OctoInformBeta

OpenAI preprint, September 2026 · Section 4. Finite blocks and prediction

Lemma 4.2: Uniform separation and arbitrary successive lower bounds for finitely many tests with L1L^1 coefficients

Separated frequencies approximate independent phase averages.

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

Integer frequencies couple several torus phases to a single circle variable. Write 𝕋=ℝ/ℤ\mathbb T=\mathbb R/\mathbb Z and e(t)=exp⁡(2πit)e(t)=\exp(2\pi i t). The phase Haar measure dzdz is normalized Haar measure on 𝕋\mathbb T; on 𝕋n\mathbb T^n, let dydy denote its product.✦

Let KK and II be finite sets and n∈ℕn\in\mathbb N, with coordinates indexed by 0,…,n−10,\ldots,n-1. Given coefficient functions cki:𝕋→ℂc_{ki}:\mathbb T\to\mathbb C and exponent vectors νki∈ℤn\nu_{ki}\in\mathbb Z^n for k∈Kk\in K and i∈Ii\in I, a phase test Pk:𝕋×𝕋n→ℂP_k:\mathbb T\times\mathbb T^n\to\mathbb C, for k∈Kk\in K, is the finite Fourier sum (1): Pk(z,y)=∑i∈Icki(z)e(∑j=0n−1νki,jyj).P_k(z,y)=\sum_{i\in I}c_{ki}(z) e\!\left(\sum_{j=0}^{n-1}\nu_{ki,j}y_j\right).(1)✦

For integrable coefficients cki∈L1(𝕋)c_{ki}\in L^1(\mathbb T) and an integer vector L=(L0,…,Ln−1)∈ℤnL=(L_0,\ldots,L_{n-1})\in\mathbb Z^n, the sampled average Ak(L)A_k(L) and independent average MkM_k are the integrals in (2): Ak(L)=∫𝕋Pk(z,(L0z,…,Ln−1z))dz,Mk=∫𝕋∫𝕋nPk(z,y)dydz.\begin{aligned} A_k(L)&=\int_{\mathbb T}P_k(z,(L_0z,\ldots,L_{n-1}z))\,dz,\\ M_k&=\int_{\mathbb T}\int_{\mathbb T^n}P_k(z,y)\,dy\,dz. \end{aligned}(2)

The Claim

Theorem 1. For the phase tests with integrable coefficients above, let ε>0\varepsilon>0. There are integers R,D≥0R,D\geq0 such that every L∈ℤ≥0nL\in\mathbb Z_{\geq0}^n satisfying R+D∑i<jLi<LjR+D\sum_{i<j}L_i<L_j for each 0≤j<n0\leq j<n obeys the bounds (3): |Ak(L)−Mk|<ε(k∈K).\bigl|A_k(L)-M_k\bigr|<\varepsilon \qquad(k\in K).(3)✦

Let Bj:ℤj→ℤB_j:\mathbb Z^j\to\mathbb Z be any prescribed function for each j∈ℕj\in\mathbb N. There is a vector L∈ℤ>0nL\in\mathbb Z_{>0}^n satisfying (4): Bj(L0,…,Lj−1)<Lj(0≤j<n),|Ak(L)−Mk|<ε(k∈K).\begin{aligned} B_j(L_0,\ldots,L_{j-1})&<L_j &&(0\leq j<n),\\ \bigl|A_k(L)-M_k\bigr|&<\varepsilon &&(k\in K). \end{aligned}(4)✦

In the paper, the result reads: for finitely many Laurent polynomials Pk(z,w)=∑ν∈Ekck,ν(z)∏j=0n−1wjνjP_k(z,w)=\sum_{\nu\in E_k}c_{k,\nu}(z)\prod_{j=0}^{n-1}w_j^{\nu_j} in nn phase variables, with finite Ek⊂ℤnE_k\subset\mathbb Z^n and ck,ν∈L1(𝕋)c_{k,\nu}\in L^1(\mathbb T), and every η>0\eta>0, successively large positive integers L0,…,Ln−1L_0,\ldots,L_{n-1} can be chosen so that each integral ∫𝕋Pk(z,e(L0z),…,e(Ln−1z))dz\int_{\mathbb T}P_k(z,e(L_0z),\ldots,e(L_{n-1}z))\,dz differs in absolute value by less than η\eta from its independent product-torus average, while each LjL_j exceeds any finite lower bound determined by the previous choices and the then-current construction data.

What Is Checked

The assertions (3) and (4) are stated for phase tests with L1(𝕋)L^1(\mathbb T) coefficients and arbitrary integer exponent vectors, with all averages taken in normalized Haar measure.