Let say if we wish to filter an input sequence x[n1, n2, n3] of NxNxN points using an Linear Shift Invariance system with impulse response h[n1, n2, n3], where the filter is a separable sequence, h[n1, n2, n3] = a[n1]b[n2]c[n3].

How can we develop a computationally efficient way to obtain the output y[n1, n2, n3]?

My thought: since we can represent the Fourier Transform of the Filter as: $$ H(w_1, w_2, w_3) = \sum \sum \sum h[n_1,n_2,n_3]*e^{-jw_1n_1} *e^{-jw_2n_2}*e^{-jw_3n_3} $$ $$ H(w_1, w_2, w_3) = A(w_1)B(w_2)C(w_3)$$, where

$$ A(w_1) = \sum h[n_1,n_2,n_3]*e^{-jw_1n_1} $$ $$ B(w_2) = \sum h[n_1,n_2,n_3]*e^{-jw_2n_2} $$ $$ C(w_3) = \sum h[n_1,n_2,n_3]*e^{-jw_3n_3}$$


$$ y(n_1, n_2, n_3) = x(n_1, n_2, n_3)*h(n_1, n_2, n_3)$$ $$ Y(w_1, w_2, w_3) = X(w_1, w_2, w_3)H(w_1, w_2, w_3)$$ $$ Y(w_1, w_2, w_3) = X(w_1, w_2, w_3)A(w_1)B(w_2)C(w_3)$$

I got stuck in this point and I am not too sure how to proceed through the process. I don't quite see how to separate the sequence and how to multiply through the vectors to get the convolution sum of $Y(w_1, w_2, w_3)$

can anyone please guide me to the right direction? Thanks!


x * h = IFT[FT[x * h] and FT[x * h] is Ft[x] FT[h]. Since h is separable, it might be computationally efficient to expand its transform as a convolution Ft[x]*FT[b]*FT[c].


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.