Contribution from: Mathematical Physics Studies Vol. 27 Perspectives in Analysis Essays in Honor of Lennart Carleson’s 75th Birthday Michael Benedicks, Peter W. Jones, Stanislav Smirnov (Eds.) © Springer Verlag Berlin Heidelberg 2005 ISBN 3-540-30432-0 On Scaling Properties of Harmonic Measure Peter W. Jones Department of Mathematics, Yale University, New Haven, CT 06520, USA jones-peter@yale.edu 1 Introduction This short article gives an exposition of some open problems concerning harmonic measure for domains in Euclidean space. This is an area where Lennart Carleson has been a leading figure for about fifty years. Like many other analysts, this author had his career shaped to a large extent by Lennart, and this provides me an opportunity to both wish him a happy birthday and thank him for all the assistance he has given me for nigh thirty years. Recall that the harmonic measure for a base point z0 ∈ Ω is the (unique) probability measure ω on ∂Ω so that Z F (z0 ) = F (z)dω(z) for all harmonic functions F on Ω that are (modulo small technicalities) continuous up to ∂Ω. Despite decades of intensive research, we are far from understanding the “fine structure” of harmonic measure. Certainly the greatest achievement here is N. Makarov’s theorem [11] that on any simply connected planar domain, ω (for any base point) is supported on a set of Hausdorff dimension one, and ω puts no mass on any set of dimension less than one. In modern parlance this is written as dim(ω) = 1. A related, surprising theorem due to Carleson [5] is that for “square” Cantor sets (as boundary) in the plane, dim(ω) < 1, no matter how large the dimension of the Cantor set. In Rd , d ≥ 3, our knowledge is much less complete, though notable results due to Bourgain [4], Wolff [12], and others provide some knowledge of basic scaling laws. It was understood in the 1990’s (see [6]) that dynamical systems and Conformal Field Theory were intimately tied to these problems, at least in dimension two. It is a curious state of affairs that CFT as it is understood today, makes no notable predictions for these problems. This is perhaps even more surprising in light of the great success CFT has had with SLE. Exactly how 74 Peter W. Jones the full story will unfold remains mysterious. My guess is that we are missing fundamental concepts, and that the future will bring some unsuspected, closer connections between analysis and statistical physics. If the situation is murky for d = 2, it is much worse for d ≥ 3. All results known there are the end products of arguments of real variable type, and therefore cannot yield sharp constants. Whether there can be a unified approach to all dimensions seems now to be an untouchable problem. If we are very lucky, the new generation may give us the solution due to the rise of a young star of Lennart Carleson’s caliber and vision. 2 Harmonic Measure in the Plane We discuss here various conjectures concerning the fine structure of harmonic measure. In the plane one has some extra measure of confidence in these conjectures due to experience from conformal mappings as well as Conformal Field Theory. A guiding principle taken from CFT is that all expressions should be simple and analytic (in the parameter values). We start with some examples from dimension two and then turn to d ≥ 3 in the following section. The first conjecture we discuss concerns conformal mappings h : D → Ω, where Ω is a bounded domain. (All conjectures are the same for mappings h of {|z| > 1} to Ω, where h(∞) = ∞.) Let t ∈ C be a complex number and consider the growth of integral means for |(h0 )t |. Define βh (t) ≥ 0 to be the smallest real number such that Z 2π |h0 (reiθ )t |dθ = O (1 − r)−βh (t) . 