Blind block partitioned tensor decomposition for radio activity estimation by a single sensor

We consider a scenario where a single sensor is observing a bandwidth of the radio spectrum occupied by several signals which are non-orthogonal in both time and frequency. We form a tensor using anti-diagonal slices of the trispectrum and show that it can be modeled by a block-partitioned tensor (BPT) structure. Current BPT decomposition algorithms assume known factor-matrix partitioning which does not apply in our setting. Hence we develop a blind BPT decomposition strategy to estimate each individual signals´ power spectrum and activity-in-time. We verify the functioning of the algorithm through Monte Carlo simulations and observe that it can accurately estimate the BPT paritioning and factor matrices. The accuracy of the estimates is a function of the correlation between factor matrix columns and the frequency resolution used.