Log-optimal independence testing by betting
We construct anytime-valid tests of independence whose log wealth per observation converges to mutual information for every joint distribution on a pair of standard Borel spaces. The same data-driven procedure works without knowledge of either marginal distribution or a common dominating measure. A conditional predictive likelihood, normalized by the maximum marginal likelihood, gives an e-process in the full observation filtration. Refining finite partitions and multiclass kernel learning yield universal attainment; arbitrary probabilistic classifiers give practical implementations that can be trained throughout the test. For two complex data types, a normalized sequential assignment rule gives an exact matching e-value with quadratic score-evaluation cost. We connect these constructions to classical fixed-sample efficiency. Experiments compare practical implementations with sequential kernelized and classifier-based tests, permutation HSIC, and held-out classifier tests.
Authors
- Aaditya Ramdas (ORCID: https://orcid.org/0000-0003-0497-311X)
Institutions
- Stanford University (US)
Publication Details
- Journal
- Stanford Digital Repository
- Published
- 2026-11-14
- DOI
- https://doi.org/10.25740/pg979gk9581
- Primary Topic
- Statistical Methods and Inference
- Type
- article
- Field-Weighted Citation Impact
- 0.00