Matrix-product-state-assisted variational Gibbs-state preparation

arXiv:2510.23546 · quant-ph · Submitted 2025-10-27 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: I'm Kai, and with me are Mira and Lev, guest researcher.

Mira: Today's paper: "Matrix-product-state-assisted variational Gibbs-state preparation".

Kai: This work introduces an MPS-assisted variational framework designed for efficiently preparing quantum Gibbs states on digital quantum processors,

Mira: First, who's behind it and why it matters.

Title and authors: Kai: So, to summarize what this "Matrix-product-state-assisted variational Gibbs-state preparation" paper is about, it’s essentially proposing a method that uses MPS techniques to make the classical evaluation of the Helmholtz free energy much more efficient for preparing quantum Gibbs states.

Mira: It's about taking these scalable tensor network techniques and combining them with a hardware-efficient ansatz to create a hybrid variational framework that can accurately approximate thermal states for one- and two-dimensional systems.

Lev: I see how that combination addresses the limitations we've seen before where circuit depth and measurement overhead became bottlenecks when trying to prepare high-fidelity thermal states on NISQ devices.

Kai: Right, because they show this approach allows them to compute both the energy and the von Neumann entropy without needing resource-intensive quantum measurements or state tomography, which is a big deal for experimental work.

Mira: The core idea hinges on using MPS to compactly represent low-temperature quantum states that obey an area law entanglement, allowing them to compute S(rho) using Equation four after tracing out ancilla qubits.

Lev: That reliance on the area law entanglement is interesting because it suggests they are targeting states that have a structure amenable to MPS compression, which is a strong constraint for how hard they are to prepare.

Kai: They then benchmark this method against two different ansatzes, the Thermofield Double Ansatz and the Hardware-Efficient Ansatz, showing how they compare in practice.

Mira: And their summary points out that the TFDA is model-dependent while the HEA is more generic and compatible with different qubit layouts, which highlights a practical consideration for implementation.

Lev: If you can use a more generic ansatz like the HEA, it makes sense because physical hardware often has constraints on qubit connectivity that a specific, fixed ansatz might not respect well.

The paper's summary: Kai: What they suggest as improvements centers around the choice of ansatz and how they leverage MPS for classical cost-function evaluation to enhance fidelity and scalability across different system sizes.

Mira: They also point out that the HEA outperforms the TFDA in the low-temperature regime, which is a significant finding because it suggests that for preparing states at lower temperatures, we should favor ansätze like the HEA.

Lev: That performance difference between high and low temperatures tells us exactly where we should expect to see more noise or simulation error when running these experiments on actual quantum hardware.

Kai: Furthermore, they showed that using Digital Zero-Noise Extrapolation is the most effective way to mitigate noise for both energy and susceptibility estimates on IBM hardware, reducing relative error by over fifty percent.

Mira: That error mitigation technique is a practical improvement because it directly tackles the noise inherent in NISQ devices by systematically extrapolating results.

Lev: That fifty percent reduction in relative error is substantial; that’s the kind of empirical data we need to see when we start thinking about how many times we can trust these thermal state preparations for real applications.

Kai: They also highlight that the method scales the representation of quantum states linearly with system size using MPS compression, which means they can simulate systems much larger than previously possible.

Mira: That linear scaling is what makes this approach promising because it moves beyond exponential complexity limitations when dealing with increasingly large many-body problems.

The paper's improvements: Kai: So, wrapping up the discussion on this paper, the main implication is that we now have a variational framework that uses MPS to efficiently calculate the free energy for preparing thermal states, allowing us to tackle larger systems and improve accuracy through techniques like ZNE.

Mira: It opens up a clear pathway for using quantum simulation to train quantum Boltzmann machines or perform optimization via thermal sampling on complex systems, provided we choose an appropriate ansatz like the HEA.

Lev: From a researcher's standpoint, this framework gives us a concrete toolset that moves beyond just demonstrating small-scale states and provides reliable estimates for thermodynamic properties across a range of temperatures.

Kai: I think the paper on "Matrix-product-state-assisted variational Gibbs-state preparation" shows that hybrid methods combining scalable entanglement entropy computation with hardware efficient ansätze are a viable path forward.

Mira: It really suggests that focusing on how to map physical constraints onto the ansatz structure, like the HEA's flexibility, is key to making these thermal state preparations practical.

Lev: We still have limitations, though; they noted that challenges remain in the intermediate temperature regime around beta about one/, suggesting deeper circuits might be needed there.

