Convergence issues in derivatives of Monte Carlo nullcollision integral formulations: A solution
Abstract
When a Monte Carlo algorithm is used to evaluate a physical observable A, it is possible to slightly modify the algorithm so that it evaluates simultaneously A and the derivatives ∂_{ς} A of A with respect to each problemparameter ς. The principle is the following: Monte Carlo considers A as the expectation of a random variable, this expectation is an integral, this integral can be derivated as function of the problemparameter to give a new integral, and this new integral can in turn be evaluated using Monte Carlo. The two Monte Carlo computations (of A and ∂_{ς} A) are simultaneous when they make use of the same random samples, i.e. when the two integrals have the exact same structure. It was proven theoretically that this was always possible, but nothing ensures that the two estimators have the same convergence properties: even when a large enough samplesize is used so that A is evaluated very accurately, the evaluation of ∂_{ς} A using the same sample can remain inaccurate. We discuss here such a pathological example: nullcollision algorithms are very successful when dealing with radiative transfer in heterogeneous media, but they are sources of convergence difficulties as soon as sensitivityevaluations are considered. We analyse theoretically these convergence difficulties and propose an alternative solution.
 Publication:

Journal of Computational Physics
 Pub Date:
 July 2020
 DOI:
 10.1016/j.jcp.2020.109463
 arXiv:
 arXiv:1903.06508
 Bibcode:
 2020JCoPh.41309463T
 Keywords:

 Monte Carlo method;
 Direct derivatives;
 Nullcollision algorithm;
 Sensitivity;
 Integral formulation;
 Physics  Computational Physics
 EPrint:
 doi:10.1016/j.jcp.2020.109463