OctoInformBeta

OpenAI preprint, September 2026 · §2. Families of directions

Direction families and chart maps: For a large positive integer mm, m3m^3 families of m6m^6 directions in ℝm\mathbb{R}^m, operator and column norm bounds, and inverse power parameters

Chart maps of pivoted direction families have left pseudoinverses.

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 large positive integer mm, and set N=m6N = m^6 and S=m3S = m^3. For SS families of NN direction vectors in ℝm\mathbb{R}^m, normalized inner products assemble each family into a linear chart map from ℝm\mathbb{R}^m into ℝN\mathbb{R}^N. Alongside an inverse power parameter, the question is how to construct a linear left inverse for each chart map with bounds on its operator norm and column norms.

Direction Families

The empirical covariance Cov⁡(g,v)\operatorname{Cov}(g, v) of an NN-tuple g=(g1,…,gN)g = (g_1, \dots, g_N) of vectors in a real inner product space EE at a vector v∈Ev \in E, for a positive integer NN, is the real number Cov⁡(g,v)=1N∑i=1N⟨v,gi⟩2.\operatorname{Cov}(g, v) = \frac{1}{N} \sum_{i=1}^N \langle v, g_i \rangle^2.(1)✦

The empirical sine transform Ψ(g,w)\Psi(g, w) of an NN-tuple g=(g1,…,gN)∈ENg = (g_1, \dots, g_N) \in E^N at a vector w∈Ew \in E, for a real inner product space EE and a positive integer NN, is the vector in EE Ψ(g,w)=1N∑i=1Nsin⁡(⟨gi,w⟩)gi.\Psi(g, w) = \frac{1}{N} \sum_{i=1}^N \sin(\langle g_i, w \rangle)\, g_i.(2)✦

The bad projection set 𝒜(q,v)\mathcal{A}(q, v) of a threshold q∈ℝq \in \mathbb{R} and a vector vv in a real inner product space EE, for a nonnegative integer NN, is the subset of ENE^N defined by 𝒜(q,v)={x=(x1,…,xN)∈EN:∃1≤i≤N,|⟨v,xi⟩|>q}.\mathcal{A}(q, v) = \{ x = (x_1, \dots, x_N) \in E^N : \exists\, 1 \le i \le N,\ |\langle v, x_i \rangle| > q \}.(3)✦

The number of bad samples ν(A,x)\nu(A, x) of a subset AA of a set XX and an SS-tuple x=(x1,…,xS)∈XSx = (x_1, \dots, x_S) \in X^S, for a nonnegative integer SS, is the nonnegative integer counting the indices 1≤s≤S1 \le s \le S such that xs∈Ax_s \in A.✦

A direction family g=(gs,i)1≤s≤S,1≤i≤Ng = (g_{s,i})_{1 \le s \le S,\, 1 \le i \le N} of parameter mm, for a positive integer mm with N=m6N = m^6 and S=m3S = m^3, is an indexed collection of vectors in ℝm\mathbb{R}^m, written gs=(gs,1,…,gs,N)g_s = (g_{s,1}, \dots, g_{s,N}) for each 1≤s≤S1 \le s \le S, that satisfies five conditions using (1), (2), and (3): m2≤∥gs,i∥≤2m(1≤s≤S,1≤i≤N),gs,a≠tgs,b(1≤s≤S,1≤a,b≤N,a≠b,t∈ℝ),12≤Cov⁡(gs,v)≤2(1≤s≤S,v∈ℝm,∥v∥=1),∥Ψ(gs,w)−e−∥w∥2/2w∥≤2m(1≤s≤S,w∈ℝm,∥w∥≤m42),ν(𝒜(11log⁡m,v),g)≤S2(v∈ℝm,∥v∥=1).\begin{aligned} \frac{\sqrt{m}}{2} \le \|g_{s,i}\| &\le 2\sqrt{m} && (1 \le s \le S,\ 1 \le i \le N), \\ g_{s,a} &\neq t g_{s,b} && (1 \le s \le S,\ 1 \le a, b \le N,\ a \neq b,\ t \in \mathbb{R}), \\ \frac{1}{2} \le \operatorname{Cov}(g_s, v) &\le 2 && (1 \le s \le S,\ v \in \mathbb{R}^m,\ \|v\| = 1), \\ \left\| \Psi(g_s, w) - e^{-\|w\|^2/2} w \right\| &\le \frac{2}{\sqrt{m}} && (1 \le s \le S,\ w \in \mathbb{R}^m,\ \|w\| \le m^{42}), \\ \nu\bigl(\mathcal{A}(11\sqrt{\log m}, v),\, g\bigr) &\le \frac{S}{2} && (v \in \mathbb{R}^m,\ \|v\| = 1). \end{aligned}(4)✦

