WorksMathematics
Anchored isoperimetry in percolation
Abstract
We show that, for any infinite transitive graph , the infinite cluster after supercritical percolation almost surely inherits the anchored isoperimetry of . Moreover, the probability of violating the corresponding isoperimetric inequality decays exponentially in the isoperimetric profile of the host graph. This was previously known only for nonamenable transitive graphs, , and very recently for transitive graphs of polynomial growth. Our proof combines the thinning-and-sprinkling technique with a novel counting algorithm.
Introduction
Percolation provides a natural way to probe the geometry of an infinite graph : randomly delete its edges and ask which structural properties survive. In the supercritical regime, the resulting random graph contains infinite connected components. A fundamental question is to what extent these clusters retain the large-scale geometric properties of the original graph. One particularly important aspect of this question is isoperimetry, which measures how the boundary of a vertex set grows with its size.
More precisely, let be an infinite, connected, locally finite, vertex-transitive graph of degree . In Bernoulli bond percolation, each edge of is retained independently with probability . This produces a random subgraph with distribution , whose connected components are called clusters. The critical probability is defined as
The regimes and are called subcritical and supercritical, respectively. Percolation was introduced in 1957 by Broadbent and Hammersley [2], and was initially studied mainly on the hypercubic lattice . An important step towards a more general understanding was made by Benjamini and Schramm [4], who initiated the systematic study of percolation on arbitrary infinite transitive graphs in 1996.
An important quantitative measure of the global geometry of a graph is its isoperimetric profile. For a finite set , let denote the set of edges of with exactly one endpoint in . The isoperimetric profile of is defined by
For example, when is nonamenable, whereas for . In the study of percolation, one might naturally hope that infinite supercritical clusters inherit the isoperimetric profile of , up to multiplicative constants. However, such inequalities cannot hold uniformly over all finite subsets of a cluster. Indeed, when , infinite clusters contain arbitrarily long bare paths, and taking a long subpath of their internal vertices yields a large connected set with only two boundary edges (see, e.g., [11, 13]). The natural way to overcome this difficulty is to consider an anchored version of isoperimetry, where we restrict attention to connected sets containing a fixed vertex. Given an infinite connected graph and a vertex , its anchored isoperimetric profile is defined by
Let denote the cluster containing in . When and is infinite, it is natural to ask whether its anchored isoperimetric profile is comparable to that of . More precisely, for infinite connected graphs containing , we say that has anchored isoperimetry with respect to if
We say that has anchored isoperimetry if, for every , the cluster almost surely has anchored isoperimetry with respect to , conditionally on being infinite. When is nonamenable, we have , and anchored isoperimetry of with respect to simply means that there exists such that every sufficiently large finite connected set containing satisfies . This property is called anchored expansion. Beyond their geometric significance, these notions have important consequences for random walks and phase transitions. In particular, anchored expansion implies positive speed of the simple random walk on bounded-degree graphs [19]; see also [10, 11, 12].
The notion of anchored expansion was introduced by Benjamini, Lyons, and Schramm [3], who asked whether it is preserved under supercritical percolation. Chen and Peres [7], with an appendix by Pete, established anchored expansion for percolation parameters sufficiently close to one. Subsequently, Pete [17], using ideas from that appendix, proved the corresponding anchored isoperimetric inequality for supercritical percolation on . The question for nonamenable transitive graphs was eventually resolved for every by Hermon and Hutchcroft [11], who also formulated the general anchored isoperimetry problem for arbitrary transitive graphs (Conjecture 5.4 in [11]). More recently, Martineau and Panagiotis [16] established anchored isoperimetry for transitive graphs of polynomial growth.
Our main result settles this problem for all infinite connected transitive graphs. Moreover, it provides a quantitative strengthening: the probability that an infinite cluster violates the desired anchored isoperimetric inequality at scale decays exponentially in .
Theorem 1. Fix any infinite connected transitive graph with a vertex . For every , there is a constant such that, for every ,
Previous approaches to anchored isoperimetry have relied on supercritical sharpness, which provides an upper bound on the probability that a vertex belongs to a large but finite cluster. Very recently, supercritical sharpness was established for all infinite transitive graphs [8]. More precisely, it was shown that, for every , there exists such that
for every . This result naturally suggests deriving anchored isoperimetry directly from the decay of finite cluster sizes, as in the previously known cases [17, 11, 16]. However, the general approach to converting such tail bounds into anchored isoperimetry incurs a logarithmic loss in (see [13 Proposition 3.1]), and therefore does not yield the desired bound for arbitrary transitive graphs.
