A log spectral approach to bidirectional deconvolution |
Here we fill in more details of the algorithm. After we are certain of its behavior we would naturally switch over to conjugate directions.
D(omega,x) = FT d(t,x) u=0; iteration { U = FT(u) remove mean from U(omega) exp(U(Z)) dU = 0 for all x r(t,x) = IFT( D exp(U) ) softclip( r ) dU += conjg(FT(r)) * FT(softclip) # "*" means multiply remove mean from dU(omega) for all x dR = FT(r) * dU # "*" means multiply dr = IFT(dR) argmin(alpha) = H(r+alpha*dr) u = u + alpha du }