Simple-regret rates and minimax optimality of fixed-prior expected improvement in Matérn and squared-exponential RKHSs
Authors: Emmanuel Vazquez, Sébastien Petit
Organizations: Universit´e Paris-Saclay, CNRS, CentraleSup´elec, Laboratoire des Signaux et Syst`emes, 91190 Gif-sur-Yvette, France · Laboratoire national de m´etrologie et d’essais (LNE), 1 rue Gaston Boissier, 75724 Paris Cedex 15, France
We study expected improvement (EI) for minimizing a deterministic function f in the RKHS Hk of a continuous positive-semidefinite kernel k on a nonempty compact set X⊂Rd. Function values are observed exactly, and EI is computed from a fixed zero-mean Gaussian-process model with covariance σ2k, σ>0. A weak-EI policy queries a point whose EI is at least a fixed positive fraction of its maximum. We introduce a notion of sequential separation radius relating ranked selected-point innovation norms to Kolmogorov widths, drawing on greedy approximation. Standard power-function estimates from scattered-data approximation and a finite-budget regret argument yield the rates. After N post-initial queries, every weak-EI policy has simple regret O(N−ν/d) for isotropic Matérn kernels of smoothness ν>0 and O(exp[−c1min{N,N1/dlog(eN)}]) for the isotropic squared-exponential kernel, with c1>0. For d=1, the sharper bound O(exp[−c2Nlog(eN)]) holds for exact EI, with c2>0. These bounds are uniform over each fixed RKHS ball. If X has nonempty interior and B>0, the exact EI policy is minimax-rate optimal over the RKHS ball of radius B for Matérn kernels, even among randomized strategies whose final recommendation need not be a query point. For the squared-exponential kernel, it is minimax-rate optimal up to constants in the exponent among deterministic methods whose final recommendation may be any point of X.