Our proof takes a different route. It combines the thinning-and-sprinkling method, developed in recent papers [9, 1 CDLS26, 6], with a novel counting algorithm. The key idea is to count the possible ways in which connected sets can touch infinite clusters, rather than counting the sets themselves. This ‘reduction in entropy’ allows us to avoid the logarithmic loss and obtain a short, unified proof of anchored isoperimetry for all transitive graphs. Moreover, the argument uses transitivity only through a probabilistic estimate from [8 Proposition 2.1], without requiring any detailed structural information about the underlying graph.
The rest of this paper is organised as follows. We present a high-level overview of the proof in § 2, and give the full details in § 3.
Proof outline
Rather than invoking supercritical sharpness directly, our proof relies on the following key estimate from [8], which concerns the number of edges leaving a finite set that can connect to infinity without returning to the set.
For a finite set , let denote the minimum size of an edge cutset separating from infinity. Given , we also write
The following is a convenient consequence of [8 Proposition 2.1].
Proposition 2.1. Fix an infinite connected transitive graph of degree . For every and every , there exists such that, for every finite and every positive integer ,
The relation between this notion of “touches” (that is, edges between and the infinite clusters in ) and anchored isoperimetry goes back to Pete [17]. Roughly speaking, Pete conjectured [18 Conjecture 12.32] that the probability under -percolation that the cluster is finite and satisfies should decay exponentially in (note that this is, morally, the opposite direction of Proposition 2.1 for a special random set ). This conjecture, combined with supercritical sharpness and the surgery approach of [7 Appendix A], would imply Theorem 1; indeed, this was the approach taken in [17, 16] for and for transitive graphs of polynomial growth. However, for arbitrary transitive graphs, the conjecture of Pete remains open; we take a different path instead.
Call a connected set boundary-minimiser if there is no finite connected superset with . (Formally, we will work inside a large finite ball, and this notion will be restricted to this ball.) Note that is nondecreasing under inclusion and therefore, for Theorem 1 to fail, there must be a boundary-minimiser set in containing with rather large boundary in but rather small boundary in .
To analyse this event, we will use a variant of the thinning-and-sprinkling method. Fix . Since we are interested in sets with few boundary edges in , requiring that also has no open boundary edges in costs, for a fixed -configuration, a factor of . We further restrict our attention to sets with many -touches; by Proposition 2.1 and FKG, this costs a factor of at most when . Moreover, it follows from [15] that the number of boundary-minimisers with at most boundary edges is at most . Our goal will thus be to show that it is very unlikely that contains a boundary-minimiser set (in ) with many -touches, no boundary edges in and too few boundary edges in .
Fix , and consider the following setup. First, reveal the entire -configuration and let be the union of its infinite clusters. If , the obstruction we seek is impossible. Otherwise, reveal the -edges within , and let be the component of in the resulting graph restricted to a large finite ball. Let be the vertices with a neighbour in . For a connected set containing , its fingerprint , together with the possible touches at , determines all its touch edges. We want to bound the probability that there is such a set with no -open boundary edges, and , where and for a sufficiently small constant . (We note that the upper bound provided here will hold not only for boundary-minimiser sets, but more generally.)
A natural first attempt is to union bound over all possible sets ; however, this approach leads to a severe overcounting. The first key observation is that we can “reduce the entropy” by counting instead the possible fingerprints , rather than the sets themselves (see Figure 1).
Figure 1. Two different connected sets (blue) and (green) meeting the same touches, drawn in red.
Indeed, once the -configuration has been revealed, each edge between and its complement becomes -open independently with probability . Thus, for every fixed choice of such edges, the probability that at most of them are in is
for some by Chernoff’s bound, provided . This point justifies why we require that : while we only “paid” for closing edges in the thinning stage of the argument, the “price” for opening at most edges among upon sprinkling is much more significant. Now, we would like to conclude by a union bound argument. Say that is -feasible if for some connected set containing with . It suffices to control the number of such fingerprints of each size : indeed, the number of edges from touching lies in the interval , and the sprinkling probability (1) decays exponentially with this number.
