Which SVM kernel to use when there is a structured covariance among the features?

Question

I am trying to use SVM for multi-class classification. The input features are assumed to be generated from a multi-variate Gaussian distribution. Each class corresponds to a