Claims that quantum links beat classical ones are usually checked in ideal simulators. Examples include sending two bits per transmitted qubit using pre-shared entanglement, or winning a distributed computing game more often than any classical strategy allows. In a real link, two kinds of error stand between the theoretical number and a trustworthy measurement. The first is statistical: there are only finitely many measurement shots, so the error shrinks like $1/\sqrt{N}$. The second is systematic: photon detectors lose efficiency over time, dark-count rates wander, and quantum memories lose coherence while idle. These biases do not shrink with more data. We propose to measure quantum advantage the way reactor-neutrino physicists measure oscillations. They do not trust absolute event counts. They take ratios between two measurements that share the same nuisance factors, so those factors cancel.
Concretely, the quantum protocol and a matched classical reference protocol are run interleaved, round by round, on the same emulated physical channel. The advantage is reported as the ratio of their performance. This cancels "common-mode" systematics such as loss and detector efficiency, which hit both protocols equally. It leaves only the "differential" systematics, such as decoherence and misaligned joint measurements, which hit only the quantum side and should count against it. The result is a procedure with three parts: an estimator, an error budget that says which effects cancel and which remain, and a stopping rule that says when an advantage is certified and when collecting more shots is futile.
Core Claim
Our main claim is about how the advantage should be estimated and certified. Let $S_q$ and $S_c$ be the success rates or delivered bits per channel use of the quantum and classical protocols, measured in interleaved rounds on one channel. Then the ratio estimator $\hat A = \hat S_q / \hat S_c$ is insensitive to any slowly varying multiplicative factor $\eta(t)$, such as detector efficiency or transmission, that scales both protocols. This holds as long as $\eta$ changes little within one interleaving period.
What remains is a differential systematic $\delta$, and we will bound it explicitly. Advantage is certified only when
$$\hat A - z\,\sigma_A/\sqrt{N} - \delta > 1,$$
where $\sigma_A$ is the per-round standard deviation of the ratio, $N$ is the number of interleaved rounds, and $z$ sets the confidence level. The same quantities give a crossover point, $N^* = (\sigma_A/\delta)^2$, past which extra shots buy nothing. This is the same equality of statistical and systematic error that limits the saturated instruments discussed in Performance of the MAGIC stereo system obtained with Crab Nebula data. The form of $N^*$ follows directly from setting $\sigma_A/\sqrt{N}=\delta$.
The interesting outcomes are two. One is a measured map of where the claimed advantages survive, across memory coherence time, drift rate and shot budget. The other is evidence that the naive absolute-count estimator gives overconfident, biased advantage claims in exactly the regimes theorists care about.
Relation to Prior Work
The methodological template comes from Model-independent ν̄e short-baseline oscillations from reactor spectral ratios. That work divides antineutrino spectra measured at two baselines (DANSS at 10.7 m and 12.7 m, and NEOS normalized to Daya Bay) so that reactor-flux normalization cancels. It also exposed a roughly 5% overestimate of the uranium-235 flux that absolute counting would have absorbed. We reuse the logic of shared-nuisance cancellation. We change what plays the role of the "baseline": instead of two distances, we pair two protocols on one channel.
The claims we test come from three theoretical papers. Quantum Advantages of Communication Complexity from Bell Nonlocality maps Bell tests onto distributed computing games and reports success 3/4 versus a classical 1/2 for the four-cycle game, and $(2m-1)/2m$ for $2m$-cycles. It assumes ideal measurements and analyzes neither shot noise nor detector drift. The Multiple-Access Channel with Entangled Transmitters proves that entanglement enlarges the capacity region of a shared classical channel. Its result is asymptotic and assumes perfectly characterized, stationary channels. We use it only as motivation, because its regularized formulas are not directly measurable. Exponential Quantum Communication Advantage in Distributed Inference and Learning proves polylogarithmic communication for distributed learning. Its experiments, however, are classical simulations of a polynomial-activation network, and the authors state that fault-tolerant hardware is required. That puts it outside what our testbed can validate. We treat it as a constraint on scope, not a target.
