This page intentionally left blank

VARIATIONAL PRINCIPLES AND METHODS IN

THEORETICAL PHYSICS AND CHEMISTRY

This book brings together the essential ideas and methods behind current applica-
tions of variational theory in theoretical physics and chemistry. The emphasis is on
understanding physical and computational applications of variational methodology
rather than on rigorous mathematical formalism.
The text begins with an historical survey of familiar variational principles in
classical mechanics and optimization theory, then proceeds to develop the vari-
ational principles and formalism behind current computational methodology for
bound and continuum quantum states of interacting electrons in atoms, molecules,
and condensed matter. It covers multiple scattering theory, as applied to electrons
in condensed matter and in large molecules. The specific variational principles
developed for electron scattering are then extended to include a detailed presenta-
tion of contemporary methodology for electron-impact rotational and vibrational
excitation of molecules. The book also provides an introduction to the variational
theory of relativistic fields, including a detailed treatment of Lorentz and gauge
invariance for the nonabelian gauge field of modern electroweak theory.
Ideal for graduate students and researchers in any field that uses variational
methodology, this book is particularly suitable as a backup reference for lecture
courses in mathematical methods in physics and theoretical chemistry.

Robert K. Nesbet obtained his BA in physics from Harvard College in 1951

and his PhD from the University of Cambridge in 1954. He was then a research
associate in MIT for two years, before becoming Assistant Professor of Physics
at Boston University. He did research at RIAS, Martin Company, Baltimore, the
Institut Pasteur, Paris and Brookhaven National Laboratory, before becoming a
Staff Member at IBM Almaden Research Center in San Jose in 1962. He acted as
an Associate Editor for the Journal of Computational Physics and the Journal of
Chemical Physics, between 1969 and 1974, and was a visiting professor at several
universities throughout the world. Professor Nesbet officially retired in 1994, but
has continued his research and visiting since then. Over the years he has written
more than 270 publications in computational physics, atomic and molecular physics,
theoretical chemistry, and solid-state physics.
VARIATIONAL PRINCIPLES AND
METHODS IN THEORETICAL
PHYSICS AND CHEMISTRY

ROBERT K. NESBET
IBM Almaden Research Center
         
The Pitt Building, Trumpington Street, Cambridge, United Kingdom

  

The Edinburgh Building, Cambridge CB2 2RU, UK
40 West 20th Street, New York, NY 10011-4211, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
Ruiz de Alarcón 13, 28014 Madrid, Spain
Dock House, The Waterfront, Cape Town 8001, South Africa

http://www.cambridge.org

© Robert K. Nesbet 2004

First published in printed format 2002

ISBN 0-511-04162-4 eBook (netLibrary)

ISBN 0-521-80391-8 hardback
To Anne, Susan, and Barbara
 Contents

Preface page xiii

I Classical mathematics and physics 1
1 History of variational theory 3
1.1 The principle of least time 4
1.2 The variational calculus 5
1.2.1 Elementary examples 7
1.3 The principle of least action 8
2 Classical mechanics 11
2.1 Lagrangian formalism 11
2.1.1 Hamilton’s variational principle 12
2.1.2 Dissipative forces 12
2.1.3 Lagrange multiplier method for constraints 13
2.2 Hamiltonian formalism 14
2.2.1 The Legendre transformation 14
2.2.2 Transformation from Lagrangian to Hamiltonian 15
2.2.3 Example: the central force problem 16
2.3 Conservation laws 17
2.4 Jacobi’s principle 18
2.5 Special relativity 20
2.5.1 Relativistic mechanics of a particle 21
2.5.2 Relativistic motion in an electromagnetic field 22
3 Applied mathematics 25
3.1 Linear systems 25
3.2 Simplex interpolation 26
3.2.1 Extremum in n dimensions 27
3.3 Iterative update of the Hessian matrix 28
3.3.1 The BFGS algorithm 29
3.4 Geometry optimization for molecules 30

vii
viii Contents

3.4.1 The GDIIS algorithm 31

3.4.2 The BERNY algorithm 31
II Bound states in quantum mechanics 33
4 Time-independent quantum mechanics 35
4.1 Variational theory of the Schrödinger equation 36
4.1.1 Sturm–Liouville theory 36
4.1.2 Idiosyncracies of the Schrödinger equation 38
4.1.3 Variational principles for the Schrödinger equation 40
4.1.4 Basis set expansions 41
4.2 Hellmann–Feynman and virial theorems 43
4.2.1 Generalized Hellmann–Feynman theorem 43
4.2.2 The hypervirial theorem 43
4.2.3 The virial theorem 44
4.3 The N -electron problem 45
4.3.1 The N -electron Hamiltonian 45
4.3.2 Expansion in a basis of orbital wave functions 46
4.3.3 The interelectronic Coulomb cusp condition 48
4.4 Symmetry-adapted functions 49
4.4.1 Algorithm for constructing symmetry-adapted
functions 50
4.4.2 Example of the method 51
5 Independent-electron models 53
5.1 N -electron formalism using a reference state 54
5.1.1 Fractional occupation numbers 55
5.1.2 Janak’s theorem 56
5.2 Orbital functional theory 57
5.2.1 Explicit components of the energy functional 57
5.2.2 Orbital Euler–Lagrange equations 58
5.2.3 Exact correlation energy 59
5.3 Hartree–Fock theory 61
5.3.1 Closed shells – unrestricted Hartree–Fock (UHF) 61
5.3.2 Brillouin’s theorem 62
5.3.3 Open-shell Hartree–Fock theory (RHF) 62
5.3.4 Algebraic Hartree–Fock: finite basis expansions 64
5.3.5 Multiconfiguration SCF (MCSCF) 64
5.4 The optimized effective potential (OEP) 65
5.4.1 Variational formulation of OEP 67
5.5 Density functional theory (DFT) 68
5.5.1 The Hohenberg–Kohn theorems 68
5.5.2 Kohn–Sham equations 70
 Contents ix

5.5.3 Functional derivatives and local potentials 71

5.5.4 Thomas–Fermi theory 72
5.5.5 The Kohn–Sham construction 74
6 Time-dependent theory and linear response 77
6.1 The time-dependent Schrödinger equation for one electron 78
6.2 The independent-electron model as a quantum field theory 79
6.3 Time-dependent Hartree–Fock (TDHF) theory 81
6.3.1 Operator form of Hartree–Fock equations 81
6.3.2 The screening response 81
6.4 Time-dependent orbital functional theory (TOFT) 83
6.4.1 Remarks on time-dependent theory 83
6.4.2 Exact linear response theory 84
6.4.3 Definition of the response kernel 84
6.5 Reconciliation of N -electron theory and orbital models 85
6.6 Time-dependent density functional theory (TDFT) 86
6.7 Excitation energies and energy gaps 89
III Continuum states and scattering theory 91
7 Multiple scattering theory for molecules and solids 93
7.1 Full-potential multiple scattering theory 95
7.1.1 Definitions 96
7.1.2 Two-center expansion 96
7.1.3 Angular momentum representation 97
7.1.4 The surface matching theorem 99
7.1.5 Surface integral formalism 100
7.1.6 Muffin-tin orbitals and atomic-cell orbitals 101
7.1.7 Tail cancellation and the global matching function 102
7.1.8 Implementation of the theory 103
7.2 Variational principles 104
7.2.1 Kohn–Rostoker variational principle 104
7.2.2 Convergence of internal sums 106
7.2.3 Schlosser–Marcus variational principle 108
7.2.4 Elimination of false solutions 111
7.3 Energy-linearized methods 113
7.3.1 The LMTO method 113
7.3.2 The LACO method 115
7.3.3 Variational theory of linearized methods 116
7.4 The Poisson equation 118
7.5 Green functions 120
7.5.1 Definitions 121
7.5.2 Properties of the Green function 124
x Contents

7.5.3 Construction of the Green function 125

8 Variational methods for continuum states 129
8.1 Scattering by an N -electron target system 129
8.1.1 Cross sections 132
8.1.2 Close-coupling expansion 133
8.2 Kohn variational theory 134
8.2.1 The matrix variational method 135
8.2.2 The Hulthén–Kohn variational principle 137
8.2.3 The complex Kohn method 139
8.3 Schwinger variational theory 140
8.3.1 Multichannel Schwinger theory 143
8.3.2 Orthogonalization and transfer invariance 145
8.4 Variational R-matrix theory 147
8.4.1 Variational theory of the R-operator 154
8.4.2 The R-operator in generalized geometry 156
8.4.3 Orbital functional theory of the R-matrix 157
9 Electron-impact rovibrational excitation of molecules 161
9.1 The local complex-potential (LCP) model 163
9.1.1 The projection-operator method 164
9.2 Adiabatic approximations 166
9.2.1 The energy-modified adiabatic
approximation (EMA) 168
9.3 Vibronic R-matrix theory 169
9.3.1 Phase-matrix theory 172
9.3.2 Separation of the phase matrix 173
9.3.3 Phase-matrix formalism: EMAP 174
9.3.4 Nonadiabatic theory: NADP 175
IV Field theories 179
10 Relativistic Lagrangian theories 181
10.1 Classical relativistic electrodynamics 182
10.1.1 Classical dynamical mass 184
10.1.2 Classical renormalization and the Dirac equation 185
10.2 Symmetry and Noether’s theorem 186
10.2.1 Examples of conservation laws 187
10.3 Gauge invariance 189
10.3.1 Classical electrodynamics as a gauge theory 190
10.3.2 Noether’s theorem for gauge symmetry 191
10.3.3 Nonabelian gauge symmetries 192
10.3.4 Gauge invariance of the SU (2) field theory 195
 Contents xi

10.4 Energy and momentum of the coupled fields 197

10.4.1 Energy and momentum in classical
electrodynamics 197
10.4.2 Energy and momentum in SU (2) gauge theory 199
10.5 The Standard Model 201
10.5.1 Electroweak theory (EWT) 202
10.5.2 Quantum chromodynamics (QCD) 203
References and bibliography 205
Index 225
 Preface

As theoretical physics and chemistry have developed since the great quantum rev-
olution of the 1920s, there has been an explosive speciation of subfields, perhaps
comparable to the late Precambrian period in biological evolution. The result is
that these life-forms not only fail to interbreed, but can fail to find common ground
even when placed in proximity on a university campus. And yet, the underlying
intellectual DNA remains remarkably similar, in analogy to the findings of recent
research in biology. The purpose of this present text is to identify common strands
in the substrate of variational theory and to express them in a form that is intelligi-
ble to participants in these subfields. The goal is to make hard-won insights from
each line of development accessible to others, across the barriers that separate these
specialized intellectual niches.
Another great revolution was initiated in the last midcentury, with the introduction
of digital computers. In many subfields, there has been a fundamental change
in the attitude of practicing theoreticians toward their theory, primarily a change
of practical goals. There is no longer a well-defined barrier between theory for
the sake of understanding and theory for the sake of predicting quantitative data.
Given modern resources of computational power and the coevolving development
of efficient algorithms and widely accessible computer program tools, a formal
theoretical insight can often be exploited very rapidly, and verified by quantitative
implications for experiment. A growing archive records experimental controversies
that have been resolved by quantitative computational theory.
It has been said that mathematics is queen of the sciences. The variational branch
of mathematics is essential both for understanding and predicting the huge body
of observed data in physics and chemistry. Variational principles and methods lie
in the bedrock of theory as explanation, and theory as a quantitative computational
tool. Quite simply, this is the mathematical foundation of quantum theory, and
quantum theory is the foundation of all practical and empirical physics and chem-
istry, short of a unified theory of gravitation. With this in mind, the present text is

xiii
xiv Preface

subdivided into four parts. The first reviews the variational concepts and formalism
that developed over a long history prior to the discovery of quantum mechanics,
subdivided into chapters on history, on classical mechanics, and on applied math-
ematics (severely truncated out of respect for the vast literature already devoted
to this subject). The second part covers variational formalism and methodology in
subfields concerned with bound states in quantum mechanics. There are separate
chapters on time-independent quantum mechanics, on independent-electron mod-
els, which may at some point be extended to independent-fermion models as the
formalism of the Standard Model evolves, and on time-dependent theory and linear
response. The third part develops the variational theory of continuum states, includ-
ing chapters on multiple scattering theory (the essential formalism for electronic
structure calculations in condensed matter), on scattering theory relevant to the
true continuum state of a quantum target and an external fermion (with emphasis
on methodology for electron scattering by atoms and molecules), continuing to a
separate chapter on the currently developing theory of electron-impact rotational
and vibrational excitation of molecules. The fourth part develops variational theory
relevant to relativistic Lagrangian field theories. The single chapter in this part de-
rives the nonquantized field theory that underlies the quantized theory of the current
Standard Model of elementary particles.
This book grew out of review articles in specialized subfields, published by
the author over nearly fifty years, including a treatise on variational methods in
electron–atom scattering published in 1980. Currently relevant topics have been
extracted and brought up to date. References that go more deeply into each of the
topics treated here are included in the extensive bibliography. The purpose is to
set out the common basis of variational formalism, then to open up channels for
further exploration by any reader with specialized interests. The most recent source
of this text is a course of lectures given at the Scuola Normale Superiore, Pisa, Italy
in 1999. These lectures were presented under the present title, but concentrated
on the material in Parts I and II here. The author is indebted to Professor Renato
Colle, of Bologna and the Scuola Normale, for making arrangements that made
these lectures possible, and to the Scuola Normale Superiore for sponsoring the
lecture series.
 I
Classical mathematics and physics

This part is concerned with variational theory prior to modern quantum

mechanics. The exception, saved for Chapter 10, is electromagnetic the-
ory as formulated by Maxwell, which was relativistic before Einstein,
and remains as fundamental as it was a century ago, the first example of a
Lorentz and gauge covariant field theory. Chapter 1 is a brief survey of the
history of variational principles, from Greek philosophers and a religious
faith in God as the perfect engineer to a set of mathematical principles that
could solve practical problems of optimization and rationalize the laws
of dynamics. Chapter 2 traces these ideas in classical mechanics, while
Chapter 3 discusses selected topics in applied mathematics concerned
with optimization and stationary principles.
 1
History of variational theory

The principal references for this chapter are:

[5] Akhiezer, N.I. (1962). The Calculus of Variations (Blaisdell, New York).
[26] Blanchard, P. and Brüning, E. (1992). Variational Methods in Mathematical Physics
(Springer-Verlag, Berlin).
[78] Dieudonné, J. (1981). History of Functional Analysis (North-Holland, Amsterdam).
[147] Goldstine, H.H. (1980). A History of the Calculus of Variations from the 17th
through the 19th Century (Springer-Verlag, Berlin).
[210] Lanczos, C. (1966). Variational Principles of Mechanics (University of Toronto
Press, Toronto).
[322] Pars, L.A. (1962). An Introduction to the Calculus of Variations (Wiley, New York).
[436] Yourgrau, W. and Mandelstam, S. (1968). Variational Principles in Dynamics and
Quantum Theory, 3rd edition (Dover, New York).

The idea that laws of nature should satisfy a principle of simplicity goes back
at least to the Greek philosophers [436]. The anthropomorphic concept that the
engineering skill of a supreme creator should result in rules of least effort or of most
efficient use of resources leads directly to principles characterized by mathematical
extrema. For example, Aristotle (De Caelo) concluded that planetary orbits must be
perfect circles, because geometrical perfection is embodied in these curves: “. . . of
lines that return upon themselves the line which bounds the circle is the shortest.
That movement is swiftest which follows the shortest line”. Hero of Alexandria
(Catoptrics) proved perhaps the first scientific minimum principle, showing that
the path of a reflected ray of light is shortest if the angles of incidence and reflection
are equal.
The superiority of circular planetary orbits became almost a religious dogma
in the Christian era, intimately tied to the idea of the perfection of God and of
His creations. It was replaced by modern celestial mechanics only after centuries in
which the concept of esthetic perfection of the universe was gradually superseded by
a concept of esthetic perfection of a mathematical theory that could account for the

3
4 1 History of variational theory

actual behavior of this universe as measured in astronomical observations. Aspects

of value-oriented esthetics lay behind Occam’s logical “razor” (avoid unnecessary
hypotheses), anticipating the later development of observational science and the
search for an explanatory theory that was both as general as possible and as simple
as possible. The path from Aristotle to Copernicus, Brahe, Kepler, Galileo, and
Newton retraces this shift from a priori purity of concepts to mathematical theory
solidly based on empirical science. The resulting theory of classical mechanics
retains extremal principles that are the basis of the variational theory presented
here in Chapter 2.
Variational principles have turned out to be of great practical use in modern
theory. They often provide a compact and general statement of theory, invariant
or covariant under transformations of coordinates or functions, and can be used to
formulate internally consistent computational algorithms. Symmetry properties are
often most easily derived in a variational formalism.

1.1 The principle of least time

The law of geometrical optics anticipated by Hero of Alexandria was formulated
by Fermat (1601–1655) as a principle of least time, consistent with Snell’s law of
refraction (1621). The time for phase transmission from point P to point Q along
a path x(t) is given by

Q
ds
T = , (1.1)
P v(s)

where ds is a path element, and v is the phase velocity. Fermat’s principle is that
the value of the integral T should be stationary with respect to any infinitesimal
deviation of the path x(t) from its physical value. This is valid for geometrical optics
as a limiting case of wave optics. The mathematical statement is that δT = 0 for
all variations induced by displacements δx(t). In this and subsequent variational
formulas, differentials defined by the notation δ · · · are small increments evaluated
in the limit that quadratic infinitesimals can be neglected. Thus for sufficiently small
displacements δx(t), the integral T varies quadratically about its physical value. For
planar reflection consider a ray path from P : (−d, −h) to the observation point
Q : (−d, h) via an intermediate point (0, y) in the reflection plane x = 0. Elapsed
time in a uniform medium is
  
T (y) = d + (h + y) + d + (h − y)
2 2 2 2 v, (1.2)

to be minimized with respect to displacements in the reflection plane parametrized

 1.2 The variational calculus 5

by y. The angle of incidence θi is defined such that

h+y
sin θi = 
d2 + (h + y)2
and the angle of reflection θr is defined by
h−y
sin θr =  .
d + (h − y)2
2

The law of planar reflection, sin θi = sin θr , follows immediately from

∂T
= (sin θi − sin θr )/v = 0.
∂y
To derive Snell’s law of refraction, consider the ray path from point P : (−d, −h)
to Q : (d, h) via point (0, y) in a plane that separates media of phase velocity
vi (x < 0) and vr (x > 0). The elapsed time is
 
T ( y) = vi−1 d 2 + (h + y)2 + vr−1 d 2 + (h − y)2 . (1.3)

The variational condition is

∂T
= sin θi /vi − sin θr /vr = 0.
∂y
This determines parameter y such that
sin θi vi
= , (1.4)
sin θr vr
giving Snell’s law for uniform refractive media.

1.2 The variational calculus

Derivation of a ray path for the geometrical optics of an inhomogeneous medium,
given v(r) as a function of position, requires a development of mathematics beyond
the calculus of Newton and Leibniz. The elapsed time becomes a functional T [x(t)]
of the path x(t), which is to be determined so that δT = 0 for variations δx(t)
with fixed end-points: δx P = δx Q = 0. Problems of this kind are considered in the
calculus of variations [5, 322], proposed originally by Johann Bernoulli (1696),
and extended to a full mathematical theory by Euler (1744). In its simplest form,
the concept of the variation δx(t) reduces to consideration of a modified function
x (t) = x(t) + w(t) in the limit  → 0. The function w(t) must satisfy conditions
of continuity that are compatible with those of x(t). Then δx(t) = w(t) d and the
variation of the derivative function is δx (t) = w (t) d.
6 1 History of variational theory

The problem posed by Bernoulli is that of the brachistochrone. If two points are
connected by a wire whose shape is given by an unknown function y(x) in a vertical
plane, what shape function minimizes the time of descent of a bead sliding without
friction from the higher to the lower point? The mass of a bead moving under gravity
is not relevant. It can easily be verified by trial and error that a straight line does not
give the minimum time of passage. Always in such problems, conditions appropriate
to physically meaningful solution functions must be specified. Although this is a
vital issue in any mathematically rigorous variational calculus, such conditions
will be stated as simply as possible here, strongly dependent on each particular
application of the theory. Clearly the assumed wire in the brachistochrone problem
must have the physical properties of a wire. This requires y(x) to be continuous,
but does not exclude a vertical drop. Since no physical wire can have an exact
discontinuity of slope, it is reasonable to require velocity of motion along the wire
to be conserved at any such discontinuity, so that the hypothetical sliding bead does
not come to an abrupt stop or bounce with undetermined loss of momentum. It can
easily be verified that a vertical drop followed by a horizontal return to the smooth
brachistochrone curve always increases the time of passage. Thus such deviations
from continuity of the derivative function do not affect the optimal solution.
The calculus of variations [5, 322] is concerned with problems in which a function
is determined by a stationary variational principle. In its simplest form, the problem
is to find a function
 x1 y(x) with specified values at end-points x0 , x1 such that the
integral J = x0 f (x, y, y  )d x is stationary. The variational solution is derived
from

 
∂f  ∂f
δJ = δy + δy dx = 0
∂y ∂y
after integrating by parts to eliminate δy  (x). Because



 ∂f ∂ f x1 d ∂f
δy 
d x = δy  − δy d x,
∂y ∂ y x0 dx ∂y
δ J = 0 for fixed end-points δy(x0 ) = δy(x1 ) = 0 if
∂f d ∂f
− = 0. (1.5)
∂y dx ∂y
This is a simple example of the general form of Euler’s equation (1744), derived
directly from a variational expression.
Blanchard and Brüning [26] bring the history of the calculus of variations into
the twentieth century, as the source of contemporary developments in pure math-
ematics. A search for existence and uniqueness theorems for variational problems
engendered deep studies of the continuity and compactness of mathematical entities
 1.2 The variational calculus 7

that generalize the simple intuitive definitions assumed by Euler and Lagrange. The
seemingly self-evident statement that, for free variations of the function y(x),



∂f d ∂f
− δyd x = 0
∂y dx ∂y
implies Euler’s equation, was first proven rigorously by Du Bois-Reymond in 1879.
With carefully stated conditions on the functions f and y, this made it possible to
prove the fundamental theorem of the variational calculus [26], on the existence of
extremal solutions of variational problems.

1.2.1 Elementary examples

A geodesic problem requires derivation of the shortest path connecting two points
in some system for which distance is defined, subject to constraints that can be
either geometrical or physical in nature. The shortest path between two points in a
plane follows from this theory. The problem is to minimize


x1
x1
2
 dy
J= f (x, y, y )d x = dx 1 + ,
x0 x0 dx

where
∂f ∂f ∂f y
= 0, = 0, =  .
∂x ∂y ∂y 1 + y 2
In this example, Euler’s equation takes the form of the geodesic equation
d y
 = 0.
d x 1 + y 2

The solution is y  = const, or

x1 − x x − x0
y(x) = y0 + y1 ,
x1 − x0 x1 − x0
a straight line through the points x0 , y0 and x1 , y1 .
In Johann Bernoulli’s problem, the brachistochrone, it is required to find the
shape of a wire such that a bead slides from point 0, 0 to x1 , y1 in the shortest
time T under the force of gravity. The energy equation 12 mv 2 = −mgy implies
√
v = −2gy, so that

x1
x1
ds
T = = f (y, y  ) d x,
0 v 0
8 1 History of variational theory

where f (y, y  ) = −(1 + y 2 )/2gy. Because ∂ f /∂ x = 0, the identity



d ∂f d ∂f ∂f
y  − f = y − ,
dx ∂y dx ∂y ∂y
and the Euler equation imply an integral of motion,
∂f −1
y − f = = const.
∂y −2gy(1 + y 2 )
On combining constants into the single parameter a this implies

2
dy −2a
1+ = .
dx y
The solution for a bead starting from rest at the coordinate origin is a cycloid,
determined by the parametric equations x = a(φ − sin φ) and y = a(cos φ − 1).
This curve is generated by a point on the perimeter of a circle of radius a that
rolls below the x-axis without slipping. The lowest point occurs for φ = π , with
x1 = πa and y1 = −2a. By adding a constant φ0 to φ, a can be adjusted so that the
curve passes through given points x0 , y0 and x1 , y1 .

1.3 The principle of least action

Variational principles for classical mechanics originated in modern times with the
principle of least action, formulated first imprecisely by Maupertuis and then as
an example of the new calculus of variations by Euler (1744) [436]. Although not
stated explicitly by either Maupertuis or Euler, stationary action is valid only for
motion in which energy is conserved. With this proviso, in modern notation for
generalized coordinates,

Q
δ p · dq = 0, (1.6)
P

for a path from system point P to system point Q.

For a particle of mass m moving in the (x, y) plane with force per mass (X, Y ),
instantaneous motion is described by velocity v along the trajectory. An instanta-
neous radius of curvature ρ is defined by angular momentum  = mvρ such that
the centrifugal force mv 2 /ρ balances the true force normal to the trajectory. Hence,
following Euler’s derivation, Newtonian mechanics implies that
v2 Y dx − Xdy
=
ρ d x 2 + dy 2
along the trajectory. The principle of least action requires the action integral
 1.3 The principle of least action 9

per unit mass





2
dy
v ds = v dx 1 +
dx
to be stationary. The variation of v along the trajectory is determined for fixed
energy E = T + V by


1 ∂V ∂V
v dv = − dx + dy = X d x + Y dy.
m ∂x ∂y
Thus v ∂∂vx = X and v ∂v
∂y
= Y . Euler’s equation then takes the form

d vy  Y
 − 1 + y  2 = 0,
dx 1 + y 2 v

where y  = dy/d x. The local curvature of a trajectory is defined by

1 d   1
y /(1 + y 2 ) 2 = y  /(1 + y 2 ) 2 .
3
=
ρ dx
X +Yy
Using this formula and dv
dx
= v
, Euler’s equation implies
v (X + Y y  )y  Y 
+  − 1 + y 2 = 0.
ρ v 1+y 2 v
This reproduces the formula derived directly from Newtonian mechanics:
v2 Y − Xy Y dx − Xdy
= = .
ρ 1+y 2 d x 2 + dy 2
Euler’s proof of the least action principle for a single particle (mass point in mo-
tion) was extended by Lagrange (c. 1760) to the general case of mutually interacting
particles, appropriate to celestial mechanics. In Lagrange’s derivation [436], action
along a system path from initial coordinates P to final coordinates Q is defined by

Q 
Q
A= ma va dsa = ma ẋa · d xa . (1.7)
a P a P

Variations about a true dynamical path are defined by coordinate displacements

δxa . Velocity displacements δ ẋa are constrained so as to maintain invariant total
energy. This implies modified time values at the displaced points [146]. The energy
constraint condition is

∂V

δE = m a ẋa · δ ẋa + · δxa = 0.
a ∂xa
10 1 History of variational theory

The induced variation of action is


Q
δA = ma (ẋa · dδxa + δ ẋa · dxa )
a P
 
Q
= m a ẋa · δxa | QP − ma (d ẋa · δxa − ẋa dt · δ ẋa ),
a a P

on integrating by parts and using dxa = ẋa dt. The final term here can be replaced,
using the energy constraint condition. Then, using d ẋa = ẍa dt,
 Q

Q
∂V

δA = m a ẋa · δxa | P − m a ẍa + · δxa dt.
a a P ∂xa
If the end-points are fixed, the integrated term vanishes, and A is stationary if
and only if the final integral vanishes. Since δxa is arbitrary, the integrand must
vanish, which is Newton’s law of motion. Hence Lagrange’s derivation proves that
the principle of least action is equivalent to Newtonian mechanics if energy is
conserved and end-point coordinates are specified.
 2
Classical mechanics

The principal references for this chapter are:

[26] Blanchard, P. and Brüning, E. (1992). Variational Methods in Mathematical
Physics: A Unified Approach (Springer-Verlag, Berlin).
[146] Goldstein, H. (1983). Classical Mechanics, 2nd edition (Wiley, New York).
[187] Jeffreys, H. and Jeffreys, B.S. (1956). Methods of Mathematical Physics,
3rd edition (Cambridge University Press, New York).
[208] Kuperschmidt, B.A. (1990). The Variational Principles of Dynamics (World
Scientific, New York).
[210] Lanczos, C. (1966). Variational Principles of Mechanics (University of Toronto
Press, Toronto).
[240] Mercier, A. (1959). Analytical and Canonical Formalism in Physics (Interscience,
New York).
[323] Pauli, W. (1958). Theory of Relativity, tr. G. Field (Pergamon Press, New York).
[393] Synge, J.L. (1956). Relativity: the Special Theory (Interscience, New York).
[436] Yourgrau, W. and Mandelstam, S. (1968). Variational Principles in Dynamics and
Quantum Theory, 3rd edition (Dover, New York).

2.1 Lagrangian formalism

Newton’s equations of motion, stated as “force equals mass times acceleration”,
are strictly true only for mass points in Cartesian coordinates. Many problems of
classical mechanics, such as the rotation of a solid, cannot easily be described in
such terms. Lagrange extended Newtonian mechanics to an essentially complete
nonrelativistic theory by introducing generalized coordinates q and generalized

forces Q such that the work done in a dynamical process is k Q k dqk [436]. Since
this must be the same when expressed in Cartesian coordinates, it follows that

Q k = a Xa · ∂X a
∂qk
, where the Newtonian force is Xa = − ∂X∂V
a
. Equivalently, if the
∂V
potential function is V ({q}) in generalized coordinates, then Q k = − ∂q . The New-
 ∂T
k
tonian kinetic energy T = 2 a m a ẋa defines momenta pa = ∂ ẋa = m a ẋa , which
1 2

becomes pk = ∂∂q̇T when kinetic energy is expressed as T ({q, q̇}). The equations
k
of motion ṗa = X transform into ṗ k = Q k . Although this can be shown by direct

11
12 2 Classical mechanics

transformation, a more elegant and ultimately more general procedure is to prove

the variational principle of Hamilton in a form that is valid for any choice of gen-
eralized coordinates.

2.1.1 Hamilton’s variational principle

The principle of least action suffers from the awkward constraint that energy must be
fixed on nonphysical displaced trajectories. Following the introduction by Lagrange
of a dynamical formalism using generalized coordinates, it was shown by Hamilton
that a revised variational principle could be based on a new definition of action that
had the full generality of Lagrange’s theory. Hamilton’s definition of the action
integral is

t1
I = L dt, (2.1)
t0

where the Lagrangian L is an explicit function of generalized coordinates {q} and

their time derivatives {q̇},

L = T (q, q̇) − V (q, q̇). (2.2)

Generalized momenta are defined by pk = ∂∂q̇Lk and generalized forces are defined
∂L
by Q k = ∂q k
. When applied using t as the independent variable, Euler’s variational
equation for the action integral I takes the form of Lagrange’s equations of motion
d ∂L ∂L
− = 0. (2.3)
dt ∂ q̇ ∂q
Hamilton considered variations δq(t) that define continuous generalized trajectories
with fixed end-points, q(t0 ) = q0 and q(t1 ) = q1 . The variational expression is

t1  
∂L ∂L
δI = dt δ q̇(t) + δq(t) = 0.
t0 ∂ q̇ ∂q

Following the logic of Euler, after integrating by parts to replace the term in δ q̇ by
one in δq, this implies the Lagrangian equations of motion.

2.1.2 Dissipative forces

Hamilton’s principle exploits the power of generalized coordinates in problems with
static or dynamical constraints. Going beyond the principle of least action, it can also
treat dissipative forces, not being restricted to conservative systems. If energy loss
 2.1 Lagrangian formalism 13

during system motion due to a nonconservative force is −dW , where d W =


 t1k Q̂ k dqk , then a dissipative action function is defined by its variation δ I =
t0 (δL + δW )dt. Hamilton’s variational principle implies modified equations of
motion
d ∂L ∂L
− = Q̂ k .
dt ∂ q̇ k ∂qk

The dissipative term takes the form of a nonconservative force, Q̂ k , acting on the
generalized coordinate qk .

2.1.3 Lagrange multiplier method for constraints


The differential form d X j = k ak j dqk + akt dt = 0 defines a general linear
constraint condition. For an integrable or holonomic constraint, expressed by
∂X
X j ({qk }, t) = 0, the coefficients are partial derivatives of X j such that ak j = ∂qkj
∂X
and akt = ∂t j . Nonintegrable or nonholonomic constraints are defined by the differ-
ential form. In applications of Hamilton’s principle, δ X j differs from d X j because
the displacement of each point on a trajectory is defined such that δt = 0. Hence
the coefficients akt drop out of δ X j .
Lagrange introduced the very powerful method of Lagrange multipliers to incor-
porate such constraints into the formalism of analytical dynamics. The basic idea
is that if a term X j λ j is added to L, its gradients provide a generalized internal
force that dynamically enforces the desired constraint. The Lagrange multiplier
λ j provides a generalized coordinate whose value is to be determined so that the
constraint condition is satisfied. This artificial coordinate characterizes an effective
potential well whose extremum stabilizes a system conformation that satisfies the
constraint condition. Because the added term X j λ j vanishes when the constraint
condition is satisfied, it does not change any physical properties of the system except
by avoiding dynamical states incompatible with the constraint.
The equations of motion, generalized to include holonomic or nonholonomic
constraints with Lagrange multipliers, are
d ∂L ∂L 
− − ak j λ j = 0.
dt ∂ q̇ k ∂qk j

These equations are to be supplemented by the alternative constraint conditions:

X j ({qk }, t) = 0,

ak j q̇ k + akt = 0,
k
14 2 Classical mechanics

for holonomic and nonholonomic constraints, respectively. The number of equa-

tions equals the number of unknown quantities {qk , λ j } in either case. Forces of

constraint Q̂ k = j ak j λ j appear in the modified Lagrange equations.
A simple example of the use of Lagrange multipliers is a hoop of mass M
and radius r rolling without slipping down an inclined plane [146]. Appropri-
ate generalized coordinates are x, the distance from the top of the plane to the
point of contact and θ , the angle of rotation of the hoop. If the plane is at
a fixed angle φ from the horizontal, kinetic and potential energy are, respec-
2
tively, T = 12 M ẋ 2 + 12 Mr 2 θ̇ , V = −Mgx sin φ subject to the differential con-
straint d X = r dθ − d x = 0. δ(T − V + X λ) replaces δL, where λ is the Lagrange
multiplier for the constraint. The coefficients in the differential constraint form are
aθ = r and ax = −1. The equations of motion and constraint are

M ẍ − Mg sin φ + λ = 0,
Mr 2 θ¨ − λr = 0,
r θ̇ − ẋ = 0.

The solution is simplified because r θ̈ = ẍ implies M ẍ = λ, so that ẍ = g sin φ/2.

Hence λ = Mg sin φ/2, and −λ is the translational force of constraint. The angu-
lar acceleration is θ̈ = g sin φ/2r . The hoop rolls with half the unconstrained
translational acceleration.

2.2 Hamiltonian formalism

Generalizing Newton’s equations of motion, Lagrange’s equations also set the time
rate of change of momenta p equal to forces Q. Hamilton recognized that these
generalized momenta could replace the time derivatives q̇ as fundamental variables
of the theory. This is most directly accomplished by a Legendre transformation, as
described in the following subsection.

2.2.1 The Legendre transformation

Given F(x, y) such that
∂F ∂F
d F = u(x, y) d x + v(x, y) dy = dx + dy,
∂x ∂y
it is often desirable to transform to different independent variables (u, y). The
alternative function G(u, y) = F(x, y) − ux characterizes a Legendre trans-
formation, such that dG = d F − u d x − x du = v dy − x du. Thus the partial
 2.2 Hamiltonian formalism 15

derivatives are
∂G ∂G
= v(x, y), = −x,
∂y ∂u
expressed as functions of the original variables.
This transformation is used to define thermodynamic functions that depend
on easily measurable variables such as pressure and temperature [57, 403]. For
example, the thermodynamic function X :enthalpy is determined by V :volume,
S:entropy, T :temperature, and P:pressure through the differential form d X =
T dS + V dP. A Legendre transformation from the abstractly defined entropy S to
the directly measurable temperature T defines the Gibbs function G = X − T S.
Changes dG = −S dT + V dP of the Gibbs function are determined by temperature
and pressure changes. In an isothermal process, ∂G∂P
= V is the system volume. In
∂G
an isobaric process, ∂ T = −S directly measures the entropy.

2.2.2 Transformation from Lagrangian to Hamiltonian

Independent variables {q̇} in a general time-dependent Lagrangian L are replaced
by momenta { p} using a Legendre transformation that defines the Hamiltonian

function H ( p, q, t) = k pk q̇ k − L(q, q̇, t). Using the definition pk = ∂∂q̇L and the
k
Lagrangian equations of motion, the transformation removes terms in d q̇ k from the

differential form d H = k {q̇ k dpk − ṗ k dqk } − ∂∂tL dt. The implied partial deriva-
tives give Hamilton’s equations of motion
∂H
q̇ k = ,
∂ pk
(2.4)
∂H
− ṗ k = .
∂qk
Alternatively, an auxiliary variable wk can be introduced, constrained to be
dynamically equal to q̇ k using a Lagrange multiplier that turns out to be the
substituted variable pk . The constraint condition in this ingenious procedure is
X k = q̇ k − wk = 0. The modified Lagrangian is

L  = L(q, w) + pk (q̇ k − wk )
k
  
= pk q̇ k − pk wk − L(q, w)
k k

= pk q̇ k − H ( p, q, w),
k

in agreement with the Legendre transformation. The constrained equations of

16 2 Classical mechanics

motion are

∂ L ∂H
= q̇ k − wk = q̇ k − = 0;
∂ pk ∂ pk
d ∂ L ∂ L ∂H
− = ṗ k + = 0;
dt ∂ q̇ k ∂qk ∂qk
∂ L ∂L ∂H
= − pk = − = 0.
∂wk ∂wk ∂wk

The first and second of these equations are Hamilton’s equations of motion. The
∂L
first and third establish wk = q̇ k and pk = ∂w k
as dynamical conditions, equivalent
∂L
to pk = ∂ q̇ .
k

2.2.3 Example: the central force problem

In spherical polar coordinates, for one mass point moving in a central potential
V = V (r ),

1 2
T = m(ṙ 2 + r 2 θ̇ ).
2

p2 p2
The Hamiltonian is H = 2mr + 2mrθ 2 + V (r ), where pr = ∂∂Tṙ = m ṙ , and pθ = ∂∂Tθ̇ =
mr 2 θ̇ . Using the relevant partial derivatives, Hamilton’s equations of motion are

pθ
ṗ θ = 0; θ̇ = ;
mr 2
2 ∂V pr
ṗr = mr θ̇ − ; ṙ = .
∂r m

For specified energy E and angular momentum pθ = const = , on solving H = E

for pr ,

1
m ṙ = pr = [2m(E − V (r )) − 2 /r 2 ] 2 .

A trajectory r (t) is obtained by integrating ṙ between turning points, where the

argument of the square root vanishes. Using dθ = mr 2 drṙ , this formula for ṙ provides
an equation to be integrated for a classical orbit,

dθ 
= 2 [2m(E − V (r )) − 2 /r 2 ]− 2 .
1

dr r
 2.3 Conservation laws 17

2.3 Conservation laws

Conservation of energy follows directly from Hamilton’s equations. The differential
change of the Hamiltonian along a trajectory is
 ∂H
dH = (q̇ k d pk − ṗ k dqk ) + dt,
k
∂t

using the Hamiltonian equations of motion. The total time derivative is

dH  ∂H ∂H
= (q̇ k ṗ k − ṗ k q̇ k ) + = .
dt i
∂t ∂t

Hence, stated as a theorem,

∂H
= 0 ⇒ H (t) = E = const,
∂t
and energy is conserved unless H is explicitly time-dependent.
∂H
Since Hamilton’s equations imply that ṗ k = 0 if ∂q k
= 0, pk is a constant of
motion if qk is such an ignorable coordinate. An ingenious choice of generalized
coordinates can produce such constants and simplify the numerical or analytic task
of integrating the equations of motion.
The close connection between symmetry transformations and conservation laws
was first noted by Jacobi, and later formulated as Noether’s theorem: invariance
of the Lagrangian under a one-parameter transformation implies the existence of a
conserved quantity associated with the generator of the transformation [304]. The
equations of motion imply that the time derivative of any function ( p, q) is

∂ ∂
 

∂ ∂ H ∂ ∂ H

=
˙ q̇ + ṗ = − ,
k
∂qk k ∂ pk k k
∂qk ∂ pk ∂ pk ∂qk

which defines a classical Poisson bracket {, H }. A constant of motion is charac-

terized by {, H } = 0. Invariance under a symmetry transformation implies such
a vanishing Poisson bracket for the symmetry generator.
By Noether’s theorem, invariance of the Lagrangian under an infinitesimal time
displacement implies conservation of energy. This is consistent with the direct proof
of energy conservation given above, when L and by implication H have no explicit
time dependence. Define a continuous time displacement by the transformation t =
t  + α(t  ) where α(t0 ) = α(t1 ) = 0, subject to α → 0. Time intervals on the original
and displaced trajectories are related by dt = (1 + α  )dt  or dt  = (1 − α  )dt. The
transformed Lagrangian is
 ∂L
L(q, q̇) = L(q, q  (1 − α  )) = L(q, q  ) − q̇ k α  .
k
∂ q̇ k
18 2 Classical mechanics

The transformed action integral is

t1
t1
Iα = L(q, q̇) dt = L(q, q  (1 − α  ))(1 + α  ) dt  .
t0 t0

In the limit α → 0 this is


t1  ∂L
Iα→0 = I − q̇ k − L α  dt  .
t0 k
∂ q̇ k

Treating α as a generalized coordinate, its equation of motion is


d ∂ L d 
=−  pk q̇ k − L = 0,
dt  ∂α  dt k

so that H = k pk q̇ k − L = const = E.
Similarly for translational invariance, an infinitesimal coordinate translation is
defined for a single particle by

xi = xi + α(t), α → 0.

Assuming invariant potential energy, the transformed kinetic energy is

1 
Tα = m i (ẋi + α̇)2 = T + m i ẋi · α̇ + O(α 2 ).
2 i i

Treating the components of α as generalized coordinates, an equation of motion is

implied in the form
∂ Lα 
= m i ẋi = const = P,
∂ α̇ i

the conservation law for linear momentum.

2.4 Jacobi’s principle

As an introduction to relativistic dynamics, it is of interest to treat time as a dy-
namical variable rather than as a special system parameter distinct from particle
coordinates. Introducing a generic global parameter τ that increases along any
generalized system trajectory, the function t(τ ) becomes a dynamical variable. In
special relativity, this immediately generalizes to ti (τ ) for each independent particle,
associated with spatial coordinates xi (τ ). Hamilton’s action integral becomes


I = Lt  dτ, (2.5)
 2.4 Jacobi’s principle 19

where t  = dt/dτ . Limiting the discussion to conservative systems, with fixed

energy E, the modified Lagrangian Lt  does not depend on t. Hence the generalized
momentum pt = ∂t∂  (Lt  ) is a constant of motion. In detail,
 ∂ L ∂
q  
pt = L + t  i
= L − pi q̇ i = −H = −E.
i
∂ q̇ i ∂t t i

Again anticipating relativistic dynamics, energy is related to momenta as time is

related to spatial coordinates.
Since time here is an “ignorable” variable, it can be eliminated from the dynamics
by subtracting pt t  from the modified Lagrangian and by solving H = E for t  as
a function of the spatial coordinates and momenta. This produces Jacobi’s version
of the principle of least action as a dynamical theory of trajectories, from which
time dependence has been removed. The modified Lagrangian is

 = Lt  − pt t  = (T − V + E )t  = 2T t  ,

such that
∂ ∂
 =  (Lt  ) − pt = 0.
∂t  ∂t

Since kinetic energy is positive, the action integral A = 2T t  dτ is nondecreas-
ing. This suggests using a global parameter τ = s defined by the Riemannian line
element

ds 2 = m i j dqi dq j ,
i, j

where the positive-definite matrix m i j defines a mass tensor as a function of gener-

alized coordinates. Then

ṡ 2 = m i j q̇ i q̇ j = 2T.
i, j

This makes it possible to express t  as a function of the generalized coordinates and

momenta. ṡ 2 = (s  /t  )2 = 2T implies that ds = (2T ) 2 dt, or t  = (2T )− 2 s  . The
1 1

reduced action integral, originally derived by Jacobi, is

 1
 1
A = 2T t dτ = (2T ) s dτ = (2T ) 2 ds.
2

Because variations are restricted to those that conserve energy, 2T = 2(E − V ) and
1
ds = (2(E − V )) 2 dt along the varied trajectories. The action integral becomes




A = 2(E − V )dt = m i j q̇ i q̇ j dt = pi dqi .
i, j i
20 2 Classical mechanics

δ A = 0 subject to the energy constraint restates the principle of least action. When
the external potential function is constant, the definition of ds as a path element
implies that the system trajectory is a geodesic in the Riemann space defined by
the mass tensor m i j . This anticipates the profound geometrization of dynamics
introduced by Einstein in the general theory of relativity.

2.5 Special relativity

The concept of a mass point remains valid, but a time interval dt can no longer
be treated as a nondynamical parameter. Einstein’s basic postulate [323, 393] is
that the interval ds between two space-time events is characterized by the invariant
expression

ds 2 = c2 dt 2 − d x 2 − dy 2 − dz 2 ,

where c is the constant speed of light. The space and time intervals can be measured
in any reference frame in uniform linear motion, referred to as an inertial frame.
Space and time intervals measured in two such inertial frames are related by a
linear Lorentz transformation, parametrized by the velocity and direction of relative
motion. It is convenient to introduce Minkowski space-time coordinates

d xµ = (d x1 , d x2 , d x3 , d x4 ) = (d x, dy, dz, ic dt),

such that ds 2 = −d xµ d xµ , using the summation convention of Einstein. A covariant

4-vector Aµ is defined by four quantities that are related by the appropriate Lorentz
transformation when measured in two different inertial frames.
Since dt cannot be singled out for special treatment, the covariant generalization
of Hamilton’s variational principle for a single particle requires an invariant action
integral


I = (xµ , u µ , τ ) dτ,

where c2 dτ 2 = ds 2 > 0 for a timelike interval defines an invariant proper time

interval dτ . u µ = d xµ /dτ defines a covariant velocity 4-vector. The variational
condition δ I = 0, for fixed end-points xµ0 and xµ1 separated by a timelike interval,
implies relativistic Lagrangian equations of motion

d ∂ ∂
− = 0.
dτ ∂u µ ∂ xµ
 2.5 Special relativity 21

2.5.1 Relativistic mechanics of a particle

As measured in some inertial reference frame, the instantaneous particle velocity
is v = (d x/dt, dy/dt, dz/dt). It is customary to define noninvariant parameters
β = v/c and γ = (1 − β 2 )− 2 . The definition of the invariant interval ds 2 implies
1

that dτ 2 = dt 2 (1 − β 2 ), where dt is the measured time interval. Hence dτ dt

= γ , as
defined above. The spatial components of the 4-velocity are u = γ v and the time
component is u 4 = iγ c. Hence u µ u µ = (v 2 − c2 )γ 2 = −c2 , a space-time invariant.
Unlike classical mechanics, the 4-acceleration is not independent of the 4-
velocity. Because u µ u µ is invariant, its τ -derivative must vanish, so that
du µ
uµ = 0.
dτ
This implies a relationship between classical velocity and acceleration,
d dγ
v· (γ v) − c2 = 0,
dt dt
which is an immediate consequence of the definition of γ .
The variational formalism makes it possible to postulate a relativistic Lagrangian
that is Lorentz invariant and reduces to Newtonian mechanics in the classical limit.
Introducing a parameter m, the proper mass of a particle, or mass as measured in its
own instantaneous rest frame, the Lagrangian for a free particle can be postulated
to have the invariant form  = 12 mu µ u µ = − 12 mc2 . The canonical momentum is
pµ = mu µ and the Lagrangian equation of motion is
d d
pµ = (mu µ ) = 0,
dτ dτ
which clearly reduces to Newton’s equation when β → 0.
As measured in an inertial reference frame, the spatial components of the canon-
ical momentum are p = mγ v and the time component is p4 = imγ c. This can be
related to classical quantities by defining the relative energy as E = mγ c2 so that
p4 = ci E. Thus pµ is an energy–momentum vector, for which pµ pµ = −m 2 c2 , a
constant for unaccelerated motion. In terms of classical quantities, this invariant
2
norm is expressed by p 2 − Ec2 = −m 2 c2 , or E 2 = m 2 c4 + p 2 c2 , which implies the
famous Einstein formula E = mc2 in the instantaneous rest frame of a particle.
In the limit β → 0, E = mc2 + 12 mv 2 + · · ·, verifying the classical formula for
kinetic energy.
The free-particle Lagrangian  is a space-time constant − 12 mc2 . If terms are
added that are invariant functions of xµ , the equations of motion become
d
pµ = X µ
dτ
22 2 Classical mechanics

defining a covariant Minkowski 4-force X µ = ∂∂ xµ

. The canonical momenta are
modified if these added terms depend explicitly on the 4-velocity u µ . It can no
longer be assumed that the rest mass is constant.
If m is not constant, the equations of motion are


d d xµ dm d xµ d 2 xµ
m = +m = X µ.
dτ dτ dτ dτ dτ 2
On multiplying by u µ = d xµ /dτ and summing, and using u µ u µ = −c2 , which
d
implies u µ dτ u µ = 0, this reduces to
d
(mc2 ) = −X µ u µ .
dτ
The rest mass remains constant if the 4-force and 4-velocity are orthogonal in the
Minkowski sense that X µ u µ = 0.
In relativistic theory, energy is a component of the 4-momentum, and is conserved
only under particular circumstances. Given dτ d
= γ dtd , the equation of motion for
p4 = i E/c is
d
γ (i E/c) = X 4 .
dt
Consider a 4-force X µ that does not change the rest mass m. As shown above,
X µ u µ = iγ cX 4 + γ X · v = 0, since u 4 = iγ c and u = γ v. Then X 4 = ci X · v for
such a force, and
d
γ E = X · v.
dt

2.5.2 Relativistic motion in an electromagnetic field

The classical electromagnetic force acting on a particle of charge q is the Lorentz
force (in Gaussian units)
 
1
Q = q E + (v × B) ,
c
where, in terms of scalar and vector potentials φ, A,
1 ∂A
B = ∇ × A, E = −∇φ − .
c ∂t
After expanding the triple vector product v × ∇ × A = v × B,
 
q ∂ d
Q= (A · v − φc) − A . (2.6)
c ∂x dt
 2.5 Special relativity 23

This is of the form

∂W d ∂W
Q= − ,
∂x dt ∂v
where W = qc (A · v − φc).
Given the electromagnetic 4-vector field Aµ = (A, iφ) and the 4-velocity u µ =
(γ v, iγ c), W is the classical limit of a relativistic invariant γ W = qc Aµ u µ . This
term augments the free-particle relativistic Lagrangian to give
1 q
 = mu µ u µ + Aµ u µ . (2.7)
2 c
The canonical momentum is pµ = mu µ + qc Aµ . If defined such that p4 = ci E,
the energy is E = mc2 γ + qφ, adding electrostatic energy to rest energy mc2 and
relativistic kinetic energy T = mc2 (γ − 1). Because mu µ mu µ = ( p − qc A)µ ( p −
q
c
A)µ = −m 2 c2 , energy and momentum are related by (p − qc A)2 − (mc2 +
T )2 /c2 = −m 2 c2 , or
 q 2
(mc2 + T )2 = p − A c2 + m 2 c4 .
c
Written as a formula for iteration, this is
 q 2 
T = p− A (2m + T /c2 ).
c
The Lagrangian equations of motion are
d  q  q ∂
mu µ + Aµ = (Aν u ν ).
dτ c c ∂ xµ
Using d
dτ
= γ dtd , these equations can be written as
 
d q ∂ d
γ (mu µ ) = X µ = (Aν u ν ) − γ Aµ . (2.8)
dt c ∂ xµ dt
Comparing the classical force Q, given above, the spatial components of X µ are
X = γ Q, verifying the historical fact that Maxwell’s theory is covariant under
Lorentz transformations.
 3
Applied mathematics

The principal references for this chapter are:

[77] Dennis, J.E. and Schnabel, R.B. (1983). Numerical Methods for Unconstrained
Optimization and Nonlinear Equations (Prentice-Hall, Englewood Cliffs,
New Jersey).
[125] Fletcher, R. (1987). Practical Methods of Optimization, 2nd edition (Wiley,
New York).
[168] Hestenes, M.R. (1966). Calculus of Variations and Optimal Control Theory
(Wiley, New York).
[332] Pulay, P. (1987). Analytical derivative methods in quantum chemistry, Adv. Chem.
Phys. 69, 241–286.
[439] Zerner, M.C. (1989). Analytic derivative methods and geometry optimization, in
Modern Quantum Chemistry, eds. A. Szabo and N.S. Ostlund (McGraw-Hill,
New York), pp. 437–458.

Before undertaking the major subject of variational principles in quantum mech-

anics, the present chapter is intended as a brief introduction to the extension of
variational theory to linear dynamical systems and to classical optimization meth-
ods. References given above and in the Bibliography will be of interest to the reader
who wishes to pursue this subject in fields outside the context of contemporary theor-
etical physics and chemistry. The specialized subject of optimization of molecular
geometries in theoretical chemistry is treated here in some detail.

3.1 Linear systems

Any multicomponent system whose dynamical behavior is governed by coupled
linear equations can be modelled by an effective Lagrangian, quadratic in the sys-
tem variables. Hamilton’s variational principle is postulated to determine the time
behavior of the system. A dynamical model of some system of interest is valid if it
satisfies the same system of coupled equations. This makes it possible, for example,

25
26 3 Applied mathematics

to construct an electrical circuit model of a mechanical system, or to reduce either

to a computer model with the appropriate choice of parameters.
For example [146], a system of interconnected electrical circuits and a mechanical
system of masses connected by springs satisfy the same linear equations if system
parameters are related by the following definitions:

Symbol Electrical circuit Masses on springs

I current displacement
M inductance reciprocal mass tensor
R resistance viscous force
C capacitance inverse spring constant
E external emf driving force

The Lagrangian for this linear system is

1 1 1 2 
L= M jk I˙ j I˙k − I + Ė j (t)I j .
2 j,k 2 j Cj j j

A dissipation function adds a nonintegrable term


δW = − R j I˙ j δ I j
j

to δL. The implied equations of motion are


M jk I¨k + R j I˙ j + I j /C j = Ė j (t).
k

3.2 Simplex interpolation

In many practical applications of nonlinear optimization, a linear approximation
is iterated until some vector of system parameters has negligible norm. At each
stage of such a process, previous steps provide a set of m ≤ n trial vectors in an
n-dimensional parameter space, each associated with an output vector of gradients
of the quantity to be optimized. The next iterative step is facilitated if a linear com-
bination of such vectors can be found that produces an output vector of minimum
norm.
Linear interpolation can be described geometrically in terms of an m-simplex in
the n-dimensional parameter space. An m-simplex in a hyperspace of dimension
n ≥ m is a set of m + 1 points that do not lie in a subspace of dimension less
than m. For example, a triangle is a 2-simplex and a tetrahedron is a 3-simplex.

The interior of a simplex is the set of points X = i xi λi , such that 0 ≤ λi ≤ 1 and

i λi = 1.
 3.2 Simplex interpolation 27

A typical optimization problem is to interpolate a vectorial function f(x) to


a given vector F by a least-square fit F  i fi λi , where fi = F(xi ). The varia-

tional problem is to minimize 12 | i fi λi − F|2 , subject to the linear constraint

i λi = 1, using a Lagrange multiplier µ. The variational functional is Iµ =
1 
2 i, j (fi λi − F) · (f j λ j − F) − ( i λi − 1)µ. The variational equations, for i =
1, . . . , m, are
 
fi · f j λ j − µ = fi · F; λ j = 1.
j j

The simplest possible example is linear interpolation in a 1-simplex (x0 , x1 ). The

system of linear equations is

f 0 f 0 λ0 + f 0 f 1 λ1 − µ = f 0 F
f 1 f 0 λ0 + f 1 f 1 λ1 − µ = f 1 F
λ0 + λ1 = 1.

The solution λ0 = ff11−−Ff0 ; λ1 = F− f 0

f1 − f0
; µ = 0 determines the coefficients in the in-
terpolation formula
x0 ( f 1 − F) + x1 (F − f 0 )
X= .
f1 − f0

3.2.1 Extremum in n dimensions

For an extremum in n dimensions, δW (q) = 0, where q is a vector of n gener-
alized coordinates. Generalized gradients, defined by pi = ∂∂qWi for i = 1, . . . , n,

must vanish at an extremum. A linear interpolating function P = i pi λi is to be

interpolated to P  0, at the corresponding point Q = i qi λi in the coordinate
hyperspace. The variational equations for the simplex method are
 
pi · p j λ j − µ = 0; λ j = 1.
j j

On solving these equations for the coefficients {λ j }, the solution of minimum norm
is the interpolated gradient vector P, such that µ = |P|2  0, at the interpolated
coordinate vector Q. The Lagrange multiplier µ in this method provides an estimate
of the residual error.
At each step of an iterative method for nonlinear optimization, the subsequent
coordinate step q must be estimated. The vector Q, interpolated in a selected
m-simplex of prior coordinate vectors, must be combined with an iterative esti-
mate of the component of q orthogonal to the hyperplane of the simplex.
28 3 Applied mathematics

A Hessian matrix
∂2W
Fi j = ,
∂qi ∂q j
is defined by small displacements about a given point, such that p = Fq. A
generalization of Newton’s method can be used to estimate the location of a zero
value of a vector from its given value and derivative. If an approximation to F −1 is
maintained and updated at each iterative step, then

q = q⊥ + qi λi ,
i

where

q⊥ = −F −1 p + (F −1 p · qi )[(q · q)]i−1
j qj.
i, j≤m

Alternatively, the interpolation and extrapolation steps can be combined, using the
formula [75]

q = Q − F −1 P = (qi − F −1 pi )λi .
i

3.3 Iterative update of the Hessian matrix

Quadratic expansion of a function W (q) about a local extremum takes the form
1 0
W  W0 + F qi q j .
2 i, j i, j

At a general point, displaced from the extremum, gradients pi = ∂∂qWi do not vanish.
Newton’s formula estimates a displacement toward the extremum q = −Gp0 ,
where G is the inverse of the Hessian matrix F. Since F is not constant in a nonlinear
problem, if it cannot be computed directly it must be deduced from an initial
estimate followed by iterative updates. In many circumstances, the gradients at a
general point can be computed directly or estimated with useful accuracy. Then
each successive coordinate increment q is associated with a gradient increment
p. If the Hessian matrix is known, Fq = p for sufficiently small increments.
Given an estimated F 0 , it must in general be updated to be consistent with the
computed gradient increment p. This implies a linear formula for the increment of
F, Fq = p − F 0 q. Since this provides only n equations for the n 2 elements
of F, the incremental matrix cannot be determined from data obtained in a single
iterative step.
The practical problem is to find a way to use the information obtained in each
iterative step without making unjustified changes of the Hessian. This could be
 3.3 Iterative update of the Hessian matrix 29

accomplished if the Hessian were projected onto two orthogonal subspaces, one of
which corresponds to a subset of m linearly independent increments q and the
corresponding p. A rank-m (Rm) update in the vector space spanned by these
coordinate increments is defined such that q† Fq = q† (p − F 0 q). This
condition is satisfied by

F(Rm) = p(q† p)−1 q† (p − F 0 q)(p† q)−1 p† .

Alternatively, the inverse matrix G can be updated directly,

G(Rm) = q(p† q)−1 p† (q − G 0 p)(q† p)−1 q† .

These formulas update the rank-m projection of F 0 or G 0 , using the nonhermitian

†
projection operator Pm = p(q† p)−1 q† , such that Pm q = q, Pm p =
p, and Pm Pm = Pm . This operator projects onto the m-dimensional vector space
spanned by the specified set of gradient vectors. The Rm update has the undesirable
property of altering the complementary projection of the updated matrix.

3.3.1 The BFGS algorithm

The BFGS (Broyden [42], Fletcher [124], Goldfarb [145], Shanno [379]) algor-
ithm is an update procedure for the Hessian matrix that is widely used in iterative
optimization [125]. The simpler Rm update takes the form

F = Fmm + F 0 − Pm F 0 Pm† ,

where Fmm q = p. Replacing this by a form that leaves the complementary
projection of F 0 unchanged,

F = Fmm + (I − Pm )F 0 (I − Pm† ),

the BFGS update of F 0 is

F(BFGS) = p(q† p)−1 q† (p + F 0 q) (p† q)−1 p†

− p(q† p)−1 q† F 0 − F 0 q(p† q)−1 p† .

Alternatively, the BFGS update of G 0 is

G(BFGS) = q(p† q)−1 p† (q + G 0 p) (q† p)−1 q†

− q(p† q)−1 p† G 0 − G 0 p(q† p)−1 q† .

For exact infinitesimal increments, the matrix

d p † dq = dq † d p = dq † F dq
30 3 Applied mathematics

is symmetric and nonsingular at an extremum of W . When m > 1, the matrices

p† q and q† p are not necessarily symmetric or even nonsingular when eval-
uated with approximate gradients away from an extremum. Any practical algorithm
must modify these matrices to remove singularities.

