Identification of almost invariant aggregates in reversible nearly uncoupled Markov chains (Q1583716)

The paper is devoted to a model identification problem, rising from a concrete application task. The key idea is to synthesize an algorithm for identification of almost invariant componentwise aggregates, via the sign structure of the eigenvectors corresponding to the Perron cluster of eigenvalues. Based on the latest results in the area of reversible uncoupled Markov chains, a particular discriminating sign structure is derived, for the identification of invariant aggregates. A robust identification algorithm is derived. Part of it is transformed into a solution of the graph coloring problem, which is known as NP-complete. As a consequence, heuristics are justified to play a crucial role in the computational solution and algorithm implementation. Two examples are given, to illustrate the performance of the suggested identification algorithm. One of them is simple artificial but the second concerns molecular dynamics and particular detection of the known chemical structure, without explicit use of chemical insight.