Estimates of random walk exit probabilities and application to loop-erased random walk

We prove an estimate for the probability that a simple random walk in a simply connected subset A of Z^2 starting on the boundary exits A at another specified boundary point. The estimates are uniform over all domains of a given inradius. We apply these estimates to prove a conjecture of S. Fomin in 2001 concerning a relationship between crossing probabilities of loop-erased random walk and Brownian motion.


Introduction
In the last few years a number of results have been proved about scaling limits of twodimensional lattice systems in statistical mechanics. Site percolation on the triangular lattice [15], loop-erased random walk [13,16], uniform spanning trees [13], and the harmonic explorer [14] have all been shown to have limits that can be described using the Schramm-Loewner evolution. In the last three cases, the proofs use a version of the well-known fact that simple random walk has a scaling limit of Brownian motion, which is conformally invariant in two dimensions. What is needed is a strong version of this result which holds uniformly over a wide class of domains where the errors do not depend on the smoothness of the boundary. In this paper, we present another result of this type. It differs from the lemmas in [13] in two ways: we give explicit error bounds that show that the error decays as a power of the "inradius" of the domain, and the continuous domain that we compare to the discrete domain is slightly different.
We give an application of our result to loop-erased walk by proving a conjecture of S. Fomin [4] and giving a quick derivation of a crossing exponent first proved by R. Kenyon [6]. Fomin showed that a certain crossing probability for loop-erased walk can be given in terms of a determinant of hitting probabilities for simple random walk. Our estimate shows that this determinant approaches a corresponding determinant for Brownian motion. We then estimate the determinant for Brownian motion to derive the crossing exponent.
We will start by discussing the main results, leaving some of the precise definitions until Section 2. The only Euclidean dimension that will concern us is d = 2; consequently, we associate C ∼ = R 2 in the natural way. Points in the complex plane will be denoted by any of w, x, y, or z. A domain D ⊂ C is an open and connected set.
Throughout this paper, B t , t ≥ 0, will denote a standard complex Brownian motion, and S n , n = 0, 1, . . ., will denote two-dimensional simple random walk, both started at the origin unless otherwise noted. We write B[0, t] := {z ∈ C : B s = z for some 0 ≤ s ≤ t}, and S[0, n] := [S 0 , S 1 , . . . , S n ] for the set of lattice points visited by the random walk. We will generally use T for stopping times for Brownian motion, and τ for stopping times for random walk. We write E x and P x for expectations and probabilities, respectively, assuming B 0 = x or S 0 = x, as appropriate.

Main results
Let A n denote the collection of simply connected subsets A of Z 2 such that n ≤ inrad(A) ≤ 2n, i.e., such that n ≤ sup{|z| : z ∈ Z 2 \ A} ≤ 2n.
Associated to A is a simply connected domainÃ ⊂ C which is obtained by identifying each lattice point in A with the square of side one centred at that point. By the Riemann mapping theorem, there is a unique conformal transformation f A ofÃ onto the unit disk with f A (0) = 0, f ′ A (0) > 0. We let θ A (x) := arg(f A (x)). We can extend θ A to ∂A in a natrual way.
If x ∈ A, y ∈ ∂A, let h A (x, y) be the probability that a simple random walk starting at x leaves A at y. If x, y ∈ ∂A, let h ∂A (x, y) be the probability that a simple random walk starting at x takes its first step into A and then leaves A at y.
In this theorem, and throughout this paper, we will use O(·) for uniform error terms that depend only on n. For example, the statement above is shorthand for the following: there is a constant c < ∞ such that for all A ∈ A n and all x, y ∈ ∂A with |θ A (x) − θ A (y)| ≥ n −1/16 log 2 n, When discussing k-fold determinants, we will have error terms that depend also on a positive integer k; we will write these as O k (·). We do not believe the error term O(n −1/16 log n) is optimal, and we probably could have improved it slightly in this paper. However, our methods are not strong enough to give the optimal error term. The importance of this result is that the error is bounded uniformly over all simply connected domains and that the error is in terms of a power of n. For domains with "smooth" boundaries, one can definitely improve the power of n.
To help understand this estimate, one should consider h ∂A (x, y) as having a "local" and a "global" part. The local part, which is very dependent on the structure of A near x and y, is is the conformal invariant and depends only on the image of the points under the conformal transformation ofÃ onto the unit disk. In contrast to the discrete case, Example 2.14 shows that the Brownian version of this result is exact.
As part of the proof, we also derive a uniform estimate for G A (x), the expected number of visits to x before leaving A of a simple random walk starting at 0. Let a denote the potential kernel for two-dimensional simple random walk. It is known that there is a k 0 such that | is the Green's function for Brownian motion inÃ and k x := k 0 + 2 π log |x| − a(x).

