OctoInformBeta

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

Lemma 4.1: Coherent state errors are controlled by the event probability and the number of records

Mixed exposure preserves conditional answer statistics.

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

Conditioning a repeated game on wins at a tested set changes the question and answer statistics at the remaining coordinates. Mixed exposure associates a joint amplitude and one amplitude for each player’s question with each position in an ordering of those coordinates.

Let X,YX,Y be finite nonempty question alphabets and II a finite coordinate set of size kk. A question profile z∈(X×Y)Iz\in(X\times Y)^I assigns a question pair to each coordinate.✦ The question law μ\mu on X×YX\times Y is a family of nonnegative weights summing to one; write π(z)=∏i∈Iμ(zi)\pi(z)=\prod_{i\in I}\mu(z_i) for its product law, and 𝔼πf=∑zπ(z)f(z)\mathbb E_\pi f=\sum_z\pi(z)f(z) for expectation. The conditional laws μ(y∣x)\mu(y\mid x) and μ(x∣y)\mu(x\mid y) normalize the corresponding joint weights; at a zero marginal they are the point masses at fixed defaults y0y_0 and x0x_0, respectively. A point mass assigns weight one to its chosen point and zero elsewhere.✦ For nonnegative weights ww on a finite set, their normalized probability law assigns weight w(x)/∑yw(y)w(x)/\sum_y w(y) to xx when the total is positive, and is the point mass at a chosen default when the total is zero.

For a coefficient matrix SS between finite local spaces, the associated bipartite vector has coefficients SuvS_{uv} in the product basis.✦ For local operators E,HE,H, the squared Hilbert-Schmidt norm ∥S∥22\|S\|_2^2 of the coefficient matrix SS and its Born functional χS(E,H)\chi_S(E,H) are the real quantities given by (1); the latter is the real part of the bipartite operator expectation.✦ ∥S∥22=∑u,v|Suv|2,χS(E,H)=Re⁡⟨S|E⊗H|S⟩.\|S\|_2^2=\sum_{u,v}|S_{uv}|^2, \qquad \chi_S(E,H)=\operatorname{Re}\langle S|E\otimes H|S\rangle.(1) A local positive operator-valued measure (POVM) PP with finite outcome alphabet AA on a finite-dimensional Hilbert space is a family of positive semidefinite effects PaP^a, indexed by a∈Aa\in A, summing to the identity.✦ A coarsening of a POVM by an outcome map is a POVM whose effects are the sums over the map’s fibers✦; a block-diagonal POVM on an additional local register is a POVM with one POVM in each register block.✦ For local POVMs P=(Pa)aP=(P^a)_a and Q=(Qb)bQ=(Q^b)_b on a normalized state with coefficient matrix SS, the joint measurement law assigns weight χS(Pa,Qb)\chi_S(P^a,Q^b) to (a,b)(a,b).✦ The measurement payoff is the sum of these weights over accepted answer pairs.✦

