A team of researchers has demonstrated a method for putting error bars on the results of a quantum simulator, addressing a problem that becomes more acute as quantum machines outgrow the conventional computers used to check them. The approach reconstructs how a device actually behaves from experimental measurements, then propagates the remaining uncertainty into a bound on the accuracy of its predictions.
The demonstration was carried out on trapped-ion simulators in Innsbruck, Austria. The team first validated the method on a system of ten ions, small enough that a classical computer can still calculate the answer independently. They then applied it to a chain of 51 ions, beyond the scale where direct classical comparison is straightforward.
The work was led by Tristan Kraft of the Technical University of Munich and Peter Zoller of the University of Innsbruck and the Institute for Quantum Optics and Quantum Information of the Austrian Academy of Sciences, together with Barbara Kraus of the Technical University of Munich. Their paper appears in Physical Review X under the title "Bounded-Error Quantum Simulation via Hamiltonian and Lindbladian Learning."
Quantum simulators are purpose-built physical systems that reproduce the behaviour of other quantum systems. They are particularly useful for studying large numbers of interacting particles, where the required calculations can overwhelm even powerful classical machines. But a simulation is only useful if researchers can tell how much to trust it.
The usual check is comparison with a classical calculation. That works while the simulated system is small. Once the quantum device explores regimes that classical machines cannot reproduce, the comparison disappears. The new method is designed for that regime: it uses measurements of the device itself to estimate which interactions are present and how noise affects them, then calculates the resulting limits on the simulation's accuracy.
The authors do not claim to have solved verification for all quantum computers. They have shown a route to bounded-error results on a specific platform, and they are now adapting it to two-dimensional simulators that can model more complex interactions. The broader ambition is to change how quantum advantage is assessed, so that accuracy as well as speed becomes part of the comparison.
The verification problem
Classical computers are the default reference for checking calculations. If a new machine produces a result, scientists normally ask whether a conventional machine can reproduce it. For quantum simulators, that strategy has a natural endpoint. The devices are built precisely because some quantum systems are too complex for classical calculation.
This creates what researchers call the verification problem. A simulator may produce an answer that looks plausible, but without an independent benchmark there is no obvious way to know whether the device implemented the intended model correctly. Small differences in interactions, unwanted coupling to the environment and imperfect measurements can all shift the result.
The problem is not hypothetical. Quantum simulators are used to investigate phenomena such as magnetism in materials, phase transitions and the behaviour of many interacting particles. If the device's Hamiltonian, the mathematical description of its energies and interactions, differs from the intended one, the simulated physics may be subtly wrong.
Until now, many experiments have relied on a combination of small-scale benchmarking and theoretical assumptions. Researchers test a device on systems they can calculate, then assume that its behaviour scales in a predictable way. That assumption becomes harder to justify as the apparatus grows and as the questions asked of it change.
The new approach tackles the issue by making the uncertainty explicit. Rather than reporting a single number as the simulation result, it reports a value with an error margin derived from experimental data. That margin is intended to quantify how much the result could change given what is known and unknown about the device.
What the team built
The method combines Hamiltonian learning and Lindbladian learning. In plain terms, the researchers use measurements to infer two things: the coherent interactions that govern how the quantum system evolves, and the incoherent processes such as noise and dissipation that disturb it.
A Hamiltonian describes the energy structure of a quantum system and the interactions among its parts. A Lindbladian extends that description to open systems, where the environment can leak energy or introduce randomness. Learning both from data gives a more realistic picture of an experiment than assuming the device follows an ideal model.
The team's starting point is experimental measurements of the simulator. Those measurements are used to reconstruct the relevant interactions and the main influences of fluctuations and noise. Once that reconstruction is in place, the researchers calculate how uncertainty in the reconstructed model affects the quantity the simulator is meant to estimate.
The output is a bounded result. Instead of saying that a simulation produced value X, the researchers can say that the value lies within a specified range, given the data and the model. The range reflects both statistical uncertainty in the measurements and uncertainty about which physical processes are present.
The authors describe the device as providing not just a single value but a result with error margins that quantify its accuracy. That distinction is central to the idea of verifiable quantum simulation: the user receives an estimate together with a statement about how far off it could be.
Learning the simulator from measurements
Reconstructing a quantum simulator's behaviour is itself a measurement challenge. Researchers cannot simply look at every component of a many-particle quantum state. They must choose measurements that reveal the parameters they care about while leaving enough information to estimate the surrounding uncertainty.
In this study, the team used data from the ion-trap simulator to identify the interactions that mattered most for the quantity being simulated. That selection is important, because a full characterisation of a large quantum device is impractical. The goal is not to know everything about the machine, but to know enough to bound the error on a particular answer.
