Lecture notes for the course

554017 Advanced Computational Chemistry

Pekka Manninen

2009
2
Contents

1 Many-electron wave functions 7

1.1 Molecular Schrödinger equation . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.1 Helium-like atom . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2 Molecular orbitals: the first contact . . . . . . . . . . . . . . . . . . . . . . 10
1.2.1 Dihydrogen ion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 Spin functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Antisymmetry and the Slater determinant . . . . . . . . . . . . . . . . . . 13
1.4.1 Approximate wave functions for He as Slater determinants . . . . . 14
1.4.2 Remarks on “interpretation” . . . . . . . . . . . . . . . . . . . . . . 15
1.4.3 Slater method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4.4 Helium-like atom revisited . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.5 Slater’s rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.6 Exercises for Chapter 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2 Exact and approximate wave functions 19

2.1 Characteristics of the exact wave function . . . . . . . . . . . . . . . . . . 19
2.2 The Coulomb cusp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 Hamiltonian for a He-like atom . . . . . . . . . . . . . . . . . . . . 21
2.2.2 Coulomb and nuclear cusp conditions . . . . . . . . . . . . . . . . . 21
2.2.3 Electron correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Variation principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.1 Linear ansatz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.2 Hellmann–Feynman theorem . . . . . . . . . . . . . . . . . . . . . . 24
2.3.3 The molecular electronic virial theorem . . . . . . . . . . . . . . . . 25
2.4 One- and N-electron expansions . . . . . . . . . . . . . . . . . . . . . . . . 26
2.5 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.6 Exercises for Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3 Molecular integral evaluation 31

3.1 Atomic basis functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.1.1 The Laguerre functions . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.2 Slater-type orbitals . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3
 3.1.3 Gaussian-type orbitals . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2 Gaussian basis sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3 Integrals over Gaussian basis sets . . . . . . . . . . . . . . . . . . . . . . . 37
3.3.1 Gaussian overlap distributions . . . . . . . . . . . . . . . . . . . . . 38
3.3.2 Simple one-electron integrals . . . . . . . . . . . . . . . . . . . . . . 38
3.3.3 The Boys function . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.4 Obara–Saika scheme for one-electron Coulomb integrals . . . . . . . 40
3.3.5 Obara–Saika scheme for two-electron Coulomb integrals . . . . . . . 41
3.4 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.5 Exercises for Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4 Second quantization 45
4.1 Fock space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1.1 Creation and annihilation operators . . . . . . . . . . . . . . . . . . 46
4.1.2 Number-conserving operators . . . . . . . . . . . . . . . . . . . . . 48
4.2 The representation of one- and two-electron operators . . . . . . . . . . . . 49
4.3 Commutators and anticommutators . . . . . . . . . . . . . . . . . . . . . . 51
4.3.1 Evaluation of commutators and anti-commutators . . . . . . . . . . 52
4.4 Orbital rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.5 Spin in second quantization . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.5.1 Operators in the orbital basis . . . . . . . . . . . . . . . . . . . . . 54
4.5.2 Spin properties of determinants . . . . . . . . . . . . . . . . . . . . 57
4.6 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.7 Exercises for Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

5 Hartree–Fock theory 61
5.1 The Hartree–Fock approximation . . . . . . . . . . . . . . . . . . . . . . . 61
5.2 Restricted and unrestricted Hartree–Fock theory . . . . . . . . . . . . . . . 63
5.2.1 Hartree–Fock treatment of H2 . . . . . . . . . . . . . . . . . . . . . 63
5.3 Roothaan–Hall equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.3.1 On the convergence of the SCF method . . . . . . . . . . . . . . . . 66
5.3.2 Integral-direct SCF . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.4 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.5 Exercises for Chapter 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

6 Configuration interaction 73
6.1 Configuration-interaction wave function . . . . . . . . . . . . . . . . . . . . 73
6.1.1 Single-reference CI wave functions . . . . . . . . . . . . . . . . . . . 74
6.1.2 Multi-reference CI wave functions . . . . . . . . . . . . . . . . . . . 74
6.1.3 Optimization of the CI wave function . . . . . . . . . . . . . . . . . 75
6.1.4 On the disadvantages of the CI approach . . . . . . . . . . . . . . . 75
6.2 Multi-configurational SCF theory . . . . . . . . . . . . . . . . . . . . . . . 76
6.2.1 MCSCF wave function of H2 . . . . . . . . . . . . . . . . . . . . . . 76

4
 6.2.2 Selection of the active space . . . . . . . . . . . . . . . . . . . . . . 77
6.3 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.4 Exercises for Chapter 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

7 Description of dynamical correlation 81

7.1 Coupled-cluster theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
7.1.1 The CC Schrödinger equation . . . . . . . . . . . . . . . . . . . . . 81
7.1.2 The coupled-cluster exponential ansatz . . . . . . . . . . . . . . . . 82
7.1.3 CI and CC models compared . . . . . . . . . . . . . . . . . . . . . 83
7.1.4 On coupled-cluster optimization techniques . . . . . . . . . . . . . . 86
7.1.5 The closed-shell CCSD model . . . . . . . . . . . . . . . . . . . . . 87
7.1.6 The equation-of-motion coupled-cluster method . . . . . . . . . . . 88
7.1.7 Orbital-optimized coupled-cluster theory . . . . . . . . . . . . . . . 88
7.2 Perturbation theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7.3 Local correlation methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.4 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.5 Exercises for Chapter 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

8 Performance of the electronic-structure models 97

8.1 Molecular equilibrium geometries . . . . . . . . . . . . . . . . . . . . . . . 97
8.2 Molecular dipole moments . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
8.3 Reaction enthalpies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
8.4 Notes on total electronic energies . . . . . . . . . . . . . . . . . . . . . . . 102

5
6
Chapter 1

Many-electron wave functions

1.1 Molecular Schrödinger equation

The Schrödinger equation for a system of N electrons is1

ĤΨ(x1 , x2 , . . . , xN ) = EΨ(x1 , x2 , . . . , xN ), (1.1)

where Ĥ is the molecular Hamiltonian, which is for a molecule that contains N electrons
and K nuclei (assumed to be fixed in space)
N
X N X
X N
Ĥ = ĥi + ĝij , (1.2)
i=1 i=1 j=i+1

where ĥi is the one-electron Hamiltonian of electron i moving in an electrostatic potential

caused by the nuclei, and ĝij is the interaction between the electrons i and j:
K
1 X ZI 1
ĥi = − ∇2i − ; ĝij = . (1.3)
2 I=1
|ri − RI | |ri − rj |

The eigenfunctions or wave functions Ψ(x1 , . . . , xN ) describe the electronic structure and
properties of the molecule. {x} are used to collectively symbolize all variables needed to
refer to particles; here the position and the internal degree of freedom, spin. It is necessary
that the wave function is square-integrable,
Z
Ψ⋆ (x1 , x2 , . . . , xN )Ψ(x1 , x2 , . . . , xN )dx1 dx2 · · · dxN = K (K finite). (1.4)

Usually we assume that the wave function is normalized, i.e., K = 1.

The corresponding eigenvalues are in the present context the quantized energy levels of
the electronic “cloud” of the molecule. Any two eigenfunctions that correspond to different
1
In atomic units used throughout this course, where ~ = me = 4πǫ0 ≡ 1; c = 1/α ≈ 137.

7
energy values Em , En possess the orthogonality property
Z
Ψ⋆m Ψn dx1 dx2 · · · dxN = 0. (1.5)

Even if two or more eigenvalues happen to be identical, corresponding to a degenerate

state of the system, the different eigenfunctions may still be assumed orthogonal.
If we would like to include the effects of nuclear motion, the molecular Hamiltonian
becomes (now restricting to electrostatic forces only)

Ĥ = Ĥe + Ĥn + Ĥen (1.6)

1 X 2 1 X′
Ĥe = − ∇i + ĝij
2 i 2 ij
1 X 1 2 1 X′
Ĥn = − ∇ + ZI ZJ ĝIJ
2 I mI I 2 IJ
X X ZI
Ĥen = − .
I i
r Ii

The wave function is correspondingly dependent on both the electron and nuclear variables.
Usually, because of the large ratio mn /me , it is a good approximation to separate these,
and we may turn our attention to the electronic problem. This is referred to as the Born–
Oppenheimer approximation. The nuclear-motion Schrödinger equation plays, however, a
central role in molecular spectroscopy. The motion of nuclei can sometimes affect even
those properties that are considered electronic. This is the case e.g. in the Jahn–Teller
effect and the Renner effect. For fixed nuclei, we can write the total energy of the system
simply as
1 X ′ ZI ZJ
Etot = Eelectronic + . (1.7)
2 IJ |RI − RJ |

Eq. (1.1) provides a basis for molecular quantum mechanics and all the static electronic
properties of the molecule may be obtained by expectation values of the appropriate Her-
mitian operators or using perturbation theory. However, we should note the limitations of
this model:

• Limitations in the Hamiltonian: non-relativistic kinematics, instant interactions, and

the Born–Oppenheimer approximation (that implies T = 0)

• Eq. (1.1) is soluble exactly only for a few trivial one-electron systems, such as
hydrogen-like atom or the molecule H+2 (see Exercise 1.1)

• Real experiments are not carried out for an isolated molecule, and observation always
involves interaction with the system, and thus a time-independent description is not
sufficient; but the time-independent form of (1.1), ĤΨ = i∂Ψ/∂t, should be applied.

8
1.1.1 Helium-like atom
For a helium-like atom with a point-like nucleus of charge Z the electronic Hamiltonian,
Eq. (1.2), is
1 Z 1 Z 1
Ĥ = ĥ1 + ĥ2 + ĝ12 = − ∇21 − − ∇22 − + .
2 r1 2 r2 r12
Due to ĝ12 , this is a three-body problem, and thereby no closed form solution exists for
the eigenvalue equation (1.1). We start our consideration by an approximation, where ĝ12
is set to zero, i.e. Ĥ0 = ĥ1 + ĥ2 . We can separate the variables in the eigenvalue equation
H0 Ψ0 = EΨ0 by substituting2

Ψ0 = Φ(r1 , r2 ) = χ1 (r1 )χ2 (r2 ),

to the eigenvalue equation and dividing by Φ, obtaining

ĥ1 χ1 (r1 ) ĥ2 χ2 (r2 )

+ = E,
χ1 (r1 ) χ2 (r2 )
meaning that Φ is a solution if (denoting E = ǫ1 + ǫ2 )

ĥ1 χ1 (r1 ) = ǫ1 χ1 (r1 )

ĥ2 χ2 (r2 ) = ǫ2 χ2 (r2 ),

meaning that χ1 and χ2 are solutions of the one-electron eigenvalue problem ĥχ = ǫχ. Φ
is a simultaneous eigenstate of the operators ĥ1 and ĥ2 ,

ĥ1 Φ(r1 , r2) = ǫ1 Φ(r1 , r2 )

ĥ2 Φ(r1 , r2) = ǫ2 Φ(r1 , r2 ).

The 1-electron 1-center eigenfunctions that are solutions to the one-electron eigenvalue
problem (this holds only in the absence of ĝ) are the same with the solutions of the
Schrödinger equation for a hydrogen-like atom (note the factor Z),

χnlm (r, θ, ρ) = Rnl (r)Ylm (θ, ρ), (1.8)

consisting of the radial part

 3/2 s  l    
2Z (n − l − 1)! 2Zr 2Zr Zr
Rnl (r) = L2l+1
n−l−1 exp − (1.9)
n 2n(n + 1)! n n n

and a spherical harmonic function, explicitly expressed as

s
2l + 1 (l − m)! m
Ylm (θ, ρ) = P (cos θ) exp(imρ). (1.10)
4π (l + m)! l
2
Considering only the variables r (position vectors), which in this case are most conveniently represented as the variables of spherical
coordinates, r = (r, θ, ρ)

9
In the above, L denotes the associated Laguerre polynomials

1 dn  
Lαn (x) = exp(x)x−α n exp(−x)xn+α (1.11)
n! dx
and P the Legendre polynomials

(−1)m 2 m/2 d
l+m
(l − m)! m
Plm (x) = (1 − x ) (x2 − 1)l ; Pl−m = (−1)m P (x). (1.12)
2l l! dxl+m (l + m)! l

These shall be referred to as atomic orbitals (AO). However, this is used also for other
atomic-centered functions with different radial parts, such as sc. Slater- or Gaussian-type
functions.

1.2 Molecular orbitals: the first contact

When moving from atoms to molecules, the first complication is that even the one-electron
eigenfunctions, even when ĝ = 0, are not obtainable in closed form. Instead of the
1s, 2s, 2p, . . . AOs we need polycentric molecular orbitals (MO) that describe the states of
a single electron spread around the whole nuclear framework.

1.2.1 Dihydrogen ion

The one-electron eigenvalue equation in the case of hydrogen molecule ion H+
2 is ĥφ = ǫφ,
where  
1 2 1 1
ĥ = − ∇ − + .
2 rA rB
This may be solved with high accuracy in confocal elliptic coordinates, the solution being
a product of three factors that are in turn solutions of three separate differential equations.
We will, however, go immediately to approximations. It is reasonable to expect that

φ(r) ∼ cA χA
100 (r) (densities close to nucleus A)
φ(r) ∼ cB χB
100 (r) (densities close to nucleus B)

where cA and cB are some numerical factors. An appropriate trial MO would then be

φ(r) = cA χA B
100 (r) + cB χ100 (r).

Due to the symmetry of the molecule, it is reasonably to assume that there is no higher
density in the other end than in the other; therefore

|φ(r)|2 = c2A [χA 2 2 B 2 A B

100 (r)] + cB [χ100 (r)] + 2cA cB χ100 (r)χ100 (r),

10
 n=1 n=2 n=3
l=0 l=0 l=1 l=0 l=1 l=2

m=0
11

m=1

m=2

Table 1.1: Some lowest-energy solutions χnlm (Z = 1) (blue for positive, orange for negative phase)
 Figure 1.1: The two lowest MOs of H+
2.

implies that cB = ±cA . Thus, it appears that we can construct two rudimentary MOs from
the 1s AOs on the two centers, denoted by

φ1 (r) = c′A [χA B

100 (r) + χ100 (r)]
φ2 (r) = c′B [χA B
100 (r) − χ100 (r)],

where the remaining numerical factors c′A and c′B are to be chosen so that the wave functions
will be normalized. These are sketched in Figure 1.1.
MOs build up this way, in linear combination of atomic orbitals (LCAO), are of great
importance throughout molecular quantum mechanics.

1.3 Spin functions

It is well-known that a single electron is not completely characterized by its spatial wave
function φ(r), but an “intrinsic” angular momentum, spin, is required to explain e.g. line
emission spectra of atoms in an external magnetic field. Therefore we must introduce a
spin variable ms in addition to r. We further associate operators ŝx , ŝy and ŝz with the
three components of spin angular momentum. We will adopt x as a variable taking into
account both degrees of freedom, ms and r, and then introduce spin-orbitals

ψ(x) = φ(r)σ(ms ). (1.13)

σ describes the spin state and is a solution of an eigenvalue equation ŝz σ = λσ.
If the orbital state is of energy ǫ, then

ĥψ = ǫψ (1.14)
ŝz ψ = λψ, (1.15)

and ψ is a state in which an electron simultaneously has definite energy and z-component
of spin.

12
 According to observations, the eigenvalue equation (1.15) has only two solutions for
electrons:
1 1
ŝz α = α; ŝz β = − β
2 2
often referred to as “spin-up” and “spin-down”. Therefore an orbital φ yields two possible
spin-orbitals, ψ = φα and ψ ′ = φβ with the normalization and orthogonality properties
Z Z Z Z
⋆ ⋆
α αdms = β βdms = 1, α βdms = β ⋆ αdms = 0.
⋆
(1.16)

Also the total spin operators

N
X
Ŝz = ŝi,z (1.17)
i=1
XN N
X
2


Ŝ = ŝ2i = ŝ2i,x + ŝ2i,y + ŝ2i,z (1.18)
i=1 i=1

with eigenvalues of M and S(S + 1), respectively, are needed in the discussion of many-
electron system. They commute with a spin-free electronic Hamiltonian; and any exact
stationary state Ψ is an eigenfunction of them.

1.4 Antisymmetry and the Slater determinant

The fact that electrons are indistinguishable, i.e. |Ψ(x1 , x2 )|2 = |Ψ(x2 , x1 )|2 , leads to (when
generalized for N electrons) two possibilities for a wave function:

P̂ Ψ(x1 , x2 , . . . , xN ) = Ψ(x1 , x2 , . . . , xN ) symmetric wave function

P̂ Ψ(x1 , x2 , . . . , xN ) = εP Ψ(x1 , x2 , . . . , xN ) antisymmetric wave function

where the operator P̂ permutes the arguments and εP is +1 for an even and −1 for an odd
number of permutations.
The choice for an system consisting of N electrons is once and for all determined by the
antisymmetry principle (Pauli exclusion principle): The wave function Ψ(x1 , x2 , . . . , xN )
that describes any state of an N-electron system is antisymmetric under any permutation
of the electrons.
Let us now work out the N-electron wave function. We start by introducing a product
function of one-electron wave functions (spin-orbitals):

Θ(x1 , x2 , . . . , xN ) = ψ1 (x1 )ψ2 (x2 ) · · · ψN (xN ). (1.19)

To yield a fully antisymmetric wave function, A number of differently permuted prod-

uct functions must be combined. The combination of permutations is most conveniently

13
expressed as a Slater determinant
 
 ψ1 (x1 ) ψ2 (x1 ) · · · ψN (x1 ) 
 
 ψ1 (x2 ) ψ2 (x2 ) · · · ψN (x2 ) 
Φ(x1 , x2 , . . . , xN ) = M  .. .. .. .. , (1.20)
 . . . . 

 ψ (x ) ψ (x ) · · · ψ (x ) 
1 N 2 N N N

by which all symmetric permutations are automatically rejected, and thus the Pauli prin-
ciple is fulfilled. Expanded in Slater determinants the exact wave function is written as
X
Ψ(x1 , x2 , . . . , xN ) = ck Φk (x1 , x2 , . . . , xN ). (1.21)
k

The normalization factor M is reasoned in the following way: when expanded, Φ consists
of N! products, and thus an integrand Φ⋆ Φ would consist of (N!)2 products, each of a form
±[ψ1 (xi )ψ2 (xj ) · · · ψN (xk )][ψ1 (xi′ )ψ2 (xj ′ ) · · · ψN (xk′ )]. It is evident that unless i = i′ , j =
j ′ etc. in a particular product, the product will give no contribution to the result, due to the
orthogonality of spin-orbitals. Therefore, there are N! non-vanishing contributions, R ⋆ and as
every contribution is equal to 1 with normalized spin-orbitals, we have Φ Φdx1 · · · dxN =
M 2 N!, hence the normalization factor has the value M = (N!)−1/2 .

1.4.1 Approximate wave functions for He as Slater determinants

The Slater determinant of the configuration (1s)2 of a helium atom in the minimal orbital
basis is  
1  χ100 (r1 )α(ms1 ) χ100 (r1 )β(ms1 )  1
Φ= √  ≡ √ |1sα 1sβ|.
2 χ100 (r2 )α(ms2 ) χ100 (r2 )β(ms2 )  2
Note the abbreviated notation that will be used quite extensively hereafter: only the
diagonal elements are displayed, and the explicit expression of variables is omitted – the
left-right order of the variables in this notation will be 1, 2, 3, .... Also, note that we have
denoted χnlm ’s with their corresponding AO labels.
The proper antisymmetric wave functions for the excited 1s2s configuration are in turn

1
Φ2 = √ |1sα 2sα|
2
1
Φ3 = (|1sα 2sβ| + |1sβ 2sα|)
2
1
Ψ4 = √ |1sβ 2sβ|
2
1
Ψ5 = (|1sα 2sβ| − |1sβ 2sα|) .
2

14
1.4.2 Remarks on “interpretation”
What is the physical meaning of the wave functions introduced earlier? First of all, the
general formulation of quantum mechanics is concerned only with symbolic statements in-
volving relationships between operators and operands, and is independent of the employed
“representation” or “interpretation”.

• In a very “nihilistic” interpretation, MOs as well as spin-orbitals are just mathemat-

ical means for the solution of the molecular Schrödinger equation and no meaning
can be given for them or even their squared norms.

• Perhaps more commonly, a spin-orbital ψ(x) itself is thought to be just a mathemat-

ical entity but |ψ(x)|2 dx is taken to be the probability of a point-like electron being
in an element dx, that is, in volume element dr with spin between ms and ms + dms .
|φ(r)|2dr is the probability of finding an electron with any spin in the volume ele-
ment dr. Also Ψ(x1 , x2 , . . . , xN ) cannot be interpreted; but |Ψ|2 dx1 dx2 · · · dxN is a
probability of the electron 1 in dx1 , electron 2 simultaneously in dx2 , etc. Then, the
probability ofR electron 1 being in dx1 while other electrons are anywhere would be
equal to dx1 Ψ⋆ Ψdx2 · · · dxN .

