OctoInformBeta

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

Section 4 (continuation in Lean): Quantum correlated sampling controls the conditional excess through the event probability and the number of records

Conditioning bounds the remaining success.

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 on successful tests selects a quantum event as well as a question profile. The issue is how large the conditional success on an untested coordinate can be compared with the entangled value of the base game.

A two-player nonlocal game GG has finite nonempty question alphabets X,YX,Y, finite nonempty answer alphabets A,BA,B, a question distribution qq on X×YX\times Y, and an acceptance predicate V(x,y,a,b)∈{0,1}V(x,y,a,b)\in\{0,1\}.✦ A finite probability law μ\mu on a finite sample space is a function assigning nonnegative weights to its points, with the weights summing to one.✦ For a finite probability law μ\mu and a function ff on its sample space, the expectation of ff under μ\mu and the product law of μ\mu on profiles z:I→X×Yz:I\to X\times Y, for a finite coordinate set II, are 𝔼μ[f]=∑uμ(u)f(u),μ⊗I(z)=∏i∈Iμ(zi).\mathbb E_\mu[f]=\sum_u\mu(u)f(u), \qquad \mu^{\otimes I}(z)=\prod_{i\in I}\mu(z_i).(1) The game question distribution is the law with weights q(x,y)q(x,y); the independent question profiles below have the product law in (1).

For a rectangular coefficient matrix Γ\Gamma on finite local bases m,nm,n, its state vector, squared Hilbert-Schmidt norm and bipartite operator expectation are those in (2). The Born probability Pr⁡Γ(E,F)\Pr_\Gamma(E,F) is the real part of the expectation of local operators E,FE,F: |Γ⟩=∑u∈m,w∈nΓuw|u⟩|w⟩,∥Γ∥HS2=∑u,w|Γuw|2,ev⁡Γ(E,F)=⟨Γ|E⊗F|Γ⟩,PrΓ(E,F)=Re⁡ev⁡Γ(E,F).\begin{aligned} |\Gamma\rangle&=\sum_{u\in m,\,w\in n}\Gamma_{uw}|u\rangle|w\rangle, &\|\Gamma\|_{\mathrm{HS}}^2&=\sum_{u,w}|\Gamma_{uw}|^2,\\ \operatorname{ev}_\Gamma(E,F) &=\langle\Gamma|E\otimes F|\Gamma\rangle, &\Pr_\Gamma(E,F)&=\operatorname{Re}\operatorname{ev}_\Gamma(E,F). \end{aligned}(2) A positive operator-valued measure (POVM) P=(Pa)a∈AP=(P^a)_{a\in A} is a family of positive semidefinite operators summing to the identity. For local POVMs P=(Pa)a∈AP=(P^a)_{a\in A} and Q=(Qb)b∈BQ=(Q^b)_{b\in B} on the two spaces of Γ\Gamma, the measurement payoff Π(Γ,P,Q,U)\Pi(\Gamma,P,Q,U) for an acceptance predicate UU on A×BA\times B is Π(Γ,P,Q,U)=∑a,bU(a,b)PrΓ(Pa,Qb).\Pi(\Gamma,P,Q,U)=\sum_{a,b}U(a,b)\Pr_\Gamma(P^a,Q^b).(3)✦

An entangled strategy 𝒮\mathcal S consists of a unit pure state ψ\psi on finite nonzero local Hilbert spaces and local POVMs Px,QyP_x,Q_y for the respective questions.✦ A single-round strategy uses the question and answer alphabets of GG.✦ The Born probabilities and winning probability of a single-round entangled strategy 𝒮\mathcal S for GG, and the finite-dimensional entangled value vv of GG, are defined by (4): p𝒮(a,b∣x,y)=Re⁡⟨ψ|Pxa⊗Qyb|ψ⟩,Succ⁡G(𝒮)=∑x,yq(x,y)∑a,bV(x,y,a,b)p𝒮(a,b∣x,y),v=sup𝒮Succ⁡G(𝒮).\begin{aligned} p_{\mathcal S}(a,b\mid x,y)&= \operatorname{Re}\langle\psi|P_x^a\otimes Q_y^b|\psi\rangle,\\ \operatorname{Succ}_G(\mathcal S)&= \sum_{x,y}q(x,y)\sum_{a,b}V(x,y,a,b)p_{\mathcal S}(a,b\mid x,y), &v&=\sup_{\mathcal S}\operatorname{Succ}_G(\mathcal S). \end{aligned}(4)

Local Events

Left coordinate replacement at ii changes Alice’s question there and retains all other questions.✦ A function ff of question profiles has left coordinate invariance at coordinate ii when replacing Alice’s question at ii while retaining all other questions leaves ff unchanged.✦ Right coordinate replacement at coordinate ii changes Bob’s question there and retains all other questions; a function ff of question profiles has right coordinate invariance at ii when every right coordinate replacement at ii leaves ff unchanged.

