Stochastic Loewner evolution with branching and the Dyson superprocess

Transcription

1 Stochastic Loewner evolution with branching and the Dyson superprocess Govind Menon (Brown University) Vivian Olsiewski Healey (Brown/University of Chicago) This talk is based on Vivian's Ph.D thesis (May 2017) as well as more recent joint work. The help of Steffen Rohde (U. of Washington) is gratefully acknowledged. Random Matrix Theory, IAS/PCMI summer school, NSF , NSF

2 Preview Galton-Watson trees The Loewner equation

3 Preview Galton-Watson trees Describe the genealogy of birthdeath processes. The Loewner equation

4 Preview Galton-Watson trees Describe the genealogy of birthdeath processes. The Loewner equation

5 Preview Galton-Watson trees Describe the genealogy of birthdeath processes. The Loewner equation

6 Preview Galton-Watson trees Describe the genealogy of birthdeath processes. The Loewner equation

7 Preview Galton-Watson trees Describe the genealogy of birthdeath processes. The Loewner equation Continuum Random Tree The CRT is a random metric space that is a universal scaling limit of these trees, conditioned to be large.

8 Preview Galton-Watson trees Describe the genealogy of birthdeath processes. The Loewner equation Gives a bijection between (certain) growth processes in the upper half-plane and (certain) evolving measures on the real line. Continuum Random Tree The CRT is a random metric space that is a universal scaling limit of these trees, conditioned to be large.

9 Preview Galton-Watson trees Describe the genealogy of birthdeath processes. The Loewner equation Gives a bijection between (certain) growth processes in the upper half-plane and (certain) evolving measures on the real line. Continuum Random Tree The CRT is a random metric space that is a universal scaling limit of these trees, conditioned to be large.

10 Preview Galton-Watson trees Describe the genealogy of birthdeath processes. The Loewner equation Gives a bijection between (certain) growth processes in the upper half-plane and (certain) evolving measures on the real line. Continuum Random Tree The CRT is a random metric space that is a universal scaling limit of these trees, conditioned to be large.

11 Preview Galton-Watson trees Describe the genealogy of birthdeath processes. The Loewner equation Gives a bijection between (certain) growth processes in the upper half-plane and (certain) evolving measures on the real line. Continuum Random Tree The CRT is a random metric space that is a universal scaling limit of these trees, conditioned to be large.

12 The main questions Q1. Can we use the Loewner equation to construct natural graph embeddings of Galton-Watson trees in the upper half plane? Q2. Can we construct a graph embedding of the CRT as a scaling limit of these embeddings of finite Galton-Watson trees? Q3. What does this construction say about "true trees" (conformally balanced embeddings) and the Brownian map? At first sight, there is no random matrix theory here. But the above problems are closely related to map enumeration.

13 Outline (a) A brief introduction to the CRT. (b) Loewner evolution with branching (tree embedding). (c) Scaling limits: the SPDE in the case of a Feller diffusion. (d) Some remark on true trees and the Brownian map. (a) is (necessary) background. Basic reference: Le Gall (1999).

14 The continuum random tree Definition: A plane tree is a rooted combinatorial tree for which the edges are assigned a cyclic order about each vertex.

15 The continuum random tree Definition: A plane tree is a rooted combinatorial tree for which the edges are assigned a cyclic order about each vertex. Each plane tree has a corresponding contour function.

16 The continuum random tree Definition: A plane tree is a rooted combinatorial tree for which the edges are assigned a cyclic order about each vertex. Each plane tree has a corresponding contour function.

17 The continuum random tree Def: A real tree is a pointed compact metric space with the tree property.

18 The continuum random tree Def: A real tree is a pointed compact metric space with the tree property. Construction: Let f :[0, 1] R 0 be continuous, and f (0) = f (1) = 0. Consider and d(u, v) =f(u)+f(v) 2 min f(s) s [u,v] u f v d(u, v) =0.

19 The continuum random tree Def: A real tree is a pointed compact metric space with the tree property. Construction: Let f :[0, 1] R 0 be continuous, and f (0) = f (1) = 0. Consider and d(u, v) =f(u)+f(v) 2 min f(s) s [u,v] u f v d(u, v) =0. Def: T f =[0, 1]/ f is the real tree coded by f.

