Tag
The paper analyzes why clipping is crucial in AdaGrad under generalized smoothness and heavy-tailed noise, proving it prevents anisotropic miscalibration and provides high-probability convergence guarantees.