Quantum Hamiltonian Learning Using Imperfect Quantum Resources

• Tomography and classical simulation become infeasible as quantum systems grow, raising the problem of learning models that cannot themselves be efficiently simulated.Source: §I · p. 1
• Existing and near-future experiments require independent verification even when their behavior exceeds classical simulation capabilities [3] [6] [8] [9].Source: §I · p. 1
• Quantum Hamiltonian learning uses a trusted quantum simulator to predict dynamics under hypothetical Hamiltonian models and infer which model explains experimental outcomes.Source: §I · p. 1
• The approach uses Bayesian inference, with quantum simulators evaluating likelihoods; interactive learning adds a swap-gate connection and is more stable and efficient than the non-interactive approach [10].Source: §I · p. 1
• Unlike related methods that classically propagate efficient state representations, this approach evaluates likelihood functions using quantum resources.Source: §I · p. 2
• The paper studies Bayesian inference, quantum-simulation-based likelihood evaluation, numerical learning of Ising models, noise robustness, and mismatched model classes.Source: §I · p. 2

• Bayesian inference estimates the Hamiltonian most likely to produce observed data and naturally provides uncertainty estimates through posterior distributions.Source: §II · pp. 2–3
• The prior encodes initial confidence, while the likelihood gives the probability of observed data under a Hamiltonian according to the Born rule.Source: §II · p. 2
• Repeated Bayesian updates sharpen the distribution around the true parameters under informative, nondegenerate experiments; a few hundred experiments can outperform traditional methods requiring billions of outcomes [10] [11].Source: §II · p. 2
• Bayesian inference reduces Hamiltonian inference to likelihood evaluation, making it well suited when likelihoods are easy to compute but motivating quantum Hamiltonian learning when classical evaluation is expensive.Source: §II · pp. 2–3
• Sequential Monte Carlo discretizes the probability distribution with weighted particles so Bayesian updating remains tractable.Source: §II · p. 2
• Liu–West resampling refreshes particles while preserving the posterior mean and covariance, increasing effective sample size as the posterior contracts.Source: §II · pp. 2–3

Motivation and comparison

• Quantum resources address a severe classical bottleneck: simulating the likelihoods for a 100-qubit system is estimated to require roughly 5×10^80 years on the stated workstation.The estimate assumes 2,000 particles and 200 experiments, illustrating why a quantum simulator is needed for scalable characterization.Source: §III · p. 4

Quantum simulation resources

• Quantum simulators estimate likelihoods for candidate Hamiltonians, replacing classically intractable dynamical simulation with sampling from the relevant outcome distribution.The simulator may be universal or specialized, provided it approximately samples Pr(D|x) and can swap its state with the uncharacterized system.Source: §III · pp. 3–4

QLE and IQLE experiments

• QLE repeats the uncharacterized experiment on a trusted simulator, whereas IQLE swaps states and approximately inverts the unknown evolution before measurement.IQLE offers a check for whether the current Hamiltonian hypothesis is close to the true model, although it requires more complex operations.Source: §III · p. 4, Figure 1 · p. 4

Adaptive experiments

• The particle guess heuristic adaptively chooses evolution times from sampled hypotheses, using the Loschmidt echo to make experiments informative about whether the inferred model is close to the true one.The heuristic sets t = ∥H(x−) − H(x′)∥^-1 and can yield a constant amount of information per experiment in well-posed problems.Source: §III · pp. 4–5

Algorithm

• The QHL algorithm maintains weighted Hamiltonian particles, computes likelihoods with a trusted simulator, resamples when necessary, and returns a Bayes estimate of the unknown parameters.Its inputs include particle locations and weights, sampling and experiment counts, state preparation, measurement, and resampling protocols.Source: §III · p. 5

Scaling

• The uncertainty can decrease exponentially with the number of experiments, while total computational time scales as e^(γN) and therefore as δ^-γ when the uncertainty is δ.For less well-posed problems, the cited analysis instead gives shot-noise-like scaling when the relevant exponent is one.Source: §III · p. 5

Experimental setup

• The benchmark uses 20,000 prior particles, the Liu–West resampling algorithm with a = 0.9, an initial |+⟩^⊗n state, and 320 randomly chosen Ising models.The quadratic loss uses d = n − 1 model parameters.Source: §IV · p. 6

Scaling result

• The fitted learning rate γ agrees with the O(1/d) scaling reported previously, implying that these Ising Hamiltonians remain easy to learn as n grows.Here d = n − 1, so the relevant parameter dimension grows linearly with system size.Source: §IV · p. 6

State preparation

• Random local Clifford operations can supply maximally naive initial states while allowing QHL to learn the couplings directly.The paper also describes refocusing as a more structured strategy for suppressing interaction terms while learning transverse-field terms first.Source: A. Non–Commuting Models · p. 7

