betaFeedback

WorksPhysics and Astronomy

Closest Wannier functions to a given set of localized orbitals

Abstract

A non-iterative method is presented to calculate the closest Wannier functions (CWFs) to a given set of localized guiding functions, such as atomic orbitals, hybrid atomic orbitals, and molecular orbitals, based on minimization of a distance measure function. It is shown that the minimization is directly achieved by a polar decomposition of a projection matrix via singular value decomposition, making iterative calculations and complications arising from the choice of the gauge irrelevant. The disentanglement of bands is inherently addressed by introducing a smoothly varying window function and a greater number of Bloch functions, even for isolated bands. In addition to atomic and hybrid atomic orbitals, we introduce embedded molecular orbitals in molecules and bulks as the guiding functions, and demonstrate that the Wannier interpolated bands accurately reproduce the targeted conventional bands of a wide variety of systems including Si, Cu, the TTF-TCNQ molecular crystal, and a topological insulator of Bi2_2Se3_3. We further show the usefulness of the proposed method in calculating effective atomic charges. These numerical results not only establish our proposed method as an efficient alternative for calculating WFs, but also suggest that the concept of CWFs can serve as a foundation for developing novel methods to analyze electronic structures and calculate physical properties.

Introduction

The Wannier functions (WFs) [1, 2, 3] play a central role in analyzing electronic structures of real materials and advancing electronic structure methods alongside density functional theory (DFT) [4] and the other electronic structure theories [5, 6]. A widely adopted method for calculating WFs involves maximizing their localization, which can be reformulated as minimizing the spread function characterizing the variance of WFs in real space [7]. The concept of maximal localization leads to an elegant formulation, resulting in maximally localized Wannier functions (MLWFs), and the compact representation of the Hamiltonian that follows allows us to analyze the electronic structures of real materials using a localized orbital picture. However, it is also recognized that the minimization of the spread function can often encounter local minima, particularly in large-scale systems having complicated electronic structures [3]. To overcome this challenge, methods aimed at automated high-throughput Wannierisation have been developed [8, 9, 10]. In addition, the other methods for generating (nonorthogonal) WFs have been proposed based on projection methods [11, 12, 13], which produce compact atomic like orbitals. The orthogonalization of such nonorthogonal orbitals obtained by the projection can be achieved by the Löwdin orthogonalization procedure [14, 15]. These orthogonalized orbitals are utilized as an initial guess of WFs in the minimization of the spread function to obtain MLWFs [7, 16]. Apart from the role of initial guess, it is worth noting that the Löwdin orthogonalized orbitals possess an important variational property, which in particular minimizes the sum of squared distances between the orthogonal and original nonorthogonal orbitals in the Hilbert space [17]. As long as the original nonorthogonal orbitals are localized, the localization of the Löwdin orthogonalized orbitals in real space is guaranteed in a sense of minimization of the distance.

In quantum chemistry, this property is leveraged to develop methods for generating localized orthogonal orbitals. These methods, which include the natural atomic and bond orbital method [18] and the intrinsic atomic orbital method [19], involve obtaining orthogonalized orbitals through Löwdin type orthogonalization, giving heavy weight to occupied states via the density matrix. However, in solid state physics the variational property inherent in the Löwdin orthogonalization has not been fully explored yet. A reason for this is that the Löwdin orthogonalization requires calculation of S−1/2S^{-1/2} for the overlap matrix SS, which causes numerical instability for nearly ill-conditioned matrices. Due to the difficulty, advancements in methodologies along these lines appear to be slow. Our study aims to develop an efficient and robust method to generate the closest Wannier functions (CWFs) to a given set of localized guiding functions. We will demonstrate that exploiting the variational property leads to a versatile method that eliminates the need for iterative optimization and avoids complications related to gauge choice.

The structure of this paper is as follows: In Section II, we present the theory of calculating CWFs. Section III provides a detailed discussion on the implementation of the method. A series of benchmark calculations are presented in Section IV. Finally, in Section V, we summarize the theory of CWFs and suggest a potential role for CWFs as a foundation for developing novel methods.

Theory

Let us start by introducing Bloch functions {ϕ}\{\phi\}, which can be obtained by solving the Kohn-Sham (KS) equation [20] within the density functional theory (DFT) [4], normalized as

⟨ϕk1μ∣ϕk2ν⟩=NBCδk1k2δμν,(1)\langle\phi_{k_1\mu} \mid\phi_{k_2\nu} \rangle= N_{\mathrm{BC}}\delta_{k_1 k_2}\delta_{\mu\nu}, \tag*{(1)}

where k\mathbf{k} and μ\mu are the k\mathbf{k}-vector and the band index, respectively, and NBCN_{\mathrm{BC}} is the number of primitive cells in the Born-von Karman (BvK) boundary condition. We consider a projection of a localized guiding function QQ onto the Bloch functions {ϕ}\{\phi\} weighted with a window function w(ε)w(\varepsilon) as

∣LRp⟩=1NBC∑k,μe−ik⋅R∣ϕkμ⟩akμ,p,(2)|L_{\mathbf{R}p}\rangle= \frac{1}{N_{\mathrm{BC}}}\sum_{\mathbf{k},\mu}e^{-i\mathbf{k}\cdot\mathbf{R}}|\phi_{\mathbf{k}\mu}\rangle a_{\mathbf{k}\mu,p}, \tag*{(2)}

with

akμ,p=w(εkμ)⟨ϕkμ∣Q0p⟩,(3)a_{\mathbf{k}\mu,p}=w(\varepsilon_{\mathbf{k}\mu})\langle\phi_{\mathbf{k}\mu}|Q_{0p}\rangle, \tag*{(3)}

where R\mathbf{R} and pp are a translational lattice vector and the index of localized orbital, respectively, εkμ\varepsilon_{\mathbf{k}\mu} is the eigenvalue of the KS equation, and Q0p≡QRpQ_{0p}\equiv Q_{\mathbf{R}p} (R=0\mathbf{R}=0). In the summation of Eq. (2), the numbers of k\mathbf{k}-point and the Bloch functions to be included are NBCN_{\mathrm{BC}} and NbandN_{\mathrm{band}} per k\mathbf{k}-point, respectively. Throughout the paper we do not consider the spin dependency on the formulation for sake of simplicity, but the generalization is straightforward. The localized functions {QRp}\{Q_{\mathbf{R}p}\}, such as atomic orbitals, hybrid orbitals, or molecular orbitals (MOs), are assumed to localize in the unit cell specified by R\mathbf{R}. When the window function w(ε)w(\varepsilon) is taken to be unity, Eq. (2) is nothing but an expansion of QRpQ_{\mathbf{R}p} with Bloch functions from the identity relation

1NBC∑kμ∣ϕkμ⟩⟨ϕkμ∣=I^.\frac{1}{N_{\mathrm{BC}}}\sum_{\mathbf{k}\mu}|\phi_{\mathbf{k}\mu}\rangle\langle\phi_{\mathbf{k}\mu}|=\hat{I}.

To introduce an expansion of QRpQ_{\mathbf{R}p} by the Bloch functions {ϕkμ}\{\phi_{\mathbf{k}\mu}\} in a subspace, we use the following window function w(ε)w(\varepsilon):

w(ε)=1−exp⁡(x0+x1)(1+exp⁡(x0))(1+exp⁡(x1))+δ(4)w(\varepsilon)=\frac{1-\exp(x_0+x_1)}{(1+\exp(x_0))(1+\exp(x_1))}+\delta \tag*{(4)}

with the definition of x0≡β(ε0−ε)x_0\equiv\beta(\varepsilon_0-\varepsilon) and x1≡β(ε−ε1)x_1\equiv\beta(\varepsilon-\varepsilon_1), where β=1kBT\beta=\frac{1}{k_{\mathrm{B}}T} with the Boltzmann constant of kBk_{\mathrm{B}} and a temperature of TT. The window function of Eq. (4) is obtained by subtracting 11 from the sum of two functions of 1/(1+exp⁡(x))1/(1+\exp(x)). The last term δ\delta is a small constant, e.g., 10−1210^{-12}, which is introduced to avoid the ill-conditioning of the matrix consisting of akμ,pa_{\mathbf{k}\mu,p} by Eq. (3) as discussed later on. As shown in Figure 1, the parameters of ε0\varepsilon_0 and ε1\varepsilon_1 (ε0<ε1\varepsilon_0<\varepsilon_1) determine the range of energy where the Bloch states are included in the expansion of Eq. (2) with a large weight, and the temperature TT gives the degree of smearing around ε0\varepsilon_0 and ε1\varepsilon_1 in w(ε)w(\varepsilon). To control the degree of smearing, we introduce the temperature of TT so that one can intuitively understand the degree of smearing. It is noted that the choice of the window function by Eq. (4) is not unique, and the other choices should give almost equivalent results as long as they are chosen properly. The function LRpL_{\mathbf{R}p} defined by Eq. (2) must be similar to the localized function QRpQ_{\mathbf{R}p}, but they are non-orthogonal to each other in general. We now consider generating a set of closest Wannier functions (CWFs) {WRp}\{W_{\mathbf{R}p}\} to a set of the localized functions {LRp}\{L_{\mathbf{R}p}\} in a sense that the sum of the squared distance between LL and WW in the Hilbert

