betaFeedback

WorksPhysics and Astronomy

Computing the QRPA Level Density with the Finite Amplitude Method

Abstract

We describe a new algorithm to calculate the vibrational nuclear level density of an atomic nucleus. Fictitious perturbation operators that probe the response of the system are generated by drawing their matrix elements from some probability distribution function. We use the Finite Amplitude Method to explicitly compute the response for each such sample. With the help of the Kernel Polynomial Method, we build an estimator of the vibrational level density and provide the upper bound of the relative error in the limit of infinitely many random samples. The new algorithm can give accurate estimates of the vibrational level density. Since it is based on drawing multiple samples of perturbation operators, its computational implementation is naturally parallel and scales like the number of available processing units.

Introduction

The first attempts to calculate the nuclear level density were based on the statistical model [1, 2, 3, 4]. The formulas for the nuclear level density and spin distribution of excited states that are implemented in modern nuclear reaction codes such as TALYS [5] or EMPIRE [6] are largely based on these early formulations. While such an approach gives the correct asymptotic, it relies on adjustable parameters and usually fails to capture the low-lying, discrete part of the spectrum. Direct methods based on the nuclear shell model are much more predictive but can only be applied in light-to-medium mass nuclei or with schematic interactions [7, 8, 9, 10, 11, 12, 13].

What is often referred to as the combinatorial method is a good compromise between predictive power and feasibility [14]. It is based on the general framework of the nuclear energy density functional theory, where several many-body methods can be leveraged to compute nuclear properties at a reasonable computational cost [15]. The combinatorial method relies on counting nuclear excitations using generating function techniques [16]. In the earlier version of the method, nuclear excited states were computed as nn-particle nn-holes excitations [14]. The theory was extended to include vibrational and rotational states [17, 18] and the effect of nuclear excitation energy on the single-particle structure [19]. Most recently, particle-hole excitations were replaced by the exact eigenvalues of the Quasiparticle Random Phase (QRPA) matrix, giving a much more realistic description of low-lying states [20].

While the latest version of the combinatorial method contains a lot of new physics, it has become computationally expensive in deformed nuclei, where the QRPA matrix may easily reach sizes of 106×10610^{6} \times10^{6} for realistic harmonic oscillator basis. Computing all these matrix elements is only doable on a case-by-case basis or with additional approximations such as cut-offs in quasiparticle space or smaller bases [21, 22]. The Finite Amplitude Method (FAM) offers a computationally efficient alternative to computing the QRPA linear response for a given excitation operator without explicitly constructing the QRPA matrix, but does not have direct access to the QRPA eigenfrequencies, i.e. the vibrational spectrum [23, 24].

In this paper, we show that we can still use the Finite Amplitude Method to efficiently probe the vibrational spectrum of nuclei of arbitrary size and compute estimates of the density of phonons with controlled precision.

In Section 2, we summarize very briefly the QRPA and linear response theory to introduce the vibrational level density. Our new solution method is outlined in Section 3. Practical aspects related to its implementation, which requires introducing an additional algorithm known as the Kernel Polynomial Method, are discussed in Section 4. In Section 5, we illustrate the performance of the method with several benchmarks, both with synthetic or realistic data.

QRPA level density

In this section, we recall some of the basic properties of the QRPA matrix and of the linear response theory formalism which is used to compute the response of the nucleus to a given excitation operator.

Notations

Throughout the paper we use A†A^{\dagger} for the Hermitian conjugate matrix, ATA^{T} for the transposed matrix and A∗A^{*} for the element-wise complex conjugate matrix of an arbitrary matrix A∈Cm×nA \in\mathbb{C}^{m \times n}. We also use MATLAB notation for diagonal matrices, i.e. if x∈Cn\mathbf{x} \in\mathbb{C}^{n} is a vector, then diag⁡[x]∈Cn×n\operatorname{diag}[\mathbf{x}] \in\mathbb{C}^{n \times n} is a diagonal matrix having elements of the vector x\mathbf{x} as diagonal elements, i.e. diag⁡[x]ij=\operatorname{diag}[\mathbf{x}]_{ij} =1

Preprint submitted to Computer Physics Communications

LLNL-JRNL-866460 δijxj\delta_{ij}x_j, for i,j=1,…,ni,j=1,\ldots,n. For a matrix A∈Cm×nA\in\mathbb{C}^{m\times n}, the Frobenius norm is denoted by: ∥A∥F=∑i=1m∑j=1n∣Aij∣2\lVert A\rVert_F=\sqrt{\sum_{i=1}^{m}\sum_{j=1}^{n}\lvert A_{ij}\rvert^2}, while the Euclidean norm of a vector v∈Cnv\in\mathbb{C}^n is ∥v∥=∑i=1n∣vi∣2\lVert v\rVert=\sqrt{\sum_{i=1}^{n}\lvert v_i\rvert^2}. We denote the imaginary unit as ii. For an nn-dimensional random vector X\mathbf{X} that follows the multivariate Gaussian distribution with expected value μ∈Rn\boldsymbol{\mu}\in\mathbb{R}^n and covariance matrix Σ∈Rn×n\boldsymbol{\Sigma}\in\mathbb{R}^{n\times n}, we use the notation: X∼Nn(μ,Σ)\mathbf{X}\sim\mathcal{N}_n(\boldsymbol{\mu},\boldsymbol{\Sigma}). For a random variable XX, E[X]\mathbb{E}[X] is its expected mean value and Var⁡(X)\operatorname{Var}(X) its variance.

QRPA matrix

The goal of this paper is to compute the QRPA level density function ρ(ω)\rho(\omega), for all ω≥0\omega\geq0, defined as

ρ(ω)=∑i=1Npδ(ω−Ωi),(1)\rho(\omega)=\sum_{i=1}^{N_p}\delta(\omega-\Omega_i), \tag*{(1)}

where Np∈NN_p\in\mathbb{N} is the number of quasiparticle pairs μ<ν\mu<\nu, and the Ωi\Omega_i are the QRPA eigenfrequencies defined as the positive eigenvalues of the QRPA matrix SS

S=[+I00−I][ABB∗A∗],(2)S=\begin{bmatrix}+I&0\\0&-I\end{bmatrix}\begin{bmatrix}A&B\\B^*&A^*\end{bmatrix}, \tag*{(2)}

where the matrices A,B∈CNp×NpA,B\in\mathbb{C}^{N_p\times N_p} satisfy A†=AA^\dagger=A and BT=BB^T=B and involve the second derivatives of the energy functional with respect to the normal and pairing densities [15, 25]. It is customary to introduce the QRPA eigenmodes X,Y∈CNp×NpX,Y\in\mathbb{C}^{N_p\times N_p} to define the eigenvectors of the QRPA matrix. If Ω=diag⁡[Ωi]i=1Np∈RNp×Np\Omega=\operatorname{diag}[\Omega_i]_{i=1}^{N_p}\in\mathbb{R}^{N_p\times N_p} is the diagonal matrix made of its eigenvalues and is such that Ωi>0\Omega_i>0 for each i=1,…,Npi=1,\ldots,N_p, then one can show that

\begin{bmatrix}+I&0\\0&-I\end{bmatrix}\begin{bmatrix}A&B\\B^*&A^*\end{bmatrix} = \begin{bmatrix}X&Y^*\\Y&X^*\end{bmatrix} \begin{bmatrix}+\Omega&0\\0&-\Omega\end{bmatrix} \begin{bmatrix}X&Y^*\\Y&X^*\end{bmatrix}^{-1}, \tag*{(3)}\end{bmatrix}

with

\begin{bmatrix}X&Y^*\\Y&X^*\end{bmatrix}^{-1} = \begin{bmatrix}+I&0\\0&-I\end{bmatrix} \begin{bmatrix}X&Y^*\\Y&X^*\end{bmatrix}^{\dagger} \begin{bmatrix}+I&0\\0&-I\end{bmatrix}. \tag*{(4)}\end{bmatrix}

In the context of numerical linear algebra, the function ρ(ω)\rho(\omega) is known as the spectral density of the QRPA matrix SS. The practical calculation of this function in the case of the QRPA matrix (2) faces two problems

  • In spherical nuclei, the QRPA matrix SS can be easily constructed and diagonalized as in Eqs. (3) and (4). In that case, one has an explicit access to the QRPA eigenfrequencies {Ωi}i=1Np\{\Omega_i\}_{i=1}^{N_p}, and therefore the computation of the level density function ρ(ω)\rho(\omega) becomes trivial. However, in heavy deformed nuclei the dimension NpN_p becomes prohibitively large as it can reach Np∼106N_p\sim10^6 for large harmonic oscillator bases. In those cases, the direct determination and diagonalization of the QRPA matrix has only been achieved in a few cases [21, 22].

  • Although numerous numerical methods for estimating the spectral density exist [26, 27], they all assume that the matrix for which the spectral density is computed is Hermitian. In our case, the QRPA matrix SS is not Hermitian and the matrix of its eigenvectors is not unitary but involves the orthonormality condition of Eq. (4). Therefore, those spectral density methods cannot be directly used.

Linear response theory

While the QRPA approximation is the small amplitude limit of the time-dependent Hartree-Fock-Bogoliubov theory, linear response theory is the same limit in the presence of an external, time-dependent field F^(t)\hat{F}(t) [25]. In the quasiparticle basis represented by the quasiparticle creation and annihilation operators β^μ†\hat{\beta}^{\dagger}_{\mu} and β^ν\hat{\beta}_{\nu}, this field is represented as [15]

F^=12∑μν[Fμν11β^μ†β^ν+Fμν20β^μ†β^ν†+Fμν02β^μβ^ν−Fμν11∗β^ν†β^μ].(5)\hat{F}=\frac{1}{2}\sum_{\mu\nu}\left[F^{11}_{\mu\nu}\hat{\beta}^{\dagger}_{\mu}\hat{\beta}_{\nu}+F^{20}_{\mu\nu}\hat{\beta}^{\dagger}_{\mu}\hat{\beta}^{\dagger}_{\nu}+F^{02}_{\mu\nu}\hat{\beta}_{\mu}\hat{\beta}_{\nu}-F^{11*}_{\mu\nu}\hat{\beta}^{\dagger}_{\nu}\hat{\beta}_{\mu}\right]. \tag*{(5)}

Let F20,F02∈CNpF^{20},F^{02}\in\mathbb{C}^{N_p} be the vectorized strict upper triangle parts (μ<ν\mu<\nu) of the matrices Fμν20,Fμν02F^{20}_{\mu\nu},F^{02}_{\mu\nu}. For ω∈R\omega\in\mathbb{R} and smearing γ>0\gamma>0, we denote the complex frequency ωγ=ω+γi\omega_\gamma=\omega+\gamma i in the upper complex plane.