To do so, we develop a novel counting algorithm specialised for counting sets with small boundary. Roughly speaking, given a witness for a feasible set , the algorithm grows an explored region from . The region stays connected in and is maximised conditionally on keeping the edge cut separating it from of minimum size. Then, the algorithm examines a prescribed edge leaving that side. It writes a 1 and enlarges the region if the other endpoint belongs to ; otherwise it writes a 0 and deletes the edge. The crucial feature is that the resulting binary word determines , though not necessarily itself (Lemma 3.4). Submodularity shows that the word has at most zeroes, and the estimate allows to bound its length by (Proposition 3.3). Counting these words gives the required entropy bound; combined with the exponential sprinkling estimate (1) and a union bound over the possible sizes , this completes the argument.
Anchored isoperimetry via thinning and sprinkling
Fix , choose , and consider the natural coupling . We write and for the corresponding edge boundaries, and use for the joint law of the coupled configurations. Choose sufficiently small (its precise choice will be determined by comparing the thinning and sprinkling estimates at the end of the proof). By Proposition 2.1, there exists such that whenever . Fix and set . Let be sufficiently small (with its precise value to be chosen later) and set .
Recall that is the event that there exists a -connected set containing with and . To work in a finite graph, fix , put , and let be the same event with the additional restriction . Since increases to as , it suffices to obtain a bound on which is independent of .
We compare with the (simpler to analyse) event that there exists a -connected set satisfying , , and . Thus, describes a set with few -open boundary edges, none of them -open, but many potential connections to infinity. We first show that thinning from to transforms a bad anchored set into a configuration satisfying each of these properties at a controlled probability cost.
Lemma 3.1. We have
Proof. Call a -connected set containing a -boundary-minimiser if every proper -connected superset of satisfies . In other words, cannot be enlarged within without increasing its boundary. Since is increasing under inclusion, whenever occurs, it has a -boundary-minimiser witnessing it. (Let us note here that this property is the reason to restrict our graph to the large ball .)
To exploit thinning, we count the number of such witnesses that also have many touches and whose boundaries become entirely -closed. We first bound from below their expected number, and then use a standard counting result for small edge cuts to turn this into a probability estimate.
Let denote this number, namely
Fix a deterministic set appearing in this sum, and reveal only the - and -states of the edges with at least one endpoint in . Conditional on this exposure and the connectivity of , the event that is a -boundary-minimiser with is increasing in the remaining -edges. Indeed, suppose that adding edges outside creates a connected superset whose boundary is no larger than that of . Let be the component containing in the subgraph induced by before these edges were added. Since no edge incident to has changed, . Moreover, every previously open edge leaving also leaves , since otherwise its other endpoint would belong to . Thus, the original boundary of is no larger than the new boundary of , contradicting the boundary-minimality of . Furthermore, the event is increasing in the -edges outside . Since the event depends only on these unrevealed edges, FKG and Proposition 2.1 give
Now, reveal the entire -configuration and thin each -open edge independently with retention probability . For every -boundary-minimiser with , the probability that its entire boundary becomes -closed is at least . This allows us to bound the right-hand side of (2) from below and deduce that
By summing over all such possible sets , and using that guarantees at least one such boundary-minimiser, we obtain .
It remains to relate to the probability of . Here we use the notion of important separators from [15]. An important edge separator between a vertex and a target set is an inclusion-minimal cut which cannot be replaced by a cut of at most the same size that leaves a strictly larger connected component containing . Marx [15 Theorem 2] proved that there are at most important separators of size at most . (See also Remark 3.5, where we recover this result using the algorithm we present in the next subsection.) We apply this result in the finite graph , taking as the ‘source’ and the exterior neighbourhood of as the ‘target’. By the maximality property of a -boundary-minimiser, its boundary is precisely such an important separator. Consequently, in every -configuration, there are at most possible -boundary-minimisers with boundary at most . Thus, pointwise. Furthermore, can be positive only if occurs. Hence . Combining this with the lower bound on proves the lemma. □
We now bound the probability of the same obstruction from the sprinkling viewpoint. The only additional input is Proposition 3.3, which counts the possible fingerprints of connected sets with small boundary on a prescribed set of marked vertices. Its proof is postponed to Subsection 3.1.
