Available via license: CC BY 4.0
Content may be subject to copyright.
Inverse Volume Scaling of Finite-Size Error in Periodic Coupled Cluster Theory
Xin Xing 1and Lin Lin 1,2
1Department of Mathematics, University of California, Berkeley, California 94720, USA
2Applied Mathematics and Computational Research Division,
Lawrence Berkeley National Laboratory, Berkeley, California 94720, USA
(Received 18 April 2023; revised 17 January 2024; accepted 13 February 2024; published 28 March 2024)
Coupled cluster theory is one of the most popular post-Hartree-Fock methods for ab initio molecular
quantum chemistry. The finite-size error of the correlation energy in periodic coupled cluster calculations
for three-dimensional insulating systems has been observed to satisfy the inverse volume scaling, even in
the absence of any correction schemes. This is surprising, as simpler theories that utilize only a subset of the
coupled cluster diagrams exhibit much slower decay of the finite-size error, which scales inversely with the
length of the system. In this study, we review the current understanding of finite-size error in quantum
chemistry methods for periodic systems. We introduce new tools that elucidate the mechanisms behind this
phenomenon in the context of coupled cluster doubles calculations. This reconciles some seemingly
paradoxical statements related to finite-size scaling. Our findings also show that singularity subtraction can
be a powerful method to effectively reduce finite-size errors in practical quantum chemistry calculations for
periodic systems.
DOI: 10.1103/PhysRevX.14.011059 Subject Areas: Chemical Physics,
Computational Physics,
Quantum Physics
I. INTRODUCTION
In the past few decades, ab initio methods for quantum
many-body systems, such as density functional theory
(DFT), quantum Monte Carlo methods, and quantum
chemistry wave function methods, are becoming increas-
ingly accurate and applied to ever larger range of systems
[1,2]. Unlike molecular systems, periodic systems, includ-
ing solids and surfaces, require calculating properties in the
thermodynamic limit (TDL), a theoretical state in which the
system size approaches infinity. However, the TDL cannot
be directly accessed in practical applications. Finite-sized
computational supercells are employed to approximate this
limit, which introduces finite-size errors into the calcula-
tions. Finite-size errors can significantly affect the accuracy
of calculations, even for systems with thousands of atoms.
An extreme case is a moir´e system such as magic angle
twisted bilayer graphene (MATBG), where each computa-
tional unit cell consists of approximately 10 000 atoms, and
the supercell needs to have more than 100 000 atoms to
capture subtle correlation effects [3,4]. Directly tackling
finite-size effects by enlarging the supercell size is
very demanding, even for relatively inexpensive DFT
calculations with modern-day supercomputers. For more
accurate theories, this task is often computationally intrac-
table. Understanding the finite-size scaling, i.e., the scaling
of the finite-size error with respect to the system size, and
employing finite-size error correction schemes are, there-
fore, crucial for obtaining accurate results using moderate-
sized calculations.
The sources of finite-size errors in ab initio calculations
are multifaceted and complex [5–7]. These errors are
influenced by numerous factors, including system charac-
teristics such as whether it is insulating or metallic or
whether it is a three-dimensional bulk system versus a low-
dimensional system. Calculations of electron kinetic
energy, electron-ion interaction energy, Hartree energy,
Fock exchange energy, and electron correlation energy
can all contribute to finite-size errors. The first four types
are predominantly single particle in nature, while the
electron correlation energy is significantly more complex.
To a large extent, electronic correlation is short-ranged, and
this characteristic has spurred the development of local
correlation methods, whose computational cost may scale
linearly with system size. However, for ab initio methods to
be accurate, they must also effectively account for van der
Waals (vdW) interactions [8]. In solids, the cumulative
effect of weak van der Waals interactions can become a
significant contributor to the energy. The convergence of
vdW energy follows an inverse volume scaling, implying
that the finite-size error is inversely proportional to the
volume of the supercell. The origin of finite-size error is not
Published by the American Physical Society under the terms of
the Creative Commons Attribution 4.0 International license.
Further distribution of this work must maintain attribution to
the author(s) and the published article’s title, journal citation,
and DOI.
PHYSICAL REVIEW X 14, 011059 (2024)
2160-3308=24=14(1)=011059(31) 011059-1 Published by the American Physical Society
solely confined to the characteristics of a physical system.
The finite-size scaling can vary significantly across differ-
ent theoretical frameworks even when applied to the same
physical system. Sometimes, the finite-size scaling can
manifest differently even when using the same theoretical
framework for the same system, simply by altering the
shape of the computational unit cell. All these complexities
require careful analysis under varying scenarios, and this is
important for accurately extrapolating toward the TDL or,
preferably, for devising improved correction schemes
aimed at reducing finite-size errors.
In recent years, there has been a growing interest in
employing quantum chemistry wave function methods,
such as Møller-Plesset perturbation theory and coupled
cluster (CC) theory [9], to compute ground-state and
excited-state properties for periodic systems [7,10–25].
Originally developed to study nuclear physics [26,27],
CC theory has become one of the most popular methods
in quantum chemistry that involves electronic correlation
[28]. The coupled cluster singles, doubles, and perturbative
triples [CCSD(T)] theory is often referred to as the “gold
standard”in molecular quantum chemistry. In these meth-
ods, there are two primary strategies to approximate the
TDL. The first involves expanding the computational
supercells within real space. The second strategy involves
performing calculations using a fixed unit cell and refining
discretization of the Brillouin zone within reciprocal space
using a k-point mesh, such as the Monkhorst-Pack mesh
[29]. This paper focuses on the latter approach, where the
number of kpoints is denoted by Nk. If the Monkhorst-Pack
mesh includes the Gamma point of the Brillouin zone, this
approach is equivalent to using a supercell comprised of Nk
unit cells. The convergence toward the TDL can be studied
by increasing this single parameter Nktoward infinity. The
computational expense of quantum chemistry methods can
rise sharply with respect to the system size. For example, in
coupled cluster singles and doubles (CCSD) calculations,
the computational cost scaling is OðN6
kÞin real-space
implementations [9] and OðN4
kÞin reciprocal-space imple-
mentations [30]. Therefore, even moderate-sized calcula-
tions with Nk¼2×2×2or 3×3×3can already be
computationally challenging. In CCSD(T) calculations, the
computational cost of the real-space implementation can
escalate to OðN7
kÞin the worst-case scenario. This implies
that even a moderate refinement of the Monkhorst-Pack grid
by a factor of 2 along each dimension can lead to an increase
in computational cost by ð23Þ7≈2million folds. Given this
computational challenge imposed by system size, it is
mostly impractical to estimate the finite-size scaling using
power-law fitting over calculations on increasingly large
systems and then determine the TDL value by extrapolation.
Instead, a more feasible approach is to acquire the exact
finite-size scaling through rigorous mathematical analysis
and subsequently utilize power-law extrapolation to esti-
mate the TDL value.
Although numerous empirical studies have examined the
finite-size scaling of some quantum chemistry methods,
there has been a notable lack of rigorous analysis. To the
best of our knowledge, the first rigorous analysis of finite-
size error in Hartree-Fock (HF) theory and second-order
Møller-Plesset perturbation theory (MP2) for insulating
systems has been conducted only recently [31]. The
principal findings in Ref. [31] include the following.
(1) In the absence of any corrections, the finite-size error
in HF scales as OðN−1=3
kÞ.AsNkis proportional to
the supercell volume, N1=3
kis proportional to the
length of the supercell (we always assume uniform
refinement of the k-point mesh along all three
dimensions for any shaped unit cell). This scaling
is, thus, also referred to as inverse length scaling.
(2) By applying the Madelung constant correction, the
finite-size error in HF improves to OðN−1
kÞ, i.e., the
inverse volume scaling.
(3) The finite-size error of the MP2 correlation energy
also satisfies the inverse volume scaling OðN−1
kÞ.
As can be seen from HF and MP2, the specific finite-size
scaling depends on the level of computational theory and
numerical treatment employed in the calculations. In
addition to confirming the empirically observed scalings,
this rigorous analysis offers significant additional insights.
It elucidates the nature of the singularity of the electron
repulsion integral (ERI) due to the Coulomb kernel and its
impact on finite-size scaling; it explains why the finite-size
error can depend on the shapes and symmetries of the unit
cell; and it has led to the development of new correction
schemes, such as the staggered mesh method, which can
expedite the convergence of HF, MP2, and random phase
approximation (RPA) calculations toward TDL [32,33].
An innovative theoretical observation provided in
Ref. [31] is that the finite-size error can be largely compre-
hended if assuming HF single-particle orbitals can be
acquired exactly at any given kpoint in the Monkhorst-
Pack mesh. This perspective helps to disentangle the
contribution due to the relatively manageable single-particle
effects from the collective and more complex electron
correlation effects. Based on this assumption, the finite-size
error can be rigorously examined from a numerical quad-
rature perspective. Specifically, the value of a physical
observable at the TDL can often be written as a multidi-
mensional integral over the Brillouin zone. As the
Monkhorst-Pack mesh forms a uniform grid discretizing
the Brillouin zone, the analysis simplifies to investigating
the quadrature error of certain trapezoidal rule in a periodic
region, a topic widely discussed in numerical analysis
literature. The novelty here lies in the recognition that the
associated integrands possess a unique singularity structure
that is asymptotically of a specific fractional form.
Reference [31], therefore, develops a new Euler-
MacLaurin type of analysis that facilitates the study of
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-2
the finite-size error associated with HF and MP2 methods,
taking into account this fractional form singularity.
Compared to HF and MP2, the finite-size error analysis
and finite-size error correction methods in CC methods
remain nascent. Early works in Refs. [12,34] focus on an
important intermediate quantity called the structure factor
and develop finite-size correction methods based on differ-
ent types of polynomial interpolation of the structure factor
near the Coulomb singularity. Subsequent studies on the
uniform electron gas system [35,36] numerically examine
the power-law scaling of the finite-size error in CC theory,
offering new analytical perspective on the structure factor.
These works also provide new error correction methods,
such as twist-angle techniques [37,38] and structure factor
interpolation [39] for these metallic systems. Very recently,
Ref. [40] replaces the Coulomb kernel in CC calculations
with an averaged Coulomb kernel in each quadrature
element to reduce the finite-size error for anisotropic
systems. Despite these developments, theoretical under-
standing of the finite-size error scaling in CC methods (with
and without error correction methods) has been lacking.
Recently, by applying a similar approach, we have
expanded the finite-size error analysis to include the
correlation energy of the third-order Møller-Plesset pertur-
bation theory (MP3) and CC theory [41]. The simplest CC
theory is the coupled cluster singles (CCS) theory.
However, due to Thouless’s theorem [42], CCS only rotates
the Slater determinant to another Slater determinant. For a
nondegenerate closed-shell Hartree-Fock reference, the
correlation energy from the CCS theory vanishes.
Therefore, we focus on the coupled cluster doubles
(CCD) theory, which is mathematically the simplest and
representative form of CC theory. From a diagrammatic
perspective, CC diagrams encompass all Møller-Plesset
perturbation diagrams. When the CCD amplitude equation
is solved iteratively with nfixed point iterations [referred to
as the CCDðnÞscheme], the MP2 and MP3 diagrams can
be generated from CCD(1) and CCD(2), respectively. More
generally, CCDðnÞconsists of a finite subset of Møller-
Plesset perturbation diagrams. It is worth pointing out that
there is a method called CCSD(2) which uses second-order
perturbation to rectify CCSD energies for multireference
and open-shell systems [43], and its meaning should,
therefore, be distinguished from our CCDðnÞ. Our analysis
uncovers that the finite-size error in MP3 and CC is
fundamentally different from that in MP2. Specifically,
theories such as MP2 and RPA incorporate only “particle-
hole”types of diagrams. The integrands corresponding to
these diagrams are singular, but the singularity is relatively
mild. However, starting from MP3, additional perturbative
terms, namely, the “particle-particle”and “hole-hole”
diagrams, must be considered. These terms introduce much
stronger singularities, necessitating the development of new
analytical tools. Our quadrature error analysis adapts the
Poisson summation formula in a new setting and aligns
with a recently developed trapezoidal quadrature analysis
for certain singular integrals [44]. This approach provides
more accurate estimates and can be more applicable than
our previous quadrature analysis based on the Euler-
MacLaurin formula for HF and MP2 in Ref. [31]. In the
absence of finite-size correction schemes and assuming
exact orbital energies at any kpoint, the study in Ref. [41]
concludes that the finite-size errors of MP3 and CCD both
satisfy the inverse length scaling OðN−1=3
kÞ.
Interestingly, for CCD, earlier numerical calculations did
not provide conclusive evidence regarding its finite-size
scaling, with different studies suggesting either an inverse
volume scaling [15,34] or inverse length scaling [17]. More
recent calculations demonstrate that the electron correlation
energy in periodic coupled cluster calculations should
follow an inverse volume scaling, even in the absence of
finite-size correction schemes. This observation points to a
significant gap in the theoretical understanding of the finite-
size error and prompts the central question we address in
this paper: How can we reconcile the following seemingly
paradoxical facts?
(1) Without finite-size corrections, the finite-size error
in CCD exhibits inverse volume scaling.
(2) Without finite-size corrections, the finite-size error
in MP3 exhibits inverse length scaling. This rate is
sharp and cannot be improved.
(3) All MP3 diagrams are encompassed within the CCD
formulation.
There are several often-cited physical justifications for
expecting that the CCD method, and CC theory more
generally, may exhibit superior behavior when applied to
periodic systems. One such reason stems from the size
extensivity of CC theory. A theory is size extensive when
the total energy of two noninteracting identical systems,
calculated as a combined system, equates to twice the
energy of one system computed independently. Unlike
theories such as truncated configuration interaction meth-
ods, which are not size extensive, truncated CC theory
(such as CCD) possesses this advantageous characteristic.
Another reason is that CC theory can be formulated in such
a way that it does not explicitly depend on orbital energies.
One practical consequence of this is that, upon the con-
vergence of the coupled cluster iterations, the Madelung
constant correction, often used to reduce finite-size effects
in many-body simulations, cancels out naturally. Therefore,
in this scenario, CC without the Madelung constant
correction is equivalent to that with the Madelung constant
correction.
We would like to clarify that neither size extensivity nor
the cancellation of the Madelung constant alone is suffi-
cient to address the aforementioned question. While size
extensivity is indeed a desirable property, many methods
such as HF, MP2, and MP3, among others, are all size
extensive. In fact, given that periodic systems are infinitely
large, size extensivity should be viewed as a necessary
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-3
condition for the applicability of any numerical method to
such systems. However, this property does not provide
insight into the convergence rate of the finite-size error. The
cancellation of the Madelung constant plays a pivotal role
here. Nevertheless, demonstrating that CCD calculation (or
even MP3) with the Madelung constant correction exhibits
inverse volume scaling in finite-size error is itself a
significant challenge. This requires the development of
new technical tools not currently available in existing
literature. Indeed, the development and application of these
tools are the main technical contribution of this paper.
In this paper, we elucidate the origin of the inverse
volume scaling behavior. Our analysis consists of two
steps. First, we investigate the structure of the CCD
amplitude equation. We show that the Madelung constant
correction, commonly used to reduce finite-size errors in
Fock exchange energy and orbital energies, can also be
applied to reduce the finite-size error in ERI contractions
within the CCD amplitude equation. We establish a con-
nection between the Madelung constant correction and a
quadrature error reduction technique known as the singu-
larity subtraction method [45]. By subtracting the leading
singular terms from the integrands in the numerical quad-
ratures, the Madelung constant correction reduces the
finite-size errors in both the ERI contractions and the
orbital energies from OðN−1=3
kÞto OðN−1
kÞ. Furthermore,
we demonstrate that, with the Madelung constant correc-
tion, the finite-size errors in CCDðnÞand converged CCD
calculations satisfy the desired inverse volume scaling.
In the second step of our analysis, we observe that, upon
convergence of the CCD amplitude equations, the
Madelung constant corrections to both orbital energies
and ERI contractions perfectly cancel each other out for any
finite-sized system. This cancellation ensures that the CCD
correlation energy remains the same, regardless of whether
the Madelung constant correction is applied. Combining
this result with the first step, we conclude that the finite-size
error of the correlation energy in converged CCD calcu-
lations satisfies the desired inverse volume scaling without
the need for any additional correction schemes. However,
prior to the convergence of the amplitude equations, this
perfect cancellation does not occur, and the finite-size error
of CCDðnÞcalculations remains OðN−1=3
kÞ. A similar lack
of cancellation occurs when the orbital energies take their
exact value at the TDL but the ERI contractions are not
corrected, resulting in an OðN−1=3
kÞfinite-size error for both
converged CCD and CCDðnÞcalculations studied
in Ref. [41].
To validate our theoretical analysis, we perform CCD
calculations on a 3D periodic hydrogen dimer system using
the
PYSCF
software package [46]. Our numerical results
support the conclusions drawn from the theoretical analysis
and provide further evidence for the finite-size scaling
behavior summarized in Table I.
The paper is organized as follows. Section II introduces
background knowledge and basic notations. Section III first
decomposes the finite-size error in CCDðnÞcalculations
into the errors in four basic components and then describes
how the four components contribute to the overall finite-
size error with the possible Madelung constant correction.
