betaFeedback

WorksPhysics and Astronomy

Quantum Monte Carlo calculations in the nuclear shell model by the complex Langevin method

Abstract

The nuclear shell model is known to describe the properties of various nuclei extremely well. However, the auxiliary-field quantum Monte Carlo calculations cannot be applied to it with general interactions due to the sign problem. The model has therefore been investigated primarily by variational methods, where the accuracy of the results depends crucially on the ansatz for the wave function. Here we perform the auxiliary-field quantum Monte Carlo calculations in the case of small systems at finite temperature using the complex Langevin method (CLM), which has been successfully applied to various interesting systems with the sign problem over the decade. In particular, we show the existence of a parameter region in which the validity criterion for the CLM is satisfied and the expectation value of the energy obtained by exact diagonalization is correctly reproduced. Thus the CLM can be a complementary approach to the variational method for large systems.

Introduction

The nuclear shell model is one of the most powerful theoretical models that describe not only low-lying energy spectra of nuclei including those near the neutron drip line [1–3] but also thermal nuclear properties [4]. In this model, a frozen inert core is assumed and active particles in the valence orbits are treated by the so-called model space. The quantum many-body problem of the active particles is then solved by diagonalizing the Hamiltonian matrix in the subspace spanned by a huge number of Slater determinants, each of which represents how the active particles occupy the single-particle states. However, the dimension of the Hamiltonian matrix increases combinatorially as the active particles and the single-particle states increase in number. While the conventional Lanczos diagonalization method can handle the system with the so-called MM-scheme for dimension up to O(1011)O(10^{11}) [5], the explosive growth of the dimension hampers the shell-model calculations, in particular, in the heavy-mass region.

In order to circumvent the problem of the explosive growth of the dimension, various efforts have been made. (See, for instance, Refs. [6, 7].) In 1990s, the auxiliary-field Monte Carlo (AFMC) method was introduced in the shell-model calculations. Based on this approach, which was named the shell model Monte Carlo (SMMC) method, pfpf-shell nuclei were studied successfully [8]. However, practical applications of the method turn out to be severely restricted by the notorious sign problem. In particular, a realistic effective interaction in the shell model, which is given, for instance, by the GG-matrix theory or by the valence-space in-medium similarity renormalization group (VS-IMSRG) method [9], cannot be used directly because of the sign problem. In order to circumvent this problem, one can prepare some artificial interaction without the sign problem and make an extrapolation [10], but such an extrapolation procedure often causes large uncertainties. This led to the common use of a schematic interaction without the sign problem instead [11, 12]. As another attempt, the constrained-path approximation was tested in sdsd- and pfpf-shell nuclei [13].

On the other hand, various attempts have been made to obtain approximate wave functions using variational approaches such as conventional particle-hole truncations, the VAMPIR method [14] and the Monte Carlo shell model [15]. While these methods are indeed useful in practical applications, they only provide us with the upper bound or the extrapolated value within the variational ansatz, and furthermore, they cannot be used for finite-temperature physics. Thus, it is still desirable to develop quantum Monte Carlo techniques, which do not have these problems of the variational approaches.

In this paper, we attempt to solve the sign problem in the shell-model calculations with realistic interactions by the complex Langevin method (CLM) [16,17]. (See Ref. [18] for an earlier attempt using the Lipkin model.) While the method is known to have the wrong convergence problem, the conditions for the correct convergence have recently been clarified [19–28] and various new techniques to satisfy these conditions have been proposed. Based on these new developments, the CLM has been applied successfully to lattice QCD at finite density [29–38]. (See Refs. [39–41] for recent reviews.) See also Ref. [42] for an application to the Hubbard model.

We discuss two formalisms, which differ in how we treat the projection onto a fixed number of active particles. Our results for finite temperature show that the correct expectation value of energy is obtained only for sufficiently high temperature, which can be understood by the validity criterion of the CLM. In order to obtain results at lower temperature, we propose to introduce a tunable parameter in the interaction terms and to make an extrapolation using the data obtained in the parameter region in which the validity criterion of the CLM is satisfied. Thus we conclude that the CLM can be a useful approach to larger nuclei, which is complementary to the variational method. We hope that this approach is useful also in abab initio nuclear structure theories based on the configuration interaction method, such as the no-core shell-model method [43] and the VS-IMSRG method [9].

This paper is organized as follows. In section 2, we briefly review how the CLM works and discuss the validity criterion of the method. In section 3, we discuss how we apply the CLM to the nuclear shell model, and present our results in section 4. Section 5 is devoted to a summary and discussions. In Appendix A, we discuss how to implement the projection onto a fixed number of active particles in one of the formalisms discussed in this paper. In Appendix B, we present some results for the case without the sign problem to confirm that the real Langevin simulation works in that case. In Appendix C, we discuss the gg-extrapolation as an alternative way to perform extrapolations from the region in which correct results are available.

Brief review of the complex Langevin method

In this section, we briefly review how the complex Langevin method works. For that, let us consider the integral

⟨O(X)⟩=1Z∫dX O(X)e−S(X),Z=∫dX e−S(X)(1)\langle O(X)\rangle= \frac{1}{Z}\int dX\,O(X)e^{-S(X)}, \qquad Z=\int dX\,e^{-S(X)} \tag*{(1)}

where XX represents a set of dynamical variables and S(X)S(X) represents its function. If S(X)S(X) is a real function, e−S(X)/Ze^{-S(X)}/Z can be interpreted as a probabilistic weight so that one can numerically evaluate the integral by stochastic methods.

One of such stochastic methods is the so-called (real) Langevin method, which is based on the equivalence between the Fokker-Planck equation and the Langevin equation. Let us consider the following Fokker-Planck equation

∂∂tP(X;t)=∂∂X(∂∂X−v(X))P(X;t)(2)\frac{\partial}{\partial t}P(X;t)=\frac{\partial}{\partial X}\left(\frac{\partial}{\partial X}-v(X)\right)P(X;t) \tag*{(2)}

where P(X;t)P(X;t) is the probability distribution for XX at time tt and v(X)v(X) is the drift term defined by v(X)=−∂S∂Xv(X)=-\frac{\partial S}{\partial X}. If SS is real and non-singular, the solution to the Fokker-Planck equation uniquely asymptotes to e−S(X)/Ze^{-S(X)}/Z in the large tt limit. It is well known that this Fokker-Planck equation is equivalent to the Langevin equation

dx(t)dt=v(x(t))+η(t)(3)\frac{dx(t)}{dt}=v(x(t))+\eta(t) \tag*{(3)}

where η(t)\eta(t) is the white noise that satisfies

⟨η(t)⟩η=0,⟨η(t)η(t′)⟩η=2δ(t−t′)(4)\langle\eta(t)\rangle_{\eta}=0,\qquad\langle\eta(t)\eta(t')\rangle_{\eta}=2\delta(t-t') \tag*{(4)}

with ⟨⋯ ⟩η\langle\cdots\rangle_{\eta} being an average with respect to the white noise, which can be generated by the Gaussian distribution ∝exp⁡[−14∫dt η(t)2]\propto\exp\left[-\frac{1}{4}\int dt\,\eta(t)^2\right]. The solution x(t)x(t) to the Langevin equation then corresponds to the solution P(X;t)P(X;t) to the Fokker-Planck equation through P(X;t)=⟨δ(X−x(t))⟩ηP(X;t)=\langle\delta(X-x(t))\rangle_{\eta}. Thus one just needs to solve the Langevin equation for a long enough tt to realize the desired probability distribution e−S(X)/Ze^{-S(X)}/Z.

The complex Langevin method is an extension of the real Langevin method to a system with a complex-valued S(x)S(x). In that case, it is natural to consider a complex version of the Langevin equation

dz(t)dt=v(z(t))+η(t),(5)\frac{dz(t)}{dt}=v(z(t))+\eta(t), \tag*{(5)}

where z(t)z(t) represents a set of complexified dynamical variables, v(z)=−∂S∂zv(z)=-\frac{\partial S}{\partial z} represents the drift term, and η(t)\eta(t) represents the real-valued white noise that satisfies (2.4). The equivalent Fokker-Planck equation can be obtained as

∂∂tP(Z,Zˉ;t)={∂∂Z(∂∂Zˉ−v(Z))+∂∂Zˉ(∂∂Z−v(Z)‾)+2∂2∂Z∂Zˉ}P(Z,Zˉ;t)=(∂2∂X2−∂∂XRe⁡[v(Z)]−∂∂YIm⁡[v(Z)])P(Z,Zˉ;t),(6)\frac{\partial}{\partial t}\mathcal{P}(Z,\bar{Z};t)=\left\{\frac{\partial}{\partial Z}\left(\frac{\partial}{\partial\bar{Z}}-v(Z)\right)+\frac{\partial}{\partial\bar{Z}}\left(\frac{\partial}{\partial Z}-\overline{v(Z)}\right)+2\frac{\partial^2}{\partial Z\partial\bar{Z}}\right\}\mathcal{P}(Z,\bar{Z};t) = \left(\frac{\partial^2}{\partial X^2}-\frac{\partial}{\partial X}\operatorname{Re}[v(Z)]-\frac{\partial}{\partial Y}\operatorname{Im}[v(Z)]\right)\mathcal{P}(Z,\bar{Z};t), \tag*{(6)}

where XX and YY are the real and imaginary parts of ZZ, respectively, and P(Z,Zˉ;t)\mathcal{P}(Z,\bar{Z};t) is the probability distribution for ZZ and Zˉ\bar{Z} at time tt. If the following equation holds

∫dX P(X;t)O(X)=∫dZ dZˉ P(Z,Zˉ;t)O(Z),(7)\int dX\,P(X;t)\mathcal{O}(X)=\int dZ\,d\bar{Z}\,\mathcal{P}(Z,\bar{Z};t)\mathcal{O}(Z), \tag*{(7)}

where P(X;t)P(X;t) is a complex-valued function that satisfies (2.2) with the complex-valued S(X)S(X), a solution to the complex Langevin equation (2.5) gives the expectation value of observables with the complex weight P(X;t)P(X;t).

Let us note first that, unlike the real Langevin case, P(X;t)P(X;t) does not necessarily converge to e−S/Ze^{-S}/Z. This is not a problem, however, as far as P(Z,Zˉ;t)\mathcal{P}(Z,\bar{Z};t) that satisfies (2.7) converges uniquely to some finite function in the t→∞t\to\infty limit since in that case, one can prove that the t→∞t\to\infty limit of P(X;t)P(X;t) should be e−S/Ze^{-S}/Z [23].

Thus the question boils down to whether (2.7) holds, and if so, under which condition. By formal integration of the Fokker-Planck equation (2.6), the right-hand side of (2.7) can be rewritten as [20, 21]

∫dZ dZˉ P(Z,Zˉ;t)O(Z)=∫dZ dZˉ {etLTP(Z,Zˉ;0)}O(Z),(8)\int dZ\,d\bar{Z}\,\mathcal{P}(Z,\bar{Z};t)\mathcal{O}(Z)=\int dZ\,d\bar{Z}\,\{e^{tL^{T}}\mathcal{P}(Z,\bar{Z};0)\}\mathcal{O}(Z), \tag*{(8)}

where LL is the Langevin operator for (2.5) given by

L=∂2∂X2+Re⁡[v(Z)]∂∂X+Im⁡[v(Z)]∂∂Y,(9)L=\frac{\partial^2}{\partial X^2}+\operatorname{Re}[v(Z)]\frac{\partial}{\partial X}+\operatorname{Im}[v(Z)]\frac{\partial}{\partial Y}, \tag*{(9)}

and its transpose LTL^{\mathsf{T}} reads

LT=∂2∂X2−∂∂XRe⁡[v(Z)]−∂∂YIm⁡[v(Z)] .(10)L^{\mathsf{T}} = \frac{\partial^2}{\partial X^2} - \frac{\partial}{\partial X}\operatorname{Re}[v(Z)] - \frac{\partial}{\partial Y}\operatorname{Im}[v(Z)] \, . \tag*{(10)}

Let us here assume that there is no boundary term in the following integration by parts

∫dZ dZˉ {etLTP(Z,Zˉ;0)}O(Z)=∫dZ dZˉ P(Z,Zˉ;0){etLO(Z)} ,(11)\int dZ\,d\bar{Z}\,\{e^{tL^{\mathsf{T}}}\mathcal{P}(Z,\bar{Z};0)\}\mathcal{O}(Z) = \int dZ\,d\bar{Z}\,\mathcal{P}(Z,\bar{Z};0)\{e^{tL}\mathcal{O}(Z)\} \, , \tag*{(11)}

where the right-hand side is well-defined in the t→∞t \to\infty limit. We also assume that the operator O(Z)\mathcal{O}(Z) is a holomorphic function of ZZ and that the initial condition is P(Z,Zˉ;0)=P(X;0)δ(Y)\mathcal{P}(Z,\bar{Z};0) = P(X;0)\delta(Y). Under these assumptions, one can derive (2.7) as

∫dZ dZˉ P(Z,Zˉ;0){etLO(Z)}=∫dX P(X;0){etL0O(X)}=∫dX {etL0TP(X;0)}O(X)=∫dX P(X;t)O(X) ,(12)\begin{aligned} \int dZ\,d\bar{Z}\,\mathcal{P}(Z,\bar{Z};0)\{e^{tL}\mathcal{O}(Z)\} &= \int dX\,P(X;0)\{e^{tL_0}\mathcal{O}(X)\} \\ &= \int dX\,\{e^{tL_0^{\mathsf{T}}}P(X;0)\}\mathcal{O}(X) \\ &= \int dX\,P(X;t)\mathcal{O}(X) \, , \tag*{(12)} \end{aligned}

where we have used LO(Z)∣Y=0=L0O(X)L\mathcal{O}(Z)|_{Y=0}=L_0\mathcal{O}(X) for holomorphic O(Z)\mathcal{O}(Z) and defined

L0=∂2∂X2+v(X)∂∂X ,L0T=∂2∂X2−∂∂Xv(X) .(13)L_0 = \frac{\partial^2}{\partial X^2} + v(X)\frac{\partial}{\partial X} \, , \qquad L_0^{\mathsf{T}} = \frac{\partial^2}{\partial X^2} - \frac{\partial}{\partial X}v(X) \, . \tag*{(13)}

Thus the key point in justifying the CLM is the validity of (2.11), which depends on the model as well as its parameter region. There are some criteria [44] [21, 25, 28] for the absence of the boundary term in (2.11). In this paper, we use the one proposed in Ref. [25], which is a practical and convenient criterion based on a sufficient condition for (2.11). This criterion focuses on the probability distribution of the magnitude of the drift term v(z)v(z) generated by P(Z,Zˉ;t)\mathcal{P}(Z,\bar{Z};t). If the probability distribution falls off exponentially or faster at large magnitude, the formal expression of the right-hand side in (2.11) is well defined and the boundary term in the integration by parts vanishes.

In practice, we simulate the model with the discretized complex Langevin equation

z(t+Δt)=z(t)+v(z(t))Δt+η~(t)Δt ,(14)z(t+\Delta t) = z(t) + v(z(t))\Delta t + \widetilde{\eta}(t)\sqrt{\Delta t} \, , \tag*{(14)}

where η~(t)\widetilde{\eta}(t) is generated by the Gaussian distribution ∝exp⁡[−14η~(t)2]\propto\exp[-\frac{1}{4}\widetilde{\eta}(t)^2], monitoring the maximum value of the magnitude of the drift term. The expectation value of an observable is obtained by the average of the observable measured at each Langevin time after thermalization. We check the validity criterion by making a histogram for the maximum value of the magnitude of the drift term.

Even when the CLM is supposed to work, it may fail because of some instability in the time-evolution by the discretized complex Langevin equation. This can be controlled by using an adaptive step-size for the discretized time-evolution, which depends on the magnitude of the drift term [19]. In this paper, we choose the step-size Δt\Delta t such that

Δt=min⁡(Δt0,Δt0v0max⁡(∣v(z(t))∣)),\Delta t=\min\left(\Delta t_0,\Delta t_0\frac{v_0}{\max(\lvert v(z(t))\rvert)}\right),

where Δt0\Delta t_0 and v0v_0 are positive parameters that we set at the beginning of a simulation. Since the model we simulate has more than one dynamical variables, we use max⁡(∣v(z(t))∣)\max(\lvert v(z(t))\rvert) for the adaptive step-size, which is the maximum value of the absolute value of the elements of the drift term. Note that one needs to take the adaptive step-size into account when one makes the drift histogram for the criterion.

Applying the CLM to the shell model

In this section we apply the CLM reviewed in the previous section to the nuclear shell-model calculations. After defining the shell model, we discuss how it can be investigated by the CLM.

Let us consider a many-body problem of the active particles in the model space, in which the number of the single-particle states and that of the active particles are NsN_s and NvN_v, respectively. The many-body shell-model Hamiltonian up to the two-body interaction is written in the density-decomposed form as [8]

H^=∑i,j=1NsTijc^i†c^j+12∑αVα(O^α)2,(15)\hat{H}=\sum_{i,j=1}^{N_s}T_{ij}\hat{c}_i^\dagger\hat{c}_j+\frac{1}{2}\sum_{\alpha}V_{\alpha}\left(\hat{O}_{\alpha}\right)^2, \tag*{(15)}

where c^i†\hat{c}_i^\dagger represents the creation operator of a single-particle state and O^α\hat{O}_{\alpha} represents the density operator of the form c^i†c^j\hat{c}_i^\dagger\hat{c}_j. We have introduced TijT_{ij}, which denotes the strength of the one-body term, and VαV_{\alpha}, which denotes the strength of the two-body interaction.

We simulate this theory at finite temperature by using the standard imaginary-time formalism, where the imaginary time β\beta, which corresponds to the inverse temperature, is sliced up into nβn_{\beta} segments. The partition function for a fixed number NvN_v of valence nucleons is written as

Z=Tr⁡Nv[(e−ΔβH^)nβ],(16)Z=\operatorname{Tr}_{N_v}\left[\left(e^{-\Delta\beta\hat{H}}\right)^{n_{\beta}}\right], \tag*{(16)}

where Δβ=β/nβ\Delta\beta=\beta/n_{\beta} and Tr⁡Nv\operatorname{Tr}_{N_v} is a trace in the Fock space with NvN_v particles. Here we use the Hubbard-Stratonovich transformation at each imaginary time to linearize the Hamiltonian with respect to O^α\hat{O}_{\alpha} by introducing new auxiliary scalar fields σαn\sigma_{\alpha n}, where nn runs from 1 to nβn_{\beta}. Then the partition function becomes [8]

Z≃∫[dσ](Tr⁡NvU^)e−12Δβ∑αn∣Vα∣σαn2,(17)Z\simeq\int[d\sigma]\left(\operatorname{Tr}_{N_v}\hat{U}\right)e^{-\frac{1}{2}\Delta\beta\sum_{\alpha n}\lvert V_{\alpha}\rvert\sigma_{\alpha n}^{2},} \tag*{(17)}
U^=e−Δβh^nβe−Δβh^nβ−1⋯e−Δβh^1,(18)\hat{U}=e^{-\Delta\beta\hat{h}_{n_\beta}}e^{-\Delta\beta\hat{h}_{n_{\beta-1}}}\cdots e^{-\Delta\beta\hat{h}_1}, \tag*{(18)}
h^n=∑i,jTijc^i†c^j+∑αsαVασαnO^α,(19)\hat{h}_n=\sum_{i,j}T_{ij}\hat{c}_i^\dagger\hat{c}_j+\sum_{\alpha}s_\alpha V_\alpha\sigma_{\alpha n}\hat{O}_\alpha, \tag*{(19)}

where sα=1s_\alpha=1 for Vα<0V_\alpha<0 and sα=is_\alpha=i for Vα>0V_\alpha>0. The integration measure for σ\sigma in (17) is defined by

[dσ]=∏α,n(dσαnΔβ∣Vα∣2π).[d\sigma]=\prod_{\alpha,n}\left(d\sigma_{\alpha n}\sqrt{\frac{\Delta\beta|V_\alpha|}{2\pi}}\right).