• We may also think an electron of being delocalized, i.e. not having a more precise
position than that indicated by its spatial part of the wave function φ(r). Then
|φ(r)|2dr would be the amount of charge in an element dr. In other words, in this
point of view a MO can be thought to approximately describe an electron spread
around the nuclear framework. However, as we will see later, MOs can be unitarily
“rotated”, i.e. their form shaped, and they still correspond to the same definite
energy. Which MOs are then closest to the physical ones?

1.4.3 Slater method

It is possible to find reasonable approximations for the wave function with only a small
number of determinants. The optimal coefficients ck may, in principle, be determined by
solving a secular problem
Hc = ESc, (1.22)
where the matrices H and S have the matrix elements
Z
Hµν = Ψ⋆µ (x1 , x2 , . . . , xN )ĤΨν (x1 , x2 , . . . , xN )dx1 dx2 · · · dxN
D   E
 
= Ψµ Ĥ  Ψν (1.23)
Z
Sµν = Ψ⋆µ (x1 , x2 , . . . , xN )Ψν (x1 , x2 , . . . , xN )dx1 dx2 · · · dxN = hΨµ |Ψν i. (1.24)

S is referred to as an overlap matrix. The equation (1.22) is the essence of the Slater
method. Observe the introduced Dirac notation for integrals. The coefficients c are then

15
determined by a condition called the secular determinant
 
 H11 − E H12 ... H1n 
 
 H21 H22 − E ... H2n 
 .. .. .. ..  = 0. (1.25)

 . . . . 

 H Hn2 . . . Hnn − E 
n1

1.4.4 Helium-like atom revisited

Consider the 2-electron Hamiltonian, now including electron–electron interaction, Ĥ =
ĥ1 + ĥ2 + ĝ. Let us assume that the wave function Ψ1 in Example 1.2 approximates the
(1s)2 configuration
D   E of the He atom. The corresponding energy is obtained as an expectation
 
value Ψ1 Ĥ  Ψ1 , or by writing the normalization explicitly, as
D   E
 
Ψ1 Ĥ  Ψ1
E1 = . (1.26)
hΨ1 |Ψ1 i
Now
Z
1
H11 = {χ⋆100 (r1 )χ100 (r2 ) [α(ms1 )β(ms2 ) − α(ms2 )β(ms1 )]⋆
2
o
×Ĥχ100 (r1 )χ100 (r2 ) [α(ms1 )β(ms2 ) − α(ms2 )β(ms1 )] dr1 dr2dms1 dms2 ,

but since the Hamiltonian does not operate on spins, we may perform the spin-integration
immediately [c.f. Eq. (1.16)],
Z
[α(ms1 )β(ms2 ) − α(ms2 )β(ms1 )]⋆ [α(ms1 )β(ms2 ) − α(ms2 )β(ms1 )] dms1 dms2 = 2.

Thus we have
Z
H11 = χ⋆100 (r1 )χ⋆100 (r2 )Ĥχ100 (r1 )χ100 (r2 )dr1 dr2
Z Z
= ⋆
χ100 (r1 )ĥ1 χ100 (r1 )dr1 χ⋆100 (r2 )χ100 (r2 )dr2
Z Z
+ χ100 (r2 )ĥ2 χ100 (r2 )dr2 χ⋆100 (r1 )χ100 (r1 )dr1
⋆

Z
+ χ⋆100 (r1 )χ⋆100 (r2 )ĝχ100 (r1 )χ100 (r2 )dr1 dr2
= 2ǫ1s + J

This result is still a crude approximation, the only difference to the independent-particle
model being the contribution J from the two-electron integral. The next step in the
determination of He energy levels would be to include more and more Slater determinants

16
corresponding to excited state configurations and evaluate the respective matrix elements.
As anticipated earlier, although the exact solution would require an infinite number of
included determinants, the energy spectrum would in practise converge towards the correct
one rather rapidly. However, the analysis of the many-electron problem using this kind of
direct expansion would obviously be extremely tedious, and a more general approach is
needed. Also, the evaluation of integrals gets far more complicated in the case of many
atomic cores.

1.4.5 Slater’s rules

We may indeed devise more systematic rules for the matrix elements. We begin byPnoting
that the expectation value over the 1-electron part of the molecular Hamiltonian, N i=1 ĥi ,
will reduce to the sum of N identical terms, since the coordinates of each electron appear
symmetrically in the corresponding integral:
*   +
X  X D   E
 
Φ ĥi  Φ = ψr ĥ ψr . (1.27)
  r
i

The expectation value of the two-electron part is a bit more complicated, as contribu-
tions may arise from terms that differ by an interchange of two electrons; but essentially a
similar argumentation leads to
*   +
X  X
 ′  ′
Φ ĝij  Φ = (hψr ψs |ĝ| ψr ψs i − hψr ψs |ĝ| ψs ψr i) , (1.28)
 
ij rs

where the former part is referred to as the Coulomb integral and the latter as the exchange
integral. the expectation value of the energy being simply the sum of (1.27) and (1.28).
We also need the off-diagonal elements. Fortunately enough, there are non-vanishing
elements only in two cases:

• one spin-orbital is different between Φ and Φ′ (ψr′ 6= ψr ) when

D   E D   E X
Φ′ Ĥ  Φ = ψr′ ĥ ψs + (hψr′ ψs |ĝ| ψr ψs i − hψr′ ψs |ĝ| ψs ψr i) ,
   
(1.29)
s6=r

• two spin-orbitals are different (ψr′ 6= ψr , ψs′ 6= ψs ) when

D   E
Φ′ Ĥ  Φ = (hψr′ ψs′ |ĝ| ψr ψs i − hψr′ ψs′ |ĝ| ψs ψr i) ,
 
(1.30)

These results are called Slater’s rules. In Slater’s rules, the spin-orbitals are required to be
orthonormal, but the rules may be generalized for matrix elements even in a non-orthogonal
spin-orbital basis.

17
1.5 Further reading
• R. McWeeny, Methods of Molecular Quantum Mechanics (Academic Press, 1992)

1.6 Exercises for Chapter 1

1. A general singlet two-electron state may be expanded as
X
Ψ(x1 , x2 ) = Cnlm,nlm |φnlmα φnlm β|
nlm
X
+ Cn1 l1 m1 ,n2 l2 m2 (|φn1l1 m1 α φn2 l2 m2 β| − |φn1 l1 m1 β φn2 l2 m2 α|),
(n1 l1 m1 )>(n2 l2 m2 )

where |φ1 φ2 | is the two-electron Slater determinant. Show that we may separate the
singlet function into a symmetric spatial part and an antisymmetric spin part:
1
Ψ(x1 , x2 ) = Ψ(r1 , r2 ) √ [α(1)β(2) − β(1)β(2)]
2

2. ⋆ Show, that the overlap integral between two normalized s orbitals centred on
nuclei A and B is
Z
1
S = 1sA (r)1sB (r)dr = (1 + R + R2 ) exp(−R)
3
where R is the internuclear separation. Hint: use the confocal elliptic coordinate
system (µ, ν, φ) with
rA + rB
µ = 1≤µ≤∞
R
rA − rB
ν = −1≤µ≤1
R
and 0 ≤ φ ≤ π. In it, the volume element is given by dr = 18 R3 (µ2 − ν 2 )dµdνdφ.

3. ⋆ Consider the solution of the Schrödinger equation for the hydrogen molecule using
a minimal basis, that is, 1s function for both of the atoms. Sketch the obtained 1σg
and 1σu orbitals, using the expression for the overlap integral given above.

18
Chapter 2

Exact and approximate wave

functions

2.1 Characteristics of the exact wave function

We shall need to approximate the molecular wave function – perhaps drastically – in order
to reach the applicability to molecules with multiple atomic cores and several electrons.
The approximations should, however, carried out with care. For this reason, we will now list
down the properties (either rationalized earlier or simply taken for granted) that an exact
wave function would possess, and on this basis we should seek an approximate solution
that retains as many of the following properties as possible.

The exact molecular electronic wave function Ψ

1. is antisymmetric with respect to the permutation of any pair of the electrons:

P̂ Ψ(x1 , . . . , xN ) = ǫP Ψ(x1 , . . . , xN ) (2.1)

where ǫP = ±1 for even and odd number of permutations, respectively.

2. is square-integrable everywhere in space,

hΨ|Ψi = K (K finite) (2.2)

3. is variational in the sense that for all possible variations δΨ, which are orthogonal to
the wave function, the energy remains unchanged:

hδΨ|Ψi = 0 ⇒ hδΨ|H|Ψi = 0 (2.3)

4. is size-extensive; for a system containing non-interacting subsystems the total energy

is equal to the sum of the energies of the individual systems.

19
 0.8

0.6

0.4

0.2

-0.2
Ψ1σ
-0.4 Ψ1σ*

-0.6
-3 -2 -1 0 1 2 3

Figure 2.1: Nuclear cusps. Wave functions corresponding to the two lowest eigenvalues of
H2 molecule, along the H–H bond (obtained from a FCI calculation in cc-pVDZ basis).

5. Within the non-relativistic theory, the exact stationary states are eigenfunctions of
the total and projected spin operators,

Ŝ 2 Ψ = S(S + 1)Ψ (2.4)

Ŝz Ψ = MΨ (2.5)

6. The molecular electronic Hamiltonian (1.2) is singular for ri = rj and thus the exact
wave function must possess a characteristic non-differentiable behavior for spatially
coinciding electrons, known as the electronic Coulomb cusp condition,
 
∂Ψ 1
lim = Ψ(rij = 0), (2.6)
rij →0 ∂rij ave 2
the description of which is a major obstacle in the accurate practical modelling of the
electronic wave function. There exists a condition similar to (2.6) also for electrons
coinciding point-like nuclei, known as the nuclear cusp condition.

7. It may shown that at large distances the electron density decays as

√
ρ(r) ∼ exp(−23/2 Ir), (2.7)

where I is the ionization potential of the molecule.

8. The exact wave function transforms in a characteristic manner under gauge trans-
formations of the potentials associated with electromagnetic fields, ensuring that all

20
 molecular properties described by the wave function are unaffected by the transfor-
mations.
As we saw in the introductionary survey, some of these – square integrability and Pauli
principle – are included straightforwardly, whereas size-extensivity and the cusp condition
are more difficult to impose but still desirable. Others are of interest only in special
situations.

2.2 The Coulomb cusp

2.2.1 Hamiltonian for a He-like atom
The non-relativistic Hamiltonian of the helium-like atom with the origin at the nucleus,
1 1 Z Z 1
Ĥ = − ∇21 − ∇22 − − +
2 2 |r1 | |r2 | |ri − r2 |
has singularities for r1 = 0, r2 = 0, and for r1 = r2 . At these points, the exact solution
to the Schrödinger equation must provide contributions that balance the singularities such
that the local energy
ĤΨ(r1 , r2 )
ǫ(r1 , r2) = (2.8)
Ψ(r1 , r2 )
remains constant and equal to the eigenvalue E. The only possible “source” for this
balancing is the kinetic energy. It is convenient to employ the symmetry of the helium
atom and express the kinetic energy operator in terms of three radial coordinates r1 , r2
and r12 , such that the Hamiltonian is written as
2    2 
1 X ∂2 2 ∂ 2Z ∂ 2 ∂ 1
Ĥ = − + + − + −
2 i=1 ∂ri2 ri ∂ri ri ∂r122
r12 ∂r12 r12
 
r1 r12 ∂ r2 r21 ∂ ∂
− · + · . (2.9)
r1 r12 ∂r1 r2 r21 ∂r2 ∂r12

2.2.2 Coulomb and nuclear cusp conditions

In Eq. (2.9, the terms that multiply 1/r12 at r12 = 0 must vanish in ĤΨ, which imposes a
condition 
∂Ψ  1
= Ψ(r12 = 0) (2.10)
∂r12 r12 =0 2
on the wave function. Similarly, the singularities at the nucleus are now seen to be balanced
by the kinetic energy terms proportional to 1/ri:

∂Ψ 
= −ZΨ(ri = 0). (2.11)
∂ri ri =0

21
Eq. (2.10) describes the situation when the electrons coincide in space and is referred to as
the Coulomb cusp condition; whereas Eq. (2.11) establishes the behavior of the ground-state
wave function in the vicinity of the nucleus and is known as the nuclear cusp condition.
Expanding the ground-state helium wave function around r2 = r1 and r12 = 0 we obtain
 
∂Ψ  ∂Ψ 
Ψ(r1 , r2 , 0) = Ψ(r1 , r1 , 0) + (r2 − r1 ) + r12 + ...,
∂r2 r2 =r1 ∂r12 r2 =r1

which gives, when the cusp condition (2.10) is applied


∂Ψ  1
Ψ(r1 , r2 , 0) = Ψ(r1 , r1 , 0) + (r2 − r1 ) + |r2 − r1 |Ψ(r1 , r1 , 0) + . . .
∂r2 r2 =r1 2


Therefore, the cusp condition leads to a wave function that is continuous but not smooth
(discontinuous first derivative) at r12 .
The nuclear cusp condition for the “first” electron when the wave function does not
vanish at r1 = 0 (such as the helium ground state) is satisfied if the wave function exhibits
an exponential dependence on r1 close to the nucleus:

Ψ(r1 , r2 , r12 ) = exp(−Zr1 )Ψ(0, r2, r12 ) ≈ (1 − Zr1 )Ψ(0, r2 , r12 ).

Molecular electronic wave functions are usually expanded in simple analytical functions
centered on the atomic nuclei (AOs), and close to a given nucleus, the behavior of the
wave function is dominated by the analytical form of the AOs. In particular, the Slater-
type orbitals (introduced later), are compatible with the nuclear cusp condition, while the
Gaussian-type orbitals are not.
It should be noted that the cusp conditions in Eqs. (2.10) and (2.11) are written for the
totally symmetric singlet ground state of the helium atom. The cusp conditions in a more
general situation (but still for a wave function that does not vanish at the singularities)
should be written as
 
∂Ψ 1
lim = Ψ(rij = 0) (2.12)
rij →0 ∂rij ave 2
 
∂Ψ
lim = −ZΨ(ri = 0), (2.13)
ri →0 ∂ri ave
where the averaging over all directions is implied.

2.2.3 Electron correlation

Let us at this stage review some nomenclature of electron correlation used in molecular
electronic-structure theory.
• A commonly-used concept, correlation energy, has a pragmatic definition

Ecorr = Eexact − ESD ,

22
 where ESD is a best single-determinant total energy and Eexact the “exact” energy in
the same basis. This definition is most usable when speaking of molecular ground
states and equilibrium geometries, while outside them it is untenable.
• Fermi correlation arises from the Pauli antisymmetry of the wave function and is
taken into account already at the single-determinant level.
• Static correlation, also known as near-degeneracy or nondynamical correlation, arises
from the near-degeneracy of electronic configurations.
• Dynamical correlation is associated with the instantaneous correlation among the
electrons arising from their mutual Coulombic repulsion. It is useful to distinguish
between
– long-range dynamical correlation and
– short-range dynamical correlation, which is related to the singularities in the
Hamiltonian and giving rise to the Coulomb cusp in the wave function.

2.3 Variation principle

According to the variation principle, the solution of the time-independent Schrödinger
equation Ĥ |Ψi = E |Ψi is equivalent to an optimization of the energy functional
D   E
 
Ψ̃ Ĥ  Ψ̃
E[Ψ̃] = D  E , (2.14)
Ψ̃ Ψ̃

where |Ψ̃i is some approximation to the eigenstate |Ψi. It provides a simple and powerful
procedure for generating approximate wave functions: for some proposed model for the
wave function, we express the electronic state |Ci in terms of a finite set of numerical
parameters; the stationary points of energy function
D   E
 
C Ĥ  C
E(C) = (2.15)
hC |Ci
are the approximate electronic states |Ci and the values E(C) at the stationary points
the approximate energies. Due to the variation principle, the expectation value of the
Hamiltonian is correct to second order in the error.

2.3.1 Linear ansatz

A simple realization of the variation method is to make a linear ansatz for the wave function,
m
X
|Ci = Ci |ii ,
i=1

23
i.e. the approximate state is expanded in m-dimensional set of Slater determinants. We
further assume here that the wave function is real. The energy function of this state is
given by Eq. (2.15). In order to locate and to characterize the stationary points, we shall
employ the first and second derivatives with respect to the variational parameters:
D   E
 
∂E(C) i Ĥ  C − E(C) hi |Ci
(1)
Ei (C) = =2
∂Ci hC |Ci
D   E
 
∂ 2 E(C) i Ĥ  j − E(C) hi |ji hj |Ci hi |Ci
(2) (1) (1)
Eij (C) = =2 − 2Ei (C) − 2Ej (C) ,
∂Ci ∂Cj hC |Ci hC |Ci hC |Ci

known as the electronic

D   gradient
E and electronic Hessian, respectively. The condition for
 
stationary points, i Ĥ  C = E(C) hi |Ci is in a matrix form equal to HC = E(C)SC,
D   E
 
where the Hamiltonian and overlap matrix are given by Hij = i Ĥ  j and Sij = hi |ji.
Assuming that S = 1, we have to solve a standard m-dimensional eigenvalue problem with
m orthonormal solutions CK = (C1K C2K · · · CmK )T , CTK CK = δKL , with the associated
real eigenvalues EK = E(CK ), E1P ≤ E2 ≤ · · · ≤ Em . The eigenvectors represent the
approximate wave functions |Ki = M i=1 CiK |ii with a corresponding approximate energy
EK . To characterize
D  the E stationary points, we note that the Hessian is at these points
K (2)  
Eij (CK ) = 2( i Ĥ  j − Ek hi |ji). We may also express the Hessian in the basis formed
of the eigenvectors:
D   E
K (2) (2)
EM N = 2( M Ĥ  N − EK hM |Ni) = 2(EM − EN )δM N ⇒ K EM M = 2(EM − EK ).
 

K (2)
EM M is thus singular and K th state has exactly K − 1 negative eigenvalues. Therefore,
in the space orthogonal to CK , the first solution is a minimum, the second a first-order
saddle point, the third a second-order saddle point, and so forth.

2.3.2 Hellmann–Feynman theorem

Many of the theorems for exact wave functions hold also for the approximate ones that
are variationally determined. One of the most important is the Hellmann–Feynman theo-
rem, which states that the first-order change in the energy due to a perturbation may be
calculated as the expectation value of the perturbation operator V̂ : Let |Ψα i be the wave
function associated with Ĥ + αV̂ and |Ψi unperturbed wave function at α = 0. Then
D   E
 

dE(α)  ∂ Ψ αĤ + α V̂  Ψα  
∂Ψ

    D   E
α    
=  = 2 Re  Ĥ − E(0) Ψ + Ψ V̂  Ψ
dα α=0 ∂α hΨα |Ψα i 
 ∂α α=0

α=0
D   E
 
= Ψ V̂  Ψ . (2.16)

24
 In fact, for the Hellmann–Feynman theorem to hold, we need to demand from the
approximate wave functions that they are optimized with respect to the changes induced by
the perturbation: |Ψi → |Ψi + α |∂Ψ/∂αi. Usually in molecular calculations we construct
our wave function from a finite set of analytical functions attached to the atomic nuclei.
For example, when we distort the molecular geometry, we change the basis set in terms
of which our electronic wave function is expanded, and a wave function optimized at one
particular geometry is not accurately represented in terms of a basis set associated with
another geometry. As a result, the unperturbed electronic is not variational with respect
to these changes and the conditions of the Hellmann–Feynman theorem is not satisfied.
However, this is an artificial difficulty arising from the coordinate representation of
quantum mechanics we presently operate in. The picture presented by the second quan-
tization, the necessary conditions are fulfilled and we are able to employ the theorem for
approximate wave functions in finite bases in exactly the same manner as for exact wave
functions.