Fomin's identity for loop-erased walk
We briefly review the definition of the loop-erased random walk; see [10,Chapter 7] and [11] for more details. Since simple random walk in Z 2 is recurrent, it is not possible to construct loop-erased random walk by erasing loops from an infinite walk. However, the following looperasing procedure makes perfect sense since it assigns to each finite simple random walk path a self-avoiding walk. Let S := S[0, m] := [S 0 , S 1 , . . . , S m ] be a simple random walk path of length m. We construct L(S), the loop-erased part of S, recursively as follows. If S is already self-avoiding, set L(S) = S. Otherwise, let s 0 = max{j : S j = S 0 }, and for i > 0, let Suppose that A ∈ A n and x 1 , . . . , x k , y k , . . . , y 1 are distinct points in ∂A, ordered counterclockwise. For i = 1, . . . , k, let L i = L(S i ) be the loop erasure of the path , and let C = C(x 1 , . . . , x k , y k , . . . , y 1 ; A) be the event that both and The following theorem was proved in [4] which relates a determinant of simple random walk probabilities to a "crossing probability" for loop-erased random walk. Theorem 1.3 (Fomin). If C is the event defined above, and This is a special case of an identity that Fomin established for general discrete stationary Markov processes. In his paper, he made the following conjecture.
In order for the statement of Theorem 1.3 to make sense, the Markov process under consideration does not have to be discrete. . . . The proofs can be obtained by passing to a limit in the discrete approximation. The same limiting procedure can be used to justify the well-definedness of the quantities involved; notice that in order to define a continuous analogue of Theorem 1.3, we do not need the notion of loop-erased Brownian motion. Instead, we discretize the model, compute the probability, and then pass to the limit. One can further extend these results to densities of the corresponding hitting distributions. Technical details are omitted.
With Theorem 1.1, we have taken care of the "technical details" in the case of simply connected planar domains. Note that represents the conditional probability that (3) holds given (2) holds. Suppose D is a smooth Jordan domain, and that x 1 , . . . , x k , y k , . . . , y 1 are distinct points on ∂D ordered counterclockwise. The "Brownian motion" analogue of the determinant (4) is where H ∂D (z, w) denotes the excursion Poisson kernel. In the case D = D, if z = e iθ , w = e iθ ′ , then .
Conformal covariance of the excursion Poisson kernel shows that where f is a conformal transformation of D onto D and θ D (z) := arg(f (z)).
Corollary 1.4. Suppose A ∈ A n and x 1 , . . . , x k , y k , . . . , y 1 are distinct points in ∂A ordered counterclockwise. Let If m ≥ n −1/16 log 2 n, then But, if |δ j,l | ≤ ε, multilinearity of the determinant and the estimate det Using the corollary, we know that we can approximate the determinant for random walks, and hence the probability of the crossing event C, in terms of the corresponding quantity for Brownian motion, at least for simply connected domains. We will consider the asymptotics of Λ D (x 1 , . . . , x k , y k , . . . , y 1 ) when x 1 , . . . , x k get close and y 1 , . . . , y k get close. Since this quantity is a conformal invariant, we may assume that D = R L , where R L = { z : 0 < Re(z) < L, 0 < Im(z) < π }, and x j = iq j , y j = L + iq ′ j , where 0 < q k < · · · < q 1 < π and 0 < q ′ k < · · · < q ′ 1 < π. Proposition 1.5. As L → ∞, This crossing exponent k(k − 1)/2 was first proved by Kenyon [6] for loop-erased walk. We describe this result in the framework of Brownian excursion measure.

Outline of the paper
In the first five subsections of Section 2, we review facts about random walk, Brownian motion, and conformal mapping that we will need. In the remaining subsections, we review Brownian excursion measure, define the analogue of the Fomin determinant for excursion measure, and then derive Proposition 1.5. In Section 3, we begin with a brief review of strong approximation, before proving Theorem 1.2 in the second subsection. The final two subsections contain the proof of the other main result, Theorem 1.1.

Background, notation, and preliminary results
In this section we review some basic material that will be needed in subsequent parts and standardize our notation. Almost all of the complex analysis is well-known, and may be found in a variety of sources; we prove several elementary results, but often refer the reader to the literature for details. The material on the excursion Poisson kernel is not difficult, but these results are not widespread.

Simply connected subsets of C and Z 2
We will use D to denote domains, i.e., open connected subsets of C. We write D := {z ∈ C : |z| < 1} to denote the open unit disk, and H := {z ∈ C : Im(z) > 0} to denote the upper half plane. An analytic, univalent (i.e, one-to-one) function 1 is called a conformal mapping.
We say that f : D → D ′ is a conformal transformation if f is a conformal mapping that is onto D ′ . It follows that f ′ (z) = 0 for z ∈ D, and f −1 : D ′ → D is also a conformal transformation. We write S to denote the set of functions f which are analytic and univalent in D satisfying the normalizing conditions f (0) = 0 and f ′ (0) = 1.
If D ⊂ C with 0 ∈ D, we define the radius (with respect to the origin) of D to be rad(D) := sup{|z| : z ∈ ∂D}, and the inradius (with respect to the origin) of D to be inrad(D) := dist(0, ∂D) := inf{|z| : z ∈ ∂D}. The diameter of D is given by diam(D) := sup{|x − y| : x, y ∈ D}. If D ⊂ C, then we say that a bounded D is a Jordan domain if ∂D is a Jordan curve (i.e., homeomorphic to a circle). A Jordan domain is nice if the Jordan curve ∂D can be expressed as a finite union of analytic curves. Note that Jordan domains are simply connected. For each r > 0, let D r be the set of nice Jordan domains containing the origin of inradius r, and write D := r>0 D r . We also define D * to the be set of Jordan domains containing the origin, and note that D ∈ D ⊂ D * . If D, D ′ ∈ D * , let T (D, D ′ ) be the set of all f : D → D ′ that are conformal transformations of D onto D ′ . The Riemann mapping theorem implies that T (D, D ′ ) = ∅, and since ∂D, ∂D ′ are Jordan, the Carathéodory extension theorem tells us that f ∈ T (D, D ′ ) can be extended to a homeomorphism of D onto D ′ . We will use this fact repeatedly throughout, without explicit mention of it. For statements and details on these two theorems, consult [3, §1.5].
A subset A ⊂ Z 2 is connected if every two points in A can be connected by a nearest neighbour path staying in A. We say that a finite subset A is simply connected if both A and Z 2 \ A are connected. There are three standard ways to define the "boundary" of a proper subset A of Z 2 : • (outer) boundary: ∂A := {y ∈ Z 2 \ A : |y − x| = 1 for some x ∈ A}; • inner boundary: To each finite, connected A ⊂ Z 2 we associate a domainÃ ⊂ C in the following way. For each edge (x, y) ∈ ∂ e A, considered as a line segment of length one, let ℓ x,y be the perpendicular line segment of length one intersecting (x, y) in the midpoint. Let ∂Ã denote the union of the line segments ℓ x,y , and letÃ denote the domain with boundary ∂Ã containing A. Observe that That is, S x is the closed square of side length one centred at x whose sides are parallel to the coordinate axes. Also, note thatÃ is simply connected if and only if A is a simply connected subset of Z 2 . We refer toÃ as the "union of squares" domain associated to A. Let A denote the set of all finite simply connected subsets of Z 2 containing the origin. If A ∈ A, let inrad(A) := min{|z| : z ∈ Z 2 \ A} and rad(A) := max{|z| : z ∈ A} denote the inradius and radius (with respect to the origin), respectively, of A, and define A n to be the set of

