CIRCLE PATTERNS WITH THE COMBINATORICS OF THE SQUARE GRID

Oded Schramm

The Weizmann Institute

Abstract. Explicit families of entire circle patterns with the combinatorics of the square grid are constructed, and it is shown that the collection of entire, locally univalent circle patterns on the sphere is in nite dimensional. In Particular, Doyle's conjecture is false in this setting. Mobius invariants of circle patterns are introduced, and turn out to be discrete analogs of the Schwarzian derivative. The invariants satisfy a nonlinear discrete version of the Cauchy-Riemann equations. A global analysis of the solutions of these equations yields a rigidity theorem characterizing the Doyle spirals. It is also shown that by prescribing boundary values for the Mobius invariants, and solving the appropriate Dirichlet problem, a locally univalent meromorphic function can be approximated by circle patterns.

1991 Mathematics Subject Classi cation. 30C99, 05B40, 30D30, 31A05, 31C20, 30G25. Key words and phrases. Meromorphic functions, Schwarzian derivative, rigidity, error function, Dirichlet problem. Some of the work exposed here was done while the author visited UCSD, and the author wishes to express thanks to the UCSD mathematics department for its hospitality. This research was partially supported by NSF grant DMS94-03548. Typeset by AMS-TEX

ODED SCHRAMM

1. Introduction

Some aspects of the relatively new theory of circle patterns (or packings) and their relations to analytic functions can be considered as well understood, while other aspects are still very mysterious and enigmatic. The roots of the topic are in the circle packing theorem 18] which is a uniformization result, and the uniformization theory of circle patterns is now well developed 29], 1], 16]. There has also been steady progress in understanding the convergence of circle packings to conformal maps 22], 12], 7], 17]. Branched packings were considered, and analogues of polynomials and nite Blaschke products have been studied 3], 9], 10]. Also, a satisfying crop of applications and by-products has emerged 21], 4], 20], 24], 15], 13], 2]. However, there is also a darker side: very little is known about circle pattern analogues of entire functions. It seems that Peter Doyle was the rst to look into this area, and constructed entire immersed hexagonal circle patterns analogous to the exponential map. Doyle conjectured that these immersed packings, which came to be known as Doyle spirals, are the only entire immersed hexagonal circle packings. To date, the only other contribution to this topic seems to be Callahan and Rodin's 5]. They showed that entire immersed hexagonal circle packings form regularly exhaustible surfaces, and therefore satisfy Ahlfors' value distribution theory. In particular, Picard's theorem is valid in this setting. We have found that a framework of circle patterns based on the square grid, which we call SG patterns, is more tractable than the traditional hexagonal patterns. It turns out that Doyle's conjecture is false for SG patterns: there is 2an SG ?p R pattern analogous to the entire function erf iz , where erf(z) = z e?w dw is the error function. We also show that the collection of entire SG patterns in the sphere is in nite dimensional, and exhibit some explicit nite dimensional families. \Explicit" means that there is a closed form expression for the radius of any given circle in the pattern. Quite possibly, more explicit nite dimensional families could be found. This would be very interesting. In joint work, yet unpublished, with Rick Kenyon, we have discovered explicit branched SG patterns analogous to polynomials and to the function log z . Much of the theory of the hexagonal packings can be carried over to SG patterns. In particular, Callahan and Rodin's work 5] applies with virtually no modi cations. On the other hand, many of the results given below for SG patterns are not known for hexagonal packings, and proving (or disproving) them in the hexagonal setting would be a worthy challange. We start with the traditional approach of studying the radius function. In may respects, the `discrete nonlinear Laplace equation' that the radius function of an SG pattern satis es is more amenable than the corresponding equation in the hexagonal ?p setting. This is what permits the construction of the erf iz pattern. We show ?p that the surface detrmined by the erf iz SG pattern is isometric to the surface determined by the classical error function, up to a constant scaling factor. It is also demonstrated that in a limited but well de ned sense, there is no SG pattern p analog of the erf function itself, without i pre-rotation. But the radius function can only take us so far. The analysis of the radius function is the analysis of the circle pattern invariants for the group of isometries

SG CIRCLE PATTERNS

of the plane. Our main progress is in the analysis of the circle pattern invariants for the group of Mobius transformations. We introduce Mobius invariants and , and discover that they are closely analogous to the real and imaginary parts of the Schwarzian derivative. Naturally, it turns out that and satisfy a discrete nonlinear version of the Cauchy-Riemann equations. After the invariant is eliminated from these equations, one obtains a nonlinear discrete version of the Laplace equation that the invariant satis es. Conversely, any that satis es this equation has a one parameter family of companion invariants as co-solutions of the nonlinear Cauchy-Riemann system. We show that the nonlinear SG -Laplace equation satis es the Dirichlet principle. This gives a parametrization of SG patterns corresponding to nite `simply connected' sub-pieces of the square lattice, in terms of the boundary values of the invariant, and one value of the invariant. That's the local analysis. Two global rigidity results are presented. First, the uniqueness of the embedded SG pattern is established. That is analogous to the rigidity of the hexagonal circle packing, and shows that the Rodin-Sullivan theorem is valid in the SG setting. The second rigidity result is more novel: if the invariant of an entire SG pattern in the sphere is bounded, then the pattern is Mobius equivalent to a Doyle spiral (or to the at pattern). It follows that if the radius function of a planar SG pattern satis es c?1 < r(z + 1)=r(z) < c for some constant c , then the pattern is a Doyle spiral. Suppose that f : W ! S 2 is a locally injective meromorphic function, where W is simply connected. We show that on compact subsets of W , f can be approximated by an SG pattern, which is obtained by specifying the value of the invariant at the boundary according to Re Sf , the real part of the Schwarzian derivative of f , extending to be a solution of the SG -Dirichlet problem, nding an SG -Cauchy-Riemann co-solution that matches the value of Im Sf at some interior point, and taking the SG pattern with invariant and . This is a Mobius geometry analogue of a similar theorem by Carter and Rodin 6], with Re Sf taking the role of jf 0 j . The last section of the paper contains a collection of several open problems, but we brie y describe a few problems now. Suppose that f : C ! S 2 is a locally univalent open map. Then the metric on S 2 can be pulled back to C , and the resulting Riemannian surface is called the surface determined by f . Similarly, an entire SG pattern, or an entire immersed hexagonal packing, determines a Riemannian surface S , and the circle pattern is embedded on that surface. An important and fundamental question is: what Riemannian surfaces support entire circle patterns? It would also be interesting to understand to what degree the answer depends on the underlying combinatorics. One might conjecture that all the schemes that are based on doubly periodic combinatorics should behave in the same manner, but in our work below there is a hint suggesting that this might not be the case. (Some assumption, like double periodicity, is clearly necessary. Consider, for example, a circle pattern whose combinatorics is described by the hexagonal grid modulo a translation taking the grid to itself. It is not hard to see that such a pattern cannot lie on the universal cover of the once punctured plane. In other words, in that setting Doyle spirals are not present.) The author would like to thank Peter Doyle, Tomasz Dubejko, Alex Eremenko,

ODED SCHRAMM

Zheng-Xu He, Yakar Kannai, Rick Kenyon, Al Marden, and Burt Rodin for valuable discussions and advice. Let SG be the cell complex whose vertices are the Gaussian integers, V (SG) = Z+ iZ, whose edges are the pairs z; z 0 ] such that jz ? z 0 j = 1 and z; z 0 2 V (SG), and whose 2-cells are the squares z + x + iy : x; y 2 0; 1] , z 2 V (SG). When z; z0 are neighbors in SG , we write z z0 . The vertices z; z0 are called half-neighbors if they belong to the same square of SG , but are not neighbors. Let G be a subgraph of the 1-skeleton of SG . Suppose that for every z 2 V (G) (the vertices of G ) there is associated an oriented circle Cz in the Riemann sphere ^ S 2 = C . We now formalize what it means for the circles to be combinatorially like the pattern in Figure2.1. Consider the following three conditions. (1) Whenever z z1 are neighbors in G , the corresponding circles Cz ; Cz1 intersect orthogonally. (2) If z1; z2 are neighbors of a vertex z in G , and they belong to the same square of SG , then the circles Cz1 ; Cz2 are distinct and tangent. It follows from (1) and (2) that the tangency point of Cz1 and Cz2 lies in Cz . Hence there will be three special points on Cz , namely, Cz1 \ Cz2 ; Cz \ Cz1 ? Cz2 ; Cz \ Cz2 ? Cz1 . (3) Whenever the situation is as in (2) and z2 is the neighbor of z which is one step counterclockwise from z1 (that is, z2 = z + i(z2 ? z)), the circular order of the triplet of points Cz \ Cz1 ? Cz2 ; Cz1 \ Cz2 ; Cz \ Cz2 ? Cz1 agrees with the orientation of Cz .
2. Notations and Terminology

Figure 2.1. The regular SG pattern.

Observe that when the circles in Figure 2.1 are oriented positively with respect to the planar disks they bound, these three conditions are satis ed. De nitions. A circle pattern for G is an indexed collection C = ?Cz : z 2 V (G) of oriented circles in S 2 that satis es (1){(3) above. The pattern is planar, if each Cz C and is positively oriented with respect to the bounded component of C ? Cz . The pattern is embedded, provided that whenever z; z0 2 V (G) are

SG CIRCLE PATTERNS

not neighbors the open disks determined by Cz and Cz0 are disjoint. (The open ^ disk determined by an oriented circle C is the component D of C ? C such that @D = C with agreement of orientation.) If G is the whole 1-skeleton of SG , then a circle pattern for G is called an entire circle pattern. A vertex in G is called an interior vertex, if it has 4 neighbors in G , otherwise, it is a boundary vertex. The set of boundary vertices of G is denoted by V@ (G), and the set of interior vertices is denoted Vint(G). Let G be a subgraph of the 1-skeleton of SG . The next proposition allows the studying of a circle pattern through its radius function, the function which assigns to every vertex in G the radius of the corresponding circle. If C = (Cz : z 2 V (G)) is a planar circle pattern for a graph G SG , we let rC (z) denote the radius of the circle Cz . Sometimes, when no confusion is likely, we will use the notation r(z) in place of rC (z). 3.1. Proposition. Let Y be a union of squares of SG and let G be the graph which is the intersection of Y and the 1 -skeleton of SG . (1) Suppose that C = (Cz : z 2 V (G)) is a planar circle pattern for G , Then for every interior vertex z0 2 Vint(G) , we have ? r(z0 ) = H r(z0 + 1); r(z0 + i); r(z0 ? 1); r(z0 ? i) ; (3.1) where r = rC and
?1 ?1 + ?1 ?1 2 H (r1 ; r2 ; r3 ; r4 ) = (r1 + rr + rr3+ + r4 r)r1 r2 r3 r4 : r +
1 2 3 4

3. The Radius Function

(2) Conversely, if the interior of Y is connected, Y is simply connected, and r : V (G) ! (0; 1) satis es (3.1) for every z 2 Vint(G) , then there is a planar circle pattern for G such that radius(Cz ) = r(z) for all z 2 V (G) . This pattern is unique, up to isometries of C . Proof. Results of this kind are quite standard in the eld. We therefore only give a proof of the existence part of (2), which is the harder claim. Even the nonexpert should have no trouble in supplying the missing arguments after seeing the proof of (2). The following construction is essentially the same as the one in Thurston's notes 28, Chap. 13]. Consider two neighboring z0 ; z1 2 V (G). Let A(z0 ; z1 ) be the right angled triangle with vertices 0; r(z0 ); ir(z1 ). The vertex r(z0 ) repectively, ir(z1 )] will be called the z0 respectively, z1 ] vertex of A(z0 ; z1 ). The edge 0; r(z0 )] of A(z0 ; z1 ) will be called its z0 edge, and the edge 0; ir(z1 )], its z1 edge. Each edge of A(z0 ; z1 ) gets an orientation as part of @A(z0 ; z1). Note that the circle of radius r(z0 ) with center at the z0 vertex of A(z0 ; z1) is orthogonal to the circle of radius r(z1 ) with center at the z1 vertex of A(z0 ; z1 ). Let X be the disjoint union of the A(z0 ; z1) over all oriented edges z0; z1 ] in G . Consider the space X 0 obtained by pasting together edges of X as follows. The hypothenuse of A(z0 ; z1 ) is pasted to the hypothenuse of A(z1 ; z0 ), with orientations

