Quantum Hamiltonian-Based Models and the Variational Quantum Thermalizer Algorithm

I. INTRODUCTION

• Mixed states combine classical and quantum correlations, so their generative models need a representation that can learn both.The introduction frames hybridization as a way to use classical and quantum processors for their respective strengths.Source: §I · p. 1

II. QUANTUM HAMILTONIAN-BASED MODELS

• QHBMs represent mixed states by passing a simple parameterized latent mixed state through a unitary quantum neural network.The latent state supplies classical statistics, while the unitary incorporates quantum correlations; the output is the variational visible state.Source: §II · p. 2

A. Modular Hamiltonians and the Exponential Ansatz

• A QHBM is a thermal state of a parameterized modular Hamiltonian, whose exponential form supports computing relative entropies against target distributions.The latent and visible states share a partition function, and their entropies are identical because unitary transformations preserve von Neumann entropy.Source: A. Modular Hamiltonians and the… · pp. 2–3
• QHBMs define two tasks: learning a target mixed state from copies of it, and generating a thermal state from a known Hamiltonian.The paper names these tasks Quantum Modular Hamiltonian Learning (QMHL) and Variational Quantum Thermalization (VQT), respectively.Source: A. Modular Hamiltonians and the… · p. 3

1. Factorized Latent Space Models

• Factorized latent states comprise uncorrelated subsystems; their partition functions and entropies decompose across subsystems, simplifying entropy estimation relative to the full Hilbert space.The number of parameters grows linearly with the number of subsystems.Source: §1 · pp. 3–4
• A completely decorrelated latent representation has a natural orthogonal basis, supporting interpolation, independent component analysis, compression-code learning, and principal component analysis.The decorrelation applies to both entanglement and classical correlations.Source: §1 · p. 4

2. Examples of Latent Subsystem Modular Hamiltonians

• Latent subsystem modular Hamiltonians include diagonal qudit operators, whose parameters define softmax-like multinoulli distributions, and bosonic number operators, whose thermal states have Gaussian Wigner representations.For bosonic modes, one parameter modulates the Gaussian variance; fermionic particle-number operators are another stated choice.Source: §2 · p. 4

3. General Classical Latent Space Models

• Factorized latent priors may be inadequate when decorrelated representations cannot yield accurate results; the paper cites this as a reason uncorrelated-Gaussian-prior VAEs underperform GANs on several tasks.The passage also names Neural ODEs, normalizing flows, and neural energy-based models as classical alternatives.Source: §3 · p. 4
• A general categorical latent distribution scales with the global Hilbert-space dimension, which is typically exponential in qubit count, motivating more efficient classical parameterizations.The classical processor learns the latent distribution and supplies samples for preparing computational-basis states.Source: §3 · pp. 4–5
• Neural energy-based models parameterize an energy function whose exponential distribution defines latent probabilities, and enable sampling and estimation of energy expectations and entropy.The paper identifies Langevin dynamics or Hamiltonian Monte Carlo for sampling and uses energy evaluations to estimate modular-Hamiltonian expectations.Source: §3 · p. 5
• An arbitrary mixed state can be represented in principle, but an unrestricted multinoulli latent distribution paired with a universal random quantum neural network has excessive capacity and exponential parameter-estimation overhead.The paper says physics-informed priors can adapt the latent space and unitary transformation to the situation’s complexity.Source: §3 · p. 5

III. QUANTUM GENERATIVE MODELLING OF MIXED & THERMAL STATES

• QMHL learns the modular Hamiltonian of an unknown mixed state so the trained model can reproduce copies; VQT instead generates a known Hamiltonian’s thermal state.The paper frames the former as quantum machine learning and the latter as quantum simulation.Source: §III · p. 5