Green's functions on C
If D is a domain whose boundary includes a curve, let g D (x, y) denote the Green's function for Brownian motion. If x ∈ D, we can define g D (x, ·) as the unique harmonic function on D \ {x}, vanishing on ∂D (in the sense that g D (x, y) → 0 as y → y 0 for every regular In the case D = D, we have Note that g D (0, x) = − log |x|, and g D (x, y) = g D (y, x). The Green's function is a well-known example of a conformal invariant (see, e.g., [3, §1.8]).
Hence, to determine g D for arbitrary D ∈ D * , it is enough to find f D ∈ T (D, D). On the other hand, suppose D ⊂ C is a simply connected domain containing the origin with Green's function g D (z, w). Then we can write the unique conformal transformation of where An equivalent formulation of the Green's function for D ∈ D * can be given in terms of Brownian motion. Using (8) Additional details may be found in [12,Chapter 2]. If A ∈ A, and if we let g A (x, y) := gÃ(x, y) be the Green's function (for Brownian motion) inÃ, then by Proposition 2.1 and (9) we have that If we write θ A := θÃ, then (10) implies that

Green's functions on Z 2
Suppose that S n is a simple random walk on Z 2 and A is a proper subset of Z 2 . If τ A := min{j ≥ 0 : S j ∈ A}, then we let denote the Green's function for random walk on A.
In analogy with the Brownian motion case (see [10, Proposition 1.6.3]), we have where a is the potential kernel for simple random walk defined by It is also known [10, Theorem 1.6.2] that as |x| → ∞, where k 0 := (2ς + 3 ln 2)/π and ς is Euler's constant. The error above will suffice for our purposes, even though stronger results are known. The asymptotic expansion of a(x) given in [5] shows that the error is O(|x| −2 ).

Consequences of the Koebe theorems
We now recall some standard results from the study of univalent functions. Proofs may be found in [ A number of useful consequences may now be deduced.
Proof. If we combine the first estimate in Theorem 2.3 with the estimate of |f ′ (z)| in the third statement of that theorem, then we can obtain a uniform bound on |f ′′ (z)| over all f ∈ S and |z| ≤ r.

Recall that f
Using the Koebe one-quarter theorem and the Schwarz lemma, we see that if F : D → D is a conformal transformation with F (0) = 0, then . Along with Corollary 2.4, the growth and distortion theorem yields the following. We remark that this corollary implies lim |x|→0 (g A (x) + log |x|) = − log f ′ A (0) which shows the size of the error term in (8).

Beurling estimates and related results
Throughout this subsection, suppose that A ∈ A n with associated "union of squares" domaiñ A ⊂ C, and write T A := TÃ := inf{t : B t ∈Ã}. From the Beurling projection theorem [2, Theorem (V.4.1)] the following may be derived.
Proof. Without loss of generality, we may assume by Brownian scaling that x = 0 and inrad(Ã) =: In particular, if |x| > n/2, the probability starting at x of reaching D n/2 before leaving A i s bounded above by cn −1/2 dist(x, ∂Ã) 1/2 . From the Koebe one-quarter theorem it easily follows that g A (x) ≤ c for |x| ≥ n/4; hence we get Recall from (13) The Beurling estimate and a simple Harnack principle show that There are analogous Beurling-type results in the discrete case; the following is a corollary of [10, Theorem 2.5.2]. Let τ A := min{j ≥ 0 : S j ∈ A}. Corollary 2.9 (Discrete Beurling Estimate). There is a constant c < ∞ such that if In particular, if |x| > n/2, the probability starting at x of reaching D n/2 before leaving A is bounded above by cn −1/2 dist(x, ∂A) 1/2 . It is easy to show that G A (x) ≤ c for |x| ≥ n/4; hence in this case we get We can replace A \ {x} in the above formula with the connected component of A \ {x} containing the origin. In particular, if A ∈ A n and x ∈ ∂ i A, then we conclude that