3.4 Geometry optimization for molecules

Quantum mechanical calculations of the electronic structure of molecules for fixed
nuclear coordinates involve lengthy calculations even using the sophisticated com-
putational methods that have evolved over half a century of computational quantum
chemistry. A principal output of such calculations is the variational energy as a func-
tion of the nuclear coordinates. Current methodology makes it possible to compute
the energy gradients or effective forces as well as the energy itself. For chemical
applications, equilibrium geometries and the coordinates of transition states must
be deduced from such data [357, 332, 439].
For such applications of classical optimization theory, the data on energy and
gradients are so computationally expensive that only the most efficient optimization
methods can be considered, no matter how elaborate. The number of quantum
chemical wave function calculations must absolutely be minimized for overall
efficiency. The computational cost of an update algorithm is always negligible
in this context. Data from successive iterative steps should be saved, then used to
reduce the total number of steps. Any algorithm dependent on line searches in the
parameter hyperspace should be avoided.
Molecular geometry can be specified either in Cartesian coordinates xa for each of
N nuclei, or in generalized internal coordinates qk , where k = 1, . . . , 3N − 6 for a
general polyatomic molecule. Neglecting kinetic energy of nuclear motion because
of the large nuclear masses, six generalized coordinates corresponding to translation
and rotation are subject to no internal forces, and can be fixed or removed from the
internal coordinates considered in geometry optimization. Cartesian coordinates are
subject to six constraint conditions, which may be imposed directly or indirectly
using Lagrange multipliers. This number reduces to five for a linear molecule, since
rotation about the molecular axis is not defined.
An equilibrium state is defined for generalized coordinates such that the total
energy E({q}) is minimized. The energy gradients p = ∂∂qE are forces with reversed
sign. Coordinate and gradient displacements from m successive iterations are saved
as m × n column matrices, qk and pk , respectively. The Hessian matrix is Fi j =
∂2 E
∂qi ∂q j
, such that Newton’s extrapolation formula is q = −Gp0 , where G = F−1 .
In quasi-Newton methods, a parametrized estimate of F or G is used initially,
then updated at each iterative step. Saved values of q, p can be used in standard
algorithms such as BFGS, described above.
 3.4 Geometry optimization for molecules 31

3.4.1 The GDIIS algorithm

This method [75] uses simplex interpolation combined with an update step such as
BFGS. The variational functional is
2 
1 m
k
m
Iµ = p ck − ck − 1 µ.
2 k=0 k=0

The variational equations determined by δ Iµ = 0 are

 
pk · p j c j − µ = 0; c j = 1.
j j

The interpolated gradient vector is p̄ = k pk ck and the interpolated coordinate

vector is q̄ = k qk ck . In the GDIIS algorithm, a currently updated estimate of the
inverse Hessian is used to estimate a coordinate step based on these interpolated
vectors. This gives
q̄ = qupd − q̄ = −(G 0 + G)p̄
or, equivalently, with G = G 0 + G,

qupd = q̄ − G p̄ = (qk − Gpk )ck .
k

In test calculations [75] this algorithm was found to produce rapid convergence.
When combined with single-step (m = 1) BFGS update of the inverse Hessian, this
is a very efficient algorithm.
The GDIIS algorithm can locate saddle points (transition states), because it
searches specifically for a point at which all gradients vanish, independently of
the sign of the second derivative. A reaction path can be followed by selecting the
eigenvector of G that belongs to the eigenvalue of greatest magnitude [332].

3.4.2 The BERNY algorithm

This algorithm, standard in the widely used GAUSSIAN program system, is a
rank-m update of the Hessian matrix, in an orthonormal basis [356]. A basis of unit
vectors is constructed in the m-dimensional vector space spanned by the increments
q. For k = 1, m, define

k−1
d = q −
k k
e j (e j · qk ).
j=1

Then orthonormal unit vectors are constructed such that

ek = dk (dk · dk )− 2 .
1
32 3 Applied mathematics

A significant advantage of this procedure is that nearly linearly dependent vectors

can be eliminated at this stage, simply reducing the value of m. The resulting unit
vectors define an n × m column matrix. Using these unit vectors, the algorithm
solves an m × m system of linear equations in the e-space,
(q† e)(e† F̄e) = (p† e).
The matrix (q† e) is lower-triangular by construction. For the index range j ≤ i ≤ m,
solution matrix elements are determined by
 

i−1
i† i −1
(e F̄e ) = (q e )
i† j
(p e ) −
i† j
(q e )(e F̄e ) .
i† k k† j

k=1

The Hessian matrix F 0 + F is updated and symmetrized, using


F = ei [(ei† F̄e j ) j≤i + (e j† F̄ei ) j>i − (ei† F 0 e j )]e j† .
i, j

Several aspects of the BERNY algorithm have a somewhat inconsistent math-

ematical basis. A modified algorithm may be more efficient. The Hessian matrix
can be constructed directly from gradient vectors but not from coordinate vectors.
The defining relation Fq = p implies that an Hermitian matrix F must have
the form pAp† . In principle, the unit vectors of the BERNY algorithm should
be in the p space. However, the orthonormalization process ensures numerical
stability by systematizing rejection of redundant coordinate increments q. One
might then propose an alternative algorithm, in which this basis is used to update
the inverse Hessian G. Because Gp = q, this matrix has the formal expansion
qBq† . The equations to be solved are
(p† e)(e† Ḡe) = (q† e).
In these equations, the matrix (p† e) is no longer triangular. However, the additional
computational effort may be unimportant if the number of iterative steps can be
reduced, thus saving energy and gradient evaluations. The matrix e† Ḡe is to be
symmetrized before updating G. The incremental matrix is
 1 1 j† i

G = ei
(e Ḡe ) + (e Ḡe ) − (e G e ) e j† .
i† j i† 0 j

i, j
2 2
 II
Bound states in quantum mechanics

This part introduces variational principles relevant to the quantum mech-

anics of bound stationary states. Chapter 4 covers well-known varia-
tional theory that underlies modern computational methodology for elec-
tronic states of atoms and molecules. Extension to condensed matter is
deferred until Part III, since continuum theory is part of the formal basis of
the multiple scattering theory that has been developed for applications in
this subfield. Chapter 5 develops the variational theory that underlies
independent-electron models, now widely used to transcend the practical
limitations of direct variational methods for large systems. This is ex-
tended in Chapter 6 to time-dependent variational theory in the context
of independent-electron models, including linear-response theory and its
relationship to excitation energies.
 4
Time-independent quantum mechanics

The principal references for this chapter are:

[74] Courant, R. and Hilbert, D. (1953). Methods of Mathematical Physics (Interscience,

New York).
[101] Epstein, S.T. (1974). The Variation Method in Quantum Chemistry (Academic
Press, New York).
[180] Hurley, A.C. (1976). Electron Correlation in Small Molecules (Academic Press,
New York).
[242] Merzbacher, E. (1961). Quantum Mechanics (Wiley, New York).
[365] Schrödinger, E. (1926). Quantisiering als Eigenwertproblem, Ann. der Physik 81,
109–139.
[394] Szabo, A. and Ostlund, N. S. (1982). Modern Quantum Chemistry; Introduction to
Advanced Electronic Structure Theory (McGraw-Hill, New York).

In 1926, Schrödinger [365] recognized that the variational theory of elliptical differ-
ential equations with fixed boundary conditions could produce a discrete eigenvalue
spectrum in agreement with the energy levels of Bohr’s model of the hydrogen atom.
This conceptually startling amalgam of classical ideas of particle and field turned
out to be correct. Within a few years, the new wave mechanics almost completely
replaced the ad hoc quantization of classical mechanics that characterized the “old”
quantum theory initiated by Bohr. Although the matrix mechanics of Heisenberg
was soon shown to be logically equivalent, the variational wave theory became the
standard basis of methodology in the physics of electrons.
The nonrelativistic Schrödinger theory is readily extended to systems of N in-
teracting electrons. The variational theory of finite N-electron systems (atoms and
molecules) is presented here. In this context, several important theorems that follow
from the variational formalism are also derived.
Hartree atomic units will be used here. In these units, the unit of action h, the
mass m of the electron, and the magnitude e of the electronic charge −e are all set
equal to unity. The velocity of light c is 1/α, where α is the fine-structure constant.

35
36 4 Time-independent quantum mechanics

The unit of length is a0 , the first Bohr radius of atomic hydrogen. The Hartree unit
of energy is e2 /a0 , approximately 27.212 electron volts.

4.1 Variational theory of the Schrödinger equation

The Schrödinger equation for one electron is

{t̂ + v(r) − }ψ(r) = 0,

where t̂ = − 12 ∇ 2 , the kinetic energy operator of Schrödinger. For historical reasons,

this is written as

{H − }ψ = 0,

where H = t̂ + v(r) is the Hamiltonian operator of the theory. It is assumed that

physically meaningful potential functions v(r) vanish for large r → ∞. Solutions
ψ are required to be bounded in R3 , 3-dimensional Euclidean space, and to be
continuous with continuous gradients except at Coulomb singularities of the po-
tential function. These conditions define the Hilbert space of trial functions for the
variational theory. Eigenfunctions of the one-electron Hamiltonian that lie in this
Hilbert space will be called orbital wave functions here.
Because H is Hermitian, the energy eigenvalues  are real numbers. For any phys-
ically realizable potential function v(r), there is a lowest eigenvalue 0 . For bound
states, discrete eigenvalues  < 0 are determined by the condition that the wave
function ψ must vanish as r → ∞. Continuum states, with  ≥ 0, are bounded,
oscillatory functions at large r , but must be regular at the coordinate
 3 origin. The
∗
orbital Hilbert space is characterized by a scalar product (i| j) = d rψi (r)ψ j (r).

4.1.1 Sturm–Liouville theory

It is generally true that the normalized eigenfunctions of an Hermitian operator such
as the Schrödinger H constitute a complete orthonormal set in the relevant Hilbert
space. A completeness theorem is required in principle for each particular choice of
v(r) and of boundary conditions. To exemplify such a proof, it is helpful to review
classical Sturm–Liouville theory [74] as applied to a homogeneous differential
equation of the form

L[ f (x)] + λρ(x) f (x) = 0,

where L[ f ] = ( p f  ) − q f , for continuous p, p  , q with ρ, p > 0 in the interval

x0 ≤ x ≤ x1 . The Hilbert space of solutions f (x) is defined by real-valued con-
tinuous functions with continuous first derivatives over the interval x0 ≤ x ≤ x1 ,
 4.1 Variational theory of the Schrödinger equation 37

subject to homogeneous boundary conditions (specified logarithmic derivatives) at

the end-points x0 , x1 .
A version of Green’s theorem follows from partial integration of the symmetric
integral



( p f 1 f 2 + q f 1 f 2 ) d x = − f 2 L[ f 1 ] d x + p f 1 f 2 |xx10 ,
 

which implies


( f 1 (x)L[ f 2 (x)] − f 2 (x)L[ f 1 (x)]) d x = p( f 1 f 2 − f 1 f 2 )|xx10 .

For homogeneous boundary conditions, the logarithmic derivatives f 1 / f 1 and f 2 / f 2

are equal at both end-points x0 , x1 . Hence the integrated term vanishes, and the
differential expression L[ f ] is self-adjoint with these boundary conditions. The
1
weighting function ρ can be eliminated by converting to u = ρ 2 f . Then [u] =
(Pu  ) − Qu, where


p q − 12 d d −1
P= ; Q = −ρ p ρ 2 ,
ρ ρ dx dx
and

[u(x)] + λu(x) = 0.

A Green function is defined as a solution of the inhomogeneous equation (using

a Dirac delta-function)

[G(x, ξ )] = −δ(x, ξ ),

subject to the specified boundary conditions. Then

x1
u(x) = λ G(x, ξ )u(ξ ) dξ
x0

is a solution of the differential equation with these boundary conditions. Because

[u] is self-adjoint, u satisfies an homogeneous integral equation with a symmet-
rical kernel [74].
Starting from an assumed minimum eigenvalue λ0 , for state u 0 (x), a nondecreas-
ing sequence can be built up one by one. A global constant can be added so that
λ0 > 0, and then removed when all eigenvalues are determined.
 This construction
uses the theorem that the maximum value of (u|G|u) = u(s)G(s, t)u(t) dt ds,
subject to (u|u) = 1, is given by the eigenfunction u 0 corresponding to the mini-
mum eigenvalue λ0 . This follows from δ[(u|G|u) − µ((u|u) − 1)] = 0, for which
38 4 Time-independent quantum mechanics

the Euler equation is the homogeneous integral equation

G(s, t)u(t) dt = µu(s),

where the Lagrange multiplier µ = λ−1 . The maximized value (u|G|u) is µ(u|u) =
λ−1
0 , selecting the lowest eigenvalue. Given u 0 and λ0 , consider the modified Green
function G 1 (s, t) = G(s, t) − u 0 (s)λ−1
0 u 0 (t). If (u|G 1 |u) is maximized, subject to
the orthogonality constraint (u|u 0 ) = 0 and to normalization (u|u) = 0, the same
procedure obtains u 1 and the eigenvalue λ1 . Equivalently, u n+1 maximizes (u|G|u)
subject to orthogonality to all eigenfunctions u i with i ≤ n. Thus an entire countable
sequence of eigenvalues and orthonormal eigenfunctions can be constructed.
If the full set of eigenfunctions is complete, any function χ in the Hilbert space
that is orthogonal to all eigenfunctions must vanish identically. To prove this, sup-
pose that some function χ exists such that (u i |χ ) = 0, i ≤ n, but (χ|χ ) = 1 and
λ(χ|G|χ ) = 1 for finite positive λ. The construction given above develops the
expansion

n
G(s, t) = lim u i (s)u i (t)/λi,
n→∞
i=0

which can be proven to converge uniformly for a positive-definite sequence

of eigenvalues (Mercer’s theorem) [74]. Because the summation in (χ|G|χ) =
n
limn→∞ i=0 (χ |u i )(u i |χ)/λi vanishes term by term, due to the orthogonality
condition, the sum must converge to zero. Since λ is assumed to be finite, and
(χ |χ ) = λ(χ |G|χ ), this contradicts the hypothesis that χ (x) is normalizable to
unity, implying that χ (x) must vanish identically for x0 ≤ x ≤ x1 if it satisfies the
defining properties for functions u(x) in the Hilbert space. This proves completeness
of the orthonormal set of eigenfunctions in the designated Hilbert space [74].

4.1.2 Idiosyncracies of the Schrödinger equation

In time-dependent quantum mechanics, wave functions contain an inherently com-
plex time-factor. There is a significant notational advantage in treating wave func-
tions as complex fields even in bound-state applications of the theory. In particular,
complex spherical harmonics are assumed in the vector-coupling formalism re-
quired for systematic treatment of orbital and spin angular momentum. The Bloch
waves appropriate to regular periodic solids are also inherently complex fields. The
extension from real fields is straightforward, using a notation for matrix elements
such that



(i| j) = d 3 r ψi∗ (r)ψ j (r); (i|H| j) = d 3 r ψi∗ (r)Hψ j (r).
 4.1 Variational theory of the Schrödinger equation 39

This implies ( j|i)∗ = (i| j). Orthonormality is expressed by (i| j) = δi j . If H is

Hermitian (self-adjoint), (i|H| j) − ( j|H|i)∗ = ( j − i∗ )(i| j) = 0. Since (i|i) = 1,
i must be real. If j = i, either  j = i or (i| j) = 0.
For a single electron moving in a fixed potential function v(r) the coordinate range
is infinite in R3 and positive energy eigenvalues lie in a continuum. This spectrum
can be discretized by placing the system in a polyhedral or spherical box. The first
alternative is appropriate to solid-state physics. Assuming a regular periodic lattice
composed of space-filling polyhedral atomic cells, periodic boundary conditions
are appropriate on each cell boundary. This situation will be discussed in detail in
the chapter on multiple scattering theory. A spherical boundary is appropriate to an
isolated atom. However, the relevant physics is not that of bound states but rather
the theory of electron scattering, which will also be discussed in a separate chapter.
For the variational theory of bound states, box normalization makes it possible to
consider completeness in terms of an artificially discrete eigenvalue spectrum, but is
not necessarily the most direct way to achieve practical completeness in variational
calculations. A widely used practical alternative is to retain the typical bound-state
boundary condition that functions vanish for r → ∞, but to use discrete basis sets
constructed from spherical harmonics, powers of r , and exponential or Gaussian
exponential factors in radial wave functions. In general, practical completeness can
be achieved within some effective convergence radius that depends on the choice of
decay parameters for such exponential functions. If the boundary conditions allow
a true continuum, then the subset of bound states cannot be complete. The bound-
state sum must be augmented by an integral over a parameter associated with the
continuum.
The usual Hilbert-space requirement of continuous gradients is not appropriate to
Coulombic point-singularities of the potential function v(r) [196]. This is illustrated
by the cusp behavior of hydrogenic bound-state wave functions, for which the
Hamiltonian operator is

H = t̂ + v(r) = t̂ − Z /r.

It is convenient to use spherical polar coordinates (r, θ, φ) for any spherically

symmetric potential function v(r ). The surface spherical harmonics Ym satisfy
Sturm–Liouville equations in the angular coordinates and are eigenfunctions of the
orbital angular momentum operator ˆ such that

ˆ · Y
ˆ m = ( + 1)Ym
ˆz Ym = mYm .
40 4 Time-independent quantum mechanics

The kinetic energy operator takes the form


1 ∂ 2 ∂ ˆ2
t̂ = − 2 r + 2.
2r ∂r ∂r 2r
For a spatial wave function of the form

ψ(r, θ, φ) = r −1 u(r )Ym (θ, φ),

the radial wave equation is

 
d2 ( + 1)
− + + v(r ) −  u(r ) = 0.
2dr 2 2r 2
∞
A radial wave function is normalized so that (u|u) = 0 u ∗ (r )u(r )dr = 1.
A solution regular at the origin must have the form

u(r ) = r +1 (u 0 + r u 1 + · · ·).

The leading terms in the radial equation for r −1 u are

 
−2 1
r [−( + 1) + ( + 1)]u 0 +
2
 
−1 1
r [−( + 2)( + 1) + ( + 1)]u 1 − Z u 0 + · · · .
2
The first term vanishes due to the leading factor in u(r ). To avoid a singularity, the
second term must vanish for  = 0. This implies that
Z
u1 = − u0,
+1
which for  = 0 is the Coulomb cusp condition. An external potential that is regular
at the origin cannot affect this r −1 term even if  > 0. Hence both even and odd
powers of r must be present in the wave function, which implies a discontinuous
gradient for  = 0. The orbital Hilbert space must be defined so that the  = 0
angular components of wave functions satisfy this cusp condition when expanded
about Coulomb singularities in spherical polar coordinates.

4.1.3 Variational principles for the Schrödinger equation

Schrödinger [365] defined the kinetic energy functional as the positive-definite
form T = 12 τ d 3 r ∇ψ ∗ (r) · ∇ψ(r), valid in a finite volume τ enclosed by a sur-
∗
 σ .∗ When integrated by parts, defining t̂ = − 2 ∇ , T = τ d r ψ (r)t̂ψ(r) +
1 2 3
face
1
2 σ
ψ (σ)∇ψ(σ) · dσ. With the usual boundary conditions for a large volume,
 4.1 Variational theory of the Schrödinger equation 41

the surface term vanishes, leaving the volume term as defined in subsequent liter-
ature, T = (ψ|t̂|ψ). The surface term becomes important in scattering theory and
in solid-state theory using local atomic cells.
Defining the functional V = (ψ|v|ψ), where v(r) is a local potential for which
E = T + V is bounded below, the Schrödinger variational principle requires E
to be stationary subject to normalization (ψ|ψ) = 1. The variation δ E induced
δE
by δψ is expressed using the functional derivative δψ ∗ = Hψ, defined such that
 3 ∗ δE
δ E = δ(ψ|H|ψ) = d r{δψ δψ ∗ + cc}. The variational condition is


  
∗ δE
δ(E − [(ψ|ψ) − 1]) = d r δψ 3
− ψ + cc = 0,
δψ ∗
for unconstrained δψ in the orbital Hilbert space, which implies the Schrödinger
equation
δE
= Hψ = {t̂ + v}ψ = ψ.
δψ ∗
Use of the Rayleigh quotient [ψ] = (ψ|H|ψ)/(ψ|ψ) as a variational func-
tional is an alternative to using a Lagrange multiplier to ensure normalization.
Assuming that the eigenfunctions {ψi } of H are complete and orthonormal in
 
the relevant Hilbert space, (ψ|ψ) = i |(i|ψ)|2 and (ψ|H|ψ) = i i |(i|ψ)|2 ≥
0 (ψ|ψ). This implies that [ψ] ≥ 0 , a variational principle for the ground
state. For other eigenstates, {H − i }ψi = 0, so that [φi ] = (i|i)/(i|i) = i .
Let ψ = ψi + δψ. Then δ(ψ|ψ) = (δψ|i) + (i|δψ), and δ(ψ|H|ψ) = (δψ|H|i) +
(i|H|δψ) = i δ(ψ|ψ). It then follows that

δ(ψ|H|ψ)(ψ|ψ) − (ψ|H|ψ)δ(ψ|ψ)
δ|ψ→ψi =
(ψ|ψ)2 ψ→ψi
δ(ψ|H|ψ) − i δ(ψ|ψ)
= = 0.
(ψi |ψi )
By construction, this stationary principle holds for any eigenstate.

4.1.4 Basis set expansions

For expansion in a given set of basis functions that lie in the relevant Hilbert
 
space, such that ψi  a ηa xai , (ψ|H|ψ) = a,b xa∗ (a|H|b)xb is to be made sta-
tionary with respect to variations for which the normalization integral (ψ|ψ) =
 ∗
a,b x a (a|b)x b is constant. For variations of the expansion coefficients,

δxa∗ {(a|H|b) − (a|b)}xb = 0,
a,b
42 4 Time-independent quantum mechanics

which implies

{(a|H|b) − (a|b)}xb = 0.
b

In terms of the matrices (a|H|b) and (a|b), the set of coefficients xb is a null vector
of the matrix (a|H|b) − (a|b) when  is determined so that the determinant of this
matrix vanishes. If the basis functions are orthonormalized, this condition becomes
det(h ab − δab ) = 0 and the coefficient vector is an eigenvector of the Hermitian
matrix h ab .
Suppose that h (n−1) is diagonalized in a basis of dimension n − 1, and this basis
is extended by adding an orthonormalized function ηn . The diagonalized matrix
is augmented by a final row and column, with elements h ni , h in respectively, for
i < n. The added diagonal element is h nn . Modified eigenvalues are determined
by the condition that the bordered determinant of the augmented matrix h (n) − 
should vanish. This is expressed by
 
 (n)   |h ni |2
det h −  = i<n (i − ) (h nn − ) − = 0.
 −
i<n i

If the matrix elements h ni do not vanish, the final factor may vanish at values of 
different from the original eigenvalues, giving the equation
 |h ni |2
h nn −  = .
 −
i<n i

The function on the right here has poles at each of the original eigenvalues. It has
positive slope everywhere, because
d |h ni |2 |h ni |2
=
d i −  (i − )2
for each term in the sum. This function increases from −∞ to ∞ between each
adjacent pair of unequal original eigenvalues, and must cross the straight line h nn − 
once in each such interval.
Hence the new set of eigenvalues interleaves with the the original set. One new
eigenvalue lies below all of the originals, one lies above, and the rest each occur
in one of the intermediate intervals. This construction shows that the mth old eigen-
value is an upper bound for the mth new eigenvalue, for m < n. It follows that each
of the n eigenvalues of a variational matrix eigenvalue problem is an upper bound
for the corresponding eigenvalue of the Hamiltonian operator. If the eigenvalues
are indexed in nondecreasing sequence, and h nn falls between two adjacent nonde-
generate eigenvalues k−1 and k , the new eigenvalues i(n) for i = k are displaced
away from h nn . Thus i(n) ≤ i (i < k), and i+1
(n)
≥ i (i ≥ k). This displacement
 4.2 Hellmann–Feynman and virial theorems 43

rule is evident from the 2 × 2 submatrix for ηn and any ψi . The secular equation is
(h nn − i(n) )(i − i(n) ) = |h ni |2 , which requires both factors to have the same sign.

4.2 Hellmann–Feynman and virial theorems

4.2.1 Generalized Hellmann–Feynman theorem
The derivation here follows Hurley [179]. Given H (ξ ) for some real parameter
ξ , and a variational approximation such that δ(!|H |!) = 0 and δ(!|!) = 0, the
following theorem can be proved: if variations δ! include all variations induced
by δξ , then
 
∂E ∂H
= .
∂ξ ∂ξ
The notation · · · here denotes a mean value, (!| · · · |!)/(!|!). To prove the
theorem, let each trial function depend on ξ , and require δ! = δξ ∂! ∂ξ
to lie in
the Hilbert space of variational trial functions. For variational wave functions and
energy values, (δ!|H − E|!) + (!|H − E|δ!) = 0. For variations driven by δξ ,
δ(!|H − E|!) = (δ!|H − E|!) + (!|H − E|δ!)


∂ H ∂ E
+ δξ ! − ! = 0
∂ξ ∂ξ
implies

  
∂E ∂ H ∂ H
= ! ! (!|!) = ,
∂ξ ∂ξ ∂ξ
which proves the theorem.
If the Hellmann–Feynman theorem is to be valid for forces on nuclei, the
Coulomb cusp condition must be satisfied. However, if the nuclei are displaced,
the orbital Hilbert space is modified. Hurley [179] noted this condition for
finite basis sets, and introduced the idea of “floating” basis functions, with cusps
that can shift away from the nuclei, in order to validate the theorem for such
forces.

4.2.2 The hypervirial theorem

As an extension of Noether’s theorem to quantum mechanics, the hypervirial theo-
rem [101] derives conservation laws from invariant transformations of the theory.
Consider a unitary transformation of the Schrödinger equation, U (H − E)! =
U (H − E)U † U ! = 0, and assume the variational Hilbert space closed under a
unitary transformation U (ξ ) = eiξ Ĝ , where Ĝ is an Hermitian operator. Thus i Ĝ!
44 4 Time-independent quantum mechanics

lies in the Hilbert space of variational trial functions. In the limit of an infinitesimal
transformation, ξ → 0, (i Ĝ!|!) + (!|i Ĝ!) = 0, and

(i Ĝ!|H − E|!) + (!|H − E|i Ĝ!) = 0

⇒ (!|[H, Ĝ]|!) = 0,

where [H, Ĝ] = H Ĝ − Ĝ H . Since normalization is preserved, this implies the

hypervirial theorem:

[H, Ĝ] = 0.

If U (ξ ) = eiG(ξ ) such that G(0) = 0, then

 
∂G
H, = 0.
∂ξ ξ →0
As an example, suppose for an N-electron system that energy E is approximated
by an orbital functional E[{φi , n i }], which depends on one-electron orbital wave
functions φi and on occupation numbers n i through a variational N-electron trial
 !. A momentum displacement is generated by U = exp( i h π · D̂),
1
wave function
where D̂ = i ri . In the momentum representation of the orbital wave functions,
U !(. . . , φi (pi ), . . .) = !(. . . , φi (pi + π), . . .). If the Hamiltonian is invariant
under such a displacement, then hi [H, D̂] = P̂ and the hypervirial theorem implies
conservation of total linear momentum P̂ = 0.
A coordinate displacement is generated by U = exp( hi ξ · P̂) such that
U !(. . . , φi (ri ), . . .) = !(. . . , φi (ri + ξ ), . . .). In this case, hi [H, P̂] =

− i ∂∂rHi = F̂, the total force acting on the electrons. The hypervirial theo-
rem implies that the mean value of this force vanishes for an isolated system,
F̂ = 0.

4.2.3 The virial theorem

An N-electron virial operator is defined by
1 
V= (ri · pi + pi · ri ) = ri · pi − 3ih(N /2).
2 i i

− V)e h ξ V !|ξ →0 = 0, this operator generates a scale transformation

i
d
Using ( hi dξ
e h ξ V !(. . . , ri , . . .) = τ 3N /2 !(. . . , τ ri , . . .), where τ = eξ . The commutator in the
i

hypervirial theorem is
i 
∂ H ∂H

[H, V] = · pi − ri · .
h i
∂pi ∂ri
 4.3 The N-electron problem 45

If the Hilbert space of the variational basis set is invariant under scale transforma-
tion, then the hypervirial theorem implies
 

∂ H ∂H
· pi − ri · = 0. (4.1)
i
∂pi ∂ri

In the nonrelativistic Hamiltonian, the kinetic energy T is homogeneous of de-

gree 2 in the electronic momenta pi , and the Coulombic energy, including nuclear
repulsions, is homogeneous of degree −1 in the electronic coordinates ri and the nu-
clear coordinates Ra . From Euler’s theorem on partial derivatives of homogeneous
functions,
 ∂H
· pi = 2T,
i
∂pi
 ∂H  ∂H (4.2)
· ri = −V − · Ra .
i
∂ri a ∂Ra

∂H
The force acting on nucleus a is Fa = − ∂Ra
. From Eqs. (4.1) and (4.2), the general
virial theorem for a molecule is

2T  + V  − Fa  · Ra = 0.
a
For a diatomic molecule with internuclear distance R,
∂E
2T  + V  + R = 0.
∂R

4.3 The N-electron problem

4.3.1 The N-electron Hamiltonian
The N-electron wave function ! is an antisymmetric function of N sets of spatial
and spin coordinates ri , si for individual electrons, all evaluated at a common time t.
In postulating a time-dependent Schrödinger equation of the form
∂
H ! = ih !,
∂t
the nonrelativistic N-electron theory is inherently inconsistent with the underlying
relativistic field theory of electrons. Relativistic covariance in principle requires
assigning a separate time variable ti to each electron. The time-independent non-
relativistic theory of stationary states has a deeper justification, since assuming
instantaneous interactions is not inconsistent with time-averaged interactions. The
theory will be developed here primarily for stationary states. In this context, the
N-electron Schrödinger equation is an eigenvalue equation in 4N spatial and spin
46 4 Time-independent quantum mechanics

coordinates,
(H − E)! = 0,
where H = T̂ + Û + V̂ . The individual operators, for kinetic energy, interelec-
tronic Coulomb interaction, and the external potential, respectively, are
1 2
T̂ = − ∇i ,
2 i
1 1 1
Û = = u(ri j ),
2 i, j ri j 2 i, j

V̂ = v(ri ).
i

Spin indices and summations will be suppressed in the notation here, but are to
be understood in connection with coordinate indices and integrations over coor-
dinates. Thus d 3 r · · · is intended to imply an implicit summation over a spin
index. The energy
 3 functional 3 for an antisymmetric wave function !(1, . . . , N ) is
∗
(!|H
 3 |!) = d r
 13 1 ∗ N· · · d r N ! H ! and the normalization integral is (!|!) =
1 d r1 · · · N d r N ! !.
The Schrödinger variational principle requires (!|H |!) to be stationary subject
to constant normalization (!|!). Introducing a Lagrange multiplier, the variational
condition is
δ(!|H − E|!) = (δ!|H − E|!) + (!|H − E|δ!) = 0,
for variations in the N-electron Hilbert space. The Hermitian character of H implies
that (!|H − E|δ!) = (δ!|H − E|!)∗ . Since δ! is arbitrary, it can be multiplied
by i. Then (!|H − E|δ!) = −(δ!|H − E|!)∗ . Together, these equations imply
that (!|H − E|δ!) = 0 if and only if (δ!|H − E|!) = 0. Hence if δ! is unre-
stricted in the Hilbert space, the variational condition implies the time-independent
Schrödinger equation (H − E)! = 0.

4.3.2 Expansion in a basis of orbital wave functions

The N-electron Hamiltonian is a sum of one- and two-electron operators, Ĥ =
N 1 N
i h(i) + 2 i, j u(i, j). A reference state " is defined as a single Slater deter-
minant constructed as a normalized antisymmetrized product of N orthonormal
orbital wave functions of the one-electron coordinates, " = det{φ1 (1) · · · φ N (N )},
implicitly including a normalization factor (N !)− 2 . A complete set of such Slater
1

determinants is determined by virtual excitations from this reference state, defined

by replacing one or more of the N occupied orbital functions of " by functions
 4.3 The N-electron problem 47

from the residual unoccupied set. The general form is "µ = det{φ1µ (1) · · · φ Nµ (N )}.
If the orbital functions are indexed by i, j, . . . ≤ N for the occupied set, and by
N < a, b, . . . for the residual unoccupied set, a consistent notation for virtual ex-
citations is given by

"ia = det{1, . . . , a, . . . , N }, "iabj = det{1, . . . , a, . . . , b, . . . , N }, . . . .

Taking spin indices into account, all two-electron integrals are of the form
(i j|ū|kl) = (i j|u|kl) − (i j|u|lk), with the convention that orbitals with different
spin indices are orthogonal. It is convenient to truncate summations by the use of
occupation numbers n i , which are in principle determined by Fermi–Dirac statis-
tics. At zero temperature, occupation numbers are determined by the structure of the
reference state. Then n i = 1, n a = 0 for i ≤ N < a. A convention used for double
summation indices is i j : i < j ≤ N , ab : N < a < b.
With these conventions of notation, simple formulas exist for all matrix elements
of H in the basis of Slater determinants generated by virtual excitations from
a reference state [72]. Denoting the latter by "0 , and defining an effective one-

electron operator H = h + j n j (i j|ū|i j) for "0 , the sequence of nondiagonal
elements is
a 
|H |0 = (a|H|i),
i
ab   
ij |H |0 = n i n j (1 − n a )(1 − n b )(ab|ū|i j),
ij ab
abc 
i jk |H |0 = 0, etc.

All such matrix elements vanish if more than two occupied orbitals are replaced,
because H contains only a two-electron interaction operator. Given (0|H |0) =
 1
i n i (i|h|i) + 2 i, j n i n j (i j|ū|i j), diagonal elements follow a simple rule:

a 
i |H |ia − (0|H |0) = (a|H|a) − (i|H|i) − (ai|ū|ai),
ab 
i j |H |i j − (0|H |0) = (a|H|a) + (b|H|b) − (i|H|i) − ( j|H| j)
ab

+ (ab|ū|ab) + (i j|ū|i j) − (ai|ū|ai) − (bi|ū|bi)

− (a j|ū|a j) − (bj|ū|bj).

For (µ|H |µ) − (0|H |0) defined by a virtual excitation i, j, . . . → a, b, . . . , the

general rule is to add all elements (a|H|a), subtract all elements (i|H|i), add all
elements of the form (ab|ū|ab) or (i j|ū|i j), and subtract all elements of the form
(ai|ū|ai).
48 4 Time-independent quantum mechanics

4.3.3 The interelectronic Coulomb cusp condition

The electronic Coulomb interaction u(r12 ) = r112 greatly complicates the task of
formulating and carrying out accurate computations of N-electron wave functions
and their physical properties. Variational methods using fixed basis functions can
only with great difficulty include functions expressed in relative coordinates. Unless
such functions are present in a variational basis, there is an irreconcilable conflict
with Coulomb cusp conditions at the singular points r12 → 0 [23, 196]. No finite
sum of product functions or Slater determinants can satisfy these conditions. Thus
no practical restricted Hilbert space of variational trial functions has the correct
structure of the true N-electron Hilbert space. The consequence is that the full
effect of electronic interaction cannot be represented in simplified calculations.
The interelectronic Coulomb cusp can be analyzed by transforming a two-
electron Hamiltonian to relative coordinates. The one-electron potential function
is regular at the singularity r12 → 0 and does not affect the cusp behavior. Given
coordinates r1 and r2 , mean and relative coordinates are defined, respectively, by
1
r = (r1 + r2 ),
2
q = r2 − r1 .

The Hamiltonian is
1 1 1
H = − ∇12 − ∇22 + + O(1)
2 2 r12
1 1
= − ∇r2 − ∇q2 + + O(1).
4 q
Given relative angular momentum , the singular part of the Schrödinger equation is

∂2 2 ∂ ( + 1) 1
Hq = − − + + + O(1).
∂q 2 q ∂q q2 q

A power-series solution is f  (q) = q  ( f 0 + q f 1 + · · ·), where f 0 = 0 for  ≥ 0.

When substituted into the differential equation, this gives

Hq f  = q −2 {[−( + 1) + ( + 1)] f 0 }

+ q −1 {[−( + 1)( + 2) + ( + 1)] f 1 + f 0 } + O(q  ).

To avoid a singularity that cannot be cancelled by any one-electron potential, the co-
efficient of q −1 must vanish when  = 0. This implies the Coulomb cusp condition
f 0 (q) = f 00 (1 + 12 q + · · ·). A similar expansion is valid for any  > 0. Because the
 4.4 Symmetry-adapted functions 49

expansion includes odd powers of q, it describes a cusp in three dimensions, but

implies a discontinuous gradient only for  = 0.

4.4 Symmetry-adapted functions

For systems with high symmetry, in particular for atoms, symmetry properties can
be used to reduce the matrix of the N-electron Hamiltonian to separate noninter-
acting blocks characterized by global symmetry quantum numbers. A particular
method will be outlined here [263], to complete the discussion of basis-set ex-
pansions. A symmetry-adapted function is defined by # = O", where O is an
Hermitian projection operator (O2 = O) that characterizes a particular irreducible
representation of the symmetry group of the electronic Hamiltonian. Thus H com-
mutes with O. This implies the “turnover rule” (O"|H |O") = ("|H |O"), which
removes the projection operator from one side of the matrix element. Since the ex-
pansion of O" may run to many individual terms, this can greatly simplify formulas
and computing algorithms. Matrix elements (#µ |H |#ν ) simplify to ("µ |H |#ν ) or
(#µ |H |"ν ).
For a general symmetry group, symmetry-adapted orbital functions can be in-
dexed by νλµ, where ν designates a Hilbert space of dimension dνλ . Each of
the orthonormal basis orbitals belongs to an irreducible representation λ of the
group, with subspecies index µ. Group elements are represented by matrices
indexed by λµ in this basis, which is closed under group operations. A con-
figuration in the N-electron Hilbert space is characterized by a set of subshell
 
occupation numbers n νλ ≤ dνλ such that ν λ n νλ = N . Closed subshells are
invariant and only open subshells (n νλ < dνλ ) are relevant to the subsequent analy-
sis. A configuration defines a Hilbert space whose basis is the set of Slater de-
terminants defined by all independent choices of subspecies indices µ for the
indicated open subshells. This set of basis determinants {"i , i = 1, n} is closed
under group operations and defines a group representation of n × n transformation
matrices.
The representation of the N-electron symmetry group generated by a given
configuration is block-diagonalized into irreducible representations indexed by
, with subspecies index M. In general there will be some number m ≤ n
of independent functions with the same symmetry indices. If the configuration
basis is {"i , i = 1, n}, then a set of m ≤ n unnormalized projected determinants
n
# j = O" j = −1 i=1 y ji "i are to be constructed, where O is a projector onto
j 1 n
the symmetry species , M. The function  j # j is normalized if  j = i=1
2
|y ji |2
and y j j =1 1. The projected determinants can be transformed into an orthonormal set
− n n
!µ = kµ 2 i=1 xµi "i , where kµ = i=1 |xµi |2 . Alternatively, as an expansion in
50 4 Time-independent quantum mechanics
1 
the basis of projected determinants, !µ = kµ2 mj=1 aµj1# j . Using the turnover rule,
m 2 ∗ − 12
the implied overlap integrals are (!µ |!ν ) = i=1 (kµ aµi )(kν xνi ) = δµν , where
m ∗
i=1 aµi x νi = δµν .
In the basis of these orthonormal symmetry-adapted functions, matrix elements
of the invariant Hamiltonian are given for two different configurations A and B by

1

mA 
nB
(!µA |H |!ν B ) = (kµ /kν ) 2 xνi (" j A |H |"i B )
A∗ B
aµj
j=1 i=1

1

nA 
mB
= (kν /kµ ) 2 aν j ("i A |H |" j B ).
A∗ B
xµi (4.3)
i=1 j=1

As an option, the normalization constants can be eliminated using

(!µA |H |!ν B ) =
  ! 12
A  B A 
B
xνi (" j A |H |"i B )
A∗ B
aµj aν j ("i A |H |" j B )
A∗ B
xµi . (4.4)
j i i j

4.4.1 Algorithm for constructing symmetry-adapted functions

In order to construct orthonormal projected determinants, the symmetry conditions
should be expressed as a set of homogeneous linear equations. For example, the
projector O for a finite group is a weighted sum of the representation matrices.
This implies homogeneous matrix equations {O − I }# = 0. For a rotation group,
with step-up and step-down operators J± , functions obtained for M = J by solving
J+ # = 0 determine those with M < J by using the step-down operator J− . The
algorithm consists of the following steps [263]:

1. Reduce the equations by Gaussian elimination from right to left.

2. Use back-substitution to determine the solution whose leading coefficient corresponds
to the rightmost undetermined pivot element. Set this coefficient to unity.
3. Append this solution vector to the set of equations and remove elements to the right of
the new pivot element by Gauss elimination.
4. Repeat to obtain a trapezoidal array of m orthogonal null vectors {xµi }, each with leading
coefficient unity.
5. The last line is the projection of the first basis function.
6. Invert the leading triangle xµi to get aµj .
 4.4 Symmetry-adapted functions 51

Table 4.1. a/x table for functions 2S (abc)

Equations
i =2 3 4
1 −1 1

Coefficients kµ

− 12 1∗ 1 2
∗
1 1
2
− 12 3
2

4.4.2 Example of the method

Consider the total spin eigenstates constructed from the three-electron configuration
abc, consisting of all eight Slater determinants obtained by assigning spin α :
m s = 12 or spin β : m s = − 12 to each of three orthonormal spatial orbital functions
φa (r), φb (r), φc (r). Because total Sz is quantized, the full configuration basis can
be separated into subsets indexed by M S . Thus for M S = 32 there is only one basis
function "1 = det(aα bα cα ). The algorithm determines a single quartet (S = 32 ) spin
eigenfunction, for which !1 = #1 = "1 with coefficient vectors a11 = x11 = 1,
and normalization constant k1 = 1.
For M S = 12 , there are three basis functions,

"2 = det(aα bα cβ ),
"3 = det(aα cα bβ ),
"4 = det(bα cα aβ ).

Coefficients ("1 |S+ |"i ), i = 2, 3, 4 in the linear equation S+ # = 0 are shown in

Table (4.1), together with the vectors xµi generated using the present algorithm.
Nondiagonal elements of the upper left-hand triangle of xµi , all of which vanish by
construction, are replaced in such a/x tables by elements of the triangular matrix
aµi . Diagonal elements common to both matrices a and x are marked by an asterisk.
If the one-electron and Coulomb matrix elements, which are the same for all
basis functions of the configuration, are denoted by E 0 , matrix elements of the
Hamiltonian in the M S = 12 subconfiguration are

H22 = E 0 − (ab|u|ba),
H23 = (bc|u|cb),
52 4 Time-independent quantum mechanics

H33 = E 0 − (ac|u|ca),
H24 = −(ac|u|ca),
H34 = (ab|u|ba),
H44 = E 0 − (bc|u|cb).
Matrix elements for the two orthonormal projected doublet (S = 12 ) functions,
computed using Eq. (4.4), are
1 1
(!2 |H |!2 ) = H22 + H23 − H24
2 2
1 1
= E 0 − (ab|u|ba) + (ac|u|ca) + (bc|u|cb),
2 2


1 1 1
(!3 |H |!2 ) = − H22 + H23 − H24
2 2 2

  12
1 1
+ H32 + H33 − H34 [(H23 + H24 )]
2 2
  12
3
= [−(ac|u|ca) + (bc|u|cb)][−(ac|u|ca) + (bc|u|cb)]
4

 12
3
=− [(ac|u|ca) − (bc|u|cb)],
4
1
(!3 |H |!3 ) = − (H23 + H24 ) + (H33 + H34 )
2
1 1
= E 0 + (ab|u|ba) − (ac|u|ca) − (bc|u|cb).
2 2
 5
Independent-electron models

The principal references for this chapter are:

[26] Blanchard, P. and Brüning, E. (1992). Variational Methods in Mathematical

Physics: A Unified Approach (Springer-Verlag, Berlin).
[90] Dreizler, R.M. and Gross, E.K.U. (1990). Density Functional Theory
(Springer-Verlag, Berlin).
[102] Eschrig, H. (1996). The Fundamentals of Density Functional Theory (Teubner,
Stuttgart).
[130] Froese Fischer, C. (1977). The Hartree-Fock Method for Atoms (Wiley, New York).
[185] Janak, J.F. (1978). Proof that ∂ E/∂n i = i in density-functional theory, Phys. Rev.
B18, 7165–7168.
[189] Jones, R.O. and Gunnarsson, O. (1989). The density functional formalism, its
applications and prospects, Rev. Mod. Phys. 61, 689–746.
[211] Landau, L.D. (1956). The theory of a Fermi liquid, Zh. Eksp. Teor. Fiz. 30,
1058–1064. [Sov. Phys. JETP 3, 920–925 (1956)]
[212] Landau, L.D. (1957). Oscillations in a Fermi liquid, Zh. Eksp. Teor. Fiz. 32, 59–66.
[Sov. Phys. JETP 5, 101–108 (1957)]
[239] McWeeny, R. (1989). Methods of Molecular Quantum Mechanics (Academic Press,
New York).
[321] Parr, R.G. and Yang, W. (1989). Density-Functional Theory of Atoms and
Molecules (Oxford University Press, New York).
[386] Slater, J.C. (1972). Statistical exchange-correlation in the self-consistent field, Adv.
Quantum Chem. 6, 1–92.
[388] Slater, J.C. and Wood, J.H. (1971). Statistical exchange and the total energy of a
crystal, Int. J. Quantum Chem. Suppl. 4, 3–34.

Despite the simple and universal structure of the nonrelativistic Hamiltonian for
N interacting electrons, it produces a broad spectrum of physical and chemical
phenomena that are difficult to conceptualize within the full N-electron theory.
Starting with the work of Hartree [162] in the early years of quantum mechanics,
it was found to be very rewarding to develop a model of electrons that interact
only indirectly with each other, through a self-consistent mean field. A deeper
motivation lies in the fact that the relativistic quantum field theory of electrons is

53
54 5 Independent-electron models

explicitly described by a field operator that corresponds more closely to a one-

particle model wave function than to that of the Schrödinger N-electron theory.
The fundamental characterization of this electron field by Fermi–Dirac statistics is
directly applicable to the mean-field theory, using concepts of statistical occupation
numbers determined by effective one-electron orbital energy values. The variational
theory appropriate to such independent-electron models is developed in this chapter.
Hartree’s original idea of the self-consistent field involved only the direct
Coulomb interaction between electrons. This is not inconsistent with variational
theory [163], but requires an essential modification in order to correspond to the
true physics of electrons. In neglecting electronic exchange, the pure Coulombic
Hartree mean field inherently allowed an electron to interact with itself, one of
the most unsatisfactory aspects of pre-quantum theories. Hartree simply removed
the self-interaction by fiat, at the cost of making the mean field different for each
electron. Orbital orthogonality, necessary to the concept of independent electrons,
could only be imposed by an artificial variational constraint. The need for an ad hoc
self-interaction correction (SIC) persists in recent theories based on approximate
local exchange potentials.
In a theory using antisymmetric wave functions [127, 79] the interaction of an
electron with itself disappears due to cancellation between classical Coulombic and
exchange interactions. The Hartree–Fock theory, including exact exchange, became
an essential methodology in atomic and molecular physics. Computational proce-
dures and results for atoms have been reviewed by Hartree [163] and by Froese
Fischer [130]. The electronic correlation energy, neglected in the Hartree–Fock
approximation, is important for an exact description of electronic phenomena in
atoms, molecules, and solids, but is difficult to treat in practical computational the-
ory. Before discussing Hartree–Fock theory and other mean-field approximations
here, a formally exact independent-electron theory will be developed. This orbital
functional theory (OFT) embodies the correlation theory of Brueckner [43], origi-
nally a rationalization of the nuclear shell model, which becomes an exact theory
when applied to the explicit two-electron Coulomb interaction.

5.1 N-electron formalism using a reference state

Although a Slater-determinant reference state " cannot describe such electronic
“correlation” effects as the wave-function modification required by the interelec-
tronic Coulomb singularity, a variationally based choice of an optimal reference
state can greatly simplify the N-electron formalism. " defines an orthonormal
set of N occupied orbital functions φi with occupation numbers n i = 1. While
("|") = 1 by construction, for any full N-electron wave function ! that is to
be modelled by " it is convenient to adjust (!|!) ≥ 1 to the unsymmetrical
 5.1 N-electron formalism using a reference state 55

normalization ("|!) = 1. This implies that ("|! − ") = 0, so that ! is parti-

tioned into complementary orthogonal components " and ! − ". Because E! =
H ! for any energy eigenstate, E = ("|H |!) = E 0 + E c , where E 0 = ("|H |")
is the mean energy of the reference state, while E c = ("|H |! − ") provides a
natural definition of correlation energy. Thus any rule ! → " that determines a
reference state also defines the correlation energy.
Concerned with reconciling the strong internucleon interaction, described em-
pirically by a hard-core interaction, with the success of the “independent-particle”
nuclear shell model, Brueckner [43] proposed that for any given !, with arbi-
trary normalization, an optimal reference state could be defined by maximizing
the projection ("|!). Since H ! = E! for an energy eigenstate, the Brueckner
variational condition (δ"|!) = 0 implies that (δ"|H |!) = 0. Following an argu-
ment used to derive Hartree–Fock theory, variations that just mix occupied orbitals
do not affect the antisymmetrized function ". Meaningful infinitesimal variations
δ" can be expressed as a sum of single virtual excitations "ia . The corresponding
variational condition in Brueckner theory can be expressed in terms of an effec-
tive one-electron “Hamiltonian” G defined such that ("ia |H |!) = (a|G|i) = 0 for
i ≤ N < a. This heuristic derivation from Brueckner’s variational condition [275]
does not fully determine the operator G, and explicitly assumes that ! is given in
advance. In any model theory that expresses both E 0 and E c as orbital function-
als, depending only on occupied orbitals of a reference state " and on occupation
numbers n i , the linear operator G is determined by the orbital functional derivative
δE
n i δφi∗
= Gφi . Introducing Lagrange multipliers for orthonormality, the variational

condition δ{E − i, j [(φi |φ j ) − δi j ]λ ji } = 0 implies the orbital Euler–Lagrange
(OEL) equations of orbital functional theory,

{G − i }φi = 0,

for the occupied orbitals of the reference state in a diagonalized representation

of the matrix λ ji . An iterative “self-consistent” procedure is required in general,
because both E 0 and E c depend nonlinearly on the occupied orbital set. When G is
Hermitian, Gφi = φi i and φi∗ G = i φi∗ . Defining the Dirac density matrix ρ̂, whose

kernel is i φi (r)n i φ ∗ (r ), this implies [G, ρ̂] = 0 as an alternative statement of
the OEL equations.

5.1.1 Fractional occupation numbers

For applications to open-shell states of atoms and molecules or to metallic solids,
and for systems at finite temperature, it is convenient to treat occupation numbers as
parameters that can vary freely in the range 0 ≤ n i ≤ 1. If the occupation numbers
56 5 Independent-electron models

{n i } in orbital functional theory are treated as continuous variables, the extended

energy functional E[{φ, n i }] provides a natural interpolation between quantum
states characterized by an explicit reference state ". Formal theory based on this
interpolation will be considered here, setting aside the somewhat controversial
issue of physical realization of such interpolations. Landau [211, 212] introduced
this concept as appropriate to a model of fermionic quasiparticles analogous to the
dressed electrons of renormalized quantum electrodynamics. The theory is charac-
terized by empirical expressions for total energy and entropy that are parametric
functions of orbital occupation numbers. The orbital functional theory developed
below, especially in its time-dependent extension, is a particular realization of Lan-
∂E
dau’s concept. Energy relationships are determined by Janak’s theorem, i = ∂n i
,
which will be proven below for general OFT models.
Orbital functional theory defines energy functionals for fractional occupation
numbers, providing a smooth interpolation between the values 0, 1 appropriate to
specific wave functions. Following Slater and Wood [388], changes such as addition
or removal of an electron can be described by integrating fractional occupation
numbers between the appropriate limits, using derivatives ∂ E/∂n i derived from
solution of OEL equations with intermediate fractional occupation numbers. An
excitation energy can be estimated by a single calculation of the “transition state”
defined with occupation numbers averaged over the initial and final states [386].

5.1.2 Janak’s theorem

When the variational energy is a functional of the reference state ", it can always

be expressed in terms of integrals over the Dirac matrix i φi (r)n i φi∗ (r ), which
remains well-defined for nonintegral occupation numbers. This implies two chain
rules, for n i = 0:
δE δE
∗ = Gφi = ,
n i δφi δ(n i φi∗ )



∂E ∗ δE
= d r φi
3
= d 3 r φi∗ Gφi .
∂n i δ(n i φi∗ )
Hence if n i is allowed to vary,



∗
δE = δn i d r φi Gφi +
3
n i d 3 r{δφi∗ Gφi + cc},
i i
  
where φi∗ Gφi = i and δφi∗ Gφi = i δφi∗ φi = 0 for normalized occupied or-
bital solutions of the OEL equations. This implies that a theorem proved by Janak
[185] is generally valid in OFT. For variations of occupation numbers in which
 5.2 Orbital functional theory 57

occupied orbitals are “relaxed” to satisfy the OEL equations, Janak’s theorem takes
the form
 ∂E 
dE = dn i = i dn i .
i
∂n i i

This theorem justifies the procedure of Slater and Wood for any version of orbital
functional theory.

5.2 Orbital functional theory

5.2.1 Explicit components of the energy functional
E 0 = ("|H |") is a sum of explicit orbital functionals,

T = ("|T̂ |") = n i (i|t̂|i),
i
1
U = ("|Û |") = n i n j (i j|ū|i j),
2 i, j

V = ("|V̂ |") = n i (i|v|i),
i

where tˆ = − 12 ∇ 2 and ū = r112 (1 − P12 ) incorporates an operator P12 that exchanges

space-spin coordinates of two electrons. This defines Coulomb and exchange func-
tionals such that ("|Û |") = E h + E x , where
1
Eh = n i n j (i j|u|i j),
2 i, j
1
Ex = − n i n j (i j|u| ji).
2 i, j

The corresponding orbital functional derivatives are defined by

δT
= tˆφi ,
n i δφi∗
δV
= v(r)φi ,
n i δφi∗
δ Eh 
= v h (r)φi = n j ( j|u| j)φi ,
n i δφi∗ j

δ Ex 
= v̂ x φi = − n j ( j|u|i)φ j .
n i δφi∗ j
58 5 Independent-electron models

The correlation energy E c , defined in general by H  − ("|H |"), takes the

particular form E c = ("|H |Q!), using the unsymmetrical normalization ("|!) =
("|") = 1 and defining the projection operator Q = I − ""† , determined by ".
In any practical theoretical model, E c must be approximated, and the functional
derivative δ E c /n i δφi∗ = v̂c φi is to be evaluated from this approximate expression.

5.2.2 Orbital Euler–Lagrange equations

The reference state " is a Slater determinant constructed as a normalized antisym-
metrized product of N orthonormal spin-indexed orbital functions φi (r). The orbital
energy functional E = E 0 + E c is to be made stationary, subject to the orbital or-
thonormality constraint (i| j) = δi j , imposed by introducing a matrix of Lagrange
multipliers λ ji . The variational condition is


 !

δ E − ni n j φi∗ φ j d 3 r − δi j λ ji =
ij

! 
 δE 
ni δφi∗ − n j φ j λ ji d r + cc
3

i
n i δφi∗ j


δE 3
+ δn i φi∗ d r = 0. (5.1)
i
n i δφi∗

With fixed n i = 0, for variations of occupied orbitals that are unconstrained in

the orbital Hilbert space, the variational condition implies orbital Euler–Lagrange
equations

δE 
= Gφi = n j φ j λ ji . (5.2)
n i δφi∗ j

The orbital functional derivative here defines an effective Hamiltonian

1
G = − ∇ 2 + v(r) + vh (r) + v̂x + v̂c .
2
The theory is usually expressed in terms of canonical equations

{G − i }φi = 0,

obtained by diagonalizing the matrix of Lagrange multipliers for the occupied

orbitals.
 5.2 Orbital functional theory 59

5.2.3 Exact correlation energy

It was shown by Brenig [35] that (δ"|!) = 0 removes all particle–hole vir-
tual excitations "ia from the wave function !. This follows because (δ"|!) =

i,a δci ("i |!) implies ("i |!) = 0 for i ≤ N < a. Since there are only two-
a∗ a a

particle interactions in the N-electron Hamiltonian, an exact formula for the

correlation energy can be expressed with only two-electron matrix elements:
  
E c = ("|H |! − ") = 0|H |iabj ciabj = (i j|ū|ab)ciabj .
ij ab ij ab

By construction |ciabj | = |("iabj |!)| ≤ ("|!) = 1. ciabj can be considered to define

a two-electron operator with antisymmetric matrix elements (ab|c̄|i j), such that
 
Ec = ni n j (1 − n a )(1 − n b )(i j|ū|ab)(ab|c̄|i j).
ij ab

The coefficients ciabj can be obtained only by constructing the wave function !. An
exact formal expression can be derived by partitioning the N-electron Hilbert space
using the projection operator P = ""† and its orthogonal complement Q = I −
""† . Using the definition given above, the correlation energy is E c = ("|H |Q!).
A more explicit expression is obtained from the partitioned Schrödinger equation
P(H − E)P! + P(H − E)Q! = 0,
Q(H − E)P! + Q(H − E)Q! = 0.
Here QH Q defines a reduced Hamiltonian operator Ĥ . On substituting a formal
solution of the second set of equations into the first, it follows for η → 0+ that
E = ("|H − H ( Ĥ − E − iη)−1 H |"),
where ( Ĥ − E − iη)−1 denotes an inverse operator in the Q-space. This expresses
the correlation energy as an exact implicit functional of " and hence of its occupied
orbital functions [290],
E c = −("|H ( Ĥ − E 0 − E c − iη)−1 H |"). (5.3)
The coefficients ciabj are given by an expression that depends only on the reference
state and on a parametric value of E c ,
 
ciabj = "iabj |Q!
 
= − "iabj |( Ĥ − E 0 − E c − iη)−1 H |" .
Since H is specified, Eq. (5.3) defines E c as a functional of the occupied orbitals
of ", although it cannot be expressed as an explicit closed formula. The orbital
functional derivative nδi δφ
Ec
∗ = v̂c φi defines a nonlocal correlation potential v̂c in
i
60 5 Independent-electron models

exact OEL equations. Iterative solution for E c combines a search for a root of
the partitioned N-electron secular determinant with the Brueckner condition that
defines ". This would be a legitimate variational method if the inverse of matrix
Q(H − E)Q could be evaluated exactly. In practice this is not possible, but standard
approximations of many-body theory are available for particular applications of the
formalism.
An infinitesimal unitary transformation of the orbital basis that modifies " must
mix occupied and unoccupied orbital functions. For a typical orbital variation,
δ" = "ia δcia and δφi = φa δcia . Unitarity induces δφa = −φi δcia∗ = −φi (δφi |φa ).
In terms of functional derivatives δ/δ̃φ defined for independent variations of occu-
pied and unoccupied orbitals,


δ Ec 
δ Ec = n i δφi
i n i δ̃φi∗


 δ Ec
+ (1 − n a ) δφa + cc. (5.4)
a (1 − n a )δ̃φa

For such variations,



δ Ec
δ Ec = ni δφi
i n δ̃φ ∗
i i


  δ Ec
− ni (1 − n a ) φi (δφi |φa ) + cc. (5.5)
i a (1 − n a )δ̃φa

Treating the coefficients ciabj as constants in the current cycle of an iteration loop,
and using Eq. (5.5), (a|δ E c /n i δφi∗ ) defines the matrix element
1 
(a|v̂c |i) = nj (1 − n c )(1 − n b )(a j|ū|cb)(cb|c̄|i j)
2 j b,c

1 
− nk n j (1 − n b )(k j|ū|ib)(ab|c̄|k j), (5.6)
2 j,k b

in agreement with matrix elements deduced from the double virtual excitation terms
in Q! [275]. The effective correlation “potential” is a linear operator v̂c with the
kernel
1 
vc (r, r ) = nj (1 − n c )(1 − n b )( j|ū|b)φc (r)φc∗ (r )(b|c̄| j)
2 j b,c

1 
− nk n j (1 − n b )(b|c̄| j)φk (r)φk∗ (r )( j|ū|b). (5.7)
2 j,k b
 5.3 Hartree–Fock theory 61

This operator is consistent with the leading terms in quasiparticle self-energies

implied by many-body theory [275, 407].

5.3 Hartree–Fock theory

5.3.1 Closed shells – unrestricted Hartree–Fock (UHF)
The theory is based on an optimized reference state " that is a Slater deter-
minant constructed as a normalized antisymmetrized product of N orthonormal
spin-indexed orbital functions φi (r). This is the simplest form of the more general
orbital functional theory (OFT) for an N-electron system. The energy functional
E = ("|H |") is required to be stationary, subject to the orbital orthonormality con-
straint (i| j) = δi j , imposed by introducing a matrix of Lagrange multipliers λ ji .
The general OEL equations derived above reduce to the UHF equations if correla-
tion energy E c and the implied correlation potential v̂c are omitted. The effective
Hamiltonian operator is
1
H = − ∇ 2 + v(r) + vh (r) + v̂x .
2
The theory is usually expressed in terms of canonical Hartree–Fock equations
{H − i }φi = 0,
obtained by diagonalizing the matrix of Lagrange multipliers for the occupied
orbitals.
Janak’s theorem, valid for general OEL equations when occupation numbers are
varied, holds for the UHF theory in the form


∂E
= φi∗ Hφi d 3 r = i ,
∂n i
for infinitesimal changes of occupation numbers for which the occupied orbitals
are relaxed as H changes. If orbital relaxation is neglected, the change of energy
due to removing an occupied orbital φi is given by the general rule for diagonal
elements cited in Section 4.3.2:
(i |H |i ) − (0|H |0) = −(i|H|i) = −i ,
which is a statement of Koopmans’ theorem [206] for ionization potentials. The
corresponding expression for electron affinities,
(a |H |a ) − (0|H |0) = (a|H|a) = a ,
must be used with caution, since typical variational calculations do not correctly
represent the continuum of electronic states outside a neutral atom or molecule.
62 5 Independent-electron models

This subject will be treated in a later chapter on continuum states and scattering
theory.

5.3.2 Brillouin’s theorem

Because of antisymmetry, variations of " that are simply linear transformations of
occupied orbitals have no effect other than a change of normalization. For orbital
functions with fixed normalization, a general variation of " takes the form δ" =
 
a (1 − n a )"i δci . The variational condition is
a a
i ni
  "   #
δ("|H |") = ni (1 − n a ) δcia∗ ia |H |0 + cc
i a
  " #
= ni (1 − n a ) δcia∗ (a|H|i) + cc = 0.
i a

This is a statement of Brillouin’s theorem [37], that (a|H|i) = 0, i ≤ N < a is a

necessary condition for ("|H |") to be stationary. The normalization of occupied
variables must also be varied in order to determine the Lagrange multipliers i .
Definition of the effective Hamiltonian H requires diagonal matrix elements deter-
mined by δ E/n i δφi∗ for unconstrained variations δφi .

