(v,r)$ will relate to a ball in the $\mathbb{R}^{d}$
789: distance, while a ball in the metric $\delta$ will be denoted by
790: $B_{\delta}$.
791:
792:
793: \subsection{Harnack's inequality}
794:
795: \begin{lem}
796: \label{lem:Harnack}A Euclidean net satisfies the elliptic Harnack
797: inequality (\ref{eq:ellipt-Harnack}) for $r$ sufficiently large.
798: \end{lem}
799: Two comments should be made on the formulation of the lemma. First
800: is that in the definition of the elliptic Harnack inequality (\ref{eq:ellipt-Harnack})
801: we mean balls in the $\mathbb{R}^{d}$ metric and not in the graph
802: metric. The second is about the constant in (\ref{eq:ellipt-Harnack}).
803: We implicitly assume that the constant $C=C(G)$ depends only on the
804: following parameters:\label{page:eucCstruct}
805:
806: \begin{enumerate}
807: \item The bounds for $\omega$;
808: \item The constants of the rough isometries $i$ and $f$ between $G$ and
809: $\mathbb{R}^{d}$;
810: \item The constants of the rough identities $i\circ f$ and $f\circ i$;
811: \item The lower bound for $|v-w|$;
812: \item $d$.
813: \end{enumerate}
814: We call the aggregation of these parameters the \textbf{Euclidean
815: net structure constants}. Whenever we use the notation $C(G)$ we
816: mean a constant depending only on these parameters. Similarly, constants
817: implicit in the notations $o,O$ and $\approx$ notation are not universal
818: but may depend on the structure constants of the Euclidean net. The
819: phrase {}``sufficiently large'' means {}``larger than a constant
820: depending on the Euclidean structure constants only''.
821:
822: In chapter \ref{sec:Isotropic-gluing} we shall apply results obtained
823: up to that point to families of graphs with uniformly bounded structure
824: constants, hence it is important that $C$ does not depend on other
825: properties of $G$.
826:
827: \begin{proof}
828: Construct an auxiliary graph $G^{*}$ with the same vertex set as
829: $G$ and with the weights defined by $\omega_{G^{*}}(v,v)=\omega_{G}(v,v)+\omega_{G}(v)$.
830: In other words, the random walk on $G^{*}$ is a random walk on $G$
831: with a probability of $\frac{1}{2}$ to stay at the same spot at each
832: step (additional to any such probability already existing for $G$).
833: The random walk on $G^{*}$ is sometimes called the lazy walk on $G$.
834: Clearly $\mathbb{Z}^{d}$ satisfies the volume doubling property and
835: it is easy to see that $\mathbb{Z}^{d}$ satisfies the weak Poincaré
836: inequality --- every group does, see e.g.~\cite[4.1.1]{PSC99} (the
837: other conditions of Delmotte's theorem are also easy to verify, if
838: you prefer). Since volume doubling and the weak Poincaré inequality
839: are preserved by rough isometries, $G^{*}$ satisfies them. Hence
840: by Delmotte's theorem it satisfies the parabolic Harnack inequality
841: with respect to the graph metric $\delta$. Hence it satisfies the
842: elliptic Harnack inequality, and since $G$ and $G^{*}$ have the
843: same harmonic functions, $G$ also satisfies the elliptic Harnack
844: inequality with respect to $\delta$.
845:
846: Now, to prove Harnack's inequality for the $\mathbb{R}^{d}$ metric,
847: cover $B(v,r)$ by a constant number of balls $B_{\delta}(w_{i},cr)$
848: such that $B_{\delta}(w_{i},4cr)\subset B(v,2r)$. It is easy to see
849: that this can be done with the number of balls uniformly bounded.
850: From the above discussion we have, for every $f$ harmonic and positive
851: on $B(v,2r)$ that \[
852: \max\{ f:B_{\delta}(w_{i},2cr)\}\leq C\min\{ f:B_{\delta}(w_{i},2cr)\}\]
853: for every $i$. Therefore if we have that $B_{\delta}(w_{j},2cr)\cap B_{\delta}(w_{j+1},2cr)\neq\emptyset$
854: for $j=1,\dotsc,k$ then we get\[
855: \max\Big\{ f:\bigcup_{j}B_{\delta}(w_{j},2cr)\Big\}\leq C^{k}\min\Big\{ f:\bigcup_{j}B_{\delta}(w_{j},2cr)\Big\}.\]
856: Now, if $r$ is sufficiently large then the balls $B_{\delta}(w_{i},2cr)$
857: form a connected graph with respect to intersection, so we get the
858: required result.
859: \end{proof}
860: \begin{lem}
861: \label{lem:Harnack-general}Let $\mathcal{E},\mathcal{D}$ be domains
862: in $\mathbb{R}^{d}$ such that $\overline{\mathcal{E}}\subset\mathcal{D}$.
863: Then Harnack's inequality holds with respect to $\mathcal{E}$ and
864: $\mathcal{D}$ i.e. for any $v\in\mathbb{R}^{d}$, any $r>C(\mathcal{E},\mathcal{D},G)$
865: and any function $f$ positive and harmonic on $(r\mathcal{D}+v)\cap G$,
866: \[
867: \max\{ f:(r\mathcal{E}+v)\cap G\}\leq C(\mathcal{E},\mathcal{D},G)\min\{ f:(r\mathcal{E}+v)\cap G\}.\]
868: Further, if $\mathcal{K}$ be a family of $(\mathcal{E},\mathcal{D})$
869: with $\diam\mathcal{E}$ bounded above and $d(\mathcal{E},\partial\mathcal{D})$
870: bounded below, then all $C(\mathcal{E},\mathcal{D},G)$ are bounded
871: by some constant $C(\mathcal{K},G)$.
872: \end{lem}
873: \begin{proof}
874: Use the same covering trick as above.
875: \end{proof}
876: \begin{rem*}
877: The fact that a graph roughly isometric to $\mathbb{R}^{d}$ satisfies
878: the elliptic Harnack inequality was known before \cite{D99}. In the
879: setting of graphs, it was proved concurrently by Delmotte \cite{D97}
880: and Holopainen-Soardi \cite{HS97} (who proved it for the $p$-Laplacian
881: for any $p$). In the setting of manifolds this goes back to Kanai
882: \cite{K85}, who proved that a manifold roughly isometric to $\mathbb{R}^{d}$
883: satisfies the (continuous) Harnack inequality by showing that it follows
884: from a $d$-dimensional isoperimetric inequality.
885: \end{rem*}
886:
887: \subsection{Green's function}
888:
889: Let $H$ be any graph (possibly directed). Let $v,w\in H$ and $S\subset H$.
890: Then Green's function with respect to $S$ is defined by\[
891: G(v,w;S)=\sum_{t=0}^{\infty}\mathbb{P}^{v}(R(t)=w,\, R[0,t]\subset S)\]
892: or in other words, the expected number of visits to $w$ before leaving
893: $S$. If $S=H$ we shall omit it in the notation and write $G(v,w)$.
894: In general there is nothing forcing $G$ to be finite.
895:
896: If $G$ is finite then it is zero outside $S$ and inside $S$ satisfies\begin{equation}
897: \Delta G(\cdot,w;S)=-\delta_{w}\label{eq:GreenLap}\end{equation}
898: i.e.~$G$ is harmonic on $S\setminus\{ w\}$ and $\Delta G(w,w)=-1$.
899: These conditions uniquely determine $G(v,w;S)$. The symmetry of random
900: walk (\ref{eq:sym}) translates to a symmetry of $G$ in the form\[
901: \omega(v)G(v,w;S)=\omega(w)G(w,v;S).\]
902:
903:
904: \begin{lem}
905: \label{lem:a}Let $H$ be a $d$-dimensional Euclidean net. Then
906: \begin{enumerate}
907: \item If $d=2$ then $H$ is recurrent and Green's function satisfies\begin{equation}
908: G(v,w;S)\leq C(H)\log\diam S.\label{eq:arecur}\end{equation}
909:
910: \item If $d\geq3$ then $H$ is transient and Green's function satisfies\begin{equation}
911: \begin{split}G(v,w;S) & \leq C(H)|v-w|^{2-d}\quad\forall v\neq w.\\
912: G(v,v;S) & \leq C(H).\end{split}
913: \label{eq:atrans}\end{equation}
914: If $B(x,2r)\subset S$ and $r$ is sufficiently large then inside
915: $B(x,r)$ a lower bound also holds, \begin{equation}
916: \begin{split}G(v,w;S) & \approx|v-w|^{2-d}\quad\forall v\neq w\in B(x,r)\\
917: G(v,v;S) & \approx1.\end{split}
918: \label{eq:GvwS-lower-bound}\end{equation}
919:
920: \end{enumerate}
921: \end{lem}
922: \begin{rem*}
923: For $d=2$ we have (from recurrence) that $G(v,w)=\infty$ for all
924: $v$ and $w$. The natural analog of $G(v,w)$ in this case is the
925: harmonic potential of $H$ defined by $a(v,w)=\lim_{r\to\infty}G(v,v;B(v,r))-G(v,w;B(v,r))$.
926: It is possible to show that for any two dimensional Euclidean net
927: $a(v,w)$ is well defined and $a(v,w)\approx\log|v-w|$, but we will
928: have no use for this fact.
929: \end{rem*}
930: \begin{proof}
931: We start with the case of $d\geq3$. Let $H^{*}$ be the lazy version
932: of $H$ as in the proof of lemma \ref{lem:Harnack}. By Delmotte's
933: theorem,\[
934: ct^{-d/2}e^{-C|v-w|^{2}/t}\leq\mathbb{P}_{H^{*}}^{v}(R(t)=w)\leq Ct^{-d/2}e^{-c|v-w|^{2}/t}.\]
935: Summing this over all $t$ we get \[
936: G_{H^{*}}(v,w)\approx|v-w|^{2-d}.\]
937: Now, $G_{H}=\frac{1}{2}G_{H^{*}}$ because one may couple the walks
938: on $H$ and $H^{*}$ so that each step of $H$ the walker on $H^{*}$
939: walks the same step and then waits for an expected time of $1$. Therefore
940: we get $G_{H}(v,w)\approx|v-w|^{2-d}$. We will not need the graph
941: $H^{*}$ again, so all Green functions henceforth are with respect
942: to $H$.
943:
944: Now, $G(v,w;S)\leq G(v,w)$ gives us (\ref{eq:atrans}). To get the
945: lower bound under the assumption $B(x,2r)\subset S$ take $f=f_{S,w}$
946: to be a harmonic function on $S$ with $f(v)=G(v,w)$ for all $v\in\partial S$.
947: $G(\cdot,w)-f$ will satisfy (\ref{eq:GreenLap}) which defines $G(\cdot,w;S)$
948: so they are equal. By the maximum principle we get\[
949: f(v)\leq\max_{y\in\partial S}f(y)\leq Cr^{2-d}\quad\forall v\in S\]
950: so we get $G(v,w)\approx|v-w|^{2-d}$ inside a ball $B(v,\lambda r)$
951: for some constant $\lambda=\lambda(G)$ sufficiently small. Using
952: Harnack's inequality (lemma \ref{lem:Harnack-general}) for the domains
953: $B(0,1)\setminus B(0,\lambda)\subset B(0,2)\setminus B(0,\frac{1}{2}\lambda)$
954: proves (\ref{eq:GvwS-lower-bound}).
955:
956: The two dimensional case follows from electrical resistance arguments.
957: See \cite{S94} for background on this topic. The maximum principle
958: shows that $G(v,w)\leq G(v,v)$ and the latter is equal to the resistance
959: between $v$ and $\partial S$. The electrical resistance is preserved
960: (up to a constant) by rough isometries, and so we get \[
961: G(v,v;S)\leq G(v,v;B(v,2\diam S))\approx G_{\mathbb{Z}^{2}}(0,0;B(0,2\diam S))\approx\log\diam S.\qedhere\]
962:
963: \end{proof}
964: \begin{rem*}
965: The use of Delmotte's theorem here is somewhat of an overkill. The
966: fact that the a graph roughly isometric to $\mathbb{R}^{d}$ has a
967: $d$-dimensional heat kernel decay follows essentially from Varopoulos
968: \cite{V85}. To get an estimate for the Green function one can apply
969: e.g.~Hebisch and Saloff-Coste \cite{HS93} which gives a square exponential
970: decay upper bound.
971: \end{rem*}
972: \begin{lem}
973: \label{lem:r2}Let $G$ be a Euclidean net and let $v\in G$ and $r>1$.
974: Then for some constant $\lambda(G)$ sufficiently large,\[
975: \mathbb{P}^{w}(R(\left\lfloor \lambda r^{2}\right\rfloor )\in B(v,r))\leq\frac{1}{2}\quad\forall w.\]
976:
977: \end{lem}
978: \begin{proof}
979: Let $G^{*}$ be the lazy version of $G$, as in lemma \ref{lem:Harnack}.
980: Again we use Delmotte's theorem to show that for any $\lambda$, \[
981: \mathbb{P}_{G^{*}}^{w}(R(n)=x)\leq C\left\lfloor \lambda r^{2}\right\rfloor ^{-d/2}\quad\forall w,\forall n\geq\left\lfloor \lambda r^{2}\right\rfloor .\]
982: We have that $\# B(v,r)\leq C(G)r^{d}$. Hence summing gives that
983: for $\lambda$ a constant sufficiently large \[
984: \mathbb{P}_{G^{*}}^{w}(R(n)\in B(v,r))\leq\frac{1}{2}\quad\forall w,\forall n\geq\left\lfloor \lambda r^{2}\right\rfloor .\]
985: But the coupling between the walk and the lazy walk shows that after
986: the walk did $n$ steps the lazy walk did at least $n$ steps. Hence\begin{align*}
987: \lefteqn{\mathbb{P}_{G}^{w}(R(\left\lfloor \lambda r^{2}\right\rfloor )\in B(v,r))=}\qquad\\
988: & =\sum_{n\geq\left\lfloor \lambda r^{2}\right\rfloor }\mathbb{P}(\textrm{the lazy walk did }n\textrm{ steps and is in }B(v,r))=\\
989: & =\sum_{n\geq\left\lfloor \lambda r^{2}\right\rfloor }\mathbb{P}(\textrm{the lazy walk did }n\textrm{ steps})\mathbb{P}_{G^{*}}(R(n)\in B(v,r))\leq\\
990: & \leq\frac{1}{2}\sum_{n\geq\left\lfloor \lambda r^{2}\right\rfloor }\mathbb{P}(\textrm{the lazy walk did }n\textrm{ steps})=\frac{1}{2}.\qedhere\end{align*}
991:
992: \end{proof}
993: Here too Delmotte's theorem can be replaced by Varopoulos \cite{V85}.
994:
995: The following lemma basically states that a random walk has positive
996: probability to hit large objects. We will only use the lemma for simple
997: domains with piecewise smooth boundary, so the requirements of clause
998: \ref{enu:lemDD-K} will always be satisfied.
999:
1000: \begin{lem}
1001: \label{lem:DD}Let $H$ be a $d$-dimensional Euclidean net.\newC{C:DD}\newc{c:DDi}\newc{c:DDo}
1002: \begin{enumerate}
1003: \item \label{enu:lemDD-hit}Let $\mathcal{D}$, $\mathcal{S}$ (start) and
1004: $\mathcal{H}$ (hit) be domains in $\mathbb{R}^{d}$ with $\overline{\mathcal{S}},\mathcal{H}\subset\mathcal{D}$,
1005: $\mathcal{H}\neq\mathcal{D}$. Then there exists a $C_{\ref{C:DD}}(\mathcal{D},\mathcal{S},\mathcal{H},H)$
1006: such that for all $r>C_{\ref{C:DD}}$; all $v\in H$ and all $w\in\left(v+r\mathcal{S}\right)\cap H$,
1007: if $R$ is a random walk starting from $w$ then\begin{equation}
1008: \mathbb{P}(T(\partial(v+r\mathcal{H}))<T(\partial(v+r\mathcal{D})))>c_{\ref{c:DDi}}(\mathcal{D},\mathcal{S},\mathcal{H},H).\label{eq:hitinside}\end{equation}
1009:
1010: \item \label{enu:lemDD-avoid}If $\overline{\mathcal{S}}\cap\overline{\mathcal{H}}=\emptyset$
1011: and $\mathcal{S}$ is a subset of the unbounded component of $\mathbb{R}^{d}\setminus\overline{\mathcal{H}}$
1012: then in addition\begin{equation}
1013: \mathbb{P}(T(\partial(v+r\mathcal{H}))\geq T(\partial(v+r\mathcal{D})))>c_{\ref{c:DDo}}(\mathcal{D},\mathcal{S},\mathcal{H},H).\label{eq:hitoutside}\end{equation}
1014:
1015: \item \label{enu:lemDD-K}Let $\mathcal{K}$ is a family of triplets $(\mathcal{D},\mathcal{S},\mathcal{H})$
1016: such that for every $x\in\mathcal{S}$ there exists a path $\gamma$
1017: with $\len\gamma$ bounded above leading from $x$ to $\partial_{\textrm{cont}}\mathcal{H}$
1018: with $d(\gamma,\partial\mathcal{D})$ bounded below (case \ref{enu:lemDD-hit})
1019: or from $x$ to $\partial\mathcal{D}$ with $d(\gamma,\partial_{\textrm{cont}}\mathcal{H})$
1020: bounded below (case \ref{enu:lemDD-avoid}). Then $C_{\ref{C:DD}},c_{\ref{c:DDi}}$
1021: and $c_{\ref{c:DDo}}$ are bounded on $\mathcal{K}$.
1022: \end{enumerate}
1023: \end{lem}
1024: \begin{proof}
1025: Let us start with (\ref{eq:hitinside}). Assume first that $\mathcal{D}=B(0,1)$,
1026: $\mathcal{H}=B(0,\frac{1}{22})$ and $\mathcal{S}=B(0,\frac{1}{2})\setminus B(0,\frac{2}{22})$,
1027: and that $d\geq3$. Examine the Green function $G(w)=G(v,w;v+r\mathcal{D})$.
1028: Let $w\in(v+r\mathcal{S})\cap H$, let $R$ be a random walk starting
1029: from $w$ and let $T$ be its stopping time on $\partial(v+r\mathcal{H})\cup\partial(v+r\mathcal{D})$.
1030: Then, since $G$ is harmonic on $(v+r\mathcal{D})\setminus(v+r\mathcal{H})$
1031: we get (from (\ref{eq:fEfRT})) that\[
1032: G(w)=\mathbb{E}G(R(T)).\]
1033: Denote $p=\mathbb{P}(R(T)\in\partial(v+r\mathcal{H}))$ which is the
1034: probability we want to estimate. Now for every $x\in\partial(v+r\mathcal{D})$,
1035: $G(x)=0$ while for $x\in\partial(v+r\mathcal{H})$ we have from (\ref{eq:atrans})
1036: that $G(x)\leq C(H)r^{2-d}$. At $w$ itself we have from (\ref{eq:GvwS-lower-bound})
1037: that $G(w)\geq cr^{2-d}$ for some $c(H)$ and $r$ sufficiently large.
1038: We get\begin{equation}
1039: cr^{2-d}\leq G(w)=\mathbb{E}G(R(T))\leq p\cdot Cr^{2-d}\label{eq:1213to113}\end{equation}
1040: so $p\geq c$ for $r$ sufficiently large.
1041:
1042: To see that the same holds for $d\leq2$, construct an auxiliary graph
1043: $H'=H\times\mathbb{Z}^{3-d}$ weighted so that the projection of the
1044: random walk on $H'$ on $H$ is (a time change of) the random walk
1045: on $H$. Then the fact that there is a positive probability to hit
1046: $\partial B(v,\frac{1}{22}r)\subset H'$ before hitting $\partial B(v,r)\subset H'$
1047: immediately implies the same for $H$.
1048:
1049: We now consider general domains. Let $x\in\mathcal{S}\setminus\mathcal{H}$
1050: and choose an $\epsilon>0$ and a sequence of points $\left\{ x_{i}\right\} _{i=0}^{n}$
1051: with the following properties:
1052: \begin{enumerate}
1053: \item $x_{0}=x$ and $B(x_{n},\epsilon)\subset\mathcal{H}$.
1054: \item $B(x_{i},2\epsilon)\subset\mathcal{D}$.
1055: \item $|x_{i+1}-x_{i}|\in\left]\frac{6}{22}\epsilon,\frac{7}{22}\epsilon\right[$.
1056: \end{enumerate}
1057: It is a simple exercise to show that $\epsilon$ and $x_{i}$ can
1058: always be chosen, and furthermore that both $\epsilon$ and $n$ may
1059: be bounded on all $\mathcal{S}\setminus\mathcal{H}$ and in case \ref{enu:lemDD-K}
1060: on all $\mathcal{K}$ i.e.~for any $x$ in any $\mathcal{S}\setminus\mathcal{H}$
1061: such that $(\mathcal{D},\mathcal{S},\mathcal{H})\in\mathcal{K}$ we
1062: have $\epsilon(x,\mathcal{H},\mathcal{D})>c(\mathcal{K})$ and $n(x,\mathcal{H},\mathcal{D})<C(\mathcal{K})$.
1063: Assume now that $r$ is sufficiently large such that the following
1064: are satisfies:
1065: \begin{enumerate}
1066: \item \label{enu:every-ball-point}every ball of radius $\frac{1}{22}\epsilon r$
1067: contains at least one point of $H$;
1068: \item $\partial B(w,s)\subset B(w,s+\frac{1}{22}\epsilon r)$ for any $s>0$
1069: and $w\in H$;
1070: \item $\epsilon r$ is sufficiently large so as to satisfy (\ref{eq:1213to113}).
1071: \end{enumerate}
1072: Clearly this is an assumption of the type $r>C(\mathcal{D},\mathcal{S},\mathcal{H},H)$.
1073: Condition \ref{enu:every-ball-point} allows to choose a point $w_{i}\in H\cap B(v+rx_{i},\frac{1}{22}\epsilon r)$
1074: for every $i$. Let $T_{0}=0$ and \[
1075: T_{i}=\min\{ t>T_{i-1}:R(t)\in\partial B(w_{i},{\textstyle \frac{1}{22}}\epsilon r)\}.\]
1076: We wish to use the case already established for the portion of the
1077: random walk after $T_{i-1}$, with the radius being $\epsilon r$
1078: instead of $r$ and the center of the balls being $w_{i}$ instead
1079: of $v$. We may do this because $R(T_{i-1})\in\partial B(w_{i-1},\frac{1}{22}\epsilon r)\subset B(w_{i-1},\frac{2}{22}\epsilon r)\subset B(v+rx_{i-1},\frac{3}{22}\epsilon r)\subset B(v+rx_{i},\frac{10}{22}\epsilon r)\setminus B(v+rx_{i},\frac{3}{22}\epsilon r)\subset B(w_{i},\frac{1}{2}\epsilon r)\setminus B(w_{i},\frac{2}{22}\epsilon r)$.
1080: However, since $B(w_{i},\epsilon r)\subset B(v+rx_{i},2\epsilon r)\subset v+r\mathcal{D}$
1081: then not hitting $\partial B(w_{i},\epsilon r)$ means staying inside
1082: $v+r\mathcal{D}$ and so we get\[
1083: \mathbb{P}\big(T_{i}<T(\partial(v+r\mathcal{D}))\,|\, T_{i-1}<T(\partial(v+r\mathcal{D}))\big)>c(H).\]
1084: This immediately gives\[
1085: p\stackrel{(*)}{\geq}\mathbb{P}(T_{n}<T(\partial(v+r\mathcal{D})))>c^{n}\]
1086: where $(*)$ comes from the fact that $\partial B(w_{n},\frac{1}{22}\epsilon r)\subset B(v+rx_{n},\frac{3}{22}\epsilon r)\subset v+r\mathcal{H}$.
1087: This proves the direction (\ref{eq:hitinside}) for $x\not\in\mathcal{H}$.
1088: The case $x\in\mathcal{H}$ is proved identically but taking the $x_{i}$-s
1089: from $x$ to $\mathcal{D}\setminus\mathcal{H}$. For the direction
1090: (\ref{eq:hitoutside}) take the $x_{i}$-s outside, i.e.~with $B(x_{n},\epsilon)\cap\mathcal{D}=\emptyset$
1091: and $B(x_{i},2\epsilon)\cap\mathcal{H}=\emptyset$.
1092: \end{proof}
1093: \begin{lem}
1094: \label{lem:DDunbounded}With the notations of the previous lemma,
1095: if $d\geq3$ then (\ref{eq:hitinside}) and (\ref{eq:hitoutside})
1096: hold even if $\mathcal{D}$ is allowed to be unbounded. If $d\leq2$
1097: then only (\ref{eq:hitinside}) holds for unbounded $\mathcal{D}$.
1098: \end{lem}
1099: \begin{proof}
1100: The only part not following directly from lemma \ref{lem:DD} is the
1101: proof of (\ref{eq:hitoutside}) when $d\geq3$. We start with the
1102: case that $\mathcal{D}=\mathbb{R}^{d}$ (so $T(\partial(v+r\mathcal{D}))$
1103: is always $\infty$), $\mathcal{H}=B(0,1)$ and $\mathcal{S}\cap B(0,\lambda)=\emptyset$
1104: where $\lambda(G)>2$ is some constant sufficiently large that will
1105: be fixed later. Let $s>\lambda r$ and denote \[
1106: p(s)=\mathbb{P}^{w}(T_{v,s}<T_{v,r})\quad w\in v+r\mathcal{S}.\]
1107: Let $G(x)=G(v,x;B(v,s))$ i.e.~Green's function. From (\ref{eq:GvwS-lower-bound})
1108: we see that for $x\in\partial B(v,r)$ we have $G(x)\geq c(H)r{}^{2-d}$
1109: if $r$ is sufficiently large. At $w$ itself we have $G(w)\leq C(H)(\lambda r)^{2-d}$
1110: and by definition $G|_{\partial B(v,s)}\equiv0$ so\[
1111: C(\lambda r)^{2-d}\geq G(w)=\mathbb{E}G(R(T))\geq(1-p)c(H)r^{2-d}\]
1112: so $p\geq1-C\lambda^{2-d}$ and for $\lambda$ sufficiently large
1113: this would be $\geq\frac{1}{2}$. Fix $\lambda$ to be some such constant.
1114: We get that $\lim_{s\to\infty}p\geq\frac{1}{2}$. Hence\begin{equation}
1115: \mathbb{P}(T(\partial(v+r\mathcal{H}))=\infty)=\lim_{s\to\infty}\mathbb{P}(T(\partial(v+r\mathcal{H}))>T_{v,s})\geq{\textstyle \frac{1}{2}}\label{eq:infcG}\end{equation}
1116: and this case is finished. For general $\mathcal{D}$, $\mathcal{S}$
1117: and $\mathcal{H}$, let $B(0,\rho)$ ($\rho>1$) be a ball sufficiently
1118: large such that \emph{$\mathcal{D}\cap B(0,\rho)$} contains $\mathcal{S}$,
1119: $\mathcal{H}$ and a path between them. Let $\mathcal{D}_{2}$ be
1120: the component of $\mathcal{D}\cap B(0,\lambda\rho)$ containing $\mathcal{S}$
1121: and $\mathcal{H}$. Then lemma \ref{lem:DD} shows that for $r$ sufficiently
1122: large we have \begin{equation}
1123: \mathbb{P}(T(\partial(v+r\mathcal{H}))\geq T(\partial(v+r\mathcal{D}_{2})))>c(\mathcal{D}_{2},\mathcal{S},\mathcal{H},G).\label{eq:HD2c}\end{equation}
1124: If $R(T(\partial(v+r\mathcal{D}_{2})))\not\in\partial(v+r\mathcal{D})$
1125: then it is $\in\partial B(v,\lambda\rho r)$ and then the previous
1126: case (rescaled by $\rho$) with the strong Markov property shows that
1127: \begin{eqnarray*}
1128: \lefteqn{\mathbb{P}(T(\partial(v+r\mathcal{H}))\geq T(\partial(v+r\mathcal{D}))\,|\, T(\partial(v+r\mathcal{H})\geq T(\partial(v+r\mathcal{D}_{2})))=}\\
1129: & & =\mathbb{P}(T(\partial(v+r\mathcal{H}))\geq T(\partial(v+r\mathcal{D})),\, T(\partial(v+r\mathcal{D}_{2}))=T_{v,\lambda\rho r}\,|\,\\
1130: & & \qquad\qquad T(\partial(v+r\mathcal{H})\geq T(\partial(v+r\mathcal{D}_{2})))=\\
1131: & & =\mathbb{E}\mathbb{P}^{R(T_{v,\lambda\rho r})}(T(\partial(v+r\mathcal{H}))\geq T(\partial(v+r\mathcal{D})))\geq\\
1132: & & \geq\mathbb{E}\mathbb{P}^{R(T_{v,\lambda\rho r})}(T_{v,\rho r}=\infty)\stackrel{(\ref{eq:infcG})}{\geq}\mathbb{E}c(G)=c(G)\end{eqnarray*}
1133: and together with (\ref{eq:HD2c}) we get (\ref{eq:hitoutside}).\newC{C:DDp}
1134: \end{proof}
1135: \begin{lem}
1136: \label{lem:DDp}Let $\mathcal{D}$, $\mathcal{S}$ and $\mathcal{H}$
1137: be domains in $\mathbb{R}^{d}$ with $\overline{\mathcal{S}},\overline{\mathcal{H}}\subset\mathcal{D}$
1138: and $\overline{\mathcal{S}}\cap\overline{\mathcal{H}}=\emptyset$.
1139: Then there exists a $C_{\ref{C:DDp}}(\mathcal{D},\mathcal{S},\mathcal{H},G)$
1140: such that for all $r>C_{\ref{C:DDp}}$; all $v\in G$; all $w\in\left(v+r\mathcal{S}\right)\cap G$
1141: and all $x\in(v+r\mathcal{H})\cap G$, if $R$ is a random walk starting
1142: from $w$ then\[
1143: \mathbb{P}(T(\{ x\})<T(\partial(v+r\mathcal{D})))\approx\begin{cases}
1144: r^{2-d} & d\geq3\\
1145: 1/\log r & d=2\end{cases}\]
1146: where the constants implicit in the $\approx$ may depend on $\mathcal{D},\mathcal{S},\mathcal{H}$
1147: and $G$. Further, if $\mathcal{K}$ is family of $(\mathcal{D},\mathcal{H},\mathcal{S})$
1148: triplets satisfying the conditions of lemma \ref{lem:DD}, clause
1149: \ref{enu:lemDD-K} then $C_{\ref{C:DDp}}$ and the implicit constants
1150: are bounded on $\mathcal{K}$.
1151: \end{lem}
1152: \begin{proof}
1153: Let $\epsilon=\epsilon(\mathcal{K},G)$ be sufficiently small. Use
1154: lemma \ref{lem:DD} to show that the probability to hit a ball of
1155: radius $r\epsilon$ around $x$ is $\approx1$ and then the same Green's
1156: function calculations as in that lemma to show that the probability
1157: to hit a point before exiting from a ball containing $\mathcal{D}$
1158: are $\approx r^{2-d}$ for $d\geq3$ and $\approx1/\log r$ for $d=2$.
1159: \end{proof}
1160: \begin{lem}
1161: \label{lem:hitbsame}Let $G$ be a $d$-dimensional Euclidean net,
1162: $d\geq3$. Let $v\in G$, $s>4r>C(G)$, let $A\subset B(v,r)$ and
1163: $w\in\partial B(v,2r)$. Then\[
1164: \mathbb{P}^{w}(T(A)<T_{v,4r})\approx\mathbb{P}^{w}(T(A)<T_{v,s}).\]
1165: Further, if $B\subset A$ then\[
1166: \mathbb{P}^{w}(R(T(A\cup\partial B(v,4r)))\in B)\approx\mathbb{P}^{w}(R(T(A\cup\partial B(v,s)))\in B).\]
1167:
1168: \end{lem}
1169: \begin{proof}
1170: We shall only show the first estimate, the second one is proved identically.
1171: Clearly $\mathbb{P}^{w}(T(A)<T_{v,4r})\leq\mathbb{P}^{w}(T(A)<T_{v,s})$,
1172: so we need to show the other direction. Define stopping times $T_{0}=0$
1173: and\begin{align*}
1174: T_{2i+1} & :=\min\{ t>T_{2i}:R(t)\in\partial B(v,4r)\cup A)\}\\
1175: T_{2i} & :=\min\{ t>T_{2i-1}:R(t)\in\partial B(v,2r)\cup\partial B(v,s)\}.\end{align*}
1176: Let $I$ be the first time when $R(T_{I})\in A$ (for $i$ odd) or
1177: $R(T_{i})\in\partial B(v,s)$ (for $i$ even). We consider the process
1178: stopped at $I$. From lemma \ref{lem:DDunbounded} we see that \begin{equation}
1179: \mathbb{P}(R(T_{2i})\in\partial B(v,s)\,|\, I>2i-1)\geq\mathbb{EP}^{R(T_{2i-1})}(T_{v,2r}=\infty)\geq c(G).\label{eq:RT2iexp}\end{equation}
1180: From Harnack's inequality we get that\begin{multline*}
1181: \mathbb{P}^{w}(R(T_{2i+1})\in A\,|\, I>2i)=\mathbb{E}\mathbb{P}^{R(T_{2i})}(T(A)<T_{v,4r})\approx\\
1182: \approx\min_{x\in\partial B(v,2r)}\mathbb{P}^{x}(T(A)<T_{v,4r})\end{multline*}
1183: and hence this is (up to a constant) independent of $i$. Hence we
1184: get\begin{align*}
1185: \mathbb{P}^{w}(T(A)<T_{v,s}) & =\sum_{i=0}^{\infty}\mathbb{P}^{w}(I>2i,\, R(T_{2i+1})\in A)\leq\\
1186: & \leq C(G)\sum_{i=0}^{\infty}\mathbb{P}^{w}(R(T_{1})\in A)\mathbb{P}(I>2i)\leq\\
1187: & \!\!\stackrel{(\ref{eq:RT2iexp})}{\leq}C(G)\sum_{i=0}^{\infty}\mathbb{P}^{w}(R(T_{1})\in A)(1-c(G))^{i}=\\
1188: & =C(G)\mathbb{P}^{w}(T(A)<T_{v,4r}).\qedhere\end{align*}
1189:
1190: \end{proof}
1191:
1192: \subsection{\label{sub:Beurling}The discrete Beurling projection in three dimensions}
1193:
1194: The Beurling projection theorem says that the probability of a two
1195: dimensional Brownian motion starting at $0$ to hit a given set $K\subset B(0,1)$
1196: before hitting $\partial B(0,1)$ is larger than the probability to
1197: hits its \emph{angular projection}, namely the set $\{|z|:z\in K\}$.
1198: In particular, if $K$ is connected and intersects both $\partial B(0,\epsilon)$
1199: and $\partial B(0,1)$ then the probability to avoid it is maximal
1200: when $K=[\epsilon,1]$, and in this case it may be calculated explicitly
1201: from the conformal invariance of Brownian motion and is $\approx\sqrt{\epsilon}$.
1202: A discrete version of this result (up to constants) was achieved by
1203: Kesten \cite{K87}. In this section we shall prove a three dimensional
1204: variation on this result, namely the following lemma:
1205:
1206: \begin{lem}
1207: \label{lem:Beurling}Let $H$ be a three dimensional Euclidean net.
1208: Then there exists a constant $C(H)$ such that for all $v\in H$,
1209: for all $r>C(H)$ and for all connected sets $A\subset H$ that intersect
1210: both $B(v,r)$ and $H\setminus B(v,2r)$ one have\[
1211: \mathbb{P}^{v}(R[0,T_{v,4r}]\cap A\neq\emptyset)\geq\frac{c}{\log r}.\]
1212:
1213: \end{lem}
1214: While Kesten's version of Beurling's arguments may be applied to three
1215: dimensions without much change, in the setting of lemma \ref{lem:Beurling}
1216: the notion of capacity, particularly of Martin capacity, can be used
1217: to shorten the argument significantly. We shall first give the relevant
1218: definitions, and the proof will follow after.
1219:
1220:
1221: \subsubsection{Martin capacity}
1222:
1223: \begin{defn*}
1224: Let $H$ be a countable set and let $K(v,w)$ be some function ({}``the
1225: kernel''). The capacity of a set $S\subset H$ with respect to $K$
1226: is defined by \[
1227: \capa_{K}(S):=\Big(\inf_{\mu(S)=1}\int_{S}\int_{S}K(v,w)\, d\mu(v)\, d\mu(w)\Big)^{-1}.\]
1228: The infimum here is over all probability measures $\mu$ supported
1229: on $S$.
1230: \end{defn*}
1231:
1232:
1233: \begin{defn*}
1234: Let $H$ be a directed graph and let $v\in H$. Then the \textbf{Martin
1235: capacity} of the graph with respect to $v$ is the capacity with respect
1236: to the Martin kernel, defined by\[
1237: K(w,x):=\frac{G(w,x)}{G(v,x)}.\]
1238: where $G$ is Green's function.
1239: \end{defn*}
1240: \begin{bpp}Let $H$ be a directed graph, let $v\in H$ and let $S\subset H$
1241: satisfy that Green's function $G(w,x)$ is finite for all $w,x\in S$.
1242: Let $\capa$ be the Martin capacity with respect to $v$. Then for
1243: any $S$ we have\[
1244: {\textstyle \frac{1}{2}}\capa(S)\leq\mathbb{P}^{v}(R\left[0,\infty\right[\cap S\neq\emptyset)\leq\capa(S).\]
1245: \end{bpp}The nice and simple proof may be found in \cite{BPP95},
1246: theorem 2.2.
1247:
1248: \begin{proof}
1249: [Proof of lemma \ref{lem:Beurling}]Let $\lambda=\lambda(H)$ be some
1250: constant such that every edge of $H$ has length $\leq\lambda$. Then
1251: $A$ intersects every spherical shell $B(v,s+\lambda)\setminus B(v+s)$,
1252: $r\leq s\leq2r$. Let $a_{i}\in A\cap(B(v,r+(i+1)\lambda)\setminus B(v,r+i\lambda))$
1253: for $i=0,\dotsc,\left\lfloor r/\lambda\right\rfloor $. Let $A^{*}=\{ a_{i}\}$.
1254: The lemma will be proved if we show\[
1255: \mathbb{P}^{v}(T(A^{*})<T_{v,4r})\geq\frac{c}{\log r}.\]
1256: Let $H'$ be the directed graph given by taking $\overline{H\cap B(v,4r)}$
1257: and making each point of $\partial B(v,4r)$ a {}``sink'' i.e. a
1258: point with the only exit being a self loop. By definition, $G_{H'}(w,x)=G_{H}(w,x;B(v,4r))$.
1259: We get the equivalent formulation $\mathbb{P}_{H'}^{v}(T(A^{*})<\infty)\geq c/\log r$.
1260: We now use Benjamini-Pemantle-Peres on $H'$. We get that it is enough
1261: to estimate $\capa(A^{*})\geq c/\log r$. By the definition of capacity
1262: we that we need to show that there exists a $\mu$ on $A^{*}$ such
1263: that \begin{equation}
1264: \int_{A^{*}}\int_{A^{*}}\frac{G(w,x)}{G(v,x)}d\mu(w)d\mu(x)\leq C\log r.\label{eq:req-mu}\end{equation}
1265: Let $\mu$ be the uniform measure on $A^{*}$. Then by (\ref{eq:GvwS-lower-bound})
1266: we have that \[
1267: K(a_{i},a_{j})=\frac{G(a_{i},a_{j})}{G(v,a_{j})}\leq C\frac{|a_{i}-a_{j}|^{-1}}{r^{-1}}\leq Cr|i-j|^{-1}.\]
1268: Summing gives (\ref{eq:req-mu}) and the lemma.
1269: \end{proof}
1270:
1271: \subsection{Intersection probabilities}
1272:
1273: \begin{lem}
1274: \label{lem:Rintrsct}Let $H$ be a Euclidean net of dimension $\leq3$
1275: and let $\epsilon\in\left]0,1\right[$. Let $v\in H$ and $r>0$ and
1276: let $R^{1}$ and $R^{2}$ be random walks starting from vertices $v^{1}$
1277: and $v^{2}$, $v^{i}\in B(v,(1-\epsilon)r)$. Then\newC{C:minr}\newc{c:intrsct}\[
1278: \mathbb{P}(R^{1}[0,T_{v^{1},\epsilon r}^{1}]\cap R^{2}[0,T_{v,r}^{2}]\neq\emptyset)>c_{\ref{c:intrsct}}(\epsilon,H)\]
1279: if only $r>C_{\ref{C:minr}}(\epsilon,H)$.
1280: \end{lem}
1281: \begin{proof}
1282: We shall only show the case $d=3$ --- the case $d=2$ is identical
1283: and will be left to the reader. Let $\lambda=\lambda(\epsilon,H)$
1284: be some constant whose value will be fixed later. Let $s\geq\epsilon\lambda$
1285: be some number, $w\in H$ and $\delta>0$. For any $x\in B(w,(1-\delta)s$),
1286: let $\Gamma(y)$ be the probability that a random walk starting from
1287: $y$ will hit $x$ before hitting $\partial B(w,s)$, and define $\Gamma(x):=1$.
1288: $\Gamma$ is harmonic on $B(w,s)\setminus\{ x\}$ and $0$ on $\partial B(w,s)$.
1289: Hence $\Gamma(y)=G(y,x;B(w,s))/G(x,x;B(w,s))$ where $G$ is Green's
1290: function. Using (\ref{eq:atrans}), (\ref{eq:GvwS-lower-bound}) and
1291: Harnack's inequality (lemma \ref{lem:Harnack-general}) we get\begin{align*}
1292: \Gamma(y) & \geq c(\delta,H)/|y-x| & y & \in B(w,(1-\delta)s)\setminus\{ x\}\\
1293: \Gamma(y) & \leq C(\delta,H)/|y-x| & y & \in B(w,s)\setminus\{ x\}.\end{align*}
1294: for $s$ sufficiently large.
1295:
1296: Define now $A:=B(v^{1},\frac{1}{2}\epsilon r)\setminus B(v^{1},\frac{1}{4}\epsilon r)$
1297: and $X:=A\cap R^{1}[0,T_{v^{1},\epsilon r}^{1}]\cap R^{2}[0,T_{v,r}^{2}]$.
1298: For any $x\in A$, the preceding calculation (used once for $w=v^{1}$,
1299: $s=\epsilon r$ and $\delta=\frac{1}{2}$ and a second time for $w=v^{2}$,
1300: $s=r$ and $\delta=\frac{1}{2}\epsilon$) shows that $\mathbb{P}(x\in X)\approx r^{-2}$.
1301: The $\approx$ sign here and below may depend on $\epsilon$ and on
1302: the Euclidean net structure constants of $H$. Rough isometry preserves
1303: (up to a constant) volumes of balls and shells, hence if $\epsilon r$
1304: is sufficiently large we get $\# A\approx r^{3}$ and hence $\mathbb{E}\# X\approx r$.
1305: Next we want to calculate $\mathbb{E}(\# X)^{2}$. For any $x\neq y\in A$
1306: we have\begin{equation}
1307: \mathbb{P}(x,y\in R^{i}[0,T^{i}])\approx r^{-1}|x-y|^{-1}.\label{eq:probxy}\end{equation}
1308: Indeed, this probability is $\geq$ the probability to hit $x$ first
1309: (which is $\approx1/r$) and then to hit $y$ (which is $\approx1/|x-y|$).
1310: On the other hand it is $\leq$ the sum of this probability and its
1311: symmetric image. So (\ref{eq:probxy}) is explained. This shows that
1312: $\mathbb{P}(x,y\in X)\approx r^{-2}|x-y|^{-2}$ and summing over $y$
1313: we get\begin{align*}
1314: \sum_{y:y\neq x}\mathbb{P}(x,y\in X) & \stackrel{(*)}{=}\sum_{n=-C(H)}^{\left\lfloor \log_{2}r\right\rfloor }\sum_{y\in B(x,2^{n+1})\setminus B(x,2^{n})}\mathbb{P}(x,y\in X)\\
1315: & \leq C(\epsilon,H)r^{-2}\sum_{n=-C(H)}^{\left\lfloor \log_{2}r\right\rfloor }4^{-n}\#(B(x,2^{n+1})\setminus B(x,2^{n}))\\
1316: & \!\!\stackrel{(**)}{\leq}\!\! C(\epsilon,H)r^{-2}\sum_{n=-C(H)}^{\left\lfloor \log_{2}r\right\rfloor }2^{n}\leq C(\epsilon,H)r^{-1}\end{align*}
1317: where $(*)$ comes from the fact that $H$ is separated in $\mathbb{R}^{3}$
1318: hence $|x-y|\geq c(H)$ and $(**)$ uses again the fact that rough
1319: isometry preserves volumes. Summing over $x$ we get \[
1320: \mathbb{E}(\# X)^{2}\leq\# A\cdot C(\epsilon,H)r^{-1}+\sum_{x\in A}\mathbb{P}(x\in X)\leq C(\epsilon,H)r^{2}.\]
1321: Hence the well known inequality $\mathbb{P}(\# X>0)\geq(\mathbb{E}\# X)^{2}/\mathbb{E}(\# X)^{2}$
1322: finishes the lemma.
1323: \end{proof}
1324: \begin{lem}
1325: \label{lem:roughxi}Let $G$ be a Euclidean net of dimension $d\leq3$
1326: and let $v^{1}$ and $v^{2}\in G$. Let $R^{i}$ be random walks starting
1327: from $v^{i}$ and stopped on $\partial B(v^{1},r)$. Then\[
1328: \mathbb{P}(R^{1}\cap R^{2}=\emptyset)\leq C(G)\left(\frac{|v^{1}-v^{2}|}{r}\right)^{c(G)}.\]
1329:
1330: \end{lem}
1331: \begin{proof}
1332: Let $a_{j}:=2^{j}|v^{1}-v^{2}|$ and assume without loss of generality
1333: that $r=a_{n}$ for some integer $n$. Let $T_{j}^{i}$ be the stopping
1334: time of $R^{i}$ on the shell $\partial B(v^{1},a_{j})$. Examine
1335: the events \[
1336: \mathcal{E}_{j}:=\{ R^{1}[T_{j}^{1},T_{j+1}^{1}]\cap R^{2}[T_{j}^{2},T_{j+1}^{2}]\neq\emptyset\}.\]
1337: For $j>C(G)$ we have that $\partial B(v^{1},a_{j})\subset B(v^{1},\frac{2}{3}a_{j+1})$
1338: and we may use lemma \ref{lem:Rintrsct} with $\epsilon=\frac{1}{3}$
1339: and the strong Markov property to get\[
1340: \mathbb{P}(\mathcal{E}_{j}\,|\, R^{1}(T_{j}^{1}),R^{2}(T_{j}^{2}))\geq c_{\ref{c:intrsct}}({\textstyle \frac{1}{3}},G)\quad\forall j>C(G).\]
1341: However, $\mathcal{E}_{j}$ may depend on $\mathcal{E}_{0},\dotsc,\mathcal{E}_{j-1}$
1342: only through $R^{i}(T_{j}^{i})$ so we get \[
1343: \mathbb{P}(\mathcal{E}_{j}\,|\,\mathcal{E}_{0},\dotsc,\mathcal{E}_{j-1})\geq c(G).\]
1344: And hence\[
1345: \mathbb{P}(R^{1}\cap R^{2}=\emptyset)\leq\mathbb{P}\Big(\bigcap_{j=C}^{n-1}\neg\mathcal{E}_{j}\Big)\leq(1-c(G))^{n-C}\leq C\left(\frac{|v^{1}-v^{2}|}{r}\right)^{c(G)}\]
1346: and the lemma is proved.
1347: \end{proof}
1348: This basic proof method is known as the {}``Wiener shell test''.
1349:
1350: \begin{rem*}
1351: Given theorem \ref{thm:kof} below (page \pageref{thm:kof}) it might
1352: be tempting to conjecture that $c(G)$ is in effect $\xi$, the non-intersection
1353: exponent of $d$-dimensional Brownian motion. However, this is not
1354: true. Indeed, the intersection exponent is a {}``conformally invariant''
1355: property rather than a metric property. Unfortunately, I don't know
1356: any example sufficiently simple to explain here.
1357: \end{rem*}
1358: \begin{lem}
1359: \label{lem:plane}Let $G$ be a Euclidean net and let $H\subset\mathbb{R}^{d}$
1360: be a half-space and let $v\in G\setminus\partial H$. Let $R$ be
1361: a random walk starting from $v$. Then\[
1362: \mathbb{P}(T_{v,r}<T(\partial H))\leq C(G)\left(\frac{d(v,\partial H)}{r}\right)^{c(G)}\quad\forall r>2d(v,\partial H).\]
1363:
1364: \end{lem}
1365: \begin{proof}
1366: Let $s>2d(v,\partial H)$ and examine a random walk $R$ starting
1367: from any point in $\partial B(v,s)$ and stopped on $\partial B(v,2s)$.
1368: We use lemma \ref{lem:DD} with $\mathcal{D}=B(0,2)$, $\mathcal{S}=B(0,\frac{3}{2})$
1369: and $\mathcal{H}=((H-v)/s)\cap\mathcal{D}$. If $s>C(G)$ then $\partial B(v,s)\subset v+s\mathcal{S}$
1370: and $\mathcal{H}$ is non empty so the lemma applies to our $R$.
1371: We get that the probability of $R$ to hit $\partial(v+s\mathcal{H})\subset\partial H\cap B(v,2s)$
1372: before $\partial B(v,2s)$ is $\geq c_{\ref{c:DDi}}(\mathcal{D},\mathcal{S},\mathcal{H},G)$
1373: if only $s>C_{\ref{C:DD}}(\mathcal{D},\mathcal{S},\mathcal{H},G)$.
1374: Further, if $s$ is sufficiently large then the family of possible
1375: $\mathcal{H}$-s satisfies the requirements of clause \ref{enu:lemDD-K}
1376: of lemma \ref{lem:DD} and these $c$-s and $C$-s are bounded. The
1377: Wiener shell test now gives the lemma.
1378: \end{proof}
1379: Again, it is not necessarily true that $c=1$ as in $\mathbb{Z}^{d}$.
1380: This is only true with additional assumptions, such as isotropicity,
1381: see theorem \ref{lem:escape} below (page \pageref{lem:escape}).
1382: \textbf{}A counterexample may be constructed as follows: Let $\phi:\overline{\mathbb{H}}\to\mathbb{C}$
1383: be defined by $\phi(re^{i\theta})=re^{2i\theta}$ where $\mathbb{H}$
1384: is as usual the upper half space $\{\mathrm{Im\,}z>0\}$. Let $G:=\phi(\mathbb{Z}^{2}\cap\overline{\mathbb{H}})$
1385: and identify $\phi(n)$ and $\phi(-n)$ so that $G$ contains edges
1386: from $n$ to both $\sqrt{n^{2}+1}\, e^{\pm i2\arctan1/n}$. Then it
1387: is easy to see that $G$ is a Euclidean net while the escape probability
1388: from, say, $v=i\sqrt{2}$ to $\partial B(v,r)$ without hitting $\partial\mathbb{H}$
1389: are the same as the escape probabilities of a random walk on $\mathbb{Z}^{2}$
1390: from a corner, which are well known to be $\approx r^{-2}$.
1391:
1392: An argument identical to that of lemma \ref{lem:plane} works for
1393: any polyhedron:
1394:
1395: \begin{lem}
1396: \label{lem:polyh}Let $G$ be a Euclidean net and let $Q\subset\mathbb{R}^{d}$
1397: be a polyhedron. Let $1<r_{1}<r_{2}<s$ be some numbers and let $v\in G$
1398: satisfy that $d(v,sQ)\leq r_{1}$. Let $R$ be a random walk starting
1399: from $v$. Then \[
1400: \mathbb{P}(T_{v,r_{2}}<T(sQ))\leq C(Q,G)\left(\frac{r_{1}}{r_{2}}\right)^{c(Q,G)}.\]
1401:
1402: \end{lem}
1403: We omit the proof.
1404:
1405: The next lemma is technical and is only here for completeness. In
1406: fact all the graphs we will consider in this paper have no {}``dangling
1407: ends'' and it is straightforward to see that for every $v^{1}\neq v^{2}$
1408: one can construct disjoint paths going in opposite directions (so
1409: clause \ref{enu:gamisep} of the lemma is satisfied).
1410:
1411: \begin{lem}
1412: \label{lem:0n0}Let $G$ be an Euclidean net of dimension $\geq2$
1413: and let $0<\epsilon<\frac{1}{2}$, $s>0$. Then there exists a $\kappa=\kappa(\epsilon,s,G)$
1414: such that for any $v,v^{1}\neq v^{2}\in G$ with $|v-v^{i}|\leq s$
1415: one of the following holds:
1416: \begin{enumerate}
1417: \item There are no two disjoint paths starting from $v^{i}$ and ending
1418: outside $B(v,\kappa)$.
1419: \item \label{enu:gamisep}For any $r\geq\kappa$ there exists two disjoint
1420: simple paths $\gamma^{i}\subset B(v,r)$ starting from $v^{i}$ and
1421: ending in some $w^{i}$ such that \begin{equation}
1422: B(w^{i},\left(1-\epsilon\right)r)\cap\gamma^{3-i}=\emptyset,\quad w^{i}\in B(v,r)\setminus\overline{B(v,(1-\epsilon)r)}\label{eq:lem0n0}\end{equation}
1423:
1424: \end{enumerate}
1425: \end{lem}
1426: \begin{proof}
1427: Let $\lambda=\lambda(G)$ satisfy that any ball in $\mathbb{R}^{d}$
1428: with radius $\geq\lambda$ contains at least one point of $G$, and
1429: such that no edge of $G$ has length $>\lambda$. Let $\mu$ satisfy
1430: that for any $x$, $y$ in $G$ such that $|x-y|\leq4\lambda$ there
1431: is a path $\gamma$ from $x$ to $y$, $\gamma\subset B(x,\mu)$.
1432: Now take any point $x\in G$ and any direction $\theta$ and construct
1433: an infinitely long {}``ray'' $x\in R\subset G$ by taking a sequence
1434: of tangent balls on the half line in direction $\theta$, taking a
1435: point of $G$ in every ball and connecting them by short paths as
1436: above. The result is that $R$ is contained in the open infinite cylinder
1437: whose basis is a $d-1$ dimensional ball of radius $\mu+\lambda$
1438: orthogonal to $\theta$, centered at $x$. Actually, $R$ is contained
1439: only in the half cylinder starting $\mu$ before $x$.
1440:
1441: It is now clear that if $|v^{1}-v^{2}|>2\mu$ then we may simply extend
1442: such rays in the directions $\pm(v^{1}-v^{2})$ and they will not
1443: intersect. This allows to prove the case $s>2\mu$ given the case
1444: $s=2\mu$ --- this is a simple geometric exercise (I got that it is
1445: enough to define $\kappa(s,\epsilon)=(s+C(G)+\kappa(\frac{1}{10},2\mu))/\epsilon$).
1446: Hence we will assume $s=2\mu$.
1447:
1448: Define now $\nu:=2\mu+2\lambda$ and impose the condition $\kappa\geq16\nu$.
1449: Let now $\delta^{1}$ and $\delta^{2}$ be two disjoint paths starting
1450: from $v^{i}$ and going to a distance of $\kappa$. The lemma will
1451: be proved once we construct $\gamma^{i}$ satisfying \ref{enu:gamisep}.
1452: Let $x^{i}=\delta^{i}(j^{i})$ be the first point of $\delta^{i}$
1453: outside $B(v,16\nu)$. A simple exercise in plane geometry shows that,
1454: if $|v-x^{i}|\geq\alpha$ then one can find an $\eta$ satisfying
1455: that $\left|\langle\eta,v\rangle-\langle\eta,x^{i}\rangle\right|\geq\frac{1}{2}\alpha$
1456: for $i=1,2$. Applying this in our case gives $\left|\langle\eta,v\rangle-\langle\eta,x^{i}\rangle\right|\geq8\nu$.
1457: Define now\[
1458: I^{i}=\left[\min_{j\leq j^{i}}\left\langle \eta,\delta^{i}(j)\right\rangle ,\max_{j\leq j^{i}}\left\langle \eta,\delta^{i}(j)\right\rangle \right].\]
1459:
1460:
1461: In the sequel, we will say about a (half-)cylinder $\mathcal{C}$
1462: in a direction orthogonal to $\eta$ that it is {}``in elevation
1463: $e$'' if \[
1464: \min_{v\in\mathcal{C}}\left\langle \eta,v\right\rangle =\left\langle \eta,x\right\rangle +e.\]
1465:
1466:
1467: \noindent \emph{Case 1}. Consider the case that $|I^{1}\cap I^{2}|<6\nu$.