Excursion Poisson kernel
Let D be a domain in C. We say that a connected Γ is an (open) analytic arc of ∂D if there is a domain D ′ ⊂ C that is symmetric about the real axis and a conformal map f : where H denotes the upper half plane. We say that ∂D is locally analytic at x ∈ ∂D, if there exists an analytic arc of ∂D containing x. If Γ is an analytic arc of ∂D and x, y ∈ Γ, we write x < y if f −1 (x) < f −1 (y). We let n x := n x,D be the unit normal at x pointing into D. If x ∈ D and ∂D is locally analytic at y ∈ ∂D, then both harmonic measure P x {B T D ∈ dy}, and its density with respect to arc length, the Poisson kernel H D (x, y), are well-defined. Also, recall that for fixed z ∈ D, the function y → H D (z, y) is continuous in y, and that for fixed y ∈ ∂D, the function z → H D (z, y) is harmonic in z. If x ∈ D and Γ ⊂ ∂D is an analytic arc, then write The Riemann mapping theorem and Lévy's theorem on the conformal invariance of Brownian motion [2] allow us to describe the behaviour of the Poisson kernel under a conformal transformation.
Proposition 2.10. If f : D → D ′ is a conformal transformation, x ∈ D, ∂D is locally analytic at y ∈ ∂D, and ∂D ′ is locally analytic at f (y), then P For each ε > 0, let µ x,ε,D denote the probability measure on paths obtained by starting a Brownian motion at x + εn x and stopping the path when it reaches ∂D. The excursion measure of D at x is defined by Excursion measure on Γ is defined by If Υ is another analytic arc on ∂D, we define E D (Γ, Υ) to be the excursion measure E D (Γ) restricted to curves that end at Υ and whose endpoints are different. The excursion boundary measure is defined by where | · | denotes total mass. We can write We can also write where E # D (z, w) := E D (z, w)/H ∂D (x, y) is the excursion measure between x and y normalized to be a probability measure. We will consider E D and E # D as measures on curves modulo reparametrization. Conformal invariance of complex Brownian motion implies that E # D is conformally invariant, i.e., if f : D → D ′ is a conformal transformation, then f (y)). Proposition 2.11. Suppose f : D → D ′ is a conformal transformation and x, y are distinct points on ∂D. Suppose that ∂D is locally analytic at x, y and ∂D ′ is locally analytic at Proof. By definition, H ∂D (x, y) := lim ε→0+ ε −1 H D (x + εn x , y). Therefore, where we have written ε 1 := ε|f ′ (x)|, and have noted that n f (x) = f ′ (x)n x /|f ′ (x)| and |n f (x) | = |n x | = 1. Example 2.13. If R L = {z : 0 < Re(z) < L, 0 < Im(z) < π}, then separation of variables can be used to show that for 0 < r < L, 0 < q, q ′ < π, sinh(nr) sin(nq) sin(nq ′ ) sinh(nL) , Example 2.14. In the Brownian excursion case, there is an exact form of Theorem 1.1. Suppose that D ∈ D, and x, y ∈ ∂D with ∂D locally analytic at x and y. .

We finish this subsection by stating a formula relating the excursion Poisson kernel and the Green's function for Brownian motion:
For the simply connected case, which is all that we will use, this follows from a straightforward computation. We omit the proof in the general case.
A straightforward extension of Proposition 2.11 is that the determinant of the hitting matrix of excursion Poisson kernels is conformally covariant. a conformal transformation, x 1 , . . . , x k ,y 1 , . . . , y k are distinct points at which ∂D is locally analytic, and ∂D ′ is locally analytic at f (x 1 ), . . . , f (x k ), f (y 1 ), . . . , f (y k ), then Proof. By the definition of determinant, where the sum is over all permutations σ of {1, . . . , k}, and sgn σ denotes the sign of the permutation.
It follows from this proposition that Λ D (x 1 , . . . , x k , y k , . . . , y 1 ) given by (5) is a conformal invariant, and hence it can be defined for any Jordan domain by conformal invariance.

