Tensor Expander Chernoff Bounds

Abstract

The Chernoff bound is an important inequality relation in probability theory. The original version of the Chernoff bound is to give an exponential decreasing bound on the tail distribution of sums of independent random variables. Recent years, several works have been done by extending the original version of the Chernoff bound to high-dimensional random objects, e.g., random matrices, or/and to consider the relaxation that there is no requirement of independent assumptions among random objects. In this work, we generalize the matrix expander Chernoff bound studied by Garg et al. at work: A Matrix Expander Chernoff Bound, to tensor expander Chernoff bounds. Our main tool is to develop new tensor norm inequalities based on log-majorization techniques. These new tensor norm inequalities are used to bound the expectation of Ky Fan norm of the random tensor exponential function, then tensor expander Chernoff bounds can be established. Compared with the matrix expander Chernoff bound, the tensor expander Chernoff bounds proved at this work contributes following aspects: (1) the random objects dimensions are increased from matrices (two-dimensional data array) to tensors (multidimensional data array); (2) this bound generalizes the identity map of the random objects summation to any polynomial function of the random objects summation; (3) Ky Fan norm, instead only the maximum or the minimum eigenvalues, for the function of the random objects summation is considered; (4) we remove the restriction about the summation of all mapped random objects is zero, which is required in the matrix expander Chernoff bound derivation.