Lemma 3.2. There exists a constant such that
Proof. Reveal the entire -configuration, and let be the union of its infinite clusters. Next, reveal the -states of all edges with both endpoints in . Denote by the configuration revealed so far. If , then is impossible: a finite set containing cannot have -closed boundary. Otherwise, let be the component of in the subgraph of induced by , and mark the vertices .
Say that is -feasible if for some -connected set containing with . For such a set , write . Every witness to lies in , since its -boundary is closed. Hence is -feasible, , and at most of these edges are -open.
Conditioned on the revealed configuration , all edges between and are -closed, and they become -open independently with probability . Assuming , a Chernoff bound and the relation ensure that, for every fixed feasible with ,
We will now union-bound over the possible choices of a -feasible set , rather than over all possible witnesses . For this, we use the following counting proposition; its proof is delayed to the next subsection.
Proposition 3.3. Given a finite graph with maximum degree at most containing a vertex and a vertex set , the number of -feasible vertex subsets of of size is at most
We apply the proposition to the exposed graph and the marked set . Note that, by excluding from and including it automatically in , we avoid any restriction on whether has neighbours in . Notice that . Thus, implies , and for every relevant . Conditioning on , the proposition and a union bound give
Using that the binomial expression in the term corresponding to in the above sum is bounded from above by , and choosing sufficiently small with respect to and shows that the right-hand side of (3) is at most
for a suitably large constant . Since the bound is uniform in , averaging completes the proof.
Combining the two lemmas, and recalling that and , we choose sufficiently small with respect to , and sufficiently small compared with , so that
for some , and note that this bound holds uniformly in .
Finally, we choose in the exponent such that the factor can be absorbed into it. This is indeed possible: note that, for small , taking the constant in Theorem 1 sufficiently small produces a trivial bound on the event that . Letting and taking completes the proof.
The algorithm and proof of Proposition 3.3
In this section, we prove the aforementioned Proposition 3.3. To prove this statement, we develop an algorithm that encodes each -feasible set by a short binary word. Given a connected set witnessing the feasibility of , the algorithm produces a binary word from which can be uniquely recovered. The key properties of this encoding are that the word contains at most zeroes and has length at most . Counting such words then gives the desired bound on the number of feasible sets.
We now describe the algorithm and then establish these properties. In this description, parameters called dynamic are going to evolve during the algorithm; we denote these by bold letters and we use the symbol to update their values.
Algorithm 1. Input: a finite connected graph containing a vertex with an ordering of its edge set , an -connected vertex set containing , and a set of marked vertices . Initiate a dynamic graph , dynamic vertex sets , and , a dynamic edge set , and a dynamic integer .
Output: we construct a binary word initially set to .
Procedure: at every iteration, first check whether . If this is the case, terminate the algorithm and output the word . Otherwise, update to be the size of a minimum edge-cutset in between and . Next, update to be the connected vertex set in which:
contains but is disjoint from the set ,
satisfies ,
has maximum size given the previous two conditions, and
if more than one set satisfies the conditions above, is chosen so that is minimal with respect to the lexicographic ordering on the edge subsets of induced by .
Set and let be the first edge of in the ordering , with .
(a) If , append a 0 to , update and .
(b) If , append a 1 to and update .
(c) If , append a 1 to and update and .
Finally, proceed to the next iteration.

Figure 2. Initial state (left) and state after one iteration (right) of Algorithm 1. The fixed witness is shaded grey; squares indicate the marked vertices . The boundary of on the left is . Since , case (b) gives (coloured in teal) and becomes 1.
Observe that the algorithm is well-defined and terminates. Indeed, only edges with an endpoint outside are deleted, so any vertex remains connected to in ; hence, as long as the algorithm is still running, the next edge always exists. Moreover, each iteration either deletes an edge or enlarges . Let us note here that the choice of a largest boundary-minimising set is crucial for controlling the length of the word: case (b) forces the minimum-cut size to increase, while the possible decreases in cases (a) and (c) will be shown to be controlled by and , respectively.
We first establish that the output word determines .
Lemma 3.4. Given the input , , , and the output of the algorithm upon termination, the set can be uniquely determined.
