OctoInformBeta

OpenAI preprint, September 2026 · §3. Common mass, smoothing, and score moments

Lemma 3.3: Common events give finite source oracles with recursive error budgets

Clipping controls loss and score moments.

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

Clipping two trace experiments on the same event controls the mass discarded from both laws and the likelihood score of the resulting pair. The finite source experiments here consist of paired side reports and recursively selected leaf laws.

A finite probability law PP on a finite set Ω\Omega is a nonnegative mass function with ∑x∈ΩP(x)=1\sum_{x\in\Omega}P(x)=1.✦ For a real function hh on Ω\Omega, an event A⊆ΩA\subseteq\Omega and a finite probability law QQ on Ω\Omega, expectation, event probability and total variation distance are given in (1). 𝔼Ph=∑xP(x)h(x),P(A)=∑x∈AP(x),TV⁡(P,Q)=12∑x|P(x)−Q(x)|.\mathbb E_P h=\sum_xP(x)h(x),\qquad P(A)=\sum_{x\in A}P(x),\qquad \operatorname{TV}(P,Q)=\frac12\sum_x|P(x)-Q(x)|.(1) For a finite reference law RR on Ω\Omega and a nonnegative density ff on Ω\Omega with 𝔼Rf=1\mathbb E_R f=1, the density law PfP_f is the probability law with mass Pf(x)=R(x)f(x)P_f(x)=R(x)f(x). Conditioning a law PP on an event AA of positive probability gives the law PAP_A in (2). For an event E⊆ΩE\subseteq\Omega and a density ff on Ω\Omega, the common clipping event A(E,f)A(E,f) is the subset of Ω\Omega given in (2). PA(x)=𝟏A(x)P(x)P(A),A(E,f)=E∩{x:f(x)≥12}.P_A(x)=\frac{\mathbf1_A(x)P(x)}{P(A)},\qquad A(E,f)=E\cap\{x:f(x)\geq\tfrac12\}.(2) The squared even moment norm of a real function hh under PP, with half-order k≥1k\geq1, is the probability-weighted quantity in (3). ∥h∥P,2k,2=(𝔼P|h|2k)1/k.\|h\|_{P,2k}^{,2} =\bigl(\mathbb E_P|h|^{2k}\bigr)^{1/k}.(3)✦ A clipped oracle for laws P,QP,Q, with parameters k,e,Vk,e,V, is a pair of finite laws μ,ν\mu,\nu and a real score Z*Z_* satisfying (4); denote the set of these triples by 𝒞k(P,Q;e,V)\mathscr C_k(P,Q;e,V). ν(x)=μ(x)(1+Z*(x)),1+Z*(x)≥0,𝔼μZ*=0,∥Z*∥μ,2k,2≤V,TV⁡(μ,P)≤e,TV⁡(ν,Q)≤e.\begin{gathered} \nu(x)=\mu(x)(1+Z_*(x)),\qquad 1+Z_*(x)\geq0,\qquad \mathbb E_\mu Z_*=0,\\ \|Z_*\|_{\mu,2k}^{,2}\leq V,\qquad \operatorname{TV}(\mu,P)\leq e,\qquad \operatorname{TV}(\nu,Q)\leq e. \end{gathered}(4)✦

