A framework for constrained adaptive time-frequency kernel design

In this paper, t-f kernels are updated every data sample using constrained adaptive techniques. The kernel elements along each lag in the time-lag domain are considered as FIR filter coefficients operating on time-series of the data bilinear products. Linearly constrained minimum variance and constrained linear prediction adaptive techniques are used to allow effective reduction of the noise and crossterms without distorting the signal autoterms. Two different cases are considered for which the data-dependent kernel design via linearly constrained minimization proves useful and leads to significant improvements over fixed kernel design