Chemical Physics Letters 393 (2004) 5157

www.elsevier.com/locate/cplett

A new hybrid exchangecorrelation functional using the

Coulomb-attenuating method (CAM-B3LYP)
Takeshi Yanai
a

a,*

, David P. Tew b, Nicholas C. Handy

Oak Ridge National Laboratory, P.O. Box 2008, MS6367, Oak Ridge, TN 37831, USA
b
Department of Chemistry, University of Cambridge, Cambridge, CB2 1EW, UK
Received 15 March 2004; in nal form 4 June 2004

Abstract
A new hybrid exchangecorrelation functional named CAM-B3LYP is proposed. It combines the hybrid qualities of B3LYP and
the long-range correction presented by Tawada et al. [J. Chem. Phys., in press]. We demonstrate that CAM-B3LYP yields atomization energies of similar quality to those from B3LYP, while also performing well for charge transfer excitations in a dipeptide
model, which B3LYP underestimates enormously. The CAM-B3LYP functional comprises of 0.19 HartreeFock (HF) plus 0.81
Becke 1988 (B88) exchange interaction at short-range, and 0.65 HF plus 0.35 B88 at long-range. The intermediate region is smoothly
described through the standard error function with parameter 0.33.
2004 Elsevier B.V. All rights reserved.

1. Introduction
In density functional theory (DFT), as it is used for
computational chemistry, the hybrid functional B3LYP
[2,3] appears to oer the greatest contribution (if measured by the number of applications which have been
published). However it is unsuccessful in a number of
important applications: (i) the polarizability of long
chains, (ii) excitations using time dependent theory
(TDDFT) [46] for Rydberg states, and perhaps most
important (iii) charge transfer (CT) excitations [79].
The reason for these failures is understood, at longrange the exchange potential behaves as
0:2r
1 , instead of the exact value
r
1 . Even so, the potential is
an improvement over that LDA and BLYP, where there
is no r
1 dependence in the potential.
Recently Tsuneda and co-workers [1] have overcome

1
this deciency through an Ewald split of r12
into
1
1
erf lr12 erf lr12

:
r12
r12
r12

The rst term accounts for the short-range interaction, and the second term accounts for the long-range
*

Corresponding author. Fax: +1-865-574-0680.

E-mail address: [email protected] (T. Yanai).

0009-2614/$ - see front matter 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.cplett.2004.06.011

interaction. The key in the long-range-corrected (LC)

exchange functional scheme for DFT [1013] is that the
DFT exchange interaction is included using the rst
term (short-range), and the long-range orbitalorbital
exchange interaction is described with the Hartree
Fock (HF) exchange integral via the complementary
term. Specically, the short-range part of the exchange
interaction is incorporated by modifying
R usual exP the
3
change functional form, Ex
1=2 r q4=3
r Kr d r,
into
Z
1X
Exsr

q4=3
r Kr
2 r






p
8
1
 1
ar perf
2ar br
cr d3 r;
3
2ar
2
where ar , br and cr are:
lK 1=2
ar pr 1=3 ;
6 pqr


1
br exp
2
1;
4ar
cr 2a2r br 12:

4
5

52

T. Yanai et al. / Chemical Physics Letters 393 (2004) 5157

To reach this form, erfclr12 has been multiplied by

the square of the one-particle density matrix for the
uniform electron gas, and then integrated [12]. Use of Kr
allows the incorporation of generalized gradient approximation functionals [10]. The long-range part of the
exchange interaction is expressed with the HF exchange
integral [12],
occ X
occ Z Z
1XX
lr
Ex

