Laplace and saddlepoint approximations in high dimensions (Q6978637)

Paper investigates the theoretical accuracy of two widely used asymptotic approximation methods -- Laplace and saddlepoint approximations -- in settings, where the parameter dimension~\(p\) increases with the sample size~\(n\). While these methods are well understood in fixed-dimensional regimes, their behavior in high-dimensional contexts has remained largely unexplored.\N\NThe authors establish general convergence rates for approximating joint densities, marginal posterior densities, and conditional densities under mild regularity conditions. A central result shows that the relative error of both Laplace and saddlepoint approximations is of order \(O(p^4/n)\) when \(p=o(n^{1/4})\), with improved rates \(O(p^3/n)\) achievable under stronger assumptions on higher-order derivatives of the log-likelihood.\N\NThe paper further demonstrates that improved accuracy can be obtained in structured models, such as logistic regression or generalized linear models, where symmetry or additional assumptions reduce higher-order error terms. Moreover, when approximating ratios of integrals -- such as marginal posterior densities -- certain error terms cancel, allowing for faster growth of \(p\) relative to \(n\).\N\NAnalogous results are derived for saddlepoint approximations using complex analytic techniques, including contour deformation and Laplace's method. The authors show that saddlepoint approximations achieve similar error rates to Laplace approximations under comparable assumptions.\N\NOverall, the paper provides rigorous high-dimensional error bounds for two fundamental approximation methods and highlights conditions under which their accuracy improves. These findings are particularly relevant for modern statistical applications involving high-dimensional models, where exact computations are often infeasible.