Communication Complexity of Distributed High Dimensional Correlation Testing

Abstract

We consider a two-party distributed hypothesis testing problem for correlated Gaussian random variables. For a d-dimensional random vector X and a scalar random variable Y, where X and Y are jointly Gaussian with an unknown correlation vector ρ, parties <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">P</i> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub> and <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">P</i> <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> observe independent copies of X and Y, respectively. The parties seek to test if their observations are correlated or not, namely they seek to test if ||ρ|| <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> exceeds τ or is it 0. To that end, they communicate interactively and declare the test output. We show that roughly order d/τ <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> bits of communication are sufficient and necessary for resolving the distributed correlation testing problem above. Furthermore, we establish a lower bound of roughly d <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> /τ <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> bits for the communication needed for distributed estimation of ρ, implying that distributed correlation testing requires less communication than distributed estimation. Both our lower bounds for testing and estimation hold for an arbitrary d and interactive communication with shared randomness, while our distributed test requires only one-way communication with shared randomness. For the one-dimensional case, with one-way communication and with probability of one of the error-types fixed, our bounds are more refined in the dependence on the other error-type and are tight even in the constant.