Systolic arrays for spherical subspace tracking

Real-time subspace tracking demands fast parallel algorithms with localized communication. Here we parallelize the two-level signal averaged (SA2) spherical subspace tracker by converting it to a systolic flow of inner-products and plane rotations. We similarly parallelize the four level signal averaged (SA4) subspace tracker. The SA4 tracker provides a means to adaptively adjust the signal rank estimate as the number of sources change. We call this on-the-fly process of subspace rank estimate changing “gear-shifting” and use a customized version of MDL to trigger the gear-shifts in parallel. The systolic algorithms all run on our UNIX based iWarp simulator