We present multivariate penalized least squares regression estimates. We use Vapnik-Chervonenkis (see Statistical Learning Theory 1998) theory and bounds on the covering numbers to analyze convergence of the estimates. We show strong consistency of the truncated versions of the estimates without any conditions on the underlying distribution.