pubs.acs.

org/JCTC Article

All-Electron BSE@GW Method with Numeric Atom-Centered Orbitals

for Extended Periodic Systems
Ruiyi Zhou,# Yi Yao,# Volker Blum, Xinguo Ren,* and Yosuke Kanai*
Cite This: J. Chem. Theory Comput. 2025, 21, 291−306 Read Online

ACCESS Metrics & More Article Recommendations *

sı Supporting Information
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

ABSTRACT: Green’s function theory has emerged as a powerful

many-body approach not only in condensed matter physics but also
in quantum chemistry in recent years. We have developed a new
all-electron implementation of the BSE@GW formalism using
numeric atom-centered orbital basis sets (Liu, C. et al. J. Chem.
Downloaded via Yosuke Kanai on January 15, 2025 at 12:05:08 (UTC).

Phys. 2020, 152, 044105). We present our recent developments in

implementing this formalism for extended periodic systems. We
discuss its numerical implementation and various convergence tests
pertaining to numerical atom-centered orbitals, auxiliary basis sets
for the resolution-of-identity formalism, and Brillouin zone
sampling. Several proof-of-principle examples are presented to
compare with other formalisms, illustrating the new all-electron BSE@GW method for extended periodic systems.

1. INTRODUCTION systems. Our previous work has also shown that the all-electron
In the past decade, the many-body perturbation theory based on BSE@GW method using atom-centered orbital basis sets
the Green’s function formalism has found its way into the provides high accuracy for modeling core−electron excitations,
chemistry community from the condensed matter physics comparable to the state-of-the-art EOM-CCSD method.24 This
community. By going beyond well-accepted approximations for computational capability to quantitatively predict X-ray
condensed matter systems (e.g., plasmon-pole approximation, absorption spectra of molecules is of great interest as many
etc.), a number of groups have shown that the so-called GW and light-source facilities have undergone great advancement in
Bethe−Salpeter equation (BSE) methods can be made quite recent years.
promising for studying excited state properties of isolated Building on our recent effort on developing all-electron many-
molecules.1−4 Solving the particle-hole two-particle Green’s body perturbation theory methods using numeric atom-
function, the BSE method5−8 has become increasingly popular centered orbital (NAO) basis sets (e.g., GW methods for
for calculating neutral excitations of molecules in recent years,4 extended periodic systems25 and the BSE for isolated
adding to a history of successes for condensed matter systems24,26), we here extend the BSE@GW approach to
systems.9−11 It is now widely recognized as a promising periodic systems with Brillouin zone (BZ) integration. For
alternative12−14 to density functional theory (DFT)-based studying extended periodic systems, taking into account the
approaches such as linear-response time-dependent density dependence on the reciprocal wave vector requires careful
function theory (LR-TDDFT)15,16 and traditional wave consideration. Most GW/BSE method developments originat-
function-based methods like the equation-of-motion coupled ing in condensed matter physics are based on plane-waves or
cluster (EOM-CC).17,18 real-space-grids with nonlocal pseudopotentials, and the
The BSE method is based on the many-body perturbation numerical formulations of these Green’s function methods are
theory in the Green’s function (G) framework, and the screened largely incompatible with the mathematical/numerical frame-
Coulomb interaction, W, is used to model the interaction works used in many quantum-chemistry developments/codes.
between the excited electron and hole. Combined with the GW The present theoretical method and algorithm will benefit the
approximation to the self-energy, the BSE method has been quantum chemistry field, which is largely based on all-electron
shown to successfully yield the optical spectra of solids and
nanostructured systems, and more recent work also shows
similar applicability to low-energy electronic excitation of Received: September 22, 2024
molecular systems.19−23 The BSE@GW approach yields the Revised: December 5, 2024
accuracy of 0.1−0.2 eV, being comparable to the EOM-CCSD Accepted: December 6, 2024
method.13,14 Due to its favorable scaling of N4 in terms of the Published: December 30, 2024
number of electrons, the approach is highly promising for
studying excited-state properties of increasingly complex

© 2024 American Chemical Society https://doi.org/10.1021/acs.jctc.4c01245

291 J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

implementations with atom-centered basis functions including (x1, x 2 ; x 1, x 2 ; )

Gaussian-type orbitals. Furthermore, some of us have recently ÄÅ ÉÑ
demonstrated significant efficiency increases of numerically ÅÅÅ AS (x1, x 1)AS*(x 2, x 2) AS (x 2, x 2)AS*(x 1, x1) ÑÑÑ
=i ÅÅ ÑÑ
precise exact exchange and hybrid DFT for periodic system sizes ÅÅ +
+ S i 0+ ÑÑÑÖ
S 0Å Ç S + i0
exceeding 10,000 atoms in size, using the NAO formalism.27
(6)
Thus, the groundwork laid in the present paper for BSE@GW
should be extendable to significantly larger scales as well in where 0+ is a positive infinitesimal, AS represents the amplitude
future developments. Overall, our development will enable the associated with the coupled electron−hole pair for the S-th
quantum chemistry community to take advantage of recent excited state: AS(x, x′) = −⟨N, 0|ψ̂ †(x)ψ̂ (x′)|N, S⟩ and ωm
exciting advances in the methods based on Green’s function corresponds to the excitation energy from the ground state to
theory and promote further synergies with traditional post- the S-th excited state ωS = EN,S − EN,0.
Hartree−Fock methods. The Bethe−Salpeter equation (BSE) relates the two-body
correlation function to the noninteracting one as28
2. GREEN’S FUNCTION THEORY
(1, 2; 1 , 2 ) = 0(1, 2; 1 , 2 )
2.1. Bethe−Salpeter Equation. For describing electron−
hole pairs in a many-electron system, the two-body correlation + d(3456) 0(1, 4; 1 , 3)K (3, 5; 4, 6)
function plays a central role in many-body perturbation theory
based on Green’s functions. The one-body Green’s function G1 (6, 2; 5, 2 ) (7)
and the two-body Green’s function G2 are defined as
where 0(1, 2; 1 , 2 ) is noninteracting correlation function,
†
iG1(1, 2) = N , 0|T[ (1) (2)]|N , 0 (1) given by

i 2G2(1, 2; 1 , 2 ) = N , 0|T[ (1)[ (2) †

(2 ) †
(1 )]|N , 0 0(1, 2; 1 , 2 ) = G1(1, 2 )G1(2, 1 ) (8)
(2) The electron−hole interaction kernel, K(35; 46), is defined as
where the creation operator ψ̂ (1) and the annihilation operator [vH (3) (3, 4) + (3, 4)]
ψ̂ †(2) are the field operators written in the Heisenberg picture: K (3, 4; 5, 6) =
̂ ̂ ̂ ̂
ψ̂ (1) = eiHt1 ψ̂ (x1)e−iHt1 and ψ̂ †(2) = eiHt2 ψ̂ †(x2)e−iHt2, 1 ≡ (x1, t1), G(6, 5) (9)
2 ≡ (x2, t2) denotes a composite variable encompassing space, where vH(1) = −i∫ d2v(1, 2)G(2, 2+) is the Hartree potential
spin, and time. Here, |N, 0⟩ is the ground state state for a N- and Σ(3, 4) is the self-energy. 1+ is used to indicates t+1 = t1 + 0+.
electron system and T is the time ordering operator. The By adopting the widely used GW approximation29 to the self-
correlation function, , is formally given as a functional energy Σ(3, 4) = iG(3, 4) W(3, 4), the kernel simplifies to30
derivative of the one-body Green’s function G(1, 1′) with
respect to an external nonlocal perturbation U(2′, 2) K (3, 4; 5, 6) = i (3, 4) (5 , 6)v(3, 6)
G(1, 1 ) + i (3, 6) (4, 5)W (3+ , 4; ) (10)
(1, 2; 1 , 2 ) =
U (2 , 2) (3) where δ represents Dirac delta function. Although the frequency
In terms of the Greens’ functions, it is expressed as dependence of W has been considered in solving BSE,30−32 most
standard implementations adapt the so-called static screening
(1, 2; 1 , 2 ) = G2(1, 2; 1 , 2 ) + G1(1, 1 )G1(2, 2 ) effect approximation by neglecting the frequency dependence30
(4) such that eq 10 reduces to
(1, 2; 1 , 2 ) describes the probability amplitude of an electron K (r3, r5, r4, r6) = iv(r3 r5) (r3 r4) (r5 r6)
propagating from 1′ to 2 and a hole propagating from 1 to 2′. By
incorporating eqs 1 and 2 into eq 4 and using the completeness + iW (r3, r4) (r3 r6) (r4 r5) (11)
property of N-electron system in an excited state S |N, S⟩ (i.e.,
∑S|N, S⟩⟨N, S| = I), we can expand the correlation function as In our subsequent discussion, we use the notation W(r, r′) to
following represent the static screened interaction in the limit of ω = 0 for
brevity. The BSE interaction kernel includes two physically
(x1, x 2 ; x 1, x 2 ; ) distinct terms: first, the exchange term, which results from the
† †
bare Coulomb potential v, and second, the direct interaction
= ( ) N , 0| (x 1) (x1)|N , S N , S| (x 2 ) (x 2) term coming from the screened exchange interaction W. Note
S 0 that this is in contrast with the one-electron self-energy case, and
|N , 0 e i(EN ,S EN ,0)
+ ( ) N , 0| can be seen from the structure of Feynman diagrams. The direct
S 0
interaction term is responsible for the attractive nature of
electron−hole interaction and formulation of bound electron−
† †
(x 2) (x 2)|N , S N , S| (x 1) (x1)|N , 0 hole states (exciton states). On the other hand, the exchange
i(EN , S EN ,0) (5) interaction term controls details of the excitation spectrum, such
e
as the splitting between spin-singlet and spin-triplet excitations.7
where EN,S is the energy for |N, S⟩. Due to time translation In the limit in which W approaches the value of v as the dielectric
invariance, the linear-response function depends only on τ = t1 − term ε approaches one, it reduces to time-dependent Hartree−
t2. By employing Fourier transformation, we arrive at the Fock theory.
Lehmann representation of the two-body correlation function 2.2. Bethe−Salpeter Equation as Eigenvalue Problem.
within the frequency domain For the first-principles theory implementation of BSE, it is
292 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

