Testing Dependency of Databases

Wasim Huleihel
Tel Aviv University
Electrical Engineering

In this talk, we will investigate the problem of detecting the dependency between two random databases represented as matrices. This is formalized as a hypothesis testing problem, where under the null hypothesis, the two databases are independently generated; under the alternative, the two databases are dependent under some latent row permutation/alignment, but have the same marginal distributions as the null. We determine sharp thresholds (information-theoretic and computational) at which optimal testing error probability exhibits a phase transition from zero to one, as a function of the database's dimensions and their generative distributions.

Presentation (PDF File)

Back to EnCORE Workshop on Computational vs Statistical Gaps in Learning and Optimization