Publications

5 selected recent publications

(one/year over the last 5 years)
  • Universal generalization guarantees for WDRO
    Tam Le, Jerome Malick
    Int. Conf. Learning Representations (ICLR) (2025) [preprint]

  • What is the Long-Run Distribution of Stochastic Gradient Descent?
    W. Azizian, F. Iutzeler, J. Malick, P. Mertikopoulos
    Int. Conf. on Machine Learning (ICML) (2024) [paper] [preprint]

  • Harnessing structure in composite nonsmooth minimization
    Gilles Bareilles, Franck Iutzeler, Jerome Malick
    SIAM Journal on Optimization (2023) [paper] [preprint]

  • Multi-agent online optimization with delays: Asynchronicity, adaptivity, and optimism
    YG. Hsieh, F. Iutzeler, J. Malick, P. Mertikopoulos
    Journal of Machine Learning Research (2022) [paper][preprint]

  • Proximal gradient methods with adaptive subspace sampling
    Dmitry Grishchenko, Franck Iutzeler, Jerome Malick
    Mathematics of Operations Research (2021) [paper][preprint]

My personal top-5

(just my favorite ones)
  • Local linear convergence for alternating nonconvex projections
    Adrian Lewis, Russel Luke, Jerome Malick
    Foundations of Computational Mathematics (2009) [paper][preprint]
    The first study in the non-convex setting of the simple and fundamental algorithm of alternating projections. This gives nice illustrations on how geometry controls convergence: see e.g. the idea of the proof of Theorem 5.2 (page 14 in the preprint).

  • Cut-generating functions and S-free sets
    M. Conforti, G. Cornuejols, A. Daniilidis, C. Lemarechal, J. Malick
    Mathematics of Operations Research (2014) [paper][preprint]
    This paper establishes the basis of a theory of cut-generating functions, a fundamental tool in combinatorial optimization. Nice fact: this theory is grounded on convex geometry -- and even subtle new results (we needed the smallest representation of pre-polars of convex sets; see Theorem 3.8). The proof of Theorem 5.1 required creativity and abnegation; we got both from geometrical constructions.

  • U-Newton methods for nonsmooth convex minimization
    Scott Miller, Jerome Malick
    Mathematical Programming (2005) [paper][preprint]
    Newton is everywhere ! This paper reveals the connection between (i) U-Newton methods from nonsmooth optimization, (ii) Riemannian Newton methods, and (iii) standard Newton methods (SQP) of non-linear optimization The paper is not easy to read; you may prefer this one, which is a follow-up treating the non-convex case.

  • Decomposition algorithm for large-scale two-stage unit-commitment
    Wim van Ackooij, Jerome Malick
    Annals of Operations Research (2016) [paper][preprint]
    How to save carbon emission (and money)? Include stochasticity in the electricity park management model. We propose an effective algorithm to solve resulting large-scale optimization problem.

  • Harnessing structure in composite nonsmooth minimization
    Gilles Bareilles, Franck Iutzeler, Jerome Malick
    SIAM Journal on Optimization (2021)[paper][preprint]
    If you think you know all about the proximal mapping, check out Theorem 3.2 (page 9) and you'll be surprised! The prox implicity identifies relevant sub-manifolds, for a specific range of prox-parameters, as illustrated on this picture for max-functions. Property both surprising and useful to design superlinear algorithms minimizing nonsmooth functions, as the max-eigenvalue of a parametrized matrice.

Last highlight: "apply mechanics to maths" as JJ Moreau said... or vice-versa ?

  • A formulation of the linear discrete Coulomb friction problem via convex optimization
    Vincent Acary, Florent Cadoux, Claude Lemarechal, Jerome Malick
    ZAMM (Journal of Applied Maths and Mechanics) [paper][preprint]
    I worked on applications of optim' in computer vision, electricity, learning obviously - but also in mechanics, with my former colleagues of the bipop/tripop team, where I belong back in my inria days :) In this paper, we isolate the convexity in the problems with contact/friction.