A theoretical guarantee that bounds a model's test error based on its training error and complexity, used to formally connect geometry to generalization.