2.3.3 The molecular electronic virial theorem

The exact molecular energy is variational with respect to an arbitrary change in the wave
function, particularly of interest are the uniform, normalization-preserving scaling of the
electron coordinates
Ψ(ri ) → Ψα (ri ) = α3N/2 Ψ(αri ).
Let us partition the electronic Hamiltonian in a bit different manner than earlier:
Ĥ(R) = T + V (R) (2.17)
1X 2
T̂ = − ∇i (2.18)
2 i
X 1 X ZI X ZI ZJ
V̂ (R) = − + , (2.19)
i>j
|r i − r j | iI
|r i − R I | I>J
|R I − R J |

and applying this decomposition we may find the stationary points of the energy function
are found at6
∂ D   E ∂ h 2 D   E D   Ei
   
0 = Ψα Ĥ(R) Ψα  = α Ψ T̂  Ψ + α Ψ V̂ (αR) Ψ 
∂α α=1 ∂α α=1
D   E D   E      
     dV (αR)   H−F dE(αR) 
⇒ 2 Ψ T̂  Ψ + Ψ V̂ (R) Ψ = − Ψ     Ψ = −  (2.20)
.
dα α=1  dα α=1
This result (2.20) is the quantum-mechanical virial theorem for a field-free non-relativistic
molecular Hamiltonian. The most important observation is that in the molecular equilib-
rium geometries (R = Re ), and in the case of atoms
D   E 1 D   E
  
Ψ T̂  Ψ = − Ψ V̂ (Re ) Ψ . (2.21)
2
6
Note that the unperturbed wave function corresponds here to α = 1 instead of α = 0.

25
Furthermore,
D   E D   E
   
Ψ T̂  Ψ |R=Re = −E(Re ), Ψ V̂ (Re ) Ψ = 2E(Re ). (2.22)

Also these expressions hold for any approximate wave function that is variational with
respect to a uniform scaling of the nuclear as well as electronic coordinates. Note that
according to the last result, no stationary points of positive energy may exist.
The scaling force may be easily related to classical Cartesian forces on the nuclei,
FI (R) = −dE(R)/dRI , by invoking the chain rule
  
dE(αR)  X d(αRI )  dE(αR I )  X
=   = − RI · FI (R)
dα α=1 I
dα 
α=1 d(αR I

α=1 I

and combining this with virial theorem, we may extract the Cartesian forces experienced
by the nuclei, D   E D   E
X    
RI · FI (R) = 2 Ψ T̂  Ψ + Ψ V̂ (R) Ψ . (2.23)
I

2.4 One- and N -electron expansions

Let us recall: the Slater determinants do not represent an accurate solution to the molecular
Schrödinger equation, but we may express the exact wave function as a superposition
of determinants. Furthermore, it is necessary to expand the molecular orbitals in the
determinants as a linear combination of atomic orbitals
X
φa = Cµa χµ (r). (2.24)
µ

As we will see later, atomic orbitals are further approximated with a set of simpler
functions that are also functions of the coordinates of a single electron. This three-step
procedure is often referred to as the “standard model of quantum chemistry”. To arrive
at the exact solution, i) orbitals in terms of which the determinants are constructed must
form a complete set in one-electron space, and ii) in the expansion of N-electron wave
function, we must include all determinants that can be generated from these orbitals.
All approximations in the solution of the Schrödinger equation should be unambigious
and precisely defined. Ideally, they should be improvable in a systematic fashion; and to
yield more and more elaborate solutions that approach the exact solution. Thus we speak
of different models, levels of theory and hierarchies of approximations.
Computational electronic-structure theory has a few standard models for the construc-
tion of approximate electronic wave functions. At the simplest level, the wave function is
represented by a single Slater determinant (the Hartree–Fock approximation); and at the
most complex level the wave function is represented as a variationally determined superpo-
sition of all determinants in the N-electron Fock space (full configuration-interaction, FCI).

26
 Figure 2.2: The two-way hierarchy of approximate molecular wave functions

Between these extremes, there are a vast number of intermediate models with variable cost
and accuracy.
It should be noted that none of the models is applicable to all systems. However, each
model is applicable in a broad range of molecular systems, providing solutions of known
quality and flexibility at a given computational cost.

27
2.5 Further reading
• D. P. Tew, W. Klopper, and T. Helgaker, Electron correlation: the many-body problem
at the heart of chemistry, J. Comput. Chem. 28, 1307 (2007)

2.6 Exercises for Chapter 2

1. Here we consider, how the error in the energy can be estimated from the norm of the
gradient.

(a) Show that the electronic energy, gradient and Hessian can be expanded around
a stationary point as
1
E(λ) = E + λT E(2) λ + O(kλk3 )
2
E(1) (λ) = E(2) λ + O(kλk2 )
E(2) (λ) = E(2) + O(kλk)

In the above, λ is the vector containing the wave-function parameters; λ = 0

represents the stationary point. E and E(2) are the energy and the Hessian at
λ = 0.
(b) Use these expansions to show that the errors in the wave function and in the
energy are related to the gradient as

λ = [E(2) (λ)]−1 E(1) (λ) + O(kE(1) (λ)k2 )

1 (1)
E(λ) − E = [E (λ)]T [E(2) (λ)]−1 E(1) (λ) + O(kE(1) (λ)k3 )
2
2. ⋆ Consider the uniform scaling of the electron coordinates,

Ψ(ri ) → Ψα (ri ) = α3N/2 Ψ(αri ).

Here α is a real scaling factor and N the number of electrons. We assume that Ψ(ri )
is normalized.

(a) Shot that the overlap is unaffected by the scaling, i.e.

hΨα |Ψα i = 1.

(b) Show that the expectation values of the kinetic and potential energies transform
as
D   E D   E
Ψα T̂  Ψα = α2 Ψ T̂  Ψ
   

hΨα |V (R)| Ψα i = α hΨ |V (αR)| Ψi .

28
3. ⋆ Here, we will consider the use of correlating functions to describe the Coulomb
cusp of many-electron systems. For a system with two electrons and K nuclei, with
the Hamiltonian

Ĥ = T̂ + V̂ + ĝ
1 1
T̂ = T̂1 + T̂2 = − ∇21 − ∇22
2 2
X ZI X ZI
V̂ = V̂1 + V̂2 = − −
I
|r1 − RI | I
|r2 − RI |
1
ĝ = ,
r12
write the wave function in the form Ψ = γΦ, where γ is a continuous correlating
function that depends only on r12 , and Φ is a smooth function of r1 and r2 . The
functions γ and Φ are, of course, finite.

(a) Show that the Schrödinger equation can be written in the form

ĝγΦ + [T̂ , γ]Φ = (E − V̂ )γΦ − γ T̂ Φ.

(b) Consider the former result as r12 → 0 and ĝ becomes singular. Assuming that
the electrons and nuclei do not coincide, show that the singularity in ĝ must be
compensated for by a singularity in [T̂ , γ] such that ĝγΦ + [T̂ , γ] remains finite.
(c) Show that
2 r12
[T̂ , r12 ] = − − · (∇1 − ∇2 ),
r12 r12
n −1
and that for n > 1, [T̂ , r12 ] contains no terms proportional to r12 .
2
(d) Consider the expansion of γ = 1 + βr12 + O(r12 ) around r12 = 0; determine the
parameter β such that ĝγ + [T̂ , γ] becomes nonsingular.
(e) For more than two electrons, we may again model the wave function by γΦ,
where γ depends on the interelectronic coordinates rij and where Φ is a smooth
function of the electronic coordinates ri . Show that the form
X
γ =1+β rij
i>j

eliminates the two-electron singularities in the many-electron Schrödinger equa-

tion if β is chosen as for the two-electron system.
(f) Determine, which of the following correlating functions remove the singularities
at rij = 0:
X
γ = 1+β rij
i>j

29
 Y
γ = (1 + βrij )
i>j
!
X
γ = exp β rij
i>j
!
X X
γ = 1+ αp exp βp rij2
p i>j

30
Chapter 3

Molecular integral evaluation

3.1 Atomic basis functions

Molecular orbitals can be constructed either numerically or algebraically, of which the for-
mer provides greater flexibility and accuracy, but is tractable only for atoms and diatomic
molecules; for polyatomic systems we are forced to expand the MOs φa (r) in a set of some
simple analytical one-electron functions.
Perhaps surprisingly, it turns out that the hydrogenic functions Eq. (1.8) are not ideal
for this purpose; as they do not constitute a complete set by themselves but must be
supplemented by the unbounded continuum states. Secondly, they become very diffuse
and a large number of them is needed for a proper description of both the core and valence
regions of a many-electron atom, especially when located in a molecule.
Therefore, in the molecular electronic-structure theory, χ’s are usually functions of
slightly more unphysical origin. Ideally, they should

1. allow orderly and systematic extension towards completeness with respect to one-
electron square-integrable functions

2. allow rapid convergence to any atomic or molecular electronic state, requiring only a
few terms for a reasonably accurate description of molecular electronic distributions

3. have an analytical form for easy and accurate manipulation especially for molecular
integrals

4. be orthogonal or at least their non-orthogonality should not pose problems related

to numerical instability.

Let us begin by reviewing the properties of the hydrogenic functions. As we noted

earlier, they are the bounded eigenfunctions of the Hamiltonian

1 Z
H = − ∇2 − .
2 r
31
and are written as
 3/2 s  l    
2Z (n − l − 1)! 2Zr 2l+1 2Zr Zr
χnlm (r, θ, ϕ) = Ln−l−1 exp −
n 2n(n + 1)! n n n
| {z }
Rnl (r)
s
2l + 1 (l − m)! m
× P (cos θ) exp(imϕ).
4π (l + m)! l
| {z }
Ylm (θ,ϕ)

The energy of the bounded hydrogenic state χnlm is given by

Z2
En = − . (3.1)
2n2
The degeneracy of the hydrogenic states of different angular momenta – that is, 2l + 1
degenerate states for every l – is peculiar to the spherical symmetry of the Coulomb
potential and is lifted in a many-atom system. The notable features of the radial functions
include
• the presence of the exponential

• that Laguerre polynomials introduce n − l − 1 radial nodes in the wave function.

The diffusiveness of χnlm for large n is seen from the expectation value of r̂
3n2 − l(l + 1)
hχnlm |r̂| χnlm i = (3.2)
2Z
In the following we consider basis functions that retain the product form of the hydrogenic
wave functions but have a more compact radial part.

3.1.1 The Laguerre functions

In order to retain the correct exponential behavior but avoid the problems with continuum
states we introduce the following class of functions:

χLF LF
nlm (r, θ, ϕ) = Rnl (r)Ylm (θ, ϕ) (3.3)
s
LF (n − l − 1)!
Rnl = (2ζ)3/2 (2ζr)l L2l+2
n−l−1 (2ζr) exp(−ζr), (3.4)
(n + l + 1)!

referred to as the Laguerre functions. For fixed l and ζ the radial functions (3.4) with n > l
constitute a complete orthonormal set of functions, for which

LF  2n + 1
χnlm |r̂| χLF
nlm = , (3.5)
2ζ

32
 1s 2s 2p 3s

0 5 10 0 5 10 0 5 10 0 5 10

Figure 3.1: Lagurerre functions.

being thus considerably more compact than the hydrogenic functions for large n and inde-
pendent of l. The Laguerre functions exhibit the same nodal structure as the hydrogenic
ones but their radial distributions R2 r 2 are quite different (ζ = 1/n, Z = 1).
One should note that the orthogonality of the Laguerre functions is valid only with
fixed exponents. It turns out, however, that the expansion of the orbital of a given l-value
of an atomic system requires a large number of fixed-exponent Laguerre functions with
different n. For non-orthogonal, variable-exponent functions Laguerre functions are not
an optimal choice, but we need to introduce functions specifically tailored to reproduce as
closely as possible the different orbitals of each atom.

3.1.2 Slater-type orbitals

LF
In simplifying the polynomial structure of the Laguerre functions, we note that in Rnl r
occurs to power l and to powers n − l − 1 and lower. Thus a nodeless one-electron function
that resembles χLF
nlm and the hydrogenic functions closely for large r is obtained by retaining
in the Laguerre function only the term of the highest power of r, yielding the Slater-type
orbitals (STO)

χSTO STO
nlm (r, θ, ϕ) = Rnl (r)Y (θ, ϕ) (3.6)
(2ζ)3/2
RnSTO (r) = p (2ζr)n−1 exp(−ζr) (3.7)
Γ(2n + 1)
For the STOs we use the same 1s, 2p,... notation as for the hydrogenic functions, keeping
in mind that STOs are nodeless and now n only refers to the monomial factor r n−1 . The
expectation value of r̂ and the maximum in the radial distribution curve are given by

STO  2n + 1
χnlm |r̂| χSTO
nlm = (3.8)
2ζ
STO n
rmax = (3.9)
ζ
Also STOs would constitute a complete set of functions for a fixed exponent ζ, and we
could choose to work with a single exponent and include in our basis orbitals (1s, 2s, . . .),
(2p, 3p, . . .), (3d, 4d, . . .), all with the same exponent. Alternatively, we may describe the
radial space by functions with variable exponents. For each l, we employ only the functions

33
of the lowest n, yielding a basis of the type ((1s(ζ1s ), 1s(ζ2s ), . . .), (2p(ζ1p , 2p(ζ2p ), . . .), . . .)
These variable-exponent STOs are given by

χζnl lm (r, θ, ϕ) = RζSTO

nl l
(r)Ylm (θ, ϕ) (3.10)
3/2
(2ζnl )
RζSTO
nl l
= p (2ζnl r)l exp(−ζnl r). (3.11)
(2l + 2)!

In practice, a combined approach is employed. For example, for the carbon atom, we would
introduce sets of 1s, 2s and 2p functions, all with variable exponents that are chosen to
ensure an accurate representation of the wave function.

3.1.3 Gaussian-type orbitals

The STO basis sets are very successful for atoms and diatomic molecules, but the evaluation
of many-center two-electron integrals becomes very complicated in terms of them. The
basis sets that have established to be the standard practice in molecular electronic-structure
theory are less connected to the hydrogenic functions or actual charge distributions in
molecules. They are called Gaussian-type orbitals (GTO) as they employ the Gaussian
distributions exp(−αr 2 ) instead of the exp(−ζr) as in STOs. A large number of GTOs are
needed to describe an STO (or an AO) properly, but this is more than compensated by a
fast evaluation of the molecular integrals.
The spherical-harmonic GTOs are

χGTO GTO
nlm (r, θ, ϕ) = Rnl (r)Ylm (θ, ϕ) (3.12)
s
GTO 2(2α) 3/4
22n−l−2 √
Rnl (r) = 1/4
( 2αr)2n−l−2 exp(−αr 2 ). (3.13)
π (4n − 2l − 3)!!

They form a complete set of nonorthogonal basis functions. In above, the double factorial
function is 

 1 n=0
 n(n − 2)(n − 4) · · · 2 even n > 0
n!! = (3.14)

 n(n − 2)(n − 4) · · · 1 odd n > 0
 1
odd n < 0
(n+2)(n+4)···1

The spatial extent of the GTOs are

r

GTO  2n − l − 2
χnlm |r̂| χGTO
nlm ≈ (3.15)
r 2α
GTO 2n − l − 1
rmax = (3.16)
2α
When we compare these results with the case of STOs, we note that much larger number
of GTOs are needed for a flexible description of outer regions of the electron distribution.

34
 2.5 5.0 5.0 7.5
1s STO (ζ=1) 1s GTO (α=1) 1s 1s
2s STO (ζ=1) 2s GTO (α=1) STO (ζ=1) GTO (α=3.42)
2p GTO (α=1) GTO (α=1) GTO (α=0.62)
GTO (α=0.16)
5.0

2.5 2.5

2.5

0.0 0.0 0.0 0.0

0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5

Figure 3.2: Slater and Gaussian-type orbitals compared.

In practice, the use of GTOs with variable exponents differs from the procedure for
STOs. For GTOs, we describe the radial space exclusively by means of variable exponents,
using for this purpose only spherical-harmonic GTOs with n = l + 1, only powers of r
introduced are thus associated with l of the spherical harmonics. The following set of
spherical-harmonic GTOs is used:

χGTO GTO
αnl lm (r, θ, ϕ) = Rαnl l (r)Ylm (θ, ϕ) (3.17)
s
2(2αnl ) 3/4
22n−l−2 √
RαGTO
nl l
(r) = ( 2αr)2n−l−2 exp(−αr 2 ). (3.18)
π 1/4 (4n − 2l − 3)!!

Let us finally consider the Gaussian product rule, that is the one of the main reasons
2
why GTOs are so widely used. The product of two s GTOs χA = exp(−αrA ) and χB =
2
exp(−βrB ) centered on rA and rB is a third Gaussian
 
αβ 2  
χA χB = χC = exp − rAB exp −(α + β)rC2 (3.19)
α+β

centered on rC given by
αrA + βrB
rC = .
α+β

3.2 Gaussian basis sets

Of the three general features determining the accuracy of quantum chemical methods–the
choice of the Hamiltonian, electron correlation treatment or truncation of the many-particle
space, and the description of the one-particle space, i.e., the basis set used–the choice of the
basis set is the most crucial one, as insufficient basis sets yield erroneous results regardless
of the level of theory, whereas with a proper basis set, at least qualitatively correct results

35
can be obtained for many problems already at the uncorrelated level of theory and using
the non-relativistic Hamiltonian.
The majority of quantum chemical applications employ contracted GTO (CGTO) sets,
i.e. linear combinations of GTO functions with coefficients optimized regarding some
criterion (usually SCF energy) that significantly increase efficiency. There are two widely
used contraction schemes: segmented contraction, where each GTO contributes only to a
single CGTO, and general contraction, where such a restriction is not applied. Examples of
the segmented scheme include the Pople-style basis set families (3-21G, 6-31G,...) and the
Karlsruhe basis sets. The correlation-consistent (cc) basis set families by Dunning and co-
workers are the most widely used generally contracted basis sets. The basic idea behind the
cc sets is that functions that contribute approximately the same amount to the correlation
energy are added to the basis in groups. The cc basis sets provide smooth, monotonic
convergence for the electronic energy as well as for many molecular properties, especially
those originating in the valence region. The polarized valence (cc-pVXZ, X = D, T, Q, 5,
6 corresponding to the number of CGTOs used to represent an occupied atomic orbital)
and core-valence (cc-pCVXZ) sets can be augmented with diffuse functions (aug-cc-pVXZ
and aug-cc-pCVXZ; d-aug-cc-pVXZ and t-aug-cc-pVXZ for doubly and triply augmented
sets, respectively). While these features are favorable and the sets are widely used, the cc-
basis sets become very large at large X, and it is not straightforward to extend the family
beyond the published X values or to new elements. The atomic natural orbital (ANO) basis
sets provide another general contraction approach. The contraction coefficients therein are
atomic ANO coefficients that are obtained by optimizing atomic energies.
The use of increasingly large basis sets that produce results converging to some partic-
ular value is usually regarded as a solution to the problem of basis set incompleteness. In
calculations of molecular properties that originate, e.g., in the region close to the nuclei
(examples include indirect spin-spin couplings and hyperfine couplings), approaching the
basis-set limit using the cc or comparable energy-optimized paradigms may lead to exces-
sively large basis sets prohibiting calculations of large molecules. An alternative and often
used approach is to uncontract the basis set and to supplement it by additional steep basis
functions in the l-shells relevant for the property under examination.
The performance of the different basis sets for a certain molecular property can be
qualitatively understood within a concept, which can be measured by the completeness
profiles
X
Y(α) = hg(α)|χmi2 . (3.20)
m

Here {χ} denotes a set of orthonormalized, contracted or primitive, basis functions, and
g(α) is a primitive “test” GTO with the exponent α, used to probe the completeness of
{χ}. The completeness profile becomes unity for all α, for all l-values in a CBS. Y(α)
is typically plotted in the logarithmic scale, against x = log α. In a certain exponent
interval [αmin , αmax ] this quantity is intuitively connected to the possibility – from the
point of view of one-particle space – of describing all details of the wave function in the
corresponding distance range from the atomic nuclei. Put simply, atomic properties that

36
 1.0
s s
0.75 (a) p (b) p
Y(α)

0.5 d d
f
0.25

1.0
s s
0.75 (c) p (d) p
d
Y(α)

0.5 d f
f g
0.25 g h

1.0
s s
0.75 (e) p (f) p
d d
Y(α)

0.5 f f
g
0.25 h
i
1.0
s s
0.75 (g) p (h) p
d d
Y(α)

0.5 f f
0.25

4 3 2 1 0 -1 -2 -3 -4 4 3 2 1 0 -1 -2 -3 -4
Log (α) Log (α)

Figure 3.3: Completeness profiles of (a) cc-pVDZ (b) cc-pVTZ (c) cc-pVQZ (d) cc-pV5Z
(e) cc-pV6Z (f) cc-pCVTZ (g) aug-cc-pVTZ and (h) aug-cc-pCVTZ basis sets of fluorine.

√ √
obtain relevant contributions roughly within [1/ αmax , 1/ αmin ] from the nucleus can be
reproduced by a basis that has Y(α) = 1 in this interval. This way of thinking can be
generalized to molecular properties that may be dominated by phenomena occurring close
to the expansion centers of the basis functions, i.e., atomic nuclei (region described by
high-exponent basis functions) and/or in the valence region, further away from the nuclei.