A. Quantum Modular Hamiltonian Learning of Mixed Quantum States

• The learning setup assumes access to several copies of the target mixed state in quantum memory.Source: A. Quantum Modular Hamiltonian… · p. 6
• QMHL minimizes forward quantum relative entropy, which is nonnegative and reaches zero exactly when the model equals the data state.QHBMs make evaluation of this objective tractable.Source: A. Quantum Modular Hamiltonian… · p. 6
• The quantum cross-entropy loss omits the data entropy, which is independent of the variational parameters and unknown beforehand.The loss combines the data-state expectation of the model’s modular Hamiltonian with the model’s log partition function.Source: A. Quantum Modular Hamiltonian… · p. 6
• The modular-energy term can be evaluated by applying the inverse unitary to the data state and measuring the latent modular Hamiltonian.The latent partition function is estimated from the known latent distribution.Source: Figure 2 · p. 6, A. Quantum Modular Hamiltonian… · p. 7
• At convergence to the data state, the loss equals the data entropy, providing an entropy estimate as a by-product of training.The paper notes potential use of this property with quantum simulation to estimate entropies and other information-theoretic quantities.Source: A. Quantum Modular Hamiltonian… · p. 7

B. Variational Quantum Thermalizer Algorithm: Quantum Simulation of Thermal States

• VQT targets an approximation to the thermal state specified by a Hamiltonian and inverse temperature β = 1/T.The target is σβ = e^(-βH)/Zβ, with Zβ the thermal partition function.Source: B. Variational Quantum Thermalizer… · p. 7
• Minimizing the model’s free energy provides the VQT variational principle because the thermal state achieves the minimum free energy.The free-energy difference from the thermal state is proportional to the quantum relative entropy.Source: B. Variational Quantum Thermalizer… · p. 7
• With a QHBM ansatz, the latent distribution’s entropy is computed classically, leaving only the energy expectation to be evaluated on the quantum device.The required run count is expected to be similar to that of VQE and other expectation-value-based variational algorithms.Source: B. Variational Quantum Thermalizer… · pp. 7–8, Figure 3 · p. 7
• If training converges to the true thermal state, the loss equals its free energy, which is −log Zβ.Source: B. Variational Quantum Thermalizer… · p. 8
• In the zero-temperature limit, the VQT recovers the VQE loss and ground-state variational principle, with no need for latent-space parameters.At that limit there is no entropy, and the latent state is unitarily equivalent to any starter pure state.Source: B. Variational Quantum Thermalizer… · p. 8

A. Variational Quantum Thermalizer for the 2D Heisenberg Model

• The spin-system ansatz uses a factorized latent distribution and a simulated 1D qubit array with single-qubit rotations and staggered nearest-neighbor two-qubit gates.The circuit has three layers, with no shared parameters or periodic boundary gates.Source: A. Variational Quantum Thermalizer for… · p. 8
• After 200 variational steps, VQT reconstructs the 2D Heisenberg target thermal state; a few iterations later the differences become too small to visualize and training converges.The density-matrix comparison uses Nx = Ny = 2, β = 2.6, Jx = 1.0, and Jy = 0.6.Source: A. Variational Quantum Thermalizer for… · p. 9, Figure 4 · p. 9
• Across 100 random initializations, VQT trains generically and learns trace distance and fidelity while optimizing relative entropy.Figure 5 reports mean values and 95% confidence intervals over 100 training iterations.Source: A. Variational Quantum Thermalizer for… · p. 9, Figure 5 · p. 9

B. Learning Quantum Compression Codes

• Compression encodes reduced-chain states in decoupled latent modes, where high-temperature modes carry most entropy and low-temperature modes can be set to their ground state.The approach approximates quantum compression codes; the plotted covariance-matrix errors compare single-shot compression ratios from 0.1 to 0.9.Source: B. Learning Quantum Compression Codes · pp. 9–10, Figure 6 · p. 10
• Even after removing almost all latent-mode information, the reduced harmonic-chain state can still be approximately reconstructed.For Bosonic Gaussian systems, QHBM directly learns modular modes, compared with exact modes from Williamson decomposition.Source: B. Learning Quantum Compression Codes · p. 10, Figure 7 · p. 10

