Introduction to Electron Correlation

C. David Sherrill
School of Chemistry and Biochemistry
Georgia Institute of Technology
Molecular Orbitals Arent Real...
Molecular orbitals are the eigenvectors of the Fock matrix in
Hartree-Fock theory. Hartree-Fock is just an approximation,
even if a complete orbital basis is used. Why? Two reasons,
equivalent to each other:
It assumes the wavefunction can be written as one Slater
determinant. This is wrong!
It assumes that each electron interacts with an average
charge distribution due to the other electrons. This is
wrong!
Electron Correlation
Electrons repel each other according to Coulombs law,
with the repulsion energy r
1
ij
.
Hartree-Fock replaces this instantaneous electron-electron
repulsion with the repulsion of each electron with an
average electron charge cloud.
This introduces an error in the wavefunction and the
energy. The energy error is called the total correlation
energy. The error in the total energy is about 1 eV per
electron pair in a bond or lone pair.
Dispersion forces are due to electron correlation; expect
Hartree-Fock to fail badly for van der Waals complexes.
Correlation Energy
The correlation energy is dened as the dierence between
the true energy and the Hartree-Fock energy in a complete
basis (Hartree-Fock limit)
E
corr
= c
exact
E

HF
Usually we dont know the exact energy c
exact
, but
sometimes we can compute the exact energy for a given
one-electron basis set. This lets us compute the basis set
correlation energy (or just correlation energy for short...)
E
basis
corr
= E
basis
exact
E
basis
HF
Why its called Electron Correlation
The Hartree Product

HP
(x
1
, x
2
, , x
N
) =
i
(x
1
)
j
(x
2
)
k
(x
N
)
is completely uncorrelated, in the sense that the probability of
simultaneously nding electron 1 at x
1
, electron 2 at x
2
, etc.,
is given by [
HP
(x
1
, , x
N
)[
2
dx
1
dx
N
and is just
[
i
(x
1
)[
2
dx
1
[
j
(x
2
)[
2
dx
2
[
k
(x
N
)[
2
dx
N
,
which is the probability of nding electron 1 at x
1
times the
probability of nding electron 2 at x
2
, etc....the product of the
probabilities. This makes the Hartree Product an independent
particle model. Electrons move independently; their motion is
uncorrelated.
Correlation in Hartree-Fock
Hartree-Fock is usually dened as uncorrelated. However,
the electron motions are no longer completely independent.
For two electrons with the same spin
For two electrons with dierent spins, [
1
(r
1
)(
1
)
2
(r
2
)(
2
)),
the probability of nding electron 1 at r
1
and electron 2 at
r
2
is
P(r
1
, r
2
)dr
1
dr
2
= dr
1
dr
2
_
d
1
d
2
[[
2
=
1
2
_
[
1
(r
1
)[
2
[
2
(r
2
)[
2
+ [
1
(r
2
)[
2
[
2
(r
1
)[
2
_
dr
1
dr
2
The electrons are uncorrelated.
For two electrons with the same spin, [
1
(r
1
)(
1
)
2
(r
2
)(
2
)),
the probability of nding electron 1 at r
1
and electron 2 at
r
2
is
P(r
1
, r
2
)dr
1
dr
2
=
1
2
_
[
1
(r
1
)[
2
[
2
(r
2
)[
2
+ [
1
(r
2
)[
2
[
2
(r
1
)[
2
[

1
(r
1
)
2
(r
1
)

2
(r
2
)
1
(r
2
)
+
1
(r
1
)

2
(r
1
)
2
(r
2
)

1
(r
2
)]) dr
1
dr
2
Now P(r
1
, r
1
) = 0. No 2 electrons with same spins can be
at the same place. Fermi hole. Same-spin electrons are
correlated in Hartree-Fock, dierent-spin electrons are not.
The N-electron Basis
A collection of atom-centered Gaussian functions can be
used as a basis set for expanding one-electron functions
(molecular orbitals).
We need to solve the electronic Schr odinger equation to
get
e
(x
1
, x
2
, , x
N
), a function of N electrons. What
can we use as a basis for expanding
e
?
Slater determinants are proper N-electron basis functions:
they are functions which can be used to expand any
(antisymmetric) N-electron function.
In the limit of an innite number of Slater determinants,
any N-electron function can be expanded exactly.
What Other Determinants?
If it takes an innite number of determinants to exactly
represent the wavefunction, taking only one, as in Hartree-Fock
theory, is not necessarily a decent approximation!
For a given set of MOs, other determinants can be formed by
putting electrons in dierent orbitals. These other determinants
will generally belong to dierent electron congurations.
Example: Minimal Basis H
2
There are four possible determinants, one for each spin-orbital
conguration:
(
g
)
2
(
g
)(