20 The continuum random tree Definition: The continuum random tree (CRT) is the random real tree coded by the normalized Brownian excursion.

21 The continuum random tree Definition: The continuum random tree (CRT) is the random real tree coded by the normalized Brownian excursion. As the uniform distribution on rescaled Dyck paths of length 2n converges to, the uniform distribution on plane trees conditioned to have n edges converges to the CRT. 1 2n C n(2nt) 0 t 1 n (d) ( t ) 0 t 1 The CRT was introduced by Aldous ( ).

22 Chordal Loewner evolution Let γ : (0, T] H be a simple curve with γ(0) R. (t) g t ) U(t)

23 Chordal Loewner evolution Let γ : (0, T] H be a simple curve with γ(0) R. (t) g t ) The Riemann mapping theorem implies that for each t there is a unique conformal mapping g t such that 1) g t : H \ γ((0, t]) H U(t)

24 Chordal Loewner evolution Let γ : (0, T] H be a simple curve with γ(0) R. (t) g t ) The Riemann mapping theorem implies that for each t there is a unique conformal mapping g t such that 1) g t : H \ γ((0, t]) H 2) g t (z) =z + b(t) z + O 1 z 2, z. U(t)

25 Chordal Loewner evolution Let γ : (0, T] H be a simple curve with γ(0) R. (t) g t ) The Riemann mapping theorem implies that for each t there is a unique conformal mapping g t such that 1) g t : H \ γ((0, t]) H 2) g t (z) =z + b(t) z Loewner (1920s): + O 1 z 2, z. U(t) g t satisfies the initial value problem ġ t (z) = ḃ(t) g t (z) U(t), g 0(z) =z.

26 Chordal Loewner evolution General version: Let g t (z) denote the solution to the initial value problem ġ t (z) = R μ t (du) g t (z) u, g 0(z) =z. Let H t = {z H} for which g t (z) H is well defined. Then g t is the unique conformal map from H t onto H with the hydrodynamic normalization. The hull is the set H\H t = K t. Idea: The measure is supported on points that are escaping H. Conditions: {µ t } t 0 is a family of nonnegative Borel measures on R that is right continuous with left limits in the weak topology. For each t, µ s (R) and supp (µ s ) are each uniformly bounded for 0 s t.

27 Examples 1) $ t = δ U(t) N 2) μ t = produces the multislit equation (Schleissinger 13): i=1 δ Ui (t) ġ t (z) = N i=1 1 g t (z) U i (t). 3) μ t = δ κbt generates SLE κ. Question: Which measures generate embeddings of trees?

28 What's new The general form of Loewner evolution is rarely used. However, it is central to our approach. The one line summary of our work is the following conjecture: Graph embeddings of continuum trees are generated by Loewner evolution when the driving measure is a suitable superprocess. The simplest example in this class is what we call the Dyson superprocess. It is the free probability analogue of the Dawson-Watanabe superprocess.

29 Loewner evolution driven by Dyson BM with branching. (Simulation courtesy of Vivian Healey and Brent Werness.)

30 SPDE for the Dyson superprocess µ t (dx) = (x, t) dx, x (, ),t>0, t + x ( H )= Ẇ, where Ẇ is space-time white noise and is the Hilbert transform H (Hµ t )(x) = p.v. 1 x s µ t(ds). The SPDE is formal, but convenient. The measure valued process is actually defined through a martingale problem.

31 Comparison with the Dawson-Watanabe superprocess The Dawson-Watanabe superprocess is the scaling limit of branching Brownian motion when the discrete branching processes converges to the Feller diffusion. The spatial motion of each particle is independent. t = Ẇ, x R d,t>0. The Dyson superprocess is the free probability version of this SPDE: t + x ( H )= Ẇ, x R,t>0. Unlike Dawson-Watanabe, this is a superprocess of interacting particles.

32 The SPDE and stochastic Loewner evolution Consider the Cauchy-Stieltjes transform f(z,t) = 1 z s µ t(ds), z H. Define the Gaussian analytic function h(z,t),z H with covariance kernel E h(z,t) h(w, t ) = (t t ) z 1 s w 1 s µ t(ds).

