Unbiased Monte Carlo evaluation of certain functional integrals (Q1089719): Difference between revisions

The paper is concerned with the numerical evaluation of functional integrals of the form E exp\(\int^{T}_{0}c(s,X(s))ds\) where X is a Markov process in \(R^ d\) with initial condition \(X(0)=0\) and explicitly known transition function, c is a measurable function, and E denotes the mathematical expectation in the underlying probability space. For instance, functional integrals of this form are related, via Feynman- Kac's type formulas, to the solutions of partial differential equations. Classical methods for the numerical evaluation of functional integrals are based on various approximations by means of finite dimensional integrals; since these integrals are high dimensional, they are often computed by Monte Carlo methods. The paper presents a Monte Carlo method that is not based on approximation by finite dimensional integrals, but on a transformation of the functional integral into an integral over a countable union of finite dimensional spaces. This method avoids the introduction of the systematic error due to the finite dimensional approximation, so that only the statistical error is present. Therefore the error analysis becomes easier, and can be performed by means of confidence intervals during the process of generation of independent samples of the estimator. Finally, some variance reduction techniques are considered, and connection with the von Neumann-Ulam scheme for solving linear integral equations is discussed.