convenient to write the BSE amplitudes in eq 6 in terms of the excitations. The explicit forms of the matrix elements of V and W
particle-hole basis will be discussed in a later subsection. Meanwhile,we can
observe that the eigenvalues of H2p in eq 16 correspond precisely
AS (x , x ) = X vc , S c(x) v*(x ) + Ycv , S v(x) c*(x ) to the poles in the Lehmann representation of (see eq 6). The
vc (12) eigenvalues represent the excitation energies, ωS = EN,S − EN,0, of
where ψc and ψv represent single-particle orbitals of the the N-electron system in terms of the particle-hole excitation.
conduction band (i.e., unoccupied) and valence band (i.e., The corresponding eigenvectors give the amplitudes in the BSE
occupied), respectively. ψc(x)ψ*v (x′) and ψv(x)ψ*c (x′) represent as defined in eq 12. To simplify the eigenvalue problem of H2p,
the particle-hole basis functions, and they correspond to we can separate the particle-hole basis pair (n1, n2) into resonant
resonant and antiresonant transitions of electron−hole pairs, pairs (i, a) and antiresonant pairs (a, i), the particle-hole basis
respectively. Typically, by adapting the G0W0 approximation for pair (n3, n4) into (j, b) and (b, j), where the orbital indices i, j
the self-energy calculation, the orbitals from mean-field theories refer to the valence band states (i.e., occupied orbitals) while a, b
such as Hartree−Fock (HF) and Kohn−Sham density func- refer to the conduction band states (i.e., unoccupied orbitals).
tional theory (KS-DFT) are used for the single-particle orbitals As discussed by Rohlfing et al.30 and Strinati et al.,5 the BSE is
in practice. Thus, the matrices X and Y represent the solutions to finally written in a numerically convenient form as an eigenvalue
the BSE, which needs to be solved. equation as often encountered in quantum chemistry and
The noninteracting correlation function 0 (eq 8) can be condensed matter physics
ÄÅ ÉÑÄÅ ÉÑ ÄÅ ÉÄ É
written using the Lehmann representation, analogously to the
ÅÅÅ A B ÑÑÑÅÅÅÅ X m ÑÑÑÑ ÅÅ 0 ÑÑÑÅÅÅÅ XS ÑÑÑÑ
ÅÅ Ñ Å ÑÑÅÅ ÑÑ
ÅÅ B* A*ÑÑÑÅÅÅÅ Y ÑÑÑÑ = SÅÅÅÅ 0
derivation of eq 6 from eq 5 ÑÑÅÅY ÑÑ
ÅÇ ÑÖÅÇ m ÑÖ Ç ÑÖÅÅÇ S ÑÑÖ (18)
0(x1,
x 2; x 1, x 2 ; )
ÅÄÅ Here ωS is the excitation energy (see eq 6), and (XS, YS)
ÅÅÅ c(x1) v(x2) v*(x1) c*(x 2) correspond to the eigenvectors defined in eq 12. The matrix
=i ÅÅ
Å ( QP QP +
v,c Å v ) + i0
blocks denoted by A correspond to the Hamiltonian describing
ÅÇ c
ÑÉ
resonant transitions from occupied to unoccupied orbitals, while
(x ) (x ) *(x 1) v*(x 2) ÑÑÑ the matrix blocks represented by −A* correspond to the
v 1 c 2 c ÑÑ
Ñ
+ ( QP
c
QP
v ) i 0+ ÑÑÑÖ (13)
Hamiltonian for antiresonant transitions from unoccupied to
occupied orbitals. Similarly, the blocks represented by B and
where εQP n is the quasi-particle energy for the quasi-particle −B* account for the coupling between resonant and
orbital indexed by n. antiresonant pairs. The matrices A and B are given by
For convenience, we introduce the variable z = ω + i0+sgn( f n1 Aiajb = ( QP QP S /T
a i ) ij ab + ia|V |jb ij|W |ab
d

− f n2), where f n is the occupation number for orbital index n. In

d

(19)
terms of the particle-hole basis (see eq 12), we can arrive at the
numerically convenient matrix representation for 0 Biabj = S/T
ia|V |bj ib|W |aj (20)
The operators V̂ and Ŵ represent the bare Coulomb operator
0(x1, x 2; x 1, x 2 ; )
and static screened Coulomb operator, respectively, and
=( *
0)n1n 2 ; n3n4 (z) n1(x1) n 2(x 2) n3(x 1) n4 (x 2)
* numerical evaluation of these matrices are discussed in the
(14)
subsequent sections. This BSE eigenvalue problem has a
where the diagonal matrix 0 is defined as mathematically similar form as the widely known Casida
fn fn equation of LR-TDDFT15 while their theoretical origins are
( 2 1 quite different as discussed above.
0)n1n 2 ; n3n4 (z) =i QP QP n1n4 n 2n3
z ( n1 n2 ) (15) Although the blocks A and −A* are Hermitian, the overall
matrix is non-Hermitian. Numerically, solving this non-
1
Noting (z) = 0 1(z) K , the BSE (eq 7) is expressed in Hermitian eigenvalue problem is quite complicated even though
the matrix representation as a wide range of efficient eigensolvers have been developed for
this specific type of numerical problem in recent years.34−37 In
( )n1n2; n3n4 (z) = [ 0(z) K ]n11n2; n3n4 practical implementation, the Tamm−Dancoff approximation
(TDA) is widely used. The TDA amounts to neglecting the
= i[H Ip z ]n11n2; n3n4 (fn fn ) (16) coupling matrix between excitation and de-excitation pairs (i.e.,
2 4

2p B and −B*), reducing the problem to a Hermitian eigenvalue

where is the identity matrix, and H is the two-particle problem
Hamiltonian33 given by
AXS = S XS (21)
(H2p)n1n2 ; n3n4 = ( QP
n2
QP
n1 ) n1n4 n2n3 + (fn fn )
1 3 This simplification is justifiable as long as the energy associated
s /t with particle-hole interaction remains significantly smaller than
( V W )n1n2 ;n3n4 (17)
the quasi-particle energy gap. As discussed in a previous
In this work, we neglect the spin−orbit coupling. All the study,4,37−39 the TDA works particularly well especially for
benchmark absorption spectra presented in this work are solids in the optical limit. In quantum chemistry, the TDA is also
specifically for singlet excitations in closed-shell systems, such as used in the context of Casida’s equation of LR-TDDFT and the
silicon (Si) and magnesium oxide (MgO). Thus, the spin degree Configuration Interaction Singles (CIS) methods.40,41
of freedom can be integrated out. For transitions with different For extended periodic systems, it is straightforward to extend
spins, we have αs = 2 for singlet excitations and αt = 0 for triplet this formalism to the BZ with Bloch states. Particle-hole pairs
293 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

here include those excitations from the valence band at the k- agreement within 0.2 eV has been observed for the G0W0
point k1 to the conduction band at the k-point k1 + k0 such that benchmark calculation under “tier 2” basis sets with respect to
k0 represents the momentum change in the excitation. Equation LAPW+lo results for periodic systems.25 Meanwhile, the
18 then becomes molecular BSE benchmark calculations showed that “tier 2”
basis set is adequate for accurately capturing the electron−hole
Aiajbkk1 2 = ( QP
ak1+ k0
QP
i k1 ) ij ab k1k 2 + S/T
i k1a k1 + k0| interaction of low-lying valence excited states when additional
diffuse augmentation Gaussian functions “aug”53 are included in
V |j k 2b k 2 + k0 i k1j k 2|W |a k1 + k0bk 2 + k0 the basis set.26 These recent developments pave the way for the
(22) work presented here. Building on our all-electron NAO-based
In a typical calculation of the optical absorption spectrum, where GW method for extended periodic systems25 and all-electron
electron−phonon coupling can be neglected, it is assumed that NAO-based BSE method for isolated systems,24,26 we introduce
there is no momentum change involved. Therefore, k0 vector is a new all-electron NAO implementation of the BSE method for
generally taken to be zero. With this consideration, we construct extended periodic systems in this work.
the Bethe−Salpeter equation (BSE) Hamiltonian and solve the
associated Hermitian eigenvalue problem of matrix A in eq 21. 3. ALL-ELECTRON IMPLEMENTATION WITH NUMERIC
with ATOMIC ORBITALS BASIS FOR EXTENDED
PERIODIC SYSTEMS
Aiajbkk12 = ( QP
a k1
QP
i k1 ) ij ab k1k 2 + S/T
i k1a k1|V |j k 2b k 2
Throughout this section, we utilize the following indices: i, j, k, l
i k1j k 2|W |a k1b k 2 (23) for denoting occupied Kohn−Sham (KS) orbitals, and a, b, c, d
for denoting unoccupied KS orbitals. For the atomic orbital
i k1ak1|V |j k 2b k 2 = drdr k1
(r) k1* k 2* (AO) basis, we employ the indices m and n, while the Greek
i a (r)v(r , r) j (r )
letters μ, ν, α, and β are used for auxiliary basis functions (ABFs)
k2 applied in the resolution of identity approach. k1 and k2 are used
b (r ) (24)
for k-point sampling in the BZ for KS orbitals while q is used for
i k1j k 2|W |a k1b k 2 the grid sampling in the BZ for ABFs.
3.1. Numeric Atomic Orbital Basis Representation.
k1 k 2* k1* k2 NAO basis functions have the general form
= drdr i
(r) j
(r)W (r, r) a
(r ) b
(r )
(25)
un(r )
The absorption spectrum, given by ε2, can be obtained from the
n(r) = Yl , ml( )
(28)
excitation energy ωS and the exciton wave function coefficients r
XS as
where un(r) is the radial part and numerically tabulated. Yl,ml(Ω) d

16 2e 2 2 are real-valued functions, comprised of either the real parts (ml =

2( )= 2
e· 0|v|S | ( S) 0, ..., l) or the imaginary parts (ml = −l, ..., −1) of the complex-
S (26)
valued spherical harmonics. The indices l and ml are quantum
where numbers, describing the angular momentum quantities of
spherical harmonic functions Yl,ml(Ω) associated with the basis
0|v |S = vk|v |ck X vck, S d

vck (27) function index n.

The definition of the NAO basis functions allows for using a
and v̂ is the velocity operator and e is the direction of the wide range of shapes, including both analytically and numeri-
polarization of light. cally defined functions. This includes traditional quantum
First-principles computational methods based on Green’s chemistry’s analytically defined Gaussian-type or Slater-type
function theory like GW and BSE originate formally from orbitals. A key advantage of the NAO basis is the flexibility
application of quantum field theory in condensed matter associated with choosing un(r), and one can select numerical
physics, and they have traditionally been formulated with solutions for the Schrödinger-like radial equations54
plane waves as the basis sets along with the use of
pseudopotentials.6,7,30,39,42 In recent years, there has been a ÅÄÅ ÑÉÑ
ÅÅ 1 d2 l(l + 1) ÑÑ
ÅÅ + + v ( r ) + v ( r ) ÑÑun(r ) = nun(r )
growing interest in formulating GW and BSE methods using ÅÅÅÇ 2 dr 2 2r 2
n cut
ÑÑÑÖ
atom-centered basis sets in the context of traditional molecular
(29)
quantum chemistry.3,4,24−26,43−51 Gaussian48 and NAO-based25
GW methods have been also demonstrated for extended This Schrödinger-like equation includes a potential term,
periodic systems, with reciprocal space summations over the vn(r), which determines the primary behavior of un(r), and
BZ. As an important prerequisite, we emphasize that the valence another steeply increasing confining potential, vcut(r). The
and low-lying conduction energy band structures at the Kohn− confining potential can be chosen to ensure that each radial
Sham level of theory are numerically completely converged function, un(r), decays smoothly and becomes strictly zero
using “tier 2” NAO basis sets. In a broad benchmark of DFT-KS beyond a specific confinement radius. For a comprehensive
PBE band structures between FHI-aims code and another all- discussion on the NAO basis set, readers are referred to ref 54.
electron Wien2k code (full-potential (linearized) augmented For a periodic system, the Kohn−Sham (KS) orbital, denoted
plane-wave ((L)APW) + local orbitals (lo) method), average as ψki/a(r), can be represented as a linear combination of Bloch-
deviations are shown to be <0.01 and <0.02 eV for the valence adapted atomic orbitals as the basis set functions
band range and for the conduction band range up to 5 eV above k
the conduction band minimum, respectively.52 In Green’s i/a
(r) = e ik·R cmk , i / a m(r m R)
function theory calculations using NAO basis functions, an m R (30)

