What Hardware Noise Does to Interference: A Dalitz-Plot Probe of the Simulation-to-Hardware Gap

Quantum circuits trained or checked in an ideal simulator often fail when run on a real noisy quantum processor. Nearly everything we know about this gap comes from classification tasks such as low-resolution MNIST, where the only question is whether the predicted label survives. Many of the quantum algorithms people actually care about are different. They output an interference pattern, a probability built from adding complex amplitudes and squaring them, and the scientifically meaningful quantities are the relative phases inside that sum. We propose to study the gap on a problem where those phases are the whole point and the correct answer is known exactly: the Dalitz plot. This is particle physics' standard picture of a three-body decay, and it shows how several short-lived intermediate states ("resonances") interfere across a two-dimensional space of pair masses.

The core idea is simple. A two- or three-qubit circuit can compute the Dalitz intensity $|\sum_j c_j A_j(x)|^2$ at any point $x$ as the probability of a single measurement outcome. Under that construction, standard hardware noise channels turn into familiar physics systematics. Depolarizing noise, which randomly scrambles the qubit state, turns the coherent sum of resonances into a partly incoherent sum, so it erases interference without moving any peaks. Coherent over-rotations on the register that labels the resonance channels add fake relative phases, which look exactly like the physical "strong phases" that experiments measure. Because this mapping is exact, the noise can be absorbed as a few extra nuisance parameters in the physics fit and pinned down with calibration circuits whose answer is known. That replaces generic output normalization, which breaks probability conservation, with a correction that respects physics. We do not claim a quantum speedup for Dalitz fitting. The claim is about measurement: how real noise biases phases relative to magnitudes, and how to remove that bias.

Core Claim

Write a channel register of $d = 2^n$ basis states $|j\rangle$, one per resonance. A circuit prepares the parameter state $|c\rangle = \sum_j c_j|j\rangle$, where the $c_j$ are complex isobar coefficients normalized to unit length. It then applies the inverse of a unitary that loads the normalized amplitude vector $a(x) = A(x)/\|A(x)\|$ at kinematic point $x$. Here $A_j(x)$ is the Breit-Wigner lineshape of resonance $j$ including angular and barrier factors, as in the isobar formalism of Aubert et al.. The probability of reading all zeros is $p_0(x) = |\langle a(x)|c\rangle|^2$, so the physical intensity is $I(x) = \|A(x)\|^2 p_0(x)$. With four to eight channels this needs only two or three qubits and roughly a dozen two-qubit gates, which fits current devices comfortably.

Suppose the hardware acts on this circuit as a global depolarizing channel with survival probability $\lambda$, together with a per-channel coherent phase error $\phi_j$. The measured intensity then becomes

$$\tilde I(x) = \lambda\,\Big|\sum_j c_j e^{i\phi_j} A_j(x)\Big|^2 + \frac{1-\lambda}{d}\sum_j |A_j(x)|^2 .$$

The first term is the true coherent intensity with fake phases added. The second term is the incoherent sum of resonance intensities, the shape a physicist would see if interference were switched off. Two predictions follow directly. Depolarizing-type noise biases fit fractions and interference fractions but leaves fitted phases mostly unbiased. Coherent channel-dependent errors bias phases one for one and leave rates untouched. A noise component that acts identically on every channel is a global phase and drops out of the observable completely.

This decomposition is only exact for idealized channels. On real hardware, crosstalk and spatially uneven decoherence (Medina Bandic et al.) will add terms it does not capture. Measuring how much they add is part of the study.

Research Program

The work has three parts.

The first part is the probe itself. We build the intensity circuit above, with amplitude loading through standard two- and three-qubit state preparation. We generate toy Dalitz models with exactly known parameters, patterned after published analyses, so every fitted phase and magnitude has a known true value. The patterns are a $D^0 \to K_S^0\pi^+\pi^-$-like model with a $\rho$, a $K^*$ and a broad S-wave (Aubert et al.), a $D^0 \to K^-\pi^+\eta$-like model (Dalitz analysis of D0 → K−π+η), and a sparse $D^0 \to K_S^0\eta\pi^0$-like model sized to the 155-candidate CLEO sample (Rubin et al.). Fitting every event without binning would need one circuit per event, which no device can afford. We therefore fit binned Dalitz plots, following the binned approach behind the model-independent determination of γ (First model-independent determination of the relative strong phase between D 0 and D ¯ 0 → K S 0 π + π − and its impact on the CKM angle γ / ϕ 3 measurement). The circuit measures the intensity of each bin, and the fit is a Poisson likelihood over bins.