Section IV explains the key ideas of how the Madelung
constant correction can reduce the finite-size error in orbital
energies and ERI contractions from a numerical quadrature
perspective. Section Villustrates the numerical results that
corroborate our error estimate. Lastly, Sec. VI discusses the
implication of our theoretical study and future directions.
TABLE I. Finite-size scaling of different computational theories with and without corrections to orbital energies
and ERI contractions. The first two rows refer to the finite-size scaling of the Hartree-Fock (HF) exchange energy,
and the remaining rows refer to the finite-size scaling of the correlation energy (excluding the HF exchange). While
we focus on the Madelung constant correction in Ref. [31] and this work, any correction schemes that reduces the
finite-size errors in orbital energies and ERI contractions to OðN−1
kÞcan also be applied, and the conclusions drawn
here remain valid. In particular, exact values of single particle orbital energies at the TDL satisfy the condition
above. N/A means that Madelung constant correction does not apply within the theory.
Theory
Correction to
orbital energies
Correction to
ERI contractions
Finite-size
scaling References
HF N/A ✗N−1=3
k[31, Theorem 3.1] and [6,17,47]
HF N/A ✓N−1
k[31, Theorem 5.1] and [47]
MP2 ✓N/A N−1
k[31, Theorem 4.1] and [15,34]
RPA, SOSEX, drCCD ✓N/A N−1
k[33,34]
MP3 ✓✗
N−1=3
k[41]
MP3 ✓✓
N−1
kTheorem 1
CCDðnÞ=CCD ✓✗
N−1=3
k=N−1=3
k[41]
CCDðnÞ=CCD ✓✓
N−1
k=N−1
kTheorem 1/Theorem 2
CCDðnÞ=CCD ✗✓
N−1=3
k=N−1=3
kTheorem 1/Theorem 2
CCDðnÞ=CCD ✗✗
N−1=3
k=N−1
kTheorem 1/Corollary 3
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-4
II. BACKGROUND
Consider a unit cell and its Brillouin zone denoted as Ω
and Ω, respectively. Denote the associated real- and
reciprocal-space lattices as Land L, respectively. To
model such a periodic system, the Brillouin zone Ωis
commonly discretized using a uniform mesh Kof size Nk
(known as the Monkhorst-Pack grid [29]). The orbitals and
orbital energies (also called bands and band energies)
fψnk;εnkgindexed by orbital indices nand momentum
vectors k∈Kcan be solved by the HF method. As a
common practice, n∈fi; jgrefers to an occupied orbital
and n∈fa; bgrefers to an virtual orbital. Throughout this
paper, we use the normalized ERI:
hn1k1;n
2k2jn3k3;n
4k4i
¼4π
jΩjX0
G∈L
1
jqþGj2
ˆ
ϱn1k1;n3ðk1þqÞðGÞˆ
ϱn2k2;n4ðk2−qÞð−GÞ;
ð1Þ
where q¼k3−k1is the momentum transfer vector,
the crystal momentum conservation k1þk2−k3−
k4∈Lis assumed implicitly, and ˆ
ϱn0k0;nkðGÞ¼
hψn0k0jeiðk0−k−GÞ·rjψnkiis Fourier representation of the pair
product. The primed summation over Gmeans that the
possible term with qþG¼0is excluded in the numerical
calculation. Using a finite mesh K, the HF orbital energy
without any correction is computed as
εNk
nk¼hnkj^
H0jnki
þ1
NkX
ki∈KX
i
ð2hiki;nkjiki;nki−hiki;nkjnk;ikiiÞ;
ð2Þ
where ^
H0refers to the single-particle component of the many
body Hamiltonian.
In this paper, we focus on three-dimensional insulating
systems with an indirect gap, i.e., εaka−εiki>0,
∀i; a; ki;ka. To simplify the analysis, we assume that
the orbitals are exact at any kpoint and the number of
virtual orbitals are truncated to a finite number. In addition,
we assume that the exact orbitals and orbital energies in the
TDL are smooth and periodic with respect to their
momentum vector index k∈Ω. This assumption is a
restriction in our current analysis. For systems free of
topological obstructions [48,49], these conditions may be
replaced by weaker ones using techniques based on Green’s
functions or Hamiltonians defined in the atomic orbital
basis instead of the band basis [50].
A. CCD theory
Based on the reference HF determinant jΦi, the CCD
theory represents the wave function as
jΨi¼eTjΦi;
T¼1
NkX
ijab X
ki;kj;ka∈K
Tijabðki;kj;kaÞa†
akaa†
bkbajkjaiki;
where a†
nkand ankare creation and annihilation operators,
respectively, for ψnk,Tijabðki;kj;kaÞ(commonly denoted
as taka;bkb
iki;jkjin the literature) is the normalized CCD double
amplitude, and kb∈Kis uniquely determined using the
crystal momentum conservation kiþkj−ka−kb∈L.
The amplitude tensor TNk
¼fTijab ðki;kj;kaÞg is defined
as the root of a nonlinear amplitude equation that consists
of constant, linear, and quadratic terms.
In practice, the amplitude equation can be solved using a
quasi-Newton method [9,51], which is equivalent to apply-
ing fixed point iteration to
T¼1
εNk
ANkðTÞ:ð3Þ
Here, εNkdenotes a diagonal operator with entries
εNk
iki;jkj;aka;bkb¼εNk
ikiþεNk
jkj−εNk
aka−εNk
bkb, and 1=εNkgives
the diagonal operator with 1=εNk
iki;jkj;aka;bkb. The operator
ANkðTÞis referred to as the ERI-contraction map (see the
definition in Appendix A). It consists of contractions
between ERIs and Tand does not involve orbital energies.
Note that both Tand ANkðTÞare tensors indexed by
ði; j; a; bÞand ðki;kj;kaÞ∈K×K×K. The CCD corre-
lation energy is then defined as
ENk
CCD ¼1
N3
kX
kikjka∈KX
ijab
Wijabðki;kj;kaÞTijab ðki;kj;kaÞ
≔GNkðTNk
Þ;
where Wijabðki;kj;kaÞdenotes the antisymmetrized
ERI 2hiki;jkjjaka;bkbi−hiki;jkjjbkb;akai.
In the TDL with Kreplaced by Ω,TNk
¼
fTijabðki;kj;kaÞg with ðki;kj;kaÞ∈K×K×Kshould
be replaced by t¼ftijabðki;kj;kaÞg, where each
tijabðki;kj;kaÞis a function of ðki;kj;kaÞin Ω×
Ω×Ω. The converged TDL amplitude tsatisfies a
similar amplitude equation
t¼1
εTDL ATDLðtÞ;ð4Þ
where ATDL can be obtained from ANkby taking the limit
ð1=NkÞPk∈K→ð1=jΩjÞ RΩdkwith Nk→∞.For
example, the four-hole–two-particle (4h2p)linear term
in ½ANkðTÞijabðki;kj;kaÞconverges in ATDLðtÞas
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-5
1
NkX
kk∈KX
kl
hkkk;lkljiki;jkjiTklabðkk;kl;kaÞ
⟶
K→Ω1
jΩjZΩ
dkkX
kl
hkkk;lkljiki;jkjitklabðkk;kl;kaÞ:
ð5Þ
The image of ATDL is a set of functions of ðki;kj;kaÞ
indexed by ði; j; a; bÞ. The CCD correlation energy in the
TDL is defined in a similar way as
ETDL
CCD ¼1
jΩj3ZΩ×Ω×Ω
dkidkjdkaWijabðki;kj;kaÞ
×tijabðki;kj;kaÞ
≔GTDLðtÞ:
Applying nsteps of fixed point iteration over Eqs. (3)
and (4) with zero initial guess (i.e., t¼0,T¼0), we
obtain the CCDðnÞamplitude and the CCDðnÞenergy:
ENk
CCDðnÞ¼GNkðTNk
nÞETDL
CCDðnÞ¼GTDLðtnÞ
TNk
m¼ðεNkÞ−1ANkðTNk
m−1Þ→tm¼ðεTDL Þ−1ATDLðtm−1Þ
TNk
0¼0t0¼0:
ð6Þ
CCDðnÞis related to the perturbative description of CCD
and consists of a subset of finite order perturbation energy
terms in the Møller-Plesset perturbation theory. For exam-
ple, CCD(1) can be identified with MP2, and CCD(2)
contains all the terms in MP2 and MP3 and a subset of
terms in MP4.
One main result of this paper is the rigorous analysis of
the finite-size error in CCDðnÞcalculation with any fixed
n>0. If the fixed point iterations in both the finite and
TDL cases converge to the CCD amplitudes as n→∞, i.e.,
TNk
n→TNk
and tn→t(the technical definition of this
convergence is provided in Appendix C), the finite-size
error analysis for CCDðnÞalso applies to the converged
CCD calculation. In other words, we analyze the finite-size
error in CCD calculation using a perturbative approach
based on the analysis on CCDðnÞ.
B. Madelung constant correction
Reference [41] shows that the finite-size errors in
CCDðnÞand CCD both scale as OðN−1=3
kÞwhen assuming
that exact orbital energies are used in the amplitude
equation (3). The same finite-size scaling also appears in
Fock exchange energy and occupied orbital energy calcu-
lations. One common correction to reduce the finite-size
errors in the latter two calculations is to add a Madelung
constant shift [6,47,52] to the Ewald kernel. This shift
introduces a correction to all involved ERIs in the calcu-
lations as
hn1k1;n
2k2jn3k3;n
4k4i−δn1n3δn2n4δk1k3δk2k4Nkξ:
Such a correction is triggered only in ERIs which have fully
matched orbital indices, i.e., n1¼n3;n
2¼n4, and zero
momentum transfer, i.e., k1¼k3. The Madelung constant
ξis defined uniquely by the unit cell and the k-point mesh
Kas
ξ¼1
NkX
q∈Kq
−
1
jΩjZΩ
dqX0
G∈L
4π
jΩj
e−σjqþGj2
jqþGj2−
4πσ
Nk
þX0
R∈LK
erfcðσ−1=2jRj=2Þ
jRj;ð7Þ
where Kqis a uniform mesh that is of the same size as K
and contains q¼0and LKis the real-space lattice
associated with the reciprocal-space lattice qþGwith
q∈Kq;G∈L. Note that ξdoes not vary with respect to
σ>0[6] and this parameter σis commonly tuned to
control the lattice cutoffs in the summation over Land LK
in Eq. (7) when numerically computing ξ.
For finite-size orbital energy calculation in Eq. (2), the
Madelung constant correction gives
εNk;ξ
nk¼εNk
nkþξnis occupied;
εNk
nknis virtual:ð8Þ
For the ERI contractions in ANkðTÞ, this correction
should be applied to six linear amplitude terms (see
Appendix B2). For example, the 4h2plinear term in
Eq. (5) should be modified to
1
NkX
kk∈KX
kl
hkkk;lkljiki;jkjiTklabðkk;kl;kaÞ
−ξTijabðki;kj;kaÞ:ð9Þ
Collecting the corrections to all six terms together, ANkðTÞ
is modified to
ANk;ξðTÞ¼ANkðTÞþ2ξT: ð10Þ
In the finite-size CCDðnÞand converged CCD calculations,
the Madelung constant correction can be applied to the
orbital energies, ERI contractions, or both in the amplitude
equation (3). As a result, we may have three correction
schemes compared to the standard calculation without any
correction.
Particularly, applying the Madelung constant correction
to both components gives the amplitude equation
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-6
T¼1
εNkþ2ξ½ANkðTÞþ2ξT¼ 1
εNk;ξANk;ξðTÞ:ð11Þ
It can be easily verified that this amplitude equation has the
same solution as Eq. (3) without any correction. Its CCD
solution is, thus, identical to the one without correction.
However, the associated CCDðnÞcalculation differs and
can be interpreted as solving the original amplitude
equation (3) using a quasi-Newton method that has a 2ξ-
diagonal shift to the Jacobian matrix.
III. MAIN STATEMENTS
We start our error analysis with the CCDðnÞcalculation
with a fixed number of iterations n>0and then general-
ize the analysis to the fully converged CCD calcu-
lation. In CCDðnÞ, the finite-size error is quantified by
ETDL
CCDðnÞ−ENk
CCDðnÞ. According to Eq. (6), this error can be
traced back to two sources: the difference in energy oper-
ators, GTDL versus GNk, and the difference in amplitudes, tn
versus TNk
n.Recallthattnand TNk
nare the amplitudes of the
system in TDL and finite-size cases, respectively. Let us
consider the evaluation map,denotedbyMK.Thismap
evaluates a tensor valued function, initially defined on the
product space Ω×Ω×Ω, on a finite-sized grid
K×K×K. Consequently, the values of the TDL amplitude
tnon this finite-size grid, K×K×K, are given by MKtnand
are approximated by the finite-size amplitude TNk
n.
By applying the triangle inequality, we can decompose
the finite-size error into two distinct sources: the errors
arising from the discretized energy calculation using the
exact amplitude MKtnand the errors stemming from the
amplitude calculation itself:
jETDL
CCDðnÞ−ENk
CCDðnÞj
≤jGTDLðtnÞ−GNkðMKtnÞj þjGNkðMKtnÞ−GNkðTNk
nÞj
≤jGTDLðtnÞ−GNkðMKtnÞj
þC
N3
kX
ijab X
ki;kj;ka∈K
j½MKtn−TNk
nijabðki;kj;kaÞj:ð12Þ
Here, we use the fact that jWijabðki;kj;kaÞj can be upper
bounded uniformly by a constant.
To further break down the error in amplitude calculation,
we note that tnand TNk
nare recursively constructed by
Eq. (6) with initial values t0¼0and TNk
0¼0. As a result,
the error in the CCDðnÞamplitude calculation can
also be recursively decomposed using the same strategy
above as
MKtn−TNk
n¼1
εTDL ½MKATDLðtn−1Þ−ANkðMKtn−1Þ
þANkðMKtn−1Þ1
εTDL −
1
εNk
þ1
εNkhANkðMKtn−1Þ−ANkðTNk
n−1Þi:ð13Þ
The three error terms from this dissection can be interpreted
as the errors in ERI contractions using exact CCDðn−1Þ
amplitudes, orbital energies, and CCDðn−1Þamplitude
calculation composed with ANk, respectively. For the
last term, it can be shown that entries in ANkðMKtn−1Þ−
ANkðTNk
n−1Þhave the same scaling with respect to Nkas
those in MKtn−1−TNk
n−1. Replacing the last term by
MKtn−1−TNk
n−1and applying the dissection recursively,
we find that the error in the CCDðnÞamplitude calculation
is determined by those in the ERI contractions and orbital
energies, i.e., the first two terms in Eq. (13).
Overall, the finite-size error of CCDðnÞcalculation
can be decomposed into errors in three basic factors:
(i) energy calculation using exact CCDðnÞamplitude,
(ii) ERI contractions using exact CCDðn−1Þamplitude,
and (iii) orbital energies. This error decomposition is also
valid when applying the Madelung constant correction to
orbital energies [Eq. (8)] or ERI contractions [Eq. (10)]. By
analyzing these three error sources with or without cor-
rections separately, we can obtain the finite-size error
estimate for CCDðnÞwith various correction schemes.
The Madelung constant correction can reduce the finite-
size error in orbital energies from OðN−1=3
kÞto OðN−1
kÞ.
This correction is at the HF level. One main technical result
of this work is to show that the Madelung constant
correction can also reduce the finite-size error in most
(but not all) entries of the ERI contraction [note that
MKATDLðtÞ−ANkðMKtÞis a tensor] from OðN−1=3
kÞto
OðN−1
kÞ; see Sec. IV B. As a result, when applying the
Madelung constant correction to both orbital energies and
ERI contractions in the CCDðnÞcalculation, the overall
finite-size error can be successfully reduced to OðN−1
kÞ.
Theorem 1. In CCDðnÞcalculation, the finite-size error
in the correlation energy scales as OðN−1=3
kÞin each of the
following scenarios: (i) there is no finite-size correction,
(ii) the Madelung constant correction is applied only to the
ERI contraction ANk, and (iii) the Madelung constant
correction is applied only to the orbital energy εNk
nk.
When the Madelung constant correction is applied to
both ANkand εNk
nkin the CCDðnÞcalculation, the overall
finite-size error scales as OðN−1
kÞ.
As a special case, the same conclusion applies to MP3
calculations.
Proof. See Appendix B.▪
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-7
Now what happens to the finite-size error of the CCD
calculation as n→∞? For gapless and small-gap systems,
it has been observed in practice that the fixed point iteration
might not converge or the amplitude equation might have
multiple solutions. While we focus on systems with an
indirect gap, it is worth noting that, even in such favorable
scenarios, the existence and uniqueness of solutions to the
CCD amplitude equations in both finite and TDL cases
remain an open question and are beyond the scope of
this paper.
To study the finite-size error of CCD via the above
results on CCDðnÞ, we make additional technical assump-
tions (see Appendix C) that can guarantee the convergence
of CCDðnÞto CCD. Under these assumptions, we show
that the finite-size scaling of CCD calculation is upper
bounded by those of its associated converging CCDðnÞ
calculations. Numerical observations (see Sec. V) further
show that this finite-size scaling estimate through CCDðnÞ
is asymptotically sharp for CCD calculation with the
Madelung constant correction applied to orbital energies,
ERI contractions, or both.