3.3 Integrals over Gaussian basis sets

In the following, we will evaluate the one- and two-electron integrals
 E Z

Oµν = hχµ Ô χν = drχµ (r)Ô(r)χν (r) (3.21)

37
  E

Oµνλσ = hχµ (1)χν (1) Ô χλ (2)χσ (2)
Z
= dr1 dr2 χµ (r1 )χν (r1 )Ô(r)χλ (r2 )χσ (r2 ) (3.22)

such that the AOs (χ) are taken as fixed linear combinations of real-valued primitive
Cartesian GTOs
Gijk (r, a, R) = xiR yRj zRk exp(−arR
2
), (3.23)
where the ”Cartesian quantum numbers” i, j, k are greater than zero and l = i + j + k for
a given total angular momentum quantum number.
The integrals over primitive Cartesian GTOs may subsequently be transformed to in-
tegrals over contracted and/or spherical-harmonic GTOs by linear combinations.

3.3.1 Gaussian overlap distributions

An important property of the primitive Cartesian GTOs is that they can be factorized in
the three Cartesian directions,

Gijk (r, a, R) = Gi (x, a, Rx )Gj (y, a, Ry )Gz (z, a, Rz ), (3.24)

where for example Gi (x, a, Rx ) = xiR exp(−ax2R ). We will need also a concept of Gaussian
overlap distribution
Ωab (r) = Gijk (r, a, RA )Glmn (r, b, RB ), (3.25)
that may also be factorized in the same way

Ωab (r) = Ωxij (x, a, b, RA,x , RB,x )Ωykl (y, a, b, RA,y , RB,y )Ωzmn (z, a, b, RA,z , RB,z ), (3.26)

where the x component (for example) is given by

Ωxij (x, a, b, RA,x , RB,x ) = Gi (x, a, RA,x )Gj (x, b, RB,x ). (3.27)

3.3.2 Simple one-electron integrals

We are now ready to consider simple one-electron integrals, by which we mean molecular
integrals that do not involve the singular Coulombic 1/r interaction. All such integrals can
be factorized in the three directions as

Sab = Sij Skl Smn . (3.28)

According to the Gaussian product rule (3.19), any Cartesian component of the overlap
distribution can be written as a single Gaussian at the centre of charge Px , and we may
thus evaluate the integral
Z ∞ Z ∞ r
x 2 2 π 2
Ω00 dx = exp(−µXAB ) exp(−pxp )dx = exp(−µXAB ), (3.29)
−∞ −∞ p

38
where µ denotes the reduced exponent ab/(a + b), p = a + b, and

aRA,x + bRB,x
Px = (3.30)
p
XAB = RA,x − RB,x . (3.31)

This result provides a basis for a set of recurrence relations by which we may evaluate
simple integrals – overlap or more complicated ones – over GTOs of arbitrary quantum
numbers. This procedure is known as Obara–Saika scheme. There exist other methods
for molecular integral evaluation but we will not discuss them here. The relations are
obtained by considering the behaviour of the integral under coordinate transformation and
their detailed derivation is omitted here but recommended further reading.
The simplest case is the overlap integral

Sab = hGa |Gb i , (3.32)

for an Cartesian component of which the Obara–Saika relations are written as

1
Si+1,j = XP A Sij + (iSi−1,j + jSi,j−1 ) (3.33)
2p
1
Si,j+1 = XP B Sij + (iSi−1,j + jSi,j−1). (3.34)
2p
For the linear and angular momentum as well as for kinetic energy integrals we need
the integrals over differential operators
  e f g  
ef g
∂ ∂ ∂ 
Dab = Ga  e f g  Gb . (3.35)
∂x ∂y ∂z

In one direction the relations are

e 1
Di+1,j = XP A Dije + e
(iDi−1,j e
+ jDi,j−1 − 2beDije−1 ) (3.36)
2p
e 1
Di+1,j e
= XP B Dije + (iDi−1,j e
+ jDi,j−1 + 2aeDije−1 ), (3.37)
2p
with the e = 0 case providing the starting point and being equivalent with the overlap
integral.

3.3.3 The Boys function

We will now introduce a special function that has a key role in the evaluation of one- and
two-electron Coulomb integrals, the Boys function of order n, defined as
Z 1
Fn (x) = exp(−xt2 )t2n dt. (3.38)
0

39
 1

0.8

0.6

0.4

0.2

2 4 6 8 10

Figure 3.4: The three lowest-order Boys functions. The zeroth-order function in cyan,
first-order in blue and second-order in dark blue.

The Boys functions of orders 0, 1 and 2 are illustrated in Figure 3.4. It is a strictly positive
and decreasing function. Its values at x = 0 have a closed-form expression
Z 1
1
Fn (0) = t2n dt = . (3.39)
0 2n + 1
Different methods for evaluating the Boys function has been introduced. One way is
through recursion; by integrating the function by parts, we see that the different orders
are related by
(2n + 1)Fn (x) − exp(−x)
Fn+1 (x) = (3.40)
2x
or
2xFn+1 (x) − exp(−x)
Fn (x) = . (3.41)
2n + 1
We would therefore need to calculate the Boys function for the highest or the lowest order
needed, and the others are obtained through the downward or upward recursion, of which
the former is numerically more robust.

3.3.4 Obara–Saika scheme for one-electron Coulomb integrals

The Cartesian one-electron Coulomb integrals,
Z

   Ωab (r)
Vab = Ga rC−1  Gb = dr, (3.42)
rC
needed for the one-electron part of the molecular Hamiltonian, may also be evaluated by
using the Obara–Saika recursion relations.

40
 We again factorize the distribution according to (3.26), and are about to obtain the
integrals through the recursion relations (c.f., below) for auxillary integrals ΘN
ijklmn :

1
ΘN N
i+1 jklmn = XP A Θijklmn + (iΘN N
i−1 jklmn + jΘi j−1 klmn )
2p
1
−XP C ΘN +1
ijklmn − (iΘN +1 N +1
i−1 jklmn + jΘi j−1 klmn ) (3.43)
2p
1
ΘN
i j+1 klmn = XP B Θ N
ijklmn + (iΘN N
i−1 jklmn + jΘi j−1 klmn )
2p
1
−XP C ΘN +1
ijklmn − (iΘN +1 N +1
i−1 jklmn + jΘi j−1 klmn ) (3.44)
2p
with the special cases (that also serve the starting points of the recursion)

Θ0ijklmn = Vijklmn = Vab (3.45)

2π xyz
ΘN000000 = Kab FN (pRP2 C ), (3.46)
p
where FN is the Boys function, and we have denoted
x 2
Kab = exp(−µXAB ). (3.47)

3.3.5 Obara–Saika scheme for two-electron Coulomb integrals

The evaluation of electron-electron repulsion integrals is a highly non-trivial task and due
to their bottleneck status, under intense study since the early days of quantum chemistry.
We will discuss here only one of the numerous schemes that is in accordance with the
earlier discussion, namely the Obara–Saika scheme for Cartesian two-electron integrals,
Z Z

 −1   Ωab (r1 )Ωcd (r2 )
gabcd = Ga (r1 )Gb (r1 ) r12 Gc (r2 )Gd (r2 ) =
  dr1 dr2 . (3.48)
r12
The relations feature similarly to the one-electron corresponds a set of auxillary integrals
Θ. Again, two special cases of them are required:
2π 5/2 xyz xyz
ΘN
0000 0000 0000 = √ Kab Kcd FN (αRP2 Q ) (3.49)
pq p + q
Θ0ix jx kx lx iy jy ky ly iz jz kz lz = gix jx kx lx iy jy ky ly iz jz kz lz = gabcd . (3.50)
xyz x y z
In the above, we have denoted Kab = Kab Kab Kab [c.f. Eq. (3.47)] and α is the reduced
exponent pq/(p + q). FN is again the Boys function. In the following, the indices i, j,
k and l are the Cartesian quantum numbers for orbitals a, b, c and d in the x-direction.
Starting from these, first a set of integrals with j = k = l = 0 is generated by
 
N α N +1 i N α N +1
Θi+1 000 = XPA Θi000 − XPQ Θi000 + Θi−1 000 − Θi−1 000 . (3.51)
p 2p p

41
Then, the second electron is treated using the integrals generated in (3.51):

bXAB + dXCD N i k N p
Θi0 k+1 0 = − Θi0k0 − ΘNi−1 0k0 + Θi0 k−1 0 − ΘN . (3.52)
q 2q 2q q i+1 0k0
In the final step, the Cartesian powers are ”transferred” between the orbitals of the same
electron:

Θi j+1 kl = Θi+1 jkl + XAB ΘN

ijkl (3.53)
Θijkl+1 = Θij k+1 l + XCD ΘN
ijkl . (3.54)

In this way, we may construct the full set of Cartesian two-electron Coulomb integrals.

42
3.4 Further reading
• T. Helgaker and P. Taylor, Gaussian basis sets and molecular integrals, in Modern
Electronic Structure Theory, Part II, D. R. Yarkony (ed.), (World Scientific 1995),
pp. 725–856

• R. Lindh, Integrals of electron repulsion, in P. v. R. Schleyer et al. (eds.), Encyclo-

pedia of Computational Chemistry Vol. 2, (Wiley 1998) p. 1337.

• M. C. Strain, G. E. Scuseria and M. J. Frisch, Achieving linear scaling for the elec-
tronic quantum Coulomb problem, Science 271, 51 (1996).

• S. Obara and M. Saika, Efficient recursive computation of molecular integrals over

Cartesian Gaussian functions, Journal of Chemical Physics 84, 3963 (1986); 89, 1540
(1988).

3.5 Exercises for Chapter 3

1. Verify the expressions for radial extents
Z ∞
hri = Rnl (r)rRnl (r)r 2 dr
0

of (a) the hydrogenic wave functions, (b) the Laguerre functions, (c) the radial STOs.

2. ⋆ In this exercise, we study the radial form of the GTOs, the most frequently used
basis functions in molecular electronic-structure theory:
GTO GTO
√
Rαl (r) = Nαl ( 2αr)l exp(−αr 2 ).
GTO
The related functions Rnl are obtained by replacing l by 2n − l − 2.

(a) Using the integral

Z ∞ r
2n 2 (2n − 1)!! π
x exp(−αx )dx = n+1 2n+1
, n ≥ 0, α > 0
0 2 α
GTO
verify that the normalization constant Nαl is given by
s
3/4
GTO 2(2α) 2l
Nαl = .
π 3/4 (2l + 1)!!

(b) Determine the radial overlap between two GTOs:

Z ∞
Sα1 α2 l = RαGTO
1l
(r)RαGTO
2l
(r)r 2dr.
0

43
 (c) By setting α2 = βα1 verify that the overlap Sα1 βα1 l is independent of α1 .
(d) Using the integral
Z ∞
n!
x2n+1 exp(−αx2 )dx = n ≥ 0, α > 0
0 2αn+1
show that the expectation value of r is
Z ∞
r
GTO GTO 2 2l (l + 1)!
hri = Rαl (r)rRαl (r)r 2dr = .
0 πα (2l + 1)!!
√
(e) Use the Stirling approximation for factorials k! ∼ k k+1/2 2πe−k (k ≫ 1) to
show that hri is for large l r
l
hri ∼ .
2α
(f) Verify that the radial distribution function 4πr 2 [Rαl
GTO
(r)]2 has a maximum at
r
GTO l+1
rmax = .
2α

(g) Consider two GTOs of different l and exponents. Determine the relation between
α1 and α2 so that the radial distribution of these GTOs have the same maximum.

3. Using the Obara–Saika scheme, evaluate the self-overlap integrals hpx |px i and hdx2 |dx2 i.

4. Construct the Obara–Saika recursion relations for kinetic-energy integrals

1
 2  
Tab = − Ga ∇ Gb .
2

5. ⋆ Derive the Obara–Saika relations for overlap and differential operator integrals,
Eqs. (3.33-3.34) and (3.36-3.37).

44
Chapter 4

Second quantization

4.1 Fock space

We now introduce an abstract linear vector space – the Fock space – where each Slater
determinant (1.20) of M orthonormal spin-orbitals is represented by an occupation number
vector (ON) k,

1 ψP occupied
|ki = |k1 , k2 , . . . , kM i , kP = (4.1)
0 ψP unoccupied

For an orthonormal set of spin-orbitals, we define the inner product between two ON vectors
as
YM
hm |ki = δmP kP . (4.2)
P =1

This definition is consistent with the overlap between two Slater determinants, but has a
well-defined but zero overlap between states with different electron numbers is a special
feature of the Fock-space formulation. It allows for a unified description of systems with
variable numbers of electrons.

• In a given spin-orbital basis, there is a one-to-one mapping between the Slater de-
terminants with spin-orbitals in a canonical order and the ON vectors in the Fock
space.

• However, ON vectors are not Slater determinants: ON vectors have no spatial struc-
ture but are just basis vectors in an abstract vector space.

• The Fock space can be manipulated as an ordinary inner-product vector space.

The ON vectors constitute an orthonormal basis in the 2M dimensional Fock space

F (M), that can be decomposed as a direct sum of subspaces,

F (M) = F (M, 0) ⊕ F (M, 1) ⊕ · · · ⊕ F (M, N), (4.3)

45
where F (M, N) contains all ON vectors obtained by distributing N electrons among the
M spin-orbitals, in other words, all ON vectors for which the sum of occupation number
is N. The subspace F (M, 0) is the vacuum state,

|vaci = |01 , 02 , . . . , 0M i . (4.4)

Approximations to an exact N-electron wave function

 are
 expressed in terms of vectors in
M
the Fock subspace F (M, N) of dimension equal to .
N

4.1.1 Creation and annihilation operators

In second quantization, all operators and states can be constructed from a set of elementary
creation and annihilation operators. The M creation operators are defined by

a†P |k1 , k2 , . . . , 0P , . . . , kM i = ΓkP |k1 , k2 , . . . , 1P , . . . , kM i (4.5)

a†P |k1 , k2 , . . . , 1P , . . . , kM i = 0 (4.6)

where
P
Y −1
ΓkP = (−1)kQ . (4.7)
Q=1

The definition can be also combined a single equation:

a†P |k1 , k2 , . . . , 0P , . . . , kM i = δkP 0 ΓkP |k1 , k2 , . . . , 1P , . . . , kM i (4.8)

Operating twice with a†P on an ON vector gives

a†P a†P |k1 , k2 , . . . , 0P , . . . , kM i = a†P δkP 0 ΓkP |k1 , k2 , . . . , 1P , . . . , kM i = 0,

therefore
a†P a†P = 0. (4.9)
For Q > P ,

a†P a†Q |. . . , kP , . . . , kQ , . . .i = a†P δkQ 0 ΓkQ |. . . , kP , . . . , 1Q , . . .i

= δkP 0 δkQ 0 ΓkP ΓkQ |. . . , 1P , . . . , 1Q , . . .i .

Reversing the order of the operators,

a†Q a†P |. . . , kP , . . . , kQ , . . .i = a†Q δkP 0 ΓkP |. . . , 1P , . . . , kQ , . . .i

= δkP 0 δkQ 0 ΓkP (−ΓkQ ) |. . . , 1P , . . . , 1Q , . . .i .

and combining the results, we obtain

a†P a†Q + a†Q a†P |ki = 0.

46
Substitution of dummy indices shows that this holds for Q < P as well, and is true also
for Q = P . Since |ki is an arbitrary ON vector, we conclude the anticommutation relation
a†P a†Q + a†Q a†P = [a†P , a†Q ]+ = 0. (4.10)
The properties of the adjoint or conjugate operators aP can be reasoned from those of
the creation operators. Thus, the adjoint operators satisfy the anticommutation relation
[aP , aQ ]+ = 0. (4.11)
Let us invoke the resolution of the identity:
X
aP |ki = |mi hm |aP | ki ,
m

where, using Eq. (4.8),


∗ δmP 0 Γm
P if kQ = mQ + δQP
hm |aP | ki = hk |aP | mi = .
0 otherwise
From the definition of Γ and from kQ = mQ + δQP we see that Γm K
P = ΓP . We may therefore
write the equation as

δkP 1 ΓkP if mQ = kQ − δQP
hm |aP | ki = .
0 otherwise
Hence, only one term in aP |ki survives:
aP |ki = δkP 1 ΓkP |k1 , . . . , 0P , . . . , kM i . (4.12)
The operator aP is called the annihilation operator.
Combining Eqs. (4.8) and (4.12), we get
a†P aP |ki = δkP 1 |ki (4.13)
aP a†P |ki = δkP 0 |ki (4.14)
that leads to
a†P aP + aP a†P = 1. (4.15)
For P > Q, we have
a†P aQ |ki = −δkP 0 δkQ 1 ΓkP ΓkQ |k1 , . . . , 0Q , . . . , 1P , . . . , kM i
aQ a†P |ki = δkP 0 δkQ 1 ΓkP ΓkQ |k1 , . . . , 0Q , . . . , 1P , . . . , kM i ,
thus we have the operator identity
a†P aQ + aQ a†P = 0 P > Q. (4.16)
Hence,
a†P aQ + aQ a†P = [a†P , aQ ]+ = δP Q . (4.17)

47
4.1.2 Number-conserving operators
The occupation-number operator that counts the number of electrons in spin-orbital P is
introduced as
N̂P = a†P aP , N̂P |ki = kP |ki (4.18)
The occupation-number operators are Hermitian, N̂ † = N̂ , as well as idempotent, N̂P2 =
N̂P . Using the basic anticommutation relations, we obtain
[N̂P , a†Q ] = δP Q a†Q (4.19)
h i
N̂P , aQ = −δP Q aQ (4.20)

So for an arbitrary string X̂ consisting of creation and annihilation operators, e.g. X̂ =

a†P aQ aP a†R aS , the commutators with N̂ become
[N̂P , X̂] = NPX X̂, (4.21)
where NPX is the number of times a†P occurs in X̂ minus the number of times aP occurs.
All occupation-number operators added together in the Fock space gives the Hermitian
operator
M
X
N̂ = a†P aP , (4.22)
P =1

which returns the number of electrons in an ON vector, N̂ |ki = N |ki, and is therefore
known as the particle-number operator. We find that the number operator commutes with
any string T̂ that contains an equal number of creation and annihilation operators. Such
strings are called number-conserving. In general, the application of the string X̂ to a
Fock-space vector increases the number of electrons by N X .
The simplest number-conserving operators are the elementary excitation operators
ÊQP = a†P aQ , (4.23)
which give as applied to an ON vector
a†P aQ |ki = a†P δkQ 1 ΓkQ |k1 , . . . , 0Q , . . . , kP , . . . , kM i
= δkP 0 δkQ 1 ΓkQ ΓkP εP Q |k1 , . . . , 0Q , . . . , 1P , . . . , kM i
 
k k
 kP → 1
= εP Q ΓQ ΓP (1 − kP )kQ   (4.24)
kQ → 0
where 
1 P ≤Q
εP Q = .
−1 P > Q
The case P < Q differs from this result only in the interpretation of εP Q , whereas the case
P = Q is covered by Eq. (4.18). Therefore we may write
 
† k k
 kP → 1
aP aQ |ki = εP QΓP ΓQ (1 − kP + δP Q )kQ 
 . (4.25)
kQ → δP Q

48
 First quantization Second quantization
• One-electron operator: • One-electron operator:
X
N
X fˆ = fP Q a†P aQ
fˆc = f (xi ) PQ
i=1

• Two-electron operator: • Two-electron operator:

N X
N 1 X
c
ĝ =
X
g(xi, xj ) ĝ = gP QRS a†P a†R aS aQ
2 P QRS
i=1 j=i+1

• Operators are independent of the spin- • Operators depend on the spin-orbital ba-
orbital basis sis
• Operators depend on the number of elec- • Operators are independent of electrons
trons
• Operators are exact • Projected operators

Table 4.1: Comparison between first and second quantization representations.

The application of a single such operator gives a single excitation, two a double excitation
and so forth.

