0
$\begingroup$

I am currently studying complex FastICA and the paper says that

Suppose $\mathbf{s}$ is a $n\times1$ complex random vector. If $\mathbf{s}$ has zero mean, unit variance, and uncorrelated real and imaginary part of equal variances, then $E[\mathbf{s}\mathbf{s}^H]=\mathbf{I}_n$ and $E[\mathbf{s}\mathbf{s}^T]=\mathbf{0}_n$.

I don't quite get how $E[\mathbf{s}\mathbf{s}^T]=\mathbf{0}_n$ come about from the conditions.

We have the covariance matrix as \begin{align} \operatorname{cov}(\mathbf{s}) &= E[\mathbf{s}\mathbf{s}^H]-E[\mathbf{s}]E[\mathbf{s}^H] \\ &= E[\mathbf{s}\mathbf{s}^H]-\mathbf{0}_{n\times1}\mathbf{0}_{1\times n}\\ &= E[\mathbf{s}\mathbf{s}^H]\\ \end{align} and the pseudocovariance \begin{align} \operatorname{pcov}(\mathbf{s}) &= E[\mathbf{s}\mathbf{s}^T]-E[\mathbf{s}]E[\mathbf{s}^T] \\ &= E[\mathbf{s}\mathbf{s}^T]-\mathbf{0}_{n\times1}\mathbf{0}_{1\times n}\\ &= E[\mathbf{s}\mathbf{s}^T]\\ \end{align} I don't quite get how to equate the last line of covariance matrix to identity and the pseucovariance to zero.

If I were to write out the matrix, \begin{align} E[\mathbf{s}\mathbf{s}^H] &=E\left\{\begin{bmatrix} s_1s_1^* & s_1s_2^* &\cdots & s_1s_n^*\\ s_2s_1^* & s_2s_2^* &\cdots & s_2s_n^*\\ \vdots & \vdots &\ddots & \vdots\\ s_ns_1^* & s_ns_2^* &\cdots & s_ns_n^*\\ \end{bmatrix}\right\} \end{align} and \begin{align} E[\mathbf{s}\mathbf{s}^T] &=E\left\{\begin{bmatrix} s_1s_1 & s_1s_2 &\cdots & s_1s_n\\ s_2s_1 & s_2s_2 &\cdots & s_2s_n\\ \vdots & \vdots &\ddots & \vdots\\ s_ns_1 & s_ns_2 &\cdots & s_ns_n\\ \end{bmatrix}\right\} \end{align}

I still can't quite figure how all of these eventually becomes identity and zeros.

$\endgroup$
0
$\begingroup$

Okay now that I ponder for a few more hours.

I know for sure that $E[s_is_j^*]=E[s_i]E[s_j]=0$ and $E[s_is_j]=E[s_i]E[s_j]=0$ where $i\ne j$ so I have

\begin{align} E[\mathbf{s}\mathbf{s}^H] &=\operatorname{diag}(E[s_is_i^*])\\ &=\operatorname{diag}(E[(a+ib)(a-ib)]) &&(\because s_i:=a+ib)\\ &=\operatorname{diag}(E[a^2+b^2])\\ &=\operatorname{diag}(E[a^2]+E[b^2])\\ &=\operatorname{diag}(\operatorname{Var}[a]+E[a]^2+\operatorname{Var}[b]+E[b]^2) && (\because \operatorname{Var}[X]=E[X^2]-E[X]^2)\\ &=\operatorname{diag}(\operatorname{Var}[s_i]+0) && (\because E[a]=E[b]=0,\\ &&& \qquad\operatorname{Var}[s]:=\operatorname{Var}[a]+\operatorname{Var}[b])\\ &=\operatorname{diag}(1) && (\because \operatorname{Var}[s]:=1)\\ &=\mathbf{I}_n \end{align} and \begin{align} E[\mathbf{s}\mathbf{s}^T] &=\operatorname{diag}(E[s_is_i])\\ &=\operatorname{diag}(E[s_i^2])\\ &=\operatorname{diag}(E[a^2+2abi-b^2])\\ &=\operatorname{diag}(E[a^2]+2E[abi]-E[b^2])\\ &=\operatorname{diag}(E[a^2]-E[b^2])\\ &=\operatorname{diag}(\operatorname{Var}[a]-\operatorname{Var}[b])\\ &=\operatorname{diag}(0)\\ &=\mathbf{0}_n \end{align}

I will just leave my answer here in case someone were to be in my position in the future since I can't find any good reference for this.

$\endgroup$

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.