Non-commuting models

• QHL continues to learn Hamiltonian parameters for non-commuting transverse Ising models, although the numerical study is restricted to two- or three-qubit systems.The restriction reflects the greater classical cost of approximately exponentiating non-commuting Hamiltonians.Source: A. Non–Commuting Models · pp. 6–7

Learning-rate behavior

• For short experiments learning is rapid, but later progress slows to quadratic-loss scaling δ ∝ 1/N, giving total simulation time proportional to 1/ε^2.The slowdown occurs because longer-time dynamics make the effects of different Hamiltonian parameters less distinguishable.Source: A. Non–Commuting Models · p. 7

Experiment design

• The slowdown arises from an approximate degeneracy between onsite and interaction terms that makes individual onsite contributions difficult to resolve.Naive initial-state and measurement choices may therefore provide too little information in the non-commuting setting.Source: A. Non–Commuting Models · p. 7

Breaking degeneracy

• Assuming translationally invariant onsite and interaction terms breaks the degeneracy and produces strikingly better performance despite removing only one parameter when n = 2.The result emphasizes that informative experiments with high parameter contrast are crucial for non-commuting Hamiltonian learning.Source: A. Non–Commuting Models · p. 7

• State transfer between uncharacterized and trusted systems is a key practical noise source, with severity depending on whether the hardware is naturally coupled.The paper identifies minimal noise as possible in integrated architectures but potentially catastrophic errors when flying qubits mediate the coupling.Source: §V · p. 7
• Known depolarizing noise and realistic swap-gate errors do not substantially impede quantum Hamiltonian learning in the studied simulations.The analysis considers depolarizing noise theoretically and numerically, alongside noise models for quantum dots and superconducting circuits.Source: §V · p. 7
• The non-commuting transverse-Ising benchmark evaluates median quadratic loss versus experiment count for two- and three-qubit systems with 50% confidence regions.Source: Figure 3 · p. 8

• 50% depolarizing noise slows learning by only a constant factor, while 5% noise negligibly changes the quadratic-loss scaling for random four-qubit Hamiltonians.Source: A. Depolarizing Noise · p. 8, Figure 5 · p. 9
• This robustness relies on using many SMC particles and knowing the depolarizing-noise strength, which permits likelihood calculations for the noisy experiments.Source: A. Depolarizing Noise · p. 8
• Figure 4 compares median quadratic loss across experiments for two- and three-qubit translationally invariant transverse-Ising models, with shaded 50% confidence regions.Source: Figure 4 · p. 8
• Under the stated assumptions, each observation has a relative posterior impact smaller by 1−N, implying a learning-rate parameter γ that scales as 1−N.Source: A. Depolarizing Noise · p. 9
• For 200 random Ising models, the fitted γ scales as 1−N and roughly as d^-1, suggesting efficient scaling with the number of qubits in the chain.Source: A. Depolarizing Noise · p. 9, Figure 6 · p. 9
• Depolarizing noise may be more problematic for QLE than IQLE because QLE outcome distributions are flatter, reducing visibility and making observations less informative.Source: A. Depolarizing Noise · p. 9

• The cumulant expansion models stochastic environmental processes, open-system dynamics, and finite pulse fidelity as a superoperator noise map for the swap gate.Source: B. Realistic Models for swap Gate Errors · pp. 9–10
• GRAPE-designed swap implementations and XY4 decoupling are evaluated for quantum-dot and superconducting models, including deliberately lower-fidelity superconducting cases.Source: B. Realistic Models for swap Gate Errors · p. 10
• The simulated swap channel is composed with state preparation, evolution, and partial-trace maps to reproduce imperfect coupling within the trusted-register simulation.Source: B. Realistic Models for swap Gate Errors · pp. 10–11
• Honest approximations can trade exaggerated channel error against SMC sampling accuracy, but simulation cost limits the numerics to two qubits.Source: B. Realistic Models for swap Gate Errors · pp. 10–11
• After approximately 200 experiments, the unknown coupling J is estimated to 6–7 digits despite swap-gate infidelities 4–5 orders of magnitude larger than that uncertainty.The result applies to the realistic noise levels studied in Figure 7 and relies substantially on knowing the swap-gate noise model.Source: B. Realistic Models for swap Gate Errors · p. 11

Model mismatch

• QHL is designed to remain stable when the true system contains couplings absent from the trusted simulator, rather than requiring an exactly specified model family.The algorithm infers the best possible parameters within the restricted model and bounds the resulting quadratic loss.Source: §VI · p. 11
• The saturation threshold grows at most linearly with the number of neglected terms when their individual norms remain independent of that number.This scaling preserves tractability for physically motivated high-dimensional systems.Source: §VI · pp. 11–12
• Learning proceeds exponentially before saturating at an error approximately ||H−˜H||_2 when the simulator uses an inexact Hamiltonian model.The approximation does not substantially degrade learning until inference uncertainty becomes comparable to the neglected couplings.Source: §VI · p. 12

Model selection

• Bayesian inference also addresses overparameterized models by implicitly penalizing unnecessary parameters through an Occam-razor effect.The dual model-selection problem tests whether a simpler Hamiltonian should replace a model containing extra coupling terms.Source: §VI · p. 12