4.2 The representation of one- and two-electron op-

erators
Expectation values of Hermitian operators correspond to physical observables and should
thus be independent of the representation of operators and states. Therefore, we require
the matrix element of a second-quantization operator between two ON vectors to be equal
to its first-quantization counterpart.
To determine the numerical parameters fP Q in the second-quantization representation
of one-electron operators, we evaluate the matrix elements of fˆ between two ON vectors
and compare with the usual Slater rules for matrix elements between determinants. We
are lead to the identification
D   E
fP Q = ψP fˆc  ψQ
 
(4.26)
gP QRS = hψP ψR |ĝ c | ψQ ψS i . (4.27)

The parameters gP QRS possess the particle-interchange symmetry,

gP QRS = gRSP Q . (4.28)

49
In the case of real spin-orbitals, the integrals exhibit the following permutational symme-
tries:
fP Q = fQP (4.29)
gP QRS = gQP RS = gP QSR = gQP SR, (4.30)
of which the latter may be further combined with Eq. (4.28) to yield an eight-fold symmetry.
Using the introduced second-quantization representation, we may now construct the
(non-relativistic, spin-free) molecular electronic Hamiltonian in the Born–Oppenheimer
approximation,
X 1
Ĥ = hP Q a†P aQ + gP QRS a†P a†R aS aQ + hnuc , (4.31)
PQ
2
where
Z !
1 X ZI
hP Q = ψP∗ (x) − ∇2 − ψQ (x)dx (4.32)
2 I
riI
Z
ψP∗ (x)ψR∗ (x)ψQ (x)ψS (x)
gP QRS = dx1 dx2 (4.33)
r12
1 X ZI ZJ
hnuc = . (4.34)
2 I6=J RIJ

The form of the second-quantization Hamiltonian can be interpreted in the following way:
Applied to an electronic state, the Hamiltonian produces a linear combination of the orig-
inal state with states generated by single and double electron excitations from this state.
With each such excitation, there is an associated amplitude hP Q or gP QRS , which represents
the probability of this event happening.
Let Âc , B̂ c and Ĉ c be one-electron operators in first quantization and Â, B̂ and Ĉ be
the corresponding second-quantization representations.
• The first-quantization operator aÂc + bB̂ c , where a and b are numbers, is represented
by a + bB̂.
• The standard relations
Â(B̂ Ĉ) = (ÂB̂)Ĉ (4.35)
(ÂB̂)† = B̂ † † (4.36)
are valid.
• For a complete one-electron basis
Ĉ c = Âc B̂ c ⇒ Ĉ = ÂB̂
but for finite basis sets this expression does not hold. The second-quantization op-
erators are projections of the exact operators onto a basis of spin-orbitals. For an
incomplete basis, the second-quantization representation of an operator product de-
pends on when the projection is made.

50
 Second quantization treats operators and wave functions in a unified way – they are all
expressed in terms of the elementary creation and annihilation operators: any ON vector
may be written as "M #
Y †
|ki = X̂k |vaci = (aP )kP |vaci (4.37)
P =1

and therefore the matrix elements can be viewed as the vacuum expectation value of an
operator D   E D   E
k Ô  m = vac X̂k† ÔX̂m  vac ,
   
(4.38)

and expectation values become linear combinations of vacuum expectation values.

4.3 Commutators and anticommutators

In the manipulation of operators and matrix elements in second quantization, the commu-
tator [Â, B̂] = ÂB̂ − B̂ Â and the anticommutator [Â, B̂]+ = ÂB̂ + B̂ Â of two operators are
often encountered. Referring to the basic relations of the elementary creation and annihi-
lation operators, it is usually possible to simplify the commutators and anticommutators
between strings of elementary operators considerably.
We need to introduce the concepts of operator rank and rank reduction.

• The (particle) rank of a string of creation and annihilation operators is simply the
number of elementary operators divided by 2.

• Rank reduction is said to occur when the rank of a commutator or anticommutator

is lower than the combined rank of the operators commuted or anticommuted.

A simple rule

• Rank reduction follows upon anticommutation of two strings of half-integral rank

and upon commutation of all other strings

is useful for simplifying the expressions.

One useful strategy for the evaluation of commutators and anticommutators is based
on their linear expansion in simpler commutators or anticommutators according to e.g. the
following operator identities (see Exercise 3), and aim at greatest possible rank reduction:

h i
Â, B̂1 B̂2 = [Â, B̂1 ]B̂2 + B̂1 [Â, B̂2 ] (4.39)
h i n
X
Â, B̂1 · · · B̂n = B̂1 · · · B̂k−1 [Â, B̂k ]B̂k+1 · · · B̂n (4.40)
k=1
h i
Â, B̂1 B̂2 = [Â, B̂1 ]+ B̂2 − B̂1 [Â, B̂2 ]+ (4.41)

51
 h i n
X
Â, B̂1 · · · B̂n = (−1)k−1 B̂1 · · · [Â, B̂k ]+ · · · B̂n (n even) (4.42)
k=1
h i
Â, B̂1 B̂2 = [Â, B̂1 ]B̂2 + B̂1 [Â, B̂2 ]+ = [Â, B̂1 ]+ B̂2 − B̂1 [Â, B̂2 ] (4.43)
+
h i n
X
Â, B̂1 · · · B̂n = (−1)k−1 B̂1 · · · [Â, B̂k ]+ · · · B̂n (n odd). (4.44)
+
k=1

4.3.1 Evaluation of commutators and anti-commutators

Let us consider the simplest nontrivial commutator, [a†P , a†Q aR ]. Moving one of the opera-
tors after the comma out of the commutators by using (4.41),

[a†P , a†Q aR ] = [a†P , a†Q ]+ aR − a†Q [a†P , aR ]+ (4.45)

and invoking the basic anticommutation relation we have

[a†P , a†Q aR ] = −δP R a†Q . (4.46)

Similarly,
[aP , a†Q aR ] = δP Q aR . (4.47)

Then, let us evaluate the commutator between two excitation operators, using the
results from the previous example,

[a†P aQ , a†R aS ] = [a†P , a†R a†S ]aR + a†P [aQ , a†R aS ] = δQR a†P aS − δP S a†R aQ . (4.48)

A slightly more complicated commutator is evaluated by applying Eq. (4.39) and the
previous result:

[a†P aQ , a†R aS a†M aN ] = [a†P aQ , a†R aS ]a†M aN + a†R aS [a†P aQ , a†M aN ]

= δQR a†P aS a†M aN − δP S a†R aQ a†M aN
+δQM a†R aS a†P aN − δP N a†R aS a†M aQ . (4.49)

The following double commutator can be evaluated by invoking the result (4.48) twice

[a†P aQ , [a†R aS , a†M aN ]] = δSM [a†P aQ , a†R a†N ] − δRN [a†P aQ , a†M aS ]
= δSM δQR a†P aN − δSM δP M a†R aQ − δRN δQM a†P aS
+δRN δP S a†M aQ . (4.50)

This gives rise to a rank reduction by 2.

52
4.4 Orbital rotations
In many situations, e.g., during the optimization of an approximate electronic state or in
the calculation of the response of an electronic state to an external perturbation, it becomes
necessary to carry out transformations between different sets of orthonormal spin-orbitals.
We shall encounter especially cases where the occupation numbers refer to a set of
orthonormal spin-orbitals ψ̃P that is obtained from another set ψP by a unitary transfor-
mation X
ψ̃P = ψQ UQP . (4.51)
Q

Here, a unitary matrix U is a matrix that fulfills the relation

U† U = UU† = 1. (4.52)

The spectral theorem states that any unitary matrix can be diagonalized as

U = VǫV† ,

where V is unitary and ǫ a complex diagonal matrix, ǫkk = exp(iδk ). Therefore any unitary
matrix can be written as the matrix exponential1 of an anti-Hermitian matrix:

U = V exp(iδ)V† = exp(iVδV†)
≡ exp(X), X = −X† . (4.53)

By an introduction of the anti-Hermitian operator

X
κ̂ = κP Q a†P aQ , κ̂† = −κ̂, (4.54)
PQ

where the parameters κP Q are the elements of an anti-Hermitian

  matrix κ, for which
†
U = exp(−κ), the elementary operators ãP , ãP and state 0̃ generated by the unitary
transformation (4.51) can be expressed in terms of the untransformed operators and states
as

ã†P = exp(−κ̂)a†P exp(κ̂) (4.55)

ã = exp(−κ̂)aP exp(κ̂) (4.56)
 P
0̃ = exp(−κ̂) |0i . (4.57)

4.5 Spin in second quantization

In the formalism of second quantization, there is no reference to electron spin. However,
in non-relativistic theory, many important simplifications follow by taking spin explicitly
1 P∞ n
The matrix exponential is defined as exp(A) = n=0 A /n!.

53
into account. Therefore, we shall separate the orbital space and the spin space,2 i.e. a spin
orbital will be denoted by ψP (x) = φp (r)σ(ms ) = φpσ . The theory of second quantization
holds unchanged; for example the basic anticommutator now becomes

[a†pσ , aqτ ]+ = δpσ,qτ = δpq δστ , (4.58)

where, for example, a†pσ is the creation operator associated with the spin-orbital φpσ .

4.5.1 Operators in the orbital basis

We may classify the quantum-mechanical operators according to how they affect the orbital
and spin parts of wave functions: Spin-free operators, spin operators and mixed operators.
One-electron are written in the spin-orbital basis as
XX
fˆ = fpσ,qτ a†pσ aqτ .
pσ qτ

The integrals vanish for opposite spins in the case of spin-free (or spinless) operators:
Z
fpσ,qτ = φ⋆p (r)σ ∗ (ms )f c (r)φq (r)τ (ms )drdms
Z X
= δστ φ⋆p (r)f c (r)φq (r)dr = fpq δστ = fpq Êpq ,
pq

where we have introduced the singlet excitation operator

Êpq = a†pα aqα + a†pβ aqβ . (4.59)

The two-electron operators are given by

1 X
ĝ = gpσ,qτ,rµ,sν a†pσ a†rµ asν aqτ .
2 pσqτ rµsν

Here, the orthogonality of the spin functions reduce the integrals to

gpσ,qτ,rµ,sν = gpqrsδστ δµν , (4.60)

where gpqrs are the two-electron integrals in ordinary space, that are always symmetric
with respect to the particle-interchange, gpqrs = grspq , and have the Hermitian symmetry
⋆
gpqrs = gqpsr in the case of complex orbitals; and in the case of real orbitals, we have a
permutational symmetry
gpqrs = gqprs = gpqsr = gqpsr .
2
Hereafter, the spin functions α and β will be denoted generically by σ, τ , µ and ν.

54
The second-quantization representation of a spin-free two-electron operator is then
1X X 1X 1X
ĝ = gpqrs a†pσ a†rτ asτ aqσ = gpqrs (Êpq Êrs − δqr Eps ) = gpqrsêpqrs . (4.61)
2 pqrs στ
2 pqrs
2 pqrs

We have introduced the two-electron singlet excitation operator in the last equality. We can
now express the spin-free, non-relativistic molecular electronic Hamiltonian in the orbital
basis:
X 1X
Ĥ = hpq Êpq + gpqrs êpqrs . (4.62)
pq
2 pqrs

The one- and two-electron integrals are the same as in (4.31), except that the integrations
are over the spatial coordinates only.
The one-electron second-quantization operators associated to operators that work in
spin space only may be written in the general form
XZ
ˆ
f = φ⋆p (r)σ ⋆ (ms )f c (ms )φq (r)τ (ms )drdms a†pσ aqτ
pσqτ
XZ X
= σ ⋆ (ms )f c (ms )τ (ms )dms a†pσ apτ . (4.63)
στ p

Three important examples of pure spin operators are the raising and lowering operators
Ŝ+c and Ŝ−c , and Ŝzc . The effect of these operators on the (one-particle state) spin functions
is

S+c β = α Ŝ+c α = 0
S−c β = 0 Ŝ−c α = β
1 1
Szc β = − β Ŝzc α = + α.
2 2
From these we arrive at the following expressions for the basic spin operators:
X
Ŝ+ = a†pα apβ (4.64)
p
X
Ŝ− = a†pβ apα (4.65)
p
1 X † †

Ŝz = apα apα − apβ apβ . (4.66)
2 p

The lowering operator is seen to be the Hermitian adjoint of the raising operator:
X X †
Ŝ+† = (a†pα apβ )† = apβ apα = Ŝ− .
p p

55
The operators for the x and y components of the spin angular momentum can be written
in terms of the raising and lowering operators as

1 c

Sxc = S+ + S−c
2
1

Syc = S+c − S−c ,
2i
from which we obtain the second-quantization counterparts,

1 X † 
Ŝx = apα apβ + a†pβ apα (4.67)
2 p
1 X † †

Ŝy = apα apβ − apβ apα , (4.68)
2i p

which may be combined with (4.66) to give the operator for the total spin, Ŝ 2 = Ŝx2 +Ŝy2 +Ŝz2 .
As it contains products of one-electron operators, it is a two-electron operator and therefore
somewhat tedious to manipulate. However, the explicit form is seldom needed, because we
are able to employ the standard operator identities: In contrast to the orbital basis, the
spin basis is complete, hence the usual first-quantization commutation relations hold also
for the second-quantization spin operators. One such example is
" #
X X † X †

† †
[Ŝ+ , Ŝ− ] = apα apβ , aqβ , aqα = apα apα − apβ apβ = 2Ŝz . (4.69)
p q p

For example, the fine-structure and hyperfine-structure operators of first quantization

affect both the spatial and spin parts of the wave function. An example would be the
spin-orbit interaction operator
N
X
c
VSO = ξ(ri )ℓci · Sci .
i=1

The second-quantization representation is obtained in the usual manner as

XZ
V̂SO = φ⋆p (r)σ ⋆ (ms )VSO
c
(r, ms )φq (r)τ (ms )drdms a†pσ aqτ
pσqτ
X 
= Vpqx T̂pq
x
+ Vpqy T̂pq
y
+ Vpqz T̂pq
z
, (4.70)
pq

where Z
Vpqµ = φ⋆p (r)ξ(r)ℓcµφq (r)dr

56
and where the Cartesian components of the triplet excitation operators are given by
1 † 
x
T̂pq = apα aqβ + a†pβ aqα (4.71)
2
1  † 
y
T̂pq = apα aqβ − a†pβ aqα (4.72)
2i
1 † 
z
T̂pq = apα aqα − a†pβ aqβ . (4.73)
2
The second-quantization representation of

• spin-free operators depend on the orbitals but have the same amplitudes (=integrals)
for alpha and beta spins

• pure spin operators are independent of the functional form of the orbitals

• mixed operators depend on both the spin of the electron and the functional form of
the orbitals.

4.5.2 Spin properties of determinants

Slater determinants are in general not eigenfunctions of the non-relativistic Hamiltonian,
but are instead the non-degenerate eigenfunctions of the spin-orbital occupation-number
operators:
N̂pσ |ki = kpσ |ki . (4.74)
We also note that the spin-orbital ON operators commute with the spin-projection opera-
tor, i.e. [Ŝz , N̂pσ ] = 0, because the ON operators commute among themselves, and Ŝz is a
linear combination of spin-orbital ON operators. Added by the observation that there is
no degeneracies in (4.74), the Slater determinants are eigenfunctions of the projected spin:

Ŝz |ki = M |ki . (4.75)

By contrast, the Slater determinants are not eigenfunctions of the total spin. However, it
is possible to determine spin eigenfunctions as simple linear combinations of determinants.
A hint for the procedure is given by an observation that both the total and projected-spin
operators commute with the sum of the ON operators for alpha and beta spins:
h i
Ŝz , N̂pσ + N̂pβ = 0
h i
Ŝ 2 , N̂pσ + N̂pβ = 0,

because N̂pσ + N̂pβ are singlet operators. Such spin-adapted functions are known as con-
figuration state functions.

57
4.6 Further reading
• T. Helgaker, P. Jørgensen, J. Olsen, Molecular Electronic-Structure Theory (Wiley
2002), pp. 1–106.

• W. Kutzelnigg, Quantum chemistry in Fock space, in D. Mukherjee (ed), Aspects of

Many-Body Effects in Molecules and Extended Systems, Lecture Notes in Chemistry
Vol. 50 (Springer-Verlag 1989).

• P. R. Surján, Second Quantized Approach to Quantum Chemistry (Springer-Verlag

1989).

4.7 Exercises for Chapter 4

1. Verify the commutation and anticommutation relations (142-147) in the lectures.

2. Let κ̂ and fˆ be one-electron operators. Show that the commutator [κ̂, fˆ] can be
written as a one-electron operator
X
ˆ =
[κ̂, f] fPκ Q a†P aQ
PQ

with the modified integrals

X
fPκ Q = (κP R fRQ − fP R κRQ ).
R

3. One of the operators that describe the interaction between nuclear magnetic dipoles
and electrons is the Fermi contact operator
XX
c
HFC = γK δ(ri − RK )sci · IK
K i

where sci is the spin of electron i, IK and γK the nuclear spin and the magnetogyric
ratio of nucleus K, respectively, and δ the Dirac delta function. Determine the
second-quantization representation of the Fermi contact operator.

4. ⋆ Consider two operators appearing in the quasi-relativistic Breit–Pauli Hamiltonian,

the two-electron Darwin operator and the spin-spin contact operator:
c π X
HDar(2) = − 2 δ(ri − rj )
c i>j
c 8π X
HSSC = − 2 δ(ri − rj )si · sj .
3c i>j

58
 (a) Show that the second-quantization representation of the two-electron Darwin
operator becomes X
ĤDar(2) = Spqrs a†pα a†rβ asβ aqα ,
pqrs

where Z
π
Spqrs =− 2 φ⋆p (r)φq (r)φ⋆r (r)φs (r)dr.
c
(b) Show that in second quantization these operators are related as

ĤSSC = −2ĤDar(2) .

5. ⋆ Consider a set of spin-orbitals φ̃P obtained from another set φP by a unitary

trasformation X
φ̃P = φQ UQP , U unitary
Q

When U is written in terms of an anti-Hermitian matrix κ as

U = exp(−κ), κ† = −κ,
 
show, that the elementary operators ã†P , ãP as well as state 0̃ generated by the
unitary transformation may be expressed in terms of the untransformed operators
and states as

ã†P = exp(−κ̂)a†P exp(κ̂)

ã = exp(−κ̂)aP exp(κ̂)
 P
0̃ = exp(−κ̂) |0i .

59
60
Chapter 5

Hartree–Fock theory

5.1 The Hartree–Fock approximation

In the Hartree–Fock approximation the electronic wave function is approximated by a single
Slater determinant, and the energy is optimized with respect to variations of these spin-
orbitals. Thus, the wave function may be written in a form

|κi = exp(−κ̂) |0i , (5.1)

where |0i is some reference configuration and exp(−κ̂) an operator that carries out unitary
transformations among the spin orbitals; these orbital-rotation parameters for the Hartree–
Fock ground state are obtained by minimizing the energy:
D   E
 
EHF = min κ Ĥ  κ . (5.2)
κ
The optimization could be carried out using standard numerical analyisis; however, for the
most of purposes more “spesific” methods are required in order to cope with the compu-
tational cost. We recall that the Slater determinant represents a situation where electrons
behave as independent particles, but are subject to the Fermi correlation. Therefore, the
optimal determinant can be found by solving a set of effective one-electron Schrödinger
equations – the Hartree–Fock equations – for the spin-orbitals, with the associated Hamil-
tonian being the Fock operator X
fˆ = fP Q a†P aQ , (5.3)
PQ

where the elements fP Q are called the Fock matrix. A very essential feature of the Fock
operator is that the one-electron part of the “true” molecular Hamiltonian is retained,
but the two-electron part is replaced with the effective Fock operator, so that fˆ = ĥ + V̂ .
The Fock operator incorporates in an average sense the Coulomb interaction among the
electrons, corrected for Fermi correlation:
X
V̂ = VP Q a†P aQ (5.4)
PQ

61
 X
VP Q = (gP QII − gP IIQ ) , (5.5)
I

where the integrals g are given in Eq. (4.33), with I running over the occupied and P and
Q all spin-orbitals. The first term in VP Q describes the classical Coulomb interaction of the
electron with the charge distribution caused by the others, and the second one, exchange
term, is a correction that arises from the antisymmetry of the wave function.
• The Hartree–Fock equations are solved by diagonalizing the Fock matrix. The re-
sulting eigenvectors are called the canonical spin-orbitals and the the eigenvalues of
the Fock matrix are the orbital energies, fP Q = δP Q ǫP .

• Since the Fock matrix is defined in terms of its own eigenvectors, the canonical spin-
orbitals and the orbital energies can only be obtained using an iterative procedure;
where the Fock matrix is constructed and diagonalized in every iteration until the
spin-orbitals obtained in a particular iteration become identical to those from which
the Fock matrix was constructed. This procedure is referred to as the self-consistent
field (SCF) method and the resulting wave function the SCF wave function.
• In the canonical representation, i.e., in the frame where the Fock matrix is diagonal,
the electrons occupy the spin-orbitals in accordance with the Pauli principle and
move independently one another in the electrostatic field caused by the stationary
nuclei and by the charce distribution of the N − 1 electrons. The orbital energies
are the one-electron energies associated with the independent electrons and may thus
be interpreted as the energies required to remove a single electron, that is, to ionize
the system. The identification of the orbital energies with the negative ionization
potentials is known as the Koopman’s theorem.

