Neural networks can now reconstruct small quantum states from measurement counts faster and more accurately than classical estimators. Quantum state tomography means recovering the density matrix, the full statistical description of a quantum state, from repeated measurements. Networks trained once on simulated data map counts to a density matrix in about a millisecond. The weak point is the simulator they learn from. Every such network is trained on data from an idealized measurement apparatus with honest sampling noise. A real device then applies a measurement channel that is miscalibrated, drifts between calibrations, and decoheres the qubits while they are being read out. Separately, a line of quantum machine learning work shows that injecting physical noise during training makes quantum classifiers more robust. Nobody has carried that idea over to tomography.
The transfer is not straightforward, and that is what this note is about. In classification the label never changes when noise is injected. In tomography it does, depending on where the noise acts. Noise acting on the state before measurement changes the correct answer, because the decohered state is the thing we want to reconstruct. Noise acting inside the measurement apparatus leaves the correct answer fixed. We propose to train neural reconstructors on a randomized family of measurement-side channels while keeping the label equal to the prepared state. Separately, we sample preparation-side decoherence to build a realistic prior over mixed states. We will test whether this makes reconstructors robust to measurement drift without recalibration, and whether the gain comes from true regularization or merely from wider data coverage.
Core Claim
A reconstructor trained with label-preserving measurement noise, drawn from a physically parameterized family rather than a single calibrated snapshot, will beat three alternatives when the device's measurement channel differs from the training assumption. The first alternative is the same architecture trained on ideal measurements. The second is that architecture fed counts corrected with a calibrated readout matrix. The third is maximum-likelihood estimation with the calibrated detector model. We also claim that randomizing preparation-side noise helps only as a prior over mixed states. Randomizing measurement-side noise helps in a second way: it makes the network insensitive to nuisance distortions.
Concretely, training pairs are generated as
$$ \rho = \mathcal{E}_\theta(\sigma), \qquad f \sim \mathrm{Multinomial}\big(S,\ p(\rho, \mathcal{M}_\phi)\big), \qquad \theta \sim \Pi_{\text{prep}},\ \phi \sim \Pi_{\text{meas}}, $$
where the symbols are as follows. $\sigma$ is a pure target state. $\mathcal{E}_\theta$ is a decoherence channel (phase and amplitude damping with strength $\theta$) that defines the label $\rho$. $\mathcal{M}_\phi$ is a perturbed measurement described by $\phi$, which covers asymmetric readout flips, small basis-rotation errors, and damping during readout. $p$ gives the outcome probabilities. $S$ is the shot count, itself randomized. $f$ is the frequency vector the network sees. The network $g$ is trained to minimize a distance between $g(f)$ and $\rho$, never between $g(f)$ and anything that depends on $\phi$.
Relation to Prior Work
Two lines of tomography work supply the method we build on. DNN-QST maps frequency vectors to a Cholesky factor $L$ with $\rho = LL^\dagger/\mathrm{Tr}(LL^\dagger)$, which guarantees a valid state. It is trained on Haar and Ginibre states with simulated sampling noise and small unitary-rotation errors, and it beats maximum-likelihood and linear-regression estimation at $S=10$ shots. The Quantum-Aware Transformer (QAT) adds cross-attention over the measurement operators and a Bures-distance loss, and it was run on IBM hardware. We reuse both architectures unchanged.
DNN-QST already performs a narrow form of noise augmentation, so our contribution is not "add shot noise." It is three things together. We use a structured, randomized measurement-channel family. We keep labels fixed under that family. We use decoherence-generated mixed states rather than Ginibre states as the label distribution.
The noise-regularization literature motivates the approach. QuantumNAT/RoQNN shows that hardware noise acts roughly as a scale-and-shift on measurement outcomes. It also shows that injecting device-calibrated Pauli errors during training moves classifiers to flatter minima and recovers up to 43% accuracy on real devices. That work applies only to classification, where the label question never arises. Noisy HQNNs constrains the design. It found that phase-type noise can regularize, while depolarizing and amplitude damping above $p\approx 0.4$ destroy performance. So our measurement family must stay in a physically moderate range, and we will map where randomization stops helping.
Khanal et al. show that noise shrinks the effective dimension of quantum models. That result concerns quantum-circuit models, not classical reconstructors. We borrow only the question it raises: does noise reduce the generalization gap? We do not borrow the bound.
Palmieri et al. is the closest competitor. It learns a supervised filter that removes one fixed optical state-preparation-and-measurement (SPAM) error, raising fidelity from 0.82 to 0.91. It needs labelled experimental states from that fixed setup, and it does not address drift. The shadow-based neural methods, NSQST and bootstrapped shadows, fit a separate model to each state rather than training a reusable reconstructor. NSQST notes that noise biases its loss without changing the gradient direction, but restricts itself to pure states. We stay in the reusable, small-system regime, where hardware validation is feasible.
Research Program
We will build a simulator for the generative process above covering 2 to 4 qubits. It will support Pauli and mutually-unbiased-basis measurement schemes, matching the DNN-QST and QAT setups. The measurement family $\Pi_{\text{meas}}$ is fitted as a spread around repeated readout-calibration data collected from IBM devices over several days. This means the randomization covers observed drift, not arbitrary corruption.
We then train DNN-QST and QAT under four regimes. The first uses ideal measurement only. The second adds preparation noise only. The third adds measurement noise only. The fourth, the full proposal, adds both. The shot budget and the optimizer tuning are identical across all four.