The closest empirical work is Integrating Quantum Simulation for Quantum-Enhanced Classical Network Emulation. It bridges the quantum network simulator QuNetSim into the classical network emulator ComNetsEmu. It then shows that buffering entangled pairs during idle periods (the GEWI protocol, "Generate Entanglement While Idle") raises average bits per data qubit from 1 toward the superdense-coding limit of 2 on bursty 168-bit packets. That evaluation is noiseless, assumes unlimited memory coherence, and models no detector drift. Our work keeps its testbed and protocol but replaces its implicit assumption, that the measured rate equals the true rate, with an explicit estimator and error budget.
Entanglement is Necessary for Optimal Quantum Property Testing plays a constraining role. It proves that single-copy measurements need $\Omega(d^{4/3}/\epsilon^2)$ copies where entangled ones need $O(d/\epsilon^2)$. Protocols restricted to single-copy readout are therefore pushed to large $N$, which is exactly where $N^*$ bites.
Research Program
We will extend the QuNetSim–ComNetsEmu bridge with a drift layer. This layer injects time-varying detector efficiency $\eta(t)$, dark-count rates, asymmetric loss between the two receiver arms, and exponential memory decoherence during idle periods, all under user-specified spectra. Because the true values are known inside the emulator, every estimator can be scored against ground truth. That is the main reason to start in emulation rather than on hardware.
On top of the drift layer we will build three things. The first is an interleaving scheduler that alternates quantum and classical-reference rounds over the same simulated channel. For superdense coding, the reference sends one bit per photon over the same path. For the communication-complexity games, the reference is the optimal classical strategy with the same communication budget.
The second is an analytic error budget. It classifies each injected effect as common-mode (cancels in the ratio) or differential (survives), and derives $\delta$ for each. Detector efficiency that is identical across protocols cancels exactly. Dark counts cancel only partly, because they hurt a Bell-state measurement differently from a single-bit readout. Decoherence is purely differential. The budget will spell these cases out.
The third is the certification rule above, with $N^*$ reported for every run. We will also characterize when the method breaks. It should break when the drift correlation time approaches the interleaving period. We will measure the coverage loss as a function of that ratio, so users know the minimum interleaving rate their hardware requires.
Baselines
The naive reference point is the ideal-simulator claim itself. GEWI in Integrating Quantum Simulation for Quantum-Enhanced Classical Network Emulation reports bits per qubit rising from 1 toward 2. That paper gives no exact finite value, and no reported result under noise was found on Lacuna. The communication-complexity games have the theoretical 3/4 versus 1/2 from Quantum Advantages of Communication Complexity from Bell Nonlocality. We re-run these exactly in our emulator with drift switched off, to confirm we reproduce them. We then switch drift on and report how far the ideal number misstates the true advantage.
The strongest competitor is the procedure the field implicitly uses. Calibrate the channel once beforehand, as QuNetSim-style simulators treat channel estimation, then compute advantage from absolute counts of the quantum protocol divided by the calibrated classical expectation. No reported result for this estimator under drift was found on Lacuna, so we will establish it ourselves. To keep the comparison fair, it receives exactly the same total shot budget as our method, counting calibration shots. It uses the same emulated drift realizations and the same confidence-interval construction. We also grant it periodic recalibration at a tuned interval chosen on held-out drift seeds, with the same tuning budget we give our interleaving period.
We beat it if, at equal shots, our certified intervals cover the true advantage at their nominal rate where its intervals do not. A second criterion is that our false-certification rate stays at its nominal value when the true advantage is at or below 1, and its rate does not. Ties in benign, slow-drift regimes are expected and will be reported as such.
Two ablations isolate the key ingredient. The first, ratio without interleaving, runs quantum and classical rounds in long sequential blocks and then divides. It tests whether cancellation needs temporal proximity. The second, interleaving without the ratio, runs interleaved rounds but analyzes each protocol's absolute rate separately. It tests whether the ratio itself, not merely the scheduling, does the work.
Why not just…?
