Approximation accuracy, gradient methods, and error bound for structured convex optimization