A quantum classifier trained on a perfect simulator loses accuracy on real hardware for two reasons that look identical from the outside. The device itself may have drifted: its error rates differ from the ones it was trained under. Or the incoming data may have changed: different class proportions, different input styles. Both reasons move the numbers the circuit outputs. The most effective current repair for the hardware gap re-standardizes those outputs using statistics from the current batch of inputs. That repair cannot tell whether the outputs moved because the chip changed or because the data changed. When the data has genuinely changed, it erases the very signal the classifier needs.
We propose to measure the hardware's distortion separately from the data. A small, fixed set of reference inputs, called probes, is run both on the simulator and on the device at each calibration cycle. The probes come from the training distribution, so their ideal outputs are known exactly. Fitting the device-to-simulator map on probes alone gives a noise correction that does not depend on whatever the target data looks like. Any shift that remains after this correction can then be attributed to the data and handled by a domain-adaptation step built for that job. The claim is that this separation helps most when both shifts happen at once, and that this compound regime is where batch-statistics normalization breaks.
Core Claim
The motivating observation comes from RoQNN: Noise-Aware Training for Robust Quantum Neural Networks, the paper behind QuantumNAT. It found that hardware noise acts approximately as a per-output scaling and shift on measured expectation values (the averaged qubit readouts that serve as the network's outputs). It corrects this by standardizing each output across the batch, injecting random Pauli errors during training, and quantizing outputs. With these steps it reaches 80% on 4-class MNIST on real IBMQ machines.
Write the affine model as $y_{hw}(x) \approx a \odot y_{sim}(x) + b$. Here $y_{sim}(x)$ is the vector of ideal expectation values for input $x$, $y_{hw}(x)$ is what the device returns, and $a, b$ are per-output scale and offset vectors set by the device's current noise. Batch standardization subtracts the batch mean $a \odot \mu_T + b$, where $\mu_T$ is the mean ideal output over the target data.
If the target domain has a different class balance than the source, then $\mu_T$ differs from the source mean $\mu_S$ for purely semantic reasons. Standardization removes that difference as if it were noise, and it pushes decision scores toward the source-trained boundary under the wrong prior. Estimating $a$ and $b$ instead from probes $p_1,\dots,p_K$ gives
$$ \hat a, \hat b = \arg\min_{a,b} \sum_{k=1}^{K} \lVert y_{hw}(p_k) - a \odot y_{sim}(p_k) - b \rVert^2 . $$
This fit never touches target data, so it cannot absorb $\mu_T$. We invert it, $\tilde y = (y_{hw} - \hat b)/\hat a$, and pass $\tilde y$ to a separate domain-adaptation step.
The idea is simple, and it is only useful if two things hold. First, the affine noise map must be approximately input-independent, so that probes from the source transfer to target inputs. Second, the cost of the probes must be small compared with the alternatives. We test both directly.
Relation to Prior Work
Three lines of work address the hardware gap with static datasets. QuantumNAT, described above, normalizes noise away using batch statistics. QuantumNAS: Noise-Adaptive Search for Robust Quantum Circuits searches circuit architectures and qubit placements matched to one chip's error profile, reaching over 95% on 2-class and 85% on 4-class MNIST on IBMQ, and it notes that calibration drift forces re-search. QMLP builds error-tolerant gate primitives that hold 63–67% on 10-class MNIST under bit- and phase-flip noise. All three evaluate on a fixed test distribution. This proposal supplies what they omit, a shift in the data. We reuse QuantumNAT's affine noise finding as our method and its pipeline as our main baseline. QuantumNAS and QMLP play the role of upstream architectures that our correction sits on top of.
On the domain side, Transfer Learning Analysis of Variational Quantum Circuits gives a one-step analytical parameter update (QVA) for data shift. On a 1-qubit Two Moons task it recovers accuracy from 49.8% to 77.2%, matching 17 epochs of gradient descent, but it assumes noise-free simulation and says hardware is future work. That paper supplies our second-stage adapter. Our contribution is the noise-correction step that makes its noise-free assumption approximately true on hardware.
Quantum transfer component analysis for domain adaptation proposes quantum domain-invariant projections but runs no noisy experiments. A comparative analysis and noise robustness evaluation in quantum neural networks constrains the design. Its finding that classical ResNet features fed into a quantum head collapse to about 20% on 4-class MNIST warns that domain adaptation done purely before encoding does not carry over cleanly. QEEGNet motivates a real cross-dataset use case, but it was validated only in simulation.
The classical domain-sensitivity literature supplies the conceptual separation. DAFE keeps domain-specific offsets apart from base parameters so they can be swapped independently. We do the same for noise offsets versus domain offsets, estimating each from its own source of evidence.
Research Program
The work proceeds in three stages. The first stage tests whether the assumption holds. Using TorchQuantum's simulator with noise models built from real IBMQ calibration data (as QMLP did with ibmq_quito), and real devices where access allows, we measure how well a single $(a, b)$ fitted on source probes predicts device outputs on shifted target inputs. We do this as circuit depth and entanglement grow. If the remaining error grows sharply on target inputs, the affine noise map depends on the input. In that case we would move to a probe-conditioned correction indexed by a coarse input feature, and report the depth beyond which separation fails.
The second stage builds the pipeline. Probes are chosen once from source training data to span the output range; we compare a stratified selection against a random one. They are re-run whenever the device recalibrates, and the affine correction is refitted. Corrected outputs then go to one of two domain steps. The first is an unsupervised class-prior re-estimation by expectation-maximization on target outputs. The second is the QVA one-shot update when a few labeled target points exist.
