OctoInformBeta

OpenAI preprint, September 2026 · Section 4. Conditioning on a tested set

Section 4 (proof counterpart): Relative entropy controls sequential exposure with a common unit fallback

Local normalization costs under quantum conditioning.

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 conditioned local purification need not have unit length. Its normalization cost is measured by the squared Hilbert-Schmidt distance to a unit vector, averaged over the original question distribution.

A finite probability law PP on a finite set α\alpha assigns nonnegative weights P(a)P(a) with sum one.✦ The expectation 𝔼P[g]\mathbb E_P[g] of a vector-valued function gg under a finite probability law PP is the weighted sum in (1). 𝔼P[g]=∑a∈αP(a)g(a).\mathbb E_P[g]=\sum_{a\in\alpha}P(a)g(a).(1) The mixture of laws QaQ_a with mixing law PP is the probability law assigning bb the weight ∑aP(a)Qa(b)\sum_a P(a)Q_a(b). The point mass δb\delta_b at bb is the probability law assigning weight one to bb and zero elsewhere.✦ The pushforward f*Pf_*P of a probability law PP under a function ff is the mixture of the point masses δf(a)\delta_{f(a)} with mixing law PP.✦ For a finite coordinate set ι\iota and i∈ιi\in\iota, the coordinate marginal PiP_i of a probability law PP on αι\alpha^{\iota} is the pushforward of PP by evaluation at ii.✦ The product law ⨂iQi\bigotimes_i Q_i of probability laws QiQ_i is the probability law assigning zz the weight ∏iQi(zi)\prod_i Q_i(z_i).✦

Let X,YX,Y be finite nonempty question alphabets, II a finite coordinate set, and RR a finite nonempty record set. A question profile zz on the coordinate set II with question alphabets X,YX,Y is a function z:I→X×Yz\colon I\to X\times Y; write 𝒵=(X×Y)I\mathcal Z=(X\times Y)^I.✦ Fix a law μ\mu on X×YX\times Y and its product law Π=μ⊗I\Pi=\mu^{\otimes I}. Choose default questions x0∈Xx_0\in X, y0∈Yy_0\in Y and an ordered list ℓ=(i0,…,iL−1)\ell=(i_0,\ldots,i_{L-1}) of distinct coordinates.

For a complex matrix TT, its squared Hilbert-Schmidt norm is ∥T∥22=∑a,b|Tab|2\|T\|_2^2=\sum_{a,b}|T_{ab}|^2.✦ A coefficient matrix Γ\Gamma on finite local bases m,nm,n determines the bipartite vector |Γ⟩=∑a,bΓab|a⟩⊗|b⟩|\Gamma\rangle=\sum_{a,b}\Gamma_{ab}|a\rangle\otimes|b\rangle.✦ For local operators AA and BB on Alice’s and Bob’s spaces, respectively, the bipartite operator expectation eΓ(A,B)e_\Gamma(A,B) and its Born probability bΓ(A,B)b_\Gamma(A,B) are the scalars defined by (2): eΓ(A,B)=⟨Γ|A⊗B|Γ⟩,bΓ(A,B)=Re⁡eΓ(A,B).e_\Gamma(A,B)=\langle\Gamma|A\otimes B|\Gamma\rangle, \qquad b_\Gamma(A,B)=\operatorname{Re}e_\Gamma(A,B).(2)

A local quantum event system on question profiles z∈(X×Y)Iz\in(X\times Y)^I, with record set RR and local bases m,nm,n, consists of a coefficient matrix Γ\Gamma with ∥Γ∥2=1\|\Gamma\|_2=1, local effects 0≤Fr(z)≤Im0\le F_r(z)\le I_m and 0≤Hr(z)≤In0\le H_r(z)\le I_n, and weights 0≤θr(z)≤10\le\theta_r(z)\le1. Alice’s effect FrF_r is independent of all Bob questions, and Bob’s effect HrH_r is independent of all Alice questions. The weight θr\theta_r is independent of both questions at each coordinate in ℓ\ell. The event likelihood aa of a local quantum event system is the function in (3), with values at most one and positive mean pp; the conditioned density ρ\rho of the system is the function defined there: a(z)=∑r∈Rθr(z)bΓ(Fr(z),Hr(z))≤1,p=𝔼Π[a]>0,ρ(z)=a(z)p.a(z)=\sum_{r\in R}\theta_r(z)b_\Gamma(F_r(z),H_r(z))\le1, \qquad p=\mathbb E_\Pi[a]>0, \qquad \rho(z)=\frac{a(z)}p.(3)

Conditional Exposure

Normalized nonnegative weights are divided by their total mass when that mass is positive; at zero mass they give a point mass at the chosen default.✦ The conditional laws μ(⋅∣y)\mu(\cdot\mid y) on XX and μ(⋅∣x)\mu(\cdot\mid x) on YY are normalized slices of μ\mu, with defaults x0x_0 and y0y_0, respectively.

For a question profile zz, coordinate i∈Ii\in I and question x∈Xx\in X, the left coordinate replacement zAi←xz_A^{i\leftarrow x} is the question profile obtained by changing Alice’s question at ii to xx, retaining Bob’s question and all other coordinates.✦ The right coordinate replacement zBi←yz_B^{i\leftarrow y} of a question profile zz, at coordinate i∈Ii\in I with question y∈Yy\in Y, changes Bob’s question at ii to yy and retains every other question.✦ A function gg on 𝒵\mathcal Z is left coordinate invariant at i∈Ii\in I when its value is unchanged by every replacement zAi←xz_A^{i\leftarrow x} with x∈Xx\in X, and right coordinate invariant at ii when its value is unchanged by every replacement zBi←yz_B^{i\leftarrow y} with y∈Yy\in Y. For zi=(xi,yi)z_i=(x_i,y_i) and a function g:𝒵→Eg\colon\mathcal Z\to E into a vector space EE, the left and right conditional averaging operators are given by (4): (𝖫ig)(z)=𝔼x∼μ(⋅∣yi)[g(zAi←x)],(𝖱ig)(z)=𝔼y∼μ(⋅∣xi)[g(zBi←y)].(\mathsf L_i g)(z)=\mathbb E_{x\sim\mu(\cdot\mid y_i)} [g(z_A^{i\leftarrow x})], \qquad (\mathsf R_i g)(z)=\mathbb E_{y\sim\mu(\cdot\mid x_i)} [g(z_B^{i\leftarrow y})].(4) Iterated conditional averaging along a list composes these operators in list order, with the last operator applied first and the empty composition equal to the identity. Alice’s and Bob’s exposure chains are the families in (5), for 0≤j≤L0\le j\le L: fj(r,⋅)=𝖫ij⋯𝖫iL−1Fr,hj(r,⋅)=𝖱ij−1⋯𝖱i0Hr.f_j(r,\cdot)=\mathsf L_{i_j}\cdots\mathsf L_{i_{L-1}}F_r, \qquad h_j(r,\cdot)=\mathsf R_{i_{j-1}}\cdots\mathsf R_{i_0}H_r.(5) The exposure index space is J={0,…,L}×R×𝒵J=\{0,\ldots,L\}\times R\times\mathcal Z.✦ Alice’s exposure effect family 𝒜\mathcal A assigns (j,r,z)(j,r,z) the matrix fj(r,z)f_j(r,z)✦; Bob’s exposure effect family ℬ\mathcal B assigns it hj(r,z)h_j(r,z).✦

Common Purifications

For a finite family 𝒯=(Tt)t∈J\mathcal T=(T_t)_{t\in J} of positive effects, the resolvent of TT at s>0s>0 is (T+sI)−1(T+sI)^{-1}✦ and its purification kernel is KT(s)K_T(s) in (6). The joint purification kernel 𝒦𝒯(s)\mathcal K_{\mathcal T}(s) is the horizontal concatenation of the kernels of the family.✦ The common Gram matrix G𝒯G_{\mathcal T} of the effect family 𝒯\mathcal T is the matrix defined by the integral in (6): KT(s)=T(T+sI)−1,𝒦𝒯(s)=[KTt(s)]t∈J,G𝒯=∫0∞𝒦𝒯(s)*𝒦𝒯(s)ds.K_T(s)=T(T+sI)^{-1},\qquad \mathcal K_{\mathcal T}(s)=[K_{T_t}(s)]_{t\in J},\qquad G_{\mathcal T}=\int_0^\infty \mathcal K_{\mathcal T}(s)^*\mathcal K_{\mathcal T}(s)\,ds.(6) The finite common purification at index tt is the ttth block column of G𝒯1/2G_{\mathcal T}^{1/2}.✦ The common resolvent purification FT𝒯F_T^{\mathcal T} of an effect TT in the family 𝒯\mathcal T is the block column of G𝒯1/2G_{\mathcal T}^{1/2} corresponding to a chosen occurrence of TT in 𝒯\mathcal T.✦ The local purification maps Uj,VjU_j,V_j are defined by (7): Uj(r,z)=Ffj(r,z)𝒜,Vj(r,z)=Fhj(r,z)ℬ.U_j(r,z)=F_{f_j(r,z)}^{\mathcal A}, \qquad V_j(r,z)=F_{h_j(r,z)}^{\mathcal B}.(7)

The weighted block state 𝒲(w,Q)\mathcal W(w,Q) with weights wtw_t and coefficient matrices QtQ_t is the block diagonal matrix in (8): 𝒲(w,Q)=⨁twtQt.\mathcal W(w,Q)=\bigoplus_t\sqrt{w_t}\,Q_t.(8)✦ Write zi←qz^{i\leftarrow q} for replacement of the entire question pair at ii by q∈X×Yq\in X\times Y. The conditional block state 𝒮i[U,V](q)\mathcal S_i[U,V](q) associated with local maps U,VU,V and question pair qq at coordinate ii is the coefficient matrix in (9), with block index t=(r,z)∈R×𝒵t=(r,z)\in R\times\mathcal Z: 𝒮i[U,V](q)=𝒲((Π(z)θr(z)p)r,z,(U(r,zi←q)ΓV(r,zi←q)𝖳)r,z).\mathcal S_i[U,V](q)=\mathcal W\left( \left(\frac{\Pi(z)\theta_r(z)}p\right)_{r,z}, \left(U(r,z^{i\leftarrow q})\Gamma V(r,z^{i\leftarrow q})^{\mathsf T}\right)_{r,z} \right).(9) For 0≤j<L0\le j<L, the joint conditional state DjD_j and the local marginal states ΨjA,ΨjB\Psi_j^A,\Psi_j^B are defined by (10): Dj(x,y)=𝒮ij[Uj+1,Vj](x,y),ΨjA(x)=𝒮ij[Uj+1,Vj+1](x,y0),ΨjB(y)=𝒮ij[Uj,Vj](x0,y).\begin{aligned} D_j(x,y)&=\mathcal S_{i_j}[U_{j+1},V_j](x,y),\\ \Psi_j^A(x)&=\mathcal S_{i_j}[U_{j+1},V_{j+1}](x,y_0),\\ \Psi_j^B(y)&=\mathcal S_{i_j}[U_j,V_j](x_0,y). \end{aligned}(10) These coefficient matrices have row index (J×m)×(R×𝒵)(J\times m)\times(R\times\mathcal Z) and column index (J×n)×(R×𝒵)(J\times n)\times(R\times\mathcal Z).

For a unit coefficient matrix EE on the enlarged coefficient spaces, Hilbert-Schmidt normalization 𝒩E\mathcal N_E with fallback EE is the map in (11). Alice’s and Bob’s normalized purifications are MjEM_j^E and NjEN_j^E: 𝒩E(Q)={Q/∥Q∥2,∥Q∥2>0,E,∥Q∥2=0,MjE(x)=𝒩E(ΨjA(x)),NjE(y)=𝒩E(ΨjB(y)).\mathcal N_E(Q)= \begin{cases} Q/\|Q\|_2,&\|Q\|_2>0,\\ E,&\|Q\|_2=0, \end{cases} \qquad M_j^E(x)=\mathcal N_E(\Psi_j^A(x)), \qquad N_j^E(y)=\mathcal N_E(\Psi_j^B(y)).(11)

Relative Entropy

A law PP is absolutely continuous with respect to QQ, written P≪QP\ll Q, when P(a)>0P(a)>0 entails Q(a)>0Q(a)>0.✦ Their finite relative entropy is the sum in (12), with terms having P(a)Q(a)=0P(a)Q(a)=0 assigned zero. All logarithms are natural. D(P‖Q)=∑a∈αP(a)log⁡P(a)Q(a).D(P\Vert Q)=\sum_{a\in\alpha}P(a)\log\frac{P(a)}{Q(a)}.(12)✦ For a nonnegative function gg with 𝔼P[g]=1\mathbb E_P[g]=1, the tilted law PgP^g on α\alpha is defined by (13): Pg(a)=P(a)g(a).P^g(a)=P(a)g(a).(13)✦

The Claim

Theorem 1. For a local quantum event system as in (3), the conditioned density satisfies (14) at every question profile: ρ(z)≤1p(z∈𝒵).\rho(z)\le\frac1p\qquad(z\in\mathcal Z).(14)✦

Let PP be a finite probability law and gg a nonnegative function with 𝔼P[g]=1\mathbb E_P[g]=1. If B>0B>0 and g(a)≤Bg(a)\le B for every aa, the tilted relative entropy satisfies (15): D(Pg‖P)≤log⁡B.D(P^g\Vert P)\le\log B.(15)✦

Let ν\nu be a law on αι\alpha^{\iota} for a finite coordinate set ι\iota, and let QiQ_i be laws on α\alpha. If ν≪⨂iQi\nu\ll\bigotimes_i Q_i, its coordinate marginals obey the relative entropy tensorization inequality (16): ∑i∈ιD(νi‖Qi)≤D(ν‖⨂i∈ιQi).\sum_{i\in\iota}D(\nu_i\Vert Q_i) \le D\!\left(\nu\,\middle\Vert\,\bigotimes_{i\in\iota}Q_i\right).(16)✦

For every unit fallback matrix EE on the enlarged spaces, the local normalization costs satisfy (17): ∑j=0L−1𝔼(x,y)∼μ[∥ΨjA(x)−MjE(x)∥22]≤log⁡(1/p),∑j=0L−1𝔼(x,y)∼μ[∥ΨjB(y)−NjE(y)∥22]≤log⁡(1/p).\begin{aligned} \sum_{j=0}^{L-1}\mathbb E_{(x,y)\sim\mu} \bigl[\|\Psi_j^A(x)-M_j^E(x)\|_2^2\bigr]&\le\log(1/p),\\ \sum_{j=0}^{L-1}\mathbb E_{(x,y)\sim\mu} \bigl[\|\Psi_j^B(y)-N_j^E(y)\|_2^2\bigr]&\le\log(1/p). \end{aligned}(17)✦ The normalized approximation errors satisfy (18): ∑j=0L−1𝔼(x,y)∼μ[∥Dj(x,y)−MjE(x)∥22]≤4log⁡(|R|/p),∑j=0L−1𝔼(x,y)∼μ[∥Dj(x,y)−NjE(y)∥22]≤4log⁡(|R|/p).\begin{aligned} \sum_{j=0}^{L-1}\mathbb E_{(x,y)\sim\mu} \bigl[\|D_j(x,y)-M_j^E(x)\|_2^2\bigr]&\le4\log(|R|/p),\\ \sum_{j=0}^{L-1}\mathbb E_{(x,y)\sim\mu} \bigl[\|D_j(x,y)-N_j^E(y)\|_2^2\bigr]&\le4\log(|R|/p). \end{aligned}(18)✦

What Is Checked

The library’s statements concern finite probability laws with the stated support conditions and local quantum event systems with distinct exposure coordinates, using common resolvent purifications and an arbitrary unit fallback matrix.