Presentation is loading. Please wait.

Presentation is loading. Please wait.

Phase Transitions In Reconstruction Yuval Peres, U.C. Berkeley

Similar presentations


Presentation on theme: "Phase Transitions In Reconstruction Yuval Peres, U.C. Berkeley"— Presentation transcript:

1 Phase Transitions In Reconstruction Yuval Peres, U.C. Berkeley
11/13/2018

2 Markov Chains on trees 2 generalizations of Markov property to infinite trees: Markov random field. (~ two sided) Markov chain on the tree (~ one sided) + A B + + + - + + + - - + - + + XA and XB are cond. ind. given X (B finite). ’[] is given by [(0)] £ {e = u ! v} Me(u),(v) Definitions extend to infinite trees. For finite trees definitions are equivalent. 11/13/2018

3 The reconstruction problem
Fix an infinite rooted tree T, an initial distribution  and Markov chains Me on the edges. Let n = values of the chain at level n. We want to know if n is “correlated” with 0 as n ! 1 If T=1/2-line and Me=M for all e where M is ergodic, then the correlation between 0 and n decays exponentially in n. For general trees, delicate balance between exponential decay of correlation and exponential growth of tree. Remark: The problem is different from the standard setting in statistical physics, since we cannot set the boundary conditions. 0 0 n n 11/13/2018

4 Formal Definition: Reconstruction problem
T = an infinite rooted tree. Ln = { v : d(v,0) = n }. ns = projection/marginals of ’s on Ln (the measure conditioned on (0) = s) : ns[] =  { ’s[] :  | Ln = } The reconstruction problem is solvable if one of the following equivalent conditions hold: 9 i,j: limn ! 1 |in - jn|TV := limn ! 1  |in() - jn()|> 0.   has a non trivial tail   is non-extremal. + Which properties of T and M determine solvability? + + + - + + 11/13/2018 L3 + - - + - + +

5 Statistical physics on trees
The Ising model on the binary tree can be defined: Set σr, the root spin, to be +/- with probability ½. For all pairs of (parent, child) = (v, w), set σw = σv, with probability , otherwise σw = +/- with probability ½. This is exactly the CFN model. Studied in statistical physics [Spitzer 75, Higuchi 77,Chayes-Chayes-Sethna-Thouless-86,Bleher-Ruiz-Zagrebnov 95, Evans-Kenyon-Peres-Schulman 2000, Ioffe 99, M 98, Haggstrom-M 2000, Kenyon-M-Peres 2001, Martinelli-Sinclair Weitz 2003, Martin 2003] + + + + - + + 11/13/2018 - + - + - + +

6 Recursive Reconstruction
= Binary symmetric channel (BSC) = Ising model (no external field) T = 3-ary regular tree with Me = M for all edges. Consider the recursive majority function. Let pn := P[n-fold rec-maj(n) = 0] . Let (p) = (1-) p +  (1-p) and g(p) = 3(p)+32(p)(1-(p)) p0 = 1 and pn+1 = g(pn) ) pn ! ½ if and only if (1-2) > 2/3. ) reconstruction problem is solvable if  < 1/6. Von-Neumann (56) for reliable noisy-computation. Later: Evans-Schulmann93, Steel94, Mossel98,00. 11/13/2018

7 Spectral Reconstruction
Let M be the Ising (BSC) model on a b-ary tree T. Is 0 correlated with f(n) = sign({(v) : v 2 Ln}) ? Theorem (Higuchi 77): limn P[0 = f(n)] > ½ if b(1-2)2 > 1. ) reconstruction for ternary tree if  < ½ - (1/3)1/2. Let M be any chain and T the b-ary tree Let  be the 2nd eigenvalue of M in absolute value. Claim[Kesten-Stigum66] b ||2 > 1 ) reconstruction. Mossel-Peres03: b ||2 =1 is the threshold for reconstruction using the census. Janson-Mossel04: This is also the threshold for “robust” reconstruction (where level n is perturbed). Can replace b by “branching number” for general trees. 11/13/2018

8 Non-reconstruction - Coupling down
Copying rule. For i =+,-: P[i ! i] =  = 1 – 2  P[i ! Uniform] = 1– = 2  Continuing down the tree, non-coupled elements form a branching process with parameter . + / - + / - = = + / - = = = = = = = = = = If 2  · 1, branching process dies ) coupling. ) for  ¸ ¼ no reconstruction (this is not tight!) The threshold for reconstruction is only known for Ising (BSC) model and is given by 22 = 1. Seen: 2 2 > 1 ) reconstruction (spectral argument) 11/13/2018