The second part is a noise-aware fit. The isobar fit gains nuisance parameters $\lambda$ and $\phi_j$, plus a readout-error matrix. These are constrained by calibration runs that execute the same circuits with one-hot parameter states $|c\rangle = |j\rangle$. For those runs the true intensity is $|A_j(x)|^2$ with no interference, so any interference-like or floor-like structure that appears is caused by the hardware. The calibrations are what keep a real physical non-resonant component from being confused with depolarizing noise. Without them the two are degenerate under the model above.

The third part is shot allocation. Measurement shots are the scarce resource. Following the idea behind iCANS (Kübler et al.), which sets the shot count for each gradient component adaptively, we give shots to each bin in proportion to its estimated Fisher information about the relative phases. That information is concentrated where resonance bands overlap. Every method gets the same total shot budget.

The first part motivates the work and the second and third parts supply the method. The motivation comes from Li Ding et al., who show that noise reshapes cost landscapes, and from the larger-than-60% simulation-to-hardware drop reported by RoQNN/QuantumNAT. QuantumNAT also constrains the design. Its fix shifts and rescales outputs to zero mean and unit variance and then quantizes them. That produces negative "probabilities" and erases narrow interference fringes, so it cannot be applied directly to an intensity. Our nuisance model keeps QuantumNAT's insight that noise acts roughly as an affine map on outputs, but expresses that map inside the physics model, where the result stays a valid probability. QMLP argues that entangling data loading is fragile under noise. Our loading is entangling by necessity, which is a reason to measure its fragility directly. Error Interference in Quantum Simulation suggests that some coherent errors cancel. Our framework predicts which ones will: those that act uniformly across channels. The ablation in the Baselines section tests this.

The HEP-quantum literature, Delgado et al., the 2024 roadmap and the 10-qubit LHC classifier, treats only event classification, tracking and shower generation. None of it measures phases on hardware.

Baselines

The reference point for statistics is a classical isobar fit using exact intensities, applied to the same binned data. It sets the irreducible statistical error on each phase and magnitude, and every hardware result is quoted as bias divided by this error, which is a pull. No quantum method is expected to beat it. The question is how close each method gets.

The "do nothing" baseline runs the intensity circuits on hardware and fits them with the plain isobar model, with no mitigation. Its pulls show the raw size of the gap. We expect phase pulls of several standard deviations wherever coherent errors depend on the channel.

The standard-mitigation baseline adds readout-error correction and zero-noise extrapolation, which estimates the noiseless answer by running at deliberately amplified noise and extrapolating back to zero. It uses the same total shot budget, with the extra shots for amplified-noise runs counted against that budget. This is the strongest generic competitor because it needs no physics knowledge. No reported result found on Lacuna for mitigated phase extraction from interference distributions.

The QuantumNAT-style baseline trains or fits under simulated noise injected to match device error rates, and replaces the zero-mean normalization with renormalization to unit integral so the comparison is legitimate. On classification, RoQNN/QuantumNAT reports accuracy gains of up to 43% and a noise-free-to-hardware drop above 60% on MNIST-4. No reported result found on Lacuna for this approach on continuous densities or phases.

Ablations isolate the key ingredient. One fit uses only the floor nuisance $\lambda$. One uses only the phase nuisances $\phi_j$. One uses both nuisances but no one-hot calibrations. A further ablation reorders the channel encoding so that dominant hardware coherent errors act as a global phase. The proposal counts as beating the field if, at equal shots on real hardware, all phase and fit-fraction pulls have a mean within ±0.3 and a width near 1 across the three models, while the standard-mitigation baseline still shows at least one phase pull with magnitude above 2.

Why not just…?

Why not just fit Dalitz plots on GPUs? That is correct for physics production, and we do not compete with it. The classical fit is our reference. The Dalitz problem is used here because it is the cleanest available interference observable with exact truth and physically meaningful phases. The deliverable is a calibrated understanding of how hardware corrupts phases, which carries over to any quantum algorithm whose output is a squared amplitude. If the reader's question is purely whether quantum hardware helps Dalitz analysis today, the file supports no such claim. As Meyer et al. note, hardware remains too slow and noisy for high-volume training.

Why not just use generic error mitigation such as zero-noise extrapolation? This is a real threat, and it is the main baseline. The difference is how each method treats noise. Extrapolation assumes noise scales smoothly and spends shots on the extra amplified-noise runs. The nuisance fit spends shots on calibrations whose physics meaning is known. If extrapolation reaches the same pulls at equal shot budget, we will report that the physics-structured model adds nothing beyond interpretability. The prediction that coherent phase offsets do not shrink much under noise amplification gives a specific place where extrapolation should fail.

Why not just simulate with uniform depolarizing noise? Under our model, uniform depolarizing noise produces only the incoherent-floor term. It therefore predicts unbiased phases and would miss the dominant failure mode. Medina Bandic et al. and Chu et al. document that uniform models leave out crosstalk and spatial unevenness. Our comparison of uniform simulation, device-calibrated simulation and real hardware measures exactly how much each simulator underpredicts phase bias.

Why not just a generic quantum generative model benchmark? Generic Born-machine benchmarks score the overall distribution, for example with divergences. They cannot tell a lost interference term apart from a shifted phase. The Dalitz parameterization separates these by construction, which is the entire reason to use it.

Proposed Experiments

We will run the three toy models in four settings: noiseless simulation, uniform-depolarizing simulation, device-calibrated simulation, and an IBM-class superconducting device. Each model is fitted with 20 to 50 bins and matched shot budgets.

The first decisive check tests the core prediction. We sweep injected coherent channel-phase errors and depolarizing strength separately in simulation. Phase pulls should track the coherent errors one for one and stay flat as depolarizing strength rises. Fit fractions should do the opposite. If both quantities are biased by both noise types, the mapping is wrong.

The second check tests identifiability. We add a genuine flat non-resonant physical amplitude to the toy model. We then show that without one-hot calibrations it cannot be told apart from $\lambda$, and that with calibrations it is recovered without bias.

The third check is the hardware comparison against all baselines, using the pull criterion stated in the Baselines section. A drift test supports it: calibrations and physics runs are separated by hours, and we track how fast the coherent nuisances go stale.

References

  1. Measurement of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>cos</mml:mi><mml:mo></mml:mo><mml:mn>2</mml:mn><mml:mi>β</mml:mi></mml:math>in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>B</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>D</mml:mi><mml:mrow><mml:mo stretchy="false">(</mml:mo><mml:mo>*</mml:mo><mml:mo stretchy="false">)</mml:mo></mml:mrow></mml:msup><mml:msup><mml:mi>h</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math>Decays with a Time-Dependent Dalitz Plot Analysis of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>D</mml:mi><mml:mo>→</mml:mo><mml:msubsup><mml:mi>K</mml:mi><mml:mi>s</mml:mi><mml:mn>0</mml:mn></mml:msubsup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>−</mml:mo></mml:msup></mml:math>
  2. Profiling quantum circuits for their efficient execution on single- and multi-core architectures
  3. Constraints on the CKM angleγinB0→D¯0K*0andB0→D0K*0from a Dalitz analysis ofD0andD¯0decays toKSπ+π−
  4. Dalitz analysis of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo stretchy="false">→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mi>η</mml:mi></mml:math> decays at Belle
  5. First Observation and Dalitz Analysis of the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:msubsup><mml:mi>K</mml:mi><mml:mi>S</mml:mi><mml:mn>0</mml:mn></mml:msubsup><mml:mi>η</mml:mi><mml:msup><mml:mi>π</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math>Decay
  6. First model-independent determination of the relative strong phase between<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math>and<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mover accent="true"><mml:mi>D</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:msubsup><mml:mi>K</mml:mi><mml:mi>S</mml:mi><mml:mn>0</mml:mn></mml:msubsup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>−</mml:mo></mml:msup></mml:math>and its impact on the CKM angle<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>γ</mml:mi><mml:mo>/</mml:mo><mml:msub><mml:mi>ϕ</mml:mi><mml:mn>3</mml:mn></mml:msub></mml:math>measurement
  7. An Adaptive Optimizer for Measurement-Frugal Variational Algorithms
  8. Evolutionary quantum architecture search for parametrized quantum circuits
  9. QuantumNAT: Quantum Noise-Aware Training with Noise Injection, Quantization and Normalization
  10. QMLP: An Error-Tolerant Nonlinear Quantum MLP Architecture using Parameterized Two-Qubit Gates
  11. Error Interference in Quantum Simulation
  12. Quantum computing for data analysis in high energy physics
  13. Quantum Computing for High-Energy Physics: State of the Art and Challenges
  14. Application of quantum machine learning using the quantum variational classifier method to high energy physics analysis at the LHC on IBM quantum computer simulator and hardware with 10 qubits
  15. Benchmarking Quantum Reinforcement Learning