ODED SCHRAMM

reversed. If z; z + ij ; z + ij+1 are all in V (G) ( j = 0; 1; 2 or 3), then the z edge of A(z; z + ij ) is pasted with orientation reversed to the z edge of A(z + ij+1 ; z). In this way every edge of X is pasted to at most one other edge, which has the same length, and X 0 is an oriented surface with boundary. We choose the gluing maps to be isometries. Hence X 0 has a metric. We now show that the metric on X 0 is locally Euclidean in the interior of X 0 ; that is, in X 0 ? @X 0 . Let p be a point in X , whose image p0 2 X 0 is not in @X 0 . Suppose rst that p is the vertex of A(z0 ; z1) where the right angle is. Since p0 2 @X 0 , it follows that four triangles are glued at p0 . Note that the angles glued = at p0 are all right angles. Hence X 0 is locally Euclidean near p0 . Now suppose that p is the z0 vertex of A(z0 ; z1 ). Since p0 is interior to X 0 , it follows that z0 is an interior vertex of G . Hence 8 triangles are pasted at p0 , namely A(z0 ; z0 + ij ); A(z0 + ij ; z0); j = 0; 1; 2; 3. The sum of the angles at p0 is X =2 arctan rr((zz)) : (3.2) 0 z z0 Hence,

ei =2 =

From (3.1), it now immediately follows that

z z0 r(z0 ) + ir(z )

Y r(z0 ) + ir(z) :
(3.3)

Im ei =2 = 0:

Because each of the eight angles pasted at p0 is in the range (0; =2), it follows that =2 2 (0; 2 ). Hence, (3.3) implies that = 2 . So X 0 is locally Euclidean near p0 . It is easy to see that X 0 is locally Euclidean near p0 if p is not a vertex of any of the triangles. So the interior of X 0 is locally Euclidean. There is some open nonempty set W X 0 that's isometric to a disk in the plane. Let f be such an isometry. Because X 0 is locally Euclidean, given any simple curve with initial point in W , f can be analytically continued along , and the result of this analytic continuation is a local isometry de ned in a neighborhood of . Since X 0 is simply connected, the Monodromy Theorem tells us that f has an analytic extension, de ned in the interior of X 0 . It follows that there is a local isometry F : X 0 ! C . For every z 2 V (G), we choose Cz to be the circle of radius r(z) centered at F (p0z ), where p0z is the image in X 0 of the z -vertices of the triangles A(z; z0 ), with ? z0 any neighbor of z . It is easy to check that C = Cz : z 2 V (G) is a circle pattern for G . The uniqueness part is left to the reader. 3.2. De nition. The surface X 0 in the proof will be called the surface determined by the circle pattern C . The map F : X 0 ! C is called the developing map. It is worth while to note, for future use, some properties of the function H from 3.1.

SG CIRCLE PATTERNS

3.3. Properties of H . In the following, r1 ; r2 ; r3 ; r4 > 0 are arbitrary.

(1) (2) (3) (4) (5)

Homogeneity: H (tr1 ; tr2 ; tr3 ; tr4) = tH (r1 ; r2 ; r3 ; r4 ) holds for all t > 0 . Symmetry: H is invariant under permutations of its arguments. ?? ? ? ? Reciprocity: H r1 1 ; r2 1 ; r3 1 ; r4 1 = H (r1 ; r2 ; r3 ; r4 )?1 . Monotonicity: H is strictly monotone increasing in each of its arguments. As p1 ! 1 and r3 ! 0 , while r2 ; r4 are kept xed, H (r1 ; r2 ; r3 ; r4 ) tends r to p r2 r4 , the geometric p of r2 and r4 . pr r , then mean = H (r ; r ; r ; r ) . (6) If r1 r3 = 2 4 r1 r3 1 2 3 4

The veri cation of these properties is left to the reader. They all have a geometric proof, as well as a direct proof from the de nition of H . Note that properties (2) and (3) do not hold for the corresponding function for circle packing immersions of the hexagonal lattice. In joint work with R. Kenyon we shall show that property (3) can be used to obtain explicit constructions for circle patterns corresponding to the polynomials zd .
4. Exp and Erf

P. Doyle has constructed circle packing immersions based on the hexagonal combinatorics that are analogous to the exponential functions eaz , later to be called Doyle spirals, and conjectured that these, and the regular hexagonal packings, are the only entire circle packing immersions with the combinatorics of the in nite hexagonal grid. See 5]. We now show that Doyle spirals are present in the setting of SG -circle patterns, and Doyle's conjecture is false in this setting. Suppose that z0 2 V (SG) and z1 ; z2; z3 ; z4 are its neighbors. Let r : V (SG) ! (0; 1). From 3.3.(6) it follows that for every a 2 C nonzero r(z) = jeaz j satises (3.1) at every z 2 V (SG). Consequently, r corresponds to an entire circle pattern. See Figure 4.1.

Figure 4.1. An SG Doyle spiral.

We shall call these patterns SG Doyle spirals, since they are clearly the analogues of the Doyle spirals for SG circle patterns. The surface determined by an SG Doyle spirals (or by a hexagonal Doyle spiral) is easily seen to be the universal cover of

8
C

ODED SCHRAMM

derivative of an exponential. Now consider the function r(z) = eaxy , where z = x + iy and a is a real constant. It is immediate that r(z) is equal to the geometric mean of r(z + 1) and r(z ? 1), and to the geometric mean of r(z + i) and r(z ? i). Consequently, by property 3.3.(6), r gives rise to an entire planar SG circle pattern. Note that p r(z) = cjf 0 (z)j for f (z)R= erf( iaz) and an appropriate c , where erf is the error function erf(z) = (2= ) e?z2 dz . Also note that the picture of the circle pattern is similar to the image of the rotated and scaled square grid under f , as indicated in p Figures 4.2.a, and 4.2.b. We shall call these circle patterns the iSG erf patterns.

? f0g , which is the same as C with the metric pulled back by the exponential map exp : C ! C ? f0g . This is one sense in which these spirals are analogues of the exponentials. Also, the function jeaz j is the same as the absolute value of the

Figure 4.2.a. The erf-image of a square grid, erf ?piSG .

Figure 4.2.b. The piSG erf pattern.

The Figures 4.2.a, and 4.2.b are not misleading, the erf circle pattern is indeed geometrically related to the pattern given by the error function. Before we make

SG CIRCLE PATTERNS

this statement precise, we need to recall a de nition. De nition. Let f : C ! X be holomorphic, where X is either the sphere or C . Then the metric on X can be pulled back via f to C . The resulting surface is called the surface determined by f : C ! X . 4.1. Theorem. The surface Sa determined by the piSG erf pattern with radius function r(z) = eaxy (a > 0) is isometric to the surface R determined by c erf : C ! C , where c is a constant which will be determined below. Proof. To simplify notations, only the case a = 1 will be discussed. Set S = S1 . ~ The circle in S corresponding to a vertex z 2 V (SG) will be denoted Cz . Let S denote the completion of S . We now show that S ? S contains precisely ~ 2 points. Let S (n) denote the part of S covered by circles Cx+iy with xy > ?n , and their interiors. Then S (n) is clearly complete. The complement S ? S (n) has two components: let A1 (n) be the component of S ? S (n) that contains circles ~ Cx+iy with y < 0 < x , and let A2 (n) be the component of S ? S (n) that contains ~ circles Cx+iy with x < 0 < y . It is easy to see that the diameters of A1 (n) and A2 (n) tend to zero as n ! 1 . So any sequence pn with pn 2 A1(n) is a Cauchy sequence. Consequently, there is a point q1 2 S such that pn ! q1 . It is immediate that q1 2 S , so q1 2 S ? S . Similarly, any sequence p0n with p0n 2 A2 (n) tends = to a point q2 2 S ? S . The points q1; q2 are distinct, because the distance from A1 (1) to A2 (1) is clearly positive. It follows that S = S fq1; q2 g . Let F : S ! C denote the developing map. Then F has a continuous extension 0 0 F : S ! C . Set qj0 = F (qj ), for j = 1; 2. We now determine the points q1 ; q2 . Note that the circles C0; C1; C?1 ; Ci; C?i all have radius 1. Since we are free to modify the developing map by an isometry of C , we assume withoutp p generalp loss of p ity that the centers the circles C0; C1 ; C?1; Ci; C?i are at 0; 2; ? 2; i 2; ?i 2, respectively. Because of the symmetry r(x + iy) = r(?y ? ix) of the radius function r(z) = xy , the circle Cx+iy will be the re ection of the circle C?y?ix in the diagonal x = ?y . Consequently, the centers of the circles Cn(1?i) will all lie on the diagonal x = ?y . Since Cn(1?i) is tangent to C(n+1)(1?i) it follows by induction that for n > 0 the center of Cn(1?i) is at the point 1 ? i 1 + 2e?1 + j1 ? ij Set + 2e?(n?1)2 + e?n2 :
1 X
n=1

c=1+2

e?n2 :

0 0 Then it follows that q1 = c(1 ? i)=j1 ? ij and similarly q2 = c(i ? 1)=ji ? 1j . Suppose that p0 is any point on S and : 0; 1] ! C is a path in C starting at 0 0 (0) = F (p0 ). Assume that does not pass through the points q1 and q2 . Then has a lift ~ : 0; 1] ! S with ~(0) = p0 . (That ~ is a lift of means = F ~ .) Indeed, let Z be the set of t 2 0; 1] so that the restriction of to 0; t] has a lift. Because F is a local isometry, the set Z is open in 0; 1], and when has a lift in 0; t] it is unique. The set Z is not empty, since 0 2 Z . Set t1 = sup Z . Then the restriction of to 0; t1 ) has a lift : 0; t1) ! S . Because S is complete, it follows

10

ODED SCHRAMM

that has a continuous extension : 0; t1 ] ! S . Therefore, F (t1 ) = (t1). 0 0 Since avoids q1 ; q2 , we have (t1) 6= q1 ; q2 , and is a path in S . So has a lift to 0; t1] and sup Z = t1 2 Z . Because Z is open in 0; 1], it follows that Z = 0; 1]. ~ Now let p0 be the center of C0 in S , and let l be the lift of the line x = y passing through p0 . Then S ? l has two components. Let B1 be the component of S ? l whose closure in S contains q1 , and let B2 be the component of S ? l whose closure in S contains q2 . 0 = 0 = 4.2. Lemma. q1 2 F (B1) and q2 2 F (B2 ) . 0 Proof. Suppose, for example, that q1 2 F (B1). Let p be some point in B1 with 0 F (p) = q1 . 0 Since F is a local isometry, a small disk B(q1 ; ) can be lifted to a small disk 0 B(p; ) in B1 . Every line segment of the form q1 ; q] contained in the half plane x > y will then have a lift to a segment in B1 starting at p . By taking the union of all these lifts, it follows that the identity map from the half plane x > y onto itself has a lift g : fz : x > yg ! B1 . The boundary of the image of g is a lift of the line x = y . Suppose rst that this lift is not l . Then g can be extended beyond the line x = y ; that is, there is a lift of the whole plane C to B1 whose restriction to x > y is g . This means that B1 contains a subset isometric to C , which implies that B1 is isometric to ? C , a contradiction. On the other hand, if the boundary of g fx > y g is l , then it follows that the image of g is all of B1 , which is false, because B1 l is not complete. Now let erf (z) = c(1 ? i) erf(z)=j1 ? ij . Then R is isometric to the surface R determined by erf . Using limz!+1 erf(z) = 1, a discussion entirely analogous to the above shows that the completion R of R consists of R with two additional 0 0 points, one, say q1 , over q1 and one, say q2 , over q2 . Let l be the lift to R that passes through 0 2 R of the line x = y . (Recall that R = C with the metric pulled back via erf .) Let B1 be the component of R ? l whose closure in R contains q1 , and let B2 be the component of R ? l whose closure in R 0 = contains q2 . As in the proof of the lemma, it is easy to see that q1 2 erf (B1 ) and 0 2 erf (B2 ). q2 = Let p be any point in B1 l . Take any path that connects p and 0 in 0 B1 l . Set = erf . We know that does not pass through q1 . Hence has a (unique) lift ~ to B1 l that intersects l . Let f1 (p) be the endpoint of ~ corresponding to p . Since B1 l is simply connected, f1(p) does not depend on the choice of . It is clear that f1 : B1 l ! B1 l is an isometry. Similarly, we de ne the isometry f2 : B2 l ! B2 l . These two isometries can be glued along l , to give an isometry from R to S . This proves the theorem. p One can slightly generalize the iSG erf pattern, by taking r(z) = eaxy+bx+cy with real a; b; c . When b; c are integer multiples of a , this corresponds to pretranslating the square grid. When they are not, you get an essentially new pattern, which is very similar to the pattern for the case b = c = 0. Altogether, up to p similarity, this gives a 3-dimensional family of iSG erf patterns.

