David Croydon (University of Warwick) THE GEOMETRY OF DISCRETE RANDOM STRUCTURES EPSRC SYMPOSIUM WORKSHOP UNIVERSITY OF WARWICK, 18-22 JUNE 2012 critical random trees on random paths and Biased random walks 1 β (β − 1)ne1 Xn = + cβ,dN (0, I)n1/2 + o(n1/2). β + 2d − 1 That is, the random walk that is β times more likely to jump in first-coordinate direction than in any other direction. We have 1 1 Consider the β-biased random walk (Xn)n≥0 on the integer lattice Zd (we will always assume β ≥ 1): BIASED RANDOM WALKS If β > 1, then the walk is directionally transient for P-a.e. environment. Moreover, there exists a βc ∈ (1, ∞) such that: if β < βc, then the biased random walk has positive speed, if β > βc, then the biased random walk has zero speed, [Berger/Gantert/Peres, Sznitman, Fribergh/Hammond]. If β = 1, then the random walk is diffusive for P-a.e. environment [Barlow, Sidoravicius/Sznitman, Biskup/Berger, Mathieu/Piatnitski]. Bond percolation on integer lattice Zd (d ≥ 2), parameter p > pc: e.g. p = 0.53, SUPERCRITICAL PERCOLATION Motivated by this, we consider biased random walk on: • a critical Galton-Watson tree conditioned to survive, • the range of a random walk. ‘Traps along the backbone’ ‘Trapping in branches’ Close to p = pc, physicists [Barma/Dhar] have identified two types of trapping: CRITICAL PERCOLATION How does the biased random walk behave? (Joint with Alexander Fribergh and Takashi Kumagai.) αn ∼ n(α−1)/αℓ(n). converges in distribution to a non-trivial diffusion, where ) −1 αn X⌊tnαn⌋ , t≥0 ( [cf. C.] The unbiased walk (β = 1), when rescaled as T ∗ - family tree of critical Galton-Watson process conditioned to survive. Offspring distribution in domain of attraction of α-stable law, α ∈ (1, 2]. Bias away from root. TRAPPING IN CRITICAL BRANCHES x∼y∈T ∑ c(x, y) ≈ β h(T ), h(T ) has exponential tails ⇒ time in traps has polynomial ⇒ polynomial rate of escape. [cf. Zindy] where c(x, y) := β min{gen(x),gen(y)}. 1 + β −i Expected time to leave trap with base at level i: Linear progress on backbone ⇒ regeneration arguments. Decompose into backbone (solid lines) and traps (dashed lines). SUPERCRITICAL ARGUMENT [BEN AROUS/FRIBERGH/GANTERT/HAMMOND] 1 β > βc := ′ , f (q) log |Xn| log βc → γ := , Pρ-a.s. log n log β Finer subsequential distributional limits established. Also distributional limits in randomly biased case [Ben Arous/Hammond, Hammond]. where q is the extinction probability and f (x) := ExZ , then and the drift satisfies: EZ > 1, EZ 2 < ∞, P(Z = 0) > 0, If offspring distribution Z satisfies: SUPERCRITICAL RESULT [BEN AROUS/FRIBERGH/GANTERT/HAMMOND] ρ1 ) ... ρi1 ρij ρi Tij are unconditioned Galton-Watson trees, offspring dist. Z. k ≥ 1. ... i ρi(Z̃ −1) Ti1 . . . Tij . . . Ti(Z̃i−1) P Z̃ = k = kP(Z = k), ( Backbone ρ = ρ0 Buds Leaves Let T ∗ be a critical Galton-Watson tree (i.e. EZ = 1), with offspring distribution in domain of attraction of α-stable law, α ∈ (1, 2], conditioned to survive: CRITICAL STRUCTURE } P (Nn(i) = 0) ∼ 1 − ρi α , (α − 1)hn i ρi(Z̃ −1) the number of backbone vertices from which big traps emanate up to level n is distributed as Binomial(n, cαn−1 log n). In particular, it grows like cα log n. Then, since ρi1 Nn(i) := # 1 ≤ j ≤ Z̃i − 1 : h(Tij ) ≥ hn . { Define a level n critical height hn := n(log n)−1, and set BIG TRAP LOCATIONS i=0 ti1{Nn(i)≥1}. Since jump process on backbone does not backtrack more than (log n)2, and big traps are separated by at least nε, summands are asymptotically independent. ∆n ≈ n−1 ∑ Let ti be total time spent by random walk in traps emanating from backbone vertex ρi. Can show time spent in small traps (height less than hn) is negligible. Therefore ∆n := inf{m ≥ 0 : Xm = ρn}. Define a hitting time ∆n by setting APPROXIMATION BY I.I.D. SUM ρi ρij xij yij Tij zi } Vi := j ∈ Bi : τxij < τzi , { Bi := j = 1, . . . , Z̃i − 1 : h(Tij ) ≥ hn { } ∗ PρTi (Vi = A) = (#B )−1 1 i . 