C. Preparation of Fermionic Thermal States

• QHBM can prepare approximate thermal states for quantum simulation, while modular Hamiltonian learning may facilitate out-of-equilibrium time evolution.The paper identifies these as possible applications across several quantum hardware implementations.Source: C. Preparation of Fermionic Thermal… · p. 10
• The method restricts the BCS mean-field Hamiltonian to Fermionic Gaussian states and learns in the Majorana basis.The example assumes uniform Cooper pairing and sets the chemical potential to zero.Source: C. Preparation of Fermionic Thermal… · pp. 10–11
• VQT minimizes relative entropy while learning the orthogonal transformation and effective frequencies of decoupled Fermionic modes; Givens rotations parameterize the transformation.A discrete-variable implementation additionally requires qubit-to-Fermion and Majorana-basis transformations.Source: C. Preparation of Fermionic Thermal… · p. 11
• After 150 steps, the d-wave target and VQT reconstruction are indistinguishable by fidelity, entropy, and trace distance.Classical simulations learned successfully for systems up to 50 Fermions, although their covariance matrices were too large to discern individual elements clearly.Source: C. Preparation of Fermionic Thermal… · p. 11, Figure 8 · p. 11

V. DISCUSSION & FUTURE WORK

• Unlike methods tailored to low-rank states [11], QHBMs use relative entropy and apply to mixed and thermal states of any rank.They also estimate entropy and free energy and provide a diagonalizing transformation that supports modular time evolution and quantum simulation.Source: §V · pp. 11–12
• Latent-space factorization can support entropy-based quantum compression and may also serve interpolation, clustering, and denoising applications.The authors also suggest layer-wise compression as a possible analogue of layer-wise pretraining.Source: §V · p. 12
• Relative-entropy-based diagonalization can address higher-rank density matrices than prior Hilbert-Schmidt approaches, though it does not remove state-preparation requirements such as QRAM or a special oracle.The authors describe the approach as a possible variational alternative to quantum principal component analysis [103] and related algorithms [8, 104].Source: §V · p. 12
• The authors propose future loss-function variants, including a quantum ELBO, and identify learning mixed-state representations from partial tomography as future work.They state that combining classical neural-network energy-based models with QHBMs could extend beyond full-rank multinoulli estimation and unlock scalability to highly non-trivial large-scale systems.Source: §V · p. 12
• QHBMs represent physical quantities directly and, in preliminary numerics, appear robust to train with few iterations and require less circuit depth than quantum GANs.The discussion notes that GANs can be difficult to train and to use for extracting physical quantities.Source: §V · p. 12
• VQT can prepare a thermal state using polynomial resources given its Hamiltonian, while QMHL can reuse a learned modular Hamiltonian to generate samples at other temperatures.The modular Hamiltonian also provides information related to topological properties, thermalization, and nonequilibrium dynamics.Source: §V · pp. 12–13
• The authors frame QHBMs as applicable to any task involving learning a quantum distribution, not only physical-system applications.This generality is presented alongside the new quantum-probabilistic hybrid variational algorithm class.Source: §V · p. 13

Appendix A: Quantum Neural Networks and Gradients

• A quantum neural network is built from layers combining non-parametric unitaries with parameterized unitaries, whose generators can be Hermitian operators on n qubits.Commuting Pauli terms can be exponentiated separately; noncommuting terms may use a Trotter-Suzuki decomposition or other simulation methods.Source: Appendix A · pp. 16–17
• A cost Hamiltonian can be decomposed into weighted operators, and its expectation is evaluated as a linear combination of operator expectation values.For non-commuting terms, the expectations are estimated over separate runs.Source: Appendix A · p. 17
• Central finite differences use 2M objective evaluations for M continuous parameters and yield gradient estimates with O(ε^2) precision.Forward differences use M + 1 queries with O(ε) error, while a five-point stencil uses 4M queries with O(ϵ^4) error.Source: Appendix A · p. 17
• The numerical experiments use first-order central finite differences, although analytic-gradient approaches are also described for single-term and multi-term Hamiltonians.The multi-term method cited uses a linear combination of unitaries, and later methods use stochastic estimation to reduce runs.Source: Appendix A · p. 17

1. Gradients of model parameters

• The gradient with respect to circuit parameters φ is the gradient of the latent modular-Hamiltonian expectation on the data state transformed by the reverse circuit U†(φ).The passage relates this to gradients of multi-term observables in VQE and quantum neural networks.Source: §1 · p. 17

2. Gradients of variational parameters

• Gradients with respect to latent parameters θ include both modular-Hamiltonian expectation terms and the partition function; for a simple parameterized Hamiltonian, the partition-function gradient can be evaluated classically from its known spectrum.The text states that training on quantum cross-entropy can theoretically train the generative model using gradient-based techniques.Source: §2 · pp. 17–18