Why not just collect more shots? More shots shrink only the $1/\sqrt{N}$ term. Precision experiments in this file repeatedly hit a floor where more data stops helping. Examples are luminosity and alignment limits in Exclusive J/ψ detection and physics with ECCE and instrument-response limits in the MAGIC Crab data. Worse, beyond $N^*$ a biased absolute estimator becomes more confidently wrong. Our crossover experiment shows this directly by plotting coverage against $N$ for both estimators.
Why not just add depolarizing noise to the simulator and read off the advantage? Static symmetric noise is what existing simulators already assume. It produces a lower but still exactly known advantage, and it never tests whether the measurement procedure can recover the truth when nuisance parameters drift. The question here concerns estimation under unknown, time-varying systematics, which static noise cannot pose.
Why not just recalibrate the detectors frequently? This is the real threat. Frequent recalibration is close in spirit to interleaving, and with enough calibration shots it could match our coverage. The difference is cost: calibration shots carry no protocol information, whereas interleaved classical rounds are both reference and calibration. The equal-budget comparison with a tuned recalibration interval settles this. If recalibration matches us at equal shots across drift regimes, our contribution shrinks to the error budget and the $N^*$ rule, and we will say so.
Why not just compare quantum distributed learning against classical compression instead? EXACT: Scalable Graph Neural Networks Training via Extreme Activation Compression reports up to 32x activation compression with 0.2–0.5% accuracy loss. That is a fair challenge to the distributed-learning claim. But that claim cannot yet be run on any channel, by its own authors' account. We therefore restrict ourselves to protocols that can actually be executed and measured, and leave that comparison to work that has hardware.
Proposed Experiments
The first experiment runs GEWI superdense coding on bursty 168-bit traffic in the extended emulator. We sweep memory coherence time against idle duration, detector-efficiency drift amplitude and correlation time, and dark-count rate. For each setting we report true advantage, both estimators' point estimates, interval coverage over many drift seeds, and the predicted versus observed $N^*$. The output is a phase diagram of where the advantage in bits per qubit remains certifiably above 1 once finite memory is included.
The second experiment repeats this for the four-cycle and larger $2m$-cycle communication games. Larger $m$ raises the ideal quantum success but adds measurement settings, and so more differential systematics. That makes these games a natural test of whether the error budget predicts which $m$ remains certifiable.
The decisive checks are two. The first is coverage under null: when injected decoherence drives the true advantage to exactly 1, the false-certification rate must stay at its nominal level for our estimator. We expect it to rise with $N$ for the static-calibration estimator. The second is the break test, which sweeps the ratio of drift correlation time to interleaving period. Cancellation should degrade as this ratio approaches one, and if it does not degrade where the budget predicts, the error model is wrong.
A further hypothesis, which we will test rather than assume, is that the estimator carries over to multi-sender links of the kind studied in The Multiple-Access Channel With Entangled Transmitters. On those links, unequal drift across senders would make maximal-error and average-error performance diverge again.
References
- Performance of the MAGIC stereo system obtained with Crab Nebula data
- Model-independent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:msub><mml:mrow><mml:mover accent="true"><mml:mrow><mml:mi>ν</mml:mi></mml:mrow><mml:mrow><mml:mo stretchy="false">¯</mml:mo></mml:mrow></mml:mover></mml:mrow><mml:mrow><mml:mi>e</mml:mi></mml:mrow></mml:msub></mml:math> short-baseline oscillations from reactor spectral ratios
- Quantum Advantages of Communication Complexity from Bell Nonlocality
- The Multiple-Access Channel With Entangled Transmitters
- Exponential Quantum Communication Advantage in Distributed Inference and Learning
- Integrating Quantum Simulation for Quantum-Enhanced Classical Network Emulation
- Entanglement is Necessary for Optimal Quantum Property Testing
- Exclusive J/<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline" id="d1e654" altimg="si6.svg"><mml:mi>ψ</mml:mi></mml:math> detection and physics with ECCE
- EXACT: Scalable Graph Neural Networks Training via Extreme Activation Compression