SG CIRCLE PATTERNS

11

It is natural to ask for pre-rotated versions of these patterns. To be precise, these patterns are not analogues of the true erf function, but of the functions ?p erf ai(z + c) , with a a real constant, and c a complex constant. One may ask for a circle pattern which would mimick erf(z), without pre-rotation. Here is another view of the situation. The function H can be thought of as a nonlinear version of the discrete Laplacian for log r(z). The linear discrete Laplacian, ? g(z) = (1=4) g(z + 1) + g(z + i) + g(z ? 1) + g(z ? i) ? g(z). behaves in many respects like the Laplacian in R2 . In particular, the space of discrete harmonic ? functions on the square grid that have O jz2 j growth is spanned by the functions 1; x; y; xy; x2 ?y2 . We have seen that any linear combination of 1; x; y;2xy 2is log r(z) for some planar circle pattern. Direct inspection shows that r(z) = ex ?y does not satisfy equation (3.1). The question arises if there is some radius function similar to ex2?y2 which does. The corresponding pattern may be an erf analog, without pre-rotation. It will follow from the following theorem that, in some well de ned sense, such a pattern does not exist. 4.3. Theorem. Let G be the intersection of SG with the quadrant fx + iy : x > jyjg , and let r : V (G) ! (0; 1) be a function that satis es (3.1) at every interior vertex of G . Then

r(z) 6 C maxfr(v) : v 2 V@ (G); Re v 6 xg e3x log x;

holds for every z 2 V (G) . Here x = Re z , and C is some absolute constant. Proof. Suppose that R : V (G) ! (0; 1) is a function that satis es the the inequality ? R(z) > H 1; R(z + i); R(z ? 1); R(z ? i) ; (4.1) for every interior vertex z 2 Vint(G). (Here, the expression H (1; r2 ; r3 ; r4 ) stands for the limit of H (r1 ; r2 ; r3 ; r4 ) as r1 ! 1 .) Let n be some positive integer. Set Vn = fz 2 V (G) : Re z 6 ng , and let M be the maximum of r(z)=R(z) for z 2 Vn . Let z 2 Vn \ Vint(G). Then we have from the properties of H ,

r(z) = H r(z + 1); r(z + i); r(z ? 1); r(z ? i) ? = MH r(z + 1)=M; r(z + i)=M; r(z ? 1)=M; r(z ? i)=M ? 6 MH r(z + 1)=M; R(z + i); R(z ? 1); R(z ? i) ? < MH 1; R(z + i); R(z ? 1); R(z ? i) 6 MR(z):
Therefore, the maximum of r(z)=R(z) in Vn is attained in Vn \ @V (G). We now choose the function R , and prove that it satis es (4.1). Set

a(x) = log(x + 3);

and de ne b(n) inductively by b(0) = 0 and

b(n + 1) = b(n) + a(n);

12

ODED SCHRAMM

for positive integers n . For negative n , let b(n) = b jnj . Now set

R(z) = e3b(x)?b(y);
for every z = x + iy 2 V (G). We will show that R satis es (4.1). Note that (4.1) is the same as

R(x + iy)2 > ? ? ? ? R x ? 1 + iy R x + i(y ? 1) + R x ? 1 + iy R x + i(y + 1) + ? ? + R x + i(y + 1) R x + i(y ? 1) :

For our particular R , this evaluates to

e6b(x)?2b(y) > e3b(x?1)?b(y)+3b(x)?b(y?1) + e3b(x?1)?b(y)+3b(x)?b(y+1) + e6b(x)?b(y+1)?b(y?1): (4.2)

It is clearly su cient to prove this for 0 6 y 6 x ? 1, since both sides do not change if we replace y by ?y . Assume rst that y 2 1; x ? 1]. Then (4.2) reduces to 1 > e?3a(x?1)+a(y?1) + e?3a(x?1)?a(y) + ea(y?1)?a(y): We estimate the right hand side, as follows,

e?3a(x?1)+a(y?1) + e?3a(x?1)?a(y) + ea(y?1)?a(y) = (x + 2)?3 (y + 2) + (x + 2)?3 (y + 3)?1 + (y + 2)(y + 3)?1 1 6 (y + 2)?2 + (y + 3)?4 + 1 ? y + 3 6 1:
So (4.2) holds for y 2 1; x ? 1]. For y = 0, (4.2) reduces to

e6b(x) > 2e3b(x?1)e3b(x)?b(1) + e3b(x)?b(1)e3b(x)?b(1);

which is the same as 1 > 2e?3a(x?1)?a(0) + e?2a(0): The right hand side is easily estimated, as follows, 2e?3a(x?1)?a(0) + e?2a(0) = 2(x + 2)?3 =3 + 1=9 6 1: Hence (4.1) holds throughout Vint(G). Now note that
0

?jnj 6 Z jnj log(t + 3) dt = (jnj + 3) log ?jnj + 3 ? jnj ? 3 log 3: b(n) = b

SG CIRCLE PATTERNS

13

Therefore, it is easy to conclude that

R(z) 6 Ce3x log x; R(z) > 1:

(4.3) (4.4)

holds in V (G), for an appropriate constant C < 1 . On the other hand, we have Fix any z 2 V (G), and take n = x = Re z . We have seen that the maximum of r(v)=R(v) on Vn is achieved on Vn \ V@ (G). Hence max r r(z) 6 R(z) minffR(w) :: w 2 Vn \ V@ (G)g : (w) w 2 V \ V (G)g
n @

The theorem now follows from (4.3) and (4.4). Because erf 0 (z) = (2= )e?z2 , we have erf 0 (z) = 2= on the diagonals j Re zj = j Im zj . At a bounded distance from the diagonals we have erf 0 (z) 6 ec1 jzj , for an appropriate constant c1 . So it is reasonable to require an SG pattern analogue of erf (without pre-rotation) to have a radius function r satisfying r(z) 6 ec2jzj on the diagonals. By Theorem 4.3, r would then satisfy r(z) 6 Cec2jzj+3jzj log jzj . Hence r would deviate from the behaviour of j erf 0 j on the imaginary axis. It follows that there is no entire SG pattern that mimicks erf in the sense of the growth rate of r(z) along the axis and diagonals. Similarly, Theorem 4.3 is applicable to many other functions in place of erf , R for example, ew4 dw . It is also applicable to non-locally injective functions and branched circle patterns (which we have not de ned here), e.g., ez2 . Notwithstanding the above, we cannot conclude that there is no entire SG pattern analogue of erf ; it is conceivable that there exists some SG pattern possessing many of the geometric properties of erf . It seems that entire SG circle patterns cannot be expected to parallel many entire locally injective functions. The non-isotropy of the square grid is the underlying reason. There is every reason to expect that the circle packings based on the hexagonal grid would do no better. Perhaps, circle patterns based on more isotropic graphs (e.g., graphs with some randomness) might be better. We end this section with the moral that computer simulations cannot be trusted when searching for entire circle patterns. For example, it is easy to create pictures suggesting the existence of a true SG erf pattern (without pre-rotation). See Figure4.3. The radius function is a handy tool for studying SG circle patterns in the plane. However, it is not useful for studying SG patterns on the sphere. We now present Mobius invariants for?SG pattern in the sphere. Suppose that C = Cz : z 2 V (SG) is a circle pattern for SG on the Riemann ^ sphere C . The set of centers of squares of SG will be denoted by V (SG); that is, V (SG) = fz + 1=2 + i=2 : z 2 V (SG)g . Let w 2 V (SG). Note that the
5. The Mobius Invariants

14

ODED SCHRAMM

Figure 4.3. A fake SG erf pattern?

intersection of the four circles Cw 1=2 i=2 is a single point. We denote this point by pC (w), or sometimes just p(w), when there can be no confusion about the implied C . Now let z 2 V (SG). The four points p(z 1=2 i=2) are all on the circle Cz . Hence their cross ratio is real. Recall that the cross ratio of four points ^ p1; p2 ; p3 ; p4 2 C is de ned by (p ? p )(p ? p ) cr p1; p2 ; p3 ; p4 = (p1 ? p3)(p2 ? p4 ) ; 1 4 2 3 and does not change when a Mobius transformation is applied to all four points. Let the invariant of C be the function : V (SG) ! R de ned by
C (z ) = (z ) = ? cr p(z +1=2+i=2); p(z ?1=2?i=2); p(z +1=2?i=2); p(z ?1=2+i=2) :

(The reason for the awkward order is that it makes formulas below a bit simpler.) The invariant does not change if we modify C by a Mobius transformation. Because fp(z + 1=2 + i=2); p(z ? 1=2 ? i=2)g separates p(z + 1=2 ? i=2) and p(z ? 1=2 + i=2) on Cz , it follows that (z) > 0 for z 2 V (G). It turns out that the invariant is not su cient to determine C up to Mobius transformations. Another invariant is necessary. For each w 2 V (SG) we de ne
C (w) = (w) = ? cr p(w + 1); p(w ? 1); p(w ? i); p(w + i) :

This is the invariant for C . We now show that (w) is real and positive. Let m be a Mobius transformation ? that take p(w) to 1 . The four circles m Cw 1=2 i=2 are lines, and together form a rectangle with corners p (w + 1); p (w + i); p (w ? 1); p (w ? i), where p = pm(C) . See Figure 5.1. Hence p (w + 1) ? p (w + i) = p (w ? i) ? p (w ? 1),

SG CIRCLE PATTERNS

p (w ? 1) ? p (w + i) = p (w ? i) ? p (w + 1), and the ratio p (w + 1) ? p (w + ? i) = p (w + 1) ? p (w ? i) is pure imaginary. Therefore, p (w + 1) ? p (w ? i) 2 = p (w + 1) ? p (w ? i) 2 : C (w) = m(C ) (w) = ? p (w + 1) ? p (w + i) p (w + 1) ? p (w + i) 2 (5.1)
p(w+i) p(w+1) p*(w-i) p*(w+1)

15

p(w) p(w-1) p(w-i) p*(w-1) p*(w+i)

Figure 5.1. The con guration determining (w).

