OctoInformBeta

OpenAI preprint, September 2026 · Section 5. Uniform moment bounds for Fourier lifts

Lemma 5.3: Fixed fibers of real functions and the Fourier expansion of frozen residue lifts

Freezing coordinates bounds moments and removes support.

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

Fixing some coordinates of a product leaves a function on the remaining coordinates. The relevant size is the cardinality of the fixed coordinate space, while exact coordinate support records which coordinates have zero conditional average.

Let JJ be a finite coordinate set, let XjX_j be finite sets, and write XS=∏j∈SXjX_S=\prod_{j\in S}X_j for S⊆JS\subseteq J and X=XJX=X_J. For a finite set YY and a function HH on YY, averages are uniform probability averages, with the empty average assigned value zero: 𝔼y∈YH(y)=1|Y|∑y∈YH(y).\mathbb E_{y\in Y}H(y)=\frac{1}{|Y|}\sum_{y\in Y}H(y).(1) The sum in (1) is a counting sum.

The coordinate replacement TB,z:X→XT_{B,z}:X\to X, for B⊆JB\subseteq J and z∈Xz\in X, is the map fixing the coordinates in BB at their values in zz: (TB,zx)j={zj,j∈B,xj,j∉B.(T_{B,z}x)_j= \begin{cases} z_j,&j\in B,\\ x_j,&j\notin B. \end{cases}(2)✦ A real function h:X→ℝh:X\to\mathbb R has exact coordinate support U⊆JU\subseteq J if its average in each coordinate of UU is zero and it is invariant under every replacement outside UU; writing xj←tx^{j\leftarrow t} for coordinate replacement, these conditions are 𝔼t∈Xjh(xj←t)=0(j∈U),h(xj←t)=h(x)(j∉U),\mathbb E_{t\in X_j}h(x^{j\leftarrow t})=0\quad(j\in U), \qquad h(x^{j\leftarrow t})=h(x)\quad(j\notin U),(3)✦ for every x∈Xx\in X and every replacement value in the indicated coordinate. We denote the property in (3) by Exact⁡U(h)\operatorname{Exact}_U(h).

Residue Lifts and Fourier Coefficients

Fix positive integer moduli pjp_j, indexed by JJ. The product residue space XpX_p is the product of the residue class rings ℤ/pjℤ\mathbb Z/p_j\mathbb Z, and qSq_S is the cardinality of the coordinate subproduct indexed by S⊆JS\subseteq J: Xp=∏j∈Jℤ/pjℤ,qS=∏j∈Spj.X_p=\prod_{j\in J}\mathbb Z/p_j\mathbb Z, \qquad q_S=\prod_{j\in S}p_j.(4) On the space in (4), coordinate freezing is the map (2).✦ The unfrozen modulus family is (pj)j∉B(p_j)_{j\notin B}✦, and the outside coordinate assembly AB,z:Xp,J\B→XpA_{B,z}:X_{p,J\setminus B}\to X_p is defined by Xp,S=∏j∈Sℤ/pjℤ,(AB,zy)j={zj,j∈B,yj,j∉B.X_{p,S}=\prod_{j\in S}\mathbb Z/p_j\mathbb Z, \qquad (A_{B,z}y)_j= \begin{cases} z_j,&j\in B,\\ y_j,&j\notin B. \end{cases}(5)✦

For a positive integer qq and a support U⊆JU\subseteq J, the centered residue delta Kq:ℤ/qℤ→ℝK_q:\mathbb Z/q\mathbb Z\to\mathbb R and the centered residue product KU:Xp→ℝK_U:X_p\to\mathbb R are Kq(t)=q𝟏{0}(t)−1,KU(z)=∏j∈UKpj(zj).K_q(t)=q\mathbf 1_{\{0\}}(t)-1, \qquad K_U(z)=\prod_{j\in U}K_{p_j}(z_j).(6) For a natural length LL and a real function f:ℕ→ℝf:\mathbb N\to\mathbb R, the residue lift component ℒUf:Xp→ℝ\mathcal L_Uf:X_p\to\mathbb R is the normalized counting sum built from (6): ℒUf(z)=1L∑0≤n<Lf(n)KU(z−n_),n_=(nmod⁡pj)j∈J.\mathcal L_Uf(z)=\frac1L\sum_{0\le n<L}f(n)K_U(z-\underline n), \qquad \underline n=(n\bmod p_j)_{j\in J}.(7)✦ The sum in (7) is zero at L=0L=0; the public integer index on {1,…,L}\{1,\ldots,L\} is n+1n+1. For a family ℱ\mathcal F of complete supports U⊆JU\subseteq J, the partial lift ℒℱf:Xp→ℝ\mathcal L^{\mathcal F}f:X_p\to\mathbb R is the sum of the corresponding residue lift components: ℒℱf=∑U∈ℱℒUf.\mathcal L^{\mathcal F}f=\sum_{U\in\mathcal F}\mathcal L_Uf.(8)

The exponential phase is e(t)=exp⁡(2πit)e(t)=\exp(2\pi i t).✦ For a,z∈Xpa,z\in X_p, the product character χa\chi_a, frequency angle θ(a)\theta(a), and nonzero coordinate support supp⁡(a)\operatorname{supp}(a) are defined by χa(z)=∏j∈Je(ajzj̃pj),θ(a)=∑j∈Jaj̃pj,supp⁡(a)={j∈J:aj≠0},\chi_a(z)=\prod_{j\in J}e\!\left(\frac{\widetilde{a_jz_j}}{p_j}\right), \qquad \theta(a)=\sum_{j\in J}\frac{\widetilde{a_j}}{p_j}, \qquad \operatorname{supp}(a)=\{j\in J:a_j\ne0\},(9) where tildes denote least nonnegative representatives. The Fourier polynomial 𝒫c:Xp→ℂ\mathcal P_c:X_p\to\mathbb C with coefficients c:Xp→ℂc:X_p\to\mathbb C is the sum of the product characters weighted by those coefficients: 𝒫c(x)=∑a∈Xpc(a)χa(x).\mathcal P_c(x)=\sum_{a\in X_p}c(a)\chi_a(x).(10) For h:ℕ→ℂh:\mathbb N\to\mathbb C, a support S⊆JS\subseteq J, and residue points z,a∈Xpz,a\in X_p, the gate Fourier coefficient CL,h,S(z;a)C_{L,h,S}(z;a) is the complex number given by the normalized sum using the phase and angle in (9): CL,h,S(z;a)=1L∑0≤n<Lh(n)e(−nθ(a))KS(z−n_).C_{L,h,S}(z;a)=\frac1L\sum_{0\le n<L} h(n)e(-n\theta(a))K_S(z-\underline n).(11) The normalized sum in (11) is assigned value zero at L=0L=0.

The Claim

Theorem 1. Let F:X→ℝF:X\to\mathbb R be nonnegative, let B⊆JB\subseteq J, and fix z∈Xz\in X. Then 𝔼x∈XF(TB,zx)≤(∏j∈B|Xj|)𝔼x∈XF(x).\mathbb E_{x\in X}F(T_{B,z}x) \le \left(\prod_{j\in B}|X_j|\right)\mathbb E_{x\in X}F(x).(12)✦ For a real function F:Xp→ℝF:X_p\to\mathbb R and z∈Xpz\in X_p, the assembly in (5) satisfies 𝔼y∈Xp,J\BF(AB,zy)2≤qB𝔼x∈XpF(x)2.\mathbb E_{y\in X_{p,J\setminus B}}F(A_{B,z}y)^2 \le q_B\,\mathbb E_{x\in X_p}F(x)^2.(13)✦ Let h:X→ℝh:X\to\mathbb R have exact coordinate support UU, and fix z∈Xz\in X. Then Exact⁡U\B(h∘TB,z).\operatorname{Exact}_{U\setminus B}(h\circ T_{B,z}).(14)✦ Let LL be a natural length, let f:ℕ→ℝf:\mathbb N\to\mathbb R be a real function, let U,B⊆JU,B\subseteq J be coordinate sets, and let z,x∈Xpz,x\in X_p be residue points. With the polynomial (10) and coefficients (11), one has ℒUf(TB,zx)=𝒫c(x),c(a)={CL,f,U∩B(z;a),supp⁡(a)=U\B,0,supp⁡(a)≠U\B.\begin{aligned} \mathcal L_Uf(T_{B,z}x)&=\mathcal P_c(x),\\ c(a)&= \begin{cases} C_{L,f,U\cap B}(z;a),&\operatorname{supp}(a)=U\setminus B,\\ 0,&\operatorname{supp}(a)\ne U\setminus B. \end{cases} \end{aligned}(15)✦ Here the real function ff and the real value of its lift are regarded as complex.

In the paper, the result reads: for a prime coordinate family JJ, B⊆JB\subseteq J, zB∈Xp,Bz_B\in X_{p,B}, G:Xp→ℂG:X_p\to\mathbb C, and every finite real exponent s≥1s\ge1, ∥G(zB,⋅)∥Ls(Xp,J\B)≤qB1/s∥G∥Ls(Xp)\|G(z_B,\cdot)\|_{L^s(X_{p,J\setminus B})} \le q_B^{1/s}\|G\|_{L^s(X_p)}; under the hypotheses of Lemma 5.2, with its moment parameter rr and parameters M,k,ArM,k,A_r, this yields ∥(ℒℱf)(zB,⋅)∥2r≤2MqB1/(2r)(k+1)Ark\|(\mathcal L^{\mathcal F}f)(z_B,\cdot)\|_{2r} \le2M q_B^{1/(2r)}(k+1)A_r^k, with the same bound for the whole-support subfamilies qU\B≤Hq_{U\setminus B}\le H and qU\B>Hq_{U\setminus B}>H, for any H≥1H\ge1, whose remaining denominators after specialization are respectively at most HH and greater than HH.

What Is Checked

The average and support statements concern real functions on finite coordinate products with the uniform averages (1); the Fourier expansion concerns the residue lift components (7) for a fixed modulus family.