OctoInformBeta

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

Section 4 (setting): For a tested set CC and dd joint answers per coordinate, mixed exposure chains and coordinate refinements use d|C|d^{|C|} records

Local effects for tested answer 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

Winning every coordinate in a tested set conditions the questions and answers of a repeated entangled game. Local effects indexed by the tested answers describe this event, while conditional averaging describes the exposure of the remaining questions.

Fix finite question alphabets X,YX,Y, answer alphabets A,BA,B, a coordinate set II with |I|=k|I|=k, and finite-dimensional local Hilbert spaces. A repeated quantum experiment consists of a question law μ\mu on X×YX\times Y, a normalized bipartite state, local positive operator-valued measures (POVMs) PxP_x and QyQ_y with outcomes in AIA^I and BIB^I for each x∈XIx\in X^I and y∈YIy\in Y^I, and a verifier V:X×Y×A×B→{0,1}V:X\times Y\times A\times B\to\{0,1\}.✦ A local POVM is a finite family of positive operators whose sum is the identity.✦

In fixed orthonormal bases {|u⟩}\{|u\rangle\} and {|v⟩}\{|v\rangle\} of the local Hilbert spaces, a coefficient matrix Ψ\Psi specifies the state vector in (1); its squared Hilbert-Schmidt norm is the sum of the squared moduli of its entries. For local operators S,TS,T, write eψ(S,T)e_\psi(S,T) for their bipartite operator expectation and bψ(S,T)b_\psi(S,T) for its real part, the Born probability convention: |ψ⟩=∑u,vΨuv|u⟩⊗|v⟩,∥Ψ∥HS2=∑u,v|Ψuv|2=1,eψ(S,T)=⟨ψ|S⊗T|ψ⟩,bψ(S,T)=Re⁡eψ(S,T).\begin{aligned} |\psi\rangle&=\sum_{u,v}\Psi_{uv}|u\rangle\otimes|v\rangle, &\|\Psi\|_{\mathrm{HS}}^2&=\sum_{u,v}|\Psi_{uv}|^2=1,\\ e_\psi(S,T)&=\langle\psi|S\otimes T|\psi\rangle, &b_\psi(S,T)&=\operatorname{Re}e_\psi(S,T). \end{aligned}(1)

A finite probability law pp on a finite set Ω\Omega is a function assigning nonnegative weights of total one to the elements of Ω\Omega.✦ For a function ff on Ω\Omega, its expectation under pp and the product law of coordinate laws pip_i are given by (2). A question profile is a function z:I→X×Yz:I\to X\times Y, written zi=(xi,yi)z_i=(x_i,y_i); the independent question law is ν\nu. 𝔼p[f]=∑ω∈Ωp(ω)f(ω),(⨂i∈Ipi)(z)=∏i∈Ipi(zi),ν=μ⊗I.\mathbb E_p[f]=\sum_{\omega\in\Omega}p(\omega)f(\omega), \qquad \Bigl(\bigotimes_{i\in I}p_i\Bigr)(z)=\prod_{i\in I}p_i(z_i), \qquad \nu=\mu^{\otimes I}.(2)

A question and answer transcript t=(z,a,b)t=(z,a,b) is a triple consisting of a question profile zz and answer vectors a∈AIa\in A^I and b∈BIb\in B^I.✦ The conditional answer distribution αz\alpha_z is the joint measurement distribution of Px,QyP_x,Q_y in the state |ψ⟩|\psi\rangle.✦ The joint transcript law JJ includes the independent questions: αz(a,b)=bψ(Px(a),Qy(b)),J(z,a,b)=ν(z)αz(a,b).\alpha_z(a,b)=b_\psi\bigl(P_x(a),Q_y(b)\bigr), \qquad J(z,a,b)=\nu(z)\alpha_z(a,b).(3) The coordinate winning predicate is wi(t)=V(xi,yi,ai,bi)w_i(t)=V(x_i,y_i,a_i,b_i).✦ For a tested set C⊆IC\subseteq I, the event ECE_C, its winning indicator and its probability pCp_C are given by (4). EC={t:wi(t)=1 for every i∈C},𝟏EC(t)={1,t∈EC,0,t∉EC,pC=𝔼J[𝟏EC].E_C=\{t:w_i(t)=1\text{ for every }i\in C\}, \qquad \mathbf 1_{E_C}(t)= \begin{cases}1,&t\in E_C,\\0,&t\notin E_C,\end{cases} \qquad p_C=\mathbb E_J[\mathbf 1_{E_C}].(4)

