A Non-convex Optimization Approach of Searching Algebraic Degree Phase-type Representations for General Phase-type Distributions
Yujie Liu, Dacheng Yao, Hanqin Zhang
TL;DR
This work addresses the problem of when a PH distribution has a minimal representation whose order matches the algebraic degree of its Laplace-Stieltjes transform ${\cal L}(s)=\frac{p(s)}{q(s)}$, by deriving an iff condition encoded as a nonconvex quadratic optimization. The authors develop an alternating minimization algorithm that alternates between solving convex subproblems for matrices $P$ and $A$, with a proof of convergence to a critical point and a clear translation from the transfer-function form to a PH representation via the Jordan form ${\cal J}$ and residue vector $\beta$. They show that the same approach extends to discrete-time PH representations through a rigorous CT↔DT equivalence, yielding a unified method for minimal PH representations across time scales. Numerical experiments demonstrate the method’s effectiveness when a minimal algebraic-degree representation exists (zero objective) and its limitations when such a representation does not exist or when poles are complex, while highlighting robustness and sensitivity to initialization. Overall, the paper provides a practical, convergent framework for identifying minimal PH representations and connects CT and DT cases in a cohesive theory.
Abstract
For a continuous-time phase-type distribution, starting with its Laplace-Stieltjes transform, we obtain a necessary and sufficient condition for its minimal phase-type representation to have the same order as the algebraic degree of the Laplace-Stieltjes transform. To facilitate finding this minimal representation, we transform this condition equivalently into a quadratic nonconvex optimization problem, which can be effectively addressed using an alternating minimization algorithm. The algorithm convergence is also proved. Moreover, the method we develop for the continuous-time phase-type distributions can be directly used to the discrete-time phase-type distributions after establishing an equivalence between the minimal representation problems for continuous-time and discrete-times phase-type distributions.
