The component graph of the uniform spanning forest: transitions in dimensions 9,10,11,…\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$9,10,11,\ldots $$\end{document}

We prove that the uniform spanning forests of Zd\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {Z}^d$$\end{document} and Zℓ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {Z}^{\ell }$$\end{document} have qualitatively different connectivity properties whenever ℓ>d≥4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ell >d \ge 4$$\end{document}. In particular, we consider the graph formed by contracting each tree of the uniform spanning forest down to a single vertex, which we call the component graph. We introduce the notion of ubiquitous subgraphs and show that the set of ubiquitous subgraphs of the component graph changes whenever the dimension changes and is above 8. To separate dimensions 5, 6, 7, and 8, we prove a similar result concerning ubiquitous subhypergraphs in the component hypergraph. Our result sharpens a theorem of Benjamini, Kesten, Peres, and Schramm, who proved that the diameter of the component graph increases by one every time the dimension increases by four.

the WUSF was given by Benjamini, Lyons, Peres, and Schramm [3], who proved that the WUSF of a graph is connected if and only if two independent random walks on G intersect infinitely often a.s. Extending Pemantle's result, Benjamini, Kesten, Peres, and Schramm [2] (henceforth referred to as BKPS) discovered the following surprising theorem.
Theorem (BKPS [2]) Let F be a sample of the USF of Z d . For each x, y ∈ Z d , let N (x, y) be the minimal number of edges that are not in F used by a path from x to y in Z d . Then almost surely.
In particular, this theorem shows that every two trees in the uniform spanning forest of Z d are adjacent almost surely if and only if d ≤ 8. Similar results have since been obtained for other models [4,18,[22][23][24]. The purpose of this paper is to show that, once d ≥ 5, the uniform spanning forest undergoes qualitative changes to its connectivity every time the dimension increases, rather than just every four dimensions.
In order to formulate such a theorem, we introduce the component graph of the uniform spanning forest. Let G be a graph and let ω be a subgraph of G. The component graph C 1 (ω) of ω is defined to be the simple graph that has the connected components of ω as its vertices, and has an edge between two connected components k 1 and k 2 of ω if and only if there exists an edge e of G that has one endpoint in k 1 and the other endpoint in k 2 . More generally, for each r ≥ 1, we define the distance r component graph C r (ω) to be the graph which has the components of ω as its vertices, and has an edge between two components k 1 and k 2 of ω if and only if there is path in G from k 1 to k 2 that has length at most r .
When formulated in terms of the component graph, the result of BKPS states that the diameter of C 1 (F) is almost surely (d − 4)/4 for every d ≥ 1. In particular, it implies that C 1 (F) is almost surely a single point for all 1 ≤ d ≤ 4 (as follows from Pemantle's theorem), and is almost surely a complete graph on a countably infinite number of vertices for all 5 ≤ d ≤ 8.
We now introduce the notion of ubiquitous subgraphs. We define a graph with boundary H = (∂ V , V • , E) = (∂ V (H ), V • (H ), E(H )) to be a graph H = (V , E) whose vertex set V is partitioned into two disjoint sets, V = ∂ V ∪ V • , which we call the boundary and interior vertices of H , such that ∂ V = ∅. Given a graph G, a graph with boundary H , and collection of distinct vertices (x u ) u∈∂ V of G indexed by the boundary vertices of H , we say that H is present at (x u ) u∈∂ V if there exists a collection of vertices (x u ) u∈V • of G indexed by the interior vertices of H such that x u ∼ x v or x u = x v for every u ∼ v in H . (Note that, in this definition, we do not require that x u and x v are not adjacent in G if u and v are not adjacent in H .) We say that H is faithfully present at (x u ) u∈∂ V if there exists a collection of distinct vertices (x u ) u∈V • of G, disjoint from (x u ) u∈∂ V , indexed by the interior vertices of H such that x u ∼ x v for every u ∼ v in H . In figures, we will use the convention that boundary vertices are white and interior vertices are black. We say that H is ubiquitous in G if it is present at every collection of distinct vertices (x u ) u∈∂ V in G, and that H is faithfully ubiquitous in G if it is faithfully present at every collection of distinct vertices (x u ) u∈∂ V in G.
For example, if H is a path of length n with the endpoints of the path as its boundary, then H is ubiquitous in a graph G if and only if G has diameter less than or equal to n. The same graph is faithfully ubiquitous in G if and only if every two vertices of G can be connected by a simple path of length exactly n. If H is a star with k leaves set to be in the boundary and the central vertex set to be in the interior, then H is ubiquitous in a graph G if and only if every k vertices of G share a common neighbour, and in this case H is also faithfully ubiquitous.
The main result of this paper is the following theorem. We say that a transitive graph G is d-dimensional if there exist positive constants c and C such that cn d ≤ |B(x, n)| ≤ Cn d for every vertex x of G and every n ≥ 1, where B(x, n) denotes the graph-distance ball of radius n around x in G. The WUSF and FUSF of any ddimensional transitive graph coincide [3], and we speak simply of the USF of G. Note that the geometry of a d-dimensional transitive graph may be very different from that of Z d . (Working at this level of generality does not add any substantial complications to the proof, however.) Theorem 1.1 Let G 1 and G 2 be transitive graphs of dimension d 1 and d 2 respectively, and let F 1 and F 2 be uniform spanning forests of G 1 and G 2 respectively. Then the following claims hold for every r 1 , r 2 ≥ 1: (1) (Universality and monotonicity.) If d 1 ≥ d 2 ≥ 9, then every finite graph with boundary that is ubiquitous in C r 1 (F 1 ) is also ubiquitous in C r 2 (F 2 ) almost surely. (2) (Distinguishability of different dimensions.) If d 1 > d 2 ≥ 9, then there exists a finite graph with boundary H such that H is almost surely ubiquitous in C r 2 (F 2 ) but not in C r 1 (F 1 ).

Moreover, the same result holds with 'ubiquitous' replaced by 'faithfully ubiquitous'.
In order to prove item (2) of Theorem 1.1, it will suffice to consider the case that H is a tree. In this case, the following theorem allows us to calculate the dimensions for which H is ubiquitous in the component graph of the uniform spanning forest. The corresponding result for general H is given in Theorem 1.4. Examples of trees that can be used to distinguish between different dimensions using Theorem 1.2 are given in Figs. 1 and 2. Note that (d − 4)/(d − 8) is a decreasing function of d for d > 8. The theorem of BKPS follows as a special case of Theorem 1.2 by taking T to be a path. Figure 2 gives an example of a family of trees that can be used to deduce item (2) of Theorem 1.1 from Theorem 1.2. See Fig. 3 for another example application.
The next theorem shows that uniform spanning forests in different dimensions between 5 and 8 also have qualitatively different connectivity properties. The result is more naturally stated in terms of ubiquitous subhypergraphs in the component hypergraph of the USF; see the following section for definitions and Fig. 4 for an illustration of the relevant hypergraphs.

Ubiquity of general graphs and hypergraphs in the component graph.
In this section, we extend Theorem 1.2 to the case that H is not a tree. In order to formulate this extension, it is convenient to consider the even more general setting in which H is a hypergraph with boundary. Indeed, it is a surprising feature of the resulting theory that one is forced to consider hypergraphs even if one is interested only in graphs. We define a hypergraph H = (V , E, ⊥) to be a triple consisting of a set of vertices V , a set of edges E, and a binary relation ⊥⊆ V × E such that the set {v ∈ V : (v, e) ∈⊥} is nonempty for every e ∈ E. We write v ⊥ e or e ⊥ v and say that v is incident to e if (v, e) ∈⊥. Note that this definition is somewhat nonstandard, as it allows multiple edges with the same set of incident vertices. We say that a hypergraph is simple if it does not contain two distinct edges whose sets of incident vertices are equal. Every graph is also a hypergraph. A hypergraph with boundary H = (∂ V , V • , E, ⊥) is defined to be a hypergraph H = (V , E, ⊥) together with a partition of V into disjoint subsets, V = ∂ V ∪ V • , the boundary and interior vertices of H , such that ∂ V = ∅. The degree of a vertex in a hypergraph is the number of edges that are incident to it, and the degree of an edge in a hypergraph is the number of vertices it is incident to. To lighten notation, we will often write simply H = (∂ V , V • , E) for a hypergraph with boundary, leaving the incidence relation ⊥ implicit.
If H = (∂ V , V • , E, ⊥) is a hypergraph with boundary, a subhypergraph (with boundary) of H is defined to be a hypergraph with boundary of the form H = We say that a hypergraph with boundary This terminology used here arises from the following analogy: We imagine that from each vertex-edge pair (v, e) of H with v ⊥ e we hang a weight exerting a downward force of (d − 4), while from each edge and each interior vertex of H we attach a balloon exerting an upward force of either d or (d − 4) respectively. The net force is equal to the apparent weight. The hypergraph is buoyant (i.e., floats) if the apparent weight is non-positive. Theorem 1.4 is best understood as a special case of a more general theorem concerning the component hypergraph. Given a subset ω of a graph G and r ≥ 1, we define the component hypergraph C hyp r (ω) to be the simple hypergraph that has the components of ω as vertices, and where a finite set of components W is an edge of C hyp r (ω) if and only if there exists a set of diameter r in G that intersects every component of ω in the set W . Presence, faithful presence, ubiquity and faithful ubiquity of a hypergraph with boundary H in a hypergraph G are defined similarly to the graph case. For example, we say that a finite hypergraph with boundary indexed by the interior vertices of H such that for each e ∈ E there exists an edge f of G that is incident to all of the vertices in the set {x v : v ⊥ e}. Given a d-dimensional graph G and M ≥ 1, we let R G (M) be minimal such that there exists a set of vertices in G of diameter R G (M) that intersects M distinct components of the uniform spanning forest of G with positive probability. Given a hypergraph with boundary H , we let R G (H ) = R G (max e∈E deg(e)).