The reconstruction accounts for disturbances. In a real experiment, interactions may differ slightly from the design, the system is influenced by its surroundings, and measurements carry their own uncertainty. Kraft and his colleagues treat those effects as part of the model rather than ignoring them.
The method is also designed to be tested. On the ten-ion system, the team could check both the reconstructed model and the calculated error limits against independent measurements and classical calculations. That allowed them to ask whether the quoted bounds were sensible: not so narrow that they excluded the true value, and not so broad that they were useless.
Validating error bars is a different task from validating a single result. A bound has to be reliable across repeated experiments and across the range of conditions under which the simulator will be used. The ten-ion tests provided that first check before the team moved to a larger system.
Turning uncertainty into error bars
The central technical achievement is the propagation of model uncertainty into a bound on the simulation result. Once the researchers have reconstructed the interactions and noise processes, they ask how much the predicted quantity could vary within the set of models consistent with the data.
That set can be large. Many different combinations of small interaction changes and noise levels may fit the measurements equally well. The method considers that range rather than picking one best-fit model and pretending it is exact. The resulting interval is meant to cover the answers those plausible models would produce.
This is why the paper's title refers to bounded-error simulation. The bound is not a generic estimate of experimental quality; it is a statement about a specific simulated quantity, derived from a specific set of measurements and assumptions about the physics involved.
The approach differs from simply repeating an experiment and calculating the spread of outcomes. Repeated measurements reduce statistical noise, but they do not by themselves reveal a systematic mismatch between the intended Hamiltonian and the real one. Learning the Hamiltonian and the dissipative processes addresses that mismatch directly.
The researchers caution that the bounds depend on the model class used in the reconstruction. If an important physical process is missing from that class, it will not be reflected in the error margin. The method is therefore strongest when the relevant sources of error are understood well enough to be included.
From ten ions to fifty-one
The experimental work had two stages. The first used a relatively small quantum system containing ten trapped ions. At that size, a conventional computer can still calculate the behaviour of the system, so the team could compare their reconstructed models and error limits with an independent reference.
That comparison served as a validation step. It showed that the method could produce bounds that were consistent with what a classical calculation said about the same system. Without that check, there would be no way to know whether the error margins were meaningful.
The second stage expanded the experiment to a chain of 51 trapped ions. The larger demonstration was carried out by a team led by Manoj Joshi and Christian Roos using an ion-trap quantum simulator. At 51 qubits, direct classical verification becomes substantially more difficult, and the problem the method is designed to solve becomes real.
The jump from ten to fifty-one ions is not just a change in number. Larger systems have more interactions, more possible error channels and a much larger space of models that could fit limited measurements. Showing that the technique still works at that scale is a significant part of the paper's contribution.
The authors present the 51-ion result as a step toward regimes where classical checks are unavailable. They do not claim that the same bounds would automatically hold for every larger device or every simulated quantity. Each application requires its own reconstruction and its own validation of the model assumptions.
Why trapped ions suit the test
In an ion-trap quantum simulator, charged atoms are held in a vacuum by electromagnetic fields. Their internal states serve as qubits, and their interactions can be controlled with lasers. The platform is widely used for quantum simulation because the particles are identical, well isolated and measurable with high precision.
Those properties matter for error bounding. Because the ions are individually addressable, researchers can gather detailed data about the system's behaviour. Because the platform is relatively clean, the main sources of noise can be identified and included in the model.
The Innsbruck group has a long history of work with trapped ions, and the 51-ion demonstration draws on that experience. The experiment required not only preparing the system but also performing the measurements needed to learn its Hamiltonian and dissipative processes.
The team acknowledges that no real experiment is perfect. Interactions may turn out slightly different from what was intended, the environment influences the system, and measurements carry uncertainty. The method is built around that reality rather than assuming an ideal device.
Other quantum platforms may present different challenges. Superconducting circuits, neutral atoms and photons each have their own dominant error sources and measurement constraints. The general idea of learning a model and bounding its predictions can in principle transfer, but each platform requires adaptation.
The role of noise and fluctuations
Noise is not merely an annoyance in quantum simulation. In an open system, the interaction with the environment can change the physics being studied. Dissipation, dephasing and unwanted coupling can shift the outcome just as a change in the intended Hamiltonian would.
That is why the team included Lindbladian learning alongside Hamiltonian learning. The Hamiltonian captures the coherent evolution; the Lindbladian captures the incoherent processes that cause the system to lose information or energy. Treating both on an equal footing gives a more honest account of what the device did.