In the small amplitude limit, the external perturbation induces oscillations of the Hamiltonian δHμν20(ωγ)\delta H^{20}_{\mu\nu}(\omega_\gamma) and δHμν02(ωγ)\delta H^{02}_{\mu\nu}(\omega_\gamma) according to the equations [24],

(Eμ+Eν−ωγ)Xμν(ωγ)+δHμν20(ωγ)=Fμν20(ωγ),(E_{\mu}+E_{\nu}-\omega_\gamma)X_{\mu\nu}(\omega_\gamma)+\delta H^{20}_{\mu\nu}(\omega_\gamma)=F^{20}_{\mu\nu}(\omega_\gamma),
(Eμ+Eν+ωγ)Yμν(ωγ)+δHμν02(ωγ)=Fμν02(ωγ),(6)(E_{\mu}+E_{\nu}+\omega_\gamma)Y_{\mu\nu}(\omega_\gamma)+\delta H^{02}_{\mu\nu}(\omega_\gamma)=F^{02}_{\mu\nu}(\omega_\gamma), \tag*{(6)}

where the antisymmetric matrices Xμν(ω)X_{\mu\nu}(\omega) and Yμν(ω)Y_{\mu\nu}(\omega) are the corresponding forward and backward amplitudes. Let the vectors X(ωγ),Y(ωγ)∈CNpX(\omega_\gamma),Y(\omega_\gamma)\in\mathbb{C}^{N_p} denote the vectorized strict upper triangle parts of the matrices Xμν(ωγ),Yμν(ωγ)X_{\mu\nu}(\omega_\gamma),Y_{\mu\nu}(\omega_\gamma). By expanding δHμν20(ωγ)\delta H^{20}_{\mu\nu}(\omega_\gamma) and δHμν02(ωγ)\delta H^{02}_{\mu\nu}(\omega_\gamma) in terms of the amplitudes Xμν(ωγ)X_{\mu\nu}(\omega_\gamma) and Yμν(ωγ)Y_{\mu\nu}(\omega_\gamma) we obtain the linear response equation,

([ABB∗A∗]−ωγ[+I00−I])[X(ωγ)Y(ωγ)]=−[F20F02].(7)\left(\begin{bmatrix}A&B\\B^*&A^*\end{bmatrix} -\omega_\gamma\begin{bmatrix}+I&0\\0&-I\end{bmatrix}\right) \begin{bmatrix}X(\omega_\gamma)\\Y(\omega_\gamma)\end{bmatrix} = -\begin{bmatrix}F^{20}\\F^{02}\end{bmatrix}. \tag*{(7)}

Traditionally, Eq. (7) is solved by explicitly constructing and diagonalizing the matrix SS of Eq. (2), which can become computationally prohibitive for heavy deformed nuclei. However, it is in fact possible to solve for the X(ωγ)X(\omega_\gamma) and Y(ωγ)Y(\omega_\gamma) without having to compute the matrix elements of SS using the Finite Amplitude Method (FAM) [23, 24]. The key idea of the FAM is to iteratively solve Eq. (6), independently, for each frequency ω\omega on a mesh. As was shown in Ref. [28], the FAM actually only accesses the mapping

\begin{bmatrix}x\\y\end{bmatrix}\in\mathbb{C}^{2N_p}\mapsto \begin{bmatrix}A&B\\B^*&A^*\end{bmatrix} \begin{bmatrix}x\\y\end{bmatrix}\in\mathbb{C}^{2N_p}, \tag*{(8)}\end{bmatrix}

and therefore the FAM could be classified as a matrix-free method for solving the linear system (7). In the rest of this paper, we will make extensive use of this property and design a method to estimate the spectral density ρ(ω)\rho(\omega) by only using the mapping (8).

For any given operator F^\hat{F}, we define the strength function S(ωγ,F^)S(\omega_\gamma,\hat{F}) as

S(ωγ,F^)=[F20F02]†[X(ωγ)Y(ωγ)],(9)S(\omega_\gamma,\hat{F}) = \begin{bmatrix} F^{20} \\ F^{02} \end{bmatrix}^{\dagger} \begin{bmatrix} X(\omega_\gamma) \\ Y(\omega_\gamma) \end{bmatrix}, \tag*{(9)}

and the response function dB(ω,F^)/dωdB(\omega,\hat{F})/d\omega as

dB(ω,F^)dω=lim⁡γ→0+−1πIm⁡[S(ωγ,F^)].(10)\frac{dB(\omega,\hat{F})}{d\omega} = \lim_{\gamma\to0^{+}} -\frac{1}{\pi}\operatorname{Im}\left[S(\omega_\gamma,\hat{F})\right]. \tag*{(10)}

One can show that the response function can be written as

dB(ω,F^)dω=∑i=1Np∣⟨i∣F^∣0⟩∣2δ(ω−Ωi)−∑i=1Np∣⟨0∣F^∣i⟩∣2δ(ω+Ωi),(11)\frac{dB(\omega,\hat{F})}{d\omega} = \sum_{i=1}^{N_p} \left|\langle i|\hat{F}|0\rangle\right|^2 \delta(\omega-\Omega_i) - \sum_{i=1}^{N_p} \left|\langle0|\hat{F}|i\rangle\right|^2 \delta(\omega+\Omega_i), \tag*{(11)}

where ∣0⟩|0\rangle is the QRPA ground state, and ∣i⟩=O^i†∣0⟩|i\rangle= \hat{O}_i^\dagger|0\rangle is the ii-th excited state defined from the phonon creation operators O^i†\hat{O}_i^\dagger

O^i†=∑μ<ν[Xμνiβ^μ†β^ν†−Yμνiβ^νβ^μ].(12)\hat{O}_i^\dagger= \sum_{\mu<\nu}\left[X_{\mu\nu}^{i}\hat{\beta}_{\mu}^{\dagger}\hat{\beta}_{\nu}^{\dagger} - Y_{\mu\nu}^{i}\hat{\beta}_{\nu}\hat{\beta}_{\mu}\right]. \tag*{(12)}

Using Eqs. (5) and (12) we can compute the transition strengths [29]

Ti(F^)=⟨i∣F^∣0⟩=⟨Φ0∣[O^i,F^]∣Φ0⟩=∑μ<ν[(Xμνi)∗Fμν20+(Yμνi)∗Fμν02],T_i(\hat{F}) = \langle i|\hat{F}|0\rangle= \langle\Phi_0|[\hat{O}_i,\hat{F}]|\Phi_0\rangle= \sum_{\mu<\nu}\left[(X_{\mu\nu}^{i})^*F_{\mu\nu}^{20}+(Y_{\mu\nu}^{i})^*F_{\mu\nu}^{02}\right],
T~i(F^)=⟨0∣F^∣i⟩=⟨Φ0∣[O^i†,F^]∣Φ0⟩=∑μ<ν[YμνiFμν20+XμνiFμν02],(13)\widetilde{T}_i(\hat{F}) = \langle0|\hat{F}|i\rangle= \langle\Phi_0|[\hat{O}_i^\dagger,\hat{F}]|\Phi_0\rangle= \sum_{\mu<\nu}\left[Y_{\mu\nu}^{i}F_{\mu\nu}^{20}+X_{\mu\nu}^{i}F_{\mu\nu}^{02}\right], \tag*{(13)}

where ∣Φ0⟩|\Phi_0\rangle is the HFB state. A shorthand, vectorized form of these transition strengths reads

\begin{bmatrix} T(\hat{F}) \\ \widetilde{T}(\hat{F}) \end{bmatrix} = \begin{bmatrix} X & Y^* \\ Y & X^* \end{bmatrix}^{\dagger} \begin{bmatrix} F^{20} \\ F^{02} \end{bmatrix}. \tag*{(14)}\end{bmatrix}

Using Eq. (4), we can also invert the previous equation:

\begin{bmatrix} F^{20} \\ F^{02} \end{bmatrix} = \begin{bmatrix} +I & 0 \\ 0 & -I \end{bmatrix} \begin{bmatrix} X & Y^* \\ Y & X^* \end{bmatrix} \begin{bmatrix} +I & 0 \\ 0 & -I \end{bmatrix} \begin{bmatrix} T(\hat{F}) \\ \widetilde{T}(\hat{F}) \end{bmatrix}. \tag*{(15)}\end{bmatrix}

Notice that if we choose the excitation operator F^′\hat{F}' (i.e. picked Fμ,ν20,Fμ,ν02∈CF_{\mu,\nu}^{20}, F_{\mu,\nu}^{02} \in\mathbb{C}) such that

\begin{bmatrix} +F^{20} \\ -F^{02} \end{bmatrix} = \begin{bmatrix} X & Y^* \\ Y & X^* \end{bmatrix} \begin{bmatrix} 1 \\ 1 \end{bmatrix}, \tag*{(16)}\end{bmatrix}

where 1=[1,…,1]T∈RNp1=[1,\ldots,1]^T\in\mathbb{R}^{N_p}, then we would have:

∣Ti(F^′)∣2=∣⟨i∣F^′∣0⟩∣2=1and∣T~i(F^′)∣2=∣⟨0∣F^′∣i⟩∣2=1,(17)|T_i(\hat{F}')|^2=|\langle i|\hat{F}'|0\rangle|^2=1 \quad\text{and}\quad |\widetilde{T}_i(\hat{F}')|^2=|\langle0|\hat{F}'|i\rangle|^2=1, \tag*{(17)}

for all i=1,…,Npi=1,\ldots,N_p, and consequently the response function for that specific operator F^′\hat{F}' would be equal to the QRPA level density function

dB(ω,F^′)dω=ρ(ω),∀ ω≥0.(18)\frac{dB(\omega,\hat{F}')}{d\omega}=\rho(\omega),\qquad\forall\ \omega\geq0. \tag*{(18)}

In practice, such operator F^′\hat{F}' is impossible to construct because it essentially involves computing the sum of all eigenvectors in Eq. (16), and therefore requires the full diagonalization of the QRPA matrix. However, this simple example demonstrates that by carefully picking the excitation operator and afterwards solving only the linear response equation (7), one can, at least in principle, obtain an approximation of the QRPA level density function. Therefore, the main question is how to efficiently construct an operator F^′\hat{F}' which excites all the eigenmodes with equal strength.

Method description

We consider a collection of NsN_s random excitation operators F^is\hat{F}_{i_s}, for 1≤is≤Ns1\leq i_s\leq N_s, with matrix elements Fis20,Fis02F_{i_s}^{20}, F_{i_s}^{02} drawn from a multivariate Gaussian distribution. We will show how to estimate the QRPA level density function ρ(ω)\rho(\omega) by introducing a specific averaging procedure over the collection of the obtained response functions dB(ω,F^is)/dωdB(\omega,\hat{F}_{i_s})/d\omega computed for each sample operator. Before we proceed, we recall a well-known Lemma describing how a Gaussian random vector is affected by a linear transformation.

Lemma 3.1. Let X∼Nn(0,Σ)X\sim\mathcal{N}_n(0,\Sigma) be a nn-dimensional random vector having multivariate Gaussian distribution with zero expected value and covariance matrix Σ\Sigma. Then for any matrix A∈Rm×nA\in\mathbb{R}^{m\times n}, the mm-dimensional random vector AXAX has a multivariate Gaussian distribution with zero expected value and covariance matrix AΣATA\Sigma A^T, i.e. AX∼Nm(0,AΣAT)AX\sim\mathcal{N}_m(0,A\Sigma A^T).

Randomized excitation operator

According to Theorem 1 in Appendix B from Ref. [28], there exist C,D∈CNp×NpC,D\in\mathbb{C}^{N_p\times N_p} unitary matrices and diagonal matrix θ=diag⁡[θi]i=1Np\theta=\operatorname{diag}[\theta_i]_{i=1}^{N_p}, for θ1,…,θNp≥0\theta_1,\ldots,\theta_{N_p}\geq0, such that

X=Dcosh⁡θCandY=D∗sinh⁡θC.(19)X=D\cosh\theta C \quad\text{and}\quad Y=D^*\sinh\theta C. \tag*{(19)}

Inserting the previous relations into Eq. (14), we obtain:

\begin{bmatrix} T(\hat{F}) \\ \widetilde{T}(\hat{F}) \end{bmatrix} = \begin{bmatrix} X & Y^* \\ Y & X^* \end{bmatrix}^{\dagger} \begin{bmatrix} F^{20} \\ F^{02} \end{bmatrix} = \begin{bmatrix} C^* & 0 \\ 0 & C^T \end{bmatrix} \begin{bmatrix} \cosh\theta& \sinh\theta\\ \sinh\theta& \cosh\theta \end{bmatrix} \begin{bmatrix} D^\dagger& 0 \\ 0 & D^T \end{bmatrix} \begin{bmatrix} F^{20} \\ F^{02} \end{bmatrix}. \tag*{(20)}\end{bmatrix}

Suppose that F20,F02∈CNpF^{20},F^{02}\in\mathbb{C}^{N_p} are random complex vectors, such that all the components

Re⁡[Fi20], Re⁡[Fi02], Im⁡[Fi20], Im⁡[Fi02]∼N1(0,σ2),(21)\operatorname{Re}[F_i^{20}],\ \operatorname{Re}[F_i^{02}],\ \operatorname{Im}[F_i^{20}],\ \operatorname{Im}[F_i^{02}] \sim\mathcal{N}_1(0,\sigma^2), \tag*{(21)}

are independent and identically distributed Gaussian random variables with zero expected value and the same variance σ2>0\sigma^2>0, for each i=1,…,Npi=1,\ldots,N_p. Then it is easy to show that the 4Np4N_p-dimensional random vector

\begin{bmatrix} \operatorname{Re}F^{20} \\ \operatorname{Re}F^{02} \\ \operatorname{Im}F^{20} \\ \operatorname{Im}F^{02} \end{bmatrix} \sim\mathcal{N}_{4N_p}(0,\sigma^2 I), \tag*{(22)}\end{bmatrix}

has multivariate Gaussian distribution with zero expected value and covariance matrix σ2I\sigma^2 I.

Let us introduce the real and imaginary parts: C1C_1, C2C_2, D1D_1, D2∈RNp×NpD_2 \in\mathbb{R}^{N_p \times N_p} of CC and DD matrices, i.e. C=C1+C2iC = C_1 + C_2i and D=D1+D2iD = D_1 + D_2i. It is easy to see that Eq. (20) can be recast into

\begin{bmatrix} \operatorname{Re} T(\hat{F}) \\ \operatorname{Re} \widetilde{T}(\hat{F}) \\ \operatorname{Im} T(\hat{F}) \\ \operatorname{Im} \widetilde{T}(\hat{F}) \end{bmatrix} = \mathcal{X} \begin{bmatrix} \operatorname{Re} F^{20} \\ \operatorname{Re} F^{02} \\ \operatorname{Im} F^{20} \\ \operatorname{Im} F^{02} \end{bmatrix}, \tag*{(23)}\end{bmatrix}

where the real matrix X∈R4Np×4Np\mathcal{X} \in\mathbb{R}^{4N_p \times4N_p} is given by

X=C[cosh⁡θsinh⁡θ00sinh⁡θcosh⁡θ0000cosh⁡θsinh⁡θ00sinh⁡θcosh⁡θ]D,(24)\mathcal{X} = \mathcal{C} \begin{bmatrix} \cosh\theta& \sinh\theta& 0 & 0 \\ \sinh\theta& \cosh\theta& 0 & 0 \\ 0 & 0 & \cosh\theta& \sinh\theta\\ 0 & 0 & \sinh\theta& \cosh\theta \end{bmatrix} \mathcal{D}, \tag*{(24)}

with

C=[C1T0C2T00C1T0−C2T−C2T0C1T00C2T0C1T],(25)\mathcal{C} = \begin{bmatrix} C_1^T & 0 & C_2^T & 0 \\ 0 & C_1^T & 0 & -C_2^T \\ -C_2^T & 0 & C_1^T & 0 \\ 0 & C_2^T & 0 & C_1^T \end{bmatrix}, \tag*{(25)}
D=[D1T0D2T00D1T0−D2T−D2T0D1T00D2T0D1T].(26)\mathcal{D} = \begin{bmatrix} D_1^T & 0 & D_2^T & 0 \\ 0 & D_1^T & 0 & -D_2^T \\ -D_2^T & 0 & D_1^T & 0 \\ 0 & D_2^T & 0 & D_1^T \end{bmatrix}. \tag*{(26)}

Both matrices C\mathcal{C} and D\mathcal{D} are orthogonal since CC and DD are unitary.

According to Lemma 3.1, we know that the following 4Np4N_p-dimensional random vector,

\begin{bmatrix} \operatorname{Re} T(\hat{F}) \\ \operatorname{Re} \widetilde{T}(\hat{F}) \\ \operatorname{Im} T(\hat{F}) \\ \operatorname{Im} \widetilde{T}(\hat{F}) \end{bmatrix} \sim\mathcal{N}_{4N_p}(0,\Sigma), \tag*{(27)}\end{bmatrix}

has multivariate Gaussian distribution with zero expected value and (real) covariance matrix Σ=X(σ2I)XT\Sigma= \mathcal{X}(\sigma^2 I)\mathcal{X}^T. We can easily extract an arbitrary component of the random vector (27). As an example, for any j=1,…,Npj = 1,\ldots,N_p, we have

Re⁡Tj(F^)=Re⁡[⟨j∣F^∣0⟩]=[ejT000][Re⁡T(F^)Re⁡T~(F^)Im⁡T(F^)Im⁡T~(F^)].(28)\operatorname{Re} T_j(\hat{F}) = \operatorname{Re}[\langle j|\hat{F}|0\rangle] = \begin{bmatrix} e_j^T & 0 & 0 & 0 \end{bmatrix} \begin{bmatrix} \operatorname{Re} T(\hat{F}) \\ \operatorname{Re} \widetilde{T}(\hat{F}) \\ \operatorname{Im} T(\hat{F}) \\ \operatorname{Im} \widetilde{T}(\hat{F}) \end{bmatrix}. \tag*{(28)}

Invoking Lemma 3.1 again, we see that the random variable Re⁡⟨j∣F^∣0⟩∼N(0,σj2)\operatorname{Re}\langle j|\hat{F}|0\rangle\sim\mathcal{N}(0,\sigma_j^2) has Gaussian distribution with zero expected value and variance σj2\sigma_j^2 given by

σj2=[ejT000]Σ[ej000].(29)\sigma_j^2 = \begin{bmatrix} e_j^T & 0 & 0 & 0 \end{bmatrix} \Sigma \begin{bmatrix} e_j \\ 0 \\ 0 \\ 0 \end{bmatrix}. \tag*{(29)}

Using the result Σ=X(σ2I)XT\Sigma= \mathcal{X}(\sigma^2 I)\mathcal{X}^T, this can be easily be computed explicitly as

σj2=σ2+2σ2∥Yej∥2.(30)\sigma_j^2 = \sigma^2 + 2\sigma^2\lVert Y e_j\rVert^2. \tag*{(30)}

Repeating a similar procedure for the other components of the random vector (27), we find that for any i=1,…,Npi = 1,\ldots,N_p,

Re⁡⟨i∣F^∣0⟩, Re⁡⟨0∣F^∣i⟩, Im⁡⟨i∣F^∣0⟩, Im⁡⟨0∣F^∣i⟩∼N1(0,σi2)(31)\operatorname{Re}\langle i|\hat{F}|0\rangle,\ \operatorname{Re}\langle0|\hat{F}|i\rangle,\ \operatorname{Im}\langle i|\hat{F}|0\rangle,\ \operatorname{Im}\langle0|\hat{F}|i\rangle\sim\mathcal{N}_1(0,\sigma_i^2) \tag*{(31)}

with the variance σi2\sigma_i^2 given by Eq. (30). It is worth mentioning that all these components are not necessarily independent random variables.

We can now for any i=1,…,Npi = 1,\ldots,N_p compute the following expected value

E[∣⟨i∣F^∣0⟩∣2]=E[Re⁡⟨i∣F^∣0⟩2]+E[Im⁡⟨i∣F^∣0⟩2].(32)\mathbb{E}[|\langle i|\hat{F}|0\rangle|^2] = \mathbb{E}[\operatorname{Re}\langle i|\hat{F}|0\rangle^2] + \mathbb{E}[\operatorname{Im}\langle i|\hat{F}|0\rangle^2]. \tag*{(32)}

Since for any random variable XX with zero expected value its variance can be expressed as Var⁡(X)=E[X2]\operatorname{Var}(X) = \mathbb{E}[X^2], result (31) yields:

E[∣⟨i∣F^∣0⟩∣2]=σi2+σi2=2σ2+4σ2∥Yei∥2.\mathbb{E}[|\langle i|\hat{F}|0\rangle|^2] = \sigma_i^2 + \sigma_i^2 = 2\sigma^2 + 4\sigma^2\lVert Y e_i\rVert^2.

The same procedure can be repeated for ∣⟨0∣F^∣i⟩∣2|\langle0|\hat{F}|i\rangle|^2. Overall, we conclude that if one draws the elements of F20F^{20} and F02F^{02} from a Gaussian distribution as in Eq. (21), then for any i=1,…,Npi = 1,\ldots,N_p there holds:

E[∣⟨i∣F^∣0⟩∣2]=E[∣⟨0∣F^∣i⟩∣2]=4σ2(12+∥Yei∥2).(33)\mathbb{E}[|\langle i|\hat{F}|0\rangle|^2] = \mathbb{E}[|\langle0|\hat{F}|i\rangle|^2] = 4\sigma^2\left(\frac{1}{2} + \lVert Y e_i\rVert^2\right). \tag*{(33)}

Level density estimator

Recall that the response function expression (11) for non-negative frequencies ω≥0\omega\ge0 reads

dB(ω≥0,F^)dω=∑i=1Np∣⟨i∣F^∣0⟩∣2δ(ω−Ωi).(34)\frac{dB(\omega\ge0,\hat{F})}{d\omega} = \sum_{i=1}^{N_p} |\langle i|\hat{F}|0\rangle|^2 \delta(\omega-\Omega_i). \tag*{(34)}

Notice that now the response function is a random variable because the excitation operator F^\hat{F} is a random variable. Also notice that the following integral over non-negative frequencies

∫0+∞dB(ω′≥0,F^)dω′dω′=∑i=1Np∣⟨i∣F^∣0⟩∣2(35)\int_0^{+\infty} \frac{dB(\omega' \ge0,\hat{F})}{d\omega'}d\omega' = \sum_{i=1}^{N_p}|\langle i|\hat{F}|0\rangle|^2 \tag*{(35)}

is also a random variable.

Since both random variables in Eqs. (34) and (35) are well defined, we can compute their expected values. Therefore, we can define the QRPA level density estimator ρ^(ω)\hat{\rho}(\omega), for non-negative frequencies ω≥0\omega\ge0, as

ρ^(ω)=NpE[dB(ω≥0,F^)dω]E[∫0+∞dB(ω′≥0,F^)dω′dω′].(36)\hat{\rho}(\omega) = N_p \frac{\mathbb{E}\left[\dfrac{dB(\omega\ge0,\hat{F})}{d\omega}\right]} {\mathbb{E}\left[\int_0^{+\infty}\dfrac{dB(\omega' \ge0,\hat{F})}{d\omega'}d\omega'\right]}. \tag*{(36)}

Inserting Eq. (33) into Eq. (34) and Eq. (35), we get

E[dB(ω≥0,F^)dω]=4σ2∑i=1Np(12+∥Yei∥2)δ(ω−Ωi).(37)\mathbb{E}\left[\frac{dB(\omega\ge0,\hat{F})}{d\omega}\right] = 4\sigma^2\sum_{i=1}^{N_p}\left(\frac{1}{2}+\lVert Y e_i\rVert^2\right)\delta(\omega-\Omega_i). \tag*{(37)}

and

E[∫0+∞dB(ω′≥0,F^)dω′dω′]=4σ2Np(12+∥Y∥F2Np).(38)\mathbb{E}\left[\int_0^{+\infty}\frac{dB(\omega' \ge0,\hat{F})}{d\omega'}d\omega'\right] = 4\sigma^2N_p\left(\frac{1}{2}+\frac{\lVert Y\rVert_F^2}{N_p}\right). \tag*{(38)}

