betaFeedback

WorksMathematics

Internal diameter bounds for DLA on random planar maps under invariance principles

Abstract

We establish sublinear bounds for the internal diameter of External Diffusion Limited Aggregation (DLA) clusters on random planar maps. We develop two conditional DLA growth results for deterministic planar maps: the weaker condition yielding sublinear growth for DLA, and the stronger yielding polynomially sublinear growth for DLA. We then verify the strong assumption a.s. for the γγ-mated CRT maps, and the weak condition a.s. for ergodic scale-free random environments, establishing the corresponding a.s bounds.

Introduction

External Diffusion Limited Aggregation (DLA) is a discrete-time random growth process on a graph, consisting of a sequence of increasing subgraphs, in which new edges and vertices are added at each step according to the harmonic measure from a target vertex or ∞\infty. While a number of variants of DLA have been studied, rigorous results regarding the process have been scarce. This paper is principally concerned with establishing bounds on the size and shape of DLA on random planar geometries as time tends to ∞\infty. Two notions of size can be considered: 1) external diameter, in which the DLA cluster is viewed as a subset of the ambient graph, and diameter of the cluster is measured with respect to the ambient graph metric, and 2) internal diameter, in which the DLA cluster is viewed as a graph in its own right, and the diameter of that graph is measured. The latter notion is more informative of the shape of the DLA cluster. Indeed, consider a spiral curve on a graph versus a radially expanding ball. These increasing subgraphs may have the same external diameter, but from the perspective of the internal diameter these are polar opposites.

While there does exist limited literature on the external diameter of DLA on random planar maps, to our knowledge no sublinear upper bounds have been established for the internal diameter. The purpose of this paper is to derive these bounds on the internal diameter of DLA. The primary tool used in this paper is quenched control of the random walk on the random planar map, i.e. control of the random walk conditional on the random environment. Such properties have been established for large classes of random planar maps (see e.g. [28]), and so the results of this paper are expected to generalize to a number of models.

As the inputs to our arguments are almost sure quenched statements, we typically work with a deterministic underlying geometry, and indeed our main results (Theorems 1.6 and 1.10) are stated as black boxes for deterministic planar maps. Transfer to random planar map models then follows by a simple application of the deterministic theorem to the relevant random planar map model on the appropriate event (§ 1.3).

We deal with two notions of quenched control of random walk: 1) almost sure quenched convergence of random walk to Brownian motion (Assumption 1.4), and 2) almost sure quenched mesoscopic control of random walk (Assumption 1.8). The first assumption is sufficient to establish that the internal diameter of the DLA cluster is o(n)o(n), and the second (stronger) assumption allows one to prove that the internal diameter of the DLA cluster is O(n1−ζ)O(n^{1-\zeta}) with ζ>0\zeta> 0. For our principal example, which is the γ\gamma-mated continuum random tree (CRT) map (Definition 1.11), the exponent ζ\zeta will be related to the Liouville quantum gravity (LQG) parameter γ\gamma.

We now define DLA, and follow the presentation in [29]. Below, V(G)\mathcal{V}(G) denotes the vertices of the graph GG; see § 1.6 for a more detailed discussion of notation.

Definition 1.1 (External DLA). Let GG be an infinite, connected graph, and let v0,v∗∈V(G)v_0, v_* \in\mathcal{V}(G). External DLA from v0v_0 to v∗v_* is a sequence (Kn)n∈N0(K_n)_{n\in\mathbb{N}_0} of random subgraphs of GG with K0={v0}K_0=\{v_0\}, and with KnK_n defined as follows:

  1. If v∗∈Kn−1v_* \in K_{n-1}, terminate the process, i.e. Kn+k−1=Kn−1K_{n+k-1}=K_{n-1} for all k∈Nk\in\mathbb{N}.

  1. Otherwise, run a simple random walk from v∗v_*, conditioned to hit Kn−1K_{n-1}, and consider the edge ene_n first traversed from G∖Kn−1G\setminus K_{n-1} into Kn−1K_{n-1}. Define KnK_n as the union of Kn−1K_{n-1}, ene_n and the endpoint vnv_n of ene_n in G∖Kn−1G\setminus K_{n-1}.

External DLA targeted at ∞\infty is the same process with the harmonic measure from v∗∈Gv_* \in G replaced with harmonic measure from ∞\infty, where such a measure is well defined.

Unless otherwise stated, we work with recurrent rooted graphs (G,o)(G,o), DLA starts from v0=ov_0=o and is targeted at ∞\infty, and harmonic measure from ∞\infty is assumed to be unique and well-defined. A rooted set is a vertex subset containing oo. For a fixed deterministic graph, almost sure statements about DLA refer to its growth randomness.

There are several notions of diameter for which DLA has been studied. [29] concerns the diameter of KnK_n considered as a subset of GG, which is the so-called external diameter, and is denoted diam⁡(Kn;G)\operatorname{diam}(K_n;G). This paper concerns the internal diameter, which considers KnK_n as a graph in its own right, and is denoted diam⁡(Kn;Kn)\operatorname{diam}(K_n;K_n). As this paper only concerns the internal diameter, we write diam⁡(Kn)\operatorname{diam}(K_n) for short.

Definition 1.2 (DLA growth exponent). The (internal) growth exponent of DLA from oo to ∞\infty on a connected rooted graph (G,o)(G,o) is defined as the limit

lim⁡n→∞log⁡diam⁡(Kn)log⁡n\lim_{n\to\infty}\frac{\log\operatorname{diam}(K_n)}{\log n}

when it exists.

We now proceed to state our main results, and the properties on which they rely.

Property 1: Quenched convergence of random walk to Brownian motion

First, we require a few definitions, beginning with an embedding of a graph into C\mathbb{C}. For the definition of a proper embedding in the plane, we follow the presentation of [35 Section 2.4].

Definition 1.3 (Proper embedding in the plane). Define a proper embedding Φ\Phi of a graph GG into the plane as a locally finite topological embedding of vertices V(G)\mathcal{V}(G) and edges E(G)\mathcal{E}(G) into C\mathbb{C}.

Subsequently, we refer to the object in Definition 1.3 as an embedding. We normalize the embedding so that Φ(o)=0\Phi(o)=0. A random walk on GG is mapped into C\mathbb{C} under the embedding in the obvious way; we refer to this as the embedded walk. For finite curves parametrized by [0,1][0,1], set

dF(γ1,γ2):=inf⁡ϕsup⁡t∈[0,1]∣γ1(t)−γ2(ϕ(t))∣d_F(\gamma_1,\gamma_2):=\inf_{\phi}\sup_{t\in[0,1]}\lvert\gamma_1(t)-\gamma_2(\phi(t))\rvert

where ϕ\phi ranges over increasing homeomorphisms of [0,1][0,1]. The function dFd_F induces a metric on curves viewed modulo parametrization, with curves of zero dFd_F-distance considered equivalent [1].

For R,L>0R,L>0, let Yx;R,LY^{x;R,L} be the embedded walk started from xx and stopped at its first exit from BLR(0)B_{LR}(0) (which may be in the interior of an edge). Subsequently, we will choose LL as a radius multiplier and RR as the non-normalized scale for the embedded graph. A normalized scale LL corresponds to the radius r=LRr=LR before rescaling. Let Bz;LB^{z;L} be standard planar Brownian motion started from zz and stopped on leaving BL(0)B_L(0). We write L\mathcal{L} for law and Lx\mathcal{L}_x for law with starting point xx. We can now state the first of the aforementioned invariance principles in the form of Assumption 1.4.

Assumption 1.4 (Random walk convergence hypothesis). For every 0<C<L<∞0<C<L<\infty, we have

sup⁡x:∣Φ(x)∣≤CRdPr(L(R−1Yx;R,L),L(BΦ(x)/R;L))⟶0as R→∞\sup_{x:\lvert\Phi(x)\rvert\le CR}d_{\mathrm{Pr}}\left(\mathcal{L}(R^{-1}Y^{x;R,L}),\mathcal{L}(B^{\Phi(x)/R;L})\right)\longrightarrow0\quad\text{as }R\to\infty

where dPrd_{\mathrm{Pr}} is the Prokhorov distance induced by 1∧dF1\wedge d_F.

The above assumption states that the rescaled embedded random walk converges in law to Brownian motion (modulo parametrization) uniformly over starting points in compact subsets of C\mathbb{C}. Using Assumption 1.4 we can establish an analogue of Kesten’s Beurling estimate on Zd\mathbb{Z}^d. Below, harm⁡A∞\operatorname{harm}^{\infty}_A is the harmonic measure on the edges out of AA viewed from ∞\infty, and EA(v)\mathcal{E}_A(v) is the set of edges out of AA incident to v∈Av\in A; see § 1.6 for a more precise discussion.

Theorem 1.5 (Weak Beurling estimate). Under Assumption 1.4, we have

lim⁡m→∞sup⁡A∋o finite, connected∣A∣≥msup⁡v∈Aharm⁡A∞(EA(v))=0.\lim_{m\to\infty}\sup_{\substack{A\ni o\text{ finite, connected}\\\lvert A\rvert\ge m}}\sup_{v\in A}\operatorname{harm}^{\infty}_A(\mathcal{E}_A(v))=0.

Using Theorem 1.5, we can prove sublinearity of the internal diameter of the DLA cluster. We state this result as a theorem rather than a corollary, as DLA is the focus of this paper.

Theorem 1.6 (Sublinear growth). Let (Kn)(K_n) be DLA started from the root oo and targeted at ∞\infty. Under Assumption 1.4, we have

diam⁡(Kn)=o(n)\operatorname{diam}(K_n)=o(n)

almost surely.

The conclusion also holds for DLA started from any other vertex vv: adjoining a fixed path from oo to vv transfers the harmonic measure bound from sets AA with o∈Ao\in A to sets v∈Av\in A, up to a finite multiplicative constant depending on the path (see Equation (4)).

Property 2: Quenched mesoscopic control of random walk

The assumption and results of this subsection require slightly more information than that of Assumption 1.4. Our planar map, rather than being presented by merely an embedding, is given by a cellular embedding. We combine the presentation from [27, Definition 1.6 and [26].

Definition 1.7 (Cellular embedding in the plane). A cellular embedding into C\mathbb{C} consists of:

  1. A locally finite collection of compact connected subsets of C\mathbb{C} with nonempty interiors, denoted by H\mathcal{H}, such that ⋃A∈HA=C\bigcup_{A\in\mathcal{H}} A = \mathbb{C} and Leb⁡(A∩B)=0\operatorname{Leb}(A \cap B) = 0 for all distinct A,B∈HA, B \in\mathcal{H}.

  2. A planar graph GG, consisting of vertices V(G)\mathcal{V}(G) and unweighted edges E(G)\mathcal{E}(G).

  3. A bijection ϕ\phi from V(G)\mathcal{V}(G) to H\mathcal{H} such that a∼ba \sim b only if ϕ(a)∩ϕ(b)≠∅\phi(a) \cap\phi(b) \ne\varnothing.

We write Hx:=ϕ(x)\mathcal{H}_x := \phi(x) and normalize so that the root cell Ho\mathcal{H}_o contains 00.

For D⊂CD \subset\mathbb{C} and A⊂V(G)A \subset\mathcal{V}(G), write VH(D):={x∈V(G):Hx∩D≠∅}\mathcal{V}_{\mathcal{H}}(D) := \{x \in\mathcal{V}(G) : \mathcal{H}_x \cap D \ne\varnothing\} and UAH:=⋃x∈AHxU_A^{\mathcal{H}} := \bigcup_{x\in A} \mathcal{H}_x. A path is said to disconnect two vertex sets if its vertices intersect every path in the graph joining them.

We choose to present the mesoscopic result in the context of a cellular embedding because it is slightly more natural for the γ\gamma-mated CRT map application, but the subsequent assumption and theorems can be readily adapted to the planar embedding data of Assumption 1.4.

The above definition is implicit in [26], and is a special case of [28 Definition 1.15]. The more general definition, which is called a cell configuration, has in place of items 2 and 3 in Definition 1.7 a symmetric adjacency relation ∼\sim with A∼BA \sim B only if A≠BA \ne B and A∩B≠∅A \cap B \ne\varnothing, and a symmetric function c:{(A,B):A,B∈H and A∼B}→(0,∞)c : \{(A, B) : A, B \in\mathcal{H} \text{ and } A \sim B\} \to(0,\infty) representing conductance. The graph structure can be encoded by c(A,B)=∣{e∈E(G):e is between ϕ−1(A) and ϕ−1(B)}∣c(A, B) = |\{e \in\mathcal{E}(G) : e \text{ is between } \phi^{-1}(A) \text{ and } \phi^{-1}(B)\}|. We leave the adjacency and conductances implicit in the notation H\mathcal{H}, but will discuss the more general notion of cell configuration in § 7, in which we sketch a proof for an analogue of Theorems 1.5 and 1.6 with more general edge weights (called conductances).

We can now introduce the mesoscopic invariance principle assumption. This assumption involves three functions which depend on the non-normalized scale RR:

  1. m(R)m(R), which is the mesoscopic scale lower bound at non-normalized scale RR.

  2. e(R)e(R), which is the net approximation error at non-normalized scale RR.

  3. d(R)d(R), which is the maximal cell diameter at non-normalized scale RR.

Assumption 1.8 (Mesoscopic control hypothesis). There are constants p∈(0,1)p \in(0,1) and σ>0\sigma> 0, a minimal non-normalized scale R0<∞R_0 < \infty, and functions

m,e,d:[R0,∞)→(0,∞)m,e,d : [R_0,\infty) \to(0,\infty)

such that the following hold.

  1. m(R)≤R1−σm(R) \le R^{1-\sigma} and e(R)+d(R)m(R)→0\frac{e(R)+d(R)}{m(R)} \to0.

  1. Every cell intersecting B4R(0)B_{4R}(0) has Euclidean diameter at most d(R)d(R).

  1. For every R≥R0R \ge R_0 there exist finite e(R)e(R)-nets ZR⊂B2R(0)\mathcal{Z}_R \subset B_{2R}(0) and QR⊂[m(R),R/100]Q_R \subset[m(R),R/100] such that for every (z,r)∈ZR×QR(z,r) \in\mathcal{Z}_R \times Q_R and every x∈VH(∂Br(z))x \in\mathcal{V}_{\mathcal{H}}(\partial B_r(z)) a simple random walk started from xx has probability at least pp to disconnect VH(∂B3r/4(z))\mathcal{V}_{\mathcal{H}}(\partial B_{3r/4}(z)) from VH(∂B5r/4(z))\mathcal{V}_{\mathcal{H}}(\partial B_{5r/4}(z)) before exiting VH(B5r/4(z))\mathcal{V}_{\mathcal{H}}(B_{5r/4}(z)), before exiting VH(Br/2,3r/2(z))\mathcal{V}_{\mathcal{H}}(B_{r/2,3r/2}(z)), with the union of cells visited up to and including exit contained in Ar/3,5r/3r/3(z)A_{r/3,5r/3r/3}(z).

  1. D(A):=diam⁡(UAH)≥cvol∣A∣νD(A) := \operatorname{diam}(U_A^{\mathcal{H}}) \ge c_{\mathrm{vol}}|A|^\nu for some cvol,ν>0c_{\mathrm{vol}},\nu> 0, for every sufficiently large finite connected AA containing $o.

We now briefly describe the above assumption in words. Item 1 controls the size of the mesoscopic scale lower bound relative to the non-normalized scale, net error and diameter, item 2 states the cell diameter control enforced by d(R)d(R), item 3 provides a lower bound on the disconnection probability for each fixed scale and center in the nets, and 4 provides control of the diameter of the cellular embedding of a vertex subset in terms of its cardinality which enables the passage to intrinsic graph geometry from embedded graph geometry.

