OctoInformBeta

OpenAI preprint, September 2026 · Section 3. Retaining more variance than the degree cost

Section 3.3 (setting): transfer to finite inputs and deterministic realization: The threshold is fixed by Gaussian moments, while the selector and leaf tests depend only on finite labels

Normalized finite scores admit deterministic table gain.

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

For a centered score of unit second moment on a finite alphabet, the question is whether deterministic tests of finite labels can retain a fixed positive quadratic deficit. The labels may be observations of the score through a prescribed map.

For a finite alphabet AA and real functions f,T:A→ℝf,T:A\to\mathbb R, the uniform mean 𝔼Af\mathbb E_A f is the average and the normalized sum Un:An→ℝU_n:A^n\to\mathbb R is the function defined in (1); the mean on the empty alphabet is zero. 𝔼Af=1|A|∑a∈Af(a),Un(x)=1n∑j=1nT(xj)(n>0).\mathbb E_A f=\frac{1}{|A|}\sum_{a\in A}f(a), \qquad U_n(x)=\frac{1}{\sqrt n}\sum_{j=1}^n T(x_j) \quad(n>0).(1) All expectations below are uniform on the indicated finite product alphabet. For a finite leaf set II and real weights si,wi:A→ℝs_i,w_i:A\to\mathbb R, the soft event weight Ws,w:A×AI→ℝW_{s,w}:A\times A^I\to\mathbb R is the function in (2). Ws,w(x0,(xi)i∈I)=∑i∈Isi(x0)∏j≠iwj(xj)−∏j∈Iwj(xj).W_{s,w}(x_0,(x_i)_{i\in I}) =\sum_{i\in I}s_i(x_0)\prod_{j\ne i}w_j(x_j) -\prod_{j\in I}w_j(x_j).(2)✦ For a score U:A→ℝU:A\to\mathbb R, a real center bb and a real weight a:A×AI→ℝa:A\times A^I\to\mathbb R, the quadratic deficit DU,b(a)D_{U,b}(a) is the weighted average in (3). DU,b(a)=𝔼A×AI[a(x)(1−(U(x0)+∑i∈IU(xi)−b)2)].D_{U,b}(a)=\mathbb E_{A\times A^I} \left[a(x)\left(1-\left(U(x_0)+\sum_{i\in I}U(x_i)-b\right)^2\right)\right].(3)✦

The Fixed Gaussian Tests

The Gaussian construction uses the grid half-size NN, radius RR, spacing δ\delta, tent width hh and leaf set I*I_* in (4); its middle index is m=Nm=N. N=(2128)8,R=96,δ=R/N,h=(2128)−12,I*={0,…,2N},m=N.N=(2^{128})^8,\qquad R=96,\qquad \delta=R/N,\qquad h=(2^{128})^{-12},\qquad I_*=\{0,\ldots,2N\},\qquad m=N.(4) Clipping κ\kappa is a function from ℝ\mathbb R to [0,1][0,1]. Write 𝟏E\mathbf1_E for the indicator of an event EE. The grid ramp difference rir_i, selector partition weight σi\sigma_i, leaf center aia_i and leaf tent test ℓi\ell_i are the functions and numbers in (5); τa,q\tau_{a,q} denotes a tent with center aa and width parameter qq. κ(z)=max⁡{0,min⁡{1,z}},ri(z)=κ(z/δ+N−i+1)−κ(z/δ+N−i),σi(z)=ri(z)+𝟏{i=m}(1−∑j∈I*rj(z)),ai=1+δ(i−N),τa,q(z)=max⁡{0,1−|z−a|/q},ℓi(z)=τai,h(z).\begin{aligned} \kappa(z)&=\max\{0,\min\{1,z\}\},\\ r_i(z)&=\kappa(z/\delta+N-i+1)-\kappa(z/\delta+N-i),\\ \sigma_i(z)&=r_i(z)+\mathbf1_{\{i=m\}} \left(1-\sum_{j\in I_*}r_j(z)\right),\\ a_i&=1+\delta(i-N),\qquad \tau_{a,q}(z)=\max\{0,1-|z-a|/q\},\qquad \ell_i(z)=\tau_{a_i,h}(z). \end{aligned}(5)

