Thèse Inférence Statistique Sous Confideniélité Différentielle pour Données Matricielles Structurées H/F - Doctorat.Gouv.Fr
Thèse Inférence Statistique Sous Confideniélité Différentielle pour Données Matricielles Structurées H/F
Doctorat.Gouv.Fr · Paris - 75 · CDD
Les missions du poste
Établissement : Institut Polytechnique de Paris École nationale de la statistique et de l'administration économique
École doctorale : Mathématiques Hadamard
Laboratoire de recherche : CREST - Centre de recherche en économie et statistique
Direction de la thèse : Cristina BUTUCEA ORCID 0009000177639510
Début de la thèse : 2027-09-01
Date limite de candidature : 2027-01-15T23:59:59
This project aims to develop a theoretical framework for differentially private statistical inference with structured matrix-valued data. It considers both independent observations and temporally dependent matrix-valued processes, focusing on three fundamental problems: matrix completion, compressed sensing and low-rank matrix estimation. The central objective is to establish non-asymptotic statistical guarantees and minimax optimality results that characterise the trade-off between privacy, estimation accuracy and sample complexity. For independent observations, the project will investigate private recovery from incomplete entries and noisy linear measurements, exploiting low-rank, sparse and low-rank-plus-sparse structures. For time series, it will address private estimation of covariance and spectral density matrices under suitable dependence assumptions. By combining privacy mechanisms, high-dimensional statistical methods and information-theoretic lower bounds, the project seeks to identify fundamental limits and develop optimal or near-optimal procedures for private matrix estimation and recovery.
The privacy of data is a major issue of our society and the mathematical concept is now well formalized and it is common knowledge that individual privacy deteriorates the behaviour of inference methods (estimation and tests). It is even more so in the case of complex data such as matrix-valued independent or time-dependent series. Algorithms have been suggested in a few papers, however theoretical guarantees still lack in many such setups.
Le profil recherché
Profil mathématique et mathématiques appliquées, en particulier: statistique mathématique, théorie de l'information et probabilités.