Theorem 2 (informal). Under additional conditions on
the convergence of CCDðnÞto CCD, the finite-size error of
the CCD correlation energy scales as OðN−1=3
kÞin each of
the following scenarios: (i) the Madelung constant correc-
tion is applied only to the ERI contraction ANk, and (ii) the
Madelung constant correction is applied only to the orbital
energy εNk
nk.
When the Madelung constant correction is applied to
both ANkand εNk
nkin the CCD calculation, the overall finite-
size error scales as OðN−1
kÞ.
Compared to Theorem 1, Theorem 2 does not address the
finite-size error of the CCD calculation when no finite-size
correction is applied. The natural conclusion from Theorem
1 is that this finite-size scaling should be OðN−1=3
kÞas well.
However, this error estimate is loose and inconsistent with
the numerical observations which suggest an inverse
volume scaling. To obtain a tight estimate, we now use
the observation that the CCD calculation without any
correction is equivalent to the CCD calculation with the
Madelung constant correction applied to both ERI con-
tractions and orbital energies. Specifically, when applying
the Madelung constant corrections Eqs. (8) and (10), the
CCD amplitude equation in Eq. (11) can be formulated as
ðεNkþ2ξÞT¼ANkðTÞþ2ξT⇔εNkT¼ANkðTÞ:
With 2ξTon both sides canceling each other, this refor-
mulation is exactly reduced to the original amplitude
equation (3) without corrections. Therefore, the roots of
the two amplitude equations with and without the correc-
tions are the same, and the correlation energy in converged
CCD calculation without any corrections is the same as that
with the Madelung constant correction. In other words,
when investigating the finite-size error of CCD calculation
without corrections, we should apply Theorem 2 with the
Madelung constant correction applied to both ANkand εNk
nk.
This yields a sharp estimate OðN−1
kÞand explains the origin
of the inverse volume scaling of the finite-size error.
Corollary 3. Under the same additional conditions as in
Theorem 2, the finite-size error of the CCD correlation
energy without finite-size correction scales as OðN−1
kÞ.
We provide the proof of Theorem 2 in Appendix C, and
Corollary 3 follows directly from Theorem 2 and the
reasonings above.
IV. KEY IDEAS
As demonstrated in the previous section, the finite-size
errors in CCDðnÞand CCD calculations can be reduced to
the errors in three simpler basic calculations: energy
calculation using exact amplitudes, ERI contraction using
exact amplitudes, and orbital energies. A key observation is
that the finite-size errors in all three calculations can be
interpreted and analyzed from a numerical quadrature
perspective.
For a function gover a hypercube V, we denote a
(generalized) trapezoidal rule using a uniform mesh Xof V
and its quadrature error as
QVðg; XÞ¼jVj
jXjX
x∈X
gðxÞ;
EVðg; XÞ¼ZV
dxgðxÞ−QVðg; XÞ:
Under the assumption of exact orbitals at any kpoint in
finite-size calculation, the errors in (i) orbital energy,
(ii) energy, and (iii) ERI contraction ANkcan be, respec-
tively, formulated by their definitions as
εTDL
nk−εNk
nk¼1
jΩjEΩX
i
Wininðki;k;kiÞ;K;
GTDLðtÞ−GNkðMKtÞ¼ 1
jΩj3EΩ×Ω×ΩX
ijab
Wijabtijab ðki;kj;kaÞ;K×K×K;
MKATDLðtÞ−ANkðMKtÞ¼ 1
jΩjEΩX
kl
hkkk;lkljiki;jkjitklabðkk;kl;kaÞ;Kþ:ð14Þ
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-8
For ERI contraction in the last line, the error is denoted by a
tensor with indices ði; j; a; b; ki;kj;kaÞ, and the error for
the 4h2plinear term Eq. (5) is detailed with kkbeing the
integration variable. Meanwhile, terms not shown account
for errors from computing other linear and quadratic
amplitude terms. From a numerical quadrature perspective,
all three finite-sized calculations approximate the corre-
sponding integrals in the TDL using the trapezoidal rule
and a finite mesh Kto discretize Ω. Therefore, their finite-
size errors can be estimated systematically by quadrature
error analysis.
In general, the quadrature error associated with a
trapezoidal rule is influenced by the integrand’s smoothness
and boundary conditions. If we take Δhas the mesh size
along each dimension, the quadrature error for a smooth
integrand is typically of order OðΔh2Þ. Interestingly, if the
integrand is also periodic, the error diminishes much more
rapidly than it does in a nonperiodic scenario. The decay
rate is faster than any finite power of Δh, showcasing a
superalgebraic decay [53]. However, for an integrand
periodic but marked by singularities, its quadrature error
tends to taper off at a slower rate.
For quadrature errors in Eq. (14), the involved integrands
are all periodic across their integration domains. However,
many of these integrands have singularities within the
domains, affecting the scaling of their quadrature errors. As
demonstrated next, these integrands include point singu-
larities arising from both ERIs and amplitudes, resulting in
low-order power-law decay of the corresponding quad-
rature errors as Nkincreases.
A. Singularity structure and quadrature
error estimate
All the quadrature errors in Eq. (14) have integrands
comprising of either ERIs or contractions between ERIs
and exact CCDðnÞamplitudes. The integration variables
are momentum vectors sampled in the Brillouin zone Ω.
Consequently, understanding the singularity structure of
ERIs and exact CCDðnÞamplitudes with respect to their
momentum vector indices is crucial in comprehending the
singularity structure of these integrands and ultimately
estimating the quadrature error in the three basic
calculations.
First, consider a generic ERI hn1k1;n
2k2jn3k3;n
4k4i
with fixed band indices ðn1;n
2;n
3;n
4Þand treat it as a
function of k1,k2, and q¼k3−k1in Ω. By its
definition, the ERI can be separated as
hn1k1;n2k2jn3k3;n4k4i
¼4π
jΩj
ˆ
ϱn1k1;n3ðk1þqÞð0Þˆ
ϱn2k2;n4ðk2−qÞð0Þ
jqj2
þ4π
jΩjX
G∈Lnf0g
ˆ
ρn1k1;n3ðk1þqÞðGÞˆ
ρn2k2;n4ðk2−qÞð−GÞ
jqþGj2:ð15Þ
Since we assume all orbitals ψnkperiodic and smooth with
respect to k∈Ω, this ERI has a point singularity at q¼0
only due to the first fraction term. Specifically, this term is of
fractional form fðk1;k2;qÞ=jqj2with a smooth numerator f.
As can be verified by direct calculation, such a fraction term
can have its point singularity at q¼0described by the
following general concept called algebraic singularity.
Definition 4 (informal). A function fðxÞhas algebraic
singularity of order γ∈Rat x0∈Rdif every lth-order
derivative near x0is bounded asymptotically by jx−x0jγ−l,
i.e.,
∂α
∂xαfðxÞ
≤Cjx−x0jγ−jαj;∀α≥0;
where αdenotes a non-negative d-dimensional derivative
multi-index. For brevity, fis said to be singular at x0with
the order of γ. See Definition 8 in Appendix Bfor the
rigorous mathematical definition.
A representative example of such a singular function of
order γis gðxÞ=jxj2, where gðxÞis smooth and scales as
Oðjxj2þγÞnear x¼0. Using orbital orthogonality, the
generic ERI exhibits singularities at q¼0with the order
of 0 when band indices mismatch (n1≠n3,n2≠n4), −1
when they partially match (n1¼n3;n
2≠n4or n1≠n3;
n2¼n4), and −2when they fully match (n1¼n3;n2¼n4).
We can now characterize the singularity structure of
integrand PiWininðki;k;kiÞdefined by orbital energy
εTDL
nkin Eq. (14). Fixing ðn; kÞ, the leading singularity in
each Wininðki;k;kiÞ(as a function of ki) comes from the
exchange term hiki;nkjnk;ikiiwhich has algebraic sin-
gularity at ki¼kof order 0 when i≠nand −2when
i¼n. In computing εTDL
nk, the overall integrand is, thus,
singular at ki¼kof order 0 for a virtual band nand −2for
an occupied band n.
For such periodic functions with one point of algebraic
singularity, the conventional textbook analysis of the
trapezoidal rule is overly pessimistic. A key technical tool
in this work is Lemma 5 below, which provides a rigorous
and sharp quadrature error estimate linking its error scaling
to the singularity order. (See Lemma 22 in Appendix Dfor
a more general statement.)
Lemma 5. Let fðxÞbe periodic with respect to V¼
½−1
2;1
2dand smooth everywhere except at x¼0with
singularity order γ≥−dþ1.Atx¼0,fðxÞis set to 0.
The quadrature error of a trapezoidal rule using an md-sized
uniform mesh Xthat contains x¼0can be estimated as
jEVðf; XÞj ≤Cm−ðdþγÞ:
If fð0Þis set to an Oð1Þvalue in the calculation, it
introduces additional Oðm−dÞquadrature error.
Combining this error estimate with the integrand singu-
larity structure for orbital energies (with m¼N1=3
k,d¼3,
and γ¼−2;0for occupied and virtual orbitals, respec-
tively), we obtain the finite-size error estimate for orbital
energy calculation as
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-9
εTDL
nk−εNk
nk
≤CN−1=3
knis occupied;
N−1
knis virtual:ð16Þ
To adapt the above approach for examining energy and
ERI-contraction calculations with exact amplitude, we need
to characterize the singularity structure of the exact
CCDðnÞamplitude entries. First, we note that the exact
CCD(1) amplitude is just the MP2 amplitude
nhaka;bkbjiki;jkjiðεTDL
iki;jkj;aka;bkbÞ−1o:
As a function of ðki;kj;kaÞ, each CCD(1) amplitude entry
indexed by ði; j; a; bÞhas the same singularity structure as
the included ERI term, which is singular at ka−ki¼0of
order 0. It turns out that the exact CCDðnÞamplitude with
any n>0all has the same singularity structure
as haka;bkbjiki;jkji.
Lemma 6 (singularity structure of the amplitude,
Lemma 4 in Ref. [41]). In CCDðnÞcalculation with
n>0, each entry of the exact double amplitude tn¼
f½tnijabðki;kj;kaÞg belongs to the following func-
tion space:
TðΩÞ¼ffðki;kj;kaÞ∶fis periodic with respect to ki;kj;ka∈Ω;
fis smooth everywhere except at ka¼kiwith algebraic singularity of order 0;
fis smooth with respect to ki;kjat the singularity ka¼kig:
Combining the above singularity structure characterizations of ERIs and exact CCDðnÞamplitudes, we are able to
analyze the integrands for energy and ERI-contraction calculations in Eq. (14). Take the CCDðnÞexchange energy term as
an example whose finite-size error can be formulated as
−
1
jΩj3EΩ×Ω×ΩX
ijab
hiki;jkjjbkb;akaitijabðki;kj;kaÞ;K×K×K;
with integration variables ðki;kj;kaÞ. For each set of
ði; j; a; bÞ, both the associated ERI and exact amplitude
in the integrand exhibit an algebraic singularity of order 0 at
ka−kj¼0and ka−ki¼0, respectively. For such a
product of two functions with algebraic singularities, we
also provide a rigorous quadrature error estimate similar to
Lemma 5. Specifically, for all the integrands in the energy
and ERI-contraction calculations, their quadrature errors
are determined by the most singular product components in
the integrand. The quadrature error still scales as m−ðdþγÞ
similar to Lemma 5 but with γdenoting the minimum
algebraic singularity order of all ERIs and exact ampli-
tudes.
For the exchange term above, all involved ERIs and
exact amplitudes have point singularities of order 0 and
similar for the direct term. Thus, we can get the finite-size
error estimate for the energy calculation using exact
CCDðnÞamplitude as
jGTDLðtÞ−GNkðMKtÞj ≤CN−1
k:ð17Þ
Similar analysis is also applicable to ERI contractions
using exact amplitudes. The key distinction lies in the fact
that the ERI contractions involve integrands formed by
ERIs with stronger singularities, such as those defined by
particle-particle or hole-hole diagrams. A prominent exam-
ple is the 4h2plinear term in Eq. (5), where the involved
ERI hkkk;lkljiki;jkjiexhibits an algebraic singularity of
order −2at kk¼kiwhen k¼iand l¼j. Consequently,
as per the above analysis, the finite-size error in the 4h2p
linear term calculation alone scales as OðN−1=3
kÞ. This error
turns out to dominate the overall finite-size error in the ERI-
contraction calculation, and we have
½MKATDL ðtÞ−ANkðMKtÞijabðki;kj;kaÞ
≤CN−1=3
k;∀i; j; a; b; ∀ki;kj;ka∈K:ð18Þ
B. Madelung constant correction as a quadrature
error reduction method
To reduce the quadrature error for a singular integrand,
one common numerical quadrature technique is the singu-
larity subtraction method [45]. Essentially, this method
involves constructing an auxiliary function hthat possesses
the same leading singularity as the integrand g. The integral
is then approximated as
ZV
dxgðxÞ≈jVj
jXjX
x∈X
ðg−hÞðxÞþZV
dxhðxÞ
¼jVj
jXjX
x∈X
gðxÞþZV
dx−jVj
jXjX
x∈XhðxÞ:ð19Þ
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-10
This approximation consists of the numerical quadrature of
g−hand the exact integral of h(which can be computed
analytically or numerically with high precision). It is also
equivalent to adding a correction EVðh; XÞto the numerical
quadrature of g. By this correction, the quadrature
error changes from EVðg; XÞto EVðg−h; XÞ. Since h
removes the leading singularity of gin the subtraction
g−h,EVðg−h; XÞcan be asymptotically smaller than
EVðg; XÞ.
The Madelung constant defined in Eq. (7) can be
reformulated using an arbitrary σ>0as
ξ¼−
1
jΩjZΩ
dq−jΩj
NkX
q∈KqX
G∈L
4π
jΩj
e−σjqþGj2
jqþGj2
þOðN−1
kÞ:ð20Þ
Compared to Eq. (19), this representation connects ξto the
singularity subtraction correction defined by an auxiliary
function
hσðqÞ¼ X
G∈L
4π
jΩj
e−σjqþGj2
jqþGj2:
The effectiveness of the Madelung constant correction to
reduce the finite-size error can be rigorously explained by
this connection.
Taking the occupied orbital energy εnkas an example, its
exchange portion leads to the dominant finite-size error,
and the associated Madelung constant correction modifies
the calculation as (with a change of variable ki→k−q)
1
NkX
q∈KqX
i
hiðk−qÞ;nkjnkiðk−qÞi−ξ:
Comparing this calculation with Eqs. (19) and (20), the
Madelung constant correction exactly uses the auxiliary
function hσðqÞto remove the leading singularity (i.e.,
4π=jΩjjqj−2) of the target integrand and, thus, reduces
the associated finite-size error asymptotically to OðN−1
kÞ.
One major technical contribution of this paper is to
rigorously prove the effectiveness of the Madelung constant
correction for reducing the finite-size error in the ERI-
contraction calculations following the same singularity
subtraction interpretation. For instance, consider the
4h2plinear term calculation in Eq. (5) with any fixed
entry index ði; j; a; b; ki;kj;kaÞ. Using the change of
variable kk→ki−q, this term with the Madelung con-
stant correction can be detailed as
1
NkX
q∈KqX
kl
hkðki−qÞ;lðkjþqÞjiki;jkjitklabðki−q;kjþq;kaÞ−ξtijabðki;kj;kaÞ:ð21Þ
The leading singularity of the integrand comes from the
product with ðk; lÞ¼ði; jÞ. In this product, the ERI is
singular at q¼0of order −2, and the amplitude is singular
at q¼ki−kaof order 0. The correction in Eq. (21)
defines a singularity subtraction with an auxiliary function
hσðqÞtijabðki;kj;kaÞ. This auxiliary function shares ex-
actly the same leading singularity as the integrand at q¼0
due to the ERI term, i.e.,
4π
jΩj
1
jqj2tijabðki;kj;kaÞ:
Similar to the orbital energy analysis, the finite-size error in
this 4h2plinear entry can be reduced to OðN−1
kÞ. However,
the key difference here is that this error reduction is the case
only for most but not all the amplitude entries. The
exception is when the amplitude singularity q¼ka−ki
is close or equal to the ERI singularity q¼0, i.e., when
computing an ERI-contraction entry whose momentum
vector indices kiand kaare close or identical. In the worst
case when ki¼ka, the two singularities overlap and the
finite-size error of such a 4h2plinear entry can be shown to
be still of scale OðN−1=3
kÞ. Similar analysis also applies to
other terms in the ERI-contraction calculation.
In summary, the Madelung constant correction does not
uniformly reduce the finite-size errors in the ERI-contrac-
tion tensor. This is the case for the 4h2plinear term and
also for other terms in the ERI contraction. More precisely,
we have the following technical error estimate (see Lemma
11 in Appendix Bfor the general statement and proof).