5.3.3 Open-shell Hartree–Fock theory (RHF)

The UHF formalism becomes inconvenient for open-shell configurations of atoms
or molecules with point-group symmetry. Unless specific restrictions are imposed,
the self-consistent occupied orbitals fall into sets that are nearly but not quite
transformable into each other by operations of the symmetry group. By imposing
“equivalence” and “symmetry” restrictions, these sets become symmetry-adapted
basis states for irreducible representations of the symmetry group. This makes
it possible to construct symmetry-adapted N-electron functions, as described in
Section 4.4. The constraints in general invalidate the theorems of Brillouin and
Koopmans. This “restricted” theory (RHF) is described in detail for atoms by
Hartree [163] and by Froese Fischer [130].
To illustrate the modifications of UHF formalism, it is convenient to consider pure
spin symmetry for a single Slater determinant with Nc doubly occupied spatial or-
bitals χic and No singly occupied orbitals χio . The corresponding UHF state has Nα :
β
m s = 12 occupied spin orbitals φiα and Nβ : m s = − 12 occupied spin orbitals φi .
The number of open-shell and closed-shell orbitals are, respectively No = Nα −
Nβ > 0 and Nc = Nβ . Occupation numbers for the spatial orbitals are n c = 2,
n o = 1. If all orbital functions are normalized, a canonical form of the RHF refer-
ence state is defined by orthogonalizing the closed- and open-shell sets separately.
 5.3 Hartree–Fock theory 63

Nonzero Lagrange multipliers are required for orthogonalization of orbitals χ o to

χ c . The energy functional ("|H |") is
!
 1 1 1c,c 
o,c 
o,o
n i (i|h|i) + n i n j (i j|u|i j) − + + n i n j (i j|u| ji),
i
2 i, j 2 2 i, j i, j i, j

expressed in terms of the spatial orbitals. Alternatively,

c  o
1c,c
("|H |") = 2(i|h|i) + (i|h|i) + [4(i j|u|i j) − 2(i j|u| ji)]
i i
2 i, j

o,c
1o,o
+ [2(i j|u|i j) − (i j|u| ji)] + [(i j|u|i j) − (i j|u| ji)],
i, j
2 i, j

on substituting explicit spatial-orbital occupation numbers. This implies different

effective Hamiltonians for closed and open shells, respectively, defined by func-
tional derivatives
δ("|H |")  c
 
= hχ c
+ 2( j|u| j)χic − ( j|u|i)χ cj
n i δχi
c c∗ i
j

 o  
1
+ ( j|u| j)χic − ( j|u|i)χ oj = Hc χic ,
j
2

δ("|H |") 
c
 
= hχ o
+ 2( j|u| j)χio − ( j|u|i)χ cj
n i δχi
o o∗ i
j


o
 
+ ( j|u| j)χio − ( j|u|i)χ oj = Ho χio .
j
For comparison, the UHF effective Hamiltonians, indexed by spin α and β, are
δ("|H |") 
α
   β
α
α α∗ = hφ + ( j|u| j)φiα − ( j|u|i)φ αj + ( j|u| j)φiα = Hα φiα ,
n i δφi i
j j

δ("|H |") β

α
β

β
 β β β
β β∗
= hφi + ( j|u| j)φi + ( j|u| j)φi − ( j|u|i)φ j = Hβ φi .
n i δφi j j

In a basis of RHF orbitals, Hc = 12 (Hα + Hβ ) and Ho = Hα . The constraint con-

ditions (χic |χ cj ) = δi j , (χio |χ cj ) = 0, (χio |χ oj ) = δi j are incorporated into the RHF
Euler–Lagrange equations for the spatial orbital functions
 c   o
 o  o  c
H − ic χic = χ oj λoc
ji ; H − i
o
χi = χ cj λco
ji ,
j j

using nonvanishing Lagrange multipliers λoc

ji = (χ oj |Hc |χic ) and λco
ji = (χ cj |Ho |χio ).
64 5 Independent-electron models

For the 1s 2 2s 2 S ground state of atomic Li, using the common notation Jn χx =
(n|u|n)χx , K n χx = (n|u|x)χn , the RHF Hamiltonians are
1
Hc = h + 2J1s + J2s − K 1s − K 2s ,
2
Ho = h + 2J1s + J2s − K 1s − K 2s ,

and the UHF Hamiltonians are

Hα = h + 2J1s + J2s − K 1s − K 2s ,
Hβ = h + 2J1s + J2s − K 1s .

5.3.4 Algebraic Hartree–Fock: finite basis expansions

The OEL equations derived here as integrodifferential equations can be converted to
linear algebraic equations by expanding the orbital functions in a basis set {η p } that
can be extended to completeness in the orbital Hilbert space [347, 262, 349]. This

expansion takes the general form φi = p η p upi , so that orbital variations δφi =

p η p δupi are specified by variations
 of∗ the coefficients upi . If the basis set were
 ∗ δφi {H − i }φ3i d r = 0 would be equivalent
3
complete, the variational condition
 ∗
to algebraic equations, p,q δupi η p {H − i }ηq d r uqi = 0. Free variation of the
coefficients in a finite or countable basis implies matrix eigenvalue equations

{( p|H|q) − i ( p|q)}uqi = 0.
q

This is the stationary condition for the energy functional, when orbital functions
are expanded in the specified basis set. RHF equations are related to UHF equations
exactly as they are in the theory based on integrodifferential OEL equations.

5.3.5 Multiconfiguration SCF (MCSCF)

In some cases of strong correlation, or in general for open-shell states, an N-electron
state may not be dominated by a single reference state. In these circumstances, an
N-electron basis of Slater determinants "µ constructed from an occupied set of more
than N orbital functions defines a variational basis for the N-electron problem. A
configuration-interaction (CI) calculation in the N-electron basis "µ is combined
with variational optimization of these occupied orbital functions. Following the
general formalism developed here, a particular reference state " = "0 is singled out

as “first among equals”, giving the expansion ! = " + µ=0 "µ cµ with the usual
unsymmetrical normalization. The coefficients cµ are determined by diagonalizing
the N-electron matrix Hµν , or by constructing symmetry-adapted functions for
 5.4 The optimized effective potential (OEP) 65

open-shell configurations. Several of these coefficients may have unit magnitude.


The energy functional for orbital variations is E = ("|H |") + µ=0 ("|H |"µ )cµ .
For this limited CI problem, it is convenient to diagonalize the one-electron
density matrix, defining a set of natural orbitals that vary at each step of an iterative
procedure. The one-electron density operator is defined by



 1
ρ(r, r ) = · · · d 3 r2 · · · d 3 r N i=2
N
δ(ri , ri )
(!|!)
× !(r1 , . . . , r N )! ∗ (r1 , . . . , rN )|r=r1 ;r =r1 ,
such that

ρ(r, r ) = φi (r) di j φ ∗j (r ).
i

Natural orbitals {φi } and occupation numbers n i are determined by diagonalizing

the density matrix di j . The reference state " is constructed from N orbitals with
the largest occupation numbers. The energy functional is of the form
 1
E= n i (i|h|i) + πi jkl (i j|ū|kl).
i
2 i jkl

Orbital Euler–Lagrange equations are determined by functional derivatives

δE  πi jkl
∗ = hφi (r) + ( j|ū|l)φk (r) = Hi φi (r),
n i δφi (r) jkl
ni

giving in general a different effective Hamiltonian for each orbital. The Euler–
Lagrange equations are to solved subject to the orthonormality constraints
(i| j) = δi j . Introducing Lagrange multipliers λ ji = ( j|Hi |i) = λi j = (i|H j | j),
these equations take the general form

Hi φi = φ j λ ji .
j

Orbital construction and diagonalization of the CI matrix are alternated until the
calculations converge.

5.4 The optimized effective potential (OEP)

The computational effort of solving orbital Euler–Lagrange (OEL) equations is
significantly reduced if the generally nonlocal exchange-correlation potential v̂xc
can be replaced or approximated by a local potential vxc (r). A variationally defined
optimal local potential is determined using the optimized effective potential (OEP)
method [380, 398]. This method can be applied to any theory in which the model
66 5 Independent-electron models

exchange-correlation energy E xc is an explicit orbital functional, so that the func-

tional derivative δ E xc /n i δφi∗ = v̂xc φi is specified for occupied orbital functions φi
of a reference state ". Variations of the orbitals φi induce


δE = n i d 3 r{δφi∗ (r)Gφi (r) + cc},
i

expressed in terms of an effective Hamiltonian

1
G = − ∇ 2 + v(r) + vh (r) + v̂xc
2
defined for the OEL equations by the functional derivative
δE
= Gφi .
n i δφi∗
A modified effective Hamiltonian GOEP is defined by replacing v̂xc by a model
local potential vxc (r). The energy functional is made stationary with respect to
variations of occupied orbitals φi that are determined by modified OEL equa-
tions in which G is replaced by GOEP . δφi is determined by variations δvxc (r)
in these modified OEL equations. To maintain orthonormality, δφi can be con-
strained to be orthogonal to all occupied orbitals of the OEP trial state ", so

that δφi (r) = a (1 − n a )φa (r)(a|δφi ). First-order perturbation theory for the OEP
Euler–Lagrange equations implies that

{GOEP − i }δφi = −{δvxc − (i|δvxc |i)}φi ,

or

(a − i )(a|δφi ) = −(a|δvxc |i),

for i ≤ N < a. Expressed in terms of a Green’s function, this implies

δφi (r) = − gi (r, r )δvxc (r )φi (r ) d 3 r ,

where

gi (r, r ) = (1 − n a )φa (r)(a − i )−1 φa∗ (r ).
a

This defines δφi (r)/δvxc (r ) = −gi (r, r )φi (r ) as the kernel of a linear operator.
The variational equation induced by δvxc → δφi is
 
ni (1 − n a )(i|δvxc |a)(a − i )−1 (a|G|i) = 0.
i a
 5.4 The optimized effective potential (OEP) 67

If δvxc is arbitrary, this implies that

 
n i φi∗ (r) (1 − n a )φa (r)(a − i )−1 (a|G|i) = 0.
i a

On substituting (a|GOEP |i) = 0 from the OEP orbital equation, and using (a|GOEP −
G|i) = (a|vxc − v̂xc |i), this gives the OEP integral equation
 
n i φi∗ (r) (1 − n a )φa (r)(a − i )−1 (a|vxc − v̂xc |i) = 0.
i a

5.4.1 Variational formulation of OEP

The OEP integral equation is implied if the nonnegative integral

1 
I = ni (1 − n a )(i|vxc (r) − v̂xc |a)(a − i )−1 (a|vxc (r) − v̂xc |i),
2 i a

evaluated for solutions of the OEP orbital equations, is stationary for variations
of v . If this local potential function is expanded in a basis set πq , vxc (r) =
 xc
q πq (r)Lq , the coefficient vector Lq is determined by linear algebraic equations
[69, 151]
 
(i| p|a)(a − i )−1 (a|q|i)Lq = (i| p|a)(a − i )−1 (a|v̂xc |i),
q i,a i,a

where (a|q|i) = φa∗ (r)πq (r)φi (r)d 3 r. The variational formalism [151, 181] pro-
vides a practical methodology for molecular wave functions by introducing a basis-
set expansion of the optimized local potential. Because only off-diagonal elements
(a|vxc |i) occur in the OEP equations, the local effective potential is determined
only up to an additive constant. The physical boundary condition that potentials
for an isolated system must vanish at infinite distance fixes a specific value of this
constant.
The variational energy E OEP = ("OEP |H |"OEP ) is minimized by the OEP equa-
tions subject only to the constraint of locality. Hence E OEP ≥ E UHF , the Hartree–
Fock ground-state energy. If an exact local exchange potential existed for UHF
ground states, it would be determined by the OEP, implying E OEP = E UHF . Calcu-
lations for the closed-shell atoms He, Be, and Ne [1, 97] obtain OEP energies E OEP
and UHF energies E UHF that agree for He, but E OEP , in Hartree units, is −14.5724
for Be and −128.5455 for Ne, above E UHF , −14.5730 for Be and −128.5471 for Ne,
by amounts greater than the residual computational errors. When E OEP > E UHF ,
an exact local exchange potential cannot exist.
68 5 Independent-electron models

5.5 Density functional theory (DFT)

The density functional theory of Hohenberg, Kohn and Sham [173, 205] has become
the standard formalism for first-principles calculations of the electronic structure
of extended systems. Kohn and Sham postulate a model state described by a single-
determinant wave function whose electronic density function is identical to the
ground-state density of an interacting N-electron system. DFT theory is based
on Hohenberg–Kohn theorems, which show that the external potential function
v(r) of an N-electron system is determined by its ground-state electron density.
The theory can be extended to nonzero temperatures by considering a statistical
electron density defined by Fermi–Dirac occupation numbers [241]. The theory is
also easily extended to the spin-indexed density characteristic of UHF theory and
of the two-fluid model of spin-polarized metals [414].

5.5.1 The Hohenberg–Kohn theorems

For N-electron ground states, Hohenberg and Kohn [173] (HK) proved theorems that
remain valid for spin-indexed electron density functions ρ(r) and external potential
functions v(r): (i) the electron density determines the external potential and hence
the ground-state wave function ! and all physical properties; and (ii) a universal
functional F[ρ] is defined such that the energy functional E v = F + vρ d 3 r is
minimized by the ground-state density function, and yields the ground-state energy
as its minimum value. The most straightforward proof, due to Levy [222], is a
constrained-search construction. Given the N-electron Hamiltonian operator H =
N
T̂ + Û + V̂ , where
 V̂ = i=1 v(ri ), and a specified density function ρ normalized
to N electrons, ρ d r = N , and using the notation · · ·t = (!t | · · · |!t )/(!t |!t ),
3

then

F[ρ] = min T̂ + Û t

!t →ρ

defines a universal functional. Thus the mean value of H − V̂ = T̂ + Û is mini-

mized over all N-electron wave functions whose density function is the specified ρ.
This requires that the given density must be physically realizable in a ground state.
The constraint ρt = ρ can be imposed using a Lagrange-multiplier field that is just
the external potential function v(r). Any trial external potential defines a functional



Fv [ρ] = min T̂ + Û t + v(ρt − ρ)d r = E[v] − vρ d 3 r,
3
!t

where E[v] is the ground-state energy for external potential v(r). The minimizing
wave function is !v and the density is ρv . If v = vρ is chosen such that ρv = ρ,
 5.5 Density functional theory (DFT) 69

this defines


F[ρ] = Fv [ρv ] = E[v] − vρv d 3 r,

which can be computed in the ground state for any v.

A density function ρ is said to be “v-representable” if the Lagrange multiplier
 Theorem (i) by construction. vρ contains an
field v = vρ exists. If so, this proves HK
arbitrary additive constant because (ρt − ρ)d 3 r = N − N = 0. This constant is
fixed by the physical requirement that interactions vanish at infinite separation. For
a nondegenerate ground state, v determines a unique density ρ. Uniqueness of the
inverse mapping, from ρ to v, is proved by contradiction [173]. In the Levy construc-
tion [222], suppose that v1 and v2 define different ground-state wave functions !1
and !2 and distinct functionals
 F1 and F2 for the same ρ. The construction implies
that F1 [ρ] = E[v1 ] − v1 ρ ≤ F2 [ρ] because F2 is just the mean value T̂ + Û 2
evaluated for !2 . But the theorem also implies F2 [ρ] = E[v2 ] − v2 ρ ≤ F1 [ρ].
This is a contradiction unless v[ρ] and the functional F[ρ] are uniquely de-
fined for nondegenerate
 ground states. The Hohenberg–Kohn energy functional
is E v [ρ] = F[ρ] + vρ d 3 r. Suppose that F[ρ] corresponds to !ρ but ρ = ρv .
Then

E v [ρ] = T̂ + Û + V̂ ρ ≥ E[v],

which proves HK Theorem (ii).

Since the exact ground-state electronic wave function and density can only be
approximated for most N-electron systems, a variational theory is needed for the
practical case exemplified by an orbital functional theory. As shown in Section 5.1,
any rule ! → " defines an orbital functional theory that in principle is exact for
ground states. The reference state " for any N-electron wave function ! determines
an orbital energy functional E = E 0 + E c , in which E 0 = T + E h + E x + V is a
sum of explicit orbital functionals, and E c is a residual correlation energy functional.
In practice, the combination of exchange and correlation energy is approximated
by an orbital functional E xc .
The Levy construction [222] can be used to prove Hohenberg–Kohn theorems
for the ground state of any such theory. It should be noted that any explicit model of
the Hohenberg–Kohn functional F[ρ] implies a corresponding orbital functional
theory. The relevant density function ρ(r) is that constructed from an OFT ground

state. This has the orbital decomposition i n i φi∗ φi , as postulated by Kohn and
Sham [205]. Unlike the density ρ! for an exact N-electron wave function !, which
cannot be determined for most systems of interest, the OFT ground-state density
function is constructed from explicit solutions of the orbital Euler–Lagrange equa-
tions, and the theory is self-contained.
70 5 Independent-electron models

With fixed occupation numbers n i = 1 for i ≤ N , an orbital functional is a func-

tional of a model or reference state ". For any model functional E xc ["], an orbital
functional F = T + E h + E xc is defined such that min(F + V ) = E[v] approx-
imates an N-electron ground-state energy. The occupied orbital functions of the
reference state " satisfy orbital Euler–Lagrange (OEL) equations of the form given
above. Considering density functions ρ" constructed from such orbital functions,
suppose that one such function ρ is specified. The orbital functional F["] is to be
minimized subject to ρ" = ρ, following the Levy construction [222]. The density
constraint is enforced using a Lagrange-multiplier field v(r). For any given field
v(r) a functional of ρ is defined by



Fv [ρ] = min F["t ] + v(ρt − ρ)d 3 r = E[v] − vρ d 3 r.
"t

The minimizing model function "v determines ρv . A ground-state density func-

tional Fs [ρ] is defined by Fv [ρ] when v is chosen such that ρv = ρ. The no-
tation Fs distinguishes a density functional restricted to ground states from the
equivalent orbital functional F, defined for all functions in the orbital Hilbert
space. This construction determines v for any given ρ if the variational prob-
lem has a solution. E v [ρ] = Fs [ρ] + vρ d 3 r defines an energy functional for
arbitrary v. When ρ = ρv , E v [ρv ] = E[v], its minimum value. When ρ = ρv ,
E v [ρ] = F["] + vρ d 3 r for some "(→ ρ) = "v . Hence E v [ρ] ≥ E v [ρv ]. This
establishes the variational property of E v [ρ], verifying the Hohenberg–Kohn the-
orems. Uniqueness of the Lagrange multiplier field v is implied for nondegenerate
OFT solutions that minimize the orbital functional E.

5.5.2 Kohn–Sham equations

In any practical application of the Hohenberg–Kohn theory, a specified density
functional Fs [ρ] restricted to ground-state densities defines an equivalent orbital
functional F[{φi , n i }] that can be extended to all functions in the orbital Hilbert
space. The OEL equations for occupied orbitals of the reference state of an
N-electron ground state take the form, for i ≤ N ,

{tˆ + vh (r) + v̂xc }φi = {i − v(r)}φi ,

derived from the orbital functional derivative

δF
= fˆφi = {tˆ + vh (r) + v̂xc }φi .
n i δφi∗
 5.5 Density functional theory (DFT) 71

For exchange-correlation energy expressed as an orbital functional,

δ E xc
v̂xc φi =
n i δφi∗
defines a generally nonlocal potential v̂xc . The Kohn–Sham equations [205] for the
model state corresponding to the same ground state are

{tˆ + vh (r) + vxc (r)}φi = {i − v(r)}φi ,

if a functional derivative δ δρ
E xc
= vxc (r) exists in the form of a local potential function.
Assuming the existence of such a Fréchet functional derivative [26, 102] constitutes
the locality hypothesis. If this hypothesis were valid, the OEL and Kohn–Sham
equations would be equivalent, determining the same model or reference state.
It might appear paradoxical to invoke the locality hypothesis for vxc (r) while
using the Schrödinger operator tˆ rather than an assumed Fréchet derivative δT δρ
=
vT (r) for the kinetic energy. Applying Hohenberg–Kohn theory to a noninteracting
N-electron system, Kohn and Sham [205] show that the kinetic energy can be
expressed as a functional Ts [ρ] of the same ground-state density used to define
Fs [ρ]. Following the constrained-search logic of Levy [222], the density constraint
can be enforced by a Lagrange multiplier field w(r) that acts as a local effective
potential in the Kohn–Sham orbital equations. Thus it would appear that ground-
state theory for an interacting N-electron system can be replaced by a noninteracting
model constructed to obtain the same density, and that the theory requires only a
local potential w(r). Such a conclusion depends on the locality hypothesis, that
Fréchet functional derivatives exist for density functionals. This question is most
easily examined for a noninteracting system in the context of Thomas–Fermi theory,
where only the kinetic energy is relevant. Before examining this issue, consistency
conditions are derived for orbital and density functional derivatives of the same
functional.

5.5.3 Functional derivatives and local potentials

The variation of an orbital functional induced by infinitesimal orbital variations is

  δF

∗
δF = d r 3
n i δφi (r) + cc .
i
n i δφi∗ (r)

If this becomes a density functional Fs [ρ] for ground states, the density functional
variation is

  δ Fs

∗
δ Fs = d r 3
n i δφi (r) φi (r) + cc .
i
δρ(r)
72 5 Independent-electron models

Consistency between these variations of a density functional and of the equiva-

δ Fs δF
lent orbital functional implies the chain rule δρ(r) φi (r) = ni δφ ∗ , regardless of any
i (r)
constraint on variations. This implies the sum rule
 δ Fs  δF
n i φi∗ φi = n i φi∗ .
i
δρ i
n i δφi∗

An orbital functional derivative in general defines a linear operator such that niδδφ F
∗ =
δ Fs i
v̂ F φi . If the locality hypothesis were valid, then δρ(r) = v F (r), and the implied local
potential function could be computed directly from the sum rule,

v F (r)ρ(r) = n i φi∗ v̂ F φi .
i

This formula was used by Slater [385] to define an effective local exchange potential.
The generally unsatisfactory results obtained in calculations with this potential
indicate that the locality hypothesis fails for the density functional derivative of the
exchange energy E x [294].

5.5.4 Thomas–Fermi theory

The Hohenberg–Kohn theory of N-electron ground states is based on consideration
of the spin-indexed density function. Much earlier in the development of quantum
mechanics, Thomas–Fermi theory [402, 108] (TFT) was formulated as exactly
such a density-dependent formalism, justified as a semiclassical statistical theory
[231, 232]. Since Hohenberg–Kohn theory establishes the existence of an exact
universal functional Fs [ρ] for ground states, it apparently implies the existence
of an exact ground-state Thomas–Fermi theory. The variational theory that might
support such a conclusion is considered here.
The orbital theory of Kohn and Sham [205] differs from Thomas–Fermi theory
 
in that the density function ρ = i n i ρi = i n i φi∗ φi is postulated to have an
orbital structure, with occupation numbers n i that are consistent with Fermi–Dirac
statistics. The orbital functions φi are determined by orbital Kohn–Sham (KS)
equations, in which the kinetic energy is represented by the operator tˆ = − 12 ∇ 2
of Schrödinger. The Thomas–Fermi equations ignore such orbital structure and
represent kinetic energy by a local effective potential vT (r), defined in principle
as the functional derivative δTs /δρ of a kinetic-energy functional Ts [ρ]. If such a
Fréchet derivative exists, the two theories are just alternative ways of describing
the same ground state.
In Kohn–Sham theory, densities are postulated to be sums of orbital densities,
for functions φi in the orbital Hilbert space. This generates a Banach space [102]
of density functions. Thomas–Fermi theory can be derived if an energy functional
E[ρ] = Fs [ρ] + V [ρ] is postulated to exist, defined for all normalized ground-state
 5.5 Density functional theory (DFT) 73

densities in this density function space, including all infinitesimal neighborhoods

of such densities.
 Introducing a Lagrange multiplier µ to enforce the normalization
constraint ρ d r = N , the variational equation is
3

"  #
δ Fs [ρ] + V [ρ] − µ ρ d 3 r − N = 0,
or, if a Fréchet functional derivative δδρFs exists,

 
δ Fs
δρ + v(r) − µ d 3 r = 0.
δρ
If Fs is defined for unrestricted variations of ρ in any infinitesimal function neigh-
borhood of a solution, this implies the Thomas–Fermi (TF) equation
δ Fs
= µ − v(r).
δρ
The Lagrange multiplier µ, determined by normalization, is the chemical potential
[232], such that µ = ∂ E/∂ N when the indicated derivative is defined. This deriva-
tion requires the locality hypothesis, that a Fréchet derivative of Fs [ρ] exists as a
local function v F (r).
The locality hypothesis can be tested in a noninteracting model, in which the func-

tional Fs is replaced by Ts . The kinetic energy orbital functional is T = i n i (i|tˆ|i)
and the OEL equations are just orbital Schrödinger equations
tˆφi = {i − v(r)}φi . (5.8)
If the locality hypothesis is valid, then ∂∂ρTs = vT (r), and the Thomas–Fermi equation
is
vT (r) = µ − v(r). (5.9)

The trace sum i n i (i|vT − tˆ|i) must vanish if the OEL and TF equations are
equivalent. Equations (5.8) and (5.9), multiplied by appropriate factors, summed
over orbitals, and integrated, imply the sum rule [288]

n i i = N µ.
i

Since i n i = N and all i ≤ µ, this implies that all i are equal, in violation of
the exclusion principle for any compact system with more than two electrons. The
failure of this sum rule implies that in general the assumed Fréchet derivative of
Ts [ρ] cannot exist for more than two electrons, and there can be no exact Thomas–
Fermi theory.
If the hypothetical Fréchet derivative vT (r) could be replaced by the operator t̂
when acting on occupied orbital functions φi , there would be no contradiction. It
74 5 Independent-electron models

can be shown that this is the correct implication of variational theory. The functional
derivative in question can be shown to be a Gâteaux derivative [26, 102], the analog
in functional analysis of a partial derivative, equivalent to a linear operator that acts
on orbital functions. Consider variations of Ts induced by unconstrained orbital
variations about ground-state solutions of the OEL equations



δTs = d 3 r n i (δφi∗ t̂φi + cc) = d 3 r n i (δφi∗ {i − v}φi + cc)
i i


= d 3r n i {i − v(r)}δρi (r).
i

Since ρ = i n i ρi , this equation determines partial functional derivatives
δTs
= i − v(r). (5.10)
n i δρi
This defines a Gâteaux functional derivative [26, 102], whose value depends on a
“direction” in the function space, reducing to a Fréchet derivative only if all i are
equal. Defining H = t̂ + v, an explicit orbital index is not needed if Eq. (5.10) is
interpreted to define a linear operator acting on orbital wave functions, H − v =
t̂. The elementary chain rule nδT s
i δρi
= ni∂ρ∂ρi δT
δρ
s
= δT
δρ
s
is valid when the functional
derivatives are interpreted as linear operators. This confirms the chain rule, δT s
δρ i
φ =
δT
n i δφi∗
= t̂φi .
This argument shows that the locality hypothesis fails for more than two electrons
because the assumed Fréchet derivative must be generalized to a Gâteaux derivative,
equivalent in the context of OEL equations to a linear operator that acts on orbital
wave functions. The conclusion is that the use by Kohn and Sham of Schrödinger’s
operator t̂ is variationally correct, but no equivalent Thomas–Fermi theory exists
for more than two electrons. Empirical evidence (atomic shell structure, chemical
binding) supports the Kohn–Sham choice of the nonlocal kinetic energy operator,
in comparison with Thomas–Fermi theory [288]. A further implication is that if
an explicit approximate local density functional E xc is postulated, as in the local-
density approximation (LDA) [205],  the resulting Kohn–Sham theory is variation-
ally correct. Typically, for E xc = exc (ρ)ρ d r, the density functional derivative
3

is a Fréchet derivative, the local potential function vxc = exc + ρ dexc /dρ.

5.5.5 The Kohn–Sham construction

A particular mapping ! → " is determined by the Kohn–Sham construction

(KSC): minimize the kinetic energy orbital functional T = i (i|t̂|i) for speci-
fied spin-indexed electron density ρ. This applies Hohenberg–Kohn logic to a
 5.5 Density functional theory (DFT) 75

noninteracting system and constructs a ground-state kinetic energy functional Ts [ρ].

If the Levy constrained-search algorithm has a solution, the specified density ρ is
said to be “noninteracting v-representable”, and a Lagrange-multiplier field w(r) is
determined that acts as a local potential in the resulting noninteracting Kohn–Sham
equations. If the specified density is that of a nondegenerate ground state of an
interacting system, and if E xc [ρ] for this system has a Fréchet functional derivative
that defines a local potential vxc (r), then Hohenberg–Kohn logic implies that the
orbital equations for the interacting and noninteracting systems must be equiva-
lent. Because the local potentials must be equal if both vanish at infinity, vxc (r) is
determined by subtraction:
vxc (r) = w(r) − v(r) − vh (r).
Hohenberg–Kohn theorems for any orbital functional model imply that the Kohn–
Sham construction must result in noninteracting KS equations equivalent to the
OEL equations if the latter contain only local potential functions. The relevant
density function is that computed from the orbital functional ground state. Since
the exchange-correlation term in the OEL equations is nδiEδφxc∗ = v̂xc φi , the existence
i
of a Fréchet derivative vxc (r) is not assured, and must be proven for each particular
model of E xc . However, vxc (r) always exists in the local-density approximation.
If a local potential vxc does not exist, there is no clear relationship between the
noninteracting orbital equations obtained by the KS construction and the ground-
state OEL equations which minimize the energy functional.
The unrestricted Hartree–Fock theory (UHF) for closed-shell atoms provides an
exchange-only orbital model for which Hohenberg–Kohn theorems can be proved
[324]. Ground-state orbital wave functions and energies and total variational energy
are known to high numerical accuracy [130]. The Kohn–Sham construction (KSC)
has been carried out for the atoms He, Be, and Ne with sufficient numerical accuracy
to test consequences of the locality hypothesis for the exchange potential [152, 294].
These calculations verify by construction that the UHF densities have the property
of “noninteracting v-representability”, but the computed wave functions are not the
ground states of the variational UHF model, for more than two electrons. Results
can be compared with OEP calculations (optimized effective potential) for these
atoms [1, 97]. Because both KSC and OEP constrain the exchange potential to be
local, while KSC also constrains the density function, these nested variational
conditions imply that E KSC ≥ E OEP ≥ E UHF [152]. The computed energies are
shown in Table (5.1). These computed energies are consistent with the existence of
a local exchange potential for He but not for Be and Ne. This confirms the discus-
sion given above, which indicates that Fréchet functional derivatives can exist for
two electrons, but not in general otherwise, because density functional derivatives
differ for partial orbital densities that correspond to different orbital energies.
76 5 Independent-electron models

Table 5.1. Computed total energies for typical atoms

(Hartree units)

Atom E UHF E OEP E KSC

He −2.8617 −2.8617 −2.8617
Be −14.5730 −14.5724 −14.5724
Ne −128.5471 −128.5455 −128.5454

If an exact local exchange potential does not exist, there is no reason for UHF,
OEP, and KSC results to be the same in the UHF model. That the OEP density is
not exactly equal to that of the UHF ground state is indicated by an analysis of OEP
results [412], and by recent test calculations [69]. This would imply that the density
constraint in KSC is a true variational constraint for more than two electrons, so that
E KSC > E OEP . The calculations considered here may not have sufficient numerical
accuracy to establish this evidently small energy difference.
 6
Time-dependent theory and linear response

The principal references for this chapter are:

[79] Dirac, P.A.M. (1930). Note on exchange phenomena in the Thomas atom, Proc.
Camb. Phil. Soc. 26, 376–385.
[94] Ehrenreich, H. and Cohen, M.H. (1959). Self-consistent field approach to the
many-electron problem, Phys. Rev. 115, 786–790.
[155] Gross, E.K.U., Dobson, J.F. and Petersilka, M. (1996). Density functional theory of
time-dependent phenomena, in: Density Functional Theory, ed. R.F. Nalewajski
(Springer, Berlin).
[238] McLachlan, A.D. and Ball, M.A. (1964). Time-dependent Hartree-Fock theory for
molecules, Rev. Mod. Phys. 36, 844–855.
[292] Nesbet, R.K. (2001). Orbital functional theory of linear response and excitation, Int.
J. Quantum Chem. 86, 342–346.
[329] Petersilka, M., Gossmann, U.J. and Gross, E.K.U. (1996). Excitation energies from
time-dependent density-functional theory, Phys. Rev. Lett. 76, 1212–1215.
[351] Runge, E. and Gross, E.K.U. (1984). Density-functional theory for time-dependent
systems, Phys. Rev. Lett. 52, 997–1000.
[407] Thouless, D.J. (1961). The Quantum Mechanics of Many-body Systems (Academic
Press, New York).

In the quantum theory of interacting electrons, a physically correct theory of time

dependence should in principle be formulated as a relativistic quantum field the-
ory. The physical model is that of electrons, each characterized by a probability
distribution over space-time events xi , ti , that interact indirectly through the quan-
tized electromagnetic field. This theory is simplified for particular applications by
neglecting true radiative effects of quantum electrodynamics, and by passing to the
limit of large c, the velocity of light in vacuo. Although this theory cannot be de-
veloped with adequate detail in the present context, the discussion will emphasize
independent-electron models that are consistent with a physical picture of instanta-
neous direct interactions replaced by a mean field that varies with a time parameter
that is the same for all electrons.

77
78 6 Time-dependent theory and linear response

6.1 The time-dependent Schrödinger equation for one electron

For a single electron, the time-independent Schrödinger eigenvalue problem is
determined by the variational condition


d 3 x[δψ ∗ {H − }ψ + cc] = 0,

for variations of trial functions ψ that belong to the usual Hilbert space and satisfy
specified boundary conditions on the surface of the volume of integration.  3 The
Lagrange multiplier  is determined so that the normalization integral d xψ ∗ ψ
remains constant. The resulting Schrödinger equation
{H − }ψ = 0,
is modified in time-dependent theory to
∂
ih ψ(xt) = H(xt)ψ(xt).
∂t
The simplified notation (xt) is used here to denote (x, t). If H is independent of
time, these two equations are equivalent. The wave function ψ(x) is modified by a
time-dependent phase factor exp(t/i h), which has no physical consequences.
Following Hamilton’s principle in classical mechanics, the required time depen-
dence can be derived from a variational principle based on a seemingly artificial
Lagrangian density, integrated over both space and time to define the functional

t1
 
∂
A[ψ] = dt d 3 xψ ∗ i h − H ψ.
t0 ∂t
Treating variations of ψ and ψ ∗ as independent, because they lead to equivalent
Euler–Lagrange equations, and integrating the time integral by parts as in Euler’s
theory, the resulting variational expression is

t1
   
∗ ∂
δA = dt d x δψ i h − H ψ + cc = 0,
3
t0 ∂t
 3 ∗
if the normalization integral d xψ ψ is held constant. This implies the time-
dependent Schrödinger equation in the time interval t0 ≤ t ≤ t1 .
 t integral A is not changed if the trial function ψ is multiplied by a phase
The action
factor exp( γ (t  )dt  /i h), while H is increased by a time-dependent but spatially
uniform potential γ (t). This is an example of gauge invariance, taken out of the
usual context of electromagnetic theory. Indicating the modified wave function by
ψγ , the modified action integral is

t1
   
∗ ∂
Aγ = dt d x δψγ i h − H − γ (t) ψγ + cc .
3
t0 ∂t
 6.2 The independent-electron model as a quantum field theory 79

An alternative interpretation of this equation is that γ (t) is a time-dependent

Lagrange multiplier, introduced to enforce constant normalization (ψ|ψ). By the
conventional definition of an isolated system, asymptotic values of potential func-
tions in H should become negligible outside some enclosing surface. Since for
finite systems this is a property of potentials derived from Coulombic interactions,
the otherwise arbitrary function γ (t) is set to zero for such potentials.

6.2 The independent-electron model as a quantum field theory

The reference state " of N-electron theory becomes a reference vacuum state |" in
the field theory. A complete orthonormal set of spin-indexed orbital functions φ p (x)
is defined by eigenfunctions of a one-electron Hamiltonian H, with eigenvalues  p .
The reference vacuum state corresponds to the ground state of a noninteracting
N-electron system determined by this Hamiltonian. N occupied orbital functions
† †
(i ≤ µ) are characterized by fermion creation operators ai such that ai |" = 0.
Here µ is the chemical potential or Fermi level. A complementary orthogonal set
of unoccupied orbital functions are characterized by destruction operators aa such
that aa |" = 0 for a > µ and a > N . A fermion quantum field is defined in this
orbital basis by

ψ(x, t) = φ p (x)a p (t).
p
† †
The fermion creation and destruction operators are defined such that a p aq + aq a p =
δ pq . In analogy to relativistic theory, and more appropriate to the linear response
theory to be considered here, the elementary fermion operators a p can be treated
as algebraic objects fixed in time, while the orbital functions are solutions of a
time-dependent Schrödinger equation
∂
ih φ p (xt) = H(xt)φ p (xt).
∂t
The fermion field operator takes the form

ψ(xt) = φ p (xt)a p .
p

†
Occupation numbers are defined by n p = "|a p a p |", such that n i = 1, n a = 0
for i ≤ N < a. An (N−1)-electron basis state is defined such that |"i  = ai |" if
i ≤ N . The orthonormality condition is verified by
†
"i |" j  = "|ai a j |" = n i δi j .
†
An (N+1)-electron basis state is defined such that |"a  = aa |" if a > N . Here
80 6 Time-dependent theory and linear response

the orthonormality condition is

†
"a |"b  = "|aa ab |" = (1 − n a )δab .

These definitions remain valid in relativistic theory.

A Dirac density operator is defined at specified time by its matrix kernel

ρ̂(x, x ) = φ p (x)n p φ ∗p (x ) = "|ψ † (x )ψ(x)|".
p

The Dirac density operator for the reference state is idempotent:

 
ρ̂ 2 = φ p n p ( p|q)n q φq∗ = φ p (x)n 2p φ ∗p (x ) = ρ̂,
p,q p

because n 2p = n p for either occupied or unoccupied orbitals. Normalization is such


that Trρ̂ = p n p = N .
Reduction of the time-dependent N-electron problem to an independent-electron
model appears to be consistent with quantum field theory, if true radiative effects are
neglected. In this approximation, the theory can be developed in Coulomb gauge,
with instantaneous interactions. Except for effects due to quasiparticle lifetimes,
which must be considered whenever there are continuum states, the mean field
of the independent-particle model changes instantaneously with changes of local
electron density. The orbital wave functions satisfy Schrödinger or Dirac equations
determined by a time-dependent effective mean field. It appears plausible that the
time development of such a system should be described as in time-independent
theory at each instant, carried forward in time by a continuously developing unitary
change of representation of the self-consistent orbital basis. This is the physical
implication of Dirac’s density operator, which remains idempotent so long as the
effective Hamiltonian is Hermitian.
Ultimately, the theory must be consistent with quantum electrodynamics, which
reduces in the absence of radiative terms to a time-dependent equation, in which ψ

is interpreted as a fermion field operator φ p a p acting on the state function |".
The field equation of motion is equivalent to simultaneous equations
 
∂
i h − G φi (xt)ai |" = 0,
∂t

for noninteracting electrons in the self-consistent mean field described by an

operator G. In the nonrelativistic orbital theories considered here, G reduces to
an effective one-electron Hamiltonian operator.
 6.3 Time-dependent Hartree–Fock (TDHF) theory 81

6.3 Time-dependent Hartree–Fock (TDHF) theory

6.3.1 Operator form of Hartree–Fock equations
When the idempotent density operator ρ̂ is constructed from orbital solutions of
the Hartree–Fock equations, (H − i )φi = 0, it satisfies the commutator equation

[H, ρ̂] = Hρ̂ − ρ̂H = 0.

This follows from

 
Hρ̂ = {Hφi (x)}n i φi∗ (x ) = φi (x)i n i φi∗ (x ).
i i

Similarly, operating to the left with H,


ρ̂H = φi (x)n i i∗ φi∗ (x ).
i

Hence, for real eigenvalues, Hρ̂ − ρ̂H = 0.

As shown by Dirac [79], the corresponding time-dependent equation takes the
form
∂
ih ρ̂(t) = H(t)ρ̂(t) − ρ̂(t)H(t) = [H(t), ρ̂(t)],
∂t
consistent with a time-dependent Hartree–Fock equation,
∂
ih φi = Hφi .
∂t
The time-dependent Hartree–Fock equation is expressed formally by
t
H(u)du/i h
φi (xt) = e 0 φi (x0).

This implies consistency conditions for Hermitian H,


 t H(u)du/i h ∗   t 
φi∗ (yt)φ j (yt) = e0 φi (y0) e 0 H(u)du/i h φ j (y0)
y y

t †
= φi∗ (y0)e 0 (H(u)−H (u))du/i h
φ j (y0) = δi j .
y

If H(t) ≡ H(0), φi (xt) = ei t/i h φi (x0).

6.3.2 The screening response

For a stationary state at t = 0, [H(0), ρ̂(0)] = 0. When t ≥ 0, a weak perturbing
potential v(t) modifies the effective Hamiltonian by an induced screening term
82 6 Time-dependent theory and linear response

such that H = v + Hs . The first-order perturbation equations for ρ(t) =

ρ̂(t) − ρ̂(0) are
∂
ih ρ = [H(0), ρ] + [H, ρ̂(0)].
∂t
In a basis set of occupied and unoccupied orbitals of a reference state ", variation
of an occupied orbital φi affects " only through an incremental sum of unoccupied

orbitals φi (t) = a (1 − n a )φa cia (t). The corresponding variation of the density
matrix is
  
ρ(x, x ; t) = n i (1 − n a ) φa (x)cia (t)φi∗ (x ) + φi (x)cia (t)∗ φa∗ (x ) .
i a

From the Fock operator j n j ( j|ū| j),
  
Hs (t) = n j (1 − n b ) ( j|ū|b)cbj (t) + (b|ū| j)cbj (t)∗ .
j b

Using ( p|H(0)| j) =  p δ pj and ( j|ρ̂(0)|q) = n q δ jq ,

( p|[H(0), ρ]|q) = ( p − q )( p|ρ|q);

( p|[H, ρ̂(0)]|q) = (n q − n p )( p|H|q).

The linear response to v(xt) = 2w(x)(e−iωt ), a weak time-dependent perturb-

ing potential, is given for indices i, j ≤ N < a, b by

i hċia (t) = (a − i )cia (t) + (n i − n a )(a|v(xt)|i)

 
+ (n i − n a ) (a j|ū|ib)cbj (t) + (ab|ū|i j)cbj (t)∗ .
j b

∗
For cia (t) = X ia e−iωt + Yia∗ eiω t , treating frequency ω as a complex number,
 
(a − i − hω)X ia + (n i − n a ) (a j|ū|ib)X bj + (ab|ū|i j)Y jb
j b
= − (n i − n a )(a|w|i),
 
(a − i + hω)Yia + (n i − n a ) (i j|ū|ab)X bj + (ib|ū|a j)Y jb
j b
= − (n i − n a )(i|w|a).

If there is no driving term w(x) in these TDHF equations, discrete excitation

energies are determined by values of hω for which the determinant of the residual
homogeneous equations vanishes. This gives the equations of the random-phase
approximation (RPA) for excitation energies, in an exact-exchange model [94, 407].
 6.4 Time-dependent orbital functional theory (TOFT) 83

6.4 Time-dependent orbital functional theory (TOFT)

6.4.1 Remarks on time-dependent theory
Dirac’s development of TDHF theory invokes the Heisenberg equation of motion
for the density operator as a basic postulate,

∂
ih ρ̂ = [H, ρ̂].
∂t
Equivalently, time-dependent canonical Hartree–Fock equations are assumed to
take the same form as the time-dependent Schrödinger equation.
These concepts, inherent in the TDHF formalism, generalize immediately to
orbital functional theory, when electronic correlation energy is included in the
model. Given some definition ! → " that determines a reference state " for any
N-electron state !, correlation energy can be defined for any stationary state by
E c = E − E 0 , where E 0 = ("| Ĥ |") and E = ("| Ĥ |!). Conventional normaliza-
tion ("|!) = ("|") = 1 is assumed. A formally exact functional E c ["] exists for
stationary states, for which a mapping " → ! is established by the Schrödinger
equation [292]. Because both " and ρ̂ are defined by the occupied orbital func-
tions {φi }, for fixed occupation numbers n i , E["], E[ρ̂] and E[{φi }] are equivalent
functionals. Since E 0 is an explicit orbital functional, any approximation to E c as
an orbital functional defines a TOFT theory. Because a formally exact functional
E c exists for stationary states, linear response of such a state can also be described
by a formally exact TOFT theory. In nonperturbative time-dependent theory, total
energy is defined only as a mean value E(t), which lies outside the range of def-
inition of the exact orbital functional E c [{φi }] for stationary states. Although this
may preclude a formally exact TOFT theory, the formalism remains valid for any
model based on an approximate functional E c .
Any postulated orbital functional E[{φi }] defines an action integral,

t1


∗ ∂φi
A[{φi }] = dt n i d x φi i h
3
− E[{φi }] ,
t0 i
∂t

that is stationary for t0 ≤ t ≤ t1 if and only if the occupied orbital functions satisfy
the time-dependent Euler–Lagrange equations

∂φi (xt) δE
ih = = G(xt)φi (xt).
∂t n i δφi∗ (xt)

Here G = H + v̂ c , where H is defined by δ E 0 /n i δφi∗ = Hφi as in Hartree–Fock

theory. v̂ c is defined by the functional derivative nδi δφ
Ec
∗ = v̂ c φi . If E c is omitted, the
i
theory reduces to TDHF.
84 6 Time-dependent theory and linear response

6.4.2 Exact linear response theory

When extended to include electronic correlation, for which an exact but implicit
orbital functional was derived above, the TDHF formalism becomes a formally
exact theory of linear response. In practice, some simplified orbital functional
E c [{φi }] must be used, and the accuracy of results is limited by this choice. The
Hartree–Fock operator H is replaced by G = H + v̂ c . Dirac defines an idempo-

tent density operator ρ̂ whose kernel is i φi (r)n i φi∗ (r ). The OEL equations
are equivalent to [G, ρ̂] = 0. The corresponding time-dependent equations are
i h ∂t∂ ρ̂ = [G(t), ρ̂(t)]. Dirac proved, for Hermitian G, that the time-dependent equa-
tion i h ∂t∂ φi (rt) = G(rt)φi (rt) implies that ρ̂(t) is idempotent. Hence ρ̂(t) corre-
sponds to a normalized time-dependent reference state.
For t ≥ 0, a weak perturbing potential v(rt) induces a screening potential such
that G = v + Gs . The first-order perturbation equations are
∂
ihρ̂ = [G(0), ρ̂] + [G, ρ̂(0)].
∂t

φi (t) = a (1 − n a )φa cia (t) in this orbital representation. Hence the kernel of ρ̂
is
  
ρ(r, r ; t) = n i (1 − n a ) φa (r)cia (t)φi∗ (r ) + φi (r)cia (t)∗ φa∗ (r ) .
i a

Using ( p|G(0)| j) =  p δ pj and ( j|ρ̂(0)|q) = n q δ jq , in the basis of eigenfunctions

of G(0)
( p|[G(0), ρ̂]|q) = ( p − q )( p|ρ̂|q);
( p|[G, ρ̂(0)]|q) = (n q − n p )( p|G|q).
The equation of motion implied for cia (t) is

i hċia (t) = (a − i )cia (t) + (n i − n a )(a|G|i),

for i ≤ N < a.

6.4.3 Definition of the response kernel

Because G itself is defined by an orbital functional derivative, the increment Gs
is proportional to a functional second derivative. It is convenient to define a re-
 
sponse kernel fˆ such that φ j (t) induces v̂ = j n j b (1 − n b )[( j| fˆ|b)cbj (t) +
δF
j (t)(b| f | j)]. If v̂ = n i δφi∗ , this equation defines a functional second derivative
cb∗ ˆ
in the form fˆφ j (r) = n jδδφ v̂
∗ . In agreement with Dirac [79] and with the second-
j
quantized Hamiltonian, the response kernel for Hartree and exchange energy
functionals is fˆh + fˆx = ū = r112 (1 − P̂ 12 ).
 6.5 Reconciliation of N-electron theory and orbital models 85

The response kernel fˆc is a linear operator such that fˆc φ j = n δj δφ

v̂ c
∗ . Variations of
j
unoccupied orbitals δφa (N < a) in the functional E c are induced by variations of
occupied orbitals δφi (i ≤ N ) through unitarity. The combined total response kernel
is the linear operator fˆh + fˆx + fˆc = fˆhxc = ū + fˆc .
The equations of motion for the coefficients cia (t) are

i hċia (t) = (a − i )cia (t) + (n i − n a ) (a|v(t)|i)
" 
#
+ (a j| fˆhxc |ib)cbj (t) + (ab| fˆhxc |i j)cbj (t)∗ ,
j b

for indices i, j ≤ N < a, b. Setting v(rt) = 2w(r)(e−iωt ) for complex fre-

quency ω, the independent coefficients defined by
∗
cia (t) = X ia e−iωt + Yia∗ eiω t

satisfy linear inhomogeneous equations

 
(a − i − hω)X ia + (n i − n a ) (a j| fˆhxc |ib)X bj + (ab| fˆhxc |i j)Y jb
j b
= −(n i − n a )(a|w|i),
 
(a − i + hω)Yia + (n i − n a ) (i j| fˆhxc |ab)X bj + (ib| fˆhxc |a j)Y jb
j b
= −(n i − n a )(i|w|a).
In the exchange-only limit, these are the TDHF or RPA equations (Thouless [407],
p. 89). If w = 0, excitation energies are values of hω for which the determinant
of the residual homogeneous equations vanishes.

6.5 Reconciliation of N-electron theory and orbital models

The action integral defined in standard N-electron theory is

t1  
t1   
∂ ∂
A N [!] = dt i h − Ĥ = dt i h − E[!] ,
t0 ∂t t0 ∂t
where the notation · · · denotes (!| · · · |!)/(!|!) for an arbitrarily normalized
N-electron wave function !(t). The theory postulates that this action integral is
stationary subject to fixed normalization. For comparison, the action integral of
orbital functional theory is

t1 
 
∂
A["] = dt
" i h " − E["] .
t0 ∂t
86 6 Time-dependent theory and linear response

While E["] = E[!], the mean values of the time derivative are not equal. Even if

Ĥ is independent of time, these values are constants i n i i and E, respectively.

These constants can be reconciled by inserting a gauge potential γ = E − i n i i
into A, determined by equating the time-dependent phase factors. This also pre-
serves the normalization ("|!) = ("|") = 1 for the time-dependent wave func-
tions, so that the definition of correlation energy E c = ("| Ĥ |! − ") remains un-
changed. When the Hamiltonian is time-dependent, γ (t) must be determined so
that ("γ |i h ∂t∂ |"γ ) = i h ∂t∂ , where "γ incorporates the gauge-dependent phase
factor. For any rule ! → " applied t for all t0 ≤ t ≤ t1 , the time derivative terms in
the action integrals are #s = t01 dt("|i h ∂t∂ |") and # = t01 dti h ∂t∂ . The present
t

argument indicates that a gauge transformation can be introduced such that #s − #

drops out of the variational formalism. This is invoked below to remove such a term
from the equations of time-dependent density functional theory.

6.6 Time-dependent density functional theory (TDFT)

An extension of density functional theory to time-dependent external potentials has
recently been derived [351, 155]. Hohenberg and Kohn [173] proved, for nondegen-
erate ground states, that the electronic density function ρ(r) and the external poten-
tial function v(r) mutually determine each other. Easily extended to spin-indexed
density and potential, this implies that all properties of an N-electron ground state
are functionals of the density function. A slightly more restrictive proof, requiring
the existence of a Taylor expansion of the potential function in time, establishes an
analogous 1–1 correspondence between the (spin-indexed) time-dependent density
function ρ(xt) and an external potential v(xt) [351, 155]. This shows that all prop-
erties of the time-dependent state that evolves from a given initial state !(t0 ) in the
time interval t0 ≤ t ≤ t1 are functionals of ρ(xt). This theorem is used to justify
a time-dependent density functional theory (TDFT), based on a time-dependent
generalization of the Kohn–Sham equations. These are modified time-dependent
Schrödinger equations
∂
ih φi (xt) = {tˆ + w(xt)}φi (xt),
∂t
for each of N occupied orbital functions of a time-dependent model state ", itself
a single Slater determinant. If a Fréchet functional derivative [26] of a density
functional analogous to E − T exists in this time-dependent theory, it defines an
effective local potential w(xt) in the orbital equations. If the assumed local effective
potential is replaced by a more general linear operator or nonlocal potential, this
formalism becomes identical to TOFT, as derived above, which reduces, for an
exchange-only model, to the TDHF theory of Dirac [79].
 6.6 Time-dependent density functional theory (TDFT) 87

Generalized Hohenberg–Kohn theorems are proved by developing the action

integrals appropriate to the time-dependent theory [351, 155]. Assumptions can
be minimized by carrying out a time-dependent version of the constrained-search
construction of Levy [222, 201]. The variational condition that a certain action
integral should be stationary subject to the constraint of specified density ρ(xt)
is implemented here by introducing the time-dependent external potential v(xt)
as a Lagrange-multiplier field for the density constraint. If these equations have
a stationary solution, this establishes the property of “v-representability” of the
specified density. An action integral is defined for arbitrarily normalized time-
dependent N-electron trial functions !(t) by

t1  
∂
B[!(t)] = dt i h − T̂ − Û .
t0 ∂t
The N-electron Hamiltonian is Ĥ = T̂ + Û + V̂ , where the three terms represent
kinetic energy, interelectronic Coulomb interaction, and an external field, respec-
tively. The variational condition that determines  for t0 ≤ t ≤ t1 is


t1

δ B[!(t)] − dt d x v(xt)[ρ! (xt) − ρ(xt)] = 0,
3
t0

for variations of ! within the N-electron Hilbert space. Given a particular field v(xt),
the stationary condition determines !v (t) through the time-dependent Schrödinger
equations, but in general the density ρv (xt) computed from this wave function dif-
fers from the specified ρ(xt). A Lagrange-multiplier
 t1  3 field vρ (xt) is to be chosen such
that ρv (xt) = ρ(xt).
 Since vρ = t0 dt d x v(xt)ρ(xt) is constant for given v,
Av [ρ] = B[ρ] + vρ is defined as a functional of ρ, where B[ρ] = B[!v ], eval-
uated for v = vρ . This construction must succeed whenever the time-dependent
Schrödinger equation has a solution in the interval t0 ≤ t ≤ t1 with !(t0 ) = !0 ,
and if the Lagrange multiplier field vρ (xt) exists. This latter condition is “v-
representability”. These conditions seem very plausible for variations about well-
defined physical solutions, but there is no rigorous theory giving general uniqueness
and existence conditions. When this construction succeeds, it determines a func-
tional Av [ρ] that has a stationary value when ρ(xt) is determined by a solution of
the time-dependent Schrödinger equation defined by the external potential function
vρ (xt). Uniqueness follows from uniqueness of the solution of this time-dependent
equation [351], within a class of equivalent solutions corresponding to gauge trans-
formations. The trial functional B is “universal” in the sense that it depends only
on Ĥ − V̂ , the same for all external potentials. However, the variational condition
implies different results for different t initial conditions, !(t0 ) = !0 .
Defining the functional # = t01 dti h ∂t∂ , then Av = # − , where  =
 t1
t0 dt E v (t). Using E v (t) =  Ĥ  =  f + V̂  = F + V , stationary Av implies
ˆ
88 6 Time-dependent theory and linear response

Euler–Lagrange equations of the form

δ# δF
= + v(xt),
δρ δρ
if the indicated density functional derivatives exist as Fréchet derivatives, equal to
multiplicative local potential functions.
Applying the same argument to a noninteracting N-electron system with the same
density defines an action integral

t1

∂
Bs ["(t)] = dt " i h − T̂ " ,
t0 ∂t
where the N-electron trial functions are model functions ", in the form of single
Slater determinants. A Lagrange-multiplier field wρ (xt) is determined such that the
solution of the variational condition


t1

δ Bs ["(t)] − dt d x w(xt)[ρ" (xt) − ρ(xt)] = 0,
3
t0

determines a density equal to the specified ρ(xt) when w(xt) = wρ (xt). “Nonin-
teracting v-representability” is established by any solution of these equations. This
defines a “universal” density functional Bs [ρ] when evaluated for ρ and wρ . For
fixed w(xt) but variable ρ, the functional Asw [ρ] = Bs [ρ] + wρ is stationary
when ρ = ρw , the density obtained by solving the noninteracting time-dependent
Schrödinger equation in which the time-dependent potential is wρ = w.
Defining the functional #s = t01 dt("|i h ∂t∂ |"), then Asw = #s − s , where
t
 t1
s = t0 dt("|T̂ + Ŵ |"). Stationary Asw implies Euler–Lagrange equations of
the form
δ#s δTs
= + w(xt),
δρ δρ
if the indicated density functional derivatives exist as Fréchet derivatives.
In the noninteracting problem both required functional derivatives can be eval-
uated from explicit orbital functional derivatives. In detail, nδ# s
i δφi
∗ = ih
∂
φ and
∂t i
δTs ∗
n i δφi∗
= t φi . For variations of the partial densities ρi = φi φi , this implies Gâteaux
ˆ
functional derivatives [26, 102] nδ# s
i δρi
φi = i h ∂t∂ φi and nδT s
i δρi
φi = tˆφi . The implied or-
bital Euler–Lagrange equations are the time-dependent Schrödinger equations for
the occupied orbitals i ≤ N ,
∂
ih φi (xt) = (tˆ + w(xt))φi (xt).
∂t
Thus for noninteracting electrons, density functional analysis reproduces the usual
Schrödinger equations, but a theory based solely on densities (a time-dependent
 6.7 Excitation energies and energy gaps 89

Thomas–Fermi theory) requires a Fréchet derivative for the kinetic energy, and
cannot exist for more than two electrons [288].
For interacting electrons, the required functional derivatives cannot be evaluated
explicitly. In the TDFT derivation [155], the density functional derivative of the
kinetic energy term in  is treated correctly as equivalent to the Schrödinger operator
tˆ. Nonetheless, the functional derivative of the exchange-correlation functional is
assumed without proof to be a Fréchet derivative, equal to a time-dependent local
potential function vxc (xt). It is argued above that the difference functional # − #s
can consistently be set to zero by a gauge transformation. Taking this into account,
and assuming that vxc exists, both interacting and noninteracting equations have
the same form, and by construction produce the same density function ρ(xt). It
can then by inferred that the local potentials in these equations are equal, up to a
gauge potential γ (t). Then the Kohn–Sham potential could be decomposed into
w(xt) = v(xt) + vh (xt) + vxc (xt). Hence if a Fréchet derivative vxc did exist, the
correct interacting ρ(xt) would be obtained by solving the implied time-dependent
Kohn–Sham equation.
This proposition has been tested in the exact-exchange limit of the implied
linear-response theory [329]. The TDFT exchange response kernel disagrees qual-
itatively with the corresponding expression in Dirac’s TDHF theory [79, 289]. This
can be taken as evidence that an exact local exchange potential does not exist
in the form of a Fréchet derivative of the exchange energy functional in TDFT
theory.

6.7 Excitation energies and energy gaps

Time-dependent OFT implies matrix equations for excitation energies hω,
 
(a − i − hω)X ia + (n i − n a ) (a j| fˆhxc |ib)X bj + (ab| fˆhxc |i j)Y jb = 0,
j b
 
(a − i + hω)Yia + (n i − n a ) (i j| fˆhxc |ab)X bj + (ib| fˆhxc |a j)Y jb = 0.
j b

The simplest internally consistent approximation to an excitation energy is ob-

tained by limiting the summation to the diagonal term j, b = i, a. The second line
vanishes because of antisymmetry, (ii| fˆhxc |aa) ≡ 0, and the first line reduces to a
single equation,
{a − i − hω − (ai| fˆhxc |ai)}X ia = 0.
Neglecting correlation response, this implies a well-known formula for zeroeth-
order hole–particle excitation energies [261, 149], hω = a − i − (ai|ū|ai). The
two-electron integral here depends strongly on orbital localization. Since the lower
90 6 Time-dependent theory and linear response

state is stationary, this energy difference should be minimized to represent a

stationary excited state. This can be done by separate localization transformations
of occupied and unoccupied orbitals. This provides a mechanism to reduce Hartree–
Fock band gaps, which are systematically too large. The correlation response kernel
has the physical effect of screening the interelectronic Hartree and exchange terms.
This implies direct effects on energy gaps, evident in the TOFT formalism.
 III
Continuum states and scattering theory

This part extends quantum variational theory to continuum states. In par-

ticular, variational principles are developed for wave function continu-
ity at specified energy, which is the usual context of scattering theory.
Chapter 7, concerned with multiple scattering theory, lies somewhere be-
tween the theory of bound states and true scattering theory. Formalism
appropriate to the latter is adapted to computing the electronic structure of
large molecules and periodic solids, whose energy levels are determined
by consistency conditions for wave function continuity. A variational
formalism is derived for energy linearization. Chapter 8 develops vari-
ational principles and methods suitable for the true continuum problem
of electron scattering at specified energy. Chapter 9 presents method-
ology, some very recent, that allows rotational and vibrational effects in
electron–molecule scattering to be treated as a practicable extension of
fixed-nuclei variational theory.
 7
Multiple scattering theory for molecules and solids

The principal references for this chapter are:

[52] Butler, W.H. and Zhang, X.-G. (1991). Accuracy and convergence properties of
multiple-scattering theory in three dimensions, Phys. Rev. B 44, 969–983.
[113] Ferreira, L.G. and Leite, J.R. (1978). General formulation of the variational cellular
method for molecules and crystals, Phys. Rev. A 18, 335–343.
[148] Gonis, A. and Butler, W.H. (2000). Multiple Scattering in Solids (Springer, New
York).
[203] Kohn, W. (1952). Variational methods for periodic lattices, Phys. Rev. 87, 472–481.
[204] Kohn, W. and Rostoker, N. (1954). Solution of the Schrödinger equation in periodic
lattices with an application to metallic lithium, Phys. Rev. 94, 1111–1120.
[277] Nesbet, R.K. (1990). Full-potential multiple scattering theory, Phys. Rev. B 41,
4948–4952.
[279] Nesbet, R.K. (1992). Variational principles for full-potential multiple scattering
theory, Mat. Res. Symp. Proc. 253, 153–158.
[359] Schlosser, H. and Marcus, P. (1963). Composite wave variational method for
solution of the energy-band problem in solids, Phys. Rev. 131, 2529–2546.
[384] Skriver, H.L. (1984). The LMTO Method (Springer-Verlag, Berlin).
[432] Williams, A.R. and Morgan, J. van W. (1974). Multiple scattering by non-muffin-tin
potentials: general formulation, J. Phys. C 7, 37–60.

For direct N-electron variational methods, the computational effort increases so

rapidly with increasing N that alternative simplified methods must be used for
calculations of the electronic structure of large molecules and solids. Especially
for calculations of the electronic energy levels of solids (energy-band structure),
the methodology of choice is that of independent-electron models, usually in the
framework of density functional theory [189, 321, 90]. When restricted to local
potentials, as in the local-density approximation (LDA), this is a valid variational
theory for any N-electron system. It can readily be applied to heavy atoms by
relativistic or semirelativistic modification of the kinetic energy operator in the
orbital Kohn–Sham equations [229, 384].

