Status: open (2026-10-24). Area: analysis, difference equations · Prerequisites: Fourier analysis on , norms · Related: Universal Approximation for Implicit Diffusion Learners §3, Implicit Diffusion Learners.
Sources: original to this vault (from Universal Approximation for Implicit Diffusion Learners §3).
Question
Given a smooth function on a compact set , shifts and positive weights , find a smooth with
How large must be? With two antithetic shifts (, ) the solution is an alternating series whose norm grows at most linearly in . For the slab recurrence used in the vault only gives an exponential bound . Is there always a solution with polynomial growth as the shifts shrink, and if not, for which shift configurations?
Setting
The equation arises when the training field of an implicit diffusion learner is evaluated at a finite set of noise draws at one noise level: the shifts are (up to scaling), and is the potential the learner must represent to reproduce a target relation exactly (Universal Approximation for Implicit Diffusion Learners §3). Fix , shifts of size , and ask for on a neighbourhood of with the equation holding on .
What is known
- : explicit alternating solution, linear growth (the note, §3).
- , equal weights: the slab recurrence has amplification factor per slab of width , giving the exponential bound; whether it is attained is open (the note marks this as a conjecture).
- If only approximation is needed, the unshifted rescaling has error in as (uniform continuity of ), which suffices for the universal approximation theorem as stated; for , and only for centred shifts () and (the note, §3). Exact solvability is a separate question.
What would count as an answer
A construction of with polynomial norm growth for generic shifts; or a lower bound showing exponential growth is unavoidable for some configuration (e.g. via the symbol having zeros near the real axis).
Why it matters here
It tells whether exact recovery of a relation by a finite-noise-draw training field is well-conditioned, which bears on how many noise draws per level are worth using.
Leads
The Fourier symbol : division by is well-conditioned where is bounded below; the question is about its zeros and the localisation to . Mean-periodic functions and convolution equations (Schwartz) are the classical setting.