Separating populations with wide data: A spectral analysis