A Risk Minimization Principle for a Class of Parzen Estimators