Kai: So, for now, we have a solid methodology for preparing Gibbs states on current hardware, but pushing into those specific intermediate temperature regimes will require further refinement of the ansatz itself.

Mira: Indeed, this paper lays a good foundation by showing the power of MPS in this context before we try to push those specific limits further.

Conclusion: Kai: So we've seen how this paper on "Matrix-product-state-assisted variational Gibbs-state preparation" uses MPS to make classical evaluations for preparing thermal states, which is pretty interesting because it tackles a major bottleneck in simulating these systems.

Mira: Exactly; the way they use MPS to compute the von Neumann entropy without needing full tomography is a really clever trick that addresses the computational cost head-on.

Lev: From my side, I'm looking at how this framework would run on real hardware, and if you can get those energy and susceptibility estimates improved by over fifty percent with ZNE, that makes the whole process much more feasible for us.

Kai: It really does; the fact that they managed to scale up to thirty spins in 1D and six times six sites in 2D using this method is a big step forward for what we can actually build on current NISQ devices.

Mira: And when we look at the results, seeing how the HEA performs better than the TFDA at low temperatures gives us a clear hint about which ansatz structure might be more physically relevant for those specific thermal regimes.

Lev: That comparison between TFDA and HEA is crucial for error correction research because it tells us which physical state preparation method we should even bother optimizing circuit depth around.

Kai: I think the overall implication here is that we're getting a robust toolset to predict thermodynamic properties of these quantum magnets with higher fidelity than we could before.

Mira: It means we can start screening material candidates based on these high-quality thermal state preparations, which has huge potential for condensed matter physics simulations.

Lev: And for error correction, having a clearer picture of the noise sensitivity in susceptibility estimates helps us design better error mitigation strategies tailored to those specific operators.

Kai: It's wild to think about what this means for training quantum Boltzmann machines or doing optimization problems that rely on sampling these states accurately at finite temperatures.

Mira: Those applications are where the real payoff is, as accurate thermal state sampling is fundamental for training QML models in complex distributions.

Lev: I'm cautiously optimistic; while the ZNE improvement is solid, we still have those gaps around beta about one/ that point toward needing more sophisticated ansatzes in future work.

Kai: So, we've got a solid method here for preparing Gibbs states on current hardware using MPS assistance, and it sets a high bar for what kind of scalable thermal state preparation we can expect moving forward.

Mira: I think the paper "Matrix-product-state-assisted variational Gibbs-state preparation" provides a very practical roadmap for bridging the gap between theoretical thermal states and actual NISQ implementation.

Lev: It’s a solid piece of work that gives us concrete performance benchmarks we can actually use to guide our next round of hardware testing.

Kai: We're really excited about the direction this points us in, and I think we need to keep an eye on how they tackle those intermediate temperature challenges next.

Rui-Hao Li, *Semeon Valgushev, Khadijeh Najafi

Center for Computational Life Sciences, Cleveland Clinic · Department of Physics, National Tsing Hua University · Department of Physics and Astronomy, Iowa State University · IBM Quantum T.J. Watson Research Center · MIT-IBM Watson AI Lab

quant-ph

Submitted: 2025-10-27

Updated: 2026-09-29

Comments: v3: 18-page main text + 10-page appendices, 14 figures; substantially revised numerical benchmarks with new ansatz and expanded analyses

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 81/100

The gist: This work introduces an MPS-assisted variational framework designed for efficiently preparing quantum Gibbs states on digital quantum processors, combining scalable tensor-network compression with

Key concepts

MPS (Matrix Product State)
MPS is a scalable tensor network technique used to compactly represent low-temperature quantum states that obey an area law entanglement. This allows researchers to compute properties like the von Neumann entropy using Equation four after tracing out ancilla qubits, enabling efficient state representation.
Variational Gibbs State Preparation
This is a hybrid framework combining MPS techniques with hardware-efficient ansätze to create a variational method for preparing quantum Gibbs states. It aims to accurately approximate thermal states for one- and two-dimensional systems on NISQ devices.
Zero-Noise Extrapolation (ZNE)
ZNE is a noise mitigation technique shown to be effective for estimating energy and susceptibility on IBM hardware. It reduces relative error by over fifty percent by systematically extrapolating results, making thermal state preparations more feasible.
Ansätze (Ansatzes)
These are the specific circuit structures used in the variational framework, such as the Thermofield Double Ansatz (TFDA) and the Hardware-Efficient Ansatz (HEA). The discussion highlights that HEA is more generic and compatible with different qubit layouts, while TFDA is model-dependent.

