Quantum coherence is the "superposition-ness" of a quantum state, measured by how large the off-diagonal entries of its density matrix are in a chosen reference basis. Physicists quantify it with several inequivalent formulas, such as the $l_1$-norm and the relative entropy of coherence, and these formulas sometimes give opposite answers. One experiment found that the same photon state certifies a nonlocal coherence advantage under one measure and fails under another. Those verdicts, however, were computed from a single reconstructed density matrix, and that matrix came from an optical apparatus with small calibration errors: slightly misset waveplates, unequal detector efficiencies, dark counts. Errors of this kind do not shrink when more photons are collected. This proposal asks whether the reported disagreements between measures are physical, or whether they partly come from systematic errors that each measure amplifies in its own way.
We propose to bring the toolkit that precision particle and neutrino physics uses against irreducible systematic errors into coherence certification. That toolkit has three parts: nuisance parameters for known calibration effects, calibration of the test statistic against null pseudo-experiments, and cross-checks through independent measurement channels with different error budgets. The result would be a measure-by-measure verdict, with honest error bars, on whether a state has a given coherence advantage, together with a formal test of whether two measures truly disagree on that state.
Core Claim
Different calibration errors act on different parts of the density matrix. A small misrotation of the measurement frame is a unitary error. It leaves the eigenvalues of the state unchanged but reshuffles weight between diagonal and off-diagonal entries. Dark counts and loss act more like mixing with noise, which changes the eigenvalues. The $l_1$-norm, $C_{l_1}=\sum_{i\neq j}|\rho_{ij}|$, is linear in off-diagonal magnitudes. The relative entropy of coherence, $C_R=S(\rho_{\rm diag})-S(\rho)$ with $S$ the von Neumann entropy, depends on the diagonal entries and on the spectrum. The two measures therefore respond to the same systematic with different sign and size.
Our claim is that, near a certification threshold, this difference in response can open or close the region where the measures disagree. The size of the effect should be first order in the miscalibration, because the coherence of a generic state in a basis rotated by a small angle $\epsilon$ changes linearly in $\epsilon$. If the claim is right, some reported "measure-dependent advantages" will not survive a systematics-aware analysis, while others will become stronger, statistically certified disagreements. Either outcome is informative, and no existing work separates the two.
The closest work is the nonlocal advantage of quantum coherence (NAQC) experiment of Experimental investigation of the nonlocal advantage of quantum coherence. NAQC is a steering-type task: measuring one photon of an entangled pair lets an observer prepare more coherence on the other photon than any classical correlation could allow. That paper reconstructs the second photon's states by 16-setting tomography on 28 Bell-diagonal states, with average fidelity 0.9978, and shows that the NAQC verdict depends on the measure. We reuse its setup and its question. We differ in treating the reconstructed state as an estimate with systematic uncertainty rather than as an exact point.
Research Program
The first component is a forward model of the polarization tomography apparatus. Its nuisance parameters are waveplate angle offsets and retardance errors, the relative efficiency of the two detectors in each arm, dark-count and accidental rates, and polarization-dependent loss. For each state we fit raw counts by maximum likelihood with these parameters included. External calibration measurements constrain them, which plays the role of the calibration standards that bound precision experiments.
For a coherence measure $C$ and the NAQC criterion, the steered coherence summed over the three mutually unbiased measurement settings on the first photon is compared with a fixed threshold $T_C$. We define the verdict margin
$$\Delta_C(\theta) = \sum_{k} C\big(\rho_{B|k}(\theta)\big) - T_C,$$
where $\theta$ collects the nuisance parameters and $\rho_{B|k}$ is the reconstructed conditional state of the second photon for setting $k$. We report a profiled interval for $\Delta_C$ under each measure.
A disagreement between measures $C$ and $C'$ is then a joint statement: $\Delta_C>0$ and $\Delta_{C'}<0$ simultaneously over the systematic band. We test it with a statistic whose null distribution we calibrate by simulating states placed exactly on the boundary. This mirrors how Robust resonant anomaly detection with NPLM calibrates its minimum p-value over a threshold scan against signal-free pseudo-experiments. We borrow the calibration logic, not the kernel classifier.
The second component is an orthogonal cross-check channel. Independent measurement of the neutrino mixing angle \theta_{13} via neutron capture on hydrogen at Daya Bay validated its main result with a second detection channel, neutron capture on hydrogen rather than on gadolinium, which has an independent error budget. Combining the channels gave $\sin^2 2\theta_{13}=0.089\pm0.008$. By analogy, we reconstruct the same states with a second tomographic measurement set whose sensitivity to each nuisance parameter differs by design. One candidate is an informationally complete set of waveplate settings rotated from the standard one. Another is a symmetric non-projective measurement such as those discussed in Set Coherence: Basis-Independent Quantification of Quantum Coherence. A disagreement between measures counts as physical only if both channels reproduce it.
The third component is a simple differential observable that cancels a common-mode frame rotation. It takes its inspiration from the spectral ratios in Model-independent \bar{\nu}_e short-baseline oscillations from reactor spectral ratios, where dividing spectra measured at two distances cancelled the uncertainty in absolute reactor flux. Here the observable is the difference in measure-specific margins between two states measured back-to-back in the same apparatus configuration. We will test whether this observable keeps the physics while removing the shared bias.
A secondary testbed checks generality: the non-Markovianity simulator of Detecting non-Markovianity via quantified coherence: theory and experiments. A process is non-Markovian when the environment retains memory and information can flow back to the system. That paper showed that ordinary coherence measures rise spuriously under memoryless dynamics because of basis dependence. It introduced an ancilla-assisted measure that provably never increases in the memoryless case, and it observed a temporary rise at $t\approx178\lambda_0$, the signature of information flowing back. The paper assumes an ideal detector response. We ask whether that rise survives a realistic miscalibration model of the frame and the cavity transmission.
Baselines
The naive reference point is the published practice. It reconstructs one density matrix per state by maximum likelihood, plugs it into each measure, and reads the verdict directly, as in Experimental investigation of the nonlocal advantage of quantum coherence. The only reported quality figure is the average fidelity of 0.9978 over 28 states. No reported result was found on Lacuna for verdict error rates, or for uncertainty intervals on NAQC margins, under any measure. The baseline therefore has no published error bar to beat; we will compute its empirical verdict-flip rate under our simulations.
The strongest practical competitor is the same reconstruction with statistical error bars from Poisson resampling of counts, the standard bootstrap. It is strongest because it is what a careful experimentalist already does, and it captures every error that falls as more photons are collected. It will run on identical count data with the same number of resamples as our pseudo-experiments. Our method beats it if, on simulated data with injected calibration offsets at realistic magnitudes, the bootstrap intervals undercover the true margin while ours reach nominal coverage, and the bootstrap's false-disagreement rate is measurably higher at fixed power.
A second strong competitor calibrates the apparatus once, corrects the settings, and then performs point-estimate tomography. It receives the same calibration data our priors use. We beat it if residual drift after calibration still flips verdicts that our profiled intervals flag as uncertain.
Two ablations isolate the key ingredients. The first removes the frame-rotation nuisance parameters and keeps only the loss and dark-count ones; this tests whether unitary misalignment drives the measure-specific asymmetry, as we predict. The second drops the orthogonal second channel; this tests whether the cross-check adds rejection power beyond profiling alone.
Why not just…?
Why not just collect more photons? More counts shrink only the statistical part of the error. A waveplate offset biases every count in the same direction no matter how many are collected. The 217-day Daya Bay dataset still needed an independent detection channel and a 50 cm cut between the two signals of each event, which rejected 98% of accidental coincidences, to control backgrounds; volume alone was not enough. Our simulations will show the bootstrap interval shrinking toward a biased value as counts grow, while the true margin stays outside it.
Why not just adopt the relative entropy of coherence as the single standard? Maximal coherence in a generic basis shows that complex Hadamard bases maximize it universally. Quantifying Resources in General Resource Theory with Catalysts gives the regularized relative entropy an operational meaning as the asymptotic cost of erasing a resource. Choosing one measure, however, does not remove systematic bias; it only hides how that bias depends on the measure. It would also discard the $l_1$-certified advantages reported by Ding et al. Our analysis reports which measure's verdict is most robust to the actual apparatus errors. If relative entropy turns out most robust, that is a finding in its favor that we would state plainly.
Why not just use set coherence, which is basis-independent? Set coherence (Set Coherence: Basis-Independent Quantification of Quantum Coherence) quantifies the coherence of a collection of states or measurements, not of an individual steered state. It is also developed mainly for qubits, and its robustness quantities model uniform noise rather than structured calibration error. It cannot answer the NAQC question as posed. We use it only as a design source for the second, non-projective tomography channel.
Why not just do full self-calibrating tomography of the measurement apparatus first? This is the most serious alternative. If calibration were perfect and stable, the residual systematic would vanish and the naive baseline would be adequate. Our second baseline is exactly this approach. The experiments settle the question by measuring how much drift remains after calibration over the timescale of a run. If the residual flips no verdicts, we will report that the reported disagreements are physical and that this approach suffices.
Proposed Experiments
The first study is simulation-based and needs no new hardware. We model the Ding et al. apparatus, using 28 Bell-diagonal states chosen to match their fidelity regime and placed densely near the NAQC thresholds of the $l_1$-norm, relative entropy, and skew information. We inject calibration offsets over a realistic grid. For each method we measure the verdict-flip rate, the interval coverage of the true margin, the false-disagreement rate under boundary-null pseudo-experiments, and the power to detect true disagreements. The decisive check is a dose-response curve. If the width of the disagreement region grows linearly with the misrotation angle, and differently for $l_1$ than for relative entropy, the mechanism is confirmed. If the width stays flat, systematics do not drive the disagreement, and the published conclusions stand on firmer ground.
The second study requests the raw counts behind the published states, or remeasures with the same type of setup, adding deliberate waveplate offsets and the second tomography channel. The decisive check asks whether each state that Ding et al. classified as $l_1$-yes and relative-entropy-no keeps that classification across both channels and across the profiled systematic band.
The third study applies the same pipeline to the non-Markovianity simulator. It tests whether the backflow signature at $t\approx178\lambda_0$ in the ancilla-assisted measure remains significant once frame misrotation and the uncertainty in cavity transmission are profiled. It also tests whether the ordinary measures' spurious rises grow or shrink under the same model.
References
- Experimental investigation of the nonlocal advantage of quantum coherence
- Robust resonant anomaly detection with NPLM
- Independent measurement of the neutrino mixing angle<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>θ</mml:mi></mml:mrow><mml:mrow><mml:mn>13</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>via neutron capture on hydrogen at Daya Bay
- Set Coherence: Basis-Independent Quantification of Quantum Coherence
- 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
- Detecting non-Markovianity via quantified coherence: theory and experiments
- Maximal coherence in a generic basis
- Quantifying Resources in General Resource Theory with Catalysts