OctoInformBeta

OpenAI preprint, September 2026 · Section 4. Two other reporting rules

Section 4.2 (setting): switching to an independent score: An independent fallback score reduces prediction error while table-switched reports have controlled cell degree

A positive quadratic deficit yields switching gains.

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 report can reveal the old block labels or reveal an independent fallback label on an event determined by the old labels. The comparison is between the retained variance of the total score and the Fourier degree of the report cells.

For a finite set XX and a real function f:X→ℝf:X\to\mathbb R, write 𝔼Xf\mathbb E_X f for the uniform mean of ff, with value zero when XX is empty.✦ The fiber indicator CyGC_y^G of a report G:X→BG:X\to B at a label y∈By\in B is the indicator function of the fiber over yy✦, and the fiber probability μG(y)\mu_G(y) is its uniform mean, as in (1).✦ For a report G:X→BG:X\to B, a real observable g:X→ℝg:X\to\mathbb R and a label y∈By\in B, the restricted moment MG,g(y)M_{G,g}(y) is the uniform mean of gg weighted by the indicator of the fiber over yy.✦ The fiber conditional mean mG,g(y)m_{G,g}(y) of a real observable g:X→ℝg:X\to\mathbb R under a report G:X→BG:X\to B, at a label y∈By\in B, is the real number obtained by dividing the restricted moment by the fiber probability, with value zero on an empty fiber.✦ For a report G:X→BG:X\to B and a real observable g:X→ℝg:X\to\mathbb R, the conditional expectation 𝔼[g∣G]:X→ℝ\mathbb E[g\mid G]:X\to\mathbb R is the function in (1). 𝔼Xf=1|X|∑x∈Xf(x),CyG(x)=𝟏{G(x)=y},μG(y)=𝔼XCyG,MG,g(y)=𝔼X[CyGg],mG,g(y)=MG,g(y)μG(y),𝔼[g∣G](x)=mG,g(G(x)).\begin{aligned} \mathbb E_X f&=\frac{1}{|X|}\sum_{x\in X}f(x),& C_y^G(x)&=\mathbf 1_{\{G(x)=y\}},& \mu_G(y)&=\mathbb E_X C_y^G,\\ M_{G,g}(y)&=\mathbb E_X[C_y^Gg],& m_{G,g}(y)&=\frac{M_{G,g}(y)}{\mu_G(y)},& \mathbb E[g\mid G](x)&=m_{G,g}(G(x)). \end{aligned}(1)✦

Let α\alpha be a finite block alphabet, β\beta a label set, and II a finite leaf index set with m=|I|m=|I| and q=m+1q=m+1. The selector and leaf space is Eα=α×αIE_\alpha=\alpha\times\alpha^I✦; the full block space is Xα=Eα×αX_\alpha=E_\alpha\times\alpha, whose last coordinate is the fallback block.✦ For T:α→ℝT:\alpha\to\mathbb R, the old score LTL_T and full score HTH_T are the sums in (2). For F:α→βF:\alpha\to\beta, the old label tuple RFR_F applies FF to each selector and leaf block. LT(z)=T(z0)+∑i∈IT(zi),HT(z,u)=LT(z)+T(u),RF(z)=(F(z0),(F(zi))i∈I).\begin{aligned} L_T(z)&=T(z_0)+\sum_{i\in I}T(z_i),& H_T(z,u)&=L_T(z)+T(u),\\ R_F(z)&=\bigl(F(z_0),(F(z_i))_{i\in I}\bigr). \end{aligned}(2)✦

Write Eβ=β×βIE_\beta=\beta\times\beta^I for the old label tuple space.✦ For an event A⊆EβA\subseteq E_\beta, the event-switched observation OF,A:Xα→Eβ⊔βO_{F,A}:X_\alpha\to E_\beta\sqcup\beta is the tagged report in (3); the tags old\mathrm{old} and fallback\mathrm{fallback} distinguish its two kinds of value. OF,A(z,u)={(fallback,F(u)),RF(z)∈A,(old,RF(z)),RF(z)∉A.O_{F,A}(z,u)= \begin{cases} (\mathrm{fallback},F(u)),&R_F(z)\in A,\\ (\mathrm{old},R_F(z)),&R_F(z)\notin A. \end{cases}(3)✦

The quadratic deficit Δb(T,a)\Delta_b(T,a), for a real weight a:Eα→ℝa:E_\alpha\to\mathbb R and b∈ℝb\in\mathbb R, is the weighted mean in (4).✦ For a report F:α→βF:\alpha\to\beta and an event A⊆EβA\subseteq E_\beta, the event weight aA:Eα→ℝa_A:E_\alpha\to\mathbb R is the function given by aA(z)=𝟏A(RF(z))a_A(z)=\mathbf 1_A(R_F(z)). Δb(T,a)=𝔼Eα[a(z)(1−(LT(z)−b)2)].\Delta_b(T,a)=\mathbb E_{E_\alpha} \left[a(z)\bigl(1-(L_T(z)-b)^2\bigr)\right].(4)

