Fast Direct Methods for Gaussian Processes

摘要

A number of problems in probability and statistics can be addressed using the multivariate normal (Gaussian) distribution. In the one-dimensional case, computing the probability for a given mean and variance simply requires the evaluation of the corresponding Gaussian density. In the -dimensional setting, however, it requires the inversion of an covariance matrix, , as well as the evaluation of its determinant, . In many cases, such as regression using Gaussian processes, the covariance matrix is of the form , where is computed using a specified covariance kernel which depends on the data and additional parameters (hyperparameters). The matrix is typically dense, causing standard direct methods for inversion and determinant evaluation to require work. This cost is prohibitive for large-scale modeling. Here, we show that for the most commonly used covariance functions, the matrix can be hierarchically factored into a product of block low-rank updates of the identity matrix, yielding an algorithm for inversion. More importantly, we show that this factorization enables the evaluation of the determinant , permitting the direct calculation of probabilities in high dimensions under fairly broad assumptions on the kernel defining . Our fast algorithm brings many problems in marginalization and the adaptation of hyperparameters within practical reach using a single CPU core. The combination of nearly optimal scaling in terms of problem size with high-performance computing resources will permit the modeling of previously intractable problems. We illustrate the performance of the scheme on standard covariance kernels.