PAC learning theory and VC dimension provide a theoretical framework for understanding generalization; higher model complexity yields better fitting on the training set but may reduce generalization ability, which is the bias-variance tradeoff.