The use of nets in item 3 is to enable a union bound argument which provides a quantitative upper bound on the failure probability of the above assumption for random planar maps, which is necessary to apply Borel-Cantelli and conclude mesoscopic control at all sufficiently large non-normalized scales. In principle, one could strengthen item 3 to hold for all (z,r)∈B2R(0)×[m(R),R/100](z,r) \in B_{2R}(0) \times[m(R),R/100], which would allow one to weaken the limit condition in item 1 to d(R)/m(R)→0d(R)/m(R) \to0, but this becomes more difficult to verify for e.g. the γ\gamma-mated CRT maps.

As with Assumption 1.4, we can prove a Beurling-type estimate using Assumption 1.8, albeit a stronger one than that of Theorem 1.5.

Theorem 1.9 (Strong Beurling estimate). Under Assumption 1.8, there exists θ>0\theta> 0 and a finite threshold N=N(G,o)N=N(G,o) such that

harm⁡A∞(EA(v))≤∣A∣−θ\operatorname{harm}^{\infty}_{A}(\mathcal{E}_{A}(v)) \leq|A|^{-\theta}

for all finite rooted connected AA with ∣A∣>N|A| > N and all v∈Av \in A.

Using Theorem 1.9, we establish the polynomially sublinear bounds on the internal diameter of DLA, again stated as a theorem.

Theorem 1.10 (Polynomially sublinear growth). Let (Kn)(K_n) be DLA started from the root oo and targeted at ∞\infty. Under Assumption 1.8, there exists a deterministic ζ>0\zeta> 0 such that

diam⁡(Kn)=O(n1−ζ)\operatorname{diam}(K_n)=O(n^{1-\zeta})

almost surely.

As with Theorem 1.6, this result can be transferred to an arbitrary initial vertex. The stronger mesoscopic control allows one to upper bound the growth exponent by 1−ζ1-\zeta.

Transfer to random planar maps

The primary random planar map model of interest in this paper will be the so-called mated CRT map, which we now define. We follow the presentation of [26 Section 1.1].

Definition 1.11 (Mated CRT map). Fix γ∈(0,2)\gamma\in(0,2) and let (L,R)(L,R) be a two sided two-dimensional Brownian motion such that

Var⁡(Lt)=Var⁡(Rt)=∣t∣andCov⁡(Lt,Rt)=−∣t∣cos⁡(πγ24)\operatorname{Var}(L_t)=\operatorname{Var}(R_t)=|t| \quad\text{and} \quad\operatorname{Cov}(L_t,R_t)=-|t|\cos\left(\frac{\pi\gamma^2}{4}\right)

and normalized with L0=R0=0L_0=R_0=0. The ϵ\epsilon-mated CRT map for (L,R)(L,R) is the random graph with vertices ϵZ\epsilon\mathbb{Z}, and where vertices x1,x2∈ϵZx_1,x_2 \in\epsilon\mathbb{Z} with x1<x2x_1<x_2 connected by an edge if and only if

(inf⁡t∈[x1−ϵ,x1]Lt)∨(inf⁡t∈[x2−ϵ,x2]Lt)≤inf⁡t∈[x1,x2−ϵ]Lt\left(\inf_{t\in[x_1-\epsilon,x_1]}L_t\right)\vee\left(\inf_{t\in[x_2-\epsilon,x_2]}L_t\right)\leq\inf_{t\in[x_1,x_2-\epsilon]}L_t

or the same condition holds for RR. If both conditions hold for x1+ϵ<x2x_1+\epsilon<x_2, then the vertices are connected by two edges. We denote this random planar map, rooted at o=0o=0, by GϵG^\epsilon. We observe that by identifying x∈ϵZx\in\epsilon\mathbb{Z} with x/ϵ∈Zx/\epsilon\in\mathbb{Z} the law of GϵG^\epsilon does not depend on ϵ\epsilon by Brownian scaling.

When the parameter γ\gamma is obvious or unimportant, we refer to the above graph simply as the mated CRT map. Note that the above definition does not immediately specify a planar embedding of GϵG^\epsilon. However, due to the relationship between the mated CRT map, Schramm-Loewner evolution (SLE), and LQG, it does come equipped with a canonical cellular embedding; this is presented in § 2.2.

By an application of Theorem 1.10 to the mated CRT map a.s., one can achieve an upper bound on the DLA growth exponent (Definition 1.2). As random planar maps are the principal concern of this paper, rather than deterministic ones, we record the consequences of Theorems 1.6 and 1.10 in the following theorem.

Theorem 1.12 (DLA on the γ\gamma-mated CRT). Fix γ∈(0,2)\gamma\in(0,2). Then there exists a ζ=ζ(γ)>0\zeta=\zeta(\gamma)>0 such that DLA started from the root o=0o=0 and targeted at ∞\infty on the γ\gamma-mated CRT map satisfies diam⁡(Kn)=O(n1−ζ)\operatorname{diam}(K_n)=O(n^{1-\zeta}) a.s.

Using [28] one can make a rather strong and general claim: growth of DLA on an arbitrary ergodic scale free random environment (in particular, those obeying the hypothesis of Theorem 1.5 of [28]) is a.s. sublinear. We now define the general class of ergodic scale-free random environments, which we call good embedded lattices, to which our theory applies, following the presentation of Section 1.1 of [28].

Definition 1.13. Let M\mathcal{M} be a random, infinite, locally finite planar, undirected graph (with loops and multiple edges allowed) with conductances c:E(M)→(0,∞)c:\mathcal{E}(\mathcal{M})\to(0,\infty), embedded so that each edge is a simple curve with zero Lebesgue measure, edges intersect only at their endpoints, each compact subset of C\mathbb{C} intersects at most finitely many edges, and each connected component of C∖M\mathbb{C}\setminus\mathcal{M} has compact closure. We call this class of graphs embedded lattices, and denote it by M\mathbb{M}. A good embedded lattice satisfies:

  1. (Translation invariance modulo scaling) For every nonnegative measurable F:M×C2→[0,∞)F:\mathbb{M}\times\mathbb{C}^{2}\to[0,\infty) on the space of doubly-marked embedded lattices such that

F(a(M−z),a(u−z),a(v−z))=a−2F(M,u,v),a>0,z,u,v∈CF(a(\mathcal{M}-z),a(u-z),a(v-z))=a^{-2}F(\mathcal{M},u,v),\qquad a>0,\quad z,u,v\in\mathbb{C}

we have

E[∫CF(M,0,w) dw]=E[∫CF(M,w,0) dw].\mathbb{E}\left[\int_{\mathbb{C}}F(\mathcal{M},0,w)\,dw\right]=\mathbb{E}\left[\int_{\mathbb{C}}F(\mathcal{M},w,0)\,dw\right].
  1. (Ergodicity modulo scaling) Every measurable event invariant under all transformations of the form M↦a(M−z)\mathcal{M}\mapsto a(\mathcal{M}-z) for a>0a>0 and z∈Cz\in\mathbb{C} has probability 00 or 11.

  1. (Finite expectation) Faces are the closures of the connected components of C∖M\mathbb{C}\setminus\mathcal{M}. For z∈Cz\in\mathbb{C}, write HzH_z for a face of M\mathcal{M} containing zz, chosen arbitrarily if there is more than one, and denote the collection of faces by FM\mathcal{F}_{\mathcal{M}}. For distinct faces, write H∼H′H\sim H' if they share an edge, and c(H,H′)c(H,H') for the sum of the conductances of their shared edges. Additionally, define

π(H):=∑H′∈FMH∼H′c(H,H′)andπ∗(H):=∑H′∈FMH∼H′c(H,H′)−1.\pi(H):=\sum_{\substack{H'\in\mathcal{F}_{\mathcal{M}}\\ H\sim H'}}c(H,H')\qquad\text{and}\qquad\pi^*(H):=\sum_{\substack{H'\in\mathcal{F}_{\mathcal{M}}\\ H\sim H'}}c(H,H')^{-1}.

For the origin-containing face H0H_0 we have

E[diam⁡(H0)2Area⁡(H0)(π(H0)+π∗(H0))]<∞.\mathbb{E}\left[\frac{\operatorname{diam}(H_0)^2}{\operatorname{Area}(H_0)}\left(\pi(H_0)+\pi^*(H_0)\right)\right]<\infty.

The main invariance principle pertaining to a good embedded lattice, stated as Proposition 2.8 in the preliminaries, is the convergence of the random walk on the embedded lattice to a non-isotropic Brownian motion [28]. Our methods for isotropic Brownian motion adapt to this setting in a relatively straightforward manner to produce sublinear internal diameter bounds for DLA.

Here, DLA is defined by replacing the simple random walk with the conductance weighted random walk (see Definition 2.7), assuming harmonic measure from infinity is unique.

Theorem 1.14 (DLA on an ergodic scale-free random environment). Let M∗\mathcal{M}^* be the dual of a good embedded lattice M\mathcal{M} rooted at o=H0o=H_0, and let (Kn)(K_n) be External DLA on M∗\mathcal{M}^* started from oo and targeted at ∞\infty on M∗\mathcal{M}^*. Then

diam⁡(Kn)=o(n)\operatorname{diam}(K_n)=o(n)

a.s.

The almost sure conclusions of Theorems 1.12 and 1.14 are joint in the environment and DLA growth. A proof sketch of this result is provided in § 7. Note that in both Theorems 1.12 and 1.14, the required control of the random walk is almost sure convergence of the law of the random walk conditioned on the environment. For a number of random planar maps (RPMs), such as spanning-tree-weighted maps [34], only convergence in probability of the conditionals is available, and one may wonder if the techniques of this paper might be extended to this weakened assumption. In § 8, we present an example which possesses this weaker property, but for which DLA in fact has a.s. linear internal diameter growth. This demonstrates the tightness of the arguments of this paper.

Related work

The literature on random growth processes is broad and diverse, and in particular a number of distinct setups and questions for DLA have been addressed, as well as for continuum analogues and candidate scaling limits. We now briefly review this work, and group these contributions by whether they deal with a discrete or continuum growth process, and a random or deterministic underlying geometry.

Discrete growth on deterministic geometry. Witten and Sander introduced DLA [65, 64] to model processes of accumulation in nature. Prior rigorous work especially relevant to this paper is the work of Kesten on hitting probabilities and so-called long arms [37, 38, 39], which produces Beurling-type bounds on the harmonic measure of the boundary of a connected lattice subset, uniformly over the particular boundary vertex and subsets of a minimal size, and uses these bounds to control the likelihood that the next accumulated particle is adjacent to the most recently accumulated one. Subsequent work has extended the technique of Kesten to various other graph geometries [5, 6, 7, 42]. This line of work heavily exploits the use of the harmonic measure in DLA.

DLA belongs to a more general class of models called the Dielectric Breakdown Models (DBM) [53, 32, 47, 48, 31, 17, 59, 55], in which the harmonic measure is replaced by an η\eta’th power of the harmonic measure; DBM includes models such as the Eden model, which is of independent interest [23, 30, 58, 18, 16, 19, 20]. Recent work achieves η\eta-dependent arm bounds for DBM on Z2\mathbb{Z}^2 and Z3\mathbb{Z}^3, recovering the classical DLA bounds at η=1\eta= 1, by departing to an extent from the technique of Kesten [46]. Other variants of DLA include one-dimensional long-range DLA [2], stationary half-plane DLA [57, 56], and Internal DLA [41, 36, 4]. The latter model is closely related to the Abelian sandpile [11, 12, 14, 13, 45]. Both the work on Beurling-type bounds and on long-arms control the internal diameter of DLA.

Continuum growth on deterministic geometry. A number of continuum aggregation growth models on deterministic geometries have been studied, which provide candidate scaling limits for DLA on deterministic geometries; these may be regarded as “semi-discrete” models, as they are discrete-time models but accumulate continuum segments. The Hastings-Levitov (HL) model is a family of conformal aggregation models [31, 55], which for a particular choice of parameter is expected to be related to DLA [49 Section 1.2.3]; some scaling limit results have been established for these models, as well as for the broader class of growth models called Aggregate Loewner Evolution (ALE) [54, 63, 33, 9], which encompasses the HL model. However, rigorous relationships between HL/ALE and DLA have yet to be achieved.

Discrete growth on random geometry. An early contribution to the study of DLA on random geometries is [7]. For the uniform infinite spanning-tree-weighted random planar map (UITM), Gwynne and Pfeffer prove that the DLA cluster has external diameter n2/d+o(1)n^{2/d+o(1)}, for an exponent dd related to volume of the UITM [29]; their work exploits heavily the relationship between DLA and the Loop-Erased Random Walk (LERW) on the UITM. On the mated CRT maps, which are studied in this paper, shape theorems for internal DLA and the divisible sandpile have been established [15].

Continuum growth on random geometry. Miller and Sheffield introduced Quantum Loewner Evolution (QLE) as a candidate scaling limit for the DBM models on an RPM converging to γ\gamma-LQG [49]. Their construction is motivated by the relationship between the η\eta-DBM model and the η\eta-Laplacian Random Walk (LRW), which generalizes that between DLA and LERW on the UITM. It is expected that DLA on the UITM converges jointly to QLE(2, 1) on a 2\sqrt{2}-LQG surface. The so-called “quantum natural time” QLE(8/3, 0) was constructed and used to construct the 8/3\sqrt{8/3}-LQG metric [50, 51]. Ang and the author subsequently construct quantum natural time QLE for a range of parameters including the (8/3, 0) construction of [50], establish the qualitative phases of the model, as well as stationarity and Markov properties of the unexplored surface. Their time parametrization is conjectured to arise from the time parametrization of the discrete DBM models, but thus far, aside from the Eden model on uniform RPMs (see [3]), convergence of DBM models on RPMs to QLE on LQG is not established.

Outline

The proof strategy of both Theorems 1.6 and 1.10 is identical: bound the maximal harmonic measure (from ∞\infty) of EKn(v)E_{K_n}(v) over any v∈Knv \in K_n. Indeed, this is identified as the core mechanism for internal diameter bounds by Gwynne and Pfeffer in [29 Section 8]. To do this, one needs to identify the size of annulus AA required such that the probability that the random walk, upon entering AA, disconnects the relevant vertex sets before exiting AA, possesses a positive lower bound. The larger the number of disjoint annuli which can be identified, the stronger the upper bound on harmonic measure, which is what produces the differing conclusions of Theorems 1.6 and 1.10.

In § 2, we present preliminaries on the mated CRT map required to verify Assumption 1.8. In § 3 we prove a bound on the internal diameter of DLA conditional on an annular intersection probability via a harmonic measure bound. In § 4 and § 5 we use the blackbox theorem of § 3 to establish Theorems 1.6 and 1.10 under assumptions 1.4 and 1.8 respectively. In § 6, we verify Assumption 1.8. In § 7 we sketch the proof of Theorem 1.14. Lastly, in § 8 we discuss open problems related to DLA on RPMs, including the mated CRT and UITM, and present a counterexample illustrating an obstruction to a straight-forward application of known, weaker invariance principles to the study of DLA.