33 The Dyson superprocess and Loewner evolution Let Ḃ denote white-noise (in time alone). Then (formally) tf + f z f = hḃ, z H,t>0. This SPDE may be solved by the method of characteristics dz dt = f(z,t), df = h(z,t) db, z H. The stochastic Loewner evolution is given by the subordination formula ġ t (z) =f(g t (z),t), g 0 (z) =z, z H.

34 Absolute continuity w.r.t. Lebesgue measure µ t (dx)? = (x, t) dx Dawson-Watanabe superprocess: µ t d =1 is absolutely continuous (Konno-Shiga, Reimers, 1988). µ t d 2 is singular (Perkins, 1988). Dyson superprocess: we don't know yet. The basic regularity estimates for free convolution with a semicircular law were obtained by Biane (1997). Unlike Dawson-Watanabe we would like µ t to be singular with respect to Lebesgue measure. If a density exists, then the hull cannot be a tree, so this is a crucial property.

35 Tree Embedding Question: In the deterministic setting, what conditions guarantee that the hull is a tree? Fundamental step: What conditions on the driving measure guarantee that the generated hull is a union of two simple curves in H that meet at a single point on R at nontrivial angles (not 0 or #)?

36 Tree Embedding: (α, β)-approach Setup: Let U 1,..., U n be n continuous functions U i : [0,T] R that are mutually nonintersecting U i (t) < U i+1 (t) for all i = 1,..., n and all t [0,T], except for U j (0) = U j+1 (0). Let $ t be the discrete measure μ t = c n i=1 δ Ui (t)

37 Tree Embedding: (α, β)-approach Setup: Let U 1,..., U n be n continuous functions U i : [0,T] R that are mutually nonintersecting U i (t) < U i+1 (t) for all i = 1,..., n and all t [0,T], except for U j (0) = U j+1 (0). Let $ t be the discrete measure μ t = c Definition: Let α, β (0, #) such that α + β < '. n i=1 δ Ui (t)

38 Tree Embedding: (α, β)-approach Setup: Let U 1,..., U n be n continuous functions U i : [0,T] R that are mutually nonintersecting U i (t) < U i+1 (t) for all i = 1,..., n and all t [0,T], except for U j (0) = U j+1 (0). Let $ t be the discrete measure μ t = c n i=1 δ Ui (t) Definition: Let α, β (0, #) such that α + β < '. We say that K t approaches R at U j (0) in (α, β)-direction if for each ε > 0 there is s = s ε > 0 such that there are exactly two connected components of K s that have U j (0) as a boundary point, and K j s ε {z H : π β ε < arg(z U j (0)) < π β + ε}, K j+1 s ε {z H : α ε < arg(z U j (0)) < α + ε}. (Motivated by Schleissinger 12.)

39 Tree Embedding: (α, β)-approach β α

40 Tree Embedding: (α, β)-approach Theorem (Healey): In the setting above, the hulls K t approach R in (α, β)-direction at U j (0) if U j (t) U j (0) lim t 0 t U j+1 (t) U j+1 (0) lim t 0 t = φ 1 (α, β) φ 2 (α, β) = φ 1 (α, β)+φ 2 (α, β), where φ 1 (α, β) and φ 2 (α, β) are explicitly computable functions.

41 Tree Embedding: (α, β)-approach Balanced case: If 0 < α = β < '/2, then φ 1 and φ 2 simplify to φ 1 (α, α) =0 and φ 2 (α, α) = 2c π 2α. α Intuitively: Loewner scaling If $ t generates hulls K t, then ρ$ t/ρ 2 generates the hulls ρk t. So we expect to see t whenever a hull is preserved under dilation.

42 A driving measure for any tree Let T = {ν, h(ν)} be a marked plane tree. (Think of h(ν) as the time of death of ν.) Let $ t be indexed by the elements of T alive at t: μ t = c ν Δ t T δ Uν (t).

43 Tree Embedding Let T = {ν, h(ν)} be a marked plane tree. (Think of h(ν) as the time of death of ν.) Let $ t be indexed by the elements of T alive at t: μ t = c δ Uν (t). ν Δ t T On time intervals without branching, chose the U ν to evolve according to U ν (t) = ν=η Δ t T c 1 U ν (t) U η (t).