93
94 7 Multiple scattering theory for molecules and solids

Solution of the one-electron Schrödinger or modified Dirac equation for a system

of many atoms is still a difficult computational task. This has been simplified and
made tractable by the methodology of multiple scattering theory (MST), recently
reviewed by Gonis and Butler [148]. These authors concentrate on methods in which
orbital energy levels are identified by searching for roots of a secular determinant
appropriate to the Kohn–Sham equations. Although MST uses the formalism of
scattering theory, it is used in practice as a bound-state method, computing en-
ergy levels by varying a specified energy until continuity conditions are satisfied.
In particular, the electronic energy bands of regular periodic solids for specified
translational quantum numbers (momentum vectors in a reduced Brillouin zone
[384, 148]) are defined as a discrete set of energy values that cause a secular deter-
minant to vanish. True continuum scattering theory is discussed in the following
chapter.
At the cost of introducing an additional level of approximation, root-search
methodology can be replaced within some fixed interval of electronic energy lev-
els (an energy panel or window) by an energy-linearized method which requires
solution of a matrix eigenvalue problem for energy eigenvalues within the panel.
Because of the very substantial relative gain in efficiency, such linearized methods
[384] are by far the most widely used in energy-band calculations. Both root-search
and energy-linearized MST methods require iteration to self-consistency, converg-
ing when the computed density function is sufficiently close to that used to construct
the potential function in the current iteration. This requires accurate solution of the
Poisson equation determined by the input density function, which is also simplified
by an adaptation of MST [148].
In addition to the model approximation implied by a restriction to local exchange
and correlation potentials, the methodology has often been further simplified by the
muffin-tin model. In this model, each atom of a molecule or solid is surrounded by
a muffin-tin sphere, of sufficiently small radius that adjacent spheres do not over-
lap. The local potential function is spherically averaged within each such sphere,
and set to a constant value (the muffin-tin zero) in the interstitial volume between
these spheres. This mathematical model greatly simplifies the required computa-
tional effort. It is physically justifiable for metals, since the interstitial region is
described in terms of electronic free wave functions, but it is counterintuitive for
atoms in molecules, since electronic exchange and correlation in chemical bonds
must be treated accurately. Earlier work restricted to the muffin-tin model has been
supplanted by full-potential methodology in recent years, based on the full local
potential of the Kohn–Sham model, without simplifying approximations.
The most natural geometrical framework for full-potential methodology is that
of space-filling polyhedral atomic cells, mathematically a Voronoi lattice construc-
tion in three dimensions. Space is subdivided by perpendicular bisector planes of
 7.1 Full-potential multiple scattering theory 95

the internuclear coordinate vectors. This lattice construction corresponds to the

Wigner–Seitz [429] cellular model. For open lattice structures (e.g. the diamond
lattice), it is common practice to introduce empty cells, with no enclosed atomic nu-
cleus, to provide a more uniform space-filling structure. Variational principles and
methods appropriate to such space-filling cellular lattices will be discussed here.
The muffin-tin model is a simplified special case in this general theory. A closely
related approximation, the atomic sphere model [384], extends the spherically aver-
aged local potential of the muffin-tin sphere throughout a concentric sphere whose
volume equals that of the Wigner–Seitz polyhedral atomic cell. This can be ex-
tended to the full enclosing sphere of each polyhedral atomic cell, which greatly
simplifies the computation of atomic basis functions required in full-potential MST.
These spherically averaged potentials can be considered as initial approximations
to be used during self-consistency iterations, to be modified following convergence
by nonspherical potential terms to be treated as perturbations in the context of linear
response theory. Nonlocal corrections to an initial local-potential model may also
be included in principle within such a linear-response theory.

7.1 Full-potential multiple scattering theory

Since the principal applications of MST are in solid-state theory, which traditionally
uses Rydberg rather than Hartree units of energy, this convention will be followed
in the present chapter. Restricting the formalism to local potential functions as in
Kohn–Sham theory, the Schrödinger equation for an electron of positive energy is
(∇ 2 + κ 2 )ψ(x) = v(x)ψ(x),
where the one-electron energy is defined by  = κ 2 (Ryd). The momentum κ is
a parameter that characterizes electronic energies above a zero value assigned to
the lowest energy of an energy band. Since this Schrödinger equation is formally
a modified Helmholtz equation, it is convenient to introduce a Helmholtz Green
function defined such that
(∇ 2 + κ 2 )G 0 (x, x ) = δ(x, x ).
The Green function must satisfy boundary conditions at large distances consis-
tent with the wave function ψ. The Schrödinger equation can be replaced by an
equivalent Lippmann–Schwinger integral equation


ψ(x) = χ (x) + G 0 (x, x )v(x )ψ(x ) d 3 x ,
3

where χ is a solution of the homogeneous Helmholtz equation. In an operator

notation this is ψ = χ + G 0 vψ. For a solid or molecule, coordinate space 3 can
96 7 Multiple scattering theory for molecules and solids

be subdivided into space-filling cells, each containing no more than one atomic
nucleus. A spherical polar coordinate system in each cell is centered about this
Coulombic singularity if it exists. The closed surface of cell τµ is denoted by σµ .

7.1.1 Definitions
µ µ
Solid-harmonic solutions JL and N L of the homogeneous Helmholtz equation in
cell µ are products of spherical Bessel functions and spherical harmonics. Specific
functional forms for the regular and irregular solid harmonics, respectively, are [188]
1
JL = κ 2 j (κr )Y,m (θ, φ),
1
N L = −κ 2 n  (κr )Y,m (θ, φ).

The standard angular momentum indices , m, and an implicit spin index, are here
summarized by a single collective index L. Relative normalization is such that
(N L |Wσ |JL  ) = δ L ,L  , expressed in terms of Wronskian integrals over surface σ
defined by


(φ1 |Wσ |φ2 ) = [φ1∗ ∇n φ2 − (∇n φ1 )∗ φ2 ] dσ.
σ

Here ∇n denotes the outward normal gradient on σ . It can easily be verified from
the defining differential equation that the Helmholtz Green function is given by the
expansion

G 0 (r, r ) = − JL (r)N L∗ (r ), r < r ,
L

in any specified atomic cell [188]. The definition of irregular solid harmonics N L
here establishes boundary conditions appropriate to the principal value Green func-
tion in scattering theory.

7.1.2 Two-center expansion

MST is based on expanding the total wave function in the local coordinate system
of each atomic cell. The Helmholtz Green function has an explicit expansion of
this sort. Given two distinct cells τµ and τν , with coordinate origins Xµ and Xν ,
respectively, global coordinates take the local form x = Xµ + r in τµ and x =
Xν + r in τν . The Green function is required in the local coordinates r, r . Now
µ
N L (x − Xµ ) is a regular solution of the Helmholtz equation in a sphere of radius
|Xµ − Xν | about the origin of cell ν = µ. This determines a local expansion of the
 7.1 Full-potential multiple scattering theory 97

form
µ
 νµ
N L (r + Xν − Xµ ) = − JLν (r )g L  ,L ,
L

for r  < |Xµ − Xν |. On substituting this into the one-center expansion,

 µ µν
G 0 (Xµ + r, Xν + r ) = JL (r)g L ,L  JLν∗ (r ),
L L

µν
for r, r  < |Xµ − Xν |. The Hermitian matrix g L ,L  defines structure constants for
the two-center expansion of the Green function in locally regular solid harmonics.
Through the Lippmann–Schwinger equation, matrix g propagates a given wave
function in one cell into incremental wave functions in the other cell. MST collects
the total incremental wave function in each cell as a multicenter expansion, to
establish a linear consistency condition for a molecule or solid. For a regular periodic
lattice, contributions from outside an elementary polyatomic translational cell are
multiplied by translational phase factors and summed over an infinite space-lattice.

7.1.3 Angular momentum representation

 ψ of the Schrödinger equation satisfies the homogeneous integral

An eigenfunction
equation ψ = 3 G 0 vψ. In cell τ this defines a locally regular solution of the
Helmholtz equation



χ = ψ − G 0 vψ = G 0 vψ,
τ 3 −τ

consistent with the Lippmann–Schwinger equation.

The function χ can be represented in τ by a sum of regular solid harmonic
functions,

χ (r) = JL (r)C L .
L

Similarly, because ψ is regular at the local coordinate origin, it can be represented

at this origin by

ψ(r) = JL (r)c L .
L

From the Lippmann–Schwinger equation and the expansion of G 0 in some spec-

µ
ified cell τµ , the coefficients in these expansions must be related by C L =
µ
c L + τµ N L∗ vψ. Consistency conditions are derived by considering the alternative
98 7 Multiple scattering theory for molecules and solids

expansion of χ in cell τµ ,


χ (r) = G 0 vψ.
ν=µ τν

From the two-center expansion of the Green function, this equation implies
   µν
χ (r) = − JL (r) g L ,L  SLν  ,
L L  ν=µ

where SLν = − τν JL∗ vψ. On equating both expressions for χ , and matching coef-
ficients in the regular expansion about the coordinate origin in cell τµ ,
µ
  µν
CL + g L ,L  SLν  = 0.
L  ν=µ

This derivation implies that these equations are valid under the condition that no
adjacent nucleus should lie within the enclosing sphere of a given local cell.
The wave function ψ can be expanded as a linear combination of basis functions
defined separately in each atomic cell. It is convenient to define specific basis func-
tions φ L (r) constructed by solving the local Schrödinger equation at a specified
orbital energy and matching to JL (r) at the cell origin. Given these primitive basis
functions, in one-to-one correspondence with locally defined regular solid harmon-

ics, the local expansion is ψ = L φ L γ L . If the local potential is nonspherical,
each basis function φ L becomes a sum of spherical-harmonic components as it is
integrated outward from the local cell origin. Nevertheless, it is completely char-
acterized by the single index L, and by a spin index that is assumed in the notation
and discussion here. In this basis,
  µν µν 
δ L ,L  C Lν  ,L  + g L ,L  SLν  ,L  γ Lν = 0,
ν,L  ,L 

where



C L ,L  = δ L ,L  + N L∗ vφ L  , SL ,L  = − JL∗ vφ L  . (7.1)
τ τ

If the matrix C is not singular, which requires the number of basis functions to
match the number of solid harmonics used to expand the Green function, a local
t-matrix is defined by t = −SC −1 . The consistency condition expressed above in
terms of C and S matrices then reduces to the simple matrix expression

(t −1 − g)Sγ = 0.

This is the fundamental equation of multiple scattering theory. It has the re-
markable property of concentrating effects of the local potential function into the
 7.1 Full-potential multiple scattering theory 99

t-matrix, separately determined in each local atomic cell, while the geometry of
the polyatomic superstructure is entirely characterized by the matrix g of structure
constants. For a regular periodic solid, the matrices here are summed over equiv-
alent translational cells, with coefficient phase factors that depend on an effective
momentum vector k, to give

{t −1 () − g(; k)}S()γ (; k) = 0.

The Korringa–Kohn–Rostoker (KKR) method [207, 204] is implemented by search-

ing for the zeroes of the secular determinant det(t −1 − g), or det(I − tg) if t is
singular.
The single-site matrix t corresponds to a scattering operator t̂ defined by
t̂χ L = vφ L for any primitive basis orbital φ L in a given atomic cell. Here

χ L = φ L − τ G 0 vφ L = L  JL  C L  ,L . From this definition it follows that t̂ JL =

v L  φ L  (C −1 ) L  ,L . The t-matrix is

(JL |t̂|JL  ) = (JL |v|φ L  )(C −1 ) L  ,L  .
L 

This implies that t = −SC −1 , since (JL |v|φ L  ) = −SL ,L  .

7.1.4 The surface matching theorem

Consider an atomic cell τ with closed boundary surface σ . Given a function ψ and its
normal gradient ∇n ψ on the surface σ , a unique solution of the Helmholtz equation
is determined by either inward or outward integration. Because this global function
is completely determined on σ , it in general will be singular at the coordinate
origin of cell τ and unbounded at infinite distance from this center. Inward integra-
tion in general gives a singular function, which includes irregular functions N .
For a function that is regular at the origin, either ψ (classical Dirichlet problem),
or ∇n ψ (Neumann problem) can be specified on σ , but not both [74, 255]. An
expansion valid on surface σ will be established here, determined by both the value
and normal gradient of ψ. Consistency of the surface Wronskian integrals implies
that coefficients of an expansion in both JL and N L must be given by

ψ(rσ ) ≡σ [JL (N L |Wσ |ψ) − N L (JL |Wσ |ψ)]
L

if the sums converge. The notation here denotes matching of both function value
and normal gradient. Even if the separate series in functions JL and N L do not
converge, it can be shown that well-defined functions exist, providing summation
formulas for the possibly divergent series. This justifies formal operations in MST
that depend on Wronskian integrals of this expansion.
100 7 Multiple scattering theory for molecules and solids

An interior function χ (r) = −(G 0 (r, rσ )|Wσ |ψ) is defined that is regular at the
cell origin. The one-center expansion of the Green function converges in the largest
sphere enclosed by surface σ , sphere S0 of radius r0 . This determines a series ex-

pansion χ(r) = L JL (N L |Wσ |ψ), valid for r < r0 . Similarly, an exterior func-
tion η(r) = (G 0 (r, rσ )|Wσ |ψ) is defined that is bounded at infinite separation.
Again using the convergent expansion of the Green function, but in the exterior

region outside an enclosing sphere S1 of radius r1 , η(r) = − L N L (JL |Wσ |ψ)
for r > r1 .
The auxiliary functions χ and η have finite values on σ because the surface area
element removes the singularity of the Green function in their defining surface inte-
grals. Both functions are defined throughout the region r0 ≤ r ≤ r1 . The boundary
conditions determine unique solutions of the Helmholtz equation for integration of
χ outward from S0 and of η inward from S1 . Similarly, the function ψ given on σ
can be extended throughout this region by integrating the Helmholtz equation both
inwards to S0 and outwards to S1 . Green’s theorem implies for such solutions of
the Helmholtz equation that Wronskian surface integrals are conserved on nested
closed surfaces. Hence the difference function ψ − χ − η has vanishing Wronskian
integrals with all regular and irregular solid harmonics on both spherical surfaces
S0 and S1 . It follows from Green’s theorem that this difference function must vanish
identically in the region between these two surfaces, and in particular on surface σ .
This establishes the surface matching theorem [280],

lim χ + lim η ≡σ ψ(rσ ).

r→rσ − r→rσ +

The series expansion of χ may diverge in the interior moon region [148], between
the enclosed sphere S0 and σ , while the corresponding series for η may diverge
between σ and S1 . Nonetheless, these functions are well defined by Wronskian
integrals of the Green function, independently of such expansions. The implied
surface integrals are properties of the integrated functions, and justify formal use
of the surface series expansion in boundary matching conditions based on these
Wronskian integrals. If applied to basis functions that are solutions of the same
Schrödinger equation in τ , for which (φi |Wσ |φ j ) = 0, surface expansion of either
function implies
 
(φi |Wσ |JL )(N L |Wσ |φ j ) − (φi |Wσ |N L )(JL |Wσ |φ j ) = 0.
L L

7.1.5 Surface integral formalism

Formalism from scattering theory, to be developed in a later chapter, shows that
a function ψ and its normal gradient on a closed surface σ can be matched to a
 7.1 Full-potential multiple scattering theory 101

linear combination of regular internal solutions of a Schrödinger equation if and

only if


ψ(rσ ) = (; rσ , rσ )∇n ψ(rσ ) dr σ .
σ

Here  is an Hermitian linear integral operator over σ that can be constructed

variationally from basis solutions {φi } of the Schrödinger equation that are regular
in the enclosed volume. The functions φi do not have to be defined outside the
enclosing surface and in fact must not be constrained by a fixed boundary condi-
tion on this surface [270]. This is equivalent to the Wronskian integral condition
(φi |Wσ |ψ) = 0, for all such φi . When applied to particular solutions of the form
ψ = J − N t on σ ,

(φi |Wσ |JL ) − (φi |Wσ |N L  )t L  ,L = 0.
L

This equation determines the local t-matrix. Any local basis function φi has the
formal expansion on σ

φi ≡σ [JL (N L |Wσ |φi ) − N L (JL |Wσ |φi )].
L

Canonical basis orbitals are defined here such that φ L → JL at the local cell origin.
Introducing a matrix notation for surface Wronskian integrals,
C L ,L  = (N L |Wσ |φ L  ), SL ,L  = −(JL |Wσ |φ L  ), (7.2)
the formal expansion of a primitive basis orbital φ L on σ is

φ L ≡σ (JL  C L  ,L + N L  SL  ,L ).
L

The surface Wronskian integrals C and S defined here are identical with the volume
integrals defined in Eqs. (7.1). Using (∇ 2 + κ 2 )φ L = vφ L , integration by parts of
the MST volume integrals, Eqs. (7.1), gives equivalent surface integrals, verifying
Eqs. (7.2),


SL ,L  = − JL∗ (∇ 2 + κ 2 )φ L  = −(JL |Wσ |φ L  ),
τ


C L ,L  = δ L ,L  + N L∗ (∇ 2 + κ 2 )φ L  = (N L |Wσ |φ L  ).
τ

7.1.6 Muffin-tin orbitals and atomic-cell orbitals

Deriving an energy-linearized version of MST, Andersen [12, 9] introduced muffin-
tin orbital (MTO) basis functions. These functions have the form φ − χ inside S0 ,
102 7 Multiple scattering theory for molecules and solids

the enclosed sphere of an atomic cell, and η outside. Here φ is a local solution of
the Schrödinger equation, while χ and η are sums of regular and irregular solid
harmonics, respectively. The matching on S0 is exact in the muffin-tin model,
so that these functions provide a basis set suitable for variational calculations.
In practice, because the long-range η functions are not orthogonal to inner shell
functions in displaced atoms [193], they are expanded in the energy derivatives φ̇ of
the basis functions φ. The resulting linearized muffin-tin orbital (LMTO) method
[13, 9, 384] solves the Schrödinger secular equation in an MTO basis, constructed
using KKR/MST structure constants.
The surface matching theorem makes it possible to generalize the idea of muffin-
tin orbitals to a nonspherical Wigner–Seitz cell τ . Each local basis orbital is rep-
resented as φ ≡σ χ + η on the cell surface σ , where χ and η are the auxiliary
functions defined by the surface matching theorem. An atomic-cell orbital (ACO)
is defined as the function φ − χ , regular inside τ . By construction, the smooth
continuation of this ACO outside τ is the function η. The specific functional
forms are
 
χL = JL  C L  ,L , ηL = N L  SL  ,L .
L L

Either MTO or ACO functions are valid as basis functions for expanding a global
wave function ψ in all atomic cells. By construction, they are regular in τ , smooth
at σ , and bounded outside. When the matrix C is nonsingular, modified canonical
basis functions can be defined such that

φ̂ L = φ L  (C −1 ) L  ,L ,
L

so that the auxiliary regular function is χ̂ L = JL . The corresponding ACO function

takes the especially simple form φ̂ L − JL , matched on the cell surface σ to η̂ L =

− L  N L  t L  ,L .

7.1.7 Tail cancellation and the global matching function

As discussed by Andersen [9, 10] for muffin-tin orbitals, the locally regular com-
ponents χ defined in each muffin-tin sphere are cancelled exactly if expansion
coefficients satisfy the MST equations (the tail-cancellation condition) [9, 384].
The standard MST equations for space-filling cells can be derived by shrinking the
interstitial volume to a honeycomb lattice surface that forms a common boundary
for all cells. The wave function and its normal gradient evaluated on this honeycomb
interface define a global matching function ξ (σ ).
 7.1 Full-potential multiple scattering theory 103

In any cell τ = τµ , with surface σ = σµ , any solution ψ of the Lippmann–

Schwinger
 equation defines auxiliary functions χ in = ψ − τ G 0 vψ and χ out =
3 −τ G 0 vψ, which are equal by construction. After integration by parts,
 µ µ 
χ in = −(G 0 |Wµ |ψ) = JL N L |Wµ |ψ .
L

The notation Wµ defines a Wronskian integral over σµ . By the surface match-

ing theorem, χ in = χ , the interior component of ψ. Since ψ is a solution of the
Lippmann–Schwinger equation, this implies χ out = χ in when evaluated in the inte-
rior of τµ . This is a particular statement of the tail-cancellation condition. To show
this in detail, after integration by parts
   µ µν  
χ out = (G 0 |Wν |ψ) = JL  g L  ,L JLν |Wν |ψ ,
ν=µ ν=µ L  ,L

which determines the standard MST equations.

 µ µν
On substituting the expansion N Lν = − L  JL  g L  ,L ,
  
χ out = − N Lν JLν |Wν |ψ .
ν=µ L
 µ µ
For comparison, η = − L N L (JL |Wµ |ψ)
is the exterior component of ψ on σ =
σµ . Since χ out ≡σ χ on σ and ξ ≡σ χ + η, this determines ξ everywhere on the
global matching surface,
 
ξ ≡σ ην = N Lν β Lν , (7.3)
ν ν L

where β Lν= −(JLν |Wν |ψ).If the sum here diverges, specific integrated functions
ην should be used.
All Wronskian integrals of ψ − ξ should vanish on the honeycomb lattice. Given
 µ µ µ
the local expansion ψ = L φ L γ L in any particular cell τµ , (JL |Wµ |ψ − ξ ) = 0
µ  µ µ µ
determines β L = L  SL ,L  γ L  . (N L |Wµ |ψ − ξ ) = 0 implies that
  µν µν 
δ L ,L  C Lν  ,L  + g L ,L  SLν  ,L  γ Lν = 0. (7.4)
ν,L  ,L 

These are the standard MST equations, in the form (t −1 − g)Sγ = 0, where
t = −SC −1 = (J |Wσ |φ)(N |Wσ |φ)−1 . (7.5)

7.1.8 Implementation of the theory

The computational survey of electronic structure of metals by Moruzzi et al. [256] is
a landmark example of the original KKR formalism, using the muffin-tin model and
104 7 Multiple scattering theory for molecules and solids

a root-search algorithm. The historical development of MST, from Huygens’ princi-

ple and the original formulation by Lord Rayleigh [337, 338] to recent applications
of full-potential theory, is surveyed by Gonis and Butler [148]. The full-potential
theory was first proposed by Morgan and Williams [431, 432]. Nonspherical cell
boundaries defined by the walls of polyhedral atomic cells were replaced by non-
spherical enclosing potentials. This methodology was simplified when it was rec-
ognized [39, 40] that viable basis orbitals could be computed within the enclosing
sphere S1 , and then matched after computation to a variational wave function at
the polyhedral cell boundaries. Ultimately, it was shown that these two apparently
disparate approaches necessarily produce the same orbital basis functions within
the enclosed atomic cell [277]. In the surface integral formalism developed here, it
is clear that the continuation of basis functions outside a local cell boundary cannot
affect the C and S matrices that characterize MST. This formalism also provides
a proof of the validity of MST for nonspherical atomic cells and potentials, which
was disputed over a decade [148].

7.2 Variational principles

7.2.1 Kohn–Rostoker variational principle
Kohn and Rostoker [204] (KR) derived the KKR method using variational theory
with a functional designed to produce the Lippmann–Schwinger equation. Local
basis solutions can be computed to high accuracy within each unit cell. The vari-
ational formalism optimizes the construction of linear combinations of these local
solutions that are compatible with the functions propagated from all adjacent cells.
In this version of MST, local solutions are coupled through the free-scattering
Green function of the Helmholtz equation. The Kohn–Rostoker variational func-
tional [204],




∗
= ψ v ψ− G 0 vψ
3 3

is stationary for infinitesimal variations of the one-electron wave function ψ if

and only if it satisfies the homogeneous Lippmann–Schwinger equation, ψ =
3 G 0 vψ. The Green function G 0 satisfies the inhomogeneous Helmholtz equation
at a specified energy  = κ 2 (Ryd),

(∇ 2 + κ 2 )G 0 (x, x ) = δ(x, x ).

The KR variational principle determines a wave function with correct boundary

conditions at a specified energy, the typical conditions of scattering theory. Energy
values are deduced from consistency conditions.
 7.2 Variational principles 105

The KR variational principle and the resulting KKR/MST formalism determine

wave functions for bound energy levels of a molecule or of a regular periodic
solid or two-dimensional periodic surface. In these applications, the homogeneous
Lippmann–Schwinger equation has no valid solutions except at a discrete set of
energy levels. Hence either κ must be adjusted to meet the conditions for such
a solution, or else the potential function must be altered to v(x) −  so that
the parameter  can be varied while retaining the same G 0 . Using the positive
parameter κ 2 to define G 0 already implies that energy values are measured relative to
the minimum of an energy band. Because it simplifies functional forms of the Green
function and solid harmonics, κ 2 = 0 is often used in LMTO calculations [384].
In a true scattering problem, an incident wave is specified, and scattered wave
components of ψ are varied. In MST or KKR  theory, the fixed term χ in the full
Lippmann–Schwinger equation, ψ = χ + G 0 ψ, is required to vanish. χ is a solu-
tion of the Helmholtz equation. In each local atomic cell τ of a space-filling cellular
model, any variation of ψ in the orbital Hilbert space induces an infinitesimal vari-
ation of the KR functional of the form δ = τ δψ ∗ v(ψ − 3 G 0 vψ) + hc. This
can be expressed in the form


δ = δψ ∗ v(χ in − χ out ) + hc,
τ

where, integrating the finite volume integrals by parts, and using the coefficient
matrices defined in MST,

 µ
χ =ψ−
in
G 0 vψ = JL C L ,
τµ L

 µν
χ out = G 0 vψ = − JL g L ,L  SLν  .
ν=µ τν ν=µ L ,L 

µ
In a finite basis of primitive functions φ L , the stationary condition is
 µ†
 µ†
 µν µν 
−δ = δγ L SL L 1 δ L 1 L 2 C Lν 2 L  + g L 1 L 2 SLν 2 L  γ Lν + hc = 0,
L ,L  L 1 ,L 2 ν

which implies the standard MST equations. Energy eigenvalues are determined by
adjusting the trial energy so that the secular determinant of the matrix of MST
equations is singular. Continuity conditions at cell boundaries are not addressed
by this variational principle. It is usually assumed that ψ lies in the orbital Hilbert
space (continuous function value and normal gradient at any cell interface). These
conditions are required for consistency with the Schrödinger equation, but are not
required by the KR variational principle and the MST equations.
106 7 Multiple scattering theory for molecules and solids

7.2.2 Convergence of internal sums

In the Kohn–Rostoker variational equation, the summation indices refer to two dif-
ferent classes of functions: orbital basis functions φ L , and solid harmonics JL for the
matrix of structure constants. For smooth matching at cell boundaries, the JL sums
should be carried to completion. In practice, finite sets of basis orbital functions and
of solid harmonics are used. The set φ L is truncated at some maximum orbital an-
gular momentum ¯φ in each inequivalent atomic cell, and the local expansion of the
Helmholtz Green function is limited to some maximum ¯g . In the muffin-tin model
[204], a spherically averaged local potential does not couple different spherical
harmonics in a single cell, but a propagated wave function in general has an infinite
expansion in other cells because of the Green function. The k-dependent Green
function for a periodic lattice structure introduces angular-momentum coupling
into the translational cell even if it is spherical.
Model studies of full-potential KKR theory were carried out for both two-
dimensional [51] and three-dimensional [52] space-lattices. The essential conclu-
sion is that the truncation parameters ¯φ and ¯g influence convergence in different
ways, and should not ordinarily be set to a common value . ¯ Adequately converged
results in general require essentially complete convergence of the geometrical ex-
pansion, which determines the Green function in the KR variational principle. In
contrast, high angular momentum values are less significant for the shell structure
of an atomic cell, considered to be a distorted atom. This can be understood be-
cause high orbital -values imply a centrifugal potential that excludes any orbital
wave function if the classical turning radius is greater than the enclosing radius of
the cell. Thus Rydberg orbitals in solids have no direct physical meaning if the
Rydberg radius extends into the valence shell of adjacent atoms. For these reasons,
the C and S matrices should be treated as rectangular column matrices, with the
row index ¯g significantly greater than the column index ¯φ [52].
Because rectangular matrices are singular, effective closure of these internal sums
over solid-harmonic indices requires a generalized definition of t = −SC −1 and
of other expressions that involve inverse matrices in MST [280]. As the index L
increases, the standard coefficient matrices S and C have very different behavior.
The effect of the centrifugal barrier for high orbital -values is to make the regular
solid harmonics JL very small on any finite cell boundary, while standard normal-
ization forces the irregular harmonics to be very large. In this Born-approximation
limit, the low-order matrix C is extended by a unit matrix, while S is extended
by a matrix of zeroes [281]. Thus the matrix t = −SC −1 is augmented by a null
matrix, while t −1 is augmented by elements of rapidly increasing magnitude, whose
computational numerical errors can overwhelm the computationally accurate low-
order elements. This can account for the strikingly different behavior found in
 7.2 Variational principles 107

empty-lattice calculations on the square-planar lattice using the KKR expression

t −1 − g [107] and using the formally equivalent expression I − tg [438]. The lat-
ter calculations demonstrated exponential convergence with increasing L indices,
while reproducing the much slower and less systematic convergence of the former
calculations when the inverse matrix t −1 was explicitly truncated.
This problem can be avoided by expressing the MST equations in terms of
the square matrix S † C = C † S, Hermitian in consequence of the surface-matching
theorem. This matrix has full rank because it is contracted over the larger index g .
From the definitions of the C and S matrices, the matrix product S † C is a specific
integral involving the Helmholtz Green function [281],



 µ† µ µ∗ µ µ
SL ,L  C L  ,L  = − φL v φL  − G 0 vφ L  ,
L  τµ τµ

which can be evaluated in closed form. The contracted matrix product gS can
also be evaluated in closed form by defining a generalized C-matrix, for displaced
µν µ µ
irregular solid harmonics N , C̃ L 1 ,L = (N L 1 |Wν |φ Lν ). The function N L 1 defined in
cell τµ is evaluated on the surface σν of a displaced cell. This is justified because
N -functions are valid outside their cell of definition. Using the Green function to
µν  µν  µν
propagate N µ , this becomes C̃ L 1 ,L = − L 2 g L 1 L 2 (JLν2 |Wν |φ Lν ) = L 2 g L 1 L 2 SLν 2 ,L .
In a matrix notation omitting L indices and summations, the KKR/MST equation is
 
S µ† (δ µν C ν + g µν S ν )γ ν = µν γ ν = 0,
ν ν

where all matrix products are contracted over the long dimension. In the Hermitian
µν †
matrix  the triple product S µ† g µν S ν = S µ† C̃ = (C̃ )µν S ν . Multiplying on the
†
left by S(S † C)−1 , and using (C̃ )µν = S µ† g µν , this takes the form β = 0, defining
a modified secular matrix
†
µν = δ µν + S µ (S µ† C µ )−1 (C̃ )µν .
The coefficients β = Sγ are the expansion coefficients of the global matching
 µ µ
function ξ (σ ) = µ,L N L (rσ )β L . Defining t = −S(S † C)−1 S † , which reduces to
t = −SC for square matrices [281], µν takes the standard MST form I − tg,
−1

without implying inversion of a rectangular matrix. Using these closed forms, the
coefficients β and γ are related by
 †
γ µ = −(S µ† C µ )−1 (C̃ )µν β ν .
ν=µ

In alloy theory, this equation determines the variational wave function for an atomic
cell τµ embedded in a statistical medium defined by the vector of coefficients β for
all other cells [157, 281].
108 7 Multiple scattering theory for molecules and solids

7.2.3 Schlosser–Marcus variational principle

Since a large truncation parameter ¯g implies large unsymmetrical C and S ma-
trices, and a very large matrix of structure constants, it might be more efficient to
find a variational condition directly applicable to the matching conditions for wave
functions on the common interface between adjacent atomic cells. Schlosser and
Marcus [359] (SM) extended the Schrödinger variational principle to the case of
basis functions that might be discontinuous or have discontinuous normal gradients
across atomic cell boundaries. It will be shown here that this variational principle
is equivalent to that of Kohn and Rostoker if the internal sums over solid harmonic
L-indices are carried to completion in the MST formalism. The SM variational
principle has been used by Ferreira and Leite [112, 113, 110] to formulate the vari-
ational cellular method (VCM), which eliminates the need for structure constants.
A matrix of Wronskian surface integrals over all shared cell facets is constructed.
Optimal matching conditions correspond to zeroes of the determinant of this matrix.
The Schlosser–Marcus variational principle is derived for a single surface σ that
subdivides coordinate space 3 into two subvolumes τin and τout . This generalizes
immediately to a model of space-filling atomic cells, enclosed for a molecule by an
external cell extending to infinity. The continuity conditions for the orbital Hilbert
space require ψ out ≡σ ψ in . This implies a vanishing Wronskian surface integral


(ψ |Wσ |ψ ) = − [(ψ in )∗ ∇n ψ out + (∇n ψ in )∗ ψ out ] = 0,
in out
σ

where σ is the enclosing boundary of τin . The sign of the first term here is reversed
because the notation Wσ implies an outward normal derivative in → out, while
by convention, ∇n ψ denotes the outward normal gradient across the boundary that
encloses the region of definition of ψ. The notation Wµ will be used for a Wronskian
integral with outward normal gradients over the full bounding surface of atomic cell
τµ . The SM variational functional adds a Wronskian surface integral Z σ = 12 (ψ in +
ψ out |Wσ |ψ in − ψ out ) to the Schrödinger functional Z τ = 3 ψ ∗ (H − )ψ. This
interface surface integral vanishes if ψ in ≡σ ψ out for independent trial functions
ψ in and ψ out . Then the global trial function ψ is in the usual Hilbert space and
the variational condition δ Z = 0 is just the Schrödinger condition for stationary
energy .
Schlosser and Marcus [359] showed that for variations about such a continuous
trial function, the induced first-order variations of Z τ and Z σ exactly cancel, even if
the orbital variations are discontinuous at σ or have discontinuous normal gradients.
After integration by parts, the variation of Z τ about an exact solution is a surface
Wronskian integral

δ Z τ = −(ψ|Wσ |δψ in − δψ out ).

 7.2 Variational principles 109

The variation of Z σ is

δ Z σ = (ψ|Wσ |δψ in − δψ out ),

which exactly cancels δ Z τ [359].

Because of this cancellation, variation about an arbitrary trial function gives
1
δ(Z τ + Z σ ) = (δψ|H − |ψ) + (δψ in + δψ out |Wσ |ψ in − ψ out ).
2
This can vanish only if (H − )ψ = 0 in both τin and τout . Moreover, this requires
that both (δψ in |Wσ |ψ in − ψ out ) and (δψ out |Wσ |ψ in − ψ out ) must vanish when ψ in
and ψ out are varied independently. By an extension of the surface matching theorem,
both these Wronskian integrals must vanish in order to eliminate the value and
normal gradient of ψ in − ψ out on σ . Practical applications of this formalism use
independent truncated orbital basis expansions in adjacent atomic cells, so that the
continuity conditions cannot generally be satisfied exactly.
In a space-filling cellular model, the SM variational functional can be expanded
in a local basis in each atomic cell. Variation of the expansion coefficients of the
 µ µ
trial orbital function ψ = L φ L γ L in cell τµ induces the variation
 µ†  µν
δZ = δγ L Z L L  γ Lν + hc.
L ,L  ν
µν
Matrix elements are Wronskian surface integrals on interface σµν between
Z L L
µ
adjacent cells τµ and τν , evaluated for basis functions φ L , φ Lν  . If the basis functions
in cell µ are all evaluated by integrating the local Schrödinger equation at the same
energy, the site-diagonal matrix elements Z µµ vanish. The interface elements are
1
Z µν = − (φ µ |Wµν |φ ν ), Z νµ = Z µν† ,
2
where Wµν implies normal gradients in the sense µ → ν on the cell interface σµν .
By this convention, matrix Z µν is Hermitian.
The variational cellular method (VCM) [113, 110] is an application of the
Schlosser–Marcus variational principle. The VCM variational matrix omits cell-
diagonal Wronskian integrals, which vanish by construction in a basis of orbital
functions all computed at the same energy. Because this matrix is Hermitian, its
null eigenvalues define solutions of the variational problem [113]. The SM sur-
face functional vanishes not only if ψ out and ψ in match in both value and normal
gradient at σ , but also if they are both equal and opposite. This implies that the
VCM determinant vanishes both for “true” solutions, giving an optimal solution of
the surface matching conditions, and for “false” solutions, giving a maximal mis-
match. This can be understood in a one-dimensional model, in which the Wronskian
110 7 Multiple scattering theory for molecules and solids

condition matches logarithmic derivatives across a boundary. The correct logarith-

mic derivative is obtained even if both function value and gradient have the wrong
sign. The VCM formalism avoids such false solutions by retaining only those zeroes
of the secular determinant for which the energy derivative is negative, a criterion
that selects the “true” solutions [113].
If cell-diagonal terms are retained, the VCM variational equations are
 µ 
φ L |Wµν |ψ µ − ψ ν = 0,
ν=µ

for each atomic cell surface σµ , summed over all adjacent cells. Although two
surface integral conditions are needed in general to eliminate both value and gradient
discontinuities, these equations can be justified because each interface σµν occurs
twice, for cells µ and ν, respectively. Expansion in a local orbital basis in each cell
gives

 µ  µ µ 
ν ν
φ L |Wµν | φL  γL  − φ L  γ L  = 0.
ν=µ L L

For an exact solution, the external function ψ ν would be identical to the global
matching function ξ on each interface. An alternative algorithm can be based
on fitting a linear combination of basis functions in each cell to ξ , which is
uniquely determined. The VCM equations for an orbital basis at fixed energy reduce
to [282]

µ
 µ µ  µ 
φ L |Wµ | φ L  γ L  − ξ = − φ L |Wµ |ξ = 0.
L

The global matching function determined by MST is


 µ µ  µ µ µ
  µν
ν
ξ= N L βL = N L β L − JL g L ,L  β L  ,
µ,L L ν=µ L 

expanded using coefficients

    
β Lν = − JLν |Wν |ξ = − JLν |Wν |φ Lν  γ Lν = SLν ,L  γ Lν .
L L

The variation of the SM functional in τµ then reduces to

 

2δ Z = δγ µ† (φ µ |Wµ |J µ ) (N µ |Wσ |φ µ )γ µ − g µν (J ν |Wν |φ ν )γ ν + hc,
ν=µ
 7.2 Variational principles 111

omitting L-indices and summations to simplify the notation. The first term is derived
using the general formula
 
(φi |Wσ |JL )(N L |Wσ |φ j ) = (φi |Wσ |N L )(JL |Wσ |φ j ),
L L

from the surface matching theorem. 2δ Z agrees with the KR variational expression
if intermediate sums over angular indices are complete [148]. Because the SM
variational principle implies stationary energy eigenvalues when the continuity
conditions are exactly satisfied, this is also true for the KKR/MST equations when
variational trial functions are continuous with continuous gradients, as for muffin-
tin and atomic-cell orbitals.

7.2.4 Elimination of false solutions

In a root-search procedure, false zeroes of the VCM secular determinant can easily
be identified by testing its energy derivative [113]. However, this is a major obstacle
to energy linearization of the method, following the logic of the LMTO method as
a linearization of KKR [384]. Two simplified variants of VCM have been proposed
that eliminate “false” solutions at the cost of abandoning a variational formalism
related to Schrödinger’s condition of stationary energy [287]. The equations of the
Green-function cellular method (GFCM) [53, 444, 282], or NVCM [287], are

 µ
 µ µ

N L |Wµν | φL  γL  − φ Lν  γ Lν = 0. (7.6)
ν=µ L L

An alternative formalism (JVCM [287]) substitutes the equations [443]


 µ
 µ µ 
JL |Wµν | φL  γL  − φ Lν  γ Lν = 0. (7.7)
ν=µ L L

A variational alternative, not yet implemented, is to augment the SM variational

functional by a Lagrange-multiplier term designed to enforce the matching condi-
tion at each atomic cell boundary. The proposed
 variational functional  supple-
ments the Schrödinger functional τ = 3 ψ ∗ (H − )ψ by a Wronskian integral
over the global matching surface,


1 in 1 out
σ = ψ + ψ − ξ |Wσ |ψ − ψ in out
.
2 2
ξ and its normal gradient, defined on the global matching surface, are both
Lagrange-multiplier fields, to be determined such that ψ in ≡σ ψ out . Comparison
with MST shows that if this condition is satisfied, ξ is the global matching function
112 7 Multiple scattering theory for molecules and solids

ξ ≡σ ψ out ≡σ ψ in . Then σ vanishes quadratically and  reduces to the Schrödinger

functional for trial functions ψ in the usual Hilbert space.
If σ is stationary with respect to free variations of ψ in , ψ out , and ξ ,


1 in 1 out
δσ = ψ + ψ − ξ |Wσ |δψ − δψ in out
2 2


1 in 1 out
+ δψ + δψ − δξ |Wσ |ψ − ψ in out
= 0.
2 2
This implies
1 in
ξ ≡σ (ψ + ψ out ); ψ in ≡σ ψ out ,
2
or ψ in ≡σ ψ out ≡σ ξ . Hence ξ is the global matching function, and a global trial
function ψ is defined in the usual Hilbert space. Variation of this function does not
produce any surface terms in δ. The implied practical procedure, as in the LMTO
method [384], is to use these variational conditions to construct trial functions as a
basis for the Schrödinger eigenvalue equations.
The variational equations imply ψ − ξ ≡σ 0 on each cell boundary σµ . Given
 µ µ
independent expansions ψ µ = L φ L γ L within each atomic cell, and ξ (σ ) =
 µ µ
µ,L N L (rσ )β L on the global matching surface, the coefficients are determined
by the implied variational equations. The surface matching theorem implies two
independent Wronskian integral conditions for each atomic cell,
 µ   µ µ µ
JL |Wµ |ψ µ − ξ = − SL L  γ L  + β L = 0,
L
 µ   µ µ
  µ 
N L |Wµ |ψ µ − ξ = C L L  γL  − N L |Wµ |N Lν 1 SLν 1 L  γ Lν = 0.
L L 1 ,L  ν=µ

The first condition determines the expansion coefficients β µ = S µ γ µ . On substi-

µ  µ µ µν µν
tuting (N L |Wµ |N Lν 1 ) = − L 2 (N L |Wµ |JL 2 )g L 2 L 1 = −g L L 1 , the second condition
reduces to the standard MST equations,
 µ µ   µν
C L L  γL  + g L L 1 SLν 1 L  γ Lν = 0.
L ν=µ L 1 ,L 

An attempt to fit ξ on the boundary of any cell τµ by a single expansion in

the local primitive basis functions φ L fails if these functions all satisfy the same
Schrödinger equation at a single energy. The problem is that such basis functions all
have vanishing Wronskian surface integrals with each other, and cannot match both
value and normal gradient of an arbitrary function on the surface σµ . In MST the
set of irregular solid harmonics N µ from all cells is not limited by this constraint,
and thus provides a suitable basis for ξ . In contrast, representation as a sum of local
primitive basis orbitals in a single cell leads to the problem of “false” solutions of
 7.3 Energy-linearized methods 113

the SM variational equations. The MST expansion of ξ is suggestive that, on any

interface segment σµν , ξ must be represented by two distinct terms, one from the
indexed cell µ and the other from cell ν. In general this ensures that the number of
matching conditions (value and normal gradient for each basis function) equals the
number of independent basis functions.

7.3 Energy-linearized methods

Because of the nonlinear energy dependence of the KKR secular equation, roots
of the secular determinant must be computed by a root-search procedure, requir-
ing repeated evaluation of the secular matrix for each root. Despite the inherent
inefficiency of this procedure, simplifications implicit in the KKR–DFT theory for
muffin-tin potential functions can be exploited. This methodology was used [256]
to compute self-consistent energy-band structures for 32 elemental metals, ranging
from H to In in the periodic table, including spin-polarized structures for the tran-
sition metals. This reference provides a bibliography giving technical details of the
analytical and computational methods used. In order to carry out self-consistent
calculations on more complex materials with polyatomic translational cells, or to
examine effects of lattice distortion, it is desirable to convert this formalism into a
variational form in which energy eigenvalues can be computed from a linearized
secular equation.

7.3.1 The LMTO method

Starting from the idea of “muffin-tin orbitals” (MTO) [7], the linear muffin-tin
orbital method (LMTO), developed by Andersen and collaborators, became one of
the most widely used computational methods in energy-band theory. Publications
on this method include a general description, with source listings of computer
programs [384], and a review of methodology and applications [11]. The LMTO
method differs from the KKR muffin-tin model of multiple scattering theory in three
important aspects. In addition to conversion of the theory to an energy-linearized
form, the method also replaces the original muffin-tin model by an atomic-sphere
approximation (ASA), in which the local atomic potential function is extended
from the enclosed muffin-tin sphere S0 to a larger “atomic” sphere, defined to have
the same volume as the Wigner–Seitz atomic cell. The third element of difference
from KKR is that structure constants are taken to be independent of energy. This is
done by subtracting an energy shift from both sides of the Schrödinger equation,
   
∇ 2 + κ02 ψ(x) = v(x) − κ 2 + κ02 ψ(x). (7.8)
114 7 Multiple scattering theory for molecules and solids

The Green function G 0 and solid-harmonic functions JL and N L are required only
for the fixed energy κ02 , usually taken to be zero. This amounts to adding a linear
energy term to the potential function in the interstitial region, justified in full-
potential theory or the ASA because the net interstitial volume is reduced to zero.
The practical effect is slower convergence of the angular momentum expansions. Al-
though the ASA violates the original KKR conditions that ensure an exact solution
of the muffin-tin model problem, the ASA model of overlapping atomic spheres is
in many ways a more satisfactory approximation to full-potential MST. In particu-
lar, in the surface integral formalism, approximating Wronskian integrals over the
surface of a Wigner–Seitz cell by the corresponding integrals over an equivalent-
volume sphere can be considered as a first approximation to a surface quadrature
scheme for polyhedral cells. The full DFT local potential function can be used within
the atomic sphere. The muffin-tin model, which assumes a constant potential in the
near-field region, has no special justification as a computational approximation for
atomic potentials.
In the usual ASA, the t-matrix is real and diagonal, but the elements have both
zeroes and poles as functions of energy. The original proposal was to simplify KKR
calculations by using constant-energy structure constants while fitting diagonal ele-
ments of t −1 in the standard secular equation, (t −1 − g)β = 0, to rational functions
of energy [8]. This also simplifies the energy dependence of muffin-tin orbitals,
which can be used as basis functions in the standard Schrödinger variational prin-
ciple for the energy eigenvalues. The resulting LMTO formalism (called LCMTO
in the original presentation) produces a linear eigenvalue problem that includes
a “combined correction” matrix representing the difference of the overlap matrix
evaluated in atomic spheres versus atomic polyhedra together with a correction for
basis functions of higher  quantum numbers. LMTO energy-band calculations ob-
tain all energy eigenvalues at a given k-point from a single matrix diagonalization.
These eigenvalues are valid within an energy panel in which the rational fit to the
t −1 -matrix is sufficiently accurate.
In considering nonspherical local potentials, a simple rational fit to diagonal ele-
ments of t −1 is no longer valid. From the general definition of the t-matrix as −SC −1
in terms of the standard MST matrices, it is clear that for energy-independent solid-
harmonic functions, linear energy expansion of the local basis functions φ L (; r) is
equivalent to fitting elements of the t-matrix by a simple rational formula. In the
ASA this produces the parametrization proposed by Andersen. It was recognized at
the same time [13] that a linear energy expansion of the basis functions also solves

the practical problem in MTO theory that in the local representation χ = φ − JC
of primitive MTOs the regular solid-harmonic functions are not orthogonal to inner-
shell occupied orbitals. The practical solution is to replace the solid harmonics by
linear combinations of the energy-derivative functions denoted here by φ̇ L . In the
 7.3 Energy-linearized methods 115

standard LMTO method, the basis functions are energy-independent MTOs ex-
pressed as linear combinations of functions φ L and φ̇ L computed at a fixed panel
energy [9, 384]. Basis functions of this kind are used in alternative energy-linearized
methods [9, 430] which will not be considered here.

7.3.2 The LACO method

The linearized atomic-cell orbital (LACO) method is a full-potential analog of
LMTO [274]. Energy-independent structure constants are computed using the
LMTO program STR [384]. On a grid of energy values suitable for accurate interpo-
lation, basis functions are computed by numerical integration within the enclosing
sphere of each atomic cell. The basis functions and their normal gradients are inter-
polated to atomic-cell boundaries so that the standard MST C and S matrices can
be evaluated by two-dimensional quadrature over the surfaces of polyhedral atomic
Wigner–Seitz cells. Energy linearization is accomplished in two stages, each with
several options as to the specific procedure. In the first stage, energy-independent

atomic-cell orbitals are constructed in the form φ + φ̇ω. Here, φ is a single
µ
primitive basis function φ L in a particular cell τµ . In this cell, the defining ACO

form φ − J C is modified by representing the regular solid-harmonic functions
J as linear combinations of energy-derivative functions φ̇ L . In all other cells, the

matched external function N S is expanded in the local φ̇ L basis. The coefficient
matrix ω L ,L  satisfies a set of inhomogeneous linear equations, obtained from one
of the linear forms of full-potential MST equations, in the extended fixed-energy
basis of functions φ L and φ̇ L . In the second stage of LACO calculations, these
energy-independent ACO functions are used in the standard Rayleigh–Schrödinger
or Schlosser–Marcus variational equations to determine energy eigenvalues.
In an energy range or panel sufficiently narrow that linear energy interpolation
of the basis functions is accurate, it can be shown that this general linearization
procedure produces eigenfunctions identical to the null vectors of the MST secu-
lar equations. To prove this for basis functions all at the same energy 0 , consider
right-eigenvectors of the complex, unsymmetric matrix ω. The eigenvalue equation
ωck = ck k defines a displacement  from the fixed panel energy. Each eigen-
vector of the ω-matrix determines a linear combination of energy-independent
ACO functions which, by construction, is an optimized fit to the orbital continuity
conditions. This function is
    
ψk = φ+ φ̇ω ck = (φ + φ̇k )ck = φ(k )ck . (7.9)

This agrees with the form of an MST null-vector expanded in the energy-dependent
basis at k = 0 + k . Because of numerical approximations inherent in practi-
cal calculations, it is generally inconvenient to diagonalize matrix ω directly. The
116 7 Multiple scattering theory for molecules and solids

eigenvalues are complex numbers if there are any residual numerical errors. For
this reason, a standard variational principle based on Hermitian matrices is used
in the second stage of linearized calculations. The φ, φ̇ expansion must be used
with caution unless all elements of the ω-matrix are small. A large energy shift 
may imply that basis functions fall outside the energy panel used in a particular
calculation, so that the interpolated functions are not accurate. So-called “ghost”
bands can occur, especially for higher angular quantum numbers, when functions
φ and φ̇ tend to become linearly dependent within a local cell [11].

7.3.3 Variational theory of linearized methods

The Kohn–Rostoker variational principle [204] implies variational equations in the
extended basis {φ, φ̇}. Each energy-independent ACO basis function φ̃ λ is defined
by one fixed function φ λ modified by a sum of φ̇ µ functions with coefficients ωµλ .
Suppressing L-indices, the KKR/MST equations indexed by φ̃ κ in cell τµ for the
coefficients in φ̃ λ are

(S µ δ µκ + Ṡ µ ωµκ )† [δ µν (C ν δ νλ + Ċ ν ωνλ ) + g µν (S ν δ νλ + Ṡ ν ωνλ )] = 0. (7.10)
ν

Matrices C, S, Ċ and Ṡ here are to be considered as rectangular matrices. The

internal sums over solid-harmonic L-indices should be carried to convergence. The
L  , L indices of the square matrix ω are basis function indices and may have a
smaller range.
If the KKR functional  were treated as a functional of the coefficient matrix
ω, the derived variational equations would be a set of linear equations of the form

Ṡ µ† ν [· · ·] = 0, where the bracketed term is the same as in Eqs. (7.10). The
solution of these simplified equations for a given value of λ, L and all values of
µ, L  is a column vector of the ω-matrix. These simplified equations were tested
by empty-lattice calculations on an fcc space-lattice [280].
The Schlosser–Marcus variational principle [359] provides an alternative that
does not use structure constants. On substituting the expansion of an energy-
independent ACO into the SM variational functional, the variational equations
indexed by φ̃ κ in cell τµ are

(φ µ δ µκ + φ̇ µ ωµκ )|Wµν |(φ µ δ µλ + φ̇ µ ωµλ ) − (φ ν δ νλ + φ̇ ν ωνλ ) = 0, (7.11)
ν=µ

where L-indices and sums are suppressed to simplify the notation. In these equa-
tions, the internal summations in Eqs. (7.10) are replaced by direct matching
across adjacent cell interfaces. If the generally complex unsymmetric matrix ω
µ
is diagonalized, with eigenvector elements c Lk , Eqs. (7.11) reduce to the VCM
 7.3 Energy-linearized methods 117
 µ µ
equations for ψ µ = L φ L (k )c Lk ,
 µ 
φ L (k )|Wµν |ψ µ − ψ ν = 0,
ν=µ

whose null vectors determine equivalent linear combinations of ACO functions or

of atomic basis functions at energy values displaced by k . In this representa-
tion, the LACO method is an energy-linearized version of the variational cellular
method (VCM) [112]. However, Eqs. (7.11) are quadratic in the matrix ω, which
is nonhermitian in general and complex for periodic lattice structures, due to trans-
lational phase factors. The problem of “false” solutions, inherent in the SM/VCM
variational principle, remains in these quadratic matrix equations, for which no
computationally viable solution algorithm has been found. Approximate linear
equations are obtained if the left-hand element is simplified to either just φ µ or φ̇ µ ,
as done in tests [280] of the corresponding KKR/MST equations. As in the case
of Eqs. (7.10), each solution of the simplified equations for given λ, L is a column
vector of the ω-matrix. The two alternative sets of simplified equations cannot be
satisfied simultaneously.
These difficulties with the linearized VCM can be resolved by using the extended
functional, defined over the global matching surface σ in Subsection 7.2.4, above,


1 in 1 out
σ = ψ + ψ − ξ |Wσ |ψ − ψ in out
.
2 2
Following LMTO theory, an energy-independent ACO function φ̃ λ based on atomic
cell τλ is represented in all cells τµ by the form φ µ δ µλ + φ̇ µ ωµλ . The Lagrange-
multiplier field ξ , which becomes the global matching function in the variational
equations, must be expanded in a basis of functions that can be extended to complete-
ness on the global matching surface. In MST, the expansion on each interface sector
 µ
σµν of the matching surface consists of two terms. One is the local sum, L N L , and
the other is the corresponding sum over irregular solid harmonics extended to σµν
from all other cells. Both sums are needed in order to match both value and normal
gradient of the wave function on the local cell surface. Because the primitive basis
µ
functions φ L are strictly truncated on σµ , the surface of cell τµ , the energy-derivative
functions φ̇ µ should be defined for consistency as identically zero outside σµ . By
implication, replacing solid-harmonic functions by the local basis {φ̇} in each cell,
the matching function should be expanded on each σµν as ξ = φ̇ µ β µλ + φ̇ ν β νλ .
The Wronskian integrals of ψ µ − ξ that must vanish for each cell τµ are

(φ µ |Wµν |(φ µ δ µλ + φ̇ µ ωµλ ) − (φ̇ µ β µλ + φ̇ ν β νλ )) = 0,
ν=µ
 (7.12)
(φ̇ µ |Wµν |(φ µ δ µλ + φ̇ µ ωµλ ) − (φ̇ µ β µλ + φ̇ ν β νλ )) = 0.
ν=µ
118 7 Multiple scattering theory for molecules and solids
µ
If the total number of basis functions φ L for all cells is Nφ , then for each global
solution index λ there are 2Nφ equations for the 2Nφ elements of the column vectors
ωµλ and β µλ . Thus the variational equations derived from σ provide exactly the
number of inhomogeneous linear equations needed to determine the two coefficient
matrices, ω and β. These equations have not yet been implemented, but they
promise to provide an internally consistent energy-linearized full-potential MST.

7.4 The Poisson equation

For a solid or molecule subdivided into atomic cells, the Poisson equation, given
in Rydberg units by

∇ 2 v = −8πρ, (7.13)

can be solved by full-potential multiple scattering theory methods. Here ρ is the

local number density of electrons, and the negative electronic charge is factored
out of both charge density and potential. Adopting the classical theory of inhomo-
geneous linear differential equations to a cellular model, particular solutions are
first obtained for each atomic cell and then modified by the addition of regular
solutions of the homogeneous equation to satisfy continuity and boundary condi-
tions. The original full-potential MST [432] was modified for the Poisson equation
and applied to a periodic charge distribution on an fcc space-lattice [247]. In this
example all atomic cells are geometrically equivalent, so the actual calculation is
for a single cell with periodic boundary conditions. The given charge distribution,
with Heaviside cutoff factors at the polyhedral cell boundary, was expanded in
spherical harmonics at radii up to that of the enclosing sphere. A local particu-
lar solution v0 was obtained by integrating equations analogous to the KKR/MST
equations in spherical polar coordinates, subject to the boundary condition that the
particular solution should vanish at large r . The external continuation of v0 is a
sum of multipole potentials. Within a particular cell, the sum of these potentials
due to all other cells defines a local potential v, expressed in terms of structure
constants that are just the coefficients in the two-center expansion of the Coulomb
interaction. v is added to v0 to give the required global solution of the Poisson
equation. The surface integral MST method described below has been applied to the
Poisson equation, and is incorporated in the LACO program package [278], still
using structure constants. More recently, MST cellular methods, using direct match-
ing theory with NVCM (GFCM) and JVCM equations rather than structure con-
stants, were applied to this problem in a detailed numerical study [443]. This study
finds that the JVCM is the most accurate and reliable of the methods considered,
which included standard MST with structure constants.
 7.4 The Poisson equation 119

Multiple scattering theory treats the Poisson equation as an inhomogeneous

Helmholtz equation of zero energy (Laplace equation). The Green function G 0
is proportional to the Coulomb potential. In Rydberg units, 8π G 0 = −2/r12 . A
particular solution v0 and complementary solution v within a local atomic cell τ
are defined by subdividing the Lippmann–Schwinger equation such that



v = v0 + v = −8π G 0 ρ − 8π G 0 ρ. (7.14)
τ R3 −τ

The choice of solid-harmonic functions

JL = r  Y L /(2 + 1); N L = r −−1 Y L

gives the well-known one-center expansion of 1/r12 [188]. It is assumed that the
local density function is subdivided into spherical harmonic components as ρ =

L ρ L . Each of these components defines a particular solution within the enclosing
sphere of the local cell. A very efficient numerical algorithm is available for solving
the radial Poisson equation [229], generalized to arbitrary -values [278]. Denoting
these primitive solutions by v̂ L , the ACO construction can be used to define a

particular solution v0L for each ρ L such that v0L = v̂ L − J ĉ in the local cell τ .

v0L matches onto an external function N ŝ on the cell surface σ . The coefficient
matrices here are defined by

ĉ L  ,L = (N L  |Wσ |v̂ L ), ŝ L  ,L = −(JL  |Wσ |v̂ L ). (7.15)

The complementary potential v or generalized Madelung term is expanded in

regular solid harmonics as
µ
 µ µ
v L = − JL  c L  ,L . (7.16)
L

The coefficients here are

  
cµ = − g µν J ν |Wν |v0ν , (7.17)
ν

obtained by substituting the two-center expansion of N µ into the second term in


Eq. (7.14), or, equivalently, into N ŝ.
Alternative equations for the coefficients c are given by direct matching at
cellular interfaces, bypassing the need for structure constants. The two alternatives
considered by [443] are the NVCM or GFCM equations,
  
(N µ |Wµν |J ν )cν − cµ = N µ |Wµν |v0ν , (7.18)
ν=µ ν=µ
120 7 Multiple scattering theory for molecules and solids

and the JVCM equations,

    µ
(J µ |Wµν |J ν )cν = J µ |Wµν |v0ν − J µ |Wµ |v0 . (7.19)
ν=µ ν=µ

Exploiting the high efficiency of the integration algorithm for the radial Poisson
equation [229], MST surface integral formalism [274, 278] reduces the computation
of multipole moments of cellular charge densities to the evaluation of Wronskian
surface integrals. Electrostatic multipole moments Q L are defined by the asymp-
totic external potential for each cell τ . For the particular choice of solid-harmonic
functions given above,
 
lim r →∞ v0L = N L  ŝ L  ,L = −8π NL Q L . (7.20)
L

A factor −2 included in the last term here compensates for the use of Rydberg units
and for the omission of the negative electronic charge in potential functions derived
from Eq. (7.14). Hence the electrostatic multipole moments of atomic cell τµ are

µ 1  µ 1 µ
 µ
QL = − ŝ  = JL |Wσµ | v̂ L  . (7.21)
8π L  L ,L 8π L

This equation is used for  = 0 to evaluate normalization integrals in the LACO

program package [278], avoiding numerical volume quadrature.

