Tobias Rydén, Statistiska institutionen, Stockholms universitet

SEMINARIUM
Datum: onsdag 18 mars 2026
Tid: 13:00 - 14:00
Plats: Campus Albano, lärosal 26, hus 4, plan 2

Regularised Estimation of Large Covariance Matrices with Convex Optimisation

Consider the problem of estimating a covariance matrix of size p from N data points in R^p. The setting where p both and N are large, and p/N tends to a limit, has received considerable attention over the last 20 years so. Various approaches to handling this situation exist, two of which are eigenvalue shrinkage, and sparsity.

Eigenvalue shrinkage means that a preliminary estimator (usually the sample covariance matrix) is taken as a base estimate, and then its eigenvalues are shrunk to obtain an improved estimate. In the talk I will outline a method that models the function mapping preliminary eigenvalues to shrunk ones as a cubic spline, estimated using a variation of cross-validation. There is a considerable literature on asymptotically optimal estimators in this class; the proposed method gives potentially more flexibility in the choice a suitable loss function and the choice of preliminary estimator.

A small numerical example illustrates that the approach can beat linear shrinkage for a variety of loss functions and, depending on the loss function, be on par with or beat analytically derived non-linear estimators. The second part of the talk (joint work with Samuel Erickson, KTH) will be about a method ("joint partial regression", JPR) that computes an estimate for which the precision matrix (inverse covariance matrix) is sparse (has relatively few non-zero elements). The method is based on the observation that columns of the precision matrix (less the diagonal elements) provide optimal linear predictors of the variables held out. JPR computes simultaneous LASSO regressions for each column, while constraining the estimated matrix to be symmetric and positive semi-definite. The convergence rate of JPR beats those of many previously proposed methods.

Senast uppdaterad: 2026-03-10

Sidansvarig: Statistiska institutionen