Proof of Proposition 1.5
By applying Proposition 2.16 and judiciously choosing D = R L , we now give the asymptotics as points gets close, and complete the proof of Proposition 1.5.
Proof of Proposition 1.5. Let q = (q 1 , . . . , q k ), q ′ = (q ′ 1 , . . . , q ′ k ), Using (25), we see that (π/2) k det H ∂R L (iq, L + iq ′ ) can be written as By multilinearity of the determinant, we can write the determinant above as The determinants in the last sum equal zero if the indices are not distinct. Also it is not difficult to show that Hence, except for an error of O k (e −(k 2 +k+2)L/2 ), we see that (π/2) k det H ∂R L (iq, L + iq ′ ) equals where the sum is over all permutations σ of {1, . . . , k}. But which up to an error of O k (e −(k 2 +k+2)L/2 ) equals To finish the proof, note that from (25), we can also write so that ].

2.9
The discrete excursion Poisson kernel Definition 2.17. Suppose that A is a proper subset of Z 2 , and let τ A := min{j > 0 : S j ∈ ∂A}. For x ∈ A and y ∈ ∂A, define the discrete Poisson kernel (or hitting probability) to be We define the discrete analogue of the excursion Poisson kernel to be the probability that a random walk starting at x ∈ ∂A takes its first step into A, and then exits A at y. Definition 2.18. Suppose that A is a proper subset of Z 2 , and let τ A := min{j > 0 : S j ∈ ∂A}. For x, y ∈ ∂A, define the discrete excursion Poisson kernel to be h ∂A (x, y) : Also, a last-exit decomposition gives where G A is the Green's function for simple random walk on A as in (14). Hence, which is a discrete analogue of (26).

Proofs of the Main Results
After briefly reviewing strong approximation and establishing some ancillary results in the first subsection, we devote the second subsection to the proof of Theorem 1.2 which relates the Green's function for Brownian motion to the Green's function for simple random walk in certain domains. While this result holds in general, it is most useful for points away from the boundary. In the final two subsections, we prove Theorem 1.1 by obtaining better estimates for the case of points near the boundary, provided they are not too close to each other. Throughout this section, suppose that A ∈ A n with associated "union of squares" domaiñ A ∈ D. As in Section 2.1, let where g A is the Green's function for Brownian motion inÃ.

Strong approximation
In order to establish our Green's function estimates, we will need to establish a strong approximation result making use of the theorem of Komlós, Major, and Tusnády [7,8].
Since we are concerned exclusively with complex Brownian motion and simple random walk on Z 2 , the results noted in [1] suffice.

Theorem 3.1 (Komlós-Major-Tusnády).
There exists c < ∞ and a probability space (Ω, F , P) on which are defined a two-dimensional Brownian motion B and a two-dimensional simple random walk S with B 0 = S 0 , such that for all λ > 0 and each n ∈ N, Here S t is defined for noninteger t by linear interpolation. The one-dimensional proof may be found in [8], and the immediate extension to two dimensions is written down in [1, Lemma 3]. For our purposes, we will need to consider the maximum up to a random time, not just a fixed time. The following strong approximation will suffice. Our choice of n 8 is arbitrary and will turn out to be good enough.
There exists a constant c such that for every n, a Brownian motion B and a simple random walk S can be defined on the same probability space so that if A ∈ A n , 1 < r ≤ n 20 , and x ∈ A with |x| ≤ n 3 , then Proof. For any given n, letB and S be defined as in Corollary 3.2 above, and let C be the constant in that corollary. Define T ′ A := inf{t ≥ 0 : dist(B t / √ 2, ∂Ã) ≤ 2C log n}, τ ′ A := inf{t ≥ 0 : dist(S t , ∂A) ≤ 2C log n},T A = inf{t ≥ 0 :B t / √ 2 ∈ ∂A}. and consider the events By the Beurling estimates (Corollaries 2.8 and 2.9), and the strong Markov property (applied separately to the random walk and the Brownian motion), it follows that P( 3.2 Proof of Theorem 1.2 Lemma 3.4. There exists a c such that if A ∈ A n and |x| ≤ n 2 , then Proof. For any n, let B and S be as in Proposition 3.3, and let V = V (n) be the event that |B T A − S τ A | ≤ n 2/3 log n. By that proposition, P(V ) ≥ 1 − cn −1/3 . Since inrad(A) ≥ n, we know that on the event V , Using the Beurling estimates it is easy to see that and similarly for log |S τ A | in the last two estimates. Hence, . For any A ∈ A n , let A * ,n be the set The choice of 1/16 for the exponent is somewhat arbitrary, and slightly better estimates might be obtained by choosing a different exponent. However, since we do not expect the error estimate derived here to be optimal, we will just make this definition.
Corollary 3.5. If A ∈ A n , and x ∈ A * ,n , y ∈ A, then Proof. From the Beurling estimates, (20), and (22), it follows that if A ∈ A n and x ∈ A * ,n , then dist(x, ∂A) ≥ cn 7/8 . That is, if we translate A to make x the origin, then the inradius of the translated set is at least cn 7/8 . Hence, we can use Theorem 1.2 to deduce (30).

