English

Shallower ReLU Network Representations via Exact Linear Algebra

Machine Learning 2026-07-22 v1 Neural and Evolutionary Computing Combinatorics

Abstract

We prove that the maximum of nn real numbers is exactly representable by a ReLU network with two hidden layers for every n10n\le 10. The constructions are obtained by reducing the problem to exact rational linear algebra: after a symmetry reduction, the necessary cancellations are encoded in finite linear systems over Q\mathbb{Q}, which we solve and verify computationally. The representation of max10\max_{10} has a structured first hidden layer consisting only of pairwise maxima, a feature that allows it to be recursively substituted into larger networks. We use this to show that for every n>10n>10, the maximum maxn\max_{n} can be exactly represented with log5(n/2)+1<log5(n)+1.5694\lceil{\log_5 (n / 2)\rceil}+1 < \log_5(n) +1.5694 hidden layers. Via the generalized hinging-hyperplane representation [Wang, Sun, IEEE Trans. Inf. Theory 2005], the same depth bound holds for all continuous piecewise-linear functions on Rd\mathbb{R}^d, with d+1d+1 in place of nn. In particular, every continuous piecewise-linear function on Rd\mathbb{R}^d for d9d\le 9 admits a two-hidden-layer ReLU representation. Our results improve on [Bakaev, Brunck, Hertrich, Stade, Yehudayoff, STOC'26]. In that work, the authors established a two-hidden-layer representation for max5\max_{5} and an upper bound of log3(n2)+1\lceil{\log_3 (n-2)\rceil}+1 hidden layers for maxn\max_{n}.

Cite

@article{arxiv.2607.21651,
  title  = {Shallower ReLU Network Representations via Exact Linear Algebra},
  author = {Kilian Rueß and Gennadiy Averkov and Florestan Brunck and Moritz Grillo and Christoph Hertrich and Georg Loho and Jack Stade and Moritz Stargalla and Matthew Sun and Martin Winter},
  journal= {arXiv preprint arXiv:2607.21651},
  year   = {2026}
}