The Hartree–Fock state is invariant to unitary transformations among the occupied

spin-orbitals. Therefore, the spin-orbitals of the Hartree–Fock state are not uniquely de-
termied by the condition (5.2) and the canonical orbitals are just one possible choice of
spin-orbitals for the optimized N-particle state. Any set of energy-optimized orbitals de-
composes the Fock matrix into two non-interacting blocks: one for the occupied, and
another for the unoccupied spin-orbitals. When these subblocks are diagonalized, the
canonical spin-orbitals are obtained.
The final electronic state is obtained as an antisymmetrized product of the canonical
spin orbitals !
Y † †
|HFi = aiα aiβ |vaci . (5.6)
i

Finally, let us make a remark that the Hartree–Fock state is an eigenfunction of the Fock
operator with an eigenvalue equal to the sum of the orbital energies of the canonicial
spin-orbitals X
fˆ |HFi = 2 ǫi |HFi . (5.7)
i

62
 Figure 5.1: The self-consistent field procedure.

5.2 Restricted and unrestricted Hartree–Fock theory

The exact wave function is an eigenfunction of the total and projected spin operators. The
Hartree–Fock wave function, which is not an eigenfunction of the Hamiltonian (but of the
effective one, the Fock operator), does not possess these symmetries, they must be imposed
on the Hartree–Fock solution.
• In the restricted Hartree–Fock (RHF) approximation, the energy is optimized subject
to the condition that the wave function is an eigenfunction of the total and projected
spins. In a practical implementation of the RHF scheme, it is advantageous to write
the wave function as a configuration state function rather than a Slater determinant.
• In the unrestricted Hartree–Fock (UHF) approximation, the wave function is not
required to be a spin eigenfunction, and different spatial orbitals are used for different
spins.
• In some systems, usually close to the equilibrium geometry, the symmetries of the
exact state are present in the UHF state as well. In such cases the UHF and RHF
states coincide.

5.2.1 Hartree–Fock treatment of H2

For the hydrogen molecule, the RHF wave function is a singlet spin eigenfunction with
doubly occupied symmetry-adapted 1σg orbitals;

|RHFi = a†1σg α a†1σg β |vaci

|UHFi = a†φ1 α a†φ2 β |vaci .

63
 Figure 5.2: Perfomance of RHF and UHF models in H2 .

Let us plot these wave functions as well as the FCI wave function for comparison using the
LCAO approximation for the orbital basis (cc-pVQZ basis set) in Figure 5.2.
At short internuclear distances the RHF and UHF wave functions are identical, pro-
viding a crude but reasonable representation of the potential energy surface. At larger
separations, the RHF wave function behaves poorly for energy, whereas the UHF wave
function falls towards the correct FCI curve. On the other hand, while the RHF wave
function remains a true singlet at all distances, the UHF wave function behaves incorrectly
with respect to the spin, yielding a spin intermediate of a singlet and a triplet.

5.3 Roothaan–Hall equations

We will now consider the classical Roothaan–Hall formulation P of the Hartree–Fock the-
ory, in which the MOs are expanded in a set of AOs, φp = µ χµ Cµp , whose expansion
coefficients C are used as the variational parameters.
The Hartree–Fock energy is the expectation value of the true molecular electronic
Hamiltonian (4.62), and is given in the orbital basis by
D   E X 1X
E (0) = HF Ĥ  HF =
 
hpq Dpq + gpqrs dpqrs (5.8)
pq
2 pqrs

where
D   E
 
Dpq = HF Êpq  HF (5.9)
dpqrs = hHF |êpqrs | HFi . (5.10)
For a closed-shell state, the only nonzero elements of (5.9) and (5.10) are
Dij = 2δij
1
dijkl = Dij Dkl − Dil Dkj = 4δij δkl − 2δil δkj ,
2
64
and thus the total Hartree–Fock energy is
X X
E (0) = 2 hii + (2giijj − gijji). (5.11)
i ij

The energy must be optimized subject to the orthonormality of the MOs. We introduce
the Hartree–Fock Lagrangian
X
L(C) = E(C) − 2 λij (hφi |φj i − δij )
ij

and the variational conditions are written in terms of it as

∂L(C) X X
0= = 4hµk + 4 (2gµkjj − gµjjk ) − 4 Sµj λjk ,
∂Cµk j j

which gives the condition for the optimized Hartree–Fock state

X
fµk = Sµj λjk . (5.12)
j

Since the multiplier matrix λ is symmetric (for real orbitals), it can be diagonalized by
an orthogonal transformation among the occupied orbitals, λ = UǫUT , the Hartree–Fock
energy being invariant to such transformations. The off-diagonal Lagrange multipliers may
thus be eliminated by an orthogonal transformation to a set of occupied MOs that satisfy
the canonical conditions. In the canonical basis, the variational conditions become
X X
AO
fµν Cνk = ǫk Sµν Cνk , (5.13)
ν ν

where the elements of the AO Fock matrix are

X
AO
fµν = hµν + (2gµνii − gµiiν ). (5.14)
i

Since f AO is symmetric, we can extend the variational conditions to yield a set of orthonor-
mal virtual MOs that satisfy the same canonical conditions as the occupied MOs, i.e.,
fab = δab ǫa . Now, we are able to write the Hartree–Fock variational conditions as
f AO C = SCǫ, (5.15)
where ǫ is a diagonal matrix that contains the orbital energies. This equation is called the
Roothaan-Hall equations.
The AO Fock matrix can be evaluated entirely in the AO basis,
 
AO
X
AO 1
fµν = hµν + Dρσ gµνρσ − gµσρν , (5.16)
ρσ
2

where DAO is the AO representation of the one-electron density,

DAO = CDCT . (5.17)

65
 Figure 5.3: Roothaan–Hall procedure.

5.3.1 On the convergence of the SCF method

A straightforward implementation of the Roothaan–Hall SCF method (given in the figure)
may fail to converge or may converge slowly. Several schemes to improve the SCF conver-
gence have been introduced; of which the direct inversion in the iterative subspace (DIIS)
method is amongst the simplest and one of the most successful. In the DIIS method, the
information from the preceding iterations is used instead of generating the next density
from the last AO Fock matrix:
X n
AO
f̄n = wi fiAO , (5.18)
i=1

where the DIIS

Pn weights w are obtained by minimizing the norm of the averaged error
vector ēn = i=1 wi ei ,
n X
X n
kek2 = wi hei |ej i wj ,
| {z }
i=1 j=1
Bij

Pn
subject to the constraint that i=1 wi = 1. For this purpose, we construct the Lagrangian

n n
!
X X
L= wi Bij wj − 2λ wi − 1 ,
i,j=1 i=1

66
 100
SCF
10 DIIS

1
SCF
0.1 DIIS

∆ Eh 0.01

0.001

1e-04

1e-05
10 20 10 20
Iteration Iteration

Figure 5.4: On the left: H2 O in its equilibrium geometry, on the right: H2 O with its
bonds stretched to double (angle remaining constant). Calculations are carried out in the
cc-pVQZ basis.

the minimization of which leads to the set of linear equations:

    
B11 B12 · · · B1n −1 w1 0
 B21 B22 · · · B2n −1   w2   0 
 . .. .. .. ..   .   . 
 ..  .   . 
 . . .  .  =  . 
. 
B
n1 Bn2 · · · Bnn −1   wn   0 
−1 −1 · · · −1 0 λ −1

from which the weights are solved.

P
• The C2-DIIS method is similar to the DIIS, but the constraint is ni=1 wi2 = 1. This
leads to a solution of an eigenvalue problem instead of the set of linear equa tions as
in DIIS; the scheme may handle singularities better.

• In the EDIIS (“energy DIIS”) scheme, the error-vector-norm minimization is replaced

by a minimization of an approximation to the true energy function. It improves the
convergence in cases where the start guess has a Hessian structure far from the
optimal one. It also can converge in cases where DIIS diverges. For non-problematic
cases EDIIS has actually a slower convergence rate than DIIS.

• These are based on either the diagonalization of the Fock matrix or on a direct
optimization of the AO density matrix. Nevertheless, in some cases, even the DIIS
method (and alike) is difficult to converge, when it might be better to use the second-
order Newton method to the optimization of the Hartree–Fock energy. This has been
applied in the literature to the optimization of the AO density matrix (the basis for

67
 the trust-region or restricted-step method) as well as in a method that carries out
rotations among the MOs.

• The “trust-region SCF” (TRSCF) is the state-of-the-art model in SCF convergence.

It is basically a globally convergent black-box method, where each SCF iteration
consists of a trust-region density subspace minimization and a trust-region Roothaan-
Hall step.

5.3.2 Integral-direct SCF

In every Roothaan–Hall SCF iteration a large number of two-electron integrals is needed
to construct the Fock matrix. Since the same AO integrals are used in each iteration, these
integrals can be written to disk and read in when needed.
In large-scale calculations, this approach becomes impractical due to the large number
of needed integrals. In such cases, the AO integrals are instead recomputed in every
iteration – as soon as a batch of two-electron integrals has been produced, it is contracted
with the appropriate densities. This procedure is known as the direct SCF method.
If the value of an integral is very small, or if it is combined with small density-matrix el-
ements, its contribution to the Fock matrix may be neglected. Therefore, by pre-estimating
the integrals by having some kind of upper limit for them, the number of calculated in-
tegrals in the direct SCF scheme can be reduced dramatically. The computational cost
of the SCF method scales formally as O(n4 ), where n is the number of basis set func-
tions, but it can be shown that only O(n2 ) of the two-electron integrals are significant.
Therefore, the direct SCF scheme combined with the pre-screening of the integrals is much
cheaper for large systems. Furthermore, by using so-called fast multipole methods, only
O(n) two-electron integrals is needed, while the rest can be either neglected or evaluated
classically. The bottleneck is then not anymore in the Fock-matrix construction, but in
the diagonalization step (which is an O(n3 ) procedure).
To overcome the diagonalization bottleneck and to achieve a SCF method that would
scale linearly with respect of the size of the system, and thus be applicable to large molecular
systems such as biomolecules, we would need to optimize directly the AO density matrix.

68
5.4 Further reading
• T. Helgaker, P. Jørgensen, J. Olsen, Molecular Electronic-Structure Theory (Wiley
2002), pp. 433–522.

• J. Almlöf, Notes on Hartree–Fock theory and related topics, in B. O. Roos (ed),

Lecture Notes in Quantum Chemistry II (Springer-Verlag 1994).

• J. Almlöf, Direct methods in electronic structure theory, in D. R. Yarkony (ed),

Modern Electronic Structure Theory (World Scientific 1995), p. 110.

5.5 Exercises for Chapter 5

1. Here we consider the total energies of the singlet and triplet states obtained by
promoting one electron from the occupied canonical orbital φi to the virtual canonical
orbital φa of a closed-shell Hartree–Fock state. The total energies of the singlet and
triplet states can be written as
D   E
1  0,0 0,0 
Eai = HF Ŝia Ĥ Ŝai  HF
D   E
3
Eai = HF T̂ia1,0 Ĥ T̂ai1,0  HF ,
 

where
1
0,0
Ŝpq = √ (a†pα aqα + a†pβ aqβ )
2
1
1,0
T̂pq = √ (a†pα aqα − a†pβ aqβ ).
2

(a) Show that the singlet total energy can be expressed as

D   E 1D   E
1    
Eai = HF Ĥ  HF + HF Eia [Ĥ, Eai ] HF .
2

(b) Show, that the singlet and triplet energies are related as
D   E
3  β
Eai =1 Eai − 2 HF Eia α
ĤEai  HF ,

β
α
where Eai = a†aα aiα and Eai = a†aβ aiβ .

2. In this exercise, we shall determine the gradient for the closed-shell Hartree–Fock
energy D   E
 
∂ cs exp(κ̂)Ĥ exp(−κ̂) cs
(1)
EP Q (κ) =
∂κP Q

69
 for a non-vanishing κ̂ that contains only real parameters κP Q . According to exercise
4.3⋆ , the derivative of an exponential operator is
∂ exp(−κ̂) X exp [i(λR − λS )] − 1

=− UP⋆ R UQS − UQR
⋆
UP S ā†R āS exp(−κ̂),
∂κP Q RS
i(λR − λS )

where UP Q are the elements of the unitary matrix U that diagonalizes κ; and λP are
the real eigenvalues of iκ:
iκ = UλU† .
The ā†P are the creation operators in the eigenvector-basis of κ,
X †
ā†P = aR URP .
R

(a) Show that the elements of the gradient are given by

(1)
X exp [i(λR − λS )] − 1

EP Q (κ) = UP⋆ R UQS − UQR
⋆
UP S
RS
i(λR − λS )

(b) Specialize this result to the case κ = 0.

⋆
3. Consider the orbital-rotation operator
X
κ̂ = κP Q (a†P aQ − a†Q aP ),
P >Q

where κP Q are real parameters.

(a) Verify the relation

Z 1
d
exp[−(κ̂ + δκ̂)] − exp(−κ̂) exp[−(κ̂ + δκ̂)z] exp(κz)dz exp(−κ̂).
0 dz

(b) Using this result, write the first-order variation of exp(−κ̂) with respect to δκ̂
as ∞
X (−1)n+1
δ exp(−κ̂) = [κ̂, [κ̂, . . . [κ̂, δκ̂] . . .]] exp(−κ̂).
n=0
(n + 1)! | {z }
n commutators

(c) Show, that the real antisymmetric matrix κ with elements


κP Q P ≥ Q
[κ]P Q =
−κP Q P < Q
can be diagonalized as
κ = −iUλU†
where U is a unitary matrix and λ a real diagonal matrix.

70
 (d) The eigenvectors U define a basis
X
ā†P = a†R URP .
R

Show that, in this eigenvector basis, the orbital-rotation operator and its first-
order variation may be expressed as
X
κ̂ = −i λP ā†P āP
P
X 
δκ̂ = U† δκU P Q ā†P āQ .
PQ

(e) Using the previous expansions, rewrite the expression for δ exp(−κ̂) as
X exp[i(λP − λQ )] − 1
δ exp(−κ̂) = − [U† δκU]P Q ā†P āQ exp(−κ̂).
PQ
i(λ P − λ Q )

(f) Show that the first derivatives of the orbital-rotation operator are given by
∂ exp(−κ̂) X exp [i(λR − λS )] − 1

=− UP⋆ R UQS − UQR
⋆
UP S ā†R āS exp(−κ̂),
∂κP Q RS
i(λR − λS )

⋆
4. In the optimization of a closed-shell electronic system, the nonredundant orbital-
rotation parameters  
0 κTAI
κ=
−κAI 0
determine the transformation
P from the original MOs φp (0) to a set of improved MOs
φp (κ) as φp (κ) = φ
q p (0) exp(−κ) qp . A further rotation of φp (κ) involving the
redundant parameters  
red κIJ 0
ρ=κ =
0 κAB
P
gives a new MO basis ψp (ρ) = q φq (κ) exp(−ρ)qp which is equivalent to φp (κ) in the
sense that it gives the same energy; in other words, the total energy is invariant to this
transformation. In this exercise, we show that the choice of redundant parameters
ρ = 0 minimizes the function
X
D(ρ) = kφp (0) − ψp (ρ)k2 ,
p

which represents the difference between the original and optimized MOs.
(a) Show that the difference measure D(ρ) can be expressed in the matrix form

D(ρ) = 2Tr [1 − exp(−κ) exp(−ρ)] .

71
(b) Show that the first-order variation of D(ρ) at ρ = 0 becomes

δD(ρ) = 2TrWδρ,

where    
WIJ 0 exp(−κ)IJ 0
W= = .
0 WAB 0 exp(−κ)AB
(c) Show that the exponential of κ can be written as
 
cos A −κTAI B−1 sin B
exp(−κ) = ,
κAI A−1 sin A cos B

where A = (κTAI κ)1/2 , and B = (κAI κT )1/2 .

(d) Use the previous result for exp(−κ) to show that ρ = 0 represents a stationary
point.
(e) Show that, at the stationary point, the second-order variation of D(ρ) becomes

δ 2 D = −TrWδρδρ.

(f) Show that W is positive definite for small κ and use this to show that the second-
order variation is positive, implying that the stationary point is a minimum.
Thus, for a closed-shell wave functions, ρ = 0 gives the maximum overlap
between the rotated and original MOs.

72
Chapter 6

Configuration interaction

6.1 Configuration-interaction wave function

The configuration-interaction (CI) wave function is constructed as a linear combination of
Slater determinants (or configuration-state functions):
X
|Ci = Ci |ii (6.1)
i

with the coefficients determined by a variational optimization of the expectation value of

the electronic energy, i.e., by minimization for the ground state
D   E
 
C Ĥ  C
ECI = min . (6.2)
C hC |Ci
As discussed earlier, this is equivalent to the solution of an eigenvalue equation HC =
ECI C. The construction of the DCI wave
 Efunction may thus be carried out by diagonalizing
 
the Hamiltonian matrix Hij = i Ĥ  j , but more often iterative techniques are used.
In particular, we have to construct the CI wave function in a finite one-electron basis.
The full CI (FCI) wave function consists of all the configurations that can be generated
in a given basis, and it is most convenient to think it as generated from a single reference
configuration, which dominates the wave function and usually is the Hartree–Fock state,
by the application of a linear combination of spin-orbital excitation operators
!
X 1 X
|FCIi = 1 + X̂IA + AB
X̂IJ + · · · |HFi , (6.3)
AI
2 ABIJ

where for example

X̂IA |HF i = CIA a†A aI |HFi

AB AB † †
X̂IJ |HF i = CIJ aA aB aI aJ |HFi .

73
 R = Rref R = 2Rref
E − EFCI Weight E − EFCI Weight
RHF 0.2178 0.9410 0.3639 0.5896
CISD 0.0120 0.9980 0.0720 0.9487
CISDT 0.0090 0.9985 0.0560 0.9590
CISDTQ 0.0003 1.0000 0.0058 0.9988
CISDTQ5 0.0001 1.0000 0.0022 0.9995

Table 6.1: The electronic energies of truncated CI wave functions for the water molecule
(cc-pVDZ basis) relative to the FCI energy.

Thus, we may characterize the determinants in the FCI expansion as single (S), double (D),
triple (T), quadruple (Q), quintuple (5), and so forth excitations relative to the Hartree–
Fock state.
The number of determinants in an FCI wave function is
 2  2
n n!
Ndet = = , (6.4)
k k!(n − k)!
for a system with n orbitals containing k alpha and k beta electrons. For example, when
n = 2k, Ndet ≈ 16k /kπ – for large k, the number of determinants increases by a factor of
16 for each new pair of electrons and orbitals!

6.1.1 Single-reference CI wave functions

As the FCI wave function is obtainable only for the smallest systems, it becomes necessary
to truncate the expansion, so that only a subset of the determinants are included, based
on the excitation level. This is justified, since in general the lower-order excitations are
more important than those of high orders. In principle, this procedure of hierarchal trun-
cation may be continued until the FCI wave function is recovered. Since the CI model is
variational, the FCI energy is approached from the above.
In Table 6.1 the truncated CI functions are examined. Whereas the contributions drop
monotonically, the step from an even-order excitation level to an odd-order is far more
pronounced than the step from an odd to an even-order excitation level; hence the CI
expansion is preferably truncated at even orders. For example, the first useful truncated
CI wave function – the CI singles-and-doubles wave function (CISD) – recovers 94.5% of
the correlation energy, the inclusion of triples on top of that improves the treatment to
95.9%, while as much as 99.9% of the correlation energy is recovered at the CISDTQ level.

6.1.2 Multi-reference CI wave functions

In the stretched geometry 2Rref , the Hartree–Fock determinant is less dominant. Clearly,
the strategy of defining a correlation hierarchy in terms of excitations out of a single

74
reference determinant such as the Hartree–Fock determinant does not provide the best
possible description of the dissociation process. To overcome this, we may introduce CI
wave functions based on the idea of a reference space comprising more than a single deter-
minant: The multi-reference CI (MRCI) wave functions is generated by including in the
wave function all configurations belonging to this reference space as well as all excitations
up to a given level from each reference configuration, yielding, e.g., the multi-reference
singles-and-doubles CI (MRSDCI) wave function.
The construction of a MRCI wave function begins with the generation of a set of orbitals
and the reference space of configurations. For example, in the water molecule, we would
include in our reference space for instance the configurations obtained by distributing the
eight valence electrons among the six valence MOs in all possible ways (consistent with the
spin- and space-symmetry restrictions).

• MRSDCI model is very accurate, if all the important configurations are included in
the reference space.

• The number of important configurations becomes, however, very large, making thus
the model applicable to quite small systems.

• The choice of the active space requires a lot of chemical intuition and patience.

Another way of truncating the FCI expansion is to construct the wave function from
individual configurations, the selection of which is based on physical intuition or perturba-
tional estimates.

6.1.3 Optimization of the CI wave function

For large CI expansions it is impossible to set up and diagonalize the Hamiltonian matrix
of the eigenvalue equation. However, usually only a few of the lowest eigenvalues are of
interest, and we may determine them by iterative methods, where the eigenvectors and
eigenvalues are generated by a sequence of linear transformations of the form σ = HC,
where C is some trial vector. These methods can be devised for both of the CI eigenvalue
problem and the optimization of the CI energy (6.2). The most often used strategies rely
on the second-order Newton’s method and its approximations.

