Restricted Boltzmann machine network versus Jastrow correlated wave function for the two-dimensional Hubbard model
Karthik V, Amal Medhi · 2024
Read Restricted Boltzmann machine network versus Jastrow correlated wave function for the two-dimensional Hubbard model by Karthik V, Amal Medhi (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-04103.txt.
Introduction
Karthik V and Amal Medhi present a 2024 physics paper pitting a neural network against a classic formula as the stand-in for a famous electron model. The two-dimensional Hubbard model captures electrons hopping on a grid and repelling on site, a minimal stage for high-temperature superconductivity. Both candidate wave functions describe its ground state; the question is which captures its phases better.
The neural version, a restricted Boltzmann machine layered on a pairing state, won on energy in the lightly doped region that matters most. It slightly trailed on superconducting correlations but, strikingly, produced strong antiferromagnetic patterns on its own, with no magnetism built in. The classic Jastrow versions barely magnetized at all. The authors read this as evidence that learned wave functions see more of the model's true character.
Catalog record
| CREATOR | Karthik V, Amal Medhi |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 25 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-04103
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Restricted Boltzmann machine network versus Jastrow correlated wave function for the two-dimensional Hubbard model Karthik V, Amal Medhi Abstract We consider a restricted Boltzmann Machine (RBM) correlated BCS wave function as the ground state of the two-dimensional Hubbard model and study its electronic and magnetic properties as a function of hole doping. We compare the results with those obtained by using conventional Jastrow projectors. The results show that the RBM wave function outperforms the Jastrow projected ones in the underdoped region in terms of the variational energy. Computation of superconducting (SC) correlations in the model shows that the RBM wave function gives slightly weaker SC correlations as compared to the Jastrow projected wave functions. A significant advantage of the RBM wave function is that it spontaneously gives rise to strong antiferromagnetic (AF) correlations in the underdoped region even though the wave function does not incorporate any explicit AF order. In comparison, AF correlations in the Jastrow projected wave functions are found to be very weak. These and other results obtained show that the RBM wave function provides an improved description of the phase diagram of the model. The work also demonstrates the power of neural-network quantum state (NQS) wave functions in the study of strongly correlated electron systems. I INTRODUCTION The neural-network quantum state (NQS) wave functions have recently been applied successfully to study several quantum many-body systems Carrasquilla and Melko 2017 ; Carleo and Troyer 2017 ; van Nieuwenburg et al. 2017 ; Broecker et al. 2017 ; Ch’ng et al. 2017 ; Deng et al. 2017 . The NQS wave functions are variational wave functions constructed based on an artificial neural-networks (ANNs). The power of such wave functions derives from the expressive capability of the ANNs which can represent highly nonlinear functions Hornik et al. 1989 . In contrast to the traditional Jastrow-type variational wave functions Yamaji et al. 1998 ; Edegger et al. 2007 ; Zhao et al. 2017 which are biased by the choice of a mean-field ansatz, NQS wave functions can be highly unbiased and can provide an accurate representation of quantum many-body states as has been demonstrated in several works Carleo and Troyer 2017 ; Choo et al. 2018 ; Carleo et al. 2019 ; Ch’ng et al. 2017 ; Carrasquilla 2020 ; Melko et al. 2019 ; Carrasquilla and Torlai 2021 . However, most of these applications have been limited to bosonic systems. For fermionic systems, there is a key disadvantage that though neural-networks can represent complicated functions, it generally fails to take into account the complicated sign structure of a fermionic many-body wave function Inui et al. 2021 ; Cai and Liu 2018 . Indeed, it fails to correctly represent the sign structures of the wave function of even non-interacting fermions Cai and Liu 2018 . An alternative approach in which the above problem is avoided is to use an ANN as a correlator to an antisymmetric one-body wave function which takes care of the fermionic symmetry correctly. Although such a wave function is less unbiased by construction, it still outperforms traditional Jastrow projected wave functions in terms of accuracy. For example, Nomura et al. Nomura et al. 2017 have used the restricted Boltzmann machine (RBM) network as a correlator to a generalized pair-product wave function. The resulting RBM+PP wave function was shown to give significantly lower energy compared to Jastrow type wave function for the two-dimensional Hubbard model at half-filling. In a recent work V. and Medhi 2024 , the present authors showed that a convolutional