A conditioned quantum event system ℰ\mathcal E has question law qq, a unit coefficient matrix Γ\Gamma, a finite record set RR, and a list l=(i1,…,iℓ)l=(i_1,\ldots,i_\ell) of distinct coordinates. For each record rr and profile zz, Alice’s local effect Fr(z)F_r(z) and Bob’s local effect Hr(z)H_r(z) satisfy 0≤Fr(z)≤I0\le F_r(z)\le I and 0≤Hr(z)≤I0\le H_r(z)\le I. Alice’s effects have right coordinate invariance at every coordinate, and Bob’s have left coordinate invariance. Its weights θr(z)∈[0,1]\theta_r(z)\in[0,1] have both invariances at every coordinate in ll. The event weights and their strictly positive total probability pp satisfy (5): ∑r∈Rθr(z)PrΓ(Fr(z),Hr(z))≤1,p=∑r∈R𝔼q⊗I[θr(z)PrΓ(Fr(z),Hr(z))]>0.\sum_{r\in R}\theta_r(z)\Pr_\Gamma(F_r(z),H_r(z))\le1, \qquad p=\sum_{r\in R}\mathbb E_{q^{\otimes I}} [\theta_r(z)\Pr_\Gamma(F_r(z),H_r(z))]>0.(5)

A local effect refinement 𝒯\mathcal T of ℰ\mathcal E consists of positive semidefinite families Li,r,za,Ki,r,zbL_{i,r,z}^a,K_{i,r,z}^b satisfying ∑aLi,r,za=Fr(z),∑bKi,r,zb=Hr(z).\sum_a L_{i,r,z}^a=F_r(z),\qquad \sum_b K_{i,r,z}^b=H_r(z).(6) The refined effects in (6) retain the respective right and left coordinate invariances at every coordinate. The conditional coordinate score sis_i of a local effect refinement 𝒯\mathcal T of a conditioned quantum event system ℰ\mathcal E is the refinement’s accepted mass at coordinate ii divided by the system’s event probability pp: si=1p∑r∈R𝔼q⊗I[θr(z)∑a,bV((zi)X,(zi)Y,a,b)PrΓ(Li,r,za,Ki,r,zb)].s_i=\frac1p\sum_{r\in R}\mathbb E_{q^{\otimes I}} \left[\theta_r(z)\sum_{a,b}V((z_i)_X,(z_i)_Y,a,b) \Pr_\Gamma(L_{i,r,z}^a,K_{i,r,z}^b)\right].(7) For a conditioned quantum event system ℰ\mathcal E with ℓ>0\ell>0, the entropy parameter τℰ\tau_{\mathcal E} is the real number τℰ=log⁡(|R|/p)ℓ.\tau_{\mathcal E}=\frac{\log(|R|/p)}{\ell}.(8) All logarithms are natural.

For rectangular matrix families (Mx)x∈X(M_x)_{x\in X} and (Ny)y∈Y(N_y)_{y\in Y} on the same local bases, their mean squared discrepancy DD is D=∑x,yq(x,y)∥Mx−Ny∥HS2.D=\sum_{x,y}q(x,y)\|M_x-N_y\|_{\mathrm{HS}}^2.(9)

The Claim

Theorem 1. Let 0<ε≤10<\varepsilon\le1. Let Mx,NyM_x,N_y be rectangular matrices on finite nonempty local bases, each of unit squared Hilbert-Schmidt norm, and let Px,QyP_x,Q_y be local POVMs. With the payoff (3) and discrepancy (9), quantum correlated sampling gives ∑x,yq(x,y)Π(Mx,Px,Qy,V(x,y,⋅,⋅))≤v+42ε2+rac12ε2D.\sum_{x,y}q(x,y)\Pi(M_x,P_x,Q_y,V(x,y,\cdot,\cdot)) \le v+4\sqrt{2\varepsilon^2+ rac{12}{\varepsilon}\sqrt{2D}}.(10)✦ Let ℰ\mathcal E be a conditioned quantum event system with ℓ>0\ell>0 and let 𝒯\mathcal T be a local effect refinement. For its entropy parameter (8), the conditional rounding bound is 1ℓ∑j=1ℓsij≤v+42ε2+rac12ε32τℰ+24τℰ+4τℰ.\frac1\ell\sum_{j=1}^{\ell}s_{i_j} \le v+4\sqrt{2\varepsilon^2+ rac{12}{\varepsilon} \sqrt{32\tau_{\mathcal E}}} +2\sqrt{4\tau_{\mathcal E}}+4\tau_{\mathcal E}.(11)✦ If 0<δ≤10<\delta\le1 and τℰ≤δ8/280\tau_{\mathcal E}\le\delta^8/2^{80}, the small-entropy rounding bound is 1ℓ∑j=1ℓsij≤v+δ4.\frac1\ell\sum_{j=1}^{\ell}s_{i_j}\le v+\frac\delta4.(12)✦

What Is Checked

The correlated sampling bound (10) is stated over finite nonempty local bases of possibly unequal sizes; the conditioning bounds (11) and (12) concern finite event systems with positive event probability and local effect refinements on a nonempty list of distinct coordinates.