Now we can express the estimator ρ^(ω)\widehat{\rho}(\omega) as

ρ^(ω)=∑i=1Np(12+∥Yei∥212+∥Y∥F2Np)δ(ω−Ωi).(39)\widehat{\rho}(\omega)=\sum_{i=1}^{N_p}\left(\frac{\frac{1}{2}+\lVert Y e_i\rVert^2}{\frac{1}{2}+\frac{\lVert Y\rVert_F^2}{N_p}}\right)\delta(\omega-\Omega_i). \tag*{(39)}

In the interpretation where the Dirac delta function is a non-negative function of infinitesimal width and unit area, we can easily show that the absolute error ∣ρ^(ω)−ρ(ω)∣\lvert\widehat{\rho}(\omega)-\rho(\omega)\rvert with respect to the true QRPA level density (1) is bounded as

∣ρ^(ω)−ρ(ω)∣≤ρ(ω)max⁡i=1,…,Npϵi,(40)\lvert\widehat{\rho}(\omega)-\rho(\omega)\rvert\leq\rho(\omega)\max_{i=1,\ldots,N_p}\epsilon_i, \tag*{(40)}

with

ϵi=∣∥Yei∥2−∥Y∥F2Np12+∥Y∥F2Np∣.(41)\epsilon_i=\left\lvert\frac{\lVert Y e_i\rVert^2-\frac{\lVert Y\rVert_F^2}{N_p}}{\frac{1}{2}+\frac{\lVert Y\rVert_F^2}{N_p}}\right\rvert. \tag*{(41)}

The relative error bound for any ω≥0\omega\geq0 is then

∣ρ^(ω)−ρ(ω)ρ(ω)∣≤max⁡i=1,…,Npϵi.(42)\left\lvert\frac{\widehat{\rho}(\omega)-\rho(\omega)}{\rho(\omega)}\right\rvert\leq\max_{i=1,\ldots,N_p}\epsilon_i. \tag*{(42)}

Assuming for simplicity that the eigenfrequencies Ωi\Omega_i are non-degenerate, we see that as one approaches the given QRPA eigenfrequency ω→Ωi\omega\to\Omega_i, the local relative error is

lim⁡ω→Ωi∣ρ^(ω)−ρ(ω)ρ(ω)∣=ϵi,∀i=1,…,Np.(43)\lim_{\omega\to\Omega_i}\left\lvert\frac{\widehat{\rho}(\omega)-\rho(\omega)}{\rho(\omega)}\right\rvert=\epsilon_i,\qquad\forall i=1,\ldots,N_p. \tag*{(43)}

It is worth emphasizing that (42) and (43) hold only approximately if the Dirac delta function is approximated by a finite width function.

If we imagine the level density function ρ(ω)\rho(\omega) as the linear response function with all transition probabilities being equal to one, we can see that the level density estimator ρ^(ω)\widehat{\rho}(\omega) is the linear response function with the proper QRPA eigenfrequencies Ωi\Omega_i but transition probabilities with relative error bounded by ϵi\epsilon_i for each Ωi\Omega_i. We also notice that ϵi\epsilon_i is a measure of how much the squared norm of the ii-th column of the YY matrix deviates from the average value over all columns. In practice, the norm ∥Y∥F\lVert Y\rVert_F is relatively small because the matrix YY describes the ground-state correlations [25]. If we expand the residual interaction δHμν20\delta H_{\mu\nu}^{20} and δHμν02\delta H_{\mu\nu}^{02} in terms amplitudes XμνX_{\mu\nu} and YμνY_{\mu\nu}:

δHμν20=−(Eμ+Eν)Xμν+∑μ′<ν′[Aμν,μ′ν′Xμ′ν′+Bμν,μ′ν′Yμ′ν′],δHμν02=−(Eμ+Eν)Yμν+∑μ′<ν′[Bμν,μ′ν′∗Xμ′ν′+Aμν,μ′ν′∗Yμ′ν′],(44)\begin{aligned} \delta H_{\mu\nu}^{20}&=-(E_\mu+E_\nu)X_{\mu\nu}+\sum_{\mu'<\nu'}\left[A_{\mu\nu,\mu'\nu'}X_{\mu'\nu'}+B_{\mu\nu,\mu'\nu'}Y_{\mu'\nu'}\right],\\ \delta H_{\mu\nu}^{02}&=-(E_\mu+E_\nu)Y_{\mu\nu}+\sum_{\mu'<\nu'}\left[B_{\mu\nu,\mu'\nu'}^{*}X_{\mu'\nu'}+A_{\mu\nu,\mu'\nu'}^{*}Y_{\mu'\nu'}\right], \tag*{(44)} \end{aligned}

and assume that the residual interaction is relatively weak, i.e., ∣δHμν20∣≪1\lvert\delta H_{\mu\nu}^{20}\rvert\ll1 and ∣δHμν02∣≪1\lvert\delta H_{\mu\nu}^{02}\rvert\ll1 in addition to ground-state correlations being small (∥Y∥F≪1\lVert Y\rVert_F\ll1), the QRPA matrix can be simplified to

Aμν,μ′ν′≈(Eμ+Eν)δμν,μ′ν′,(45)A_{\mu\nu,\mu'\nu'}\approx(E_\mu+E_\nu)\delta_{\mu\nu,\mu'\nu'}, \tag*{(45)}
Bμν,μ′ν′≈0,(46)B_{\mu\nu,\mu'\nu'}\approx0, \tag*{(46)}

which is known as the cranking approximation. In this case, the QRPA eigenproblem (3) has a trivial solution: X=IX=I, Y=0Y=0, and all the relative errors ϵi\epsilon_i are identically equal to zero. If instead we assume Bμν,μ′ν′≈0B_{\mu\nu,\mu'\nu'}\approx0 while retaining the full matrix AA, then the QRPA eigenproblem (3) again has a trivial solution: X=QX=Q, Y=0Y=0, where the unitary matrix Q∈CNp×NpQ\in\mathbb{C}^{N_p\times N_p} diagonalizes the Hermitian matrix: A=QΛQ†A=Q\Lambda Q^{\dagger}. The relative errors ϵi\epsilon_i are then again zero. Because AA has dominant diagonal elements Eμ+EνE_\mu+E_\nu, we can expect that the eigenvectors Yμ,νiY_{\mu,\nu}^{i} will be generally small in magnitude, except for a few modes with small eigenfrequencies Ωi\Omega_i for which the diagonal elements Eμ+EνE_\mu+E_\nu are relatively small. This is the reason why we can a priori expect very small relative errors ϵi\epsilon_i, for all i=1,…,Npi=1,\ldots,N_p, except for a few low-lying modes, and, therefore, why the introduced estimator ρ^(ω)\widehat{\rho}(\omega) is believed to be a sufficiently good approximation of ρ(ω)\rho(\omega).

Practical aspects

Sampling the level density estimator

In practice, random variables can only be sampled, and the expected values in Eq. (36) can be estimated by computing average values over a collection of samples. If we generate a collection of NsN_s samples of complex vectors Fis20,Fis02∈CNpF_{i_s}^{20},F_{i_s}^{02}\in\mathbb{C}^{N_p}, for is=1,…,Nsi_s=1,\ldots,N_s, such that their real and imaginary parts are independently drawn from a Gaussian distribution with zero expected value and variance σ2\sigma^2 as in Eq. (22), and if for each of those generated excitation operators F^is\widehat{F}_{i_s} we compute the response function dB/dω(ω,F^is)dB/d\omega(\omega,\widehat{F}_{i_s}) on the positive frequencies domain ω>0\omega>0, then we can approximate the level density estimator (36) as

ρ^(ω)=Np1Ns∑is=1NsdB(ω≥0,F^is)dω1Ns∑is=1Ns∫0+∞dB(ω′≥0,F^is)dω′ dω′.(47)\widehat{\rho}(\omega)=N_p\frac{\frac{1}{N_s}\sum_{i_s=1}^{N_s}\frac{dB(\omega\geq0,\widehat{F}_{i_s})}{d\omega}}{\frac{1}{N_s}\sum_{i_s=1}^{N_s}\int_0^{+\infty}\frac{dB(\omega'\geq0,\widehat{F}_{i_s})}{d\omega'}\,d\omega'}. \tag*{(47)}

Notice that the denominator in the previous equation is an average over all zeroth moments m0(F^is)m_0(\widehat{F}_{i_s}). Of course, the larger the sample size NsN_s, the better the approximation but the approximation of the level density function does not depend on the variance σ2\sigma^2. In principle, one can select any value; in practice, a reasonable choice is to take σ≈1\sigma\approx1, whichever measuring unit we express the excitation operator F^\widehat{F}. It is worth mentioning that (42) and (43) hold in the Ns→+∞N_s\to+\infty limit where the expected value matches the computed average value from the collection of samples.

Kernel Polynomial Method

The estimator (47) seems very practical at first, since any method which can evaluate the response function dB/dω(ω,F^)dB/d\omega(\omega,\widehat{F}) for an arbitrary excitation operator F^\widehat{F}, as well as the integral in the denominator of Eq. (47) is applicable. For example, one can solve the classical point-by-point FAM equations to compute the numerator on a fine ω\omega grid with small smearing γ\gamma, or use the iterative Arnoldi method from Ref. [30]. The denominator can be evaluated by employing a contour integration technique, as presented in Ref. [29]. However this calculation needs to be performed both over a very large ω\omega domain with small smearing γ\gamma and for sufficiently large sample size NsN_s. As a result, the computational cost of the method could become prohibitive. This issue can be largely mitigated by employing the Kernel Polynomial Method (KPM) introduced in Ref. [28]. The KPM is specifically designed to efficiently compute the response function over the entire ω≥0\omega\ge0 region of interest for arbitrary excitation operator. In the following we briefly introduce the KPM method and apply it to evaluate the expression in Eq. (47).

We define the bounding frequency Ωb\Omega_b, such that all the QRPA eigenmodes are contained within the interval: Ωi⊆⟨0,+Ωb⟩\Omega_i \subseteq\langle0,+\Omega_b\rangle. The main idea of the KPM method is to expand the response function in Chebyshev series

dB(KPM)(ω,F^)dω=2/πΩb1−(ω/Ωb)2∑n=02Nitμn(F^)Tn(ωΩb)(48)\frac{dB^{(\mathrm{KPM})}(\omega,\hat{F})}{d\omega} = \frac{2/\pi}{\Omega_b\sqrt{1-(\omega/\Omega_b)^2}} \sum_{n=0}^{2N_{\mathrm{it}}}\mu_n(\hat{F})T_n\left(\frac{\omega}{\Omega_b}\right) \tag*{(48)}

where the coefficients {μn(F^)}n=02Nit\{\mu_n(\hat{F})\}_{n=0}^{2N_{\mathrm{it}}} are computed via the recursive relations given in Ref. [28]. The computation of coefficients {μn(F^)}n=02Nit\{\mu_n(\hat{F})\}_{n=0}^{2N_{\mathrm{it}}} requires exactly NitN_{\mathrm{it}} evaluations of the QRPA mapping (8). In order to damp the Gibbs oscillations, one needs to multiply these coefficients with the kernel coefficients {gn(2Nit+1)}n=02Nit\{g_n^{(2N_{\mathrm{it}}+1)}\}_{n=0}^{2N_{\mathrm{it}}}: μn(F^)←gn(2Nit+1)μn(F^)\mu_n(\hat{F}) \leftarrow g_n^{(2N_{\mathrm{it}}+1)}\mu_n(\hat{F}). In this paper, we use the Jackson kernel [28] unless stated otherwise. Then, the obtained KPM response function is approximately equal to

dB(KPM)(ω,F^)dω≈+∑i=1Np∣⟨i∣F^∣0⟩∣2δσKPM(ω−Ωi)−∑i=1Np∣⟨0∣F^∣i⟩∣2δσKPM(ω+Ωi)(49)\frac{dB^{(\mathrm{KPM})}(\omega,\hat{F})}{d\omega} \approx +\sum_{i=1}^{N_p}|\langle i|\hat{F}|0\rangle|^2\delta_{\sigma_{\mathrm{KPM}}}(\omega-\Omega_i) -\sum_{i=1}^{N_p}|\langle0|\hat{F}|i\rangle|^2\delta_{\sigma_{\mathrm{KPM}}}(\omega+\Omega_i) \tag*{(49)}

where δσKPM(ω)\delta_{\sigma_{\mathrm{KPM}}}(\omega) is the Gaussian function:

