Adaptive Conformal Prediction for Quantum Machine Learning

AI summary

Overview

Research area: Uncertainty quantification for quantum machine learning, sitting at the intersection of conformal prediction (distribution-free statistical guarantees) and noisy intermediate-scale quantum (NISQ) hardware.

Technical level: Intermediate. The core insight is intuitive, but the paper assumes familiarity with conformal prediction theory, density-matrix formalism, and quantum noise channels.

Scope: The paper formalises how time-varying hardware noise breaks the exchangeability assumptions underpinning quantum conformal prediction, and introduces an adaptive recalibration algorithm (AQCP) that restores coverage guarantees under arbitrary noise drift.

What This Paper Is About

Quantum machine learning models run on real hardware that drifts: gate fidelities and measurement errors change hour to hour, degrading predictions in ways that are hard to characterise. Conformal prediction promises prediction sets that contain the true answer with a user-specified probability, but its guarantees rely on the statistical assumption that calibration and test data are exchangeable. This paper shows that time-varying quantum noise silently violates that assumption even when the underlying data are perfectly exchangeable, and proposes a fix by adapting the miscoverage target on the fly.

Key Contributions

1. Formalisation of time dependence: A theoretical framework demonstrating that non-stationary hardware noise invalidates the exchangeability of conformity scores, even when the calibration and test feature-target pairs are themselves exchangeable. The scores depend on shot timestamps, and those timestamps carry distributional information.

2. Adaptive Quantum Conformal Prediction (AQCP): An extension of Adaptive Conformal Inference to the quantum setting, using score functions defined over multisets of measurement shots drawn from an implicit quantum output distribution. The algorithm recalibrates the effective miscoverage level after each test point.

3. Theoretical guarantees for the quantum setting: An asymptotic almost-sure convergence result for average coverage under arbitrary hardware noise (no exchangeability required), plus a finite-sample bound adapted from Gibb & Candès that transfers the hidden-Markov-model argument to the quantum case.

4. Hardware evaluation and score-function analysis: Experimental validation on an IBM quantum processor comparing AQCP against quantum conformal prediction, together with a comparison of several sample-based score functions by the average prediction-set size they induce.

Main Findings

• Non-stationary noise breaks exchangeability despite exchangeable data: The paper shows that when a score function takes the shot multiset as an additional input, the augmented observations inherit a time dependence from the noise channel. Since timestamps and noise states are correlated, the resulting scores are not exchangeable, and the classical rank argument behind conformal coverage no longer applies.

• AQCP attains target coverage on real hardware: On an IBM quantum processor, AQCP consistently achieved the requested coverage level, matching or exceeding the performance of standard quantum conformal prediction.

• AQCP is more stable than QCP: Across the test stream, AQCP exhibited less volatility in empirical coverage than quantum conformal prediction, at the cost of requiring labels for each test point to be revealed sequentially.

• Asymptotic average coverage holds without distributional assumptions: The bound

  |(1/N) Σ err_i − α| ≤ (max{α₁, 1 − α₁} + γ) / (Nγ)

  depends only on the step size γ and the number of test points N, and does not depend on exchangeability or the form of the noise drift. As N grows, the average miscoverage converges almost surely to α.

• Step size is the central trade-off: Larger γ gives faster adaptation to shifts but noisier coverage; smaller γ gives smoother coverage but slower response. Setting γ = 0 recovers standard QCP with an updating calibration set.

• The adaptive parameter stays bounded: The update rule keeps α_i within [−γ, 1 + γ] with probability one, which is what enables the finite-sample and asymptotic guarantees.

• Finite-sample behaviour is expected but not exactly justified: The Theorem 2 bound assumes the score process is governed by a hidden Markov chain with a non-zero spectral gap, assumptions unlikely to hold exactly on quantum hardware, but the authors argue the bound is still broadly representative of empirical behaviour.

Methodology in Plain English

The authors begin with the standard split conformal prediction recipe: hold out a calibration set, score each calibration point with a model-dependent function, and use the empirical quantile of those scores as a threshold for building prediction sets at test time. This recipe's validity rests on exchangeability — the property that permuting the order of data points does not change their joint distribution.

They then model a quantum classifier as a parametrised quantum circuit whose unitary operations are interleaved with time-dependent noise channels, indexed by the effective execution time of each shot. Because each measurement shot carries its own timestamp, the output distribution of the circuit can shift from shot to shot. A score function must therefore take the shot multiset as an extra argument, and this makes the score itself dependent on time. Even if the features and labels are drawn i.i.d., the scores are not exchangeable.

