No label defined (Q4242076)

To calculate the transient solution of a continuous-time Markov process it is necessary to solve Kolmogorov's forward or backward equation \(P'(t)= P(t)Q\), \(P'(t)=QP(t)\), respectively. In these equations \(Q=(q_{ij})_{i, j=0,1, \dots}\) represents the transition rate matrix (or \(q\)-matrix) of the process, which is considered to be measurable, while \(P(t)=(P_{ij}(t))_{i,j=0,1, \dots}\) represents the unknown matrix of transition functions. This paper will generalize a method of J. A. Murphy and M. R. O'Donohoe (1974), who have numerically approximated the solution of the Kolmogorov forward equation for the generalized birth and death process by means of continued fractions. While the above process is represented by a general tridiagonal \(q\)-matrix, this paper suggests an algorithm for \(q\)-matrices of band structure \((n,1)\) that will then be generalized even further to \(q\)-matrices of lower Hessenberg form. Applications involving \(q\)-matrices of these two types include e.g. queueing systems with batch processing or birth, death and catastrophe population processes. The approach is developed using a truncation algorithm for Markov processes. Applying the Laplace transform to the linear differential-difference equation will result in an inhomogeneous linear difference equation of order \(n+1\). The components of the unique solution for this finite system are shown to be quotients of characteristic polynomials of certain subsections of the \(q\)-matrix. These rational functions are analytic in the complex half-plane to the right of the origin and give rise to a partial fraction expansion and Laplace inversion involving exponential functions. It is then shown that the algorithm is more feasible than the approach of calculating the solution of the truncated differential difference equation directly by using the exponential function.