Status: open (2026-10-24). Area: analysis, differential geometry of densities · Prerequisites: implicit function theorem, tubular neighbourhoods, basic diffusion models · Related: Universal Approximation for Implicit Diffusion Learners §7, Implicit Diffusion Learners, Open Problems in Implicit Diffusion Learning (T1, I12), Channels and Polarity.
Sources: original to this vault (question formulated from the heuristic analysis of Universal Approximation for Implicit Diffusion Learners §7 and the MNIST observation of Open Problems in Implicit Diffusion Learning I12).
Question
A single smooth potential is used to answer queries in many directions: clamp some coordinates, solve for the others as stationary points of on the slice. For which potentials, and in particular for which smoothed data densities , are the answers isolated and nondegenerate for every query direction of interest, including directions with more free coordinates than the codimension of the data?
Setting
Split the coordinates as with , (a polarity). Given a clamp , anchors and anchor precisions , the answers are the solutions of
with positive definite (regular, stable answers). Let the data lie on a smooth compact submanifold of dimension with a smooth positive density on , and let be the density of , , .
Question. For , , small and a family of polarities:
- If and the coordinate slice meets transversally, are the answers near regular, with bounded below like ?
- If , the slice meets in an -dimensional set. Are the answers exactly the nondegenerate local modes of restricted to that set, and is then of order the curvature of , independent of ?
- What can be said when is uniform on part of (no in-manifold curvature)? Is a positive necessary, and how large?
What is known
- §7 of Universal Approximation for Implicit Diffusion Learners describes the three regimes (transverse slicing, conditional mode finding, soft-anchored projection) as a heuristic analysis and conjecture, not a proof.
- One energy network answers both MNIST directions, image → label () and label → image () (Open Problems in Implicit Diffusion Learning I12): an observation that both regimes occur in practice, not a theorem.
- For the pure distance potential the Hessian on is the projector onto the normal space, so (1) holds and (2) fails: the answers form a continuum.
What would count as an answer
Theorems for (1) and (2) with explicit dependence on , the reach of and the derivatives of ; or counterexamples (a manifold and density where (2) fails for small ). A sharp answer to (3) would give a rule for choosing .
Why it matters here
It decides whether “one model, every question” holds for diffusion learners, and how to set the anchor precisions that the code currently chooses by hand.
Leads
Laplace-type expansions of near (the normal part concentrates like , the tangential part carries ); the literature on the manifold hypothesis for score-based models; Morse theory on the slice.