Just as the radius function for a circle pattern is a discrete analogue to the absolute value of the derivative of an analytic function, we shall see that log and log are analogous to the real and imaginary parts of the Schwarzian derivative, respectively. The Schwarzian derivative of a function f is de ned by f 00 0 ? 1 f 00 2 : Sf = f 0 (5.2) 2 f0 It is left-Mobius invariant; that is, Sm f = Sf for any Mobius transformation m . If f; g are meromorphic locally injective functions de ned on a connected open set, and Sf = Sg , then f = m g for some Mobius transformation m . See 19, xII.1] for a brief discussion of the Schwarzian derivative. The analogies between log + i log and the Schwarzian derivative will be pointed out from time to time, as we progress. For starters, note that when f (z) = g(iz) we have Sf (z) = ?Sg (iz), and similarly ?C (z) = C (iz)?1 ; C (z) = ? C (iz )?1 , when C = Cz : z 2 V (SG) and C = Cz : z 2 V (SG) satisfy Cz = Ciz . Examples. Before we proceed further with the general theory, let us calculate the invariants for the patterns described previously. To accomplish this, we rst relate the Mobius invariants to the radius function. Suppose that C is a?planar SG pattern. Take w 2 V (SG). The Mobius transformation m(z) = z ? p(w) ?1 takes p(w) to 1 . If C is a circle of radius r passing through p(w), then m(C ) is a line whose distance from 0 is (2r)?1 . Hence (5.1) shows that r(w + 1=2 + i=2)?1 + r(w ? 1=2 ? i=2)?1 2 (w) = r(w + 1=2 ? i=2)?1 + r(w ? 1=2 + i=2)?1 :

16

ODED SCHRAMM

To calculate (z), let c denote the center of Cz , set qj = p z + ij (1 + i)=2 ? c , and observe that ? r(z) + ir z + ij qj = qj?1 r(z) ? ir (z + ij ) : Hence, (q ? q )(q ? q ) (z) = ? cr q0 ; q2 ; q3; q1 = ? (q0 ? q3)(q2 ? q1) 0 1 2 3 r(z)+ir( r(z)+ir( q3 r(z)?ir(z+1) ? 1 q1 r(z)?ir(z?1) ? 1 z+1) z?1) =? ?1 ?1 ir(z+ ) r(z)+ir( q1 r(z)+ir(z+ii) ? 1 q3 r(z)?ir(z?ii) ? 1 r(z)? z? ) =?
r(z)+ir(z+1) ? 1 r(z)?ir(z+1) r(z)?ir(z+i) ? 1 r(z)+ir(z+i) r(z)+ir(z?1) r(z)?ir(z?1) ? 1 r(z)?ir(z?i) r(z)+ir(z?i) ? 1

r(z) + i r(z) ? i ?1 r(z) ? i ?1 : r(z) + i = ? r(z + i) r(z ? i) r(z + 1) r(z ? 1) Because is positive, taking absolute values in both sides gives

s? 2 ? r(z) + r(z + i)2 ?r(z)2 + r(z ? i)2 : (z) = ? 2 r(z) + r(z + 1)2 r(z)2 + r(z ? 1)2

(5.3)

We are now ready to compute and for the known examples of planar SG patterns. For the at SG pattern, where all circles have the same radius, (and its Mobius images), we have = 1; = 1. Take some a 2 C , and let C (a) denote the SG Doyle spiral with radius function r(z) = jeaz j = ea1 Re z?a2 Im z , where a1 = Re a , a2 = Im a . Using the above formula for and , we get ea2 + e?a2 C (a)(z ) = ea1 + e?a1 ; a2 a1 2 (w) = ee 2++ e 1 : C (a) a a1 + Let expa(z) = eaz . Then, using the de nition (5.2) of the Schwarzian derivative, Sexpa = ?a2=2 = a2=2 ? a2=2 ? ia1a2 . Up to Mobius transformations, the functions 2 1 expa are the only functions with constant Schwarzian derivative. Similarly, it will follow from results below that the SG Doyle spirals are the only SG patterns with constant and . Moreover, note the following analogies between Sexpa and log C(a) + i log C(a) . Both do not change when a is replaced by ?a , Re Sexpa and log C(a) are both monotone increasing functions of ja2j and ?ja1j , and vanish when ja2 j = ja1 j , Im Sexpa and log C(a) both vanish when a1 a2 = 0, and do not change when a1 and a2 are exchanged. It is also interesting to note that Sexpa = log C(a) + i log C(a) + O(a4 );

SG CIRCLE PATTERNS

17

near a = 0. p For the iSG erf pattern with radius function r(z) = exy (z = x + iy), we have
x + e?x (z) = ey + e?y ; e x+y (z) = 1 eex ++y1 e e
2

(5.4)

If f is meromorphic, the real and imaginary parts of Sf satisfy together the Cauchy-Riemann equations. So if log + i log is indeed analogous to Sf , we would expect and to be related in a similar way. As will be seen below, this is indeed the case. In the following, we shall determine the necessary and su cient conditions for a pair of functions : V (SG) ! (0; 1), : V (SG) ! (0; 1) to be the invariants of a circle pattern for SG . (Theorem 5.1.) As usual, it is convenient to start with `necessity'. Let C be some circle pattern for SG . We shall nd equations relating its in1 variants and . Let HG denote 2 SG + 1+i ; that is, HG is a square grid 4 1 with edge-size 2 , translated so that 0 is at a center of a square of HG . Take v 2 V (HG), and let z 2 V (SG) be the unique vertex of SG that's closest to v . Set w1 = z + 2(v ? z); w2 = z + 2i(v ? z); w3 = z ? 2i(v ? z), and note that w1; w2; w3 2 V (SG). Let mv be the Mobius transformation that takes 1; 0; 1 to p(w1); p(w2 ); p(w3 ), respectively. For any directed edge v1; v2 ] of HG , set

m v1;v2] = m?11 mv2 : (5.5) v Note that m v1 ;v2] does not change if we apply a Mobius transformation to C ; it is a Mobius invariant of C . We shall now compute the Mobius transformations m v1;v2 ] for the various types of edges v1; v2 ], in terms of the invariants and . 1 First take v1 to be of the form w ? 4 (1 + i), where w 2 V (SG), and v2 = w + i(v1 ? w) = v1 + 1 . When calculating m v1;v2] , we shall assume that mv1 is 2 the identity. This involves no loss of generality, because C can be replaced by its Mobius image m?11(C ). With this assumption, we have v p(w) = 1; p(w ? 1) = 0; p(w ? i) = 1: Because p(w) = 1 , the four points p(w 1); p(w i) form the vertices of a rectangle, and p(w + 1) ? p(w ? i) 2 : (w) = ? p(w ? i) ? p(w ? 1)
This gives

p(w + 1) = p(w ? i) ? i

(w) p(w ? i) ? p(w ? 1) = 1 ? i

(The above correct choice of sign for i = ?1 follows from an inspection of the order of the corners p(w 1); p(w i) of the rectangle.) Now, m v1 ;v2] takes 1; 0; 1 to p(w); p(w ? i); p(w + 1), respectively. Hence m v1;v2] is easily computed:

(w):

m v1;v2] ( ) = 1 ? i

(w) :

18

ODED SCHRAMM
1 4i

Similar computations yield expressions for m v1 ;v2] when v1 = w 1 4 v2 = w + i(v1 ? w), where w 2 V (SG). We record the results as follows
1 1 1 1 m w? 4 ? 1 i;w+ 4 ? 1 i] ( ) = m w+ 4 + 1 i;w? 4 + 1 i] ( ) =1 ? i (w) : 4 4 4 4 p 1 1 1 1 1 m w+ 4 ? 4 i;w+ 4 + 4 i] ( ) = m w? 1 + 1 i;w? 4 ? 1 i] ( ) =1 ? i (w)?1 : 4 4 4

and (5.6)

These formula are valid for all w 2 V (SG). 1 1 1 Now consider an edge v1; v2 ] in HG , where v1 = z + 4 + 4 i; v2 = z ? 4 + 1 i , 4 and z 2 V (SG). We assume with no loss of generality that

p(z + 1=2 + i=2) = 1; p(z ? 1=2 + i=2) = 0; p(z + 1=2 ? i=2) = 1: From the de nition of (z), we get, 1 p(z ? 1=2 ? i=2) = (z)?1 + 1 : The transformation m v1 ;v2] takes 1; 0; 1 to p(z ? 1=2+ i=2); p(z ? 1=2 ? i=2); p(z + 1=2 + i=2), respectively. Hence, m v1;v2]

? (z)?1 + 1 ?1 : ( )=
1?

Similar computations yield expressions for m v1 ;v2] when v1 = z v2 = z + i(v1 ? z). The results are,
1 1 m z? 4 ? 1 i;z+ 4 ? 1 i] ( ) = m z+ 1 + 1 i;z? 1 + 1 i] 4 4 4 4 4 4

1 4

1 4i

and

1? (5.7) ( (z) + 1)?1 ; 1 1 1 1 m z+ 4 ? 1 i;z+ 1 + 4 i]( ) = m z? 4 + 1 i;z? 4 ? 1 i]( ) = 1 ? 4 4 4 4 and are valid for all z 2 V (SG). From the de nition of the Mobius transformations m v1 ;v2] , it is clear that for any closed path v0; v1 ; : : : ; vn = v0 in V (HG) the composition of the Mobius transformations corresponding the edges of the path is the identity; that is,

? (z)?1 + 1 ?1 ; ( )=

m v0;v1] m v1;v2 ]
In particular,

m vn?1;vn] = identity:

(5.8)

m v1 ;v2] ?1 = m v2;v1] ; (5.9) for every directed edge v1 ; v2] in HG . Hence the equations (5.6) and (5.7) allow us to compute m v1;v2] for every directed edge v1; v2 ] of HG . Let a be of the form a = n + (m + 1=2)i , where n; m 2 Z, and let vj = a + ij (1 + i)=4, for j = 0; 1; 2; : : : . Then v0 ; v1 ; v2; v3 ; v4 = v0 is a closed path in HG . Consequently, we get m v4;v3] m v3;v2 ] m v2 ;v1] m v1;v0] = identity: (5.10)

SG CIRCLE PATTERNS

19

After using the appropriate formulas from (5.6) and (5.7), this becomes

0 p 1 ? i (a + 1=2)?1 @
(a + 1=2) = (a ? 1=2)

1? 1?i

p (a ? 1=2)?1
2

? (a ? i=2)?1 + 1 ?1

( (a+i=2)?1 +1)?1 1?

1 A= ;

and simpli cation yields, (a + i=2)?1 + 1 (a ? i=2)?1 + 1

for a 2 V (SG) + i=2:

(5.11)

Choosing vj = b + ij (1 + i)=4, where b has the form b = m + ni + 1=2, gives, after entirely similar computations, (b + i=2) = (b ? i=2) (b + 1=2) + 1 (b ? 1=2) + 1
2

for b 2 V (SG) + 1=2:

(5.12)

Note that the equations (5.11) and (5.12) are nonlinear discrete versions of the Cauchy-Riemann equations. We shall call them the SG -CR equations. We have established the rst part of the following

5.1. Theorem.

(1) Let C be a circle pattern for SG in the sphere, then its and invariants satisfy the SG -CR equations. (2) Conversely, suppose that and are positive functions on V (SG) and V (SG) , respectively, and suppose they satisfy the SG -CR equations. Then there is a circle pattern C for SG such that = C and = C . (3) Suppose that C and C are circle patterns for SG and C = C , C = C . Then C is a Mobius image of C .

Proof. Part (1) has been proven above. So consider part (2). Use equations (5.6), (5.7) and (5.9) to de ne Mobius transformations m v1;v2 ] for the edges v1 ; v2] of HG . We now want to verify that (5.8) holds for every closed path v0 ; v1; : : : ; vn in HG . Since (5.9) holds, it is enough to demonstrate this for paths that form the boundary of a single square tile of HG . If the path is the boundary of a tile of HG whose center a is in V (SG) + i=2, then (5.8) is equivalent to (5.11), as we have seen. Similarly, if the path is the boundary of a tile of HG whose center b is in V (SG) + 1=2, then (5.8) is equivalent to (5.12). Consider the situation where vj = w + ij (1 + i)=4, j = 0; 1; 2; 3; 4, where w 2 V (SG). Then (5.8) becomes

1?i

(w) 1 ? i

(w)?1 1 ? i

(w) 1 ? i

(w)?1

= ;

which indeed holds. Similarly, if vj = z + ij (1 + i)=4, j = 0; 1; 2; 3; 4, where z 2 V (SG), then (5.8) is easily veri ed. (In fact, in these two cases one can also argue without any computations.)