7.5 Green functions

Although electronic structure calculations are usually concerned with finding the
eigenfunctions ψi (x) and eigenvalues i of the Schrödinger equation, the corre-
sponding Green function is required for many classes of applications. Problems
dealing with impurities, transport, disordered systems, photoemission, etc. are most
naturally solved using the Green function. In principle, the Green function can be
constructed from eigenfunctions, using the spectral representation,

G(x, x ) = ψi (x)( − i )−1 ψi∗ (x ).
i

This representation converges very slowly in general, and an alternative method

is needed. Direct methods for computing the Green function are considered here,
using boundary-matching methods of multiple scattering theory. An important
consideration, in the context of extensions and applications of density functional
theory (DFT), is that any external perturbing potential will in general be modified or
screened by a self-consistent response. Hence any incremental potential considered
here is assumed to be modified or screened in the sense of linear-response theory
[79, 292].
 7.5 Green functions 121

7.5.1 Definitions
For energy κ in Rydberg units, the Green function of the Helmholtz equation, with
2

given boundary conditions, satisfies

(∇ 2 + κ 2 )G 0 (x, x ) = δ(x, x ), (7.22)

while the corresponding Schrödinger Green function satisfies

(∇ 2 + κ 2 − v(x))G(x, x ) = δ(x, x ). (7.23)

These two Green functions are related by the Lippmann–Schwinger integral

equation


 
G(x, x ) = G 0 (x, x ) + G 0 (x, y)v(y)G(y, x )d 3 y. (7.24)
3

The wave function of a perturbed system defined by v(x) is

ψ̃(x) = ψ(x) + G(x, x )v(x )ψ̃(x )d 3 x . (7.25)
3

This equation can be applied to a wide class of problems including the electronic
structure of vacancies, impurities, and of other localized perturbations of solids and
atomic clusters. The methods considered here make it possible to construct G(x, x )
for any potential function v(x) defined throughout the coordinate space 3 .
It is assumed that 3 is subdivided into space-filling cells. A typical cell τµ has
closed surface σµ . An external empty cell can be included to fill space for a finite
system. Solutions of the homogeneous Helmholtz equation about the origin of cell µ
define regular and irregular generalized solid-harmonic functions, denoted here by
µ µ
JL and N L , respectively. Each generalized solid-harmonic function is the product
of a radial factor and a spherical-harmonic function, defined here by a combined
index L that denotes the quantum numbers (m). Normalization is assumed to be
consistent with the Kronecker-delta relation,

(N L |Wσ |JL  ) = δ L ,L  . (7.26)

Expressed in local coordinates for each atomic cell, the Helmholtz Green function
can be expanded in solid-harmonic functions,

G 0 (r, r ) = − JL (r)N L∗ (r ), r < r . (7.27)
L

To verify this expansion, consider G 0 (r, r ) for a fixed point r . For r < r  , this
function is a regular solution of the Helmholtz equation and must have an expansion
122 7 Multiple scattering theory for molecules and solids

of the form

G 0 (r, r ) = JL (r)(N L |W0 |G 0 ), (7.28)
L

where the Wronskian integral is over the surface of any sphere centered at the cell
origin with radius r0 < r  . For r > r  , this function is a bounded solution of the
Helmholtz equation and must have an expansion of the form

G 0 (r, r ) = − N L (r)(JL |W1 |G 0 ), (7.29)
L

where the Wronskian integral is over the surface of any sphere centered at the
cell origin with radius r1 > r  . If Eq. (7.22) is multiplied by N L∗ (r), and the cor-
responding null term with the Helmholtz operator applied to N L∗ is subtracted, on
integrating over the volume defined by r0 < r  < r1 , between radii r0 and r1 , this
integral reduces to Wronskian surface integrals. This implies

N L∗ (r ) = (N L |W1 |G 0 ) − (N L |W0 |G 0 ). (7.30)

Since r1 > r  , Eq. (7.29) implies that the first integral on the right-hand side van-
ishes. The second integral gives the coefficient in Eq. (7.28), which reduces to
Eq. (7.27). Similarly, if Eq. (7.22) is multiplied by JL∗ (r) and integrated over the
volume between radii r0 and r1 ,

JL∗ (r ) = (JL |W1 |G 0 ) − (JL |W0 |G 0 ). (7.31)

In this case, the second integral on the right-hand side vanishes and the first integral
gives the coefficient in Eq. (7.29). This provides a formula valid for r > r  , com-
plementary to Eq. (7.27). Comparison of these two formulas shows that the linear
operator G 0 is Hermitian for real energies.
When coordinate x = Xµ + r lies in cell τµ and x = Xν + r lies in a different
µ
cell τν , a two-center expansion of the Green function is needed. Since N L (x − Xµ )
is a regular solution of the Helmholtz equation in a sphere of radius |Xµ − Xν |
about the origin of a displaced cell ν,
µ µ
 νµ
N L (r + Xν − Xµ ) = N L (r )δµν − JLν (r )g L  ,L , (7.32)
L
νµ
for r  < |Xµ − Xν |. The expansion coefficients g L  ,L are structure constants at the
given energy. This matrix is Hermitian for real energies. Using the above expansion,
the Helmholtz Green function can be written in the form,
 µ µν
 µ
G 0 (Xµ + r, Xν + r ) = JL (r)g L L  JLν∗ (r ) − JL (r)N Lν∗ (r )δµν , (7.33)
L L L
 7.5 Green functions 123

where r < r  within any single cell µ. The two-center expansion converges in gen-
eral for muffin-tin geometry, when the two coordinate values lie within nonover-
lapping spheres. It is assumed here that cell coordinate origins are chosen to satisfy
this geometric condition for some neighborhood of each cell origin.
These definitions can be generalized to give an expansion of the Green function
G(x, x ) in terms of local solutions of the Schrödinger equation in each cell τ . Local
regular solutions φ L and local irregular solutions ζ L can be defined by matching to
functions JL and N L , respectively, on an infinitesimal sphere about the cell origin,
or, alternatively, by requiring similar conditions on a local cell boundary σ . This
construction imposes the Wronskian condition

(ζ L |Wσ |φ L  ) = δ L ,L  . (7.34)

These integrals vanish between any two φ or two ζ functions. Green’s theorem
implies that this Kronecker-delta formula is valid for Wronskian integrals over
any closed surface that encloses the local cell center and excludes all other singular
points of the potential function. If singularities of the potential function occur only at
cell centers, these functions can be assumed to be have regular extensions out to the
nearest neighbor cell center, but in general no further. For this reason, the primitive
function ζ must be modified by the addition of a sum of regular functions φ in its
cell of origin to define a modified irregular function ζ̃ that is regular everywhere
except at the defining cell center. With these definitions [53, 441, 444], by analogy
to the Helmholtz Green function,

G(r, r ) = − φ L (r)ζ̃ ∗L (r ), r < r , (7.35)
L
µ µ 
 νµ
ζ̃ L (r + Xν − Xµ ) = ζ L (r )δµν − φ Lν  (r )G L  ,L , (7.36)
L

for r  < |Xµ − Xν |, and

 µ µν
 µ
G(Xµ + r, Xν + r ) = φ L (r)G L ,L  φ Lν∗ (r ) − φ L (r)ζ Lν∗ (r )δµν , (7.37)
L L L

valid for r < r  within a single cell, or for coordinates in two nonoverlapping
µν
spheres. The coefficients G L ,L  define the structural matrix of the Schrödinger Green
function. For ν = µ, ζ̃ µ − ζ µ is a sum of regular functions φ µ with coefficients
µµ
−G L ,L  .
The definition of irregular functions ζ depends on boundary conditions imposed
on the Green function. Any variation of the set of functions ζ defined by adding
linear combinations of the regular functions φ with coefficients that constitute an
Hermitian matrix simply moves the corresponding term between the two summa-
tions in Eq. (7.37), leaving the net sum invariant, and preserving the Kronecker-delta
124 7 Multiple scattering theory for molecules and solids

conditions of Eq. (7.34). The present derivations are covariant with respect to the
group of such transformations. For particular applications, the representation can be
chosen for the greatest convenience, subject only to the Kronecker-delta condition.

7.5.2 Properties of the Green function

The definition of the Schrödinger Green function can be extended to a complex-
valued energy parameter z. Then Eq. (7.23) for the Schrödinger equation can be
written as

(z − H(x))G(z; x, x ) = δ(x, x ), (7.38)

which defines G(z) = (z − H)−1 as the resolvent operator. G(z) is an analytic op-
erator function in the upper half of the complex z-plane. The singularities are poles
or a continuum branch cut corresponding to energy eigenvalues. If H is Hermitian,
the eigenvalues lie on the real energy axis. The corresponding orthonormal set of
eigenfunctions ψi defines a spectral representation

G(z; x, x ) = ψi (x)(z − i )−1 ψi∗ (x ). (7.39)
i

This formula indicates that the residue of G(z) at an eigenvalue pole is just the
density matrix, in the coordinate representation, summed over all eigenfunctions
with this eigenvalue. This residue can be extracted conveniently by using Dirac’s
formula after displacing each eigenvalue into the lower complex plane,
1 1
lim δ→0 =P − iπ δ(z − i ). (7.40)
z − i + iδ z − i
Specializing to the electronic density for electrons of one spin in the local coordi-
nates of a particular cell, the local density per unit energy is
1
n(; r) = − G(; r, r). (7.41)
π
Summing over all eigenenergies below a chemical potential or Fermi level µ, the
local density function is defined by a contour integral passing above all poles on
the real axis,


1 µ
ρ(r) = − G(z; r, r) dz. (7.42)
π −∞
In methodology based on computation of the Green function, this formula re-
places the usual sum over eigenfunctions. When integrated over coordinate space,
 7.5 Green functions 125

Eq. (7.41) gives the density of states per unit energy,

1
n() = − G(; r, r) dτ. (7.43)
π R3

These formulas must be summed over a spin index to give the corresponding total
densities. In multiple scattering theory, the Green function as given by Eq. (7.37) is
subdivided into terms valid in separate atomic cells. The elements with r = r are
single sums over the cells. Hence the density of states of given spin in a particular
cell τµ is


µ 1
n () = − G µ (; r, r), (7.44)
π τµ

where G µ denotes the terms with ν = µ in Eq. (7.37).

7.5.3 Construction of the Green function

µν
Equation (7.24) suggests that the structural matrix G L ,L  of the Schrödinger Green
function at specified energy should have a simple relationship to that of the
µν
Helmholtz Green function, which is the matrix of structure constants g L ,L  . The
basis set of regular local solutions of the Schrödinger equation can be constructed
so that matrix G depends only on g and the t-matrix. In deriving the KKR/MST
equations, primitive basis functions were defined by outward integration, starting
from specified regular solid-harmonic functions JL at the origin in each atomic
cell. A canonical basis set is defined by transforming the primitive basis by ma-
trix C −1 , where C L ,L  is defined as in MST. Hence, in the canonical basis, C is
a unit matrix. The matching equations on the local cell boundary depend only on
the t-matrix, because each basis function matches to a boundary function J − N t.
These basis functions, still indexed by L, are in one-to-one correspondence with
the regular solid-harmonic functions JL . Corresponding to the canonical regular
basis function φ L , which matches onto J − N t on cell surface σ , a paired canoni-
cal irregular function ζ L can be computed by integrating the Schrödinger equation
inwards from σ , starting from a single irregular function N L on this surface. The
full set of canonical basis functions satisfies Eq. (7.34).
The structural matrix G can be derived in this canonical basis from the matrix g of
structure constants, using the surface-matching theorem. For a modified irregular
function ζ̃ λ and the corresponding global matching function ξ = 'µ N µ β µ , the
matching conditions on the boundary of any cell τµ are
 µ µ λ   µ µ λ 
JL |Wσ |ζ̃ − ξ = 0, N L |Wσ |ζ̃ − ξ = 0. (7.45)
126 7 Multiple scattering theory for molecules and solids

Expanding the difference function ζ̃ λ − ζ λ in the form −φ µ G µλ in cell τµ , its

representation in the canonical basis on surface σµ is −(J µ − N µ t µ )G µλ . Similarly,
ζ λ is N µ δ µλ on σµ and ξ is N µ β µ − 'ν J µ g µν β ν . Then from Eqs. (7.45) the
coefficients in functions ξ and ζ̃ λ − ζ λ , respectively, are given by

β µ = δ µλ + t µ G µλ , G µλ = g µν β ν , (7.46)
ν

omitting indices L. Combining these equations, for each cell τµ ,


G µλ = g µλ + g µν t ν G νλ , (7.47)
ν

again omitting L-indices and summations. This is a Dyson equation in the form
of a matrix representation G = g + gt G of the Lippmann–Schwinger equation,
Eq. (7.24). This equation can also be derived by using the idea of atomic-cell
orbitals (ACO), generalized to irregular functions. An irregular ACO is defined in
a reference cell τλ as the canonical function ζ λ . In all other cells τµ it is the local
expansion of N λ , expressed in terms of the structure constants g µλ . Similarly, a
regular ACO takes the form φ − J in its indexed cell and Jgt in all other cells. The
ACO form of ζ̃ λ in cell τµ is
 

ζ̃ λ = ζ µ δ µλ − φ µ G µλ + J µ (δ µν − g µν t ν )G νλ − g µλ , (7.48)
ν

omitting indices L. Following tail-cancellation logic [7], the coefficients of func-

tions J must vanish in all cells. This implies Eq. (7.47).
µν
The structural matrix G L ,L  can be evaluated by direct matching across adjacent
cell interfaces, without using the matrix of structure constants. This construction
was originally derived (unpublished notes) from the Green function cellular method
(GFCM) [53, 441, 444]. The GFCM itself can be derived as a variant of the varia-
tional cellular method (VCM) [112, 110, 276], by restricting the form of allowed
variations on cell surfaces. The resulting NVCM or GFCM equations are given
above as Eqs. (7.6).
The structure constants or structural matrix for the Helmholtz Green function,
defined by Eq. (7.22), can be determined by applying Eqs. (7.6) to the global
solution of the Helmholtz equation given by Eq. (7.32), with a specified singularity
in cell τλ . For this application, the potential function vanishes everywhere except
for an implied singular potential in cell τλ , designed to specify the irregular solid
harmonic function N Lλ as an exact solution in this cell. Then the local basis in
each cell ν = λ is the set of regular solid harmonic functions JLν . Using these local
expansions, Eqs. (7.6) become a set of inhomogeneous linear equations for column
λ of the matrix of structure constants. Suppressing L-indices, these equations take
 7.5 Green functions 127

the form,

(N µ |Wµν |(N µ δ µλ − J µ g µλ ) − (N ν δ νλ − J ν g νλ )) = 0, (7.49)
ν=µ

for all cells τµ . Simplifying by use of Eq. (7.26),

 
(N µ |Wµν |J ν )g νλ − g µλ = (N µ |Wµν |N ν )δ νλ . (7.50)
ν=µ ν=µ

While these equations are consistent with the GFCM, their relationship to the
VCM variational equations suggests that the alternative formalism defined by the
JVCM might be more appropriate. In the global solution sought in the present case,
the irregular term is fixed, and all variational increments are linear combinations
of regular functions JL . Hence variation of coefficients gives the JVCM Eqs. (7.7).
A similar situation occurs for the Poisson equation, since a particular solution of
the inhomogeneous equation is fixed, and all incremental functions are regular
solid harmonics. Recent calculations comparing these methods for the Poisson
equation [443] show improved convergence in the JVCM formalism. It was shown
that extension of the Schlosser–Marcus variational principle to the Poisson equation
implies the JVCM equations [443]. In the present case, Eqs. (7.7) take the form

(J µ |Wµν |(N µ δ µλ − J µ g µλ ) − (N ν δ νλ − J ν g νλ )) = 0, (7.51)
ν=µ

for all cells τµ . Simplifying by use of Eq. (7.26),

 
(J µ |Wµν |J ν )g νλ = (J µ |Wµν |N ν )δ νλ + δ µλ . (7.52)
ν=µ ν=µ

Equation (7.52) is generalized to the Schrödinger Green function by using local

regular functions φ on the left and by using Eq. (7.36) to define an exact solution
of the Schrödinger equation in each cell, with a specified singularity in cell τλ . The
resulting linear equations take the form, suppressing L-indices,
 
(φ µ |Wµν |φ ν )G νλ = (φ µ |Wµν |ζ ν )δ νλ + δ µλ , (7.53)
ν=µ ν=µ

for all values of µ and λ. These equations determine the structural matrix G for
any set of space-filling cells.
Equations (7.53) simplify if the system has point-group or translational sym-
metry. Then local basis functions of equivalent cells are related to those of a smaller
number of generating cells by phase factors and elementary rotations. Considering
only translational symmetry, it is convenient to index cells in a reference transla-
tional cell by indices µ, etc. and to index translated equivalent cells by the corre-
sponding indices µ̃, etc., such that the displacement of cell τµ̃ relative to cell τµ
128 7 Multiple scattering theory for molecules and solids

is a lattice translational vector dµ̃µ . Using this notation, Eqs. (7.53) imply separate
equations for each k-vector in the reduced Brillouin zone, in the form
 µ ν̃ νλ
ν̃=µ (φ |Wµν̃ |φ ) exp(ik · dν̃ν )G (k) =
 µ ν̃ νλ
ν̃=µ (φ |Wµν̃ |ζ ) exp(ik · dν̃ν )δ + δ µλ , (7.54)
for all values of µ and λ that index cells in the reference translational cell. The
corresponding equations derived from Eq. (7.50) have been verified by comput-
ing structure constants for an fcc lattice along the ( − K line in the reduced
Brillouin zone, checked against a published structure-constant program STR [384].
Equations (7.53) have not been tested, but are expected to show improved conver-
gence because they are variationally correct for this problem.
 8
Variational methods for continuum states

The principal references for this chapter are:

[3] Adhikari, S.K. (1998). Variational Principles and the Numerical Solution of
Scattering Problems (Wiley, New York).
[47] Burke, P.G. and Robb, W.D. (1975). The R-matrix theory of atomic processes, Adv.
At. Mol. Phys. 11, 143–214.
[178] Huo, W.M. and Gianturco, F.A. (1995). Computational Methods for Electron-
Molecule Collisions (Plenum, New York).
[202] Kohn, W. (1948). Variational methods in nuclear collision problems, Phys. Rev. 74.
1763–1772.
[214] Lane, A.M. and Thomas, R.G. (1958). R-matrix theory of nuclear reactions, Rev.
Mod. Phys. 30, 257–353.
[215] Lane, N.F. (1980). The theory of electron–molecule collisions, Rev. Mod. Phys. 52,
29–119.
[246] Moiseiwitsch, B.L. (1966). Variational Principles (Interscience, New York).
[257] Mott, N.F. and Massey, H.S.W. (1965). The Theory of Atomic Collisions (Oxford
University Press, New York).
[270] Nesbet, R.K. (1980). Variational Methods in Electron-Atom Scattering Theory
(Plenum, New York).
[371] Schwinger, J. (1947). A variational principle for scattering problems, Phys. Rev. 72,
742.
[428] Wigner, E.P. and Eisenbud, L. (1947). Higher angular momenta and long range
interaction in resonance reactions, Phys. Rev. 72, 29–41.

8.1 Scattering by an N-electron target system

Scattering by an N-electron atom or molecule with fixed nucleus is described by
an (N +1)-electron Schrödinger wave function of the form
 
!s = A# p ψ ps + "µ cµs . (8.1)
p µ

The index s here denotes a particular degenerate solution, specified by boundary

conditions, at given total energy E. The N-electron function # p is a target state with

129
130 8 Variational methods for continuum states

energy E p . It is assumed that all target functions # p are defined by orthonormal

eigenvectors of a Hamiltonian matrix in a common set of N-electron basis functions.
The eigenvalues are the energies E p . ψ ps is a one-electron channel orbital wave
function, antisymmetrized into # p by the operator A. The functions "µ are an
orthonormal set of (N +1)-electron Slater determinants. The channel orbital ψ ps
is characterized by orbital and spin angular momentum quantum numbers and by
asymptotic energy E − E p . This energy is positive for open scattering channels,
defining a wave vector or electron momentum k p such that E − E p = 12 k 2p . Hartree
atomic units are used here. If E − E p < 0 the channel is closed. Then k p is replaced
by iκ p , for κ p > 0. In practice, the basis functions "µ are replaced by symmetry-
adapted linear combinations of simple Slater determinants, defined for total spin
and atomic orbital angular momentum or for molecular point-group symmetry. The
expansion indicated in Eq. (8.1) remains completely general. For heavy atoms, the
notation can be adapted to j j-coupling.
For an atom, or outside a sphere that completely encloses a target molecule,
channel orbital functions are of the form
ψ ps = r −1 f ps (r )Ym  (θ, φ)u m s ,
−1
where, for r → ∞, f ps (r ) ∼ k p 2 sin(k p r − 12  p π + η p ), for single-channel scat-
tering by a neutral target. This functional form must be modified for long-range
Coulomb and dipole potentials. The normalization implies unit radial flux den-
sity for a free electron. For interacting scattering channels, the multichannel wave
functions are


 
− 12 1 1
f ps (r ) ∼ k p sin k p r −  p π α0 ps + cos k p r −  p π α1 ps . (8.2)
2 2
The coefficient matrices α0 and α1 determine scattering matrices and cross sections.
Closed-channel orbital functions must vanish for r → ∞, in a way determined
by the long-range potentials in the Schrödinger equation. The term in exp(−κ p r )
implied by analytic continuation of an open-channel orbital through the continuum
threshold is generally dominated by terms in reciprocal powers of r due to such
long-range potentials.
It is assumed that ψ ps is orthogonal to all orbital functions used to construct
# p and "µ . This ensures that Eq. (8.1) describes projection of !s onto orthog-
onal target-state components ! p = A# p ψ ps and a residual orthogonal compo-

nent ! Q = µ "µ cµs that is quadratically integrable. The coefficients cµs are
determined variationally. For any calculation using a finite orbital basis, the open-
channel terms A# p ψ ps remain distinct from the Hilbert-space component ! Q . The
oscillatory function ψ ps with nonvanishing asymptotic amplitude cannot be repre-
sented as a finite superposition of quadratically integrable functions. In contrast,
 8.1 Scattering by an N-electron target system 131

closed-channel terms can be included in either sum. In the close-coupling formal-

ism, part of ! Q can be replaced by A#γ ψγ s , where #γ represents a target pseudo-
state with energy E γ > E, corresponding to a closed channel. Such a function lies
in the electronic continuum. Polarization response can be represented in terms of
such polarization pseudostates #γ ( p) , computed as the first-order perturbation of
# p by a polarizing field. The corresponding closed-channel orbital function ψγ s
may be computed from the close-coupling equations derived here.
Using the projection-operator formalism of Feshbach [115, 116], an implicit vari-
ational solution for the coefficients cµs in ! Q can be incorporated into an equivalent
partitioned equation for the channel orbital functions. This is a multichannel variant
of the logic used to derive the correlation potential operator v̂c in orbital-functional
theory. Define a projection operator Q such that

! Q = Q!s = "µ ("µ |!s ),
µ

and the orthogonal complement operator P, such that


! P = P!s  A# p ψ ps .
p

Defining a Schrödinger functional  = (!|H − E|!), the Euler–Lagrange equa-

tions for fixed energy E are
δ δ
= 0; = 0.
δ! P∗ δ! Q∗
The resulting coupled (N +1)-electron equations are
P(H − E)P! P + PHQ! Q = 0
QHP! P + Q(H − E)Q! Q = 0.
The reduced operator Q(H − E)Q in the Q-space can be inverted to give a formal
solution of the second equation,
! Q = −{Q(H − E)Q}−1 QHP! P .
The operator {Q(H − E)Q}−1 is an (N +1)-electron linear operator whose kernel
is

"µ (H − E)−1 ∗
µν "ν .
µ,ν

When substituted into the first equation, this gives an effective multichannel
Schrödinger equation of the form
{ H̃ − E}! P = {P(H − E)P − PHQ{Q(H − E)Q}−1 QHP}! P = 0. (8.3)
132 8 Variational methods for continuum states

Equation (8.3) provides a common basis for the major computational methods used
in low-energy electron scattering theory. The terms containing {Q(H − E)Q}−1 in
this equation describe polarization response and other correlation effects in terms
of a matrix optical potential. Singularities due to this energy denominator cancel
out exactly [264, 265, 270].
Effective one-electron equations for the channel orbital functions can be obtained
either by evaluating orbital functional derivatives of the variational functional 
or more directly by projecting Eq. (8.3) onto the individual target states # p . With
appropriate normalizing factors, (# p |!s ) = ψ ps . Equations for the radial channel
functions f ps (r ) are obtained by projecting onto spherical harmonics and elemen-
tary spin functions. The matrix operator acting on channel orbitals is

m̂ pq = (# p | H̃ − E|A#q ). (8.4)

The projection of this operator onto spherical harmonics and spin functions defines
a radial operator m pq = ĝ pq −  p δ pq , where  p = E − E p .

8.1.1 Cross sections

If there are n o open channels at energy E, there are n o linearly independent degen-
erate solutions of the Schrödinger equation. Each solution !s is characterized by
a vector of coefficients αi ps , for i = 0, 1, defined by the asymptotic form of the
multichannel wave function in Eq. (8.1). The rectangular column matrix α consists
of the two n o × n o coefficient matrices α0 , α1 . Any nonsingular linear combination
of the column vectors of α produces a physically equivalent set of solutions. When
multiplied on the right by the inverse of the original matrix α0 , the transformed
α-matrix takes the canonical form

α0 = I, α1 = K ,

which defines the reactance matrix K = α1 (α0 )−1 . For exact solutions of the
projected Schrödinger equation, the open-channel K-matrix is real and sym-
metric [257, 302]. It can be diagonalized by an orthogonal transformation to
define eigenchannels, defined by the eigenvectors. The eigenvalues are tan ησ , for
σ = 1, . . . , n o , which defines the eigenphases ησ up to a multiple of π radians. The
scattering matrix S is defined by [257]

S = (I + i K )(I − i K )−1 ,

and the transition matrix T is defined by

1
T = (S − I ) = K (I − i K )−1 .
2i
 8.1 Scattering by an N-electron target system 133

In terms of eigenchannels and eigenphases,


no
K ps = x pσ xsσ tan ησ
σ =1
no
S ps = x pσ xsσ exp(2iησ )
σ =1
no
T ps = x pσ xsσ exp(iησ ) sin(ησ ).
σ =1

The S-matrix is unitary and symmetric, while the T-matrix is symmetric. This
particular definition of the T-matrix reduces for scattering by a central potential to
the phase-shift factor in the scattering amplitude,

1 ∞
f (θ ) = (2 + 1) exp(iη ) sin η P (cos θ ),
k =0

such that the differential cross section is

dσ
= | f (θ)|2 .
d)
If k is in atomic units a0−1 , the differential cross section is in units a02 per steradian.
Differential and total cross sections for multichannel scattering by an atomic tar-
get can be derived from general formulas [27, 184]. The partial cross section for
scattering from channel q to channel p is
2
4π 4π 
σqp = 2 |T pq | = 2
2
x pσ xqσ exp(iησ ) sin(ησ ) .
kq kq σ

The total cross section for unpolarized scattering is obtained by summing σqp over
degenerate final states and by averaging over initial states.

8.1.2 Close-coupling expansion

If the set of target states # p in Eq. (8.1) could be extended to completeness, solution
of the coupled equations for the channel orbitals ψ ps would provide a quantitative
solution of the scattering problem for an N-electron target. This is the strategy
of the close-coupling method [374, 325, 48, 376]. The coupled integrodifferen-
tial equations for the channel orbitals are the close-coupling equations. Because
the expansion in target states must be truncated for any practical calculation, the
second term in Eq. (8.1) is required in order to include polarization and correlation
effects as well as to assure consistent orbital orthogonality conditions. No generally
134 8 Variational methods for continuum states

satisfactory way has been found to include the continuum of unbound target states
required for completeness of this expansion.
Since only a relatively small number of coupled equations can be solved in
practice, the target states # p must be selected carefully. Virtual target excitation
into the ionization continuum must be approximated by inclusion of closed-channel
pseudostates that cannot be target eigenstates but have the character of wave packets
in the continuum. Target polarization response is treated by including polarization
pseudostates #γ ( p) .
If only a single target state is included, itself represented as a single-determinant
model state " p , the formalism reduces to the static-exchange (SE) model. Because
of the complexity of more detailed scattering models, much work in fixed-nuclei
electron scattering by molecules has been carried out at this level of approximation
[215]. This model is generally unsatisfactory for low-energy scattering, and must
be augmented by a valid approximation to the dipole polarization response of the
target system [270, 135]. Models that augment static exchange with polarization
response are denoted by the acronym SEP. It will be shown below that a model with
this computational structure can be derived from formally exact orbital functional
theory (OFT), if the correlation potential operator v̂c is approximated in terms of
polarization pseudostates [290].

8.2 Kohn variational theory

Variational theory relevant to the representation of an (N +1)-electron continuum
wave function by Eq. (8.1) will be considered here. Direct computational imple-
mentation of this formalism defines the matrix variational method. Since the widely
applied close-coupling method can be described by Eq. (8.1), the general argument
here also applies to that formalism, with some modifications of detail. In the al-
gebraic close-coupling method, the variational formalism, using an orbital basis
expansion, replaces direct numerical solution of integrodifferential close-coupling
equations.
In bound-state calculations, the Rayleigh–Ritz or Schrödinger variational princi-
ple provides both an upper bound to an exact energy and a stationary property that
determines free parameters in the wave function. In scattering theory, the energy is
specified in advance. Variational principles are used to determine the wave function
but do not generally provide variational bounds. A variational functional is made
stationary by choice of variational parameters, but the sign of the residual error is not
determined. Because there is no well-defined bounded quantity, there is no simple
absolute standard of comparison between different variational trial functions. The
present discussion will develop a stationary estimate of the multichannel K-matrix.
Because this matrix is real and symmetric for open channels, it provides the most
 8.2 Kohn variational theory 135

convenient of several alternative representations of the asymptotic part of a scatter-

ing wave function. Recent developments that exploit aspects of the complex-valued
S- or T-matrices will be discussed below.
Variational principles for scattering matrices and wave functions were originally
introduced by Hulthén [175, 176] and by Kohn [202]. Quantitative applications of
variational methods to electron–atom scattering theory began with calculations of
e− –H elastic scattering by Schwartz [369, 370], using the Kohn method for the
K-matrix. Further progress required understanding and resolution of the problem
of spurious singularities in the Kohn formalism. These developments have been
discussed in detail elsewhere [270]. More recently, this problem has largely been
bypassed in the complex Kohn formalism, which has been very effectively used for
calculations of electron–molecule scattering [341].

8.2.1 The matrix variational method

The asymptotic form of radial open-channel orbitals f ps (r ) is given by Eq. (8.2).
Functions of this form can be represented as linear combinations of two independent
continuum basis functions for each open channel. These basis functions must be
regular at the coordinate origin, but have the asymptotic forms


−1 1
F0 p (r ) ∼
kp 2 sin k p r −  p π ,
2

 (8.5)
− 12 1
F1 p (r ) ∼ k p cos k p r −  p π .
2

A variational approximation with the required asymptotic form

f ps (r ) = F0 p α0 ps + F1 p α1 ps (8.6)

determines the coefficients αi ps defined by Eq. (8.2). In the matrix variational

method [270], the quadratically integrable orbital basis is augmented for each open
channel by adding two functions φ0 p , φ1 p whose radial factors are r −1 F0 p (r ) and
r −1 F1 p (r ), respectively. These functions are orthogonalized to the quadratically
integrable orbital basis {φi ; φa }.
For an exact scattering wave function, the coefficients cµs in Eq. (8.1) would
be determined as linear functions of the matrix elements αi ps , since the algebraic
equations are linear. By factoring these coefficients

cµs = cµi p αi ps ,
ip
136 8 Variational methods for continuum states

the variational approximation to the (N +1)-electron wave function takes the form

 
!s = A# p φi p + "µ cµ αi ps .
ip

ip µ

ip
In the matrix variational method, the coefficients cµ are determined separately
for each set of indices i, p from the matrix equations, for all µ,


"µ |H − E|A#q φ jq + "ν cνjq = 0, (8.7)
ν

which imply the variational condition ("µ |H − E|!s ) = 0. For arbitrary values of
the coefficients αi ps these equations follow from the variational condition
∂st
i p∗
= 0,
∂cµ
for all µ, i, p. The variational functional st = (!s |H − E|!t ) here is an n o × n o
matrix.
When Eq. (8.7) is satisfied, the variational functional becomes an explicit
quadratic function of the coefficients αi ps , which can be assumed to be real numbers.
Thus
 pq
st = αi ps m i j α jqt ,
ip jq

where, as a consequence of Eq. (8.7),


Mi p,µ (M −1 )µν Mν, jq .
pq pq
m i j = Mi j −
µ,ν

This equation combines several submatrices of H − E: the Hermitian bound–bound

matrix Mµν , the rectangular bound–free matrix Mµ,i p , and the nonhermitian free–
pq
free matrix Mi j , where
Mµν = ("µ |H − E|"ν ),
Mµ,i p = ("µ |H − E|A# p φi p ),
pq
Mi j = (A# p φi p |H − E|A#q φ jq ).
pq pq
Phase factors can be chosen so that both Mi j and m i j are real but unsymmetrical.
If !t were an exact solution of the Schrödinger equation, and total energy E is
above the thresholds for n o open channels, there would be n o linearly independent
solutions of the equations
 pq
m i j α jqt = 0, (8.8)
jq
 8.2 Kohn variational theory 137

for all indices i, p. Each such solution defines a 2n o element column vector of
the rectangular matrix of coefficients α. In a matrix notation, suppressing channel
indices, the variational functional is
† †
 = α † mα = α0 (m 00 α0 + m 01 α1 ) + α1 (m 10 α0 + m 11 α1 ).
In this notation, Eq. (8.8) is



m 00 m 01 α0
mα = = 0.
m 10 m 11 α1

The reactance matrix K is α1 α0−1 . Exact solutions require the matrix m i j to be of

pq

rank n o , implying n o linearly independent null-vectors as solutions of the homo-

geneous equations mα = 0. Because this algebraic condition is not satisfied in
general by approximate wave functions, a variational method is needed in order to
specify in some sense an optimal approximate solution matrix α.
If α were an exact solution of the matrix equation mα = 0, the functional  =
α † mα would vanish. For a variational approximation, consider the variation of 
induced by an infinitesimal variation δα. This is
δ = δα † mα + (mα)† δα + α † (m − m † )δα.
pq
The last term here does not vanish even if mα = 0. The nonhermitian part of m i j
comes from the free–free matrix of H − E, which is characteristic of scattering the-
ory. When channel orbital functions are normalized to unit flux, with the asymptotic
forms specified above, this nonhermitian term is given by
pq qp pq qp 1
m i j − m ji = Mi j − M ji = δ pq (δi0 δ j1 − δi1 δ j0 ).
2
This formula expresses the surface integral obtained by integrating the kinetic en-
ergy integral by parts for open-channel orbital functions of the specified asymptotic
form. In matrix notation,
† 1
m 01 − m 10 =I,
2
where I is the n o × n o unit matrix. The matrices m 00 and m 11 are Hermitian (real
and symmetric). When substituted into the expression for δ this gives
1 † †
δ = δα † mα + (mα)† δα + (α0 δα1 − α1 δα0 ). (8.9)
2

8.2.2 The Hulthén–Kohn variational principle

Although δ does not vanish even if the variational equation mα = 0 is satisfied,
the nonvanishing terms can be combined with variations of the various scattering
138 8 Variational methods for continuum states

matrices to give stationary expressions for these matrices. This logic produces
multichannel versions of the variational principle originally derived by Hulthén
[175, 176], Kohn [202], and Kato [194, 195]. As applied to the K-matrix, Kohn’s
variational principle can be derived by considering variations about an exact solution
that maintain the canonical form α0 = I, α1 = K . Then δα0 = 0 and δα1 = δ K ,
which implies δ = 12 δ K . This implies that the Kohn functional

[K ] = K − 2

is stationary for all such variations.

In the matrix variational method, the equations mα = 0 do not in general have a
solution. For variations about an estimated matrix K t , and restricted to the canonical
form,

1
δ = δ K † (m 10 + m 11 K t ) + (m 10 + m 11 K t )† δ K + δ K .
2

If K t = −m −1 11 m 10 , such that m 10 + m 11 K t = 0, this implies that [K ] = K t −

2(K t ) is stationary. This defines
† †
[K ] = K t − 2(m 00 + m 01 K t + K t m 10 + K t m 11 K t )
 † 
= −2 m 00 − m 10 m −1
11 m 10 . (8.10)

This is a real symmetric matrix by construction, and remains valid for an exact
variational solution, when  = 0. Hence this derivation proves that an exact open-
channel K-matrix is real and symmetric.
The difficulty with this formula is that there is no guarantee that the auxiliary
matrix m 11 is not singular [264, 265]. At an energy for which an eigenvalue of m 11
vanishes, K t is not defined unless all columns of the matrix m 10 are orthogonal
to the null eigenvector. Specific examples [265] show that m 10 does not have this
property in general for otherwise valid variational trial wave functions. The number
of null values of m 11 increases with the number of basis functions, leading to energy
and basis-dependent anomalies in the Kohn formalism [369, 370]. In an alternative
version of the theory [176, 350], variations are restricted to α0 = K −1 and α1 = I .
This implies that anomalies occur at null points of m 00 , which in general do not co-
incide with those of m 11 . This fact can be exploited to generate various anomaly-free
variants of the Kohn theory [265]. Especially in the matrix variational formalism,
this method has been applied to many calculations of electron–atom scattering of
high accuracy [270]. This method was used to compute e− –He scattering cross
sections that established a calibration standard for subsequent experimental work
[268, 269].
 8.2 Kohn variational theory 139

8.2.3 The complex Kohn method

The derivation given above of the stationary Kohn functional [K ] depends on
logic that is not changed if the functions F0 and F1 of Eq. (8.5) are replaced in
†
each channel by any functions for which the Wronskian condition m 01 − m 10 =
1
2
I is satisfied [245, 191]. The complex Kohn method [244, 237, 440] exploits
this fact by defining continuum basis functions consistent with the canonical form
α0 = I, α1 = T , where T is the complex-symmetric multichannel transition matrix.
These continuum basis functions have the asymptotic forms


− 12 1
u 0 p (r ) ∼ k p sin k p r −  p π ,
2


−1 1
u 1 p (r ) ∼ k p 2 exp i k p r −  p π .
2
This transformed representation of the asymptotic wave functions can easily be
verified for potential scattering. The asymptotic radial wave function in a given
-channel satisfies the identity

 
 
− 12 iη 1 − 12 1 i(kr − 12 π )
k e sin kr − π + η = k sin kr − π + e T ,
2 2
where T = (e2iη − 1)/2i = eiη sin η.
Using these asymptotic continuum functions, the derivation given above implies
that for Tt = −m −1
11 m 10 , such that m 10 + m 11 Tt = 0, then
 † 
[T ] = Tt − 2(Tt ) = −2 m 00 − m 10 m −1 11 m 10

is stationary for infinitesimal variations of Tt , and for infinitesimal variations about

an exact solution for which  = 0. The significance of this revised formalism is
that the matrix m 11 , constructed from complex exponential basis functions, is a
symmetric matrix but no longer real. Its null values are displaced away from the
real energy axis, which essentially eliminates the problem of Kohn anomalies [237].
This can be rationalized from the theory of scattering resonances: null values of m 11
correspond to pure outgoing-wave solutions of the projected Schrödinger equation.
This condition characterizes resonances that in general are displaced below the real
energy-axis by a finite width or reciprocal lifetime [192, 382].
Introduced in the context of heavy-particle reactive collisions [440], the complex
Kohn method has been successfully applied to electron–molecule scattering [341].
It is accurate but computationally intensive, since continuum basis orbitals do not
have the Gaussian form that is exploited in most ab initio molecular bound-state
studies. The method has been implemented using special numerical methods [341]
developed for these integrals. These numerical methods mitigate another practical
140 8 Variational methods for continuum states

limitation of the matrix variational method [270], in that analytic formulas are
available only for specialized forms of the long-range potentials that can occur,
especially for molecules.

8.3 Schwinger variational theory

In the Hulthén–Kohn formalism of the matrix variational method, a large part
of the computational effort is concerned with the evaluation and inversion of the
bound–bound matrix of the (N +1)-electron operator H − E. Because powerful
computational methods for this matrix can be taken over directly from bound-state
atomic physics and theoretical chemistry, there is a great practical advantage in
using continuum variational methods that incorporate this as a central aspect of the
overall algorithm. A fundamentally different approach replaces the Schrödinger
equation by a formally equivalent Lippmann–Schwinger integral equation [228],
developing stationary expressions for the resulting scattering matrices. In such
methods, a Green function is introduced for the asymptotic electron continuum
part of the overall wave function, while short-range effects are treated through an
incremental potential v. This formalism makes use of the Schwinger variational
principle [371]. Formulated as a variational method using exponential or Gaussian
orbital basis functions [419, 418], and extended to a multichannel formalism [270,
396], it has been widely applied to calculations of electron–molecule scattering
[230, 177].
The theory will be developed here with reference to a spherical coordinate system,
appropriate to a target atom or to a single-center representation of a molecular wave
function. Extension to a multicenter representation is straightforward, following
the same formal argument, but greatly complicates the notation and the detailed
form of the Green function used in this theory. The formalism can most easily be
understood in the simple model of potential scattering in a single -channel. A model
Schrödinger equation, determined by a radial Hamiltonian operator h = t̂ + v(r ),
is assumed to be modified by a difference potential v, which may be a nonlocal
linear operator. At energy  = 12 k 2 , in the scattering continuum, regular and irregular
solutions w0 (r ) and w1 (r ), respectively, of the model equation

(h − )wi (r ) = 0

are assumed to be known. If normalized to unit Wronskian,

w1 (r )w0 (r ) − w1 (r )w0 (r ) = 1,

these functions define a Green function G, a linear operator whose kernel is

g(r, r  ) = 2w0 (r< )w1 (r> ).

 8.3 Schwinger variational theory 141

Acting on any function F(r ) that vanishes for r → ∞ and is regular at the origin
r → 0,

∞
r 

G F(r ) = 2 w0 (r ) w1 (r ) + w1 (r ) w0 (r ) F(r  ) dr  .

r 0

This implies that (h − )G F = F, which can be verified by applying the operator

h −  to G F(r ). Thus the Green function is a formal inverse of h − . The kernel
function g(r, r  ) is symmetric in r, r  , regular at the origin in either variable, and
continuous. The derivative in either variable is discontinuous at r  = r , so that
(h − )G is equivalent to a Dirac δ-function. G depends on  by construction.

∞
G F ∼ 2w0 (r ) w1 (r  )F(r  ) dr 
0

at the coordinate origin, r → 0, and as r → ∞,

∞
G F ∼ 2w1 (r ) w0 (r  )F(r  ) dr  .
0

Thus G F is regular at the origin but is asymptotically proportional to the irregu-

lar function w1 (r ). It has the properties assumed for the second continuum basis
function required for each open channel in the matrix variational method.
These properties of the model Green function imply that the Lippmann–
Schwinger equation [228],
f = w0 − Gv f,
defines a solution of the radial Schrödinger equation (h −  + v) f (r ) = 0. Reg-
ular and irregular functions with asymptotic forms, respectively,



1 1
w0 ∼ k − 2 sin kr − π a, w1 ∼ k − 2 cos kr − π ,
1 1

2 2
are used to construct the principal value Green function. This Green function im-
poses the asymptotic form f (r ) ∼ w0 (r ) + w1 (r ) tan η, where

∞
tan η = −2 w0 (r  )v(r  ) f (r  ) dr  .
0

Other asymptotic forms consistent with unit Wronskian define different but equally
valid Green functions, with different values of the asymptotic coefficient of w1 . In
particular, if w1 ∼ k − 2 exp i(kr − 12 π ), this determines the outgoing-wave Green
1

function, and the asymptotic coefficient of w1 is the single-channel T-matrix,

eiη sin η. This is the basis of the T-matrix method [342, 344], which has been
used for electron–molecule scattering calculations [126]. It is assumed that v f
is regular at the origin and that v vanishes more rapidly than r −2 for r → ∞.
142 8 Variational methods for continuum states

Coulomb or static dipole potentials should be included in the model potential whose
solutions w0 , w1 are used to construct the Green function.
Specializing the present derivation to the principal value Green function, the
unsymmetrical expression tan η = −2(w0 |v| f ) is exact for an exact solution of
the Lippmann–Schwinger equation, but it is not stationary with respect to infinites-
imal variations about such a solution. Since w0 = f + Gv f for such a solution,
this can be substituted into the unsymmetrical formula to give an alternative, sym-
metrical expression tan η = −2( f |v + vGv| f ), which is also not stationary.
However, these expressions can be combined to define the Schwinger functional

[tan η] = −2(w0 |v| f )( f |v + vGv| f )−1 ( f |v|w0 ),

which is stationary for variations of the trial function f about an exact solution.
The variation δ[tan η] induced by an infinitesimal variation of f about an exact
solution is
 
(δ f |v|w0 ) (δ f |v + vGv| f )
− + hc [tan η]
( f |v|w0 ) ( f |v + vGv| f )
= −2(δ f v|w0 − f − Gv f ) + hc.

This implies that [tan η] is stationary if and only if f , in the range of v, satisfies
the Lippmann–Schwinger equation.
The Schwinger functional has several remarkable properties. It combines three
formulas that may have different numerical values:

tan η  −2(w0 |v| f ),

tan η  −2( f |v + vGv| f ),
tan η  −2( f |v|w0 ).

The function f can be arbitrarily normalized, because [tan η] is homogeneous in

f . There is no constraint on the asymptotic form of trial functions, because they
have no effect outside the range of the assumed short-range potential v. Only
the regular model functions w0 occur explicitly in the theory, although the irregular
functions w1 must be consistent with the asymptotic character of the Green function,
and may be used to construct it.
An algebraic theory is obtained if f is expanded within the range of v as a

linear combination f (r ) = a ηa (r )(a| f ) of orthonormal basis functions {ηa }. v
is expanded in the same basis as a linear integral operator with the kernel,

v(r, r  ) = ηa (r )vab ηb∗ (r  ).
a,b

For [tan η] to be stationary with respect to variations of the coefficients (a| f ), this
 8.3 Schwinger variational theory 143

requires, for all indices a,

(a| f − w0 + Gv f ) = 0.
This is a system of linear equations that determine

−1
(a| f ) = (I + Gv)ab (b|w0 )
b

−1
= (v + vGv)ab vbc (c|w0 ).
b,c

Combining these equations, all three expressions for tan η give the same result for
the stationary functional,

[tan η] = (w0 |a)vac (v + vGv)−1 cd vdb (b|w0 ).
a,b c,d

8.3.1 Multichannel Schwinger theory

For practical use in electron scattering theory, the Lippmann–Schwinger equation
and the Schwinger variational principle must be generalized to a multichannel
formalism. In applying this formalism, the orthogonality conditions that distin-
guish the bound and free components of an (N +1)-electron continuum !s , as
given in Eq. (8.1), must be taken into account. The argument here summarizes
an earlier derivation [270]. The multichannel close-coupling equations appropri-
ate to the (N +1)-electron problem require orthogonality constraints expressed by
off-diagonal Lagrange multipliers. Reduction to a form in which these parameters
vanish by construction will be considered here after generalizing the Schwinger
stationary principle to such equations in their simpler homogeneous form.
Consider the multichannel radial close-coupling equations, without Lagrange
multipliers for orthogonality,

m pq u qs (r ) = 0, (8.11)
q

where
m pq = (h − ) pq + v pq .
It is assumed that all closed-channel components have been eliminated by the
partitioning transformations that determine Eq. (8.4), from which the radial operator
m pq is derived. Equations of this form can be solved using the Green function of
the model problem

(h − ) ps wisq (r ) = 0.
s
144 8 Variational methods for continuum states

The multichannel model functions w0sq and w1sq are real-valued continuum solu-
tions of the model problem, in a K-matrix formalism. They are normalized to the
matrix Wronskian condition


(wisp w jsq − wisp w jsq ) = δ pq (δi0 δ j1 − δi1 δ j0 ).
s

The equations here are simplified by working in an eigenchannel representation,

in which the model K-matrix is diagonalized, with eigenvalues tan ησ . Defining
matrices C and S, diagonal in the eigenchannel representation, with eigenvalues
cos ησ and sin ησ , respectively, the physical K-matrix implied by an incremental
matrix K  in this representation is [375]
K = (S + C K  )(C − S K  )−1 .
The multichannel generalization of the principal value Green function is a linear
operator G σpq , defined by its kernel
g σpq (r, r  ) = 2w0 pσ (r< )w1qσ (r> ),
indexed by eigenchannel σ of the model continuum functions.
If {Fq (r )} is a vector of functions, one for each open channel, that are regular at
the coordinate origin and quadratically integrable, then
  
∞
r 
σ 
G pq Fq (r ) = 2 w0 pσ (r ) w1qσ (r ) + w1 pσ (r ) w0qσ (r ) Fq (r  ) dr  ,


q q r 0

such that
 
(h − ) pp G σp q Fq (r ) = F p (r ).
σ p,q 

Using the matrix Wronskian condition, the proof is the same as for the single-
channel problem.
This multichannel matrix Green function determines a multichannel Lippmann–
Schwinger equation
 
u pτ = w0 pτ − G σpp v p q u qτ
σ p  ,q

that determines solutions of the multichannel Schrödinger equation. The asymptotic

form of the second term here is

−2 w1 pσ (w0 pσ |v pq |u qτ ),
p,q

which implies that the incremental K-matrix in the model eigenchannel representa-
tion is K  = −2(w0 |v|u), in a matrix notation. The Hermitian conjugate relation
is K  = −2(u|v|w0 ).
 8.3 Schwinger variational theory 145

The multichannel Lippmann–Schwinger equation implies that w0 = u + Gvu.

Substituting this into the second expression for K  gives an alternative formula,

K  = −2(u|v + vGv|u).

The multichannel Schwinger functional is defined by the matrix product

[K  ] = −2(w0 |v|u)(u|v + vGv|u)−1 (u|v|w0 ).

This expression is stationary for variations of u if and only if u is an exact solution

of the multichannel Lippmann–Schwinger equation. The proof follows exactly as
in the single-channel case, if the order of matrix products is maintained in the
derivation. If the multichannel continuum solution is expanded as

u ps (r ) = η pa ( pa| ps),
a

where {η pa } are orthonormal radial basis functions, then [K  ] is a function of

the coefficients ( pa| ps). Following the logic of the single-channel derivation, the
stationary expression obtained by varying [K  ] with respect to these coefficients is

[K  ]σ τ =

−2 (w0 pσ |v|η p a )[(η|v + vGv|η)]−1
p  a,q  b (ηq  b |v|w0qτ ).
a,b p, p  q,q 

The same result is obtained by substituting a formal solution for u into any of the
three approximate expressions for K  combined in the Schwinger functional matrix.

8.3.2 Orthogonalization and transfer invariance

Equations (8.11), the multichannel radial close-coupling equations, are derived
from Eq. (8.3) by projecting onto orbital spherical harmonics for a single spin
component. The partitioning argument that leads to Eq. (8.3) explicitly requires
each channel orbital wave function ψ ps to be orthogonal to all orbital functions
used to construct target states # p and (N +1)-electron basis states "µ . In the matrix
variational formalism, this condition is imposed by orthogonalizing each continuum
basis function to the full set of quadratically integrable basis orbitals. Because open-
channel continuum functions have finite amplitudes outside any sphere enclosing
the target system, such orthogonalization cannot eliminate them. In practice, this is
also true for pseudostate closed-channel orbitals generated by long-range potentials.
This orthogonality condition is implemented in close-coupling theory by
p
introducing Lagrange multipliers λas into the radial integrodifferential equations,
146 8 Variational methods for continuum states

which become
 
m pq u qs (r ) = η pa λas
p
. (8.12)
q a

The λ multipliers are determined by the orthogonality condition. If the quadratically

integrable radial basis in channel p is the orthonormal set {η pa }, values of the
Lagrange multipliers are given by

λas
p
= (η pa |m pq |u ps ).
q

In the Lippmann–Schwinger formalism, the Green function can be defined in a

projected orbital space consistent with this orthogonality condition. An alternative,
assumed in the derivation given above, is to use a Green function appropriate to
close-coupling equations from which the Lagrange multipliers have been elimi-
nated. This can be done if each (N +1)-electron trial wave function has the property
of transfer invariance [270]. This means that !s , in Eq. (8.1), is unchanged under
any transfer of quadratically integrable terms A# p φ ps between its free and bound
 
components, p A# p ψ ps and µ "µ cµs , respectively. If a valid system of close-
coupling equations exists for any arbitrary transfer of such terms, these equations
are transfer covariant. Projection of the Schrödinger equation onto the bound space
determines the coefficients cµs through linear equations
("µ |H − E|!s ) = 0
for all µ, or
 
(H − E)µν cνs = − ("µ |H − E|A# p ψ ps ).
ν p

This is transfer covariant if all quadratically integrable functions are represented

in the same orbital basis. Requiring ψ ps to be orthogonal to all φ pa (radial factor
η pa (r )) enforces a unique representation, but introduces Lagrange multipliers in
the close-coupling equations. An alternative is to require
(# p |H − E|!s ) = 0
to be valid for all # p , and to adjust the Lagrange multipliers accordingly. This
implies that
(A# p φ pa |H − E|!s ) = 0
for all p, a. This reduces to

(η pa |m pq |u qs ) = λas
p
= 0,
q
 8.4 Variational R-matrix theory 147

which removes the Lagrange multipliers, if the operator m pq is determined from

(# p |H − E|!s ) as in the matrix variational formalism. This analysis shows that a
representation of the transfer-covariant close-coupling equations exists in which the
Lagrange multipliers for orthogonalization vanish. This representation is assumed
in the development of the multichannel Schwinger theory given above.
Use of the model Green function in the Lippmann–Schwinger equation is the most
serious weakness of the Schwinger variational formalism. The method requires v
to be a short-range potential or operator, so that it can be represented quantitatively
in a quadratically integrable orbital basis. This requires all long-range potentials
to be included in the model Hamiltonian used to construct the Green function.
This is a formidable difficulty in the context of electron–molecule scattering, since
much of the structure of low-energy cross sections is due to polarization response
and to long-range dipole and quadrupole potentials [215, 178]. A molecular Green
function must satisfy Coulomb cusp conditions and be regular at each atomic center,
precluding the use of standard asymptotic forms for model long-range potentials.
Construction of such a Green function and evaluation of the implied integrals
(η|vGv|η) requires analytic or computational methodology beyond present
capabilities.

8.4 Variational R-matrix theory

The Wigner–Eisenbud [428] R-matrix, or derivative matrix, is defined by the rela-
tionship between radial channel orbitals f ps (r ) and their derivatives on some sphere
of radius r1 that surrounds a target system [214, 33]. Assuming spherical geometry,
the dimensionless radial R-matrix R pq at r1 is defined by


f ps (r1 ) = R pq r1 f qs (r1 ).
q

The indices refer to both open and closed channels, which do not have to be dis-
tinguished at the finite radius r1 . As defined for solutions of radial close-coupling
equations, R pq is a real symmetric matrix. Extension of the theory and definitions to
more general geometry will be considered below. For electron–molecule scattering,
a molecular center is chosen somewhat arbitrarily such that the electronic charge
is approximately enclosed in some smaller sphere of radius r0 , while bound–free
exchange can be neglected outside the larger R-matrix radius r1 .
The theory of the R-matrix was developed in nuclear physics. As usually pre-
sented, the theory makes use of a Green function to relate value and slope of the
radial channel orbitals at r1 , expanding these functions for r < r1 as linear com-
binations of basis functions that satisfy fixed boundary conditions at r1 . The true
logarithmic derivative (or reciprocal of the R-matrix in multichannel formalism)
148 8 Variational methods for continuum states

is computed from Green’s theorem, despite the use of basis functions whose log-
arithmic derivatives at r1 have a fixed but arbitrary value. Because of the inherent
discontinuity of the boundary derivative this expansion tends to converge slowly,
and requires correction by an approximate method due to Buttle [54].
In nuclear physics, the specifically nuclear interaction is of short range, so that
full scattering information is obtained by matching the R-matrix at r1 to asymptotic
external wave functions. The method has been extended to electron–atom scatter-
ing, where long-range potentials are important, by combining basis expansion of
the channel orbital functions within r1 with explicit numerical solution of close-
coupling equations outside [50, 47]. This method makes it possible to process
algebraic equations defined by matrix elements of nonlocal operators within r1 ,
while solving the simple asymptotic close-coupling equations, without exchange,
outside r1 . This requires r1 to be large enough that exchange can be neglected
and that the nonlocal optical potential can be approximated by an asymptotic local
potential.
The R-matrix can be matched at r1 to external channel orbitals, solutions in
principle of external close-coupling equations, to determine scattering matrices.
Radial channel orbital vectors, of standard asymptotic form for the K-matrix,


− 12 1
w0 pq (r ) ∼ k p sin k p r −  p π δ pq ,
2


− 12 1
w1 pq (r ) ∼ k p cos k p r −  p π δ pq ,
2
are defined by integrating inwards for r > r1 . The asymptotic forms must be modi-
fied appropriately for Coulomb or fixed-dipole scattering. The function to be fitted
at r1 is

f ps (r1 ) = [w0 pq (r1 )δqs + w1 pq (r1 )K qs ].
q

The R-matrix relation between function value and gradient can be solved for the
K-matrix,
 −1
 

K st = − w1 ps (r1 ) − r1 R pq (r1 )w1qs (r1 )
p q
  sp


× w0 pt (r1 ) − r1 R pq (r1 )w0qt (r1 ) .
q pt

As written, these equations refer to open channels only. When external closed
channels are considered, an external closed channel orbital that vanishes as r → ∞
must be included for each such channel. The indices p, q, s run over all channels,
 8.4 Variational R-matrix theory 149

but the index t refers to open channels only. K st is a rectangular matrix, whose
open-channel submatrix can be shown to be real and symmetric.
The theory of the R-matrix can be understood most clearly in a variational for-
mulation. The essential derivation for a single channel was given by Kohn [202],
as a variational principle for the radial logarithmic derivative. If h is the radial
Hamiltonian operator, the Schrödinger variational functional is

∞
= f (h − ) f dr.
0

If f (r ) is an exact solution for r > r1 , this reduces the functional to the finite integral
considered by Kohn,

r1
r1 
 
1 2 1 2 1
λ = f (h − ) f dr = ( f ) + v(r ) − k f 2 dr − λ f 2 (r1 ),
0 0 2 2 2
obtained after integration by parts. The parameter λ is the logarithmic derivative
f  / f = (r1 R)−1 , defining the parameter R as the single-channel R-matrix.
For an infinitesimal variation δ f , with fixed λ, the variation of λ is

r1
δλ = 2 δ f (h − ) f dr + δ f (r1 )[ f  (r1 ) − λ f (r1 )].
0

This vanishes for unconstrained variations if and only if (h − ) f = 0 for 0 ≤

r ≤ r1 and λ = f  (r1 )/ f (r1 ).
If f (r ) is approximated by a finite expansion in linearly independent basis func-

tions {ηa }, such that f = a ηa ca , variation of λ with respect to the coefficients
ca gives the linear equation
 1
Aab cb = ληa (r1 ) f (r1 ), (8.13)
b
2

where

r1 
 
1   1
Aab = ηa ηb + ηa v(r ) − k 2 ηb dr. (8.14)
0 2 2
The parameter λ is uniquely determined as a consistency condition [202]. From
Eq. (8.13),
1  −1
ca = λ (A )ab ηb (r1 ) f (r1 ),
2 b

so that
 1 
f (r1 ) = ηa (r1 )ca = λ ηa (r1 )(A−1 )ab ηb (r1 ) f (r1 ).
a 2 a,b
150 8 Variational methods for continuum states

Unless f (r1 ) vanishes, this requires as a consistency condition

1 
R = (r1 λ)−1 = ηa (r1 )(A−1 )ab ηb (r1 ),
2r1 a,b

which is the standard formula for the single-channel R-matrix.

Several important aspects of this formula generalize immediately to multichannel
scattering. The matrix Aab is the Hermitian residue of  after integration by parts,
which separates off the boundary term as a surface integral. This boundary term
determines the R-matrix, which contains all information needed to extend an internal
solution by matching to external wave functions. Thus the internal Schrödinger
equation, which may contain nonlocal potentials such as an optical potential, is
solved by the typical bound-state method of expanding in a set of basis functions,
while the external problem may be solved by some completely independent method.
If r1 is large enough that asymptotic potentials are valid for r > r1 , then specialized
and highly efficient analytic methods may be used, without affecting the variational
expansion appropriate to the inner region.
A multichannel generalization of the variational R-matrix theory was derived by
Jackson [183]. In the context of the generalized close-coupling equations, incor-
porating an optical potential derived by Feshbach partitioning, this is in principle
an exact theory. Exact solutions of the coupled equations are assumed for r ≥ r1 ,
determined by specifying their values at r1 in terms of an R-matrix. For each chan-
nel p, projected onto spin and spatial symmetry (angular) functions, there are two
linearly independent radial functions. If these are denoted by u 0 ps (r ) and u 1 ps (r ),
respectively, for a global wave function indexed by s, the boundary conditions at
r1 chosen by Jackson are
− 12
u 0 ps (r1 ) = 0, u 0 ps (r1 ) = δ ps r1
+1
u 1 ps (r1 ) = δ ps r1 2 , u 1 ps (r1 ) = 0.
In this R-matrix theory, open and closed channels are not distinguished, but the
eventual transformation to a K-matrix requires setting the coefficients of expo-
nentially increasing closed-channel functions to zero. Since the channel functions
satisfy the unit matrix Wronskian condition, a generalized Kohn variational princi-
ple is established [195], as in the complex Kohn theory. In this case the canonical
form of the multichannel coefficient matrices is
α0 = I, α1 = R
and the radial channel orbital function is

f ps (r ) = (u 0 pt δts + u 1 pt Rts ), r ≥ r1 .
t
 8.4 Variational R-matrix theory 151

Then, by construction,
1
 −1
f ps (r1 ) = R ps r12 , f ps (r1 ) = δ ps r1 2 ,

in agreement with the defining equation for the R-matrix,



f ps (r1 ) = R pq r1 f qs (r1 ).
q

Since an exact solution is assumed outside r1 , the Schrödinger matrix functional

is

r1
†
st = f ps (r )m pq f qt (r ) dr,
p,q 0

whose variation is given by Eq. (8.9). This defines a trial R-matrix Rt = −m −1 11 m 10 ,

such that m 10 + m 11 Rt = 0 in this representation of the channel orbitals, and implies
that [R] = Rt − 2(Rt ) is stationary. This defines
 † 
[R] = −2 m 00 − m 10 m −1 11 m 10 . (8.15)

This expression is stationary with respect to variations about an exact solution.