9 Reconstruction for the CFN model
Thm: The reconstruction problem for the Ising model on the (b+1)-regular tree is solvable if and only if b 2 > 1. “Easy direction” [Higuchi 77]: prove that a certain reconstruction algorithm works when b 2 > 1. Higuchi argument extends to general chains and general trees. Will also show an argument from [M98] useful for phylogeny. “Hard direction” [¸ 95]: Non-reconstruction? 6 different proofs! All involve a magic. None extends to other markov models. 11/13/2018

10 Ising Model on Binary Trees
low interm. high bias bias no bias bias no bias “typical” boundary 2 2 > 1 “typical” boundary 2  > 1 22 < 1 2  < 1 Unique Gibbs measure The transition at 2 2 = 1 was proved by: Bleher-Ruiz-Zagrebnov95,Evans-Kenyon-Peres-Schulman2000,Ioffe99, Kenyon-Mossel-Peres-2001,Martinelli-Sinclair-Weitz2004. 11/13/2018

11 Reconstruction for Markov models
So the threshold b  2 = 1 is important. But [M-2000] it is not the threshold for the reconstruction problem. Not even for 2 £ 2 markov chains, Or symmetric markov chains on q symbols. Moreover, there exists a markov chain M s.t.  = 0, but the reconstruction problem is solvable for some b. Open: What is the threshold for q=3 Potts on binary tree? 11/13/2018

12 Reconstruction for other Markov models
Leaving the Ising model … Reconstruction for other models is more interesting. The “natural” bound for reconstruction is b ||2 > 1, where  is the second eigen-value of M (in absolute value). In “census reconstruction” we reconstruct from Yn = (Yn(i))i 2 A, where Yn(i) = # of times color i appears at the n’th level. Theorem [M-Peres 2002]: The count reconstruction problem is solvable if b ||2 > 1 and unsolvable if b ||2 < 1. 11/13/2018

13 Reconstruction for other Markov models
Theorem [M-Peres 2002]: The census reconstruction problem is solvable if b ||2 > 1 and unsolvable if b ||2 < 1. Proof uses [Kesten-Stigum-66] theorem. 11/13/2018

14 Reconstruction for Markov models
In “robust reconstruction”, instead of the n’th level, n, we are given n, where for each v at level n n(v) = n(v) with probability , n(v) = an independent color from a distribution  with probability 1 - . Similar to “robust phase transition” Pemantle-Steif 99. Easy: if b 2 > 1 then robust reconstruction is solvable for all  > 0. Theorem [Janson-M 2003]: If b 2 < 1, then for  > 0 small, robust reconstruction is unsolvable. Same is true with br(T) instead of b. 11/13/2018

15 Glauber dynamics for Ising models on trees
Consider the following dynamics on +/- configurations on the tree. Start with a configuration σ. At rate 1: Pick a vertex v uniformly at random, and update σ(v) according to the conditional probability given {σ(w): w ~ v}. Converges to Ising distribution; but how fast? Note: It is trivial to sample from the distribution we are interested in. Note: The same process can be defined on any Markov Random Field. 11/13/2018

16 Ising Model on Binary Trees
low interm. high bias bias no bias bias no bias “typical” boundary 2 2 > 1 “typical” boundary 2  > 1 22 < 1 2  < 1 Unique Gibbs measure 2 = (n1 + 2 log2  ) 2 = O(1) 11/13/2018

17 Temporal mixing spatial mixing
Thm [BKMP]: Let G be an ∞-graph of bounded degree; (Gr) balls of radius r around o. Consider a particle system (e.g. Ising; Coloring) on G s.t. Glauber dynamics on Gr satisfy τ2 = O(1). Then, for any finite set A, if f is a function of (σv)v 2 A and g a function of (σv)|v| > r, then Cov(f,g) < exp(-Ω(r))Var1/2(f) Var1/2(g). Open problem: spatial ) temporal? MSW: Yes for trees. g r A f 11/13/2018

18 Phylogeny Here the tree is unknown.
Given a sequence of collections of random variables at the leaves (“species”). Collections are i.i.d.! Want to reconstruct the tree (un-rooted). 11/13/2018

19 Phylogeny Algorithmically “hard”. 11/13/2018


Download ppt "Phase Transitions In Reconstruction Yuval Peres, U.C. Berkeley"

Similar presentations


Ads by Google