Estimates of the Approximation Error via Rademacher Complexity: Learning Vector-Valued Functions