wir r1 wjr r1
2 r
i
j


erflr12
wir r2 wjr r2 d3 r1 d3 r2 ;
r12

where wir is the ith r-spin molecular orbital. The parameter l determines the balance of DFT to HF exchange at intermediate r12 . If l 0, the LC DFT
calculation corresponds to the pure (non-LC) DFT
calculation, and conversely l 1 corresponds to the
standard HF calculation. Tsuneda and co-workers have
demonstrated that their LC method for the Becke 1988
[14] or PBE [15] generalized-gradient approximation
(GGA) exchange functionals in conjunction with their
one-parameter progressive (OP) correlation functional
[16] can address the above notorious problems in DFT.
The incorrect long-range exchange interaction delivered
by the standard DFT exchange functionals seems to lead
to the underestimation of 4s3d interaction energies of
the rst-row transition metals, the overestimation of the
longitudinal polarizabilities of p-conjugated polyenes,
the poor description of the weak interaction of van der
Waals bonding of rare-gas dimers, and the underestimations of Rydberg excitation energies, oscillator
strengths, and charge-transfer excitation energies.
Tsuneda and co-workers have shown that their LC
method with l 0:33 gives greatly improved results for
all of these phenomena [1,10,17].
Unfortunately, we nd that LC-BOP does not work
well for the more standard energy calculations on which
the parameters for B3LYP were derived. Indeed, the
mean absolute error in the atomization energies of 53
molecules with the high quality basis sets, augmented ccpVQZ, increases from 2.5 kcal mol
1 (B3LYP) to
9.5 kcal mol
1 (LC-BOP), which seriously detracts from
the quality of LC-BOP. The purpose of this Letter is to
show how it is possible to combine the ideas behind
B3LYP and LC-BOP to deliver a functional which has
the energetic qualities of B3LYP and the asymptotic
qualities of LC-BOP.

Fig. 1. Plots for f r 1
erflr and f r 1
a berflr.

where the relations 0 6 a b 6 1, 0 6 a 6 1, and 0 6 b 6 1

should be satised. We term this the Coulomb-attenuating method (CAM) approach. Fig. 1 illustrates the
schematic plots of two functions, Eqs. (1) and (7). The
parameter a allows us to incorporate the HF exchange
contribution over the whole range by a factor of a, and
the parameter b allows us to incorporate the DFT
counterpart over the whole range by a factor of
1
a b. We note that the widely-used hybrid B3LYP
functional [2,3] takes CAM potential partitioning of Eq.
(7) with a 0:2 and b 0:0 for the mixing of Slater exchange EXSlater and the HF exchange EXHF as follows:
EXB3 1
aEXSlater aEXHF cB88 DEXB88 ;

DEXB88

is Beckes 1988 gradient

where the additional term
correction for exchange [14] with the semiempirical parameter cB88 0:72, which Becke obtained by a linear
least-square t to experimental data [2]. Also, the original LC corresponds to the CAM with a 0:0 and
b 1:0. Figs. 2ac show the contributions to exchange

1
from r12
, apportioned into DFT and HF, for B3LYP,
LC and CAM methodologies.
The extra exibility arising from two extra parameters a and b allows us to look at how important the HF
exchange contribution is for the short-range region and
the DFT counterpart is for the long-range region. In the
original form of the LC decomposition (Eq. (1)), either
HF or DFT exchange vanishes at r 0 and r 1.
We note that CAM with the Gaussian-type basis
implementation requires the same types of the Coulomb-attenuated and non-attenuated (standard) twoelectron integrals [12] as the original LC approach of
Eq. (1). The details about modifying the DFT exchange
functionals and the HF exchange integral to involve the
error function are described in [1,10].

2. Coulomb-attenuating method with three parameters

Now, we generalize the form of Eq. (1) using two
extra parameters a and b as,
1
1
a b  erf lr12  a b  erf lr12

;
r12
r12
r12

3. Procedure: trial CAM exchangecorrelation functionals

In this Letter, we investigate the performance of the
Coulomb-attenuating method with several existing
GGA functionals. Table 1 summarizes the exchange

T. Yanai et al. / Chemical Physics Letters 393 (2004) 5157

(a)

B3LYP (=0.2, +=0.2)

(b)
0

LC ( =0.0, +=1.0)

(c)
0

53

CAM (=0.2, +=0.6)

0.2

0.4

0.4
DFT

0.6

0.2

0.6
DFT

0.8

1
0

r12

DFT

0.8

HF

0.4

HF

0.6

0.8

HF
0.2

r12

1
0

r12

1
Fig. 2. Schematic plots of the contributions to exchange from r12
, apportioned into DFT and HF, for: (a) B3LYP, (b) LC, and (c) CAM.