6.1.4 On the disadvantages of the CI approach

CI wave functions truncated at a given excitation level relative to a reference configuration
do not provide size-extensive energies.

• For example, if two fragments are described at the CISD level, then a size-extensive
treatment of the compound system requires that the wave function is augmented
with certain triples and quadruples; more precisely those that represent products of
single and double excitations in the two subsystems.

75
 • An approach called quadratic CI (QCI) is a size-extensive revision of the CI model.
Other widely used approach to overcome the lack of size-extensivity is the Davidson
correction applied on top of the truncated CI energy.

The CI description of the electronic system is not at all compact. Even though higher
excitations are less important than those of lower orders, their number is enormous. There-
fore, the CI wave function converges slowly with the number of variational parameters.
Both of these problems are due to the linear parameterization in the CI model (6.1).

6.2 Multi-configurational SCF theory

The multi-configurational self-consistent field (MCSCF) theory is a generalization of the
Hartree–Fock wave function to systems dominated by more than one electronic configura-
tion. This method is particularly useful in the description of bond breakings and molecular
dissociation processes.
In MCSCF theory, the wave function is written as a linear combination of determinants,
whose expansion coefficients Ci are optimized simultaneously with the MOs according to
the variation principle: X
|κ, Ci = exp(−κ̂) Ci |ii . (6.5)
i

The ground-state MCSCF wave function is obtained by minimizing the energy with respect
to the variational parameters,
D   E
 
κ, C Ĥ  κ, C
EMCSCF = min . (6.6)
κ,C hκ, C |κ, Ci

6.2.1 MCSCF wave function of H2

A two-configuration MCSCF wave function of the hydrogen molecule is

|MCi = C1 a†1σg α a†1σg β |vaci + C2 a†1σu α a†1σu β |vaci ,

where the variation principle is invoked to optimize the configuration coefficients as well
as the orbitals. Its behaviour in H2 dissociation is sketched in Figure 6.2.1. The solid line
is the MCSCF potential energy curve in the cc-pVQZ basis set. The thick and the dashed
lines are for FCI and RHF, respectively.
For any larger system than the H2 molecule the selection of configuration space is
the greatest difficulty in the MCSCF calculation. It is very often impossible to generate
a sufficient MCSCF configuration, which is computationally tractable. The selection is
usually carried out by dividing the MO space in inactive, active, and secondary orbitals.

• The inactive orbitals are set to be doubly occupied.

76
 Figure 6.1: MCSCF potential energy curve of H2 .

• The active orbitals have variable occupancies of 0, 1, or 2. The MCSCF expansion

is then obtained by distributing the active electrons in all possible ways among the
active orbitals. In the final optimized state, the active orbitals have non-integer
occupation numbers between 0 and 2.
• The secondary orbitals are unoccupied throughout the optimization.
This scheme is called the complete active space SCF method (CASSCF). We note that the
CASSCF wave function is identical to the FCI wave function, when all the orbitals of the
system are active; and reduces to the Hartree–Fock wave function when the active space
is empty.
In restricted active space SCF (RASSCF) calculations, the active orbital space is further
divided into three subspaces: RAS1, with an upper limit on the allowed number of holes in
each configuration; RAS2, with no restrictions enforced; and RAS3, with an upper limit on
the allowed number of electrons in each configuration. This allows for larger active spaces
than CASSCF.
In practise, the simultaneous optimization of orbitals and CI coefficients is a difficult
nonlinear optimization problem, which restricts the MCSCF expansions to be significantly
smaller than those encountered in the CI wave functions. By itself, the MCSCF model
is not suited to the treatment of dynamical correlation, that requires large basis sets and
long configuration expansions.

6.2.2 Selection of the active space

One popular approach for the choice of the MCSCF active spaces is based on the natural
occupation numbers obtained from a preceding second-order Møller–Plesset perturbation
theory calculation. For the water molecule in the cc-pVTZ basis in the order (A1 , B1 , B2 ,
A2 ) we obtain

77
 Natural orbital occupation numbers, symmetry 1

1.99953982 1.98393843 1.96476967 0.02261760 0.01156239

0.00648257 0.00552982 0.00113424 0.00083070 0.00060446
0.00047604 0.00036853 0.00035173 0.00019527 0.00013628
0.00008516 0.00007619 0.00005837 0.00004768 0.00003078
0.00002067 0.00001341 0.00000675

Natural orbital occupation numbers, symmetry 2

1.96724668 0.01959000 0.00552930 0.00068405 0.00064037

0.00042732 0.00035178 0.00013127 0.00005653 0.00004947
0.00001646

Natural orbital occupation numbers, symmetry 3

1.96415036 0.02452309 0.00641995 0.00104230 0.00086705

0.00065869 0.00039618 0.00019746 0.00013435 0.00011930
0.00006217 0.00005180 0.00003331 0.00001667 0.00001255
0.00000755 0.00000429

Natural orbital occupation numbers, symmetry 4

0.00626166 0.00079588 0.00038289 0.00012271 0.00006547

0.00006019 0.00001431

Table 6.2: Natural occupation numbers of C2 v water molecule in cc-pVTZ basis.

We can choose active spaces of different flexibility:

• Smallest possible active space seems to be (2,0,2,0), i.e. four electrons in four MOs
3a1 , 4a1 , 1b2 , and 2b2 . This corresponds to a combination of 12 determinants.

• For more accurate description, eight electrons distributed on (3,1,2,0) active space,
corresponding to 37 configurations, gives almost similar potential curve.

• For further improvement, an eight-electron (6,3,3,1) active space is of high quality

and sufficient for incorporating dynamical correlation effects as well. However, we get
almost the same curve by supplementing the CAS(3,1,2,0) wave function by RAS3
space (3,2,1,0) limiting the number of electrons in RAS3 space to be 2 at maximum.
This reduces the number of determinants from 36240 to 2708.

78
6.3 Further reading
• T. Helgaker, P. Jørgensen, J. Olsen, Molecular Electronic-Structure Theory (Wiley
2002), pp. 523–647.
• I. Shavitt, The Method of Configuration Interaction, in H. F. Schaefer III (ed), Meth-
ods of Electronic Structure Theory (Plenum Press 1977).
• H.-J. Werner, Matrix-Formulated Direct Multiconfiguration Self-Consistent Field and
Multiconfiguration Reference Configuration-Interaction Methods, in K. P. Lawley
(ed), Ab Initio Methods in Quantum Chemistry (Wiley 1987).

6.4 Exercises for Chapter 6

1. Calculate the non-vanishing one- and two-electron density-matrix elements
Dpq = h0 |Epq | 0i
X
  
dpqrs = h0 |epqrs | 0i = 0 a†pσ a†qτ arσ asτ  0
στ

for the following states of the H2 molecule in the minimal basis:

(a) the single-configuration singlet states
1 2 
σ
g = a†1α a†1β |vaci
1 2 
 σu = a†2α a† |vaci
2β

(b) the two-configuration singlet state

1 +     
 Σg (τ ) = cos(τ ) 1σg2 + sin(τ ) 1σu2

(c) the triplet and singlet ungerade states

3 +  1  † † † †

Σ
u = √ a a
2α 1β + a2β 1α |vaci
a
2
 
 Σu = √1 a†2α a† − a† a†1α |vaci
1 + 
1β 2β
2
2. ⋆ Consider a dimer of two noninteracting H2 molecules, each described in a minimal
basis.
(a) Show that the ground-state FCI wave function can be parametrized as
 
|FCIi = 1 + c1 D̂1 + c2 D̂2 + c12 D̂1 D̂2 |HFi ,

where the operator D̂I = a†uI α a†uI β agI β agI α creates a double excitation from the
Hartree–Fock state for molecule I.

79
 (b) Show that the FCI eigenvalue problem becomes
    
0 K K 0 c0 c0
 K 2∆ 0 K   c1  = Ecorr  c1 
   

 K 0 2∆ K   c2   c2 
0 K K 4∆ c12 c12

where Ecorr is the FCI correlation energy of the H2 dimer and

K = ggugu
2∆ = 2huu − 2hgg + guuuu − ggggg .

(c) Find the ground-state FCI solution.

3. ⋆ Assume that we are carrying out an MCSCF calculation of the electronic state
X (0)
|0i = Ci |ii
i

where the determinants share a commonDset of inactive

E orbitals. Show that, for this
 
state, the Hamiltonian matrix elements i Ĥ  0 may be written in the form

D   E X I D   E 1X
gvwxy hi |êvwxy | 0i + Ci0 I E
   
i Ĥ  0 = Fvw i Êvw  0 +
vw
2 vwxy

where I F is the inactive Fock matrix

X
I
Fmn = hmn + (2gmnii − gmiin )
i

and I E the inactive energy. Obtain an expression for the MCSCF electronic gradient.

4. ⋆ Show that the FCI wave function is size-extensive but the truncated CI-doubles
(CID) wave function is not.

80
Chapter 7

Description of dynamical correlation

7.1 Coupled-cluster theory

As noted earlier, the shortcomings of the CI hierarchy – the lack of size-extensivity and
the slow convergence towards the FCI limit – are due to the linear expansion. However,
we may recast the FCI expansion (6.3) in the form of a product wave function
" #
Y
|CCi = (1 + tµ τ̂µ ) |HFi , (7.1)
µ

where we have introduced a generic notation τ̂µ for an excitation operator of unspecified
excitation level (D, T, Q,...) and the associated cluster amplitude tµ . This (full) coupled-
cluster (CC) wave function differs from the FCI wave function by the presence of terms
that are nonlinear in the excitation operators. Of course, the CC and FCI wave functions
are completely equivalent provided that all the excitations are included.
Since the τ̂ s always excite from the set of occupied Hartree–Fock spin-orbitals to the
virtual ones, [τ̂µ , τ̂ν ] = 0. Due to the presence of the product excitations in the coupled-
cluster state (7.1), each determinant can thus be reached in several different ways. For
example, the determinant |µνi = τ̂µ τ̂ν |HFi = τ̂ν τ̂µ |HFi may be reached with an overall
amplitude equal to the sum of the individual amplitude. With respect to this determinant,
the double-excitation amplitude tµν is referred to as a connected cluster amplitude and
subsequent singles excitation amplitudes tµ tν as a disconnected cluster amplitude.

7.1.1 The CC Schrödinger equation

By analogy with the CI theory, we could attempt to determine the CC state by minimizing
the expectation value of the Hamiltonian with respect to the amplitudes. Due to the
nonlinear parameterization of the CC model, the analogous variational conditions would
give rise to an intractable set of nonlinear equations. Hence, in practice coupled-cluster
models are not solved using the variation principle, and the CC energies are therefore
non-variational.

81
 Instead, by multiplying the eigenvalue equation Ĥ |CCi = E |CCi by hµ| = hHF| τ̂µ† we
obtain the projected coupled-cluster equations
D   E
 
µ Ĥ  CC = E hµ |CCi , (7.2)

where the CC energy is obtained by projection against the Hartree–Fock state (note that
hHF |CCi = 1) D   E
 
E = HF Ĥ  CC . (7.3)
These equations are nonlinear in amplitudes as well, but unlike the variational conditions,
this expansion will, due to the Slater’s rules, terminate after few terms, as we shall see
later.

7.1.2 The coupled-cluster exponential ansatz

Because the excitation operators in (7.1) commute, i.e. τ̂µ2 = 0, the correlating operators
can be written in more convenient form, known as the exponential ansatz for the coupled-
cluster wave function,
|CCi = exp(T̂ ) |HFi , (7.4)
where we have introduced the cluster operator
X
T̂ = tµ τ̂µ . (7.5)
µ

In CC theory, a hierarchy of approximations is established by partitioning the cluster

operator in the form
T̂ = T̂1 + T̂2 + T̂3 + · · · + T̂n , (7.6)
where, for example,
X X †
T̂1 = tA A
I τ̂I = tA
I aA aI (7.7)
AI AI
X 1 X AB †
T̂2 = tAB AB
IJ τ̂IJ = tIJ aA aI a†B aJ . (7.8)
A>B,I>J
4 AIBJ

Each excitation operator in T̂ excites at least one electron from an occupied Hartree–Fock
spin-orbital to a virtual one. It should be noted that for a system with N electrons,
the expansion terminates after N terms. We further note that [T̂i , T̂j ] = 0 as a trivial
consequence of the individual excitation operator commutation relations. Moreover,

hHF| τ̂µ = 0
τ̂µ† |HFi = 0,

because it is impossible to excite an electron from an unoccupied orbital.

82
 To compare the CC model with the CI model, we expand the exponential operator and
collect terms to the same order in excitation level:
N
X
exp(T̂ ) |HFi = Ĉi |HFi .
i=0

The lowest-order operators Ĉi are given by

Ĉ0 = 1
Ĉ1 = T̂1
1 2
Ĉ2 = T̂2 + T̂
2! 1
1 3
Ĉ3 = T̂3 + T̂1 T̂2 + T̂
3! 1
1 1
Ĉ4 = T̂4 + T̂3 T̂1 + T̂12 T̂2 + T̂14 ,
2! 4!
demonstrating the composition of the excitation processes. The advantages of the cluster
parameterization arise in truncated wave functions and are related to the fact that even at
the truncated level, the coupled-cluster state contains contributions from all determinants
in the FCI wave function, with weights obtained from the different excitation processes.
Similarly with the CI theory, a hierarchy of CC models is introduced by truncating the
cluster operator at different excitation levels. The simplest successful CC wave function
is the coupled-cluster singles-and-doubles (CCSD) model. In this model, the T̂2 operator
describes the electron-pair interactions and T̂1 carries out the corresponding orbital re-
laxations. Its computational cost scales as O(N 6 ). For higher accuracy, the connected
triple excitations have to be taken into account, which leads to the coupled-cluster singles-
doubles-and-triples (CCSDT) model. This model describes already the dynamical correla-
tion almost perfectly, but is computationally very demanding, O(N 8 ) and is thus applicable
to smallest molecules only. The triples excitations can be estimated in a perturbational
way, giving rise to the CCSD(T) method, which is a good compromise between the com-
putational cost (exhibiting an O(N 7 ) scaling) and accuracy. It is often considered to be
the best single-reference treatment of molecular Schrödinger equation that is applicable in
practise.

7.1.3 CI and CC models compared

Let us at this point compare the CI and CC hierarchies. Owing to the disconnected
clusters, CC wave functions truncated at a given excitation level also contain contributions
from determinants corresponding to higher-order excitations; whereas CI wave functions
truncated at the same level contain contributions only from determinants up to this level.
This is reflected in the convergence towards the FCI limit, which is typical of small systems.
For larger systems, CI starts to behave very badly, while the CC description is unaffected
by the number of electrons.

83
Figure 7.1: On the left: The CC dissociation curves of the C2v water molecule in the
cc-pVDZ basis. On the right: corresponding differences between the CC and FCI energies.
RHF reference state is plotted the full line and UHF with the dashed.

84
Figure 7.2: The error with respect to FCI of CC and CI wave functions for the water
molecule at the equilibrium geometry in the cc-pVDZ basis.

According to the previous discussion, the projected coupled-cluster equations are writ-
ten as
D   E
 
HF Ĥ exp(T̂ ) HF = E (7.9)
D   E D   E
   
µ Ĥ exp(T̂ ) HF = E µ exp(T̂ ) HF . (7.10)

• The excited projection manifold hµ| comprises the full set of all determinants up to
the chosen truncation level.

• For the full CC wave function, the number of equations is equal to the number
of determinants and the solution of the projected equations recovers the FCI wave
function.

• The nonlinear equations (7.9) must be solved iteratively, substituting in every itera-
tion the CC energy calculated from (7.10).

It is convenient to rewrite the projected CC equations using the non-Hermitian similarity-

transformed Hamiltonian H̄ ˆ = exp(−T̂ )Ĥ exp(T̂ ):
D   E
 
HF exp(−T̂ )Ĥ exp(T̂ ) HF = E (7.11)
D   E
 
µ exp(−T̂ )Ĥ exp(T̂ ) HF = 0 (7.12)

The equations (7.10) and (7.12) are completely equivalent (yet the equivalence is nontrivial
for all the truncated CC models!), meaning that they yield same amplitudes and energy
upon solution. They are referred to as the unlinked and linked coupled-cluster equations,
respectively. Their computational cost is practically the same, but the linked equations

85
have some advantages: Although both yield the same, size-extensive wave function, the
linked equations are size-extensive term by term, that simplifies some perturbational treat-
ments; in addition, the CC approach excited states in mind is more fruitful to start from
the linked equations. The main reason is however the following:
• When the Baker–Campbell–Hausdorff (BCH) expansion
1 1
exp(A)B exp(−A) = B + [A, B] + [A, [A, B]] + [A, [A, [A, B]]] + · · · (7.13)
2! 3!
is applied to the similarity-transformed Hamiltonian, one obtains
1 1 1
exp(−T̂ )Ĥ exp(T̂ ) = Ĥ+[Ĥ, T̂ ]+ [[Ĥ, T̂ ], T̂ ]+ [[[Ĥ, T̂ ], T̂ ], T̂ ]+ [[[[Ĥ, T̂ ], T̂ ], T̂ ], T̂ ],
2 6 24
(7.14)
the coupled-cluster equations are therefore no higher than quartic in the cluster am-
plitudes – for any truncation level or even for the full CC expansion.
• Due to the Brillouin theorem, the one-particle operators contribute only to second
order. Furthermore, cluster operators higher than the doubles do not contribute to
the CC energy since Ĥ is a two-particle operator. As a result, only singles and doubles
amplitudes contribute directly to the CC energy – irrespective of the truncation level
of the cluster operator. Of course, the higher-order excitations contribute indirectly,
since all amplitudes are coupled by the projected equations.

7.1.4 On coupled-cluster optimization techniques

The elements of the left-hand side of the CC amplitude equation (7.12) constitute the
coupled-cluster vector function Ωµ (t), which can be expanded around the set of amplitudes
of the current iteration t(n) :

Ω(t(n) + ∆t) = Ω(0) (t(n) ) + Ω(1) (t(n) ∆t + · · · ,

where Ω(0) is the vector function calculated from the amplitudes t(n) and Ω(1) the CC
Jacobian D   E
Ω(1) (n) (n) 

µν = µ exp(−T̂ )[Ĥ, τ̂ν ] exp( T̂ )  HF . (7.15)
Then an iterative scheme is established:
• Solve the set of linear equations

Ω(1) (t(n) )∆t(n) = −Ω(0) (t(n) ) (7.16)

for ∆t(n)
• Form the improved estimate

t(n+1) = t(n) + ∆t(n) (7.17)

86
 • Iterate Equations (7.16) and (7.17) until convergence in the amplitudes.
In more sophisticated implementations, this Newton-method-like approach is replaced by
a scheme, where the solution of the linear equations as well as the construction of the
Jacobian are avoided by the approximation
∆t(n) −1 (0) (n)
µ = −εµ Ωµ (t ), (7.18)
where, e.g. εab
ij = (ǫa + ǫb − ǫi − ǫj ). Usually, the convergence is accelerated by the DIIS
method.

7.1.5 The closed-shell CCSD model

We consider now in detail the important special case of CC theory: the closed-shell CCSD
model. We recall that the singlet CCSD state is generated by |CCi = exp(T̂1 + T̂2 ) |HFi,
hence only those terms in T̂i that transform as singlet operators should be retained in the
cluster operator. It is easy to show that
h i
Ŝ± , T̂i = 0 (7.19)
h i
Ŝz , T̂i = 0. (7.20)
These impose important constraints on the cluster amplitudes in the spin-orbital basis.
Due to these, the singles cluster operator becomes
X
T̂1 = tai Êai (7.21)
ai

and the doubles operator

1 X ab
T̂2 = t Eai Ebj . (7.22)
2 aibj ij
The cluster amplitudes are symmetric: tab ba
ij = tji .
The CCSD energy is obtained from Eq. (7.11). By noting that hHF| exp(−T̂ ) = hHF|,
expanding the cluster operators and after some algebra we obtain the form
1D   E D   E
   
ECCSD = EHF + HF [[Ĥ, T̂1 ], T̂1 ] HF + HF [Ĥ, T̂2 ] HF , (7.23)
2
D   E
 
where the Hartree–Fock energy is given by EHF = HF Ĥ  HF . By inserting the CCSD
operators to the energy expression and calculating the commutators explicitly we obtain
1X a bD   E 1X D   E
  ab  
ECCSD = EHF + ti tj HF [[Ĥ, Êai ], Êbj ] HF + tij HF [Ĥ, Êai Êbj ] HF
2 aibj 2 aibj
1 X ab
D   E
tij + tai tbj HF [[Ĥ, Êai ], Êbj ] HF
 
= EHF +
2 aibj
X

= EHF + tab a b
ij + ti tj Liajb , (7.24)
aibj

87
where we have denoted Lpqrs = 2gpqrs − gpsrq .

