This action might not be possible to undo. Are you sure you want to continue?
http://chem.ps.uci.edu/˜kieron/dft/book/
The ABC of DFT
Kieron Burke and friends
Department of Chemistry, University of California, Irvine, CA 92697
April 10, 2007
2
Contents
I Background 13
1 Introduction 15
1.1 Importance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2 What is a KohnSham calculation? . . . . . . . . . . . . . . . . . . . . . . 16
1.3 Reinterpreting molecular orbitals . . . . . . . . . . . . . . . . . . . . . . . . 19
1.4 Sampler of applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.5 Particle in a box . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.6 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2 Functionals 27
2.1 What is a functional? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Functional derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3 EulerLagrange equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3 One electron 33
3.1 Variational principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Trial wavefunctions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Three dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.4 Diﬀerential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.5 Virial theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.6 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4 Two electrons 39
4.1 Antisymmetry and spin . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 HartreeFock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5 Many electrons 47
5.1 Ground state . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 HartreeFock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3
4 CONTENTS
5.3 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.4 Atomic conﬁgurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.5 Atomic densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.6 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
II Basics 55
6 Density functional theory 57
6.1 One electron . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.2 HohenbergKohn theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.3 ThomasFermi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6.4 Particles in boxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.5 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7 KohnSham 65
7.1 KohnSham equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.2 Exchange . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
7.3 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
7.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
8 The local density approximation 71
8.1 Local approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
8.2 Local density approximation . . . . . . . . . . . . . . . . . . . . . . . . . . 72
8.3 Uniform electron gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
8.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
9 Spin 77
9.1 KohnSham equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
9.2 Spin scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
9.3 LSD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
9.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
10 Properties 81
10.1 Total energies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
10.2 Densities and potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
10.3 Ionization energies and electron aﬃnities . . . . . . . . . . . . . . . . . . . 83
10.4 Dissociation energies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
10.5 Geometries and vibrations . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
10.6 Transition metals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
10.7 Weak bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
CONTENTS 5
10.8 Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
10.9 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
III Analysis 87
11 Simple exact conditions 89
11.1 Size consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
11.2 One and two electrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
11.3 LiebOxford bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
11.4 Bond breaking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
11.5 Uniform limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
11.6 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
12 Scaling 93
12.1 Wavefunctions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
12.2 Density functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
12.3 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
12.4 Correlation inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
12.5 Virial theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
12.6 Kinetic correlation energy . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
12.7 Potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
12.8 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
13 Adiabatic connection 103
13.1 One electron . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
13.2 Adiabatic connection formula . . . . . . . . . . . . . . . . . . . . . . . . . 105
13.3 Relation to scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
13.4 Static correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
13.5 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
14 Discontinuities 113
14.1 Koopman’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
14.2 Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
14.3 Derivative discontinuities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
14.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
15 Analysis tools 117
15.1 Enhancement factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
15.2 Density analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
15.3 Energy density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6 CONTENTS
15.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
16 Exchangecorrelation hole 121
16.1 Density matrices and holes . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
16.2 Hooke’s atom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
16.3 Transferability of holes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
16.4 Old faithful . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
IV Beyond LDA 135
17 Gradients 137
17.1 Perimeter problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
17.2 Gradient expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
17.3 Gradient analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
17.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
18 Generalized gradient approximation 147
18.1 Fixing holes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
18.2 Visualizing and understanding gradient corrections . . . . . . . . . . . . . . 151
18.3 Eﬀects of gradient corrections . . . . . . . . . . . . . . . . . . . . . . . . . 153
18.4 Satisfaction of exact conditions . . . . . . . . . . . . . . . . . . . . . . . . 154
18.5 A brief history of GGA’s . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
18.6 Questions about generalized gradient approximations . . . . . . . . . . . . . 155
19 Hybrids 157
19.1 Static, strong, and strict correlation . . . . . . . . . . . . . . . . . . . . . . 157
19.2 Mixing exact exchange with GGA . . . . . . . . . . . . . . . . . . . . . . . 157
19.3 Questions about adiabatic connection formula and hybrids . . . . . . . . . . 157
20 Orbital functionals 159
20.1 Selfinteraction corrections . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
20.2 Optimized potential method . . . . . . . . . . . . . . . . . . . . . . . . . . 159
20.3 G¨ orlingLevy perturbation theory . . . . . . . . . . . . . . . . . . . . . . . 159
20.4 Meta GGA’s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
20.5 Jacob’s ladder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
V Timedependent DFT 161
21 Timedependence 163
CONTENTS 7
21.1 Schr¨ odinger equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
21.2 Perturbation theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
21.3 Optical response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
21.4 Questions about timedependent quantum mechanics . . . . . . . . . . . . . 170
22 Timedependent density functional theory 171
22.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
22.2 RungeGross theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
22.3 KohnSham equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
22.4 Adiabatic approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
22.5 Questions on general principles of TDDFT . . . . . . . . . . . . . . . . . . 176
23 Linear response 179
23.1 Dysonlike response equation and the kernel . . . . . . . . . . . . . . . . . 179
23.2 Casida’s equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
23.3 Singlepole approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
24 Performance 183
24.1 Sources of error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
24.2 Poor potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
24.3 Transition frequencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
24.4 Atoms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
24.5 Molecules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
24.6 Strong ﬁelds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
25 Exotica 187
25.1 Currents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
25.2 Initialstate dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
25.3 Lights, camera, and...Action . . . . . . . . . . . . . . . . . . . . . . . . . . 187
25.4 Solids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
25.5 Back to the ground state . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
25.6 Multiple excitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
25.7 Exact conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
A Math background 191
A.1 Lagrange multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
A.2 Properties of the δfunction . . . . . . . . . . . . . . . . . . . . . . . . . . 192
A.3 Fourier transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8 CONTENTS
B Results for simple oneelectron systems 195
B.1 1d H atom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
B.2 Harmonic oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
B.3 H atom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
C Green’s functions 199
D Further reading 201
E Discussion of questions 203
F Solutions to exercises 209
F.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
F.2 Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
F.3 One electron . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
F.4 Two electrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
F.5 Many electrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
F.6 Density functional theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
F.7 KohnSham . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
F.8 Local density approximation . . . . . . . . . . . . . . . . . . . . . . . . . . 255
F.9 Spin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
F.10 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
F.11 Simple exact conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
F.12 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
G Answers to extra problems 285
CONTENTS 9
Deﬁnitions and notation
Coordinates
Position vector: r = (x, y, z), r = r.
Spin index: σ =↑ or ↓ = α or β.
Spacespin vector: x = (r, σ).
Sums:
dx =
¸
σ
d
3
r
Operators
Kinetic energy:
ˆ
T = −
1
2
N ¸
i
∇
2
i
.
Potential energy:
ˆ
V =
ˆ
V
ee
+
ˆ
V
ext
Coulomb repulsion:
ˆ
V
ee
=
1
2
¸
i=j
1/r
i
−r
j
.
External potential:
ˆ
V
ext
=
N ¸
i=1
v
ext
(r
i
)
Wavefunctions
Physical wavefunction: Ψ[n](x
1
...x
N
) has density n and minimizes
ˆ
T +
ˆ
V
ee
KohnSham: Φ[n](x
1
...x
N
) has density n and minizes
ˆ
T
Φ(x
1
...x
N
) =
¸
p
(−1)
p
φ
1
(x
p1
)...φ
N
(x
pN
)
Φ(x
1
...x
N
) =
¸
p
(−1)
p
φ
1
(x
p1
)...φ
N
(x
pN
)
where φ
i
(x) and
i
are the ith KS orbital and energy, with i = α, σ.
Energies
Universal functional: F[n] = min
Ψ→n
Ψ
ˆ
T +
ˆ
V
ee
Ψ = Ψ[n]
ˆ
TΨ[n]
Kinetic energy: T[n] = Ψ[n]
ˆ
TΨ[n].
Noninteracting kinetic energy: T
s
[n] = min
Φ→n
Φ
ˆ
TΦ = Φ[n]
ˆ
TΦ[n] =
N ¸
i=1
d
3
r∇φ
i
(r)
2
Coulomb repulsion energy: V
ee
[n] = Ψ[n]
ˆ
V
ee
Ψ[n].
Hartree energy: U[n] =
1
2
d
3
r
d
3
r
n(r) n(r
)/r −r

Exchange: E
X
= ΦV
ee
Φ −U = −
1
2
¸
σ
¸
i,j
occ
d
3
r
d
3
r
φ
∗
iσ
(r) φ
∗
jσ
(r
) φ
iσ
(r
) φ
jσ
(r)/r −r

Kineticcorrelation: T
C
[n] = T[n] −T
S
[n]
Potentialcorrelation: U
XC
[n] = V
ee
[n] −U[n]
Exchangecorrelation: E
XC
[n] = T[n] − T
S
[n] + V
ee
[n] − U[n] = Ψ[n]
ˆ
T +
ˆ
V
ee
Ψ[n] −
Φ[n]
ˆ
T +
ˆ
V
ee
Φ[n]
Potentials
Functional derivative: F[n + δn] −F[n] =
d
3
r δn(r) δF[n]/δn(r)
KohnSham: v
S
(r) = −δT
S
/δn(r) +µ
Hartree: v
H
(r) = δU/δn(r) =
d
3
r
n(r
)/r −r

Exchangecorrelation: v
ext
(r) = v
S
(r) +v
H
(r) +v
XC
(r)
Densities and density matrices
Spin density: n(x) = n
σ
(r) = N
dx
2
. . . dx
N
Ψ(x, x
2
. . . , x
N
)
2
=
N ¸
i=1
φ
i
(x)
2
properties: n(x) ≥ 0,
dx n(x) = N
10 CONTENTS
Density matrix: γ(x, x
) = N
dx
2
. . .
dx
N
Ψ
∗
(x, x
2
, . . . , x
N
)Ψ(x
, x
2
, . . . , x
N
)
properties: γ(x, x) = n(x), γ(x
, x) = γ(x, x
)
kinetic energy: T = −
1
2
d
3
r∇
2
γ(r, r
)
r=r
KohnSham: γ
S
(x, x
) = δ
σσ
Nσ ¸
i=1
φ
∗
iσ
(r)φ
iσ
(r
)
Pair Density: P(x, x
) = N(N −1)
dx
3
. . .
dx
N
Ψ(x, x
, x
3
, . . . , x
N
)
2
properties:
dx
P(x, x
) = (N −1)n(x), P(x
, x) = P(x, x
), P(x, x
) ≥ 0
potential energy:V
ee
=
1
2
d
3
r
d
3
r
P(r, r
)/r −r

KohnSham: P
X
(x, x
) = n(x)n(x
) −γ
S
(x, x
)
2
Exchangecorrelation hole around r at coupling constant λ:
P
λ
(r, r +u) = n(r) [n(r +u) +n
λ
XC
(r, r +u)]
KohnSham: n
X
(x, x
) = −γ
S
(x, x
)
2
/n(x)
properties: n
X
(r, r +u) ≤ 0,
d
3
u n
X
(r, r +u) = −1,
d
3
u n
λ
C
(r, r +u) = 0
Pair correlation function: g
λ
(x, x
) = P
λ
(x, x
)/(n(x)n(x
))
Electronelectron cusp condition: dg
λ
(r, u)/du
u=0
= λg
λ
(r, u = 0)
Uniform coordinate scaling
Density: n(r) →n
γ
(r) = γ
3
n(γr)
Wavefunction: Ψ
γ
(r
1
. . . r
N
) = γ
3N/2
Ψ(γr
1
. . . γr
N
)
Ground states: Φ[n
γ
] = Φ
γ
[n], but Ψ[n
γ
] = Ψ
γ
[n]
Fundamental inequality: F[n
γ
] ≤ γ
2
T[n] + γV
ee
[n]
Noninteracting kinetic energy: T
S
[n
γ
] = γ
2
T
S
[n]
Exchange and Hartree energies: E
X
[n
γ
] = γE
X
[n], U[n
γ
] = γU[n]
Kinetic and potential: T[n
γ
] < γ
2
T[n], V
ee
[n
γ
] > γV
ee
[n] (γ > 1)
Correlation energies: E
C
[n
γ
] > γE
C
[n], T
C
[n
γ
] < γ
2
T
C
[n] (γ > 1)
Virial theorem: 2T = r · ∇
ˆ
V
N electrons: 2T +V
ee
=
d
3
r n(r) r · ∇v
ext
(r)
XC: E
XC
+T
C
=
d
3
r n(r) r · ∇v
XC
(r)
Spin scaling
Kinetic: T
S
[n
↑
, n
↓
] =
1
2
(T
S
[2n
↑
] +T
S
[2n
↓
])
Exchange: E
X
[n
↑
, n
↓
] =
1
2
(E
X
[2n
↑
] +E
X
[2n
↓
])
Adiabatic connection
HellmannFeynman: E = E
λ=0
+
1
0
dλΨ
λ
dH
λ
/dλΨ
λ
Wavefunction: Ψ
λ
[n] has density n and mininimizes
ˆ
T + λ
ˆ
V
ee
;
relation to scaling: Ψ
λ
[n] = Ψ
λ
[n
1/λ
]
Energies: E
λ
[n] = λ
2
E[n
1/λ
]
kinetic: T
λ
S
[n] = T
S
[n]
exchange: E
λ
X
[n] = λE
X
[n], U
λ
[n] = λU[n]
correlation: E
λ
C
[n] = λ
2
E
C
[n
1/λ
] = λ
2
E
(2)
C [n] + λE
(3)
S [n] +. . .
for small λ
ACF: E
XC
=
1
0
dλ U
XC
(λ), where U
XC
(λ) = U
λ
XC
/λ
CONTENTS 11
Finite systems
Kato’s cusp at nucleus: dn/dr
r=Rα
= −2Z
α
n(R
α
)
Large r in Coulombic system:
n(r) →Ar
β
e
−
√
2Ir
, β =
¸
α
Z
α
−N + 1/
√
2I
Exchange potential: v
X
(r) →−1/r
Correlation potential: v
C
(r) →−α(N −1)/2r
4
, where α(N −1) is the polarizability of the
N −1 electron system.
VonWeisacker: T
VW
S
[n] =
d
3
r ∇n
2
/(8n) (exact for N = 1, 2)
Exchange: E
X
= −U/N for N = 1, 2
Correlation: E
C
= 0 for N = 1
Highdensity limit: E
C
[n
γ
] = E
(2)
C [n] +E
(3)
C [n]/γ +. . . as γ →∞
Lowdensity limit: E
C
[n
γ
] = γB[n] + γ
3/2
C[n] +. . . as γ →0.
Uniform gas and LSD
Measures of the local density WignerSeitz radius: r
s
(r) = (3/(4πn(r))
1/3
Fermi wave vector: k
F
(r) = (3π
2
n(r))
1
3
ThomasFermi wavevector: k
s
(r) =
4k
F
(r)/π
Measure of the local spinpolarization:
Relative polarization: ζ(r) = (n
↑
(r) −n
↓
(r))/n(r)
Kinetic energy: t
unif
S
(n) = (3/10)k
2
F
(n)n
exchange energy: e
unif
X
(n) = n
unif
X
(n), where
unif
X
(n) = (3k
F
(n)/4π),
correlation energy:
unif
C
(r
s
) →0.0311 ln r
s
−0.047 + 0.009r
s
ln r
s
−0.017r
s
(r
s
→0)
ThomasFermi: T
TF
S
[n] = A
S
d
3
r n
5/3
(r) where A
s
= 2.871.
LSD: E
LDA
X
[n] = A
X
d
3
r n
4/3
(r) where A
X
= −(3/4)(3/π)
1/3
= −0.738.
E
LSD
C
[n
↑
, n
↓
] =
d
3
r n(r)
unif
C
(r
s
(r), ζ(r)
Gradient expansions
Gradient expansion: A[n] =
d
3
r
a(n(r)) +b(n(r)∇n(r)
2
. . .
¸
Gradient expansion approximation: A
GEA
[n] = A
LDA
[n] +∆A
GEA
[n]
Reduced density gradient: s(r) = ∇n(r)/(2k
F
(r)n(r)
Correlation gradient: t(r) = ∇n(r)/(2k
s
(r)n(r)
Polarization enhancement: φ(ζ) = ((1 + ζ)
2/3
+ (1 −ζ)
2/3
)/2
Kinetic energy: T
S
[n] = A
S
d
3
r n
5/3
1 + 5s
2
/27
or ∆T
GEA
S
[n] = T
VW
[n]/9.
Exchange energy: E
X
[n] = A
X
d
3
r n
4/3
1 + 10s
2
/81
Highdensity correlation energy: ∆E
GEA
C
= (2/3π
2
)
d
3
r n(r)φ(ζ(r))t
2
(r)
Generalized gradient approximation: A
GGA
[n] =
d
3
r a(n, ∇n)
Enhancement factor: E
GGA
XC
=
d
3
r e
unif
X
(n(r)) F
XC
(r
s
(r), s(r))
12 CONTENTS
Part I
Background
13
Chapter 1
Introduction
In which we introduce some of the basic concepts of modern density functional theory,
including the KohnSham description of a system, and give a simple but powerful example
of DFT at work.
1.1 Importance
1
Density functional theory (DFT) has long been the mainstay of electronic structure calcula
tions in solidstate physics. In the 19990’s it became very popular in quantum chemistry. This
is because approximate functionals were shown to provide a useful balance between accuracy
and computational cost. This allowed much larger systems to be treated than by traditional
ab initio methods, while retaining much of their accuracy. Nowadays, traditional wavefunc
tion methods, either variational or perturbative, can be applied to ﬁnd highly accurate results
on smaller systems, providing benchmarks for developing density functionals, which can then
be applied to much larger systems
But DFT is not just another way of solving the Schr¨ odinger equation. Nor is it simply a
method of parametrizing empirical results. Density functional theory is a completely diﬀerent,
formally rigorous, way of approaching any interacting problem, by mapping it exactly to a
much easiertosolve noninteracting problem. Its methodology is applied in a large variety of
ﬁelds to many diﬀerent problems, with the groundstate electronic structure problem simply
being the most common.
The aim of this book is to provide a relatively gentle, but nonetheless rigorous, introduction
to this subject. The technical level is no higher than any graduate quantum course, or many
advanced undergraduate courses, but leaps at the conceptual level are required. These are
almost as large as those in going from classical to quantum mechanics. In some sense, they
are more diﬃcult, as these leaps are usually made when we are more advanced (i.e., set) in
our thinking.
Students from all areas of modern computational science (chemistry, physics, materials
science, biochemistry, geophysics, etc.) are invited to work through the material. The only
1
c 2000 by Kieron Burke. All rights reserved.
15
16 CHAPTER 1. INTRODUCTION
necessary requirements are a good background in elementary quantum mechanics, no fear
of calculus of more than one variable, and a desire to learn. The student should end up
knowing what the KohnSham equations are, what functionals are, how much (or little) is
known of their exact properties, how they can be approximated, and how insight into all these
things produces understanding of the errors in electronic structure calculations. Preliminary
forms of my lecture notes have been used throughout the world during 5 years its taken to
complete this book. The hope is that these notes will be used by students worldwide to gain
a better understanding of this fundamental theory. I ask only that you send me an email (to
kieron@rutchem.rutgers.edu) if you use this material. In return, I am happy to grade problems
and answer questions for all who are interested. These notes are copyright of Kieron Burke.
No reproduction for purposes of sale is allowed.
The book is laid out in the form of an undergraduate text, and requires the working of
many exercises (although many of the answers are given). The idea is that the book and
exercises should be easy reading, but leave the reader with very clear concepts of modern
density functional theory. Throughout the text, there are exercises that must be performed
to get full value from the book. Also, at the end of each chapter, there are questions aimed
at making you think about the material. These should be thought about and answere as you
go along, but answers don’t need to be written out as explicitly as for the problems.
1.2 What is a KohnSham calculation?
To give an idea of what DFT is all about, and why it is so useful, we start with a very simple
example, the hydrogen molecule, H
2
. Throughout this book, we make the BornOppenheimer
approximation, in which we treat the heavy nuclei as ﬁxed points, and we want only to solve
the groundstate quantum mechanical problem for the electrons.
In regular quantum mechanics, we must solve the interacting Schr¨ odinger equation:
−
1
2
¸
i=1,2
∇
2
i
+
1
r
1
−r
2

+
¸
i=1,2
v
ext
(r
i
)
Ψ(r
1
, r
2
) = EΨ(r
1
, r
2
), (1.1)
where the index i runs over the two electrons, and the external potential, the potential
experienced by the electrons due to the nuclei, is
v
ext
(r) = −Z/r −Z/r −Rˆz, (1.2)
where Z = 1 is the charge on each nucleus, ˆz is a unit vector along the bond axis, and R is
a chosen internuclear separation. Except where noted, we use atomic units throughout this
text, so that
e
2
= ¯ h = m = 1, (1.3)
where e is the electronic charge, ¯ h is Planck’s constant, and m is the electronic mass. As
a consequence, all energies are in Hartrees (1 H = 27.2114 eV= 627.5 kcal/mol) and all
1.2. WHAT IS A KOHNSHAM CALCULATION? 17
distances are given in Bohr radii (a
o
= 0.529
˚
A). In this example, the electrons are in a
spin singlet, so that their spatial wavefunction Ψ(r
1
, r
2
) is symmetric under interchange of
r
1
and r
2
. Solution of Eq. (1.1) is complicated by the electrostatic repulsion between the
particles, which we denote as V
ee
. It couples the two coordinates together, making Eq. (1.1)
a complicated partial diﬀerential equation in 6 coordinates, and its exact solution can be quite
demanding. In Fig. 1.1 we plot the results of such a calculation for the total energy of the
molecule, E+1/R, the second term being the Coulomb repulsion of the nuclei. The position
of the minimum is the equilibrium bond length, while the depth of the minimum, minus the
zero point vibrational energy, is the bond energy. More generally, the global energy minimum
determines all the geometry of a molecule, or the lattice structure of a solid, as well as all the
vibrations and rotations. But for larger systems with N electrons, the wavefunction depends
on all 3N coordinates of those electrons.
150
100
50
0
50
0.4 0.6 0.8 1 1.2 1.4 1.6
exact
HF
LDA
GGA
E
(
H
2
)
+
1
/
R

E
(
2
H
)
(
k
c
a
l
/
m
o
l
)
R(
˚
A)
Figure 1.1: The total energy of the H2 molecule as a function of internuclear separation.
We note at this point that, with an exact groundstate wavefunction, it is easy to calculate
the probability density of the system:
n(r) = 2
d
3
r
Ψ(r, r
)
2
. (1.4)
The probability density tells you that the probability of ﬁnding an electron in d
3
r around r is
n(r)d
3
r. For our H
2
molecule at equilibrium, this would look like the familiar two decaying
exponentials centered over the nuceli, with an enhancement in between, where the chemical
bond has formed.
Next, imagine a system of two noninteracting electrons in some potential, v
S
(r), chosen
somehow to mimic the true electronic system. Because the electrons are noninteracting,
their coordinates decouple, and their wavefunction is a simple product of oneelectron wave
functions, called orbitals, satisfying:
−
1
2
∇
2
+v
S
(r)
¸
φ
i
(r) =
i
φ
i
(r), (1.5)
18 CHAPTER 1. INTRODUCTION
where Φ(r
1
, r
2
) = φ
0
(r
1
)φ
0
(r
2
). This is a much simpler set of equations to solve, since it
only has 3 coordinates. Even with many electrons, say N, one would still need to solve only
a 3D equation, and then occupy the ﬁrst N/2 levels, as opposed to solving a 3Ncoordinate
Schr¨ odinger equation. If we can get our noninteracting system to accurately ’mimic’ the
true system, then we will have a computationally much more tractable problem to solve.
How do we get this mimicking? Traditionally, if we think of approximating the true wave
function by a noninteracting product of orbitals, and then minimize the energy, we ﬁnd the
HartreeFock equations, which yield an eﬀective potential:
2
v
HF
S
(r) = v
ext
(r) +
1
2
d
3
r
n(r
)
r −r

. (1.6)
The correction to the external potential mimics the eﬀect of the second electron, in particular
screening the nuclei. For example, at large distances from the molecule, this potential decays
as −1/r, reﬂecting an eﬀective charge of Z − 1. Note that insertion of this potential into
Eq. (1.5) now yields a potential that depends on the electronic density, which in turn is
calculated from the solution to the equation. This is termed therefore a selfconsistent set
of equations. An initial guess might be made for the potential, the eigenvalue problem is
then solved, the density calculated, and a new potential found. These steps are repeated
until there is no change in the output from one cycle to the next – selfconsistency has been
reached. Such a set of equations are often called selfconsistent ﬁeld (SCF) equations. In
Fig. 1.1, we plot the HartreeFock result and ﬁnd that, although its minimum position is very
accurate, it underbinds the molecule signiﬁcantly. This has been a wellknown deﬁciency of
this method, and traditional methods attempt to improve the wavefunction to get a better
energy. The missing piece of energy is called the correlation energy.
In a KohnSham calculation, the basic steps are very much the same, but the logic is
entirely diﬀerent. Imagine a pair of noninteracting electrons which have precisely the same
density n(r) as the physical system. This is the KohnSham system, and using density
functional methods, one can derive its potential v
S
(r) if one knows how the total energy
E depends on the density. A single simple approximation for the unknown dependence of
the energy on the density can be applied to all electronic systems, and predicts both the
energy and the selfconsistent potential for the ﬁctitious noninteracting electrons. In this
view, the KohnSham wavefunction of orbitals is not considered an approximation to the
exact wavefunction. Rather it is a preciselydeﬁned property of any electronic system, which
is determined uniquely by the density. To emphasize this point, consider our H
2
example in
the united atom limit, i.e., He. In Fig. 1.2, a highly accurate manybody wavefunction for
the He atom was calculated, and the density extracted. In the bottom of the ﬁgure, we plot
both the physical external potential, −2/r, and the exact KohnSham potential.
3
Two non
interacting electrons sitting in this potential have precisely the same density as the interacting
2
For the wellinformed, we note that the correction to the external potential consists of the Hartree potential, which is double that shown above,
less the exchange potential, which in this case cancels exactly half the Hartree.
3
It is a simple exercise to extract the exact KohnSham potential from the exact density, as in section ??.
1.3. REINTERPRETING MOLECULAR ORBITALS 19
8
6
4
2
0
2
4
0 0.2 0.4 0.6 0.8 1
He atom
density
2/r
KohnSham potential
r
Figure 1.2: The external and KohnSham potentials for the He atom, thanks to Cyrus’ Umrigar’s very accurate density.
electrons. If we can ﬁgure out some way to approximate this potential accurately, we have
a much less demanding set of equations to solve than those of the true system. Thus we
are always trying to improve a noninteracting calculation of a noninteracting wavefunction,
rather than that of the full physical system. In Fig. 1.1 there are also plotted the local
density approximation (LDA) and generalized gradient approximation (GGA) curves. LDA
is the simplest possible density functional approximation, and it already greatly improves on
HF, although it typically overbinds by about 1/20 of a Hartree (or 1 eV or 30 kcal/mol),
which is too inaccurate for most quantum chemical purposes, but suﬃciently reliable for many
solidstate calculations. More sophisticated GGA’s (and hybrids) reduce the typical error in
LDA by about a factor of 5 (or more), making DFT a very useful tool in quantum chemistry.
In section 1.5, we show how density functionals work with a simple example from elementary
quantum mechanics.
1.3 Reinterpreting molecular orbitals
The process of bonding between molecules is shown in introductory chemistry textbooks as
linear combinations of atomic orbitals forming molecular orbitals of lower energy, as in Fig.
1.3.
But later, in studying computational chemistry, we discover this is only the HartreeFock
picture, which, as stated above, is rarely accurate enough for quantum chemical calculations.
In this picture, we need a more accurate wavefunction, but then lose this simple picture of
chemical bonding. This is a paradox, as chemical reactivity is usually thought of in terms of
frontier orbitals.
In the KohnSham approach, the orbitals are exact and unique, i.e., there exists (at most)
one external potential that, when doubly occupied by two noninteracting electrons, yields
the exact density of the H
2
molecule. So in this view, molecular orbital pictures retain their
signiﬁcance, if they are the exact KohnSham orbitals, rather than those of HartreeFock.
20 CHAPTER 1. INTRODUCTION
1s
1s
1s
!
!
Figure 1.3: Orbital diagram of H2 bond formation.
And a highly accurate approximate density functional calculation produces the full electronic
energy from these orbitals, resolving the paradox.
1.4 Sampler of applications
In this section, we highlight a few recent applications of modern DFT, to give the reader a
feeling for what kinds of systems can be tackled.
Example from biochemistry, using ONIOM.
Example from solidstate, eg ferromagnetism.
Example using CP molecular dynamics, eg phasediagram of C.
Example from photochemistry, using TDDFT.
1.5 Particle in a box
In this section, we take the simplest example from elementary quantum mechanics, and apply
density functional theory to it. This provides the underlying concepts behind what is going
on in the more previous sections of this chapter. Much of the rest of the book is spent
connecting these two.
In general, we write our Hamiltonian as
ˆ
H =
ˆ
T +
ˆ
V (1.7)
where
ˆ
T denotes the operator for the kinetic energy and
ˆ
V the potential energy. We begin
with the simplest possible case. The Hamiltonian for a 1D 1electron system can be written
as
ˆ
H = −
1
2
d
2
dx
2
+V (x) (1.8)
1.5. PARTICLE IN A BOX 21
The timeindependent Schr¨ odinger equation has solutions:
ˆ
Hφ
i
= ε
i
φ
i
, i = 0, 1, ... (1.9)
Thus ε
0
denotes the ground state energy and φ
0
the ground state wave function. Because
the operator
ˆ
H is hermitian the eigenstates can be chosen orthonormal:
∞
−∞
dx φ
∗
i
(x)φ
j
(x) = δ
ij
(1.10)
Thus each eigenfunction is normalized. The electron probability density is just n(x) =
φ
0
(x)
2
.
Exercise 1 Particle in a onedimensional box
Consider the elementary example of a particle in a box, i.e., V (x) = ∞ everywhere, except
0 ≤ x ≤ L, where V = 0. This problem is given in just about all elementary textbooks on
quantum mechanics. Show that the solution consists of trigonometric functions (standing
waves), and making them vanish at the boundaries quantizes the energy, i.e.,
φ
i
=
2
L
sin(k
i
x), i = 1, 2, ... (1.11)
where k
i
= πi/L, and the energies are
i
= k
2
i
/2. Check the orthonormality condition, Eq.
(1.10).
Next we consider the following approximate density functional for the kinetic energy of
noninteracting electrons in onedimension. Don’t worry if you never heard of a functional
before. In fact, chapter ?? discusses functionals in some detail. For now, we need only the
fact that a functional is a rule which maps a function onto a number. We use square brackets
[..] to indicate a functional dependence. The functional we write down will be the local
approximation to the kinetic energy of noninteracting spinless fermions in one dimension.
We do not derive it here, but present it for calculational use, and derive it later in the chapter:
T
loc
S
[n] = 1.645
∞
−∞
dx n
3
(x). (1.12)
A local functional is one which is a simple integral over a function of its argument. Now,
since for the particle in a box, all its energy is kinetic, we simply estimate the energy using
T
loc
S
[n]. Since the electron density is the square of the orbital:
n
1
(x) =
2
L
sin
2
(k
1
x), (1.13)
we ﬁnd T
loc
S
= 4.11/L
2
, a 17% underestimate of the true value, π
2
/2L
2
= 4.93/L
2
. What
is so great about that? There is not much more work in ﬁnding the exact energy (just take
the second derivative of φ
1
, and divide by φ
1
) as there is in evaluating the integral.
But watch what happens if there are two particles, fermions of the same spin (the general
case for which T
loc
S
was designed). We ﬁnd now E =
1
+
2
= 5π
2
/2L
2
= 24.7/L
2
. If we
evaluate the approximate kinetic energy again, using n(x) = n
1
(x) +
2
L
sin
2
(k
2
x), we ﬁnd
T
loc
S
= 21.8/L
2
, an 11% underestimate.
22 CHAPTER 1. INTRODUCTION
Exercise 2 Three particles in a box
Calculate the exact energy of three identical fermions in a box of length L. Plot the total
density for one, two, and three particles. Compare the exact energy with that from the local
approximation. What happens as the number of electrons grows? Answer: E = 7π
2
/L
2
,
E
loc
= 63/L
2
, a 9% underestimate.
These results are gotten by the evaluation of a simple integral over the density, whereas
the exact numbers require solving for all the eigenvalues of the box, up to N, the number of
electrons in it. Thus approximate density functionals produce reliable but inexact results at
a fraction of the usual cost.
To see how remarkable this is, consider another paradigm of textbook quantum mechanics,
namely the onedimensional harmonic oscillator. In the previous example, the particles were
sine waves; in this one, they are Gaussians, i.e., of a completely diﬀerent shape. Applying
the same approximate functional, we ﬁnd merely a 20% overestimate for one particle:
Exercise 3 Kinetic energy of 1d harmonic oscillator
What is the groundstate energy of a 1d harmonic oscillator? What is its kinetic energy?
Calculate its density, and from it extract the local density approximation to its kinetic energy.
Repeat for two samespin electrons occupying the lowest two levels of the well.
Congratulations. You have performed your ﬁrst elementary density functional calculations,
and are well on the way to becoming an expert!
0
2
4
6
8
10
12
14
0 0.2 0.4 0.6 0.8 1
10 particles in a box
Figure 1.4: Density of 10 particles in a box.
In fact, there’s a simple way to see how this magic has been performed. Imagine the box
with a large number of particles. In Fig. 1.4, we have plotted the case for N = 10. As the
number of particles grows, the density becomes more and more uniform (independent of the
shape of the box!). If we assume it actually becomes uniform as N → ∞, with errors of
order 1/N or smaller, this gives us the constant in the local approximation. To see this, note
1.6. QUESTIONS 23
that the exact energy is
E =
π
2
2L
2
N
¸
j=1
j
2
=
π
2
2L
2
N(N + 1)(2N + 1)/6 (1.14)
whose leading contribution is simply π
2
N
3
/6. Since the density has become uniform (to this
order), the local approximation becomes exact, yielding the constant.
Later in the book, we will see how, if we had an extremely accurate noninteracting kinetic
energy density functional in three dimensions, we could revolutionize electronic structure
calculations, by making them very fast. This is because we would then have a method for
calculating the density and groundstate energy of systems which involved solving a single
selfconsistent integrodiﬀerential equation for any electronic system. This would be a true
density functional calculation. Most present calculations do not do this. By solving an
eﬀective singleparticle equation for the orbitals (the KohnSham equation), they ﬁnd T
S
exactly, but at large computational cost (for large systems). So now you even know an
unsolved problem in density functional theory.
Now answer the following simple questions as well as you can, justifying your answers in
each case.
1.6 Questions
We begin with conceptual questions, that are designed to test your understanding of the
material presented in the chapter.
1. Why is a KohnSham calculation much faster than a traditional wavefunction calculation?
2. If you evaluate the kinetic energy of a KohnSham system, is it equal to the physical
kinetic energy?
3. Repeat above question for 1/r, which can be measured in scattering experiments.
4. Why is the density cubed in the local approximation for T
S
? (see section ?? for the
answer).
We continue with more creative questions, designed to make you think about what is in
the chapter.
5. Suppose you’d been told that T
loc
S
is proportional to
dxn
3
(x), but were not told the
constant of proportionality. If you can do any calculation you like, what procedure might
you use to determine the constant in T
loc
S
? (see section 11.5 for the answer).
6. In what way will T
loc
S
change if spin is included, e.g., for two electrons of opposite spin
in a box? (see section 9.2 for the answer).
7. If we add an inﬁnitesimal to n(x) at a point, i.e., δ(x) as →0, how does T
loc
S
change?
(see section 2.2 for the answer).
24 CHAPTER 1. INTRODUCTION
8. The simplest density functional approximation is a local one. What form might a cor
rection to T
loc
S
take? (see section 17.2 for the answer).
9. In the chapter and exercises, T
loc
S
does very well. But there are cases where it fails quite
badly. Can you ﬁnd one, and say why?
10. Consider a single electron in an excited state, with density n
∗
(x). Will T
loc
S
[n
∗
] provide
an accurate estimate of its kinetic energy?
1.6. QUESTIONS 25
Extra exercises
1. For a δfunction potential, v(x) = −Zδ(x), the groundstate kinetic energy is Z
2
/2 and
the density is Z exp(−2Zx). Before calculating the local approximation, ﬁrst guess if it
will be an overestimate or an underestimate, and by how much.
2. Another famous example is v(x) = −a
2
sech
2
(ax), with groundstate energy −a
2
/2 and
density a sech
2
(ax)/2. Repeat above problem.
26 CHAPTER 1. INTRODUCTION
Chapter 2
Functionals
1
In this chapter, we introduce in a more systematic fashion what exactly a functional
is, and how to perform elementary operations on functionals. This mathematics will
be needed when we describe the quantum mechanics of interacting electrons in terms of
density functional theory.
2.1 What is a functional?
A function maps one number to another. A functional assigns a number to a function. For
example, consider all functions r(θ), 0 ≤ θ ≤ 2π, which are periodic, i.e., r(θ + 2π) = r(θ).
Such functions describe shapes in twodimensions of curves which do not “doubleback” on
themselves, such as in Fig. 2.1. For every such curve, we can deﬁne the perimeter P as the
Figure 2.1: A 2D curve which is generated by a function r = r(θ)
length of the curve, and the area A as the area enclosed by it. These are functionals of r(θ),
in the sense that, for a given curve, such as the ellipse
r(θ) = 1/
sin
2
(θ) + 4 cos
2
(θ) (2.1)
there is a single welldeﬁned value of P and of A. We write P[r] and A[r] to indicate this
functional dependence. Note that, even if we don’t know the relation explicitly, we do know
it exists: Every bounded curve has a perimeter and an area.
For this simple example, we can use elementary trigonometry to deduce explicit formulas
for these functionals. Consider the contribution from an inﬁnitesimal change in angle dθ, and
1
c 2000 by Kieron Burke. All rights reserved.
27
28 CHAPTER 2. FUNCTIONALS
integrate over the entire range of angles. The contribution to the area is that of a triangle
of base r cos(dθ) ≈ r and height r(θ) sin(dθ) ≈ rdθ, yielding
A[r] =
1
2
rdθr =
1
2
2π
0
dθ r
2
(θ) (2.2)
For an abitrary r(θ) the integral above can be evaluated to give the area that is enclosed by
the curve. Thus this functional maps a real function of one argument to a number, in this
case the area enclosed by the curve. Similarly, the inﬁnitesimal change in the perimeter is
just the line segment in polar coordinates:
δP
2
= dr
2
+r
2
(θ)dθ. (2.3)
Now, since r = r(θ), we can write dr = dθ (dr/dθ), yielding
P[r] =
2π
0
dθ
r
2
(θ) + (dr/dθ)
2
. (2.4)
In both cases, once we know r(θ), we can calculate the functional.
The area is a local functional of r(θ), since it can be written in the form:
A[r] =
2π
0
dθ f(r(θ)), (2.5)
where f(r) is a function of r. For the area, f = r
2
/2. These are called local because,
inside the integral, one needs only to know the function right at a single point to evaluate
the contribution to the functional from that point. On the other hand, P is a semilocal
functional of the radius, as it depends not only on r, but dr/dθ.
We have already come across a few examples of density functionals. In the opening
chapter, the local approximation to the kinetic energy is a local functional of the density,
with f(n) = 1.645n
3
. On the other hand, for any oneelectron system, the exact kinetic
energy is a semilocal functional, called the von Weisacker functional:
T
VW
S
[n] =
1
8
∞
−∞
dx
n
2
n
, (2.6)
where n
(x) = dn/dx. Later we will see some examples of fully nonlocal functionals.
2.2 Functional derivatives
When we show that the groundstate energy of a quantum mechanical system is a functional
of the density, we will then want to minimize that energy to ﬁnd the true groundstate density.
To do this, we must learn how to diﬀerentiate functionals, in much the same way as we learn
how to diﬀerentiate regular functions in elementary calculus.
To begin with, we must deﬁne a functional derivative. Imagine making a tiny increase in
a function, localized to one point, and asking how the value of a functional has changed due
2.2. FUNCTIONAL DERIVATIVES 29
to this increase, i.e., we add an inﬁnitesimal change δr(θ) = δ(θ − θ
0
).
2
How does A[r]
change?
A[r +δr] −A[r] =
1
2
2π
0
dθ
(r + δ(θ −θ
0
))
2
− r
2
¸
=
2π
0
dθ rδ(θ −θ
0
) = r(θ
0
) (2.7)
The functional derivative, denoted δA/δr(θ) is just the change in A divided by , or just r(θ)
in this case. Since this is linear in the change in r, the general deﬁnition, for any inﬁnitesimal
change in r, is just
A[r + δr] −A[r] =
2π
0
dθ
δA
δr(θ)
δr(θ), (2.8)
where the functional derivative δA/δr(θ) is that function of θ which makes this formula
exact for any small change in r(θ). Just as the usual derivative df/dx of a function f(x)
tells you how much f changes when x changes by a small amount, i.e., f(x +dx) −f(x) =
(df/dx) dx + O(dx
2
), so does the functional derivative, a function, tell you how much a
functional changes when its arguments changes by a small “amount” (in this case, a small
function). Thus
v
A
[r](θ) ≡
δA[r]
δr(θ)
= r(θ), (2.9)
where the notation v
A
[r](θ) emphasizes that the functional derivative of A[r] is a θdependent
functional.
Exercise 4 Derivative of a local functional
Show that, for any local functional A[n] =
dx a(n(x)), the functional derivative is δA/δn(x) =
a
(n(x)), where a
(n) = da/dn.
In general, the way to ﬁnd a functional derivative is to evaluate the expression A[r+δr]−A[r]
to leading order in δr, and the resulting integral must be cast in the form of a function times
δr. Often it is necessary to do an integration by parts to identify the functional derivative.
For the perimeter, such complications arise. We have
P[r + δr] =
2π
0
dθ
(r + δr)
2
+ (r
+ δr
)
2
(2.10)
where r
= dr/dθ and all functions are assumed to have argument θ. To ﬁrst order in δr, we
ﬁnd
P[r + δr] =
2π
0
dθ
√
r
2
+r
2
¸
1 +
rδr +r
δr
r
2
+r
2
¸
= P[r] +
2π
0
dθ
¸
r
√
r
2
+r
2
δr +
r
√
r
2
+r
2
δr
¸
(2.11)
2
For purists, we take a Gaussian of tiny but ﬁnite width, so that as →0, the maximum change becomes arbitrarily small. Then we take the
width to zero!
30 CHAPTER 2. FUNCTIONALS
The ﬁrst integral is in just the right form for identifying the contribution to the functional
derivative, but the second is not. But we can perform an integration by parts to ﬁnd
δP
δr(θ)
= p[r](θ) =
r
√
r
2
+r
2
−
d
dθ
¸
r
√
r
2
+r
2
¸
(2.12)
Note that the endpoint term of the integrationbyparts vanishes, because r(θ) is periodic.
A little calculus and algebra ﬁnally yields
p[r](θ) =
r
3
+ 2rr
2
−r
2
r
)
(r
2
+r
2
)
3/2
(2.13)
Exercise 5 Derivative of a semilocal functional
Show that the functional derivative of a semilocal functional B[n] =
∞
−∞
dx b(n, dn/dx),
assuming n and its derivatives vanish rapidly as x →∞, is
v
B
[n](x) =
∂b
∂n
−
d
dx
∂b
∂n
(2.14)
where n
(x) = dn/dx. Use your answer to show
δT
VW
S
δn(x)
= −
n
4n
+
n
2
8n
2
. (2.15)
The density functionals we use in practice are threedimensional:
Exercise 6 Hartree potential
The Hartree (or classical electrostatic) energy of a charge distribution interacting with itself
via Coulomb’s law is given by
U[n] =
1
2
d
3
r
d
3
r
n(r) n(r
)
r −r

. (2.16)
Show that its functional derivative, the Hartree potential, is
v
H
[n](r) =
d
3
r
n(r
)
r −r

. (2.17)
2.3 EulerLagrange equations
The last piece of functional technology we need for now is how to solve a constrained opti
mization problem, i.e., how to maximize (or minimize) one functional, subject to a constraint
imposed by another.
For example, we might want to know what is the maximum area we can enclose inside a
loop of string of ﬁxed length l. Thus we need to maximize A[r], subject to the constraint
that P[r] = l. There is a wellknown method for doing such problems, called the method
2.4. QUESTIONS 31
of Lagrange multipliers. Following the example in the Appendix (A.1), construct a new
functional
B[r] = A[r] −µP[r], (2.18)
where µ is at this point an unknown constant. Then extremize the new functional B[r], by
setting its functional derivative to zero:
δB
δr
=
δA
δr
−µ
δP
δr
= r(θ) −µ
r
3
+r
2
(2r −r
)
(r
2
+r
2
)
3/2
= 0, (2.19)
A simple solution to this equation is r(θ) = µ, a constant. This tells us that the optimum
shape is a circle, and the radius of that circle can be found by inserting the solution in the
constraint, P[r = µ] = 2πµ = l. The largest area enclosable by a piece of string of length l
is A[r = µ] = l
2
/(4π).
Exercise 7 Change in a functional
Suppose you know that a functional E[n] has functional derivative v(r) = −1/r for the
density n(r) = Z exp(−2Zr), where Z = 1. Estimate the change in E when Z becomes
1.1.
Exercise 8 Second functional derivative
Find the second functional derivative of (a) a local functional and (b) T
VW
S
.
2.4 Questions
1. Compare the functional derivative of T
VW
S
[n] with T
loc
S
[n] for some sample oneelectron
problem. Comment.
2. If someone just tells you a number for any density you give them, e.g., the someone might
be Mother Nature, and the number might be the total energy measured by experiment,
devise a method for deducing if Mother Nature’s functional is local or not.
3. Is there a simple relationship between T
S
and
dx n(x)δT
S
/δn(x)? First consider the
local approximation, then the Von Weisacker. Comment on your result.
4. For ﬁxed particle number, is there any indeterminancy in the functional derivative of a
density functional?
32 CHAPTER 2. FUNCTIONALS
Extra exercises
1. Consider a square of side 1. Check that the circle of the same perimeter has a larger
area.
2. For the unit square centered on the origin, check that Eqs. (2.2) and (2.4). are correct.
Hint: You need only do the ﬁrst octant, x > y > 0.
Chapter 3
One electron
1
In this chapter, we review the traditional wavefunction picture of Schr¨odinger for one
particle, but introducing our own speciﬁc notation.
3.1 Variational principle
We start oﬀ with the simplest possible case, one electron in one dimension. Recall, from
basic quantum mechanics, the RayleighRitz variational principle:
E = min
φ
φ
ˆ
H  φ,
∞
−∞
dx φ(x)
2
= 1. (3.1)
We will use this basic, extremely powerful principle throughout this book. This principle says
that we can use any normalized wavefunction to calculate the expectation value of the energy
for our problem, and we are guaranteed to get an energy above the true groundstate energy.
For example, we can use the same wavefunction for every 1d oneelectron problem, and get
an upper bound on the groundstate energy.
Having learnt about functionals in Chapter ??, we may write this principle in the following
useful functional form. For example, the kinetic energy of the particle can be considered as
a functional of the wavefunction:
T[φ] =
∞
−∞
dx φ
∗
(x)
¸
−
1
2
d
2
dx
2
¸
φ(x) =
1
2
∞
−∞
dx φ
(x)
2
(3.2)
where the second form is gotten by integration by parts, and the prime denotes a spatial
derivative. (This second form is much handier for many calculations, as you only need take
one derivative.) Thus, for any given normalized wavefunction, there is a single number T,
which can be calculated from it, via Eq. (3.2). Similarly, the potential energy is a very simple
functional:
V [φ] =
∞
−∞
dx V (x) φ(x)
2
. (3.3)
The kinetic energy of a given wavefunction is always the same, no matter what the prob
lem we are applying that wavefunction to. The kinetic energy is a universal functional,
1
c 2000 by Kieron Burke. All rights reserved.
33
34 CHAPTER 3. ONE ELECTRON
independent of the particular problem, i.e., we apply the same operation on a given trial
wavefunction, no matter what our physical problem is. But the potential energy functional
diﬀers in each problem; it is not universal. This is analogous to trying to ﬁnd the minimum
over x of f(x, y) = x
2
/2 − xy. The ﬁrst term is independent of y, so we could make a list
of it value for every x, and then use that same list to ﬁnd the minimum for any y, without
having to evaluate it again for each value of y.
Now our variational principle can be written by writing
E = min
φ
{T[φ] +V [φ]},
∞
−∞
dx φ(x)
2
= 1 (3.4)
We need simply evaluate the energy of all possible normalized wavefunctions, and choose the
lowest one.
3.2 Trial wavefunctions
Next, we put trial wavefunctions into the variational principle to ﬁnd upper bounds on the
groundstate energy of a system.
Exercise 9 Gaussian trial wavefunction
An interesting potential in onedimension is just V (x) = −δ(x), where δ is the Dirac delta
function. Using a Gaussian wavefunction, φ
G
(x) = π
−1/4
exp(−x
2
/2), as a trial wavefunc
tion, make an estimate for the energy, and a rigorous statement about the exact energy.
We can improve on our previous answers in two distinct ways. The ﬁrst method is to
include an adjustable parameter, e.g., the length scale, into our trial wavefunction.
Exercise 10 Adjustable trial wavefunction
Repeat Ex. (9) with a trial wavefunction φ
G
(αx) (don’t forget to renormalize). Find T (α)
and V (α) and plot their sum as a function of α. Which value of α is the best, and what
is your estimate of the groundstate energy? Compare with previous calculation.
Note that this has led to a nonlinear optimization problem in α. Varying exponents in
trial wavefunctions leads to diﬃcult optimization problems.
To see how well we did for this problem, the next exercise yields the exact answer.
Exercise 11 1d Hydrogen atom
Use φ
E
(αx), where φ
E
(x) = exp(−x) as a trial wavefunction for the problem in Ex. (9).
Find the lowest energy. This is the exact answer. Calculate the errors made in the previous
exercises, and comment.
An alternative, simpler approach to improving our trial wavefunction is to make a linear
combination with other trial wavefunctions, but not vary the length scale. Thus we write
φ
trial
(x) = c
0
φ
0
(x) +c
1
φ
1
(x), (3.5)
3.3. THREE DIMENSIONS 35
and minimize the expectation value of
ˆ
H, subject to the constraint that φ be normalized.
This leads to a set of linear equations
H−ES = 0, (3.6)
where
H
ij
=
∞
−∞
dx φ
∗
i
(x)
ˆ
Hφ
j
(x) (3.7)
is the Hamiltonian matrix, and
S
ij
=
∞
−∞
dx φ
∗
i
(x)φ
j
(x) (3.8)
is the overlap matrix. In general, Eq. (3.6) is a generalized eigenvalue equation. Only if the
basis of trial wavefunctions is chosen as orthonormal (Eq. 1.10) will S = 1, and the equation
be an ordinary eigenvalue equation.
The standard textbook problem of chemical bonding is the formation of molecular orbitals
from atomic orbitals in describing the molecular ion H
+
2
.
Exercise 12 1d H
+
2
For a potential V (x) = −δ(x) − δ(x + a), use φ
mol
(x) = c
1
φ
E
(x) + c
2
φ
E
(x + a) as a trial
wavefunction, and calculate the bonding and antibonding energies as a function of atomic
separation a.
3.3 Three dimensions
There are only slight complications in the real threedimensional world. We will consider
only spherical 3d potentials, v(r), where r = r. Then the spatial parts of the orbitals are
characterized by 3 quantum numbers:
φ
nlm
(r) = R
nl
(r)Y
lm
(θ, ϕ), (3.9)
where the Y
lm
’s are spherical harmonics. Inserting (3.9) into the Schr¨ odinger Equation gives
the radial equation
−
1
2r
2
d
dr
r
2
d
dr
+
l(l + 1)
2r
2
+v(r)
R
nl
(r) = ε
nl
R
nl
(r). (3.10)
These are written about in almost all textbooks on quantum mechanics, with emphasis on the
Coulomb problem, v(r) = −Z/r. Less frequently treated is the threedimensional harmonic
oscillator problem, v(r) = kr
2
/2.
Exercise 13 Three dimensions
Use the trial wavefunctions of an exponential and a Gaussian in r to deduce the groundstate
energy and orbital of (a) the hydrogen atom (Z = 1), and (b) the harmonic oscillator with
k = 1.
36 CHAPTER 3. ONE ELECTRON
3.4 Diﬀerential equations
We have just seen how to construct various trial wavefunctions and ﬁnd their parameters by
minimizing the energy. In a few lucky cases, we’ve found the exact solution to the given
problem.
But in the real world, where analytic solutions are few and far between, we cannot be
relying on lucky guesses. When we’re given a variational problem to solve, we must ﬁnd the
exact answer, and be sure we’ve done so.
Happily, we can do so, by applying the techniques of Chapter 2 to the variational problem of
Eq. (3.1). We choose the wavefunction real, and want to minimize E[φ], given the constraint
that the wavefunction be normalized. Thus we must minimize:
F[φ] = T[φ] +V [φ] −µ(N[φ] −1) (3.11)
where N[φ] =
dx φ
2
(x), and µ is a Lagrange multiplier. Taking functional derivatives, we
ﬁnd:
δT
δφ(x)
= −φ
(x),
δV
δφ(x)
= 2v(x) φ(x),
δN
δφ(x)
= 2φ(x) (3.12)
Thus
δF
δφ(x)
= −φ
(x) + 2(v(x) −µ)φ(x) (3.13)
At the minimum, this should vanish, yielding:
−
1
2
φ
(x) +v(x) φ(x) = µφ(x) (3.14)
and we can identify µ = E
0
by integrating both sides with φ(x).
This should all make a lot of sense. From the RayleighRitz variational principle, we have
rededuced the Schr¨ odinger equation. Note however that all eigenstates satisfy the Schr¨ odinger
equation, but only the ground state satisﬁes the variatonal principle (all others are only local
minima).
The general rule is that, given a variational principle, trial wavefunctions can be used to
get approximate answers, but the exact solution can often be found by solving the diﬀerential
equation that results from extremizing the functional, given the constraints.
3.5 Virial theorem
We close this chapter with a statement, without proof, of the virial theorem in one dimension.
The proof will be given later in the book. The virial theorem says that, for an eigenstate of
a given potential,
2T = x
dV
dx
. (3.15)
3.6. QUESTIONS 37
In the special case in which v(x) = Ax
p
, x
dV
dx
= p V , so that
2T = pV (3.16)
This can save a lot of time when doing calculations, because one only needs either the kinetic
or the potential energy to deduce the total energy. Note that, for these purposes, the δfn
acts as if it had power p = −1.
The theorem is also true for threedimension problems, and also true for interacting prob
lems, once the potential is interpreted as all the potential energy in the problem. Thus, for
any atom, p = −1, and E = −T = −V/2, where V is the sum of the external potential and
the electronelectron repulsion.
Exercise 14 Virial theorem for one electron
Check Eq. (3.16) for the ground state of (a) the 1d H atom, (b) the 1d harmonic oscillator,
(c) the threedimensional Hydrogen atom. Also check it for Ex. (9) and (10).
Some approximations automatically satisfy the virial theorem, others do not. In cases that
do not, we can use the virial theorem to judge how accurate the solution is. In cases that do,
we can use the virial theorem to check we dont have an error, or to avoid some work.
3.6 Questions
All the questions below are conceptual.
1. Suggest a good trial wavefunction for a potential that consists of a negative delta function
in the middle of a box of width L.
2. Sketch, on the same ﬁgure, the 1d Hatom for Z = 1/2, 1, and 2. What happens as
Z →∞ and as Z →0?
3. What is the exact kinetic energy density functional for one electron in onedimension?
4. How would you check to see if your trial wavefunction is the exact groundstate wave
function for your problem?
5. Consider what happens to the potential v(x) = Ax
p
as p gets large. Can the virial
theorem be applied to the particle in the box (even if it yields trivial answers)?
38 CHAPTER 3. ONE ELECTRON
Chapter 4
Two electrons
1
In this chapter, we review the traditional wavefunction picture of Schr¨odinger for two
electrons. We keep everything as elementary as possible, avoiding sophistry such as the
interaction picture, second quantization, Matsubara Green’s functions, etc. These are all
valuable tools for studying advanced quantum mechanics, but are unneccessary for the
basic logic of our presentation.
4.1 Antisymmetry and spin
When there is more than one particle in the system, the Hamiltonian and the wavefunction
includes one coordinate for each particle, as in Eqn. (1.1) for the H
2
molecule. Furthermore,
electrons have two possible spin states, up or down, and so the wavefunction is a function both
of spatial coordinates and spin coordinates. A general principle of manyelectron quantum
mechanics is that the wavefunction must be antisymmetric under interchange of any two sets
of coordinates. All this is covered in elementary textbooks. For our purposes, this implies the
2electron wavefunction satisﬁes:
Ψ(r
2
, σ
2
, r
1
, σ
1
) = −Ψ(r
1
, σ
1
, r
2
, σ
2
). (4.1)
The groundstates of our systems will be spinsinglets, meaning the spinpart of their wave
function will be antisymmetric, while their spatial part is symmetric:
Ψ(r
1
σ
1
, r
2
σ
2
) = Ψ(r
1
, r
2
) χ
Singlet
(σ
1
, σ
2
), (4.2)
where the spatial part is symmetric under exchange of spatial coordinates and the spin part
χ
Singlet
(σ
1
, σ
2
) =
1
√
2
(  ↑↓ −  ↓↑ ) is antisymmetric.
4.2 HartreeFock
For more than one electron, the operators in the Hamiltonian depend on all the coordinates,
but in a simple way. Both the kinetic energy and the external potential are onebody operators,
1
c 2000 by Kieron Burke. All rights reserved.
39
40 CHAPTER 4. TWO ELECTRONS
meaning that they are sums of terms which each depend on only one coordinate at a time:
ˆ
T = −
1
2
¸
d
2
dx
2
1
+
d
2
dx
2
2
¸
ˆ
V
ext
= v
ext
(x
1
) +v
ext
(x
2
) (4.3)
The electronelectron repulsion operator is a twobody operator, each term depending on
two coordinates simultaneously. Note that it is this term that complicates the problem,
by coupling the two coordinates together. The interaction between two electrons in three
dimensions is Coulombic, i.e., 1/r − r
. This is homogeneous of degree 1 in coordinate
scaling. However, 1/x − x
 in onedimension is an exceedingly strong attraction, and we
prefer to use δ(x − x
), which has the same scaling property, but is much weaker and more
shortranged.
We will now consider the problem of onedimensional He as a prototype for two electron
problems in general. The Hamiltionian is then:
ˆ
H = −
1
2
¸
d
2
dx
2
1
+
d
2
dx
2
2
¸
−Zδ(x
1
) −Zδ(x
2
) + δ(x
1
−x
2
). (4.4)
To ﬁnd an exact solution to this, we might want to solve the Schr¨ odinger equation with
this Hamiltonian, but to ﬁnd approximate solutions, its much easier to use the variational
principle:
E = min
Ψ
Ψ
ˆ
H  Ψ,
dx
1
dx
2
Ψ(x
1
, x
2
)
2
= 1. (4.5)
The most naive approach might be to ignore the electronelectron repulsion altogether.
Then the diﬀerential equation decouples, and we can write:
Ψ(x
1
, x
2
) = φ(x
1
)φ(x
2
) (ignoring V
ee
) (4.6)
where both orbitals are the same, and satisfy the oneelectron problem. We know from
chapter 3 that this yields
φ(x) =
√
Z exp(−Zx). (4.7)
This decoupling of the coordinates makes it possible to handle very large systems, since we
need only solve for one electron at a time. However, we have made a very crude approximation
to do this. The contribution of the kinetic energy and the potential energy is then just twice
as big as in the single electron system:
T
s
= 2
Z
2
/2
V
ext
= 2(−Z
2
) V
ee
= 0 (4.8)
and therefore the total energy becomes
E = 2
0
= −Z
2
(4.9)
which gives for He (Z=2) E = −4. This is in fact lower than the true groundstate energy
of this problem, because we have failed to evaluate part of the energy in the Hamiltonian.
4.2. HARTREEFOCK 41
To improve on our estimate, and restore the variational principle, we should evaluate the
expectation value of V
ee
on our wavefunction:
V
ee
=
∞
−∞
dx
∞
−∞
dx
φ(x)
2
φ(x
)
2
δ(x −x
) = Z
2
∞
−∞
dx exp(−4Zx) =
Z
2
. (4.10)
For Helium (Z=2), V
ee
= +1 so that the ground state energy becomes E = −4 + 1 = −3.
Because we evaluated all parts of the Hamiltonian on our trial wavefunction, we know that
the true E ≤ −3.
However, using the variational principle, we see that, with almost no extra work, we can
do better still. We chose the length scale of our orbital to be that of the V
ee
= 0 problem,
i.e., α = Z, the nuclear charge. But we can treat this instead as an adjustable parameter,
and ask what value minimizes the energy. Inserting this orbital into the components of the
energy, we ﬁnd:
T
s
= 2
α
2
/2
V
ext
= −2Zα V
ee
= α/2 (4.11)
so that the total energy, as a function of α, is
E(α) = α
2
−2α(Z −1/4). (4.12)
Minimizing this, we ﬁnd α
min
= Z −1/4 and thus
E
min
= −α
2
min
= −(Z −1/4)
2
= −
7
4
2
= −3.0625. (4.13)
We have lowered the energy by 1/16 of a Hartree, which may not seem like much, but is
about 1.7 eV or 40 kcal/mol. Chemical accuracy requires errors of about 1 or 2 kcal/mol.
The best solution to this problem is to ﬁnd the orbital that produces the lowest energy.
We could do this by including many variational parameters in a trial orbital, and minimize
the energy with respect to each of them, giving up when addition of further parameters
has negligible eﬀect. A systematic approach to this would be to consider an inﬁnite set
of functions, usually of increasing kinetic energy, and to include more and more of them.
However, having learned some functional calculus, we can provide a more direct scheme.
The energy, as a functional of the orbital, is:
E[φ] = 2T
S
[φ] + 2V
ext
[φ] +U[φ]/2, U[φ] =
1
2
∞
−∞
dx n
2
(x) (4.14)
where n(x) = 2φ(x)
2
, T
S
and V
ext
are the oneelectron functionals mentioned in chapter 3,
and U is the 1d equivalent of the Hartree energy (discussed more below). For two electrons
in HF, V
ee
= U/2, and we discuss this much more later. If we simply minimize this functional
of the orbital, subject to the restriction that the orbital is normalized, we ﬁnd:
−
1
2
d
2
dx
2
−Zδ(x) +
1
2
v
H
(x)
φ(x) = φ(x), (4.15)
42 CHAPTER 4. TWO ELECTRONS
where
v
H
(x) =
δU
δn(x)
=
∞
−∞
dx
n(x
)δ(x −x
) = n(x) (4.16)
is the Hartree potential. This is also known as the classical or electrostatic potential, as it
is the electrostatic potential due to the charge distribution in classical electrostatics (if the
electrostatic interaction were a δfunction, for this case). (Compare with Eq. (1.6) of the
introduction.) This equation is the HartreeFock equation for this problem.
Exercise 15 HartreeFock equations for two electrons
Derive the HartreeFock equation for two electrons, Eq. (4.15), by minimizing the energy as
a functional of the orbital, Eq. (4.14), keeping the orbital normalized, using the techniques
of Chapter ??.
There are several important aspects of this equation that require comment. First, note the
factor of 2 dividing the Hartree potential in the equation. This is due to exchange eﬀects:
The Hartree potential is the electrostatic potential generated by the total charge density due
to all the electrons. A single electron should only see the potential due to the other electron,
hence the division by two.
Next, note that these are selfconsistent equations, since the potential depends on the
density, which in turn depends on the orbital, which is the solution of the equation, etc.
These can be solved in practice by the iterative method described in the introduction.
To get an idea of what the eﬀective potential looks like, we construct it for our approximate
solution above. There α = 1.75 for He, so the potential is a delta function with an added
Hartree potential of 4 exp(−7x). This contribution is positive, pushing the electron farther
away from the nucleus. We say the other electron is screening the nucleus. The eﬀective
nuclear charge is no longer Z, but rather Z −1/4.
It is straightforward to numerically solve the orbital equation Eq. (4.15) on a grid, essen
tially exactly. But there is an underlying analytic solution in this case. Write
φ(x) = γ cosech (γ x + β) (4.17)
where coth(β) = Z/γ, and γ = Z −1/2. Insertion into Eq. (4.15) yields
= −γ
2
/2, (4.18)
but this does not mean E = −γ
2
. The energy of the eﬀective noninteracting system is not
the energy of the original interacting system. In this case, from the derivation, we see E =
2−U/2 (since = T
S
+V
ext
+U/2). We ﬁnd U/2 = Z/2−1/6, or E = −Z
2
+Z/2−1/12 as
the exact HF energy, slightly lower than our crude estimate using hydrogenic wavefunctions.
But this slight diﬀerence is 1/48, or 0.6 eV or 13 kcal/mol, which is still a signiﬁcant error
by modern standards.
In Fig. 4.1, we plot both the exact orbital and the scaled Hydrogenic orbital, showing there
is very little diﬀerence.
4.2. HARTREEFOCK 43
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 0.5 1 1.5 2 2.5 3 3.5 4
1d He atom in HF
exact
approx
φ
(
x
)
x
Figure 4.1: HF orbital for 1d He atom, both exact and simple exponential with eﬀective charge.
Exercise 16 Analytic solution to HF for 1d He
This exercise is best done with a table of integrals, a mathematical symbolic program, or
simply numerically on a grid.
1. Show that the orbital of Eq. (4.17) is normalized.
2. Show that the orbital of Eq. (4.17) satisﬁes the HF equation with the correct eigenvalue.
3. Check that the HF solution satisﬁes the virial theorem (for power 1).
Exercise 17 Ionization energy of 1d He
The ionization energy of He is deﬁned as
I = E
1
−E
2
(4.19)
where E
N
is the groundstate energy with N electrons in the potential.
1. What is the estimate of the ionization energy given by the crude Hydrogenic orbital
approximation?
2. What is the ionization energy given by the HF solution?
3. Can you think of another way to estimate the ionization energy from the HF solution?
Exercise 18 He in 3d
Repeat the approximate HF calculation for He in three dimensions. What is the orbital energy,
the eﬀective nuclear charge, and the total physical energy? Make a rigorous statement about
both the exact HF energy and the true groundstate energy. Plot the eﬀective potential
(external plus half Hartree) using your approximate solution. (For help, this calculation is
done in many basic quantum books).
44 CHAPTER 4. TWO ELECTRONS
atom E
HF
E E
X
E
C
H
−
0.486 0.528 0.381 0.042
He 2.862 2.904 1.025 0.042
Be
++
13.612 13.656 2.277 0.044
Table 4.1: Energies for twoelectron ions.
4.3 Correlation
The approximate solution of the HartreeFock equations is not exact, because the true wave
function is not a product of two orbitals, but is rather a complicated function of both vari
ables simultaneously. The true wavefunction satisﬁes the exact Schr¨ odinger equation, and
also minimizes the groundstate energy functional for the given external potential. In tra
ditional quantum chemistry, the correlation energy is deﬁned as the diﬀerence between the
HartreeFock energy and the exact groundstate energy, i.e.,
E
trad
C
= E −E
HF
. (4.20)
We illustrate the eﬀects of correlation on the simplest system we have, He. We can
adjust the relative importance of correlation by considering various values of Z. Perhaps
counterintuitively, for large Z, the external potential dominates over the repulsion, and so the
noninteracting solution becomes highly accurate. As Z is reduced, the interaction becomes
more important. For larger Z values, Hartree and exchange dominate over correlation. But
when Z is of order 1, correlation becomes important. Eventually, if Z is made too small, the
system becomes unstable to ejecting one of the electrons, i.e., its energy becomes less than
that of the oneelectron ion, i.e., its ionization potential passes through zero. For Z just a
little greater than that, the system has signiﬁcant correlation. We illustrate these results in
Table 4.3. For large Z, the total energy grows as −Z
2
(the Hydrogenic result), while the
exchange energy grows as −5Z/8. Thus the magnitude of the exchange energy grows with
Z, but its relative size diminishes. Similarly, clearly E
C
is almost independent of Z. So E
C
becomes an ever smaller fraction of E
X
as Z grows. Finally, notice that, for H
−
, the exchange
energy is so small that correlation is 11% of it. The twoelectron ion unbinds at the critical
value of Z = 0.911.
Returning to the 1d world, all the same remarks apply qualitatively. The groundstate
energy for 1d He is 3.154. Since the HF energy is exactly 3 1/12, the correlation energy is
71 mH. Thus we need to estimate correlation energies to within about 10% for useful accuracy
in quantum chemistry. This may seem daunting, but we’ve made great strides so far: Even
our crudest method was good to within 5%. This is a gift of the variational principle, which
implies that all errors in the groundstate energy are at least secondorder in the error in the
wavefunction (don’t expect other expectation values to be as good). Its just that we need
about 0.2%!
4.4. QUESTIONS 45
ﬁxed Z adjustable Z HF exact
E 3 3.0625 3.0833 3.154
% err 5 3 2 0
Table 4.2: Approximate Energies for 1d He.
Exercise 19 Ions on the verge of ionization
1. Using the simple exponential approximation, calculate the critical Z at which 1d He
becomes unstable.
2. Repeat using the exact HF result. Is there an alternative method of estimating Z
c
in this
case? Which is better?
3. The exact energy is accurately reproduced by
E =
1
2
−2Z
2
+Z −α +
β
Z
(4.21)
where α = 3/4 −4/(3π) = 0.3255868 and β = 0.0342. Find the critical value of Z.
4. Calculate the ratio of E
C
to E
X
at Z
c
.
4.4 Questions
All the questions below are conceptual.
1. Is the HF estimate of the ionization potential for 1d He an overestimate or underesti
mate?
2. Consider the approximate HF calculation given in section 4.2. Comment on what it does
right, and what it does wrong. Suggest a simple improvement.
3. Which is bigger, the kinetic energy of the true wavefunction or that of the HF wavefunc
tion for 1d He?
46 CHAPTER 4. TWO ELECTRONS
Chapter 5
Many electrons
In this chapter, we generalize the concepts and ideas introduced for two electrons to the
N electron case.
5.1 Ground state
We deﬁne carefully our notation for electronic systems with N electrons. First, we note
that the wavefunction for N electrons is a function of 3N spatial coordinates and N spin
coordinates. Writing x
i
= (r
i
, σ
i
) to incorporate both, we normalize our wavefunction by:
dx
1
. . .
dx
N
Ψ(x
1
, . . . , x
N
)
2
= 1, (5.1)
where
dx denotes the integral over all space and sum over both spins. Note that the
antisymmetry principle implies
Ψ(x
1
, . . . , x
j
, . . . , x
i
, . . .) = −Ψ(x
1
, . . . , x
i
, . . . , x
j
, . . .). (5.2)
The electronic density is deﬁned by
n(r) = N
¸
σ
dx
2
. . .
dx
N
Ψ(r, σ, x
2
, . . . , x
N
)
2
, (5.3)
and retains the interpretation that n(r)d
3
r is the probability density for ﬁnding any electron
in a region d
3
r around r. The density is normalized to the number of electrons
d
3
r n(r) = N. (5.4)
Our favorite operators become sums over one and twoparticle operators:
ˆ
T = −
1
2
N
¸
i=1
∇
2
i
, (5.5)
ˆ
V
ext
=
N
¸
i=1
v
ext
(r
i
), (5.6)
and
ˆ
V
ee
=
1
2
N
¸
i=j
1
r
i
−r
j

(5.7)
47
48 CHAPTER 5. MANY ELECTRONS
Note that the factor of 2 in the electronelectron repulsion is due to the sum running over all
pairs, e.g., including (1,2) and (2,1). The Schr¨ odinger equation is then
ˆ
T +
ˆ
V
ee
+
ˆ
V
ext
Ψ(x
1
, . . . , x
N
) = EΨ(x
1
, . . . , x
N
). (5.8)
The groundstate energy can be extracted from the variational principle:
E = min
Ψ
Ψ
ˆ
T +
ˆ
V
ee
+
ˆ
V
ext
Ψ, (5.9)
once the minimization is performed over all normalized antisymmetic wavefunctions.
5.2 HartreeFock
In the special case of noninteracting particles, we denote the wavefunction by Φ instead of
Ψ, and this will usually be a single Slater determinant of occupied orbitals, i.e.,
Φ(x
1
, . . . , x
N
) =
φ
1
(x
1
) · · · φ
N
(x
1
)
.
.
.
.
.
.
φ
1
(x
N
) · · · φ
N
(x
N
)
(5.10)
For systems with equal numbers of up and down particles in a spinindependent external
potential, the full orbitals can be written as a product, φ
i
(x) = φ
i
(r)σ, and each spatial
orbital appears twice. The HartreeFock energy is
E = min
Φ
Φ
ˆ
T +
ˆ
V
ee
+
ˆ
V
ext
Φ, (5.11)
The ﬁrst contribution is just the noninteracting kinetic energy of the many orbitals, and the
last is their external potential energy.
The Coulomb interaction for a single Slater determinant yields, due to the antisymmetric
nature of the determinant, two contributions:
Φ
ˆ
V
ee
Φ = U[Φ] +E
X
[Φ]. (5.12)
The ﬁrst of these is called the direct or Coulomb or electrostatic or classical or Hartree
contribution, given in Eq. F.4, with the density being the sum of the squares of the occupied
orbitals. This is the electrostatic energy of the charge density in electromagnetic theory,
ignoring its quantum origin. The second is the Fock or exchange integral, being
E
X
[φ
i
] = −
1
2
¸
σ
¸
i,j
occ
d
3
r
d
3
r
φ
∗
iσ
(r) φ
∗
jσ
(r
) φ
iσ
(r
) φ
jσ
(r)
r −r

(5.13)
for a determinant of doublyoccupied orbitals. This is a purely Pauliexclusion principle eﬀect.
Exercise 20 Exchange energies for one and two electrons
Argue that, for one electron, E
X
= −U, and show that, for two electrons in a singlet,
E
X
= −U/2.
5.3. CORRELATION 49
atom Z E
HF
E E
C
H 1 0.5 0.5 0
He 2 2.862 2.904 0.042
Li 3 7.433 7.478 0.045
Be 4 14.573 14.667 0.094
B 5 24.529 24.654 0.125
C 6 37.689 37.845 0.156
N 7 54.401 54.589 0.188
O 8 74.809 75.067 0.258
F 9 99.409 99.733 0.324
Ne 10 128.547 128.937 0.39
Ar 18 526.817 527.539 0.722
Kr 36 2752.055 2753.94 1.89
Xe 54 7232.138 7235.23 3.09
Rn 86 22866.745 22872.5 5.74
Table 5.1: Total energies in HF, exactly, and correlation energies across ﬁrst row and for noble gas atoms.
To ﬁnd the orbitals in HartreeFock, we must minimize the energy as a functional of
each orbital φ
i
(r), subject to the constraint of orthonormal orbitals. Doing this, using the
techniques of chapter 2 yields the HartreeFock equations
−
1
2
∇
2
+v
ext
(r) +v
H
(r)
φ
i
(r) +f
F,i
[{φ
i
}](r) =
i
φ(r) (5.14)
where the last contribution to the potential is due to the Fock operator, and is deﬁned by
f
F,i
[{φ
i
}](r) = −
N
¸
j
d
3
r
φ
∗
j
(r
) φ
i
(r
)
r −r

