Dear All,

We have the pleasure thanks to the support of the ESSEC IDS dpt, Institut des Actuaires, LabEx MME-DII, the group BFA (SFdS), to invite you to the seminar by:

We consider a penalty framework based on regularizing the squared distance to set-based constraints for several core statistical tasks. These distance-to-set penalties convert problems cast as constrained optimization problems to simpler, more tractable unconstrained ones. These formulations can be more flexible than many existing algebraic and regularization penalties. We will see that they often avoid drawbacks that arise from popular alternatives such as shrinkage. We present a general strategy for eliciting effective algorithms in this framework using majorization-minimization (MM), a principle that transfers difficult problems onto a sequence of more manageable subproblems through the use of surrogate functions. Methods derived from this perspective feature monotonicity, are often amenable to acceleration, and come with global convergence guarantees. We showcase new progress on classical problems including constrained generalized linear models and sparse covariance estimation using this approach, and discuss connections to constraint relaxation from a Bayesian perspective.