44 The driving measure Let T = {ν, h(ν)} be a marked plane tree. (Think of h(ν) as the time of death of ν.) Let $ t be indexed by the elements of T alive at t: μ t = c δ Uν (t). ν Δ t T On time intervals without branching, chose the U ν to evolve according to U ν (t) = ν=η Δ t T c 1 U ν (t) U η (t). Theorem (Healey) : If T is a binary tree such that h ν h η, then for each 0 s max{h(ν)}, the hull K s generated at time s by the Loewner equation driven by $ t is a graph embedding of the subtree T s = {ν T : h(p(ν)) < s} in H, with the image of the root on the real line, and K s K s if s < s.

45 Tree Embedding Proof (idea): The proof relies on ODE results about the particle system U ν (t) = ν=η Δ t T c 1 U ν (t) U η (t). Extend the solution backward to the initial condition U j (0) = U j+1 (0). Show that the solution generates curves away from t = 0. Show that the generated hull approaches R in (α, α)-direction for α = π 2 + c 1 2c.

46 Example: Galton-Watson Trees with deterministic repulsion If θ is distributed as a binary Galton-Watson tree with exponential lifetimes, then the theorem guarantees that the Loewner equation driven by $ t generates a graph embedding of θ with probability one. A sample of a binary Galton- Watson tree with exponential lifetimes. (Simulation courtesy of Brent Werness.)

47 Embedding the CRT? Question 2: Let {θ k } be a sequence of random trees that (when appropriately rescaled) converges in distribution to the CRT when θ k is conditioned on having k edges. Does the law of the generated hulls converge to a scaling limit?

48 Embedding the CRT? Question 2: Let {θ k } be a sequence of random trees that (when appropriately rescaled) converges in distribution to the CRT when θ k is conditioned on having k edges. Does the law of the generated hulls converge to a scaling limit? First step: Find the scaling limit of the corresponding sequence of random driving measures.

49 Choosing a Sequence of Measures Let {T k } be a sequence of random trees, and let {c k } and {c k 1} be two sequences with elements in R +. For each k, define μ k t = ck ν Δ t T k δ Uν (t),

50 Choosing a Sequence of Measures Let {T k } be a sequence of random trees, and let {c k } and {c k 1} be two sequences with elements in R +. For each k, define μ k t = ck ν Δ t T k δ Uν (t), where the U ν (t) evolve according to U ν (t) = ν=η Δ t T k ck 1 U ν (t) U η (t).

51 Choosing a Sequence of Measures Let {T k } be a sequence of random trees, and let {c k } and {c k 1} be two sequences with elements in R +. For each k, define μ k t = ck ν Δ t T k δ Uν (t), where the U ν (t) evolve according to U ν (t) = ν=η Δ t T k ck 1 U ν (t) U η (t). Same setting as tree embedding theorem. How do we choose the trees {T k }?

52 Galton-Watson trees to the CRT Theorem (Aldous): If θ k is distributed as a critical binary Galton-Watson 1 tree with exponential lifetimes of mean 2 k, conditioned to have k edges, then θ k converges in distribution to the CRT as k.

53 Galton-Watson trees to the CRT Theorem (Aldous): If θ k is distributed as a critical binary Galton-Watson 1 tree with exponential lifetimes of mean 2 k, conditioned to have k edges, then θ k converges in distribution to the CRT as k. Question 2a: Can we find a scaling limit of {μ k t } defined by μ k t = ck ν Δ t θ k δ Uν (t), where the U ν (t) evolve according to U ν (t) = ν=η Δ t θ k ck 1 U ν (t) U η (t), if for each k, θ k is distributed as a critical binary Galton-Watson tree 1 with exponential lifetimes of mean, conditioned to have k edges? 2 k

54 The Scaling Limit For each k, the driving measure μ k t is really a measure-valued process defined for t [0, ). By convergence of driving measures we mean convergence in the Skorokhod space D Mf (R)[0, ) of functions from [0, ) to M f (R) with càdlàg paths (right continuous with left limits). Called superprocesses. How to prove convergence of superprocesses? Tightness Prokhorov: tight relatively compact subset of D Mf (R)[0, ). In particular, there is at least one limit point. Uniqueness of the limit point. (Conv. of finite dimensional marginals.)