For a measure PP on a measurable space Ω\Omega, a real score U:Ω→ℝU:\Omega\to\mathbb R and a family v=(vi)v=(v_i) of real functions on ℝ\mathbb R, the weighted moment array MP(U,v)M^P(U,v) is the real array of orders k∈{0,1,2}k\in\{0,1,2\} defined in (6). For a finite alphabet AA, a real score U:A→ℝU:A\to\mathbb R and real weights wi:A→ℝw_i:A\to\mathbb R, the finite weighted moment array RA(U,w)R^A(U,w) is the real array defined in (6). Mk,iP(U,v)=∫Ωvi(U(x))U(x)kdP(x),Rk,iA(U,w)=𝔼A[wi(x)U(x)k].M^P_{k,i}(U,v)=\int_\Omega v_i(U(x))U(x)^k\,dP(x), \qquad R^A_{k,i}(U,w)=\mathbb E_A[w_i(x)U(x)^k].(6) The weighted means μi(l)\mu_i(l) and variances vi(l)v_i(l) of a moment array ll are the ratios in (7). μi(l)=l1,il0,i,vi(l)=l2,il0,i−μi(l)2.\mu_i(l)=\frac{l_{1,i}}{l_{0,i}},\qquad v_i(l)=\frac{l_{2,i}}{l_{0,i}}-\mu_i(l)^2.(7) For real moment arrays l,sl,s of orders zero, one and two, the moment gain G(l,s)G(l,s) is the real number defined by (8). G(l,s)=(∏il0,i)(1+∑ivi(l)−∑is2,i+2(1−μi(l))s1,i+((1−μi(l))2+∑j≠ivj(l))s0,il0,i).G(l,s)=\left(\prod_i l_{0,i}\right) \left(1+\sum_i v_i(l)-\sum_i \frac{s_{2,i}+2(1-\mu_i(l))s_{1,i} +\left((1-\mu_i(l))^2+\sum_{j\ne i}v_j(l)\right)s_{0,i}} {l_{0,i}}\right).(8) Writing γ\gamma for the standard Gaussian law, the Gaussian leaf and selector moment arrays lG,sGl^G,s^G and the gain threshold η\eta are those of (9). lG=Mγ(id⁡,ℓ),sG=Mγ(id⁡,σ),η=12G(lG,sG).l^G=M^\gamma(\operatorname{id},\ell),\qquad s^G=M^\gamma(\operatorname{id},\sigma),\qquad \eta=\tfrac12 G(l^G,s^G).(9) The finite tests and their centering value for the normalized score in (1) are denoted by (10). si(n)=σi∘Un,wi(n)=ℓi∘Un,l(n)=RAn(Un,w(n)),bn=∑i∈I*μi(l(n))−1.s_i^{(n)}=\sigma_i\circ U_n,\qquad w_i^{(n)}=\ell_i\circ U_n, \qquad l^{(n)}=R^{A^n}(U_n,w^{(n)}),\qquad b_n=\sum_{i\in I_*}\mu_i(l^{(n)})-1.(10)

Deterministic Tables

For finite sets I,BI,B, a selector table SS is a map from BB to II, and a leaf test table L=(Li)i∈IL=(L_i)_{i\in I} is a family of maps Li:B→{0,1}L_i:B\to\{0,1\}, with 11 meaning that the test passes. The unique failing leaf event ES,LE_{S,L} associated with a selector table S:B→IS:B\to I and a leaf test table L=(Li)i∈IL=(L_i)_{i\in I} on BB is the subset of B×BIB\times B^I defined in (11). ES,L={y:LS(y0)(yS(y0))=0,Li(yi)=1 for every i≠S(y0)}.E_{S,L}=\left\{y:L_{S(y_0)}(y_{S(y_0)})=0, \quad L_i(y_i)=1\text{ for every }i\ne S(y_0)\right\}.(11) For real weights pip_i on a finite set CC, the product weight PpP_p is the real function on CIC^I defined in (12). The Bernoulli weight ρp\rho_p for a real number pp is a real function on {0,1}\{0,1\}, and the leaf table weight HpH_p for a real function p:B→ℝp:B\to\mathbb R is a real function on tables L:B→{0,1}L:B\to\{0,1\}; both are defined in (12). Pp(z)=∏i∈Ipi(zi),ρp(e)={pe=1,1−pe=0,Hp(L)=∏y∈Bρp(y)(L(y)).P_p(z)=\prod_{i\in I}p_i(z_i),\qquad \rho_p(e)=\begin{cases}p&e=1,\\1-p&e=0,\end{cases} \qquad H_p(L)=\prod_{y\in B}\rho_{p(y)}(L(y)).(12) For real arrays si,wi:B→ℝs_i,w_i:B\to\mathbb R, the rounding table weight πs,w\pi_{s,w} on pairs (S,L)(S,L) is the product in (13). πs,w(S,L)=(∏y∈BsS(y)(y))∏i∈IHwi(Li).\pi_{s,w}(S,L)=\left(\prod_{y\in B}s_{S(y)}(y)\right) \prod_{i\in I}H_{w_i}(L_i).(13)✦ A label map F:A→BF:A\to B acts on all selector and leaf coordinates through ΦF\Phi_F; its coordinatewise map on length nn samples is FnF_n, as in (14). Write sF=(si∘F)is_F=(s_i\circ F)_i and wF=(wi∘F)iw_F=(w_i\circ F)_i. ΦF(x0,(xi)i)=(F(x0),(F(xi))i),Fn(x)=(F(xj))j=1n.\Phi_F(x_0,(x_i)_i)=(F(x_0),(F(x_i))_i),\qquad F_n(x)=(F(x_j))_{j=1}^n.(14)