φ
j
(r). (5.15)
Note that this oddlooking animal is orbitaldependent, i.e., it is a diﬀerent function of r for
each occupied orbital. Also, for just one electron, the Hartree and Fock terms in the potential
cancel, as they should. In order to discuss the pro’s and con’s of HF calculations, we must
ﬁrst discuss the errors it makes.
5.3 Correlation
The deﬁnition of correlation energy remains the same for N electrons as for two: It is the
error made by a HartreeFock calculation. Table 5.3 lists a few correlation energies for
atoms. We see that correlation energies are a very small (but utterly vital) fraction of the
total energy of systems. They are usually about 2040 mH/electron, a result we will derive
later.
50 CHAPTER 5. MANY ELECTRONS
Exercise 21 Atomic energies
This exercise uses the numbers in Table 5.3.
1. By plotting ln(−E) versus ln(Z), ﬁnd the dependence of the total energy on Z for large
Z.
2. Repeat for the correlation energy.
3. Plot the correlation energy (a) across the ﬁrst row and (b) down the last column of the
periodic table. Comment on the results.
But we strive for chemical accuracy in our approximate solutions, i.e., errors of less than
2 mH per bond. Another important point to note is that in fact, we often do not need total
energies to this level of accuracy, but rather only energy diﬀerences, e.g., between a molecule
and its separated constituent atoms, in which the correlation contribution might be a much
larger fraction, and in which compensating errors might occur in the separate calculation of
the molecule and the atoms. An example of this is the core electrons, those electrons in closed
subshells with energies below the valence electrons, which are often relatively unchanged in
a chemical reaction. A large energy error in their contribution is irrelevant, as it cancels out
of energy diﬀferences.
Quantum chemistry has developed many interesting ways in which to calculate the corre
lation energy. These can mostly be divided into two major types: perturbation theory, and
wavefunction calculations. The ﬁrst is usually in the form of MollerPlesset perturbation the
ory, which treats the HartreeFock solution as the starting point, and performs perturbation
theory in the Coulomb interaction. These calculations are nonvariational, and may produce
energies below the groundstate energy. Perhaps the most common type of wavefunction
calculation is conﬁguration interaction (CI). A trial wavefunction is formed as a linear com
bintation of products of HF orbitals, including excited orbitals, and the energy minimized.
In electronic structure calculations of weaklycorrelated solids, most often density functional
methods are used. Green’s function methods are often applied to stronglycorrelated systems.
Exercise 22 Correlation energy
Show that the correlation energy is never positive. When is it zero?
An interesting paradox to note in chemistry is that most modern chemists think of reac
tivity in terms of frontier orbitals, i.e., the HOMO (highest occupied molecular orbital) and
the LUMO (lowest unoccupied), and their energetic separation, as in Fig. 1.3. In the wave
function approach, these are entirely constructs of the HF approximation, which makes errors
of typically 0.2 Hartrees in binding energies. Thus, although this approximation obviously
contains basic chemical information, needed for insight into chemical reactivity, the resulting
thermochemistry is pretty bad. On the other hand, to obtain better energetics, one adds
many more terms to the approximate wavefunction, losing the orbital description completely.
5.4. ATOMIC CONFIGURATIONS 51
We will see later how DFT resolves this paradox, by showing how an orbital calculation can
in principle yield the exact energetics.
5.4 Atomic conﬁgurations
We have seen how the HartreeFock equations produce a selfconsistent set of orbitals with
orbital energies. In many cases, the true wavefunction has a strong overlap with the HF
wavefunction, and so the behavior of the interacting system can be understood in terms of
the HF orbitals. We occupy the orbitals in order.
This is especially true for the atoms, and the positions of the HF orbitals largely determine
the strucutre of the periodic table. If we ﬁrst ignore interaction, and consider the hydrogenic
levels, we get the basic idea. For Hydrogenic atoms (N = 1), there is an exact degeneracy
between all orbitals of a given principal quantum number. For each n, there are n values of
l, ranging from 0 to n − 1. For each l, there are 2l + 1 values of m
l
= −l, ..., 0, ..l. Thus
each closed shell contains
g
n
=
n−1
¸
l=0
(2l + 1) = n(n −1) +n = n
2
(5.16)
Since each orbital can hold 2 electrons, the nthshell contains 2n
2
electrons. The ﬁrst shell
closes at N = 2, the next at 10, the next at 28, and so forth.
This scheme works up to Ar, but then fails to account for the transition metals. To
understand these, we must consider the actual HF orbitals, which are not hydrogenic. Since
the HF potentials are not Coulombic, the degeneracy w.r.t. diﬀerent lvalues is lifted, and
we have
nl
. As expected, for ﬁxed n, the higher l, the less negative the energy. Thus the 2s
orbitals lie lower than 2p. This eﬀect does not change the ﬁlling order described above.
But for Z between 7 and 20, the 4s orbital dips below the 3d. So it gets ﬁlled before the
3d, and, for example, the groundstate conﬁguration of Ca is [Ar]4s
2
. Then the 3d starts
ﬁlling, and so the transition metals appear a little late. Even when the 3d orbital energy drops
below the 4s, the total energy remains lower with the 4s ﬁlled. The situation is similar for
higher n.
A list of the ﬁlled orbitals is called the electronic conﬁguration of a system. This still does
not determine the groundstate entirely, as many diﬀerent angular momenta combinations,
both spin and orbital, (terms) can have the same conﬁguration. Hund’s rules are used to
choose which one has the lowest energy. Our treatment will not require details beyond
knowing the lowest conﬁguration.
5.5 Atomic densities
Since we will be constructing a formally exact theory of interacting quantum mechanics based
on the oneelectron density, it seems appropriate here to discuss and show some pictures of
52 CHAPTER 5. MANY ELECTRONS
this quantity for atoms. All atomic densities, when plotted as function of r with the nucleus
0
5
10
15
20
25
0 0.5 1 1.5 2 2.5 3
Ar atom
4
π
r
2
n
(
r
)
r
Figure 5.1: Radial density in Ar atom.
at the origin, have been found to decay monotonically, although that’s never been generally
proven. Close to a nucleus, the external potential dominates in the Schr¨ odinger equation,
and causes a cusp in the density, whose size is proportional to the nuclear charge:
dn
dr

r=0
= −2Z n(0) (5.17)
for a nucleus of charge Z at the origin. This is Kato’s cusp condition, and we already saw
examples of this in earlier exercises. It is also true in our onedimensional examples with
deltafunction external potentials.
Typically, spherical densities are plotted multiplied by the phasespace factor 4πr
2
. This
means the area under the curve is precisely N. Furthermore, the diﬀerent electronic shells
are easily visible. We take the Ar atom as an example. In Fig. 5.1, we plot its radial density.
This integrates to 18 electrons. We ﬁnd that the integral up to r = 0.13 contains 2 electrons,
up to 0.25 contains 4, 0.722 contains 10, and 1.13 contains 12. These correspond to the 2
1s electrons, 2 2s electrons, 6 2p electrons, and 2 3s electrons, respectively. Note that the
peaks and dips in the radial density roughly correspond to these shells.
The decay at large distances is far more interesting. When one coordinate in a wavefunction
is taken to large distances from the nuclei, the Nelectron groundstate wavefunction collapses
to the product of the squareroot of the density times the (N−1)electron wavefunction. This
means the squareroot of the density satisﬁes a Schr¨ odingerlike equation, whose eigenvalue
is the diﬀerence in energies between the two systems:
n(r) = Ar
β
exp(−αr) (r →∞) (5.18)
where α =
√
2I, and
I = E(N −1) −E(N) (5.19)
is ﬁrst ionization potential. In fact, the power can also be deduced, β = (Z −N +1)/α−1,
and A is some constant. Thus, a useful and sensitive function of the density to plot for
5.6. QUESTIONS 53
spherical systems is
κ(r) = (1/2)
d log(n(r))
dr
, (5.20)
since κ(0) = −Z, while κ(r) → −α, as r → ∞. The groundstate density can always be
reconstructed from
n(r) = C exp(
∞
r
dr
2κ(r
)) (5.21)
where the constant is determined by normalization.
In a onedimensional world with delta function interactions, these conditions remain true,
except for the details of the power law in front of the exponential in Eq. (5.18). In particular,
we have already seen the cusp in the orbital for 1d hydrogenic atoms. We may see dramatic
evidence for these conditions in our HF solution of the 1d He atom, by plotting κ(x) on the
left of Fig. 5.2. Near the nucleus, the exact cusp condition is satisﬁed, so that κ equals 2,
but at large distances, it tends to a smaller constant, determined by the ionization potential.
We can also clearly see the diﬀerence between the approximate and exact HF solutions here,
which was not so visible in the last ﬁgure. On the right, we have also plotted κ(r) for real
He, and see that the eﬀect of correlation on the density is extremely small: The HF and exact
densities are very close.
2
1.9
1.8
1.7
1.6
1.5
1.4
1.3
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
1d He atom in HF
exact
approx
κ
x
2
1.9
1.8
1.7
1.6
1.5
1.4
1.3
0 2 4 6 8 10
He atom
exact
HF
r
Figure 5.2: κ in (a) 1d He atom in HF approximation, both exact and approximate; (b) in real He, in HF and exactly
5.6 Questions
All the questions below are conceptual.
1. Does each orbital in a HF calculation satisfy the same Schr¨ odinger equation?
2. What is the relationship between E
HF
and
¸
N
i=1
i
?
3. Speculate on why the correlation energy of Li is about the same as that of He.
54 CHAPTER 5. MANY ELECTRONS
4. Which is bigger, the kinetic energy of the true wavefunction or that of the HF wavefunc
tion, for an atom?
5. What do you expect happens to the correlation energy of the twoelectron ions as Z →
∞?
6. What do you expect happens to the correlation energy of the fourelectron ions as Z →
∞?
Part II
Basics
55
Chapter 6
Density functional theory
1
This chapter deals with the foundation of modern density functional theory as an exact
approach (in principle) to systems of interacting particles. In our case, electrons, i.e.,
fermions interacting via the Coulomb repulsion.
6.1 One electron
To illustrate the basic idea behind density functional theory, we ﬁrst formulate the problem
of one electron as a density functional problem. We know that the groundstate energy can
be written as a density functional instead of as a wavefunction functional. This functional is
E[n] =
1
8
d
3
r
∇n
2
n
+
d
3
r v(r) n(r) (6.1)
in any number of dimensions. We must minimize it, subject to the constraint that
d
3
r n(r) =
1. Using Lagrange multipliers, we construct the auxillary functional:
H[n] = E[n] −µN (6.2)
where N =
d
3
r n(r). Minimizing this yields
δH
δn(r)
=
δT
VW
δn(r)
+v(r) −µ = 0 (6.3)
Using the derivative from Ex. 5, we ﬁnd the Schr¨ odinger equation for the density:
¸
−
∇
2
4
+
∇n
2
8n
2
+v(r)
¸
n(r) = µn(r) (6.4)
with boundary conditions that n(r) and ∇n(r) → 0 at the edges. We can identify the
Lagrange multiplier by integrating both sides over all space at the solution. Since the integral
of any Laplacian vanishes (with the given boundary conditions), we see that µ = E.
Exercise 23 Checking equation for density
Show that the groundstate density for a particle in a box satisﬁes Eq. (6.4).
1
c 2000 by Kieron Burke. All rights reserved.
57
58 CHAPTER 6. DENSITY FUNCTIONAL THEORY
6.2 HohenbergKohn theorems
It is selfevident that the external potential in principle determines all the properties of the sys
tem: this is the normal approach to quantum mechanical problems, by solving the Schr¨ odinger
equation for the eigenstates of the system.
The ﬁrst HohenbergKohn theorem demonstrates that the density may be used in place
of the potential as the basic function uniquely characterizing the system. It may be stated
as: the groundstate density n(r) uniquely determines the potential, up to an arbitrary
constant.
In the original HohenbergKohn paper, this theorem is proven for densities with non
degenerate ground states. The proof is elementary, and by contradiction. Suppose there
existed two potentials diﬀering by more than a constant, yielding the same density. These
would have two diﬀerent groundstate wavefunctions, Ψ
1
and Ψ
2
. Consider Ψ
2
as a trial
wavefunction for potential v
ext,1
(r). Then, by the variational principle,
Ψ
2

ˆ
T +
ˆ
V
ee
+
ˆ
V
ext,1
Ψ
2
≥ Ψ
1

ˆ
T +
ˆ
V
ee
+
ˆ
V
ext,1
Ψ
1
. (6.5)
But since both wavefunctions have the same density, this implies
Ψ
2

ˆ
T +
ˆ
V
ee
Ψ
2
≥ Ψ
1

ˆ
T +
ˆ
V
ee
Ψ
1
. (6.6)
But we can always swap which wavefunction we call 1 and which we call 2, which reverses
this inequality, leading to a contradiction, unless the total energies of the two wavefunctions
are the same, which implies they are the same wavefunction by the variational principle and
the assumption of nondegeneracy. Then, simple inversion of the Schr¨ odinger equation, just
as we have done several times for the oneelectron case, yields
N
¸
i=1
v
ext
(r
i
) = −
1
2
N
¸
i=1
∇
2
i
Ψ/Ψ+
1
2
¸
j=i
1
r
i
−r
j

(6.7)
This determines the potential up to a constant.
Exercise 24 Finding the potential from the density
For the smoothed exponential, φ(x) = C(1+x) exp(−x), ﬁnd the potential for which this
is an eigenstate, and plot it. Is it the ground state of this potential?
An elegant constructive proof was found later by Levy, which automatically includes degen
erate states. It is an example of the constrained search formalism. Consider all wavefunctions
Ψ which yield a certain density n(r). Deﬁne the functional
F[n] = min
Ψ→n
Ψ
ˆ
T +
ˆ
V
ee
Ψ (6.8)
where the search is over all antisymmetric wavefunctions yielding n(r). Then, for any n(r),
any wavefunction minimizing
ˆ
T +
ˆ
V
ee
is a groundstate wavefunction, since the groundstate
energy is simply
E = min
n
F[n] +
d
3
r v
ext
(r) n(r)
, (6.9)
6.2. HOHENBERGKOHN THEOREMS 59
from the variational principle, where the search is over all normalized positive densities. Any
such wavefunction can then be fed into Eq. (6.7) to construct the unique corresponding
potential. We denote the minimizing wavefunction in Eq. (6.8) by Ψ[n]. This gives us a
verbal deﬁnition of the groundstate wavefunction. The exact groundstate wavefunction of
density n(r) is that wavefunction that yields n(r) and has minimizes T + V
ee
. We may
also deﬁne the exact kinetic energy functional as
T[n] = Ψ[n]
ˆ
T
Ψ[n], (6.10)
and the exact electronelectron repulsion functional as
V
ee
[n] = Ψ[n]
ˆ
V
ee
Ψ[n]. (6.11)
The second HohenbergKohn theorem states that the functional F[n] is universal, i.e.,
it is the same functional for all electronic structure problems. This is evident from Eq.
(6.8), which contains no mention of the external potential. To understand the content of
this statement, consider some of our previous problems. Recall our orbital treatment of
oneelectron problems. The kinetic energy functional, T[φ] =
1
2
dx φ
(x)
2
, is the same
functional for all oneelectron problems. When we evaluate the kinetic energy for a given
trial orbital, it is the same for that orbital, regardless of the particular problem being solved.
Similarly, when we treated the two electron case within the HartreeFock approximation, we
approximated
F[n] ≈ F
HF
[n] =
1
8
d
3
r
∇n
2
n
+
1
4
d
3
r
d
3
r
n(r)n(r
)
r −r

. (6.12)
Then minimization of the energy functional in Eq. (6.9) with F
HF
[n] yields precisely the
HartreeFock equations for the twoelectron problem. The important point is that this single
formula for F
HF
[n] is all that is needed for any twoelectron HartreeFock problem. The
second HohenbergKohn theorem states that there is a single F[n] which is exact for all
electronic problems.
Exercise 25 Errors in HartreeFock functional
Comment, as fully as you can, on the errors in F
HF
[n] relative to the exact F[n] for two
electrons.
The last part of the HohenbergKohn theorem is the EulerLagrange equation for the
energy. We wish to mininize E[n] for a given v
ext
(r) keeping the particle number ﬁxed. We
therefore minimize E[n] −µN, as in chapter ??, and ﬁnd the EulerLagrange equation:
δF
δn(r)
+v
ext
(r) = µ. (6.13)
We can identify the constant µ as the chemical potential of the system, since µ = ∂E/∂N.
The exact density is such that it makes the functional derivative of F exactly equal to the
60 CHAPTER 6. DENSITY FUNCTIONAL THEORY
negative of the external potential (up to a constant). Note that it would be marvellous if
we could ﬁnd an adequate approximation to F for our purposes, so that we could solve Eq.
(6.13) directly. It would yield a single integrodiﬀerential equation to be solved, probably
by a selfconsistent procedure, for the density, which could then be normalized and inserted
back into the functional E[n], to recover the groundstate energy. In the next section, we
will examine the original crude attempt to do this (ThomasFermi theory), and ﬁnd that,
although the overall trends are sound, the accuracy is insuﬃcient for modern chemistry and
materials science. Note also that insertion of F
HF
[n] will yield an equation for the density
equivalent to the orbital HF equation.
An important formal question is that of vrepresentability. The original HK theorem was
proven only for densities that were groundstate densities of some interacting electronic prob
lem. The constrained search formulation extended this to any densities extracted from a
single wavefunction, and this domain has been further extended to densities which result
from ensembles of wavefunctions. The interested reader is referred to Dreizler and Gross for
a thorough discussion.
6.3 ThomasFermi
In this section, we brieﬂy discuss the ﬁrst density functional theory (1927), its successes, and
its limitations. Note that it predates HartreeFock by three years.
In the ThomasFermi theory, F[n] is approximated by the local approximation for the
(noninteracting) kinetic energy of a uniform gas, plus the Hartree energy
F
TF
[n] = A
S
d
3
rn
5/3
(r) +
1
2
d
3
r
d
3
r
n(r)n(r
)
r −r

. (6.14)
Several points need to be clariﬁed. First, these expressions were developed for a spin
unpolarized system, i.e., one with equal numbers of up and down spin electrons, in a spin
independent external potential. Second, in the kinetic energy, we saw in Chapter ?? how the
power of n can be deduced by dimensional analysis, while the coeﬃcient is chosen to agree
with that of a uniform gas, yielding A
S
= (3/10)(3π
2
)
2/3
. We will derive these numbers in
detail in Chapter 8. They can also be derived by classical arguments applied to the electronic
ﬂuid.
Insertion of this approximate F into the EulerLagrange equation yields the ThomasFermi
equation:
5
3
A
S
n
2/3
(r) +
d
3
r
n(r
)
r −r

+v
ext
(r) = µ (6.15)
We will focus on its solution for spherical atoms, which can easily be achieved numerically with
a simple ordinary diﬀerential equation solver. We can see immediately some problems. As
r →0, because of the singular Coulomb external potential, the density is singular, n →1/r
3/2
(although still integrable, because of the phasespace factor 4πr
2
). Again, as r → ∞,
6.3. THOMASFERMI 61
the density decays with a power law, not exponentially. But overall, the trends will be
approximately right.
The ThomasFermi equation for neutral densities, Eq. (6.15), is usually given in term of a
dimensionless function Φ(x) satisfying:
Φ
(x) =
Φ
3
x
(6.16)
where Φ(0) = 1 and Φ(∞) = 0. The distance x = Z
1/3
r/a, where
a =
1
2
3π
4
2/3
= 0.885341, (6.17)
and the density is given by
n(r) =
Z
2
4πa
3
Φ
x
3/2
(6.18)
Its been found that at the solution for neutral atoms, Φ
(0) = −B, where B = 1.58807102261,
and the asymptotic behavior is the trivial solution:
Φ(x) =
144
x
3
, x →∞. (6.19)
These results, and the density of Fig. 6.1, are needed for the next exercise.
0
10
20
30
40
50
60
70
80
90
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Xe atom in LDA
4
π
r
2
n
(
r
)
r
Figure 6.1: Radial density in the Xe atom, calculated using LDA.
Exercise 26 ThomasFermi for atoms
1. Using the diﬀerential equation and integrating by parts, ﬁnd the TF kinetic energy. Using
the virial theorem for atoms, show that
E = −
3
7
B
a
Z
7/3
= −.7845 Z
7/3
(6.20)
This is the exact limiting behavior for nonrelativistic atoms as Z →∞.
62 CHAPTER 6. DENSITY FUNCTIONAL THEORY
2. Show that the external potential goes as
V
ext
= −
B
a
Z
7/3
= 1.79374 Z
7/3
(6.21)
3. By using the virial theorem, deduce the behavior of the Hartree energy as Z →∞.
4. Using table 5.3, compare with energies along ﬁrst row and for the noble gas atoms.
Comment.
5. An accurate radial density for the Xe atom is given in Fig. 6.1. An accurate parametriza
tion of Φ is
Φ(y) = (1 + αy + β y
2
exp(−γ y))
2
exp(−2αy) (6.22)
where y =
√
x and α = 0.7280642371, β = −0.5430794693, and γ = 0.3612163121.
Plot the approximate TF density on the same ﬁgure, and comment.
6. Calculate the expectation value 1/r for ThomasFermi. Do atomic radii grow or shrink
with Z? Explain why.
To see the quality of our TF solution, we calculate the κ(r) function of the last chapter,
and compare with that of the exact solution. We see that, although it messes up all the
20
15
10
5
0
0 0.5 1 1.5 2 2.5 3
Kappa in Ar atom
exact
approx TF
Figure 6.2: κ(r) for the Ar atom, both exactly and in ThomasFermi calculation of Exercise ??.
details (nuclear cusp, shell structure, decay at large distances), it has a remarkably good
overall shape for such a simple theory.
Lastly, Teller showed that molecules do not bind in TF theory, because the errors in TF
are far larger than the relatively small binding energies of molecules.
6.4 Particles in boxes
Before closing this chapter, let us return to the examples of the introduction and understand
how they are connected to the ThomasFermi model. They highlight more clearly the relative
6.5. QUESTIONS 63
advantages and disadvantages of local approximations, since they apply in one dimension (no
shell structure) and without interaction.
First we note that, for the particles in a box, we cheated slightly, by applying the local
approximation to the exact density. But if we’re making the local approximation for the
energy, where would we have gotten the exact density? A more honest calculation is to
ﬁnd the selfconsistent density within the given approximation. Repeating the arguments of
section 6.1, we ﬁnd
n(x) =
1
π
2(µ −v
ext
(x)) (6.23)
wherever the argument of the root is positive, zero otherwise, and µ is to be found by
normalization (
dx n(x) = N.
For the particle in the box, v(x) vanishes inside the box, yielding
n(x) = n =
N
L
(6.24)
This is quite diﬀerent from the exact density for one electron, but the true density approaches
it as N gets large. Using exactly this density, the energy contains only one term, the
asymptotic one, E = π
2
N
3
/(6L
2
).
Exercise 27 Energy from local approximation on exact density in box
A ‘fun’ math problem is to show that, using the local approximation evaluated on the exact
density for a particle in a box, the energy can be found analytically to be
E
N
=
π
2
6
N
2
+
9
8
N +
3
8
(6.25)
You can check how it compares to your answers to Exercise X.
In particular, for one electron, the selfconsistent solution is too small by a factor of 3!
Compare this with the 25% error using the exact density. This is because, although the local
approximation is good for the energy, its derivative is not so good, and the density is quite
sensitive to this.
Note that our reasoning applies to all bounded 1d problems, not just particles in boxes.
Exercise 28 Large N for harmonic oscilator
Calculate the selfconsistent density of the local approximation for the harmonic oscillator,
and compare with your exact densities. Also, compare the energies.
6.5 Questions
These are all conceptual.
64 CHAPTER 6. DENSITY FUNCTIONAL THEORY
1. How would you go about ﬁnding the potential from the wavefunction of a twoelectron
system? For two electrons, can more than one potential have the same groundstate
density?
2. Deﬁne the groundstate wavefunction generating density n(r) without mentioning the
external potential.
3. When we say F[n] is a universal functional, exactly what do we mean? After all, since
v
ext
(r) itself is a functional of n(r), is not the entire energy universal?
4. Why do the HK theorems specify the groundstate and not, say, the ﬁrst excited state?
5. Using the local approximation for the onedimensional kinetic energy developed in Chapter
1, ﬁnd the density for the onedimensional harmonic oscillator, and evaluate the total
energy on that density. Compare with your results for the HO problem there.
Chapter 7
KohnSham
The most important step, establishing modern density functional theory as a useful tool,
comes with the introduction of the KohnSham equations, which allow almost all the
kinetic energy to be calculated exactly.
7.1 KohnSham equations
The major error in the ThomasFermi approach comes from approximating the kinetic energy
as a density functional. We saw in chapter one that local approximations to the kinetic
energy are typically good to within 1020%. However, since for Coulombic systems, the
kinetic energy equals the absolute value of the total energy, errors of 10% are huge. Even
if we could reduce errors to 1%, they would still be too large. A major breakthrough in this
area is provided by the KohnSham construction of noninteracting electrons with the same
density as the physical system, because solution of the KohnSham equations produces the
exact noninteracting kinetic energy, which includes almost all the true kinetic energy.
We now have the theoretical tools to immediately write down these KS equations. Recall
from the introduction, that the KS system is simply a ﬁctitious system of noninteracting
electrons, chosen to have the same density as the physical system. Then its orbitals are given
by Eq. (1.5), i.e.,
−
1
2
∇
2
+v
S
(r)
¸
φ
i
(r) =
i
φ
i
(r), (7.1)
and yield
n(r) =
N
¸
i=1
φ
i
(r)
2
. (7.2)
The subscript s denotes singleelectron equations. But the Euler equation that is equivalent
to these equations is
δT
S
δn(r)
+v
S
(r) = µ, (7.3)
where
T
S
[n] = min
Φ→n
Φ
ˆ
T
Φ, (7.4)
65
66 CHAPTER 7. KOHNSHAM
is the kinetic energy of noninteracting electrons. We have implicitly assumed that the Kohn
Sham wavefunction is a single Slater determinant, which is true most of the time. We denote
the minimizing wavefunction in Eq. (7.4) by Φ[n]. This gives us a verbal deﬁnition of the
KohnSham wavefunction. The KohnSham wavefunction of density n(r) is that wave
function that yields n(r) and has least kinetic energy. Obviously T
S
[n] = Φ[n]
ˆ
T
Φ[n],
which diﬀers from T[n].
Exercise 29 Kinetic energies
Show that T[n] ≥ T
S
[n]. What is the relation between T
HF
[n] and T
S
[n]? Is there a simple
relation between T[n] and T
HF
[n]?
Now, write the groundstate functional of an interacting system in terms of the non
interacting kinetic energy:
F[n] = T
S
[n] +U[n] +E
XC
[n]. (7.5)
We have explicitly included the Hartree energy, as we know this will be a large part of the
remainder, and we know it explictly as a density functional. The rest is called the exchange
correlation energy. Inserting F[n] into Eq. (6.13) and comparing with Eq. (7.3), we ﬁnd
v
S
(r) = v
ext
(r) +
d
3
r
n(r)
r −r

+v
XC
[n](r) v
XC
(r) =
δE
XC
δn(r)
(7.6)
This is the ﬁrst and most important relationship of exact density functional theory: from the
functional dependence of F[n], we can extract the potential felt by noninteracting electrons
of the same density.
We note several important points:
• The KohnSham equations are exact, and yield the exact density. For every physical
system, the KohnSham alter ego is welldeﬁned and unique (recall Fig. 1.2). There is
nothing approximate about this.
• The KohnSham equations are a set of singleparticle equations, and so are much easier
to solve than the coupled Schr¨ odinger equation, especially for large numbers of electrons.
However, in return, the unknown exchangecorrelation energy must be approximated.
(We do not know this functional exactly, or else we would have solved all Coulomb
interacting electronic problems exactly.)
• While the KS potential is unique by the HohenbergKohn theorem (applied to non
interacting electrons), there are known examples where such a potential cannot be found.
In common practice, this has never been a problem.
• The great advantage of the KS equations over ThomasFermi theory is that almost all
the kinetic energy (T
S
) is treated exactly.
7.2. EXCHANGE 67
• The KS orbitals supercede the HF orbitals, in providing an exact molecular orbital theory.
With exact KS theory, we see now how an orbital calculation can provide exact energetics.
The HF orbitals are better thought of as approximations to exact KS orbitals. So the
standard ﬁgure, Fig. 1.3, and other orbital pictures, are much more usefully interpreted
as pictures of KS atomic and molecular orbitals.
• Note that pictures like that of Fig. 1.2 tell us nothing about the functional dependence of
v
S
(r). That KS potential was found simply by inverting the noninteracting Schr¨ odinger
equation for a single orbital. This gives us the exact KS potential for this system, but
tells us nothing about how that potential would change with the density.
• By subtracting T
S
and U from F, what’s left (exchange and correlation) will turn out to
be very amenable to localtype approximations.
7.2 Exchange
In density functional theory, we deﬁne the exchange energy as
E
X
[n] = Φ[n]
ˆ
V
ee
Φ[n] −U[n], (7.7)
i.e., the electronelectron repulsion, evaluated on the KohnSham wavefunction, yields a direct
contribution (the Hartree piece) and an exchange contribution. In most cases, the KohnSham
wavefunction is simply a single Slater determinant of orbitals, so that E
X
is given by the Fock
integral, Eq. (5.13) of the KS orbitals. These diﬀer from the HartreeFock orbitals, in that
they are the orbitals that yield a given density, but that are eigenstates of a single potential
(not orbitaldependent). Note that Eq. (5.13) does not give us the exchange energy as an
explicit functional of the density, but only as a functional of the orbitals. The total energy in
a HartreeFock calculation is extremely close to T
S
+U +V
ext
+E
X
.
The diﬀerences between HF exchange and KSDFT exchange are subtle. They can be
thought of as having two diﬀerent sources.
1. The KS exchange is deﬁned for a given density, and so the exact exchange of a system
is the exchange of the KS orbitals evaluated on the exact density. The HF exchange is
evaluated on the HF orbitals for the system.
2. To eliminate the density diﬀerence, we can compare KS E
X
[n
HF
] with that from HF. The
remaining diﬀerence is due to the local potential for the KS orbitals.
Exercise 30 HF versus DFT exchange
Prove E
HF
≤ T
S
[n
HF
] +U[n
HF
] +V
ext
[n
HF
] +E
X
[n
HF
] for any problem.
Exercise 31 Exchange energies
Calculate the exchange energy for the hydrogen atom and for the threedimensional harmonic
oscillator. If I construct a spin singlet, with both electrons in an exponential orbital, what is
the exchange energy then?
68 CHAPTER 7. KOHNSHAM
atom E T V
ext
V
ee
T
S
U E
X
T
C
U
C
E
C
He 2.904 2.904 6.753 0.946 2.867 2.049 1.025 0.037 0.079 0.042
Be 14.667 14.667 33.710 4.375 14.594 7.218 2.674 0.073 0.169 0.096
Ne 128.94 128.94 311.12 53.24 128.61 66.05 12.09 0.33 0.72 0.39
Table 7.1: Energy components for ﬁrst three noble gas atoms, found from the exact densities.
7.3 Correlation
In density functional theory, we then deﬁne the correlation energy as the remaining unknown
piece of the energy:
E
C
[n] = F[n] −T
S
[n] −U[n] −E
X
[n]. (7.8)
We will show that it usually better to approximate exchangecorrelation together as a single
entity, rather than exchange and correlation separately. Inserting the deﬁnition of F above,
we ﬁnd that the correlation energy consists of two separate contributions:
E
C
[n] = T
C
[n] +U
C
[n] (7.9)
where T
C
is the kinetic contribution to the correlation energy,
T
C
[n] = T[n] −T
S
[n] (7.10)
(or the correlation contribution to kinetic energy), and U
C
is the potential contribution to the
correlation energy,
U
C
[n] = V
ee
[n] −U[n] −E
X
[n]. (7.11)
For many systems, T
C
∼ −E
C
∼ −U
C
/2.
Exercise 32 Correlation energy from Hamiltonian
Deﬁne the DFT correlation energy in terms of the Hamiltonian evaluated on wavefunctions.
Exercise 33 Correlation energy
Prove E
C
≤ 0, and say which is bigger: E
C
or E
trad
C
.
Finally, to get an idea of how large these energies are for real systems, in Table 7.3 we
give accurate values for three noble gas atoms. These numbers were found as follows. First,
a highly accurate solution was found of the full Schr¨ odinger equation for each atom. From
it, the total groundstate energy and its various components could be calculated. Also, the
groundstate density was extracted from the wavefunction. Next, a search was made for the
unique KS potential that corresponded to that density. This can be thought of as guessing
the potential, solving for its density, and comparing with the exact density. Then changing the
potential to shift the computed density toward the exact one, and repeating until converged.
Once the KS potential and its orbitals are found, it is straightforward to evaluate its kinetic,
Hartree, and exchange energies. The last set of columns were found by subtracting KS
quantities from their exact counterparts.
There are many points to note in this table:
7.4. QUESTIONS 69
• The exact kinetic energy is simply E, by the virial theorem for atoms.
• The interaction with the nucleus, V
ext
, is a little more than twice the kinetic energy, as
it must also overcome the (positive) V
ee
.
• The electronelectron repulsion is typically less than half the kinetic energy, but all three
are on the same scale. Recall from ThomasFermi theory that V
ext
= −7T/3 = −7V
ee
.
• The KS kinetic energy T
S
is almost as large as the true kinetic energy, and they diﬀer by
less than 2%.
• The Hartree energy is typically a small overestimate of V
ee
.
• The exchange energy cancels a fraction of the Hartree energy, that fraction getting smaller
as Z increases.
• The kinetic contribution to the correlation energy, which is also the correlation contribu
tion to the kinetic energy, is on the same scale as −E
C
.
• The potential contribution to the correlation energy, which is also the correlation contri
bution to the potential is energy, is a little more than −2T
C
.
• The magnitude of all energy components grows with Z.
We end with an exercise designed to make you think about the trends.
Exercise 34 Energy components for atoms
Use the data in Table 7.3 to answer the following:
1. How close are the ratios of T to V
ee
to V
ext
to their ThomasFermi ideal values? What
is the trend with Z? Comment.
2. What is the ratio of E
X
 to U? How does it change with Z? Comment.
3. Repeat above for T
C
versus T
S
. Comment.
4. What is the ratio T
C
versus U
C
? As a function of Z? Comment.
7.4 Questions
All questions are conceptual.
1. Deﬁne the groundstate KohnSham wavefunction generating density n(r).
2. Consider two normalized orbitals in one dimension, φ
1
(x) and φ
2
(x), and the density
n = φ
1

2
+ φ
2

2
. What is the KohnSham kinetic energy of that density? How does
it change when we alter one of the orbitals? Repeat the question for the KohnSham
exchange energy.
70 CHAPTER 7. KOHNSHAM
3. The exchange energy of the orbitals is the same in HartreeFock and KohnSham theory.
What is the relation between the two for the HartreeFock density?
4. Considering the exact relations for the asymptotic decay of the density of Coulombic
systems, is there any signiﬁcance to
HOMO
for the KS system? (See section 14.2).
Chapter 8
The local density approximation
1
This chapter deals with the mother of all density functional approximations, namely
the local approximation, for exchange and correlation.
8.1 Local approximations
To illustrate why a local approximation is the ﬁrst thing to try for a functional, we return
to the problem in Chapter 2 of the perimeter of a curve r(θ). Suppose we live in a ﬂatland
in which it is very important to estimate the perimeters of curves, but we haven’t enough
calculus to ﬁgure out the exact formula.
To study the problem, we ﬁrst consider a smoother class of curves,
r(θ) = r
0
(1 + cos(nθ)). (8.1)
where n = 1, 2, 3... is chosen as an integer to preserve periodicity. Here determines how
large the deviation from a circle is, while n is a measure of how rapidly the radius changes
with angle. To illustrate, we show in Fig. 8.1 the case where n = 20 and = 0.1.
smooth curves
n=20, eps=0.1
circle
Figure 8.1: Smooth parametrized curve, r = 1 + 0.1 cos(20θ).
1
c 2000 by Kieron Burke. All rights reserved.
71
72 CHAPTER 8. THE LOCAL DENSITY APPROXIMATION
Now, in ﬂatland, we wish to approximate the perimeter with a local functional of r, so we
write
P
loc
=
2π
0
dθ f(r) (8.2)
To determine the function f, we note that the perimeter of a circle is 2πr. The only place
our approximation can be exact is for the circle itself, since its r is independent of θ. Thus
we should choose f = r if we want our functional to be correct whenever it can be. Thus
P
loc
=
2π
0
dθ r(θ) (8.3)
is the local approximation to the perimeter. Our more advanced mathematicians might be
able to prove simple rules, such as P
loc
≤ P, etc. Also, we can test this approximation on
the squares we really care about, and ﬁnd, for the unit square, P
loc
= 4 ln(1 +
√
2), a 12%
underestimate, just like the kinetic energy approximation from Chapter 1.
8.2 Local density approximation
So now its simple to see how to construct local density approximations.
Recall the example of Chapter 1, the noninteracting kinetic energy of samespin electrons,
as a warmup exercise. We ﬁrst write
T
loc
S
[n] =
∞
−∞
dx t
s
(n(x)) (8.4)
Next we deduced that t
s
, the kinetic energy density, must be proportional to n
3
, from dimen
sional analysis, i.e.,
t
s
(n) = a
s
n
3
(8.5)
Last, we deduced a
s
by ensuring T
loc
S
is exact for the only system for which it can be exact,
namely a uniform gas of electrons. We did this by looking at the leading contribution to the
energy of the electrons as the number became large. This yielded a
s
= π
2
/6.
At this point, we introduce a useful concept, called the Fermi wavevector. This is the
wavevector of the highest occupied orbital in our system. Since we have N electrons in the
box,
k
F
=
Nπ
L
= nπ (1D polarized) (8.6)
Then
T
loc
S
=
∞
−∞
dx n(x) τ
S
(n(x)), τ
S
(n) = k
2
F
/6 (8.7)
Thus, for 3D Coulombinteracting problems, we write
E
loc
XC
[n] =
d
3
r f(n(r)) (8.8)
where f(n) is some function of n. To determine this function, we look to the uniform electron
gas.
8.3. UNIFORM ELECTRON GAS 73
8.3 Uniform electron gas
The local approximation is exact for the special case of a uniform electronic system, i.e.,
one in which the electrons sit in an inﬁnite region of space, with a uniform positive external
potential, chosen to preserve overall charge neutrality. The kinetic and exchange energies
of such a system are easily evaluated, since the KohnSham wavefunctions are simply Slater
determinants of plane waves. The correlation energy is extracted from accurate Monte Carlo
calculations, combined with known exact limiting values.
If we repeat the exercise above in three dimensions, we ﬁnd that it is simpler to use
planewaves with periodic boundary conditions. The states are then ordered energetically by
momentum, and the lowest occupied levels form a sphere in momentumspace. Its radius is
the Fermi wavevector, and is given by
k
F
= (3π
2
n)
1/3
, (8.9)
where the diﬀerent power and coeﬃcient come from the diﬀerent dimensionality. Note that we
have now doublyoccupied each orbital, to account for spin. Dimensional analysis produces
the same form for the kinetic energy as in 1D, but matching to the uniform gas yields a
diﬀerent constant:
T
TF
S
[n] =
3
10
d
3
r k
2
F
(r)n(r) = A
S
d
3
r n
5/3
(r) (3D unpolarized) (8.10)
where A
s
= (3/10)(3π
2
)
2/3
= 2.871.
For the exchange energy of the uniform gas, we simply note that the Coulomb interaction
has dimensions of inverse length. Thus we know that its energy density must be proportional
to k
F
, leading to
E
LDA
X
[n] = A
X
d
3
r n
4/3
(r) (8.11)
Evaluation of the Fock integral Eq. (5.13) for a Slater determinant of planewave orbitals
yields the exchange energy per electron of a uniform gas as
unif
X
(n) =
3k
F
4π
, (8.12)
so that A
X
= −(3/4)(3/π)
1/3
= −0.738.
Correlation is far more sophisticated, as it depends explictly on the physical groundstate
wavefunction of the uniform gas. Another useful measure of the density is the WignerSeitz
radius
r
s
=
3
4πn
1/3
=
1.919
k
F
(8.13)
which is the radius of a sphere around electron such that the volume of all the spheres matches
the total density of electrons. Thus r
S
→0 is the highdensity limit, and r
s
→∞ is the low
density limit.
74 CHAPTER 8. THE LOCAL DENSITY APPROXIMATION
We thus write
E
LDA
C
=
d
3
r n(r)
unif
C
(r
s
(r)) (8.14)
where
unif
C
(r
s
) is the correlation energy per electron of the uniform gas. A simple way to
think about correlation is simply as an enhancement over exchange:
unif
XC
(r
s
) = F
XC
(r
s
)
unif
X
(r
s
) (8.15)
and this enhancement factor is plotted in Fig. 15.1. There are several important features to
the curve:
• At r
s
= 0 (inﬁnite density), exchange dominates over correlation, and F
X
= 1.
• As r
s
→0 (high density), there is a sharp dive toward 1. This is due to the longranged
nature of the Coulomb repulsion in an inﬁnite system, leading to
unif
C
(r
s
) →0.0311 ln r
s
−0.047 + 0.009r
s
ln r
s
−0.017r
s
(r
s
→0) (8.16)
The logarithmic divergence is not enough to make correlation as large as exchange in this
limit, but we will see later that ﬁnite systems have ﬁnite correlation energy in this limit.
• In the large r
s
limit (lowdensity)
unif
C
(r
s
) →−
d
0
r
s
+
d
1
r
3/2
s
−.... (r
s
→∞) (8.17)
where the constants are d
0
= 0.896 and d
1
= 1.325. The constant d
0
was ﬁrst deduced
by Wigner from the Wigner crystal for this system. Note that this means correlation
becomes (almost) as large as exchange here, and so F
unif
XC
(r
s
→∞) = 1.896. Note also
that the approach to the lowdensity limit is extremely slow.
0.3
0.25
0.2
0.15
0.1
0.05
0
0 1 2 3 4 5 6
uniform gas
C
X
Figure 8.2: Exchange and correlation energies per particle for a uniform electron gas.
The uniform gas exchange and correlation energies/particle are plotted in Fig 8.2, as a
function of r
s
. The exchange is very simple, being proportional to 1/r
s
. The sharp downturn
8.3. UNIFORM ELECTRON GAS 75
in the correlation as r
s
→0 is due to the logarithmic term (but is not as singular as exchange,
which is diverging as 1/r
s
). At large r
s
, the two curves become comparable, but r
s
must be
larger than shown here. For typical valence electrons, r
s
is between 1 and 6, and correlation
is much smaller than exchange.
Over the years, this function has become very wellknown, by combining limiting informa
tion like that above, with accurate quantum Monte Carlo data for the uniform gas. An early
popular formula in solidstate physics is PZ81, while the parametrization of VoskoWilkes
Nusair (VWN) has been implemented in quantum chemical codes. Note that in the Gaussian
codes, VWN refers to an older formula from the VWN paper, while VWNV is the actual
parametrization recommended by VWN. More recently, Perdew and Wang reparametrized the
data. These accurate parametrizations diﬀer only very slightly among themselves.
An early but inaccurate extrapolation from the lowdensity gas, was given by Wigner:
e
C
(r
s
) = −a/(b +r
s
) (Wigner) (8.18)
where a = −.44 abd b = 7.8. This gives a ballpark number, but misses the logarithmic
singularity as r
s
→0.
Unfortunately, we must postpone a general review of the performance of LDA until after
the next chapter. This is because we have only developed density functional theory so far,
and not accounted for spin eﬀects. A slight generalization of LDA, called the local spin
density approximation, is what is actually used in practice. In fact, many (if not most)
practical problems require this. For example, pure LDA performs badly whenever there is an
odd number of electrons, since it makes no distinction between polarized and unpolarized
densities. We emphasize this fact in the next few exercises.
Exercise 35 Hartree energy for an exponential
1. For an exponential density, n(r) =
Z
3
/π exp(−2Zr), calculate (a) the Hartree poten
tial and (b) the Hartree energy.
2. Find the exact exchange energy for the Hydrogen atom.
3. Find the exact exchange energy for the Helium atom, using the exponential approximation
of Chapter 4.
Exercise 36 LDA exchange energy for 1 or 2 electrons
1. Calculate the LDA exchange energy for the Hydrogen atom, and compare with exact
answer.
2. Find the LDA exchange energy for the Helium atom, using the exponential approximation
of Chapter X, and compare with exact answer.
76 CHAPTER 8. THE LOCAL DENSITY APPROXIMATION
8.4 Questions
1. Following the previous question, can you deduce the asymptotic form of the exchange
correlation potential, i.e., v
XC
(r) as r → ∞? Would a local approximation capture this
behavior? (See section 14.2).
Chapter 9
Spin
In practice, people don’t use KohnSham density functional theory, they use KohnSham
spindensity functional theory. For many problems, separate treatment of the up and
down spin densities yields much better results.
9.1 KohnSham equations
We next introduce spin density functional theory, a simple generalization of density functional
theory. We consider the up and down densities as separate variables, deﬁned as
n
σ
(r) = N
dx
2
. . .
dx
N
Ψ(r, σ, x
2
, . . . , x
N
)
2
, (9.1)
with the interpretation that n
σ
(r)d
3
r is the probability for ﬁnding an electron of spin σ
in d
3
r around r. Then the HohenbergKohn theorems can be proved, showing a oneto
one correspondence between spin densities and spindependent external potentials, v
ext,σ
(r).
1
Similarly, the KohnSham equations can be developed with spindependent KohnSham
potentials. All modern density functional calculations are in fact spindensity functional
calculations. This has several advantages:
1. Systems in collinear magnetic ﬁelds are included.
2. Even when the external potential is not spindependent, it allows access to magnetic
response properties of a system.
3. Even if not interested in magnetism, we will see that the increased freedom in spin DFT
leads to more accurate functional approximations for systems that are spinpolarized,
e.g., the Li atom, whose spin densities are plotted in Fig. 9.1.
We will see in the next sections that it is straightforward to turn some density functionals
into spindensity functionals, while others are more complicated.
1
This is not quite true. There is a little more wriggle room than in the DFT case. For example, for a spinup H atom, the spin down potential
is undeﬁned, so long as it does not produce an orbital whose energy is below 1/2.
77
78 CHAPTER 9. SPIN
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
0 1 2 3 4 5 6 7 8
Li atom
up
down
Figure 9.1: Radial spin densities in the Li atom.
9.2 Spin scaling
We can easily deduce the spin scaling of orbital functionals, while little is known of the exact
spinscaling of correlation in general. By orbital functionals, we mean those that are a sum
of contributions from the orbitals of each spin seperately.
We do spinscaling of the kinetic energy functional as an example. Suppose we know T
S
[n]
as a density functional for spinunpolarized systems. Let’s denote it T
unpol
S
[n] to be very
explicit. But since the kinetic energy is the sum of contributions from the two spin channels:
T
S
[n
↑
, n
↓
] = T
S
[n
↑
, 0] +T
S
[0, n
↓
] (9.2)
i.e., the contributions come from each spin separately. Applying this to the spinunpolarized
case, we ﬁnd
T
unpol
S
[n] = T
S
[n/2, n/2] = 2T
S
[n/2, 0] (9.3)
or T
S
[n, 0] = T
unpol
S
[2n]/2. Inserting this result back into Eq. (9.2), we ﬁnd
T
S
[n
↑
, n
↓
] =
1
2
T
unpol
S
[2n
↑
] +T
unpol
S
[2n
↓
]
¸
(9.4)
This clearly yields a consistent answer for an unpolarized system, and gives the result for a
fullypolarized system:
T
pol
S
[n] = T
unpol
S
[2n]/2 (9.5)
Analogous formulas apply to exchange:
E
X
[n
↑
, n
↓
] =
1
2
E
unpol
X
[2n
↑
] +E
unpol
X
[2n
↓
]
¸
(9.6)
and
E
pol
X
[n] = E
unpol
X
[2n]/2 (9.7)
9.3. LSD 79
9.3 LSD
The local spin density (LSD) approximation is the spinscaled generalization of LDA. As input,
we need the energies per particle of a spinpolarized uniform gas. The formula for the exchange
energy is straightforward, but the correlation energy, as a function of spinpolarization, for
the uniform gas, must be extracted from QMC and known limits.
Begin with exchange. In terms of the energy density of a uniform gas, Eq. (9.6) becomes:
e
X
(n
↑
, n
↓
) =
1
2
e
unpol
X
(2n
↑
) +e
unpol
X
(2n
↓
)
¸
(9.8)
Now we know e
unpol
X
(n) = A
X
n
4/3
, so
e
X
(n
↑
, n
↓
) = 2
1/3
A
X
n
4/3
↑
+n
4/3
↓
(9.9)
Now we introduce a new, useful concept, that of relative spin polarization. Deﬁne
ζ =
n
↑
−n
↓
n
(9.10)
as the ratio of the spin diﬀerence density to the total density. Thus ζ = 0 for an unpolarized
system, and ±1 for a fully polarized system. In terms of ζ,
n
↑
= (1 + ζ) n/2, n
↓
= (1 −ζ) n/2, (9.11)
Inserting these deﬁnition into Eq. (9.9), we ﬁnd, for the energy density:
X
(n, ζ) =
unpol
X
(n)
(1 + ζ)
4/3
+ (1 −ζ)
4/3
2
(9.12)
Thus our LSD exchange energy formula is
E
LSD
X
[n
↑
, n
↓
] = A
X
d
3
r n
4/3
(r)
(1 + ζ(r))
4/3
+ (1 −ζ(r))
4/3
2
(9.13)
where ζ(r) is the local relative spin polarization:
ζ(r) =
n
↑
(r) −n
↓
(r)
n(r)
(9.14)
Unfortunately, the case for correlation is much more complicated, and we merely say that
there exists wellknown parametrizations of the uniform gas correlation energy as a function of
spin polarization,
unif
C
(n, ζ). In Fig. 9.2, we show what happens. When the gas is polarized,
exchange gets stronger, because electrons exchange only with like spins. On the other hand,
exchange keeps the electrons further apart, reducing the correlation energy. This eﬀect is
typical of most systems, but this has not been proven generally.
Exercise 37 Relative spin polarization
Sketch ζ(r) for the Li atom, as shown in Fig. 9.1.
Exercise 38 Spin polarized ThomasFermi
Convert the ThomasFermi local approximation for the kinetic energy in 3D to a spindensity
functional. Test it on the hydrogen atom, and compare to the regular TF.
80 CHAPTER 9. SPIN
0.3
0.25
0.2
0.15
0.1
0.05
0
0 1 2 3 4 5 6
C
X
uniform gas
pol
unpol
Figure 9.2: Eﬀect of (full) spin polarization on exchange and correlation energies of the uniform electron gas.
9.4 Questions
All are conceptual.
1. Why does LSD yield a more accurate energy for the Hydrogen atom than LDA?
2. State in words the relation between v
S
[n](r) and v
Sσ
[n
↑
, n
↓
](r). When do they coincide?
3. Deduce the formula of T
loc
S
for unpolarized electrons in a onedimensional box.
Chapter 10
Properties
In this chapter, we survey some of the more basic properties that are presently being
calculated with electronic structure methods, and include how well or badly LDA does for
these. The local density approximation to exchangecorrelation was introduced by Kohn
and Sham in 1965, and (the spindependent generalization) has been one of the most
succesful approximations ever. Until the early 90’s, it was the standard approach for all
density functional calculations, which were often called ab initio in solid state physics. It
remains perhaps the most reliable approximation we have.
10.1 Total energies
For atoms and molecules, the total exchange energy is typically underestimated by about
10%. On the other hand, the correlation energy is overestimated by about a factor of 2 or 3.
Since for many systems of physical and chemical interest, exchange is about 4 times bigger
than correlation, the overestimate of correlation compliments the underestimate of exchange,
and the net exchangecorrelation energy is typically underestimated by about 7%.
Exercise 39 Wigner approximation for He atom correlation energy:
By approximating the density of the He atom as a simple exponential with Z
eff
= 7/4 as we
found in Chapter 4, calculate the LDA correlation energy using the Wigner approximation,
Eq. (8.18).
E
X
E
C
E
XC
atom LDA exact error % LDA exact error % LDA exact error %
He 0.883 1.025 14 % 0.112 0.042 167 % 0.996 1.067 7 %
Be 2.321 2.674 13 % 0.225 0.096 134 % 2.546 2.770 8 %
Ne 11.021 12.085 9 % 0.742 0.393 89 % 11.763 12.478 6 %
Table 10.1: Accuracy of LDA for exchange, correlation, and XC for noble gas atoms.
81
82 CHAPTER 10. PROPERTIES
10.2 Densities and potentials
The selfconsistent LSD density is always close to the exact density, being hard to distinguish
by eye, as seen in Fig. 10.1. However, when we calculate κ(r), as deﬁned earlier, we can
0
2
4
6
8
10
12
0 0.5 1 1.5 2
Ne atom
exact
LDA
4
π
r
2
n
(
r
)
r
Figure 10.1: Radial density of the Ne atom, both exactly and from an LDA calculation
see the diﬀerence in the asymptotic region: Note that the density integrates to 9.9 at about
2
1.8
1.6
1.4
1.2
1
0 1 2 3 4 5
Ne atom
exact
LDA
κ
(
r
)
r
Figure 10.2: Logarithmic derivative of density for Ne atom, both exactly and in LDA. The curves are indistinguishable close
to the nucleus, where they both reach 10.
r = 2.68. So for all the occupied region, even κ(r) is extremely accurate in LDA, and all
integrals over the density are very accurate. For example, 1/r is 31.1 exactly, but 31.0 in
LDA.
The KohnSham potentials reﬂect this similarity. In Fig. 10.3, we show both the LDA
and exact potentials, and how close they appear. However, note that they are dominated
by terms that are (essentially) identical in the two cases. The largest term is the external
potential, which is −10/r. This determines the scale of the ﬁgure. The next term is the
Hartree potential, plotted in Fig. 10.4. This is essentially the same in both cases, since the
densities are almost identical. Again note the scale.
10.3. IONIZATION ENERGIES AND ELECTRON AFFINITIES 83
100
80
60
40
20
0
0 0.2 0.4 0.6 0.8 1
Ne atom
exact
LDA
v
S
(
r
)
r
Figure 10.3: KohnSham potential of Ne atom, both exactly and in LDA.
0
5
10
15
20
25
30
35
40
0 0.5 1 1.5 2 2.5 3 3.5 4
Ne atom
exact
LDA
v
H
(
r
)
r
Figure 10.4: Hartree potential of Ne atom, both exactly and in LDA.
The LDA exchangecorrelation potential, however, looks very diﬀerent from the exact
quantities for any ﬁnite systems, as shown in Fig. 10.5. This in turn means that the
orbital eigenvalues can be very diﬀerent from exact KohnSham eigenvalues. Thus ionization
potentials from orbital energy diﬀerences are very poor. This will be discussed in great detail
in chapter X.
10.3 Ionization energies and electron aﬃnities
In Table ??, we list ionization potentials for the ﬁrst twenty atoms. There are many things
to see in this table:
• Comparing the exchange results (which are essentially identical to HF) with the exact
results, we ﬁnd that HF underestimates ionization potentials by about 1 eV . In fact, the
mean error is 0.9 eV.
• Repeating with the LSD numbers, we ﬁnd that the errors vary in sign. Now we must
84 CHAPTER 10. PROPERTIES
atom LSDX X LSD exact
h 12.40 13.61 13.00 13.61
he 22.01 23.45 24.28 24.59
li 5.02 5.34 5.46 5.39
be 7.63 8.04 9.02 9.32
b 7.52 7.93 8.57 8.30
c 10.72 10.81 11.73 11.26
n 13.91 14.00 14.93 14.55
o 11.78 11.87 13.85 13.62
f 16.14 15.66 17.98 17.45
ne 20.35 19.81 22.07 21.62
na 4.84 4.94 5.32 5.13
mg 6.48 6.59 7.70 7.64
al 5.14 5.49 5.99 5.99
si 7.40 7.65 8.26 8.16
p 9.63 10.04 10.51 10.53
s 8.79 9.03 10.52 10.37
cl 11.67 11.77 13.22 12.98
ar 14.42 14.76 15.90 15.84
Table 10.2: Ionization energies of ﬁrst twenty atoms, nonrelativistic, using exactexchange densities, in eV. The X results
use Engel’s code, and the LSD results are evaluated on those densities. The ’exact’ results are from Davidson.
10.4. DISSOCIATION ENERGIES 85
8
7
6
5
4
3
2
1
0
0 0.2 0.4 0.6 0.8 1
Ne atom
exact
LDA
v
X
C
(
r
)
r
Figure 10.5: Exchangecorrelation potential of Ne atom, both exactly and in LDA.
take the mean absolute error (MAE), which is 0.25 eV. This is much better than HF,
but not accurate enough for many purposes.
• Interestingly, if we compare LSDX with the exact exchange numbers, we ﬁnd that LSDX
almost always underestimates the ionization energy, and has an MAE of 0.4 eV. Thus
LSDX is a poorer approximation to exact exchange than LSD is to the exact result.
Again, there is a cancellation of errors between exchange and correlation.
Electron aﬃnities
Electronegativity and Hardness
10.4 Dissociation energies
We consider ﬁrst atomization energies, being the energy diﬀerence between molecules or
solids at equilibrium, and their constituent atoms. These are called cohesive energies of
solids. To calculate these, the zeropoint energy of the molecule (or solid) must be added to
the minimum in its total energy curve. Typically, HF underestimates such binding energies
by about 100 kcal/mol, while LDA overbinds by about 30 kcal/mol, or 1.3 eV, or about 50
millihartrees. This meant LDA was never adopted as a general tool in quantum chemistry. For
similar reasons, this also leads to transition state barriers that are too low, often nonexistent.
Finally, qualitative errors occur for highly correlated systems, such as the solid NiO or the
molecules Cr
2
, or H
2
stretched to large distances. These systems are extremely diﬃcult to
study with density functional approximations, as will be discussed below.
10.5 Geometries and vibrations
However, bond lengths are extremely good in LDA, usually (but not always) being underes
timated by about 12%.
86 CHAPTER 10. PROPERTIES
10.6 Transition metals
In solid state physics, an infamous failure is the prediction that the nonmagnetic structure
of iron is of slightly lower energy than the magnetic one.
In general, density functionals are much less reliable for transition metal complexes than
ﬁrst and secondrow elements, with bondlength errors more likely in the 0.05
˚
Arange.
10.7 Weak bonds
The remarks above apply to covalent, metallic, and ionic bonds. But many processes are
determined by much weaker bonds, such as hydrogen bonds and dispersion forces. LDA
signiﬁcantly overbinds both hydrogen bonds, and van der Waals dimers.
Between two isolated pieces of matter (no signiﬁcant density overlap) there are always van
der Waals forces due to ﬂuctuating dipoles. In absence of permanent dipoles, and ignoring
relativistic eﬀects, the energy between two such pieces behaves as
E
bind
= −
C
6
R
6
, R →∞ (10.1)
where R is their separation. The exchange energy is always repulsive, so the existence of C
6
is a purely correlation eﬀect.
LDA cannot reproduce this asymptotic behavior, since any contribution to the energy
diﬀerence must come from a density diﬀerence. Thus the LDA correlation energy vanishes
exponentially in this limit.
10.8 Gaps
On the other hand, for solids, these eigenvalues are often plotted as the band structure in
solidstate texts. In these cases, the overall shape and position is good, but the band gap,
between HOMO and LUMO, is consistently underestimated by at least a factor of 2. In some
cases, some semiconductors have no gap in LDA, so it makes the incorrect prediction that
they are metals.
10.9 Questions
1. Does the reliability of LSD show that most systems are close to uniform?
Part III
Analysis
87
Chapter 11
Simple exact conditions
1
In this chapter, we introduce various simple exact conditions, and see how well LDA
meets these conditions.
11.1 Size consistency
An important exact property of any electronic structure method is size consistency. Imagine
you have a system which consists of two extremely separated pieces of matter, called A and
B. Their densities are given, n
A
(r) and n
B
(r), and the overlap is negligible. Then a size
consistent treatment should yield the same total energy, whether the two pieces are treated
separately or as a whole, i.e.,
E[n
A
+n
B
] = E[n
A
] +E[n
B
]. (11.1)
Conﬁguration interaction calculations with a ﬁnite order of excitations are not sizeconsistent,
but coupledcluster calculations are. Most density functionals are sizeconsistent, but some
with explicit dependence on N are not. Any local or semilocal functional is sizeconsistent.
11.2 One and two electrons
For any one electron system
V
ee
= 0 E
X
= −U E
C
= 0 (N = 1) (11.2)
This is one of the most diﬃcult properties for local and semilocal density functionals to get
right, because the exact Hartree energy is a nonlocal functional. The error made for one
electron systems is called the selfinteraction error, as it can be thought of as arising from
the interaction of the charge density with itself in U. By spinscaling, we ﬁnd
E
X
= −U/2 (N = 2, unpolarized) (11.3)
for two spinunpolarized electrons. LSD does have a selfinteraction error, as can be seen for
the H atom results in the tables.
1
c 2000 by Kieron Burke. All rights reserved.
89
90 CHAPTER 11. SIMPLE EXACT CONDITIONS
Exercise 40 FermiAmaldi correction
An approximation for the exchange energy is E
X
[n] = −U[n]/N. This is exact for one or
two (spinunpolarized) electrons. Show that it is not sizeconsistent.
This problem becomes acute when there is more than one center in the external potential.
Consider, e.g. H
+
2
. As the nuclear separation grows from zero, the charge density spreads out,
so that a local approximation produces a smaller and smaller fraction of the correct result.
Exercise 41 Stretched H
+
2
Calculate the LSD error in the exchange energy of H
+
2
in the limit of inﬁnite separation.
11.3 LiebOxford bound
Another exact inequality for the potential energy contribution to the exchangecorrelation
functional is the LiebOxford bound:
U
XC
[n] ≥ C
LO
E
LDA
X
[n] (11.4)
where C
LO
≤ 2.273. It takes a lot of work to prove this result, but it provides a strong bound
on how negative the XC energy can become.
Exercise 42 LiebOxford bound
What does the LiebOxford bound tell you about
unif
C
(r
s
)? Does LDA respect the LiebOxford
bound?
11.4 Bond breaking
Another infamous diﬃculty of DFT goes all the way back to the ﬁrst attempts to understand
H
2
using quantum mechanics. In the absence of a magnetic ﬁeld, the ground state of two
electrons is always a singlet. At the chemical bond length, a HartreeFock single Slater deter
minant is a reasonable approximation to the true wavefunction, and yields roughly accurate
numbers:
Φ
HF
(r
1
, r
2
) = φ(r
1
)φ(r
2
) ≈ Ψ(r
1
, r
2
) (11.5)
However, when the bond is stretched to large distances, the correct wavefunction becomes
a linear combination of two Slater determinants (the HeitlerLondon wavefunction), one for
each electron on each Hatom:
Ψ(r
1
, r
2
) =
1
√
2
(φ
A
(r
1
)φ
B
(r
2
) + φ
B
(r
1
)φ
A
(r
2
)) , (11.6)
where φ
A
is an atomic orbital centered on nucleus A and similarly for B. The total density is
just n(r) = n
A
(r) +n
B
(r), where n
I
(r) is the hydrogen atom density centered on nucleus I.
But note that the density remains unpolarized.
11.5. UNIFORM LIMIT 91
If we apply LSD to the exact density, we make a large error, because E
LSD
[n] = E
LDA
[n
A
]+
E
LDA
[n
B
], and we saw in a previous chapter the huge error LDA makes for the Hatom,
because it does not account for spinpolarization. We would much rather ﬁnd E
LSD
=
2E
LSD
(H).
There is a way to do this, that has been wellknown in quantum chemistry for years. If
one allows a HartreeFock calculation to spontaneously break spinsymmetry, i.e., do not
require the upspin spatial orbital to be the same as that of the downspin, then at a certain
separation, called the CoulsonFischer point, the unrestricted solution will have lower energy
than the restricted one, and will yield the correct energy in the separatedatom limit, by
producing two H atoms that have opposite spins. Thus we speak of RHF and UHF. The
trouble with this is that the UHF solution is no longer a pure spineigenstate. Thus the
symmetry dilemma is that an RHF calculation produces the correct spin symmetry with the
wrong energy, while the UHF solution produces the right energy but with the wrong spin
symmetry.
KohnSham DFT calculations with approximate density functionals have a similar problem.
The LSD calculation will spontaneously break symmetry at a CoulsonFischer point (further
than that of HF) to yield the appropriate energy. Now the dilemma is no longer the spin
eigenvalue of the wavefunction, but rather the spindensities themselves. In the unrestricted
LDA calculation, which correctly dissociates to two LSD H atoms, the spindensities are
completely polarized, which is utterly incorrect.
11.5 Uniform limit
We have already used the argument that any local approximation must use the value for the
uniform limit as its argument. But more generally, we require
E
XC
[n] →V e
unif
XC
(n) n(r) →n (constant) (11.7)
where V is the volume of the system. This can be achieved, for example, by taking ever
larger boxes, but keeping the density of particles ﬁxed. This limit is obviously gotten right by
LDA, but only if we use the correct energy density. Getting this right is relevant to accuracy
for valence electrons of simple bulk metals, such as Li or Al.
11.6 Questions
1. If I approximate the correlation energy as about 1 eV per electron, is that a sizeconsistent
approximation?
2. A spinrestricted LSD calculation for two H atoms 10
˚
Aapart yields a diﬀerent answer to
twice the answer for one H atom. Does this mean LSD is not size consistent?
3. Does LDA satisfy the LO bound pointwise? Does LSD?
92 CHAPTER 11. SIMPLE EXACT CONDITIONS
4. What does the LO bound say for one electron?
Chapter 12
Scaling
1
In this chapter, we introduce the concept of scaling the density, as the natural way to
understand the most important limits about density functionals.
12.1 Wavefunctions
It will prove very useful to ﬁgure out what happens to various quantities when the coordinates
are scaled, i.e., when x is replaced by γx everywhere in a problem, with γ a positive number.
For one electron in one dimension, we deﬁne the scaled wavefunction as
φ
γ
(x) = γ
1
2
φ(γx), 0 < γ < ∞, (12.1)
where the scaling factor out front has been chosen to preserve normalization:
1 =
∞
−∞
dx φ(x)
2
=
∞
−∞
dx γφ(γx)
2
=
∞
−∞
dx
φ(x
)
2
. (12.2)
A scale factor of γ > 1 will squeeze the wavefunction, and of γ < 1 will stretch it out.
For example, consider φ(x) = exp(−x). Then φ
γ
(x) =
√
γ exp(−γx) is a diﬀerent
wavefunction for every γ. For example, φ
2
(x) =
√
2 exp(−2x). In Fig. 12.1, we sketch
a density for the onedimensional H
2
molecule of a previous exercise, and the same density
scaled by γ = 2,
n
γ
(x) = γn(γx). (12.3)
Note how γ > 1 squeezes the density toward the origin, keeping the number of electrons
ﬁxed. We always use subscripts to denote a function which has been scaled.
What is the kinetic energy of such a scaled wavefunction?
T[φ
γ
] =
1
2
∞
−∞
dx φ
γ
(x)
2
=
γ
2
∞
−∞
dx φ
(γx)
2
=
γ
2
2
∞
−∞
dx
φ
(x
)
2
= γ
2
T[φ], (12.4)
i.e., the kinetic energy grows quadratically with γ. The potential energy is not so straightfor
ward in general, because it is not a universal functional, i.e., it is diﬀerent for every diﬀerent
problem.
V [φ
γ
] =
∞
−∞
dx φ
γ
(x)
2
V (x) =
∞
−∞
dx
φ(x
)
2
V (
x
γ
) = V (x/γ). (12.5)
1
c 2000 by Kieron Burke. All rights reserved.
93
94 CHAPTER 12. SCALING
!5.0 5.0
0.0
2.0
Figure 12.1: Cartoon of the onedimensional H2 molecule density, and scaled by a factor of 2.
However, if the potential energy is homogeneous of degree p, i.e.,
V (γx) = γ
p
V (x), (12.6)
then
V [φ
γ
] = γ
−p
V [φ]. (12.7)
Consider any set of trial wavefunctions. If scaling any member of the set leads to another
member of the set, then we say that the set admits scaling.
Exercise 43 Scaling of trial wavefunctions
Do the sets of trial wavefunctions in Ex. 10 admit scaling?
The only important change when going to three dimensions for our purposes in the pre
ceding discussion is that the scaling normalization factors change:
φ
γ
(r) = γ
3
2
φ(γr) n
γ
(r) = γ
3
n(γr) (12.8)
When including many electrons, one gets a similar factor for each coordinate:
Ψ
γ
(r
1
, . . . , r
N
) = γ
3N/2
Ψ(γr
1
, . . . , γr
N
), (12.9)
suppressing spin indices. We can they easily evaluate our favorite operators. It is simple to
show:
T[Ψ
γ
] = γ
2
T[Ψ] (12.10)
and
V
ee
[Ψ
γ
] = γ V
ee
[Ψ] (12.11)
while V
ext
has no simple rule in general.
12.2. DENSITY FUNCTIONALS 95
12.2 Density functionals
Now we turn our attention to density functionals. Functionals deﬁned as explicit operators
on the KohnSham wavefunction are far simpler than those including correlation eﬀects. In
particular, there is both the noninteracting kinetic energy and the exchange energy. We shall
see that both their uniform scaling and their spinscaling are straightforward.
Consider uniform scaling of an Nelectron wavefunction, as in Eq. (12.9). The density of
the scaled wavefunction is
n(r) = N
d
3
r
2
. . .
d
3
r
N
Ψ
γ
(r, r
2
, . . . , r
N
)
2
, = γ
3
n(γr) = n
γ
(r). (12.12)
Now, a key question is this. If Ψ[n] is the groundstate wavefunction with density n(r), is
Ψ[n
γ
] = Ψ
γ
[n]? That is, is the scaled wavefunction the same as the wavefunction of the
scaled density? We show below that the answer is no for the physical wavefunction, but is
yes for the KohnSham wavefunction.
Consider the latter case ﬁrst. We already know that T
S
[Φ
γ
] = γ
2
T
S
[Φ]. Thus if Φ minimizes
T
S
and yields density n, then Φ
γ
also minimizes T
S
, but yields density n
γ
. Therefore, Φ
γ
is
the KohnSham wavefunction for n
γ
, or
Φ
γ
[n] = Φ[n
γ
]. (12.13)
This result is central to understanding the behaviour of the noninteracting kinetic and ex
change energies. We can immediately use it to see how they scale, since we can turn density
functionals into orbital functionals, and vice versa:
T
S
[n
γ
] = T
S
[Φ[n
γ
]] = T
S
[Φ
γ
[n]] = γ
2
T
S
[Φ[n]] = γ
2
T
S
[n], (12.14)
using Eq. (12.24), and
E
X
[n
γ
] = E
X
[Φ[n
γ
]] = E
X
[Φ
γ
[n]] = γE
X
[Φ[n]] = γE
X
[n], (12.15)
using Eq. (F.15), and the fact that
U[n
γ
] = γ U[n] (12.16)
These exact conditions are utterly elementary, saying only that if the length scale of a sys
tem changes, these functionals should change in a trivial way. Any approximation to these
functionals must therefore satisfy these relations. However, they are also extremely powerful
in limiting the possible forms functional approximations can have.
The next exercise shows us that scaling relations determine the functional forms of the
local approximation to these simple functionals, and the uniform electron gas is referred to
only for the values of the coeﬃcients.
Exercise 44 Local density approximations for T
S
and E
X
:
Show that, in making a local approximation for threedimensional systems, the power of the
density in both the noninteracting kinetic and exchange energy approximations are deter
mined by scaling considerations.
96 CHAPTER 12. SCALING
12.3 Correlation
Levy has shown that for ﬁnite systems, E
C
[n
γ
] tends to a negative constant as γ →∞. From
Fig. 12.2, we see that for the He atom E
C
varies little with scaling. We may write power
0.3
0.25
0.2
0.15
0.1
0.05
0
1 2 3 4 5 6 7 8 9 10
He atom
exact
LDA
E
C
[
n
γ
]
γ
Figure 12.2: Correlation energy of the He atom, both exactly and within LDA, as the density is squeezed.
series for E
C
[n
γ
] around the highdensity limit:
E
C
[n
γ
] = E
(2)
C
[n] +E
(3)
C
[n]/γ +. . . , γ →∞ (12.17)
We can easily scale any approximate functional, and so extract the separate contributions
to the correlation energies, and test them against their exact counterparts, check limiting
values, etc. A simple example is correlation in LDA. Then when n →n
γ
, r
S
→r
S
/γ. Thus
E
LDA
C
[n
γ
] =
d
3
r n
γ
(r)
unif
C
(r
s,γ
(r)) =
d
3
rn(r)
unif
C
(r
s
(r)/γ) (12.18)
In the highdensity limit, γ → ∞,
unif
C
(r
s
/γ) →
unif
C
(0). This is logarithmically singular in
true LDA, as we saw in Chapter 8, and is an error made by LDA.
From the ﬁgure, we see that for the He atom, as is the case with most systems of chemical
interest, E
C
[n
γ
] is almost constant. In fact, it is very close to linear in 1/γ, showing that the
system is close to the highdensity limit. The highdensity property is violated by LDA (see
the ﬁgure), and diverges logarithmically as γ → ∞. This is another way to understand the
LDA overestimate of correlation for ﬁnite systems. Since the LDA correlation energy diverges
as γ →∞, it will be overestimated at γ = 1.
On the other hand, the correlation energy is believed to have the following expansion under
scaling to the lowdensity limit:
E
C
[n
γ
] = γB[n] + γ
3/2
C[n] +. . . , γ →0 (12.19)
Note that in this lowdensity regime, correlation is so strong it scales the same as exchange.
This limit is satisﬁed within LDA, as seen from Eqs. (??), although it might not be numerically
terribly accurate here. In the low density limit,
unif
C
(r
s
) = −d
0
/r
s
+ d
1
/r
3/2
s
+ . . ., so that
12.4. CORRELATION INEQUALITIES 97
E
LDA
C
[n
γ
] = −γd
0
d
3
r n(r)/r
s
(r), vanishing linearly with γ, correctly. But most systems of
chemical interest are closer to the highdensity limit, where LDA fails.
Exercise 45 Extracting exchange
If someone gives you an exchangecorrelation functional E
XC
[n], deﬁne a procedure for
extracting the exchange contribution.
As usual, one should not focus too much on the failings of the correlation energy alone.
Remember, exchange scales linearly, and dominates correlation. When we add both contri
butions, we ﬁnd Fig. 12.3. Clearly, the XC energy is best at γ = 1, and worsens as γ grows,
10
9
8
7
6
5
4
3
2
1
0
1 2 3 4 5 6 7 8 9 10
He atom
exact
LDA
E
X
C
[
n
γ
]
γ
Figure 12.3: Exchangecorrelation energy of the He atom, both exactly and within LDA, as the density is squeezed.
because of the cancellation of errors.
Exercise 46 Scaling LDA in Wigner approximation:
Using the correct exchange formula and the simple Wigner approximation for the correlation
energy of the uniform gas, Eq. (8.18), and the simple exponential density for the He atom,
calculate the curves shown in Figs. 12.2 and 12.3.
12.4 Correlation inequalities
This reasoning breaks down when correlation is included, because this involves energies eval
uated on the physical wavefunction. As mentioned above, a key realization there is that the
scaled groundstate wavefunction is not the groundstate wavefunction of the scaled density,
because the physical wavefunction minimizes both T and V
ee
simultaneously. However, we
can still use the variational principle to deduce an inequality. We write
F[n
γ
] = Ψ[n
γ
]
ˆ
T +
ˆ
V
ee
Ψ[n
γ
] ≤ Φ
γ
[n]
ˆ
T +
ˆ
V
ee
Φ
γ
[n] (12.20)
or
F[n
γ
] ≤ γ
2
T[n] + γV
ee
[n] (12.21)
98 CHAPTER 12. SCALING
This is the fundamental inequality of uniform scaling, as it tells us inequalities about how
correlation contributions scale.
To see an example, apply Eq. (12.21) to n
= n
γ
and write γ
= 1/γ, to yield
F[n
] ≤ T[n
γ
]/γ
2
+V
ee
[n
γ
]/γ
. (12.22)
But we can just drop the primes in this equation, and multiply through by γ, and add
(γ
2
−γ)T[n] to both sides:
γ
2
T[n] + γV
ee
[n] ≤ T[n
γ
]/γ +V
ee
[n
γ
] + (γ
2
−γ)T[n]. (12.23)
Now the lefthandside equals the righthandside of Eq. (12.21), so we can combine the two
equations, to ﬁnd
(γ −1) T[n
γ
] ≤ γ
2
(γ −1) T[n]. (12.24)
Note that for γ > 1, we can cancel γ − 1 from both sides, to ﬁnd T[n
γ
] < γ
2
T[n]. As
we scale the system to high density, the physical kinetic energy grows less rapidly than the
noninteracting kinetic energy. For γ < 1, the reverse is true, i.e.,
T[n
γ
] ≤ γ
2
T[n] γ > 1, T[n
γ
] ≥ γ
2
T[n] γ < 1. (12.25)
Exercise 47 Scaling V
ee
Show
V
ee
[n
γ
] ≥ γV
ee
[n] γ > 1, V
ee
[n
γ
] ≤ γV
ee
[n] γ < 1. (12.26)
These inequalities actually provide very tight bounds on these large numbers. But of
greater interest are the much smaller diﬀerences with KohnSham values, i.e., the correlation
contributions. So
Exercise 48 Scaling E
C
Show
E
C
[n
γ
] ≥ γE
C
[n], γ > 1, E
C
[n
γ
] ≤ γE
C
[n], γ < 1, (12.27)
and
T
C
[n
γ
] ≤ γ
2
T
C
[n], γ > 1, T
C
[n
γ
] ≥ γ
2
T
C
[n], γ < 1. (12.28)
Exercise 49 Scaling inequalities for LDA
Give a oneline argument for why LDA must satisfy the scaling inequalities for correlation
energies.
In fact, for most systems we study, both E
C
and T
C
are relatively insensitive to scaling the
density toward the highdensity limit, so these inequalities are less useful.
12.5. VIRIAL THEOREM 99
12.5 Virial theorem
Any eigenstate wavefunction extremizes the expectation value of the Hamiltonian. Thus any
small variations in such a wavefunction lead only to second order changes, i.e.,
E[φ
sol
+ δφ] = E[φ
sol
] +O(δφ
2
). (12.29)
Assume φ is the groundstate wavefunction of
ˆ
H. Then φ
γ
is not, except at γ = 1. But
small variations in γ near 1 lead only to second order changes in
ˆ
H , so
d
dγ

