arXiv Science⌕ Search

arXiv subjects

Omar Kassi

Publications and source records attributed to Omar Kassi.

4 recordsLinked to original sources

Posterior contraction rates for banded approximations of Gaussian processes

Gaussian process regression models are widely used in modern statistics and machine learning due to their flexibility, interpretability, built-in uncertainty quantification, and strong theoretical foundations. However, training Gaussian processes (GPs) has a computational cost of $O(n^3)$, while the memory and prediction costs are also $O(n^2)$. For large datasets commonly encountered in practice, this becomes infeasible and necessitates the use of approximate posterior methods. In this work, we propose a class of approximations based on sparse covariance matrices, where the kernel matrix is sparsified according to the spatial proximity of the design points. We analyze the accuracy of the resulting approximate posterior distribution and show that the proposed approach preserves optimal statistical inference properties, provided that the dependencies among a sufficient number of neighboring points for each feature are retained. We derive explicit sufficient and necessary conditions on the number of neighbors required in a general setting and apply these results to standard kernels, including the squared exponential and Matérn kernels. Our theoretical findings are supported by numerical experiments.

math.ST↗

Structural adaptation and rate accelerated estimation in bivariate functional data

We introduce directional regularity, a new definition of anisotropy for multivariate functional data. Instead of taking the conventional view, which determines anisotropy as a notion of smoothness along a dimension, directional regularity additionally views anisotropy through the lens of directions. We show that faster rates of convergence for smoothing can be obtained through a change-of-basis by adapting to the anisotropy of a bivariate process. An algorithm for the estimation and identification of the change-of-basis matrix is constructed, made possible due to the replication structure of functional data. Non-asymptotic bounds are provided for our algorithm, supplemented by numerical evidence from an extensive simulation study. Finally, a real-world rainfall measurement dataset is analyzed with our methods.

stat.ME↗

Optimal inference for the mean of random functions

We study estimation and inference for the mean of real-valued random functions defined on a hypercube. The independent random functions are observed on a discrete, random subset of design points, possibly with heteroscedastic noise. We propose a novel optimal-rate estimator based on Fourier series expansions and establish a sharp non-asymptotic error bound in $L^2-$norm. Additionally, we derive a non-asymptotic Gaussian approximation bound for our estimated Fourier coefficients. Pointwise and uniform confidence sets are constructed. Our approach is made adaptive by a plug-in estimator for the Hölder regularity of the mean function, for which we derive non-asymptotic concentration bounds.

math.ST↗

Learning the regularity of multivariate functional data

Combining information both within and between sample realizations, we propose a simple estimator for the local regularity of surfaces in the functional data framework. The independently generated surfaces are measured with errors at possibly random discrete times. Non-asymptotic exponential bounds for the concentration of the regularity estimators are derived. An indicator for anisotropy is proposed and an exponential bound of its risk is derived. Two applications are proposed. We first consider the class of multi-fractional, bi-dimensional, Brownian sheets with domain deformation, and study the nonparametric estimation of the deformation. As a second application, we build minimax optimal, bivariate kernel estimators for the reconstruction of the surfaces.

math.ST↗