Conditional Averaging

Fix default questions x0∈Xx_0\in X and y0∈Yy_0\in Y. Normalizing nonnegative weights means dividing by their total when it is positive and using a point mass at a specified default point when the total is zero.✦ The point mass δs\delta_s at s∈Ωs\in\Omega is the probability law on Ω\Omega assigning weight one to ss and zero elsewhere.✦ The conditional laws μ(⋅∣y)\mu(\cdot\mid y) on XX and μ(⋅∣x)\mu(\cdot\mid x) on YY normalize the corresponding slices of μ\mu, with defaults x0x_0 and y0y_0, respectively.

Write zL,i←x′z^{L,i\leftarrow x'} for the profile obtained by replacing only Alice’s question at ii by x′x', and zR,i←y′z^{R,i\leftarrow y'} for Bob’s corresponding replacement. A profile function ff is left coordinate invariant at ii if Alice’s question replacement at ii leaves ff unchanged for every profile and replacement question; right coordinate invariance has the analogous meaning for Bob. Left and right conditional averaging of a vector-valued profile function ff are the operators in (5). (𝖫if)(z)=∑x′∈Xμ(x′∣yi)f(zL,i←x′),(𝖱if)(z)=∑y′∈Yμ(y′∣xi)f(zR,i←y′).\begin{aligned} (\mathsf L_i f)(z)&=\sum_{x'\in X}\mu(x'\mid y_i) f(z^{L,i\leftarrow x'}),\\ (\mathsf R_i f)(z)&=\sum_{y'\in Y}\mu(y'\mid x_i) f(z^{R,i\leftarrow y'}). \end{aligned}(5) For a list s=(i1,…,it)s=(i_1,\ldots,i_t), the iterated averages in (6) compose these operators in the listed order. For a coordinate list ℓ\ell and a nonnegative integer jj, let ℓ≥j\ell_{\geq j} denote the suffix obtained by removing the first jj entries of ℓ\ell, and ℓ<j\ell_{<j} its prefix of at most jj entries. 𝖫s=𝖫i1∘⋯∘𝖫it,𝖱s=𝖱i1∘⋯∘𝖱it,𝖫⌀=𝖱⌀=id.\mathsf L_s=\mathsf L_{i_1}\circ\cdots\circ\mathsf L_{i_t}, \qquad \mathsf R_s=\mathsf R_{i_1}\circ\cdots\circ\mathsf R_{i_t}, \qquad \mathsf L_{\varnothing}=\mathsf R_{\varnothing}=\mathrm{id}.(6)

Tested Answer Effects

A tested answer record is a pair r=(rA,rB)r=(r_A,r_B) in RC=AC×BCR_C=A^C\times B^C.✦ Restriction sends full answer vectors to a|Ca|_C and b|Cb|_C. Coarsening a POVM along a map assigns to each new outcome the sum of the effects in its inverse image.✦ Alice’s and Bob’s tested answer POVMs are the coarsenings in (7). PzC(rA)=∑a:a|C=rAPx(a),QzC(rB)=∑b:b|C=rBQy(b).P_z^C(r_A)=\sum_{a:\,a|_C=r_A}P_x(a), \qquad Q_z^C(r_B)=\sum_{b:\,b|_C=r_B}Q_y(b).(7) The coordinate refinement POVMs PzC,iP_z^{C,i} and QzC,iQ_z^{C,i} coarsen the full measurements to (a|C,ai)(a|_C,a_i) and (b|C,bi)(b|_C,b_i), respectively. The tested wins indicator θC\theta_C depends on a record and a question profile as in (8). θC(r,z)={1,V(xi,yi,rA(i),rB(i))=1 for every i∈C,0,otherwise.\theta_C(r,z)= \begin{cases} 1,&V(x_i,y_i,r_A(i),r_B(i))=1\text{ for every }i\in C,\\ 0,&\text{otherwise}. \end{cases}(8)✦ Alice’s restricted answers a|C∈ACa|_C\in A^C are the map on CC that takes each tested coordinate to its answer in the full vector a∈AIa\in A^I. Bob’s restricted answers b|C∈BCb|_C\in B^C are the map on CC that takes each tested coordinate to its answer in the full vector b∈BIb\in B^I.