7.1.6 The equation-of-motion coupled-cluster method

In CC theory, we arrive at a good description of the ground state. The excited states
could be examined by carrying out separate, independent calculation for each state of
interest, using in each case some appropriate zero-order reference determinant. However,
it is usually difficult or even impossible to determine these adequate reference determinants;
it is expensive to carry out individual calculations; and the states obtained in this manner
are not orthogonal, which hinders the identification of the excited states.
Instead, we calculate the excited states in the spirit of CI theory, by a linear expansion
in the space spanned by all states of the form
X X
|ci = cµ τ̂µ |CCi = exp(T̂ ) cµ τ̂µ |HFi , (7.25)
µ µ

where the summation is carried out over all the excitation operators present in the cluster
operator. This expansion is referred to as the equation-of-motion coupled-cluster (EOM-
CC) model. The expansion parameters are optimized by the minimization of the energy
D   E
 
c̄ exp(−T̂ )Ĥ exp(T̂ ) c
E(c, c̄) = . (7.26)
hc̄ |ci

The EOM-CC theory can be thought as conventional CI theory with a similarity-

transformed Hamiltonian, which carries the information about electron correlation; while
the configuration expansions carry the information about the excitation structure of the
electronic states.

7.1.7 Orbital-optimized coupled-cluster theory

In the standard CC theory, we use the Hartree–Fock orbitals and then determine a set of
non-zero single-excitation amplitudes together with higher-excitation amplitudes. Alter-
natively, we could use exp(T̂1 ) to generate an orbital transformation to a basis in which the
single-excitation amplitudes vanish. This is the idea behind the orbital-optimized (OCC)
theory, in which the orbital-rotation operator exp(−κ̂) is used instead of exp(T̂1 ) (as they
generate the same state to first order),

|OCCi = exp(−κ̂) exp(T̂O ) |0i , (7.27)

where T̂O = T̂2 + T̂3 + · · · + T̂N . The OCC energy is obtained in the usual manner by
projecting the OCC Schrödinger equation against the reference state
D   E
 
EOCC = 0 exp(−T̂O ) exp(κ̂)Ĥ exp(−κ̂) exp(T̂O ) 0 , (7.28)

88
whereas the cluster amplitudes are determined by projection against the manifold |µO i
spanned by T̂O D   E
 
µO exp(−T̂O )Ĥ exp(−κ̂) exp(T̂O ) 0 = 0. (7.29)
The Brueckner coupled-cluster theory (BCC) is closely related to OCC and differs in some
details of the solution of orbital rotation parameters. Unlike in standard CC theory, the
OCC and BCC orbitals are optimized simultaneously with the optimization of the cluster
amplitudes – compare with the MCSCF theory. Surprisingly, the differences between stan-
dard and OCC/BCC wave functions are small. Only in the rare cases that are characterized
by the sc. Hartree–Fock singlet instability the OCC/BCC are clearly more successful than
the standard formulation.

7.2 Perturbation theory

Finally, we consider less rigorous but computationally more appealing improvements of the
Hartree–Fock description, based on perturbation theory. They are rather successful, when
the Hartree–Fock wave function is reasonably accurate.
In Møller–Plesset perturbation theory (MPPT) the electronic Hamiltonian is partitioned
as
Ĥ = fˆ + Φ̂ (7.30)
where fˆ is the Fock operator (5.3) and Φ̂ the fluctuation potential, which is the difference
between the true two-electron Coulomb potential in Ĥ and the effective one-electron po-
tential V̂ of the Fock operator, Φ̂ = ĝ − V̂ . The zero-order state is Prepresented by the
ˆ
Hartree–Fock state in the canonical representation, that is f |HFi = I ǫI |HFi, and by
applying the standard machinery of perturbation theory, we obtain to second order in the
perturbation
D   E X
EMP = HF fˆ HF =
(0)  
ǫI (7.31)
I
D   E
(1)  
EMP = HF Φ̂ HF (7.32)
(2)
X |gAIBJ − gAJBI |2
EMP = − . (7.33)
A>B,I>J
ǫA + ǫB − ǫI − ǫJ

Thus, the Hartree–Fock energy is equal to the sum of the zero- and first-order contribu-
tions; and by adding the second-order correction, we obtain the second-order Møller–Plesset
energy (MP2)
X |gAIBJ − gAJBI |2
EMP2 = EHF − , (7.34)
A>B,I>J
ǫA + ǫB − ǫI − ǫJ

which is very popular and successful approach to introducing electron correlation on top of
the Hartree–Fock wave function; providing surprisingly accurate and size-extensive correc-
tion at low computational cost. Higher-order corrections are derived in a similar manner,

89
Figure 7.3: The MPn dissociation curves of the C2v water molecule in the cc-pVDZ basis.
Restricted MP2 with full and unrestricted with dashed line.

yielding MP3 and MP4 models, but these become rather expensive and are not as appeal-
ing compromise between cost and accuracy as MP2 is. Furthermore, the MPPT series have
been shown to be inherently divergent.
In the coupled-cluster perturbation theory (CCPT) the Hamiltonian in the CC equa-
tions (7.11) and (7.12) is partitioned as in Eq. (7.30). The Fock and the cluster operator
obey the following results:
h i X
fˆ, T̂ = εµ tµ τ̂µ
µ
hh i i
f , T̂ , T̂ = [[[fˆ, T̂ ], T̂ ], T̂ ] = · · · = 0.
ˆ

Thus we may write the non-Hermitian similarity-transformed Fock operator as

X
fˆ¯ = fˆ + εµ tµ τ̂µ , (7.35)
µ

90
and we arrive at the working equations of the CCPT:
D   E
ˆ
E = E0 + HF Φ̄ HF (7.36)
D   E
ˆ
εµ tµ = − µ Φ̄  HF . (7.37)

By expanding the cluster operator in the orders of the fluctuation potential we arrive at
the amplitude equations D   E
 ˆ (n) 
εµ tµ(n) = − µ [Φ̄]  HF (7.38)
ˆ ](n) contains the nth order part of the similarity-transformed fluctuation potential.
where [Φ̄
Then the amplitudes would be determined self-consistently from the projected equations
after truncating the cluster operator at some excitation level. The CCPT is most often
used to establish approximations to contributions from some excitations – the most famous
is the non-iterative triples, or (T) correction to the CCSD wave function. Also very suc-
cessful iterative CCPT schemes have been set up, e.g. the CC2 model approximates the
CCSD model with the computational cost O(N 5 ) and CC3 the CCSDT model with the
computational cost O(N 7 ). The CCPT series provides more systematic improvement to
the wave function than the MPPT.
Both MPPT and CCPT are limited to the ground states of systems dominated by a
single configuration. To overcome this, multiconfigurational perturbation theory has been
developed. Most important such theory is CAS perturbation theory (CASPT), where the
zero-order state is taken to be a CASSCF wave function. At the moment, it is the only
generally applicable method for the ab initio calculation of dynamical correlation effects of
open- and closed-shell multiconfigurational electronic system.

7.3 Local correlation methods

Traditional correlated electronic wave function methods, which are built upon the de-
localized Hartree–Fock molecular orbitals (MO) do not exploit the fact that the elec-
tron correlation—i.e., the electron-electron Coulomb cusp—is spatially a rather local phe-
nomenon.
In the recent literature, localized correlation methods have been pursued to overcome
the “scaling wall” encountered in the correlated wave function methods. However, despite
recent progress, there is still a need to formulate and implement new approaches that
exploit the locality of the electron correlation problem.

• Approaches based on introducing locality in the correlation treatment using localized

MOs (LMO) appeared already in 1960s and in the context of CC methods in 1980s.

• An alternative approach was presented by Head-Gordon et al., who formulated the

CC equations in a general non-orthogonal [e.g., the atomic orbital] basis using a
general tensor-based formulation. Scuseria and Ayala presented a transformation

91
 of the standard MO CC doubles (CCD) equations into the AO basis. Scuseria and
Ayala argued that the solution of the CC equations in the AO basis scales in principle
linearly with respect to N, but their approach seems so far not to have received any
widespread use.

• A fully operational local CCSD implementation have been presented by Werner et al.
This implementation, referred to as the local CCSD (LCCSD) method, is based on
localizing the occupied space as suggested by Pulay and coworkers, using traditional
MO localization schemes, and using for the virtual space a projected atomic orbital
(PAO) basis. LCCSD further relies on a spatial division of the system into domains.
Restrictions on the extent to which the interactions between different domains are
accounted for is important for achieving the low-power scaling of LCCSD making the
results deviate from the standard CCSD results.

• The equations of correlated wave function theories may also be derived in a biorthog-
onal projected atomic orbital basis, aiming at a black-box formulation that could be
applied to large systems. This formulation also constitutes a good platform to intro-
duce efficient approximation schemes to reduce the number of parameters and thus
the computational scaling. The locality of the PAO basis is reflected in the fact
that the cluster amplitudes, the integrals as well as the correlation energy contri-
butions decay exponentially with respect to the distance between the centers of the
participating PAOs.

92
7.4 Further reading
• T. Helgaker, P. Jørgensen, J. Olsen, Molecular Electronic-Structure Theory (Wiley
2002), pp. 648–816.

• T. J. Lee and G. E. Scuseria, Achieving chemical accuracy with coupled-cluster theory,

in S. R. Langhoff (ed), Quantum-Mechanical Electronic Structure Calculations with
Chemical Accuracy (Kluwer 1995).

• O. Christiansen, Coupled cluster theory with emphasis on selected new developments,

Theor. Chim. Acta 116, 106 (2006).

7.5 Exercises for Chapter 7

1. Evaluate the commutator [Ĥ, Êai ] and the expression for this commutator applied to
the Hartree–Fock state [Ĥ, Êai ] |HFi.

2. Verify the overlap between the doubly excited states:

  
ab  cd
= 4δac δbd δik δjl + 4δad δbc δil δjk − 2δac δbd δil δjk − 24δad δbc δik δjl
ij  kl
= 2Pijab (2δaibj,ckdl − δajbi,ckdl ).

The operator P carries out the permutation Pijab Aab ab ba

ij = Aij + Aji .

⋆
3. Verify the following commutation relations:
hh i i X
Ĥ, Êai , Êbj = −Pijab hib Êaj − Pijab (gpbiq êpjaq + gpqjaêpqbi )
pq
X
+ (gpbqa êpjqi + gjpiq êbpaq )
pq
hhh i i i X
abc
Ĥ, Êai , Êbj , Êck = Pijk (gkbip êpbic + gpbic êpjak ) ,
p

and the expressions for these commutators applied to the Hartree–Fock state:
hh i i
Ĥ, Êai , Êbj |HFi = 2Liajb |HFi
!
X X
−Pijab I Fib Êaj + Likjb Êak − Lcajb Êci |HFi
k c
!
X X
+ gbjcaÊbj Êci − gbjik Êbj Êak |HFi
bjc bjk

93
 hhh i i i
abc
Ĥ, Êai , Êbj , Êck |HFi = −Pijk Ljbic Êak |HFi
!
X X
abc
+Pijk giljcÊal Êbk − gidbc Êaj Êdk |HFi .
l d

In the above, we have introduced the notation Lpqrs = 2gpqrs − gpsrq as well as used
the three-index permutation operator,
abc abc
Pijk Aijk = Aabc acb bac bca cba
ijk + Aikj + Ajik + Ajki + Akji .

⋆
4. Consider, once again, the H2 molecule in a minimal basis.

(a) Show that the cluster operator becomes

1
T̂ = T̂2 = tuu Êug Êug
2 gg
and that the biorthogonal basis can be expressed as
 
uu 
= hvac| auβ auα .
gg 

(b) Show that the unlinked coupled-cluster equations can be expressed as

D   E
 
HF Ĥ(1 + T̂2 ) HF = E
       
uu   uu  
Ĥ(1 + T̂2 ) HF = E T̂  HF .
gg gg 2

(c) Show that the linked coupled-cluster equations are

D   E
 
HF Ĥ(1 + T̂2 ) HF = E
   
uu  
(1 − T̂2 )Ĥ(1 + T̂2 ) HF = 0.
gg

(d) Show, that the linked and unlinked coupled-cluster equations are equivalent.

5. ⋆ Construct expressions for the CCSD vector functions

   
a 
Ωai = exp(−T̂2 )Ĥ exp(T̂2 ) HF
i
   
ab  
Ωaibj = exp(−T̂2 )Ĥ exp(T̂2 ) HF
ij

94
6. ⋆ The basic idea behind the quantum-mechanical perturbation theory (known as the
Rayleigh–Schrödinger perturbation theory, RSPT) is to partition the Hamiltonian
into two parts, Ĥ = Ĥ0 + Û , where Ĥ0 is some zero-order Hamiltonian and Û the
perturbation
 Ĥ − Ĥ0 . The orthonormal eigensolutions to the zero-order Hamiltonian,
Ψ(0) with the corresponding energy E (0) are then used to expand the eigenfunction
|Ψi of the exact Hamiltonian.
(a) Show, that
n
X
(0)
 (n)   (n−1)   
(Ĥ0 − E ) Ψ
 = −Û Ψ
 ++ E (k) Ψ(n−k)
k=1

and that all corrections to the zero-order energy are obtained as

D   E
(n) (0)   (n−1)
E = Ψ Û  Ψ .

(b) Write out explicitly three lowest-order corrections to the energy and to the wave
function.
7. ⋆ Consider the wave function correct to nth order in RSPT:
Xn
 ( 
|Ψn i = Ψ i) .
i=0

Show that the energy calculated as the expectation value

D   E
 
Ψn Ĥ  Ψn
E=
hΨn |Ψn i
is correct to order 2n + 1. This result is known as the Wigner’s rule. In  addition,
  as
a special case, consider the wave function correct to first order, |Ψ1 i = Ψ(0) + Ψ(1)
and show that the expectation value may then be expressed in terms of the RSPT
corrections as D   E
 
Ψ1 Ĥ  Ψ1 E (2) + E (3)
= E (0) + E (1) + .
hΨ1 |Ψ1 i 1 + hΨ(1) |Ψ(1) i
⋆
8. Here we generalize the RSPT formalism to the expansion of the perturbed states
 (n)  X (n)
Ψ = Ci |ii
i

in a nonorthogonal basis, whose overlap elements differ from the Kronecker delta:
hi |ji =
6 δij .
(a) Show that the zero-order equations and the condition for the intermediate nor-
malization become
H0 C(0) = E (0) SC(0)
(C(0) )T SC(0) = 0.

95
96
Chapter 8

Performance of the
electronic-structure models

Having now considered the most important approximations to the molecular electronic
wave function we will now address their performance in terms of a few numerical examples.
It should be remembered that none of these models (besides FCI, which is inapplicable for
most cases) is an all-around-good one: good performance in some problem and/or in some
system does not guarantee good performance in another situation. In addition, rehearse
the graph concerning the approximations to the one- and N-electron spaces.

8.1 Molecular equilibrium geometries

One of the most common applications of molecular electronic-structure calculations is
the determination of the molecular equilibrium structures. Good geometries are also of
fundamental importance in study of molecular properties.
We consider the mean errors and standard deviations of calculated bond distances and
bond angles as compared to experimental results using a hierarchy of N-electron models
and a hierarchy of Gaussian basis sets. The error is given by ∆ = Picalc − Piexp , and the
mean error and standard deviation by
n
¯ = 1
X
∆ ∆i (8.1)
n i=1
v
u n
u 1 X
∆std = t ¯ 2.
(∆i − ∆) (8.2)
n − 1 i=1

They can be conveniently represented graphically in terms of the normal distribution

"  #
1 P −∆ ¯ 2
ρ(P ) = Nc exp − . (8.3)
2 ∆std

97
Figure 8.1: Performance of various N-electron models in conjunction with one-electron
bases in 29 molecular bond lengths.

98
The calculations were carried out for sets of small molecules containing main-group ele-
ments. 29 bond distances and 10 bond angles were calculated, and the results are presented
in Figures 8.1 and 8.2.
Observations:
• The Hartree–Fock model is suitable for only preliminary calculations.
• The CCSD(T) model is capable of providing structures – both bond distances and
angles – of high quality.
• The MP2 model is less accurate but more generally applicable as it is considerably less
expensive. Higher-order MPPT is significantly less accurate but hideously expensive,
i.e., useless.
• The performance of the CCSD model is disappointing. It however performs more
robust than MP2 in some selected cases with complicated electronic structure.
• The CISD model is also rather useless, showing no improvement when the basis set
is increased.
• The basis-set requirements are different for different models. Whereas there is no
much use to use QZ-level basis sets with the Hartree–Fock model (as the results
are not improved), the correlated models require at least a TZ-level basis set, but
preferably the cc-pVQZ basis.
The results suggest the following recipe for the determination of molecular structures:
obtain an initial structure with HF using a DZ-level basis and improve it by using MP2
and a TZ basis. If ever possible, calculate the final geometry in the cc-pVQZ basis using
the CCSD(T) model.

8.2 Molecular dipole moments

The comparison between the experimental and calculated molecular dipole moments is
difficult, as the experiments are measuring the dipole moment in the vibrational ground
state µ0 , whereas the calculations are carried out for the equilibrium dipole moment µe ,
and thus we would have to carry out a vibrational averaging in order to speak of the same
quantity. However, there are a few experimental values for µe , and we shall compare the
performance of our standard models with those. Note, that due to the outer-valence origin
of this property, we have to augment our basis sets with extra diffuse primitive Gaussians;
hence we are using the aug-cc-pVXZ basis sets. The geometries are optimized using the
same wave function as for the calculation of the dipole moment.
We observe from Table 8.1, that the quality of the calculation is more dependent on
the correlation treatment than the cardinal number of the basis set; yet it is necessary to
augment the basis set. The further augmentation (d-aug-cc-pVXZ basis sets) of the basis
set does not affect the calculated dipole moments significantly.

99
Figure 8.2: Performance of various N-electron models in conjunction with one-electron
bases in 10 molecular bond angles.

100
 ¯ of the cc-pVQZ basis.
Figure 8.3: Changes in bond lengths calculated from ∆

Exp. HF MP2 CCSD CCSD(T)

QZ QZ QZ DZ TZ QZ
NH3 1.561(5) 1.544 1.517 1.530 1.541 1.513 1.521
HF 1.803(2) 1.884 1.811 1.813 1.799 1.797 1.800
H2 O 1.8473(10) 1.936 1.864 1.870 1.859 1.845 1.853

Table 8.1: Calculated dipole moments.

101
 Figure 8.4: Reaction enthalpies.

8.3 Reaction enthalpies

Next, we study the reaction enthalpies of 13 chemical reactions among closed-shell molecules.
The calculations are carried out in the HF–MP2–CCSD–CCSD(T) hierarchy, using the
correlation-consistent basis set hierarchy supplemented with steep functions to provide
core-polarization, which – somewhat surprisingly – must be included in the high-precision
calculation of reaction enthalpies. The results are illustrated in Figure 8.4.

8.4 Notes on total electronic energies

For the total molecular energy, i.e., E in the molecular Schrödinger equation, there is no
experimental counterpart. and we examine it in order to establish a feeling on the sever-
ity of the approximations involved in the calculation. We should recall that there were a
third class of approximations in addition to the truncation of one- and N-electron spaces:
approximations in the molecular Hamiltonian Ĥ. To investigate the validity of the use of
the non-relativistic Hamiltonian, we include the leading-order one-electron relativistic cor-
rections that include the spin-orbit interaction (SO), mass-velocity (MV), and the Darwin
(Dar) corrections. The leading-order two-electron contributions, such as the two-electron
Darwin contribution and the spin-spin contact interaction, are smaller by at least one order
of magnitude. The MV and Dar corrections are always of opposite sign. The calculation
is carried out using the CCSDT model for the water molecule in the cc-pCVXZ bases at
a CCSD(T)/cc-pCVQZ geometry.
Observations:

102
 Total energy [Eh ] Energy corrections [mEh ]
CCSDT HF HF→SD SD→(T) (T)→T SO MV Dar
DZ -76.24121 -76.02680 -211.21 -3.04 -0.16
TZ -76.33228 -76.05716 -267.40 -7.65 -0.08
QZ -76.35981 -76.06482 -285.98 -8.99 -0.01 0.00 -255.87 203.69
5Z -76.36899 -76.06708 -292.44 -9.52 0.04

Table 8.2: Decomposition of the CCSDT energy of water.

• It is remarkable how a tiny, less than one per cent fraction of the total energy, plays
such an important role in the determination of molecular properties.

• The leading-order relativistic correction, -52.18 mEh , is larger than the triples con-
tribution already in water, and gets far more pronounced in systems with heavier
elements present!

• (T) is a very good approximation to full triples.

103