How Can We Predict Complex Quantum Systems from Only a Few Experiments?
Quantum computers are becoming larger and more capable, but understanding what happens inside them remains a difficult scientific problem. When a quantum device performs an unknown operation, one natural approach is to reconstruct the entire process. Yet as the number of qubits increases, the amount of information required can grow extremely quickly.
A recent study introduces a different strategy: instead of reconstructing everything a quantum device does, learn only the relationship between the quantum states entering the device and the measurements we actually care about. The approach combines quantum channels, Gaussian processes, and Bayesian inference to make predictions from a limited number of experiments. The researchers demonstrate the method on systems of up to 64 qubits and show that a rescaled version of the model can recover learnability even for global quantum channels.
Key Takeaways
* Quantum Gaussian processes can predict measurements produced by an unknown quantum evolution without reconstructing the complete quantum process.
* The method incorporates physical information about quantum channels into its statistical model through a kernel related to the overlap between quantum states.
* A rescaled kernel addresses an important scalability problem that appears when quantum channels act globally on large systems.
* Numerical experiments reached systems of up to 64 qubits, while experiments on noisy quantum hardware demonstrated robustness under realistic conditions.
* The same framework can also serve as a surrogate model for Bayesian optimization, helping select useful quantum experiments while reducing the number of evaluations required.
The Context
To understand why this approach matters, consider the problem of characterizing a quantum device.
Suppose a device receives a quantum state, applies some unknown evolution, and produces another quantum state. We may then measure a particular observable, such as a Pauli operator or another physical quantity. If we want to completely characterize the device, we could attempt to reconstruct the full quantum process through quantum process tomography.
The difficulty is that a complete description becomes increasingly expensive as the system grows. A quantum system with many qubits has an enormous state space, and reconstructing every aspect of the transformation can require resources that quickly become impractical.
But complete characterization is not always necessary.
If the actual goal is to predict a particular measurement, it may be much more efficient to learn the input-output relationship directly. We can prepare a limited number of input states, send them through the device, record the resulting measurements, and use those observations to predict what would happen for new states.
This changes the question from:
"What exactly is the quantum device doing?"
to:
"What measurement should we expect when this quantum state enters the device?"
That distinction is at the heart of the approach.
The Main Idea
The researchers use a quantum Gaussian process to model the relationship between input quantum states and the expectation value of an observable measured after an unknown quantum evolution.
A Gaussian process is a probabilistic machine-learning model that can make predictions from relatively small datasets. An important feature is that it provides not only a predicted value but also an estimate of the uncertainty surrounding that prediction.
For quantum systems, however, the model needs to understand something more than ordinary numerical similarity between data points.
Two quantum states can have a physical relationship that is naturally expressed through their state overlap. This information can be incorporated into the kernel of the Gaussian process.
The kernel determines how strongly different input states should be correlated within the model. Instead of learning this relationship entirely from a large dataset, the quantum version derives important aspects of the kernel from the physical structure of quantum channels.
This effectively gives the model a physics-based inductive bias.
In other words, the model is not starting from the assumption that every possible relationship between two quantum states is equally likely. It uses information about quantum mechanics to guide its predictions.
The new work extends quantum Gaussian-process regression beyond ideal unitary evolutions to the more general setting of quantum channels. This is important because real quantum devices interact with their environment and experience noise, meaning that their evolution cannot always be described as a perfectly isolated unitary transformation.
The Scaling Problem
There is, however, a serious challenge.
The kernel derived from a broad uniform prior over quantum channels contains a dimensional factor that determines the scale of the observable correlations. For local channels acting on only a small subsystem, this factor can remain manageable.
For global channels, the situation is very different.
As the number of qubits increases, the relevant dimensional factor can become exponentially suppressed. The resulting signal becomes so small that learning the relationship between input states and measurements can require an impractical number of observations.
This creates a fundamental scalability problem.
The researchers address it by introducing a rescaled kernel.
Rather than forcing the dimensional factor to remain fixed according to the original prior, they replace it with a learnable scale parameter. The model can then determine the appropriate scale from the available observations while preserving the physically meaningful state-overlap structure of the kernel.
The scale is estimated using empirical Bayes, specifically through maximization of the marginal likelihood. This optimization can be performed classically and does not require additional quantum operations.
Why It Matters
The significance of this approach is that it changes how we can think about characterizing large quantum systems.
A quantum computer does not necessarily have to reveal its complete internal process for us to make useful predictions about its behavior.
Instead, we can treat the device as a physical system that can be queried experimentally.
The quantum hardware produces measurements.
The classical model learns from those measurements.
The model then predicts the outcomes of experiments that have not yet been performed and quantifies the uncertainty of those predictions.
This creates a feedback loop between quantum experiments and classical machine learning:
Quantum state → Quantum experiment → Measurement → Model update → Prediction → Experiment selection → New measurement
Such a strategy can be particularly valuable when quantum measurements are expensive.
Every additional experiment may require repeated circuit executions, multiple measurement shots, and valuable hardware time. If a statistical model can identify which experiments are most informative, researchers may be able to extract more information without proportionally increasing the experimental workload.
The study also demonstrates this idea through Bayesian optimization.
Here, the quantum Gaussian process acts as a surrogate model. Instead of evaluating every possible experimental configuration, the model uses previous observations to estimate where promising configurations are likely to exist and guides the next experiment accordingly.
In the reported optimization experiment involving noisy XXZ dynamics, the researchers used the rescaled quantum Gaussian process for state preparation. The optimization runs reached the global minimum within 100 evaluations, while the gradient-descent comparison did not reach the same minimum even after 800 evaluations.
What the Experiments Showed
The researchers tested the framework on systems of different sizes, including simulations involving up to 64 qubits.
For local channels acting on a small four-qubit subsystem of a 64-qubit system, the original kernel showed a strong inductive bias that allowed the model to extrapolate successfully from limited training data.
The more difficult case was a global channel acting across the large system.
Here, the original kernel suffered from the exponential suppression associated with system dimensionality. The rescaled kernel substantially improved the situation, restoring learnability and producing increasingly accurate predictions as the number of measurement shots increased.
The study also tested the approach on a noisy quantum computer. These experiments showed that the method can remain useful under experimental conditions rather than existing only as an idealized numerical technique.
These results are important because scalability is one of the central challenges facing quantum machine learning. A method that performs well on a handful of qubits but becomes unusable as the system grows has limited practical value. The 64-qubit simulations therefore provide an important test of whether the approach can remain useful beyond small toy systems.
What To Watch Next
The results are promising, but several questions remain open.
First, the rescaled kernel introduces a practical solution to the scaling problem, but its success depends on whether the chosen statistical structure accurately represents the physical problem being studied. The learnable scale improves flexibility, but it does not remove the fundamental difficulty of extracting information from a large quantum system.
Second, the experiments reported so far do not mean that arbitrary 64-qubit quantum processes can now be efficiently characterized. The goal is much more specific: predict selected observables from limited measurements rather than reconstruct the complete quantum channel.
This distinction is crucial.
The approach is therefore best viewed as a targeted prediction framework rather than a replacement for full quantum process tomography in every application.
Another important direction is experimental implementation on increasingly large and noisy quantum processors. As hardware improves, it will be interesting to see how well the method performs when measurement noise, device drift, finite sampling, and other experimental imperfections become more significant.
There is also a broader question about the role of physics-informed machine learning in quantum computing.
How much physical structure should be built directly into a learning model?
When does a physics-based inductive bias provide a major advantage over a more general machine-learning model?
And can similar ideas be developed for other classes of quantum dynamics, observables, and experimental optimization problems?
These questions could determine whether quantum Gaussian processes become a specialized technique or part of a broader framework for intelligent quantum experimentation.