Organisation
In Sect. 2, we give background on uniform spanning forests, establish notation, and prove some simple preliminaries that will be used throughout the rest of the paper. In Sect. 3, we outline some of the key steps in the proof of the main theorems; this section is optional if the reader prefers to go straight to the fully detailed proofs. Section 4 is the computational heart of the paper, where the quantitative estimates needed for the proof of the main theorems are established. In Sect. 5, we deduce the main theorems from the estimates of Sect. 4 together with the multicomponent indistinguishability theorem of [12], which is used as a zero-one law. This section is quite short, most the work having already been done in Sect. 4. We conclude with some open problems and remarks in Sect. 6.

Basic notation
Let G be a d-dimensional transitive graph with vertex set V, and let F be the uniform spanning forest of G. For each set W ⊆ V, we write F (W ) for the event that the vertices of W are all in the same component of F. Let r ≥ 1 and let H = (∂ V , V • , E) be a finite hypergraph with boundary. We definê We write , , and for inequalities or equalities that hold up to a positive multiplicative constant depending only on some fixed data that will be clear from the context, usually G, H , and r , and write , and ≈ for inequalities or equalities that hold up to an additive constant depending only on the same data. In particular a b if and only if log 2 a ≈ log 2 b.
We sometimes write exp 2 (a) to mean 2 a .
For each two vertices x and y of G, we write x y = d G (x, y) + 1, where d G is the graph metric on G. For each vertex x of G and ∞ ≥ N > n ≥ 0, we define the dyadic shell for all n ≥ 0 and N ≥ n + 1. The upper bound is immediate, while the lower bound follows because x (n, N ) contains both some point y with x 0 y = 2 N −1 + 2 N −2 and the ball of radius 2 N −2 around this point y.

Uniform spanning forests
Given a finite connected graph G, we define UST G to be the uniform probability measure on the set of spanning trees of G, that is, connected subgraphs of G that contain every vertex of G and do not contain any cycles. Now suppose that G = (V , E) is an infinite, connected, locally finite graph, and let (V i ) i≥1 be an exhaustion of V by finite sets, that is, an increasing sequence of finite, connected subsets of V such that i≥1 V i = V . For each i ≥ 1, let G i be the subgraph of G induced 1 by V i , and let G * i be the graph formed from G by contracting V \V i down to a single vertex and deleting all of the self-loops that are created by this contraction. The free and wired uniform spanning forest (FUSF and WUSF) measures of G, denoted FUSF G and WUSF G , are defined to be the weak limits of the uniform spanning tree measures of G i and G * i respectively. That is, for every finite set S ⊂ E, Both limits were proven to exist by Pemantle [21] (although the WUSF was not considered explicitly until the work of Häggström [9]), and do not depend on the choice of exhaustion.
Benjamini, Lyons, Peres, and Schramm [3] proved that the WUSF and FUSF of G coincide if and only if G does not admit harmonic functions of finite Dirichlet energy, from which they deduced that the WUSF and FUSF coincide on any amenable transitive graph. In particular, it follows that the WUSF and FUSF coincide for every transitive d-dimensional graph, and in this context we refer to both the FUSF and WUSF measures on G as simply the uniform spanning forest measure, USF G , on G. We say that a random spanning forest of G is a uniform spanning forest of G if it has law USF G .

Wilson's algorithm
Wilson's algorithm [29] is a way of generating the uniform spanning tree of a finite graph by joining together loop-erased random walks. It was extended to generate the wired uniform spanning forests of infinite, transient graphs by Benjamini, Lyons, Peres, and Schramm [3].
Recall that, given a path γ = (γ n ) n≥0 in a graph G that is either finite or visits each vertex of G at most finitely often, the loop-erasure of γ is defined by deleting loops from γ chronologically as they are created. The loop-erasure of a simple random walk path is known as loop-erased random walk and was first studied by Lawler [17]. Formally, we define the loop-erasure of γ to be LE(γ ) = (γ τ i ) i≥0 , where τ i is defined recursively by setting τ 0 = 0 and (If G is not simple, then we also keep track of which edges are used by LE(γ ).) Let G be an infinite, connected, transient, locally finite graph. Wilson's algorithm rooted at infinity allows us to sample the wired uniform spanning forest of G as follows. Let (v i ) i≥1 be an enumeration of the vertices of G. Let F 0 = ∅, and define a sequence of random subforests (F i ) i≥0 of G as recursively follows.
(1) Given F i , let X i+1 be a random walk started at v i+1 , independent of F i .
(2) Let T i+1 be the first time X i+1 hits the set of vertices already included in F i , where T i+1 = ∞ if X i+1 never hits this set. Note that T i+1 will be zero if v i+1 is already included in F i . (3) Let F i+1 be the union of F i with the loop-erasure of the stopped random walk path n=0 . Finally, let F = i≥1 F i . This is Wilson's algorithm rooted at infinity: the resulting random forest F is a wired uniform spanning forest of G.

The main connectivity estimate
Let K be a finite set of vertices of G. Following [2], we define the spread of K , denoted K , to be x y : τ = (W , E) is a tree with vertex set K .
Note that the tree τ being minimized over in the definition of K need not be a subgraph of G. If we enumerate the vertices of K as x 1 , . . . , x n , then we have the simple estimate [2, Lemma 2.6] where the implied constant depends on the cardinality of K . In practice we will always use (2.2), rather than the definition, to estimate the spread. The main tool in our analysis of the USF is the following estimate of BKPS. Recall that F (K ) is the event that every vertex of K is in the same component of the uniform spanning forest F.
BKPS proved the theorem in the case G = Z d . The general case follows from the same proof by applying the heat kernel estimates of Hebisch and Saloff-Coste [10] (see Theorem 4.18), as stated in [2, Remark 6.12]. These heat kernel estimates imply in particular that the Greens function estimate Proof We may assume that I = {1, . . . , k} for some k ≥ 1. Given a collection of independent random walks X 1 , . . . , X n , let A(X 1 , . . . , X n ) be the indicator of the event that the forest generated by running the first n steps of Wilson's algorithm using the walks X 1 , . . . , X n , in that order, is connected. Thus, given a finite set K ⊂ V, we have P(F (K )) = P A X 1 , . . . , X |K | = 1 where X 1 , . . . , X |K | are independent random walks started at the vertices of K . Now suppose that (K i ) i∈I is a collection of finite sets, and suppose we generate a sample F of the USF, starting with independent random walks X 1,1 , . . . , X 1,|K 1 | , X 2,1 , . . . , X k,|K k | , where X i, j starts from the jth element of K i . Then we observe that and hence that The claim now follows from Theorem 2.1. It is also possible to prove (2.6) using the negative association property of the USF, see e.g. [8].

Witnesses
Let H be a finite hypergraph with boundary, let r ≥ 1, and let x = (x v ) v∈∂ V be a collection of vertices in G. We say that H is r -faithfully present at x if it is faithfully present at the components of x in C hyp r (F). We define r -presence of H at x similarly. Let E • be the set of pairs (e, v), where e ∈ E is an edge of H and v ⊥ e is a vertex of H incident to e. We say that ξ = (ξ (e,v) ) (e,v)∈E • ∈ V E • is a witness for the r -faithful presence of H at x if the following conditions hold: (1) For every e ∈ E and every u, v ⊥ e we have that ξ (e,v) ξ (e,u) ≤ r − 1. See Fig. 7  Often, x, r and H will be fixed. In this case we will speak simply of 'faithful presence' to mean 'r -faithful presence', 'robustly faithfully present' to mean 'r -robustly faithfully present', 'witnesses' to mean 'witnesses for the r -faithful presence of H at x', and so on.
x v 1 x v 3 } are edges of G and there exist three distinct trees of F each containing one of the sets It will be useful to define the following sets in which witnesses must live. For every (x v ) v∈∂ V , n ≥ 0 and N > n, let for every e ∈ E and every u, v ⊥ e} , so that ξ ∈ x (n, N ) E • is a witness for the faithful presence of H if and only if ξ ∈ •x (n, N ) and conditions (2), (3), and (4) in the definition of witnesses, above, hold.

