Foundations of quantum mechanics · Theory and numerics (AI-assisted)
Linear-optical nonlocality without postselection
arXiv:2611.19570v1 ↗PDF ↗Submitted 14 November 2026
- Question
- Single photons, passive interferometers and photon-number detectors are the resources of BosonSampling, which is believed not to be universal for computation. They cannot make a Bell pair deterministically, and a long-known two-photon scheme reaches a CHSH value of \(1+\sqrt2\) only if every run is kept, against Tsirelson's bound of \(2\sqrt2\). Is \(1+\sqrt2\) the best that two photons can do without postselection, and do more photons get closer to \(2\sqrt2\)?
- Result
- With two input photons, any interferometer for state preparation, any split of the modes and any measurements that conserve each party's photon number, the CHSH value is at most \(1+\sqrt2\approx2.414\). With \(2k\) photons, running \(k\) copies and reporting the first successful one gives \(2\sqrt2-(\sqrt2-1)/2^{k-1}\), which tends to \(2\sqrt2\). Six photons in two three-mode Fourier interferometers violate the Coiteux-Roy–Wolfe–Renou inequality for genuine tripartite nonlocality (10.24 against a bound of 10), and ladders of eight-mode interferometers take the Mermin value to its maximum of 4 asymptotically.
The paper in detail
A Bell test asks two parties, each choosing one of two \(\pm1\)-valued measurements, to produce correlations whose CHSH combination \(S\) exceeds 2, the limit for any local realistic model; quantum mechanics allows up to \(2\sqrt2\). With an entangled pair and arbitrary local measurements that maximum is easy. The authors ask what happens with the restricted toolkit of linear optics: single photons in separate modes, beam splitters and phase shifters, and detectors that count photons, with no feedforward and no discarded runs. This toolkit defines BosonSampling, thought to be neither classically simulable nor universal, and it cannot make a Bell pair on demand.
The standard scheme sends two photons through two beam splitters and a swap, giving \(\tfrac1{\sqrt2}(\ket{\mathrm{Bell}}+\ket{\mathrm{fail}})\), where in the failure branch one party receives both photons. Discarding failures (postselection) gives the full \(2\sqrt2\); the discard is benign because photon-number conservation makes both parties see the failure, but it turns into an ordinary detection loophole once photons are lost. Keeping every run, the parties output fixed answers when they see 0 or 2 photons, worth exactly 2, and the average is \(\tfrac12\cdot2\sqrt2+\tfrac12\cdot2=1+\sqrt2\). A recent experiment measured \(2.23\pm0.07\) this way.
The first theorem shows that no choice of interferometers does better with two photons. A local photon-number-conserving measurement cannot create coherence between sectors with different local photon numbers, so \(S\) is a weighted average over sectors, and the sectors where one party has no photon contribute at most 2. The only useful sector, one photon each, is a sum of two product vectors, and its entanglement is limited by how the two input photons are routed. Making that sector likely forces it to be weakly entangled, and the optimum of the trade-off is \(1+\sqrt2\) at success probability one half.
More photons remove the ceiling. With \(2k\) photons the parties run \(k\) copies of the two-photon scheme side by side, measure the same setting in each and report the first copy in which each received one photon. Failures are perfectly correlated within each copy, so they always report the same Bell pair, and only the all-fail event, of probability \(2^{-k}\), costs anything. The value approaches \(2\sqrt2\) exponentially, which contradicts the earlier suggestion that superselection forbids a maximal violation at any particle number. The successful pairs are not heralded, though: success is read only at measurement time.
For three parties the authors feed six photons into two three-mode Fourier interferometers (tritters) and give one output of each to every party. A suppression law forbids any tritter output with exactly two photons in one mode, so a party that sees 1 or 4 photons knows the run produced one of three GHZ states, in which one qubit is encoded in the four-photon basis \(\{\ket{31},\ket{13}\}\); this happens with probability 4/9 and all parties agree. Because a phase shifter rotates that encoding by twice the angle, one setting cannot be optimal for every GHZ branch, and the values reached are 2.67 (Mermin), 4.44 (Svetlichny, above its bipartite bound of 4) and 10.24 (CWR, above 10). Ladders raise these towards 3.5, 5 and 10.54; replacing tritters by four-mode Sylvester interferometers removes the angle obstruction and takes Mermin to 4 and CWR to about 10.65. Whether CWR can reach its quantum maximum without postselection, and how loss and distinguishability affect the tripartite flags, remain open.
Transferable methods
Sector decomposition and a Schmidt-rank-two trade-off
The bound needs to tie how often both parties hold one photon to how entangled that branch can be. Splitting the state by local photon number reduces the problem to one branch, and writing that branch as a sum of two product vectors bounds its Schmidt coefficients by the routing probabilities of the two input photons.
After an \(M\)-mode interferometer \(U\) acting on \(a_1^\dagger a_2^\dagger\ket0\), the state splits into branches \(\ket{\psi_{20}}\), \(\ket{\psi_{02}}\), \(\ket{\psi_{11}}\) by how many photons Alice holds, with probabilities \(p_{20}+p_{02}+p_{11}=1\). Local measurements conserve photon number, so there is no interference between branches, and outputs fixed at the classical optimum give \(S_{20}=S_{02}=2\): \[S=2(1-p_{11})+p_{11}S_{11}.\]
Write \(\ket{u_k}\) and \(\ket{v_k}\) for the parts of input photon \(k\) routed to Alice and Bob, with \(t_k=\|u_k\|^2\). Then \(\ket{\psi_{11}}=\ket{u_1}\ket{v_2}+\ket{u_2}\ket{v_1}\) has Schmidt rank at most two, with coefficients \(\alpha,\beta\), and Gisin's bound reads \(p_{11}S_{11}\le2\sqrt{p_{11}^2+4|\alpha\beta|^2}\). The product of the Schmidt coefficients is a product of Gram determinants, \[|\alpha\beta|^2=\bigl(t_1t_2-|\langle u_1|u_2\rangle|^2\bigr)\bigl((1-t_1)(1-t_2)-|\langle v_1|v_2\rangle|^2\bigr)\le\tfrac1{16}.\]
For \(p_{11}\le\tfrac12\), Tsirelson's bound in the branch gives \(S\le2+(2\sqrt2-2)p_{11}\le1+\sqrt2\). For \(p_{11}\ge\tfrac12\), \[S\le2(1-p_{11})+2\sqrt{p_{11}^2+\tfrac14},\] whose derivative \(-2+2p_{11}/\sqrt{p_{11}^2+1/4}\) is negative, so the maximum is again \(1+\sqrt2\), at \(p_{11}=\tfrac12\). Editorial check: on 2000 random interferometers with 3 to 6 modes and random splits, the Gram-determinant identity holds to \(10^{-16}\) and the bound \(1/16\) is never exceeded.
How to reuse it
When a conserved local quantity splits a state into sectors that cannot interfere, bound each sector separately and look for a trade-off between a sector's weight and its quality. For any state that is a sum of two product vectors, Gram determinants give the product of Schmidt coefficients directly, which is often all an entanglement bound needs.
Correlated failure flags and first-success ladders
Without postselection, failed runs must still yield outputs. If every party can see failure locally and all flags agree, a failure costs only the classical value, and running copies in parallel while reporting the first success makes failure exponentially rare. Suppression laws supply such flags in the tripartite case.
Let a run succeed with probability \(s\), giving the value \(S_{\mathrm{succ}}\), and let failures give the local bound \(L\). With \(k\) independent copies whose flags agree within each copy, reporting the first success gives \[S_k=S_{\mathrm{succ}}-(S_{\mathrm{succ}}-L)(1-s)^k.\] For CHSH, \(s=\tfrac12\), \(S_{\mathrm{succ}}=2\sqrt2\) and \(L=2\) give \(2\sqrt2-(\sqrt2-1)/2^{k-1}\), or \((1+3\sqrt2)/2\approx2.621\) at \(k=2\). For Svetlichny with tritters, \(s=4/9\), \(S_{\mathrm{succ}}=5\) and \(L=4\) give \(5-(5/9)^k\).
The flag comes from the tritter suppression law: three photons entering a three-mode Fourier interferometer one per mode never leave with exactly two photons in a mode. Editorial check: computing the output amplitudes as permanents gives \(\ket{111}\) with probability 1/3 and each of \(\ket{300},\ket{030},\ket{003}\) with probability 2/9, and two tritters give every party 1 or 4 photons with total probability 4/9, as stated.
Loss makes flags disagree. With detection efficiency \(\eta\) per photon, a copy succeeds with \(s=\eta^2/2\), fails for both parties with \(f=(1-\eta)^2+\eta^2/2\) and leaves one party with a lone photon with \(d=2\eta(1-\eta)\), which contributes nothing. Then \[S=2\sqrt2\,s\,\frac{1-f^k}{1-f}+2f^k,\] which exceeds 2 exactly when \(\sqrt2\,s>s+d\), that is \(\eta>4/(3+\sqrt2)\), for every \(k\), with limit \(2\sqrt2\,\eta/(4-3\eta)\). Editorial check: at \(\eta=4/(3+\sqrt2)\approx0.9062\), \(\sqrt2\,s-(s+d)\) vanishes to machine precision.
How to reuse it
Any protocol with a success flag that every party can read locally and that all parties agree on can be amplified by parallel copies and a fixed reporting order, at linear cost in resources for exponentially small failure. The check that matters is whether noise desynchronises the flags, which is where the detection-efficiency threshold comes from.
Context & limitations
Why it matters. It proves for the first time that the value \(1+\sqrt2\), reported many times for particular schemes, is optimal for two photons, and it shows that postselection-free linear optics has no ceiling below Tsirelson's bound: the price of a stronger violation is photon number, not a missing operation. Interferometric suppression laws, so far used to certify indistinguishability, become a way for every party to know locally that a run succeeded.
Limitations. All schemes rely on each party reading success from its own photon count, so photon loss and partial distinguishability break the flags. Loss is analysed only for the bipartite schemes (a violation survives for detection efficiency above \(4/(3+\sqrt2)\approx0.906\)); the tripartite schemes are left for later work. The two-photon bound assumes measurements that conserve local photon number, which excludes coherent-state references. The tripartite settings come from numerical optimisation, and the tritter ladder stays below the quantum CWR maximum of \(2\sqrt2+8\approx10.83\). The authors state that Claude Opus 5.5, Fable 5.1 and Gemini Carbon aided all the numerical optimisation and the polishing of text and figures, and that they verified the results themselves.