Bell tests certify that two measurement devices share correlations no classical mechanism could produce. The standard escape route for a classical explanation is the "detection loophole": if detectors miss enough events, a classical model can fake the quantum statistics by quietly failing to fire at convenient moments. Laplante and coauthors showed that a broad family of Bell inequalities built from communication complexity can be made immune to this loophole. In their construction, a missed detection is treated as an extra outcome, "abort", and the quantum violation shrinks by at most a constant factor (Robust Bell inequalities from communication complexity). Their guarantee rests on one quiet assumption. Every detector fails with the same probability $\eta$, whatever outcome it would have recorded.
Real single-photon detectors do not behave this way. The detectors used in the best photonic Bell tests are superconducting nanowire single-photon detectors (SNSPDs), wires held just below their superconducting transition that "click" when a small energy deposit breaks superconductivity locally. Their efficiency depends on how much energy is deposited, and it falls off sharply near a threshold. The dark-matter community has characterized exactly this near-threshold behavior while pushing SNSPDs down to sub-eV recoil detection: QROCODILE reached a 0.11 eV threshold (First Sub-MeV Dark Matter Search with the QROCODILE Experiment). This note proposes to extend Laplante-style inefficiency resistance to detectors whose failure probability depends on the outcome's energy. The motivating case is high-dimensional Bell tests whose outcomes are frequency or energy bins, where some bins necessarily sit on the detector's efficiency turn-on curve. The contribution is a sound certificate for that setting, together with a quantitative answer to whether it beats the obvious fix of adding loss until the efficiency is flat.
Core Claim
When the chance of an abort depends on which outcome would have occurred, a classical model gains new freedom. It can suppress the outcomes that would expose it and let the others through. Laplante's constant-factor guarantee does not cover this case. Applying their functionals with an average efficiency can certify nonclassicality that is not there.
We model the observed statistics as follows. Alice's detector registers outcome $a$ with probability $\eta_A(a)$ and aborts otherwise; Bob's registers $b$ with probability $\eta_B(b)$. Here $P_Q(a,b\mid x,y)$ is the ideal quantum distribution for measurement settings $x,y$. The observed probability of the pair $(a,b)$ is then
$$P_{\mathrm{obs}}(a,b\mid x,y)=\eta_A(a)\,\eta_B(b)\,P_Q(a,b\mid x,y).$$
The remaining probability mass goes to abort events. Each efficiency $\eta(a)$ is drawn from a turn-on curve $\eta(E_a)$ evaluated at the energy $E_a$ that encodes outcome $a$.
Our hypothesis is that the right transformation of a normalized Bell functional still exists in this setting. We further hypothesize that its violation degrades by a factor controlled by the mismatch ratio $\kappa=\eta_{\min}/\eta_{\max}$ rather than by $\eta_{\min}$ alone. If that holds, keeping all clicks and reweighting outcomes is strictly better than attenuating every channel down to the worst one. If it fails, the flattening fix is optimal, and the paper reports that negative result.
What This Is Not
The research file's candidate gaps suggest using Bell or entanglement witnesses to separate dark-matter recoils from micro-discharge backgrounds in PandaX-4T's ionization-only data (Search for Light Dark Matter with 259 Days of Data in PandaX-4T). We reject that direction. A single particle recoil deposits energy through a classical cascade, so no Bell functional can be violated by it. A test that no signal can pass cannot discriminate signal from background.
The genuine link runs the other way. Low-energy recoil physics supplies a realistic model of the detector's threshold behavior. Bell theory needs such a model, and has so far used only a flat $\eta$.
Direction 2 therefore plays two roles here. First, it is the source of the detector model. QROCODILE's threshold-limited SNSPD operation, and the phonon-level description of sub-threshold energy deposition in solids (Phonon-mediated Migdal effect in semiconductor detectors), motivate a smooth energy-dependent turn-on rather than a step. Second, it constrains a minor noise term. QROCODILE recorded 15 unexplained pulses in 415.15 hours. Spread over nanosecond coincidence windows, that rate is negligible, and we treat recoil-like false clicks as a secondary check rather than a central effect.
Research Program
The work has three stages.
Stage 1: exact local bounds by linear programming. For small scenarios, we compute the exact local bound with outcome-dependent aborts. The deterministic local strategies are those that may output any outcome or abort. For each strategy, the achievable abort pattern is constrained by the outcome-dependent efficiency profile. The local bound of any candidate functional is then a linear program over mixtures of these strategies. We run this for CHSH, for the CGLMP family (a standard high-dimensional Bell inequality) up to $N=10$ outcomes, and for the smallest instances of Laplante's Gap Orthogonality and Vector in Subspace functionals.
We choose $N\ge 8$ deliberately. Atkin and Zohren showed that random states violate CGLMP-type inequalities almost surely from $N\ge 8$ outcomes (Violations of Bell inequalities from random pure states). Frequency-bin encodings with that many bins necessarily span a wide energy range, which is where mismatch bites.
