⚛️ quantum physics

Quantum Gaussian processes for prediction of channel observations

This paper extends Quantum Gaussian Process (QGP) regression to predict outputs of unknown quantum channels by deriving a closed-form kernel under a uniform prior and introducing an empirical Bayes heuristic to overcome scalability limitations, thereby enabling faithful learning and optimization for both local and global multi-qubit systems even under experimental noise.

Original authors: Jonas Jäger, Yaroslav Khmelnitskiy, Paolo Braccia, Artur Miroszewski, Diego García-Martín, M. Cerezo, Piotr Czarnik

Published 2026-08-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jonas Jäger, Yaroslav Khmelnitskiy, Paolo Braccia, Artur Miroszewski, Diego García-Martín, M. Cerezo, Piotr Czarnik

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to understand how a complex machine works by watching what comes out of it, without ever being allowed to open the case or see the gears inside. This is the daily reality for scientists working with quantum computers. These machines manipulate the fundamental building blocks of matter, but they are notoriously fragile and difficult to control. Often, researchers do not need to know every single detail of how the machine processes information; they simply need to predict the outcome of a specific measurement, such as the average magnetism of a group of atoms, after the machine has run a certain sequence of operations. The challenge is that to learn the machine's behavior by testing every possible input would take so much time and energy that it becomes impossible, especially as the machines grow larger.

To solve this, scientists have turned to a statistical tool called a Gaussian process. Think of this tool as a highly intelligent guesser. If you show it a few examples of how a system behaves, it can predict how it will behave in situations it has never seen before, while also telling you how confident it is in that guess. For years, this method worked well for quantum systems that were perfectly isolated and behaved in a predictable, reversible way. However, real-world quantum devices are messy. They interact with their environment, lose information, and behave in ways that are not perfectly reversible. This paper tackles the difficult problem of applying this intelligent guessing tool to those messy, real-world quantum machines, known as quantum channels.

The researchers began by asking a fundamental question: if we know nothing about a specific quantum machine other than that it is a valid physical process, what is the most reasonable way to guess its behavior? They decided to treat every possible way the machine could evolve as equally likely, a concept known as a uniform prior. Using this broad assumption, they mathematically proved that the machine's outputs would follow a specific statistical pattern. They derived a precise rule, or kernel, that describes how the output for one input is related to the output for another. This rule relies on a simple physical idea: the more similar two starting states are, the more similar their results will be. This created a new, mathematically rigorous framework for predicting quantum outcomes without needing to fully understand the machine's internal workings.

However, when the team tested this new framework on large systems, they hit a wall. The mathematical rule they derived included a factor that made the predicted signals incredibly tiny as the system size grew. For small machines, this was manageable. But for a system with sixty-four quantum bits, the signal became so faint that it was drowned out by the natural noise of measurement. It was as if the tool was telling them that, statistically, the machine should produce no signal at all, making it impossible to learn anything from the data. The researchers realized that while their assumption of total ignorance was mathematically sound, it was too pessimistic for real experiments. In a controlled lab, even a noisy machine usually produces a clear, non-zero signal.

To fix this, the team introduced a clever adjustment. Instead of letting the rigid mathematical rule dictate the strength of the signal, they proposed a method to learn the signal's scale directly from the data. They kept the part of the rule that describes how similar inputs lead to similar outputs, but they replaced the tiny, fixed scaling factor with a flexible number that could be tuned. This allowed the model to trust the data it was seeing. When they tested this adjusted approach, the results were striking. For the large sixty-four-qubit system, the original method failed completely, unable to improve its predictions no matter how many measurements were taken. The new, adjusted method, however, learned the system's behavior successfully. As they increased the number of measurements, the predictions became systematically more accurate, eventually matching the true behavior of the machine.

The team did not stop at simulations. They took their method to a real quantum computer built by IBM, located in Boston. They programmed a small four-qubit machine to act as a noisy quantum channel and used their tool to predict the outcome of measurements on inputs the machine had never seen before. The predictions matched the experimental results with high precision, proving that the method works even when the hardware is imperfect and noisy. This demonstrated that the approach is robust enough for the messy reality of current quantum technology.

Beyond simply predicting outcomes, the researchers showed that this tool could be used to guide the machine toward a desired goal. They set up a task where the goal was to prepare a specific quantum state by adjusting the machine's settings. Using their statistical model as a guide, they were able to find the best settings much faster than standard optimization methods. This suggests that the technique could become a vital part of controlling future quantum devices, helping scientists navigate the complex landscape of quantum behavior without needing to run millions of expensive experiments. The work bridges the gap between theoretical probability and practical application, offering a way to learn from quantum machines even when they are large, noisy, and largely unknown.

Technical Summary: Quantum Gaussian Processes for Prediction of Channel Observations

Problem Statement

Characterizing the action of an unknown quantum channel on a family of input states is a fundamental experimental task. In many scenarios, the objective is not to reconstruct the full quantum process (which requires resources exponential in the number of qubits via process tomography) but to predict the expectation values of a specific set of observables at the output. This task arises in probing many-body dynamics, variational quantum algorithms, and device benchmarking. While Gaussian Process (GP) regression has been successfully applied to closed-system unitary dynamics, extending this framework to open-system quantum channels remains an open challenge. Specifically, there is a need to analytically derive a physics-informed GP prior for general quantum channels that allows for efficient learning from limited measurement data.

