log Cρ dual!, whereLρ = λmax(W) µ+ρλmax(W),µρ = λ+min(W) L+ρλ+min(W). (Eq. 5onpage5) Ignoringthelogfactor,thisquantityisproportionaltotheregularizedconditionnumber q

To apply6 variance reduction methods, we need to further assume that the loss function on each node can be decomposed7 into a finite-sum structure, say each loss is a sum ofmfunctions. In short, we have applied the same warm-start strategy as21 IDEAL+AGD in the complexity analysis of SSDA+AGD, the higher computation cost of SSDA is due to an22 intrinsic weakness of the method.