294 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

where φm is the NAO basis function centered at the atomic P q(r) = P q (r R )e iq·R
position τm, from which the m-th atomic basis originates within R (37)
the unit cell, and the sum runs over all unit cells R in the Born−
von Karman (BvK) supercell. Cμm,n(k + q, k) is the atomic orbital (AO) based RI expansion
3.2. BSE in the Auxiliary Basis Set of Resolution of coefficient, which depends on two independent Bloch wave-
Identity. Construction of the particle-hole kernel, through eqs vectors, k + q and k. Following refs 50 and 25 the matrix
24 and 25 is a major computational task. The direct evaluation of representations of the Coulomb operator V̂ and static screened
four-center integrals has historically posed challenges due to Coulomb operator in terms of the ABFs read
their significant computational and memory requirements in
first-principles theory. The so-called Resolution of Identity (RI) P q *(r)P q(r )
V (q) = drdr
approximation, also known as the density fitting, is a commonly |r r|
employed method to alleviate the large computational cost in
calculations with atom-centered orbitals like NAOs and W (q) = P q *(r)W (r , r )P q(r )drdr
(38)
Gaussians as basis functions.55,56 Hartree−Fock57,58 and other
post-Hartree−Fock methods such as second-order Møller− With the definition of the screened Coulomb operator Ŵ , the
Plesset perturbation theory (MP2)59,60 and coupled cluster matrix can be computed from the static dielectric matrix25 such
(CC)61,62 often utilize the RI method. The RI approximation that
streamlines the calculation by reducing all four-center two-
electron Coulomb integrals to precomputed three- and two- W (q) = V1/2(q) 1
, (q)V1/2(q)
center integrals.55,63 (39)
The present all-electron NAO-based implementation also
employs the RI approximation through constructing a set of where V1/2 represents the square root of the Coulomb matrix V
NAO auxiliary basis functions to expand the products of two and ε represents the symmetrized static dielectric function,
NAO orbitals, as described in refs 25 and 50. For isolated whose matrix elements are computed as
systems (nonperiodic case), the product of two NAO basis
functions can be approximated within the RI approximation as a (q) = V1/2(q) 0, (q)V1/2(q)
linear combination of auxiliary basis functions (ABF) as (40)

* χ0 is the noninteracting static response function, according to

m (r) n(r) = Cm , nP (r )
(31) the Adler−Wiser formula65,66
k+ q* k
where Pμ(r) represents the μ-th auxiliary basis function and Cμm,n 1BZ
i
(r) a
(r) ak *(r ) i
k+q
(r )
is the expansion coefficient for the three-orbital (triple) 0 (r , r ) = 2wk wq k+q k
expansion. The expansion coefficient Cμm,n is given by i , a k, q i a
(41)
1
Cm , n = mn|V | V
(32) We need this response function given in the basis of ABFs, χ0,αβ,
in eq 40. To this end, we introduce the molecular orbital (MO)
where ⟨mn|V̂ |ν⟩ is the three center-Coulomb integral given by based RI expansion coefficients C̃ (k + q, k) such that
*(r) (r)P (r ) Naux
m n
mn|V | = drdr k+q* k
Ci / a , j / b(k + q, k)P q *(r)
|r r| (33) i / a (r) j / b(r) =
(42)
and Vμν is the two-center Coulomb integral
where ψk are KS orbitals. The MO-based expansion coefficients
P (r)P (r ) are related to the AO-based expansion coefficients by
V = drdr
|r r| (34)
Ci , j(k1, k 2) = cm*, i(k1)cn , j(k 2)Cm , n(k1, k 2)
The four-centered two-electron Coulomb integral, for instance,
m,n
can be conveniently calculated as
Ca , b(k1, k 2) = cm*, a(k1)cn , b(k 2)Cm , n(k1, k 2)
mn|V |pq = Cm , nV Cp, q
m,n (43)
(35)
This approach applicable in the nonperiodic case is referred to as where cm,i/a are molecular orbital (KS) coefficients which depend
the “RI-V” method in the subsequent discussion. only on a single wave vector (see eq 30) and Cμm,n(k1, k2) are the
3.3. Extended Periodic Systems. For extended periodic AO-based expansion coefficients (see eq 36). Both the AO-
systems, the products of Bloch-based atomic orbitals can be based and MO-based expansion coefficients depend on two
expanded using Bloch-based atom-centered Auxiliary Basis momentum vectors k1 and k2. Then, the noninteracting
Functions (ABFs) response function χ0 in the auxiliary basis is given by
Naux Ci , a(k + q, k)Ca , i(k, k + q)
k+ q*
m (r) nk (r) = Cm , n(k + q, k)P q*(r) 0,
( q) = wk k+q k
(36) i ,a k i a (44)

where Naux represents the number of ABFs within each unit cell Finally, the matrix elements of the Coulomb operator V̂ and the
and the Bloch-based atom-centered ABFs, Pqμ(r), are defined static screened Coulomb operator Ŵ needed for constructing
through Bloch theorem as64 the BSE Hamiltonian (see eq 23) can be computed as
295 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

* μ(Rn) μ(0)
Ci , a (k1, k1)V (0)Cj ,b(k 2, k 2) systems, i.e., Cm(Rm),n(Rn)
= Cm(R m−Rn),n(0)
, and eq 47 can be
i k1a k1|V |j k 2b k 2 = d d d d

(45) transformed as
*
i k1j k 2|W |a k1bk 2 = Ci , j (k1, k 2)W (k 2 k1)Ca , b(k1, k 2) m(r Rm m) n(r Rn n)

Cm((00),) n(R n R m)P (r Rm m)

(46)
M
3.3.1. Local RI Technique. In contrast to the molecular case,
where the RI-V method (eq 32) can be directly applied, solving + Cm((R0)m R n), n(0)P (r Rn n)
for these coefficients in extended periodic systems presents N (49)
significant difficulties. One major obstacle arises from the long- Through Fourier transformation, the product of Bloch-based
range nature of two-centered and three-centered integrals, atomic orbitals in LRI approximation can be derived as
necessitating the use of Ewald summation techniques in the
k+ q*
integral construction. While notable progress has been made in
m (r) nk (r) = e i(k + q)·R m ik·R n
e m( r Rm m) n
recent years on addressing this challenge,67,68 it remains a highly R m, R n
nontrivial task to implement them. Additionally, it is worth
noting that the computational cost of computing and storing AO (r Rn n)
triple coefficients scales as O(NauxN2bN2k) where Nb represents e i(k + q)·R m ik·R n
e
the number of basis functions, and Nk represents the number of R m, R n
k-points, and thus the RI-V formalism is computationally quite ÄÅ
ÅÅ
expensive. To address these issues, FHI-aims implementation ÅÅ
ÅÅ Cm((00),) n(R n R m)P (r Rm m)
utilizes the LRI (Local Resolution of Identity) approximation,, ÅÅ
called RI-LVL in ref 69. Within the LRI approximation, the ÅÅÇ M
ÉÑ
ABFs are used to expand the product of two NAOs are limited to ÑÑ
ÑÑ
those ABFs centered on the same two atoms on which the NAOs + Cm((R0) R n), n(0)P (r Rn n)Ñ
Ñ
ÑÑ
are centered. In the quantum chemistry community, this two- N
m
ÑÑÖ
center LRI scheme is also referred to as the Pair-Atom RI ÄÅ
(PARI) approximation.70,71 In the context of extended periodic ÅÅ
ÅÅ
= ÅÅ e iq·R m
P (r Rm m)
systems, the LRI approximation has been implemented for ÅÅ
hybrid exchange-correlation functionals for DFT,27,72,73 ÅÅÅ R m
M
Ç
MP2,50,74 RPA,50,75 and GW25 methods within the NAO basis ÑÉÑ
Ñ Ñ
framework. e ik·(R n R m)
Cm((00),) n(R R )ÑÑÑÑ
m Ñ
In real space, the two NAOs, labeled as m and n, can originate n
ÑÑÑ
from different unit cells, denoted by two Bravais lattice vectors
Rn
Ö
ÅÄÅ
Rm and Rn. The LRI approximation for extended periodic ÅÅ
systems implies that72 ÅÅ e iq·R n
+ ÅÅ P (r Rn n)
ÅÅ R
N ÅÅÇ n
m(r Rm m) n(r Rn n) ÉÑ
ÑÑ
ÑÑ
(R m) i(k + q)·(R m R n)
Cm((R0)m Ñ
R n), n(0)Ñ
Cm(R m), n(R n)P (r Rm m) e ÑÑ
ÑÑ
M Rm ÑÖ
+ Cm((RR mn)), n(R n)P (r Rn n) = Cm((0)k q*
q), n(0)P (r)
N (47) M

where M and N constitute the atoms on which AO basis + Cm((00),) n(k)P q *(r)
functions φm and φn are centered, and the summation over the N (50)
ABFs is restricted to those ABFs centered on those atoms. By
minimizing the self Coulomb repulsion of the expansion error Here, as shown in the above equation, the terms Cμ(0)m(−k−q),n(0) and
given by eq 47, the expansion coefficient can be determined as69 Cμ(0)
m(0),n(k) can be determined through the Fourier transformation
of the real-space term Cμ(0) μ(0)
m(R),n(0) and Cm(0),n(R). Therefore, by
l
o
o
o m(0), n(R)|V | (V MN ) 1
comparing eqs 36 and 50, we can obtain the atomic centered
o
o
o {M , N(R)} expansion coefficients in the reciprocal space using the LRI
Cm((00),) n(R) = o
m
o approximation as25
o
o for M
o
o
o
o l
o
n 0 otherwise (48) o
o C m((0)k q), n(0) M
o
o
o
Instead of solving the inverse matrix of the entire two-electron Cm , n(k + q, k) = o
m
o Cm((00),) n(k) N
Coulombic matrix as seen in eq 32, the LRI method computes o
o
o
o
the inverse of a local metric within the domain of ν ∈ M, N(R), o
o
centered either on the atom M in the original cell 0 or on the n 0 otherwise (51)
atom N in the cell specified by R as shown in the above equation. In summary, using the above set of working formulae, one can
Detailed discussion on the LRI approximation can be found in efficiently calculate the AO-based expansion coefficients in the
ref 69. reciprocal space and subsequently, through a linear trans-
To further derive the expansion coefficient in reciprocal space, formation, derive the MO-based expansion coefficients. This
we utilize the transnational symmetry property of periodic approach enables an efficient computation of the matrix
296 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

elements for Coulombic and static screened Coulombic

interactions in the BSE formalism, as expressed in eqs 45 and form, represented as vcut(|r − r′|). The expression for vcut(|
46 within the LRI approximation. An important advantage of r − r′|) is given by
ÄÅ É
this approximation is that the AO-based expansion coefficients
erfc( r) 1 ÅÅ ln(r ) ln(R cut) ÑÑÑ
become dependent on only a single vector, either −k−q or k, cut
+ erfcÅÅ Å
Å ÑÑ
v (r ) = ÑÑ
within the BZ, rather than both simultaneously. As a result, the r 2 ÅÅÇ ln(R ) ÑÑÖ
computational cost and memory storage requirements asso- erfc( r )
ciated with the RI coefficients can be significantly reduced. ×
r (53)
3.3.2. Singularity Treatment at Γ Point. One outstanding
technical challenge for periodic BSE calculation are the where Rcut represents the cutoff radius. This expression
singularities that appear in the Coulomb term V and the static retains the short-range part of the Coulomb potential
screened Coulomb term W of the BSE kernel at the Γ point. In while rapidly suppressing the long-range part beyond Rcut.
three-dimensional (3D) systems, the inherent nature of the bare The value of Rcut is determined as the radius of a sphere
Coulomb potential, characterized by the 1/r behavior, leads to a inscribed inside the Born−von Kármán (BvK) supercell.
1/q2 divergence as q approaches 0 in the reciprocal space. Within As the density of the k-point mesh increases, Rcut gradually
the traditional plane-wave basis functions, eiG·r, the Coulomb grows, allowing for the restoration of the full bare
operator is well-known to have a specific matrix form given by Coulomb operator. To optimize performance, the
(G, G )
VG, G (q) = 4 |q + G|2 . This divergence is observed in the matrix screening parameter γ and the width parameter Rω in eq
53 can be adjusted. In this work, we utilized the values that
element where both indices, G and G′, are equal to 0. This had been optimized and reported in previous work on
particular element is often referred to as the “head” term of the hybrid XC functionals72 and G0W025 implementations.
matrix with indices G and G′. In physical terms, this term reflects 2. Proper Treatment of Symmetrized Dielectric Function
the interaction energy of an infinitely extended, periodic array of In the calculation of the screened Coulomb interaction
charges of the same sign, which is infinite even per unit cell. In given by eq 39, the bare Coulomb interaction enters in the
the atom-centered ABF representation, this divergence carries numerator through V1/2(q) and the denominator through
over to the matrix elements between two nodeless s-type the dielectric function ε(q). In the numerator, replacing
functions, resulting in a 1/q2 divergence. Similarly, between one the bare V by its truncated counterpart Vcut works very
nodeless s-type and one nodeless p-type function, it leads to a 1/ well, but doing so for the dielectric function ε is not a good
q divergence.25 The analytical form of the Coulomb operator strategy as the screening property is not properly
can be expressed as described. Therefore, following ref 25, we adopted a
mixed scheme where in the numerator the bare Coulomb
v(2) v(1) interaction is truncated as is in the case of exact-exchange
V , (q) = + + V , (q)
q2 q (52) calculations,72 where the full Coulomb operator is used to
calculate the dielectric function. Note that the dielectric
where V̅ μ,ν(q) represents the analytic part of the Coulomb function is regular and finite everywhere in the BZ except
operator as q approaches 0 while v(2) (1)
μν and vμν are the coefficients for q = 0 where the divergence of the bare Coulomb
of the matrix elements exhibiting 1/q2 and 1/q asymptotic operator needs to be analytically treated. This can be done
behaviors, commonly referred to as the “head” and “wing” terms, most conveniently in the basis set representation of the
respectively. This divergence behavior of Vμν(q) also extends to eigenvectors of the Coulomb matrix, as discussed
the screened Coulomb matrix Wμν(q) in nonmetallic systems. previously in the literature. 78−80 Within such a
For addressing this issue, two numerical schemes are well- representation, similar to the plane-wave case, the
known in the context of the Coulomb singularity within periodic divergence in the bare Coulomb potential as q → 0 is
HF calculation. The first scheme, known as the Gygi− canceled by the corresponding asymptotic behavior of χ0,
Baldereschi (GB) scheme,76 incorporates an analytically and the “head” and “wing” terms of the dielectric function
integrable compensating function to eliminate the diverging ε at q = 0 can thus be properly treated. Afterward, one can
term and subtracts it separately. The second scheme, referred to transform ε(q = 0) back to the ABF representation. Via
as the Spencer−Alavi scheme,77 employs a truncated Coulomb such a treatment, the εμν(q) matrix becomes regular
operator that avoids the Coulomb singularity, while ensuring everywhere in the BZ, and can be numerically inverted.
systematic convergence to the correct limit as the number of k- The screened Coulomb matrix calculated via eq 39 is thus
point increases. In this study, we deal with the singularity issue also regular everywhere in the BZ and can now be
associated with both the bare Coulomb and screened Coulomb conveniently employed in setting up the BSE equation.
operators by adopting a similar approach employed for G0W0 We note that, in the present work, only the “head” term of the
method by Ren et al.25 A brief outline is provided here and, one dielectric function at q = 0 within the basis representation of the
can find a more comprehensive procedure in refs 25,72. Our eigenvectors of the Coulomb matrix was explicitly treated, while
approach primarily consists of two main steps: the “wing” term, which has a secondary effect, was left
1. Modified Spencer−Alavi Scheme: untouched. This is because a proper treatment of the “wing”
To address the singularity problem of the bare term was not available when the present research work was
Coulomb operator, a truncated Coulomb operator is conducted. Very recently, a rigorous treatment of the “wing”
term became available in FHI-aims code as part of a separate,
introduced. In this way, the regularity of the bare
ongoing study, and preliminary test calculations show that this
Coulomb matrix as q → 0 can be ensured. This truncated entails a blue shift of the G0W0 band gap by 0.03 eV for Si and
operator, denoted as Vcut(q), is obtained within the 0.13 eV for MgO. Therefore, we do not expect a significant
auxiliary basis by replacing the 1 term with a truncated influence of the “wing” correction on the BSE@GW results
|r r|
297 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

Figure 1. (a) Comparison of the absorption spectrum of silicon (Si) ε2 and (b) absolute error Δε2 in BSE@G0W0 calculations using different auxiliary
numerical atomic orbitals (NAO) basis sets. The OBS+4f5g6h auxiliary NAO basis is used as the reference for the errors in (b). All calculations are
performed with 8 × 8 × 8 Γ-centered BZ sampling and the tier 2 NAO basis set.84

presented in the present work. Further investigation of this issue the LRI approximation, a larger number of ABFs is required than
will be presented in future work. in the nonlocal RI-V scheme, since high angular momentum
components of the pair density to be expanded must be
4. DEMONSTRATION OF NAO IMPLEMENTATION accounted for by ABFs on the same two centers on which the
AND CONVERGENCE orbital basis functions in questions are centered.69 One practical
strategy to create accurate auxiliary basis sets for LRI is to
The new all-electron NAO-based periodic BSE method is supplement the OBS with additional functions of higher angular
implemented within the FHI-aims code.54,81 For the BZ
momenta. It is important to note that these additional orbital
sampling, we employ an even-sampled Γ-centered grid with
basis functions are exclusively employed in the construction of
equally spaced n1 × n2 × n3 k-points. The integration grid for the
ABFs and do not participate in the preceding self-consistent KS-
NAO basis for single-particle matrix elements and other real-
DFT calculations. In the input files of FHI-aims, the
space integrals employs FHI-aims’ “tight” settings. The detailed
supplemental functions to the OBS that generate the extended
choice and convergence of the NAO basis set and auxiliary basis
set are discussed in subsequent subsections. We use crystalline ABF basis set are denoted by the keyword for_aux.
silicon (Si), 2-atom primitive cell, as an example to demonstrate Following the nomenclature introduced in ref 69, we refer to
our implementation, particularly focused on the NAO basis set, the set of OBS plus the additional supplemental functions as the
auxiliary basis set, and the BZ sampling. We also provide a direct enhanced orbital basis set (OBS+). It was previously
comparison of our all-electron NAO results to those obtained demonstrated that for an OBS containing at least up to f
using the traditional PlaneWave Pseudopotential (PW−PP) functions, the inclusion of an extra 5 g hydrogenic function in the
approach implemented in the BerkeleyGW11 code and enhanced orbital basis set (OBS+) can yield an ABF basis set
Quantum Espresso82 code. with sufficient accuracy for various computational methods,
The computational procedure starts with Kohn−Sham (KS) including Hartree−Fock (HF), MP2, and RPA, in molecular
DFT calculations using the local density approximation (LDA) calculations69 Furthermore, based on benchmark calculations of
in the Perdew−Wang (PW) parameterization.83 Subsequently, periodic G0W0, it was found that the inclusion of 4f or even 5g
G0W0 calculations are performed on top of the KS orbitals and hydrogenic functions in the OBS+ is necessary to achieve
energies to obtain the quasi-particle energies. The BSE convergence of the auxiliary basis.25 To assess the convergence
calculation is then performed with the results of the G0W0 of the auxiliary basis for the periodic BSE with the LRI scheme,
calculation. For the frequency-dependent dielectric function, we we compute the absorption spectrum for crystalline silicon (Si)
utilize 80 frequency points in the Pade approximation for using an 8 × 8 × 8 BZ sampling. In defining augmented
analytic continuation. In constructing the BSE Hamiltonian, we hydrogenic functions within the enhanced orbital basis set (OBS
used 4 valence bands and 6 conduction bands. (this choice is for +), an additional parameter, Z is introduced, which represents an
Si and it would need separate validation for other systems). an effective charge for a hydrogen-like generating potential vn(r)
4.1. Convergence of Auxiliary Basis Set. As discussed in in eq 29 and governs the shape and spatial extent of the solution
Section 3.3, our periodic BSE method is implemented based on to the radial Schrödinger equation.54,69 In this work, we set Z = 0
the LRI approximation, thus the accuracy depends on the quality as the default value, resulting in the utilization of spherical Bessel
of the auxiliary basis functions (ABFs). In the FHI-aims code, functions, confined by the confining potential introduced in eq
standard ABFs are generated from the one-electron orbital basis 29, for constructing the auxiliary basis.
set (OBS) employed in the preceding KS calculations, As depicted in Figure 1, we employ the exhaustive OBS
summarized in Figure 1 of Reference.69 To construct the +4f5g6h result as the reference standard and compute the
auxiliary basis from OBS, the radial components are derived relative errors for OBS, OBS+4f, and OBS+4f5g.84 In terms of
directly from the products of the one-electron orbital basis sets, the optical energy gap, given by the lowest BSE eigenvalue, using
and then Gram-Schmidt orthogonalization is applied to the regular OBS to construct ABFs yields the value of 3.083 eV,
eliminate linear dependencies within the auxiliary basis.50 In which is already in excellent agreement with the reference value,
298 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

differing only by 1 meV. However, for the features spanning As shown in Figure 2(a), the optical spectrum converges
from 3 to 8 eV in the optical absorption spectrum, the ABFs quickly in the tier n basis set, with only a marginal shift in
derived from the tier 2 OBS only lead to exhibits a relatively large
error in the absorption peak intensity as seen in Figure 1(b).
Introducing additional OBS+ auxiliary basis functions as in OBS
+4f and OBS+4f5g make the LRI-related error negligible as they
are converged with respect to the reference OBS+4f5g6h result
(see Figure 1(b)). We note that, when comparing the BSE
convergence with the G0W0 convergence test presented in ref 25,
the sensitivity to the auxiliary basis in the BSE calculations is less
pronounced than in the G0W0 calculations. The primary
difference arises from how the screened interaction is handled
differently in BSE and G0W0 calculations. In BSE calculations,
the screened interaction among electron−hole pairs near the
Fermi level is important. In contrast, the self-energy evaluation
in G0W0 calculations involve the screened interaction for higher-
energy orbitals, necessitating a more extensive set of delocalized
basis functions.
To summarize, we here demonstrated the robustness of the
LRI scheme within our periodic BSE framework. To achieve the
full convergence of auxiliary basis set, OBS+4f or more diffuse
spherical Bessel functions with a confinement potential are
found necessary. Thus, in all simulations presented in this work,
OBS+4f is employed as the default setting for the auxiliary basis.
4.2. Convergence of NAO Basis Set. In terms of NAO
basis sets convergence, very high precision can be reached in all-
electron ground-state density functional theory (DFT) calcu-
lations when only occupied KS states need to be evaluated.54,85
However, similar to the case of Gaussian Type Orbitals (GTOs)
and other atom-centered basis sets, larger basis sets may be
needed when applied to correlated calculations such as MP2 and
RPA. The standard FHI-aims-2009 NAO basis set series (“tier
n” basis), while originally designed for ground-state DFT
calculations, can actually produce acceptable results also for
molecular MP2 and RPA calculation when counterpoise
corrections are employed.50 Enhanced accuracy can be achieved
by employing the alternative, so-called valence-correlation
consistent (VCC) NAO-VCC-nZ basis sets,86 which allow for
results to be extrapolated to the complete basis set (CBS) limit
through a two-point extrapolation procedure.74 In densely
packed solids, however, it was observed that the original NAO- Figure 2. Convergence of the absorption spectrum of Si from BSE@
VCC-nZ basis sets, initially designed for molecules, lead to G0W0 calculation with respect to different NAO basis sets. A Lorentzian
overlap matrices with impractically large condition numbers broadening of η = 0.10 eV is used. The calculations are performed with
compared to the tier n basis sets, and numerical instabilities in 7 × 7 × 7 Γ-centered BZ sampling and (a) tier n (n = 1, 2, 3) and (b)
standard linear algebra (specifically, eigenvalue solutions) can be loc-NAO-VCC-nZ (n = 2, 3, 4) NAO basis sets complemented by the
OBS+4f auxiliary NAO basis set.
a consequence. To address this limitation, Zhang et al.74
optimized these basis sets by eliminating the so-called
“enhanced minimal basis” and tightening the cutoff radius of excitation peaks as we increase the basis set size from tier 1 to tier
the basis functions. The resulting basis sets, referred to as 3. For instance, the energy of the first peak changes by only 0.021
localized NAO-VCC-nZ (loc-NAO-VCC-nZ) here, have been eV from 3.173 to 3.152 eV as we move up from tier 1 to tier 3.
demonstrated to yield accurate MP2 and RPA energies for Conversely, when employing the loc-NAO-VCC-nZ basis set,
simple solids when used in conjunction with an appropriate the absorption spectrum for the 2Z basis set is not fully
extrapolation procedure in ref 74 Separately, in BSE@G0W0 converged as seen in Figure 2(b). There is some qualitatively
benchmarks for molecules, our team observed excellent incorrect behavior as evidenced, for example, in the erroneous
numerical convergence with the tier n basis sets, augmented prediction of multiple peaks around 4 eV as well as the sizable
with two extended Gaussian orbital basis functions from blue-shift of the first excitation at 3.238 eV. These erroneous
Dunning’s augmented correlation-consistent basis sets.26 To features can be attributed to the omission of the “enhanced
assess basis set convergence in the present work, we conducted minimal basis” in the loc-NAO-VCC-2Z basis, rendering it
calculations of the silicon (Si) absorption spectrum using the insufficient for accurately describing the occupied states in the
BSE@G0W0 method with two different sets of basis functions: preceding self-consistent field (SCF) calculations. At the same
(a) tier n (n = 1, 2, 3) and (b) loc-NAO-VCC-nZ (n = 2, 3, 4). time, the larger basis sets in this series, loc-NAO-VCC-3Z and
These basis set convergence tests were performed using a 7 × 7 loc-NAO-VCC-4Z basis sets, yield converged results, and they
× 7 Γ-centered BZ sampling. also agree well with the converged results using the “tier n”
299 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

default basis sets of the FHI-aims code. To summarize, with the

largest NAO-VCC-4Z and tier 3 basis sets, we find precise
agreement within 30 meV for the BSE@G0W0 absorption
spectrum. Furthermore, even the smaller tier 1 and tier 2 NAO
basis sets lead to results that are remarkably well converged. This
behavior is a significant and highly promising success. In
particular, this behavior is qualitatively different from the small-
molecule case,26 in which extended augmentation functions
were shown to be needed, as is standard in molecular quantum
chemistry. We attribute this encouraging difference to the
molecular case to the fact that in solids, there is a significantly
higher density of basis functions that are nonzero at any given
point, compared to finite molecules. The reason is the overlap of
basis functions associated with neighboring unit cells or,
equivalently, the basis functions with different Bloch phase
factors in the BvK cell. In our view, this higher density of
nonzero basis functions lends itself to a more finely resolved
description of the two-particle correlation function than is Figure 3. Absorption spectrum of Si with Γ-centered BZ sampling of
possible in molecules, where describing the two-particle dimensions n × n × n (n = 7, 10, 12, 14) calculated using the tier 2 NAO
response is problematic especially in regions that are relatively basis set and OBS+4f auxiliary NAO basis set. A Lorentzian broadening
distant from the atoms (covered by augmentation functions). In of η = 0.15 eV is used.
summary, it seems that FHI-aims’ standard NAO basis sets can
be used for well converged BSE@GW calculations in solids. In Our current implementation does not exploit space group
the subsequent sections, we employ the default tier 2 basis set of symmetry, k-space interpolation strategies, or similar simplifi-
the FHI-aims code as specified in ref 25. cations. Therefore, our current computational resources do not
4.3. Convergence of Brillouin Zone (BZ) Sampling. One allow us to achieve the full convergence with with a significantly
of the most important and also challenging aspects of calculating denser BZ sampling at this time. We intend to explore explore
the optical absorption spectrum of extended condensed matter incorporating enhanced BZ sampling techniques in our future
systems is achieving the convergence with respect to the BZ research effort, such as coarse-grained k-grid interpolation,11
sampling.11,39,87,88 This is a particularly important consideration Wannier interpolation,87 and other related schemes.39,88
for many inorganic solids in which there exist strong band 4.4. Comparison with Planewave Basis-Set Result.
dispersion and the band gap is indirect. To accelerate the Having examined the convergence behavior of the BSE
convergence of the absorption spectrum in BSE calculations, the calculation for our all-electron NAO based implementation,
BZ sampling with a randomly shifted k-grid around the Γ point we now turn to the comparison to the BSE calculation based on
has been used for practical calculations as discussed in the the traditional plane-wave pseudopotential (PW−PP) imple-
literature.30,42,89−91 For the purpose of benchmarking our new mentation. We performed the PW−PP based BSE@GW
NAO-based BSE implementation, we do not consider such calculation using the well-documented BerkeleyGW package.11
accelerated k-point sampling techniques in this work, but we The BSE@GW calculation was performed on top of the DFT
focus on the convergence behavior of the direct equally spaced calculation using the Quantum Espresso92,93 code.
BZ sampling centered at Γ point. To generate quasi-particle To benchmark our result with BerkeleyGW, we first
(QP) energies for BSE calculation, we employ G0W0 calculation performed a one-shot G0W0 calculation to obtain quasi-particle
for all k-grids instead of applying a consistent scissors shift to energies. The frequency dependence of the inverse dielectric
Kohn−Sham orbital energies as is sometimes done in the matrix was described using full frequency integration based on
literature. In the previous benchmark on the NAO-based G0W0 the Contour-deformation formalism (Adler−Wiser formula).65
method,25 it was found that 7 × 7 × 7 BZ sampling for the We used 80 points for the numerical integration and a 50 Ry
dielectric matrix is adequate to achieve convergence of QP cutoff for the plane wave expansion of the dielectric matrix. Both
energies within 5 meV. the static RPA polarizability and the Coulomb-hole self-energy
Figure 3 shows the optical absorption spectrum of crystalline term included 200 bands. Then, for the BSE calculation, the
silicon, utilizing Γ-centered uniform BZ sampling with n × n × n electron−hole interaction kernel was constructed on a 10 × 10 ×
where n = 7, 10, 12, 14. Neither the shape of the absorption 10 k-point grid and interpolated to a finer 14 × 14 × 14 grid. We
spectrum nor the excitation energies are converged when included 4 valence bands and 14 conduction bands for the
employing the 7 × 7 × 7 BZ sampling, despite the convergence coarse grid, and 4 valence bands and 4 conduction bands for the
observed for the G0W0 QP energies.25 With increased BZ fine grid.
samplings, the absorption spectrum generally becomes blue- In Figure 4(a), we present the optical absorption spectrum of
shifted, particularly for the first absorption peak. At the same crystalline silicon, using a 14 × 14 × 14 Γ-centered BZ sampling.
time, the peak at around 5.0 eV is seen to red-shift, and the Both spectra use the same Lorentzian broadening of 0.15 eV.
spectrum does not converge uniformly at all energies. In general, Note that Contour Deformation (not the widely used
the shape of the absorption spectrum tends to converge with the generalized plasmon-pole model)94 technique is employed
increased BZ sampling from n = 10 to n = 14. At the same time, here for the numerical (frequency) integration of the self-energy
we note that achieving the complete convergence of the in the G0W0 calculation with BerkeleyGW, in order to compare
absorption spectrum for cystalline silicon, as discussed in ref 39, the two implementations on a similar footing. Our NAO-based
may necessitate an exceedingly fine sampling of the BZ to the BSE@G0W0 results closely match the PW−PP result in terms of
value of n = 40. the peak positions while some differences in the amplitude are
300 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

Figure 4. Comparison of the (a) optical absorption spectra and (b) joint density of state (JDOS) divided by ω2 for Si between our NAO
implementation in FHI-aims code and the PW−PP implementation in BerkeleyGW code, performed with a 14 × 14 × 14 Γ-centered BZ sampling. The
FHI-aims calculation uses the tier 2 NAO basis set and OBS+4f auxiliary NAO basis set. A Lorentzian broadening of η = 0.15 eV is used.

Figure 5. Optical absorption spectrum of MgO from BSE@G0W0 calculations. A Lorentzian broadening of η = 0.30 eV is used in all cases. (a) Optical
absorption spectrum of MgO with Γ-centered BZ sampling with dimensions n × n × n (n = 11, 13, 15). Calculations are performed using tier 2 NAO
basis set complemented by the OBS+4f auxiliary NAO basis set. (b) Comparison of MgO absorption spectra from our NAO implementation in FHI-
aims code with the result from VASP package (PW+PAW) using the 15 × 15 × 15 Γ-centered BZ sampling. The result from the Exciting code (LAPW
+lo) with a 11 × 11 × 11 randomly shifted BZ sampling is also compared. Both the PW+PAW and LAPW+lo results are taken from ref 101.

observed at around 4.0−5.0 eV. We note that HOMO−LUMO is observed. This rather small deviation in this energy range has a
QP energy gap values from the G0W0 calculations agree very significant impact on the differences observed in the absorption
closely between our NAO-based result and the PW−PP result, spectrum as seen in Figure 4(a).
within 6 meV. BSE eigenvalues (i.e., excitation energies) from As discussed above regarding basis set convergence, the tier 2
our NAO-based calculation also closely match with the PW−PP basis is sufficient to converge the optical spectrum at a 7 × 7 × 7
results, exhibiting an average difference of 28 meV. BZ sampling. The difference between tier 3 and tier 2 is almost
To better understand the origin of the observed differences in negligible (see Figure 2), indicating that the basis set is not the
the optical absorption spectrum, Figure 4(b) shows the
primary issue here. While extrapolation techniques have been
comparison using the scaled joint density of states (JDOS)
used for calculating molecular excited states95,96 and exciton97 in
16 2e 2 atom-centered orbital basis, they have not been widely applied
2
JDOS ( )/ = 2
( m) to the calculation of optical spectra for extended systems.
m (54) Given the fully converged basis set for this particular case of
crystalline silicon as discussed above, these differences could be
where ωm is the energy of excited state m from BSE calculation.
Instead of having the transition amplitudes convoluted as part of attributed to the difference between the two implementations,
the calculation of the absorption spectrum, the JDOS allows for e.g., due to the difference in the frequency integration treatments
a direct comparison of electronic transitions as a function of the utilized in the G0W0 calculation. In our benchmark calculation,
excitation energy. As seen in Figure 4(b), our NAO-based BSE the contour deformation technique is employed for the
excitation energies closely align with the PW−PP results, except numerical (frequency) integration of the self-energy in the
in the energy range of 4.0−4.5 eV, where a noticeable deviation G0W0 calculation using BerkeleyGW code, whereas Pade
301 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

approximation is utilized in FHI-aims for analytic continu- based on the PW+PAW scheme,106,107 and the Exciting code is
ation.98 based on LAPW+lo scheme.108 Lorentzian broadening param-
eter of 0.30 eV was used for the optical absorption spectrum in
5. APPLICATION TO EXCITONS WITH LARGE BINDING all calculations. The VASP (PP+PAW) result was obtained using
ENERGY: MGO the BZ sampling of 15 × 15 × 15 centered at the Γ point while
The BSE formalism explicitly solves for the two-particle the Exciting (LAPW+lo) result was obtained with a randomly
correlation function, making it an ideal approach for studying shifted 11 × 11 × 11 k-grids centered at (0.09, 0.02, 0.04) of the
extended periodic systems with strongly excitonic character. In BZ. For the lowest excited state, we obtain a large value of 385
this section, we compare our all-electron NAO-based meV for the exciton binding energy (i.e., the difference between
implementation of BSE@GW with other BSE@GW calculations the excited energy and the quasi-particle energy gap), which is a
reported in literature for crystalline magnesium oxide (MgO). good agreement with the reported value of 442 meV using the
MgO is one of the extensively investigated oxides.99,100 VASP code and 435 meV using the exciting code.101
Developing a comprehensive understanding of its spectroscopic Figure 5(b) shows the optical absorption spectrum for our
features from first-principles necessitates an accurate modeling NAO-based implementation, along with the results using the
of electronic excitation of this extended material with a large VASP (PP+PAW) and Exciting (LAPW+lo) codes as reported
exciton binding energy.101 MgO also serves as an ideal in ref 101. The VASP result and the Exciting result agree well
benchmark system for studying optical properties using many- above 10 eV but they show noticeable differences between 8 and
body perturbation theory, and excellent agreement with 10 eV. Using the exactly the same BZ sampling, our NAO-based
experimental measurement has been reported.101−103 We result and the VASP (PP+PAW) result agree quite well in terms
compare our all-electron NAO implementation of BSE@GW of the peak positions overall. At the same time, a variation
with two other corresponding calculations based on the between our result and the reported VASP/Exciting results is
planewaves with the projector-augmented-waves method (PW observed for the amplitudes of some prominent peaks. To
+PAW) and linearized augmented planewave + local orbital investigate this difference, we examined the optical absorption
(LAPW+lo) formalism reported in ref 101. spectrum within the independent-particle (IP) approximation.
To make a direct comparison, we employ the same Unlike the BSE formalism described in eq 26, the IP absorption
computational settings for XC functional, the lattice structure, spectrum is directly calculated by enumerating all excitation
as well as the BSE active space while BZ sampling varies pairs between the valence band and the conduction band of the
somewhat for the LAPW+lo calculation reported in ref 101. The Kohn−Sham equations
DFT-KS calculation is used as the starting point, employing the 16 2e 2
PBEsol Generalized Gradient Approximation (GGA).104,105 2( )= 2
e· vk|v |ck |2 ( ck + vk )
Within this XC functional, the DFT-optimized equilibrium vck
lattice constant is 4.21 Å.101 To compute QP energies, a single- (55)
shot G0W0 calculation was performed, utilizing a 7 × 7 × 7 Γ- where v̂ is the velocity operator and e is the direction of the
centered BZ sampling for the dielectric matrix in the self-energy polarization of light. As seen in Figure 6, the absorption spectra
calculation. The analytical continuation with the Pade within the IP approximation using KS states agree well, with
approximation with 100 parameters was performed for GW some residual discrepancies between our result and the reported
self-energy on the imaginary axis. For the BSE calculation, we result101 in the peak intensities, similar to what is seen in the
employed the n × n × n Γ-centered BZ sampling with n = 11, 13, BSE@GW calculation (Figure 5(b)).
15. In constructing the BSE Hamiltonian, 4 valence bands and 5
conduction bands are included. Here, all calculations were
carried out using the tier 2 NAO basis set54 and the OBS+4f
auxiliary basis set25,69 for both G0W0 and BSE calculations. In
Figure 5(a), we present the optical absorption spectrum with
different n for the BZ sampling. Similarly to the case of the
crystalline silicon, the shape and the relative peak heights of the
absorption spectrum are sensitive to the BZ sampling even
though the dielectric function is converged in the G0W0
calculation already with 7 × 7 × 7 Γ-centered BZ sampling
(see Figure S1 in Supporting Information (SI)). As we increase
the number of the BZ sampling points, the contribution from
high-symmetry k-points decreases for calculating the absorption
spectrum. Consequently, we observe the increased splitting of
the absorption peaks in the energy range of 8−11 eV.
Nevertheless, the absorption spectrum tends to converge as
the BZ sampling grid reaches 15 × 15 × 15, which is the largest
grid size used here due to the computational cost and memory,
and this is also the most dense BZ sampling reported in the
literature.101 Figure 6. Comparison of absorption spectrum of MgO within
Lastly, we compare the BSE@GW result from our NAO independent particle (IP) approximation from DFT Kohn−Sham
implementation to the results reported in the literature using the eigenvalues between FHI-aims calculation (this work) and VASP ref
planewave-based formulations as shown in Figure 5(b). In 101. FHI-aims calculations are performed under tier 2 NAO basis set
particularly, we compare to the results reported using the VASP complemented by the OBS+4f auxiliary NAO basis set and Γ-centered
code and the Exciting code from ref 101. The VASP code is 15 × 15 × 15 BZ sampling.

302 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

As noted earlier, both the valence states and the low-lying

unoccupied states of MgO using FHI-aims NAO “tier 2” basis
■ AUTHOR INFORMATION
Corresponding Authors
sets are fully converged in KS-DFT calculations52 and, for G0W0
calculations, are converged to approximately 0.2 eV or better.25 Xinguo Ren − Institute of Physics, Chinese Academy of Sciences,
These factors would not contribute to the discrepancies seen Beijing 100190, China; orcid.org/0000-0002-3360-2281;
between our results and those reported in the referenced Email: [email protected]
literature.101 Therefore, the small residual differences in the Yosuke Kanai − Department of Chemistry, University of North
BSE@GW calculations can be partially ascribed to the Carolina at Chapel Hill, Chapel Hill, North Carolina 27599,
differences in the high-lying unoccupied states and the United States; Department of Physics and Astronomy,
corresponding transition matrix elements on which BSE@GW University of North Carolina at Chapel Hill, Chapel Hill,
calculation method relies. North Carolina 27599, United States; orcid.org/0000-
0002-2320-4394; Email: [email protected]
6. CONCLUSIONS Authors
We described the formulation and algorithms of a new all- Ruiyi Zhou − Department of Chemistry, University of North
electron periodic BSE implementation within the numeric atom- Carolina at Chapel Hill, Chapel Hill, North Carolina 27599,
centered orbital (NAO) framework. To our knowledge, this is United States; orcid.org/0000-0002-7732-6338
the first all-electron NAO-based BSE implementation that works Yi Yao − Department of Chemistry, University of North
with periodic boundary conditions. Our implementation was Carolina at Chapel Hill, Chapel Hill, North Carolina 27599,
carried out within the FHI-aims code package.54,81 We use United States; Thomas Lord Department of Mechanical
computed absorption spectra for crystalline silicon (Si) as an Engineering and Materials Science, Duke University, Durham,
example to demonstrate our implementation and performed North Carolina 27708, United States
systematic convergence tests with respect to the computational Volker Blum − Thomas Lord Department of Mechanical
parameters including the NAO basis set size, auxiliary basis set, Engineering and Materials Science, Duke University, Durham,
and Brillouin zone sampling. With the fully converged result in North Carolina 27708, United States; Department of
hand, we make a direct comparison of our all-electron NAO Chemistry, Duke University, Durham, North Carolina 27708,
result to absorption spectra computed using the traditional United States; orcid.org/0000-0001-8660-7230
PW−PP approach implemented in the BerkeleyGW11 code with Complete contact information is available at:
Quantum Espresso82 code. Having established the excellent https://pubs.acs.org/10.1021/acs.jctc.4c01245
agreement with the well-established implementation of BSE@
GW based on the PW−PP formalism, we also performed Author Contributions
calculations on crystalline MgO with a large exciton binding #
R.Z. and Y.Y. led and contributed equally to the work. All
energy. We compare our NAO-based BSE@GW calculation to authors discussed the results and contributed to the final
available BSE@GW results based on PP+PAW and LAPW+lo manuscript.
formulations from the literature. The calculations show good
agreement for the optical absorption spectra, again demonstrat- Notes
ing the accuracy of our approach. However, achieving complete The authors declare no competing financial interest.
convergence of the optical absorption spectrum may also
necessitate an exceedingly fine sampling of the BZ as widely
discussed in the literature.39 Our current implementation
■ ACKNOWLEDGMENTS
We thank the Research Computing at the University of North
supports the standard Γ-centered sampling. In future research, Carolina at Chapel Hill for computer resources. We thank
we plan to explore incorporating enhanced BZ sampling Jianhang Xu, Minye Zhang, and Tianyu Zhu for helpful
techniques for BSE calculation, such as coarse-grained k-grid discussion.
interpolation,11 Wannier interpolation,87 and other related
methods.39,88 As mentioned in the introduction, another
important future opportunity is that the groundwork for
periodic BSE@GW laid out in the present work should be
■ REFERENCES
(1) van Setten, M. J.; Weigend, F.; Evers, F. The GW-method for
extendable to much larger systems, by using optimization quantum chemistry applications: Theory and implementation. J. Chem.
strategies similar to those that recently enabled accurate periodic Theory Comput. 2013, 9, 232−246.
exact exchange and hybrid DFT beyond 10,000 atoms using the (2) Leng, X.; Jin, F.; Wei, M.; Ma, Y. GW method and Bethe−Salpeter
same underlying NAO-based framework.27 equation for calculating electronic excitations. WIREs Comput. Mol. Sci.
2016, 6, 532−550.

■ ASSOCIATED CONTENT
Data Availability Statement
(3) Golze, D.; Dvorak, M.; Rinke, P. The GW compendium: A
practical guide to theoretical photoemission spectroscopy. Front. Chem.
2019, 7, No. 377.
(4) Blase, X.; Duchemin, I.; Jacquemin, D.; Loos, P.-F. The Bethe−
The data that support the findings of this study are available Salpeter equation formalism: From physics to chemistry. J. Phys. Chem.
from the corresponding author upon reasonable request. Lett. 2020, 11, 7371−7382.
*
sı Supporting Information (5) Strinati, G. Effects of dynamical screening on resonances at inner-
shell thresholds in semiconductors. Phys. Rev. B 1984, 29, No. 5718.
The Supporting Information is available free of charge at (6) Albrecht, S.; Reining, L.; Del Sole, R.; Onida, G. Ab initio
https://pubs.acs.org/doi/10.1021/acs.jctc.4c01245. calculation of excitonic effects in the optical spectra of semiconductors.
Phys. Rev. Lett. 1998, 80, No. 4510.
Convergence test of Brillouin zone sampling in G0W0 (7) Rohlfing, M.; Louie, S. G. Electron-hole excitations in semi-
dielectric function (PDF) conductors and insulators. Phys. Rev. Lett. 1998, 81, No. 2312.

303 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

(8) Onida, G.; Reining, L.; Rubio, A. Electronic excitations: density- (30) Rohlfing, M.; Louie, S. G. Electron-hole excitations and optical
functional versus many-body Green’s-function approaches. Rev. Mod. spectra from first principles. Phys. Rev. B 2000, 62, No. 4927.
Phys. 2002, 74, No. 601. (31) Loos, P.-F.; Blase, X. Dynamical correction to the Bethe−
(9) Marsili, M.; Mosconi, E.; De Angelis, F.; Umari, P. Large-scale G Salpeter equation beyond the plasmon-pole approximation. J. Chem.
W-BSE calculations with N 3 scaling: Excitonic effects in dye-sensitized Phys. 2020, 153, No. 114120, DOI: 10.1063/5.0023168.
solar cells. Phys. Rev. B 2017, 95, No. 075415. (32) Zhang, X.; Leveillee, J. A.; Schleife, A. Effect of dynamical
(10) Vorwerk, C.; Aurich, B.; Cocchi, C.; Draxl, C. Bethe−Salpeter screening in the Bethe-Salpeter framework: Excitons in crystalline
equation for absorption and scattering spectroscopy: implementation naphthalene, arXiv:2302.07948. arXiv.org e-Print archive, 2023
in the exciting code. Electron. Struct. 2019, 1, No. 037001. https://arxiv.org/abs/2302.07948.
(11) Deslippe, J.; Samsonidze, G.; Strubbe, D. A.; Jain, M.; Cohen, M. (33) Martin, R. M.; Reining, L.; Ceperley, D. M. Interacting Electrons;
L.; Louie, S. G. BerkeleyGW: A massively parallel computer package for Cambridge University Press, 2016; pp 345−388.
the calculation of the quasiparticle and optical properties of materials (34) Shao, M.; Yang, C. BSEPACK User’s Guide, arXiv:1612.07848.
and nanostructures. Comput. Phys. Commun. 2012, 183, 1269−1289. arXiv.org e-Print archive, 2016. https://arxiv.org/abs/1612.07848.
(12) Faber, C.; Boulanger, P.; Attaccalite, C.; Duchemin, I.; Blase, X. (35) Benner, P.; Dolgov, S.; Khoromskaia, V.; Khoromskij, B. N. Fast
Excited states properties of organic molecules: From density functional iterative solution of the Bethe−Salpeter eigenvalue problem using low-
theory to the GW and Bethe−Salpeter Green’s function formalisms. rank and QTT tensor approximation. J. Comput. Phys. 2017, 334, 221−
Philos. Trans. R. Soc., A 2014, 372, No. 20130271. 239.
(13) Jacquemin, D.; Duchemin, I.; Blase, X. Is the Bethe−Salpeter (36) Benner, P.; Marek, A.; Penke, C. Improving the Performance of
formalism accurate for excitation energies? Comparisons with TD- Numerical Algorithms for the Bethe-Salpeter Eigenvalue Problem.
DFT, CASPT2, and EOM-CCSD. J. Phys. Chem. Lett. 2017, 8, 1524− PAMM 2018, 18, No. e201800255.
1529. (37) Ljungberg, M. P.; Koval, P.; Ferrari, F.; Foerster, D.; Sanchez-
(14) Blase, X.; Duchemin, I.; Jacquemin, D. The Bethe−Salpeter Portal, D. Cubic-scaling iterative solution of the Bethe-Salpeter
equation in chemistry: relations with TD-DFT, applications and equation for finite systems. Phys. Rev. B 2015, 92, No. 075422.
challenges. Chem. Soc. Rev. 2018, 47, 1022−1043. (38) Rocca, D.; Vörös, M.; Gali, A.; Galli, G. Ab initio optoelectronic
(15) Casida, M. E. Recent Advances In Density Functional Methods: properties of silicon nanoparticles: Excitation energies, sum rules, and
(Part I); World Scientific, 1995; pp 155−192. Tamm-Dancoff approximation. J. Chem. Theory Comput. 2014, 10,
(16) Ullrich, C. A. Time-Dependent Density-Functional Theory: 3290−3298.
Concepts and Applications; OUP: Oxford, 2011; pp 124−132. (39) Sander, T.; Maggio, E.; Kresse, G. Beyond the Tamm-Dancoff
(17) Bartlett, R. J. Coupled-cluster theory and its equation-of-motion approximation for extended systems using exact diagonalization. Phys.
extensions. WIREs Comput. Mol. Sci. 2012, 2, 126−138. Rev. B 2015, 92, No. 045209.
(18) Krylov, A. I. Equation-of-motion coupled-cluster methods for (40) Hirata, S.; Head-Gordon, M. Time-dependent density functional
theory within the Tamm-Dancoff approximation. Chem. Phys. Lett.
open-shell and electronically excited species: The hitchhiker’s guide to
1999, 314, 291−299.
Fock space. Annu. Rev. Phys. Chem. 2008, 59, 433−462.
(41) Chantzis, A.; Laurent, A. D.; Adamo, C.; Jacquemin, D. Is the
(19) Yang, L.; Cohen, M. L.; Louie, S. G. Excitonic effects in the
Tamm-Dancoff approximation reliable for the calculation of absorption
optical spectra of graphene nanoribbons. Nano Lett. 2007, 7, 3112−
and fluorescence band shapes? J. Chem. Theory Comput. 2013, 9, 4517−
3115.
4525.
(20) Perebeinos, V.; Tersoff, J.; Avouris, P. Radiative lifetime of
(42) Rocca, D.; Ping, Y.; Gebauer, R.; Galli, G. Solution of the Bethe-
excitons in carbon nanotubes. Nano Lett. 2005, 5, 2495−2499.
Salpeter equation without empty electronic states: Application to the
(21) Jacquemin, D.; Duchemin, I.; Blase, X. Benchmarking the
absorption spectra of bulk systems. Phys. Rev. B 2012, 85, No. 045116.
Bethe−Salpeter formalism on a standard organic molecular set. J. Chem. (43) Blase, X.; Attaccalite, C.; Olevano, V.; First-principles, G. W.
Theory Comput. 2015, 11, 3290−3304. calculations for fullerenes, porphyrins, phtalocyanine, and other
(22) Jacquemin, D.; Duchemin, I.; Blondel, A.; Blase, X. Assessment of molecules of interest for organic photovoltaic applications. Phys. Rev.
the accuracy of the Bethe−Salpeter (BSE/GW) oscillator strengths. J. B 2011, 83, No. 115103.
Chem. Theory Comput. 2016, 12, 3969−3981. (44) Faber, C.; Duchemin, I.; Deutsch, T.; Attaccalite, C.; Olevano,
(23) Körbel, S.; Boulanger, P.; Duchemin, I.; Blase, X.; Marques, M. V.; Blase, X. Electron-phonon coupling and charge-transfer excitations
A.; Botti, S. Benchmark many-body GW and Bethe−Salpeter in organic systems from many-body perturbation theory: The Fiesta
calculations for small transition metal molecules. J. Chem. Theory code, an efficient Gaussian-basis implementation of the GW and
Comput. 2014, 10, 3934−3943. Bethe−Salpeter formalisms. J. Mater. Sci. 2012, 47, 7472−7481.
(24) Yao, Y.; Golze, D.; Rinke, P.; Blum, V.; Kanai, Y. All-electron (45) Wilhelm, J.; Del Ben, M.; Hutter, J. GW in the Gaussian and
BSE@GW method for K-edge core electron excitation energies. J. plane waves scheme with application to linear acenes. J. Chem. Theory
Chem. Theory Comput. 2022, 18, 1569−1583. Comput. 2016, 12, 3623−3635.
(25) Ren, X.; Merz, F.; Jiang, H.; Yao, Y.; Rampp, M.; Lederer, H.; (46) Wilhelm, J.; Hutter, J. Periodic G W calculations in the Gaussian
Blum, V.; Scheffler, M. All-electron periodic G 0 W 0 implementation and plane-waves scheme. Phys. Rev. B 2017, 95, No. 235123.
with numerical atomic orbital basis functions: Algorithm and (47) Bruneval, F.; Rangel, T.; Hamed, S. M.; Shao, M.; Yang, C.;
benchmarks. Phys. Rev. Mater. 2021, 5, No. 013807. Neaton, J. B. molgw 1: Many-body perturbation theory software for
(26) Liu, C.; Kloppenburg, J.; Yao, Y.; Ren, X.; Appel, H.; Kanai, Y.; atoms, molecules, and clusters. Comput. Phys. Commun. 2016, 208,
Blum, V. All-electron ab initio Bethe-Salpeter equation approach to 149−161.
neutral excitations in molecules with numeric atom-centered orbitals. J. (48) Zhu, T.; Chan, G. K.-L. All-electron Gaussian-based G 0 W 0 for
Chem. Phys. 2020, 152, No. 044105, DOI: 10.1063/1.5123290. valence and core excitation energies of periodic systems. J. Chem.
(27) Kokott, S.; Merz, F.; Yao, Y.; Carbogno, C.; Rossi, M.; Havu, V.; Theory Comput. 2021, 17, 727−741.
Rampp, M.; Scheffler, M.; Blum, V. Efficient all-electron hybrid density (49) Lei, J.; Zhu, T. Gaussian-based quasiparticle self-consistent GW
functionals for atomistic simulations beyond 10,000 atoms. J. Chem. for periodic systems. J. Chem. Phys. 2022, 157, No. 214114,
Phys. 2024, 161, No. 024112. DOI: 10.1063/5.0125756.
(28) Nakanishi, N. A general survey of the theory of the Bethe- (50) Ren, X.; Rinke, P.; Blum, V.; Wieferink, J.; Tkatchenko, A.;
Salpeter equation. Prog. Theor. Phys. Suppl. 1969, 43, 1−81. Sanfilippo, A.; Reuter, K.; Scheffler, M. Resolution-of-identity approach
(29) Hedin, L. New method for calculating the one-particle Green’s to Hartree-Fock, hybrid density functionals, RPA, MP2 and GW with
function with application to the electron-gas problem. Phys. Rev. 1965, numeric atom-centered orbital basis functions. New J. Phys. 2012, 14,
139, No. A796. No. 053020.

304 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

(51) Bruneval, F.; Hamed, S. M.; Neaton, J. B. A systematic (73) Lin, P.; Ren, X.; He, L. Accuracy of localized resolution of the
benchmark of the ab initio Bethe-Salpeter equation approach for low- identity in periodic hybrid functional calculations with numerical
lying optical excitations of small organic molecules. J. Chem. Phys. 2015, atomic orbitals. J. Phys. Chem. Lett. 2020, 11, 3082−3088.
142, No. 244101, DOI: 10.1063/1.4922489. (74) Zhang, I. Y.; Logsdail, A. J.; Ren, X.; Levchenko, S. V.;
(52) Huhn, W. P.; Blum, V. One-hundred-three compound band- Ghiringhelli, L.; Scheffler, M. Main-group test set for materials science
structure benchmark of post-self-consistent spin-orbit coupling treat- and engineering with user-friendly graphical tools for error analysis:
ments in density functional theory. Phys. Rev. Mater. 2017, 1, systematic benchmark of the numerical and intrinsic errors in state-of-
No. 033803. the-art electronic-structure approximations. New J. Phys. 2019, 21,
(53) Papajak, E.; Zheng, J.; Xu, X.; Leverentz, H. R.; Truhlar, D. G. No. 013025.
Perspectives on basis sets beautiful: Seasonal plantings of diffuse basis (75) Shi, R.; Lin, P.; Zhang, M.-Y.; He, L.; Ren, X. Subquadratic-
functions. J. Chem. Theory Comput. 2011, 7, 3027−3034. scaling real-space random phase approximation correlation energy
(54) Blum, V.; Gehrke, R.; Hanke, F.; Havu, P.; Havu, V.; Ren, X.; calculations for periodic systems with numerical atomic orbitals. Phys.
Reuter, K.; Scheffler, M. Ab initio molecular simulations with numeric Rev. B 2024, 109, No. 035103.
atom-centered orbitals. Comput. Phys. Commun. 2009, 180, 2175− (76) Gygi, F.; Baldereschi, A. Self-consistent Hartree-Fock and
2196. screened-exchange calculations in solids: Application to silicon. Phys.
(55) Whitten, J. L. Coulombic potential energy integrals and Rev. B 1986, 34, No. 4405.
approximations. J. Chem. Phys. 1973, 58, 4496−4501. (77) Spencer, J.; Alavi, A. Efficient calculation of the exact exchange
(56) Dunlap, B. I.; Connolly, J.; Sabin, J. On some approximations in energy in periodic systems using a truncated Coulomb potential. Phys.
applications of X α theory. J. Chem. Phys. 1979, 71, 3396−3402. Rev. B 2008, 77, No. 193110.
(57) Vahtras, O.; Almlöf, J.; Feyereisen, M. Integral approximations (78) Friedrich, C.; Schindlmayr, A.; Blügel, S. Efficient calculation of
for LCAO-SCF calculations. Chem. Phys. Lett. 1993, 213, 514−518. the Coulomb matrix and its expansion around k = 0 within the FLAPW
(58) Weigend, F. A fully direct RI-HF algorithm: Implementation, method. Comput. Phys. Commun. 2009, 180, 347−359.
optimized auxiliary basis sets, demonstration of accuracy and efficiency. (79) Friedrich, C.; Blügel, S.; Schindlmayr, A. Efficient implementa-
Phys. Chem. Chem. Phys. 2002, 4, 4285−4291. tion of the G W approximation within the all-electron FLAPW method.
(59) Feyereisen, M.; Fitzgerald, G.; Komornicki, A. Use of Phys. Rev. B 2010, 81, No. 125102.
approximate integrals in ab initio theory. An application in MP2 (80) Jiang, H.; Gómez-Abal, R. I.; Li, X.-Z.; Meisenbichler, C.;
energy calculations. Chem. Phys. Lett. 1993, 208, 359−363. Ambrosch-Draxl, C.; Scheffler, M. FHI-gap: A GW code based on the
(60) Weigend, F.; Häser, M.; Patzelt, H.; Ahlrichs, R. RI-MP2: all-electron augmented plane wave method. Comput. Phys. Commun.
optimized auxiliary basis sets and demonstration of efficiency. Chem. 2013, 184, 348−366.
(81) Blum, V.; Rossi, M.; Kokott, S.; Scheffler, M. The FHI-aims
Phys. Lett. 1998, 294, 143−152.
(61) Hättig, C. Geometry optimizations with the coupled-cluster Code: All-electron, ab initio materials simulations towards the exascale,
arXiv:2208.12335. arXiv.org e-Print archive, 2022 https://arxiv.org/
model CC2 using the resolution-of-the-identity approximation. J.
abs/2208.12335.
Chem. Phys. 2003, 118, 7751−7761.
(82) Giannozzi, P.; Baseggio, O.; Bonfà, P.; Brunato, D.; Car, R.;
(62) Schütz, M.; Manby, F. R. Linear scaling local coupled cluster
Carnimeo, I.; Cavazzoni, C.; De Gironcoli, S.; Delugas, P.; Ferrari
theory with density fitting Part I: 4-external integrals. Phys. Chem. Chem.
Ruffino, F.; et al. Quantum ESPRESSO toward the exascale. J. Chem.
Phys. 2003, 5, 3349−3358.
Phys. 2020, 152, No. 154105, DOI: 10.1063/5.0005082.
(63) Sierka, M.; Hogekamp, A.; Ahlrichs, R. Fast evaluation of the
(83) Perdew, J. P.; Wang, Y. Accurate and simple analytic
Coulomb potential for electron densities using multipole accelerated
representation of the electron-gas correlation energy. Phys. Rev. B
resolution of identity approximation. J. Chem. Phys. 2003, 118, 9136− 1992, 45, No. 13244.
9148. (84) Knuth, F.; Carbogno, C.; Atalla, V.; Blum, V.; Scheffler, M. All-
(64) Bloch, F. Ü ber die quantenmechanik der elektronen in electron formalism for total energy strain derivatives and stress tensor
kristallgittern. Z. Phys. 1929, 52, 555−600. components for numeric atom-centered orbitals. Comput. Phys.
(65) Adler, S. L. Quantum theory of the dielectric constant in real Commun. 2015, 190, 33−50.
solids. Phys. Rev. 1962, 126, No. 413. (85) Jensen, S. R.; Saha, S.; Flores-Livas, J. A.; Huhn, W.; Blum, V.;
(66) Wiser, N. Dielectric constant with local field effects included. Goedecker, S.; Frediani, L. The elephant in the room of density
Phys. Rev. 1963, 129, No. 62. functional theory calculations. J. Phys. Chem. Lett. 2017, 8, 1449−1457.
(67) Sun, Q.; Berkelbach, T. C.; McClain, J. D.; Chan, G. K. Gaussian (86) Zhang, I. Y.; Ren, X.; Rinke, P.; Blum, V.; Scheffler, M. Numeric
and plane-wave mixed density fitting for periodic systems. J. Chem. Phys. atom-centered-orbital basis sets with valence-correlation consistency
2017, 147, No. 164119, DOI: 10.1063/1.4998644. from H to Ar. New J. Phys. 2013, 15, No. 123033.
(68) Ye, H.-Z.; Berkelbach, T. C. Fast periodic Gaussian density fitting (87) Kammerlander, D.; Botti, S.; Marques, M. A.; Marini, A.;
by range separation. J. Chem. Phys. 2021, 154, No. 131104, Attaccalite, C. Speeding up the solution of the Bethe-Salpeter equation
DOI: 10.1063/5.0046617. by a double-grid method and Wannier interpolation. Phys. Rev. B 2012,
(69) Ihrig, A. C.; Wieferink, J.; Zhang, I. Y.; Ropo, M.; Ren, X.; Rinke, 86, No. 125203.
P.; Scheffler, M.; Blum, V. Accurate localized resolution of identity (88) Gillet, Y.; Giantomassi, M.; Gonze, X. Efficient on-the-fly
approach for linear-scaling hybrid density functionals and for many- interpolation technique for Bethe-Salpeter calculations of optical
body perturbation theory. New J. Phys. 2015, 17, No. 093020. spectra. Comput. Phys. Commun. 2016, 203, 83−93.
(70) Merlot, P.; Kjærgaard, T.; Helgaker, T.; Lindh, R.; Aquilante, F.; (89) Schmidt, W. G.; Glutsch, S.; Hahn, P.; Bechstedt, F. Efficient O
Reine, S.; Pedersen, T. B. Attractive electron-electron interactions (N 2) method to solve the Bethe-Salpeter equation. Phys. Rev. B 2003,
within robust local fitting approximations. J. Comput. Chem. 2013, 34, 67, No. 085307.
1486−1496. (90) Marini, A.; Hogan, C.; Grüning, M.; Varsano, D. Yambo: an ab
(71) Wirz, L. N.; Reine, S. S.; Pedersen, T. B. On resolution-of-the- initio tool for excited state calculations. Comput. Phys. Commun. 2009,
identity electron repulsion integral approximations and variational 180, 1392−1403.
stability. J. Chem. Theory Comput. 2017, 13, 4897−4906. (91) Alvertis, A. M.; Champagne, A.; Del Ben, M.; da Jornada, F. H.;
(72) Levchenko, S. V.; Ren, X.; Wieferink, J.; Johanni, R.; Rinke, P.; Qiu, D. Y.; Filip, M. R.; Neaton, J. B. Importance of nonuniform
Blum, V.; Scheffler, M. Hybrid functionals for large periodic systems in Brillouin zone sampling for ab initio Bethe-Salpeter equation
an all-electron, numeric atom-centered basis framework. Comput. Phys. calculations of exciton binding energies in crystalline solids. Phys. Rev.
Commun. 2015, 192, 60−69. B 2023, 108, No. 235117.

305 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306
Journal of Chemical Theory and Computation pubs.acs.org/JCTC Article

(92) Giannozzi, P.; Baroni, S.; Bonini, N.; Calandra, M.; Car, R.;
Cavazzoni, C.; Ceresoli, D.; Chiarotti, G. L.; Cococcioni, M.; Dabo, I.;
et al. QUANTUM ESPRESSO: a modular and open-source software
project for quantum simulations of materials. J. Phys.: Condens. Matter
2009, 21, No. 395502.
(93) Giannozzi, P.; Andreussi, O.; Brumme, T.; Bunau, O.; Nardelli,
M. B.; Calandra, M.; Car, R.; Cavazzoni, C.; Ceresoli, D.; Cococcioni,
M.; et al. Advanced capabilities for materials modelling with Quantum
ESPRESSO. J. Phys.: Condens. Matter 2017, 29, No. 465901.
(94) Hybertsen, M. S.; Louie, S. G. Electron correlation in
semiconductors and insulators: Band gaps and quasiparticle energies.
Phys. Rev. B 1986, 34, No. 5390.
(95) Zuev, D.; Jagau, T.-C.; Bravaya, K. B.; Epifanovsky, E.; Shao, Y.;
Sundstrom, E.; Head-Gordon, M.; Krylov, A. I. Complex absorbing
potentials within EOM-CC family of methods: Theory, implementa-
tion, and benchmarks. J. Chem. Phys. 2014, 141, No. 024102,
DOI: 10.1063/1.4885056.
(96) Ambroise, M. A.; Jensen, F. Probing basis set requirements for
calculating core ionization and core excitation spectroscopy by the Δ
self-consistent-field approach. J. Chem. Theory Comput. 2019, 15, 325−
337.
(97) Wang, X.; Berkelbach, T. C. Excitons in solids from periodic
equation-of-motion coupled-cluster theory. J. Chem. Theory Comput.
2020, 16, 3095−3103.
(98) Vidberg, H.; Serene, J. Solving the Eliashberg equations by means
of N-point Padé approximants. J. Low Temp. Phys. 1977, 29, 179−192.
(99) Maruyama, T.; Shiota, Y.; Nozaki, T.; Ohta, K.; Toda, N.;
Mizuguchi, M.; Tulapurkar, A. A.; Shinjo, T.; Shiraishi, M.; Mizukami,
S.; Ando, Y.; Suzuki, Y. Large voltage-induced magnetic anisotropy
change in a few atomic layers of iron. Nat. Nanotechnol. 2009, 4, 158−
161.
(100) Yang, S.; Park, H.-K.; Kim, J.-S.; Kim, J.-Y.; Park, B.-G.
Magnetism of ultrathin Fe films on MgO (001). J. Appl. Phys. 2011, 110,
No. 093920, DOI: 10.1063/1.3662179.
(101) Begum, V.; Gruner, M. E.; Vorwerk, C.; Draxl, C.; Pentcheva, R.
Theoretical description of optical and x-ray absorption spectra of MgO
including many-body effects. Phys. Rev. B 2021, 103, No. 195128.
(102) Wang, N.-P.; Rohlfing, M.; Krüger, P.; Pollmann, J. Electronic
excitations of CO adsorbed on MgO (001). Appl. Phys. A: Mater. Sci.
Process. 2004, 78, 213−221.
(103) Schleife, A.; Fuchs, F.; Furthmüller, J.; Bechstedt, F. First-
principles study of ground-and excited-state properties of MgO, ZnO,
and CdO polymorphs. Phys. Rev. B 2006, 73, No. 245212.
(104) Perdew, J. P.; Ruzsinszky, A.; Csonka, G. I.; Vydrov, O. A.;
Scuseria, G. E.; Constantin, L. A.; Zhou, X.; Burke, K. Restoring the
density-gradient expansion for exchange in solids and surfaces. Phys.
Rev. Lett. 2008, 100, No. 136406.
(105) Csonka, G. I.; Perdew, J. P.; Ruzsinszky, A.; Philipsen, P. H.;
Lebègue, S.; Paier, J.; Vydrov, O. A.; Á ngyán, J. G. Assessing the
performance of recent density functionals for bulk solids. Phys. Rev. B
2009, 79, No. 155107.
(106) Kresse, G.; Furthmüller, J. Efficient iterative schemes for ab
initio total-energy calculations using a plane-wave basis set. Phys. Rev. B
1996, 54, No. 11169.
(107) Kresse, G.; Joubert, D. From ultrasoft pseudopotentials to the
projector augmented-wave method. Phys. Rev. B 1999, 59, No. 1758.
(108) Gulans, A.; Kontur, S.; Meisenbichler, C.; Nabok, D.; Pavone,
P.; Rigamonti, S.; Sagmeister, S.; Werner, U.; Draxl, C. Exciting: a full-
potential all-electron package implementing density-functional theory
and many-body perturbation theory. J. Phys.: Condens. Matter 2014, 26,
No. 363202.

306 https://doi.org/10.1021/acs.jctc.4c01245
J. Chem. Theory Comput. 2025, 21, 291−306