The expectation value of an observable O^\hat{O} with fixed NvN_v can be written as

⟨O⟩Nv=1Z∫[dσ]Tr⁡NvU^O^Tr⁡NvU^(Tr⁡NvU^)e−12Δβ∑αn∣Vα∣σαn2.(20)\langle O\rangle_{N_v}=\frac{1}{Z}\int[d\sigma]\frac{\operatorname{Tr}_{N_v}\hat{U}\hat{O}}{\operatorname{Tr}_{N_v}\hat{U}}\left(\operatorname{Tr}_{N_v}\hat{U}\right)e^{-\frac{1}{2}\Delta\beta\sum_{\alpha n}|V_\alpha|\sigma_{\alpha n}^{2}}. \tag*{(20)}

Let us call this formalism (20) the canonical formalism in this paper.

In fact, the canonical formalism is not very convenient since Tr⁡NvU^\operatorname{Tr}_{N_v}\hat{U} takes different forms for different NvN_v, and it becomes more and more complicated as NvN_v increases. From this point of view, it is beneficial to rewrite the path integral into the grand canonical ensemble with the number projection [8]. Using the projection onto a fixed NvN_v given by

δ(N^−Nv)=∫02πdϕ2πe(βμ+iϕ)(N^−Nv),(21)\delta(\hat{N}-N_v)=\int_{0}^{2\pi}\frac{d\phi}{2\pi}e^{(\beta\mu+i\phi)(\hat{N}-N_v)}, \tag*{(21)}

we rewrite (20) as

⟨O⟩Nv=1Z∫[dσ]∫dϕ2πTr⁡[e(βμ+iϕ)(N^−Nv)U^O^]e−12Δβ∑αn∣Vα∣σαn2,Z=∫[dσ]∫dϕ2πTr⁡[e(βμ+iϕ)(N^−Nv)U^]e−12Δβ∑αn∣Vα∣σαn2,(22)\begin{aligned} \langle O\rangle_{N_v}&=\frac{1}{Z}\int[d\sigma]\int\frac{d\phi}{2\pi}\operatorname{Tr}\left[e^{(\beta\mu+i\phi)(\hat{N}-N_v)}\hat{U}\hat{O}\right]e^{-\frac{1}{2}\Delta\beta\sum_{\alpha n}|V_\alpha|\sigma_{\alpha n}^{2}},\\ Z&=\int[d\sigma]\int\frac{d\phi}{2\pi}\operatorname{Tr}\left[e^{(\beta\mu+i\phi)(\hat{N}-N_v)}\hat{U}\right]e^{-\frac{1}{2}\Delta\beta\sum_{\alpha n}|V_\alpha|\sigma_{\alpha n}^{2}}, \tag*{(22)} \end{aligned}

where N^\hat{N} is the number operator of valence nucleons and μ\mu is the chemical potential. Here Tr⁡\operatorname{Tr} is a trace in the entire Fock space of the model space, which sums over all possible NvN_v. In the CLM, μ\mu is redundant since ϕ\phi is a complex dynamical variable, and therefore we set μ=0\mu=0 from now on. It is straightforward to rewrite the weight Tr⁡eiϕ(N^−Nv)U^\operatorname{Tr}e^{i\phi(\hat{N}-N_v)}\hat{U} for the grand canonical ensemble by single-particle operators as

Tr⁡eiϕ(N^−Nv)U^=e−iϕNvdet⁡(1+eiϕU),(23)\operatorname{Tr}e^{i\phi(\hat{N}-N_v)}\hat{U}=e^{-i\phi N_v}\det\left(1+e^{i\phi}U\right), \tag*{(23)}

where UU is a one-body matrix defined by Uij=(exp⁡[−βh(tot)])ijU_{ij}=(\exp[-\beta h^{(\mathrm{tot})}])_{ij} with the matrix h(tot)h^{(\mathrm{tot})} defined by U^=exp⁡[−β∑i,jhij(tot)c^i†c^j]\hat{U}=\exp[-\beta\sum_{i,j}h_{ij}^{(\mathrm{tot})}\hat{c}_i^\dagger\hat{c}_j]. Thus, the expectation value in the case of a one-body operator O^\hat{O} can be written as

⟨O⟩Nv=1Z∫[dσ]∫dϕ2πe−iϕNvdet⁡(1+eiϕU)e−12Δβ∑αn∣Vα∣σαn2tr⁡[UO],\langle O\rangle_{N_v}=\frac{1}{Z}\int[d\sigma]\int\frac{d\phi}{2\pi}e^{-i\phi N_v}\det\left(1+e^{i\phi}U\right)e^{-\frac{1}{2}\Delta\beta\sum_{\alpha n}|V_\alpha|\sigma_{\alpha n}^{2}}\operatorname{tr}[UO],
Z=∫[dσ]∫dϕ2πe−iϕNvdet⁡(1+eiϕU)e−12Δβ∑αn∣Vα∣σαn2(24)Z=\int[d\sigma]\int\frac{d\phi}{2\pi}e^{-i\phi N_v}\det\left(1+e^{i\phi}U\right)e^{-\frac{1}{2}\Delta\beta\sum_{\alpha n}|V_\alpha|\sigma_{\alpha n}^{2}} \tag*{(24)}

where tr⁡\operatorname{tr} is the matrix trace for Ns×NsN_s\times N_s matrices and we have defined

U=(1+eiϕU)−1eiϕU.(25)\mathcal{U}=\left(1+e^{i\phi}U\right)^{-1}e^{i\phi}U . \tag*{(25)}

Let us call (24) the grand-canonical formalism in this paper.