55 Scaling limits: tightness Theorem (Healey, M.): (a) For each k, let θ k be distributed as a critical 1 binary Galton-Watson tree with exponential lifetimes of mean 2 k, conditioned to have k edges, and let {$ k } be the corresponding sequence of measures. If the scaling constants are c k = c k 1 = 1 k then the sequence {$ k } is tight in D Mf (ˆR) [0, ).

56 Scaling limits: tightness Theorem (Healey, M.): (a) For each k, let θ k be distributed as a critical 1 binary Galton-Watson tree with exponential lifetimes of mean 2 k, conditioned to have k edges, and let {$ k } be the corresponding sequence of measures. If the scaling constants are c k = c k 1 = 1 k then the sequence {$ k } is tight in D Mf (ˆR) [0, ).

57 Scaling limits: tightness Theorem (Healey, M. ): For each k, let θ k be distributed as a critical 1 binary Galton-Watson tree with exponential lifetimes of mean 2 k, conditioned to have k edges, and let {$ k } be the corresponding sequence of measures. If the scaling constants are c k = c k 1 = 1 k then the sequence {$ k } is tight in D Mf (ˆR) [0, ). Why these constants? Choose c k = c k 1, since the ratio c k 1/c k determines the branching angle. c k = 1/ k is the rescaling for which the total population process of θ k converges to L t, the local time at level t of the normalized Brownian excursion.

58 The CRT scaling limit Contour function for critical binary GW trees time

59 The CRT scaling limit Contour function for critical binary GW trees time

60 The CRT scaling limit Contour function for critical binary GW trees time Galton-Watson process

61 The Scaling limit Contour function for critical binary GW trees time Galton-Watson process Normalized Brownian excursion ( ) 0 t 1 Local time at level t of normalized Brownian excursion: L t

62 The Scaling limit Theorem (Pitman): If for each k, N k is the total population process of θ k, then N k t L t, k as k, in the sense of convergence in distribution of random variables in D Mf (R)[0, ).

63 Conditioned v. unconditioned limits Theorem (Pitman): If for each k, N k is the total population process of θ k, then N k t L t, k as k, in the sense of convergence in distribution of random variables in D Mf (R)[0, ). Standard unconditioned result: For each k, let N k t be a discrete critical Galton-Watson process (all lifetimes of length one) whose offspring distribution has finite variance. Then N k kt X t, k as k, in the sense of convergence in distribution of random variables in D Mf (R)[0, ), where X t is the Feller diffusion. (Need N k 0/k x 0 > 0, since the Feller diffusion is absorbing at 0.)

64 The scaling limit: characterization Theorem (Healey, M. ): In the unconditioned case, each subsequential limit solves the martingale problem for the Dyson superprocess: µ t, = µ 0, + t 0 R R (x) x y (y) µ s (dx)µ s (dy)+m t ( ), where the local martingale M has quadratic variation [M( )] t = 0 t µ s, 2 ds. Remark: (a) Don't know yet if the solution to the martingale problem is unique. (b) Similar result for the conditioned case (Pitman, Serlet).

65 Conformally balanced trees in C Joel Barnes (Ph.D, 2014, U. Washington) A planar tree is conformally balanced if (a) each edge has equal harmonic measure from infinity (b) edge subsets have the same measure from either side. Balanced trees are in 1-1 correspondence with Shabat polynomials (Bishop, Biane). However, these polynomials are poorly understood.

66 Conformally balanced trees in C ( ) D C B E A F H G Fig. 1. A noncrossing partition B D C E A B H F G C A Fig. 3. The domain B C A Fig. 2. The planar tree Biane (2009): builds a conformal mapping of the exterior domain by welding edges in pairs. End result is a conformal mapping, continuous onto the boundary, that gives exactly the non-crossing partition.

67 The CVS bijection and the Brownian map t t 111 t t 112 t t 12 2 t t t 2 3 t 3 t 2 t 1 t 2 t t 1? 1 t Figure 2. The Cori-Vauquelin-Schae er bijection. On the left side, a well-labeled tree (the framed numbers are the labels assigned to the vertices). On the right side, the edges of the associated quadrangulation Q appear in thick curves. Image from Le Gall (ICM, 2014).

68 Conformal CVS

69 Conformal CVS

70 Given the conformal map, each labeling gives a nested family of geodesics in the upper half plane with endpoints on its preimage. The image of these geodesics under the conformal map is a quadrangulation of the upper half plane.