Indistinguishability of tuples of trees
In this section we provide background on the notion of indistinguishability theorems, including the indistinguishability theorem of [12] which will play a major role in the proofs of our main theorems. Indistinguishability theorems tell us that, roughly speaking, 'all infinite components look alike'. The first such theorem was proven in the context of Bernoulli percolation by Lyons and Schramm [20]. Indistinguishability of components in uniform spanning forests was conjectured by Benjamini, Lyons, Peres, and Schramm [3] and proven by Hutchcroft and Nachmias [13]. (Partial progress was made independently at the same time by Timár [27].) All of the results just mentioned apply to individual components. In this paper, we will instead apply the indistinguishability theorem of [12], which yields a form of indistinguishability for multiple components in the uniform spanning forest. We will use this theorem as a zero-one law that allows us to pass from an estimate showing that certain events occur with positive probability to knowing that these events must occur with probability one.
We now give the definitions required to state this theorem. Let G = (V , E) be a graph, and let k ≥ 1. We define k (G) = {0, 1} E × V k , which we equip with its product σ -algebra and think of as the set of subgraphs of G rooted at an ordered k-tuple That is, A is a k-component property if it is stable under replacing the root vertices with other root vertices from within the same components. Given a k-component property A , we say that a k-tuple of components (K 1 , . . . , We say that A is a multicomponent property if it is a k-component property for some k ≥ 1. For our purposes, the key example of a tail multicomponent property is the property that some finite hypergraph with boundary H is r -robustly faithfully present at (x v ) v∈∂ V . Applying Theorem 2.3, we will deduce that if H is r -robustly faithfully present at some (x v ) v∈∂ V with positive probability then it must be almost surely r -robustly faithfully present at every (x v ) v∈∂ V for which the vertices {x v } v∈∂ V are all in distinct components of F.

Optimal coarsenings
In this section we study the min max problem appearing in Theorems 1.4 and 1.5, proving the following.
where [e] denotes the equivalence class of e under . It is easily seen that every coarsening of H can be uniquely represented in this way. We say that a coarsening H / of a hypergraph with boundary H is proper if there exist at least two nonidentical edges of H that are related under . Let It follows that Remark Lemma 2.6 yields a brute force algorithm for computing the value of the relevant max min problem that is exponentially faster than the trivial brute force algorithm, although still taking superexponential time in the number of edges of H .

Sketch of the proof
In this section we give a detail-free overview of the most important components of the proof. This section is completely optional; all the arguments and definitions mentioned here will be repeated in full detail later on.

Non-ubiquity in high dimensions
Let G be a d-dimensional transitive graph, let H be a finite hypergraph with boundary, and let F be the uniform spanning forest of G. We wish to show that if every coarsening of H has a subhypergraph that is not d-buoyant, then H is not faithfully ubiquitous in C hyp r (F) for any r ≥ 1 a.s. By Lemma 2.4, this condition is equivalent to there existing a subhypergraph of H none of whose coarsenings are d-buoyant. If H is faithfully ubiquitous then so are all of its subhypergraphs, and so it suffices to consider the case that H does not have any d-buoyant coarsenings, i.e., thatη d (H ) > 0.
To show that H is not faithfully ubiquitous, it would suffice to show that if the vertices x = (x v ) v∈∂ V are far apart from each other, then the expected total number of witnesses for the faithful presence of H at x is small. As it happens, we are not able to control the total number of witnesses without making further assumptions on H . Nevertheless, the most important step in our argument is to show that if x is contained in x (0, n−1), then the expected number of witnesses in x (n, n+1) is exponentially small as a function of n. Once we have done this, we will control the expected number of witnesses that occur 'at the same scale' as x by a similar argument. We are not finished at this point, of course, since we have not ruled out the existence of witnesses that are spread out across multiple scales. However, given the single-scale estimates, we are able to handle multi-scale witnesses of this form via an inductive argument on the size of H (Lemmas 4.7-4.9), which allows us to reduce from the multi-scale setting to the single-scale setting.
Let us briefly discuss how the single-scale estimate is attained. Write = x (n, n+ 1). Proposition 2.2 implies that the expected number of witnesses in x (n, n + 1) is at most a constant multiple of To control this sum, we split it as follows. Let L be the set of symmetric functions : For each ∈ L, let = ξ ∈ : 2 (e,e ) ≤ ξ e ξ e ≤ 2 (e,e )+2 for all e, e ∈ E , so that = ∈L . The advantage of this decomposition is that W is approximately constant on each set : On the other hand, by considering the number of choices we have for ξ e i at each step given our previous choices, it follows that whereˆ is the largest ultrametric on E that is dominated by . ( could be much smaller than this of course-it could even be empty.) We deduce that We have that log 2 |L| = E 2 log 2 (n + 1), which will be negligible compared with the rest of the expression in the case thatη d (H ) > 0. From here, the problem is to identify the ∈ L achieving the maximum above. We will argue, by invoking a general lemma (Lemma 4.4) about optimizing linear combinations of minima of distances on the ultrametric polytope, that there is an ∈ L maximizing the expression such that is an ultrametric and (e, e ) ∈ {0, n} for every e, e ∈ E.
giving the desired exponential decay.

Ubiquity in low dimensions
We now sketch the proof of ubiquity in low dimensions. Here we will only discuss  Let us suppose for now that every subhypergraph of H is d-buoyant (i.e., that we do not have to pass to a coarsening for this to be true). To prove that H has a positive probability of being robustly faithfully present at some x, we perform a first and second moment analysis on the number of witnesses in dyadic shells. Suppose that x is contained in x (0, n − 1). Since we are now interested in existence rather than nonexistence, we can make things easier for ourselves by considering only ξ that are both contained in a dyadic shell x (n, n + 1), and such that ξ (e,u) ξ (e ,u ) ≥ 2 n−C 1 whenever e = e , for some appropriate chosen constant C 1 . Furthermore, for each e ∈ E the points {ξ (e,u) : u ⊥ e} must be sufficiently well separated that there are not local obstructions to ξ being a witness-this is where we need that r ≥ R G (H ). Call such a ξ good, and denote the set of good ξ by x (n). We then argue that for good ξ , the probability that ξ is a witness is comparable to where ξ e is chosen arbitrarily from {ξ (e,u) : u ⊥ e} for each e, and hence that the expected number of witnesses in x (n) is comparable to 2 −η d (H )n . In other words, we have that the upper bound on the probability that ξ is a witness provided by Proposition 2.2 is comparable to the true probability when ξ is good. Our proof of this estimate appears in Sect. 4.3; unfortunately it is quite long.
Taking this lower bound on trust for now, the rest of the analysis proceeds similarly to that sketched in Sect. 3.1, and is in fact somewhat simpler thanks to our restriction to good configurations. The bound implies that the expected number of good witnesses in x (n, n + 1) is comparable to exp 2 [−η d (H ) n]. Estimating the second moment is equivalent to estimating the expected number of pairs ξ, ζ such that ξ and ζ are both good witnesses. Observe that if ξ and ζ are both good witnesses then the following hold: (e,v) and ζ (e ,v ) are in the same component of F for some (and hence every) e ⊥ v and e ⊥ v .
To account for the degrees of freedom given by (1), we define to be the set of Here and elsewhere, we use as a dummy symbol so that we can encode partial bijections by functions.) For each φ ∈ , we defineW φ (ξ, ζ ) to be the event that ξ and ζ are both witnesses, and that ξ (e,v) and ζ (e ,v ) Next, to account for the degrees of freedom given by (2), we define to be the set of functions ψ : E → E ∪ { } such that the preimage ψ −1 (e) has at most one element for each e ∈ E. For each ψ ∈ and k = (k e ) e∈E ∈ {0, . . . , n} E , let for all e ∈ E such that ψ(e) = , and ζ e ξ e ≥ 2 n−C 1 −2 for all e, e ∈ E such that e = ψ(e) .
We can easily upper bound the volume Using this together with Proposition 2.2, is straightforward to calculate that for every φ ∈ , ψ ∈ and k ∈ {0, . . . , n} E . We now come to some case analysis. Observe that for every ψ ∈ and e ∈ E, we have that is not an integer, the middle case cannot occur and we obtain that From here, our task is to show that the expression on the right hand side is maximized when φ ≡ and ψ ≡ , in which case it is equal to −2dη d (H )n. To do this, we identify optimal choices of φ and ψ with subhypergraphs of H , and use the assumption that every subhypergraph of H is d-buoyant. This should be compared to how, in the proof of non-ubiquity sketched in the previous subsection, we identified optimal choices of with coarsenings of H . Once we have this, since there are only a constant number of choices for φ and ψ, we deduce that the second moment of the number of good witnesses is comparable to the square of the first moment. Thus, it follows from the Cauchy-Schwarz inequality that the probability of there being a good witness in each sufficiently large dyadic shell is bounded from below by some ε > 0, and we deduce from Fatou's lemma that there are good witnesses in infinitely many dyadic shells with probability at least ε. This completes the proof that robust faithful presence occurs with positive probability.
It remains to remove the simplifying assumption we placed on H , i.e., to allow ourselves to pass to a coarsening of H all of whose subhypergraphs are d-buoyant before proving faithful ubiquity. To do this, we introduce the notion of constellations of witnesses. These are larger collections of points, defined in such a way that every constellation of witness for H contains a witness for each refinement of H . In the actual, fully detailed proof we will work with constellations from the beginning. This does not add many complications.

