OctoInformBeta

OpenAI preprint, September 2026 · §4. Spreading Fourier coefficients

Curvatures, spreading and truncation: Products ∏j|fi(bj)|\prod_j|f_i(b_j)| tuned to positive targets; for large BB, bounds on spread coefficients, their boxes, their decay and the truncated sum

Tuning curvature products and bounding quadratically spread Fourier coefficients.

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 real trigonometric polynomial 2Re⁡∑s∈Scse(s⋅y)2\operatorname{Re}\sum_{s\in S}c_s\,e(s\cdot y) on the torus 𝕋n\mathbb T^n, with a finite set S⊂ℤnS\subset\mathbb Z^n of modes, coefficients cs∈ℂc_s\in\mathbb C and e(x)=e2πixe(x)=e^{2\pi ix}. Spreading it along vectors bj∈ℝnb_j\in\mathbb R^n, for jj in a finite set JJ, multiplies each mode e(s⋅y)e(s\cdot y) by a product of Fourier series in new circle variables zjz_j, built from a fixed bump gg on ℝ\mathbb R and quadratic phases x↦e(tx2/2−ux)x\mapsto e(tx^2/2-ux) of curvature t=Bs⋅bjt=B\,s\cdot b_j, where BB is a large parameter. One asks how small the coefficients of the spread polynomial are, how many frequencies each mode occupies, how fast the coefficients decay outside the frequencies uu with |uj|≤B|s⋅bj|/2|u_j|\le B\,|s\cdot b_j|/2 for every j∈Jj\in J, and whether, for a given δ>0\delta>0, the polynomial cut to those frequencies has modulus at most 1+δ1+\delta.

The vectors bjb_j are at one’s disposal, and a prior question is whether they can be chosen so that the products ∏j∈J|s⋅bj|\prod_{j\in J}|s\cdot b_j|, one for each mode ss, stand in prescribed proportions.

The Spread Polynomial

The polynomial that is spread is the binary cosine form. For each natural number nn, the binary cosine form Φn\Phi_n is the real function on 𝕋n\mathbb T^n defined by Φ0=0\Phi_0=0 and Φn+1(x)=Φn(x′)+12(1−Φn(x′)2)cos⁡(2πx0),\Phi_{n+1}(x)=\Phi_n(x')+\tfrac12\bigl(1-\Phi_n(x')^2\bigr)\cos(2\pi x_0),(1)✦ where x=(x0,…,xn)∈𝕋n+1x=(x_0,\dots,x_n)\in\mathbb T^{n+1} and x′=(x1,…,xn)x'=(x_1,\dots,x_n) is its tail.

The row form of an integer vector s∈ℤns\in\mathbb Z^n is the linear form x↦s⋅x=∑k<nskxkx\mapsto s\cdot x=\sum_{k<n}s_kx_k on ℝn\mathbb R^n.✦ For a mode s∈ℤns\in\mathbb Z^n and a velocity v∈ℝnv\in\mathbb R^n, the curvature of ss along vv is the number s⋅vs\cdot v.

The quadratic phase of curvature tt and linear frequency uu, for real tt and uu, is the function x↦e(tx2/2−ux)x\mapsto e(tx^2/2-ux) on ℝ\mathbb R.✦ For an integrable function gg on ℝ\mathbb R we write gˆt(u)=∫ℝg(x)e(tx2/2−ux)dx\hat g_t(u)=\int_{\mathbb R}g(x)\,e\bigl(tx^2/2-ux\bigr)\,dx(2) for its integral against that phase, the Fourier transform at uu of gg times the chirp e(tx2/2)e(tx^2/2).

The box UT(β)U_T(\beta) of real numbers TT and β\beta is the set of integers ww with |w|≤T|β|/2|w|\le T|\beta|/2.✦ For a finite set II, the product box UT(β)U_T(\beta) of a vector β∈ℝI\beta\in\mathbb R^I is the set of integer vectors u∈ℤIu\in\mathbb Z^I with ui∈UT(βi)u_i\in U_T(\beta_i) for every i∈Ii\in I.✦ The determinant of the product box UT(β)U_T(\beta) is the number Δ(β)=∏i∈I|βi|\Delta(\beta)=\prod_{i\in I}|\beta_i|.✦

The truncated row of a function gg on ℝ\mathbb R, at scale TT and curvature β\beta, is the trigonometric polynomial RT,βgR_{T,\beta}\,g on 𝕋\mathbb T given by RT,βg(z)=∑w∈UT(β)gˆTβ(w)e(wz).R_{T,\beta}\,g(z)=\sum_{w\in U_T(\beta)}\hat g_{T\beta}(w)\,e(wz).(3)✦

Given a finite set S⊂ℤnS\subset\mathbb Z^n of modes with coefficients cs∈ℂc_s\in\mathbb C, a function gg on ℝ\mathbb R and vectors bj∈ℝnb_j\in\mathbb R^n for jj in a finite set JJ, the truncated spread polynomial at scale BB is the trigonometric polynomial on 𝕋n×𝕋J\mathbb T^n\times\mathbb T^J FB(y,z)=∑s∈Scse(s⋅y)∏j∈JRB,s⋅bjg(zj).F_B(y,z)=\sum_{s\in S}c_s\,e(s\cdot y)\prod_{j\in J}R_{B,\,s\cdot b_j}\,g(z_j).(4)

The Claim

Theorem 1.

