Exact distribution of selected multivariate test criteria by numerical inversion of their characteristic functions

TL;DR

This paper presents a numerical inversion method for the exact distribution of multivariate test criteria based on their characteristic functions, providing precise results where analytical solutions are complex or unavailable.

Contribution

It introduces a practical numerical inversion approach for characteristic functions to obtain exact distributions of multivariate test statistics, filling a gap in existing statistical software.

Findings

Exact distribution of Bartlett's test statistic computed

Wilks's Lambda distribution accurately derived

Numerical inversion proves efficient and precise

Abstract

Application of the exact statistical inference frequently leads to a non-standard probability distributions of the considered estimators or test statistics. The exact distributions of many estimators and test statistics can be specified by their characteristic functions. Typically, distribution of many estimators and test statistics can be structurally expressed as a linear combination or product of independent random variables with known distributions and characteristic functions, as is the case for many standard multivariate test criteria. The characteristic function represents complete characterization of the distribution of the random variable. However, analytical inversion of the characteristic function, if possible, frequently leads to a complicated and computationally rather strange expressions for the corresponding distribution function (CDF/PDF) and the required quantiles. As an efficient alternative, here we advocate to use the well-known method based on numerical inversion of the characteristic functions --- a method which is, however, ignored in popular statistical software packages. The applicability of the approach is illustrated by computing the exact distribution of the Bartlett's test statistic for testing homogeneity of variances in several normal populations and the Wilks's $\Lambda$-distribution used in multivariate hypothesis testing.