correlation functionals examined in this study. The

Becke 1988 exchange functional is used in all of the LC
and CAM functionals, and is mixed with the HF exchange according to Eq. (7). For the partner correlation
functionals, we use the OP correlation functional, the
LeeYangParr (LYP) [19], and the correlation functional employed in B3LYP, which is 0.19 VWN5 + 0.81
LYP, where the VWN5 functional is the local correlation functional of Vosko, Wilk and Nusair (VWN) [20]
parameterized with the data of Ceperley and Alder [21].
Note that this is dierent to the standard B3LYP implemented in GA U S S I A N which uses VWN1 instead of
VWN5 [18], we refer to this functional as B3LYP(G).
The possible combinations of the exchangecorrelation
functionals are termed CAM-BOP, CAM-BLYP, CAMB3LYP, LC-BOP, and LC-BLYP. For the parameter l,
the same value is used as in Tawadas study [1],
l 0:33. The parameter a, which determines the contribution of the HF exchange at the short-range region,
was chosen to be 0.2 for the three functionals, CAMBOP, CAM-BLYP, and CAM-B3LYP. We vary the
parameter b so that the HF exchange could contribute
to the long-range region with a b 0:6; 0:8; or 1:0 for
three functionals.

We compare the present functionals with four kinds

of the widely used, well-examined exchangecorrelation
functionals, HCTH/93 [22], BLYP, B3LYP(G) (VWN1),
and B3LYP (VWN5). We used the IN T E G R A [23] as a
part of the UT C H E M 2004 program package [24,25] to
carry out KohnSham self-consistent eld (KS-SCF)
calculations with the LC and CAM methods. The KSSCF calculations with the standard BLYP, HCTH,
B3LYP(G), and B3LYP were performed using
NW C H E M program package version 4.5 [26].

4. Results
4.1. Atomization energies, ionization potentials, and
atomic energies
We calculated 53 atomization energies and 22 ionization potentials from the molecules of the G2 set
[27,28]. All calculations were performed with suciently
accurate correlation-consistent aug-cc-pVQZ Gaussian
basis sets. Tables 2 and 3 show the statistical data for
atomization energies and ionization potentials with
comparison to the experimental data, which are taken

Table 1
Summary of the exchangecorrelation functionals
Name

Exchange functional

ab

LC-BOP
LC-BLYP
CAM-BOP

Becke88
Becke88
Becke88

0.0
0.0
0.2

CAM-BLYP

Becke88

0.2

CAM-B3LYP

Becke88

0.2

B3LYP(G)
B3LYP
BLYP
HCTH

Slater
Slater
Becke88
xHCTH

0.2
0.2
0.0
0.0

1.0
1.0
1.0
0.8
0.6
1.0
0.8
0.6
1.0
0.8
0.6
0.2
0.2
0.0
0.0

Additional exchange

Correlation functional
OP
LYP
OP

LYP

0.19 VWN5 + 0.81 LYP

0.72 DBecke88
0.72 DBecke88

0.19 VWN1(RPA) + 0.81 LYP

0.19 VWN5 + 0.81 LYP
LYP
cHCTH

54

T. Yanai et al. / Chemical Physics Letters 393 (2004) 5157

along with zero point energies from [2,29]. Table 4

summarizes the atomic energy data of the rst-row
atoms H through Ne with comparison to the exact
energies [2].
Within the standard exchangecorrelation functionals, B3LYP(G) gives the smallest errors for the atomization energies with 2.54 kcal mol
1 for the mean
absolute error (MAE) and 3.38 kcal mol
1 for the root
mean square deviation (RMS). The HCTH functional
gives the best performance for the ionization potentials
within the standard functionals with 0.154 eV for MAE
and 0.187 eV for RMS. B3LYP performs best for the
atomization energies with 0.005Eh for both MAE and
RMS.
The LC-based exchangecorrelation functionals, LCBOP and LC-BLYP, enormously overestimate the atomization energies with MAEs of 67 kcal mol
1 and
RMS errors of 89 kcal mol
1 , while they perform
comparably well for computing ionization potentials.
Both of the LC functionals systematically overestimate

