Source-linked AI summary
Tomography and Generative Data Modeling via Quantum Boltzmann Training
Maria Kieferova, Nathan Wiebe
TL;DR
Quantum Boltzmann training seeks to model complex data, while earlier transverse-Ising methods could not learn quantum terms from classical data and used models considered efficiently simulatable. This paper develops broader training approaches and reports better data fitting by the quantum model in every examined instance, while noting that generalization and overfitting need further study.
Problem
Earlier transverse-Ising training could not learn quantum terms from classical data, and its models were considered simulatable by quantum Monte Carlo, leaving the performance of manifestly quantum models an open question.
Source: p. 1Method
The paper develops quantum Boltzmann training methods, including non-diagonal POVM elements for learning quantum Hamiltonian terms and gradient-based training from a supplied quantum state.
Source: p. 4Results
In every instance examined, the quantum model fitted the data better than the classical model.
Source: p. 4Takeaways & Limitations
The learned Hamiltonian provides a state model that supports generating copies of the reconstructed state, unlike conventional tomographic methods.
Source: p. 4Further study is needed to establish whether these models generalize better than classical Boltzmann machines and to assess their susceptibility to overfitting.
Source: pp. 4–5
Abstract
from arXiv · showhide
The promise of quantum neural nets, which utilize quantum effects to model complex data sets, has made their development an aspirational goal for quantum machine learning and quantum computing in general. Here we provide new methods of training quantum Boltzmann machines, which are a class of recurrent quantum neural network. Our work generalizes existing methods and provides new approaches for training quantum neural networks that compare favorably to existing methods. We further demonstrate that quantum Boltzmann machines enable a form of quantum state tomography that not only estimates a state but provides a perscription for generating copies of the reconstructed state. Classical Boltzmann machines are incapable of this. Finally we compare small non-stoquastic quantum Boltzmann machines to traditional Boltzmann machines for generative tasks and observe evidence that quantum models outperform their classical counterparts.
Appendix A: Preparing Thermal States
Thermal-state preparation is NP-hard, and this appendix reviews quantum approaches with different complexity trade-offs. Their suitability depends on the partition function, spectral gap, simulation cost, and required precision.
Appendix A: Preparing Thermal States
- Preparing a thermal state is NP-hard; classical methods approximate it with contrastive divergence, while quantum approaches include high-precision preparation methods [10].The appendix also notes quantum solutions proposed in [6] [8].Source: Appendix A · p. 5
- Chowdhury–Somma preparation approximates the exponential as a linear combination of unitaries, with efficiency depending on the partition function and required error.It is roughly quadratically better for constant error, or exponentially better when 1/ϵ is large; efficiency also depends on Z ∈ Θ(N/polylog(N)).Source: Appendix A · p. 5
- The Yung–Aspuru-Guzik approach uses a Szegedy walk and phase estimation, and its relative appeal depends on the transition-matrix gap, simulation cost, and desired precision.The appendix favors Eq. (A1) for a large partition function or high accuracy, and Eq. (A2) when the gap is small, simulation is inexpensive, and low precision suffices.Source: Appendix A · p. 5
Appendix B: Relative Entropy Training
Relative-entropy training rapidly fits many mixed states, but pure states require more training and some cases retain much larger errors. The appendix relates these patterns to the target states’ properties and to concentration expectations.
Appendix B: Relative Entropy Training
- After 35 epochs at η = 1, mixed-state median error reaches numerical precision, while the 97.5th percentile is 10^-5 after 60 epochs.That upper percentile decays much more slowly than the median.Source: Appendix B · p. 7
- Pure states require many more epochs than highly mixed states because approximating them requires large Hamiltonian weights, with a rougher training landscape expected near that limit.Pure states are possible only as ∥H∥→∞.Source: Appendix B · p. 7
- Nearly pure examples may account for the fat tails in the mixed-state error distribution.Source: Appendix B · p. 7
- For Haar-random pure targets, Levy’s lemma predicts confidence-interval widths shrinking exponentially with qubit count, but the same concentration is not established for mixed states.The stated width is O(1/N) ⊆ O(2^-n/2).Source: Appendix B · p. 7
Appendix C: Applications to Hamiltonian Learning
The appendix tests whether relative-entropy training can reconstruct a transverse-Ising Hamiltonian from thermal expectation values. Learning difficulty depends strongly on Hamiltonian norm, or equivalently thermal-state inverse temperature.
Appendix C: Applications to Hamiltonian Learning
- The method learns a local transverse-Ising model on a complete graph from the expectation values of its Hamiltonian terms.The experiment tests reconstruction of the true Hamiltonian from those values.Source: Appendix C · pp. 7–8
- Because transverse-Ising models can often be simulated efficiently with quantum Monte Carlo, quantum computers are not necessary for estimating gradients in these cases.Source: Appendix C · p. 8
- Learning is much easier at high temperature than at low temperature, and the error varies more with Hamiltonian norm than with qubit count.Figure 4 compares relative entropies and Hamiltonian errors for normalized and unnormalized transverse-Ising Hamiltonians.Source: Appendix C · p. 8
1. Mean Field Approximations
The appendix applies relative-entropy training to construct mean-field approximations and finds that their quality depends on the underlying Hamiltonian’s correlations. A complete-graph transverse-Ising example is approximated closely, whereas the tested n-local cases are not.