δσKPM(ω)=12πσKPMexp⁡(−12ω2σKPM2)(50)\delta_{\sigma_{\mathrm{KPM}}}(\omega)= \frac{1}{\sqrt{2\pi}\sigma_{\mathrm{KPM}}} \exp\left(-\frac{1}{2}\frac{\omega^2}{\sigma_{\mathrm{KPM}}^2}\right) \tag*{(50)}

of width σKPM\sigma_{\mathrm{KPM}}

σKPM=Ωbπ2Nit+1.(51)\sigma_{\mathrm{KPM}}=\Omega_b\frac{\pi}{2N_{\mathrm{it}}+1}. \tag*{(51)}

The number NitN_{\mathrm{it}} of KPM iterations is determined by the desired resolution σKPM\sigma_{\mathrm{KPM}} according to Eq. (51). In the limit Nit→+∞N_{\mathrm{it}}\to+\infty, we have indeed: dB(KPM)/dω→dB/dωdB^{(\mathrm{KPM})}/d\omega\to dB/d\omega. It is not difficult to see that the zeroth moment m0(F^)m_0(\hat{F}) of (48) is

m0(F^)=∫0+ΩbdB(KPM)(ω′,F^)dω′ dω′=∑n=02Nitμn(F^)sin⁡(nπ2)(nπ2).(52)m_0(\hat{F})= \int_0^{+\Omega_b}\frac{dB^{(\mathrm{KPM})}(\omega',\hat{F})}{d\omega'}\,d\omega' =\sum_{n=0}^{2N_{\mathrm{it}}}\mu_n(\hat{F})\frac{\sin\left(\frac{n\pi}{2}\right)}{\left(\frac{n\pi}{2}\right)}. \tag*{(52)}

In order to evaluate the level density estimator ρ^(ω)\hat{\rho}(\omega) given in Eq. (47), one runs the KPM calculation for a collection of NsN_s random operators {F^is}is=1Ns\{\hat{F}_{i_s}\}_{i_s=1}^{N_s}. After obtaining the collection of coefficients: {(μn(F^is))n=02Nit}is=1Ns\{(\mu_n(\hat{F}_{i_s}))_{n=0}^{2N_{\mathrm{it}}}\}_{i_s=1}^{N_s}, the QRPA level density estimator is given by

ρ^(ω)=Np∑n=02Nitμn(F^is)‾2/πΩb1−(ω/Ωb)2Tn(ωΩb)∑n=02Nitμn(F^is)‾sin⁡(nπ2)(nπ2).(53)\hat{\rho}(\omega)=N_p \frac{ \sum_{n=0}^{2N_{\mathrm{it}}}\overline{\mu_n(\hat{F}_{i_s})} \frac{2/\pi}{\Omega_b\sqrt{1-(\omega/\Omega_b)^2}} T_n\left(\frac{\omega}{\Omega_b}\right) }{ \sum_{n=0}^{2N_{\mathrm{it}}}\overline{\mu_n(\hat{F}_{i_s})} \frac{\sin\left(\frac{n\pi}{2}\right)}{\left(\frac{n\pi}{2}\right)} }. \tag*{(53)}

where μn(F^is)‾\overline{\mu_n(\hat{F}_{i_s})} are the averaged coefficients over the collection of randomly generated samples

μn(F^is)‾=1Ns∑is=1Nsμn(F^is),∀n=0,1,…,2Nit.(54)\overline{\mu_n(\hat{F}_{i_s})} =\frac{1}{N_s}\sum_{i_s=1}^{N_s}\mu_n(\hat{F}_{i_s}),\qquad\forall n=0,1,\ldots,2N_{\mathrm{it}}. \tag*{(54)}

The KPM computation for each isi_s is independent from the others and therefore can be run fully in parallel.

Spurious modes removal

In practical calculations, spurious modes, also know as the Nambu-Goldstone modes [31], often contaminate the response function due to their high transition probabilities at positive near-zero frequencies. In most nuclei, the spurious pairing Kπ=0+K^\pi=0^+, rotational Kπ=1+K^\pi=1^+ and translational Kπ=0−,1−K^\pi=0^-,1^- modes in the response function will result in a corresponding spurious contribution to the calculated level density function.

The idea is to eliminate the contribution of spurious modes in the low-energy region by shifting them to high-energy region. Once the spurious mode with eigenfrequency Ω1>0\Omega_1>0 and eigenvector x1,y1∈CNpx_1,y_1\in\mathbb{C}^{N_p} is identified, the only thing that needs to be modified is the QRPA mapping (8) which now becomes

\begin{bmatrix}x\\y\end{bmatrix} \mapsto \begin{bmatrix}A&B\\B^*&A^*\end{bmatrix} \begin{bmatrix}x\\y\end{bmatrix} +(\widetilde{\Omega}_1-\Omega_1) \left\{ (x_1^\dagger x-y_1^\dagger y) \begin{bmatrix}+x_1\\-y_1\end{bmatrix} +(x_1^T y-y_1^T x) \begin{bmatrix}-y_1^*\\+x_1^*\end{bmatrix} \right\}, \tag*{(55)}\end{bmatrix}

where Ω~1>0\widetilde{\Omega}_1>0 is the position of the shifted spurious mode. In practice most spurious modes have Ω1<0.5\Omega_1<0.5 MeV. By choosing large values of Ω~1>0\widetilde{\Omega}_1>0, e.g. Ω~1≥200\widetilde{\Omega}_1\ge200 MeV, one can shift these modes to arbitrarily high energy. Additional details on the practical implementation of this technique are given in Appendix A.

Algorithm

Algorithm 1 summarizes the main steps of our method. The total number #QFAM\#_{\mathrm{QFAM}} of FAM iterations required is

#QFAM=NsNit=Ns⌈12(ΩbπσKPM−1)⌉(56)\#_{\mathrm{QFAM}}=N_sN_{\mathrm{it}}=N_s\left\lceil\frac{1}{2}\left(\frac{\Omega_b\pi}{\sigma_{\mathrm{KPM}}}-1\right)\right\rceil \tag*{(56)}

where NitN_{\mathrm{it}} is obtained from Eq. (51). We emphasize again that the loop in Algorithm 1 is naturally parallelizable. For example, in the case of a non-relativistic FAM solver, if one picks Ωb≈250\Omega_b\approx250 MeV and σKPM=0.05\sigma_{\mathrm{KPM}}=0.05 MeV, it takes around Nit≈8000N_{\mathrm{it}}\approx8000 FAM iterations per sample. If we assume that each FAM iteration lasts around ≈5\approx5 s, and that we can launch Ns≈200N_s\approx200 in parallel, one can expect to compute the level density function in ≈11\approx11 hours, which makes the method feasible in practice.

Algorithm 1: QRPA level density with FAM+KPM

Input:

  • Mapping: [xy]∈C2Np↦[ABB∗A∗][xy]∈C2Np\begin{bmatrix}x\\y\end{bmatrix}\in\mathbb{C}^{2N_p}\mapsto\begin{bmatrix}A&B\\B^*&A^*\end{bmatrix}\begin{bmatrix}x\\y\end{bmatrix}\in\mathbb{C}^{2N_p}, provided by a FAM solver.

  • Width σKPM\sigma_{\mathrm{KPM}} of Gaussian approximation of Dirac delta function.

  • Bounding frequency Ωb\Omega_b: ∀Ωi, Ωi∈⟨0,+Ωb⟩\forall\Omega_i,\ \Omega_i\in\langle0,+\Omega_b\rangle.

  • Optionally provide: spurious mode eigenfrequency Ω1>0\Omega_1>0, spurious mode eigenvector x1,y1∈CNpx_1,y_1\in\mathbb{C}^{N_p}, and a desired position of the shifted spurious eigenfrequency Ω~1\widetilde{\Omega}_1 such that Ω~1∈⟨0,+Ωb⟩\widetilde{\Omega}_1\in\langle0,+\Omega_b\rangle. How to obtain {Ω1,x1,y1}\{\Omega_1,x_1,y_1\} is explained in Appendix A.

  • Number of samples NsN_s in a collection of randomly generated excitation operators {F^is}is=1Ns\{\hat{F}_{i_s}\}_{i_s=1}^{N_s}.

for is=1,2,…,Nsi_s=1,2,\ldots,N_s do

  • Generate a random operator F^is\hat{F}_{i_s} defined by vectors: Re⁡[Fis20],Re⁡[Fis02],Im⁡[Fis20],Im⁡[Fis02]∈RNp\operatorname{Re}[F_{i_s}^{20}],\operatorname{Re}[F_{i_s}^{02}],\operatorname{Im}[F_{i_s}^{20}],\operatorname{Im}[F_{i_s}^{02}]\in\mathbb{R}^{N_p}, where each component is drawn independently from N(0,σ2)\mathcal{N}(0,\sigma^2).

  • Compute coefficients μn(F^is)\mu_n(\hat{F}_{i_s}) using spurious-corrected mapping (55) as described in Ref. [28]. The number NitN_{\mathrm{it}} of FAM iterations needed is computed from Eq. (51).

  • Apply the kernel transformation: μn(F^is)←gn(2Nit+1)μn(F^is)\mu_n(\hat{F}_{i_s})\leftarrow g_n^{(2N_{\mathrm{it}}+1)}\mu_n(\hat{F}_{i_s}) with the Jackson kernel coefficients gn(2Nit+1)g_n^{(2N_{\mathrm{it}}+1)} given in Ref. [28].

end

  • Compute the averaged coefficients μn(F^is)\mu_n(\hat{F}_{i_s}) as in Eq. (54).

  • Using Eq. (53), evaluate ρ^(ω)\widehat{\rho}(\omega) on a grid of frequencies ω≥0\omega\geq0, e.g., with the FFT-based method of Ref. [28].

Output:

  • QRPA level density estimator ρ^(ω)\widehat{\rho}(\omega) evaluated on a ω≥0\omega\geq0 grid.

Numerical tests

In this section, we show a series of numerical tests to evaluate the performance of the new method. We first compute the QRPA level density for synthetically generated QRPA-like matrices to illustrate the convergence of the method as we increase the number of samples NsN_s. We then use the open-source code skyrme_rpa [32] to solve the RPA equation in heavy spherical nuclei and benchmark the QRPA level density of specific KπK^\pi modes.

Examples using synthetically generated QRPA matrix

In this section we synthetically generate QRPA-like matrices AA and BB by employing the procedure described in Appendix B of Ref. [28]. The matrix dimension and the bounding frequency are set to Np=200N_p=200 and Ωb=250\Omega_b=250 MeV, respectively. 100 random eigenfrequencies Ωi\Omega_i are generated uniformly in the range from 0 MeV to 200 MeV and combined with another 100 random eigenfrequencies Ωi\Omega_i generated uniformly in the range from 0 MeV to 30 MeV, with the intent of making the resulting QRPA spectrum more dense in the low energy region.