An estimate for hitting the boundary
Suppose that A is any finite, connected subset of Z 2 , not necessarily simply connected, and let V ⊂ ∂A be non-empty. Recall thatÃ ⊂ C is the "union of squares" domain associated to A as in Section 2.1. For every y ∈ V , consider the collection of edges containing y, namely E y := {(x, y) ∈ ∂ e A}, and set E V := y∈V E y . If ℓ x,y is the perpendicular line segment of length one intersecting (x, y) in the midpoint as in Section 2.1, then defineṼ := to be the associated boundary arc in ∂Ã. Suppose that τ A := min{j : S j ∈ ∂A}, T A = TÃ := inf{t : B t ∈ ∂Ã}, and throughout this subsection, write and where h A and HÃ are the discrete Poisson kernel and the Poisson kernel, respectively.
Definition 3.6. If F : Z 2 → R, let L denote the discrete Laplacian defined by Let ∆ denote the usual Laplacian in C, and recall that F : C → R is harmonic at x if ∆F (x) = 0. Note that L is a natural discrete analogue of ∆. If r > 1, F : C → R, and ∆F (x) = 0 for all x ∈ C with |x| < r, then Taylor's series and uniform bounds on the derivatives of harmonic functions [2] imply that Note that h defined by discrete (31) is discrete harmonic in A, and H defined by (32), is harmonic inÃ. Our goal in the remainder of this subsection is to prove the following proposition.
We first note that for every n < ∞, there is a δ ′ = δ ′ (n) > 0 such that the proposition holds for all A of cardinality at most n and all ε > 0. This is because h and H are strictly positive (since V = ∅) and the collection of connected subsets of Z 2 containing the origin of cardinality at most n is finite. Hence, we can choose where the minimum is over all finite connected A of cardinality at most n, all x ∈ A, and all y ∈ ∂A. We now extend this to all A for x near the boundary. Proof. By the recurrence of planar Brownian motion, we can find an N = N(M, ε) such that and the lemma holds with δ = δ ′ (3N), say, where δ ′ is defined as in (34) Proof. For any function F on A and any x ∈ A, (Note that F is bounded since A is finite.) Applying (35) to H gives Since H is harmonic and bounded by 1, (33) implies |LH(y)| ≤ c dist(y, ∂A) −3 . A routine estimate shows that there is a constant c such that for all A ∈ A, x ∈ A, and r ≥ 1, if a simple random walk is within distance r of the boundary, then the probability that it will hit the boundary in the next r 2 steps is bounded below by c/ log r. Consequently, we have r≤dist(y,∂A)≤2r Combining these estimates gives The proof is completed by choosing M sufficiently large which will guarantee that the right side of (36) is smaller than ε.
Thus, in view of the estimates (37) and (38) above, there is nothing surprising about the leading terms in (40) and (41). Proposition 3.10 essentially says that these relations are valid, at least in the dominant term, if we replace g A with (π/2) G A . Theorem 1.1 follows from this proposition and (28).
The hardest part of the proof is a lemma that states that if the random walk starts at a point x with f A (x) near ∂D, then, given that the walk does not leave A, f A (S j ) moves a little towards the centre of the disk before its argument changes too much.
Lemma 3.11. For A ∈ A n , let η = η(A, n) := min{j ≥ 0 : S j ∈ A * ,n ∪ A c }. There exist constants c, α such that if A ∈ A n , x ∈ A \ A * ,n , and r > 0, then In particular, there is a c 0 such that if In order to prove this lemma, we will need to establish several ancillary results. Therefore, we devote Section 3.4.1 which follows to the complete proof of this lemma, and then prove Proposition 3.10 in the separate Section 3.4.2.