1468: In this case one of the $I^{i}$-s --- without loss of generality,
1469: we may assume $I^{1}$ --- contains $\left\langle \eta,v\right\rangle +[\nu,8\nu]$
1470: and $I^{2}$ contains $\left\langle \eta,v\right\rangle +[-8\nu,-\nu]$
1471: (otherwise replace $\eta$ with $-\eta$). Therefore $\left\langle \eta,v\right\rangle \pm[7\nu,8\nu]$
1472: is contained in $I^{1}\setminus I^{2}$ and $I^{2}\setminus I^{1}$
1473: respectively. Let $y^{i}=\delta^{i}(k^{i})$ satisfy that \[
1474: \Big|\left|\left\langle \eta,y^{i}\right\rangle -\left\langle \eta,v\right\rangle \right|-{\textstyle \frac{15}{2}}\nu\Big|\leq{\textstyle \frac{1}{2}}\lambda.\]
1475: Such $y^{i}$ always exist since every edge in $G$ has length $\leq\lambda$.
1476: Let $\theta$ be some vector orthogonal to $\eta$. Let $R^{1}$ (respectively
1477: $R^{2}$) be an infinite path starting from $w^{1}$ (respectively
1478: $w^{2}$) and contained in the half cylinder of radius $\frac{1}{2}\nu$
1479: in the direction $\theta$ (respectively $-\theta$) and in elevation
1480: $7\nu$ (respectively $-8\nu$) . Since the cylinders are disjoint
1481: so are the $R^{i}$-s. See figure \ref{cap:I1uI2d}, left. %
1482: \begin{figure}
1483: \input{I1I2.pstex_t}
1484:
1485: \setcaptionmargin{0.5in}
1486:
1487:
1488: \caption{\label{cap:I1uI2d}On the left, case 1, the case of $|I^{1}\cap I^{2}|$
1489: small. On the right, case 2. The points denoted by $i,j$ are $\delta^{j}(k^{i,j})=R^{i}(m^{i,j})$.}
1490: \end{figure}
1491: Since $I^{1}\cap\left\langle \eta,v\right\rangle +[-8\nu,-7\nu]=\emptyset$
1492: we get that $\delta^{1}\cap R^{2}=\emptyset$ and symmetrically we
1493: have $\delta^{2}\cap R^{1}=\emptyset$. Let $l^{i}$ be the first
1494: time such that $R^{i}(l^{i})\not\in B(x,r$). Then we define\[
1495: \gamma^{i}=\LE(\delta^{i}[0,k^{i}]\cup R^{i}[0,l^{i}-1]).\]
1496: Clearly $\delta^{i}\cup R^{i}$ are disjoint and the operation of
1497: taking $\LE$ conserves this, hence the $\gamma^{i}$-s are simple
1498: and disjoint. (\ref{eq:lem0n0}) now follows from simple plane geometry
1499: (recall that $\LE$ conserves the end points) if only $\kappa$ is
1500: sufficiently large, so this case is finished.
1501:
1502: \medskip{}
1503: \noindent \emph{Case 2}. We now assume that $|I^{1}\cap I^{2}|\geq6\nu$.
1504: Without loss of generality we may assume that for some $a\geq\left\langle \eta,v\right\rangle +\nu$
1505: we have $[a,a+2\nu]\subset I^{1}\cap I^{2}$ (if not, replace $\eta$
1506: with $-\eta$). Let $y\in\delta^{1}$ be the first that satisfies
1507: that $|\left\langle \eta,y\right\rangle -(a+\frac{1}{2}\nu)|\leq\frac{1}{2}\lambda$
1508: and let $z^{i}\in\delta^{i}$ satisfy that $|\left\langle \eta,z^{i}\right\rangle -(a+\frac{3}{2}\nu)|\leq\frac{1}{2}\lambda$.
1509:
1510: This time define $\theta$ be the projection of $z^{1}-z^{2}$ into
1511: the (hyper) plane orthogonal to $\eta$ (if $z^{1}-z^{2}$ is collinear
1512: with $\eta$, just pick $\theta$ an arbitrary vector orthogonal to
1513: $\eta$). Let $\mathcal{C}^{2}$ be a cylinder of radius $\frac{1}{2}\nu$
1514: in the direction $\theta$ with elevation $a+\nu$ such that both
1515: $z^{i}$ are in the middle cylinder of side length $\lambda$. Let
1516: $R^{2}$ be an infinite path in a half of $\mathcal{C}^{2}$ in the
1517: direction $\theta$ starting from $z^{2}$ and containing $z^{1}$.
1518: Let $\mathcal{C}^{1}$ be a similar cylinder in elevation $a$ containing
1519: $y$ in its middle and let $R^{1}$ be an infinite path in the half
1520: of $\mathcal{C}^{1}$ in the direction $-\theta$ starting from $y$.
1521: Also let $R^{1}$ be simple, which can be done, say by taking $\LE$.
1522: See figure \ref{cap:I1uI2d}, right. Let $k^{2,i}$ be the first time
1523: when $\delta^{i}(k^{2,i})\in R^{2}$ and let $m^{2,i}$ be such that
1524: $R^{2}(m^{2,i})=\delta^{i}(k^{2,i})$. Let \[
1525: m^{1,i}:=\max\left\{ m:R^{1}(m)\in\delta^{i}\left[0,k^{2,i}\right[\right\} \quad k^{1,i}:=\min\{ k:\delta^{i}(k)=R^{1}(m^{1,i})\}.\]
1526: By definition $R^{1}$ always intersects $\delta^{1}[0,k^{2,1}[$
1527: hence $m^{1,1}$ and $k^{1,1}$ are well defined. If $R^{1}$ does
1528: not intersect $\delta^{2}[0,k^{2,2}[$ we consider $m^{1,2}$ to be
1529: $-\infty$ and $k^{1,2}$ to be undefined. As before define $l^{i}$
1530: to be the first time when $R^{i}(l^{i})\not\in B(v,r)$.
1531:
1532: We can now define $\gamma^{i}$ by connecting a $\delta^{i}$ to an
1533: $R^{i}$ according to the relation between $m^{1,1}$ and $m^{1,2}$.
1534: In formulas, if $m^{1,1}<m^{1,2}$ define\[
1535: \gamma^{1}:=\LE(\delta^{1}[0,k^{2,1}[\;\cup\; R^{2}[m^{2,1},l^{2}-1])\quad\gamma^{2}:=\LE(\delta^{2}[0,k^{1,2}[\;\cup\; R^{1}[m^{1,2},l^{1}-1])\]
1536: while if $m^{1,1}>m^{1,2}$ (which includes $m^{1,2}=-\infty$) define\[
1537: \gamma^{1}:=\LE(\delta^{1}[0,k^{1,1}[\;\cup\; R^{1}[m^{1,1},l^{1}-1])\quad\gamma^{2}:=\LE(\delta^{2}[0,k^{2,2}[\;\cup\; R^{2}[m^{2,2},l^{2}-1]).\]
1538: In both cases it is easy to verify that $\gamma^{1}\cap\gamma^{2}=\emptyset$
1539: and that \ref{enu:gamisep} holds, if $\kappa$ is sufficiently large,
1540: just like in case 1. Hence the lemma is concluded.
1541: \end{proof}
1542:
1543: \section{\label{sec:Brownian-graphs}Isotropic graphs}
1544:
1545:
1546: \subsection{\label{sub:Brownian}Preliminaries}
1547:
1548: In this chapter we will need to compare random walk and Brownian motion.
1549: For definition and basic properties of Brownian motion see any standard
1550: text book, e.g.~\cite{RW94,B95}.
1551:
1552: To avoid confusion with the use of the letter $B$ for a ball in a
1553: metric space, we will denote Brownian motion by $W$, giving homage
1554: to Wiener, even though he seems to have only been interested in the
1555: one dimensional case.
1556:
1557: The equivalent of the stopping times $T(X)$ and $T_{v,r}$ will be
1558: denoted by $S$ i.e. $S(X)$ is the time when the Brownian motion
1559: hits $X$ for the first time and $S_{v,r}=S(\partial_{\textrm{cont}}B(v,r))$.
1560: Similarly we shall use $S^{i}(X)$ and $S_{v,r}^{i}$ when we have
1561: more than one Brownian motion involved.
1562:
1563: We will also need in a few places the following {}``Hausdorff distance
1564: from a subset'', defined by\[
1565: \subset_{\Haus}(A,B):=\sup_{a\in A}d(a,B)=\inf_{C\subset B}d_{\Haus}(A,C).\]
1566: $\subset_{\Haus}$ is monotone in the sense that if $A_{1}\subset A_{2}$
1567: and $B_{1}\subset B_{2}$ then \[
1568: \subset_{\Haus}(A_{1},B_{2})\leq\;\subset_{\Haus}(A_{2},B_{1}).\]
1569:
1570:
1571:
1572: \subsection{\label{sub:Background-exponent}Background on the non-intersection
1573: exponent}
1574:
1575: We shall need some known results about the non-intersection exponent
1576: $\xi_{d}$, so let us start with a quick survey of this topic (mostly
1577: developed by Lawler and coauthors).
1578:
1579:
1580: \subsubsection*{1}
1581:
1582: Let $x$ and y be two points on $\partial B(0,1)$ and let $W^{x}$
1583: and $W^{y}$ be independent Brownian motions in $\mathbb{R}^{d}$
1584: starting from $x$ and $y$ respectively and stopped when hitting
1585: $\partial B(0,r)$. Define the non-intersection probability by \[
1586: p(x,y,r):=\mathbb{P}(W^{x}\cap W^{y}=\emptyset)\quad P(r):=\max_{x,y}p(x,y,r).\]
1587: The scaling invariance of Brownian motion and the strong Markov property
1588: easily give that $P$ is submultiplicative in $r$ (i.e.~$P(rs)\leq P(r)P(s)$,
1589: or $\log P$ is subadditive in $\log r$) and we get \begin{equation}
1590: P(r)=r^{-\xi_{d}+o(1)}\textrm{ as }r\rightarrow\infty.\label{eq:defxi}\end{equation}
1591: $\xi_{d}$ is the well known non-intersection exponent. The invariance
1592: of Brownian motion to rotations, scaling and translations allows to
1593: map $\vec{1}:=(1,0\dotsc,0)$ and $-\vec{1}$ to the $x$ and $y$
1594: where the maximum occurs and conclude that $p(\vec{1},-\vec{1},r)=r^{-\xi_{d}+o(1)}$.
1595:
1596: We will also need a generalization of this quantity: let $W^{x_{i}}$
1597: ($i=1,\dotsc,k)$ and $W^{y_{i}}$ ($i=1,\dotsc,l$) be independent
1598: Brownian motions in $\mathbb{R}^{d}$ starting from $x_{1},\dotsc,x_{k}$
1599: and $y_{1},\dotsc,y_{l}$ in $\partial B(0,1)$ respectively and stopped
1600: on $\partial B(0,r)$. Define equivalently \begin{align}
1601: p(x_{1},\dotsc,x_{k},y_{1},\dotsc,y_{l},r) & :=\mathbb{P}\Big(\Big(\bigcup_{i=1}^{k}W^{x_{i}}\Big)\cap\Big(\bigcup_{i=1}^{l}W^{y_{i}}\Big)=\emptyset\Big)\nonumber \\
1602: P(k,l,r) & :=\max_{x_{1},\dotsc,x_{k},y_{1},\dotsc,y_{l}}p(x_{1},\dotsc,x_{k},y_{1},\dotsc,y_{l},r).\label{eq:defPmax}\end{align}
1603: Again, submultiplicativity shows that $P=r^{-\xi_{d}(k,l)+o(1)}$
1604: and $\xi_{d}(k,l)$ is called the $k,l$-nonintersection exponent.
1605: With this notation $\xi_{d}=\xi_{d}(1,1)$. A simple {}``choosing
1606: the best point'' argument shows that the maximum in (\ref{eq:defPmax})
1607: is achieved, up to a factor $\leq kl$ when all the $x_{i}$-s are
1608: the same and all the $y_{i}$-s are the same. And invariance again
1609: shows us that\[
1610: p(\underbrace{\vec{1},\dotsc,\vec{1}}_{k},\underbrace{-\vec{1},\dotsc,-\vec{1}}_{l},r)=r^{-\xi_{d}(k,l)+o(1)}.\]
1611:
1612:
1613: These $\xi$-s are non-trivial only in dimensions $2$ and $3$. In
1614: dimension $1$ it follows from the {}``gambler ruin problem'' that
1615: $\xi_{1}(k,l)=k+l$ while in dimensions $\geq4$ Brownian motions
1616: never intersect \cite{DEK50} so $\xi\equiv0$. See \cite{L91} for
1617: a more detailed explanation of these facts. Hence from now on we will
1618: only relate to dimensions $2$ and $3$.
1619:
1620:
1621: \subsubsection*{2}
1622:
1623: In \cite{BL90a} it was shows that the same $\xi(k,l)$ hold for the
1624: equivalent problem for random walks (see also \cite{CM91}). Other
1625: relevant variations consider using Brownian motions (or random walks)
1626: $W_{i}^{*}$ with fixed length $t$, or with the length an exponential
1627: variable with expectation $t$ ({}``a random walk with killing rate
1628: $1/t$''). In either case,\begin{equation}
1629: \mathbb{P}\Big(\Big(\bigcup_{i=1}^{k}W_{i}^{x}\Big)\cap\Big(\bigcup_{i=1}^{l}W_{i}^{y}\Big)=\emptyset\Big)=t^{-\xi_{d}(k,l)/2+o(1)}.\label{eq:2zeta}\end{equation}
1630: For example, notice that, if $\tau_{R}$ is the stopping time when
1631: $W$ exits $\partial B(0,r)$, then the probability that either $\tau_{R}>Cr^{2}\log r$
1632: or $\tau_{R}<cr^{2}/\log r$ are negligible, which explains (\ref{eq:2zeta}).
1633:
1634:
1635: \subsubsection*{3}
1636:
1637: In \cite{L89} it was shown that $\xi_{d}(2,1)=4-d$. Very roughly,
1638: the proof uses the fact that two random walks starting from the same
1639: point can be thought of as one bi-directional walk, which allows to
1640: {}``reduce one parameter'' and get an estimate for the probability.
1641: We remark that a similar technique was used in \cite[section 12.5]{L99}
1642: to calculate some intersection exponents for combinations of random
1643: walks and loop-erased random walks.
1644:
1645:
1646: \subsubsection*{4}
1647:
1648: In \cite{BL90b} it was shown that the $\xi(k,l)$ are strictly increasing,
1649: and in particular that $\xi_{3}(1,1)<1$. The proof uses the Wiener
1650: shell test, somewhat like the techniques we will use in chapter \ref{sec:Quasi-loops}.
1651:
1652:
1653: \subsubsection*{5}
1654:
1655: In \cite{L96a} the estimate (\ref{eq:defxi}) was improved to\begin{equation}
1656: P(x,y,r)\approx r^{-\xi(1,1)}\label{eq:xiConly}\end{equation}
1657: i.e.~the error was shown to be in a constant only (for better comparison,
1658: write $P(x,y,r)=r^{-\xi(1,1)+O(1/\log r)}$). Roughly, this follows
1659: by proving {}``supermultiplicativity'' in the sense that $P(rs)\geq cP(r)P(s)$.
1660: This, in turn, follows after proving that two Brownian motions conditioned
1661: not to intersect will also be quite far along the path and in their
1662: end points. In \cite{L96b} this result was extended to simple random
1663: walk via the so-called Skorokhod embedding, a coupling of Brownian
1664: motion and random walk on the same probability space so as to be quite
1665: close.
1666:
1667: The analog of \cite{L96a} for general $\xi(k,l)$ was proved in \cite{L98}
1668: while the analog of \cite{L96b} is \cite{LP00}. See also \cite{LSW02b}.
1669:
1670:
1671: \subsubsection*{6}
1672:
1673: Both \cite{L96a} and \cite{L96b} used the estimate (\ref{eq:xiConly})
1674: to prove the existence of many cut times or cut points for random
1675: walk, using relatively straightforward second moment methods. Since
1676: $\xi_{3}(1,1)<1$ we get that the Hausdorff dimension of the cut points
1677: is strictly bigger than $1$, which implies that the set of cut points
1678: of Brownian motion is hittable by a second Brownian motion meaning
1679: that the hitting probability is positive. See \cite[section 12.4]{L99}
1680: for the corresponding calculation for random walk.
1681:
1682:
1683: \subsubsection*{7}
1684:
1685: While we will not use it, it is impossible not to mention that in
1686: dimension $2$ there is a precise formula for $\xi_{2}(n,k)$, conjectured
1687: by Duplantier and Kwon \cite{DK88} and proved by Lawler, Schramm
1688: and Werner \cite{LSW02a}. Both the heuristic arguments and the final
1689: proof depend crucially on the Riemann conformal mapping theorem and
1690: are therefore specifically two dimensional.
1691:
1692:
1693: \subsection{\label{sub:DefinitionIsotropic}Definition}
1694:
1695: Let $G$ be a $d$-dimensional Euclidean net. Let $v\in G$ and $r>0$.
1696: Let $A$ be a $d-1$ dimensional spherical simplex (since we are only
1697: interested in $d=2,3$ we have in effect an arc or a spherical triangle)
1698: in $\partial_{\textrm{cont}}B(v,r)$. Let $|A|$ be $(d-1)$-volume
1699: of $A$ normalized so that $|\partial_{\textrm{cont}}B(v,r)|=1$.
1700: We wish to define discrete versions of $A$. For this purpose, identify
1701: each edge $(v,w)$ of $G$ with the linear segment in $\mathbb{R}^{d}$
1702: between the two vertices, and say that $w\in A^{-}$ if $w\in\partial B(v,r)$
1703: and all edges $(v,w)$, $v\in B(v,r)$ intersect $A$. Say that $w\in A^{+}$
1704: if $w\in\partial B(v,r)$ and some edge $(v,w)$ intersects $A$.
1705: Any set $A^{*}$ between $A^{-}$ and $A^{+}$ will be called a discrete
1706: version of $A$. Denote by $p_{A^{*}}=\mathbb{P}^{v}(R(T_{v,r})\in A^{*})$.
1707: We call $G$ \textbf{isotropic} if \begin{equation}
1708: \big|p_{A^{*}}-|A|\big|\leq Kr^{-\alpha}\quad\forall v,r,A,A^{*}.\label{eq:defbrown}\end{equation}
1709: $K>0$ and $\alpha>0$ are parameters of $G$, so it would be more
1710: precise to call $G$ $(d,\alpha,K)$-isotropic. We will rarely need
1711: to do so, though. As in the previous chapter, when we write $C(G)$
1712: we mean a constant that depends only on the isotropicity parameters
1713: $d,\alpha,K$ and the Euclidean net structure constants (see page
1714: \pageref{page:eucCstruct}), but not on other properties of $G$.
1715: Together we call these numbers the \textbf{isotropicity structure
1716: constants}.
1717:
1718: We haven't defined whether we are talking about an open, closed or
1719: other simplex because by expanding or contracting slightly it is obvious
1720: that if (\ref{eq:defbrown}) holds for one than it holds for any and
1721: all. We also remark that by examining triangles intersecting no edge
1722: of $G$, it is obvious that $\alpha\leq d-1$, and if $G$ is a grid
1723: then $\alpha\leq1$. This last inequality tight: it is possible to
1724: show that the grid $\mathbb{Z}^{d}$ is isotropic with $\alpha=1$,
1725: though we will have no use for this fact. It would be interesting
1726: to construct an example in $d\geq3$ of an isotropic graph with $\alpha>1$,
1727: even if one weakens the definition to require that (\ref{eq:defbrown})
1728: holds only for a specific choice of discrete version of $A$.
1729:
1730:
1731: \subsection{Coupling with Brownian motion}
1732:
1733: In this section with shall show how to couple random walk on $G$
1734: with Brownian motion on $\mathbb{R}^{d}$. This will be the main tool
1735: for using isotropic graphs and indeed, it is probably possible to
1736: define isotropic graphs via the coupling. However, we will need some
1737: specific properties of the coupling (see below) that are cumbersome
1738: to formulate.
1739:
1740: We will construct the coupled walk and motion by considering a random
1741: walk $R$ on $G$ and constructing an appropriate Brownian motion
1742: $W$. Let therefore $R$ be given and define inductively a sequence
1743: of stopping times, $\tau_{0}=0$ and\begin{equation}
1744: \tau_{i}:=\min\{ t>\tau_{i-1}:|R(t)-R(\tau_{i-1})|>r_{i}\}\quad r_{i}:=i^{4/\alpha}.\label{eq:defri}\end{equation}
1745: The reason behind the choice of $r_{i}$ will become evident later
1746: on, during the proof of lemma \ref{lem:RWclose} --- we remark only
1747: that the connection between $4$ and the dimension $d$ is $4\geq2(d-1)$.
1748: Construct now fixed divisions $\Delta_{i}$ of the sphere $\mathbb{S}^{d-1}$
1749: into $D_{i}:=\left\lfloor r_{i}^{\alpha/2}\right\rfloor +4$ disjoint
1750: spherical simplices of $(d-1)$-normalized volume $\approx1/D_{i}$
1751: and diameter $\approx D_{i}^{-1/(d-1)}$ (associate the boundaries
1752: of the simplices to them as you please --- this is not important).
1753: In two dimensions one may just take $\Delta_{i}$ to be a collection
1754: of (half-closed half-open) arcs of length $1/D_{i}$. In three dimensions
1755: it is an easy geometric exercise to show that such a {}``triangulation''
1756: exists, knowing only that $D_{i}$ is $\geq4$. For every $\delta\in\Delta_{i}$
1757: define $\delta^{*}$ which, unlike $\delta$, may depend on $v$ and
1758: on the walk up to $R(\tau_{i-1})$, to be a discrete version of $\delta$
1759: such that the $\delta^{*}$ cover $\partial B(R(\tau_{i-1}),r_{i})$
1760: and are disjoint%
1761: \footnote{This definition is not unique, but everything will do will not depend
1762: on the choice of which {}``boundary vertex'' to associate with which
1763: $\delta^{*}$. If one prefers a uniquely defined coupling, just order
1764: $\Delta_{i}$ and then associate each boundary vertex to the $\delta^{*}$
1765: first in this order.%
1766: }. Define, \begin{equation}
1767: p_{i,\delta}:=\mathbb{P}(R(\tau_{i})\in\delta^{*}\,|\, R(\tau_{i-1})).\label{eq:defexit}\end{equation}
1768: We get from (\ref{eq:defbrown}) that\begin{equation}
1769: \big|p_{i,\delta}-|\delta|\big|\leq Kr_{i}^{-\alpha}.\label{eq:pidel_areadel}\end{equation}
1770: Define therefore $\eta_{i}:=\min_{\delta}p_{i,\delta}/|\delta|$ and
1771: get \[
1772: 1\geq\eta_{i}\geq1-CKr_{i}^{-\alpha/2}.\]
1773: We can now construct $W$, and we shall do so in parts, in parallel
1774: with times $\sigma_{i}$ which would be the analogs of $\tau_{i}$.
1775: Define $W(0):=v$ and $\sigma_{0}:=0$. Assume $R(\tau_{i})\in\delta^{*}$.
1776: Throw a random independent coin $X_{i}$ with probability $\eta_{i}|\delta|/p_{i,\delta}$
1777: for $1$. The definition of $\eta_{i}$ ensures that this number is
1778: $\in[0,1]$. If $X_{i}=1$, define $W'_{i}$ to be a Brownian motion
1779: starting from $0$ and conditioned to exit $B(0,r_{i})$ at $r_{i}\delta$.
1780: If $X_{i}=0$, let $W'_{i}$ be an unconditioned Brownian motion.
1781: In both cases define $\sigma_{i}'$ to be the time when $W_{i}'$
1782: exits $B(0,r_{i})$. Finally define $\sigma_{i}=\sigma_{i-1}+\sigma_{i}'$
1783: and $W$ on the interval $\left]\sigma_{i-1},\sigma_{i}\right]$ by
1784: $W(t):=W(\sigma_{i-1})+W_{i}'(t-\sigma_{i-1})$.
1785:
1786: \begin{lem}
1787: The $W$ constructed above is regular Brownian motion.
1788: \end{lem}
1789: \begin{proof}
1790: Since $\mathbb{E}\sigma_{i}'>c>0$ and they are independent we get
1791: that almost surely $\sum\sigma_{i}=\infty$ and hence $W$ is an almost
1792: surely well defined function $[0,\infty[\to\mathbb{R}^{d}$. Now compare
1793: $W$ to a regular Brownian motion $W^{*}$. Let $\sigma_{i}^{*}$
1794: be stopping times defined by\[
1795: \sigma_{i}^{*}=\inf\{ t>\sigma_{i-1}^{*}:W^{*}(t)\not\in B(W^{*}(\sigma_{i-1}^{*},r_{i}))\}.\]
1796: Using the strong Markov property \cite[page 21]{RW94} inductively
1797: gives that $W^{*}(t-\sigma_{i-1}^{*})-W^{*}(\sigma_{i-1}^{*})$ is
1798: distributed like Brownian motion starting from $0$ and stopped when
1799: exiting $B(0,r_{i})$. On the other hand, it follows from the definition
1800: that each $W_{i}'$ has probability $\eta_{i}|\delta|$ to be a Brownian
1801: motion conditioned to hit $r_{i}\delta$ (for every $\delta\in\Delta_{i}$)
1802: and probability $1-\eta_{i}(\sum p_{i})$ to be unconditioned, hence
1803: $W_{i}'$ is also a regular Brownian motion starting from $0$ and
1804: stopped on $\partial B(0,r_{i})$. Hence $W[0,\sigma_{i}]\sim W^{*}[0,\sigma_{i}^{*}]$
1805: for all $i$. Taking limit as $i\to\infty$ shows that $W\sim W^{*}$.
1806: \end{proof}
1807: \begin{lem}
1808: \label{lem:RWclose}Let $G$ be an isotropic graph and $v\in G$.
1809: Let $R$ and $W$ be the coupled walk and motion starting from $v$.
1810: Let $r_{i}$, $\tau_{i}$ and $\sigma_{i}$ be as in the definition
1811: of the coupling (\ref{eq:defri}). Then\[
1812: \mathbb{P}(\exists j\leq i:|R(\tau_{j})-W(\sigma_{j})|\geq\lambda r_{i})\leq C(G)\exp(-c(G)\lambda)\]
1813: for any $\lambda>0$.
1814: \end{lem}
1815: \begin{proof}
1816: We use the notation $X_{i}$ from the definition of the coupling.
1817: If $X_{j}=1$ for some $j$ then we get\[
1818: R(\tau_{j-1})-R(\tau_{j})\in r_{j}\delta+B(0,C(G))\quad W(\sigma_{j-1})-W(\sigma_{j})\in r_{j}\delta\]
1819: Hence\[
1820: |R(\tau_{j-1})-R(\tau_{j})-W(\sigma_{j-1})+W(\sigma_{j})|\leq Cr_{j}^{1-\alpha/2(d-1)}+C(G),\]
1821: and since $d\leq3$ we may simply write $\leq Cr_{j}^{1-\alpha/4}+C$.
1822: Summing we get that if $X_{j}=1$ for all $j\leq i$ then\[
1823: |R(\tau_{i})-W(\sigma_{i})|\leq C\sum_{j=1}^{i}j^{4/\alpha-1}+C\leq C(G)i^{4/\alpha}=C(G)r_{i}.\]
1824: Hence we need to estimate \[
1825: \Sigma:=\sum_{j:X_{j}=0}|R(\tau_{j-1})-R(\tau_{j})-W(\sigma_{j-1})+W(\sigma_{j})|\leq\sum_{j:X_{j}=0}2r_{j}+C.\]
1826: Divide this sum into blocks \[
1827: \Sigma_{k}:=\!\!\!\sum_{\substack{j:X_{j}=0\\
1828: 2^{k-1}<r_{j}\leq2^{k}}
1829: }\!\!2r_{j}+C.\]
1830:
1831:
1832: Now, each $\Sigma_{k}$ contains $\leq C(G)2^{k\alpha/4}$ summands,
1833: and each summand is zero with probability $\geq1-C(G)2^{-k\alpha/2}$
1834: independently so a very rough estimate gives\[
1835: \mathbb{P}(\Sigma_{k}>\lambda2^{k})\leq C(G)\exp(-c(G)\lambda).\]
1836: Define $l:=\left\lceil \log_{2}r_{j}\right\rceil $ and sum over $k$
1837: from $0$ to $l$ to get \begin{align*}
1838: \mathbb{P}\left(\Sigma>\lambda r_{i}\right) & \stackrel{(*)}{\leq}\mathbb{P}(\exists k:\Sigma_{k}>c\lambda2^{(l+k)/2})\leq C(G)\sum_{k}\exp(-c(G)\lambda2^{(l-k)/2})\\
1839: & \leq C(G)\exp(-c(G)\lambda).\end{align*}
1840: $(*)$ comes from the fact that if $\Sigma_{k}\leq\lambda2^{(k+l)/2}$
1841: for all $k$ then $\Sigma=\sum\Sigma_{k}\leq(2+\sqrt{2})\lambda2^{l}\leq2(2+\sqrt{2})\lambda r_{i}$
1842: so one may take $c=1/2(2+\sqrt{2})$ on the right hand side of $(*)$.
1843: Since $\Sigma\leq\lambda r_{i}$ implies $|R(\tau_{j})-W(\sigma_{j})|\leq(\lambda+C(G))r_{i}$
1844: for all $j\leq i$, the lemma is proved.
1845: \end{proof}
1846: Lemma \ref{lem:RWclose} is not really convenient to use as is, because
1847: one needs to relate $i$ to more natural events. Here is one such
1848: useful relation:
1849:
1850: \begin{lem}
1851: \label{lem:rismall}Let $G$ be an isotropic graph and let $v\in G$.
1852: Let $R$ and $W$ be the coupled walk and motion starting from $v$.
1853: Let $r>1$ and let $T=T_{v,r}$ and $S=S_{v,r}$. Let $r_{i}$, $\tau_{i}$
1854: and $\sigma_{i}$ be as in the definition of the coupling. Let\newc{c:RWclose}
1855: \[
1856: I:=\min\{ i:\tau_{i}\geq T,\,\sigma_{i}\geq S\}.\]
1857: Then for some constant $c_{\ref{c:RWclose}}(G)$,\[
1858: \mathbb{P}(r_{I}>\lambda r^{1-c_{\ref{c:RWclose}}(G)})\leq C(G)\exp(-\lambda^{c(G)}).\]
1859:
1860: \end{lem}
1861: \begin{proof}
1862: Since $W(\sigma_{i})-W(\sigma_{i-1})$ are independent variables with
1863: mean zero and variance $cr_{i}$ we get (say by second moment methods)
1864: that for some $c>0$, \[
1865: \mathbb{P}\Big(|W(\sigma_{j})-W(\sigma_{i})|>c\Big(\sum_{k=i}^{j}r_{k}^{2}\Big)^{1/2}\Big)>c\quad\forall j>i.\]
1866: Lemma \ref{lem:RWclose} allows us to replace $W$ with $R$: we get
1867: that for some constants $\mu=\mu(G)$ and $\nu=\nu(G)$,\[
1868: \mathbb{P}\Big(|R(\tau_{j})-R(\tau_{i})|>\mu\Big(\sum_{k=i}^{j}r_{k}^{2}\Big)^{1/2}\Big)>c\quad\forall j>i+\nu.\]
1869: In particular, if $r_{i}\sqrt{j-i}>(2/\mu)r$ then $\mathbb{P}(R(\tau_{j})\not\in B(v,r))>c$
1870: for any $R(\tau_{i})\in B(v,r)$. Hence if we define $J:=C(G)r^{1/(4/\alpha+1/2)}$
1871: for some $C$ sufficiently large we get both $r_{J}\sqrt{J}>(2/\mu)r$
1872: as well as $J>\nu$. Hence \[
1873: \mathbb{P}(R(\tau_{(n+1)J})\not\in B(v,r)\,|\, R[0,\tau_{nJ}],\, R(\tau_{nj})\in B(v,r))>c\quad\forall n>1\]
1874: and hence $\mathbb{P}(T>\tau_{nJ})\leq Ce^{-cn}$. An identical calculation
1875: shows that $\mathbb{P}(S>\sigma_{nJ})\leq Ce^{-cn}$. Hence $\mathbb{P}(r_{I}>r_{nJ})\leq Ce^{-cn}$
1876: and since\[
1877: r_{nJ}=(nJ)^{4/\alpha}=n^{C(G)}r^{1-c(G)}\]
1878: the lemma is proved.
1879: \end{proof}
1880: \begin{cor*}
1881: With the notations of lemma \ref{lem:rismall}, \[
1882: \mathbb{P}(\exists j\leq I:d_{\Haus}(R[0,\tau_{j}],W[0,\sigma_{j}])>\lambda r^{1-c_{\ref{c:RWclose}}})\leq C(G)\exp(-\lambda^{c(G)}).\]
1883:
1884: \end{cor*}
1885: \begin{proof}
1886: Clearly we may assume $\lambda>5$. Let $i_{*}$ by the maximal $i$
1887: such that $r_{i_{*}}\leq\sqrt{\lambda}r^{1-c_{\ref{c:RWclose}}}$.
1888: Then lemma \ref{lem:RWclose} shows that \[
1889: \mathbb{P}(\exists j\leq i_{*}:|R(\tau_{j})-W(\sigma_{j})|\geq(\sqrt{\lambda}-2)r_{i_{*}})\leq C(G)\exp(-c(G)(\sqrt{\lambda}-2)).\]
1890: Now, the point $R(\tau_{j})$ are an approximation (in the Hausdorff
1891: distance) of the entire path, i.e. \[
1892: d_{\Haus}(R[0,\tau_{j}],\{ R(0),\dotsc,R(\tau_{j})\})\leq r_{j}\]
1893: and similarly for $W$. Thus we get\[
1894: \mathbb{P}(\exists j\leq i_{*}:d_{\Haus}(R[0,\tau_{j}],W[0,\sigma_{j}])\geq\sqrt{\lambda}r_{i_{*}})\leq C(G)\exp(-c(G)(\sqrt{\lambda}-2))\]
1895: and from the definition of $i_{*}$, \[
1896: \mathbb{P}(\exists j\leq i_{*}:d_{\Haus}(R[0,\tau_{j}],W[0,\sigma_{j}])\geq\lambda r^{1-c_{\ref{c:RWclose}}})\leq C(G)\exp(-c(G)(\sqrt{\lambda}-2))\]
1897: Estimating the probability that $I>i_{*}$ using lemma \ref{lem:rismall}
1898: proves the corollary.
1899: \end{proof}
1900: \begin{lem}
1901: \label{lem:subhaus}Let $G$ be an isotropic graph, let $\nu>1$ and
1902: let $v\in G$. Let $R$ and $W$ be the coupled walk and motion starting
1903: from $v$. Let $T_{r}=T_{v,r}$ and $S_{r}=S_{v,r}$. Then for all
1904: $r>\max1,s$, and for all $\lambda>0$,\begin{align}
1905: \mathbb{P}(\sh(R[T_{s},T_{r}],W[S_{s/\nu},S_{\nu r}])>\lambda r^{1-c_{\ref{c:RWclose}}(G)}) & \leq C(\nu,G)\exp(-\lambda^{c(G)})\label{eq:RinW2}\\
1906: \mathbb{P}(\sh(W[S_{s},S_{r}],R[T_{s/\nu},T_{\nu r}])>\lambda r^{1-c_{\ref{c:RWclose}}(G)}) & \leq C(\nu,G)\exp(-\lambda^{c(G)}).\label{eq:WinR2}\end{align}
1907:
1908: \end{lem}
1909: We explicitly include the case $s=0$ in which case we define $T_{0}=S_{0}=0$.
1910:
1911: \begin{proof}
1912: Let us prove (\ref{eq:RinW2}). Let $r_{i}$, $\tau_{i}$ and $\sigma_{i}$
1913: be as in the definition of the coupling (\ref{eq:defri}). Define\[
1914: I_{1}:=\max\{ i:\tau_{i}<T_{s}\},\quad I_{2}:=\min\{ i:\tau_{i}\geq T_{r}\}.\]
1915: Then lemma \ref{lem:rismall} gives that\begin{align}
1916: \mathbb{P}(r_{I_{1}}>\lambda s^{1-c_{\ref{c:RWclose}}}) & \leq C(G)\exp(-\lambda^{c(G)}),\label{eq:flem22s}\\
1917: \mathbb{P}(r_{I_{2}}>\lambda r^{1-c_{\ref{c:RWclose}}}) & \leq C(G)\exp(-\lambda^{c(G)}).\label{eq:flem22r}\end{align}
1918: Denote by $i_{1}$ the last $i$ such that $r_{i}\leq\lambda s^{1-c_{\ref{c:RWclose}}}$
1919: and by $i_{2}$ the last $i$ such that $r_{i}\leq\lambda r^{1-c_{\ref{c:RWclose}}}$.
1920: Lemma \ref{lem:RWclose} shows that \begin{align}
1921: \mathbb{P}(\exists j\leq i_{k}:|R(\tau_{j})-W(\sigma_{j})|\geq\lambda r_{i_{k}}) & \leq C(G)\exp(-c(G)\lambda)\quad k=1,2.\label{eq:flem23rs}\end{align}
1922: Together with the estimates of $r_{I_{k}}$ this gives the following
1923: corollary:\begin{align}
1924: \mathbb{P}(d_{\Haus}(R[\tau_{I_{1}},\tau_{I_{2}}],W[\sigma_{I_{1}},\sigma_{I_{2}}])\geq\lambda^{2}r^{1-c_{\ref{c:RWclose}}}) & \leq C(G)\exp(-\lambda^{c(G)}).\label{eq:RtauI1I2W}\end{align}
1925: We can replace $\lambda^{2}$ with $\lambda$ on the left hand side
1926: paying only in the constant inside the exponent on the right hand
1927: side.
1928:
1929: \medskip{}
1930: \noindent \textbf{case 1:} If $s<r^{1-c_{\ref{c:RWclose}}}$ then
1931: the fact that $|W(S_{s/\nu})-R(T_{s})|\leq C(G)+2s$ shows that \[
1932: \sh(R[T_{s},T_{r}],W[0,S_{\nu r}])\geq\sh(R[T_{s},T_{r}],W[S_{s/\nu},S_{\nu r}])-(C(G)+2s)\]
1933: which allows to estimate\begin{align*}
1934: \lefteqn{\mathbb{P}(\sh(R[T_{s},T_{r}],W[S_{s/\nu},S_{\nu r}])>\lambda r^{1-c_{\ref{c:RWclose}}})\leq}\quad\\
1935: & \leq\mathbb{P}(\sh(R[T_{s},T_{r}],W[0,S_{\nu r}])>(\lambda-C(G))r^{1-c_{\ref{c:RWclose}}})\leq\\
1936: & \leq\mathbb{P}(\sh(R[\tau_{I_{1}},\tau_{I_{2}}],W[0,\sigma_{I_{2}}])>(\lambda-C(G))r^{1-c_{\ref{c:RWclose}}})+\mathbb{P}(\sigma_{I_{2}}>S_{\nu r})\leq\\
1937: & \!\!\stackrel{(\ref{eq:RtauI1I2W})}{\leq}\! C(G)\exp(-(\lambda-C(G))^{c(G)})+\mathbb{P}(\sigma_{I_{2}}>S_{\nu r}).\end{align*}
1938: Now, to estimate $\mathbb{P}(\sigma_{I_{2}}>S_{\nu r})$ we use the
1939: fact that $|R(\tau_{i})-v|\leq r$ for any $i<I_{2}$ and get \begin{alignat*}{2}
1940: \mathbb{P}(\sigma_{I_{2}}>S_{\nu r}) & & \leq\;\;\: & \mathbb{P}(\exists j\leq i_{2}:|W(\sigma_{j})-R(\tau_{j})|\geq(\nu-1)r-\lambda r^{1-c_{\ref{c:RWclose}}})+\mathbb{P}(I_{2}>i_{2})\\
1941: & & \stackrel{(\ref{eq:flem22r},\ref{eq:flem23rs})}{\leq} & C(G)\exp(-c(G)((\nu-1)\lambda^{-1}r^{c_{\ref{c:RWclose}}}-1))+C(G)\exp(-\lambda^{c(G)}).\end{alignat*}
1942: Hence, if $\lambda<r^{c(G)}$, (\ref{eq:RinW2}) is proved.
1943:
1944: \medskip{}
1945: \noindent \textbf{case 2:} If $s\geq r^{1-c_{\ref{c:RWclose}}}$ then
1946: we estimate\begin{align*}
1947: \lefteqn{\mathbb{P}(\sh(R[T_{s},T_{r}],W[S_{s/\nu},S_{\nu r}])>\lambda r^{1-c_{\ref{c:RWclose}}})\leq}\\
1948: & \quad\leq\mathbb{P}(\sh(R[\tau_{I_{1}},\tau_{I_{2}}],W[\sigma_{I_{1}},\sigma_{I_{2}}])>\lambda r^{1-c_{\ref{c:RWclose}}})+\mathbb{P}(\sigma_{I_{1}}<S_{s/\nu})+\mathbb{P}(\sigma_{I_{2}}>S_{\nu r})\\
1949: & \quad\!\!\stackrel{(\ref{eq:RtauI1I2W})}{\leq}\! C(G)\exp(-\lambda^{c(G)})+\mathbb{P}(\sigma_{I_{1}}<S_{s/\nu})+\mathbb{P}(\sigma_{I_{2}}>S_{\nu r}).\end{align*}
1950: Now, the estimate of $\mathbb{P}(\sigma_{I_{2}}>S_{\nu r})$ is as
1951: in case 1. The estimate of $\mathbb{P}(\sigma_{I_{1}}<S_{s/\nu})$
1952: is similar. Using $|R(\tau_{I_{1}})-v|\geq s-r_{I_{1}}-C(G)$, we
1953: get \begin{align*}
1954: \lefteqn{\mathbb{P}(\sigma_{I_{1}}<S_{s/\nu})\leq\mathbb{P}(|W(\sigma_{I_{1}})-v|<s/\nu)\leq}\\
1955: & \quad & \leq\;\;\: & \mathbb{P}(\exists j\leq i_{1}:|W(\sigma_{j})-R(\tau_{j})|\geq s(1-1/\nu)-\lambda s^{1-c_{\ref{c:RWclose}}}-C(G))+\\
1956: & & & \qquad+\mathbb{P}(I_{1}>i_{1})\leq\\
1957: & & \stackrel{(\ref{eq:flem22s},\ref{eq:flem23rs})}{\leq} & C(G)\exp(-c(G)((1-1/\nu)\lambda^{-1}s^{c_{\ref{c:RWclose}}}-C(G))+C(G)\exp(-\lambda^{c(G)})\end{align*}
1958: and again, if $\lambda\leq s^{c(G)}$, (\ref{eq:RinW2}) is proved.
1959: Since in this case $s>r^{c(G)}$ it is enough to assume $\lambda\leq r^{c(G)}$
1960: in order to get $\lambda\leq s^{c(G)}$ and consequently (\ref{eq:RinW2}).
1961:
1962: \medskip{}
1963: \noindent \textbf{case 3:}\newc{c:lamr}The previous calculations
1964: proved the case $\lambda\leq r^{c_{\ref{c:lamr}}}$ for some $c_{\ref{c:lamr}}(G)$.
1965: However, this implies that for any $\lambda\leq2r^{c_{\ref{c:RWclose}}}$,\[
1966: \mathbb{P}(\sh(R[T_{s},T_{r}],W[S_{s/\nu},S_{\nu r}])>\lambda r^{1-c_{\ref{c:RWclose}}(G)})\leq C(\nu,G)\exp\left(-\left({\textstyle \frac{1}{2}}\lambda\right)^{c(G)c_{\ref{c:lamr}}/c_{\ref{c:RWclose}}}\right)\]
1967: or in other words, (\ref{eq:RinW2}) holds with different constants
1968: on the right hand side. However, for $\lambda>2r^{c_{\ref{c:RWclose}}}$,
1969: (\ref{eq:RinW2}) holds trivially because \[
1970: \sh(R[T_{s},T_{r}],W[S_{s/\nu},S_{\nu r}])\leq d_{\Haus}(R[T_{s},T_{r}],W[S_{s/\nu},S_{r}])\leq2r\]
1971: so the probability in (\ref{eq:RinW2}) is zero. This finishes the
1972: proof of (\ref{eq:RinW2}). The proof of (\ref{eq:WinR2}) is identical.
1973: \end{proof}
1974: \begin{cor*}
1975: With the notations of lemma \ref{lem:subhaus}, if $R$ and $W$ start
1976: from a $w$, $|v-w|\leq\frac{1}{4}s(1-1/\nu)$, then (\ref{eq:RinW2})
1977: and (\ref{eq:WinR2}) still hold, possibly with different constants.
1978: Further, this holds if $s=0$ and $|v-w|\leq\frac{1}{4}r(\nu-1)$.
1979: \end{cor*}
1980: This follows from lemma \ref{lem:subhaus} and the monotonicity of
1981: $\sh$.
1982:
1983:
1984: \subsection{Hitting of small balls}
1985:
1986: From now on we will prove {}``natural'' facts about walk on isotropic
1987: graphs, natural in the sense that they don't need the coupling (or
1988: other special notations) to be stated. In this section we shall prove
1989: two lemmas about the hitting probability of {}``intermediate scale''
1990: objects, i.e. of the size $r^{1-c}$ (both will be used in chapter
1991: \ref{sec:Isotropic-gluing}). In the next sections we shall focus
1992: on more delicate facts.
1993:
1994: \begin{lem}
1995: \label{lem:sball}Let $G$ be an isotropic graph, and let $\epsilon>0$.
1996: Let $r>C(\epsilon,G)$ and let $v,w\in B(x,r(1-\epsilon))$, $|v-w|>\frac{1}{2}\epsilon r$.
1997: Let \[
1998: p:=\mathbb{P}^{v}(T_{w,s}<T_{x,r})\quad q:=\mathbb{P}^{v}(S_{w,s}<S_{x,r}).\]
1999: Where $r^{1-c_{\ref{c:RWclose}}/2}\leq s\leq|v-w|-\frac{1}{2}\epsilon r$,
2000: $c_{\ref{c:RWclose}}(G)$ from lemma \ref{lem:rismall}. Then\[
2001: \left|p-q\right|\leq C(\epsilon,G)r^{-c(G)}\max p,q.\]
2002:
2003: \end{lem}
2004: \begin{proof}
2005: The first step is to get a simple lower bound for $q$. In the case
2006: $|v-w|\leq\frac{3}{4}d(w,\partial_{\textrm{cont}}B(x,r))$ then a
2007: calculation using the continuous analog of (\ref{eq:fEfRT}) with
2008: the Newtonian potential \cite[chapter II 3]{B95} around $w$ shows
2009: that\begin{equation}
2010: q\geq c(\epsilon,d)\begin{cases}
2011: s/r & d=3\\
2012: 1/\log(r/s) & d=2\end{cases}\label{eq:lowq}\end{equation}
2013: ($d$ being the dimension). Removing the condition $|v-w|\leq\frac{3}{4}d(w,\partial B(0,r))$,
2014: we still have (\ref{eq:lowq}), perhaps with a different constant.
2015: Indeed, using the continuous Harnack inequality \cite[chapter II 1]{B95}
2016: for the domain $B(x,r)\setminus B(w,s+\frac{1}{4}\epsilon r)$ shows
2017: (\ref{eq:lowq}) for all $w$ and $v$.
2018:
2019: Next define \[
2020: q^{\pm}=\mathbb{P}^{v}(S_{w,s^{\pm}}<S_{x,r})\quad s^{\pm}=s\pm r^{1-(3/4)c_{\ref{c:RWclose}}}\]
2021: so $q^{-}\leq q\leq q^{+}$. Now, the strong Markov property gives
2022: us that\[
2023: q-q^{-}\leq\mathbb{P}^{v}\left(S_{w,s}<S_{x,r}\right)\cdot\max_{y\in\partial_{\textrm{cont}}B(w,s)}\mathbb{P}^{y}\left(W[0,S_{x,r}]\cap B(w,s^{-})=\emptyset\right)\]
2024: and again, similar calculations with the continuous Newtonian potential
2025: gives, for any $y\in\partial_{\textrm{cont}}B(w,s)$, \[
2026: \mathbb{P}^{y}\left(W[0,S_{x,r}]\cap B(w,s^{-})=\emptyset\right)\leq C(\epsilon,d)\frac{r^{1-(3/4)c_{\ref{c:RWclose}}}}{s}\leq Cr^{-c_{\ref{c:RWclose}}/4}\]
2027: So we get $q-q^{-}\leq Cqr^{-c}$. A similar calculation shows that
2028: $q^{+}-q\leq Cq^{+}r^{-c}$ and for $r$ sufficiently large we may
2029: write $q^{+}-q\leq Cqr^{-c}$.
2030:
2031: To extract from these inequalities conclusions about $|q-p|$, couple
2032: $R$ and $W$ as above. Let $r_{i}$, $\tau_{i}$ and $\sigma_{i}$
2033: be as in the definition of the coupling (\ref{eq:defri}). Let\[
2034: I:=\min\{ i:\tau_{i}\geq T_{v,r},\,\sigma_{i}\geq S_{v,r}\}.\]
2035: When comparing $p$ to $q^{+}$ we need to consider two cases: the
2036: first that $R[0,\tau_{I}]$ and $W[0,\sigma_{I}]$ are not very close;
2037: and the second is as in figure \ref{cap:RBnWB}. %
2038: \begin{figure}
2039: \input{SvrTvr.pstex_t}
2040:
2041:
2042: \caption{$R[0,T_{x,r}]$ intersects the ball $B(w,s)$ but $W[0,S_{x,r}]$
2043: is quite far from it.\label{cap:RBnWB}}
2044: \end{figure}
2045: In a formula: if $\lambda$ is the length of the longest edge in $G$
2046: then\begin{align}
2047: p-q^{+} & \leq\mathbb{P}\left(\left\{ R[0,T_{x,r}]\cap\partial B(w,s)\neq\emptyset\right\} \cap\left\{ W[0,S_{x,r}]\cap\partial B(w,s^{+})=\emptyset\right\} \right)\leq\nonumber \\
2048: & \leq\mathbb{P}(d_{\Haus}(R[0,\tau_{I}],W[0,\sigma_{I}])\geq r^{1-(3/4)c_{\ref{c:RWclose}}}-\lambda)\;+\nonumber \\
2049: & \qquad\mathbb{P}(W[S_{x,r},S_{x,r+r^{1-(3/4)c_{\ref{c:RWclose}}}}]\cap B(w,s^{+})\neq\emptyset).\label{eq:WSAnn}\end{align}
2050: Now, the corollary to lemma \ref{lem:rismall} for the ball $B(v,2r)$
2051: gives, since $B(x,r)\subset B(v,2r)$, that\[
2052: \mathbb{P}(d_{\Haus}(R[0,\tau_{I}],W[0,\sigma_{I}])\geq r^{1-(3/4)c_{\ref{c:RWclose}}}-\lambda)\leq C(G)\exp(-r^{c(G)}).\]
2053: while for the second summand we have from the strong Markov property,\begin{eqnarray*}
2054: \lefteqn{\mathbb{P}(W[S_{x,r},S_{x,r+r^{1-(3/4)c_{\ref{c:RWclose}}}}]\cap B(w,s^{+})\neq\emptyset)\leq}\\
2055: & \qquad & \leq\max_{y\in\partial_{\textrm{cont}}B(x,r)}\mathbb{P}^{y}(W[0,S_{x,r+r^{1-(3/4)c_{\ref{c:RWclose}}}}]\cap B(x,r(1-\epsilon/4))\neq\emptyset)\cdot\\
2056: & & \qquad\max_{y\in\partial_{\textrm{cont}}B(x,r(1-\epsilon/4))}\mathbb{P}^{y}(W[0,S_{x,r+r^{1-(3/4)c_{\ref{c:RWclose}}}}]\cap B(w,s^{+})\neq\emptyset)\leq\\
2057: & & \stackrel{(*)}{\leq}Cr^{-(3/4)c_{\ref{c:RWclose}}}\begin{cases}
2058: s/r & d=3\\
2059: 1/\log(r/s) & d=2\end{cases}\leq Cr^{-c}q.\end{eqnarray*}
2060: Both estimates of $(*)$ follow from the continuous Newtonian potential.
2061: Note that we assumed here that $\frac{1}{8}\epsilon r>r^{1-(3/4)c_{\ref{c:RWclose}}}$
2062: which we may, if $r$ is sufficiently large. This finishes the proof
2063: that $p\leq q+Cr^{-c}q$.
2064:
2065: The proof of the other direction is similar. We have\begin{align*}
2066: q^{-}-p & \leq\mathbb{P}\left(\left\{ W[0,S_{x,r}]\cap\partial B(w,s^{-})\neq\emptyset\right\} \cap\left\{ R[0,T_{x,r}]\cap\partial B(w,s)=\emptyset\right\} \right)\leq\\
2067: & \leq\mathbb{P}(d_{\Haus}(R[0,\tau_{I}],W[0,\sigma_{I}])\geq r^{1-(3/4)c_{\ref{c:RWclose}}})\;+\\
2068: & \qquad\mathbb{P}(W[S_{x,r-r^{1-(3/4)c_{\ref{c:RWclose}}}},S_{x,r}]\cap B(w,s^{-})\neq\emptyset)\end{align*}
2069: and an identical calculation finishes this case, and the lemma.
2070: \end{proof}
2071: \begin{lem}
2072: \label{lem:striag}Let $G$ be an isotropic graph and let $\epsilon>0$.
2073: Let $r>C(\epsilon,G)$ and let $v\in B(x,(1-\epsilon)r)$. Let $\Delta$
2074: be a spherical triangle on $\partial_{\textrm{cont}}B(x,r)$ and let
2075: $\Delta^{*}$ be a discrete version of it. Let $q:=\mathbb{P}^{v}(W(S_{x,r})\in\Delta)$
2076: and $p:=\mathbb{P}^{v}(R(T_{x,r})\in\Delta^{*})$. Then\newc{c:striag}\[
2077: \left|p-q\right|\leq C(G,\epsilon)r^{-c_{\ref{c:striag}}(G)}.\]
2078:
2079: \end{lem}
2080: \noindent (the only difference between lemma \ref{lem:striag} and
2081: the definition of an isotropic graph is that here the starting point
2082: of the walk $v$ might be different from the center of the stopping
2083: ball $x$). The proof is very similar to the proof of the previous
2084: lemma, so we indicate only the differences. We define $q^{\pm}$ as
2085: the probabilities of $W$ to hit $\partial B(x,r)$ at $\Delta^{\pm}$
2086: where\begin{align*}
2087: \Delta^{+} & :=\left(\Delta+B(0,r^{1-c_{\ref{c:RWclose}}/2})\right)\cap\partial_{\textrm{cont}}B(x,r)\\
2088: \Delta^{-} & :=\left\{ x\in\Delta:B(x,r^{1-c_{\ref{c:RWclose}}/2})\cap\partial_{\textrm{cont}}B(x,r)\subset\Delta\right\} .\end{align*}
2089: The proof that $q^{+}-q$, $q-q^{-}\leq Cr^{-c}$ is direct calculation
2090: for $v=x$, and for general $v$ follows from the continuous Harnack
2091: inequality. The proof that $p-q^{+}\leq Cr^{-c}$ is similar, except
2092: the last term on the right should be replaced, for example, with\begin{multline*}
2093: \mathbb{P}\left(\left\{ W[0,S_{x,r}]\cap\left(\Delta+B(0,r^{1-(3/4)c_{\ref{c:RWclose}}})\right)\neq\emptyset\right\} \cap\left\{ W(S_{x,r})\not\in\Delta^{+}\right\} \right)\leq\\
2094: \leq\max_{d(y,\Delta)\leq r^{1-(3/4)c_{\ref{c:RWclose}}}}\mathbb{P}^{y}(W(S_{x,r})\not\in\Delta^{+})\leq Cr^{-c}.\end{multline*}
2095: The proof that $q^{-}-p\leq Cr^{-c}$ and the rest of the lemma are
2096: similar.\qed
2097:
2098:
2099: \subsection{Escape probabilities}
2100:
2101: In the section we move from the {}``intermediate scale'' objects
2102: of the previous section to single points. This is more delicate, and
2103: we shall employ techniques similar to those of Lawler \cite{L96b}.
2104: Our main goal is theorem \ref{lem:escape}, but first we need to state
2105: and prove two simple claims.