20

ODED SCHRAMM

The cases we have checked are su cient to guarantee (5.8) for any closed path v0; v1 ; : : : ; vn . Now take an arbitrary vertex v0 in HG , and set mv0 = identity. For every other v 2 V (HG) set mv = m v0;v1 ] m v1 ;v2] m vn?1;vn] ; where v0; v1 ; : : : ; vn is an arbitrary path in HG from v0 to v = vn . Because (5.8) is valid for closed paths in HG , the transformation mv does not depend on the choice of path from v0 to v . For any w 2 V (SG), de ne p(w) = mv (1), where v is one of the vertices of HG closest to w . It does not matter which of these v is chosen, because the transformations in (5.6) preserve 1 . For any z 2 V (SG), let Cz be the image of R f1g under mv , where v is one of the vertices of HG closest to z . It does not matter which of these v is chosen, because the transformations in (5.7) preserve R f1g . The orientation of the circle Cz is taken as the image of the orientation of R f1g . It is immediate to verify that for any z 2 V (SG) the points p(z 1=2 i=2)? appear on Cz in the correct cyclic order. One easily shows that the collection C = Cz : z 2 V (SG) is a circle pattern for SG , which has invariants and . The details of this, as well as part (3), are left to the reader. An interesting and useful property of the SG -CR equations is that the invariant can be eliminated from them. More precisely, we have the following, (1) Suppose that : V (SG) ! (0; 1) and : V (SG) ! (0; 1) satisfy the SG -CR equations. Then satis es the equation z 1) + 1) ( (z ? 1) 1) (z)2 = ( (z (+ +?1 + 1) ( (z ? i)?+ + 1) for every z 2 V (SG): (5.13) 1 ( i) (2) Conversely, suppose that : V (SG) ! (0; 1) satis es (5.13). Then there is a : V (SG) ! (0; 1) such that and together satisfy the SG CR equations. Moveover, is unique, up to multiplication by a positive constant. This is similar to the situation with the Cauchy-Riemann equations @y v = @x u; @xv = ?@y u , where v may be eliminated, and the result is the Laplace equation for u . Given a solution u of the laplace equation, a companion solution to the Cauchy-Riemann equation exists, and is unique up to an additive scalar. So the SG situations seems analogous. In fact, equation (5.13) is a nonlinear discrete version of the Laplace equation. Hence, we call it the SG -Laplace equation. Unfortunately, it seems there is no way to eliminate from the SG -CR equations and get an equation involving only . Herep an application of the theorem. Recall the formula (5.4) for the invariants is of the iSG erf pattern. From the theorem it follows that if we multiply the invariant by a positive constant b , then the pair ; b still corresponds to an SG immersion on the sphere. In this way, we get a new 2-parameter family of Mobius inequivalent entire SG patterns in the sphere, with invariants given by, 2 x+y x + a?x (5.14) (z) = b aax ++y1 ; (z) = ay + a?y ; a a

5.2. Theorem.

SG CIRCLE PATTERNS

21

where a; b are arbitrary positive numbers. (In fact, this family embeds in a 4parameter family, since z may be translated by an arbitrary complex constant.) Let us call these circle patterns erf-like patterns, for lack of a better name. To guess what meromorphic functions the erf-like patterns might correspond to, p the Schwarzian derivative should be examined. Recall that our iSG erf pattern is R analogous to the function f = e?iz2 . The Schwarzian derivative of this function is 00 0 00 2 Sf = f 0 ? 1 f 0 = ?2i + 2z2: f 2 f

It is then reasonable to guess that the erf-like patterns would correspond to the functions that have Schwarzian derivative S = z2 + i , where > 0 and 2 R is arbitrary. Figures 5.2.a and 5.2.b show the erf-like pattern with a = 1:152 and b = 1 and the corresponding meromorphic function with Schwarzian equal to z2 (which is a ratio of two Bessel functions).

Figure 5.2.a.

Figure 5.2.b.

One striking feature of these gures is the =2 rotational symmetry. This symmetry can be read o from the Schwarzian. Indeed, let f be some meromorphic function, and set f^(z) = f (!z), where j!j = 1. Then Sf^(z) = !2Sf (!z). Consequently, f^ = m f for some Mobius transformation m i !2Sf (!z) = Sf (z). This holds when ! = i and Sf (z) = z2 . Similarly, the symmetry of the erf-like pattern with b = 1 can be read from the relations (iz) = (z)?1 and (iz) = (z)?1 , which are easy to verify. Proof of 5.2. We start with (1). For any z 2 V (SG), (z + 1=2 + i=2) (z + 1=2 ? i=2) (z ? 1=2 ? i=2) (z ? 1=2 + i=2) 1 = (z + 1=2 ? i=2) (z ? 1=2 ? i=2) (z ? 1=2 + i=2) (z + 1=2 + i=2) : Taking square roots, and using the SG -CR equations gives z ?1 1 z ?1 1 (5.15) 1 = (z(+)1) + 1 (z(?) i)?++ 1 (z(?)1) + 1 (z(+) i)?++ 1 ; 1 1 z +1 z +1

22

ODED SCHRAMM

which simpli es to (5.13). The proof of (2) is straightforward. One chooses (1=2 + i=2) arbitrarily. If v; v0 2 V (SG) are neighbors and (v) is de ned, then one of the SG -CR equations de nes (v0 ). So given any v 2 V (SG), we de ne (v) by walking along a path from 1=2 + i=2 to v and using either of these equations for every edge. As the computation of the previous paragraph shows, equation (5.13) shows that the choice of path does not in uence the value of (v). This then implies that the SG -CR equations are valid everywhere. The remaining details are left to the reader. For convenience, we introduce the function
1 e ? 1 H ( 1 ; 2; 3; 4) = ? (?1 + 1) ( 3?+ 1) : 2 +1 4 +1 Using this notation, equation equation (5.13) can be written as e? (z) = H (z + 1); (z + i); (z ? 1); (z ? i) :

(5.16)

e 5.3. Properties of H . In the following, 1; 2; 3 ; 4 > 0 are arbitrary. ? ; ; ; . e (1) 1 = H 1 1 1 1 e (2) Monotonicity: H is strictly monotone increasing in each of its arguments. e e e (3) Partial symmetry: H ( 1 ; 2; 3;? 4 ) = H ( 1; 4 ; 3; 2) = H ( 3 ; 2; 1; 4 ) . ?1 . e e ? ? ? (4) Rotation: H ( 4 ; 1; 2; 3) = H 1 1; 2 1; 3 1; 4?1 p 1 3. e( (5) H ? 1 ; 1; 3; 3 ) = ? e ? (6) H 1; 1 1; 3 ; 3 1 = 1 . e e (7) lim !0 H ( 1; 2; 3 ; 4) > 0 and lim !1 H ( 1; 2 ; 3; 4) < 1 .
1 2

The proof is left to the reader. Property (4) is related to the fact that when ? ? two circle patterns C = Cz : z 2 V (SG) and C = Cz : z 2 V (SG) satisfy Cz = Ciz , then C (z) = C (iz)?1 . One important consequence of (1) and (2) is the validity of the following maximum principle. Suppose that G is a nite subgraph of SG and : V (G) ! (0; 1) satis es (5.13) at every interior vertex of G . Then attains its maximum and its minimum on V@ (G).
6. The Local Theory

The goal of this section is to understand the collection of all circle patterns for a nite `simply connected' subgraphs of SG . First, a `local' version of Theorem 5.1 will be presented. 6.1. Theorem. Let S be a non-empty union of squares of SG so that the interior of S is connected, and the closure simply connected. Let G be the intersection of the 1 -skeleton of SG with S , let VS denote the vertices of G that have at least three neighbors in G , and let V (S ) be the centers of the squares of S . (1) Let C be a circle pattern for G in the sphere. Then its invariants : VS ! (0; 1) and : V (S ) ! (0; 1) satisfy the SG -CR equations across every edge of SG that borders two squares of S .

SG CIRCLE PATTERNS

23

(2) Conversely, suppose that : VS ! (0; 1) and : V (S ) ! (0; 1) satisfy the SG -CR equations across every edge of SG that borders two squares of S , then there is a circle pattern for G in the sphere, which is unique, up to Mobius transformations. The proof is left to the reader. Similarly, there is also a local version of Theorem 5.2, whose formulation we also omit. Using these results, the study of circle patterns for nite graphs G which are as in Theorem 6.1 reduces to the study of solutions of equation (5.13). These are now described. 6.2. Theorem ( SG Dirichlet Principle). Let G be a nite subgraph of the 1 -skeleton of SG , and let : V@ (G) ! (0; 1) be some positive function on the boundary vertices of G . Then can be extended to V (G) in such a way that equation (5.13) holds at any interior vertex z 2 Vint(G) . Moreover, the extension is unique. Proof. A function : V (G) ! (0; 1) that satis es the inequality

e (z) 6 H (z + 1); (z + i); (z ? 1); (z ? i) ;

(6.1)

e at every z 2 Vint (G) will be called H -subharmonic in G . Let a be the minimum of on V@ (G), and let 1 be equal to on V@ (G) and equal to a in Vint (G). e e Then 1 is clearly H -subharmonic. Let be the supremum of all H -subharmonic functions on G that coincide with on V@ (G). The maximum principle shows that < 1 . It is left to the reader to nish the existence proof by checking that satis es (5.13) at every z 2 Vint(G). Suppose that : V (G) ! (0; 1) is any positive function. Let z; z + ij ] be some edge of G , and set
z; z + ij ] = (?1)j log (z)(?1)j + 1 + (?1)j+1 log (z + ij )(?1)j + 1 : (6.2)
Note that z; z + ij ] = ? z + ij ; z], and that z; z + ij ] is strictly monotone increasing in (z) and strictly monotone decreasing in (z + ij ). It is straightforward to verify that equation (5.13) for at z 2 Vint(G) is equivalent to the equation
3 X

j =0

z; z + ij ] = 0:

(6.3)

Now suppose that 1; 2 : V (G) ! (0; 1) satisfy (5.13) at every z 2 Vint (G), and 1 = 2 on V@ (G). Let V0 be a connected component of the set of v 2 V (G) such that 1 < 2 , if this set is not empty. Clearly, V0 Vint(G), and hence,
3 XX

z2V0 j =0

z; z + ij ] = 0;

24

ODED SCHRAMM

for k = 1; 2. Each edge that connects two vertices in V0 contributes a total of 0 to any of these sums. Therefore,

z z0 z02V0 z 2V0 =

z; z0 ] = 0;

k = 1; 2:

But this is impossible, because monotonicity implies that each such edge z; z0 ] contributes more to the 2 sum than to the 1 sum. Hence we see that V0 does not exist, and 1 > 2 everywhere. Similarly, 1 6 2 everywhere, and uniqueness holds. As we shall see in the next section, SG patterns in S 2 are highly non rigid: there is an in nite dimensional family of them. But under certain restrictions, rigidity and uniqueness can be guaranteed. The fundamental application of such rigidity is in the setting of the celebrated Rodin-Sullivan Theorem 22]. There, the uniqueness (up to similarity) of the circle packing with the combinatorics of the hexgonal grid is used to show that Thurston's scheme of approximating conformal maps by nite circle packings works. Rodin and Sullivan prove uniqueness by applying the quasiconformal deformation theory to the group generated by inversions in the circles. There were several generalizations of the Rodin-Sullivan rigidity result, and a few di erent proofs. Of these, we mention 23], which applies also to packings by shapes other than circles, Dubejko's 11], which proves rigidity of branched packings, and He's recent 14], which applies to `embedded circle patterns' with prescribed intersection angles. In fact, He's result is relavant for SG patterns, and gives the following 7.1. First Rigidity Theorem. Any planar embedded SG circle pattern is the image of the at pattern (where all circles have radius 1 ) under a similarity. This theorem implies that the Rodin-Sullivan theorem is applicable to circle patterns based on the combinatorics of the square grid. We now sketch two more proofs of the Rigidity Theorem, and then prove a stronger theorem, which will provide yet another proof. The motivation for providing multiple proofs is to show a spectrum of techniques available for dealing with SG patterns. Adaptation of the Rodin-Sullivan proof. Here we must assume that the reader is familiar with the original argument from 22]. Suppose that two such patterns C 1; C 2 are given. Let f0 : C ! C be a quasi-conformal map that takes each circle of C 1 to the corresponding circle of C 2 . Denote the inversion in a circle c by jc . Let z 2 C . If there is only one circle c of C 1 whose interior contains z , de ne
7. Rigidity of SG Patterns

f1(z) = jf (c) f0 jc(z): If there are two such circles, c; c0 , then set f1 (z) = jf (c) jf (c0) f0 jc0 jc(z):

