fb2e203234df6dee15934e448ee88971-AuthorFeedback.pdf

Regarding the restrictions of the state-feedback case, we agree that the output12 feedback case isimportant. Wenotethateveninthe23 zero-sum LQ game context, our proofs are the first correct ones that carry rigorousrobust controlimplications; (c)24 Our algorithms are not "minor extensions" ofthose in[44]. For nonconvex32 optimization without additional problem structure, e.g., the gradient domination property of the objective for the33 inner-loop LQR, thisglobal sublinear rate is something one can hardly improve in general. But note that in our34 simulations (Figure 4), convergence of this double-loop algorithm is not that bad (sublinear only in the beginning).35 We will add the runtime discussions.