| Popis: |
In many fields, including biology, medicine, physics, chemistry, economy and actuarial science, data can be represented as the time-to-event of a finite state Markov chain model. The distribution of time intervals between successive recorded events is known as a phase-type distribution. We demonstrate that in cases where the eigenvalues of the Markov chain transition matrix are distinct (non-degenerate), the phase-type distribution is multi-exponential. We then pose and solve an inverse problem: given knowledge of the phase-type distribution, can we determine the transition rate parameters of the underlying Markov chain? To tackle this challenge, we initially convert the inverse problem into a computer algebraic task, involving the solution of a system of polynomial equations. These equations are symmetrized with respect to the parameters of the phase-type distribution. For a specific subset of Markov models that we refer to as "solvable," the inverse problem yields a unique solution, up to transformations by finite symmetries. We outline a recursive approach to compute these solutions for specific families of models, regardless of their number of states. Additionally, we use the Thomas decomposition technique to calculate solutions for models possessing 2, 3, or 4 states. Interestingly, we show that models having the same number of states but different transition graphs can yield identical phase-distributions. In such "Rashomon effect" scenarios, time-to-event data permits the formulation of multiple models and interpretations, all of which are consistent with the observed experimental data. To differentiate among these models, we propose additional distinguishing properties beyond just the time until the next event. In order to prove its applicability, we discuss how this method can be used to infer models of transcription regulation from transcriptional bursting data. |
