Provable and scalable quantum Gaussian processes for quantum learning

Abstract: Despite rapid recent advances in quantum machine learning, the field is in many ways stuck. Existing approaches can exhibit serious limitations, and we still lack learning frameworks that are simple, interpretable, scalable, and naturally suited to quantum data. To address this, here we introduce quantum Gaussian processes, a Bayesian framework for learning from quantum systems through priors over unknown quantum transformations. We show that, under suitable conditions, unitary quantum stochastic processes define Gaussian processes, thereby enabling regression, classification, and Bayesian optimization directly on quantum data. The key ingredient in this framework is sufficient knowledge of a quantum process's structure and symmetries to define an informative prior through its corresponding quantum kernel, effectively injecting a strong, physics-informed inductive bias into the learning model. We then prove that matchgate, or free-fermionic, evolutions give rise to provable and scalable quantum Gaussian processes, providing the first family in our framework where the unknown unitary acts non-trivially on all qubits. Finally, we demonstrate accurate long-range extrapolation, phase-diagram learning in many-body systems, and sample-efficient Bayesian optimization in a quantum sensing task. Our results identify quantum Gaussian processes as a promising route toward simpler and more structured forms of quantum learning.

• The paper demonstrates the construction of quantum Gaussian processes (QGPs) from matchgate circuits that yield provable and scalable inference for many-qubit systems.
• It introduces a method using group-theoretic priors and closed-form kernel estimation to maintain predictive accuracy under sparse and noisy data conditions.
• The framework is applied to quantum regression, classification, and Bayesian optimization, outperforming traditional approaches with efficient, resource-aware learning.

Provable and Scalable Quantum Gaussian Processes for Quantum Learning

Introduction and Motivation

Quantum machine learning (QML) has experienced sustained activity but still lacks frameworks that combine interpretability, scalability, and an inductive bias suited to quantum data. Contemporary approaches focus either on quantum algorithms built from quantum linear algebra primitives, which depend crucially on data access models and are susceptible to dequantization, or variational quantum learning with parametrized circuits, which faces issues including barren plateaus, optimization instability, limited interpretability, and a lack of scaling guarantees for large quantum systems.

This work advances a rigorous and structured QML framework by formulating quantum Gaussian processes (QGPs)—nonparametric Bayesian models built from priors derived from physically consistent quantum stochastic processes. The framework encompasses regression, classification, and Bayesian optimization directly on quantum data and, crucially, introduces the first provable and scalable QGP family where the unknown quantum process nontrivially acts on the entire many-qubit system, specifically exploiting matchgate (free-fermionic) evolutions. This direction establishes QGPs as a systematic route toward quantum learning protocols with strong inductive biases and robust generalization from sparse data.

Figure 1: Schematic representation of the QGP framework, integrating structural knowledge of the quantum process, observable, and available data to select a kernel from a QGP toolbox and construct a Bayesian surrogate for quantum learning.

Theoretical Foundations

Quantum Gaussian Processes and Prior Construction

A Gaussian process (GP) is a distribution over random functions specified by mean $\mu(t)$ and covariance (kernel) $\kappa(t, t')$. Given a finite set of training points, noisy observations, and a GP prior, Bayesian conditioning yields posterior predictive means and variances for unmeasured points with closed-form expressions. In the quantum setting, the random function is instantiated as $f_U(t) = \mathrm{Tr}[U\rho(t)U^\dagger O]$, with $U$ an unknown unitary, $\rho(t)$ parameterized quantum states, and $O$ a Hermitian observable. The prior on $U$—often the Haar measure over a subgroup—imposes a distribution over the functions, and, under specific structural constraints, can yield a GP over $f_U$.

The critical technical advances in this work are twofold:

1. Demonstrating provable emergence of QGPs from well-defined quantum processes, especially those with informative group-theoretic structure;
2. Establishing criteria for the scalability of QGPs, i.e., conditions under which all GP kernels and inference quantities are efficiently estimable ($O(\mathrm{poly}(n))$) in the number of qubits $n$.

The work unifies and extends known families of provable QGPs—from few-qubit $\kappa(t, t')$0, $\kappa(t, t')$1, and $\kappa(t, t')$2 groups—to systems-wide matchgate circuits.

Matchgate/Free-Fermionic QGPs: Provable and Scalable Construction

A central result is the formal proof that matchgate circuits (the image of the group $\kappa(t, t')$3 via the Jordan-Wigner transformation) with observables in specific operator modules $\kappa(t, t')$4 produce QGPs for arbitrary system sizes, provided $\kappa(t, t')$5 is kept small (e.g., $\kappa(t, t')$6). The observable $\kappa(t, t')$7 must belong to a specific sector $\kappa(t, t')$8, corresponding to products of $\kappa(t, t')$9 distinct Majorana operators. The QGP kernel is then analytically derived as:

$f_U(t) = \mathrm{Tr}[U\rho(t)U^\dagger O]$0

Scalability follows since these overlaps are measurable with $f_U(t) = \mathrm{Tr}[U\rho(t)U^\dagger O]$1 experimental resources for $f_U(t) = \mathrm{Tr}[U\rho(t)U^\dagger O]$2.

