stochastic second-order methods achieve fast local convergence in strongly
convex optimization by using noisy Hessian estimates to precondition the
gradient. However, these methods typically reach superlinear convergence only
when the stochastic Hessian noise diminishes, increasing per