Iterative thresholding for blind block partitioned tensor decomposition

We consider a model where a single sensor observes a bandwidth of spectrum occupied by several non-orthogonal in both time and frequency signals. The sensor constructs a tensor based on the trispectrum and which can be modeled as a block partitioned tensor (BPT). By decomposing this tensor we can estimate signal activity in both time and frequency. We develop a partition-blind BPT decomposition algorithm using iterative thresholding and parallel coordinate descent. We test the algorithm in simulations and find that it succeeds in estimating received signal PSDs and time activity substantially faster than previous algorithms.