A central difficulty is evaluation on hardware, where the true prepared state is unknown. The fidelities QAT reports on hardware are measured against the intended target state, which mixes reconstruction error with preparation error. We will therefore add a reference-free metric: held-out-setting likelihood. The reconstructed $\rho$ must predict outcome distributions for measurement settings the reconstructor never saw, scored under that session's measured readout model. We will also compare against a high-shot reference estimate.
To separate regularization from coverage, we will measure two quantities. The first is the train-test gap on clean in-distribution data. The second is the sharpness of the loss around the trained weights. Measurement-side randomization might lower sharpness and the clean-test gap even when there is no mismatch. If so, it is acting as a regularizer in the QuantumNAT sense. If it helps only under mismatch, it is domain randomization, and we will say so.
Baselines
The naive reference is linear-regression estimation followed by projection onto valid states, per Qi et al.. On ibmq_manila at 1,000 shots, QAT reports mean fidelity 0.9179 for this method and 0.9173 for IBM's built-in tomography. It is run on exactly the same counts as every other method.
The strongest existing method is QAT itself, which reports mean fidelity 0.9948 on ibmq_manila at 1,000 shots. We retrain it on the ideal-measurement regime with its published loss. It receives the same hyperparameter-search budget and training-set size as our variant. Because that specific device may no longer be available, we will rerun both methods on current devices and not rely on the reported number for the head-to-head comparison. For DNN-QST at $S=10$ we found no reported numerical infidelity on Lacuna, so we will reproduce it.
The strongest calibration-based competitor is maximum-likelihood estimation with a detector model measured by readout calibration. We run it twice, once with a calibration taken in the same session and once with a stale calibration taken 24 or more hours earlier. A second calibration-based competitor feeds readout-corrected counts into the ideal-trained QAT.
Isolating ablations come from the regime grid. The key ablation is measurement-only versus preparation-only noise. We also add a variant that randomizes the measurement channel but lets the label follow the measured distribution, which deliberately violates label preservation. It tests whether keeping the label fixed is what matters.
Beating the field means two things. Under stale calibration, our method must exceed every baseline in held-out-setting likelihood and in fidelity to the reference estimate, with bootstrap confidence intervals that do not overlap. Under fresh calibration, it must be no worse than calibrated maximum-likelihood estimation.
Why not just…?
Why not just calibrate the readout and invert it? This is the real threat. Under a fresh calibration, with errors that are purely classical bit-flips, readout inversion plus a clean-trained network may well match our method. We expect it to fall short in two cases. One is coherent basis-rotation errors, which a readout confusion matrix cannot represent. The other is drift after calibration. The fresh-versus-stale comparison settles this directly. If our method only ties even under a stale calibration, the proposal reduces to "recalibrate more often," and we will report that result.
Why not just use dropout and weight decay? Generic regularizers limit the network's capacity but do not encode which input distortions leave the answer fixed. We include a dropout-plus-weight-decay arm, tuned with the same budget as the ideal-trained network. If it matches the measurement-randomized arm under mismatch, the physical structure of the noise adds nothing.
Why not just fine-tune on calibration states from the device, as Palmieri et al. do? Their filter works well for a fixed optical SPAM error, lifting fidelity from 0.82 to 0.91. However, it needs labelled device data each time the channel changes. We compare against fine-tuning on $k$ known product states per session, at several values of $k$, and report the $k$ at which fine-tuning catches up. Robustness without per-session data has value only below that crossover.
Why not just use classical shadows? The file gives no evidence that raw shadows fail at 2 to 4 qubits for full-state reconstruction. The documented failures, from Kokaew et al. and NSQST, concern high-weight observables at larger scale. Shadows are also not immune to readout miscalibration. We include projected shadows (Qin et al.) as a secondary baseline on the same shots.
Proposed Experiments
In simulation, we sweep the gap between the training measurement family and a held-out test channel. The gap is set by scaling the drift beyond the fitted spread and by adding coherent rotation angles. We also sweep $S$ from 10 to 1,000 on 2- to 4-qubit Haar, Ginibre and decoherence-generated states. We plot infidelity and Bures distance against the size of the gap. The decisive simulated check has two parts. The measurement-randomized reconstructor must degrade more slowly with the gap than readout-inversion plus the clean network. The label-violating ablation must do worse than the label-preserving version.
On hardware, we prepare Bell, GHZ and random two-layer states on 2 to 4 qubits. We insert idle delays to produce controlled mixed states, which addresses the untested open-system regime flagged for NSQST. Counts are collected at 1,000 shots across multiple days. Every method is scored with held-out-setting likelihood, with fidelity to a high-shot reference estimate, and, for comparability with QAT, with fidelity to the target state.
The decisive hardware check has two parts. On days when the calibration is stale, our method must beat calibrated maximum-likelihood estimation and ideal-trained QAT. The clean-test generalization gap and the sharpness measurements then tell us whether the gain comes from regularization or from coverage alone. A negative result would still be informative. It would show that, for tomography, physical noise is best handled by calibration rather than by training-time injection. That would bound how far the noise-as-regularizer story extends beyond classification.
References
- Neural networks for quantum state tomography with constrained measurements
- Tomography of Quantum States From Structured Measurements via Quantum-Aware Transformer
- QuantumNAT: Quantum Noise-Aware Training with Noise Injection, Quantization and Normalization
- Noisy HQNNs: A Comprehensive Analysis of Noise Robustness in Hybrid Quantum Neural Networks
- Data-dependent generalization bounds for parameterized quantum models under noise
- Experimental neural network enhanced quantum tomography
- Neural-shadow quantum state tomography
- Bootstrapping Classical Shadows for Neural Quantum State Tomography
- Quantum State Tomography via Linear Regression Estimation
- Enhancing quantum state reconstruction with structured classical shadows