- Full text:
- PDF (eReader) / PDF (Download) (823 kB)
Recently, quantum-state representation using artificial neural networks has started to be recognized as a powerful tool. However, due to the black-box nature of machine learning, it is difficult to analyze what machine learns or why it is powerful. Here, by applying one of the simplest neural networks, the restricted Boltzmann machine (RBM), to the ground-state representation of the one-dimensional (1D) transverse-field Ising (TFI) model, we make an attempt to directly analyze the optimized network parameters. In the RBM optimization, a zero-temperature quantum state is mapped onto a finite-temperature classical state of the extended Ising spins that constitute the RBM. We find that the quantum phase transition from the ordered phase to the disordered phase in the 1D TFI model by increasing the transverse field is clearly reflected in the behaviors of the optimized RBM parameters and hence in the finite-temperature phase diagram of the classical RBM Ising system. The present finding of a correspondence between the neural-network parameters and quantum phases suggests that a careful investigation of the neural-network parameters may provide a new route to extracting nontrivial physical insights from the neural-network wave functions.
Quantum many-body wave functions are key quantities in understanding quantum many-body phenomena. Indeed, good ansatzes for the quantum many-body wave functions such as the Bardeen–Cooper–Schrieffer1) and Laughlin2) wave functions have promoted our understanding of physics. Although these wave functions were constructed by human brains, the recent rapid development of numerical methods has offered new routes to construct accurate wave functions. In particular, using machine learning techniques for this purpose is quite interesting because the use of “machine brains” might shed new light on unsolved problems in physics.
Such a trend began in 2017 when Carleo and Troyer introduced quantum many-body wave functions constructed from artificial neural networks for quantum spin systems without frustration.3) Since then, much effort has been made to improve the performance and extend the applicability of the method. Now, the neural-network method has been extended, e.g., to the simulations of spin systems with geometrical frustration,4–13) itinerant boson systems,14,15) fermion systems,4,16–24) fermion-boson coupled systems,25) topologically nontrivial quantum states,26–32) excited states,10,22,25,33–35) real-time evolution,3,36,37) open quantum systems,38–41) and finite-temperature properties.42,43) Through various benchmarks using small system sizes that allow the exact diagonalization and special Hamiltonians for which the quantum Monte Carlo calculations can be performed without the sign problem, the usefulness of the artificial neural networks in quantum many-body problems has started to be recognized. When the accuracy is ensured, one can go beyond benchmark calculations and apply the artificial neural networks to analyze unsettled quantum many-body problems. Indeed, recently, neural-network wave functions have been applied to investigate the physics of frustrated quantum spin systems.11,12)
While there is growing numerical evidence that artificial neural networks are useful for accurately approximating quantum states, there is little understanding of what the machines have learned. This is due to the general problem that the machine learning process is a black box. In order to extract non-trivial physical insights from the many-body wave functions obtained via machine learning, we need to make machine learning “white box”. If such a thing becomes possible, it will provide a new perspective in understanding the quantum many-body problems.
In this paper, as a primitive step, we take one of the simplest neural networks, the restricted Boltzmann machine (RBM), to learn the ground state of the one-dimensional (1D) transverse-field Ising (TFI) model, whose ground state properties are well understood. The RBM is a generative model, which exploits the Boltzmann weights of the Ising spin systems consisting of visible and hidden units. When the RBM is applied to approximate the ground state wave function of the 1D TFI model, the configurations of physical and visible spins are identified, and the zero-temperature quantum state is mapped onto an ensemble of finite-temperature classical states of the extended Ising spin system (see Sect. 2 for more details).
Although the RBM wave function is successfully applied to the ground-state representation of quantum spin models,3,10) the coupling parameters of the optimized RBM network have seldom been investigated. Here, by applying the RBM wave function to the 1D TFI model, we directly investigate the property of the RBM parameters. We find that a quantum phase transition in the 1D TFI model is indeed reflected in the optimized RBM; the behavior of the tail of the RBM coupling parameters changes qualitatively associated with the quantum phase transition. As a result, the finite-temperature phase diagram of the RBM spin system changes qualitatively at the quantum phase transition point. The present result suggests the potential importance of analyzing the optimized neural-network parameters themselves.
This paper is organized as follows. Section 2 introduces the RBM wave function and shows that a zero-temperature quantum state is mapped onto a finite-temperature classical state of the extended Ising spins that constitute the RBM. In Sect. 3, we apply the RBM wave function to approximate the ground state of the 1D TFI model. We then show the results of the investigation of the optimized RBM parameters and the finite-temperature phase diagram of the RBM spin system. Section 4 is devoted to the discussion and summary.