Methodology

1. Theoretical Derivation of Channel QGPs

The authors extend the Quantum Gaussian Process (QGP) framework to general quantum channels. They model the unknown channel $C$ as a stochastic process drawn from an ensemble.

Prior Selection: In the absence of specific channel knowledge, the authors adopt the Lebesgue measure over the convex set of quantum channels as the prior. This corresponds to a uniform sampling over all possible channels, avoiding bias toward any specific region of channel space.
Stinespring Dilation: To evaluate the prior mean and covariance, the channel is represented via Stinespring dilation, mapping the channel to a Haar-random unitary on a dilated Hilbert space (system + environment).
Kernel Derivation: Using Haar integration and Weingarten calculus in the limit of large system size, the authors derive a closed-form QGP prior:
- Mean: The prior mean is zero ( $\mu(t) = 0$ ).
- Kernel: The covariance kernel is given by $\kappa(t_i, t_j) = \frac{1}{d} \text{Tr}[\rho(t_i)\rho(t_j)]$ , where $d = 2^{n_A + 2n_B}$ is the dimension of the dilated space, and $\rho(t)$ are the input states. The correlations are determined entirely by the pairwise overlaps (fidelity) of the input states.

2. Scalability Analysis

The authors analyze the scalability of the derived kernel with respect to finite-shot measurements.

Dimensional Suppression: The kernel contains a prefactor $1/d$ . For channels acting on small subsystems (where $n_A + 2n_B \in O(\log n)$ ), $d$ grows polynomially, and the signal-to-noise ratio remains manageable.
Exponential Barrier: For channels acting extensively on the system (where $n_A + 2n_B \in \Theta(n)$ ), $d$ grows exponentially. This suppresses the kernel scale exponentially, requiring an exponential number of measurement shots to distinguish the signal from noise. Consequently, the Lebesgue prior lacks the inductive bias necessary to learn non-zero observations in large, global systems.

3. Empirical Bayes Heuristic

To address the scalability issue in global systems, the authors propose a rescaled channel kernel.

Mechanism: They replace the fixed dimensional prefactor $1/d$ with a learnable scale parameter $\eta$ . The new kernel is $\kappa_\eta(t_i, t_j) = \eta \text{Tr}[\rho(t_i)\rho(t_j)]$ .
Optimization: The parameter $\eta$ is determined via an empirical Bayes approach by maximizing the log-marginal likelihood of the observations. This is a strictly classical optimization routine that leverages the closed-form objective of GPs, incurring no additional quantum computing overhead.
Rationale: This approach retains the physically motivated state-overlap correlation structure derived from the Lebesgue prior while allowing the overall scale to be adapted to the specific experimental data, effectively restoring learnability for favorable datasets.

Key Results

Numerical Simulations

Small Systems (4 Qubits): For a global channel acting on a 4-qubit system, the standard channel QGP (with the Lebesgue kernel) successfully interpolates observable expectation values. It achieves comparable accuracy to a row Pauli Transfer Matrix (PTM) method but requires one to two orders of magnitude fewer measurement shots.
Large Systems (64 Qubits):
- Local Channel: When the channel acts only on a 4-qubit subsystem of a 64-qubit system, the standard channel QGP successfully extrapolates the observable dependence, demonstrating the utility of the subsystem inductive bias.
- Global Channel: For a channel acting globally on 64 qubits, the standard channel QGP fails to improve upon the zero-mean prior regardless of the shot budget due to exponential suppression. In contrast, the rescaled kernel restores learnability. Predictions improve systematically with the shot budget, and the model successfully reproduces the observable dependence.
Bayesian Optimization: The rescaled channel QGP was deployed as a surrogate model for Bayesian optimization (BO) in a noisy XXZ spin-chain state preparation task. The BO approach, using the rescaled QGP, successfully located the global minimum of the loss function within 100 evaluations, whereas gradient descent failed to reach the minimum even after 800 evaluations.

Hardware Demonstration

The authors implemented a 4-qubit global channel on IBM's ibm_boston quantum computer. The channel was generated by Trotterized transverse-field Ising dynamics coupling the system to a 4-qubit environment. Using 6 training observations, the QGP regression accurately interpolated the expectation value $\langle X_1 \rangle(t)$ over the interval $t \in [0, 1]$ , matching the experimental outcomes despite unknown hardware noise.

Significance and Claims

The paper claims to provide the first analytical extension of Quantum Gaussian Processes from unitary dynamics to general quantum channels. Its primary contributions are:

Provable Convergence: It proves that channel outputs converge to a QGP under a uniform (Lebesgue) prior, deriving a closed-form, hyperparameter-free kernel based on state overlaps.
Scalability Limits: It identifies a fundamental limitation where the uniform prior leads to exponential suppression of the signal for global channels, rendering learning intractable without modification.
Practical Solution: It introduces a rescaled kernel heuristic that combines the physical insights of the analytic derivation with data-driven scale adjustment. This allows the framework to be applied to large-scale, global quantum channels where the standard prior fails.
Efficiency: The method demonstrates superior sample efficiency compared to traditional PTM learning and enables efficient Bayesian optimization for state preparation in noisy environments.

The authors emphasize that while the Lebesgue prior is theoretically "uninformative" for large systems, the rescaled approach leverages the correlation structure inherent in the physics of the problem while adapting to the specific device response, offering a practical path for learning unknown quantum channels in experimental settings.