1. Mean Field Approximations
- For a 5-qubit complete-graph transverse-Ising thermal state, 100 epochs yield a graphically close mean-field approximation with Tr(ρσ) ≈ 0.71.Source: §1 · p. 8
- Across 1,000 random instances, median relative entropy approximately reaches its optimal value after one training epoch, while the saturated error tends to rise with qubit count.The authors associate the increase partly with more correlated complete-graph Hamiltonians as qubit count grows.Source: §1 · p. 8
- For n-local Hamiltonians with n = 2,…,5, mean-field approximations fail qualitatively and quantitatively to capture correlations, even when Hamiltonian norm is reduced.The authors attribute this to the expected strong correlations and conclude fidelity depends more on Hamiltonian properties than qubit count.Source: §1 · p. 8
Appendix D: Commutator Training
Commutator training can provide better gradients than Golden–Thompson training, but its accuracy depends on a tractable low-order expansion and comes with stability and measurement-cost trade-offs.
Appendix D: Commutator Training
- Commutator training approximates the series in Eq. (10) as a commutator series using Hadamard’s lemma.The approximation is tractable when truncated at low order if ∥[H, ∂θH]∥ is small.Source: Appendix D · p. 8
- Low-order truncation is expected to outperform Golden–Thompson training with L2 regularization on quantum terms, but is less broadly applicable.The stated condition for avoiding substantial truncation error is ∥[H, ∂θH]∥≪1.Source: Appendix D · p. 8
- At a fixed learning rate, fifth-order commutator gradients outperform Golden–Thompson gradients, and switching methods dramatically improves the objective.The improvement comes at the cost of measuring more expectation values.Source: Appendix D · p. 9
- Commutator training is less stable and can fail abruptly when its expansion does not converge or the learning rate is too large.Its optimal learning rate can differ substantially from Golden–Thompson’s; the authors suggest using it in the final training epochs after approximate training reaches a local optimum.Source: Appendix D · p. 9
Appendix E: Additional Experiments for POVM-based Training
In the tested cases, the all-visible quantum model outperforms the classical cases, while adding quantum hidden units yields limited further gains; the role of hidden units warrants further study.
Appendix E: Additional Experiments for POVM-based Training
- The experiment examines training performance as hidden-unit count varies for a six-visible-unit example.The paper calls for more systematic study of hidden units in Fermionic Boltzmann machines and related models.Source: Appendix E · p. 9
- The all-visible quantum model outperforms every classical case considered.Adding a hidden unit substantially helps the 4-visible-unit quantum model, but additional units provide little further power on this training set.Source: Appendix E · p. 9
- Hidden units can dramatically improve classical learning, while the limited quantum gains suggest Hpqrs terms may already represent the training correlations.The passage notes that the impact of these terms is diminished in the 4-visible-unit case.Source: Appendix E · p. 9
- Figure 6 reports median relative entropies between mean-field and true distributions for 2–5-qubit transverse Ising thermal states.Its panels correspond to 2, 3, 4, and 5 qubits, with dashed 95% confidence intervals.Source: Figure 6 · p. 10
Appendix F: Complexity Analysis
The appendix bounds training-query costs and shows that general high-accuracy generative training faces a tomography-derived lower bound. Its analysis also specifies oracle-cost assumptions and a gradient-estimation sampling scale.
Appendix F: Complexity Analysis
- The cost model assumes approximate Gibbs-state preparation and charges the state-preparation and relative-entropy data oracles equivalently; POVM-element preparation is uncosted.The authors justify the latter choice by noting that elementary POVM elements are often simple projectors and that charging for them would require opening the state-preparation oracle.Source: Appendix F · p. 10
- Switching from Golden–Thompson to commutator training produces a sudden accuracy increase in the four-visible-unit all-visible machine, with parameters chosen for stable commutator training.Source: Figure 7 · p. 11
- Ω(Dr/[ϵ^2 log(D/rϵ)]) queries to Fρ are required to train an arbitrary QBM whose generated state is within trace distance ϵ of a rank-r state, with failure probability Θ(1).The lower bound follows from tomography; the appendix states that training cannot be efficient in general when the generative model must be highly accurate.Source: Appendix F · pp. 10–11
- A complete-term QBM can approximate any density operator, regardless of rank, within trace distance ϵ using a bounded Hamiltonian.Source: Appendix F · p. 11
- General efficient Boltzmann training is precluded without further assumptions about the training data or less onerous requirements on model-output precision.Source: Appendix F · p. 11
- For gradient estimation, n = M/ϵ^2 samples suffice for overall variance ϵ^2, and each of M components requires O(n) queries.Each sample requires one thermal-state preparation or one training-data query.Source: Appendix F · p. 12