1 + #Bi #A where xij is ‘trap entrance’ and zi := ρi+1+hδ , then n and Let # BIG TRAPS VISITED j=1,...,Z̃i −1 max ( ) α . (α − 1)x log β . ≥x ≈ (α − 1) log x ) ∼ αqxα−1L(qx) ∼ maxj∈Vi h(Tij ) h(Tij ) ≥ x Pρ (ti ≥ x) ≈ Pρ β It follows that Pρ ( NB. This is different to By first conditioning on Bi and then Vi, it is possible to check that ( ) 1 Pρ max h(Tij ) ≥ x = qxα−1L(qx) ∼ . j∈Vi (α − 1)x TIME SPENT IN BIG TRAP CLUSTER 1 nt ∑ where (m(t))t≥0 is an extremal process. In particular, m can be defined as the maximum process of the Poisson point process with intensity measure x−2dxdt. See [Darling, Kasahara]. L Xi → (m(t))t≥0 , n i=1 t≥0 for any v > 0, and F̄ (u) → 0 as u → ∞. If L(x) := 1/F̄ (x), then F̄ (uv) lim = 1, u→∞ F̄ (u) Let (Xi)∞ i=1 be independent random variables, with distributional tail F̄ (u) = 1 − F (u) = P(Xi > u) satisfying: F̄ (0) = 1, F̄ (u) > 0 for all u > 0, SUMS OF SLOWLY-VARYING VARIABLES (α − 1) log+ ∆nt → (m(t))t≥0 . n log β t≥0 ) (α − 1)n π(X⌊ent⌋) log β ) t≥0 ) −1 → m (t) . t≥0 ( Arguments also demonstrate localisation and extremal aging. ( Moreover, if (π(Xn))n≥0 is the projection of the biased random walk onto the backbone, then ( If β > 1, then (∆n)n≥0 satisfies EXTREMAL CONVERGENCE [C./FRIBERGH/KUMAGAI] ( ) ( ) 1 −1 XF̄ −1(1/nt) → m (t) . t≥0 n t≥0 and c(x, y) = 0 otherwise, then ( ) β −1 , if y = x + 1, τ x ) c(x, y) := ( β+1 1 −1 , if y = x − 1, τ x β+1 Conditional on τ , let X be a Markov chain on Z with jump rates Let τ = (τx)x∈Z be a family of independent and identically distributed strictly positive (and finite) random variables whose distribution has a slowly-varying tail F̄ (u) = P(τ0 > u). RELATED ONE-DIMENSIONAL TRAP MODEL ‘Traps along the backbone’ ‘Trapping in branches’ RECALL CRITICAL TRAPPING MECHANISMS { } For P-a.e. random walk path, the graph G is infinite, connected and clearly has bounded degree. E(G) := {Sn, Sn+1} : n ∈ Z . and edge set V (G) := {Sn : n ∈ Z} , Let S = (Sn)n∈Z be the two-sided simple random walk on Zd starting from 0, built on an underlying probability space with probability measure P. Define the range of the random walk S to be the graph G = (V (G), E(G)) with vertex set RANGE OF A RANDOM WALK NB. Result does not hold in d = 3, 4 [C., Shiraishi]. tκ2 (d) (d) under P, converges as n → ∞ to the law of (WB ) −1/4 , n X⌊tn⌋ t≥0 ( )t≥0. under PG 0 , converges as n → ∞ to the law of (Btκ1(d))t≥0. Furthermore, the law of ) −1/2 n sgn(X⌊tn⌋)(dG (0, X⌊tn⌋) , t≥0 ( Let d ≥ 5. For P-a.e. realisation of G, the law of UNBIASED RANDOM WALK ON G [C.] Gn−1 Gn Cn := STn [Bolthausen/Sznitman/Zeitouni] applied this kind of decomposition of the path of S to deduce the diffusivity of a random walk in a particular high-dimensional random environment. Gn−2 Under P̂ := P(·|0 ∈ T ), G is made up of a string of stationary ergodic finite graphs. which will be denoted . . . T−2 < T−1 < T0 ≤ 0 < T1 < T2 < . . . . For d ≥ 5, P-a.s., the two-sided process S admits an infinite set of cut-times T := {n : S(−∞,n] ∩ S[n+1,∞) = ∅}, GEOMETRY OF TWO-SIDED RANGE (1) is a stationary, ergodic sequence. −, ω 0, ω +) (ωn n n n∈Z From this, one can check that ± := PG (J ωn 0 m+1 = n ± 1|Jm = n) = , 1 . µ({Cn})RG (Cn, Cn±1) is the first coordinate of e±. For jump chain, we have where e± µe := β max{e− ,e+ } (1) (1) Given β ≥ 1, assign to each edge e = {e+, e−} ∈ E(G) a conductance BIASED RANDOM WALK ON G i=1 log ρi, (1) (1) log β ∼ −Snτ log β, a(n) b(n) c(n) log n which has Brownian scaling ⇒ Sinai’s regime, in which the walker is trapped for long periods in valleys of the potential [Sinai]. Rn = log RG (Cn, Cn+1)−log RG (C0, C1) ∼ −Cn where ρi := ωi−/ωi+. This satisfies Rn := n ∑ A key quantity in understanding the behaviour of a RWRE is the potential function: POTENTIAL OF RWRE L , log β in distribution under P, where L is a random variable taking values in Rd whose distribution can be characterized explicitly. Ln → Lβ := log n for any ε > 0. Moreover, ( ) Xn P − Ln > ε → 0, Fix a bias parameter β > 1 and d ≥ 5. If X = (Xn)n≥0 is the biased random walk on the range G of the two-sided simple random walk S in Zd, then there exists an S-measurable random variable Ln taking values in Rd such that LOCALIZATION RESULT [C.]