γ=1
φ
γ

ˆ
H  φ
γ
= 0. (12.30)
But, from Eq. (12.4), dT[φ
γ
]/dγ = 2γT[φ], and from Eq. (12.5), dV [φ
γ
]/dγ = −γ
−2
x
dV
dx
at γ = 1, yielding
2T = x
dV
dx
. (12.31)
The righthand side is called the virial of the potential. So this must be satisﬁed by any
solution to the Schr¨ odinger equation, and can be used to test how accurately an approximate
solution satisﬁes it.
Note that this is true for any eigenstate and even for any approximate solution, once that
solution is a variational extremum over a class of wavefunctions which admits scaling, i.e., a
class in which coordinate scaling of any member of the class yields another member of that
class. Linear combinations of functions do not generally form a class which admits scaling.
Exercise 50 Using the virial theorem
Of the exercises we have done so far, which approximate solutions satisfy the virial theorem
Eq. (12.31) exactly, and which do not? What can you deduce in the two diﬀerent cases?
We now derive equivalent formulae for the density functional case. We will ﬁrst generalize
Eq. (12.29), and then Eq. (12.31). As always, instead of considering the groundstate energy
itself, we consider just the construction of the universal functional F[n]. Write
ˆ
F =
ˆ
T +
ˆ
V
ee
.
Then
F[n] = Ψ[n]
ˆ
FΨ[n], (12.32)
and, if we insert any other wavefunction yielding the density n(r), we must get a higher
number. Consider especially the wavefunctions Ψ
γ
[n
1/γ
]. These yield the same density, but
are not the minimizing wavefunctions. Then
d
dγ

γ=1
Ψ
γ
[n
1/γ
]
ˆ
FΨ
γ
[n
1/γ
] = 0. (12.33)
But expectation values of operators scale very simply when the wavefunction is scaled, so
that
d
dγ

γ=1
γ
2
T[n
1/γ
] + γV
ee
[n
1/γ
]
= 0, (12.34)
100 CHAPTER 12. SCALING
or, swapping γ to 1/γ in the derivative,
2T[n] +V
ee
[n] =
d
dγ

γ=1
(T[n
γ
] +V
ee
[n
γ
]) . (12.35)
This then is the generalization of Eq. (12.29) for density functional theory: Because Ψ[n]
minimizes
ˆ
F, variations induced by scaling (but keeping the density ﬁxed) have zero derivative
around the solution, leading to Eq. (12.35).
12.6 Kinetic correlation energy
As usual, the more interesting statement comes from the KS quantities. We apply Eq. (12.35)
to both the interacting and KohnSham systems, and subtract, to ﬁnd:
E
XC
[n] +T
C
[n] =
dE
XC
[n
γ
]
dγ

γ=1
(12.36)
Exercise 51 Virial expressed as scaling derivative
Prove Eq. (12.36).
This statement is trivially true for Hartree and Exchange energies, so we subtract oﬀ the
exchange contribution, and write:
T
C
[n] =
dE
C
[n
γ
]
dγ

γ=1
−E
C
[n] (12.37)
This is an extremely useful and powerful statement. It tells us how, if we’re given a functional
for the correlation energy, we can extract the kinetic contribution simply by scaling.
Exercise 52 LDA kinetic correlation energy:
Assuming the correlation energy density of a uniform gas is known, e
unif
C
(r
s
), give formulas
for both t
unif
C
(r
s
) and u
unif
C
(r
s
).
Exercise 53 Scaling to ﬁnd correlation energies:
1. By applying Eq. (12.37) to n
γ
(r), show
T
C
[n
γ
] = γ
dE
C
[n
γ
]
dγ
−E
C
[n
γ
] (12.38)
2. Next, consider Eq. (12.38) as a ﬁrstorder diﬀerential equation in γ for E
C
[n
γ
]. Show
that
E
C
[n
γ
] = −γ
∞
γ
dγ
γ
2
T
C
[n
γ
], (12.39)
using the fact that E
C
[n
γ
] vanishes as γ →0.
12.7. POTENTIAL 101
3. Show
U
C
[n
γ
] = 2E
C
[n
γ
] −γ
dE
C
[n
γ
]
dγ
(12.40)
and
E
C
[n
γ
] = γ
2
∞
γ
dγ
γ
3
U
C
[n
γ
]. (12.41)
4. Combine Eqs. (12.40) and (12.39) to ﬁnd diﬀerential and integral relations between U
C
and T
C
.
Exercise 54 Scaling of potentials:
If v
C
[n](r) = δE
C
[n]/n(r), and I deﬁne E
C
[n](γ) = E
C
[n
γ
], how is v
C
[n](γ, r) = δE
C
[n](γ)/n(r)
related to v
C
[n](r)?
Exercise 55 Scaling LDA in Wigner approximation:
Within the simple Wigner approximation for the correlation energy of the uniform gas,
Eq. (8.18), deduce both the potential and kinetic contributions, u
unif
C
(r
S
) and t
unif
C
(r
S
),
respectively. Then check all 6 relations of the previous exercise.
12.7 Potential
After our earlier warmup exercises, it is trivial to derive the virial theorem for the exchange
correlation potential. For an arbitrary system, the virial theorem for the groundstate yields:
2T =
N
¸
i=1
Ψ
r
i
· ∇
i
ˆ
V
Ψ, (12.42)
no matter what the external potential. For our problems, the electronelectron repulsion is
homogeneous of order 1, and so
2T +V
ee
=
d
3
r n(r)r · ∇v
ext
(r). (12.43)
This theorem applies equally to the noninteracting and interacting systems, and by subtract
ing the diﬀerence, we ﬁnd
E
XC
[n] +T
C
[n] = −
d
3
r n(r) r · ∇v
XC
(r). (12.44)
Since we can either turn oﬀ the coupling constant or scale the density toward the highdensity
limit, this applies separately to the exchange contributions to both sides, and the remaining
correlation contributions:
E
C
[n] +T
C
[n] = −
d
3
r n(r) r · ∇v
C
(r). (12.45)
For almost any approximate functional, once a KohnSham calculation has been cycled to
selfconsistency, Eq. (12.44) will be satisﬁed. It is thus a good test of the convergence of
such calculations, but is rarely performed in practice.
102 CHAPTER 12. SCALING
Exercise 56 Virial for LDA:
Show that the XC virial theorem is satisﬁed in an LDA calculation.
12.8 Questions
1. What is the relation between scaling and changing Z for the Hatom?
2. Is the virial theorem satisﬁed for the groundstate of the problem V (x) = −exp(−x)?
3. For a homogeneous potential of degree p, e.g., p=1 for the delta function well, we know
E = −T. Can we ﬁnd the groundstate by simply maximizing T?
4. What is the exact kinetic energy density functional for one electron in onedimension?
How does it scale?
5. Does T
loc
S
[n] from chapter ?? satisfy the correct scaling relation? (Recall question 1 from
chapter ??).
6. If Ψ is the groundstate for an interacting electronic system of potential v
ext
(r), of what
is Ψ
γ
a groundstate?
Chapter 13
Adiabatic connection
1
In this chapter, we introduce an apparently new formal device, the coupling constant of
the interaction for DFT. This diﬀers from that of regular manybody theory. But we see
at the end that this is very simply related to scaling.
13.1 One electron
Introduce a parameter in
ˆ
H, say λ. Then all eigenstates and eigenvalues depend on λ, i.e.,
Eq. (1.9) becomes:
ˆ
H
λ
φ
λ
= ε
λ
φ
λ
. (13.1)
Most commonly,
ˆ
H
λ
=
ˆ
T + λ
ˆ
V , 0 < λ < ∞. (13.2)
A λ in front of the potential is often called the coupling constant, and then all eigenvalues
and eigenfunctions depend on λ. For example, for the harmonic oscillator of force constant
k = ω
2
, the groundstate wavefunction is φ(x) = (ω/π)
1/4
exp(−ωx
2
/2). Since V (x) =
kx
2
/2, λV (x) = λkx
2
/2, i.e., the factor of λ multiplies k. Then ω →
√
λω, and so
φ
λ
(x) =
¸
√
λω
π
¸
1/4
exp(−
√
λωx
2
/2) (harmonic oscillator). (13.3)
For a λdependent Hamiltonian, we can diﬀerentiate the energy with respect to λ. Be
cause the eigenstates are variational extrema, the contributions due to diﬀerentiating the
wavefunctions with respect to λ vanish, leaving
dE
λ
dλ
= φ
λ

∂
ˆ
H
λ
∂λ
 φ
λ
. (13.4)
Thus
E = E
λ=1
= E
λ=0
+
1
0
dλ φ
λ

∂
ˆ
H
λ
∂λ
 φ
λ
. (13.5)
1
c 2000 by Kieron Burke. All rights reserved.
103
104 CHAPTER 13. ADIABATIC CONNECTION
Furthermore, if λ simply multiplies
ˆ
V , then
dE
λ
dλ
= φ
λ

ˆ
V  φ
λ
, (13.6)
so that
E = E
λ=0
+
1
0
dλ φ
λ

ˆ
V  φ
λ
. (13.7)
This is often called the HellmannFeynman theorem.
Exercise 57 HellmannFeynman theorem
Show that the 1d Hatom and 1d harmonic oscillator satisfy the HellmanFeynman theorem.
How does the approximate solution using a basis set do?
In modern density functional theory, one of the most important relationships is between
the coupling constant and coordinate scaling. We can see this for these simple 1d problems.
The Schr¨ odinger equation at coupling constant λ is
ˆ
T + λ
ˆ
V (x)
φ
λ
(x) = E
λ
φ
λ
(x) (13.8)
If we replace x by γx everywhere, we ﬁnd
1
γ
2
ˆ
T + λ
ˆ
V (γx)
φ
λ
(γx) = E
λ
φ
λ
(γx) (13.9)
Furthermore, if
ˆ
V is homogeneous of degree p,
ˆ
T + λγ
2
γ
p
ˆ
V (x)
φ
λ
(γx) = γ
2
E
λ
φ
λ
(γx) (13.10)
Thus, if we choose λγ
p+2
= 1, then Eq. (13.10) is just the normal Schr¨ odinger equation,
and we can identify φ
λ
(γx), which is equal (up to normalization) to φ
λ
γ
(x) with φ(x). If we
scale both of these by 1/γ, we ﬁnd
φ
λ
(x) = φ
1/γ
(x) = φ
λ
1/(p+2) (x) (13.11)
In this case, changing coupling constant by λ is equivalent to scaling by λ
1/4
, but this
relation depends on the details of the potential.
Exercise 58 λdependence
Show that, for a homogeneous potential of degree p,
E
λ
= λ
2
p+2
E (13.12)
Exercise 59 λdependence of 1d H
Find the couplingconstant dependence for the wavefunction and energy for the 1d Hatom,
V (x) = −δ(x), and for the 1d harmonic oscillator, V (x) =
1
2
x
2
.
13.2. ADIABATIC CONNECTION FORMULA 105
13.2 Adiabatic connection formula
Now we apply the same thinking to DFT. When doing so, we always try to keep the density
ﬁxed (or altered only in some simple way), and not include the external potential. The
adiabatic connection formula is a method for continuously connecting the KohnSham system
with the physical system. Very simply, we introduce a coupling constant λ into the universal
functional, to multiply V
ee
:
F
λ
[n] = min
Ψ→n
Ψ
ˆ
T + λ
ˆ
V
ee
Ψ (13.13)
For λ = 1, we have the physical system. For λ = 0, we have the KohnSham system. We
can even consider λ →∞, which is a highly correlated system, in which the kinetic energy is
negligible. But for all values of λ, the density remains that of the physical system. Note that
this implies that the external potential is a function of λ: v
λ
ext
(r). We denote Ψ
λ
[n] as the
minimizing wavefunction for a given λ. We can generalize all our previous deﬁnitions, but we
must do so carefully. By a superscript λ, we mean the expecation value of an operator on
the system with couplingconstant λ. Thus
T
λ
[n] = Ψ
λ
[n]
ˆ
TΨ
λ
[n], (13.14)
but
V
λ
ee
[n] = Ψ
λ
[n]λ
ˆ
V
ee
Ψ
λ
[n]. (13.15)
The KohnSham quantities are independent of λ, so
E
λ
XC
[n] = Ψ
λ
[n]
ˆ
T + λ
ˆ
V
ee
Ψ
λ
[n] −Φ[n]
ˆ
T + λ
ˆ
V
ee
Φ[n] (13.16)
Exercise 60 Coupling constant dependence of exchange
Show that
U
λ
[n] = λU[n], E
λ
X
[n] = λE
X
[n], (13.17)
i.e., both Hartree and exchange energies have a linear dependence on the couplingconstant.
Our next step is to write the HellmannFeynman theorem for this λdependence in the
Hamiltonian. We write
F[n] = T
S
[n] +
1
0
dλ Ψ
λ
[n]
ˆ
V
ee
Ψ
λ
[n] (13.18)
where F[n] = F
λ=1
[n], T
S
[n] = F
λ=0
[n]. The derivative w.r.t. λ of F
λ
[n] is just the
derivative of the operator w.r.t. λ, because Ψ
λ
[n] is a minimizing wavefunction at each λ,
just as in the oneelectron case. Inserting the deﬁnition of the correlation energy, we ﬁnd
E
XC
[n] =
1
0
dλ Ψ
λ
[n]
ˆ
V
ee
Ψ
λ
[n] −U[n]
=
1
0
dλ
λ
U
λ
XC
[n] =
1
0
dλU
XC
[n](λ), (13.19)
106 CHAPTER 13. ADIABATIC CONNECTION
where we have used V
λ
ee
= λΨ
λ
[n]
ˆ
V
ee
Ψ
λ
[n] and U
λ
= λU, and introduced the simple
notation U
XC
(λ) = U
λ
XC
/λ for later convenience. This is the celebrated adiabatic connection
formula. We have written the exchangecorrelation energy as a solely potential contribution,
but at the price of having to evaluate it at all intermediate coupling constants. We will see
shortly that this price actually provides us with one of our most important tools for analyzing
density functionals.
A cartoon of the integrand in the adiabatic connection formula for a given physical system
is sketched in Fig. 13.1. The curve itself is given by the solid line, and horizontal lines have
been drawn at the value at λ = 0 and at λ = 1. We can understand this ﬁgure by making
0 0.2 0.4 0.6 0.8 1
U
X
C
(
λ
)
λ
E
X
E
C
−T
C
0
E
X
U
XC
Figure 13.1: Cartoon of the adiabatic connection integrand.
several key observations:
• The value at λ = 0 is just E
X
.
• The value at λ = 1 is just E
X
+U
C
.
• The area between the curve and the xaxis is just E
XC
.
• The area beneath the curve, between the curve and a horizontal line drawn through the
value at λ = 1, is just T
C
.
Thus the adiabatic connection curve gives us a geometrical interpretation of many of the
energies in density functional theory.
To demonstrate the usefulness of adiabatic decomposition, we show this curve for the He
atom in Fig. 13.2. Note ﬁrst the solid, exact line. It is almost straight on this scale. This
is telling us that this system is weakly correlated, as we discuss below. The numbers, from
Table X, are E
X
= −1.025, E
C
= −0.042, U
C
= −0.079, and T
C
= 0.037. Thus E
C
is only
slightly more negative than −T
C
.
This produces yet another way to understand the cancellation of errors. We see that LDA
improves as λ grows from 0 to 1. This is characteristic of a very general trend. But once we
13.3. RELATION TO SCALING 107
1.1
1.05
1
0.95
0.9
0.85
0 0.2 0.4 0.6 0.8 1
He atom
exact
LDA
U
X
C
(
λ
)
λ
Figure 13.2: Adiabatic decomposition of exchangecorrelation energy in He atom, both exactly and in LDA.
claim that the LDA error drops with λ, then the usual cancellation of exchange and correlation
errors follows, since the area under the curve will then be more accurate than the value at
λ = 0. Why this happens will be discussed later. But for now, we simply note that the
statement that LDA improves with λ is equivalent to the cancellation of errors statement,
but is much more transparent.
We can use the adiabatic decomposition (λdependence of U
XC
(λ)) to analyze any approx
imate functional, and ﬁnd out why it behaves as it does. The next section shows how easy
it is to ﬁnd U
XC
(λ) from E
XC
. In fact, we already know.
13.3 Relation to scaling
A third important concept is the relation between scaling and coupling constant. Consider
Ψ
λ
[n], which minimizes
ˆ
T+λ
ˆ
V
ee
and has density n(r). Then Ψ
λ
γ
[n] minimizes
ˆ
T/γ
2
+(λ/γ)
ˆ
V
ee
and has density n
γ
(r). If we choose γ = 1/λ, we ﬁnd Ψ
λ
1/λ
[n] minimizes λ
2
(
ˆ
T +
ˆ
V
ee
) and
has density n
1/λ
(r). But if λ
2
(
ˆ
T +
ˆ
V
ee
) is minimized, then
ˆ
T +
ˆ
V
ee
is minimized, so that we
can identify this wavefunction being simply Ψ[n
1/λ
]. If we scale both wavefunctions by λ, we
ﬁnd the extremely simple but important result
Ψ
λ
[n] = Ψ
λ
[n
1/λ
], (13.20)
which tells us how to construct a wavefunction of coupling constant λ by ﬁrst scaling the
density by 1/λ, ﬁnding the groundstate wavefunction for the scaled density, and then scaling
that wavefunction back to the original size.
This solves a mystery from the previous chapter: The groundstate wavefunction of a
scaled density is the scaled groundstate wavefunction, but only if the coupling constant is
changed, i.e.,
Ψ[n
γ
] = Ψ
1/γ
γ
[n]. (13.21)
108 CHAPTER 13. ADIABATIC CONNECTION
Our argument also yields a simple relation for the energies:
F
λ
[n] = λ
2
F[n
1/λ
], (13.22)
which shows how the λdependence of the universal functional (or, as we shall see, any energy
component) is completely determined by its scaling dependence. Contrast Eq. (13.22) with
Eq. (12.21).
These relations prove to be extremely useful in analyzing functionals and their behavior.
First note that we can take any functional and ﬁnd out its scaling behavior quite easily. Then,
through Eq. (13.22), applied to that functional, we can now deduce its couplingconstant
dependence. For example, the adiabatic connection integrand of Eq. (13.19) is simply
U
XC
(λ)[n] = λU
XC
[n
1/λ
]. (13.23)
The simplest example in this regard is the noninteracting kinetic energy, which we know
scales quadraticly with scaling parameter. Then
T
λ
S
[n] = λ
2
T
S
[n
1/λ
] = T
S
[n] (13.24)
i.e., the kinetic energy is independent of coupling constant, as it should be. Similarly,
E
λ
X
[n] = λ
2
E
X
[n
1/λ
] = λE
X
[n] (13.25)
which again makes sense, since E
X
is constructed from the λindependent KohnSham orbitals,
integrated with λ/r −r
.
Much less trivial is the relation between diﬀerent components of the correlation energy.
Consider Eq. (13.19) in diﬀerential form:
dE
λ
XC
[n]
dλ
= Ψ
λ
[n]
ˆ
V
ee
Ψ
λ
[n] −U[n] (13.26)
Then, from Eq. (13.22),
V
λ
ee
[n] = λΨ
λ
[n]
ˆ
V
ee
Ψ
λ
[n] = λ(U[n] +E
X
[n]) +U
λ
C
[n], (13.27)
where we have used the linear dependence of exchange on λ, and we have written
E
C
[n] = (T[n] −T
S
[n]) + (V
ee
[n] −U[n]) = T
C
[n] +U
C
[n] (13.28)
and T
C
[n] is the kinetic contribution to the correlation energy, while U
C
[n] is the potential
contribution. Inserting Eq. (13.27) into Eq. (13.26) and cancelling the trivial exchange
contributions to both sides, we ﬁnd
dE
λ
C
[n]
dλ
= U
λ
C
[n]/λ (13.29)
Finally, we rewrite this as a scaling relation. Since E
λ
C
[n] = λ
2
E
C
[n
1/λ
] and U
λ
C
[n] =
λ
2
U
C
[n
1/λ
] and writing γ = 1/λ, we ﬁnd, after a little manipulation
γ
dE
C
[n
γ
]
dγ
= E
C
[n
γ
] +T
C
[n
γ
]. (13.30)
13.4. STATIC CORRELATION 109
which is just Eq. (12.38). Thus, the adiabatic connection formula may be thought of as an
integration of this scaling formula between γ = 1 and γ = ∞.
Exercise 61 E
C
from T
C
Derive a relation to extract E
C
[n] from T
C
[n
γ
] alone. Rewrite this to get E
λ
C
[n] from T
λ
C
[n].
Exercise 62 Use Eq. (16.5) and the scaling relations of the previous chapter to write E
C
in
terms of density matrices of diﬀerent λ.
We can easily scale any approximate functional, and so extract the separate contributions
to the correlation energies, and test them against their exact counterparts, check limiting
values, etc. A simple example is LDA. Then when n →n
γ
, r
S
→r
S
/γ. Thus
E
LDA
C
[n
γ
] =
d
3
r n
γ
(r)
unif
C
(r
s,γ
(r)) =
d
3
rn(r)
unif
C
(r
s
(r)/γ) (13.31)
In the highdensity limit, γ →∞,
unif
C
(r
s
/γ) →
unif
C
(0). This is singular in true LDA, and is
an error made by LDA. However, in the low density limit,
unif
C
(r
s
) = −d
0
/r
s
+d
1
/r
3/2
s
+. . .,
so that E
LDA
C
[n
γ
] = −γd
0
d
3
r n(r)/r
s
(r), vanishing linearly with γ, correctly. Note that in
this lowdensity regime, correlation is so strong it scales the same as exchange.
Exercise 63 Adiabatic connection for He atom:
Using the Wigner approximation and the eﬀective exponential density, draw the adiabatic
connection curve for He.
Exercise 64 Adiabatic connection for H atom:
Draw the adiabatic connection curve for the H atom.
Exercise 65 Changing λ:
Derive E
LDA,λ
XC
[n] and T
LDA
C
[n].
13.4 Static correlation
We have seen how, for the total XC energy of many systems of interest, the adiabatic
connection curve is close to linear. In fact, it has always been found that the adiabatic
connection curve is (slightly) concave upwards:
d
2
U
C
(λ)
dλ
2
≥ 0 (unproven) (13.32)
Given the geometric interpretation of Fig X, this means T
C
≤ E
C
. To quantify the fraction
of correlation that is kinetic, we deﬁne
b =
T
C
U
C

(13.33)
110 CHAPTER 13. ADIABATIC CONNECTION
Thus, it has always been found that 0 < b ≤ 1/2, but b is always close to 0.5 for real systems.
An alternative way to write it is
E
C
= (1 −b)U
C
(13.34)
or that if we simplify the adiabatic curve as a step down at some value of λ from E
X
to
E
X
+U
C
, then the step occurs at λ = b.
Considering both the ﬁgures and the numbers, we see that LDA clearly signiﬁcantly un
derestimates b. This is again because of the logarithmic divergence as r
s
→ 0, and we will
see that this does not happen for more sophisticated approximations.
The reason that most systems have b close to 1/2 is that they are not very far from
the highdensity limit. As one moves towards lower densities (much lower than in realistic
systems), b reduces, even coming close to zero. Thus, even for the uniform gas, T
C
eventually
becomes small relative to U
C
, as r
s
→∞. Thus we can speak of b as measuring the amount
of dynamic correlation, i.e., the fraction of correlation energy that is kinetic, with a value
of 1/2 being the maximum. Typical systems are close to this value, and only for very low
densities does b become small.
0.65
0.6
0.55
0.5
0.45
0.4
0 0.2 0.4 0.6 0.8 1
exact (R=5)
U
X
C
(
λ
)
λ
Figure 13.3: Adiabatic decomposition of exchangecorrelation energy in stretched H2.
To appreciate why we are introducing this parameter, we next consider what happens to
a H
2
molecule when we stretch it. In Figure 13.3, we plot the adiabatic connection curve
at R = 5. It looks very diﬀerent from that of He, or of H
2
at equilibrium. To understand
it, we note that at λ = 1, the curve is almost ﬂat, and just about equal to −5/8, i.e., the
exchange(correlation) energy of two separate H atoms. On the other hand, at λ = 0, we
have the exact exchange value of KS theory. For this problem, the electrons are always in a
singlet and the exchange energy is about 0.42. The value of b is about 0.16, and this system
has strong static correlation.
There is no way for LDA to produce a curve that looks anything like this. Because the
average density will be almost the same as a single H atom, the adiabatic curve is very ﬂat,
with essentially no static correlation. Since we are beyond the CoulsonFischer point, we
13.5. QUESTIONS 111
must chose if we wish to break spinsymmetry or not. If we allow symmetry breaking, we will
get the λ = 1 end about right, but have a ﬂat curve, and so overestimate the magnitude of
the correlation energy. On the other hand, if we don’t, we will get a curve that bends just a
little down from the λ = 0 end, and so underestimate the correlation.
As we will see throughout the rest of this book, this is a very important problem for DFT
in quantum chemistry. Many molecules exhibit similar behavior, not in their total energies,
but in their energy diﬀerences between bonded and dissociated states.
Exercise 66 Adiabatic connection for stretched H
2
Draw an accurate adiabatic connection curve for H
2
with a bond length of 20
˚
A.
13.5 Questions
1. Using the highdensity limit of Wigner correlation, estimate roughly how large the corre
lation energy usually is within LDA.
112 CHAPTER 13. ADIABATIC CONNECTION
Chapter 14
Discontinuities
14.1 Koopman’s theorem
Since the large distance condition Eq. (5.18) is true for both the interacting wavefunction
and an independent particle description, the following approximate theorem was ﬁrst noted
by Koopmans (who later won a Nobel prize in economics). In a HartreeFock calculation, if
you ignore relaxation, meaning the change in the selfconsistent potential, when an electron
is removed from the system, you ﬁnd
I
HF
= E
HF
(N −1) −E
HF
(N) ≈ −
N
(14.1)
where
N
is the eigenvalue of the highest occupied orbital.
Exercise 67 Koopman’s theorem for 1d He
For the accurate HF 1d He calculation, test Koopman’s theorem.
But, in exact density functional theory, Koopman’s theorem is exact. To see why, note
that Eq. (5.18) is true independent of the strength of the interaction. Consider the coupling
constant λ, keeping the density ﬁxed. Since the density is ﬁxed, its asympototic decay is the
same for any value of λ, i.e., the ionization potential is independent of λ. In particular, at
λ = 0, its just −
HOMO
, where we now refer to the HOMO of the KS potential. Thus
I = E(N −1) −E(N) = −
HOMO
. (14.2)
We will see below that this condition is violated by all our approximate functionals.
Exercise 68 Asymptotic behavior of the density
From the information in this and earlier chapters, deduce the limiting values of the asymptotes
on the right of Fig. 5.2.
14.2 Potentials
It is straightforward to argue for the exact asymptotic behaviour of the exchangecorrelation
potential for a Coulombic system. Consider ﬁrst exchange. Far from the nucleus, and
113
114 CHAPTER 14. DISCONTINUITIES
electron in an orbital should see an eﬀective nuclear charge of Z − (N − 1), the N − 1
because there are N − 1 electrons close to the nucleus. Since v
ext
(r) = −Z/r, and at
large distances v
H
(r) = N/r, this implies v
X
(r) = −1/r. Since this eﬀect will occur in
any wavefunction producing the exact density, this must also be true for v
XC
(r), so that the
correlation contribution must decay more rapidly. In fact,
v
X
(r) →−1/r, v
C
(r) →−α(N −1)/2r
4
(r →∞) (14.3)
where α(N − 1) is the polarizability of the (N − 1)electron species. The decay of the
correlation potential is so rapid that it has only ever been clearly identiﬁed in an exact
calculation for H
−
.
Exercise 69 Asymptotic potentials
Derive the asymptotic condition on v
X
(r) from the exact decay of the density. How does
v
LDA
X
(r) behave at large distances?
As mentioned earlier, LDA does not provide very realistic looking exchangecorrelation
potentials. Because the density decays exponentially in the tail region, so too will the LDA
potential. In Fig. 14.1, we illustrate this eﬀect. Of course, LDA also misses the shell
structure between the 1s and 2s and 2p electrons, just as smooth approximations for T
S
miss
shell structure. This is partially explained by the above argument: the longrange decay of
8
7
6
5
4
3
2
1
0
0 0.2 0.4 0.6 0.8 1
Ne atom
exact
LDA
v
X
C
(
r
)
r
Figure 14.1: XC potential in the Ne atom, both exact and in LDA
v
X
(r) is due to nonselfinteraction in the exact theory, an eﬀect missed by LDA. This has a
strong eﬀect on the HOMO orbital energy (and all those above it), as will be discussed in
more detail below.
14.3 Derivative discontinuities
We can understand the apparent diﬀerences between LDA and exact potentials in much more
detail. Consider what happens when two distincct subsystems are brought a large distance
14.3. DERIVATIVE DISCONTINUITIES 115
from each other. For example, take the HHe
+
case. At large distances, the interacting
wavefunction for the two electrons should become two (unbalanced) Slater determinants,
making up a HeitlerLondon type wavefunction, yielding a density that is the sum of the
isolated atom densities, centered on each nucleus. In our world of 1d illustrations, this will
be
n(x) = exp(−2x) + 2 exp(−4x −L) (14.4)
where the Hatom is at 0 and the He ion is at L. For this two electron system, we ﬁnd
the molecular orbital as φ(x) =
n(x)/2 and the KohnSham potential by inversion. This
12
8
4
0
4
1 0 1 2 3 4 5
1d HHe+ potential
0
1
2
3
1 0 1 2 3 4 5
1d HHe+ density
v
S
(
x
)
n
(
x
)
x
Figure 14.2: Density and KohnSham potential for 1d HHe
+
.
is shown in Fig. 14.2, for L = 4. The apparent step in the potential between the atoms
occurs where the dominant exponential decay changes. This is needed to ensure KS system
produces two separate densities with separate decay constants. Elementary math then tells
you that the step will be the size of the diﬀerence in ionization potentials between the two
systems.
More generally, for e.g., NaCl being separated, the step must be the diﬀerence in IE.A.,
where E.A. is the electron aﬃnity. This is the energy cost of transferring an inﬁnitesimal of
charge from one system to another, and the KS potential must be constructed so that the
molecule dissociates into neutral atoms. But the same reasoning can be applied to a single
116 CHAPTER 14. DISCONTINUITIES
atom, in contact with a bath of electrons, eg a metal. One ﬁnds that the step in the potential
induced by adding an inﬁnitesimal of charge is IEA.
But LDA is a smooth functional, with a smooth derivative. Thus its potential cannot
discontinuously change when an inﬁnitesimal of charge is added to the atom. So it will be in
error about (IEA)/2 for the neutral, and the reverse for the slightly negatively charged ion.
Comparing the potentials for the Ne atom, we see in Fig. 14.3 how close the LDA potential
is in the region around r = 1, relevant to the 2s and 2p electrons. However, it has been
shifted down by 0.3, the LDA error in the HOMO orbital energy. Since Ne has zero electron
aﬃnity, this is close to (IEA)/2 for this system, since I=0.8. This rationalizes one feature of
the poorlooking potentials within LDA.
2
1.5
1
0.5
0
0 0.5 1 1.5 2
Ne atom
exact
LDA0.3
v
X
C
(
r
)
r
Figure 14.3: XC potential in the Ne atom, both exact and in LDA, in the valence region, but with LDA shifted down by
0.3.
If the functional is generalized to include fractional particle numbers via ensemble DFT,
these steps are due to discontinuities in the slope of the energy with respect to particle number
at integer numbers of electons. Hence the name.
14.4 Questions
1. What kind of chemical systems will suﬀer from strong selfinteraction error?
2. Which is better, to get the right energy and wrong symmetry, or vice versa?
3. Is Koopmans’ theorem satisﬁed by LDA?
4. If one continues beyond x = 5 in Fig. 14.2, what happens?
Chapter 15
Analysis tools
In this chapter, we introduce a variety of tools that help analyze how approximate density
functionals work.
15.1 Enhancement factor
1
1.1
1.2
1.3
1.4
1.5
0 1 2 3 4 5 6
uniform gas
unpolarized
polarized
F
u
n
i
f
X
C
(
r
s
)
rs
Figure 15.1: Enhancement factor for correlation in a uniform electron gas as a function of WignerSeitz radius for unpolarized
(solid line) and fully polarized (dashed line) cases.
15.2 Density analysis
What we can do is ask how accurately we need to know the uniform gas inputs. Recall the
radial density plot of the Ar atom, Fig. 5.1. Now we plot the local WignerSeitz radius,
r
s
(r), in Fig. 15.2, and ﬁnd that the shells are less obvious. The core electrons have r
s
≤ 1,
while the valence electrons have r
s
≥ 1, with a tail stretching toward r
s
→ ∞. To see the
distribution of densities better, we deﬁne the density of r
s
’s:
g
1
(r
s
) =
d
3
r n(r) δ(r
s
−r
s
(r)). (15.1)
117
118 CHAPTER 15. ANALYSIS TOOLS
0
0.5
1
1.5
2
2.5
3
0 0.5 1 1.5 2 2.5 3
Ar atom
r
s
(
r
)
r
Figure 15.2: rs(r) in Ar atom.
This has the simple interpretation: g(r
s
)dr
s
is the number of electrons in the system with
Seitz radius between r
s
and r
s
+dr
s
, and so satisﬁes:
∞
0
dr
s
g(r
s
) = N. (15.2)
This is plotted in Fig. 15.3. The diﬀerent peaks represent the s electrons in each shell. If
0
5
10
15
20
25
30
35
40
45
0 0.5 1 1.5 2 2.5 3
Ar atom
g
1
(
r
s
)
rs
Figure 15.3: g1(rs) in Ar atom.
we integrate this function forward from r
s
= 0, we ﬁnd that it reaches 2 at r
s
= 0.15, 4
at 0.41, 10 at 0.7, and 12 at 0.88; these numbers represent (roughly) the maximum r
s
in a
given shell. The 1s core electrons live with r
s
≤ 1.5, and produce the ﬁrst peak in g
1
; the
2s produce the peak at r
s
= 0.2, and the 2p produce no peak, but stretch the 2s peak upto
0.7. The peak at about 0.82 is the 3s electrons, and the long tail includes the 3p electrons.
Why is this analysis important? We usually write
E
LDA
XC
=
d
3
r n(r)
unif
XC
(r
s
(r)), (15.3)
where
unif
XC
is the exchangecorrelation energy per particle in the uniform gas. But armed with
15.3. ENERGY DENSITY 119
our analysis, we may rewrite this as
E
LDA
XC
=
∞
0
dr
s
g(r
s
)
unif
XC
(r
s
). (15.4)
Thus the contribution to E
LDA
XC
for a given r
s
value is given by g
1
(r
s
) times the weighting
factor
unif
XC
(r
s
). Clearly, from Fig. 15.3, we see that, to get a good value for the exchange
correlation energy of the Ar atom, LDA must do well for r
s
≤ 2, but its performance for
larger r
s
values is irrelevant.
Typical r
s
values are small for core electrons (at the origin, a hydrogenic atom has r
s
=
0.72/Z), but valence electrons have r
s
between about 1 and 6. These produce the dominant
contribution to chemical processes, such as atomization of molecules, but core relaxations
with r
s
1 can also contribute. The valence electrons in simple metal solids have r
s
between
2 and 6.
15.3 Energy density
15.4 Questions
120 CHAPTER 15. ANALYSIS TOOLS
Chapter 16
Exchangecorrelation hole
1
In this chapter, we explore in depth the reliability of LSD. Although not accurate enough for
thermochemistry, LSD has proven remarkably systematic in the errors that it makes. Improved
functionals should incorporate this reliability. We will see that in fact, understanding the
inherent reliability of LDA is the ﬁrst step toward useful generalized gradient approximations.
16.1 Density matrices and holes
We now dig deeper, to understand better why LSD works so reliably, even for highly inho
mogeneous systems. We must ﬁrst return to the manybody wavefunction. We deﬁne the
following (ﬁrstorder) density matrix:
γ(x, x
) = N
dx
2
. . .
dx
N
Ψ
∗
(x, x
2
, . . . , x
N
)Ψ(x
, x
2
, . . . , x
N
) (16.1)
The diagonal elements of this density matrix are just the spindensities:
γ(x, x) = n(x) (16.2)
and the exact kinetic energy can be extracted from the spinsummed (a.k.a reduced) density
matrix:
γ(r, r
) =
¸
σσ
γ(rσ, r
σ
) (16.3)
and
T = −
1
2
d
3
r∇
2
γ(r, r
)r=r
. (16.4)
Note that, although γ(x, x
) is a onebody property, only its diagonal elements are deter
mined by the density. For example, we know that γ
S
, the density matrix in the KohnSham
system, diﬀers from the true density matrix, since the true kinetic energy diﬀers from the
noninteracting kinetic energy:
TC = −
1
2
d
3
r∇
2
(γ(r, r
) −γS(r, r
)) r=r
(16.5)
1
c 2000 by Kieron Burke. All rights reserved.
121
122 CHAPTER 16. EXCHANGECORRELATION HOLE
For a single Slater determinant, we have the following simple result:
γ
S
(x, x
) = δ
σσ
Nσ
¸
i=1
φ
∗
iσ
(r)φ
iσ
(r
) (16.6)
where φ
iσ
(r) is the ith KohnSham orbital of spin σ. This is diagonal in spin.
Exercise 70 For samespin electrons in a large box (from 0 to L, where L → ∞) in one
dimension, show that
γ
S
(x, x
) =
k
F
π
¸
sin(k
F
(x −x
))
k
F
(x −x
)
−
sin(k
F
(x +x
))
k
F
(x +x
)
¸
. (16.7)
Prove this density matrix has the right density, and extract t
S
(x) from it.
We will, however, focus more on the potential energy, since that appears directly in the
usual adiabatic connection formula. To this end, we deﬁne the pair density:
P(x, x
) = N(N −1)
dx
3
. . .
dx
N
Ψ(x, x
, x
3
, . . . , x
N
)
2
(16.8)
which is also known as the (diagonal) secondorder density matrix. This quantity has an
important physical interpretation: P(rσ, r
σ
)d
3
rd
3
r
is the probability of ﬁnding an electron
of spin σ in d
3
r around r, and a second electron of spin σ
in d
3
r
around r
. Thus it contains
information on the correlations among the electrons. To see this, we may write:
P(x, x
)d
3
rd
3
r
=
n(x)d
3
r
¸
n
2
(x, x
)d
3
r
¸
(16.9)
where n
2
(x, x
)d
3
r
is the conditional probability of ﬁnding the second electron of spin σ
in d
3
r
around r
, given that an electron of spin σ has already been found in d
3
r around r.
Now, if the ﬁnding of the second electron were completely independent of the ﬁrst event,
then n
2
(x, x
) = n(x
). This can never be true for any wavefunction, since, by deﬁnition:
dx
P(x, x
) = (N −1) n(x) (16.10)
implying:
dx
n
2
(x, x
) = N −1, (16.11)
i.e., the fact that we have found one electron already, implies that the remaining conditional
probability density integrates up to one less number of electrons.
Exercise 71 Show that
dx
P(x, x
) = (N −1) n(x) (16.12)
and state this result in words. What is the pair density of a oneelectron system? Deduce
what you can about the parallel and antiparallel pair density of a spinunpolarized two electron
system, like the He atom.
16.1. DENSITY MATRICES AND HOLES 123
Why is this an important quantity? Just as the kinetic energy can be extracted from
the ﬁrstorder density matrix, the potential energy of an interacting electronic system is a
twobody operator, and is known once the pair density is known:
V
ee
=
1
2
dx
dx
P(x, x
)/r −r
 (16.13)
Although it can be very interesting to separate out the parallel and antiparallel spin contribu
tions to the pair density, we won’t do that here, since the Coulomb repulsion does not. The
reduced pair density is the spinsummed pair density:
P(r, r
) =
¸
σσ
P(rσ, r
σ
), (16.14)
and is often called simply the pair density. Then
V
ee
=
1
2
d
3
r
d
3
r
P(r, r
)
r −r

. (16.15)
As we have seen, a large chunk of V
ee
is simply the Hartree electrostatic energy U, and
since this is an explicit density functional, does not need to be approximated in a KohnSham
calculation. Thus it is convenient to subtract this oﬀ. We deﬁne the (potential) exchange
correlation hole density around an electron at r of spin σ:
P(x, x
) = n(x) (n(x
) +n
XC
(x, x
)) . (16.16)
The hole is usually (but not always) negative, and integrates to exactly −1:
dx
n
XC
(x; x
) = −1 (16.17)
Since the pair density can never be negative,
n
XC
(x, x
) ≥ −n(x
). (16.18)
This is not a very strong restriction, especially for many electrons. Again, while the spin
decomposition of the hole is interesting, our energies depend only on the spinsum, which is
less trivial for the hole:
n
XC
(r, r
) =
¸
¸
σσ
n(x)n
XC
(x, x
)
¸
/n(r). (16.19)
The pair density is a symmetrical function of r and r
:
P(r
, r) = P(r, r
), (16.20)
but the hole is not. It is often useful to deﬁne u = r
− r as the distance away from the
electron, and consider the hole as function of r and u. These ideas are illustrated in Fig.
(16.1). It is best to think of the hole as a function of u = x
− x, i.e., distance from the
electron point. The hole is often (but not always) deepest at the electron point, and decays
124 CHAPTER 16. EXCHANGECORRELATION HOLE
1
0
1
2
3
4
5
6 4 2 0 2 4 6
cartoon
n
(
x
)
x
Figure 16.1: Cartoon of hole in onedimensional exponential density. Plotted are the actual density n(x
) (long dashes), the
conditional density n2(x = 2, x
) (solid line), and their diﬀerence, the hole density, nXC(x = 2, x
) (dashed line).
rapidly with distance, as the electrons avoid the electron at x. The product of densities in
P(r, r
) gives rise to U, the Hartree energy, while
U
XC
=
d
3
r n(r)
d
3
u
n
XC
(r, u)
2u
. (16.21)
Thus we may say, in a wavefunction interpretation of density functional theory, the exchange
correlation energy is simply the Coulomb interaction between the charge density and its
surrounding exchangecorrelation hole.
The exchange hole is a special case, and arises from the KohnSham wavefunction. For a
single Slater determinant, one can show
P
X
(x, x
) =
Nσ
¸
i=1
N
σ
¸
j=1
φ
∗
iσ
(r)φ
∗
jσ
(r
) (φ
iσ
(r)φ
jσ
(r
) −φ
iσ
(r
)φ
jσ
(r)) (16.22)
The ﬁrst (direct) term clearly yields simply the product of the spindensities, n(x)n(x
), while
the second (exchange) term can be expressed in terms of the density matrix:
P
X
(x, x
) = n(x)n(x
) −γ
S
(x, x
)
2
(16.23)
yielding
n
X
(x, x
) = −γ
S
(x, x
)
2
/n(x). (16.24)
This shows that the exchange hole is diagonal in spin (i.e., only likespins exchange) and is
everywhere negative. Since exchange arises from a wavefunction, it satisﬁes the sumrule, so
that
d
3
u n
X
(r, u) = −1 (16.25)
The exchange hole gives rise to the exchange energy:
E
X
=
1
2
d
3
r
d
3
r
n(r)n
X
(r, r
)
r −r

(16.26)
16.2. HOOKE’S ATOM 125
Exercise 72 Spinscaling the hole
Deduce the spinscaling relation for the exchange hole.
Exercise 73 Holes for oneelectron systems
For the H atom (in 3d), plot P(r, r
), n
2
(r, r
) and n
X
(r, r
) for several values of r as a
function of r
, keeping r
parallel to r.
The correlation hole is everything not in the exchange hole. Since the exchange hole
satisﬁes the sumrule, the correlation hole must integrate to zero:
d
3
u n
C
(r, u) = 0. (16.27)
This means the correlation hole has both positive and negative parts, and occasionally the
sum of exchange and correlation can be positive. It also has a universal cusp at u = 0, due
to the singularity in the electronelectron repulsion there.
An alternative way of representing the same information that is often useful is the pair
correlation function:
g(x, x
) = P(x, x
)/(n(x)n(x
)) (16.28)
This paircorrelation function contains the same information as the hole, but can make some
results easier to state. For example, at small separations, the Coulomb interaction between
the electons dominates, leading to the electronelectron cusp in the wavefunction:
dg
sph.av.
(r, u)/du
u=0
= g
sph.av.
(r, u = 0), (16.29)
where the superscript denotes a spherical average in u. Similarly, at large separations, g →1
as u → ∞ in extended systems, due to the screening eﬀect, but not so for ﬁnite systems.
Furthermore, the approach to unity diﬀers between metals and insulators.
16.2 Hooke’s atom
A useful and interesting alternative external potential to the Coulomb attraction of the nucleus
for electrons is the harmonic potential. Hooke’s atom consists of two electrons in a harmonic
well of force constant k, with a Coulomb repulsion. It is useful because it is exactly solvable,
so that many ideas can be tested, and also (more importantly), many general concepts can
be illustrated.
Exercise 74 Hooke’s atom: Approximate HF
Repeat the approximate HF calculation of chapter ?? on Hooke’s atom, using appropriate
orbitals. Make a deﬁnite statement about the groundstate and HF energies.
Exercise 75 Hooke’s atom: Separation of variables
Show that, for two electrons in external potential v
ext
(r) = kr
2
/2, and with Coulomb inter
action, the groundstate wavefunction may be written as
Ψ(r
1
, r
2
) = Φ(R)φ(u), (16.30)
126 CHAPTER 16. EXCHANGECORRELATION HOLE
where R = (r
1
+ r
2
)/2, u = r
2
− r
1
, Φ(R) is the groundstate orbital of a 3d harmonic
oscillator of mass 2, and φ(u) satisﬁes
−∇
2
u
+
ku
2
4
+
1
u
φ(u) =
u
φ(u) (16.31)
Exercise 76 Hooke’s atom exactly
Show that the function φ(u) = C(1 +u/2) exp(−u
2
/4) satisﬁes the Hooke’s atom equation
(16.31) with k = 1/4, and ﬁnd the exact groundstate energy. How big was your error in
your estimated energy above? Make a rigorous statement about the correlation energy.
Exercise 77 Hooke’s atom: Exchange and correlation holes
Use the exact wavefunction for Hooke’s atom at k = 1/4 to plot both n
X
(r, u) and n
C
(r, u)
for r = 0 and 1, with u parallel to r. The exact density is:
n(r) =
√
2C
2
exp(−r
2
/2)/r ×
7r +r
3
+ 8r exp(−r
2
/2)/
√
2π + 4(1 +r
2
)erf(r/
√
2)
(16.32)
where the error function is
erf(x) =
2
√
π
x
0
dy exp(−y
2
) (16.33)
and C
−1
= 2
5π + 8
√
π
Exercise 78 Hooke’s atom: Electronelectron cusp
Repeat the above exercise for the paircorrelation function, both exchange and exchange
correlation. Where is the electronelectron cusp? What happens as u →∞?
16.3 Transferability of holes
The pair density looks very diﬀerent from one system to the next. But let us consider two
totally diﬀerent systems: the 1d Hatom and the (samespin) 1d uniform electron gas. For
any oneelectron system, the pair density vanishes, so
n
X
(r, r
) = −n(r
) (N = 1). (16.34)
For our 1d H atom, this is just n(x+u), where n(x) = exp(−2x), and is given by the solid
curve in Fig. 16.2, with x = 0. On the other hand, we can deduce the exchange hole for
the uniform gas from the bulk value of the ﬁrstorder density matrix. From Eq.(16.7), taking
x, x
to be large, we see that the density matrix becomes, in the bulk,
γ
unif
S
(u) = nsin(k
F
u)/(k
F
u) (16.35)
leading to
n
X
(u) = −nsin
2
(k
F
u)/(k
F
u)
2
(16.36)
16.3. TRANSFERABILITY OF HOLES 127
1
0.8
0.6
0.4
0.2
0
3 2 1 0 1 2 3
1d Hatom
exact
LDA
n
X
(
0
,
u
)
u
Figure 16.2: Exhange hole at center of onedimensional Hatom, both exactly and in LDA.
These depend only on the separation between points, as the density is constant in the system.
They also are symmetric in u, as there cannot be any preferred direction. This hole is also
plotted in Fig. 16.2, for a density of n = 1, the density at the origin in the 1d Hatom.
Comparison of the uniform gas hole and the Hatom hole is instructive. They are remarkably
similar, even though their pair densities are utterly diﬀerent. If the electronelectron repulsion
is taken as δ(u), U = 1/8 for the 1d Hatom, while U diverges for the uniform gas. Why are
they so similar? Because they are both holes of some quantum mechanical system. So both
holes are normalized, and integrate to 1. They are also equal at the ontop value, u = 0.
Exercise 79 Ontop exchange hole Show that the ontop exchange hole is a localspin
density functional, and give an expression for it.
Why is this important? We may think of the local approximation as an approximation to
the hole, in the following way:
n
LDA
X
(x, x +u) = n
unif
X
(n(x); u), (16.37)
i.e., the local approximation to the hole at any point is the hole of a uniform electron gas,
whose density is the density at the electron point. Fig 16.2 suggests this will be a pretty good
approximation. Then the energy per electron due to the hole is just
X
(x) =
∞
−∞
du v
ee
(u) n
X
(x, x +u), (16.38)
with
E
X
=
∞
−∞
dx n(x)
X
(x). (16.39)
Since, for the uniform gas,
unif
X
(n) =
∞
−∞
du v
ee
(u) n
unif
X
(n; u), (16.40)
we may consider E
LDA
X
[n] as arising from a local approximation to the hole at x. Thus, the
general similarities tell us that LDA should usually be in the ball park. In fact, if v
ee
(u) =
128 CHAPTER 16. EXCHANGECORRELATION HOLE
−δ(u) as we have used in the past, and since the exact ontop hole is just −n(x), then LDA
would be exact for X in that 1d world. But since the real Coulomb interaction in 3d is not a
contact interaction, we will imagine a 1d world in which the repulsion is exp(−2u). Then,
because the overall hole shapes are similar, we ﬁnd (0) = −0.5 exactly, and = −0.562 in
LDA.
1
0.8
0.6
0.4
0.2
0
3 2 1 0 1 2 3
1d Hatom
exact
LDA
n
X
(
0
.
5
,
u
)
u
Figure 16.3: Exhange hole at x = 0.5 of onedimensional Hatom, both exactly and in LDA.
We have shown above how to understand qualitatively why LDA would give results in the
right ballpark. But can we use this simple picture to understand quantitatively the LDA
exchange? If we calculate the exchange energy per electron everywhere in the system, we
get Fig. 16.4. This ﬁgure resembles that of the real He atom given earlier. Looked at more
0.6
0.4
0.2
0
0 1 2
1d H atom
X
LDA
X
(
x
)
x
Figure 16.4: Exhange energy per electron in the onedimensional Hatom, both exactly and in LDA, for repulsion exp(−2u).
closely, our naive hopes are dashed. In fact, our x = 0 case is more an anomaly than typical.
We ﬁnd, for our exponential repulsion, E
X
= −0.351 in LDA, and −0.375 exactly, i.e., the
usual 10% underestimate in magnitude by LDA. But at x = 0, the LDA overestimates the
contribution, while for large x, it underestimates it.
At this point, we do well to notice the diﬀerence in detail between the two holes. The
16.3. TRANSFERABILITY OF HOLES 129
exact Hatom hole contains a cusp at u = 0, missing from the uniform gas hole. This causes
it to deviate quickly from the uniform gas hole near u = 0. At large distances, the Hatom
hole decays exponentially, because the density does, while the uniform gas hole decays slowly,
being a power law times an oscillation. These oscillations are the same Friedel oscillations
we saw in the surface problem. For an integral weighted exp(−2u), the rapid decay of the
exact hole leads to a smaller energy density than LDA.
Now watch what happens when we move the hole point oﬀ the origin. In Fig. 16.3, we
plot the two holes at x = 0.5. The Hatom hole is said to be static, since it does not change
position with x. By plotting it as a function of u, there is a simple shift of origin. The hole
remains centered on the nucleus, which is now at u = −0.5. On the other hand, we will
see that for most many electron systems, the hole is typically quite dynamic, and follows the
position where the ﬁrst electron was found. This is entirely true in the uniform gas, whose
hole is always symetrically placed around the electron point, as seen in this ﬁgure. Note that,
although it is still true that both holes are normalized to −1, and that the ontop values still
agree, the strong diﬀerence in shape leads to more diﬀerent values of (x) (0.312 in LDA,
and 0.368 exactly). In Fig. 16.4, we plot the resulting exchange energy/electron throughout
the atom. The LDA curve only loosely resembles the exact curve. Near the nucleus it has a
cusp, and overestimates
X
(x). As x gets even larger, n(x) decays exponentially, so that the
LDA hole becomes very diﬀuse, since its length scale is determined by 1/k
F
, where k
F
= πn,
while the exact hole never changes, but simply moves further away from the electron position.
large x. When weighted by the electron density, to deduce the contribution to the exchange
energy, there is a large cancellation of errors between the region near the nucleus and far
away. (To see the similarities with real systems, compare this ﬁgure with that of Fig. 17.7).
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.5 1 1.5 2 2.5 3
1d Hatom
exact
LDA
n
s
y
m
X
(
0
.
5
,
u
)
u
Figure 16.5: Symmetrized exhange hole at x = 0.5 of onedimensional Hatom, both exactly and in LDA.
But, while all the details of the hole are clearly not wellapproximated in LDA, especially as
we move around in a ﬁnite system, we now show that the important averages over the hole
130 CHAPTER 16. EXCHANGECORRELATION HOLE
are. To begin with, both positive and negative values contribute equally, so we may write:
E
X
=
∞
−∞
dx n(x)
∞
0
du n
sym
X
(x, u), (16.41)
where
n
sym
X
(x, u) =
1
2
(n
X
(x, u) +n
X
(x, −u)) (16.42)
This symmetrizing has no eﬀect on the LDA hole, since it is already symmetric, but Fig.
?? shows that it improves the agreement with the exact case: both holes are now parabolic
around u = 0, and the maximum deviation is much smaller. However, the cusp at the nucleus
is clearly missing from the LDA hole, and still shows up in the exact hole.
0.5
0.4
0.3
0.2
0.1
0
0 0.5 1 1.5 2 2.5 3
1d Hatom
exact
LDA
n
X
(

u

)
u
Figure 16.6: Systemaveraged symmetrized exhange hole of onedimensional Hatom, both exactly and in LDA.
Our last and most important step is to point out that, in fact, it is the systemaveraged
symmetrized hole that appears in E
X
, i.e.,
n
X
(u) =
∞
−∞
dx n(x)n
sym
X
(x; u) (16.43)
where now u always taken to be positive, since
E
X
=
∞
0
du v
ee
(u)n
X
(u) (16.44)
The systemaveraged symmetrized hole is extremely wellapproximated by LDA. Note how no
cusps remain in the exact hole, because of the systemaveraging. Note how both rise smoothly
from the same ontop value. Note how LDA underestimates the magnitude at moderate u,
which leads to the characteristic underestimate of LDA exchange energies. Finally, multiplying
by v
ee
= exp(−2u), we ﬁnd Fig. 16.7. Now the 10% underestimate is clearly seen, with no
cancellation of errors throughout the curve.
We can understand this as follows. For small u, a local approximation can be very accurate,
as the density cannot be very diﬀerent at x + u from its value at x. But as u increases, the
density at x + u could diﬀer greatly from that at x, especially in a highly inhomogeneous
system. So the shortranged hole is wellapproximated, but the longrange is not. In fact, for
16.4. OLD FAITHFUL 131
0.5
0.25
0
0 0.5 1 1.5
1d Hatom
exact
LDA
n
X
(