Non-ubiquity in high dimensions
The goal of this section is to prove the following. Let G be a d-dimensional graph with d > 4, and let F be the uniform spanning For each (ξ e ) e∈E ∈ V E , we also define and so that, if we choose a vertex u(e) ⊥ e arbitrarily for each e ∈ E and set (ξ e ) e∈E = (ξ (e,u(e)) ) e∈E , it follows from Proposition 2.2 that for every x, n, and N . To avoid trivialities, in the case that H does not have any edges we define W H x (n, N ) = 1 for every x ∈ V ∂ V and N > n.
In order to prove Proposition 4.1, it will suffice to show that if H has a subhypergraph with boundary that does not have any d-buoyant coarsenings, then for every ε > 0 there exists a collection of vertices (x u ) u∈∂ V such that all the vertices x u are in a different component of F with probability at least 1/2 (which, by Theorem 2.1, will be the case if the vertices are all far away from each other), but P(H is faithfully present at In order to prove this, we seek to obtain upper bounds on the quantity W H x (n, N ). We begin by considering the case of a single distant scale. That is, the case that |N −n| is a constant and all the points of x are contained in x (0, n−1).
It will be useful for applications in Sect. 4.3 to prove a more general result. A graph G is said to be d-Ahlfors regular if there exists a positive constant c such that c −1 r d ≤ |B(x, r )| ≤ cr d for every r ≥ 1 and every x ∈ V (in which case we say G is d-Ahlfors regular with constant c). Given α > 0 and a finite hypergraph with boundary H , we define and, for each N > n, x , so that Lemma 4.2 follows as a special case of the following lemma.
Before proving this lemma, we will require a quick detour to analyze a relevant optimization problem.

Optimization on the ultrametric polytope
Recall that a (semi)metric space (X , d) is an ultrametric space if d(x, y) ≤ max{d(x, z), d(z, y)} for every three points x, y, z ∈ X . For each finite set A, the ultrametric polytope on A is defined to be which is a closed convex subset of R A 2 . We consider U A to be the set of all ultrametrics on A with distances bounded by 1. We write P(A 2 ) for the set of subsets of A 2 .

Lemma 4.4 Let A be a finite non-empty set, and let F :
. Then the maximum of F on U A is obtained by an ultrametric for which all distances are either zero or one. That is, Proof We prove the claim by induction on |A|. The case |A| = 1 is trivial. Suppose that the claim holds for all sets with cardinality less than that of A. We may assume that (a, a) / ∈ W k for every 1 ≤ k ≤ K and i ∈ A, since if (a, a) ∈ W k for some 1 ≤ k ≤ K then the term c k min{x a,b : (a, b) ∈ W k } is identically zero on U A . We write 1 for the vector It is easily verified that . We may assume that F(y) > F(1) and that F(y) > F(0) = 0, since otherwise the claim is trivial. Let m = min{y a,b : a, b ∈ A, a = b}, which is less than one by assumption. We have that and so we must have m = 0 since y maximizes F. Define an equivalence relation on A by letting a and b be related if and only if y a,b = 0. We writeâ for the equivalence class of b under . Let C be the set of equivalence classes of , and let φ : for every x ∈ U n . For each 1 ≤ k ≤ K , letŴ k be the set of pairsâ,b ∈ C such that (a, b) ∈ W k for some a in the equivalence classâ and b in the equivalence classb.
We have thatF = F • φ, and, since y maximized F, we deduce that, by the induction hypothesis, completing the proof.
We will also require the following generalisation of Lemma 4.4. For each finite collection of disjoint finite sets {A i } i∈I with union A = i∈I A i , we define U {A i } i∈I = {x ∈ U A : x a,b = 1 for every distinct i, j ∈ I and every a ∈ A i and b ∈ A j .}.

Lemma 4.5
Let {A i } i∈I be a finite collection of disjoint, finite, non-empty sets with union A = i∈I A i , and let F : R A 2 → R be of the form where K < ∞, c 1 , . . . , c K ∈ R, and W 1 , . . . , W K ∈ P(A 2 ). Then the maximum of F on U A is obtained by an ultrametric for which all distances are either zero or one. That is, Proof We prove the claim by fixing the index set I and inducting on |A|. The case |A| = |I | is trivial. Suppose that the claim holds for all collections of finite disjoint sets indexed by I with total cardinality less than that of A. We may assume that (i, i) / ∈ W k for every 1 ≤ k ≤ K and i ∈ A, since if (i, i) ∈ W k for some 1 ≤ k ≤ K then the term c k min{x i, j : (i, j) ∈ W k } is identically zero on U A . Furthermore, we may assume that W k contains more than one element of at least one of the sets A i for each 1 ≤ k ≤ K , since otherwise the term c k min{x i, j : (i, j) ∈ W k } is equal to the constant c k on U {A i } i∈I . We write 1 and i for the vectors It is easily verified that The rest of the proof is similar to that of Lemma 4.4.

Back to the uniform spanning forest
We now return to the proofs of Proposition 4.1 and Lemma 4.3.
Let e 1 , . . . , e |E| be an enumeration of E. For every ∈ L, every 1 ≤ j < i ≤ |E| and every ξ ∈ we have that By considering the number of choices we have for ξ e i at each step given our previous choices, it follows that Now, for every ξ ∈ , we have that Thus, from (4.5) and (4.2) we have that Let Q : L → R be defined to be the expression on the right hand side of (4.3). We clearly have that Q(ˆ ) ≥ Q( ) for every ∈ L, and so there exists ∈ L maximizing Q such that is an ultrametric. It follows from Lemma 4.4 (applied to the normalized ultrametric /n) that there exists ∈ L maximizing Q such that is an ultrametric and every value of is in {0, n}. Fix one such , and define an equivalence relation on E by letting e e if and only if (e, e ) = 0, which is an equivalence relation since is an ultrametric. Observe that, for every 2 ≤ i ≤ |E|, min{ (e i , e j ) : j < i} = 1[e j is not in the equivalence class of e i for any j < i] n, and hence that Similarly, we have that, for every vertex u of H , where we say that an equivalence class of is incident to u if it contains an edge that is incident to u. Thus, we have that Let H = H / be the coarsening of H associated to as in Sect. 2.7. We can rewrite (4.4) as Since |L| ≤ (n + 1) |E| 2 , we deduce that as claimed.
Next, we consider the case that the points x v are roughly equally spaced and we are summing over points ξ that are on the same scale as the spacing of the x v .
Proof We may assume that E = ∅, the case E = ∅ being trivial. For notational convenience, we will write ξ v = x v , and consider v ⊥ v for every vertex v ∈ ∂ V . Write = x (0, n + m 2 ), and observe that for each ξ ∈ and e ∈ E there exists at most one v ∈ ∂ V for which log 2 ξ e ξ v < n − m 1 − 1. To account for these degrees of freedom, we define to be the set of functions φ : and observe that = φ∈ ∈L φ φ, . Now, for each φ ∈ and ∈ L φ , letˆ be the largest ultrametric on E ∪ ∂ V that is dominated by . Observe thatˆ ∈ L φ , and that, as in the previous lemma, we have that for every e, e ∈ E ∪ ∂ V . Let e 1 , . . . , e |E| be an enumeration of E, and let e 0 , e −1 , . . . , e −|∂ V |+1 be an enumeration of ∂ V . As in the proof of the previous lemma, we have the volume estimate Now, for every ξ ∈ φ, , we have that, similarly to the previous proof, min{ (e i , e j ) : j < i, e j ⊥ u}.
(Recall that we are considering u ⊥ u for each u ∈ ∂ V .) Thus, we have (4.6) Let Q : L φ → R be defined to be the expression on the right hand side of (4.6). Similarly to the previous proof but applying Lemma  Since d > 4 and each equivalence class of can contain at most one vertex of v, we see that Q increases if we remove a vertex v ∈ ∂ V from its equivalence class. Since was chosen to maximize Q, we deduce that the equivalence class of v under is a singleton for every v ∈ ∂ V . Thus, there exists an ultrametric ∈ L φ maximizing Q such that (e, e ) ∈ {0, n} for every e, e ∈ E and (e, v) = n for every e ∈ E and v ∈ ∂ V . Letting be the equivalence relation on E (rather than E ∪ ∂ V ) corresponding to such an optimal , we have

Lemma 4.7 (Induction estimate) Let G be a d-dimensional transitive graph and let H be a finite hypergraph with boundary. Then there exists a constant c = c(G, H ) such that
Note that when |E| ≥ 1 we must consider the term E = ∅ when taking the maximum in this lemma, which gives −η d (H )N + |E| 2 log 2 N .