Because radial derivatives of the functions u 1 ps all vanish at r = r1 , null values of
the determinant of m 11 characterize eigensolutions of the multichannel Schrödinger
equation that satisfy a homogeneous Neumann boundary condition on the R-matrix
sphere. In contrast to the Kohn formalism for the K-matrix, these null values are not
anomalous, but are a characteristic feature of any system of integrodifferential equa-
tions as energy or other parameters are varied. In one dimension, this simply means
that the logarithmic derivative passes through zero, as a continuous nondecreasing
function of energy.
The derivation by Kohn [202] can readily be extended to the multichannel
R-matrix. The underlying logic depends on the variational principle that the multi-
channel Schrödinger functional is stationary for variations of a trial function that
satisfies the correct boundary conditions if and only if that function satisfies the
Schrödinger equation. In a matrix notation, suppressing summations and indices,
the variational functional of Schrödinger is


1 †
= f † Â f − f (r1 )R −1 f (r1 ),
2r1
when integrated
 by parts to separate the surface term from the residual Hermitian
volume term f † Â f . Here R is a trial R-matrix, analogous to a Lagrange multiplier,
to be adjusted so that the trial wave function satisfies R-matrix boundary conditions
152 8 Variational methods for continuum states

at r = r1 . Â defines the modified Hamiltonian operator of Bloch [30], such that

1 d
h B −  = h −  + δ(r − r1 ) .
2 dr
Its representation in an orbital basis {ηa } defines the matrix A given by Eq. (8.14).
Given f = ηc, the functional becomes
 
1 †
 = c† A − η (r1 )R −1 η(r1 ) c.
2r1
Variation with respect to the coefficients c† , for fixed R, gives the linear equations
1 †
Ac = η (r1 )R −1 f (r1 ),
2r1
whose solution is
1 −1 †
c= A η (r1 )R −1 f (r1 ).
2r1
This implies that  = 0 and
1
f (r1 ) = η(r1 )c = η(r1 )A−1 η† (r1 )R −1 f (r1 ).
2r1
Unless A is singular, as it is at the null values of m 11 , this implies the consistency
condition, restoring summation indices,
1  pq †
R pq = ηa (r1 )(A−1 )ab ηb (r1 ).
2r1 a,b
Since matrix A is Hermitian by construction, the R-matrix is also Hermitian, and
can be made real by suitable choice of phase factors for the basis functions. Because
 vanishes, R pq is the stationary value of the matrix [R] in the basis {ηa }.
These early derivations [202, 183] have an important feature, not emphasized at
the time, that has subsequently turned out to be of great practical significance. No
conditions are imposed on the basis functions η other than regularity at the coordi-
nate origin and linear independence. The nonvariational theory [428], which uses
Green’s theorem, introduces a complete basis set in the interval 0 ≤ r < r1 , obtained
as the set of eigenfunctions of some model Hamiltonian, with a fixed boundary con-
dition, r1 η (r1 ) = bη(r1 ). Several standard derivations follow this logic [214, 47].
If the trial function f (r ) is expanded as a sum of functions all with the same log-
arithmic derivative at r1 , it must satisfy this same boundary condition. The value
of λ computed from the variational formula in the Kohn theory differs in gen-
eral from the imposed value b/r1 . Unless the energy parameter is an exact eigen-
value of the A-matrix, with this imposed boundary condition, the trial function in
standard R-matrix theory has a discontinuous derivative at r1 , if the function value
 8.4 Variational R-matrix theory 153

is matched at the boundary. This discontinuity leads to slow convergence of the

basis set expansion [187]. Standard procedure is to remove the gradient disconti-
nuity by a nonvariational Buttle correction [54]. This methodology has been used
in an extended series of ab initio calculations of atomic [50, 47] and molecular
[138, 139, 249] scattering cross sections. An alternative method [106, 220] is to
adjust the parameter b iteratively so that it agrees with the computed value of λ.
Other calculations have used the variational formalism directly, implemented with
orbital basis functions that are not constrained at r1 [310, 311, 299, 298].
The R-matrix radius r1 must be large enough that all nonlocal interactions can
be neglected or replaced by local asymptotic potentials outside it. In electron–
molecule scattering, a much smaller target radius r0 can be defined, within which
the total target electronic density has converged sufficiently to represent molecular
multipole moments and polarizabilities within the overall accuracy of the full scat-
tering calculation. In the intermediate region r0 ≤ r ≤ r1 , basis orbital functions for
the target system die off exponentially, while external multipole and polarization
response potentials are well approximated by their asymptotic forms. This behav-
ior has several very important implications for quantitative ab initio calculations.
Crucially, orbital basis sets and multiconfiguration expansions adapted to the target
system cannot adequately represent the intermediate region. Open-channel orbital
functions retain a constant amplitude throughout this region, and closed-channel
(polarization pseudostate) orbitals decay as inverse powers of r , not exponentially.
The most direct way to represent this region is to add orthogonalized fixed-energy
continuum orbital functions, two for each open channel, to the basis set. When
such basis orbitals are orthogonalized to the exponential basis set, this induces an
orthogonal projection of the (N +1)-electron variational basis, as indicated in the
definition of the (N +1)-electron wave function, Eq. (8.1). Fixed linear combinations
of Gaussian basis orbitals (with small exponents) can be constructed that model
spherical Bessel functions or Coulomb functions quite accurately in the inter-
mediate region [300]. An alternative basis for the intermediate region is the set
of numerical asymptotic functions (NAFs), continuum eigenfunctions of the model
Hamiltonian constructed from the long-range asymptotic potential functions, in-
tegrated outwards from r0 [311]. Since the centrifugal potential barrier increases
rapidly with orbital angular momentum , orbital functions with high  values be-
come insensitive to the region inside r0 . This justifies the asymptotic distorted-wave
(ADW) approximation [273], in which the R-matrix is computed by numerical inte-
gration of the asymptotic close-coupling equations from r0 to r1 . The high- channel
sectors of this matrix are used to augment a variational R-matrix computed in an
orbital basis cut off at some relatively smaller -value. NAF orbitals and the ADW
approximation were used in accurate calculations of fixed-nuclei e− –H2 scattering
cross sections [299, 298].
154 8 Variational methods for continuum states

8.4.1 Variational theory of the R-operator

Let τ denote a spatial cell of finite volume, enclosed by a convex bounding sur-
face σ . Variational R-matrix theory can be put into the general context of the
classical theory of elliptical partial differential equations by considering the three-
dimensional Schrödinger equation (for a given spin index) within such a bounding
surface [272]. For a scalar wave function ψ(x) with an implicit spin index, the
Schrödinger functional is an integral over the volume τ



∗ 1
 = ψ (h − )ψ = A − ψ ∗ ∇n ψ,
τ 2 σ

integrating the kinetic energy term by parts. ∇n ψ here is the outward normal gradient
on σ of the trial function ψ. The residual volume integral A, incorporating a Bloch-
modified [30] kinetic energy term, is

 
1 ∗ ∗
A= ∇ψ · ∇ψ + ψ (v − )ψ .
τ 2

In any representation basis {φa (x)}, the matrix

 
1 ∗ ∗
Aab = ∇φa · ∇φb + φa (v − )φb
τ 2

is Hermitian if the potential v is real (or nonlocal and Hermitian).

If ψ and ∇n ψ are given on the enclosing surface σ , the R operator is defined
such that


ψ(σ 1 ) = R(σ 1 , σ 2 )∇n ψ(σ 2 )d 2 σ 2 .
σ

This will be symbolized here by ψ(σ ) = R∇n ψ(σ ) = Rξ (σ ). If the normal gradi-
ent ξ is specified, this defines a classical Neumann boundary condition on σ , which
determines a unique solution of the Schrödinger equation in the enclosed volume
τ . The value of the boundary integral is


1
A1 = ψ ∗ ξ.
2 σ
The functional  must vanish for an exact solution of the Schrödinger equation
in τ . This implies that A1 and A are equal, and because A is real,

A = A1 = A∗1 .

Infinitesimal variations of ψ about an exact solution give


∗ 1 ∗
δA = δψ (h − )ψ + δψ ∇n ψ + cc ,
τ 2 σ
 8.4 Variational R-matrix theory 155

and


1
δ A1 = δψ ∗ ξ.
2 σ

In analogy to the Schwinger variational principle, consider the product functional

[A] = A1 A−1 A∗1 ,

which is real for any trial function. From the variations of A1 and A,




 
−1 1 ∗ −1 1 ∗ ∗
δ[A] = [A] A1 δψ ξ − A δψ ∇n ψ + δψ (h − )ψ + cc .
2 σ 2 σ τ

For variations about an exact solution, for which A = A1 = A∗1 , this reduces to




1 ∗ ∗
δ[A] = δψ (ξ − ∇n ψ) − δψ (h − )ψ + cc .
2 σ τ

For variations of ψ that are unconstrained throughout τ and on σ , this implies that
[A] is stationary if and only if (h − )ψ = 0 in τ and ∇n ψ = ξ on σ .

When ψ(x) = a φa (x)ca is expressed as a sum of basis functions, variations
δψ are driven by variations of the coefficients ca . If δ[A] vanishes, this implies


1


∗ ∗
φa (h − )φb cb = φ ξ− ∇n φb cb .
b τ 2 σ a b

Integrating by parts, in terms of matrix elements Aab ,



1
Aab cb = φa∗ ξ, all a.
b
2 σ

Explicitly,


1  −1
cb = [A ]ba φa∗ ξ, all b.
2 a σ

When these values are substituted into the definitions of A, A1 , and [A], all of these
quantities are equal to



1
[A] = ξ ∗ (1)φa (1)[A−1 ]ab φ ∗ (2)b ξ (2).
4 a,b σ1 σ2

In terms of the R-operator, the definition of A1 is

1
A1 = ξ ∗ (1)R(1, 2)ξ (2).
2 σ1 σ2
156 8 Variational methods for continuum states

Comparison with the stationary value of [A] = A1 indicates that

1
R(1, 2) = φa (1)[A−1 ]ab φb∗ (2).
2 a,b

Because [A] is stationary and the surface function ξ is arbitrary, the operator
R must itself be stationary. It is evidently real and symmetric by construction.
The basis functions φa (x) are required only to be linearly independent in τ . En-
forcing orthonormality through a fixed boundary condition on σ imposes a lack
of completeness with respect to determining the R-operator. The factor 1/r1 is
omitted from the definition given above, making the operator R dimensionally a
length. This cannot be avoided in general geometry, since no unique radius is de-
fined for a nonspherical surface. The usual R-matrix in spherical geometry is the
surface-harmonic representation of the R-operator as defined here, divided by the
dimensional scale factor r1 .

8.4.2 The R-operator in generalized geometry

In generalized geometry, it may be useful to introduce different coordinate systems
on different sectors of a boundary hypersurface σ . This extended definition was
introduced by Light and Walker [225] to define successive extensions of a prop-
agated R-matrix, and implicitly by Schneider et al. [363] in extending R-matrix
theory to vibronic interactions in molecules. It provides the formal basis for recent
applications of adiabatic phase-matrix theory to rovibronic excitation in molecules
[284]. It is assumed that coordinates can be defined such that the generalized kinetic
energy operator in the enclosed hypervolume τ takes the form
1
T̂ = ∇ · ∇,
2
for a generalized scalar product defined by

p·q = (1/µi ) pi qi .
i

The R-operator is defined on hypersurface σ by

ψ(σ 1 ) = R(σ 1 , σ 2 )∇ψ(σ 2 ) · dσ 2 .
σ

The classical Neumann boundary problem generalizes directly to the hyperspace

if the kinetic energy operator can be put into this canonical form. The argument
given above, when generalized to nonspherical geometry, remains valid. Given the
 8.4 Variational R-matrix theory 157

normal gradient ξ on σ , the stationary value of the variational functional is

1
[A] = dσ 1 ξ ∗ (σ 1 )R(σ 1 , σ 2 )ξ (σ 2 ) dσ 2 .
2 σ σ
Integration by parts of the Schrödinger functional is equivalent to using a Bloch-
modified Schrödinger equation


(H B − E)ψ(x) = L(x, σ)ψ(σ) dσ.
σ

Here the Bloch-modified Hamiltonian H B is obtained by integration by parts of the

kinetic energy integral,




∗ 1 ∗ 1
φα T̂ φβ dτ = ∇φα · ∇φβ dτ − φα∗ ∇φβ · dσ.
τ 2 τ 2 σ

This can be expressed in terms of a Bloch surface operator,

1
L = δ(x, σ)n(σ) · ∇
2
such that



1
φα∗ ∇φβ · dσ = φα∗ Lφβ dτ.
2 σ τ

Expansion in a linear independent orbital basis implies

1 
R(σ 1 , σ 2 ) = φα (σ 1 )(H B − E)−1 ∗
αβ φβ (σ 2 ).
2 α β

8.4.3 Orbital functional theory of the R-matrix

Variational R-matrix theory has been developed here in the context of multichannel
close-coupling equations derived by Feshbach partitioning from the (N +1)-electron
wave function given by Eq. (8.1). The short-range (exponentially decreasing) part of
this wave function is eliminated in this formalism, transformed into a multichannel
optical potential in the residual coupled orbital equations. Bound target states and
pseudostates have been projected out to define the coupled open and closed “free”
electronic channels. These equations take the form of a multichannel generalization
of the ground-state orbital Euler–Lagrange equations of orbital functional theory,
derived in Chapter 5 here. This relationship has been used to consider multipole
polarization response as an example of the OFT formalism for electronic correlation
[290].
158 8 Variational methods for continuum states

Specializing to electronically elastic scattering, there is only one target reference

state, which will be denoted by ", a normalized N-electron Slater determinant con-
structed from occupied OFT target orbitals. The target OEL equations are assumed
to contain a correlation potential providing an accurate approximation to the true
ground state. The (N +1)-electron reference state takes the form A"φκ , where A
is an antisymmetrizing operator. While all bound target orbital functions are nor-
malized within the R-matrix boundary r = r1 , the continuum orbital φκ extends
beyond this boundary. Its normalization requires some discussion.
Channel orbital functions are considered here to be orthogonalized to the bound-
state orbital basis. If this basis is complete within the target radius r0 , these or-
thogonalized channel orbitals must effectively vanish inside the target system. In
R-matrix formalism, if the effective potentials in the close-coupling equations do
not depend on the multichannel orbital functions, these orbitals may be arbitrarily
normalized [270]. Conventionally, they are normalized within the R-matrix volume
characterized by radius r1 , where they are matched to logarithmic derivatives of ex-
ternal continuum wave functions normalized to unit flux density. This conventional
normalization obscures the fact that a continuum orbital, extending throughout an
infinite volume, must vanish in any finite volume if normalized to unity over all
space. In the context of a continuum OFT, a continuum orbital may either be nor-
malized to an infinitesimal within the R-matrix boundary or normalized to unity
but assigned an infinitesimal occupation number. The latter option is followed here.
In a set of N +1 occupied orbitals, the orbital wave functions are orthonormal for

r ≤ r1 , but one occupation number n κ → 0+. Thus i n i = N , and bound–free
exchange is described by a Fock operator that depends only on the occupied target
orbitals. Similarly, in the theory developed here, correlation energy of the target is
not affected by a continuum orbital, but bound–free correlation energy acts as an
effective potential for the continuum electron.
In the case of a scattering resonance, bound–free correlation is modified by a
transient “bound” state of N +1 electrons. In a finite matrix representation, the
projected (N +1)-electron Hamiltonian Ĥ has positive energy eigenvalues, which
define possible scattering resonances if they interact sufficiently weakly with the
scattering continuum. In resonance theory [270], this transient discrete state is
multiplied by an energy-dependent coefficient whose magnitude is determined by
that of the channel orbital in the resonant channel. Thus the normalization of the
channel orbital establishes the absolute amplitude of the transient discrete state,
and arbitrary normalization of the channel orbital cannot lead to an inconsistency.

Bound–free correlation
For an N-electron target state, E c is a sum of pair-correlation energies. In a two-
electron system, such as atomic He, a major contribution to the correlation energy
 8.4 Variational R-matrix theory 159

arises from replacement of a product function by one in relative coordinates that

satisfies the Kato cusp condition at the singularity r12 → 0. In a CI representation,
this implies a slowly convergent sum of configurations, including relatively high
angular quantum numbers. Methods used to simplify this problem include use of
a free-electron-gas (FEG) local correlation potential [134] and an Ansatz wave
function that specifically corrects the cusp condition [70]. Complementary to this
short-range effect, the emphasis here is on long-range correlation potentials that
arise from multipole response of the target to a scattered electron.
The N-electron target wave function is coupled to a continuum orbital φκ for
which n κ → 0. Vanishing n κ implies that the continuum electron does not modify
the effective Hamiltonian G that acts on occupied target orbitals (n i = 1). G also
acts on φκ because n κ → 0 cancels out of the functional derivatives in n κδδφ E
∗ . This
κ
implies that φκ is orthogonal to the occupied target orbitals. The result is to augment
standard static-exchange equations with a nonlocal correlation potential v̂c .
From Eq. (5.7), and Janak’s theorem [185], the contribution of correlation energy
to the mean energy of the continuum orbital within the R-matrix boundary is
 
(κ|v̂c |κ) = nj (1 − n c )(1 − n b )(κ j|ū|cb)(cb|c̄|κ j)
j c<b
 
− nk n j (1 − n b )(k j|ū|κb)(κb|c̄|k j), (8.16)
k< j b

where all integrals are evaluated for r ≤ r1 . The second line of this general formula
is a correction to target correlation energy due to removal of φκ from the unoccupied
set. The first line can be shown to include a multipole polarization potential.
To model the effect of close-coupling equations including a closed pseudostate
channel, suppose that a pseudostate orbital φ p j represents the first-order pertur-
bation of a particular target orbital φ j in a multipole field of order λ, and that
p
the pseudostate excitation energy is E j − E 0 . Because φ p j is selected to interact
strongly within the inner radius r0 , the residual orthogonalized set of unoccupied
orbitals φq remain approximately complete in r0 ≤ r ≤ r1 for functions orthogonal
to occupied orbitals. The close-coupling equations imply that a closed-channel or-
bital φqκ coupled to φ p j is approximately of the form of a slowly decaying function
times the open-channel orbital φκ . Hence the average energy of the two-electron
p
virtual excitations j, κ → p j , q can be approximated by E j − E 0 . This argument
justifies a closure approximation, and the first line of Eq. (8.16) becomes
 
(κ|v̂c |κ)  nj (1 − n q )(κ j|ū|q p j )(q p j |c̄|κ j)
j q
   p −1 
− n j κ|( j|u| p j ) E j − E 0 ( p j |u| j)|κ . (8.17)
j
160 8 Variational methods for continuum states

Table 8.1. Partial wave phase shifts for He

SE SEP
k (au) Present Ref. [383] Present Ref. [383]
=0
0.3 2.7049 2.7037 2.7424 2.7459
0.5 2.4357 2.4325 2.4812 2.4836
0.7 2.1943 2.1963 2.2433 2.2518
0.9 1.9836 1.9808 2.0339 2.0426
=1
0.3 0.0140 0.0108 0.0340 0.0297
0.5 0.0447 0.0426 0.0866 0.0847
0.7 0.0980 0.0947 0.1629 0.1567
0.9 0.1570 0.1552 0.2366 0.2311

This is the diagonal matrix element of a multipole polarization potential. For

multipole index λ > 0, transition matrix elements ( p j |u| j) vary as r λ for small
r and as 1/r λ+1 for r  r0 , so that the multipole polarization potential varies as
r 2λ and 1/r 2λ+2 , respectively, in these limits. This approximate polarization poten-
tial vanishes as r → 0, eliminating the need for an ad hoc cutoff for small
r . When λ = 1, the spherically averaged static dipole polarizability is αd =
( j|x| p j ) · ( p j |x| j)(E j − E 0 )−1 , a sum over 2λ + 1 = 3 dipole pseudostates. The
2 p
3
isotropic polarization potential implied by Eq. (8.17) takes the well-known form
−αd /2r 4 for large r . Due to the orthogonality conditions, this approximation is
valid only for the long-range part of bound–free correlation, outside the target
charge distribution. The fact that the derived multipole polarization potential van-
ishes for r → 0 may justify simply adding it to a modeled short-range correlation
potential, or fitting at an estimated r0 to such an effective potential [133].
Table (8.1) shows results of test calculations of e–He partial wave phase shifts,
compared with earlier variational calculations [383]. The polarization pseudostate
was approximated here for He by variational scaling of the well-known hydro-
gen pseudostate [76]. The present method is no more difficult to implement for
polarization response (SEP) than it is for static exchange (SE).
 9
Electron-impact rovibrational excitation of molecules

The principal references for this chapter are:

[49] Burke, P.G. and West, J.B., eds. (1987). Electron–Molecule Scattering and
Photoionization (Plenum, New York).
[82] Domcke, W. (1991). Theory of resonance and threshold effects in electron–
molecule collisions: the projection-operator approach, Phys. Rep. 208, 97–188.
[178] Huo, W.M. and Gianturco, F.A., eds. (1995). Computational Methods for Electron–
Molecule Collisions (Plenum, New York).
[215] Lane, N.F. (1980). The theory of electron–molecule collisions, Rev. Mod. Phys. 52,
29–119.
[284] Nesbet, R.K. (1996). Nonadiabatic phase-matrix method for vibrational excitation
and dissociative attachment in electron–molecule scattering, Phys. Rev. A 54,
2899–2905.
[363] Schneider, B.I., LeDourneuf, M. and Burke, P.G. (1979). Theory of vibrational
excitation and dissociative attachment: an R-matrix approach, J. Phys. B 12,
L365–L369.

Electron–molecule scattering data, observed experimentally or computed with

methodology available as of 1980, was reviewed in detail by Lane [215]. If there
were no nuclear motion, electron–molecule scattering would differ from electron–
atom scattering only because of the loss of spherical symmetry and because of the
presence of multiple Coulomb potentials due to the atomic nuclei. This is already
a formidable challenge to theory, exemplified by the qualitative increase in com-
putational difficulty and complexity between atomic theory and molecular theory
for electronic bound states. While bound-state molecular computational methods
have been extended to fixed-nuclei electron scattering [49, 178], an effective and
computationally practicable treatment of rovibrational (rotational and vibrational)
excitation requires a significant and historically challenging extension of bound-
state theory. The small ratio between electron and nuclear masses is exploited in
bound-state theory through the Born–Oppenheimer separation [31], leading to the
qualitative physical principle that in the lowest order of a perturbation expansion,

161
162 9 Electron-impact rovibrational excitation of molecules

transfer of kinetic energy between electronic and nuclear motion can be neglected.
1
In the expansion parameter (m/M) 4 , where M is a typical nuclear mass, elec-
tronic energy is of zeroeth order, vibrational energy of second order, and rotational
energy of fourth order. Far from neglecting these Born–Oppenheimer corrections,
a valid theory of electron-impact rovibrational excitation must compute the scat-
tering effects of such terms with quantitative accuracy, because the phenomenon
of interest is precisely such an energy transfer between electronic and nuclear
motion.
Rotational level spacing is in fact so small for most molecules that it can be
resolved only by precise spectroscopic techniques. Typical electron scattering data
does not have sufficient energy resolution, and most published data is for rotationally
averaged scattering cross sections. Special methods based on an adiabatic theory
of slowly moving nuclei are valid under these circumstances, not dependent on
variational theory and methods. However, near excitation or scattering thresholds,
rotational analysis may play an essential role in interpreting complex scattering
phenomena due to dynamical long-range potentials, such as the dipole potential
for heteropolar diatomic molecules, quenched by molecular rotation. For these
reasons, the present discussion of computational methodology will concentrate on
threshold effects characteristic of low-energy electron scattering and on vibrational
excitation. The phenomenon of dissociative attachment, when an incident electron is
captured by a target molecule, which then dissociates to produce an ionic fragment,
is a particular example of such energy transfer, and will be considered here as an
application of rovibrational threshold theory.
An electron scattering resonance, in abstract but general mathematical terms, is
characterized by a singularity of the scattering matrix when analytically continued to
a complex energy value res − 2i γ for relatively small width parameter γ > 0 [270].
In simple effective-potential models, a so-called shape resonance occurs when an
electronic energy level in a potential well is degenerate with an energy continuum
outside a potential barrier that confines the well. An electron initially bound in
the state corresponding to this energy level leaks out through the potential barrier
with a time constant h/γ . Such poles of the analytically continued scattering matrix
can also cause significant energy-dependent Wigner cusp or rounded step structures
[427, 32] in scattering cross sections just below or just above the energy threshold at
which a continuum becomes energetically accessible. This behavior is characteristic
of long-range potentials and does not require a well-defined potential barrier [270].
A typical phenomenon is a virtual state, which occurs in simple potential models
at the threshold for a continuum with orbital angular momentum  = 0, hence no
centrifugal barrier.
Resonance and threshold structures in fixed-nuclei electron-scattering cross sec-
tions are replicated, with characteristic energy shifts, in rovibrational excitation
 9.1 The local complex-potential (LCP) model 163

cross sections. One of the most striking examples is provided by the prominent
multiple peaks observed in both the electronically elastic and vibrational excita-
tion cross sections for electron scattering by the N2 molecule [368, 143, 95, 215].
These peaks are associated with a fixed-nuclei resonance at approximately 2 eV,
but their width arises from electron–vibrational coupling and their spacing does
not correspond precisely to the vibrational level structure of either the neutral
target molecule or of a vibrating transient negative ion. A quantitative theory of
these excitation structures has been developed through several advances in both
formalism and computational technique, ultimately based on variational theory of
the interacting electron–vibrational system. Another striking example is the ob-
servation of excitation peaks associated with each successive vibrational excita-
tion threshold in electron scattering by dipolar molecules such as HF, HCl, and
HBr [346]. Because of the electric dipole moment of such target molecules, if
low-energy scattering theory is to be relevant it must provide a detailed analy-
sis of rotational excitation and of rotational screening of the long-range dipole
potential.

9.1 The local complex-potential (LCP) model

Considerable understanding of molecular resonance and threshold phenomena has
come from the conceptually simple model of an electronic resonance state that de-
termines a complex-valued effective internuclear potential function, characterized
by a decay width whenever the real part of the resonance energy lies above the
fixed-nuclei potential of the target molecule. The vibrational levels of this complex
potential are a first approximation to resonance energies such as those observed in
N2 . As shown in the “boomerang” model of Herzenberg [167], the implied lifetime
of these model vibrational states is so short that they do not survive a single oscilla-
tion of a classical wave packet. Hence a dynamical model arises in which excitation
from the ground-state vibrational state of the neutral molecule to a wave-packet state
of the transient negative ion is followed by a single vibration out and back, travers-
ing a region of relatively long lifetime. Because the decay width increases rapidly
for small internuclear distances, this wave packet decays into vibrational states of
the neutral molecule as it makes this return trip, and disappears before completing
a full oscillation. With appropriate parametrization, this model gives a convincing
explanation of the observed vibrational resonance peak shapes and separations in
e−N2 scattering [24, 91].
The complex-potential model originated as a qualitative explanation of dissocia-
tive attachment [18]. A fixed-nuclei electronic resonance is described by a complex
potential whose width goes to zero at the crossing point with the molecular ground-
state potential, and remains zero if the resonance potential remains below the target
164 9 Electron-impact rovibrational excitation of molecules

potential for large internuclear separation. In this case, an outgoing wave packet
becomes a true bound state of the negative ion after crossing outwards, and disso-
ciates without returning. Excitation and decay processes inside the crossing point
are the same as in the boomerang model. An incident electron, captured into ex-
cited ionic vibrational levels, creates a vibronic wave packet that moves outwards
while decaying into vibrational excitations, but eventually simply crosses outwards
as a molecular negative ion that dissociates into stable fragments. Applied to the
dissociative attachment of H2 [16, 19, 415], this parametrized model accurately
reproduces available experimental data.
A very significant practical problem with the LCP model is the lack of a unique
and well-defined theory of the required resonance state when its decay width be-
comes large. In the case of H−2 , serious discrepancies between various parametrized
or computed values of the energy and width of the 2 'u+ shape resonance in the LCP
model of dissociative attachment indicate that a more fundamental, first-principles
theory is needed to give fully convincing results. The difficulty is that as the
decay width becomes large, the concept of a well-defined LCP model becomes
questionable, since the scattering resonance fades into the background scattering
continuum [271]. Analytic theory implies an energy shift which may qualitatively
change the character of the assumed resonance potential curve [59, 82].

9.1.1 The projection-operator method

The semiclassical picture of nuclear motion inherent in the LCP model has a
quantum-mechanical foundation that is most directly developed in the projection-
operator formalism of Feshbach [115, 116, 270]. Domcke [82] reviews the applica-
tion of this formalism to resonant effects in electron–molecule scattering. Neglect-
ing the kinetic energy of nuclear motion, T̂n , an electronic resonance at nuclear
coordinate(s) q is characterized by a pole of the analytically continued scattering
matrix S(q, ) at complex energy  = res (q) − 12 iγ (q), where γ is the decay width
of the resonance. The set of values res (q) defines the real part of the LCP effec-
tive potential curve. The Feshbach formalism in electron–molecule scattering is
ordinarily applied to the static-exchange model of fixed-nuclei scattering, coupled
to the Hamiltonian Hn for nuclear motion [67, 314, 315, 15]. The electronic wave
function is represented by a model state " (Slater determinant), whose orbital func-
tions are orthogonalized to a normalized, localized orbital function φd (q; x) that
interacts with a continuum orbital wave function ψk to produce the scattering reson-
ance. For fixed nuclei, this formalism is exactly the same as resonance theory for
atoms, resulting in a complex-valued nonlocal optical potential. It is most directly
described by a Green function that is formally orthogonalized to the postulated
localized function φd [270, 82].
 9.1 The local complex-potential (LCP) model 165

In the LCP model of vibrational excitation, parametrized functions r es (q) and

γ (q), together with the ground-state potential curve, V0 (q), suffice to determine
cross sections averaged over rotational substructure. The full Feshbach formalism
requires parametrized transition matrix elements between φd and orthogonalized
background continuum orbitals φ̂ k , as well as V0 (q) and the mean electronic energy
Vd (q) of a postulated discrete state φd . In a one-electron model [82], the Feshbach
optical potential is added to the effective electronic Hamiltonian for static exchange.
This formalism is readily extended to an orbital-functional theory that includes elec-
tronic correlation [290]. The transition element Vdk , parametrized as a function of
both q and the electronic continuum energy , models the (N +1)-electron matrix
element (!d |H |A#0 ψk ), where !d is a postulated discrete (N +1)-electron state,
#0 is the target electronic ground state, and ψk is a continuum orbital function
at energy  = 12 k 2 . The full function A#0 ψk must be orthogonal to !d . Thus the
formalism defined by such matrix elements applies to the original case consid-
ered by Feshbach, where !d is a core-excited state such as He:(1s3s 2 )2S at
22.45 eV excitation energy, as well as to the explicitly modeled one-electron
attached state.
The vibronic Hamiltonian in the one-electron model is H = H0 + V . The kernels
of these operators are

 
∗ 1 2 ∗
h 0 = φd [T̂ n + Vd (q)]φd + k dk d)k φ̂ k φ̂ n + V0 (q) + k φ̂ k ;
2


v = k dk d)k φd Vdk (q)φ̂ ∗k + hc.

The coupled Schrödinger equations can be projected onto the φd · · · φd∗ subspace
by Feshbach partitioning, giving an equation for the coefficient function χd (q)
in the component φd χd (q) of the total wave function. The effective Hamiltonian
in this equation is T̂ n + Vd (q) + V̂opt , which contains an optical potential that is
nonlocal in the q-space. This operator is defined by its kernel in the φd · · · φd∗
subspace,


−1
 1 ∗
vopt (q, q ; E) = k dk d)k Vdk (q) E − k 2 − T̂ n − V0 (q) + iη Vdk (q  ).
2

This defines a nonlocal complex energy shift vopt =  − 12 i( such that

  −1
 1
(q, q ; E) = k dk Vdk (q)χv (q) E − k 2 − E v χv∗ (q  )Vdk
∗
(q  ),
v 2



((q, q  ; E) = 2π Vdkv (q)χv (q)χv∗ (q  )Vdk
∗
v
(q  ).
v
166 9 Electron-impact rovibrational excitation of molecules

Here kv2 = 2(E − E v ) for the bound or continuum vibrational state indexed by v.
Thus the Feshbach formalism implies energy-dependent, nonlocal energy-shift and
width functions for a resonance.
Neglecting nonresonant scattering, the resonant contribution to the transition
matrix is [82]
 ∗ 
T (k f , v f ; ki , vi ) = v f Vdk f
G(E)Vdki
vi ,

deduced from the Lippmann–Schwinger equation. Here G(E) is the resolvent

operator in the φd · · · φd∗ space,

G(E) = (E − T̂ n − Vd (q) − V̂ opt + iη)−1 ,

evaluated in the limit η → 0+ for outgoing-wave boundary conditions. The tran-

sition matrix can be computed as
 ∗ 
T (k f , v f ; ki , vi ) = v f Vdk f
χd ,

since the projected equation for the coefficient function χd (q) implies χd (q) =
G(E)Vdki χvi . This projected equation is

{T̂ n + Vd (q) + V̂ opt − H }χd (q) = −Vdki χvi (q), (9.1)

which is nonlocal in the nuclear coordinates [15].

Although in earlier applications Vd (q) + V̂ opt was approximated by an empirical
local complex potential (LCP), more recent work has solved this nonlocal equation
directly [82], obtaining detailed resonance excitation and near-threshold cross sec-
tions in excellent agreement with experimental data. Model calculations, using the
nonlocal formalism, indicate that a local approximation can yield accurate results for
vibrational excitations [59, 258]. The local model is inadequate for very broad res-
onances [259] and for near-threshold singularities [84, 86]. The neglect of nonres-
onant scattering implies significant discrepancies for elastic scattering. Despite the
practical success of the projection-operator formalism for resonant scattering, the
underlying electronic theory is still restricted to a static-exchange model, augmented
by an optical potential derived from empirical functions Vd (q) and (d (q) or Vdk (q).

9.2 Adiabatic approximations

The small mass ratio m/M suggests various levels of approximation based funda-
mentally on the inherently small kinetic energy of nuclear motion T̂ n . The physical
concept that justifies the LCP model is that the electronic wave function adjusts
essentially instantaneously to displacements of the nuclei, because electronic vel-
ocities are much greater than nuclear velocities in a time-dependent semiclassical
 9.2 Adiabatic approximations 167

model. Thus nuclear motion is described in an adiabatic picture, assuming instanta-

neous electronic relaxation. A more fundamental rationalization can be based on a
generalized independent-particle model, including nuclei as well as electrons. This
implies a formally exact theory if electronic correlation and electron–nuclei inter-
actions are included. However, the small mass ratio m/M justifies simplification
of the self-consistent mean field that acts on each nucleus even though the nuclear
motion is described by quantum mechanics and T̂ n is treated as a Schrödinger
operator that acts on rovibrational wave functions. Whether or not this mean field
can be reduced to a static electronic potential computed for fixed nuclei depends on
the limit of linear response theory for the electrons as m/M → 0. Thus a complete
quantitative theory must use the full variational theory of interacting electrons, but
can be expected to justify significant simplifications in the treatment of nuclear
motion.
Because molecular rotational kinetic energy and rotational level spacings are
small in all cases, a rotationally adiabatic model has been widely and successfully
used for electron–molecule scattering [215]. The essence of this adiabatic nuclei
(ADN) approximation [313, 63, 64, 401] is that a scattering amplitude or matrix
computed for fixed nuclei is treated as an operator in the nuclear coordinates. Then
the rotational state-to-state scattering amplitude or matrix is estimated by matrix
elements of this operator in the basis of rotational states. The basic assumption is
that of the Born–Oppenheimer separation: the commutator between the fixed-nuclei
scattering operator and the operator T̂ n is neglected. This approximation becomes
difficult to justify only when the precise energy of individual rotational levels is
important, near thresholds or in the presence of long-range potentials. Because of
the well-defined internuclear geometry in a body-fixed reference frame, fixed-nuclei
scattering matrices are most directly computed in this frame. In contrast, rotational
scattering structure is observed in a fixed laboratory frame. Unless an explicit
rotational frame transformation is carried out, strong rotational coupling occurs as
a computational artifact [215]. For this reason, applications of the rotational ADN
model require such a frame transformation [62].
The ADN approximation has been much less successful for vibrational excita-
tion, as might be expected from the much larger vibrational level spacings and the
strong variation of fixed-nuclei resonance parameters with nuclear displacements.
This motivated the proposal [61] of a hybrid close-coupling model. As applied
to e−N2 vibrational excitation, this model combines fixed-nuclei electronic close-
coupling calculations for nonresonant body-frame molecular symmetry states with
extended vibronic close-coupling calculations in the 2 g resonant symmetry of
N−2 . Although multipeaked vibrational excitation and vibrationally elastic cross
sections are obtained in qualitative agreement with experiment, the peak shapes
and spacings are not in good agreement. This can be attributed to truncation of the
168 9 Electron-impact rovibrational excitation of molecules

electronic partial-wave expansion and of the limited number of vibrational states

included in the close-coupling basis. A deeper problem is that any complete set of
vibrational states, such as the eigenfunctions of a parametrized Morse potential,
must include the vibrational continuum. No practical way has been found to do this
in the close-coupling formalism.

9.2.1 The energy-modified adiabatic approximation (EMA)

In many scattering processes, energy levels of the target system are split by a
perturbation that is weak relative to the interaction responsible for the scattering. The
adiabatic approximation neglects the effect on threshold scattering structures and
on resonances of the energy-level splitting and energy shifts of the perturbed target
states. Especially for threshold structures, this can lead to qualitatively incorrect
results. For example, transition matrix elements computed below a rovibrational
excitation threshold do not vanish if only the fixed-nuclei target energy is taken
into account. This qualitatively incorrect behavior of adiabatic cross sections near
rotational excitation thresholds can be compensated simply by modifying electron
momenta in the adiabatic cross-section formula [64]. With these corrections, the
ADN theory appears to be adequate for rotational excitation [144].
Much larger anomalous effects occur in vibrational excitation because the energy
shifts are larger. The energy-modified adiabatic approximation (EMA) [267] was
introduced in order to provide a systematic treatment of such effects, while retaining
the computational efficiency of the adiabatic approximation. The usual adiabatic
approximation is modified by allowing for the dependence of unperturbed scat-
tering matrices on the kinetic energy of the perturbed target state, using formulas
that are qualitatively correct for threshold and resonance structures. For molecules,
this means that for target geometry determined by a generalized coordinate q the
energy (q) = E − V (q) of a continuum electron must be replaced in principle by
op (q) = E − Hn (q), an operator in the nuclear coordinates. Fixed-nuclei calcula-
tions produce scattering matrices that are functions of numerical parameters q, .
In the EMA, the integrals that project these matrices onto rovibrational states are
evaluated by approximations that replace the parameter  by the operator op . The
simplest such approximation, for diagonal matrix elements in a rovibronic state
indexed by µ, is to replace  by µ = E − E µ . If nondiagonal matrix elements
between states indexed by µ and ν are evaluated for the geometric mean energy
1
µν = [(E − E µ )(E − E ν )] 2 , this implies correct threshold behavior for general
short-range potentials. When applied to the energy-denominator characteristic of a
fixed-nuclei molecular resonance, this state-dependent modification of  to op was
shown to give qualitatively correct results for the multipeaked vibrational excita-
tion structures observed in e−N2 scattering [267]. Because the EMA formalism
 9.3 Vibronic R-matrix theory 169

replaces fixed-nuclei scattering matrices by operators that are represented by rovi-

bronic scattering matrices, a single fixed-nuclei or threshold structure becomes a set
of overlapping scattering structures, displaced with possibly irregular energy shifts
by the discrete vibronic energy level structure of the target molecule [267]. This
repetition and displacement of underlying structures is characteristic of observed
electron scattering cross sections for molecular targets. A more recent extension of
the EMA formalism to the context of variational R-matrix theory is discussed below.

9.3 Vibronic R-matrix theory

Recognizing that exact quantum electron–molecule scattering theory for interacting
nuclei and electrons is and will remain computationally intractable, except for the
simplest diatomic molecules, Schneider [361] initiated reconsideration of the Born–
Oppenheimer approximation as a logical foundation for the adiabatic nuclei (ADN)
formalism. For a long-lived (narrow) resonance, neglect of vibrational derivatives
of the fixed-nuclei electronic wave function should be no less valid for electron
scattering at low energies than it is for molecular bound states. This argument is
the basis of an electronic R-matrix methodology in which eigenstates, defined
by eigenvalues of the Bloch-modified electronic Hamiltonian from which the
R-matrix is constructed, are used to define effective molecular potential functions,
parametrized by nuclear coordinates [363]. The small mass ratio m/M justifies
neglecting derivatives of the electronic eigenfunctions of this fixed-nuclei Bloch-
modified Hamiltonian with respect to the nuclear coordinates.
This work introduced the concept of a vibronic R-matrix, defined on a hyper-
surface in the joint coordinate space of electrons and internuclear coordinates. In
considering the vibronic problem, it is assumed that a matrix representation of
the Schrödinger equation for N +1 electrons has been partitioned to produce an
equivalent set of multichannel one-electron equations coupled by a matrix array of
nonlocal optical potential operators [270]. In the body-fixed reference frame, par-
tial wave functions in the separate channels have the form # p (q; x N )Y L (θ, φ)χv (q),
multiplied by a radial channel orbital function ψ(q; r ) and antisymmetrized in the
electronic coordinates. Here # is a fixed-nuclei N -electron target state or pseudo-
state and Y L is a spherical harmonic function. Both # and ψ are parametric functions
of the internuclear coordinate q. It is assumed that the target states # for each value
of q diagonalize the N -electron Hamiltonian matrix and are orthonormal.
An electronic R-matrix radius a is chosen such that exchange can be neglected
for r > a. An upper limit qd for the internuclear coordinate q is chosen so that
a dissociating electronic state "d is bound for q ≥ qd . This defines a vibronic
hypercylinder [284] with two distinct surface regions: an electronic wall with
r = a for 0 ≤ q ≤ qd and a dissociation cap defined by the enclosed volume of
170 9 Electron-impact rovibrational excitation of molecules

the electronic sphere for q = qd . For nondissociating molecules, qd should be large

enough to enclose the highest vibrational state to be considered. The R-matrix is
defined by matrix elements of the variational operator R in a complete basis of
surface functions [284], such as the spherical harmonics on the wall surface.
A matrix of operators R pp is defined by projection of R into the multichannel
representation indexed by N -electron target states # p . This defines the vibrational
excitation submatrix of the R-matrix as
( pLv|R| p  L  v  ) =





dq dq  d) d) χv∗ (q)Y L∗ (θ φ)R pp (qθ φ; q  θ  φ  )Y L  (θ  φ  )χv (q  ).

This matrix can be computed from the general variational formula derived in
Chapter 8, using a complete set of vibronic basis functions

!α (x N +1 , q) = A# p (q; x N )Y L (θ, φ)ψα (q; r )χα (q).

It is assumed that target states # p are indexed for each value of q such that a smooth
diabatic energy function E p (q) is defined. This requires careful analysis of avoided
crossings. The functions χα should be a complete set of vibrational functions for
the target potential V p = E p , including functions that represent the vibrational con-
tinuum. All vibrational basis functions are truncated at q = qd , without restricting
their boundary values. The radial functions ψα should be complete for r ≤ a.
Free boundary conditions are not allowed in the formulation of the theory
by Schneider et al. [363], based on the nonvariational theory of Wigner and
Eisenbud [428]. Specific boundary conditions are imposed using a Bloch opera-
tor. This determines boundary conditions correctly at energy poles of the R-matrix
determinant, but requires a Buttle correction [54] for energy values between such
poles [363]. This becomes problematic for the internuclear coordinate, because the
physical model of the dissociating state is a complex potential function for q ≤ qd ,
so that fixed boundary conditions imply complex energy eigenvalues. Nevertheless,
in the usual case that R-matrix poles are associated with homogeneous Neumann
boundary conditions on the R-matrix boundary, the Wigner–Eisenbud theory and
variational R-matrix theory derive the same equations for the vibronic R-matrix.
The projection integrals on the electronic wall are

qd
( pLv|α) = (A# p χv |!α )r =a = χv∗ (q)ψα (q; a)χα (q) dq.
0

In agreement with [363], the R-matrix for vibrational excitation is

1
( pLv|R| p  L  v  ) = ( pLv|α)(H B − E)−1   
αβ (β| p L v ),
2 α,β
 9.3 Vibronic R-matrix theory 171

where H B is the Bloch-modified vibronic Hamiltonian. This requires vibrational

dχ ∗
1
kinetic energy matrix elements to be evaluated as the Hermitian form 2µ dq dqv dχ
dq
v
.
The derivation up to this point involves no approximations if the vibronic basis
set is complete in the closed hypervolume, including its surface. If a dissociation
channel exists, it can be approximated by projection onto a single diabatic state
"d (q; x N +1 ), assumed to be well defined as a discrete state on the cap surface
q = qd . The projection integrals on this surface are
(d|α) = ("d |!α )q=qd .
The R-matrix connecting wall and cap surfaces, obtained by projecting the R
operator onto both surfaces, is
1
(d|R| pLv) = (d|α)(H B − E)−1
αβ (β| pLv),
2 α,β

which determines excitation–dissociation transitions. The R-matrix in the cap sur-

face is
1
(d|R|d) = (d|α)(H B − E)−1αβ (β|d).
2 α,β

This is the reciprocal of the logarithmic derivative of the wave function χd (q) in the
dissociation channel, for q = qd . At given total energy E these R-matrix elements
are matched to external scattering wave functions by linear equations that determine
the full scattering matrix for all direct and inverse processes involving nuclear
motion and vibrational excitation. Because the vibronic R-matrix is Hermitian by
construction (real and symmetric by appropriate choice of basis functions), the
vibronic S-matrix is unitary.
Schneider et al. [363] use Born–Oppenheimer vibronic basis functions as indi-
cated above, and neglect Born–Oppenheimer corrections determined by the inter-
nuclear momentum operator acting on the electronic wave function. Radial basis
functions ψk (q; r ) correspond to R-matrix pole states, whose energy values E k (q)
define an indexed vibrational potential. Vibrational basis functions χkµ (q) are com-
puted as eigenfunctions of the corresponding Hamiltonian. Since resonance states
are not treated by projection, the method depends on effective completeness of the
double expansion in electronic and vibrational eigenfunctions. The first application
of this method, to the multipeaked vibrational excitation structure observed in e−N2
scattering, was remarkably successful [364], in much closer agreement with exper-
iment than were comparable calculations using the hybrid ADN close-coupling
formalism [61]. Electronic wave functions with fixed boundary conditions at r = a
were used for the four lowest R-matrix pole states, and vibrational wave func-
tions were computed without considering the vibrational continuum. Boundary
172 9 Electron-impact rovibrational excitation of molecules

values on the cap surface are not relevant since dissociation is not involved. A
Buttle correction for the electronic basis was computed using adiabatic theory.
These calculations were later extended up to 30 eV [138] scattering energy, in-
cluding differential cross sections for elastic scattering and vibrational excitation.
The original method was modified to use a fixed electronic basis set for all inter-
nuclear distances, in order to mitigate problems arising from avoided crossings of
R-matrix pole state potential curves. Vibrational wave functions were represented
in a basis of shifted Legendre polynomials. These procedures were used for cal-
culations of integral and differential cross sections in e−HF scattering [250] and
for similar calculations on HCl [251], both examples of threshold excitation peaks
due to the molecular dipole moment. Because results computed with this method,
based on a fixed boundary condition at q = qd , depend strongly on the choice of qd ,
a modified theory has been proposed in which the evidently successful R-matrix
theory of vibrational excitation is combined with resonance-state theory for nu-
clear motion [103]. The approximation of neglecting derivatives of electronic wave
functions with respect to internuclear coordinates appears to be satisfactory in all
of these applications [362].

9.3.1 Phase-matrix theory

The general success of projection-operator methods indicates that quantitative cal-
culations can be based on the strategy of separating singularities of scattering
matrices from a smoothly varying background [82]. The R-matrix is in general a
real symmetric matrix with isolated real energy poles, analogous to the K -matrix
in scattering theory. For the K -matrix, these poles have no special physical signifi-
cance, simply indicating that the sum of eigenphases of the corresponding unitary
S-matrix passes through an odd multiple of π/2 radians. Similarly, the choice of
pole states of the R-matrix to define “vibrational” potential functions in the method
of Schneider et al. [363] is an arbitrary construction whose principal effect is to pro-
duce linearly independent vibronic basis states that can be extended to a complete
set inside the R-matrix hypersurface.
In order to establish a better-motivated connection to resonance theory, the fixed-
nuclei R-matrix can be converted to a phase matrix ", defined such that tan " =
k(q)R, or to the corresponding unitary matrix
U = (I + ik(q)R)/(I − ik(q)R),
the analog of the scattering S-matrix, whose complex energy poles define scattering
1
resonances and bound states. The factor k(q) = [2(E − V0 (q))] 2 makes " dimen-
sionless. A resonance is characterized for real energies by a point of most rapid
increase of the eigenphase sum [389, 270], which is the trace of the matrix tan−1 K .
 9.3 Vibronic R-matrix theory 173

Evaluated at specified r = a, the phase matrix " of the R-matrix has analogous
properties. For real energies, " has a monotonically increasing trace, which can be
made continuous by suitable choice of the branch of each multivalued eigenphase
function, adding or subtracting integral multiples of π at each energy value. A
point of maximum slope defines a “precursor resonance” [284] independently of
any specific physical model, which corresponds to a pole of the analytically con-
tinued S-matrix as the R-matrix boundary is increased. The proposed methodology
uses time-delay analysis [389] to separate a given fixed-nuclei phase matrix into
a rapidly varying resonant part and a slowly varying background part. The rovi-
bronic phase matrix is constructed by applying resonance theory to the resonant
phase matrix only, treating the nonresonant background part by energy-modified
adiabatic theory. By separating out rapid variations of " this methodology reduces
the completeness requirement from the double expansion inherent in the method of
Schneider et al. [363] to a single expansion for a well-defined precursor resonance
state. Vibrational completeness is obtained by using a special basis of “spline-delta”
functions that make no distinction between bound and continuum vibrational wave
functions [295].

9.3.2 Separation of the phase matrix

With current computational methods, accurate fixed-nuclei R-matrices R FN can be
obtained that interpolate smoothly in a vibrational coordinate q and in the electronic
continuum energy . The fixed-nuclei phase matrix "FN is defined such that

tan "FN (q; ) = k(q)R FN (q; ),

where eigenphases are adjusted by multiples of π to make matrix elements contin-

uous in both  and q. For the vibronic phase matrix,
1 1
tan "vv = kv2 Rkv2 ,
1
where kv = [2(E − E v )] 2 . A precursor resonance corresponds to rapid variation of
the trace of the phase matrix. Single-pole parameters for a resonance can be deter-
mined from the energy derivative of the phase matrix [389]. Analysis of a single-
channel resonance [270] shows that an S-matrix pole at complex energy res − 12 iγ
implies that the energy derivative of the phase shift η() has a maximum value at
res . Assuming constant background phase variation, res (q) = E res (q) − V0 (q) is
defined by a local maximum of d d
T r ". The maximum eigenvalue of the matrix
d
d
" at res is 2/γ [389]. The eigenvector y defines a resonance eigenchannel. This
analysis determines the parameters in the Breit–Wigner formula for an isolated
multichannel resonance [270], consistent with a single pole of the analytically
174 9 Electron-impact rovibrational excitation of molecules

continued U -matrix as defined above. This implies an analytic formula for the
resonant phase matrix "1 , which will be considered in more detail below. Given
the phase matrix ", this construction of "1 defines a background matrix "0 by
subtraction, such that " = "0 + "1 . This procedure can be repeated for several
neighboring precursor resonances if necessary. Since " is unchanged, no informa-
tion is lost.

9.3.3 Phase-matrix formalism: EMAP

After separation into resonant and background parts, the nonresonant fixed-nuclei
phase matrix "0 is converted to a vibronic or rovibronic phase matrix by the
energy-modified adiabatic phase-matrix method (EMAP) [409]. This is simply
an adaptation of the EMA formalism to the phase matrix "0 . The implied vibronic
background phase matrix is


  
( pLv|"0 | p L v ) = χv∗ (q) pL "FN
  
0 (q; vv  ) p L χv  (q) dq,

1
where vv = [(E − E v )(E − E v )] 2 . The geometric mean is appropriate to the
dimensionless vibronic phase matrix for general short-range potentials [236], as
in the earlier EMA theory [267].
The EMAP method has been used in ab initio calculations of near-threshold rota-
tional and vibrational excitation in electron scattering by polar molecules [409, 410].
Computed differential cross sections are in quantitative agreement with available
experimental data for e−HF scattering. Because of the computational simplifica-
tions resulting from combining R-matrix theory with the adiabatic approximation,
this methodology was able to obtain results equivalent to converged close-coupling
calculations including both vibrational and rotational degrees of freedom. Specific
treatment of rotational structure is essential for such molecules because of dynam-
ical rotational screening of the long-range dipole potential. These calculations pro-
vide a detailed analysis of the striking threshold peak structures observed for such
dipolar molecules. For e−H2 vibrational excitation, in a direct comparison with
the FONDA (first-order nondegenerate adiabatic) approximation [252, 2, 254], and
with benchmark vibrational close-coupling results [411], the EMAP method was
found to be computationally efficient and reasonably accurate at energies somewhat
above threshold and away from a scattering resonance [234].
The EMAP method has been used to compute elastic scattering and symmetric-
stretch vibrational excitation cross sections for electron scattering by CO2 [235].
This is one of the first ab initio calculations of vibrational excitation for a poly-
atomic molecule. The results are in good agreement with experiment, which shows
unusually large low-energy cross sections. The theory identifies a near-threshold
 9.3 Vibronic R-matrix theory 175

singularity in the fixed-nuclei scattering matrix, changing from a virtual state to a

bound state as the vibrational coordinate varies [249, 235].

9.3.4 Nonadiabatic theory: NADP

For low collision energies, especially when electronic resonances occur, and for
processes such as dissociative attachment, adiabatic theory is not adequate for
vibrational excitation and energy transfer. The strategy of separating the fixed-
nuclei phase matrix into resonant and nonresonant parts makes it possible to apply
resonance analysis, analogous to the very successful projection-operator method
[82], to the rapidly varying part of the phase matrix. This nonadiabatic phase-matrix
(NADP) formalism is derived here. Since adiabatic theory is generally adequate
for rotational coupling and excitation, the discussion here is limited to the model
considered above, specifically for a diatomic molecule in the body-frame, with only
one internuclear coordinate q.
As described above, time-delay analysis [389] of the energy derivative of the
phase matrix " determines parametric functions that characterize the Breit–Wigner
formula for the fixed-nuclei resonant R-matrix R1FN (q; ). The resonance energy
res (q), the decay width γ (q), and the channel-projection vector y(q) define R1FN
and its associated phase matrix "FN 1 , such that tan "1 = k(q)R1 , where
FN FN

1
y(q)γ 2 (q)[res (q) − (q)]−1 γ 2 (q)y † (q).
1 1
R1FN (q; ) = (9.2)
2
Using the basic rationale of EMA theory [267], the parametric function (q) be-
comes op (q) = E−Hn when the kinetic energy of nuclear motion cannot be neg-
lected. However, the operator (res (q) − op (q)) has a well-defined c-number value
in vibrational eigenstates determined by the eigenvalue equation

(T̂ n + E res (q))χs (q) = χs (q)E s . (9.3)

Defining s (q) = E s − V0 (q), this implies

(res (q) − op (q))χs (q) = χs (q)(s (q) − (q)).

Here s (q) − (q) = E s − E, independent of q. Thus the energy denominator in

Eq. (9.2) can be replaced by E s − E in the vibrational eigenstate χs . In the NADP
method, Eq. (9.3) is solved in a basis of spline delta-functions [295], which de-
termines bound and continuum vibrational eigenfunctions to graphical accuracy
(a cubic spline fit) in the coordinate range 0 ≤ q ≤ qd . Substituting this c-number
energy denominator into Eq. (9.2), and evaluating matrix elements of the resulting
operator in the vibrational coordinate q, the NADP resonant vibronic R-matrix for
176 9 Electron-impact rovibrational excitation of molecules

vibrational excitation is
1   1   1 † 
( pLv|R1 | p  L  v  ) = v y pL γ 2 s (E s − E)−1 s γ 2 y p L  v  . (9.4)
2 s

For comparison with projection-operator theory, this corresponds to a Born–

Oppenheimer precursor resonance state

!s (x N +1 , q) = "d (q; x N +1 )χs (q),

where "d is a postulated discrete state, defined for q ≤ qd , that interacts with
the background electronic continuum. For this precursor state, the projection
integrals on the electronic wall of the vibronic hypercylinder are ( pLv|s) =
(A# p Y L χv |!s )r =a , and the corresponding projection integral on the dissociation
cap is (d|s) = ("d |!s )q=qd = χs (qd ), a normalized eigenfunction of Eq. (9.3). The
vibronic R-matrix for this precursor resonance state is
1
( pLv|Rres | p  L  v  ) = ( pLv|s)(E s − E)−1 (s| p  L  v  ), (9.5)
2 s

the R-matrix connecting wall and cap surfaces is

1
(d|Rres | pLv) = χs (qd )(E s − E)−1 (s| pLv),
2 s

and the R-matrix on the cap surface is

1
(d|Rres |d) = χs (qd )(E s − E)−1 χs∗ (qd ).
2 s

As a consistency check, if a Green function is defined by

1
G(q, q  ) = χs (q)(E s − E)−1 χs∗ (q  ),
2 s

all formulas agree with Schneider et al. [363].

Comparing Eqs. (9.5) and (9.4), they would be identical if the “magic formula”
 1 
( pLv|s) ∼= v y pL γ 2 s

were valid. The NADP formalism postulates this to be true, implying that the R-
matrix connecting wall and cap surfaces is
1  1 † 
(d|R1 | pLv) = χs (qd )(E s − E)−1 s|γ 2 y pL |v ,
2 s
 9.3 Vibronic R-matrix theory 177

and the R-matrix on the cap surface is

1
(d|R1 |d) = χs (qd )(E s − E)−1 χs∗ (qd ).
2 s
Because the postulated diabatic state "d is never computed explicitly, the NADP
formalism completely avoids the conceptual difficulties associated with this state in
the LCP and projection operator methods. Reduction of the vibrational computation
to solution of Eq. (9.3) removes the difficult issue of vibrational completeness
inherent in vibrational close-coupling theory and in the method of Schneider et al.
The NADP method was first tested in calculations of e−N2 rovibrational excita-
tion. The efficiency of this formalism was demonstrated by carrying the calculations
to effective completeness for combined rotational and vibrational close-coupling.
The multipeaked vibrational excitation structure was computed to an accuracy that
appears to agree more closely with experiment than does any previous theoretical
work [153]. Computed differential vibrational excitation cross sections are in close
agreement with experiment. More recently, the NADP method has been used in a
series of calculations of e−H2 scattering intended to calibrate the method against
earlier work that was designed to give definitive results for low-energy vibrational
excitation cross sections. The 2 u shape resonance of H− 2 , which dominates elec-
tron scattering for energies below 10 eV, has traditionally been very difficult to
characterize, because the indicated decay width becomes very large at internuclear
distances near the the ground-state equilibrium q0 . It was found in NADP calcu-
lations [236, 233] that the precursor resonance considered in this methodology is
in fact very well defined, and the resulting resonance vibrational excitation cross
sections are in close agreement with the best available prior calculations.
 IV
Field theories

This part is concerned with variational principles underlying field

theories. Chapter 10 develops the nonquantized theory of interacting
relativistic fields, emphasizing Lorentz and gauge invariant Lagrangian
formalism. The theory of a classical nonabelian gauge field is carried to the
point of proving gauge invariance and of deriving the local conservation
law for field energy and momentum densities.
 10
Relativistic Lagrangian theories

The principal references for this chapter are:

[73] Cottingham, W.N. and Greenwood, D.A. (1998). Introduction to the Standard
Model of Particle Physics (Cambridge University Press, New York).
[93] Edelen, D.G.B. (1969). Nonlocal Variations and Local Invariance of Fields
(American Elsevier, New York).
[121] Feynman, R.P. (1961). Quantum Electrodynamics (Benjamin, New York).
[217] Leader, E. and Predazzi, E. (1996). An Introduction to Gauge Theories and Modern
Particle Physics, Vols. 1 and 2 (Cambridge University Press, New York).
[336] Ramond, P. (1989). Field Theory: A Modern Primer, 2nd edition (Addison-Wesley,
Redwood City, California).
[340] Renton, P. (1990). Electroweak Interactions (Cambridge University Press,
New York).
[352] Ryder, L.H. (1985). Quantum Field Theory (Cambridge University Press,
New York).
[373] Schwinger, J. (1958). Quantum Electrodynamics (Dover, New York).
[391] Swanson, M.S. (1992). Path Integrals and Quantum Processes (Academic Press,
New York).
[421] Weinberg, S. (1995). The Quantum Theory of Fields, Vol. I (Cambridge University
Press, New York).
[422] Weinberg, S. (1996). The Quantum Theory of Fields, Vol. II (Cambridge University
Press, New York).

In quantum electrodynamics (QED), the classical electromagnetic field Aµ of

Maxwell and the electronic field ψ of Dirac are given algebraic properties (Bose–
Einstein and Fermi–Dirac quantization, respectively), and through their interaction
account for almost all physical phenomena that can be observed in ordinary human
circumstances. The relativistic theory is derived from Hamilton’s principle for an
action defined by the space-time integral of a Lorentz invariant Lagrangian density
[373]. This same action integral can be used to develop the diagrammatic pertur-
bation theory of Feynman [121]. The cited references describe the formalism and
methodology which demonstrate that QED is in remarkable agreement with all

181
182 10 Relativistic Lagrangian theories

empirical data to which it is applicable. Classical and quantized QED will be used
here to introduce the basic formalism of field theory, including the variational the-
ory of invariance properties. This theory, especially gauge invariance, is central to
recent developments of electroweak theory (EWT) and quantum chromodynamics
(QCD).
Because both the Maxwell and the Dirac fields appear to be truly elementary, not
simply approximate models of inaccessible underlying structures, QED serves as a
model for a more general field theory that might ultimately describe all phenomena.
In recent years, some of the mystery has been removed from nuclear and high-energy
physics by the discovery (or invention) of quarks and gauge boson fields, as analogs
and generalizations of electrons and the electromagnetic field, respectively. Starting
from QED as a generic gauge theory, the formalism of electroweak theory (EWT)
is developed here, following the implications of nonabelian gauge symmetries in
field theory. The same principles apply to the quantum chromodynamics (QCD) of
quarks and gluons, which will not be considered in detail.

10.1 Classical relativistic electrodynamics

The term “classical” here implies “nonquantized”, treating the Dirac spinor field ψ
as a classical field in space-time. The theory is simplified by considering a repre-
sentation in terms of truly elementary unrenormalized fields, corresponding to an
electron with no “bare” mass, and a classical Maxwell field, interacting through an
unrenormalized coupling constant. Physically observed electronic mass and charge
can be considered to be dynamical results of such a model, when extended by field
quantization and augmented by additional terms in the Lagrangian density. The bare
fermion field ψ is described by a Pauli 2-spinor, with definite left-handed helicity
[25, 336]. Elementary massless fermions with left-handed helicity have positive
energy, and those with right-handed helicity have negative energy. In the bare vac-
uum state of the quantized theory, all right-handed chiral states are fully occupied.
Holes in these states are equivalent to right-handed antifermion states of positive
energy and reversed charge. Thus the massless QED model includes electrons and
positrons from the outset. The effect of renormalization, to be considered below, is
to mix bare left-handed electrons of positive energy with bare right-handed elec-
trons of negative energy, to produce the “dressed” electron described by the Dirac
4-component spinor, with the observed mass and charge by construction. In com-
parison with EWT and QCD, this model of QED neglects only weak interactions,
since electrons do not interact with the gluon fields.
The QED Lagrangian density in mixed Gaussian units is

L = −(1/16π )F µν Fµν + i hcψ † γ 0 γ µ Dµ ψ,

 10.1 Classical relativistic electrodynamics 183

using the customary summation convention for repeated indices. Covariant 4-

vectors are defined here by

xµ = (ct, −r), ∂µ = (∂/c∂t, ∇),

Aµ = (φ, −A), jµ = (cρ, −j).

The signs of spatial components are reversed in the corresponding contravariant

4-vectors, indicated by x µ , etc. Dirac matrices are represented in a form appropriate
to a 2-component fermion theory [38], in which helicity γ 5 is diagonal for zero-mass
fermions,




µ 0 I 0 σ −I 0
γ = , ; γ =
5
. (10.1)
I 0 −σ 0 0 I

The covariant electromagnetic field is defined in terms of the electromagnetic

4-potential Aµ by the antisymmetric form

Fµν = ∂µ Aν − ∂ν Aµ .

The Dirac and Maxwell fields are coupled through the covariant derivative

Dµ = ∂µ + i(−e/hc)Aµ ,

where −e is the electronic charge.

