OctoInformBeta

OpenAI preprint, September 2026 · §2. Reduction to approximation and rounding

Proof of Theorem 1.1: With μ(X)≤4δN\mu(X) \le 4\delta N at energy at least (1−8δ)N(1-8\delta)N, an entropy inequality, δ\delta chosen for each η>0\eta > 0, and z=e(t)z = e(t) for |z|=1|z| = 1

Rounding a real coefficient vector to a Littlewood polynomial.

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

A signing ε\varepsilon of length NN is a vector (ε0,…,εN−1)(\varepsilon_0, \dots, \varepsilon_{N-1}) of real numbers, each equal to −1-1 or 11.✦ For a vector ε\varepsilon of NN real numbers and a complex number zz, put Pε(z)=∑k<Nεkzk;P_\varepsilon(z) = \sum_{k<N} \varepsilon_k z^k ;(1)✦ when ε\varepsilon is a signing, PεP_\varepsilon is its Littlewood polynomial. Let mNm_N be the least value of N−1/2max⁡|z|=1|Pε(z)|N^{-1/2} \max_{|z|=1} |P_\varepsilon(z)| over the signings ε\varepsilon of length NN. By Parseval’s identity mN≥1m_N \ge 1, and the flatness problem asks how close to 11 the value mNm_N comes for large NN.

The item is the deduction of Theorem 1.1 of the OpenAI preprint Asymptotically minimal maxima of real Littlewood polynomials (September 2026) from its reduction to approximation and rounding, in which a vector of real coefficients in [−1,1][-1,1] with a nearly flat trigonometric polynomial is rounded to a signing. The library states the deduction as five separate facts: a parametrization of the unit circle, a choice of a small positive parameter δ\delta, a bound on the defect mass of a vector (defined below), an entropy inequality, and the rounding itself.

Phases, Energy and Defect Mass

The phase of a real number xx is the point eix=cos⁡x+isin⁡xe^{ix} = \cos x + i \sin x of the unit circle, and e(x)e(x) denotes the phase of 2πx2\pi x: e(x)=e2πix.e(x) = e^{2\pi i x}.(2)✦

The energy of a vector X=(X0,…,XN−1)X = (X_0, \dots, X_{N-1}) in [−1,1]N[-1,1]^N is ∑k<NXk2\sum_{k<N} X_k^2, and its trigonometric polynomial is the function QXQ_X on [0,1][0,1] given by QX(t)=∑k<NXke(kt).Q_X(t) = \sum_{k<N} X_k \, e(kt).(3)

The defect mass of a vector XX of NN real numbers is the real number μ(X)=12∑k<N(1−|Xk|).\mu(X) = \frac{1}{2} \sum_{k<N} \bigl(1 - |X_k|\bigr).(4)✦

For a parameter δ\delta with 0<δ<1/200 < \delta < 1/20, the constant KδK_\delta is Kδ=(1+δ)31−δ.K_\delta = \sqrt{\frac{(1+\delta)^3}{1-\delta}}.(5)

The Claim

Theorem 1.

  1. For every complex number zz with |z|=1|z| = 1 there is a t∈[0,1]t \in [0,1] with z=e(t).z = e(t).(6)✦

  2. For every η>0\eta > 0 there is a δ\delta with 0<δ<1/200 < \delta < 1/20 and, with KδK_\delta as in (5), Kδ+393216δlog⁡(20/δ)<1+η2.K_\delta + 393216 \sqrt{\delta \log(20/\delta)} < 1 + \frac{\eta}{2}.(7)✦

  3. Let N≥1N \ge 1, let δ\delta be a real number, and let X∈[−1,1]NX \in [-1,1]^N have energy at least (1−8δ)N(1 - 8\delta)N. Then its defect mass satisfies 0≤μ(X)≤4δN.0 \le \mu(X) \le 4\delta N.(8)✦

  4. Let δ\delta, NN and uu be real numbers with 0<δ<1/200 < \delta < 1/20, N>0N > 0 and 0≤u≤4δN0 \le u \le 4\delta N. Then, the left side being 00 when u=0u = 0, ulog⁡80Nu≤4δNlog⁡20δ.u \log\frac{80N}{u} \le 4\delta N \log\frac{20}{\delta}.(9)✦

  5. Let N≥1N \ge 1, let δ\delta satisfy 0<δ<1/200 < \delta < 1/20, let KK and BB be real numbers, and let X∈[−1,1]NX \in [-1,1]^N have energy at least (1−8δ)N(1 - 8\delta)N and satisfy N−1/2|QX(t)|≤K+B/NN^{-1/2} |Q_X(t)| \le K + B/N for every t∈[0,1]t \in [0,1]. Then there is a signing ε\varepsilon of length NN such that, for every complex number zz with |z|=1|z| = 1, |Pε(z)|N≤K+BN+196608(1N+2δlog⁡(20/δ)).\frac{|P_\varepsilon(z)|}{\sqrt{N}} \le K + \frac{B}{N} + 196608 \Bigl( \frac{1}{\sqrt{N}} + 2\sqrt{\delta \log(20/\delta)} \Bigr).(10)✦

In the paper, the result reads: a vector XNX_N of length NN with energy at least (1−8δ)N(1-8\delta)N has μ(XN)≤4δN\mu(X_N) \le 4\delta N; ulog⁡(80N/u)≤4δNlog⁡(20/δ)u \log(80N/u) \le 4\delta N \log(20/\delta) for u≤4δNu \le 4\delta N; hence limsup⁡N→∞mN≤Kδ+2Cδlog⁡(20/δ)\limsup_{N\to\infty} m_N \le K_\delta + 2C\sqrt{\delta\log(20/\delta)} for a constant CC, and one lets δ↓0\delta \downarrow 0.

What Is Checked

The statements are made over vectors XX in [−1,1]N[-1,1]^N, with their energy, defect mass μ(X)\mu(X) and trigonometric polynomial QXQ_X in the phase e(t)e(t), and over signings ε\varepsilon with the values of their Littlewood polynomials PεP_\varepsilon on the unit circle.