u

)
e
x
p
(
−
2

u

)
u
Figure 16.7: Systemaveraged symmetrized exhange hole of onedimensional Hatom, both exactly and in LDA, weighted by
exp(−2u).
large u, the oscillating powerlaw LDA behavior is completely diﬀerent from the exponentially
decaying exact behavior:
Exercise 80 Show that the exact systemaveraged symmetrized hole in the 1d H atom is:
n
X
(u) = −exp(−2u) (1 + 2u), (= f(πu) LSD), (16.45)
where
f(z) = −
2
z
2
z
0
dz
sin
2
(z
)/z
. (16.46)
But the constraint of the sumrule and the wonders of system and spherical (in the 3d case)
averaging lead to a very controlled extrapolation at large u. This is the true explanation of
LDA’s success. Notice also that this explanation requires that uniform gas values be used:
Nothing else implicitly contains the information about the hole.
16.4 Old faithful
We are now in a position to understand why LSD is such a reliable approximation. While it
may not be accurate enough for most quantum chemical purposes, the errors it makes are
very systematic, and rarely very large.
We begin from the ansatz that LSD is a model for the exchangecorrelation hole, not just
the energy density. We denote this hole as n
unif
XC
(r
s
, ζ; u), being the hole as a function of
separation u of a uniform gas of density (4πr
3
s
/3)
−1
and relative spinpolarization ζ. Then,
by applying the technology of the previous section to the uniform gas, we know the potential
exchangecorrelation energy density is given in terms of this hole:
u
unif
XC
(r
s
, ζ) = 2π
∞
0
duu n
unif
XC
(r
s
, ζ; u). (16.47)
This then means that, for an inhomogeneous system,
U
LSD
XC
[n] =
d
3
r u
unif
XC
(r
s
(r), ζ(r)) = N
∞
0
duu n
LSD
XC
(u), (16.48)
132 CHAPTER 16. EXCHANGECORRELATION HOLE
where
n
LSD
XC
(u) =
1
N
d
3
r n(r)
dΩ
u
4π
n
unif
XC
(r
s
(r), ζ(r); u) (16.49)
is the system and sphericallyaveraged hole within LSD.
Figure 16.8: Universal curve for the systemaveraged ontop hole density in spinunpolarized systems. The solid curve is for
the uniform gas. The circles indicate values calculated within LSD, while the crosses indicate essentially exact results, and
the plus signs indicate less accurate CI results.
Why should this hole look similar to the true hole? Firstly, note that, because the uniform
gas is an interacting manyelectron system, its hole satisﬁes the same conditions all holes
satisfy. Thus the uniform gas exchange hole integrates to 1, and its correlation hole integrates
to zero. Furthermore, its exchange hole can never be negative. Also, the ontop hole (u = 0)
is very well approximated within LSD. For example, in exchange,
n
X
(r, r) = −
n
2
↑
(r) +n
2
↓
(r)
/n(r) (16.50)
i.e., the ontop exchange hole is exact in LSD. Furthermore, for any fully spinpolarized or
highlycorrelated system,
n
XC
(r, r) = −n(r) (16.51)
i.e., the ontop hole becomes as deep as possible (this makes P(r, r) = 0), again making
it exact in LSD. It has been shown to be highly accurate for exchangecorrelation for most
systems, although not exact in general. Then, with an accurate ontop value, the cusp
condition, which is also satisﬁed by the uniform gas, implies that the ﬁrst derivative w.r.t. u
at u = 0 is also highly accurate. In Fig. 16.8, we plot the systemaveraged ontop hole for
several systems where it is accurately known. For these purposes, it is useful to deﬁne the
16.4. OLD FAITHFUL 133
systemaveraged density:
n(u) =
1
N
dΩ
u
4π
d
3
r n(r) n(r +u). (16.52)
and a systemaveraged mean r
s
value:
r
s
=
d
3
r n
2
(r) r
s
(r)
d
3
r n
2
(r)
. (16.53)
Clearly, the systemaveraged ontop hole is very accurate in LSD.
That LSD is most accurate near u = 0 can be easily understood physically. At u grows r
gets further and further away from r. But in LSD, our only inputs are the spindensities at r,
and so our ability to make an accurate estimate using only this information suﬀers. Indeed, as
mentioned above, at large separations the paircorrelation function has qualitatively diﬀerent
behavior in diﬀerent systems, so that LSD is completely incorrect for this quantity. Thus, we
may expect LSD to be most accurate for small u, as indeed it is. But even at large u, its
behavior is constained by the hole normalization sumrule.
0.5
0.4
0.3
0.2
0.1
0
0 1 2 3
He atom
HF
LDA
2
π
u
n
X
(
u
)
u
Figure 16.9: Systemaveraged exchange hole density (in atomic units) in the He atom, in LDA (dashed) and exactly (HF 
solid). The area under each curve is the exchange energy.
Two other points are salient. The exchangecorrelation hole in the uniform gas must be
spherical, by symmetry, whereas the true hole is often highly aspherical. But this is irrelevant,
since it is only the sphericalaverage that occurs in E
XC
. Furthermore, the accuracy of LSD
can fail in regions of extreme gradient, such as near a nucleus or in the tail of a density. But
in the former, the phasespace in the systemaverage is small, while in the latter, the density
itself is exponentially small. The same argument applies to large separations: even if g is
badly approximated by LSD, it is −n(r + u)g(r, r + u) that appears in the hole, and this
vanishes rapidly, both exactly and in LSD. Thus limitations of LSD in extreme situations do
not contribute strongly to the exchangecorrelation energy.
This then is the explanation of LSD’s reliability. The energydensity of the uniform gas
contains, embedded in it, both the sumrule and the accurate ontop hole information. For
134 CHAPTER 16. EXCHANGECORRELATION HOLE
0.08
0.04
0
0 1 2 3
He atom
exact
LDA
2
π
u
n
λ
=
1
C
(
u
)
u
Figure 16.10: Systemaveraged correlation hole density at full coupling strength (in atomic units) in the He atom, in LSD,
numerical GGA, and exactly (CI). The area under each curve is the full couplingstrength correlation energy.
most systems, this implies that their exact systemaveraged hole looks very like its LSD
approximation, with only small diﬀerences in details, but never with very large diﬀerences.
In Figs. 16.916.10, we plot the systemaveraged exchange and potentialcorrelation holes
for the He atom, both exactly and in LSD. Each is multiplied by a factor of 2πu, so that
its area is just the exchange or correlation energy contribution (per electron). We see that
LSD is exact for small u exchange, and typically underestimates the hole in the allimportant
moderate u region, while ﬁnally dying oﬀ too slowly at large u. For correlation, we see that
the LSD hole is always negative in the energetically signiﬁcant regions, whereas the true hole
has a signiﬁcant positive bump near u = 1.5. This is yet another explanation of the large
overestimate of LSD correlation. In the next chapter, we will show how to use the failed
gradient expansion to improve the description of the hole, and so construct a generalized
gradient approximation.
Part IV
Beyond LDA
135
Chapter 17
Gradients
In this chapter, we explore the ”obvious” correction to the local approximation, namely
the inclusion of gradient information. We’ll see that its not as easy as it looks, and why
it took about a quarter of a century to accomplish.
17.1 Perimeter problem
To illustrate the interesting complexities of the use of gradients, we return to our problem
from Chapter 2, in which we needed to know the perimeter of a shape, given its deﬁnition
in terms of polar coordinates, r(θ). In all of what follows, we assume we do not know the
exact answer.
We already saw in chapter 8 that the local approximation is
P
loc
=
2π
0
dθ r(θ). (17.1)
Let us ﬁrst consider the family of smooth curves introduced in that chapter:
r(θ) = 1 + cos(nθ) (17.2)
Here 0 ≤ < 1, while n is an integer. For these curves, the local approximation yields exactly
2π for the perimeter.
In table 17.1, I have compared results for n = 1, as a function of . The local approximation
is doing rather well, with a maximum error of −15%. Notice how its always an underestimate.
n local % error exact
1 0.0 6.2832 0.00 6.2832
1 0.1 6.2832 0.25 6.2989
1 0.2 6.2832 0.99 6.3462
1 0.4 6.2832 3.88 6.5371
1 0.8 6.2832 14.37 7.3376
Table 17.1: Perimeter’s of shapes for n = 1, and various , both exactly and in local approximation.
137
138 CHAPTER 17. GRADIENTS
n local % error exact
1 0.2 6.2832 1 6.3462
2 0.2 6.2832 4 6.5297
4 0.2 6.2832 13 7.1989
8 0.2 6.2832 32 9.2984
16 0.2 6.2832 57 14.7104
32 0.2 6.2832 77 26.7835
64 0.2 6.2832 88 51.9081
Table 17.2: Perimeter’s of shapes for = 0.2, and various n, both exactly and in local approximation.
One could imagine a clever person, not knowing the true formula, being able to prove an
exact condition like:
P
loc
≤ P (17.3)
We also notice that the error increases with . Obviously = 0 is a circle, where the local
approximation is exact. For small , the curve is in some sense close to a circle, and the
local approximation should be accurate. In Fig. 17.1, we plot the = 0.8 case, and see that,
n=1, eps=0.8
Figure 17.1: Shape r = 1 + 0.8 cos(θ).
although the radius varies from 0.2 to 1.8, the curve looks quite circular.
On the other hand, in Table 17.1, we ﬁx = 0.2, and increase n. The error grows almost
quadratically with n. Again, n = 0 is the circular case, and so small n is somehow more
circular than large n.
To understand the ever increasing error, we plot the n = 64 curve in Fig. 17.2. The local
approximation yields the perimeter of a circle with the average radius, r = 1. Obviously, the
64 wiggles between r = 1.2 and r = 0.8 as one goes round the circle are not accounted for
in the local approximation, but add greatly to the perimeter.
Thus we can speak of two distinct ways in which a curve might be ’close’ to a circle. The
ﬁrst, more familiar way, is when is small, and so the perturbation on r = 1 is weak. This can
17.1. PERIMETER PROBLEM 139
n=64, eps=0.2
Figure 17.2: Shape r = 1 + 0.2 cos(64 θ).
usually be treated by response theory. Perturbation theory will always be accurate once the
pertubation is suﬃciently week, no matter how rapidly the curve is varying (i.e., no matter
how large n is). The second, less familiar way is when the perturbation is slowlyvarying, but
can be arbitrarily large. This is the case when n is small, but need not be. In Fig. 17.1,
the local approximation works quite well, even though is enormous. The ratio of maximum
to minimum r is 9!
Let us ﬁrst analyze the weak perturbation case. We can cheat by using our knowledge of the
exact functional, but I’m sure there are ways to derive the result without it. If r = r
0
+ f(θ),
then
P = 2πr
0
+
2
2r
0
2π
0
dθ
df
dθ
2
+O(
4
) (17.4)
For our standard curve, the integral is n
2
π. There is no linear term, because it vanishes by
periodicity requirements. You will ﬁnd this formula quite accurate, even at = 0.2, and
hence the (near) quadratic growth in the error in the local approximation in Table 17.1.
Note that, in the local approximation, there is no
2
term. Thus the local approximation,
while being exact for the circle, is hopeless for weak perturbations around the circle.
The other good approximation is called the gradient expansion. We ﬁrst note that the
only way to make a dimensionless gradient is by dividing the derivative by the radius. We
deﬁne s = dr/dθ/r. Then, for a slowlyvarying shape, we can write
P =
dθr (1 +C s
2
+. . .) (17.5)
where C is yet to be determined. There is no term linear in s, because its integral would
vanish. Thus the GEA, or gradient expansion approximation, consists of keeping just the ﬁrst
two terms, once C has been found.
Finding C is easy, once we know the linear response. We simply imagine a perturbation
that is both weak and slow. Then s = dr/dθ/r
0
, and we see that C must be 1/2. Thus the
140 CHAPTER 17. GRADIENTS
n local % error GEA % error exact
1 0.0 6.2832 0.00 6.2832 0.00 6.2832
1 0.1 6.2832 0.25 6.2989 0.00 6.2989
1 0.2 6.2832 1 6.3467 0.01 6.3462
1 0.4 6.2832 4 6.5455 0.1 6.5371
1 0.8 6.2832 14 7.5398 3 7.3376
2 0.2 6.2832 48 6.5371 0.1 6.5297
4 0.2 6.2832 13 7.2988 1 7.1989
8 0.2 6.2832 32 10.3455 11 9.2984
16 0.2 6.2832 57 22.5323 53 14.7104
32 0.2 6.2832 77 71.2795 166 26.7835
64 0.2 6.2832 88 266.2640 400 51.9081
Table 17.3: Perimeter’s of shapes for = 0.2, and various n, both exactly and in local approximation and in GEA.
GEA in this case is
P
GEA
= P
loc
+
1
2
2π
0
dθ
1
r
dr
dθ
2
(17.6)
In Table 17.1, we have added the results of the gradient expansion. We see that for
n = 1, where the local approximation was quite good, GEA reduces the error by at least
a factor of 4, and sometimes much more. It also overestimates the perimeter in all cases.
Even up to n = 8, for = 0.2, the error is still reduced by a factor of 3. But for larger n,
meaning larger gradients, the GEA overcorrects, eventually producing larger errors than the
local approximation! Our conclusion is that, for slowlyvarying shapes, the gradient expansion
greatly improves on the local approximation, but does not work for rapid variations, and can
even worsen the results.
To make the point very clear, suppose we live in a world where most of the shapes we care
about have n between 2 and 8, while varies between .4 and .8. The results of the local and
gradient expansion approximations are listed in Table ??. Only for the most slowlyvarying
cases does the GEA really improve over the local approximation. In all cases, it overcorrects,
usually by more than the original error. For these systems, the GEA is hardly an improvement.
17.2 Gradient expansion
Way back when, in the original KohnSham paper, it was feared that LSD might not be too
good an approximation (it turned out to be one of the most successful ever), and a simple
suggestion was made to improve upon its accuracy. The idea was that, for any suﬃciently
slowly varying density, an expansion of a functional in gradients should be of ever increasing
accuracy:
A
GEA
[n] =
d
3
r
a(n(r)) +b(n(r))∇n
2
+. . .
(17.7)
17.2. GRADIENT EXPANSION 141
n local % error GEA % error exact
2 0.4 6.28 13 7.33 2 7.22
2 0.6 6.28 24 8.80 7 8.26
2 0.8 6.28 34 11.31 18 9.55
4 0.4 6.28 33 10.48 12 9.34
4 0.6 6.28 48 16.34 36 12.04
4 0.8 6.28 58 26.39 76 15.00
6 0.4 6.28 47 15.73 32 11.94
6 0.6 6.28 61 28.90 77 16.32
6 0.8 6.28 70 51.52 146 20.93
8 0.4 6.28 57 23.07 56 14.76
8 0.6 6.28 70 46.50 124 20.80
8 0.8 6.28 77 86.71 221 27.04
Table 17.4: Perimeter’s of shapes for various and n, both exactly and in local approximation and in GEA.
Then, if LSD was moderately accurate for inhomogeneous systems, GEA, the gradient ex
pansion approximation, should be more accurate. The form of these gradient corrections can
be determined by scaling, while coeﬃcients can be determined by several techniques.
For both T
S
and E
X
, deﬁned on the KohnSham wavefunction, the appropriate measure of
the density gradient is given by
s(r) = ∇n(r)/(2k
F
(r)n(r)) (17.8)
This measures the gradient of the density on the length scale of the density itself, and has
the important property that s[n
γ
](r) = s[n](γr), i.e., it is scale invariant. It often appears in
the chemistry literature as x = ∇n/(n
4/3
), so that x ≈ 6s. The gradient expansion of any
functional with powerlaw scaling may be written as:
A
GEA
[n] =
d
3
r a(n(r))
1 +Cs
2
(r)
(17.9)
The coeﬃcient may be determined by a semiclassical expansion of the KohnSham density
matrix, whose terms are equivalent to a gradient expansion. One ﬁnds
C
S
= 5/27, C
X
= 10/81 (17.10)
(In fact, a naive zerotemperature expansion gives C
X
= 7/81, but a more sophisticated
calculation gives the accepted answer above). The gradient correction generally improves
both the noninteracting kinetic energy and the exchange energy of atoms. The improvement
in the exchange energy is to reduce the error by about a factor of 2.
Exercise 81 For the Hatom, calculate the gradient corrections to the kinetic and exchange
energies. (Don’t forget to spinscale).
142 CHAPTER 17. GRADIENTS
For correlation, the procedure remains as simple in principle, but more diﬃcult in practice.
Now, there are nontrivial density and spin dependences, and the terms are much harder to
calculate in perturbation theory. In the highdensity limit, the result may be written as
∆E
GEA
C
=
2
3π
2
d
3
r n(r)φ(ζ(r))t
2
(r). (17.11)
Here t = ∇n/(2k
s
n), where k
s
=
4k
F
/π is the ThomasFermi screening length, the
natural wavevector scale for correlations in a uniform system. The spinpolarization factor is
φ(ζ) = ((1 +ζ)
2/3
+ (1 −ζ)
2/3
)/2. This correction behaves poorly for atoms, being positive
and sometimes larger in magnitude than the LSD correlation energy, leading to positive
correlation energies!
Exercise 82 GEA correlation energy in H atom
Calculate the GEA correction to the H atom energy, and show how ∆E
GEA
C
scales in the
highdensity limit.
17.3 Gradient analysis
Having seen in intimate detail which r
s
values are important to which electrons in Fig. 15.2,
we next consider reduced gradients. Fig. 17.3 is a picture of s(r) for the Ar atom, showing
0
0.5
1
1.5
2
0 0.5 1 1.5 2 2.5 3
Ar atom
s
(
r
)
r
Figure 17.3: s(r) in Ar atom.
how s changes from shell to shell. Note ﬁrst that, at the nucleus, s has a ﬁnite moderate
value. Even though the density is large in this region, the reduced gradient is reasonable.
Exercise 83 Reduced gradient at the origin
Show that at the origin of a hydrogen atom, s(0) = 0.38, and argue that this value wont
change much for any atom.
For any given shell, e.g. the core electrons, s grows exponentially. When another shell begins
to appear, there exists a turnover region, in which the gradient drops rapidly, before being
dominated once again by the decay of the new shell. We call these intershell regions.
17.3. GRADIENT ANALYSIS 143
We deﬁne a density of reduced gradients, analogous to our density of r
s
’s:
g
3
(s) =
d
3
r n(r) δ(s −s(r)), (17.12)
also normalized to N, and plotted for the Ar atom in Fig. 17.4. Note there is no contribution
0
5
10
15
20
25
30
35
40
0 0.5 1 1.5 2
Ar atom
g
3
(
s
)
s
Figure 17.4: g3(s) in Ar atom.
to atoms (or molecules) from regions of s = 0. Note also that almost all the density has
reduced gradients less than about 1.2, with a long tail stretching out of larger s.
Exercise 84 Gradient analysis for H atom
Calculate r
s
(r) and s(r) for the H atom, and make the corresponding g
1
(r
s
) and g
3
(s) plots.
Why does the exponential blow up at large r not make the LDA or GEA energies diverge?
We now divide space around atoms into spheres. We will (somewhat arbitrarily) denote
the valence region as all r values beyond the last minimum in s(r), containing 6.91 electrons
in Ar. We will denote the region of all r smaller than the position of the last peak in s as
being the core region (9.34 electrons). The last region in which s decreases with r is denoted
the corevalence region, in which the transition occurs (1.75 electrons). We further denote
the tail region as being that part of the valence region where s is greater than its maximum
in the core (1.54 electrons).
To see how big s can be in a typical system that undergoes chemical reactions, consider
Fig. 17.5. On the left is plotted values of r
s
(top) and of s (bottom) as a function of r in an N
atom. Then on the right are plotted the distribution of values (turned sideways). We see that
the core electrons have r
s
≤ 0.5, while the valence electrons have 0.5 ≤ r
s
≤ 2. The svalues
are more complex, since they are similar in both the core and valence regions. We see that
almost all the density has 0.2 ≤ s ≤ 1.5. Because s is not monotonic, many regions in space
can contribute to g
3
(s) for a given s value. To better distinguish them, one can perform a
pseudopotential calculation, that replaces the core electrons by an eﬀective potential designed
only to reproduce the correct valence density. This is marked in the two righthand panels
by a longdashed line. In the g
3
(r
s
) ﬁgure, we see indeed that the pseudopotential is missing
144 CHAPTER 17. GRADIENTS
Figure 17.5: Distribution of densities and gradients in the N atom.
precisely the inner core peak, while in the g(s) ﬁgure, we see the contribution of the valence
electrons alone to the peak from the combined corevalence region.
We may use the s curve to make more precise (but arbitrary) deﬁnitions of the various
shells. Up to the ﬁrst maximum in s we call the core. The small region where s drops rapidly
is called the corevalence intershell region. The region in which s grows again, but remains
below its previous maximum, we call the valence region. The remainer is denoted the tail.
We make these regions by vertical dashed lines, which have been extended to the upper r
s
curve. Because that curve is monotonic, we can draw horizontal lines to meet the vertical
lines along the curve, and thus divide the g(r
s
) curve into corresponding regions of space.
But we are more interested in atomization energies than atomic energies. In the case of a
molecule, there is no simple monotonic function, but one can easily extract the densities of
variables. Having done this for the molecule, we can make a plot of the diﬀerences between
(twice) the atomic curves and the molecular curve. Now the curves are both positive and
negative. We see immediately that the core does not change on atomization of the molecule,
and so is irrelevant to the atomization energy. Interestingly, it is also clear that it is not
17.3. GRADIENT ANALYSIS 145
Figure 17.6: Diﬀerences in distribution of densities and gradients upon atomization of the N2 molecule.
only the valence electrons (r
s
> 0.8) that contribute, but also the corevalence region, which
include r
s
values down to 0.5. Finally, note that the r
s
values contributing to this energy also
extend to higher r
s
regions.
To summarize what we have learned from the gradient analysis:
• Note again that nowhere in either the N or Ar atoms (or any other atom) has s << 1,
the naive requirement for the validity of LSD. Real systems are not slowlyvarying. Even
for molecules and solids, only a very small contribution to the energy comes from regions
with s < 0.1.
• A generalized gradient approximation, that depends also on values of s, need only do
well for s ≤ 1.5 to reproduce the energy of the N atom. To get most chemical reactions
right, it need do well only for s ≤ 3.
We end by noting that while this analysis can tell us which values of r
s
and s are relevant
to real systems, it does not tell us how to improve on LSD. To demonstrate this, in Fig.
17.7, we plot the exchange energy per electron in the He atom. One clearly sees an apparent
146 CHAPTER 17. GRADIENTS
1.2
0.8
0.4
0
0 1 2 3
He atom
X
LDA
X
(
r
)
r
Figure 17.7: Exchange energy per electron in the He atom, both exactly and in LDA. The nucleus is at r = 0.
cancellation of errors throughout the system, i.e., LDA is not describing this quantity well at
each point in the system, but it does a good job for its integral. As we will show below, a
pointwise analysis of energy densities turns out to be, well, somewhat pointless...
17.4 Questions
1. If GEA overcorrects LDA, what can you say about how close the system is to uniform?
2. Do you expect LDA to yield the correct linear response of a uniform electron gas?
3. Sketch the gradient and density analysis for the ionization energy for Li.
Chapter 18
Generalized gradient approximation
1
In this chapter, we ﬁrst discuss in more detail why the gradient expansion fails for ﬁnite
systems, in terms of the exchangecorrelation hole. Fixing this leads directly to a generalized
gradient approximation (GGA) that is numerically deﬁned, and corrects many of the limita
tions of GEA for ﬁnite systems. We then discuss how to represent some of the many choices
of forms for GGA, using the enhancement factor. We can understand the enhancement factor
in terms of the holes, and then understand consequences of gradient corrections for chemical
and physical systems. Finally we mention other popular GGAs not constructed in this fashion.
18.1 Fixing holes
We begin with exchange. We have seen that LSD works by producing a good approximation
to the system and sphericallyaveraged hole near u = 0, and then being a controlled extrap
olation into the large u region. The control comes from the fact that we are modelling the
hole by that of another physical system, the uniform gas. Thus the LSD exchange hole is
normalized, and everywhere negative. When we add gradient corrections, the model for the
hole is no longer that of any physical system. In particular, its behavior at large separations
goes bad. In fact, the gradient correction to the exchange hole can only be normalized with
the help of a convergence factor.
To see this, consider Fig. 18.1, which shows the spherically averaged exchange hole for a
given reduced gradient of s = 1. Everything is plotted in terms of dimensionless separation,
z = 2k
F
u. The region shown is the distance to the ﬁrst maximum in the LDA hole, where it
just touches zero for the ﬁrst time. The GEA correction correctly deepens the hole at small
u, but then contains large oscillations at large u. These oscillations cause the GEA hole to
unphysically become positive after z = 6. These oscillations are so strong that they require
some damping to make them converge. Without a damping factor, the gradient correction
to the exchange energy from this hole is undeﬁned.
The realspace cutoﬀ construction of the generalized gradient approximation (GGA), is
to include only those contributions from the GEA exchange hole that are negative, and to
1
c 2000 by Kieron Burke. All rights reserved.
147
148 CHAPTER 18. GENERALIZED GRADIENT APPROXIMATION
Figure 18.1: Sphericallyaveraged exchange hole density n
sph.av.
X (u) for s = 1.
truncate the resulting hole at the ﬁrst value of u which satisﬁes the sumrule. In the ﬁgure, for
small z, the GGA hole becomes more negative than GEA, because in some directions, the GEA
hole has become positive, and these regions are simply sliced out of GGA, leading to a more
negative sphericallyaveraged hole. Then, at about 2k
F
u = 6.4, the hole is truncated, because
here the normalization integral equals 1, and its energy density contribution calculated.
We note the following important points:
• Simply throwing away positive contributions to the GEA exchange hole looks very ugly
in real space. But remember we are only trying to construct a model for the spherical
and systemaverage. Thus although that has happened in some directions for small z,
the sphericallyaveraged result remains smooth.
• By throwing out postive GEA contributions, a signiﬁcantly larger energy density is found
for moderate values of s. Even for very small s, one ﬁnds corrections. Even when GEA
corrections to the LSD hole are small, there are always points at which the LSD hole
vanishes. Near these points, GEA can make a postive correction, and so need ﬁxing.
This means that even as s →0, the GGA energy diﬀers from GEA.
• For very large s (e.g. greater than 3), the wild GEA hole produces huge corrections
to LSD, the normalization cutoﬀ becomes small, and limits the growth of the energy
density. However, this construction clearly cannot be trusted in this limit, since even a
small distance away from u = 0 may have a completely diﬀerent density.
The results of the procedure are shown in Fig. 18.2. We see that the LDA underestimate
is largely cured by the procedure, resulting in a systemaveraged hole that matches the exact
one better almost everywhere.
18.1. FIXING HOLES 149
0.5
0.4
0.3
0.2
0.1
0
0 1 2 3
He atom
HF
LDA
GGA
2
π
u
n
X
(
u
)
u
Figure 18.2: Systemaveraged exchange hole density (in atomic units) in the He atom. The exact curve (solid) is Hartree
Fock, the long dashes denote LDA, and the short dashes are realspace cutoﬀ GGA. The area under each curve is the exchange
energy.
There is a simple way to picture the results. We have seen how the spinscaling re
lation means that one need only devise a total density functional, and its corresponding
spindependence follows. Furthermore, since it scales linearly, the only way in which it can
depend on the ﬁrstorder gradient is through s, the reduced dimensionless gradient. Thus we
may write any GGA for exchange, that satisﬁes the linear scaling relation, as
E
GGA
X
[n] =
d
3
r e
unif
X
(n(r))F
X
(s(r)) (18.1)
The factor F
X
is the exchange enhancement factor, and tells you how much exchange is
enhanced over its LDA value. It is the analog of F
unif
XC
(r
s
) from before. In Fig. 18.3, we plot
1
1.2
1.4
1.6
1.8
0 0.5 1 1.5 2 2.5 3
Exchange
hole
GEA
B88
PBE
F
X
(
s
)
s = ∇n/(2kF n)
Figure 18.3: Exchange enhancement factors for diﬀerent GGA’s
various enhancement factors. First note that F
X
= 1 corresponds to LDA. Then the dashed
line is GEA, which has an indeﬁnite parabolic rise. The solid line is the result of the real
space cutoﬀ procedure. It clearly produces a curve whose enhancement is about double that
of GEA for small and moderate s, because of the elimination of positive contributions to the
150 CHAPTER 18. GENERALIZED GRADIENT APPROXIMATION
hole. But the holeGGA curve also grows less rapidly at large s, because of the normalization
cutoﬀ.
Also included in the ﬁgure are two modern GGAs, (Becke 88 and PerdewWang 91) that
we discuss more later. They both agree with the general shape of the holeGGA for moderate
values of s, making the realspace cutoﬀ procedure a justiﬁcation for any of them. The most
signiﬁcant deviation in this region is that the holeGGA has an upward bump near s = 1.
This is due to the negativity and normalization cutoﬀs meeting, as is about to happen in Fig.
?? for slightly larger s, and is an artifact of the crude construction. Recalling that GGA is
still an approximation, the diﬀerences between diﬀerent GGA’s are probably of the order of
the intrinsic error in this approximate form.
Figure 18.4: Sphericallyaveraged correlation hole density nC for rs = 2 and ζ = 0. GEA holes are shown for four values of
the reduced density gradient, t = ∇n/(2ksn). The vertical lines indicate where the numerical GGA cuts oﬀ the GEA hole
to make
vC
0
dv 4π v
2
nC(v) = 0.
The case for correlation is very similar. Here there is no negativity constraint, and the
realspace construction of the GGA correlation hole simply corrects the lack of normalization
in the GEA correlation hole. Fig. 18.4 shows what happens. The line marked 0 is the LDA
correlation hole. Note its diﬀuseness. Here the natural length scale is not k
F
u, but rather k
s
u,
because we are talking about correlation. Similarly reduced gradients are measured relative
to this length scale. The other curves are GEA holes for increasing reduced gradients. Note
that the GEA hole is everywhere positive, leading to consistently positive corrections to LDA
energies. Note also how large it becomes at say t = 1.5, producing a stongly positive energy
contribution. For small reduced gradients, the gradient correction to the hole is slight, and
has little eﬀect on the LSD hole, until at large distances, it must be cutoﬀ. As the gradient
grows, the impact on the LSD hole becomes much greater, and the cutoﬀ shrinks toward
zero. Since the energy density contribution is weighted toward u = 0 relative to this picture,
18.2. VISUALIZING AND UNDERSTANDING GRADIENT CORRECTIONS 151
and all curves initially drop, the GGA correlation hole always produces a negative correlation
energy density.
0.08
0.04
0
0 1 2 3
He atom
exact
LDA
GGA
2
π
u
n
λ
=
1
C
(
u
)
u
Figure 18.5: Systemaveraged correlation hole density at full coupling strength (in atomic units) in the He atom. The exact
curve (solid) is from a CI calculation, the long dashed is LDA, and the shortdashes are realspace cutoﬀ GGA. The area
under each curve is the potential contribution to correlation energy.
In Fig. 18.5, we see the results for the systemaveraged hole, which are even more dramatic
than the exchange case. At small separations, the GEA correclty makes the LDA hole less
deep, as expected. But now, instead of producing a wild (i.e., unnormalized) postive peak,
the postive contribution is much less pronounced, and is controlled by the normalization.
The upward bump and rapid turnoﬀ beyond u = 2 are artifacts of the crude realspace cutoﬀ
procedure, and remind us of how little we know of the details of the corrections to LDA. But
the overall eﬀect on the energy is very satisfying. The LDA hole spreads out enormously,
giving rise to a correlation energy that is at least a factor of 2 too deep, while the realspace
cutoﬀ hole has produced something similar to the true hole.
18.2 Visualizing and understanding gradient corrections
To understand gradient corrections for exchangecorrelation, we begin with the spinunpolarized
case, where we write
E
GGA
XC
[n] =
d
3
r e
unif
X
(n(r)) F
XC
(r
s
(r), s(r)). (18.2)
This enhancement factor contains all our others as special cases:
F
XC
(r
s
, s = 0) = F
unif
XC
(r
s
), F
XC
(r
s
= 0, s) = F
X
(s). (18.3)
We plot it for the case of the PBE functional in Fig 18.6. We will discuss the various kinds of
GGA functionals that have been developed and are in use later, but for now we simply note
that this GGA was designed to recover the GGA for the realspace correction of the GEA hole
for moderate s, and includes also several other energetically signiﬁcant constraints.
152 CHAPTER 18. GENERALIZED GRADIENT APPROXIMATION
1
1.1
1.2
1.3
1.4
1.5
1.6
0 0.5 1 1.5 2 2.5 3
rs
0
1/2
2
4
10
F
X
C
(
r
s
,
s
)
s = ∇n/(2k
F
n)
Figure 18.6: Exchangecorrelation enhancement factors for the PBE GGA.
These curves contain all the physics (and ultimately) chemistry behind GGA’s. We make
the following observations:
• Along the yaxis, we have F
unif
XC
(r
s
), the uniform gas enhancement factor.
• The eﬀect of gradients is to enhance exchange. We have already seen this in action in
the previous sections. Essentially, in the positive direction of the gradient, the density
is increased, while dropping on the opposite side. This allows the center of the hole to
move in that direction, becoming deeper due to the higher density, and yielding a higher
exchange energy density.
• The eﬀect of gradients is turn oﬀ correlation relative to exchange. This is because, in
regions of high gradient the exchange eﬀect keeps electrons apart, so that their correlation
energy becomes relatively smaller. In our realspace analysis of the hole, we see that large
gradients ultimately eliminate correlation.
• A simple picture of the correlation eﬀects is given by the observation that for t up to
about 0.5, the GGA make little correction to LDA, but then promptly cuts oﬀ correlation
as t grows, as in Fig. ??. But t is the reduced gradient for correlation, deﬁned in terms
of the screening wavevector, not the Fermi wavevector. For an unpolarized system:
t =
∇n
2k
s
n
=
k
F
k
s
s ≈
1.5
r
s
s (18.4)
Thus, for r
s
about 2, t = s, and we see correlation turning oﬀ at t about 0.5. For
r
s
= 10, t ≈ 0.4s, so that correlation turns oﬀ about s = 1.2, while for r
s
= 1/2,
t ≈ 1.7s, so that correlation turns oﬀ very quickly, at s about 0.3.
• An interesting consequence of the point above is that pure exchange is the least local
curve, and that as correlation is turned on, the functional becomes more local, i.e., closer
to the LDA value, both in the sense that gradient corrections become signiﬁcant at larger
s values, and also that even when they do, their magnitude is less.
18.3. EFFECTS OF GRADIENT CORRECTIONS 153
We can repeat this analysis to understand the eﬀects of spin polarization. Must ﬁll this in.
Exercise 85 Deduce the maximum value for F
XC
allowed by the LiebOxford bound.
Exercise 86 The lines of diﬀerent r
s
values never cross in Fig. Y. What does this imply
about the functional?
18.3 Eﬀects of gradient corrections
In this section, we survey some of the many properties that have been calculated in DFT
electronic structure calculations, and try to understand the errors made LSD, and why GGA
corrects them the way it does.
• Both the large underestimate in the magnitude of total exchange energies and the over
estimate in the magnitude of correlation energies can be seen immediately from Fig.
18.6. Ignoring gradients ignores the enhancement of exchange and the truncation of
correlation.
• In studying ionization energies of atoms, one ﬁnds LSD overbinds s electrons relative to
p, p relative to d, and so forth. Since core electrons have higher density than outer ones,
the local approximation is less accurate for them.
• When a molecule is formed from atoms, the density becomes more homogeneous than
before. The region of the chemical bond is ﬂatter than the corresponding regions of
exponential decay in atoms. Thus LDA makes less of an underestimate of the energy of
a molecule than it does of the constituent atoms. Write
E
XC
(system) = E
LDA
XC
(system) +∆E
LDA
XC
(system) (18.5)
Then ∆E
XC
(atoms) < ∆E
XC
(molecules), remebering that both are negative. The
atomization energy of a molecule is
E
atmiz
XC
= E
XC
(molecule) −E
XC
(atoms) (18.6)
so that
E
XC
(atom) = E
LDA
XC
(atom) +∆E
LDA
XC
(molecule) −E
LDA
XC
(atom) (18.7)
Thus molecules are overbound in LDA, typically by as much as 30 kcal/mol.
• Transition state barriers can be understood in much the same way. A typical transition
state in quantum chemistry is one of higher symmetry and coordination than the reac
tants. Thus comparing the transition state to the reactants, in just the way done above
for a bond and its constituent atoms, we ﬁnd that the transitionstate barrier, deﬁned as
E
trans
XC
= −E
XC
(reactants) +E
XC
(transitionstate) (18.8)
154 CHAPTER 18. GENERALIZED GRADIENT APPROXIMATION
will be too low in an LDA calculation, and be raised by a GGA calculation. Notice though
that this depends on the coordination argument. If the transition state is less coordinated,
the result is the reverse: LDA overestimates the barrier, and GGA will weaken it. This is
often the case for, eg. surface diﬀusion, where the transition state (e.g. bridgebonded)
has lower symmetry than the initial state (e.g. atop bonded).
• By the same reasoning as for atomization energies, consider the process of stretching a
bond. As a bond is stretched, LDA’s overbinding tendancy will reduce. Thus equilibrium
bond lengths are usually too small in LDA, and GGA stretches them. The only exception
to this is the case of bonds including H atoms. One can show that typically GGA favors
a process in which
∆s
s
≥ P
∆r
s
2r
s
+Q∆ζ, (18.9)
where s, etc., are precisely chosen averages, and where P is typically close to 1, and
Q ≈ 0. Usually fractional changes in the density are very small, due to the presence of
core electrons, and the right hand side is thought of as zero. This leads to the generic
claim that GGA prefers inhomogeneity, i.e., LDA overfavors homogeneity. For example, in
our cases above, when a molecule is atomized, its mean gradient increases. The inequality
means that GGA will like this more than LDA, and so have a smaller atomization energy.
Neglecting the righthandside works for all atomization energies, in which the changes
are large. But for a bondstretch, the fractional gradient changes are much smaller. In
the case of H atoms, the density is suﬃciently low as to make its fractional change larger
than that of the gradient, when stretching a bond with an H atom. Thus in that case,
the gradient corrections have the reverse eﬀect, and those bonds are shortened.
18.4 Satisfaction of exact conditions
We can now look back on the previous chapters, and check that the realspace cutoﬀ GGA
satisﬁes conditions that LDA satisﬁes, and maybe a few more.
• Sizeconsistency Obviously, GGA remains size consistent.
• The LiebOxford bound.
The LO bound gives us a maximum value for the enhancement factor of 1.804. The
realspace cutoﬀ will eventually violate this bound, but only at large s, where we do not
trust it anyhow.
• Coordinate scaling
For exchange, clear cos only s dependence.
Show F
XC
lines never cross.
• Virial theorem Trivially true.
18.5. A BRIEF HISTORY OF GGA’S 155
• Highdensity correlation
The numerical GGA satisﬁes this condition. As we show below, many chemical systems
are close to this, and, by truncating the longranged Coulomb correlation hole, numerical
GGA yields a ﬁnite value.
• Selfinteraction error
Simple density functionals like LDA and GEA naturally have a selfinteraction error,
because they cannot tell when there is only one electron in the system. While GGA
might improve numerically the value of E
X
or E
C
for that one electron, it will still have
a residual selfinteraction error.
• Symmetry dilemma
Just as above, nothing in GGAs construction helps here.
• Potentials
From the naive sense...
• Koopman’s theorem No real improvement here.
18.5 A brief history of GGA’s
We have seen how the realspace cutoﬀ construction imposed on the gradient expansion
produces a numericallydeﬁned GGA. This is not unique, in that, e.g. a smoother cutoﬀ will
lead to slightly diﬀerent curves for moderate gradients, and perhaps much diﬀerent curves
for large gradients.
18.6 Questions about generalized gradient approximations
156 CHAPTER 18. GENERALIZED GRADIENT APPROXIMATION
Chapter 19
Hybrids
19.1 Static, strong, and strict correlation
19.2 Mixing exact exchange with GGA
19.3 Questions about adiabatic connection formula and hybrids
1. Does the HellmannFeynman theorem apply to the problem of the particle in a potential
V (x) = exp(−x)?
2. What happens to a system as λ →∞?
3. What is the formula for v
λ
ext
[n](r) in terms of a scaled density?
157
158 CHAPTER 19. HYBRIDS
Chapter 20
Orbital functionals
20.1 Selfinteraction corrections
20.2 Optimized potential method
20.3 G¨ orlingLevy perturbation theory
20.4 Meta GGA’s
20.5 Jacob’s ladder
159
160 CHAPTER 20. ORBITAL FUNCTIONALS
Part V
Timedependent DFT
161
Chapter 21
Timedependence
21.1 Schr¨ odinger equation
We learn in kindergarten that the evolution of the wavefunction is governed by the time
dependent Schr¨ odinger equation:
i
˙
Ψ(t) =
ˆ
H(t) Ψ(t), Ψ(0) given (21.1)
where
ˆ
H(t) =
ˆ
T +
ˆ
V (t), and a dot implies diﬀerentiation with respect to time. We will
generally be concerned with a timedependence of the form
ˆ
H(t) =
ˆ
H
o
+
ˆ
V (t) (21.2)
where the unperturbed Hamiltonian, H
o
has eigenstates
ˆ
H
o
j = E
j
j (21.3)
and
ˆ
V (t) is zero for t < 0. We may then expand the timedependent wavefunction in terms
of these unperturbed eigenstates:
Ψ(t) =
¸
j
C
j
(t) exp(−iE
j
t) j (21.4)
where the phase factor for undisturbed evolution has been extracted. Insertion into Eq. (21.1)
yields
i
˙
C
j
(t) =
¸
k
C
k
exp(iω
jk
t) V
jk
(t), (21.5)
a coupled set of ﬁrstorder diﬀerential equations, with V
jk
= j
ˆ
V k, ω
jk
= E
j
− E
k
and
with initial values
C
j
(0) = Ψ(0)j. (21.6)
We usually begin in the groundstate, so that C
j
(0) = δ
j0
.
To have a feeling for what timedependent quantum mechanics really means, we immedi
ately show some examples for a simple case. We take our particle in a box and subject it to
a particular driving potential,
V (xt) = x E(t) (21.7)
163
164 CHAPTER 21. TIMEDEPENDENCE
This is the potential due to an external timedependent electric ﬁeld in the dipole approxi
mation.
Exercise 87 Matrix elements of timedependent potentials
Calculate the matrix elements V
jk
(t) for (a) a harmonic oscillator and (b) a particle in a box,
when the perturbation is an electric ﬁeld.
We choose
E(t) = A sin(ωt) (1 −exp(−ωt)) (21.8)
This form allows a slow turnon of a periodic electric ﬁeld of frequency ω and amplitude
A. Our ﬁrst example is taken with generic values, A = 10, ω = 4, and the time run up to
t = 4. In Fig. 21.1, we show how the system responds. The density barely changes from its
groundstate value, but simply jiggles to left and right. The mean position begins to oscillate
immediately, not with the frequency of the driving ﬁeld, but rather with a period of about
1/3. This corresponds to the lowest transition, since
1
= π
2
/2 = 4.93, and
2
= 4
1
, so
that ω
21
= 14.8, and the period is 2π/ω ∼ 0.4. Thus the timedependent evolution of the
10
8
6
4
2
0
2
4
6
8
10
0 0.5 1 1.5 2 2.5 3 3.5 4
E
(
t
)
t
6
5
4
3
2
1
0
1
2
3
4
0 0.5 1 1.5 2 2.5 3 3.5 4
1
0
0
<
x
(
t
)
>
t
Figure 21.1: Electric ﬁeld and expectation value of x for a particle in a box of length 1 subjected to the timedependent ﬁeld
shown and given in the text.
expectation value of operators contains information about transition frequencies.
Next, we illustrate the concept of resonance. We run our calculation with a = 1, but
sweep the driving frequency through ω = ω
12
, as shown in Fig. 21.2. We see that, the nearer
the driving frequency is to the transition frequency, the greater the eﬀect of the driving ﬁeld
on the particle. As ω →ω
21
, the perturbation grows indeﬁnitely and pushes more and more
energy into the particle.
Another interesting regime of timedependent problems is when the time variation is very
slow. This is the adiabatic limit. For our particle in a box in an external electric ﬁeld, if the
external frequency ω ω
12
, then the chances of exciting out of the groundstate in any small
interval are exponentially small. Over long times, as the potential changes signiﬁcantly from
its original, these small chances add up, and the wavefunction does evolve. But if, instead
21.1. SCHR
¨
ODINGER EQUATION 165
1
0
1
2
3
4
5
6
7
0 0.5 1 1.5 2
1
0
0
<
x
(
t
)
>
t
t=13
t=14
t=15
t=16
4.9
5
5.1
5.2
5.3
5.4
0 0.5 1 1.5 2
<
H
(
t
)
>
t
t=13
t=14
t=15
t=16
Figure 21.2: Expectation value of x and energy of a particle in a box of length 1 subjected to timedependent ﬁelds of
frequencies ω = 13, 14, 15, and 16.
of using the unperturbed eigenfunctions as our basis, we used the instantaneous eigenstates,
i.e., those satisfying
ˆ
H(t) Ψ
n
(t) = E
n
(t) Ψ(t) (21.9)
these chances would remain exponentially small, and so the system stays in the adiabatic
ground state. Essentially, the potential is moving so slowly that the wavefunction of the
particle simply rearranges itself to the new ground state at any given time.
In Fig X, I’ve calculated the case where ω = 1, t = 0.1, and a = 2000. Thus ω is much
lower than the lowest transition, and ωt 1. This yields a ﬁnal electric ﬁeld of a(ωt)
2
/2
or 10. The particle adiabatically follows the groundstate of the potential. This limit can be
found by simply turning on a static external potential, and asking what the groundstate is.
Exercise 88 Harmonic oscillator in a timedependent electric ﬁeld
Consider a harmonic oscillator of frequency ω
0
in an electric ﬁeld. If it starts out at rest at
the origin, show that its classical position is given by
x
cl
(t) = −
t
0
dt
E(t
)
mω
0
sin(ω
0
(t −t
)) (21.10)
Then show that
φ(xt) = φ
0
(x −x
cl
(t)) exp(iα(t)) (21.11)
is a solution to the timedependent Schr¨ odinger equation. How important is α(t)?
In this framework, the timedependent density is given by the usual formula of integrating
the square of the wavefunction over all coordinates but one. The density operator may be
written as
ˆ n(r) =
N
¸
i=1
δ(r −r
i
) (21.12)
and
n(rt) = Ψ(x)ˆ n(r)Ψ(x) (21.13)
166 CHAPTER 21. TIMEDEPENDENCE
. For our particle in a box, it is simply
n(xt) = Ψ(xt)
2
=
∞
¸
j,k=1
c
∗
j
(t) c
k
(t) φ
j
(x) φ
k
(x) (21.14)
A new quantity, that will play a central role in the development of TDDFT, is the current
density. The operator is deﬁned as
ˆ
j(r) =
1
2
N
¸
i=1
(p
i
δ(r −r
i
) + δ(r −r
i
)p
i
) (21.15)
and its expectation value given by evaluation on the timedependent wavefunction. For our
particle in a box,
j(xt) = φ
∗
(xt)dφ(xt)/dx (21.16)
Since, in any dx around a point x, the change in the number of electrons per unit time must
equal the diﬀerence between the ﬂow of electrons into dx from the left and out from the
right,
˙ n(xt) = −dj(xt)/dx (21.17)
This is the equation of continuity. More generally, the time evolution of any (timeindependent)
operator is
ˆ
˙
A = i[
ˆ
H(t),
ˆ
A] (21.18)
Applying this to the density yields
˙ n(rt) = Ψ(t)i[
ˆ
T, ˆ n(r)]Ψ(t) = −∇j(x) (21.19)
where the density commutes with all but the kinetic operator in the Hamiltonian, and its
commutation (with a little algebra) yields the divergence of the current operator.
21.2 Perturbation theory
This concept is best seen in perturbation theory. Imagine integrating the equations of motion
for the coeﬃcients just a small amount in time, so that the excited states are only inﬁnites
imally populated. Then one may insert the initial condition into the righthandside of Eq.
X to ﬁnd the probability of excitation. Next consider the eﬀect of varying the external fre
quency, and one ﬁnds a huge response as it passes through a transition frequency. We ﬁnd
the transition rate, i.e., the probability of excitation per unit time, to be
W(k) =
˙
P(k) = ... (21.20)
which is known as Fermi’s golden rule.
The full range of behaviors of timedependent systems is unnervingly large. In many cases,
we are interested only in the response to a weak potential, i.e.,
ˆ
H(t) =
ˆ
T +
ˆ
V + δ
ˆ
V (t) (21.21)
21.2. PERTURBATION THEORY 167
where δ
ˆ
V (t) is in some sense small. Usually we will consider potentials whose timedependence
is exp(iωt +ηt), where η →0, and the perturbation begins at −∞, and is fully turned on by
t = 0. Under these conditions, we can ask what is the change in the system to ﬁrstorder in
the perurbation? More precisely, we want to know the change in observables of the system.
Since this is a book on density functional theory, a key role will be played by the change in
density in response to such a perturbation. We deﬁne the susceptibility of the system by
δρ(xt) =
dx
dt
χ(x, x
; t −t
) δv(x
t
) (21.22)
i.e., χ tells you the density change throughout the system induced by any external potential.
It is also known as the densitydensity response function and the nonlocal (meaning depends
on x and x
) susceptibility. Since we are doing DFT, we will write it more neatly as
χ[n
0
](x, x
; t −t
) = δρ(xt)/δv(x
t
) (21.23)
i.e., it is a functional derivative, and depends only on the groundstate density (for a non
degenerate system). Note that it is a function of t −t
alone, since the unperturbed system
is static. Furthermore, we will speak of the retarded function, meaning χ = 0 for t
< t. We
show in the rest of this section that χ contains most of what we want to know about the
response of a system.
Next we specialize further to the speciﬁc case of the optical response of a system, i.e., to
a long wavelength electric ﬁeld. We want to ﬁnd out the timedependent dipole moment of
our system, deﬁned by
d(t) =
dx ρ(xt)x (21.24)
ie the ﬁrst moment of the density. This is because the optical absorption is given by ... If we
set v(x
ω) = Ex
, we ﬁnd
α(ω) =
dx x
dx
x
χ(x, x
ω) (21.25)
for the dynamic polarizability, and
σ(ω) = α(ω)ω/c (21.26)
ie the moments of the susceptibility yield the complex dynamic susceptibility, which in turn
give the optical response of the system. One can show this satisﬁes the wellknown Thomas
ReichKuhn sumrule,
∞
0
dω σ(ω) =
2π
2
c
N (21.27)
Thre is also the static sumrule, which follows from the general KramersKronig rule for the
polarizability:
2
π
∞
0
dω
ω
α(ω) = α(0) (21.28)
168 CHAPTER 21. TIMEDEPENDENCE
To see better what this means, we can write χ in its Lehman representation, i.e., in terms
of the levels of the system:
χ(xx
ω) = φ
0
(x)φ
0
(x
)
¸
k=j
... (21.29)
We deﬁne:
n
jk
(x) = jˆ n(x)k (21.30)
the generalized transition dipole. Thus, taking the moments to construct α, we ﬁnd
x
jk
=
dx x n
jk
(x) (21.31)
are the transition dipole moments for each transition. Then
α(ω) =
¸
j=0
x
0j

2
ω
2
−ω
2
j0
(21.32)
To see the eﬀect of speciﬁc transitions on the optical spectrum, we must use the wellknown
formula:
1
ω −ω
0
= P
1
ω −ω
0
+iπ δ(ω −ω0) (21.33)
ie the real part is a principal value, and the imaginary part a deltafunction. Thus
σ(ω) =
¸
j=0
2ω
j0
x
j0

2
δ(ω −ω
0
) (21.34)
The optical spectrum consists of a series of delta functions, whose intensity is determined by
the dipole matrix elements.
Exercise 89 Dipole matrix elements
Calculate the dipole matrix elements for optical absorption for (a) the harmonic oscillator and
(b) the particle in a box.
We deﬁne the oscillator strength of a given transition as
f
ji
= 2ω
ji
x
ji

2
/3 (21.35)
These satisfy the TRK sumrule,
¸
k=j
f
j
jk = N (21.36)
Note that all oscillator strengths from the groundstate are positive, but from an excited
state, some are negative.
Exercise 90 Oscillator strengths
From above, now ﬁnd oscillator strengths for both systems and show that they satisfy the
ThomasReichKuhn sumrule.
21.3. OPTICAL RESPONSE 169
When discussing ﬁrstorder perturbation theory, it is always useful to deﬁne response func
tions of the system at hand. A wellknown one is the Green’s function, which, for a one
electron system, may be written in operator form as
ˆ
G(z) = (z −
ˆ
H)
(−1)
(21.37)
where z is a complex number with positive imaginary part, and boundary conditions are
chosen so that G(x, x
) vanishes at large distances, just like the boundstates. In real space,
G satisﬁes the diﬀerential equation:
−
1
2
∇
2
+v(r) −z
G(rr
; z) = δ
(3)
(r −br
) (21.38)
Formally, one may write G in terms of sums over the eigenstates
G(rr
; z) =
¸
i
φ
∗
i
(r) φ
i
(r
)
z −
i
. (21.39)
Note that the Green’s function is illdeﬁned whenever its argument matches that of an eigen
value. In such cases, we deﬁne
G
±
(rr
, λ) = G(rr
, λ ±iη) (21.40)
where + refers to the retarded Green’s function, while − refers to the advanced one. When
Fouriertransformed, the Green’s function yields the timeevolution of any initial state, for a
timeindependent Hamiltonian:
Ψ(rt) =
d
3
r
dt G
+
(rr
, t −t
) Ψ(r
t
) (21.41)
since
ˆ
G
+
(t −t
) = exp(−i
ˆ
H(t −t
)) (21.42)
Simple expression for G in terms of L and R phi.
Relate density matrix to equaltime G
Deﬁne χ and discuss meaning
Relate χ and G for noninteracting systems.
Relate pair density to equaltime χ.
21.3 Optical response
Deﬁne polarizability, static and dynamic.
Deﬁne oscillator strength.
Deﬁne optical spectrum.
170 CHAPTER 21. TIMEDEPENDENCE
21.4 Questions about timedependent quantum mechanics
1. If a system is in a linear combination of two eigenstates, does its density change?
2. Calculate the oscillator strengths for a particle in a box, and show that they satisfy the
ThomasReich Kuhn sum rule.
3. Calculate the imaginary part of the frequencydependent susceptibility of the 1d Hatom.
4. Calculate the transmission and reﬂection amplitudes of the 1d Hatom, assuming a par
ticle incident from the left.
5. Calculate the retarded green’s function for the 1d Hatom.
6. Using the ﬁrst two levels of the particle in a box, calculate the Rabi frequency, and
compare its time evolution with that of Fig X.
7. Show how a harmonic oscillator evolves in a timedependent electric ﬁeld.
8. Calculate the oscillator strengths for the harmonic oscillator, and show that they satisfy
the ThomasReich Kuhn sum rule.
Chapter 22
Timedependent density functional theory
So far, we have restricted our attention to the ground state of nonrelativistic electrons in time
independent potentials. This is where most DFT research has been done, and has achieved its
great successes, by providing a general tool for tackling electronic structure problems. In this
chapter, we extend our view, to include timedependent potentials. In subsequent chapters,
we apply the methods of DFT to such potentials.
22.1 Overview
We will see below that, under certain quite general conditions, one can establish a oneto
one correspondence between timedependent densities n(rt) and timedependent onebody
potentials v
ext
(rt), for a given initial state. That is, a given evolution of the density can be
generated by at most one timedependent potential. This is the timedependent analog of
the HohenbergKohn theorem, and is called the RungeGross theorem. Then one can deﬁne a
ﬁctitious system of noninteracting electrons moving in a KohnSham potential, whose density
is precisely that of the real system. The exchangecorrelation potential (deﬁned in the usual
way) is then a functional of the entire history of the density, n(x), the initial interacting
wavefunction Ψ(0), and the initial KohnSham wavefunction, Φ(0). This functional is a very
complex one, much more so than the groundstate case. Knowledge of it implies solution of
all timedependent Coulombinteracting problems.
An obvious and simple approximation is adiabatic LDA (ALDA), sometimes called time
dependent LDA, in which we use the groundstate potential of the uniform gas with that
instantaneous and local density, i.e., v
XC
[n](x) ≈ v
unif
XC
(n(x)). This gives us a working Kohn
Sham scheme, just as in the ground state. We can then apply this DFT technology to all
problems of timedependent electrons. These applications fall into three general categories:
nonperturbative regimes, linear (and higherorder) response, and groundstate applications.
The ﬁrst of these involves atoms and molecules in intense laser ﬁelds, in which the ﬁeld
is so intense that perturbation theory does not apply. In these situations, the perturbing
electric ﬁeld is comparable to or much greater than the static electric ﬁeld due to the nuclei.
Experimental aims would be to enhance, e.g., the 27th harmonic, i.e., the response of the
171
172 CHAPTER 22. TIMEDEPENDENT DENSITY FUNCTIONAL THEORY
system at 27 times the frequency of the perturbing electric ﬁeld, or to cause a speciﬁc chemical
reaction to occur (quantum control). Previously, only one and two electron systems could
be handled computationally, as the full timedependent wavefunction calculations are very
demanding. Crude and unreliable approximations had to be made to tackle larger systems.
But with the advent of TDDFT, which has become very popular in this community, larger
systems with more electrons can now be tackled.
When the perturbing ﬁeld is weak, as in normal spectroscopic experiments, perturbation
theory applies. Then, instead of needing knowledge of v
XC
for densities that are chang
ing signiﬁcantly with time, we only need know this potential in the vicinity of the initial
state, which we take to be a nondegenerate groundstate. These changes are character
ized by a new functional, the exchangecorrelation kernel. While still a more complex beast
than the groundstate exchangecorrelation potential, the exchangecorrelation kernel is much
more manageable than the full timedependent exchangecorrelation potential, because it is a
functional of the groundstate density alone. Analysis of the linear response then shows that
it is (usually) dominated by the response of the groundstate KohnSham system, but cor
rected by TDDFT via matrix elements of the exchangecorrelation kernel. In the absence of
Hartreeexchangecorrelation eﬀects, the allowed transitions are exactly those of the ground
state KohnSham potential. But the presence of the kernel shifts the transition frequencies
away from the KohnSham values to the true values. Moreover, the intensities of the optical
transitions can be extracted in the same calculations, so these also are aﬀected by the kernel.
Chemists and physicists have devised separate approaches to extracting excitations from
TDDFT for atoms, molecules, and clusters. The chemists way is to very eﬃciently convert
the search for poles of response functions into a large eigenvalue problem, in a space of
the singleexcitations of the system. The eigenvalues yield transition frequencies while the
eigenvectors yield oscillator strengths. This allows use of many existing fast algorithms to
extract the lowest few excitations. In this way, TDDFT has been programmed into most
standard quantum chemical packages and, after a molecule’s structure has been found, it is
usually not too costly to extract its lowlying spectrum. Physicists, on the other hand, prefer
to simply solve the timedependent KohnSham equations when a weak ﬁeld has been turned
on. Fouriertransform of the timedependent dipole matrix element then yields the optical
spectrum. Using either methodology, the number of these TDDFT response calculations
for transition frequencies is growing exponentially at present. Overall, results tend to be
fairly good (0.1 to 0.2 eV errors, typically), but little is understood about their reliability.
Diﬃculties remain for the application of TDDFT to solids, because the present generation of
approximate functionals (local and semilocal) lose important eﬀects in the thermodynamic
limit, but much work is currently in progress.
The last class of application of TDDFT is, perhaps surprisingly, to the groundstate prob
lem. This is because one can extract the groundstate exchangecorrelation energy from a
response function, in the same fashion as perturbation theory yields expressions for ground
22.2. RUNGEGROSS THEOREM 173
state contributions in terms of sums over excited states. Thus any approximation for the
exchangecorrelation kernel yields an approximation to E
XC
. Such calculations are signiﬁ
cantly more demanding than regular groundstate DFT calculations. Most importantly, this
produces a natural method for incorporating timedependent ﬂuctuations in the exchange
correlation energy. In particular, as a system is pulled apart into fragments, this approach
includes correlated ﬂuctuations on the two separated pieces. While in principle all this is
included in the exact groundstate functional, in practice TDDFT provides a natural method
ology for modeling these ﬂuctuations. Simple ALDA approximations for polarizabilities of
atoms and molecules allow C
6
coeﬃcients to be accurately calculated by this method, and
attempts exist to generate entire molecular energy curves, from the covalent bond distance
to dissociation, including van der Waals. Such calculations are much more demanding than
simple selfconsistent groundstate calculations, but may well be worth the payoﬀ.
22.2 RungeGross theorem
The analog of the HohenbergKohn theorem for timedependent problems is the RungeGross
theorem[?]. We consider N nonrelativistic electrons, mutually interacting via the Coulomb
repulsion, in a timedependent external potential. The RungeGross theorem states that the
densities n(x) and n
(x) evolving from a common initial state Ψ
0
= Ψ(t = 0) under the
inﬂuence of two potentials v
ext
(x) and v
ext
(x) (both Taylor expandable about the initial time
0) are always diﬀerent provided that the potentials diﬀer by more than a purely timedependent
(rindependent) function:
∆v
ext
(x) = v(x) −v
(x) = c(t) . (22.1)
If so, there is a onetoone mapping between densities and potentials, and one can construct
a density functional theory.
We prove this theorem by ﬁrst showing that the corresponding current densities must diﬀer.
Because the Hamiltonians only diﬀer in their external potentials, the equation of motion for
the diﬀerence of the two current densities is:
∆
˙
j(x)
t=0
= −iΨ
0

¸
ˆ
j(r), ∆
ˆ
H(t
0
)
Ψ
0
= −n
0
(r)∇∆v
ext
(r, 0), (22.2)
where n
0
(r) = n(r, 0) is the initial density. One can go further, by repeatedly using the
equation of motion, and considering t = 0, to ﬁnd[?]
∂
k+1
0
∆j(x) = −n
0
(r) ∇∂
k
0
∆v
ext
(x) , (22.3)
where ∂
k
0
= (∂
k
/∂t
k
)
t=0
, i.e., the kth derivative evaluated at the initial time. If Eq. (22.1)
holds, and the potentials are Taylor expandable about t
o
, then there must be some ﬁnite k
for which the right hand side of (22.2) does not vanish, so that
∆j(x) = 0. (22.4)
174 CHAPTER 22. TIMEDEPENDENT DENSITY FUNCTIONAL THEORY
This is the ﬁrst part of the RungeGross theorem, which establishes a onetoone correspon
dence between current densities and external potentials.
In the second part, we extend the proof to the densities. From continuity,
∆˙ n(x) = −∇· ∆j(x) (22.5)
so that
∂
k+2
0
∆n(x) = ∇·
n
0
(r)∇∂
k
0
∆v
ext
(x)
. (22.6)
Now, if not for the divergence on the righthandside, we would be done, i.e., if f(r) =
∂
k
0
∆v
ext
(x) is nonzero for some k, then the density diﬀerence must be nonzero, since n
0
(r)
is nonzero everywhere.
Does the divergence allow some escape from this conclusion? The answer is no, for any
physical density. To see this, write
d
3
r f(r)∇· (n
0
(r)∇f(r)) =
d
3
r
∇· (f(r)n
0
(r)∇f(r)) −n
0
(r)∇f(r)
2
(22.7)
The ﬁrst term on the right may be written as a surface integral at r = ∞, and vanishes for
any realistic potential (which fall oﬀ at least as fast as −1/r). If we imagine that ∇f(r)
is nonzero somewhere, then the second term on the right is deﬁnitely negative, so that the
integral on the left cannot vanish, and its integrand must be nonzero somewhere. Thus
there is no way for ∇f(r) to be nonzero, and yet have f∇(n
0
∇f) vanish everywhere.
Since the density determines the potential up to a timedependent constant, the wavefunc
tion is in turn determined up to a timedependent phase, which cancels out of the expectation
value of any operator.
22.3 KohnSham equations
We will see below that, under certain quite general conditions, one can establish a oneto
one correspondence between timedependent densities n(rt) and timedependent onebody
potentials v
ext
(rt), for a given initial state. That is, a given history of the density can be
generated by at most one history of the potential. This is the timedependent analog of the
HohenbergKohn theorem. Then one can deﬁne a ﬁctious system of noninteracting electrons
that satisfy timedependent KohnSham equations
i
˙
φ
j
(rt) =
¸
−
∇
2
2
+v
S
[n](x)
¸
φ
j
(x) . (22.8)
whose density is
n(x) =
N
¸
j=1
φ
j
(x)
2
, (22.9)
precisely that of the real system. We deﬁne the exchangecorrelation potential via:
v
S
(x) = v
ext
(x) +v
H
(x) +v
XC
(x) (22.10)
22.4. ADIABATIC APPROXIMATION 175
where the Hartree potential has the usual form, but for a timedependent density. The
exchangecorrelation potential is then a functional of the entire history of the density, n(x),
the initial interacting wavefunction Ψ(0), and the initial KohnSham wavefunction, Φ(0).
This functional is a very complex one, much more so than the groundstate case. Knowledge
of it implies solution of all timedependent Coulomb interacting problems.
Exercise 91 KohnSham potential for one electron
By inverting the timedependent KohnSham equation, give an explicit expression for the
timedependent KohnSham potential in terms of the density for one electron.
In groundstate DFT, the exchangecorrelation potential is the functional derivative of
E
XC
[n]. It would be nice to ﬁnd a functional of n(rt) for which v
XC
(rt) was the functional
derivative, called the exchangecorrelation action. However, as shown in Sec. X, this turns
out not to be possible.
22.4 Adiabatic approximation
As we have noted, the exact exchangecorrelation potential depends on the entire history of
the density, as well as the initial wavefunctions of both the interacting and the KohnSham
systems:
v
XC
[n; Ψ(0), Φ(0)](rt) = v
S
[n; Φ(0)](rt) −v
ext
[n; Ψ(0)](rt) −v
H
[n](rt). (22.11)
However, in the special case of starting from a nondegenerate ground state (both interacting
and non), the initial wavefunctions themselves are functionals of the initial density, and so
the initialstate dependence disappears. But the exchangecorrelation potential at r and t
has a functional dependence not just on n(r
t) but on all n(r
t
) for 0 ≤ t
≤ t. We call this
a historydependence.
The adiabatic approximation is one in which we ignore all dependence on the past, and
allow only a dependence on the instantaneous density:
v
adia
XC
[n](rt) = v
approx
XC
[n(t)](r), (22.12)
i.e., it approximates the functional as being local in time. If the timedependent potential
changes very slowly (adiabatically), this approximation will be valid. But the electrons will
remain always in their instantaneous ground state. To make the adiabatic approximation
exact for the only systems for which it can be exact, we require
v
adia
XC
[n](rt) = v
gs
XC
[n
0
](r)
n0(r)=n(rt)
(22.13)
where v
gs
XC
[n
0
](r) is the exact groundstate exchangecorrelation potential of the density n
0
(r).
This is the precise analog of the argument made to determine the function used in LDA for
the groundstate energy (see Sec. X).
176 CHAPTER 22. TIMEDEPENDENT DENSITY FUNCTIONAL THEORY
In practice, the spatial locality of the functional is also approximated. As mentioned above,
ALDA is the mother of all TDDFT approximations, but any groundstate functional, such
as a GGA or hybrid, automatically yields an adiabatic approximation for use in TDDFT
calculations. A simple way to go beyond the adiabatic approximation would be to include
some dependence on, say, ˙ n(rt).
Bureaucratically, ALDA should only work for systems with very small density gradients in
space and time. It may seem like a drastic approximation to neglect all nonlocality in time.
However, we’ve already seen how well LDA works beyond its obvious range of validity for the
groundstate problem, due to satisfaction of sum rules, etc., and how its success can be built
upon with GGA’s and hybrids. We’ll soon see if some of the same magic applies to TDDFT.
Exercise 92 Continuity in ALDA
Does a KohnSham ALDA calculation satisfy continuity? Prove your answer.
22.5 Questions on general principles of TDDFT
1. If a system is in a linear combination of two eigenstates, does its density change?
2. In a timedependent KohnSham calculation starting from the ground state, will electronic
transitions occur at frequencies that are diﬀerences between groundstate KohnSham
orbital energies?
3. Is there anything special about the response of a particle in a harmonic well to an external
electic ﬁeld?
4. Does the RungeGross theorem guarantee that the current density in a TDDFT calcula
tion is correct?
5. Can more than one potential produce the same timedependent density?
6. If a system is subjected to a driving potential that, after a time, returns its density to its
initial value, will the potential return to its initial value?
7. Repeat question above, but replace density with wavefunction.
8. Repeat both questions above, within the adiabatic approximation. What happens if your
functional has some memory?
9. Consider a He atom sitting in its groundstate. Imagine you have a timedependent
potential that pushes it into its ﬁrst excited state (1s2s). Now imagine the exact Kohn
Sham calculation. Describe the ﬁnal situation.
22.5. QUESTIONS ON GENERAL PRINCIPLES OF TDDFT 177
Bibliography
The original proof of the RungeGross theorem is in Densityfunctional theory for time
dependent systems, E. Runge and E.K.U. Gross, Phys. Rev. Lett. 52, 997 (1984).
The extension to piecewise analytic potentials and the initialstate dependence is addressed
in Memory in timedependent density functional theory, N.T. Maitra, K. Burke, and C.
Woodward, Phys. Rev. Letts, 89, 023002 (2002).
For a simple demonstration of initialstate dependence, see Demonstration of initialstate
dependence in timedependent density functional theory, N.T. Maitra and K. Burke, Phys.
Rev. A 63, 042501 (2001); 64 039901 (E).
There was some controversy over the last step of the RungeGross proof, requiring the inte
gration by parts, but this is settled in E.K.U. Gross and W. Kohn, in Advances in Quantum
Chemistry, Vol. 21: Density Functional Theory of ManyFermion Systems, edited by S.
B. Trickey (Academic Press, San Diego, 1990).
178 CHAPTER 22. TIMEDEPENDENT DENSITY FUNCTIONAL THEORY
Chapter 23
Linear response
Note that the diﬀerence between n(x) and n
(x) is nonvanishing already in ﬁrst order of
v(x) −v
(x), ensuring the invertibility of the linear response operators of section ??.
23.1 Dysonlike response equation and the kernel
When the perturbing ﬁeld is weak, as in normal spectroscopic experiments, perturbation
theory applies. Then, instead of needing knowledge of v
XC
for densities that are changing
signiﬁcantly with time, we only need know this potential in the vicinity of the initial state,
which we take to be a nondegenerate groundstate. Writing n(x) = n(r) + δn(x), we have
v
XC
[n + δn](x) = v
XC
[n](r) +
dt
d
3
r
f
XC
[n](rr
, t −t
) δn(r
t
), (23.1)
where f
XC
is called the exchangecorrelation kernel, evaluated on the groundstate density:
f
XC
[n](rr
, t −t
) = δv
XC
(rt)/δn(r
t
). (23.2)
While still a more complex beast than the groundstate exchangecorrelation potential, the
exchangecorrelation kernel is much more manageable than the full timedependent exchange
correlation potential, because it is a functional of the groundstate density alone. To un
derstand why f
XC
is important for linear response, we deﬁne the pointwise susceptibility
χ[n
0
](rr
, t −t
) as the response of the ground state to a small change in the external poten
tial:
δn(x) =
dt
d
3
r
χ[n
0
](rr
, t −t
) δv
ext
(r
t
), (23.3)
i.e., if you make a small change in the external potential at point r
and time t
, χ tells you
how the density will change at point r and later time t. Now, the groundstate KohnSham
system has its own analog of χ, which we denote by χ
s
, which says how the noninteracting
KS electrons would respond to δv
S
(r
t
), which is quite diﬀerent from the interacting case.
But both must yield the same density response, so using the deﬁnition of the KS potential,
and of the exchangecorrelation kernel, we ﬁnd the key equation of TDDFT linear response:
χ(rr
ω) = χ
S
(rr
ω) +
d
3
r
1
d
3
r
2
χ
S
(rr
1
ω)
¸
1
r
1
−r
2

+f
XC
(r
1
r
2
ω)
¸
χ(r
2
r
ω), (23.4)
179
180 CHAPTER 23. LINEAR RESPONSE
where all objects are functionals of the groundstate density. This is called a Dysonlike
equation, because it has the same mathematical form as the Dyson equation relating the
oneparticle Green’s function to its free counterpart, with a kernel that is the onebody
potential. However, it is an oddity of DFT that it has this form, since χ is a twobody
response function.
This equation contains the key to electronic excitations via TDDFT. This is because when
ω matches a true transition frequency of the system, the response function χ blows up, i.e.,
has a pole as a function of ω. Thus χ
S
has a set of such poles, at the singleparticle excitations
of the KS system:
χ
S
(rr
ω) = 2
¸
iocc,aunocc
φ
∗
i
(r) φ
a
(r) φ
∗
a
(r
) φ
i
(r
)
ω −(
i
−
a
) +i0
+
−c.c.(ω →−ω) (23.5)
In the absence of Hartreeexchangecorrelation eﬀects, χ = χ
S
, and so the allowed transi
tions are exactly those of the groundstate KS potential. But the presence of the kernel in
Eq. (23.4) shifts the transitions away from the KS values to the true values. Moreover,
the strengths of the poles can be simply related to optical absorption intensities (oscillator
strengths) and so these also are aﬀected by the kernel.
In linear response, this implies that the exchangecorrelation kernel has the form:
f
adia
XC
(rr
, t −t
) =
δv
gs
XC
[n
0
](r)
δn
0
(r)

n0(r)=n(rt)
δ(t −t
). (23.6)
When Fouriertransformed, this implies f
adia
XC
is frequencyindependent.
23.2 Casida’s equations
The various methods for extracting excitations have been described in the overview. Here we
focus on the results. Casida showed that, ﬁnding the poles of χ is equivalent to solving the
eigenvalue problem:
¸
q
˜
Ω
qq
(ω)v
q
= Ωv
q
, (23.7)
where q is a double index, representing a transition from occupied KS orbital i to unoccupied
KS orbital a, ω
q
=
a
−
i
, Ω = ω
2
, and Φ
q
(r) = φ
∗
i
(r)φ
a
(r). The matrix is
˜
Ω
qq
(ω) = δ
qq
Ω
q
+ 2
ω
q
ω
q
qf
HXC
(ω)q
, (23.8)
where
qf
HXC
(ω)q
=
d
3
r
d
3
r
Φ
∗
q
(r)f
HXC
(r, r
, ω)Φ
q
(r
). (23.9)
In this equation, f
HXC
is the Hartreeexchangecorrelation kernel, 1/r − r
 + f
XC
(r, r
, ω).
A selfconsistent solution of Eq. (23.7) yields the excitation energies ω and the oscillator
strengths can be obtained from the eigenvectors [?]. In the special case of an adiabatic
approximation, these equations are a straightforward matrix equation. Algorithms exist for
extracting just the lowest transitions.
23.3. SINGLEPOLE APPROXIMATION 181
23.3 Singlepole approximation
182 CHAPTER 23. LINEAR RESPONSE
Chapter 24
Performance
24.1 Sources of error
24.2 Poor potentials
A problem was noticed early on. Our favorite approximations to the exchangecorrelation
energy yield fairly lousy exchangecorrelation potentials. These potentials are especially un
pleasant at large distances from Coulombic systems. They are too shallow, making Koopman’s
theorem fail drastically. This means that all the higher states are too shallow. In fact the
Rydberg series, i.e., an inﬁnite sequence of excitations as E → 0 from below, characteristic
of the −1/r decay, does not exist at all. How can one correct KS transition frequencies if
they do not exist?
The standard answer has been to asymptotically correct the potentials, to make them mimic
the true KS potential more accurately. A variety of schemes exist that do this to varying
degrees of accuracy. However, we note that all such schemes then produce a potential that
is not a functional derivative of an exchangecorrelation energy.
24.3 Transition frequencies
In any event, for many systems of real interest, it is the lowlying valence excitations that are
of interest, and in these regions, the potentials are quite accurate. Table X shows...
Hybrid functionals in TDDFT: An important point to note here concerns imple
mentation of hybrid functionals in TDDFT. There are two ways a hybrid might be coded,
which in the case of excitations, can yield very diﬀerent results. In the ﬁrst, ’correct’ DFT
approach, the optimized eﬀective potential method should be used, to guarantee that the
eﬀective potential is always a local, multiplicative one, as required by KohnSham theory.
However, in practice, HartreeFock calculations are much easier, and have been around since
time immemorial. Thus the potential produced in many quantum chemical codes consists
of a fraction of the HF nonlocal potential. This has negligible eﬀect for groundstate and
occupied orbitals, but a large eﬀect on the unoccupied levels.
183
184 CHAPTER 24. PERFORMANCE
At this point, we have surveyed all that one needs to perform a TDDFT calculation
of excitation energies using a modern code, either a quatum chemistry one or a realtime
evolution. The remainder of this chapter, and all of the next, are devoted to delving deeper
into understanding how TDDFT works and what are its current limitations.
In this section, we report calculations for the He and Be atoms using the exact ground
state KohnSham potentials, the three approximations to the kernel mentioned in the previous
section, and including many boundstate poles in Eq. (??), but neglecting the continuum.
The technical details are given in Ref. [?]. Table I lists the results, which are compared with
a highly accurate nonrelativistic variational calculations[?, ?] In each symmetry class (s, p,
and d), up to 38 virtual states were calculated. The errors reported are absolute deviations
from the exact values. The second error under the Be atom excludes the 2s →2p transition,
for reasons discussed in the next section.
The eﬀect of neglecting continuum states in these calculations has been investigated by van
Gisbergen et al.[?], who performed ALDA calculations from the exact KohnSham potential in
a localized basis set. These calculations were done including ﬁrst only bound states, yielding
results identical to those presented here, and then including all positive energy orbitals allowed
by their basis set. They found signiﬁcant improvement in He singletsinglet excitations,
especially for 1s → 2s and 1s → 3s. Other excitations barely changed. Assuming inclusion
of the continuum aﬀects results with other approximate kernels similarly, these results do not
change the basic reasoning and conclusions presented below, but suggest that calculations
including the continuum may prove to be more accurate than those presented here.
24.4 Atoms
24.5 Molecules
24.6 Strong ﬁelds
24.6. STRONG FIELDS 185
Table 24.1: Singlet/triplet excitation energies in the helium and beryllium atoms, calculated from the exact KohnSham
potential by using approximate xc kernels (in millihartrees), and using the lowest 34 unoccupied orbitals of s and p symmetry
for He, and the lowest 38 unoccupied orbitals of s, p, and d symmetry for Be. Exact values from Ref. [?] for He and from
Ref. [?] for Be.
Singlet/triplet shifts
ω
KS
ALDA X SIC hybrid exact
Transitions from the 1s state in He atom
2s 746.0 22/11 20/25 19/16 14/19 12/18
3s 839.2 6.9/2.4 5.8/4.9 5.6/3.6 5.1/4.6 3.3/4.2
4s 868.8 3.1/0.9 2.5/1.7 2.4/1.3 2.4/2.0 1.3/1.6
5s 881.9 1.6/0.4 1.3/0.8 1.3/0.6 1.3/1.0 0.6/0.8
6s 888.8 1.0/0.3 0.8/0.5 0.7/0.4 0.8/0.7 0.4/0.5
2p 777.2 0.8/7.4 7.2/8.4 6.1/ 0.2 2.7/3.4 2.7/6.6
3p 847.6 0.7/1.9 2.5/2.3 2.2/0.5 1.4/1.3 1.0/2.0
4p 872.2 0.4/0.7 1.1/0.9 1.0/0.2 0.6/0.5 0.5/0.8
5p 883.6 0.2/0.4 0.6/0.5 0.5/0.1 0.4/0.3 0.2/0.4
6p 889.8 0.1/0.3 0.3/0.3 0.3/0.1 0.2/0.2 0.1/0.3
err 57 32 31 29 12 
Transitions from the 2s state in Be atom
3s 244.4 7.1/5.7 10.9/10.6 10.3/1.4 6.6/4.6 4.7/7.1
4s 295.9 2.5/1.6 3.6/2.5 3.5/0.6 2.6/1.6 1.4/2.0
5s 315.3 1.1/0.7 1.7/1.0 1.6/0.3 1.2/0.7 0.6/0.9
6s 324.7 0.6/0.4 0.9/0.5 0.9/0.2 0.7/0.4 0.3/0.5
2p 132.7 56/42 55/133 53/53 10/88 61/32
3p 269.4 2.0/4.3 6.4/4.2 5.6/ 1.1 4.2/2.8 4.8/1.5
4p 304.6 0.3/1.4 2.1/1.2 1.9/ 0.3 1.3/0.8 1.7/4.1
5p 319.3 0.1/0.6 1.0/0.5 0.9/ 0.1 0.6/0.2 0.2/ 0.0
6p 326.9 0.0/0.4 0.5/0.3 0.4/ 0.0 0.3/0.2 0.1/0.1
3d 283.3 5.4/2.8 1.8/2.0 0.9/ 3.2 1.4/ 1.2 10.3/0.6
4d 309.8 1.4/1.1 0.8/0.9 0.5/ 0.9 0.2/ 0.6 3.6/0.2
err 138 56 144 73 136 
err’ 45 41 37 44 29 
186 CHAPTER 24. PERFORMANCE
Chapter 25
Exotica
25.1 Currents
25.2 Initialstate dependence
Not much attention is paid in the literature to the initialstate dependence in the RungeGross
theorem, and its often not mentioned explicitly. But TDDFT would be useless if we always
had to account for it, since we’d have a diﬀerent functional for every initial wavefunction.
In practice, this is not really a problem. Often, one restricts oneself to starting in a
nondegenerate ground state. In that case, by virtue of the groundstate HohenbergKohn
theorem, the initial wavefunction is an (implicit) density functional, and so the entire potential
is a density functional.
In fact, one can be even more general, and include any initial wavefunction that can be
generated by some timeevolution of the system from a nondegenerate groundstate. Suppose
one can ﬁnd such a prehistory, using an external potential v(x), with t running from −t
P
to
0. In that case, simply tack on this prehistory, and apply TDDFT with the initial ground
state functional, begining from −t
P
. The original initial potential (at t = 0) diﬀers from
what it would be if we began in the groundstate, because of the memorydependence of the
functional.
Can all initialstates be generated by a pseduoprehistory begining in a groundstate? This
seems unlikely, since one only has the freedom to vary v
ext
(rt), but the wavefunction is a
function of many variables. But presumably most of physical interest are of this kind.
We will not discuss further the initialstate dependence, and assume from here on that the
initial state is a nondegenerate ground state.
25.3 Lights, camera, and...Action
Before discussing the action in timedependent DFT, I ﬁrst point out that, for TDDFT, there
is no analog of the groundstate energy. That is, there is no one functional in which we have
an overarching interest. In a timedependent problem, we wish to know the timedependent
expectation value of the density, which is determined by the KohnSham potential. But in
187
188 CHAPTER 25. EXOTICA
the groundstate problem, what we really care about is the energy, which will determine the
geometry of our molecules and solids.
Nevertheless, the action is the mathematically analogous quantity, deﬁned as
A[Ψ] =
tf
0
dt Ψ(t)
ˆ
H −i
∂
∂t
Ψ(t) (25.1)
The principle of least action says that
∂A
∂Ψ
= 0 (25.2)
is satisﬁed by the wavefunction obeying the timedependent Schrodinger equation. It would
be very nice to deﬁne an exact A[n] whose variations with respect to the density yielded the
timedependent Schroedinger equation.
The original RungeGross paper deﬁned this action simply as
A[n] = A[Ψ[n]] (25.3)
However, this has since been shown to lead to several inconsistencies. Essentially, variations
in the density restrict variations in the wavefunction to only those that are vrepresentable,
i.e., form
25.4 Solids
25.5 Back to the ground state
The last class of application of TDDFT is, perhaps surprisingly, to the groundstate problem.
This is because one can extract the groundstate exchangecorrelation energy from a response
function, in the same fashion as perturbation theory yields expressions for groundstate con
tributions in terms of sums over excited states. To do this, we simply note that the equaltime
susceptibility, χ(rr
, t − t
= 0), yields the densitydensity correlation function, i.e., the pair
density of chapter X. Thus
n
λ
XC
(r, r
) = −χ
λ
(rr
, t →t
)/n(r) −δ
(3)
(r −r
), (25.4)
where we have now generalized Eq. (23.4) to arbitrary λ by using a kernel λ/r
1
− r
2
 +
f
λ
XC
(r
1
r
2
ω). Inserting this expression for the hole into the adiabatic connection formula for
the energy, we ﬁnd:
E
XC
=
d
3
r
d
3
r
∞
0
dω
π
.... (25.5)
giving the exchangecorrelation energy directly in terms of χ
λ
. Any approximation for f
XC
,
whose λdependence can be simply extracted via scaling, yields an approximation to E
XC
.
Most importantly, Eq. (25.5) produces a natural method for incorporating timedependent
ﬂuctuations in the exchangecorrelation energy. In particular, as a system is pulled apart
25.6. MULTIPLE EXCITATIONS 189
into fragments, Eq. (25.5) includes correlated ﬂuctuations on the two separated pieces.
While in principle all this is included in the exact functional, in practice TDDFT provides
a natural methodology for modelling these ﬂuctuations. Simple ALDA approximations for
polarizabilities of atoms and molecules allow C
6
coeﬃcients to be accurately calculated by
this method, and attempts exist to generate entire molecular enery curves, from the covalent
bond distance to dissociation, including van der Waals. Such calculations are much more
demanding than simple selfconsistent groundstate calculations, but may well be worth the
payoﬀ.
25.6 Multiple excitations
25.7 Exact conditions
190 CHAPTER 25. EXOTICA
Appendix A
Math background
A.1 Lagrange multipliers
To illustrate how Lagrange multipliers work, we do a simple example on a function of two
variables, instead of functionals.
Suppose we wish to maximize the function f(x, y) = x + y subject to the constraint
x
2
+ y
2
= 1. The inelegant approach is to simply solve the constraint equation for y as a
function of x, ﬁnding y =
√
1 −x
2
, and substitute into f, yielding
˜
f(x) = x +
√
1 −x
2
(A.1)
Then we ﬁnd the extrema of f by diﬀerentiation:
d
˜
f
dx
= 1 −
x
√
1 −x
2
(A.2)
which we set to zero, ﬁnding x = 1/
√
2. Then y = 1/
√
2 also, yielding the maximum
f =
√
2.
But a much more elegant way is to introduce a Lagrange multiplier. Write the constraint
as g(x, y) = 0, where g = x
2
+y
2
−1. Then we deﬁne
h = f −µg (A.3)
where µ is a constant, independent of x and y. We then minimize h freely, with no restriction
between x and y. Write
∂h
∂x
= 1 −2µx
∂h
∂y
= 1 −2µy (A.4)
and set both to zero, yielding (x, y) = (1, 1)/(2µ). Then we enforce the condition g = 0,
yielding µ = ±1/
√
2, and choose the positive sign for a maximum. This method avoids
diﬀerentiating square roots, etc. This is especially important if we cannot solve the constraint
equation analytically.
191
192 APPENDIX A. MATH BACKGROUND
A.2 Properties of the δfunction
We review brieﬂy the useful properties of δfunction. The δfunction is an inﬁnitely narrow,
inﬁnitely tall peak at x = 0, chosen so that its area is exactly 1. Thus we can write several
limiting formulas:
δ(x) = lim
a→0
(Θ(x +a/2) −Θ(x −a/2)) /a,
= lim
γ→0
1
γ
√
π
exp(−x
2
/γ
2
)
= lim
γ→0
1
π
γ
x
2
+ γ
2
(A.5)
where Θ(x) is the step function (= 0 for x < 0, 1 for x > 0), and the δfunction looks like
(with a ﬁnite width) what’s shown in Fig X.
The usefulness of the δ function is that, for any reasonably smooth function:
dx f(x)δ(x) = f(0) (A.6)
assuming the integration interval includes the origin. This can be used to extract the value
of the function at any point x
0
:
dx f(x)δ(x −x
0
) = f(x
0
) (A.7)
again assuming the integration interval includes the point x
0
.
Another handy formula is the integral over the δfunction:
x
−∞
dx
δ(x
) = Θ(x) (A.8)
which is just the step function, i.e., you get zero unless the interval includes the δfunction,
when you get 1. Another way to see this is that the ﬁrst formula of Eq. (A.5) implies that
the δ function is the derivative of the step function.
The Fourier transform of the δfunction is
δ(x) =
∞
−∞
dp
2π
exp(ipx) (A.9)
i.e., it’s just a constant.
Finally, one can show:
δ(f(x)) =
¸
i,f(xi)=0
δ(x −x
i
)
f
(x
i
)
(A.10)
where f
(x) = df/dx.
Lastly, with some care,
δ
(x)f(x) = −f
(0) (A.11)
assuming the interval includes 0.
In fact, its not strictly a function at all, but properly called a distribution.
A.3. FOURIER TRANSFORMS 193
A.3 Fourier transforms
Include KramersKronig relation.
194 APPENDIX A. MATH BACKGROUND
Appendix B
Results for simple oneelectron systems
B.1 1d H atom
The Hamiltonian for this problem is
H = −
1
2
d
2
dx
2
−Zδ(x) (B.1)
Integrating the Schr¨ odinger equation through x = 0 yields the cusp condition:
φ

0+
0−
= −Zφ(0) (B.2)
The groundstate energy is just
0
= −Z
2
/2, and T = −
0
, V = 2
0
. The normalized orbital
is:
φ(x) =
√
Z exp(−Zx) (B.3)
and groundstate density
n(x) = Z exp(−2Zx) (B.4)
We can also extract continuum excited states, which are useful for TDDFT. With the usual
scattering boundary conditions:
φ
s
(x) = exp(ikx) +r exp(−ikx), x < 0
= t exp(ikx), x > 0 (B.5)
where r and t are given by:
r =
iZ
k −iZ
, t =
ik
Z +ik
(B.6)
and k =
√
2, E > 0. Furthermore, when
ˆ
H is given by (B.1), the solution to:
(
+
−
ˆ
H)g
>
(x, x
; ) = δ(x −x
) (B.7)
is the outgoing Green’s function:
g
>
(x, x
; ) =
1
i
√
2
¸
¸e
i
√
2x−x

−
Ze
i
√
2(x+x
)
i
√
2 +Z
¸
. (B.8)
195
196 APPENDIX B. RESULTS FOR SIMPLE ONEELECTRON SYSTEMS
and in terms of g
>
, the susceptibility is
χ(x, x
; ) =
n(x)n(x
) [g
>
(x, x
; +
0
) +g
>∗
(x, x
, − +
0
)] (B.9)
From this we can calculate both the static and dynamic polarizabilites:
α(0) = −
5
4Z
4
(B.10)
and
Im(α(ω)) =
−16
(˜ ω −1)
Z
4
˜ ω
4
(B.11)
where ˜ ω = ω/
0

