We studied the subsampled Newton (SSN) method and its variants, GIANT and SSPN, and established stronger convergence guarantees than the prior works.
In particular, we showed that a sample size of s=~Θ(dγeff) suffices, where γ is the ℓ2 regularizati…