English

Tractability from overparametrization: The example of the negative perceptron

Machine Learning 2025-03-25 v3 Probability Statistics Theory Statistics Theory

Abstract

In the negative perceptron problem we are given nn data points (xi,yi)({\boldsymbol x}_i,y_i), where xi{\boldsymbol x}_i is a dd-dimensional vector and yi{+1,1}y_i\in\{+1,-1\} is a binary label. The data are not linearly separable and hence we content ourselves to find a linear classifier with the largest possible \emph{negative} margin. In other words, we want to find a unit norm vector θ{\boldsymbol \theta} that maximizes mininyiθ,xi\min_{i\le n}y_i\langle {\boldsymbol \theta},{\boldsymbol x}_i\rangle. This is a non-convex optimization problem (it is equivalent to finding a maximum norm vector in a polytope), and we study its typical properties under two random models for the data. We consider the proportional asymptotics in which n,dn,d\to \infty with n/dδn/d\to\delta, and prove upper and lower bounds on the maximum margin κs(δ)\kappa_{\text{s}}(\delta) or -- equivalently -- on its inverse function δs(κ)\delta_{\text{s}}(\kappa). In other words, δs(κ)\delta_{\text{s}}(\kappa) is the overparametrization threshold: for n/dδs(κ)εn/d\le \delta_{\text{s}}(\kappa)-\varepsilon a classifier achieving vanishing training error exists with high probability, while for n/dδs(κ)+εn/d\ge \delta_{\text{s}}(\kappa)+\varepsilon it does not. Our bounds on δs(κ)\delta_{\text{s}}(\kappa) match to the leading order as κ\kappa\to -\infty. We then analyze a linear programming algorithm to find a solution, and characterize the corresponding threshold δlin(κ)\delta_{\text{lin}}(\kappa). We observe a gap between the interpolation threshold δs(κ)\delta_{\text{s}}(\kappa) and the linear programming threshold δlin(κ)\delta_{\text{lin}}(\kappa), raising the question of the behavior of other algorithms.

Keywords

Cite

@article{arxiv.2110.15824,
  title  = {Tractability from overparametrization: The example of the negative perceptron},
  author = {Andrea Montanari and Yiqiao Zhong and Kangjie Zhou},
  journal= {arXiv preprint arXiv:2110.15824},
  year   = {2025}
}

Comments

107 pages; 7 pdf figures

R2 v1 2026-06-24T07:17:54.518Z