A pivoted direction family g=(gs,i)1≤s≤S,1≤i≤Ng = (g_{s,i})_{1 \le s \le S,\, 1 \le i \le N} of parameter mm, for a positive integer mm with N=m6N = m^6 and S=m3S = m^3, is a direction family satisfying the conditions (4) together with the coordinate lower bound |(gs,i)j|>1m30(1≤s≤S,1≤i≤N,1≤j≤m).|(g_{s,i})_j| > \frac{1}{m^{30}} \qquad (1 \le s \le S,\ 1 \le i \le N,\ 1 \le j \le m).(5)✦

Left Pseudoinverses and Frame Bounds

Let EE be a finite-dimensional real Hilbert space, let FF be a real Hilbert space, let J:E→FJ : E \to F be a continuous linear map with adjoint J*:F→EJ^* : F \to E, and let a>0a > 0 satisfy a∥x∥2≤∥Jx∥2a \|x\|^2 \le \|J x\|^2 for all x∈Ex \in E; then the Gram operator J*J:E→EJ^* J : E \to E is the invertible continuous linear map given by composing the adjoint of JJ with JJ.✦

The left pseudoinverse J+:F→EJ^+ : F \to E of a continuous linear map J:E→FJ : E \to F with adjoint J*:F→EJ^* : F \to E, for a finite-dimensional real Hilbert space EE, a real Hilbert space FF, and a constant a>0a > 0 satisfying a∥x∥2≤∥Jx∥2a \|x\|^2 \le \|J x\|^2 for all x∈Ex \in E, is the continuous linear map J+=(J*J)−1J*.J^+ = (J^* J)^{-1} J^*.(6)✦

The frame lower bound constant ama_m for a large positive integer mm is the positive real number am=m62m,a_m = \frac{m^6}{2m},(7)✦

The Claim

Definition 1. The inverse power parameter p(x,n)p(x, n) of a real number xx and a nonnegative integer nn is the real number p(x,n)=1xn.p(x, n) = \frac{1}{x^n}.(8)✦ The frame chart Ju:E→ℝNJ_u : E \to \mathbb{R}^N of a tuple u=(u1,…,uN)∈ENu = (u_1, \dots, u_N) \in E^N, for a real inner product space EE and a nonnegative integer NN, is the continuous linear map with components (Juθ)i=⟨ui,θ⟩(J_u \theta)_i = \langle u_i, \theta \rangle for 1≤i≤N1 \le i \le N.✦ Let mm be a large positive integer with N=m6N = m^6 and S=m3S = m^3, and let g=(gs,i)1≤s≤S,1≤i≤Ng = (g_{s,i})_{1 \le s \le S,\, 1 \le i \le N} be a pivoted direction family of parameter mm as in (5). For 1≤s≤S1 \le s \le S and 1≤i≤N1 \le i \le N, the normalized direction us,iu_{s,i} is the vector gs,i/m∈ℝmg_{s,i} / \sqrt{m} \in \mathbb{R}^m✦, and the chart map Js:ℝm→ℝNJ_s : \mathbb{R}^m \to \mathbb{R}^N is the continuous linear map JusJ_{u_s} for us=(us,1,…,us,N)u_s = (u_{s,1}, \dots, u_{s,N}). With the positive constant ama_m of (7), every chart map JsJ_s satisfies the frame energy lower bound am∥θ∥2≤∥Jsθ∥2a_m \|\theta\|^2 \le \|J_s \theta\|^2 for all θ∈ℝm\theta \in \mathbb{R}^m✦, and the chart pseudoinverse Js+:ℝN→ℝmJ_s^+ : \mathbb{R}^N \to \mathbb{R}^m of JsJ_s is the continuous linear map given by the left pseudoinverse (6).✦ Then for each 1≤s≤S1 \le s \le S, the chart pseudoinverse Js+J_s^+ satisfies the left inverse identity (9) for all θ∈ℝm\theta \in \mathbb{R}^m, the column norm bound (10) for each standard basis vector ei∈ℝNe_i \in \mathbb{R}^N (1≤i≤N1 \le i \le N), and the operator norm bound (11) for all y∈ℝNy \in \mathbb{R}^N. Js+(Jsθ)=θ,J_s^+ (J_s \theta) = \theta,(9)✦ ∥Js+ei∥≤4mm6,\|J_s^+ e_i\| \le 4 \frac{m}{m^6},(10)✦ ∥Js+y∥≤∥y∥am.\|J_s^+ y\| \le \frac{\|y\|}{\sqrt{a_m}}.(11)✦

What Is Checked

The inverse power parameter (8) is stated over any real number xx and nonnegative integer nn, the frame chart over any real inner product space EE, nonnegative integer NN, and tuple u∈ENu \in E^N, and the normalized chart constructions together with the bounds (9), (10), and (11) over a large positive integer mm with N=m6N = m^6 and S=m3S = m^3, a pivoted direction family gg of parameter mm, indices 1≤s≤S1 \le s \le S and 1≤i≤N1 \le i \le N, and vectors θ∈ℝm\theta \in \mathbb{R}^m and y∈ℝNy \in \mathbb{R}^N.