2106:
2107: Henceforth $R$ and $W$ will always be a random walk and a Brownian
2108: motion coupled as above.
2109:
2110: \begin{lem}
2111: \label{lem:Brownescape}Let $H\subset\mathbb{R}^{d}$ be a half space.
2112: Let $W$ be a Brownian motion on $\mathbb{R}^{d}$ starting from some
2113: vertex $v\not\in\partial_{\textrm{cont}}H$. Let $r>2d(v,\partial_{\textrm{cont}}H)$.
2114: Then\[
2115: p:=\mathbb{P}(S_{v,r}<S(\partial_{\textrm{cont}}H))\approx\frac{d(v,\partial_{\textrm{cont}}H)}{r}.\]
2116:
2117: \end{lem}
2118: \begin{proof}
2119: By translation, scaling and rotation invariance we may assume that
2120: $r=1$, that $v=\epsilon e_{1}$ for some $\frac{1}{2}>\epsilon>0$
2121: ($e_{1}$ being the first basis element) and that $H=\left\{ x:\left\langle x,e_{1}\right\rangle >0\right\} $.
2122: Examine two positive harmonic functions: $f(x)=\mathbb{P}^{x}(S_{v,r}<S(\partial H_{i}))$
2123: and $g(x)=\left\langle x,e_{1}\right\rangle $. Both $f$ and $g$
2124: are zero on $\partial_{\textrm{cont}}H$ so the boundary Harnack principle
2125: for Lipschitz domains \cite[theorem III.1.2, page 178]{B95} shows
2126: that $g(v)/f(v)\approx g(x_{0})/f(x_{0})$ where $x_{0}$ is any reference
2127: point. But $g(x_{0})/f(x_{0})$ is just a number, and the lemma is
2128: proved.
2129: \end{proof}
2130: \begin{lem}
2131: \label{lem:Ers}Let $G\subset\mathbb{R}^{d}$ be an isotropic graph.
2132: Let $H\subset\mathbb{R}^{d}$ be a half space. Let $v\in G\setminus\partial_{\textrm{cont}}H$.
2133: Let $\frac{1}{2}\rho>s>r>2d(v,\partial_{\textrm{cont}}H)$. Let $R$
2134: and $W$ be coupled walk and motion starting from $v$. Let $\mathcal{E}$
2135: be an event depending on $R[0,T_{v,r}]$ and $W[0,S_{v,r}]$ only.
2136: Then\[
2137: \mathbb{P}(\mathcal{E}\cap\{ W[S_{v,r},S_{v,s}]\cap\partial_{\textrm{cont}}H=\emptyset\})\leq C\frac{r}{s}(\mathbb{P}(\mathcal{E})+C(G)\exp(-r^{c(G)})).\]
2138:
2139: \end{lem}
2140: \begin{proof}
2141: We may assume $s>4r$ and (using lemma \ref{lem:Brownescape}) $r>1$.
2142: Let $r_{i}$, $\tau_{i}$ and $\sigma_{i}$ be as in the definition
2143: of the coupling. Define $I:=\min\{ i:\tau_{i}\geq T_{v,r},\,\sigma_{i}\geq S_{v,r}\}$
2144: and examine $W$ after $\sigma_{I}$. This is a regular Brownian motion,
2145: starting from $W(\sigma_{I})$ and \emph{independent of both $W[0,\sigma_{I}]$
2146: and $R[0,\tau_{I}]$.} Therefore, if we denote by $\mathcal{F}$ the
2147: event $W[\sigma_{I},S_{v,s}]\cap\partial_{\textrm{cont}}H=\emptyset$
2148: we get from lemma \ref{lem:Brownescape}, \begin{equation}
2149: \mathbb{P}(\mathcal{F}\,|\, W(\sigma_{I}))\leq C\frac{d(W(\sigma_{I}),\partial_{\textrm{cont}}H)}{d(W(\sigma_{I}),\partial_{\textrm{cont}}B(v,s))}.\label{eq:S1SHWsigI}\end{equation}
2150: Actually, this holds only if $W(\sigma_{I})\in B(v,s)$ --- in the
2151: other case we will simply estimate $\mathbb{P}\leq1$. This shows
2152: that, \begin{equation}
2153: \mathbb{P}(\mathcal{E}\cap\mathcal{F}\cap\{|W(\sigma_{I})-v|\leq2r\})\leq C\mathbb{P}(\mathcal{E})\frac{3r}{s-2r}\leq C\mathbb{P}(\mathcal{E})\frac{r}{s}.\label{eq:smallWv}\end{equation}
2154: For the case that $|W(\sigma_{I})-v|>2r$ we use the corollary to
2155: lemma \ref{lem:rismall} to get $\mathbb{P}(|R(\tau_{I})-W(\sigma_{I})|>\lambda r)\leq C\exp(-c(\lambda r)^{c}).$
2156: By the definition of $I$, either $|W(\sigma_{I})-v|=r$ or $|R(\tau_{I})-v|\leq r+C(G)$
2157: so we get \begin{equation}
2158: \mathbb{P}(|W(\sigma_{I})-v|>\lambda r)\leq C(G)\exp(-(\lambda r)^{c(G)}).\label{eq:WsigIvlog}\end{equation}
2159: Hence we get that \begin{eqnarray}
2160: \lefteqn{\mathbb{P}(\{|W(\sigma_{I})-v|>2r\}\cap\mathcal{F})=}\nonumber \\
2161: & \qquad & =\sum_{\lambda=3}^{\infty}\mathbb{P}(\{|W(\sigma_{I})-v|\in\left](\lambda-1)r,\lambda r\right]\}\cap\mathcal{F})\nonumber \\
2162: & & \stackrel{(*)}{\leq}C(G)\sum_{\lambda=3}^{\infty}\exp(-(\lambda r)^{c(G)})\cdot\begin{cases}
2163: 2(\lambda+1)r/s & \lambda r<\frac{1}{2}s\\
2164: 1 & \textrm{otherwise}\end{cases}\nonumber \\
2165: & & \leq C(G)\exp(-r^{c(G)})\frac{r}{s}\label{eq:largeWv}\end{eqnarray}
2166: where the inequality $(*)$ comes from (\ref{eq:WsigIvlog}) for the
2167: left multiplicand and (\ref{eq:S1SHWsigI}) for the right multiplicand.
2168: Combining (\ref{eq:smallWv}) and (\ref{eq:largeWv}) ends the lemma.
2169: \end{proof}
2170: \begin{thm}
2171: \label{lem:escape}\newC{C:mindvH}Let $G\subset\mathbb{R}^{d}$
2172: be an isotropic graph. Let $H\subset\mathbb{R}^{d}$ be a half space.
2173: Then there exists a constant $C_{\ref{C:mindvH}}(G)$ such that for
2174: any such $H$, any $v$ with $d(v,\partial_{\textrm{cont}}H)>C_{\ref{C:mindvH}}$
2175: and any $r>2d(v,\partial_{\textrm{cont}}H)$ one has that the probability
2176: $p$ that a random walk $R$ on $G$ starting from $v$ will hit $\partial B(v,r)$
2177: before hitting $\partial H$ satisfies \[
2178: p\approx\frac{d(v,\partial_{\textrm{cont}}H)}{r}.\]
2179:
2180: \end{thm}
2181: Remember that the constants implicit in the $\approx$ above, like
2182: $C_{\ref{C:mindvH}}$, may depend on the Brownian structure constants
2183: of $G$. Actually, the proof below shows that the lower bound does
2184: not depend on $G$ at all, and the upper bound can be shown to do
2185: the same easily. However, we will have no use for these facts.
2186:
2187: \begin{proof}
2188: Denote $\mu=d(v,\partial_{\textrm{cont}}H)/r$. We may assume w.l.o.g.~$\mu=2^{-M}$
2189: for some integer $M$. Before starting with estimates for $R$ we
2190: need to know a fact about Brownian motion, roughly speaking that Brownian
2191: motion conditioned to have $\{ S_{v,4r}<S(\partial_{\textrm{cont}}H)\}$
2192: also avoids $\partial_{\textrm{cont}}H$ along (most of) its path.
2193: To formulate precisely, let $\beta<1$ be some parameter which will
2194: be fixed later, and denote \[
2195: a_{k}=d(v,\partial_{\textrm{cont}}H)2^{k}\quad b_{k}=a_{k}^{\beta}.\]
2196: So that $r=a_{M}$. Define stopping times $S_{k}:=S_{v,a_{k}}$ for
2197: $k\in\left]-\infty,M+2\right]$ and let $\mathcal{E}:=\bigcup_{k=-\infty}^{M+1}\mathcal{E}_{k}$
2198: be defined by \[
2199: \mathcal{E}_{k}:=\{\exists S_{k-1}<t\leq S_{k}:d(W(t),\partial_{\textrm{cont}}H)\leq b_{k}\}\cap\{ S_{M+2}<S(\partial_{\textrm{cont}}H))\}.\]
2200: Note that because we end at $\mathcal{E}_{M+1}$ we actually ignore
2201: the event of getting close to $\partial_{\textrm{cont}}H$ on the
2202: last stretch of the Brownian motion, namely $\left]S_{M+1},S_{M+2}\right]$
2203: (this makes the proof a little simpler). We assume $C_{\ref{C:mindvH}}\geq1$
2204: and then $\mathcal{E}_{k}$ is empty for $k\leq-1$. For $0\leq k\leq M+1$
2205: we shall use lemma \ref{lem:Brownescape} in the form $\mathbb{P}(S_{k-1}<S(\partial_{\textrm{cont}}H))\leq C2^{-k}$
2206: and again (together with the strong Markov property) to get\[
2207: \mathbb{P}(S_{M+2}<S(\partial_{\textrm{cont}}H)\,|\, W[0,t^{*}])\le C\frac{b_{k}}{r}\quad t^{*}:=\inf_{t>S_{k-1}}d(W(t),\partial_{\textrm{cont}}H)\leq b_{k}.\]
2208: (clearly $t^{*}$ is a stopping time hence we may use the strong Markov
2209: property). Hence we get\[
2210: \mathbb{P}(\mathcal{E}_{k})\leq C2^{-k}\frac{b_{k}}{r}=C\mu a_{k}^{\beta-1}\]
2211: and summing (remember that $\beta<1$)\begin{equation}
2212: \mathbb{P}(\mathcal{E})\leq C\mu\frac{d(v,\partial_{\textrm{cont}}H)^{\beta-1}}{1-2^{\beta-1}}.\label{eq:Fsmall}\end{equation}
2213:
2214:
2215: We now move to examine the random walk $R$. Couple $W$ with $R$
2216: as above. Let $T_{k}=T_{v,a_{k}}$ that is the $R$-equivalents of
2217: the $S_{k}$. Assume $C_{\ref{C:mindvH}}$ is sufficiently large so
2218: that $R[0,T_{-1}]\cap\partial H=\emptyset$ always. We start with
2219: a lower bound for $p$. A little set calculus gives \begin{align}
2220: \mathbb{P}(T_{M}<T(\partial H)) & \geq\mathbb{P}(S_{M+2}<S(\partial_{\textrm{cont}}H))-\mathbb{P}(\mathcal{E})-\nonumber \\
2221: & \quad-\mathbb{P}(\{ S_{M+2}<S(\partial_{\textrm{cont}}H)\}\setminus(\{ T_{M}<T(\partial H)\}\cup\mathcal{E})).\label{eq:setcalc}\end{align}
2222: Now, if $T_{M}\geq T(\partial H)$ then we have that $R(n)\in\partial H$
2223: for some $n\in[T_{k-1},T_{k}]$. If $S_{M+2}<S(\partial_{\textrm{cont}}H)$
2224: but $\mathcal{E}$ did not happen then we must have that $d(W[S_{k-2},S_{k+1}],\linebreak[4]\partial_{\textrm{cont}}H)>b_{k-1}$
2225: and hence $\sh(R[T_{k-1},T_{k}],W[S_{k-2},S_{k+1}])>b_{k-1}-C$. Thus
2226: we arrive at\begin{eqnarray}
2227: \lefteqn{\mathbb{P}(\{ S_{M+2}<S(\partial_{\textrm{cont}}H)\}\setminus(\{ T_{M}<T(\partial H)\}\cup\mathcal{E}))\le}\label{eq:RScapT}\\
2228: & & \le\sum_{k=0}^{M}\mathbb{P}(\{\sh(R[T_{k-1},T_{k}],W[S_{k-2},S_{k+1}])>b_{k-1}-C\}\cap\nonumber \\
2229: & & \qquad\qquad\cap\;\{ S_{M+2}<S(\partial_{\textrm{cont}}H)\}).\nonumber \end{eqnarray}
2230: The estimate of (\ref{eq:RScapT}) follows from lemma \ref{lem:subhaus}
2231: but first we need to chose $\beta$ and we choose $\beta=1-\frac{1}{2}c_{\ref{c:RWclose}}(G)$
2232: where $c_{\ref{c:RWclose}}(G)$ comes from lemma \ref{lem:subhaus}.
2233: The lemma then claims \begin{equation}
2234: \mathbb{P}(\sh(R[T_{k-1},T_{k}],W[S_{k-2},S_{k+1}])>b_{k-1}-C)\leq C(G)\exp(-a_{k}^{c(G)}).\label{eq:noSMSH}\end{equation}
2235: The condition $S_{M+2}<S(\partial_{\textrm{cont}}H)$ can be added
2236: via lemma \ref{lem:Ers}, and we get \begin{multline*}
2237: \mathbb{P}\Big(\{\sh(R[T_{k-1},T_{k}],W[S_{k-2},S_{k+1}])>b_{k-1}-C\}\:\cap\\
2238: \cap\{ S_{M+2}<S(\partial_{\textrm{cont}}H)\}\Big)\leq C(G)\exp(-a_{k}^{c(G)})\mu2^{k}.\end{multline*}
2239: Plugging this into (\ref{eq:RScapT}) and summing we get \[
2240: \mathbb{P}(\{ S_{M+2}<S(\partial_{\textrm{cont}}H)\}\setminus(\{ T_{M}<T(\partial H)\}\cup\mathcal{E}))\le C(G)\exp(-d(v,\partial_{\textrm{cont}}H)^{c(G)})\mu.\]
2241: This we may plug into (\ref{eq:setcalc}) together with (\ref{eq:Fsmall})
2242: and lemma \ref{lem:Brownescape} and get\[
2243: \mathbb{P}(T_{M}<T(H))\geq\mu\left(c-C\frac{d(v,\partial_{\textrm{cont}}H)^{-c_{\ref{c:RWclose}}(G)/2}}{1-2^{-c_{\ref{c:RWclose}}(G)/2}}-C(G)\exp(-d(v,\partial_{\textrm{cont}}H)^{c(G)})\right)\]
2244: and it is now clear that if $C_{\ref{C:mindvH}}$ is chosen sufficiently
2245: large, then $d(v,\partial_{\textrm{cont}}H)>C_{\ref{C:mindvH}}$ would
2246: give that everything inside the parenthesis is $>c$ and the direction
2247: $p\geq c\mu$ is proved.
2248:
2249: The proof that $p\leq C\mu$ is, generally speaking, a mirror image
2250: exchanging the roles of $R$ and $W$ in the proof of $p\geq c\mu$.
2251: Since our a-priori knowledge about the random walk is smaller (it
2252: is, essentially, lemma \ref{lem:plane}), the proof is somewhat rearranged.
2253: Here are the details: Define \[
2254: g_{i}:=2^{i}\mathbb{P}(T_{i}<T(\partial H)).\]
2255: Fix one $i\geq4$. For every $j\in[2,i-2]$ examine the event\[
2256: \mathcal{F}_{j}:=\{ W[S_{j-1},S_{j}]\cap\partial_{\textrm{cont}}H\neq\emptyset\}\cap\{ W\left]S_{j},S_{i-2}\right]\cap\partial_{\textrm{cont}}H=\emptyset\}.\]
2257: The event $\{ T_{i}<T(\partial H)\}\cap\mathcal{F}_{j}$ has a number
2258: of consequences:
2259:
2260: \smallskip{}
2261: \noindent \textbf{(i)} $R[0,T_{j-2}]\cap\partial H=\emptyset$. By
2262: definition this event has probability $2^{2-j}g_{j-2}$.
2263:
2264: \smallskip{}
2265: \noindent \textbf{(ii)} $R[T_{j-2},T_{j+1}]\cap\partial H=\emptyset$.
2266: Lemma \ref{lem:subhaus} shows that \[
2267: \mathbb{P}(\sh(W[S_{j-1},S_{j}],R[T_{j-2},T_{j+1}])>a_{j}^{1-c_{\ref{c:RWclose}}/2})\leq C(G)\exp(-a_{j}^{c(G)})\]
2268: and since $W[S_{j-1},S_{j}]\cap\partial_{\textrm{cont}}H\neq\emptyset$
2269: we get that with probability $1-C\exp(-a_{j}^{c})$, $d(R[T_{j-2},T_{j+1}],\partial H)\leq a_{j}^{1-c_{\ref{c:RWclose}}/2}+C(G)$.
2270: Denote this event by $\mathcal{G}$. Lemma \ref{lem:plane} and the
2271: strong Markov property now show that\[
2272: \mathbb{P}(\mathcal{G}\cap\{ T_{j+2}>T(\partial H)\}\,|\, R[0,T_{j-2}])\leq C(G)\left(\frac{a_{j}^{1-c_{\ref{c:RWclose}}/2}}{a_{j}}\right)^{c(G)}.\]
2273: Together with clause (i) we get\begin{eqnarray*}
2274: \lefteqn{\mathbb{P}(\{ R[T_{0},T_{j+2}]\cap\partial H=\emptyset\}\cap\{ W[S_{j-1},S_{j}]\cap\partial_{\textrm{cont}}H\neq\emptyset\})\leq}\\
2275: & \qquad & \leq C(G)2^{2-j}g_{j-2}a_{j}^{-c(G)}+C(G)\exp(-a_{j}^{c(G)})\stackrel{(*)}{\leq}C(G)2^{2-j}g_{j-2}a_{j}^{-c(G)}\end{eqnarray*}
2276: where in $(*)$ we used the lower bound $g_{j}\geq c$ already established.
2277:
2278: \smallskip{}
2279: \noindent \textbf{(iii)} $W[S_{j+2},S_{i}]\cap\partial_{\textrm{cont}}H=\emptyset$.
2280: Here we employ lemma \ref{lem:Ers} and get \begin{align*}
2281: \mathbb{P}(\{ T_{i}<T(\partial H)\}\cap\mathcal{F}_{j}) & \leq C2^{j+2-i}\big(C(G)2^{2-j}g_{j-2}a_{j}^{-c(G)}+C(G)\exp(-a_{j}^{c(G)})\big)\\
2282: & \leq C(G)2^{-i}g_{j-2}a_{j}^{-c(G)}.\end{align*}
2283:
2284:
2285: With the estimate of $\mathbb{P}(\{ T_{i}<T(\partial H)\}\cap\mathcal{F}_{j})$
2286: complete we need only sum on $j$ and use lemma \ref{lem:Brownescape}
2287: to get\[
2288: \mathbb{P}(\{ T_{i}<T(\partial H)\}\setminus\bigcup_{j=2}^{i-2}\mathcal{F}_{j})\leq\mathbb{P}(W[S_{1},S_{i-2}]\cap\partial_{\textrm{cont}}H=\emptyset)\leq C2^{-i}\leq C2^{-i}g_{0}\]
2289: so we get\[
2290: g_{i}\leq C(G)\sum_{j=0}^{i-4}g_{j}\left(2^{-c(G)}\right)^{j}.\]
2291: By lemma 4.5 of \cite{L96b}, $g_{i}$ are bounded and the bound depends
2292: only on the isotropic structure constants of $G$ (we use here that
2293: $g_{0}\leq1$) and the theorem is proved. \newC{C:arc}
2294: \end{proof}
2295: \begin{lem}
2296: \label{lem:arc}With the notations of theorem \ref{lem:escape} (but
2297: $d(v,\partial_{\textrm{cont}}H)\geq C_{\ref{C:arc}}$), let $p$ be
2298: the probability that a random walk $R$ on $G$ starting from $v$
2299: will hit $\partial B(v,r)\cup\partial H$ in the arc\[
2300: \alpha:=\partial B(v,r)\cap\{ x:d(x,\partial_{\textrm{cont}}H)>{\textstyle \frac{1}{2}}r\}.\]
2301: Then $p\approx d(v,\partial_{\textrm{cont}}H)/r$.
2302: \end{lem}
2303: \begin{proof}
2304: $p\leq C(G)d(v,\partial_{\textrm{cont}}H)/r$ is an immediate consequence
2305: of theorem \ref{lem:escape}. For the other direction, first assume
2306: w.l.o.g.~that $r>4d(v,\partial_{\textrm{cont}}H)$, which can be
2307: done by lemma \ref{lem:DD}. Let $\lambda>0$ be some parameter, and
2308: let $\beta$ be the two arcs $\partial B(v,r/2)\cap\{ x:d(x,\partial_{\textrm{cont}}H)<\lambda r\}$.
2309: Examine the event\[
2310: \mathcal{E}:=\{ R(T(\partial B(v,r/2)\cup\partial H))\in\beta\}\cap\{ T_{v,r}<T(\partial H)\}.\]
2311: Theorem \ref{lem:escape} shows that \[
2312: \mathbb{P}(R(T(\partial B(v,r/2)\cup\partial H))\in\beta)\leq\mathbb{P}^{v}\left(T_{v,r/2}<T(\partial H)\right)\leq C(G)d(v,\partial_{\textrm{cont}}H)/r.\]
2313: For any $x\in\beta$ we have, again from theorem \ref{lem:escape}
2314: \[
2315: \mathbb{P}^{x}(T_{v,r}<T(\partial H))\leq\mathbb{P}^{x}(T_{x,r/4}<T(\partial H))\leq C(G)\lambda\]
2316: if only $\lambda<\frac{1}{8}$ and $\lambda r>C(G)$. Hence we get\[
2317: \mathbb{P}(\mathcal{E})\leq C(G)\lambda d(v,\partial_{\textrm{cont}}H)/r.\]
2318: Combining this with the lower bound of theorem \ref{lem:escape} we
2319: get \begin{eqnarray*}
2320: \lefteqn{c(G)\frac{d(v,\partial_{\textrm{cont}}H)}{r}\leq\mathbb{P}\left(T_{v,r/2}<T(\partial H)\right)\leq}\\
2321: & \qquad & \leq\mathbb{P}(R(T(B(v,r/2)\cup\partial H))\in\partial B(v,r/2)\setminus\beta)+\lambda C(G)\frac{d(v,\partial_{\textrm{cont}}H)}{r}.\end{eqnarray*}
2322: Choose $\lambda=\lambda(G)$ some constant sufficiently small and
2323: get that \[
2324: \mathbb{P}(R(T(B(v,r/2)\cup\partial H))\in\partial B(v,r/2)\setminus\beta)>c(G)\frac{d(v,\partial_{\textrm{cont}}H)}{r}.\]
2325: Lemma \ref{lem:DD} now shows that there is a probability $>c(G)$
2326: to hit $\partial B(v,r)\cup\partial H$ at $\alpha$ if you start
2327: from any point of $\partial B(v,r/2)\setminus\beta$, and we are done.
2328: \end{proof}
2329:
2330: \subsection{Lower bound for the non-intersection probability}
2331:
2332: The proof of theorem \ref{lem:escape} in the previous section was
2333: modeled roughly on Lawler \cite{L96b}. In contrast, theorem \ref{thm:kof},
2334: which will be proved in this section and the next, is a completely
2335: straightforward generalization of \cite{L96b}.
2336:
2337: \begin{thm}
2338: \label{thm:kof}Let $G$ be an isotropic graph of dimension $2$
2339: or $3$. Then for any $v^{1},v^{2}\in G$ with $|v^{1}-v^{2}|>C(G)$,
2340: If $R^{1}$ and $R^{2}$ are two walks with $R^{i}$ starting from
2341: $v^{i}$ and stopped on $\partial B(v^{1},r)$, $r>2|v^{1}-v^{2}|$,
2342: then \begin{equation}
2343: c(G)\left(\frac{|v_{1}-v_{2}|}{r}\right)^{-\xi}\leq\mathbb{P}(R^{1}\cap R^{2}=\emptyset)\leq C(G)\left(\frac{|v_{1}-v_{2}|}{r}\right)^{-\xi}\label{eq:thmkof}\end{equation}
2344: where $\xi=\xi_{d}(1,1)$ is the intersection exponent from (\ref{eq:defxi}).
2345: \end{thm}
2346: We shall not repeat the argumentation of \cite{L96b}, we shall only
2347: note the pieces that require changes. Hence the rest of the chapter
2348: should be read side by side with \cite{L96b}. Chapter 2 of \cite{L96b}
2349: requires almost no changes: the following lemma, which is a replacement
2350: for (7) in lemma 2.5 is perhaps worth proving here.\newC{C:min7}
2351:
2352:
2353:
2354: \begin{lem}
2355: \label{lem:lwlr25}Let $\epsilon>0$ and let $G$ be an isotropic
2356: graph. Then there exists a $\delta=\delta(\epsilon,G)$ such that
2357: for any $r>C_{\ref{C:min7}}(\epsilon,G)$; any $v\in G$ and any $w\in G$
2358: with $|v-w|\leq r$ one has\[
2359: \mathbb{P}^{1,w}\left(\inf_{|z-v|<r}\mathbb{P}^{2,z}\big(R^{2}[0,T_{v,2r}^{2}]\cap R^{1}[T_{v,r}^{1},T_{v,2r}^{1}]\neq\emptyset\,|\, R^{1}[T_{v,r}^{1},T_{v,2r}^{1}]\big)<\delta\right)<\epsilon\]
2360: where $R^{i}$ are two independent random walks.
2361: \end{lem}
2362: In words, if we consider a path to be {}``$\delta$-hittable from
2363: $z$'' if the probability of a random walk ($R^{2}$) starting from
2364: $z$ to hit it is $\geq\delta$, then what we prove here is that random
2365: walk ($R^{1}$) is, with probability $1-\epsilon$, $\delta$-hittable
2366: from any $z\in B(v,r)$. (to understand the formula formally, remember
2367: that the conditional probability $\mathbb{P}(\cdot|*)$ is a function
2368: of $*$ and note that the $\inf$ relates to a pointwise infimum of
2369: these functions).
2370:
2371: \begin{proof}
2372: Let $x\in G$ and let $s>\lambda$ where $\lambda=\lambda(G)$ will
2373: be some constant sufficiently large that will be fixed later. Assume
2374: for now that $\lambda>C_{\ref{C:minr}}(\frac{1}{8},G)$ where $C_{\ref{C:minr}}$
2375: comes from lemma \ref{lem:Rintrsct}. Hence we use lemma \ref{lem:Rintrsct}
2376: and get for any $y\in B(x,\frac{7}{4}s)$, \begin{equation}
2377: \mathbb{P}^{1,x,2,y}(R^{1}[0,T_{x,s/4}^{1}]\cap R^{2}[0,T_{x,2s}^{2}]\neq\emptyset)>c(G).\label{eq:P12}\end{equation}
2378: For any path $\gamma$ from $x$ to $\partial B(x,s/4)$ define $Y(y,\gamma):=\mathbb{P}^{2,y}(\gamma\cap R^{2}[0,T_{x,2s}^{2}]\neq\emptyset)$.
2379: Then (\ref{eq:P12}) implies that\[
2380: \mathbb{P}^{1,x}(Y(y,R^{1}[0,T_{x,s/4}^{1}])>c(G))>c(G)\quad\forall y\in B(x,{\textstyle \frac{7}{4}}s).\]
2381: If $\lambda$ is sufficiently large then we have $\partial B(x,\frac{1}{4}s)\subset B(x,\frac{1}{2}s)$
2382: and then $Y(\cdot,\gamma)$ is harmonic on $\{ y:\frac{1}{2}s\leq|x-y|\leq\frac{3}{2}s\}$,
2383: and we may use Harnack's inequality (lemma \ref{lem:Harnack-general})
2384: to show that for some constant $\mu=\mu(G)$, \begin{equation}
2385: \mathbb{P}^{1,x}\Big(\inf_{\frac{3}{4}s\leq|x-y|\leq\frac{5}{4}s}Y(y,R^{1}[0,T_{x,s/2}^{1}])>\mu\Big)>c(G).\label{eq:defZ}\end{equation}
2386: Denote by $Z(\gamma)$ the event $\{\inf_{3s/4\leq|x-y|\leq5s/4}Y(y,\gamma)>\mu\}$
2387: where $x$ is the beginning of the path $\gamma$.
2388:
2389: Next, let $N=N(\epsilon,G)$ be an integer parameter which will be
2390: fixed later. For $i=1,\dotsc,N-1$ define $T_{i}:=T_{v,(1+i/N)r}^{1}$.
2391: Let $x_{i}=R^{1}(T_{i})$. Let $s=r/4N$. Define $U_{i}$ to be the
2392: stopping times\[
2393: U_{i}:=\min\{ t>T_{i}:R^{1}(t)\in\partial B(x_{i},{\textstyle \frac{1}{4}}s)\}.\]
2394: Finally define $Z_{i}=Z(R^{1}[T_{i},U_{i}])$. Then (\ref{eq:defZ})
2395: says that $\mathbb{P}(Z_{i}\,|\, x_{i})>c(G)$. Since the only effect
2396: of $Z_{1},\dotsc,Z_{i-1}$ on $Z_{i}$ is through $x_{i}$ we get
2397: in fact that\[
2398: \mathbb{P}(Z_{i}\,|\, Z_{1},\dotsc,Z_{i-1})>c(G)\]
2399: and hence\[
2400: \mathbb{P}\Big(\bigcap_{i=1}^{N-1}\neg Z_{i}\Big)<(1-c(G))^{N-2}.\]
2401: Denote $\mathcal{Z}:=\cup_{i=1}^{N-1}Z_{i}$ and choose our parameter
2402: $N$ such that $\mathbb{P}(\mathcal{Z})>1-\epsilon$. Lemma \ref{lem:DD}
2403: shows that for $r$ bigger than some constant $\nu(N,G)$ we have
2404: that the probability of $R^{2}$ to hit $\partial B(x_{i},s)$ for
2405: any $i$ and for any starting point $z$ of $R^{2}$ is $\geq c(N,G)$.
2406: If $\lambda$ is sufficiently large then $\partial B(x_{i},s)\subset B(x_{i},\frac{5}{4}s)$.
2407: Hence we get for any $i\in\{1,\dotsc,N\}$,\begin{eqnarray*}
2408: \lefteqn{\mathbb{P}(R^{2}[0,T_{v,2r}^{2}]\cap R^{1}[T_{v,r}^{1},T_{v,2r}^{1}]\neq\emptyset\,|\, Z_{i})\geq}\\
2409: & \qquad & \geq\mathbb{P}(R^{2}[0,T_{v,2s}^{2}]\cap R^{1}[T_{i},U_{i}]\neq\emptyset\,|\, Z_{i})\geq\\
2410: & & \stackrel{(*)}{\geq}\mathbb{P}(T_{x_{i},s}^{2}<T_{v,2r}^{2})\cdot\mathbb{E}(Y(R^{2}(T_{x_{i},s}^{2}),R^{1}(T_{i},U_{i}))\,|\, Z_{i})\geq\\
2411: & & \geq\mu\mathbb{P}(T_{x_{i},s}^{2}<T_{v,2s}^{2})\geq\mu c(N,G)\end{eqnarray*}
2412: where $(*)$ comes from the strong Markov property at the stopping
2413: time $T_{x_{i},s}^{2}$. Hence we get \[
2414: \mathbb{P}(R^{2}[0,T_{v,2s}^{2}]\cap R^{1}[T_{v,s}^{1},T_{v,2s}^{1}]\,|\,\mathcal{Z})\geq\mu c(N,G).\]
2415: This finishes the lemma: we fix $\lambda$ and define $\delta(\epsilon,G):=\mu c(N,G)$
2416: and $C_{\ref{C:min7}}(\epsilon,G):=\max(8N\lambda,\nu(N,G))$ and
2417: we are done.
2418: \end{proof}
2419: \begin{lem}
2420: \label{lem:lwlr26}\newC{C:MKepsG}Let $G$ be an isotropic graph
2421: and let $M$, $K$ and $\epsilon$ be some parameters. Then there
2422: exists a $\delta(M,K,\epsilon,G)$ and a $C_{\ref{C:MKepsG}}(M,K,\epsilon,G)$
2423: such that for all $v,w\in G$, and all $r>|v-w|$,\[
2424: \mathbb{P}^{1,w}\left(\inf_{(*)}\mathbb{P}^{2,z}\big(R^{2}[0,T_{v,2r}^{2}]\cap R^{1}[0,T_{v,2r}^{1}]\neq\emptyset\,|\, R^{1}[0,T_{v,2r}^{1}]\big)\geq r^{-\delta}\right)<C_{\ref{C:MKepsG}}r^{-M}\]
2425: where the $(*)$ stands for all the $z\in B(v,r)$ such that $d(z,R^{1}[0,T_{v,2r}^{1}])\leq Kr^{1-\epsilon}$.
2426: \end{lem}
2427: The proof is identical to that of lemma 2.6 from \cite{L96b} and
2428: we shall omit it. Very roughly, it uses the previous lemma and the
2429: Wiener shell test.
2430:
2431: Chapter 3 of \cite{L96b} has no real equivalence here. The Skorokhod
2432: embedding used in \cite{L96b} has the convenient property that the
2433: random walk and the Brownian motion have comparable times, that is
2434: $|R(t)-W(t)|\ll t^{1/4+\epsilon}$ (after linear calibration). This
2435: is just not true in our case, or anyway would require non-linear adaptive
2436: calibration which is not worth messing with --- measuring the Hausdorff
2437: distance between $R$ and $W$ is a completely adequate replacement.
2438: Hence we shall make no effort to give analogs of the results of chapter
2439: 3 of \cite{L96b} and continue immediately to chapter 4. Lemma \ref{lem:Lwlr41}
2440: is a replacement for Lawler's lemma 4.1, but first an auxiliary result:
2441:
2442: \begin{lem}
2443: \label{lem:Erskof}Let $G\subset\mathbb{R}^{d}$ be an isotropic graph.
2444: Let $R^{1},W^{1}$ and $R^{2},W^{2}$ be two independent pairs of
2445: coupled random walk and Brownian motion on $G$ starting from $v^{1}$
2446: and $v^{2}$ respectively. Let $s>r>2|v^{1}-v^{2}|$. Let $\mathcal{E}$
2447: be an event depending on $R^{i}[0,T_{v,r}^{i}]$ and $W^{i}[0,S_{v,r}^{i}]$
2448: only. Then\[
2449: \mathbb{P}(\mathcal{E}\cap\{ W^{1}[S_{v,r}^{1},S_{v,s}^{1}]\cap W^{2}[S_{v,r}^{2},S_{v,s}^{2}]=\emptyset\})\leq C\left(\frac{r}{s}\right)^{\xi}(\mathbb{P}(\mathcal{E})+C(G)\exp(-r^{c(G)})).\]
2450: where $\xi$ is from (\ref{eq:defxi}).
2451: \end{lem}
2452: The proof is identical to that of lemma \ref{lem:Ers}, with the use
2453: of lemma \ref{lem:Brownescape} replaced by estimates for the non-intersection
2454: probability of two Brownian motion, see \cite[(2)]{L96a}. We omit
2455: the details.
2456:
2457: \begin{lem}
2458: \label{lem:Lwlr41}Let $G$ be an isotropic graph, let $v\in G$ and
2459: let $R^{1},W^{1}$ and $R^{2},W^{2}$ be two independent pairs of
2460: coupled random walk and Brownian motion on $G$ starting from $v^{1}$
2461: and $v^{2}$ respectively, $|v-v^{i}|\leq2^{m}$ and stopped on $\partial B(v,2^{n})$.
2462: Define $T_{j}^{i}:=T_{v,2^{j}}^{i}$ and $S_{j}^{i}:=S_{v,2^{j}}^{i}$
2463: and\begin{alignat*}{2}
2464: Q_{j}^{i} & := & \; & \{\sh(R^{i}[T_{j-1}^{i},T_{j}^{i}],W^{i}[S_{j-2}^{i},S_{j+1}^{i}])\geq2^{j(1-c_{\ref{c:RWclose}}/2)}\}\;\cup\\
2465: & & & \{\sh(W^{i}[S_{j-1}^{i},S_{j}^{i}],R^{i}[T_{j-2}^{i},T_{j+1}^{i}])\geq2^{j(1-c_{\ref{c:RWclose}}/2)}\},\\
2466: Q_{*}^{i} & := & & \{\sh(R^{i}[0,T_{m+2}^{i}],W^{i}[0,S_{m+3}^{i}])\geq2^{(m+2)(1-c_{\ref{c:RWclose}}/2)}\}\;\cup\\
2467: & & & \{\sh(W^{i}[0,S_{m+2}^{i}],R^{i}[0,T_{m+3}^{i}])\geq2^{(m+2)(1-c_{\ref{c:RWclose}}/2)}\}.\\
2468: \mathcal{Q} & := & & Q_{*}^{1}\cup Q_{*}^{2}\cup\bigcup_{j=m+3}^{n-1}Q_{j}^{1}\cup Q_{j}^{2}.\end{alignat*}
2469: Then\[
2470: \mathbb{P}(\mathcal{Q}\cap\{ W^{1}[0,S_{n+1}^{1}]\cap W^{2}[0,S_{n+1}^{2}]=\emptyset\})\leq C(G)\exp(-2^{mc(G)})2^{-(n-m)\xi}.\]
2471:
2472: \end{lem}
2473: \begin{proof}
2474: The corollary to lemma \ref{lem:subhaus} with $\nu=2$ shows that\[
2475: \mathbb{P}(Q_{j}^{i})\leq C(G)\exp(-2^{jc(G)}).\]
2476: Next, lemma \ref{lem:Erskof} shows that\[
2477: \mathbb{P}(Q_{j}^{i}\cap\{ W^{1}[0,S_{n+1}^{1}]\cap W^{2}[0,S_{n+1}^{2}]=\emptyset\})\leq C(G)\exp(-2^{jc(G)})2^{-(n-j)\xi}.\]
2478: $Q_{*}^{i}$ have a similar estimate. Summing on $i$ and $j$ we
2479: get the lemma.
2480: \end{proof}
2481: We now prove a lemma, the equivalent of corollary 4.2 of \cite{L96b},
2482: somewhat stronger than the direction $\mathbb{P}(R^{1}\cap R^{2}=\emptyset)\geq c\left(|v_{1}-v_{2}|/r\right)^{-\xi}$
2483: of (\ref{eq:thmkof}). We will need the strengthening in the next
2484: chapter.\newC{C:cor42}
2485:
2486: \begin{lem}
2487: \label{lem:lwlr42}Let $G$ be an isotropic graph of dimension $2$
2488: or $3$ and let $v\in G$ and $s\geq C_{\ref{C:cor42}}(G)$. Let $v^{1},v^{2}\in G\cap\big(B(v,s)\setminus B(v,\frac{7}{8}s)\big)$,
2489: $|v^{1}-v^{2}|>\frac{1}{4}s$, let $r>4s$ and let $\eta$ be a unit
2490: vector in $\mathbb{R}^{d}$ and define two subsets of $G$,\begin{align}
2491: & U^{1}:=\big(B(v,r/2)\setminus B(v,s)\big)\cup B(v^{1},{\textstyle \frac{1}{4}}s)\cup\left(\overline{B(v,r)}\cap\left\{ w:\left\langle w,\eta\right\rangle \geq{\textstyle \frac{1}{4}}r\right\} \right),\label{eq:defMi}\\
2492: & U^{2}:=\big(B(v,r/2)\setminus B(v,s)\big)\cup B(v^{2},{\textstyle \frac{1}{4}}s)\cup\left(\overline{B(v,r)}\cap\left\{ w:\left\langle w,\eta\right\rangle \leq-{\textstyle \frac{1}{4}}r\right\} \right).\nonumber \end{align}
2493: Let $R^{1}$ and $R^{2}$ are two walks with $R^{i}$ starting from
2494: $v^{i}$ and stopped on $\partial B(v,r)$. Then\begin{equation}
2495: \mathbb{P}(\{ R^{1}\cap R^{2}=\emptyset\}\cap\{ R^{i}\subset U^{i},i=1,2\})\geq c(G)\left(\frac{s}{r}\right)^{\xi}.\label{eq:mshrm}\end{equation}
2496:
2497: \end{lem}
2498: \begin{proof}
2499: This is now immediate. Indeed, consider slightly smaller domains (but
2500: extended outward), \begin{align*}
2501: & V^{1}:=\big(B(v,{\textstyle \frac{5}{12}}r)\setminus B(v,{\textstyle \frac{25}{24}}s)\big)\cup B(v^{1},{\textstyle \frac{5}{24}}s)\cup\left\{ w\in B(v,2r):\left\langle w,\eta\right\rangle \geq{\textstyle \frac{1}{3}}r\right\} ,\\
2502: & V^{2}:=\big(B(v,{\textstyle \frac{5}{12}}r)\setminus B(v,{\textstyle \frac{25}{24}}s)\big)\cup B(v^{2},{\textstyle \frac{5}{24}}s)\cup\left\{ w\in B(v,2r):\left\langle w,\eta\right\rangle \leq-{\textstyle \frac{1}{3}}r\right\} .\end{align*}
2503: Consider also the event $\mathcal{F}$ that $W^{1}$ and $W^{2}$
2504: are reasonably far apart along their paths, namely\begin{align*}
2505: \mathcal{F}_{j}^{i} & :=\{ d(W^{i}[S_{v,2^{j-3}}^{i},S_{v,2^{j}}^{i}],W^{3-i}[0,S_{v,2^{j}}^{3-i}])>2^{j(1-c_{\ref{c:RWclose}}/4)}\}\\
2506: \mathcal{F} & :=\bigcap_{i=1,2}\bigcap_{j=\left\lfloor \log_{2}s\right\rfloor }^{\left\lceil \log_{2}2r\right\rceil }\mathcal{F}_{j}^{i}\end{align*}
2507: Then it follows using techniques similar to \cite{L96a}, see corollaries
2508: 3.9, 3.11 and 3.12 ibid. and lemma 2.8 of \cite{L96b} that for $s>C(G)$,
2509: \[
2510: \mathbb{P}(\mathcal{F}\cap\{ W^{i}[0,2r]\subset V^{i},\, i=1,2\})\geq c\left(\frac{s}{r}\right)^{\xi}.\]
2511: We couple $W^{i}$ to $R^{i}$ such that $(R^{1},W^{1})$ is independent
2512: from $(R^{2},W^{2})$, and consider the event $\mathcal{Q}$ from
2513: lemma \ref{lem:Lwlr41}. If $\mathcal{Q}$ did not occur, then $R^{i}$
2514: is sufficiently close to $W^{i}$ such that $W^{i}[0,S_{v,2r}^{i}]\subset V^{i}$
2515: implies $R^{i}[0,T_{v,r}^{i}]\subset U^{i}$, if $s>C(G)$. Further,
2516: $\mathcal{F}\setminus\mathcal{Q}$ also implies that $R^{1}[0,T_{v,r}^{1}]\cap R^{2}[0,T_{v,r}^{2}]=\emptyset$.
2517: Finally, lemma \ref{lem:Lwlr41} shows that if $s>C(G)$ then $\mathbb{P}(\mathcal{Q})\leq\frac{1}{2}c(s/r)^{\xi}$
2518: which finishes the lemma.
2519: \end{proof}
2520: \begin{cor*}
2521: Let $G$ be an isotropic graph, let $v^{1},v^{2}\in G$ and let $r>4|v^{1}-v^{2}|$.
2522: Let $R^{1}$ and $R^{2}$ be two walks starting from $v^{i}$ and
2523: stopped on $\partial B(v^{1},r)$. Then \[
2524: \mathbb{P}(R^{1}\cap R^{2}=\emptyset)\begin{cases}
2525: >c(G)(|v^{1}-v^{2}|/r)^{\xi}\\
2526: =0\end{cases}.\]
2527:
2528: \end{cor*}
2529: \begin{proof}
2530: Using lemma \ref{lem:lwlr42} we can fix a constant $\lambda=\lambda(G)$
2531: such that for all $|v^{1}-v^{2}|>\lambda$ the first choice happens.
2532: Hence assume $|v^{1}-v^{2}|\leq\lambda$ and use lemma \ref{lem:0n0}
2533: with $\epsilon=\frac{1}{8}$ and $s=\lambda$ and get that either
2534: \begin{enumerate}
2535: \item There are no two disjoint paths going from $v^{1}$ and $v^{2}$ to
2536: the exterior of $B(v^{1},\kappa(\frac{1}{8},\lambda,G))$, $\kappa$
2537: from lemma \ref{lem:0n0}. In this case the probability is $0$ for
2538: every $r>\kappa$.
2539: \item For $\mu=\max\kappa(\frac{1}{8},\lambda,G),C_{\ref{C:cor42}}$ there
2540: are two disjoint simple paths $\gamma^{i}$ starting from $v^{i}$
2541: and ending at $w^{i}$ satisfying $w^{i}\in B(v^{1},\mu)\setminus\overline{B(v^{1},\frac{7}{8}\mu)}$
2542: and $B(w^{i},\frac{7}{8}\mu)\cap\gamma^{3-i}=\emptyset$.
2543: \end{enumerate}
2544: In the second case we use lemma \ref{lem:lwlr42} and get\[
2545: \mathbb{P}^{1,w^{1},2,w^{2}}((\gamma^{1}\cup R^{1}[0,T_{v,r}^{1}])\cap(\gamma^{2}\cup R^{2}[0,T_{v,r}^{2}])=\emptyset)>c(G)(\mu/r)^{\xi}\]
2546: where $\gamma^{i}\cap R^{3-i}=\emptyset$ is satisfied because $B(w^{i},\frac{1}{4}\mu)\cap\gamma^{3-i}=\emptyset$,
2547: $\gamma^{i}\subset B(v^{1},\mu)$ and the event of lemma \ref{lem:lwlr42}
2548: includes that $R^{i}\cap B(v^{1},\mu)\subset B(w^{i},\frac{1}{4}\mu)$.
2549: In the case that the $R^{i}$ start from $v^{i}$, the probability
2550: that both follow $\gamma^{i}$ until its end is $>c(G)$, which proves
2551: the corollary for $r>4\mu$. For $r<4\mu$ the lemma will hold automatically
2552: for a sufficiently small constant in its definition.
2553: \end{proof}
2554:
2555: \subsection{The upper bound}
2556:
2557: Having settled the lower bound in theorem \ref{thm:kof}, we need
2558: only the following lemma, which is slightly stronger than the upper
2559: bound (again, we will need the stronger version in the next chapter).
2560:
2561: \begin{lem}
2562: \label{lem:kofup}Let $G$ be an isotropic graph of dimension $2$
2563: or $3$. Then for any $v^{1},v^{2}\in G$, if $R^{1}$ and $R^{2}$
2564: are two walks with $R^{i}$ starting from $v^{i}$ and stopped on
2565: $\partial B(v^{1},r)$ then \[
2566: \mathbb{P}(R^{1}\cap R^{2}=\emptyset)\leq C(G)\left(\frac{|v_{1}-v_{2}|}{r}\right)^{-\xi}\]
2567:
2568: \end{lem}
2569: \begin{proof}
2570: Assume that $\mathbb{P}(R^{1}\cap R^{2}=\emptyset)>0$ (in particular
2571: that $v^{1}\neq v^{2}$). Also assume w.l.o.g.~that $r>4|v^{1}-v^{2}|$.
2572: Let $a_{j}=2^{j}|v^{1}-v^{2}|$, $b_{j}=a_{j}^{1-c_{\ref{c:RWclose}}/2}$
2573: and $T_{j}^{i}:=T_{v^{1},a_{j}}^{i}$. Define\[
2574: g_{j}:=2^{j\xi}\mathbb{P}(R^{1}[0,T_{j}^{1}]\cap R^{2}[0,T_{j}^{2}]=\emptyset).\]
2575: The corollary to lemma \ref{lem:lwlr42} shows that $g_{j}>c(G)$.
2576: We need to show that $g_{j}\leq C(G)$. Let $W^{1}$ and $W^{2}$
2577: be Brownian motions coupled to $R^{1}$ and $R^{2}$ respectively,
2578: i.e.~the couples $(R^{1},W^{1})$ and $(R^{2},W^{2})$ are independent.
2579: Let $S_{j}^{i}=S_{v^{1},a_{j}}^{i}$. Examine the event $\mathcal{F}_{j}$
2580: that $j$ is the last step where the $W^{i}$ intersect, namely,\[
2581: \mathcal{F}_{j}^{1}=\{ W^{1}\left]S_{j}^{1},S_{n}^{1}\right]\cap W^{2}\left]S_{j}^{2},S_{n}^{2}\right]=\emptyset\}\cap\{ W^{1}[S_{j-1}^{1},S_{j}^{1}]\cap W^{2}[0,S_{j}^{2}]\neq\emptyset\}\]
2582: Define $\mathcal{F}_{j}^{2}$ replacing the roles of $W^{1}$ and
2583: $W^{2}$. $\{ R^{1}[0,T_{n}^{1}]\cap R^{2}[0,T_{n}^{2}]=\emptyset\}\cap\mathcal{F}_{j}^{1}$
2584: has a number of consequences:
2585:
2586: \smallskip{}
2587: \noindent \textbf{(i)} $R^{1}[0,T_{j-2}^{1}]\cap R^{2}[0,T_{j-2}^{2}]=\emptyset$.
2588: By definition this event has probability $2^{(2-j)\xi}g_{j-2}$.
2589:
2590: \smallskip{}
2591: \noindent \textbf{(ii)} Next we use the fact that the $W^{i}$ intersect
2592: while the $R^{i}$ don't. The corollary to lemma \ref{lem:subhaus}
2593: shows that \begin{align*}
2594: \mathbb{P}(\sh(W^{1}[S_{j-1}^{1},S_{j}^{1}],R^{1}[T_{j-2}^{1},T_{j+1}^{1})>b_{j}) & \leq C(G)\exp(-a_{j}^{c(G)})\\
2595: \mathbb{P}(\sh(W^{2}[0,S_{j}^{2}],R^{2}[0,T_{j+1}^{2}])>b_{j}) & \leq C(G)\exp(-a_{j}^{c(G)})\end{align*}
2596: and hence if we define $\mathcal{A}:=\{ d(R^{1}[T_{j-2}^{1},T_{j+1}^{1}],R^{2}[0,T_{j+1}^{2}])\leq2b_{j}\}$
2597: ($\mathcal{A}$ standing for {}``almost intersecting'') we get\begin{equation}
2598: \mathbb{P}\left(\left\{ W^{1}[S_{j-1}^{1},S_{j}^{1}]\cap W^{2}[0,S_{j}^{2}]\neq\emptyset\right\} \setminus\mathcal{A}\right)\leq C(G)\exp(-a_{j}^{c(G)}).\label{eq:WIRA}\end{equation}
2599: Next define the event $\mathcal{N}:=\left\{ R^{1}[T_{j-2}^{1},T_{j+2}^{1}]\cap R^{2}[0,T_{j+2}^{2}]=\emptyset\right\} $
2600: ($\mathcal{N}$ standing for {}``not intersecting''). Lemma \ref{lem:lwlr26}
2601: allows as to estimate $\mathbb{P}(\mathcal{A}\cap\mathcal{N})$: we
2602: use it with the parameters $v=v^{1}$, $w=R^{1}(T_{j-2}^{1})$, $r=a_{j+1}$,
2603: $M=2\xi$, $K=2$ and $\epsilon=c_{\ref{c:RWclose}}/2$. We get that
2604: with probability $\geq1-C_{\ref{C:MKepsG}}(2\xi,2,c_{\ref{c:RWclose}}/2,G)a_{j+1}^{-2\xi}$
2605: in $R^{2}$, \begin{equation}
2606: \mathbb{P}^{1}\left(\mathcal{A}\cap\mathcal{N}\Bigm|R^{2}\left[0,T_{j+2}^{1}\right],R^{1}(T_{j-2}^{1})\right)\leq a_{j}^{-\delta(2\xi,2,c_{\ref{c:RWclose}}/2,G)}.\label{eq:ANsmall}\end{equation}
2607: Notice that we used the strong Markov property from the stopping time
2608: $\min\big\{ t>T_{j-2}^{1}:R^{1}(t)\in B(v^{1},a_{j+1}),\, d(R^{1}(t),R^{2}[0,T_{j+2}^{1}])\leq2b_{j}\big\}$.
2609: Since the events that $R^{i}$ do not intersect up to $T_{j-2}^{i}$
2610: and everything that happens after the $T_{j-2}^{i}$ are dependant
2611: only through $R^{i}(T_{j-2}^{i})$, and since (\ref{eq:ANsmall})
2612: holds for any values of $R^{1}(T_{j-2}^{1})$ we get\begin{align*}
2613: \mathbb{P}(\mathcal{A}\cap\{ R^{1}[0,T_{j+2}^{1}]\cap R^{2}[0,T_{j+2}^{2}]=\emptyset\}) & \leq C(G)a_{j+1}^{-2\xi}+2^{(2-j)\xi}g_{j-2}a_{j}^{-c(G)}\leq\\
2614: & \stackrel{(*)}{\leq}C(G)2^{-j\xi}g_{j-2}a_{j}^{-c(G)}\end{align*}
2615: where in $(*)$ we used the lower bound. Adding (\ref{eq:WIRA}) we
2616: get\textbf{}\begin{multline}
2617: \mathbb{P}(\left\{ W^{1}[S_{j-1}^{1},S_{j}^{1}]\cap W^{2}[0,S_{j}^{2}]\neq\emptyset\right\} \cap\{ R^{1}[0,T_{j+1}^{1}]\cap R^{2}[0,T_{j+1}^{2}]=\emptyset\})\leq\\
2618: \leq C(G)2^{-j\xi}g_{j-2}a_{j}^{-c(G)}+C(G)\exp(-a_{j}^{c(G)})\leq C(G)2^{-j\xi}g_{j-2}a_{j}^{-c(G)}.\label{eq:kjsmall}\end{multline}
2619:
2620:
2621: \smallskip{}
2622: \noindent \textbf{(iii)} Finally we use the condition $W^{1}\left]S_{j+2}^{1},S_{n}^{1}\right]\cap W^{2}\left]S_{j+2}^{2},S_{n}^{2}\right]=\emptyset$.
2623: Here we employ lemma \ref{lem:Erskof} and together with (\ref{eq:kjsmall})
2624: we get\begin{equation}
2625: \mathbb{P}\left(\left\{ R^{1}[0,T_{n}^{1}]\cap R^{2}[0,T_{n}^{2}]\neq\emptyset\right\} \cap\mathcal{F}_{j}^{1}\right)\leq C(G)2^{-n\xi}g_{j-2}a_{j}^{-c(G)}.\label{eq:Fnk1}\end{equation}
2626:
2627:
2628: An identical calculation holds for $\mathcal{F}_{j}^{2}$. We are
2629: almost done! We need only remark that\begin{eqnarray*}
2630: \lefteqn{\mathbb{P}\bigg(\left\{ R^{1}[0,T_{n}^{1}]\cap R^{2}[0,T_{n}^{2}]=\emptyset\right\} \setminus\bigcup_{\substack{2\leq j\leq n-2\\
2631: i=1,2}
2632: }\mathcal{F}_{j}^{i}\bigg)\leq}\\
2633: & \qquad\qquad & \leq\mathbb{P}\left(W^{1}[S_{1}^{1},S_{n-2}^{1}]\cap W^{2}[S_{1}^{2},S_{n-2}^{2}]=\emptyset\right)\leq C2^{-n\xi}\leq C2^{-n\xi}g_{0}\end{eqnarray*}
2634: and we get\[
2635: g_{n}\leq C(G)\sum_{j=0}^{n-4}g_{j}\left(2^{-c(G)}\right)^{j}.\]
2636: By lemma 4.5 of \cite{L96b}, $g_{i}$ are bounded and the bound depends
2637: only on the isotropic structure constants of $G$ (we use here that
2638: $g_{0}\leq1$) so the lemma and theorem \ref{thm:kof} are proved.
2639: \end{proof}
2640:
2641: \section{\label{sec:Quasi-loops}Quasi-loops}
2642:
2643: Let $\gamma$ be a path in a $d$-Euclidean net and let $v\in\mathbb{R}^{d}$.
2644: We say that $\gamma$ has an $(s,r)$-quasi-loop near $v$ if there
2645: exists a couple of points $\gamma(i),\gamma(j)\in B(v,s)$ such that
2646: $\diam\gamma[i,j]\geq r$. In this case we write $v\in\mathcal{QL}(s,r,\gamma)$.
