Foundations of quantum mechanics · Theory and numerics (AI-assisted)
Randomized routing: ruling out classical networks without trusting source independence
arXiv:2611.20866v1 ↗PDF ↗Submitted 16 November 2026
- Question
- In network nonlocality without inputs, three laboratories joined pairwise by sources each perform one fixed measurement, and some outcome distributions cannot come from three independent classical sources. The argument rests on the sources being independent, which no data can test: a single hidden classical source feeding all three parties reproduces any distribution. Can every classical explanation, including correlated sources, be excluded using only assumptions that an experiment can enforce?
- Result
- Yes, by randomized routing. A spare copy of one source is built, and the two parties it connects choose at random in each round which copy to read. Rounds in which both read the same copy reproduce the triangle data; rounds in which they read different copies give new data. A single classical source must commit in advance to an answer for every routing choice, so whether one exists is a linear program, and its infeasibility certificate is a linear inequality on the data. Every classical network, with independent or correlated sources, is such a source. For the RGB4 family the routed data admit no classical explanation for \(u^2\) in about \((0.08,0.92)\), except at \(u^2=\tfrac12\), and the verdict survives 3% white noise at every source. The routed data of the Fritz distribution and of the elegant joint measurement still admit a single classical source.
The paper in detail
In the triangle network, source \(\alpha\) connects Bob and Charlie, \(\beta\) connects Charlie and Alice, and \(\gamma\) connects Alice and Bob. Each party measures the two systems it receives in a fixed basis, so there are no settings, yet some distributions \(p(abc)\) cannot be produced by three independent classical sources wired the same way. This network nonlocality without inputs has been used to certify randomness. Its weak point is independence: if one classical source samples \((a,b,c)\) from \(p\) and hands each party its outcome, it fakes any network, and without inputs no data distinguish the two. Earlier responses bounded the allowed dependence between sources, but those bounds are themselves unobservable.
Randomized routing adds hardware rather than an assumption. Source \(\alpha\) is duplicated into \(\alpha_1\) and \(\alpha_2\); in each round Bob picks \(j\in\{1,2\}\) and Charlie picks \(k\in\{1,2\}\), and each feeds the system from the chosen copy into the same measurement as before, discarding the other. Closed rounds (\(j=k\)) reproduce the triangle data; in open rounds (\(j\ne k\)) Bob and Charlie hold halves of different entangled pairs and see new statistics. A hidden source that does not know which copies will be read must decide in advance what each party would answer for each choice. The data are then marginals of a single distribution over these counterfactual answers, and whether such a distribution exists is a linear program, the routed analogue of Fine's theorem.
Three examples show the range of outcomes. For the Fritz distribution, in which the triangle embeds a CHSH test, the open data make Bob and Charlie independent given Alice's outcome, and an explicit model passes the test. For the elegant joint measurement, whose triangle data were recently shown to need nonclassical sources, an exact model also passes at this level. The RGB4 family, with every source emitting \(\ket{\psi^+}\) and every party measuring in a partly entangled basis with parameter \(u\), fails for \(u^2\) in about \((0.08,0.92)\) apart from \(u^2=\tfrac12\). That includes \(u^2\approx0.785\), where the triangle data alone have a numerical model with independent sources; conversely, for \(u^2\in(0.92,1)\) the triangle data are network nonlocal but the routed data pass. With white noise of visibility \(v\) at all four sources, the test still fails somewhere at \(v=0.97\), most robustly near \(u^2\approx0.69\), and nowhere at \(v=0.965\).
More copies give more constraints, and infeasibility at one level persists at all higher ones. For a single duplicated source the whole hierarchy has a clean limit: the data pass at every level exactly when, given Alice's outcome and a shared variable, each copy of \(\alpha\) acts as an independent classical source of the pair \((b,c)\). The open questions are which quantum explanations routing can rule out, whether higher levels expose the elegant joint measurement, how finite statistics and memory across rounds change the analysis, and whether protocols built on network nonlocality, such as randomness certification, can drop the independence assumption by routing.
Transferable methods
Counterfactual tables as a linear program
A model that cannot see the routing choices has to fix one answer per party for every choice that party might make. The observed statistics are then marginals of a single distribution over those tables. Existence is a linear feasibility problem, and when it fails, linear programming duality returns an inequality that every classical model satisfies and the data violate.
A single classical source means \(p(abc|jk)=\int p(a|\lambda)\,p(b|j,\lambda)\,p(c|k,\lambda)\,\mu(d\lambda)\), with \(\mu\) independent of the routing choices (free choice) and each response depending only on its own party's choice (locality). At level \((2,1,1)\), with four outcomes per party, this holds if and only if there is \(q(a,b_1,b_2,c_1,c_2)\ge0\) with \[\sum_{b_{\bar j},\,c_{\bar k}}q(a,b_1,b_2,c_1,c_2)\Big|_{b_j=b,\,c_k=c}=p(abc|jk)\quad\text{for all }j,k,a,b,c,\] that is, \(4^5=1024\) unknowns and 256 equality constraints. Necessity: take \(q=\int p(a|\lambda)\prod_jp(b_j|j,\lambda)\prod_kp(c_k|k,\lambda)\,\mu(d\lambda)\). Sufficiency: the source samples a full table from \(q\) and each party reads off the entry for its choice.