Proof
The claim is trivial in the case E = ∅, so suppose that |E| ≥ 1. Let = For each E E and every 1 ≤ m ≤ |E| + 1, let Observe that if ξ ∈ then, by the Pigeonhole Principle, there must exist 1 ≤ m ≤ |E| + 2 such that ξ e is not in x (N − m − 1, N − m) for any e ∈ E, and we deduce that Thus, to prove the lemma it suffices to show that For each ξ ∈ E ,m , let ξ = (ξ e ) e∈E = (ξ e ) e∈E and ξ = (ξ e ) e∈E = (ξ e ) e∈E . Then the above displays imply that for every ξ ∈ E ,m . Thus, summing over ξ ∈ ( x (0, N + m − 1)) E and ξ ∈ ( x (N + m, N + |E| + 2)) E , we obtain that where the second inequality follows from Lemma 4.2.
To deduce (4.8) from (4.9), it suffices to show that We now use Lemmas 4.6 and 4.7 to perform an inductive analysis of W. Although we are mostly interested in the non-buoyant case, we begin by controlling the buoyant case.  If every subhypergraph of H has a d-buoyant coarsening, then there exists a  constant c = c(G, H , m) such that for each proper subhypergraph H of H , and hence that (Note that the implicit constants depending on H from the induction hypothesis are bounded by a constant depending on H since H has only finitely many subhypergraphs.) Observe that whenever E E we have that and so we deduce that for every proper subhypergraph H of H . Thus, we have that for all N ≥ n, where we applied Lemma 4.7 in the second inequality. Summing from n to N we deduce that Using Lemma 4.6 to control the term W H x (0, n) completes the induction.
We are now ready to perform a similar induction for the non-buoyant case. Note that in this case the induction hypothesis concerns probabilities rather than expectations. This is necessary because the expectations can grow as N → ∞ for the wrong reasons if H has a buoyant coarsening but has a subhypergraph that does not have a buoyant coarsening (e.g. the tree in Fig. 3).

Lemma 4.9 (Every scale, non-buoyant case) Let H be a finite hypergraph with boundary such that E = ∅, let m ≥ 1, and suppose that H has a subhypergraph that does not have any d-buoyant coarsenings. Then there exist positive constants c
Proof We induct on the number of edges in H . For the base case, suppose that H has a single edge. In this case we must have that η d (H ) > 0, and we deduce from Lemmas 4.2 and 4.6 that so that the claim follows from Markov's inequality. This establishes the base case of the induction. Now suppose that |E| > 1 and that the claim holds for all finite hypergraphs with boundary that have fewer edges than H . If H has a proper subhypergraph H witĥ ∞) is, and so the claim follows from the induction hypothesis, letting c 1 (G, H , m) = c 1 (G, H , m) and c 2 (G, H , m) = c 2 (G, H , m).
Thus, it suffices to consider the case thatη d (H ) > 0 but thatη d (H ) ≤ 0 for every proper subhypergraph H of H . In this case, we apply Lemma 4.7 to deduce that Lemma 4.8 then yields that Finally, combining this with Lemma 4.6 yields that, sinceη d (H ) > 0,

Positive probability of robust faithful presence in low dimensions
Recall that if G is a d-dimensional transitive graph, H = (∂ V , V • , E) is a finite hypergraph with boundary, that r ≥ 1 and that (x v ) v∈∂ V is a collection of points in G, we say that H is r -robustly faithfully present at We say that a set W ⊂ V is well-separated if the vertices of W are all in different components of the uniform spanning forest F with positive probability.

Lemma 4.10 Let G be a d-dimensional transitive graph with d > 4, and let F be the uniform spanning forest of G. Then a finite set W ⊂ V is well-separated if and only if when we start a collection of independent simple random walks {X
Proof We will be brief since the statement is intuitively obvious from Wilson's algorithm and the details are somewhat tedious. The 'if' implication follows trivially from Wilson's algorithm. To see the reverse implication, suppose that W is well-separated and consider the paths {( v i ) i≥0 : v ∈ W } from the vertices of W to infinity in F. Using Wilson's algorithm and the Green function estimate (2.4), it is easily verified that almost surely on the event that the vertices of W are all in different components of F. Let i ≥ 1 and consider the collection of simple random walks Y v,i started at v i and conditionally independent of each other and of F given ( v i ) v∈W , and letỸ v,i be the random path formed by concatenating ( v j ) i j=1 with Y v,i . It follows from (4.11) and Markov's inequality that 12) where we recall that F (W ) is the event that all the vertices of W are in different components of F. In particular, it follows that the probability appearing on the left hand side of (4.12) is positive for some i 0 ≥ 0. The result now follows since the walks {X v : v ∈ W } have a positive probability of following the paths v for their first i 0 steps, and on this event their conditional distribution coincides with that of The goal of this subsection is to prove criteria for robust faithful presence to occur with positive probability. We begin with the case that d/(d − 4) is not an integer (i.e., d / ∈ {5, 6, 8}), which is technically simpler. The corresponding proposition for d = 5, 6, 8 is given in Proposition 4.15. We call a set of vertices y = (y (B,b) ) of G indexed by P • (A) an A-constellation.
Given an A-constellation y, we define A r (y) to be the event that y (B,b) and y (B ,b ) are connected in F if and only if b = b , and in this case they are connected by a path in F with diameter at most r . We say that an A-constellation y in G is r -good if it satisfies the following conditions.
The proof of the following lemma is deferred to Sect. 4.3.

Lemma 4.12 Let G be a d-dimensional transitive graph with d > 4. Let A be a finite set. Then there exists r = r (|A|) such that for every vertex x of G, there exists an r -good A-constellation contained in the ball of radius r around x.
Let H = (∂ V , V • , E) be a finite hypergraph with boundary with at least one edge, and let r = r (max e deg(e)) be as in Lemma 4.12. We write P • (e) = P • ({v ∈ V : v ⊥ e}) for each e ∈ E. For each ξ = (ξ e ) e∈E ∈ V E and each e ∈ E, we let (ξ (e,B,v) ) (B,v)∈P • (e) be an r -good e-constellation contained in the ball of radius r about ξ e , whose existence is guaranteed by Lemma 4.12.
For each x = (x v ) v∈∂ V and ξ = (ξ e ) e∈E , we defineW (x, ξ) to be the event that the following conditions hold: For each n ≥ 0, let x (n) be the set where C 1 = C 1 (E) is chosen so that log 2 | x (n)| ≈ nd|E| for all n sufficiently large and all x. It is easy to see that such a constant exists using the d-dimensionality of G.
For each n ≥ 0 we defineS x (n) to be the random variablẽ so that every refinement H of H is R G (H )-faithfully present at x on the event that S x (n) is positive for some n ≥ 0, and every refinement H of H is R G (H )-robustly faithfully present at x on the event thatS x (n) is positive for infinitely many n ≥ 0.
The following lemma lower bounds the first moment ofS n .

Lemma 4.13 Let G be a d-dimensional transitive graph with d > 4.
Let H be a finite hypergraph with boundary with at least one edge, let ε > 0, and suppose that for every ξ ∈ x (n) and hence that The proofs of Lemmas 4.12 and 4.13 are unfortunately rather technical, and are deferred to Sect. 4.3. For the rest of this section, we will take these lemmas as given, and use them to prove Proposition 4.11. The key remaining step is to upper bound the second moment of the random variableS x (n).
Proof Observe that if ξ, ζ ∈ x (n) are such that the eventsW (x, ξ) andW (x, ζ ) both occur, then the following hold: As a bookkeeping tool to account for the first of these degrees of freedom, we define be the set of functions φ : V • → V • ∪{ } such that the preimage φ −1 (v) has at most one element for each v ∈ V • . We write φ −1 (v) = if v is not in the image of φ ∈ , and write φ(v) = v for every v ∈ ∂ V . (Here and elsewhere, we use as a dummy symbol so that we can encode partial bijections by functions.) For each φ ∈ , and ξ, ζ ∈ V, define the eventW φ (ζ, ξ ) to be the event that both the eventW (x, ξ) ∩W (x, ζ ) occurs, and that for any two distinct vertices u, v ∈ V • the components of and hence that

