Improving Certified Robustness via Statistical Learning with Logical Reasoning