The Claim

Theorem 1. Let AA be finite and nonempty, and let T:A→ℝT:A\to\mathbb R have uniform mean zero and second moment one. There is a positive integer nn with the soft gain (15): 𝔼Anwi(n)>0(i∈I*),η<DUn,bn(Ws(n),w(n)).\mathbb E_{A^n}w_i^{(n)}>0\quad(i\in I_*),\qquad \eta<D_{U_n,b_n}\bigl(W_{s^{(n)},w^{(n)}}\bigr).(15)✦ For finite I,BI,B and real arrays si,wi:B→ℝs_i,w_i:B\to\mathbb R, suppose that ∑isi(y)=1\sum_i s_i(y)=1 for every y∈By\in B. The table averages satisfy (16) for every y∈B×BIy\in B\times B^I, and (17) for every map F:A→BF:A\to B, real score TT and center bb: ∑S,Lπs,w(S,L)𝟏ES,L(y)=Ws,w(y).\sum_{S,L}\pi_{s,w}(S,L)\mathbf1_{E_{S,L}}(y)=W_{s,w}(y).(16)✦ ∑S,Lπs,w(S,L)DT,b(𝟏ES,L∘ΦF)=DT,b(WsF,wF).\sum_{S,L}\pi_{s,w}(S,L) D_{T,b}\bigl(\mathbf1_{E_{S,L}}\circ\Phi_F\bigr) =D_{T,b}(W_{s_F,w_F}).(17)✦ Suppose in addition that si(y)≥0s_i(y)\ge0 and 0≤wi(y)≤10\le w_i(y)\le1 for all i,yi,y. For any real c<DT,b(WsF,wF)c<D_{T,b}(W_{s_F,w_F}), there are tables S,LS,L with the deterministic deficit (18): c<DT,b(𝟏ES,L∘ΦF).c<D_{T,b}\bigl(\mathbf1_{E_{S,L}}\circ\Phi_F\bigr).(18)✦ Finally, let F:A→BF:A\to B be a map to a finite label alphabet and let t:B→ℝt:B\to\mathbb R satisfy 𝔼A[t∘F]=0\mathbb E_A[t\circ F]=0 and 𝔼A[(t∘F)2]=1\mathbb E_A[(t\circ F)^2]=1. With UnU_n formed from T=t∘FT=t\circ F, there are a positive integer nn, a real bb, and tables S:Bn→I*S:B^n\to I_* and Li:Bn→{0,1}L_i:B^n\to\{0,1\} with the gain (19): η<DUn,b(𝟏ES,L∘ΦFn).\eta<D_{U_n,b}\bigl(\mathbf1_{E_{S,L}}\circ\Phi_{F_n}\bigr).(19)✦

In the paper, the result reads: the Gaussian event and predictor are fixed before transfer, normalized finite scores have an eventual positive deficit, and the cell cost and retained variance bounds are then applied to amplify.

What Is Checked

The soft and final table gain statements concern centered scores of second moment one on uniform finite nonempty alphabets and their product samples; the averaging identities concern real selector and leaf arrays on finite label sets, with selector normalization, and deterministic extraction also assumes nonnegative selectors and leaf weights in [0,1][0,1].