B.2. HARMONIC OSCILLATOR 197
B.2 Harmonic oscillator
For the Hamiltonian
ˆ
H = −
1
2
d
2
dx
2
+
1
2
ω
2
x
2
(B.12)
the energies are
n
=
n +
1
2
ω, n = 0, 1, 2, . . . , (B.13)
The lengthscale is given by x
0
=
1
ω
, and the wavefunctions are
φ
n
(x) =
1
x
0
√
π2
n
n!
¸
¸
1
2
e
−
1
2
x
x
0
2
H
n
x
x
0
, (B.14)
where H
n
(y) are the Hermite polynomials. The ground state and ﬁrst 2 excited states wave
functions are
φ
0
(x) =
1
x
0
√
π
¸
¸
1
2
e
−
1
2
x
x
0
2
, (B.15)
φ
1
(x) =
2
x
0
√
π
¸
¸
1
2
x
x
0
e
−
1
2
x
x
0
2
, (B.16)
φ
2
(x) =
1
2x
0
√
π
¸
¸
1
2
(2
x
x
0
2
−1)e
−
1
2
x
x
0
2
. (B.17)
198 APPENDIX B. RESULTS FOR SIMPLE ONEELECTRON SYSTEMS
B.3 H atom
Appendix C
Green’s functions
199
200 APPENDIX C. GREEN’S FUNCTIONS
Appendix D
Further reading
Chapter 1
There are now a variety of sources to choose from to learn about density functional theory
in general.
• Perhaps the best overview comes from a recent summer school, A Primer in Density
Functional Theory, ed. C. Fiolhais, F. Nogueira, and M. Marques (SpringerVerlag, NY,
2003).
• A good generalpurpose book for solidstate physicists about electronic structure is
Richard Martin’s book. Another is Vignale’s.
• For chemists, a very useful guide is A chemist’s guide to density functional theory, W.
Koch and M.C. Holthausen, (WileyVCH, Weinheim, 2000). Part A deals with theory,
while part B has a very useful survey of properties and how well approximations do for
them.
• A standard in physics for many years has been: R.M. Dreizler and E.K.U. Gross, Density
Functional Theory (SpringerVerlag, Berlin, 1990). This is very careful and rigorous,
but uses some higherlevel physics concepts.
• In chemistry, we have Density Functional Theory of Atoms and Molecules, R.G. Parr
and W. Yang (Oxford, New York, 1989), which includes many concepts particularly useful
for chemistry, such as reactivitity theory.
• A nice discussion of using KohnSham orbitals to understand chemistry appears in A
quantum chemical view of density functional theory, E.J. Baerends and O.V. Gritsenko,
J. Phys. Chem. A 101, 5383 (1997).
Chapter 3
There are many books on basic quantum mechanics.
• The one I enjoy the most is Griﬃths..
201
202 APPENDIX D. FURTHER READING
• A great old one that demonstrates the unity of quantum is Morisson...
Appendix E
Discussion of questions
Chapter 1
1. Why is a KohnSham calculation much faster than a traditional wavefunction calcu
lation?
Because a KohnSham calculation requires solving only a oneelectron problem (self
consistently) and occupying N levels, rather than solving a coupled diﬀerential equation
for a function of 3N variables.
2. If you evaluate the kinetic energy of a KohnSham system, is it equal to the physical
kinetic energy?
No. The kinetic energy operator is the same for the two systems, but their wavefunctions
are diﬀerent.
3. Repeat above question for 1/r, which can be measured in scattering experiments.
Yes. 1/r =
d
3
r n(r)/r, and so depends only on the oneelectron density, which is
deﬁned to be the same in both systems.
4. Why is the density cubed in the local approximation for T
S
? (see section ?? for the
answer).
Performing a dimensional analysis, we note that n ∼ 1/L in one dimension, and since
T ∼ 1/L
2
=
dxn
p
(x), p must be 3.
5. Suppose you’d been told that T
loc
S
is proportional to
dxn
3
(x), but were not told the
constant of proportionality. If you can do any calculation you like, what procedure
might you use to determine the constant in T
loc
S
?
A physicist might answer that, since T
S
becomes more and more accurate as the number
of electrons in the box grows, take the limit as N →∞. We saw in the chapter how that
yields the π
2
/6 constant. On the other hand, a practical chemist might answer that, if
all I want to do is solve problems that look like one electron in a box, I should simply ﬁt
the constant to the value for that problem.
203
204 APPENDIX E. DISCUSSION OF QUESTIONS
Both answers have value, but reﬂect diﬀerent priorities. The chemist’s answer is likely
to be more accurate than the physicist’s for problems that are similar to the ﬁtted one.
Errors are likely to be both positive or negative. On the other hand, the physicist’s answer
is likely to be more accurate over a broad range of problems, with systematic errors (but
larger than the chemist’s on the ﬁtted systems).
6. In what way will T
loc
S
change if spin is included, e.g., for two electrons of opposite
spin in a box?
A detailed answer is given in section 9.2. Here we just doubly occupy each level in the
box. Thus...(Rudy, ﬁll in).
7. If we add an inﬁnitesimal to n(x) at a point, i.e., δ(x) as → 0, how does T
loc
S
change?
Again, a detailed answer is given in section 2.2. But we can ﬁgure it out easily in this
case. Rudy...
8. The simplest density functional approximation is a local one. What form might a cor
rection to T
loc
S
take?
A detailed answer is given in section 17.2. But we can make intelligent suggestions here.
The most obvious is that, since using the density at a point works so robustly, surely
including information on the gradient of the density can help accuracy. Rudy...
9. In the chapter and exercises, T
loc
S
does very well. But there are cases where it fails
quite badly. Can you ﬁnd one, and say why?
Consider a particle in a large box, at whose center is a tall but ﬁnite barrier. Ignoring the
exponentially small tunneling contribution, the lowest molecular orbital is
φ(x) = (φ
A
(x) + φ
B
(x))/
√
2 (E.1)
where φ
A
(x) is the groundstate wavefunction for the electron to be in the left side,
and φ
B
(x) is that for the right. The exact kinetic energy is then
1
2
(T
A
+ T
B
) = T
A
by
symmetry. On the other hand, our local functional uses n(x) = (n
A
(x) + n
B
(x))/2, so
that T
loc
S
= (T
loc
A
+T
loc
B
)/8 = T
loc
A
/4, i.e., it is about 4 times too small!
10. Does T
loc
S
[n
∗
] work for a single electron in an excited state, with density n
∗
(r)?
For ANY single excited state of the particle in the box, T
loc
S
[n
∗
] = T
loc
S
[n], so it fails very
badly in this form. Need to show explicit failure for particle in box. Rudy...
However, Rudy points out that, since our electrons are not interacting, he can ﬁnd the
energy of the ﬁrst excited state by writing n
∗
(x) = n
2
(x) − n
1
(x), where n
i
(x) is the
density with i particle in the box. Then T[n
∗
] = T
loc
2
−T
loc
1
= 17.7/L
2
205
Chapter 2
1. Compare the functional derivative of T
VW
S
[n] with T
loc
S
[n] for some sample one
electron problem. Comment.
You might take either the particle in the box or the harmonic oscillator or the one
dimensional hydrogen atom. You should ﬁnd that a plot of the two functional derivatives
gives you two very diﬀerent curves in each case. To understand this, think of functions.
Imagine one function being a horizontal line, and a second function having weak but
rapid oscillations around that line. The second function is a good approximation to ﬁrst
everywhere, but its derivative is very diﬀerent.
2. Devise a method for deducing if a functional is local or not.
Take the second functional derivative. If its proportional to a delta function, the function
is local. Note that just saying if the integrand depends only on the argument at r is not
enough, since integrands can change within functionals, e.g., by integrating by parts.
3. Is there a simple relationship between T
S
and
dx n(x)δT
S
/δn(x)?
You should have found that the integral is three times the functional for the local ap
proximation, but equal to the functional for von Weisacker.
4. For ﬁxed particle number, is there any indeterminancy in the functional derivative
of a density functional?
If the particle number does not change, then
d
3
r δn(r) = 0. Thus addition of any con
stant to a functional derivative does not alter the result, so that the functional derivative
is not determined up to a constant.
206 APPENDIX E. DISCUSSION OF QUESTIONS
Chapter 3
1. Suggest a good trial wavefunction for a potential that consists of a negative delta
function in the middle of a box of width L.
The simplest guess is just that of the particle in the box,
2/Lcos(πx/L). Its energy is
obviously π
2
/2L
2
− 2Z/L, where Z is the strength of the delta function. Much better
(in fact, exact) is to take linear combinations of the deltafunction solutions, exp(−αx)
and exp(αx) vanishing at the edges.
2. What is the eﬀect of having nuclear charge Z = 1 for the 1d Hatom?
It alters the length scale of the wavefunction without changing its normalization, i.e.,
√
Z exp(−Zx). We see shortly how this can be a handy trick.
3. What is the exact kinetic energy density functional for one electron in onedimension?
It is the von Weisacker functional,
T
VW
S
=
1
2
∞
−∞
dx φ
(x)
2
=
1
8
∞
−∞
dx n
(x)
2
/n(x). (E.2)
4. Is the HF estimate of the ionization potential for 1d He an overestimate or under
estimate?
The ionization potential is I = E
1
− E
2
for a twoelectron system, where E
N
is the
energy with N electrons. Since E
HF
≥ E, with equality for one electron, I
HF
< I.
5. Consider the approximate HF calculation given in section 4.2. Comment on what it
does right, and what it does wrong. Suggest a simple improvement.
The HF calculation correctly writes the wavefunction as a spatially symmetric singlet, but
incorrectly approximates that as a product of separate functions of x
1
and x
2
(orbitals).
It correctly has satisﬁes the cusp condition at the nucleus, and changes decay at large
distance, unlike the scaled orbital solution. But its energy is not low enough.
6. Which is bigger, the kinetic energy of the true wavefunction or that of the HF wave
function for 1d He? Hooke’s atom?
The virial theorem applies to this problem (see section X), and, since all potentials are
homogeneous of order−1, E = −T = −V/2, where V includes both the external and
the electronelectron interaction. Since E
HF
> E, then T
HF
< T. For Hooke’s atom,
the situation is more complicated, since there are two diﬀerent powers in the potentials.
Chapters 46
1. The groundstate wavefunction is that normalized, antisymmetric wavefunction that has
density n(r) and minimizes the kinetic plus Coulomb repulsion operators.
207
2. The KohnSham wavefunction of density n(r) is that normalized, antisymmetric wave
function that has density n(r) and minimizes the kinetic operator.
3. The KohnSham kinetic energy is not
1
2
dx φ
1