It follows from Proposition 2.2 that
We define R φ (ξ, ζ ) to be the expression on the right hand side of (4.13), so that We now account for the second of the two degrees of freedom above. Let be the set of functions ψ : E → E ∪ { } such that the preimage ψ −1 (e) has at most one element for every e ∈ E. For each ψ ∈ and k = (k e ) e∈E ∈ {0, . . . , n} E , let for all e ∈ E such that ψ(e) = , and ζ e ξ e ≥ 2 n−C 1 −2 for all e, e ∈ E such that e = ψ(e) , where C 1 is the constant from the definition of x (n), and observe that (4.14) For each ξ, ζ ∈ x (n) and e ∈ E, there is at most one e ∈ E such that ζ e ξ e ≤ 2 n−C 1 −2 , and it follows that where the union is taken over ψ ∈ and k ∈ {0, . . . , n} E . Now, for any ξ, ζ ∈ ψ,k and u ∈ V • with φ(u) = , we have that Meanwhile, we have that for every u ∈ ∂ V . Summing these estimates yields Thus, using the volume estimate (4.14), we have that Observe that for every ψ ∈ and e ∈ E, we have that Thus, summing over k, we see that for every ψ ∈ and φ ∈ we have that   4) is not an integer, the last term is zero, so that if we define Q : Thus, since | × | does not depend on n, we have that and so it suffices to prove that Q(φ, ψ) ≤ 0 for every (φ, ψ) ∈ × . To prove this, first observe that we can bound Let H be the subhypergraph of H with boundary vertices given by the boundary vertices of H , edges given by the set of edges of H that have |{u ⊥ e : φ(u) = }| > d/(d − 4), and interior vertices given by the set of interior vertices u of H for which φ(u) = and φ(u) ⊥ e for some e ∈ E . Then we can rewritẽ where the second inequality follows by the assumption that every subhypergraph of H is d-buoyant. This completes the proof.

Proof of Proposition 4.11
Suppose that the finite hypergraph with boundary H has a doptimal coarsening all of whose subhypergraphs are d-buoyant. Then the lower bound on the square of the first moment ofS H x (n) provided by Lemma 4.13 and the upper bound on the second moment ofS H x (n) provided by Lemma 4.14 coincide, so that the Cauchy-Schwarz inequality implies that for every n such that x u x v ≤ 2 n−1 for every u, v ∈ ∂ V . It follows from Fatou's lemma that so that H is robustly faithfully present at x with positive probability as claimed.

The cases d = 5, 6, 8.
We now treat the cases in which d/(d − 4) is an integer. This requires somewhat more care owing to the possible presence of the logarithmic term in (4.15). Indeed, we will only treat certain special 'building block' hypergraphs directly via the second moment method. We will later build other hypergraphs out of these special hypergraphs in order to prove the main theorems.
be a finite hypergraph with boundary. We say that a subhyper- is incident to at most one edge in E . For example, every full subhypergraph containing every boundary vertex is bordered. We say that a subhypergraph of H is proper if it is not equal to H and non-trivial if it has at least one edge. We say that H is d-basic if it does not have any edges of degree less than or equal to d/(d − 4) and does not contain any proper, non-trivial bordered subhypergraphs H with η d (H ) = 0.

Proposition 4.15
Let G be a d-dimensional transitive graph with d ∈ {5, 6, 8}, and let F be the uniform spanning forest of G. Let H be a finite hypergraph with boundary with at least one edge. Suppose additionally that one of the following assumptions holds: (1) H is a refinement of a hypergraph with boundary that has exactly one edge, the unique edge contains exactly d/(d − 4) boundary vertices, and every interior vertex is incident to the unique edge.
or (2) H has a d-basic coarsening with more than one edge, all of whose subhypergraphs are d-buoyant.
Then for every r ≥ R G (H ) and every well-separated collection of points (x v ) v∈∂ V in V there is a positive probability that the vertices x u are all in different components of F and that H is robustly faithfully present at x.
The Proof of Proposition 4.15 will apply the following lemma, which is the analogue of Lemma 4.14 in this context.

Lemma 4.16 Let G be a d-dimensional transitive graph with d ∈ {5, 6, 8}. Let H be a hypergraph with boundary with at least one edge such that every subhypergraph of H is d-buoyant. (1) If H has exactly one edge, this unique edge is incident to exactly d/(d−4) boundary
vertices, and every interior vertex is incident to this unique edge, then there exists a constant c = c(G, H ) such that H is d-basic, then there exists a constant c = c(G, H ) such that Proof Note that in both cases we have that every subhypergraph of H is d-buoyant. We use the notation of the Proof of Proposition 4.11. As in Eq. (4.15) of that proof, we have that where Q(φ, ψ) is defined as in (4.16). Moreover, the same argument used in that proof shows that Q(φ, ψ) ≤ 0 for every (φ, ψ) ∈ × . In case (1) of the lemma, in which H has a single edge, we immediately obtain the desired bound since η d (H ) = 0 and the coefficient of the log 2 n term is either 0 or 1. Now suppose that H is d-basic. Let L(φ, ψ) be the coefficient of log 2 n in (4.18). Note that H cannot have an edge whose intersection with ∂ V has (d − 4)/d elements or more, since otherwise the subhypergraph H of H with that single edge and with no internal vertices is proper, bordered, and has Let Isom ⊆ × be the set of all (φ, ψ) such that φ(u) ⊥ ψ(e) for every e ∈ E and v ⊥ e. Since H is d-basic we have that if (φ, ψ) ∈ Isom then We claim that Q(φ, ψ) ≤ −(d − 4) unless either φ = φ 0 or (φ, ψ) ∈ Isom. Once proven this will conclude the proof, since we will then have that for every (φ, ψ) ∈ × , from which we can conclude by summing over × as done previously.
We first prove that We claim that if φ is such that η d (H ) = 0 then H is bordered, and consequently is either equal to H or does not have any edges by our assumptions on H . To see this, suppose for contradiction that H is not bordered, so that there exists a vertex v ∈ V • \V • that is incident to more than one edge of H . Let H be the subhypergraph of It remains to show that if φ(v) = for every v ∈ V then Q(φ 1 , ψ) ≤ −(d − 4) unless (φ, ψ) ∈ Isom. Since every edge of H has degree strictly larger than d/(d −4), we have that for every e ∈ E and every (φ, ψ) ∈ × such that |{u ⊥ e : φ(u) ⊥ ψ(e)}| < deg(e). It follows easily from this and the definition of Since η d (H ) ≤ 0 by assumption, it follows that Q(φ, ψ) ≤ −(d − 4) unless (φ, ψ) ∈ Isom. This concludes the proof.
Lemma 4.14 (together with Lemma 4.13) is already sufficient to yield case (2) of Proposition 4.15. To handle case (1), we will require the following additional estimate.
Proof Let andW φ (ξ, ζ ) be defined as in the Proof of Lemma 4.14.
For every ξ ∈ x (n) and ζ ∈ x (n + m), we have that all distances relevant to our calculations are on the order of either 2 n or 2 n+m . That is, log 2 ξ e ξ e , log 2 ξ e x v ≈ n and log 2 ξ e ζ e , log 2 ζ e ζ e , log 2 ζ e x v ≈ n + m for all e, e ∈ E and v ∈ ∂ V . Thus, using (4.13), can estimate (|{e ∈ E : e ⊥ u}| n − n + |{e ∈ E : e ⊥ φ(u)}| (n + m)) which is maximized when φ(v) = for all v ∈ V • . Now, since we deduce that (H )(2n + m) as claimed.

Proof of Proposition 4.15 given Lemmas 4.12 and 4.13
The second case, in which H has a d-basic coarsening with more than one edge all of whose subhypergraphs are dbuoyant, follows from Lemma 4.12 and Lemmas 4.13 and 4.16 exactly as in the proof of Proposition 4.11. Now suppose that H is a refinement of a hypergraph with boundary H that has d/(d − 4) boundary vertices and a single edge incident to every vertex. Then η d (H ) = 0 and every subhypergraph of H is d-buoyant. Applying Lemmas 4.13, 4.16 and 4.17, we deduce that for every n such that x u x v ≤ 2 n−1 for every u, v ∈ ∂ V , from which it follows by Cauchy-Schwarz that Now, recall that two graphs G = (V , E) and G = (V , E ) are said to be (α, β)rough isometric if there exists a function φ : V → V such that the following conditions hold.
The following stability theorem for Gaussian heat kernel estimates follows from the work of Delmotte [5]; see also [15,Theorem 3.3.5]. Recall that a function h : V → R defined on the vertex set of a graph is said to be for every vertex v ∈ A, where the sum is taken with appropriate multiplicities if there are multiple edges between u and v. The graph G is said to satisfy an elliptic Harnack inequality if for every α > 1, there exist a constant c(α) ≥ 1 such that for every two vertices u and v of G and every positive function h that is harmonic on the set in which case we say that G satisfies an elliptic Harnack inequality with constants c(α).
The following theorem also follows from the work of Delmotte [5], and was implicit in the earlier work of e.g. Fabes and Stroock [6]; see also [15,Theorem 3.3.5]. Note that these references all concern the parabolic Harnack inequality, which is stronger than the elliptic Harnack inequality. G be a graph. If G satisfies (c 1 , c 1 )-Gaussian heat kernel estimates, then there exists c 2 (α) = c 2 (α, c 1 ) such that G satisfies an elliptic Harnack inequality with constants c 2 (α).