Lemma 7 (error in ERI contractions). The finite-size
error in the ERI contractions using exact CCDðnÞampli-
tude tnwith the Madelung constant correction in Eq. (10)
satisfies
½MKATDL ðtnÞ−ANk;ξðMKtnÞijabðki;kj;kaÞ
≤C(1
jqiaj2N−1
kqia ≠0;
N−1=3
kqia ¼0;
where qia ¼ka−kiþG0with G0∈Lchosen such that
qia ∈Ω.
As a result of this nonuniform error reduction, the
maximum entrywise finite-size error in the CCDðnÞ
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-11
amplitude calculation MKtn−TNk
nwith the Madelung
constant correction remains OðN−1=3
kÞ. However, the aver-
age entrywise error satisfies the bound
1
N3
kX
ijab X
ki;kj;ka∈K
j½MKtn−TNk
nijabðki;kj;kaÞj ≤CN−1
k:
From the error decomposition in Eq. (12), such a refined
bound is sufficient for our finite-size error analysis.
Therefore, the Madelung constant corrections to the ERI
contractions and the orbital energies effectively reduce the
finite-size error in the overall CCDðnÞenergy calculation
to OðN−1
kÞ.
V. NUMERICAL EXAMPLES
To validate the above theoretical analysis, we conduct
CCD and CCDðnÞcalculations on a 3D periodic system of
hydrogen dimers. One hydrogen dimer is positioned at the
center of each cubic unit cell with an edge length of 6 bohr
in the xdirection and a separation distance of 1.8 bohr. For
each uniform mesh K, we perform an HF calculation on K
to obtain the orbitals and orbital energies. We then perform
CCD and CCDðnÞcalculations under four distinct settings,
each with a different combination of the Madelung constant
correction to the orbital energies and the ERI contractions
in the amplitude equation. Specifically, we compute CCD
and CCDðnÞwith both corrections, the correction to the
orbital energies alone, the correction to the ERI contrac-
tions alone, and without any corrections. All the calcu-
lations are carried out using the
PYSCF
package [46] with a
minimal basis set gth-szv.
Figure 1illustrates the numerical results of the CCD(1),
CCD(2), CCD(3), and converged CCD calculations. For
CCD(1), which is identical to MP2, we have T0¼0, and
the Madelung constant correction to the ERI contractions
has no effect. As a result, the curves for the calculations
with and without this correction are identical. For CCD(2)
and higher, the four correction settings produce distinct
curves. Only the CCDðnÞcalculation with both corrections
exhibits a convergence rate of OðN−1
kÞ, while the other
three calculations have convergence rates of OðN−1=3
kÞ.
This highlights the importance of taking into account the
Madelung constant correction to both the orbital energies
and ERI contractions in higher-level CCDðnÞcalculations.
As CCDðnÞconverges to CCD, the difference between the
calculation with both corrections and the one without any
corrections gradually diminishes to zero, and the finite-size
error satisfies the inverse volume scaling.
2345
-0.02
-0.019
-0.018
-0.017
-0.016
-0.015
-0.014
(a) CCD(1)
2345
-0.032
-0.03
-0.028
-0.026
-0.024
-0.022
-0.02
-0.018
-0.016
(b) CCD(2)
2345
-0.04
-0.035
-0.03
-0.025
-0.02
-0.015
(c) CCD(3)
2345
-0.05
-0.045
-0.04
-0.035
-0.03
-0.025
-0.02
-0.015
(d) CCD
2345
-0.01435
-0.0143
-0.01425
-0.0142
-0.01415
-0.0141
(e) CCD(1)
2345
-0.0205
-0.0204
-0.0203
-0.0202
-0.0201
-0.02
(f) CCD(2)
2345
-0.0231
-0.023
-0.0229
-0.0228
-0.0227
-0.0226
(g) CCD(3)
2345
-0.0251
-0.02505
-0.025
-0.02495
-0.0249
-0.02485
-0.0248
-0.02475
-0.0247
(h) CCD
FIG. 1. Convergence of the CCD(1), CCD(2), CCD(3), and CCD correlation energies for a 3D periodic system of hydrogen dimers
with increasing Nk. The settings are distinguished by the presence or absence of the Madelung constant correction to the orbital energies
(“εNk;ξ
nk”) and the ERI contraction (“ANk;ξ”). In (a)–(d), the dashed curves show the power-law fitting using C0þC1N−1=3
kand data
points N1=3
k¼3, 4, 5 for the two cases with partial Madelung constant correction. (e)–(h) plot the curve fittings using N−1=3
kand N−1
k
over the calculations with correction to both components, numerically corroborating the inverse volume scaling.
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-12
VI. DISCUSSION
Recent years have witnessed significant progresses in
leveraging wave function methods to study solids. One
important driver of this trend is the potential of these methods
to provide systematically improvable results. Among the
simplest post-HF methods, MP2 has been applied to increas-
ingly large periodic systems along with the development of
efficient implementations utilizing density fitting and paral-
lelization techniques [16,23]. Notably, the next-order
method, MP3, has also been applied to periodic systems
recently for the first time [24]. Though computationally
much more expensive, CC theory has also gained momentum
in this context. It is being increasingly employed to com-
pute ground state and band structures for a variety of
solid materials including insulators and metallic systems
[17–19,22,25]. Furthermore, related theories such as equa-
tion-of-motion coupled cluster theory, GW method, and
algebraic diagrammatic construction theory also start their
utilizations in computing excited state properties of solids
[20,21,24]. Because of the omnipresence of finite-size
effects, as wave function methods continue to be increasingly
applied to periodic systems, the importance of developing a
thorough understanding of finite-size effects becomes even
more pronounced.
In this work, we fill a gap in understanding the finite-size
error in periodic coupled cluster calculation for insulating
systems. Notably, we reveal an unexpected inverse volume
scaling of this error in CCD calculation. This behavior
manifests even in the absence of any finite-size correction
schemes, owing to an error cancellation. Our findings,
together with the methodologies employed in this study,
provide valuable insights for practitioners, method devel-
opers, and theorists.
For practitioners, when applying computational quantum
chemistry methods to periodic systems, reducing finite-size
errors using techniques such as power-law extrapolation
requires an in-depth understanding of the error scaling.
This is particularly important when calculations are con-
strained to small-sized systems due to the steep increase of
the computational cost with respect to the system size and
limited resources. Many production-level quantum chem-
istry packages that support periodic systems use certain
finite-size corrections to reduce errors in the HF exchange
energy calculation and the HF orbital energies. For in-
stance, the truncated Coulomb correction scheme [54,55]
can be applied to insulating systems, so that the finite-size
error in orbital energies decays superalgebraically with
respect to Nk. However, our analysis shows that if this
correction is applied only to the orbital energy but without
any correction to ERI contractions, the finite-size error of
the correlation energy deteriorates from inverse volume
scaling to inverse length scaling. This is also a finding
consistent with numerical observations.
For method developers, a key aspect of our result is the
connection between the Madelung constant correction and
the singularity subtraction method. This relationship serves
not just as a crucial element in our theoretical proof but also
points toward new methods for further finite-size error
reduction. The Madelung constant correction operates as a
one-shot correction, while the singularity subtraction
method can be systematically improved to reduce the
finite-size error. Our analysis and correction schemes can
also be extended to more advanced coupled cluster theories
such as CCSD, and coupled cluster singles, doubles, and
triples (CCSDT). Exploring the finite-size error in lower-
dimensional systems, such as MATBG that requires a
different form of the Coulomb kernel [4,56], can lead to
different finite-size scaling patterns and novel correction
schemes. Going beyond finite-size corrections, the singu-
larity structure of the amplitude, as outlined in Lemma 6,
may be of independent interest. The singularity structure
imposes important analytic constraints that should be taken
into account when developing efficient numerical methods
for compressing the CC amplitude tensor.
For theorists, several critical questions persist: How can
finite-size error analysis be integrated with the study of
complete basis limits to tackle basis set dependence? How
should the finite-size error behavior in metallic systems be
analyzed, especially when the orbital energy difference in
the denominator could vanish? Can the scope of finite-size
analysis be broadened to include more complex systems
like disordered systems and finite-temperature alloys?
These questions present a fertile ground for future research.
ACKNOWLEDGMENTS
This material is based upon work supported by the U.S.
Department of Energy, Office of Science, Office of
Advanced Scientific Computing Research and Office of
Basic Energy Sciences, Scientific Discovery through
Advanced Computing (SciDAC) program under Grant
No. DE–SC0022364 (X. X.). This work is also partially
supported by the Applied Mathematics Program of the
U.S. Department of Energy (DOE) Office of Advanced
Scientific Computing Research under Contract No. DE-
AC02-05CH1123 (L. L.). L. L. is a Simons Investigator in
Mathematics (Grant No. 825053). We are grateful to
Timothy Berkelbach, Garnet Chan, and Alexander
Sokolov for insightful discussions and to the anonymous
reviewers for their helpful suggestions, which have been
instrumental in enhancing the presentation and clarity of
our paper.
APPENDIX A: CCD AMPLITUDE EQUATION
In this section, we use a capital letter to denote an index
pair consisting of the orbital index and the kindex. For
instance, I¼ði; kiÞ,J¼ðj; kjÞ,A¼ða; kaÞ, etc. We use
P∈fI; J; K; Lgto refer to occupied orbitals (also known as
holes) and P∈fA; B; C; Dgto refer to unoccupied orbitals
(also known as particles). Any summation PPrefers to
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-13
summing over all occupied or virtual orbital indices p
and all momentum vectors kp∈K, while the crystal
momentum conservation is enforced according to the
summand. This notation is used only in this section to
simplify the notation and also to connect the equations to
those in the molecular case [9] for better readability.
Using a finite mesh Kof size Nk, the normalized
CCD amplitude TNk
¼fTijab ðki;kj;kaÞg ≔ftAB
IJ gwith
ki;kj;ka∈Kis defined as the solution of the amplitude
equation
tAB
IJ ¼1
εNk
IJAB
½ANkðTNk
ÞIJAB
¼1
εNk
IJAB hABjIJiþPX
C
κA
CtCB
IJ −X
K
κK
ItAB
KJþ1
NkX
KL
χKL
IJ tAB
KL þ1
NkX
CD
χAB
CDtCD
IJ
þP1
NkX
KC
ð2χAK
IC −χAK
CI ÞtCB
KJ −χAK
IC tBC
KJ −χAK
CJ tBC
KI ;∀I; J; A; B; ðA1Þ
where εNk
IJAB ¼εNk
ikiþεNk
jkj−εNk
aka−εNk
bkb,½ANkðTÞIJAB ¼½ANkðTÞijabðki;kj;kaÞ, and Pis a permutation operator defined
as Pð ÞAB
IJ ¼ðÞ
AB
IJ þðÞ
BA
JI . This reformulation of the CCD amplitude equation is derived from the CCSD amplitude
equation in Ref. [30] by removing all the terms related to single amplitudes and normalizing the involved ERIs and
amplitudes (which gives the extra 1=Nkfactor in the equation and the intermediate blocks). The intermediate blocks in the
equation are defined as
κA
C¼−
1
N2
kX
KLD
ð2hKLjCDi−hKLjDCiÞtAD
KL;
κK
I¼1
N2
kX
LCD
ð2hKLjCDi−hKLjDCiÞtCD
IL ;
χKL
IJ ¼hKLjIJiþ 1
NkX
CD
hKLjCDitCD
IJ ;
χAB
CD ¼hABjCDi;
χAK
IC ¼hAKjICiþ 1
2NkX
LD
ð2hLKjDCi−hLKjCDiÞtAD
IL −hLKjDCitDA
IL ;
χAK
CI ¼hAKjCIi−1
2NkX
LD
hLKjCDitDA
IL ;
and their momentum vector indices also assume the crystal momentum conservation
κQ
P→kp−kq∈L;
χRS
PQ →kpþkq−kr−ks∈L:
In the TDL, the amplitude equation for the exact amplitude t¼ftijab ðki;kj;kaÞg ≔ftAB
IJ gas functions of
ki;kj;ka∈Ωcan be formulated by letting Kin Eq. (A1) converge to Ωas
tAB
IJ ¼1
εTDL
IJAB
½ATDLðtÞIJAB
¼1
εTDL
IJAB hABjIJiþPX
C
κA
CtCB
IJ −X
K
κK
ItAB
KJþ1
jΩjZΩ
dkkX
kl
χKL
IJ tAB
KL þ1
jΩjZΩ
dkcX
cd
χAB
CDtCD
IJ
þP1
jΩjZΩ
dkkX
kc
ð2χAK
IC −χAK
CI ÞtCB
KJ −χAK
IC tBC
KJ −χAK
CJ tBC
KI ;∀I; J; A; B; ðA2Þ
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-14
where the intermediate blocks in the TDL are defined as
κA
C¼−
1
jΩj2ZΩ×Ω
dkkdklX
kld
ð2hKLjCDi−hKLjDCiÞtAD
KL;
κK
I¼1
jΩj2ZΩ×Ω
dkcdkdX
lcd
ð2hKLjCDi−hKLjDCiÞtCD
IL ;
χKL
IJ ¼hKLjIJiþ 1
jΩjZΩ
dkcX
cd
hKLjCDitCD
IJ ;
χAB
CD ¼hABjCDi;
χAK
IC ¼hAKjICiþ 1
2jΩjZΩ
dklX
ld
ð2hLKjDCi−hLKjCDiÞtAD
IL −hLKjDCitDA
IL ;
χAK
CI ¼hAKjCIi−1
2jΩjZΩ
dklX
ld
hLKjCDitDA
IL :
APPENDIX B: PROOF OF THEOREM 1
Restatement of Theorem 1. In CCDðnÞcalculation, the
finite-size error in the correlation energy scales as
OðN−1=3
kÞin each of the following scenarios: (i) there is
no finite-size correction, (ii) the Madelung constant cor-
rection is applied only to the ERI contraction ANk, and
(iii) the Madelung constant correction is applied only to the
orbital energy εNk
nk.
When the Madelung constant correction is applied to
both ANkand εNk
nkin the CCDðnÞcalculation, the overall
finite-size error scales as OðN−1
kÞ.
As a special case, the same conclusion applies to MP3
calculations.
1. Proof outline
The main context of this paper has provided a brief
description of the main proof idea. In this proof, we recap
some of the equations and concepts discussed before to
make it self-contained. The proofs of Theorem 1 for
CCDðnÞcalculations with various types of corrections
are based on the error splitting in Eqs. (12) and (13), i.e.,
ETDL
CCDðnÞ−ENk
CCDðnÞ≤jGTDL ðtnÞ−GNkðMKtnÞj
þCkMKtn−TNk
nk1;ðB1Þ
MKtn−TNk
n¼1
εTDL ½MKATDLðtn−1Þ−ANkðMKtn−1Þ
þANkðMKtn−1Þ1
εTDL −
1
εNk
þ1
εNk½ANkðMKtn−1Þ−ANkðTNk
n−1Þ:ðB2Þ
Note that the amplitude error here is measured in the
average norm as
kTk1¼1
N3
kX
ki;kj;ka∈KX
ijab
jTijabðki;kj;kaÞj:
The finite-size error in CCDðnÞcalculation is, thus,
decomposed into the error in energy calculation using
exact amplitude, the error in ERI contractions, the error in
orbital energies, and the error accumulated from previous
iteration. The latter three errors together make up of the
error in amplitude calculation. The error in energy calcu-
lation using exact amplitude is studied previously in
Ref. [41]. For completeness, we provide a brief review
of the main results below.
a. Brief review of error in energy calculation
with exact amplitude
One basic observation is that this error in CCDðnÞ
calculation can be interpreted as the quadrature error of
a specific trapezoidal rule as
jGTDLðtnÞ−GNkðMKtnÞj
¼
1
jΩj3EΩ×Ω×ΩX
ijab
ðWijab½tnijab Þ;K×K×K
:
ðB3Þ
Since both Wijab and ½tnijab are periodic with respect to
ki;kj;ka∈Ω, the asymptotic scaling of this quadrature
error depends on the smoothness of these two components
that constitute the integrand.
A general ERI hn1k1;n
2k2jn3k3;n
4k4ican be viewed
as a function of momentum vectors k1and k2and its
momentum transfer vector q¼k3−k1. This function is
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-15
periodic with respect to each variable in Ω, and its
definition in Eq. (1) can be decomposed as
4π
jΩj
ˆ
ϱn1k1;n3ðk1þqÞð0Þˆ
ϱn2k2;n4ðk2−qÞð0Þ
jqj2þ4π
jΩjX
G∈Lnf0g
jqþGj2:
ðB4Þ
The numerators of all the fractions are smooth with respect
to k1,k2, and q[note the assumption that ψnkðrÞis smooth
with respect to k]. Therefore, this ERI example is smooth
with respect to k1and k2and has one point singularity in
Ωwith respect to qat q¼0, which is due to the first
fraction term. For any fixed kiand kj, this point singularity
can be characterized using the concept of algebraic
singularity of certain orders.
Definition 8 (algebraic singularity for univariate func-
tions). A function fðxÞhas algebraic singularity of order
γ∈Rat x0∈Rdif there exists δ>0such that
∂α
∂xαfðxÞ
≤Cα;δjx−x0jγ−jαj;∀0<jx−x0j<δ;∀α≥0;
where the constant Cα;δdepends on δand the non-negative
d-dimensional derivative multi-index α. For brevity, fis
also said to be singular at x0of order γ.
The numerator of the first fraction in Eq. (B4) scales as
OðjqjsÞwith s∈f0;1;2gnear q¼0using the orbital
orthogonality. The value of sdepends on the relation
between orbital indices ðn1;n
2Þand ðn3;n
4Þ. As a result,
the algebraic singularity of the ERI example above at
q¼0with any fixed k1and k2has order in f−2;−1;0g.
In addition, to connect this singularity with varying
k1;k2∈Ω, we introduce the algebraic singularity with
respect to one variable for a multivariate function.
Definition 9 (algebraic singularity for multivariate
functions). A function fðx;yÞis smooth with respect to
y∈VY⊂Rdyfor any fixed xand has algebraic singularity
of order γwith respect to xat x0∈Rdxif there exists δ>0
such that
∂α
∂xα∂β
∂yβfðx;yÞ
≤Cα;β;δjx−x0jγ−jαj;
∀0<jx−x0j<δ;∀y∈VY;∀α;β≥0;
where constant Cα;β;δdepends on δ,α, and β. Compared to
the univariate case in Definition 4, the key additions are the
shared algebraic singularity of partial derivatives over yat
x¼x0of order γand the independence of Cα;β;δon y∈VY.
By this definition, the ERI hn1k1;n
2k2jn3k3;n
4k4iis
smooth everywhere with respect to ki;kj;q∈Ωexcept
at q¼0, and the singularity order γ∈f−2;−1;0g.
Specifically, γequals to −2,−1, and 0, respectively, when
the orbital indices are fully matched, i.e., n1¼n3;n
2¼n4,
partially matched, i.e., n1¼n3;n2≠n4or n1≠n3;n2¼n4,
and not matched, i.e., n1≠n3;n
2≠n4. If treating the ERI
example as a function of k1,k2,k3instead, we equiv-
alently claim that the function is singular at k1¼k3of
order γ.
One key result in Ref. [41] is the singularity structure
characterization for the exact CCDðnÞamplitude tn,
which is essential for estimating the quadrature error in
Eq. (B3). It turns out that each exact amplitude entry
½tnijabðki;kj;kaÞwith any n>0has one point of
algebraic singularity of order 0 at ka−ki¼0, sharing a
similar singularity structure as the ERI hiki;jkjjaka;
bkbior the exact MP2/CCD(1) amplitude entry
ðεTDL
iki;jkj;aka;bkbÞ−1haka;bkbjiki;jkji.
Restatement of Lemma 6 (singularity structure of the
amplitude, Lemma 4 in Ref. [41]). In CCDðnÞcalculation
with n>0, each entry of the exact double amplitude tn
belongs to the following function space:
TðΩÞ¼ffðki;kj;kaÞ∶fis periodic with respect to ki;kj;ka∈Ω;
fis smooth everywhere except at ka¼kiwith algebraic singularity of order 0;
fis smooth with respect to ki;kjat the singularity ka¼kig:
Based on the above singularity structures of ERIs and
exact amplitudes, the integrand in the energy calculation in
Eq. (B3) consists of products of periodic functions where
each has one point singularity of order 0. (Recall that Wijab
is the antisymmetrized ERI and consists of two ERIs that
can be treated separately.) We provide a sharp quadrature
error bound for trapezoidal rules over periodic functions in
such a product form, and its application to Eq. (B3) gives
the following lemma.
Lemma 10 (energy error with exact amplitude, Lemma 5
in Ref. [41]). In CCDðnÞcalculation with any n>0, the
finite-size error in the energy calculation using the exact
amplitude tncan be estimated as
jGTDLðtnÞ−GNkðMKtnÞj ≤CN−1
k;
where the constant Cdepends on tn.
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-16
b. Error in amplitude calculation
Based on the error splitting in Eq. (B2), analysis of the
error in amplitude calculation is reduced to estimating the
two main error terms MKATDLðtÞ−ANkðMKtÞand
εTDL
nk−εNk
nkand understanding how ANkamplifies the
amplitude error MKtn−1−TNk
n−1from the previous iteration.
Like the energy calculation, the errors in ERI contrac-
tions and orbital energies consist of specific quadrature
errors. In both cases, we can show that the Madelung
constant correction is connected to certain singularity
subtraction methods and can significantly reduce the
corresponding dominant quadrature errors.
Lemma 11 (error in ERI contractions). In CCDðnÞ
calculation with any n>0, the finite-size error in the
entries of the ERI contractions using the exact amplitude tn
without any corrections can be bounded as
½MKATDL ðtnÞ−ANkðMKtnÞijabðki;kj;kaÞ≤CN−1=3
k:
ðB5Þ
The Madelung constant correction reduces this error to
½MKATDL ðtnÞ−ANk;ξðMKtnÞijabðki;kj;kaÞ
≤C(1
jqj2N−1
kq≠0;
N−1=3
kq¼0;
ðB6Þ
where q¼ka−kiþG0with G0∈Lchosen such that
q∈Ω. In both cases, the constant Cdepends on tnbut not on
the entry index ði; j; a; bÞand ðki;kj;kaÞ∈K×K×K.
Remark 12. The prefactor 1=jqj2in the above estimate is
important when q∈Kqis near the origin. For example, if
ki;ka∈Kare adjacent to each other, jqjis of scale
OðN−1=3
kÞand the estimate in Lemma 11 suggests an error
bound of scale OðN−1=3
kÞ.
Lemma 13 (error in orbital energies). The finite-size
error in orbital energies without any corrections is
bounded as
εTDL
nk−εNk
nk≤CN−1=3
knis occupied;
N−1
knis virtual:
The Madelung constant correction reduces this error to
εTDL
nk−εNk;ξ
nk≤CN−1
k:
In both cases, the constant Cis independent of the entry
index n; k∈K.
Note that there are three multipliers for the three error
terms in the amplitude error splitting Eq. (B2). These
prefactors are bounded by constants independent of Nk.
First, the orbital energy difference in Eq. (B2) satisfies
jεTDL
iki;jkj;aka;bkbj≥2εgby the assumed indirect gap
εTDL
aka−εTDL
iki≥εg>0. Second, since ANkðTÞconsists of
constant, linear, and quadratic terms of T, a straightforward
estimate shows (e.g., using Lemma 7 in Ref. [41]) that
max
ijab;ki;kj;ka∈Kj½ANkðMKtnÞijabðki;kj;kaÞj
≤Cmax
ijab k½tnijabk2
L∞ðΩ×Ω×ΩÞ;
where the L∞-norm of ½tnijab is finite according to
Lemma 6.
Based on the above estimates of these multipliers, we
have that the summation of the first two error terms in
Eq. (B2) is dominated by the error in the ERI contractions
and the orbital energies discussed in Lemmas 11 and 13. If
applying the Madelung constant correction to both orbital
energies and ERI contractions, the summation of the first
two error terms has its entries bounded asymptotically the
same as Eq. (B6). Otherwise, its entries are bounded
asymptotically as in Eq. (B5). Lastly, for the third error
term in Eq. (13), the application of ANk=ANk;ξcan be
proved to maintain the entrywise error scaling obtained in
Lemma 11.
Lemma 14. Consider two arbitrary bounded amplitude
tensors T;S ∈Cnocc ×nocc ×nvir×nvir ×Nk×Nk×Nkand assume their
entrywise upper bound independent of Nk. If the difference
between Tand Ssatisfies an estimate similar to that in
Eq. (B5), i.e.,
j½T−Sijabðki;kj;kaÞj ≤CN−1=3
k;
the ERI-contraction map without any corrections ANk
satisfies
j½ANkðTÞ−ANkðSÞijabðki;kj;kaÞj ≤CN−1=3
k:
If the difference between Tand Ssatisfies an estimate
similar to that in Eq. (B6), i.e.,
j½T−Sijabðki;kj;kaÞj ≤C(1
jqj2N−1
kq≠0;
N−1=3
kq¼0;
the ERI-contraction map with Madelung constant correc-
tion ANk;ξsatisfies
j½ANk;ξðTÞ−ANk;ξðSÞijabðki;kj;kaÞj ≤C(1
jqj2N−1
kq≠0;
N−1=3
kq¼0:
Combining the estimates of the three error terms in
Eq. (13) and the initial condition MKt0¼TNk
0¼0in both
cases with and without Madelung constant correction, we
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-17
can obtain the entrywise estimate of the error in the
CCDðnÞamplitude calculation recursively. First, for the
CCDðnÞcalculation without any corrections or with partial
Madelung constant corrections to either orbital energies or
ERI contractions, we have
j½MKtn−TNk
nijabðki;kj;kaÞj ≤CN−1=3
k:ðB7Þ
Accordingly, the average entrywise error can be esti-
mated as
kMKtn−TNk
nk1
¼1
N3
kX
ki;kj;ka∈KX
ijab
j½MKtn−TNk
nijabðki;kj;kaÞj
≤CN−1=3
k:
Second, for the CCDðnÞcalculation with the Madelung
constant correction to both orbital energies and ERI
contractions, we have
j½MKtn−TNk
nijabðki;kj;kaÞj ≤C(1
jqj2N−1
kq≠0;
N−1=3
kq¼0:
ðB8Þ
Accordingly, the average entrywise error can be esti-
mated as
kMKtn−TNk
nk1
¼1
N3
kX
ki;kj∈KX
q∈KqX
ijab
j½MKtn−TNk
nijabðki;kj;kiþqÞj
≤C1
NkX
q∈Kqnf0g
1
jqj2N−1
kþC1
Nk
N−1=3
k
≤CN−1
k:
Plugging the above estimate and the error estimate for
energy calculation in Lemma 10 into Eq. (12) then finishes
the proof of Theorem 1.
Remark 15. In our previous work [41], we use the
maximum entrywise norm to characterize the finite-size
error in the amplitude calculation, which loosens the
average entrywise norm used in the error splitting in
Eq. (B1) as
kTk1≤kTk∞≔max
ijab;ki;kj;ka∈KjTijabðki;kj;kaÞj:
Without the Madelung constant correction, the maximum
norm suffices since all entries in the amplitude error are of
the same scale, as shown in Eq. (B7). However, this norm is
no longer sufficient for the calculation with the Madelung
constant correction, bounded in Eq. (B8). Because the
maximum entrywise error is now of scale OðN−1=3
kÞ, while
most entries are of scale OðN−1
kÞ. In this case, the average
entrywise norm provides a necessary and tighter estimate of
amplitude error.
2. Proof of Lemma 11: Error in ERI contractions
According to the singularity structure of exact CCDðnÞ
amplitude in Lemma 6, we consider the ERI contraction
using an arbitrary exact amplitude t∈TðΩÞnocc×nocc ×nvir×nvir .
Fixing a set of entry index ði; j; a; bÞand ðki;kj;kaÞ∈
K×K×K, the error in the indexed ERI-contraction
entry can be detailed as [by comparing Eqs. (A1)
and (A2)]
½MKATDLðtÞ−ANk;ξðMKtÞijab;kikjka¼1
jΩjEΩX
kl
hkkk;lkljiki;jkjitklabðkk;kl;kaÞ;KþjΩjξtijab ðki;kj;kaÞ
þ1
jΩj2EΩ×ΩX
klcd
hkkk;lkljckc;dkditijcdðki;kj;kcÞtklabðkk;kl;kaÞ;K×K
þ;ðB9Þ
where the constant terms cancel with each other, the first
term is the error in the 4h2plinear term calculation with the
Madelung constant correction, and the second term is the
error in the 4h2pquadratic term calculation. The neglected
ones are the errors in remaining linear and quadratic term
calculations, which can all be similarly formulated as
quadrature errors of specific trapezoidal rules.
In the error analysis for CCD calculation without
corrections in ERI contractions [41], it has been shown
that without the Madelung constant correction the error
entry in the ERI contractions is uniformly bounded as
½MKATDL ðtÞ−ANkðMKtÞijab;kikjka≤CN−1=3
k:ðB10Þ
More specifically, all the quadratic terms and part of the
linear terms that contain only ERIs with mismatched orbital
indices contribute at most OðN−1
kÞerrors in both Eqs. (B9)
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-18
and (B10). The dominant error in Eq. (B10) comes from the
calculation of the following six linear terms:
1
NkX
kk∈KX
kl
hkkk;lkljiki;jkjitklabðkk;kl;kaÞ;
1
NkX
kc∈KX
cd
haka;bkbjckc;dkditijcdðki;kj;kcÞ;
−
1
NkX
kk∈KX
kc
haka;kkkjckc;ikiitkjcbðkk;kj;kcÞ;
−
1
NkX
kk∈KX
kc
hbkb;kkkjckc;jkjitkicaðkk;ki;kcÞ;
−
1
NkX
kk∈KX
kc
haka;kkkjckc;jkjitkibcðkk;ki;kbÞ;
−
1
NkX
kk∈KX
kc
hbkb;kkkjckc;ikiitkjacðkk;kj;kaÞ;
which contain ERIs with fully or partially matched orbital
indices. The Madelung constant correction in Eq. (B9) is
only triggered in these six terms. In this proof, we focus on
the error estimate for the 4h2plinear term (the first term
above) with the correction, and similar analysis can be
applied to all the other five terms.
Denote the ERI-amplitude product in the 4h2plinear
term with orbital indices ðk; lÞas
Fklðq1Þ¼hkðki−q1Þ;lðkjþq1Þjiki;jkji
×tklabðki−q1;kjþq1;kaÞ
¼Hkl
eriðq1ÞHkl
ampðq1Þ;
where q1¼ki−kkis the momentum transfer vector of the
ERI. The 4h2plinear term calculation with the Madelung
constant correction using a finite mesh Kcan be reformu-
lated as
1
NkX
kk∈KX
kl
Fklðki−kkÞ−ξHij
ampð0Þ
¼1
NkX
q1∈KqX
kl
Fklðq1Þ−ξHij
ampð0Þ;
using the change of variable kk→ki−q1and the perio-
dicity of Fklðq1Þ. The error of this calculation compared to
its TDL value, i.e., the first error term in Eq. (B9), can be
written as
1
jΩjX
kl
EΩðFklðq1Þ;KqÞþjΩjξHij
ampð0Þ:
Previously, in Ref. [41], the quadrature error for Fklðq1Þ
with varying ðk; lÞis estimated as
jEΩðFklðq1Þ;KqÞj ≤C8
>
>
<
>
>
:
N−1
kk≠i; l ≠j;
N−2=3
kk¼i; l ≠jor k≠i; l ¼j;
N−1=3
kk¼i; l ¼j;
∀i; j; a; b; ∀ki;kj;ka∈K:ðB11Þ
As demonstrated next, in the case of partially matched orbital indices (e.g., k¼i,l≠j), this error estimate turns out to be
loose when ki≠kaand can be further improved as
jEΩðFilðq1Þ;KqÞj ≤C(1
jqjN−1
kq≠0;
N−2=3
kq¼0;
∀i; j; a; b; ∀ki;kj;ka∈K;ðB12Þ
where q¼ka−kiþG0with G0∈Lchosen to make q∈Ω. In the case of fully matched orbital indices (k¼i,l¼j),
the Madelung constant correction is triggered in the ERI evaluation and can help remove the leading quadrature error when
ki≠kaas
jEΩðFijðq1Þ;KqÞþjΩjξHij
ampð0Þj ≤C(1
jqj2N−1
kq≠0;
N−1=3
kq¼0;
∀i; j; a; b; ∀ki;kj;ka∈K:ðB13Þ
In the following discussion, we prove these two error estimates Eqs. (B12) and (B13) respectively at ki≠ka.
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-19
a. Error estimate for linear terms with partially
matched orbital indices
Consider a fixed set of ði; j; a; bÞand ki;kj;ka∈Kwith
ki≠ka. Assume k¼iand l≠j. Our target is a sharper
estimate of EΩðFilðq1Þ;KqÞ. Since Kqand Ωare both
inversely symmetric over q1¼0, the quadrature error can
be symmetrized as
EΩðFilðq1Þ;KqÞ¼1
2EΩðHil
eriðq1ÞHil
ampðq1Þ
þHil
erið−q1ÞHil
ampð−q1Þ;KqÞ:
This symmetrized integrand can be further decomposed
into two terms:
½ðHil
eriðq1ÞþHil
erið−q1ÞÞHil
ampðq1Þ
−½Hil
erið−q1ÞðHil
ampðq1Þ−Hil
ampð−q1ÞÞ:ðB14Þ
For the first term in Eq. (B14), we note that the ERI term
can be detailed as
Hil
eriðq1Þ¼ 4π
jΩj
ˆ
ϱiðki−q1Þ;ikið0Þˆ
ϱlðkjþq1Þ;jkjð0Þ
jq1j2
þ4π
jΩjX
G∈Lnf0g
jq1þGj2:
The ERI nonsmoothness with q1∈Ωcomes from the first
fractional term whose numerator is smooth and scales as
Oðjq1jÞ near q1¼0using orbital orthogonality. It can be
verified directly that Hil
eriðq1Þis singular at q1¼0of order
−1, and after symmetrization Hil
eriðq1ÞþHil
erið−q1Þis
singular at q1¼0of order 0. Meanwhile, the amplitude
Hil
ampðq1Þis smooth everywhere in Ωexcept at q1¼−qof
order 0 according to Lemma 6. An error estimate lemma in
Ref. [41] (restated as Lemma 24 in Appendix D) provides a
quadrature error estimate for periodic functions in such a
product form and can show that
jEΩððHil
eriðq1ÞþHil
erið−q1ÞÞHil
ampðq1Þ;KqÞj ≤CN−1
k;
where constant Cis independent of i,j,a,band
ki;kj;ka∈K.
For the second term in Eq. (B14), direct application of
Lemma 24 leads to error estimate of scale OðN−2=3
kÞ, since
Hil
erið−q1Þis singular at q1¼0of order −1. However, note
that Hil
ampðq1Þis smooth at q1¼0[recall that Hil
ampðq1Þis
singular only at q1¼−q≠0], and, thus, the subtraction
Hil
ampðq1Þ−Hil
ampð−q1Þscales as Oðjq1jÞ near q1¼0.
Multiplication by this extra Oðjq1jÞ term improves the
algebraic singularity of Hil
eriðq1Þat q1¼0, and the overall
product can be shown to be singular at q1¼0of order 0.
Intuitively, this improved algebraic singularity at q1¼0
can lead to asymptotically smaller quadrature errors. To
rigorously justify this statement, we generalize Lemma 24
to estimate the quadrature error for this special case.
Lemma 16. Let fðxÞ¼f1ðxÞf2ðxÞ, where f1ðxÞand
f2ðxÞare periodic with respect to V¼½−1
2;1
2dand
(i) f1ðxÞis smooth everywhere except at x¼z1¼0of
order γ≤−1,
(ii) f2ðxÞis smooth everywhere except at x¼z2≠0of
order 0,
(iii) ∂α
xf2ð0Þ¼0for any derivative order jαj≤s.
Assume γ>−dfor fðxÞto be integrable in Vand γþ
sþ1≤0so the leading algebraic singularity of fðxÞis at
x¼0. Consider an md-sized uniform mesh Xin V.
Assume that Xsatisfies that z1,z2are either on the mesh
or Θðm−1Þaway from any mesh points and mis sufficiently
large that jz1−z2j¼Ωðm−1Þ.Atx¼z1and x¼z2,fðxÞ
is set to 0. The trapezoidal rule using Xhas quadrature error
jEVðf; XÞj ≤CHdþ1
V;z1ðf1ÞHdþ1
V;z2ðf2Þjz1−z2j−ðsþ1Þ
×m−ðdþγþsþ1Þ:
Remark 17. The two factors Hdþ1
V;z1ðf1Þand Hdþ1
V;z2ðf2Þ
characterize the algebraic singularities of the two functions,
and their exact definition can be found in Appendix D.
Proof of Lemma 16 is provided in Appendix D2.
In order to utlize this result, we further decompose the
second term in Eq. (B14) into
½Hil
erið−q1ÞðHil
ampðq1Þ−Hil
ampð0ÞÞ
þ½Hil
erið−q1ÞðHil
ampð0Þ−Hil
ampð−q1ÞÞ:
Applying Lemma 16 to both terms with γ¼−1and s¼0
gives
jEΩðHil
erið−q1ÞðHil
ampðq1Þ−Hil
ampð−q1ÞÞ;KqÞj ≤C1
jqjN−1
k;
where constant Cis independent of i,j,a,band
ki;kj;ka∈K. Combining the above quadrature error
estimates for the two terms in Eq. (B14), we prove a
tighter error estimate at q≠0shown in Eq. (B12) while the
previous result in Eq. (B11) at q¼0still holds.
b. Error estimate for linear terms with fully
matched orbital indices
Let k¼i,l¼jand ki≠kaand consider the corre-
sponding calculation in the 4h2plinear term. The
Madelung constant correction is applied in the ERI
evaluation at kk¼kior, equivalently, at q1¼0, and the
calculation can be written as
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-20
1
NkX
q1∈Kq
Fijðq1Þ−ξHij
ampð0Þ
¼1
NkX
q1∈KqHij
eriðq1ÞHij
ampðq1Þ−hσðq1ÞHij
ampð0Þ
þ1
jΩjZΩ
dq1hσðq1ÞHij
ampð0ÞþOðN−1
kÞ;
which uses the expansion of the Madelung constant with an
arbitrary fixed parameter σ>0in Eq. (20). Note that the
prefactor in the OðN−1
kÞremainder term above depends
only on σ. The right-hand side of the above reformation is
equivalent to a singularity subtraction method that decom-
poses the original integrand into two terms:
Fijðq1Þ¼Hij
eriðq1ÞHij
ampðq1Þ
¼Hij
eriðq1ÞHij
ampðq1Þ−hσðq1ÞHij
ampð0Þ
þhσðq1ÞHij
ampð0Þ;
and then computes the numerical quadrature of the first
term and the exact integral of the second term. The
Madelung-corrected calculation, thus, has quadrature error
only from the first term as
1
jΩjEΩðFijðq1Þ;KqÞþξHij
ampð0Þ
¼1
jΩjEΩHij
eriðq1ÞHij
ampðq1Þ−hσðq1ÞHij
ampð0Þ;Kq
þOðN−1
kÞ:
Following a similar approach as in the partially matched
case, the effective integrand after the Madelung constant
correction above can be split into two terms as
ðHij
eriðq1Þ−hσðq1ÞÞHij
ampðq1Þ
þ½hσðq1ÞðHij
ampðq1Þ−Hij
ampð0ÞÞ:ðB15Þ
For the first term in Eq. (B15), the subtraction part can be
detailed as
Hij
eriðq1Þ−hσðq1Þ
¼4π
jΩj
ˆ
ϱiðki−q1Þ;ikið0Þˆ
ϱjðkjþq1Þ;jkjð0Þ−e−σjq1j2
jq1j2
þ4π
jΩjX
G∈Lnf0g
jq1þGj2;
where the numerator of the first fraction scales as Oðjq1jÞ
near q1¼0. This subtraction is, thus, singular at q1¼0of
order −1and shares a similar form as the ERI with partially
matched orbital indices. Using the earlier inversion sym-
metry analysis for partially matched case, the quadrature
error of the first term in Eq. (B15) when ki≠kacan be
estimated as
EΩððHij
eriðq1Þ−hσðq1ÞÞHij
ampðq1Þ;KqÞ≤C1
jqjN−1
k:
For the second term in Eq. (B15), we exploit the
inversion symmetry of Kqand Ωagain, and its quadrature
error equals to that of its symmetrized version as
1
2hσðq1ÞðHij
ampðq1Þ−2Hij
ampð0ÞþHij
ampð−q1ÞÞ:
This formula uses hσðq1Þ¼hσð−q1Þafter symmetrization.
Note that hσðq1Þis singular at q1¼0of order −2, while the
term in the parentheses scales as Oðjq1j2Þnear q1¼0using
the smoothness of Hij
ampðq1Þat q1¼0. Therefore, the overall
function above is singular at q1¼0of order 0. To fit the
integrand form in Lemma 16, we further decompose the
symmetrized integrand above into
½hσðq1ÞðHij
ampðq1Þ−Hij
ampð0Þ−q1·∇Hij
ampð0ÞÞ
þ½hσðq1ÞðHij
ampð−q1Þ−Hij
ampð0Þ−ð−q1Þ·∇Hij
ampð0ÞÞ:
Applying Lemma 16 to these two terms separately with γ¼
−2and s¼1gives
EΩ1
2hσðq1ÞðHij
ampðq1Þ−2Hij
ampð0ÞþHij
ampð−q1ÞÞ;Kq
≤C1
jqj2N−1
k:
Combining the above quadrature error estimates for the
two terms in Eq. (B15), we prove a tighter error estimate at
q≠0shown in Eq. (B13) while the previous result in
Eq. (B11) at q¼0still holds.
The same error estimate can be obtained for all the six
linear terms that trigger the Madelung constant correction,
and the remaining linear and quadratic terms contribute at
most OðN−1
kÞquadrature error. Gathering the error esti-
mates for all these terms together and plugging into
Eq. (B9), we finish the proof.
3. Proof of Lemma 13: Error in orbital energies
In the TDL, the orbital energy εNk
nkwith any fixed nand
k∈Kconverges to
εTDL
nk¼hnkj^
H0jnkiþ 1
jΩjZΩ
dkiX
i
ð2hiki;nkjiki;nki
−hiki;nkjnk;ikiiÞ:
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-21
Comparing this exact orbital energy with its finite-size
calculation in Eq. (2), the finite-size error without any
corrections can be written as
εTDL
nk−εNk
nk¼2
jΩjEΩX
i
hiki;nkjiki;nki;K
−
1
jΩjEΩX
i
hiki;nkjnk;ikii;K:ðB16Þ
For the first quadrature error in Eq. (B16), i.e., the finite-
size error in the direct term, the ERI with each ican be
specified as
hiki;nkjiki;nki¼4π
jΩjX
G∈Lnf0g
ˆ
ρiki;ikiðGÞˆ
ρnk;nkð−GÞ
jGj2:
Note that the momentum transfer vector of this ERI is
always zero and the singular fraction term in this ERI is set
to 0 by definition. As a result, the integrand is periodic and
smooth with respect to ki∈Ω. Therefore, the quadrature
error of the direct term calculation, thus, decays super-
algebraically according to Lemma 20 as
EΩX
i
hiki;nkjiki;nki;K
≤ClN−l
k;∀l>0:
For the second quadrature error in Eq. (B16), i.e., the
finite-size error in the exchange term, the ERI with each i
can be written as
hiki;nkjnk;ikii¼4π
jΩj
ˆ
ρiki;nkð0Þˆ
ρnk;ikið0Þ
jqj2
þ4π
jΩjX
G∈Lnf0g
ˆ
ρiki;ikiðGÞˆ
ρnk;nkð−GÞ
jqþGj2;
which is singular at q≔k−ki¼0of order −2when n¼
iand 0 otherwise. An error estimate lemma in Ref. [41]
(restated as Lemma 22 in Appendix D) gives a tight
quadrature error estimate for such periodic functions with
one point of algebraic singularity, and its application to the
above integrand gives
jEΩðhiki;nkjnk;ikii;KÞj ≤C(N−1=3
kn¼i;
N−1
kn≠i:
Combining the estimates of the two error terms in
Eq. (B16), we obtain the overall finite-size error estimate
for orbital energies without any corrections as
εTDL
nk−εNk
nk≤C(N−1=3
knis occupied;
N−1
knis virtual:
From the above analysis, the dominant finite-size error in
an occupied orbital energy lies in the calculation of the
exchange term with i¼n. In the Madelung-corrected
orbital energy εNk;ξ
nk, the correction is applied to this
exchange term as (ignoring the prefactor −jΩj−1)
QΩðhnki;nkjnk;nkii;KÞ→QΩðhnki;nkjnk;nkii;KÞ
−jΩjξ:
Applying the change of variable ki→k−qand using the
periodicity of the ERI with respect to ki, this corrected
calculation can be reformulated as
QΩðhnki;nkjnk;nkii;KÞ−jΩjξ
¼QΩðhnðk−qÞ;nkjnk;nðk−qÞi −hσðqÞ;KqÞ
þZΩ
dq1hσðq1ÞþOðN−1
kÞ;
using the expansion of the Madelung constant with any
fixed σ>0in Eq. (2). The quadrature error after the
correction can be written as
EΩðhnki;nkjnk;nkii;KÞþjΩjξ
¼EΩðhnðk−qÞ;nkjnk;nðk−qÞi −hσðqÞ;KqÞ
þOðN−1
kÞ:
The effective integrand above after the correction can be
detailed as
hnðk−qÞ;nkjnk;nðk−qÞi −hσðqÞ
¼4π
jΩj
jˆ
ϱnðk−qÞ;nkð0Þj2−e−σjqj2
jqj2þ4π
jΩjX
G∈Lnf0g
jqþGj2:
The integrand singularity comes from the first fractional
term and is of order −1, since the ERI and hσðqÞshare the
same leading singular term. Similar to the analysis in the
ERI contraction, we can then combine this singularity
subtraction with the inverse symmetry of Ωand Kqto
show that
jEΩðhnðk−qÞ;nkjnk;nðk−qÞi −hσðqÞ;KqÞj ≤CN−1
k:
Combining this estimate with the above estimate for all
the remaining direct and exchange terms in εNk;ξ
nk, we obtain
the finite-size error estimate for orbital energies with the
Madelung constant correction as
εTDL
nk−εNk;ξ
nk≤CN−1
k:
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-22
4. Proof of Lemma 14: Error from previous iteration
Fixing a set of entry index ði; j; a; bÞand ðki;kj;kaÞ∈K×K×K, the ERI-contraction entry ½ANk;ξðTÞijab;kikjkacan
be detailed as
½ANk;ξðTÞijabðkikjkaÞ¼haka;bkbjiki;jkjiþ 1
NkX
kk∈KX
kl
hkkk;lkljiki;jkjiTklabðkk;kl;kaÞ
þ1
N2
kX
kkkc∈KX
klcd
hkkk;lkljckc;dkdiTijcdðki;kj;kcÞTklab ðkk;kl;kaÞ
þþ2ξTijabðki;kj;kaÞ;
where the neglected terms are all the other linear and quadratic terms and the Madelung constant corrections to different
terms are collected together at the end.
In the subtraction ANk;ξðTÞ−ANk;ξðSÞ, the constant terms cancel each other. The subtraction between the two Madelung
constant correction terms can be estimated directly as
j2ξ½T−Sijabðki;kj;kaÞj ≤CN−1=3
k(1
jqj2N−1
kq≠0;
N−1=3
kq¼0;
using the fact that ξ¼OðN−1=3
kÞ.
The subtraction between the two 4h2plinear terms can be formulated and bounded as
1
NkX
kk∈KX
kl
hkkk;lkljiki;jkji½T−Sklabðkk;kl;kaÞ
≤1
NkX
kk∈Knfki;kag
þ1
Nk
δkk;kiþ1
Nk
δkk;kaX
kl
hkkk;lkljiki;jkji½T−Sklabðkk;kl;kaÞ
:ðB17Þ
For the term with kk¼ki(i.e., δkk;ki), we have
1
NkX
kl
hkki;lkjjiki;jkji½T−Sklabðki;kj;kaÞ
≤1
NkX
kl
CN−1=3
k≤CN−4=3
k;
where the ERI definition at zero momentum transfer skips the singular fraction and is Oð1Þ, and according to the
assumption on T−Sit always holds that
max
ijab;ki;kj;ka∈Kj½T−Sijabðki;kj;kaÞj ≤CN−1=3
k:ðB18Þ
For the term with kk¼ka(i.e., δkk;ka), its estimate is the same as the term above when ka¼ki. When ka≠ki∈K,we
have jqj≥CN−1=3
kand
1
NkX
kl
hkka;lkbjiki;jkji½T−Sklabðka;kb;kaÞ
≤
1
NkX
kl
C
jqj2N−1=3
k≤C1
jqjN−1
k:
For the first summation term in Eq. (B17), we introduce the change of variable kk→ki−q1and write it as
1
NkX
q1∈Kqnf0;−qgX
kl
hkðki−q1Þ;lðkjþq1Þjiki;jkji½T−Sklabðki−q1;kjþq1;kaÞ
:
When q¼0, this term can be further bounded using Eq. (B18) by
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-23
ðÞ ≤1
NkX
q1∈Kqnf0g
C
jq1j2N−1=3
k≤CZΩ
dq1
1
jq1j2N−1=3
k≤CN−1=3
k:
When q≠0, this term can be further bounded as
ðÞ ≤C1
NkX
q1∈Kqnf0;−qg
1
jq1j2X
G∈L
0
1
jq1þqþGj2N−1
k
≤CN−1
kZΩ
dq1X
G∈L
0
1
jq1j2
1
jq1þqþGj2
≤CN−1
kX
G∈L
0
1
jqþGj2≤C1
jqj2N−1
k;
where L
0denotes the set of 27 lattice vectors in Laround the origin. The first inequality is based on the lemma assumption
on T−Sthat
j½T−Sklabðki−q1;kjþq1;kaÞj ≤C1
jq1þqþG0j2N−1
k≤CX
G∈L
0
1
jq1þqþGj2N−1
k;
where G0∈L
0is the unique lattice vector that makes q1þqþG0∈Ω. The third inequality can be obtained from the
nonsmoothness characterization of function
fðzÞ¼ZΩ
dq1
1
jq1j2
1
jq1þzj2:
Using Lemma 11 in Ref. [41],fðzÞis singular only at z¼0of order −2, and its value at z∈Ωnf0gis bounded by Cjzj−2.
Based on the above estimates of the first term in Eq. (B17), we obtain the estimate of the error accumulation in the 4h2p
linear term calculation as
1
NkX
kk∈KX
kl
hkkk;lkljiki;jkji½T−Sklabðkk;kl;kaÞ
≤C1
jqj2N−1
kq≠0;
N−1=3
kq¼0:
The same analysis can be applied to all the similar linear terms that contain ERIs with matched orbital indices. For other
linear terms, the analysis can be done similarly, and they all contribute at most OðN−1
kÞerror to the overall subtraction.
Taking the subtraction between the following 3h3plinear terms as an example, it can be formulated and bounded as
1
NkX
kk∈KX
kc
haka;kkkjiki;ckci½T−Skjbcðkk;kj;kbÞ
≤C1
NkX
kk∈KnfkbgX
kc
þ1
Nk
δkk;kbj½T−Skjbcðkk;kj;kbÞj
≤C1
NkX
q1∈Kqnf0g
1
jq1j2N−1
kþ1
Nk
N−1=3
k≤CN−1
k:
For the subtraction between quadratic terms, we consider the 4h2pquadratic term as an example. The subtraction
between the two 4h2pquadratic terms can be formulated and bounded as
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-24
1
N2
kX
kkkc∈KX
klcd
hkkk;lkljckc;dkdiðTijcdðki;kj;kcÞTklab ðkk;kl;kaÞ−Sijcdðki;kj;kcÞSklab ðkk;kl;kaÞÞ
≤C1
N2
kX
kkkc∈KX
klcd
jTijcdðki;kj;kcÞ½T−Sklab ðkk;kl;kaÞþ½T−Sijcd ðki;kj;kcÞSklabðkk;kl;kaÞj
≤C1
N2
kX
kkkc∈KX
klcd
j½T−Sklabðkk;kl;kaÞj þ C1
N2
kX
kkkc∈KX
klcd
j½T−Sijcdðki;kj;kcÞj
≤CN−1
k:
Similar analysis can be done to all the remaining quadratic terms, and they all contribute at most OðN−1
kÞerror to the overall
subtraction. Gathering all the estimates above together, we finish the proof.
APPENDIX C: PROOF OF THEOREM 2
To guarantee the convergence and control the regularity
of CCDðnÞcalculations with n→∞, we introduce addi-
tional technical assumptions similar to those in Ref. [41].
One key difference is that Ref. [41] measures the error in
the amplitude calculation using the maximum entrywise
norm. When Madelung constant correction is used, the
error of the amplitude should be measured by the average
entrywise norm (related to the L1norm) instead, which is
denoted by
kTk1¼1
N3
kX
ki;kj;ka∈KX
ijab
jTijabðki;kj;kaÞj;
ktk1¼1
jΩj3ZΩ×Ω×Ω
dkidkjdkaX
ijab
jtijabðki;kj;kaÞj;
where Tand tdenote a generic amplitude computed using
mesh Kand in the TDL, respectively.
Assume that the CCD amplitude equations using a
sufficiently fine mesh K(with or without the Madelung
constant correction) and in the TDL have unique solutions
and denote the solutions as TNk
and t, respectively.
Convergence of the CCDðnÞamplitudes to the CCD
amplitude is defined in the k·k1-norm sense as
lim
n→∞kTNk
n−TNk
k1¼0and lim
n→∞ktn−tk1¼0:
We impose a sufficient condition that guarantees the
convergence of fixed point iterations by requiring the target
mapping to be contractive in a domain that contains both
the solution point and the initial guess.
Assumption 18. For Nksufficiently large, the following
statements hold.
(1) The exact CCDðnÞamplitude tnconverges to the
CCD amplitude tpointwisely as n→∞, i.e.,
lim
n→∞½tnijabðki;kj;kaÞ
¼½tijabðki;kj;kaÞ;∀i; j;a; b;ki;kj;ka:ðC1Þ
(2) ðεNkÞ−1ANkis a contraction map in some domain
BNk⊂Cnocc×nocc ×nvir×nvir ×Nk×Nk×Nkthat contains TNk
and the initial guess 0, i.e.,
ðεNkÞ−1ANkðTÞ∈BNk;∀T∈BNk;
kðεNkÞ−1ðANkðTÞ−ANkðSÞÞk1
≤LkT−Sk1;∀T; S∈BNk;ðC2Þ
with a Lipschitz constant L<1.
(3) The exact CCDðnÞamplitude tnand the domain BNk
above satisfy that
MKtn∈BNk;∀n>0:ðC3Þ
(4) For all the amplitudes ftng, there exists a constant C
such that
kMKATDLðtnÞ−ANk;ξðMKtnÞk1
≤CN−1
k;∀n>0;
kMKATDLðtnÞ−ANkðMKtnÞk1
≤CN−1=3
k;∀n>0:ðC4Þ
Note that, when we consider the finite-size calculation with
Madelung constant correction, the components εNkor ANk
in the second assumption need to be changed to εNk;ξor
ANk;ξaccordingly.
Remark 19. The second assumption guarantees that
fTNk
nglies in BNkand converges to TNk
. For the third
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-25
assumption, Theorem 1 shows that with each fixed n>0
the amplitude TNk
nconverges to tnin the sense of
lim
Nk→∞kMKtn−TNk
nk1¼0;
suggesting that fTNk
ng⊂BNkconverges to tnwith
K→Ω. Therefore, BNkand MKtnshould be related to
each other, which leads to the third assumption.
For the last assumption, we note that the two related error
estimates in Lemma 11 have the prefactor Cdependent on
the amplitude tn. The assumption here is stronger in the
sense that Cneeds to be independent of tn.
Rigorous statement of Theorem 2. Under Assumption
18, the finite-size error of the CCD correlation energy
scales as OðN−1=3
kÞin each of the following scenarios:
(i) the Madelung constant correction is applied only to
the ERI contraction ANkand (ii) the Madelung constant
correction is applied only to the orbital energy εNk
nk. When
the Madelung constant correction is applied to both ANk
and εNk
nkin the CCD calculation, the overall finite-size error
scales as OðN−1
kÞ.
Proof. The finite-size error in the CCD energy calcu-
lation with or without the Madelung constant correction
can be estimated as
ETDL
CCD −ENk
CCD¼jGTDLðtÞ−GNkðTNk
Þj
≤jGNkðMKtÞ−GNkðTNk
Þj
þjGTDL ðtÞ−GNkðMKtÞj
≤CkMKt−TNk
k1þCN−1
k;ðC5Þ
where the last inequality uses the boundedness of jWijabjin
GNkand Lemma 10. To first estimate the above amplitude
error when the Madelung constant correction is applied to
both orbital energy and ERI contractions, we consider the
error splitting Eq. (B2) for the amplitude calculation at the
nth fixed point iteration as
kMKtn−TNk
nk1≤
1
εTDL ½MKATDLðtn−1Þ−ANk;ξðMKtn−1Þ
1
þ
1
εTDL −
1
εNk;ξANk;ξðMKtn−1Þ
1
þ
1
εNk;ξ½ANk;ξðMKtn−1Þ−ANk;ξðTNk
n−1Þ
1
≤CN−1
kþL
MKtn−1−TNk
n−1
1;
where the last estimate uses the assumption in Eq. (C4) for
the first term and the assumptions of contraction maps in
Eq. (C2) and MKtn−1∈BNkin Eq. (C3) for the third term.
Since the initial guesses in the finite and the TDL cases
satisfy kMKt0−TNk
0k1¼0, we can recursively derive that
kMKtn−TNk
nk1≤C1−Ln
1−LN−1
k;
and, thus, the first assumption Eq. (C1) gives
kMKt−TNk
k1¼lim
n→∞kMKtn−TNk
nk1≤CN−1
k:
Plugging this estimate into Eq. (C5) then finishes the
proof for the scenario when the Madelung constant cor-
rection is applied to both orbital energies and ERI con-
tractions.
For the two scenarios with partial Madelung constant
correction, a similar analysis as above gives
kMKtn−TNk
nk1≤CN−1=3
kþL
MKtn−1−TNk
n−1
1;
where the dominant CN−1=3
kerror comes from uncorrected
orbital energy or uncorrected ERI-contraction term.
Recursively, we can obtain the amplitude error estimate as
kMKt−TNk
k1≤CN−1=3
k;
which finishes the proof for the two scenarios with partial
correction. ▪
APPENDIX D: QUADRATURE ERROR
ESTIMATE FOR PERIODIC FUNCTIONS
WITH ALGEBRAIC SINGULARITY
This section presents a collection of lemmas that provide
quadrature error estimates for trapezoidal rules over peri-
odic functions with specific algebraic singularities, which
are used in this paper. Most of the lemmas are proven in
Ref. [41] and are restated here for completeness. In
addition, we introduce and prove a new quadrature error
estimate that is critical in describing the efficacy of the
Madelung constant correction and inverse symmetry in
reducing the quadrature error in orbital energies and ERI
contractions.
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-26
The lemmas presented in this section provide the
asymptotic scaling of the quadrature errors and also a
quantitative relationship between the prefactors in the
estimate and the algebraic singularities of the integrand.
In addition to the singularity order, we further
quantitatively characterize the algebraic singularity as
follows. For a univariate function fðxÞthat is smooth
everywhere in Vexcept at x¼x0with algebraic singu-
larity of order γ, we define a constant
Hl
V;x0ðfÞ¼min C∶j∂α
xfðxÞj ≤Cjx−x0jγ−jαj;∀jαj≤l; ∀x∈Vnfx0g
¼max
jαj≤l
ð∂α
xfðxÞÞ=jx−x0jγ−jαj
L∞ðVÞ:
For a multivariate function fðx;yÞthat is smooth everywhere in VX×VYexcept at x¼x0with algebraic singularity of
order γ, we define a constant
Hl
VX×VY;ðx0;·ÞðfÞ¼min nC∶j∂α
x∂β
yfðx;yÞj ≤Cjx−x0jγ−jαj;∀jαj;jβj≤l; ∀x∈VXnfx0g;y∈VYo
¼max
jαj≤l
ð∂α
x∂β
yfðx;yÞÞ=jx−x0jγ−jαj
L∞ðV×VÞ;
where “·”in the subscript “ðx0;·Þ”is a placeholder to indicate the smooth variable. Using these two quantities, we have the
following function estimates that are extensively used in this section:
j∂α
xfðxÞj ≤Hl
V;x0ðfÞjx−x0jγ−jαj;∀l≥jαj;∀x∈Vnfx0g;
j∂α
x∂β
yfðx;yÞj ≤Hl
VX×VY;ðx0;·ÞðfÞjx−x0jγ−jαj;∀l≥jαj;jβj;∀x∈VXnfx0g;y∈VY:
1. Existing results from Ref. [41]
Lemma 20. Let fðxÞbe smooth and periodic in
V¼½−1
2;1
2d. The quadrature error of a trapezoidal rule
using an md-sized uniform mesh Xin Vdecays super-
algebraically as
jEVðf; XÞj ≤Clm−l;∀l>0:
Remark 21. If we replace fðxÞby fðx;yÞdefined in
V×VYwhich is smooth and periodic with respect to xfor
each y∈VYand satisfies supx∈V;y∈VYj∂α
xfðx;yÞj <∞for
any α≥0, Lemma 20 can be generalized as
jEVðfð·;yÞ;XÞj ≤Clm−l;∀l>0;∀y∈VY;
where constant Clis independent of y∈VY.
Lemma 22. Let fðxÞbe periodic with respect to V¼
½−1
2;1
2dand smooth everywhere except at x¼0of order
γ≥−dþ1.Atx¼0,fðxÞis set to 0. The quadrature error
of a trapezoidal rule using an md-sized uniform mesh X
that contains x¼0can be estimated as
jEVðf; XÞj ≤CHdþmaxð1;γÞ
V;0ðfÞm−ðdþγÞ:
If fð0Þis set to an Oð1Þvalue in the calculation, it
introduces an additional Oðm−dÞquadrature error.
Remark 23. If we replace fðxÞby fðx;yÞdefined in
V×VYwhich is smooth everywhere in V×VYexcept at
x¼0of order γ, Lemma 22 can be generalized to
jEVðfð·;yÞ;XÞj ≤CHdþmaxð1;γÞ
V×VY;ð0;·ÞðfÞm−ðdþγÞ;∀y∈VY;
where the prefactor applies uniformly across all y∈VY.
Lemma 24. Let fðxÞ¼f1ðxÞf2ðxÞ, where f1ðxÞand
f2ðxÞare periodic with respect to V¼½−1
2;1
2dand
(i) f1ðxÞis smooth everywhere except at x¼z1¼0of
order γ≤0and
(ii) f2ðxÞis smooth everywhere except at x¼z2≠0of
order 0.
Consider an md-sized uniform mesh Xin V. Assume that X
satisfies that z1,z2are either on the mesh or Θðm−1Þaway
from any mesh points and mis sufficiently large that
jz1−z2j¼Ωðm−1Þ.Atx¼z1and x¼z2,fðxÞis set to
0. The trapezoidal rule using Xhas quadrature error
jEVðf; XÞj ≤CHdþ1
V;z1ðf1ÞHdþ1
V;z2ðf2Þm−ðdþγÞ:
If fðz1Þand fðz2Þare set to arbitrary Oð1Þvalues, it
introduces an additional Oðm−dÞquadrature error.
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-27
Remark 25. If we replace fiðxÞwith i¼1,2byfiðx;yÞ
defined in V×VYwhich is smooth everywhere in V×VY
except at x¼ziof order γand 0, respectively, Lemma 24
can be generalized to
jEVðf1ð·;yÞf2ð·;yÞ;XÞj
≤CHdþ1
V×VY;ð0;·Þðf1ÞHdþ1
V×VY;ð0;·Þðf2Þm−ðdþγÞ;∀y∈VY;
where the prefactor applies uniformly across all y∈VY.
2. A new quadrature error estimate
Here, we prove Lemma 16, which is used in the
quadrature error estimate of the ERI contractions with
the Madelung constant correction in Appendix B2. This
lemma is a generalization of the existing result in Lemma
24 with an additional condition.
Restatement of Lemma 16. Let fðxÞ¼f1ðxÞf2ðxÞ,
where f1ðxÞand f2ðxÞare periodic with respect to V¼
½−1
2;1
2dand
(i) f1ðxÞis smooth everywhere except at x¼z1¼0of
order γ≤−1,
(ii) f2ðxÞis smooth everywhere except at x¼z2≠0of
order 0, and
(iii) ∂α
xf2ð0Þ¼0for any derivative order jαj≤s.
Assume γ>−dfor fðxÞto be integrable in Vand γþ
sþ1≤0so the leading algebraic singularity of fðxÞis at
x¼0. Consider an md-sized uniform mesh Xin V.
Assume that Xsatisfies that z1,z2are either on the mesh
or Θðm−1Þaway from any mesh points and mis sufficiently
large that jz1−z2j¼Ωðm−1Þ.Atx¼z1and x¼z2,fðxÞ
is set to 0. The trapezoidal rule using Xhas quadrature error
jEVðf; XÞj ≤CHdþ1
V;z1ðf1ÞHdþ1
V;z2ðf2Þjz1−z2j−ðsþ1Þ
×m−ðdþγþsþ1Þ:
Proof. For z1¼0and any z2∈V, we can introduce a
proper translation fðxÞ→fðx−x0Þto move both the
singular points z1and z2to the smaller cube ½−1
4;1
4din
V. The target quadrature error can be correspondingly
reformulated as
EVðfð·Þ;XÞ¼EVþx0ðfð·−x0Þ;X−x0Þ
¼EVðfð·−x0Þ;X−x0Þ;
which is the quadrature error of the translated function
fðx−x0Þ. Without loss of generality, we assume such a
translation has been applied to fðxÞand Xand both
singular points z1and z2lie in ½−1
4;1
4d.
Define a cutoff function ψ∈C∞
cðRnÞsatisfying
ψðxÞ¼1;jxj<1
4;
0;jxj>1
2:
Denote the distance between the two singular points as
δz¼jz2−z1jand define two local cutoff functions that
isolate the two singular points as
ψδz;1ðxÞ¼ψx−z1
δz;ψδz;2ðxÞ¼ψx−z2
δz;
whose supports are both inside V. The target integrand can
be split into three parts as
fðxÞ¼fðxÞψδz;1ðxÞþfðxÞψδz;2ðxÞ
þfðxÞð1−ψδz;1ðxÞ−ψδz;1ðxÞÞ:
All three terms satisfy the periodic boundary condition on
∂V. The first term is smooth everywhere except at x¼z1
of order γþsþ1, the second term is smooth everywhere
except at x¼z2of order 0, and the last term is smooth
everywhere. Application of Lemmas 20 and 22 to these
terms suggests that
jEVðfψδz;1;XÞj ≤CHdþ1
V;z1ðfψδz;1Þm−ðdþγþsþ1Þ;ðD1Þ
jEVðfψδz;2;XÞj ≤CHdþ1
V;z2ðfψδz;2Þm−d;ðD2Þ
jEVðfð1−ψδz;1−ψδz;2Þ;XÞj
≤CjαjZV
dxj∂α
xfð1−ψδz;1−ψδz;2Þjm−jαj;∀jαj>d:
ðD3Þ
Note that the last estimate means a superalgebraic decaying
error and constant Cjαjdepends only on Vand jαj(see
Eq. (H.6) in Ref. [41] for the derivation of this detailed
prefactor for Lemma 20). Since γþsþ1≤0, the three
estimates together prove that the quadrature error of fðxÞ
scales asymptotically as m−ðdþγþsþ1Þas m→∞. In order to
describe the extreme case where the two singular points are
only Oðm−1Þaway from each other for a given mesh X,
i.e., δz¼Oðm−1Þ, we provide a precise description of the
three prefactors in the estimates above using δz.
For the prefactor in Eq. (D1), it is defined by
Hl
V;z1ðfψδz;1Þ
¼max
jαj≤l
ð∂α
xfðxÞψδz;1ðxÞÞ=jx−z1jγþ1−jαj
L∞ðVÞ;
∀l>0:
For any derivative order jαj≤l,wehave
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-28
j∂α
xðf1ðxÞf2ðxÞψδz;1ðxÞÞj ≤CX
α1þα2þα3¼α
j∂α1
xf1ðxÞjj∂α2
xf2ðxÞjj∂α3
xψδz;1ðxÞj
≤CX
α1þα2þα3¼α
Hl
V;z1ðf1Þjx−z1jγ−jα1jj∂α2
xf2ðxÞjδz−jα3jδjx−z1j<1
2δz
≤CX
α1þα2þα3¼α
Hl
V;z1ðf1Þjx−z1jγ−jα1j−jα3jj∂α2
xf2ðxÞjδjx−z1j<1
2δz:ðD4Þ
The last inequality uses jx−z1j≤1
2δzby noting that ∂α3
xψδz;1ðxÞis zero when jx−z1j>1
2δz. Next, we estimate j∂α2
xf2ðxÞj
in the ball jx−z1j≤1
2δz. When jα2j≤s,wehave
j∂α2
xf2ðxÞj ¼ X
jβj¼sþ1−jα2j
sþ1−jα2j
β!ðx−z1ÞβZ1
0
ð1−tÞs−jα2j∂α2þβ
xf2ðz1þtðx−z1ÞÞdt
≤Cjx−z1jsþ1−jα2jsup
jx−z1j≤1
2δz;jβj¼sþ1
j∂β
xf2ðxÞj
≤Cjx−z1jsþ1−jα2jHsþ1
V;z2ðf2Þδz−ðsþ1Þ:
The first equality applies the Taylor expansion of ∂α2
xf2ðxÞat x¼z1with the assumption that ∂β
xf2ðz1Þ¼0for any jβj≤s.
The last inequality uses the algebraic singularity of f2ðxÞat x¼z2. When jα2j>s, we have
j∂α2
xf2ðxÞj ≤sup
jx−z1j<1
2δz
j∂α2
xf2ðxÞj ≤CHl
V;z2ðf2Þδz−jα2j≤CHl
V;z2ðf2Þjx−z1jsþ1−jα2jδz−ðsþ1Þ:
Plugging the estimates of j∂α2
xf2ðxÞj above into Eq. (D4), we obtain
j∂α
xðf1ðxÞf2ðxÞψδz;1ðxÞÞj ≤CHl
V;z1ðf1ÞHl
V;z2ðf2Þδz−ðsþ1Þjx−z1jγþsþ1−jαj;l≥dþ1;
which suggests that
Hl
V;z1ðfψδz;1Þ≤CHl
V;z1ðf1ÞHl
V;z2ðf2Þδz−ðsþ1Þ;∀l≥dþ1:
Similar analysis can also be applied to the prefactor in Eq. (D2) to obtain
Hl
V;z2ðfψδz;2Þ≤CHl
V;z1ðf1ÞHl
V;z2ðf2Þδz−ðsþ1Þ;∀l≥dþ1:
For the estimate of the third term in Eq. (D3) with any jαj¼l≥dþ1, the prefactor can be bounded as
ZV
dxj∂α
xfðxÞð1−ψδz;1ðxÞ−ψδz;2ðxÞÞj
≤CX
α1þα2þα3¼αZV
dxj∂α1
xf1ðxÞjj∂α2
xf2ðxÞjj∂α3
xð1−ψδz;1ðxÞ−ψδz;2ðxÞÞj
≤CHl
V;z1ðf1ÞHl
V;z2ðf2ÞX
α1þα2þα3¼αZV
dxjx−z1jγ−jα1jjx−z2j−jα2jj∂α3
xð1−ψδz;1ðxÞ−ψδz;2ðxÞÞj:
For the integral with each set of ðα1;α2;α3Þ, we consider two cases.
(i) α3¼0. The integral can be bounded by
ðÞ ≤ZVnðBðz1;1
4δzÞ∪Bðz2;1
4δzÞÞ
dxjx−z1jγ−jα1jjx−z2jjα1j−jαj≤Cð1þδzγþd−jαjÞ;
using the Hölder-inequality technique developed in the proof of Lemma 24 in Ref. [41].
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-29
(ii) α3>0. The integral can be bounded by
ðÞ ≤CZðBðz1;1
2δzÞnBðz1;1
4δzÞÞ∪ðBðz2;1
2δzÞnBðz2;1
4δzÞÞ
dxjx−z1jγ−jα1jjx−z2j−jα2jδz−jα3j≤Cδzγþd−jαj:
Collecting these two terms together, the error estimate in Eq. (D3) with any jαj¼l≥dþ1can be further bounded as
jEVðfð1−ψδz;1−ψδz;2Þ;XÞj ≤ClHl
V;z1ðf1ÞHl
V;z2ðf2Þδzγþd−lm−l
≤ClHl
V;z1ðf1ÞHl
V;z2ðf2Þδz−ðsþ1Þm−ðγþsþ1þd−lÞm−l
≤ClHl
V;z1ðf1ÞHl
V;z2ðf2Þδz−ðsþ1Þm−ðγþsþ1þdÞ;
where the second inequality uses the assumption δz¼Ωðm−1Þand γþsþ1þd−l≤0. Gathering the above prefactor
descriptions for Eqs. (D1)–(D3), we obtain
jEVðf; XÞj ≤CHdþ1
V;z1ðf1ÞHdþ1
V;z2ðf2Þδz−ðsþ1Þm−ðγþsþ1þdÞ;
which finishes the proof. ▪
[1] R. Martin, Electronic Structure: Basic Theory and Practical
Methods (Cambridge University Press, Cambridge, England,
2008).
[2] R. M. Martin, L. Reining, and D. M. Ceperley, Interacting
Electrons (Cambridge University Press, Cambridge,
England, 2016).
[3] R. Bistritzer and A. H. MacDonald, Moir´e bands in twisted
double-layer graphene,Proc. Natl. Acad. Sci. U.S.A. 108,
12233 (2011).
[4] N. Bultinck, E. Khalaf, S. Liu, S. Chatterjee, A. Vishwanath,
and M. P. Zaletel, Ground state and hidden symmetry of
magic-angle graphene at even integer filling,Phys. Rev. X
10, 031034 (2020).
[5] G. Makov and M. C. Payne, Periodic boundary conditions
in ab initio calculations,Phys. Rev. B 51, 4014 (1995).
[6] L. M. Fraser, W. M. C. Foulkes, G. Rajagopal, R. J. Needs,
S. D. Kenny, and A. J. Williamson, Finite-size effects and
Coulomb interactions in quantum Monte Carlo calculations
for homogeneous systems with periodic boundary condi-
tions,Phys. Rev. B 53, 1814 (1996).
[7] T. Gruber, K. Liao, T. Tsatsoulis, F. Hummel, and A.
Grüneis, Applying the coupled-cluster ansatz to solids
and surfaces in the thermodynamic limit,Phys. Rev. X 8,
021043 (2018).
[8] J. Hermann, R. A. DiStasio Jr, and A. Tkatchenko, First-
principles models for van der Waals interactions in mol-
ecules and materials: Concepts, theory, and applications,
Chem. Rev. 117, 4714 (2017).
[9] I. Shavitt and R. J. Bartlett, Many-Body Methods in Chem-
istry and Physics: MBPT and Coupled-Cluster Theory
(Cambridge University Press, Cambridge, Engaldn, 2009).
[10] M. Marsman, A. Grüneis, J. Paier, and G. Kresse, Second-
order Møller-Plesset perturbation theory applied to ex-
tended systems. I. Within the projector-augmented-wave
formalism using a plane wave basis set,J. Chem. Phys. 130,
184103 (2009).
[11] A. Grüneis, M. Marsman, and G. Kresse, Second-order
Møller-Plesset perturbation theory applied to extended
systems. II. Structural and energetic properties,J. Chem.
Phys. 133, 074107 (2010).
[12] T. Gruber and A. Grüneis, Ab initio calculations of carbon
and boron nitride allotropes and their structural phase
transitions using periodic coupled cluster theory,Phys. Rev.
B98, 134108 (2018).
[13] A. Grüneis, A coupled cluster and Møller-Plesset pertur-
bation theory study of the pressure induced phase transition
in the LiH crystal,J. Chem. Phys. 143, 102817 (2015).
[14] C. Müller and B. Paulus, Wavefunction-based electron
correlation methods for solids,Phys. Chem. Chem. Phys.
14, 7605 (2012).
[15] G. H. Booth, A. Grüneis, G. Kresse, and A. Alavi, Towards
an exact description of electronic wavefunctions in real
solids,Nature (London) 493, 365 (2013).
[16] T. Schäfer, B. Ramberger, and G. Kresse, Quartic scaling
MP2 for solids: A highly parallelized algorithm in the plane
wave basis,J. Chem. Phys. 146, 104101 (2017).
[17] J. McClain, Q. Sun, G. K. L. Chan, and T. C. Berkelbach,
Gaussian-based coupled-cluster theory for the ground-state
and band structure of solids,J. Chem. Theory Comput. 13,
1209 (2017).
[18] I. Y. Zhang and A. Grüneis, Coupled cluster theory in
materials science,Front. Mater. 6, 123 (2019).
[19] Y. Gao, Q. Sun, J. M. Yu, M. Motta, J. McClain, A. F.
White, A. J. Minnich, and Garnet Kin-Lic Chan, Electronic
structure of bulk manganese oxide and nickel oxide from
coupled cluster theory,Phys. Rev. B 101, 165138 (2020).
[20] X. Wang and T. C. Berkelbach, Excitons in solids from peri-
odic equation-of-motion coupled-cluster theory,J. Chem.
Theory Comput. 16, 3095 (2020).
[21] T. Zhu and G. K.-L. Chan, All-electron Gaussian-based
G0W0for valence and core excitation energies of periodic
systems,J. Chem. Theory Comput. 17, 727 (2021).
XIN XING and LIN LIN PHYS. REV. X 14, 011059 (2024)
011059-30
[22] V. A. Neufeld, H.-Z. Ye, and T. C. Berkelbach, Ground-state
properties of metallic solids from ab initio coupled-cluster
theory, J. Phys. Chem. Lett. 13, 7497 (2022).
[23] S. J. Bintrim, T. C. Berkelbach, and H.-Z. Ye, Integral-
direct Hartree-Fock and Møller-Plesset perturbation theory
for periodic systems with density fitting: Application to the
benzene crystal,J. Chem. Theory Comput. 18, 5374 (2022).
[24] S. Banerjee and A. Y. Sokolov, Non-Dyson algebraic
diagrammatic construction theory for charged excitations
in solids,J. Chem. Theory Comput. 18, 5337 (2022).
[25] V. A. Neufeld and T. C. Berkelbach, Highly accurate
electronic structure of metallic solids from coupled-cluster
theory with nonperturbative triple excitations,Phys. Rev.
Lett. 131, 186402 (2023).
[26] F. Coester and H. Kümmel, Short-range correlations in
nuclear wave functions,Nucl. Phys. 17, 477 (1960).
[27] H. Kümmel, Origins of the coupled cluster method,Theor.
Chim. Acta 80, 81 (1991).
[28] J. Čížek, On the correlation problem in atomic and
molecular systems. Calculation of wavefunction compo-
nents in Ursell-type expansion using quantum-field theo-
retical methods,J. Chem. Phys. 45, 4256 (1966).
[29] H. J. Monkhorst and J. D. Pack, Special points for Brillouin-
zone integrations,Phys. Rev. B 13, 5188 (1976).
[30] S. Hirata, R. Podeszwa, M. Tobita, and R. J. Bartlett,
Coupled-cluster singles and doubles for extended systems,
J. Chem. Phys. 120, 2581 (2004).
[31] X. Xing, X. Li, and L. Lin, Unified analysis of finite-size error
for periodic Hartree-Fock and second order Møller-Plesset
perturbation theory,Math. Comput. 93, 679 (2024).
[32] X. Xing, X. Li, and L. Lin, Staggered mesh method for
correlation energy calculations of solids: Second-order
Møller-Plesset perturbation theory,J. Chem. Theory
Comput. 17, 4733 (2021).
[33] X. Xing and L. Lin, Staggered mesh method for correlation
energy calculations of solids: Random phase approximation
in direct ring coupled cluster doubles and adiabatic con-
nection formalisms,J. Chem. Theory Comput. 18, 763
(2022).
[34] K. Liao and A. Grüneis, Communication: Finite size
correction in periodic coupled cluster theory calculations
of solids,J. Chem. Phys. 145, 141102 (2016).
[35] T. N. Mihm, B. Yang, and J. J. Shepherd, Power laws used
to extrapolate the coupled cluster correlation energy to the
thermodynamic limit,J. Chem. Theory Comput. 17, 2752
(2021).
[36] T. N. Mihm, L. Weiler, and J. J. Shepherd, How the
exchange energy can affect the power laws used to
extrapolate the coupled cluster correlation energy to the
thermodynamic limit,J. Chem. Theory Comput. 19, 1686
(2023).
[37] T. N. Mihm, A. R. McIsaac, and J. J. Shepherd, An opti-
mized twist angle to find the twist-averaged correlation
energy applied to the uniform electron gas,J. Chem. Phys.
150, 191101 (2019).
[38] T. N. Mihm, W. Z. Van Benschoten, and J. J. Shepherd,
Accelerating convergence to the thermodynamic limit with
twist angle selection applied to methods beyond many-body
perturbation theory,J. Chem. Phys. 154, 024113 (2021).
[39] L. Weiler, T. N. Mihm, and J. J. Shepherd, Machine learning
for a finite size correction in periodic coupled cluster theory
calculations,J. Chem. Phys. 156, 204109 (2022).
[40] T. Schäfer, W. Z. Van Benschoten, J. J. Shepherd, and A.
Grüneis, Sampling the reciprocal Coulomb potential in
finite and anisotropic cells,J. Chem. Phys. 160, 051101
(2024).
[41] X. Xing and L. Lin, Finite-size effects in periodic coupled
cluster calculations,J. Comput. Phys. 500, 112755 (2024).
[42] D. J. Thouless, Stability conditions and nuclear rotations in
the Hartree-Fock theory,Nucl. Phys. 21, 225 (1960).
[43] S. R. Gwaltney and M. Head-Gordon, A second-order
perturbative correction to the coupled-cluster singles and
doubles method: CCSD (2),J. Chem. Phys. 115, 2014
(2001).
[44] F. Izzo, O. Runborg, and R. Tsai, High-order corrected
trapezoidal rules for a class of singular integrals,
Adv. Comput. Math. 49, 60 (2023).
[45] P. J. Davis and P. Rabinowitz, Methods of Numerical
Integration (Dover, New York, 2007).
[46] Q. Sun, T. C. Berkelbach, N. S. Blunt, G. H. Booth, S. Guo,
Z. Li, J. Liu, J. D. McClain, E. R. Sayfutyarova, S. Sharma,
S. Wouters, and G. K.-L. Chan,
PYSCF
: The Python-based
simulations of chemistry framework,Comput. Mol. Sci. 8,
e1340 (2018).
[47] S. Chiesa, D. M. Ceperley, R. M. Martin, and M. Holzmann,
Finite-size error in many-body simulations with long-range
interactions,Phys. Rev. Lett. 97, 076404 (2006).
[48] C. Brouder, G. Panati, M. Calandra, C. Mourougane, and N.
Marzari, Exponential localization of Wannier functions in
insulators,Phys. Rev. Lett. 98, 046402 (2007).
[49] D. Monaco, G. Panati, A. Pisante, and S. Teufel, Optimal
decay of Wannier functions in Chern and quantum Hall
insulators,Commun. Math. Phys. 359, 61 (2018).
[50] G. E. Scuseria and P. Y. Ayala, Linear scaling coupled
cluster and perturbation theories in the atomic orbital
basis,J. Chem. Phys. 111, 8330 (1999).
[51] R. Schneider, Analysis of the projected coupled cluster
method in electronic structure calculation,Numer. Math.
113, 433 (2009).
[52] N. D. Drummond, R. J. Needs, A. Sorouri, and W. M. C.
Foulkes, Finite-size errors in continuum quantum
Monte Carlo calculations,Phys. Rev. B 78, 125106
(2008).
[53] L. N. Trefethen and J. Weideman, The exponentially con-
vergent trapezoidal rule,SIAM Rev. 56, 385 (2014).
[54] R. Sundararaman and T. A. Arias, Regularization of the
Coulomb singularity in exact exchange by Wigner-Seitz
truncated interactions: Towards chemical accuracy in
nontrivial systems,Phys. Rev. B 87, 165122 (2013).
[55] J. Spencer and A. Alavi, Efficient calculation of the exact
exchange energy in periodic systems using a truncated
Coulomb potential,Phys. Rev. B 77, 193110 (2008).
[56] S. Ismail-Beigi, Truncation of periodic image interactions
for confined systems,Phys. Rev. B 73, 233103 (2006).
INVERSE VOLUME SCALING OF FINITE-SIZE ERROR IN …PHYS. REV. X 14, 011059 (2024)
011059-31