Proof of Lemma 3.11
If A ∈ A and x ∈ A, let d A (x) be the distance from f A (x) to the unit circle. Note that (13). As a first step in proving Lemma 3.11, we need the following.
Remark. (42) is not completely obvious since the random walk takes discrete steps.
Proof. We start by stating three inequalities whose verification we leave to the reader. These are simple estimates for conformal maps on domains that are squares or the unions of two squares. Recall that S x is the closed square of side one centred at x whose sides are parallel to the coordinate axes. There exists a constant c 2 ∈ (1, ∞) such that if A ∈ A; x, y, w ∈ A; x = y; |w − x| = 1, then d A (z) ≤ c 2 d A (x) for z ∈ S x ; |f A (z) − f A (x)| ≤ c 2 |f A (z ′ ) − f A (x)| for z, z ′ ∈ S y ; and |f A (w)−f A (x)| ≤ c 2 d A (x). The first of these inequalities implies that if z ∈ S y and d A (z) ≥ 3c 2 d A (x), then d A (y) ≥ 3d A (x). Fix A ∈ A, x ∈ A with d A (x) ≤ 1/(100c 2 2 ), and let J = J(x, A) := y ∈ A : S y ∩ {z ∈Ã : That is, y ∈ J if there is a z ∈ S y with |f A (z)−f A (x)| < 5c 2 d A (x). Note that J is a connected subset of A (although it is not clear whether it is simply connected) andJ ⊆ {z ∈Ã : There is a positive probability ρ 1 that a Brownian motion in D starting at f A (x) leaves D before leaving the disk of radius 2d A (x) about f A (x). By conformal invariance, this implies that with probability at least ρ 1 , a Brownian motion starting at x leavesJ at ∂Ã. Hence by Proposition 3.7, there is an ε 1 such that P x {S τ J ∈ A} ≥ ε 1 .
Similarly, there is a positive probability ρ 2 that a Brownian motion in the disk starting at f A (x) reaches [1 − 6c 2 2 d A (x)] D before leaving D and before leaving the set {z : x) 2 on this set. In particular, with probability at least ρ 2 , a Brownian motion starting at x leavesJ at a point z with d A (z) ≥ 4c 2 d A (x). Such a point z must be contained in an S y with d A (y) ≥ 4d A (x). Hence, again using Proposition 3.7, there is a positive probability ε 2 that a random walk starting at x reaches a point y ∈ J with d A (y) ≥ 4d A (x) before leaving J. In the notation of Lemma 3.12 choose c := 1/(100c 2 2 ), let c ′ be the c ′ mentioned above, and let ε := min{ε 1 , ε 2 }. Then we have already shown that P which implies that |θ A (w) − θ A (x)| ≤ c ′′ d A (x) for an appropriate c ′′ . This gives the last assertion in Lemma 3.12 and completes the proof. Corollary 3.13. There exist constants c, c ′ , α such that if a ∈ (0, 1/2), A ∈ A, and x ∈ A with d A (x) < a, then the probability that a random walk starting at x reaches the set {y ∈ A : d A (y) ≥ a} without leaving the set {y ∈ A : |θ A (y) − θ A (x)| ≤ c ′ a} is at least c(d A (x)/a) α .
Proof. Let Q(r, a, b) be the infimum over all A ∈ A and x ∈ A with d A (x) ≥ r of the probability that a random walk starting at x reaches the set {y ∈ A : d A (y) ≥ a} without leaving the set {y ∈ A : |θ A (y) − θ A (x)| ≤ b}. It follows from Lemma 3.12 that there exist q > 0, ρ < 1, and a c ′′ such that Q(ρ k , ρ j , b) ≥ q Q(ρ k−1 , ρ j , b − c ′′ ρ k ). By iterating this we get Q(ρ k , ρ j , 2 c ′′ (1 − ρ) −1 ρ j ) ≥ q k−j . This and obvious monotonicity properties give the result.

Remark.
A similar proof gives an upper bound of c 1 (d A (x)/a) α 1 , but this result is not needed.