How does adaptive whitening work?

52 Views Asked by At

In the Equivariant Adaptive Source Separation paper I am wondering what the intermediate steps for eq. 23 on page 5 are? I am not sure where to even begin with this and haven't been able to find any information online.

Also, this paper introduced relative gradient to me (and I think introduced it in general) which seems like a really profound thing, but the author does not dwell on it at all which leaves me wondering. Is there anything that goes more into depth on it?

1

There are 1 best solutions below

5
On BEST ANSWER

In the paper, we have the following: $$ \delta R_z = \mathcal{E}R_z+R_z\mathcal{E}^T \tag{21} $$ and $$ K(R_z+\delta R_z) = K(R_z) + \mathrm{Trace}\{(I-R_z^{-1})\delta R_z\}+o(\delta R_z) \tag{22} $$

If you replace (21) in (22), and use the fact that $R_z = R_z^T$ (symmetry), you get for what's inside the trace: $$ (I-R_z^{-1})\delta R_z = (I-R_z^{-1})(\mathcal{E}R_z+R_z\mathcal{E}^T) = (I-R_z^{-1})\mathcal{E}R_z+ (I-R_z^{-1}) R_z^T\mathcal{E}^T) = (I-R_z^{-1})\mathcal{E}R_z+ (I-R_z^{-1})(\mathcal{E}R_z)^T. $$ Using the trace properties, and $(I-R_z^{-1}) = (I-R_z^{-1})^T$ (thanks again to symmetry of $R_z$) you get \begin{equation} \mathrm{Trace}\{(I-R_z^{-1})\mathcal{E}R_z+ (I-R_z^{-1})(\mathcal{E}R_z)^T\} = \mathrm{Trace}\{(I-R_z^{-1})\mathcal{E}R_z\}+\mathrm{Trace}\{(I-R_z^{-1})^T(\mathcal{E}R_z)^T\} = \mathrm{Trace}\{(I-R_z^{-1})\mathcal{E}R_z\}+\mathrm{Trace}\{(\mathcal{E}R_z)^T(I-R_z^{-1})^T\} = \mathrm{Trace}\{(I-R_z^{-1})\mathcal{E}R_z\}+\mathrm{Trace}\{((I-R_z^{-1})(\mathcal{E}R_z))^T\} = 2\mathrm{Trace}\{(I-R_z^{-1})\mathcal{E}R_z\} = \mathrm{Trace}\{2R_z(I-R_z^{-1})\mathcal{E}\} = \mathrm{Trace}\{2(R_z-I)\mathcal{E}\} = \mathrm{Trace}\{2(R_z-I)^T\mathcal{E}\} =<2(R_z-I)|\mathcal{E}> \end{equation}

Identifying this with Eq. (14), we find $$ \nabla\phi = 2(R_z-I) \tag{23} $$