2
+ φ
2

2
, because these orbitals could
have come from anywhere. There is no reason to think they are KohnSham orbitals.
They might be, e.g., HartreeFock orbitals. The way to get T
S
is to construct the density
n(x) and then ﬁnd v
S
(x), a local potential that, with two occupied orbitals, yields that
density. The kinetic energy of those two orbitals is then T
S
. Again, if we alter one
of the original orbitals, the change in T
S
is not the kinetic contribution directly due to
that orbital. Rather, one must calculate the new density, construct the new KohnSham
potential, get the kinetic energy of its orbitals, and that tells you the change.
All the same reasoning applies for E
X
[n].
4. See above. One can do a HF calculation, yielding orbitals and density. Then E
HF
X
is
simply the Fock integral of those orbitals. But this is not E
X
[n
HF
], which could only be
found by ﬁnding the local potential v
S
(r) whose orbitals add up to n
HF
(r), and evaluating
the Fock integral on its orbitals.
5. The additional ﬂexibility of spinDFT over DFT means that its much easier to make good
approximations for spinpolarized systems.
6. Given v
Sσ
(r), solve the KohnSham equations for up and down noninteracting electrons
in these potentials, and construct the total density. Then ask what single potential all
electrons must feel in the KohnSham equations to reproduce that density. Obviously,
they coincide when the system is unpolarized, so that v
Sσ
(r) = v
S
(r), and when fully
polarized, v
S↑
(r) = v
S
(r), v
S↓
(r) = 0 or undetermined.
7. No. See further work whead.
208 APPENDIX E. DISCUSSION OF QUESTIONS
4
CONTENTS
5.3 5.4 5.5 5.6
Correlation . . . . . . Atomic conﬁgurations Atomic densities . . . Questions . . . . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
49 51 51 53
Contents
II I Background 13 Basics 55
1 Introduction 1.1 Importance . . . . . . . . . . . . . 1.2 What is a KohnSham calculation? 1.3 Reinterpreting molecular orbitals . . 1.4 Sampler of applications . . . . . . 1.5 Particle in a box . . . . . . . . . . 1.6 Questions . . . . . . . . . . . . . . 2 Functionals 2.1 What is a functional? . . 2.2 Functional derivatives . . 2.3 EulerLagrange equations 2.4 Questions . . . . . . . . . 3 One 3.1 3.2 3.3 3.4 3.5 3.6 electron Variational principle . Trial wavefunctions . . Three dimensions . . Diﬀerential equations Virial theorem . . . . Questions . . . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
15 15 16 19 20 20 23 27 27 28 30 31 33 33 34 35 36 36 37 39 39 39 44 45 47 47 48
6 Density functional theory 6.1 One electron . . . . . . . 6.2 HohenbergKohn theorems 6.3 ThomasFermi . . . . . . 6.4 Particles in boxes . . . . . 6.5 Questions . . . . . . . . . 7 KohnSham 7.1 KohnSham equations 7.2 Exchange . . . . . . . 7.3 Correlation . . . . . . 7.4 Questions . . . . . . . 8 The 8.1 8.2 8.3 8.4 . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . .
57 57 58 60 62 63 65 65 67 68 69 71 71 72 73 76 77 77 78 79 80 81 81 82 83 85 85 86 86
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
local density approximation Local approximations . . . . . Local density approximation . Uniform electron gas . . . . . Questions . . . . . . . . . . . . . . . . . . . . . . . . . . .
9 Spin 9.1 KohnSham equations 9.2 Spin scaling . . . . . 9.3 LSD . . . . . . . . . 9.4 Questions . . . . . . .
4 Two electrons 4.1 Antisymmetry and spin 4.2 HartreeFock . . . . . . 4.3 Correlation . . . . . . . 4.4 Questions . . . . . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
5 Many electrons 5.1 Ground state . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 HartreeFock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
10 Properties 10.1 Total energies . . . . . . . . . . . . . . 10.2 Densities and potentials . . . . . . . . . 10.3 Ionization energies and electron aﬃnities 10.4 Dissociation energies . . . . . . . . . . . 10.5 Geometries and vibrations . . . . . . . . 10.6 Transition metals . . . . . . . . . . . . . 10.7 Weak bonds . . . . . . . . . . . . . . .
CONTENTS
5
6
CONTENTS
10.8 Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.9 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
86 86
15.4 Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 16 Exchangecorrelation hole 16.1 Density matrices and holes . 16.2 Hooke’s atom . . . . . . . 16.3 Transferability of holes . . . 16.4 Old faithful . . . . . . . . . 121 121 125 126 131
III
Analysis
87
11 Simple exact conditions 11.1 Size consistency . . . 11.2 One and two electrons 11.3 LiebOxford bound . . 11.4 Bond breaking . . . . 11.5 Uniform limit . . . . . 11.6 Questions . . . . . . .
. . . . . .
. . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . .
89 89 89 90 90 91 91
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
IV
Beyond LDA
135
12 Scaling 12.1 Wavefunctions . . . . . . 12.2 Density functionals . . . . 12.3 Correlation . . . . . . . . 12.4 Correlation inequalities . . 12.5 Virial theorem . . . . . . 12.6 Kinetic correlation energy 12.7 Potential . . . . . . . . . 12.8 Questions . . . . . . . . . 13 Adiabatic connection 13.1 One electron . . . . 13.2 Adiabatic connection 13.3 Relation to scaling . 13.4 Static correlation . . 13.5 Questions . . . . . .
93 . 93 . 95 . 96 . 97 . 99 . 100 . 101 . 102 . . . . . . . . . 103 103 105 107 109 111 113 113 113 114 116 117 117 117 119
17 Gradients 17.1 Perimeter problem 17.2 Gradient expansion 17.3 Gradient analysis . 17.4 Questions . . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
137 137 140 142 146 147 147 151 153 154 155 155
18 Generalized gradient approximation 18.1 Fixing holes . . . . . . . . . . . . . . . . . . . . . 18.2 Visualizing and understanding gradient corrections . 18.3 Eﬀects of gradient corrections . . . . . . . . . . . . 18.4 Satisfaction of exact conditions . . . . . . . . . . . 18.5 A brief history of GGA’s . . . . . . . . . . . . . . . 18.6 Questions about generalized gradient approximations
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . formula . . . . . . . . . . . . . . . . . . . . . . .
19 Hybrids 157 19.1 Static, strong, and strict correlation . . . . . . . . . . . . . . . . . . . . . . 157 19.2 Mixing exact exchange with GGA . . . . . . . . . . . . . . . . . . . . . . . 157 19.3 Questions about adiabatic connection formula and hybrids . . . . . . . . . . 157 20 Orbital functionals 20.1 Selfinteraction corrections . . . . 20.2 Optimized potential method . . . 20.3 G¨rlingLevy perturbation theory o 20.4 Meta GGA’s . . . . . . . . . . . 20.5 Jacob’s ladder . . . . . . . . . . 159 159 159 159 159 159
14 Discontinuities 14.1 Koopman’s theorem . . . 14.2 Potentials . . . . . . . . 14.3 Derivative discontinuities . 14.4 Questions . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
15 Analysis tools 15.1 Enhancement factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.2 Density analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15.3 Energy density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
V
Timedependent DFT
161
21 Timedependence
163
. 181 24 Performance 24. . . . . . . . . . . . . . . . . . . . 24. . . . . . . .3 Singlepole approximation .. . . . . . . . . F. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25. . . . Questions about timedependent quantum mechanics . . . . . . . . . . . . . . . . . . . . . 24.3 21. .12 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Adiabatic approximation . . . . . . . . . . . . .2 Functionals . . . . . . . . . . . . 199 201 203 209 209 214 220 230 239 242 249 255 257 258 258 261 285 22 Timedependent density functional theory 22. . . . . . . . . . . 25. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . F. . . . . . . . . . . . . . . . . .CONTENTS 7 8 CONTENTS 21. . . .1 Lagrange multipliers . . . . . . . . . . . . . . . . . . . . . . . 183 183 183 183 184 184 184 187 187 187 187 188 188 189 189 . .1 Overview . . . . . F. . . . . . . . . . . . .2 21. F.5 Questions on general principles of TDDFT . . . . . . . . . . . . . . . . . . .Action 25. . . . . .3 One electron . . . .10 Properties . . . .2 Properties of the δfunction . . . . . . . . . . . . . . . F. . . 22. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Dysonlike response equation and the kernel . . . . .. . . . . . . . . . . . . . . . . . . .11 Simple exact conditions . . . . . . . . . . . .3 H atom . . . . . . . . . F.1 1d H atom . . . . . . . . . . . . . . . .6 Multiple excitations . . . . . .5 Many electrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Fourier transforms . . . . . . 22. .3 KohnSham equations . .2 Initialstate dependence . . . . . . . . . . . . . . . . . . . .8 Local density approximation F. . . . . . F. . .5 Back to the ground state . . . . . . . . . . F. .4 Solids . . . . . . . . . . . . . . . . . . . . . . . .9 Spin . . . . . . . . . . . . 193 . . . .3 Transition frequencies 24. . . . . . . . . F. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 B. . . . . . . . . . . 180 23. . . . F. 24. . . . . . . . . . . . . . . . . . . . . . . . .2 Casida’s equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Optical response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 21. . . . . . . . . . . . . . . . . . . . . . . 163 166 169 170 171 171 173 174 175 176 B Results for simple oneelectron systems 195 B. . .2 RungeGross theorem . . . . . . . . . . . . . 23 Linear response 179 23. . . . . . . . 25 Exotica 25. . . .4 Atoms . . . . . . . . . . . . . 24. 25. . . . . . 192 A. . . . . . . . . . . . . . . . . . . . . . . . and. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Lights. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . G Answers to extra problems . . . .7 Exact conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7 KohnSham . . . . . . . . . . . . . . . .2 Poor potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . o Perturbation theory .4 Schr¨dinger equation . . . .2 Harmonic oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 B. . . . . . . . . . . . . . . .5 Molecules .6 Density functional theory . . . . . . . . . . . . . . . .1 Introduction . . . . . . . . . . . . . . . . . . . . . . . A Math background 191 A. .6 Strong ﬁelds . . . . . . . . . . . . . . . . . . . . . . 22. . . camera. . . . . . . 179 23. . . . . . . . . . . . . . 25. .1 Currents .4 Two electrons . . . . . . . . . . . . . . . .1 Sources of error . . 191 A. . . . . 198 C Green’s functions D Further reading E Discussion of questions F Solutions to exercises F. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. but Ψ[nγ ] = Ψγ [n] Fundamental inequality: F [nγ ] ≤ γ 2 T [n] + γVee [n] Noninteracting kinetic energy: TS [nγ ] = γ 2 TS [n] Exchange and Hartree energies: EX [nγ ] = γEX [n]. Vee [nγ ] > γVee [n] (γ > 1) Correlation energies: EC [nγ ] > γEC [n]. x ) = δσσ Nσ ˆ External potential: Vext = vext (ri ) ˆ ˆ Noninteracting kinetic energy: Ts [n] = min ΦT Φ = Φ[n]T Φ[n] = Φ→n Wavefunctions ˆ ˆ Physical wavefunction: Ψ[n](x1 . x ) ≥ 0 potential energy:Vee = 1 d3 r d3 r P (r. . r = r. . x ) kinetic energy: T = − 1 d3 r 2 γ(r. x ) = N (N − 1) dx3 . n↓ ] = 1 (TS [2n↑ ] + TS [2n↓ ]) 2 Exchange: EX [n↑ . σ). . x) = γ(x.φN (xpN ) where φi (x) and i are the ith KS orbital and energy. . x )2 /n(x) properties: nX (r. r )/r − r  2 KohnSham: PX (x. x3 . λ relation to scaling: Ψ [n] = Ψλ [n1/λ ] Energies: E λ [n] = λ2 E[n1/λ ] kinetic: TSλ [n] = TS [n] λ exchange: EX [n] = λEX [n]. x2 . .xN ) has density n and minizes T Φ(x1 . σ.. r + u) = −1. 2 i=j N i=1 2 i i Density matrix: γ(x. x ) = n(x)n(x ) − γS (x. P (x. xN ) properties: γ(x. . dxN Ψ(x.. . x ) = −γS (x. x ) = (N − 1)n(x). Kinetic energy: T ˆ ˆ ˆ Potential energy: V = Vee + Vext ˆee = 1 Coulomb repulsion: V 1/ri − rj . x . r + u) = 0 C Pair correlation function: g λ (x. for small λ ACF: EXC = 1 0 dλ λ UXC (λ). Spacespin vector: x = (r.. . TC [nγ ] < γ 2 TC [n] (γ > 1) ˆ Virial theorem: 2T = r · V N electrons: 2T + Vee = d3 r n(r) r · vext (r) XC: EXC + TC = d3 r n(r) r · vXC (r) Spin scaling Kinetic: TS [n↑ . . 1 3 3 Hartree energy: U [n] = 2 d r d r n(r) n(r )/r − r  Exchange: EX = ΦVee Φ − U = − 1 d3 r d3 r φ∗ (r) φ∗ (r ) φiσ (r ) φjσ (r)/r − r  iσ jσ 2 σ i. z). dxN Ψ(x. . x2 . . U λ [n] = λU [n] (2) (3) λ correlation: EC [n] = λ2 EC [n1/λ ] = λ2 EC [n] + λES [n] + .. γrN ) Ground states: Φ[nγ ] = Φγ [n]. .CONTENTS 9 10 CONTENTS Deﬁnitions and notation Coordinates Position vector: r = (x. x ) = P λ (x. . . y. x ). Energies ˆ ˆ ˆ Universal functional: F [n] = min ΨT + Vee Ψ = Ψ[n]T Ψ[n] Ψ→n ˆΨ[n] ... r + u) ≤ 0. . r + u) = n(r) [n(r + u) + nλ (r. d3 r Sums: dx = σ Operators ˆ = − 1 N 2. x) = P (x. . n↓ ] = 1 (EX [2n↑ ] + EX [2n↓ ]) 2 Adiabatic connection 1 HellmannFeynman: E = E λ=0 + 0 dλ Ψλ dH λ /dλΨλ ˆ ˆ Wavefunction: Ψλ [n] has density n and mininimizes T + λVee . r + u)] XC KohnSham: nX (x.xN ) = p (−1)p φ1 (xp1 ).xN ) = p (−1)p φ1 (xp1 ). dxN Ψ∗ (x. . γ(x . with i = α. u)/duu=0 = λg λ (r.φN (xpN ) Φ(x1 .. U [nγ ] = γU [n] Kinetic and potential: T [nγ ] < γ 2 T [n]. x2 . . xN )2 properties: dx P (x. u = 0) Uniform coordinate scaling Density: n(r) → nγ (r) = γ 3 n(γr) Wavefunction: Ψγ (r1 . x )2 Exchangecorrelation hole around r at coupling constant λ: P λ (r. . r )r=r 2 KohnSham: γS (x.. xN )2 = properties: n(x) ≥ 0.. Kinetic energy: T [n] = Ψ[n]T N ˆ Coulomb repulsion energy: Vee [n] = Ψ[n]Vee Ψ[n] . . Spin index: σ =↑ or ↓ = α or β. where UXC (λ) = UXC /λ i=1 φ∗ (r)φiσ (r ) iσ Kineticcorrelation: TC [n] = T [n] − TS [n] Potentialcorrelation: UXC [n] = Vee [n] − U [n] ˆ ˆ Exchangecorrelation: EXC [n] = T [n] − TS [n] + Vee [n] − U [n] = Ψ[n]T + Vee Ψ[n] − ˆ ˆ Φ[n]T + Vee Φ[n] Potentials Functional derivative: F [n + δn] − F [n] = d3 r δn(r) δF [n]/δn(r) KohnSham: vS (r) = −δTS /δn(r) + µ Hartree: vH (r) = δU/δn(r) = d3 r n(r )/r − r  Exchangecorrelation: vext (r) = vS (r) + vH (r) + vXC (r) Densities and density matrices Spin density: n(x) = nσ (r) = N dx2 . x ) = N dx2 . .. xN )Ψ(x . x )/(n(x)n(x )) Electronelectron cusp condition: dg λ (r. dx n(x) = N N i=1 φi (x)2 . . .. . P (x .. . . d3 u nX (r. x) = n(x). rN ) = γ 3N/2 Ψ(γr1 .xN ) has density n and minimizes T + Vee ˆ KohnSham: Φ[n](x1 .j occ i=1 d3 r φi (r)2 Pair Density: P (x. d3 u nλ (r.
n↓ ] = d3 r n(r) unif (rs (r).CONTENTS 11 12 CONTENTS Finite systems Kato’s cusp at nucleus: dn/drr=Rα = −2Zα n(Rα ) √ √ Large r in Coulombic system: n(r) → Ar β e− 2Ir . Gradient expansion approximation: AGEA [n] = ALDA [n] + ∆AGEA [n] Reduced density gradient: s(r) =  n(r)/(2kF (r)n(r) Correlation gradient: t(r) =  n(r)/(2ks (r)n(r) Polarization enhancement: φ(ζ) = ((1 + ζ)2/3 + (1 − ζ)2/3 )/2 Kinetic energy: TS [n] = AS d3 r n5/3 1 + 5s2 /27 or ∆TSGEA [n] = T VW [n]/9. Exchange energy: EX [n] = AX d3 r n4/3 1 + 10s2 /81 GEA Highdensity correlation energy: ∆EC = (2/3π 2 ) d3 r n(r)φ(ζ(r))t2 (r) GGA [n] = d3 r a(n. as γ → ∞ 3/2 Lowdensity limit: EC [nγ ] = γB[n] + γ C[n] + . .0311 ln rs − 0.871. Uniform gas and LSD Measures of the local density WignerSeitz radius: rs (r) = (3/(4πn(r))1/3 1 Fermi wave vector: kF (r) = (3π 2 n(r)) 3 ThomasFermi wavevector: ks (r) = 4kF (r)/π Measure of the local spinpolarization: Relative polarization: ζ(r) = (n↑ (r) − n↓ (r))/n(r) 2 Kinetic energy: tunif (n) = (3/10)kF (n)n S exchange energy: eunif (n) = n unif (n). LDA LSD: EX [n] = AX d3 r n4/3 (r) where AX = −(3/4)(3/π)1/3 = −0. VonWeisacker: TSVW [n] = d3 r  n2 /(8n) (exact for N = 1.  n) Generalized gradient approximation: A GGA 3 unif Enhancement factor: EXC = d r eX (n(r)) FXC (rs (r).009rs ln rs − 0. . ζ(r) C Gradient expansions Gradient expansion: A[n] = d3 r a(n(r)) + b(n(r) n(r)2 . . as γ → 0.047 + 0.738. 2 Correlation: EC = 0 for N = 1 (2) (3) Highdensity limit: EC [nγ ] = EC [n] + EC [n]/γ + . 2) Exchange: EX = −U/N for N = 1. LSD EC [n↑ . β = α Zα − N + 1/ 2I Exchange potential: vX (r) → −1/r Correlation potential: vC (r) → −α(N − 1)/2r 4 . X X X (rs → 0) correlation energy: unif (rs ) → 0. where α(N − 1) is the polarizability of the N − 1 electron system.017rs C ThomasFermi: TSTF [n] = AS d3 r n5/3 (r) where As = 2. . s(r)) . . where unif (n) = (3kF (n)/4π). .
Part I Background 13 .
which can then be applied to much larger systems But DFT is not just another way of solving the Schr¨dinger equation. The hope is that these notes will be used by students worldwide to gain a better understanding of this fundamental theory.) are invited to work through the material.e.1) where Z = 1 is the charge on each nucleus. ¯ (1. way of approaching any interacting problem. These are almost as large as those in going from classical to quantum mechanics. there are questions aimed at making you think about the material. is z vext (r) = −Z/r − Z/r − Rˆ. set) in our thinking. In return. but leave the reader with very clear concepts of modern density functional theory. physics. providing benchmarks for developing density functionals.5 kcal/mol) and all c 2000 by Kieron Burke. The book is laid out in the form of an undergraduate text. r2 ). the potential experienced by the electrons due to the nuclei. These notes are copyright of Kieron Burke. what functionals are.edu) if you use this material. with the groundstate electronic structure problem simply being the most common. and a desire to learn. formally rigorous. ˆ is a unit vector along the bond axis.3) where e is the electronic charge. Except where noted. This allowed much larger systems to be treated than by traditional ab initio methods. The idea is that the book and exercises should be easy reading. or many advanced undergraduate courses. all energies are in Hartrees (1 H = 27. This is because approximate functionals were shown to provide a useful balance between accuracy and computational cost. etc. Also. including the KohnSham description of a system. In regular quantum mechanics. INTRODUCTION Chapter 1 Introduction In which we introduce some of the basic concepts of modern density functional theory. but answers don’t need to be written out as explicitly as for the problems. no fear of calculus of more than one variable. 1. and requires the working of many exercises (although many of the answers are given). and the external potential. In the 19990’s it became very popular in quantum chemistry. we make the BornOppenheimer approximation. we use atomic units throughout this text. 15 . and R is z a chosen internuclear separation. either variational or perturbative.2 What is a KohnSham calculation? To give an idea of what DFT is all about.. I am happy to grade problems and answer questions for all who are interested. Throughout the text. Preliminary forms of my lecture notes have been used throughout the world during 5 years its taken to complete this book. H2 . The student should end up knowing what the KohnSham equations are. The technical level is no higher than any graduate quantum course. geophysics.2 2 i 1 + vext (ri ) Ψ(r1 . The aim of this book is to provide a relatively gentle. as these leaps are usually made when we are more advanced (i.2114 eV= 627. + r1 − r2  i=1. The only 1 necessary requirements are a good background in elementary quantum mechanics.2) 1 − 2 i=1.1 1 Importance Density functional theory (DFT) has long been the mainstay of electronic structure calculations in solidstate physics.rutgers. in which we treat the heavy nuclei as ﬁxed points. there are exercises that must be performed to get full value from the book. by mapping it exactly to a much easiertosolve noninteracting problem. 1. introduction to this subject.2 (1. while retaining much of their accuracy. can be applied to ﬁnd highly accurate results on smaller systems. r2 ) = EΨ(r1 . In some sense. Students from all areas of modern computational science (chemistry. As ¯ a consequence. at the end of each chapter. the hydrogen molecule. materials science. so that e2 = h = m = 1. I ask only that you send me an email (to kieron@rutchem. how they can be approximated. These should be thought about and answere as you go along. Density functional theory is a completely diﬀerent. and how insight into all these things produces understanding of the errors in electronic structure calculations. but leaps at the conceptual level are required.16 CHAPTER 1. No reproduction for purposes of sale is allowed. how much (or little) is known of their exact properties. they are more diﬃcult. but nonetheless rigorous. All rights reserved. we start with a very simple example. Its methodology is applied in a large variety of ﬁelds to many diﬀerent problems. biochemistry. h is Planck’s constant. and give a simple but powerful example of DFT at work. and m is the electronic mass. traditional wavefunction methods. Nor is it simply a o method of parametrizing empirical results. Throughout this book. (1. and why it is so useful. we must solve the interacting Schr¨dinger equation: o where the index i runs over the two electrons. Nowadays. and we want only to solve the groundstate quantum mechanical problem for the electrons.
with an exact groundstate wavefunction. This is a much simpler set of equations to solve. He.e. we ﬁnd the HartreeFock equations. which yield an eﬀective potential: 2 HF vS (r) = vext (r) + 1 2 d3 r 0 50 100 150 0. at large distances from the molecule. which in turn is calculated from the solution to the equation. and their wavefunction is a simple product of oneelectron wavefunctions. i. this would look like the familiar two decaying exponentials centered over the nuceli. as well as all the vibrations and rotations. Rather it is a preciselydeﬁned property of any electronic system.8 1 1. Solution of Eq. and its exact solution can be quite demanding. this potential decays as −1/r. E + 1/R. consider our H2 example in the united atom limit. while the depth of the minimum.6 R(˚ ) A Figure 1. (1. Imagine a pair of noninteracting electrons which have precisely the same density n(r) as the physical system. and using density functional methods. making Eq. we note that the correction to the external potential consists of the Hartree potential.4 1. Such a set of equations are often called selfconsistent ﬁeld (SCF) equations. we plot the HartreeFock result and ﬁnd that. This is the KohnSham system.2. −2/r.5) now yields a potential that depends on the electronic density. 1. The missing piece of energy is called the correlation energy.2. one would still need to solve only a 3D equation. Because the electrons are noninteracting. (1. the wavefunction depends on all 3N coordinates of those electrons. the density calculated. INTRODUCTION E(H2 )+1/RE(2H) (kcal/mol) ˚ distances are given in Bohr radii (ao = 0.. and the density extracted. r − r  (1. and the exact KohnSham potential.6 0. 50 exact HF LDA GGA where Φ(r1 .1. 3 It is a simple exercise to extract the exact KohnSham potential from the exact density. In this example. These steps are repeated until there is no change in the output from one cycle to the next – selfconsistency has been reached. but the logic is entirely diﬀerent. it underbinds the molecule signiﬁcantly. is the bond energy. But for larger systems with N electrons. The position of the minimum is the equilibrium bond length. say N . we plot both the physical external potential. imagine a system of two noninteracting electrons in some potential. For example. In the bottom of the ﬁgure.1.1 we plot the results of such a calculation for the total energy of the molecule. (1. 3 Two noninteracting electrons sitting in this potential have precisely the same density as the interacting 2 For the wellinformed. In this view. If we can get our noninteracting system to accurately ’mimic’ the o true system. with an enhancement in between. so that their spatial wavefunction Ψ(r1 .1: The total energy of the H2 molecule as a function of internuclear separation. Note that insertion of this potential into Eq. the second term being the Coulomb repulsion of the nuclei. chosen somehow to mimic the true electronic system. a highly accurate manybody wavefunction for the He atom was calculated. This has been a wellknown deﬁciency of this method. A single simple approximation for the unknown dependence of the energy on the density can be applied to all electronic systems. called orbitals. (1. For our H2 molecule at equilibrium. Even with many electrons. 1. their coordinates decouple. r2 ) = φ0 (r1 )φ0 (r2 ). the basic steps are very much the same. In Fig. and a new potential found. How do we get this mimicking? Traditionally. More generally. the KohnSham wavefunction of orbitals is not considered an approximation to the exact wavefunction.5) . although its minimum position is very accurate.2 1. This is termed therefore a selfconsistent set of equations. it is easy to calculate the probability density of the system: n(r) = 2 d3 r Ψ(r. To emphasize this point.4 0. r )2 . less the exchange potential. then we will have a computationally much more tractable problem to solve. and traditional methods attempt to improve the wavefunction to get a better energy. the eigenvalue problem is then solved. In Fig. minus the zero point vibrational energy. It couples the two coordinates together. if we think of approximating the true wavefunction by a noninteracting product of orbitals. r2 ) is symmetric under interchange of r1 and r2 . as in section ??. v S (r). satisfying: − 1 2 2 The correction to the external potential mimics the eﬀect of the second electron.6) + vS (r) φi (r) = i φi (r). reﬂecting an eﬀective charge of Z − 1.1) a complicated partial diﬀerential equation in 6 coordinates. and then minimize the energy. which is double that shown above. Next. 1. (1. the global energy minimum determines all the geometry of a molecule. which in this case cancels exactly half the Hartree. one can derive its potential vS (r) if one knows how the total energy E depends on the density. An initial guess might be made for the potential. In Fig.1) is complicated by the electrostatic repulsion between the particles.4) The probability density tells you that the probability of ﬁnding an electron in d 3 r around r is n(r)d3 r. n(r ) . and then occupy the ﬁrst N/2 levels. the electrons are in a spin singlet. where the chemical bond has formed.529A). In a KohnSham calculation. and predicts both the energy and the selfconsistent potential for the ﬁctitious noninteracting electrons. WHAT IS A KOHNSHAM CALCULATION? 17 18 CHAPTER 1. in particular screening the nuclei. which is determined uniquely by the density. or the lattice structure of a solid. since it only has 3 coordinates. as opposed to solving a 3N coordinate Schr¨dinger equation. We note at this point that. which we denote as Vee .
but then lose this simple picture of chemical bonding. REINTERPRETING MOLECULAR ORBITALS 19 20 ! CHAPTER 1.6 0. which is too inaccurate for most quantum chemical purposes.1 there are also plotted the local density approximation (LDA) and generalized gradient approximation (GGA) curves. the orbitals are exact and unique. LDA is the simplest possible density functional approximation.8 1 Figure 1. Example from photochemistry. So in this view. eg phasediagram of C. If we can ﬁgure out some way to approximate this potential accurately. thanks to Cyrus’ Umrigar’s very accurate density. 1. i.3. we have a much less demanding set of equations to solve than those of the true system.4 Sampler of applications In this section. Much of the rest of the book is spent connecting these two. INTRODUCTION 4 2 0 2 4 6 He atom 8 0 0. In the KohnSham approach. we need a more accurate wavefunction. but suﬃciently reliable for many solidstate calculations. we highlight a few recent applications of modern DFT. rather than that of the full physical system. if they are the exact KohnSham orbitals. More sophisticated GGA’s (and hybrids) reduce the typical error in LDA by about a factor of 5 (or more). although it typically overbinds by about 1/20 of a Hartree (or 1 eV or 30 kcal/mol).5 Particle in a box The process of bonding between molecules is shown in introductory chemistry textbooks as linear combinations of atomic orbitals forming molecular orbitals of lower energy. Figure 1. when doubly occupied by two noninteracting electrons. We begin with the simplest possible case.. we discover this is only the HartreeFock picture.2: The external and KohnSham potentials for the He atom. Example from solidstate. which. as stated above. there exists (at most) one external potential that.3: Orbital diagram of H2 bond formation. 1.4 KohnSham potential 1s density 1s 1s 2/r ! r 0. using TDDFT. we show how density functionals work with a simple example from elementary quantum mechanics. using ONIOM. 1.7) ˆ ˆ where T denotes the operator for the kinetic energy and V the potential energy.5. and apply density functional theory to it. molecular orbital pictures retain their signiﬁcance. 1.8) H=− 2 dx2 . In section 1. making DFT a very useful tool in quantum chemistry. rather than those of HartreeFock. electrons. This is a paradox. Example using CP molecular dynamics. This provides the underlying concepts behind what is going on in the more previous sections of this chapter. yields the exact density of the H2 molecule. eg ferromagnetism. we write our Hamiltonian as ˆ ˆ ˆ H =T +V (1. and it already greatly improves on HF.3. Thus we are always trying to improve a noninteracting calculation of a noninteracting wavefunction. resolving the paradox. In this section. In general. is rarely accurate enough for quantum chemical calculations. as chemical reactivity is usually thought of in terms of frontier orbitals. Example from biochemistry. The Hamiltonian for a 1D 1electron system can be written as 1 d2 ˆ + V (x) (1.3 Reinterpreting molecular orbitals And a highly accurate approximate density functional calculation produces the full electronic energy from these orbitals. to give the reader a feeling for what kinds of systems can be tackled. we take the simplest example from elementary quantum mechanics. In Fig. But later. In this picture. in studying computational chemistry.2 0.e. as in Fig. 1.1.
Plot the total density for one. they are Gaussians. As the number of particles grows. If we 2 evaluate the approximate kinetic energy again. Repeat for two samespin electrons occupying the lowest two levels of the well. . and divide by φ1 ) as there is in evaluating the integral. Figure 1.1. there’s a simple way to see how this magic has been performed. What happens as the number of electrons grows? Answer: E = 7π 2 /L2 . Check the orthonormality condition..645 loc ∞ −∞ dx n (x). and making them vanish at the boundaries quantizes the energy. (1..10). chapter ?? discusses functionals in some detail. To see how remarkable this is. using n(x) = n1 (x) + L sin2 (k2 x). consider another paradigm of textbook quantum mechanics. For now. all its energy is kinetic. E loc = 63/L2 . In the previous example. In fact.2 0.e..4: Density of 10 particles in a box.5. 1.11/L ..8 1 10 particles in a box dx φ∗ (x)φj (x) = δij i (1.8/L . Since the electron density is the square of the orbital: 2 n1 (x) = sin2 (k1 x).] to indicate a functional dependence.. V (x) = ∞ everywhere. In Fig.. i = 0. . These results are gotten by the evaluation of a simple integral over the density. we simply estimate the energy using TSloc [n].e. Imagine the box with a large number of particles. i. Exercise 1 Particle in a onedimensional box Consider the elementary example of a particle in a box.10) Thus each eigenfunction is normalized. (1. (1. and are well on the way to becoming an expert! 14 12 10 8 6 4 2 0 0 0.. we need only the fact that a functional is a rule which maps a function onto a number. (1. Don’t worry if you never heard of a functional before.4.11) sin(ki x). If we assume it actually becomes uniform as N → ∞. the density becomes more and more uniform (independent of the shape of the box!). 1. π /2L = 4. and from it extract the local density approximation to its kinetic energy. the particles were sine waves. Now. φi = Next we consider the following approximate density functional for the kinetic energy of noninteracting electrons in onedimension. Congratulations. in this one. the number of electrons in it. PARTICLE IN A BOX 21 22 CHAPTER 1. two. But watch what happens if there are two particles. i. We use square brackets [. but present it for calculational use. To see this. What is so great about that? There is not much more work in ﬁnding the exact energy (just take the second derivative of φ1 .13) L loc 2 2 2 2 we ﬁnd TS = 4. The electron probability density is just n(x) = φ0 (x)2 . Show that the solution consists of trigonometric functions (standing waves). L 2 where ki = πi/L. and three particles. Eq. Compare the exact energy with that from the local approximation. and derive it later in the chapter: TS [n] = 1.e. an 11% underestimate.9) Thus ε0 denotes the ground state energy and φ0 the ground state wave function. whereas the exact numbers require solving for all the eigenvalues of the box. this gives us the constant in the local approximation. We ﬁnd now E = 1 + 2 = 5π 2 /2L2 = 24. we have plotted the case for N = 10. a 17% underestimate of the true value.4 0.. since for the particle in a box. of a completely diﬀerent shape. note . and the energies are i = ki /2. where V = 0. 2. except 0 ≤ x ≤ L. we ﬁnd merely a 20% overestimate for one particle: Exercise 3 Kinetic energy of 1d harmonic oscillator What is the groundstate energy of a 1d harmonic oscillator? What is its kinetic energy? Calculate its density. we ﬁnd loc 2 TS = 21. up to N .6 0. a 9% underestimate.7/L2 . Applying the same approximate functional.93/L . The functional we write down will be the local approximation to the kinetic energy of noninteracting spinless fermions in one dimension. fermions of the same spin (the general case for which TSloc was designed). INTRODUCTION The timeindependent Schr¨dinger equation has solutions: o ˆ Hφi = εi φi . namely the onedimensional harmonic oscillator. This problem is given in just about all elementary textbooks on quantum mechanics. You have performed your ﬁrst elementary density functional calculations. In fact. i.12) A local functional is one which is a simple integral over a function of its argument. Thus approximate density functionals produce reliable but inexact results at a fraction of the usual cost. 3 (1. Because ˆ the operator H is hermitian the eigenstates can be chosen orthonormal: ∞ −∞ Exercise 2 Three particles in a box Calculate the exact energy of three identical fermions in a box of length L. We do not derive it here. with errors of order 1/N or smaller. 2 i = 1.
If you evaluate the kinetic energy of a KohnSham system. 4. by making them very fast. the local approximation becomes exact. designed to make you think about what is in the chapter. We continue with more creative questions. Later in the book. Will TSloc [n∗ ] provide an accurate estimate of its kinetic energy? whose leading contribution is simply π 2 N 3 /6. But there are cases where it fails quite badly. If you can do any calculation you like. . The simplest density functional approximation is a local one. that are designed to test your understanding of the material presented in the chapter. This is because we would then have a method for calculating the density and groundstate energy of systems which involved solving a single selfconsistent integrodiﬀerential equation for any electronic system. if we had an extremely accurate noninteracting kinetic energy density functional in three dimensions.2 for the answer). 9. e. but at large computational cost (for large systems). TSloc does very well.g. but were not told the constant of proportionality. If we add an inﬁnitesimal to n(x) at a point. Why is a KohnSham calculation much faster than a traditional wavefunction calculation? 2. justifying your answers in each case.6 Questions We begin with conceptual questions..2 for the answer). Can you ﬁnd one. we could revolutionize electronic structure calculations. Suppose you’d been told that TSloc is proportional to dxn3 (x). 1.2 for the answer).5 for the answer).6.e. δ(x) as → 0. Since the density has become uniform (to this order). Consider a single electron in an excited state.14) 8. In the chapter and exercises. INTRODUCTION that the exact energy is π2 N 2 π2 E= j = N (N + 1)(2N + 1)/6 2 2L j=1 2L2 (1. they ﬁnd T S exactly. we will see how. Now answer the following simple questions as well as you can. This would be a true density functional calculation. 6. for two electrons of opposite spin in a box? (see section 9..1. with density n∗ (x). Why is the density cubed in the local approximation for TS ? (see section ?? for the answer). Repeat above question for 1/r . is it equal to the physical kinetic energy? 3. and say why? 10. QUESTIONS 23 24 CHAPTER 1. i. which can be measured in scattering experiments. 5. In what way will TSloc change if spin is included. 1. yielding the constant. how does TSloc change? (see section 2. Most present calculations do not do this. what procedure might you use to determine the constant in TSloc ? (see section 11. So now you even know an unsolved problem in density functional theory. What form might a correction to TSloc take? (see section 17. 7. By solving an eﬀective singleparticle equation for the orbitals (the KohnSham equation).
For a δfunction potential. INTRODUCTION Extra exercises 1. QUESTIONS 25 26 CHAPTER 1.6. v(x) = −Zδ(x). . ﬁrst guess if it will be an overestimate or an underestimate. the groundstate kinetic energy is Z 2 /2 and the density is Z exp(−2Zx).1. Before calculating the local approximation. with groundstate energy −a2 /2 and density a sech2 (ax)/2. and by how much. Another famous example is v(x) = −a2 sech2 (ax). Repeat above problem. 2.
yielding P [r] = 2π 0 (2. which are periodic. inside the integral.2) Functionals In this chapter. These are called local because. We write P [r] and A[r] to indicate this functional dependence. with f (n) = 1. such as the ellipse r(θ) = 1/ sin2 (θ) + 4 cos2 (θ) (2. Similarly. localized to one point. The contribution to the area is that of a triangle of base r cos(dθ) ≈ r and height r(θ) sin(dθ) ≈ rdθ. For the area.1 What is a functional? 1 For an abitrary r(θ) the integral above can be evaluated to give the area that is enclosed by the curve. Note that. 2. (2. On the other hand. P is a semilocal functional of the radius. For example. A functional assigns a number to a function. since r = r(θ). FUNCTIONALS integrate over the entire range of angles. Later we will see some examples of fully nonlocal functionals. yielding A[r] = Chapter 2 1 2 rdθr = 1 2 2π 0 dθ r2 (θ) (2.28 CHAPTER 2. To do this. for a given curve.645n3 . We have already come across a few examples of density functionals. we can deﬁne the perimeter P as the dθ f (r(θ)). the local approximation to the kinetic energy is a local functional of the density. For this simple example. one needs only to know the function right at a single point to evaluate the contribution to the functional from that point. f = r 2 /2. we can calculate the functional.4) In both cases. (2.2 Functional derivatives there is a single welldeﬁned value of P and of A. This mathematics will be needed when we describe the quantum mechanics of interacting electrons in terms of density functional theory. the exact kinetic energy is a semilocal functional. in this case the area enclosed by the curve. we can use elementary trigonometry to deduce explicit formulas for these functionals. we must deﬁne a functional derivative.e. For every such curve. Consider the contribution from an inﬁnitesimal change in angle dθ. such as in Fig.3) dθ r2 (θ) + (dr/dθ)2 .5) where f (r) is a function of r. and 1 c 2000 by Kieron Burke. we introduce in a more systematic fashion what exactly a functional is. Such functions describe shapes in twodimensions of curves which do not “doubleback” on themselves. On the other hand. in much the same way as we learn how to diﬀerentiate regular functions in elementary calculus. for any oneelectron system. we do know it exists: Every bounded curve has a perimeter and an area. the inﬁnitesimal change in the perimeter is just the line segment in polar coordinates: δP 2 = dr2 + r2 (θ)dθ.1) where n (x) = dn/dx.. These are functionals of r(θ). In the opening chapter. and the area A as the area enclosed by it. consider all functions r(θ). Now. once we know r(θ). 2. When we show that the groundstate energy of a quantum mechanical system is a functional of the density. i. Imagine making a tiny increase in a function. n (2. r(θ + 2π) = r(θ). we will then want to minimize that energy to ﬁnd the true groundstate density. and asking how the value of a functional has changed due 27 . 0 ≤ θ ≤ 2π. we can write dr = dθ (dr/dθ). since it can be written in the form: A[r] = 2π 0 A function maps one number to another. called the von Weisacker functional: TSVW [n] = 1 8 ∞ −∞ Figure 2. as it depends not only on r. Thus this functional maps a real function of one argument to a number. in the sense that. 2. even if we don’t know the relation explicitly. All rights reserved. The area is a local functional of r(θ). and how to perform elementary operations on functionals. but dr/dθ.1.6) length of the curve. To begin with.1: A 2D curve which is generated by a function r = r(θ) dx n2 . we must learn how to diﬀerentiate functionals.
tell you how much a functional changes when its arguments changes by a small “amount” (in this case. is just 2π δA A[r + δr] − A[r] = δr(θ). To ﬁrst order in δr. (2. change? A[r + δr] − A[r] = 1 2 2π 0 How does A[r] The ﬁrst integral is in just the right form for identifying the contribution to the functional derivative. (2.e. where a (n) = da/dn. or just r(θ) in this case. δn(x) 4n 8n The density functionals we use in practice are threedimensional: Exercise 6 Hartree potential The Hartree (or classical electrostatic) energy of a charge distribution interacting with itself via Coulomb’s law is given by U [n] = 1 2 d3 r d3 r n(r) n(r ) .15) Show that its functional derivative.10) where r = dr/dθ and all functions are assumed to have argument θ.13) Exercise 5 Derivative of a semilocal functional Show that the functional derivative of a semilocal functional B[n] = assuming n and its derivatives vanish rapidly as x → ∞. Then we take the The last piece of functional technology we need for now is how to solve a constrained optimization problem.16) (2. i. (2. subject to the constraint that P [r] = l. denoted δA/δr(θ) is just the change in A divided by . For the perimeter.12) (2. the way to ﬁnd a functional derivative is to evaluate the expression A[r +δr]−A[r] to leading order in δr. Use your answer to show δTSVW n n2 = − + 2.7) The functional derivative. how to maximize (or minimize) one functional. the maximum change becomes arbitrarily small. so does the functional derivative. is vH [n](r) = 2. for any local functional A[n] = dx a(n(x)). Since this is linear in the change in r. Often it is necessary to do an integration by parts to identify the functional derivative. the general deﬁnition. the Hartree potential. we add an inﬁnitesimal change δr(θ) = δ(θ − θ0 ). In general.. and the resulting integral must be cast in the form of a function times δr.e. the functional derivative is δA/δn(x) = a (n(x)). Thus we need to maximize A[r]. such complications arise. Thus δA[r] vA [r](θ) ≡ = r(θ).2. subject to a constraint imposed by another.14) where n (x) = dn/dx.11) 0 r2 + r 2 r +r2 2 For purists. because r(θ) is periodic.9) δr(θ) where the notation vA [r](θ) emphasizes that the functional derivative of A[r] is a θdependent functional. Exercise 4 Derivative of a local functional Show that. we take a Gaussian of tiny but ﬁnite width. Just as the usual derivative df /dx of a function f (x) tells you how much f changes when x changes by a small amount.3 EulerLagrange equations d3 r (2.. a small function). There is a wellknown method for doing such problems. dn/dx)..2. r − r  n(r ) . for any inﬁnitesimal change in r. so that as width to zero! → 0. A little calculus and algebra ﬁnally yields p[r](θ) = r3 + 2rr 2 − r2 r ) (r2 + r 2 )3/2 ∞ −∞ dx r r2 + r 2 (2. i. i.8) dθ 0 δr(θ) where the functional derivative δA/δr(θ) is that function of θ which makes this formula exact for any small change in r(θ). FUNCTIONAL DERIVATIVES 29 2 30 CHAPTER 2. FUNCTIONALS to this increase.17) dθ (r + δr)2 + (r + δr )2 (2. is vB [n](x) = d ∂b ∂b − ∂n dx ∂n b(n. a function. But we can perform an integration by parts to ﬁnd δP d r − = p[r](θ) = √ 2 δr(θ) r + r 2 dθ √ dθ (r + δ(θ − θ0 ))2 − r2 = 2π 0 dθ r δ(θ − θ0 ) = r(θ0 ) (2. r − r  (2. We have P [r + δr] = 2π 0 Note that the endpoint term of the integrationbyparts vanishes. f (x + dx) − f (x) = (df /dx) dx + O(dx2 ). called the method .e. we might want to know what is the maximum area we can enclose inside a loop of string of ﬁxed length l. we ﬁnd √ 2π rδr + r δr P [r + δr] = dθ r2 + r 2 1 + 2 0 r +r2 2π r r √ = P [r] + dθ δr + √ 2 δr (2. For example. but the second is not.
Exercise 8 Second functional derivative Find the second functional derivative of (a) a local functional and (b) TSVW . This tells us that the optimum shape is a circle. check that Eqs. For ﬁxed particle number. FUNCTIONALS of Lagrange multipliers. 2. The largest area enclosable by a piece of string of length l is A[r = µ] = l2 /(4π). a constant. the someone might be Mother Nature. If someone just tells you a number for any density you give them.g. Compare the functional derivative of TSVW [n] with TSloc [n] for some sample oneelectron problem. construct a new functional B[r] = A[r] − µP [r]. then the Von Weisacker. devise a method for deducing if Mother Nature’s functional is local or not. Is there a simple relationship between TS and dx n(x)δTS /δn(x)? First consider the local approximation. and the number might be the total energy measured by experiment. Check that the circle of the same perimeter has a larger area. (2. (2. 4.4 Questions 1. is there any indeterminancy in the functional derivative of a density functional? . P [r = µ] = 2πµ = l.19) A simple solution to this equation is r(θ) = µ. = −µ = r(θ) − µ δr δr δr (r2 + r 2 )3/2 Extra exercises 1. by setting its functional derivative to zero: δB δA δP r3 + r 2 (2r − r ) = 0. QUESTIONS 31 32 CHAPTER 2. e. (2.. Hint: You need only do the ﬁrst octant. where Z = 1.4.2) and (2. Following the example in the Appendix (A. Then extremize the new functional B[r].1). Estimate the change in E when Z becomes 1. Comment on your result. x > y > 0. 2. Consider a square of side 1.18) where µ is at this point an unknown constant. Comment. are correct. Exercise 7 Change in a functional Suppose you know that a functional E[n] has functional derivative v(r) = −1/r for the density n(r) = Z exp(−2Zr).4).2. and the radius of that circle can be found by inserting the solution in the constraint. 3. 2.1. For the unit square centered on the origin.
simpler approach to improving our trial wavefunction is to make a linear combination with other trial wavefunctions. y) = x2 /2 − xy. φ ∞ −∞ Next. but not vary the length scale. (3. To see how well we did for this problem. ONE ELECTRON Chapter 3 One electron 1 independent of the particular problem. via Eq. Having learnt about functionals in Chapter ??. as you only need take one derivative. (3. The ﬁrst method is to include an adjustable parameter. Find T (α) and V (α) and plot their sum as a function of α. where φE (x) = exp(−x) as a trial wavefunction for the problem in Ex.2) −∞ c 2000 by Kieron Burke. Thus we write φtrial (x) = c0 φ0 (x) + c1 φ1 (x). and choose the lowest one. we review the traditional wavefunction picture of Schr¨dinger for one o particle. This principle says that we can use any normalized wavefunction to calculate the expectation value of the energy for our problem. The ﬁrst term is independent of y. All rights reserved.2 Trial wavefunctions We start oﬀ with the simplest possible case. Calculate the errors made in the previous exercises. the RayleighRitz variational principle: ˆ E = min φ  H  φ .34 CHAPTER 3. we apply the same operation on a given trial wavefunction. no matter what the problem we are applying that wavefunction to. Note that this has led to a nonlinear optimization problem in α.5) dx φ(x)2 = 1. and what is your estimate of the groundstate energy? Compare with previous calculation. the potential energy is a very simple functional: ∞ V [φ] = dx V (x) φ(x)2 . and we are guaranteed to get an energy above the true groundstate energy. φ ∞ −∞ dx φ(x)2 = 1 (3. (9). we may write this principle in the following useful functional form. The kinetic energy is a universal functional. we put trial wavefunctions into the variational principle to ﬁnd upper bounds on the groundstate energy of a system. make an estimate for the energy.. (3. without having to evaluate it again for each value of y. This is the exact answer. the kinetic energy of the particle can be considered as a functional of the wavefunction: T [φ] = ∞ −∞ where the second form is gotten by integration by parts. 3. But the potential energy functional diﬀers in each problem. and a rigorous statement about the exact energy. it is not universal.) Thus. Now our variational principle can be written by writing E = min {T [φ] + V [φ]}. For example. for any given normalized wavefunction. (3.2). 1 dx φ∗ (x) − 1 d2 1 φ(x) = 2 dx2 2 ∞ −∞ dx φ (x)2 (3. extremely powerful principle throughout this book. (This second form is much handier for many calculations. An alternative. Exercise 11 1d Hydrogen atom Use φE (αx). one electron in one dimension. 33 . we can use the same wavefunction for every 1d oneelectron problem. as a trial wavefunction. there is a single number T . and the prime denotes a spatial derivative. no matter what our physical problem is. the next exercise yields the exact answer.g. e.. from basic quantum mechanics. 3. For example. i. Exercise 9 Gaussian trial wavefunction An interesting potential in onedimension is just V (x) = −δ(x).1 Variational principle We need simply evaluate the energy of all possible normalized wavefunctions.e. and then use that same list to ﬁnd the minimum for any y. Recall.1) We will use this basic. Find the lowest energy. Using a Gaussian wavefunction. which can be calculated from it. and comment. We can improve on our previous answers in two distinct ways. into our trial wavefunction.3) The kinetic energy of a given wavefunction is always the same. the length scale. Exercise 10 Adjustable trial wavefunction Repeat Ex. Similarly. (9) with a trial wavefunction φG (αx) (don’t forget to renormalize). but introducing our own speciﬁc notation. Which value of α is the best. φG (x) = π −1/4 exp(−x2 /2).4) In this chapter. and get an upper bound on the groundstate energy. Varying exponents in trial wavefunctions leads to diﬃcult optimization problems. so we could make a list of it value for every x. This is analogous to trying to ﬁnd the minimum over x of f (x. where δ is the Dirac delta function.
where Hij = is the Hamiltonian matrix. we cannot be relying on lucky guesses. we’ve found the exact solution to the given problem. From the RayleighRitz variational principle.14) There are only slight complications in the real threedimensional world. and µ is a Lagrange multiplier. In general. Eq. Only if the basis of trial wavefunctions is chosen as orthonormal (Eq. and be sure we’ve done so. by applying the techniques of Chapter 2 to the variational problem of Eq. with emphasis on the Coulomb problem. 3.1).12) δφ(x) δφ(x) δφ(x) Thus δF = −φ (x) + 2(v(x) − µ)φ(x) δφ(x) 1 − φ (x) + v(x) φ(x) = µ φ(x) 2 (3.4 Diﬀerential equations (3. − l(l + 1) 1 d 2d r + + v(r) Rnl (r) = εnl Rnl (r).9) into the Schr¨dinger Equation gives o the radial equation These are written about in almost all textbooks on quantum mechanics. = 2v(x) φ(x). (3.5 Virial theorem We close this chapter with a statement. This should all make a lot of sense. The general rule is that. where analytic solutions are few and far between. this should vanish.8) is the overlap matrix. dV .6) (3. trial wavefunctions can be used to get approximate answers. When we’re given a variational problem to solve. Exercise 13 Three dimensions Use the trial wavefunctions of an exponential and a Gaussian in r to deduce the groundstate energy and orbital of (a) the hydrogen atom (Z = 1). we can do so.3. where r = r. The virial theorem says that.10) will S = 1. Inserting (3. ONE ELECTRON ˆ and minimize the expectation value of H. ϕ). Less frequently treated is the threedimensional harmonic oscillator problem. The standard textbook problem of chemical bonding is the formation of molecular orbitals + from atomic orbitals in describing the molecular ion H2 . 2 dr 2 2r dr 2r (3. but the exact solution can often be found by solving the diﬀerential equation that results from extremizing the functional. Taking functional derivatives. and want to minimize E[φ]. of the virial theorem in one dimension. given the constraint that the wavefunction be normalized. Exercise 12 1d H+ 2 For a potential V (x) = −δ(x) − δ(x + a). and calculate the bonding and antibonding energies as a function of atomic separation a. given a variational principle. use φmol (x) = c1 φE (x) + c2 φE (x + a) as a trial wavefunction. v(r) = kr 2 /2. Then the spatial parts of the orbitals are characterized by 3 quantum numbers: φnlm (r) = Rnl (r)Ylm (θ. Happily. and (b) the harmonic oscillator with k = 1. This leads to a set of linear equations H − ES = 0. subject to the constraint that φ be normalized. But in the real world. but only the ground state satisﬁes the variatonal principle (all others are only local minima).13) At the minimum. we must ﬁnd the exact answer.7) ˆ dx φ∗ (x)Hφj (x) i dx φ∗ (x)φj (x) i (3. Thus we must minimize: F [φ] = T [φ] + V [φ] − µ(N [φ] − 1) (3. (3. and the equation be an ordinary eigenvalue equation. v(r). In a few lucky cases.3. We choose the wavefunction real.11) where N [φ] = dx φ2 (x). v(r) = −Z/r. given the constraints.15) 2T = x dx . (3.3 Three dimensions We have just seen how to construct various trial wavefunctions and ﬁnd their parameters by minimizing the energy. (3. We will consider only spherical 3d potentials. The proof will be given later in the book. 1. = 2φ(x) (3. we ﬁnd: δN δV δT = −φ (x).9) where the Ylm ’s are spherical harmonics. we have rededuced the Schr¨dinger equation. and Sij = ∞ −∞ ∞ −∞ 3. 3. THREE DIMENSIONS 35 36 CHAPTER 3. without proof.6) is a generalized eigenvalue equation. for an eigenstate of a given potential.10) and we can identify µ = E0 by integrating both sides with φ(x). yielding: (3. Note however that all eigenstates satisfy the Schr¨dinger o o equation.
(c) the threedimensional Hydrogen atom. 1. for these purposes. once the potential is interpreted as all the potential energy in the problem. Some approximations automatically satisfy the virial theorem. What happens as Z → ∞ and as Z → 0? 3. Note that. and 2. 2. Exercise 14 Virial theorem for one electron Check Eq. (9) and (10). Thus. where V is the sum of the external potential and the electronelectron repulsion.6. How would you check to see if your trial wavefunction is the exact groundstate wavefunction for your problem? 5. In cases that do. for any atom.16) for the ground state of (a) the 1d H atom. Consider what happens to the potential v(x) = Axp as p gets large. Sketch. the 1d Hatom for Z = 1/2. QUESTIONS 37 38 CHAPTER 3. Suggest a good trial wavefunction for a potential that consists of a negative delta function in the middle of a box of width L. 3. because one only needs either the kinetic or the potential energy to deduce the total energy.6 Questions All the questions below are conceptual. and also true for interacting problems. ONE ELECTRON In the special case in which v(x) = Axp . on the same ﬁgure. (3. In cases that do not. 1. The theorem is also true for threedimension problems. p = −1. Also check it for Ex.16) This can save a lot of time when doing calculations. others do not. so that dx 2T = pV (3.3. or to avoid some work. Can the virial theorem be applied to the particle in the box (even if it yields trivial answers)? . the δfn acts as if it had power p = −1. we can use the virial theorem to judge how accurate the solution is. and E = −T = −V /2. What is the exact kinetic energy density functional for one electron in onedimension? 4. x dV = p V . (b) the 1d harmonic oscillator. we can use the virial theorem to check we dont have an error.
7) φ(x) = Z exp(−Zx). and satisfy the oneelectron problem. while their spatial part is symmetric: Ψ(r1 σ1 . The contribution of the kinetic energy and the potential energy is then just twice as big as in the single electron system: Ts = 2 Z 2 /2 and therefore the total energy becomes E=2 0 where the spatial part is symmetric under exchange of spatial coordinates and the spin part 1 χSinglet (σ1 . We know from chapter 3 that this yields √ (4. but is much weaker and more shortranged. 39 .e. r2 σ2 ) = Ψ(r1 . σ2 ).9) c 2000 by Kieron Burke.1 Antisymmetry and spin 1 The electronelectron repulsion operator is a twobody operator. This is in fact lower than the true groundstate energy of this problem. which has the same scaling property. because we have failed to evaluate part of the energy in the Hamiltonian. by coupling the two coordinates together. the operators in the Hamiltonian depend on all the coordinates. Then the diﬀerential equation decouples.5) Ψ The most naive approach might be to ignore the electronelectron repulsion altogether. but to ﬁnd approximate solutions. meaning the spinpart of their wavefunction will be antisymmetric. and we prefer to use δ(x − x ). but are unneccessary for the basic logic of our presentation. we have made a very crude approximation to do this. (4. However.2 HartreeFock Vext = 2(−Z 2 ) Vee = 0 (4. but in a simple way. and we can write: Ψ(x1 . The interaction between two electrons in three dimensions is Coulombic. we might want to solve the Schr¨dinger equation with o this Hamiltonian.1) for the H2 molecule. etc. 1/x − x  in onedimension is an exceedingly strong attraction. and so the wavefunction is a function both of spatial coordinates and spin coordinates. σ1 . x2 )2 = 1. 4.6) The groundstates of our systems will be spinsinglets. its much easier to use the variational principle: ˆ E = min Ψ  H  Ψ .8) For more than one electron. x2 ) = φ(x1 )φ(x2 ) (ignoring Vee ) (4. These are all valuable tools for studying advanced quantum mechanics. r1 . This decoupling of the coordinates makes it possible to handle very large systems.. (4. For our purposes. the Hamiltonian and the wavefunction includes one coordinate for each particle. However. second quantization. electrons have two possible spin states. Note that it is this term that complicates the problem. Matsubara Green’s functions. r2 .4) When there is more than one particle in the system.40 CHAPTER 4.3) Chapter 4 Two electrons In this chapter. as in Eqn. each term depending on two coordinates simultaneously. dx1 dx2 Ψ(x1 . since we need only solve for one electron at a time. r2 ) χSinglet (σ1 . this implies the 2electron wavefunction satisﬁes: Ψ(r2 . All rights reserved. (1. 2 dx1 dx2 (4. A general principle of manyelectron quantum mechanics is that the wavefunction must be antisymmetric under interchange of any two sets of coordinates. We will now consider the problem of onedimensional He as a prototype for two electron problems in general. (4. We keep everything as elementary as possible. TWO ELECTRONS meaning that they are sums of terms which each depend on only one coordinate at a time: 1 d2 d2 ˆ T = − 2 + 2 2 dx1 dx2 ˆ Vext = vext (x1 ) + vext (x2 ) (4. All this is covered in elementary textbooks. σ2 ) = √2 (  ↑↓ −  ↓↑ ) is antisymmetric. which gives for He (Z=2) E = −4. up or down. i. σ2 .2) where both orbitals are the same. This is homogeneous of degree 1 in coordinate scaling. 4. Both the kinetic energy and the external potential are onebody operators. 1/r − r . σ2 ). avoiding sophistry such as the interaction picture. 1 = −Z 2 (4. The Hamiltionian is then: d2 1 d2 ˆ H = − 2 + 2 − Zδ(x1 ) − Zδ(x2 ) + δ(x1 − x2 ). we review the traditional wavefunction picture of Schr¨dinger for two o electrons.1) To ﬁnd an exact solution to this. σ1 ) = −Ψ(r1 . Furthermore.
we ﬁnd: Vext = −2Zα Vee = α/2 (4. note that these are selfconsistent equations. (Compare with Eq. for this case). giving up when addition of further parameters has negligible eﬀect.15). We chose the length scale of our orbital to be that of the Vee = 0 problem. HARTREEFOCK 41 42 CHAPTER 4. 2 2 dx 2 (4. from the derivation. where vH (x) = Exercise 15 HartreeFock equations for two electrons Derive the HartreeFock equation for two electrons. we can do better still. A single electron should only see the potential due to the other electron.18) (4. For two electrons in HF. If we simply minimize this functional of the orbital. (4. Because we evaluated all parts of the Hamiltonian on our trial wavefunction. we ﬁnd αmin = Z − 1/4 and thus 2 Emin = −αmin = −(Z − 1/4)2 = − ∞ δU = dx n(x )δ(x − x ) = n(x) (4. (4. since the potential depends on the density. keeping the orbital normalized. the nuclear charge. This is due to exchange eﬀects: The Hartree potential is the electrostatic potential generated by the total charge density due to all the electrons. We could do this by including many variational parameters in a trial orbital. (4. = −γ 2 /2.16) −∞ δn(x) is the Hartree potential.17) (4.. A systematic approach to this would be to consider an inﬁnite set of functions. Minimizing this. The energy. is: E[φ] = 2TS [φ] + 2Vext [φ] + U [φ]/2. (1. as a functional of the orbital. or 0. using the techniques of Chapter ??. Vee = U/2. as a function of α. The energy of the eﬀective noninteracting system is not the energy of the original interacting system. It is straightforward to numerically solve the orbital equation Eq. with almost no extra work. essentially exactly. showing there is very little diﬀerence. Inserting this orbital into the components of the energy. we know that the true E ≤ −3. we should evaluate the expectation value of Vee on our wavefunction: Vee = ∞ −∞ dx ∞ −∞ dx φ(x)2 φ(x )2 δ(x − x ) = Z 2 ∞ −∞ dx exp(−4Zx) = Z . Write φ(x) = γ cosech (γ x + β) where coth(β) = Z/γ. Insertion into Eq. This is also known as the classical or electrostatic potential. which may not seem like much. (4. and we discuss this much more later.6 eV or 13 kcal/mol.6) of the introduction.0625. which is the solution of the equation. so the potential is a delta function with an added Hartree potential of 4 exp(−7x). and γ = Z − 1/2.13) We have lowered the energy by 1/16 of a Hartree.10) For Helium (Z=2).1.12) 7 4 2 = −3.15) on a grid. (4. as it is the electrostatic potential due to the charge distribution in classical electrostatics (if the electrostatic interaction were a δfunction. There are several important aspects of this equation that require comment.15) yields but this does not mean E = −γ 2 . 2 (4. but rather Z − 1/4. using the variational principle. Chemical accuracy requires errors of about 1 or 2 kcal/mol. and U is the 1d equivalent of the Hartree energy (discussed more below). subject to the restriction that the orbital is normalized. we ﬁnd: − 1 d2 1 − Zδ(x) + vH (x) φ(x) = φ(x).14) where n(x) = 2φ(x)2 . In this case. The best solution to this problem is to ﬁnd the orbital that produces the lowest energy.11) Ts = 2 α2 /2 so that the total energy. But we can treat this instead as an adjustable parameter. which in turn depends on the orbital. TWO ELECTRONS To improve on our estimate. To get an idea of what the eﬀective potential looks like. We say the other electron is screening the nucleus.2. pushing the electron farther away from the nucleus. is E(α) = α2 − 2α(Z − 1/4). 1 U [φ] = 2 ∞ −∞ dx n (x) 2 (4. These can be solved in practice by the iterative method described in the introduction. and ask what value minimizes the energy. hence the division by two. 4. note the factor of 2 dividing the Hartree potential in the equation.14). we construct it for our approximate solution above. and restore the variational principle. But there is an underlying analytic solution in this case. α = Z. Next. i. we can provide a more direct scheme. Vee = +1 so that the ground state energy becomes E = −4 + 1 = −3. Eq. TS and Vext are the oneelectron functionals mentioned in chapter 3. or E = −Z 2 +Z/2−1/12 as the exact HF energy. which is still a signiﬁcant error by modern standards. There α = 1. First. The eﬀective nuclear charge is no longer Z. Eq. we see that. having learned some functional calculus. and minimize the energy with respect to each of them.15) . usually of increasing kinetic energy. etc.) This equation is the HartreeFock equation for this problem. However. but is about 1. and to include more and more of them. (4. by minimizing the energy as a functional of the orbital. we see E = 2 −U/2 (since = TS +Vext +U/2). we plot both the exact orbital and the scaled Hydrogenic orbital. We ﬁnd U/2 = Z/2−1/6.e. This contribution is positive.75 for He.4.7 eV or 40 kcal/mol. slightly lower than our crude estimate using hydrogenic wavefunctions. But this slight diﬀerence is 1/48. However. In Fig.
Since the HF energy is exactly 3 1/12. the external potential dominates over the repulsion.. What is the estimate of the ionization energy given by the crude Hydrogenic orbital approximation? 2. the system has signiﬁcant correlation.277 0. (For help. 1.486 0.612 13. both exact and simple exponential with eﬀective charge.4 1. but we’ve made great strides so far: Even our crudest method was good to within 5%. Check that the HF solution satisﬁes the virial theorem (for power 1). i. the interaction becomes more important.5 2 2. and so the noninteracting solution becomes highly accurate.4 0. 1. Its just that we need about 0. clearly EC is almost independent of Z. But when Z is of order 1. notice that. while the exchange energy grows as −5Z/8.025 0. or simply numerically on a grid.8 0.4.17) satisﬁes the HF equation with the correct eigenvalue.5 3 3.2 0 0 0. the eﬀective nuclear charge.5 1 1.e. Thus we need to estimate correlation energies to within about 10% for useful accuracy in quantum chemistry. In traditional quantum chemistry.1: Energies for twoelectron ions. but is rather a complicated function of both variables simultaneously.19) The approximate solution of the HartreeFock equations is not exact. The true wavefunction satisﬁes the exact Schr¨dinger equation. for H − . the correlation energy is 71 mH. Exercise 16 Analytic solution to HF for 1d He This exercise is best done with a table of integrals. because the true wavefunction is not a product of two orbitals. Returning to the 1d world.3 Correlation x Figure 4..3. Plot the eﬀective potential (external plus half Hartree) using your approximate solution.2 1 exact approx 1d He atom in HF φ(x) 0.6 0. Perhaps counterintuitively. which implies that all errors in the groundstate energy are at least secondorder in the error in the wavefunction (don’t expect other expectation values to be as good). 4. Finally.911.2. a mathematical symbolic program.e.904 1. Hartree and exchange dominate over correlation. What is the ionization energy given by the HF solution? 3. and o also minimizes the groundstate energy functional for the given external potential.042 He 2.1: HF orbital for 1d He atom. As Z is reduced. correlation becomes important.. This is a gift of the variational principle. We illustrate these results in Table 4. the system becomes unstable to ejecting one of the electrons. For Z just a little greater than that.528 0.044 Table 4. Can you think of another way to estimate the ionization energy from the HF solution? Exercise 18 He in 3d Repeat the approximate HF calculation for He in three dimensions. TWO ELECTRONS 1. The groundstate energy for 1d He is 3. (4. This may seem daunting. i. (4. if Z is made too small. For larger Z values. He.381 0. Show that the orbital of Eq. the total energy grows as −Z 2 (the Hydrogenic result). but its relative size diminishes. the exchange energy is so small that correlation is 11% of it. (4. the correlation energy is deﬁned as the diﬀerence between the HartreeFock energy and the exact groundstate energy.154.5 4 atom E HF E EX EC − H 0. So EC becomes an ever smaller fraction of EX as Z grows. Thus the magnitude of the exchange energy grows with Z.862 2. Show that the orbital of Eq.17) is normalized.20) We illustrate the eﬀects of correlation on the simplest system we have. and the total physical energy? Make a rigorous statement about both the exact HF energy and the true groundstate energy. We can adjust the relative importance of correlation by considering various values of Z. trad EC = E − E HF .042 Be++ 13.656 2. its ionization potential passes through zero. HARTREEFOCK 43 44 CHAPTER 4. What is the orbital energy.e. all the same remarks apply qualitatively. 3. for large Z. The twoelectron ion unbinds at the critical value of Z = 0. its energy becomes less than that of the oneelectron ion. (4. i.2%! . Eventually. Similarly. 2. Exercise 17 Ionization energy of 1d He The ionization energy of He is deﬁned as I = E 1 − E2 where EN is the groundstate energy with N electrons in the potential. For large Z. this calculation is done in many basic quantum books).
and what it does wrong.4 Questions All the questions below are conceptual. calculate the critical Z at which 1d He becomes unstable. 4. Suggest a simple improvement.0342.154 5 3 2 0 Table 4. TWO ELECTRONS E % err ﬁxed Z adjustable Z HF exact 3 3.3255868 and β = 0.4. Comment on what it does right. 4. Is the HF estimate of the ionization potential for 1d He an overestimate or underestimate? 2.2: Approximate Energies for 1d He. the kinetic energy of the true wavefunction or that of the HF wavefunction for 1d He? . 3. Consider the approximate HF calculation given in section 4. Exercise 19 Ions on the verge of ionization 1.4.0625 3.2. Calculate the ratio of EC to EX at Zc . Repeat using the exact HF result. Find the critical value of Z.21) where α = 3/4 − 4/(3π) = 0. Using the simple exponential approximation. 1. Is there an alternative method of estimating Z c in this case? Which is better? 3.0833 3. 2. QUESTIONS 45 46 CHAPTER 4. Which is bigger. The exact energy is accurately reproduced by E= β 1 −2Z 2 + Z − α + 2 Z (4.
x2 . . . . . xj . xN ) = EΨ(x1 . 5. . due to the antisymmetric nature of the determinant.48 CHAPTER 5. . This is a purely Pauliexclusion principle eﬀect. e. xN ). . . Exercise 20 Exchange energies for one and two electrons Argue that. . . . The Coulomb interaction for a single Slater determinant yields. . 5. including (1. for a determinant of doublyoccupied orbitals. we normalize our wavefunction by: dx1 .).6) d3 r i. MANY ELECTRONS Note that the factor of 2 in the electronelectron repulsion is due to the sum running over all pairs. . the full orbitals can be written as a product.3) The ﬁrst contribution is just the noninteracting kinetic energy of the many orbitals. we note that the wavefunction for N electrons is a function of 3N spatial coordinates and N spin coordinates. dxN Ψ(r. 3 (5. xN ) = 1.j occ d3 r φ∗ (r) φ∗ (r ) φiσ (r ) φjσ (r) iσ jσ r − r  (5. . φi (x) = φi (r)σ . .4. This is the electrostatic energy of the charge density in electromagnetic theory. . (5. . .8) Chapter 5 The groundstate energy can be extracted from the variational principle: ˆ ˆ ˆ E = min ΨT + Vee + Vext Ψ . . .e. . The Schr¨dinger equation is then o ˆ ˆ ˆ T + Vee + Vext Ψ(x1 . (5. . Ψ (5. . . . . The second is the Fock or exchange integral.) = −Ψ(x1 . EX = −U .2) and (2. . xi . . xi .1). with the density being the sum of the squares of the occupied orbitals. for two electrons in a singlet. we generalize the concepts and ideas introduced for two electrons to the N electron case. Φ where dx denotes the integral over all space and sum over both spins. . . . xN ) = φ1 (x1 ) · · · φN (x1 ) . . F. . The electronic density is deﬁned by n(r) = N σ (5. Φ(x1 . xN )2 . and the last is their external potential energy. . . (5.13) vext (ri ). . and this will usually be a single Slater determinant of occupied orbitals.1 Ground state once the minimization is performed over all normalized antisymmetic wavefunctions. .5) (5. two contributions: ˆ ΦVee Φ = U [Φ] + EX [Φ]. 2 (5.4) The ﬁrst of these is called the direct or Coulomb or electrostatic or classical or Hartree contribution. we denote the wavefunction by Φ instead of Ψ.7) . The HartreeFock energy is ˆ ˆ ˆ E = min ΦT + Vee + Vext Φ .9) Many electrons In this chapter. . i. . .and twoparticle operators: 1 ˆ T =− 2 i=1 ˆ Vext = and N i=1 N 2 i.11) (5.12) and retains the interpretation that n(r)d3 r is the probability density for ﬁnding any electron in a region d3 r around r.10) We deﬁne carefully our notation for electronic systems with N electrons. .2 HartreeFock In the special case of noninteracting particles. . . for one electron. . EX = −U/2. dxN Ψ(x1 . σ. .2) dx2 . First. Our favorite operators become sums over one. Writing xi = (ri . . and show that. . given in Eq. . .1) For systems with equal numbers of up and down particles in a spinindependent external potential. 1 1 N ˆ Vee = 2 i=j ri − rj  47 (5. σi ) to incorporate both. being EX [φi ] = − 1 2 σ (5. Note that the antisymmetry principle implies Ψ(x1 . and each spatial orbital appears twice. ignoring its quantum origin.. The density is normalized to the number of electrons d r n(r) = N.. . xj .g. . φ1 (xN ) · · · φN (xN ) (5.
5 2 2. These can mostly be divided into two major types: perturbation theory.529 6 37. we often do not need total energies to this level of accuracy. using the techniques of chapter 2 yields the HartreeFock equations − 1 2 2 + vext (r) + vH (r) φi (r) + fF. and the energy minimized.e..5 2. needed for insight into chemical reactivity. Green’s function methods are often applied to stronglycorrelated systems. 5.478 14. Table 5. Exercise 22 Correlation energy Show that the correlation energy is never positive.23 22872. exactly.i [{φi }](r) = − d3 r φ∗ (r ) φi (r ) j φj (r). 2. i. Plot the correlation energy (a) across the ﬁrst row and (b) down the last column of the periodic table.045 0.39 0. and is deﬁned by fF. one adds many more terms to the approximate wavefunction..817 36 2752.845 54.138 86 22866.i [{φi }](r) = N j i φ(r) (5.324 0. most often density functional methods are used.g. They are usually about 2040 mH/electron.3. which are often relatively unchanged in a chemical reaction.89 3. .5 EC 0 0.937 527.433 4 14.409 10 128. in which the correlation contribution might be a much larger fraction.09 5. it is a diﬀerent function of r for each occupied orbital.e. r − r  (5. and wavefunction calculations. and correlation energies across ﬁrst row and for noble gas atoms. and their energetic separation. When is it zero? An interesting paradox to note in chemistry is that most modern chemists think of reactivity in terms of frontier orbitals. An example of this is the core electrons.125 0. Table 5. In order to discuss the pro’s and con’s of HF calculations. as it cancels out of energy diﬀferences. 1. By plotting ln(−E) versus ln(Z). A trial wavefunction is formed as a linear combintation of products of HF orbitals.904 7. the resulting thermochemistry is pretty bad.667 24. i.539 2753. These calculations are nonvariational. the HOMO (highest occupied molecular orbital) and the LUMO (lowest unoccupied). these are entirely constructs of the HF approximation.654 37.15) Note that this oddlooking animal is orbitaldependent. and may produce energies below the groundstate energy. e. Comment on the results. Quantum chemistry has developed many interesting ways in which to calculate the correlation energy. subject to the constraint of orthonormal orbitals. In electronic structure calculations of weaklycorrelated solids.547 18 526. although this approximation obviously contains basic chemical information.. including excited orbitals.862 3 7. 1.258 0.e. for just one electron. the Hartree and Fock terms in the potential cancel. those electrons in closed subshells with energies below the valence electrons. which treats the HartreeFock solution as the starting point. to obtain better energetics. and in which compensating errors might occur in the separate calculation of the molecule and the atoms. We see that correlation energies are a very small (but utterly vital) fraction of the total energy of systems.5.1: Total energies in HF. as in Fig.188 0.042 0. ﬁnd the dependence of the total energy on Z for large Z.3.589 75.3 Correlation The deﬁnition of correlation energy remains the same for N electrons as for two: It is the error made by a HartreeFock calculation. In the wavefunction approach. To ﬁnd the orbitals in HartreeFock.2 Hartrees in binding energies.74 Exercise 21 Atomic energies This exercise uses the numbers in Table 5. as they should. errors of less than 2 mH per bond. Also.401 8 74. losing the orbital description completely. between a molecule and its separated constituent atoms. but rather only energy diﬀerences.809 9 99. and performs perturbation theory in the Coulomb interaction.14) where the last contribution to the potential is due to the Fock operator. Another important point to note is that in fact. i.3 lists a few correlation energies for atoms. Perhaps the most common type of wavefunction calculation is conﬁguration interaction (CI).573 5 24. we must minimize the energy as a functional of each orbital φi (r).94 7235.156 0. Doing this.055 54 7232. 3.745 E 0. we must ﬁrst discuss the errors it makes. On the other hand.733 128. The ﬁrst is usually in the form of MollerPlesset perturbation theory. Thus.094 0..067 99.722 1. which makes errors of typically 0. A large energy error in their contribution is irrelevant. a result we will derive later. Repeat for the correlation energy. MANY ELECTRONS atom H He Li Be B C N O F Ne Ar Kr Xe Rn Z E HF 1 0.689 7 54.3. CORRELATION 49 50 CHAPTER 5. But we strive for chemical accuracy in our approximate solutions.
for example.13 contains 2 electrons. Thus each closed shell contains n−1 0 0. All atomic densities. although that’s never been generally proven. we get the basic idea. whose size is proportional to the nuclear charge: dn  = −2Z n(0) (5.. there are n values of l. and the positions of the HF orbitals largely determine the strucutre of the periodic table. (terms) can have the same conﬁguration.17) dr r=0 for a nucleus of charge Z at the origin. So it gets ﬁlled before the 3d. . The decay at large distances is far more interesting. For Hydrogenic atoms (N = 1).19) Since we will be constructing a formally exact theory of interacting quantum mechanics based on the oneelectron density. Furthermore. the next at 28. and. and we have nl . there are 2l + 1 values of ml = −l. it seems appropriate here to discuss and show some pictures of is ﬁrst ionization potential.r. In many cases. In Fig. Typically. When one coordinate in a wavefunction is taken to large distances from the nuclei. Our treatment will not require details beyond knowing the lowest conﬁguration. and so the transition metals appear a little late. and A is some constant. It is also true in our onedimensional examples with deltafunction external potentials. whose eigenvalue o is the diﬀerence in energies between the two systems: where α = √ n(r) = Ar β exp(−αr) 2I. as many diﬀerent angular momenta combinations. spherical densities are plotted multiplied by the phasespace factor 4πr 2 . As expected. The ﬁrst shell closes at N = 2. 2 2s electrons. Hund’s rules are used to choose which one has the lowest energy. the less negative the energy. This is especially true for the atoms. For each n.1. and 2 3s electrons. We take the Ar atom as an example.5 1 1. the 4s orbital dips below the 3d.722 contains 10. Thus the 2s orbitals lie lower than 2p. and 1. To understand these. and consider the hydrogenic levels. respectively. 0. and so forth. and so the behavior of the interacting system can be understood in terms of the HF orbitals. 5.18) (5. the nthshell contains 2n2 electrons. we must consider the actual HF orbitals. have been found to decay monotonically. and we already saw examples of this in earlier exercises. the total energy remains lower with the 4s ﬁlled. Since the HF potentials are not Coulombic. the power can also be deduced. there is an exact degeneracy between all orbitals of a given principal quantum number. the diﬀerent electronic shells are easily visible.5 Atomic densities at the origin.1: Radial density in Ar atom. for ﬁxed n. by showing how an orbital calculation can in principle yield the exact energetics.5. This eﬀect does not change the ﬁlling order described above. the true wavefunction has a strong overlap with the HF wavefunction. .4 Atomic conﬁgurations this quantity for atoms. A list of the ﬁlled orbitals is called the electronic conﬁguration of a system. Note that the peaks and dips in the radial density roughly correspond to these shells. o and causes a cusp in the density.13 contains 12. Then the 3d starts ﬁlling. The situation is similar for higher n. If we ﬁrst ignore interaction. We ﬁnd that the integral up to r = 0. This is Kato’s cusp condition. 5.4. both spin and orbital. 6 2p electrons.. Even when the 3d orbital energy drops below the 4s. In fact. This scheme works up to Ar.25 contains 4. and (r → ∞) (5. gn = l=0 (2l + 1) = n(n − 1) + n = n2 (5. Close to a nucleus. 0. up to 0. a useful and sensitive function of the density to plot for I = E(N − 1) − E(N ) . we plot its radial density. ranging from 0 to n − 1.l. the N electron groundstate wavefunction collapses to the product of the squareroot of the density times the (N −1)electron wavefunction. the higher l. Thus. We occupy the orbitals in order. This means the area under the curve is precisely N . but then fails to account for the transition metals.5 2 2.t. This still does not determine the groundstate entirely.. This means the squareroot of the density satisﬁes a Schr¨dingerlike equation. the groundstate conﬁguration of Ca is [Ar]4s2 . diﬀerent lvalues is lifted.. 5. the degeneracy w. ATOMIC CONFIGURATIONS 51 52 CHAPTER 5. the next at 10.16) Since each orbital can hold 2 electrons. the external potential dominates in the Schr¨dinger equation. which are not hydrogenic. These correspond to the 2 1s electrons. This integrates to 18 electrons. But for Z between 7 and 20.5 3 r Figure 5. β = (Z − N + 1)/α − 1. when plotted as function of r with the nucleus 25 Ar atom 20 4πr2 n(r) 15 10 5 0 We have seen how the HartreeFock equations produce a selfconsistent set of orbitals with orbital energies. MANY ELECTRONS We will see later how DFT resolves this paradox. For each l.
5. for an atom? 5.8 2 x r Figure 5. (b) in real He.20) dr since κ(0) = −Z. The groundstate density can always be reconstructed from ∞ dr 2κ(r )) (5. the exact cusp condition is satisﬁed.6. QUESTIONS 53 54 CHAPTER 5.6 Questions All the questions below are conceptual.2 1.9 2 0 0.9 2 0 2 4 He atom 6 8 10 exact HF κ 1. Speculate on why the correlation energy of Li is about the same as that of He.5 1. 1.4 0. Does each orbital in a HF calculation satisfy the same Schr¨dinger equation? o 2.8 1d He atom in HF 1.8 1 1.2. in HF and exactly 5. so that κ equals 2. Which is bigger. which was not so visible in the last ﬁgure. we have also plotted κ(r) for real He. (5. We can also clearly see the diﬀerence between the approximate and exact HF solutions here. while κ(r) → −α.6 1. both exact and approximate.6 0. MANY ELECTRONS spherical systems is d log(n(r)) κ(r) = (1/2) .21) n(r) = C exp( r 4.2 0. it tends to a smaller constant. by plotting κ(x) on the left of Fig. 5.4 1.6 1. On the right. we have already seen the cusp in the orbital for 1d hydrogenic atoms. the kinetic energy of the true wavefunction or that of the HF wavefunction. What do you expect happens to the correlation energy of the twoelectron ions as Z → ∞? 6. What is the relationship between E HF and N i=1 i ? 3. 1.4 1. (5. except for the details of the power law in front of the exponential in Eq. In particular.8 1. determined by the ionization potential. and see that the eﬀect of correlation on the density is extremely small: The HF and exact densities are very close. What do you expect happens to the correlation energy of the fourelectron ions as Z → ∞? where the constant is determined by normalization.6 exact approx 1.5 1.2: κ in (a) 1d He atom in HF approximation. Near the nucleus. these conditions remain true. .4 1. We may see dramatic evidence for these conditions in our HF solution of the 1d He atom.7 1. but at large distances.7 1.18). as r → ∞. In a onedimensional world with delta function interactions.3 1.3 1.
Part II Basics 55 .
. electrons.1 Ψ1 .1 (r). Minimizing this yields H[n] = E[n] − µ N VW But we can always swap which wavefunction we call 1 and which we call 2. we ﬁnd the Schr¨dinger equation for the density: o 2 − Exercise 24 Finding the potential from the density For the smoothed exponential. Ψ1 and Ψ2 . by the variational principle. All rights reserved. simple inversion of the Schr¨dinger equation. Using Lagrange multipliers. which reverses this inequality.3) 1 1 2 j=i ri − rj  (6.6) To illustrate the basic idea behind density functional theory. we see that µ = E.9) E = min F [n] + d3 r vext (r) n(r) . We know that the groundstate energy can be written as a density functional instead of as a wavefunction functional. It may be stated as: the groundstate density n(r) uniquely determines the potential. Consider all wavefunctions Ψ which yield a certain density n(r). 1  n2 + v(r) n(r) = µn(r) 8n2 (6. It is an example of the constrained search formalism. Since the integral of any Laplacian vanishes (with the given boundary conditions). this theorem is proven for densities with nondegenerate ground states. Exercise 23 Checking equation for density Show that the groundstate density for a particle in a box satisﬁes Eq. In the original HohenbergKohn paper.5) Ψ2 T + Vee + Vext. just o as we have done several times for the oneelectron case. this implies ˆ ˆ ˆ ˆ Ψ2 T + Vee Ψ2 ≥ Ψ1 T + Vee Ψ1 . for any n(r). which implies they are the same wavefunction by the variational principle and the assumption of nondegeneracy.4) (6.1 One electron 1 It is selfevident that the external potential in principle determines all the properties of the system: this is the normal approach to quantum mechanical problems. In our case. (6. (6. 6. up to an arbitrary constant.8) c 2000 by Kieron Burke. (6. and plot it. yielding the same density.58 CHAPTER 6. Then. n 57 . since the groundstate energy is simply (6. But since both wavefunctions have the same density. unless the total energies of the two wavefunctions are the same. Suppose there existed two potentials diﬀering by more than a constant. Deﬁne the functional ˆ ˆ F [n] = min Ψ T + Vee Ψ Ψ→n 4 + with boundary conditions that n(r) and n(r) → 0 at the edges. ﬁnd the potential for which this is an eigenstate. 5. which automatically includes degenerate states. These would have two diﬀerent groundstate wavefunctions. We must minimize it. This functional is 1  n2 E[n] = d3 r (6.4). We can identify the Lagrange multiplier by integrating both sides over all space at the solution.7) δH δT = + v(r) − µ = 0 δn(r) δn(r) Using the derivative from Ex. yields N i=1 (6. leading to a contradiction. where the search is over all antisymmetric wavefunctions yielding n(r). ˆ ˆ any wavefunction minimizing T + Vee is a groundstate wavefunction. The ﬁrst HohenbergKohn theorem demonstrates that the density may be used in place of the potential as the basic function uniquely characterizing the system. Then. Is it the ground state of this potential? An elegant constructive proof was found later by Levy. Consider Ψ2 as a trial wavefunction for potential vext. i. we ﬁrst formulate the problem of one electron as a density functional problem.1) + d3 r v(r) n(r) 8 n in any number of dimensions. we construct the auxillary functional: where N = d3 r n(r). DENSITY FUNCTIONAL THEORY 6. Then.2 HohenbergKohn theorems Chapter 6 Density functional theory This chapter deals with the foundation of modern density functional theory as an exact approach (in principle) to systems of interacting particles.1 Ψ2 ≥ Ψ1 T + Vee + Vext.e. and by contradiction.2) vext (ri ) = − 1 N 2 i=1 2 i Ψ/Ψ + This determines the potential up to a constant. ˆ ˆ ˆ ˆ ˆ ˆ (6. by solving the Schr¨dinger o equation for the eigenstates of the system. subject to the constraint that d3 r n(r) = 1. fermions interacting via the Coulomb repulsion. φ(x) = C(1 + x) exp(−x). The proof is elementary.
13) directly. We may also deﬁne the exact kinetic energy functional as ˆ T [n] = Ψ[n] T Ψ[n] . Note that it predates HartreeFock by three years. probably by a selfconsistent procedure.e. (6.8). consider some of our previous problems. The kinetic energy functional. Insertion of this approximate F into the EulerLagrange equation yields the ThomasFermi equation: 5 n(r ) AS n2/3 (r) + d3 r + vext (r) = µ (6. it is the same functional for all electronic structure problems. T [φ] = 1 dx φ (x)2 . as in chapter ??. and ﬁnd that. because of the singular Coulomb external potential. as fully as you can.7) to construct the unique corresponding potential. we approximated F [n] ≈ F HF [n] = 1 8 d3 r  n2 1 + n 4 d3 r d3 r n(r)n(r ) .9) with F HF [n] yields precisely the HartreeFock equations for the twoelectron problem.2. We can see immediately some problems. First. which could then be normalized and inserted back into the functional E[n]. We wish to mininize E[n] for a given vext (r) keeping the particle number ﬁxed. This is evident from Eq. when we treated the two electron case within the HartreeFock approximation. so that we could solve Eq. The exact density is such that it makes the functional derivative of F exactly equal to the Several points need to be clariﬁed. In the next section. the density is singular. since µ = ∂E/∂N . This gives us a verbal deﬁnition of the groundstate wavefunction. in the kinetic energy.13) We can identify the constant µ as the chemical potential of the system. (6. where the search is over all normalized positive densities.3 ThomasFermi In this section. Again. Similarly. Any such wavefunction can then be fed into Eq. (6. The original HK theorem was proven only for densities that were groundstate densities of some interacting electronic problem. and its limitations. we saw in Chapter ?? how the power of n can be deduced by dimensional analysis. An important formal question is that of vrepresentability.12) negative of the external potential (up to a constant). F [n] is approximated by the local approximation for the (noninteracting) kinetic energy of a uniform gas. As r → 0.6. i.14) Then minimization of the energy functional in Eq. Exercise 25 Errors in HartreeFock functional Comment. these expressions were developed for a spinunpolarized system. Second. plus the Hartree energy F TF [n] = AS d3 rn5/3 (r) + 1 2 d3 r d3 r n(r)n(r ) . because of the phasespace factor 4πr 2 ). which can easily be achieved numerically with a simple ordinary diﬀerential equation solver. The exact groundstate wavefunction of density n(r) is that wavefunction that yields n(r) and has minimizes T + V ee . The second HohenbergKohn theorem states that there is a single F [n] which is exact for all electronic problems. on the errors in F HF [n] relative to the exact F [n] for two electrons. the accuracy is insuﬃcient for modern chemistry and materials science. 6. we will examine the original crude attempt to do this (ThomasFermi theory). in a spinindependent external potential. DENSITY FUNCTIONAL THEORY from the variational principle. It would yield a single integrodiﬀerential equation to be solved.15) 3 r − r  We will focus on its solution for spherical atoms. They can also be derived by classical arguments applied to the electronic ﬂuid. . Note also that insertion of F HF [n] will yield an equation for the density equivalent to the orbital HF equation. we brieﬂy discuss the ﬁrst density functional theory (1927). We therefore minimize E[n] − µN . We will derive these numbers in detail in Chapter 8. The important point is that this single formula for F HF [n] is all that is needed for any twoelectron HartreeFock problem.8) by Ψ[n]. regardless of the particular problem being solved. as r → ∞. yielding AS = (3/10)(3π 2 )2/3 . The interested reader is referred to Dreizler and Gross for a thorough discussion. r − r  (6. The last part of the HohenbergKohn theorem is the EulerLagrange equation for the energy.. (6. Note that it would be marvellous if we could ﬁnd an adequate approximation to F for our purposes. HOHENBERGKOHN THEOREMS 59 60 CHAPTER 6. and this domain has been further extended to densities which result from ensembles of wavefunctions.10) The second HohenbergKohn theorem states that the functional F [n] is universal. and the exact electronelectron repulsion functional as ˆ Vee [n] = Ψ[n] Vee Ψ[n] . δn(r) (6. is the same 2 functional for all oneelectron problems. it is the same for that orbital. to recover the groundstate energy. although the overall trends are sound. one with equal numbers of up and down spin electrons. while the coeﬃcient is chosen to agree with that of a uniform gas.. When we evaluate the kinetic energy for a given trial orbital. r − r  (6. The constrained search formulation extended this to any densities extracted from a single wavefunction. Recall our orbital treatment of oneelectron problems. and ﬁnd the EulerLagrange equation: δF + vext (r) = µ. n → 1/r 3/2 (although still integrable. its successes. (6. We denote the minimizing wavefunction in Eq. In the ThomasFermi theory. which contains no mention of the external potential. (6.11) (6.e. for the density. To understand the content of this statement. i.
They highlight more clearly the relative .17) Z 2 Φ 3/2 (6. Figure 6. Eq. By using the virial theorem.21) 3.3612163121.1: Radial density in the Xe atom.4 1. and the asymptotic behavior is the trivial solution: 144 x → ∞. But overall. The ThomasFermi equation for neutral densities. 6.16) 2. and comment.79374 Z 7/3 a (6.6 1.885341.22) √ where y = x and α = 0. (6. show that 3 B 7/3 E=− Z = −. details (nuclear cusp. and γ = 0. compare with energies along ﬁrst row and for the noble gas atoms. both exactly and in ThomasFermi calculation of Exercise ??. decay at large distances). Exercise 26 ThomasFermi for atoms 1.2: κ(r) for the Ar atom. Do atomic radii grow or shrink with Z? Explain why. β = −0. x These results.5430794693. where B = 1. Plot the approximate TF density on the same ﬁgure. ﬁnd the TF kinetic energy. To see the quality of our TF solution.4 Particles in boxes Before closing this chapter. although it messes up all the 0 where Φ(0) = 1 and Φ(∞) = 0.3. An accurate radial density for the Xe atom is given in Fig.15). Show that the external potential goes as Vext = − B 7/3 Z = 1. 6. Calculate the expectation value 1/r for ThomasFermi.7280642371.8 2 20 r Figure 6. Using table 5. We see that. (6. are needed for the next exercise.6.5 2 2. not exponentially. Φ (0) = −B. DENSITY FUNCTIONAL THEORY the density decays with a power law. 6. THOMASFERMI 61 62 CHAPTER 6.6 0. and compare with that of the exact solution.5 1 1. Comment. 5. is usually given in term of a dimensionless function Φ(x) satisfying: Φ (x) = Φ3 x (6.4 0. Lastly.19) Φ(x) = 3 . Teller showed that molecules do not bind in TF theory. deduce the behavior of the Hartree energy as Z → ∞.2 0. and the density of Fig.20) 7 a This is the exact limiting behavior for nonrelativistic atoms as Z → ∞. where 1 a= 2 and the density is given by 3π 4 2/3 = 0.5 3 0 0 0. The distance x = Z 1/3 r/a. n(r) = 90 80 70 5 Xe atom in LDA 4πr2 n(r) 60 50 40 30 20 10 10 Kappa in Ar atom 15 exact approx TF 0 0.1. 6.18) 4πa3 x Its been found that at the solution for neutral atoms.7845 Z 7/3 (6. 4. we calculate the κ(r) function of the last chapter. Using the virial theorem for atoms. An accurate parametrization of Φ is Φ(y) = (1 + α y + β y 2 exp(−γ y))2 exp(−2α y) (6. because the errors in TF are far larger than the relatively small binding energies of molecules. calculated using LDA.58807102261. it has a remarkably good overall shape for such a simple theory.8 1 1.2 1. let us return to the examples of the introduction and understand how they are connected to the ThomasFermi model.3. the trends will be approximately right. shell structure.1. Using the diﬀerential equation and integrating by parts. (6.
This is because.5. for the particles in a box. 6. Also. we cheated slightly. the selfconsistent solution is too small by a factor of 3! Compare this with the 25% error using the exact density. n(x) = n = Exercise 27 Energy from local approximation on exact density in box A ‘fun’ math problem is to show that. DENSITY FUNCTIONAL THEORY advantages and disadvantages of local approximations.25) 1. the energy contains only one term. Deﬁne the groundstate wavefunction generating density n(r) without mentioning the external potential.5 Questions (6. where would we have gotten the exact density? A more honest calculation is to ﬁnd the selfconsistent density within the given approximation. for one electron.23) π wherever the argument of the root is positive. Exercise 28 Large N for harmonic oscilator Calculate the selfconsistent density of the local approximation for the harmonic oscillator. although the local approximation is good for the energy. In particular. These are all conceptual. yielding N (6. can more than one potential have the same groundstate density? 2. zero otherwise. Using exactly this density. E = π 2 N 3 /(6L2 ). by applying the local approximation to the exact density. and µ is to be found by normalization ( dx n(x) = N . Why do the HK theorems specify the groundstate and not. Note that our reasoning applies to all bounded 1d problems. But if we’re making the local approximation for the energy. Compare with your results for the HO problem there. ﬁnd the density for the onedimensional harmonic oscillator. its derivative is not so good. and compare with your exact densities. using the local approximation evaluated on the exact density for a particle in a box. exactly what do we mean? After all. and evaluate the total energy on that density. For the particle in the box. Repeating the arguments of section 6.6. not just particles in boxes. the asymptotic one.24) L This is quite diﬀerent from the exact density for one electron. Using the local approximation for the onedimensional kinetic energy developed in Chapter 1. and the density is quite sensitive to this. we ﬁnd 1 n(x) = 2(µ − vext (x)) (6. When we say F [n] is a universal functional. How would you go about ﬁnding the potential from the wavefunction of a twoelectron system? For two electrons. since they apply in one dimension (no shell structure) and without interaction. v(x) vanishes inside the box.1. the energy can be found analytically to be π2 3 E 9 = N2 + N + N 6 8 8 You can check how it compares to your answers to Exercise X. . but the true density approaches it as N gets large. since vext (r) itself is a functional of n(r). compare the energies. 3. First we note that. QUESTIONS 63 64 CHAPTER 6. say. the ﬁrst excited state? 5. is not the entire energy universal? 4.
