A fast algorithm for numerical solutions to Fortet's equation (Q939562)

Recently, a tendency of the problems concerning default-risk literature is to show the analytic tractability and the ease of econometric estimation. In the present paper the considered process of firm asset forms a continuous multivariate Markov process with known transition probabilities. A fast algorithm for computation of default times of multiple firms in a structural model is presented. This algorithm uses a multivariate extension of the Fortet's equation, established for an one-dimensional setting. By discretization the Fortet's equation the hitting time density of the Markov asset process is presented as a solution of a linear system. Using the Levinson-Durbin algorithm this system is transformed into a block diagonal matrix in which all blocks are in a Toeplitz form. This technique allows to significantly improve the computation time. In Section 2 a time-homogeneous continuous diffusion process in \({\mathbb R}^{M}\) with transition probabilities given in an integral form, which is multivariate extension of the Fortet's equation is considered. In subsection 2.1 discretization of the so-called fundamental equation is developed. Theorem 2 states the convergence in Euclidean norm of the discretization of the fundamental equation to the exact solution of this equation. In subsection 2.2 the details of an effective algorithm for a computation of hatting time densities are exposed. In subsection 2.3 a computational speed-up for octant-like hatting boundaries is given. The concepts of a system \(A_{t}\) and a system \(A_{t}^{\text{ap}}\) of a Toeplitz block structure are presented. Theorem 4 gives the accuracy of the solution of using \(A_{t}^{\text{ap}}\) instead of \(A_{t}.\) In Section 3 the algorithm developed in section 2 is used to determine the default time probabilities and securities price of firms in a network market model. Theorem 6 gives the computational demands of the algorithm. In the case of \(M\) firms, \(N\) discretization points in every space dimension and \(n\) discretization points in the time dimension, the total number of operations in algorithm is \({\mathcal O}(n \cdot \log n \cdot M \cdot M! \cdot N^{{\max}(2, (M-1)M / 2)}).\) In subsection 3.2 the network model is realized in the case of two firms. In Appendix A the proofs of the main results are presented.