確率的二段階最適化に必要な勾配問い合わせ回数の下限

We study the oracle complexity of finding $\epsilon$-stationary points of smooth bilevel optimization problems with a nonconvex upper-level objective and a strongly convex lower-level problem. We consider a stochastic first-order oracle that returns unbiased stochastic gradients of both the upper- and lower-level objectives, with variance bounded by $\sigma^2$. We prove that, for any initial optimality gap $\Delta>0$ and all sufficiently small $\epsilon>0$, every adaptive randomized first-order algorithm requires $\Omega\!\left(\Delta \kappa_y^2 \epsilon^{-2}\max\{1,\sigma^2\kappa_y^6\epsilon^{-4}\}\right)$ oracle queries to find an $\epsilon$-stationary point of its hyper-objective function, where $\kappa_y$ denotes the condition number of the lower-level problem. In particular, in the noise-dominated regime, the lower bound is $\Omega\!\left(\Delta\sigma^2\kappa_y^8\epsilon^{-6}\right)$. This establishes the optimality of the $\epsilon^{-6}$ dependence achieved by the best-known first-order stochastic methods.