On model reduction by L2-optimal pole retention