The Elements of Statistical Learning says on page 113:

Sphere the data with respect to the common covariance estimates $\hat{\Sigma}$: $X^* \leftarrow D^{-1/2}U^TX$ where $\hat{\Sigma} = UDU^T$. The common covariance estimate of $X^*$ will now be the identity.

Can someone help me understand why $X^* \leftarrow D^{-1/2}U^TX$ spheres the data?

  • 1
    $\begingroup$ Have you tried working out what the covariance of $X^{\star}$ will be? $\endgroup$ – cardinal Feb 8 '12 at 3:42

I think I figured out the answer after seeing cardinal's suggestion and reading the Wikipedia page on whitening.

$cov(X^*) = E[X^*X^{*T}]$

$= E[D^{-\frac{1}{2}}U^TXX^TUD^{-\frac{1}{2}T}]$

$D^{-\frac{1}{2}T} = D^{-\frac{1}{2}}$ because it's a diagonal matrix

$= D^{-\frac{1}{2}}U^TE[XX^T]UD^{-\frac{1}{2}}$

$= D^{-\frac{1}{2}}U^T\hat{\Sigma}UD^{-\frac{1}{2}}$

$= D^{-\frac{1}{2}}U^TUDU^TUD^{-\frac{1}{2}}$

The $U^TU = 1$ because U's have unit length.

$= D^{-\frac{1}{2}}DD^{-\frac{1}{2}}$

$= I$


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.