Polynomial Chaos Expansion for general multivariate distributions with correlated variables

TL;DR

This paper introduces a novel Polynomial Chaos Expansion method capable of handling correlated multivariate distributions, enabling accurate uncertainty quantification without significant additional computational costs.

Contribution

It develops a basis for orthogonal polynomials applicable to correlated variables, extending PCE applicability beyond independent variables.

Findings

Convergence rate similar to independent case

Accurate propagation of experimental errors in correlated systems

Significant differences when accounting for true correlations

Abstract

Recently, the use of Polynomial Chaos Expansion (PCE) has been increasing to study the uncertainty in mathematical models for a wide range of applications and several extensions of the original PCE technique have been developed to deal with some of its limitations. But as of to date PCE methods still have the restriction that the random variables have to be statistically independent. This paper presents a method to construct a basis of the probability space of orthogonal polynomials for general multivariate distributions with correlations between the random input variables. We show that, as for the current PCE methods, the statistics like mean, variance and Sobol' indices can be obtained at no significant extra postprocessing costs. We study the behavior of the proposed method for a range of correlation coefficients for an ODE with model parameters that follow a bivariate normal distribution. In all cases the convergence rate of the proposed method is analogous to that for the independent case. Finally, we show, for a canonical enzymatic reaction, how to propagate experimental errors through the process of fitting parameters to a probabilistic distribution of the quantities of interest, and we demonstrate the significant difference in the results assuming independence or full correlation compared to taking into account the true correlation.