Numerical illustration

• Figure 8 evaluates QHL with an Ising simulator on a line against a true complete-graph Ising Hamiltonian containing non-nearest-neighbor interactions.The experiment uses small omitted couplings, including coefficients drawn with standard deviation 10^-4, to separate learning from saturation.Source: §VI · p. 12, Figure 8 · p. 13

Model misspecification

• The algorithm can learn the best parameters available within a misspecified model, even when the true parameters lie outside its allowed set.The resulting parameter distance is bounded rather than driven to zero by an impossible model specification.Source: A. Learning the Best Hamiltonian Model · p. 13

Bayesian model selection

• QHL compares competing models using posterior odds and Bayes factors computed from their marginalized likelihoods.With unbiased prior odds, a Bayes factor above one favors the alternate model, while values below one favor the null model.Source: A. Learning the Best Hamiltonian Model · pp. 13–14
• Bayesian model selection penalizes extra parameters asymptotically through the same complexity-control principle underlying BIC.The exact sequential Monte Carlo procedure approximates the non-asymptotic marginal likelihood rather than relying only on the asymptotic form.Source: A. Learning the Best Hamiltonian Model · pp. 13–14
• Marginal likelihood is already available from the sequential Monte Carlo weight normalization used by QHL, avoiding a separate costly calculation.The method evaluates the required expectations using the likelihood values generated during inference.Source: A. Learning the Best Hamiltonian Model · p. 14
• Two QHL instances can use the same experimental settings to compare a restricted model with an alternative model, switching experimental control when the odds favor the alternative.The procedure repeats until posterior odds sufficiently reduce uncertainty in the favored Hamiltonian.Source: A. Learning the Best Hamiltonian Model · p. 14

Numerical illustration

• In the complete-graph versus line-Ising example, the incorrect reduced model learns rapidly but then saturates, while posterior odds identify the true model.By the 200th measurement, the odds against the reduced model reach at least 10^120:1; the dual experiment also rapidly rejects overfitting.Source: Figure 9 · p. 13, A. Learning the Best Hamiltonian Model · p. 14

Conclusion

• QHL tolerates substantial depolarizing noise, realistic swap-gate errors, noncommuting models, and small unmodeled Hamiltonian terms.These results are supported by numerical and theoretical evidence across the paper’s error and dynamics settings.Source: §VII · p. 14
• When the model omits true interactions, QHL finds approximate Hamiltonians that are maximally close to the true Hamiltonian within the allowed model family.Model selection can additionally identify how the assumed model fails by distinguishing competing models.Source: §VII · p. 14
• The results provide a proof of principle that Bayesian inference paired with realistic quantum simulators can address seemingly intractable Hamiltonian-estimation problems.The authors frame quantum resources as useful for characterizing and controlling quantum information processors.Source: §VII · pp. 14–15
• Figures 10 and 11 illustrate posterior-odds model selection for underfit and overfit Hamiltonians, respectively.The first compares a true model with non-nearest-neighbor terms against a nearest-neighbor reduction, while the second tests the dual case.Source: Figure 10 · p. 15, Figure 11 · p. 15

Limitations and outlook

• The approach is not a final solution or replacement for strong experimental intuition, which remains important for optimizing QHL’s learning rate.The paper presents future extensions including better experiment design, quantum acceleration, and recovery from rare cases where learning stops.Source: §VII · p. 15

Conclusion

• The authors argue that quantum-resource-based characterization may become increasingly valuable as quantum devices exceed classical simulation capacity.They connect this capability to developing larger quantum computers and simulators.Source: §VII · p. 15

Superconducting gates

• The simulations use a superconducting-qubit model whose Hamiltonian contains single-qubit and two-qubit interaction terms.The model includes time-varying controls and stochastic Gaussian noise processes with 1/f power spectra.Source: Appendix A · p. 17
• Table I lists the noise parameters used for the superconducting-gate numerical experiments.The table supports the controlled variation of physical error conditions across simulations.Source: Table I · p. 18, Appendix A · p. 18

Noise-control comparison

• An XY4 decoupling sequence is included to compare primitive shaped-pulse gates with higher-fidelity gates and to refocus stochastic control-field contributions.This separates the effects associated with Lindblad dissipation from those caused by fluctuating controls.Source: Appendix A · p. 18

Quantum-dot model

• The quantum-dot example introduces stochastic voltage-detuning and Zeeman-splitting noise and uses parameters for two-qubit gates from the cited physical model.It provides a relatively low-fidelity swap and shows that the results are not tied to one specific hardware model.Source: Appendix A · p. 18

Noise range

• Noise strength is varied by shortening T1 and increasing 1/f-noise amplitudes Γ, producing swap fidelities from approximately 0.9 to nearly ideal.Shorter T1 strengthens dissipation, while larger Γ increases stochasticity in the control fields.Source: Appendix A · p. 18