Markov chains are mathematical models that describe how a system moves between different states over time, where the next state depends only on a limited amount of history. A standard (first order) Markov chain only needs to know the current state to predict the next one. A higher order Markov chain uses several recent states to make that prediction, which makes it more flexible but also more mathematically complex.
The paper focuses on a subtle problem that arises with higher order Markov chains. When you convert such a chain into an equivalent standard chain (a technical step often used for analysis), the resulting standard chain can behave poorly, specifically it may not settle into a single long-run stable pattern. Despite this, the higher order chain itself does converge to a unique long-run distribution over just the current state. The authors give this a precise name, calling it the "marginal stationary distribution," and prove several of its mathematical properties, essentially showing that even when the underlying machinery looks messy, the observable behavior of the system is well-behaved.
The paper also addresses how quickly this convergence happens. Knowing that a system eventually stabilizes is useful, but knowing how long it takes to get close to that stable pattern is often more practically important. The authors derive bounds on the convergence speed and introduce two new measures called "marginal mixing times," which capture how long a higher order chain needs to run before its current-state behavior is essentially indistinguishable from its long-run pattern. These concepts extend familiar tools from standard Markov chain theory into the higher order setting, giving researchers better ways to analyze systems with memory.