Theorem 4.20 Let
We remark that the elliptic Harnack inequality has recently been shown to be stable under rough isometries in the breakthrough work of Barlow and Murugan [1].
Recall that a graph is said to be d-Ahlfors regular if there exists a positive constant c such that c −1 r d ≤ |B(x, r )| ≤ cr d for every r ≥ 1 and every x ∈ V (in which case we say G is d-Ahlfors regular with constant c). Ahlfors regularity is clearly preserved by rough isometry, in the sense that if G and G are (α, β)-rough isometric graphs for some positive α, β, and G is d-Ahlfors regular with constant c, then there exists a constant c = c (α, β, c) such that G is d-Ahlfors regular with constant c .
Observe that if the graph G is d-Ahlfors regular for some d > 2 and satisfies a Gaussian heat kernel estimate, then summing the estimate (4.19) yields that for every vertex v, and that for all vertices u and v of G.
We now turn to the proofs of Lemmas 4.12 and 4.13. The key to both proofs is the following lemma.  Clearly this is possible for sufficiently large r . We have by independence that On the other hand, it follows easily from the Greens function estimate (2.4) that if r is sufficiently large (depending on |A| and ε) then and we deduce that for all n sufficiently large and ξ ∈ x (n). Let G ξ be the graph obtained by contracting the tree T (e,v) (ξ ) down to a single vertex for each (e, v) ∈ E • . The spatial Markov property of the USF (see e.g. [14, Section 2.2.1]) implies that the law of F given the event B(ξ ) is equal to the law of the union of (e,v)∈E • T (e,v) (ξ ) with the uniform spanning forest of G ξ . Observe that G ξ and G are rough isometric, with constants depending only on G and H , and that G ξ has degrees bounded by a constant depending only on G and H . Thus, it follows from Theorem 4.18-4.20 that G ξ is d-Ahlfors regular, satisfies Gaussian heat kernel estimates, and satisfies an elliptic Harnack inequality, each with constants depending only on H and G.
Let ) be the vertex of G ξ that was formed by contracting T (e,v) (ξ ), and let x ( ,v) completing the proof.
We now start working towards the Proof of Lemma 4.21. We begin with the following simple estimate. for every c ≥ C, every vertex x, every n ≥ 1, and every u, w ∈ x (n + c, n + 2c).

Proof
The upper bound follows immediately from (4.20). We now prove the lower bound. For every c ≥ 1 and every u, w ∈ x (n + c, ∞), we have that Thus, we have that P u (hit w and x (0, n)) ≤ P u (hit x (0, n) after hitting w) + P u (hit w after hitting x (0, n)) where the second term is bounded by conditioning on the location at which the walk hits x (0, n) and then using the strong Markov property. By the triangle inequality, we must have that at least one of ux or wx is greater than 1 2 uw . This yields the bound On the other hand, if u, w ∈ x (n + c, n + 2c) then conditioning on the location at which the walk hits x (n + 3c, ∞) yields that The claim now follows easily.

Proof of Lemma 4.13
For each 1 ≤ i ≤ N , let x i be chosen arbitrarily from the set K i . Let (X x ) x∈K be a collection of independent random walks on G, where X x is started at x for each x ∈ K , and write X i = X x i . Let K i = K i \{x i } for each 1 ≤ i ≤ N and let K = N i=1 K i . In this proof, implicit constants will be functions of |K |, N , c 0 , and d. We take n such that 2 n−1 ≤ diam(K ) ≤ 2 n .
Let c 1 , c 2 , c 3 be constants to be determined. For each y = (y x ) x∈K ∈ ( (n + c 1 , n + c 3 )) K , let Y y be the event Let C (c 2 ) be the event that none of the walks X x intersect each other before time 2 2(n+c 2 ) , so that P(C (c 2 )) ≥ ε for every c 2 ≥ 0 by assumption. For each x ∈ K , let D x (c 1 , c 3 ) be the event that X x 2 2(n+c 2 ) is in (n+c 1 , n+c 3 ) and that X x m ∈ (n, ∞) for all m ≥ 2 2(n+c 2 ) , and let D(c 1 , c 3 ) = D x (c 1 , c 3 ). It follows by an easy application of the Gaussian heat kernel estimates that we can choose c 2 = c 2 (G, N , ε) and c 3 = c 3 (G, N , ε) sufficiently large that (4.25) for every y = (y x ) x∈K ∈ ( (n + c 1 , n + c 3 )) K , and in particular so that P(C (c 2 ) ∩ D(c 1 , c 3 )) ≥ ε. We fix some such sufficiently large c 1 , c 2 , and c 3 , and also assume that c 1 is larger than the constant from Lemma 4.22. We write For each 1 ≤ i ≤ N and x ∈ K i , we define I x to be the event that the walk X x hits the set before hitting (n + 6c 3 , ∞), and let I = x∈K I x .
For each x and x in K , we define E x,x to be the event that the walks X x and X x intersect, and let These events have been defined so that, if we sample F using Wilson's algorithm, beginning with the walks {X v : v ∈ V } (in any order) and then the walks {X x : x ∈ K } (in any order), we have that if and only if i = j , and each two points in K i are connected by a path in F of diameter at most 2 6c Thus, it suffices to prove that We break this estimate up into the following two lemmas: one lower bounding the probability of the good event C ∩D ∩I , and the other upper bounding the probability of the bad event C ∩ D ∩ I ∩ E .

Lemma 4.23
The estimate holds for every x ∈ K and y = (y x ) x∈K ∈ ( (n + c 1 , n + c 3 )) K .
Moreover, for every m, ≥ 2 2(n+c 2 ) and every y ∈ ( (n + c 1 , n + c 3 )) K , the walks Y k k≥m and Z k k≥ have the same distribution conditional on the event Thus, we deduce that whenever the event being conditioned on has positive probability, and therefore that ·P y x ( hit w before (n + 6c 3 , ∞) | do not hit (0, n)) On the other hand, we have that Meanwhile, decomposing E[I 2 | Y y ] according to the location of the intersections and applying the Gaussian heat kernel estimates yields that where the two different terms come from whether Y and Z hit the points of intersection in the same order or not. With the possible exception of wz , all the distances involved in this expression are comparable to 2 n . Thus, we obtain that For each w ∈ V, considering the contributions of dyadic shells centred at w yields that, since d > 4, and we deduce that Thus, the Cauchy-Schwarz inequality implies that as claimed.
We next use the elliptic Harnack inequality to pass from an estimate on I x to an estimate on I .
Let X be the σ -algebra generated by the random walks (X i ) N i=1 . Observe that for each x ∈ K we have , never leave (n, ∞) P y x (never leave (n, ∞)) P y x hit L i good before (0, n + 6c 3 ), never leave (n, ∞) .
The right hand side of the second line is a positive harmonic function of y x on (n + c 1 , n + c 3 + 1), and so the elliptic Harnack inequality implies that for every y, y ∈ ( (n + c 1 , n + c 3 )) K and every x ∈ K , we have that Furthermore, if y is obtained from y by swapping y x and y x for some 1 ≤ i ≤ N and x, x ∈ K i , then clearly Therefore, it follows that for all 1 ≤ i ≤ N and x, x ∈ K i . Since the events I x are conditionally independent given the σ -algebra X and the event C ∩ D ∩ Y y , we deduce that Now, the random variables P(I x i | X , C ∩D ∩Y y ) |K i | are independent conditional on the event C ∩ D ∩ Y y , and so we have that as claimed, where the second line follows from Jensen's inequality.
Finally, it remains to show that the probability of getting unwanted intersections in addition to those that we do want is of lower order than the probability of just getting the intersections that we want.

Lemma 4.25
We have that Proof For each w ∈ V and x, x ∈ K , let E x,x (w) be the event that X x and X x both hit w. Let ζ = (ζ x ) x∈K and let σ = (σ i ) N i=1 be such that σ v is a bijection from {1, . . . , |K i |} to K i for each 1 ≤ i ≤ N . We define R σ (ζ ) to be the event that for each 1 ≤ i ≤ N the walk X i passes through the points {ζ x : x ∈ K i } in the order given by σ and that for each x ∈ K the walk X x hits the point ζ x . We also define so that P(R σ (ζ )) R σ (ζ ) for every ζ ∈ V K . Let ζ = (n + c 1 , n + c 1 + c 2 ) K , w,1 = (n, n + c 2 + 1), w,2 = (n + c 2 + 1, ∞), and w = w,1 ∪ w,2 . (Note that these sets are not functions of ζ or w, but rather are the sets from which ζ and w will be drawn.) We also define To be the set of pairs of points at least one of which must have their associated pair of random walks intersect in order for the event E to occur. Define the random variables M σ,0 , M σ,1 , and M σ,2 to be Observe that σ (M σ,1 + M σ,2 ) ≥ 1 on the event C ∩ B ∩ I ∩ E , and so to prove Lemma 4.25 it suffices to prove that for every σ . We will require the following estimate.

Lemma 4.26 The estimate
holds for every (x, x ) ∈ O, every ζ ∈ ζ , every w ∈ w , and every collection Proof Unfortunately, this proof requires a straightforward but tedious case analysis. We will give details for the simplest case, in which both x, x ∈ K . A similar proof applies in the cases that one or both of x or x is not in K , but there are a larger amount of subcases to consider according to when the intersection takes place. In the case that x, x ∈ K , let E −,− (ζ, w), E −,+ (ζ, w), E +,− (ζ, w) and E +,+ (ζ, w) be the events defined as follows: The event R σ (ζ ) occurs, and X x and X x both hit w before they hit ζ x and ζ x respectively. E −,+ (ζ, w): The event R σ (ζ ) occurs, X x hits w before hitting ζ x , and X x hits w after hitting ζ x . E +,− (ζ, w): The event R σ (ζ ) occurs, X x hits w after hitting ζ x , and X x hits w before hitting ζ x . E +,+ (ζ, w): The event R σ (ζ ) occurs, and X x and X x both hit w after they hit ζ x and ζ x respectively.
We have the estimates , and In all cases, a bound of the desired form follows since wx ζ x x and wx ζ x x for every x, x ∈ K , ζ ∈ ζ , and w ∈ w , and we conclude by summing these four bounds.
Our aim now is to prove Eq. (4.26) by an appeal to Lemma 4.3. To do this, we will encode the combinatorics of the potential ways that the walks can intersect via hypergraphs. To this end, let H σ be the finite hypergraph with boundary that has vertex set See Fig. 8 and hence that by Lemma 4.3. Indeed, suppose that H σ / is a proper coarsening of H σ corresponding to some equivalence relation on E(H σ ), and that the edge corresponding to x = σ i ( j) ∈ K is maximal in its equivalence class in the sense that there does not exist σ i ( j ) in the equivalence class of σ i ( j) with j > j. Clearly such a maximal x must exist in every equivalence class. Moreover, for such a maximal x = σ i ( j) there can be at most one edge of H σ that it shares a vertex with and is also in its class, namely the edge corresponding to σ i ( j − 1). Thus, if x is maximal and its equivalence class is not a singleton, let H σ / be the coarsening corresponding to the equivalence relation obtained from by removing x from its equivalence class. Then we have that (H σ / ) ≤ (H σ / ) + 1 and that |E(H σ / )| = |E(H σ / )| + 1, so that (4.30) and the claim follows by inducting on the number of edges in non-singleton equivalence classes.
To obtain a bound on the expectation of M σ,2 , considering the contribution of each shell (m, m + 1) yields the estimate for every ζ ∈ ζ , and it follows from Lemma 4.26 and (4.29) that (4.31) It remains to bound the expectation of M σ,1 . For each two distinct x, x ∈ K , let H σ (x, x ) be the hypergraph with boundary obtained from H σ by adding a single vertex, , and adding this vertex to the two edges corresponding to x and x respectively. These hypergraphs are defined in such a way that, by Lemma  If x x then H σ must be a proper coarsening of H σ , and we deduce from (4.28) that the inequality η d,2 (H σ (x, x )) ≥ η d,2 (H σ ) + 2 holds for every coarsening H σ (x, x ) of H σ (x, x ), yielding the claimed inequality (4.32). Using (4.32), we deduce from

Proof of the main theorems
We now complete the Proof of Theorem 1.5. We begin with the simpler case in which d/(d − 4) is not an integer.
Proof of Theorem 1.5 for d / ∈ {5, 6, 8} We begin by analyzing faithful ubiquity. Let G be a d-dimensional transitive graph, and let H be a finite hypergraph with boundary. If H has a subhypergraph none of whose coarsenings are d-buoyant, then Proposition 4.1 implies that H is not faithfully ubiquitous in C hyp r (F) almost surely for any r ≥ 1.
E(H )\{e 0 }. By the induction hypothesis, every refinement H 1 of H 1 is faithfully ubiquitous in C hyp R G (H 1 ) (F) almost surely. Let H 2 be a refinement of H , and let H 3 be obtained from H 2 by deleting every edge of H 2 which corresponds to e 0 under the refinement. Then H 3 is a refinement of H 1 , and so is faithfully ubiquitous in C hyp R G (H 3 ) (F) almost surely. On the other hand, every edge of H 2 that was deleted to form H 3 has degree at most d/ (d − 4), and since C hyp R G (H 2 ) (F) contains every possible edge of these sizes almost surely, we deduce that H 2 is faithfully ubiquitous on C Proof of Theorem 1. 2 We begin by proving the claim about faithful ubiquity. Applying Theorem 1.4 and Lemma 2.4, and since every subgraph of a tree is a forest, it suffices to prove that if T is a finite forest with boundary then η d (T ) ≥ η d (T ) whenever d ≥ 4 and T is a coarsening of T , so that, in particular, Indeed, suppose that T = T / is a proper coarsening of a finite forest with boundary T . Since T is a finite forest, the subgraph of T spanned by each equivalence class of is also a finite forest, and therefore must contain a leaf. Choose a nonsingleton equivalence class of and an edge e of this equivalence relation that is incident to a leaf of the spanned forest. Thus, e has the property that one of the endpoints of e is not incident to any other edge in e's equivalence class. Let be the equivalence relation obtained from by removing e from its equivalence class and placing it in a singleton class by itself. Then we have that |E(T / )| = |E(T / )|+1 and (T / ) ≤ (T / ) + 1 so that Thus, it follows by induction on the number of edges of T in non-singleton equivalence classes that η d (T / ) ≥ η d (T ) for every coarsening T / of T as claimed. This establishes the claim about faithful ubiquity.
We now turn to ubiquity. Let G be a d-dimensional transitive graph for some d > 8, let r ≥ 1, and let F be the uniform spanning forest of G. Let T be a finite tree with boundary that is not faithfully ubiquitous in C r (F), and let T be a subgraph of T such that (d − 8)|E(T )| − (d − 4)|V • (T )| > 0, which exists by the previous paragraph. Since We easily deduce that η d (S) ≥ η d (T ) > 0, and consequently that S is not faithfully ubiquitous in C r (F) almost surely. On the other hand, since S is a subgraph of H , we have that if H is faithfully ubiquitous in C r (F) almost surely then S is also. Since the quotient H was arbitrary, it follows from Theorem 1.4 that T is ubiquitous in C r (F) if and only if it is faithfully ubiquitous in C r (F) almost surely, completing the proof.  Fig. 5, but is also very easy to prove directly.)

Further questions about the component graph of the USF
It is natural to wonder whether Theorem 1.4 determines the component graph up to isomorphism. It turns out that this is not the case. Indeed, observe that faithful ubiquity of a finite graph with boundary H can be expressed as a first order sentence in the language of graphs: Ubiquity of H can be expressed similarly. However, even if we knew the almost-sure truth value of every first order sentence in the language of graphs, this still would not suffice to determine the graph up to isomorphism. Indeed, recall that a graph G = (V , E) is quasi-k-transitive if the action of its automorphism group on V k has only finitely many orbits. The model-theoretic Ryll-Nardzewski Theorem [11,Theorem 7.3.1] implies that a countably infinite graph is determined up to isomorphism by its first order theory if and only if it is oligomorphic, i.e., quasi-k-transitive for every k ≥ 1. By considering sizes of cliques as in Sect. 6.1, it follows from the discussion in that section that the component graph of the uniform spanning forest of Z d is a.s. not quasi-(d − 4)/(d − 8) -transitive when d > 8, and hence is a.s. not oligomorphic when d > 8. We conjecture that in fact the component graph has very little symmetry indeed.
Conjecture 6.1 Let G be a d-dimensional transitive graph for some d > 8, and let r ≥ 1. Then C r (F) has no non-trivial automorphisms almost surely. Moreover, there does not exist a deterministic graph G such that C r (F) is isomorphic to G with positive probability.
Although we do not believe the component graphs of the USF on different transitive graphs of the same dimension to be isomorphic, it seems nevertheless that most properties of the component graph should be determined by the dimension. One way of formalizing such a statement would be to axiomatize entire the almost-sure first order theory of the component graph of the uniform spanning forest and show that this first order theory is the same for different transitive graphs of the same dimension. We expect that Theorem 1.4, or a slightly stronger variation of it, should play an important role in this axiomatization. See [25] for the development of such a theory in the mean-field setting of Erdős-Rényi graphs. In particular, we believe the following. Conjecture 6.2 Let G 1 and G 2 be d-dimensional transitive graphs, let r 1 , r 2 ≥ 1, and let F 1 and F 2 be the uniform spanning forests of G 1 and G 2 respectively. Then