SG CIRCLE PATTERNS

25

0 Note that jc and jc commute, since the circles c and c0 are orthogonal. Also observe that f1 is K -quasi-conformal, with the same K as f0 , and that f1(C 1 ) = C 2 . Similarly, f2 can be de ned from f1 as f1 was de ned from f0 . Continuing in this manner, a sequence f0 ; f1 ; f2 ; : : : is de ned. For any point z which is the orbit of some tangency of two circles of C 1 under the group generated by inversions in circles of C 1 , the image fn (z) remains the same when n is large enough, that is fn (z) = fn+1(z) for all su ciently large n . Since these points are dense in C n , it follows that the limit f = lim fn exists. This limit f must conjugate the group generated by inversions of circles in C 1 to the group generated by inversions of circles in C 2 . It then follows that f is a Mobius transformation from Sullivan's rigidity theorems for quasiconformal deformations of Mobius groups, as in 22]. A proof using the Mobius invariants. Suppose that C is such an embedded pattern. Any circle c 2 C has a chain of eight circles surrounding it in a chain, of which four are half-neighbors, see Figure 7.1. From this is is easy to see that the Ring Lemma of 22] applies, and shows that the ratio of the radius of any circle to the radius of a half-neighbor is bounded from above. The same would then also be true of the full neighbors. It then follows that the Mobius invariants and of C are bounded.

Figure 7.1. The ring lemma.

Let M be the supremum of the value of , and let fzng be a sequence in V (SG) such that limn (zn) = M . For any subsequence such that (zn + ij ) converges for j = 0; 1; 2; 3, we must have

M = lim (zn) e? = H lim (zn + 1); lim (zn + i); lim (zn ? 1); lim (zn ? i) e? e 6 H lim (zn + 1); M; M; M 6 H (M; M; M; M ) = M:
So equality must hold throughout, which implies that lim (zn + 1) = M . A similar argument gives lim (zn + ij ) = M for j = 1; 2; 3, and it follows that lim (zn + v) = M for every v 2 V (SG). Consequently, given any r; > 0, there is some z 2 V (SG) such that (z0 ) 2 M ? ; M ] for every z0 2 V (SG) satisfying

26

ODED SCHRAMM

jz0 ? zj < r . It follows that is also almost constant in the r -ball around z . Therefore, the pattern C has arbitrarily large parts where is arbitrarily close to M and is almost xed. In these parts, C approximates an SG -Doyle spiral with

= M to arbitrary accuracy or the at SG circle pattern. But the SG -Doyle spirals are not embedded. Consequently, sup = M = 1. Similarly, we see that inf = 1, and because C is embedded, must also be 0 everywhere. 7.2. Second Rigidity Theorem. Let C be an SG circle pattern on the sphere S 2 , and let ; be its invariants. Suppose that supz (z) < 1 or inf z (z) > 0 . Then C is Mobius equivalent to an SG Doyle spiral, or to the at SG circle pattern. The proof uses methods of random walks and harmonic functions, as in 26], 27], 2], 11], but is most similar to an argument from 14]. We refer the reader to 8], 25] or 30] for background on random walks and harmonic functions. Proof. It is enough to treat the case sup < 1 , because (z) = (iz)?1 is also a -invariant of a circle pattern (by 5.3.(4)). So assume sup < 1 . We must show that is constant. Assume that it is not, and let a 2 (inf ; sup ) be arbitrary. Set Va = fz 2 V (SG) : (z) > ag . Then Va and V (SG) ? Va are both nonempty. Let Ea be the set of edges of SG that connect two vertices in Va . Recall the de nition (6.2) of z; z + ij ]. We have noted that z; z + ij ] = ? z + ij ; z], which means that can be thought of as a ow on the network SG . Moreover, the ow is preserving, in the sense that (6.3) holds for . De ne z; ij z; z + ij ] = (z) ? z(++ ]ij ) : (7.1) z (If (z) = (z + ij ) set z; z + ij ] = 1.) It is clear that z; z + ij ] = z + ij ; z] > 0. Because (z); (z)?1 are bounded in Va , and the partial derivative of z; z + ij ] with respect to (z) is positive and continuous, it follows that sup z; z0 ] < 1: 0
z;z ]2Ea

A consequence of (6.3) is that the function is then a harmonic function on the network (SG; ) with conductances given by , which just means that P3 z; z + ij ] (z + ij ) (z) = j=0 3 (7.2) P z; z + ij ] : j =0 Let z; z0 ] = z; z0 ] for edges z; z0 ] 2 Ea and z; z0 ] = 0 otherwise. Polya's theorem tells us that the network (SG; 1) is recurrent, and the same is clearly true for (SG; c) if c > 0 is any constant. Since reducing the conductances of a recurrent network produces a recurrent network, the network (Va ; Ea; ) is recurrent. However, because 6 a outside of Va and > a in Va , it follows from (7.2) that is subharmonic in (Va; Ea; ). Since any bounded subharmonic function on a recurrent network is constant, is constant on Va . It follows that is constant on V (SG), which proves the theorem. As remarked earlier, for a planar embedded pattern the invariant is bounded, so the second rigidity theorem gives another proof of the rst rigidity theorem. But we actually have more,

SG CIRCLE PATTERNS

27

7.3. Corollary. Suppose that the radius function r = rC of a planar SG pattern ?1

C satis es c 6 r(z + 1)=r(z) 6 c for some constant c . Then C is Mobius

equivalent to an SG Doyle spiral, or to the at pattern. It is quite surprizing that we do not need to assume anything about the ratios r(z + i)=r(z). Note that there are Doyle spirals with r(z + i)=r(z) arbitrarily large, but r(z + 1)=r(z) = 1 for all z . Proof. From (5.3), we get,

?1 + r(z + i)2 =r(z)2 ?1 + r(z ? i)2 =r(z)2 1 (z)2 = ? 2 =r(z )2 ?1 + r(z ? 1)2 =r(z )2 > (1 + c2 )2 > 0; 1 + r(z + 1)

and therefore inf > 0. It would be very interesting to nd other natural conditions that imply that a given circle pattern must belong to a speci c nite dimensional family.
8. Immersions in the Sphere

In addition to the erf-like patterns, we will exhibit below another explicit nite dimensional family of non-planar SG circle patterns. But rst, Theorems 5.2 and 5.1 are used to show that the collection of all SG patterns in the sphere is in nite dimensional. 8.1. Lemma. There is some c > 0 such that the following is true. Set (z) = 1 on the diagonals fz 2 V (SG) : Re z = Im zg . On the set fz 2 V (SG) : Re z = j Im zj + 1g de ne arbitrarily but with > c2 there. Then can be extended to all of V (SG) in such a way that > 0 and equation (5.13) is satis ed everywhere. Proof. Note that has already been de ned at z = 1 and three neighbors of 1. Hence equation (5.13) determines the value of at the remaining neighbor, 2. Now, the equation at 1 + i and 1 ? i determines the value of at 2 + i and 2 ? i . Proceeding in this manner, the value of is determined on fz 2 V (SG) : Re z = j Im zj + 2g . Then one can continue, and determine the value of on all of the quadrant fz 2 V (SG) : Re z > j Im zjg . Let us assume for the moment that this goes smoothly, that is, all the values of in that quadrant are positive. Then indeed will satisfy (5.13) at any fz 2 V (SG) : Re z > j Im zjg . We now de ne on all of V (SG) by requiring that (Rz) = (z)?1 (8.1) will hold for any z 2 V (SG), where R is a re ection in either of the diagonals Re z = Im z . Clearly, there is precisely one extension of to V (SG) that satis es that equivariance condition (8.1). Property 5.3.(6) shows that the equation (5.13) holds on the diagonals fz 2 V (SG) : Re z = Im zg . One can also verify, using 5.3.(4), that the equivariance relation (8.1) will force (5.13) to hold on the rest of V (SG). So all that remains is to check that constructed in this manner will be positive in the quadrant fz 2 V (SG) : Re z > j Im zjg .

28

ODED SCHRAMM

Assume that c satis es

c2

1 Y?

Clearly, this holds for c su ciently large. Set

j =1

1 + c1?j ?1 ?
n Y?

1 X
j =1

c?j?1 > c:
n X j =1

(8.2)

B(n) = c2

We shall prove by induction on Re z that (z + 1)= (z) > B(Re z ? j Im zj); (8.3) holds for every fz 2 V (SG) : Re z > j Im zjg . This is clear whenever Re z = j Im zj . So we move on to the induction step. Let z satisfy z 2 V (SG) and Re z > j Im zj , and suppose that (8.3) holds for every appropriate w with Re w < Re z . In particular, it follows that (w) > 0 for w 2 V (SG) with j Im wj 6 Re w 6 Re z . Set zk = j Im zj + k + i Im z , and let n = Re z ? j Im zj . Then z = zn . Also set ak = (zk ), and bk = ak+1=ak . The inductive assumption and (8.2) imply that bk > c for k = 1; 2; : : : ; n ? 1. Because b0 > c2 , it follows that ak > ck+1 for k = 1; : : : ; n . Since (z i) > 0, equation (5.13) gives a2 < (an+1 + 1)(an?1 + 1): n This implies, bn > ?1 1 ?1 ? a?1 > ?1 1 ?n?1 ? c?n?1: (8.4) n bn?1 + an bn?1 + c By the inductive hypothesis, we have bn?1 > B(n ? 1): Since the right hand side of (8.4) is monotone increasing in bn?1 , and since B(n ? 1) 6 c2 , we get 1 B( ? 1) bn > B(n ? 1)?1 + c?n?1 ? c?n?1 = 1 + B(nn? 1)c?n?1 ? c?n?1 Q?? ?1 P ? B(n ? 1) ? c?n?1 = c2 n=11 1 + c1?j ? n=11 c?j?1 ? c?n?1 j j > 1 + c1?n 1?n 1+c

j =1

1 + c1?j ?1 ?

c?j?1:

> c2

n Y?

j =1

n X 1 + c1?j ?1 ? c?j?1 = B(n); j =1

which completes the induction step and the proof. An interesting corollary of the proof is that for any > 0 de ned in fz 2 V (SG) : j Im zj 6 1; j Re zj > 0g and satisfying (5.13) in fz 2 V (SG) : Im z = 0; j Re zj > 0g , if (0) > 1 and (1)= (0) > c2 , then (n) > cn for every positive integer n . This is a special nonlinear e ect.

SG CIRCLE PATTERNS

29

in S . Proof. The theorem follows from 8.1, 5.2 and 5.1. The erf-like patterns are analogous to functions having Schwarzian az2 + bi , b; a 2 R . We shall now exhibit a family of patterns that are analogous to functions having Schwarzian a(1 ? i)z , where a 2 R . The formuli for the invariants are, ? (z) = eax+ay; (z) = 1 e?ax + eay 2 : (8.5) 4 It is immediate to verify that these satisfy the nonlinear discrete Cauchy-Riemann equations (5.11) and (5.12). The collection described by (8.5) is 1-dimensional, but we may translate z by a complex constant, and obtain a 3-dimensional family. In Figures 8.1.a and 8.1.b you can see the circle pattern and the corresponding function that has Schwarzian equal to (1 ? i)z (which is a ratio of two Airy functions). It is interesting to note that the pictures are nearly invariant by a 2 =3 rotation. The explanation is the same as given in Section 5 for the =4 rotational symmetry of the function whose Schwarzian is z2 . The special feature here is that while the combinatorics of SG does not have 2 =3 rotational symmetry, the circle pattern does `try' to be symmetric.

