OctoInformBeta

OpenAI preprint, September 2026 · Appendix A. Elementary certificates

Appendix A: Exact arithmetic enclosures: Taylor bounds and admissible charge regimes for the five-cell and larger scalar cases

Exact enclosures for Gaussian scalar bounds.

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

The elementary estimates in the Gaussian propeller bound concern upper tail probabilities and quadratic charges in scalar radii. The relevant numerical bounds involve the Gaussian exponential and its integral, together with specified scalar inequalities.

Write Φ\Phi for the standard Gaussian distribution function. The unnormalized Gaussian density ρ\rho and its normalizing constant cc are defined by (1); the standard Gaussian density is φ=cρ\varphi=c\rho. ρ(x)=e−x2/2,c=12π.\rho(x)=e^{-x^2/2},\qquad c=\frac{1}{\sqrt{2\pi}}.(1) The unnormalized Gaussian upper tail TT is the integral over the open upper half-line, and the Gaussian upper tail probability pp is its normalized value, as in (2). T(x)=∫(x,∞)ρ(t)dt,p(x)=cT(x)=Φ(−x).T(x)=\int_{(x,\infty)}\rho(t)\,dt, \qquad p(x)=cT(x)=\Phi(-x).(2) The inverse Mills ratio λ\lambda is the density divided by the upper tail.✦ The inverse Mills derivative GG in (3) uses the reflected argument q=−xq=-x. λ(x)=ρ(x)T(x)=φ(x)Φ(−x),G(q)=φ(q)(qΦ(q)+φ(q))Φ(q)2,G(−x)=λ(x)(λ(x)−x).\begin{aligned} \lambda(x)&=\frac{\rho(x)}{T(x)}=\frac{\varphi(x)}{\Phi(-x)},\\ G(q)&=\frac{\varphi(q)\bigl(q\Phi(q)+\varphi(q)\bigr)}{\Phi(q)^2}, &G(-x)&=\lambda(x)\bigl(\lambda(x)-x\bigr). \end{aligned}(3) The Gaussian loss is the function loss\operatorname{loss} in (4); write BB for the propeller constant given there. loss⁡(x)=p(x)2G(−x),B=322π,B2=98π.\operatorname{loss}(x)=p(x)^2G(-x), \qquad B=\frac{3}{2\sqrt{2\pi}},\qquad B^2=\frac{9}{8\pi}.(4)✦

Taylor Sums and Scalar Charges

For a nonnegative integer nn, the exponential Taylor sum EnE_n and the Gaussian integral sum JnJ_n are the polynomials defined in (5). En(y)=∑j=0n(−y)jj!,Jn(x)=∑j=0n(−1)jx2j+12jj!(2j+1).E_n(y)=\sum_{j=0}^{n}\frac{(-y)^j}{j!},\qquad J_n(x)=\sum_{j=0}^{n} \frac{(-1)^j x^{2j+1}}{2^j j!(2j+1)}.(5)

For real AA and uu, the scalar charge qA(u)q_A(u) is the quadratic function in (6). For a real parameter AA, the charge threshold τA\tau_A and the clipping coefficient κA\kappa_A are the real quantities defined in (6). qA(u)=0.415u−(A−0.15)u2,τA=0.415A−0.15,κA=(A−0.15)0.42τA+0.42.\begin{aligned} q_A(u)&=0.415u-(A-0.15)u^2,\\ \tau_A&=\frac{0.415}{A-0.15}, &\kappa_A&=\frac{(A-0.15)0.42}{\tau_A+0.42}. \end{aligned}(6) For real parameters AA and kk and a real argument tt, the smaller budget root tA,kt_{A,k} and the terminal charge budget HA,k(t)H_{A,k}(t) are the functions defined in (7). tA,k=0.415−0.4152−4(A−0.15)(1−A)/k2(A−0.15),HA,k(t)=qA(2.6t)+kqA(t).\begin{aligned} t_{A,k} &=\frac{0.415- \sqrt{0.415^2-4(A-0.15)(1-A)/k}}{2(A-0.15)},\\ H_{A,k}(t)&=q_A\bigl(\sqrt{2.6t}\bigr)+kq_A(t). \end{aligned}(7) An admissible scalar regime is a pair (A,k)∈ℝ×ℕ(A,k)\in\mathbb R\times\mathbb N satisfying all the conditions in (8), where t=tA,kt=t_{A,k}; write ℛ(A,k)\mathcal R(A,k) for this condition. 0.884≤A<1,2≤k,0.066<t<0.12,kqA(t)=1−A,2.6t<τA2,(A−0.15)(0.42+t)<0.415,1−A<kqA(0.42),qA(0.066)≤qA(0.45),0<qA(0.066),1−A<κA(4τA2−1+(k−1)0.0662)+(k−1)qA(0.066),1−A<qA(0.23)+kqA(0.066),1−A<HA,k(0.066).\begin{aligned} &0.884\le A<1,\qquad 2\le k,\\ &0.066<t<0.12,\qquad kq_A(t)=1-A,\\ &2.6t<\tau_A^2,\qquad (A-0.15)(0.42+t)<0.415,\\ &1-A<kq_A(0.42),\qquad q_A(0.066)\le q_A(0.45),\\ &0<q_A(0.066),\\ &1-A<\kappa_A\bigl(4\tau_A^2-1+(k-1)0.066^2\bigr) +(k-1)q_A(0.066),\\ &1-A<q_A(0.23)+kq_A(0.066),\\ &1-A<H_{A,k}(0.066). \end{aligned}(8)✦

The Claim

Theorem 1. Let nn be a nonnegative integer and y≥0y\ge0. The odd and even truncations satisfy (9) and (10), respectively. E2n+1(y)≤e−y.E_{2n+1}(y)\le e^{-y}.(9)✦ e−y≤E2n(y).e^{-y}\le E_{2n}(y).(10)✦ For x≥0x\ge0, the Gaussian integral satisfies (11). J2n+1(x)≤∫0xe−t2/2dt≤J2n(x).J_{2n+1}(x)\le\int_0^x e^{-t^2/2}\,dt\le J_{2n}(x).(11)✦ Let 0<r≤2/30<r\le2/3 and x≥0x\ge0, and suppose that B2r2/(1+1−r2)≤loss⁡(x)B^2r^2/(1+\sqrt{1-r^2})\le\operatorname{loss}(x). Then the probability bound is (12). 0.415r+0.15r2<p(x).0.415r+0.15r^2<p(x).(12)✦ The admissible regimes include (13) and (14). ℛ(0.929,2).\mathcal R(0.929,2).(13)✦ ℛ(0.884,3).\mathcal R(0.884,3).(14)✦

In the paper, the result reads: for y≥0y\ge0, E15(y)≤e−y≤E14(y)E_{15}(y)\le e^{-y}\le E_{14}(y), and integrating gives corresponding J15,J14J_{15},J_{14} enclosures for Φ\Phi, together with rational π\pi and square-root bounds, cap estimates and finite interval comparisons.

What Is Checked

The statements are over the finite sums in (5) and nonnegative Gaussian thresholds, with 0<r≤2/30<r\le2/3 in the loss constraint and all the conditions of (8) for the scalar charge constructions in (6) and (7).