Computing first passage times for Markov-modulated fluid models using numerical PDE problem solvers

A popular method to compute first-passage probabilities in continuous-time Markov chains is by numerically inverting their Laplace transforms. Past decades, the scientific computing community has developed excellent numerical methods for solving problems governed by partial differential equations (PDEs), making the availability of a Laplace transform not necessary here for computational purposes. In this study we demonstrate that numerical PDE problem solvers are suitable for computing first passage times, and can be very efficient for this purpose. By doing extensive computational experiments, we show that modern PDE problem solvers can outperform numerical Laplace transform inversion, even if a transform is available. When the Laplace transform is explicit (e.g. does not require the computation of an eigensystem), numerical transform inversion remains the primary method of choice.