OctoInformBeta

OpenAI preprint, September 2026 · §3. Probability bounds for the cells

Lemma 3.2: Bounds and monotonicity for the inverse Mills derivative in reflected coordinates

Gaussian loss decreases with nonnegative thresholds.

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

Gaussian cell probability bounds involve the upper-tail probability at a real threshold and a loss weighted by its square. Let φ\varphi and Φ\Phi denote the standard Gaussian density and distribution function on ℝ\mathbb R.

The unnormalized Gaussian density d:ℝ→ℝd:\mathbb R\to\mathbb R and the Gaussian normalizing constant cc are given by (1), which also fixes the normalization of φ\varphi: d(x)=e−x2/2,c=12π,φ(x)=cd(x).d(x)=e^{-x^2/2},\qquad c=\frac{1}{\sqrt{2\pi}},\qquad \varphi(x)=c\,d(x).(1) The unnormalized Gaussian tail T:ℝ→ℝT:\mathbb R\to\mathbb R is the integral of dd over the open upper half-line, as in (2): T(x)=∫(x,∞)d(t)dt.T(x)=\int_{(x,\infty)}d(t)\,dt.(2)✦ The Gaussian upper tail p:ℝ→ℝp:\mathbb R\to\mathbb R is the probability function in (3): p(x)=cT(x)=Φ(−x).p(x)=c\,T(x)=\Phi(-x).(3)✦

The inverse Mills ratio λ:ℝ→ℝ\lambda:\mathbb R\to\mathbb R is the density divided by the upper tail, with either normalization in (4): λ(x)=d(x)T(x)=φ(x)Φ(−x).\lambda(x)=\frac{d(x)}{T(x)}=\frac{\varphi(x)}{\Phi(-x)}.(4)✦ The inverse Mills derivative G:ℝ→ℝG:\mathbb R\to\mathbb R is the function given in (5), and its reflected form G̃\widetilde G is the function obtained by setting x=−qx=-q: G(q)=φ(q)(qΦ(q)+φ(q))Φ(q)2,G̃(x)=G(−x)=λ(x)(λ(x)−x).G(q)=\frac{\varphi(q)\bigl(q\Phi(q)+\varphi(q)\bigr)}{\Phi(q)^2}, \qquad \widetilde G(x)=G(-x)=\lambda(x)\bigl(\lambda(x)-x\bigr).(5) The Gaussian loss loss⁡:ℝ→ℝ\operatorname{loss}:\mathbb R\to\mathbb R is the function in (6): loss⁡(x)=p(x)2G(−x)=p(x)2G̃(x).\operatorname{loss}(x)=p(x)^2G(-x)=p(x)^2\widetilde G(x).(6)✦ The inverse Gaussian distribution function Φ−1\Phi^{-1} is defined on probabilities in (0,1)(0,1).

The Claim

Theorem 1. Let x≤yx\le y be nonnegative real thresholds. The reflected inverse Mills derivative satisfies (7) and (8): G̃(x)≤G̃(y).\widetilde G(x)\le\widetilde G(y).(7)✦ 0≤G̃(x).0\le\widetilde G(x).(8)✦ For every real threshold tt, its upper bound is (9): G̃(t)≤1.\widetilde G(t)\le1.(9)✦ For the ordered nonnegative thresholds x,yx,y, the loss satisfies (10): loss⁡(y)≤loss⁡(x).\operatorname{loss}(y)\le\operatorname{loss}(x).(10)✦

In the paper, the result reads: GG is nonincreasing on ℝ\mathbb R and satisfies 0<G(q)<10<G(q)<1 for every real qq, and P↦P2G(Φ−1(P))P\mapsto P^2G(\Phi^{-1}(P)) is increasing on (0,1/2](0,1/2].

What Is Checked

The assertions are stated over the real Gaussian density and upper-tail integral, with the inverse Mills ratio and its derivative defined in reflected coordinates, and the loss defined using the normalized upper-tail probability.