  1. Let VV be a finite-dimensional real vector space, let (fi)i∈I(f_i)_{i\in I} be a finite family of nonzero linear forms on VV, no two of them proportional, and let Ai>0A_i>0 for i∈Ii\in I and 0<δ<10<\delta<1. Then there are a number τ>0\tau>0 and vectors bj∈Vb_j\in V, one for each j∈Ij\in I, with fi(bj)≠0f_i(b_j)\neq0 for all i,j∈Ii,j\in I and (1−δ)τAi≤∏j∈I|fi(bj)|≤(1+δ)τAi(i∈I).(1-\delta)\,\tau A_i\le\prod_{j\in I}|f_i(b_j)|\le(1+\delta)\,\tau A_i \qquad(i\in I).(5)✦

  2. Let gg be a smooth, compactly supported complex function on ℝ\mathbb R with |g|≤1|g|\le1, let β∈ℝI\beta\in\mathbb R^I, for a finite set II, have no zero entry, and let ε>0\varepsilon>0. Then there is B0B_0 such that for all B≥B0B\ge B_0 and all u∈ℤIu\in\mathbb Z^I B|I|Δ(β)|∏i∈IgˆBβi(ui)|2≤1+ε.B^{|I|}\,\Delta(\beta)\,\Bigl|\prod_{i\in I}\hat g_{B\beta_i}(u_i)\Bigr|^2\le1+\varepsilon.(6)✦

  3. Let β∈ℝI\beta\in\mathbb R^I, for a finite set II, have no zero entry, and let ε>0\varepsilon>0. Then there is B0B_0 such that for all B≥B0B\ge B_0 the product box satisfies #UB(β)≤(1+ε)B|I|Δ(β).\#U_B(\beta)\le(1+\varepsilon)\,B^{|I|}\,\Delta(\beta).(7)✦

  4. Let gg be a smooth complex function on ℝ\mathbb R supported in [−α,α][-\alpha,\alpha], where 0≤α<1/20\le\alpha<1/2, let β≠0\beta\neq0 be real and let ll be a natural number. Then there is a constant C≥0C\ge0 such that for all real T>0T>0 and ww with |w|>T|β|/2|w|>T|\beta|/2 (T+|w|)l|gˆTβ(w)|≤C.(T+|w|)^l\,\bigl|\hat g_{T\beta}(w)\bigr|\le C.(8)✦

  5. Let S⊂ℤnS\subset\mathbb Z^n be a finite set and cs∈ℂc_s\in\mathbb C coefficients with Φn(y)=2Re⁡∑s∈Scse(s⋅y)\Phi_n(y)=2\operatorname{Re}\sum_{s\in S}c_s\,e(s\cdot y) for every y∈𝕋ny\in\mathbb T^n. Let gg be a smooth function on ℝ\mathbb R with 0≤g≤10\le g\le1, supported in [−α,α][-\alpha,\alpha] where 0≤α<1/20\le\alpha<1/2, let bj∈ℝnb_j\in\mathbb R^n, for jj in a finite set JJ, satisfy s⋅bj≠0s\cdot b_j\neq0 for all s∈Ss\in S and j∈Jj\in J, and let δ>0\delta>0. Then there is B0B_0 such that for all B≥B0B\ge B_0 the truncated spread polynomial (4) satisfies 2|Re⁡FB(y,z)|≤1+δ(y∈𝕋n,z∈𝕋J).2\,\bigl|\operatorname{Re}F_B(y,z)\bigr|\le1+\delta \qquad(y\in\mathbb T^n,\ z\in\mathbb T^J).(9)✦

In the paper, the result reads: the products Δs=∏j|s⋅bj|\Delta_s=\prod_j|s\cdot b_j| lie within a factor 1±δ1\pm\delta of τ|cs|2/ρs\tau|c_s|^2/\rho_s, for the paper’s numbers ρs\rho_s, the spread coefficients obey |Gˆ(s,u)|≤(1+δ)|cs|/BDΔs|\hat G(s,u)|\le(1+\delta)|c_s|/\sqrt{B^D\Delta_s}, the boxes UsU_s have at most (1+δ)BDΔs(1+\delta)B^D\Delta_s points, and the off-box coefficients decay with a sum tending to zero, where Gˆ(s,u)\hat G(s,u) is the coefficient of the spread polynomial at the mode ss and the frequency vector uu, UsU_s is the box of frequency vectors of the mode ss, and DD is the number of vectors bjb_j.

What Is Checked

The five parts are stated over linear forms on a finite-dimensional real vector space and over the library’s binary cosine form (1), curvatures s⋅bjs\cdot b_j of row forms, integrals (2) against quadratic phases, boxes UT(β)U_T(\beta) with their products and determinants, and truncated rows (3).