8.2. 2Theorem. There exists an in nite dimensional family of SG circle patterns

Figure 8.1.a.

Figure 8.1.b.

Property 5.3.(5) led to the observation that (z) = eax+ay is a solution of (5.13). However, there is another possible route to the discovery of the pattern given by (8.5). Recall the erf-like pattern described by (5.14). Substituting there z + c(1 ? i) for z and a2c for b , where c 2 R , and taking the limit as c ! 1 gives (z) = ax+y ;

? (z) = a?x + ay 2;

which is a basically a translated variant of (8.5). In other words, the patterns described by (8.5) are locally uniform limits of erf-like patterns.

30

ODED SCHRAMM

It would be very interesting to nd new explicit SG patterns, in addition to those that we have, and to nd patterns based on the hexagonal grid, or other combinatorics. In joint work with Rick Kenyon, we have found explicit patterns analogous to log z and to polynomials. (Some of these are conjectural, but the pictures show that they are valid.) However, the only explicit, immersed (locally univalent) patterns currently known are those described here, and some simple at patterns. In the hexagonal combinatorics, the only explicit patterns known are the Doyle spirals and the at pattern. (Using the partial symmetry of the e function H , it is possible to create variations on the patterns we know. For example, x?y ; (z ) = ?a?x + a?y ?2 can be obtained from (8.5). But these trivial (z) = a modi cations will not produce new pictures.)
9. Convergence of Circle Patterns to Meromorphic Functions

Carter and Rodin have shown that prescribing the boundary radii of a circle packing according to jf 0 j , where f is a locally injective analytic function, gives an approximation of f 6]. The following is a similar theorem for the Schwarzian derivative in place of f 0 . 9.1. Theorem. Let D C be a simply connected bounded domain, and let W C ^ be an open set that contains the closure of D . Suppose that f : W ! C is a locally injective meromorphic function. Assume, for convenience, that 0 2 D . For n 2 N , let SG(nD) be the union of squares of SG that are contained in nD , and let SGn be the closure of the connected component of the interior of SG(nD) that contains 0 . At boundary vertices v 2 V@ (SGn) set
n (v ) = 1 + n2 Re Sf (v=n);

where Sf is the Schwarzian derivative of f . Using Theorem 6.2, extend n to a solution of the SG -Dirichlet problem on SGn . Let n be the companion of n in the solution of the SG -CR equations that satis es

? n ^ Let C n = Cz : z 2 V (SGn) be the SGn circle pattern in C that has n and n as its Mobius invariants, whose existence is guaranteed by Theorem 6.1. Suppose that C n is normalized by Mobius transformations so that
pn (1 + i)=2 = f (1 + i)=2n ; ? ? pn (?1 + i)=2 = f (?1 + i)=2n ; ? ? pn (1 ? i)=2 = f (1 ? i)=2n ;

n (1 + i)=2 = 1 + n2 Im Sf (1 + i)=2n

(9.1)

(9.2)

where pn = pCn . For points z 2 D set fn (z) = pn(w) , where w is a vertex of V (SGn) closest to nz . Then, fn ! f uniformly on compact subsets of D as n ! 1.

SG CIRCLE PATTERNS

31

9.2. Lemma. Set

for every v 2 V (SGn) . Then

hn(v) = Re Sf (v=n); ? tn (v) = n2 n(v) ? 1 ;

? tn (v) ? hn(v) = O n?2 : ?n?4 :

Here and below, the notation g1 = O(g2 ) means that there is a constant C , which may depend on W; D; f , but may not depend on n and v , such that jg1j 6 g2 wherever g1 is de ned. A consequence of the lemma is,
n(v ) = 1 + n?2 Re Sf (v=n) + O

Proof. We shall rst show that hn = O n?4 , where is the discrete Laplacian, ? g(z) = (1=4) g(z + 1) + g(z + i) + g(z ? 1) + g(z ? i) ? g(z). Indeed, expand Sf in power series about z0 = v=n to get

(9.3)

Sf (z) = b0 + b1 (z ? z0 )1 + b2 (z ? z0)2 + b3 (z ? z0 )3 + O jz ? z0j4 :

This gives,

0 1 3 X ? j A 1 hn(v) = Re @Sf (z0 ) ? 4 Sf z0 + i =n j =0 0 3 1 X ?1 j ?2 2j ?3 3j A ? ?4 1 = Re @ 4 n b1i + n b2i + n b2i +O n

j =0

? = O n?4 ;
(9.4) (9.5)

P because 3=0 ikj = 0 for k = 1; 2; 3. j

Consider the function

where 2 0; n?2 is some function of n . We want to determine how large must be so that g will have no maximum in Vint (SGn). So suppose that g has a maximum at z , which is in Vint(SGn). Then for j = 0; 1; 2; 3,

g(z) = tn(z) ? hn(z) + jzj2 ;

tn z + ij 6 qj ;
where

(9.6)

(9.7) qj = tn(z) + hn z + ij ? hn(z) ? z + ij 2 + jzj2 : We prepare for the following computation by gathering some information about the qj 's. Note that jzj = O(n), because z 2 Vint(SGn). Since < n?2 , this gives

z + ij 2 ? jzj2 = O n?1 ; ? z + ij 2 + z + ij+2 2 ? 2 jzj2 = O n?2 :

32

ODED SCHRAMM

When using these estimates, the smoothness of Re Sf and the maximum principle for n , we get qj = O (1) ; ? qj ? tn(z) = O n?1 ; (9.8) ?n?2 : qj + qj+2 ? 2tn(z) = O Finally, from (9.4) it easily follows that

? 4tn(z) ? q0 ? q1 ? q2 ? q3 = 4 + O n?4 : ?

(9.9)

We are now ready for the real work. From (9.6), the de nition of tn and the e monotonicity of H 5.3.(2), we get

e 1 + n?2tn(z) 6 H 1 + n?2q0 ; 1 + n?2q1 ; 1 + n?2q2 ; 1 + n?2q3 : e After using the de nition of H and taking logarithems, this gives,
2 log 1 + n?2tn(z) ? ? 6 log 2 + n?2 q0 ? log 1 + 1 + n?2q1 ?1 +

? We expand log(1 + x) = x ? x2 =2 + x3 =3 : : : , and make an O n?6 analysis:

2n?2tn(z) ? n?4tn(z)2 + O n?6 2 2 6 n?2q0 =2 ? n?4 q0 =8 + n?2q1 =2 ? 3n?4 q1 =8 + ? 2 2 + n?2q2 =2 ? n?4q2 =8 + n?2q3 =2 ? 3n?4q3 =8 + O n?6 : Rearranging and applying (9.9) gives,

+ log 2 + n?2q2 ? log 1 + 1 + n?2q3 ?1 ? ? ? = log 1 + n?2 q0=2 + log 2 ? log 1 + n?2 q1=2 ? log 2 + log 1 + n?2q1 + ? ? ? + log 1 + n?2q2 =2 + log 2 ? log 1 + n?2q3 =2 ? log 2 + log 1 + n?2q3 ? ? ? = log 1 + n?2 q0=2 ? log 1 + n?2q1 =2 + log 1 + n?2q1 + ? ? ? + log 1 + n?2q2 =2 ? log 1 + n?2 q3=2 + log 1 + n?2q3 :

? 1 ? 3 ? 2 2 2 2 2n?2 6 8n4 2tn(z)2 ? q0 ? q2 + 8n4 2tn(z)2 ? q1 ? q3 + O n?6 : (9.10)

2 2 We now estimate the expression 2tn(z)2 ? q0 ? q2 , using (9.8). 2 2 2t2 ? q0 ? q2 = (2tn ? q0 ? q2) (tn + q0 ) + (q2 ? q0 ) (tn ? q2) n ? ? ? = O n?2 + (q2 ? tn) + (tn ? q0 ) O n?1 ? ? ? ? = O n?2 + O n?1 O n?1 = O n?2 :

? 2 2 Similarly, 2t2 ? q1 ? q3 = O n?2 . From these estimates and (9.10), we conclude n that ? = O n?4 : This means that if we choose = Cn?4 with C > 0 a su ciently large constant, then g will have ?no maxima in Vint (SGn?). In that case, on V@ (SGn) we have g(z) = jzj2 = O n?2 . Hence, g(z) 6 O n?2 in V (SGn), which implies, ? tn(z) 6 hn(z) + O n?2 for z 2 V (SGn): (9.11) The proof for the reverse inequality is almost the same. The only modi cations needed are reversing the sign of and a few of the inequalities. It is left to the reader to verify the details. 9.3. Lemma. Let w 2 V (SGn ) , and let n(w) denote the combinatorial distance in SGn from w to (1 + i)=2 ; that is, the least k such that there is a sequence fw1; w2; : : : ; wk = wg V (SGn ) such that w1 = (1 + i)=2 and jwm+1 ? wm j = 1 for m = 1; : : : ; k ? 1 . Then, ? (9.12) n (w) = 1 + n?2 Im Sf (w=n) + n (w)O n?4 : Note that if @D is smooth, then n = O(n). In general, n = O(n) on compact subsets K D , where the constant implicit in the O(n) notation may depend on K . In any case, on compact subsets of D we have ? (9.13) n (w) = 1 + n?2 Im Sf (w=n) + O n?3 : Proof. Taking logarithms in equation (5.11), and using Lemma 9.2 gives, log (a + 1=2) ? log (a ? 1=2) ? ? = 2 log (a + i=2)?1 + 1 ? 2 log (a ? i=2)?1 + 1 ? ? = 2 log 1 + n?2tn (a + i=2) ?1 + 1 ? 2 log 1 + n?2 tn(a ? i=2) ?1 + 1 ? ? ? = 2 log 2 ? n?2 tn(a + i=2) ? 2 log 2 ? n?2tn(a ? i=2) + O n?4 ? = 2 log 1 ? n?2 tn(a + i=2)=2 + 2 log 2 ? ? ? 2 log 1 ? n?2tn(a ? i=2)=2 ? 2 log 2 + O n?4 ? = n?2tn(a ? i=2) ? n?2 tn(a + i=2) + O n?4 ? ? ? ? = n?2 Re Sf (a ? i=2)=n ? Sf (a + i=2)=n + O n?4 ? ? 0 = n?2 Re (?i=n)Sf (a=n) + O n?4 ? 0 = n?3 Im Sf (a=n) + O n?4 ? ? ? = n?2 Im Sf (a + 1=2)=n ? n?2 Im Sf (a ? 1=2)=n + O n?4 : (9.14) Similarly, equation (5.12) implies log (b + i=2) ? log (b ? i=2) ? ? = 2 log 2 + n?2tn (b + 1=2) ? 2 log 2 + n?2 tn(b ? 1=2) ? = n?2tn(b + 1=2) ? n?2tn(b ? 1=2) + O n?4 (9.15) ?S 0 (b=n) + O ?n?4 = n?3 Re f ? ? ? = n?2 Im Sf (b + i=2)=n ? n?2 Im Sf (b ? i=2)=n + O n?4

SG CIRCLE PATTERNS

33

34

ODED SCHRAMM

Note that

? ? ? log (1 + i)=2 = n?2 Im Sf (1 + i)=2n + O n?4 ; ? log (w) ? n?2 Im Sf (w=n) = n(w)O n?4 ;

(9.16)

follows from (9.1). Then induction shows that (9.14), (9.15) and (9.16) imply and the lemma follows. Proof (of 9.1). For each v 2 V (SGn) let Mv denote the Mobius transformation that takes the three points (1 + i)=2n; (?1 + i)=2n; (1 ? i)=2n to the points pn (v + (1 + i)=2) ; pn (v + (?1 + i)=2) ; pn (v + (1 ? i)=2), respectively. b We shall use the matrix notation a d for a Mobius transformation (az + c b)=(cz + d). The matrix representing a Mobius transformation is well de ned only up to a scalar factor, and we will freely cancel scalar factors from our matrices. Let A = An be the Mobius transformation that takes 1; 0; 1 to (1 + i)=2n; (?1 + i)=2n; (1 ? i)=2n , respectively. Then

