English

Effective Sample Size and Generalization Bounds for Temporal Networks

Machine Learning 2026-03-05 v3 Artificial Intelligence

Abstract

Learning from time series is fundamentally different from learning from i.i.d.\ data: temporal dependence can make long sequences effectively information-poor, yet standard evaluation protocols conflate sequence length with statistical information. We propose a dependence-aware evaluation methodology that controls for effective sample size NeffN_{\text{eff}} rather than raw length NN, and provide end-to-end generalization guarantees for Temporal Convolutional Networks (TCNs) on β\beta-mixing sequences. Our analysis combines a blocking/coupling reduction that extracts B=Θ(N/logN)B = \Theta(N/\log N) approximately independent anchors with an architecture-aware Rademacher bound for 2,1\ell_{2,1}-norm-controlled convolutional networks, yielding O(Dlogp/B)O(\sqrt{D\log p / B}) complexity scaling in depth DD and kernel size pp. Empirically, we find that stronger temporal dependence can \emph{reduce} generalization gaps when comparisons control for NeffN_{\text{eff}} - a conclusion that reverses under standard fixed-NN evaluation, with observed rates of Neff0.9N_{\text{eff}}^{-0.9} to Neff1.2N_{\text{eff}}^{-1.2} substantially faster than the worst-case O(N1/2)O(N^{-1/2}) mixing-based prediction. Our results suggest that dependence-aware evaluation should become standard practice in temporal deep learning benchmarks.

Keywords

Cite

@article{arxiv.2508.06066,
  title  = {Effective Sample Size and Generalization Bounds for Temporal Networks},
  author = {Barak Gahtan and Alex M. Bronstein},
  journal= {arXiv preprint arXiv:2508.06066},
  year   = {2026}
}
R2 v1 2026-07-01T04:40:29.790Z