This formula (24) can be generalized to the case in which O^\hat{O} is a product of one-body operators O^1,⋯ ,O^n\hat{O}_1,\cdots,\hat{O}_n using

Tr⁡[eiϕ(N^−Nv)U^O^1⋯O^n]=e−iϕNv∂∂ϵ1⋯∂∂ϵndet⁡(1+eiϕUϵ1,…,ϵn)∣ϵ1=⋯=ϵn=0,(26)\operatorname{Tr}\left[e^{i\phi(\hat{N}-N_v)}\hat{U}\hat{O}_1\cdots\hat{O}_n\right] =e^{-i\phi N_v}\left.\frac{\partial}{\partial\epsilon_1}\cdots\frac{\partial}{\partial\epsilon_n}\det\left(1+e^{i\phi}U_{\epsilon_1,\ldots,\epsilon_n}\right)\right|_{\epsilon_1=\cdots=\epsilon_n=0}, \tag*{(26)}

where U^ϵ1,…,ϵn=exp⁡[−β∑i,jhij(tot)c^i†c^j]exp⁡[ϵ1O^1]⋯exp⁡[ϵnO^n]\hat{U}_{\epsilon_1,\ldots,\epsilon_n}=\exp\left[-\beta\sum_{i,j}h_{ij}^{(\mathrm{tot})}\hat{c}_i^\dagger\hat{c}_j\right]\exp[\epsilon_1\hat{O}_1]\cdots\exp[\epsilon_n\hat{O}_n] corresponding to a one-body matrix (Uϵ1,…,ϵn)ij=(exp⁡[−βh(tot)]exp⁡[ϵ1O1]⋯exp⁡[ϵnOn])ij(U_{\epsilon_1,\ldots,\epsilon_n})_{ij}=\left(\exp[-\beta h^{(\mathrm{tot})}]\exp[\epsilon_1O_1]\cdots\exp[\epsilon_nO_n]\right)_{ij}. For instance, the energy (15), which involves two-body operators, can be written as

⟨H⟩Nv=1Z∫[dσ]∫dϕ2πe−iϕNvdet⁡(1+eiϕU)e−12Δβ∑αn∣Vα∣σαn2(tr⁡[UT]+12∑αVα{(tr⁡[UOα])2+tr⁡[UOα2]−tr⁡[UOαUOα]}).(27)\langle H\rangle_{N_v}=\frac{1}{Z}\int[d\sigma]\int\frac{d\phi}{2\pi}e^{-i\phi N_v}\det\left(1+e^{i\phi}U\right)e^{-\frac{1}{2}\Delta\beta\sum_{\alpha n}|V_\alpha|\sigma_{\alpha n}^{2}} \left(\operatorname{tr}[U T]+\frac{1}{2}\sum_\alpha V_\alpha\left\{\left(\operatorname{tr}[UO_\alpha]\right)^2+\operatorname{tr}[UO_\alpha^2]-\operatorname{tr}[UO_\alpha UO_\alpha]\right\}\right). \tag*{(27)}

The standard procedure in the literature is to reduce the integration over ϕ\phi to a summation over NsN_s integers [8]. In the CLM, however, we leave the integration over ϕ\phi as it is and treat ϕ\phi as another integration variable.

In our simulations, we change the normalization of the variables as σαn=σ~αnΔβ∣Vα∣\sigma_{\alpha n}=\frac{\widetilde{\sigma}_{\alpha n}}{\sqrt{\Delta\beta|V_\alpha|}}. The effective action then becomes

S=∑α,nσ~αn22−ln⁡det⁡[1+eiϕU]+iNvϕ.(28)S=\sum_{\alpha,n}\frac{\widetilde{\sigma}_{\alpha n}^{2}}{2}-\ln\det\left[1+e^{i\phi}U\right]+iN_v\phi. \tag*{(28)}

Using this effective action, the Langevin equations are obtained as

dσ~αndt=−∂S∂σ~αn+ηαn(t),dϕdt=−∂S∂ϕ+ηϕ(t),(29)\frac{d\widetilde{\sigma}_{\alpha n}}{dt}=-\frac{\partial S}{\partial\widetilde{\sigma}_{\alpha n}}+\eta_{\alpha n}(t),\qquad \frac{d\phi}{dt}=-\frac{\partial S}{\partial\phi}+\eta_\phi(t), \tag*{(29)}

where ηαn\eta_{\alpha n} and ηϕ\eta_\phi represent the Gaussian noise, which satisfies ⟨ηαn(t)ηβm(0)⟩=2δαβδnmδ(t)\langle\eta_{\alpha n}(t)\eta_{\beta m}(0)\rangle=2\delta_{\alpha\beta}\delta_{nm}\delta(t) and ⟨ηϕ(t)ηϕ(0)⟩=2δ(t)\langle\eta_\phi(t)\eta_\phi(0)\rangle=2\delta(t). The drift terms are calculated as

∂S∂σ~αn=σ~αn−tr⁡[(1+eiϕU)−1eiϕ∂U∂σ~αn]=σ~αn−tr⁡[UU−1∂U∂σ~αn],\frac{\partial S}{\partial\widetilde{\sigma}_{\alpha n}} =\widetilde{\sigma}_{\alpha n}-\operatorname{tr}\left[\left(1+e^{i\phi}U\right)^{-1}e^{i\phi}\frac{\partial U}{\partial\widetilde{\sigma}_{\alpha n}}\right] =\widetilde{\sigma}_{\alpha n}-\operatorname{tr}\left[U U^{-1}\frac{\partial U}{\partial\widetilde{\sigma}_{\alpha n}}\right],
∂S∂ϕ=−itr⁡[(1+eiϕU)−1eiϕU]+iNv=−i(tr⁡[U]−Nv).(30)\frac{\partial S}{\partial\phi} =-i\operatorname{tr}\left[\left(1+e^{i\phi}U\right)^{-1}e^{i\phi}U\right]+iN_v =-i\left(\operatorname{tr}[U]-N_v\right). \tag*{(30)}

See Appendix § 7 for the implementation of the canonical formalism in the CLM.

The computational cost for the grand-canonical formalism1 grows as O((Ns)5nβ)O((N_s)^5n_\beta). The dominant part comes from the calculation of the drift terms (30) in the complex Langevin equation, which contains the calculations of UU, U\mathcal{U} and ∂U∂σ~αn\frac{\partial U}{\partial\widetilde{\sigma}_{\alpha n}}. The calculation of UU involves matrix multiplications in the imaginary-time direction, which cost O((Ns)3nβ)O((N_s)^3n_\beta). The calculation of U\mathcal{U} involves taking the inverse of Ns×NsN_s \times N_s matrices, which costs O((Ns)3)O((N_s)^3). The most dominant part comes from the calculation of ∂U∂σ~αn\frac{\partial\mathcal{U}}{\partial\widetilde{\sigma}_{\alpha n}} since it involves matrix multiplications at each α=1,…,Nα\alpha= 1,\ldots,N_\alpha, which costs O(Nα(Ns)3nβ)O(N_\alpha(N_s)^3n_\beta). Using Nα=(Ns)2N_\alpha= (N_s)^2, which is the case in our implementation, the computational cost becomes O((Ns)5nβ)O((N_s)^5n_\beta).

The results of the CLM

In this section, we present our numerical results for the shell model obtained by the CLM. We focus on the case with Ns=20N_s = 20 in the pfpf shell, considering only valence neutrons for simplicity. The interaction we use in the Hamiltonian (15) is taken from GXPF1A in [45]. Throughout this paper, we set the Langevin step-size to Δt=5×10−3\Delta t = 5 \times10^{-3}, which is found to be small enough for the accuracy required in this paper.