Paired Reports and Source Windows

The product law P⊗QP\otimes Q and the coordinate product law ⨂iPi\bigotimes_iP_i are laws with the masses in (5). For a probability kernel KK, the joint law P⋉KP\ltimes K has the mass there; a mixture of laws QbQ_b with mixing law RR has mass ∑bR(b)Qb(x)\sum_bR(b)Q_b(x). (P⊗Q)(x,y)=P(x)Q(y),(⨂iPi)(x)=∏iPi(xi),(P⋉K)(x,y)=P(x)Kx(y).(P\otimes Q)(x,y)=P(x)Q(y),\qquad \left(\bigotimes_iP_i\right)(x)=\prod_iP_i(x_i),\qquad (P\ltimes K)(x,y)=P(x)K_x(y).(5) A family of paired mechanisms on a finite side space SS consists of laws Pℓ,Qℓ,HℓP_\ell,Q_\ell,H_\ell and weights ωℓ≥0\omega_\ell\geq0 with ωℓHℓ(x)≤Pℓ(x),Qℓ(x)\omega_\ell H_\ell(x)\leq P_\ell(x),Q_\ell(x) at every outcome.✦ For a paired mechanism with laws Pℓ,QℓP_\ell,Q_\ell on SS and an orientation β∈{0,1}\beta\in\{0,1\}, the oriented pair law OℓβO_\ell^\beta on S2S^2 is Oℓ0=Pℓ⊗QℓO_\ell^0=P_\ell\otimes Q_\ell or Oℓ1=Qℓ⊗PℓO_\ell^1=Q_\ell\otimes P_\ell.✦ The hidden pair mass hℓh_\ell, flagged pair mass FℓβF_\ell^\beta, and paired report law JℓJ_\ell are defined by (6); ⋆\star denotes an unreported orientation, with the side pair still returned. hℓ(x,y)=ωℓ2Hℓ(x)Hℓ(y),Fℓβ(d,x,y)={hℓ(x,y),d=⋆,Oℓβ(x,y)−hℓ(x,y),d=β,0,d∈{0,1}\{β},Jℓ=12(Fℓ0+Fℓ1).\begin{gathered} h_\ell(x,y)=\omega_\ell^2H_\ell(x)H_\ell(y),\\ F_\ell^\beta(d,x,y)= \begin{cases} h_\ell(x,y),&d=\star,\\ O_\ell^\beta(x,y)-h_\ell(x,y),&d=\beta,\\ 0,&d\in\{0,1\}\setminus\{\beta\}, \end{cases} \qquad J_\ell=\tfrac12(F_\ell^0+F_\ell^1). \end{gathered}(6) The report space for mm mechanisms is ℛm=({⋆,0,1}×S2)m\mathcal R_m=(\{\star,0,1\}\times S^2)^m✦, and its joint report law is 𝒥=⨂ℓ=0m−1Jℓ\mathcal J=\bigotimes_{\ell=0}^{m-1}J_\ell.✦ For a reported flag d∈{⋆,0,1}d\in\{\star,0,1\}, the orientation posterior πd\pi_d is the probability law on {0,1}\{0,1\} that is uniform for d=⋆d=\star and is the point mass at dd otherwise.✦ For a report vector zz, the joint orientation posterior is Πz=⨂ℓπdℓ\Pi_z=\bigotimes_\ell\pi_{d_\ell}.✦

For leaf laws P,QP,Q on a finite set UU and a Boolean labelling ϑ=(ϑv)v\vartheta=(\vartheta_v)_v of a finite set of leaves indexed by vv, the selected leaf law is ℒϑ=⨂v(P if ϑv=0,Q if ϑv=1)\mathcal L_\vartheta=\bigotimes_v(P\text{ if } \vartheta_v=0,\ Q\text{ if }\vartheta_v=1). A rule θ\theta from orientation vectors to leaf labellings defines the canonical source oracle 𝒪θ\mathcal O_\theta, the joint law in (7). 𝒪θ(z,u)=𝒥(z)∑β∈{0,1}mΠz(β)ℒθ(β)(u).\mathcal O_\theta(z,u) =\mathcal J(z)\sum_{\beta\in\{0,1\}^m} \Pi_z(\beta)\mathcal L_{\theta(\beta)}(u).(7)✦

Let Th,thT_h,t_h be natural window and local trial counts. All window slots and nesting indices are zero-based; write [m]={0,…,m−1}[m]=\{0,\ldots,m-1\}. The scheduled global trial counts ghg_h, window leaf sets WhW_h, trial sets IhI_h, and collection sets ChC_h are the recursive objects in (8). g0=0,gh+1=⌊Th+1/4⌋th,W0={*},Wh+1=[Th]×Wh,I0=⌀,Ih+1=[th]⊔[gh]⊔([Th]×Ih),C0=⌀,Ch+1={0,1}⊔([Th]×Ch).\begin{gathered} g_0=0,\qquad g_{h+1}=\lfloor T_{h+1}/4\rfloor t_h,\\ W_0=\{*\},\qquad W_{h+1}=[T_h]\times W_h,\\ I_0=\varnothing,\qquad I_{h+1}=[t_h]\sqcup[g_h]\sqcup([T_h]\times I_h),\\ C_0=\varnothing,\qquad C_{h+1}=\{0,1\}\sqcup([T_h]\times C_h). \end{gathered}(8) The source leaf and trial counts are ℓh=|Wh|\ell_h=|W_h|✦ and mh=|Ih|m_h=|I_h|✦. A window orientation assignment at depth h+1h+1 is a choice of tht_h local bits, ghg_h global bits and one depth-hh assignment for each child; at depth zero the assignment is unique.✦ These assignments are identified with Boolean functions on IhI_h✦, and enumeration gives vectors in {0,1}mh\{0,1\}^{m_h}.✦ The recursive source leaf orientation propagates the integer template label to each child, using the local assignments and incoming shift for the node anchor and the global assignments for the child’s scheduled shift. With motif parameter ss, local centers ch=⌊Th/2⌋−1+⌊th/2⌋c_h=\lfloor T_h/2\rfloor-1+\lfloor t_h/2\rfloor, incoming shift zero and root label τ\tau, denote its leaf vector by θhτ(β)∈{0,1}ℓh\theta_h^\tau(\beta)\in\{0,1\}^{\ell_h}; at depth zero its label is τ\tau.✦

The Recursive Budgets

A collection v∈Chv\in C_h has trial count νh(v)\nu_h(v): the two root collections have counts th−1t_{h-1} and gh−1g_{h-1}, and a descendant collection keeps its count in the child window.✦ For paired mechanisms with common hidden weight ω\omega, the hidden metadata tail Hh(ω)H_h(\omega) and fitting tail FhF_h are the quantities in (9). Hh(ω)=∑v∈Chνh(v)>0exp⁡(−νh(v)ω4/16),F0=F1=0,Fh+2=Th+1[2exp(−2(Th/8)2(Th+1+1)th)+Fh+1].\begin{gathered} H_h(\omega)=\sum_{\substack{v\in C_h\\\nu_h(v)>0}} \exp(-\nu_h(v)\omega^4/16),\qquad F_0=F_1=0,\\ F_{h+2}=T_{h+1}\left[ 2\exp\left(-\frac{2(T_h/8)^2}{(T_{h+1}+1)t_h}\right)+F_{h+1}\right]. \end{gathered}(9) For a natural depth bb and positive real parameters a,La,L, the source sparsity count SbS_b, centered moment budget Ub,k(L)U_{b,k}(L) and clipping cost eb,k(ω,a,L)e_{b,k}(\omega,a,L) are defined by (10), with w=6(2k)Lw=6(2k)L. Sb=2s∏j=1bTj,Ub,k(L)=w2Sb1−w2Sb,eb,k(ω,a,L)=Hb+1(ω)+Fb+1a+6(4Ub,k(L))k.\begin{gathered} S_b=2^s\prod_{j=1}^bT_j,\qquad U_{b,k}(L)=\frac{w^2S_b}{1-w^2S_b},\\ e_{b,k}(\omega,a,L) =H_{b+1}(\omega)+\frac{F_{b+1}}a+ \sqrt{6\bigl(4U_{b,k}(L)\bigr)^k}. \end{gathered}(10) The hidden binomial cap κm(ω)\kappa_m(\omega) and hidden smoothing bound ℋm,s(ω)\mathcal H_{m,s}(\omega) are the functions in (11). κm(ω)=4mω2+2,ℋm,s(ω)=2s2(24(s+1)mω2+2)s4(s+1)mω2+2.\begin{gathered} \kappa_m(\omega)=\sqrt{\frac4{m\omega^2+2}},\\ \mathcal H_{m,s}(\omega)=2^{s^2} \left(2\sqrt{\frac{4(s+1)}{m\omega^2+2}}\right)^s \sqrt{\frac{4(s+1)}{m\omega^2+2}}. \end{gathered}(11) For a natural branching bound rr, a finite ordered role shape at depth zero is unique; at depth h+1h+1 it is an ordered list of at most rr depth-hh shapes, each assigned one of three roles. Let Kr,hK_{r,h} be the number of such shapes.✦ The interior label budget Ib,rlabI_{b,r}^{\mathrm{lab}} and binary tree budget Bh(c)B_h(c) are defined in (12), where (r−2b+1)+=max⁡{r−2b+1,0}(r-2^{b+1})_+=\max\{r-2^{b+1},0\}. Ib,rlab=(∏j=1bTj)(r−2b+1)+,B0(c)=1,Bh+1(c)=chBh(c)2.I_{b,r}^{\mathrm{lab}}=\left(\prod_{j=1}^bT_j\right)^{(r-2^{b+1})_+},\qquad B_0(c)=1,\qquad B_{h+1}(c)=c_hB_h(c)^2.(12) For chhid=4(2s)6κgh+1(ω)κth(ω)c_h^{\mathrm{hid}}=4(2^s)^6\kappa_{g_{h+1}}(\omega)\kappa_{t_h}(\omega), the forced contribution budget Ab,rA_{b,r} is given by (13). Ab,r=Kr,b+1Ib,rlab(2s)2(r−2b+1)+2(2s)3κtb(ω)Bb(chid).A_{b,r}=K_{r,b+1}I_{b,r}^{\mathrm{lab}}(2^s)^{2(r-2^{b+1})_+} \,2(2^s)^3\kappa_{t_b}(\omega)B_b(c^{\mathrm{hid}}).(13)✦ For a smoothing bound η\eta and a natural degree cutoff R*R_*, the sharp degree budget Db,r(ω,η,a)D_{b,r}(\omega,\eta,a) and sharp energy budget ℰb,R*(ω,η,a,w)\mathcal E_{b,R_*}(\omega,\eta,a,w) are the quantities in (14). Db,r=43(b+1)r[(ℓb+1r)3(b+1)r(η2+a2)+𝟏{r≥2b+1}Ab,r],ℰb,R*(ω,η,a,w)=∑1≤r<R*w2rDb,r+4(w2Sb)R*1−w2Sb.\begin{gathered} D_{b,r}=4\,3^{(b+1)r}\left[ \binom{\ell_{b+1}}r3^{(b+1)r}(\eta^2+a^2) +\mathbf1_{\{r\geq2^{b+1}\}}A_{b,r}\right],\\ \mathcal E_{b,R_*}(\omega,\eta,a,w) =\sum_{1\leq r<R_*}w^{2r}D_{b,r} +\frac{4(w^2S_b)^{R_*}}{1-w^2S_b}. \end{gathered}(14)

For block parameter BB and depth DD, the warmup block count is 5⋅2B5\cdot2^B.✦ The dimension bi+1b_{i+1} of zero-based block ii, scale slack ε\varepsilon, scale exponent α\alpha, block gain GG and total score exponent dB,Dd_{B,D} are defined in (15). bi+1={2,i<5⋅2B,B,i≥5⋅2B,ε=124B+100(D+1),α(b,j;ε)=(1+ε)2j2b−ε,G(b)=b−14−120,dB,D=115+∑i=0D−1G(bi+1).\begin{gathered} b_{i+1}=\begin{cases}2,&i<5\cdot2^B,\\B,&i\geq5\cdot2^B,\end{cases} \qquad \varepsilon=\frac1{2^{4B+100}(D+1)},\\ \alpha(b,j;\varepsilon)=(1+\varepsilon)\frac{2^j}{2^b}-\varepsilon, \qquad G(b)=\frac{b-1}4-\frac1{20},\\ d_{B,D}=\frac{11}5+\sum_{i=0}^{D-1}G(b_{i+1}). \end{gathered}(15) For problem size nn, the scale exponent product pip_i, backwards target scale NiN_i, half moment order kik_i, and gap floor gBg_B are the quantities in (16). Here ⌈⋅⌉+\lceil\cdot\rceil_+ is the ceiling truncated below at zero, and the even score exponent is ri=2kir_i=2k_i. pi=∏j=iD−1α(bj+1,0;ε),Ni=npi,χ=dB,D+B+1,ki=⌈108210Bχlog⁡nlog⁡Ni⌉+,gB=112822B.\begin{gathered} p_i=\prod_{j=i}^{D-1}\alpha(b_{j+1},0;\varepsilon),\qquad N_i=n^{p_i},\\ \chi=d_{B,D}+B+1,\qquad k_i=\left\lceil10^8\,2^{10B}\chi\frac{\log n}{\log N_i}\right\rceil_+, \qquad g_B=\frac1{128\,2^{2B}}. \end{gathered}(16)

The Claim

Theorem 1. Let RR be a finite law and let f,gf,g be nonnegative densities of mean one under RR. Fix k≥1k\geq1.

Suppose Pf(Ec),Pg(Ec)≤eP_f(E^c),P_g(E^c)\leq e, 0≤U≤10\leq U\leq1, and 𝔼R|f−1|2k,𝔼R|g−1|2k≤Uk\mathbb E_R|f-1|^{2k},\mathbb E_R|g-1|^{2k}\leq U^k. The common clipping loss satisfies (17). max⁡{1−Pf(A(E,f)),1−Pg(A(E,f))}≤e+6(4U)k.\max\{1-P_f(A(E,f)),\ 1-P_g(A(E,f))\} \leq e+\sqrt{6(4U)^k}.(17)✦ Suppose an event AA satisfies f≥12f\geq\tfrac12 on AA, and put pf=Pf(A)≥12p_f=P_f(A)\geq\tfrac12, pg=Pg(A)≥12p_g=P_g(A)\geq\tfrac12. For V≥0V\geq0, assume 𝔼R|𝟏A(g−f)|2k≤Vk\mathbb E_R|\mathbf1_A(g-f)|^{2k}\leq V^k. The conditioned score satisfies (18). ∥pfpggf−1∥(Pf)A,2k,2≤64V.\left\|\frac{p_f}{p_g}\frac gf-1\right\|_{(P_f)_A,2k}^{,2} \leq64V.(18)✦

For the source construction, let all ThT_h be positive and let the paired mechanisms have mb+1m_{b+1} coordinates and common hidden weight ω\omega, with 0≤ω2≤10\leq\omega^2\leq1. Let P,QP,Q be leaf laws with Q(x)=P(x)(1+Z(x))Q(x)=P(x)(1+Z(x)), 1+Z(x)≥01+Z(x)\geq0 and 𝔼P|Z|2k≤L2k\mathbb E_P|Z|^{2k}\leq L^{2k}, where a,L>0a,L>0. Assume tjt_j is positive and even for j<bj<b; for j≤bj\leq b, assume TjT_j is even, Tj≥24T_j\geq24, and Tj≥24⋅2sT_j\geq24\cdot2^s. Suppose also that 4(tb+1)≤Tb4(t_b+1)\leq T_b, w2Sb≤12w^2S_b\leq\tfrac12, ℋtj,s(ω)≤η\mathcal H_{t_j,s}(\omega)\leq\eta for 0≤j≤b0\leq j\leq b, and eb,k(ω,a,L)≤12e_{b,k}(\omega,a,L)\leq\tfrac12, with w=6(2k)Lw=6(2k)L. There is a clipped source oracle as in (19), for every natural cutoff R*R_*. 𝒞k(𝒪θb+10,𝒪θb+11;eb,k(ω,a,L),64ℰb,R*(ω,η,a,6(2k)L))≠⌀.\mathscr C_k\left( \mathcal O_{\theta_{b+1}^0},\mathcal O_{\theta_{b+1}^1}; e_{b,k}(\omega,a,L),\, 64\mathcal E_{b,R_*}(\omega,\eta,a,6(2k)L)\right) \ne\varnothing.(19)✦

For the level budget, let B≥2B\geq2, n≥4n\geq4, and let the real gap parameter γ\gamma satisfy γ≥gB\gamma\geq g_B. Fix a natural index ii and put wi=6riLw_i=6r_iL. Suppose each of Hb+1(ω)H_{b+1}(\omega), Fb+1/aF_{b+1}/a and Fb+1F_{b+1} is at most n−200χn^{-200\chi}, and wi2Sb≤min⁡{12,Ni−γ/2/8}w_i^2S_b\leq\min\{\tfrac12,N_i^{-\gamma/2}/8\}. The level clipping budgets satisfy (20). eb,ki(ω,a,L)≤12,eb,ki(ω,a,L)+Fb+1≤n−100χ.\begin{gathered} e_{b,k_i}(\omega,a,L)\leq\tfrac12,\\ e_{b,k_i}(\omega,a,L)+F_{b+1}\leq n^{-100\chi}. \end{gathered}(20)✦

In the paper, the result reads: for fixed global parameters, a level 0≤i≤D0\leq i\leq D with score exponent did_i, and error scale Δ\Delta, preliminary laws with common side metadata and conditional nonnegative unit-integral densities Gx,GyG_x,G_y relative to a common conditional reference, whose centered LriL^{r_i} norms are at most Ni−ρN_i^{-\rho} for ρ≥2−4B−10\rho\geq2^{-4B-10} and whose difference has squared LriL^{r_i} norm at most Ni−di−0.01N_i^{-d_i-0.01} uniformly on a metadata event of probability at least 1−o(Δ)1-o(\Delta), can both be conditioned on the joint event consisting of that metadata event and Gx≥12G_x\geq\tfrac12, changing each law by o(Δ)o(\Delta) in total variation and yielding laws ν≪μ\nu\ll\mu with ∥dν/dμ−1∥μ,ri,2≤Ni−di\|d\nu/d\mu-1\|_{\mu,r_i}^{,2}\leq N_i^{-d_i} uniformly over levels.

What Is Checked

The claims concern finite density experiments and canonical source oracles formed from paired mechanisms, recursive window assignments and selected product leaf laws. For finite laws P,Q,HP,Q,H and any real weight ω\omega, the paired report masses sum to one. If ω≥0\omega\geq0 and ωH(x)≤P(x),Q(x)\omega H(x)\leq P(x),Q(x) at every outcome, those masses are nonnegative.