A label score t:β→ℝt:\beta\to\mathbb R determines the switched sum predictor Pt,bP_{t,b} on tagged reports by (5), using the old sum LtL_t on EβE_\beta. Pt,b(old,y)=Lt(y),Pt,b(fallback,v)=b+t(v).P_{t,b}(\mathrm{old},y)=L_t(y),\qquad P_{t,b}(\mathrm{fallback},v)=b+t(v).(5)✦

Table Reports on the Sign Cube

The sign cube on a finite coordinate set Γ\Gamma is 𝒞Γ={−1,1}Γ\mathcal C_\Gamma=\{-1,1\}^{\Gamma}.✦ For a real function f:𝒞Γ→ℝf:\mathcal C_\Gamma\to\mathbb R and U⊆ΓU\subseteq\Gamma, the Fourier character χU\chi_U and real Fourier-Walsh coefficient fˆ(U)\widehat f(U) are given in (6). For a real function f:𝒞Γ→ℝf:\mathcal C_\Gamma\to\mathbb R, the degree deg⁡f\deg f is its ordinary real multilinear Fourier degree in (6), with degree zero for constants and the zero function. χU(x)=∏j∈Uxj,fˆ(U)=𝔼𝒞Γ[fχU],deg⁡f=max⁡({|U|:fˆ(U)≠0}∪{0}).\chi_U(x)=\prod_{j\in U}x_j,\qquad \widehat f(U)=\mathbb E_{\mathcal C_\Gamma}[f\chi_U],\qquad \deg f=\max\bigl(\{\,|U|:\widehat f(U)\ne0\,\}\cup\{0\}\bigr).(6)✦

For a cube observation F:𝒞Γ→βF:\mathcal C_\Gamma\to\beta, its conditional cube sum, or score, σF\sigma_F is the conditional expectation of the coordinate sign sum CΓC_\Gamma.✦ For a cube observation F:𝒞Γ→βF:\mathcal C_\Gamma\to\beta, the retained variance v(F)v(F) is the second moment of its centered conditional score σF\sigma_F, as in (7). CΓ(x)=∑j∈Γxj,σF=𝔼[CΓ∣F],v(F)=𝔼𝒞Γ[σF2].C_\Gamma(x)=\sum_{j\in\Gamma}x_j,\qquad \sigma_F=\mathbb E[C_\Gamma\mid F],\qquad v(F)=\mathbb E_{\mathcal C_\Gamma}[\sigma_F^2].(7)✦

A selector table S:β→IS:\beta\to I assigns a leaf to each selector label. For Boolean leaf tests ℓi:β→{0,1}\ell_i:\beta\to\{0,1\}, the unique failing leaf event AS,ℓ⊆EβA_{S,\ell}\subseteq E_\beta is defined by (8). y∈AS,ℓ⇔ℓS(y0)(yS(y0))=0 and ℓi(yi)=1for every i≠S(y0).y\in A_{S,\ell} \quad\Longleftrightarrow\quad \ell_{S(y_0)}(y_{S(y_0)})=0 \ \text{ and }\ \ell_i(y_i)=1\quad\text{for every }i\ne S(y_0).(8)✦

For observations F0,Fw,Fi:𝒞Γ→βF_0,F_w,F_i:\mathcal C_\Gamma\to\beta, the selector and leaf tuple RR and the table-switched observation OtabO^{\mathrm{tab}} are the maps in (9). R(x)=(F0(x),(Fi(x))i∈I),Otab(x)={(fallback,Fw(x)),R(x)∈AS,ℓ,(old,R(x)),R(x)∉AS,ℓ.\begin{aligned} R(x)&=\bigl(F_0(x),(F_i(x))_{i\in I}\bigr),\\ O^{\mathrm{tab}}(x)&= \begin{cases} (\mathrm{fallback},F_w(x)),&R(x)\in A_{S,\ell},\\ (\mathrm{old},R(x)),&R(x)\notin A_{S,\ell}. \end{cases} \end{aligned}(9)

The switching block index set J={0}⊔I⊔{w}J=\{0\}\sqcup I\sqcup\{w\} has a selector index and a fallback index in addition to the leaves.✦ The block coordinate identification sends x∈𝒞J×Γx\in\mathcal C_{J\times\Gamma} to the blocks xjx_j, where xj(r)=x(j,r)x_j(r)=x(j,r).✦ The switching block identification groups those blocks as an old tuple and a fallback block.✦ The switching cube identification Φ\Phi is the map from 𝒞J×Γ\mathcal C_{J\times\Gamma} to X𝒞ΓX_{\mathcal C_\Gamma} in (10), obtained by composing the block coordinate identification with the switching block identification✦; for a block observation F:𝒞Γ→βF:\mathcal C_\Gamma\to\beta, the table-switched cube observation Õ\widetilde O uses FF in every slot. Φ(x)=((x0,(xi)i∈I),xw),Õ=OF,AS,ℓ∘Φ.\Phi(x)=\bigl((x_0,(x_i)_{i\in I}),x_w\bigr),\qquad \widetilde O=O_{F,A_{S,\ell}}\circ\Phi.(10)✦

