ProbXiv
Explorersign in
Problem archiveProblem record

Statement

Consider asymptotic hypothesis-testing problems between probability measures mathbbPmathbb P and mathbbQmathbb Q. For degree at most DD, define the low-degree advantage by

Adv≤D(dPdQ):=sup⁡f∈PDEP[f]EQ[f2].\mathsf{Adv}_{\leq D}\left(\frac{\mathrm{d}\mathbb P}{\mathrm{d}\mathbb Q}\right):=\sup_{f\in\mathcal P_D}\frac{\mathbb E_{\mathbb P}[f]}{\sqrt{\mathbb E_{\mathbb Q}[f^2]}}.

Here mathcalPDmathcal P_D is the set of real polynomials of degree at most DD, and TV\mathrm{TV} denotes total variation distance. An algorithm achieves strong detection if the sum of its type-I and type-II errors tends to zero, and achieves weak detection if that sum is bounded above by 1−ϵ1-\epsilon for some fixed epsilon>0epsilon>0. Hopkins's low-degree conjecture. For “natural” high-dimensional hypothesis-testing problems between mathbbPmathbb P and mathbbQmathbb Q, the following statements hold: (1) if

Adv≤D(dP′dQ′)=O(1)\mathsf{Adv}_{\leq D}\left(\frac{\mathrm{d}\mathbb P'}{\mathrm{d}\mathbb Q'}\right)=O(1)

as n→∞n\to\infty for some mathbbP′,mathbbQ′mathbb P',mathbb Q' satisfying TV(mathbbP,mathbbP′),TV(mathbbQ,mathbbQ′)=o(1)\mathrm{TV}(mathbb P,mathbb P'),\mathrm{TV}(mathbb Q,mathbb Q')=o(1), then there exists a constant CC such that no algorithm with running time nD/(log⁡n)Cn^{D/(\log n)^C} achieves strong detection between mathbbPmathbb P and mathbbQmathbb Q; and (2) if

Adv≤D(dP′dQ′)=1+o(1)\mathsf{Adv}_{\leq D}\left(\frac{\mathrm{d}\mathbb P'}{\mathrm{d}\mathbb Q'}\right)=1+o(1)

under the same total-variation conditions, then there exists a constant CC such that no algorithm with running time nD/(log⁡n)Cn^{D/(\log n)^C} achieves weak detection between mathbbPmathbb P and mathbbQmathbb Q. The conjecture is a proposed bridge from low-degree polynomial analysis to computational lower bounds for natural high-dimensional testing problems. The source does not provide evidence resolving it in full generality.

Record

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.