A receiver separates multiple component signals carried in a composite signal that have been spread over time and frequency. Multiple diversity branches receive the composite signal. A tensor generation circuit arranges samples of that signal along time, frequency, and diversity dimensions to obtain a composite signal tensor. A set of cascaded least squares estimators then separates the component signals by fitting a constrained tensor model of the composite signal to the composite signal tensor in an alternating least squares estimation process, alternating during any given iteration of the process between fitting different unfolded representations of the constrained tensor model to corresponding unfolded representations of the composite signal tensor. The tensor model models the composite signal as a linear combination of rank-one tensors associated with the component signals, and is constrained in the sense that each rank-one tensor remains constrained according to the spreading of its associated component signal over time.