The fat-shattering dimension of norm-bounded two-layer ReLU networks in fixed dimension
Consider the functions $x\mapsto u^{T}\sigma(Wx)$ on the ball of radius b_x in $\mathbb{R}^d$, where σ is the ReLU function, W has n rows, $\|W\|_F\le B$ and |u| ≤ b, and put R = bBb_x/ε. For every width and dimension the fat-shattering dimension of this class at scale ε is at most R^2 up to logarithmic factors, and Vardi, Shamir and Srebro asked whether a better bound holds when d is small. We prove that for fixed d ≥ 2 the fat-shattering dimension is at most a constant times $R^{2(d-1)/(d+2)}\log(2+R)$ for every width, and at least a constant times R^2(d-1)/(d+2) when R exceeds a constant depending on d and the width is at least that large. The upper bound writes the networks through support functions of zonoids, as Bach did, and uses the approximation of zonoids by zonotopes due to Matou\v{s}ek and Siegel. The lower bound restricts the networks to an affine hyperplane and uses the embedding of Sobolev spaces into the variation space of ReLU networks with biases due to Mao, Siegel and Xu.
Authors
- Yitzchak Shmalo (ORCID: https://orcid.org/0000-0002-2886-3049)
Institutions
- Hebrew University of Jerusalem (IL)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24284959
- Primary Topic
- Stochastic Gradient Optimization Techniques
- Type
- preprint