A Model of Double Descent for High-Dimensional Logistic Regression

Zeyu Deng, Abla Kammoun, Christos Thrampoulidis

Research output: Chapter in Book/Report/Conference proceedingConference contribution

9 Scopus citations


We consider a model for logistic regression where only a subset of features of size p is used for training a linear classifier over n training samples. The classifier is obtained by running gradient-descent (GD) on the logistic-loss. For this model, we investigate the dependence of the classification error on the overparameterization ratio κ = p/n. First, building on known deterministic results on convergence properties of the GD, we uncover a phase-transition phenomenon for the case of Gaussian features: the classification error of GD is the same as that of the maximum-likelihood (ML) solution when κ < κ⋆, and that of the max-margin (SVM) solution when κ > κ⋆. Next, using the convex Gaussian min-max theorem (CGMT), we sharply characterize the performance of both the ML and SVM solutions. Combining these results, we obtain curves that explicitly characterize the test error of GD for varying values of κ. The numerical results validate the theoretical predictions and unveil "double-descent" phenomena that complement similar recent observations in linear regression settings.
Original languageEnglish (US)
Title of host publicationICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
ISBN (Print)978-1-5090-6632-2
StatePublished - 2020

Bibliographical note

KAUST Repository Item: Exported on 2020-10-01


Dive into the research topics of 'A Model of Double Descent for High-Dimensional Logistic Regression'. Together they form a unique fingerprint.

Cite this