The RBM consists of visible and hidden layers (see Fig. 1 for the structure of the RBM employed in this paper). The visible and hidden units can be viewed as classical Ising spins (
Figure 1. (Color online) Structure of RBM employed in the present study. To represent the wave function of the 1D TFI model, we identify the visible spin configuration with that of spins on the periodic L-spin chain (
In this study, as a first step, we consider a particular case, where the wave function is positive definite [
In this study, we investigate the 1D TFI model, whose Hamiltonian reads
In the 1D TFI model at zero temperature, as the transverse field Γ increases, there exists a quantum phase transition at
The analyses are performed as follows: First, we optimize the RBM wave function in Eq. (2) with the RBM energy given by Eq. (4) to approximate the ground state wave function of the 1D TFI model. Since the ground state is the lowest-energy eigenstate of the Hamiltonian, we optimize the RBM parameters to minimize the total energy (loss function). For the initial parameters, we put small random numbers. To stabilize the optimization, we employ the stochastic reconfiguration (SR) method;44) the SR method reproduces the imaginary-time Hamiltonian evolution as accurately as possible within the representability of the RBM variational wave function (see Ref. 16 for further technical details of the optimization). For simplicity, in this paper, the number of hidden spins
When the optimized RBMs are derived, we analyze the finite-temperature properties of the RBM Ising systems, which consist of
As is described above, we first optimize the RBM wave function with L hidden spins to approximate the ground-state wave function of the TFI model on the 1D L-spin chain with the periodic boundary condition. Figure 2(a) shows the transverse-field Γ dependence of the RBM energy at
Figure 2. (Color online) Energy of the optimized RBM wave function as a function of the transverse field Γ for the 1D TFI model on the periodic 256-spin chain (
Then, it is of great interest to investigate how the optimized
Figure 3. (Color online) Tail of
Figure 4. (Color online) Long-range tail (
To further analyze the behavior of the long-range tail, we define a quantity
We then analyze the finite-temperature property of the RBM spin system (Fig. 1) consisting of
Figure 5 shows the system size dependence of the specific heat C computed from the energy variance as
Figure 5. (Color online) Specific heat C computed from the energy variance as
Figure 6 shows the energy variance
Figure 6. (Color online) Energy variance
In the present study, by applying the RBM wave function to learn the ground state of the 1D TFI model, we have shown that the quantum phase transition is encoded in the RBM spin system. We note that Refs. 46–48 had already shown that one can detect (quantum) phase transition(s) from the neural-network parameters optimized to relate spin configurations and the corresponding parameters characterizing the spin system, such as temperature and Hamiltonian parameters. The current approach is different from those previous studies. Here, the output of the neural network is not the parameters of the system but the quantum many-body wave function. If the optimization (learning) is successful and the accuracy of the RBM wave function is good, the RBM wave function should change its property at the quantum phase transition. However, in the present study, we also investigate the RBM parameters that produce the wave function and find that the RBM parameters themselves change their qualitative property.
This is noteworthy because such qualitative change is not seen in, e.g., the path-integral formalism, another quantum-to-classical mapping than the present approach (we can view the RBM representation of quantum states as a kind of quantum-to-classical mapping within the representability of the RBM, see Sect. 2.1). In the path-integral formalism, the 1D TFI model is mapped onto a classical two-dimensional spin system, whose coupling parameters change gradually as a function of the transverse field (crossover-like behavior). The qualitative change at the quantum phase transition in the present case indicates that the mapping using the RBM, which is a more compact mapping using
The present finding of the close relationship between the quantum phases and the neural-network parameters is encouraging because not only the wave function (output of the neural network) but also the neural-network parameters themselves have some useful information. Therefore, the careful analysis of the neural-network parameters may provide a new route to extracting nontrivial physical insights from the neural-network wave functions. It is of great interest to also investigate other neural networks and other quantum-spin Hamiltonians and investigate how the neural-network parameters behave.
Acknowledgment
We are grateful for the helpful discussions with Nobuyuki Yoshioka. This work was supported by Grant-in-Aids for Scientific Research (JSPS KAKENHI) (Grants Nos. 16H06345, 20K14423, and 21H01041) and MEXT as “Program for Promoting Researches on the Supercomputer Fugaku” (Basic Science for Emergence and Functionality in Quantum Matter —Innovative Strongly-Correlated Electron Science by Integration of “Fugaku” and Frontier Experiments—, JPMXP1020200104).
