ProbXiv
Explorersign in
Problem archiveProblem record

Statement

Fix 1<p<∞1<p<\infty and let q=p/(p−1)q=p/(p-1). Given a first-order local value--gradient oracle for an objective f:Rn→Rf:\mathbb{R}^n\to\mathbb{R} whose gradient is LL-Lipschitz from (Rn,∥⋅∥p)(\mathbb{R}^n,\|\cdot\|_p) to (Rn,∥⋅∥q)(\mathbb{R}^n,\|\cdot\|_q), an initial point x0x_0 with a solution x∗x^* satisfying ∥x0−x∗∥p≤R\|x_0-x^*\|_p\le R, and a target ε>0\varepsilon>0, determine whether there is a parameter-free algorithm that does not know LL, RR, or f∗f^* and returns a queried point x^\widehat{x} satisfying ∥∇f(x^)∥q≤ε\|\nabla f(\widehat{x})\|_q\le\varepsilon, with dimension-free oracle complexity governed by K‾=max⁡{1,LR/ε}\overline K=\max\{1,LR/\varepsilon\}. Under a nondegenerate secant initialization, the claimed bounds are Op(K‾1/2)O_p(\overline K^{1/2}) post-initialization queries for 1<p≤21<p\le2 and Op(K‾p/(p+2))O_p(\overline K^{p/(p+2)}) for p>2p>2, together with an additive calibration cost Op(log⁡(e+L/M0))O_p(\log(e+L/M_0)).

Record

Added
Links

Comments

Nothing has been published against this problem yet, and nothing has been said about it.

Sign in with an institutional address to take part in the discussion. Reading every thread stays open to everyone.

Sign in

Solve with an agent

Open the statement in a chat, with the problem and the ground rules already written into the prompt.

This opens a third-party site. Nothing is posted back to ProbXiv and nothing you write there is recorded here — what a model gives you is an attempt, which a person still has to check.