0 Similarly we define the pressure spectrum, βsc (t) = sup βh (t) , where the supremum is taken over all conformal maps h to bounded domains. (The “sc” stands for simply connected.) It is a nontrivial fact (the “snowflake construction”) that for every ε > 0 and every t there is an h such that Z 2π |h0 (reiθ )t |dθ ≥ c(1 − r)ε−βsc (t) 0 holds for all r < 1. (See e.g. [6] for the case t = 1.) The first conjecture we list is βsc (t) = |t|2 /4 , |t| ≤ 2 (1a) βsc (t) = |t| − 1 , |t| ≥ 2 . (1b) and On Scaling Properties of Harmonic Measure 75 The phase transition at |t| = 2 can be (partially) understood by looking at t = 2 and noticing the bound |h0 (r exp{iθ})| ≤ c(1 − r)−1 , which is due to the finite area of the image domain. (Thus (1b) holds for t ≥ 2.) The conjecture in (1) was introduced by Kraetzer [9], but now carries the slightly ungainly name of the Brennan–Carleson–Jones–Kraetzer Conjecture. (This is because Brennan’s Conjecture was for the special value t = −2, while Carleson and Jones had made the conjecture for t = 1.) Actually, I. Binder’s name should be added to the list of names mentioned because he had the idea of introducing complex values of t. If we restrict to real values of t ∈ [−2, 2], the conjecture has two equivalent forms. The first of these is in terms of the multifractal formalism, which has the additional feature of adding Beurling’s estimate on harmonic measure in a more transparent form. Recall that if ω is a (locally finite) Borel measure, the f (α) spectrum is defined by f (α) = Dimension(Eα ) , where we understand the dimension in question to be the Hausdorff dimension and where (without getting too technical) Eα = z | ω D(z, r) ∼ rα . Here D(z, r) is the disk centered at z and of radius r, and the ∼ symbol means that this holds as r → 0. The reformulation of (1) (for t ≥ −2) is that for ω the harmonic measure on a bounded, simply connected domain, f (α) ≤ 2 − 1/α . (2) Again, the conjecture implicitly states that the estimate in (2) is optimal. We define as before fsc (α) = sup f (α) , where the supremum is now taken over all simply connected domains. (Unlike βsc , the estimates for f (α) do not depend on whether the domain is bounded or not.) Thus the conjecture is fsc (α) = 2 − 1/α , α ≥ 1/2 . (3) The proof that conjectures (1) and (3) are equivalent involves a stopping time argument followed by a renormalization procedure. The algebra involved yields the relationship t = 2 − 2/α for α ≥ 1/2. One builds by a snowflake construction a conformal map h where |h0 (reiθ )| ≤ c(1 − r)ε+1−1/α and where only this maximal growth is used to estimate βh (t). Conversely, given h one estimates f (α) by a similar construction. 76 Peter W. Jones This reasoning can be extended to give another (equivalent) formulation for Hölder domains. Recall that a simply connected domain Ω is called a γ Hölder domain, 0 < γ < 1, if (any) Riemann map h : D → Ω is in the Hölder space Λγ , |f (z) − f (w)| ≤ c|z − w|γ , z, w ∈ D (equivalently D). This is the same as |f 0 (z)| ≤ c0 (1 − |z|)γ−1 . The conjectures (1) and (3) are equivalent (for the appropriate range of t) to the conjecture Ω is a γ Hölder domain =⇒ dim(∂Ω) ≤ 2 − γ . (4) The algebra used here is the relation α = 1/γ, and the understanding of (4) is that the estimate 2 − γ is best possible. There is yet another formulation that is believed to be equivalent (for the range t ∈ [−2, 2]) in terms of quasiconformal mappings. For a globally quasiconformal map F let k be the bound for the lower dilatation. (So 0 ≤ k < 1 and k = 0 if and only if F is conformal.) Let βk (t) be the pressure function where the supremum is taken over all h having a globally quasiconformal extension F with bound k. The conjecture for t ∈ [−2, 2] is that βk (t) = β(kt) . (5) Some support for this is given by a remarkable theorem of Smirnov (unpublished), which is one of the few conjectured upper bounds that has been proven. Smirnov’s result is that if h has an extension F with bound k, then dim F (S 1 ) ≤ 1 + k 2 . Recall that Beurling’s theorem states that for a simply connected domain, ω D(z, r) ≤ cr1/2 where c depends only on the inradius of the base point for harmonic measure. Here D(z, r) is the disk centered at z with radius r. A trivial consequence of Beurling’s theorem is fsc (α) = 0 , α < 1/2 , because the sets in question are empty. What is the evidence for conjectures (1), (3) and (4)? First we note that (1) is true for t ≥ 2. For t = −2 this is Brennan’s conjecture, which is supposed to correspond to the (Beurling-type) conjecture