Dependence on the temperature

Let us discuss how our results depend on the temperature. In Figure 1, we show our results for Nv=2N_v = 2 in the grand-canonical formalism with various inverse temperature β\beta. The number of steps in the imaginary time direction is set to nβ=10n_\beta= 10. In the Top panel, we plot the expectation value of the energy in units of MeV against β\beta in units of MeV−1\mathrm{MeV}^{-1}. We observe good agreement with the results obtained by the Hamiltonian diagonalization for β≲0.6\beta\lesssim0.6. On the other hand, the energy ceases to decrease monotonically for β≳0.6\beta\gtrsim0.6, which represents an unphysical behavior.

Image not extracted

Figure 1. (Top) The expectation value of the energy is plotted against the inverse temperature β\beta in the grand-canonical formalism with Ns=20N_s = 20, Nv=2N_v = 2, nβ=10n_\beta= 10. The purple points with error bars represent the CLM results, whereas the gray points represent the results obtained by the Hamiltonian diagonalization. (Bottom) The drift histogram for the ϕ\phi field (Left) and the σ\sigma field (Right).

In the Bottom panel, we show the histogram of the drift terms. Since the histogram for β=0.3\beta= 0.3 seems to fall off exponentially for both ϕ\phi and σ\sigma fields, we can safely say that the CLM for β≲0.3\beta\lesssim0.3 is justified, which is consistent with the agreement with the results of the Hamiltonian diagonalization in that region. On the other hand, the drift histogram for the σ\sigma field does not show a fast fall-off for β≳0.4\beta\gtrsim0.4, which suggests that the agreement with the results of the Hamiltonian diagonalization in the 0.4≲β≲0.60.4 \lesssim\beta\lesssim0.6 is rather accidental.

Let us recall that the grand-canonical formalism allows us to obtain results for different NvN_v in a straightforward manner unlike the canonical formalism. In Figure 2 we show our results for Nv=10N_v = 10 with nβ=10n_\beta= 10. We observe good agreement with the result of the Hamiltonian diagonalization for β≲0.4\beta\lesssim0.4, which is consistent with the result of the drift histogram.

Image not extracted

Figure 2. Similarly to Figure 1, we present the results for the grand-canonical formalism with Ns=20N_s = 20, Nv=10N_v = 10, nβ=10n_\beta= 10.

As we have seen above, the CLM works only at sufficiently small β\beta (high temperature), at least for the Ns=20N_s = 20 case with valence neutrons. However, if we are interested in the ground state energy, we have to take the β→∞\beta\to\infty limit. Since the result for finite β\beta is a weighted sum over the low-lying energy levels EnE_n with the Boltzmann weight e−βEne^{-\beta E_n}, it is not straightforward to extract the energy levels EnE_n from the observed β\beta dependence of the energy expectation values.

In Appendix B, we present some results for the case without the sign problem, where we find that the real Langevin simulation works as we discussed in § 2.

Extrapolation with respect to the interaction term

In this section, we attempt to extract the results for larger β\beta (lower temperature) by introducing an additional parameter and making an extrapolation.2

Here we introduce a tunable parameter tt in the Hamiltonian (3.1) as

H^t=∑i,jTijc^i†c^j+t2∑αVα(O^α)2.(31)\hat{H}_t = \sum_{i,j} T_{ij}\hat{c}_i^\dagger\hat{c}_j + \frac{t}{2}\sum_\alpha V_\alpha\left(\hat{O}_\alpha\right)^2 . \tag*{(31)}

Since the observable at small tt is described by perturbation theory around t=0t = 0, it is given by a polynomial in tt in that region. Hence we can make an extrapolation to t=1t = 1 using a polynomial as a natural fitting ansatz. To obtain the expectation value of the energy, we calculate the expectation value of the original Hamiltonian H^t=1\hat{H}_{t=1} for various tt.

In Figure 3 (Top-Left), we plot the expectation value of the energy against the tunable parameter tt for Nv=2N_v = 2 and β=1\beta= 1 in the canonical and grand-canonical formalisms with for g≤0g \le0 to g=1g = 1. See Appendix § 9 for more details.

Image not extracted

Figure 3. (Top-Left) The expectation value of the energy is plotted against the tunable parameter tt in the canonical and grand-canonical formalisms with Ns=20N_s = 20, Nv=2N_v = 2, β=1\beta= 1, nβ=10n_\beta= 10. (Top-Right) The drift histogram for the σ\sigma field in the canonical formalism. (Bottom) The drift histogram for the ϕ\phi field (Left) and the σ\sigma field (Right) in the grand-canonical formalism.

nβ=10n_{\beta}=10. We find good agreement with the result of the Hamiltonian diagonalization for t≲0.6t \lesssim0.6. The drift histogram indicates that the region of validity is t≲0.4t \lesssim0.4 for the grand-canonical formalism and t≲0.5t \lesssim0.5 for the canonical formalism.

We fit the data in these regions to a quadratic function in Figure 4. The fitting curve obtained in this way is consistent with the result obtained by the Hamiltonian diagonalization within the fitting errors up to the target value t=1t=1. Thus we can obtain the expectation value of the energy at β=1\beta=1. Let us note that the difference of the energy of the ground state and the first excited state is δE≈1.438(MeV)\delta E \approx1.438(\mathrm{MeV}) in the present case, which means that e−βδE≈0.237⋯e^{-\beta\delta E} \approx0.237\cdots. In order to extract the ground state energy, it would be desirable to obtain results up to β∼3\beta\sim3. The fitting ansatz in such cases would need high orders in tt, which makes the extrapolation less reliable.

Image not extracted

Figure 4. The expectation value of the energy is plotted against the tunable parameter tt for the canonical (Left) and grand-canonical (Right) formalisms, with Ns=20N_s=20, Nv=2N_v=2, β=1\beta=1 and nβ=10n_{\beta}=10. The solid orange line represents a fit to a quadratic function, using the data within 0≤t≤0.50 \leq t \leq0.5 for the canonical formalism and the data within 0≤t≤0.40 \leq t \leq0.4 for the grand-canonical formalism. The shaded region represents the fitting errors. The gray points represent the results obtained by the Hamiltonian diagonalization, which are consistent with the fitting curves within the fitting errors.

Summary

We have discussed how the nuclear shell model can be investigated without the variational ansatz by applying the CLM, which overcomes the sign problem in quantum Monte Carlo methods. In particular, we find that the grand-canonical formalism (24) is convenient to simulate the case with various NvN_v including the region Nv∼Ns/2N_v \sim N_s/2, where the canonical formalism (20) becomes too complicated to implement.

