Pointwise Hessian Stationarity Lower Bound — Stationarity with Exact Hessian Access
Descriptions may be stale. Consult the linked Lean sources.
Stationarity with Exact Hessian Access · Lean proof
Stationarity with Exact
Hessian Access
We seek approximate stationarity of a possibly nonconvex function
, where
is a nonnegative integer. The
function is three times continuously differentiable, with
third-derivative norm bounded by at each point
, where and . At the reference point
, the objective gap above the
infimum is at most and
the gradient norm is at most . The oracle returns the exact
value, gradient, and Hessian. For , the goal is a
measurable output with
, where is the law of the algorithm’s
randomness.
The underlying optimization problem is We use the
Euclidean norm on .
Derivative norms are the induced
operator norms; in particular, is a linear operator on , and is a trilinear form. The
algorithm receives , and .
The parameter regime is restricted by Here is the least integer
greater than or equal to .
Our assumptions:
Assumption 1 (Third-derivative bound). The function
belongs to and
satisfies
Assumption 2 (Initial objective and gradient
control). At the reference point ,
Queries may be made at any . Each oracle call
returns After the allotted oracle calls, the output must be measurable
under , its gradient norm must be
integrable, and it must satisfy
Proved result
OracleTemplate.NonConvergence LongTails.Setups.PointwiseHessianStationarity.spec
LongTails.Theorems.PointwiseHessianStationarityLowerBound.budget
Relevant previous results
No qualifying prior reference verified; this is not a proof of novelty.