Many models in economics, finance, and engineering use systems that randomly switch between different "regimes" or states. Think of an economy that alternates between expansion and recession, where the rules governing how variables evolve are different in each state. A common mathematical framework for this is called a Markov-switching recurrence: at each time step, a hidden state (like "boom" or "bust") flips according to fixed probabilities, and the system updates itself using whichever set of rules corresponds to the current state. A central question is whether such a system settles into a stable long-run pattern, and if so, whether that pattern has well-behaved statistical properties like finite variance, which is needed to compute things like correlations and standard errors.
The subtlety the paper resolves is that there are actually two different kinds of stability, and they can come apart. The first kind, called strict stationarity, is governed by a quantity called the top Lyapunov exponent, which roughly measures the average rate at which the system expands or contracts over time. If this exponent is negative, the system stabilizes in the long run. But stability in this sense does not guarantee that the variance of the stationary distribution is finite. Variance requires a stricter condition, and the paper identifies exactly what that condition is. The answer involves a matrix built from the transition probabilities between regimes and the squared versions of the regime-specific coefficient matrices. The spectral radius of this matrix, meaning its largest eigenvalue in magnitude, must be less than one. When it is, the covariance matrix can be computed by solving a straightforward system of linear equations.
This matters because practitioners routinely report correlations, standard errors, and confidence intervals for Markov-switching models, implicitly assuming finite variances exist. The paper shows this assumption can fail even when the model looks stable by the usual criterion. A researcher might have a model that settles into a long-run distribution but where the variance of that distribution is technically infinite, making those reported statistics meaningless. By giving a clean, checkable condition tied directly to a computable matrix, the paper provides both a warning and a practical tool: you can now verify whether your model's stationary distribution actually supports the second-order statistics you plan to use.