Transcript ppt-2
Gibbs sampler - simple properties It’s not hard to show that this MC chain is aperiodic. Often is reversible distribution. If in addition the chain is irreducible (which depends on which elements of have non-zero probability), then this MC is a correct MCMC algorithm for simulating random variables with distribution . Systematic Gibbs sampler A commonly used variant of the Gibbs sampler is the MC obtained by systematic cycle trough the vertex set. For instance, for we may decide to update vertex: Back to the Boolean cube In this case: , . In each step we pick and flip the appropriate coordinate of with probability . Stationary distribution for the Boolean Cube We will show that the uniform distribution is reversible (and therefore stationary). Denote by the transition probability from state to . We need to show that: for any two states. Denote by the number of indexes in which and differ: ◦ Case no.1: ◦ Case no.2: ◦ Case no.3: Reminder - Coupling For a given MC with state space and transitions matrix , a coupling for the MC is a MC on the state space with transition probabilities defined by: Reminder (cont.) - The coupling lemma For a coupling , based on a ground MC on . Suppose is a function satisfying the condition: for all and all : then the mixing time for is bounded by . Fast convergence of the Gibbs sampler for the Boolean cube Coupling: we run two Markov chains and simultaneously. We go threw the coordinates systematically. If at some integer time it holds that: We flip both of them with probability Otherwise we flip each one separately with same probability. Fast convergence of the Gibbs sampler for the Boolean cube For any coordinate : after one cycle threw the coordinates: After cycles threw the chain: The event implies that for at least one coordinate : meaning that: Fast convergence of the Gibbs sampler for the Boolean cube Setting and solving for gives: Now, using the coupling lemma, it means that the mixing time is bounded by: Gibbs sampler for q-colorings For a vertex and an assignment of colors to the vertices other than , the conditional distribution of the color at is uniform over the set of all colors that are not attained in at some neighbor of . A Gibbs Sampler for random q-colorings is therefore an -valued MC where at each time , transitions take place as follows: ◦ Pick a vertex uniformly at random. ◦ Pick according to the uniform distribution over the set of colors that are not attained at any neighbor of . ◦ Leave the color unchanged at all other vertices, i.e. let for all vertices except . Gibbs sampler for q-colorings It has as a stationary distribution, the proof is similar to the hard-core model case, based on the fact that it is reversible. Whether or not the chain is irreducible depends on and , and it is a nontrivial question to generally determine this. But for big enough ( is always enough) it is. Fast convergence of Gibbs sampler for q-coloring Theorem [HAG 8.1]: Let be a graph. Let be the number of vertices in , and suppose any vertex has at most neighbors. Suppose furthermore that Then, for any fixed , the number of iterations needed for the systematic sweep Gibbs sampler described above (starting from any fixed -coloring ) to come within total variation distance of the target distribution is at most: Fast convergence of Gibbs sampler for q-coloring Comments: ◦ Can be proved for (rather than ◦ The important bound is: some . ). for Fast convergence of Gibbs sampler for q-coloring Coupling: we run two -valued Markov chains and simultaneously. When a vertex is chosen to be updated, we pick a new color for it uniformly at random among the set of colors that are not attained by any neighbor of . Concretely: Let be an i.i.d sequence of random permutations of , each of them uniformly distributed on the set the of such permutations. Fast convergence of Gibbs sampler for q-coloring (cont.) At each time , the updates of the 2 chains use the permutation and the vertex to be updated is assigned the new value where: in the first chain. In the second chain, we set where: Successful and not-successful updates If at some (random, hopefully not too large) time , we will also have for all First consider the probability that: for a given vertex . We call an update of the 2 chains at time successful if it results in having: Otherwise – we say that the update failed. Fast convergence of Gibbs sampler for q-coloring Notations: The event of having successful update has probability: Respectively: Using that and that we get: (we only used and some algebra). Fast convergence of Gibbs sampler for q-coloring Hence, we have, after steps of the MC-s, that for each vertex : Now, consider updates during the 2nd sweep of the Gibbs sampler, i.e. between times and . For an update at time during the 2nd sweep to fail the configurations and need to differ in at least one neighbor of . Each neighbor has with probability at most By summing over at most neighbors we get that: Fast convergence of Gibbs sampler for q-coloring “discrepancy” means that there exists a neighbor of with . By Repeating the arguments above we get: Hence, after steps of the Markov chains, each vertex has probability at most: By arguing the same for the third sweep: Fast convergence of Gibbs sampler for q-coloring and for any : The event implies that for at least one vertex , meaning that: Fast convergence of Gibbs sampler for q-coloring Summary: By setting and solving for , we get by using the coupling lemma that for: It holds that the mixing time after is bounded by . scans Fast convergence of Gibbs sampler for q-coloring To go from the number of scans to the number of steps of the MC, we have to multiply by , giving that: should be enough. In order to make sure that gets an integer value we should take to be: