OctoInformBeta

OpenAI preprint, September 2026 · §3. Weighted Fourier grading

The weighted grading: For a weight w>0w > 0 on Ω\Omega, the layers EkE_k have dimension (nk)\binom{n}{k}, and their projections define the gradings LL and MM

Weighted Fourier layers give an orthogonal family of projections.

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

For a positive integer nn, the cube Ω={−1,1}n\Omega = \{-1,1\}^n has N=2nN = 2^n points, and a point x∈Ωx \in \Omega has real coordinates x1,…,xn∈{−1,1}x_1, \dots, x_n \in \{-1, 1\}. Complex functions on Ω\Omega form the space ℂΩ\mathbb{C}^\Omega, with the inner product ⟨g,h⟩=∑x∈Ωg(x)h(x)¯\langle g, h \rangle = \sum_{x \in \Omega} g(x)\overline{h(x)}. The character χS\chi_S of a set S⊆{1,…,n}S \subseteq \{1, \dots, n\} is the function on Ω\Omega given by χS(x)=∏j∈Sxj,\chi_S(x) = \prod_{j \in S} x_j,(1)✦ and the space VkV_k of functions of degree at most kk is the subspace of ℂΩ\mathbb{C}^\Omega spanned by the characters of sets with at most kk elements: Vk=span⁡ℂ{χS:|S|≤k}.V_k = \operatorname{span}_{\mathbb{C}} \{\chi_S : |S| \le k\}.(2)✦

Without a weight, the degree grading of ℂΩ\mathbb{C}^\Omega comes from the filtration V0⊆V1⊆⋯⊆VnV_0 \subseteq V_1 \subseteq \cdots \subseteq V_n: the orthogonal complement of Vk−1V_{k-1} in VkV_k is spanned by the characters χS\chi_S with |S|=k|S| = k, and the grading acts on χS\chi_S as multiplication by |S||S|. The weighted grading repeats the construction for a weight w>0w > 0 on Ω\Omega, after multiplying the filtration by w\sqrt{w}.

Matrices and Projection Families

The functions δx\delta_x, equal to 11 at the point xx and 00 elsewhere, form the point basis of ℂΩ\mathbb{C}^\Omega, which is orthonormal. Operators on ℂΩ\mathbb{C}^\Omega are matrices in this basis. The matrix of an operator BB in an orthonormal basis (bi)(b_i) of a finite-dimensional complex inner product space is the square matrix with entries Bij=⟨Bbj,bi⟩,B_{ij} = \langle B b_j, b_i \rangle,(3)✦ which in the point basis reads Bxy=(Bδy)(x)B_{xy} = (B\delta_y)(x).

A family (Pk)k∈I(P_k)_{k \in I} of complex square matrices of one size, indexed by a finite set II, is an orthogonal projection family when every PkP_k is Hermitian and idempotent, PrPs=0P_r P_s = 0 whenever r≠sr \neq s, and ∑k∈IPk=Id\sum_{k \in I} P_k = \mathrm{Id}.✦

The Claim

Definition 1. Let w:Ω→ℝw \colon \Omega \to \mathbb{R} be a weight with w(x)>0w(x) > 0 at every point xx. The operator DD on ℂΩ\mathbb{C}^\Omega is multiplication by w\sqrt{w}: (Dg)(x)=w(x)g(x)(g∈ℂΩ,x∈Ω).(Dg)(x) = \sqrt{w(x)}\,g(x) \qquad (g \in \mathbb{C}^\Omega,\ x \in \Omega).(4)✦ For k≥0k \ge 0, the weighted degree space KkK_k is the image of VkV_k under DD, Kk=DVk={wg:g∈Vk},K_k = D V_k = \{\sqrt{w}\,g : g \in V_k\},(5)✦ and with K−1={0}K_{-1} = \{0\} these spaces form the weighted Fourier flag K−1⊆K0⊆K1⊆⋯K_{-1} \subseteq K_0 \subseteq K_1 \subseteq \cdots. The layer EkE_k of the weighted Fourier flag is the orthogonal complement of Kk−1K_{k-1} in KkK_k, Ek=Kk∩Kk−1⟂.E_k = K_k \cap K_{k-1}^{\perp}.(6) For k=0,…,nk = 0, \dots, n, the weighted projection Πk\Pi_k is the orthogonal projection of ℂΩ\mathbb{C}^\Omega onto EkE_k, the operator with Πkg∈Ek,g−Πkg⟂Ek(g∈ℂΩ).\Pi_k g \in E_k, \qquad g - \Pi_k g \perp E_k \qquad (g \in \mathbb{C}^\Omega).(7)✦ Then for every k≥0k \ge 0 dim⁡ℂEk=(nk),\dim_{\mathbb{C}} E_k = \binom{n}{k},(8)✦ the matrices Π0,…,Πn\Pi_0, \dots, \Pi_n form an orthogonal projection family, Πk*=Πk=Πk2,ΠrΠs=0(r≠s),∑k=0nΠk=Id,\Pi_k^* = \Pi_k = \Pi_k^2, \qquad \Pi_r \Pi_s = 0 \ \ (r \neq s), \qquad \sum_{k=0}^{n} \Pi_k = \mathrm{Id},(9)✦ and their traces have real parts Re⁡tr⁡Πk=(nk)(k=0,…,n).\operatorname{Re} \operatorname{tr} \Pi_k = \binom{n}{k} \qquad (k = 0, \dots, n).(10)✦ For complex square matrices P0,…,PnP_0, \dots, P_n of one size, the grading LL and the centered grading MM are the matrices L=∑k=0nkPk,L = \sum_{k=0}^{n} k\,P_k,(11)✦ M=∑k=0n(k−n2)Pk;M = \sum_{k=0}^{n} \Bigl(k - \frac{n}{2}\Bigr) P_k;(12)✦ for Pk=ΠkP_k = \Pi_k they are the weighted grading LL and its centering M=L−n2IdM = L - \frac{n}{2}\,\mathrm{Id}. The squared Hilbert-Schmidt norm of a complex matrix B=(Bxy)B = (B_{xy}), with rows and columns indexed by finite sets, is the real number ∥B∥HS2=∑x,y|Bxy|2.\|B\|_{\mathrm{HS}}^2 = \sum_{x, y} |B_{xy}|^2.(13)✦

What Is Checked

The definition is stated over the cube Ω\Omega with its real coordinates xix_i, the characters χS\chi_S and the spaces VkV_k they span, the space ℂΩ\mathbb{C}^\Omega with its inner product and point basis, and a weight ww that is positive at every point of Ω\Omega.

About these constructions the library proves that multiplication by a complex function with no zero on Ω\Omega is a linear bijection of ℂΩ\mathbb{C}^\Omega, inverted by multiplication by the reciprocal function, so that DD in (4) is invertible with inverse multiplication by 1/w1/\sqrt{w}.✦ It proves as well that taking matrices in an orthonormal basis, as in (3), is an isomorphism of algebras from the operators onto the square matrices.✦