For finite answer alphabets A,BA,B, a repeated quantum experiment consists of μ\mu, a coefficient matrix SS with ∥S∥22=1\|S\|_2^2=1, local POVMs PxIP_{x^I} and QyIQ_{y^I} with outcomes in AIA^I and BIB^I, and a Boolean verifier V(x,y,a,b)V(x,y,a,b).✦ A transcript (z,aI,bI)(z,a^I,b^I) records all questions and answers.✦ Its conditional answer law and joint transcript law are given by (2). Pr⁡(aI,bI∣z)=χS(PxIaI,QyIbI),Pr⁡(z,aI,bI)=π(z)Pr⁡(aI,bI∣z),z=(xI,yI).\Pr(a^I,b^I\mid z) =\chi_S(P_{x^I}^{a^I},Q_{y^I}^{b^I}), \qquad \Pr(z,a^I,b^I)=\pi(z)\Pr(a^I,b^I\mid z), \quad z=(x^I,y^I).(2) For a tested set C⊆IC\subseteq I, let ECE_C be acceptance at every coordinate in CC, and let pC=Pr⁡(EC)p_C=\Pr(E_C), the expectation of its indicator under (2).✦ A tested answer record is r=(α,γ)∈RC=AC×BCr=(\alpha,\gamma)\in R_C=A^C\times B^C, obtained by restricting the full answer strings to CC.✦ The tested wins indicator θC\theta_C for the tested set CC is the function given by θC(r,z)=𝟏{V(xi,yi,αi,γi)=1 for all i∈C}\theta_C(r,z)=\mathbf 1\{V(x_i,y_i,\alpha_i,\gamma_i)=1\text{ for all }i\in C\}.✦ The tested answer POVMs coarsen PxIP_{x^I} and QyIQ_{y^I} by these restriction maps; write their effects as Fr(z)F_r(z) and Hr(z)H_r(z). Their coordinate refinement POVMs retain both the restricted answers and the answer at ii; write the corresponding effects as Li,ra(z)L_{i,r}^a(z) and Ni,rb(z)N_{i,r}^b(z). For C≠IC\ne I with pC>0p_C>0 define the parameters in (3), with natural logarithms. d=|A||B|,τ(C)=log⁡(1/pC)+|C|log⁡dk−|C|.d=|A||B|, \qquad \tau(C)=\frac{\log(1/p_C)+|C|\log d}{k-|C|}.(3) The coordinate winning predicate at ii for a transcript (z,aI,bI)(z,a^I,b^I) is the verifier decision V(xi,yi,ai,bi)V(x_i,y_i,a_i,b_i) on the questions and answers recorded there.✦

Mixed Exposure and Coherent States

Write zi:xz^{i:x} and zi:yz^{i:y} for replacement of only Alice’s or only Bob’s question at ii, and zi:qz^{i:q} for replacement of the pair. A matrix family ff is left coordinate invariant at ii if f(zi:x)=f(z)f(z^{i:x})=f(z) for every z,xz,x; it is right coordinate invariant at ii under the corresponding condition for zi:yz^{i:y}. The conditional averaging operators ℒi,ℛi\mathcal L_i,\mathcal R_i on matrix families ff are given by (4). (ℒif)(z)=∑xμ(x∣yi)f(zi:x),(ℛif)(z)=∑yμ(y∣xi)f(zi:y).(\mathcal L_i f)(z)=\sum_x\mu(x\mid y_i)f(z^{i:x}), \qquad (\mathcal R_i f)(z)=\sum_y\mu(y\mid x_i)f(z^{i:y}).(4) For a list t=(t0,…,ts−1)t=(t_0,\ldots,t_{s-1}), let ℒt=ℒt0⋯ℒts−1\mathcal L_t=\mathcal L_{t_0}\cdots\mathcal L_{t_{s-1}} and define ℛt\mathcal R_t in the same order, so the last listed operator acts first. For a list l=(i0,…,iℓ−1)l=(i_0,\ldots,i_{\ell-1}) of distinct coordinates, Alice’s exposure chain fj,rf_{j,r} of FrF_r and Bob’s exposure chain hj,rh_{j,r} of HrH_r are the matrix families in (5); the suffix and prefix are truncated at the length of the list. fj,r=ℒl≥jFr,hj,r=ℛrev⁡(l<j)Hr,j≥0.f_{j,r}=\mathcal L_{l_{\ge j}}F_r, \qquad h_{j,r}=\mathcal R_{\operatorname{rev}(l_{<j})}H_r, \qquad j\ge0.(5) Here FrF_r is independent of all Bob questions and HrH_r of all Alice questions. The real weights θr(z)\theta_r(z) are invariant under either replacement at every coordinate in ll.

A conditioned quantum event system for the question law μ\mu and exposure list ll consists of local matrix families Fr,HrF_r,H_r, real weights θr\theta_r invariant under both question replacements along ll, a normalized coefficient matrix SS, and a positive event mass pp, with 0≤Fr,Hr≤I0\le F_r,H_r\le I, 0≤θr≤10\le\theta_r\le1, and pointwise event weight ∑rθrχS(Fr,Hr)≤1\sum_r\theta_r\chi_S(F_r,H_r)\le1 whose π\pi expectation is pp.✦ The conditioned density ρ\rho and coordinate marginal density ρi\rho_i of a conditioned quantum event system are the functions given by (6). ρ(z)=1p∑rθr(z)χS(Fr(z),Hr(z)),ρi(q)=𝔼z∼πρ(zi:q).\rho(z)=\frac1p\sum_r\theta_r(z)\chi_S(F_r(z),H_r(z)), \qquad \rho_i(q)=\mathbb E_{z\sim\pi}\rho(z^{i:q}).(6) The tested event system uses RCR_C, the tested effects and indicator, p=pC>0p=p_C>0, and an ordering of I\CI\setminus C.✦