2647: We take the $v$-s in a grid such that the balls $B(v,s)$ cover $\mathbb{R}^{d}$
2648: and define\[
2649: \QL(s,r,\gamma):=\#(\mathcal{QL}(s,r,\gamma)\cap{\textstyle \frac{1}{d}}s\mathbb{Z}^{d}).\]
2650:
2651:
2652: Our purpose in this chapter is to prove that loop-erased random walk
2653: has no quasi loops in the following sense:
2654:
2655: \begin{thm}
2656: \label{thm:QL}Let $G$ be an isotropic graph of dimension two or
2657: three, and let $0<\epsilon<1$. Then there exists a $\delta=\delta(\epsilon,G)>0$
2658: such that for all $v\in G$, all $r>C(\epsilon,G)$ and any subset
2659: $v\in\mathcal{D}\subset B(v,r)$, \[
2660: \mathbb{E}^{v}\QL(r^{1-\epsilon},r^{1-\delta},\LE(R[0,T(\partial\mathcal{D})]))\leq C(\epsilon,G)r^{-\delta}.\]
2661:
2662: \end{thm}
2663: Dimensions two and three are very different. The proof for dimension
2664: two was done in the case of $\mathbb{Z}^{2}$ by Schramm \cite[lemma 3.4]{S00}
2665: and is practically the same in our more general settings (\cite[lemma 18]{K}
2666: is another variation on Schramm's argument). Therefore we shall only
2667: sketch the required elements in the end of the chapter. We shall concentrate
2668: on dimension three. It turns out that the techniques we use will rely
2669: heavily on the non-intersection exponent and therefore work only for
2670: isotropic graphs. Hence an interesting conjecture appears
2671:
2672: \begin{conjecture*}
2673: Theorem \ref{thm:QL} holds for any Euclidean net.
2674: \end{conjecture*}
2675: Again, this is true in dimension two, hence the interesting case is
2676: dimension three.
2677:
2678: It will be convenient in many places to consider discontinuous paths.
2679: Therefore, if $\gamma:\{1,\dotsc,n\}\to G$ is some function (without
2680: the restriction that $\gamma(i)$ and $\gamma(i+1)$ are neighbors),
2681: $\LE(\gamma)$ will be defined using the formula (\ref{eq:defLE})
2682: literally, and is a simple discontinuous path. Likewise we will define
2683: $\gamma_{1}\cup\gamma_{2}$ even if $\gamma_{1}(\len\gamma_{1})$
2684: is not a neighbor of $\gamma_{2}(1)$. If $\gamma$ is a (possibly
2685: discontinuous) path and $A$ is some set, then $\gamma\cap A$ would
2686: stand for the discontinuous path created in the natural way from the
2687: parts of $\gamma$ inside $A$, in order.
2688:
2689: Here and below when we say {}``$\gamma$ is a discontinuous path'',
2690: we do not exclude the possibility that it is in effect continuous.
2691:
2692:
2693: \subsection{Cut times}
2694:
2695: For any path $\gamma$ we define\[
2696: \cut(\gamma):=\{\gamma(i):\gamma[0,i]\cap\gamma[i+1,\len\gamma]=\emptyset\}.\]
2697: $i$-s satisfying the condition will be called \textbf{cut times}
2698: and the $\gamma(i)$-s will be called \textbf{cut points}. It is clear
2699: that $\cut\gamma\subset\LE(\gamma)$, indeed $\cut\gamma$ is contained
2700: in any connected subset of $\gamma$. It will also be convenient to
2701: define \[
2702: \cut(\gamma;t):=\{\gamma(i):i<t,\,\gamma[0,i]\cap\gamma[i+1,\len\gamma]=\emptyset\}.\]
2703: It has the useful property that $\cut(\gamma;t)$ is increasing in
2704: $t$ and decreasing as $\gamma$ is extended.
2705:
2706: For a random walk $R$, $\cut R$ is intimately related with the non-intersection
2707: exponent $\xi$ via time symmetry. Lemma \ref{lem:CARb0} below has
2708: the details, but first we need some simple preparations.
2709:
2710: \begin{lem}
2711: \label{lem:condnomtr}Let $G$ be a three dimensional Euclidean net,
2712: let $v\in\mathbb{R}^{3}$ and let $r>C(G)$. Let $R^{i}$ be random
2713: walks on $G$ starting from points in $B(v,r)$. Let $\mathcal{E}$
2714: be an event with depends only on $R^{i}[0,T_{v,r}^{i}]$ and let $\mathcal{F}$
2715: be an event that depends only on $R^{i}[T_{v,2r}^{i},\infty[$. Then\[
2716: \mathbb{P}(\mathcal{E}\cap\mathcal{F})\approx\mathbb{P}(\mathcal{E})\mathbb{P}(\mathcal{F}).\]
2717: The constant implicit in the $\approx$ notation may depend on the
2718: number of walks, and on the isotropic structure constants.
2719: \end{lem}
2720: \begin{proof}
2721: For every $x\in\partial B(v,r)$ and $y\in\partial B(v,2r)$ let $\pi_{x,y}$
2722: be the probability that a random walk starting from $x$ will hit
2723: $\partial B(v,2r)$ in $y$, and let $\pi_{y}$ be the probability
2724: that $R(T_{v,2r})=y$. By Harnack's inequality (lemma \ref{lem:Harnack})
2725: we have that $\pi_{x,y}\approx\pi_{x',y}$ for any $x,x'\in\partial B(v,r)$.
2726: Hence \[
2727: \pi_{y}\approx\pi_{x,y}.\]
2728: This gives\begin{align*}
2729: \mathbb{P}(\mathcal{E}\cap\mathcal{F}) & =\sum_{x^{i},y^{i}}\mathbb{P}\left(\mathcal{E}\cap\left\{ R^{i}(T_{v,r}^{i})=x^{i}\right\} _{i}\right)\prod_{i}\pi_{x^{i},y^{i}}\mathbb{P}(\mathcal{F}\,|\, R^{i}(T_{v,2r}^{i})=y^{i}\:\forall i)\\
2730: & \approx\sum_{x^{i},y^{i}}\mathbb{P}\left(\mathcal{E}\cap\left\{ R^{i}(T_{v,r}^{i})=x^{i}\right\} _{i}\right)\prod_{i}\pi_{y^{i}}\mathbb{P}(\mathcal{F}\,|\, R^{i}(T_{v,2r}^{i})=y^{i}\:\forall i)\\
2731: & =\mathbb{P}(\mathcal{E})\mathbb{P}(\mathcal{F})\qedhere\end{align*}
2732:
2733: \end{proof}
2734: \begin{lem}
2735: \label{lem:CARb0}Let $G$ be a three dimensional isotropic graph.
2736: Let $v\in G$ and $r>C(G)$. Define the annulus $A:=B(v,2r)\setminus\overline{B(v,r)}$.
2737: Let $w\in B(v,\frac{1}{2}r)$ and let $R^{1}$ be a random walk starting
2738: from $w$. Let \[
2739: \mathcal{C}:=\cut\left(R^{1}[0,\infty[;T_{v,4r}^{1}\right).\]
2740: Let $z\in B(v,\frac{1}{2}r)$ and let $R^{2}$ be a random walk starting
2741: from $z$ and stopped on $\partial B(v,4r)$. Then\newc{c:CARb0}\[
2742: \mathbb{P}(\mathcal{C}\cap R^{2}[0,T_{v,4r}^{2}]\cap A\neq\emptyset)>c_{\ref{c:CARb0}}(G).\]
2743:
2744: \end{lem}
2745: The proof is a relatively straightforward application of second moment
2746: methods, but is quite long. Hence we shall divide it into several
2747: shorter claims.
2748:
2749: \begin{sublem}\label{sublem:0n0}\newC{C:sub0n0}There exists a $C_{\ref{C:sub0n0}}(G)$
2750: such that for any $x\in G$ one of the following holds:
2751:
2752: \begin{enumerate}
2753: \item There are no two disjoint paths leading from $x$ to $\partial B(x,C_{\ref{C:sub0n0}})$.
2754: \item For any $r\geq C_{\ref{C:sub0n0}}$ there exists two disjoint simple
2755: paths $\gamma^{i}\subset B(x,r)$ that satisfy that if $y^{i}$ is
2756: the end point of $\gamma^{i}$ then \begin{equation}
2757: B(y^{i},{\textstyle \frac{1}{4}}r)\cap\gamma^{3-i}=\emptyset,\quad y^{i}\in B(x,r)\setminus\overline{B(x,{\textstyle \frac{7}{8}}r)}\label{eq:gamiB3i}\end{equation}
2758:
2759: \end{enumerate}
2760: \end{sublem}
2761:
2762: {}``disjoint paths'' here mean except the point $x$ common to both
2763:
2764: \begin{proof}
2765: [Subproof]Let $\lambda=\lambda(G)$ satisfy that any edge in $G$
2766: has length $\leq\lambda$. Then lemma \ref{lem:0n0} for $\epsilon=\frac{1}{8}$,
2767: $s=\lambda$ and all neighbors of $x$ gives the result with $C_{\ref{C:sub0n0}}=\kappa(\frac{1}{8},\lambda,G)$.
2768: \end{proof}
2769: Points $x$ for which there exist two disjoint paths leading outside
2770: $B(x,C_{\ref{C:sub0n0}})$ will be called $\mathcal{C}$\textbf{-capable}.
2771:
2772: \begin{sublem}\label{sublem:anyball}\newC{C:hasx}Any ball $B$
2773: of radius $C_{\ref{C:hasx}}(G)$ contains at least one $\mathcal{C}$-capable
2774: point $x$.
2775:
2776: \end{sublem}
2777:
2778: \begin{proof}
2779: [Subproof]Let $\gamma$ be a path in $B\cap G$ that the distance
2780: between its two ends $y^{1},y^{2}$ is $\geq2C_{\ref{C:hasx}}-C(G)$.
2781: We may assume $\gamma$ is simple (say by taking its loop-erasure).
2782: Let $x$ be the point of $\gamma$ closest to the plane exactly between
2783: $y^{1}$ and $y^{2}$. Then clearly the portions of $\gamma$ up to
2784: $x$ and from $x$ on are disjoint paths that lead to distance at
2785: least $C_{\ref{C:hasx}}-C(G)$, which proves the sublemma, if $C_{\ref{C:hasx}}$
2786: is sufficiently large.
2787: \end{proof}
2788: \begin{sublem}
2789:
2790: \label{sublem:NERxi}Let $x\in G$ and $\rho>C(G)$. Let $R^{1}$
2791: and $R^{2}$ be two random walks starting from $x$. Define subsets
2792: similar to (\ref{eq:defMi}) as follows:\begin{align}
2793: & V^{1}:=B(x,\rho/2)\cup\left(\overline{B(x,\rho)}\cap\left\{ y:\left\langle y,(1,0,0)\right\rangle \geq{\textstyle \frac{1}{4}}\rho\right\} \right),\nonumber \\
2794: & V^{2}:=B(x,\rho/2)\cup\left(\overline{B(x,\rho)}\cap\left\{ y:\left\langle y,(1,0,0)\right\rangle \leq-{\textstyle \frac{1}{4}}\rho\right\} \right).\label{eq:defMisimp}\end{align}
2795: Notice that $\overline{B(x,\rho)}$ above refers to closure in $G$.
2796: Define further events $\mathcal{V}:=\{ R^{i}[0,T_{x,\rho}^{i}]\subset V^{i}\}_{i=1,2}$
2797: and $\mathcal{N}:=\{ R^{1}[0,T_{x,\rho}^{1}]\cap R^{2}[1,T_{x,\rho}^{2}]=\emptyset\}$.
2798: Then\[
2799: \mathbb{P}^{x}(\mathcal{N})\approx\mathbb{P}^{x}\big(\mathcal{N}\cap\mathcal{V})\begin{cases}
2800: \approx\rho^{-\xi} & x\textrm{ is }\mathcal{C}\textrm{-capable}\\
2801: =0 & \textrm{otherwise}\end{cases}.\]
2802:
2803:
2804: \end{sublem}
2805:
2806: \begin{proof}
2807: [Subproof] The case that $x$ is not $\mathcal{C}$-capable is obvious
2808: if $\rho>C_{\ref{C:sub0n0}}$. In the second case, use sublemma \ref{sublem:0n0}
2809: with its $r$ equal to $\sigma:=\max\{ C_{\ref{C:sub0n0}},C_{\ref{C:cor42}}\}$
2810: ($C_{\ref{C:cor42}}$ from lemma \ref{lem:lwlr42}) and get two disjoint
2811: paths $\gamma^{i}$ ending in $y^{i}$ satisfying (\ref{eq:gamiB3i}).
2812: This allows to use lemma \ref{lem:lwlr42} with walks starting from
2813: the $y^{i}$, the $v$, $s$ and $r$ of lemma \ref{lem:lwlr42} equal
2814: to $x$, $\sigma$ and $\frac{1}{2}\rho$ respectively, and with $\eta=(1,0,0)$.
2815: We get\begin{equation}
2816: \mathbb{P}^{1,y_{1},2,y_{2}}\left(\mathcal{N}\cap\left\{ R^{i}[0,T_{x,\rho}^{i}]\subset U^{i}\right\} _{i=1,2}\right)\approx\rho^{-\xi}\label{eq:rhoxiM}\end{equation}
2817: where $U^{i}$ is defined in (\ref{eq:defMi}). In particular, $R^{i}\subset U^{i}$
2818: shows that $R^{i}\cap B(x,\sigma)\subset B(y^{i},\frac{1}{4}\sigma)$
2819: and hence from (\ref{eq:gamiB3i}) $R^{i}\cap\gamma^{3-i}=\emptyset$.
2820: Further, $R^{i}\subset U^{i}$ implies $\gamma^{i}\cup R^{i}\subset V^{i}$.
2821: Finally, since the probability that the $R^{i}$-s starting from $x$
2822: follow $\gamma^{i}$ until $y^{i}$ is $\approx1$ we get $\mathbb{P}(\mathcal{N}\cap\mathcal{V})\approx\rho^{-\xi}.$
2823: To finish the sublemma, notice that $\mathbb{P}(\mathcal{N})\leq C(G)\rho^{-\xi}$
2824: follows from lemma \ref{lem:kofup}.
2825: \end{proof}
2826: \begin{sublem}\label{sublem:NtHRt}
2827:
2828: Let $x\in A$ and let $R^{1}$ and $R^{2}$ be two random walks starting
2829: from $x$. Define $\mathcal{H}:=T^{1}(w)<T_{v,4r}^{1}$ and $\mathcal{N}':=\{ R^{1}[0,T^{1}(w)]\cap R^{2}[1,\infty]=\emptyset\}$.
2830: Then\[
2831: \mathbb{P}(\mathcal{N}'\cap\mathcal{H})\begin{cases}
2832: \geq c(G)r^{-1-\xi} & x\textrm{ is }\mathcal{C}\textrm{-capable}\\
2833: =0 & \textrm{otherwise}.\end{cases}\]
2834:
2835:
2836: \end{sublem}
2837:
2838: \begin{proof}
2839: [Subproof]We use sublemma \ref{sublem:NERxi} with $\rho=\frac{1}{4}r$
2840: and get that (assuming $x$ is $\mathcal{C}$-capable), that \begin{equation}
2841: \mathbb{P}(\mathcal{N}\cap\mathcal{V})\approx r^{-\xi}.\label{eq:NVRnz}\end{equation}
2842: Examining the structure of the $V^{i}$-s it is not difficult to see
2843: that one may construct six domains $\mathcal{S}^{i},\mathcal{H}^{i},\mathcal{D}^{i}\subset B(0,5)$
2844: with the following properties (see figure \ref{cap:SDH})%
2845: \begin{figure}
2846: \input{SDH2.pstex_t}
2847:
2848:
2849: \caption{\label{cap:SDH}The $\mathcal{S}^{i}$, $\mathcal{H}^{i}$ and $\mathcal{D}^{i}$
2850: in the figure are actually $v+r\mathcal{S}^{i}$, $v+r\mathcal{H}^{i}$
2851: and $v+r\mathcal{D}^{i}$ respectively. $\mathcal{H}^{2}$ is not
2852: shown, imagine it {}``far away'' inside $\mathcal{D}^{2}$.}
2853: \end{figure}
2854:
2855: \begin{enumerate}
2856: \item $\overline{\mathcal{S}^{i}},\overline{\mathcal{H}^{i}}\subset\mathcal{D}^{i}$
2857: and $\overline{\mathcal{S}^{i}}\cap\overline{\mathcal{H}^{i}}=\emptyset$.
2858: \item \label{enu:SiSi}If $r>C(G)$ then $\partial V^{i}\cap\partial B(x,\frac{1}{4}r)\subset v+r\mathcal{S}^{i}$.
2859: \item \label{enu:MiD3i}$\mathcal{D}^{1}\cap\mathcal{D}^{2}=\emptyset$
2860: and, if $r>C(G)$ then $V^{i}\cap(v+r\mathcal{D}^{3-i})=\emptyset$.
2861: \item $\mathcal{D}^{1}\subset B(0,3)$\emph{,} $B(0,\frac{1}{2})\subset\mathcal{H}^{1}$,
2862: and $B(0,4)\cap\mathcal{H}^{2}=\emptyset$.
2863: \item The collection of $\mathcal{S}^{i},\mathcal{H}^{i},\mathcal{D}^{i}$
2864: for all $x,v$ and $r$ satisfies the conditions of lemma \ref{lem:DD},
2865: \ref{enu:lemDD-K}.
2866: \end{enumerate}
2867: Condition \ref{enu:SiSi} ensures that under the event $\mathcal{V}$
2868: we have $R^{i}(T_{x,\rho}^{i})\in v+r\mathcal{S}^{i}$. Hence we can
2869: apply lemma \ref{lem:DDp} with $\mathcal{D}^{1},\mathcal{S}^{1},\mathcal{H}^{1}$and
2870: lemma \ref{lem:DD} with $\mathcal{D}^{2},\mathcal{S}^{2},\mathcal{H}^{2}$
2871: for the continuation of $R^{i}$ after $T_{x,\rho}^{i}$. We get that\begin{align*}
2872: \mathbb{P}(T^{1}(w)<T^{1}(\partial(v+r\mathcal{D}^{1}))\,|\, R^{1}[0,T_{x,\rho}^{1}]\subset V^{1}) & >c(G)/r.\\
2873: \mathbb{P}(T_{v,4r}^{2}<T^{2}(\partial(v+r\mathcal{D}^{2}))\,|\, R^{2}[0,T_{x,\rho}^{2}]\subset V^{2}) & >c(G)\end{align*}
2874: On the other hand, condition \ref{enu:MiD3i} ensures that if $R^{i}[T_{x,\rho}^{i},T(v+r\mathcal{H}^{i})]\subset v+r\mathcal{D}^{i}$
2875: then $R^{1}[0,T(w)]\cap R^{2}[T_{x,\rho}^{2},T(v+r\mathcal{H}^{2})]=\emptyset$
2876: and vice versa. Together with (\ref{eq:NVRnz}) we get\begin{gather*}
2877: \mathbb{P}\left(\mathcal{N}''\cap\left\{ R^{1}[0,T^{1}(w)]\subset B(v,3r)\right\} \right)>c(G)r^{-1-\xi}.\\
2878: \mathcal{N}'':=\left\{ R^{1}[0,T^{1}(\{ w\})]\cap R^{2}[1,T_{v,4r}^{2}]=\emptyset\right\} \end{gather*}
2879: This ends the sublemma since lemma \ref{lem:DDunbounded} with $\mathcal{D}=\mathbb{R}^{3}$
2880: and $\mathcal{H}=B(0,3)$ shows that the probability of $R^{1}$ to
2881: never hit $B(v,3r)$ after hitting $\partial B(v,4r)$ is $\geq c(G)$.
2882: \end{proof}
2883: Let $\mathcal{E}$ be an event on a space of curves. We say that $\mathcal{E}$
2884: is \textbf{loop-monotone} if $\mathcal{E}(\gamma)\Rightarrow\mathcal{E}(\gamma')$
2885: whenever $\gamma$ is $\gamma'$ with some loops added. In other words,
2886: adding loops can only hurt $\mathcal{E}$. A typical example of a
2887: loop-monotone event is $\{ x\in\mathcal{C}\}$ for some $x$ ($\mathcal{C}$
2888: from the statement of lemma \ref{lem:CARb0}). We shall use loop-monotonicity
2889: to encapsulate the idea of time reversal in a convenient way in the
2890: following sublemma:
2891:
2892: \begin{sublem}\label{sublem:timerevers}
2893:
2894: Let $v,w,x\in G$ and let $\mathcal{E}\subset\{ T(x)<T_{v,r}\}$ be
2895: a loop-monotone event on the space of curves on $G$ starting from
2896: $w$. Then\[
2897: \mathbb{P}^{w}(\mathcal{E}(R\left[0,\infty\right[))\approx\mathbb{P}^{1,x,2,x}\left(\mathcal{E}\left(\overleftarrow{R^{1}}\cup R^{2}\right)\cap\left\{ T^{1}(w)<T_{v,r}^{1}\right\} \right)\]
2898: where the notation $\overleftarrow{R^{1}}\cup R^{2}$ means taking
2899: $R^{1}[0,T^{1}(w)]$, reversing it (so that it starts from $w$ and
2900: ends at $x$), and concatenating $R^{2}[1,\infty]$ at its end $x$.
2901:
2902: \end{sublem}
2903:
2904: \begin{proof}
2905: [Subproof]Denote $\mathcal{W}=\#\{ t\geq1:R(t)=w\}$. The loop monotonicity
2906: of $\mathcal{E}$ gives \begin{equation}
2907: \mathbb{P}^{w}\left(\mathcal{E}\,|\,\mathcal{W}=k\right)\leq\mathbb{P}\left(\mathcal{E}\,|\,\mathcal{W}=0\right)\quad\forall k>0\label{eq:FWksW0}\end{equation}
2908: since conditioning by $\mathcal{W}=k$ is equivalent to adding $k$
2909: closed paths from $w$ to itself and then starting a walk conditioned
2910: to have $\mathcal{W}=0$. Hence we get\begin{align}
2911: \mathbb{P}(\mathcal{E}) & =\sum_{k\geq0}\mathbb{P}(\mathcal{E}\,|\,\mathcal{W}=k)\mathbb{P}(\mathcal{W}=k)\stackrel{(\ref{eq:FWksW0})}{\leq}\mathbb{P}(\mathcal{E}\,|\,\mathcal{W}=0)\leq\nonumber \\
2912: & \stackrel{(*)}{\leq}C(G)\mathbb{P}(\mathcal{E}\cap\{\mathcal{W}=0\}).\label{eq:noWW2}\end{align}
2913: where $(*)$ comes from the transience of $G$.
2914:
2915: Next we use the time-symmetry of random walk in the form (\ref{eq:symT})
2916: for the portion of the walk between $w$ and $x$. We get\begin{equation}
2917: \mathbb{P}^{w}(\mathcal{E}\cap\{\mathcal{W}=0\})\approx\mathbb{P}^{1,x,2,x}\left(\mathcal{E}\left(\overleftarrow{R^{1}}\cup R^{2}\right)\cap\left\{ T^{1}(w)<T^{1}(\{ x\}\cup\partial B(v,r))\right\} \right)\label{eq:timesym}\end{equation}
2918: where the $\approx$ sign hides the bounded quantity $\omega(w)/\omega(x)$.
2919:
2920: The last step is defining $\mathcal{G}:=\mathcal{E}\left(\overleftarrow{R^{1}}\cup R^{2}\right)\cap\left\{ T^{1}(w)<T_{v,r}^{1}\right\} $
2921: and $\mathcal{X}:=\#\{ t\in[1,T^{1}(w)]:R^{1}(t)=x\}$. The loop-monotonicity
2922: of $\mathcal{E}$ gives\[
2923: \mathbb{P}^{1,x,2,x}\left(\mathcal{G}\,|\,\mathcal{X}=k,\, R^{2}=\gamma\right)\leq\mathbb{P}\left(\mathcal{G}\,|\,\mathcal{X}=0,\, R^{2}=\gamma\right)\quad\forall k>0,\,\gamma\]
2924: which gives, like (\ref{eq:noWW2}), \[
2925: \mathbb{P}(\mathcal{G}\,|\, R^{2}=\gamma)\approx\mathbb{P}(\mathcal{G}\cap\{\mathcal{X}=0\}\,|\, R^{2}=\gamma)\quad\forall\gamma\]
2926: and summing over all paths $\gamma$ starting from $x$ we get\begin{equation}
2927: \mathbb{P}(\mathcal{G})\approx\mathbb{P}(\mathcal{G}\cap\{\mathcal{X}=0\}).\label{eq:returnloops}\end{equation}
2928: (\ref{eq:noWW2}), (\ref{eq:timesym}) and (\ref{eq:returnloops})
2929: together finish the proof.
2930: \end{proof}
2931: \begin{sublem}\label{sublem:condintrst}For any $x\in A$,\[
2932: \mathbb{P}(x\in\mathcal{C})\begin{cases}
2933: >c(G)r^{-1-\xi} & x\textrm{ is }\mathcal{C}\textrm{-capable}\\
2934: =0 & \textrm{otherwise.}\end{cases}\]
2935: \end{sublem}
2936:
2937: \begin{proof}
2938: [Subproof]This is an immediate consequence of sublemmas \ref{sublem:timerevers}
2939: and \ref{sublem:NtHRt}.
2940: \end{proof}
2941: This completes what we would need for the estimate of the first moment,
2942: and we move to the second moment, which is not really all that more
2943: complicated --- the complication from the fact that it is second moment
2944: are partially compensated by the fact that we need an upper bound
2945: rather than a lower.
2946:
2947: \begin{sublem}\label{sublem:1mlow}
2948:
2949: For any $x\in A$,\[
2950: \mathbb{P}(x\in\mathcal{C})\leq C(G)r^{-1-\xi}.\]
2951:
2952:
2953: \end{sublem}
2954:
2955: \begin{proof}
2956: [Subproof]Sublemma \ref{sublem:timerevers} shows that\begin{align*}
2957: \mathbb{P}^{w}(x\in\mathcal{C}) & \approx\mathbb{P}^{1,x,2,x}\left(\left\{ T^{1}(w)<T_{v,4r}^{1}\right\} \cap\left\{ R^{1}[0,T^{1}(w)]\cap R^{2}\left[1,\infty\right[=\emptyset\right\} \right)\leq\\
2958: & \leq\mathbb{P}^{1,x,2,x}\left(\left\{ T^{1}(w)<T_{v,4r}^{1}\right\} \cap\left\{ R^{1}[0,T_{x,r/4}^{1}]\cap R^{2}[1,T_{x,r/4}^{2}]=\emptyset\right\} \right)\\
2959: \intertext{\textrm{and lemma \ref{lem:condnomtr} shows that}} & \approx\mathbb{P}^{1,x}\left(T^{1}(w)<T_{v,4r}^{1}\right)\mathbb{P}^{1,x,2,x}\left(R^{1}[0,T_{x,r/4}^{1}]\cap R^{2}[1,T_{x,r/4}^{2}]=\emptyset\right).\end{align*}
2960: Sublemma \ref{sublem:NERxi} shows that the term on the right is
2961: $\leq C(G)r^{-\xi}$ while (\ref{eq:GvwS-lower-bound}) shows that
2962: the term on the left is $\approx r^{-1}$.
2963: \end{proof}
2964: \begin{sublem}\label{sublem:2m}
2965:
2966: For any $x_{1},x_{2}\in A$,\[
2967: \mathbb{P}(x_{1},x_{2}\in\mathcal{C})\leq C(G)(r|x_{1}-x_{2}|)^{-1-\xi}.\]
2968:
2969:
2970: \end{sublem}
2971:
2972: \begin{proof}
2973: [Subproof]First let us note that it is possible to assume $|x_{1}-x_{2}|>C(G)$
2974: since otherwise $\mathbb{P}(x_{1},x_{2}\in\mathcal{C})\leq\mathbb{P}(x_{1}\in\mathcal{C})$
2975: and then sublemma \ref{sublem:1mlow} applies. Moreover, it is enough
2976: to prove that\begin{gather*}
2977: \mathbb{P}\left(\left\{ x_{1},x_{2}\in\mathcal{C}\right\} \cap\mathcal{O}\right)\leq C(G)(r|x_{1}-x_{2}|)^{-1-\xi}\end{gather*}
2978: where $\mathcal{O}:=\{ T_{1}<T_{2}\}$ and $T_{i}$ is the last time
2979: $R$ is in $x_{i}$. The other case is just a renaming of $x_{1}$
2980: and $x_{2}$.
2981:
2982: Define now $\rho:=\frac{1}{16}|x_{1}-x_{2}|$ and $x=(x_{1}+x_{2})/2$.
2983: Denote $\mathcal{X}=\{ x_{1},x_{2}\in\mathcal{C}\}\cap\mathcal{O}$.
2984: $\mathcal{X}$ is loop-monotone, hence we may use sublemma \ref{sublem:timerevers}
2985: for $x_{1}$ and get\begin{align*}
2986: \mathbb{P}^{w}(\mathcal{X}) & \approx\mathbb{P}^{1,x_{1},2,x_{1}}\Big(\left\{ T^{1}(w)<T_{v,4r}^{1}\right\} \cap\left\{ R^{1}\cap R^{2}=\emptyset\right\} \cap\\
2987: & \quad\Big\{\exists t<T_{v,4r}^{2}:\left\{ R^{2}(t)=x_{2}\right\} \cap\left\{ \left(R^{1}\cup R^{2}[0,t])\right)\cap R^{2}\left[t+1,\infty\right[=\emptyset\right\} \Big\}\Big).\end{align*}
2988: where the $R^{i}$-s in the expression $R^{1}\cap R^{2}=\emptyset$
2989: stand for the walks until their natural ending, namely $R^{1}[0,T^{1}(w)]$
2990: and $R^{2}[1,\infty[$ respectively. Denote $R^{2}(T_{x_{1},2\rho}^{2})$
2991: by $y$ and {}``stop'' $R^{2}$ there, and consider the rest of
2992: $R^{2}$ as a new random walk $R^{3}$ starting from $y$. We get\begin{align}
2993: \mathbb{P}^{w}(\mathcal{X}) & \approx\sum_{y\in\partial B(x_{1},2\rho)}\mathbb{P}^{1,x_{1},2,x_{1},3,y}\Big(\left\{ T^{1}(w)<T_{v,4r}^{1}\right\} \cap\left\{ R^{2}(T_{x_{1},2\rho}^{2})=y\right\} \cap\nonumber \\
2994: & \qquad\cap\left\{ R^{1}\cap(R^{2}\cup R^{3})=\emptyset\right\} \cap\Big\{\exists t<T_{v,4r}^{3}:\left\{ R^{3}(t)=x_{2}\right\} \cap\nonumber \\
2995: & \qquad\cap\left\{ \left(R^{1}\cup R^{2}\cup R^{3}[0,t]\right)\cap R^{3}\left[t+1,\infty\right[=\emptyset\right\} \Big\}\Big)\label{eq:onesym}\end{align}
2996: where $R^{2}$ stands for $R^{2}[1,T_{x_{1},2\rho}^{2}]$ and $R^{3}$
2997: stands for $R^{3}[0,\infty[$. We use sublemma \ref{sublem:timerevers}
2998: again, this time for the random walk $R^{3}$ and the point $x_{2}$
2999: (it is easy to see that the corresponding event is loop-monotone for
3000: any value of $R^{1}$ and $R^{2}$). We get\begin{align}
3001: \mathbb{P}^{w}(\mathcal{X}) & \approx\sum_{y}\mathbb{P}^{1,x_{1},2,x_{1},3,x_{2},4,x_{2}}\Big(\left\{ T^{1}(w)<T_{v,4r}^{1}\right\} \cap\left\{ R^{2}(T_{x_{1},2\rho}^{2})=y\right\} \cap\nonumber \\
3002: & \cap\left\{ T^{3}(y)<T_{v,4r}^{3}\right\} \cap\left\{ R^{1}\cap(R^{2}\cup R^{3}\cup R^{4})=\emptyset\right\} \cap\nonumber \\
3003: & \cap\left\{ (R^{1}\cup R^{2}\cup R^{3})\cap R^{4}=\emptyset\right\} \Big)\label{eq:R1234}\end{align}
3004: where $R^{3}$ stands for $R^{3}[0,T^{3}(y)]$ and $R^{4}$ stands
3005: for $R^{4}[1,\infty[$. Reducing slightly the non-intersecting sections
3006: we may write\begin{align*}
3007: \mathbb{P}^{w}(\mathcal{X}) & \leq C(G)\sum_{y}\mathbb{P}\Big(\textrm{same first three conditions }\cap\\
3008: & \cap\left\{ R^{1}[0,T_{x_{1},\rho}^{1}]\cap R^{2}[1,T_{x_{1},\rho}^{2}]=\emptyset\right\} \cap\left\{ R^{3}[0,T_{x_{2},\rho}^{3}]\cap R^{4}[1,T_{x_{2},\rho}^{4}]=\emptyset\right\} \cap\\
3009: & \cap\left\{ R^{1}[T_{x,18\rho}^{1},T_{x,r/4}^{1}]\cap R^{4}[T_{x,18\rho}^{4},T_{x,r/4}^{4}]=\emptyset\right\} \Big).\end{align*}
3010:
3011:
3012: Denote the three non-intersection events above by $\mathcal{N}_{1}$,
3013: $\mathcal{N}_{2}$ and $\mathcal{N}_{3}$ by order. We understand
3014: that if $18\rho>r/4$ then $\mathcal{N}_{3}$ is considered to always
3015: be satisfied. Now, sublemma \ref{sublem:NERxi} shows that \[
3016: \mathbb{P}(\mathcal{N}_{1})\leq C(G)\rho^{-\xi}\quad\mathbb{P}(\mathcal{N}_{2})\leq C(G)\rho^{-\xi}\]
3017: and since these events are independent the probability of their intersection
3018: is $\leq C(G)\rho^{-2\xi}$. Assume for a moment that $18\rho\leq r/4$.
3019: Then we use lemma \ref{lem:condnomtr} for the ball $B(x,9\rho$)
3020: and get\[
3021: \mathbb{P}(\mathcal{N}_{1}\cap\mathcal{N}_{2}\cap\mathcal{N}_{3})\approx\mathbb{P}(\mathcal{N}_{1}\cap\mathcal{N}_{2})\mathbb{P}(\mathcal{N}_{3})\]
3022: and theorem \ref{thm:kof} shows that $\mathbb{P}(\mathcal{N}_{3})\leq C(G)(\rho/r)^{\xi}$,
3023: so in total \begin{equation}
3024: \mathbb{P}(\mathcal{N}_{1}\cap\mathcal{N}_{2}\cap\mathcal{N}_{3})\leq C(G)(r\rho)^{-\xi}.\label{eq:N123xi}\end{equation}
3025: If $18\rho>r/4$ then $(r\rho)^{-\xi}\approx\rho^{-2\xi}$ and (\ref{eq:N123xi})
3026: is again satisfied, so we can continue without the assumption $18\rho\leq r/4$.
3027:
3028: Finally we need to accommodate the various hitting and exit conditions
3029: in (\ref{eq:R1234}). Let $\mathcal{E}$ be the end points of the
3030: portions of the $R^{i}$-s needed for the $\mathcal{N}_{i}$-s, namely\[
3031: \mathcal{E}:=(R^{1}(T_{x,r/4}^{1}),R_{x_{1},\rho}^{2},R_{x_{2},\rho}^{3}).\]
3032: For the condition $T^{3}(y)<T_{v,4r}^{3}$ we use the fact that for
3033: any point $z$ where $R^{3}$ exits $B(x_{2},\rho)$ we have $|z-y|\geq\rho$
3034: and therefore the estimate of the harmonic potential (\ref{eq:atrans})
3035: gives\begin{equation}
3036: \mathbb{P}(T^{3}(y)<T_{v,4r}^{3}\,|\,\mathcal{E})\leq\mathbb{P}(T^{3}(y)<\infty\,|\,\mathcal{E})\leq C(G)\rho^{-1}.\label{eq:EE1}\end{equation}
3037: A similar argument for $R^{1}$ gives\begin{equation}
3038: \mathbb{P}(T^{1}(w)<T_{v,4r}^{1}\,|\,\mathcal{E})\leq C(G)r^{-1}.\label{eq:EE2}\end{equation}
3039: Conditioning over $\mathcal{E}$ the events of (\ref{eq:EE1}) and
3040: (\ref{eq:EE2}); $\mathbb{P}(R^{2}(T_{x_{1},2\rho}^{2})=y)$ and $\mathcal{N}_{1}\cap\mathcal{N}_{2}\cap\mathcal{N}_{3}$
3041: are all independent. Hence we get \begin{align*}
3042: \mathbb{P}^{w}(\mathcal{X}) & \stackrel{(*)}{\leq}\sum_{E}C(G)\rho^{-1}r^{-1}\mathbb{P}(\mathcal{N}_{1}\cap\mathcal{N}_{2}\cap\mathcal{N}_{3}\cap\{\mathcal{E}=E\})\:\cdot\\
3043: & \qquad\qquad\cdot\:\sum_{y\in\partial B(x_{1},2\rho)}\mathbb{P}(R^{2}(T_{x_{1},2\rho}^{2})=y\,|\,\mathcal{E}=E)=\\
3044: & =C(G)\rho^{-1}r^{-1}\mathbb{P}(\mathcal{N}_{1}\cap\mathcal{N}_{2}\cap\mathcal{N}_{3})\stackrel{(\ref{eq:N123xi})}{\leq}C(G)(\rho r)^{-1-\xi}.\end{align*}
3045: where in $(*)$ we used (\ref{eq:EE1}), (\ref{eq:EE2}) and independence.
3046: \end{proof}
3047:
3048:
3049: \begin{proof}
3050: [Proof of lemma \ref{lem:CARb0}]Let \[
3051: \mathcal{X}=\#\{\mathcal{C}\cap R^{2}[0,T_{v,4r}^{2}]\cap A\}.\]
3052: Sublemma \ref{sublem:condintrst} shows that\begin{align*}
3053: \mathbb{E}\mathcal{X} & =\sum_{x\in A}\mathbb{P}\left(x\in\mathcal{C}\right)\mathbb{P}\left(x\in R^{2}[0,T_{v,4r}^{2}]\right)\geq\\
3054: & \stackrel{(*)}{\geq}c(G)r^{-2-\xi}\#\{ x\in A:x\textrm{ is }\mathcal{C}\textrm{ capable}\}\stackrel{(**)}{\geq}c(G)r^{1-\xi}\end{align*}
3055: where in $(*)$ we used sublemma \ref{sublem:condintrst} to estimate
3056: $\mathbb{P}(x\in\mathcal{C})$ and (\ref{eq:GvwS-lower-bound}) to
3057: estimate $\mathbb{P}\left(x\in R^{2}[0,T_{v,4r}^{2}]\right)$; and
3058: $(**)$ follows from sublemma \ref{sublem:anyball}. Correspondingly
3059: we have\begin{align*}
3060: \mathbb{E}\mathcal{X}^{2} & =\sum_{x_{1},x_{2}\in A}\mathbb{P}(x_{1},x_{2}\in\mathcal{C})\mathbb{P}\left(x_{1},x_{2}\in R^{2}[0,T_{v,4r}^{2}]\right)\leq\\
3061: & \stackrel{(*)}{\leq}C(G)\sum_{x_{1},x_{2}\in A}(r|x_{1}-x_{2}|)^{-2-\xi}\stackrel{(**)}{\leq}C(G)r^{-2-\xi}\sum_{x_{1}\in A}\sum_{n=1}^{\log_{2}r}2^{n(1-\xi)}\leq\\
3062: & \stackrel{(\dagger)}{\leq}C(G)r^{2-2\xi}\\
3063: \end{align*}
3064: where $(*)$ follows from sublemma \ref{sublem:2m} for $\mathbb{P}(x_{1},x_{2}\in\mathcal{C})$
3065: and (\ref{eq:atrans}) for $\mathbb{P}(x_{1},x_{2}\in R^{2}[0,T_{v,4r}^{2}])$;
3066: where $(**)$ comes from the volume estimate $\#\{ x_{2}:|x_{1}-x_{2}|\in[2^{n},2^{n+1}[\}\approx2^{3n}$
3067: for $n>C(G)$ since our graph $G$ is roughly isometric to $\mathbb{R}^{3}$;
3068: and where $(\dagger)$ comes from the same volume estimate since $r>C(G)$,
3069: and (finally!) from $\xi<1$. The well known inequality $\mathbb{P}(\mathcal{X}>0)\geq(\mathbb{E}\mathcal{X})^{2}/\mathbb{E}\mathcal{X}^{2}$
3070: now finishes the lemma.
3071: \end{proof}
3072: \begin{cor*}
3073: Under the assumptions of lemma \ref{lem:CARb0}, \[
3074: \mathbb{P}^{1,w}(\forall z\in B(v,{\textstyle \frac{1}{2}}r),\,\mathbb{P}^{2,z}(\mathcal{C}\cap R^{2}[0,T_{v,4r}^{2}]\cap A\neq\emptyset)>c(G))>c(G).\]
3075:
3076: \end{cor*}
3077: \begin{proof}
3078: Denote the event inside the inner $\mathbb{P}$ symbol by $\mathcal{E}$.
3079: Then lemma \ref{lem:CARb0} shows that $\mathbb{P}^{1,w,2,v}(\mathcal{E})\geq c(G)$.
3080: This shows that\[
3081: \mathbb{P}^{1,w}(\mathbb{P}^{2,v}(\mathcal{E})>c(G))>c(G).\]
3082: Now, for any infinite path $\gamma$ starting from $w$, the function
3083: \[
3084: f(z)=\mathbb{P}^{2,z}(\mathcal{E}\,|\, R^{1}[0,\infty[=\gamma)\]
3085: is harmonic outside $A$ and in particular in $B(v,\frac{1}{2}r)$.
3086: Hence Harnack's inequality (lemma \ref{lem:Harnack}) shows that $\min_{z\in B(v,\frac{1}{2}r)}f(z)\geq cf(y)$
3087: which proves the corollary.
3088: \end{proof}
3089:
3090:
3091:
3092: \subsection{Conditioned random walks}
3093:
3094: \begin{lem}
3095: \label{lem:RAH}Let $G$ be a three dimensional isotropic graph. Let
3096: $v\in G$ and let $H\subset\mathbb{R}^{3}$ be a closed half space
3097: with $v\in\partial_{\textrm{cont}}H$. Let $r>C(G)$ and let $\Gamma\subset B(v,r)$,
3098: $d(\Gamma,H)>C_{\ref{C:arc}}$. Let $R$ be a random walk starting
3099: from $v$. Then\newc{c:RAH}\[
3100: \mathbb{P}(R(T_{v,r})\in H\,|\, R[0,T_{v,r}]\cap\Gamma=\emptyset)\geq c_{\ref{c:RAH}}(G).\]
3101:
3102: \end{lem}
3103: ($C_{\ref{C:arc}}$ is from lemma \ref{lem:arc}, page \pageref{lem:arc}.
3104: In particular $d(\Gamma,H)>C_{\ref{C:arc}}$ implies that the set
3105: of paths from $v$ to $\partial B(v,r)$ not intersecting $\Gamma$
3106: is non-empty --- use the lemma for a translation of $H$ by $C_{\ref{C:arc}}$)
3107:
3108: \begin{proof}
3109: The equivalent question for a Brownian motion can be solved by reflecting
3110: through $\partial_{\textrm{cont}}H$ the last section of the motion
3111: not intersecting $\partial_{\textrm{cont}}H$, with the result that
3112: the corresponding probability is $\geq\frac{1}{2}$. Our proof is
3113: a discrete version of this idea. Formally, denote by $\mathcal{H}$
3114: (respectively $\mathcal{H}^{-}$) the space of all paths from $v$
3115: to $\partial B(v,r)\cap H$ (respectively $\partial B(v,r)\setminus H$)
3116: not intersecting $\Gamma$. We shall dissect $\mathcal{H}^{-}$ to
3117: disjoint sets $N_{\gamma}^{-}$ indexed by $\mathcal{G}$:\[
3118: \mathcal{H}^{-}=\bigcup_{\gamma\in\mathcal{G}}N_{\gamma}^{-}\]
3119: and map each $N_{\gamma}^{-}$ into a set $N_{\gamma}\subset\mathcal{H}$
3120: such that the following holds:
3121: \begin{enumerate}
3122: \item \label{enu:prob}$\mathbb{P}(N_{\gamma})\geq c(G)\mathbb{P}(N_{\gamma}^{-})$.
3123: \item \label{enu:cover}Every path $h\in\mathcal{H}$ is contained in at
3124: most $C(G)$ different $N_{\gamma}$-s.
3125: \end{enumerate}
3126: Together these properties show that $\mathbb{P}(\mathcal{H})\geq c(G)\mathbb{P}(\mathcal{H}^{-})$
3127: or equivalently\[
3128: \mathbb{P}(R(T_{v,r})\in H\,|\, R[0,T_{v,r}]\cap\Gamma=\emptyset)\geq c(G)\mathbb{P}(R(T_{v,r})\not\in H\,|\, R[0,T_{v,r}]\cap\Gamma=\emptyset)\]
3129: which would conclude the lemma.
3130:
3131: The set $\mathcal{G}$ is the set of all paths $\gamma$ in $B(v,r)$
3132: avoiding $\Gamma$ such that $x:=\gamma(\len\gamma)\not\in H$ but
3133: its neighbor $x':=\gamma(\len\gamma-1)\in H$. For each $\gamma\in\mathcal{G}$
3134: the set $N_{\gamma}^{-}$ is the set of all paths that follow $\gamma$
3135: until its end and then avoid hitting $H\cup\Gamma$ until hitting
3136: $\partial B(v,r)$. It is clear that $\left\{ N_{\gamma}^{-}\right\} _{\gamma\in\mathcal{G}}$
3137: are disjoint sets covering $\mathcal{H}^{-}$. Take one $\gamma\in\mathcal{G}$,
3138: let $x$ be its end and denote $\rho:=d(x,\partial B(v,r))$. Clearly
3139: $\mathbb{P}(N_{\gamma}^{-})$ is the probability that $R$ follows
3140: $\gamma$ (denote it by $p_{\gamma}$) multiplied by the escape probability\[
3141: \mathbb{P}^{x}(T_{v,r}<T(\Gamma\cup H))\leq\mathbb{P}^{x}(T_{v,r}<T(H))\leq\mathbb{P}^{x}(T_{x,\rho}<T(H))\stackrel{(*)}{\leq}C(G)/\rho\]
3142: where $(*)$ comes from theorem \ref{lem:escape}.
3143:
3144: We shall now construct $N_{\gamma}$ under the assumption that $\rho$
3145: is bigger than some constant $\rho_{0}(G)$. The value of $\rho_{0}$
3146: will be fixed later on, but for now we need $\rho_{0}>4C_{\ref{C:arc}}$.
3147: We use lemma \ref{lem:arc} with the point $x'$, the radius $4C_{\ref{C:arc}}$
3148: and with the half-space $H'=H+B(0,C_{\ref{C:arc}})$ and we get that
3149: there exists a simple path $\delta'\subset\overline{B(x',4C_{\ref{C:arc}})}\setminus H'$
3150: from $x'$ to $\partial B(x',4C_{\ref{C:arc}})\cap\{ y\in H:d(y,\partial_{\textrm{cont}}H>C_{\ref{C:arc}}\}$.
3151: Let $\delta=\gamma\cup(x,x')\cup\delta'$. Note that $\Gamma\cap H'=\emptyset$
3152: and therefore also $\Gamma\cap\delta=\emptyset$. Let $N_{\gamma}$
3153: be the family of all paths that follow $\delta$ until its end and
3154: then stay inside $H$ until they exit $B(v,r)$. If $\rho\leq\rho_{0}$
3155: simply let $\delta'$ be the shortest path from $x$ to $\partial B(v,r)\setminus H$
3156: not intersecting $\Gamma$ and let $N_{\gamma}$ to contain only the
3157: path $\gamma\cup\delta'$.
3158:
3159: The lemma will be concluded once we show \ref{enu:prob} and \ref{enu:cover}.
3160: To see \ref{enu:prob}, first note that the case when $\rho\leq\rho_{0}$
3161: is obvious since then $\mathbb{P}(N_{\gamma}^{-})\approx p_{\gamma}\approx\mathbb{P}(N_{\gamma})$.
3162: In the case $\rho>\rho_{0}$, the length of $\delta'$ is $\leq C(G)$
3163: so the probability to follow $\delta$ is $\geq c(G)p_{\gamma}$.
3164: We use lemma \ref{lem:arc} again to get that the probability of a
3165: random walk starting from $y:=\delta(\len\delta)$ to hit \[
3166: \alpha:=\partial B(y,\rho)\cap\{ z:d(z,\partial Q)>{\textstyle \frac{1}{2}}\rho\}\]
3167: before $\partial H$ is $\geq C(G)/\rho$. Finally, lemma \ref{lem:DD}
3168: shows that for any $z\in\alpha$ a random walk starting from $z$
3169: has a probability $>c(G)$ to exit $B(v,r)$ before hitting $\partial H$.
3170: To use lemma \ref{lem:DD} we need to assume that $\rho$ is large
3171: enough, and this is the condition for $\rho_{0}$ which can now be
3172: fixed. All three together give \ref{enu:prob}.
3173:
3174: As for \ref{enu:cover}, it is easy to see that every $h\in\mathcal{H}$
3175: can belong to only boundedly many $N_{\gamma}$ for which $\rho\leq\rho_{0}$.
3176: Hence examine the case $\rho>\rho_{0}$ and let $h\in\mathcal{H}$.
3177: If $h\in N_{\gamma}$ then $x\not\in H$ but after $y$ all points
3178: of $h$ are in $H$ and the path between $x$ and $y$ is in $B(x,C(G))$.
3179: Therefore if we define $e(h)$ as the last vertex in $h\setminus H$
3180: we know that $x\in B(e(h)),C(G))$ and in particular has just $C(G)$
3181: possibilities. Since $\gamma$ is simply the part of $h$ up to $x$
3182: we see that it too has only $C(G)$ possibilities which shows \ref{enu:cover}
3183: and the lemma.
3184: \end{proof}
3185: \begin{lem}
3186: \label{lem:IGamIC}Let $G$ be a three dimensional isotropic graph
3187: and let $\epsilon>0$. Then there exist a $q=q(G)>0$ and a $\delta=\delta(\epsilon,G)>0$
3188: such that the following holds: Let $v\in G$, let $r>C(\epsilon,G$)
3189: and let $s\in[r,2r-\epsilon r]$. Let $\Gamma\subset B(v,s)$ be some
3190: set such that \begin{equation}
3191: \mathbb{P}^{v}(R[0,T_{v,4r}]\cap\Gamma\neq\emptyset)\leq\delta.\label{eq:IGamICassm}\end{equation}
3192: Let $w\in\partial B(v,s)$ be admissible (see below). Then \begin{gather}
3193: \mathbb{P}^{1,w}(\forall y\in B(w,\epsilon r),\,\mathbb{P}^{2,y}(\mathcal{I})>q)\,|\, R^{1}[0,T_{v,4r}^{1}]\cap\Gamma=\emptyset)>q,\label{eq:IGamIC}\\
3194: \mathcal{I}:=\left\{ \cut(R^{1}[0,T_{v,4r}^{1}];T_{w,\epsilon r}^{1})\cap R^{2}[0,T_{v,4r}^{2}]\neq\emptyset\right\} .\nonumber \end{gather}
3195:
3196: \end{lem}
3197: We call $w$ admissible if there exists a path $\gamma\subset\overline{B(w,16C_{\ref{C:arc}})}$
3198: starting from $w$ and ending outside $B(v,|v-w|+2C_{\ref{C:arc}})$
3199: which does not intersect $\Gamma$ (the constant $16C_{\ref{C:arc}}$
3200: will be used in lemma \ref{lem:L1L2} below to show that many admissible
3201: points exist).
3202:
3203: In words, the lemma says that the fact that $\cut(R^{1})$ is hittable
3204: does not change if one condition by not hitting $\Gamma$, even if
3205: one starts very close to $\Gamma$ --- the only condition is that
3206: $\Gamma$ is not very hittable ($<\delta$) from far away ($v$).
3207: The fact that $\epsilon$ affects only $\delta$ but not $q$ will
3208: play a significant role later on.
3209:
3210: \begin{proof}
3211: Let $\lambda=\lambda(G)$ be some parameter that will be fixed later.
3212: Denote also $\mu:=\epsilon/4\lambda$ and $\rho=\mu r$. Let $\gamma$
3213: be the path from the definition of admissibility of $w$ and assume
3214: w.l.o.g.~that it is simple (say by taking $\LE$). We use lemma \ref{lem:RAH}
3215: with the starting point being $w':=\gamma(\len\gamma)$; with $H'$
3216: being the half space orthogonal to the segment $[v,w]$ such that
3217: $w'\in\partial_{\textrm{cont}}H'$; and with the radius some $\rho'$
3218: to be fixed later. (Note that the condition of lemma \ref{lem:RAH}
3219: $d(\Gamma,H')>C_{\ref{C:arc}}$ will be fulfilled if $r$ is sufficiently
3220: large). We get that if $\rho'>C(G)$ then \[
3221: \mathbb{P}^{1,w'}(R^{1}(T_{w',\rho'}^{1})\in H'\,|\, R^{1}[0,T_{w',\rho'}^{1}]\cap\Gamma=\emptyset)\geq c(G).\]
3222: Let $H$ be the translation of $H'$ such that $w\in\partial_{\textrm{cont}}H$.
3223: On one side, the probability that a random walk $R^{1}$ starting
3224: from $w$ will follow $\gamma$ until $w'$ is $\geq c(G)$. On the
3225: other side, if $\rho'=\rho-C(G)$ for some $C(G)$ sufficiently large
3226: then for any point $x\in\partial B(w',\rho')\cap H'$ there is a probability
3227: $\geq c(G)$ that a random walk starting from $x$ will hit $\partial B(w,\rho)\cup\Gamma$
3228: in $\partial B(w,\rho)\cap H$. All these allow us to drop the $'$
3229: notations and we get\begin{equation}
3230: \mathbb{P}^{1,w}(R^{1}(T_{w,\rho}^{1})\in H\,|\, R^{1}[0,T_{w,\rho}^{1}]\cap\Gamma=\emptyset)\geq c(G).\label{eq:HHc}\end{equation}
3231: Denote this event by $\mathcal{H}$.
3232:
3233: Next define $\mathcal{C}=\cut(R^{1}[T_{w,\rho}^{1},T_{v,4r}^{1}];T_{w,\epsilon r}^{1})$
3234: and $A=B(w,\frac{1}{2}\epsilon r)\setminus\overline{B(w,\frac{1}{4}\epsilon r)}$.
3235: The corollary to lemma \ref{lem:CARb0} shows that, for any $x\in\partial B(w,\rho)$
3236: and some $c(G)$,\begin{eqnarray}
3237: \lefteqn{\mathbb{P}^{1,x}(\forall y\in B(w,{\textstyle \frac{1}{8}}\epsilon r),\,\mathbb{P}^{2,y}(\mathcal{C}\cap R^{2}\cap A\neq\emptyset)>c(G))\geq}\nonumber \\
3238: & & \geq\mathbb{P}^{1,x}(\forall y\in B(w,{\textstyle \frac{1}{8}}\epsilon r),\nonumber \\
3239: & & \qquad\mathbb{P}^{2,y}(\cut(R^{1}[0,\infty[;T_{w,\epsilon r}^{1})\cap R^{2}[0,T_{w,\epsilon r}^{2}]\cap A\neq\emptyset)>c(G))\geq\nonumber \\
3240: & & \geq c(G)\label{eq:yBwer8}\end{eqnarray}
3241: where the notation $R^{i}$ (e.g.~$R^{2}$ above) stands for $R^{i}[0,T_{v,4r}^{i}]$,
3242: and where the $T_{w,\rho}^{1}$ in the definition of $\mathcal{C}$
3243: is considered to be $0$ when starting from $x$. Since I promised
3244: to prove the lemma for any $y\in B(w,\epsilon r)$, just note that
3245: for any such $y$ we have that with probability $>c(G)$ the walk
3246: $R^{2}$ hits $B(w,\frac{1}{8}\epsilon r)$ and therefore (\ref{eq:yBwer8})
3247: holds for the larger ball too, i.e.\begin{equation}
3248: \mathbb{P}^{1,x}(\forall y\in B(w,\epsilon r),\,\mathbb{P}^{2,y}(\mathcal{C}\cap R^{2}\cap A\neq\emptyset)>c(G))\geq c(G).\label{eq:JJc}\end{equation}
3249: Denote this event by $\mathcal{J}$.
