OctoInformBeta

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

Blocks, the uniform maximum and the mean square: Coefficient vectors XN∈[−1,1]NX_N \in [-1,1]^N with |QXN|≤(K+C/N)N|Q_{X_N}| \le (K + C/N)\sqrt N on the circle and energy limit at least (1−δ)∑i|ci|2(1-\delta)\sum_i |c_i|^2

Sampling a torus polynomial in blocks controls maximum and energy.

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

Write e(x)=exp⁡(2πix)e(x) = \exp(2\pi i x) and let 𝕋=ℝ/ℤ\mathbb{T} = \mathbb{R}/\mathbb{Z} be the circle, parametrized by t∈[0,1]t \in [0,1]. A real coefficient vector X=(X0,…,XN−1)X = (X_0, \dots, X_{N-1}) in [−1,1]N[-1,1]^N has the polynomial QX(t)=∑k<NXke(kt)Q_X(t) = \sum_{k<N} X_k e(kt), whose mean square over the circle is ∑kXk2\sum_k X_k^2 by Parseval’s identity. The question is how small the maximum of |QX|/N|Q_X|/\sqrt N on the circle can be kept while the energy 1N∑kXk2\frac1N \sum_k X_k^2 stays large.

Here the vectors come from a trigonometric polynomial on a torus. Let SS and JJ be finite sets, and for each i∈Si \in S let ai∈ℤJa_i \in \mathbb{Z}^J be an integer frequency vector and ci∈ℂc_i \in \mathbb{C} a coefficient, giving f(y)=∑i∈Scie(ai⋅y),y∈𝕋J.f(y) = \sum_{i \in S} c_i\, e(a_i \cdot y), \qquad y \in \mathbb{T}^J.(1) Call aia_i the ii-th mode of ff; for a velocity v∈ℝJv \in \mathbb{R}^J its curvature is λi=ai⋅v\lambda_i = a_i \cdot v. An integer H≥1H \ge 1 cuts [0,1][0,1] into the blocks (h/H,(h+1)/H)(h/H, (h+1)/H), 0≤h<H0 \le h < H, and each block carries a shift θh∈ℝJ\theta_h \in \mathbb{R}^J.

The library states the sampler that turns these data into coefficient vectors, together with the seven facts behind it, from smooth cutoffs on the blocks to vanishing averages of sampled quadratic phases.

Circle Intervals, Quadratic Phases and Modes

For a point cc of 𝕋\mathbb{T} and a real number ℓ\ell, the closed interval I(c,ℓ)I(c,\ell) of the circle with center cc and length ℓ\ell is the set I(c,ℓ)={c+xmod⁡1:x∈ℝ,|x|≤ℓ/2}.I(c,\ell) = \{\, c + x \bmod 1 : x \in \mathbb{R},\ |x| \le \ell/2 \,\}.(2)✦

The quadratic phase of curvature tt and linear frequency uu is the function x↦e(tx2/2−ux)x \mapsto e(tx^2/2 - ux) of a real variable xx.✦

A family of modes indexed by a finite set PP assigns to each mode i∈Pi \in P a profile gi:ℝ→ℂg_i \colon \mathbb{R} \to \mathbb{C}, an amplitude ci∈ℂc_i \in \mathbb{C}, a curvature λi\lambda_i, a frequency bib_i and an offset did_i, the last three real. The mode sum of length NN of the family is the vector MN∈ℂNM_N \in \mathbb{C}^N with entries MN(k)=∑i∈Pcie(Ndi)gi(k/N)e(λik22N+bik),0≤k<N.M_N(k) = \sum_{i \in P} c_i\, e(N d_i)\, g_i(k/N)\, e\Bigl(\frac{\lambda_i k^2}{2N} + b_i k\Bigr), \qquad 0 \le k < N.(3)✦