For the finite record set RR indexing the event families and weights, the exposure index set is J={0,…,ℓ}×R×(X×Y)IJ=\{0,\ldots,\ell\}\times R\times(X\times Y)^I. Alice’s event effect family 𝒜\mathcal A lists every fj,r(z)f_{j,r}(z) for these indices; Bob’s family ℬ\mathcal B lists every hj,r(z)h_{j,r}(z). For a finite matrix family 𝒯=(Tt)t∈J\mathcal T=(T_t)_{t\in J}, the resolvent kernel KT(s)K_T(s) of a member TT, the joint kernel 𝒦𝒯(s)\mathcal K_{\mathcal T}(s) of the family, and its common Gram matrix G𝒯G_{\mathcal T} are given by (7). KT(s)=T(T+sI)−1,𝒦𝒯(s)=[KTt(s)]t∈J,G𝒯=∫s>0𝒦𝒯(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_{s>0}\mathcal K_{\mathcal T}(s)^* \mathcal K_{\mathcal T}(s)\,ds.(7) The finite common purification FT𝒯F_T^{\mathcal T} is the block column of G𝒯\sqrt{G_{\mathcal T}} indexed by a chosen occurrence of TT in the family. Alice’s and Bob’s common purification maps are Uj(r,z)=Ffj,r(z)𝒜U_j(r,z)=F_{f_{j,r}(z)}^{\mathcal A} and Vj(r,z)=Fhj,r(z)ℬV_j(r,z)=F_{h_{j,r}(z)}^{\mathcal B}. The resolvent of a square complex matrix TT at a real parameter ss is (T+sI)−1(T+sI)^{-1}.

For scalar weights wrw_r and matrices CrC_r indexed by a finite set RR, the weighted block state is the block-diagonal matrix whose block at rr is wrCr\sqrt{w_r}C_r.✦ The conditional block state 𝖡i[U,V](q)\mathsf B_i[U,V](q) associated with maps U,VU,V at coordinate ii and question pair qq is the matrix in (8), with the same record and profile register on both local spaces. 𝖡i[U,V](q)=diag⁡(r,z)[π(z)θr(z)pU(r,zi:q)SV(r,zi:q)𝖳].\mathsf B_i[U,V](q)= \operatorname{diag}_{(r,z)}\left[ \sqrt{\frac{\pi(z)\theta_r(z)}p}\, U(r,z^{i:q})S V(r,z^{i:q})^{\mathsf T}\right].(8)✦ At exposure position jj, the joint conditional state Dj(x,y)D_j(x,y) and the one-question states ΨjA(x),ΨjB(y)\Psi^A_j(x),\Psi^B_j(y) are the block matrices given by (9); ΨjA\Psi^A_j depends only on xx, and ΨjB\Psi^B_j only on yy. Dj(x,y)=𝖡ij[Uj+1,Vj](x,y),ΨjA(x)=𝖡ij[Uj+1,Vj+1](x,y0),ΨjB(y)=𝖡ij[Uj,Vj](x0,y).\begin{aligned} D_j(x,y)&=\mathsf B_{i_j}[U_{j+1},V_j](x,y),\\ \Psi^A_j(x)&=\mathsf B_{i_j}[U_{j+1},V_{j+1}](x,y_0),\\ \Psi^B_j(y)&=\mathsf B_{i_j}[U_j,V_j](x_0,y). \end{aligned}(9)