3250:
3251: Next we take into consideration $\lambda$. Using (\ref{eq:fEfRT})
3252: with Green's function $G(\cdot,w;\linebreak[4]B(v,4r))$ shows that\begin{equation}
3253: \mathbb{P}^{1,x}(R^{1}[T_{w,\epsilon r/4}^{1},T_{v,4r}^{1}]\cap B(w,\rho)\neq\emptyset)\leq\frac{C(G)}{\lambda}.\label{eq:KKClam}\end{equation}
3254: Denote this event by $\mathcal{K}$.
3255:
3256: The next step is saying, roughly, {}``if $\Gamma$ is not hittable,
3257: then conditioning by not hitting $\Gamma$ has no effect''. Formally,
3258: we assume that, for some $\nu=\nu(G)$ to be fixed later \begin{equation}
3259: \mathbb{P}^{1,x}(R^{1}\cap\Gamma\neq\emptyset)\leq\nu\quad\forall x\in\partial B(w,\rho)\cap H.\label{eq:assmpni}\end{equation}
3260: As we shall see later, this assumption will be satisfied with a proper
3261: choice of $\delta$. For now, this allows us to preform the following
3262: calculation, which will return us to a walk starting from $w$:\begin{align}
3263: \lefteqn{\mathbb{P}^{1,w}(\mathcal{J}\setminus\mathcal{K}\,|\, R^{1}\cap\Gamma=\emptyset)\geq}\nonumber \\
3264: & \quad\geq\mathbb{P}^{1,w}((\mathcal{J}\setminus\mathcal{K})\cap\mathcal{H})\,|\, R^{1}\cap\Gamma=\emptyset)=\nonumber \\
3265: & \quad=\sum_{x\in\partial B(w,\rho)\cap H}\mathbb{P}^{1,w}((\mathcal{J}\setminus\mathcal{K})\cap\left\{ R^{1}(T_{w,\rho}^{1})=x\right\} \,|\, R^{1}\cap\Gamma=\emptyset)=\nonumber \\
3266: & \quad\!\stackrel{(*)}{=}\!\sum_{x\in\partial B(w,\rho)\cap H}\mathbb{P}^{1,w}(R^{1}(T_{w,\rho}^{1})=x\,|\, R^{1}[0,T_{w,\rho}^{1}]\cap\Gamma=\emptyset)\cdot\nonumber \\
3267: & \qquad\qquad\cdot\frac{\mathbb{P}^{1,x}\left((\mathcal{J}\setminus\mathcal{K})\cap\left\{ R^{1}\cap\Gamma=\emptyset\right\} \right)}{\mathbb{P}^{1,w}(R^{1}\cap\Gamma=\emptyset\,|\, R^{1}[0,T_{w,\rho}^{1}]\cap\Gamma=\emptyset)}\geq\nonumber \\
3268: & \quad\!\!\stackrel{(**)}{\geq}\left(c(G)-\frac{C(G)}{\lambda}-\nu\right)\sum_{x}\mathbb{P}^{1,w}(R^{1}(T_{w,\rho}^{1})=x\,|\, R^{1}[0,T_{w,\rho}^{1}]\cap\Gamma=\emptyset)=\nonumber \\
3269: & \quad=\left(c(G)-\frac{C(G)}{\lambda}-\nu\right)\mathbb{P}^{1,w}(\mathcal{H}\,|\, R^{1}[0,T_{w,\rho}^{1}]\cap\Gamma=\emptyset)\geq\nonumber \\
3270: & \quad\!\stackrel{(\ref{eq:HHc})}{\geq}\!\left(c(G)-\frac{C(G)}{\lambda}-\nu\right)c(G)\label{eq:long}\end{align}
3271: where $(*)$ comes from the definition of conditioned probability;
3272: and $(**)$ comes from the estimates (\ref{eq:JJc}) for $\mathcal{J}$
3273: and (\ref{eq:KKClam}) for $\mathcal{K}$, from the assumption (\ref{eq:assmpni})
3274: and from bounding the denominator by $1$. Picking $\lambda$ sufficiently
3275: large and $\nu$ sufficiently small we get that the result of the
3276: computation is positive and dependant on the isotropic structure constants
3277: of $G$ only.
3278:
3279: Finally, notice that if $\mathcal{K}$ did not occur, i.e.~if $R^{1}[T_{w,\epsilon r/4}^{1},T_{v,4r}^{1}]$
3280: doesn't return to the ball $B(w,\rho)$ then \[
3281: \cut(R^{1}[T_{w,\rho}^{1},T_{v,4r}^{1}];T_{w,\epsilon r}^{1})\cap A\subset\cut(R^{1}[0,T_{v,4r}^{1}];T_{w,\epsilon r}^{1})\]
3282: so (\ref{eq:long}) gives us (\ref{eq:IGamIC}) with an appropriate
3283: choice of $q$. Hence we need only justify (\ref{eq:assmpni}).
3284:
3285: To see (\ref{eq:assmpni}) we use Harnack's inequality (lemma \ref{lem:Harnack-general})
3286: for the family of domains (see figure \ref{cap:DzEz})%
3287: \begin{figure}
3288: \input{DzEz.pstex_t}
3289:
3290:
3291: \caption{\label{cap:DzEz}The domains $\mathcal{D}_{z}$ and $\mathcal{E}_{z}$.}
3292: \end{figure}
3293: \begin{align*}
3294: \mathcal{D}_{z} & =B(0,2)\setminus\overline{B(0,|z|)}\\
3295: \mathcal{E}_{z} & =\left(B(z,2\mu)\setminus\overline{B(z,{\textstyle \frac{1}{2}}\mu)}\right)\cap\left\{ y:\langle y,z\rangle>|z|^{2}\right\} \end{align*}
3296: defined for $z\in B(0,2-\frac{1}{2}\epsilon)\setminus B(0,1)$. The
3297: function $f(x)=\mathbb{P}^{x}(R[0,T_{v,4r}]\cap\Gamma\neq\emptyset)$
3298: is harmonic in $x$ in the domain $v+r\mathcal{D}_{z}$, $z=(w-v)/r$,
3299: so we get for $r>C(\mu,G)$ that it is (up to constants depending
3300: on $G,\mu$) independent of $x\in v+r\mathcal{E}_{z}$. Note that
3301: for $r>C(\epsilon,G)$, $||z||\leq2-\epsilon-C(G)/r\leq2-\frac{1}{2}\epsilon$
3302: and $\partial B(w,\rho)\cap H\subset v+r\mathcal{E}_{z}$.
3303:
3304: We want to compare $f(x)$ with $f(v)$. We use (clause \ref{enu:lemDD-K}
3305: of) lemma \ref{lem:DD} with $\mathcal{D}=B(0,2)$, $\mathcal{S}=B(0,\frac{1}{2})$
3306: and $\mathcal{H}=\mathcal{E}_{z}$. We get\newc{c:RAHtmp}\newc{c:RAHtmp2}\begin{align}
3307: f(v) & \stackrel{(*)}{\geq}\mathbb{P}^{v}(T(v+r\mathcal{E}_{z})<T_{v,4r})\mathbb{E}f(R(T(v+r\mathcal{E}_{z})))\geq\nonumber \\
3308: & \!\stackrel{(**)}{\geq}\! c(\mu,G)\mathbb{E}f(R(T(v+r\mathcal{E}_{z})))\geq\nonumber \\
3309: & \stackrel{(\dagger)}{\geq}c(\mu,G)f(x)\qquad\forall x\in\partial B(w,\rho)\cap H\label{eq:fvfy}\end{align}
3310: where $(*)$ comes from the strong Markov property, $(**)$ comes
3311: from lemma \ref{lem:DD} and $(\dagger)$ comes from Harnack's inequality.
3312: This finishes the lemma --- choose $\delta(\epsilon,G)=\nu c(\mu,G)$
3313: and the assumption (\ref{eq:IGamICassm}) will imply (\ref{eq:assmpni}).
3314: \end{proof}
3315:
3316: \subsection{Wiener's shell test}
3317:
3318: For the next lemma we need to introduce a few notations. Let $v\in G$
3319: and $r>1$ be some number and let \begin{equation}
3320: A_{1}:=B(v,2r)\setminus\overline{B(v,r)}\quad A_{2}:=B(v,4r)\setminus\overline{B(v,{\textstyle \frac{1}{2}}r)}.\label{eq:defA12}\end{equation}
3321: The notation $\overline{B(v,r)}$ relates here to closure in $G$,
3322: not in $\mathbb{R}^{3}$. We shall denote $\partial A_{1}=\partial B(v,2r)\cup\partial B(v,r)$
3323: (the general definition of boundary in $G$ might be a little smaller).
3324: Both conventions apply to any annulus in this section.
3325:
3326: If $\gamma_{1},\dotsc,\gamma_{n}$ are discontinuous paths (usually
3327: we will consider $2$ or $3$) then we will consider $\LE(\gamma_{1}\cup\dotsb\cup\gamma_{n})$
3328: as composed of $n$ pieces, each one {}``coming from some $\gamma_{i}$'',
3329: and will denote them by $\LE_{i}(\gamma_{1}\cup\dotsb\cup\gamma_{n})$.
3330: Formally, denote $t_{0}=1$, $t_{n}=\len\LE(\gamma_{1}\cup\dotsb\cup\gamma_{n})$
3331: and $t_{i}$, $i=1,\dotsc,n-1$ to be the first $t$ such that $j_{t}>\len(\gamma_{1})+\dotsb+\len(\gamma_{i-1})$
3332: ($j_{t}$ from the definition of $\LE$, (\ref{eq:defLE})). Then\[
3333: \LE_{i}(\gamma_{1}\cup\dotsb\cup\gamma_{n}):=\begin{cases}
3334: \LE(\gamma_{1}\cup\dotsb\cup\gamma_{n})[t_{i-1},t_{i}-1] & t_{i}>t_{i-1}\\
3335: \emptyset & \textrm{otherwise.}\end{cases}\]
3336: Note that $\LE_{i}$ are simple and disjoint and that $\LE(\gamma_{1}\cup\dotsb\cup\gamma_{n})=\LE_{1}\cup\dotsb\cup\LE_{n}$.
3337:
3338: \begin{lem}
3339: \label{lem:L1L2}Let $G$ be a three dimensional isotropic graph and
3340: let $\epsilon>0$ and $\eta>0$. Then there exists a $\delta=\delta(\epsilon,\eta,G)$
3341: such that the following holds: Let $v\in G$, $r>C(\epsilon,\eta,G)$
3342: and $s\in[r,2r-\eta r]$. Let $A_{1}$ and $A_{2}$ be as in (\ref{eq:defA12}).
3343: Let $\gamma\subset A_{2}$ be a discontinuous path starting from $\partial B(v,\frac{1}{2}r)$
3344: and let $R^{1}$ be a random walk starting from some point in $\partial A_{1}$
3345: and stopped at $\partial A_{2}$. Let\[
3346: L=\LE(\gamma\cup R^{1}),\quad L_{1}=\LE_{1}(\gamma\cup R^{1}),\quad L_{2}=\LE_{2}(\gamma\cup R^{1}).\]
3347: Let $\mathcal{X}$ be the event that the following three events hold:
3348: \begin{enumerate}
3349: \item \label{enu:XX1}$L_{1}\subset B(v,s)$;
3350: \item \label{enu:XX2}$L_{2}\not\subset B(v,s+\eta r)$;
3351: \item \label{enu:XX3}$\mathbb{P}^{2,v}(R^{2}[0,T_{v,4r}^{2}]\cap L'\neq\emptyset)<\delta$
3352: where $L'=L[1,t]$ and $t$ is the first time when $L$ hits $\partial B(v,s+\eta r)$
3353: i.e.~$t:=\min\{ i:L(i)\not\in B(v,s+\eta r)\}$.
3354: \end{enumerate}
3355: Then\[
3356: \mathbb{P}(\mathcal{X})<\epsilon.\]
3357:
3358: \end{lem}
3359: In words, the probability that $R^{1}$ extends $\gamma$ even by
3360: a little ($\eta$) without being \linebreak[4]($\delta$-)hittable,
3361: is small. To get a clearer geometric picture, think about $\gamma$
3362: as the restriction of a continuous path to $A_{2}$, i.e.~as a sequence
3363: of paths coming in and (except the last one) ending in $\partial A_{2}$;
3364: and think about $R^{1}$ as starting from the end of $\gamma$.
3365:
3366: We remark that $\gamma$ may be empty, in which case \ref{enu:XX1}
3367: always holds, and in \ref{enu:XX2} one may replace $L_{2}$ with
3368: $L$.
3369:
3370: \begin{proof}
3371: Let $q=q(G)$ be from lemma \ref{lem:IGamIC}. Let $K=K(\epsilon,G)>2$
3372: be an integer such that \[
3373: (1-q)^{K}<\epsilon.\]
3374: Now fix the parameter $\epsilon$ of lemma \ref{lem:IGamIC} to be
3375: $\eta/2K$ and denote by $\lambda=\lambda(\epsilon,G)$ the result.
3376:
3377: \noindent \begin{center}\begin{tabular}{|c|c|c|c|}
3378: \hline
3379: Lemma \ref{lem:IGamIC}&
3380: $\epsilon$&
3381: $q$&
3382: $\delta$\tabularnewline
3383: \hline
3384: here&
3385: $\eta/2K$&
3386: $q$&
3387: $\lambda$\tabularnewline
3388: \hline
3389: \end{tabular}\end{center}
3390:
3391: Define $s_{i}=s+r\eta i/(K+2)$ for $i=1,\dots,K+1$. Let $j_{k}$
3392: be as in the definition of loop-erasure (\ref{eq:defLE}) so that
3393: $R^{1}[j_{k}+1,T^{1}(\partial A_{2})]$ is a random walk conditioned
3394: not to hit $\LE(\gamma\cup R^{1}[0,j_{k}])$. Define \[
3395: \tau_{i}':=\max\{ j_{k}\leq T^{1}(\partial A_{2}):\LE(\gamma\cup R^{1}[0,j_{k}])\subset B(v,s_{i})\}.\]
3396: As in lemma \ref{lem:IGamIC}, let $j_{k}$ be admissible if there
3397: exists a path $\delta\subset\overline{B(R^{1}(j_{k}),16C_{\ref{C:arc}})}$
3398: from $R^{1}(j_{k})$ to $\partial B(v,|v-R^{1}(j_{k})|+2C_{\ref{C:arc}})$
3399: which does not intersect $\LE(\gamma\cup R^{1}[0,j_{k}])$ and define
3400: \[
3401: \tau_{i}:=\min\{ j_{k}\geq\tau_{i}':j_{k}\textrm{ admissable}\}.\]
3402: We note that if $L_{1}\subset B(v,s)$ and $L_{2}\not\subset B(v,s_{i}+C_{\ref{C:arc}})$
3403: then $\tau_{i}$ is well defined. Indeed, let $t>\tau_{i}'$ ($\tau_{i}'$
3404: is obviously well defined) be the first $j_{k}$ such that $R^{1}(t)\not\in B(v,s_{i}+C_{\ref{C:arc}})$,
3405: and denote $s^{*}=|R^{1}(t)-v|-C_{\ref{C:arc}}$ so that $s^{*}>s_{i}$.
3406: We use lemma \ref{lem:arc} with the radius being $8C_{\ref{C:arc}}$
3407: and with $H$ being the half space tangent to $B(v,s^{*})$ and orthogonal
3408: to the segment $[v,R^{1}(t)]$. We get that there exists a path $\delta'\subset B(R^{1}(t),8C_{\ref{C:arc}})\setminus H$
3409: from $R^{1}(t)$ to some $x\in\partial B(R^{1}(t),8C_{\ref{C:arc}})\cap\{ x:d(x,\partial_{\textrm{cont}}H)>4C_{\ref{C:arc}}\}$.
3410: If $r$ is sufficiently large then this implies $d(x,v)>s^{*}+3C_{\ref{C:arc}}$.
3411: Let\[
3412: q_{0}:=\max\{ q:\delta'(q)\in\LE(\gamma\cup R^{1}[0,t])\}\]
3413: and let $t_{0}$ be the $j_{k}$ such that $R^{1}(t_{0})=\delta'(q_{0})$.
3414: Clearly $\delta'(q_{0})\not\in B(v,s_{i})$ so $t_{0}>\tau_{i}'$.
3415: Further, $\delta:=\delta'[q_{0},\len\delta']$ is a path from $R^{1}(t_{0})$
3416: to $x$ not intersecting $\LE(\gamma\cup R^{1}[0,t_{0}])$. Since
3417: $R^{1}(t_{0})\in B(v,s^{*}+C_{\ref{C:arc}})$ and $x\not\in B(v,s^{*}+3C_{\ref{C:arc}})$
3418: and since $B(R^{1}(t),8C_{\ref{C:arc}})\subset B(R^{1}(t_{0}),16C_{\ref{C:arc}})$
3419: we see that all the admissibility requirements are satisfied. Therefore
3420: $t_{0}$ is admissible and hence $\tau_{0}$ is well defined and $\leq t_{0}$.
3421:
3422: Fix some $i$ and assume that $\tau_{i}$ is well defined and that
3423: $\tau_{i}<T^{1}(\partial A_{2})$. Examine $R^{1}$ after $\tau_{i}$.
3424: The definition of $\tau_{i}$ considers only $\LE(R[0,\tau_{i}])$
3425: therefore it only affects $R^{1}$ after $\tau_{i}$ by conditioning
3426: it to not intersect $\Gamma_{i}:=\LE(\gamma\cup R^{1}[0,\tau_{i}]).$
3427: Hence lemma \ref{lem:IGamIC} applies. We get\begin{equation}
3428: \mathbb{P}(\mathcal{B}_{i}\,|\,\tau_{i},R^{1}[0,\tau_{i}])<1-q\label{eq:expcond}\end{equation}
3429: where $\mathcal{B}_{i}$ is the event that\renewcommand{\theenumi}{(\alph{enumi})}
3430: \begin{enumerate}
3431: \item \label{enu:BB1}$\tau_{i}$ is well defined,
3432: \item \label{enu:BB2}$\mathbb{P}^{2,v}(R^{2}[0,T_{v,4r}^{2}]\cap\Gamma_{i}\neq\emptyset)\leq\lambda$,
3433: \item \label{enu:BB3}For some $y\in B(w_{i},\rho)$, $\rho:=\eta r/2K$,
3434: \begin{equation}
3435: \mathbb{P}^{2,y}(\cut(R^{1}[\tau_{i},T_{v,4r}^{1}];T_{i}^{*})\cap R^{2}[0,T_{v,4r}^{2}]\neq\emptyset)\leq q.\label{eq:tilTv4r}\end{equation}
3436: where $T_{i}^{*}:=\min\{ t>\tau_{i}:R^{1}(t)\in\partial B(w_{i},\rho)\}$.
3437: \end{enumerate}
3438: \renewcommand{\theenumi}{(\roman{enumi})}The notation in (\ref{eq:expcond})
3439: might deserve some explanation: we are conditioning here on the fact
3440: that $\tau_{i}$ is well defined (so \ref{enu:BB1} is satisfied automatically),
3441: on its value which gives some information on $R$ beyond $\tau_{i}$
3442: and on the entire path from $0$ to $\tau_{i}$ (which gives $\Gamma_{i}$
3443: and $w_{i}$).
3444:
3445: It will be convenient to replace \ref{enu:BB1} with the stronger
3446: \begin{enumerate}
3447: \item [(a')]$\tau_{i+1}$is well defined
3448: \end{enumerate}
3449: and then replace (\ref{eq:tilTv4r}) with a slightly stronger condition:\begin{equation}
3450: \mathbb{P}^{2,y}(\cut(R^{1}[\tau_{i},\tau_{i+1}];T_{i}^{*})\cap R^{2}[0,T_{v,4r}^{2}]\neq\emptyset)\leq q.\label{eq:TilTauip1}\end{equation}
3451: This is possible since $T_{i}^{*}<\tau_{i+1}$ by our choice of $\rho$.
3452: Denote $\mathcal{B}_{i}$ with \ref{enu:BB1} and (\ref{eq:tilTv4r})
3453: replaced with (a') and (\ref{eq:TilTauip1}) by $\mathcal{B}_{i}'$
3454: and get $\mathbb{P}(\mathcal{B}_{i}'\,|\,\tau_{i},R^{1}[0,\tau_{i}])<1-q$.
3455: $\mathcal{B}_{i}'$ depends only on $\tau_{i+1}$ and $R[0,\tau_{i+1}]$
3456: and hence we can write\[
3457: \mathbb{P}(\mathcal{B}_{i}'\,|\,\mathcal{B}'_{0},\dotsc,\mathcal{B}_{i-1}')=\mathbb{EP}(\mathcal{B}_{i}'\,|\,\tau_{1},\dotsc,\tau_{i},R[0,\tau_{i}])=\mathbb{EP}(\mathcal{B}_{i}'\,|\,\tau_{i},R[0,\tau_{i}])<1-q\]
3458: (the $\mathbb{E}$ signs above stand for conditional expectation with
3459: respect to $\mathcal{B}_{i-1}',\dotsc,\mathcal{B}_{0}'$). Hence\[
3460: \mathbb{P}\Big(\bigcap_{i=0}^{K-1}\mathcal{B}_{i}'\Big)<(1-q)^{K}<\epsilon.\]
3461:
3462:
3463: The lemma will be finished when we show that for an appropriate choice
3464: of $\delta$ we have $\mathcal{X}\subset\cap\mathcal{B}_{i}'$. As
3465: explained above, conditions \ref{enu:XX1} and \ref{enu:XX2} in the
3466: definition of $\mathcal{X}$, show that all $\tau_{i}$ are well defined.
3467: Hence condition (a') in the definition of $\mathcal{B}_{i}'$ is satisfied
3468: for all $i$. Setting $\delta<\lambda$ will ensure condition \ref{enu:BB2}
3469: for all $i$, since $\Gamma_{i}\subset L'$ for all $i$. Finally,
3470: lemma \ref{lem:DD} shows that for some $\nu(\epsilon,\eta,G)$, \[
3471: \mathbb{P}^{v}(T(B(w_{i},\rho))<T_{v,4r})>\nu\quad\forall i\forall w_{i}\in\partial B(v,s_{i})\]
3472: which shows that setting $\delta<q\nu$ ensures condition \ref{enu:BB3}
3473: (we used here (\ref{eq:TilTauip1}) and $\cut(R^{1}[\tau_{i}+1,\tau_{i+1}];T_{i}^{*})\subset L'$).
3474: Hence $\mathcal{X}\subset\cap\mathcal{B}_{i}'$ and the lemma is proved.
3475: \end{proof}
3476: \begin{lem}
3477: \label{lem:wienersimp}Let $G$ be a three dimensional isotropic graph,
3478: and let $K$ be some number. Then there exists a $\delta=\delta(K,G)>0$
3479: such that the following holds. Let $v\in G$ and let $R^{1}$ be a
3480: random walk starting from $v$. Let $r>s>1$. Let $\mathcal{B}$ be
3481: the event that there exists some $T\geq0$ such that
3482: \begin{enumerate}
3483: \item $\LE(R^{1})\not\subset B(v,r)$ where $R^{1}:=R^{1}[0,T]$; and
3484: \item $\mathbb{P}^{2,v}(R^{2}[0,T_{v,r}^{2}]\cap\LE(R^{1})\cap(B(v,r)\setminus\overline{B(v,s)})=\emptyset)>(s/r)^{\delta}$.
3485: \end{enumerate}
3486: Then $\mathbb{P}(\mathcal{B})<C(K,G)(s/r)^{K}$.
3487: \end{lem}
3488: It is not difficult to see that the probability that $\LE(R^{1})\subset B(v,s)$
3489: is $\approx s/r$ (for $T\approx r^{2}$ which is the range that interests
3490: us). Therefore in fact $\LE(R)$ has a rather big probability to be
3491: {}``unhittable'' because it is small. The point about the lemma
3492: is that if it isn't small, this probability is negligible.
3493:
3494: \begin{proof}
3495: We may assume w.l.o.g.~that both $r/s$ and $s$ are sufficiently
3496: large (and the bound may depend on $K$). A Brownian motion starting
3497: from a point in $\partial B(0,1)$ has probability $\frac{2}{3}$
3498: to reach $\partial B(0,2)$ before $\partial B(0,\frac{1}{2})$. Hence
3499: by lemma \ref{lem:sball}, if $\rho>C(G)$ then random walk starting
3500: from $x\in\partial B(v,\rho)$ has probability $>\frac{7}{12}$ to
3501: reach $\partial B(v,2\rho)$ before $\partial B(v,\frac{1}{2}\rho)$.
3502: Hence if we define $\rho_{i}=s2^{i}$ and stopping times $\tau_{0}'=0$,\[
3503: \tau_{j}'=\min\{ t>\tau_{j-1}':R^{1}(t)\in\partial B(v,\rho_{i'(j)})\textrm{ for }i'(j)\textrm{ s.t. }R^{1}(\tau_{j-1}')\not\in\partial B(v,\rho_{i'(j)})\}\]
3504: then the process $i'(j)$ dominates a random walk on $\mathbb{N}$
3505: with a drift to infinity. In particular, it follows that if we denote
3506: \begin{equation}
3507: n_{i}':=\#\{ j:i'(j)=i\}\quad N':=\sum_{i=0}^{I}n_{i}'\label{eq:defniNp}\end{equation}
3508: then there exists a $\lambda=\lambda(K)$ such that\begin{equation}
3509: \mathbb{P}(N'>\lambda I)\leq C2^{-2IK}.\label{eq:NlamK}\end{equation}
3510: (\ref{eq:NlamK}) holds for any value of $I$, but we will define
3511: $I:=\left\lfloor \log_{2}r/s\right\rfloor $ since we are interested
3512: in what happens until $r$, and get $\mathbb{P}(N'>\lambda I)\leq C(s/r)^{K}$.
3513:
3514: Next we take the $I$ annuli between $s$ and $r$ and denote the
3515: even ones as {}``$A_{1}$-s'', i.e.~\[
3516: A_{1,i}:=B(v,2\rho_{2i})\setminus\overline{B(v,\rho_{2i})}\quad A_{2,i}:=B(v,4\rho_{2i})\setminus\overline{B(v,{\textstyle \frac{1}{2}}\rho_{2i})}\]
3517: and define stopping times $\tau_{j}$ {}``from one $A_{1}$ to the
3518: next'', i.e. $\tau_{1}:=\tau_{1}'$ and\[
3519: \tau_{j}=\min\{ t>\tau_{j-1}:R^{1}(t)\in\partial A_{1,i(j)}\textrm{ for some }i(j)\textrm{ such that }R^{1}(\tau_{j-1})\not\in\partial A_{1,i(j)}\}.\]
3520: It is clear that the $\tau_{j}$ are a subsequence of the $\tau_{j}'$
3521: hence if we define $n_{i}$ and $N=\sum_{i=0}^{\left\lfloor (I-1)/2\right\rfloor }n_{i}$
3522: analogously to (\ref{eq:defniNp}) then the analog of (\ref{eq:NlamK})
3523: will also hold. Define now $M:=\left\lceil 6\lambda\right\rceil $
3524: and get that \[
3525: \mathbb{P}(\#\{ i<{\textstyle \frac{1}{2}}I:n_{i}>M\}>{\textstyle \frac{1}{6}}I)\leq C(s/r)^{K}.\]
3526:
3527:
3528: Next use lemma \ref{lem:L1L2} with \[
3529: \eta=\frac{1}{M}\quad\epsilon=\frac{\mu}{M^{2}}\]
3530: where $\mu=\mu(K)$ is some parameter which will be fixed later. Call
3531: the $\delta$ of lemma \ref{lem:L1L2} $\nu$. Fix one $j$ and some
3532: $m\in\{0,\dotsc,M-1\}$. Denote $r_{j}=\rho_{i(j)}$ and $s_{j,m}=r_{j}(1+m/M)$.
3533: Define $\gamma_{i,j}:=\LE(R^{1}[0,\tau_{j}])\cap A_{2,i}$ and $\gamma_{j}:=\gamma_{i(j),j}$.
3534: Define $\mathcal{X}_{j,m}$ the events that the following conditions
3535: are all fulfilled:
3536: \begin{enumerate}
3537: \item \label{enu:YYj2}$L_{1,j}\subset B(v,s_{j,m})$ where\[
3538: L_{k,j}:=\LE_{k}(\gamma_{j}\cup R^{1}[\tau_{j},\tau_{j+1}])\quad k=1,2;\]
3539:
3540: \item \label{enu:YYj3}$L_{2,j}\not\subset B(v,s_{j,m+1})$;
3541: \item \label{enu:YYj4}$\mathbb{P}^{2,v}(R^{2}[0,T_{v,4r_{j}}^{2}]\cap L_{j,m}'\neq\emptyset)<\nu$
3542: where $L_{j,m}'=L_{1,j}\cup L_{2,j}[1,t]$ and $t$ is the first time
3543: when $L_{2,j}$ hits $\partial B(v,s_{j,m+1})$ i.e.~$t:=\min\{ i:L_{2,j}(i)\not\in B(v,s_{j,m+1})\}$.
3544: \end{enumerate}
3545: This completes all the parameters of lemma \ref{lem:L1L2} (see table
3546: \ref{cap:lemYYjm}) %
3547: \begin{table}
3548: \begin{tabular}{|c|c|c|c|c|c|c|}
3549: \hline
3550: Lemma \ref{lem:L1L2}&
3551: $\epsilon$&
3552: $\eta$&
3553: $\delta$&
3554: $r$&
3555: $s$&
3556: $A_{1}$\tabularnewline
3557: \hline
3558: here&
3559: $\frac{\mu}{M^{2}}$&
3560: $\frac{1}{M}$&
3561: $\nu$&
3562: $r_{j}$&
3563: $s_{j,m}$&
3564: $A_{1,i(j)}$\tabularnewline
3565: \hline
3566: Lemma \ref{lem:L1L2}&
3567: $A_{2}$&
3568: $\gamma$&
3569: $R^{1}$&
3570: $L_{k}$&
3571: $L'$&
3572: \emph{$\mathcal{X}$}\tabularnewline
3573: \hline
3574: here&
3575: $A_{2,i(j)}$&
3576: $\gamma_{j}$&
3577: $R^{1}[\tau_{j},\tau_{j+1}]$&
3578: $L_{k,j}$&
3579: $L_{j,m}'$&
3580: $\mathcal{X}_{j,m}$\tabularnewline
3581: \hline
3582: \end{tabular}
3583:
3584:
3585: \caption{\label{cap:lemYYjm}Use of lemma \ref{lem:L1L2} in lemma \ref{lem:wienersimp}.}
3586: \end{table}
3587: and we may use it to get $\mathbb{P}(\mathcal{X}_{j,m}\,|\, R[0,\tau_{j}])<\mu/M^{2}$
3588: and hence\[
3589: \mathbb{P}\Big(\bigcup_{m}\mathcal{X}_{j,m}\,\Big|\, R[0,\tau_{j}]\Big)<\frac{\mu}{M}.\]
3590: Denote this event by $\mathcal{X}_{j}$. The $\mathcal{X}_{j}$-s
3591: are therefore dominated by a sequence of independent random variables
3592: with probability $\mu/M$. A simple (and standard) calculation now
3593: shows that if $\mu$ is taken sufficiently small one has \[
3594: \mathbb{P}\Big(\#\{ j\leq\lambda I:\mathcal{X}_{j}\}>\frac{\lambda I}{M}\Big)\leq C(G)(s/r)^{K}.\]
3595: Fix $\mu$ to satisfy this condition. Thus we may define $\mathcal{G}$
3596: as the event that $N\leq\lambda I$ and $\mathcal{X}_{j}$ occurred
3597: less than $\frac{\lambda}{M}I\leq\frac{1}{6}I$ times before $\lambda I$.
3598: Our calculations show that $\mathbb{P}(\neg\mathcal{G})\leq C(s/r)^{K}$.
3599: The lemma will be finished once we show that with an appropriate choice
3600: of $\delta$ one has $\mathcal{B}\cap\mathcal{G}=\emptyset$.
3601:
3602: Assume therefore from now on that both $\mathcal{B}$ and $\mathcal{G}$
3603: occurred. There are $\left\lceil \frac{1}{2}I\right\rceil $ different
3604: $A_{1,i}$-s, and under the assumption $\mathcal{G}$, in $\leq\frac{1}{6}I$
3605: we have $n_{i}>M$ and in $\leq\frac{1}{6}I$ of them we have that
3606: some $\mathcal{X}_{j}$ occurred when $R(\tau_{j})\in\partial A_{1,i}$.
3607: Hence we get that at least $\frac{1}{6}I$ are {}``good'' in the
3608: sense that they are visited $\leq M$ times and are {}``$\mathcal{X}_{j}$-free''.
3609: Fix $i$ to be one such good index. We will now show that $A_{1,i}\cap\LE(R^{1})$
3610: is {}``hittable'', and we shall show that by induction.
3611:
3612: \begin{sublem}For any $j$ define \[
3613: n_{i,j}:=\#\{ k<j:R^{1}(\tau_{k})\in A_{1,i}\},\]
3614: $u_{i,j}:=\rho_{2i}(1+n_{i,j}/M)$ and $U_{i,j}:=B(v,u_{i,j})\setminus\overline{B(v,\rho_{2i})}$.
3615: Let also $\gamma_{i,j}^{*}$ be $\gamma_{i,j}[0,t_{i,j}]$ where $t_{i,j}$
3616: is the first time when $\gamma_{i,j}(t)\in\partial B(v,u_{i,j})$,
3617: or empty if $\gamma_{i,j}\cap\partial B(v,u_{i,j})=\emptyset$. Then
3618: for all $j$, either $\gamma_{i,j}^{*}=\emptyset$ or \[
3619: \mathbb{P}^{2,w}\left(\gamma_{i,j}^{*}\cap U_{i,j}\cap R^{2}[0,T_{4\rho_{2i}}^{2}]=\emptyset\right)\geq\nu.\]
3620:
3621:
3622: \end{sublem}
3623:
3624: Note that $n_{i}\leq M$ (because $i$ is a good index) and hence
3625: $u_{i,j}\leq2\rho_{2i}$ throughout the induction.
3626: \begin{proof}
3627: [Subproof]We use induction over $j$. If $R^{1}(\tau_{j})\not\in\partial A_{1,i}$,
3628: then $R^{1}[\tau_{j},\tau_{j+1}]$ does not enter $A_{1,i}$ and hence
3629: can only affect $\gamma_{i,j}^{*}\cap A_{1,i}$ by removing some components
3630: from its end. In this case it must remove the last component of $\gamma_{i,j}^{*}\cap A_{1,i}$,
3631: which is the component intersecting $\partial B(w,u_{i,j})$, and
3632: all the components of $\gamma_{i,j}$ after $\gamma_{i,j}^{*}$. Hence
3633: we get that $\gamma_{i,j+1}\cap\partial B(w,u_{i,j})=\emptyset$ (here
3634: $u_{i,j}=u_{i,j+1}$) and the induction holds in this case. Therefore
3635: we need only be interested in the case $R^{1}(\tau_{j})\in\partial A_{1,i}$.
3636: We know $\mathcal{X}_{j}$ did not occur, and in particular $\mathcal{X}_{j,n_{i,j}}$
3637: did not occur. Hence one of the three constituents of $\mathcal{X}_{j,n_{i,j}}$
3638: must have failed. Let us review them in order.
3639: \begin{enumerate}
3640: \item If $L_{1,j}\not\subset B(v,u_{i,j})$ (note that $s_{j,n_{i,j}}=u_{i,j}$)
3641: then
3642:
3643: \begin{itemize}
3644: \item $\gamma_{i,j}\cap\partial B(v,u_{i,j})\neq\emptyset$ and by the induction
3645: hypothesis, $\gamma_{i,j}^{*}$ is hittable.
3646: \item $R^{1}$ must have hit $\gamma_{i,j}$, if at all, after exiting $B(v,u_{i,j})$.
3647: Therefore, if $\gamma_{i,j+1}^{*}\neq\emptyset$ then it contains
3648: $\gamma_{i,j}^{*}$ so $\gamma_{i,j+1}^{*}\cap U_{i,j+1}$ is hittable
3649: and the induction holds.
3650: \end{itemize}
3651: \item If $L_{1,j}\cup L_{2,j}\subset B(v,u_{i,j+1})$ then $\gamma_{i,j+1}\cap\partial B(v,u_{i,j+1})=\emptyset$
3652: and the induction holds.
3653: \item Finally, if \ref{enu:YYj2}-\ref{enu:YYj3} of the definition of $\mathcal{X}_{j}$
3654: happened then we must have $\mathbb{P}^{2,w}(R^{2}[0,T_{w,4\rho_{2i}}^{2}]\cap L_{j,n_{i,j}}'\neq\emptyset)\geq\nu$
3655: and then the induction holds because $L_{j,n_{i,j}}'=\gamma_{i,j+1}^{*}$.
3656: \end{enumerate}
3657: Hence the sublemma is proved.
3658: \end{proof}
3659: Assume now that $\mathcal{B}$ happened and let $T$ be the {}``bad''
3660: time. Let $J$ be such that $\tau_{J}\leq T<\tau_{J+1}$. As in the
3661: sublemma, the part of the walk on $[\tau_{j},T]$ can affect in an
3662: adverse way only the part of the walk inside $A_{1,i(J)}$. Hence
3663: the sublemma shows that for any good $i\neq i(J)$, either $\LE(R^{1})$
3664: does not contain a crossing of $A_{1,i}$, or this crossing is hittable.
3665: However, if $\mathcal{B}$ happened then $\LE(R^{1})$ crosses $B(v,r)\setminus\overline{B(v,s)}$
3666: and hence all $A_{1,i}$-s so we get for all good $i\neq i(J)$ that\[
3667: \mathbb{P}^{2,v}\left(\LE(R^{1})\cap A_{1,i}\cap R^{2}[0,T_{v,4\rho_{2i}}^{2}]\neq\emptyset\right)\geq\nu.\]
3668: Harnack's inequality (lemma \ref{lem:Harnack-general}) shows that
3669: \begin{equation}
3670: \mathbb{P}^{2,w}\left(\LE(R^{1})\cap A_{1,i}\cap R^{2}[0,T_{v,4\rho_{2i}}^{2}]\neq\emptyset\right)>c\nu\quad\forall w\in B(v,{\textstyle \frac{1}{2}}\rho_{2i}).\label{eq:GBLARcnu}\end{equation}
3671: Hence let $i_{1}<i_{2}<\dotsc<i_{n}$ be good indices with $i_{k+1}-i_{k}\geq2$,
3672: $i_{k}\neq I,i(J)$. We can take $n:=\left\lfloor I/12\right\rfloor -2$.
3673: Then\begin{eqnarray*}
3674: \lefteqn{\mathbb{P}^{2,v}\left(R^{2}[0,T_{v,r}^{2}]\cap\LE(R^{1})\cap\left(B(v,r)\setminus\overline{B(v,s)}\right)=\emptyset\right)\leq}\\
3675: & & \leq\prod_{k=1}^{n-1}\mathbb{P}^{2,v}(R^{2}[T_{v,\frac{1}{2}\rho_{2i_{k}}}^{2},T_{v,4\rho_{2i_{k}}}^{2}]\cap\LE(R^{1})=\emptyset\,|\\
3676: & & \qquad\qquad\qquad\qquad|\, R^{2}[0,T_{v,4\rho_{2i_{k-1}}}^{2}]\cap\LE(R^{1})=\emptyset)\leq\\
3677: & & \leq\prod_{k=1}^{n-1}\max_{w\in B(v,\frac{1}{2}\rho_{2i_{k}})}\mathbb{P}^{2,w}(R^{2}[0,T_{v,4\rho_{2i_{k}}}^{2}]\cap\LE(R^{1})=\emptyset)<\\
3678: & & \!\stackrel{(\ref{eq:GBLARcnu})}{<}\!\prod_{k=1}^{n-1}(1-c\nu)\leq C(1-c\nu)^{I/12}\leq C(s/r)^{(\log(1-c\nu))/12\log2}.\end{eqnarray*}
3679: Therefore taking $\delta=(\log(1-c\nu))/24\log2$ and $r/s$ sufficiently
3680: large we get a contradiction, so $\mathcal{G}\cap\mathcal{B}=\emptyset$,
3681: and the lemma is proved.
3682: \end{proof}
3683:
3684: \subsection{Hittable sets}
3685:
3686: The next step (lemmas \ref{lem:outgoing} and \ref{lem:incoming})
3687: is to show that $\LE(R)$ is {}``hittable'' in a rather strong sense.
3688:
3689: \begin{defn*}
3690: Let $\lambda>0$ and $1\leq\rho<\sigma$. We define $\mathcal{H}^{\textrm{out}}(\lambda,\rho,\sigma)$
3691: to be the family of paths $\beta$ satisfying that for every $w$,
3692: and every subpath $\gamma\subset\beta$ which is an outgoing crossing
3693: of $B(w,\sigma)\setminus\overline{B(w,\rho)}$ i.e.~$\gamma(0)\in\partial B(w,\rho)$
3694: and $\gamma(\len\gamma)\in\partial B(w,\sigma)$ one has\[
3695: \mathbb{P}^{w}(R[0,T_{w,\sigma}]\cap\gamma)=\emptyset)\leq\lambda.\]
3696: Define $\mathcal{H}^{\textrm{out}}(\lambda,\mu)=\cap_{\rho\geq1}\mathcal{H}^{\textrm{out}}(\lambda,\rho,\mu\rho)$.
3697: \end{defn*}
3698: Similarly we define $\mathcal{H}^{\textrm{in}}$ to be the set where
3699: this holds for $\gamma$ incoming, i.e.~$\gamma(0)\in\partial B(w,\sigma)$
3700: and $\gamma(\len\gamma)\in\partial B(w,\rho)$. We define $\mathcal{H}:=\mathcal{H}^{\textrm{out}}\cap\mathcal{H}^{\textrm{in}}$.
3701: Note that these properties are hereditary: if $\beta\subset\gamma\in\mathcal{H}^{\textrm{out/in}}$
3702: then $\beta\in\mathcal{H}^{\textrm{out/in}}$.
3703:
3704: \begin{lem}
3705: \label{lem:expescape}Let $G$ be a Euclidean net and let $v\in G$
3706: and $r>1$. Then\[
3707: \mathbb{P}^{v}(R[0,t]\subset B(v,r))\leq C(G)e^{-c(G)t/r^{2}}.\]
3708:
3709: \end{lem}
3710: \begin{proof}
3711: This follows immediately by applying lemma \ref{lem:r2} repeatedly.
3712: \end{proof}
3713: \begin{lem}
3714: \label{lem:outgoing}Let $G,\epsilon,v,r$ and $\mathcal{D}$ be as
3715: in theorem \ref{thm:QL} and let $K>0$. Then there exists a $\delta=\delta(\epsilon,K,G)$
3716: such that \[
3717: \mathbb{P}^{1,v}(\exists T\leq T^{1}(\partial\mathcal{D}):\LE(R^{1}[0,T])\not\in\mathcal{H}^{\textrm{out}}(r^{-\delta},r^{\epsilon}))<C(\epsilon,K,G)r^{-K}.\]
3718:
3719: \end{lem}
3720: \begin{proof}
3721: Clearly we may assume $r$ is sufficiently large (and the bound may
3722: depend on $K$ and $\epsilon$). Since $\mathcal{H}^{\textrm{out}}(r^{-\delta},\rho,\sigma)$
3723: is increasing in $\sigma$ and decreasing in $\rho$, it is enough
3724: to show that for any fixed $\sigma=\frac{1}{2}\rho r^{\epsilon}$
3725: we have \begin{equation}
3726: \mathbb{P}^{1,v}(\exists T\leq T^{1}(\partial\mathcal{D}):\LE(R^{1}[0,T])\not\in\mathcal{H}^{\textrm{out}}(r^{-\delta},\rho,\sigma))\leq Cr^{-K-1}.\label{eq:fixrhosig}\end{equation}
3727: Once (\ref{eq:fixrhosig}) is established, we can apply it for $\rho=1,2,4,\dotsc,\log r$
3728: and bound the probability that (\ref{eq:fixrhosig}) holds for one
3729: such $\rho$ by the sum, which is $\leq Cr^{-K-1}\log r\leq Cr^{-K}$
3730: and the lemma would be proved.
3731:
3732: Fix therefore one $\rho$ and $\sigma=\frac{1}{2}\rho r^{\epsilon}$.
3733: Let $w\in\mathcal{D}$ be some point and let $A=B(w,\sigma)\setminus\overline{B(w,\rho)}$.
3734: Let $x\in\overline{B(w,\rho)}$ be some point, and let $\rho':=\max2|x-w|,\rho_{0}$
3735: for some constant $\rho_{0}(\epsilon,K,G)$ to be fixed later. Let
3736: $A':=B(x,\frac{1}{2}\sigma)\setminus\overline{B(x,\rho')}$ so that
3737: $A'\subset A$ for all $r$ sufficiently large. Let $t$ be some number.
3738: Use lemma \ref{lem:wienersimp} with the parameters and notations
3739: in table \ref{cap:lemWtW}. %
3740: \begin{table}
3741: \begin{tabular}{|c|c|c|c|c|c|c|c|}
3742: \hline
3743: Lemma \ref{lem:wienersimp}&
3744: $K$&
3745: $\delta$&
3746: $v$&
3747: $T$&
3748: $r$&
3749: $s$&
3750: \emph{$\mathcal{B}$}\tabularnewline
3751: \hline
3752: Here&
3753: $(K+10)/\epsilon$&
3754: $\lambda$&
3755: $x$&
3756: $t$&
3757: $\frac{1}{2}\sigma$&
3758: $\rho'$&
3759: $\mathcal{B}_{x}$\tabularnewline
3760: \hline
3761: \end{tabular}
3762:
3763:
3764: \caption{\label{cap:lemWtW}Use of lemma \ref{lem:wienersimp} in lemma \ref{lem:outgoing}.}
3765: \end{table}
3766: We get for the event $\mathcal{B}_{x}$ that there exists some $t$
3767: such that
3768: \begin{enumerate}
3769: \item $\LE(R^{1}[0,t])\not\subset B(x,\frac{1}{2}\sigma)$;
3770: \item \label{enu:BBx2}$\mathbb{P}^{2,x}(R^{2}[0,T_{x,\sigma/2}^{2}]\cap\LE(R^{1}[0,t])\cap A'=\emptyset)>C(\epsilon,K,G)r^{-\lambda\epsilon}$
3771: \end{enumerate}
3772: that $\mathbb{P}^{1,x}(\mathcal{B}_{x})<C(\epsilon,K,G)r^{-K-10}$.
3773: If $\rho_{0}$ is sufficiently large we can use Harnack's inequality
3774: to change in \ref{enu:BBx2} the starting point of $R^{2}$ from $x$
3775: to $w$ and pay only by increasing the constant.
3776:
3777: Returning to a random walk starting from $v$, we define an event
3778: $\mathcal{E}=\mathcal{E}(t_{1},t_{2},x)$ by
3779: \begin{enumerate}
3780: \item \label{enu:EExt1t2}$R^{1}(t_{1})=x$;
3781: \item \label{enu:EExt1t22}$\LE(R^{1}[0,t_{1}])\cap R^{1}[t_{1}+1,t_{2}]=\emptyset$;
3782: \item $\LE(R^{1}[t_{1}+1,t_{2}])\not\subset B(x,\sigma)$;
3783: \item $\mathbb{P}^{2,w}(R^{2}[0,T_{w,\sigma}^{2}]\cap\LE(R^{1}[t_{1}+1,t_{2}])\cap A=\emptyset)>c(\epsilon,K,G)r^{-\lambda\epsilon}$.
3784: \end{enumerate}
3785: and get $\mathbb{P}^{1,v}(\bigcup_{t_{2}\geq t_{1}}\mathcal{E})\leq C(\epsilon,K,G)r^{-K-10}$.
3786: Summing we get for any $U$,\begin{equation}
3787: \mathbb{P}\Big(\bigcup_{(*)}\mathcal{E}(t_{1},t_{2},x)\Big)\leq Cr^{-K-4}U\label{eq:Et1t2x}\end{equation}
3788: where the union is over all $w\in\mathcal{D}$, $x\in\overline{B(w,\rho)}$,
3789: $t_{1}\leq U$ and $t_{2}\geq t_{1}$. Now, lemma \ref{lem:expescape}
3790: shows that \begin{align*}
3791: \mathbb{P}(T(\partial\mathcal{D})>\mu r^{2}\log r) & \leq\mathbb{P}(T_{v,r}>\mu r^{2}\log r)\leq C(G)e^{-c(G)\mu\log r}\leq\\
3792: \intertext{\textrm{and for $\mu=\mu(K,G)$ sufficiently large}} & \leq Cr^{-K-1}.\end{align*}
3793: Therefore using $U=\mu r^{2}\log r$ in (\ref{eq:Et1t2x}) gives\[
3794: \mathbb{P}\Big(\bigcup_{(*)}\mathcal{E}(t_{1},t_{2},x)\Big)\leq Cr^{-K-1}\]
3795: where here the union is over all $w,x$ and $t_{1}\leq t_{2}\leq T^{1}(\partial\mathcal{D})$.
3796: This finishes the lemma by taking $\delta=\epsilon\lambda/2$ and
3797: $r$ sufficiently large, since if $\LE(R[0,T])\not\in\mathcal{H}^{\textrm{out}}$
3798: occurred then one of the $\mathcal{E}(t_{1},t_{2},x)$ must have occurred.
3799: Namely, let $\gamma$ be a subpath of $\LE(R[0,T])$ satisfying the
3800: requirements from the definition of $\mathcal{H}^{\textrm{out}}$
3801: and minimal (basically this means $\gamma\subset\overline{A}$) and
3802: assume $\gamma=\LE(R[0,T])[l,m]$. Then we may take $t_{1}:=j_{l}$
3803: ($j$ from the definition of $\LE$, (\ref{eq:defLE})), $t_{2}:=j_{m}$
3804: and $x=R^{1}(t_{1})$, and directly from the definitions, $\mathcal{E}(t_{1},t_{2},x)$
3805: will be satisfied. Therefore (\ref{eq:fixrhosig}) holds for any appropriate
3806: $\rho$ and $\sigma$ and the lemma is proved.
3807: \end{proof}
3808: The following lemma will be used only formally --- in effect the previous
3809: lemma is enough. We include it here mainly for completeness.
3810:
3811: \begin{lem}
3812: \label{lem:incoming}Lemma \ref{lem:outgoing} holds with $\mathcal{H}^{\textrm{out}}$
3813: replaced by $\mathcal{H}^{\textrm{in}}$.
3814: \end{lem}
3815: \begin{proof}
3816: Let $x\in\overline{\mathcal{D}}$ satisfy that \begin{equation}
3817: \mathbb{P}^{v}(R(T(\partial\mathcal{D}\cup\{ x\}))=x)\geq r^{-K-3}.\label{eq:xgood}\end{equation}
3818: The transience of $G$ shows that (compare to (\ref{eq:noWW2}))
3819: that\[
3820: \mathbb{P}^{v}(R(T(\partial\mathcal{D}\cup\{ v,x\}))=x)\geq c(G)r^{-K-3}.\]
3821: Using the symmetry of random walk in the form (\ref{eq:symT}) we
3822: get\begin{equation}
3823: \mathbb{P}^{x}(R(T(\partial\mathcal{D}\cup\{ v,x\}))=v)\geq c(G)r^{-K-3}.\label{eq:xgoodx}\end{equation}
3824: Next use lemma \ref{lem:outgoing} with the parameters in table \ref{cap:outgoing}
3825: (if $x\in\partial\mathcal{D}$ then use the lemma with each neighbor
3826: $x'$ of $x$ in $\mathcal{D}$ serving as the starting point instead
3827: of $x$). %
3828: \begin{table}
3829: \begin{tabular}{|c|c|c|c|c|}
3830: \hline
3831: Lemma \ref{lem:outgoing}&
3832: $v$&
3833: $\mathcal{D}$&
3834: $K$&
3835: $\delta$\tabularnewline
3836: \hline
3837: Here&
3838: $x$&
3839: $\mathcal{D}$&
3840: $2K+7$&
3841: $\delta$\tabularnewline
3842: \hline
3843: \end{tabular}
3844:
3845:
3846: \caption{\label{cap:outgoing}Use of lemma \ref{lem:outgoing} in lemma \ref{lem:incoming}.}
3847: \end{table}
3848: For any time or stopping time $t$, define an event\begin{equation}
3849: \mathcal{O}(t)=\{\LE(R[0,t])\in\mathcal{H}^{\textrm{out}}(r^{-\delta},r^{\epsilon})\}.\label{eq:defOt}\end{equation}
3850: With this notation, the conclusion of lemma \ref{lem:outgoing} is
3851: that\[
3852: \mathbb{P}^{x}(\exists t\leq T(\partial\mathcal{D}):\mathcal{O}(t))\leq C(K,G)r^{-2K-7}.\]
3853: In particular, $\mathbb{P}^{x}(\mathcal{O}(T(\partial\mathcal{D}\cup\{ v,x\})))\leq Cr^{-2K-7}$.
3854: Hence\[
3855: \mathbb{P}^{x}(\mathcal{O}(T(v))\,|\, R(T(\partial\mathcal{D}\cup\{ v,x\}))=v)\leq\frac{\mathbb{P}^{x}(\mathcal{O}(T(\partial\mathcal{D}\cup\{ v,x\})))}{\mathbb{P}^{x}(R(T(\partial\mathcal{D}\cup\{ v,x\})=v)}\stackrel{(\ref{eq:xgoodx})}{\leq}Cr^{-K-4}.\]
3856: Now, loop-erased random walk conditioned on the end vertex is symmetric
3857: (see e.g.~\cite[lemma 2]{K}) so,\[
3858: \mathbb{P}^{v}(\mathcal{I}(T(x))\,|\, R(T(\partial\mathcal{D}\cup\{ v,x\}))=x)=\mathbb{P}^{x}(\mathcal{O}(T(v))\,|\, R(T(\partial\mathcal{D}\cup\{ v,x\}))=v)\]
3859: where $\mathcal{I}$ is $\mathcal{O}$ for the reversed path, or,
3860: in other words, with $\mathcal{H}^{\textrm{out}}$ replaced by $\mathcal{H}^{\textrm{in}}$
3861: in (\ref{eq:defOt}). Now, since $\LE(R[0,T(\partial\mathcal{D}\cup\{ x\})])$
3862: does not depend on how many times we returned to $v$, we get\[
3863: \mathbb{P}^{v}(\mathcal{I}(T(x))\,|\, R(T(\partial\mathcal{D}\cup\{ x\}))=x)=\mathbb{P}^{v}(\mathcal{I}(T(x))\,|\, R(T(\partial\mathcal{D}\cup\{ v,x\}))=x)\]
3864: Lemma \ref{lem:omerB} now shows that, if $T_{n}=T_{n}(x)$ is the
3865: time of the $n$-th return to $x$ before hitting $\partial\mathcal{D}$
3866: then $\LE(R[0,T_{1}])\sim\LE(R[0,T_{n}])$ and therefore\begin{equation}
3867: \mathbb{P}^{v}(\mathcal{I}(T_{n})\,|\, T_{n}<T(\partial\mathcal{D}))\leq Cr^{-K-4}\quad n=1,2,\dotsc\label{eq:ITn}\end{equation}
3868:
3869:
3870: To finish the lemma we need only sum over $n$ and $x$. The transience
3871: of $G$ shows that \[
3872: \mathbb{P}^{v}(T_{n}>\lambda)\leq Ce^{-c\lambda}\]
3873: and therefore for $\lambda=C(G)\log r$ with $C$ sufficiently large
3874: we have that this probability is $\leq Cr^{-K-3}$. Therefore we get\begin{align*}
3875: \mathbb{P}\Big(\bigcup_{n}\{ T_{n}<T(\partial\mathcal{D})\}\cap\mathcal{I}(T_{n})\Big) & \leq Cr^{-K-3}+\sum_{n=1}^{\lambda}\mathbb{P}(\mathcal{I}(T_{n})\,|\, T_{n}<T(\partial\mathcal{D}))\\
3876: & \leq Cr^{-K-3}+(C\log r)r^{-K-4}\leq Cr^{-K-3}.\end{align*}
3877: Denoting by $\mathcal{X}$ the $x$-s satisfying (\ref{eq:xgood})
3878: we can sum over $x$ and get\begin{eqnarray*}
3879: \lefteqn{\mathbb{P}\Big(\bigcup_{n,x}\{ T_{n}(x)<T(\partial\mathcal{D})\}\cap\mathcal{I}(T_{n})\Big)\leq}\\
3880: & \qquad & \leq\sum_{x\in\mathcal{X}}\mathbb{P}\Big(\bigcup_{n}\{ T_{n}(x)<T(\partial\mathcal{D})\}\cap\mathcal{I}(T_{n})\Big)+\sum_{x\not\in\mathcal{X}}\mathbb{P}(T_{1}(x)<T(\partial\mathcal{D}))\\
3881: & & \leq(\#\mathcal{X})\cdot Cr^{-K-3}+\#(\overline{\mathcal{D}}\setminus\mathcal{X})\cdot Cr^{-K-3}\leq Cr^{-K}.\end{eqnarray*}
3882: Which finishes the lemma, since the event $\bigcup_{n,x}\{ T_{n}(x)<T(\partial\mathcal{D})\}\cap\mathcal{I}(T_{n})$
3883: is exactly the desired event.