Figure 1. Window functions of Eq. (4) with kBT=0.1k_{\mathrm{B}}T=0.1, 1.01.0, and 3.03.0 eV. ε0\varepsilon_0 and ε1\varepsilon_1 were set to be −5.0-5.0 and 5.05.0 eV, respectively, and δ=10−12\delta=10^{-12} was used.

space is minimized. As well as Eq. (2), noting that the CWFs can be expressed [3] as

∣WRp⟩=1NBC∑k,μe−ik⋅R∣ϕkμ⟩bkμ,p,(5)|W_{\mathbf{R}p}\rangle=\frac{1}{N_{\mathrm{BC}}}\sum_{\mathbf{k},\mu}e^{-i\mathbf{k}\cdot\mathbf{R}}|\phi_{\mathbf{k}\mu}\rangle b_{\mathbf{k}\mu,p}, \tag*{(5)}

and defining the residual function RR:

∣RRp⟩=∣LRp⟩−∣WRp⟩,(6)|R_{\mathbf{R}p}\rangle=|L_{\mathbf{R}p}\rangle-|W_{\mathbf{R}p}\rangle, \tag*{(6)}

the distance measure (DM) function FF we minimize is given by

F[B]=∑p⟨R0p∣R0p⟩=1NBC∑kX[B,k],(7)\begin{aligned} F[B]=\sum_p\langle R_{0p}|R_{0p}\rangle\\ &=\frac{1}{N_{\mathrm{BC}}}\sum_{\mathbf{k}}X[B,\mathbf{k}], \tag*{(7)} \end{aligned}

with

X[B,k]=tr⁡[(A†(k)−B†(k))(A(k)−B(k))],(8)X[B,\mathbf{k}]=\operatorname{tr}\left[(A^{\dagger}(\mathbf{k})-B^{\dagger}(\mathbf{k}))(A(\mathbf{k})-B(\mathbf{k}))\right], \tag*{(8)}

where the elements of the matrices A(k)A(\mathbf{k}) and B(k)B(\mathbf{k}) are given by akμ,pa_{\mathbf{k}\mu,p} and bkμ,pb_{\mathbf{k}\mu,p}, respectively. The second line of Eq. (7) is obtained by considering the orthonormality of Eq. (1), and X[B,k]X[B,\mathbf{k}] of Eq. (8) is regarded as the squared Frobenius norm of (A(k)−B(k))(A(\mathbf{k})-B(\mathbf{k})) [21]. The same Bloch functions as in Eq. (2) are included in the summation of Eq. (5). The number of CWFs to be generated is NCWFN_{\mathrm{CWF}} per unit cell, being equivalent to the number of the guiding functions per cell, and NCWFN_{\mathrm{CWF}} should be smaller than or equal to NbandN_{\mathrm{band}} to guarantee the linear independency of the subspace spanned by the CWFs, resulting in size of Nband×NCWFN_{\mathrm{band}}\times N_{\mathrm{CWF}} for the matrices A(k)A(\mathbf{k}) and B(k)B(\mathbf{k}). Assuming B†(k)B(k)=IB^{\dagger}(\mathbf{k})B(\mathbf{k})=I, where II is of size NCWF×NCWFN_{\mathrm{CWF}}\times N_{\mathrm{CWF}}, the CWFs are ensured to form a set of orthonormal functions. Under this constraint, we consider the optimization of the DM function FF. Furthermore, the matrix B(k)B(k) is regarded as a part of a unitary matrix of size Nband×NbandN_{\mathrm{band}} \times N_{\mathrm{band}}, and will be referred to as a partial unitary matrix in subsequent discussions.

The minimum of the DM function of Eq. (7) is obtained by choosing B(k)B(k) as U(k)U(k) which is calculated by the polar decomposition of A(k)=U(k)P(k)A(k) = U(k)P(k), where U(k)U(k) and P(k)P(k) are unitary and hermitian, respectively [22]. Let us prove the statement below. The polar decomposition of A(k)A(k) is obtained via the singular value decomposition (SVD) of A(k)A(k) as

A=WΣV†=WV†VΣV†=UP,(9)A = W\Sigma V^\dagger= WV^\dagger V\Sigma V^\dagger= UP, \tag*{(9)}

where we dropped the dependency on kk for simplicity of the notation, and hereafter we will denote the dependency if necessary. WW and VV are the left and right singular matrices in size of Nband×NCWFN_{\mathrm{band}} \times N_{\mathrm{CWF}} and NCWF×NCWFN_{\mathrm{CWF}} \times N_{\mathrm{CWF}}, respectively, and Σ\Sigma is the singular value diagonal matrix in size of NCWF×NCWFN_{\mathrm{CWF}} \times N_{\mathrm{CWF}}. Note that U≡WV†U \equiv WV^\dagger and P≡VΣV†P \equiv V\Sigma V^\dagger. It is worth mentioning that WW and UU are partial unitary matrices, and hold W†W=IW^\dagger W = I (WW†≠IWW^\dagger\ne I) and U†U=IU^\dagger U = I (UU†≠IUU^\dagger\ne I), and that VV is a full unitary matrix holding V†V=IV^\dagger V = I (VV†=IVV^\dagger= I). We evaluate XX of Eq. (8) for each kk with both the matrix UU obtained by the polar decomposition and an arbitrary partial unitary matrix BB, and calculate the difference as

X[U]−X[B]=2tr⁡[12(A†B+B†A)−P],=2tr⁡[ΣD−Σ],=2∑nσn(dnn−1)(10)\begin{aligned} X[U] - X[B] &= 2\operatorname{tr}\left[\frac{1}{2}\left(A^\dagger B + B^\dagger A\right) - P\right],\\ &= 2\operatorname{tr}\left[\Sigma D - \Sigma\right],\\ &= 2\sum_n \sigma_n\left(d_{nn} - 1\right) \tag*{(10)} \end{aligned}

with

D=12(V†U†BV+V†B†UV).(11)D = \frac{1}{2}\left(V^\dagger U^\dagger BV + V^\dagger B^\dagger UV\right). \tag*{(11)}

The second line of Eq. (10) is derived by using the polar decomposition of AA and P=VΣV†P = V\Sigma V^\dagger, and σn\sigma_n and dnnd_{nn} in the third line are diagonal elements of Σ\Sigma and DD, respectively. The upper bound of the diagonal elements of the hermitian matrix DD is found to be unity, since the matrix DD consists of the sum of the product of partial unitary matrices of V†U†V^\dagger U^\dagger (V†B†V^\dagger B^\dagger) and BVBV (UVUV). Another proof for the upper bound is also given based on the Cauchy-Schwarz inequality in the appendix. By considering the upper bound of the diagonal elements of DD and 0≤σn0 \le\sigma_n, the third line of Eq. (10) leads to the following inequality:

X[U]≤X[B].(12)X[U] \le X[B]. \tag*{(12)}

The equality of Eq. (12) holds if B=UB = U. If some of σn\sigma_n is zero, the corresponding dnnd_{nn} in Eq. (10) can be arbitrarily chosen. So, we see from Eq. (11) that UU is not uniquely determined. If all the singular values of σn\sigma_n are positive, all the dnnd_{nn}s should be unity when X[U]=X[B]X[U] = X[B]. The uniqueness of the solution can be confirmed as follows: Since the matrix DD consists of the products of the partial unitary matrices as discussed above, the case that all the diagonal elements dnnd_{nn} are unity happens when

U†B+B†U=2I,(13)U^\dagger B + B^\dagger U = 2I, \tag*{(13)}

which is derived by equating D=ID = I and multiplying VV and V†V^\dagger from the left and right of the equation, respectively. By noting again that BB and UU in Eq. (13) are partial unitary matrices, it is found that Eq. (13) holds if and only if B=UB = U. Thus, we have proven the statement that the minimum of the DM function of Eq. (7) is uniquely determined when B=UB = U as long as all the singular values of σn\sigma_n are positive, since F[B]F[B] of Eq. (7) consists of the sum of X[B,k]X[B,k] over kk. The uniqueness of UU itself is related to the defect of the SVD for the matrix AA of Eq. (3). If AA has NCWFN_{\mathrm{CWF}} positive singular values which are non-degenerate, the SVD is uniquely determined except for the non-unique phases in WW and VV. Even if there are degenerate singular values, the matrix UU is invariant, since the freedom of the unitary transformation KK for the degenerate left and right singular vectors is canceled out as U=WKK†V†=WV†U = WK K^\dagger V^\dagger= WV^\dagger with KK†=IKK^\dagger= I. Further, noting that the non-unique phases in WW and VV are canceled out when the matrix UU is computed as WV†WV^\dagger, we conclude that UU is uniquely determined if AA has the NCWFN_{\mathrm{CWF}} positive singular values. The violation from the positive definiteness of the singular values can be avoided by the small constant of δ\delta in Eq. (4). On the other hand, if δ\delta in Eq. (4) is set to be 0 and a narrow window, including fewer than NCWFN_{\mathrm{CWF}} eigenstates for a given kk, is used with a small kBTk_B T, the matrix AA has NPSVN_{\mathrm{PSV}} positive singular values, where NPSV<NCWFN_{\mathrm{PSV}} < N_{\mathrm{CWF}}. In this case, a subspace of the (NCWF−NPSV)(N_{\mathrm{CWF}} - N_{\mathrm{PSV}}) dimension is arbitrary chosen to form WW and VV of the dimension of NCWFN_{\mathrm{CWF}} in addition to the subspace of the NPSVN_{\mathrm{PSV}} dimension, which breaks the symmetry of CWFs. The situation can be avoided by introducing a small constant δ\delta in the window function of Eq. (4). The simple treatment restores the symmetry of CWFs, allowing us to calculate symmetry preserving CWFs for a wide variety of choices in the energy range of the window function. It is also noted that the non-unique phase of the Bloch functions is canceled out via the polar decomposition of AA and Eq. (5). Thus, we see that the proposed method is free from complications arising from the choice of gauge.

Using Eqs. (8) and (9), the minimum of the DM function FF at B=UB = U is given by

F[U]=1NBC∑k,p(σkp−1)2,(14)F[U] = \frac{1}{N_{\mathrm{BC}}}\sum_{k,p}\left(\sigma_{kp} - 1\right)^2, \tag*{(14)}

where σkp\sigma_{kp} is a singular value of the matrix A(k)A(k). From Eq. (14), we find that the mean squared distance between LL and WW is related to the deviation of σ\sigma from unity.

The proposed method based on the polar decomposition of AA is closely related to the Löwdin orthogonalization [14] as shown below. The Fourier transform of the overlap integrals for {L}\{L\} is given by

∑Reik⋅R⟨L0p∣LRq⟩=∑μakμ,p∗akμ,q,(15)\sum_{\mathbf{R}}e^{i\mathbf{k}\cdot\mathbf{R}}\langle L_{0p}|L_{Rq}\rangle=\sum_{\mu}a^{*}_{\mathbf{k}\mu,p}a_{\mathbf{k}\mu,q}, \tag*{(15)}

which is obtained by noting that 1NBC∑Rei(k−k′)⋅R=δkk′\frac{1}{N_{\mathrm{BC}}}\sum_{\mathbf{R}}e^{i(\mathbf{k}-\mathbf{k}')\cdot\mathbf{R}}=\delta_{\mathbf{k}\mathbf{k}'}. By writing Eq. (15) as S(k)=A†(k)A(k)S(\mathbf{k})=A^{\dagger}(\mathbf{k})A(\mathbf{k}) in a matrix form, and using Eq. (9), we have the following relation:

S(k)=A†(k)A(k)=V†(k)Σ2(k)V(k).(16)S(\mathbf{k})=A^{\dagger}(\mathbf{k})A(\mathbf{k})=V^{\dagger}(\mathbf{k})\Sigma^{2}(\mathbf{k})V(\mathbf{k}). \tag*{(16)}

Comparing Eq. (16) with P(k)=V(k)Σ(k)V†(k)P(\mathbf{k})=V(\mathbf{k})\Sigma(\mathbf{k})V^{\dagger}(\mathbf{k}), one obtains a relation P(k)=S1/2(k)P(\mathbf{k})=S^{1/2}(\mathbf{k}). If P(k)P(\mathbf{k}) is invertible, the matrix U(k)U(\mathbf{k}) is given with Eq. (9) by

U(k)=A(k)S−1/2(k).(17)U(\mathbf{k})=A(\mathbf{k})S^{-1/2}(\mathbf{k}). \tag*{(17)}

The matrix UU obtained by Eq. (17) is exactly equivalent to that by the Löwdin orthogonalization [14], and the closest property of the Löwdin orthogonalized orbitals to a given set of orbitals was proven in Ref. [17]. On the other hand, the proposed method does not require the calculation of S−1/2S^{-1/2}, and can be applied in a numerical stable manner even to the case that the matrix AA is nearly ill-conditioned. The definition of AA by Eq. (3) with the window function allows us to calculate CWFs in various respects, e.g., heavily weighting to occupied states and disentangling of localized states embedded in the wider bands as demonstrated in § 4. In this sense, the method we propose can be regarded as a generalization of the Löwdin orthogonalization.

The disentanglement of bands can be easily performed by properly selecting the window function of Eq. (3). Since the Nband(≥NCWF)N_{\mathrm{band}}(\ge N_{\mathrm{CWF}}) Bloch functions per kk-point are included in Eqs. (2) and (5), it should be emphasized that the proposed method always disentangles NbandN_{\mathrm{band}} states to generate NCWFN_{\mathrm{CWF}} Wannier functions. A couple of examples will be shown in § 4, including valence and low-lying conduction bands for the diamond Si and narrow 3d3d-bands embedded in the wider 4s4s-band for the face centered cubic (FCC) Cu.

Once the kk-dependent U(k)U(\mathbf{k}) are obtained by the polar decomposition, tight-binding (TB) parameters are calculated as expectation values of the KS Hamiltonian H^KS\hat{H}_{\mathrm{KS}} by summing contributions over kk and μ\mu as

t0p,Rq=⟨W0p∣H^KS∣WRq⟩=∑k,μεkμukμ,p∗ukμ,qe−ik⋅R,(18)t_{0p,Rq}=\langle W_{0p}|\hat{H}_{\mathrm{KS}}|W_{Rq}\rangle =\sum_{\mathbf{k},\mu}\varepsilon_{\mathbf{k}\mu}u^{*}_{\mathbf{k}\mu,p}u_{\mathbf{k}\mu,q}e^{-i\mathbf{k}\cdot\mathbf{R}}, \tag*{(18)}

where ukμ,qu_{\mathbf{k}\mu,q} is the matrix element of U(k)U(\mathbf{k}). Like with MLWFs, the TB parameters can be used for the Wannier interpolation in calculations of band structures and physical quantities.

The computational procedure to generate the CWFs is summarized as follows:

  1. Determining ε0\varepsilon_{0}, ε1\varepsilon_{1}, and kBTk_{\mathrm{B}}T in Eq. (4) by checking the band dispersion of a system of interest. The parameters should be chosen properly so that the targeted eigenstates can be included.

  1. Choosing a set of localized orbitals {Q}\{Q\}. Atomic orbitals, hybrid orbitals, and MOs might be possible choices. Depending on the symmetry of the targeted eigenstates, a proper set of localized orbitals needs to be chosen, e.g., dd-orbitals need to be employed in the generation of CWFs for the dd-bands in FCC Cu as shown later on.

  1. Calculation of A(k)A(\mathbf{k}) by Eq. (3). The calculation of the overlap ⟨ϕkμ∣Qop⟩\langle\phi_{\mathbf{k}\mu}|Q_{\mathrm{op}}\rangle depends on the implementation of the KS method.

  1. Performing the SVD of A(k)A(\mathbf{k}). A proper numerical library might be used.

  1. Calculation of U(k)U(\mathbf{k}). Using the left and right singular matrices, W(k)W(\mathbf{k}) and V(k)V(\mathbf{k}), of A(k)A(\mathbf{k}), U(k)U(\mathbf{k}) is calculated as W(k)V†(k)W(\mathbf{k})V^{\dagger}(\mathbf{k}).

  1. Calculations of the CWFs and TB parameters. The CWFs can be calculated using Eq. (5) with B=UB=U obtained by the step 5. The calculation is expressed by the sum of the matrix-matrix product over kk. Also, the TB parameters are obtained by Eq. (18).

