Time-inhomogeneous random Markov chains

TL;DR

This paper studies time-inhomogeneous Markov chains with random, time-fluctuating transition probabilities, analyzing their spectral properties and eigenvalue behavior through exact and numerical methods.

Contribution

It provides an exact analysis of the spectral properties of products of random Markov matrices with Dirichlet-distributed columns, including eigenvalue convergence and Lyapunov exponent statistics.

Findings

Eigenvalues decay exponentially to zero for large t

Largest Lyapunov exponent approximates a Gamma distribution

Spectrum concentrates on the real line as t increases

Abstract

We consider Markov chains with random transition probabilities which, moreover, fluctuate randomly with time. We describe such a system by a product of stochastic matrices, $U(t)=M_t\cdots M_1$, with the factors $M_i$ drawn independently from an ensemble of random Markov matrices, whose columns are independent Dirichlet random variables. The statistical properties of the columns of $U(t)$, its largest eigenvalue and its spectrum are obtained exactly for $N=2$ and numerically investigated for general $N$. For large $t$, the columns are Dirichlet-distributed, however the distribution is different from the initial one. As for the spectrum, we find that the eigenvalues converge to zero exponentially fast and investigate the statistics of the largest Lyapunov exponent, which is well approximated the a Gamma distribution. We also observe a concentration of the spectrum on the real line for large $t$.