3884: \end{proof}
3885: The combination of lemmas \ref{lem:outgoing} and \ref{lem:incoming}
3886: is that for an appropriate $\delta=\delta(\epsilon,K,G)$ we have\begin{equation}
3887: \mathbb{P}(\exists t\leq T(\partial\mathcal{D}):\LE(R[0,t])\in\mathcal{H}(r^{-\delta},r^{\epsilon}))\leq Cr^{-K}.\label{eq:inout}\end{equation}
3888:
3889:
3890:
3891: \subsection{Proof of theorem \refs{thm:QL} in three dimensions}
3892:
3893: The proof works in three different scales, which we will denote by
3894: $\rho\ll\sigma\ll\tau\ll r$. $\rho=r^{1-\epsilon}$ is the scale
3895: of the {}``closeness'' of the two ends of the quasi-loop. $\tau$
3896: is the scale of the diameter of the quasi-loop, so after all parameters
3897: are fixed we shall fix some $\delta>0$ and then define $\tau:=r^{1-\delta}$.
3898: $\sigma$ is an auxiliary scale --- rather than estimate for some
3899: point $w\in\frac{1}{3}\rho\mathbb{Z}^{3}$ that the probability that
3900: $w\in\mathcal{QL}(\rho,\tau,\LE(R[0,T(\partial\mathcal{D})]))$, we
3901: shall show it simultaneously for all points in $B(w,\sigma)\cap\frac{1}{3}\rho\mathbb{Z}^{3}$.
3902:
3903: \begin{lem}
3904: \label{lem:onegam}Let $G$ be a three dimensional isotropic graph,
3905: and let $\epsilon>0$ and $\delta>0$ be given. Let $1\leq\rho\leq\sigma\leq\frac{1}{2}\tau\leq\frac{1}{2}r$
3906: and $\sigma\geq2\rho(r^{\epsilon}+2)$. Let $w$ be some vertex and
3907: let $b\subset B(w,\tau)$ be some set. Let $\gamma\in\mathcal{H}=\mathcal{H}(r^{-\delta},r^{\epsilon})$
3908: be a path starting from some point in $\overline{B(w,\tau)}$ and
3909: ending on $\partial B(w,\tau)$. Let $R$ be a random walk starting
3910: from some $x\in\partial B(w,\sigma)$ and stopped at $\partial B(w,\tau)\cup b$;
3911: and define \[
3912: \Delta(\gamma):=\#\{ y\in{\textstyle \frac{1}{3}}\rho\mathbb{Z}^{3}\cap B(w,\sigma):d(y,\gamma)\leq\rho,\, d(y,R)\leq\rho\}\cdot\mathbf{1}_{\{ R\cap\gamma=\emptyset\}}\]
3913: Then \begin{align}
3914: \mathbb{E}^{x}\Delta(\gamma)\mathbf{1}_{\{ R\cap b=\emptyset\}} & \leq C(G)r^{-\delta}\log^{2}r,\label{eq:Delta1}\\
3915: \mathbb{E}^{x}\Delta(\gamma)\mathbf{1}_{\{ R\cap b\neq\emptyset\}} & \leq C(G)\mathbb{P}^{x}(R\cap b\neq\emptyset)\log^{2}r.\label{eq:Delta2}\end{align}
3916:
3917: \end{lem}
3918: \begin{proof}
3919: Denote the first by $\Delta_{1}(\gamma)$ and the second by $\Delta_{2}(\gamma)$.
3920: Define stopping times $s_{1}\leq s_{1}^{*}\leq s_{2}\leq\dotsc\leq T(\partial B(w,\tau)\cup b)$
3921: by \[
3922: s_{1}:=\min\{ t:d(R(t),\gamma\cap B(w,\sigma+\rho))\leq2\rho\},\]
3923: and for $i\geq1$ \begin{align*}
3924: s_{i}^{*} & :=\min\{ t\geq s_{i}:R(t)\in\partial B(R(s_{i}),8\rho)\}\\
3925: s_{i} & :=\min\{ t\geq s_{i-1}^{*}:d(R(t),\gamma\cap B(w,\sigma+\rho))\leq2\rho\}\end{align*}
3926: (for clarity we removed the conditions $s_{i},s_{i}^{*}\leq T(\partial B(w,\tau)\cup b)$
3927: --- if $R$ hits $\partial B(w,\linebreak[1]\tau)\cup b$ after any
3928: of them, consider the sequence stabilized at this point). Let $I$
3929: be the number of $s_{i}$'s defined before the process is stopped
3930: i.e.~$R(s_{I+1})\in\partial B(w,\tau)\cup b$. Lemma \ref{lem:Beurling}
3931: gives for $i<I$ that the probability to intersect $\gamma$ between
3932: $s_{i}$ and $s_{i}^{*}$ is $\geq c(G)/\log r$. This gives \begin{equation}
3933: \mathbb{P}(I\geq i)\leq\left(1-\frac{c(G)}{\log r}\right)^{i-1}\quad\forall i.\label{eq:lexpologr}\end{equation}
3934: Further, for the time period between $s_{i}$ and $s_{I+1}$ the
3935: definition of $\mathcal{H}$ gives that \[
3936: \mathbb{P}^{x}(\left\{ R[s_{i},s_{I+1}]\cap\gamma=\emptyset\right\} \cap\left\{ R(s_{I+1})\not\in b\right\} \,|\, I\geq i)\leq r^{-\delta}.\]
3937: Here is where we use the condition $\sigma\geq2\rho(r^{\epsilon}+2)$,
3938: since $d(R(s_{i}),\partial B(w,\tau))\geq\tau-\sigma-3\rho\geq\sigma-3\rho$.
3939: Finally, it is clear that $\Delta(\gamma)<CI$. Therefore\begin{align}
3940: \mathbb{E}^{x}\Delta_{1}(\gamma) & =\sum_{i=0}^{\infty}\mathbb{E}^{x}(\Delta_{1}(\gamma)\cdot\mathbf{1}_{\{ I=i\}})\leq\sum_{i=0}^{\infty}Ci\cdot\mathbb{P}^{x}(\left\{ \Delta_{1}(\gamma)>0\right\} \cap\left\{ I\geq i\right\} )\leq\nonumber \\
3941: & \leq C\sum_{i=1}^{\infty}i\mathbb{P}^{x}(\left\{ R(s_{I+1})\not\in b\right\} \cap\{ R\cap\gamma=\emptyset\}\cap\{ I\geq i\})\leq\nonumber \\
3942: & \leq C\sum_{i=1}^{\infty}i\left(1-\frac{c(G)}{\log r}\right)^{i-1}\cdot Cr^{-\delta}\leq C(G)r^{-\delta}\log^{2}r.\label{eq:Xvg_Y}\end{align}
3943:
3944:
3945: As for $\mathbb{E}^{x}\Delta_{2}(\gamma)$, we first notice that if
3946: $b$ is empty then there is nothing to prove. If $b$ is not empty,
3947: then the estimate of Green's function (lemma \ref{lem:a}) shows that
3948: $\mathbb{P}(R\cap b\neq\emptyset)>C/r$. Therefore we may use (\ref{eq:lexpologr})
3949: to show that If $\lambda=\lambda(G)$ is a sufficiently large constant
3950: then\[
3951: \mathbb{P}(I>\lambda\log^{2}r)\leq r^{-4}\]
3952: Hence we get\begin{align*}
3953: \mathbb{E}^{x} & \Delta_{2}(\gamma)\leq\mathbb{E}^{x}\Delta_{2}(\gamma)\mathbf{1}_{\{ I>\lambda\log^{2}r\}}+\mathbb{E}^{x}\Delta_{2}(\gamma)\mathbf{1}_{\{ I\leq\lambda\log^{2}r\}}\leq\\
3954: & \leq C\left(\frac{\sigma}{\rho}\right)^{3}r^{-4}+C\lambda(\log r)^{2}\mathbb{P}^{x}(R\cap b\neq\emptyset)\leq C\mathbb{P}^{x}(R\cap b\neq\emptyset)\log^{2}r.\qedhere\end{align*}
3955:
3956: \end{proof}
3957: Returning to the proof of theorem \ref{thm:QL}, we shall from this
3958: point on assume that $2r^{1-\epsilon/2}\leq\sigma\leq r^{1-2\delta}$.
3959: We fix some $w\in\mathcal{D}$ and some stopping time $T$ and define\[
3960: \mathcal{X}(T):=\mathcal{X}_{w}(T):=\#(\mathcal{QL}(\rho,\tau,\LE(R[0,T]))\cap B(w,\sigma)\cap{\textstyle \frac{1}{3}}\rho\mathbb{Z}^{3})\]
3961: so $\mathcal{X}:=\mathcal{X}(T(\partial\mathcal{D}))$ is what we
3962: need to estimate. Define exit and entry stopping times by $T_{0}=0$
3963: and\begin{align*}
3964: T_{2i+1} & =\min\{ t\geq T_{2i}:R(t)\in\partial B(w,2\sigma)\cup\partial\mathcal{D}\}\\
3965: T_{2i} & =\min\{ t\geq T_{2i-1}:R(t)\in\partial B(w,4\sigma)\cup\partial\mathcal{D}\}.\end{align*}
3966: Let $I$ be the first $i$ such that $R(T_{i})\in\partial\mathcal{D}$.
3967: We note that lemma \ref{lem:DDunbounded} shows that \[
3968: \mathbb{P}(R(T_{2i+1})\in\partial\mathcal{D}\,|\, R[0,T_{2i}])\geq\mathbb{P}^{R(T_{2i})}(T_{w,2\sigma}=\infty)\geq c(G)\]
3969: and hence we get that the probability that $I$ is large drops exponentially,
3970: and hence if $\lambda=\lambda(G)$ is sufficiently large we get\begin{equation}
3971: \mathbb{P}(I>\lambda\log r)\leq\frac{1}{r^{4}}.\label{eq:Ilamr3}\end{equation}
3972: Denote $M:=\left\lfloor \lambda\log r\right\rfloor $. With this notation
3973: we can write\begin{equation}
3974: \mathbb{E}\mathcal{X}\leq C\left(\frac{\sigma}{\rho}\right)^{3}\mathbb{P}(I>M)+\sum_{i=1}^{M}\mathbb{E}(\mathcal{X}\cdot\mathbf{1}_{\{ I=i\}})\stackrel{(\ref{eq:Ilamr3})}{\leq}Cr^{-1}+\sum_{i=1}^{M}\mathbb{E}(\mathcal{X}\cdot\mathbf{1}_{\{ I=i\}}).\label{eq:X_sum_Xk}\end{equation}
3975: In other words, we can ignore the first summand in (\ref{eq:X_sum_Xk}).
3976: Let us therefore define the number of quasi-loops up to the $i$th
3977: time $\mathcal{X}_{i}:=\mathcal{X}(T_{i})$. The process of loop-erasing
3978: between $T_{2i}$ and $T_{2i+1}$ can only destroy quasi loops in
3979: $B(w,\sigma)$ so we get \[
3980: \mathcal{X}\cdot\mathbf{1}_{\{ I=2i+1\}}\leq\mathcal{X}_{2i}\cdot\mathbf{1}_{\{ I=2i+1\}}\leq\mathcal{X}_{2i}\cdot\mathbf{1}_{\{ I>2i\}}.\]
3981: With this in mind we define \[
3982: \Delta_{i}:=(\mathcal{X}_{2i+2}-\mathcal{X}_{2i})\cdot\mathbf{1}_{\{ I>2i+2\}}.\]
3983:
3984:
3985: \begin{lem}
3986: \label{lem:Deltai}With the notations above and \begin{equation}
3987: \eta:=\min\delta({\textstyle \frac{1}{2}}\epsilon,4,G),1\label{eq:defdelf}\end{equation}
3988: where $\delta(\cdot)$ is given by (\ref{eq:inout}),\[
3989: \mathbb{E}\Delta_{i}\leq Cir^{-\eta}\log^{2}r.\]
3990:
3991: \end{lem}
3992: \begin{proof}
3993: Examine some $y$,\[
3994: y\in\mathcal{QL}(\rho,\tau,\LE(R[0,T_{2i+2}]))\setminus\mathcal{QL}(\rho,\tau,\LE(R[0,T_{2i}])).\]
3995: Directly from the definitions, we must have
3996: \begin{enumerate}
3997: \item $y$ is $\rho$-near at least one component $\gamma$ of $\LE(R[0,T_{2i}])\cap B(w,4\sigma)$.
3998: \item $R[T_{2i},T_{2i+1}]$ gets $\rho$-near $y$ and then fails to intersect
3999: at least one of the segments $\gamma$ from (i), as well as $\partial\mathcal{D}\cap B(w,4\sigma)$.
4000: \end{enumerate}
4001: In other words, the number of such $y$'s in $\frac{1}{3}\rho\mathbb{Z}^{3}\cap B(w,\sigma)$
4002: with respect to a specific $\gamma$, can be bounded by $\Delta(\gamma)$,
4003: $\Delta$ from lemma \ref{lem:onegam}, with the parameters in the
4004: following table:
4005:
4006: \noindent \begin{center}\begin{tabular}{|c|c|c|c|c|c|c|}
4007: \hline
4008: Lemma \ref{lem:onegam}&
4009: $\rho$&
4010: $\sigma$&
4011: $\tau$&
4012: $\epsilon$&
4013: $\delta$&
4014: $b$\tabularnewline
4015: \hline
4016: here&
4017: $\rho$&
4018: $2\sigma$&
4019: $4\sigma$&
4020: $\frac{1}{2}\epsilon$&
4021: $\eta$&
4022: $\partial\mathcal{D}\cap B(w,4\sigma)$\tabularnewline
4023: \hline
4024: \end{tabular}\end{center}
4025:
4026: \noindent Note that (\ref{eq:inout}) shows that with probability
4027: $\geq1-Cr^{-4}$ we have $\gamma\in\mathcal{H}=\mathcal{H}(r^{-\eta},\linebreak[1]r^{\epsilon/2})$
4028: for all $\gamma\subset\LE(R[0,T_{2i+1}])$.
4029:
4030: With this in mind we denote by $\Gamma_{i}$ the collection of connected
4031: components $\gamma$ of $\LE(R[0,T_{2i+1}])\cap B(w,4\sigma)$ satisfying
4032: $\gamma\cap B(w,\sigma+\rho)\neq\emptyset$ and get \begin{align*}
4033: \mathbb{E}\Delta_{i}\cdot\mathbf{1}_{\{\Gamma_{i}\subset\mathcal{H}\}} & \leq\max_{x\in\partial B(w,2\sigma)}\sum_{\gamma\in\Gamma_{i}}\mathbb{E}^{x}\Delta(\gamma)\cdot\mathbf{1}_{\{ R\cap\partial\mathcal{D}=\emptyset\}}\stackrel{(\ref{eq:Delta1})}{\leq}C(\#\Gamma_{i})r^{-\eta}\log^{2}r\\
4034: \mathbb{E}\Delta_{i}\cdot\mathbf{1}_{\{\Gamma_{i}\not\subset\mathcal{H}\}} & \leq C(\sigma/\rho)^{3}\mathbb{P}(\Gamma_{i}\not\subset\mathcal{H})\leq Cr^{3}\cdot r^{-4}.\end{align*}
4035: It easy to see that $\#\Gamma_{i}\leq i$, and the lemma is finished.
4036: \end{proof}
4037: Summing lemma \ref{lem:Deltai} up to $i$ we get\[
4038: \mathbb{E}\mathcal{X}_{2i}\cdot\mathbf{1}_{\{ I>2i\}}\leq Cr^{-\eta}(\log r)^{2}i^{2}\]
4039: ($\eta$ being defined by (\ref{eq:defdelf})). Another summation,
4040: up to $M$, will give us\begin{align}
4041: \mathbb{E}\mathcal{X} & \leq Cr^{-\eta}\log^{5}r+\mathbb{E}\Delta'\label{eq:EX_not_boundary}\\
4042: \Delta' & :=(\mathcal{X}_{I}-\mathcal{X}_{I-2})\cdot\mathbf{1}_{\{ I\textrm{ even}\}}.\nonumber \end{align}
4043: Thus we are left with the estimate of $\mathbb{E}\Delta'$, which
4044: is the behavior near the boundary --- if $\partial\mathcal{D}\cap B(w,4\sigma)=\emptyset$
4045: then of course $I$ is always odd and we get $\Delta'\equiv0$. It
4046: is at this point that we utilize the difference between $\tau$ and
4047: $\sigma$.
4048:
4049: \begin{lem}
4050: \label{lem:Delpharm}Let $\omega=\omega(v,\mathcal{D})$ be the discrete
4051: harmonic measure from $v$ i.e.~$\omega(A):=\mathbb{P}^{v}(R(T(\partial\mathcal{D}))\in A)$
4052: for all $A\subset\partial\mathcal{D}$. Then\[
4053: \mathbb{E}^{v}(\Delta')\leq C\frac{\sigma}{\tau}\omega(B(w,16\sigma))\log^{2}r.\]
4054:
4055: \end{lem}
4056: \begin{proof}
4057: Define stopping times $U_{0}\leq U_{1}\leq\dotsb$ using $U_{0}=0$
4058: and\begin{align*}
4059: U_{2j+1} & :=\min\{ t\geq U_{2j}:R(t)\in\partial B(w,8\sigma)\cup\partial\mathcal{D}\}\\
4060: U_{2j} & :=\min\{ t\geq U_{2j-1}:R(t)\in\partial B(w,{\textstyle \frac{1}{2}}\tau)\cup\partial\mathcal{D}\}.\end{align*}
4061: We assume here that $\sigma<\frac{1}{32}\tau$ which will hold if
4062: $r$ is sufficiently large. Define $J$ to be the first $j$ such
4063: that $R(U_{j})\in\partial\mathcal{D}$. Our first target is to connect
4064: $J$ and $\omega(B(16,\sigma))$. Denote\[
4065: p(x):=\mathbb{P}^{x}(T(\partial\mathcal{D}\cap B(w,4\sigma))<T_{w,\tau/2})\quad x\in\partial B(w,8\sigma).\]
4066: We have\begin{align}
4067: \omega(B(w,16\sigma)) & =\sum_{j=1}^{\infty}\mathbb{P}(\{ J=j\}\cap\{ R(U_{j})\in\partial\mathcal{D}\cap B(w,16\sigma)\})\geq\nonumber \\
4068: & \geq\mathbb{P}(\{ J=2\}\cap\{ R(U_{2})\in\partial\mathcal{D}\cap B(w,16\sigma)\})\geq\nonumber \\
4069: & \geq\mathbb{P}(J>1)\mathbb{E}\mathbb{P}^{R(U_{1})}(T(\partial\mathcal{D})<T_{w,16\sigma})\geq\nonumber \\
4070: & \geq\mathbb{P}(J>1)\min_{x\in\partial B(w,8\sigma)}\mathbb{P}^{x}(T(\partial\mathcal{D}\cap B(w,4\sigma))<T_{w,16\sigma})\geq\nonumber \\
4071: & \!\stackrel{(*)}{\geq}\! c(G)\mathbb{P}(J>1)\min p(x).\label{eq:harmBws4sig}\end{align}
4072: where $(*)$ comes from using lemma \ref{lem:hitbsame} to change
4073: from $T_{w,16\sigma}$ to $T_{w,\tau/2}$. $\min$ here and later
4074: on also $\max$ always refers to the minimum over $\partial B(w,8\sigma)$.
4075:
4076: The estimate of $\Delta'$ follows from the clause (\ref{eq:Delta2})
4077: of lemma \ref{lem:onegam}, which we now use with the parameters as
4078: follows:
4079:
4080: \noindent \begin{center}\begin{tabular}{|c|c|c|c|c|c|c|}
4081: \hline
4082: Lemma \ref{lem:onegam}&
4083: $\rho$&
4084: $\sigma$&
4085: $\tau$&
4086: $\epsilon$&
4087: $\delta$&
4088: $b$\tabularnewline
4089: \hline
4090: here&
4091: $\rho$&
4092: 8$\sigma$&
4093: $\frac{1}{2}\tau$&
4094: $\frac{1}{2}\epsilon$&
4095: $\eta$&
4096: $\partial\mathcal{D}\cap B(w,4\sigma)$\tabularnewline
4097: \hline
4098: \end{tabular}\end{center}
4099:
4100: \noindent And an argumentation identical to that of the previous lemma
4101: gives that, \[
4102: \mathbb{E}(\Delta'\,|\, J>2j-1)\leq C(G)(\#\Gamma_{j}')\max p(x)\log^{2}r\]
4103: where $\Gamma_{j}'$ is the collection of connected components $\gamma$
4104: of $\LE(R[0,U_{2j-1}])\cap B(w,\frac{1}{2}\tau)$ satisfying $\gamma\cap B(w,\sigma+\rho)\neq\emptyset$.
4105: Again, it is easy to see that $\#\Gamma_{j}'\leq j-1$ and in particular
4106: $\#\Gamma_{1}'=0$. Hence we get\begin{equation}
4107: \mathbb{E}(\Delta')=\sum_{j=1}^{\infty}\mathbb{E}\Delta'\cdot\mathbf{1}_{\{ J=2j\}}\leq C(G)\max p(x)\log^{2}r\sum_{j=2}^{\infty}(j-1)\mathbb{P}(J>2j-1)\label{eq:XXp}\end{equation}
4108: The Green's function estimate (\ref{eq:atrans}) also shows that \begin{align*}
4109: \mathbb{P}(J>2j+1\,|\, J>2j) & =\mathbb{EP}^{R(U_{2i})}(R(T(\partial B(w,8\sigma)\cup\partial\mathcal{D}))\in B(w,8\sigma))\leq\\
4110: & \leq\mathbb{EP}^{R(U_{2i})}(T_{w,8\sigma}<\infty)\leq C\frac{\sigma}{\tau}.\end{align*}
4111: Hence we have (for $r$ sufficiently large),\[
4112: \sum_{j=1}^{\infty}j\mathbb{P}(J>2j+1)\leq\mathbb{P}(J>1)\sum_{j=1}^{\infty}j\left(C\frac{\sigma}{\tau}\right)^{j}\leq C\mathbb{P}(J>1)\frac{\sigma}{\tau}\]
4113: and with (\ref{eq:harmBws4sig}), (\ref{eq:XXp}) and Harnack's inequality
4114: (lemma \ref{lem:Harnack-general}) which shows that $\max p(x)\leq C(G)\min p(x)$,
4115: we get\[
4116: \mathbb{E}(\Delta')\leq C\frac{\sigma}{\tau}\omega(B(w,16\sigma))\log^{2}r.\qedhere\]
4117:
4118: \end{proof}
4119: Lemma \ref{lem:Delpharm} and (\ref{eq:EX_not_boundary}) give\begin{equation}
4120: \mathbb{E}(\mathcal{X}_{w})\leq Cr^{-\eta}\log^{5}r+C\left(\frac{\sigma}{\tau}\right)\omega(B(w,16\sigma))\log^{2}r.\label{eq:XXfinal}\end{equation}
4121: For any point in $z\in B(v,r)\cap\frac{1}{3}\sigma\mathbb{Z}^{3}$,
4122: let $w_{z}$ be the point closest to $z$ in $G$. Clearly, if $r$
4123: is sufficiently large, the balls $B(w_{z},\sigma)$ form a cover of
4124: $\mathcal{D}$ and furthermore\[
4125: \sum_{z}\omega(B(w_{z},16\sigma))\leq C\omega(\partial\mathcal{D})=C.\]
4126: Hence we can sum (\ref{eq:XXfinal}) over $z$ and get\[
4127: \mathbb{E}\QL(\rho,\tau,\LE(R[0,T(\partial\mathcal{D})]))\leq C\left(\frac{r}{\sigma}\right)^{3}r^{-\eta}\log^{5}r+C\left(\frac{\sigma}{\tau}\right)\log^{2}r.\]
4128: Now is the time to pick $\delta$. We take $\delta=\min\frac{1}{11}\eta,\frac{1}{7}\epsilon$
4129: and define $\tau=r^{1-\delta}$ and $\sigma=r^{1-3\delta}$. For $r$
4130: sufficiently large the condition $\sigma>2r^{1-\epsilon/2}$ would
4131: be fulfilled. We get\[
4132: \mathbb{E}\QL(r^{1-\epsilon},r^{1-\delta},\LE(R[0,T(\partial\mathcal{D})]))\leq Cr^{-2\delta}\log^{5}r+Cr^{-2\delta}\log^{2}r\leq Cr^{-\delta}\]
4133: and the theorem is proved.\qed
4134:
4135:
4136: \subsection{Proof of theorem \refs{thm:QL} in two dimensions}
4137:
4138: \begin{lem}
4139: \label{lem:ann2}Let $G$ be a two dimensional Euclidean net. Let
4140: $v\in G$, let $r>C(G)$ and let $\gamma$ be a path from $\partial B(v,r)$
4141: to $\partial B(v,2r)$. Let $w\in B(v,r)$ and let $R$ be a random
4142: walk starting from $w$ and stopped on $\partial B(v,2r)$. Then\[
4143: \mathbb{P}(R\cap\gamma\neq\emptyset)>c(G).\]
4144:
4145: \end{lem}
4146: \begin{proof}
4147: It is easy to see, applying lemma \ref{lem:DD}, say three times,
4148: that there is a probability $>c(G)$ that $R$ does a loop around
4149: the annulus $B(v,\frac{3}{2}r)\setminus\overline{B(v,r)}$. Two dimensional
4150: geometry shows that in this case the linear extensions of $R$ and
4151: $\gamma$ intersect (the linear extension of a path in $G$ is a path
4152: in $\mathbb{R}^{2}$ which is composed of all points of $\gamma$
4153: connected by linear segments). Since the length of edges in $G$ is
4154: bounded, we get\[
4155: \mathbb{P}(\exists t\leq T_{v,2r}:d(R(t),\gamma\cap B(v,{\textstyle \frac{3}{2}}r))<C(G))>c(G).\]
4156: However, the first such $t$ is a stopping time, so we can consider
4157: the walk after it as a regular random walk, and of course it has a
4158: positive probability to hit $\gamma$.
4159: \end{proof}
4160: \begin{lem}
4161: \label{lem:kesten}Let $G$ be a two dimensional Euclidean net. Let
4162: $v\in G$ and let $s>r>1$. Let $\gamma$ be a path from $\partial B(v,r)$
4163: to $\partial B(v,s)$. Let $R$ be a random walk starting from $v$.
4164: Then\[
4165: \mathbb{P}(R\cap\gamma=\emptyset)\leq C(G)\left(\frac{r}{s}\right)^{c(G)}.\]
4166:
4167: \end{lem}
4168: This follows immediately from the previous lemma and the Wiener shell
4169: test.
4170:
4171: The proof of theorem \ref{thm:QL} now proceeds exactly as in the
4172: three dimensional case, with lemma \ref{lem:ann2} serving as a replacement
4173: for lemma \ref{lem:Beurling} and lemma \ref{lem:kesten} serving
4174: as a replacement to lemmas \ref{lem:outgoing} and \ref{lem:incoming}.
4175: The only complication is lemma \ref{lem:Delpharm}, which no longer
4176: holds as stated. It is necessary at this point to estimate $\mathbb{P}(J>2i+1\,|\, J>2i)$
4177: by lemma \ref{lem:kesten} for the incoming walk, with the result
4178: that one gets only $(\sigma/\tau)^{c}$ in the formulation of the
4179: lemma. See \cite{K}, sublemmas 18.2 \& 18.3 for details.
4180:
4181:
4182: \subsection{Continuity in the starting point}
4183:
4184: We will need one additional corollary of the techniques of this chapter:
4185:
4186: \begin{lem}
4187: \label{lem:contsp}Let $G$ be an isotropic graph and let $v,w\in\mathcal{E}\subset\mathcal{D}\subset G$,
4188: $\mathcal{D}$ finite. Denote\[
4189: p^{x}:=\mathbb{P}^{x}(\LE(R[0,T(\partial\mathcal{D})])\subset\mathcal{E}),\quad x=v,w.\]
4190: Then \[
4191: |p^{v}-p^{w}|\leq C(G)\left(\frac{|v-w|}{d(v,\partial\mathcal{E})}\right)^{c(G)}.\]
4192:
4193: \end{lem}
4194: \begin{proof}
4195: Denote $\mu=|v-w|/d(v,\partial\mathcal{E})$. We may assume w.l.o.g
4196: that $\mu$ is sufficiently small. Let $H$ be the graph generated
4197: by taking $\overline{\mathcal{D}}$ and identifying all the points
4198: of $\partial\mathcal{D}$ (this process is often called {}``wiring
4199: the boundary''). Let $T$ be the uniform spanning tree on $H$. Then
4200: by Pemantle \cite{P91}, the distribution of the path in $T$ from
4201: $x$ to $\partial\mathcal{D}$ is identical to the distribution of
4202: a loop-erased random walk on \emph{}$H$ starting from $x$ and stopped
4203: on $\partial\mathcal{D}$, which is identical to the distribution
4204: on the graph $G$. Hence the two branches $\beta^{v}$ and $\beta^{w}$
4205: of $T$ from $v$ and $w$ to $\partial\mathcal{D}$ is a coupling
4206: of the two loop-erased walks. Denote by $\gamma^{x}$ the portion
4207: of $\beta^{x}$ from $x$ until the unique intersection of $\beta^{v}$
4208: and $\beta^{w}$; and by $p$ the probability that $\gamma^{v}\cup\gamma^{w}\subset\mathcal{E}$.
4209: Then clearly, \[
4210: |p^{v}-p^{w}|\leq1-p.\]
4211: Now, by Wilson's algorithm \cite{W96}, $\gamma^{w}$ may be constructed
4212: by first constructing $\beta^{v}$ and then taking a random walk starting
4213: from $w$, stopping it when it first hits $\beta^{v}$ (possibly at
4214: time $0$) and performing loop-erasure on the result. Hence we need
4215: to show that\[
4216: \mathbb{P}^{w}(T(\beta^{v})>T(\partial\mathcal{E}))\leq C(G)\mu^{c(G)}.\]
4217: In three dimensions we use lemma \ref{lem:wienersimp}. Let $K=\frac{1}{2}$
4218: and let $\delta$ be the $\delta(\frac{1}{2},G)$ of lemma \ref{lem:wienersimp}.
4219: Let the $r$ and $s$ of lemma \ref{lem:wienersimp} be $d(w,\partial\mathcal{E})$
4220: and $\max|v-w|,1$ respectively --- if $\mu$ is sufficiently small
4221: we would get $r>s$. We get that (except for probability $C\mu^{1/2}$
4222: in the walk starting from $v$)\[
4223: \mathbb{P}^{2,w}(R^{2,w}[0,T_{w,r}^{2}]\cap\LE(R^{1}[0,T^{1}(\partial\mathcal{D})])\cap B(w,r)\setminus\overline{B(w,s)}=\emptyset)\leq C\mu^{\delta}.\]
4224: Therefore $\mathbb{P}(\gamma^{w}\not\subset\mathcal{E})\leq C\mu^{\delta}+C\mu^{1/2}$.
4225: The same holds for $\mathbb{P}(\gamma^{v}\not\subset\mathcal{E})$
4226: and the three dimensional case is finished. The two dimensional case
4227: follows similarly from lemma \ref{lem:kesten}.
4228: \end{proof}
4229: \begin{rem*}
4230: The use of lemma \ref{lem:wienersimp} to estimate the probability
4231: that a loop-erased random walk and a random walk starting from close
4232: points will hit is somewhat an overkill. For example, in $\mathbb{Z}^{3}$,
4233: if they start from the same point then the nice symmetry argument
4234: of \cite{L99} can show that this probability is $>1-Cr^{-1/3}$.
4235: Presumably, an equivalent argument would work in our case as well.
4236: The arguments of \cite{AB99} should also give a usable estimate.
4237: \end{rem*}
4238:
4239: \section{\label{sec:Isotropic-gluing}Isotropic interpolation}
4240:
4241: The purpose of this chapter is to compare random walks on two or more
4242: graphs all of which are isotropic, with uniformly bounded structure
4243: constants. We shall call such a collection $\mathcal{G}$ an \textbf{isotropic
4244: family} and denote by $C(\mathcal{G})$, $c(\mathcal{G})$ etc.~constants
4245: which depend only on the maximum of the isotropic structure constants
4246: of all $G\in\mathcal{G}$.
4247:
4248:
4249: \subsection{Hitting probabilities}
4250:
4251: In this section we will compare probabilities by proving inequalities
4252: of the sort $|p-q|\leq\epsilon\max p,q$. It will be convenient to
4253: denote this by\[
4254: p\stackrel{\epsilon}{\simeq}q.\]
4255: When $\epsilon=Cr^{-c}$ for some constants $C(\mathcal{G})$ and
4256: $c(\mathcal{G})$ we will usually omit it, and just write $p\simeq q$
4257: ($r$ will be clear from the context). Occasionally we will prove
4258: instead that $|p-q|\leq\epsilon p$ or $|p-q|\leq\epsilon\min p,q$.
4259: We will always assume $\epsilon\leq\frac{1}{2}$, and then they are
4260: all equivalent up to constants.
4261:
4262: We will often use the following version of differentiation of product:
4263: assume\[
4264: p^{1}\stackrel{\alpha}{\simeq}p^{2}\quad q^{1}\stackrel{\beta}{\simeq}q^{2}\quad\alpha,\beta\leq\frac{1}{2}.\]
4265: Then\begin{align}
4266: |p^{1}q^{1}-p^{2}q^{2}| & \leq|p^{1}-p^{2}|q^{1}+p^{2}|q^{1}-q^{2}|\leq\alpha q^{1}\max p^{1},p^{2}+\beta p^{2}\max q^{1},q^{2}\nonumber \\
4267: & \leq\alpha q^{1}(1+2\alpha)p^{1}+\beta p^{2}(1+2\beta)q^{2}\leq\nonumber \\
4268: & \leq(\alpha+\beta+2(\alpha^{2}+\beta^{2}))\max p^{1}q^{1},p^{2}q^{2}\leq\label{eq:ddm}\\
4269: & \leq2(\alpha+\beta)\max p^{1}q^{1},p^{2}q^{2}\label{eq:ddm2}\end{align}
4270: Another useful fact: if $p^{i}=\sum_{n}q_{n}^{i}$ then\begin{align}
4271: |p^{1}-p^{2}| & \leq\sum_{n}|q_{n}^{1}-q_{n}^{2}|\leq\sum_{n}\beta\max q_{n}^{1},q_{n}^{2}\leq\sum_{n}\beta(q_{n}^{1}+q_{n}^{2})=\beta(p^{1}+p^{2})\nonumber \\
4272: & \leq2\beta\max p^{1},p^{2}.\label{eq:dds}\end{align}
4273:
4274:
4275: \begin{lem}
4276: \label{lem:nodir}Let $G$ be an isotropic graph and let $r_{1}>s>r_{2}>1$.
4277: Let $v\in G$, $v^{1},v^{2}\in\partial B(v,s)$ and let $w\in A:=\partial B(v,r_{1})\cup\partial B(v,r_{2})$.
4278: Let \[
4279: p^{i}=\mathbb{P}^{v^{i}}(R(T(A))=w).\]
4280: Then \begin{equation}
4281: \left|p^{1}-p^{2}\right|\leq C(G)\left(\max\frac{s}{r_{1}},\frac{r_{2}}{s}\right)^{c(G)}\max p_{1},p_{2}.\label{eq:lemnodir}\end{equation}
4282: Similarly if $A:=\partial B(v,r_{1})\cup\{ v\}$ then $|p_{1}-p_{2}|\leq C\left(\max\frac{s}{r_{1}},\frac{1}{s}\right)^{c(G)}\max p_{1},p_{2}$.
4283: \end{lem}
4284: \begin{proof}
4285: Assume first that $r_{2}\geq r_{1}^{1-c_{\ref{c:RWclose}}/2}$ where
4286: $c_{\ref{c:RWclose}}(G)$ is from lemma \ref{lem:rismall}, and assume
4287: also that $r_{1}=2^{N}s=2^{2N}r_{2}$ for some integer $N$ (removing
4288: both assumptions is easy, and we do it in the end of the lemma). Assume
4289: also that $w\in\partial B(v,r_{2})$. Define \begin{align*}
4290: a_{j} & :=s2^{-j}\quad j=0,1,\dotsc,N.\\
4291: X_{j} & :=\partial B(v,a_{j})\cup\partial B(v,r_{1})\end{align*}
4292: Define stopping times $T_{0}=0$ and \[
4293: T_{j}:=\min\{ t\geq T_{j-1}:R(t)\in X_{j}\}.\]
4294: Define probability measures $\mu_{j}^{i}$ on $\partial B(v,a_{j})$
4295: by \[
4296: \mu_{j}^{i}(x)=\mathbb{P}^{v^{i}}(R(T_{j})=x\,|\, R(T_{j})\in\partial B(v,a_{j})).\]
4297:
4298:
4299: We proceed by examining how $\mu_{j}^{i}$ evolves with $j$. It will
4300: be useful to use $L^{1}$ estimates during intermediate stages, so
4301: define \[
4302: \epsilon_{j}:=\sum_{x\in\partial B(v,a_{j})}|\mu_{j}^{1}(x)-\mu_{j}^{2}(x)|.\]
4303: Clearly $\epsilon_{0}=2$. For $x\in\partial B(v,a_{j})$ and $y\in\partial B(v,a_{j+1})$
4304: we define \[
4305: \pi(x,y)=\mathbb{P}^{x}(R(T(X_{j+1}))=y\,|\, R(T(X_{j+1}))\in\partial B(0,a_{j+1}))\]
4306: and then get\[
4307: \mu_{j+1}^{i}(y)=\sum_{x\in\partial B(0,a_{j})}\mu_{j}^{i}(x)\pi(x,y).\]
4308:
4309:
4310: Define now \begin{gather*}
4311: A^{+}:=\{ x\in\partial B(0,a_{j}):\mu_{j}^{1}(x)\geq\mu_{j}^{2}(x)\},\quad A^{-}:=\partial B(0,a_{j})\setminus A^{+},\\
4312: D^{\pm}(y):=\sum_{x\in A^{\pm}}|\mu_{j}^{1}(x)-\mu_{j}^{2}(x)|\pi(x,y).\end{gather*}
4313: Clearly, \begin{equation}
4314: \sum_{y}D^{\pm}(y)=\sum_{x\in A^{\pm}}|\mu_{j}^{1}(x)-\mu_{j}^{2}(x)|\sum_{y}\pi(x,y)=\sum_{x\in A^{\pm}}|\mu_{j}^{1}(x)-\mu_{j}^{2}(x)|={\textstyle \frac{1}{2}}\epsilon_{j}.\label{eq:sumDpmeps}\end{equation}
4315: Next, Harnack's inequality (lemma \ref{lem:Harnack-general}) shows
4316: that $\pi(x,y)\approx\pi(x',y)$ for any $x,x'\in\partial B(0,a_{j})$
4317: so $D^{+}(y)\approx D^{-}(y)$ ($\pi(x,y)$ is a quotient of two harmonic
4318: functions and we use Harnack's inequality for both). This gives that\[
4319: |D^{+}(y)-D^{-}(y)|\leq(1-c(G))(D^{+}(y)+D^{-}(y))\]
4320: and hence\[
4321: \epsilon_{j+1}=\sum_{y\in\partial B(v,a_{j+1})}|D^{+}(y)-D^{-}(y)|\leq(1-c)\sum_{y}D^{+}(y)+D^{-}(y)\stackrel{(\ref{eq:sumDpmeps})}{=}(1-c)\epsilon_{j}.\]
4322: Therefore we get $\epsilon_{N-1}\leq2(1-c)^{N-1}=C(s/r_{1})^{c}$.
4323: This establishes the $L^{1}$ estimate. We now use the last step (from
4324: $N-1$ to $N$) to move to a uniform estimate. Return to our $w\in\partial B(v,r_{2})$.
4325: We have \[
4326: \mu_{N}^{i}(w)=\sum_{x\in\partial B(v,a_{N-1})}\mu_{N-1}^{i}(x)\pi(x,w)\geq\min_{x}\pi(x,w)\]
4327: and on the other hand\begin{align*}
4328: |\mu_{N}^{1}(w)-\mu_{N}^{2}(w)| & \leq\sum_{x\in\partial B(v,a_{N-1})}|\mu_{N-1}^{1}(x)-\mu_{N-1}^{2}(x)|\pi(x,w)\leq\\
4329: & \leq C(s/r_{1})^{c}\max_{x}\pi(x,w)\end{align*}
4330: and by Harnack's inequality (lemma \ref{lem:Harnack-general}),\begin{equation}
4331: |\mu_{N}^{1}(w)-\mu_{N}^{2}(w)|\leq C(s/r_{1})^{c}\max\mu_{N}^{1}(w),\mu_{N}^{2}(w).\label{eq:muxmuy}\end{equation}
4332:
4333:
4334: The difference between $\mu_{N}^{i}(w)$ and $p^{i}$ is just that
4335: $\mu_{N}^{i}(w)$ is conditioned, i.e.\[
4336: p^{i}=\mu_{N}^{i}(w)\mathbb{P}^{v^{i}}(R(T_{N})\in\partial B(v,r_{2})).\]
4337: To estimate the second term we use lemma \ref{lem:sball}. The probability
4338: $q^{i}$ that Brownian motion starting from $v^{i}$ hits $\partial B(v,r_{2})$
4339: before $\partial B(v,r_{1})$ is \begin{equation}
4340: q^{i}=\frac{a(\left|v^{i}\right|)-a(r_{1})}{a(r_{2})-a(r_{1})}\quad a(t)=\begin{cases}
4341: t^{2-d} & d\geq3\\
4342: \log t & d=2\end{cases}\label{eq:qiprecis}\end{equation}
4343: and in either case we get $|q^{1}-q^{2}|\leq C(G)s^{-1}\max q^{1},q^{2}$.
4344: Therefore lemma \ref{lem:sball} (take e.g.\ the $\epsilon$ of lemma
4345: \ref{lem:sball} to be $\frac{1}{2}$) gives that \begin{equation}
4346: \mathbb{P}^{v^{1}}(R(T_{N})\in\partial B(0,r_{2}))\simeq\mathbb{P}^{v^{2}}(R(T_{N})\in\partial B(0,r_{2})).\label{eq:Pvir2same}\end{equation}
4347: Together with (\ref{eq:muxmuy}) the lemma is proved in this case.
4348:
4349: The case that $w\in\partial B(v,r_{1})$ is identical, with this time
4350: defining $a_{j}=s2^{j}$ and $X_{j}=\partial B(v,a_{j})\cup\partial B(v,r_{2})$.
4351: The argument about the exponential decrease of the $\epsilon_{j}$
4352: works identically. Finally, (\ref{eq:qiprecis}) shows that $\mathbb{P}^{v^{i}}(R(T_{N})\in\partial B(v,r_{1}))>c$
4353: and therefore (\ref{eq:Pvir2same}) implies an identical estimate
4354: for $\mathbb{P}^{v^{i}}(R(T_{N})\in\partial B(v,r_{1}))$. Hence this
4355: case is finished too.
4356:
4357: Finally, assume one of the assumptions on $r_{1}$, $s$ and $r_{2}$
4358: fails. If $r_{1}/s$ or $s/r_{2}<2$ the lemma holds trivially for
4359: sufficiently large constants. Hence assume that both are $\geq2$.
4360: It now follows easily that one can find $u_{i}$ such that $r_{1}>u_{1}>s>u_{2}>r_{2}$,
4361: $\frac{u_{1}}{s}=\frac{s}{u_{2}}=2^{N}$ and $u_{2}\geq u_{1}^{1-c_{\ref{c:RWclose}}/2}$,
4362: and further, \[
4363: \frac{s}{u_{1}}\leq\left(\max\frac{s}{r_{1}},\frac{r_{2}}{s}\right)^{c(G)}.\]
4364: We use the case already established and find that for any $x\in\partial B(v,u_{1})\cup\partial B(v,u_{2})$
4365: one has\[
4366: |p^{1}(x)-p^{2}(x)|\leq C(G)\left(\frac{s}{u_{1}}\right)^{c}\max p^{1}(x),p^{2}(x)\]
4367: where we define $p^{i}(x):=\mathbb{P}^{v^{i}}(R(T(\partial B(v,u_{1})\cup\partial B(v,u_{2})))=x)$.
4368: Define \[
4369: \pi(x)=\mathbb{P}^{x}(R(T(\partial B(v,r_{1})\cup\partial B(v,r_{2})))=w)\]
4370: and get\[
4371: p^{i}=\sum_{x}p^{i}(x)\pi(x).\]
4372: Therefore\begin{align*}
4373: |p^{1}-p^{2}| & \leq\sum_{x}|p^{1}(x)-p^{2}(x)|\pi(x)\leq C(G)\left(\frac{s}{u_{1}}\right)^{c}\sum_{x}(p^{1}(x)+p^{2}(x))\pi(x)=\\
4374: & =C(G)\left(\frac{s}{u_{1}}\right)^{c}(p^{1}+p^{2})\leq C\left(\max\frac{s}{r_{1}},\frac{r_{2}}{s}\right)^{c}\max p^{1},p^{2}.\end{align*}
4375: The case of a single point is proved identically.
4376: \end{proof}
4377: \begin{lem}
4378: \label{lem:nodir2}Let $G,r_{1},s,v,v^{1},v^{2}$ be as in lemma \ref{lem:nodir},
4379: and let $w\in\partial B(v,r_{1})$. Let $p^{i}=\mathbb{P}^{v^{i}}(R(T_{v,r_{1}})=w)$.
4380: Then \[
4381: |p^{1}-p^{2}|\leq C(G)\left(\frac{s}{r_{1}}\right)^{c(G)}\max p^{1},p^{2}.\]
4382:
4383: \end{lem}
4384: The proof is a simplified version of the proof of lemma \ref{lem:nodir}
4385: and we shall omit it.
4386:
4387: \begin{cor*}
4388: The conclusion of lemma \ref{lem:nodir2} holds if $v^{1},v^{2}\in B(0,s)$
4389: (and not on its boundary).
4390: \end{cor*}
4391: \begin{proof}
4392: Apply lemma \ref{lem:nodir2} after the stopping time on $\partial B(0,s)$.
4393: \end{proof}
4394: \begin{lem}
4395: \label{lem:ifhity}Let $\mathcal{G}=(G^{1},G^{2})$ be an isotropic
4396: family, let $v\in\mathbb{R}^{d}$ and assume that for some $r>C(\mathcal{G})$
4397: one has \[
4398: G^{1}\cap(\overline{B(v,4r)}\setminus B(v,r))=G^{2}\cap(\overline{B(v,4r)}\setminus B(v,r)).\]
4399: Let $w\in\overline{B(v,\frac{5}{2}r)}\setminus B(v,\frac{3}{2}r)$
4400: and $x\in\overline{B(v,3r)}\setminus B(v,r)$ and let $R^{i}$ be
4401: random walks on $G^{i}$ starting from $x$ (which is contained in
4402: both $G^{i}$) and stopped on $\partial B(v,4r)$. Let \[
4403: p^{i}=\mathbb{P}(w\in R^{i}[1,T_{v,4r}^{i}]).\]
4404: Then $|p^{1}-p^{2}|\leq C(\mathcal{G})r^{-c(\mathcal{G})}\min\{ p^{1},p^{2}\}$.
4405: \end{lem}
4406: Here $G\cap(\overline{B(v,4r)}\setminus B(v,r))$ is the subgraph
4407: of $G$ containing all vertices of $(\overline{B(v,4r)}\setminus B(v,r))$
4408: and all edges between them; and the $=$ sign refers to equality of
4409: graphs.
4410:
4411: Notice that in the definition of $p^{i}$ we consider that $w$ was
4412: hit only starting from the first step. Hence the lemma gives a non-trivial
4413: estimate even if $x=w$ (a case that is actually important).
4414:
4415: \begin{proof}
4416: Let $X=\partial B(v,4r)\cup\{ w\}$. Let $s=r^{1-c_{\ref{c:RWclose}}(\mathcal{G})/2}$
4417: where $c_{\ref{c:RWclose}}(\mathcal{G})=\min c_{\ref{c:RWclose}}(G_{1}),\linebreak[1]c_{\ref{c:RWclose}}(G_{2})$
4418: and $c_{\ref{c:RWclose}}(G)$ is from lemma \ref{lem:rismall}. Define
4419: stopping times as follows: $T_{0}^{i}=0$ and \begin{align*}
4420: T_{2n+1}^{i} & :=\min\{ t\geq T_{2n}^{i}:R^{i}(t)\in\partial B(w,s)\cup X\}\\
4421: T_{2n}^{i} & :=\min\{ t\geq T_{2n-1}^{i}:R^{i}(t)\in\partial B(w,{\textstyle \frac{1}{4}}r)\cup X\}.\end{align*}
4422: Let $p_{n}^{i}=\mathbb{P}(R^{i}(T_{n}^{i})\not\in X).$ The core of
4423: the lemma is showing that\begin{equation}
4424: |p_{n}^{1}-p_{n}^{2}|\leq Knr^{-k}\max p_{n}^{1},p_{n}^{2}\label{eq:pn1pn2}\end{equation}
4425: for any $n$ such that $Kn^{2}r^{-k}\leq\frac{1}{4}$. Here $K=K(\mathcal{G})$
4426: and $k=k(\mathcal{G})$ are some sufficient constants (by which we
4427: mean that $K$ is sufficiently large and $k>0$ sufficiently small)
4428: which will be fixed later.
4429:
4430: The proof of (\ref{eq:pn1pn2}) will be done by induction over $n$.
4431: For $n=1$, (\ref{eq:pn1pn2}) follows from lemma \ref{lem:sball}
4432: with $\epsilon=\frac{1}{8}$, if only $r$, $K$ and $k$ are sufficient.
4433: Next, if $n\geq1$, define\[
4434: \mu_{2n}^{i}(y):=\mathbb{P}(R^{i}(T_{2n}^{i})=y\,|\, T_{2n-1}^{i}\not\in X)\quad y\in\partial B(w,{\textstyle \frac{1}{4}}r)\cup\{ w\}\]
4435: and then lemma \ref{lem:nodir} gives that the portion of the walk
4436: between $T_{2n-1}$ and $T_{2n}$, which is the same since $G^{1}\cap B(v,\frac{1}{4}r)=G^{2}\cap B(v,\frac{1}{4}r)$,
4437: erases most of the difference between the $\mu^{i}$-s and we have\begin{equation}
4438: \mu_{2n}^{1}(y)\simeq\mu_{2n}^{2}(y).\label{eq:mu2n1mu2n2}\end{equation}
4439: Denote $\alpha^{i}=\sum_{y\neq w}\mu_{2n}^{i}(y)$ so $\alpha^{1}\simeq\alpha^{2}$.
4440: Now, $p_{2n}^{i}=p_{2n-1}^{i}\alpha^{i}$ so we can use (\ref{eq:ddm}),
4441: note that $Cr^{-c}<\frac{1}{2}$ for $r$ sufficiently large and $K(2n-1)r^{-k}\leq\frac{2n-1}{4(2n)^{2}}<\frac{1}{8n}<\frac{1}{2}$,
4442: and we get \begin{align}
4443: \lefteqn{|p_{2n}^{1}-p_{2n}^{2}|\leq}\nonumber \\
4444: & \:\:\:\stackrel{(\ref{eq:ddm})}{\leq}\left(K(2n-1)r^{-k}+Cr^{-c}+2\left(\left(K(2n-1)r^{-k}\right)^{2}+\left(Cr^{-c}\right)^{2}\right)\right)\max p_{2n}^{1},p_{2n}^{2}\nonumber \\
4445: & \quad\leq\left(K\left(2n-{\textstyle \frac{1}{2}}\right)r^{-k}+Cr^{-c}\right)\max p_{2n}^{1},p_{2n}^{2}\stackrel{(*)}{\leq}K(2n)r^{-k}\max p_{2n}^{1},p_{2n}^{2}\label{eq:palpha1}\end{align}
4446: where $(*)$ holds if only $K$ and $k$ are sufficient. Hence the
4447: induction holds when moving from $2n-1$ to $2n$.
4448:
4449: For the case of moving from $2n$ to $2n+1$, define, for any $y\in\partial B(w,\frac{1}{4}r$),
4450: \[
4451: \pi^{i}(y)=\mathbb{P}^{y}(R^{i}(T^{i}(\partial B(w,s)\cup X))\in\partial B(w,s)).\]
4452: Then \[
4453: p_{2n+1}^{i}=p_{2n}^{i}\sum_{y}\mu_{2n}^{i}(y)\pi^{i}(y).\]
4454: As in the previous part, we define $\alpha^{i}:=\sum\mu_{2n}^{i}(y)\pi^{i}(y)$.
4455: Lemma \ref{lem:sball} shows that $\pi^{1}(y)\simeq\pi^{2}(y)$. Together
4456: with (\ref{eq:mu2n1mu2n2}) we get, \[
4457: |\alpha^{1}-\alpha^{2}|\stackrel{(\ref{eq:ddm2})}{\leq}\sum_{y}Cr^{-c}\max\mu_{2n}^{1}(y)\pi^{1}(y),\mu_{2n}^{2}(y)\pi(y)\stackrel{(\ref{eq:dds})}{\leq}Cr^{-c}\max\alpha^{1},\alpha^{2}\]
4458: for $r$ sufficiently large. Thus we can repeat the calculations of
4459: (\ref{eq:palpha1}) and get again that (\ref{eq:pn1pn2}) is preserved
4460: if only $K$ and $k$ are sufficient. Hence the proof of (\ref{eq:pn1pn2})
4461: is completed and we may fix the values of $K$ and $k$.
4462:
4463: Next we need to ask how many $n$-s are actually relevant. Lemma \ref{lem:DD}
4464: shows that \[
4465: p_{2n+1}^{i}\leq(1-c(\mathcal{G}))p_{2n}^{i}\quad\forall n\geq1.\]
4466: Hence the $p^{i}$ decrease exponentially. On the other hand, the
4467: Green's function estimates (\ref{eq:arecur},\ref{eq:atrans}) shows
4468: that \[
4469: p^{i}>c(\mathcal{G})\begin{cases}
4470: r^{2-d} & d\geq3\\
4471: 1/\log r & d=2\end{cases}.\]
4472: Define therefore $N=C(\mathcal{G})\log r$ for some $C$ sufficiently
4473: large and get\[
4474: \sum_{N+1}^{\infty}p_{n}^{i}<r^{-1}p^{i}.\]
4475: We can now calculate\[
4476: p^{i}=\mathbb{P}(R^{i}(T_{1}^{i})=w)+\sum_{n=1}^{\infty}p_{2n-1}^{i}\mu_{2n}^{i}(w).\]
4477: The first summand is non-zero only if $x\in B(w,s)$ and in this case
4478: it is independent of $i$. Hence we may write, \begin{align*}
4479: \lefteqn{|p^{1}-p^{2}|\leq}\\
4480: & \quad\stackrel{(\ref{eq:ddm2})}{\leq}\sum_{n=1}^{N}(2Knr^{-k}+Cr^{-c})\max p_{2n-1}^{1}\mu_{2n}^{1}(w),p_{2n-1}^{2}\mu_{2n}^{2}(w)+2r^{-1}\max p^{1},p^{2}\\
4481: & \quad\stackrel{(\ref{eq:dds})}{\leq}(4KNr^{-k}+Cr^{-c}+2r^{-1})\max p^{1},p^{2}\end{align*}
4482: if only $r$ is sufficiently large such that we get $KN^{2}r^{-k}<\frac{1}{4}$.
4483: This finishes the lemma.
4484: \end{proof}
4485: \begin{rem*}
4486: In three dimensions this lemma may be simplified significantly, since
4487: the probability to hit $w$ after $T_{2}$ is significantly smaller
4488: than between $0$ and $T_{2}$, so there is no need for the induction.
