Why do we whiten the data before running ICA? Why do we usually pre-whiten the data before doing independent components analysis (ICA), like making all eigenvalues equal? Doesn't that take away some information regarding the data?
 A: ICA assumes that the observed data $\mathbf x$ are a linear combination of independent components: $\mathbf x = \mathbf A \mathbf s$, where $\mathbf A$ is usually called mixing matrix.
Covariance matrix of the data is then given by $\mathbf A \mathbf A^\top$, because the covariance matrix of independent components $\mathbf s$ is identity (indeed, independent components must have zero correlation between each other and are assumed to be scaled to have unit variance each). This means that if the data are pre-whitened, then $\mathbf A$ must be an orthogonal matrix. This is often convenient.
Here is e.g. a quote from Hyvärinen & Oja, 2000, Independent Component Analysis: Algorithms and Applications, section 5.2, called "Whitenening":

Here we see that whitening reduces the number of parameters to be estimated. Instead of having to
  estimate the $n^2$ parameters that are the elements of the original matrix $\mathbf A$ , we only need to estimate
  the new, orthogonal mixing matrix $ \tilde{\mathbf A}$ . An orthogonal matrix contains $n ( n - 1)/ 2$ degrees of freedom.
  For example, in two dimensions, an orthogonal transformation is determined by a single angle parameter.
  In larger dimensions, an orthogonal matrix contains only about half of the number of parameters of an
  arbitrary matrix. Thus one can say that whitening solves half of the problem of ICA. Because whitening is
  a very simple and standard procedure, much simpler than any ICA algorithms, it is a good idea to reduce
  the complexity of the problem this way.

A: It does take away some information, but that information is irrelevant for ICA.  The whitening step make algorithms for ICA converge faster.
A: ICA is a invertible problem, so ideally W should be rotation matrix. If you Whiten the signal the covariant matrix will be rotation. So you can take advantage of Natural Stochastic Gradient Descent by using grassmann manifold or stiefel manifold which can promise fast coverage.