The proof leverages asymptotic Weingarten calculus for moments over $f_U(t) = \mathrm{Tr}[U\rho(t)U^\dagger O]$3 and Brauer algebra expansions, revealing that, under group and observable structure, the random process $f_U(t) = \mathrm{Tr}[U\rho(t)U^\dagger O]$4 is a GP in the large-$f_U(t) = \mathrm{Tr}[U\rho(t)U^\dagger O]$5 limit, and non-pairwise contractions are suppressed.

Empirical Analysis: Regression, Extrapolation, and Classification

Inductive Bias and Extrapolation

The criticality of physics-informed priors is highlighted in an extrapolation regime: training only on a narrow window of quantum data, QGPs based on the unstructured unitary kernel fail entirely at predicting well outside the training region. Conversely, when the correct free-fermionic (matchgate) kernel is used, the model reliably extrapolates far from training data and robustly quantifies uncertainty, even under realistic shot noise.

Figure 2: QGP extrapolation for a 50-qubit regression task: only the correct matchgate kernel supports accurate predictive signal and uncertainty far beyond the training window.

Quantum Phase Prediction and Classification in Many-Body Systems

Applying QGPs to ground-state phase diagrams in quantum spin systems, the model accurately regresses the phase boundary of a bond-alternating XXX model and reproduces high-fidelity phase classifications with only one training point per phase. For classification, the QGP latent function exhibits near-perfect correspondence with the underlying order parameter, even when trained purely on qualitative labels.

Figure 3: QGP accurately predicts phase diagrams and achieves high test accuracy in phase classification with minimal training points in a 50-qubit bond-alternating XXX model.

Figure 4: QGP reconstruction of the magnetization landscape in a 100-qubit XXZ model, using minimal (one or two) training examples and capturing nontrivial staircase phase structures.

Bayesian Optimization with Quantum Surrogates

The Bayesian framework, combined with closed-form QGP posteriors, enables efficient active learning and sample-efficient Bayesian optimization. In both synthetic multi-parameter landscapes and quantum sensor calibration scenarios, QGP-BO rapidly converges to global optima, robustly overcoming local traps that defeat standard optimizers. The QGP surrogate—the matchgate kernel—is strictly necessary; other surrogates cannot match its sample efficiency or global guidance.

Figure 5: QGP-BO identifies the global optimum of rugged quantum objective landscapes in far fewer queries than baseline optimizers, and does so robustly across random initializations.

Numerical Validation, Scalability, and Kernel Correction

Extensive numerical benchmarking demonstrates the shot efficiency of QGP regression versus classical (e.g., interpolation, ridge regression) and quantum Hamiltonian learning baselines. Active learning, driven by the QGP's uncertainty quantification, significantly enhances performance in low-data regimes. In practice, kernel matrix estimation from finite shots encounters noise-induced non-PSD corrections; a robust, theoretically justified correction using the Wigner semicircle law is introduced, resolving issues for large, noisy kernel matrices.

Figure 6: Comparison of QGP, classical, and quantum learning baselines in regression, with mean-squared error (MSE) versus training set size, under realistic shot budgets. Active learning (purple) yields the steepest accuracy gain for sparse data.

Figure 7: Pareto front analysis of regression error versus shot complexity: QGP models dominate Hamiltonian learning (HL) in the data-efficient regime.

Figure 8: Robustness of QGPs under different PSD kernel correction schemes as a function of shot allocation; the Wigner-shift approach avoids error inflation in highly imbalanced noise regimes.

Implications, Limitations, and Outlook

This work realizes a rigorously constructed, data- and resource-efficient quantum learning paradigm anchored in Bayesian inference and group-theoretic priors. Compared to generic parametric quantum models, QGPs encode strong physics-driven inductive bias, yielding generalization and reliability even from sparse or heavily noisy data. The all-qubit matchgate QGP is a foundational example, but the formalism is extensible; the framework invites further theoretical work on new QGP families for interacting systems and other operator algebras.

Practical implications are immediate for quantum experiment design, quantum phase detection, quantum sensor calibration, and closed-loop quantum control, where query or measurement costs are dominant. The methodology's compatibility with Bayesian optimization and active learning further multiplies its utility in resource-limited quantum devices.

Future directions include integrating QGP surrogates into hybrid quantum-classical pipelines (e.g., variational ansatz learning and quantum kernel methods), as well as developing efficient kernel estimation via fermionic shadows. Rigorous expansion and classification of QGP families for interacting non-free models remains a significant theoretical open problem.

Conclusion

By establishing matchgate circuits as the first provable and scalable all-qubit QGP, and by demonstrating broad applicability to quantum regression, classification, and optimization tasks, this work significantly advances the Bayesian and structured approach to quantum learning. The results have substantial consequences for efficient quantum data modeling, guided quantum query complexity, and the principled exploitation of physical knowledge in quantum machine learning.

• (2605.00099), "Provable and scalable quantum Gaussian processes for quantum learning"

---

Note: Figures referenced correspond to the main body figure identifiers provided. Additional technical details (e.g., kernel formulas, active learning methods, and corrections for finite-shot kernel estimation) are available in the cited paper’s main text and supplemental appendices.