Asymptotics of Non-Convex Generalized Linear Models in High-Dimensions: A proof of the replica formula
Matteo Vilucchio, Yatin Dandi, Matéo Pirio Rossignol, Cedric Gerbelot, Florent Krzakala · 2025
Read Asymptotics of Non-Convex Generalized Linear Models in High-Dimensions: A proof of the replica formula by Matteo Vilucchio, Yatin Dandi, Matéo Pirio Rossignol, Cedric Gerbelot, Florent Krzakala (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-20003.txt.
Description
Physicists predicted decades ago how high-dimensional learning problems should behave, using a borrowed trick called the replica method. Matteo Vilucchio and co-authors finally prove those predictions for a hard class of models in a 2025 paper. Their targets are generalized linear models with non-convex losses, where the optimization landscape has many valleys.
Convex cousins like ridge and logistic regression were settled long ago; the non-convex case resisted proof. The team welds two tools together: a Gaussian min-max theorem supplies the floor, and approximate message passing, an iterative algorithm, reaches it. Out comes a rigorous replica formula, valid exactly when the physicists' replicon condition holds. Applications include a verdict in favor of the Tukey loss over the Huber loss under certain contamination. The paper closes a long-standing gap between physics intuition and mathematics.
Catalog record
| CREATOR | Matteo Vilucchio, Yatin Dandi, Matéo Pirio Rossignol, Cedric Gerbelot, Florent Krzakala |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 142 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-20003