which allow almost all the kinetic energy to be calculated exactly. since for Coulombic systems. The KohnSham wavefunction of density n(r) is that waveˆ function that yields n(r) and has least kinetic energy. 7.3) δn(r) where ˆ TS [n] = min Φ T Φ . and yield the exact density.4) by Φ[n]. especially for large numbers of electrons. because solution of the KohnSham equations produces the exact noninteracting kinetic energy. KOHNSHAM Chapter 7 is the kinetic energy of noninteracting electrons.6) This is the ﬁrst and most important relationship of exact density functional theory: from the functional dependence of F [n]. that the KS system is simply a ﬁctitious system of noninteracting electrons. We now have the theoretical tools to immediately write down these KS equations. We note several important points: • The KohnSham equations are exact. Obviously T S [n] = Φ[n] T Φ[n] .. we ﬁnd vS (r) = vext (r) + d3 r n(r) + vXC [n](r) r − r  vXC (r) = δEXC δn(r) (7. We have implicitly assumed that the KohnSham wavefunction is a single Slater determinant. or else we would have solved all Coulombinteracting electronic problems exactly.e. Then its orbitals are given by Eq. and we know it explictly as a density functional. Exercise 29 Kinetic energies Show that T [n] ≥ TS [n].3).1) 2 and yield N We have explicitly included the Hartree energy. Recall from the introduction.66 CHAPTER 7. This gives us a verbal deﬁnition of the KohnSham wavefunction.2) The subscript s denotes singleelectron equations.13) and comparing with Eq.4) Φ→n 65 . (7. n(r) = i=1 φi (r)2 . In common practice. the unknown exchangecorrelation energy must be approximated. establishing modern density functional theory as a useful tool. (7. comes with the introduction of the KohnSham equations. The rest is called the exchangecorrelation energy. o However. write the groundstate functional of an interacting system in terms of the noninteracting kinetic energy: F [n] = TS [n] + U [n] + EXC [n]. and so are much easier to solve than the coupled Schr¨dinger equation. chosen to have the same density as the physical system. • The KohnSham equations are a set of singleparticle equations.5). as we know this will be a large part of the remainder. A major breakthrough in this area is provided by the KohnSham construction of noninteracting electrons with the same density as the physical system. the kinetic energy equals the absolute value of the total energy. this has never been a problem. There is nothing approximate about this. (We do not know this functional exactly. Inserting F [n] into Eq. But the Euler equation that is equivalent to these equations is δTS + vS (r) = µ. which diﬀers from T [n]. they would still be too large. which includes almost all the true kinetic energy. we can extract the potential felt by noninteracting electrons of the same density. (7. We saw in chapter one that local approximations to the kinetic energy are typically good to within 1020%. there are known examples where such a potential cannot be found. (7. What is the relation between T HF [n] and TS [n]? Is there a simple relation between T [n] and T HF [n]? Now. 1. (7. in return. (1. • The great advantage of the KS equations over ThomasFermi theory is that almost all the kinetic energy (TS ) is treated exactly. (7. 1 − 2 + vS (r) φi (r) = i φi (r). (6. For every physical system. Even if we could reduce errors to 1%. the KohnSham alter ego is welldeﬁned and unique (recall Fig. (7.1 KohnSham equations The major error in the ThomasFermi approach comes from approximating the kinetic energy as a density functional.2). We denote the minimizing wavefunction in Eq. errors of 10% are huge. i.5) KohnSham The most important step.) • While the KS potential is unique by the HohenbergKohn theorem (applied to noninteracting electrons). However. which is true most of the time.
1.2 tell us nothing about the functional dependence of o vS (r). we then deﬁne the correlation energy as the remaining unknown piece of the energy: EC [n] = F [n] − TS [n] − U [n] − EX [n]. There are many points to note in this table: . With exact KS theory. Fig.e. Once the KS potential and its orbitals are found.2. Then changing the potential to shift the computed density toward the exact one. The remaining diﬀerence is due to the local potential for the KS orbitals. That KS potential was found simply by inverting the noninteracting Schr¨dinger equation for a single orbital.904 6. The KS exchange is deﬁned for a given density.169 0. UC [n] = Vee [n] − U [n] − EX [n].72 0. 7. the electronelectron repulsion. For many systems. Inserting the deﬁnition of F above. (7. what is the exchange energy then? Finally. Also. but only as a functional of the orbitals. and say which is bigger: EC or EC . but that are eigenstates of a single potential (not orbitaldependent). • By subtracting TS and U from F .375 14. KOHNSHAM • The KS orbitals supercede the HF orbitals. in providing an exact molecular orbital theory. we ﬁnd that the correlation energy consists of two separate contributions: EC [n] = TC [n] + UC [n] where TC is the kinetic contribution to the correlation energy. and repeating until converged.096 Ne 128. in that they are the orbitals that yield a given density. we see now how an orbital calculation can provide exact energetics.037 0. a highly accurate solution was found of the full Schr¨dinger equation for each atom.09 0. Exercise 33 Correlation energy trad Prove EC ≤ 0.33 0.7. and comparing with the exact density.94 128. Hartree.594 7.12 53.753 0.8) In density functional theory. These diﬀer from the HartreeFock orbitals. 1. From o it.10) i. EXCHANGE 67 68 CHAPTER 7.24 128. In density functional theory. • Note that pictures like that of Fig. solving for its density. evaluated on the KohnSham wavefunction.94 311. it is straightforward to evaluate its kinetic.7) (7.073 0. Next.. and exchange energies.3 we give accurate values for three noble gas atoms. The HF orbitals are better thought of as approximations to exact KS orbitals. Eq. To eliminate the density diﬀerence. yields a direct contribution (the Hartree piece) and an exchange contribution. TC ∼ −EC ∼ −UC /2.3 Correlation We will show that it usually better to approximate exchangecorrelation together as a single entity. Note that Eq. These numbers were found as follows.218 2. the groundstate density was extracted from the wavefunction. 7. (or the correlation contribution to kinetic energy). The diﬀerences between HF exchange and KSDFT exchange are subtle.904 2.11) Exercise 32 Correlation energy from Hamiltonian Deﬁne the DFT correlation energy in terms of the Hamiltonian evaluated on wavefunctions. and UC is the potential contribution to the correlation energy. The last set of columns were found by subtracting KS quantities from their exact counterparts.946 2.025 0. If I construct a spin singlet.710 4. and other orbital pictures. So the standard ﬁgure.079 0. (5. in Table 7. the total groundstate energy and its various components could be calculated.867 2.667 14.13) does not give us the exchange energy as an explicit functional of the density. the KohnSham wavefunction is simply a single Slater determinant of orbitals.049 1. Exercise 30 HF versus DFT exchange Prove E HF ≤ TS [nHF ] + U [nHF ] + Vext [nHF ] + EX [nHF ] for any problem. The total energy in a HartreeFock calculation is extremely close to TS + U + Vext + EX . (7.9) (7. we can compare KS EX [nHF ] with that from HF. found from the exact densities. They can be thought of as having two diﬀerent sources. but tells us nothing about how that potential would change with the density.61 66.05 12. 2. we deﬁne the exchange energy as ˆ EX [n] = Φ[n] Vee Φ[n] − U [n]. to get an idea of how large these energies are for real systems. with both electrons in an exponential orbital.13) of the KS orbitals. TC [n] = T [n] − TS [n] Exercise 31 Exchange energies Calculate the exchange energy for the hydrogen atom and for the threedimensional harmonic oscillator. The HF exchange is evaluated on the HF orbitals for the system. First. (5. so that EX is given by the Fock integral.2 Exchange atom E T Vext Vee TS U EX TC UC EC He 2.667 33.3.1: Energy components for ﬁrst three noble gas atoms. what’s left (exchange and correlation) will turn out to be very amenable to localtype approximations. In most cases. This can be thought of as guessing the potential.042 Be 14. This gives us the exact KS potential for this system. rather than exchange and correlation separately. a search was made for the unique KS potential that corresponded to that density. and so the exact exchange of a system is the exchange of the KS orbitals evaluated on the exact density. are much more usefully interpreted as pictures of KS atomic and molecular orbitals. 1. (7.39 Table 7.674 0.
• The interaction with the nucleus. All questions are conceptual. Deﬁne the groundstate KohnSham wavefunction generating density n(r). • The potential contribution to the correlation energy. is a little more than −2TC . φ1 (x) and φ2 (x). Considering the exact relations for the asymptotic decay of the density of Coulombic systems. What is the ratio TC versus UC ? As a function of Z? Comment. 7. How close are the ratios of T to Vee to Vext to their ThomasFermi ideal values? What is the trend with Z? Comment. Exercise 34 Energy components for atoms Use the data in Table 7. which is also the correlation contribution to the kinetic energy. Repeat above for TC versus TS . We end with an exercise designed to make you think about the trends. but all three are on the same scale. 3. and the density n = φ1 2 + φ2 2 . 2.3 to answer the following: 1.7. 2. by the virial theorem for atoms. is there any signiﬁcance to HOMO for the KS system? (See section 14.4. 4. • The electronelectron repulsion is typically less than half the kinetic energy. Consider two normalized orbitals in one dimension. • The Hartree energy is typically a small overestimate of Vee . QUESTIONS 69 70 CHAPTER 7. is on the same scale as −EC . is a little more than twice the kinetic energy. KOHNSHAM • The exact kinetic energy is simply E. What is the ratio of EX  to U ? How does it change with Z? Comment. and they diﬀer by less than 2%. Comment. The exchange energy of the orbitals is the same in HartreeFock and KohnSham theory. as it must also overcome the (positive) Vee . Recall from ThomasFermi theory that Vext = −7T /3 = −7Vee . What is the relation between the two for the HartreeFock density? 4.2). Vext . . • The kinetic contribution to the correlation energy.4 Questions 3. which is also the correlation contribution to the potential is energy. • The magnitude of all energy components grows with Z. • The exchange energy cancels a fraction of the Hartree energy. What is the KohnSham kinetic energy of that density? How does it change when we alter one of the orbitals? Repeat the question for the KohnSham exchange energy. that fraction getting smaller as Z increases. • The KS kinetic energy TS is almost as large as the true kinetic energy. 1.
called the Fermi wavevector. we ﬁrst consider a smoother class of curves. To illustrate. r(θ) = r0 (1 + cos(nθ)). we wish to approximate the perimeter with a local functional of r. To study the problem. namely a uniform gas of electrons.1. etc. such as P loc ≤ P . we look to the uniform electron gas.4) where n = 1. while n is a measure of how rapidly the radius changes with angle. we note that the perimeter of a circle is 2πr. This yielded as = π 2 /6. Nπ kF = = nπ (1D polarized) (8. namely the local approximation. 1 c 2000 by Kieron Burke. r = 1 + 0. The only place our approximation can be exact is for the circle itself.3) is the local approximation to the perimeter. Our more advanced mathematicians might be able to prove simple rules.1 circle Next we deduced that ts .1 Local approximations 1 dθ r(θ) (8. 3. the noninteracting kinetic energy of samespin electrons. To determine this function. 8. All rights reserved.1 cos(20θ). i. must be proportional to n3 . Thus P loc = 2π 0 The local density approximation This chapter deals with the mother of all density functional approximations. This is the wavevector of the highest occupied orbital in our system. the kinetic energy density. where f (n) is some function of n. THE LOCAL DENSITY APPROXIMATION Now. for 3D Coulombinteracting problems. eps=0. 2. n=20. we introduce a useful concept. 8. so we write 2π dθ f (r) (8. We ﬁrst write TSloc [n] = ∞ −∞ dx ts (n(x)) (8. in ﬂatland..1: Smooth parametrized curve.7) TSloc = Thus. as a warmup exercise.e. 71 . Since we have N electrons in the box. is chosen as an integer to preserve periodicity. for exchange and correlation. since its r is independent of θ.8) Figure 8. 8. from dimensional analysis. we can test this approximation on √ the squares we really care about. Recall the example of Chapter 1.1 the case where n = 20 and = 0. Suppose we live in a ﬂatland in which it is very important to estimate the perimeters of curves. ts (n) = as n3 (8.6) L Then ∞ 2 dx n(x) τS (n(x)). P loc = 4 ln(1 + 2). for the unit square.72 CHAPTER 8. a 12% underestimate. but we haven’t enough calculus to ﬁgure out the exact formula..2 Local density approximation To illustrate why a local approximation is the ﬁrst thing to try for a functional. (8. We did this by looking at the leading contribution to the energy of the electrons as the number became large. Also. we show in Fig. we deduced as by ensuring TSloc is exact for the only system for which it can be exact. we write loc EXC [n] = −∞ smooth curves d3 r f (n(r)) (8. Thus we should choose f = r if we want our functional to be correct whenever it can be. just like the kinetic energy approximation from Chapter 1. and ﬁnd. At this point.2) P loc = 0 Chapter 8 To determine the function f .5) Last.1) So now its simple to see how to construct local density approximations. we return to the problem in Chapter 2 of the perimeter of a curve r(θ). Here determines how large the deviation from a circle is.. τS (n) = kF /6 (8.
leading to LDA EX [n] = AX d3 r n4/3 (r) (8. we ﬁnd that it is simpler to use planewaves with periodic boundary conditions.047 + 0. This is due to the longranged nature of the Coulomb repulsion in an inﬁnite system. Note that this means correlation unif becomes (almost) as large as exchange here. If we repeat the exercise above in three dimensions. 15. A simple way to think about correlation is simply as an enhancement over exchange: unif XC (rs ) = FXC (rs ) unif X (rs ) (8. 4π (8. being proportional to 1/rs .15) and this enhancement factor is plotted in Fig.11) Evaluation of the Fock integral Eq.017rs (rs → 0) (8.15 0. Thus rS → 0 is the highdensity limit.3 Uniform electron gas We thus write unif The local approximation is exact for the special case of a uniform electronic system.. chosen to preserve overall charge neutrality. but matching to the uniform gas yields a diﬀerent constant: 3 2 TSTF [n] = d3 r kF (r)n(r) = AS d3 r n5/3 (r) (3D unpolarized) (8. and FX = 1. leading to unif C where the diﬀerent power and coeﬃcient come from the diﬀerent dimensionality. (rs ) → − 0 0. since the KohnSham wavefunctions are simply Slater determinants of plane waves.13) 4πn kF which is the radius of a sphere around electron such that the volume of all the spheres matches the total density of electrons.8. Correlation is far more sophisticated. There are several important features to the curve: • At rs = 0 (inﬁnite density).17) rs r s where the constants are d0 = 0.009rs ln rs − 0. The exchange is very simple.9) LDA EC = d3 r n(r) unif C (rs (r)) (8.. The sharp downturn .1.14) where C (rs ) is the correlation energy per electron of the uniform gas. Note that we have now doublyoccupied each orbital. Dimensional analysis produces the same form for the kinetic energy as in 1D. one in which the electrons sit in an inﬁnite region of space.25 0. The uniform gas exchange and correlation energies/particle are plotted in Fig 8.3.0311 ln rs − 0.325.. and so FXC (rs → ∞) = 1. • In the large rs limit (lowdensity) unif C d0 d1 + 3/2 − . and the lowest occupied levels form a sphere in momentumspace. • As rs → 0 (high density). (5. The kinetic and exchange energies of such a system are easily evaluated. Its radius is the Fermi wavevector. as a function of rs .12) so that AX = −(3/4)(3/π)1/3 = −0.10) 10 where As = (3/10)(3π 2 )2/3 = 2. as it depends explictly on the physical groundstate wavefunction of the uniform gas. (8..896 and d1 = 1.919 rs = = (8. Note also that the approach to the lowdensity limit is extremely slow.3 0 1 2 uniform gas C X 3 4 5 6 (n) = 3kF . UNIFORM ELECTRON GAS 73 74 CHAPTER 8.738.2: Exchange and correlation energies per particle for a uniform electron gas. Thus we know that its energy density must be proportional to kF . with a uniform positive external potential. exchange dominates over correlation. Figure 8. to account for spin.05 0.2. combined with known exact limiting values. The constant d0 was ﬁrst deduced by Wigner from the Wigner crystal for this system.16) The logarithmic divergence is not enough to make correlation as large as exchange in this limit. and rs → ∞ is the low density limit. and is given by kF = (3π 2 n)1/3 . (rs → ∞) (8.2 0. For the exchange energy of the uniform gas. i.1 0.871. Another useful measure of the density is the WignerSeitz radius 3 1/3 1. but we will see later that ﬁnite systems have ﬁnite correlation energy in this limit. we simply note that the Coulomb interaction has dimensions of inverse length. The correlation energy is extracted from accurate Monte Carlo calculations.13) for a Slater determinant of planewave orbitals yields the exchange energy per electron of a uniform gas as unif X (rs ) → 0. there is a sharp dive toward 1.896. The states are then ordered energetically by momentum. THE LOCAL DENSITY APPROXIMATION 8.e.
4 Questions 1. calculate (a) the Hartree potential and (b) the Hartree energy. i. An early but inaccurate extrapolation from the lowdensity gas. can you deduce the asymptotic form of the exchangecorrelation potential.18) 8.e.44 abd b = 7.2). which is diverging as 1/rs ). For example. and correlation is much smaller than exchange. using the exponential approximation of Chapter 4.. vXC (r) as r → ∞? Would a local approximation capture this behavior? (See section 14. while the parametrization of VoskoWilkesNusair (VWN) has been implemented in quantum chemical codes.3. many (if not most) practical problems require this. with accurate quantum Monte Carlo data for the uniform gas. 3. This gives a ballpark number. More recently. For typical valence electrons. using the exponential approximation of Chapter X. An early popular formula in solidstate physics is PZ81. Note that in the Gaussian codes. we must postpone a general review of the performance of LDA until after the next chapter. this function has become very wellknown. since it makes no distinction between polarized and unpolarized densities. was given by Wigner: eC (rs ) = −a/(b + rs ) (W igner) (8. Exercise 35 Hartree energy for an exponential 1. and compare with exact answer. but misses the logarithmic singularity as rs → 0. while VWNV is the actual parametrization recommended by VWN. is what is actually used in practice. 2. We emphasize this fact in the next few exercises. and not accounted for spin eﬀects. called the local spindensity approximation. Exercise 36 LDA exchange energy for 1 or 2 electrons 1. Perdew and Wang reparametrized the data. . VWN refers to an older formula from the VWN paper. Find the exact exchange energy for the Helium atom. Over the years. Unfortunately. THE LOCAL DENSITY APPROXIMATION in the correlation as rs → 0 is due to the logarithmic term (but is not as singular as exchange. and compare with exact answer. In fact. rs is between 1 and 6. but rs must be larger than shown here. A slight generalization of LDA. Following the previous question. where a = −. the two curves become comparable. by combining limiting information like that above. These accurate parametrizations diﬀer only very slightly among themselves. Find the exact exchange energy for the Hydrogen atom. Calculate the LDA exchange energy for the Hydrogen atom. At large rs . n(r) = Z 3 /π exp(−2Zr). UNIFORM ELECTRON GAS 75 76 CHAPTER 8. pure LDA performs badly whenever there is an odd number of electrons. 2. Find the LDA exchange energy for the Helium atom.8. For an exponential density.8. This is because we have only developed density functional theory so far.
σ. 0] = TSunpol [2n]/2. it allows access to magnetic response properties of a system. we ﬁnd (9. .1.78 CHAPTER 9. so long as it does not produce an orbital whose energy is below 1/2. .4) This clearly yields a consistent answer for an unpolarized system. 1 This i. By orbital functionals.σ (r).6 1. the KohnSham equations can be developed with spindependent KohnSham potentials. we mean those that are a sum of contributions from the orbitals of each spin seperately. Even when the external potential is not spindependent. the spin down potential is undeﬁned. people don’t use KohnSham density functional theory. 9. We consider the up and down densities as separate variables. 3. 9. and gives the result for a fullypolarized system: TSpol [n] = TSunpol [2n]/2 (9. deﬁned as nσ (r) = N dx2 .2 0 0 1 2 3 4 5 6 7 8 Figure 9. We will see in the next sections that it is straightforward to turn some density functionals into spindensity functionals. while little is known of the exact spinscaling of correlation in general. Applying this to the spinunpolarized case.7) 77 . . Even if not interested in magnetism. For example. But since the kinetic energy is the sum of contributions from the two spin channels: TS [n↑ .2 Li atom up down Chapter 9 1 0. Systems in collinear magnetic ﬁelds are included. 9. for a spinup H atom. separate treatment of the up and down spin densities yields much better results. n↓ ] = and pol unpol EX [n] = EX [2n]/2 1 unpol unpol E [2n↑ ] + EX [2n↓ ] 2 X (9. n↓ ] (9. We do spinscaling of the kinetic energy functional as an example.. we will see that the increased freedom in spin DFT leads to more accurate functional approximations for systems that are spinpolarized. Let’s denote it TSunpol [n] to be very explicit. a simple generalization of density functional theory. There is a little more wriggle room than in the DFT case.1: Radial spin densities in the Li atom. x2 .6) (9.2). 2. the Li atom. 0] + TS [0. 1 Similarly.e. xN )2 .g. n/2] = 2TS [n/2. the contributions come from each spin separately. dxN Ψ(r.8 0. . For many problems. .3) TSunpol [n] = TS [n/2. showing a onetoone correspondence between spin densities and spindependent external potentials. n↓ ] = TS [n↑ .2 Spin scaling We next introduce spin density functional theory.2) with the interpretation that nσ (r)d3 r is the probability for ﬁnding an electron of spin σ in d3 r around r.6 Spin In practice. This has several advantages: 1. n↓ ] = 1 unpol T [2n↑ ] + TSunpol [2n↓ ] 2 S (9. e. they use KohnSham spindensity functional theory.4 0. we ﬁnd TS [n↑ . Inserting this result back into Eq. while others are more complicated.4 1. v ext.1 KohnSham equations 0. (9. is not quite true. . (9.1) We can easily deduce the spin scaling of orbital functionals. Then the HohenbergKohn theorems can be proved. All modern density functional calculations are in fact spindensity functional calculations.. 0] or TS [n. SPIN 1. whose spin densities are plotted in Fig. Suppose we know T S [n] as a density functional for spinunpolarized systems.5) Analogous formulas apply to exchange: EX [n↑ .
(9. and we merely say that there exists wellknown parametrizations of the uniform gas correlation energy as a function of spin polarization. as shown in Fig. must be extracted from QMC and known limits.2: Eﬀect of (full) spin polarization on exchange and correlation energies of the uniform electron gas. Now we introduce a new.9).3 0 1 2 uniform gas pol unpol 3 4 5 6 C X The local spin density (LSD) approximation is the spinscaled generalization of LDA. we show what happens. useful concept. Test it on the hydrogen atom. the case for correlation is much more complicated. d3 r n4/3 (r) Exercise 37 Relative spin polarization Sketch ζ(r) for the Li atom. 1. reducing the correlation energy. This eﬀect is typical of most systems. (9. Exercise 38 Spin polarized ThomasFermi Convert the ThomasFermi local approximation for the kinetic energy in 3D to a spindensity functional. n↓ ](r). so X eX (n↑ . Deduce the formula of TSloc for unpolarized electrons in a onedimensional box. n↓ ] = AX n↓ = (1 − ζ) n/2. Why does LSD yield a more accurate energy for the Hydrogen atom than LDA? 3. When the gas is polarized.1 0. 2. 9.2. we ﬁnd. exchange keeps the electrons further apart. The formula for the exchange energy is straightforward. 9.25 0. we need the energies per particle of a spinpolarized uniform gas. that of relative spin polarization. Thus ζ = 0 for an unpolarized system.8) X 2 X Now we know eunpol (n) = AX n4/3 . n↓ ) = 21/3 AX n↑ + n↓ 4/3 4/3 (9. . On the other hand.12) (1 + ζ(r))4/3 + (1 − ζ(r))4/3 (9.9. as a function of spinpolarization. In terms of ζ. In Fig. State in words the relation between vS [n](r) and vSσ [n↑ . When do they coincide? Inserting these deﬁnition into Eq.3 LSD 0 0. LSD 79 80 CHAPTER 9.15 0. n↓ ) = e (2n↑ ) + eunpol (2n↓ ) (9. unif (n. n↑ = (1 + ζ) n/2.3.1. but this has not been proven generally. ζ) = X 2 Thus our LSD exchange energy formula is LSD EX [n↑ . In terms of the energy density of a uniform gas. and ±1 for a fully polarized system. and compare to the regular TF.14) n(r) Unfortunately. C exchange gets stronger. As input.2 0.11) (9.4 Figure 9.9) 9. for the energy density: (1 + ζ)4/3 + (1 − ζ)4/3 (n) X (n.10) n as the ratio of the spin diﬀerence density to the total density. but the correlation energy. Deﬁne n↑ − n ↓ ζ= (9. Begin with exchange. ζ). Eq.13) 2 where ζ(r) is the local relative spin polarization: n↑ (r) − n↓ (r) ζ(r) = (9. because electrons exchange only with like spins. unpol Questions All are conceptual. for the uniform gas. (9. SPIN 9.05 0.6) becomes: 1 unpol eX (n↑ .
This determines the scale of the ﬁgure.770 8 % Ne 11. For example. The KohnSham potentials reﬂect this similarity. the total exchange energy is typically underestimated by about 10%. the correlation energy is overestimated by about a factor of 2 or 3.996 1.883 1.5 2 r Figure 10. So for all the occupied region. Again note the scale. calculate the LDA correlation energy using the Wigner approximation. note that they are dominated by terms that are (essentially) identical in the two cases.3. r = 2. However. but 31.0 in LDA. 10. However.2: Logarithmic derivative of density for Ne atom.1: Accuracy of LDA for exchange. which were often called ab initio in solid state physics. and XC for noble gas atoms. since the densities are almost identical. and the net exchangecorrelation energy is typically underestimated by about 7%.2 10. The curves are indistinguishable close to the nucleus. . and how close they appear. as deﬁned earlier.085 9 % 0. which is −10/r. plotted in Fig. and (the spindependent generalization) has been one of the most succesful approximations ever. exchange is about 4 times bigger than correlation. and all integrals over the density are very accurate.9 at about 1 1. The largest term is the external potential.4. 1/r is 31. being hard to distinguish by eye. where they both reach 10.393 89 % 11.021 12. we can Chapter 10 4πr2 n(r) 12 10 Ne atom exact LDA Properties In this chapter.067 7 % Be 2. both exactly and in LDA.8 2 Ne atom exact LDA 0 1 2 3 4 5 For atoms and molecules. 10. On the other hand. when we calculate κ(r).025 14 % 0.2 Densities and potentials The selfconsistent LSD density is always close to the exact density. correlation.18).1 Total energies κ(r) 1.1: Radial density of the Ne atom. we survey some of the more basic properties that are presently being calculated with electronic structure methods. This is essentially the same in both cases. Until the early 90’s.763 12. 10. it was the standard approach for all density functional calculations. even κ(r) is extremely accurate in LDA. Since for many systems of physical and chemical interest.225 0. In Fig.1 exactly. The next term is the Hartree potential. 8 6 4 2 0 0 0.6 1.112 0.742 0. as seen in Fig.82 CHAPTER 10. The local density approximation to exchangecorrelation was introduced by Kohn and Sham in 1965. Eq. Exercise 39 Wigner approximation for He atom correlation energy: By approximating the density of the He atom as a simple exponential with Zef f = 7/4 as we found in Chapter 4. It remains perhaps the most reliable approximation we have. (8.096 134 % 2.1. and include how well or badly LDA does for these.042 167 % 0. EX EC EXC atom LDA exact error % LDA exact error % LDA exact error % He 0. we show both the LDA and exact potentials.4 1.674 13 % 0. 81 r Figure 10. PROPERTIES 10.68. the overestimate of correlation compliments the underestimate of exchange.5 1 1.478 6 % Table 10.546 2. both exactly and from an LDA calculation see the diﬀerence in the asymptotic region: Note that the density integrates to 9.321 2.
04 9.62 17. as shown in Fig.07 5.3 Ionization energies and electron aﬃnities atom h he li be b c n o f ne na mg al si p s cl ar LSDX 12. nonrelativistic.30 11.00 24.02 8.63 7. Thus ionization potentials from orbital energy diﬀerences are very poor.32 8.2: Ionization energies of ﬁrst twenty atoms.52 13.52 10.79 11.39 9. 10.35 4.81 4.5 1 1.42 X 13.99 8. in eV.26 10.34 8. looks very diﬀerent from the exact quantities for any ﬁnite systems.40 22.04 7.77 14.5 3 3.3: KohnSham potential of Ne atom.63 8.91 11.9 eV.5 2 2.78 16.14 7.66 19. There are many things to see in this table: • Comparing the exchange results (which are essentially identical to HF) with the exact results. Now we must . using exactexchange densities.99 8.5.70 5.45 5.2 0.93 10.85 17.84 Table 10. both exactly and in LDA.76 LSD 13.8 1 r Figure 10. The ’exact’ results are from Davidson.32 7.81 14.40 9.37 12.46 9.4 Ne atom exact LDA 0. we list ionization potentials for the ﬁrst twenty atoms. 40 35 30 Ne atom exact LDA vH (r) 25 20 15 10 5 0 0 0. and the LSD results are evaluated on those densities.28 5.62 5. however.5 4 r Figure 10.00 11.48 5. we ﬁnd that HF underestimates ionization potentials by about 1 eV .13 7. both exactly and in LDA. The LDA exchangecorrelation potential.61 24.10.87 15.94 6.67 14.3.22 15.26 14. • Repeating with the LSD numbers.98 15. IONIZATION ENERGIES AND ELECTRON AFFINITIES 0 20 83 84 CHAPTER 10.53 10.93 13.03 11.6 0. This will be discussed in great detail in chapter X.61 23.01 5.59 5.72 13.73 14.16 10. 10. the mean error is 0.51 10.59 5. we ﬁnd that the errors vary in sign.90 exact 13.45 21. PROPERTIES vS (r) 40 60 80 100 0 0.4: Hartree potential of Ne atom. The X results use Engel’s code.57 11.02 7. In Table ??. This in turn means that the orbital eigenvalues can be very diﬀerent from exact KohnSham eigenvalues.65 10.14 20.55 13. In fact.49 7.64 5.84 6.98 22.
Thus the LDA correlation energy vanishes exponentially in this limit. These are called cohesive energies of solids.6 Transition metals vXC (r) 3 4 5 6 7 8 0 0. qualitative errors occur for highly correlated systems. 10.8 Gaps 10. we ﬁnd that LSDX almost always underestimates the ionization energy. This is much better than HF. or H2 stretched to large distances. or about 50 millihartrees.7 Weak bonds r Figure 10. usually (but not always) being underestimated by about 12%. but not accurate enough for many purposes.4. But many processes are determined by much weaker bonds.25 eV. but the band gap. and ignoring relativistic eﬀects. and van der Waals dimers.3 eV. these eigenvalues are often plotted as the band structure in solidstate texts. LDA cannot reproduce this asymptotic behavior. R→∞ (10. For similar reasons. Thus LSDX is a poorer approximation to exact exchange than LSD is to the exact result. Electron aﬃnities Electronegativity and Hardness The remarks above apply to covalent. HF underestimates such binding energies by about 100 kcal/mol. The exchange energy is always repulsive. and has an MAE of 0. the overall shape and position is good. To calculate these. often nonexistent. LDA signiﬁcantly overbinds both hydrogen bonds.9 Questions 1. Typically. being the energy diﬀerence between molecules or solids at equilibrium. metallic. as will be discussed below. In these cases. DISSOCIATION ENERGIES 0 1 2 85 86 CHAPTER 10. both exactly and in LDA.10.05˚range. • Interestingly. Does the reliability of LSD show that most systems are close to uniform? However. some semiconductors have no gap in LDA. there is a cancellation of errors between exchange and correlation. take the mean absolute error (MAE). . Ebind = − 10. an infamous failure is the prediction that the nonmagnetic structure of iron is of slightly lower energy than the magnetic one. Between two isolated pieces of matter (no signiﬁcant density overlap) there are always van der Waals forces due to ﬂuctuating dipoles. and ionic bonds. or 1. In absence of permanent dipoles. 10. In some cases. is consistently underestimated by at least a factor of 2. Finally.5: Exchangecorrelation potential of Ne atom. the zeropoint energy of the molecule (or solid) must be added to the minimum in its total energy curve. density functionals are much less reliable for transition metal complexes than ﬁrst. such as the solid NiO or the molecules Cr2 . so the existence of C 6 is a purely correlation eﬀect. and their constituent atoms.6 0.5 Geometries and vibrations On the other hand. so it makes the incorrect prediction that they are metals. if we compare LSDX with the exact exchange numbers. Again. These systems are extremely diﬃcult to study with density functional approximations.4 Dissociation energies We consider ﬁrst atomization energies. this also leads to transition state barriers that are too low. PROPERTIES 10. the energy between two such pieces behaves as C6 .8 1 In solid state physics.4 eV. for solids.2 0. A 10. This meant LDA was never adopted as a general tool in quantum chemistry. bond lengths are extremely good in LDA. since any contribution to the energy diﬀerence must come from a density diﬀerence. while LDA overbinds by about 30 kcal/mol. such as hydrogen bonds and dispersion forces. with bondlength errors more likely in the 0.1) R6 where R is their separation. which is 0. In general.4 Ne atom exact LDA 0.and secondrow elements. between HOMO and LUMO.
Part III Analysis 87 .
1 Size consistency 1 Another exact inequality for the potential energy contribution to the exchangecorrelation functional is the LiebOxford bound: LDA UXC [n] ≥ CLO EX [n] An important exact property of any electronic structure method is size consistency.e. the correct wavefunction becomes a linear combination of two Slater determinants (the HeitlerLondon wavefunction). r2 ) = √ (φA (r1 )φB (r2 ) + φB (r1 )φA (r2 )) .4) where CLO ≤ 2. 2 11. As the nuclear separation grows from zero. Any local or semilocal functional is sizeconsistent. Most density functionals are sizeconsistent. we introduce various simple exact conditions. Their densities are given. and yields roughly accurate numbers: (11. Imagine you have a system which consists of two extremely separated pieces of matter.4 Bond breaking unif C (rs )? Does LDA respect the LiebOxford Conﬁguration interaction calculations with a ﬁnite order of excitations are not sizeconsistent. the ground state of two electrons is always a singlet. one for each electron on each Hatom: 1 Ψ(r1 . This is exact for one or two (spinunpolarized) electrons. Exercise 41 Stretched H+ 2 Calculate the LSD error in the exchange energy of H+ in the limit of inﬁnite separation. SIMPLE EXACT CONDITIONS Exercise 40 FermiAmaldi correction An approximation for the exchange energy is EX [n] = −U [n]/N . By spinscaling.6) 2 where φA is an atomic orbital centered on nucleus A and similarly for B. a HartreeFock single Slater determinant is a reasonable approximation to the true wavefunction.g. r2 ) = φ(r1 )φ(r2 ) ≈ Ψ(r1 .5) ΦHF (r1 .2) This is one of the most diﬃcult properties for local and semilocal density functionals to get right. r2 ) However.273. Consider. called A and B. But note that the density remains unpolarized. the charge density spreads out. 1 c 2000 by Kieron Burke. because the exact Hartree energy is a nonlocal functional. Then a sizeconsistent treatment should yield the same total energy. where nI (r) is the hydrogen atom density centered on nucleus I. In the absence of a magnetic ﬁeld. All rights reserved. E[nA + nB ] = E[nA ] + E[nB ].90 CHAPTER 11.3 LiebOxford bound Simple exact conditions In this chapter. 11. 89 . H+ . as can be seen for the H atom results in the tables. whether the two pieces are treated separately or as a whole.3) Another infamous diﬃculty of DFT goes all the way back to the ﬁrst attempts to understand H2 using quantum mechanics. (11. and the overlap is negligible. when the bond is stretched to large distances. (11. At the chemical bond length. i. as it can be thought of as arising from the interaction of the charge density with itself in U . Show that it is not sizeconsistent. It takes a lot of work to prove this result. but coupledcluster calculations are. e. for two spinunpolarized electrons. we ﬁnd EX = −U/2 (N = 2.2 One and two electrons For any one electron system Vee = 0 EX = −U EC = 0 (N = 1) (11.1) (11.. and see how well LDA meets these conditions. Chapter 11 This problem becomes acute when there is more than one center in the external potential. 11. but it provides a strong bound on how negative the XC energy can become. but some with explicit dependence on N are not. The error made for oneelectron systems is called the selfinteraction error. nA (r) and nB (r). Exercise 42 LiebOxford bound What does the LiebOxford bound tell you about bound? 11. The total density is just n(r) = nA (r) + nB (r). unpolarized) (11. LSD does have a selfinteraction error. 2 so that a local approximation produces a smaller and smaller fraction of the correct result.
11. that has been wellknown in quantum chemistry for years. We would much rather ﬁnd E LSD = 2E LSD (H). KohnSham DFT calculations with approximate density functionals have a similar problem. UNIFORM LIMIT 91 92 CHAPTER 11. for example. Thus the symmetry dilemma is that an RHF calculation produces the correct spin symmetry with the wrong energy. There is a way to do this. If one allows a HartreeFock calculation to spontaneously break spinsymmetry. The LSD calculation will spontaneously break symmetry at a CoulsonFischer point (further than that of HF) to yield the appropriate energy. we require EXC [n] → V eunif (n) XC n(r) → n (constant) (11. we make a large error. do not require the upspin spatial orbital to be the same as that of the downspin. which correctly dissociates to two LSD H atoms. What does the LO bound say for one electron? 11. while the UHF solution produces the right energy but with the wrong spin symmetry. by producing two H atoms that have opposite spins. A spinrestricted LSD calculation for two H atoms 10˚apart yields a diﬀerent answer to A twice the answer for one H atom. but only if we use the correct energy density. Thus we speak of RHF and UHF. the unrestricted solution will have lower energy than the restricted one.11. This can be achieved. is that a sizeconsistent approximation? 2. and we saw in a previous chapter the huge error LDA makes for the Hatom. because E LSD [n] = E LDA [nA ]+ [nB ]. then at a certain separation. Getting this right is relevant to accuracy for valence electrons of simple bulk metals. Does this mean LSD is not size consistent? 3. But more generally. which is utterly incorrect. SIMPLE EXACT CONDITIONS If we apply LSD to the exact density. LDA 4. by taking ever larger boxes. The trouble with this is that the UHF solution is no longer a pure spineigenstate.7) where V is the volume of the system. E because it does not account for spinpolarization. In the unrestricted LDA calculation. but rather the spindensities themselves.e. Does LDA satisfy the LO bound pointwise? Does LSD? . but keeping the density of particles ﬁxed.6 Questions 1. called the CoulsonFischer point.. This limit is obviously gotten right by LDA.5. and will yield the correct energy in the separatedatom limit. If I approximate the correlation energy as about 1 eV per electron. such as Li or Al. Now the dilemma is no longer the spin eigenvalue of the wavefunction. i. the spindensities are completely polarized.5 Uniform limit We have already used the argument that any local approximation must use the value for the uniform limit as its argument.
one gets a similar factor for each coordinate: Ψγ (r1 . All rights reserved.1. 1= ∞ −∞ 1 where the scaling factor out front has been chosen to preserve normalization: dx φ(x)2 = ∞ −∞ 0 < γ < ∞. We always use subscripts to denote a function which has been scaled.. −∞ −∞ γ 1 nγ (r) = γ 3 n(γr) (12. . .94 CHAPTER 12.11) (12. Exercise 43 Scaling of trial wavefunctions Do the sets of trial wavefunctions in Ex. . and scaled by a factor of 2. it is diﬀerent for every diﬀerent problem. ∞ −∞ (12. The potential energy is not so straightforward in general. . 12. Consider any set of trial wavefunctions.2) dx γφ(γx)2 = dx φ(x )2 . and of γ < 1 will stretch it out.0 Figure 12.1) (12.. keeping the number of electrons ﬁxed. nγ (x) = γn(γx). . (12. i. because it is not a universal functional. (12.e. 12. when x is replaced by γx everywhere in a problem. (12.7) (12.1 Wavefunctions 1 However. if the potential energy is homogeneous of degree p. i. and the same density scaled by γ = 2. For one electron in one dimension.5) V [φγ ] = dx φγ (x)2 V (x) = dx φ(x )2 V ( ) = V (x/γ). with γ a positive number. γrN ). as the natural way to understand the most important limits about density functionals. In this chapter. For example.0 Chapter 12 Scaling 0.3) Note how γ > 1 squeezes the density toward the origin. What is the kinetic energy of such a scaled wavefunction? 1 ∞ γ ∞ γ2 ∞ dx φγ (x)2 = dx φ (γx)2 = dx φ (x )2 = γ 2 T [φ].e. 93 . the kinetic energy grows quadratically with γ.e.10) c 2000 by Kieron Burke. √ For example.8) When including many electrons.1: Cartoon of the onedimensional H2 molecule density. (12.6) It will prove very useful to ﬁgure out what happens to various quantities when the coordinates are scaled. V (γx) = γ p V (x). (12. It is simple to show: T [Ψγ ] = γ 2 T [Ψ] and Vee [Ψγ ] = γ Vee [Ψ] while Vext has no simple rule in general. We can they easily evaluate our favorite operators. we deﬁne the scaled wavefunction as φγ (x) = γ 2 φ(γx). then V [φγ ] = γ −p V [φ]. then we say that the set admits scaling. 10 admit scaling? The only important change when going to three dimensions for our purposes in the preceding discussion is that the scaling normalization factors change: φγ (r) = γ 2 φ(γr) 3 A scale factor of γ > 1 will squeeze the wavefunction.0 !5. .e.. i. . φ2 (x) = 2 exp(−2x). ∞ ∞ x (12. we sketch a density for the onedimensional H2 molecule of a previous exercise.4) T [φγ ] = −∞ −∞ 2 2 2 −∞ i. rN ) = γ 3N/2 Ψ(γr1 . consider φ(x) = exp(−x). Then√ γ (x) = γ exp(−γx) is a diﬀerent φ wavefunction for every γ. SCALING 2.. we introduce the concept of scaling the density. If scaling any member of the set leads to another member of the set. . In Fig.0 5.9) suppressing spin indices.
we see that for the He atom. and so extract the separate contributions to the correlation energies. they are also extremely powerful in limiting the possible forms functional approximations can have.. unif (rs ) = −d0 /rs + d1 /rs + . Consider the latter case ﬁrst.15). (12. we see that for the He atom EC varies little with scaling. . so that C . it will be overestimated at γ = 1. but is yes for the KohnSham wavefunction. rN ) . EC [nγ ] is almost constant. in making a local approximation for threedimensional systems. 3 d rN Ψγ (r. Therefore.2. as is the case with most systems of chemical interest. On the other hand. and the fact that U [nγ ] = γ U [n] (12. We already know that TS [Φγ ] = γ 2 TS [Φ]. both exactly and within LDA. Functionals deﬁned as explicit operators on the KohnSham wavefunction are far simpler than those including correlation eﬀects. However. . . and vice versa: TS [nγ ] = TS [Φ[nγ ]] = TS [Φγ [n]] = γ 2 TS [Φ[n]] = γ 2 TS [n]. is the scaled wavefunction the same as the wavefunction of the scaled density? We show below that the answer is no for the physical wavefunction. is Ψ[nγ ] = Ψγ [n]? That is.9). and EX [nγ ] = EX [Φ[nγ ]] = EX [Φγ [n]] = γEX [Φ[n]] = γEX [n]. using Eq. . Thus if Φ minimizes TS and yields density n. using Eq. (12.18) In the highdensity limit. series for EC [nγ ] around the highdensity limit: (2) (3) EC [nγ ] = EC [n] + EC [n]/γ + . as seen from Eqs. . these functionals should change in a trivial way.19) Note that in this lowdensity regime. . Exercise 44 Local density approximations for TS and EX : Show that. . or Φγ [n] = Φ[nγ ]. Consider uniform scaling of an N electron wavefunction.3 1 exact LDA 2 3 4 5 6 7 8 9 10 He atom γ Figure 12. In fact. as in Eq. We can immediately use it to see how they scale.15) (12.2. and the uniform electron gas is referred to only for the values of the coeﬃcients.05 0. This is logarithmically singular in C C true LDA. . (F.2: Correlation energy of the He atom. Thus LDA EC [nγ ] = d3 r nγ (r) unif C (rs. SCALING 12. γ → ∞. .2 0.15 0. rS → rS /γ. and test them against their exact counterparts. and diverges logarithmically as γ → ∞. γ→0 (12. and is an error made by LDA. . The density of the scaled wavefunction is (12. If Ψ[n] is the groundstate wavefunction with density n(r).12) Now. the correlation energy is believed to have the following expansion under scaling to the lowdensity limit: EC [nγ ] = γB[n] + γ 3/2 C[n] + . (12. From the ﬁgure. but yields density nγ .2 Density functionals 12. r2 . although it might not be numerically 3/2 terribly accurate here. saying only that if the length scale of a system changes. EC [nγ ] tends to a negative constant as γ → ∞. = γ n(γr) = nγ (r). the power of the density in both the noninteracting kinetic and exchange energy approximations are determined by scaling considerations. In the low density limit. We shall see that both their uniform scaling and their spinscaling are straightforward. showing that the system is close to the highdensity limit. This limit is satisﬁed within LDA.1 0.14) Levy has shown that for ﬁnite systems. γ→∞ (12.3 Correlation n(r) = N d r2 . .12. . as the density is squeezed. it is very close to linear in 1/γ. DENSITY FUNCTIONALS 95 96 CHAPTER 12. a key question is this. We may write power 0 0. since we can turn density functionals into orbital functionals. The highdensity property is violated by LDA (see the ﬁgure). This is another way to understand the LDA overestimate of correlation for ﬁnite systems. From Fig.17) We can easily scale any approximate functional. there is both the noninteracting kinetic energy and the exchange energy. 12.γ (r)) = d3 rn(r) unif C (rs (r)/γ) (12. A simple example is correlation in LDA. 3 2 3 EC [nγ ] Now we turn our attention to density functionals. In particular. correlation is so strong it scales the same as exchange. Any approximation to these functionals must therefore satisfy these relations. (12. . etc.13) This result is central to understanding the behaviour of the noninteracting kinetic and exchange energies. (??). unif (rs /γ) → unif (0). as we saw in Chapter 8. Since the LDA correlation energy diverges as γ → ∞.16) These exact conditions are utterly elementary. Φγ is the KohnSham wavefunction for nγ . Then when n → nγ . The next exercise shows us that scaling relations determine the functional forms of the local approximation to these simple functionals.25 0. then Φγ also minimizes TS .24). check limiting values. .
(12.2 and 12. (12. and multiply through by γ. TC [nγ ] ≥ γ 2 TC [n]. the correlation contributions. so we can combine the two equations. Eq. correctly. Clearly. He atom EXC [nγ ] exact LDA 1 2 3 4 5 6 7 8 9 10 Note that for γ > 1. Exercise 45 Extracting exchange If someone gives you an exchangecorrelation functional EXC [n]. exchange scales linearly.e. 12. both exactly and within LDA. But of greater interest are the much smaller diﬀerences with KohnSham values. As usual. 0 1 2 3 4 5 6 7 8 9 10 This is the fundamental inequality of uniform scaling.28) Exercise 49 Scaling inequalities for LDA Give a oneline argument for why LDA must satisfy the scaling inequalities for correlation energies. T [nγ ] ≤ γ 2 T [n] γ > 1.21). SCALING LDA EC [nγ ] = −γd0 d3 r n(r)/rs (r). Exercise 46 Scaling LDA in Wigner approximation: Using the correct exchange formula and the simple Wigner approximation for the correlation energy of the uniform gas. 12. apply Eq. and dominates correlation. However.27) This reasoning breaks down when correlation is included. because this involves energies evaluated on the physical wavefunction. (8. as the density is squeezed. Exercise 47 Scaling Vee Show Vee [nγ ] ≥ γVee [n] γ > 1. to ﬁnd (12. deﬁne a procedure for extracting the exchange contribution. Vee [nγ ] ≤ γVee [n] γ < 1.22) But we can just drop the primes in this equation.3. a key realization there is that the scaled groundstate wavefunction is not the groundstate wavefunction of the scaled density.21) γ > 1. EC [nγ ] ≤ γEC [n]. i.24) (γ − 1) T [nγ ] ≤ γ 2 (γ − 1) T [n].4..4 Correlation inequalities and These inequalities actually provide very tight bounds on these large numbers.26) because of the cancellation of errors.12. γ < 1. for most systems we study. vanishing linearly with γ. To see an example. (12. because the physical wavefunction minimizes both T and Vee simultaneously. i. to yield F [n ] ≤ T [nγ ]/γ 2 + Vee [nγ ]/γ . As mentioned above.. and the simple exponential density for the He atom. In fact. So Exercise 48 Scaling EC Show EC [nγ ] ≥ γEC [n].3: Exchangecorrelation energy of the He atom. γ < 1. we can cancel γ − 1 from both sides. Remember.20) (12.21) to n = nγ and write γ = 1/γ. γ > 1. 12. TC [nγ ] ≤ γ 2 TC [n]. and add (γ 2 − γ)T [n] to both sides: γ 2 T [n] + γVee [n] ≤ T [nγ ]/γ + Vee [nγ ] + (γ 2 − γ)T [n]. to ﬁnd T [nγ ] < γ 2 T [n]. we can still use the variational principle to deduce an inequality. . We write ˆ ˆ ˆ ˆ F [nγ ] = Ψ[nγ ] T + Vee Ψ[nγ ] ≤ Φγ [n] T + Vee Φγ [n] or F [nγ ] ≤ γ 2 T [n] + γVee [n] (12. As we scale the system to high density. one should not focus too much on the failings of the correlation energy alone. CORRELATION INEQUALITIES 97 98 CHAPTER 12. (12. (12.25) γ Figure 12.3. and worsens as γ grows. both EC and TC are relatively insensitive to scaling the density toward the highdensity limit. When we add both contributions. But most systems of chemical interest are closer to the highdensity limit. the physical kinetic energy grows less rapidly than the noninteracting kinetic energy. (12. the reverse is true. T [nγ ] ≥ γ 2 T [n] γ < 1.e. where LDA fails. we ﬁnd Fig. calculate the curves shown in Figs.23) Now the lefthandside equals the righthandside of Eq. the XC energy is best at γ = 1. as it tells us inequalities about how correlation contributions scale. (12.18). (12. For γ < 1. so these inequalities are less useful.
These yield the same density.35) to both the interacting and KohnSham systems.35) Any eigenstate wavefunction extremizes the expectation value of the Hamiltonian. dγ As usual. except at γ = 1. eunif (rs ).29). By applying Eq. We apply Eq. yielding dV 2T = x . Write F = T + Vee . a class in which coordinate scaling of any member of the class yields another member of that class. if we insert any other wavefunction yielding the density n(r). give formulas C for both tunif (rs ) and uunif (rs ). i. show TC [nγ ] = γ dEC [nγ ] − EC [nγ ] dγ (12. and subtract.39) γ γ2 using the fact that EC [nγ ] vanishes as γ → 0. Linear combinations of functions do not generally form a class which admits scaling. leading to Eq. swapping γ to 1/γ in the derivative. Then d ˆ γ=1 Ψγ [n1/γ ]F Ψγ [n1/γ ] = 0. C C Exercise 53 Scaling to ﬁnd correlation energies: 1.5 Virial theorem or. (12. Exercise 52 LDA kinetic correlation energy: Assuming the correlation energy density of a uniform gas is known. but are not the minimizing wavefunctions. (12.36). (12. Thus any small variations in such a wavefunction lead only to second order changes..33) But expectation values of operators scale very simply when the wavefunction is scaled.32) and.6 Kinetic correlation energy But. (12. Note that this is true for any eigenstate and even for any approximate solution.37) This is an extremely useful and powerful statement. This statement is trivially true for Hartree and Exchange energies. and write: TC [n] = dEC [nγ ] − EC [n]  dγ γ=1 (12. instead of considering the groundstate energy ˆ ˆ ˆ itself. .31) exactly. so d ˆ  φγ  H  φγ = 0. consider Eq. we consider just the construction of the universal functional F [n]. we can extract the kinetic contribution simply by scaling. 12.31). the more interesting statement comes from the KS quantities. (12. As always.31) dx The righthand side is called the virial of the potential.4). We will ﬁrst generalize Eq. and from Eq. which approximate solutions satisfy the virial theorem Eq. and which do not? What can you deduce in the two diﬀerent cases? We now derive equivalent formulae for the density functional case.5). dT [φγ ]/dγ = 2γT [φ]. Exercise 50 Using the virial theorem Of the exercises we have done so far. E[φsol + δφ] = E[φsol ] + O(δφ2 ). Then φγ is not. Next.5. variations induced by scaling (but keeping the density ﬁxed) have zero derivative around the solution. dV [φγ ]/dγ = −γ −2 x dV dx at γ = 1. VIRIAL THEOREM 99 100 CHAPTER 12.37) to nγ (r). 2T [n] + Vee [n] = d γ=1 (T [nγ ] + Vee [nγ ]) . So this must be satisﬁed by any solution to the Schr¨dinger equation.. (12. dγ (12. (12. Consider especially the wavefunctions Ψγ [n1/γ ]. But ˆ small variations in γ near 1 lead only to second order changes in H . (12. (12. if we’re given a functional for the correlation energy. dγ 2. and can be used to test how accurately an approximate o solution satisﬁes it. SCALING 12. i. It tells us how. once that solution is a variational extremum over a class of wavefunctions which admits scaling. (12. Then ˆ F [n] = Ψ[n]F Ψ[n] .30) This then is the generalization of Eq.12. (12. (12. from Eq. (12. to ﬁnd: EXC [n] + TC [n] = dEXC [nγ ] γ=1 dγ (12. dγ γ=1 (12.29) for density functional theory: Because Ψ[n] ˆ minimizes F .e.29) ˆ Assume φ is the groundstate wavefunction of H.38) (12. (12. so we subtract oﬀ the exchange contribution.e. and then Eq.38) as a ﬁrstorder diﬀerential equation in γ for EC [nγ ].34) γ=1 γ 2 T [n1/γ ] + γVee [n1/γ ] = 0.36) Exercise 51 Virial expressed as scaling derivative Prove Eq. we must get a higher number. Show that ∞ dγ EC [nγ ] = −γ TC [nγ ]. (12.35). so that d (12.
how is vC [n](γ. POTENTIAL 101 102 CHAPTER 12. once a KohnSham calculation has been cycled to selfconsistency. (12. the virial theorem for the groundstate yields: N 2T = i=1 Ψ ri · iV ˆ Ψ. we know E = −T . What is the relation between scaling and changing Z for the Hatom? 2. (12. 12. Can we ﬁnd the groundstate by simply maximizing T ? 4. (12. Does TSloc [n] from chapter ?? satisfy the correct scaling relation? (Recall question 1 from chapter ??). r) = δEC [n](γ)/n(r) related to vC [n](r)? Exercise 55 Scaling LDA in Wigner approximation: Within the simple Wigner approximation for the correlation energy of the uniform gas. Eq. Then check all 6 relations of the previous exercise. Exercise 54 Scaling of potentials: If vC [n](r) = δEC [n]/n(r). C C respectively.40) and (12. (12. but is rarely performed in practice. Eq. (8.43) . What is the exact kinetic energy density functional for one electron in onedimension? How does it scale? 5.8 Questions dγ UC [nγ ]. 6. and the remaining correlation contributions: For almost any approximate functional. p=1 for the delta function well. of what is Ψγ a groundstate? After our earlier warmup exercises.7 Potential 1.7.39) to ﬁnd diﬀerential and integral relations between U C and TC . uunif (rS ) and tunif (rS ). Is the virial theorem satisﬁed for the groundstate of the problem V (x) = − exp(−x)? 3.18).12.40) (12. 12. and so 2T + Vee = This theorem applies equally to the noninteracting and interacting systems. Combine Eqs. e.41) Exercise 56 Virial for LDA: Show that the XC virial theorem is satisﬁed in an LDA calculation..44) d3 r n(r)r · vext (r).44) will be satisﬁed. It is thus a good test of the convergence of such calculations. For a homogeneous potential of degree p. (12. If Ψ is the groundstate for an interacting electronic system of potential v ext (r). SCALING 3. this applies separately to the exchange contributions to both sides. deduce both the potential and kinetic contributions. and I deﬁne EC [n](γ) = EC [nγ ]. the electronelectron repulsion is homogeneous of order 1. For our problems. we ﬁnd Since we can either turn oﬀ the coupling constant or scale the density toward the highdensity limit. (12. it is trivial to derive the virial theorem for the exchangecorrelation potential. and by subtracting the diﬀerence. For an arbitrary system. Show UC [nγ ] = 2EC [nγ ] − γ and EC [nγ ] = γ 2 ∞ γ dEC [nγ ] dγ (12. EC [n] + TC [n] = − d3 r n(r) r · vC (r).45) EXC [n] + TC [n] = − d3 r n(r) r · vXC (r).g. γ3 4.42) no matter what the external potential.
∂λ (13. and for the 1d harmonic oscillator.12) 1 0 dλ φλ  ˆ ∂H λ λ φ . and we can identify φλ (γx). The Schr¨dinger equation at coupling constant λ is o ˆ ˆ T + λV (x) φλ (x) = E λ φλ (x) If we replace x by γx everywhere. If we γ scale both of these by 1/γ. for the harmonic oscillator of force constant 2 /2). the factor of λ multiplies k.10) (13. Exercise 58 λdependence Show that. Because the eigenstates are variational extrema. i. but this relation depends on the details of the potential. All rights reserved. Since V (x) = k = ω 2 . the coupling constant of the interaction for DFT. Eq. (1. and then all eigenvalues and eigenfunctions depend on λ. ˆ ˆ ˆ H λ = T + λV ..2) A λ in front of the potential is often called the coupling constant. E λ = λ p+2 E 2 (13. This diﬀers from that of regular manybody theory.5) c 2000 by Kieron Burke. V (x) = 1 x2 . and so √ 1/4 √ λω λ φ (x) = exp(− λωx2 /2) (harmonic oscillator). 2 103 . dλ (13. This is often called the HellmannFeynman theorem. λV (x) = λkx /2. one of the most important relationships is between the coupling constant and coordinate scaling. if λ simply multiplies V . leaving ˆ ∂H λ λ dE λ = φλ  φ .. Exercise 59 λdependence of 1d H Find the couplingconstant dependence for the wavefunction and energy for the 1d Hatom.11) In this case. Most commonly. for a homogeneous potential of degree p. the groundstate wavefunction is φ(x) = (ω/π)1/4 exp(−ωx√ 2 2 kx /2. Then ω → λω. say λ.1 One electron ˆ Introduce a parameter in H. then dE λ ˆ = φλ  V  φ λ . (13.3) π For a λdependent Hamiltonian. then Eq. V (x) = −δ(x). Then all eigenstates and eigenvalues depend on λ.1) H λ φλ = ελ φλ . But we see at the end that this is very simply related to scaling.104 CHAPTER 13.9) becomes: ˆ (13.4) (13. 0 < λ < ∞.7) Adiabatic connection 1 In this chapter. the contributions due to diﬀerentiating the wavefunctions with respect to λ vanish. we ﬁnd φλ (x) = φ1/γ (x) = φλ1/(p+2) (x) (13. How does the approximate solution using a basis set do? In modern density functional theory. we ﬁnd 1 ˆ ˆ T + λV (γx) φλ (γx) = E λ φλ (γx) γ2 ˆ Furthermore.e. i.9) (13. (13. changing coupling constant by λ is equivalent to scaling by λ1/4 .10) is just the normal Schr¨dinger equation. (13. (13. dλ ∂λ Thus E = E λ=1 = E λ=0 + 1 o Thus. if V is homogeneous of degree p. Exercise 57 HellmannFeynman theorem Show that the 1d Hatom and 1d harmonic oscillator satisfy the HellmanFeynman theorem. We can see this for these simple 1d problems. we can diﬀerentiate the energy with respect to λ.e. ADIABATIC CONNECTION ˆ Furthermore.6) Chapter 13 so that E = E λ=0 + 1 0 ˆ dλ φλ  V  φλ . we introduce an apparently new formal device. which is equal (up to normalization) to φλ (x) with φ(x).8) 13. For example. if we choose λ γ p+2 = 1. ˆ ˆ T + λγ 2 γ p V (x) φλ (γx) = γ 2 E λ φλ (γx) (13.
To demonstrate the usefulness of adiabatic decomposition.15) but The KohnSham quantities are independent of λ. 13. EC = −0.2.079. and not include the external potential. Thus ˆ T [n] = Ψ [n]T Ψ [n] . Inserting the deﬁnition of the correlation energy. By a superscript λ. Note ﬁrst the solid. and introduced the simple λ notation UXC (λ) = UXC /λ for later convenience. We denote Ψλ [n] as the minimizing wavefunction for a given λ.14) (13. because Ψλ [n] is a minimizing wavefunction at each λ. we ﬁnd EXC [n] = = 1 0 1 0 ˆ dλ Ψλ [n] Vee Ψλ [n] − U [n] dλ λ U [n] = λ XC 1 0 dλUXC [n](λ).t. • The area beneath the curve. We see that LDA improves as λ grows from 0 to 1.2 Adiabatic connection formula Now we apply the same thinking to DFT. both Hartree and exchange energies have a linear dependence on the couplingconstant.r. The adiabatic connection formula is a method for continuously connecting the KohnSham system with the physical system. λ of F λ [n] is just the derivative of the operator w. We can even consider λ → ∞. It is almost straight on this scale.042.037. But once we i. λ λ λ λ ˆ where we have used Vee = λ Ψλ [n] Vee Ψλ [n] and U λ = λU . For λ = 0. But for all values of λ.t. exact line.4 0.8 1 (13. This is characteristic of a very general trend. • The area between the curve and the xaxis is just EXC . (13.. which is a highly correlated system. TS [n] = F λ=0 [n]. Note that λ this implies that the external potential is a function of λ: vext (r). are EX = −1. ADIABATIC CONNECTION 13.1.19) . ADIABATIC CONNECTION FORMULA 105 106 CHAPTER 13. When doing so. We can understand this ﬁgure by making 0 UXC (λ) EX −TC 0 0. and TC = 0. Thus EC is only slightly more negative than −TC . we always try to keep the density ﬁxed (or altered only in some simple way).13) For λ = 1. EX [n] = λEX [n].r. Exercise 60 Coupling constant dependence of exchange Show that λ U λ [n] = λU [n]. 13. just as in the oneelectron case. (13.13. λ ˆ ˆ ˆ ˆ EXC [n] = Ψλ [n] T + λVee Ψλ [n] − Φ[n] T + λVee Φ[n] (13. We can generalize all our previous deﬁnitions. is just TC . λ Figure 13. Very simply. in which the kinetic energy is negligible.16) several key observations: • The value at λ = 0 is just EX . This is telling us that this system is weakly correlated. Our next step is to write the HellmannFeynman theorem for this λdependence in the Hamiltonian. as we discuss below. The curve itself is given by the solid line. This is the celebrated adiabatic connection formula.e. the density remains that of the physical system. This produces yet another way to understand the cancellation of errors. from Table X. λ. we mean the expecation value of an operator on the system with couplingconstant λ. but at the price of having to evaluate it at all intermediate coupling constants. between the curve and a horizontal line drawn through the value at λ = 1. Thus the adiabatic connection curve gives us a geometrical interpretation of many of the energies in density functional theory. UC = −0. so λ ˆ Vee [n] = Ψλ [n]λVee Ψλ [n] . we have the physical system. to multiply Vee : ˆ ˆ F λ [n] = min Ψ T + λVee Ψ Ψ→n (13. we show this curve for the He atom in Fig. A cartoon of the integrand in the adiabatic connection formula for a given physical system is sketched in Fig. we have the KohnSham system.2 0.025.1: Cartoon of the adiabatic connection integrand. We write F [n] = TS [n] + 1 0 ˆ dλ Ψλ [n] Vee Ψλ [n] (13. and horizontal lines have been drawn at the value at λ = 0 and at λ = 1.18) where F [n] = F λ=1 [n]. We will see shortly that this price actually provides us with one of our most important tools for analyzing density functionals. but we must do so carefully.6 EX UXC EC 0. The derivative w. The numbers. we introduce a coupling constant λ into the universal functional.2.17) • The value at λ = 1 is just EX + UC . We have written the exchangecorrelation energy as a solely potential contribution.
as it should be. and we have written EC [n] = (T [n] − TS [n]) + (Vee [n] − U [n]) = TC [n] + UC [n] (13. λ λ ˆ Vee [n] = λ Ψλ [n] Vee Ψλ [n] = λ(U [n] + EX [n]) + UC [n].19) is simply UXC (λ)[n] = λUXC [n1/λ ]. (13. (13.. we ﬁnd.1 0 0. If we choose γ = 1/λ. but is much more transparent. through Eq. claim that the LDA error drops with λ. both exactly and in LDA. (13.4 0. since EX is constructed from the λindependent KohnSham orbitals. RELATION TO SCALING 0.27) where we have used the linear dependence of exchange on λ. Much less trivial is the relation between diﬀerent components of the correlation energy. (13.21). as we shall see. Similarly. (13. (13.3. First note that we can take any functional and ﬁnd out its scaling behavior quite easily.22) with Eq. which minimizes T +λVee and has density n(r). after a little manipulation which tells us how to construct a wavefunction of coupling constant λ by ﬁrst scaling the density by 1/λ. (13. Then TSλ [n] = λ2 TS [n1/λ ] = TS [n] λ EX [n] = λ2 EX [n1/λ ] = λEX [n] UXC (λ) 0. (13. the adiabatic connection integrand of Eq.30) . (13. which again makes sense.2 0. Then Ψλ [n] minimizes T /γ 2 +(λ/γ)Vee γ λ 2 ˆ ˆee ) and and has density nγ (r).26) and cancelling the trivial exchange contributions to both sides. We can use the adiabatic decomposition (λdependence of UXC (λ)) to analyze any approximate functional. we simply note that the statement that LDA improves with λ is equivalent to the cancellation of errors statement. (13. and ﬁnd out why it behaves as it does.6 0.27) into Eq. we ﬁnd λ dEC [n] λ (13. Contrast Eq. This solves a mystery from the previous chapter: The groundstate wavefunction of a scaled density is the scaled groundstate wavefunction. In fact. These relations prove to be extremely useful in analyzing functionals and their behavior.22) which shows how the λdependence of the universal functional (or. while UC [n] is the potential contribution.28) and TC [n] is the kinetic contribution to the correlation energy.25) i. ADIABATIC CONNECTION He atom 0. since the area under the curve will then be more accurate than the value at λ = 0. we can now deduce its couplingconstant dependence.21) γ γ dEC [nγ ] = EC [nγ ] + TC [nγ ].e. we already know..29) = UC [n]/λ dλ λ λ Finally. The next section shows how easy it is to ﬁnd UXC (λ) from EXC . from Eq. Why this happens will be discussed later. applied to that functional. 13. (13.22). But if λ2 (T + Vee ) is minimized. we rewrite this as a scaling relation.9 Our argument also yields a simple relation for the energies: F λ [n] = λ2 F [n1/λ ]. i.24) (13. Since EC [n] = λ2 EC [n1/λ ] and UC [n] = 2 λ UC [n1/λ ] and writing γ = 1/λ. any energy component) is completely determined by its scaling dependence. Consider Eq. But for now. the kinetic energy is independent of coupling constant.26) A third important concept is the relation between scaling and coupling constant. but only if the coupling constant is changed.8 1 λ Figure 13. Ψ[nγ ] = Ψ1/γ [n]. Inserting Eq. and then scaling that wavefunction back to the original size. dγ (13. Then.20) Then. For example. then the usual cancellation of exchange and correlation errors follows.23) The simplest example in this regard is the noninteracting kinetic energy. (13. (12.e.95 1 1. so that we can identify this wavefunction being simply Ψ[n1/λ ]. then T + Vee is minimized.05 1.13. ﬁnding the groundstate wavefunction for the scaled density.3 Relation to scaling (13. which we know scales quadraticly with scaling parameter. If we scale both wavefunctions by λ. we ﬁnd the extremely simple but important result Ψλ [n] = Ψλ [n1/λ ].2: Adiabatic decomposition of exchangecorrelation energy in He atom.19) in diﬀerential form: λ dEXC [n] ˆ = Ψλ [n] Vee Ψλ [n] − U [n] dλ (13.85 107 exact LDA 108 CHAPTER 13. Consider ˆ ˆ ˆ ˆ Ψλ [n]. we ﬁnd Ψ1/λ [n] minimizes λ (T + V ˆ ˆ ˆ ˆ has density n1/λ (r).22). integrated with λ/r − r .
it has always been found that the adiabatic connection curve is (slightly) concave upwards: d2 UC (λ) ≥0 (unproven) (13. Exercise 65 Changing λ: LDA. The value of b is about 0.31) In the highdensity limit. then the step occurs at λ = b. we next consider what happens to a H2 molecule when we stretch it. This is again because of the logarithmic divergence as rs → 0. There is no way for LDA to produce a curve that looks anything like this. Typical systems are close to this value. b reduces.16. we plot the adiabatic connection curve at R = 5. Exercise 61 EC from TC λ λ Derive a relation to extract EC [n] from TC [nγ ] alone. We can easily scale any approximate functional. as rs → ∞. An alternative way to write it is EC = (1 − b)UC (13. and so extract the separate contributions to the correlation energies. with essentially no static correlation. Then when n → nγ . correctly. A simple example is LDA..2 0. this means TC ≤ EC . check limiting values. and this system has strong static correlation. we . To quantify the fraction of correlation that is kinetic. C LDA 3 so that EC [nγ ] = −γd0 d r n(r)/rs (r). rS → rS /γ. Considering both the ﬁgures and the numbers.38). Because the average density will be almost the same as a single H atom.45 d3 r nγ (r) unif C (rs.e.5 for real systems.4. This is singular in true LDA. Thus LDA EC [nγ ] = Thus.13.4 Static correlation 0. To understand it. Thus we can speak of b as measuring the amount of dynamic correlation.4 0.33) UC  To appreciate why we are introducing this parameter.e. Exercise 62 Use Eq. (12. we have the exact exchange value of KS theory. and we will see that this does not happen for more sophisticated approximations. Note that in this lowdensity regime. the curve is almost ﬂat. 13.3: Adiabatic decomposition of exchangecorrelation energy in stretched H 2 . draw the adiabatic connection curve for He.55 0. it has always been found that 0 < b ≤ 1/2. we see that LDA clearly signiﬁcantly underestimates b. On the other hand. and just about equal to −5/8. For this problem.γ (r)) = d3 rn(r) unif C (rs (r)/γ) (13. but b is always close to 0. the adiabatic connection formula may be thought of as an integration of this scaling formula between γ = 1 and γ = ∞.λ LDA Derive EXC [n] and TC [n]. . we note that at λ = 1. γ → ∞.6 0.4 0. at λ = 0. and only for very low densities does b become small. with a value of 1/2 being the maximum. TC eventually becomes small relative to UC . However. or of H2 at equilibrium. we deﬁne TC b= (13. in the low density limit.. As one moves towards lower densities (much lower than in realistic systems). Exercise 63 Adiabatic connection for He atom: Using the Wigner approximation and the eﬀective exponential density. even coming close to zero. unif (rs /γ) → unif (0).32) dλ2 Given the geometric interpretation of Fig X. i. STATIC CORRELATION 109 110 CHAPTER 13.65 0 0. In Figure 13. etc. and is C C 3/2 an error made by LDA. Thus. We have seen how. (16. Thus.5 0. vanishing linearly with γ.6 exact (R=5) Draw the adiabatic connection curve for the H atom. the adiabatic connection curve is close to linear.. correlation is so strong it scales the same as exchange. the fraction of correlation energy that is kinetic. 0. the exchange(correlation) energy of two separate H atoms. even for the uniform gas. It looks very diﬀerent from that of He. unif (rs ) = −d0 /rs + d1 /rs + .34) or that if we simplify the adiabatic curve as a step down at some value of λ from EX to EX + UC . Rewrite this to get EC [n] from TC [n]. for the total XC energy of many systems of interest.8 1 λ Figure 13.3. and test them against their exact counterparts.42.5) and the scaling relations of the previous chapter to write E C in terms of density matrices of diﬀerent λ. In fact. the adiabatic curve is very ﬂat. ADIABATIC CONNECTION which is just Eq. . Exercise 64 Adiabatic connection for H atom: UXC (λ) 0. The reason that most systems have b close to 1/2 is that they are not very far from the highdensity limit. the electrons are always in a singlet and the exchange energy is about 0. i. Since we are beyond the CoulsonFischer point.
if we don’t. If we allow symmetry breaking. Exercise 66 Adiabatic connection for stretched H2 Draw an accurate adiabatic connection curve for H2 with a bond length of 20 ˚.5. Using the highdensity limit of Wigner correlation. A 13. . estimate roughly how large the correlation energy usually is within LDA. QUESTIONS 111 112 CHAPTER 13. but in their energy diﬀerences between bonded and dissociated states.13. and so underestimate the correlation. On the other hand. we will get a curve that bends just a little down from the λ = 0 end. we will get the λ = 1 end about right. As we will see throughout the rest of this book. but have a ﬂat curve. not in their total energies. this is a very important problem for DFT in quantum chemistry.5 Questions 1. ADIABATIC CONNECTION must chose if we wish to break spinsymmetry or not. and so overestimate the magnitude of the correlation energy. Many molecules exhibit similar behavior.
note that Eq. the ionization potential is independent of λ. i. LDA does not provide very realistic looking exchangecorrelation potentials. so that the correlation contribution must decay more rapidly. deduce the limiting values of the asymptotes on the right of Fig. Thus I = E(N − 1) − E(N ) = − HOMO 1 2 3 4 5 6 7 8 0 0. if you ignore relaxation. in exact density functional theory.1 Koopman’s theorem where α(N − 1) is the polarizability of the (N − 1)electron species.114 CHAPTER 14. Exercise 69 Asymptotic potentials Derive the asymptotic condition on vX (r) from the exact decay of the density. Since vext (r) = −Z/r.3) Discontinuities 14. this must also be true for v XC (r). keeping the density ﬁxed.8 1 .2 Potentials vX (r) is due to nonselfinteraction in the exact theory. In a HartreeFock calculation. vX (r) → −1/r.1. In fact. just as smooth approximations for T S miss shell structure. test Koopman’s theorem. This is partially explained by the above argument: the longrange decay of 0 Since the large distance condition Eq. where we now refer to the HOMO of the KS potential. 14. Consider ﬁrst exchange.. Because the density decays exponentially in the tail region. Consider the coupling constant λ. the N − 1 because there are N − 1 electrons close to the nucleus.2 0. Far from the nucleus.1: XC potential in the Ne atom. as will be discussed in more detail below. The decay of the correlation potential is so rapid that it has only ever been clearly identiﬁed in an exact calculation for H− . DISCONTINUITIES Chapter 14 electron in an orbital should see an eﬀective nuclear charge of Z − (N − 1). This has a strong eﬀect on the HOMO orbital energy (and all those above it). and 113 We can understand the apparent diﬀerences between LDA and exact potentials in much more detail. at λ = 0. Consider what happens when two distincct subsystems are brought a large distance . 14.18) is true independent of the strength of the interaction. Koopman’s theorem is exact. (5. In Fig. an eﬀect missed by LDA.18) is true for both the interacting wavefunction and an independent particle description. In particular. we illustrate this eﬀect. both exact and in LDA We will see below that this condition is violated by all our approximate functionals. vXC (r) Exercise 67 Koopman’s theorem for 1d He For the accurate HF 1d He calculation. this implies vX (r) = −1/r. To see why. Exercise 68 Asymptotic behavior of the density From the information in this and earlier chapters. so too will the LDA potential.4 Ne atom exact LDA 0. Of course. vC (r) → −α(N − 1)/2r 4 (r → ∞) (14. How does LDA vX (r) behave at large distances? As mentioned earlier. its asympototic decay is the same for any value of λ. (5. 14. the following approximate theorem was ﬁrst noted by Koopmans (who later won a Nobel prize in economics).2.1) is the eigenvalue of the highest occupied orbital.3 Derivative discontinuities It is straightforward to argue for the exact asymptotic behaviour of the exchangecorrelation potential for a Coulombic system.2) r Figure 14.6 0. and at large distances vH (r) = N/r. meaning the change in the selfconsistent potential. LDA also misses the shell structure between the 1s and 2s and 2p electrons. Since this eﬀect will occur in any wavefunction producing the exact density.e. But. (14. when an electron is removed from the system. Since the density is ﬁxed. you ﬁnd I HF = E HF (N − 1) − E HF (N ) ≈ − where N N (14. its just − HOMO . 5.
or vice versa? 3. in contact with a bath of electrons.5 2 1 0 1 2 3 4 5 2 vS (x) 4 r Figure 14.2.4 Questions 12 1 0 1 2x 3 4 5 Figure 14.3: XC potential in the Ne atom. Which is better. The apparent step in the potential between the atoms occurs where the dominant exponential decay changes. 14. the step must be the diﬀerence in IE. for e. is the electron aﬃnity. Elementary math then tells you that the step will be the size of the diﬀerence in ionization potentials between the two systems. However. In our world of 1d illustrations.3. for L = 4. this is close to (IEA)/2 for this system. eg a metal. One ﬁnds that the step in the potential induced by adding an inﬁnitesimal of charge is IEA. This is needed to ensure KS system produces two separate densities with separate decay constants.3. But the same reasoning can be applied to a single .5 1 1 1.3 0 0. what happens? is shown in Fig.3. to get the right energy and wrong symmetry.3 how close the LDA potential is in the region around r = 1. making up a HeitlerLondon type wavefunction. in the valence region. This rationalizes one feature of the poorlooking potentials within LDA. Comparing the potentials for the Ne atom. and the KS potential must be constructed so that the molecule dissociates into neutral atoms.. NaCl being separated. At large distances. This is the energy cost of transferring an inﬁnitesimal of charge from one system to another. it has been shifted down by 0. yielding a density that is the sum of the isolated atom densities. these steps are due to discontinuities in the slope of the energy with respect to particle number at integer numbers of electons. the interacting wavefunction for the two electrons should become two (unbalanced) Slater determinants. Thus its potential cannot discontinuously change when an inﬁnitesimal of charge is added to the atom.g. the LDA error in the HOMO orbital energy. 14..5 1 1. Since Ne has zero electron aﬃnity. both exact and in LDA. and the reverse for the slightly negatively charged ion. with a smooth derivative. centered on each nucleus.5 0 Ne atom exact LDA0.8.A. So it will be in error about (IEA)/2 for the neutral. For this two electron system.14. 14. we see in Fig. DISCONTINUITIES where the Hatom is at 0 and the He ion is at L.2: Density and KohnSham potential for 1d HHe+ . but with LDA shifted down by 0. 1. DERIVATIVE DISCONTINUITIES 115 116 CHAPTER 14. relevant to the 2s and 2p electrons. since I=0. 0 4 1d HHe+ potential 8 If the functional is generalized to include fractional particle numbers via ensemble DFT. More generally. Is Koopmans’ theorem satisﬁed by LDA? 4. this will be n(x) = exp(−2x) + 2 exp(−4x − L) (14.A. If one continues beyond x = 5 in Fig. For example. where E. What kind of chemical systems will suﬀer from strong selfinteraction error? 2. 14. But LDA is a smooth functional. 0 vXC (r) 1d HHe+ density n(x) 2 0. we ﬁnd the molecular orbital as φ(x) = n(x)/2 and the KohnSham potential by inversion.2. take the HHe+ case.4) atom. This 3 from each other. Hence the name.
10 at 0.2 1. and the long tail includes the 3p electrons.7.2) This is plotted in Fig. The diﬀerent peaks represent the s electrons in each shell.2: rs (r) in Ar atom. (15. 15.5 2 2. Recall the radial density plot of the Ar atom. and ﬁnd that the shells are less obvious.1. But armed with . the 2s produce the peak at rs = 0.3: g1 (rs ) in Ar atom. 0 0.2.88. This has the simple interpretation: g(rs )drs is the number of electrons in the system with Seitz radius between rs and rs + drs . Why is this analysis important? We usually write LDA EXC = d3 r n(r) unif XC (rs (r)). we deﬁne the density of rs ’s: g1 (rs ) = d r n(r) δ(rs − rs (r)). To see the distribution of densities better. 5.15.5 1. these numbers represent (roughly) the maximum rs in a given shell.118 3 2. 15. and 12 at 0. 15. ANALYSIS TOOLS Ar atom Chapter 15 rs (r) 1. 117 3 we integrate this function forward from rs = 0.5. Now we plot the local WignerSeitz radius. Fig. and so satisﬁes: 0 ∞ drs g(rs ) = N.5 1 1. and the 2p produce no peak.1 Enhancement factor 1.5 1 Analysis tools In this chapter.82 is the 3s electrons. 15.3) (15.7. but stretch the 2s peak upto 0. while the valence electrons have rs ≥ 1.5 3 r Figure 15. The peak at about 0. The 1s core electrons live with rs ≤ 1.1) where unif XC is the exchangecorrelation energy per particle in the uniform gas.5 2 2. and produce the ﬁrst peak in g1 . If 45 40 35 30 1.3. we ﬁnd that it reaches 2 at rs = 0. in Fig. (15. 4 at 0.2 Density analysis What we can do is ask how accurately we need to know the uniform gas inputs.5 0 0 0.5 3 rs Figure 15.4 unif FXC (rs ) 0. with a tail stretching toward rs → ∞. we introduce a variety of tools that help analyze how approximate density functionals work.41. rs (r).1 1 uniform gas unpolarized polarized 0 1 2 3 4 5 6 Ar atom g1 (rs ) 25 20 15 10 5 0 rs Figure 15.1: Enhancement factor for correlation in a uniform electron gas as a function of WignerSeitz radius for unpolarized (solid line) and fully polarized (dashed line) cases.2.5 1 1. The core electrons have rs ≤ 1.3 1.5 2 CHAPTER 15.
Clearly.15. LDA must do well for rs ≤ 2. from Fig.3.72/Z).3 15. ENERGY DENSITY 119 120 CHAPTER 15. but its performance for larger rs values is irrelevant. but valence electrons have rs between about 1 and 6. we see that.4) LDA Thus the contribution to EXC for a given rs value is given by g1 (rs ) times the weighting unif factor XC (rs ). 15. Typical rs values are small for core electrons (at the origin. The valence electrons in simple metal solids have rs between 2 and 6. (15. but core relaxations with rs 1 can also contribute. such as atomization of molecules. a hydrogenic atom has rs = 0. to get a good value for the exchangecorrelation energy of the Ar atom. we may rewrite this as LDA EXC = 0 ∞ drs g(rs ) unif XC (rs ). 15.3.4 Energy density Questions . ANALYSIS TOOLS our analysis. These produce the dominant contribution to chemical processes.
. we know that γS . and a second electron of spin σ in d3 r around r . For example. xN ) (16. x ) = (N − 1) n(x) implying: dx n2 (x. since that appears directly in the usual adiabatic connection formula. even for highly inhomogeneous systems. focus more on the potential energy. Now. given that an electron of spin σ has already been found in d3 r around r. r σ )d3 rd3 r is the probability of ﬁnding an electron of spin σ in d3 r around r.7) We will. x ) = kF π Exchangecorrelation hole 1 In this chapter.3) γ(r. To this end.1) which is also known as the (diagonal) secondorder density matrix. x ) = N (N − 1) dx3 .1 Density matrices and holes (16. (16.5) c 2000 by Kieron Burke. To see this. This can never be true for any wavefunction. only its diagonal elements are determined by the density.122 CHAPTER 16. the density matrix in the KohnSham system. x ) = N dx2 .2) and the exact kinetic energy can be extracted from the spinsummed (a. Improved functionals should incorporate this reliability. EXCHANGECORRELATION HOLE For a single Slater determinant. . x ) is a onebody property. . if the ﬁnding of the second electron were completely independent of the ﬁrst event. r )r=r . . x . Although not accurate enough for thermochemistry. x ) = N − 1. where L → ∞) in onedimension. . dxN Ψ(x. xN ) 2 16. 121 .e. . . implies that the remaining conditional probability density integrates up to one less number of electrons. since. diﬀers from the true density matrix.6) Chapter 16 where φiσ (r) is the ith KohnSham orbital of spin σ. . x ) = n(x ). . This quantity has an important physical interpretation: P (rσ. we may write: P (x.4) Note that. All rights reserved. by deﬁnition: dx P (x. x3 . What is the pair density of a oneelectron system? Deduce what you can about the parallel and antiparallel pair density of a spinunpolarized two electron system.a reduced) density matrix: (16. We must ﬁrst return to the manybody wavefunction. x ) = (N − 1) n(x) (16.12) and state this result in words. We will see that in fact. x2 . the fact that we have found one electron already. (16. . to understand better why LSD works so reliably. we deﬁne the pair density: P (x.. . r )) r=r (16.8) We now dig deeper. x ) = δσσ i=1 φ∗ (r)φiσ (r ) iσ (16. . r ) = γ(rσ. however. r ) − γS (r. x) = n(x) (16. This is diagonal in spin. . we explore in depth the reliability of LSD.11) and T =− 1 2 d3 r 2 γ(r. show that γS (x. x2 . We deﬁne the following (ﬁrstorder) density matrix: γ(x. we have the following simple result: Nσ γS (x. like the He atom. sin(kF (x − x )) sin(kF (x + x )) − . x )d3 r (16. 1 2 d3 r 2 (γ(r. r σ ) σσ where n2 (x. x )d3 r is the conditional probability of ﬁnding the second electron of spin σ in d3 r around r . then n2 (x. x )d3 rd3 r = n(x)d3 r n2 (x. Exercise 70 For samespin electrons in a large box (from 0 to L. .9) The diagonal elements of this density matrix are just the spindensities: γ(x. .10) (16. LSD has proven remarkably systematic in the errors that it makes. since the true kinetic energy diﬀers from the noninteracting kinetic energy: TC = − 1 i. Thus it contains information on the correlations among the electrons. although γ(x. understanding the inherent reliability of LDA is the ﬁrst step toward useful generalized gradient approximations.k. . dxN Ψ∗ (x. kF (x − x ) kF (x + x ) (16. Exercise 71 Show that dx P (x. xN )Ψ(x . Prove this density matrix has the right density. and extract tS (x) from it.
nXC (x. Thus it is convenient to subtract this oﬀ. r σ ). r ) r − r  (16.17) (16.1: Cartoon of hole in onedimensional exponential density.e.23) (16.14) x Figure 16. and their diﬀerence. For a single Slater determinant. the conditional density n2 (x = 2. the hole density. x )2 yielding nX (x. (16. r ) = σσ The ﬁrst (direct) term clearly yields simply the product of the spindensities.. (16. x ) = −γS (x. Plotted are the actual density n(x ) (long dashes). DENSITY MATRICES AND HOLES 123 124 CHAPTER 16. and is known once the pair density is known: 1 Vee = dx dx P (x. x )2 /n(x). Since exchange arises from a wavefunction. It is best to think of the hole as a function of u = x − x.22) This is not a very strong restriction. which is less trivial for the hole: nXC (r.19) P (r .20) but the hole is not. It is often useful to deﬁne u = r − r as the distance away from the electron. since the Coulomb repulsion does not. u) = −1 (16. our energies depend only on the spinsum. the potential energy of an interacting electronic system is a twobody operator. the exchangecorrelation energy is simply the Coulomb interaction between the charge density and its surrounding exchangecorrelation hole. (16. r) = P (r. while the spindecomposition of the hole is interesting. The hole is often (but not always) deepest at the electron point. as the electrons avoid the electron at x. we won’t do that here. x ) (dashed line). while the second (exchange) term can be expressed in terms of the density matrix: PX (x.26) . x ) (solid line). x ) ≥ −n(x ). one can show UXC = d3 r n(r) d3 u Nσ N σ PX (x. does not need to be approximated in a KohnSham calculation.1). and decays This shows that the exchange hole is diagonal in spin (i. it satisﬁes the sumrule. and since this is an explicit density functional. Again. and consider the hole as function of r and u.21) 2u Thus we may say.. nXC (x = 2.e. r − r  (16. in a wavefunction interpretation of density functional theory. r ) = σσ 5 4 3 cartoon n(x ) 2 1 0 1 6 4 2 0 2 4 6 P (rσ. These ideas are illustrated in Fig. and arises from the KohnSham wavefunction. r ) . We deﬁne the (potential) exchangecorrelation hole density around an electron at r of spin σ: P (x. (16.25) The exchange hole gives rise to the exchange energy: EX = 1 2 d3 r d3 r n(r)nX (r. and integrates to exactly −1: dx nXC (x. a large chunk of Vee is simply the Hartree electrostatic energy U . the Hartree energy. x ) /n(r).16. especially for many electrons. EXCHANGECORRELATION HOLE Why is this an important quantity? Just as the kinetic energy can be extracted from the ﬁrstorder density matrix.16) rapidly with distance. x ) = n(x)n(x ) − γS (x. (16. x ) = n(x) (n(x ) + nXC (x. so that d3 u nX (r. r ). The product of densities in P (r. (16. x )) . while nXC (r.15) As we have seen. n(x)n(x ). and is often called simply the pair density. (16. u) .18) (16. The hole is usually (but not always) negative.13) 2 Although it can be very interesting to separate out the parallel and antiparallel spin contributions to the pair density.1. r ) gives rise to U . The exchange hole is a special case. The reduced pair density is the spinsummed pair density: P (r. x )/r − r  (16. x ) = −1 Since the pair density can never be negative. i. only likespins exchange) and is everywhere negative. Then Vee = 1 2 d3 r d3 r P (r.24) The pair density is a symmetrical function of r and r : n(x)nXC (x. x ) = i=1 j=1 φ∗ (r)φ∗ (r ) (φiσ (r)φjσ (r ) − φiσ (r )φjσ (r)) iσ jσ (16. distance from the electron point.
HOOKE’S ATOM 125 126 CHAPTER 16. x ) = P (x. Exercise 73 Holes for oneelectron systems For the H atom (in 3d). but can make some results easier to state. Exercise 74 Hooke’s atom: Approximate HF Repeat the approximate HF calculation of chapter ?? on Hooke’s atom. Since the exchange hole satisﬁes the sumrule. this is just n(x + u).28) g(x. u)/duu=0 = g sph. How big was your error in your estimated energy above? Make a rigorous statement about the correlation energy. u) = 0.32) where the error function is 2 erf(x) = √ π √ and C −1 = 2 5π + 8 π Exercise 78 Hooke’s atom: Electronelectron cusp Repeat the above exercise for the paircorrelation function. u = 0). the approach to unity diﬀers between metals and insulators. both exchange and exchangecorrelation. An alternative way of representing the same information that is often useful is the paircorrelation function: (16.36) leading to nX (u) = −n sin2 (kF u)/(kF u)2 . For any oneelectron system. where n(x) = exp(−2x). The exact density is: √ 2 2C exp(−r 2 /2)/r × n(r) = √ √ 7r + r3 + 8r exp(−r 2 /2)/ 2π + 4(1 + r 2 )erf(r/ 2) (16.29) where the superscript denotes a spherical average in u.av.16. u) and nC (r. (r. with x = 0. at small separations. For example. due to the screening eﬀect. Exercise 75 Hooke’s atom: Separation of variables Show that. (16.2. the pair density vanishes. Φ(R) is the groundstate orbital of a 3d harmonic oscillator of mass 2. It is useful because it is exactly solvable.31) with k = 1/4. using appropriate orbitals. r ) = −n(r ) (N = 1). but not so for ﬁnite systems.(16.31) This means the correlation hole has both positive and negative parts. Where is the electronelectron cusp? What happens as u → ∞? 16. the correlation hole must integrate to zero: d3 u nC (r. the Coulomb interaction between the electons dominates. u) for r = 0 and 1. the groundstate wavefunction may be written as Ψ(r1 . The correlation hole is everything not in the exchange hole. plot P (r. r2 ) = Φ(R)φ(u).3 Transferability of holes x 0 − 2 u + ku2 1 + φ(u) = 4 u u φ(u) (16. so that many ideas can be tested. x )/(n(x)n(x )) This paircorrelation function contains the same information as the hole. Hooke’s atom consists of two electrons in a harmonic well of force constant k. EXCHANGECORRELATION HOLE Exercise 72 Spinscaling the hole Deduce the spinscaling relation for the exchange hole.2 Hooke’s atom dy exp(−y 2 ) (16. (r. at large separations. On the other hand. 16. g → 1 as u → ∞ in extended systems. (16. and is given by the solid curve in Fig.2. for two electrons in external potential vext (r) = kr2 /2.av. Furthermore. From Eq. and also (more importantly). we see that the density matrix becomes.30) The pair density looks very diﬀerent from one system to the next. keeping r parallel to r. Make a deﬁnite statement about the groundstate and HF energies. x to be large. so nX (r. n2 (r. unif γS (u) = n sin(kF u)/(kF u) (16. r ) for several values of r as a function of r . Similarly.33) A useful and interesting alternative external potential to the Coulomb attraction of the nucleus for electrons is the harmonic potential. with a Coulomb repulsion.27) where R = (r1 + r2 )/2. due to the singularity in the electronelectron repulsion there. But let us consider two totally diﬀerent systems: the 1d Hatom and the (samespin) 1d uniform electron gas. It also has a universal cusp at u = 0. with u parallel to r. u = r2 − r1 .34) For our 1d H atom. (16.7). many general concepts can be illustrated. r ). and occasionally the sum of exchange and correlation can be positive. leading to the electronelectron cusp in the wavefunction: dg sph. r ) and nX (r. and φ(u) satisﬁes Exercise 76 Hooke’s atom exactly Show that the function φ(u) = C(1 + u/2) exp(−u2 /4) satisﬁes the Hooke’s atom equation (16. Exercise 77 Hooke’s atom: Exchange and correlation holes Use the exact wavefunction for Hooke’s atom at k = 1/4 to plot both nX (r. and with Coulomb interaction.35) (16. 16. we can deduce the exchange hole for the uniform gas from the bulk value of the ﬁrstorder density matrix. and ﬁnd the exact groundstate energy. taking x. (16. in the bulk.
for a density of n = 1. Thus. EXCHANGECORRELATION HOLE nX (0. both exactly and in LDA.3. But at x = 0. for the uniform gas.4 1d H atom X LDA 0 1 2 (x) = ∞ −∞ du vee (u) nX (x. u) 0. and integrate to 1. i. the local approximation to the hole at any point is the hole of a uniform electron gas. This ﬁgure resembles that of the real He atom given earlier.2: Exhange hole at center of onedimensional Hatom.e. for repulsion exp(−2u).37) (x) 0. At this point. u). unif (16.2 suggests this will be a pretty good approximation.5 exactly. we ﬁnd (0) = −0. Figure 16. Then the energy per electron due to the hole is just X X nLDA (x. In fact.6 with Since. u) 0.40) LDA we may consider EX [n] as arising from a local approximation to the hole at x.4 0. We have shown above how to understand qualitatively why LDA would give results in the right ballpark.5. EX = −0. unif X x dx n(x) X EX = (x). Comparison of the uniform gas hole and the Hatom hole is instructive.5 of onedimensional Hatom. whose density is the density at the electron point.562 in LDA. we get Fig. the LDA overestimates the contribution.4.38) (16. as there cannot be any preferred direction. while for large x. Fig 16. even though their pair densities are utterly diﬀerent.375 exactly. So both holes are normalized.. both exactly and in LDA. Exercise 79 Ontop exchange hole Show that the ontop exchange hole is a localspin density functional. x + u). our x = 0 case is more an anomaly than typical. and since the exact ontop hole is just −n(x).3: Exhange hole at x = 0. in the following way: u Figure 16.4 0. the usual 10% underestimate in magnitude by LDA. we will imagine a 1d world in which the repulsion is exp(−2u).2 127 128 CHAPTER 16. x + u) = nunif (n(x). If the electronelectron repulsion is taken as δ(u). They are also equal at the ontop value. it underestimates it. TRANSFERABILITY OF HOLES 0 0. In fact. then LDA would be exact for X in that 1d world. But since the real Coulomb interaction in 3d is not a contact interaction. Looked at more 0 i. for our exponential repulsion. The . (n) = ∞ −∞ du vee (u) nX (n. 0 0. 16.8 1 1d Hatom exact LDA 3 2 1 0 1 2 3 u Figure 16. They also are symmetric in u. 16. both exactly and in LDA. X X (16. and −0. the general similarities tell us that LDA should usually be in the ball park. because the overall hole shapes are similar. They are remarkably similar.. Why are they so similar? Because they are both holes of some quantum mechanical system. and give an expression for it. u).351 in LDA. Then.16.8 1 3 2 1 0 1d Hatom exact LDA 1 2 3 −δ(u) as we have used in the past. if v ee (u) = closely. the density at the origin in the 1d Hatom.6 0.39) 0. Why is this important? We may think of the local approximation as an approximation to the hole.e. as the density is constant in the system. We ﬁnd. and = −0.2. while U diverges for the uniform gas. U = 1/8 for the 1d Hatom. u = 0. This hole is also plotted in Fig. ∞ −∞ (16. But can we use this simple picture to understand quantitatively the LDA exchange? If we calculate the exchange energy per electron everywhere in the system.2 nX (0.6 0. These depend only on the separation between points.2 0. our naive hopes are dashed.4: Exhange energy per electron in the onedimensional Hatom. we do well to notice the diﬀerence in detail between the two holes.
4 0. we plot the two holes at x = 0. because the density does. i. TRANSFERABILITY OF HOLES 129 130 CHAPTER 16. since EX = 0 du vee (u) nX (u) (16. In Fig. nsym (x.5 1 1. Now watch what happens when we move the hole point oﬀ the origin.368 exactly). For small u. Note how LDA underestimates the magnitude at moderate u. large x. Note that. the rapid decay of the exact hole leads to a smaller energy density than LDA.1 are. being a power law times an oscillation.e. but the longrange is not. Now the 10% underestimate is clearly seen. u) + nX (x.16.3. 16. However. The hole remains centered on the nucleus.5. Note how both rise smoothly from the same ontop value.4 0. the Hatom hole decays exponentially.7.7).2 0..3 0.5 1 1d Hatom exact LDA 1. as the density cannot be very diﬀerent at x + u from its value at x. We can understand this as follows.42) (nX (x. (To see the similarities with real systems. since its length scale is determined by 1/kF .1 nX (u) 0. Note how no cusps remain in the exact hole. both positive and negative values contribute equally. we ﬁnd Fig. which is now at u = −0. a local approximation can be very accurate. Near the nucleus it has a cusp.5 2 2. while all the details of the hole are clearly not wellapproximated in LDA. missing from the uniform gas hole. both exactly and in LDA. On the other hand. with no cancellation of errors throughout the curve. u) = X 0 0. At large distances.5. X (16. This is entirely true in the uniform gas. the cusp at the nucleus is clearly missing from the LDA hole. For an integral weighted exp(−2u). we will see that for most many electron systems.5: Symmetrized exhange hole at x = 0. although it is still true that both holes are normalized to −1. and 0.6 0 0. In Fig. especially in a highly inhomogeneous system. for . and still shows up in the exact hole. The LDA curve only loosely resembles the exact curve. and the maximum deviation is much smaller. 16. since it does not change position with x.5. especially as we move around in a ﬁnite system. and that the ontop values still agree. whose hole is always symetrically placed around the electron point. But as u increases. To begin with. while the exact hole never changes. we now show that the important averages over the hole The systemaveraged symmetrized hole is extremely wellapproximated by LDA. 17. so that the LDA hole becomes very diﬀuse. in fact.5 3 where now u always taken to be positive. as seen in this ﬁgure. By plotting it as a function of u. but Fig. compare this ﬁgure with that of Fig. n(x) decays exponentially. where kF = πn. and overestimates X (x).5 3 Figure 16. so we may write: EX = where ∞ −∞ dx n(x) 0 ∞ du nsym (x. When weighted by the electron density. These oscillations are the same Friedel oscillations we saw in the surface problem. to deduce the contribution to the exchange energy.2 0. we plot the resulting exchange energy/electron throughout the atom. This causes it to deviate quickly from the uniform gas hole near u = 0.5 1d Hatom exact LDA 0 0. but simply moves further away from the electron position. it is the systemaveraged symmetrized hole that appears in EX . multiplying by vee = exp(−2u). both exactly and in LDA.6: Systemaveraged symmetrized exhange hole of onedimensional Hatom.4.5 2 2. Finally. the hole is typically quite dynamic.43) nsym (0. u). u) X 0. EXCHANGECORRELATION HOLE exact Hatom hole contains a cusp at u = 0. ?? shows that it improves the agreement with the exact case: both holes are now parabolic around u = 0.41) 1 (16. there is a large cancellation of errors between the region near the nucleus and far away.3 0. 0 0. because of the systemaveraging.44) u Figure 16. 16.312 in LDA. But. since it is already symmetric.5 of onedimensional Hatom.3. The Hatom hole is said to be static. the density at x + u could diﬀer greatly from that at x. u) X (16. −u)) 2 This symmetrizing has no eﬀect on the LDA hole. So the shortranged hole is wellapproximated. In fact. As x gets even larger.5 0. which leads to the characteristic underestimate of LDA exchange energies. the strong diﬀerence in shape leads to more diﬀerent values of (x) (0. there is a simple shift of origin. nX (u) = ∞ −∞ ∞ dx n(x)nsym (x. while the uniform gas hole decays slowly. u Our last and most important step is to point out that. and follows the position where the ﬁrst electron was found.
