1
$\begingroup$

I am using complex-valued ICA to extract sources for complex-valued sensor data. One of the three ambiguities for complex ICA is phase ambiguity, i.e., phase rotation $\exp(i\theta_k)$ of the sources $s_k$ (in addition to permutation and scaling ambiguities, as for real ICA). Yet the complex linear ICA model $\bf{z}=A\cdot \bf{s}$ starts from the general form that all three quantities $\bf{z},A,\bf{s}$ are complex, so the merit of this assumption for $\bf{s}$ eludes me. Is there any meaning that can be given to the phases of the $s_k$, or is $|s_k|$ all that can be reported?

$\endgroup$
0
$\begingroup$

ICA is run on multiple samples: for a matrix $A$ and collections of vectors $\{s_n\}_{n=1}^N$ and $\{z_n\}_{n=1}^N$, the model says $z_n \approx As_n$. The $z$'s are observed, while $A$ and the $s$'s are inferred. If each entry of $s$ is allowed to be complex, then the algorithm may discover that $s_{1,1}$ and $s_{2, 1}$ are out of phase, meaning their ratio is not real. If entries of $s$ are all real, that situation cannot arise. So, it does make a difference; you can't just (for example) squeeze an extra diagonal matrix in there and make the phase difference of the latent variables go away.

$\endgroup$

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.