4489: \end{rem*}
4490: \begin{lem}
4491: \label{lem:iiexit}Let $\mathcal{G},r,v,w$ and $R^{i}$ be as in
4492: lemma \ref{lem:ifhity} (perhaps with a different constant bounding
4493: $r$ from below). Let $x\in\partial B(v,4r)$ and let \[
4494: p^{i}:=\mathbb{P}^{w}(R^{i}(T_{v,4r}^{i})=x).\]
4495: Then $p^{1}\simeq p^{2}$.
4496: \end{lem}
4497: \begin{proof}
4498: The symmetry of random walk in the form (\ref{eq:symT}) shows that
4499: \begin{equation}
4500: p^{i}=\frac{\mathbb{P}^{x}(w\in R^{i}[1,T_{v,4r}^{i}])}{1-\mathbb{P}^{w}(w\in R^{i}[1,T_{v,4r}^{i}])}\nu\label{eq:pisym}\end{equation}
4501: where the constant $\nu$ is the ratio of the degrees of $w$ and
4502: $x$ and is independent of $i$. Denote the denominator by $1-a^{i}$.
4503: Lemma \ref{lem:ifhity} shows that $a^{1}\simeq a^{2}$. The Green's
4504: function estimates (\ref{eq:arecur},\ref{eq:atrans}) show that $1-a^{i}\geq c/\log r$
4505: and therefore \begin{equation}
4506: \left|\frac{1}{1-a^{1}}-\frac{1}{1-a^{2}}\right|\leq Cr^{-c}\log r\max\frac{1}{1-a^{1}},\frac{1}{1-a^{2}}\label{eq:1oa11oa2}\end{equation}
4507: and we can drop the $\log$ factor from (\ref{eq:1oa11oa2}) and pay
4508: in the constants only.
4509:
4510: Next denote the nominator of (\ref{eq:pisym}) by $b^{i}$. For any
4511: $y\in\partial B(v,3r)$ denote \[
4512: \pi(y):=\mathbb{P}^{x}(R^{i}(T^{i}(\partial B(v,4r)\cup\partial B(v,3r)))=y)\quad\rho^{i}(y):=\mathbb{P}^{y}(w\in R^{i}[1,T_{v,4r}^{i}])\]
4513: ($\pi$ does not depend on $i$ because $G^{1}$ and $G^{2}$ are
4514: identical on the relevant annulus). Hence we get $b^{i}=\sum_{y}\pi(y)\rho^{i}(y)$.
4515: Lemma \ref{lem:ifhity} shows that $\rho^{1}(y)\simeq\rho^{2}(y)$
4516: and therefore by (\ref{eq:dds}) we get $b^{1}\simeq b^{2}$. Together
4517: with (\ref{eq:1oa11oa2}) the lemma is proved (again we use here (\ref{eq:ddm2})).
4518: \end{proof}
4519: \begin{cor*}
4520: Let $\mathcal{G},r,v$ and $x$ be as in lemma \ref{lem:iiexit},
4521: and let $v^{1},v^{2}\in\partial B(v,2r)$. Let\[
4522: p^{i}:=\mathbb{P}^{v^{i}}(R^{i}(T_{v,4r}^{i})=x).\]
4523: Then \[
4524: |p^{1}-p^{2}|\leq C(\mathcal{G})\left(\frac{|v^{1}-v^{2}|}{r}\right)^{c(\mathcal{G})}\max p^{1},p^{2}.\]
4525:
4526: \end{cor*}
4527: \begin{proof}
4528: If $|v^{1}-v^{2}|\leq\frac{1}{8}r$ then we can use lemma \ref{lem:nodir2}
4529: to shows that the walk up to $B(v^{1},\frac{1}{4}r)$ erases the difference
4530: between $v^{1}$ and $v^{2}$, and then use lemma \ref{lem:iiexit}
4531: to show that the fact that the graphs are different has a small effect.
4532: If $|v^{1}-v^{2}|>\frac{1}{8}r$ then for a $C$ sufficiently large,
4533: there is nothing to prove.
4534: \end{proof}
4535: \begin{lem}
4536: \label{lem:iiexit2}Lemma \ref{lem:iiexit} holds also when $w\in B(v,\frac{3}{2}r)$.
4537: \end{lem}
4538: Here $w$ is some point in $\mathbb{R}^{d}$ and the notation $\mathbb{P}^{w}$
4539: refers to a random walk starting from the point of the relevant graph
4540: closest to $w$.
4541:
4542: \begin{proof}
4543: Let $\{\Delta_{j}\}$ be a triangulation of $B(v,2r)$ by spherical
4544: triangles such that\[
4545: |\Delta_{j}|\geq cr^{-c_{\ref{c:striag}}/2}\quad\diam\Delta_{j}\leq Cr^{1-c_{\ref{c:striag}}/2d}\]
4546: where $c_{\ref{c:striag}}$ is from lemma \ref{lem:striag} and $|\cdot|$
4547: is the normalized volume. Let $\Delta_{j}^{*}$ be disjoint discrete
4548: versions of the $\Delta_{j}$ covering $\partial B(v,2r)$. Let $w^{i}$
4549: be the point of $G^{i}$ closest to $w$, let $p_{j}^{i}:=\mathbb{P}^{w^{i}}(R^{i}(T_{v,2r})\in\Delta_{j}^{*})$
4550: and $q_{j}^{i}:=\mathbb{P}^{w^{i}}(W(S_{v,2r})\in\Delta_{j})$ the
4551: Brownian motion analogs. Now, $q_{j}^{i}$ has a formula given from
4552: the surface integral over the Poisson kernel \cite[II theorem 1.17]{B95}:\[
4553: q_{j}^{i}=r^{d-2}\int_{\Delta_{j}}\frac{r^{2}-\left\Vert w^{i}\right\Vert ^{2}}{\left\Vert x-w^{i}\right\Vert ^{d}}\, dx.\]
4554: which immediately shows, since $||w^{1}-w^{2}||\leq C(G)$ that $q_{j}^{1}\simeq q_{j}^{2}$.
4555: Lemma \ref{lem:striag} shows that \[
4556: |q_{j}^{i}-p_{j}^{i}|\leq Cr^{-c_{\ref{c:striag}}}\leq Cr^{-c_{\ref{c:striag}}/2}|\Delta_{j}|\leq Cr^{-c_{\ref{c:striag}}/2}q_{j}^{i}\]
4557: So also $p_{j}^{1}\simeq p_{j}^{2}$.
4558:
4559: Next, let $y^{1},y^{2}\in\Delta_{j}^{*}$. Denote $\pi^{i}(y)=\mathbb{P}^{y}(R^{i}(T_{v,4r}^{i})=x)$.
4560: Because $\diam\Delta_{j}^{*}\leq Cr^{1-c}$ we can use the corollary
4561: after lemma \ref{lem:iiexit} to get $\pi^{1}(x^{1})\simeq\pi^{2}(x^{2})$.
4562: This finishes the lemma, since the probabilities to hit a given $\Delta_{j}^{*}$
4563: are $Cr^{-c}$ similar, and the point in which you hit is unimportant
4564: up to $Cr^{-c}$ error.
4565: \end{proof}
4566:
4567: \subsection{\label{sub:Definition-II}Definition}
4568:
4569: Let $G^{1}$ and $G^{2}$ be two $d$-dimensional isotropic graphs,
4570: and let $\alpha>0$. We say that $G^{1}$ and $G^{2}$ have an $\alpha$-\textbf{isotropic
4571: interpolation} if the following holds. Let $L$ and $M$, $M\geq1$
4572: integer, satisfy \begin{equation}
4573: L>C(G^{i})\quad M\leq L^{\alpha}.\label{eq:iirMreq}\end{equation}
4574: We assume $C(G^{i})\geq2$ always. Let $\xi\in\{1,2\}^{M^{d}}$ be
4575: any configuration. Then there exist graphs $G(L,M,\xi)$ such that:
4576:
4577: \begin{enumerate}
4578: \item \label{enu:isinI}If all the coordinates of $\xi$ are $i$ then $G(L,M,\xi)\equiv G^{i}$.
4579: \item \label{enu:isinII}If $I:=[a_{1},b_{1}]\times\dotsb\times[a_{d},b_{d}]\subset\{0,\dotsc,M-1\}^{d}$
4580: is some box in the configuration space , and if $\left.\xi_{1}\right|_{I}\equiv\left.\xi_{2}\right|_{I}$
4581: then $G(L,M,\xi_{1})$ is equal to $G(L,M,\xi_{2})$ on a corresponding
4582: box in $\mathbb{R}^{d}$ i.e.\begin{gather}
4583: G(L,M,\xi_{1})\cap J=G(L,M,\xi_{2})\cap J,\nonumber \\
4584: J:=[L(a_{1}+{\textstyle \frac{1}{3}}),L(b_{1}+{\textstyle \frac{2}{3}})]\times\dotsb\times[L(a_{d}+{\textstyle \frac{1}{3}}),L(b_{d}+{\textstyle \frac{2}{3}})].\label{eq:influ}\end{gather}
4585:
4586: \item $G(L,M,\xi)$ is isotropic with the isotropic structure constants
4587: bounded independently of $L$, $M$ and $\xi$.
4588: \end{enumerate}
4589: Notice that this definition gives special importance to the point
4590: zero. This is just for convenience --- in practice, this fact will
4591: have no significance.
4592:
4593: The core of this paper is the proof of the following theorem.
4594:
4595: \begin{thm}
4596: \label{thm:interp}Let $G^{1}$ and $G^{2}$ be two $d$-dimensional
4597: graphs with an $\alpha$-isotropic interpolation, $d=2,3$. Let $\mathcal{D}\subset[\frac{1}{4},\frac{3}{4}]^{d}$
4598: be an open polyhedron and let $\mathcal{E}$ be some open set. Let
4599: $a\in\mathcal{E}\cap\mathcal{D}$ be some point. Let $s>0$ be some
4600: number, and let $R^{i}$ be random walks on $G^{i}$ starting from
4601: $sa$ and stopped when hitting $\partial(s\mathcal{D})$. Then \[
4602: \mathbb{P}(\LE(R^{i})\subset s(\mathcal{E}+B(0,Cs^{-c})))>\mathbb{P}(\LE(R^{3-i})\subset s\mathcal{E})-Cs^{-c}.\]
4603:
4604: \end{thm}
4605: Remember that {}``a random walk starting from $sa$'', means that
4606: it starts from the point of $G^{i}$ closest to $sa$; that $\mathcal{E}+B(0,Cs^{-c})$
4607: refers to the set of points within distance $Cs^{-c}$ from $\mathcal{E}$;
4608: and that an {}``open polyhedron'' is any open set whose boundary
4609: is made of non-degenerate linear polyhedra of dimension $d-1$, and
4610: that we do not require that the boundary of the polyhedron be connected,
4611: but we do not allow slits.
4612:
4613: A comment is due on the use of constants here. They all depend on
4614: $a$, $\mathcal{D}$ and on the graphs $G^{1}$ and $G^{2}$ (in fact
4615: they don't really depend on $a$ but we will have no use for this
4616: fact). Like in previous chapters, they depend only on $\alpha$ and
4617: the global bound for the isotropic structure constants over all of
4618: $G(L,M,\xi)$ and not on other properties of $G^{i}$. However, there
4619: is no need to continue to point this fact out --- we only did so in
4620: previous chapters in order to be able to analyze walks on $G(L,M,\xi)$
4621: simultaneously, and we will not have families of isotropic interpolations
4622: in the future.
4623:
4624:
4625: \subsection{Proof of theorem \refs{thm:interp}}
4626:
4627: The argumentation in this section is very similar to that of \cite[section 4]{K},
4628: so we will be brief.
4629:
4630: \begin{lem}
4631: \label{lem:xizhit}Let $G^{i},a,s$ be as in theorem \ref{thm:interp}.
4632: Assume $s\leq LM$ with $L$ and $M$ satisfying (\ref{eq:iirMreq})
4633: and $L$ sufficiently large (in addition to the restriction of (\ref{eq:iirMreq})).
4634: Let $\xi^{1}$,$\xi^{2}$ be two configurations which differ only
4635: in one point $z$, so in particular $H:=G(L,M,\xi^{i})\cap([\frac{1}{3}L,s-\frac{1}{3}L]^{d}\setminus B(Lz,16L))$
4636: does not depend on $i$. Let $\mathcal{B}\subset H$ be any subset
4637: such that $a$ is in a finite component of $G(L,M,\xi^{i})\setminus\mathcal{B}$
4638: (think about \emph{$\mathcal{B}$} as the boundary of some $\mathcal{D}$,
4639: $a\in\mathcal{D}$). Let $R^{i}$ be random walks on $G(L,M,\xi^{i})$
4640: starting from $sa$. Let $b\in\mathcal{B}$ and define\[
4641: p^{i}=\mathbb{P}(R^{i}(T^{i}(\mathcal{B}))=b).\]
4642: Then $|p^{1}-p^{2}|\leq C(\mathcal{G})L^{-c(\mathcal{G})}\max p^{1},p^{2}$.
4643: \end{lem}
4644: \begin{proof}
4645: [Proof sketch]Define stopping times $T_{j}^{i}$ on $\partial B(Lz,3L)$
4646: and $\partial B(Lz,12L)$ alternatively. The graphs $G(L,M,\xi^{i})$
4647: are identical outside $B(Lz,3L)$ hence lemma \ref{lem:iiexit2} shows
4648: that the transition probabilities, up to $CL^{-c}$ do not depend
4649: on $i$. Further, the probability to reach $T_{j}^{i}$ without hitting
4650: $\mathcal{B}$ drops like $e^{-cj}$ in three dimensions and like
4651: $e^{-cj/\log M}$ in two dimensions. Harnack's inequality on $B(Lz,16L)$
4652: shows that the probabilities to hit $b$ after $T_{j}$ conditioned
4653: on not hitting it before are, up to a constant, independent of $j$.
4654: Hence these $CL^{-c}$ errors do not accumulate to more than $CL^{-c}\log M\leq CL^{-c}$.
4655: See \cite[lemma 16]{K} for a detailed argument.
4656: \end{proof}
4657: \begin{lem}
4658: \label{lem:xizLE}Let $G^{i},a,s,L,M,\xi^{i},z,\mathcal{B}$ and $b$
4659: be as in lemma \ref{lem:xizhit}. Let $R^{i}$ be random walks on
4660: $G(L,M,\xi^{i})$ starting from $sa$ and conditioned to hit $\mathcal{B}$
4661: in $b$. Let $\zeta^{i}$ be the segment of $\LE(R^{i})$ until first
4662: hitting $\partial B(Lz,16L)$ or all of $\LE(R^{i})$ if $\LE(R^{i})\cap\partial B(Lz,16L)=\emptyset$.
4663: Then\[
4664: \sum_{\gamma}|\mathbb{P}(\zeta^{1}=\gamma)-\mathbb{P}(\zeta^{2}=\gamma)|\leq C(\mathcal{G})L^{-c(\mathcal{G})}.\]
4665: where the sum is over all simple paths $\gamma$ from $sa$ to $\partial B(Lz,16L)\cup\partial(s\mathcal{D})$.
4666: \end{lem}
4667: \begin{proof}
4668: [Proof sketch]The crucial point here is that $\zeta^{i}$ depends
4669: only on what happens outside $B(Lz,16L)$ (quite unlike the other
4670: portion of $\LE(R^{i})$). Therefore the same argument as in the previous
4671: lemma works here. Conditioning on all the entry exit points from $\partial B(Lz,3L)$
4672: and $\partial B(Lz,12L)$ the probabilities are identical, and in
4673: average in $\gamma$ it is enough to consider $\log^{2}L$ such points.
4674: See \cite[lemma 17]{K} for a detailed argument.
4675: \end{proof}
4676: The proof of theorem \ref{thm:interp} is also detailed in \cite{K}
4677: where it is called the {}``main lemma''. Since this is a crucial
4678: part of the argument, I prefer to bring it here in full.
4679:
4680: Since the theorem is symmetric in $i$, fix it to be $1$. Let $\epsilon>0$
4681: be some constant (depending on $\mathcal{G}$) which will be fixed
4682: later. We define $L$ by \begin{equation}
4683: 34L=s^{1-\epsilon}\label{eq:defepst3}\end{equation}
4684: and $M=\left\lceil s/L\right\rceil $. We assume $L$ and $M$ satisfy
4685: the requirements (\ref{eq:iirMreq}) of the isotropic interpolations,
4686: which will hold if $\epsilon<\min\frac{1}{2}\alpha,\frac{1}{2}$ and
4687: $L$ is large enough. We will also assume $L$ is sufficiently large
4688: such that lemmas \ref{lem:xizhit} and \ref{lem:xizLE} hold, and
4689: also that all edges of all graphs $G(L,M,\xi)$ are shorter than $L$
4690: and every ball of radius $L$ in $\mathbb{R}^{d}$ contains at least
4691: one point from every $G(L,M,\xi$). All these requirements translate
4692: to $s>C(\epsilon,\mathcal{G})$.
4693:
4694: Let $\delta=\delta(\epsilon,\mathcal{G})$ be the quantity given by
4695: theorem \ref{thm:QL} (page \pageref{thm:QL}) for the $\epsilon$
4696: from (\ref{eq:defepst3}), for all the graphs $G(L,M,\xi)$ simultaneously.
4697: In other words, we have, if $s>C(\epsilon,\mathcal{G})$, \[
4698: \mathbb{E}^{sa}\QL(34L,s^{1-\delta},\LE(R[0,T(\partial\mathcal{D})]))\leq C(\mathcal{G})s^{-\delta}\]
4699: which holds for $R$ a random walk on any $G(L,M,\xi)$. With this
4700: $\delta$, define {}``bad'' subsets of $\{ x\in\{0,\dotsc,M-1\}^{d}:x/M\in\mathcal{D}\}$
4701: as follows:\begin{align}
4702: \Phi & :=\left\{ x:d(Lx,\partial_{\textrm{cont}}s\mathcal{E})\leq s^{1-\delta}+18L\right\} \label{eq:defPHI}\\
4703: \Psi & :=\left\{ x:d(Lx,\partial_{\textrm{cont}}s\mathcal{D})\leq17L\right\} .\label{eq:defPSI}\\
4704: \Theta & :=\left\{ x:d(Lx,sa)\leq17L\right\} \nonumber \end{align}
4705: It should be noticed that any $x\not\in\Psi$ satisfies $d(Lx,\partial_{G(L,M,\xi)}(s\mathcal{D}))>17L$
4706: and any $x\not\in\Phi$ satisfies $d(Lx,\partial_{G(L,M,\xi)}(s\mathcal{E}))>s^{1-\delta}+17L$,
4707: both for any configuration $\xi$.
4708:
4709: \begin{lem}
4710: \label{lem:0_to_Y}With the definitions above, let $Y\subset\{0,\dotsc,M-1\}^{d}\setminus(\Phi\cup\Psi\cup\Theta)$.
4711: Let $\xi^{1}$ and $\xi^{2}$ be two configurations such that $\xi^{i}|_{Y}\equiv i$
4712: but $\xi^{1}|_{Y^{c}}\equiv\xi^{2}|_{Y^{c}}$. Let $R^{i}$ be random
4713: walks on $G(L,M,\xi^{i})$ starting from $sa$ and stopped on $\partial\mathcal{D}$.
4714: Let \[
4715: p^{i}:=\mathbb{P}(\LE(R^{i})\subset s\mathcal{E}).\]
4716: Then\[
4717: |p^{1}-p^{2}|\leq C(\mathcal{G})s^{-\delta}\log\# Y+C(\mathcal{G})L^{-c(\mathcal{G})}\# Y.\]
4718:
4719: \end{lem}
4720: \begin{proof}
4721: For every $0\leq k\leq\# Y$, let $D_{k}$ be a random subset of $Y$
4722: of size $k$, let $\xi_{k}$ be the configuration which is 2 on $D_{k}$,
4723: 1 on $Y\setminus D_{k}$ and identical to $\xi^{1}$ outside $Y$.
4724: Define $H_{k}:=G(L,M,\xi_{k})$. Let $a_{k}$ be the point of $H_{k}$
4725: closest to $sa$. Let $S_{k}$ be a random walk on $H_{k}$ starting
4726: from $a_{k}$ and stopped on $B_{k}:=\partial_{H_{k}}s\mathcal{D}$.
4727: Let \[
4728: p_{k}:=\mathbb{P}(\LE(S_{k})\subset s\mathcal{E})\]
4729: where $\mathbb{P}$ here is over both the walk and the randomness
4730: of the graph (notice that $p^{1}=p_{0}$ and $p^{2}=p_{\# Y}$). The
4731: lemma will be proved once we show that
4732:
4733: \begin{equation}
4734: |p_{k}-p_{k+1}|\leq Cs^{-\delta}\left(\frac{1}{k+1}+\frac{1}{\# Y-k}\right)+CL^{-c}.\label{step1}\end{equation}
4735:
4736:
4737: For this purpose, couple $D_{k}$ and $D_{k+1}$ such that $D_{k}\subset D_{k+1}$.
4738: Let $\Delta_{k}\subset\Delta_{k+1}$ be subsets of $Y$ of sizes $k$
4739: and $k+1$ and let $z=\Delta_{k+1}\setminus\Delta_{k}$. For most
4740: of the rest of the lemma, we condition on the event (denote it by
4741: $\mathcal{X}$) that $D_{k}=\Delta_{k}$ and $D_{k+1}=\Delta_{k+1}$.
4742: Let $Z=B(Lz,16L)$. We construct $\LE(S_{k})$ as follows:
4743: \begin{itemize}
4744: \item let $b_{k}$ be a random point on $B_{k}$ chosen with the hitting
4745: probabilities of $S_{k}$.
4746: \item Let $\check{S}_{k}$ be a random walk from $a_{k}$ to $B_{k}$ conditioned
4747: to hit $b_{k}$.
4748: \item Let $\check{\gamma}_{k}$ be a random simple path from $a_{k}$ to
4749: $\partial Z\cup\{ b_{k}\}$, which has the same distribution as the
4750: segment of $\LE(\check{S}_{k})$ until $\partial Z$ (including the
4751: first point in $\partial Z$), or all of $\LE(\check{S}_{k})$ if
4752: $\LE(\check{S}_{k})\cap\partial Z=\emptyset$ (notice that $a_{k}\not\in Z$
4753: from the requirement $Y\cap\Theta=\emptyset$).
4754: \item Let $c_{k}$ be the point where $\check{\gamma}_{k}$ hits $\partial Z$
4755: if it does.
4756: \item Let $\check{T}_{k}$ be a random walk on $H_{k}$ starting from $b_{k}$
4757: and conditioned to hit $B_{k}\cup\check{\gamma}_{k}$ in $c_{k}$,
4758: or $\emptyset$ if $\check{\gamma}_{k}$ never hits $\partial Z$.
4759: \item Let $\gamma_{k}=\check{\gamma}_{k}\cup\LE(\check{T}_{k})$.
4760: \end{itemize}
4761: An easy application of lemma \ref{lem:condLE_sym} (symmetry of conditioned
4762: loop-erased random walk) shows that $\gamma_{k}\sim\LE(S_{k})$. Lemma
4763: \ref{lem:xizhit} shows that, \begin{gather}
4764: \sum_{b}|q_{k}^{1}-q_{k+1}^{1}|\leq CL^{-c}\label{bk}\\
4765: q_{k}^{1}(b):=\mathbb{P}(b_{k}=b\,|\,\mathcal{X}).\nonumber \end{gather}
4766: Next we use lemma \ref{lem:xizLE} for the random walk on $H_{k}$
4767: starting from $a_{k}$, stopped on $B_{k}$, and conditioned to hit
4768: $b_{k}$. The definition of $\Psi$ (\ref{eq:defPSI}) ensures the
4769: condition $B_{k}\cap Z=\emptyset$ required by lemma \ref{lem:xizLE}.
4770: This shows that \begin{gather}
4771: \sum_{\gamma}|q_{k}^{2}-q_{k+1}^{2}|\leq CL^{-c}\qquad\forall b\in B_{k}\label{gamma_k}\\
4772: q_{k}^{2}(b,\gamma):=\mathbb{P}(\check{\gamma}_{k}=\gamma\,|\,\mathcal{X},\, b_{k}=b).\nonumber \end{gather}
4773: Thirdly, we again use lemma \ref{lem:xizLE}, this time for a random
4774: walk starting from $b_{k}$, stopped on $B_{k}\cup\check{\gamma}_{k}$
4775: and conditioned to hit $\check{c}_{k}$ to show that, when $\check{\gamma}_{k}'$
4776: is the portion of $\LE(\check{T}_{k})$ up to $Z$,\begin{gather}
4777: |q_{k}^{3}-q_{k+1}^{3}|\leq CL^{-c}\qquad\forall b,\gamma\label{gamma_k_prime}\\
4778: q_{k}^{3}(b,\gamma):=\mathbb{P}(\check{\gamma}_{k}'\subset s\mathcal{E}\,|\,\mathcal{X},\, b_{k}=b,\,\check{\gamma}_{k}=\gamma).\nonumber \end{gather}
4779: Summing (\ref{bk}), (\ref{gamma_k}) and (\ref{gamma_k_prime}) gives\begin{eqnarray}
4780: \lefteqn{|\mathbb{P}((\check{\gamma}_{k}\cup\check{\gamma}_{k}')\subset s\mathcal{E}\,|\,\mathcal{X})-\mathbb{P}((\check{\gamma}_{k+1}\cup\check{\gamma}_{k+1}')\subset s\mathcal{E}\,|\,\mathcal{X})|=}\nonumber \\
4781: & \qquad & =\left|\sum_{b,\gamma\subset s\mathcal{E}}q_{k}^{1}q_{k}^{2}q_{k}^{3}-q_{k+1}^{1}q_{k+1}^{2}q_{k+1}^{3}\right|\leq\nonumber \\
4782: & & \leq CL^{-c}+\left|\sum_{b,\gamma\subset s\mathcal{E}}(q_{k}^{1}q_{k}^{2}-q_{k+1}^{1}q_{k+1}^{2})q_{k}^{3}\right|\leq\dotsb\leq\nonumber \\
4783: & & \leq CL^{-c}.\label{no_quasi_loop}\end{eqnarray}
4784: In other words, we have proved that the probabilities (for $k$ and
4785: $k+1$) that both segments of $\LE(S_{k})$, leading up to $Z$ and
4786: from $Z$ to $B_{k}$ to be in $s\mathcal{E}$ are close. Thus the
4787: only case we have not covered is that $\check{\gamma}_{k}\cup\check{\gamma}_{k}'\subset s\mathcal{E}$
4788: but $\LE(S_{k})\not\subset s\mathcal{E}$. But $Lz$ is far from $\partial s\mathcal{E}$
4789: (because $Y\cap\Phi=\emptyset$ and the definition of $\Phi$ (\ref{eq:defPHI}))
4790: so we get a quasi loop near $Lz$, namely\[
4791: Lz\in\mathcal{QL}(17L,s^{1-\delta},\LE(S_{k})).\]
4792: Denote therefore\[
4793: q_{k}^{4}(x):=\mathbb{P}(Lx\in\mathcal{QL}(17L,s^{1-\delta},\LE(S_{k}))\,|\,\mathcal{X})\]
4794: and then write (\ref{no_quasi_loop}) as\[
4795: |\mathbb{P}(\LE(S_{k})\subset s\mathcal{E}\,|\,\mathcal{X})-\mathbb{P}(\LE(S_{k+1})\subset s\mathcal{E}\,|\,\mathcal{X})|\leq CL^{-c}+q_{k}^{4}(z)+q_{k+1}^{4}(z).\]
4796: We now integrate over $\mathcal{X}$. We get\begin{equation}
4797: |p_{k}-p_{k+1}|\leq CL^{-c}+\mathbb{E}q_{k}^{4}(z)+\mathbb{E}q_{k+1}^{4}(z).\label{eq:quasi_loop_left}\end{equation}
4798:
4799:
4800: The estimate of (\ref{eq:quasi_loop_left}) is where the random choice
4801: of the sets $D_{k}$ plays its part. Theorem \ref{thm:QL} gives us
4802: that \[
4803: \sum_{x\in Y\setminus D_{k}}q_{k}^{4}(x)\leq\sum_{x\in Y}q_{k}^{4}(x)\leq C\mathbb{E}\QL(34L,s^{1-\delta},\LE(R[0,T(\partial\mathcal{D})]))\leq Cs^{-\delta}.\]
4804: Now, if we think about the coupling of $D_{k}$ and $D_{k+1}$ as
4805: {}``$D_{k+1}$ is the addition of a random $z$ to $D_{k}$'', then
4806: $z$ is is obviously independent of the walk on $G(r,M,\xi_{k})$
4807: so we have \[
4808: \mathbb{E}q_{k}(z)\leq\frac{Cs^{-\delta}}{\#(Y\setminus D_{k})}=\frac{Cs^{-\delta}}{\# Y-k}.\]
4809: For $q_{k+1}(z)$ we similarly think about $D_{k}$ as the removal
4810: of a random $z$ from $D_{k+1}$ and get\[
4811: \mathbb{E}q_{k+1}(z)\leq\frac{Cs^{-\delta}}{k+1}\]
4812: and the lemma is proved.
4813: \end{proof}
4814: Continuing the proof of the theorem, we first apply lemma \ref{lem:0_to_Y}
4815: to the configurations $\xi^{1}\equiv1$ and \[
4816: \xi^{2}=\begin{cases}
4817: 1 & \Phi\cup\Psi\cup\Theta\\
4818: 2 & \textrm{otherwise}\end{cases}.\]
4819: Denoting by \[
4820: p^{i}:=\mathbb{P}_{G(L,M,\xi^{i})}^{sa}(\LE(R)\subset s\mathcal{E})\]
4821: ($R$ here is $R[0,T(\partial\mathcal{D})]$ and will stay so for
4822: a while). We get \begin{equation}
4823: |p^{1}-p^{2}|\leq Cs^{-\delta}\log s+CL^{-c}M^{d}.\label{eq:step1}\end{equation}
4824:
4825:
4826: Next, we wish to remove $\Phi$. For this purpose we define \begin{eqnarray*}
4827: p^{3} & := & \mathbb{P}_{G(L,M,\xi^{2})}^{sa}(\LE(R)\subset s\mathcal{E}_{2})\\
4828: \mathcal{E}_{2} & := & \mathcal{E}+B(0,2s^{-\delta}+37L/s)\end{eqnarray*}
4829: and get $p^{3}>p^{2}$. The definition of $\mathcal{E}_{2}$ gives
4830: us that\[
4831: d(Lx,\partial_{\textrm{cont}}s\mathcal{E}_{2})>s^{1-\delta}+18L\quad\forall x\in\Phi.\]
4832: so we can use lemma \ref{lem:0_to_Y} with $\xi^{2}$, \[
4833: \xi^{3}=\begin{cases}
4834: 1 & \Psi\cup\Theta\\
4835: 2 & \textrm{otherwise}\end{cases}\]
4836: and the domain $\mathcal{E}_{2}$ to get \begin{equation}
4837: |p^{3}-p^{4}|\leq Cs^{-\delta}\log s+CL^{-c}M^{d}.\label{eq:step2}\end{equation}
4838: where $p^{4}=\mathbb{P}_{G(L,M,\xi^{3})}^{sa}(\LE(R)\subset s\mathcal{E}_{2})$.
4839:
4840:
4841: The third step is to get rid of $\Theta$. For this purpose we define
4842: \[
4843: \mathcal{E}_{3}:=\mathcal{E}_{2}+B(0,s^{-\epsilon/2})\]
4844: and $p^{5}=\mathbb{P}_{G(L,M,\xi^{3})}^{sa}(\LE(R)\subset s\mathcal{E}_{3})$
4845: so that $p^{5}>p^{4}$. Find some point $b\in\mathcal{E}_{3}$ with
4846: $|a-b|=35L/s$ (which can always be done if $s$ is sufficiently large)
4847: and use lemma \ref{lem:contsp} to find that\begin{multline}
4848: |p^{5}-p^{6}|\leq C\left(\frac{35L}{sd(a,\partial_{\textrm{cont}}(\mathcal{E}_{3}\cap\mathcal{D})}\right)^{c}\leq\\
4849: \leq C\left(\frac{35L}{\min s^{1-\epsilon/2},sd(a,\partial_{\textrm{cont}}\mathcal{D})}\right)^{c}=C(\epsilon,a,\mathcal{D},\mathcal{G})s^{-c(\epsilon,a,\mathcal{D},G)}.\label{eq:step3a}\end{multline}
4850: where $p^{6}=\mathbb{P}_{G(L,M,\xi^{3})}^{sb}(\LE(R)\subset s\mathcal{E}_{3})$.
4851: We can now apply lemma \ref{lem:0_to_Y} with $\xi^{3}$, \[
4852: \xi^{4}=\begin{cases}
4853: 1 & \Psi\\
4854: 2 & \textrm{otherwise,}\end{cases}\]
4855: the domain $\mathcal{E}_{3}$ and the point $b$ to get \begin{equation}
4856: |p^{6}-p^{7}|\leq Cs^{-\delta}+CL^{-c}.\label{eq:step3b}\end{equation}
4857: where $p^{7}=\mathbb{P}_{G(L,M,\xi^{4})}^{sb}(\LE(R)\subset s\mathcal{E}_{3})$.
4858: We apply lemma \ref{lem:contsp} again to return to $a$: we get $|p^{7}-p^{8}|\leq Cs^{-c}$
4859: where $p^{8}=\mathbb{P}_{G(L,M,\xi^{4})}^{sa}(\LE(R)\subset s\mathcal{E}_{3})$.
4860:
4861: \label{page:begstep4}Finally, we need to get rid of $\Psi$. Let
4862: $\mathcal{D}_{2}$ be a shrinking of $\mathcal{D}$ by $21L/s$ i.e.\[
4863: \mathcal{D}_{2}:=\{ x:B(x,21L/s)\subset\mathcal{D}\}\]
4864: so we have $L\Psi\cap s\mathcal{D}_{2}=\emptyset$. Let $\delta_{2}$
4865: be the value given by theorem \ref{thm:QL} for $\epsilon/2$, $\epsilon$
4866: from (\ref{eq:defepst3}), for all the graphs $G(L,M,\xi)$ simultaneously,
4867: and define\[
4868: \mathcal{E}_{4}:=\mathcal{E}_{3}+B(0,s^{-\delta_{2}}+s^{-\epsilon/2}).\]
4869: Let \[
4870: p^{9}=\mathbb{P}_{G(L,M,\xi^{4})}^{sa}(\LE(R[0,T(\partial\mathcal{D}_{2})])\subset s\mathcal{E}_{4})\]
4871:
4872:
4873: \begin{lem}
4874: \label{sublem:step4}\begin{equation}
4875: p^{9}>p^{8}-Cs^{-c}\label{eq:step4}\end{equation}
4876: where $C$ and $c$ may depend on $\mathcal{D}$ and $\epsilon$ in
4877: addition to $\mathcal{G}$.
4878: \end{lem}
4879: \begin{proof}
4880: $\mathbb{R}^{d}\setminus\mathcal{D}$ has a finite number of connected
4881: components, $\{ Q_{i}\}$ which are all polyhedra. For each $Q_{i}$
4882: we may use lemma \ref{lem:polyh} to get that, for every $1<r_{1}<r_{2}<s$,
4883: every $\xi$ and every $v\in G(L,M,\xi)$, $d(v,sQ_{i})\leq r_{1}$,
4884: \begin{equation}
4885: \mathbb{P}(T_{v,r_{2}}<T(sQ_{i}))\leq C(\mathcal{G},Q_{i})\left(\frac{r_{1}}{r_{2}}\right)^{c(\mathcal{G},Q_{i})}.\label{eq:defKk}\end{equation}
4886: Let $K=\max_{i}C(\mathcal{G},Q_{i})$ and $k=\min_{i}c(\mathcal{G},Q_{i})$.
4887:
4888: Now $p^{8}$ and $p^{9}$ measure walks on the same graph stopped
4889: at $\partial s\mathcal{D}$ and $\partial s\mathcal{D}_{2}$ respectively.
4890: Therefore we may couple these walks so that the first is a continuation
4891: of the second. In other words, define $t_{1}>t_{2}$ be the stopping
4892: times of $R$ on $\partial s\mathcal{D}$ and $\partial s\mathcal{D}_{2}$
4893: (define $t_{2}=0$ if $a\not\in s\mathcal{D}_{2}$) then the question
4894: reduces to an estimate of \begin{equation}
4895: \mathbb{P}\left(\left\{ \LE(R[0,t_{2}])\not\subset s\mathcal{E}_{4}\right\} \cap\left\{ \LE(R[0,t_{1}])\subset s\mathcal{E}_{3}\right\} \right).\label{eq:St2_large_St1_small}\end{equation}
4896: Now, the definition of $\mathcal{D}_{2}$ gives that the distance
4897: of $R(t_{2})$ from the closest connected component of $\mathbb{R}^{d}\setminus(s\mathcal{D})$
4898: is $\leq21L$. From the definition of $K$ and $k$ we get\begin{equation}
4899: \mathbb{P}(R[t_{2},t_{1}]\text{ exits }B(R(t_{2}),{\scriptstyle \frac{1}{2}}s^{1-\epsilon/2}))\leq K\left(\frac{42L}{s^{1-\epsilon/2}}\right)^{k}=C(\mathcal{D},\mathcal{G})s^{-\epsilon c(\mathcal{D},\mathcal{G})}.\label{eq:r_large}\end{equation}
4900: On the other hand, if $R[t_{2},t_{1}]\subset B(R(t_{2}),{\scriptstyle \frac{1}{2}}s^{1-\epsilon/2})$
4901: and in addition the event of (\ref{eq:St2_large_St1_small}) holds
4902: then we can conclude that $R(t_{2})\in\mathcal{QL}({\scriptstyle \frac{1}{2}}s^{1-\epsilon/2},s^{1-\delta_{2}},\LE(S_{3}[0,t_{2}]))$
4903: which means that \[
4904: \QL(s^{1-\epsilon/2},s^{1-\delta_{2}},\LE(S_{3}[0,t_{2}]))\geq1.\]
4905: By theorem \ref{thm:QL} and Markov's inequality, the probability
4906: for that is $\leq Cs^{-\delta_{2}}$. This ends the lemma.
4907: \end{proof}
4908: The definition of isotropic interpolation shows $\overline{G(L,M,\xi^{4})\cap\mathcal{D}_{2}}=\overline{G^{2}\cap\mathcal{D}_{2}}$.
4909: Therefore we may define $R$ to be a random walk on $G^{2}$ starting
4910: from $sa$ and get that $p^{9}=\mathbb{P}(\LE(R[0,T(\partial s\mathcal{D}_{2})])\subset s\mathcal{E}_{4})$.
4911: We only need to return from $\mathcal{D}_{2}$ to $\mathcal{D}$ so
4912: write\[
4913: p^{10}:=\mathbb{P}(\LE(R[0,T(\partial s\mathcal{D})])\subset s\mathcal{E}_{5})\]
4914: where\[
4915: \mathcal{E}_{5}:=\mathcal{E}_{4}+B(0,s^{-\epsilon/2}).\]
4916:
4917:
4918: \begin{lem}
4919: With the definitions above\begin{equation}
4920: p^{10}>p^{9}-Cs^{-c}.\label{eq:step5}\end{equation}
4921: where again $C$ and $c$ may depend on $\mathcal{D}$ and $\epsilon$
4922: in addition to $\mathcal{G}$.
4923: \end{lem}
4924: \begin{proof}
4925: As in lemma \ref{sublem:step4}, it is enough to show that\[
4926: \mathbb{P}(\LE(R[0,t_{1}])\not\subset s\mathcal{E}_{5},\,\LE(R[0,t_{2}])\subset s\mathcal{E}_{4})\leq Cs^{-c}.\]
4927: with the same $t_{1}$ and $t_{2}$. Unlike in sublemma \ref{sublem:step4},
4928: this requires no recourse to theorem \ref{thm:QL} but rather follows
4929: directly from lemma \ref{lem:polyh} since this event implies that
4930: $R[t_{2},t_{1}]\not\subset R(t_{2})+B(0,s^{1-\epsilon/2})$ whose
4931: probability can be bounded by $K(21L/\linebreak[4]s^{1-\epsilon/2})^{k}=Cs^{-c}$
4932: with the same $K$ and $k$ as in lemma \ref{sublem:step4}.\label{page:endstep5}
4933: \end{proof}
4934: Summing up (\ref{eq:step1}), (\ref{eq:step2}), (\ref{eq:step3a}),
4935: (\ref{eq:step3b}), (\ref{eq:step4}) and (\ref{eq:step5}) we get\newc{c:Lpow}
4936: for some $c_{\ref{c:Lpow}}(\mathcal{G})$, \begin{equation}
4937: p^{10}>p^{1}-C(\epsilon,a,\mathcal{D},\mathcal{G})s^{-c(\epsilon,a,\mathcal{D},\mathcal{G})}-C(\mathcal{G})M^{d}L^{-c_{\ref{c:Lpow}}}.\label{just_choose_M}\end{equation}
4938: The only thing left now is to choose $\epsilon$. For $\epsilon<c_{\ref{c:Lpow}}/2d$,
4939: we will have $M^{d}L^{-c_{\ref{c:Lpow}}}\leq Cs^{-c_{\ref{c:Lpow}}/2}$.
4940: This finishes the proof of theorem \ref{thm:interp}.\qed
4941:
4942:
4943: \subsection{The limit process}
4944:
4945: In this section we derive consequences from theorem \ref{thm:interp}:
4946: we will prove the following theorems:
4947:
4948: \begin{thm}
4949: \label{thm:scallim}Let $G$ be a $d$-dimensional isotropic graph
4950: with an isotropic interpolation to $2G$. Let $\mathcal{D}\subset\mathbb{R}^{d}$
4951: be a polyhedron and let $a\in\mathcal{D}$. Let $\mathbb{P}_{n}$
4952: be the distribution of the loop-erasure of a random walk on $2^{n}\mathcal{D}\cap G$
4953: starting from $2^{n}a$ and stopped when hitting $\partial2^{n}\mathcal{D}$,
4954: multiplied by $2^{-n}$. Then $\mathbb{P}_{n}$ converge in the space
4955: $\mathcal{M}(\mathcal{H}(\overline{\mathcal{D}}))$.
4956: \end{thm}
4957: Recall that $\mathcal{H}(\mathcal{X})$ is the space of compact subsets
4958: of $\mathcal{X}$ with the Hausdorff metric, and $\mathcal{M}(\mathcal{X})$
4959: is the space of measures on $\mathcal{X}$ with the topology of weak
4960: convergence. $2G$ is the graph gotten by stretching $G$ by $2$
4961: uniformly and possibly multiplying all weights by a constant $\alpha$
4962: (this last action does not change the process of course). In other
4963: words, the theorem holds if for any $\alpha$ there is an isotropic
4964: interpolation between $G$ and $2G$. Strangely enough, in 3 dimensions
4965: $\alpha$ will usually not be 1.
4966:
4967: The limit of $\mathbb{P}_{n}$ is called the {}``scaling limit''
4968: (of the loop-erased random walk) and $G$ is said to {}``have a scaling
4969: limit''.
4970:
4971: \begin{thm}
4972: \label{thm:univ}Let $G$ and $H$ be two $d$-dimensional isotropic
4973: graphs with an isotropic interpolation, and assume $G$ has a scaling
4974: limit. Then $H$ has a scaling limit and they are identical.
4975: \end{thm}
4976: We start with a proof of theorem \ref{thm:scallim}. In the following
4977: we shall abuse notations by denoting, for any subset $A$ in a metric
4978: space $X$ and any $\epsilon>0$,\begin{align*}
4979: A+B(\epsilon) & :=\{ x\in X:d(x,A)<\epsilon\}\\
4980: A+\overline{B}(\epsilon) & :=\{ x\in X:\exists a\in A,\, d(x,a)\leq\epsilon\}.\end{align*}
4981: We shall also need the following notation: For a set $E\subset\overline{\mathcal{D}}$
4982: relatively open, define subsets of $\mathcal{H}(\overline{\mathcal{D}})$,\begin{equation}
4983: \mathcal{S}(E):=\{ K\subset E\},\quad\mathcal{I}(E):=\{ K\cap E\neq\emptyset\}.\label{eq:defSIE}\end{equation}
4984:
4985:
4986: \begin{lem}
4987: \label{lem:montrik}Let $G,a$ and $\mathcal{D}$ be as in theorem
4988: \ref{thm:scallim} and let $\mathcal{E}_{1},\dotsc,\mathcal{E}_{k}\subset\mathcal{D}$
4989: be open or closed. Then for almost every $\epsilon>0$ the limit\[
4990: \lim_{n\to\infty}\mathbb{P}_{n}\Big(\mathcal{S}\Big(\bigcap_{i=1}^{k}\mathcal{E}_{i}+\overline{B(\epsilon)}\Big)\Big)\]
4991: exists.
4992: \end{lem}
4993: Here and below {}``almost every'' can be replaced with {}``except
4994: for a countable set''.
4995:
4996: \begin{proof}
4997: We may replace $\overline{B(\epsilon)}$ with $B(\epsilon)$: for
4998: $\mathcal{E}_{i}$ open $\mathcal{E}_{i}+\overline{B(\epsilon)}=\mathcal{E}_{i}+B(\epsilon)$
4999: and for $\mathcal{E}_{i}$ closed $\mathcal{E}_{i}+\overline{B(\epsilon)}=\overline{\mathcal{E}_{i}+B(\epsilon)}$
5000: and there is only a countable number of $\epsilon$-s for which the
5001: closing operation affects the problem at all. Further, we may also
5002: assume all $\mathcal{E}_{i}$ are open, since replacing each with
5003: $\mathcal{E}_{i}+B(\delta)$ and then taking $\delta$ to zero will
5004: prove the general case.
5005:
5006: Denote by $\LE$ the expression $\LE(R[0,T(\partial2^{n}\mathcal{D})])$
5007: when $n$ is assumed to be clear from the context. Denote\[
5008: \overline{p}(\mathcal{E}):=\varlimsup_{n\to\infty}\mathbb{P}_{n}(\mathcal{S}(\mathcal{E}))\]
5009: and similarly $\underline{p}(\mathcal{E})$. By theorem \ref{thm:interp}
5010: we have\begin{align*}
5011: \mathbb{P}_{G}^{2^{n}a}(\LE\subset2^{n}\cap\mathcal{E}_{i}) & =\mathbb{P}_{2G}^{2^{n+1}a}(\LE\subset2^{n+1}\cap\mathcal{E}_{i})<\\
5012: & <\mathbb{P}_{G}^{2^{n+1}a}\big(\LE\subset2^{n+1}((\cap\mathcal{E}_{i})+B(C2^{-nc}))\big)+C2^{-nc}\\
5013: \intertext{\textrm{and inductively for any $m$,}} & <\mathbb{P}_{G}^{2^{n+m}a}\big(\LE\subset2^{n+m}((\cap\mathcal{E}_{i})+B(C(2^{-nc}+\dotsb\\
5014: & \qquad+2^{-(n+m)c})))\big)+C(2^{-nc}+\dotsb+2^{-(n+m)c})<\\
5015: & <\mathbb{P}_{G}^{2^{n+m}a}\big(\LE\subset2^{n+m}((\cap\mathcal{E}_{i})+B(C2^{-nc}/(1-2^{-c})))\big)\;+\\
5016: & \qquad+C2^{-nc}/(1-2^{-c})\leq\\
5017: & \leq\mathbb{P}_{G}^{2^{n+m}a}\big(\LE\subset2^{n+m}\cap(\mathcal{E}_{i}+B(C2^{-nc}))\big)+C2^{-nc}.\end{align*}
5018: Taking $m$ to $\infty$ we get\[
5019: \mathbb{P}_{G}^{2^{n}a}(\LE\subset2^{n}\cap\mathcal{E}_{i})\leq\underline{p}(\cap(\mathcal{E}_{i}+B(C2^{-nc})))+C2^{-nc}\]
5020: and taking $n$ to $\infty$ gives\[
5021: \overline{p}(\cap\mathcal{E}_{i})\leq\lim_{\epsilon\to0^{+}}\underline{p}(\cap(\mathcal{E}_{i}+B(\epsilon))).\]
5022: Now, $\underline{p}(\cap(\mathcal{E}_{i}+B(\epsilon)))$ is a monotone
5023: function hence it is continuous except at a countable number of points.
5024: At each point $x$ of continuity we have\[
5025: \overline{p}(\cap(\mathcal{E}_{i}+B(x)))\leq\lim_{\epsilon\to0^{+}}\underline{p}(\cap(\mathcal{E}_{i}+B(x+\epsilon)))=\underline{p}(\cap(\mathcal{E}_{i}+B(x)))\leq\overline{p}(\cap(\mathcal{E}_{i}+B(x)))\]
5026: hence all the inequalities are equalities.
5027: \end{proof}
5028: \begin{rem*}
5029: It is not very difficult to construct an example of an open set $\mathcal{E}$
5030: (say with $G=\mathbb{Z}^{d}$, $\mathcal{D}=[\frac{1}{4},\frac{3}{4}]^{d}$
5031: and $a=\vec{\frac{1}{2}}$) such that $\mathbb{P}_{n}(\mathcal{S}(\mathcal{E}))$
5032: does not converge (construct $\mathcal{E}$ that it contains areas
5033: $\mathcal{E}_{k}$ which become connected only for $n$ above some
5034: $N_{k}$ and affect the probability significantly).
5035: \end{rem*}
5036: \begin{lem}
5037: \label{lem:OVOeps}Let $G,a$ and $\mathcal{D}$ be as in theorem
5038: \ref{thm:scallim}. Let $\mathcal{O}\subset\mathcal{H}(\overline{\mathcal{D}})$
5039: be an open set and let $\epsilon>0$. Then there exists an open set
5040: $\mathcal{V}\subset\mathcal{H}(\overline{\mathcal{D}})$ with $\mathcal{O}\subset\mathcal{V}\subset\mathcal{O}+B(\epsilon)$
5041: such that \begin{equation}
5042: \lim_{n\to\infty}\mathbb{P}_{n}(\mathcal{V})\label{eq:lemOVOeps}\end{equation}
5043: exists.
5044: \end{lem}
5045: \begin{proof}
5046: Let $E\subset\overline{\mathcal{D}}$ be relatively open, let $v_{1},\dotsc,v_{k}\in\overline{\mathcal{D}}$
5047: and let $a>b>0$. We define \[
5048: \mathcal{Q}(E,v_{1},\dotsc,v_{k},a,b):=\mathcal{S}(E+B(a+b))\cap\bigcap_{i=1}^{k}\mathcal{I}(B(v_{i},a-b)).\]
5049: Our main goal is to show that for $\mathcal{Q}_{1},\dotsc,\mathcal{Q}_{l}$
5050: with a common $a$ and $b$,\[
5051: \mathcal{Q}_{i}=\mathcal{Q}(E_{i},v_{1,i},\dotsc,v_{k_{i},i},a,b)\]
5052: one has, for every $a$ and almost every $b$, that\begin{equation}
5053: \lim_{n\to\infty}\mathbb{P}_{n}(\cap\mathcal{Q}_{i})\label{eq:capQi}\end{equation}
5054: exists. Collect all the $v_{i,j}$-s into a single list, $v_{1},\dotsc,v_{m}$,
5055: and let $I\subset\{1,\dotsc,m\}$ be arbitrary. We use lemma \ref{lem:montrik}
5056: for $E_{1}+B(a),\dotsc,E_{l}+B(a)$ and for $\mathbb{R}^{d}\setminus B(v_{i},a)$
5057: for all $i\in I$. We get that for almost every $0<b<a$, \begin{equation}
5058: \lim_{n\to\infty}\mathbb{P}_{n}\Big(\mathcal{S}\Big(\Big(\bigcap_{i=1}^{k}E_{i}+B(a+b)\Big)\setminus\Big(\bigcup_{i\in I}B(v_{i},a-b)\Big)\Big)\Big)\label{eq:oneI}\end{equation}
5059: exists. Now take some $b$ such that the limit (\ref{eq:oneI}) exists
5060: for all $I$. Subtracting (\ref{eq:oneI}) for a given $I$ from (\ref{eq:oneI})
5061: for $I=\emptyset$ gives that the limit\[
5062: \lim_{n\to\infty}\mathbb{P}_{n}\Big(\mathcal{S}\Big(\bigcap_{i=1}^{k}E_{i}+B(a+b)\Big)\cap\bigcup_{i\in I}\mathcal{I}(B(v_{i},a-b))\Big)\]
5063: exists. Since this holds for all $I$, we can use the inclusion-exclusion
5064: principle to show that the limit\[
5065: \lim_{n\to\infty}\mathbb{P}_{n}\Big(\mathcal{S}\Big(\bigcap_{i=1}^{k}E_{i}+B(a+b)\Big)\cap\bigcap_{i=1}^{m}\mathcal{I}(B(v_{i},a-b))\Big)\]
5066: exists, which is equivalent to (\ref{eq:capQi}).
5067:
5068: Proving the lemma is now easy. Take a finite set of $\left\{ K_{i}\right\} _{i=1}^{L}\in\mathcal{O}$
5069: such that $B(K_{i},\frac{1}{4}\epsilon)$ cover $\mathcal{O}$ (this
5070: is possible from the compactness of $\mathcal{H}$). For each $K_{i}$
5071: one can take $E_{i}=K_{i}+B(\frac{1}{4}\epsilon)$ and $v_{1,i},\dotsc,v_{k_{i},i}$
5072: to be a $\frac{1}{4}\epsilon$-net in $K_{i}$ and get that \[
5073: B(K_{i},{\textstyle \frac{1}{4}}\epsilon)\subset\mathcal{Q}(E_{i},v_{1,i},\dotsc,v_{k_{i},i},{\textstyle \frac{1}{2}}\epsilon,b)\subset B(K_{i},\epsilon)\quad\forall0<b<{\textstyle \frac{1}{4}}\epsilon.\]
5074: Denote this set by $\mathcal{Q}_{i}$. For every $I\subset\{1,\dotsc,L\}$
5075: we use (\ref{eq:capQi}) to see that for almost every $0<b<\frac{1}{4}\epsilon$,
5076: \[
5077: \lim_{n\to\infty}\mathbb{P}_{n}\Big(\bigcap_{i\in I}\mathcal{Q}_{i}\Big)\]
5078: exists. Take a $b$ such that the limit exists for all $I$. By the
5079: inclusion-exclusion principle we get that\[
5080: \lim_{n\to\infty}\mathbb{P}_{n}\Big(\bigcup_{i=1}^{L}\mathcal{Q}_{i}\Big)\]
5081: exists. Define $\mathcal{V}=\cup\mathcal{Q}_{i}$ and the lemma is
5082: finished.
5083: \end{proof}
5084:
5085:
5086: \begin{proof}
5087: [Proof of theorem \ref{thm:scallim}]Let $f:\mathcal{H}(\overline{\mathcal{D}})\to[0,\infty)$
5088: be a continuous function. Let $\epsilon>0$ and define\[
5089: \mathcal{O}_{i}=f^{-1}[\epsilon i,\infty).\]
5090: Then\[
5091: \epsilon\sum_{i=1}^{\infty}\mathbf{1}_{\mathcal{O}_{i}}\leq f<\epsilon\Big(1+\sum_{i=1}^{\infty}\mathbf{1}_{\mathcal{O}_{i}}\Big).\]
5092: By the compactness of $\mathcal{H}$ there exists some $\delta>0$
5093: such that $\mathcal{O}_{i}+B(\delta)\subset\mathcal{O}_{i-1}$ for
5094: all $i\geq1$ (note that $\mathcal{O}_{i}=\emptyset$ for $i$ sufficiently
5095: large). For every $i\geq1$ such that $\mathcal{O}_{i}\neq\emptyset$
5096: use lemma \ref{lem:OVOeps} to find a $\mathcal{O}_{i}\subset\mathcal{V}_{i}\subset\mathcal{O}_{i-1}$
5097: such that the limit (\ref{eq:lemOVOeps}) exists (for larger $i$-s
5098: define $\mathcal{V}_{i}=\emptyset$). We get\begin{align*}
5099: \varlimsup_{n\to\infty}\int fd\mathbb{P}_{n} & \leq\varlimsup_{n\to\infty}\int\epsilon\Big(1+\sum_{i=1}^{\infty}\mathbf{1}_{\mathcal{V}_{i}}\Big)d\mathbb{P}_{n}=\varliminf_{n\to\infty}\int\epsilon\Big(1+\sum_{i=1}^{\infty}\mathbf{1}_{\mathcal{V}_{i}}\Big)d\mathbb{P}_{n}\\
5100: & \leq\varliminf_{n\to\infty}\int\epsilon\Big(1+\sum_{i=0}^{\infty}\mathbf{1}_{\mathcal{O}_{i}}\Big)d\mathbb{P}_{n}\leq2\epsilon+\varliminf_{n\to\infty}\int fd\mathbb{P}_{n}.\end{align*}
5101: Since $\epsilon$ was arbitrary we see that the limit $\int fd\mathbb{P}_{n}$
5102: exists for any positive $f$. Any function $f$ can be written as
5103: $f^{+}-f^{-}$ hence we see that the limit exists for any continuous
5104: $f$. This finishes the theorem since by compactness, if $\mathbb{P}_{n}$
5105: does not converge it must have two subsequences converging to different
5106: values, which is a contradiction.