Refined Measurements

A local effect refinement is a pair of positive semidefinite families Li,ra(z)L_{i,r}^a(z) and Ni,rb(z)N_{i,r}^b(z), indexed by finite answer alphabets, summing to Fr(z)F_r(z) and Hr(z)H_r(z), respectively, with each family independent of the other player’s questions.✦ The averaged fine effects L¯j,ra,N¯j,rb\overline L_{j,r}^a,\overline N_{j,r}^b of a local effect refinement L,NL,N are the matrix families given by (10). L¯j,ra=ℒl≥j+1Lij,ra,N¯j,rb=ℛrev⁡(l<j)Nij,rb.\overline L_{j,r}^a =\mathcal L_{l_{\ge j+1}}L_{i_j,r}^a, \qquad \overline N_{j,r}^b =\mathcal R_{\operatorname{rev}(l_{<j})}N_{i_j,r}^b.(10) The coordinate refinement of the repeated experiment is a local effect refinement of its tested event system.✦

The support inverse T+T^+ of a square complex matrix TT is the operator defined by continuous functional calculus with the reciprocal function, taking its value at zero to be zero. The range projection ΠM\Pi_M of a map MM and the transported effects 𝒯M(G)a\mathcal T_M(G)^a of a family GaG^a are the operators given by (11); the transported POVM applies this construction to positive effects with ∑aGa=M*M\sum_aG^a=M^*M, assigning the support complement to a distinguished answer a0a_0. ΠM=M(M*M)+M*,𝒯M(G)a=M(M*M)+Ga(M(M*M)+)*+𝟏a=a0(I−ΠM).\Pi_M=M(M^*M)^+M^*,\qquad \mathcal T_M(G)^a=M(M^*M)^+G^a(M(M^*M)^+)^* +\mathbf1_{a=a_0}(I-\Pi_M).(11) Alice’s transported fine POVM Lˆj,r,z\widehat L_{j,r,z} is obtained by transporting her averaged fine effects along Uj+1(r,z)U_{j+1}(r,z), and Bob’s transported fine POVM Nˆj,r,z\widehat N_{j,r,z} by transporting his averaged fine effects along Vj(r,z)V_j(r,z), assigning the support complements to fixed answers a0,b0a_0,b_0. Alice’s conditional block POVM 𝖯j(x)\mathsf P_j(x) places Lˆj,r,zij:(x,y0)\widehat L_{j,r,z^{i_j:(x,y_0)}} in block (r,z)(r,z); Bob’s 𝖰j(y)\mathsf Q_j(y) uses Nˆj,r,zij:(x0,y)\widehat N_{j,r,z^{i_j:(x_0,y)}}.

For any local families L,NL,N, the conditional coordinate mass mi(q;L,N)m_i(q;L,N) is defined by (12); the event weight is evaluated at the original profile. mi(q;L,N)=1p∑r𝔼z∼πθr(z)χS(Lr(zi:q),Nr(zi:q)).m_i(q;L,N)=\frac1p\sum_r\mathbb E_{z\sim\pi} \theta_r(z)\chi_S(L_r(z^{i:q}),N_r(z^{i:q})).(12)✦ For a Boolean verifier VV, the conditional coordinate score si(V)s_i(V) is the weighted sum over accepted refined answers in (13). si(V)=1p∑r𝔼z∼πθr(z)∑a,b:V(xi,yi,a,b)=1χS(Li,ra(z),Ni,rb(z)).s_i(V)=\frac1p\sum_r\mathbb E_{z\sim\pi}\theta_r(z) \sum_{a,b:V(x_i,y_i,a,b)=1}\chi_S(L_{i,r}^a(z),N_{i,r}^b(z)).(13)✦

The Claim

Theorem 1. Let ll have distinct entries, let F,HF,H be local matrix families, and let the real weights θ\theta be invariant under both question replacements along ll. For every stage j≥0j\ge0, the mass identity is (14). ∑r𝔼πθr(z)χS(fj,r(z),hj,r(z))=∑r𝔼πθr(z)χS(Fr(z),Hr(z)).\sum_r\mathbb E_\pi\theta_r(z)\chi_S(f_{j,r}(z),h_{j,r}(z)) =\sum_r\mathbb E_\pi\theta_r(z)\chi_S(F_r(z),H_r(z)).(14)✦ Suppose additionally that F,HF,H are positive contractions, 0≤θ≤10\le\theta\le1, ∥S∥22≤1\|S\|_2^2\le1, and the right side of (14) is p>0p>0. For the common purification maps, the chain error bounds are (15). ∑j<ℓ∑r𝔼πθr(z)∥(Uj+1−Uj)SVj𝖳∥22≤plog⁡|R|p,∑j<ℓ∑r𝔼πθr(z)∥Uj+1S(Vj−Vj+1)𝖳∥22≤plog⁡|R|p,\begin{aligned} \sum_{j<\ell}\sum_r\mathbb E_\pi\theta_r(z) \|(U_{j+1}-U_j)S V_j^{\mathsf T}\|_2^2 &\le p\log\frac{|R|}{p},\\ \sum_{j<\ell}\sum_r\mathbb E_\pi\theta_r(z) \|U_{j+1}S(V_j-V_{j+1})^{\mathsf T}\|_2^2 &\le p\log\frac{|R|}{p}, \end{aligned}(15)✦ where all maps inside each expectation are evaluated at (r,z)(r,z).

For a conditioned quantum event system, the coherent errors satisfy (16). ∑j<ℓ𝔼(x,y)∼μ∥Dj(x,y)−ΨjA(x)∥22≤log⁡|R|p,∑j<ℓ𝔼(x,y)∼μ∥Dj(x,y)−ΨjB(y)∥22≤log⁡|R|p.\begin{aligned} \sum_{j<\ell}\mathbb E_{(x,y)\sim\mu} \|D_j(x,y)-\Psi^A_j(x)\|_2^2&\le\log\frac{|R|}{p},\\ \sum_{j<\ell}\mathbb E_{(x,y)\sim\mu} \|D_j(x,y)-\Psi^B_j(y)\|_2^2&\le\log\frac{|R|}{p}. \end{aligned}(16)✦ At every position j<ℓj<\ell, the joint norm identity is (17), Alice’s marginal norm identity is (18), and Bob’s is (19). ∥Dj(x,y)∥22=ρij(x,y).\|D_j(x,y)\|_2^2=\rho_{i_j}(x,y).(17)✦ ∥ΨjA(x)∥22=∑yμ(y∣x)ρij(x,y).\|\Psi^A_j(x)\|_2^2=\sum_y\mu(y\mid x)\rho_{i_j}(x,y).(18)✦ ∥ΨjB(y)∥22=∑xμ(x∣y)ρij(x,y).\|\Psi^B_j(y)\|_2^2=\sum_x\mu(x\mid y)\rho_{i_j}(x,y).(19)✦ For a local effect refinement and any distinguished answers a0,b0a_0,b_0, the transport identity is (20), for every r,z,a,br,z,a,b. χUj+1(r,z)SVj(r,z)𝖳(Lˆj,r,za,Nˆj,r,zb)=χS(L¯j,ra(z),N¯j,rb(z)).\chi_{U_{j+1}(r,z)S V_j(r,z)^{\mathsf T}} (\widehat L_{j,r,z}^a,\widehat N_{j,r,z}^b) =\chi_S(\overline L_{j,r}^a(z),\overline N_{j,r}^b(z)).(20)✦ The conditional block measurements satisfy (21), and their expected payoff satisfies (22) for every Boolean verifier VV. χDj(x,y)(𝖯j(x)a,𝖰j(y)b)=mij((x,y);Lija,Nijb).\chi_{D_j(x,y)}(\mathsf P_j(x)^a,\mathsf Q_j(y)^b) =m_{i_j}((x,y);L_{i_j}^{a},N_{i_j}^{b}).(21)✦ 𝔼(x,y)∼μ∑a,b:V(x,y,a,b)=1χDj(x,y)(𝖯j(x)a,𝖰j(y)b)=sij(V).\mathbb E_{(x,y)\sim\mu} \sum_{a,b:V(x,y,a,b)=1} \chi_{D_j(x,y)}(\mathsf P_j(x)^a,\mathsf Q_j(y)^b) =s_{i_j}(V).(22)✦ For the coordinate refinement of any repeated experiment and any tested set with pC>0p_C>0, the score at every coordinate i∈Ii\in I is given by (23). si(V)=pC∪{i}pC.s_i(V)=\frac{p_{C\cup\{i\}}}{p_C}.(23)✦

