Foundations of quantum mechanics · Theory and experiment (AI-assisted)

Quantum violation of Causal Friendliness in the instrumental scenario

Emanuele Polino, Giulia Lelli, Aaditya Datar, Marwan Haddara, Davide Poderini, Giorgio Milani, Sergei Slussarenko, Howard M. Wiseman, Nora Tischler, Eric G. Cavalcanti, Fabio Sciarrino

arXiv:2611.17431v1 ↗PDF ↗Submitted 12 November 2026

Question
The Local Friendliness theorem showed that quantum predictions in Wigner's-friend versions of a Bell test conflict with any theory in which observed events are absolute, even when causes are allowed to be nonclassical. But its key assumption is a relativistic one, that a distant choice cannot influence a local outcome, and it says nothing about networks in which parties communicate. Does the same strong nonclassicality appear in a causal network where one party's outcome is sent to the other, so that spacelike separation plays no role?
Result
In the instrumental network (Alice's outcome sets Bob's measurement), the authors combine absoluteness of observed events with the constraints that generalized, nonclassical causal models impose on this network, and call the conjunction Causal Friendliness. Its correlation polytope has 36 vertices and 48 facets, 8 of them new. The genuine inequality \(I_{\mathrm{inst}}\le2\) is violated by a singlet with three settings, reaching \(1+\sqrt2\), which is optimal by the NPA hierarchy. A photonic experiment with feed-forward measured \(I_{\mathrm{inst}}=2.266\pm0.002\), more than 120 standard deviations above the bound, and a noisier run produced correlations that defeat classical causal models but remain inside the Causal Friendliness polytope.

The paper in detail

Bell's theorem is now often read as a conflict between quantum correlations and four assumptions: observed events are absolute, choices are free, causes cannot act faster than light, and causal explanations are classical. Dropping the last one and allowing quantum or more general causes rescues the Bell correlations. Local Friendliness put a Wigner's friend inside one laboratory and showed that even nonclassical causes fail if every observed event, including the friend's, has a single absolute value. Its decisive assumption, however, is relativistic: a distant setting cannot affect a local outcome.

The instrumental network asks what happens without that crutch. Alice's setting \(X\) controls her measurement, her outcome \(A\) is sent to Bob, and Bob's outcome \(B\) may depend on it; the two share a source. Here Alice is a superobserver. Inside her sealed laboratory a friend, Charlie, measures and records \(C\). For \(X=0\) Alice simply reads Charlie's record, so \(A=C\); for \(X\neq0\) she measures the whole laboratory in another basis.

Causal Friendliness is four constraints on a joint distribution \(P(a,c,b\mid x)\) whose marginal must reproduce the data: Charlie's outcome exists (absoluteness), it is read faithfully when \(X=0\), it is independent of \(X\), and, because there is no arrow from \(X\) to \(B\) except through \(A\), the generalized Pearl inequalities \(P(a,b{=}{-1}\mid x,c)+P(a,b{=}{+1}\mid x',c)\le1\) hold for each \(c\). These are strictly weaker than classical causal models, which lie inside them, and strictly stronger than the general operational theories, which reach the algebraic value 3.

With binary outcomes and three settings, the resulting polytope has eight facets that belong to neither of the other two, all equivalent to \(\langle AB\rangle_1-\langle AB\rangle_0+\langle B\rangle_0-\langle B\rangle_2\le2\). A singlet with suitable Pauli measurements gives \(1+\sqrt2\). In the experiment, polarization-entangled photons are produced in a Sagnac source; the friend's measurement is a beam displacer that writes polarization into path, undone by a second displacer when \(X\neq0\), and a fast optical switch routes Bob's delayed photon according to Alice's detector click. The measured 2.266 corresponds to a visibility of about 0.94, against the 0.83 needed for violation.

The paper is careful about scope. The theorem is theory-independent and does not depend on the friend story, but the story is what motivates absoluteness, and a convincing test would need an observer more substantial than a photon's path, under coherent control. What stays open is whether Causal-Friendliness violations exist in other networks with communication and whether they certify randomness, as Local Friendliness violations do.

Transferable methods

  1. A partially deterministic causal polytope

    To compare quantum correlations with a theory class that allows nonclassical causes, the paper needs that class as a finite set of linear constraints. Absoluteness supplies a hidden variable \(C\), the reading rule makes the outcome deterministic for one setting, and the network's missing arrows give linear inequalities conditioned on \(C\). The result is a polytope whose facets can be enumerated.