It should be stressed that the minimization of FF can be efficiently performed in a non-iterative manner by the steps 1 to 6 using the polar decomposition via the SVD. The computational cost of each step is estimated to be O(NBCNbandNCWFNbasis)O(N_{\mathrm{BC}}N_{\mathrm{band}}N_{\mathrm{CWF}}N_{\mathrm{basis}}), O(NBCNband3)O(N_{\mathrm{BC}}N_{\mathrm{band}}^{3}), O(NBCNbandNCWF)O(N_{\mathrm{BC}}N_{\mathrm{band}}N_{\mathrm{CWF}}), or O(NBCNbandNCWFNbasis)O(N_{\mathrm{BC}}N_{\mathrm{band}}N_{\mathrm{CWF}}N_{\mathrm{basis}}) for the step 3, 4, 5, or 6, respectively, where NbasisN_{\mathrm{basis}} is the number of basis functions to expand the KS orbitals. If the localized basis functions are used, the computational cost of the step 3 becomes O(NBCNbandNCWF)O(N_{\mathrm{BC}}N_{\mathrm{band}}N_{\mathrm{CWF}}).

For isolated systems, the Brillouin zone sampling is limited to only the Γ\Gamma point. No modification of the proposed method is needed to calculate the CWFs.

The theoretical framework we have discussed so far is general for any implementation of the DFT-KS method. However, the choice of the localized functions {Q}\{Q\} in Eq. (3) may depend on the implementation. So, we will discuss how the localized functions {Q}\{Q\} can be properly generated in § 3.

Implementation

General

We have implemented the proposed method into the OpenMX DFT software package [23, 24, 25] which is based on norm-conserving pseudopotentials (PPs) [26, 27] and optimized pseudo-atomic orbitals (PAOs) [28, 29] as basis set. All the benchmark calculations were performed with a computational condition of a production level. The basis functions used are listed in Table I. In the abbreviation of basis functions such as H7.0-s2p2d1, H stands for the atomic symbol, 7.0 the cutoff radius (Bohr) in the generation by the confinement scheme [28, 29], and s2p2d1 means the employment of two, two, and one optimized radial functions for the ss-, pp-, and dd-orbitals, respectively. The radial functions were optimized by a variational optimization method [28]. These basis functions we used can be regarded as at least double zeta plus double polarization basis sets if we follow the terminology of Gaussian or Slater-type basis functions. Valence states included in the PPs are listed in Table I. All the PPs and PAOs we used in the study were taken from the database (2019) in the OpenMX website [23], which were benchmarked by the delta gauge method [30]. Real space grid techniques were used for the numerical integrations and the solution of the Poisson equation using fast Fourier transform (FFT) with the energy cutoff of 250 to 500 Ryd [24]. We used a generalized gradient approximation (GGA) proposed by Perdew, Burke, and Ernzerhof to the exchange-correlation functional [31]. An electronic temperature of 300 K was used to count the number of electrons by the Fermi-Dirac function for all the systems we considered. For all the calculations, δ=10−12\delta= 10^{-12} was used in the window function of Eq. (4), and NbandN_{\mathrm{band}} was set to be equilavent to the number of basis functions in both the summation of Eqs. (2) and (5).

ElementBasis functionsValence states in PP
HH7.0-s2p2d11s1s
CC6.0-s3p2d22s2s, 2p2p
NN6.0-s3p2d22s2s, 2p2p
Na*Na9.0-s3p3d2f22s2s, 2p2p, 3s3s
SiSi7.0-s2p2d2f13s3s, 3p3p
Cl*Cl9.0-s3p3d2f23s3s, 3p3p
SS7.0-s3p2d2f13s3s, 3p3p
CuCu6.0-s3p3d3f13s3s, 3p3p, 3d3d, 4s4s
SeSe7.0-s3p2d24s4s, 4p4p
BiBi8.0-s3p2d2f15d5d, 6s6s, 6p6p

Table I. Basis functions and valence states included in PPs. *For the calculations of atomic charges, the basis functions listed in Table III were used.

Choice of Guiding Functions {Q}\{Q\}

Depending on the targeted eigenstates in the window function of Eq. (4), one can choose either atomic orbitals, hybrid atomic orbitals, or MOs as the guiding functions {Q}\{Q\} in Eq. (3). In this subsection we discuss the three kinds of the localized guiding functions {Q}\{Q\}, i.e., atomic orbitals, hybrid atomic orbitals, and embedded MOs in molecules and bulks, and how they can be generated in our implementation.

In our implementation, the Bloch function ϕkμ\phi_{k\mu} is expanded by PAOs χ\chi [28, 29] as

∣ϕkμ⟩=∑Reik⋅R∑iαckμ,iα∣χRiα⟩,(19)|\phi_{k\mu}\rangle= \sum_{\mathbf{R}} e^{i\mathbf{k}\cdot\mathbf{R}} \sum_{i\alpha} c_{k\mu,i\alpha} |\chi_{\mathbf{R}i\alpha}\rangle, \tag*{(19)}

where ii and α\alpha are atomic and orbital indices, and cc is a linear combination of PAO (LCPAO) coefficient. Also note that ⟨r∣χRiα⟩≡χiα(r−τi−R)\langle\mathbf{r}|\chi_{\mathbf{R}i\alpha}\rangle\equiv\chi_{i\alpha}(\mathbf{r}-\boldsymbol{\tau}_i-\mathbf{R}), where τi\boldsymbol{\tau}_i is the position of atom ii. We use PAOs as the guiding functions {Q}\{Q\} of atomic orbitals, corresponding to valence orbitals or specific orbitals such as localized dd-orbitals in a narrow energy window.

The choice of the atomic orbitals gives a good guiding function, however, the resultant CWFs may not recover the symmetry of CWFs respecting the bond direction to the neighboring atoms unlike hybrid atomic orbitals due to the closest property of CWFs. In this case, it would be better to use the hybrid atomic orbitals rather than the atomic orbitals as explained below. Let us introduce a projection operator for the occupied space by

P^=1NBC∑kμ∣ϕkμ⟩f(εkμ)⟨ϕkμ∣,(20)\hat{P} = \frac{1}{N_{\mathrm{BC}}} \sum_{k\mu} |\phi_{k\mu}\rangle f(\varepsilon_{k\mu}) \langle\phi_{k\mu}|, \tag*{(20)}

where ff is the Fermi-Dirac function. With the projection operator, the density matrix is calculated by

ρRiα,R′jβ=∑kμ⟨χ~Riα∣P^∣χ~R′jβ⟩,=1NBC∑kμeik⋅(R−R′)f(εkμ)ckμ,iαckμ,jβ∗,(21)\begin{aligned} \rho_{\mathbf{R}i\alpha,\mathbf{R}'j\beta} = \sum_{k\mu} \langle\widetilde{\chi}_{\mathbf{R}i\alpha}|\hat{P}|\widetilde{\chi}_{\mathbf{R}'j\beta}\rangle, \\ &= \frac{1}{N_{\mathrm{BC}}} \sum_{k\mu} e^{i\mathbf{k}\cdot(\mathbf{R}-\mathbf{R}')} f(\varepsilon_{k\mu}) c_{k\mu,i\alpha} c^{*}_{k\mu,j\beta}, \tag*{(21)} \end{aligned}

where χ~\widetilde{\chi} is the dual orbital defined by

∣χ~Riα⟩=∑R′jβ∣χR′jβ⟩SR′jβ,Riα−1(22)|\widetilde{\chi}_{\mathbf{R}i\alpha}\rangle= \sum_{\mathbf{R}'j\beta} |\chi_{\mathbf{R}'j\beta}\rangle S^{-1}_{\mathbf{R}'j\beta,\mathbf{R}i\alpha} \tag*{(22)}

holding the following orthonormality:

⟨χ~Riα∣χ~R′jβ⟩=⟨χ~Riα∣χR′jβ⟩=δRR′δijδαβ.(23)\langle\widetilde{\chi}_{\mathbf{R}i\alpha}|\widetilde{\chi}_{\mathbf{R}'j\beta}\rangle= \langle\widetilde{\chi}_{\mathbf{R}i\alpha}|\chi_{\mathbf{R}'j\beta}\rangle= \delta_{\mathbf{R}\mathbf{R}'}\delta_{ij}\delta_{\alpha\beta}. \tag*{(23)}

By setting R=R′=0\mathbf{R} = \mathbf{R}' = 0 and i=ji = j in Eq. (21), and diagonalizing the diagonal block element consisting of ρ0iα,0iβ\rho_{0i\alpha,0i\beta}, which is associated with selected atomic orbitals on the atomic site ii, we obtain the hybrid atomic orbitals respecting the symmetry of the bond direction to the neighboring atoms. We employ the hybrid atomic orbitals as the localized functions {Q}\{Q\} in our implementation. When the same atomic orbitals are chosen to form the diagonal block element as for the case of atomic orbitals {Q}\{Q\}, the resultant CWFs span the same subspace in the Hilbert space.