QED theory is based on two distinct postulates. The first is the dynamical postu-
late that the integral of the Lagrangian density over a specified space-time region is
stationary with respect to variations of the independent fields Aµ and ψ, subject to
fixed boundary values. The second postulate attributes algebraic commutation or
anticommutation properties, respectively, to these two elementary fields. In the clas-
sical model considered here, the dynamical postulate is retained, but the algebraic
postulate and its implications will not be developed in detail.
The dynamical postulate implies the covariant field equations,

∂ µ Fµν = (4π/c) jν = (4π/c)(−ecψ † γ 0 γν ψ),

i hcγ µ Dµ ψ = 0. (10.2)

Expanded in a complete set of spinor functions, the fermion field is ψ(x) =


p u p (x)a p , where x = (x, t). In a particular Lorentz frame, Eq. (10.2) takes the
form

i h∂t ψ = Hψ = −i h{cγ 0 γ µ Dµ − ∂t }ψ. (10.3)

For zero rest mass, using Eqs. (10.1), the 2 × 2 operator that acts on 2-component
184 10 Relativistic Lagrangian theories

spinors of negative helicity is

H L = σ · (i hc∇ − eA) − e A0
 e 
= −cσ · p + A − eφ.
c

10.1.1 Classical dynamical mass

If there is no explicit external electromagnetic field, the covariant field equations
determine a self-interaction energy that can be interpreted as a dynamical electron
mass δm. Since this turns out to be infinite, renormalization is necessary in order to
have a viable physical theory. Field quantization is required for quantitative QED.
The classical field equation for the electromagnetic field can be solved explicitly
µν
using the Green function or Feynman propagator G P , whose Fourier transform is
−g µν /κ 2 , where κ = k p − kq is the 4-momentum transfer. The product of γ0 and
the field-dependent term in the Dirac Hamiltonian, Eq. (10.3), is
−e A/(x) = −eγ µ Aµ (x)


µν
= 4π e γµ d 4 x  G P (x − x  )
2
a †p u ∗p (x  )γ 0 γν u q (x  )aq .
p,q

In terms of the 4-momentum transfer κ this is

d 4 κ g µν  † 
−e A/(x) = −4πe2 γµ a d 4 x  e−iκ · (x −x) u ∗p (x  )γ 0 γν u q (x  )aq .
(2π )4 κ 2 p,q p

This is an exact result of the theory, and must retain its validity for quantized fields. If
there is no external field, any scalar field resulting from coupling u ∗p u q must vanish.
This is required in the vacuum state of the quantized theory, and a valid classical
model should omit such terms. Because the functions u p (x) are elementary Pauli
spinors, the nonscalar field A/(x) has triplet-odd character, describing a vector boson
in quantized theory. Then A/(x) is an odd-parity field, whose matrix elements vanish
except for transitions between negative and positive energies, because massless
fermions have definite helicity. Chirality-breaking virtual transitions are required
in order to produce a nonvanishing mass from the electromagnetic interaction.
In relativistic theory, it is desirable to retain covariant forms in the equations of
motion. Instead of Eq. (10.3), Eq. (10.2) can be rewritten as

i hcγ µ ∂µ ψ = −e A/ψ = δ m̂c2 ψ,

defining δ m̂c2 = −e A/. This is an operator or matrix in the Pauli spinor representa-
tion, since its nonvanishing matrix elements describe transitions between helicity
 10.1 Classical relativistic electrodynamics 185

states. If there were a bare mass m 0 , a term m 0 c2 ψ would be added to the right-hand
side of this equation, defining m̂ = m 0 + δ m̂. Since the bare mass m 0 is an unde-
termined parameter of the theory, the only nonarbitrary value that can be assigned
to it is zero, unless it can be attributed to a physical mechanism omitted from the
field theory. The dynamical mass δm is in principle determined by the theory, but
is found to be infinite in the absence of a renormalization procedure equivalent
to introducing an arbitrary cutoff for divergent sums or integrals over relativistic
momentum transfer [121].

10.1.2 Classical renormalization and the Dirac equation

The mass-density ψ † (x, t)γ 0 m̂ψ(x, t) defines an invariant mass integral for the
Dirac field,


M= ψ † (x, t)γ 0 m̂ψ(x, t) d 3 x. (10.4)

In quantized theory, this is an operator in the fermion field algebra. Assuming

m 0 = 0, the mean value 0|M|0 vanishes in the reference vacuum state be-
†
cause all momenta and currents cancel out. In a single-electron state |a = aa |0,
 3 † 0 (more precisely, self-mass) is defined by δmc = a|Mc |a =
2 2
a self-energy
a| d xψ γ (−e A/)ψ|a. Only helicity-breaking virtual transitions can contribute
to this electromagnetic self-mass.
Because the Dirac field equation is inherently nondiagonal in the chiral repre-
sentation, any eigenstate that diagonalizes the operator m̂ must be represented by a
4-component Dirac spinor, constructed from two Pauli spinors of opposite helicity.
Perturbation theory sums or integrals over momentum-transfer, due to matrix ele-
ments of m̂, are infinite without some cutoff. Following the logic of quantized QED
[121], it is assumed that a covariant cutoff can be defined such that the eigenvalues
of this matrix are mc2 , −mc2 for dressed physical electrons of positive and negative
energy, respectively. The corresponding classical renormalization replaces an inde-
terminate mass by its physical value, transforming the spinor representation space
to be consistent with this value. A cutoff parameter is to be determined such that
the integral defined by Eq. (10.4) reproduces the physical dressed mass m when
evaluated using a Dirac spinor field defined for this mass. In the presence of an
external Maxwell field, the field equation with diagonalized self-mass is just the
Dirac equation for a single dressed electron.
The field equation for a free dressed electron is ( p/ − m̂)ψ = 0, in units such
that h = c = 1. The space-time Green function G m (x1 , x0 ) for this field equation
186 10 Relativistic Lagrangian theories

is defined as the solution of

( p/ − m 0 − δ m̂)G m (x1 , x0 ) = iδ (4) (x1 , x0 )

that vanishes when t1 < t0 . If G 0 (x1 , x0 ) is defined in the same way, without δ m̂,
this implies a Lippmann–Schwinger equation


G m (x1 , x0 ) = G 0 (x1 , x0 ) − i d 4 yG 0 (x1 , y)δ m̂(y)G m (y, x0 ).

In momentum space this takes the form

i i i
= (−iδ m̂) , (10.5)
p/ − m̂ p/ − m 0 p/ − m̂
which produces an infinite sum of Feynman diagrams [120] when expanded by
iteration. Individual terms in this expansion are modified by field quantization.

10.2 Symmetry and Noether’s theorem

In the formalism of either classical or quantum field theory, symmetries of the
Lagrangian density L or of the action integral S = L d 4 x play a vital role in
establishing global conservation laws. In classical field theory, Noether’s theorem
[304, 93, 336] shows that each symmetry of the action integral implies a con-
servation law. In quantum field theory, such symmetries result in Ward identities
[417, 395] for time-ordered products and expectation values. In this context, a
symmetry of the formalism is any transformation of the fields and space-time co-
ordinates that leaves the action integral unchanged.
Noether’s theorem will be proved here for a classical relativistic theory defined
by a generic field φ, which may have spinor or tensor indices. The Lagrangian
density L(φ, ∂µ φ) is assumed to be Lorentz invariant and to depend only on scalar
forms defined by spinor or tensor fields. It is assumed that coordinate displacements
are described by Jacobi’s theorem δ(d 4 x) = d 4 x ∂µ δx µ . The most general variation
of the action integral, evaluated over a closed space-time region ), is


δS = d 4 x(L∂µ δx µ + δL).
)

Any variation of φ takes the form δφ = δ0 φ + ∂µ φδx µ , where δ0 omits coordinate

variations. The full variation δL is
∂L ∂L
δL = (∂µ L)δx µ + δ0 φ + ∂µ δ0 φ,
∂φ ∂(∂µ φ)
using δ0 ∂µ φ = ∂µ δ0 φ. As in integration by parts, a generalized gradient term can
 10.2 Symmetry and Noether’s theorem 187

be extracted such that

   
µ ∂L ∂L ∂L
δL = (∂µ L)δx + ∂µ δ0 φ + − ∂µ δ0 φ.
∂(∂µ φ) ∂φ ∂(∂µ φ)
The last term here vanishes when φ satisfies the field equations. Using L∂µ δx µ +
(∂µ L)δx µ = ∂µ (Lδx µ ), the total variation of the action about a field solution takes
the form


δS = − d 4 x ∂µ J µ ,
)

where J µ is the infinitesimal 4-current density

∂L
−Lδx µ − δ0 φ.
∂(∂µ φ)
Because the assumed hypervolume can be reduced to an infinitesimal, stationary or
invariant action implies the local form of Noether’s theorem, ∂µ J µ = 0, an equation
of continuity in space-time for the generalized current density determined by the
field φ.
If the variations of both coordinates and field are determined by parameters
denoted by α q , then
δx µ q δφ
δx µ = δα , δφ = q δα q .
δα q δα
Functional derivatives are indicated here because α q may be a parametric field, as
µ
for example a Lagrange multiplier field. Variations of α q determine J µ = Jq dα q ,
where
  ν
µ δx µ ∂L δ0 φ ∂L µ δx ∂L δφ
Jq = −L q − = ∂ν φ − gν L − .
δα ∂(∂µ φ) δα q ∂(∂µ φ) δα q ∂(∂µ φ) δα q
Green’s theorem for the action integral in a hypercylinder enclosed by space-like
end surfaces implies that the difference between integrals of J 0 over the forward and
backward end 3-surfaces is compensated by the flux J integrated over the cylinder
walls. End surfaces can be defined by a progressive time parameter τ such that
τ1 ≤ τ ≤ τ2 . If φ and its normal gradient vanish on the wall surface for sufficiently
large effective radius R, integrals of J 0 over any such nested τ -surface must be
equal. This is the integral Noether theorem. If α q is arbitrary, the integral of Jq0 over
any such 3-surface defines a constant of motion for each independent index q.

10.2.1 Examples of conservation laws

Consider an infinitesimal displacement δa µ in space-time. The action is invariant
if for x  = x + δa, φ  (x  ) = φ(x) for a field φ, where the displaced field function
188 10 Relativistic Lagrangian theories

is defined by φ  (x) = φ(x) + δ0 φ. These definitions are equivalent to

δφ(x) = 0; δx µ = δa µ .

The conserved current density determined by Noether’s theorem is

 
µ ∂L
J = ∂ν φ − gν L δa ν = Tνµ δa ν ,
µ
∂(∂µ φ)
defining the energy–momentum tensor Tνµ . Noether’s theorem determines an
energy–momentum 4-vector for the field φ, conserved if the action integral is 4-
translation invariant,


Pν = d 3 x Tν0 = (E, −cP),

where



∂L
E= d x 3
φ̇ − L
∂ φ̇


∂L
P = − d 3 x ∇φ.
∂ φ̇
These formulas are valid for the nonrelativistic one-electron Schrödinger equa-
tion. The Lagrangian density is

L = ψ ∗ (i hψ̇ − Hψ).

Then ∂L/∂ ψ̇ = i hψ ∗ implies the standard formulas

E = d x (i hψ ψ̇ − L) = d 3 x ψ ∗ Hψ
3 ∗

P = d 3 x ψ ∗ (−i h∇)ψ.

Under a Lorentz transformation , a covariant field φa (x) transforms according

to

φa (x  ) = Dab ()φb (−1 x  ).

For an infinitesimal transformation, the irreducible representation matrix is Dab =

δab + δ Dab , and the coordinate transformation returning to the original coordinate
value is (−1 )µν x ν = x µ − λµν x ν = x µ . Defining the transformation with this re-
verse step makes δx µ = 0, while the functional form of the field changes according
to

δ0 φa (x) = δ Dab φb (x) − λµν x ν ∂µ φa (x).

 10.3 Gauge invariance 189

The conserved current defined by Noether’s theorem is

∂L  
Jµ = − δ Dab φb (x) − λµν x ν ∂µ φa (x) .
∂(∂µ φa )
In the case of a scalar field, the irreducible matrix D is a unit matrix, and drops out
of J µ . For rotation through an angle δθk about the Cartesian axis êk , the rotational
submatrix of the Lorentz matrix is given by λij x j = δθk  ki j x j , where  i jk is the
totally antisymmetric Levi–Civita tensor. For the one-electron Schrödinger field ψ,
Noether’s theorem defines three conserved components of a spatial axial vector,


∂L j
L = − d 3 x  i jk
i
x ∇k ψ,
∂ ψ̇
which is just the orbital angular momentum vector


L = d 3 x ψ ∗ r × (−i h∇)ψ.

For the Dirac bispinor, the irreducible representation matrix Dab for each helicity
component is a Pauli spin matrix σ multiplied by h/2. Then
1
δ Dab ψb = hδθk σab
k
ψb .
2
Given
∂L/∂(∂µ ψ) = i hcψ † γ 0 γ µ ,

and γ 0 γ 0 = I , the conserved angular momentum is

 
† 1
J = d xψ3
hσ + r ×(−i h∇) ψ = S + L.
2
S is the elementary spin of the Dirac electron.

10.3 Gauge invariance

In classical electrodynamics, the field equations for the Maxwell field Aµ depend
only on the antisymmetric tensor Fµν , which is invariant under a gauge transfor-
mation Aµ → Aµ + hc∂µ χ (x), where χ is an arbitrary scalar field in space-time.
Thus the vector field Aµ is not completely determined by the theory. It is customary
to impose an auxiliary gauge condition, such as ∂µ Aµ = 0, in order to simplify the
field equations. In the presence of an externally determined electric current density
4-vector j µ , the Maxwell Lagrangian density is
1 µν 1
LA = − F Fµν − j µ Aµ .
16π c
190 10 Relativistic Lagrangian theories

The Euler–Lagrange equations are

∂L A ∂L A
∂µ − = 0,
∂(∂µ Aν ) ∂ Aν
which implies the covariant inhomogeneous Maxwell equations

∂µ F µν = (4π/c) j ν .

Because F µν is antisymmetric, the symmetrical derivative ∂µ ∂ν F µν must vanish.

This requires j ν to satisfy the equation of continuity, ∂ν j ν = 0, which implies
charge conservation in an enclosed volume if net current flow vanishes across its
spatial boundary.
Gauge covariance of the classical theory is due to the invariance of the field
tensor Fµν under the local gauge transformation

Aµ (x) → Aµ (x) + hc∂µ χ (x),

on the assumption that the current density j µ is not affected by this transformation.
However, the Lagrangian density L A is not invariant, since the coupling term is
modified such that

L A → L A − h j µ ∂µ χ (x).

Thus the equations of motion are gauge invariant, but the action integral is not.

10.3.1 Classical electrodynamics as a gauge theory

Analysis of the “classical” Dirac theory shows a similar inconsistency under local
phase transformations, such that ψ(x) → eieχ (x) ψ(x), corresponding to the local
infinitesimal transformation, for χ → 0,

δψ(x) = ieχ (x)ψ(x).

For the Dirac field in an externally determined Maxwell field, the Lagrangian density
including a renormalized mass term is

L D = i hcψ † γ 0 {γ µ Dµ + im}ψ.

The field interaction is expressed in terms of the covariant derivative

Dµ = ∂µ + i(−e/hc)Aµ ,

such that the Euler–Lagrange equation for the Dirac field is

i hcγ µ Dµ ψ = mc2 ψ.
 10.3 Gauge invariance 191

This is gauge-covariant only if the gauge transformation of Aµ and the phase

transformation of ψ are combined, such that

Aµ → Aµ + hc∂µ χ (x); ψ → eieχ(x) ψ.

Because
 
−ieχ (x) ie
e Dµ ψ → ∂µ + ie(∂µ χ ) − (Aµ + hc∂µ χ ) ψ = Dµ ψ,
hc

the Euler–Lagrange equation for ψ is covariant under this generalized local

gauge transformation. The electrodynamic field action, expressed in terms of the
gauge invariant tensor Fµν and the covariant derivative Dµ ψ, is gauge invariant.
This introduces a scalar gauge function χ (x) that is inherently unobservable. Gauge
invariance completely determines the interaction between the Dirac and Maxwell
fields.

10.3.2 Noether’s theorem for gauge symmetry

For a local infinitesimal gauge transformation about a solution of the field equations,
 

∂L ∂L hc
δL = ∂µ δ Aν + δψ = ∂µ − F µν ∂ν χ + h j µ χ = 0,
∂(∂µ Aν ) ∂(∂µ ψ) 4π

defining the Dirac electric current density

j µ (x) = −ecψ † (x)γ 0 γ µ ψ(x).

Because antisymmetry of F µν implies that F µν ∂µ ∂ν χ ≡ 0,


µ c νµ µ
(∂µ j )χ (x) = ∂ν F − j ∂µ χ(x) = 0.
4π

Thus, since χ(x) is arbitrary, Noether’s theorem implies ∂µ j µ = 0, as required

for consistency of the field equations. While j µ is invariant under a local phase
transformation, local gauge invariance requires an interaction described by the
covariant derivative, implying a gauge-covariant Maxwell field. This has profound
consequences in the generalization to electroweak theory [340, 217], since renor-
malizability of the quantum field theory follows from generalized gauge invariance
[404, 405]. Noether’s theorem implies that gauge fields and the corresponding
conserved current densities must exist in any renormalizable quantum theory that
can describe the weak interactions responsible for beta-decay processes.
192 10 Relativistic Lagrangian theories

10.3.3 Nonabelian gauge symmetries

In electrodynamics, the gauge function χ (x) is a real scalar function in space-
time, and the phase transformation of the fermion field takes the form ψ(x) →
eieχ (x) ψ(x). This is a multiplicative unitary transformation, characterized by U (1)
in group theory [217]. Because the representation matrix is simply the phase factor
eieχ (x) , any two such factors commute, defining an abelian group. As originally
shown by Yang and Mills [435], it is possible to construct field theories with gauge
symmetries in which U (1) phase factors are replaced by noncommuting unitary
representation matrices of a nonabelian continuous group. For example, the group
SU(2) is represented by 2 × 2 unitary matrices exp(ieχ(x) · τ ), where the matrices
τ are identical to the Pauli matrices σ.
Electroweak theory (EWT) [73, 340, 217] makes use of the gauge group U (1) ×
SU (2). The representation matrices are of the form

exp(ieχ (x)τ0 + ie χ(x) · τ ),

where τ0 is the unit matrix. This is a product of U (1) factors and 2 × 2 SU(2)
matrices. The fermion field ψ becomes an array consisting of two spinors, each
carrying an additional “weak isospin” index I3 . In EWT this index distinguishes
members of isospin doublets: electrons and neutrinos or down quarks and up quarks.
A local unitary transformation of a fermion field ψ is defined by

ψ → U ψ; ψ † → ψ †U −1 ,

where U commutes with the Dirac matrices γ µ . Interaction with a gauge field E µ
is described by a covariant derivative Dµ in the fermion Lagrangian density

L = i hcψ † γ 0 γ µ Dµ ψ = i hcψ † γ 0 γ µ (∂µ + E µ )ψ.

If the gauge field transforms according to

E µ → U E µU −1 − (∂µU )U −1 , (10.6)

this ensures gauge invariance of L,

L → i hcψ † γ 0 γ µU −1 [U ∂µ + (∂µU ) + U E µ − (∂µU )]ψ = L.

For the abelian phase factor of the U (1) group, U = exp{ieχ(x)} and E µ =
−(ie/hc)Aµ , which produces the usual gauge transformation Aµ → Aµ +
hc∂µ χ (x). In electroweak theory, a U (1) gauge field Bµ is defined such that
U = exp{− 12 ig  yχ(x)} and E µ = (ig  y/2hc)Bµ . The gauge transformation is
Bµ → Bµ + hc∂µ χ (x).
 10.3 Gauge invariance 193

The transformation matrix for the nonabelian group SU(2) is

 g 
U = exp −i χ(x) · τ .
2
The SU(2) gauge field Wµ has three components, corresponding to the isospin vec-
tor of matrices τ , with no relationship to the coordinate space ct, x. By implication,
the fermion field ψ is a set of spinors, one for each value of the isospin index. The
covariant derivative
g
Dµ = ∂µ + i τ · Wµ
2hc
ensures that the fermion Lagrangian density L = i hcψ † γ 0 γ µ Dµ ψ is invariant if the
gauge field transforms according to Eq. (10.6). This Lagrangian density describes
physical processes associated with beta decay, transitions that convert a neutrino
into a lepton or convert a down quark into an up quark (or a neutron into a proton),
coupling such transitions in second order.
Considering only the SU(2) field interaction, the Euler–Lagrange equation for
the fermion field is
1
i hcγ µ ∂µ ψ = gγ µ τ · Wµ ψ.
2
If there is no independent external gauge field, this defines an incremental fermion
mass operator δ m̂ W c2 = 12 gγ µ τ · Wµ such that
i hcγ µ ∂µ ψ = δ m̂ W c2 ψ.
As in the case of the electromagnetic self-mass, the implied dynamical mass incre-
ment is infinite unless perturbation-theory sums are truncated by a renormalization
cutoff procedure. In analogy to electrodynamics, each fermion field acquires an
incremental dynamical mass through interaction with the gauge field. This implies
in electroweak theory that neutrinos must acquire such a dynamical mass from
their interaction with the SU(2) gauge field. For a renormalized Dirac fermion in
an externally determined SU(2) gauge field, the Lagrangian density is
1 µ
L D = i hcψ † γ 0 {γ µ ∂µ + im}ψ − jW · Wµ ,
c
where
µ 1
jW = gcψ † γ 0 γ µ τ ψ.
2
This defines the fermion contribution to an isovector gauge current density. Al-
though the Euler–Lagrange equation is gauge covariant by construction, this
fermion gauge current is not invariant, because the matrix τ does not commute
with the SU(2) unitary transformation matrices. It will be shown below that the
194 10 Relativistic Lagrangian theories
µ
conserved Noether current density augments jW by a component determined by the
gauge field. The nonabelian gauge field carries a nonvanishing isospin, which is
implicit in the isovector notation used here for Wµ . It interacts with itself through
the Noether current density.
Transformation properties of Wµ can be derived by considering an infinitesimal
SU(2) gauge transformation, U = I − 12 igχ(x) · τ , for χ(x) → 0. The correspond-
ing infinitesimal transformation of the fermion field is
1
ψ → ψ − igχ(x) · τ ψ.
2
Because the matrix χ(x) · τ can convert a neutrino into an electron at any point in
space-time, a physically correct theory must couple such a transition with a field that
carries a compensating electric charge. Thus local SU(2) gauge invariance implies
existence of the gauge field Wµ with components that carry electric charge. The field
equations for Wµ contain a source term with off-diagonal elements that represent
charge transitions. Retaining only first-order terms in χ(x), Eq. (10.6) takes the
form [435, 340]
1
τ · Wµ → τ · Wµ + ig(τ · Wµ χ · τ − χ · τ τ · Wµ ) + hcτ · ∂µ χ
2
= τ · (Wµ − gWµ × χ + hc∂µ χ).
This derivation uses an algebraic identity valid for Pauli matrices and any two
vectors A, B,
(τ · A)(τ · B) = A · B + iτ · A × B.
Since the components of τ are linearly independent, the implied infinitesimal gauge
transformation is
Wµ → Wµ − gWµ × χ + hc∂µ χ.

The term proportional to the coupling constant g is typical of a nonabelian gauge

group [422].
In the presence of an externally determined fermion gauge current, the
Lagrangian density for the SU(2) gauge field is [435]
1 1 µ
LW = − Wµν · Wµν − jW · Wµ ,
16π c
expressed in terms of an isovector field defined by
g
Wµν = ∂µ Wν − ∂ν Wµ − Wµ × Wν .
hc
The first term in LW is gauge invariant by construction [340, 435]. The second
 10.3 Gauge invariance 195

term is not, because of the gauge modification of Wµ . The implied Euler–Lagrange

equations for the gauge field [435] are
g 4π ν 4π ν
∂µ Wµν = Wµ × Wµν + jW = J ,
hc c c W
defining the total isovector current density JνW . It will be shown here that this is the
conserved Noether current density for the SU(2) theory. In contrast with electro-
dynamics, a nonabelian gauge field interacts with itself through its contribution to
this conserved current density.

10.3.4 Gauge invariance of the SU(2) field theory

The Lagrangian density for a massless fermion field interacting with the SU(2)
gauge field is
1
L=− Wµν · Wµν + i hcψ † γ 0 γ µ Dµ ψ,
16π
expressed in terms of the covariant derivative
g
Dµ = ∂µ + i τ · Wµ .
2hc
Some of the algebraic complexity of the proof of gauge invariance and of derivation
of the conserved Noether gauge current for this Lagrangian density can be simplified
by considering general variations of the fermion and gauge fields about a solution
of the field equations. The logic of this derivation is unchanged if a fermion mass
is included. Stationary action implies that



1
d 4 x δL = d 4 x − Wµν · δWµν + i hc(δψ † )γ 0 γ µ Dµ ψ
8π

† 0 µ
+ i hcψ γ γ δ(Dµ ψ) = 0.

The fermion field equation i hγ µ Dµ ψ = 0 is implied by independent variation of

ψ † . The Euler–Lagrange equations for the gauge field follow from
∂L ∂Wτ σ ∂L ∂Wτ σ ∂L ∂(Dµ ψ)
∂µ · = · + .
∂Wτ σ ∂(∂µ Wν ) ∂Wτ σ ∂Wν ∂(Dµ ψ) ∂Wν
The individual terms here are
∂L ∂Wτ σ 1
· = − Wµν ,
∂Wτ σ ∂(∂µ Wν ) 4π
∂L ∂Wτ σ g
· =− Wµ × Wµν ,
∂Wτ σ ∂Wν 4π hc
∂L ∂(Dµ ψ) 1
= − gψ † γ 0 γ ν τ ψ,
∂(Dµ ψ) ∂Wν 2
196 10 Relativistic Lagrangian theories

implying the gauge field equations


µν 4π g µν 1 † 0 ν 4π ν
∂µ W = Wµ × W + gcψ γ γ τ ψ = J .
c 4π h 2 c W
Because Wµν is an antisymmetric Lorentz tensor, the total isovector current density
must satisfy ∂ν JνW = 0.
Field variations driven by the infinitesimal local gauge transformation
1
δψ = − igτ · χ(x)ψ,
2
δWµ = −gWµ × χ(x) + hc∂µ χ(x),

where χ(x) → 0, determine the conserved Noether current density implied by

gauge invariance. Dµ and the gauge field transformation are defined so that the
fermion part of L is gauge invariant. To evaluate the variation of the gauge field
part, direct substitution of δWµ into the definition of Wµν gives
g
δWµν = ∂µ δWν − ∂ν δWµ − (Wµ × δWν − Wν × δWµ )

hc 
g
= −g ∂µ Wν − ∂ν Wµ − Wµ × Wν × χ
hc
= −gWµν × χ,

using the vector identity a × (b × c) − b × (a × c) = (a × b) × c. Gauge invari-

ance follows from Wµν · Wµν × χ = Wµν × Wµν · χ ≡ 0.
For fixed coordinates, invariance of the action integral for gauge variations about
SU(2) field solutions implies
 
∂L ∂L
δL = ∂µ · δWν + δψ
∂(∂µ Wν ) ∂(∂µ ψ)
 
1 µν 1 † 0 µ
= ∂µ W · (gWν × χ − hc∂ν χ) + hgcψ γ γ τ ψ · χ
4π 2


µ hc µν
= ∂µ hJW · χ − W · ∂ν χ = 0.
4π
Because Wµν · ∂µ ∂ν χ vanishes identically, due to the antisymmetry of Wµν , this
implies Noether’s theorem in the form


µ c νµ µ
(∂µ JW ) · χ(x) = ∂ν W − JW · ∂µ χ(x) = 0.
4π
µ
Because χ(x) is arbitrary, the total isovector current satisfies ∂µ JW = 0 at all points
in space-time so long as the field equations are satisfied.
 10.4 Energy and momentum of the coupled fields 197

10.4 Energy and momentum of the coupled fields

10.4.1 Energy and momentum in classical electrodynamics
The energy–momentum tensor derived from Noether’s theorem for electrodynamics
is
∂L ∂L
Tνµ = ∂ν Aλ + ∂ν ψ − gνµ L.
∂(∂µ Aλ ) ∂(∂µ ψ)
The Lorentz momentum 4-vector is


Pν = d 3 x Tν0 = (E, −cP),

in any inertial frame, where Tν0 = Tν0 (ψ) + Tν0 (A) such that

Tν0 (ψ) = i hcψ † (∂ν ψ − δ0ν γ 0 γ λ Dλ ψ),

1 0λ 1 σλ
Tν0 (A) = − F ∂ν Aλ + δ0ν F Fσ λ .
4π 16π
Energy–momentum conservation is expressed by ∂ ν Tν0 = 0 for a closed system.
If Tνµ were a symmetric tensor (when converted to T µν ), this would be assured
because ∂µ Tνµ = 0 by construction. Since the gauge field part of the tensor deduced
from Noether’s theorem is not symmetric, this requires special consideration, as
discussed below. A symmetric energy–momentum tensor is required for any event-
ual unification of quantum field theory and general relativity [422]. The fermion
field energy and momentum are



†
E(ψ) = d x ψ Hψ,
3
P(ψ) = d 3 x ψ † (−i h∇)ψ,

where, for a massless electron of negative helicity (positive energy),

 e 
0 λ
H = −i h(cγ γ Dλ − ∂t ) = −cσ · p + A − eφ.
c
For the Maxwell field, the energy–momentum tensor Tνµ (A) derived from
Noether’s theorem is unsymmetric, and not gauge invariant, in contrast to the sym-
metric stress tensor derived directly from Maxwell’s equations [318]. Consider the
symmetric tensor # = T + T , where
1 µλ
Tνµ = F ∂λ A ν
4π
is defined such that
1 1 µλ 1
∂µ Tνµ = (∂µ F µλ )∂λ Aν + F ∂µ ∂λ Aν = j λ ∂λ Aν ,
4π 4π c
198 10 Relativistic Lagrangian theories

using the antisymmetry of F µλ and the field equations. For a noninteracting field the
residual term vanishes and the resulting symmetrical energy–momentum tensor is

1 µλ 1 σλ
#µν (A) = − F Fνλ + gνµ F Fσ λ .
4π 16π
The electric and magnetic field vectors in vacuo are, respectively,

1 ∂A
E=− − ∇φ, B = ∇ × A.
c ∂t
Using italic indices for 3-vectors, with Aµ = (φ, −A), ∂µ = (∂/c∂t, ∇),

Ei = ∂0 Ai − ∂i A0 = F0i ,
Bi = −i jk (∂ j Ak − ∂k A j ) = −i jk F jk .

In terms of these vector fields, the Maxwell field Lagrangian density is (1/8π )(E 2 −
B2 ), and the field energy and momentum are



E 2 + B2 E ×B
E(A) = 3
d x , P(A) = d3x .
8π 4π c

For interacting fields, the Maxwell field energy is not separately conserved.
A gauge-covariant derivation follows from the inhomogeneous field equations
(Maxwell equations in vacuo),
∂
∇×E + B = 0, ∇ · E = 4πρ,
c∂t
∂ 4π
∇×B − E= j, ∇ · B = 0.
c∂t c
The local conservation law follows immediately [318] from
1 ∂ 4π
B · (∇ × E) − E · (∇ × B) = ∇ · (E × B) = − (E 2 + B2 ) − E · j.
2 c∂t c
Expressed in terms of the field energy and momentum densities defined above, this
is
∂ E 2 + B2 E ×B
+ c2 ∇ · = −E · j.
∂t 8π 4π c
When integrated over a finite volume, the divergence term becomes a surface
integral of the Poynting vector [318]. Expressed in covariant notation, ∂ ν #0ν =
− 1c F 0ν jν , showing the explicit modification of energy–momentum conservation
due to the final dissipative term here.
 10.4 Energy and momentum of the coupled fields 199

As a model for classical gauge fields, the energy–momentum conservation law

can be derived directly in covariant notation. The 4-divergence
1 ν 0λ 1 1 0 0j
− ∂ (F Fνλ ) = − ∂ i (F 0 j Fi j ) − ∂ (F F0 j )
4π 4π 4π
can also be written as
1 ν 0λ 1 0λ ν 1 0 νλ 1
− (∂ F )Fνλ − F (∂ Fνλ ) = − ∂ (F Fνλ ) − F 0λ jλ ,
4π 4π 16π c
using the field equations and the identity
1
(∂ ν F 0λ )Fνλ ≡ ∂ 0 (F νλ Fνλ ).
4
This establishes, as a consequence of the field equations,
1 0 0i 1 i 0j 1
− ∂ (F F0i − F jk F jk | j<k ) − ∂ (F Fi j ) = − F 0i ji ,
8π 4π c
verifying the conservation law.
Since local energy and momentum density are well-defined for the classical
Maxwell field, respectively

E 2 + B2 E ×B
(x) = , π (x) = ,
8π 4π c
the relativistic mass density of the field is defined by

µ(x)2 c4 = (x)2 − π(x)2 c2 ,

without invoking a specific requirement for a symmetric stress tensor. For a plane-
wave radiation field in vacuo,
(E 2 − B2 )2 + 4(E · B)2
µ2 c 4 = = 0,
64π 2
because the field vectors are orthogonal and of equal magnitude (in mixed Gaussian
units).

10.4.2 Energy and momentum in SU(2) gauge theory

The Lagrangian density for a massless fermion field interacting with the SU(2)
gauge field defines the Noether energy–momentum tensor
∂L ∂L
Tνµ = · ∂ν Wλ + ∂ν ψ − gνµ L.
∂(∂µ Wλ ) ∂(∂µ ψ)
200 10 Relativistic Lagrangian theories

The fermion field energy and momentum are

E(ψ) = d 3 x ψ † HW ψ, P(ψ) = d 3 x ψ † (−i h∇)ψ.

Using italic indices for spatial 3-vectors, but retaining vector notation for the abstract
3-vector index of the matrices τ and the gauge field W, the fermion “Hamiltonian”
operator is


0 λ g 1
HW = −i h(cγ γ Dλ − ∂t ) = −cσi pi − τ · Wi + gτ · W0 .
2c 2
The SU(2) energy–momentum tensor can be symmetrized and made gauge invariant
by adding an incremental tensor


µ 1 µλ g
Tν = W · ∂λ Wν + Wν × Wλ .
4π hc
Because ∂µ Tνµ does not reduce to terms that vanish even for a noninteracting
field, this construction must be verified. The energy and 3-momentum of the gauge
field derived from the resulting symmetric energy–momentum tensor #µν are


1
E(W ) = − d 3 x(W0i · W0i − W jk · W jk | j<k ),
8π


1
Pi (W ) = d 3 x W0 j · Wi j .
4π c
The local conservation law for the interacting gauge field can be derived from
the covariant field equations, as was done above for the Maxwell field. Using the
SU(2) field equations and expanding (∂ ν W0λ ) · Wνλ as
1 0 νλ g 0 ν
∂ (W · Wνλ ) + [∂ (W × Wλ ) − 2∂ ν (W0 × Wλ )] · Wνλ ,
4 2hc
there are two distinct expansions of the 4-divergence
1 ν 0λ 1 1 0 0j
− ∂ (W · Wνλ ) = − ∂ i (W0 j · Wi j ) − ∂ (W · W0 j )
4π 4π 4π
1 0 νλ 1
=− ∂ (W · Wνλ ) − W0λ · JλW
16π c
g
− [∂ 0 (Wν × Wλ ) − 2∂ ν (W0 × Wλ )] · Wνλ .
8π hc
The final term here can be expressed in the form


1 0λ g ν 1  
W · W × Wνλ +  = W0λ · JλW − jλW + .
c 4π h c
 10.5 The Standard Model 201

The effect of this term is to remove the self-interaction current density JλW − jλW
from the dissipative term in the local conservation law. The residual invariant 
can be shown to vanish identically. Using the identity
1 g
(∂ ν Wλ ) × Wνλ = (∂ ν Wλ − ∂ λ Wν ) × Wνλ ≡ (Wν × Wλ ) × Wνλ ,
2 2hc
the residual invariant is
g
= [2∂ ν (W0 × Wλ ) − ∂ 0 (Wν × Wλ ) − 2W0λ × Wν ] · Wνλ
8π hc
g2
= W0 · [2Wλ × (Wν × Wνλ ) + (Wν × Wλ ) × Wνλ ]
8π h2 c2
g2
= W0 · [Wν (Wλ · Wνλ ) + Wλ (Wν · Wνλ )] ≡ 0.
8π h2 c2
The implied local conservation law for the interacting SU(2) field is
1
∂ ν #0ν = − W0λ · jλW .
c
Further analysis is required to conclude that the local mass density must necessarily
vanish.

10.5 The Standard Model

Because field quantization falls outside the scope of the present text, the discussion
here has been limited to properties of classical fields that follow from Lorentz
and general nonabelian gauge invariance of the Lagrangian densities. Treating the
interacting fermion field as a classical field allows derivation of symmetry properties
and of conservation laws, but is necessarily restricted to a theory of an isolated single
particle. When this is extended by field quantization, so that the field amplitude ψ
becomes a sum of fermion annihilation operators, the theory becomes applicable to
the real world of many fermions and of physical antiparticles, while many qualitative
implications of classical gauge field theory remain valid.
Quantized gauge fields play a central role in the widely accepted Standard Model
of elementary particles and their interactions. Cottingham and Greenwood [73]
provide an “elementary” introduction to the phenomological basis of this model
and to the relevant theory, together with references to the technical literature. The
postulate of generalized gauge invariance, found to imply renormalizability of the
quantized theory, has powerful implications that are evident even for the classi-
cal fields. Fermion interactions have a strictly prescribed form, which determines
the existence of gauge fields with specified internal symmetry, and determines
202 10 Relativistic Lagrangian theories

corresponding conserved current densities. The Standard Model considers three

such gauge fields, with fermion gauge transformations given by symmetry groups
U(1) (electromagnetic field), SU(2) (weak isospin), and SU (3) (strong color). There
are three generations of fermion field. The generations (electron, muon, tauon)
appear to differ only by the magnitude of the renormalized mass. The U(1) field
interacts with all fermions of nonvanishing electric charge; the SU(2) field with all
fermions (since isospin is quantized by half-integers for fermions); and the SU(3)
field (color gluons) interacts only with quark fermions, which carry a color charge.
The combination SU(2) ⊗ (U(1) ⊕ SU(3)) requires eight elementary fermions in
each generation.
The underlying symmetries of the fermion and gauge fields are broken dynami-
cally in renormalized quantum field theory. In particular, helicity is broken by the
Maxwell field, so that only neutrinos retain well-defined chirality. This implies a
canonical transformation of the fermion vacuum state into a quasiparticle repre-
sentation appropriate to the observed particles. As pointed out above, the Dirac
equation is characteristic of such a representation, mixing bare massless fermions
with different helicities. The differing physical masses of members of an isospin
doublet, for example electron and neutrino, imply broken gauge symmetry. The
SU(2) gauge field Lagrangian of renormalized electroweak theory is decomposed
into separate interaction terms for left- and right-handed components of the fermion
field, since a massless neutrino has only left-handed helicity.

10.5.1 Electroweak theory (EWT)

The minimal internally consistent theory of elementary particles is a gauge theory
of the elementary fermions of the first observed generation [340]: electron (e),
electron-neutrino (νe ), down quark (d), in three colors, and up quark (u), also
in three colors. Strong interactions do not have to be considered if stable nuclei
are treated as a passive background. If transitions between lepton generations are
neglected, the theory can be applied independently to the lepton and quark doublets
of each generation, interacting with the electromagnetic and weak gauge fields. A
universal EWT field e0Uµ = 12 (g  y Bµ + gτ · Wµ ) is coupled to fermion current
densities, characterized by weak hypercharge y and weak isospin 12 τ , through the
covariant derivative

Dµ = ∂µ + i(e0 /hc)Uµ .

Here e0 is the magnitude of the electronic charge. The equation of motion for the
massless fermion field is

i hcγ µ Dµ ψ = 0.
 10.5 The Standard Model 203

Classical gauge fields, determined by their equations of motion, are expressed

as Green function propagators acting on source current densities. The renormal-
ized theory postulates a spontaneously broken symmetry as part of a dynamical
mechanism that accounts for the large observed mass of the SU(2) gauge field
quanta. This mechanism implies a canonical transformation of the renormalized
vacuum state that mixes the bare neutral fields Wµ3 and Bµ . Defining a Weinberg
angle θW such that g  cos θW = g sin θW = e0 , the correct fermion electric charge
q0 = 12 (y + τ3 )e0 follows from an orthogonal transformation of the neutral fields
Aµ = Bµ cos θW + Wµ3 sin θW ,
Z µ0 = −Bµ sin θW + Wµ3 cos θW .

Defining Wµ± = √1 (W 1
2 µ ± Wµ2 ) and τ± = 12 (τ1 ± iτ2 ),
1 1
Uµ = (y + τ3 )Aµ + (−y tan θW + τ3 cot θW )Z µ0
2 2
1
+√ (τ+ Wµ− + τ− Wµ+ ).
2 sin θW
The parameter y is −1 for leptons and 1/3 for quarks, while τ3 = ∓1.

10.5.2 Quantum chromodynamics (QCD)

The SU (3) group is the group of all 3 × 3 complex matrices with unit determinant.
The condition det U = 1 reduces the number of independent matrices to eight, so
that there are eight independent gauge fields. These matrices act on quark fermion
fields, which carry a color index with three values. The SU (3) gluon fields corre-
spond to matrices with two such indices, and are characterized either by a transition
that changes color (six matrices) or by repeated color indices, in two independent
matrices that have only diagonal elements. The Lagrangian density for this non-
abelian gauge group implies that each gluon field contributes to the color current
density. There is no empirical evidence for gluon field mass, in contrast to the
observed large mass of the weak isospin SU(2) fields.
 References and bibliography

[1] Aashamar, K., Luke, T.M. and Talman, J.D. (1978). Optimized central potentials
for atomic ground-state wavefunctions, At. Data Nucl. Data Tables 22,
443–472.
[2] Abdolsalami, M. and Morrison, M.A. (1987). Calculating vibrational-excitation
cross sections off the energy shell: A first-order adiabatic theory, Phys. Rev. A 36,
5474–5477.
[3] Adhikari, S.K. (1998). Variational Principles and the Numerical Solution of
Scattering Problems (Wiley, New York).
[4] Aissing, G. and Monkhorst, H.J. (1993). On the removal of the exchange singularity
in extended systems, Int. J. Quantum Chem. S27, 81–89.
[5] Akhiezer, N.I. (1962). The Calculus of Variations (Blaisdell, New York).
[6] Amaldi, U. et al. (1987). Comprehensive analysis of data pertaining to the weak
neutral current and the intermediate vector meson masses, Phys. Rev. D 36,
1385–1407.
[7] Andersen, O.K. (1971). Comments on the KKR wavefunctions; extension of the
spherical wave expansion beyond the muffin tins. In Computational Methods in
Band Theory, eds. P.M. Marcus, J.F. Janak, A.R. Williams (Plenum, New York),
178–182.
[8] Andersen, O.K. (1973). Simple approach to the band-structure problem, Solid State
Commun. 13, 133–136.
[9] Andersen, O.K. (1975). Linear methods in band theory, Phys. Rev. B 12, 3060–3083.
[10] Andersen, O.K. and Jepsen, O. (1977). Advances in the theory of one-electron
states, Physica B 91, 317–328.
[11] Andersen, O.K., Jepsen, O. and Glötzel, D. (1985). Canonical description of the
band structures of metals. In Highlights of Condensed-Matter Theory, Soc. Ital. Fis.
Corso 89, 59–176.
[12] Andersen, O.K. and Kasowski, R.V. (1971). Electronic states as linear combinations
of muffin-tin orbitals, Phys. Rev. B 4, 1064–1069.
[13] Andersen, O.K. and Woolley, R.G. (1973). Muffin-tin orbitals and molecular
calculations: General formalism, Mol. Phys. 26, 905–927.
[14] Baluja, K., Burke, P.G. and Morgan, L.A. (1982). R-matrix propagation program
for solving coupled second-order differential equations, Comput. Phys. Commun.
27, 299–307.
[15] Bardsley, J.N. (1968). The theory of dissociative recombination, J. Phys. B 1,
365–380.

205
206 References and bibliography

[16] Bardsley, J.N., Herzenberg, A. and Mandl, F. (1966). Electron resonances of the
H−2 ion, Proc. Phys. Soc. (London) 89, 305–320.
[17] Bardsley, J.N., Herzenberg, A. and Mandl, F. (1966). Vibrational excitation and
dissociative attachment in the scattering of electrons by hydrogen molecules, Proc.
Phys. Soc. (London) 89, 321–340.
[18] Bardsley, J.N. and Mandl, F. (1968). Resonant scattering of electrons by molecules,
Rep. Prog. Phys. 31, 471–531.
[19] Bardsley, J.N. and Wadehra, J.M. (1979). Dissociative attachment and vibrational
excitation in low-energy collisions of electrons with H2 and D2 , Phys. Rev. A 20,
1398–1405.
[20] Berman, M. and Domcke, W. (1984). Projection-operator calculations for shape
resonances: A new method based on the many-body optical approach, Phys. Rev. A
29, 2485–2496.
[21] Berman, M., Estrada, H., Cederbaum, L.S. and Domcke, W. (1983). Nuclear
dynamics in resonant electron–molecule scattering beyond the local approximation:
The 2.3-eV shape resonance in N2 , Phys. Rev. A 28, 1363–1381.
[22] Berman, M., Mündel, C. and Domcke, W. (1985). Projection-operator calculations
for molecular shape resonances: The 2 'u+ resonance in electron-hydrogen
scattering, Phys. Rev. A 31, 641–651.
[23] Bingel, W.A. (1967). A physical interpretation of the cusp conditions for molecular
wave functions, Theor. Chim. Acta 8, 54–61.
[24] Birtwistle, D.T. and Herzenberg, A. (1971). Vibrational excitation of N2 by
resonance scattering of electrons, J. Phys. B 4, 53–70.
[25] Bjorken, J.D. and Drell, S.C. (1965). Relativistic Quantum Fields (McGraw-Hill,
New York).
[26] Blanchard, P. and Brüning, E. (1992). Variational Methods in Mathematical
Physics: A Unified Approach (Springer-Verlag, Berlin).
[27] Blatt, J.M. and Biedenharn, L.C. (1952). The angular distribution of scattering and
reaction cross sections, Rev. Mod. Phys. 24, 258–272.
[28] Blatt, J.M. and Weisskopf, V.F. (1979). Theoretical Nuclear Physics
(Springer-Verlag, New York).
[29] Bleecker, D. (1985). Gauge Theory and Variational Principles (Benjamin,
New York).
[30] Bloch, C. (1957). A unified formulation of the theory of nuclear reactions, Nucl.
Phys. 4, 503–528.
[31] Born, M. and Oppenheimer, J.R. (1927). Ann. Phys. 84, 457.
[32] Brandsden, B.H. (1970). Atomic Collision Theory (Plenum, New York).
[33] Breit, G. (1959). Handbuch der Physik 41, No.1 (Springer, Berlin).
[34] Breit, G. and Wigner, E.P. (1936). Capture of slow neutrons, Phys. Rev. 49,
519–531.
[35] Brenig, W. (1957). Zweiteilchennäherungen des Mehrkörperproblems I, Nucl. Phys.
4, 363–374.
[36] Brescansin, L.M., Leite, J.R. and Ferreira, L.G. (1979). A study of the ground states
and ionization energies of H2 , C2 , N2 , F2 , and CO molecules by the variational
cellular method, J. Chem. Phys. 71, 4923–4930.
[37] Brillouin, L. (1934). Actual. Sci. Ind. 159, 28.
[38] Brown, L.M. (1958). Two-component fermion theory, Phys. Rev. 111, 957–964.
[39] Brown, R.G. and Ciftan, M. (1983). Generalized non-muffin-tin band theory, Phys.
Rev. B 27, 4564–4579.
[40] Brown, R.G. and Ciftan, M. (1983). Phys. Rev. B 28, 5992–6007.
 References and bibliography 207

[41] Broyden, C.G. (1967). Quasi-Newton methods and their application to function
minimisation, Math. Comput. 21, 368–381.
[42] Broyden, C.G. (1970). The convergence of single-rank quasi-Newton methods,
Math. Comput. 24, 365–382.
[43] Brueckner, K.A. and Wada, W. (1956). Nuclear saturation and two-body forces:
Self-consistent solutions and the effects of the exclusion principle, Phys. Rev. 103,
1008–1016.
[44] Burke, P.G. (1973). The R-matrix method in atomic physics, Comput. Phys.
Commun. 6, 288–302.
[45] Burke, P.G. (1977). Potential Scattering in Atomic Physics (Plenum, New York).
[46] Burke, P.G. and Robb, W.D. (1972). Elastic scattering of electrons by hydrogen and
helium atoms, J. Phys. B 5, 44–54.
[47] Burke, P.G. and Robb, W.D. (1975). The R-matrix theory of atomic processes, Adv.
At. Mol. Phys. 11, 143–214.
[48] Burke, P.G. and Seaton, M.J. (1971). Numerical solutions of the integro-differential
equations of electron–atom collision theory, Methods Comput. Phys. 10, 1–80.
[49] Burke, P.G. and West, J.B., eds. (1987). Electron–Molecule Scattering and
Photoionization (Plenum, New York).
[50] Burke, P.G, Hibbert, A. and Robb, W.D. (1971). Electron scattering by complex
atoms, J. Phys. B 4, 153–161.
[51] Butler, W.H. and Nesbet, R.K. (1990). Validity, accuracy, and efficiency of
multiple-scattering theory for space-filling scatterers, Phys. Rev. B 42,
1518–1525.
[52] Butler, W.H. and Zhang, X.-G. (1991). Accuracy and convergence properties of
multiple-scattering theory in three dimensions, Phys. Rev. B 44, 969–983.
[53] Butler, W.H., Zhang, X.-G. and Gonis, A. (1992). The Green function cellular
method and its relation to multiple scattering theory, Mat. Res. Symp. Proc. 253,
205–210.
[54] Buttle, P.J.A. (1967). Solution of coupled equations by R-matrix techniques,
Phys. Rev. 160, 719–729.
[55] Cairo, L. (1965). Variational Techniques in Electromagnetism, tr. G.D. Sims
(Gordon and Breach, New York).
[56] Callaway, J. and March, N.H. (1985). Density functional methods: theory and
applications, Solid State Phys. 38, 135–221.
[57] Callen, H.B. (1963). Thermodynamics (Wiley, New York).
[58] Carathéodory, C. (1967). Calculus of Variations and Partial Differential Equations
of the First Order, Part II, tr. R.B. Dean and J.J. Brandstatter (Holden-Day,
San Francisco).
[59] Cederbaum, L.S. and Domcke, W. (1981). Local against nonlocal complex potential
in resonant electron–molecule scattering, J. Phys. B 14, 4665–4689.
[60] Ceperley, D.M. and Alder, B.J. (1980). Ground state of the electron gas by a
stochastic method, Phys. Rev. Lett. 45, 566–569.
[61] Chandra, N. and Temkin, A. (1976). Hybrid theory and calculation of e – N2
scattering, Phys. Rev. A 13, 188–215.
[62] Chang, E.S. and Fano, U. (1972). Theory of electron–molecule collisions by frame
transformation, Phys. Rev. A 6, 173–185.
[63] Chang, E.S. and Temkin, A. (1969). Rotational excitation of diatomic molecules by
electron impact, Phys. Rev. Lett. 23, 399–403.
[64] Chang, E.S. and Temkin, A. (1970). Rotational excitation of diatomic molecular
systems. II. H+2 , J. Phys. Soc. Japan 29, 172–179.
208 References and bibliography

[65] Chase, D.M. (1956). Adiabatic approximation for scattering processes, Phys. Rev.
104, 838–842.
[66] Chatwin, R.A. and Purcell, J.E. (1971). Approximate solution of a Sturm–Liouville
system using nonorthogonal expansions: Application to α–α nuclear scattering,
J. Math. Phys. 12, 2024–2030.
[67] Chen, J.C.Y. (1966). Dissociative attachment in rearrangement electron collision
with molecules, Phys. Rev. 148, 66–73.
[68] Clementi, E. and Roetti, C. (1974). Roothaan–Hartree–Fock atomic wavefunctions,
At. Data Nucl. Data Tables 14, 177–478.
[69] Colle, R. and Nesbet, R.K. (2001). Optimized effective potential in finite-basis-set
treatment, J. Phys. B 34, 2475–2480.
[70] Colle, R. and Salvetti, O. (1975). Approximate calculation of the correlation energy
for the closed shells, Theoret. Chim. Acta 37, 329–334.
[71] Collins, J.C. (1984). Renormalization (Cambridge University Press, Cambridge).
[72] Condon, E.U. and Shortley, G.H. (1935). The Theory of Atomic Spectra (Cambridge
University Press, Cambridge).
[73] Cottingham, W.N. and Greenwood, D.A. (1998). An Introduction to the Standard
Model of Particle Physics (Cambridge University Press, New York).
[74] Courant, R. and Hilbert, D. (1953). Methods of Mathematical Physics (Interscience,
New York).
[75] Csaszar, P. and Pulay, P. (1984). Geometry optimization by direct inversion in the
iterative subspace, J. Mol. Struct. 114, 31–34.
[76] Damburg, R.J. and Karule, E. (1967). A modification of the close-coupling
approximation for e – H scattering allowing for the long-range interaction, Proc.
Phys. Soc. (London) 90, 637–640.
[77] Dennis, J.E. and Schnabel, R.B. (1983). Numerical Methods for Unconstrained
Optimization and Nonlinear Equations (Prentice-Hall, Englewood Cliffs,
New Jersey).
[78] Dieudonné, J. (1981). History of Functional Analysis (North-Holland, Amsterdam).
[79] Dirac, P.A.M. (1930). Note on exchange phenomena in the Thomas atom, Proc.
Camb. Phil. Soc. 26, 376–385.
[80] Dirac, P.A.M. (1933). Physikalische Z. der Sowjetunion 3, 64.
[81] Dirac, P.A.M. (1958). Quantum Mechanics, 4th edition (Oxford University Press,
London).
[82] Domcke, W. (1991). Theory of resonance and threshold effects in electron–molecule
collisions: the projection-operator approach, Phys. Rep. 208, 97–188.
[83] Domcke, W. and Cederbaum, L.S. (1977). Theory of the vibrational structure of
resonances in electron–atom scattering, Phys. Rev. A 16, 1465–1482.
[84] Domcke, W. and Cederbaum, L.S. (1981). On the interpretation of low-energy
electron-HCl scattering phenomena, J. Phys. B 14, 149–173.
[85] Domcke, W., Cederbaum, L.S. and Kaspar, F. (1979). Threshold phenomena in
electron–molecule scattering: a nonadiabatic theory, J. Phys. B 12, L359–L364.
[86] Domcke, W. and Mündel, C. (1985). Calculation of cross sections for vibrational
excitation and dissociative attachment in HCl and DCl beyond the local complex
potential approximation, J. Phys. B 18, 4491–4509.
[87] Domps, A., Reinhard, P.-G. and Suraud, E. (1998). Time-dependent Thomas–
Fermi approach to electron dynamics in metal clusters, Phys. Rev. Lett. 80,
5520–5523.
[88] Dovesi, R., Saunders, V.R., Roetti, C., Causà, M., Harrison, N.M., Orlando, R. and
Aprà, E. (1996). CRYSTAL95 User’s Manual (University of Torino, Torino, Italy).
 References and bibliography 209

[89] Dreizler, R.M. and de Providencia, J. (1985). Density Functional Methods in

Physics (Plenum, New York).
[90] Dreizler, R.M. and Gross, E.K.U. (1990). Density Functional Theory
(Springer-Verlag, Berlin).
[91] Dubé, L. and Herzenberg, A. (1979). Absolute cross sections from the “boomerang
model” for resonant electron–molecule scattering, Phys. Rev. A 20, 194–213.
[92] Durrant, A.V. (1980). A derivation of optical field quantization from absorber
theory, Proc. Roy. Soc. (London) A370, 41–59.
[93] Edelen, D.G.B. (1969). Nonlocal Variations and Local Invariance of Fields
(American Elsevier, New York).
[94] Ehrenreich, H. and Cohen, M.H. (1959). Self-consistent field approach to the
many-electron problem, Phys. Rev. 115, 786–790.
[95] Ehrhardt, H. and Willmann, K. (1967). Angular dependence of low-energy
resonance scattering of electrons by N2 , Z. Phys. 204, 462–473.
[96] Elsgolc, L.E. (1962). Calculus of Variations (Pergamon Press, London).
[97] Engel, E. and Vosko, S.H. (1993). Accurate optimized-potential-model solutions for
spherical spin-polarized atoms: Evidence for limitations of the exchange-only local
spin-density and generalized-gradient expansions, Phys. Rev. A 47, 2800–2811.
[98] Englert, F. and Brout, R. (1964). Broken symmetry and the mass of gauge vector
mesons, Phys. Rev. Lett. 13, 321–323.
[99] Englisch, H. and Englisch, R. (1984). Exact density functionals for ground-state
energies. I General results, Phys. Stat. Sol. B 123, 711–721.
[100] Englisch, H. and Englisch, R. (1984). Exact density functionals for ground-state
energies. II Details and remarks, Phys. Stat. Sol. B 124, 373–379.
[101] Epstein, S.T. (1974). The Variation Method in Quantum Chemistry (Academic
Press, New York).
[102] Eschrig, H. (1996). The Fundamentals of Density Functional Theory (Teubner,
Stuttgart).
[103] Fabrikant, I.I. (1990). Resonance processes in e-HCl collisions: comparison of the
R-matrix and the nonlocal-complex-potential methods, Comments At. Mol. Phys.
24, 37–52.
[104] Faddeev, L.D. and Slavnov, A.A. (1980). Gauge Fields: Introduction to Quantum
Theory (Benjamin/Cummings, Reading, Massachusetts).
[105] Faisal, F.H.M. and Temkin, A. (1972). Application of the adiabatic-nuclei theory to
vibrational excitation, Phys. Rev. Lett. 28, 203–206.
[106] Fano, U. and Lee, C.M. (1973). Variational calculation of R-matrices. Application
to Ar photoabsorption, Phys. Rev. Lett. 31, 1573–1576.
[107] Faulkner, J.S. (1986). Non-muffin-tin potentials in multiple scattering theory, Phys.
Rev. B 34, 5931–5934.
[108] Fermi, E. (1927). Un metodo statistice per la determinazione di alcune proprieta
dell’atomo, Rend. Accad. Lincei 6, 602–607.
[109] Fermi, E. (1928). A statistical method for the determination of some atomic
properties and the application of this method to the theory of the periodic system of
elements, Z. Phys. 48, 73–79.
[110] Ferraz, A.C., Chagas, M.I.T., Takahashi, E.K. and Leite, J.R. (1984). Variational
cellular model of the energy bands of diamond and silicon, Phys. Rev. B 29,
7003–7006.
[111] Ferraz, A.C., Takahashi, E.K. and Leite, J.R. (1982). Application of the variational
cellular method to periodic structures. Energy band of sodium, Phys. Rev. B 26,
690–700.
210 References and bibliography

[112] Ferreira, L.G. and Leite, J.R. (1978). Variational cellular model of the molecular
and crystal electronic structure, Phys. Rev. Lett. 40, 49–52.
[113] Ferreira, L.G. and Leite, J.R. (1978). General formulation of the variational cellular
method for molecules and crystals, Phys. Rev. A 18, 335–343.
[114] Ferreira, L.G. and Leite, J.R. (1979). Self-consistent calculation of the electronic
structure of N2 and CO by the variational cellular method, Phys. Rev. A 20,
689–699.
[115] Feshbach, H. (1958). Unified theory of nuclear reactions, Ann. Phys. (New York) 5,
357–390.
[116] Feshbach, H. (1962). A unified theory of nuclear reactions, Ann. Phys. (New York)
19, 287–313.
[117] Feynman, R.P. (1939). Forces in molecules, Phys. Rev. 56, 340–343.
[118] Feynman, R.P. (1948). A relativistic cut-off for classical electrodynamics, Phys.
Rev. 74, 939–946.
[119] Feynman, R.P. (1948). Space-time approach to nonrelativistic quantum mechanics,
Rev. Mod. Phys. 20, 367–387.
[120] Feynman, R.P. (1949). Space-time approach to quantum electrodynamics, Phys.
Rev. 76, 769–789.
[121] Feynman, R.P. (1961). Quantum Electrodynamics (Benjamin, New York).
[122] Feynman, R.P. and Gell-Mann, M. (1958). Theory of the Fermi interaction, Phys.
Rev. 109, 193–198.
[123] Feynman, R.P. and A.R. Hibbs (1965). Quantum Mechanics and Path Integrals
(McGraw-Hill, New York).
[124] Fletcher, R. (1970). A new approach to variable metric algorithms, Comput. J. 13,
317–322.
[125] Fletcher, R. (1987). Practical Methods of Optimization, 2nd edition (Wiley,
New York).
[126] Fliflet, A.W. and McKoy, V. (1978). Discrete basis set method for
electron–molecule continuum wave functions, Phys. Rev. A 18, 2107–2114.
[127] Fock, V. (1930). Z. Phys. 61, 126.
[128] Fock, V. (1957). The Theory of Space, Time and Gravitation, tr. N. Kemmer
(Pergamon Press, London).
[129] Fried, H. (1972). Functional Methods and Models in Quantum Field Theory (MIT
Press, Cambridge, Massachusetts).
[130] Froese Fischer, C. (1977). The Hartree–Fock Method for Atoms (Wiley, New York).
[131] Garrod, G. and Percus, J.K. (1964). Reduction of the N -particle variational
problem, J. Math. Phys. 5, 1756–1776.
[132] Gerjuoy, E., Rau, A.R.P. and Spruch, L. (1983). A unified formulation of the
construction of variational principles, Rev. Mod. Phys. 55, 725–774.
[133] Gianturco, F.A. and Stoecklin, T. (1996). The elastic scattering of electrons from
CO2 molecules: I. Close coupling calculations of integral and differential cross
sections, J. Phys. B 29, 3933–3954.
[134] Gianturco, F.A., Jain, A. and Pantano, L. (1987). Electron-methane scattering via a
parameter-free model interaction, J. Phys. B 20, 571–586.
[135] Gianturco, F.A., Thompson, D.G. and Jain, A. (1995). Electron scattering from
polyatomic molecules using a single-center expansion formulation, in
Computational Methods for Electron–Molecule Collisions, eds. W.M. Huo and
F.A. Gianturco (Plenum, New York), pp. 75–118.
[136] Gilbert, T.L. (1975). Hohenberg–Kohn theorem for nonlocal external potentials,
Phys. Rev. B 12, 2111–2120.
 References and bibliography 211

[137] Gill, P.E., Murray, W. and Wright, M.H. (1981). Practical Optimization (Academic
Press, New York).
[138] Gillan, C.J., Nagy, O., Burke, P.G., Morgan, L.A. and Noble, C.J. (1987). Electron
scattering by nitrogen molecules, J. Phys. B 20, 4585–4603.
[139] Gillan, C.J., Tennyson, J. and Burke, P.G. (1995). The UK molecular R-matrix
scattering package: a computational perspective, in Computational Methods for
Electron–Molecule Collisions, eds. W.M. Huo and F.A. Gianturco (Plenum,
New York), pp. 239–254.
[140] Glashaw, S.L. (1961). Partial symmetries of weak interactions, Nucl. Phys. 22,
579–588.
[141] Glimm, A. and Jaffe, A. (1981). Quantum Mechanics - A Functional Integral Point
of View (Springer-Verlag, Berlin).
[142] Glowinski, R. (1983). Numerical Methods for Nonlinear Variational Problems
(Springer-Verlag, Berlin).
[143] Golden, D.E. (1966). Low-energy resonances in e− −N2 total scattering cross
sections: the temporary formation of N− 2 , Phys. Rev. Lett. 17, 847–851.
[144] Golden, D.E., Lane, N.F., Temkin, A. and Gerjuoy, E. (1971). Low energy
electron–molecule scattering experiments and the theory of rotational excitation,
Rev. Mod. Phys. 43, 642–678.
[145] Goldfarb, D. (1970). A family of variable-metric methods derived by variational
means, Math. Comput. 24, 23–26.
[146] Goldstein, H. (1983). Classical Mechanics, 2nd edition (Wiley, New York).
[147] Goldstine, H.H. (1980). A History of the Calculus of Variations from the 17th
Through the 19th Century (Springer-Verlag, Berlin).
[148] Gonis, A. and Butler, W.H. (2000). Multiple Scattering in Solids (Springer,
New York).
[149] Gonze, X. and Scheffler, M. (1999). Exchange and correlation kernels at the
resonance frequency: implications for excitation energies in density-functional
theory, Phys. Rev. Lett. 82, 4416–4419.
[150] Gordon, R.G. and Kim, Y.S. (1972). Theory for the forces between closed-shell
atoms and molecules, J. Chem. Phys. 56, 3122–3133.
[151] Görling, A. (1999). New KS method for molecules based on an exchange charge
density generating the exact KS exchange potential, Phys. Rev. Lett. 83,
5459–5462.
[152] Görling, A. and Ernzerhof, M. (1995). Energy differences between Kohn–Sham and
Hartree–Fock wave functions yielding the same electron density, Phys. Rev. A 51,
4501–4513.
[153] Grimm-Bosbach, T., Thümmel, H.T., Nesbet, R.K. and Peyerimhoff, S.D. (1996).
Calculation of cross sections for rovibrational excitation of N2 by electron impact,
J. Phys. B 29, L105–L112.
[154] Gross, E.K.U. and Kohn, W. (1985). Local density-functional theory of
frequency-dependent linear response, Phys. Rev. Lett. 55, 2850–2852.
[155] Gross, E.K.U., Dobson, J.F. and Petersilka, M. (1996). Density functional theory of
time-dependent phenomena, in Density Functional Theory, ed. R.F. Nalewajski,
Series “Topics in Current Chemistry” (Springer, Berlin).
[156] Guralnik, G.S., Hagen, C.R. and Kibble, T.W.B. (1964). Global conservation laws
and massless particles, Phys. Rev. Lett. 13, 585–587.
[157] Gyorffy, B.L. (1972). Coherent-potential approximation for a nonoverlapping
muffin-tin potential model of random substitutional alloys, Phys. Rev. B 5,
2382–2384.
212 References and bibliography

[158] Ham, F. and Segal, B. (1961). Energy bands in periodic lattices – Green’s function
method, Phys. Rev. 124, 1786–1796.
[159] Harris, J. (1979). The role of occupation numbers in HKS theory, Int. J. Quantum
Chem. S13, 189–192.
[160] Harris, J. (1984). Adiabatic-connection approach to Kohn–Sham theory, Phys. Rev.
A 29, 1648–1659.
[161] Harris, J. and Jones, R.O. (1974). The surface energy of a bounded electron gas,
J. Phys. F 4, 1170–1186.
[162] Hartree, D.H. (1928). The wave mechanics of an atom with a non-Coulomb central
field. Part II. Some results and discussion, Proc. Camb. Phil. Soc. 24, 111–132.
[163] Hartree, D.H. (1957). The Calculation of Atomic Structures (Wiley, New York).
[164] Hazi, A.U., Orel, A.E. and Rescigno, T.N. (1981). Ab initio study of dissociative
attachment of low-energy electrons to F2 , Phys. Rev. Lett. 46, 918–922.
[165] Head, J.D., Weiner, B. and Zerner, M.C. (1988). A survey of optimization
procedures for stable structures and transition states, Int. J. Quantum Chem. 33,
177–186.
[166] Head, J.D. and Zerner, M.C. (1988). Newton-based optimization methods for
obtaining molecular conformation, Adv. Quantum Chem. 20, 241–290.
[167] Herzenberg, A. (1968). Oscillatory energy dependence of resonant
electron–molecule scattering, J. Phys. B 1, 548–558.
[168] Hestenes, M.R. (1966). Calculus of Variations and Optimal Control Theory (Wiley,
New York).
[169] Higgs, P.W. (1964). Broken symmetries and the masses of gauge bosons, Phys. Rev.
Lett. 13, 508–509.
[170] Higgs, P.W. (1966). Spontaneous symmetry breakdown without massless bosons,
Phys. Rev. 145, 1156–1163.
[171] Hill, E.L. (1951). Hamilton’s principle and the conservation theorems of
mathematical physics, Rev. Mod. Phys. 23, 253–260.
[172] Höck, A. and Engel, E. (1998). Pseudopotentials from orbital-dependent
exchange-correlation functionals, Phys. Rev. A 58, 3578–3581.
[173] Hohenberg, P. and Kohn, W. (1964). Inhomogeneous electron gas, Phys. Rev. 136,
B864–B871.
[174] Hoyle, F.M. and Narlikar, J.V. (1974). Action at a Distance in Physics and
Cosmology (Freeman, San Francisco).
[175] Hulthén, L. (1944). Kgl. Fysiogr. Sällsk. Lund Förh. 14, No.21.
[176] Hulthén, L. (1948). Arkiv. Mat. Astron. Fysik 35A, No.25.
[177] Huo, W.M. (1995). The Schwinger variational method, in Computational Methods
for Electron–Molecule Collisions, eds. W.M. Huo and F.A. Gianturco (Plenum,
New York), pp. 326–355.
[178] Huo, W.M. and Gianturco, F.A. (1995). Computational Methods for Electron–
Molecule Collisions (Plenum, New York).
[179] Hurley, A.C. (1954). The electrostatic calculation of molecular energies II.
Approximate wave functions and the electrostatic method, Proc. Roy. Soc. A 226,
179–192.
[180] Hurley, A.C. (1976). Electron Correlation in Small Molecules (Academic Press,
New York).
[181] Ivanov, S., Hirata, S. and Bartlett, R.J. (1999). Exact exchange treatment for
molecules in finite-basis-set Kohn–Sham theory, Phys. Rev. Lett. 83,
5455–5458.
[182] Jackson, J.D. (1975). Classical Electrodynamics, 2nd edition (Wiley, New York).
 References and bibliography 213

[183] Jackson, J.L. (1951). A variational appeoach to nuclear reactions, Phys. Rev. 83,
301–304.
[184] Jacob, M. and Wick, G.C. (1959). On the general theory of collisions for particles
with spin, Ann. Phys. (New York) 7, 404–428.
[185] Janak, J.F. (1978). Proof that ∂ E/∂n i = i in density-functional theory, Phys. Rev.
B 18, 7165–7168.
[186] Jauch, J.M. and Rohrlich, F. (1985). The Theory of Photons and Electrons
(Springer-Verlag, Berlin).
[187] Jeffreys, H. and Jeffreys, B.S. (1956). Methods of Mathematical Physics, 3rd
edition (Cambridge University Press, New York).
[188] Joachain, C.J. (1975). Quantum Collision Theory, Vol.1 (North-Holland,
Amsterdam).
[189] Jones, R.O. and Gunnarsson, O. (1989). The density functional formalism, its
applications and prospects, Rev. Mod. Phys. 61, 689–746.
[190] Itzykson, C. and Zuber, J.-B. (1980). Quantum Field Theory (McGraw-Hill,
New York).
[191] Kamimura, M. (1977). A coupled channel variational method for microscopic study
of reactions between complex nuclei, Prog. Theor. Phys. Suppl. 62, 236–294.
[192] Kapur, P.L. and Peierls, R. (1938). The dispersion formula for nuclear reactions,
Proc. Roy. Soc. (London) A166, 277–295.
[193] Kasowski, R.V. and Andersen, O.K. (1972). Muffin tin orbitals in open structures,
Solid State Commun. 11, 799–802.
[194] Kato, T. (1950). Variational methods in collision problems, Phys. Rev. 80, 475.
[195] Kato, T. (1951). Upper and lower bounds of scattering phases, Prog. Theor. Phys. 6,
394–407.
[196] Kato, T. (1957). On the eigenfunctions of many-particle systems in quantum
mechanics, Commun. Pure Appl. Math. 10, 151–177.
[197] Kaufmann, K., Baumeister, W. and Jungen, M. (1989). Universal Gaussian basis
sets for an optimum representation of Rydberg and continuum wavefunctions,
J. Phys. B 22, 2223–2240.
[198] Kibble, T.W.B. (1967). Symmetry breaking in non-abelian gauge theories, Phys.
Rev. 155, 1554–1561.
[199] Klein, F. (1918). Nachr. Ges. Wiss. Göttingen, 171.
[200] Kline, M. (1972). Mathematical Thought from Ancient to Modern Times, 3 vols.
(Oxford University Press, New York).
[201] Kohl, H. and Dreizler, R.M. (1986). Time-dependent density-functional theory:
conceptual and practical aspects, Phys. Rev. Lett. 56, 1993–1995.
[202] Kohn, W. (1948). Variational methods in nuclear collision problems, Phys. Rev. 74,
1763–1772.
[203] Kohn, W. (1952). Variational methods for periodic lattices, Phys. Rev. 87, 472–481.
[204] Kohn, W. and Rostoker, N. (1954). Solution of the Schrödinger equation in periodic
lattices with an application to metallic lithium, Phys. Rev. 94, 1111–1120.
[205] Kohn, W. and Sham, L.J. (1965). Self consistent equations including exchange and
correlation effects, Phys. Rev. 140, A1133–A1138.
[206] Koopmans, T. (1933). Über die Zuordnung von Wellenfunktionen und Eigenwerten
zu den einzelnen Elektronen eines Atoms, Physica 1,104–113.
[207] Korringa, J. (1947). On the calculation of the energy of a Bloch wave in a metal,
Physica 13, 392–400.
[208] Kuperschmidt, B.A. (1990). The Variational Principles of Dynamics (World
Scientific, New York).
214 References and bibliography

[209] Lai, C.H. (1981). Gauge Theory of Electromagnetic and Weak Interactions, ed.
C.H. Lai (World Scientific, Singapore).
[210] Lanczos, C. (1966). Variational Principles of Mechanics (University of Toronto
Press, Toronto).
[211] Landau, L.D. (1956). The theory of a Fermi liquid, Zh. Eksp. Teor. Fiz. 30,
1058–1064. [Sov. Phys. JETP 3, 920–925 (1956)]
[212] Landau, L.D. (1957). Oscillations in a Fermi liquid, Zh. Eksp. Teor. Fiz. 32, 59–66.
[Sov. Phys. JETP 5, 101–108 (1957)]
[213] Landau, L.D. and Lifshitz, E.M. (1958). Quantum Mechanics, Non-Relativistic
Theory, translated by J.B. Sykes and J.S. Bell (Pergamon, London).
[214] Lane, A.M. and Thomas, R.G. (1958). R-matrix theory of nuclear reactions, Rev.
Mod. Phys. 30, 257–353.
[215] Lane, N.F. (1980). The theory of electron–molecule collisions, Rev. Mod. Phys. 52,
29–119.
[216] Langreth, D.C. and Mehl, M.J. (1981). Easily implementable nonlocal
exchange-correlation energy functional, Phys. Rev. Lett. 47, 446–450.
[217] Leader, E. and Predazzi, E. (1996). An Introduction to Gauge Theories and Modern
Particle Physics, Vols. 1 and 2 (Cambridge University Press, New York).
[218] Lee, B.W. and Zinn-Justin, J. (1972). Spontaneously broken gauge symmetries.
I. Preliminaries, Phys. Rev. D 5, 3121–3137.
[219] Lee, C., Yang, W. and Parr, R.G. (1988). Development of the Colle-Salvetti
correlation energy formula into a functional of the electron density, Phys. Rev. B 37,
785–789.
[220] Lee, C.M. (1974). Spectroscopy and collision theory. III. Atomic eigenchannel
calculation by a Hartree–Fock-Roothaan method, Phys. Rev. A 10, 584–600.
[221] Leibbrandt, G. (1987). Introduction to noncovariant gauges, Rev. Mod. Phys. 59,
1067–1119.
[222] Levy, M. (1979). Universal variational functionals of electron densities, first-order
density matrices, and natural spin-orbitals and solution of the v-representability
problem, Proc. Natl. Acad. Sci. 76, 6062–6065.
[223] Lieb, E.H. (1983). Density functionals for Coulomb systems, Int. J. Quantum
Chem. 24, 243–277.
[224] Lieb, E.H. and Simon, B. (1977). The Thomas–Fermi theory of atoms, molecules
and solids, Adv. Math. 23, 22–216.
[225] Light, J.C. and Walker, R.B. (1976). An R-matrix approach to the solution of
coupled equations for atom–molecule scattering, J. Chem. Phys. 65, 4272–4282.
[226] Lima, M.A.P., Leite, J.R. and Fazio, A. (1971). Theoretical study of the F2 molecule
using the variational cellular method, J. Phys. B 14, L533–L535.
[227] Lino, A.T., Leite, J.R., Ferraz, A.C. and Takahashi, E.K. (1987). Self-consistent
formulation of the variational cellular method applied to periodic structures –
results for sodium and silicon, J. Phys. Chem. Solids 48, 911–919.
[228] Lippmann, B.A. and Schwinger, J. (1950). Variational principles for scattering
processes. I, Phys. Rev. 79, 469–480.
[229] Loucks, T.L. (1967). Augmented Plane Wave Method (Benjamin, New York),
pp. 98–103.
[230] Lucchese, R.R., Takatsuka, K. and McKoy, V. (1986). Applications of the
Schwinger variational principle to electron–molecule collisions and molecular
photoionization, Phys. Rep. 131, 147–221.
[231] March, N.H. (1957). The Thomas–Fermi approximation in quantum mechanics,
Adv. Phys. 6, 1–101.
 References and bibliography 215

[232] March, N.H. (1975). Self-consistent Fields in Atoms (Pergamon Press, Oxford).
[233] Mazevet, S., Morrison, M.A., Boydstun, O. and Nesbet, R.K. (1999). Inclusion of
nonadiabatic effects in calculations on vibrational excitation of molecular hydrogen
by low-energy electron impact, Phys. Rev. A 59, 477–489.
[234] Mazevet, S., Morrison, M.A., Boydstun, O. and Nesbet, R.K. (1999). Adiabatic
treatments of vibrational dynamics in low-energy electron–molecule scattering,
J. Phys. B 32, 1269–1294.
[235] Mazevet, S., Morrison, M.A., Morgan, L.A. and Nesbet, R.K. (2001). Virtual-state
effects on elastic scattering and vibrational excitation of CO2 by electron impact.
Phys. Rev. A 64, 040701.
[236] Mazevet, S., Morrison, M.A. and Nesbet, R.K. (1998). Application of the
nonadiabatic phase matrix method to vibrational excitation near a short-lived
resonance: the case of e-H2 scattering, J. Phys. B 31, 4437–4448.
[237] McCurdy, C.W., Rescigno, T.N. and Schneider, B.I. (1987). Interrelation between
variational principles for scattering amplitudes and generalized R-matrix theory,
Phys. Rev. A 36, 2061–2066.
[238] McLachlan, A.D. and Ball, M.A. (1964). Time-dependent Hartree–Fock theory for
molecules, Rev. Mod. Phys. 36, 844–855.
[239] McWeeny, R. (1989). Methods of Molecular Quantum Mechanics (Academic Press,
New York).
[240] Mercier, A. (1959). Analytical and Canonical Formalism in Physics (Interscience,
New York).
[241] Mermin, N.D. (1965). Thermal properties of the inhomogeneous electron gas, Phys.
Rev. 137, A1441–A1443.
[242] Merzbacher, E. (1961). Quantum Mechanics (Wiley, New York).
[243] Messiah, A. (1961). Quantum Mechanics, Vol. I (North Holland, Amsterdam).
[244] Miller, W.H. and Jansen op de Haar, B.M.D.D. (1987). A new basis set for quantum
scattering calculations, J. Chem. Phys. 86, 6213–6220.
[245] Mito, Y. and Kamimura, M. (1976). The generator coordinate method for
composite-particle scattering based on the Kohn-Hulthén variational principle,
Prog. Theor. Phys. 56, 583–598.
[246] Moiseiwitsch, B.L. (1966). Variational Principles (Interscience, New York).
[247] Morgan, J. van W. (1977). Integration of Poisson’s equation for a complex system
with arbitrary geometry, J. Phys. C 10, 1181–1202.
[248] Morgan, L.A. (1984). A generalized R-matrix propagation program for solving
coupled second-order differential equations, Comput. Phys. Commun. 31, 419–422.
[249] Morgan, L.A. (1998). Virtual states and resonances in electron scattering by CO2 ,
Phys. Rev. Lett. 80, 1873–1875.
[250] Morgan, L.A. and Burke, P.G. (1988). Low-energy electron scattering by HF,
J. Phys. B 21, 2091–2105.
[251] Morgan, L.A., Burke, P.G. and Gillan, C.J. (1990). Low-energy electron scattering
by HCl, J. Phys. B 23, 99–113.
[252] Morrison, M.A. (1986). A first-order nondegenerate adiabatic theory for calculating
near-threshold cross sections for rovibrational excitation of molecules by electron
impact, J. Phys. B 19, L707–L715.
[253] Morrison, M.A. (1988). Near-threshold electron–molecule scattering, Adv. At. Mol.
Phys. 24, 51–156.
[254] Morrison, M.A., Abdolsalami, M. and Elza, B.K. (1991). Improved accuracy in
adiabatic cross sections for low-energy rotational and vibrational excitation, Phys.
Rev. A 43, 3440–3459.
216 References and bibliography

[255] Morse, P.M. and Feshbach, H. (1953). Methods of Theoretical Physics, Vols. I
and II (McGraw-Hill, New York).
[256] Moruzzi, V.L., Janak, J.F. and Williams, A.R. (1978). Calculated Electronic
Properties of Metals (Pergamon, New York).
[257] Mott, N.F. and Massey, H.S.W. (1965). The Theory of Atomic Collisions (Oxford
University Press, New York).
[258] Mündel, C. and Domcke, W. (1984). Nuclear dynamics in resonant electron–
molecule scattering beyond the local approximation: model calculations on
dissociative attachment and vibrational excitation, J. Phys. B 17, 3593–3616.
[259] Mündel, C., Berman, M. and Domcke, W. (1985). Nuclear dynamics in
resonant electron–molecule scattering beyond the local approximation: Vibrational
excitation and dissociative attachment in H2 and D2 , Phys. Rev. A 32,
181–193.
[260] Nagy, A. (1995). Coordinate scaling and adiabatic connection formula for
ensembles of fractionally occupied excited states, Int. J. Quantum Chem. 56,
225–228.
[261] Nesbet, R.K. (1955). Configuration interaction in orbital theories, Proc. Roy. Soc. A
230, 312–321.
[262] Nesbet, R.K. (1961). Approximate methods in the quantum theory of
many-Fermion systems, Rev. Mod. Phys. 33, 28–36.
[263] Nesbet, R.K. (1961). Construction of symmetry-adapted functions in the
many-body problem, J. Math. Phys. 2, 701–709.
[264] Nesbet, R.K. (1968). Analysis of the Harris variational method in scattering theory,
Phys. Rev. 175, 134–142.
[265] Nesbet, R.K. (1969). Anomaly-free variational method for inelastic scattering,
Phys. Rev. 179, 60–70.
[266] Nesbet, R.K. (1971). Where semiclassical radiation theory fails, Phys. Rev. Lett. 27,
553–556.
[267] Nesbet, R.K. (1979). Energy-modified adiabatic approximation for scattering
theory, Phys. Rev. A 19, 551–556.
[268] Nesbet, R.K. (1979). Accurate e− −He cross sections below 19 eV, J. Phys. B 12,
L243–L248.
[269] Nesbet, R.K. (1979). Variational calculations of accurate e− −He cross sections
below 19 eV, Phys. Rev. A 20, 58–70.
[270] Nesbet, R.K. (1980). Variational Methods in Electron–Atom Scattering Theory
(Plenum, New York).
[271] Nesbet, R.K. (1981). The concept of a local complex potential for nuclear motion in
electron–molecule collisions, Comments At. Mol. Phys. 11. 25–35.
[272] Nesbet, R.K. (1984). R-matrix formalism for local cells of arbitrary geometry.
Phys. Rev. B 30, 4230–4234.
[273] Nesbet, R.K. (1984). Asymptotic distorted-wave approximation for
electron–molecule scattering. J. Phys. B 17, L897–L900.
[274] Nesbet, R.K. (1986). Linearized atomic-cell orbital method for energy-band
calculations, Phys. Rev. B 33, 8027–8034.
[275] Nesbet, R.K. (1986). Nonperturbative theory of exchange and correlation in
one-electron quasiparticle states, Phys. Rev. B 34, 1526–1538.
[276] Nesbet, R.K. (1988). Variational methods for cellular models, Phys. Rev. A 38,
4955–4960.
[277] Nesbet, R.K. (1990). Full-potential multiple scattering theory, Phys. Rev. B 41,
4948–4952.
 References and bibliography 217

[278] Nesbet, R.K. (1990). Atomic cell method for total energy calculations, Bull. Am.
Phys. Soc. 35, 418.
[279] Nesbet, R.K. (1992). Variational principles for full-potential multiple scattering
theory, Mat. Res. Symp. Proc. 253, 153–158.
[280] Nesbet, R.K. (1992). Internal sums in full-potential multiple scattering theory, Phys.
Rev. B 45, 11491–11495.
[281] Nesbet, R.K. (1992). Full-potential revision of coherent-potential-approximation
alloy theory, Phys. Rev. B 45, 13234–13238.
[282] Nesbet, R.K. (1992). Full-potential multiple scattering theory without structure
constants, Phys. Rev. B 46, 9935–9939.
[283] Nesbet, R.K. (1996). Alternative density functional theory for atoms and molecules,
J. Phys. B 29, L173–L179.
[284] Nesbet, R.K. (1996). Nonadiabatic phase-matrix method for vibrational excitation
and dissociative attachment in electron–molecule scattering, Phys. Rev. A 54,
2899–2905.
[285] Nesbet, R.K. (1997). Fractional occupation numbers in density-functional theory,
Phys. Rev. A 56. 2665–2669.
[286] Nesbet, R.K. (1997). Local response model of the generalized polarization
potential, Phys. Rev. A 56. 2778–2783.
[287] Nesbet, R.K. (1997). Recent developments in multiple scattering theory and density
functional theory for molecules and solids, in Conceptual Perspectives in Quantum
Chemistry, eds. J.-L. Calais and E. Kryachko (Kluwer, Dordrecht), pp. 1–58.
[288] Nesbet, R.K. (1998). Kinetic energy in density-functional theory, Phys. Rev. A 58,
R12–R15.
[289] Nesbet, R.K. (1999). Exact exchange in linear-response theory, Phys. Rev. A 60,
R3343–R3346.
[290] Nesbet, R.K. (2000). Bound-free correlation in electron scattering by atoms and
molecules, Phys. Rev. A 62, 040701(R).
[291] Nesbet, R.K. (2001). Local potentials in independent-electron models, Int. J.
Quantum Chem. 81, 384–388.
[292] Nesbet, R.K. (2002). Orbital functional theory of linear response and excitation,
Int. J. Quantum Chem. 86, 342–346.
[293] Nesbet, R.K. and Colle, R. (1999). Does an exact local exchange potential exist?
J. Math. Chem. 26, 233–242.
[294] Nesbet, R.K. and Colle, R. (2000). Tests of the locality of exact Kohn–Sham
exchange potentials, Phys. Rev. A 61, 012503.
[295] Nesbet, R.K. and Grimm-Bosbach, T. (1993). Use of a discrete complete set of
vibrational basis functions in nonadiabatic theories of electron–molecule scattering,
J. Phys. B 26, L423–L426.
[296] Nesbet, R.K. and Sun, T. (1987). Self-consistent calculations using canonical
scaling in the linearized atomic-cell orbital method: Energy bands of fcc Cu, Phys.
Rev. B 36, 6351–6355.
[297] Nesbet, R.K., Mazevet, S. and Morrison, M.A. (2001). Procedure for correcting
variational R-matrix calculations for polarization response, Phys. Rev. A 64,
034702.
[298] Nesbet, R.K., Noble, C.J. and Morgan, L.A. (1986). Calculations of elastic electron
scattering by H2 for fixed nuclei, Phys. Rev. A 34, 2798–2807.
[299] Nesbet, R.K., Noble, C.J., Morgan, L.A. and Weatherford, C.A. (1984). Variational
R-matrix calculations of e− + H2 scattering using numerical asymptotic basis
functions, J. Phys. B 17, L891–L895.
218 References and bibliography

[300] Nestmann, B.M. and Peyerimhoff, S.D. (1990). Optimized Gaussian basis sets for
representation of continuum wave functions, J. Phys. B 22, L773–L777.
[301] Nestmann, B.M., Nesbet, R.K. and Peyerimhoff, S.D. (1991). A concept for
improving the efficiency of R-matrix calculations for electron–molecule scattering,
J. Phys. B 24, 5133–5149.
[302] Newton, R.G. (1966). Scattering Theory of Waves and Particles (Plenum,
New York).
[303] Noble, C.J. and Nesbet, R.K. (1984). CFASYM, a program for the calculation of the
asymptotic solutions of the coupled equations of electron collision theory, Comput.
Phys. Commun. 33, 399–411.
[304] Noether, E. (1918). Nachr. Ges. Wiss. Göttingen, 235.
[305] Norcross, D.W. and Padial, N.T. (1982). The multipole-extracted adiabatic nuclei
approximation for electron–molecule collisions, Phys. Rev. A 25, 226–238.
[306] Nordholm, S. and Bacskay, G. (1978). Generalizations of the finite-element and
R-matrix methods, J. Phys. B 11, 193–207.
[307] Norman, M.R. and Koelling, D.D. (1984). Towards a Kohn–Sham potential via the
optimized effective potential method, Phys. Rev. B 30, 5530–5540.
[308] Nozières, P. (1964). Theory of Fermi liquids, in 1962 Cargèse Lectures in
Theoretical Physics, ed. M. Lévy (Benjamin, New York), IV, 1–37,
[309] Nozières, P. (1964). Theory of Interacting Fermi Systems (Benjamin, New York).
[310] Oberoi, R.S. and Nesbet, R.K. (1973). Variational formulation of the R-matrix
method for multichannel scattering, Phys. Rev. A 8, 215–219.
[311] Oberoi, R.S. and Nesbet, R.K. (1973). Numerical asymptotic functions in
variational scattering theory, J. Comput. Phys. 12, 526–533.
[312] Oberoi, R.S. and Nesbet, R.K. (1974). Addendum to “Variational formulation of the
R matrix method for multichannel scattering”, Phys. Rev. A 9, 2804–2805.
[313] Oksyuk, Yu.D. (1966). Excitation of the rotational levels of diatomic molecules by
electron impact in the adiabatic approximation, Zh. Eksp. Teor. Fiz. 49, 1261–1273.
[Sov. Phys. JETP 22, 873–881 (1966)]
[314] O’Malley, T.F. (1966). Theory of dissociative attachment, Phys. Rev. 150, 14–29.
[315] O’Malley, T.F. (1967). Theory of dissociative attachment (Erratum), Phys. Rev.
156, 230.
[316] Ortiz, J.V. (1999). Toward an exact one-electron picture of chemical bonding, Adv.
Quantum Chem. 35, 33–52.
[317] Padial, N.T. and Norcross, D.W. (1984). Ro-vibrational excitation of HCl by
electron impact, Phys. Rev. A 29, 1590–1593.
[318] Panofsky, W.K.H. and Phillips, M. (1969). Classical Electricity and Magnetism
(Addison-Wesley, Reading, Mass.).
[319] Parr, R.G. (1983). Density functional theory, Ann. Rev. Phys. Chem. 34, 631–656.
[320] Parr, R.G. and Bartolotti, L.J. (1983). Some remarks on the density functional
theory of few-electron systems, J. Phys. Chem. 87, 2810–2815.
[321] Parr, R.G. and Yang, W. (1989). Density-Functional Theory of Atoms and
Molecules (Oxford University Press, New York).
[322] Pars, L.A. (1962). An Introduction to the Calculus of Variations (Wiley,
New York).
[323] Pauli, W. (1958). Theory of Relativity, tr. G. Field (Pergamon Press, New York).
[324] Payne, P.W. (1979). Density functionals in unrestricted Hartree–Fock theory,
J. Chem. Phys. 71, 490–496.
[325] Percival, I.C. and Seaton, M.J. (1957). The partial wave theory of
electron-hydrogen atom collisions, Proc. Camb. Phil. Soc. 53, 654–662.
 References and bibliography 219

[326] Perdew, J.P. and Norman, M.R. (1982). Electron removal energies in Kohn–Sham
density-functional theory, Phys. Rev. B 26, 5445–5450.
[327] Perdew, J.P. and Norman, M.R. (1984). Reply to Comment on ‘Electron removal
energies in Kohn–Sham density-functional theory’, Phys. Rev. B 30, 3525–3526.
[328] Perdew, J.P., Parr, R.G., Levy, M. and Balduz, J.L. (1982). Density-functional
theory for fractional particle number: derivative discontinuities of the energy, Phys.
Rev. Lett. 49, 1691–1694.
[329] Petersilka, M., Gossmann, U.J. and Gross, E.K.U. (1996). Excitation energies from
time-dependent density-functional theory, Phys. Rev. Lett. 76, 1212–1215.
[330] Pfingst, K., Nestmann, B.M. and Peyerimhoff, S.D. (1995). Tailoring the R-matrix
approach for application to polyatomic molecules, in Computational Methods for
Electron–Molecule Collisions, eds. W.M. Huo and F. Gianturco (Plenum,
New York), pp. 293–308.
[331] Pokorski, S. (1987) Gauge Field Theories (Cambridge University Press,
Cambridge).
[332] Pulay, P. (1987). Analytical derivative methods in quantum chemistry, Adv. Chem.
Phys. 69, 241–286.
[333] Pulay, P. and Fogarasi, G. (1992). Geometry optimization in redundant internal
coordinates, J. Chem. Phys. 96, 2857–2860.
[334] Purcell, J.E. (1969). Nuclear structure calculations in the continuum-application to
neutron-carbon scattering, Phys. Rev. 185, 1279–1285.
[335] Quigg, C. (1983). Gauge Theories of the Strong, Weak, and Electromagnetic
Interactions (Benjamin-Cummings, New York).
[336] Ramond, P. (1989). Field Theory: A Modern Primer, 2nd edition (Addison-Wesley,
Redwood City, California).
[337] Lord Rayleigh (1892). Phil. Mag. 34, 481.
[338] Lord Rayleigh (1892). The Theory of Sound, 2nd edition, Vol. 2, p. 328, reprinted
by Dover Publications, New York (1976).
[339] Reed, M. and Simon, B. (1980). Methods of Modern Mathematical Physics I:
Functional Analysis (Academic Press, New York).
[340] Renton, P. (1990). Electroweak Interactions (Cambridge University Press, New
York).
[341] Rescigno, T.N., McCurdy, C.W., Orel, A.E. and Lengsfield, B.H., III (1995). The
complex Kohn variational method, in Computational Methods for
Electron–Molecule Collisions, eds. W.M. Huo and F.A. Gianturco (Plenum,
New York), pp. 1–44.
[342] Rescigno, T.N., McCurdy, C.W. and McKoy, V. (1974). Discrete basis set approach
to nonspherical scattering, Chem. Phys. Lett. 27, 401–404.
[343] Rescigno, T.N., McCurdy, C.W. and McKoy, V. (1974). Discrete basis set approach
to nonspherical scattering. II., Phys. Rev. A 10, 2240–2245.
[344] Rescigno, T.N., McCurdy, C.W. and McKoy, V. (1975). Low energy e− −H2 elastic
cross sections using discrete basis functions, Phys. Rev. A 11, 825–829.
[345] Rivers, R.J. (1987). Path Integral Methods in Quantum Field Theory (Cambridge
University Press, New York).
[346] Rohr, K. and Linder, F. (1976). Vibrational excitation of polar molecules by
electron impact I. Threshold resonances in HF and HCl, J. Phys. B 9, 2521–2537.
[347] Roothaan, C.C.J. (1951). New developments in molecular orbital theory, Rev. Mod.
Phys. 23, 69–89.
[348] Roothaan, C.C.J. (1960). Self-consistent field theory for open shells of electronic
systems, Rev. Mod. Phys. 32, 179–185.
220 References and bibliography

[349] Roothaan, C.C.J. and Bagus, P.S. (1963). Atomic self-consistent field calculations
by the expansion method, Methods Comput. Phys. 2, 47–94.
[350] Rubinow, S.I. (1955). Variational principle for scattering with tensor forces, Phys.
Rev. 98, 183–187.
[351] Runge, E. and Gross, E.K.U. (1984). Density-functional theory for time-dependent
systems, Phys. Rev. Lett. 52, 997–1000.
[352] Ryder, L.H. (1985). Quantum Field Theory (Cambridge University Press,
Cambridge).
[353] Salam, A. (1968). In Proceedings of the Eighth Nobel Symposium, ed. N. Svartholm
(Almqvist and Wiksell, Stockholm).
[354] Sarpal, B.J., Tennyson, J. and Morgan, L.A. (1991). Vibrationally resolved electron
HeH+ collisions using the nonadiabatic R-matrix method, J. Phys. B 24,
1851–1866.
[355] Sarpal, B.J., Tennyson, J. and Morgan, L.A. (1994). Dissociative recombination
without curve crossing: study of HeH+ , J. Phys. B 27, 5943–5953.
[356] Schlegel, H.B. (1982). Optimization of equilibrium geometries and transition
structures, J. Comput. Chem. 3, 214–218.
[357] Schlegel, H.B. (1987). Optimization of equilibrium geometries and transition
structures, Adv. Chem. Phys. 67, 249–286.
[358] Schlessinger, L. and Payne, G.L. (1974). Procedure for finding the scattering
solutions of the Schrödinger equation, Phys. Rev. A 10, 1559–1567.
[359] Schlosser, H. and Marcus, P. (1963). Composite wave variational method for
solution of the energy-band problem in solids, Phys. Rev. 131, 2529–2546.
[360] Schneider, B.I. (1975). R-matrix theory for electron–atom and electron–molecule
collisions using analytic basis set expansions. Chem. Phys. Lett. 31, 237–241.
[361] Schneider, B.I. (1976). Role of the Born–Oppenheimer approximation in the
vibrational excitation of molecules by electrons, Phys. Rev. A 14, 1923–1925.
[362] Schneider, B.I. (1995). An R-matrix approach to electron–molecule collisions, in
Computational Methods for Electron–Molecule Collisions, eds. W.M. Huo and
F. Gianturco (Plenum, New York), pp. 213–226.
[363] Schneider, B.I., LeDourneuf, M. and Burke, P.G. (1979). Theory of vibrational
excitation and dissociative attachment: an R-matrix approach, J. Phys. B 12,
L365–L369.
[364] Schneider, B.I., LeDourneuf, M. and Vo Ky Lan (1979). Resonant vibrational
excitation of N2 by low-energy electrons: an ab initio R-matrix calculation, Phys.
Rev. Lett. 43, 1926–1929.
[365] Schrödinger, E. (1926). Quantisiering als Eigenwertproblem, Ann. der Physik 81,
109–139.
[366] Schulman, L.S. (1981). Techniques and Applications of Path Integration (Wiley,
New York).
[367] Schulz, G.J. (1962). Vibrational excitation of nitrogen by electron impact, Phys.
Rev. 125, 229–232.
[368] Schulz, G.J. (1964). Vibrational excitation of N2 , CO, and H2 by electron impact,
Phys. Rev. 135, A988–A994.
[369] Schwartz, C. (1961). Variational calculations of scattering, Ann. Phys. (N.Y.) 16,
36–50.
[370] Schwartz, C. (1961). Electron scattering from hydrogen, Phys. Rev. 124,
1468–1471.
[371] Schwinger, J. (1947). A variational principle for scattering problems, Phys. Rev. 72,
742.
 References and bibliography 221

[372] Schwinger, J. (1951). On the Green’s functions of quantized fields, Proc. Nat. Acad.
Sci. 37, 452–455.
[373] Schwinger, J. (1958). Quantum Electrodynamics (Dover, New York).
[374] Seaton, M.J. (1953). The Hartree–Fock equations for continuous states with
applications to electron excitation of the ground configuration terms of O I, Phil.
Trans. Roy. Soc. (London) A 245, 469–499.
[375] Seaton, M.J. (1966). A variational principle for the scattering phase matrix, Proc.
Phys. Soc. (London) 89, 469–470.
[376] Seaton, M.J. (1973). Close coupling, Comput. Phys. Commun. 6, 247–256.
[377] Seaton, M.J. (1983). Quantum defect theory, Rep. Prog. Phys. 46, 167–257.
[378] Seitz, L.I. (1955). Quantum Mechanics, 2nd edition (McGraw-Hill, New York).
[379] Shanno, D.F. (1970). Conditioning of quasi-Newton methods for function
minimisation, Math. Comput. 24, 647–656.
[380] Sharp, R.T. and Horton, G.K. (1953). A variational approach to the unipotential
many-electron problem, Phys. Rev. 90, 317.
[381] Shimamura, I. (1977). In The Physics of Electronic and Atomic Collisions, invited
papers, X ICPEAC, Paris (1977), (North-Holland, Amsterdam).
[382] Siegert, A.J.F. (1939). On the derivation of the dispersion formula for nuclear
reactions, Phys. Rev. 56, 750–752.
[383] Sinfailam, A.L. and Nesbet, R.K. (1972). Variational calculations of
electron–helium scattering, Phys. Rev. A 6, 2118–2125.
[384] Skriver, H.L. (1984). The LMTO Method (Springer-Verlag, Berlin).
[385] Slater, J.C. (1951). A simplification of the Hartree–Fock method, Phys. Rev. 81,
385–390.
[386] Slater, J.C. (1972). Statistical exchange-correlation in the self-consistent field,
Adv. Quantum Chem. 6, 1–92.
[387] Slater, J.C. (1974). The Self-Consistent Field for Molecules and Solids
(McGraw-Hill, New York).
[388] Slater, J.C. and Wood, J.H. (1971). Statistical exchange and the total energy of a
crystal, Int. J. Quantum Chem. Suppl. 4, 3–34.
[389] Smith, F.T. (1960). Lifetime matrix in collision theory, Phys. Rev. 118, 349–356.
[390] Sundermeyer, K. (1983). Constrained Dynamics (Springer-Verlag, Berlin).
[391] Swanson, M.S. (1992). Path Integrals and Quantum Processes (Academic Press,
New York).
[392] Symanzik, K. (1954). Schwinger’s functional in field theory, Z. Natürforschung 9a,
809–824.
[393] Synge, J.L. (1956). Relativity: the Special Theory (Interscience, New York).
[394] Szabo, A. and Ostlund, N. S. (1982). Modern Quantum Chemistry; Introduction to
Advanced Electronic Structure Theory (McGraw-Hill, New York).
[395] Takahashi, Y. (1957). On the generalized Ward identity, Nuovo Cimento 6, 371–375.
[396] Takatsuka, K. and McKoy, V. (1981). Extension of the Schwinger variational
principle beyond the static exchange approximation, Phys. Rev. A 24, 2473–2480.
[397] Takatsuka, K., Lucchese, R.R. and McKoy, V. (1981). Relationship between the
Schwinger and Kohn-type variational principles in scattering theory, Phys. Rev. A
24, 1812–1816.
[398] Talman, J.D. and Shadwick, W.F. (1976). Optimized effective atomic central
potential, Phys. Rev. A 14, 36–40.
[399] Taylor, J.T. (1983). Scattering Theory (Krieger, FL).
[400] Teller, E. (1962). On the stability of molecules in the Thomas–Fermi theory, Rev.
Mod. Phys. 34, 627–631.
222 References and bibliography

[401] Temkin, A. (1976). Some recent developments in electron–molecule scattering:

theory and calculation, Comments At. Mol. Phys. 5, 129–139.
[402] Thomas, L.H. (1927). The calculation of atomic fields, Proc. Camb. Phil. Soc. 23,
542–548.
[403] Thompson, C.J. (1972). Mathematical Statistical Mechanics (Princeton University
Press, Princeton, New Jersey).
[404] ’t Hooft, G. (1971). Renormalization of massless Yang–Mills fields, Nucl. Phys.
B33, 173–199.
[405] ’t Hooft, G. (1971). Renormalizable Lagrangians for massive Yang–Mills fields,
Nucl. Phys. B35, 167–188.
[406] ’t Hooft, G. and Veltman, M. (1972). Regularization and renormalization of gauge
fields, Nucl. Phys. B44, 189–213.
[407] Thouless, D.J. (1961). The Quantum Mechanics of Many-body Systems (Academic
Press, New York).
[408] Thümmel, H.T., Grimm-Bosbach, T., Nesbet, R.K. and Peyerimhoff, S.D. (1995).
Rovibrational excitation by electron impact, in Computational Methods for
Electron–Molecule Collisions, eds. W.M. Huo and F. Gianturco (Plenum,
New York), pp. 265–291.
[409] Thümmel, H.T., Peyerimhoff, S.D. and Nesbet, R.K. (1992). Near-threshold
rotational excitation in electron–polar molecule scattering, J. Phys. B 25,
4553–4579.
[410] Thümmel, H.T., Peyerimhoff, S.D. and Nesbet, R.K. (1993). Near-threshold
rovibrational excitation of HF by electron impact, J. Phys. B 26, 1223–1251.
[411] Trail, W.K., Morrison, M.A., Isaacs, W.A. and Saha, B.C. (1990). Simple
procedures for including vibrational effects in the calculation of electron–molecule
cross sections, Phys. Rev. A 41, 4868–4878.
[412] Trickey, S.B. (1984). Comment on ‘Electron removal energies in Kohn–Sham
density-functional theory’, Phys. Rev. B 30, 3523.
[413] Volterra, V. (1954). Opere Matematiche, 5 vols. (Accademia dei Lincei, Rome,
1954–1962).
[414] von Barth, U. and Hedin L. (1972). A local exchange-correlation potential for the
spin-polarized case: I, J. Phys. C 5, 1629–1642.
[415] Wadehra, J.M. (1984). Dissociative attachment to rovibrationally excited H2 , Phys.
Rev. A 29, 106–110.
[416] Wadehra, J.M. and Bardsley, J.N. (1978). Vibrational and rotational state
dependence of dissociative attachment in e−H2 collisions, Phys. Rev. Lett. 41,
1795–1798.
[417] Ward, J.C. (1950). The scattering of light by light, Phys. Rev. 77, 293.
[418] Watson, D.K. (1988). Schwinger variational methods, Adv. At. Mol. Phys. 25,
221–250.
[419] Watson, D.K. and McKoy, V. (1979). Discrete basis function approach to
electron–molecule scattering, Phys. Rev. A 20, 1474–1483.
[420] Weinberg, S. (1967). A model of leptons, Phys. Rev. Lett. 19, 1264–1266.
[421] Weinberg, S. (1995). The Quantum Theory of Fields, Vol. I (Cambridge University
Press, New York).
[422] Weinberg, S. (1996). The Quantum Theory of Fields, Vol. II (Cambridge University
Press, New York).
[423] Weiner, B. and Trickey, S.B. (1999). Time-dependent variational principle in
density functional theory, Adv. Quantum Chem. 35, 217–247.
[424] Weyl, H. (1929). Z. Phys. 56, 330.
[425] Weyl, H. (1951). Space, Time, Matter (Dover, New York).
 References and bibliography 223

[426] Wheeler, J.A. and Feynman, R.P. (1945). Interaction with the absorber as the
mechanism of radiation, Rev. Mod. Phys. 17, 157–179.
[427] Wigner, E.P. (1948). On the behavior of cross sections near thresholds, Phys. Rev.
73, 1002–1009.
[428] Wigner, E.P. and Eisenbud, L. (1947). Higher angular momenta and long range
interaction in resonance reactions, Phys. Rev. 72, 29–41.
[429] Wigner, E.P. and Seitz, F. (1933). On the constitution of metallic sodium, Phys. Rev.
43, 804–810.
[430] Williams, A.R., Kübler, K. and Gelatt, C.D. (1979). Cohesive properties of metallic
compounds: Augmented-spherical-wave calculations, Phys. Rev. B 19,
6094–6118.
[431] Williams, A.R. and Morgan, J. van W. (1972). Multiple scattering by non-muffin-tin
potentials, J. Phys. C 5, L293–L298.
[432] Williams, A.R. and Morgan, J. van W. (1974). Multiple scattering by non-muffin-tin
potentials: general formulation, J. Phys. C 7, 37–60.
[433] Wolfenstein, L. (1999). Neutrino physics, Rev. Mod. Phys. 71, S140–S144.
[434] Yang, C.N. (1977). Magnetic monopoles, fiber bundles, and gauge fields, Ann. N.Y.
Acad. Sci. 294, 86–97.
[435] Yang, C.N. and Mills, R.L. (1954). Conservation of isotopic spin and isotopic gauge
invariance, Phys. Rev. 96, 191–195.
[436] Yourgrau, W. and Mandelstam, S. (1968). Variational Principles in Dynamics and
Quantum Theory, 3rd edition (Dover, New York).
[437] Zeller, R. (1987). Multiple-scattering solution of Schrödinger’s equation for
potentials of general shape, J. Phys. C 20, 2347–2360.
[438] Zeller, R. (1988). Empty-lattice test for non-muffin-tin multiple-scattering
equations, Phys. Rev. B 38, 5993–6002.
[439] Zerner, M.C. (1989). Analytic derivative methods and geometry optimization, in
Modern Quantum Chemistry, Introduction to Advanced Electronic Structure
Theory, eds. A. Szabo and N.S. Ostlund (McGraw-Hill, New York), pp. 437–458.
[440] Zhang, J.Z.H. and Miller, W.H. (1989). Quantum reactive scattering via the
S-matrix version of the Kohn variational principle: Differential and integral cross
sections for D + H2 → HD + H, J. Chem. Phys. 91, 1528–1547.
[441] Zhang, X.-G. and Butler, W.H. (1992). Simple cellular method for the exact
solution of the one-electron Schrödinger equation, Phys. Rev. Lett. 68, 3753–3756.
[442] Zhang, X.-G. and Butler, W.H. (1992). Multiple-scattering theory with a truncated
basis set, Phys. Rev. B 46, 7433–7447.
[443] Zhang, X.-G., Butler, W.H., MacLaren, J.M. and van Ek, J. (1994). Cellular
solutions for the Poisson equation in extended systems, Phys. Rev. B 49,
13383–13393.
[444] Zhang, X.-G., Butler, W.H., Nicholson, D.M. and Nesbet, R.K. (1992).
Green-function cellular method for the electronic structure of molecules and solids,
Phys. Rev. B 46, 15031–15039.
[445] Zhao, Q. and Parr, R.G. (1993). Constrained-search method to determine wave
functions from electronic densities, J. Chem. Phys. 98, 543–548.
[446] Zhao, Q., Morrison, R.C. and Parr, R.G. (1994). From electron densities to
Kohn–Sham kinetic energies, orbital energies, exchange-correlation potentials, and
exchange-correlation energies, Phys. Rev. A 50, 2138–2142.
[447] Zumbach, G. and Maschke, K. (1983). New approach to the calculation of density
functionals, Phys. Rev. A 28, 544–554.
[448] Zvijac, D.J., Heller, E.J. and Light, J.C. (1975). Variational correction to Wigner
R-matrix theory of scattering, J. Phys. B 8, 1016–1033.
 Index

action integrals chemical potential, 73

Hamilton, 12 classical electrodynamics
Lagrangian, 12 energy–momentum
time-dependent theories, reconciliation, 85 covariant derivation, 199
atomic cell dissipative term, 198
atomic-cell orbitals, ACO, 102 local conservation law, 198
canonical basis orbitals, 102 mass density, 199
Helmholtz Green function, 121 energy–momentum tensor
muffin-tin orbitals, MTO, 101 Noether’s theorem, 197
primitive basis orbital, 101 symmetric, 197
-operator over surface, 101 momentum 4-vector, 197
surface function, solid-harmonic expansion, classical mechanics
99 Lagrangian formalism, 11
surface matching theorem, 100 special relativity, 20
t-matrix, 98 classical renormalization, Dirac equation,
defined by S and C matrices, 99 185
Wronskian integral condition, 101 classical trajectory as geodesic, 20
atomic-cell model completeness
local angular momentum, 97 infinite coordinate range, 39
multiple scattering theory, 94 Schrödinger eigenfunctions, 36
multipole moments, surface-integral formalism, Sturm–Liouville theory, 38
120 configuration interaction, CI, 64
Voronoi lattice, Wigner–Seitz, 95 conservation of energy, 17
atomic-sphere approximation, 113 conservation of generalized momenta, 17
atomic units constrained-search, Levy construction, 68
Hartree, 35 correlation energy, 54
Rydberg, 95 exact, 59
natural definition, 55
Banach space, density functions, 72 correlation potential, nonlocal, exact, 59
basis set expansions, orbital basis, 41 correlation response, TOFT, 90
bordered determinant theorem, 42 Coulomb cusp condition
Born–Oppenheimer separation, 162 interelectronic, 48
brachistochrone, 7 Kato, 39
J. Bernoulli, 6 covariant derivative
Brenig’s theorem, 59 electrodynamics, 183
Brillouin’s theorem, 62 electroweak theory, 202
covariant field equations
calculus of variations, 6 Dirac equation, dressed electron, 185
see also variational calculus fermion field
chain rule, functional derivatives, 74 bare electron, 183
channel orbital, 130 mass term, 184
closed channel, 130 Maxwell field, classical, 183
open channel, 130 current density, isovector, total, 195

225
226 Index

density function, v-representable, 69 fermion mass operator, 193

density functional derivative gauge field transformation, 194
Fréchet, local potential, 71 Noether current density, 194
Gâteaux, linear operator, 74 proof of gauge covariance, 195, 196
density functional theory unitary transformation, 193
DFT, 68 gauge group, nonabelian
Kohn–Sham construction, 74 fermion transformation, 192
Kohn–Sham equations, 71 gauge field transformation, 192
locality hypothesis, 71 gauge invariance
orbital functional theory classical electrodynamics, 189
consistency conditions, 72 Dirac field, 190
Thomas–Fermi theory, 72 interaction determined, 191
density matrix, Dirac, equation of motion, 81 Maxwell field, 190
density operator nonabelian symmetries, 192
Dirac, 80 nonabelian theory, 182
idempotent, 80 time-dependent action, 78
dissipative forces in Hamilton’s principle, 12 generalized gradients, 27
geodesic problem, 7
electrodynamics, classical relativistic theory, 182 global matching function, 102, 110
electromagnetic 4-potential, 183 MST, 110
electromagnetic field, 183 Green functions
electron mass, classical self-interaction, 184 Feynman propagator, 184
electron scattering Helmholtz, 95
atom, 129 two-center expansion, 122
molecule, fixed nuclei, 129 Helmholtz equation, 121
electroweak theory multichannel, 144
EWT, 182, 202 perturbed wave function, 121
gauge group, 192 principal value, 96
weak isospin, 192 scattering continuum, 140
energy eigenvalues, orbital Schrödinger equation, scattering theory, asymptotic forms, 141
36 Schrödinger
energy functional complex energy, 124
N -electron, 46 relation to density matrix, 124
Thomas–Fermi theory, 72 two-center expansion, 123
equations of motion Schrödinger equation, 121
classical trajectory, 16 structural matrix, 125
Lagrange, 12 direct matching on interfaces, 126
relativistic Lagrangian, 23 Lippmann–Schwinger equation, 126
Euler’s equation, 6 Sturm–Liouville theory, 37
excitation energies translational symmetry, 127
TDHF, 82 Green-function cellular method
TOFT, 89 JVCM, 111
NVCM, 111
fermion creation and destruction operators, 79
fermion field Hamilton’s equations, 16
helicity, 182 Hamiltonian
massless, Pauli spinor, 182 Bloch-modified, 154, 157
Feshbach partitioning, 165 central force, 16
field theory Hamiltonian formalism, classical mechanics, 14
Lagrangian density, Lorentz invariant, 181 Hamiltonian matrix, N -electron, Slater determinant
quantized, Feynman theory, 181 basis, 47
forces of constraint, 14 Hamiltonian operator, Schrödinger, 36
functional derivative, orbital, 55 Hartree–Fock equation, time-dependent, 81
Hellmann–Feynman theorem
gauge field forces on nuclei, 43
electroweak theory, 202 generalized, 43
nonabelian self-interaction, 195 Hessian matrix, 28
SU(2) Hessian update
covariant derivative, 193 BFGS, 29
fermion field equation, 193 rank-m, 29
fermion Lagrangian, 193 Hilbert space, orbital wave functions, 36
 Index 227

Hohenberg–Kohn theorems, 68 N -electron, 51

orbital functional theory, 69 projected basis, 51
uniqueness, 69 two-electron integrals, 47
honeycomb lattice, 102 Minkowski coordinates, 20
hypervirial theorem, 43 muffin-tin model, 94, 106
total force on electrons, 44 multiconfiguration self-consistent field, MCSCF, 64
total linear momentum, 44 multiple scattering theory, 94
atomic cell
ignorable generalized coordinate, 17 consistency condition, 98
independent-electron models, 54 primitive basis, 98
algebraic Hartree–Fock, 64 atomic-sphere model, 95
density functional theory, 68 C- and S-matrices
Hartree, 54 surface integrals, 101
Hartree–Fock, 54 volume integrals, 98
local-density approximation, 74, 93 energy-linearized, 94
optimized effective potential, 65 energy-derivative basis, 114
orbital functional theory, 54 LACO, 115
restricted Hartree–Fock, 62 LMTO, 102, 113
UHF model of DFT, 75 full-potential, 94
unrestricted Hartree–Fock, 61 full-potential theory, 104
independent-electron theory fundamental equation, 98
time-dependent density functional, TDFT, 86 Green functions, 120
time-dependent Hartree–Fock, TDHF, 81 internal sums in alloy theory, 107
time-dependent orbital functional, TOFT, 83 Korringa–Kohn–Rostoker
closure of internal sums, 106
Janak’s theorem, 56 derivation, 105
orbital functional theory, 57 homogeneous term, 105
unrestricted Hartree–Fock, 61 internal sums, 106
KKR, 99, 104
kinetic energy, Schrödinger, 36 rectangular C-, S-matrices, 106
Kohn–Sham equations, 93 secular determinant, 99
Koopmans’ theorem, unrestricted Hartree–Fock, 61 linearized methods
LACO wave function, 115
Lagrange multiplier LMTO wave function, 115
Lagrangian to Hamiltonian, 15 variational theory, 116
linear constraints, 13 root-search, 94
N -electron energy, 46 space-filling cells, 104
orbital energy, 39
potential field in Levy construction, 68 Newton’s method, 28
Lagrangian Noether’s theorem, 17
linear systems, 25 classical energy, 17
relativistic, 21 classical field theory, 186
Lagrangian constraints, rolling hoop, 14 classical relativistic field, 186
Lagrangian density, relativistic, QED, 182 conserved covariant current density, 187
least action, energy constraint condition, 9 gauge symmetry, 191
Legendre transformation, 14 fermion current density, 191
Lagrangian to Hamiltonian, 15 integral form, 187
thermodynamics, 15 linear momentum, 18
linear response theory, TOFT, 84 SU(2) theory, conserved current, 196
Lippmann–Schwinger equation, 95 Noether’s theorem, examples
Green functions, 121 Dirac angular momentum, 189
multichannel, 144 energy–momentum 4-vector, 188
local exchange potential orbital angular momentum, 189
Slater formula, 72 Schrödinger field, 188
UHF tests of locality, 75 nonlinear optimization, 26
local potential function, kinetic energy, 72
Lorentz transformation, 20 Occam’s razor, 4
occupation numbers
matrix elements fractional, 55
symmetry-adapted Landau theory, 56
example: spin eigenstates, 51 one-electron density matrix, natural orbitals, 65
228 Index

optimization multichannel projection, 131

molecular geometry, 30 multichannel R-matrix, variational formula, 151
BERNY algorithm, 31 multipole polarization potential, 159
GDIIS algorithm, 31 closure approximation, 160
quasi-Newton methods, 30 normalization of channel orbitals, 158
orbital Euler–Lagrange equations, 58 numerical asymptotic functions, NAF, 153
density functional theory, 71 phase shift, 133
MCSCF, 65 polarization response, 132
time-dependent, 83 multipole, 157
orbital functionals R-matrix
correlation, 58 derivative discontinuity, 148
Coulomb and exchange, 57 Green function formalism, 147
external potential, 41 long-range potentials, 148
kinetic energy orbital functional theory, 157
Schrödinger, 40 radii r0 and r1 , 153
orbital-functional response theory, exchange-only variational formalism, 149
limit, 85 Wigner–Eisenbud, 147
orbital wave functions, 36 R-operator, generalized geometry, 156
S-matrix, 132
Poisson bracket, classical, 17 short-range correlation, Coulomb cusp, 159
Poisson equation, 94 single-channel R-matrix, variational formula, 150
MST methodology, 118 static exchange, SE, 134
surface-integral formalism, 118 static exchange with polarization, SEP, 134
polarization pseudostate, 131 stationary R-operator, 156
T -matrix, 132
quantum chromodynamics, QCD, 182, 203 target states, 134
quantum electrodynamics, QED, 181 threshold effects, 162
quasiparticles, Landau, 56 total cross section, 133
transfer invariance, 145
reactance matrix, 137 scattering theory, resonances
reference state, 46, 54 decay width, 163
Brueckner–Brenig, 55 Feshbach formalism, 166
Hartree–Fock, 54 transition matrix, 165
Kohn–Sham, 71, 74 nonlocal Feshbach formalism, 166
quantum field theory, 79 phase matrix, separation of resonance, 173
response kernel resonance defined, 162
correlation, 85 resonance projection, rationale, 172
Hartree and exchange, 84 threshold structures, 163
TOFT, 84 time-delay analysis, 173
virtual state, 162
scattering channel scattering theory, rovibronic
closed, 130 adiabatic approximations
open, 130 adiabatic nuclei, 167
scattering theory energy-modified, 168
α-matrix, 132, 135 nuclear kinetic energy, 166
algebraic close coupling, 134 relation to Born–Oppenheimer, 169
asymptotic distorted wave, 153 vibronic R-matrix, 169
bound–free correlation, 158 dissociative attachment, 162, 163
multipole response, 159 electron–molecule
bound–free exchange, 158 boomerang model, 163
close coupling, 133 local complex-potential, 163
Lagrange multipliers, 145 projection-operator formalism, 164
cross sections, 132 frame transformation, rotational, 167
differential cross section, 133 NADP magic formula, 176
eigenchannel representation, 144 nonresonant phase matrix, EMAP method, 174
electron–molecule, 161 phase matrix
Feshbach projectors, 131 fixed nuclei, 172
K-matrix, 132 precursor resonance, 173
incremental, 144 vibrational completeness, 173
Lippmann–Schwinger, 140 precursor resonance, 176
matrix variational method, 135 resonant phase matrix, NADP method, 175
 Index 229

rovibrational excitation, 161 variational cellular method, VCM, 108, 109

vibrational close-coupling, completeness, 168 variational functional
vibrational excitation, 162 matrix variational method, 137
hybrid model, 167 Schlosser–Marcus, 108
vibronic hypercylinder, 169 Schwinger, 142
vibronic R-matrix variational principles
EMAP method, 174 algebraic Schwinger, 142
Green function formalism, 170 Hamilton, 12
NADP method, 176 space-time action integral, 181
pole states, 171 special relativity, 20
Schrödinger equation in classical mechanics, 8
idiosyncrasies, 38 least action
matrix representation, 42 Euler, 8
N -electron Jacobi, 19
time-dependent, 45 Maupertuis, 8
time-independent, 46 least time, Fermat, 4
one electron, 36 multichannel R-matrix, 152
positive energy, Helmholtz equation, 95 multichannel Schwinger, 140, 143
time-dependent, 78 N -electron, orbital basis, 46
self-consistent field, 53 optimized effective potential, 67
self-interaction correction, Hartree, 54 orbital functional theory, 58
simplex interpolation, 26 R-operator, 154
Slater determinant, 46 radial logarithmic derivative, 149
solid-harmonic functions, definition, 96 Rayleigh quotient, 41
Standard Model of elementary particles, 201 scattering theory, 150
structure constants, 97 anomalies, 138
Sturm–Liouville theory, 36 complex Kohn, 135, 139
SU(2) theory Hulthén–Kohn, 135, 137
energy–momentum Jackson, 150
dissipative term, 201 Kohn, 134, 138
local conservation law, 201 multichannel R-matrix, 151
mass density, 201 Schwinger, 140
energy–momentum tensor Schwinger Green function, 147
Noether’s theorem, 199 Schrödinger, 41
symmetric, 200 N -electron, 46
momentum 4-vector, 200 Schrödinger action integral, 78
symmetry, projection operators, turnover rule, 49 stationary, 6
symmetry-adapted functions, N -electron problem, 49 TDFT, 87
TOFT, 83
tail cancellation, 102 variational principles, MST
atomic cell, 103 Kohn–Rostoker, 104
target state, 129 φ̇ basis, 116
Thomas–Fermi equation, 73 modified Schlosser–Marcus, 110
Kohn–Sham equations, consistency test, 73 φ̇ basis, 117
time-dependent density functional global matching function, 111
interacting electrons, 89 Schlosser–Marcus, 108
noninteracting electrons, 88 φ̇ basis, 116
TDFT, 86 derivation, 108
time-dependent Hartree–Fock false solutions, 109, 111
linear response, 82 SM and KR, conditional equivalence, 108
random phase approximation, 82 virial theorem, general molecule, 45
screening response, 81
Ward identities, quantum field theory, 186
variation, infinitesimal, 5 wave functions, irregular, transformation group, 123
variational calculus, 5 wave mechanics, Schrödinger, 35
functional analysis, 7 Wronskian surface integrals, 96
fundamental theorem, 7 normalization of solid harmonics, 96