Fastest Mixing Markov Chain on a Graph

Abstract

We consider a symmetric random walk on a connected graph, where each edge is labeled with the probability of transition between the two adjacent vertices. The associated Markov chain has a uniform equilibrium distribution; the rate of convergence to this distribution, i.e., the mixing rate of the Markov chain, is determined by the second largest (in magnitude) eigenvalue of the transition matrix. In this paper we address the problem of assigning probabilities to the edges of the graph in such a way as to minimize the second largest magnitude eigenvalue, i.e., the problem of finding the fastest mixing Markov chain on the graph.We show that this problem can be formulated as a convex optimization problem, which can in turn be expressed as a semidefinite program (SDP). This allows us to easily compute the (globally) fastest mixing Markov chain for any graph with a modest number of edges (say, 1000) using standard numerical methods for SDPs. Larger problems can be solved by exploiting various types of symmetry and structure in the problem, and far larger problems (say 100000 edges) can be solved using a subgradient method we describe. We compare the fastest mixing Markov chain to those obtained using two commonly used heuristics: the maximum-degree method, and the Metropolis-Hastings algorithm. For many of the examples considered, the fastest mixing Markov chain is substantially faster than those obtained using these heuristic methods.We derive the Lagrange dual of the fastest mixing Markov chain problem, which gives a sophisticated method for obtaining (arbitrarily good) bounds on the optimal mixing rate, as well the optimality conditions. Finally, we describe various extensions of the method, including a solution of the problem of finding the fastest mixing reversible Markov chain, on a fixed graph, with a given equilibrium distribution.

A semismooth Newton-CG based dual PPA for matrix spectral norm approximation problems

“…, A t , the matrix Chebyshev polynomial approximation problem is actually a special case of the MNA problem (1). In contrast to the unconstrained example (2), some other problems may have prescribed linear constraints, for example, the fastest mixing Markov chain (FMMC) problem studied in [2,3]. Let G = (N , E) be an undirected connected graph with n nodes.…”

A local average consensus algorithm for wireless sensor networks

“…if x(k f ) =m then x(k) =m ∀k > k f . Equation (3) says that the average of all the estimates is conserved at each iteration which guarantees the final consensus value to be the average of the initial measurements. Finally, equation (4) is the contraction condition, i.e.…”

“…The convergence speed of these algorithms is related to the mixing time of Markovian chains, it is possible to show that the "error" ||x(k) −m|| 2 can be upper bounded by cλ k 2 , where λ 2 is the second largest eigenvalue of the weight matrix W and c is a constant [3], [13]. In [2], [3], the authors address the problem of selecting the optimal weights and formulate a semidefinite program to determine the matrix W with the smallest value of λ 2 . If the network topology is not known a priory (or changes in time), the implementation of this optimization procedure requires a single sensor to discover the topology of the whole network and perform the calculations.…”

“…In fact reducing the convergence time leads to a smaller number of transmissions among the sensors and then to lower energetic costs for each sensor. It is known that the set of weights that minimize the convergence time can be determined through a Semi Definite Program ( [2], [3]) that requires the knowledge of the whole topology. This approach may be unfeasible for large and dynamic sensor networks, because of sensor computational constraints and of the communication overhead due to the need to acquire the topology after each change.…”

Raptor packets: A packet-centric approach to distributed raptor code design

“…Two simple and popular heuristic methods for designing P that satisfy the above conditions are maximum-degree (MD) and Metropolis-Hastings (MH) algorithm [9] [10].…”