Proof. Starting from the known initial state, we reconstruct the algorithm one bit at a time. The current values of , together with , determine , , and the next edge . If the next bit is 0, we follow case (a). If it is 1, we follow case (b) or (c), according to whether or . Thus, each bit uniquely determines how to update each of . After reading the entire word, the recovered set equals , as required. □
We emphasise that the purpose of the algorithm and the decoding Lemma 3.4 is not to recover the set itself, but only the set . Indeed, many different sets may have the same intersection with , and distinguishing between them would lead to a substantial overcounting. The key feature of our algorithm is precisely that it encodes only the information needed to recover . Thus, for fixed , , , , each output word determines at most one subset of .
Recall that a set is -feasible if there exists an -connected set containing such that and . Call any such a witness for . We can now prove Proposition 3.3.
Proof of Proposition 3.3. Fix an arbitrary ordering of the edges of . Given a -feasible set of size , choose an arbitrary witness and let be the binary word produced by Algorithm 1. By Lemma 3.4, different feasible sets give different words. Thus, it suffices to count the possible output words.
First, we show that contains at most zeroes. Crucially, note that an edge deleted when writing a zero need not belong to since the boundary-minimising set may contain vertices outside ; we therefore track the boundary of an enlarged witness . Let , , denote the values of , , before iteration , and let , be the values of , during that iteration. Set and for (the will play the role of the enlarged witness that we track). As the sets are nested, we have . Also, (since is either a single vertex in or empty), while is disjoint from . Hence, separates from , and the minimality of gives . By submodularity of the edge boundary (see e.g. [14 Section 1.1]),
The first inequality is strict in case (a), since the deleted edge crosses the boundary of : indeed, and . Thus, the boundary of the enlarged witness never increases and decreases by at least one whenever we write a zero. Since , the word has at most zeroes.
Next, we bound the length of . If , there is only one possible set , so the proposition is immediate. Thus assume . In case (a), deleting one edge decreases the minimum-cut value by at most one, so . In case (c), we have : adding the at most edges incident to the newly recorded marked vertex to any cut for the next step gives a cut for the preceding step. Crucially, in case (b), . Indeed, the graph and target set are unchanged, so the minimum-cut value cannot decrease; if it stayed unchanged, there would be an -connected set with minimum boundary containing , contradicting the maximality of .
Let be the number of iterations, so that . Recall that the submodularity argument gives throughout the algorithm. At intermediate stages, the boundary of , together with the edges incident to the vertices in , separates from . At termination, however, we have that . Since and , the boundary of itself separates from . Thus, letting be a minimum cut between these sets in , we obtain
Writing for the numbers of iterations of types (a), (b), (c), respectively, we have and . The preceding bounds imply , and consequently .
Finally, pad every output word with 1-bits to length . Since we are counting feasible sets of fixed size , the padded word still determines : we run the same procedure described above until exactly marked vertices have been recovered. Thus, padding preserves injectivity, and the number of possible words is at most
as desired. □
Remark 3.5. Our algorithm also allows us to recover the bound on important separators from [15 Theorem 2], used earlier in the proof.
Let be a locally finite graph containing a vertex , and let be a finite vertex set. Recall that a connected set containing is a -minimiser if there is no proper connected superset , , with . We apply Algorithm 1 with being the component of in and , replacing the stopping condition by the requirement that the minimum cut-value becomes zero.
For a -boundary-minimiser , induction shows that every value of throughout the algorithm is contained in , and every deleted edge belongs to . Indeed, assuming this for previously deleted edges, the minimum-cut property and submodularity imply that adding the chosen boundary-minimiser to cannot increase its boundary. By -boundary-minimality, this side must therefore be contained in , and any subsequently deleted edge belongs to . Consequently, upon termination, the component of in the remaining graph contains , is contained in , and has all its boundary edges among those deleted. By -boundary-minimality, this component must equal , which can therefore be recovered from the output word.
Moreover, since , case (c) never occurs. If , the word contains at most zeroes, since each zero deletes a distinct edge of . Each one increases the minimum cut-value by at least one, while each zero decreases it by at most one. Since the final cut-value is zero, the number of ones is at most the number of zeroes. Thus, the word has length at most . As before, padding shorter words with ones preserves injectivity. This gives at most such minimisers.
Acknowledgements The authors are grateful to Vincent Tassion and Philip Easo for fruitful discussions. BS was supported in part by SNSF grant 20021-228014.
Use of AI: AI models were used only to generate the figures from photos of hand-drawn ones, as well as to assist with proofreading of the paper.