fsc (1/2) = 0. For t = 1 the conjecture is due to Carleson–Jones [6], who also provided some computer evidence. A small surprise there is the tie to coefficient problems for the class Σ of functions univalent on {|z| > 1} having Laurent expansion z+ ∞ X n=1 bn z −n . On Scaling Properties of Harmonic Measure 77 Unlike the case for de Brange’s theorem (|an | ≤ n) for the class S, there seems to be no natural guess for the maximal size of bn . Instead it is shown that if Bn = sup|bn |, the supremum being taken over Σ, then lim n→+∞ log Bn = γΣ log n exists and (surprise) γΣ = β(1) − 1. (Thus the conjecture that γΣ = −3/4.) Returning to other evidence, the value α = 1, which corresponds to t = 0, should be understood as the theorem of Makarov [11]: For any simply connected domain the harmonic measure is supported on a set of (exactly) Hausdorff dimension one. (More precisely, Makarov’s theorem tells one that the |t|2 /4 conjecture is correct for t = 0 in the sense that βsc (t) should be quadratic near t = 0.) Another way of rephrasing these results and conjectures is the following list: α = 1/2 ↔ t = −2 ↔ Brennan’s Conjecture and Beurling’s Theorem α=1 ↔ t = 0 ↔ Makarov’s Theorem α=2 ↔ t = 1 ↔ Carleson–Jones Conjecture and `(Γε ) ∼ ε−1/4 By the last entry we mean that if Ω is a bounded domain then Γε is the curve {z | G(z, z0 ) = ε}, where G is Green’s function for Ω with base point z0 . If we integrate |h0 | over {G(z, 0) = ε} on the disk (essentially the circle of radius 1 − ε about the origin). The |t|2 /4 conjecture says that the extremal case is the length of Γε (which is exactly the integral of |h0 |) should be O(ε−1/4 ). The upshot is that for these three special values there is a corresponding “physical” quantity that is either known (Beurling’s or Makarov’s Theorem) or reasonably conjectured (t = 1, α = 2). It would be interesting to see if there is a longer list of natural, “physical” values of t, α. (Of course t = 2, α = +∞ corresponds to finite area.) For the fsc (α) conjecture it should be noted that the asymptotics are correct. By the results of [7], there are constants c1 and c2 such that for α ≥ 1, 2 − c1 /α ≤ fsc (α) ≤ 2 − c2 /α . (6) Now the study of f (α) is also well defined for general (i.e. non-simply connected) domains in R2 (as well as in Rd ). The results of [7] are actually for general domains, and can be expressed as bounds for the “universal” spectrum: 2 − c1 /α ≤ fu (α) ≤ 2 − c2 /α , α≥1. (7) Here fu (α) = sup f (α), the supremum being taken over all planar domains. For general domains in R2 one cannot have conjecture (3) for α < 1, as the correct result there is fu (α) = α. (It pays to use Cantor sets as boundaries.) One might think that by using Cantor sets one could obtain fu (α) > fsc (α) at least in the range α > 1. (By the theorem of Jones–Wolff [8], fu (1) = 1.) However, this turns out not to be the case. Binder and Jones have announced [2]: 78 Peter W. Jones fu (α) = fsc (α) , α≥1. (8) As a consequence, the correct conjecture is fu (α) = 2 − 1/α , α>1, (9) with the value α = 1 being a theorem. Note that by (7) the asymptotic behavior is correct. Before discussing (8) let us remark on a general (d = 2) philosophy: All such expressions are analytic functions of low complexity. This philosophy is the simple consequence of adopting a CFT point of view. It seems curious that conjectures (1) and (3) do not seem to exactly fit into our present understanding of CFT, even if all arguments (e.g. renormalization) are from the philosophy of CFT. The point here is that the correct universality class of domains does not come from the canonical lists of CFT, which include SLE. Those domains are very far from extremal (for (1) and (3)) because (e.g. 0 < κ ≤ 4) the two sides of the SLE trace are statistically the same. One expects the extremal domains for (1) and (3) to be “one sided”, i.e. the exterior domain does not have extremal behavior. It is, however, not impossible that CFT could be directly brought to bear on these problems. The result (8) has a somewhat curious history, and it is unclear if the methodology used is optimal. In any case, the proof is via a two step argument. The first step is to show that if a choice of domain Ω and α > 0 are given, then (for ε > 0) there is a polynomial P , with Julia set J, such that f (α, J c ) ≥ f (α, Ω) − ε . The polynomial P has attached to it both good news and bad news. The good news is that P can be assumed to act hyperbolically on its Julia set J. In addition, the Julia set can be chosen to be of pure Cantor type, i.e. all critical points are outside of J and in the basin of infinity, P n (c) → ∞ for all critical points c. (If Ω is simply connected, one can alternatively demand that J is a Jordan curve, and P is hyperbolic on J.) The bad news is that as ε → 0, the polynomial constructed by [2] has degree tending to infinity. This “bad” behavior is indeed necessary if J is to approximate ∂Ω in the correct technical sense. Step two in the proof of (8) is to invoke a remarkable result due to Binder–Makarov–Smirnov: If P and J are of pure Cantor type, there is a polynomial Q, degree(Q) = degree(P ), such that the Julia set is connected and f (α, A∞ ) ≥ f (α, J c ) , where A∞ is the basin of infinity for Q. It is worth remarking that while this theorem of Binder, Makarov and Smirnov appears to be simply the maximum principle, no elementary proof is known at present. To sum up, to prove (6) one first “approximates” the boundary of Ω by a polynomial Julia set of pure Cantor type, so that f (α, J c ) ≥ f (α, Ω) − ε. Then one applies the result of Binder, Makarov and Smirnov to “homotopy” J c to the compliment of a ˜ This means that for α ≥ 1, connected Julia set J. On Scaling Properties of Harmonic Measure fu (α) = fdsc (α) , 79 (10) where the subscript d stands for (conformal) dynamic. It is indeed remarkable that (10) holds, even if there were early hints in [6] that this would be the case. (There one built a non-conformal dynamical system.) The boldest form of the conjecture is that for (10) one need not consider high degree polynomials, but may restrict to Julia sets for z 2 + c, where c is in the Mandelbrot set. Though the evidence for this is slight, it is positive. If this were indeed true, it would show an unexpected form of universality for quadratic Julia sets. Only slightly stronger is the conjecture that, for all α, the complement of the Mandelbrot set (M ) is extremal. This is because if c is a boundary point of M and J is the Julia set for the parameter c, “near c, M looks like J”. It should also be remarked that numerical evidence for conjecture (1) was given by Kraetzer [9]. It seems there is more work to be done here, especially for complex values of t. A very nice idea has been introduced by D. Beliaev [1]. He defines a class of “conformal snowflakes” by using a particular randomization process. He is then able to get rigorous, computer assisted proofs of lower bounds for the pressure function β(t). An amusing aspect of these problems is that it seems likely that the proofs for upper bounds (e.g. β(t) ≤ |t|2 /4, |t| ≤ 2) will be quite different from those for the corresponding lower bounds. On this point, the only thing that seems evident is that one should take unbounded domains with compact boundaries (of log-capacity one). This is of course also the natural setting for polynomial dynamics. 3 Harmonic Measure in Rd Our knowledge of harmonic measure and the f (α) spectrum is much less complete in Rd , d ≥ 3, than for d = 2. What we do know corresponds to experience from CFT: Dimension two is special. As an example of this consider the support of Harmonic measure for a domain Ω. In d = 2 the results of Makarov and Jones–Wolff show that there is always a set of dimension one that supports harmonic measure and (Makarov) if Ω is simply connected one cannot use a smaller dimension. In other words, f (α) < α for α > 1 and for simply connected domains f (1) = 1. This means f (1) is given by ∂Ω = S 1 . Now for d ≥ 3 this might lead one to expect that f (α) < α for α > d − 1, i.e. there is a set of dimension at most d − 1 that supports harmonic measure (the Øksendal Conjecture). However, Tom Wolff [12] has provided counterexamples (of “snowball-type”) showing this is false. On the positive side we know (Bourgain [4]) that harmonic measure is always supported on a set of dimension ≤ d−ε(d). The weakest conjecture that fits known properties for the dimension of the support of harmonic measure is 80 Peter W. Jones dim(ω) ≤ d − ε(d) , (11) where ε(d) → 0 as d → +∞. It is likely that this is relatively easy. As justification for this conjecture one could cite Jones–Makarov [7], where it is shown that in Rd , d − c1 /αd−1 ≤ fu (α) ≤ d − c2 /αd−1 , (12) where c1 and c2 are dimension dependent. One of the reasons that we are presently lacking precise knowledge of dim(ω) in d ≥ 3 is the lack of arguments of complex variables type. A (very) formal attempt to generalize the proof of Jones–Wolff to higher dimensions leads to the conjecture 1 (13) dim(ω) ≤ d − d−1 because (!) if ∆u = 0 in Rd , then |gradient(u)|p is subharmonic for p ≥ (d − 2)/(d − 1). Corresponding to (13) is the statement fu d − 1 1 = d− d−1 d−1 as well as f (α) < α , α>d− 1 . d−1 If one now makes the supposition that fu (α) = d − cd α−d+1 for large α, one comes to a somewhat reasonable looking formula. We first denote by α(ω) the (universal) dimension of harmonic measure, so that by conjecture (13), α(ω) = d − (d − 1)−1 . Algebra then leads one to the possibility 1 fu (α) = d − d−1 α(ω) α d−1 (14) for values α ≥ α(ω). Notice that this has the correct asymptotics as α → +∞, and has fu (α) < α for α > α(ω). If we set d = 2 we also recover the conjecture fu (α) = 2 − 1/α, α ≥ 1. On the other hand we have f 0 (α(ω)) < 1, which seems unreasonable. If we wish also to make f 0 (α(ω)) = 1, thus providing for a smoother phase transition, the next and perhaps more reasonable guess is f (α) = d − 1−d 1 α + 1 − α(ω) , d−1 (15) for α ≥ α(ω). Of course, whether one chooses (14) or (15), one should have f (α) = α for 0 ≤ α ≤ α(ω). Conjecture (15) has the advantage that f 0 exists for all α (and f 0 (α(ω)) = 1). I am unsure whether one can provide any other justification for (14) and (15) beyond the following: They fit all known facts; give us the “correct” conjecture in d = 2; are based on “simple” (here rational) On Scaling Properties of Harmonic Measure 81 functions; and interpolate analytically through dimensions. The interpretation of (14) or (15) for non-integer dimensions is unclear to me. As is the case for d = 2, it seems clear that, whatever the correct answer is for fu (α), the proofs of upper versus lower bounds will be quite different. An examination of the arguments of [6] and [12] makes it clear that near extremals for fu (α) can be built from non-conformal dynamical systems. But in d ≥ 3 there is no natural analogue of Julia sets and/or conformal dynamics, or rather it seems unlikely that this is how one should proceed. Perhaps there are natural dynamical systems in d ≥ 3 that give rise to the appropriate domain, but it would seem that if this were the case, those dynamical systems are not among the ones we now are aware of. It also seems unlikely that simply connected domains could be useful when d ≥ 3. This is due to the fact that the potential theory in d ≥ 3 puts zero capacity on line segments, and consequently fu (α) = fsc (α) actually holds for all α > 0. (As far as I am aware, no one has written down the proof of this rather elementary statement.) References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. D. Beliaev: Ph.D. Thesis, KTH, Stockholm (2005) I. Binder, P.W. Jones: to appear I. Binder, N.G. Makarov, S. Smirnov: Duke Math. J. 117, 343 (2003) J. Bourgain: Inventiones Math. 87, 477 (1987) L. Carleson: Ann. Acad. Sci. Fenn. 10, 113 (1985) L. Carleson, P.W. Jones: Duke Math. J. 66, 169 (1992) P.W. Jones, N.G. Makarov: Annals of Maths. 142, 427 (1995) P.W. Jones, T. Wolff: Acta Math. 161, 131 (1988) P. Kraetzer: Complex Variable Theory Appl. 31, 305 (1996) S. Smirnov: unpublished manuscript N.G. Makarov: Proc. London Math. Soc. 51 (1985) T. Wolff: Counterexamples with harmonic gradients in R3 . In: Essays on Fourier Analysis in Honor of E.M. Stein, no 42 (Princeton Univ. Press, Princeton, NJ 1995) pp 321–384