A = 11=ni ?(1 +i i)=2n ; ?

i i) A?1 = i ? 1 (1 +=n=2n : 1

(9.17)

Recall the de nition (near (5.5)) of the Mobius transformations mv and m v1;v2 ] . With those notations, we have

Mv A = mv+(1+i)=4:
So we get
? ?1 Mv 1 Mv+1 = A mv+(1+i)=4 mv+1+(1+i)=4 A?1 1 1 = A m?+(1+i)=4 mv+(3+i)=4 m?+(3+i)=4 mv+1+(1+i)=4 A?1 v v = A m v+(1+i)=4;v+(3+i)=4] m v+(3+i)=4;v+(5+i)=4] A?1 :

Using the formulas (5.6), (5.7) and (9.17), this gives,

=n ? Mv 1 Mv+1 = 11? i

?(1 + i)=2n
i
0 ?1 1

? (v + 1)?1 + 1 ?1

?i v + (1 + i)=2 1=2 1
?1

1 i (1 + i)=2n : i?1 1=n

We use (9.3), (9.13) to substitute for ; , then simplify (using the fact that Sf is Lipschitz), and obtain,
?1 ? ? Mv 1 Mv+1 = ?n?1S 1(v=n)=2 n1 + O n?2 : f

SG CIRCLE PATTERNS

35

Let I denote the identity matrix, and set 1 B(z) = ?Sf 0z)=2 0 : ( Then Similarly,

? ? Mv 1 Mv+1 = I + n?1B(v=n) + O n?2 :

(9.18)

(9.19) Since the statement of the theorem is Mobius invariant, we assume, with no loss of generality, that

? Mv 1 Mv+i = A m v+(1+i)=4;v+(1+3i)=4] m v+(1+3i)=4;v+(1+5i)=4] ? =n ?i v + (1 + i)=2 ?1=2 = 11? i ?(1 +i i)=2n ? (v + i) + 1 ?10 i 0 i?1 ?1 1 ?1 B (v=n) + O ?n?2 : = I + in

A?1 ?1 1 1 (1 + i)=2n 1=n

f (0) = 0;
It follows that,

f 0 (0) = 1;

f 00 (0) = 0:

(9.20)

f ij (1 + i)=2n = ij (1 + i)=2n + O n?3 ; for j = 0; 1; 2; 3. The normalization (9.2) and the de nition of Mv then give

? ? M0 ij (1 + i)=2n = ij (1 + i)=2n + O n?3 ; ? M0 = I + O n?1 :

C 0(z) = C (z)B(z);

for j = 0; 1; 3. Let Q be the Mobius transformation Q(z) = nz . The transformation Q M0 Q?1 moves each of the?three points ij (1 + i)=2 a distance of at most ? O n?2 . So Q M0 Q?1 = I + O n?2 . Consequently, (9.21) (9.22) (9.23) (9.24) Let C (z) be the matrix solution of the di erential equation with initial condition C (0) = I . We claim that for any compact K D

? Mv = C (v=n) + O n?1 ;

for v 2 V (SGn) such that v=n 2 K . Indeed, equation (9.22) implies that

? ? C (z + ij =n) = C (z) + ij =n C (z)B(z) + O n?2 ;

36

ODED SCHRAMM

for j = 0; 1; 2; 3. We also have,

? ? Mv+ij = Mv + ij =n Mv B(v=n) + Mv O n?2 :

(9.25)

For j = 0; 1, this is just (9.18) and (9.19), and for j = 2; 3, it follows by inverting both sides of these identities and noting that B(z) is Lipschitz. Consequently, we have,

? ? Since M0 = I + O n?1 = C (0) + O n?1 , it follows by induction on the combinatorial distance (v) from v to 0 in SGn that
Mv ? C (v=n) 6

Mv+ij ? C (v + ij )=n ? ? 6 Mv ? C (v=n) + n?1 (Mv ? C (v=n)) B(v=n) + 1 + jMv j O n?2 ? ? ? ? = Mv ? C (v=n) 1 + O n?1 + 1 + jMv j O n?2 :

? ? (v)O n?2 + O n?1

? 1 + O n?1

(v)

Since (v) = O(n) on any compact K

? ? Mv ? C (v=n) 6 O n?1 eO(1) = O n?1

D , this implies

on K , which proves (9.23). For v 2 V (SGn), and j = 0; 1; 3, we have

? ? ? pn v + ij (1 + i)=2 = Mv ij (1 + i)=2n = C (v=n)(0) + O n?1 : ? ? fn (z) = C (z)(0) + O n?1 ;

(z) b C (z) = a(z) d(z) : c (z)

For j = 2, this also holds, because the four points pn v + ij (1+ i)=2 , j = 0; 1; 2; 3 p ? are on the same circle, Mv z : jzj = 1= 2n . Therefore, on compacts K D . It remains to verify that C (z)(0) = f (z). Let us denote the entries of the matrix C (z) by, (9.26) (9.27)

We claim that the di erential equation (9.22) and the de nition of B(z) imply that

Sb=d = Sf :
Indeed, from (9.22) it follows that,

a0 (z) = ?b(z)Sf (z)=2; b0 (z) = a(z):

SG CIRCLE PATTERNS

37

This implies that b satis es the equation 2b00 + Sf (z)b = 0: (9.28) Similarly, d satis es the same equation. Since b and d are both solutions of this equation, it follows that (b0 d ? bd0 )0 = 0. Since b0 d ? bd0 is equal to 1 at z = 0, we must have b0 d ? bd0 = 1 everywhere. Hence (b=d)0 (z) = d(z)?2 . This gives
00 0 1 00 Sb=d = ((b=d))0 ? 2 ((b=d))0 b=d b=d
2

0 0 1 20 = ?2d ? 2 ?dd d

00 = ?2 dd ;

and (9.27) follows, since d satis es (9.28). (Compare 19, x1.2].) Observe that (b=d) (0) = 0, (b=d)0 (0) = 1, and (b=d)00 (0) = 0. So b=d agrees with f to third order at z = 0 and has the same Schwarzian as f . Consequently, C (z)(0) = b(z)=d(z) = f (z), and the proof is complete. Remark. The proof of Theorem 9.1 actually demonstrates that the circle patterns ? C n approximate f; f 0 ; f 00 with errors O n?1 on compacts of D , and if @D is smooth, then the convergence is uniform in D .
10. For Desert: Some Open Problems

Let X be the apace of all planar SG immersions (with the topology of locally uniform convergence, say). Problem 1. Is X in nite dimensional?
Problem 2. Is X contractible? Problem 3. Find interesting nite dimensional subfamilies of X . For example, nd families characterized by speed of growth or a simple geometry.

Recall that an SG immersion de nes a Riemannian surface over the sphere or the p plane (De nition 3.2). We have identi ed the surface determined by the iSG -erf pattern. It is likely that similar techniques will answer: Problem 4. Describe the surfaces determined by the other SG immersions we have discussed. Problem 5. Find su cient or necessary condition for a Riemannian surface (locally euclidean or locally spherical) to be the surface determined by an entire SG immersion. In other words, on what surfaces an entire SG pattern embeds? Note that Callahan and Rodin 5] give a necessary condition: the surface must be regularly exhaustible. In particular, Picard's theorem is satis ed. Problem 6. Suppose that the surface determined by an entire SG immersion C is the universal cover of C ? f0g . Is C necessarily a Doyle spiral?

38

ODED SCHRAMM

References
1. A. F. Beardon & K. Stephenson, The uniformization theorem for circle packings, Indiana University Math. J. 39 (1990), 1383{1425. 2. I. Benjamini & O. Schramm, Harmonic functions on planar and almost planar graphs and manifolds, via circle packings, Invent. Math. (to appear). 3. P. L. Bowers & K. Stephenson, A branched Andreev-Thurston Theorem for circle packings of the sphere, to appear, Proc. LMS. 4. R. Brooks, On the deformation theory of classical Schottky groups, Duke Math. jour. 52 (1985), 1009{1024. 5. K. Callahan & B. Rodin, Circle packing immersions form regularly exhaustible surfaces, Complex Variables 21 (1993), 171{177. 6. I. Carter & B. Rodin, An inverse problem for circle packing and conformal mapping, Trans. Amer. Math. Soc. 334 (1992), 861{875. 7. P. G. Doyle, Z.-X. He, & B. Rodin, Second derivatives of circle packings and conformal mappings., Discrete & Comput. Geom. 11 (1994), 35-49. 8. P. G. Doyle and J. L. Snell, Random walks and electric networks, The Carus Math. Monographs 22, Math. Association of America, 1984. 9. T. Dubejko, Branched circle packings and discrete Blaschke products, Trans. Amer. Math. Soc. 347 (1995), 4073{4103. 10. , In nite branched packings and discrete complex polynomials, Journal of Lond. Math. Soc. (to appear). 11. , Recurrent random walks, Liouville's theorem, and circle packings, MSRI Preprint No. 040-95 (1995). 12. Z.-X. He, An estimate for hexagonal circle packings, J. of Di erential Geometry 33 (1991), 395{412. 13. , Coarsely quasi-homogeneous circle packings in the hyperbolic plane, Michigan Math. J. 41 (1994), 175{180. , Rigidity of disk packings with overlappings I, preprint (1996). 14. 15. Z.-X. He & O. Schramm, Fixed points, Koebe uniformization and circle packings, Ann. of Math. 137 (1993), 369{406. 16. , Hyperbolic and parabolic packings, Discrete Comput. Geom. 14 (1995), 123{149. , On the convergence of circle packings to the Riemann map, Invent. math. (to appear). 17. 18. P. Koebe, Kontaktprobleme der konformen Abbildung, Berichte Verhande. Sachs. Akad. Wiss. Leipzig, Math.-Phys. Klasse 88 (1936), 141{164. 19. O. Lehto, Univalent Functions and Teichmuller Spaces, Graduate texts in mathematics, vol. 109, Springer-Verlag, New York, 1987, pp. 257. 20. S. Malitz & A. Papakostas, On the angular resolution of planar graphs, STOC 24 (1992), 527{538. 21. G. L. Miller & W. Thurston, Separators in two and three dimensions, Baltimore, May 1990, Proceedings of the 22nd Annual ACM Symposium on Theory of Computing, ACM, pp. 300309. 22. B. Rodin & D. Sullivan, The convergence of circle packings to the Riemann mapping, J. of Di erential Geometry 26 (1987), 349{360. 23. O. Schramm, Rigidity of in nite (circle) packings, J. Amer. Math. Soc. 4 (1991), 127{149. 24. , How to cage an egg, Invent. math. 107 (1992), 543{560. 25. P. M. Soardi, Potential Theory on In nite Networks, Lecture notes in mathematics, vol. 1590, Springer, Berlin, pp. 187. 26. K. Stephenson, Circle packings in the approximation of conformal mappings, research announcement, Bull. Amer. Math. Soc. 23 (1990), 407{415. , A Probablistic proof of Thurston's conjecture on circle packings, preprint. 27. 28. W. P. Thurston, The Geometry and Topology of 3-manifolds, Princeton University Notes, Princeton, New Jersey, 1982. 29. , The nite Riemann mapping theorem, invited address, International Symposium in Celebration of the Proof of the Bieberbach Conjecture, Purdue University, 1985.

SG CIRCLE PATTERNS

39

30. W. Woess, Random walks on in nite graphs and groups | a survey on selected topics, Bull. London Math. Soc. 26 (1994), 1{60.
The Weizmann Institute, Mathematics Department, Rehovot 76100, Israel

E-mail address : schramm@wisdom.weizmann.ac.il WWW: http://www.wisdom.weizmann.ac.il/~schramm