In the paper, the result reads: for every finite-dimensional repeated-game strategy, proper tested set CC with p=pC>0p=p_C>0, and fixed ordering of the k−|C|k-|C| untested coordinates, there are finite-dimensional vectors ψxyr,ψxr,ψyr\psi_{xy}^r,\psi_x^r,\psi_y^r and local POVMs such that, writing pxyr=∥ψxyr∥2p_{xy}^r=\|\psi_{xy}^r\|^2, pxr=∥ψxr∥2p_x^r=\|\psi_x^r\|^2, pyr=∥ψyr∥2p_y^r=\|\psi_y^r\|^2, and ϕxyr=ψxyr/pxyr\phi_{xy}^r=\psi_{xy}^r/\sqrt{p_{xy}^r} when pxyr>0p_{xy}^r>0, the exact conditional answer law is (24), where Gxα,a,Kyγ,bG_x^{\alpha,a},K_y^{\gamma,b} are the fine conditional answer effects, Pr⁡(a,b∣ϕxyr)=χS(Gxα,a,Kyγ,b)pxyr,pxyr>0,\Pr(a,b\mid\phi_{xy}^r) =\frac{\chi_S(G_x^{\alpha,a},K_y^{\gamma,b})}{p_{xy}^r}, \qquad p_{xy}^r>0,(24) its transcript law with weights wi(r)w_i(r) and its norm identities for the coarse conditional effects Fxα,HyγF_x^\alpha,H_y^\gamma are (25), λi(r,x,y)=wi(r)μ(x,y)∥ψxyr∥2/p,pxyr=χS(Fxα,Hyγ),pxr=∑yμ(y∣x)pxyr,pyr=∑xμ(x∣y)pxyr,\begin{aligned} \lambda_i(r,x,y)&=w_i(r)\mu(x,y)\|\psi_{xy}^r\|^2/p,\\ p_{xy}^r&=\chi_S(F_x^\alpha,H_y^\gamma),\\ p_x^r&=\sum_y\mu(y\mid x)p_{xy}^r, &p_y^r&=\sum_x\mu(x\mid y)p_{xy}^r, \end{aligned}(25) with each weighted total mass equal to pp, and, using its classical exposure law νi\nu_i and tested answer labels α,γ\alpha,\gamma in rr, the errors and their bounds are (26), with uniform averaging over untested coordinates and τ(C)\tau(C) as in (3), eA(i)=𝔼νiμ∑α,γθ∥ψxyr−ψyr∥2,eB(i)=𝔼νiμ∑α,γθ∥ψxyr−ψxr∥2,𝔼ieA(i)≤pτ(C),𝔼ieB(i)≤pτ(C),𝔼i(eA(i)+eB(i))≤2pτ(C),\begin{aligned} e_A(i)&=\mathbb E_{\nu_i\mu}\sum_{\alpha,\gamma} \theta\|\psi_{xy}^r-\psi_y^r\|^2, &e_B(i)&=\mathbb E_{\nu_i\mu}\sum_{\alpha,\gamma} \theta\|\psi_{xy}^r-\psi_x^r\|^2,\\ \mathbb E_i e_A(i)&\le p\tau(C), &\mathbb E_i e_B(i)&\le p\tau(C),\\ \mathbb E_i(e_A(i)+e_B(i))&\le2p\tau(C), \end{aligned}(26) allowing zero entries and disconnected support of the question law.

What Is Checked

The statements concern finite question and record sets, finite-dimensional coefficient matrices, local effect families and question laws with the specified default conditionals; the coherent statements use conditioned event systems and their local effect refinements.

The averaged fine effects in (10) are positive semidefinite. For any transport map and any positive semidefinite input effects, the transported effects are positive semidefinite.✦ If the input effects sum to the Gram matrix of the transport map, the transported effects sum to the identity.✦ The tested wins indicator is nonnegative and at most one. The sum over tested answer records of this indicator times the Born probability of the corresponding tested effects is at most one for every question profile✦, and its expectation under the product question law equals the joint winning probability pCp_C, as in (27).✦ ∑r∈RCθC(r,z)χS(Fr(z),Hr(z))≤1,∑r∈RC𝔼πθC(r,z)χS(Fr(z),Hr(z))=pC.\sum_{r\in R_C}\theta_C(r,z)\chi_S(F_r(z),H_r(z))\le1, \qquad \sum_{r\in R_C}\mathbb E_\pi \theta_C(r,z)\chi_S(F_r(z),H_r(z))=p_C.(27)