Next, we generate a uniformly random sequence of values θ1,…,θNp\theta_1,\ldots,\theta_{N_p} in the interval [0,θmax⁡][0,\theta_{\max}], and two unitary matrices C,D∈CNp×NpC,D\in\mathbb{C}^{N_p\times N_p} as QQ factors in the QR decomposition of two random Np×NpN_p\times N_p complex matrices. The XX and YY matrices are constructed as

X=Ddiag⁡[cos⁡θi]i=1NpCandY=D∗diag⁡[sinh⁡θi]i=1NpC,(57)X=D\operatorname{diag}\left[\cos\theta_i\right]_{i=1}^{N_p}C \quad\text{and}\quad Y=D^*\operatorname{diag}\left[\sinh\theta_i\right]_{i=1}^{N_p}C, \tag*{(57)}

and then used to generate the QRPA matrices AA and BB as

A=XΩX†+(YΩY†)∗andB=−XΩY†−(XΩY†)T.(58)A=X\Omega X^\dagger+(Y\Omega Y^\dagger)^* \quad\text{and}\quad B=-X\Omega Y^\dagger-(X\Omega Y^\dagger)^T. \tag*{(58)}

Note that the parameter θmax⁡\theta_{\max} determines the magnitudes of ∥Yi∥2\lVert Y_i\rVert^2 and therefore the corresponding theoretical relative error ϵi\epsilon_i in the limit Ns→+∞N_s\to+\infty, according to Eq. (42).

We set the width for the Gaussian approximation of the Dirac delta function equal to σKPM=0.05\sigma_{\mathrm{KPM}}=0.05 MeV, which for Ωb=250\Omega_b=250 MeV according to Eq. (56) corresponds to the number Nit≈8000N_{\mathrm{it}}\approx8000 of QFAM iterations needed in the KPM method. We draw the components of vectors Fis20,Fis02∈CNpF_{i_s}^{20},F_{i_s}^{02}\in\mathbb{C}^{N_p} according to Eq. (22) with a standard deviation of σ=1\sigma=1 MeV. Recall that the error on the estimator ρ^(ω)\widehat{\rho}(\omega) is independent of the value of σ\sigma.

Figure 1 shows the level density ρ(ω)\rho(\omega) and its estimator ρ^(ω)\widehat{\rho}(\omega) computed for a very large number of samples Ns=5000N_s=5000 and θmax⁡=0.5\theta_{\max}=0.5.

Figure 1. Comparison of the level density function ρ(ω)\rho(\omega) and its estimator ρ^(ω)\widehat{\rho}(\omega) computed for a very large number of samples Ns=5000N_s=5000 and θmax⁡=0.5\theta_{\max}=0.5. ρ(ω)\rho(\omega) is calculated by approximating the Dirac delta function with a Gaussian function, using a standard deviation of σKPM\sigma_{\mathrm{KPM}}.

Example with fixed θmax⁡\theta_{\max} and varying NsN_s

In this subsection we fix the value θmax⁡=1\theta_{\max}=1 and vary the number of randomly generated excitation operators Ns=50,200,800N_s=50,200,800. In Figure 2 we show the obtained relative error evaluated at QRPA eigenfrequencies ω=Ωi\omega=\Omega_i between the level density estimator ρ^(ω)\widehat{\rho}(\omega) compared to the true level density function ρ(ω)\rho(\omega) when the Dirac delta function in Eq. (1) is replaced by a Gaussian with standard deviation of σKPM\sigma_{\mathrm{KPM}}. Figure 2 also shows the theoretical relative error limit ϵi\epsilon_i in the case of an infinitely large collection of random excitation operators Ns→+∞N_s\to+\infty as defined in Eq. (41). We can see that for Ns=800N_s=800, the obtained relative error is of the same order ≈5−10%\approx5-10\% as the theoretical limit. Therefore, if θmax⁡=1\theta_{\max}=1, a few hundreds of randomly generated excitation operators are enough to give an estimate up to ≈10%\approx10\% relative error. We remind the reader that (42) and (43) hold in the limit of an infinitesimally small width of Dirac delta functions and the Ns→+∞N_s\to+\infty limit. When using a finite-width approximation of the Dirac delta function and a finite number of samples NsN_s, these relations are only approximately valid.

Figure 2. The blue curve with plain circles shows a relative error evaluated at eigenfrequencies ω=Ωi\omega=\Omega_i between the level density ρ(ω)\rho(\omega) and its estimator ρ^(ω)\hat{\rho}(\omega) computed using Ns=50N_s=50, 200200 and 800800 randomly generated excitation operators and θmax⁡=1\theta_{\max}=1. The red curve with open squares shows the theoretical relative error ϵi\epsilon_i in the Ns→+∞N_s\to+\infty limit.

Example with fixed NsN_s and varying θmax⁡\theta_{\max}

In this subsection we fix the number of samples NsN_s to a very large value Ns=5000N_s=5000 and vary the parameter θmax⁡=0.5\theta_{\max}=0.5, 1.01.0, 2.02.0. Recall that small values of θmax⁡\theta_{\max} correspond to small values of the matrix elements of YY, hence a smaller error ϵi\epsilon_i owing to Eq. (41). Like in the previous subsection, in Figure 3 we compare the level density estimator ρ^(ω)\hat{\rho}(\omega) with the true level density ρ(ω)\rho(\omega) by showing the theoretical relative error limit ϵi\epsilon_i and the computed relative error ∣ρ^(ω)−ρ(ω)∣/∣ρ(ω)∣|\hat{\rho}(\omega)-\rho(\omega)|/|\rho(\omega)| at ω=Ωi\omega=\Omega_i. We first notice that as the value of θmax⁡\theta_{\max} is reduced, the theoretical relative error limit ϵi\epsilon_i is reduced indeed. Secondly, even for a large number of samples Ns=5000N_s=5000, the computed relative error is primarily dominated by statistical errors due to the variations within the finite size of the collection of random samples, rather than the theoretical error limit. This implies that even under the most favorable condition of θmax⁡=0\theta_{\max}=0, the resulting error would be approximately 5%, even with a very large sample size of Ns=5000N_s=5000.

Figure 3. The blue curve with plain circles shows a relative error evaluated at eigenfrequencies ω=Ωi\omega=\Omega_i between the level density ρ(ω)\rho(\omega) and its estimator ρ^(ω)\hat{\rho}(\omega) computed using a very large number Ns=5000N_s=5000 of randomly generated excitation operators and θmax⁡=0.5\theta_{\max}=0.5, 1.01.0, 2.02.0. The red curve with open squares shows the theoretical relative error ϵi\epsilon_i in the Ns→+∞N_s\to+\infty limit.

Examples using matrix RPA code skyrme_rpa

In this section, we validate the proposed method with the realistic RPA solver skyrme_rpa [32]. This solver is implemented for closed shell spherical nuclei with Skyrme-type interactions. The Hartree-Fock equations are solved on a radial mesh using box boundary conditions and the RPA matrix is explicitly constructed and diagonalized for a given value of total angular momentum and parity JπJ^\pi.

We have modified the skyrme_rpa code to extract explicitly the matrices AA and BB for given JπJ^\pi with the intent of using them rather than synthetically generated ones as was done in the previous section. This shall provide an insight of realistic values for Ωi\Omega_i and θi\theta_i which were previously randomly generated. We use the desired width σKPM=0.05\sigma_{\mathrm{KPM}}=0.05 MeV, bounding frequency Ωb=250\Omega_{\mathrm{b}}=250 MeV, and standard deviation of randomly generated excitation operators σ=1\sigma=1 MeV. Guided by the previous synthetic examples, we choose the number of samples Ns=500N_s=500 as a compromise between computational complexity and statistical fluctuations.

J=5−J=5^{-} level density of 120Sn^{120}\mathrm{Sn}

We use the same numerical setup as in § 4.2 of Ref. [28]. Calculations were performed by employing the SLy5 Skyrme interaction on 120Sn^{120}\mathrm{Sn} in a 20 fm radius box with 0.1 fm radial step and 100 MeV cutoff energy; see Ref. [32] for details. The resulting RPA matrix for Jπ=5−J^{\pi}=5^{-} is of order Np=1310N_p=1310. First we solve the QRPA eigenvalue problem as in Eqs. (3) and (4) and obtain the eigenfrequencies Ωi\Omega_i, and matrices XX, YY. The values of θ1,…,θNp\theta_1,\ldots,\theta_{N_p} are obtained by computing the singular values of either XX or YY matrix. The maximum angle obtained is equal to: max⁡i=1,…,Npθi=0.38\max_{i=1,\ldots,N_p}\theta_i=0.38. As demonstrated in the previous section, this a favourable scenario regarding the theoretical error εi\varepsilon_i in the Ns→+∞N_s\rightarrow+\infty limit. In Figure 4 we compare the computed level density estimator against the true function. We notice that the obtained relative error is the largest for the eigenmode with the lowest eigenfrequency. This is expected because this mode has the highest ground-state correlation energy and therefore the largest ∥Yei∥2\lVert Y e_i\rVert^2, which yields largest εi\varepsilon_i. Other than the lowest mode, the obtained relative error is of order 5--10% for other modes, and is mostly influenced by statistical fluctuations.

Figure 4. Theoretical relative error εi\varepsilon_i in the limit Ns→+∞N_s\rightarrow+\infty and computed relative error: ∣ρ(Ωi)−ρ^(Ωi)∣/∣ρ(Ωi)∣|\rho(\Omega_i)-\widehat{\rho}(\Omega_i)|/|\rho(\Omega_i)| for Jπ=5−J^\pi=5^- states in 120Sn^{120}\mathrm{Sn}. Calculations were performed with Ns=500N_s=500 randomly generated excitation operators; see text for details about the RPA calculation.

J=1−J=1^{-} level density of 208Pb^{208}\mathrm{Pb}

In this example we use the SLy5 Skyrme interaction on 208Pb^{208}\mathrm{Pb} in a box of radius 24 fm with 0.1 fm radial step and 100 MeV cutoff energy. The resulting RPA matrix for Jπ=1−J^\pi=1^- is of order Np=1394N_p=1394. In this case, the maximum angle θi\theta_i is equal to: max⁡i=1,…,Npθi=1.08\max_{i=1,\ldots,N_p}\theta_i=1.08, which is why we expect a somewhat larger error than in the previous example. The top panel of Figure 5 shows the comparison between the computed level density estimator and the correct level density function. The low-lying spurious 1−1^- mode is at Ω1≈1\Omega_1\approx1 MeV.

Figure 5. Top panel: Level density estimator ρ^(ω)\widehat{\rho}(\omega), true density ρ(ω)\rho(\omega) and absolute difference of the two for Jπ=1−J^\pi=1^- states in 208Pb^{208}\mathrm{Pb} as a function of the energy ω\omega. Calculations were performed with Ns=500N_s=500 samples. Middle panel: Relative difference between ρ(ω)\rho(\omega) and ρ^(ω)\widehat{\rho}(\omega) and the theoretical relative error limit εi\varepsilon_i as a function of the energy Ωi\Omega_i of the RPA eigenstate. Bottom panel: Same as middle panel after moving the spurious mode to Ω~1=200\widetilde{\Omega}_1=200 MeV with the procedure described in Appendix A.

In the middle panel of Figure 5, we plot the theoretical relative error εi\varepsilon_i in the limit Ns→+∞N_s\rightarrow+\infty and the computed relative error ∣ρ(Ωi)−ρ^(Ωi)∣/∣ρ(Ωi)∣|\rho(\Omega_i)-\widehat{\rho}(\Omega_i)|/|\rho(\Omega_i)|. We notice that the low-lying spurious 1−1^- mode yields a very large relative error. This is expected because such spurious modes have a very large norm ∥Yei∥2\lVert Y e_i\rVert^2 and consequently large relative error εi\varepsilon_i in the Ns→+∞N_s\rightarrow+\infty limit; see Eq. (41). In this particular case, εi=334%\varepsilon_i=334\% for the spurious mode.