As a first step, we have performed explicit calculations for 40+NvCa^{40+N_v}\mathrm{Ca} systems (Ns=20N_s = 20), with a realistic potential at finite temperature. Our results agree with the results of the Hamiltonian diagonalization at high temperature but not at low temperature. This has been understood from the viewpoint of the validity criterion [25], which is a sufficient condition for the correct convergence of the CLM. The temperature at which the agreement ceases to hold is slightly lower than the temperature at which the drift histogram starts to have a slow fall-off. Thus by probing the drift histogram, we can safely estimate the validity region of the CLM calculations.

In order to obtain results at lower temperature, which is not directly accessible by the CLM, we have proposed to introduce a tunable parameter tt in the interaction terms so that the original Hamiltonian is obtained at t=1t = 1. In the case we studied, the region of tt in which the CLM works is large enough to make an extrapolation to t=1t = 1 by a quadratic fit. The result obtained by this extrapolation is consistent with the result obtained by the Hamiltonian diagonalization within the fitting errors. The extrapolation turns out to be easier for the canonical formalism (20) than the grand-canonical formalism (24) due to a slightly larger validity region of the CLM.

For heavier nuclei, the computational cost increases as ∼(Ns)5(Nv)0nβ\sim(N_s)^5(N_v)^0 n_\beta as we discussed at the end of § 3. While the sign problem becomes severer, it is possible that the validity region of the CLM is not significantly affected by the system size. We therefore expect that the CLM can be a useful approach to heavy nuclei, which is complementary to the variational method.

It is also worthwhile to mention a potential relation of our study to quantum computing. Recent developments in quantum computing provide another possible route to addressing the rapid growth of the Hilbert space in nuclear shell-model calculations. While present-day quantum devices are still limited in the size and accuracy of feasible calculations, quantum algorithms have already been investigated for nuclear structure and shell-model problems [46, 47]. Classical methods with polynomial computational scaling, such as the approach investigated here, may also provide useful benchmarks for future quantum calculations.

Acknowledgment

The computations were carried out on the PC clusters in KEK Computing Research Center and KEK Theory Center. This research was supported by MEXT as “Program for Promoting Researches on the Supercomputer Fugaku” (Simulation for basic science: from fundamental laws of particles to creation of nuclei). It is also supported by Joint Institute for Computational Fundamental Science (JICFuS). Y.A. was supported by JSPS KAKENHI Grant Number JP24K07036.

Canonical formalism

In this appendix, we discuss how to implement the canonical formalism in the CLM. As we did for the grand-canonical formalism, we change variables as σαn=σ~αnΔβ∣Vα∣\sigma_{\alpha n} = \frac{\widetilde{\sigma}_{\alpha n}}{\sqrt{\Delta\beta\lvert V_\alpha\rvert}} for simplicity.

We can write the effective action of the path integral (17) as

S=∑α,nσ~αn22−ln⁡Tr⁡NvU^.(32)S = \sum_{\alpha,n} \frac{\widetilde{\sigma}_{\alpha n}^{2}}{2} - \ln\operatorname{Tr}_{N_v} \widehat{U}. \tag*{(32)}

The complex Langevin equation is then written as

dσ~αndt=−∂S∂σ~αn+ηαn(t),(33)\frac{\mathrm{d}\widetilde{\sigma}_{\alpha n}}{\mathrm{d}t} = -\frac{\partial S}{\partial\widetilde{\sigma}_{\alpha n}} + \eta_{\alpha n}(t), \tag*{(33)}

where ηαn\eta_{\alpha n} is a Gaussian noise term that satisfies ⟨ηαn(t)ηβm(0)⟩=2δαβδnmδ(t)\langle\eta_{\alpha n}(t)\eta_{\beta m}(0)\rangle= 2\delta_{\alpha\beta}\delta_{nm}\delta(t). Unlike the grand-canonical formalism, the effective action (A.1) involves a trace in the subspace of the Fock space with NvN_v particles. Therefore, in order to derive a more explicit form of the drift term, we need to specify the value of NvN_v and to transform Tr⁡Nv\operatorname{Tr}_{N_v} into an expression written in terms of matrix traces. For instance, in the case of Nv=2N_v = 2, Tr⁡2U^\operatorname{Tr}_{2}\widehat{U} can be rewritten as

Tr⁡2U^=12[(tr⁡U)2−tr⁡U2],(34)\operatorname{Tr}_{2}\widehat{U} = \frac{1}{2}\left[(\operatorname{tr}U)^2 - \operatorname{tr}U^2\right], \tag*{(34)}

where UU is the one-body matrix defined by Uij=(exp⁡[−βh(tot)])ijU_{ij} = (\exp[-\beta h^{(\mathrm{tot})}])_{ij}, and the drift term becomes

∂S∂σ~αn=σ~αn−2(tr⁡U)2−tr⁡U2tr⁡[(tr⁡U)∂U∂σ~αn−(U∂U∂σ~αn)].(35)\frac{\partial S}{\partial\widetilde{\sigma}_{\alpha n}} = \widetilde{\sigma}_{\alpha n} - \frac{2}{(\operatorname{tr}U)^2-\operatorname{tr}U^2}\operatorname{tr}\left[(\operatorname{tr}U)\frac{\partial U}{\partial\widetilde{\sigma}_{\alpha n}}-\left(U\frac{\partial U}{\partial\widetilde{\sigma}_{\alpha n}}\right)\right]. \tag*{(35)}

The derivative of UU with respect to σ~αn\widetilde{\sigma}_{\alpha n} is written explicitly as

∂U∂σ~αn=sα∗Δβ∣Vα∣e−Δβhnβ⋯e−Δβhn+1(e−Δβadj⁡hn−1−Δβadj⁡hnOα)e−Δβhn⋯e−Δβh1,(36)\frac{\partial U}{\partial\widetilde{\sigma}_{\alpha n}} = s_\alpha^*\sqrt{\Delta\beta\lvert V_\alpha\rvert}e^{-\Delta\beta h_{n\beta}}\cdots e^{-\Delta\beta h_{n+1}}\left(\frac{e^{-\Delta\beta\operatorname{adj}h_n}-1}{-\Delta\beta\operatorname{adj}h_n}O_\alpha\right)e^{-\Delta\beta h_n}\cdots e^{-\Delta\beta h_1}, \tag*{(36)}

where h^n=∑i,j(hn)ijc^i†c^j\widehat{h}_n = \sum_{i,j}(h_n)_{ij}\hat{c}_i^\dagger\hat{c}_j and O^α=∑i,j(Oα)ijc^i†c^j\widehat{O}_\alpha= \sum_{i,j}(O_\alpha)_{ij}\hat{c}_i^\dagger\hat{c}_j. We have defined an operation (adj⁡A)B≡[A,B](\operatorname{adj} A)B \equiv[A,B]. The appearance of sα∗s_\alpha^* is due to the use of the relation sαVα=−sα∗∣Vα∣s_\alpha V_\alpha= -s_\alpha^*\lvert V_\alpha\rvert.