A local effect FF on a local Hilbert space is a positive operator bounded above by the identity. A local effect refinement Ui,r(a,z),Wi,r(b,z)U_{i,r}(a,z),W_{i,r}(b,z) of families Fr(z),Hr(z)F_r(z),H_r(z) is a pair of families of positive operators, depending only on xx and yy, respectively, whose totals satisfy (9).✦ ∑a∈AUi,r(a,z)=Fr(z),∑b∈BWi,r(b,z)=Hr(z).\sum_{a\in A}U_{i,r}(a,z)=F_r(z), \qquad \sum_{b\in B}W_{i,r}(b,z)=H_r(z).(9)

The Claim

Definition 1 (Local event systems and mixed exposure). Fix a question law μ\mu, default questions x0,y0x_0,y_0, a list ℓ\ell of distinct coordinates, and a finite record set RR. A local quantum event system for the question law μ\mu, default questions x0,y0x_0,y_0, coordinate list ℓ\ell and record set RR consists of a normalized bipartite state, local effects Fr(x),Hr(y)F_r(x),H_r(y) for r∈Rr\in R, and weights θr(z)\theta_r(z) in [0,1][0,1] that are invariant under either player’s question replacement at every coordinate of ℓ\ell, with positive event probability pp and pointwise event weights satisfying (10). q(z):=∑r∈Rθr(z)bψ(Fr(x),Hr(y))≤1,p=𝔼ν[q]>0.\begin{aligned} q(z)&:=\sum_{r\in R}\theta_r(z) b_\psi\bigl(F_r(x),H_r(y)\bigr)\leq1,\\ p&=\mathbb E_\nu[q]>0. \end{aligned}(10)✦ For vector-valued profile families fr,hrf_r,h_r and j≥0j\geq0, the left exposure chain is given by (11) and the right exposure chain by (12), where the superscript rev\mathrm{rev} denotes reversal of a coordinate list. ℒj[f]r=𝖫ℓ≥jfr.\mathcal L_j[f]_r=\mathsf L_{\ell_{\geq j}}f_r.(11)✦ ℛj[h]r=𝖱(ℓ<j)revhr.\mathcal R_j[h]_r=\mathsf R_{(\ell_{<j})^{\mathrm{rev}}}h_r.(12)✦

For a repeated quantum experiment and a tested set CC with pC>0p_C>0, choose an ordering ℓC\ell_C of I\CI\setminus C and default questions. The tested set event system uses the original question law and state, with the data in (13). R=RC,ℓ=ℓC,p=pC,Fr(z)=PzC(rA),Hr(z)=QzC(rB),θr(z)=θC(r,z).\begin{gathered} R=R_C,\qquad \ell=\ell_C,\qquad p=p_C,\\ F_r(z)=P_z^C(r_A),\qquad H_r(z)=Q_z^C(r_B), \qquad \theta_r(z)=\theta_C(r,z). \end{gathered}(13)✦ Its coordinate answer refinement is (14), for every i∈Ii\in I. Ui,r(a,z)=PzC,i(rA,a),Wi,r(b,z)=QzC,i(rB,b).U_{i,r}(a,z)=P_z^{C,i}(r_A,a), \qquad W_{i,r}(b,z)=Q_z^{C,i}(r_B,b).(14)✦ For every tested set CC, writing d=|A||B|d=|A||B|, its record alphabet has cardinality (15). |RC|=d|C|.|R_C|=d^{|C|}.(15)✦

What Is Checked

The general definitions are stated over finite question profiles and record sets, with finite-dimensional local operators and fixed default questions. The tested construction uses the repeated experiment’s local POVMs and assumes pC>0p_C>0. For its tested mask and coarse event probability, the library gives (16). 0≤θC(r,z)≤1,qC(z):=∑r∈RCθC(r,z)bψ(PzC(rA),QzC(rB))≤1,𝔼ν[qC]=pC.\begin{gathered} 0\leq\theta_C(r,z)\leq1,\\ q_C(z):=\sum_{r\in R_C}\theta_C(r,z) b_\psi\bigl(P_z^C(r_A),Q_z^C(r_B)\bigr)\leq1, \qquad \mathbb E_\nu[q_C]=p_C. \end{gathered}(16)