Notation and terminology

Write N={1,2,…}\mathbb{N} = \{1,2,\ldots\} and N0:=N∪{0}\mathbb{N}_0 := \mathbb{N}\cup\{0\}. For K∈NK \in\mathbb{N}, write [K]:={1,…,K}[K] := \{1,\ldots,K\}. We identify a subgraph with its vertex set when using vertex-set notation, and ∣A∣|A| denotes vertex cardinality. Graph distance is denoted by dist⁡(⋅,⋅;G)\operatorname{dist}(\cdot,\cdot;G); for subsets of C\mathbb{C}, dist⁡\operatorname{dist} and diam⁡\operatorname{diam} are Euclidean. Edges are unoriented, and parallel edges are distinct. Simple random walk chooses uniformly among incident edge ends, counting each self-loop twice. We denote by Ar,s(z)A_{r,s}(z) the Euclidean annulus Bs(z)∖Br(z)B_s(z) \setminus B_r(z) for r<sr < s and z∈Cz \in\mathbb{C}. When z=0z = 0, we drop the center and write Ar,sA_{r,s}. For v∈Gv \in G, define the neighborhood N(v):={x∈V(G):x∼v}∪{v}N(v) := \{x \in V(G) : x \sim v\} \cup\{v\}, and for a subset A⊂V(G)A \subset V(G) with v∈Av \in A, define the external neighborhood SA(v):={x∈V(G)∩Ac:x∼v}S_A(v) := \{x \in V(G) \cap A^c : x \sim v\}. Additionally, we let EA(v):={e∈E(G):e joins v to SA(v)}E_A(v) := \{e \in E(G) : e\text{ joins }v\text{ to }S_A(v)\}, i.e. edges with one end on vv and the other outside of AA. For finite nonempty AA and x∉Ax \notin A, harm⁡Ax\operatorname{harm}_A^x is the law of the edge traversed when a random walk from xx, conditioned to hit AA, first enters AA; harm⁡A∞\operatorname{harm}_A^\infty is its limit as dist⁡(x,A;G)→∞\operatorname{dist}(x,A;G) \to\infty, when this limit exists. For the plane data K={Jx:x∈V(G)}\mathcal{K} = \{J_x : x \in V(G)\} of § 3.1 and finite connected A⊂V(G)A \subset V(G), set UAK:=⋃x∈AJxU_A^{\mathcal{K}} := \bigcup_{x \in A} J_x. Note that the plane data K\mathcal{K} is a cell configuration H\mathcal{H}, UAKU_A^{\mathcal{K}} is connected. Set D(A):=diam⁡(UAH)D(A) := \operatorname{diam}(U_A^{\mathcal{H}}).

Acknowledgements The author would like to thank Ewain Gwynne for helpful conversations throughout the duration of this project, and Leonardo Bonanno for preparing the associated formalization of this paper.

AI Disclosure ChatGPT 5.6 and 6 Pro were used to expand a detailed proof sketch into an initial draft early in this project, for subsequent proof revisions, to generate Figures 1 and 2, and for the editing of prose, proofs and \LaTeX. All results have been verified by the author, and any errors herein are the sole responsibility of the author. A formalization of this paper prepared by Leonardo Bonanno is available in [10].

Preliminaries

Now, we discuss in greater detail the relationship between the mated CRT map and SLE/LQG, after which we present some preliminary results on the mated CRT map and ergodic scale-free random environments. Before that, we define SLE, the Gaussian free field (GFF) and LQG.

Basics of LQG and SLE

Now, we describe the basics of certain variants of SLE, GFF and LQG. We cover only the material necessary to describe the canonical embedding and cellular embedding with which the mated CRT map comes equipped. Since we rely on many established properties of the mated CRT map, but do not exploit the relationship between LQG/SLE and the mated CRT map very directly, we keep our presentation brief, and only state those definitions and properties necessary to interpret the relevant mated CRT properties. Further references on SLE and LQG are given. In what follows, fix γ∈(0,2)\gamma\in(0,2) and set Q:=2/γ+γ/2Q := 2/\gamma+ \gamma/2 and κ:=16/γ2\kappa:= 16/\gamma^2.

Space filling SLE

SLE is a family of random fractal curves that arise as scaling limits of interfaces in critical lattice models [60, 61, 44, 62] and are indexed by a parameter κ>0\kappa> 0, which varies the “roughness” of the curve. We give a very brief introduction to these random curves here, and refer to [40] for further details.

In this paper, we deal exclusively with a variant of SLE known as whole-plane space-filling SLEκ_\kappa from ∞\infty to ∞\infty with κ=16/γ2\kappa= 16/\gamma^2 when γ∈(0,2)\gamma\in(0,2) (see e.g. [21 Section 1.4.1] for the original definition in the whole-plane case, or [26 Section 2.1.3] for a more recent discussion). This is a random continuous curve η:R→C\eta: \mathbb{R} \to\mathbb{C} with η(R)=C\eta(\mathbb{R}) = \mathbb{C} and ∣η(t)∣→∞|\eta(t)| \to\infty as ∣t∣→∞|t| \to\infty. The space-filling construction is identical to ordinary whole-plane SLE when κ≥8\kappa\ge8, and is a slight modification when 4<κ<84 < \kappa< 8. Its law is invariant under translations, rotations, and positive dilations.

Whole plane GFF and LQG

Now, we introduce the GFF and LQG. Let H(C)\mathcal{H}(\mathbb{C}) be the completion of {f∈Cc∞(C):∫Cf(z) dz=0}\{f \in C^\infty_c(\mathbb{C}) : \int_{\mathbb{C}} f(z)\,dz = 0\} for the Dirichlet inner product (f,g)∇:=(2π)−1∫C∇f⋅∇g dz(f,g)_\nabla:= (2\pi)^{-1}\int_{\mathbb{C}} \nabla f \cdot\nabla g\,dz. Given an orthonormal basis (fj)(f_j) and independent standard Gaussian variables (αj)(\alpha_j), the formal sum

h:=∑j≥1αjfjh := \sum_{j\ge1} \alpha_j f_j

defines the whole-plane GFF, a random distribution modulo additive constants. For each f∈H(C)f \in\mathcal{H}(\mathbb{C}), the pairing (h,f)∇:=∑j≥1αj(fj,f)∇(h,f)_\nabla:= \sum_{j\ge1}\alpha_j(f_j,f)_\nabla converges in L2L^2, and E[(h,f)∇(h,g)∇]=(f,g)∇\mathbb{E}[(h,f)_\nabla(h,g)_\nabla] = (f,g)_\nabla; see [26 Section 2.1.1].

For z∈Cz \in\mathbb{C} and r>0r > 0, let hr(z)h_r(z) denote the average of hh with respect to the uniform probability measure on ∂Br(z)\partial B_r(z), defined in [22 Section 3.1]. We fix the additive constant by normalizing h1(0)=0h_1(0) = 0.

For a conformal map f:D→D~f : D \to\widetilde D, define f∙γh:=h∘f−1+Qlog⁡∣(f−1)′∣f \mathbin{\bullet_\gamma} h := h \circ f^{-1} + Q\log|(f^{-1})'|. Write (D,h)∼γ(D~,h~)(D,h) \sim_\gamma(\widetilde D,\widetilde h) if h~=f∙γh\widetilde h = f \mathbin{\bullet_\gamma} h for such an ff. A γ\gamma-LQG surface is an equivalence class for this relation, and a representative is called an embedding. For marked surfaces, ff must send each marked point to its corresponding marked point; see e.g. [3 Section 2.3].

The LQG area measure is formally

Ahγ:=lim⁡ϵ→0ϵγ2/2eγhϵ(z) dz.A^\gamma_h := \lim_{\epsilon\to0}\epsilon^{\gamma^2/2}e^{\gamma h_\epsilon(z)}\,dz.

where the convergence is a.s. in the local weak topology on measures on DD along dyadic ϵ\epsilon. It satisfies Af∙γhγ=f∗AhγA^\gamma_{f\mathbin{\bullet_\gamma}h} = f_*A^\gamma_h and Ah+cγ=eγcAhγA^\gamma_{h+c} = e^{\gamma c}A^\gamma_h, and is therefore well defined for a γ\gamma-LQG surface (see [22 Sections 2 and 3]). We use Area⁡\operatorname{Area} for the Lebesgue measure.

To study the mated CRT map, we will need a variant of LQG surface called an α\alpha-quantum cone with α∈{0,γ}\alpha\in\{0,\gamma\}. Let BB be standard one-dimensional Brownian motion and, independently of BB, let B^\widehat B have the law of Brownian motion conditioned so that B^t+(Q−α)t>0\widehat B_t + (Q-\alpha)t > 0 for all t>0t > 0. Let

At:={Bt+αt,t≥0,B^−t+αt,t<0.A_t := \begin{cases} B_t + \alpha t, & t \ge0,\\ \widehat B_{-t} + \alpha t, & t < 0. \end{cases}

The circle-average embedding of the α\alpha-quantum cone is defined as the field hh for which t↦he−t(0)t \mapsto h_{e^{-t}}(0) has the law of AtA_t, while h−h∣⋅∣(0)h-h_{|\cdot|}(0) is independent of AtA_t and agrees in law with the corresponding projection of a whole-plane GFF. This field satisfies sup⁡{r>0:hr(0)+Qlog⁡r=0}=1\sup\{r > 0 : h_r(0) + Q\log r = 0\} = 1 (see [21 Definition 4.10] and [27 Section 2.3.1]). In this paper, we will take α∈{0,γ}\alpha\in\{0,\gamma\}.

The mated CRT and LQG/SLE

Let hh be the circle-average embedding of the γ\gamma-quantum cone defined above, and η\eta an independent whole-plane space-filling SLEκ_\kappa. We parametrize η\eta by γ\gamma-LQG area: η(0)=0\eta(0) = 0 and Ahγ(η([s,t]))=t−sA^\gamma_h(\eta([s,t])) = t-s for s<ts < t.

This parametrization is the so-called quantum natural time parametrization of SLE, and has a number of useful properties (see e.g. [3 Section 2.5]). For ϵ>0\epsilon> 0 define

Hxϵ:=η([x−ϵ,x]),x∈ϵZ.H_x^\epsilon:= \eta([x-\epsilon,x]), \qquad x \in\epsilon\mathbb{Z}.

These are compact connected sets of measure ϵ\epsilon under Ahγ\mathcal{A}_h^\gamma. The structure graph of the cells {Hxϵ}\{H_x^\epsilon\} is defined as follows: two distinct vertices are adjacent when their cells share a nontrivial boundary arc, consecutive indices are joined by one edge, nonconsecutive indices are joined by two edges when both left and right boundary adjacency occurs [26 Section 1.2]. The resulting graph has the law of the whole-plane γ\gamma-mated-CRT map GϵG^\epsilon (see e.g. [27 Proposition 2.2], [26 Section 1.2]). This observation allows one to study the γ\gamma-mated CRT map via the γ\gamma-quantum cone and space filling SLE, and to couple the γ\gamma-mated CRT maps for different ϵ\epsilon.

For D⊂CD \subset\mathbb{C}, write VGϵ(D):={x∈ϵZ:Hxϵ∩D≠∅}V_{G^\epsilon}(D) := \{x \in\epsilon\mathbb{Z}: H_x^\epsilon\cap D \ne\varnothing\}. At unit mesh, the root is o=0o=0; set Hx:=Hx1H_x := H_x^1, H:={Hx:x∈Z}\mathcal{H} := \{H_x:x\in\mathbb{Z}\}, and UAH:=⋃x∈AHxU^\mathcal{H}_A := \bigcup_{x\in A}H_x. Distinct cells overlap only on sets of zero quantum area, so Ahγ(UAH)=∣A∣\mathcal{A}_h^\gamma(U^\mathcal{H}_A)=|A| for finite A⊂ZA\subset\mathbb{Z}. This is the cellular embedding used in the verification of Assumption 1.8.

Properties of the mated CRT map

We retain the SLE/LQG realization and notation of the preceding subsection and record the properties needed for the mesoscopic argument. We start with control of the random walk and harmonic measure.

Proposition 2.1 ([26 Theorem 1.4]). For every ϵ>0\epsilon> 0, the simple random walk on GϵG^\epsilon is recurrent almost surely.

We denote the total variation (TV) distance by ∥⋅∥TV\|\cdot\|_{\mathrm{TV}}.

Proposition 2.2 ([8 Theorem 1.1 and Section 9], [29 Proposition 1.7]). Almost surely, for every finite nonempty connected A⊂ZA\subset\mathbb{Z} we have

lim⁡m→∞sup⁡w∈V(G1)dist⁡(w,A;G1)≥m∥harm⁡Aw−harm⁡A∞∥TV=0\lim_{m\to\infty}\sup_{\substack{w\in V(G^1)\\ \operatorname{dist}(w,A;G^1)\ge m}}\left\|\operatorname{harm}_A^w-\operatorname{harm}_A^\infty\right\|_{\mathrm{TV}}=0

where harm⁡A∞\operatorname{harm}_A^\infty is a unique measure on ⋃v∈AE(v)\bigcup_{v\in A}\mathcal{E}(v).

For b>0b>0, define Rb:=sup⁡{r>0:hr(0)+Qlog⁡r=γ−1log⁡b}R_b:=\sup\{r>0:h_r(0)+Q\log r=\gamma^{-1}\log b\}, hb(⋅):=h(Rb⋅)+Qlog⁡Rb−γ−1log⁡bh^b(\cdot):=h(R_b\cdot)+Q\log R_b-\gamma^{-1}\log b, and ηb(t):=Rb−1η(bt)\eta^b(t):=R_b^{-1}\eta(bt).

Proposition 2.3 ([21 Proposition 4.13(i)]). For every b>0b>0 we have

(hb,ηb)=d(h,η).(h^b,\eta^b)\mathrel{\overset{d}{=}}(h,\eta).

As a result, the 1/b1/b-cell configuration for (hb,ηb)(h^b,\eta^b) is, after multiplication by RbR_b, the 11-cell configuration for (h,η)(h,\eta).

Define

aγ:=1γ(Q−γ)=12−γ2/2.a_\gamma:=\frac{1}{\gamma(Q-\gamma)}=\frac{1}{2-\gamma^2/2}.

The following result from [25] controls the asymptotics of RbR_b.

Proposition 2.4 ([25 Lemma 2.6 and the proof of Theorem 1.4]). Almost surely,

lim⁡k→∞log⁡R2kklog⁡2=aγ.\lim_{k\to\infty}\frac{\log R_{2^k}}{k\log2}=a_\gamma.

Next, we need control over the number of vertices in BRb(0)B_{R_b}(0) with b=2kb=2^k.

Proposition 2.5 ([25 proof of Theorem 1.4]; [24 Lemma A.3]). For every δ>0\delta>0, almost surely

∣VG1(BR2k(0))∣≤2(1+δ)k\left|V_{G^1}\left(B_{R_{2^k}}(0)\right)\right|\le2^{(1+\delta)k}

for all sufficiently large kk.

The above estimate appears immediately after Equation (3.31) in [25] with probability at least 1−O(2−αk)1-O(2^{-\alpha k}) for some α>0\alpha>0, and Borel-Cantelli gives the a.s. claim.

Mesoscopic control of random walk for mated CRT

The following estimate is extracted from [26 Proposition 3.6, Lemma 2.7] and the proof of [26 Lemma 3.11].

Proposition 2.6. There exist constants

ρ∈(0,1),p,α,β,χ,q>0,β<min⁡{χ,q},q<2(2+γ)2\rho\in(0,1), \qquad p,\alpha,\beta,\chi,q>0, \qquad\beta<\min\{\chi,q\}, \qquad q<\frac{2}{(2+\gamma)^2}

such that the following holds. Define

Zε:=(εχZ2)∩Bρ(0),Rε:=(εχZ)∩[εβ,ρ].(1)Z_\varepsilon:= (\varepsilon^\chi\mathbb{Z}^2)\cap B_\rho(0), \qquad R_\varepsilon:= (\varepsilon^\chi\mathbb{Z})\cap[\varepsilon^\beta,\rho]. \tag*{(1)}

The following hold simultaneously with probability at least 1−Oε(εα)1-O_\varepsilon(\varepsilon^\alpha):

  1. Every cell intersecting Bρ(0)B_\rho(0) has Euclidean diameter at most εq\varepsilon^q.

  1. For every (z,r)∈Zε×Rε(z,r)\in Z_\varepsilon\times R_\varepsilon with B2r(z)⊂Bρ(0)B_{2r}(z)\subset B_\rho(0), and for every x∈VGε(∂Br(z))x\in V_{\mathcal{G}_\varepsilon}(\partial B_r(z)), let XnX_n be simple random walk started from xx and τ\tau its first exit time from VGε(Ar/2,3r/2(z))V_{\mathcal{G}_\varepsilon}(A_{r/2,3r/2}(z)). With quenched probability at least pp we have:

(a) The random walk disconnects VGε(∂B3r/4(z))V_{\mathcal{G}_\varepsilon}(\partial B_{3r/4}(z)) from VGε(∂B5r/4(z))V_{\mathcal{G}_\varepsilon}(\partial B_{5r/4}(z)) before τ\tau.