Stage 2: analytic transformation. We generalize Laplante's normalization step: rescale the functional's coefficients by outcome-dependent weights so the classical bound stays at one. We then prove, or refute, a violation-loss bound expressed in $\kappa$. The linear programs from Stage 1 serve as ground truth wherever they are tractable. The analytic result matters for large communication-complexity functionals, whose violations scale as $2^{c-2q}$ and whose local polytopes are too large to enumerate.
Stage 3: realistic efficiency profiles. We parameterize turn-on curves as smooth sigmoids in deposited energy, with the threshold placed near 0.11 eV as QROCODILE reports and with varying widths. We then place frequency-bin outcomes at telecom and mid-infrared photon energies. For each placement we compute:
- the certified violation;
- the tolerance to white noise, using the same noise model as Laplante;
- tolerance to a small rate of false clicks, independent between Alice and Bob, at a rate bounded by the QROCODILE pulse rate.
Baselines
Naive reference: flat-efficiency certification. This applies Laplante's transformation unchanged, plugging in the average efficiency across outcomes. It is what an experimenter would do today with the published result. We run it on every efficiency profile in Stage 3. For each profile, we then use the Stage 1 linear program to search for a local model that exceeds its claimed bound. Any such model shows the certification is unsound. Our method "beats" this baseline by being sound wherever the naive method is not; the headline result is the fraction of realistic profiles on which the naive method certifies falsely.
Strongest simple competitor: flattening. This inserts extra loss into each channel so that every outcome is detected with efficiency $\eta_{\min}$, then applies Laplante's guarantee as written. It is sound by construction and needs no new theory, which is why it is the real threat. We compare it with the reweighted functional on identical efficiency profiles, identical quantum states, and identical measurement settings, at matched statistical confidence. Settings are optimized with the same see-saw budget for both methods. The reweighted functional counts as better only if it certifies a larger violation, or survives lower $\eta_{\min}$, at equal count budgets.
Existing detection-loophole-sensitive inequalities. CHSH and CGLMP are evaluated under the same mismatched profiles, using the exact linear-program bound. This shows whether the communication-complexity functionals' constant-factor robustness survives mismatch at all. The file offers only the textbook symmetric CHSH critical efficiency of roughly 82.8%. It gives no Lacuna-reported critical efficiency under outcome-dependent loss for any functional, so these numbers will be produced here.
Ablation. We keep the reweighted functional but replace the sigmoid turn-on with a step function at the same threshold. This isolates whether the smooth near-threshold behavior, the ingredient taken from the low-energy detector literature, changes conclusions compared with a crude cutoff.
Why not just…?
Why not just flatten the efficiencies by adding loss? This alternative may well work as well as ours, and the proposal is built to find out. Flattening throws away the difference between $\eta_{\max}$ and $\eta_{\min}$. If Laplante's violation-loss factor is dominated by the efficiency level rather than its spread, flattening loses little. The Stage 1 linear programs give the exact answer for small scenarios. We will report plainly if flattening matches the reweighted functional within statistical error. The contribution would then shrink to a soundness warning plus a recommendation to flatten.
Why not just operate the detectors far above threshold, where efficiency is flat? For qubit-level polarization tests at telecom wavelengths this is correct, and our method adds nothing there. The proposal matters only when outcomes are spread across a broad energy band, as in frequency-bin tests with $N\ge 8$ bins or mid-infrared photons. We quantify the band over which efficiency varies by less than a given tolerance. Outside that band, a flat-threshold choice is simply unavailable.
Why not just solve the linear program for each experiment and skip the theory? For CHSH and small CGLMP this suffices, and we use it as ground truth. Laplante's large-violation functionals come from communication problems such as Disjointness, and their local polytopes grow exponentially with input size. For those functionals only an analytic transformation yields certificates. The Stage 1 results check whether the analytic bound is tight where comparison is possible.
Decisive Checks
Three outcomes settle the proposal.
First, soundness. On at least some realistic sigmoid profiles, the naive flat-$\eta$ certification must admit a local model that violates it. If it never does, the soundness concern is moot.
Second, advantage. The reweighted functional must certify a strictly larger violation than flattening, at matched statistics, on profiles with $\kappa$ well below one. The crossover value of $\kappa$ is the main quantitative deliverable.
Third, the realistic regime. With turn-on parameters anchored to QROCODILE's 0.11 eV threshold, we test whether an $N\ge 8$ frequency-bin CGLMP test remains certifiable. The same test includes false clicks at QROCODILE's observed pulse rate, confirming that recoil-like backgrounds stay negligible in coincidence windows.
A clear negative result on the second check is still publishable. It would show that the simple fix is optimal and would close the question for experimental groups designing broadband Bell tests.
References
- Robust Bell inequalities from communication complexity
- First Sub-MeV Dark Matter Search with the QROCODILE Experiment Using Superconducting Nanowire Single-Photon Detectors
- Search for Light Dark Matter with 259 Days of Data in PandaX-4T
- Phonon-mediated Migdal effect in semiconductor detectors
- Violations of Bell inequalities from random pure states