Reversible Imprecise Markov Chains (2507.06374v1)

Abstract: Reversible Markov chains have identical forward and backward path distributions; this symmetry yields tractable stationary behaviour and powers algorithms from MCMC to network random-walk analysis. We extend reversibility to imprecise Markov chains, in which marginal and transition probabilities are specified by credal sets. Adopting the strong-independence interpretation of the Markov property, we reverse every precise chain compatible with a given credal specification and study the resulting ensemble. Because the reversed ensemble generally cannot be encoded by the usual forward model - a credal initial distribution pushed through a credal set of transition matrices - we introduce a symmetric representation based on credal sets of two-step joint distribution or edge-measure matrices. This framework is strictly more expressive, admits a natural reversal operation, and reduces reversibility to simple matrix symmetry. It also lets forward and reverse dynamics be described simultaneously within one closed, convex set, paving the way for efficient algorithms to evaluate expectations of path-dependent functionals. We illustrate the theory with random walks on graphs and outline procedures for computing lower expectations of path-based quantities.