Terminology

Summary

This work introduces an MPS-assisted variational framework designed for efficiently preparing quantum Gibbs states on digital quantum processors, combining scalable tensor-network compression with hybrid variational algorithms to accurately approximate thermal states in one- and two-dimensional systems. This method is crucial because preparing thermal states is fundamental for simulating complex many-body systems, training quantum Boltzmann machines, and performing optimization via thermal sampling techniques. By leveraging Matrix Product States (MPS) for the classical evaluation of the Helmholtz free energy, the authors provide a pathway to scale Gibbs state preparation beyond current limitations on NISQ devices.

Variational Framework and Free Energy Minimization

The preparation of a Gibbs state is defined by minimizing the Helmholtz free energy:

F(ρ) = E(ρ) − β−1S(ρ), where E(ρ) is the energy, β = 1/kBT is the inverse temperature, and S(ρ) is the von Neumann entropy. The goal is to find ρGibbs = arg min ρ F(ρ). When the trial state ρ is parameterized by a quantum circuit, ρ(θ), evaluating this free energy classically remains challenging because estimating S(ρ) typically requires resource-intensive tomography or stochastic reconstruction. To overcome this, the authors propose leveraging MPS to classically approximate the purified Gibbs state ψ⟩. This allows for the computation of both E(ρ) and S(ρ) without costly quantum measurements or state tomography. Specifically, after tracing out ancilla qubits from a parameterized pure state ψ⟩ represented as an MPS (Equation 3), the von Neumann entropy is computed as S(ρ) = − Tr[Xi λ2i ln λ2i] (Equation 4).

Ansatz Selection: TFDA vs. HEA

The paper compares two popular ansatzes for Gibbs state preparation:

  1. The Thermofield Double Ansatz (TFDA): This ansatz is physically motivated, built on the TFD state purification, and mimics imaginary time evolution towards the target thermal state TFD(β)⟩. The TFDA is noted to be model-dependent. However, it requires a larger number of layers to prepare Gibbs states at lower temperatures.

  2. The Hardware-Efficient Ansatz (HEA): This ansatz is designed to be compatible with the qubit connectivity of near-term quantum devices and is shown to be effective for preparing Gibbs states of local Hamiltonians. The HEA offers flexibility in terms of the number of ancilla qubits and layers, making it more hardware-friendly.

Performance Comparison on Small Systems

Small-scale numerical simulations on 1D transverse-field Ising models (TFIM) and XXZ models were conducted to compare the two ansatzes. The results showed drastically different behaviors:

(a) TFDA vs. HEA (L = N/2) for TFIM:

(b) TFDA vs. HEA for XXZ model:

The findings indicated that the HEA outperforms the TFDA in the low-temperature (large-β) regime, while the TFDA performed better at high temperatures (small β). The HEA's initial state being a product state, unlike the TFDA's maximally entangled TFD state, allows it to capture Gibbs states effectively in the low-temperature limit.

Predictive Power and Hardware Results

The MPS-assisted variational algorithm was tested on larger systems (up to 30 spins in 1D and 6x6 sites in 2D) using the HEA. Key observables assessed included:

  1. Energy density, defined as ε = Tr(ρβH) / N.

  2. Magnetic susceptibility (χ).

  3. Specific heat (cv).

  4. Two-point correlations, Czij = ⟨σzi σzⱼ⟩β − ⟨σzi⟩β⟨σzⱼ⟩β.

Noiseless simulations demonstrated that the thermal energy estimates improve with increasing β and with increasing Na and L, especially for lower β values. For susceptibility, the estimates are more accurate than the specific heat estimates on noisy quantum hardware due to the long-range nature of the susceptibility operator making it more sensitive to noise.

Hardware Implementation and Error Mitigation

The optimized circuits were executed on a 156-qubit IBM Heron processor. To mitigate noise, digital zero-noise extrapolation (ZNE) was found to be the most effective in improving the accuracy of both energy and susceptibility estimates consistently across all β values, reducing relative error by over 50% compared to unmitigated results. The paper concludes that while the HEA is promising for variational Gibbs state preparation, challenges remain in the intermediate temperature regime around β ∼ 1/∆, necessitating deeper circuits or problem-informed ansatzes for future work.

Future Directions

Potential avenues for improvement include:

Improvements for AI systems

Here are specific improvements to AI systems based on the findings of this research, focusing on applications in quantum simulation, machine learning, and optimization:

The core contribution of this paper is a scalable variational framework for preparing quantum Gibbs states using Matrix Product States (MPS) for classical evaluation. The key improvements translate directly into more accurate and efficient simulations of complex physical systems and more effective training/optimization algorithms.

Here are the specific improvements you can make to AI systems:

  1. // Improved Quantum Simulation Capabilities (for Materials Science & Condensed Matter Physics):

  2. // Scalable Preparation of Thermal States: The system can now efficiently prepare high-quality thermal states for 1D lattice models up to 30 sites and 2D systems up to 6x6 sites on current NISQ hardware.

  3. // Accurate Property Estimation: The framework allows for the accurate estimation of key observables (energy density, magnetic susceptibility, specific heat, and two-point correlations) across a wide range of inverse temperatures from low to high T.

  4. // System Size Scaling: By leveraging MPS compression, the method scales the representation of quantum states linearly with system size rather than exponentially, enabling simulations on systems previously intractable for NISQ devices (e.g., 30-spin 1D TFIM).

  5. // Enhanced Accuracy via Error Mitigation: The integration of Digital Zero-Noise Extrapolation (ZNE) significantly improves the accuracy of energy and susceptibility estimates on IBM hardware, reducing relative errors by over 50% compared to unmitigated results.

  6. // Targeted Phase Transition Studies: The framework can be used to study finite-temperature quantum phases and critical points in systems like the TFIM/XXZ models (e.g., identifying the critical temperature βc ≈ 0.5 for the 2D TFIM).

  7. // Local Property Sensitivity Analysis: The research provides insight into how operator locality affects estimation accuracy, allowing researchers to design simulations where long-range observables (like susceptibility) are handled with appropriate circuit depth and ansatz structure.

The improved AI system (or rather, the quantum simulation platform enabled by this framework) can do the following specific things:

  1. // Predictive Modeling for Material Properties: Predict the finite-temperature thermodynamic properties of novel materials or quantum magnets (like spin chains or lattices) with high fidelity, enabling rapid screening of material candidates before costly physical synthesis.

  2. // Quantum Machine Learning (QML) Training: Train quantum Boltzmann machines or other QML models by sampling from high-quality Gibbs states at finite temperatures, leading to more robust and accurate learned models for complex data distributions.

  3. // Optimization for Complex Problems: Use the thermal state preparation capability to facilitate quantum optimization tasks, such as solving semi-definite programming problems or combinatorial optimization problems (e.g., in logistics or scheduling) by sampling relevant thermal states efficiently.

  4. // Benchmarking Quantum Algorithms: Serve as a standardized tool for benchmarking new variational quantum algorithms designed for Gibbs state preparation, providing rigorous comparisons against exact analytical solutions and classical Monte Carlo simulations across various temperature regimes and system sizes.

  5. // Noise-Resilient Simulation Deployment: Deploy high-fidelity quantum simulation experiments on current noisy intermediate-scale quantum (NISQ) devices by utilizing the MPS-assisted variational framework combined with sophisticated error mitigation techniques like ZNE, ensuring that results are reliable despite hardware imperfections.

Abstract

Evaluating the von Neumann entropy is a central challenge in variational Gibbs-state preparation. We introduce and benchmark matrix-product-state (MPS) assistance for classically evaluating and optimizing the purification circuits without full-statevector storage, examining how circuit architecture affects both thermal-state accuracy and the computational resources required. Comparisons on four- and six-spin transverse-field Ising and XXZ chains reveal complementary strengths: a thermofield-double ansatz performs better at high temperature, whereas a contiguous hardware-efficient ansatz (HEA) performs better on cooling. We identify a layer-dependent entropy ceiling in the contiguous HEA and compare it with an interleaved architecture whose entropy capacity is set solely by the ancilla qubit count. MPS-based benchmarks on 10 x 1, 20 x 1, 3 x 3, and 4 x 4 transverse-field Ising systems show that the tested interleaved circuits generally lower the free-energy errors and improve thermal observables. However, greater entropy capacity and lower free energy do not guarantee improvement in every observable. The architectures also differ in classical cost: contiguous registers permit entropy evaluation at a single MPS bond, whereas the interleaved arrangement requires a dense entropy calculation that scales exponentially with the number of ancillas. These benchmarks connect the accuracy of MPS-assisted Gibbs-state preparation to circuit capacity, optimization, and the tractability of energy and entropy evaluation.

Sources

Related papers