In order to eliminate the contribution of the spurious mode to the level density in the low-energy region, we follow the method outlined in Appendix A and shift the spurious mode to high energy Ω~1=200\widetilde{\Omega}_1=200 MeV. The bottom panel of Figure 5 shows the relative error after this operation. Now the obtained relative error is of order 5--10% and is mostly influenced by statistical fluctuations.

Conclusion

In this work, we have proposed a new method to approximate the QRPA level density, that is, the density of vibrational states that result from the diagonalization of the QRPA matrix. Our approach relies on using the Finite Amplitude Method to compute the linear responses for a collection of random excitation operators that probe the QRPA spectrum. While the error on the level density decreases with the number of random operators, i.e. the number of samples, we derive a theoretical relative error bound in the case of infinite number of samples. This bound is expected to be relatively small; at the cranking approximation, it is in fact exactly zero. We use the Kernel Polynomial Method to efficiently compute the estimator of the QRPA level density. We illustrate the convergence of the method with respect to the number of samples for various magnitudes of ∥Y∥F\lVert Y\rVert_F by synthetically generating QRPA-like matrices. Using publicly available RPA code, we also show that we can approximate the RPA level density to within 5–10%.

While our method is no substitute to exact diagonalization of the QRPA matrix, it offers two main advantages: (i) it is easily applicable to heavy deformed nuclei, where the cost of building the QRPA matrix can quickly become prohibitive (ii) computationally, it is a naturally parallel problem since the linear response of each sample can be computed independently. The next step will be to use such QRPA level densities to compute the full level density of the nucleus by folding it with rotational bands.

Acknowledgements

Discussions with Tong Li are warmly acknowledged. This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under contract DE-AC52-07NA27344. Lawrence Livermore National Security, LLC. Computing support came from the Lawrence Livermore National Laboratory (LLNL) Institutional Computing Grand Challenge program.

Removal of spurious modes

The main idea stems from the following facts. If H∈Cn×nH \in\mathbb{C}^{n\times n} is a Hermitian matrix with eigenvalues: λ1,…,λn∈R\lambda_1,\ldots,\lambda_n \in\mathbb{R}, and corresponding eigenvectors: u1,…,un∈Cnu_1,\ldots,u_n \in\mathbb{C}^n,

H=∑i=1nλiuiui†,(59)H = \sum_{i=1}^{n} \lambda_i u_i u_i^\dagger, \tag*{(59)}

then for any λ~1∈R\widetilde{\lambda}_1 \in\mathbb{R}, the updated rank-1 Hermitian matrix

H~=H+(λ~1−λ1)u1u1†(60)\widetilde{H}=H+(\widetilde{\lambda}_1-\lambda_1)u_1u_1^\dagger \tag*{(60)}

has eigenvalues: λ~1,λ2,…,λn∈R\widetilde{\lambda}_1,\lambda_2,\ldots,\lambda_n \in\mathbb{R} and the same corresponding eigenvectors as the matrix HH: u1,u2,…,un∈Cnu_1,u_2,\ldots,u_n \in\mathbb{C}^n. This property can be applied to the QRPA matrix SS of Eq. (2). Let us introduce a shorthand notation for the columns of the XX and YY matrices of eigenvectors: xi=Xeix_i=Xe_i and yi=Yeiy_i=Ye_i, for i=1,…,Npi=1,\ldots,N_p. For any positive (Ω~i)i=1Np(\widetilde{\Omega}_i)_{i=1}^{N_p}, let us define the modified matrices,

A~=A+∑i=1Np(Ω~i−Ωi)[+xixi†+(yiyi†)∗],(61)\widetilde{A}=A+\sum_{i=1}^{N_p}(\widetilde{\Omega}_i-\Omega_i)\left[+x_i x_i^\dagger+(y_i y_i^\dagger)^*\right], \tag*{(61)}
B~=B+∑i=1Np(Ω~i−Ωi)[−xiyi†−(xiyi†)T].(62)\widetilde{B}=B+\sum_{i=1}^{N_p}(\widetilde{\Omega}_i-\Omega_i)\left[-x_i y_i^\dagger-(x_i y_i^\dagger)^T\right]. \tag*{(62)}

Then the new QRPA matrix where A,BA,B are replaced by A~,B~\widetilde{A},\widetilde{B} still satisfies both Eqs. (3) and (4) with the same eigenvectors XX and YY only with eigenfrequencies Ω~i\widetilde{\Omega}_i instead of Ωi\Omega_i. The QRPA mapping (8) acquires the new form

\begin{bmatrix}x\\y\end{bmatrix} \mapsto \begin{bmatrix}\widetilde{A}&\widetilde{B}\\\widetilde{B}^*&\widetilde{A}^*\end{bmatrix} \begin{bmatrix}x\\y\end{bmatrix} = \begin{bmatrix}A&B\\B^*&A^*\end{bmatrix} \begin{bmatrix}x\\y\end{bmatrix} + \sum_{i=1}^{N_p}(\widetilde{\Omega}_i-\Omega_i) \left\{ (x_i^\dagger x-y_i^\dagger y) \begin{bmatrix}+x_i\\-y_i\end{bmatrix} + (x_i^T y-y_i^T x) \begin{bmatrix}-y_i^*\\+x_i^*\end{bmatrix} \right\}. \tag*{(63)}\end{bmatrix}

In practice, we often want to shift only a single spurious mode Ω1>0\Omega_1>0 from near-zero value to some other value Ω~1>0\widetilde{\Omega}_1>0, away from the low-energy region we are typically interested in. If we somehow manage to obtain the eigenfrequency Ω1>0\Omega_1>0 of this spurious mode Ω1>0\Omega_1>0, then we can simply apply the formula (A.5) with Np=1N_p=1 to obtain exactly the same response function but with shifted spurious mode Ω1→Ω~1\Omega_1\to\widetilde{\Omega}_1.

The question is therefore how to obtain the low-lying QRPA modes Ωi\Omega_i and the corresponding (xi,yi)(x_i,y_i), i=1,…,ki=1,\ldots,k, for a subset of kk modes. There are efficient methods for computing the kk eigenvalues with the smallest norm of a general matrix H∈Cn×nH\in\mathbb{C}^{n\times n} which only require the inverse mapping: x↦H−1xx\mapsto H^{-1}x. One such method is included in ARPACK [33], a numerical software library for solving large scale eigenvalue problems using the implicitly restarted Arnoldi method.

Consider the mapping

\begin{bmatrix}f^{20}\\f^{02}\end{bmatrix} \mapsto \left( \begin{bmatrix}+I&0\\0&-I\end{bmatrix} \begin{bmatrix}A&B\\B^*&A^*\end{bmatrix} \right)^{-1} \begin{bmatrix}f^{20}\\f^{02}\end{bmatrix}. \tag*{(64)}\end{bmatrix}

According to Eq. (7), the mapping (A.6) corresponds to solving the linear response equation for zero complex frequency ωy=0\omega_y=0 and excitation operator given by F20=−f20F^{20}=-f^{20} and F02=+f02F^{02}=+f^{02}. This linear response equation is precisely what the FAM solves. Therefore, with the help of e.g. ARPACK, one can compute the 2k2k eigenvalues with the smallest norm: ±Ω1,…,±Ωk\pm\Omega_1,\ldots,\pm\Omega_k, and the corresponding eigenvectors for the following matrix

\begin{bmatrix}+I&0\\0&-I\end{bmatrix} \begin{bmatrix}A&B\\B^*&A^*\end{bmatrix} \begin{bmatrix}u_i\\v_i\end{bmatrix} = \Omega_i \begin{bmatrix}u_i\\v_i\end{bmatrix}, \qquad\text{for } i=1,\ldots,k. \tag*{(65)}\end{bmatrix}

Recall that the eigenvalues of the QRPA matrix SS of Eq. (2) come in pairs (+Ωi,−Ωi)(+\Omega_i,-\Omega_i) with respective eigenvectors

+Ωi:[uivi],+Ωi:[vi∗ui∗].(66)+\Omega_i: \begin{bmatrix}u_i\\v_i\end{bmatrix}, \qquad +\Omega_i: \begin{bmatrix}v_i^*\\u_i^*\end{bmatrix}. \tag*{(66)}

Eigensolvers such as ARPACK typically normalize the eigenvector such that ∥ui∥2+∥vi∥2=1\lVert u_i\rVert^2+\lVert v_i\rVert^2=1. We define the eigenvectors

\begin{bmatrix}x_i\\y_i\end{bmatrix} = \frac{1}{\sqrt{\lVert u_i\rVert^2-\lVert v_i\rVert^2}} \begin{bmatrix}u_i\\v_i\end{bmatrix}, \qquad\text{for } i=1,\ldots,k, \tag*{(67)}\end{bmatrix}

which are normalized with respect to the QRPA metric: ∥xi∥2−∥yi∥2=1\lVert x_i\rVert^2-\lVert y_i\rVert^2=1. Mathematically spurious zero-energy modes are actually non-normalizable [31], because for them ∥ui∥2−∥vi∥2=0\lVert u_i\rVert^2-\lVert v_i\rVert^2=0. In practice, however, spurious modes never have exactly vanishing eigenvalue and one can normalize them.

In the case where the QRPA eigenmodes are non degenerate, two eigenvectors (A.9) for i≠ji \ne j, hence Ωi≠Ωj\Omega_i \ne\Omega_j, are orthogonal with respect to the QRPA metric since it is easy to show that

Ωj[xjyj]†[+I00−I][xjyj]=Ωi[xiyi]†[+I00−I][xjyj].(68)\Omega_j \begin{bmatrix} x_j \\ y_j \end{bmatrix}^{\dagger} \begin{bmatrix} +I & 0 \\ 0 & -I \end{bmatrix} \begin{bmatrix} x_j \\ y_j \end{bmatrix} = \Omega_i \begin{bmatrix} x_i \\ y_i \end{bmatrix}^{\dagger} \begin{bmatrix} +I & 0 \\ 0 & -I \end{bmatrix} \begin{bmatrix} x_j \\ y_j \end{bmatrix}. \tag*{(68)}

Therefore, for any i,j=1,…,ki, j = 1, \ldots, k, we must have

\begin{bmatrix} x_i \\ y_i \end{bmatrix}^{\dagger} \begin{bmatrix} +I & 0 \\ 0 & -I \end{bmatrix} \begin{bmatrix} x_j \\ y_j \end{bmatrix} = \delta_{i,j}, \tag*{(69)}\end{bmatrix}

which means that eigenvectors generated from Eq. (A.9) are indeed the columns Xei,YeiX e_i, Y e_i of the X,YX, Y matrices. In cases of degeneracies, the generalized Gram–Schmidt orthogonalization procedure with QRPA metric is needed in order to orthogonalize the eigenvectors {xi,yi}\{x_i,y_i\} which belong to the degenerate eigenfrequency. It is important to emphasize that the numerical accuracy of the FAM solver, namely the mapping (8), must be high enough to avoid loss of accuracy when computing the normalization factor ∥ui∥2−∥vi∥2\sqrt{\lVert u_i\rVert^2-\lVert v_i\rVert^2} in those cases where the eigenfrequency is very close to 0.

Let us also point out the importance of carefully choosing the initial vector for the iterative algorithm in libraries such as ARPACK. This comes from the fact that typical FAM solvers assume certain symmetries such as, e.g. axialy symmetry, which means that only the excitation operators with certain selection rules are allowed. Typically this involves operators with well defined angular momentum KK and parity π\pi, which actually means that one works in a subspace of the QRPA matrix with well defined KπK^\pi. Therefore we also need to supply the eigensolver with an initial random vector for the mapping (A.6) that also satisfies the same selection rules as the ones of the FAM solver. In other words, the initial random vector must be an element of the appropriate KπK^\pi subspace.

In order to illustrate the method described above on a realistic example, we select the axially deformed (β2≈0.5\beta_2 \approx0.5) isotope 20Ne{}^{20}\mathrm{Ne}, DD-PC1 Lagrangian and use a basis of 12 oscillator shells; see Ref. [34] for a detailed discussion of the multipole response of this isotope. We implemented the spurious mode removal method in the DIRQFAM code [35, 36] using the ARPACK eigensolver, and focus on the isoscalar J=2J = 2, K=1K = 1 response which exhibits a rotational Kπ=1+K^\pi= 1^+ spurious mode.

We generated a random Kπ=1+K^\pi= 1^+ excitation operator as the initial residual vector by ARPACK. This ensures that the Arnoldi algorithm implemented in ARPACK is contained within the Kπ=1+K^\pi= 1^+ subspace of the QRPA matrix. We computed the response function with the KPM method and bounding frequency Ωb=4500\Omega_b = 4500 MeV, a Lorentz kernel with damping parameter λ=9\lambda= 9 and Nit=100000N_{\mathrm{it}} = 100000 FAM iterations. This roughly corresponds to an equivalent smear energy equal to γ=Ωbλ2Nit+1≈0.2\gamma= \Omega_b \frac{\lambda}{2N_{\mathrm{it}}+1} \approx0.2 MeV; see Ref. [28] for details. In Fig. A.6 we show the computed response functions: the black curve shows the response without shifting the spurious mode, i.e. Ω~1=Ω1\widetilde{\Omega}_1 = \Omega_1, the red curve shows the response when the spurious mode is shifted to Ω~1=40\widetilde{\Omega}_1 = 40 MeV, and the blue curve shows the response when the spurious mode is shifted to high-energy region Ω~1=200\widetilde{\Omega}_1 = 200 MeV.

Figure A.6. Response function of 20Ne{}^{20}\mathrm{Ne} for isoscalar quadrupole J=2J = 2, K=1K = 1 excitation. The black curve shows the response without spurious mode removal, the red curve shows the response when the spurious mode is shifted to Ω~1=40\widetilde{\Omega}_1 = 40 MeV, and the blue curve shows the response when the spurious mode is shifted to Ω~1=200\widetilde{\Omega}_1 = 200 MeV.

MeV. Note that even though we could shift the energy of the spurious method to any arbitrary value Ω~i\widetilde{\Omega}_i, it should still satisfy Ω~i<Ωb\widetilde{\Omega}_i < \Omega_b for the KPM method to work.

References

References

  1. [1]H. A. Bethe, An Attempt to Calculate the Number of Energy Levels of a Heavy Nucleus, Phys. Rev. 50 (4) (1936) 332. doi:10.1103/PhysRev.50.332.DOI
  2. [2]C. Bloch, Theory of Nuclear Level Density, Phys. Rev. 93 (5) (1954) 1094. doi:10.1103/PhysRev.93.1094.DOI
  3. [3]T. Ericson, The statistical model and nuclear level densities, Adv. in Phys. 9 (36) (1960) 425. doi:10.1080/00018736000101239.DOI
  4. [4]A. Bohr, B. Mottelson, Nuclear Structure, Vol. II, Nuclear Deformations, World Scientific, 1998. doi:10.1142/3530.
  5. [5]A. J. Koning, S. Hilaire, M. C. Duijvestijn, TALYS-1.0, in: Proc. Int. Conf. Nucl. Data Sci. Technol., EDP Sciences, Nice, France, 2008, p. 211. doi:10.1051/ndata:07767.DOI
  6. [6]M. Herman, R. Capote, B. V. Carlson, P. Obložinský, M. Sin, A. Trkov, H. Wienke, V. Zerkin, EMPIRE: Nuclear Reaction Model Code System for Data Evaluation, Nucl. Data Sheets 108 (12) (2007) 2655. doi:10.1016/j.nds.2007.11.003.DOI
  7. [7]H. Nakada, Y. Alhassid, Total and Parity-Projected Level Densities of Iron-Region Nuclei in the Auxiliary Fields Monte Carlo Shell Model, Phys. Rev. Lett. 79 (16) (1997) 2939. doi:10.1103/PhysRevLett.79.2939.
  8. [8]M. Horoi, J. Kaiser, V. Zelensikvky, Spin- and parity-dependent nuclear level densities and the exponential convergence method, Phys. Rev. C 67 (5) (2003) 054309. doi:10.1103/PhysRevC.67.054309.DOI
  9. [9]Y. Alhassid, G. F. Bertsch, L. Fang, Nuclear level statistics: Extending shell model theory to higher temperatures, Phys. Rev. C 68 (4) (2003) 044322. doi:10.1103/PhysRevC.68.044322.DOI
  10. [10]H. Nakada, Y. Alhassid, Isospin-projected nuclear level densities by the shell model Monte Carlo method, Phys. Rev. C 78 (5) (2008) 051304. doi:10.1103/PhysRevC.78.051304.DOI
  11. [11]R. A. Sen’kov, M. Horoi, High-performance algorithm to calculate spin- and parity-dependent nuclear level densities, Phys. Rev. C 82 (2) (2010) 024304. doi:10.1103/PhysRevC.82.024304.
  12. [12]R. A. Sen’kov, M. Horoi, V. G. Zelvensky, High-performance algorithm for calculating non-spurious spin- and parity-dependent nuclear level densities, Phys. Lett. B 702 (5) (2011) 413. doi:10.1016/j.physletb.2011.07.004.
  13. [13]Y. Alhassid, M. Bonett-Matiz, S. Liu, H. Nakada, Direct microscopic calculation of nuclear level densities in the shell model Monte Carlo approach, Phys. Rev. C 92 (2) (2015) 024307. doi:10.1103/PhysRevC.92.024307.DOI
  14. [14]S. Hilaire, J.-P. Delaroche, M. Girod, Combinatorial nuclear level densities based on the Gogny nucleon-nucleon effective interaction, Eur. Phys. J. A 12 (2) (2001) 169. doi:10.1007/s100500170025.DOI
  15. [15]N. Schunck, Energy Density Functional Methods for Atomic Nuclei., IOP Expanding Physics, IOP Publishing, Bristol, UK, 2019. URL https://doi.org/10.1088/2053-2563/aae0ed
  16. [16]J. F. Berger, M. Martinot, Shell effects on state densities with given numbers of excited protons and neutrons, Nucl. Phys. A 226 (3) (1974) 391. doi:10.1016/0375-9474(74)90491-6.
  17. [17]S. Hilaire, S. Goriely, Global microscopic nuclear level densities within the HFB plus combinatorial method for practical applications, Nucl. Phys. A 779 (2006) 63. doi:10.1016/j.nuclphysa.2006.08.014.DOI
  18. [18]S. Goriely, S. Hilaire, A. J. Koning, Improved microscopic nuclear level densities within the Hartree-Fock-Bogoliubov plus combinatorial method, Phys. Rev. C 78 (6) (2008) 064309. doi:10.1103/PhysRevC.78.064307.DOI
  19. [19]S. Hilaire, M. Girod, S. Goriely, A. J. Koning, Temperature-dependent combinatorial level densities with the D1M Gogny force, Phys. Rev. C 86 (6) (2012) 064317. doi:10.1103/PhysRevC.86.064317.DOI
  20. [20]S. Hilaire, S. Goriely, S. Péru, G. Gosselin, A new approach to nuclear level densities: The QRPA plus boson expansion, Phys. Lett. B 843 (2023) 137989. doi:10.1016/j.physletb.2023.137989.DOI
  21. [21]J. Terasaki, J. Engel, Self-consistent Skyrme quasiparticle random-phase approximation for use in axially symmetric nuclei of arbitrary mass, Phys. Rev. C 82 (3) (2010) 034326. doi:10.1103/PhysRevC.82.034326.DOI
  22. [22]M. Martini, S. Péru, S. Hilaire, S. Goriely, F. Lechaftois, Large-scale deformed quasiparticle random-phase approximation calculations of the $\ensuremath{\gamma}$-ray strength function using the Gogny force, Phys. Rev. C 94 (1) (2016) 014304. doi:10.1103/PhysRevC.94.014304.DOI
  23. [23]T. Nakatsukasa, T. Inakura, K. Yabana, Finite amplitude method for the solution of the random-phase approximation, Phys. Rev. C 76 (2) (2007) 024318. doi:10.1103/PhysRevC.76.024318.DOI
  24. [24]P. Avogadro, T. Nakatsukasa, Finite amplitude method for the quasiparticle random-phase approximation, Phys. Rev. C 84 (1) (2011) 014314. doi:10.1103/PhysRevC.84.014314.DOI
  25. [25]P. Ring, P. Schuck, The Nuclear Many-Body Problem, Texts and Monographs in Physics, Springer, 2004.
  26. [26]E. Di Napoli, E. Polizzi, Y. Saad, Efficient estimation of eigenvalue counts in an interval, Numerical Linear Algebra with Applications 23 (4) (2016) 674–692. doi:https://doi.org/10.1002/nla.2048.
  27. [27]L. Lin, Y. Saad, C. Yang, Approximating spectral densities of large matrices, SIAM Rev. 58 (1) (2016) 34–65. doi:10.1137/130934283.DOI
  28. [28]A. Bjelčić, T. Nikšić, Z. Drmač, Chebyshev kernel polynomial method for efficient calculation of the quasiparticle random phase approximation response function, Comput. Phys. Commun. 280 (2022) 108477. doi:10.1016/j.cpc.2022.108477.DOI
  29. [29]N. Hinohara, M. Kortelainen, W. Nazarewicz, Low-energy collective modes of deformed superfluid nuclei within the finite-amplitude method, Phys. Rev. C 87 (6) (2013) 064309. doi:10.1103/PhysRevC.87.064309.DOI
  30. [30]J. Toivanen, B. G. Carlsson, J. Dobaczewski, K. Mizuyama, R. R. Rodríguez-Guzmán, P. Toivanen, P. Veselý, Linear response strength functions with iterative arnoldi diagonalization, Phys. Rev. C 81 (2010) 034312. doi:10.1103/PhysRevC.81.034312. URL https://link.aps.org/doi/10.1103/PhysRevC.81.034312DOI
  31. [31]N. Hinohara, Collective inertia of the Nambu-Goldstone mode from linear response theory, Phys. Rev. C 92 (3) (2015) 034321. doi:10.1103/PhysRevC.92.034321.DOI
  32. [32]G. Colò, L. Cao, N. Van Giai, L. Capelli, Self-consistent RPA calculations with Skyrme-type interactions: The skyrme_rpa program, Comput. Phys. Commun. 184 (1) (2013) 142. doi:10.1016/j.cpc.2012.07.016.DOI
  33. [33]R. B. Lehoucq, D. C. Sorensen, C. Yang, ARPACK Users’ Guide, Society for Industrial and Applied Mathematics, 1998. doi:10.1137/1.9780898719628. URL https://epubs.siam.org/doi/abs/10.1137/1.9780898719628
  34. [34]F. Mercier, A. Bjelčić, T. Nikšić, J.-P. Ebran, E. Khan, D. Vretenar, Low-energy cluster modes in $N=Z$ nuclei, Phys. Rev. C 103 (2) (2021) 024303. doi:10.1103/PhysRevC.103.024303.DOI
  35. [35]A. Bjelčić, T. Nikšić, Implementation of the quasiparticle finite amplitude method within the relativistic self-consistent mean-field framework: The program DIRQFAM, Comput. Phys. Commun. 253 (2020) 107184. doi:10.1016/j.cpc.2020.107184.DOI
  36. [36]A. Bjelčić, T. Nikšić, Implementation of the quasiparticle finite amplitude method within the relativistic self-consistent mean-field framework (II): The program DIRQFAM v2.0.0, Comput. Phys. Commun. 287 (2023) 108689. doi:10.1016/j.cpc.2023.108689.DOI

Paper details

vlnc:W_6m3xu2uk
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

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