The third stage maps when the separation matters. It varies the size of the noise shift and the data shift independently on a grid. The aim is a picture of where batch normalization suffices, where probes win, and where neither works. This also addresses the open question in the file about when noise stops acting like a regularizer and starts destroying features out of domain.
Baselines
The do-nothing reference is the source-trained circuit run on the shifted device with shifted data and no correction. It gives the size of the compound gap.
The strongest existing competitor is the full QuantumNAT pipeline from RoQNN: batch standardization of outputs, noise injection drawn from device error rates during training, and output quantization. Its standardization uses statistics from the target batch. It is the strongest because it is the only cited method that reports large accuracy recovery on real hardware: 80% on 4-class MNIST on real IBMQ machines, and up to 43% improvement over an uncorrected baseline. We run it on the same circuits with the same training data, the same noise-injection schedule, and the same target batches, and we tune its batch size over the same grid. Beating it means higher target accuracy under compound shift at equal shot count, counting probe shots against our method. It also means no loss when only the device has drifted.
A second baseline is QuantumNAS-searched circuits, which report 85% on 4-class MNIST on IBMQ, re-searched for the drifted device. This is the "redo the specialization" option. We give it the same total device-shot budget.
Two ablations isolate the key ingredient. In the first, the affine correction is fitted on unlabeled target inputs rather than source probes, which shows whether probe provenance is what matters. In the second, probe calibration runs with no domain step, which separates the noise correction's contribution from the adapter's. An oracle that uses ideal simulator outputs on target inputs bounds what any noise correction can reach.
Why not just…?
Why not just keep QuantumNAT's batch normalization, since it already handles noise as a scale and shift? Under pure device drift it should do about as well as probes, and we expect that. The objection fails only where the target batch mean carries information. The decisive test is a label-prior shift with no device drift. If batch normalization does not lose accuracy there relative to the uncorrected circuit, the central argument of this proposal is wrong, and we will report it as such.
Why not just run standard error mitigation such as zero-noise extrapolation or readout correction? These estimate ideal expectations per input without using batch statistics. In principle they avoid the confound, which makes this a real threat. Their cost is multiple noise-amplified executions per input, whereas probes are a one-time cost per calibration cycle. RoQNN reports that inference-only mitigation left large gaps for variational classifiers, but it did not test under data shift. We therefore compare against readout mitigation and zero-noise extrapolation at matched total shots. If extrapolation matches probes at equal cost, the contribution shrinks to a cheaper approximation, and we will say so.
Why not just fine-tune on the device with a few labeled target examples? This needs labels and many device-run gradient steps. Transfer Learning Analysis of Variational Quantum Circuits shows gradient descent needed 17 epochs to match a one-step update, even noise-free. Evolutionary Quantum Architecture Search notes that noise shifts and flattens the loss landscape, which makes on-device optimization harder. We include on-device fine-tuning at matched shots as a baseline rather than assuming it loses.
Why not just adapt classically before encoding? That approach is complementary. It cannot remove shifts that the device introduces after encoding, and the about 20% QTL collapse in the comparative analysis suggests classical alignment alone does not transfer well. We stack it under both our method and QuantumNAT so the comparison stays fair.
Proposed Experiments
The task suite reuses the file's quantum ML benchmarks with shifts built in. These are 2-class and 4-class MNIST, Fashion-MNIST, and CIFAR-2 at the downsampled resolutions QuantumNAT used, plus Two Moons with the shift protocol of the QVA paper. Each source task gets three target variants: class-prior shift, where the target class proportions are skewed; covariate shift, using rotated or noise-corrupted inputs; and both together. Each is crossed with three device conditions: the training calibration, a later calibration of the same device, and a different device's noise model. Reported numbers for anchoring are 80% (QuantumNAT, 4-class MNIST, real IBMQ), 85% (QuantumNAS, 4-class MNIST, real IBMQ), and 49.8% to 77.2% (QVA, Two Moons, noise-free). No reported result found on Lacuna for any of these methods under input domain shift, so all compound-shift numbers will be new.
Metrics are target accuracy, accuracy recovered as a fraction of the oracle gap, total device shots including probes, and the probe-fit error on target inputs. The decisive checks follow from the claim. Under device-only drift, probes and batch normalization should tie. Under prior-only shift, batch normalization should fall below the uncorrected circuit while probes should not. Under compound shift, probe calibration plus a domain step should beat QuantumNAT at equal shots. Fitting the correction on target data, the first ablation, should recover QuantumNAT's failure.
A last sweep over circuit depth reports where the affine assumption breaks. That depth sets the honest limit on where this separation can be used.
References
- QuantumNAT: Quantum Noise-Aware Training with Noise Injection, Quantization and Normalization
- QuantumNAS: Noise-Adaptive Search for Robust Quantum Circuits
- QMLP: An Error-Tolerant Nonlinear Quantum MLP Architecture using Parameterized Two-Qubit Gates
- Transfer Learning Analysis of Variational Quantum Circuits
- Quantum transfer component analysis for domain adaptation
- A comparative analysis and noise robustness evaluation in quantum neural networks
- Exploring the Potential of QEEGNet for Cross-task and Cross-dataset Electroencephalography Encoding with Quantum Machine Learning
- Unsupervised Domain Adaptation for Neural Machine Translation with Domain-Aware Feature Embeddings
- Evolutionary quantum architecture search for parametrized quantum circuits