u
) (
g
)(

u
) (

u
)
2
Note that the middle two correspond to the same spatial orbital
conguration but are dierent spin orbital congurations or
Slater determinants.
Note also that the middle two determinants can be discarded in
any expansion of the
1

+
g
state of H
2
. Why??
Using Symmetry
We only need determinants with the right symmetries. This
includes: spatial symmetry,

S
z
symmetry, and

S
2
symmetry.
Determinants with values dierent than the desired state
cannot contribute to the total wavefunction.
Of the four given determinants for H
2
,
(
g
)
2
(
g
)(

u
) (
g
)(

u
) (

u
)
2
,
all have M
s
= 0. The rst and last are also eigenfunctions of

S
2
, but the middle two are not we would need to take +/-
combinations of them. Conguration State Functions (CSFs)
are combinations of one or more Slater determinants with the
proper values of spatial symmetry, M
s
, and S.
Conguration Interaction (CI)
Conguration Interaction: Express the wavefunction as a
linear combination of Slater determinants. Usually obtain
expansion coecients variationally. [) =

I
c
I
[
I
).
Complete CI: Arrange all electrons in all possible ways
(consistent with symmetry requirements) in a complete
(i.e., innite) orbital basis set. Unattainable.
Full CI: Arrange all electrons in all possible ways (w/ sym-
metry) for a nite orbital basis set. Very costly!
How Many Determinants Could There Be?
Ignoring

S
2
and spatial symmetry, but keeping

S
z
symmetry,
how many determinants can be formed? If there are N

(N

)
electrons with () spin, in n orbitals, one can keep M
s
xed
by moving the and electrons separately.
There are very many ways to do this!
N
det
=
_
_
_
n
N

_
_
_
_
_
_
n
N

_
_
_.
Number of M
s
= 0 Determinants (CSFs in parentheses)
Number of electrons
Orbitals 6 8 10 12
10 14.4 10
3
44.1 10
3
63.5 10
3
44.1 10
3
(4.95 10
3
) (13.9 10
3
) (19.4 10
3
) (13.9 10
3
)
20 1.30 10
6
23.5 10
6
240 10
6
1.50 10
9
(379 10
3
) (5.80 10
6
) (52.6 10
6
) (300 10
6
)
30 16.5 10
6
751 10
6
20.3 10
9
353 10
9
(4.56 10
6
) (172 10
6
) (4.04 10
9
) (62.5 10
9
)
Full CI is Impractical
About 1 billion determinants is the limit of what can be
reached today for Full CI computations. This requires
highly optimized computer code!
For a determinant basis, this means about 12 electrons in
20 orbitals, or 10 electrons in 30 orbitals. Mostly diatomics.
Need to separate the more important determinants
from the less important. A common way: truncate CI
expansion according to excitation level.
Classication by Substitution (Excitation) Classes
The Hartree-Fock reference determinant [
0
) should be the
leading term. Expect the importance of other congurations to
drop o rapidly as they substitute more orbitals. Let [
abc
ijk
)
denote a determinant which diers from [
0
) by replacing
orbitals ij k with ab c.
| = c
0
|
0
+

c
a
i
|
a
i
+

c
ab
ij
|
ab
ij
+

c
abc
ijk
|
abc
ijk

Reference Singles Doubles Triples
Examples of Excitation Levels
_

_
_

_
S
_

_
_

_
D
_

_
_

_
T
_

_
_

_

_

_
Q
How to obtain Coecients?
Variationally: This is the conguration interaction method.
Find the coecients which minimize the total energy.
Equivalent to diagonalizing

