45
5

Tensor denoising with trend filtering

Abstract

We extend the notion of trend filtering to tensors by considering the kthk^{\rm th}-order Vitali variation, a discretized version of the integral of the absolute value of the kthk^{\rm th}-order total derivative. We prove adaptive 0\ell^0-rates and not-so-slow 1\ell^1-rates for tensor denoising with trend filtering. For k={1,2,3,4}k=\{1,2,3,4\} we prove that the dd-dimensional margin of a dd-dimensional tensor can be estimated at the 0\ell^0-rate n1n^{-1}, up to logarithmic terms, if the underlying tensor is a product of (k1)th(k-1)^{\rm th}-order polynomials on a constant number of hyperrectangles. For general kk we prove the 1\ell^1-rate of estimation nH(d)+2k12H(d)+2k1n^{- \frac{H(d)+2k-1}{2H(d)+2k-1}}, up to logarithmic terms, where H(d)H(d) is the dthd^{\rm th} harmonic number. Thanks to an ANOVA-type of decomposition we can apply these results to the lower dimensional margins of the tensor to prove bounds for denoising the whole tensor. Our tools are interpolating tensors to bound the effective sparsity for 0\ell^0-rates, mesh grids for 1\ell^1-rates and, in the background, the projection arguments by Dalalyan et al.

View on arXiv
Comments on this paper