Lipschitz constants are widely used to measure how much a neural network's output can change when its input is slightly perturbed. They are central to robustness guarantees, but exact computation is known to be difficult even for shallow ReLU networks.
The paper on arXiv studies this problem from the perspective of parameterized complexity, focusing on input-convex neural networks. It also connects the problem to maximizing L_p norms over zonotopes, a geometric structure that appears naturally in the analysis of these networks.
By treating certain structural parameters as fixed, the authors aim to identify which instances become tractable and which remain hard. The work adds to the broader effort to understand when verification and sensitivity analysis of neural networks is computationally feasible.