H is the basis of determinants.
(Many-Body) Perturbation Theory: Treat the Hartree-Fock
determinant [
0
) as the zeroth-order solution, E
(0)
=

i
,
E
(1)
= E
HF
, doubles (and singles for open-shell) contribute
to [
(1)
) and E
(2)
, etc.
Need to truncate expansion of determinants to make it practi-
cal.
Conguration Interaction
| = c
0
|
0
+

c
a
i
|
a
i
+

c
ab
ij
|
ab
ij
+

c
abc
ijk
|
abc
ijk

Reference Singles Doubles Triples
Usually truncate expansion at a given excitation level.
CISD: CI with singles and doubles. Scales as O(N
6
).
CISDT: CI with singles, doubles, triples. Scales as O(N
8
).
Can keep going until Full CI limit; more accurate as more
terms are added, but also more costly! (Why no CIS?)
Conguration Interaction Matrices
Its easy to prove that the coecients c
I
which minimize
the total energy E
CI
are the same as the eigenvectors of the
electronic Hamiltonian in the same basis of determinants.
Simple strategy for CI: form the Hamiltonian matrix H and
diagonalize it. In actuality, we do this cleverly to avoid extra
work.
Evaluate the matrix elements H
IJ
=
I
[

H[
J
) using Slaters
Rules. Determinants diering by more than two spin orbitals
are 0.
H =
_

0
|

H|
0
0
0
|

H|D 0 0
0 S|

H|S S|

H|D S|

H|T 0
D|

H|
0
D|

H|S D|

H|D D|

H|T D|

H|Q
0 T|

H|S T|

H|D T|

H|T T|

H|Q
0 0 Q|

H|D Q|

H|T Q|

H|Q
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
_

_
Slaters Rules for Matrix Elements
Identical Determinants: If the determinants are identical,
then

1
[

H[
1
) =
N

m
m[

h[m) +
N

m>n
mn[[mn)
Determinants that Dier by One Spin Orbital:
[
1
) = [ mn )
[
2
) = [ pn )

1
[

H[
2
) = m[

h[p) +
N

n
mn[[pn)
Determinants that Dier by Two Spin Orbitals:
[
1
) = [ mn )
[
2
) = [ pq )

1
[

H[
2
) = mn[[pq)
Determinants that dier by More than Two Spin Orbitals:
[
1
) = [ mno )
[
2
) = [ pqr )

1
[

H[
2
) = 0
Percent Corr. Energy
a
Molecule CISD CISDT CISDTQ
BH 94.91 n/a 99.97
H
2
O(R
e
) 94.70 95.47 99.82
H
2
O(1.5 R
e
) 89.39 91.15 99.48
H
2
O(2.0 R
e
) 80.51 83.96 98.60
NH
3
94.44 95.43 99.84
HF 95.41 96.49 99.86
H
+
7
96.36 96.87 99.96
a
Data from Harrison et al., 1983, except H
+
7
data from Fermann et al., 1994.
Size of CI space vs Excitation Level
CSFs required
a
Molecule CISD CISDT CISDTQ FCI
BH 568 n/a 28 698 132 686
H
2
O 361 3 203 17 678 256 473
NH
3
461 4 029 19 925 137 321
HF 552 6 712 48 963 944 348
H
+
7
1 271 24 468 248 149 2 923 933
a
Data from Harrison et al., 1983, except for H
+
7
data
from Fermann et al., 1994.
Conguration Interaction Is Not Size Extensive
CISD is exact for a two-electron system like H
2
H
2

...But not exact for two noninteracting H

H
2

Would need Quadruple excitations! Should be able to express

H =

H
0
+

H

The solution is expressed as a Taylor series in , the perturba-

H
0
=

f(i)
=

_
h(i) + v
HF
(i)
_

i<j
r
1
ij
1
HF
=

i<j
r
1
ij

v
HF
(i)
v
HF
(1)
p
(x
1
) =

i
i[r
1
12
[i)
p
(x
1
)

i
i[r
1
12
[p)
i
(x
1
)
p[v
HF
[q) = v
HF
pq
=

i
pi[[qi)
Hartree-Fock energy is Sum of Zero and First Order
Energies

H
0
[
0
) = E
(0)
0
[
0
)
E
(0)
0
=

i
E
(1)
0
=
0
[H

[
0
)
=
0
[

i<j
r
1
ij
[
0
)
0
[

i
v
HF
(i)[
0
)
=
1
2

ij
ij[[ij)

i
i[v
HF
[i)
=
1
2

ij
ij[[ij)
E
HF
0
= E
(0)
0
+ E
(1)
0
Second-Order Energy Expression
E
(2)
0
=

I=0
[
0
[

H

[
I
)[
2
E
(0)
0
E
(0)
I
The numerator is zero for all other determinants except double
excitations.
E
(2)
0
=

i<j,a<b
[ab[[ij)[
2

a
+
b

i

j
Second-order energy correction is negative; electron correlation
stabilizes the energy. Usually perturbation theory is taken
through second order, MBPT(2) (or simply MP2), which scales
as O(N
5
).
Coupled-Cluster Theory
[
CC
) = e

T
[
0
)
= (1 +

T +
1
2!

T
2
+
1
3!

T
3
+ )[
0
)
CCSD:

T =

T
1
+

T
2
. Cost O(N
6
).
CCSD(T): Adds perturbative correction for

T
3
at cost O(N
7
).
Improves over CISD, etc., by using products of excitations
Size extensive version of CI: CCSD exact for N noninter-
acting H
2
s
Contains higher-order disconnected substitutions
H
2

[
A
0
)
+

T
A
2
[
A
0
)
H
2

[
B
0
)
+

T
B
2
[
B
0
)
If

H =

H
A
+

H
B
, should have E = E
A
+ E
B
and
[) = [
A
)[
B
). Unfortunately,
[
CID
) ,= [
A
CID
)[
B
CID
)
(1 +

T
A
2
+

T
B
2
)[
A
0
)[
B
0
) ,=
_
(1 +

T
A
2
)[
A
0
)
__
(1 +

T
B
2
)[
B
0
)
_
However,
e

T
A
2
+

T
B
2
[
A
0
)[
B
0
) =
_
e

T
A
2
[
A
0
)
_ _
e

T
B
2
[
B
0
)
_
[
CCD
) = [
A
CCD
)[
B
CCD
)
Excitation Operators
Coupled-cluster theory makes use of excitation operators that
act on a reference determinant to generate excited determinants
times coecients

T
1
=

ia
t
a
i
[
a
i
)

T
2
=

ijab
t
ab
ij
[
ab
ij
)
With this notation, CISD is just
[
CISD
) = (1 +

T
1
+

T
2
)[
0
)
This makes the CCSD wavefunction
[
CCSD
) = e

T
1
+

T
2
[
0
)
= (1 +

T
1
+

T
2
+
1
2

T
2
1
+

T
1

T
2
+
1
2

T
2
2
+ )[
0
)
= [
0
) +

ia
t
a
i
[
a
i
) +
ab

ij
t
ab
ij
[
ab
ij
)
+
1
2
a

i
t
a
i
b

j
t
b
j
[
ab
ij
) +
a

i
t
a
i
bc

jk
t
bc
jk
[
abc
ijk
)
+
1
2
ab

ij
t
ab
ij
cd

kl
t
cd
kl
[
abcd
ijkl
) +
Coupled-Cluster Theory: An Ab Initio Success Story
Expected Errors for Large-Basis CCSD(T)
Bond lengths 0.004

A
Bond angles 0.3
o
Harmonic frequencies +2%
Dipole moments 0.05 D
IR intensities 20%
Excitation energies 0.2 eV
Convergent ab initio Methods
Ab initio methods such as CI and MBPT can be improved,
in principle, by including more and more terms (and
simultaneously increasing one-electron basis set size).
CI and CC become more accurate as higher substitution
levels are included (although cost grows dramatically).
MBPT does not always improve at higher orders; the
series is not guaranteed to converge. Current conventional
wisdom is to stop at second order.
Density functional theory methods are nonconvergent
methods in that no one knows how to improve them
systematically to get more and more accurate results.
Convergence and Pauling Points
Basis Set
Method STO-3G 6-31G* cc-pVTZ cc-pVQZ
SCF x
MP2 x x
CISD x x
CCSD x x
CCSD(T) x x
CCSDT x x

Full CI Truth