The Claim

Theorem 1. Let F:α→βF:\alpha\to\beta, A⊆EβA\subseteq E_\beta, T:α→ℝT:\alpha\to\mathbb R and t:β→ℝt:\beta\to\mathbb R satisfy T=t∘FT=t\circ F and 𝔼αT2=1\mathbb E_\alpha T^2=1. For every b∈ℝb\in\mathbb R, the prediction error satisfies (11). 𝔼Xα[(HT−Pt,b∘OF,A)2]=1−Δb(T,aA).\mathbb E_{X_\alpha} \left[(H_T-P_{t,b}\circ O_{F,A})^2\right] =1-\Delta_b(T,a_A).(11)✦ If also 𝔼αT=0\mathbb E_\alpha T=0, every real η<Δb(T,aA)\eta<\Delta_b(T,a_A) satisfies (12). q+η<𝔼Xα[(𝔼[HT∣OF,A])2].q+\eta<\mathbb E_{X_\alpha} \left[\bigl(\mathbb E[H_T\mid O_{F,A}]\bigr)^2\right].(12)✦ For arbitrary FF, AA and g:α→ℝg:\alpha\to\mathbb R, the conditional sum identity is (13). 𝔼[Hg∣OF,A]=𝔼[H𝔼[g∣F]∣OF,A].\mathbb E[H_g\mid O_{F,A}] =\mathbb E[H_{\mathbb E[g\mid F]}\mid O_{F,A}].(13)✦

Let DD be a nonnegative integer. If every fiber indicator of F0F_0, FwF_w and each FiF_i has degree at most DD, every report value aa of the table-switched observation satisfies (14). deg⁡𝟏{Otab=a}≤qD.\deg\mathbf 1_{\{O^{\mathrm{tab}}=a\}}\le qD.(14)✦ For a block observation F:𝒞Γ→βF:\mathcal C_\Gamma\to\beta whose fiber indicators all have degree at most DD, the table-switched cube observation satisfies (15) for every report value aa. deg⁡𝟏{Õ=a}≤qD.\deg\mathbf 1_{\{\widetilde O=a\}}\le qD.(15)✦ For every block observation FF, its switched score satisfies (16). σÕ=𝔼[HσF∣OF,AS,ℓ]∘Φ.\sigma_{\widetilde O} =\mathbb E[H_{\sigma_F}\mid O_{F,A_{S,\ell}}]\circ\Phi.(16)✦ If σF=kT\sigma_F=kT for k∈ℝk\in\mathbb R and T:𝒞Γ→ℝT:\mathcal C_\Gamma\to\mathbb R, its retained variance satisfies (17). v(Õ)=k2𝔼X𝒞Γ[(𝔼[HT∣OF,AS,ℓ])2].v(\widetilde O)=k^2\mathbb E_{X_{\mathcal C_\Gamma}} \left[\bigl(\mathbb E[H_T\mid O_{F,A_{S,\ell}}]\bigr)^2\right].(17)✦

In the paper, the result reads: for q=m+1q=m+1 arbitrary independent centered variance-one original inputs with sum LL and an independent centered variance-one input ww, the tagged report O+O_+ reveals the original tuple off AA and only ww on AA, has cell cost at most qq, and, with p=ℙ(A)>0p=\mathbb P(A)>0, c=𝔼[L∣A]c=\mathbb E[L\mid A] and Δb=𝔼[𝟏A(1−(L−b)2)]\Delta_b=\mathbb E[\mathbf 1_A(1-(L-b)^2)], satisfies (18) for every real η\eta. Var⁡(𝔼[L+w∣O+])=q+p(1−Var⁡(L∣A)),p(1−Var⁡(L∣A))=Δb+p(c−b)2,Δb>η⇒Var⁡(𝔼[L+w∣O+])>q+η.\begin{aligned} \operatorname{Var}(\mathbb E[L+w\mid O_+]) &=q+p\bigl(1-\operatorname{Var}(L\mid A)\bigr),\\ p\bigl(1-\operatorname{Var}(L\mid A)\bigr) &=\Delta_b+p(c-b)^2,\\ \Delta_b>\eta&\quad\Longrightarrow\quad \operatorname{Var}(\mathbb E[L+w\mid O_+])>q+\eta. \end{aligned}(18)

What Is Checked

The switching identities are stated over uniform finite product alphabets with an event on the old label tuple; the degree bounds concern fiber indicators of observations on finite sign cubes, including the block observation obtained by regrouping the product coordinates through Φ\Phi.