For some systems the electronic structures can be rather easily understood by employing MOs as building blocks. A molecular crystal is such a case, where the eigenstates near the chemical potential can be approximately constructed by a linear combination of the MOs of constituting molecules. Another example is the bond in molecules and bulks. The bond between two atoms embedded in the system might be analyzed by MOs associated with the two atoms. Here we show how embedded MOs in molecules and bulks can be calculated in a simple procedure. Let us start by noting that using Eq. (20) and assuming that the Bloch function is expressed by Eq. (19) the total number of electrons for a spin degenerate case is given by [32]

Nele=2tr⁡[P^],=2∑Riα⟨χ~Riα∣P^∣χRiα⟩.(24)\begin{aligned} N_{\mathrm{ele}} = 2\operatorname{tr}\left[\hat{P}\right], \\ &= 2 \sum_{\mathbf{R}i\alpha} \langle\tilde{\chi}_{\mathbf{R}i\alpha}|\hat{P}|\chi_{\mathbf{R}i\alpha}\rangle. \tag*{(24)} \end{aligned}

In the second line of Eq. (24) we used the following identity operator:

I^=∑Riα∣χRiα⟩⟨χ~Riα∣=∑Riα∣χ~Riα⟩⟨χRiα∣.(25)\hat{I} = \sum_{\mathbf{R}i\alpha} |\chi_{\mathbf{R}i\alpha}\rangle\langle\tilde{\chi}_{\mathbf{R}i\alpha}| = \sum_{\mathbf{R}i\alpha} |\tilde{\chi}_{\mathbf{R}i\alpha}\rangle\langle\chi_{\mathbf{R}i\alpha}|. \tag*{(25)}

Since Nele=NBCNele(0)N_{\mathrm{ele}} = N_{\mathrm{BC}}N_{\mathrm{ele}}^{(0)} with Nele(0)=2∑iα⟨χ~0iα∣P^∣χ0iα⟩N_{\mathrm{ele}}^{(0)} = 2\sum_{i\alpha}\langle\tilde{\chi}_{0i\alpha}|\hat{P}|\chi_{0i\alpha}\rangle, it is enough to consider Nele(0)N_{\mathrm{ele}}^{(0)}. We utilize the formula for Nele(0)N_{\mathrm{ele}}^{(0)} to calculate embedded MOs in molecules and bulks, and rewrite it the sum of partial traces as

Nele(0)=2∑gtr⁡g[⟨χ~0g∣P^∣χ0g⟩]=2∑gtr⁡g[Λg],(26)N_{\mathrm{ele}}^{(0)} = 2\sum_g \operatorname{tr}_g\left[\langle\tilde{\chi}_{0g}|\hat{P}|\chi_{0g}\rangle\right] = 2\sum_g \operatorname{tr}_g\left[\Lambda_g\right], \tag*{(26)}

where gg is the index of group including PAOs on grouped atoms, e.g., a set of PAOs on a molecule. The notation of ∣χ0g⟩|\chi_{0g}\rangle stands for a subset of PAOs as

∣χ0g⟩=(⋯ ,∣χ0i1⟩,∣χ0i2⟩,⋯ ),(27)|\chi_{0g}\rangle= (\cdots,|\chi_{0i_1}\rangle,|\chi_{0i_2}\rangle,\cdots), \tag*{(27)}

where PAOs on all the atoms in the group gg are included in the subset. The notation of ∣χ~0g⟩|\tilde{\chi}_{0g}\rangle is the counterpart for the dual orbitals. Using the identity operator of Eq. (25), Λg\Lambda_g in Eq. (26) is given by

Λg=∑Rg′ρ0g,Rg′SRg′,0g(28)\Lambda_g = \sum_{\mathbf{R}g'} \rho_{0g,\mathbf{R}g'}S_{\mathbf{R}g',0g} \tag*{(28)}

with definition of block elements:

ρRg,R′g′=⟨χ~Rg∣P^∣χ~R′g′⟩,(29)\rho_{\mathbf{R}g,\mathbf{R}'g'} = \langle\tilde{\chi}_{\mathbf{R}g}|\hat{P}|\tilde{\chi}_{\mathbf{R}'g'}\rangle, \tag*{(29)}
SRg,R′g′=⟨χRg∣χR′g′⟩.(30)S_{\mathbf{R}g,\mathbf{R}'g'} = \langle\chi_{\mathbf{R}g}|\chi_{\mathbf{R}'g'}\rangle. \tag*{(30)}

We now introduce the embedded MOs orbitals in molecules and bulks by a right-singular vectors of Λg\Lambda_g. Since the elements of Λg\Lambda_g are real in case of the collinear DFT, the right-singular vectors can be obtained with real components. However, the SVD of Λg\Lambda_g tends to produce the right-singular vectors with complex components, which results in the complex CWFs making analysis complicated. To obtain the real right-singular vectors, we perform the eigendecomposition of Λg†Λg\Lambda_g^\dagger\Lambda_g as

Λg†Λg=YgΩg2Yg†,(31)\Lambda_g^\dagger\Lambda_g = Y_g\Omega_g^2Y_g^\dagger, \tag*{(31)}

where Ωg2\Omega_g^2 is the diagonal matrix consisting of the eigenvalues of Λg†Λg\Lambda_g^\dagger\Lambda_g, and YgY_g is a unitary matrix consisting of the corresponding eigenvectors {yg,ν}\{y_{g,\nu}\}. By noting that the SVD of Λg\Lambda_g is given by ZgΩgYg†Z_g\Omega_gY_g^\dagger, we see that the right-singular vectors YgY_g can be obtained by diagonalizing Λg†Λg\Lambda_g^\dagger\Lambda_g as YgY_g. The square roots of the eigenvalues of Λg†Λg\Lambda_g^\dagger\Lambda_g are singular values of Λg\Lambda_g, and can be approximately regarded as the occupation of the corresponding eigenvector yg,νy_{g,\nu}. If Λg\Lambda_g is a real matrix, YgY_g obtained by the diagonalization of Λg†Λg\Lambda_g^\dagger\Lambda_g is guaranteed to be a real unitary matrix. We further normalize the eigenvector yg,νy_{g,\nu} by considering the overlap integrals of Eq. (30) as

∣yˉg,ν⟩=∣yg,ν⟩⟨yg,ν∣S0g,0g∣yg,ν⟩,(32)|\bar{y}_{g,\nu}\rangle= \frac{|y_{g,\nu}\rangle}{\sqrt{\langle y_{g,\nu}|S_{0g,0g}|y_{g,\nu}\rangle}}, \tag*{(32)}

and calculate an expectation value of the KS Hamiltonian H^KS\hat{H}_{\mathrm{KS}} with yˉg,ν\bar{y}_{g,\nu} as

ϵg,ν=⟨yˉg,ν∣H^KS∣yˉg,ν⟩.(33)\epsilon_{g,\nu} = \langle\bar{y}_{g,\nu}|\hat{H}_{\mathrm{KS}}|\bar{y}_{g,\nu}\rangle. \tag*{(33)}

After {yˉg,ν}\{\bar{y}_{g,\nu}\} are ordered based on the expectation values, we employ selected ones among {yˉg}\{\bar{y}_g\}, e.g., ones near the chemical potential, as the guiding functions {Q}\{Q\} in Eq. (3) by monitoring the expectation values and the corresponding singular values. They are what we call embedded MOs in molecules and bulks. As demonstrated in Section IV, {yˉ}\{\bar{y}\} work as the good guiding functions to capture the electronic structure of a molecular crystal.

Numerical Results

Wannier Interpolated Bands

We present four numerical examples to demonstrate the broad applicability of the proposed method across various systems.

In Figure 2 (a) and (b), the interpolated bands of Si in the diamond structure, calculated by the TB Hamiltonian derived from CWFs, are compared with those by the conventional calculation. The hybrid atomic orbitals consisting of valence minimal orbitals, which are one ss- and a set of pxp_x-, pyp_y-, and pzp_z-orbitals per Si, were used as the guiding functions {Q}\{Q\}. For the window function of Eq. (4) we use -15.0 and 0.0, relative to the chemical potential, in eV for ϵ0\epsilon_0 and ϵ1\epsilon_1, respectively, which covers the energy range of valence bands. On the other hand, two kBTk_{\mathrm{B}}T of 0.01 and 3.0 eV were used to check how conduction bands are reproduced depending on the effect of smearing. As clearly seen, the two cases well reproduce the valence bands, and the conduction bands are also reproduced well in the case of kBT=3.0k_{\mathrm{B}}T = 3.0 eV. Evaluating the DM function per CWF using Eq. (14), we obtain values of 0.571 and 0.444 for kBTk_{\mathrm{B}}T of 0.01 and 3.0 eV, respectively. For kBT=0.01k_{\mathrm{B}}T = 0.01 eV, the first four singular values at each kk point are close to 1, while the remaining four approach δ(=10−12)\delta(=10^{-12}). Conversely, for kBT=3.0k_{\mathrm{B}}T = 3.0 eV, the eight singular values at each kk point distribute ranging from 1.21.2 to 0.10.1. These differences in the distribution of singular values explain why the value of the DM function for kBT=0.01k_{\mathrm{B}}T=0.01 eV is larger than that for kBT=3.0k_{\mathrm{B}}T=3.0 eV. One of the obtained CWFs is shown in Figure 3 (a), which represents a pp-like CWF. In case of kBT=0.01k_{\mathrm{B}}T=0.01 eV, the conduction bands are treated less importantly due to heavily weighting to the valence bands, resulting in the poor reproduction of the conduction bands. At first glance, one may consider that this is an undesirable feature in such a treatment. However, the feature enables us to generate the minimal atomic-like orthogonal orbitals, which well span the subspace by the valence bands, and to calculate an effective charge allocated to each atom using the minimal atomic-like orthogonal orbitals in an unbiased way. We will demonstrate the calculation of the effective atomic charges and stress the usefulness of the method later on.

Figure 2. Interpolated bands (red circles) of silicon in the diamond structure, calculated by the TB Hamiltonian derived from CWFs with (a) kBT=0.01k_{\mathrm{B}}T=0.01 and (b) 3.03.0 eV in the window function of Eq. (4), respectively. ε0\varepsilon_0 and ε1\varepsilon_1 relative to the chemical potential were set to be −15.0-15.0 and 0.00.0 eV, respectively. The solid line is the original one directly calculated by the conventional scheme. The number of k-points for the Brillouin zone sampling was 13×13×1313 \times13 \times13. The experimental lattice constant of 5.435.43 Å was used. The values of the DM function per CWF are 0.5710.571 and 0.4440.444 for (a) and (b), respectively.

CWFs for Si, Cu, TTF in TTF-TCNQ, and TCNQ in TTF-TCNQ

Figure 3. CWFs for (a) Si, (b) Cu, (c) TTF in TTF-TCNQ, and (d) TCNQ in TTF-TCNQ. In all the cases, isovalues of ±0.04\pm0.04 (orange: 0.040.04, blue: −0.04-0.04) are used for drawing the isosurfaces using OpenMX Viewer [33]. The computational conditions for (a), (b), (c), and (d) are the same as those in Figure 2 (b), Figure 4 (a), Figure 5 (b), and Figure 5 (b), respectively.

The disentanglement of bands in metals is demonstrated for Cu in the FCC structure as shown in Figure 4 (a) and (b). By selecting the parameters ε0\varepsilon_0 and ε1\varepsilon_1, so that the 3d3d bands are included, and only five dd-orbitals as the guiding functions {Q}\{Q\}, the five dd-bands are reproduced as shown in Figure 4 (a). We see that the 3d3d-bands are properly disentangled from the dispersive 4s4s band, and one of the obtained CWFs preserves the shape of dz2d_{z^2}-orbital as shown in Figure 3 (b). When 3d3d-, 4s4s-, and 4p4p-orbitals are used as the guiding functions {Q}\{Q\}, a broad range of bands are reproduced as shown in Figure 4 (b). Since our PP of Cu includes 3s3s, 3p3p, 3d3d, and 4s4s states, the

Figure 4. Interpolated bands (red circles) of copper in the FCC structure, calculated by the TB Hamiltonian derived from CWFs with (a) ε0=−5.5\varepsilon_0=-5.5, ε1=−1.0\varepsilon_1=-1.0 eV, kBT=1.0k_{\mathrm{B}}T=1.0 eV, and hybrid 3d-orbitals as {Q}\{Q\}, and (b) ε0=−11.0\varepsilon_0=-11.0, ε1=24.0\varepsilon_1=24.0 eV, kBT=3.0k_{\mathrm{B}}T=3.0 eV, and hybrid 3d-, 4s-, 4p-orbitals as {Q}\{Q\}. The solid line is the original one directly calculated by the conventional scheme. The number of k-points for the Brillouin zone sampling was 21×21×2121 \times21 \times21. The experimental lattice constant of 3.61 Å was used. The values of the DM function per CWF are 0.081 and 0.211 for (a) and (b), respectively.

the original band structure has the deep 3s3s- and 3p3p-bands. We set the parameters in the window function so as to discard the 3s3s- and 3p3p-bands, and use the 3d3d-, 4s4s-, and 4p4p-orbitals, not 3s3s- and 3p3p-orbitals, as the guiding functions {Q}\{Q\}. So, the 3s3s- and 3p3p-bands are not included in the interpolated bands (not shown). The examples show that the CWFs can be flexibly and easily constructed in accordance with the purpose of study without numerical difficulties.

The interpolated bands for a molecular crystal, consisting of tetrathiafulvalene (TTF) and tetracyanoquinodimethane (TCNQ), are shown in Figure 5 (a) and (b). By employing hybrid atomic orbitals as the guiding functions {Q}\{Q\}, the wide range of bands are reproduced as

Interpolated bands of TTF-TCNQ

Figure 5. Interpolated bands, red lines in (a) and red circles in (b), of TTF-TCNQ, calculated by the TB Hamiltonian derived from CWFs with (a) ε0=−22.0\varepsilon_0=-22.0 and ε1=1.0\varepsilon_1=1.0 eV relative to the chemical potential, kBT=2.0k_{\mathrm{B}}T=2.0 eV, and hybrid atomic orbitals as {Q}\{Q\}, and (b) ε0=−1.0\varepsilon_0=-1.0 and ε1=1.0\varepsilon_1=1.0 eV relative to the chemical potential, kBT=0.01k_{\mathrm{B}}T=0.01 eV, and the 26th MO and the 37th MO for TTF and TCNQ as {Q}\{Q\}. The solid line is the original one directly calculated by the conventional scheme. The number of k-points for the Brillouin zone sampling was 5×19×35 \times19 \times3. An experimental crystal structure was used [34]. The values of the DM function per CWF are 0.441 and 0.022 for (a) and (b), respectively.

TTFTCNQ
indexε\varepsilonω\omegaindexε\varepsilonω\omega
23−0.069-0.0691.3591.359340.1350.1351.0821.082
240.7190.7191.3461.346350.3750.3751.3291.329
250.7610.7611.4491.449360.4770.4771.3361.336
263.4083.4080.8640.864374.9284.9280.5350.535
275.9475.9470.0470.047389.3899.3890.0360.036
286.1976.1970.0470.047399.7959.7950.0190.019

Table 2. Expectation values ε\varepsilon (eV) of the KS Hamiltonian calculated by Eq. (33), which is relative to the chemical potential, and singular values ω\omega of Λg\Lambda_g for embedded MOs in the TTF-TCNQ molecular crystal.

shown in Figure 5 (a). On the other hand, as shown in Figure 5 (b), only four bands near the chemical potential are reproduced using the embedded MOs in the bulk as the guiding functions {Q}\{Q\}. The MOs were calculated by grouping the TTF and TCNQ molecules separately as explained in § 3 B. The expectation values calculated by Eq. (33) and the singular values ω\omega of Λg\Lambda_g are listed in Table II. It can be seen that the singular values largely change from ω26\omega_{26} to ω27\omega_{27} and from ω37\omega_{37} to ω38\omega_{38} for TTF and TCNQ, respectively. So, the 26th MO and the 37th MO for TTF and TCNQ, respectively, were used as the guiding functions {Q}\{Q\}. Since there are two TTF molecules and two TCNQ molecules in the unit cell, we have the four embedded MOs as {Q}\{Q\}, resulting in the four bands. In Figure 3 (c) and (d) two of the resultant CWFs are shown, which localize in TTF and TCNQ molecules, respectively, as expected. The value of the DM function per CWF is found to be 0.022, which implies that the guiding MOs are very close to the resultant CWFs. The example demonstrates that the proposed method provides a direct way to calculate CWFs for targeted bands in molecular crystals.

We extend the proposed method to the non-collinear DFT [36, 37] with spin-orbit interaction (SOI) [27]. The KS orbitals are expressed by two-component spinor, and the SOI is introduced by fully relativistic jj-dependent PPs [23]. The theoretical framework is readily extended to the non-collinear DFT with the SOI without any difficulty. However, it should be noted that the hybrid atomic orbitals and embedded MOs in molecules and bulks as

the guiding functions {Q}\{Q\} can be complex functions in the case. In Figure 6 (a) and (b), we show the Wannier interpolated bands of Bi2_2Se3_3 which is known to be a topological insulator when the SOI is included in the DFT calculation. For the cases without and with the SOI, corresponding to Figure 6 (a) and (b), respectively, it is confirmed that the interpolated bands accurately reproduce the original ones regardless of existence of the band inversion, demonstrating a wide variety of applicability of the proposed method.

Interpolated bands of Bi₂Se₃ without and with spin-orbit interaction

Figure 6. Interpolated bands (red circles) of Bi2_2Se3_3 (a) without and (b) with the spin-orbit interaction, calculated by the TB Hamiltonian derived from CWFs. The solid line is the original ones directly calculated by the conventional scheme. ε0=−6.0\varepsilon_0=-6.0 and ε1=3.0\varepsilon_1=3.0 eV relative to the chemical potential, kBT=1.0k_{\mathrm{B}}T=1.0 eV, and hybrid Bi 6pp-orbitals and Se 4pp-orbitals as {Q}\{Q\}, were used in both the calculations. The number of kk-points for the Brillouin zone sampling was 7×7×77 \times7 \times7. An experimental crystal structure was used [35]. The values of the DM function per CWF are 0.163 and 0.184 for (a) and (b), respectively.

Effective Atomic Charges

From the construction, one can calculate CWFs closest to (hybrid) valence atomic orbitals, while fully respecting the occupied states. The CWFs provide us a unique way to evaluate effective atomic charges. Since the CWFs are a set of orthonormal functions, the effective atomic charge κi\kappa_i of atom ii is easily calculated without arbitrariness as

κi=Ni(val)−∑p∈atom i⟨W0p∣P^∣W0p⟩,=Ni(val)−1NBC∑p∈atom i∑k,μf(εkμ)ukμ,p∗ukμ,p(34)\begin{aligned} \kappa_i &= N_i^{(\mathrm{val})} - \sum_{p\in\mathrm{atom}\ i}\langle W_{0p}|\hat{P}|W_{0p}\rangle,\\ &= N_i^{(\mathrm{val})} - \frac{1}{N_{\mathrm{BC}}}\sum_{p\in\mathrm{atom}\ i}\sum_{k,\mu} f(\varepsilon_{k\mu})u^{*}_{k\mu,p}u_{k\mu,p} \tag*{(34)} \end{aligned}

where P^\hat{P} is the projection operator defined by Eq. (20), and Ni(val)N_i^{(\mathrm{val})} is the number of valence electrons in the PP of atom ii. Though the summation over orbitals pp belonging to atom ii is taken into account in Eq. (34), one can analyze the orbital resolved charges as well. Table III shows effective atomic charges in a HCN molecule and the NaCl bulk, calculated by Eq. (34) and the Mulliken population analysis [38]. Both the systems are known to be notorious cases because of the difficulty in calculating the effective charges [19]. We see that the effective charges by the CWFs quickly converge as a function of basis functions, while those by the Mulliken population are highly dependent on the choice of the basis function. The effective atomic charges obtained from the CWFs could serve as a valuable tool for analyzing electronic structures in a manner related to the approach discussed in Ref. [19]. Further investigation in this direction will be conducted in future studies.

HCNCWFMP
BasisHCNHCN
A6.0-s1p10.077-0.052-0.0250.384-0.164-0.221
A6.0-s2p20.0690.003-0.073-0.0700.321-0.252
A6.0-s2p2d10.0660.009-0.076-0.0080.393-0.385
A6.0-s3p3d20.0660.008-0.0740.1100.298-0.408
A6.0-s3p3d2f20.0660.008-0.0740.1670.297-0.464
NaClCWFMP
BasisNaClNaCl
A9.0-s2p10.391-0.3910.716-0.716
A9.0-s3p20.422-0.4220.595-0.595
A9.0-s3p3d20.421-0.4210.158-0.158
A9.0-s3p3d2f20.421-0.421-0.0620.062

Table III. Effective atomic charges in a HCN molecule and the NaCl bulk, calculated by the proposed method (CWF) and Mulliken population analysis (MP) with a variety of basis functions, where AA represents the constituting atoms. Hybrid atomic valence orbitals are used as the guiding functions {Q}\{Q\}. For the window function, ε0=−55.0\varepsilon_0=-55.0 and ε1=0.0\varepsilon_1=0.0 eV for HCN and ε0=−35.0\varepsilon_0=-35.0 and ε1=0.0\varepsilon_1=0.0 eV for NaCl, relative to the chemical potential, and kBT=0.1k_{\mathrm{B}}T=0.1 eV are used. The number of kk-points for the Brillouin zone sampling is 9×9×99 \times9 \times9 for the NaCl bulk with the lattice constant of 5.63 Å.

Conclusions

We presented a non-iterative method to calculate the closest Wannier functions (CWFs) to a given set of localized guiding functions such as atomic orbitals, hybrid atomic orbitals, and molecular orbitals (MOs). We defined the distance measure (DM) function FF by the sum of the squared distance between the projection functions LL and Wannier functions WW in the Hilbert space, and considered minimizing the DM function. It was shown that the minimization of the DM function is achieved by the polar decomposition of the projection matrix AA

with a window function via the singular value decomposition (SVD) in a non-iterative manner. The CWFs can be uniquely constructed as long as the projection matrix AA is positive definite. It was also shown that the method is free from selective choice of the gauge. The disentanglement of bands is naturally taken into account by introducing a smoothly varying window function, and including NbandN_{\mathrm{band}} Bloch functions to generate NCWFN_{\mathrm{CWF}} CWFs, where NCWF≤NbandN_{\mathrm{CWF}} \le N_{\mathrm{band}}. Even for isolated bands, the method always performs the disentanglement of bands without any difficulty. We have implemented the proposed method into the OpenMX code, which is based on the density functional theory (DFT), numerical pseudo-atomic orbitals (PAOs), and norm-conserving pseudopotentials (PPs), and introduced three types of guiding functions, i.e., atomic orbitals, hybrid atomic orbitals, and embedded MOs in molecules and bulks. The first two are easily employed from the PAOs. For the last one, we developed a method to calculate embedded MOs in molecules and bulks, which focuses on a partial trace formula of the projection operator for the occupied space and applies SVD to the partial matrix for the projection operator. The interpolated bands by tight-binding (TB) models derived from the CWFs reproduce well the targeted conventional bands of a wide variety of systems including Si, Cu, the TTF-TCNQ molecular crystal, and a topological insulator of Bi2_2Se3_3. These successful reproduction of targeted bands clearly demonstrates a wide variety of applicability of the proposed method. We further show the usefulness of the proposed method in calculating effective atomic charges, implying that the CWFs closest to atomic orbitals can be used as a measure to analyze electronic structures from one system to the others. Thus, we conclude that the proposed method is an alternative way in efficiently calculating WFs, and the concept of CWFs will provide a basis for development of novel methods of analyzing electronic structures and calculating physical properties.

Acknowledgments

The author wishes to express gratitude to Prof. Yoshiaki Sugimoto for inspiring the development of CWFs through collaborative research. Thanks are also extended to Mr. Ryotaro Koshoji for his insightful comments on the theoretical aspect. Part of the computation in this study was carried out using the computational facility of the Institute for Solid State Physics at the University of Tokyo.

The upper bound of the diagonal elements of the matrix DD

A proof for the upper bound of the diagonal elements of the matrix DD is given in the appendix. Note that UVUV and BVBV in Eq. (11) are partial unitary matrices in size of Nband×NCWFN_{\mathrm{band}} \times N_{\mathrm{CWF}}. Since DD is hermitian, its eigenvalues are real. Let λ\lambda be an eigenvalue of DD and xx be the corresponding eigenvector. Then, the following equation holds:

12(V†U†BV+V†B†UV)∣x⟩=λ∣x⟩.(35)\frac{1}{2}\left(V^{\dagger}U^{\dagger}BV+V^{\dagger}B^{\dagger}UV\right)|x\rangle=\lambda|x\rangle. \tag*{(35)}

Defining ∣y⟩=UV∣x⟩|y\rangle=UV|x\rangle and ∣z⟩=BV∣x⟩|z\rangle=BV|x\rangle, and operating ⟨x∣\langle x| from the left side of Eq. (35), we have

⟨y∣z⟩+⟨z∣y⟩=2λ⟨x∣x⟩.(36)\langle y|z\rangle+\langle z|y\rangle=2\lambda\langle x|x\rangle. \tag*{(36)}

Noting ⟨y∣z⟩+⟨z∣y⟩≤2∣⟨y∣z⟩∣\langle y|z\rangle+\langle z|y\rangle\leq2|\langle y|z\rangle|, and using the Cauchy-Schwarz inequality ∣⟨y∣z⟩∣≤⟨y∣y⟩⟨z∣z⟩|\langle y|z\rangle|\leq\sqrt{\langle y|y\rangle\langle z|z\rangle}, we obtain

⟨y∣z⟩+⟨z∣y⟩≤2⟨y∣y⟩⟨z∣z⟩.(37)\langle y|z\rangle+\langle z|y\rangle\leq2\sqrt{\langle y|y\rangle\langle z|z\rangle}. \tag*{(37)}

Since ⟨y∣y⟩=⟨z∣z⟩=⟨x∣x⟩\langle y|y\rangle=\langle z|z\rangle=\langle x|x\rangle, combining Eq. (36) and Eq. (37) results in

λ⟨x∣x⟩≤⟨x∣x⟩.(38)\lambda\langle x|x\rangle\leq\langle x|x\rangle. \tag*{(38)}

Considering ⟨x∣x⟩≠0\langle x|x\rangle\ne0, the upper bound of the eigenvalues of DD is found to be

λ≤1.(39)\lambda\leq1. \tag*{(39)}

Noting that the matrix DD can be written by xx and λ\lambda as

D=∑ν∣xν⟩λν⟨xν∣,(40)D=\sum_{\nu}|x_{\nu}\rangle\lambda_{\nu}\langle x_{\nu}|, \tag*{(40)}

the diagonal elements dnnd_{nn} of the matrix DD is given by

dnn=∑ν∣⟨n∣xν⟩∣2λν,(41)d_{nn}=\sum_{\nu}|\langle n|x_{\nu}\rangle|^{2}\lambda_{\nu}, \tag*{(41)}

where ⟨n∣xν⟩\langle n|x_{\nu}\rangle is the nn-th element in the vector xνx_{\nu}. Since the upper bound of λ\lambda is unity, and {x}\{x\} forms a unitary matrix, we obtain the upper bound of the diagonal elements of the matrix DD as

dnn≤1.(42)d_{nn}\leq1. \tag*{(42)}

References

  1. [1]G.H. Wannier, Phys. Rev. 52, 191 (1937).
  2. [2]W. Kohn, Phys. Rev. 115, 809 (1959).
  3. [3]N. Marzari, A.A. Mostofi, J.R. Yates, I. Souza, and D. Vanderbilt, Rev. Mod. Phys. 84, 1419 (2012).
  4. [4]P. Hohenberg and W. Kohn, Phys. Rev. 136, B864 (1964).
  5. [5]D.R. Hamann and D. Vanderbilt, Phys. Rev. B 79, 045109 (2009).
  6. [6]F. Lechermann, A. Georges, A. Poteryaev, S. Biermann, M. Posternak, A. Yamasaki, and O.K. Andersen, Phys. Rev. B 74, 125120 (2006).
  7. [7]N. Marzari and D. Vanderbilt, Phys. Rev. B 56, 12847 (1997).
  8. [8]V. Vitale, G. Pizzi, A. Marrazzo, J.R. Yates, N. Marzari, and A.A. Mostofi, npj Computational Materials 6, 66 (2020).
  9. [9]A. Damle and L. Lin, Multiscale Model. Simul. 16, 1392 (2018).
  10. [10]J. Qiao, G. Pizzi, and N. Marzari, arXiv:2306.00678.DOI
  11. [11]W. Ku, H. Rosner, W.E. Pickett, and R.T. Scalettar, Phys. Rev. Lett. 89, 167204 (2002).
  12. [12]W.C. Lu, C.Z. Wang, T.L. Chan, K. Ruedenberg, and K.M. Ho, Phys. Rev. B 70, 041101(R) (2004).
  13. [13]X. Qian, J. Li, L. Qi, C.-Z. Wang, T.-L. Chan, Y.-X. Yao, K.-M. Ho, and S. Yip, Phys. Rev. B 78, 245112 (2008).
  14. [14]P.-O. Löwdin, J. Chem. Phys. 18, 365 (1950).
  15. [15]J. Des Cloizeaux, Phys. Rev. 135, A685 (1964).
  16. [16]A.A. Mostofi, J.R. Yates, Y.-S. Lee, I. Souza, D. Vanderbilt, and N. Marzari, Comput. Phys. Commun. 178, 685 (2008).
  17. [17]B.C. Carlson and J.M. Keller, Phys. Rev. 105, 102 (1957).
  18. [18]A.E. Reed, L.A. Curtiss, and F. Weinhold, Chem. Rev. 88, 899 (1988).
  19. [19]G. Knizia, J. Chem. Theory Comput. 9, 4834 (2013).
  20. [20]W. Kohn and L. J. Sham, Phys. Rev. 140, A1133 (1965).
  21. [21]Matrix Computations (3rd ed.). Baltimore: The Johns Hopkins University Press (1996), G. Golub and C.F Van Loan.
  22. [22]K. Fan and A. J. Hoffman, Proc. Amer. Math. Soc. 6, 111 (1955).
  23. [23]The code, OpenMX, pseudo-atomic basis functions, and pseudopotentials are available under terms of the GNU-GPL on a web site (http://www.openmx-square.org/).
  24. [24]T. Ozaki and H. Kino, Phys. Rev. B 72, 045121 (2005).
  25. [25]T.V.T. Duy and T. Ozaki, Comput. Phys. Commun. 185, 777 (2014).
  26. [26]I. Morrison, D.M. Bylander, L. Kleinman, Phys. Rev. B 47, 6728 (1993).
  27. [27]G. Theuerich and N.A. Hill, Phys. Rev. B 64, 073106 (2001).
  28. [28]T. Ozaki, Phys. Rev. B 67, 155108, (2003).
  29. [29]T. Ozaki and H. Kino, Phys. Rev. B 69, 195113 (2004).
  30. [30]K. Lejaeghere, G. Bihlmayer, T. Björkman, P. Blaha, S. Blügel, V. Blum, D. Caliste, I.E. Castelli, S.J. Clark, A. Dal Corso, S. de Gironcoli, T. Deutsch, J.K. Dewhurst, I. Di Marco, C. Draxl, M. Dulak, O. Eriksson, J.A. Flores-Livas, K.F. Garrity, L. Genovese, P. Giannozzi, M. Giantomassi, S. Goedecker, X. Gonze, O. Grånäs, E.K. Gross, A. Gulans, F. Gygi, D.R. Hamann, P.J. Hasnip, N.A. Holzwarth, D. Iuşan, D.B. Jochym, F. Jollet, D. Jones, G. Kresse, K. Koepperik, E. Küçükbenli, Y.O. Kvashnin, I.L. Locht, S. Lubeck, M. Marsman, N. Marzari, U. Nitzsche, L. Nordström, T. Ozaki, L. Paulatto, C.J. Pickard, W. Poelmans, M.I. Probert, K. Refson, M. Richter, G.M. Rignanese, S. Saha, M. Scheffler, M. Schlipf, K. Schwarz, S. Sharma, F. Tavazza, P. Thunstöm, A. Tkatchenko, M. Torrent, D. Vanderbilt, M.J. van Setten, V. Van Speybroeck, J.M. Wills, J.R. Yates, G.X. Zhang, and S. Cottinier, Science 351, aad3000 (2016).
  31. [31]J.P. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. 77, 3865 (1996).
  32. [32]T. Ozaki, M. Fukuda, and G. Jiang, Phys. Rev. B 98, 245137 (2018).
  33. [33]Y.-T. Lee and T. Ozaki, J. Mol. Graph. Model. 89, 192 (2019); https://www.openmx-square.org/viewer/
  34. [34]T.J. Kistenmacher, T.E. Phillips, and D.O. Cowan, Acta Cryst. B30, 763 (1974).
  35. [35]S. Nakajima, J. Phys. Chem. Solids 24, 479 (1963).
  36. [36]U. von Barth and L. Hedin, J. Phys. C: Solid State Phys. 5, 1629 (1972).
  37. [37]J. Kübler, K.-H. Hock, J. Sticht, and A. R. Williams, J. Phys. F: Met. Phys. 18, 469 (1988).
  38. [38]R.S. Mulliken, J. Chem. Phys. 23, 1833 (1955).

Paper details

vlnc:W_cqyrwa9q
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

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