Appendix C: Distance Metrics

• Appendix C introduces trace distance and generic mixed-state fidelity as distance metrics.The supplied passages do not provide their formulas.Source: Appendix C · p. 18

Appendix D: Investigation of Ansatz Performance

• The ansatz study notes that Gaussian states admit a factorized latent solution, whereas qubit systems with inter-qubit coupling do not guarantee one.The simulations are described as small in scale and restricted in the model classes considered.Source: Appendix D · p. 18
• For a four-qubit Heisenberg model, the factorized ansatz reconstructs nearly perfectly at β → 0 and approaches ideal reconstruction at β → ∞, with occasional local-minimum failures and high variance at the latter limit.At intermediate temperatures, where both quantum and classical correlations matter, factorization incurs slight losses relative to the general ansatz.Source: Appendix D · p. 18, Figure 9 · p. 19
• For random coupling models on a four-qubit chain, the factorized latent-space ansatz showed a relatively small performance loss compared with the generic ansatz after convergence or 200 training steps.Table II reports fidelity for 10 model Hamiltonians, using 10 random initializations per ansatz for each Hamiltonian.Source: Appendix D · p. 19, Table II · p. 19

Appendix E: Continuous Variable Gaussian Quantum Information

• Gaussian states are specified by their first and second moments, with higher-order correlators expressible through second-order correlators by Wick’s theorem.For thermal states, the mean displacement is zero and the density matrix is represented by the covariance matrix.Source: Appendix E · p. 19
• Bosonic quadratures obey canonical commutation relations, and the covariance matrix is constrained by the uncertainty principle; Gaussian transformations preserve the symplectic form.The appendix describes the symplectic form as a 2N × 2N skew-symmetric matrix.Source: Appendix E · p. 19
• In symplectic diagonal form, the state has independent effective oscillators and simplified entropy; the logarithm of the partition function also has a simplified form.The passage defines ωj as the frequency of the jth effective oscillator and sets β, ℏ, and kB to unity.Source: Appendix E · pp. 19–20

1. Fermions

• Fermionic Gaussian transformations use orthogonal mode rotations, parameterized classically by N(N −1)/2 antisymmetric entries or on hardware by pairwise Givens rotations.The classical parameterization antisymmetrizes a lower-diagonal matrix; Givens rotations can be systematically sequenced to represent the transformations.Source: §1 · p. 20

2. Bosons

• Bosonic symplectic transformations use a Bloch-Messiah decomposition into O1, D, and O2, with 2N^2 −N real parameters and optical implementations using passive transformations and squeezers.O1 and O2 are real orthogonal and symplectic; D is diagonal with paired parameters d_i and d_i^-1.Source: §2 · p. 20

Appendix G: Gaussian Fermionic Quantum Systems

• Fermionic implementations may use Jordan-Wigner, Bravyi-Kitaev, or BVC mappings; in one dimension, each Givens rotation uses two CNOT gates and one controlled-phase rotation.Givens rotations are close to physical device operations; higher-dimensional Jordan-Wigner implementations require techniques to handle phases.Source: Appendix G · pp. 20–21
• In the Jordan-Wigner representation, measuring Pauli-Z on a qubit measures its corresponding effective fermionic-mode occupation, and gradients can be obtained from expectation values at parameters shifted by π 2 .The derivative of a Givens rotation is given analytically, enabling exact gradient methods.Source: Appendix G · pp. 20–21
• Each physical fermion maps locally to two Majorana modes obeying canonical anticommutation relations; covariance-matrix orthogonal transformations diagonalize the thermal state into independent effective oscillators.In the Majorana basis, generic orthogonal transformations need no particle-hole transformations and require at most O(N_f^2) Givens rotations.Source: Appendix G · p. 21

• The mean-field BCS Hamiltonian is presented as a paradigmatic phenomenological model in the study of superconductors.Source: Appendix H · p. 21
• The model describes electrons on an Nx by Ny two-dimensional lattice, with two spin-orbitals per site and 2×Nx×Ny fermions in total.Source: Appendix H · p. 21
• The chemical potential µ sets filling density, t is next-neighbor hopping strength, and ∆ij parameterizes the superconducting gap and incentivizes Cooper-pair formation.Source: Appendix H · p. 21