Reproduce “Weighted Empirical Risk Minimization for Machine Learning under Long-Range Dependence: Exact Pathwise Rates and Learning-Error Geometry” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Elina Moldavskaya
For parametric empirical risk minimization driven by long-range-dependent Gaussian time series, the paper derives almost-sure Bahadur representations and exact laws of the iterated logarithm for weighted estimators. The polynomial learning rate is governed by the memory exponent and the score’s Hermite-chaos rank, while power weighting changes sharp fluctuation constants and cluster-set geometry rather than the exponent itself.
Key insights
- In the interior long-memory regime, the polynomial learning exponent is -αm/2, where α is the covariance-decay exponent and m is the score’s Hermite rank; admissible power weighting does not alter this exponent.
- Despite potentially multidimensional parameters, the leading learning error is asymptotically one-dimensional, concentrated along V^{-1}J; directions orthogonal to this vector are negligible on the LIL scale.
- Weighting affects exact pathwise fluctuation magnitudes and, for higher chaos ranks, the cluster-set shape. Even-rank errors have one-sided endpoint sets, whereas odd-rank errors have symmetric ones.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.