because the uniform gas is an interacting manyelectron system.46) z2 0 But the constraint of the sumrule and the wonders of system. we plot the systemaveraged ontop hole for several systems where it is accurately known. Notice also that this explanation requires that uniform gas values be used: Nothing else implicitly contains the information about the hole. the oscillating powerlaw LDA behavior is completely diﬀerent from the exponentially decaying exact behavior: Exercise 80 Show that the exact systemaveraged symmetrized hole in the 1d H atom is: where nX (u) = − exp(−2u) (1 + 2u)..e.8.5 0 0. weighted by exp(−2u). r) = −n(r) (16. (16. For these purposes. Then. the cusp condition. Thus the uniform gas exchange hole integrates to 1. both exactly and in LDA. ζ(r). (16.8: Universal curve for the systemaveraged ontop hole density in spinunpolarized systems. We begin from the ansatz that LSD is a model for the exchangecorrelation hole.and sphericallyaveraged hole within LSD.16. Then. nLSD (u) = XC (16.48) This then means that. not just the energy density. although not exact in general. This is the true explanation of LDA’s success. by applying the technology of the previous section to the uniform gas. In Fig. while the crosses indicate essentially exact results.4 Old faithful Figure 16. u). being the hole as a function of XC 3 separation u of a uniform gas of density (4πrs /3)−1 and relative spinpolarization ζ. It has been shown to be highly accurate for exchangecorrelation for most systems. the ontop hole (u = 0) is very well approximated within LSD. the ontop exchange hole is exact in LSD.25 0. ζ(r)) = N XC 0 duu nLSD (u) . u). we know the potential exchangecorrelation energy density is given in terms of this hole: uunif (rs . the ontop hole becomes as deep as possible (this makes P (r. nX (r. nXC (r.4. d3 r uunif (rs (r). u at u = 0 is also highly accurate. and its correlation hole integrates to zero. implies that the ﬁrst derivative w. 16. u) d3 r n(r) N 4π XC is the system.49) 0. The solid curve is for the uniform gas. which is also satisﬁed by the uniform gas. r) = 0). it is useful to deﬁne the . ζ. the errors it makes are very systematic. XC ∞ (16.50) We are now in a position to understand why LSD is such a reliable approximation. Why should this hole look similar to the true hole? Firstly. ζ) = 2π XC LSD UXC [n] = 0 ∞ i.5 Figure 16.47) (16..t. 16. r) = − n2 (r) + n2 (r) /n(r) ↑ ↓ (16. for any fully spinpolarized or highlycorrelated system. its exchange hole can never be negative. XC i. again making it exact in LSD. OLD FAITHFUL 0 131 132 CHAPTER 16.7: Systemaveraged symmetrized exhange hole of onedimensional Hatom.45) 2 z dz sin2 (z )/z . Furthermore. u large u. The circles indicate values calculated within LSD. ζ.e. For example. and rarely very large. f (z) = − (= f (πu) LSD). its hole satisﬁes the same conditions all holes satisfy.51) duu nunif (rs . and the plus signs indicate less accurate CI results. with an accurate ontop value. While it may not be accurate enough for most quantum chemical purposes.(in the 3d case) averaging lead to a very controlled extrapolation at large u. We denote this hole as nunif (rs . Also. for an inhomogeneous system.and spherical. note that.r. EXCHANGECORRELATION HOLE nX (u) exp(−2u) where 1 dΩu unif n (rs (r). Furthermore. in exchange.5 1d Hatom exact LDA 1 1.
and so our ability to make an accurate estimate using only this information suﬀers. This then is the explanation of LSD’s reliability. But in the former. Indeed. 0.04 d3 r n2 (r) rs (r) . the accuracy of LSD can fail in regions of extreme gradient. The area under each curve is the exchange energy.3 0. But in LSD. The exchangecorrelation hole in the uniform gas must be spherical. we will show how to use the failed gradient expansion to improve the description of the hole. whereas the true hole is often highly aspherical. the systemaveraged ontop hole is very accurate in LSD. so that LSD is completely incorrect for this quantity. 0 0. Thus.4. OLD FAITHFUL 133 134 CHAPTER 16. But even at large u. Each is multiplied by a factor of 2πu. That LSD is most accurate near u = 0 can be easily understood physically. Thus limitations of LSD in extreme situations do not contribute strongly to the exchangecorrelation energy. EXCHANGECORRELATION HOLE systemaveraged density: 2πu nλ=1 (u) C 1 N and a systemaveraged mean rs value: n(u) = rs dΩu 4π d3 r n(r) n(r + u). its behavior is constained by the hole normalization sumrule. this implies that their exact systemaveraged hole looks very like its LSD approximation.2 0. embedded in it. at large separations the paircorrelation function has qualitatively diﬀerent behavior in diﬀerent systems. The energydensity of the uniform gas contains.916. in LDA (dashed) and exactly (HF solid).5. The same argument applies to large separations: even if g is badly approximated by LSD. both the sumrule and the accurate ontop hole information. both exactly and in LSD. = d3 r n2 (r) (16. But this is irrelevant. both exactly and in LSD. and this vanishes rapidly. so that its area is just the exchange or correlation energy contribution (per electron). In Figs. This is yet another explanation of the large overestimate of LSD correlation.10: Systemaveraged correlation hole density at full coupling strength (in atomic units) in the He atom. For . and typically underestimates the hole in the allimportant moderate u region. r + u) that appears in the hole. Two other points are salient.5 He atom HF LDA 0 1 2 3 most systems. in LSD. 16. but never with very large diﬀerences. the density itself is exponentially small. with only small diﬀerences in details. while ﬁnally dying oﬀ too slowly at large u.52) 0. and exactly (CI). it is −n(r + u)g(r. our only inputs are the spindensities at r. numerical GGA.53) 0.9: Systemaveraged exchange hole density (in atomic units) in the He atom. as indeed it is. whereas the true hole has a signiﬁcant positive bump near u = 1. Furthermore. we may expect LSD to be most accurate for small u. the phasespace in the systemaverage is small. At u grows r gets further and further away from r. In the next chapter. 2πu nX (u) u Figure 16.4 0. as mentioned above.10. such as near a nucleus or in the tail of a density. The area under each curve is the full couplingstrength correlation energy.1 He atom exact LDA 3 u Figure 16.16. and so construct a generalized gradient approximation.08 0 1 2 Clearly. For correlation. by symmetry. we plot the systemaveraged exchange and potentialcorrelation holes for the He atom. 0 (16. We see that LSD is exact for small u exchange. we see that the LSD hole is always negative in the energetically signiﬁcant regions. since it is only the sphericalaverage that occurs in EXC . while in the latter.
Part IV Beyond LDA 135 .
138
CHAPTER 17. GRADIENTS
Chapter 17
Gradients
In this chapter, we explore the ”obvious” correction to the local approximation, namely the inclusion of gradient information. We’ll see that its not as easy as it looks, and why it took about a quarter of a century to accomplish. 17.1 Perimeter problem
n 1 2 4 8 16 32 64
0.2 0.2 0.2 0.2 0.2 0.2 0.2
local % error exact 6.2832 1 6.3462 6.2832 4 6.5297 6.2832 13 7.1989 6.2832 32 9.2984 6.2832 57 14.7104 6.2832 77 26.7835 6.2832 88 51.9081
= 0.2, and various n, both exactly and in local approximation.
Table 17.2: Perimeter’s of shapes for
One could imagine a clever person, not knowing the true formula, being able to prove an exact condition like: (17.3) P loc ≤ P We also notice that the error increases with . Obviously = 0 is a circle, where the local approximation is exact. For small , the curve is in some sense close to a circle, and the local approximation should be accurate. In Fig. 17.1, we plot the = 0.8 case, and see that,
n=1, eps=0.8
To illustrate the interesting complexities of the use of gradients, we return to our problem from Chapter 2, in which we needed to know the perimeter of a shape, given its deﬁnition in terms of polar coordinates, r(θ). In all of what follows, we assume we do not know the exact answer. We already saw in chapter 8 that the local approximation is P loc =
2π 0
dθ r(θ).
(17.1)
Let us ﬁrst consider the family of smooth curves introduced in that chapter: r(θ) = 1 + cos(nθ) (17.2)
Figure 17.1: Shape r = 1 + 0.8 cos(θ).
Here 0 ≤ < 1, while n is an integer. For these curves, the local approximation yields exactly 2π for the perimeter. In table 17.1, I have compared results for n = 1, as a function of . The local approximation is doing rather well, with a maximum error of −15%. Notice how its always an underestimate. n 1 1 1 1 1 0.0 0.1 0.2 0.4 0.8 local % error 6.2832 0.00 6.2832 0.25 6.2832 0.99 6.2832 3.88 6.2832 14.37 exact 6.2832 6.2989 6.3462 6.5371 7.3376
Table 17.1: Perimeter’s of shapes for n = 1, and various , both exactly and in local approximation. 137
although the radius varies from 0.2 to 1.8, the curve looks quite circular. On the other hand, in Table 17.1, we ﬁx = 0.2, and increase n. The error grows almost quadratically with n. Again, n = 0 is the circular case, and so small n is somehow more circular than large n. To understand the ever increasing error, we plot the n = 64 curve in Fig. 17.2. The local approximation yields the perimeter of a circle with the average radius, r = 1. Obviously, the 64 wiggles between r = 1.2 and r = 0.8 as one goes round the circle are not accounted for in the local approximation, but add greatly to the perimeter. Thus we can speak of two distinct ways in which a curve might be ’close’ to a circle. The ﬁrst, more familiar way, is when is small, and so the perturbation on r = 1 is weak. This can
17.1. PERIMETER PROBLEM
139
140
CHAPTER 17. GRADIENTS
n=64, eps=0.2
Figure 17.2: Shape r = 1 + 0.2 cos(64 θ).
n 1 1 1 1 1 2 4 8 16 32 64
0.0 0.1 0.2 0.4 0.8 0.2 0.2 0.2 0.2 0.2 0.2
local % error GEA % error exact 6.2832 0.00 6.2832 0.00 6.2832 6.2832 0.25 6.2989 0.00 6.2989 6.2832 1 6.3467 0.01 6.3462 6.2832 4 6.5455 0.1 6.5371 6.2832 14 7.5398 3 7.3376 6.2832 48 6.5371 0.1 6.5297 6.2832 13 7.2988 1 7.1989 6.2832 32 10.3455 11 9.2984 6.2832 57 22.5323 53 14.7104 6.2832 77 71.2795 166 26.7835 6.2832 88 266.2640 400 51.9081
= 0.2, and various n, both exactly and in local approximation and in GEA.
usually be treated by response theory. Perturbation theory will always be accurate once the pertubation is suﬃciently week, no matter how rapidly the curve is varying (i.e., no matter how large n is). The second, less familiar way is when the perturbation is slowlyvarying, but can be arbitrarily large. This is the case when n is small, but need not be. In Fig. 17.1, the local approximation works quite well, even though is enormous. The ratio of maximum to minimum r is 9! Let us ﬁrst analyze the weak perturbation case. We can cheat by using our knowledge of the exact functional, but I’m sure there are ways to derive the result without it. If r = r 0 + f (θ), then 2 2π df 2 P = 2πr0 + dθ + O( 4 ) (17.4) 2r0 0 dθ For our standard curve, the integral is n2 π. There is no linear term, because it vanishes by periodicity requirements. You will ﬁnd this formula quite accurate, even at = 0.2, and hence the (near) quadratic growth in the error in the local approximation in Table 17.1. Note that, in the local approximation, there is no 2 term. Thus the local approximation, while being exact for the circle, is hopeless for weak perturbations around the circle. The other good approximation is called the gradient expansion. We ﬁrst note that the only way to make a dimensionless gradient is by dividing the derivative by the radius. We deﬁne s = dr/dθ/r. Then, for a slowlyvarying shape, we can write P = dθr (1 + C s + . . .)
2
Table 17.3: Perimeter’s of shapes for
1 2π 1 dr 2 dθ (17.6) 2 0 r dθ In Table 17.1, we have added the results of the gradient expansion. We see that for n = 1, where the local approximation was quite good, GEA reduces the error by at least a factor of 4, and sometimes much more. It also overestimates the perimeter in all cases. Even up to n = 8, for = 0.2, the error is still reduced by a factor of 3. But for larger n, meaning larger gradients, the GEA overcorrects, eventually producing larger errors than the local approximation! Our conclusion is that, for slowlyvarying shapes, the gradient expansion greatly improves on the local approximation, but does not work for rapid variations, and can even worsen the results. To make the point very clear, suppose we live in a world where most of the shapes we care about have n between 2 and 8, while varies between .4 and .8. The results of the local and gradient expansion approximations are listed in Table ??. Only for the most slowlyvarying cases does the GEA really improve over the local approximation. In all cases, it overcorrects, usually by more than the original error. For these systems, the GEA is hardly an improvement. P GEA = P loc + 17.2 Gradient expansion
GEA in this case is
(17.5)
where C is yet to be determined. There is no term linear in s, because its integral would vanish. Thus the GEA, or gradient expansion approximation, consists of keeping just the ﬁrst two terms, once C has been found. Finding C is easy, once we know the linear response. We simply imagine a perturbation that is both weak and slow. Then s = dr/dθ/r0 , and we see that C must be 1/2. Thus the
Way back when, in the original KohnSham paper, it was feared that LSD might not be too good an approximation (it turned out to be one of the most successful ever), and a simple suggestion was made to improve upon its accuracy. The idea was that, for any suﬃciently slowly varying density, an expansion of a functional in gradients should be of ever increasing accuracy: (17.7) AGEA [n] = d3 r a(n(r)) + b(n(r)) n2 + . . .
17.2. GRADIENT EXPANSION
141
142
CHAPTER 17. GRADIENTS
n 2 2 2 4 4 4 6 6 6 8 8 8
0.4 0.6 0.8 0.4 0.6 0.8 0.4 0.6 0.8 0.4 0.6 0.8
local % error 6.28 13 6.28 24 6.28 34 6.28 33 6.28 48 6.28 58 6.28 47 6.28 61 6.28 70 6.28 57 6.28 70 6.28 77
GEA % error 7.33 2 8.80 7 11.31 18 10.48 12 16.34 36 26.39 76 15.73 32 28.90 77 51.52 146 23.07 56 46.50 124 86.71 221
exact 7.22 8.26 9.55 9.34 12.04 15.00 11.94 16.32 20.93 14.76 20.80 27.04
and n, both exactly and in local approximation and in GEA.
For correlation, the procedure remains as simple in principle, but more diﬃcult in practice. Now, there are nontrivial density and spin dependences, and the terms are much harder to calculate in perturbation theory. In the highdensity limit, the result may be written as 2 GEA = 2 d3 r n(r)φ(ζ(r))t2 (r). (17.11) ∆EC 3π Here t =  n/(2ks n), where ks = 4kF /π is the ThomasFermi screening length, the natural wavevector scale for correlations in a uniform system. The spinpolarization factor is φ(ζ) = ((1 + ζ)2/3 + (1 − ζ)2/3 )/2. This correction behaves poorly for atoms, being positive and sometimes larger in magnitude than the LSD correlation energy, leading to positive correlation energies! Exercise 82 GEA correlation energy in H atom GEA Calculate the GEA correction to the H atom energy, and show how ∆EC scales in the highdensity limit. 17.3 Gradient analysis
Table 17.4: Perimeter’s of shapes for various
Then, if LSD was moderately accurate for inhomogeneous systems, GEA, the gradient expansion approximation, should be more accurate. The form of these gradient corrections can be determined by scaling, while coeﬃcients can be determined by several techniques. For both TS and EX , deﬁned on the KohnSham wavefunction, the appropriate measure of the density gradient is given by s(r) =  n(r)/(2kF (r)n(r)) (17.8)
Having seen in intimate detail which rs values are important to which electrons in Fig. 15.2, we next consider reduced gradients. Fig. 17.3 is a picture of s(r) for the Ar atom, showing
2
Ar atom
1.5
This measures the gradient of the density on the length scale of the density itself, and has the important property that s[nγ ](r) = s[n](γr), i.e., it is scale invariant. It often appears in the chemistry literature as x =  n/(n4/3 ), so that x ≈ 6s. The gradient expansion of any functional with powerlaw scaling may be written as: AGEA [n] = d3 r a(n(r)) 1 + Cs2 (r) (17.9)
s(r)
1
0.5
0
0
0.5
1
1.5
2
2.5
3
The coeﬃcient may be determined by a semiclassical expansion of the KohnSham density matrix, whose terms are equivalent to a gradient expansion. One ﬁnds CS = 5/27, CX = 10/81 (17.10)
r
Figure 17.3: s(r) in Ar atom.
how s changes from shell to shell. Note ﬁrst that, at the nucleus, s has a ﬁnite moderate value. Even though the density is large in this region, the reduced gradient is reasonable. Exercise 83 Reduced gradient at the origin Show that at the origin of a hydrogen atom, s(0) = 0.38, and argue that this value wont change much for any atom. For any given shell, e.g. the core electrons, s grows exponentially. When another shell begins to appear, there exists a turnover region, in which the gradient drops rapidly, before being dominated once again by the decay of the new shell. We call these intershell regions.
(In fact, a naive zerotemperature expansion gives CX = 7/81, but a more sophisticated calculation gives the accepted answer above). The gradient correction generally improves both the noninteracting kinetic energy and the exchange energy of atoms. The improvement in the exchange energy is to reduce the error by about a factor of 2. Exercise 81 For the Hatom, calculate the gradient corrections to the kinetic and exchange energies. (Don’t forget to spinscale).
and make the corresponding g1 (rs ) and g3 (s) plots. there is no simple monotonic function. Because s is not monotonic. Note there is no contribution Ar atom g3 (s) 25 20 15 10 5 0 0 0. In the g3 (rs ) ﬁgure. many regions in space can contribute to g3 (s) for a given s value. We will denote the region of all r smaller than the position of the last peak in s as being the core region (9.91 electrons in Ar. Then on the right are plotted the distribution of values (turned sideways). containing 6. Why does the exponential blow up at large r not make the LDA or GEA energies diverge? We now divide space around atoms into spheres. We further denote the tail region as being that part of the valence region where s is greater than its maximum in the core (1. The small region where s drops rapidly is called the corevalence intershell region. The svalues are more complex.5. we can make a plot of the diﬀerences between (twice) the atomic curves and the molecular curve. one can perform a pseudopotential calculation. but one can easily extract the densities of variables. we can draw horizontal lines to meet the vertical lines along the curve.5: Distribution of densities and gradients in the N atom. We see that almost all the density has 0.5. 17. in which the transition occurs (1. and plotted for the Ar atom in Fig. we see the contribution of the valence electrons alone to the peak from the combined corevalence region. (17. and so is irrelevant to the atomization energy.2 ≤ s ≤ 1. Exercise 84 Gradient analysis for H atom Calculate rs (r) and s(r) for the H atom.12) also normalized to N .5. We see that the core electrons have rs ≤ 0. GRADIENTS We deﬁne a density of reduced gradients. precisely the inner core peak. which have been extended to the upper r s curve. The region in which s grows again. 17.5 ≤ rs ≤ 2. Having done this for the molecule. to atoms (or molecules) from regions of s = 0. that replaces the core electrons by an eﬀective potential designed only to reproduce the correct valence density. analogous to our density of rs ’s: g3 (s) = 40 35 30 d3 r n(r) δ(s − s(r)). The remainer is denoted the tail.75 electrons).34 electrons). consider Fig. To better distinguish them. In the case of a molecule. Note also that almost all the density has reduced gradients less than about 1. Because that curve is monotonic. This is marked in the two righthand panels by a longdashed line. But we are more interested in atomization energies than atomic energies. We will (somewhat arbitrarily) denote the valence region as all r values beyond the last minimum in s(r). We make these regions by vertical dashed lines. we see indeed that the pseudopotential is missing Figure 17. we call the valence region. with a long tail stretching out of larger s. GRADIENT ANALYSIS 143 144 CHAPTER 17.5 2 s Figure 17. since they are similar in both the core and valence regions.3. On the left is plotted values of rs (top) and of s (bottom) as a function of r in an N atom.4: g3 (s) in Ar atom.2. Now the curves are both positive and negative. and thus divide the g(rs ) curve into corresponding regions of space. but remains below its previous maximum. To see how big s can be in a typical system that undergoes chemical reactions. while in the g(s) ﬁgure. The last region in which s decreases with r is denoted the corevalence region.17. while the valence electrons have 0.5 1 1.54 electrons). Interestingly. We may use the s curve to make more precise (but arbitrary) deﬁnitions of the various shells. We see immediately that the core does not change on atomization of the molecule. it is also clear that it is not .4. Up to the ﬁrst maximum in s we call the core.
17.3. GRADIENT ANALYSIS
145
146
0
CHAPTER 17. GRADIENTS
X
(r)
0.4 0.8 He atom X LDA 0 1 2 3 1.2
r
Figure 17.7: Exchange energy per electron in the He atom, both exactly and in LDA. The nucleus is at r = 0.
cancellation of errors throughout the system, i.e., LDA is not describing this quantity well at each point in the system, but it does a good job for its integral. As we will show below, a pointwise analysis of energy densities turns out to be, well, somewhat pointless... 17.4 Questions
1. If GEA overcorrects LDA, what can you say about how close the system is to uniform? 2. Do you expect LDA to yield the correct linear response of a uniform electron gas? 3. Sketch the gradient and density analysis for the ionization energy for Li.
Figure 17.6: Diﬀerences in distribution of densities and gradients upon atomization of the N 2 molecule.
only the valence electrons (rs > 0.8) that contribute, but also the corevalence region, which include rs values down to 0.5. Finally, note that the rs values contributing to this energy also extend to higher rs regions. To summarize what we have learned from the gradient analysis: • Note again that nowhere in either the N or Ar atoms (or any other atom) has s << 1, the naive requirement for the validity of LSD. Real systems are not slowlyvarying. Even for molecules and solids, only a very small contribution to the energy comes from regions with s < 0.1. • A generalized gradient approximation, that depends also on values of s, need only do well for s ≤ 1.5 to reproduce the energy of the N atom. To get most chemical reactions right, it need do well only for s ≤ 3. We end by noting that while this analysis can tell us which values of rs and s are relevant to real systems, it does not tell us how to improve on LSD. To demonstrate this, in Fig. 17.7, we plot the exchange energy per electron in the He atom. One clearly sees an apparent
148
CHAPTER 18. GENERALIZED GRADIENT APPROXIMATION
Chapter 18
Generalized gradient approximation
1 In this chapter, we ﬁrst discuss in more detail why the gradient expansion fails for ﬁnite systems, in terms of the exchangecorrelation hole. Fixing this leads directly to a generalized gradient approximation (GGA) that is numerically deﬁned, and corrects many of the limitations of GEA for ﬁnite systems. We then discuss how to represent some of the many choices of forms for GGA, using the enhancement factor. We can understand the enhancement factor in terms of the holes, and then understand consequences of gradient corrections for chemical and physical systems. Finally we mention other popular GGAs not constructed in this fashion.
Figure 18.1: Sphericallyaveraged exchange hole density nsph.av. (u) for s = 1. X
18.1
Fixing holes
We begin with exchange. We have seen that LSD works by producing a good approximation to the system and sphericallyaveraged hole near u = 0, and then being a controlled extrapolation into the large u region. The control comes from the fact that we are modelling the hole by that of another physical system, the uniform gas. Thus the LSD exchange hole is normalized, and everywhere negative. When we add gradient corrections, the model for the hole is no longer that of any physical system. In particular, its behavior at large separations goes bad. In fact, the gradient correction to the exchange hole can only be normalized with the help of a convergence factor. To see this, consider Fig. 18.1, which shows the spherically averaged exchange hole for a given reduced gradient of s = 1. Everything is plotted in terms of dimensionless separation, z = 2kF u. The region shown is the distance to the ﬁrst maximum in the LDA hole, where it just touches zero for the ﬁrst time. The GEA correction correctly deepens the hole at small u, but then contains large oscillations at large u. These oscillations cause the GEA hole to unphysically become positive after z = 6. These oscillations are so strong that they require some damping to make them converge. Without a damping factor, the gradient correction to the exchange energy from this hole is undeﬁned. The realspace cutoﬀ construction of the generalized gradient approximation (GGA), is to include only those contributions from the GEA exchange hole that are negative, and to
1
truncate the resulting hole at the ﬁrst value of u which satisﬁes the sumrule. In the ﬁgure, for small z, the GGA hole becomes more negative than GEA, because in some directions, the GEA hole has become positive, and these regions are simply sliced out of GGA, leading to a more negative sphericallyaveraged hole. Then, at about 2kF u = 6.4, the hole is truncated, because here the normalization integral equals 1, and its energy density contribution calculated. We note the following important points: • Simply throwing away positive contributions to the GEA exchange hole looks very ugly in real space. But remember we are only trying to construct a model for the sphericaland systemaverage. Thus although that has happened in some directions for small z, the sphericallyaveraged result remains smooth. • By throwing out postive GEA contributions, a signiﬁcantly larger energy density is found for moderate values of s. Even for very small s, one ﬁnds corrections. Even when GEA corrections to the LSD hole are small, there are always points at which the LSD hole vanishes. Near these points, GEA can make a postive correction, and so need ﬁxing. This means that even as s → 0, the GGA energy diﬀers from GEA. • For very large s (e.g. greater than 3), the wild GEA hole produces huge corrections to LSD, the normalization cutoﬀ becomes small, and limits the growth of the energy density. However, this construction clearly cannot be trusted in this limit, since even a small distance away from u = 0 may have a completely diﬀerent density. The results of the procedure are shown in Fig. 18.2. We see that the LDA underestimate is largely cured by the procedure, resulting in a systemaveraged hole that matches the exact one better almost everywhere.
c 2000 by Kieron Burke. All rights reserved.
147
18.1. FIXING HOLES
0 0.1
149
150
CHAPTER 18. GENERALIZED GRADIENT APPROXIMATION
0.2 0.3 0.4 0.5 0 1 2 He atom HF LDA GGA 3
u
Figure 18.2: Systemaveraged exchange hole density (in atomic units) in the He atom. The exact curve (solid) is HartreeFock, the long dashes denote LDA, and the short dashes are realspace cutoﬀ GGA. The area under each curve is the exchange energy.
hole. But the holeGGA curve also grows less rapidly at large s, because of the normalization cutoﬀ. Also included in the ﬁgure are two modern GGAs, (Becke 88 and PerdewWang 91) that we discuss more later. They both agree with the general shape of the holeGGA for moderate values of s, making the realspace cutoﬀ procedure a justiﬁcation for any of them. The most signiﬁcant deviation in this region is that the holeGGA has an upward bump near s = 1. This is due to the negativity and normalization cutoﬀs meeting, as is about to happen in Fig. ?? for slightly larger s, and is an artifact of the crude construction. Recalling that GGA is still an approximation, the diﬀerences between diﬀerent GGA’s are probably of the order of the intrinsic error in this approximate form.
There is a simple way to picture the results. We have seen how the spinscaling relation means that one need only devise a total density functional, and its corresponding spindependence follows. Furthermore, since it scales linearly, the only way in which it can depend on the ﬁrstorder gradient is through s, the reduced dimensionless gradient. Thus we may write any GGA for exchange, that satisﬁes the linear scaling relation, as
GGA EX [n] =
2πu nX (u)
d3 r eunif (n(r))FX (s(r)) X
(18.1)
The factor FX is the exchange enhancement factor, and tells you how much exchange is unif enhanced over its LDA value. It is the analog of FXC (rs ) from before. In Fig. 18.3, we plot
1.8
1.6
FX (s)
Figure 18.4: Sphericallyaveraged correlation hole density nC for rs = 2 and ζ = 0. GEA holes are shown for four values of the reduced density gradient, t =  n/(2ks n). The vertical lines indicate where the numerical GGA cuts oﬀ the GEA hole v to make 0 C dv 4π v 2 nC (v) = 0.
1.4 Exchange hole GEA B88 PBE 0 0.5 1 1.5 2 2.5 3
1.2
1
s =  n/(2kF n)
Figure 18.3: Exchange enhancement factors for diﬀerent GGA’s
various enhancement factors. First note that FX = 1 corresponds to LDA. Then the dashed line is GEA, which has an indeﬁnite parabolic rise. The solid line is the result of the realspace cutoﬀ procedure. It clearly produces a curve whose enhancement is about double that of GEA for small and moderate s, because of the elimination of positive contributions to the
The case for correlation is very similar. Here there is no negativity constraint, and the realspace construction of the GGA correlation hole simply corrects the lack of normalization in the GEA correlation hole. Fig. 18.4 shows what happens. The line marked 0 is the LDA correlation hole. Note its diﬀuseness. Here the natural length scale is not kF u, but rather ks u, because we are talking about correlation. Similarly reduced gradients are measured relative to this length scale. The other curves are GEA holes for increasing reduced gradients. Note that the GEA hole is everywhere positive, leading to consistently positive corrections to LDA energies. Note also how large it becomes at say t = 1.5, producing a stongly positive energy contribution. For small reduced gradients, the gradient correction to the hole is slight, and has little eﬀect on the LSD hole, until at large distances, it must be cutoﬀ. As the gradient grows, the impact on the LSD hole becomes much greater, and the cutoﬀ shrinks toward zero. Since the energy density contribution is weighted toward u = 0 relative to this picture,
but then promptly cuts oﬀ correlation as t grows.4) 2ks n ks rs Thus. (18. . 18. the functional becomes more local. so that correlation turns oﬀ very quickly. But the overall eﬀect on the energy is very satisfying. closer to the LDA value. The area under each curve is the potential contribution to correlation energy.6 1.. 18. We will discuss the various kinds of GGA functionals that have been developed and are in use later. while the realspace cutoﬀ hole has produced something similar to the true hole. their magnitude is less. and we see correlation turning oﬀ at t about 0. and yielding a higher exchange energy density. At small separations. This allows the center of the hole to move in that direction. • A simple picture of the correlation eﬀects is given by the observation that for t up to about 0. In Fig. ??.6. For rs = 10.4 1.2 Visualizing and understanding gradient corrections To understand gradient corrections for exchangecorrelation.18. we have FXC (rs ).5.5 2 2. and the shortdashes are realspace cutoﬀ GGA.6: Exchangecorrelation enhancement factors for the PBE GGA. as in Fig.5. where we write GGA EXC [n] = d3 r eunif (n(r)) FXC (rs (r).3 1. We make the following observations: • The eﬀect of gradients is to enhance exchange. while dropping on the opposite side. and also that even when they do. the uniform gas enhancement factor. in regions of high gradient the exchange eﬀect keeps electrons apart. • The eﬀect of gradients is turn oﬀ correlation relative to exchange. the postive contribution is much less pronounced.7s.. we see the results for the systemaveraged hole. The LDA hole spreads out enormously. s = 0) = FXC (rs ). we see that large gradients ultimately eliminate correlation.2 1.2. i.3.5 3 2πu nλ=1 (u) C 0. while for rs = 1/2. for rs about 2. the GEA correclty makes the LDA hole less deep. But now. at s about 0.2. But t is the reduced gradient for correlation. t= unif • Along the yaxis. as expected. s(r)). s) = FX (s).5 = s≈ s (18. instead of producing a wild (i. and that as correlation is turned on.5: Systemaveraged correlation hole density at full coupling strength (in atomic units) in the He atom. FXC (rs = 0. the GGA correlation hole always produces a negative correlation energy density.1 rs 0 1/2 2 4 10 0 0. and remind us of how little we know of the details of the corrections to LDA. so that their correlation energy becomes relatively smaller. For an unpolarized system:  n kF 1.4s. in the positive direction of the gradient. t = s. The upward bump and rapid turnoﬀ beyond u = 2 are artifacts of the crude realspace cutoﬀ procedure. the density is increased.5 CHAPTER 18. • An interesting consequence of the point above is that pure exchange is the least local curve. X (18. t ≈ 0.08 Figure 18. s =  n/(2kF n) u Figure 18.3) We plot it for the case of the PBE functional in Fig 18. giving rise to a correlation energy that is at least a factor of 2 too deep. unnormalized) postive peak. VISUALIZING AND UNDERSTANDING GRADIENT CORRECTIONS 151 152 1. FXC (rs . deﬁned in terms of the screening wavevector. which are even more dramatic than the exchange case. not the Fermi wavevector. becoming deeper due to the higher density. GENERALIZED GRADIENT APPROXIMATION and all curves initially drop. In our realspace analysis of the hole.e. we begin with the spinunpolarized case. The exact curve (solid) is from a CI calculation. both in the sense that gradient corrections become signiﬁcant at larger s values.e. and includes also several other energetically signiﬁcant constraints. the long dashed is LDA. and is controlled by the normalization. We have already seen this in action in the previous sections.5. s) 0 1.2) This enhancement factor contains all our others as special cases: unif FXC (rs .04 He atom exact LDA GGA 0 1 2 3 1 0. This is because. Essentially. so that correlation turns oﬀ about s = 1. the GGA make little correction to LDA. but for now we simply note that this GGA was designed to recover the GGA for the realspace correction of the GEA hole for moderate s.5 1 1. t ≈ 1. These curves contain all the physics (and ultimately) chemistry behind GGA’s.
GGA remains size consistent.3 Eﬀects of gradient corrections will be too low in an LDA calculation. in which the changes are large.18. and so have a smaller atomization energy.6. due to the presence of core electrons. Y. As a bond is stretched. and the right hand side is thought of as zero. one ﬁnds LSD overbinds s electrons relative to p. Thus comparing the transition state to the reactants.5) Then ∆EXC (atoms) < ∆EXC (molecules). • Virial theorem Trivially true.8) . where we do not trust it anyhow. Exercise 85 Deduce the maximum value for FXC allowed by the LiebOxford bound. The region of the chemical bond is ﬂatter than the corresponding regions of exponential decay in atoms.6) We can now look back on the previous chapters. deﬁned as EXC trans The LO bound gives us a maximum value for the enhancement factor of 1. If the transition state is less coordinated. Thus molecules are overbound in LDA. Notice though that this depends on the coordination argument. Must ﬁll this in. and GGA will weaken it. For example.3.7) • The LiebOxford bound. Neglecting the righthandside works for all atomization energies. bridgebonded) has lower symmetry than the initial state (e. and those bonds are shortened. The atomization energy of a molecule is atmiz EXC = EXC (molecule) − EXC (atoms) (18. • Transition state barriers can be understood in much the same way. and maybe a few more. we survey some of the many properties that have been calculated in DFT electronic structure calculations. (18. clear cos only s dependence. we ﬁnd that the transitionstate barrier. For exchange. the result is the reverse: LDA overestimates the barrier. The realspace cutoﬀ will eventually violate this bound. and where P is typically close to 1. and GGA stretches them. A typical transition state in quantum chemistry is one of higher symmetry and coordination than the reactants. and be raised by a GGA calculation. i. One can show that typically GGA favors a process in which ∆ rs ∆s ≥P + Q∆ζ . and try to understand the errors made LSD. The inequality means that GGA will like this more than LDA. The only exception to this is the case of bonds including H atoms. remebering that both are negative. In the case of H atoms. are precisely chosen averages. Exercise 86 The lines of diﬀerent rs values never cross in Fig.e. the density becomes more homogeneous than before. surface diﬀusion.. GENERALIZED GRADIENT APPROXIMATION We can repeat this analysis to understand the eﬀects of spin polarization. when stretching a bond with an H atom. the fractional gradient changes are much smaller. Usually fractional changes in the density are very small. the local approximation is less accurate for them. consider the process of stretching a bond.4 Satisfaction of exact conditions In this section.g. = −EXC (reactants) + EXC (transitionstate) (18. LDA overfavors homogeneity. Since core electrons have higher density than outer ones. • In studying ionization energies of atoms. the gradient corrections have the reverse eﬀect. EFFECTS OF GRADIENT CORRECTIONS 153 154 CHAPTER 18. etc. p relative to d. Ignoring gradients ignores the enhancement of exchange and the truncation of correlation. and so forth. and why GGA corrects them the way it does. This leads to the generic claim that GGA prefers inhomogeneity. when a molecule is atomized. atop bonded). This is often the case for. the density is suﬃciently low as to make its fractional change larger than that of the gradient. typically by as much as 30 kcal/mol. eg. • By the same reasoning as for atomization energies. but only at large s. Thus equilibrium bond lengths are usually too small in LDA. 18. • Both the large underestimate in the magnitude of total exchange energies and the overestimate in the magnitude of correlation energies can be seen immediately from Fig. Write LDA LDA EXC (system) = EXC (system) + ∆EXC (system) (18. But for a bondstretch. LDA’s overbinding tendancy will reduce. where the transition state (e. • Coordinate scaling Show FXC lines never cross. in our cases above.9) s 2 rs where s . • When a molecule is formed from atoms.804. Thus in that case. and check that the realspace cutoﬀ GGA satisﬁes conditions that LDA satisﬁes. in just the way done above for a bond and its constituent atoms. 18. • Sizeconsistency Obviously. Thus LDA makes less of an underestimate of the energy of a molecule than it does of the constituent atoms. What does this imply about the functional? 18.. so that LDA LDA LDA EXC (atom) = EXC (atom) + ∆EXC (molecule) − EXC (atom) (18. and Q ≈ 0.g. its mean gradient increases.
18. and.g. by truncating the longranged Coulomb correlation hole. it will still have a residual selfinteraction error.6 Questions about generalized gradient approximations . This is not unique.. a smoother cutoﬀ will lead to slightly diﬀerent curves for moderate gradients. e. Just as above. From the naive sense. in that. 18. A BRIEF HISTORY OF GGA’S 155 156 CHAPTER 18. nothing in GGAs construction helps here. 18. numerical GGA yields a ﬁnite value. many chemical systems are close to this.. • Selfinteraction error Simple density functionals like LDA and GEA naturally have a selfinteraction error.5. and perhaps much diﬀerent curves for large gradients. • Symmetry dilemma • Potentials • Koopman’s theorem No real improvement here. As we show below. While GGA might improve numerically the value of EX or EC for that one electron.5 A brief history of GGA’s We have seen how the realspace cutoﬀ construction imposed on the gradient expansion produces a numericallydeﬁned GGA. GENERALIZED GRADIENT APPROXIMATION • Highdensity correlation The numerical GGA satisﬁes this condition. because they cannot tell when there is only one electron in the system.
2 19.158 CHAPTER 19. and strict correlation Mixing exact exchange with GGA Questions about adiabatic connection formula and hybrids 1. Does the HellmannFeynman theorem apply to the problem of the particle in a potential V (x) = exp(−x)? λ 3. What happens to a system as λ → ∞? 157 . What is the formula for vext [n](r) in terms of a scaled density? 2. HYBRIDS Chapter 19 Hybrids 19.1 19. strong.3 Static.
ORBITAL FUNCTIONALS Chapter 20 Orbital functionals 20.4 20.2 20.5 Selfinteraction corrections Optimized potential method G¨rlingLevy perturbation theory o Meta GGA’s Jacob’s ladder 159 .160 CHAPTER 20.3 20.1 20.
Part V Timedependent DFT 161 .
4) Figure 21. we illustrate the concept of resonance. To have a feeling for what timedependent quantum mechanics really means.5 2 2. The density barely changes from its groundstate value.4. where the phase factor for undisturbed evolution has been extracted. This corresponds to the lowest transition. 21. The mean position begins to oscillate immediately. with Vjk = jV k .1) ˆ ˆ ˆ where H(t) = T + V (t). (21.1 Schr¨dinger equation o E(t) = A sin(ωt) (1 − exp(−ωt)) (21.7) 163 expectation value of operators contains information about transition frequencies. We will generally be concerned with a timedependence of the form ˆ ˆ ˆ H(t) = Ho + V (t) where the unperturbed Hamiltonian. We run our calculation with a = 1. Over long times. For our particle in a box in an external electric ﬁeld.5 3 3.1: Electric ﬁeld and expectation value of x for a particle in a box of length 1 subjected to the timedependent ﬁeld shown and given in the text. the greater the eﬀect of the driving ﬁeld on the particle. But if.8. so that Cj (0) = δj0 . and the wavefunction does evolve.5 1 1. In Fig. as shown in Fig.93.6) We usually begin in the groundstate. since 1 = π 2 /2 = 4.5 4 t 100 <x(t)> 0 0.1) yields ˙ i Cj (t) = Ck exp(iωjk t) Vjk (t). Exercise 87 Matrix elements of timedependent potentials Calculate the matrix elements Vjk (t) for (a) a harmonic oscillator and (b) a particle in a box.5 3 3. as the potential changes signiﬁcantly from its original. then the chances of exciting out of the groundstate in any small interval are exponentially small. but simply jiggles to left and right.8) We learn in kindergarten that the evolution of the wavefunction is governed by the timedependent Schr¨dinger equation: o ˙ ˆ iΨ(t) = H(t) Ψ(t). We take our particle in a box and subject it to a particular driving potential. and 2 = 4 1 . ωjk = Ej − Ek and with initial values Cj (0) = Ψ(0)j . V (xt) = x E(t) (21. instead .3) 0 0.2) E(t) ˆ Ho j = Ej j (21. 21.5) k ˆ a coupled set of ﬁrstorder diﬀerential equations. We choose This form allows a slow turnon of a periodic electric ﬁeld of frequency ω and amplitude A. the perturbation grows indeﬁnitely and pushes more and more energy into the particle. TIMEDEPENDENCE This is the potential due to an external timedependent electric ﬁeld in the dipole approximation. we show how the system responds. A = 10. but rather with a period of about 1/3. (21. if the external frequency ω ω12 . the nearer the driving frequency is to the transition frequency. we immediately show some examples for a simple case. but sweep the driving frequency through ω = ω12 . and the period is 2π/ω ∼ 0. not with the frequency of the driving ﬁeld.164 CHAPTER 21. Ψ(0) given (21.5 4 t Cj (t) exp(−iEj t) j (21. when the perturbation is an electric ﬁeld. (21. and the time run up to t = 4. Insertion into Eq. these small chances add up. ω = 4. We may then expand the timedependent wavefunction in terms of these unperturbed eigenstates: Ψ(t) = j (21. Next. Thus the timedependent evolution of the 10 8 6 4 2 0 2 4 6 8 10 4 3 2 1 0 1 2 3 4 5 6 Chapter 21 Timedependence 21. Ho has eigenstates ˆ and V (t) is zero for t < 0. This is the adiabatic limit. As ω → ω21 . Our ﬁrst example is taken with generic values. We see that. and a dot implies diﬀerentiation with respect to time.2. so that ω21 = 14.1. Another interesting regime of timedependent problems is when the time variation is very slow.5 2 2.5 1 1.
so that the excited states are only inﬁnitesimally populated. A] (21. If it starts out at rest at the origin.15) Figure 21. 14.1 5 0 0. SCHRODINGER EQUATION 165 166 CHAPTER 21. and a = 2000. Essentially. The particle adiabatically follows the groundstate of the potential.1. How important is α(t)? o In this framework. The operator is deﬁned as N ˆ =1 j(r) (pi δ(r − ri ) + δ(r − ri )pi ) 2 i=1 (21.5 2 t=13 t=14 t=15 t=16 n(xt) = Ψ(xt)2 = ∞ j. and ωt 1.19) where the density commutes with all but the kinetic operator in the Hamiltonian. and so the system stays in the adiabatic ground state. those satisfying ˆ (21. (21. to be ˙ W (k) = P (k) = .3 <H(t)> 5. X to ﬁnd the probability of excitation.5 1 t 1. the change in the number of electrons per unit time must equal the diﬀerence between the ﬂow of electrons into dx from the left and out from the right. Next consider the eﬀect of varying the external frequency. 15. we are interested only in the response to a weak potential.12) (21. that will play a central role in the development of TDDFT. 21.9) H(t) Ψn (t) = En (t) Ψ(t) these chances would remain exponentially small.9 0 0.11) φ(xt) = φ0 (x − xcl (t)) exp(iα(t)) This concept is best seen in perturbation theory.17) of using the unperturbed eigenfunctions as our basis. it is simply 7 6 5 4 3 2 1 0 1 t=13 t=14 t=15 t=16 5. and one ﬁnds a huge response as it passes through a transition frequency. and asking what the groundstate is. is the current density. the probability of excitation per unit time. and its expectation value given by evaluation on the timedependent wavefunction.18) Applying this to the density yields ˆ ˆ n(rt) = Ψ(t)i[T .4 5. I’ve calculated the case where ω = 1. i..2: Expectation value of x and energy of a particle in a box of length 1 subjected to timedependent ﬁelds of frequencies ω = 13. the time evolution of any (timeindependent) operator is ˆ ˙ ˆ ˆ A = i[H(t).2 5.21) . n(r)]Ψ(t) = − j(x) ˙ (21. i.20) n(r) = ˆ i=1 δ(r − ri ) (21.14) 100 <x(t)> A new quantity.. This limit can be found by simply turning on a static external potential.16) j(xt) = φ∗ (xt)dφ(xt)/dx Since.2 Perturbation theory dt E(t ) sin(ω0 (t − t )) mω0 (21. Exercise 88 Harmonic oscillator in a timedependent electric ﬁeld Consider a harmonic oscillator of frequency ω0 in an electric ﬁeld. In many cases. The full range of behaviors of timedependent systems is unnervingly large. For our particle in a box.. n(xt) = −dj(xt)/dx ˙ (21. ˆ ˆ ˆ ˆ H(t) = T + V + δ V (t) (21. the potential is moving so slowly that the wavefunction of the particle simply rearranges itself to the new ground state at any given time. TIMEDEPENDENCE . i.1.13) and n(rt) = Ψ(x)ˆ (r)Ψ(x) n which is known as Fermi’s golden rule. the timedependent density is given by the usual formula of integrating the square of the wavefunction over all coordinates but one. This yields a ﬁnal electric ﬁeld of a(ωt) 2 /2 or 10. (21. We ﬁnd the transition rate.k=1 c∗ (t) ck (t) φj (x) φk (x) j (21.. show that its classical position is given by xcl (t) = − Then show that is a solution to the timedependent Schr¨dinger equation.¨ 21.e. we used the instantaneous eigenstates. More generally.e. in any dx around a point x.10) (21. t = 0.e. In Fig X. The density operator may be written as N t 0 This is the equation of continuity. and 16. Thus ω is much lower than the lowest transition. Imagine integrating the equations of motion for the coeﬃcients just a small amount in time. Then one may insert the initial condition into the righthandside of Eq.. and its commutation (with a little algebra) yields the divergence of the current operator. For our particle in a box.5 2 4.5 1 t 1.
in terms of the levels of the system: χ(xx ω) = φ0 (x)φ0 (x ) k=j .25) To see the eﬀect of speciﬁc transitions on the optical spectrum. Under these conditions.. we can write χ in its Lehman representation. meaning χ = 0 for t < t. Thus. and the perturbation begins at −∞.32) i. One can show this satisﬁes the wellknown ThomasReichKuhn sumrule. ∞ 2π 2 N (21.23) are the transition dipole moments for each transition. x . We deﬁne the oscillator strength of a given transition as fji = 2ωji xji 2 /3 These satisfy the TRK sumrule. and is fully turned on by t = 0.e. i. we will speak of the retarded function.34) The optical spectrum consists of a series of delta functions. we ﬁnd α(ω) = for the dynamic polarizability. we want to know the change in observables of the system. TIMEDEPENDENCE ˆ where δ V (t) is in some sense small. PERTURBATION THEORY 167 168 CHAPTER 21.28) π 0 ω Note that all oscillator strengths from the groundstate are positive. where η → 0. Furthermore. and σ(ω) = α(ω)ω/c (21. Exercise 89 Dipole matrix elements Calculate the dipole matrix elements for optical absorption for (a) the harmonic oscillator and (b) the particle in a box.27) dω σ(ω) = 0 c Thre is also the static sumrule.2..31) (21. t − t ) δv(x t ) (21. and depends only on the groundstate density (for a nondegenerate system). Since we are doing DFT. which follows from the general KramersKronig rule for the polarizability: 2 ∞ dω α(ω) = α(0) (21. Since this is a book on density functional theory. .. i.e. some are negative... taking the moments to construct α. it is a functional derivative. t − t ) = δρ(xt)/δv(x t ) (21.33) ω − ω0 ω − ω0 ie the real part is a principal value.. we will write it more neatly as χ[n0 ](x.24) d(t) = dx ρ(xt)x ie the ﬁrst moment of the density. χ tells you the density change throughout the system induced by any external potential. since the unperturbed system is static. now ﬁnd oscillator strengths for both systems and show that they satisfy the ThomasReichKuhn sumrule. and the imaginary part a deltafunction. x .. we must use the wellknown formula: 1 1 =P + iπ δ(ω − ω0) (21.. We show in the rest of this section that χ contains most of what we want to know about the response of a system. Next we specialize further to the speciﬁc case of the optical response of a system. deﬁned by (21.26) dx x dx x χ(x. we ﬁnd xjk = dx x njk (x) x0j 2 2 2 j=0 ω − ωj0 (21. which in turn give the optical response of the system. Thus σ(ω) = j=0 2ωj0 xj0 2 δ(ω − ω0 ) (21. but from an excited state. (21. a key role will be played by the change in density in response to such a perturbation. Usually we will consider potentials whose timedependence is exp(iωt + ηt).35) (21. We want to ﬁnd out the timedependent dipole moment of our system. to a long wavelength electric ﬁeld. we can ask what is the change in the system to ﬁrstorder in the perurbation? More precisely. Then α(ω) = (21.36) ie the moments of the susceptibility yield the complex dynamic susceptibility. x ω) (21.e. It is also known as the densitydensity response function and the nonlocal (meaning depends on x and x ) susceptibility. fj jk = N k=j (21. Note that it is a function of t − t alone.29) We deﬁne: n njk (x) = jˆ (x)k the generalized transition dipole.e. This is because the optical absorption is given by .21. We deﬁne the susceptibility of the system by δρ(xt) = dx dt χ(x. If we set v(x ω) = Ex . whose intensity is determined by the dipole matrix elements.30) i.22) To see better what this means. Exercise 90 Oscillator strengths From above.
Relate pair density to equaltime χ.3. If a system is in a linear combination of two eigenstates. In such cases. 3. λ) = G(rr . Using the ﬁrst two levels of the particle in a box. for a oneelectron system. and compare its time evolution with that of Fig X. x ) vanishes at large distances. which. Relate density matrix to equaltime G Deﬁne χ and discuss meaning Relate χ and G for noninteracting systems. λ ± iη) (21. 6. Deﬁne oscillator strength. G satisﬁes the diﬀerential equation: − 1 2 2 + v(r) − z G(rr . In real space. Note that the Green’s function is illdeﬁned whenever its argument matches that of an eigenvalue. z) = δ (3) (r − br ) φ∗ (r) φi (r ) i .3 Optical response (21.41) Deﬁne polarizability. TIMEDEPENDENCE When discussing ﬁrstorder perturbation theory. we deﬁne G± (rr . one may write G in terms of sums over the eigenstates G(rr .4 Questions about timedependent quantum mechanics 1. z) = i (21.42) d3 r dt G+ (rr . . calculate the Rabi frequency. and boundary conditions are chosen so that G(x. may be written in operator form as ˆ ˆ G(z) = (z − H)(−1) (21. 4. just like the boundstates. t − t ) Ψ(r t ) (21.21. it is always useful to deﬁne response functions of the system at hand. 5. and show that they satisfy the ThomasReich Kuhn sum rule. assuming a particle incident from the left. Show how a harmonic oscillator evolves in a timedependent electric ﬁeld. and show that they satisfy the ThomasReich Kuhn sum rule. the Green’s function yields the timeevolution of any initial state. for a timeindependent Hamiltonian: Ψ(rt) = since ˆ ˆ G+ (t − t ) = exp(−iH(t − t )) Simple expression for G in terms of L and R phi. Calculate the transmission and reﬂection amplitudes of the 1d Hatom. where z is a complex number with positive imaginary part. 8.37) 21. Deﬁne optical spectrum. Calculate the retarded green’s function for the 1d Hatom. OPTICAL RESPONSE 169 170 CHAPTER 21. Calculate the oscillator strengths for the harmonic oscillator. z− i (21. while − refers to the advanced one. When Fouriertransformed. Calculate the oscillator strengths for a particle in a box.40) where + refers to the retarded Green’s function.39) 7. A wellknown one is the Green’s function. 21.38) Formally. does its density change? 2. Calculate the imaginary part of the frequencydependent susceptibility of the 1d Hatom. static and dynamic.
whose density is precisely that of the real system. larger systems with more electrons can now be tackled.. for a given initial state. When the perturbing ﬁeld is weak. after a molecule’s structure has been found. These applications fall into three general categories: nonperturbative regimes. the number of these TDDFT response calculations for transition frequencies is growing exponentially at present. In subsequent chapters. In the absence of Hartreeexchangecorrelation eﬀects. Then one can deﬁne a ﬁctitious system of noninteracting electrons moving in a KohnSham potential. and is called the RungeGross theorem. Overall. results tend to be fairly good (0. An obvious and simple approximation is adiabatic LDA (ALDA). we only need know this potential in the vicinity of the initial state. i. because it is a functional of the groundstate density alone. the exchangecorrelation kernel is much more manageable than the full timedependent exchangecorrelation potential. The exchangecorrelation potential (deﬁned in the usual way) is then a functional of the entire history of the density. But with the advent of TDDFT. the allowed transitions are exactly those of the groundstate KohnSham potential. In these situations. Crude and unreliable approximations had to be made to tackle larger systems. which has become very popular in this community. or to cause a speciﬁc chemical reaction to occur (quantum control).e. typically). The eigenvalues yield transition frequencies while the eigenvectors yield oscillator strengths.1 Overview We will see below that. to the groundstate problem. TDDFT has been programmed into most standard quantum chemical packages and. perturbation theory applies. as the full timedependent wavefunction calculations are very demanding. This allows use of many existing fast algorithms to extract the lowest few excitations. and groundstate applications. on the other hand. in the same fashion as perturbation theory yields expressions for ground Timedependent density functional theory So far. we extend our view. Fouriertransform of the timedependent dipole matrix element then yields the optical spectrum. the intensities of the optical transitions can be extracted in the same calculations. we apply the methods of DFT to such potentials. Knowledge of it implies solution of all timedependent Coulombinteracting problems. molecules.172 CHAPTER 22. the 27th harmonic. instead of needing knowledge of vXC for densities that are changing signiﬁcantly with time. TIMEDEPENDENT DENSITY FUNCTIONAL THEORY Chapter 22 system at 27 times the frequency of the perturbing electric ﬁeld. Chemists and physicists have devised separate approaches to extracting excitations from TDDFT for atoms. but corrected by TDDFT via matrix elements of the exchangecorrelation kernel. This is the timedependent analog of the HohenbergKohn theorem. Physicists. The last class of application of TDDFT is.2 eV errors. in a space of the singleexcitations of the system. Φ(0). the initial interacting wavefunction Ψ(0). The ﬁrst of these involves atoms and molecules in intense laser ﬁelds. the perturbing electric ﬁeld is comparable to or much greater than the static electric ﬁeld due to the nuclei.g. In this chapter. and has achieved its great successes. and the initial KohnSham wavefunction. one can establish a onetoone correspondence between timedependent densities n(rt) and timedependent onebody potentials vext (rt). the response of the 171 . Diﬃculties remain for the application of TDDFT to solids. In this way. This is where most DFT research has been done. We can then apply this DFT technology to all problems of timedependent electrons. in which the ﬁeld is so intense that perturbation theory does not apply. The chemists way is to very eﬃciently convert the search for poles of response functions into a large eigenvalue problem. under certain quite general conditions. This is because one can extract the groundstate exchangecorrelation energy from a response function. as in normal spectroscopic experiments. a given evolution of the density can be generated by at most one timedependent potential. This gives us a working KohnSham scheme. and clusters.. because the present generation of approximate functionals (local and semilocal) lose important eﬀects in the thermodynamic limit. Moreover. perhaps surprisingly. While still a more complex beast than the groundstate exchangecorrelation potential.. But the presence of the kernel shifts the transition frequencies away from the KohnSham values to the true values. 22. Experimental aims would be to enhance.1 to 0. Previously. so these also are aﬀected by the kernel. just as in the ground state. i. vXC [n](x) ≈ vXC (n(x)). by providing a general tool for tackling electronic structure problems. it is usually not too costly to extract its lowlying spectrum. much more so than the groundstate case. only one and two electron systems could be handled computationally. but little is understood about their reliability. prefer to simply solve the timedependent KohnSham equations when a weak ﬁeld has been turned on.e. linear (and higherorder) response. to include timedependent potentials. This functional is a very complex one. sometimes called timedependent LDA. e. Then. n(x). These changes are characterized by a new functional. Using either methodology. Analysis of the linear response then shows that it is (usually) dominated by the response of the groundstate KohnSham system. but much work is currently in progress. in which we use the groundstate potential of the uniform gas with that unif instantaneous and local density. we have restricted our attention to the ground state of nonrelativistic electrons in timeindependent potentials. That is. which we take to be a nondegenerate groundstate. the exchangecorrelation kernel.
i. In the second part.7) k · n0 (r) ∂0 ∆vext (x) The analog of the HohenbergKohn theorem for timedependent problems is the RungeGross theorem[?]. a given history of the density can be generated by at most one history of the potential.e. under certain quite general conditions. if f (r) = k ∂0 ∆vext (x) is nonzero for some k. i. there is a onetoone mapping between densities and potentials. In particular. the wavefunction is in turn determined up to a timedependent phase. and one can construct a density functional theory. One can go further. (22. then the second term on the right is deﬁnitely negative. Thus any approximation for the exchangecorrelation kernel yields an approximation to EXC . if not for the divergence on the righthandside. Does the divergence allow some escape from this conclusion? The answer is no. in practice TDDFT provides a natural methodology for modeling these ﬂuctuations. ∆n(x) = − ˙ so that k+2 ∂0 ∆n(x) = · ∆j(x) . We deﬁne the exchangecorrelation potential via: vS (x) = vext (x) + vH (x) + vXC (x) (22.22. one can establish a onetoone correspondence between timedependent densities n(rt) and timedependent onebody potentials vext (rt). (22.4) . and considering t = 0.1) The ﬁrst term on the right may be written as a surface integral at r = ∞.10) ∆j(x) = 0. That is. mutually interacting via the Coulomb repulsion. from the covalent bond distance to dissociation. 0). including van der Waals.3) whose density is + vS [n](x) φj (x) ..8) k where ∂0 = (∂ k /∂tk )t=0 . Thus there is no way for f (r) to be nonzero. then there must be some ﬁnite k for which the right hand side of (22. While in principle all this is included in the exact groundstate functional. but may well be worth the payoﬀ.e. 2 (22. The RungeGross theorem states that the densities n(x) and n (x) evolving from a common initial state Ψ0 = Ψ(t = 0) under the inﬂuence of two potentials vext (x) and vext (x) (both Taylor expandable about the initial time 0) are always diﬀerent provided that the potentials diﬀer by more than a purely timedependent (rindependent) function: ∆vext (x) = v(x) − v (x) = c(t) . we would be done.2) We will see below that. write d3 r f (r) · (n0 (r) f (r)) = d3 r · (f (r)n0 (r) f (r)) − n0 (r) f (r)2 (22. We consider N nonrelativistic electrons. Then one can deﬁne a ﬁctious system of noninteracting electrons that satisfy timedependent KohnSham equations ˙ iφj (rt) = − 2 where n0 (r) = n(r. the equation of motion for the diﬀerence of the two current densities is: ˙ ˆ ∆j(x)t=0 = −i Ψ0  ˆ j(r). (22. (22. 0) is the initial density. and attempts exist to generate entire molecular energy curves.2.3 KohnSham equations If so. (22. by repeatedly using the equation of motion. and its integrand must be nonzero somewhere. 22. so that the integral on the left cannot vanish. since n0 (r) is nonzero everywhere. Because the Hamiltonians only diﬀer in their external potentials. To see this.5) (22. this approach includes correlated ﬂuctuations on the two separated pieces. which cancels out of the expectation value of any operator. RUNGEGROSS THEOREM 173 174 CHAPTER 22. we extend the proof to the densities. TIMEDEPENDENT DENSITY FUNCTIONAL THEORY state contributions in terms of sums over excited states. for a given initial state. which establishes a onetoone correspondence between current densities and external potentials. (22. the kth derivative evaluated at the initial time.1) holds. and vanishes for any realistic potential (which fall oﬀ at least as fast as −1/r). Such calculations are much more demanding than simple selfconsistent groundstate calculations. as a system is pulled apart into fragments. so that n(x) = j=1 φj (x)2 . We prove this theorem by ﬁrst showing that the corresponding current densities must diﬀer.2 RungeGross theorem This is the ﬁrst part of the RungeGross theorem. This is the timedependent analog of the HohenbergKohn theorem. N (22. If we imagine that f (r) is nonzero somewhere. 22. this produces a natural method for incorporating timedependent ﬂuctuations in the exchangecorrelation energy.6) Now. Such calculations are signiﬁcantly more demanding than regular groundstate DFT calculations. to ﬁnd[?] k+1 ∂0 ∆j(x) = −n0 (r) k ∂0 ∆vext (x) . and the potentials are Taylor expandable about to . Since the density determines the potential up to a timedependent constant. ∆H(t0 ) Ψ0 = −n0 (r) ∆vext (r. Most importantly. for any physical density.2) does not vanish.. in a timedependent external potential. and yet have f (n0 f ) vanish everywhere. then the density diﬀerence must be nonzero. Simple ALDA approximations for polarizabilities of atoms and molecules allow C6 coeﬃcients to be accurately calculated by this method.9) precisely that of the real system. If Eq. From continuity.
A simple way to go beyond the adiabatic approximation would be to include some dependence on. automatically yields an adiabatic approximation for use in TDDFT calculations. as shown in Sec. If the timedependent potential changes very slowly (adiabatically). this approximation will be valid. Imagine you have a timedependent potential that pushes it into its ﬁrst excited state (1s2s).5 Questions on general principles of TDDFT 1. and how its success can be built upon with GGA’s and hybrids. within the adiabatic approximation. Φ(0)](rt) − vext [n. Ψ(0). but for a timedependent density.11) 2. Knowledge of it implies solution of all timedependent Coulomb interacting problems. TIMEDEPENDENT DENSITY FUNCTIONAL THEORY where the Hartree potential has the usual form. Describe the ﬁnal situation. ˙ Bureaucratically. we require adia gs vXC [n](rt) = vXC [n0 ](r)n0 (r)=n(rt) (22. In groundstate DFT. As mentioned above. called the exchangecorrelation action. Now imagine the exact KohnSham calculation. Is there anything special about the response of a particle in a harmonic well to an external electic ﬁeld? 4. This is the precise analog of the argument made to determine the function used in LDA for the groundstate energy (see Sec. Repeat question above. Exercise 92 Continuity in ALDA Does a KohnSham ALDA calculation satisfy continuity? Prove your answer. but replace density with wavefunction. Exercise 91 KohnSham potential for one electron By inverting the timedependent KohnSham equation. and the initial KohnSham wavefunction. returns its density to its initial value.4..e. such as a GGA or hybrid. ALDA is the mother of all TDDFT approximations.22. In a timedependent KohnSham calculation starting from the ground state. and so the initialstate dependence disappears. n(x). Φ(0)](rt) = vS [n. . We’ll soon see if some of the same magic applies to TDDFT. If a system is in a linear combination of two eigenstates. the initial wavefunctions themselves are functionals of the initial density. (22. Does the RungeGross theorem guarantee that the current density in a TDDFT calculation is correct? 5. as well as the initial wavefunctions of both the interacting and the KohnSham systems: vXC [n. However. Φ(0). n(rt). This functional is a very complex one. it approximates the functional as being local in time. What happens if your functional has some memory? 9. say. we’ve already seen how well LDA works beyond its obvious range of validity for the groundstate problem. The exchangecorrelation potential is then a functional of the entire history of the density. this turns out not to be possible. X). 22. However. The adiabatic approximation is one in which we ignore all dependence on the past. ALDA should only work for systems with very small density gradients in space and time. will the potential return to its initial value? 7. due to satisfaction of sum rules. the spatial locality of the functional is also approximated. We call this a historydependence. after a time. and allow only a dependence on the instantaneous density: adia approx vXC [n](rt) = vXC [n(t)](r). However. To make the adiabatic approximation exact for the only systems for which it can be exact. Consider a He atom sitting in its groundstate. give an explicit expression for the timedependent KohnSham potential in terms of the density for one electron.13) gs where vXC [n0 ](r) is the exact groundstate exchangecorrelation potential of the density n 0 (r). will electronic transitions occur at frequencies that are diﬀerences between groundstate KohnSham orbital energies? 3. the initial interacting wavefunction Ψ(0). Can more than one potential produce the same timedependent density? 6. It may seem like a drastic approximation to neglect all nonlocality in time. ADIABATIC APPROXIMATION 175 176 CHAPTER 22. But the exchangecorrelation potential at r and t has a functional dependence not just on n(r t) but on all n(r t ) for 0 ≤ t ≤ t. in the special case of starting from a nondegenerate ground state (both interacting and non). If a system is subjected to a driving potential that. Repeat both questions above. much more so than the groundstate case. the exchangecorrelation potential is the functional derivative of EXC [n]. but any groundstate functional. X. 22. (22. But the electrons will remain always in their instantaneous ground state.. Ψ(0)](rt) − vH [n](rt).4 Adiabatic approximation In practice. the exact exchangecorrelation potential depends on the entire history of the density. etc.12) i. does its density change? As we have noted. 8. It would be nice to ﬁnd a functional of n(rt) for which vXC (rt) was the functional derivative.
There was some controversy over the last step of the RungeGross proof. B. 64 039901 (E).T. A 63. Vol.T. and C. San Diego.U. Phys. N. Maitra and K. Burke. Gross and W. QUESTIONS ON GENERAL PRINCIPLES OF TDDFT 177 178 CHAPTER 22. For a simple demonstration of initialstate dependence. Rev. Woodward.K. Lett. 89. 21: Density Functional Theory of ManyFermion Systems. N. Phys. in Advances in Quantum Chemistry. Letts. Gross. Trickey (Academic Press. .5. Phys. edited by S. 997 (1984). but this is settled in E. The extension to piecewise analytic potentials and the initialstate dependence is addressed in Memory in timedependent density functional theory. requiring the integration by parts.K. Rev. K. TIMEDEPENDENT DENSITY FUNCTIONAL THEORY Bibliography The original proof of the RungeGross theorem is in Densityfunctional theory for timedependent systems. E. 52. see Demonstration of initialstate dependence in timedependent density functional theory. Runge and E.U. Rev. 042501 (2001). 023002 (2002). Maitra. 1990). Burke.22. Kohn.
This is because when ω matches a true transition frequency of the system. However.e. which is quite diﬀerent from the interacting case. the exchangecorrelation kernel is much more manageable than the full timedependent exchangecorrelation potential. (23. we deﬁne the pointwise susceptibility χ[n0 ](rr . (23. Ω = ω 2 . But the presence of the kernel in Eq. (23. Thus χS has a set of such poles.1 Dysonlike response equation and the kernel where all objects are functionals of the groundstate density.e.2 Casida’s equations i. and of the exchangecorrelation kernel. 23.. we have vXC [n + δn](x) = vXC [n](r) + dt d3 r fXC [n](rr . The matrix is i ˜ Ωqq (ω) = δqq Ωq + 2 ωq ω qfHXC (ω)q . r . evaluated on the groundstate density: fXC [n](rr .2) 23.aunocc When the perturbing ﬁeld is weak. this implies fXC is frequencyindependent. In linear response. ensuring the invertibility of the linear response operators of section ??. (23. since χ is a twobody response function. r . q (23. instead of needing knowledge of vXC for densities that are changing signiﬁcantly with time. where fXC is called the exchangecorrelation kernel. (23.(ω → −ω) (23. Here we focus on the results. ω)Φq (r ).7) q where q is a double index. has a pole as a function of ω. these equations are a straightforward matrix equation. To understand why fXC is important for linear response. because it is a functional of the groundstate density alone. so using the deﬁnition of the KS potential. ω). and Φq (r) = φ∗ (r)φa (r). we ﬁnd the key equation of TDDFT linear response: χ(rr ω) = χS (rr ω) + d3 r1 d3 r2 χS (rr1 ω) 179 While still a more complex beast than the groundstate exchangecorrelation potential. Then. this implies that the exchangecorrelation kernel has the form: δv gs [n0 ](r) adia fXC (rr . (23. Casida showed that. 1/r − r  + fXC (r. But both must yield the same density response.4) shifts the transitions away from the KS values to the true values. (23. t − t ) δn(r t ). t − t ) δvext (r t ).8) q where qfHXC (ω)q = d3 r d3 r Φ∗ (r)fHXC (r. ﬁnding the poles of χ is equivalent to solving the eigenvalue problem: ˜ Ωqq (ω)vq = Ωvq . if you make a small change in the external potential at point r and time t . we only need know this potential in the vicinity of the initial state. i. which says how the noninteracting KS electrons would respond to δvS (r t ). (23. Moreover. representing a transition from occupied KS orbital i to unoccupied KS orbital a. t − t ) = XC n0 (r)=n(rt) δ(t − t ). and so the allowed transitions are exactly those of the groundstate KS potential.c. which we take to be a nondegenerate groundstate. χ = χS . the response function χ blows up. This equation contains the key to electronic excitations via TDDFT. t − t ) as the response of the ground state to a small change in the external potential: δn(x) = dt d3 r χ[n0 ](rr . LINEAR RESPONSE Chapter 23 Linear response Note that the diﬀerence between n(x) and n (x) is nonvanishing already in ﬁrst order of v(x) − v (x). fHXC is the Hartreeexchangecorrelation kernel. t − t ) = δvXC (rt)/δn(r t ).180 CHAPTER 23. Writing n(x) = n(r) + δn(x). the groundstate KohnSham system has its own analog of χ. the strengths of the poles can be simply related to optical absorption intensities (oscillator strengths) and so these also are aﬀected by the kernel.4) r1 − r2  In this equation. .7) yields the excitation energies ω and the oscillator strengths can be obtained from the eigenvectors [?]. This is called a Dysonlike equation.9) 1 + fXC (r1 r2 ω) χ(r2 r ω). with a kernel that is the onebody potential.3) The various methods for extracting excitations have been described in the overview.. (23. χ tells you how the density will change at point r and later time t. perturbation theory applies. at the singleparticle excitations of the KS system: φ∗ (r) φa (r) φ∗ (r ) φi (r ) i a − c. Algorithms exist for extracting just the lowest transitions. which we denote by χs . A selfconsistent solution of Eq. because it has the same mathematical form as the Dyson equation relating the oneparticle Green’s function to its free counterpart. Now. it is an oddity of DFT that it has this form.1) In the absence of Hartreeexchangecorrelation eﬀects. as in normal spectroscopic experiments.5) χS (rr ω) = 2 ω − ( i − a ) + i0+ iocc.6) δn0 (r) adia When Fouriertransformed. In the special case of an adiabatic approximation. ωq = a − i .
SINGLEPOLE APPROXIMATION 181 182 CHAPTER 23.23.3 Singlepole approximation . LINEAR RESPONSE 23.3.
can yield very diﬀerent results. i.4 24. in practice. Table X shows. Hybrid functionals in TDDFT: An important point to note here concerns implementation of hybrid functionals in TDDFT. for many systems of real interest. we have surveyed all that one needs to perform a TDDFT calculation of excitation energies using a modern code. Table I lists the results. Thus the potential produced in many quantum chemical codes consists of a fraction of the HF nonlocal potential. are devoted to delving deeper into understanding how TDDFT works and what are its current limitations.3 Transition frequencies At this point. especially for 1s → 2s and 1s → 3s. these results do not change the basic reasoning and conclusions presented below. There are two ways a hybrid might be coded. Our favorite approximations to the exchangecorrelation energy yield fairly lousy exchangecorrelation potentials. In fact the Rydberg series. which are compared with a highly accurate nonrelativistic variational calculations[?. How can one correct KS transition frequencies if they do not exist? The standard answer has been to asymptotically correct the potentials. ’correct’ DFT approach. an inﬁnite sequence of excitations as E → 0 from below.. but a large eﬀect on the unoccupied levels.5 24. but neglecting the continuum. to make them mimic the true KS potential more accurately. The errors reported are absolute deviations from the exact values. Assuming inclusion of the continuum aﬀects results with other approximate kernels similarly. it is the lowlying valence excitations that are of interest. to guarantee that the eﬀective potential is always a local. characteristic of the −1/r decay. However. but suggest that calculations including the continuum may prove to be more accurate than those presented here. for reasons discussed in the next section. These calculations were done including ﬁrst only bound states. 24. the optimized eﬀective potential method should be used. the potentials are quite accurate. This means that all the higher states are too shallow.184 CHAPTER 24. [?].. making Koopman’s theorem fail drastically. 183 . ?] In each symmetry class (s. multiplicative one. In this section. we note that all such schemes then produce a potential that is not a functional derivative of an exchangecorrelation energy. In the ﬁrst. and in these regions. HartreeFock calculations are much easier. A variety of schemes exist that do this to varying degrees of accuracy. which in the case of excitations. This has negligible eﬀect for groundstate and occupied orbitals. and d). the three approximations to the kernel mentioned in the previous section. Other excitations barely changed. and including many boundstate poles in Eq. we report calculations for the He and Be atoms using the exact groundstate KohnSham potentials. The remainder of this chapter. These potentials are especially unpleasant at large distances from Coulombic systems.1 24. They are too shallow. as required by KohnSham theory. p. up to 38 virtual states were calculated. The technical details are given in Ref. The eﬀect of neglecting continuum states in these calculations has been investigated by van Gisbergen et al. (??). and then including all positive energy orbitals allowed by their basis set.e. 24. yielding results identical to those presented here. However. The second error under the Be atom excludes the 2s → 2p transition. either a quatum chemistry one or a realtime evolution.2 Sources of error Poor potentials A problem was noticed early on. and all of the next. They found signiﬁcant improvement in He singletsinglet excitations. does not exist at all.[?]. and have been around since time immemorial..6 Atoms Molecules Strong ﬁelds In any event. PERFORMANCE Chapter 24 Performance 24. who performed ALDA calculations from the exact KohnSham potential in a localized basis set.
0/0.4/0.6/0.7/6.9 5.5 0. and using the lowest 34 unoccupied orbitals of s and p symmetry for He.2/ 0.9/10.8 1.4 5.2 0.4/1.7 269.5/0.3 0.6.1/0.4 295.2/0.8/1.8 1.6 872. PERFORMANCE Table 24.6 4.1: Singlet/triplet excitation energies in the helium and beryllium atoms.6/0.4/2.1 4.1/0.9 2.8 138 45 Singlet/triplet shifts ALDA X SIC hybrid exact Transitions from the 1s state in He atom 22/11 20/25 19/16 14/19 12/18 6.8/0.3 0.6/0.2 5.3 1.9 0.4 0.4 2. [?] for Be.4/0.7 0.2 3.0 0.6 3. ωKS 2s 3s 4s 5s 6s 2p 3p 4p 5p 6p err 3s 4s 5s 6s 2p 3p 4p 5p 6p 3d 4d err err’ 746.0 1.0 0.3 32 31 29 12 Transitions from the 2s state in Be atom 7.6/0.1/5.6 1.6/3.8/2.5/2.8/7.8 1.3 0.9 0. and d symmetry for Be.3/1.6 0.6/1. and the lowest 38 unoccupied orbitals of s.0 0.1 0.3/0.3/1.8 1. p.3 0.7 0.0/2.6 1.2 0.4 0.2/0.8/0.7 1.2 0.0/0.4 6.7 10.5/1.6 10. [?] for He and from Ref.5 0.2/0.6/0.4 2.3 0.4 304.2 0.4/2.1/0. Exact values from Ref.9/2.6/0.8/0.8 0.2/0.3/4.3 309.3 6.7 1.0 0.3/0.2/0.2 0.5/0.6 889.1 0.4/1.9 283.4 0.1/0.4 0.0 0.3 2.4/ 0.1/ 0.1 0.5 1.2 1.1/4.8 881.5 0.0 1.7/7.4 0.8/4.5/1.4/0.6 3.1/0.9/0.3 326.1 0.3/0. STRONG FIELDS 185 186 CHAPTER 24.6 319.2 868.5/0.2/ 0.0/0.8 777.7 132.9 315.6/2.6 1.6 5.2/8.7/1.5/ 0.24.3 324.6/0.5 0.9/0.3 1.1 2.7 2.0/4.9 2.6 2.2 10.0 839.1 0.3/1.4/4.5 0.3/1.3 1.0/0.7/4.2/2.1 5.3/0.8 57 244.4/1.4 7.3/0.9/ 0.2 883.5 56/42 55/133 53/53 10/88 61/32 2.7/0.1/1.3 2.3/0.0 1.3/0.9 1.6 1.4/ 1.9/ 3.2 2.2 56 144 73 136 41 37 44 29  .7/3.6/ 1.5 3.9 0.7/1.5 0.6 3.3/0.1/0.5/0.7/0.6/0.4 1.4 0. calculated from the exact KohnSham potential by using approximate xc kernels (in millihartrees).9/ 0.2 847.4/2.4 6.6/0.6/4.9 888.5 0.6 1.8 4.1/0.2 1.
deﬁned as Chapter 25 A[Ψ] = tf 0 ˆ dt Ψ(t)H − i ∂ Ψ(t) ∂t (25. t → t )/n(r) − δ (3) (r − r ). and so the entire potential is a density functional.Action However. and its often not mentioned explicitly. (25.. because of the memorydependence of the functional. χ(rr .. whose λdependence can be simply extracted via scaling. we simply note that the equaltime susceptibility. In fact. camera.. as a system is pulled apart . (23. Any approximation for fXC . Eq.3) Not much attention is paid in the literature to the initialstate dependence in the RungeGross theorem.e. r ) = −χλ (rr . and. the pair density of chapter X.4 25. what we really care about is the energy.4) Before discussing the action in timedependent DFT. (25.. and include any initial wavefunction that can be generated by some timeevolution of the system from a nondegenerate groundstate. Thus nλ (r...3 Lights. variations in the density restrict variations in the wavefunction to only those that are vrepresentable. simply tack on this prehistory.1 25. XC (25. we wish to know the timedependent expectation value of the density. In a timedependent problem.5) EXC = d3 r d3 r 0 π giving the exchangecorrelation energy directly in terms of χλ . i. The original RungeGross paper deﬁned this action simply as A[n] = A[Ψ[n]] (25. to the groundstate problem.2) ∂Ψ is satisﬁed by the wavefunction obeying the timedependent Schrodinger equation. and assume from here on that the initial state is a nondegenerate ground state. 25. this has since been shown to lead to several inconsistencies. In that case. EXOTICA the groundstate problem. this is not really a problem. begining from −tP . Often.188 CHAPTER 25. but the wavefunction is a function of many variables. since we’d have a diﬀerent functional for every initial wavefunction. the action is the mathematically analogous quantity. one can be even more general. and apply TDDFT with the initial groundstate functional. Inserting this expression for the hole into the adiabatic connection formula for the energy. The original initial potential (at t = 0) diﬀers from what it would be if we began in the groundstate. yields an approximation to E XC . perhaps surprisingly. there is no one functional in which we have an overarching interest. We will not discuss further the initialstate dependence. I ﬁrst point out that. Can all initialstates be generated by a pseduoprehistory begining in a groundstate? This seems unlikely..2 Currents Initialstate dependence ∂A =0 (25. for TDDFT. To do this. But presumably most of physical interest are of this kind. since one only has the freedom to vary vext (rt). by virtue of the groundstate HohenbergKohn theorem. t − t = 0). But TDDFT would be useless if we always had to account for it. one restricts oneself to starting in a nondegenerate ground state. In practice.5 Solids Back to the ground state The last class of application of TDDFT is. In particular. there is no analog of the groundstate energy.5) produces a natural method for incorporating timedependent ﬂuctuations in the exchangecorrelation energy. Nevertheless. i. in the same fashion as perturbation theory yields expressions for groundstate contributions in terms of sums over excited states. using an external potential v(x). This is because one can extract the groundstate exchangecorrelation energy from a response function.4) to arbitrary λ by using a kernel λ/r 1 − r2  + λ fXC (r1 r2 ω).1) The principle of least action says that Exotica 25. It would be very nice to deﬁne an exact A[n] whose variations with respect to the density yielded the timedependent Schroedinger equation. which will determine the geometry of our molecules and solids. with t running from −t P to 0. yields the densitydensity correlation function. Essentially. form 25. which is determined by the KohnSham potential. Suppose one can ﬁnd such a prehistory. the initial wavefunction is an (implicit) density functional.e. In that case. But in 187 where we have now generalized Eq. Most importantly. That is. we ﬁnd: ∞ dω .
6 25. including van der Waals. in practice TDDFT provides a natural methodology for modelling these ﬂuctuations.5) includes correlated ﬂuctuations on the two separated pieces. EXOTICA into fragments.6.7 Multiple excitations Exact conditions . but may well be worth the payoﬀ. 25. Eq.25. Such calculations are much more demanding than simple selfconsistent groundstate calculations. and attempts exist to generate entire molecular enery curves. from the covalent bond distance to dissociation. MULTIPLE EXCITATIONS 189 190 CHAPTER 25. While in principle all this is included in the exact functional. Simple ALDA approximations for polarizabilities of atoms and molecules allow C6 coeﬃcients to be accurately calculated by this method. (25.
independent of x and y. MATH BACKGROUND A. δ (x)f (x) = −f (0) (A. Then we deﬁne h = f − µg (A. Another handy formula is the integral over the δfunction: x −∞ (A.11) and set both to zero.3) where Θ(x) is the step function (= 0 for x < 0.1 Lagrange multipliers 1 = lim √ exp(−x2 /γ 2 ) γ→0 γ π 1 γ = lim γ→0 π x2 + γ 2 (A. The δfunction is an inﬁnitely narrow. its not strictly a function at all. Write the constraint as g(x. i.5) implies that the δ function is the derivative of the step function. But a much more elegant way is to introduce a Lagrange multiplier. This is especially important if we cannot solve the constraint equation analytically. it’s just a constant..e. Then we enforce the condition g = 0. Another way to see this is that the ﬁrst formula of Eq. inﬁnitely tall peak at x = 0. and substitute into f . yielding the maximum √ f = 2. one can show: δ(x − xi ) (A.192 APPENDIX A. chosen so that its area is exactly 1. Lastly.2 Properties of the δfunction Appendix A We review brieﬂy the useful properties of δfunction. where g = x2 + y 2 − 1. The Fourier transform of the δfunction is ∞ dp δ(x) = exp(ipx) (A.2) dx 1 − x2 √ √ which we set to zero.f (xi )=0 f (xi ) where f (x) = df /dx.5) To illustrate how Lagrange multipliers work.9) −∞ 2π i. with no restriction between x and y. y) = 0. 1)/(2µ). Write ∂h = 1 − 2µ x ∂x ∂h = 1 − 2µ y ∂y (A. but properly called a distribution. √ yielding µ = ±1/ 2. We then minimize h freely. This method avoids diﬀerentiating square roots.e.6) assuming the integration interval includes the origin. Then y = 1/ 2 also.1) Then we ﬁnd the extrema of f by diﬀerentiation: ˜ x df =1− √ (A.4) which is just the step function. we do a simple example on a function of two variables. when you get 1. Finally. and choose the positive sign for a maximum. instead of functionals. etc. . yielding (x. This can be used to extract the value of the function at any point x0 : dx f (x)δ(x − x0 ) = f (x0 ) again assuming the integration interval includes the point x0 .10) δ(f (x)) = i. y) = (1. The inelegant approach is to simply solve the constraint equation for y as a √ function of x. with some care. Suppose we wish to maximize the function f (x. yielding √ ˜ f (x) = x + 1 − x2 (A.8) where µ is a constant. a→0 Math background A. you get zero unless the interval includes the δfunction. (A.7) dx δ(x ) = Θ(x) (A. 1 for x > 0). y) = x + y subject to the constraint x2 + y 2 = 1. In fact. The usefulness of the δ function is that. 191 assuming the interval includes 0. ﬁnding x = 1/ 2. ﬁnding y = 1 − x2 .. Thus we can write several limiting formulas: δ(x) = lim (Θ(x + a/2) − Θ(x − a/2)) /a. for any reasonably smooth function: dx f (x)δ(x) = f (0) (A. and the δfunction looks like (with a ﬁnite width) what’s shown in Fig X.
3 Fourier transforms Include KramersKronig relation.A. .3. FOURIER TRANSFORMS 193 194 APPENDIX A. MATH BACKGROUND A.
) = δ(x − x ) √ i 2 x−x  e 195 (B.5) iZ ik . − + 0 )] (B. = t exp(ikx).1 1d H atom −16 (˜ − 1) ω 4ω4 Z ˜ (B.3) (B. RESULTS FOR SIMPLE ONEELECTRON SYSTEMS and in terms of g > . Furthermore. x . ) = n(x)n(x ) [g > (x.196 APPENDIX B.6) ˆ − H)g > (x. x > 0 where r and t are given by: x<0 (B.10) Results for simple oneelectron systems B. ) = √ i 2 > Zei 2 (x+x ) . The normalized orbital φ(x) = √ Z exp(−Zx) (B.9) Appendix B From this we can calculate both the static and dynamic polarizabilites: α(0) = − and Im(α(ω)) = 5 4Z 4 (B.1) φ 0+ = −Zφ(0) − The groundstate energy is just is: and groundstate density n(x) = Z exp(−2Zx) 0 2 (B. t= k − iZ Z + ik √ ˆ and k = 2 . the solution to: r= ( + (B.4) We can also extract continuum excited states. + 0 ) + g >∗ (x. V = 2 0 .7) is the outgoing Green’s function: 1 g (x. the susceptibility is χ(x. x . when H is given by (B. E > 0.1).11) where ω = ω/ 0  ˜ The Hamiltonian for this problem is 1 d2 − Zδ(x) 2 dx2 Integrating the Schr¨dinger equation through x = 0 yields the cusp condition: o H=− 0 (B. With the usual scattering boundary conditions: φs (x) = exp(ikx) + r exp(−ikx). and T = − 0 . x . which are useful for TDDFT. x .8) . − √ i 2 +Z √ (B. x .2) = −Z /2.
3 H atom For the Hamiltonian the energies are (B. x0 (B. 1 ω. x x0 2 (B. RESULTS FOR SIMPLE ONEELECTRON SYSTEMS B. .15) φ0 (x) = x0 π φ1 (x) = φ2 (x) = 2 1 −1 √ n e 2 φn (x) = x0 π2 n! Hn x . 2.17) 2x0 π 1 √ (2 x x0 2 − 1)e −1 2 . The ground state and ﬁrst 2 excited states wave functions are 1 2 2 1 x 1 √ e− 2 x0 . (B. HARMONIC OSCILLATOR 197 198 APPENDIX B. .14) 2 2 x −1 √ e 2 x0 π x0 1 2 1 x x0 2 .13) = n+ 1 ω.16) . 1. . . 2 1 n = 0.B.2 Harmonic oscillator 1 1 d2 ˆ + ω 2 x2 H=− 2 dx2 2 n B.2. The lengthscale is given by x0 = and the wavefunctions are x x0 2 where Hn (y) are the Hermite polynomials. (B.12) (B.
200 APPENDIX C. GREEN’S FUNCTIONS Appendix C Green’s functions 199 .
• A standard in physics for many years has been: R. Nogueira. 1990). Another is Vignale’s. Berlin. 2003). 5383 (1997). W.. Fiolhais.. Marques (SpringerVerlag. New York. • In chemistry. ed. Gritsenko.202 APPENDIX D. such as reactivitity theory.K. while part B has a very useful survey of properties and how well approximations do for them. Chem. but uses some higherlevel physics concepts. Holthausen.M. Koch and M. Part A deals with theory. • Perhaps the best overview comes from a recent summer school. 201 . Dreizler and E. Baerends and O.J. and M. E. • A good generalpurpose book for solidstate physicists about electronic structure is Richard Martin’s book. Parr and W. This is very careful and rigorous. Appendix D Further reading Chapter 1 There are now a variety of sources to choose from to learn about density functional theory in general. A 101. Weinheim. R. Yang (Oxford.. • For chemists. J. C. F.G. Chapter 3 There are many books on basic quantum mechanics. 1989).U. FURTHER READING • A great old one that demonstrates the unity of quantum is Morisson. Density Functional Theory (SpringerVerlag. 2000). a very useful guide is A chemist’s guide to density functional theory. Gross. (WileyVCH. we have Density Functional Theory of Atoms and Molecules. NY. A Primer in Density Functional Theory. • The one I enjoy the most is Griﬃths. which includes many concepts particularly useful for chemistry.V.C. • A nice discussion of using KohnSham orbitals to understand chemistry appears in A quantum chemical view of density functional theory. Phys.
Discussion of questions Chapter 1 1.. at whose center is a tall but ﬁnite barrier. In what way will TSloc change if spin is included. TSloc does very well. is it equal to the physical kinetic energy? No. Yes. I should simply ﬁt the constant to the value for that problem. where ni (x) is the loc loc density with i particle in the box. but were not told the constant of proportionality. But we can ﬁgure it out easily in this case.e. If you can do any calculation you like. The chemist’s answer is likely to be more accurate than the physicist’s for problems that are similar to the ﬁtted one. DISCUSSION OF QUESTIONS Appendix E Both answers have value. Thus. Does TSloc [n∗ ] work for a single electron in an excited state. i.. with systematic errors (but larger than the chemist’s on the ﬁtted systems). ﬁll in).2. Here we just doubly occupy each level in the box. which is deﬁned to be the same in both systems. Suppose you’d been told that TSloc is proportional to dxn3 (x).e. and so depends only on the oneelectron density. Errors are likely to be both positive or negative. surely including information on the gradient of the density can help accuracy.204 APPENDIX E. the lowest molecular orbital is √ (E. Repeat above question for 1/r . a practical chemist might answer that. Rudy. The simplest density functional approximation is a local one. 4.. what procedure might you use to determine the constant in TSloc ? A physicist might answer that. What form might a correction to TSloc take? A detailed answer is given in section 17. In the chapter and exercises. take the limit as N → ∞. Can you ﬁnd one. Why is a KohnSham calculation much faster than a traditional wavefunction calculation? Because a KohnSham calculation requires solving only a oneelectron problem (selfconsistently) and occupying N levels. Ignoring the exponentially small tunneling contribution. If you evaluate the kinetic energy of a KohnSham system.7/L2 5. On the other hand. a detailed answer is given in section 2. The most obvious is that.. The exact kinetic energy is then 1 (TA + TB ) = TA by 2 symmetry. On the other hand. since our electrons are not interacting. 3. On the other hand. However. Why is the density cubed in the local approximation for T S ? (see section ?? for the answer). 8. Performing a dimensional analysis.. 1/r = d3 r n(r)/r. so it fails very badly in this form.. TSloc [n∗ ] = TSloc [n].2. The kinetic energy operator is the same for the two systems. for two electrons of opposite spin in a box? A detailed answer is given in section 9. 203 . if all I want to do is solve problems that look like one electron in a box. rather than solving a coupled diﬀerential equation for a function of 3N variables. But there are cases where it fails quite badly. Rudy. he can ﬁnd the energy of the ﬁrst excited state by writing n∗ (x) = n2 (x) − n1 (x). how does T Sloc Again. Then T [n∗ ] = T2 − T1 = 17. Need to show explicit failure for particle in box. since TS becomes more and more accurate as the number of electrons in the box grows. i. the physicist’s answer is likely to be more accurate over a broad range of problems. We saw in the chapter how that yields the π 2 /6 constant. and since T ∼ 1/L2 = dxnp (x).. Rudy. but reﬂect diﬀerent priorities.(Rudy. and φB (x) is that for the right. 6.1) φ(x) = (φA (x) + φB (x))/ 2 where φA (x) is the groundstate wavefunction for the electron to be in the left side.. with density n ∗ (r)? For ANY single excited state of the particle in the box. we note that n ∼ 1/L in one dimension. Rudy points out that. But we can make intelligent suggestions here. and say why? Consider a particle in a large box. If we add an inﬁnitesimal to n(x) at a point. p must be 3. 2. δ(x) as change? → 0... 9.. 7. e. which can be measured in scattering experiments. since using the density at a point works so robustly. it is about 4 times too small! 10.2. our local functional uses n(x) = (nA (x) + nB (x))/2. but their wavefunctions are diﬀerent. so loc loc loc that TSloc = (TA + TB )/8 = TA /4.g.
We see shortly how this can be a handy trick. .2. then d3 r δn(r) = 0.g. exact) is to take linear combinations of the deltafunction solutions. the function is local. by integrating by parts. Take the second functional derivative. To understand this.e. You might take either the particle in the box or the harmonic oscillator or the onedimensional hydrogen atom. The HF calculation correctly writes the wavefunction as a spatially symmetric singlet. and. Is the HF estimate of the ionization potential for 1d He an overestimate or underestimate? The ionization potential is I = E1 − E2 for a twoelectron system. think of functions. i. so that the functional derivative is not determined up to a constant. where Z is the strength of the delta function. Note that just saying if the integrand depends only on the argument at r is not enough. Which is bigger. I HF < I. 3. 2/L cos(πx/L). e. exp(−αx) and exp(αx) vanishing at the edges. Suggest a good trial wavefunction for a potential that consists of a negative delta function in the middle of a box of width L. Comment on what it does right. where EN is the energy with N electrons. Since E HF > E. The groundstate wavefunction is that normalized. but equal to the functional for von Weisacker. since all potentials are homogeneous of order−1. You should ﬁnd that a plot of the two functional derivatives gives you two very diﬀerent curves in each case. since there are two diﬀerent powers in the potentials. Imagine one function being a horizontal line. Is there a simple relationship between TS and dx n(x)δTS /δn(x)? You should have found that the integral is three times the functional for the local approximation. and what it does wrong. Suggest a simple improvement. What is the exact kinetic energy density functional for one electron in onedimension? It is the von Weisacker functional. 4. but incorrectly approximates that as a product of separate functions of x1 and x2 (orbitals). 6. Thus addition of any constant to a functional derivative does not alter the result. What is the eﬀect of having nuclear charge Z = 1 for the 1d Hatom? It alters the length scale of the wavefunction without changing its normalization. 3. Much better (in fact. since integrands can change within functionals. Its energy is obviously π 2 /2L2 − 2Z/L. For ﬁxed particle number. 2.. with equality for one electron. the situation is more complicated.2) 4. It correctly has satisﬁes the cusp condition at the nucleus. √ Z exp(−Zx). Compare the functional derivative of TSVW [n] with TSloc [n] for some sample oneelectron problem. Chapter 3 1. is there any indeterminancy in the functional derivative of a density functional? If the particle number does not change. where V includes both the external and the electronelectron interaction. the kinetic energy of the true wavefunction or that of the HF wavefunction for 1d He? Hooke’s atom? The virial theorem applies to this problem (see section X). unlike the scaled orbital solution. For Hooke’s atom. E = −T = −V /2. TSVW = 1 2 ∞ −∞ dx φ (x)2 = 1 8 ∞ −∞ dx n (x)2 /n(x). 5. (E. 2.. Devise a method for deducing if a functional is local or not. If its proportional to a delta function. antisymmetric wavefunction that has density n(r) and minimizes the kinetic plus Coulomb repulsion operators.205 206 APPENDIX E. The second function is a good approximation to ﬁrst everywhere. and changes decay at large distance. then T HF < T . DISCUSSION OF QUESTIONS Chapter 2 1. Chapters 46 1. but its derivative is very diﬀerent. Since E HF ≥ E. and a second function having weak but rapid oscillations around that line. Consider the approximate HF calculation given in section 4. But its energy is not low enough. Comment. The simplest guess is just that of the particle in the box.
3. See further work whead. e. Obviously. See above. Again. The KohnSham wavefunction of density n(r) is that normalized. and when fully polarized. and evaluating the Fock integral on its orbitals. because these orbitals could 2 have come from anywhere. The kinetic energy of those two orbitals is then TS . One can do a HF calculation. 7. 5. vS↑ (r) = vS (r). they coincide when the system is unpolarized. No. antisymmetric wavefunction that has density n(r) and minimizes the kinetic operator. HF 4. They might be. and that tells you the change. 6. yielding orbitals and density. DISCUSSION OF QUESTIONS 2. with two occupied orbitals. a local potential that. get the kinetic energy of its orbitals. . There is no reason to think they are KohnSham orbitals. But this is not EX [n ]. Then ask what single potential all electrons must feel in the KohnSham equations to reproduce that density. which could only be found by ﬁnding the local potential vS (r) whose orbitals add up to nHF (r). Rather. All the same reasoning applies for EX [n]. solve the KohnSham equations for up and down noninteracting electrons in these potentials. vS↓ (r) = 0 or undetermined. one must calculate the new density. if we alter one of the original orbitals.g. yields that density. The additional ﬂexibility of spinDFT over DFT means that its much easier to make good approximations for spinpolarized systems. Given vSσ (r). The way to get TS is to construct the density n(x) and then ﬁnd vS (x). Then EX is HF simply the Fock integral of those orbitals. and construct the total density. the change in TS is not the kinetic contribution directly due to that orbital. so that vSσ (r) = vS (r). HartreeFock orbitals.. construct the new KohnSham potential.207 208 APPENDIX E. The KohnSham kinetic energy is not 1 dx φ1 2 + φ2 2 .
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue listening from where you left off, or restart the preview.