On the Expressivity of GNN for Solving Second Order Cone Programming

Ruizhe Li, Enming Liang, Minghua Chen

Keywords: graph neural network, second order cone programming, learning to optimize, Weisfeiler-Lehman test, universality

TL;DR: Graph neural networks can provably solve second-order cone programming.

Abstract: Graph Neural Networks (GNNs) have demonstrated both empirical efficiency and universal expressivity for solving constrained optimization problems such as linear and quadratic programming. However, extending this paradigm to more general convex problems with universality guarantees, particularly Second-Order Cone Programs (SOCPs), remains largely unexplored. We address this challenge by proposing a novel graph representation that captures the structure of conic constraints. We then establish a key universality theorem: *there exist GNNs that can provably approximate essential SOCP properties, such as instance feasibility and optimal solutions*. This work provides a rigorous foundation linking GNN expressive power to conic optimization structure, opening new avenues for scalable, data-driven SOCP solvers. Our approach extends to $p$-order cone programming for any $p \geq 1$ with universal expressivity preserved, requiring no GNN structural modifications. Numerical experiments on randomly generated SOCPs demonstrate the expressivity of the proposed GNN, which achieves better prediction accuracy with fewer parameters than fully connected neural networks.

Submission Number: 5