To address this, they borrow Adaptive Conformal Inference: instead of fixing the miscoverage level α, they maintain a running estimate α_i that is nudged upward when the previous prediction set missed the true label and downward when it covered it, with step size γ controlling responsiveness. The prediction set at each step uses the empirical (1 − α_i)-quantile of all scores seen so far, including previously revealed test points. Because the update depends only on observed errors, no assumption about the structure of the noise drift is needed.

Finally, they run the algorithm on an IBM quantum processor with hardware-efficient ansatz circuits and angle encoding, comparing the achieved empirical coverage against the target, and comparing the average prediction-set sizes induced by different k-nearest-neighbour style score functions.

Why This Matters

Impact on research. This paper identifies a specific, previously under-examined failure mode of conformity guarantees in quantum machine learning: not distribution shift in the data, but drift in the hardware itself. It bridges two literatures — adaptive conformal inference from classical statistics and quantum conformal prediction — and supplies both the negative result (exchangeability fails) and the constructive remedy (AQCP). It also gives researchers a template for reasoning about score functions that depend on stochastic measurement samples rather than deterministic model outputs.

Real-world applications.

• Drug discovery and molecular property prediction, where quantum classifiers may eventually predict binding affinities or reaction outcomes, and where an overly confident wrong prediction carries substantial cost.
• Financial risk modelling, where quantum kernels have been explored for portfolio and derivative pricing tasks and where calibrated uncertainty is a regulatory concern.
• Materials science and quantum chemistry surrogates, where a model's prediction set can guide which candidate structures are worth expensive classical simulation.
• Clinical decision support, where the downstream consumer of a model needs a set of plausible diagnoses rather than a single label, and where hardware drift during a deployment window cannot be assumed away.

Industry relevance. Cloud quantum providers such as IBM, Google, and Rigetti recalibrate their devices on hourly or daily schedules, and users submit jobs to a queue with unpredictable latency. Any deployed quantum inference service therefore inherits exactly the non-stationary noise structure this paper models. AQCP offers a monitoring-and-correction layer that a provider or an end user could bolt onto an existing quantum classifier with minimal overhead, and it requires no changes to the underlying circuit or training procedure.

Future Directions

• Tighter finite-sample theory. Theorem 2 imports assumptions from hidden-Markov-model analysis that do not match quantum hardware behaviour. A bound tailored to realistic drift processes — for instance drift driven by calibration cycles or cosmic-ray events — would strengthen the practical case.

• Optimal step-size selection. The paper treats γ as a hyperparameter set by the user. Adaptive or data-driven schemes for choosing γ online, balancing responsiveness against coverage volatility, remain open.

• Comparison with alternative shift-robust frameworks. The authors note in an appendix that the framework of Foygel Barber et al. accommodates arbitrary distribution shift with finite-sample guarantees, but would require quantifying the total variation distance induced by the noise drift. Closing that gap is a natural next step.

• Extension to other quantum learning paradigms. The current analysis targets classical-data quantum-processing with angle encoding. Whether the same non-exchangeability argument and AQCP remedy extend to quantum kernel methods, quantum reinforcement learning, or fully quantum data remains to be shown.

• Scalability and shot budget. AQCP requires sequential label revelation, which precludes batch inference. Studying how many shots M are needed per test point to keep the score estimate stable, and how that interacts with prediction-set size, is a practical open question.

Target Audience

Quantum machine learning researchers and practitioners who already understand conformal prediction at a working level and want to deploy quantum classifiers with defensible uncertainty estimates; statisticians working on conformal methods under distribution shift who are curious about hardware-induced non-exchangeability; and quantum computing engineers at hardware providers or cloud platforms who need a lightweight statistical wrapper to make downstream QML services trustworthy despite device drift. Readers without prior exposure to density matrices, Kraus operators, or conformal coverage guarantees will need to consult the background sections carefully, but the central argument is accessible to anyone comfortable with the idea of a calibration set.

Authors’ abstract

Quantum machine learning seeks to leverage quantum computers to improve upon classical machine learning algorithms. Currently, robust uncertainty quantification methods remain underdeveloped in the quantum domain, despite the critical need for reliable and trustworthy predictions. Recent work has introduced quantum conformal prediction, a framework that produces prediction sets that are guaranteed to contain the true outcome with a user-specified probability. In this work, we formalise how the time-varying noise inherent in quantum processors can undermine conformal guarantees, even when calibration and test data are exchangeable. To address this challenge, we draw on Adaptive Conformal Inference, a method which maintains validity over time via repeated recalibration. We introduce Adaptive Quantum Conformal Prediction (AQCP), an algorithm which provides asymptotic average coverage guarantees under arbitrary hardware noise conditions. Empirical studies on an IBM quantum processor demonstrate that AQCP achieves the target coverage level and exhibits greater stability than quantum conformal prediction.