For a mode ii of nonzero curvature and an angle t∈ℝt \in \mathbb{R}, the stationary points of the mode at tt are the real numbers xi,ℓ(t)=ℓ−bi−tλi,ℓ∈ℤ.x_{i,\ell}(t) = \frac{\ell - b_i - t}{\lambda_i}, \qquad \ell \in \mathbb{Z}.(4)

The Claim

Theorem 1. (i) Let ff in (1) be real-valued with |f(y)|≤1|f(y)| \le 1 for every y∈𝕋Jy \in \mathbb{T}^J, let 0<|λi|<H0 < |\lambda_i| < H for every i∈Si \in S, and let the arcs I(ai⋅θh,|λi|/H)I(a_i \cdot \theta_h, |\lambda_i|/H), for 0≤h<H0 \le h < H and i∈Si \in S, be pairwise disjoint. Let K≥0K \ge 0 with |ci|≤K|λi||c_i| \le K\sqrt{|\lambda_i|} for every i∈Si \in S, and let δ>0\delta > 0. Then there are vectors XN=(XN,0,…,XN,N−1)X_N = (X_{N,0}, \dots, X_{N,N-1}) in [−1,1]N[-1,1]^N for N≥1N \ge 1 and constants C≥0C \ge 0 and L≥(1−δ)∑i∈S|ci|2L \ge (1-\delta)\sum_{i \in S} |c_i|^2 with limN→∞1N∑k<NXN,k2=Land|QXN(t)|N≤K+CN(N≥1,t∈[0,1]).\lim_{N \to \infty} \frac1N \sum_{k<N} X_{N,k}^2 = L \qquad\text{and}\qquad \frac{|Q_{X_N}(t)|}{\sqrt N} \le K + \frac CN \quad (N \ge 1,\ t \in [0,1]).(5)✦

(ii) For every integer H≥1H \ge 1 and every δ>0\delta > 0 there are smooth functions χ0,…,χH−1:ℝ→[0,1]\chi_0, \dots, \chi_{H-1} \colon \mathbb{R} \to [0,1] with supp⁡χh⊂(h/H,(h+1)/H)\operatorname{supp}\chi_h \subset (h/H, (h+1)/H) for each hh and ∑h<H∫ℝχh(x)2dx>1−δ.\sum_{h<H} \int_{\mathbb{R}} \chi_h(x)^2\,dx > 1 - \delta.(6)✦

(iii) If s0,…,sH−1∈[0,1]s_0, \dots, s_{H-1} \in [0,1] satisfy shsh′=0s_h s_{h'} = 0 whenever h≠h′h \ne h', and w0,…,wH−1w_0, \dots, w_{H-1} are complex numbers of modulus at most 11, then |∑h<Hshwh|≤1.\Bigl|\sum_{h<H} s_h w_h\Bigr| \le 1.(7)✦

(iv) Let H≥1H \ge 1, and let each member ii of a set PP be given a block hi∈{0,…,H−1}h_i \in \{0, \dots, H-1\}, a function gi:ℝ→ℂg_i \colon \mathbb{R} \to \mathbb{C} with supp⁡gi⊂(hi/H,(hi+1)/H)\operatorname{supp} g_i \subset (h_i/H, (h_i+1)/H), a center αi∈ℝ\alpha_i \in \mathbb{R} and a curvature λi\lambda_i with 0<|λi|<H0 < |\lambda_i| < H. If the arcs I(αi,|λi|/H)I(\alpha_i, |\lambda_i|/H), i∈Pi \in P, are pairwise disjoint, then for the frequencies bi=αi−λi(hi+1/2)/Hb_i = \alpha_i - \lambda_i(h_i + 1/2)/H the stationary points (4) satisfy, for every t∈ℝt \in \mathbb{R}, #{(i,ℓ)∈P×ℤ:gi(xi,ℓ(t))≠0}≤1.\#\{(i,\ell) \in P \times \mathbb{Z} : g_i(x_{i,\ell}(t)) \ne 0\} \le 1.(8)✦