restricted Boltzmann machine (CRBM) correlated wave function provides a significantly improved description of the ground state of the half-filled Hubbard model on a square lattice. In this work, we consider an RBM correlated BCS wave function as a ground state of the two-dimensional Hubbard model and study it as a function of hole doping. We examine several electronic properties like ground state energy, superconducting and magnetic correlations etc. and compare the results with those obtained by using several standard Jastrow projectors. Specifically, the wave function we consider is of the form {\cal P}\ket{BCS}_{N} , where \ket{BCS}_{N} is the BCS state with fixed electron number and {\cal P} is the projector or the correlator which introduces electronic correlation into the wave function. We take it to be {\cal P}_{RBM} , a correlator constructed using an RBM network as well as three other Gutzwiller-Jastrow projectors. We study the resulting wave function using the variational Monte Carlo method Tahara and Imada 2008 ; Sorella 2013 ; SHIBA and YOKOYAMA 1987 . We show that compared to the Jastrow projectors, the RBM correlated wave function gives lower energy in the underdoped phase of the model. We obtain a ground state phase diagram of the model and show that the wave function give a better description of the superconducting and magnetic properties of the model. The rest of the paper is organized as follows. We describe the model and the variational wave functions considered in section II . The results are described in section III . Finally, we give the summary and concluding remarks in section IV . II Model and method We consider the fermionic Hubbard model on a two-dimensional (2D) square lattice given by, \displaystyle\mathcal{H}=-t\sum_{\langle i,j\rangle\sigma}\left(c^{\dagger}_{i\sigma}c_{j\sigma}+hc\right)+U\sum_{i}n_{i\uparrow}n_{i\downarrow} (1) The operator c^{\dagger}_{i\sigma} creates an electron at site ‘ i ’ with spin \sigma , and c_{i\sigma} is the corresponding annihilation operator. n_{i\sigma}=c^{\dagger}_{i\sigma}c_{i\sigma} is the number operator. In the first term, the summation is over the nearest-neighbor sites. The parameter t is the hopping amplitude and U is the onsite Hubbard repulsive potential. Here we take U/t=8 , a value that lies in the intermediate coupling strength regime of the model. The RBM correlated wave function we consider is given by, \displaystyle|\Psi_{RBM}\rangle={\cal P}_{RBM}\ket{BCS}_{N} (2) where \ket{BCS}_{N} is the BCS ground state in the subspace of the fixed number of particles ( 2N ), of the following mean-field pairing Hamiltonian, \displaystyle{\cal H}_{MF}= \displaystyle\sum_{\mathbf{k}\sigma}\varepsilon_{\mathbf{k}}c^{\dagger}_{\mathbf{k}\sigma}c_{\mathbf{k}\sigma}+\sum_{\mathbf{k}}\left(\Delta_{\mathbf{k}}c^{{\dagger}}_{\mathbf{k}\uparrow}c^{{\dagger}}_{-\mathbf{k}\downarrow}+hc\right) (3) where \varepsilon_{\mathbf{k}}=-2t(\cos k_{x}+\cos k_{y})-\mu . We take the superconducting (SC) pair amplitude to be \Delta_{\mathbf{k}}=\Delta_{SC}(\cos k_{x}-\cos k_{y}) ( d -wave pairing). The quantities \Delta_{SC} and \mu are the variational parameters in the mean-field part of the wave function. The BCS wave function in real-space representation can be written as, \displaystyle\ket{BCS}_{N}=\sum_{\{i_{1},\ldots i_{N},j_{1},\ldots,j_{N}\}}\begin{vmatrix}\phi(\mathbf{r}_{i_{1}}-\mathbf{r}_{j_{1}})&\cdots&\phi(\mathbf{r}_{i_{1}}-\mathbf{r}_{j_{N}})\\ \vdots&\cdots&\vdots\\ \phi(\mathbf{r}_{i_{N}}-\mathbf{r}_{j_{1}})&\cdots&\phi(\mathbf{r}_{i_{1}}-\mathbf{r}_{j_{N}})\\ \end{vmatrix} \displaystyle\times c^{\dagger}_{i_{1}\uparrow}\ldots c^{\dagger}_{i_{N}\uparrow}c^{\dagger}_{j_{1}\downarrow}\ldots c^{\dagger}_{j_{N}\downarrow}\ket{0} (4) The set \{i_{1},\ldots i_{N},j_{1},\ldots,j_{N}\} represent the set of sites occupied by the electrons. The pair amplitudes \phi(\mathbf{r}_{i}-\mathbf{r}_{j}) are functions of the variational parameters \Delta_{SC} and \mu Paramekanti et al. 2004 . Denoting the determinantal coefficients by \Psi_{BCS} , we can write, \displaystyle\ket{BCS}_{N}= \displaystyle\sum_{n_{1},n_{2},\ldots n_{2L}}\Psi_{BCS}(n_{1},n_{2},\ldots n_{2L}) \displaystyle\times(c^{{\dagger}}_{1\uparrow})^{n_{1}}\ldots(c^{{\dagger}}_{L\uparrow})^{n_{L}}(c^{{\dagger}}_{1\downarrow})^{n_{L+1}}\ldots(c^{{\dagger}}_{L\downarrow})^{n_{2L}}\ket{0} \displaystyle= \displaystyle\sum_{R}\Psi_{BCS}(R)\ket{R} (5) The occupation number n_{i} -s can values 0 or 1 . L is the number of lattice sites. The set R\equiv\{n_{1},n_{2},\ldots,n_{2L}\} specifies the electronic configurations. Now we consider a restricted Boltzmann machine (RBM) network as shown in Fig. 1 . Figure 1: Schematic diagram of a restricted Boltzmann machine (RBM) network. The input to the network are the electron occupation numbers \{n_{1},n_{2},\ldots,n_{2L}\} . The hidden layer defines variables h_{i} which can take values \pm 1 . The network consists of a visible layer with 2L neurons and a hidden layer with P=2L\alpha neurons. The parameter \alpha is called the hidden variable density. Let a_{i} , b_{j} be the bias parameters