References
- 1 J. Bardeen, L. N. Cooper, and J. R. Schrieffer, Phys. Rev. 108, 1175 (1957). 10.1103/PhysRev.108.1175 Crossref, Google Scholar
- 2 R. B. Laughlin, Phys. Rev. Lett. 50, 1395 (1983). 10.1103/PhysRevLett.50.1395 Crossref, Google Scholar
- 3 G. Carleo and M. Troyer, Science 355, 602 (2017). 10.1126/science.aag2302 Crossref, Google Scholar
- 4 Z. Cai and J. Liu, Phys. Rev. B 97, 035116 (2018). 10.1103/PhysRevB.97.035116 Crossref, Google Scholar
- 5 X. Liang, W.-Y. Liu, P.-Z. Lin, G.-C. Guo, Y.-S. Zhang, and L. He, Phys. Rev. B 98, 104426 (2018). 10.1103/PhysRevB.98.104426 Crossref, Google Scholar
- 6 K. Choo, T. Neupert, and G. Carleo, Phys. Rev. B 100, 125124 (2019). 10.1103/PhysRevB.100.125124 Crossref, Google Scholar
- 7 F. Ferrari, F. Becca, and J. Carrasquilla, Phys. Rev. B 100, 125131 (2019). 10.1103/PhysRevB.100.125131 Crossref, Google Scholar
- 8 T. Westerhout, N. Astrakhantsev, K. S. Tikhonov, M. I. Katsnelson, and A. A. Bagrov, Nat. Commun. 11, 1593 (2020). 10.1038/s41467-020-15402-w Crossref, Google Scholar
- 9 A. Szabó and C. Castelnovo, Phys. Rev. Res. 2, 033075 (2020). 10.1103/PhysRevResearch.2.033075 Crossref, Google Scholar
- 10 Y. Nomura, J. Phys.: Condens. Matter 33, 174003 (2021). 10.1088/1361-648X/abe268 Crossref, Google Scholar
- 11 Y. Nomura and M. Imada, Phys. Rev. X 11, 031034 (2021). 10.1103/PhysRevX.11.031034 Crossref, Google Scholar
- 12 N. Astrakhantsev, T. Westerhout, A. Tiwari, K. Choo, A. Chen, M. H. Fischer, G. Carleo, and T. Neupert, Phys. Rev. X 11, 041021 (2021). 10.1103/PhysRevX.11.041021 Crossref, Google Scholar
- 13 M. Li, J. Chen, Q. Xiao, Q. Jiang, X. Zhao, R. Lin, F. Wang, X. Liang, L. He, and H. An, arXiv:2108.13830. Google Scholar
- 14 H. Saito, J. Phys. Soc. Jpn. 86, 093001 (2017). 10.7566/JPSJ.86.093001 Link, Google Scholar
- 15 H. Saito and M. Kato, J. Phys. Soc. Jpn. 87, 014001 (2018). 10.7566/JPSJ.87.014001 Link, Google Scholar
- 16 Y. Nomura, A. S. Darmawan, Y. Yamaji, and M. Imada, Phys. Rev. B 96, 205152 (2017). 10.1103/PhysRevB.96.205152 Crossref, Google Scholar
- 17 D. Luo and B. K. Clark, Phys. Rev. Lett. 122, 226401 (2019). 10.1103/PhysRevLett.122.226401 Crossref, Google Scholar
- 18 J. Han, L. Zhang, and E. Weinan, J. Comput. Phys. 399, 108929 (2019). 10.1016/j.jcp.2019.108929 Crossref, Google Scholar
- 19 K. Choo, A. Mezzacapo, and G. Carleo, Nat. Commun. 11, 2368 (2020). 10.1038/s41467-020-15724-9 Crossref, Google Scholar
- 20 D. Pfau, J. S. Spencer, A. G. D. G. Matthews, and W. M. C. Foulkes, Phys. Rev. Res. 2, 033429 (2020). 10.1103/PhysRevResearch.2.033429 Crossref, Google Scholar
- 21 J. Hermann, Z. Schätzle, and F. Noé, Nat. Chem. 12, 891 (2020). 10.1038/s41557-020-0544-y Crossref, Google Scholar
- 22 N. Yoshioka, W. Mizukami, and F. Nori, Commun. Phys. 4, 106 (2021). 10.1038/s42005-021-00609-0 Crossref, Google Scholar
- 23 J. Stokes, J. R. Moreno, E. A. Pnevmatikakis, and G. Carleo, Phys. Rev. B 102, 205122 (2020). 10.1103/PhysRevB.102.205122 Crossref, Google Scholar
- 24 K. Inui, Y. Kato, and Y. Motome, Phys. Rev. Res. 3, 043126 (2021). 10.1103/PhysRevResearch.3.043126 Crossref, Google Scholar
- 25 Y. Nomura, J. Phys. Soc. Jpn. 89, 054706 (2020). 10.7566/JPSJ.89.054706 Link, Google Scholar
- 26 D.-L. Deng, X. Li, and S. Das Sarma, Phys. Rev. X 7, 021021 (2017). 10.1103/PhysRevX.7.021021 Crossref, Google Scholar
- 27 D.-L. Deng, X. Li, and S. Das Sarma, Phys. Rev. B 96, 195145 (2017). 10.1103/PhysRevB.96.195145 Crossref, Google Scholar
- 28 I. Glasser, N. Pancotti, M. August, I. D. Rodriguez, and J. I. Cirac, Phys. Rev. X 8, 011006 (2018). 10.1103/PhysRevX.8.011006 Crossref, Google Scholar
- 29 S. R. Clark, J. Phys. A 51, 135301 (2018). 10.1088/1751-8121/aaaaf2 Crossref, Google Scholar
- 30 S. Lu, X. Gao, and L.-M. Duan, Phys. Rev. B 99, 155136 (2019). 10.1103/PhysRevB.99.155136 Crossref, Google Scholar
- 31 R. Kaubruegger, L. Pastori, and J. C. Budich, Phys. Rev. B 97, 195136 (2018). 10.1103/PhysRevB.97.195136 Crossref, Google Scholar
- 32 Y. Huang and J. E. Moore, Phys. Rev. Lett. 127, 170601 (2021). 10.1103/PhysRevLett.127.170601 Crossref, Google Scholar
- 33 K. Choo, G. Carleo, N. Regnault, and T. Neupert, Phys. Rev. Lett. 121, 167204 (2018). 10.1103/PhysRevLett.121.167204 Crossref, Google Scholar
- 34 D. Hendry and A. E. Feiguin, Phys. Rev. B 100, 245123 (2019). 10.1103/PhysRevB.100.245123 Crossref, Google Scholar
- 35 T. Vieijra, C. Casert, J. Nys, W. De Neve, J. Haegeman, J. Ryckebusch, and F. Verstraete, Phys. Rev. Lett. 124, 097201 (2020). 10.1103/PhysRevLett.124.097201 Crossref, Google Scholar
- 36 S. Czischek, M. Gärttner, and T. Gasenzer, Phys. Rev. B 98, 024311 (2018). 10.1103/PhysRevB.98.024311 Crossref, Google Scholar
- 37 M. Schmitt and M. Heyl, Phys. Rev. Lett. 125, 100503 (2020). 10.1103/PhysRevLett.125.100503 Crossref, Google Scholar
- 38 A. Nagy and V. Savona, Phys. Rev. Lett. 122, 250501 (2019). 10.1103/PhysRevLett.122.250501 Crossref, Google Scholar
- 39 M. J. Hartmann and G. Carleo, Phys. Rev. Lett. 122, 250502 (2019). 10.1103/PhysRevLett.122.250502 Crossref, Google Scholar
- 40 F. Vicentini, A. Biella, N. Regnault, and C. Ciuti, Phys. Rev. Lett. 122, 250503 (2019). 10.1103/PhysRevLett.122.250503 Crossref, Google Scholar
- 41 N. Yoshioka and R. Hamazaki, Phys. Rev. B 99, 214306 (2019). 10.1103/PhysRevB.99.214306 Crossref, Google Scholar
- 42 N. Irikura and H. Saito, Phys. Rev. Res. 2, 013284 (2020). 10.1103/PhysRevResearch.2.013284 Crossref, Google Scholar
- 43 Y. Nomura, N. Yoshioka, and F. Nori, Phys. Rev. Lett. 127, 060601 (2021). 10.1103/PhysRevLett.127.060601 Crossref, Google Scholar
- 44 S. Sorella, Phys. Rev. B 64, 024512 (2001). 10.1103/PhysRevB.64.024512 Crossref, Google Scholar
- (45) Because the relative error of energy of the optimized RBM wave function for the 1D TFI model has a sharp peak at Γ = 1 [see Fig. 2(b)], the derivative of the energy of the RBM wave function with respect to the Γ parameters has a small discontinuity at Γ = 1. Accordingly, the property of the optimized RBM spin system changes discontinuously at Γ = 1 (thus, the Wtail value has a jump at Γ = 1, see Fig. 4). Therefore, although one can detect the quantum phase transition using the RBM, the critical behavior of the 1D TFI model at Γ = 1 cannot be captured in the current setup of the RBM. Google Scholar
- 46 A. Tanaka and A. Tomiya, J. Phys. Soc. Jpn. 86, 063001 (2017). 10.7566/JPSJ.86.063001 Link, Google Scholar
- 47 S. Arai, M. Ohzeki, and K. Tanaka, J. Phys. Soc. Jpn. 87, 033001 (2018). 10.7566/JPSJ.87.033001 Link, Google Scholar
- 48 Y. Miyajima, Y. Murata, Y. Tanaka, and M. Mochizuki, Phys. Rev. B 104, 075114 (2021). 10.1103/PhysRevB.104.075114 Crossref, Google Scholar