By diagonalizing hnh_n as (Pn−1hnPn)ij=λn,iδij(P_n^{-1}h_nP_n)_{ij}=\lambda_{n,i}\delta_{ij}, we can rewrite the factor with OαO_\alpha into an explicit form

(e−Δβadj⁡hn−1−Δβadj⁡hnOα)ij=∑i′,j′(Pn)ii′e−Δβ(λn,i′−λn,j′)−1−Δβ(λn,i′−λn,j′)(Pn−1OαPn)i′j′(Pn−1)j′j.(37)\left(\frac{e^{-\Delta\beta\operatorname{adj}h_n}-1}{-\Delta\beta\operatorname{adj}h_n}O_\alpha\right)_{ij} = \sum_{i',j'}(P_n)_{ii'}\frac{e^{-\Delta\beta(\lambda_{n,i'}-\lambda_{n,j'})}-1}{-\Delta\beta(\lambda_{n,i'}-\lambda_{n,j'})}(P_n^{-1}O_\alpha P_n)_{i'j'}(P_n^{-1})_{j'j}. \tag*{(37)}

Some results for the case without the sign problem

In this appendix, we show some results for the case without the sign problem to confirm that the real Langevin simulation works in that case. We use the same code as the one used for the complex Langevin simulations in this paper since Eq. (5) reduces to Eq. (3) for the real initial configuration z(0)=x(0)z(0)=x(0).

Let us consider the case with Ns=4N_s=4 in a p3/2p_{3/2} orbit and Nv=2N_v=2 as an example. The two-particle model space consists of a J=0J=0 singlet and a J=2J=2 quintuplet. For a rotationally invariant shell-model Hamiltonian in this space, the two-particle energies are given by E(J)=2e1+VJE(J)=2e_1+V_J with J=0,2J=0,2, where e1e_1 is the one-body energy and VJV_J is the strength of the two-body interaction for the two particles coupled with the total angular momentum JJ. Since only the relative energy E(0)−E(2)=V0−V2E(0)-E(2)=V_0-V_2 is relevant, we may set e1=0e_1 = 0 and choose V2=0V_2 = 0 as an energy reference. The additional choice V0=−1V_0 = -1 MeV merely fixes the energy scale and corresponds to a simple attractive interaction in the J=0J = 0 channel. In particular, there is no sign problem in this case.

In Figure 5, we show the energy obtained by real Langevin simulations, where we observe good agreement with the exact values at high temperature. Slight deviation at lower temperature can be attributed to the finite discretization effects of β\beta and the Langevin time. This figure should be compared with Figure 1 (Top) and 2 (Top), which show the results obtained by complex Langevin simulations for Ns=20N_s = 20, Nv=2N_v = 2, 10 and nβ=10n_\beta= 10.

Image not extracted

Figure 5. The expectation value of the energy is plotted against the inverse temperature β\beta in the grand-canonical formalism with Ns=4N_s=4, Nv=2N_v=2, nβ=10n_\beta=10, normalizing the ground state energy to be −1-1 MeV. The purple points with error bars represent the results of real Langevin simulations, whereas the gray curve represents the result obtained by the Hamiltonian diagonalization.

Comparison with the gg-extrapolation

In this appendix, we discuss the gg-extrapolation [10] as an alternative way to perform extrapolations from the region in which correct results are available.

Here we pay attention to the fact that the terms with Vα>0V_\alpha> 0 in the Hamiltonian (15) are the “bad” interactions, which produce imaginary factors in the partition function through sα=is_\alpha= i in (19). This motivates us to multiply a tunable parameter gg only to the interaction terms with Vα>0V_\alpha> 0 in (15) so that the gg-deformed Hamiltonian reads

H^g=f(g)(∑i,j=1NsTijc^i†c^j+12∑α,Vα<0Vα(O^α)2)+g2∑α,Vα>0Vα(O^α)2(38)\hat{H}_g = f(g) \left( \sum_{i,j=1}^{N_s} T_{ij} \hat{c}_i^\dagger\hat{c}_j + \frac{1}{2} \sum_{\alpha,V_\alpha<0} V_\alpha\left(\hat{O}_\alpha\right)^2 \right) + \frac{g}{2} \sum_{\alpha,V_\alpha>0} V_\alpha\left(\hat{O}_\alpha\right)^2 \tag*{(38)}

which is in contrast to the tt-deformed Hamiltonian (31), where all the interaction terms are commonly multiplied by tt. In Eq. (C.1), we have introduced an arbitrary function f(g)f(g) satisfying f(g)>0f(g) > 0 for g<0g < 0 and f(1)=1f(1) = 1, which is set to f(g)=1f(g) = 1 in what follows for simplicity. Note that the gg-deformed Hamiltonian (C.1) does not suffer from the sign problem for g≤0g \le0. To obtain the expectation value of the energy, we calculate the expectation value of the original Hamiltonian Hg=1H_{g=1} for various gg.

In Figure 6, we plot the expectation value of the energy obtained by the Hamiltonian diagonalization against gg for the same setting as in Figure 4. We fit the results for g≤0g \le0 to a quadratic function, where the error bands represent the fitting errors. We find that the fit does not reproduce the correct value at g=1g = 1. This has something to do with the existence of an inflection point near g=0g = 0, which does not exist in the tt-extrapolation in Figure 4. While this may not be a generic feature of the gg-extrapolation, we consider that the tt-extrapolation has a certain advantage at least in the case at hand.

Image not extracted

Figure 6. The expectation value of the energy obtained by the Hamiltonian diagonalization with Ns=20N_s = 20, Nv=2N_v = 2, β=1\beta= 1 is plotted against the tunable parameter gg. The data within the g≤0g \le0 region are fitted to a quadratic function.

References

  1. [44]L.L. Salcedo, Does the complex Langevin method give unbiased results?, Phys. Rev. D 94 (2016) 114505 [1611.06390].DOI
  2. [45]M. Honma, T. Otsuka, B.A. Brown and T. Mizusaki, Shell-model description of neutron-rich pf-shell nuclei with a new effective interaction GXPF1, Eur. Phys. J. A 25 (2005) 499.
  3. [46]I. Stetcu, A. Baroni and J. Carlson, Variational approaches to constructing the many-body nuclear ground state for quantum computing, Phys. Rev. C 105 (2022) 064308 [2110.06098].DOI
  4. [47]A. Pérez-Obiol, A.M. Romero, J. Menéndez, A. Rios, A. García-Sáez and B. Juliá-Díaz, Nuclear shell-model simulation in digital quantum computers, Sci. Rep. 13 (2023) 12291 [2302.03641].

Paper details

vlnc:W_8559cahu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents