An efficient and accurate approximation to the distribution of quadratic forms of Gaussian variables
share
Fast and accurate calculation for the distributions of Quadratic forms of centered Gaussian variables is of interest in computational statistics. This paper presents a novel numerical procedure to efficiently compute the moments of a given quadratic form. Based on that, a gamma distribution with matched skewness-kurtosis ratio is proposed to approximate its distribution. Comparing with existing methods, the new method is significantly more accurate in getting the right-tail probability. The new method may facilitate the hypothesis testing in the analysis of big data, where efficient and accurate calculation of small p-values is desired. Relevant R functions are provided in the Supplementary Materials.
READ FULL TEXT