The researchers use experimental data to identify key influences from fluctuations and noise, then calculate how uncertainty in those influences affects the simulation result. In effect, they ask how much the answer would move if the noise were somewhat stronger or weaker than estimated.
This is important for claims about quantum advantage. A simulation that appears to outperform a classical method is not meaningful if its result is less accurate than the classical answer. Quantifying the error makes the comparison fairer and more informative.
The authors argue that future benchmarks should consider which machine can solve a problem with a smaller, verifiable margin of error, not only which one finishes first. That framing puts reliability at the centre of the comparison.
Moving to two-dimensional simulators
The next experimental target is two-dimensional quantum simulators. In these devices, particles are arranged in a plane rather than a line, allowing richer patterns of interaction. That makes them valuable for studying models relevant to materials science, where geometry strongly influences behaviour.
It also makes verification harder. Classical calculations for two-dimensional systems become significantly more demanding as the number of particles increases. Independent checks that are feasible in one dimension may be out of reach in two, increasing the importance of experimentally determined error limits.
Zoller has emphasised this point in discussing the work. As classical simulation becomes more difficult, independent verification of quantum results becomes more complex, which makes bounds derived from the experiment itself more valuable.
The team is now adapting the method to newer two-dimensional simulators with improved precision and the ability to investigate larger numbers of particles. That work will require careful choices about which interactions to learn and which measurements provide useful constraints.
Success in two dimensions would broaden the method's relevance considerably. Many of the questions researchers hope to answer with quantum simulators involve two-dimensional lattices and frustrated interactions, precisely the regimes where classical benchmarking is weakest.
Rethinking quantum advantage
Quantum advantage usually refers to a quantum device solving a problem faster than a classical one, or solving a problem the classical machine cannot handle at all. The Innsbruck-led team argues that this definition is incomplete unless accuracy is included.
Their reasoning is straightforward. A classical computer and a quantum simulator may tackle the same problem, but a speed comparison means little if the quantum answer carries an unknown error. A verifiable margin of error gives users a way to judge whether the result is good enough for the scientific question being asked.
The proposed shift would change how progress in quantum computing is reported. Instead of asking only how many qubits a machine has or how quickly it runs, researchers could ask how tightly its errors are bounded on problems of interest. A smaller device with well-understood errors might be more useful than a larger one with opaque behaviour.
That idea connects to practical applications. Quantum simulators are often proposed as tools for understanding materials, chemistry and fundamental physics. In those fields, a result without an error estimate is difficult to build on, because subsequent theoretical work needs to know how much weight to give it.
The authors do not suggest that error bounds replace other benchmarks. Speed, scale and connectivity remain important. Their argument is that verifiability should be part of the definition of a useful quantum advantage.
Funding and collaboration
The research brings together institutions in Austria and Germany. Tristan Kraft and Barbara Kraus are based at the Technical University of Munich. Peter Zoller holds positions at the University of Innsbruck and the Institute for Quantum Optics and Quantum Information of the Austrian Academy of Sciences. The ion-trap demonstrations were carried out in Innsbruck.
The full author list includes Manoj K. Joshi, William T. Lam, Tobias Olsacher, Florian Kranzl, Johannes Franke, Lata Kh Joshi, Rainer Blatt, Augusto Smerzi, Daniel Stilck França, Benoît Vermersch and Christian F. Roos, alongside the lead researchers.
The paper was published in Physical Review X, volume 16, issue 3, in 2026. Its DOI is 10.1103/s96t-n8tx, and a preprint is available through arXiv under the identifier 2511.23392.
Funding came from the Austrian Science Fund, the German Ministry of Research, Technology and Space, the European Union, BMW and other organisations. The work therefore sits within a broader European effort to develop quantum technologies while maintaining rigorous standards for evaluating them.
What remains unresolved
The study is a demonstration, not a general solution to quantum verification. The bounds depend on the model class used to interpret measurements, and the method must be revalidated for each new device and simulated quantity. Larger and more complex systems will test how far the reconstruction approach can be pushed.
It is also unclear how broadly the technique transfers to other quantum platforms. Trapped ions offer favourable conditions for measurement and control, but superconducting circuits and neutral atoms may require different strategies for learning interactions and noise.
Even within ion traps, the move to two-dimensional arrays will bring new difficulties. More particles mean more possible interactions, and distinguishing among models that fit limited data becomes harder as the system grows.
Nevertheless, the work establishes an important principle: a quantum simulator can be asked to report not just an answer but a defensible estimate of its own accuracy. As the field moves toward machines that exceed classical checking, that principle is likely to become increasingly central to how results are judged.
Nexuswild welcomes factual corrections. Email contact@nexuswild.com with evidence and the article URL.