(v) Let K≥0K \ge 0, and let a family of modes indexed by a finite set PP have smooth profiles with supp⁡gi⊂(0,1)\operatorname{supp} g_i \subset (0,1) and |gi|≤1|g_i| \le 1, nonzero curvatures, and amplitudes with |ci|≤K|λi||c_i| \le K\sqrt{|\lambda_i|}. If at every angle t∈ℝt \in \mathbb{R} at most one pair (i,ℓ)∈P×ℤ(i,\ell) \in P \times \mathbb{Z} has gi(xi,ℓ(t))≠0g_i(x_{i,\ell}(t)) \ne 0, then there is a constant C≥0C \ge 0 with 1N|∑k<NMN(k)e(kt)|≤K+CN(N≥1,t∈[0,1]).\frac{1}{\sqrt N} \Bigl|\sum_{k<N} M_N(k)\, e(kt)\Bigr| \le K + \frac CN \qquad (N \ge 1,\ t \in [0,1]).(9)✦

(vi) Let a family of modes indexed by a finite set PP have smooth profiles with supp⁡gi⊂(0,1)\operatorname{supp} g_i \subset (0,1) and |gi|≤1|g_i| \le 1, and let any two distinct modes ii and jj have λi≠λj\lambda_i \ne \lambda_j, or bi−bj∉ℤb_i - b_j \notin \mathbb{Z}, or gigj¯≡0g_i \overline{g_j} \equiv 0. Then limN→∞1N∑k<N|MN(k)|2=∑i∈P|ci|2∫ℝ|gi(x)|2dx.\lim_{N \to \infty} \frac1N \sum_{k<N} |M_N(k)|^2 = \sum_{i \in P} |c_i|^2 \int_{\mathbb{R}} |g_i(x)|^2\,dx.(10)✦

(vii) Let g:ℝ→ℂg \colon \mathbb{R} \to \mathbb{C} be smooth, with support in [−A,A][-A, A] and |g|≤D|g| \le D for some A,D≥0A, D \ge 0, and let β≠0\beta \ne 0 and bb be real. Then limN→∞1N∑k∈ℤg(k/N)e(βk22N+bk)=0.\lim_{N \to \infty} \frac1N \sum_{k \in \mathbb{Z}} g(k/N)\, e\Bigl(\frac{\beta k^2}{2N} + bk\Bigr) = 0.(11)✦

(viii) Let g:ℝ→ℂg \colon \mathbb{R} \to \mathbb{C} be smooth with support in [−A,A][-A, A] for some A≥0A \ge 0, and let bb be a real number that is not an integer. Then limN→∞1N∑k∈ℤg(k/N)e(bk)=0.\lim_{N \to \infty} \frac1N \sum_{k \in \mathbb{Z}} g(k/N)\, e(bk) = 0.(12)✦

In the paper, with FF the polynomial ff of (1), the result reads: there are cutoffs χh\chi_h supported inside the hh-th block with ∑h∫χh2≥1−δ\sum_h \int \chi_h^2 \ge 1 - \delta; for points xh*x_h^*, the coefficients XN,k=∑hχh(k/N)F(kθh+N2v(k/N−xh*)2)X_{N,k} = \sum_h \chi_h(k/N)\, F\bigl(k\theta_h + \frac N2 v (k/N - x_h^*)^2\bigr) lie in [−1,1][-1,1]; at each angle at most one stationary pair (a,h,ℓ)(a, h, \ell) survives, with aa a frequency vector of ff, which gives the uniform maximum bound; and the off-diagonal sampled averages vanish, also for equal curvatures with distinct centers, which gives the energy limit.

What Is Checked

The theorem is stated over the exponential e(x)e(x), the characters y↦e(a⋅y)y \mapsto e(a \cdot y) of the torus 𝕋J\mathbb{T}^J that make up ff, the closed circle intervals (2), the quadratic phases and the mode sums (3).