OctoInformBeta

OpenAI preprint, September 2026 · §5. Sampling the auxiliary polynomial

Lemma 5.1: For smooth compactly supported gg and β≠0\beta \ne 0, the smoothed quadratic sum equals finitely many stationary waves up to C/NC/N, uniformly for |b|≤B|b| \le B

A uniform stationary phase expansion of smoothed quadratic sums.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Take a smooth function gg on the real line, with complex values and compact support, a real number β≠0\beta \ne 0 and a real number bb. At a real scale N≥1N \ge 1, sample gg at the points k/Nk/N with k∈ℤk \in \mathbb{Z}, and weight the sample at k/Nk/N by a phase quadratic in kk, with curvature β/N\beta/N tending to zero as NN grows and linear part bkbk; normalized by N−1/2N^{-1/2}, the result is a smoothed quadratic exponential sum. The question is the behavior of this sum as NN grows, uniformly for bb in a bounded interval; the answer is a stationary phase expansion to order 1/N1/N.

Quadratic Phases and Stationary Waves

For a real number θ\theta, the phase eiθe^{i\theta} is the point of the unit circle at angle θ\theta, and for real xx, e(x)=e2πixe(x) = e^{2\pi i x} is the phase at angle 2πx2\pi x.✦

The quadratic phase of curvature t∈ℝt \in \mathbb{R} and linear frequency u∈ℝu \in \mathbb{R} is the function φt,u:ℝ→ℂ\varphi_{t,u} \colon \mathbb{R} \to \mathbb{C} given by φt,u(x)=e(tx22−ux).\varphi_{t,u}(x) = e\Bigl(\frac{t x^2}{2} - u x\Bigr).(1)✦

The smoothed quadratic sum SN(b)S_N(b) of a compactly supported function g:ℝ→ℂg \colon \mathbb{R} \to \mathbb{C}, for a real β\beta, a real scale N≥1N \ge 1 and a real linear frequency bb, is the complex number SN(b)=1N∑k∈ℤg(kN)φβ/N,−b(k)=1N∑k∈ℤg(kN)e(βk22N+bk).S_N(b) = \frac{1}{\sqrt{N}} \sum_{k \in \mathbb{Z}} g\Bigl(\frac{k}{N}\Bigr)\, \varphi_{\beta/N,\,-b}(k) = \frac{1}{\sqrt{N}} \sum_{k \in \mathbb{Z}} g\Bigl(\frac{k}{N}\Bigr)\, e\Bigl(\frac{\beta k^2}{2N} + b k\Bigr).(2)

For real numbers NN, bb and β≠0\beta \ne 0 and an integer ℓ\ell, the stationary wave Wℓ(N,b)W_\ell(N,b) of a function g:ℝ→ℂg \colon \mathbb{R} \to \mathbb{C} at ℓ\ell is the complex number Wℓ(N,b)=1|β|e(sgn⁡β8−N(ℓ−b)22β)g(ℓ−bβ).W_\ell(N,b) = \frac{1}{\sqrt{|\beta|}}\, e\Bigl(\frac{\operatorname{sgn}\beta}{8} - \frac{N(\ell - b)^2}{2\beta}\Bigr)\, g\Bigl(\frac{\ell - b}{\beta}\Bigr).(3)✦ The point (ℓ−b)/β(\ell - b)/\beta at which it samples gg is the stationary point of the phase of φβ,ℓ−b\varphi_{\beta,\,\ell - b}.

The Claim

Theorem 1. Let g:ℝ→ℂg \colon \mathbb{R} \to \mathbb{C} be a smooth function with compact support, let A≥0A \ge 0 be such that the support of gg lies in [−A,A][-A, A], and let β≠0\beta \ne 0 and B≥0B \ge 0 be real numbers. Then there are a finite set JJ of integers and a constant C≥0C \ge 0 such that, for every real N≥1N \ge 1 and every real bb with |b|≤B|b| \le B, the stationary wave Wℓ(N,b)W_\ell(N,b) vanishes for every integer ℓ∉J\ell \notin J, and the sum (2) satisfies |SN(b)−∑ℓ∈JWℓ(N,b)|≤CN.\Bigl| S_N(b) - \sum_{\ell \in J} W_\ell(N,b) \Bigr| \le \frac{C}{N}.(4)✦

In the paper, the result reads, with gg, β\beta and KK standing for its χ\chi, λ\lambda and JJ: for g∈Cc∞(ℝ)g \in C_c^\infty(\mathbb{R}), a point x*∈ℝx_* \in \mathbb{R}, β≠0\beta \ne 0 and a compact set K⊂ℝK \subset \mathbb{R}, with xb,ℓ=x*+(ℓ−b)/βx_{b,\ell} = x_* + (\ell - b)/\beta and ωb,ℓ=(b−ℓ)x*−(b−ℓ)2/(2β)\omega_{b,\ell} = (b - \ell)x_* - (b - \ell)^2/(2\beta), as N→∞N \to \infty through the integers, uniformly for b∈Kb \in K, 1N∑k∈ℤg(kN)e(bk+Nβ2(kN−x*)2)=1|β|∑ℓg(xb,ℓ)e(sgn⁡β8+Nωb,ℓ)+O(1N),\frac{1}{\sqrt{N}} \sum_{k \in \mathbb{Z}} g\Bigl(\frac{k}{N}\Bigr)\, e\Bigl(bk + \frac{N\beta}{2}\Bigl(\frac{k}{N} - x_*\Bigr)^2\Bigr) = \frac{1}{\sqrt{|\beta|}} \sum_{\ell} g(x_{b,\ell})\, e\Bigl(\frac{\operatorname{sgn}\beta}{8} + N\omega_{b,\ell}\Bigr) + O\Bigl(\frac{1}{N}\Bigr),(5) the sum on the right running over one fixed finite set of ℓ\ell.

The theorem above is this statement at x*=0x_* = 0 and K=[−B,B]K = [-B, B], taken over all real N≥1N \ge 1.

What Is Checked

The theorem is stated over the quadratic phase (1), which the sum (2) takes at curvature β/N\beta/N and linear frequency −b-b, and over the stationary waves (3), both written with the phase e(x)=e2πixe(x) = e^{2\pi i x}, for a smooth gg with support in [−A,A][-A, A], a real β≠0\beta \ne 0, every real N≥1N \ge 1 and every bb with |b|≤B|b| \le B.