Effective Sample Size and Generalization Bounds for Temporal Networks
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 rather than raw length , and provide end-to-end generalization guarantees for Temporal Convolutional Networks (TCNs) on -mixing sequences. Our analysis combines a blocking/coupling reduction that extracts approximately independent anchors with an architecture-aware Rademacher bound for -norm-controlled convolutional networks, yielding complexity scaling in depth and kernel size . Empirically, we find that stronger temporal dependence can \emph{reduce} generalization gaps when comparisons control for - a conclusion that reverses under standard fixed- evaluation, with observed rates of to substantially faster than the worst-case mixing-based prediction. Our results suggest that dependence-aware evaluation should become standard practice in temporal deep learning benchmarks.
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}
}