References
References
- [1][ADLZ25] Michael Anastos, Sahar Diskin, Lyuben Lichev, and Maksim Zhukovskii. Diameter and mixing time of the giant component in the percolated hypercube. Preprint, arXiv:2510.13348, 2025.arxiv.org/abs/2510.13348
- [2][BH57] Simon R. Broadbent and John M. Hammersley. Percolation processes. I. Crystals and mazes. Proceedings of the Cambridge Philosophical Society, 53:629–641, 1957.
- [3]Itai Benjamini, Russell Lyons, and Oded Schramm. Percolation perturbations in potential theory and random walks. In Massimo A. Picardello and Wolfgang Woess, editors, Random Walks and Discrete Potential Theory, volume XXXIX of Symposia Mathematica, pages 56–84. Cambridge University Press, Cambridge, 1999. Proceedings of the conference held in Cortona, June 1997.DOI
- [4]Itai Benjamini and Oded Schramm. Percolation beyond ℤᵈ, many questions and a few answers. Electronic Communications in Probability, 1:71–82, 1996.
- [6]Micha Christoph, Zach Hunter, and Benny Sudakov. Thinning and sprinkling: from robust sampling to almost Hamiltonicity. Preprint, arXiv:2609.30165, 2026.arxiv.org/abs/2609.30165
- [7]Dayue Chen and Yuval Peres. Anchored expansion, percolation and speed. The Annals of Probability, 32(4):2978–2995, 2004. With an appendix by Gábor Pete.
- [8]Sahar Diskin, Philip Easo, Ritvik Ramanan Radhakrishnan, Benny Sudakov, and Vincent Tassion. Supercritical sharpness of percolation. Preprint, arXiv:2603.03257, 2026.arxiv.org/abs/2603.03257
- [9]Sahar Diskin and Michael Krivelevich. Expansion in supercritical random subgraphs of expanders and its consequences. Random Structures & Algorithms, 65(3):576–600, 2024.arxiv.org/abs/2205.04852
- [10]Thierry Delmotte and Clément Rau. Exit time for anchored expansion. Preprint, arXiv:0903.3892v2, 2015. Version 2, 2 July 2015; first posted 23 March 2009.
- [11]Jonathan Hermon and Tom Hutchcroft. Supercritical percolation on nonamenable graphs: isoperimetry, analyticity, and exponential decay of the cluster size distribution. Inventiones Mathematicae, 224(2):445–486, 2021.arxiv.org/abs/1904.10448
- [12]Olle Häggström, Roberto H. Schonmann, and Jeffrey E. Steif. The Ising model on diluted graphs and strong amenability. The Annals of Probability, 28(3):1111–1137, 2000.DOI
- [13]Tom Hutchcroft. Transience and anchored isoperimetric dimension of supercritical percolation clusters. Electronic Journal of Probability, 28:Paper No. 14, 2023. 15 pp.arxiv.org/abs/2207.05226
- [14]Andreas Krause and Daniel Golovin. Submodular function maximization. In Lucas Bordeaux, Youssef Hamadi, and Pushmeet Kohli, editors, Tractability: Practical Approaches to Hard Problems, chapter 3, pages 71–104. Cambridge University Press, Cambridge, 2014.
- [15]Dániel Marx. Important separators and parameterized algorithms. In Petr Kolman and Jan Kratochvíl, editors, Graph-Theoretic Concepts in Computer Science, volume 6986 of Lecture Notes in Computer Science, pages 5–10, Berlin, Heidelberg, 2011. Springer.DOI
- [16]Sébastien Martineau and Christoforos Panagiotis. Percolation on graphs of polynomial growth is local: analyticity, supercritical sharpness, isoperimetry. Preprint, arXiv:2511.01851v2, 2025. Version 2, revised 1 March 2026.arxiv.org/abs/2511.01851
- [17]Gábor Pete. A note on percolation on ℤᵈ: isoperimetric profile via exponential cluster repulsion. Electronic Communications in Probability, 13:377–392, 2008.arxiv.org/abs/math/0702474
- [18]Gábor Pete. Probability and geometry on groups. Lecture notes for a graduate course, https://math.bme.hu/~gabor/PGG.pdf, 2024. Version of 3 April 2024.
- [19]Bálint Virág. Anchored expansion and random walk. Geometric and Functional Analysis, 10(6):1588–1605, 2000.DOI