the total atomic energies of HNe with an RMS of

0.090.13 Eh .
As for the CAM exchangecorrelation functionals,
the errors of the atomization energies are signicantly
improved compared to the LC method by setting
a 0:2 for CAM-BOP, CAM-BLYP and CAMB3LYP. For ionization energies, the CAM functionals
are also better than the LC functionals. The CAM-BOP
and CAM-BLYP functionals systematically overestimate the total atomic energies of HNe with a MAE of
0.0690.030 Eh . The inclusion of the VWN5 correlation
contribution into the LYP functional greatly reduces
these errors to within an acceptable level. However, the
atomization energy errors deteriorate by 0.10.3
kcal mol
1 with the inclusion of VWN5. We consider
that due to the prevalence of hydrogen in chemical
systems, the good reproduction of the total atomic energies, hydrogen in particular, is more important than
the slight loss in quality of the atomization and ionization energies.

Table 2
Statistical data for atomization energies (kcal mol
1 ) of the small G2 set (53 molecules) with aug-cc-pVQZ Gaussian basis sets
a

ab

MAE

RMS

Maximum deviation (
)

Maximum deviation (+)

LC-BOP
LC-BLYP
CAM-BOP

0.0
0.0
0.2

CAM-BLYP

0.2

CAM-B3LYP

0.2

B3LYP(G)
B3LYP
BLYP
HCTH

0.2
0.2
0.0
0.0

1.0
1.0
1.0
0.8
0.6
1.0
0.8
0.6
1.0
0.8
0.6
0.2
0.2
0.0
0.0

9.19
9.48
4.13
3.21
2.93
3.90
2.77
2.28
3.99
2.97
2.62
2.54
2.68
4.51
3.50

11.22
12.07
5.01
4.05
3.95
4.77
3.51
3.15
4.87
3.79
3.59
3.38
3.58
5.87
4.76

)4.37 (Li2 )
)3.23 (H2 )
)8.98 (CN)
)9.33 (SO2 )
)13.12 (SO2 )
)7.00 (P2 )
)6.95 (CS)
)10.21 (SO2 )
)8.29 (P2 )
)9.03 (SO2 )
)12.82 (SO2 )
)12.01 (SO2 )
)13.48 (SO2 )
)10.42 (C2 H6 )
)9.59 (C2 H6 )

25.47 (CO2 )
29.64 (CO2 )
10.34 (N2 H4 )
8.62 (BeH)
8.58 (BeH)
13.66 (N2 H4 )
9.96 (N2 H4 )
7.60 (BeH)
13.37 (N2 H4 )
9.68 (N2 H4 )
8.33 (BeH)
8.33 (BeH)
8.21 (BeH)
15.39 (O2 )
14.96 (O2 )

Table 3
Statistical data for ionization potentials (eV) of the small G2 set (22 molecules) with aug-cc-pVQZ Gaussian basis sets
a

ab

MAE

RMS

Maximum deviation (
)

Maximum deviation (+)

LC-BOP
LC-BLYP
CAM-BOP

0.0
0.0
0.2

CAM-BLYP

0.2

CAM-B3LYP

0.2

B3LYP(G)
B3LYP
BLYP
HCTH

0.2
0.2
0.0
0.0

1.0
1.0
1.0
0.8
0.6
1.0
0.8
0.6
1.0
0.8
0.6
0.2
0.2
0.0
0.0

0.192
0.230
0.166
0.158
0.153
0.195
0.185
0.178
0.243
0.218
0.202
0.209
0.169
0.208
0.154

0.220
0.272
0.203
0.189
0.185
0.245
0.225
0.213
0.301
0.273
0.249
0.254
0.206
0.250
0.187

)0.312
)0.245
)0.337
)0.358
)0.375
)0.272
)0.291
)0.311
)0.179
)0.199
)0.223
)0.210
)0.293
)0.459
)0.359

0.386
0.552
0.488
0.367
0.291
0.558
0.446
0.408
0.658
0.537
0.499
0.532
0.423
0.544
0.340

(Be)
(Be)
(Be)
(Be)
(Be)
(Be)
(Be)
(Be)
(Be)
(Be)
(P2 )
(Be)
(Be)
(Cl2 )
(Cl2 )

(F)
(O)
(N2 )
(N2 )
(O2 )
(N2 )
(O2 )
(O2 )
(N2 )
(N2 )
(O2 )
(O)
(O)
(O)
(N)

T. Yanai et al. / Chemical Physics Letters 393 (2004) 5157

55

Table 4
Statistical data for total atomic energies (hartree) of HNe with aug-cc-pVQZ Gaussian basis sets
a

ab

MAE

RMS

Maximum deviation (
)

Maximum deviation (+)

Deviation H

LC-BOP
LC-BLYP
CAM-BOP

0.0
0.0
0.2

CAM-BLYP

0.2

CAM-B3LYP

0.2

B3LYP(G)
B3LYP
BLYP
HCTH

0.2
0.2
0.0
0.0

1.0
1.0
1.0
0.8
0.6
1.0
0.8
0.6
1.0
0.8
0.6
0.2
0.2
0.0
0.0

0.084
0.085
0.068
0.049
0.030
0.069
0.050
0.031
0.033
0.014
0.009
0.019
0.005
0.011
0.008

0.093
0.128
0.075
0.054
0.049
0.075
0.054
0.034
0.035
0.017
0.013
0.023
0.005
0.016
0.010

)0.025
)0.040
)0.006
)0.031
)0.019

0.134
0.128
0.112
0.080
0.049
0.107
0.075
0.043
0.045
0.029
0.013

0.009
0.007
0.004

0.0091
0.0091
0.0075
0.0061
0.0047
0.0075
0.0061
0.0047
0.0034
0.0019
0.0005
)0.0024
0.0010
0.0021
)0.0064

Increasing the ratio of the Becke 88 exchange contribution at long-range reduces the errors of the atomization, ionization and the total atomic energies. For
instance, the CAM-BLYP with a 0:2 and a b 0:6
is better by 1.62 kcal mol
1 for MAE and RMS than
that with a 0:2 and a b 1:0. The maximum deviations are also reduced. While the CAM-BLYP method
with a 0:2 and a b 0:6 gives the smallest errors for
the atomization energies of the three CAM functionals
with a MAE of 2.28 kcal mol
1 , and CAM-BOP with
a 0:2 and a b 0:6 gives the best ionization energies with an MAE of 0.153 eV, we recommend CAMB3LYP due to its success in reproducing total atomic
energies.
We nd that the optimal values for the two parameters a and b of CAM-B3LYP which yield the smallest
errors for the atomization energies, are found to be

(Ne)
(Ne)
(F)
(F)
(Ne)

a 0:19 and b 0:46;

(Ne)
(Ne)
(Ne)
(Ne)
(Be)
(Ne)
(Ne)
(Ne)
(Be)
(Be)
(Be)
(Be)
(Be)
(C)

a b 0:65:

This functional yields results comparable to those of

B3LYP for the atomization energies, ionization potentials and total atomic energies (Table 5). The sum a b
is rather crucial, because it gives the asymptotic coecient of
r
1 . If we set a b 0:8, (i.e., 0.15 more
HF exchange at the long-range) we nd optimal values
to be
a 0:23

and

b 0:57;

10

for CAM-B3LYP. In Section 4.2, we use these two sets

of a and b to investigate how the long-range HF exchange interaction eects the charge transfer excitations
in TDDFT calculations. Table 5 summarizes the atomization energies, ionization potentials, total atomic energies with the above two forms of CAM-B3LYP. Both

Table 5
Statistical data for atomization energies (kcal mol
1 ) (with 53 molecules), ionization potentials (eV) (with 22 atoms and molecules), and total atomic
energies (hartrees) (with H atom through Ne atom) from the small G2 set using CAM-B3LYP with a b 0:65; 0:8, compared to B3LYP and LCBLYP with aug-cc-pVQZ Gaussian basis set
ab

MAE

RMS

Maximum deviation (
)

Maximum deviation (+)

Atomization energy (kcal mol
1 )

