Statistical Mechanics Calculations Using Variational Autoregressive Networks and Quantum Annealing

Yuta Tamura,Masayuki Ohzeki

2024-05-20

Abstract:In statistical mechanics, computing the partition function is generally difficult. An approximation method using a variational autoregressive network (VAN) has been proposed recently. This approach offers the advantage of directly calculating the generation probabilities while obtaining a significantly large number of samples. The present study introduces a novel approximation method that employs samples derived from quantum annealing machines in conjunction with VAN, which are empirically assumed to adhere to the Gibbs-Boltzmann distribution. When applied to the finite-size Sherrington-Kirkpatrick model, the proposed method demonstrates enhanced accuracy compared to the traditional VAN approach and other approximate methods, such as the widely utilized naive mean field.

Disordered Systems and Neural Networks,Machine Learning

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the difficulty in calculating the partition function in statistical mechanics. Specifically, the author proposes a new approximation method that combines the Variational Autoregressive Network (VAN) and Quantum Annealing (QA) techniques to improve the accuracy of calculating the free energy of the finite - sized Sherrington - Kirkpatrick model.

### Background
In statistical mechanics, calculating the partition function is usually very difficult because as the system scale increases, the computational complexity rises sharply. Traditional methods such as the Naive Mean - Field (NMF) and Bethe approximations can simplify the calculation, but they perform poorly in strongly - coupled systems and graph structures with a large number of loops.

### Method Innovation
1. **Variational Autoregressive Network (VAN)**:
- VAN is an extended variational mean - field structure that can directly calculate the generation probability and can efficiently generate a large number of samples.
- By optimizing the parameters, VAN can provide an upper - bound estimate of the free energy.

2. **Quantum Annealing (QA)**:
- QA is an algorithm that uses quantum effects to solve combinatorial optimization problems and is especially suitable for finding the ground state of the Ising model.
- Although actual QA machines may not be able to perfectly find the ground state due to environmental factors and technical challenges, its output distribution is approximately the Gibbs - Boltzmann distribution.

3. **Combining VAN and QA**:
- The author proposes a new method, which directly uses the samples generated by QA as the input of VAN, rather than just for training the neural network.
- This method aims to use the samples generated by QA to improve the accuracy and efficiency of VAN in calculating the free energy.

### Experimental Results
- **Sherrington - Kirkpatrick model**:
- The author uses the D - Wave Advantage 4.1 system to generate QA samples and uses the Metropolis - Hastings algorithm to generate MCMC samples.
- The experimental results show that the VAN method combined with QA and the low - temperature MCMC prior distribution shows higher accuracy in calculating the free energy, especially under low - temperature conditions.
- Compared with the traditional VAN method, the new method can also maintain high performance with a small sample size.

### Conclusion
- The author demonstrates the effectiveness of the method combining VAN and QA in calculating the free energy of the Sherrington - Kirkpatrick model.
- This method not only improves the accuracy of the calculation but also broadens the application range of QA in statistical mechanics.
- Although this method still has some black - box characteristics at present, future research can further optimize the performance by adjusting the inverse temperature schedule, etc.

### Formulas
- **Boltzmann distribution**:
\[
p(s)=\frac{e^{-\beta E(s)}}{Z}
\]
where \(\beta = \frac{1}{T}\) is the inverse temperature and \(Z\) is the partition function.

- **Variational free energy**:
\[
F[q]=\frac{1}{\beta}\sum_s q_{\theta}(s)[\beta E(s)+\ln q_{\theta}(s)]
\]

- **Joint probability distribution**:
\[
q_{\theta}(s, s') = q_{\theta}(s|s')p(s')
\]

These formulas are used in the paper to describe the probability distribution and energy calculation in statistical mechanics models, as well as the optimization process of the variational autoregressive network.