5107: \end{proof}
5108: We now move to the proof of theorem \ref{thm:univ}.
5109:
5110: \begin{lem}
5111: \label{lem:SLtoeps}Let $G$ be a graph with a scaling limit and let
5112: $a,\mathcal{D}$ and $\mathcal{E}_{1},\dotsc,\mathcal{E}_{k}$ be
5113: as in lemma \ref{lem:montrik}. Then for almost every $\epsilon>0$,\[
5114: \lim_{n\to\infty}\mathbb{P}_{n}(\mathcal{S}(\cap(\mathcal{E}_{i}+\overline{B(\epsilon)})))\]
5115: exists.
5116: \end{lem}
5117: \begin{proof}
5118: As in lemma \ref{lem:montrik} we may assume all $\mathcal{E}_{i}$
5119: are open and replace $\overline{B}$ with $B$. Denote $\overline{p}(\mathcal{E})$
5120: and $\underline{p}(\mathcal{E})$ as in lemma \ref{lem:montrik}.
5121: If they are different on $\cap(\mathcal{E}_{i}+B(x))$ for an uncountable
5122: number of $x$-s, then one of them would be a continuity point for
5123: both $\overline{p}$ and $\underline{p}$. Hence we get an interval
5124: $[\alpha,\beta]$ such that \[
5125: \underline{p}(\cap(\mathcal{E}_{i}+B(x)))\leq A<B\leq\overline{p}(\cap(\mathcal{E}_{i}+B(x)))\quad\forall x\in[\alpha,\beta].\]
5126: Define now\begin{align*}
5127: \mu(K) & :=\min\{ x:K\subset\cap(\mathcal{E}_{i}+B(x))\}\quad K\in\mathcal{H}(\overline{\mathcal{D}})\\
5128: f(K) & :=\begin{cases}
5129: 1 & \mu(K)\leq\alpha\\
5130: 0 & \mu(K)\geq\beta\\
5131: \textrm{linear in }\mu(K) & \textrm{otherwise}.\end{cases}\end{align*}
5132: It is easy to see that $f$ is continuous on $\mathcal{H}$. Therefore
5133: by the definition of scaling limit, one must have that\[
5134: \lim_{n\to\infty}\int fd\mathbb{P}_{n}\]
5135: exists. However, \begin{align*}
5136: \varliminf_{n\to\infty}\int fd\mathbb{P}_{n} & \leq\varliminf_{n\to\infty}\mathbb{P}_{n}(\mathcal{S}(\cap(\mathcal{E}_{i}+B(\beta))))\leq A\\
5137: \varlimsup_{n\to\infty}\int fd\mathbb{P}_{n} & \geq\varlimsup_{n\to\infty}\mathbb{P}_{n}(\mathcal{S}(\cap(\mathcal{E}_{i}+B(\alpha))))\geq B\end{align*}
5138: which is a contradiction.
5139: \end{proof}
5140: \begin{lem}
5141: Let $G$ and $H$ be as in theorem \ref{thm:univ} and let $a,\mathcal{D},\mathcal{E}_{1},\dotsc,\mathcal{E}_{k}$
5142: and $\epsilon$ be as in lemma \ref{lem:montrik}. Then for almost
5143: all $\epsilon>0$,\[
5144: \lim_{n\to\infty}\mathbb{P}_{G}^{2^{n}a}(2^{-n}\LE\subset\cap(\mathcal{E}_{i}+\overline{B(\epsilon)}))=\lim_{n\to\infty}\mathbb{P}_{H}^{2^{n}a}(2^{-n}\LE\subset\cap(\mathcal{E}_{i}+\overline{B(\epsilon)})).\]
5145: In particular both limits exist.
5146: \end{lem}
5147: \begin{proof}
5148: [Proof sketch]Denoting \[
5149: \overline{p}_{G}(x):=\varlimsup_{n\to\infty}\mathbb{P}_{G}^{2^{n}a}(2^{-n}\LE\subset\cap(\mathcal{E}_{i}+B(x)))\]
5150: a calculation analogous to that of lemma \ref{lem:montrik} gives
5151: that\[
5152: \underline{p}_{G}(x)\leq\lim_{\epsilon\to0^{+}}\underline{p}_{H}(x+\epsilon)\quad\overline{p}_{H}(x)\leq\lim_{\epsilon\to0^{+}}\overline{p}_{G}(x+\epsilon)\]
5153: hence the lemma holds for every $x$ which is a continuity for all
5154: $\overline{p}$ and $\underline{p}$ and for which $\underline{p}_{G}(x)=\overline{p}_{G}(x)$.
5155: Lemma \ref{lem:SLtoeps} is used here to show that this last condition
5156: holds almost everywhere.
5157: \end{proof}
5158: \begin{lem}
5159: Let $G$ and $H$ be as in theorem \ref{thm:univ} and let $\mathcal{O}$
5160: and $\epsilon$ be as in lemma \ref{lem:OVOeps}. Then there exists
5161: an $\mathcal{O}\subset\mathcal{V}\subset\mathcal{O}+B(\epsilon)$
5162: such that\[
5163: \lim_{n\to\infty}\mathbb{P}_{G}^{2^{n}a}(2^{-n}\LE\in\mathcal{V})=\lim_{n\to\infty}\mathbb{P}_{H}^{2^{n}a}(2^{-n}\LE\in\mathcal{V}).\]
5164:
5165: \end{lem}
5166: The proof of this lemma is a complete analogue of the proof of lemma
5167: \ref{lem:OVOeps}. Concluding theorem \ref{thm:univ} from it is similar
5168: to the conclusion of theorem \ref{thm:scallim} from lemma \ref{lem:OVOeps}.
5169: We shall omit both.
5170:
5171:
5172: \section{\label{sec:Examples}Examples}
5173:
5174: We start with a lemma on continuous functions
5175:
5176: \begin{lem}
5177: \label{lem:harmdiff}Let $f$ be harmonic on $B(0,1)$ and continuous
5178: on $\overline{B(0,1)}$, and assume that on $\partial B(0,1)$ $f$
5179: is $C^{2}$ and satisfies\[
5180: ||f||_{\infty}\leq1,\quad||f||_{1,\infty}\leq N\quad||f||_{2,\infty}\leq N^{2}\quad\forall i,j\]
5181: for some $N\geq1$. Then $|\partial f/\partial x_{i}|\leq CN$ inside
5182: $B(0,1)$.
5183: \end{lem}
5184: The notation $||f||_{k,p}$ stands for the Sobolev norm, which in
5185: this case simply means $||f||_{1,\infty}=\max_{x,i}|(\partial f/\partial\theta_{i})(x)|$
5186: and similarly for second derivatives.
5187:
5188: \begin{proof}
5189: The value of $f$ inside $B(0,1)$ is related to the value on the
5190: boundary by the Poisson kernel,\[
5191: f(x)=\int_{\partial B(0,1)}f(y)\frac{1-||x||^{2}}{||x-y||^{d}}\, dy\]
5192: where $dy$ is the surface area measure on $\partial B(0,1)$ normalized
5193: to be a probability measure. This immediately gives an estimate in
5194: the ball $B(0,\frac{1}{2})$ since \[
5195: \frac{\partial f}{\partial x_{i}}=\int_{\partial B(0,1)}f(y)\frac{\partial}{\partial x_{i}}\frac{1-||x||^{2}}{||x-y||^{d}}\, dy\leq C\int_{\partial B(0,1)}f(y)\, dy\leq C.\]
5196: Hence we need to estimate the derivatives only near the boundary.
5197:
5198: Let now $x\in B(0,1)\setminus B(0,\frac{1}{2})$, let $\theta$ be
5199: some direction on the sphere and let $T_{\epsilon}$ be a rotation
5200: by $\epsilon$ around the pole orthogonal to $x$ and $x+\epsilon\theta$
5201: in the direction $\theta$. Then the rotational invariance of the
5202: Poisson kernel allows to write\begin{align}
5203: \frac{\partial f}{\partial\theta}(x) & =\lim_{\epsilon\to0}\int_{\partial B(0,1)}\frac{f(y)-f(T_{\epsilon}y)}{\epsilon}\cdot\frac{1-||x||^{2}}{||x-y||^{2}}\, dy\leq\nonumber \\
5204: & \leq\int_{\partial B(0,1)}N\frac{1-||x||^{2}}{||x-y||^{2}}\, dy=N.\label{eq:tangent}\end{align}
5205: Since $||x||\geq\frac{1}{2}$ we get that the tangent derivatives
5206: are $\leq2N$. Therefore the lemma will be proved once we get a similar
5207: estimate for the radial derivative.
5208:
5209: A calculation similar to (\ref{eq:tangent}) shows that $|\partial^{2}f/\partial x_{i}^{2}|\leq4N^{2}$
5210: for $x_{i}$ a tangent direction. Hence by the harmonicity of $f$
5211: we get $|\partial^{2}f/\partial r^{2}|\leq4(d-1)N^{2}$. Examine now
5212: a point $x\in\partial B(0,1)$. Since $|f(x)|\leq1$ and $|f(x(1-1/N))|\leq1$
5213: there must be a point $y$ on the interval $[x,x(1-1/N)]$ such that
5214: $|(\partial f/\partial r)(y)|\leq2N.$ With the bound on $\partial^{2}f/\partial r^{2}$
5215: this gives that $|(\partial f/\partial r)(x)|\leq CN$. This allows
5216: to bound $\partial f/\partial r$ everywhere since\begin{align*}
5217: \frac{\partial f}{\partial r}(x) & =\lim_{\epsilon\to0^{+}}\int_{\partial B(0,1)}\frac{f(y)-f(y-\epsilon/||x||)}{\epsilon}\cdot\frac{1-||x||^{2}}{||x-y||^{2}}\, dy=\\
5218: & =\frac{1}{||x||}\int_{\partial B(0,1)}\frac{\partial f}{\partial r}(y)\cdot\frac{1-||x||^{2}}{||x-y||^{2}}\, dy\leq\frac{CN}{||x||}\leq CN.\qedhere\end{align*}
5219:
5220: \end{proof}
5221: Our purpose at this point is to prove that $\mathbb{Z}^{d}$ and $2\mathbb{Z}^{d}$
5222: have an isometric interpolation, which will allow to invoke theorem
5223: \ref{thm:scallim}. We start with the case $d=3$ and for which we
5224: weight $2\mathbb{Z}^{3}$ by $2$ i.e. assume all the edges have weight
5225: $2$ (the statement of theorem \ref{thm:scallim}, page \pageref{thm:scallim},
5226: allows us to do so). We shall show this for $\alpha=\frac{1}{9}$
5227: (the $\alpha$ from the definition of isotropic interpolation, (\ref{eq:iirMreq})
5228: on page \pageref{eq:iirMreq}). Therefore assume $L>6$ and $M\leq L^{1/9}$
5229: ($6$ is our choice for the $C$ from (\ref{eq:iirMreq})). Let $\xi\in\{1,2\}^{M^{3}}$.
5230: We need to construct a graph $G=G(L,M,\xi)$. We do it as follows.
5231:
5232: \begin{lyxlist}{00.00.0000}
5233: \item [Vertices]For every $(x,y,z)\in\{0,\dotsc,M-1\}^{3}$ such that $\xi(x,y,z)=1$
5234: we take every point of $\mathbb{Z}^{3}\cap[Lx,Lx+L)\times[Ly,Ly+L)\times[Lz,Lz+L)$
5235: to be a vertex of $G$. We call such vertices {}``vertices of type
5236: $1$''. If $\xi(x,y,z)=2$ we take every point of $2\mathbb{Z}^{3}\cap[Lx,Lx+L)\times[Ly,Ly+L)\times[Lz,Lz+L)$
5237: to be a vertex of $G$ and call it a vertex of type $2$. Outside
5238: $[0,LM)^{3}$ we choose between $\mathbb{Z}^{3}$ and $2\mathbb{Z}^{3}$
5239: by a majority vote on $\xi$ (or in any other way that gives $1$
5240: for $\xi\equiv1$ and $2$ for $\xi\equiv2$).
5241: \item [Edges]Two vertices of type $1$ will have an edge if and only if
5242: their distance is $1$, and in this case the edge will have weight
5243: $1$. Two vertices of type $2$ will have an edge if and only if their
5244: distance is $2$ and in this case the edge will have weight $2$.
5245: If $v$ is of type $1$ and $w$ of type $2$ then we need to find
5246: the vertex $x$ of type $1$ closest to $w$ (usually $|x-w|$ will
5247: be either $1$ or $2$); and define the weight by\begin{equation}
5248: \omega(v,w)=\frac{1}{|w-x|}\begin{cases}
5249: 1 & v=x\\
5250: 1/2 & |v-x|=1\\
5251: 1/4 & |v-x|=\sqrt{2}\\
5252: 0 & \textrm{otherwise}\end{cases}\label{eq:defomeclose}\end{equation}
5253: ($0$ here means no edge). See figure \ref{cap:type12}.%
5254: \begin{figure}
5255: \input{vx.pstex_t}
5256:
5257:
5258: \caption{\label{cap:type12}Lower-left square are vertices of type 1, upper-right
5259: are vertices of type 2. The weights $\omega$ are under the assumption
5260: that $|w-x|=1$; otherwise divide all weights by $2$.}
5261: \end{figure}
5262:
5263: \end{lyxlist}
5264: Requirements \ref{enu:isinI} and \ref{enu:isinII} from the definition
5265: of an isotropic interpolation are obvious. Hence the only thing we
5266: need to show is that these graphs are isotropic. It is clear that
5267: $G$ is always a Euclidean net so we need to verify the more delicate
5268: condition on exit probabilities in the definition of an isotropic
5269: graph (page \pageref{sub:DefinitionIsotropic}). This will follow
5270: from a comparison of the continuous and discrete Laplacian. Let therefore
5271: $f$ be a continuously harmonic function in a ball of radius $3$
5272: around some vertex $v$. Let $\Delta$ be the discrete Laplacian.
5273: Then
5274:
5275: \begin{enumerate}
5276: \item \label{enu:regpnt}If $d(v,X)>2$ for any square $X$ of the form
5277: $\{ Ln\}\times[0,LM]\times[0,LM]$, $[0,LM]\times\{ Ln\}\times[0,LM]$
5278: or $[0,LM]\times[0,LM]\times\{ Ln\}$ ($n\in\{0,\dotsc,M\}$), then
5279: $(\Delta f)(v)\leq C||f||_{4,\infty}$, the Sobolev norm being on
5280: the ball $B(v,3)$. In a ball of radius $r$ there are no more than
5281: $C(r^{3}/L+r^{2})$ points not satisfying this condition.
5282: \item \label{enu:plnpnt}If $d(v,X)>2$ for any segment $X$ of the form
5283: $\{ Ln\}\times\{ Lm\}\times[0,LM]$, $\{ Ln\}\times[0,LM]\times\{ Lm\}$
5284: or $[0,LM]\times\{ Ln\}\times\{ Lm\}$ ($n,m\in\{0,\dotsc,M\}$),
5285: then $(\Delta f)(v)\leq C||f||_{2,\infty}$. In a ball of radius $r$
5286: there are no more than $C(r^{3}/L^{2}+r)$ points not satisfying this
5287: condition.
5288: \item \label{enu:linepnt}For any $v$, $(\Delta f)(v)\leq C||f||_{1,\infty}$.
5289: \end{enumerate}
5290: Seeing \ref{enu:regpnt}-\ref{enu:linepnt} is not difficult. Write
5291: a Taylor expansion for $f$ around $v$ of order $4$, $2$ or $1$
5292: respectively, write\[
5293: (\Delta f)(v)=\sum_{w\sim v}\omega(v,w)(f(w)-f(v))\]
5294: and calculate and see that all terms except the error term vanish.
5295: Note that case \ref{enu:plnpnt} is the one where all the fancy weights
5296: in the {}``stitching'' between $\mathbb{Z}^{3}$ and $2\mathbb{Z}^{3}$
5297: are needed, and also the reason we had to take $2\mathbb{Z}^{3}$
5298: to have all the weights $2$. In fact \ref{enu:linepnt} holds in
5299: any graph with bounded degree.
5300:
5301: \begin{lem}
5302: \label{lem:i2iiiiso}A $3$-dimensional Euclidean net satisfying \ref{enu:regpnt}-\ref{enu:linepnt}
5303: is isotropic.
5304: \end{lem}
5305: \begin{proof}
5306: By the definition of an isotropic graph, we need to take some $v\in G$,
5307: some $r>0$, and a spherical triangle $A\subset\partial_{\textrm{cont}}B(v,r)$.
5308: We may assume that $r>1$ since otherwise (\ref{eq:defbrown}) holds
5309: automatically for $K\geq2$. Let $A^{-}=\{ x\in A:d(x,\partial A)\geq r^{4/5}\}$
5310: and $A^{+}:=\{ x\in\partial_{\textrm{cont}}B(v,r):d(x,A)\leq r^{4/5}\}$.
5311: $A^{-}$ could be empty and $A^{+}$could be all of $\partial_{\textrm{cont}}B(v,r)$
5312: but anyway we always have \[
5313: \int\mathbf{1}_{A^{-}}+Cr^{-1/5}\geq\int\mathbf{1}_{A}\geq\int\mathbf{1}_{A^{+}}-Cr^{-1/5}\]
5314: where $\int$ here is with respect to the surface area measure on
5315: $\partial_{\textrm{cont}}B(v,r)$, normalized to have total area $1$.
5316: The first step is to find two $C^{5}$ functions $f^{-}$ and $f^{+}$
5317: on $\partial_{\textrm{cont}}B(v,r)$ such that $||f^{-}||_{k,\infty},||f^{+}||_{k,\infty}\leq Cr^{-4k/5}$
5318: for $k=1,\dotsc,5$ and such that\begin{equation}
5319: f^{-}\leq\mathbf{1}_{A^{-}}\leq\mathbf{1}_{A^{+}}\leq f^{+},\quad\int f^{-}+Cr^{-1/5}\geq\int\mathbf{1}_{A}\geq\int f^{+}-Cr^{-1/5}.\label{eq:f1Ag}\end{equation}
5320: It is easy to construct such an $f^{\pm}$. For example, start with
5321: a spherically symmetric function $\eta$ which is $1$ in a spherical
5322: cap of radius $r^{4/5}$ and supported in a spherical cap of radius
5323: $2r^{4/5}$, $||\eta||_{k,\infty}\leq Cr^{-4k/5}$. Cover the sphere
5324: with a locally finite family of translations $\{ T_{1},\dotsc,T_{m}\}$
5325: of the $r^{4/5}$ cap (so that $\sum_{j}T_{j}(\eta)\geq1$), and define
5326: $\nu_{i}:=T_{i}(\eta)/\sum_{j}T_{j}(\eta)$ so that the $\nu_{i}$
5327: form a division of unity. Define $f^{-}$ to be the sum of all $\nu_{i}$
5328: supported inside $A^{-}$ and $f^{+}$ to be the sum of all $\nu_{i}$
5329: such that $\supp\nu_{i}\cap A^{+}\neq\emptyset$. Verifying all the
5330: properties of $f^{-}$ and $f^{+}$ is easy.
5331:
5332: We wish to discretize (\ref{eq:f1Ag}). Let $A^{*}\subset\partial B(v,r)$
5333: be a discrete version of $A$ as in the definition of isotropic graphs.
5334: We want to find functions $F^{-}$ and $F^{+}$ satisfying (\ref{eq:f1Ag})
5335: (with the integral being with respect to the discrete harmonic measure
5336: of $\partial B(v,r)$ starting from $v$). We shall only show the
5337: construction of $F^{-}$ --- the construction of $F^{+}$ is identical.
5338:
5339: Stretch $f^{-}$ to $\partial B(v,r+\lambda)$ where $\lambda$ is
5340: some constant such that $\partial B(v,r)\subset B(v,r+\lambda)$ ---
5341: for $\mathbb{Z}^{3}$ and $2\mathbb{Z}^{3}$ we can take $\lambda=3$.
5342: Extend $f^{-}$ to a harmonic function on $B(v,r+\lambda)$ continuous
5343: on $\overline{B(v,r+\lambda)}$. Call this extension $g^{1}$ and
5344: notice that $||g^{1}||_{k,\infty}\leq Cr^{-4k/5}$ for $k=1,\dotsc,4$
5345: (this follows from using lemma \ref{lem:harmdiff} for rescaled versions
5346: of $g^{1}$ and its derivatives). Let $a(w,x):=-G(w,x;\overline{B(v,r)})$
5347: be the discrete Green function. Define the following {}``correction''
5348: for $g^{1}$:\[
5349: D^{1}(w):=\sum_{x\in B(v,r)}(\Delta g^{1})(x)a(w,x).\]
5350: Because $\Delta a(\cdot,x)$ is a delta function at $x$, we get that
5351: $g^{2}:=g^{1}-D^{1}$ is discretely harmonic on $B(v,r)$. We also
5352: note that $g^{2}\equiv g^{1}$ on $\partial B(v,r)$. What we need
5353: is to estimate $D^{1}$ at $v$. Recall the estimate $a(v,w)\leq C|v-w|^{-1}$
5354: from lemma \ref{lem:a} (\ref{eq:atrans}). Summing on points of type
5355: \ref{enu:regpnt} we get that \begin{align*}
5356: \sum_{x\textrm{ of type \ref{enu:regpnt}}}|(\Delta g^{1})(x)a(v,x)| & \leq Cr^{-16/5}\sum_{s=0}^{\left\lfloor \log_{2}r\right\rfloor }\sum_{2^{s}\leq|v-x|<2^{s+1}}a(v,x)\leq\\
5357: & \leq Cr^{-16/5}\sum_{s=0}^{\left\lfloor \log_{2}r\right\rfloor }4^{s}\leq Cr^{-6/5}.\end{align*}
5358: Points of type \ref{enu:plnpnt} have a worse estimate, but are fewer.
5359: In particular, a ball of radius $2^{s}$ around $v$ will contain
5360: no more than $C\min(8^{s}/L+4^{s},L^{2}M^{3})$ such points. Assume
5361: first that $r\leq LM$ which implies $L\geq r^{9/10}$. We get\begin{eqnarray*}
5362: \lefteqn{{\sum_{x\textrm{ of type \ref{enu:plnpnt}}}|(\Delta g^{1})(x)a(v,x)|\leq Cr^{-8/5}\sum_{s=0}^{\left\lfloor \log_{2}r\right\rfloor }\sum_{2^{s}\leq|v-x|<2^{s+1}}a(v,x)\leq}}\\
5363: & & \qquad\leq Cr^{-8/5}\sum_{s=0}^{\left\lfloor \log_{2}r\right\rfloor }2^{s}+\frac{4^{s}}{L}\leq C(r^{-3/5}+r^{2/5}/L)\leq Cr^{-1/2}.\end{eqnarray*}
5364: If $r>LM$ (which implies $M\leq r^{1/10})$ a similar calculation
5365: shows that \[
5366: \sum\leq Cr^{-8/5}\Big(LM^{2}+\sum_{s=\left\lfloor \log_{2}LM\right\rfloor }^{\left\lfloor \log_{2}r\right\rfloor }L^{2}M^{3}2^{-s}\Big)\leq Cr^{-8/5}LM^{2}\leq Cr^{-1/2}.\]
5367: Finally, for points of type \ref{enu:linepnt} we get that a ball
5368: of radius $2^{s}$ will contain no more than $C\min(8^{s}/L^{2}+2^{s},LM^{3})$
5369: and an identical calculation shows that\[
5370: \sum_{w\textrm{ of type \ref{enu:linepnt}}}|(\Delta g^{1})(v)a(v,w)|\leq Cr^{-4/5}\log r+Cr^{-3/5}.\]
5371: Summing these three terms we get $|D^{1}(v)|\leq Cr^{-1/2}$.
5372:
5373: The only reason not to use $g^{2}$ directly is that $g^{2}\not\leq\mathbf{1}_{A^{*}}$
5374: on $\partial B(v,r)$. We do have, on $\partial B(v,r)$ that $g^{2}\equiv g^{1}\leq1$,
5375: but we need to correct $g^{2}$ to be $0$ outside $A^{*}$. Let $w\in\partial B(v,r)\setminus A^{*}$
5376: and let $w'=w\frac{r}{r+\lambda}$. Then\begin{align*}
5377: g^{2}(w) & =g^{1}(w)=f^{-}(w')=r^{-1}\int_{\partial_{\textrm{cont}}B(v,r)}f^{-}(x)\frac{r^{2}-||wr'||^{2}}{||w'-x||^{3}}\, dx\leq\\
5378: & \leq r^{-1}\int_{A^{-}}\frac{r^{2}-||w'||^{2}}{||w'-x||^{3}}\, dx\leq C/d(w',A^{-})\leq Cr^{-4/5}.\end{align*}
5379: Hence defining a second correction $D^{2}$ on $\partial B(v,r)$\[
5380: D^{2}(w)=\begin{cases}
5381: f^{2}(w) & w\not\in A^{*}\\
5382: 0 & \textrm{otherwise}\end{cases}\]
5383: and extending it to a discretely harmonic function on $B(v,r)$,
5384: we get, from the discrete maximum principle that $D^{2}(v)\leq Cr^{-4/5}$.
5385:
5386: Defining $F^{-}=g^{2}-D^{2}$ the lemma is now easy: By the definition
5387: of the harmonic measure, $\int F^{-}=F^{-}(v)$ (discrete case) and
5388: $\int f^{-}=f^{-}(v)$ (continuous case). $|F^{-}(v)-f^{-}(v)|\leq|D^{1}(v)|+|D^{2}(v)|\leq Cr^{-1/2}$.
5389: Therefore\begin{align*}
5390: p_{A^{*}} & =\int\mathbf{1}_{A^{*}}\geq\int F^{-}=F^{-}(v)\geq f^{-}(v)-Cr^{-1/2}=\int f^{-}-Cr^{-1/2}\geq\\
5391: & \!\!\stackrel{(\ref{eq:f1Ag})}{\geq}\int\mathbf{1}_{A}-Cr^{-1/5}=|A|-Cr^{-1/5}.\end{align*}
5392: A similar calculation with the similarly defined $F^{+}$ will show
5393: that $p_{A}\leq|A|+Cr^{-1/5}$ so (\ref{eq:defbrown}) is proved,
5394: $G$ is isotropic and the lemma is proved.
5395: \end{proof}
5396: \begin{conclusion*}
5397: Lemma \ref{lem:i2iiiiso}, the discussion before that and theorem
5398: \ref{thm:scallim} together show that $\mathbb{Z}^{3}$ has a scaling
5399: limit. Hence theorem \ref{thm:Z3scal} is proved in three dimensions.
5400: \end{conclusion*}
5401: For the case of $\mathbb{Z}^{2}$, we define $2\mathbb{Z}^{2}$ to
5402: have all the weights $1$. We construct that graphs $G(L,M,\xi)$
5403: equivalently, but with the weights defined as follows: If $v$ and
5404: $w$ are of the same type then the edges between them are as in the
5405: three dimensional cases and all the weights are $1$. If $v$ is of
5406: type $1$ and $w$ is of type $2$ we again find the vertex $x$ of
5407: type $1$ closest to $w$, and then define\[
5408: \omega(v,w)=\frac{1}{|w-x|}\begin{cases}
5409: 1 & v=x\\
5410: 1/2 & |v-x|=1\\
5411: 0 & \textrm{otherwise}\end{cases}\]
5412: (this is the equivalent of (\ref{eq:defomeclose})). A figure can
5413: be found in \cite[figure 1, page 10]{K}. The equivalents of \ref{enu:regpnt}-\ref{enu:linepnt}
5414: from page \pageref{enu:regpnt} are
5415:
5416: \begin{enumerate}
5417: \item If $d(v,X)>2$ for any segment $X$ of the form $\{ Ln\}\times[0,LM]$
5418: or $[0,LM]\times\{ Ln\}$ ($n\in\{0,\dotsc,M\}$), then $(\Delta f)(v)\leq C||f||_{4,\infty}$.
5419: In a ball of radius $r$ there are no more than $C(r^{2}/L+r)$ points
5420: not satisfying this condition.
5421: \item If $d(v,X)>2$ for any point $X$ of the form $\{ Ln\}\times\{ Lm\}$
5422: ($n,m\in\{0,\dotsc,\linebreak[4]M\}$), then $(\Delta f)(v)\leq C||f||_{2,\infty}$.
5423: In a ball of radius $r$ there are no more than $C(r^{2}/L^{2}+1)$
5424: points not satisfying this condition.
5425: \item For any $v$, $(\Delta f)(v)\leq C||f||_{1,\infty}$.
5426: \end{enumerate}
5427: Which are easy to verify, and as in the three dimensional case, the
5428: definition of $\omega$ at the stitches is used only for \ref{enu:plnpnt}.
5429: The equivalent of lemma \ref{lem:i2iiiiso} is proved in the same
5430: way (indeed, it is simpler as the construction of $f$ and $g$ is
5431: easier; and since the estimate $a(v,w)\approx\log|v-w|$ means there
5432: is no reason to divide into shells of size $2^{s}$ as in the three
5433: dimensional case). This shows that $\mathbb{Z}^{2}$ has a scaling
5434: limit, and concludes theorem \ref{thm:Z3scal}.
5435:
5436: \begin{conjecture*}
5437: Any non-trivial stitching of $\mathbb{Z}^{d}$ and $2\mathbb{Z}^{d}$
5438: is an isotropic graph.
5439: \end{conjecture*}
5440: In other words, while condition \ref{enu:plnpnt} obviously does not
5441: hold unless we insert weights in a manner similar to (\ref{eq:defomeclose}),
5442: the conjecture states that the graph $G(L,M,\xi)$ would be isotropic
5443: even if we, for example, just connect every vertex of type $2$ to
5444: the nearest vertex of type $1$ and give the edge weight $18\frac{7}{10}$.
5445: Be forewarned that the weighting of $2\mathbb{Z}^{d}$ by $2$ (in
5446: the three dimensional case) is very much needed. A simple resistance
5447: calculation would show that, for example for $\xi\equiv1$ on $[0,M/2]\times[0,M-1]^{2}$
5448: and $2$ otherwise, the graph $G(L,M,\xi)$ can never be isotropic
5449: (no matter what you put in the connecting layer) unless the edges
5450: of length $2$ are weighted by $2$.
5451:
5452:
5453: \subsection{\label{sub:Invariance}Invariance}
5454:
5455: The definition of the scaling limit immediately implies that it is
5456: invariant to multiplication by $2$, meaning that if $\mu(\mathcal{D},a)$
5457: is the scaling limit of loop-erased random walk on $\mathbb{Z}^{3}\cap2^{n}\mathcal{D}$
5458: starting from $2^{n}a$ then $\mu(2\mathcal{D},2a)=2\mu(\mathcal{D},a)$
5459: where the notation {}``$2\mu$'' stands for a stretching of $\mu$
5460: by $2$ in the natural way.
5461:
5462: In this section we give a few examples of additional invariances $\mu$
5463: satisfies. The first example is multiplication by $3$. In other words,
5464: we want to show that $\mu(3\mathcal{D},3a)=3\mu(\mathcal{D},a)$.
5465: A moments reflection shows that this will follow if we show that $\mathbb{Z}^{3}$
5466: and $3\mathbb{Z}^{3}$ have the same scaling limit, which would follow
5467: by theorem \ref{thm:univ} if we show they have an isotropic interpolation.
5468: We follow the same guidelines as in the previous section. Define $G(L,M,\xi)$
5469: as $\mathbb{Z}^{3}$ and $3\mathbb{Z}^{3}$ with weight $3$ on the
5470: internals of the cubes. In the stitches we let $x$ be the closest
5471: vertex of type $1$ to $w$ and then define \begin{align*}
5472: \omega(v,w) & =\frac{1}{|x-w|}\begin{cases}
5473: 1 & v=x\\
5474: 2/3 & |v-x|=1\textrm{ or }\sqrt{2}\\
5475: 1/3 & |v-x|=2\textrm{ or }\sqrt{8}.\end{cases}\end{align*}
5476: As in the previous section, a calculation verifies \ref{enu:plnpnt}
5477: which implies lemma \ref{lem:i2iiiiso}, isotropic interpolation and,
5478: with theorem \ref{thm:univ}, the invariance of $\mu$.
5479:
5480: Since (as is well known) the numbers $2^{k}3^{-n}$ are dense in $\mathbb{R}$,
5481: this shows that $\mu$ is in fact invariant to a dense set of multiplications.
5482: We shall now sketch a simple continuity argument which shows that
5483: $\mu$ is in fact invariant to all multiplications. Let $\alpha>1$
5484: and examine the situation of theorem \ref{thm:interp}, i.e.~we have
5485: the graphs $\mathbb{Z}^{3}$ and $\alpha\mathbb{Z}^{3}$, some $a\in\mathcal{E}\cap\mathcal{D}$
5486: and some $s>0$. We use theorem \ref{thm:interp} for $\mathbb{Z}^{3}$
5487: and $2\mathbb{Z}^{3}$ repeatedly and rescale (as in the proof of
5488: lemma \ref{lem:montrik}) to get that for any $k$,\[
5489: \mathbb{P}_{\mathbb{Z}^{3}}^{2^{k}sa}(\LE(R)\subset2^{k}s(\mathcal{E}+B(0,Cs^{-c})))>\mathbb{P}_{\mathbb{Z}^{3}}^{sa}(\LE(R)\subset s\mathcal{E})-Cs^{-c}.\]
5490: Next we use theorem \ref{thm:interp} for $\mathbb{Z}^{3}$ and $3\mathbb{Z}^{3}$
5491: repeatedly and rescale in the opposite direction to get that, as long
5492: as $2^{k}3^{-n}>1$,\[
5493: \mathbb{P}_{\mathbb{Z}^{3}}^{2^{k}3^{-n}sa}(\LE(R)\subset2^{k}3^{-n}s(\mathcal{E}+Cs^{-c}B(0,1)))>\mathbb{P}_{\mathbb{Z}^{3}}^{sa}(\LE(R)\subset s\mathcal{E})-Cs^{-c}.\]
5494: Notice that the various constants do not depend on $k$ and $n$ ---
5495: in fact we can get as close as we want to $\alpha$, until $2^{k}3^{-n}\mathcal{D}\cap\mathbb{Z}^{3}=\alpha\mathcal{D}\cap\mathbb{Z}^{3}$
5496: (and ditto for $\mathcal{E}$), with no price to pay. There might
5497: be a problem that the point of $2^{k}3^{-n}\mathbb{Z}^{3}$ closest
5498: to $a$ might be different by $C$ from the point of $\alpha\mathcal{D}\cap\mathbb{Z}^{3}$
5499: closest to $a$, but we have lemma \ref{lem:contsp} to show us that
5500: this affects the relevant probabilities by no more than $Cs^{-c}$
5501: as well. Thus the conclusion of theorem \ref{thm:interp} holds for
5502: $\mathbb{Z}^{3}$ and $\alpha\mathbb{Z}^{3}$ and therefore so does
5503: theorem \ref{thm:univ}.
5504:
5505: We next move to rotations. Examine the lattice $G$ in $\mathbb{R}^{3}$
5506: spanned by the vectors \[
5507: (4,3,0),(3,-4,0),(0,0,5).\]
5508: Since the vectors are orthogonal and of equal length, condition \ref{enu:regpnt}
5509: will continue to hold. By now the reader should have only technical
5510: difficulties in producing a stitching of $G$ and, for example, $25\mathbb{Z}^{3}$
5511: with the weights being $5$, which will satisfy condition \ref{enu:plnpnt}:
5512: in fact our lattice is invariant to translations by $25\mathbb{Z}^{3}$
5513: which reduces the verification of \ref{enu:plnpnt} to a small number
5514: of cases. Therefore $G$ and $25\mathbb{Z}^{3}$ have an isotropic
5515: interpolation and the same scaling limit. This shows that the scaling
5516: limit is invariant to rotations by $\arctan\frac{3}{4}$ around the
5517: $z$ axis. As is well known and not difficult to see, $\arctan\frac{3}{4}/\pi$
5518: is irrational so the semigroup created by this rotation is dense,
5519: and a similar continuity argument can be used to show that $\mu$
5520: is invariant to any rotation around the $z$ axis. Since $\mu$ is
5521: clearly invariant to a change of coordinates, and since any rotation
5522: is a combinations of three rotations around the axes, we see that
5523: $\mu$ is invariant to all rotations.
5524:
5525: \begin{thebibliography}{DFGW89}
5526: \bibitem[AB99]{AB99}Michael Aizenman and Almut Burchard, \emph{Hölder regularity and dimension
5527: bounds for random curves}, Duke Mathematical Journal \textbf{99:3},
5528: (1999) 419--453.\hrlb{} \url{http://www.arxiv.org/abs/math.FA/9801027}
5529: \bibitem[BB99]{BB99}Martin T. Barlow and Richard F. Bass, \emph{Brownian motion and harmonic
5530: analysis on Sierpinski carpets}, Canadian Journal of Mathematics \textbf{54}
5531: (1999), 673--744.
5532: \bibitem[B95]{B95}Richard F. Bass, \emph{Probabilistic techniques in analysis}, Probability
5533: and its Applications (New York), Springer-Verlag, New York, 1995.
5534: \bibitem[BPP95]{BPP95}Itai Benjamini, Robin Pemantle and Yuval Peres, \emph{Martin capacity
5535: for Markov chains}, The Annals of Probability \textbf{23:3} (1995),
5536: 1332--1346.
5537: \bibitem[BB]{BB}Noam Berger and Marek Biskup, \emph{Quenched invariance principle
5538: for simple random walk on percolation clusters}.\hrlb{} \url{http://arxiv.org/abs/math.PR/0503576}
5539: \bibitem[BI03]{BI03a}David C. Brydges and John Z. Imbrie, \emph{Branched polymers and dimensional
5540: reduction}, Annals of Mathematics (2) \textbf{158:3} (2003), 1019--1039.\hrlb{}
5541: \url{http://www.arxiv.org/abs/math-ph/0107005}
5542: \bibitem[BS85]{BS85}Davic C. Brydges and Thomas Spencer, \emph{Self-avoiding walk in $5$
5543: or more dimensions}, Communications in Mathematical Physics \textbf{97:1-2}
5544: (1985), 125--148.\hrlb{} \toolong{http://projecteuclid.org/Dienst/UI/1.0/Summarize/euclid.cmp/1103941982}{http://projecteuclid.org/Dienst/UI/1.0/Summarize/...}
5545: \bibitem[BL90a]{BL90a}Krzystof Burdzy and Gregory F. Lawler, \emph{Non-intersection exponents
5546: for Brownian paths, part I. existence and an invariance principle},
5547: Probability Theory and Related Fields \textbf{84} (1990), 393--410.
5548: \bibitem[BL90b]{BL90b}Krzystof Burdzy and Gregory F. Lawler, \emph{Non-intersection exponents
5549: for Brownian paths, part II. estimates and applications to a random
5550: fractal}, Annals of Probability \textbf{18} (1990), 981--1009.
5551: \bibitem[CM91]{CM91}Michael Cranston and Thomas Mountford, \emph{An extension of a result
5552: of Burdzy and Lawler}, Probability Theory and Related Fields \textbf{89}
5553: (1991), 487--502.
5554: \bibitem[DS98]{DS98}Eric Debez and Gordon Slade, \emph{The scaling limit of lattice trees
5555: in high dimensions}, Communications in Mathematical Physics \textbf{193:1}
5556: (1998), 69--104.
5557: \bibitem[D97]{D97}Thierry Delmotte, \emph{Inégalité de Harnack elliptique sur les graphes},
5558: Colloquium Mathematicum \textbf{72:1} (1997), 19--37.
5559: \bibitem[D99]{D99}Thierry Delmotte, \emph{Parabolic Harnack inequality and estimates
5560: of Markov chains on graphs}, Revista Matemática Iberoamericana \textbf{15:1}
5561: (1999), 181--232.
5562: \bibitem[DFGW89]{DFGW89}Anna De Masi, Pablo A. Ferrari, Sheldon Goldstein and William David
5563: Wick, \emph{An invariance principle for reversible Markov processes.
5564: Applications to random motions in random environments}, Journal of
5565: Statistical Physics \textbf{55:3--4} (1989), 787--855.
5566: \bibitem[D92]{D92}Bertrand Duplantier, \emph{Loop-erased self-avoiding walks in two
5567: dimensions: exact critical exponents and winding numbers}, Physica
5568: A \textbf{191} (1992), 516--522.
5569: \bibitem[DK88]{DK88}Bertrand Duplantier and Kyung-Hoon Kwon, \emph{Conformal Invariance
5570: and Intersections of Random Walks}, Physical Review Letters \textbf{61:22}
5571: (1988), 2514--2517.
5572: \bibitem[DEK50]{DEK50}Aryeh Dvoretzky, Paul Erdös and Shizuo Kakutani, \emph{Double points
5573: of paths of Brownian motion in $n$-space}, Acta Scientiarum Mathematicarum
5574: (Szeged) \textbf{12} (1950), 75--81.
5575: \bibitem[GSC05]{GSC05}Alexander Grigor'yan and Laurent Saloff-Coste, \emph{Stability results
5576: for Harnack inequalities}, Annales de l'Institut Fourier (Grenoble)
5577: \textbf{55:3} (2005), 825--890.\hrlb{} \url{http://www.ma.ic.ac.uk/~grigor/vc1eps.pdf}
5578: \bibitem[G99]{G99}Geoffrey Grimmett, \emph{Percolation}, second edition, \foreignlanguage{german}{Grundlehren
5579: der Mathematischen Wissenschaften} 321, Springer-Verlag, 1999.
5580: \bibitem[G81]{G81}Misha Gromov, \emph{Hyperbolic manifolds, groups and actions}, in:
5581: \emph{Riemann surfaces and related topics: Proceedings of the 1978
5582: Stony Brook Conference}, Annals of Mathematics Studies 97, Princeton
5583: Univ. Press, 1981, 183--213.
5584: \bibitem[GB90]{GB90}A. J. Guttmann and R. J. Bursill, \emph{Critical exponents for the
5585: loop erased self-avoiding walk by Monte Carlo methods}, Journal of
5586: Statistical Physics \textbf{59:1/2} (1990), 1--9.
5587: \bibitem[HvdHS03]{HvdHS03}Takashi Hara, Remco van der Hofstad and Gordon Slade, \emph{Critical
5588: two-point functions and the lace expansion for spread-out high-dimensional
5589: percolation and related models}, Annals of Probability \textbf{31:1}
5590: (2003), 349--408.\hrlb{} \toolong{http://www.math.ubc.ca/people/faculty/slade/hara-hofstad-slade.pdf}{http://www.math.ubc.ca/people/faculty/slade/...}
5591:
5592: \bibitem[HS92]{HS92}Takashi Hara and Gordon Slade, \emph{Self-avoiding walk in five or
5593: more dimensions. I. The critical behaviour}, Communications in Mathematical
5594: Physics \textbf{147:1} (1992), 101--136.\hrlb{} \toolong{http://projecteuclid.org/Dienst/UI/1.0/Summarize/euclid.cmp/1104250528}{http://projecteuclid.org/Dienst/UI/1.0/Summarize/...}
5595: \bibitem[HSC93]{HS93}Waldemar Hebisch and Laurent Saloff-Coste, \emph{Gaussian estimates
5596: for Markov chains and random walks on groups}, Annals of Probability
5597: \textbf{21:2} (1993), 673--709.
5598: \bibitem[HSC01]{HS01}Waldemar Hebisch and Laurent Saloff-Coste, \emph{On the relation between
5599: elliptic and parabolic Harnack inequalities}, Annales de l'Institut
5600: Fourier (Grenoble) \textbf{51:5} (2001), 1437--1481.\hrlb{} \url{http://www.math.uni.wroc.pl/~hebisch/harnack.ps }
5601: \bibitem[vdHS03]{vdHS03}Remco van der Hofstad and Gordon Slade, \emph{Convergence of critical
5602: oriented percolation to super-Brownian motion above $4+1$ dimensions},
5603: Annales de l'Institut Henri Poincaré. Probabilités et Statistique
5604: \textbf{39:3} (2003), 413--485.\hrlb{} \url{http://www.math.ubc.ca/people/faculty/slade/op.pdf }
5605: \bibitem[HS97]{HS97}Ilkka Holopainen and Paolo M. Soardi, A \emph{strong Liouville theorem
5606: for $p$-harmonic functions on graphs}, Annales Academiae Scientiarum
5607: Fennicae Mathematica \textbf{22} (1997), 205--226.\hrlb{} \url{http://www.emis.de/journals/AASF/Vol22/holopain.html}
5608: \bibitem[J86]{J86}David Jerison, \emph{The Poincaré inequality for vector fields satisfying
5609: Hörmander's condition}, Duke Mathematical Journal \textbf{53:2} (1986),
5610: 503--523.
5611: \bibitem[K85]{K85}M. Kanai, \emph{Rough isometries and combinatorial approximations
5612: of geometries of non-compact Riemannian manifolds}, Journal of the
5613: Mathematical Society of Japan \textbf{37} (1985), 391--413.
5614: \bibitem[K00a]{K00a}Richard Kenyon, \emph{The asymptotic distribution of the discrete
5615: Laplacian}, Acta Mathematica (2) \textbf{185:2} (2000), 239--286.\hrlb{}
5616: \url{http://www.arxiv.org/abs/math-ph/0011042}
5617: \bibitem[K00b]{K00b}Richard Kenyon, \emph{Long range properties of spanning trees}, Journal
5618: of Mathematical Physics \textbf{41:3} (2000) 1338--1363.\hrlb{} \url{http://www.math.ubc.ca/~kenyon/papers/long.ps.Z}
5619: \bibitem[K87]{K87}Harry Kesten, \emph{Hitting probabilities of random walks on $\mathbb{Z}^{d}$},
5620: Stochastic Processes and their Applications \textbf{25} (1987), 165--184.
5621: \bibitem[K]{K}Gady Kozma, \emph{Scaling limit of loop erased random walk --- a naive
5622: approach}.\hrlb{} \url{http://arXiv.org/abs/math.PR/0212338}
5623: \bibitem[KS04]{KS04}Gady Kozma and Ehud Schreiber, \emph{An asymptotic expansion for the
5624: discrete harmonic potential}, Electronic Journal of Probability \textbf{9:1}
5625: (2004), 1--17.\hrlb{} \url{http://www.arxiv.org/abs/math.PR/0212156}
5626: \bibitem[L80]{L80}Gregory F. Lawler, \emph{A self-avoiding random walk}, Duke Mathematical
5627: Journal \textbf{47:3} (1980), 655--693.
5628: \bibitem[L87]{L87}Gregory F. Lawler, \emph{Loop-erased self-avoiding random walk and
5629: the Laplacian random walk}, Journal of Physics A \textbf{20:13} (1987),
5630: 4565--8.
5631: \bibitem[L89]{L89}Gregory F. Lawler, \emph{Intersections of random walks with random
5632: sets}, Israel Journal of Mathematics \textbf{65} (1989), 113--132.
5633: \bibitem[L91]{L91}Gregory F. Lawler, \emph{Intersections of random walks}, Birkh\"auser
5634: Boston, 1991.
5635: \bibitem[L95]{L95}Gregory F. Lawler, \emph{The logarithmic correction for loop-erased
5636: walk in four dimensions}, Proceedings of the conference in honor of
5637: Jean-Pierre Kahane (Orsay, 1993), special issue of Journal of Fourier
5638: Analysis and Applications (1995), 347--362.
5639: \bibitem[L96a]{L96a}Gregory F. Lawler, \emph{Hausdorff dimension of cut points for Brownian
5640: motion}, Electronic Journal of Probability \textbf{1:2} (1996).\hrlb{} \toolong{http://www.math.washington.edu/~ejpecp/viewarticle.php?id=1204&layout=abstract}{http://www.math.washington.edu/~ejpecp/...}
5641: \bibitem[L96b]{L96b}Gregory F. Lawler, \emph{Cut times for simple random walk}, Electronic
5642: Journal of Probability \textbf{1:13} (1996).\hrlb{} \toolong{http://www.math.washington.edu/~ejpecp/viewarticle.php?id=1215&layout=abstract}{http://www.math.washington.edu/~ejpecp/...}
5643: \bibitem[L98]{L98}Gregory F. Lawler, \emph{Strict concavity of the intersection exponent
5644: for Brownian motion in two and three dimensions}, Mathematical Physics
5645: Electronic Journal \textbf{4:5} (1998).\hrlb{} \url{http://www.ma.utexas.edu/mpej/Vol/4/5.ps}
5646: \bibitem[L99]{L99}Gregory F. Lawler, \emph{Loop-erased random walk}, in: \emph{Perplexing
5647: problems in probability}, Progress in Probability 44, Birkh\"auser
5648: Boston, 1999, 197--217.
5649: \bibitem[LP00]{LP00}Gregory F. Lawler and Emily E. Puckette, \emph{The intersection exponent
5650: for simple random walk}, Combinatorics, Probability and Computing
5651: \textbf{9:5} (2000), 441--464.
5652: \bibitem[LSW01a]{LSW01a}Gregory F. Lawler, Oded Schramm and Wendelin Werner, \emph{Values
5653: of Brownian intersection exponents. I. Half-plane exponents}, Acta
5654: Mathematica (2) \textbf{187:2} (2001), 237--273.\hrlb{} \url{http://arxiv.org/abs/math.PR/9911084}
5655: \bibitem[LSW01b]{LSW01b}Gregory F. Lawler, Oded Schramm and Wendelin Werner, \emph{Values
5656: of Brownian intersection exponents II: Plane exponents}, Acta Mathematica
5657: (2) \textbf{187:2} (2001), 275--308.\hrlb{} \url{http://arxiv.org/abs/math.PR/0003156}
5658: \bibitem[LSW02a]{LSW02a}Gregory F. Lawler, Oded Schramm and Wendelin Werner, \emph{Analyticity
5659: of intersection exponents for planar Brownian motion}, Acta Mathematica
5660: (2) \textbf{189:2} (2002), 179--201.\hrlb{} \url{http://arxiv.org/abs/math.PR/0005295}
5661: \bibitem[LSW02b]{LSW02b}Gregory F. Lawler, Oded Schramm and Wendelin Werner, \emph{Sharp estimates
5662: for Brownian non-intersection probabilities}, in: \emph{In and out
5663: of equilibrium (Mambucaba, 2000)}, Progress in Probability 51, Birkhäuser
5664: Boston, Boston, MA, 2002, 113--131.\hrlb{} \url{http://arxiv.org/abs/math.PR/0101247}
5665: \bibitem[LSW04a]{LSW04}Gregory F. Lawler, Oded Schramm and Wendelin Werner, \emph{Conformal
5666: invariance of planar loop-erased random walks and uniform spanning
5667: trees}, Annals of Probability \textbf{32:1B} (2004), 939--995.\hrlb{}
5668: \url{http://arxiv.org/abs/math.PR/0112234}
5669: \bibitem[LSW04b]{LSW}Gregory F. Lawler, Oded Schramm and Wendelin Werner, \emph{On the
5670: scaling limit of planar self-avoiding walk}, in: \emph{Fractal geometry
5671: and applications: a jubilee of Beno\^it Mandelbrot, Part 2}, Proceedings
5672: of Symposia in Pure Mathematics 72, Part 2, Amer. Math. Soc., Providence,
5673: RI, 2004, 339--364.\hrlb{} \url{http://arxiv.org/abs/math.PR/0204277}
5674: \bibitem[M92]{M92}S. N. Majumdar, \emph{Exact fractal dimension of the loop-erased self-avoiding
5675: walk in two dimensions}, Physical Review Letters \textbf{68:15} (1992),
5676: 2329--2331.
5677: \bibitem[MP]{MP}Pierre Mathieu and A. L. Piatnitski, \emph{Quenched invariance principle
5678: for random walks on percolation clusters}.\hrlb{} \url{http://arxiv.org/abs/math.PR/0505672}
5679: \bibitem[NRS80]{NRS80}B. Nienhuis, E. K. Riedel and M. Schick, \emph{Magnetic exponents
5680: of the two-dimensional $q$-state Potts model}, Journal of Physics
5681: A \textbf{13} (1980), L189--L192.
5682: \bibitem[dN79]{dN79}M. P. M. den Nijs, \emph{A relation between the temperature exponents
5683: of the eight-vertex and $q$-state Potts model}, Journal of Physics
5684: A \textbf{12} (1979), 1857--1868.
5685: \bibitem[NY95]{NY95}Bao Gia Nguyen and Wei-Shih Yang, \emph{Gaussian limit for critical
5686: oriented percolation in high dimensions}, Journal of Statistical Physics
5687: \textbf{78:3--4} (1995), 841--876.
5688: \bibitem[P91]{P91}Robin Pemantle, \emph{Choosing a spanning tree for the integer lattice
5689: uniformly}, Annals of Probability \textbf{19:4} (1991), 1559--1574.
5690: \bibitem[PSC]{PSC99}Christophe Pittet and Laurent Saloff-Coste, \emph{A survey on the
5691: relationships between volume growth, isoperimetry, and the behavior
5692: of simple random walk on Cayley graphs, with examples}.\hrlb{} \url{http://www.math.cornell.edu/~lsc/surv.ps.gz}
5693: \bibitem[R73]{R73}Daniel Richardson, \emph{Random growth in a tessellation}, Mathematical
5694: Proceedings of the Cambridge Philosophical Society \textbf{74} (1973),
5695: 515--528.
5696: \bibitem[RW94]{RW94}L. C. G. Rogers and David Williams, \emph{Diffusions, Markov processes
5697: and martingales,} second edition, Wiley series in probability and
5698: mathematical statistics, John Wiley \& Sons, Ltd., Chichester 1994.
5699: \bibitem[S00]{S00}Oded Schramm, \emph{Scaling limits of loop-erased random walks and
5700: uniform spanning trees}, Israeli Journal of Mathematics \textbf{118}
5701: (2000), 221--288.\hrlb{} \url{http://arxiv.org/abs/math.PR/9904022}
5702: \bibitem[SS]{SS}Oded Schramm and Scott Sheffield, \emph{The harmonic explorer and
5703: its convergence to} SLE(4).\hrlb{} \url{http://arxiv.org/abs/math.PR/0310210}
5704: \bibitem[SS04]{SS04}Vladas Sidoravicius and Alain-Sol Sznitman, \emph{Quenched invariance
5705: principles for walks on clusters of percolation or among random conductances},
5706: Probability Theory and Related Fields \textbf{129:2} (2004), 219--244.
5707: \bibitem[S99]{S99}Gordon Slade, \emph{Lattice trees, percolation and super-Brownian
5708: motion}, in: \emph{Perplexing problems in probability}, 35--51, Progress
5709: in Probability 44, Birkhäuser Boston, Boston, MA, 1999.
5710: \bibitem[S01]{S01}Stanislav Smirnov, \emph{Critical Percolation in the plane}.\hrlb{}
5711: \url{http://www.math.kth.se/~stas/papers/percol.ps}
5712: \bibitem[SW01]{SW01}Stanislav Smirnov and Wendelin Werner, \emph{Critical exponents for
5713: two-dimensional percolation}, Mathematical Research Letters \textbf{8:5-6}
5714: (2001), 729--744.\hrlb{} \url{http://arxiv.org/abs/math.PR/0109120}
5715: \bibitem[S94]{S94}Paolo M. Soardi, \emph{Potential theory on infinite networks}, Lecture
5716: Notes in Mathematics 1590, Springer-Verlag, Berlin, 1994.
5717: \bibitem[V85]{V85}Nicholas Th. Varopoulos, \emph{Isoperimetric inequalities and Markov
5718: chains}, Journal of Functional Analysis \textbf{63} (1985), 215--239.
5719: \bibitem[W96]{W96}David B. Wilson, \emph{Generating random spanning trees more quickly
5720: than the cover time}, Twenty-Eighth Annual ACM symposium on Theory
5721: of Computing, 293-303.\hrlb{} \url{http://research.microsoft.com/~dbwilson/ja/tau.ps}\end{thebibliography}
5722:
5723: \end{document}
5724: