Descriptions may be stale. Consult the linked Lean sources.
Second-Order
Stationarity in PL Minimax Problems
We consider minimizing for a possibly nonconvex function on
finite-dimensional real inner product spaces . The gradient and Hessian of are Lipschitz with constants and , respectively; every inner
maximum is attained and satisfies a maximization Polyak–Łojasiewicz
condition with .
Initial value and maximization gaps are bounded. The oracle returns
, , and at . For , the goal is twice
differentiable and an output
satisfying
and
almost surely, where is the
identity and the probability concerns algorithm randomness.
The optimization problem is Equip with the product inner product
and norm All
operator norms below are the induced operator norms. The Hessian acts on , and denotes the component of . The algorithm receives
and the given points ,
, with parameter ranges
Our assumptions:
Assumption 1 (Smoothness). The function is Fréchet differentiable everywhere,
its gradient is Fréchet differentiable everywhere, and for all ,
Assumption 2 (Attainment and maximization PL
condition). For every , there
exists such that For all and
,
Assumption 3 (Initial gaps). The given points
satisfy
Queries may be made at any point . Each oracle call returns the exact triple The required success condition includes Fréchet
differentiability of both and
everywhere on . After the allotted oracle calls, the
output must satisfy
where is the law of the algorithm’s
randomness.