If no \(q\) exists, duality gives coefficients \(y_{jk}(abc)\) with \(\sum y_{jk}(abc)\,p(abc|jk)<0\) while \(\sum_{jk}y_{jk}(a\,b_j\,c_k)\ge0\) for every deterministic table. Every single source is a mixture of deterministic tables, so it satisfies the inequality. Any classical network, with one variable per copy of each source and an arbitrary joint distribution, becomes a single source by taking \(\lambda\) to be all the variables together, which is Proposition 2. Conversely, if all rounds give the same distribution \(p_0\), the table that repeats one sample of \(p_0\) in every entry passes (Proposition 3): only the difference between closed and open data carries any power.
Editorial step for the Fritz example: if the open data factorise as \(p_{\mathrm{open}}(abc)=p(a)\,p(b|a)\,p(c|a)\), set \[q(a\,b_1b_2\,c_1c_2)=p(a)\,p_{\mathrm{closed}}(b_1c_1|a)\,p_{\mathrm{closed}}(b_2c_2|a).\] Its \((a,b_1,c_1)\) marginal is \(p_{\mathrm{closed}}\), and its \((a,b_1,c_2)\) marginal is \(p(a)\,p(b|a)\,p(c|a)\), the open data. Editorial check: recomputing the RGB4 data and solving this program reproduces the paper's verdicts at every point tested, including feasibility at \(u^2=\tfrac12\) and the loss of every violation between \(v=0.97\) and \(v=0.965\).
How to reuse it
When an assumption concerns hidden internals that the experimenter cannot inspect, look for a physical randomisation, here a switch, that turns it into a setting the experimenter controls. Consistency across settings then becomes a Fine-type linear program over counterfactual tables, with certificates that can be checked exactly.
A de Finetti limit for many copies of one source
Adding copies of \(\alpha\) adds constraints, so the levels form a hierarchy. Because the copies are interchangeable, a feasible table can be symmetrised over them, and in the limit of infinitely many copies the de Finetti theorem says what kind of classical model is left: a mixture of independent, identical emitters.
At level \((n,1,1)\), write \(q^{(n)}(a,\vec b,\vec c)\) with \(\vec b=(b_1,\dots,b_n)\) and \(\vec c=(c_1,\dots,c_n)\); the constraints fix the \((a,b_j,c_k)\) marginal to \(p_{\mathrm{closed}}\) when \(j=k\) and to \(p_{\mathrm{open}}\) when \(j\ne k\). Permuting the copies maps closed constraints to closed ones and open to open, so averaging a feasible \(q^{(n)}\) over all permutations gives a feasible exchangeable one. Compactness and a diagonal argument give consistent limits for every \(n\), and the Kolmogorov extension theorem yields an infinite exchangeable sequence of pairs \((b_m,c_m)\) given \(a\).
The de Finetti theorem then writes the limit as a mixture of product distributions, \[Q^{(n)}(a,\vec b,\vec c)=p(a)\int\prod_{m=1}^n\lambda(b_mc_m)\,\mu_a(d\lambda),\] so the data pass at every level exactly when \(p_{\mathrm{closed}}(abc)=p(a)\int r(bc|a,\lambda)\,\mu_a(d\lambda)\) and \(p_{\mathrm{open}}(abc)=p(a)\int r(b|a,\lambda)\,r(c|a,\lambda)\,\mu_a(d\lambda)\). The open data must be the same mixture of products of marginals as the closed data is of joints.
Editorial check on the elegant joint measurement at level \((2,1,1)\): the closed data are \(\tfrac{25}{256}\), \(\tfrac{1}{256}\) and \(\tfrac{5}{256}\) for three equal outcomes, exactly two equal, and all different, and the open data split further by which pair agrees. Exact rational arithmetic confirms that the paper's model, the overlap \(\min(p_{\mathrm{closed}},p_{\mathrm{open}})\) on tables with \(b_1=b_2\) and \(c_1=c_2\) plus a correction of weight \(\tfrac{2}{256}\) on tables where one copy agrees with \(a\) and the other gives three different outcomes, is nonnegative, normalised and reproduces all four routed marginals. Whether some finite level excludes this distribution is exactly the question Proposition 4 reduces to the mixture condition above.
How to reuse it
This is the classical counterpart of symmetric-extension hierarchies for entanglement. Whenever constraints come from interchangeable copies, symmetrise, pass to the limit and use de Finetti to identify the model that survives every level; finite de Finetti theorems then estimate how fast the levels converge.
Context & limitations
Why it matters. Source independence concerns hidden device internals and cannot be checked; the routing bits are generated by the experimenter and can be space-like separated from the sources and from the other parties, as settings are in loophole-free Bell tests. The proposal therefore trades an untestable assumption for locality and free choice of a switch, at the cost of one more source and two fast fibre switches. It also excludes classical explanations at a point, \(u^2\approx0.785\), where the triangle data alone have a numerically found model with independent sources.
Limitations. Only classical explanations are excluded: a single quantum source shared by all parties reproduces the routed data of any quantum network, and which quantum models the test can exclude is open. Because the routing choices act as settings, what the test certifies is Bell nonlocality with switch positions as inputs, not network nonlocality of the triangle itself; the paper notes that the two tests answer different questions. All verdicts and range endpoints are numerical, resting on dual certificates checked outside the solver. The 3% noise tolerance is at the level of current triangle sources before switching and measurement noise are added. Rounds are assumed independent and identically distributed, finite statistics are not treated, and the detection loophole must be closed. The author states that large language models assisted with literature search, numerical and analytical checks, and drafting and editing.