CAM-B3LYP
0.19
0.23
B3LYP
0.2
LC-BLYP
0.0

0.65
0.8
0.2
1.0

2.53
2.91
2.68
9.48

3.46
3.88
3.58
12.07

)10.79 (SO2 )
)12.25 (SO2 )
)13.48 (SO2 )
)3.23 (H2 )

8.34 (BeH)
8.41 (BeH)
8.21 (BeH)
29.64 (CO2 )

Ionization potentials (eV)

CAM-B3LYP
0.19
0.23
B3LYP
0.2
LC-BLYP
0.0

0.65
0.8
0.2
1.0

0.208
0.213
0.169
0.230

0.256
0.271
0.206
0.272

)0.212
)0.203
)0.293
)0.245

0.503
0.563
0.423
0.552

(O2 )
(N2 )
(O)
(O)

Total atomic energies (hartree)

0.19
CAM-B3LYPa
0.23
B3LYP
0.2
LC-BLYP
0.0

0.65
0.8
0.2
1.0

0.009
0.012
0.005
0.085

0.011
0.015
0.005
0.128

)0.017 (F)

)0.006 (F)

0.018
0.026
0.009
0.128

(Be)
(Be)
(Be)
(Ne)

Name

(Be)
(Be)
(Be)
(Be)

a
Deviation of H atom from the exact atomic energy (hartrees): 0.0010 (CAM-B3LYP a 0:19, a b 0:65), 0.0017 (CAM-B3LYP a 0:23,
a b 0:80).

7.77 (0.0816)
7.43 (0.3572)
7.24 (0.0399)

forms are similar in quality to B3LYP, and are a great

improvement over LC-BLYP.

6.98 (0.0238)
6.87 (0.2863)
5.16 (0.0043)

7.19 (0.0645)
7.23 (0.2704)
6.06 (0.0029)

6.82 (0.2397)
5.07 (0.0041)

7.57 (0.0830)
7.26 (0.3437)
7.02 (0.0430)

7.52 (0.0869)
7.32 (0.3555)
6.94 (0.0184)

4.2. Charge transfer excitations in the dipeptide model

Oscillator strengths within parentheses.

NV1
NV2
CT
1

A0
A0
1 0
A

p1 ! p1
p2 ! p2
p1 ! p2

6.32 (0.2737)
6.29 (0.3205)
6.92 (0.1168)

5.65 (0.0008)
5.88 (0.0004)
7.88 (0.0001)
Yanai et al. (this work)
5.34 (0.0009)
5.56 (0.0007)
5.62 (0.0005)
5.80 (0.0003)
4.60 (0.0002)
8.38 (0.0003)
5.43 (0.0009)
5.70 (0.0005)
4.67 (0.0000)
W1
W2
CT
1

A00
1 00
A
1 00
A

n1 ! p1
n2 ! p2
n1 ! p2

Tozer et al. [7]

5.61 (0.0006)
5.82 (0.0006)
7.92 (0.0000)

5.49 (0.0013)
5.73 (0.0009)
6.24 (0.0000)

CAM-B3LYP
a 0:19,
a b 0:65
LC-BLYP
BLYP
B3LYP
AC-HCTH
CASPT2
Band

Table 6
The electronic spectrum (eV) of a dipeptide (C5 N2 O2 H10 ) with the basis sets TZ2P (5s4p2d on rst row; 3s2p on H)

5.72 (0.0008)
5.95 (0.0004)
8.58 (0.0001)

T. Yanai et al. / Chemical Physics Letters 393 (2004) 5157

CAM-B3LYP
a 0:23,
a b 0:8

56

Excited state calculations with TDDFT were performed on a glycine dipeptide model system, which was
previously studied by Tozer and co-workers [7]. We used
the same geometry with the same basis sets TZ2P
(5s4p2d on rst row; 3s2p on H) as used in Tozers
study. Table 6 lists the excitation energies and the oscillator strengths. It is to be noted that the n ! p excitations within the same fragment are well produced by
all the DFT methods, when compared to the highly
accurate multireference perturbation calculation with
CASPT2 [30].
The charge transfer excitations of n ! p and p ! p
between dierent fragments are very poorly predicted by
the DFT calculations with the standard exchangecorrelation functionals, in particular the BLYP for n ! p
is much lower by 3.3 eV than CASPT2. B3LYP yields
values 1.7 eV lower than CASPT2. This means that the
0:2r
1 contribution via HF exchange in B3LYP improves the charge transfer excitations. Tozer and Handys asymptotic correction does not work for these
states.
It is very encouraging that the CAM-B3LYP with
a 0:19 and a b 0:65 predictions of charge transfer
excitations are in excellent agreement with the CASPT2
results. It is not surprising that the LC-BLYP method
brings the charge transfer excitations into much better
agreement with CASPT2 than the standard DFT exchangecorrelation functionals, because of the 1:0r
1
asymptotic form through the HF exchange. The CAMB3LYP functional with a 0:23 and a b 0:8 yields
almost comparable results to LC-BLYP. By comparison
between the two sets of results obtained with CAMB3LYP, we see that the long-range contribution of the
HF exchange interaction has a large impact on the
charge transfer excitations as compared with the other
calculated excitations.
The predictions for the p ! p excitation within the
same fragment are overestimated by all the DFT
methods. A clue to the problem lies in the oscillator
strengths. In DFT p2 ! p2 is dominant, whereas
p1 ! p1 and p2 ! p2 have equal strengths in CASPT2.
If the problem is truly multicongurational, then DFT
will fail. This is the most plausible explanation.

5. Conclusions
The recent work by Tsuneda and co-workers [1,10]
has shown how DFT studies can now be performed to a
useful accuracy for polarizability of long chains, excitations to Rydberg states and charge transfer excita-

T. Yanai et al. / Chemical Physics Letters 393 (2004) 5157

tions. Previously DFT calculations using GGA functionals (BLYP) and hybrid functionals (B3LYP) have
been frequently used for energetic studies and structural
studies. A combination of the two clearly goes a long
way for the provision of a very useful computational
chemistry methodology.
We are suggesting that the CAM-B3LYP functional
presented in this Letter meets these criteria. In the
CAM-B3LYP functional, we have replaced the Becke
parameter a by two parameters a, b for mixing Becke
1988 exchange and HF exchange, with l describing the
conversion from one to the other through Eq. (7). Our
best functional uses a 0:19, a b 0:65 and l 0:33
(Tawadas value). We used Beckes VWN5/LYP mixing
parameter ( 0.19) for the correlation functional without adjustment. Our calculations have shown that this
functional predicts energetic quantities to the accuracy
of B3LYP. Our rst investigations on charge transfer
energies suggest that it is possible to achieve chemical
accuracy (0.1 eV). The cost for the implementation of
this functional is no more severe than that required for
B3LYP (within a factor of 2 for computing two-electron
integrals). This functional is a hybrid functional with
improved long-range properties.

Acknowledgements
We acknowledge Dr. R.J. Harrison for the encouragement of this work, the use of the computer resources
at ORNL, and arranging TYs visit to NCHs group at
Cambridge. We would like to thank Professor K. Hirao,
Dr. T. Tsuneda, Y. Tawada and Dr. M. Sprik for
valuable discussions, Dr. D.J. Tozer for providing us
with input les for the dipeptide, Dr. A.J. Cohen for
input les of G2 set molecules. TY thanks Dr. S. Hirata
for teaching TY the implementation of TDDFT, and M.
Kamiya for his DFT quadrature code in IN T E G R A of
UT C H E M . NW C H E M Version 4.5, as developed and
distributed by Pacic Northwest National Laboratory,
P.O. Box 999, Richland, Washington 99352 USA, funded by the US Department of Energy, was used to obtain the DFT results.
References
[1] Y. Tawada, T. Tsuneda, S. Yanagisawa, T. Yanai, K. Hirao,
J. Chem. Phys. 120, in press.
[2] A.D. Becke, J. Chem. Phys. 98 (1993) 5648.

