Finite-sum nonconvex-strongly-concave minimax problems arise in adversarial training and distributionally robust learning. Known lower bounds for these problems cover only linear-span algorithms, zero-respecting algorithms, or a restricted proximal protocol. This leaves open whether unrestricted adaptivity or exact component proximal steps, which return global saddle points of regularized component problems, can improve the polynomial complexity. We show that they cannot. Consider $n\ge2$ components with joint mean-squared smoothness $L$ and condition number $κ\ge1$, and a starting point with primal gap at most $Δ$ and joint gradient norm at most $\sqrt{2LΔ}$. For $\varepsilon^2/(LΔ)$ below a universal constant, every randomized adaptive algorithm with exact component values, gradients, and proximal responses requires at least $Ω(n+\min\{\sqrt n\,κ,n^{3/4}\sqrtκ\}LΔ/\varepsilon^2)$ expected component queries in the worst case to find an $\varepsilon$-stationary point of the value function. A method that uses only component gradients attains this rate up to a factor $\log^2(2+κ)$, including its initialization cost. Zhou (arXiv:2610.04944) concurrently proves the same rate for zero-respecting first-order methods. For large condition numbers, our hard instances link primal stages through dual paths whose edges belong to random components, and with high probability a proximal query to a component reveals a path only up to the first edge owned by another component.
翻译:暂无翻译