    The variables are \(P(a,c,b\mid x)\) for \(a,b,c\in\{\pm1\}\), \(x\in\{0,1,2\}\). Constraints: \[\sum_cP(a,c,b\mid x)=\wp(a,b\mid x),\qquad P(a\mid c,x{=}0)=\delta_{a,c},\qquad P(c\mid x)=P(c),\] and, for \(x\neq x'\), \[P(a,c,b{=}{-1}\mid x)+P(a,c,b{=}{+1}\mid x')\le P(c).\] The last line is the Pearl inequality conditioned on \(c\), multiplied by \(P(c)\) so that it stays linear (editorial rewriting).

    Why the quantum value is \(1+\sqrt2\) (editorial derivation): with the singlet, \(\langle A_xB_y\rangle=-\vec a_x\cdot\vec b_y\), and Bob's setting is \(y=a\). For \(\hat A_0=-\sigma_z\), \(\hat B_{+}=\sigma_x\), \(\hat B_{-}=-\sigma_z\), one finds \(\langle AB\rangle_0=-\tfrac12\), \(\langle B\rangle_0=\tfrac12\), \(\langle AB\rangle_1=\tfrac{1}{\sqrt2}\), \(\langle B\rangle_2=-\tfrac{1}{\sqrt2}\), and \[I_{\mathrm{inst}}=\tfrac{1}{\sqrt2}+\tfrac12+\tfrac12+\tfrac{1}{\sqrt2}=1+\sqrt2.\]

    Editorial checks: maximizing \(I_{\mathrm{inst}}\) by linear programming over the constraints above gives exactly 2, and over the unconditioned Pearl inequalities alone (the operational-theory set) gives 3. The observed 2.266 is \(0.94\times(1+\sqrt2)\); white noise of visibility \(v\) violates the bound when \(v>2/(1+\sqrt2)\approx0.83\).

    How to reuse it

    Any no-go argument with a hidden record that is read out for one setting can be turned into a linear program: introduce the record as a variable, impose determinism where it is read, and condition each causal constraint on it. Facet enumeration then shows which inequalities are genuinely new compared with classical and fully general models.

  2. Bounding a partly nonclassical report

    A classical error in Charlie's report changes nothing, because the noise can be absorbed into a new deterministic record. The real worry is a nonclassical side channel from the source to \(A\) when \(X=0\). The paper relaxes the reading rule to \(P(a{=}c\mid x{=}0,c)\ge1-\epsilon\) and finds how the bound grows.

    A stochastic report \(T(a\mid c)\) can always be written as \(a=f(c,r)\) with an extra random variable \(R\); the record \(C'=f(C,R)\) is again read deterministically, so classical noise leaves the polytope unchanged. This is the paper's footnote argument.

    With the relaxed condition, the authors construct an explicit family of behaviours (Eq. (63)) reaching \(I_{\mathrm{inst}}=2+2\epsilon\) for \(\epsilon\le\tfrac12\) and argue that it is the maximum. Editorial check: a linear program with the relaxed reading rule and all other constraints unchanged returns \(2.1\), \(2.262\) and \(2.5\) at \(\epsilon=0.05\), \(0.131\) and \(0.25\), exactly \(2+2\epsilon\). The measured value minus two standard deviations, 2.262, therefore excludes \(\epsilon\) up to 13.1%, the figure the paper reports.

    How to reuse it

    When an assumption in a no-go theorem is idealized, parametrize its failure by the probability that the ideal rule is broken, recompute the bound as a function of that parameter, and quote the experiment as excluding a range of it. A linear program makes this mechanical.

Source: Sec. II.2 (Eqs. (1)–(4)); Sec. III (Eqs. (5)–(6)); Sec. IV and Appendix H. Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. It takes the Local-Friendliness type of no-go theorem beyond Bell scenarios, to a network that cannot be made spacelike-separated, and shows that it needs no more settings than the classical instrumental test. Theories that keep absolute events must then reject the network's causal structure, for example through a direct influence or a correlation between Alice's choice and the friend's outcome.

Limitations. The friend is the path of a single photon, so the experiment is a proof of principle about the assumptions, not a test with an observer anyone would call a friend; the detection loophole is open, and the counting relies on fair sampling. Allowing the friend's report to carry a fraction \(\epsilon\) of nonclassical influence raises the bound to \(2+2\epsilon\), so the data exclude only \(\epsilon\) up to about 13%. The acknowledgements state that Claude (Sonnet 5) helped with algebra and suggested the solution in Eq. (63), the explicit model for this relaxed bound, which the author verified.

Description based on the full arXiv HTML of v1, read 2026-11-16: Sections I–V, the acknowledgements with the AI statement, and the Eq. (63) passage of Appendix H; the remaining appendices (the OPT Markov condition, the polytope enumeration and experimental details) were not read. Editorial linear programs reproduced the Causal Friendliness bound 2, the OPT bound 3 and the relaxed bound \(2+2\epsilon\) (2.262 at \(\epsilon=0.131\)), and the quantum value \(1+\sqrt2\) was recomputed from the stated state and measurements.