57

[3] P.J. Stephens, J.F. Devlin, C.F. Chabalowski, M.J. Frisch,

J. Chem. Phys. 98 (1993) 11623.
[4] M.E. Casida, in: D.P. Chong (Ed.), Recent Advances in Density
Functional Methods, vol. 1, World Scientic, Singapore, 1995.
[5] R. Bauernschmitt, R. Ahlrichs, Chem. Phys. Lett. 256 (1996)
454.
[6] D.J. Tozer, N.C. Handy, J. Chem. Phys. 109 (1998) 10180.
[7] D.J. Tozer, R.D. Amos, N.C. Handy, B.O. Roos, L. SerranoAndres, Mol. Phys. 97 (1999) 859.
[8] A. Dreuw, J.L. Weisman, M. Head-Gordon, J. Chem. Phys. 119
(2003) 2943.
[9] L. Bernasconi, M. Sprik, J. Hutter, J. Chem. Phys. 119 (2003)
12417.
[10] H. Iikura, T. Tsuneda, T. Yanai, K. Hirao, J. Chem. Phys. 115
(2001) 3540.
[11] A. Savin, in: J.M. Seminario (Ed.), Recent Developments and
Applications of Modern Density Functional Theory, Elsevier,
Amsterdam, 1996.
[12] P.M.W. Gill, R.D. Adamson, J.A. Pople, Mol. Phys. 88 (1996)
1005.
[13] T. Leiniger, H. Stoll, H.-J. Werner, A. Savin, Chem. Phys. Lett.
275 (1997) 151.
[14] A.D. Becke, Phys. Rev. A 38 (1988) 3098.
[15] J.P. Perdew, K. Burke, M. Ernzerhof, Phys. Rev. Lett. 77 (1996)
3865.
[16] T. Tsuneda, T. Suzumura, K. Hirao, J. Chem. Phys. 110 (1999)
10664.
[17] M. Kamiya, T. Tsuneda, K. Hirao, J. Chem. Phys. 117 (2002)
6010.
[18] R.H. Hertwig, W. Koch, Chem. Phys. Lett. 268 (1997) 345.
[19] C. Lee, W. Yang, R.G. Parr, Phys. Rev. B 37 (1988) 785.
[20] S.J. Vosko, L. Wilk, M. Nusair, Can. J. Phys. 58 (1980) 1200.
[21] D.M. Ceperley, B.J. Alder, Phys. Rev. Lett. 45 (1980) 566.
[22] F.A. Hamprecht, A.J. Cohen, D.J. Tozer, N.C. Handy, J. Chem.
Phys. 109 (1998) 6264.
[23] H. Iikura, H. Inoue, M. Kamiya, Y. Kawashima, T. Nakajima, Y.
Nakao, K. Nakayama, H. Sekino, K. Sorakubo, T. Tawada, T.
Tsuneda, K. Yagi, S. Yanagisawa, T. Yanai, K. Hirao, IN T E G R A ,
Department of Applied Chemistry, School of Engineering, University of Tokyo, Japan.
[24] T. Yanai, M. Kamiya, Y. Kawashima, T. Nakajima, H. Nakano,
Y. Nakao, H. Sekino, J. Paulovic, T. Tsuneda, S. Yanagisawa, K.
Hirao, UT C H E M 2004, Department of Applied Chemistry, School
of Engineering, University of Tokyo, Japan.
[25] T. Yanai, H. Nakano, T. Nakajima, T. Tsuneda, S. Hirata, Y.
Kawashima, Y. Nakao, M. Kamiya, H. Sekino, K. Hirao, in:
Computational Science ICC 2003, Lecture Notes in Computer
Science, Springer, 2003, p. 84.
[26] High Performance Computational Chemistry Group, NW C H E M ,
A Computational Chemistry Package for Parallel Computers,
Version 4.5, Pacic National Laboratory, Richland, Washington
99352, USA, 2003.
[27] L.A. Curtiss, K. Raghavachari, P.C. Redfern, J.A. Pople,
J. Chem. Phys. 106 (1997) 1063.
[28] N.C. Handy, D.J. Tozer, Mol. Phys. 94 (1998) 707, References of
molecular bond lengths and bond angles therein.
[29] D. Feller, K.A. Peterson, J. Chem. Phys. 110 (1999) 8384.
 Malmqvist, B.O. Roos, A.J. Sadlej, K.
[30] K. Andersson, P.-A.
Wolinski, J. Phys. Chem. 94 (1990) 5483.