(b) Every vertex visited before τ\tau belongs to VGε(Ar/2,3r/2(z))V_{\mathcal{G}_\varepsilon}(A_{r/2,3r/2}(z)).

(c) The union of cells visited up to and including τ\tau is contained in Ar/3,5r/3(z)A_{r/3,5r/3}(z).

The proof of [26 Lemma 3.11] applies [26 Proposition 3.6] to obtain (2a) and (2b) with quenched success probability at least p>0p>0. The final union bound makes these estimates simultaneous over the grids in (1): choosing χ>0\chi>0 sufficiently small and β\beta below both χ\chi and the lower-scale exponent in that lemma gives a failure probability at most Oε(εα)O_\varepsilon(\varepsilon^\alpha). [26 Lemma 2.7] gives (1), after choosing β<q<2/(2+γ)2\beta<q<2/(2+\gamma)^2 and decreasing α\alpha if necessary. Since every admissible radius satisfies r≥εβr\ge\varepsilon^\beta, one has εq=o(r)\varepsilon^q=o(r). So, (1) and (2b) imply that the union of cells visited up to and including τ\tau is contained in Ar/3,5r/3(z)A_{r/3,5r/3}(z) for all sufficiently small ε\varepsilon, which establishes (2c).

Ergodic scale-free random environments

Lastly, we introduce the invariance principle for ergodic scale-free random environments. First, we need a definition.

Definition 2.7. For a good embedded lattice M\mathcal{M}, the conductance weighted random walk on M∗\mathcal{M}^* has transition probabilities P(H,H′)=c(H,H′)/π(H)P(H,H')=c(H,H')/\pi(H). Here M∗\mathcal{M}^* is viewed as a weighted graph with vertices given by the faces of M\mathcal{M}, with one edge of conductance c(H,H′)c(H,H') for each adjacent pair of faces. This Markov chain on the faces of M\mathcal{M} is the random walk defined in [28 Theorem 3.10].

The next result we recall is a main result from [28]; it is the invariance principle for ergodic-scale free random environments. In particular, this proposition combines Theorems 1.5 and 3.10 from [28].

Proposition 2.8. Let M\mathcal{M} be a good embedded lattice (see Definition 1.13), and choose a point pH∈int⁡(H)p_H\in\operatorname{int}(H) from each face, with pH0=0p_{H_0}=0. Let (Xn)(X_n) be the conductance weighted random walk of M∗\mathcal{M}^* in Definition 2.7. For a starting face HH, let Y^H;R,L\widehat{Y}^{H;R,L} be the piecewise-linear interpolation of n↦pXnn\mapsto p_{X_n}, started from pHp_H and stopped upon first leaving BLR(0)B_{LR}(0). There exists a positive-definite covariance matrix Σ\Sigma such that, a.s. in M\mathcal{M}, for every 0<C<L<∞0<C<L<\infty, we have

sup⁡H:∣pH∣≤CRdPr(LH(R−1Y^H;R,L∣M),L(BΣpH;R/L))→0as R→∞\sup_{H:\lvert p_H\rvert\le CR}d_{\mathrm{Pr}}\left(\mathcal{L}_{H}\left(R^{-1}\widehat{Y}^{H;R,L}\mid\mathcal{M}\right),\mathcal{L}\left(B_{\Sigma}^{p_H;R/L}\right)\right)\to0\quad\text{as }R\to\infty

where BΣz;LB_{\Sigma}^{z;L} is Brownian motion with covariance matrix Σ\Sigma started from zz and stopped upon first leaving BL(0)B_L(0), and dPrd_{\mathrm{Pr}} is induced by 1∧dF1\wedge d_F. Convergence is again with respect to the topology on curves modulo time parametrization.

Now we proceed to the proofs of the blackbox results.

Control of diameter via annular intersection probabilities

In this section, we demonstrate how to convert annular intersection probabilities into harmonic measure bounds, and harmonic measure bounds into diameter bounds. Sections 4 and 5 will invoke these intermediary results to establish the main theorems. The section concludes with a brief subsection describing the change of initial vertex of the DLA growth.

Annular intersection probabilities to harmonic measure bounds

The results of this section are stated at a fixed scale. In the next subsection, these results will be applied at an arbitrary scale, yielding the large nn conclusions.

Assumptions 1.4 and 1.8 pertain to slightly different embedding data for the planar map (a proper planar embedding versus a cellular embedding), and the arguments obey this structure. To reconcile these differences, we define the plane data as follows: let K={Jx:x∈V(G)}\mathcal{K} = \{J_x : x \in V(G)\} be a locally finite collection of nonempty bounded connected subsets of C\mathbb{C}, indexed by V(G)V(G), such that Jx∩Jy≠∅J_x \cap J_y \ne\varnothing if x∼yx \sim y. In § 4, JxJ_x will be the union of {Φ(x)}\{\Phi(x)\} and the interiors of the embedded edges incident to xx, and in § 5 we choose K\mathcal{K} as the cellular embedding. For D⊂CD \subset\mathbb{C}, define

VK(D):={x∈V(G):Jx∩D≠∅}V_{\mathcal{K}}(D) := \{x \in V(G) : J_x \cap D \ne\varnothing\}

as the vertices with planar data intersecting a given Euclidean subset.

Now, we state precisely the multiscale structure, at a fixed non-normalized scale, which will yield asymptotically shrinking harmonic measure bounds.

Definition 3.1 (Good scales). Fix K∈NK \in\mathbb{N} and p∈(0,1)p \in(0,1), and let A⊂V(G)A \subset V(G) be finite and connected. Choose v∈Av \in A. (K,p)(K,p)-good scales for (A,v)(A,v) consist of radii

r1>r2>⋯>rK>0,rj+1≤rj8,(2)r_1 > r_2 > \cdots> r_K > 0,\qquad r_{j+1} \le\frac{r_j}{8}, \tag*{(2)}

and a center z∈Cz \in\mathbb{C} such that:

  1. For all JxJ_x such that Jx∩B2r1(z)≠∅J_x \cap B_{2r_1}(z) \ne\varnothing we have diam⁡(Jx)<rK/100\operatorname{diam}(J_x) < r_K/100.

  2. ⋃x∈N(v)Jx⊂BrK/20(z)\bigcup_{x \in N(v)} J_x \subset B_{r_K/20}(z).

  3. For all j∈[K]j \in[K], there exists a graph path in AA joining VK(∂B3rj/4(z))V_{\mathcal{K}}(\partial B_{3r_j/4}(z)) to VK(∂B5rj/4(z))V_{\mathcal{K}}(\partial B_{5r_j/4}(z)).

  4. For all j∈[K]j \in[K] and all x∈VK(∂Brj(z))x \in V_{\mathcal{K}}(\partial B_{r_j}(z)), a simple random walk XnX_n started from xx and run until its first exit time τ\tau from Wj:=VK(Arj/2,3rj/2(z))W_j := V_{\mathcal{K}}(A_{r_j/2,3r_j/2}(z)) has probability at least pp that it disconnects VK(∂B3rj/4(z))V_{\mathcal{K}}(\partial B_{3r_j/4}(z)) from VK(∂B5rj/4(z))V_{\mathcal{K}}(\partial B_{5r_j/4}(z)) before τ\tau and ⋃m=0τJXm⊂Arj/3,5rj/3(z)\bigcup_{m=0}^{\tau} J_{X_m} \subset A_{r_j/3,5r_j/3}(z).

Note that the above definition admits at most logarithmically many good scales relative to the ratio r1/rKr_1/r_K. In principle one could achieve K≫log⁡(r1/rK)K \gg\log(r_1/r_K) without (3.1), but we do not encounter this scenario in this paper and so work with the slightly more restrictive definition.

As we organize scales in a Euclidean fashion, i.e. using Euclidean annuli, the bound on diameters of JxJ_x establishes a compatibility condition of the plane data with the good scales, and ensures that a random walk from ∞\infty must pass through each individual scale successively. Each scale then provides with probability at least pp the possibility of an annular disconnection, and combining with the Markov property of the random walk yields the final harmonic measure bound. This is illustrated in Figure 1, and is the content of the next proposition.

A planar graph with three disjoint annuli around $v$

Figure 1. (K,p)(K,p)-good scales for the pair (A,v)(A,v) with K=3K=3. The good scales consist of 3 disjoint annuli A1,A2A_1,A_2 and A3A_3, with inner and outer boundaries colored blue and red respectively. A segment of the ambient graph GG is drawn with gray vertices and edges, and a segment of AA is drawn with black vertices and edges.

Proposition 3.2. Suppose there exist (K,p)(K,p)-good scales for (A,v)(A,v). Then for all x∈V(G)∖Ax \in V(G) \setminus A such that Jx∩B2r1(z)=∅J_x \cap B_{2r_1}(z) = \varnothing, we have

harm⁡A∞(EA(v))≤(1−p)K.\operatorname{harm}^{\infty}_A(\mathcal{E}_A(v)) \le(1-p)^K.

and the bound follows for x=∞x=\infty.

Proof. Items 1-4 in Definition 3.1 identify KK concentric, disjoint, nonempty annuli in GG through which a random walk started outside of VK(B2r1(z))V_{\mathcal{K}}(B_{2r_1}(z)) must pass, and for which a random walk started anywhere on the boundary has a disconnection probability of at least pp.

In particular, item 2 states that vv and its immediate neighborhood are contained in the center of the concentric annuli, item 3 guarantees that a random walk trajectory disconnecting VK(∂B3rj/4(z))\mathcal{V}_K(\partial B_{3r_j/4}(z)) from VK(∂B5rj/4(z))\mathcal{V}_K(\partial B_{5r_j/4}(z)) must hit the cluster AA, item 4 lower bounds the annulus intersection probability by pp and ensures that the visited plane data is contained in the annulus, and item 1 ensures that the random walk cannot skip scales and that the sets VK(Arj/3,5rj/3(z))\mathcal{V}_K(A_{r_j/3,5r_j/3}(z)) are disjoint for differing jj.

As a random walk which first enters AA through EA(v)E_A(v) must avoid all KK disconnection events, one applies the Markov property of the random walk successively at each scale jj to conclude the final bound on the harmonic measure.

To produce a diameter bound, we will need to show the existence of (K,p)(K,p)-good scales uniformly over all v∈Av\in A. The next section shows how to convert such uniform good scales into a diameter bound.

Harmonic measure bounds to diameter bounds

We now use harmonic measure bounds to establish a diameter bound on the DLA cluster. Proposition 3.2 operates at a fixed scale, but the next proposition takes as input harmonic measure bounds which tend to 00 as the size of the subject cluster tends to ∞\infty. Establishing these bounds at arbitrarily large scales, uniformly over clusters and vertices, is the main content of § 4 and § 5, and is done using the appropriate assumption in tandem with Proposition 3.2. We prove Proposition 3.3 here since, as with Proposition 3.2, it does not inherently depend on either assumption.

Suppose that GG is a planar map with root o=v0o=v_0, and let K0={o}K_0=\{o\}, and KnK_n the DLA cluster at nn steps, evolving according to Definition 1.1. KnK_n has a tree structure, and we set height⁡(Kn):=max⁡v∈Kndist⁡(o,v;Kn)\operatorname{height}(K_n):=\max_{v\in K_n}\operatorname{dist}(o,v;K_n). Define the (non-increasing) sequence

an=an(G,o):=sup⁡A∋o finite, connected∣A∣≥nsup⁡v∈Aharm⁡A∞(EA(v)).(3)a_n=a_n(G,o):=\sup_{\substack{A\ni o\ \mathrm{finite,\ connected}\\ |A|\ge n}}\sup_{v\in A}\operatorname{harm}^{\infty}_{A}(E_A(v)). \tag*{(3)}

Note that a priori the above quantity has nothing to do with DLA, and applies to all sufficiently large finite rooted connected subgraphs of GG. The convergence of an→0a_n\to0, and in particular the rate of this convergence, is supplied by the Beurling estimates in Theorems 1.5 and 1.9; these latter results are established in subsequent sections, but here we show how to convert this harmonic measure control to internal diameter bounds on DLA.

Proposition 3.3. With ana_n as in (3), we have almost surely:

  1. If an→0a_n \to0 then diam⁡(Kn)=o(n)\operatorname{diam}(K_n)=o(n).

  1. If an≤Cn−θa_n \le Cn^{-\theta} for some C,θ>0C,\theta>0, then there exists a ζ>0\zeta>0 such that diam⁡(Kn)=O(n1−ζ)\operatorname{diam}(K_n)=O(n^{1-\zeta}).

Proof. By definition, for each v∈Kn−1v\in K_{n-1} the conditional probability that en∈EKn−1(v)e_n\in\mathcal{E}_{K_{n-1}}(v) given Kn−1K_{n-1} is at most ana_n, since ∣Kn−1∣=n|K_{n-1}|=n. Let sn:=∑j=1najs_n:=\sum_{j=1}^n a_j. If height⁡(Kn)≥k\operatorname{height}(K_n)\ge k, then there exist times 1≤t1<⋯<tk≤n1\le t_1<\cdots<t_k\le n such that v0,vt1,…,vtkv_0,v_{t_1},\ldots,v_{t_k} form a simple path from the root. For any fixed increasing sequence (t1,…,tk)(t_1,\ldots,t_k), successive conditioning on the growth history produces the bound ∏j=1katj\prod_{j=1}^k a_{t_j} on the event ⋂1≤j≤k−1{vtj∼vtj+1}∩{v0∼vt1}\bigcap_{1\le j\le k-1}\{v_{t_j}\sim v_{t_{j+1}}\}\cap\{v_0\sim v_{t_1}\}, where adjacency ∼\sim is in KnK_n. A union bound over all such sequences gives

P[height⁡(Kn)≥k]≤∑1≤t1<⋯<tk≤n∏j=1katj≤snk/k!≤(esn/k)k.\mathbb{P}[\operatorname{height}(K_n)\ge k]\le\sum_{1\le t_1<\cdots<t_k\le n}\prod_{j=1}^k a_{t_j}\le s_n^k/k!\le(es_n/k)^k.

The second inequality follows by expanding snks_n^k, and the last follows from k!≥(k/e)kk!\ge(k/e)^k.

First, suppose that an→0a_n\to0. Then sn/n→0s_n/n\to0 and so sn=o(n)s_n=o(n). Fix c>0c>0 and take k=⌈cn⌉k=\lceil cn\rceil. For all sufficiently large nn one has esn/⌈cn⌉≤1/2es_n/\lceil cn\rceil\le1/2, and so P[height⁡(Kn)≥⌈cn⌉]≤2−⌈cn⌉\mathbb{P}[\operatorname{height}(K_n)\ge\lceil cn\rceil]\le2^{-\lceil cn\rceil}. Since ∑n∈NP[height⁡(Kn)≥⌈cn⌉]<∞\sum_{n\in\mathbb{N}}\mathbb{P}[\operatorname{height}(K_n)\ge\lceil cn\rceil]<\infty, by Borel-Cantelli we have height⁡(Kn)≤O(cn)\operatorname{height}(K_n)\le O(cn) a.s., with an implied constant independent of cc. By intersecting these probability-one events over all positive rational cc, we conclude height⁡(Kn)=o(n)\operatorname{height}(K_n)=o(n) a.s.

Now suppose that an≤Cn−θa_n\le Cn^{-\theta} and let θ0:=min⁡{θ,1/2}\theta_0:=\min\{\theta,1/2\} (or θ0:=min⁡{θ,c}\theta_0:=\min\{\theta,c\} for any fixed c∈(0,1)c\in(0,1)). We have sn≤C′n1−θ0s_n\le C'n^{1-\theta_0} (with possibly C′>CC'>C). Choose ζ0∈(0,θ0)\zeta_0\in(0,\theta_0) and set kn:=⌈n1−ζ0⌉k_n:=\lceil n^{1-\zeta_0}\rceil. We then have

P[height⁡(Kn)≥kn]≤(C′′nζ0−θ0)n1−ζ0\mathbb{P}[\operatorname{height}(K_n)\ge k_n]\le(C''n^{\zeta_0-\theta_0})^{n^{1-\zeta_0}}

and Borel-Cantelli implies that a.s. height⁡(Kn)=O(n1−ζ0)\operatorname{height}(K_n)=O(n^{1-\zeta_0}). Since diam⁡(Kn)≤2⋅height⁡(Kn)\operatorname{diam}(K_n)\le2\cdot\operatorname{height}(K_n), the claimed diameter bounds follow in both cases. □

Note that KnK_n, viewed not as a tree but with all edges from GG between vertices of KnK_n retained, trivially has diameter at most that of the tree, so conforms to the asymptotics of Proposition 3.3 too.

To invoke Proposition 3.3, we must show existence of good scales for the DLA cluster KnK_n for all sufficiently large nn, with the number of good scales growing in nn, after which we can use Proposition 3.2.

Change of initial vertex

We now briefly describe the procedure for changing the starting vertex for DLA growth. For b∈V(G)b\in\mathcal{V}(G), define an(G,b)a_n(G,b) by replacing oo with bb in (3). Fix a finite path PP from oo to bb and let TS:=inf⁡{j≥0:Xj∈S}T_S:=\inf\{j\ge0:X_j\in S\} for the random walk XnX_n. Set

qP:=min⁡u∈PPu(T{b}<∞)>0.q_P:=\min_{u\in P}\mathbb{P}_u(T_{\{b\}}<\infty)>0.

For finite connected A∋bA\ni b, put B=A∪PB=A\cup P. A walk first entering AA at v∈Av\in A must first enter BB at vv or in P∖AP\setminus A. Also, for x∉Bx\notin B the Markov property gives Px(TA<∞)≥qPPx(TB<∞)\mathbb{P}_x(T_A<\infty)\ge q_P\mathbb{P}_x(T_B<\infty). Conditioning on hitting the respective sets and then sending xx to infinity gives

harm⁡A∞(EA(v))≤qP−1(harm⁡B∞(EB(v))+∑u∈P∖Aharm⁡B∞(EB(u)))≤∣P∣qPa∣A∣(G,o),\operatorname{harm}^{\infty}_A(\mathcal{E}_A(v))\le q_P^{-1}\left(\operatorname{harm}^{\infty}_B(\mathcal{E}_B(v))+\sum_{u\in P\setminus A}\operatorname{harm}^{\infty}_B(\mathcal{E}_B(u))\right)\le\frac{|P|}{q_P}a_{|A|}(G,o),

since BB is rooted and connected, ∣B∣≥∣A∣|B|\ge|A| and 1+∣P∖A∣≤∣P∣1+|P\setminus A|\le|P|. Thus

an(G,b)≤∣P∣qPan(G,o).(4)a_n(G,b)\le\frac{|P|}{q_P}a_n(G,o). \tag*{(4)}

For a recurrent random walk we have qP=1q_P=1, and the same argument applies to the conductance weighted walk. Proposition 3.3 therefore gives the same growth bounds for DLA started from bb, with the same polynomial exponent where applicable. However, the implied constants and eventual thresholds may depend on the initial vertex, and so change of initial growth point for an RPM is a little more subtle. For this reason, we largely omit the discussion of change of initial growth point in the RPM case, and focus on the rooted growth.

Proof of Theorem 1.6 under Assumption 1.4

We now use Assumption 1.4 to show the existence of good scales, and invoke Propositions 3.2 and 3.3. First, we define the relevant plane data K\mathcal{K} in this setting. For an edge ee, let Φ(e)∘\Phi(e)^\circ denote the embedded edge with endpoints removed. For x∈V(G)x \in\mathcal{V}(G), let

Jx:={Φ(x)}∪⋃e∈E(G):x∈eΦ(e)∘.J_x := \{\Phi(x)\} \cup\bigcup_{\substack{e \in E(G):\\x\in e}} \Phi(e)^\circ.

Each JxJ_x is bounded and connected, and the family K:={Jx:x∈V(G)}\mathcal{K} := \{J_x : x \in\mathcal{V}(G)\} is locally finite by properness of the embedding (see Definition 1.3). Moreover, for distinct vertices x,yx,y,

Jx∩Jy≠∅⟺x∼y.J_x \cap J_y \ne\varnothing\quad\Longleftrightarrow\quad x \sim y.

As before, for D⊂CD \subset\mathbb{C} we write

VΦ(D):={x∈V(G):Jx∩D≠∅},D(A):=diam⁡{Φ(x):x∈A}.\mathcal{V}_{\Phi}(D) := \{x \in\mathcal{V}(G) : J_x \cap D \ne\varnothing\}, \qquad D(A) := \operatorname{diam}\{\Phi(x) : x \in A\}.

Note that generally D(A)≠diam⁡(UAK)D(A) \ne\operatorname{diam}(U_A^{\mathcal{K}}).

Robust Brownian intersections

Here, we prove that with positive probability a Brownian motion intersects itself in an annulus in such a way that an appropriate approximating curve, which we will take to be a random walk trajectory on the embedded graph, also intersects itself. This allows us to infer the relevant annular intersection probability of the random walk from the Brownian motion convergence. The claims of this section are elementary and are therefore stated without proof.

Let Γ\Gamma be the polygonal path depicted in the below figure. The first vertical segment and the last horizontal segment cross transversely at (1,1/10)(1,1/10), and the loop winds once around the origin while staying in a small annulus.

The next claim is obvious.

Claim 4.1. There is η∗>0\eta_* > 0 such that every finite curve γ\gamma within η∗\eta_* of Γ\Gamma in the metric dFd_F stays in A3/4,5/4(0)A_{3/4,5/4}(0) and contains a closed curve with winding number one about B3/4(0)B_{3/4}(0).

We call an ϵ\epsilon-neighborhood of Γ\Gamma with ϵ<η∗\epsilon< \eta_* a robust neighborhood. The next claim is immediate by planarity and properties of the embedding.

Claim 4.2. If a finite piece of an embedded walk contains a closed curve with nonzero winding number about zz, then the graph walk contains an embedded graph cycle surrounding zz.

The final claim of this section is a basic property of Brownian motion (see e.g. [52 Exercise 1.8]).

Claim 4.3. There exist p0>0p_0 > 0 and η∈(0,η∗/10)\eta\in(0,\eta_*/10) such that for y=ρeiθy = \rho e^{i\theta} with 1−η≤ρ≤1+η1-\eta\leq\rho\leq1+\eta, standard Brownian motion started from yy has probability at least p0p_0 to possess an initial segment in a robust neighborhood of Γy\Gamma_y (where Γy\Gamma_y is the rotation by θ\theta of the template Γ\Gamma whose initial point (1,0)(1,0) has been replaced by (ρ,0)(\rho,0)).

We allow ρ≠1\rho\ne1 because, in our later applications to embedded random walks, the starting vertex will generally lie close to a prescribed Euclidean circle rather than exactly on it. We call the event that Brownian motion, and by abuse of terminology the random walk too, is in a robust neighborhood the robust intersection event.

Diameter bounds for embedded edges

As mentioned in § 3.1, to ensure that the embedded graph conforms to the non-normalized scales, we need to bound the diameter of the elements in the planar data. In this subsection, we prove that the maximum diameter of the edges (of a graph satisfying Assumption 1.4) intersecting a Euclidean ball is sublinear in the radius of the ball.

For C,R>0C,R>0 define

mC(R):=1Rmax⁡{diam⁡(Φ(e)):e∈E(G), Φ(e)∩BCR(0)‾≠∅}.m_C(R) := \frac{1}{R}\max\{\operatorname{diam}(\Phi(e)):e\in E(G),\ \Phi(e)\cap\overline{B_{CR}(0)}\ne\varnothing\}.

We have mC(R)<∞m_C(R)<\infty by properness of the embedding.

Proposition 4.4. Under Assumption 1.4 for every fixed C<∞C<\infty we have

mC(R)→0as R→∞.m_C(R)\to0\quad\text{as }R\to\infty.

Proof. Fix h∈(0,1)h\in(0,1). We prove that mC(R)≤5h/2m_C(R)\leq5h/2 for all sufficiently large RR. Choose finitely many points z1,…,zN∈BC(0)z_1,\ldots,z_N\in B_C(0) such that

BC(0)‾⊂⋃i=1NBh/4(zi).(5)\overline{B_C(0)}\subset\bigcup_{i=1}^{N}B_{h/4}(z_i). \tag*{(5)}

For each ii, choose a polygonal curve which starts at 00, travels to zi+(h,0)z_i+(h,0), traces a translated and scaled copy of Γ\Gamma, and then continues to ∂BL(0)\partial B_L(0). We choose L>C+10L>C+10 so that the polygonal curves and their relevant neighborhoods are contained in BL(0)B_L(0).

Let dFd_F be the metric used in Assumption 1.4 and W\mathbb{W} be the law of Brownian motion started at 00 and stopped upon exiting BL(0)B_L(0). Additionally, let μR\mu_R be the law of the rescaled embedded random walk started from the root oo and stopped upon leaving BL(0)B_L(0). Assumption 1.4 gives

dPr(W,μR)→0as R→∞.d_{\mathrm{Pr}}(\mathbb{W},\mu_R)\to0\quad\text{as }R\to\infty.

For each ii we let γi\gamma_i be the corresponding polygonal route. By Claim 4.1 there is ηi>0\eta_i>0 such that every curve γ\gamma satisfying dF(γ,γi)<2⋅ηid_F(\gamma,\gamma_i)<2\cdot\eta_i contains a curve in A3h/4,5h/4(zi)A_{3h/4,5h/4}(z_i) separating B3h/4(zi)‾\overline{B_{3h/4}(z_i)} from ∞\infty. Define the events

Ai:={γ:dF(γ,γi)<ηi},Bi:={γ:dF(γ,γi)<2⋅ηi}.\mathcal{A}_i:=\{\gamma:d_F(\gamma,\gamma_i)<\eta_i\},\qquad\mathcal{B}_i:=\{\gamma:d_F(\gamma,\gamma_i)<2\cdot\eta_i\}.

Claim 4.3 gives qi:=W(Ai)>0q_i:=\mathbb{W}(\mathcal{A}_i)>0, and since N<∞N<\infty, the number ϵC,h:=min⁡1≤i≤N{ηi,qi/2}\epsilon_{C,h}:=\min_{1\leq i\leq N}\{\eta_i,q_i/2\} is positive. By Assumption 1.4 there is a finite R∗(C,h)R_*(C,h) such that

dPr(W,μR)<ϵC,hfor every R≥R∗(C,h).d_{\mathrm{Pr}}(\mathbb{W},\mu_R)<\epsilon_{C,h}\quad\text{for every }R\geq R_*(C,h).

If dPr(W,μR)<ϵC,hd_{\mathrm{Pr}}(\mathbb{W},\mu_R)<\epsilon_{C,h} then by definition of the Lévy-Prokhorov metric we have for every Borel set AA (in the topology induced by dFd_F) that

W(A)≤μR(AϵC,h)+ϵC,h,\mathbb{W}(A)\leq\mu_R(A^{\epsilon_{C,h}})+\epsilon_{C,h},

where AϵA^\epsilon denotes the open ϵ\epsilon-neighborhood of AA under dFd_F. Taking A=AiA=A_i gives μR(AiϵC,h)≥qi−ϵC,h≥qi/2>0\mu_R(A_i^{\epsilon_{C,h}})\geq q_i-\epsilon_{C,h}\geq q_i/2>0. Moreover, since ϵC,h≤ηi\epsilon_{C,h}\leq\eta_i we have AiϵC,h⊂BiA_i^{\epsilon_{C,h}}\subset B_i and so

μR(Bi)≥qi2>0for every i and every R≥R∗(C,h).(6)\mu_R(B_i)\geq\frac{q_i}{2}>0\qquad\text{for every }i\text{ and every }R\geq R_*(C,h). \tag*{(6)}

Fix R≥R∗(C,h)R\geq R_*(C,h). Since the graph and embedding are deterministic, and the randomness in (6) is only over the random walk, positive probability implies that an appropriate trajectory exists. Choose one such trajectory for each ii. By Claims 4.1 and 4.2 each trajectory contains a graph cycle CiRC_i^R such that

R−1Φ(CiR)⊂A3h/4,5h/4(zi)andR−1Φ(CiR) surrounds B3h/4(zi)‾.R^{-1}\Phi(C_i^R)\subset A_{3h/4,5h/4}(z_i)\quad\text{and}\quad R^{-1}\Phi(C_i^R)\text{ surrounds }\overline{B_{3h/4}(z_i)}.

Since the above probabilistic argument is only existential, we can deduce the existence of such cycles for all ii simultaneously. Let ee be an edge such that Φ(e)∩BCR(0)≠∅\Phi(e)\cap B_{CR}(0)\neq\varnothing. By (5), there is an ii such that

R−1Φ(e)∩Bh/4(zi)≠∅.R^{-1}\Phi(e)\cap B_{h/4}(z_i)\neq\varnothing.

The edge ee cannot belong to CiRC_i^R since the cycle is disjoint from the disk, and by planarity the interior of ee cannot cross the cycle. So R−1Φ(e)R^{-1}\Phi(e) is contained in B5h/4(zi)B_{5h/4}(z_i), and diam⁡(R−1Φ(e))≤52h\operatorname{diam}(R^{-1}\Phi(e))\leq\frac{5}{2}h. This bound holds for all edges, and so for each fixed h>0h>0 we have lim sup⁡R→∞mC(R)≤52h\limsup_{R\to\infty}m_C(R)\leq\frac{5}{2}h and letting h↓0h\downarrow0 completes the proof.

It follows immediately that for every fixed C<∞C<\infty one has 1Rmax⁡{diam⁡(Jx):Jx∩BCR(0)≠∅}→0\frac{1}{R}\max\{\operatorname{diam}(J_x):J_x\cap B_{CR}(0)\neq\varnothing\}\to0 as R→∞R\to\infty.

Completing Theorem 1.6

Now, we verify the existence of good scales and complete the proof of Theorems 1.5 and 1.6. The next Lemma establishes control of annular intersection probabilities at a single normalized scale ρ∈(0,1)\rho\in(0,1).

Lemma 4.5. Under Assumption 1.4 there exist deterministic η,p∈(0,1)\eta,p\in(0,1) such that for every C<∞C<\infty and every fixed normalized scale ρ∈(0,1)\rho\in(0,1), there is R∗(C,ρ)<∞R_*(C,\rho)<\infty such that the following holds for all R≥R∗(C,ρ)R\geq R_*(C,\rho). Let z∈BCR(0)z\in B_{CR}(0), set r:=ρRr:=\rho R and let x∈V(G)x\in V(G) satisfy

1−η≤∣Φ(x)−z∣r≤1+η.(7)1-\eta\leq\frac{|\Phi(x)-z|}{r}\leq1+\eta. \tag*{(7)}

A simple random walk XnX_n started from xx and run until its first exit time τ\tau from VK(Ar/2,3r/2(z))V_K(A_{r/2,3r/2}(z)) has probability at least pp to disconnect VK(∂B3r/4(z))V_K(\partial B_{3r/4}(z)) from VK(∂B5r/4(z))V_K(\partial B_{5r/4}(z)) with ⋃m=0τJXm⊂Ar/3,5r/3(z)\bigcup_{m=0}^{\tau}J_{X_m}\subset A_{r/3,5r/3}(z). The probability pp is independent of C,ρ,R,zC,\rho,R,z, and xx.

Proof. Fix C<∞C<\infty and ρ∈(0,1)\rho\in(0,1), and set r=ρ⋅Rr=\rho\cdot R. Choose L>C+3L>C+3, and let η\eta be as in Claim 4.3, decreasing η\eta if necessary such that η<1\eta<1. By (7) every admissible starting vertex xx satisfies ∣Φ(x)∣≤(C+2)R|\Phi(x)|\leq(C+2)R. After translating by −z-z, rescaling by r−1r^{-1} and rotating so that the starting point lies near (1,0)(1,0), Claim 4.3 gives a p0>0p_0>0, independent of C,ρ,zC,\rho,z, and xx such that Brownian motion remains in a robust neighborhood of Γ\Gamma. In the original coordinates rescaled by R−1R^{-1}, this curve remains in A3ρ/4,5ρ/4(z/R)A_{3\rho/4,5\rho/4}(z/R) and contains a closed curve in A4ρ/5,6ρ/5(z/R)A_{4\rho/5,6\rho/5}(z/R) which separates B3ρ/4(z/R)B_{3\rho/4}(z/R) from ∞\infty. By an identical argument to that of Proposition 4.4 there exists an R1(C,ρ)<∞R_1(C,\rho)<\infty such that when R≥R1(C,ρ)R\geq R_1(C,\rho) the rescaled embedded random walk intersects robustly an equivalent annulus uniformly for all admissible zz and xx with probability at least p0/2p_0/2.

By Proposition 4.4 there is an R2(C,ρ)<∞R_2(C,\rho)<\infty such that for R≥R2(C,ρ)R\geq R_2(C,\rho), every set JuJ_u intersecting the ambient region containing the robust neighborhood of (the rotation of) Γ\Gamma has diameter smaller than a sufficiently small multiple of rr. On the robust intersection event, by Claim 4.2 the trace of the random walk contains a cycle in A4r/5,6r/5(z)A_{4r/5,6r/5}(z) disconnecting B3r/4(z)B_{3r/4}(z) from ∞\infty.

The containment claims of the Lemma follow analogously to those of Proposition 4.4, and choosing R∗(C,ρ)=max⁡(R1(C,ρ),R2(C,ρ))R_*(C,\rho)=\max(R_1(C,\rho),R_2(C,\rho)) completes the proof.

To observe that the above pp is independent of ρ\rho we use the scale invariance of Brownian motion modulo parametrization and the fact that the Brownian motion is killed upon leaving the annulus, which is itself appropriately rescaled.

Now, we show the existence of many good scales as the Euclidean diameter of the cluster diverges. The next proposition essentially uses Proposition 4.4 to show the existence of an arbitrarily large number of disjoint scales as R→∞R \to\infty and then applies Lemma 4.5 at each individual scale. Uniformity over AA and vv follows from the fact that the sets are rooted.

Proposition 4.6. Under Assumption 1.4 there is a p>0p > 0 such that for every K∈NK \in\mathbb{N} there exists an mK<∞m_K < \infty with the following property: if A∋oA \ni o is finite and connected, ∣A∣≥mK|A| \ge m_K and v∈Av \in A, then there are (K,p)(K,p)-good scales for (A,v)(A,v).

Proof. Let

ρ1:=1/100,ρj:=ρ1⋅64−(j−1),1≤j≤K.\rho_1 := 1/100,\qquad\rho_j := \rho_1 \cdot64^{-(j-1)},\qquad1 \le j \le K.

Properness of the embedding therefore implies that inf⁡A∋o:∣A∣>kD(A)→∞\inf_{A\ni o:|A|>k} D(A) \to\infty as k→∞k \to\infty over rooted connected sets; thus, any property which holds for RR large enough also holds for large enough kk.

Now, suppose that K∈NK \in\mathbb{N} and we choose ρ1\rho_1 and ρK\rho_K as above. Since the maximum diameter of all JxJ_x intersecting B2RB_{2R} is o(R)o(R), by choosing C=2C = 2 in Proposition 4.4 there exists an R(K)R(K) (where η\eta and pp are as in Lemma 4.5) such that if R≥R(K)R \ge R(K) then the maximum diameter of all JxJ_x intersecting B2RB_{2R} is less than ρKRmin⁡{1/100,η}\rho_K R \min\{1/100,\eta\}.

For AA a rooted, finite, connected set, let R=D(A)R = D(A) and fix z=Φ(v)z = \Phi(v). Since AA is rooted and Φ(o)=0\Phi(o) = 0, we have Φ(A)⊂BR(0)\Phi(A) \subset B_R(0). For every x∈N(v)∖{v}x \in N(v) \setminus\{v\}, Jx∩Jv≠∅J_x \cap J_v \ne\varnothing, so for R≥R(K)R \ge R(K), each JxJ_x intersects B2R(0)B_{2R}(0) and has diameter less than ρKR/100\rho_K R/100. If D(A)≥R(K)D(A) \ge R(K) then ⋃x∈N(v)Jx\bigcup_{x \in N(v)} J_x is contained in BρKR/20(z)B_{\rho_K R/20}(z), and every JxJ_x intersecting B2ρ1R(z)B_{2\rho_1 R}(z) has diameter less than ρKR/100\rho_K R/100 by the preceding paragraph, since B2ρ1R(z)⊂B2R(0)B_{2\rho_1 R}(z) \subset B_{2R}(0).

There exists u∈Au \in A with ∣Φ(u)−z∣≥R/2|\Phi(u)-z| \ge R/2, and an embedded graph path in AA from vv to uu crosses both circles ∂B3ρjR/4(z)\partial B_{3\rho_j R/4}(z) and ∂B5ρjR/4(z)\partial B_{5\rho_j R/4}(z) for all 1≤j≤K1 \le j \le K. Applying Lemma 4.5 with C=2C = 2 to each ρj\rho_j shows the existence of (K,p)(K,p)-good scales, and picking mKm_K such that inf⁡A∋o:∣A∣≥mD(A)>R(K)\inf_{A\ni o:|A|\ge m} D(A) > R(K) for all m≥mKm \ge m_K completes the proof. □\square

Note that the order of quantifiers to achieve uniformity over AA large and vv is somewhat delicate. For the fixed threshold R(K)R(K), an AA with D(A)>R(K)D(A) > R(K) may fail to satisfy max⁡{diam⁡(Jx):x∈A}≤ρKR(K)/100\max\{\operatorname{diam}(J_x) : x \in A\} \le\rho_K R(K)/100, since there may exist some v∈Av \in A such that Jv∩B2R(K)(0)=∅J_v \cap B_{2R(K)}(0) = \varnothing. However, max⁡{diam⁡(Jx):x∈A}≤ρKD(A)/100\max\{\operatorname{diam}(J_x) : x \in A\} \le\rho_K D(A)/100 does hold, since the numerator in the definition mC(R)m_C(R) adjusts the size of the ball with which edges intersect, and so the convergence mC(R)→0m_C(R) \to0 controls the relative mesh size at all sufficiently large scales simultaneously.

Now, we can complete Theorems 1.5 and 1.6.

Proof of Theorems 1.5 and 1.6. Let pp be as in Proposition 4.6. Given δ>0\delta> 0, choose KK with (1−p)K<δ(1-p)^K < \delta. For every sufficiently large rooted connected AA and every v∈Av \in A there are (K,p)(K,p)-good scales for (A,v)(A,v). By Proposition 3.2 we have

harm⁡A∞(EA(v))≤(1−p)K<δ.\operatorname{harm}_{A}^{\infty}(\mathcal{E}_A(v)) \le(1-p)^K < \delta.

Sending K→∞K \to\infty as δ→0\delta\to0 produces Theorem 1.5, and combining with Proposition 3.3 completes Theorem 1.6 for DLA started from oo. Equation (4) gives the extension to any other initial vertex. □\square

Proof of Theorem 1.10 under Assumption 1.8

As in § 4, we use Assumption 1.8 to show the existence of good scales, and invoke Proposition 3.3. The stronger control of Assumption 1.8 as compared to Assumption 1.4 will allow us to improve the number of good scales.

We begin with two trivial observations. The first follows from the triangle inequality, and the second follows from the reverse triangle inequality and the connectedness of the union of cells corresponding to a connected subgraph of GG.

Lemma 5.1. Suppose that for every x∈N(v)x \in N(v) the cell HxH_x has diameter at most dd, and choose zv∈Hvz_v \in H_v. Then

⋃x∈N(v)Hx⊂B2d(zv)‾.\bigcup_{x \in N(v)} H_x \subset\overline{B_{2d}(z_v)}.

Lemma 5.2. Let AA be finite and connected, with D:=diam⁡(UAH)D := \operatorname{diam}(U_A^{\mathcal{H}}), and suppose a>0a > 0 and dist⁡(z,UAH)≤a<b≤D/2\operatorname{dist}(z,U_A^{\mathcal{H}}) \le a < b \le D/2. Then AA contains a graph path joining VH(∂Ba(z))V_{\mathcal{H}}(\partial B_a(z)) to VH(∂Bb(z))V_{\mathcal{H}}(\partial B_b(z)).

The next proposition establishes good scales for every sufficiently large non-normalized scale, where the number of good scales depends logarithmically on the Euclidean diameter of the cluster. Subsequently, item 4 of Assumption 1.8 will be used to convert the lower bound on the number of good scales from a term depending on the Euclidean diameter of the cells corresponding to the cluster to a term depending only on the cardinality of the cluster.

Proposition 5.3. Under Assumption 1.8, there exists a constant c0>0c_0 > 0 and a finite threshold R1≥R0R_1 \ge R_0 such that the following holds: if A∋oA \ni o is finite and connected, R:=diam⁡(UAH)≥R1R := \operatorname{diam}(U_A^{\mathcal{H}}) \ge R_1, and v∈Av \in A, then there are (K,p)(K,p)-good scales for (A,v)(A,v) with

K≥c0log⁡R.K \ge c_0 \log R.

Proof. Fix C0≥200C_0 \ge200. By Assumption 1.8, there exists an R1≥R0R_1 \ge R_0 sufficiently large that for every R≥R1R \ge R_1 we have

e(R)+d(R)≤m(R),C0m(R)≤R/200.(8)e(R) + d(R) \le m(R), \qquad C_0m(R) \le R/200. \tag*{(8)}

Set R:=diam⁡(UAH)≥R1R := \operatorname{diam}(U_A^{\mathcal{H}}) \ge R_1, fix v∈Av \in A, and choose zv∈Hvz_v \in H_v. Since AA is rooted and 0∈Ho0 \in H_o, we have 0∈UAH0 \in U_A^{\mathcal{H}} and hence UAH⊂BR(0)U_A^{\mathcal{H}} \subset B_R(0). Next, choose z∈ZRz \in Z_R with ∣z−zv∣≤e(R)|z-z_v| \le e(R). Set

sj:=R20032−(j−1),j≥1,s_j := \frac{R}{200}32^{-(j-1)}, \qquad j \ge1,

and let KK be the largest integer with sK≥C0m(R)s_K \ge C_0m(R). For each j≤Kj \le K, choose rj∈QRr_j \in Q_R with ∣rj−sj∣≤e(R)|r_j-s_j| \le e(R). Equation (8) gives e(R)≤m(R)≤sj/C0e(R) \le m(R) \le s_j/C_0, and so we have rj≥(1−C0−1)sjr_j \ge(1-C_0^{-1})s_j and rj+1≤(32−1+C0−1)sjr_{j+1} \le(32^{-1}+C_0^{-1})s_j, so rj+1≤rj/8r_{j+1} \le r_j/8. By our choice of KK, we have K>log⁡32(R/(200C0m(R)))K > \log_{32}(R/(200C_0m(R))). Since m(R)≤R1−σm(R) \le R^{1-\sigma}, by possibly increasing R1R_1 again, we have

K≥c0log⁡R.K \ge c_0 \log R.

Now, we can verify the existence of (K,p)(K,p)-good scales for (A,v)(A,v). Since z∈B2R(0)z \in B_{2R}(0) and r1≤R/100r_1 \le R/100, all cells intersecting B2r1(z)B_{2r_1}(z) also intersect B4R(0)B_{4R}(0), so by Assumption 1.8 the cell diameters are bounded by d(R)d(R). Lemma 5.1 then gives ⋃x∈N(v)Hx⊂B2d(R)+e(R)(z)\bigcup_{x \in N(v)}H_x \subset B_{2d(R)+e(R)}(z). Now, rK≥sK−e(R)≥(C0−1)m(R)r_K \ge s_K-e(R) \ge(C_0-1)m(R), so d(R)≤rK/100d(R) \le r_K/100 and 2d(R)+e(R)≤rK/202d(R)+e(R) \le r_K/20, thus

⋃x∈N(v)Hx⊂BrK/20(z).\bigcup_{x \in N(v)}H_x \subset B_{r_K/20}(z).

This proves the first two items in Definition 3.1. Since zv∈UAHz_v \in U_A^{\mathcal{H}} we have dist⁡(z,UAH)≤e(R)<3rK/4\operatorname{dist}(z,U_A^{\mathcal{H}}) \le e(R) < 3r_K/4. Also, 5rj/4≤5r1/4<R/25r_j/4 \le5r_1/4 < R/2 after increasing R1R_1 again if necessary. Applying Lemma 5.2 with a=3rj/4a = 3r_j/4 and b=5rj/4b = 5r_j/4 gives the third item in Definition 3.1 for every RR. Finally, z∈ZRz \in Z_R and rj∈QRr_j \in Q_R, and so the third item in Assumption 1.8 gives the fourth item in Definition 3.1. This concludes the proof. □

We can now conclude Theorems 1.9 and 1.10.

Proof of Theorems 1.9 and 1.10. Let A∋oA \ni o be finite and connected with diam⁡(UAH)≥R1\operatorname{diam}(U_A^{\mathcal{H}}) \ge R_1. Propositions 3.2 and 5.3 give

harm⁡A∞(EA(v))≤(1−p)c0log⁡diam⁡(UAH)=diam⁡(UAH)−ξ,ξ:=−c0log⁡(1−p)>0.\operatorname{harm}^{\infty}_A(\mathcal{E}_A(v)) \le(1-p)^{c_0\log\operatorname{diam}(U_A^{\mathcal{H}})} = \operatorname{diam}(U_A^{\mathcal{H}})^{-\xi}, \qquad\xi:= -c_0\log(1-p) > 0.

By item four of Assumption 1.8, for AA sufficiently large we have

diam⁡(UAH)−ξ≤cvol−ξ∣A∣−νξ.\operatorname{diam}(U_A^{\mathcal{H}})^{-\xi} \le c_{\mathrm{vol}}^{-\xi}|A|^{-\nu\xi}.

Choosing any θ∈(0,νξ)\theta\in(0,\nu\xi) gives the strong Beurling estimate for sufficiently large AA. Theorem 1.9 and item two of Proposition 3.3 give the polynomially sublinear growth bound for DLA started from oo. Equation (4) transfers this bound to any other initial vertex with the same exponent ζ\zeta. □

Verifying Assumption 1.8 for the mated CRT map

With the preliminaries from § 2 we verify Assumption 1.8 for the mated CRT map, and conclude Theorem 1.12. As the assumption pertains to a deterministic geometry, we mean that the mated CRT map satisfies the assumption a.s. and with possibly random values for constants such as R0R_0.

We will need to convert the properties stated in Proposition 2.6, which pertain to a single scale, to the multiscale requirements of Assumption 1.8. Essentially, the argument boils down to scaling properties of the mated CRT map and an elementary Borel-Cantelli argument.

Proposition 6.1. Fix γ∈(0,2)\gamma\in(0,2). The γ\gamma-mated CRT map, rooted at o=0o=0, a.s. satisfies Assumption 1.8, with deterministic, possibly γ\gamma-dependent, pp and σ\sigma.

Proof. In what follows we work on a full probability event over the mated CRT map, so omit the phrase “almost surely” as regards the randomness of the environment.

Let p,α,β,χ,q,ρp,\alpha,\beta,\chi,q,\rho be as in Proposition 2.6, decreasing β\beta if necessary such that β<aγ\beta<a_\gamma. For k∈Nk\in\mathbb{N}, set bk:=2kb_k:=2^k, mk:=Rbkbk−βm_k:=R_{b_k}b_k^{-\beta}, ek:=Rbkbk−χe_k:=R_{b_k}b_k^{-\chi}, and dk:=Rbkbk−qd_k:=R_{b_k}b_k^{-q}. By quantum cone scaling (Proposition 2.3) the conclusion of Proposition 2.6 at ϵ=2−k\epsilon=2^{-k} transfers to the unit-mesh map at the (random) non-normalized scale RbkR_{b_k}. By Borel-Cantelli, simultaneously for all sufficiently large kk the conclusions of Proposition 2.6 hold, i.e. every cell intersecting BρRbk(0)B_{\rho R_{b_k}}(0) has diameter at most dkd_k and the annular random walk conclusions of Proposition 2.6 are true.

Choose λ∈(1/aγ,1/(aγ−β))\lambda\in(1/a_\gamma,1/(a_\gamma-\beta)) and set σ∗:=1−λ(aγ−β)>0\sigma_*:=1-\lambda(a_\gamma-\beta)>0. For D>1D>1, let k(D):=⌈λlog⁡2D⌉k(D):=\lceil\lambda\log_2D\rceil; we will treat DD as the non-normalized scale parameter. Proposition 2.4 gives Rbk(D)/D→∞R_{b_{k(D)}}/D\to\infty, mk(D)=D1−σ∗+oD(1)m_{k(D)}=D^{1-\sigma_*+o_D(1)}, and ek(D)+dk(D)=oD(mk(D))e_{k(D)}+d_{k(D)}=o_D(m_{k(D)}) as D→∞D\to\infty. Now, set σ:=σ∗/2\sigma:=\sigma_*/2 and define m(D):=mk(D)m(D):=m_{k(D)}, e(D):=4ek(D)e(D):=4e_{k(D)}, and d(D):=dk(D)d(D):=d_{k(D)}. This establishes item 1 of Assumption 1.8. In the remainder of the proof, take DD sufficiently large.

Now, let QD:=(ek(D)Z)∩[m(D),D/100]Q_D:=(e_{k(D)}\mathbb{Z})\cap[m(D),D/100] and ZD:=(ek(D)Z2)∩B2D(0)Z_D:=(e_{k(D)}\mathbb{Z}^2)\cap B_{2D}(0). These are e(D)e(D)-nets of [m(D),D/100][m(D),D/100] and B2D(0)B_{2D}(0) respectively, which establishes the net requirement in item 3 of Assumption 1.8.

Finally, ρRbk(D)≥5D\rho R_{b_{k(D)}}\ge5D implies B4D(0)⊂BρRbk(D)(0)B_{4D}(0)\subset B_{\rho R_{b_{k(D)}}}(0), so every cell intersecting B4D(0)B_{4D}(0) has diameter at most d(D)d(D), establishing item 2. If z∈ZDz\in Z_D and r∈QDr\in Q_D, then B2r(z)⊂B(2+1/50)D(0)⊂BρRbk(D)(0)B_{2r}(z)\subset B_{(2+1/50)D}(0)\subset B_{\rho R_{b_{k(D)}}}(0). Along with the earlier quantum cone scaling argument, this establishes item 3 of Assumption 1.8 for some deterministic p>0p>0.

For item 4, let A∋oA\ni o be finite and connected with D:=diam⁡(UAH)D:=\operatorname{diam}(U_A^{\mathcal H}) sufficiently large. Since 0∈Ho0\in H_o, we have UAH⊂BD(0)U_A^{\mathcal H}\subset B_D(0). Proposition 2.5 with δ=1\delta=1 and k=k(D)k=k(D) gives ∣A∣≤∣VG1(BRbk(D)(0))∣≤22k(D)≤4D2λ|A|\leq|\mathcal V_{G_1}(B_{R_{b_{k(D)}}}(0))|\leq2^{2k(D)}\leq4D^{2\lambda}. Local finiteness gives D→∞D\to\infty uniformly as ∣A∣→∞|A|\to\infty, so item 4 holds with ν=1/(2λ)\nu=1/(2\lambda) and cvol=4−νc_{\mathrm{vol}}=4^{-\nu}.

Finally, Proposition 6.1 and Theorem 1.10 imply that DLA grown from the root oo on the γ\gamma-mated CRT has a.s. polynomially sublinear growth. Inspection of the exponent derived in Proposition 6.1 reveals an inequality which relates the DLA growth bound to the LQG parameter γ\gamma.

DLA on an ergodic scale-free random environment

In this brief section, we sketch the proof of Theorem 1.14. Let M\mathcal M be an embedded lattice as in Definition 1.13, and let G=M∗G=\mathcal M^* be its adjacency graph, with root o=H0o=H_0 and DLA initial cluster K0={o}K_0=\{o\}. We again assume that harmonic measure from infinity is unique. In this section, good scales are defined using the conductance weighted walk in the obvious way.

Proof sketch. Let XX be the conductance weighted walk on G=M∗G=\mathcal M^*, with vertex embeddings Φ(H)=pH\Phi(H)=p_H as in Proposition 2.8. Rewritten in the notation of Assumption 1.4, [28], Theorem 3.10 gives a.s. over M\mathcal M, for every 0<C<L<∞0<C<L<\infty the convergence

sup⁡x: ∣Φ(x)∣≤CRdPr(Lx(R−1Y^x;R,L∣M),L(BΣΦ(x)/R;L))→0 as R→∞\sup_{x:\,|\Phi(x)|\le CR} d_{\mathrm{Pr}}\left(\mathcal L_x\left(R^{-1}\widehat{Y}^{x;R,L}\mid\mathcal M\right),\mathcal L\left(B_{\Sigma}^{\Phi(x)/R;L}\right)\right)\to0 \text{ as } R\to\infty

where Y^x;R,L\widehat{Y}^{x;R,L} is the piecewise-linear interpolation through chosen points pHp_H of the faces, and BΣz;LB_{\Sigma}^{z;L} is Brownian motion with positive-definite covariance matrix Σ\Sigma started from zz and stopped upon leaving BL(0)B_L(0). Draw the dual graph as a proper planar map by placing each dual vertex at its chosen face representative and drawing each dual edge inside the union of its two incident faces. By [28 Lemma 2.9], for every fixed L<∞L < \infty we have

1Rmax⁡{diam⁡(H):H∩BLR(0)≠∅}→0.\frac{1}{R}\max\{\operatorname{diam}(H): H\cap B_{LR}(0)\ne\varnothing\}\to0.

Now, fix 0<C<L<∞0<C<L<\infty and choose L′>LL'>L. Set δR:=R−1max⁡{diam⁡(H):H∩B(L′+1)R(0)≠∅}\delta_R:=R^{-1}\max\{\operatorname{diam}(H):H\cap B_{(L'+1)R}(0)\ne\varnothing\}. The preceding limit implies that δR→0\delta_R\to0 almost surely. Going forward, we fix an environment on which these almost sure conclusions hold; all walk laws below are conditional on M\mathcal{M}.

We couple the proper and piecewise-linear embeddings through the same random walk on the graph, started from xx with Φ(x)∈BCR(0)\Phi(x)\in B_{CR}(0), and denote them by YY and Y^\widehat{Y} respectively. For adjacent faces H,H′H,H', since H∩H′≠∅H\cap H'\ne\varnothing, one has diam⁡(H∪H′)≤diam⁡(H)+diam⁡(H′)\operatorname{diam}(H\cup H')\le\operatorname{diam}(H)+\operatorname{diam}(H'). It follows that dF(R−1Y,R−1Y^)≤2δRd_F(R^{-1}Y,R^{-1}\widehat{Y})\le2\delta_R, where both paths are stopped when the piecewise-linear interpolation leaves BL′R(0)B_{L'R}(0).

Write B:=BΣΦ(x)/R;L′B:=B_{\Sigma^{\Phi(x)/R};L'}. Since the Prokhorov metric is induced by 1∧dF1\wedge d_F, the above implies

dPr⁡(Lx(R−1Y),Lx(R−1Y^))≤2δR.d_{\Pr}\bigl(\mathcal{L}_x(R^{-1}Y),\mathcal{L}_x(R^{-1}\widehat{Y})\bigr)\le2\delta_R.

By the triangle inequality, dPr⁡(Lx(R−1Y),L(B))≤2δR+dPr⁡(Lx(R−1Y^),L(B))d_{\Pr}(\mathcal{L}_x(R^{-1}Y),\mathcal{L}(B))\le2\delta_R+d_{\Pr}(\mathcal{L}_x(R^{-1}\widehat{Y}),\mathcal{L}(B)). The second term tends to zero uniformly over starting vertices in BCR(0)B_{CR}(0) by Proposition 2.8, while δR→0\delta_R\to0 is already shown.

Finally, we deal with a slight mismatch in stopping times between the two walks, since the above YY is stopped when Y^\widehat{Y} exits BL′R(0)B_{L'R}(0); we prove that YY stopped when it exits BLR(0)B_{LR}(0), after multiplication by R−1R^{-1}, converges to Brownian motion stopped when exiting BL(0)B_L(0). The terminal point of YY has modulus at least (L′−2δR)R>LR(L'-2\delta_R)R>LR for all large RR, so this exit occurs before the common graph-time interval ends.

Fix a sequence Rn→∞R_n\to\infty and starting vertices xnx_n with Φ(xn)∈BCRn(0)\Phi(x_n)\in B_{CR_n}(0), and suppose first that dPr⁡(Lxn(Rn−1Y),L(BΣz;L′))→0d_{\Pr}(\mathcal{L}_{x_n}(R_n^{-1}Y),\mathcal{L}(B_{\Sigma}^{z;L'}))\to0 for some z∈BC(0)z\in B_C(0). By the Skorokhod representation theorem, we may realize this convergence as a.s. convergence under dFd_F; denote by γn\gamma_n the corresponding curves, with their parametrizations chosen so that γn→γ\gamma_n\to\gamma uniformly on [0,1][0,1]. Here γ\gamma is the curve BΣz;L′B_{\Sigma}^{z;L'} parametrized by [0,1][0,1] in the obvious way (so γ\gamma exits BL′(0)B_{L'}(0) at t=1t=1). Let τn\tau_n and τ\tau be their respective first entry into C∖BL(0)\mathbb{C}\setminus B_L(0). For each sufficiently small δ>0\delta>0, the compact set γ([0,τ−δ])\gamma([0,\tau-\delta]) lies a positive distance inside BL(0)B_L(0), so τn≥τ−δ\tau_n\ge\tau-\delta for all sufficiently large nn. On the other hand, a.s. there exists s∈(τ,τ+δ)s\in(\tau,\tau+\delta) with ∣γ(s)∣>L|\gamma(s)|>L. Uniform convergence then gives ∣γn(s)∣>L|\gamma_n(s)|>L, and so τn≤s<τ+δ\tau_n\le s<\tau+\delta, for all sufficiently large n,n, which implies τn→τ\tau_n\to\tau a.s. By continuity of γ\gamma, the parametrizations u↦γn(uτn)u\mapsto\gamma_n(u\tau_n) converge uniformly to u↦γ(uτ)u\mapsto\gamma(u\tau), so the corresponding initial curve segments converge in dFd_F. Thus the rescaled random walk under Φ\Phi, stopped upon leaving BLRn(0)B_{LR_n}(0), converges weakly to BΣz;LB_{\Sigma}^{z;L}.

To obtain uniformity over starting vertices, suppose that uniform convergence fails along some sequence Rn→∞R_n\to\infty and xnx_n, with some C′<LC'<L, with the Prokhorov distances bounded away from zero. By sequential compactness, we can choose a convergent subsequence zn:=Φ(xn)/Rn→z∈BC′(0)z_n:=\Phi(x_n)/R_n\to z\in B_{C'}(0). Write the Brownian motions as zn+Wz_n+W and z+Wz+W, where WW is the same Brownian motion of covariance Σ\Sigma started at 00. Their distance before stopping is ∣zn−z∣→0|z_n-z|\to0. The same τ−δ,τ+δ\tau-\delta,\tau+\delta comparison above, applied at each radius a∈[L,L′]a\in[L,L'], shows that their departure times and stopped curves converge, so dPr⁡(L(BΣzn;a),L(BΣz;a))→0d_{\Pr}(\mathcal{L}(B_{\Sigma}^{z_n;a}),\mathcal{L}(B_{\Sigma}^{z;a}))\to0. Now, the (non-uniform) convergence dPr⁡(Lxn(Rn−1Y∣M),L(BΣΦ(xn)/Rn;L′))→0d_{\Pr}(\mathcal{L}_{x_n}(R_n^{-1}Y\mid\mathcal{M}),\mathcal{L}(B_{\Sigma}^{\Phi(x_n)/R_n;L'}))\to0 and dPr⁡(L(BΣzn;L′),L(BΣz;L′))→0d_{\Pr}(\mathcal{L}(B_{\Sigma}^{z_n;L'}),\mathcal{L}(B_{\Sigma}^{z;L'}))\to0, along with the triangle inequality show that dPr⁡(Lxn(Rn−1Y∣M),L(BΣz;L′))→0d_{\Pr}(\mathcal{L}_{x_n}(R_n^{-1}Y\mid\mathcal{M}),\mathcal{L}(B_{\Sigma}^{z;L'}))\to0. The preceding paragraph, together with a=La=L and the triangle inequality, show that the Prokhorov distances tend to zero, which is a contradiction.

Lastly, we deal with the anisotropy of Σ\Sigma. Let η\eta be as in Claim 4.3 and let UθU_\theta denote a rotation by θ\theta and set Σθ:=UθΣUθ\Sigma_\theta:=U_\theta\Sigma U_\theta. Write PzΣz\mathbb{P}^{\Sigma_z}_z for the law of Brownian motion B^\widehat{B} started at zz with covariance Σ\Sigma. If Γρ\Gamma_\rho is the polygonal path in Figure 2 with initial radius ρ∈[1−η,1+η]\rho\in[1-\eta,1+\eta], set η0:=η/2\eta_0:=\eta/2 and

The polygonal path Gamma

Figure 2. The polygonal path Γ\Gamma. The open circle at (1,110)(1,\frac{1}{10}) indicates the initial intersection of the first vertical segment with the final horizontal segment. The dotted blue circles indicate the annulus A4/5,6/5(0)A_{4/5,6/5}(0), and the dotted red circles indicate the annulus A3/4,5/4(0)A_{3/4,5/4}(0).

p:=12inf⁡ρ∈[1−η0,1+η0]inf⁡θ∈[0,2π]P(ρ,0)Σθ[∃t>0:dF(B[0,t],Γρ)<η0].p:=\frac{1}{2}\inf_{\rho\in[1-\eta_0,1+\eta_0]}\inf_{\theta\in[0,2\pi]}\mathbb{P}^{\Sigma_\theta}_{(\rho,0)}\left[\exists t>0:d_F(B_{[0,t]},\Gamma_\rho)<\eta_0\right].

which is strictly positive by compactness and continuity.

The good scales argument now applies identically. For every KK, there exists an N(K)<∞N(K)<\infty such that for every finite, connected, rooted set AA with ∣A∣>N(K)|A|>N(K) and every v∈Av\in A, there are (K,p)(K,p)-good scales for (A,v)(A,v). This implies

harm⁡∞(EA(v))≤(1−p)K.\operatorname{harm}^{\infty}(\mathcal{E}_A(v)) \le(1-p)^K.

The argument for DLA growth then follows identically to Proposition 3.3, and a.s. sublinearity of the internal diameter follows for DLA grown from H0H_0.

Open problems

We briefly mention two natural directions of further research. The first regards the identification of the growth exponent of the internal diameter of DLA. Theorem 1.10 bounds the growth exponent of the internal diameter of DLA. As mentioned in § 1, the paper [29] provides a result for the growth exponent of the external diameter of DLA; this result is rather strong, as it identifies precisely the growth exponent, but it is limited by the special qualities of the UITM. A natural next question is the following:

Problem 8.1. With (G,o)(G,o) a whole-plane γ\gamma-mated CRT rooted at o=0o=0, and (Kn)n∈N0(K_n)_{n\in\mathbb{N}_0} a DLA process from oo to ∞\infty, show that

lim⁡n→∞log⁡diam⁡(Kn)log⁡n\lim_{n\to\infty}\frac{\log\operatorname{diam}(K_n)}{\log n}

exists, and identify its value.

The above problem is essentially [29 Problem 8.1] stated for the γ\gamma-mated CRT.

The next more open ended problem deals with a limitation of the techniques of this paper. To deduce a.s. sublinear growth of the DLA process on an RPM, we require Assumption 1.4 or 1.8 to hold a.s. for the relevant model. As demonstrated in this paper, this does indeed hold for an extremely general class of random planar geometries. However, it does fail to capture the behavior of some natural models which do exhibit slightly weaker invariance properties; in particular, the recent work [34] exhibits quenched convergence in probability of the random walk (uniform on distinct neighboring vertices) on finite spanning-tree-weighted maps to Brownian motion, and indeed the work of [29] shows that the UITM is an especially natural geometry on which to study DLA. As we will discuss, these in-probability invariance theorems alone are insufficient to prove sublinearity of the internal diameter of DLA. Thus, the following problem is a natural modification, and asks if further assumptions are sufficient to recover a.s. sublinear diameter growth.

Problem 8.2. Adapt the methods of the current paper to RPMs satisfying an in probability quenched invariance principle under further assumptions.

To observe that in-probability convergence alone cannot yield the sublinear bound on DLA, consider the following example. Set rk=42kr_k=4^{2^k} and Qr=[−r,r]2∩Z2Q_r=[-r,r]^2\cap\mathbb{Z}^2 and independently with probability 1/k1/k replace Q2rk∖QrkQ_{2r_k}\setminus Q_{r_k} in Z2\mathbb{Z}^2 by a single nearest-neighbor path with Lk=rk(2rk+1)L_k=r_k(2r_k+1) vertices, retaining only consecutive edges and the two connections to the inner and outer rings. The resulting graph is connected, recurrent, and properly embedded in the plane, with maximum degree four and unique harmonic measure from infinity. Set o=0o=0 and take K0={0}K_0=\{0\}. For each fixed L>0L>0, with probability tending to one all modifications relevant to a walk stopped on exiting BLR1/4(0)B_{LR^{1/4}}(0) lie inside B4R1/4(0)B_{4R^{1/4}}(0). The latter invariance principle and planar annulus hitting estimate [43 Theorem 3.4.2 and Proposition 6.4.1] give the convergence in Assumption 1.4 in probability. However, infinitely many modifications occur a.s. and DLA must traverse them consecutively. At their completion times NkN_k, we have Nk≤(2rk+1)2+LkN_k\le(2r_k+1)^2+L_k and diam⁡(KNk)≥Lk\operatorname{diam}(K_{N_k})\ge L_k, so

lim sup⁡n→∞diam⁡(Kn)n≥13.\limsup_{n\to\infty}\frac{\operatorname{diam}(K_n)}{n}\ge\frac{1}{3}.

a.s. For am(G,o)a_m(G,o), monotonicity makes convergence to zero in probability equivalent to a.s. convergence, so Proposition 3.3 gives sublinear DLA diameter jointly a.s. However, in the counterexample we have am(G,0)=1a_m(G,0)=1 for every mm a.s., since arbitrarily large rooted sets can end inside a long path with a single entrance from infinity. Likewise, convergence in probability of the walk-law discrepancy uniformly over all later scales recovers a.s. convergence.

Specifically, for the UITM it seems likely that one may be able to achieve analogous statements to [26] and therefore achieve the a.s. Assumption 1.8 due to the recent work [34].

Problem 8.3. Using [34], establish analogues of the quantitative quenched annular estimates of [26] on the UITM, and then verify Assumption 1.8 for the UITM.

References

References

  1. [1]M. Aizenman and A. Burchard. Hölder regularity and dimension bounds for random curves. Duke Mathematical Journal, 99(3):419–453, 1999.DOI
  2. [2]Gideon Amir, Omer Angel, Itai Benjamini, and Gady Kozma. One-dimensional long-range diffusion-limited aggregation i. The Annals of Probability, pages 3546–3579, 2016.arxiv.org/abs/0910.4416
  3. [3]Morris Ang and Deven Mithal. Quantum loewner evolution in quantum natural time: phases and markov properties. arXiv preprint arXiv:2605.03385, 2026.arxiv.org/abs/2605.03385
  4. [4]Amine Asselah and Alexandre Gaudillière. Sublogarithmic fluctuations for internal DLA. The Annals of Probability, 41(3A):1160–1179, 2013.
  5. [5]Itai Benjamini and Ariel Yadin. Diffusion limited aggregation on a cylinder. Communications in Mathematical Physics, 279(1):187–223, 2008.DOI
  6. [6]Itai Benjamini and Ariel Yadin. Harmonic measure in the presence of a spectral gap. Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, 52(3):1050–1060, 2016.arxiv.org/abs/1402.0156
  7. [7]Itai Benjamini and Ariel Yadin. Upper bounds on the growth rate of diffusion limited aggregation. arXiv preprint arXiv:1705.06095, 2017.arxiv.org/abs/1705.06095
  8. [8]Nathanaël Berestycki and Diederik van Engelengurg. Harnack inequality and one-endedness of ust on reversible random graphs: N. berestycki, d. van engelenburg. Probability Theory and Related Fields, 188(1):487–548, 2024.
  9. [9]Noam Berger, Eviatar B Procaccia, and Amanda Turner. Growth of stationary hastings–levitov. The Annals of Applied Probability, 32(5):3331–3360, 2022.DOI
  10. [10]Leonardo Bonanno and Deven Mithal. Dla-internal-diameter. GitHub repository, https://github.com/liouvillegquantumgy/Formalizations/tree/main/DLA-internal-diameter.
  11. [11]Ahmed Bou-Rabee. Convergence of the random abelian sandpile. The Annals of Probability, 49(6):3168–3196, 2021.
  12. [12]Ahmed Bou-Rabee. Dynamic dimensional reduction in the abelian sandpile. Communications in Mathematical Physics, 390(2):933–958, 2022.
  13. [13]Ahmed Bou-Rabee. Integer superharmonic matrices on the f-lattice. Advances in Mathematics, 436:109400, 2024.
  14. [14]Ahmed Bou-Rabee. A shape theorem for exploding sandpiles. The Annals of Applied Probability, 34(1A):714–742, 2024.
  15. [15]Ahmed Bou-Rabee and Ewain Gwynne. Internal dla on mated-crt maps. The Annals of Probability, 52(6):2173–2237, 2024.
  16. [16]Sébastien Bubeck and Ewain Gwynne. Asymptotic behavior of the eden model with positively homogeneous edge weights. arXiv preprint arXiv:1508.05140, 2015.arxiv.org/abs/1508.05140
  17. [17]Lennart Carleson and Nikolai Makarov. Aggregation in the plane and loewner’s equation. Communications in Mathematical Physics, 216(3):583–607, 2001.
  18. [18]J Theodore Cox and Richard Durrett. Some limit theorems for percolation processes with necessary and sufficient conditions. The Annals of Probability, pages 583–603, 1981.DOI
  19. [19]Nicolas Curien and Jean-François Le Gall. Scaling limits for the peeling process on random maps. Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, 53(1):322–357, 2017.arxiv.org/abs/1412.5509
  20. [20]Nicolas Curien and Jean-François Le Gall. First-passage percolation and local modifications of distances in random triangulations. In Annales scientifiques de l’École Normale Supérieure, volume 52, pages 631–701, 2019.arxiv.org/abs/1511.04264
  21. [21]Bertrand Duplantier, Jason R Miller, and Scott Sheffield. Liouville quantum gravity as a mating of trees. Astérisque, 427, 2021.
  22. [22]Bertrand Duplantier and Scott Sheffield. Liouville quantum gravity and kpz. Inventiones mathematicae, 185(2):333–393, 2011.arxiv.org/abs/0808.1560
  23. [23]Murray Eden et al. A two-dimensional growth process. In Proceedings of the fourth Berkeley symposium on mathematical statistics and probability, volume 4, pages 223–239. Biology and Problems of Health, University of California Press, Berkeley, CA, 1961.
  24. [24]Ewain Gwynne, Nina Holden, and Xin Sun. A distance exponent for liouville quantum gravity. Probability Theory and Related Fields, 173(3):931–997, 2019.arxiv.org/abs/1606.01214
  25. [25]Ewain Gwynne and Tom Hutchcroft. Anomalous diffusion of random walk on random planar maps. Probability Theory and Related Fields, 178(1):567–611, 2020.arxiv.org/abs/1807.01512
  26. [26]Ewain Gwynne, Jason Miller, and Scott Sheffield. Harmonic functions on mated-CRT maps. Electronic Journal of Probability, 24:Paper No. 58, 55, 2019.arxiv.org/abs/1807.07511
  27. [27]Ewain Gwynne, Jason Miller, and Scott Sheffield. The Tutte embedding of the mated-CRT map converges to Liouville quantum gravity. The Annals of Probability, 49(4):1677–1717, 2021.
  28. [28]Ewain Gwynne, Jason Miller, and Scott Sheffield. An invariance principle for ergodic scale-free random environments. Acta Mathematica, 228(2):303–384, 2022.
  29. [29]Ewain Gwynne and Joshua Pfeffer. External diffusion-limited aggregation on a spanning-tree-weighted random planar map. The Annals of Probability, 49(4):1633–1676, 2021.
  30. [30]John M Hammersley and Dominic JA Welsh. First-passage percolation, subadditive processes, stochastic networks, and generalized renewal theory. In Bernoulli 1713, Bayes 1763, Laplace 1813: Anniversary Volume. Proceedings of an International Research Seminar Statistical Laboratory University of California, Berkeley 1963, pages 61–110. Springer, 1965.DOI
  31. [31]Matthew B Hastings and Leonid S Levitov. Laplacian growth as one-dimensional turbulence. Physica D: Nonlinear Phenomena, 116(1-2):244–252, 1998.DOI
  32. [32]MB Hastings. Fractal to nonfractal phase transition in the dielectric breakdown model. Physical Review Letters, 87(17):175502, 2001.
  33. [33]Frankie Higgs. SLE scaling limits for a Laplacian random growth model. Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, 58(3):1712–1739, 2022.
  34. [34]Nina Holden. The tutte embedding of tree-weighted planar maps converges to liouville quantum gravity. arXiv preprint arXiv:2610.10501, 2026.
  35. [35]Tom Hutchcroft and Asaf Nachmias. Uniform spanning forests of planar graphs. In Forum of Mathematics, Sigma, volume 7, page e29. Cambridge University Press, 2019.arxiv.org/abs/1603.07320
  36. [36]David Jerison, Lionel Levine, and Scott Sheffield. Logarithmic fluctuations for internal dla. Journal of the American Mathematical Society, 25(1):271–301, 2012.arxiv.org/abs/1010.2483
  37. [37]Harry Kesten. Hitting probabilities of random walks on zd. Stochastic Processes and their Applications, 25:165–184, 1987.
  38. [38]Harry Kesten. How long are the arms in dla? Journal of Physics A: Mathematical and General, 20(1):L29–L33, 1987.
  39. [39]Harry Kesten. Upper bounds for the growth rate of dla. Physica A: Statistical Mechanics and its Applications, 168(1):529–535, 1990.DOI
  40. [40]Gregory F. Lawler. Conformally invariant processes in the plane, volume 114 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2005.
  41. [41]Gregory F Lawler, Maury Bramson, and David Griffith. Internal diffusion limited aggregation. The Annals of Probability, pages 2117–2140, 1992.DOI
  42. [42]Gregory F. Lawler and Vlada Limic. The Beurling estimate for a class of random walks. Electronic Journal of Probability, 9(27):846–861, 2004.DOI
  43. [43]Gregory F. Lawler and Vlada Limic. Random Walk: A Modern Introduction, volume 123 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2010.DOI
  44. [44]Gregory F. Lawler, Oded Schramm, and Wendelin Werner. Conformal invariance of planar loop-erased random walks and uniform spanning trees. Ann. Probab., 32(1B):939–995, 2004.
  45. [45]Lionel Levine and Yuval Peres. Scaling limits for internal aggregation models with multiple sources. Journal d’Analyse Mathématique, 111(1):151–219, 2010.DOI
  46. [46]Ilya Losev and Stanislav Smirnov. How long are the arms in dbm? Communications in Mathematical Physics, 406(4):93, 2025.DOI
  47. [47]Joachim Mathiesen and Mogens H. Jensen. Tip splittings and phase transitions in the dielectric breakdown model: Mapping to the diffusion-limited aggregation model. Physical Review Letters, 88(23):235505, 2002.pubmed.ncbi.nlm.nih.gov/12059377
  48. [48]Joachim Mathiesen, Mogens H. Jensen, and Jan Øystein Haavig Bakke. Dimensions, maximal growth sites, and optimization in the dielectric breakdown model. Physical Review E, 77(6):066203, 2008.pubmed.ncbi.nlm.nih.gov/18643346
  49. [49]Jason Miller and Scott Sheffield. Quantum Loewner evolution. Duke Mathematical Journal, 165(17):3241–3378, November 2016.
  50. [50]Jason Miller and Scott Sheffield. Liouville quantum gravity and the Brownian map I: The QLE(8/3,0) metric, 2019.
  51. [51]Jason Miller and Scott Sheffield. Liouville quantum gravity and the Brownian map II: Geodesics and continuity of the embedding. Ann. Probab., 49(6):2732–2829, 2021.
  52. [52]Peter Mörters and Yuval Peres. Brownian Motion, volume 30 of Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, Cambridge, 2010.
  53. [53]Lucian Niemeyer, Luciano Pietronero, and Hans J Wiesmann. Fractal dimension of dielectric breakdown. Physical Review Letters, 52(12):1033–1036, 1984.DOI
  54. [54]James Norris, Vittoria Silvestri, and Amanda Turner. Stability of regularized Hastings–Levitov aggregation in the subcritical regime. Communications in Mathematical Physics, 405(3):74, 2024.
  55. [55]James Norris and Amanda Turner. Hastings–Levitov aggregation in the small-particle limit. Communications in Mathematical Physics, 316(3):809–841, 2012.DOI
  56. [56]Eviatar B. Procaccia, J iayan Ye, and Yuan Zhang. Stationary DLA is well defined. Journal of Statistical Physics, 181(4):1089–1111, 2020.DOI
  57. [57]Eviatar B. Procaccia and Yuan Zhang. Stationary harmonic measure and DLA in the upper half plane. Journal of Statistical Physics, 176(4):946–980, 2019.arxiv.org/abs/1711.01011
  58. [58]Daniel Richardson. Random growth in a tessellation. Proceedings of the Cambridge Philosophical Society, 74(3):515–528, 1973.DOI
  59. [59]Steffen Rohde and Michel Zinsmeister. Some remarks on Laplacian growth. Topology and its Applications, 152(1–2):26–43, 2005.DOI
  60. [60]Oded Schramm. Scaling limits of loop-erased random walks and uniform spanning trees. Israel J. Math., 118:221–288, 2000.
  61. [61]Stanislav Smirnov. Critical percolation in the plane: conformal invariance, Cardy’s formula, scaling limits. C. R. Acad. Sci. Paris Sér. I Math., 333(3):239–244, 2001.DOI
  62. [62]Stanislav Smirnov. Conformal invariance in random cluster models. i. holomorphic fermions in the ising model. Annals of mathematics, pages 1435–1467, 2010.
  63. [63]Alan Sola, Amanda Turner, and Fredrik Viklund. One-dimensional scaling limits in a planar Laplacian random growth model. Communications in Mathematical Physics, 371(1):285–329, 2019.DOI
  64. [64]Thomas A. Witten and Leonard M. Sander. Diffusion-limited aggregation. Physical Review B, 27(9):5686–5697, 1983.DOI
  65. [65]Thomas A Witten Jr and Leonard M Sander. Diffusion-limited aggregation, a kinetic critical phenomenon. Physical review letters, 47(19):1400, 1981.DOI

Paper details

vlnc:W_88vskse8
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents