Lecture notes on Electrodynamics–118120

J.E. Avron1

May 5, 2013

1
Comments and typos welcome. Send to [email protected].
2

I thank my TAs Dr. Dana Levanony and Mr. Yaroslav Pollak for all their
help and Prof. Amos Ori and Dr. Oded Kenneth for all they taught me. The
students the class in 2012, especially Itai Schlesinger, who pruned many typos
and Daniel Klein who showed me how to query Wolfarm α.
Contents

1 Tensor calculus 9
1.1 Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.1 Euclidean geometry . . . . . . . . . . . . . . . . . . . . . 9
1.1.2 The metric tensor . . . . . . . . . . . . . . . . . . . . . . 10
1.1.3 Einstein summation convention . . . . . . . . . . . . . . . 11
1.1.4 Coordinate transformations . . . . . . . . . . . . . . . . . 11
1.2 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.1 Covariants components . . . . . . . . . . . . . . . . . . . 14
1.2.2 Orthogonal coordinates . . . . . . . . . . . . . . . . . . . 15
1.2.3 Contraction makes scalars . . . . . . . . . . . . . . . . . . 16
1.3 Scalars, vectors, tensors, densities . . . . . . . . . . . . . . . . . . 16
1.3.1 Symmetric and anti-symmetric tensors . . . . . . . . . . . 16
1.3.2 Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3.3 Levi-Civita symbol . . . . . . . . . . . . . . . . . . . . . . 17
1.3.4 Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.3.5 Areas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4 Mirror . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.5 Isometries of Euclidean space . . . . . . . . . . . . . . . . . . . . 19
1.6 Tensorial equations are coordinate free . . . . . . . . . . . . . . . 20
1.7 Differential operators . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.7.1 Grad . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.7.2 Div . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.7.3 Curl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.7.4 Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.8 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2 Minkowski space-time 25
2.1 The principle of relativity . . . . . . . . . . . . . . . . . . . . . . 25
2.1.1 Causality . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 Space-time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.1 Cartesian coordinates and Inertial frames . . . . . . . . . 28
2.2.2 Events, world lines, light cones, etc. . . . . . . . . . . . . 29
2.3 Everything is relative . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3.1 Time dilation . . . . . . . . . . . . . . . . . . . . . . . . . 30

3
4 CONTENTS

2.3.2 Length contraction . . . . . . . . . . . . . . . . . . . . . . 31

2.3.3 Other coordinates . . . . . . . . . . . . . . . . . . . . . . 33
2.4 Lorentz transformations . . . . . . . . . . . . . . . . . . . . . . . 35
2.4.1 Space-time translations . . . . . . . . . . . . . . . . . . . 36
2.4.2 Generators of Lorentz transformations . . . . . . . . . . . 36
2.4.3 Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4.4 Boosts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.4.5 Rapidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.5 4-Velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.5.1 4-momentum . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.5.2 4 Acceleration: . . . . . . . . . . . . . . . . . . . . . . . . 40
2.5.3 Horizons . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.6 GPS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.6.1 Two dimensional space time . . . . . . . . . . . . . . . . . 42
2.6.2 3+1 dimensions . . . . . . . . . . . . . . . . . . . . . . . . 43

3 The electromagnetic tensor 45

3.1 Amalgamating E and B . . . . . . . . . . . . . . . . . . . . . . . 45
3.1.1 Euclidean Rotations . . . . . . . . . . . . . . . . . . . . . 47
3.1.2 The overall of sign . . . . . . . . . . . . . . . . . . . . . . 48
3.1.3 Covariant formulation . . . . . . . . . . . . . . . . . . . . 49
3.2 The electromagnetic potential . . . . . . . . . . . . . . . . . . . . 50
3.2.1 Covarinat formulation . . . . . . . . . . . . . . . . . . . . 50
3.2.2 Gauge freedom . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2.3 Non-local gauge invariants Lorentz scalars . . . . . . . . . 51
3.2.4 What does a voltmeter measure? . . . . . . . . . . . . . . 52
3.2.5 Generalizations . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3 Lorentz scalars . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.1 Local scalars . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.2 Dual: F ∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.3 Second Lorentz scalar: E · B . . . . . . . . . . . . . . . . 55
3.4 The homogeneous Maxwell equations . . . . . . . . . . . . . . . . 55
3.4.1 Maxwell form . . . . . . . . . . . . . . . . . . . . . . . . . 56

4 Variational principle 57
4.1 The action begins . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.1.1 Action for a free particle . . . . . . . . . . . . . . . . . . . 58
4.1.2 Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.1.3 Gauge invariance . . . . . . . . . . . . . . . . . . . . . . . 59
4.2 Variation of the action . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2.1 Variation of Sp . . . . . . . . . . . . . . . . . . . . . . . . 60
4.2.2 Variation of Sint . . . . . . . . . . . . . . . . . . . . . . . 61
4.2.3 Euler-Lagrange equation . . . . . . . . . . . . . . . . . . . 61
4.2.4 The non-relativistic limit . . . . . . . . . . . . . . . . . . 61
4.2.5 Existence and uniqueness . . . . . . . . . . . . . . . . . . 62
4.2.6 Consistency check . . . . . . . . . . . . . . . . . . . . . . 63
CONTENTS 5

4.3 Supplement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.3.1 Fermat principle . . . . . . . . . . . . . . . . . . . . . . . 63
4.3.2 Rainbow . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.3.3 Geodesics in Curved space-time . . . . . . . . . . . . . . 65

5 Maxwell Equations 69
5.1 Technical preliminaries . . . . . . . . . . . . . . . . . . . . . . . . 69
5.1.1 Space-time volume element . . . . . . . . . . . . . . . . . 69
5.1.2 Densities . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.1.3 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.2 Charge densities and currents . . . . . . . . . . . . . . . . . . . . 72
5.2.1 4-current-density . . . . . . . . . . . . . . . . . . . . . . . 72
5.2.2 Charge conservation . . . . . . . . . . . . . . . . . . . . . 73
5.2.3 Current conservation and gauge invariance . . . . . . . . 74
5.2.4 Gauge invariance and the continuity equation . . . . . . . 75
5.3 Lagrangian field theory . . . . . . . . . . . . . . . . . . . . . . . 75
5.3.1 The Lagrangian of the electromagnetic field . . . . . . . . 76
5.4 Variation of the field: Rules of the game . . . . . . . . . . . . . . 77
5.4.1 Variation of the field: Calculations . . . . . . . . . . . . . 77
5.4.2 Variation of the interaction . . . . . . . . . . . . . . . . . 78
5.4.3 The inhomogeneous Maxwell equations . . . . . . . . . . 78
5.4.4 Current conservation . . . . . . . . . . . . . . . . . . . . . 79
5.4.5 3-D form . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.4.6 Time reversal . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.4.7 Maxwell equations: Structure . . . . . . . . . . . . . . . . 80
5.5 Dielectric and magnetic media . . . . . . . . . . . . . . . . . . . 81
5.6 New Physics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.6.1 The quantum Hall effect and Chern-Simon action . . . . . 83
5.6.2 Axion electrodynamics . . . . . . . . . . . . . . . . . . . 85
5.6.3 Quantum interface . . . . . . . . . . . . . . . . . . . . . . 86
5.6.4 Magnetic response to an electric field . . . . . . . . . . . . 88
5.6.5 Phantom monopoles . . . . . . . . . . . . . . . . . . . . . 88
5.6.6 Electrodynamics in 1+1 dimensions . . . . . . . . . . . . 90

6 Conservation laws and the Stress-Energy tensor 93

6.1 Maxwell stress energy tensor . . . . . . . . . . . . . . . . . . . . 93
6.1.1 Symmetric and traceless . . . . . . . . . . . . . . . . . . . 94
6.1.2 Conservation laws . . . . . . . . . . . . . . . . . . . . . . 94
6.1.3 Nöther currents . . . . . . . . . . . . . . . . . . . . . . . . 95
6.1.4 Shifting the field . . . . . . . . . . . . . . . . . . . . . . . 95
6.1.5 Shifting the box . . . . . . . . . . . . . . . . . . . . . . . 96
6.1.6 Joint box and field shift . . . . . . . . . . . . . . . . . . . 97
6.1.7 The momentum of the field . . . . . . . . . . . . . . . . . 97
6.1.8 The energy of a field . . . . . . . . . . . . . . . . . . . . . 98
6.1.9 Angular momentum . . . . . . . . . . . . . . . . . . . . . 98
6.2 T and variation of the metric . . . . . . . . . . . . . . . . . . . . 99
6 CONTENTS

6.2.1 Blah . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.2.2 Variation of the metric in mechanics . . . . . . . . . . . . 100
6.2.3 Variation of the metric . . . . . . . . . . . . . . . . . . . . 101
6.2.4 Matrix calculus . . . . . . . . . . . . . . . . . . . . . . . . 101
6.2.5 The stress tensor . . . . . . . . . . . . . . . . . . . . . . . 102
6.3 Energy momentum conservation . . . . . . . . . . . . . . . . . . 102
6.3.1 The source term j · F and conservation of energy . . . . . 103
6.3.2 The interpretation of T jk : Stress . . . . . . . . . . . . . . 104
6.3.3 Field lines as rubber bands . . . . . . . . . . . . . . . . . 105
6.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.4.1 Radiation pressure . . . . . . . . . . . . . . . . . . . . . . 106
6.4.2 Solar sails . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.4.3 Halbach array . . . . . . . . . . . . . . . . . . . . . . . . . 108

7 Poisson equation, Cloaking 109

7.1 Vector fields in 3D: Source and vorticity . . . . . . . . . . . . . . 109
7.1.1 Static electric fields: Ḃ = 0 . . . . . . . . . . . . . . . . . 110
7.1.2 Harmonic functions and Poisson’s equation . . . . . . . . 110
7.1.3 The case of two dimensions . . . . . . . . . . . . . . . . . 111
7.1.4 The Green function of the Laplacian . . . . . . . . . . . . 111
7.1.5 Proof of theorem 9.32 . . . . . . . . . . . . . . . . . . . . 112
7.1.6 Uniformly moving charge . . . . . . . . . . . . . . . . . . 113
7.2 Coulomb gauge . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
7.3 Magnetic fields in the case Ė = 0 . . . . . . . . . . . . . . . . . . 115
7.3.1 Biot-Savart law . . . . . . . . . . . . . . . . . . . . . . . . 116
7.3.2 Magnetic dipole . . . . . . . . . . . . . . . . . . . . . . . 117
7.3.3 Dirac monopole . . . . . . . . . . . . . . . . . . . . . . . . 119
7.3.4 Application to geometry: Linking number . . . . . . . . . 120
7.4 Cloaking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.4.1 Maxwell equation in a dielectric medium . . . . . . . . . . 121
7.4.2 Maxwell in curvilinear coordinates . . . . . . . . . . . . . 122
7.4.3 Metric and dielectric medium . . . . . . . . . . . . . . . . 122
7.4.4 Dirichlet to Neuman: Calderon problem . . . . . . . . . . 123
7.4.5 An invisible dielectric ball . . . . . . . . . . . . . . . . . . 123
7.4.6 Cloaking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

8 Electromagnetic waves 127

8.1 Electromagnetic waves . . . . . . . . . . . . . . . . . . . . . . . . 127
8.1.1 Electric and Magnetic fields . . . . . . . . . . . . . . . . . 128
8.2 Plane waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
8.2.1 Electric and magnetic fields . . . . . . . . . . . . . . . . . 129
8.2.2 Doppler . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
8.2.3 Particle production: GZK limit . . . . . . . . . . . . . . . 130
8.2.4 Laser cooling and optical molasses . . . . . . . . . . . . . 131
8.3 Polarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
8.3.1 Amplitude and phase . . . . . . . . . . . . . . . . . . . . 132
CONTENTS 7

8.3.2 Polarization . . . . . . . . . . . . . . . . . . . . . . . . . . 132

8.3.3 Poincare sphere . . . . . . . . . . . . . . . . . . . . . . . . 133
8.3.4 Stokes parameters . . . . . . . . . . . . . . . . . . . . . . 133
8.3.5 Partially polarized light . . . . . . . . . . . . . . . . . . . 134
8.3.6 3D glasses . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8.4 The wave equation . . . . . . . . . . . . . . . . . . . . . . . . . . 136
8.4.1 The wave equation in one dimension . . . . . . . . . . . . 136
8.4.2 Huygens principle . . . . . . . . . . . . . . . . . . . . . . 138
8.4.3 Covariant superposition . . . . . . . . . . . . . . . . . . . 138
8.4.4 Waves with Gaussian waists . . . . . . . . . . . . . . . . . 139
8.4.5 Monochromatic waves . . . . . . . . . . . . . . . . . . . . 139
8.4.6 Evanescent waves . . . . . . . . . . . . . . . . . . . . . . . 140
8.4.7 The eikonal equation: Geometric optics . . . . . . . . . . 140
8.5 Waves in dielectric media: Birefringence: . . . . . . . . . . . . . 140
8.6 Green’s function for the wave equation . . . . . . . . . . . . . . . 141

9 Radiation 147
9.1 Wave equation with a source term . . . . . . . . . . . . . . . . . 147
9.1.1 Retarded waves for an arbitrary source . . . . . . . . . . . 147
9.1.2 Application: Lorenz gauge . . . . . . . . . . . . . . . . . . 147
9.1.3 Scalar wave generated by a moving point source . . . . . 148
9.2 Maxwell equation in the Lorenz gauge . . . . . . . . . . . . . . . 149
9.3 Lienard-Wiechert: Retarded potentials . . . . . . . . . . . . . . 150
9.3.1 The Lorenz Gauge condition . . . . . . . . . . . . . . . . 150
9.4 Retarded Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
9.4.1 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . 153
9.5 Particle instantaneously at rest . . . . . . . . . . . . . . . . . . . 154
9.5.1 The Magnetic field: . . . . . . . . . . . . . . . . . . . . . 154
9.5.2 The electric field . . . . . . . . . . . . . . . . . . . . . . . 155
9.5.3 Slow particles . . . . . . . . . . . . . . . . . . . . . . . . . 156
9.6 Retardation from a distant source . . . . . . . . . . . . . . . . . 156
9.6.1 The dipole approximation: k`  1 . . . . . . . . . . . . . 157
9.6.2 Beyond dipole: Slow motion . . . . . . . . . . . . . . . . . 159
9.7 Power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
9.8 Classical instability of atoms . . . . . . . . . . . . . . . . . . . . 160
9.9 Orienting an antenna . . . . . . . . . . . . . . . . . . . . . . . . . 161
9.10 Reciprocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

10 Radiation reaction 163

10.1 Interacting systems . . . . . . . . . . . . . . . . . . . . . . . . . . 163
10.1.1 Non-relativistic interacting particles . . . . . . . . . . . . 163
10.1.2 Plasma and Magnetohydrodynamics . . . . . . . . . . . . 164
10.1.3 Infinities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
10.2 Radiation reaction . . . . . . . . . . . . . . . . . . . . . . . . . . 164
10.2.1 When is radiation reaction important? . . . . . . . . . . . 165
10.2.2 Friction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
8 CONTENTS

10.2.3 The Dumbbell . . . . . . . . . . . . . . . . . . . . . . . . 167

10.2.4 Conceptual difficulties . . . . . . . . . . . . . . . . . . . . 169
May 5, 2013
Chapter 1

Tensor calculus

Tensor calculus allows us to write equation without committing ourselves to a

specific coordinate systems.

1.1 Geometry
Minkowski space-time is the stage on which electrodynamics takes place. Before
describing the geometry of space-time lets us collect the tools we need from
(Riemanian) geometry.

1.1.1 Euclidean geometry

In constructing physical theories we need to know some facts about nature. For
example, we need to know various physical constants such as e, ~, c mass of
particles etc. They are all god given scalars. Besides knowledge about scalars,
we also need to know something about the space we live in. For our purposes,
physical space is to a good approximation, Euclidean, and we take this to be
another God given fact. We are still free to choose any (curvilinear in general)
coordinate systems to describe the space.

Remark 1.1 (Euclid, Gauss and Einstein). Euclid took it for granted that
physical space is Euclidean. The first to seriously entertain the possibility that
the physical space need not be Euclidean was Gauss. In a Euclidean world the
angles of all triangles sum up to π. So, when one says that the world is to a good
approximation Euclidean one means that the deviations from π are small. Gauss
who had experience in lad surveying made an experiment which was did not show
deviation from Euclidean geometry. Later Einstein taught us that space-time is
actually curved and there are many physical tests of this. However, this is a
another story.

9
10 CHAPTER 1. TENSOR CALCULUS

1.1.2 The metric tensor

1 2
In the Euclidean plane consider the three

coordinate systems: (x , x ) a Carte-
0 1 0 2
sian coordinate system; (x ) , (x ) ) a rotated (blue)

Cartesian coordinate
system and a polar coordinate systems (x”)1 , (x”)2 = (r, θ). d` is the dis-

Figure 1.1: Cartesian, rotated Cartesian and polar coordinates for the Euclidean
plane

tance measured between two neighborhood points using a standard meter. It

does not depend on the choice of coordinates
X
(d`)2 = gij dxi dxj
ij

g is called the metric tensor, also known as the Riemann metric tensor who in
his Thesis founded Riemanian geometry. It is a second rank tensor which means
it has two indices. It is also symmetric gij = gji . The components are written
downstairs. Downstairs components are called covariant components.
If the two Cartesian coordinate systems use the same yardstick then we may
choose the length scale of xj so that
gij = δij
With this choice g is dimensionless and dxj have dimension of length.
In polar coordinates, x1 = r cos θ, x2 = r sin θ
(d`)2 = (dr)2 + r2 (dθ)2
Exercise 1.2. Verify.
The moral of this is that there are many distinct metrics for a given space: As
many as coordinate transformations. The surface of the sphere is geometrically
distinct from the plane. It has the metric
(d`)2 = (dθ)2 + sin2 θ (dφ)2
The geometric distinction means that it can not accommodate Cartesian coor-
dinates. This is why maps drawn on a sheet of paper never accurately represent
regions earth.
Exercise 1.3. Compute det g in Cartesian and polar coordinates.
1.1. GEOMETRY 11

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0.2 0.4 0.6 0.8 1.0

Figure 1.2: The plot on the left shows a Cartesian y-mesh. The plot in the
middle shows a curvilinear y-mesh in the plane. The deformation is due to a
deformation of the radial polar coordinate r → r(tanh r) shown on the right.

1.1.3 Einstein summation convention

This is a short hand which says: Sum over pairs of up-down indices. For example
X
gij dxi dxj = gij dxi dxj

It is also called contraction of indices.

Remark 1.4 (Dummy indices). Summation indices are sometimes called run-
ning and sometimes dummy. They can be relabeled freely

gja v a = gjb v b

Remark 1.5 (Warning). If you get an equation where the indices are not nicely
paired, such as
va ua , va ua wa
it is a good idea to search for a typo.

1.1.4 Coordinate transformations

If g is the metric tensor in the coordinate x and x0 is different coordinate system
of the same space, then g 0 is

(d`)2 = gij dxi dxj

∂xi ∂xj
   
0 a
= gij d(x ) d(x0 )b
∂(x0 )a ∂(x0 )b
= (g 0 )ab d(x0 )a d(x0 )b

This says
∂xi
 
(g 0 )ab = Λi a Λj b gij , = Λj b gji Λi a , Λi a = (1.1)
∂(x0 )a
1
If one thinks of Λ and g as matrices the relation above can be written as

g 0 = Λt gΛ (1.2)
1 First index is row second is column.
12 CHAPTER 1. TENSOR CALCULUS

Remark 1.6. Since g is a symmetric matrix, it can be diagonalized by an

orthogonal transformation. So, at any given, fixed point x there is a Λ that
diagonalizes g. You can also rescale the coordinates so as to make gij (x) = δij
at one point x). But, in general, you can not make g the identity everywhere.
If you can, the space is the Euclidean space.
This is an expression of the fact that any (Riemannian) manifold is locally
Euclidean. Columbus needed to go far to convince everybody that Earth is a
sphere.
Exercise 1.7. By counting the number of free parameters and the number
of constraints in the Taylor expansions, show that you can choose coordinate
transformations that make g the identity and makes all its first derivative van-
ish. However, you can not make all the second derivatives vanish by coordinate
transformations.
Exercise 1.8. Write the metric of the unit sphere using the coordinates z of
the spherical projection on the plane.

1.2 Vectors
Consider a vector δx (say, in the Euclidean plane) associated with a small change
of the coordinates δxj . For the sake of masochism, we allow non-orthogonal
coordinate system.

e2

v
v2

v1
e1

Figure 1.3: Contravariant components and the basis vectors

δx = δx1 e1 + δx2 e2 (1.3)

ej are vectors pointing along the coordinate lines and δxj are the coordinates
increments. In general ej may depend on position as is the case for of polar
coordinates.
(δx1 , δx2 ) are the components of the vector. They are called the contravari-
ant components. Like the coordinate, have their indices upstairs. The length of
1.2. VECTORS 13

the vector is a scalar, and this aloows us to relate the basis vectors ej with the
metric X
δx · δx = gij δxi δxj , gij = ei · ej (1.4)
We recover that fact that g is symmetric and positive

gij = gji , g≥0

You also see form this formula that g is diagonal in orthogonal coordinate sys-
tems and that the covariant basis vectors ej are, in general, not unit vectors.

Exercise 1.9. Show that in a polar coordinate system the covariant basis vectors
and the normalized unit vectors are related by

er = r̂, eθ = rθ̂

Remark 1.10. Normalized unit vectors are defined only for orthogonal coordi-
nate systems.

In two coordinate systems we have

δx = δx1 e1 + δx2 e2 = δ(x0 )1 e0 1 + δ(x0 )2 e0 2 (1.5)

so the transformation law for the contravariant components is, by the chain rule

∂xj
 
j
δx = δ(x0 )a ⇐⇒ δxj = Λj a δ(x0 )a ⇐⇒ δx = Λδ(x0 ) (1.6)
∂(x0 )a | {z }
Contravariant

In the third expression think of the contravariant components δx as column

vector and Λ as a matrix. If you compare this with the rule for the metric
tensor you see that the prime is on the other side. Covariant and contravariant
indices have different transformation laws.

Example 1.11 (Mechanical model). A particle of unit mass moves on a ring

of fixed radius, r. Its orbit in polar coordinates is (r, θ(t)). Its velocity vector is:

v = (rθ̇) θ̂ = θ̇ eθ (1.7)

The “normalized” component is the tangential velocity, the contravariant com-

ponent is the angular velocity and its dimension is frequency, not velocity. This
is one disadvantage of contravariant components: They may screw dimensions.
They do have compensating advantages, however.

Exercise 1.12. Show that the matrices

∂xj ∂(x0 )b
   
0 b
Λj a = , (Λ ) k =
∂(x0 )a ∂xk

are inverses ΛΛ0 = 1

14 CHAPTER 1. TENSOR CALCULUS

1.2.1 Covariants components

The basis vectors ej are neither orthogonal nor normalized. In such cases one
defines the dual basis vectors ej to the basis ej by
ei · ej = δi j (1.8)
and δ is the Kronecker symbol. We could then write a vector can be represented
in two different ways
v = v a ea = va ea (1.9)
va are the covariant components and v j the contravariant. Clearly
vj = v · ej (1.10)
The covariant components are geometrically interpreted as dropping perpendic-
ulars on the coordinate mesh.
2
e
v2
v

v1

e1

Figure 1.4: Covariant components and the dual basis

From the definitions of the metric tensor, and the notion of duality we get
vk = va ea · ek = v a ea · ek = gka v a (1.11)
We learn from this that the metric tensor allows us to push indexes down.
Exercise 1.13. Show that
ej = (ej · ea ) ea
Everything one can do with contravariant components has an analog in the
covariant components. In particular, the length of a vector is evidently given
by
v · v = va vb g ab , g jk = ej · ek (1.12)
Taking the scalar product of exercise 1.13 with ek we conclude that the two
metric tensors are inversely related
δkj |{z}
= ej · ek |{z}
= (ej · ea )(ea · ek ) = g ja gak =
|{z} gj k
duality ex.10 index gym
1.2. VECTORS 15

The metric tensor with indexes up is the inverse matrix of the metric tensor
with indexes down. It raises indexes since

g ja va = g ja gab v b = δbj v b = v j (1.13)

From this one can figure the transformation rules for the covariant components

(v 0 )a = (g 0 )ab (v 0 )b
= (g 0 )ab (Λ−1 v)b
= Λi a Λj b gij (Λ−1 )b k v k
= Λi a Λj b (Λ−1 )b k gij v k
= Λi a (ΛΛ−1 )j k gij v k
= Λi a gij v j
= Λi a vi

Remark 1.14. If you think of the covariant components as a row vector then
the transfromation rule is
0
|v ={zvΛ}
covarinat

Note the similarities and differences with the rule of transfromation of the
contravariant components.

1.2.2 Orthogonal coordinates

Many of the standard curvilinear coordinate systems one encounters in practice
are orthogonal. In this case, the metric g is a diagonal matrix. Orthogonal
coordinates admit three types of components: The usual covariant and con-
travariant components and the “normalized” components. All three are given
by
V = v j ej = vj ej = vĵ nj

with ei · ej = gij = gi δij , ei · ej = g ij = (gi )−1 δij and nj · nj = δij .

Example 1.15 (Polar coordinates). In polar coordinates

v= v r er + v θ eθ = v er + v eθ = vr̂ r̂ + vθ̂ θ̂ (1.14)

| {z } | r {z θ } | {z }
contravariant components covariant components normalized

The local frames have basis vectors:

er · er = 1, eθ · eθ = r2 , er · eθ = 0
r r θ θ −2
e · e = 1, e ·e =r , er · eθ = 0
r̂ · r̂ = 1, θ̂ · θ̂ = 1, r̂ · θ̂ = 0
16 CHAPTER 1. TENSOR CALCULUS

1.2.3 Contraction makes scalars

Corollary 1.16. The contraction of a covariant index with a contravariant
index gives a (coordinate independent) scalar

u · v = u0 · v0

One easy way to see this is

0
u
| ={zuΛ} , v = Λv0 ⇔ v 0
Λ−1 v}
| ={z
covariant row contravar column

from which the assertion readily follows.

1.3 Scalars, vectors, tensors, densities

The charge of a particle, its mass, or the length of a vector is a scalar. You do not
need to decide onn coordinates to measure scalars. If you do use coordinates, the
result should be independent of the choice of coordinates. This is the defining
property of scalars.
Vectors are geometric objects and as such do not rely of a coordinate sys-
tem either. But, their representation by covariant or contravariant components
depend on the choice of coordinate system. The defining property is that their
components have a single index and the rule of transformation is:

v j = Λj k (v 0 )k , (v 0 )j = vk Λk j (1.15)

Tensors are multi-index objects and the metric tensor is an example. The num-
ber of indices is called the rank of the tensor. Each index transforms according
to whether it is up or down. For example, the second rank tensor T j k transform
like v j uk .

1.3.1 Symmetric and anti-symmetric tensors

Coordinate transformations respect the symmetry of tensors: If T is symmetric
(anti-symmetric), i.e. Tjk = ±Tkj , so is (T 0 ).
Exercise 1.17. Show that if Tjk is symmetric (anti-smmetric) so is T jk but in
general Tj k 6= Tk j . (One finds instead Tj k = T k j ).

1.3.2 Weights
There are interesting physical quantities that are neither scalars nor tensors.
det g is and example. From Eq. 1.2

2
det g 0 = det Λ det g

Objects with such a rule of transformation are called weights. det g has weight
−2.
1.3. SCALARS, VECTORS, TENSORS, DENSITIES 17

1.3.3 Levi-Civita symbol

Suppose you are in two dimensions. The highest rank anti-symmetric tensor
has rank two and it has one non trivial component which gives the sign of the
permutation
ε11 = ε22 = 0, |{z}
ε12 = −ε21 = 1
| {z }
even odd

It is a tensor under coordinate transformations provided det Λ = 1. This follows

from:
ε12 = Λ1 a Λ2 b (ε0 )ab = det Λ (ε0 )12

In general, when det Λ 6= 1, it is a tensor density, i.e. has a weight.

The same idea works in any dimension: In three dimensions the highest
rank of completely anti-symmetric tensor is three and it has one interesting
component
123 231
|ε = ε{z = ε312} = −ε321 = −ε213 = −ε132 = 1
| {z }
even permutations odd permutations

It is a tensor under cordinate transformations provided det Λ = 1. This follows

from
ε123 = Λ1 a Λ2 b Λ3 c (ε0 )abc = det Λ (e0 )123

and so, in general, is a tensor density.

Exercise 1.18. Show that

εa... √
√ , gεa...
g
are both bona-fide tensors.

Remark 1.19. The conventions for spherical coordinates is such that you get
a right handed frame provided or order the coordinates r, φ, θ, i.e.

εrφθ = 1

Remark 1.20. In even dimensions cyclic permutations are odd, while in odd
dimensions cyclic permutations are even.

Exercise 1.21. Show that in n dimensions

εij... εij... = n!

Exercise 1.22. Show that (in 3 dimensions)

εijk εiab = δaj δbk − δbj δak , εijk εijb = 2δbk

18 CHAPTER 1. TENSOR CALCULUS

1.3.4 Volume
In a 3-dimensional Euclidean space, with Cartesian coordinates xj the volume
element is
dV = dx1 dx2 dx3
and, of course, dx1 dx1 dx2 has no volume. If we also allow volume to take a
sign according to the handedness of the frame, we naturally associate volume
element with the components of the completely antisymmetric tensor:

dV = dx1 ∧ dx2 ∧ dx3

and ∧ means that order matters: dxi ∧ dxj = −dxj ∧ dxi . In Cartesian coordi-
√
nates, where g = 1, this can be re-written as
√
√ 1 2 3 g
dV = g dx ∧ dx ∧ dx = εijk dxi ∧ dxj ∧ dxk (1.16)
3!
√
We have seen in Ex.1.18 gεijk are the covariant components of a bona-fide
(completely anti-symmetric third rank) tensor. Similarly, dxi ∧ dxj ∧ dxk are
the conravariant components of a completely anti-symmetric third rank tensor.
Contracting the two we get that dV behaves like a scalar under coordinate
transformation. Hence, Eq. 1.16 holds in any curvilinear coordinate system,
(provided 1, 2, 3 is right handed frame). This works in any dimension.
Remark 1.23. It is useful to assign signs to volume (and areas): Positive for
right handed frames and negative for left handed frames.
Exercise 1.24 (Spherical coordinates). Let (x, y, z) be Cartesian coordinates
off Eucliden space with metric d`2 = (dx)2 + (dy)2 + (dz)2 . Show that the usual
spherical coordinates

z = r cos θ, x = r sin θ cos φ, y = r sin θ sin φ

have the metric tensor

d`2 = (dr)2 + r2 (dθ)2 + r2 sin2 θ (dφ)2

and that the volume element is

p
dV = det g dr dθ dφ = r2 sin θ dr dθ dφ = −r2 dr d cos θ dφ

1.3.5 Areas
The area element dS3 associated with the parallelogram (dx1 , dx2 ) is naturally
defined as
dV = dS3 dx3 = (dS3 e3 ) ·(dx3 e3 )
| {z }
area vector

Exercise 1.25. Verify the formula for the area for spherical area elements dSr .
1.4. MIRROR 19

Figure 1.5: The area element, black arrow, directed along e3 , is defined so that
its scalar product with the red arrow, e3 , gives the volume. The black arrow is
perpendicular to the blue arrows e1 and e2 .

1.4 Mirror
A mirror flips some things and does not flip others. It does not flip up and down
but does flip right and left, it flips a right handed coordinate system to a left
handed one etc.
The length of an object is a scalar. It does not flip in a mirror. However,
when one integrates it makes a sign difference if you go right or left. So, for
example, you want the volume element to flip a sign when reflected in a mirror.
For the sake of concreteness consider the 3-dimensional Euclidean space.
Under coordinate inversion (x0 )j = −xj the contravariant components of say
a position or velocity vector flip sign. (The vector still points in the same
direction.) Now consider the angular velocity. since you use your right hand
to determine the direction of the vector of angular velocity, in a mirror you’ll
use your left hand. The vector of angular velocity now points in the opposite
direction. It is a pseudo vector. Under inversion its components do not flip sign.
The cross product of two vectors is therefore a pseudo-vector. Since
(a × b)i = εijk aj bk
we conclude that the Levi-Civita symbol is also pseudo.

1.5 Isometries of Euclidean space

Euclidean space looks the same no matter where you are or how you are ori-
ented: It is homogeneous and isotropic. These symmetries reflect invariance
properties of the metric tensor of Euclidean space under suitable coordinate
transformations. In Cartesian coordinates xj a shift is:
(x0 )j = xj + aj =⇒ Λ = 1 =⇒ g 0 = g = 1,
It leaves g invariant. This reflects the homogeneity of Euclidean space. Because
of that, the components of vectors are invariant under translations: (v 0 )j = v j .
20 CHAPTER 1. TENSOR CALCULUS

This is what is meant by saying that vectors in Cartesian do not have a location.
Rotation keeps the origin fixed. This is also the case for general linear
transformation,
(x0 )j = Λj a xa (1.17)
A linear transformation is a symmetry provided it leave the metric invariant.
In Cartesian coordinates g = 1 and by Eq. (1.2)

1 = (g0 ) = Λt gΛ = Λt 1Λ = Λt Λ
This says that Λt is the inverse of Λ:

Λt Λ = 1

which is the standard definition of orthogonal transformation in Cartesian co-

ordinates. It follows that

det Λ2 = 1 =⇒ det Λ = ±1

Orthogonal transformations are associated with two types of symmetries of the

Euclidean space: When det Λ = 1 they represent rotations. When det Λ = −1
they represent mirror symmetries.
Example 1.26 (Rotations). In three dimensions, rotation by θ about the x3
axis is given by  
cos θ sin θ 0
R(θ) =  − sin θ cos θ 0 
0 0 1
Evidently, the inverse is the transpose

R(−θ) = Rt (θ) (1.18)

Example 1.27. In a two dimensional Euclidean space there are two second
rank tensors that are invariant under rotations
   
1 0 0 1
, (1.19)
0 1 −1 0

1.6 Tensorial equations are coordinate free

The nice thing about tensor equations is that once

T jk... = 0

holds in one (fixed) coordinate system, it hold in any other coordinate system.
For example, Newton’s equation

f j = maj

is a tensor equation relating force and acceleration (m is a scalar). If it holds

in one coordinate system it hold in any other.
1.7. DIFFERENTIAL OPERATORS 21

1.7 Differential operators

The equations of motions of fields are partial differential equations. This forces
us to mind how differential operators behave under change of coordinates. The
general case is complicated and one needs to introduce the notion of covaraint
derivatives. To see where the complication comes from consider the covaraint
components of a vector field Aj in Euclidean space. Then, in the exercise
1.28 you are requested to show that ∂j Ak transforms like a second rank tensor
provided x0 is a linear function of x.
Exercise 1.28. Show that under a change to general curvilinear coordinates

∂j Ak = Λa j Λb k (∂a0 A0b ) + ∂jk

2
(x0 )a A0a

A simplification however occurs for certain differential operators, and in

particular for the three differential operators we need for Maxwell’s equations:
grad, div and curl. The following exercise shows where this simplification comes
from
Exercise 1.29. Show that the anti-symmetric second rank tensor

Fjk = ∂j Ak − ∂k Aj

transforms like a tensor under general curvilinear coordinate transformations.

1.7.1 Grad
The chain rule
∂ ∂xk ∂ ∂
0
= = Λk j k
∂(x ) j ∂(x0 )j ∂xk ∂x
says that partial derivatives behave like covariant components of a vector. In
particular, if φ(x) is a scalar valued function then ∇φ give the components of a
covariant vector field.
Exercise 1.30. Show that ∇φ in spherical coordinates is

∇φ = (∂r φ) er + (∂θ φ) eθ + (∂ϕ φ) eϕ

∂θ φ ∂ϕ φ
= (∂r φ) r̂ + θ̂ + ϕ̂
r r sin θ

Here you see why covariant components often lead to simpler formulas than
normalized coordinates.

1.7.2 Div
We shall show that
1 √
∇ · E = √ ∂j ( gE j ),
g
22 CHAPTER 1. TENSOR CALCULUS

The formulas clearly hold in Cartesian coordinates where g is the identity.

The defining property of the divergence is Gauss law
Z Z
dV (∇ · E) = dS · E
V olume surf ace

Consider a smal cube in the coordinates dxj . The putative expression for div
indeed satisfies Gauss law

√
dV (∇ · E) = gdx1 dx2 dx3 (∇ · E)
√ f
= dx2 dx3 g E 1  + . . .

i
3√ 1
f
2
= dx dx ge · E 1 e1  + . . .

i
√ f
= g dx2 dx3 e1 ·E + . . .

| {z } i
dS
= dS · E

Example 1.31. In spherical coordinates div is

1
∂r (r2 sin θE r ) + ∂θ (r2 sin θE θ ) + ∂φ (r2 sin θE φ )


∇·E =
r2
sin θ
1 1
= 2 ∂r (r2 E r ) + ∂θ (sin θE θ ) + ∂φ (E φ )
r sin θ
1 1  
= 2 ∂r (r2 Er̂ ) + ∂θ (sin θEθ̂ ) + ∂φ (Eφ̂ )
r r sin θ

The last line is in terms of the normalized coordinates.

1.7.3 Curl
The last differential operator we shall need to discuss is the curl:

εijk
(∇ × E)i = √ ∂j Ek
g

The formula is evidently the standard definition in Cartesian coordinates. To

see why it is true in general we take Stokes law as defining property of the curl:
Z Z
dS · (∇ × E) = d` · E
1.7. DIFFERENTIAL OPERATORS 23

The putative formula for curl indeed gives Stokes for dS a small square dx1 ×dx2 .

dS · (∇ × E) = dS3 (∇ × E)3

ε3ij
 
√
= gdx1 dx2 √ ∂i Ej
g
= dx1 dx2 (∂1 E2 − ∂2 E1 )
f
= dx2 E2  − . . .

i
f
= (dx e2 ) · (E2 e2 ) − . . .
2 
i
f
= (dx2 e2 ) · E − . . .

i
= d` · E
2
Example 1.32. The φ components of curl in spherical coordinates (recall
Remark1.19 ) is :

1
(∇ × E)ϕ = (∂θ Er − ∂r Eθ )
r2 sin θ
and in normalized components
1
(∇ × E)ϕ̂ = (∂θ Er − ∂r Eθ )
r
1

= ∂θ Er̂ − ∂r (rEθ̂ )
r
Exercise 1.33. Compute the (∇ × E)r and (∇ × E)r̂ .

Exercise 1.34 (Vector identities). Show the vector identities

∇ × (∇φ) = 0
∇ · (∇ × E) = 0

1.7.4 Laplacian
The Laplacian of a scalar function is defined by
1 √ 1 √
∆φ = ∇ · ∇φ = √ ∂j ( gg jk ∂k φ) = √ ∂j ( g ∂ j φ)
g g

and for a vector field by

∇ × (∇ × E) = −∆E + ∇(∇ · E)
2 Note that Wolfram Mathematica notation compares with mine by interchanging ϕ ↔ θ
24 CHAPTER 1. TENSOR CALCULUS

Example 1.35. ∆φ in spherical coordinates:

    
1 2 2 ∂θ φ 2 1
∆φ = 2 ∂r (r sin θ∂r φ) + ∂θ r sin θ 2 + ∂ϕ r sin θ 2 2 ∂ϕ φ
r sin θ r r sin θ
1 2 1 1
= 2 ∂r (r ∂r φ) + 2 ∂θ (sin θ ∂θ φ) + 2 2 ∂ϕϕ φ
r r sin θ r sin θ

1.8 Bibliography
• S. Weinberg, Gravitation and Cosmology, gives all a physicist needs to
know about tensors.

• B. Schutz
• Flanders
Chapter 2

Minkowski space-time

Minkowski space-time gives a geometric description of special relativity. It

encapsulates the facts: The velocity of light is the same in all inertial frames;
Every inertial describes a homogeneous space-time; Every inertial frame frame
admits a fixed time slice which is a Euclidean 3 dimensional space; and clocks
can be synchronized in any fixed inertial frame (but disagree between frames).

2.1 The principle of relativity

Physicists unlike, say, lawyers, do not need to replace their textbooks when they
relocate. Your physics library would still be useful even if you relocated to a
different galaxy, receding from earth at large speed. You do not need to make an
adjustment for the relative motion between earth and your new home (provided
space-time is locally the same). Empty space, the vacuum, has no distinguished
inertial frame; none distinguished as being at rest and none whose origin is the
center of the universe.
The speed of light c is a property of the vacuum. It is a scalar, a constant of
nature, which takes the same value in different inertial frames. This property of
light was established around 1887 in experiments of Michelson and Morley. It
conflicts with our common intuition about adding velocities much smaller than
c.

Remark 2.1 (c is large). Human length scale, say the length of the foot, is
` = 1[meter].
p When you walk the foot behaves like a pendulum, and its period
is 2π `/g ≈ 2 [s] with g ≈ 9.8 [m/s2 ]. A human speed is then ≈ 1 [m/s].
On this scale c ≈ 3 × 108 [m/s] is essentially infinite. I do not know who first
entertained the thought that c may be finite but large, but the first to estimate c
from astronomical data was the Danish astronomer Rømer (1644-1710) .

25
26 CHAPTER 2. MINKOWSKI SPACE-TIME

2.1.1 Causality
Causality is a way to distinguish past from future: We remember the past but
can’t change it and we can affect the future but do not know it. When one takes
into account the fact that c is a constant of nature, one finds that one needs
to reconsider what one means by the notion of “future” , “now” and “past”.
Einstein said that if you could break the speed of light, you could use this to
send a message to your dead grandmother. An alternative characterization of c
is therefore the ultimate speed at which information propagates.

2.2 Space-time
Space-time is the stage on which events happen. An event, like my typing this
text, is something that happens in space and time and is labeled by 4-coordinates
xµ , µ ∈ 0, 1, 2, 3 where x0 = ct with t time. It is natural to give space and time
the same dimension. This is what we do when we say that the sun is about 8
light-minutes away from earth.

Remark 2.2. We shall use the conventions that Greek indices µ, ν run from 0
to 3, while Roman indices j, k from 1 to 3.

t t'

Future x'

Past

Figure 2.1: Space-time: The red disk is an event. The t axis is the world line
of a black clock that sits at the origin x1 = x2 = x3 = 0. The x axis represents
a Euclidean 3-space. The blue line is the world line of a blue clock moving at
constant speed c/2. In the frame of the blue clock t0 is the time axis. The red
lines represent the future and past light cones relative associated with a signal
that is emitted (or absorbed) at the origin. The Euclidean distance on the paper
reflects badly (d`)2 in Minkowski space.
2.2. SPACE-TIME 27

The “distance squared” between two nearby events in space-time is called

interval and is given by
(ds)2 = gµν dxµ dxν

with g a real, symmetric tensor. The notation is taken from geometry where
(ds)2 ≥ 0 but is potentially misleading because we allow (ds)2 to have either
sign: When the two events are simultaneous in an inertial frame the interval is
indeed the distance squared and (ds)2 > 0. When two events are separated in
time, but occur at the same point the interval is the time elapsed on a clock
squared, but (ds)2 < 0. Let us motivate this choice of signs.
We can always choose coordinates so that g is diagonal at a given point.
(Take Λ an orthogonal transformation.) By scaling we can then make the diag-
onal entries ±1. If all the entries are +1 we have a something that looks locally
like 4-dimensional Euclidean space, not space-time. Space-time has three spa-
tial coordinates, naturally associated with the three entries +1. The remaining
time coordinate is different. It comes with the −1. This gives a 3+1 space-time
manifold1 .

Definition 1. A space-time is Minkowski if it admits coordinates so that g = η

everywhere with
 
−1 0 0 0
 0 1 0 0 
ηµν = 
 0 0
 (2.1)
1 0 
0 0 0 1

A time slice x0 = const gives a Euclidean 3-space: homogeneous and isotropic.

Minkowski space-time is a good local approximation of physical space-time. It
fails at cosmological distances and at distant times or when great accuracy,
such as in GPS, is needed. Asher Peres said that physics is not an exact science
but the science of approximation. A more accurate model of space-time in the
vicinity of star (planet) of mass M is

2GM
(ds)2 = − (1 − Φ) (cdt)2 + (1 + Φ) dx · dx, Φ(x) =
c2 r

G is Newton constant. (t, x) are space-time coordinates. A clock at a fixed

location ticks at rate
√
dτ = 1 + Φ dt

Far from the star, the clock rate coincides with the coordinate time rate.

Exercise 2.3. Compute Φ at the surface of the earth. (Answer: Φ = 1.4×10−9 )

2V
Exercise 2.4. Compute the correction terms to Minkowski, c2 on earth due
the sun. (Answer: Φ = 2 × 10−8 )
28 CHAPTER 2. MINKOWSKI SPACE-TIME

Figure 2.2: Two Cartesian coordinates in Minkowski space. In both the t axis
and the x axis are Minkowsi orthogonal: The light-cone (red) bisects the angles
between the t and x axes. If you have installed Mathematica cdf player you can
view a simulation courtesy of Slava Pollak.

2.2.1 Cartesian coordinates and Inertial frames

A given Euclidean space accommodates many Cartesian coordinate system. In
all of them the metric g = 1. Similarly, Minkowski space accommodates many
coordinate systems where g = η. A coordinate system where g = η is called
Cartesian. As we shall different inertial frames are associated with different
Cartesian coordinates.
In Cartesian coordinates the (unit) vector tangent to the wold line at the
origin, t = (1, 0, 0, 0), is Minkowski orthogonal to the space-like vector x =
(0, 1, 0, 0). Now consider the blue (unit) vector tangent to the world line of an
inertial observer moving at velocity |v| < c (in the x direction) of Fig. 2.1. Its
contravariant components (in the black frame) are

1
t0 = γ(1, v/c, 0, 0), γ=p ≥1
1 − (v/c)2

What is the direction of the x0 of the blue inertial frame? Since the blue frame
is inertial, x0 and t0 must be Minkowski orthogonal and the contravariant com-
ponents of x0 must be

x0 = γ(v/c, 1, 0, 0) =⇒ t0µ x0µ = γ 2 (v/c − v/c) = 0

Example 2.5. In spherical coordinates (ct, r, θ, φ) Minkoski metric is

 
−1 0 0 0
 0 1 0 0 
ηµν =  (2.2)
 0 0 r2 0 
2 2
0 0 0 r sin θ

1 This would also be the case with three −1 and one +1.
2.3. EVERYTHING IS RELATIVE 29

2.2.2 Events, world lines, light cones, etc.

An event is an objective happening and does not depend on the choice of coor-
dinates. However, its representation in terms of coordinates x = (x0 , x1 , x2 , x3 )
depends of the choice of coordinates.
A line in space-time is called a world line. For example, the collection of
events associated with an observer—a clock moving at subluminal speed—make
a world line.
The light cone is a three-dimensional (conic) surface in space-time. It de-
scribes the collection of events x related to a given event, x0 , by a light signal.
If we denote dx = x − x0 then dx is a vector with zero length
0 = dx · dx − c2 (dt)2 = ηµν dxµ dxν = dxµ dxµ
Light-cone is made of vectors whose Minkowski length is zero.
Definition 2. A vector dx is called space like if dxµ dxµ > 0; It is called time-
like if dxµ dxµ < 0 and light-like if dxµ dxµ = 0
If the vector connecting two events is time-like, there is a clock, moving at
subluminal speed, whose world line passes through the two events. The time
elapsed on this clock dτ measures the interval
(cdτ )2 = −(d`)2 = −dxµ dxµ (2.3)
Clearly it does not depend on the choice of coordinates. dτ is known as the
proper time.
Exercise 2.6 (Measuring space like intervals with a clock). To measure space-
like intervals you would normally use meter sticks. Wigner found a clever trick
to measure space like intervals using a single clock and two light signals con-
necting the clock with the space-like event. This is illustrated in Fig. 2.3. The
interval is the product of two time intervals a and b measured by an inertial
clock.
(xO − xS )µ (xO − xS )µ = ab

2.3 Everything is relative

The ct axis represents the world line of a clock at rest at the origin of space:
xj = 0. xj = 0 is ‘here’ for the clock and the world line is (ct, 0, 0, 0). A blue
clock is moving at constant velocity v whose world line is (ct, vt, 0, 0). This
world line defines ct0 axis in Fig. 2.4. It is the origin the blue inertial frame,
(x0 )j = 0. The notion of ‘here’ is different for the two clocks: Being at the same
place is not objective but frame ( and coordinate) dependent.

Now is relative
Fig. 2.4 shows that notion of now is relative. In the black coordinate system
events separated by x occur simultaneously, in the blue ones events separated
by x0 are simultaneous.
30 CHAPTER 2. MINKOWSKI SPACE-TIME

b
S

O
a

Figure 2.3: The interval between the two space like events O and S is related to
the clock readings of an inertial clock. The red lines in the figure are light-like
vectors and a and b are the time intervals measured by the clock at O.

2.3.1 Time dilation

When you meet your high school buddies after a long time you usually note
how everybody aged and wonder, did I too age that much. This is psychology,
and I have nothing to say about it. But there is an analog objective property
of time: your clock is slowest: . If you travel fast enough and far enough when
you come back, all your friends will be dead.
Consider the interval dτ between the two time ticks of the black clock. The
time registered on a single physical clock is called proper time. It is represented
by the (cyan) vector (cdτ ), 0) in Fig. 2.5. In the blue frame the clock is moving
at v. The cyan vector has blue coordinates

dx0 = −vdt0 =⇒ (c, −v)dt0

Since the interval is a scalar, dτ and dt0 are related

(c dt0 )2
(c dτ )2 = (c2 − v 2 )(dt0 )2 = ,
γ2
where v is the (instantaneous) velocity of the clock. Proper time is the smallest
time γ is the ratio between increments in coordinate-time dt and proper time
dτ :
dt
γ=
dτ
Exercise 2.7. Suppose that you have a factory at the origin that makes identical
clocks. Explain how you can distribute the clocks while keeping them synchro-
2.3. EVERYTHING IS RELATIVE 31

t t'

x'

Figure 2.4: The black coordinate system is inertial. So is the blue coordinate
system. In both the t and x axes are Minkowski orthogonal. The two red
dots are two events which are simultaneous in the black inertial frame but not
simultaneous in the blue frame. The light cone is a coordinate independent
entity.

nized. (Hint: What happens to time delay if you half the speed and double the
travel time?)

When Paul Krugman, A Nobel Laureate in economics, was a young assistant

professor he wrote funny article where he applies time-dilation to economics. A
pleasant diversion.

2.3.2 Length contraction

Consider a rod at rest in the black frame whose length there is `. In the blue
frame the rod is moving at velocity v. A measurement of the rod in the blue
frame must be made simultaneously at the two ends. The difference of the two
events is represented by the vector (0, dx0 ), the green interval in Fig. 2.6.The t0
axis direction in the black coordinates is (1, v/c). The green vector pointing in
the x0 direction has black coordinates

(v/c, 1)`

(since the length of the rod, measured by x1 , in the frame where it is stationary
is `). The interval, being a scalar, relates the proper length ` with the apparent
32 CHAPTER 2. MINKOWSKI SPACE-TIME

ct ct¢

x¢
1

Figure 2.5: The blue frame moves to the right at u and the black frame moves
to the left at −u. Because of the symmetry, the length on the paper along t
and t0 correctly reflect the interval. The thick cyan line is one second measured
by a black clock. Its end point is on the line t = 1. The line of events t0 = 1 in
the blue frame intersects the cyan line: The moving clock is slow– Time dilates.

ct ct¢

x¢
1

Figure 2.6: The thick blue interval is a meter stick at rest in the black frame.
The black frame moves left and the blue frames righht at equal speeds. This
makes the Euclidean lengths along the x and x0 axes proportional to the interval.
The blue thick interval is longer than the green interval: The meter stick has
shorter length in the blue frame. Moving objects contract.
2.3. EVERYTHING IS RELATIVE 33

length dx0
`2
(dx0 )2 =, γ≥1
γ2
You are biggest in your own rest frame.

2.3.3 Other coordinates

Rotating frames
The rotating earth is a non-inertial frame. What is the structure of space-time
in such a frame. Let Ω be the angular frequency and (ct0 , x0 ) be the coordinate
of an inertial fame and (ct, x) the coordinates in a rotating frame. In cylindrical
coordinates the inertial and rotating frames are related by
t0 = t, ρ0 = ρ, z 0 = z, φ0 = φ + Ωt
The Euclidena metrics are the related by
dx0 · dx0 = (dρ0 )2 + (dz 0 )2 + ρ02 (dφ0 )2
= (dρ)2 + (dz)2 + ρ2 (dφ + Ωdt)2
= dx · dx + 2ρ2 Ωdφdt + ρ2 Ω2 (dt)2
Consequently, the Minwoski metric in the inertial frame is
(−cdt0 )2 + dx0 · dx0 =
 
−(dt)2 c2 − Ω2 ρ2 + 2Ωρ2 dtdφ +dx · dx
| {z } | {z }
centrif ugal Sangac ef f ect

Exercise 2.8. Earth actually rotates pretty fast. To get and idea compute Ωρ/c
at the equator . (Answer: 1.5 × 10−6 )
In the case of earth the centrifugal correction can normally be neglected, but
the Sangac term is important and to a good approximation

− (cdt)2 + 2Ωρ2 dtdφ + dx · dx (2.4)

Exercise 2.9 (Coordinate times and clock times). 1. Compare the change in
coordinate time dt in the rotating earth frame with the proper-time dτ
measured by a clock at a fixed location in the rotating frame
2. Compare the change in coordinate time dt0 in the inertial frame with the
change in coordinate time dt in the rotating coordinates
3. A clock is taken for a one year trip around earth equator. Show that the
time lag relative to a clock that stayed is
2AΩ
∆τ = ± , A = πRe2
c2
34 CHAPTER 2. MINKOWSKI SPACE-TIME

t t'

Φ'
Φ

Figure 2.7: The coordinates in the inertial frame are drawn blue and the
coordinates in the rotating frame black. The light cone in red. The gray line is
the synch line in a local inertial frame.

where Re is the earth radius and the ± depends on whether the trip was
towards the east or towards the west.

4. Compute ∆τ . (Answer: 207 [ns])

5. Explain why the result implies that one can not synchronize clocks on earth.

6. Is it still true that uµ uµ = −c2 in a rotating frame? Yes)

7. What does get modifies is γ. Show that

γ −2 = 1 − (v/c)2 + 2Ωφ̇ρ2 /c2 − Ω2 ρ2 /c2

Light-cone coordinates
Light cone coordinates in Minkowski space are
√ √
2u = x − ct, 2v = x + ct, y = y, z=z

Exercise 2.10 (Metric in light cone coordinates). Show that the Minkowski
metric tensor in light-cone ordinates is
 
0 1 0 0
 1 0 0 0 
ηµν = η µν = 
 0 0 1 0 ,


0 0 0 1
2.4. LORENTZ TRANSFORMATIONS 35

u v

Figure 2.8: Light-cone coordinates. This is not a Cartesian frame since u·v = 1.

Rindler coordinates
Rindler coordinates are the analog of polar coordinate system in a two dimen-
sional Minkowski space time

x = ρ cosh τ, t = ρ sinh τ

Note that with ρ ≥ 0 the coordinates cover 1/4 of space time.

Exercise 2.11. Show that

(dx)2 − (dt)2 = (dρ)2 − ρ2 (dτ )2

2.4 Lorentz transformations

Definition 3 (Lorentz transformations are isometries). We call Λ a Lorentz
transformation if it leaves the Minkowski metric η invariant
0
ηµν = Λµ α Λν β ηαβ = ηµν

As an equation for matrices this reads

η 0 = Λt ηΛ = η (2.5)

In particular, for any Lorentz transformation

det Λ = ±1
36 CHAPTER 2. MINKOWSKI SPACE-TIME

Figure 2.9: The red lines shows the ρ Rindler coordinate.

2.4.1 Space-time translations

Space time translations are given by

(x0 )µ = xµ + aµ

This gives
∂(x0 )µ
Λµ ν = = δµν ⇐⇒ Λ = 1
∂xν
which is the trivial Lorentz transformations. This expresses the homogeneity of
Minkowski space time.

2.4.2 Generators of Lorentz transformations

Suppose L is a 4 × 4 real matrix so that

ηL + Lt η = 0 (2.6)

We can use L to generate a one parameter family of 4 × 4 matrices Λ(t) as a

solution of the differential equation

Λ̇(t) = LΛ(t), Λ(0) = 1

The Λ(t) genertaed in this way is a Lorentz transformation for any t. This
follows from the following exercises
2.4. LORENTZ TRANSFORMATIONS 37

Exercise 2.12. Use Eq. 2.6 and the differential equation to show that det Λ(t) =
1.

Exercise 2.13. Show that

η(t) = Λ(t)t ηΛ(t)

is t independent and hence η(t) = η(0) = η. (Hint: Differentiate)

Exercise 2.14. Show that:

1. if λ is an eigenvalue of Λ so is λ∗ .

2. If λ is an eigenvalue so is 1/λ (Hint: Use Λ−1 = ηΛt η to show that

det(Λ − λ) = det(Λ−1 − λ) )

As we shall now see, there are 6 (linearly independent) generators of Lorentz

transformations: 3 are associated with (Euclidean) rotations and 3 with boosts
that relate different inertial frames.

2.4.3 Rotations
Rotation by θ about the x-axis and its generator are given by
   
1 0 0 0 0 0 0 0
 0 1 0 0   0 0 0 0 
Λyz (θ) = 
  Lyz =
 
0 0 cos θ sin θ  0 0 0 1 
0 0 − sin θ cos θ 0 0 −1 0

Similarly for rotations about the x, y space axes. The isometry expressses the
isotropy of space in of Minkowski space-time.

Remark 2.15. An airplane has three rotation controls: Stick, for pitch, rudder
for yaw, and ailerons for roll. The three are lineraly independnet, but non-
linearly dependent: You can always generate one from the other two.

Exercise 2.16. Show that

[Lyz , Lzx ] = −Lxy ,

Exercise 2.17. Calculate the residual rotation of (pairwise-cancelling) rotations

Λyz (π/2)Λzx (π/2)Λyz (−π/2)Λzx (−π/2)

Show that this is a rotation about the (−1, 1, 1) axis.

38 CHAPTER 2. MINKOWSKI SPACE-TIME

2.4.4 Boosts
Lorentz transformations relating different inertial frames are called as boosts.
A boost in the x direction and its generator are
   
cosh φ sinh φ 0 0 0 1 0 0
 sinh φ cosh φ 0 0   1 0 0 0 
Λtx (φ) =   , Ltx =   (2.7)
 0 0 1 0   0 0 1 0 
0 0 0 1 0 0 0 1
If you remember your quantum mechanics, then an easy way to see that Eq. (2.6)
is to recall that for Pauli matrices σx σz + σz σx = 0. and so are isometries of
Minkowski space. The isometry expresses the fact that Minkowski space looks
the same in all inertial frames.
Exercise 2.18. Show that the commutator of two boosts is a rotation:

[Ltx , Lty ] = Lxy ,

What can yous conclude from that? (Thomas precession).

2.4.5 Rapidity
φ of Eq. (2.7) is called the rapidity. To related the rapidity to the relative
velocity between the frames consider the origin in the primed frame x0 = 0.
0 = (x0 )1 = Λ1 µ xµ = −(sinh φ) ct + (cosh φ) x
From this we conclude that
v = c tanhφ (2.8)

vc

Rapidity

Figure 2.10: The velocity as function of the rapidity

Exercise 2.19 (Rapidity add). Show that rapidities (in the same direction) add
Λ(φ1 )Λ(φ2 ) = Λ(φ1 + φ2 )
Exercise 2.20 (Galilean transformations). Show that for small rapidities Lorentz
boosts reduce to Galilean transformation:
t0 = t + O(c−2 ), x0 = x − vt + O(c−2 )
2.5. 4-VELOCITY 39

2.5 4-Velocity
The proper time dτ is a Lorentz scalar and is non-zero for a clock that travels
slower than light. In this case we can define the velocity as a 4-vector
dxµ
uµ = (2.9)
dτ
The length of u is always -c2 , essentially by definition,
dxµ dxµ
uµ uµ = = −c2 (2.10)
(dτ )2
The 4-velocity is a time-like vector. It lies in the forward light cone. It is related
to the usual velocity by
dxµ dxµ dτ uµ
(c, v) = = =
dt dτ dt γ
Remark 2.21 (Old fashioned velocities). Whereas the 4-velocity transforms like
any 4-vector under Lorentz transformations, the usual velocity has complicated
(bad) transformation properties. This is because both the numerator and the
denominator are components of a vector.
It is convenient to think of the path xµ (τ ) as parametrized by proper time.
Since the 4-velocity is normalized we can always write it as
u = c (cosh φ, n sinh φ), n·n=1
n is the direction, which may depend on τ and φ = φ(τ ). Evidently
γ = cosh φ
from which is follows that φ is the rapidity defined in the previous section as
v = c tanh φ.
Exercise 2.22. Show this

2.5.1 4-momentum
Define the 4-momentum
pµ = muµ = mγ(c, v) = (E/c, p)
The associated scalar is
pµ pµ = −(mc)2
It is a time-like vector for massive particles. This encapsulates the most famous
equation in physics, Einstein equation for the energy in the rest frame, E = mc2 .

Remark 2.23 (Tachyons). Tachyons are (fictitious) particles with whose mass
is (fittingly) imaginary. The 4-velocity is space like, uµ uµ = (mc)2 > 0. A
Tachyon can not be a point particle: there is a Lorentz frame where the particle
is spread on a line at fixed t. If Tachyons would interact with ordinary particles
(tardyons), you could go back in time and kill your grandfather.
40 CHAPTER 2. MINKOWSKI SPACE-TIME

Figure 2.11: The dispersion relation, energy as function of momentum, for a

massive free particle.

2.5.2 4 Acceleration:
The 4-acceleration can be similarly defined as
duµ
aµ = (2.11)
dτ
It is always Minkowski orthogonal to the velocity
uµ aµ = 0 (2.12)
(Since the Minkowski length of the velocity is −1). It follows that The 4-
acceleration is always a space like vector.
Example 2.24 (Constant acceleration). The 4-velocity along some fixed direc-
tion is
u = c (cosh φ, n sinh φ)
where φ(τ ) is a parametrization of the path and n is a fixed unit vector. The
4-acceleration is then
a = c φ̇(sinh φ, n cosh φ)
Evidently
aµ aµ = c2 φ̇2
Constant acceleration g corresponds to linear dependence of φ on the proper-time
τ
gτ
φ(τ ) =
c
The red lines in Fig. 2.9 show paths of constant acceleration.
Remark 2.25 (Coincidence). An amusing phenomenological coincidence is that
the year, the gravitational acceleration on earth g, and c are simply related:
g × year/c = 1.03
Exercise 2.26. What fraction of the velocity of light would you reach in this
case. Answer: tanh 1.03 = 0.77
Exercise 2.27 (Space travel). You may worry that since c is the ultimate speed
man, living for, say 80 years, can explore at most a smallish neighborhood of 80
light-years around earth. This is wrong. A space traveler who lives for n years,
in a space ship which accelerates at g will travel a distances, cosh n, measured
in light years. The visible universe has radius of about 1011 light years. So you
will get there is about 26 years.
2.6. GPS 41

Example 2.28 (Motion with fixed velocity but changing direction). The 4-
velocity of a particle moving with fixed velocity but changing direction is
uµ = c(cosh φ, n(τ ) sinh φ)
where φ is fixed. The acceleration is
aµ = c sinh φ(0, ṅ)
Exercise 2.29. Find the orbit xµ (τ ) describing a plane circular motion with
constant angular velocity.

2.5.3 Horizons

t t¢

S
x
O

Figure 2.12: An inertial observer that lives long enough sees all the events in
Minkowski space-time

Figure 2.13: An accelerating sees only half the events in Minkowski space-time.
He never sees the black dot on the left. The red line is his horizon.

2.6 GPS
Every time you use your GPS and find, with relief, that the GPS really knows
where you are, you are testing special and general relativity. It took a century
to turn Einstein’s conceptual revolution into a palm gadget.
42 CHAPTER 2. MINKOWSKI SPACE-TIME

GPS works like that: There are about 24 GPS satellites orbiting earth at
a radius of about 26, 000 [km] and period of about half a day. Their orbit are
known (and monitored) with great precision (few centimeters). On each satellite
there is an atomic clock that measures the proper time with great precision.
Each satellite radios the coordinates of the event of transmission: The b satellite
radios the event Xbµ (τb ) and the event of reception if xµ . Since the transmission
is by electromagnetic wave that propagate at c, the two events are light-like. To
determine the four unknown coordinates xµ of teh reception event you need 4
equation. You need to see 4 GPS satellites and record 4 transmission events all
light-like separated from you. This gives 4 equations with 4 unknowns.
The GPS system is sophisticated and involved: It takes into account special
and general relativity; atmospheric effects on the veldocity off light, and the fact
you insist on having your location in a non-inertial coordinate system attached
to a rotation earth. The point I want to make here is that relativistic corrections
are large: Ignoring relativity would degrade the the accuracy to about 10 km
and make GPS useless. Relativity is regularly and routinely tested. If you want
to know more about that, then the article of Neal Ashby in Living Reviews of
General Relativity is a good place to learn. Wikipedia is, as usual, quite good
as well.
Instead, I will consider a caricature of GPS to illustrate one idea. Namely,
how one can use 4-clocks, with known position, and the fact the speed of light
is a constant of nature, to determine the reception event in space time.

Exercise 2.30 (Orders of magnitudes).

• What is a typical velocity of GPS satellite? (Anser: 3.8 [km/s])

• Compute the difference between the coordinate time and the self-time of a
GPS clock after one day. ( ∆t ≈ πRv/c2 ≈ 3.6 × 10−6 [s])

• What is the resulting positioning inaccuracy?

2.6.1 Two dimensional space time

Consider a toy GPS problem in 1+1 dimensions, shown in the Fig. 2.14. If you
see 2 satellites in 1+1 dimensions, this means that one is on your right and the
other on your left, as in Fig. 2.14. (Otherwise one would eclipse the other.)
The light-cones intersect.

Exercise 2.31
(GPS in 1+1 dimensions). Two satellites with known orbits
, a0b (τ ), a1b (τ ) , b = 1, 2, emit signals at τa and τb respectively. Assuming that
aµ − bµ is space like, show that light-cone intersect at

2x1 = ±(a01 − a02 ) + a11 + a12 , 2x0 = (a01 + a02 ) ± (a11 − a12 )

(One solution is in the past and the other in the future.)

This toy model tests:

2.6. GPS 43

ct

a
b

Figure 2.14: The world line of the two satellites are the blue lines. The intersec-
tion of the light cones is the event whose coordinates we seek. Since the orbits
of the satellites are known, the events (a0 , a1 ) and (b0 , b1 ) are known given the
proper times τa and τb .

1. Space-time is approximately Minkowski

2. Electromagnetic waves propagate at c
3. The velocity of the satellite at the transmission event is irrelevant
4. The velocity of the lost tourist at the reception event is irrelevant

2.6.2 3+1 dimensions

We have four satellites with known orbits aµb (τ ), b ∈ 1, 2, 3, 4 as functions of
their proper time. The GPS of the lost tourist receives the data τb , b ∈ 1, 2, 3, 4
and reveals the 4 events ab µ (τb ). All these events lie on the backward light cone
of the event of reception xµ . This makes all 6 pairs ab (τb ) − ab0 (τb0 ) space-like.
(The satellites do not eclipse each other.) In a Minkowski space-time the event
xµ satisfy the 4 light-cone equations
ηµν (xµ − aµb )(xµ − aµb ) = 0, b ∈ 1, 2, 3, 4
This is a system of (multivariate) quadratic equations.
Exercise 2.32. Show that if the two events xµA and xµB both lie on the future
light cone of the origin (i.e. both are light-like) then xA − xB is space like (or
light like).
It can’t be solved analytically, but Mathematica solves it numerically without
a problem. If the data are physical, ab (τb ) − ab0 (τb0 ), are all space-like, its spits
two events xµ , one in the future and one in the past. The first is the physical
one. If the data are not physical, xµ will, in general be complex valued.
Remark 2.33. If some of the satellites are close to eclipsing, it will be dif-
ficult to solve the equations accurately. You’d better look for a fifth satellite.
Optimally, you’d want the “volume’ of the space time tetrahedron eµναβ (aµ1 −
β
xµ )(aν2 − xν )(aα α β
3 − x )(a4 − x ) to be large.
44 CHAPTER 2. MINKOWSKI SPACE-TIME

Bibliography
1. B. F. Schutz, A first course in general relativity, geometric exposition of
relativity.
2. Neal Ashby in Living Reviews of General Relativity
Chapter 3

The electromagnetic tensor

In space-time the electric and magnetic fields (E, B) are amalgamated into a
second rank, anti-symmetric, tensor, Fµν . Fµν is derived from the 4-potential
Aµ . This leads to the homogeneous Maxwell equations and a neat way to write
Coulomb and Lorentz laws.

3.1 Amalgamating E and B

The basic objects of mechanics, such as velocity and accelerations can be viewed
as the 3 dimensional shadows of 4-vectors in Minkowski space-time. In both
cases the spatial 3-vectors, gained a zero component which then gives nice be-
havior under Lorentz transformations.
What about the electromagnetic fields E and B? What are they shadows
of? From a Euclidean perspective E and B are different entities. This can be
seen from their definition via the Coulomb Lorentz force law in the (defunct)
c.g.s. (Gaussian) units which we shall use here:
v
f= eE + e × B (3.1)
|{z} | c {z }
Coulomb Lorentz

The electric field is that part of the force which is independent of the velocity
of the particle and the magnetic field is the part of the force is linear in the
velocity. Of course, the partition into electric and magnetic field is different in
different inertial frames. Observers in different inertial frames do not agree on
the partition. E and B mix under change of frame.

Remark 3.1 (SI). Eq. (3.1) appears to say the magnetic forces are a relativistic
correction to the electric forces. This is misleading because it depends on what
values we take for the wo fields. In Gaussian units the unit of electric field is
300 [V /cm] and the unit of magnetic field is Gauss. In SI units the unit of
electric field is much five orders of magnitudes smaller, 1 [V /m], and the unit

45
46 CHAPTER 3. THE ELECTROMAGNETIC TENSOR

of magnetic field is four orders of magnitude larger: T esla = 104 Gauss. In SI

units1 the Coulomb-Lorentz law is

f = e (E + v × B)

The force of electric field of 1 [volt/cm] and the magnetic force of 1 [Gauss]
have comparable magnitudes at velocities of 1000 [km/sec].

Exercise 3.2 (Galilei invariance). Galilean transformations are the non-relativistic

limit of Lorentz transformations. In Newtonian mechanics the force is Galilean
invariant: f = f 0 (Why). Show that the transformation rules for the fields under
Galinean transformations with relative velocity v is

1
E0 = E + v × B, B0 = B (3.2)
c

The mixing of E and B under the change of inertial frames suggests that
they come from a single entity in space-time. This entity is not a 4-vector since
we need 6 slots and a 4-vector has too few. It is not a general second rank
tensor, since it has too many components–16. It is not a symmetric second rank
tensor since this too has too many components–10. However an anti-symmetric
rank 2 tensor
Fµν = −Fνµ (3.3)

has just the right number of components–6.

Exercise 3.3 (Symmetry invariance). Show that symmetry is a tensorial invari-

ant, e.g. if Fµν is anti-symmetric so is (F 0 )µν = Λµ α Λν β Fαβ under arbitrary
change of coordinates (and Lorentz transformation in particular). As a conse-
quence if F is anti-symmetric in Cartesian coordinates it is also anti-symmetric
in spherical coordinates.

This leaves us with the problem of how to put the two Euclidean vectors
(E, B) in Fµν . We need to identify slots in F that behave like vectors. Assuming
Cartesian coordinates, write
 
0 x1 x2 x3
 −x1 0 y3 −y 2 
Fµν =  (3.4)
 −x2 −y 3 0 y1 
−x3 y2 −y1 0

We shall show that x = (x1 , x2 , x3 ) and y = (y 1 , y 2 , y 3 ) transform like Euclidean

3-vectors. (The reason for putting some indexes up and some down will become
apparent below). This will leave us with the problem of matching the pair (x, y)
with (E, B).
1 And also if one takes units where c = 1.
3.1. AMALGAMATING E AND B 47

3.1.1 Euclidean Rotations

Consider a 3×3 rotation matrix R of Euclidean space so R−1 = Rt . In Euclidean
space contravariant components are the same as covariant components. It will
be convenient to write the indexes as one of each
column row
z}|{ z}|{
R j k = (Rt ) k
j
|{z} |{z}
row column

The lift of R to a rotation in Minkowski space is

 
1 0 0 0
 0 
ΛR = 
 0
, (3.5)
R 
0

(The identity on the right expresses the fact that Euclidean rotations correspond
to orthogonal transformation). The (covariant) components

F0j = xj

transfrom by

(x0 )j = (F 0 )0,j = Λ0 µ Λj ν Fµν

= Rj k F0k
= Rj k xk (3.6)

which is the rule of transformation of Euclidean 3-vectors under rotations.

Rotations of Fjk
Now consider the triplet 2y j = εjmn Fmn . The Levi-Civita symbol is invariant
under (proper) rotations since det R = 1. Hence

(2y 0 )j = (ε0 )jmn (F 0 )mn

= εjmn Rm a Rn b Fab
= εjmn εabk Rm a Rn b y k
j
= (R−1 )k (2y)k (3.7)

In the last step I used the formula for inverse of a 3 × 3 matrix R

j 1
(R−1 )k = εkab εjmn Rm a Rn b
2 det R
and the fact that det R = 1 for a rotation. It remains to get rid of the inverse.
To do that write
j j
(R−1 )k = (Rt )k = Rj k
48 CHAPTER 3. THE ELECTROMAGNETIC TENSOR

This gives
j
(2y 0 )j = (Rt )k (2y)k
= Rj k (2y)k (3.8)
One sees that x and y both transform as vectors. More precisely, x is a vector
while y is a pseudo-vector as the transformation rule relied on the use of Levi-
Civita (and det R = 1).
We can use these facts to identify the slots for E and B. The force f is a
vector. By the Coulomb law E must be a vector. In contrast, the Lorentz force
involves a cross product, which uses a hand rule: B is a pseudo-vector.

3.1.2 The overall of sign

It remains to choose the overall sign. This depends on choice of the Minkowski
metric2 : If η = (−1, 1, 1, 1). the right choice turns out to be (see the next
section)
 
0 −Ex −Ey −Ez
 Ex 0 B z −B y   ⇐⇒ Fk0 = Ek , Fjk = εijk B i (3.9)
Fµν = 
 Ey −B z 0 Bx 
Ez B y −B x 0
Note that the Cartesian co and contravariant components of Euclidean vectors
coincide.
Exercise 3.4 (Contravariant components). Verify that
 
0 Ex Ey Ez
µν
 −Ex 0 B z −B y

F =  −Ey −Bz
 (3.10)
0 Bx 
−Ez By −Bx 0
Exercise 3.5 (Mixed components). The matrix associated with the mixed tensor
F is neither symmetric nor anti-symmetric. Verify that
 
0 Ex Ey Ez
 Ex 0 Bz −By 
F µν =  Ey −Bz
 (3.11)
0 Bx 
Ez By −Bx 0
Remark 3.6 (Coordinate free form). In a coordinate free form

F = Fµν eµ ⊗ eν
Exercise 3.7 (Coulomb in spherical coordinates). Using the rules of tensor
calculus, show that F for Coulomb field in spherical coordinates has F0r =
−Fr0 = re2 and all other components are 0.
2 And also on the covariant form of Couloms-Lorentz law, see the next section
3.1. AMALGAMATING E AND B 49

Example 3.8. The covariant components of the field tensor in cylindrical co-
ordinates (ct, ρ, φ, z) are
 
0 −Ex c − Ey s ρ(−Ey c + Ex s) −Ez
 ... 0 ρBz −By c + Bx s 
Fcylind =
 ...

... 0 ρ(Bx c + By s) 
... ... ... 0

where c = cos φ and s = sin φ. In normalized coordinates

 
0 −Ex c − Ey s −Ey c + Ex s −Ez
 ...
 0 Bz −By c + Bx s 

 ... ... 0 Bx c + By s 
... ... ... 0

Exercise 3.9 (Magnetic field of currnet line). A line of current I along the
z-axis creates a magnetic field B = 2I
cρ θ̂ in cylindrical coordinates. Show that
in cylindrical coordinates Fρz = −Fzρ = 2I cρ and all other componnets vanish.
(Hint: It is simpler to use properties of the basis vectors eρ , ez rather than
compute the transformation matrix.)

3.1.3 Covariant formulation

It is natural to expect that the force f is a 3-dimensional shadow of a 4-force
vector. The Coulomb-Lorentz law generates this force from an anti-symmetric
second rank tensor F and from the 4-velocity u. There is essentially one way to
do that, namely
e
fµ = Fµν uν
c
Note that the 4-force has the desirable property that it is automatically Minkoski
orthogonal to the 4-velocity, no matter what the fields are:
e
fµ uµ = Fµν uµ uν = 0
c

Exercise 3.10. Explain why the last identity is true and why it is desirable.

Since this is a tensorial identity, it holds in any coordinate system. In par-

ticular, it holds in all inertial frames.
We can now use this to verify that the signs we picked in Eq. (3.9) are the
good signs: To check the sign of E consider a particle at rest: uµ = c(1, 0, 0, 0).
Then
e
fµ = Fµν uν = eFµ0 =⇒ f = eE
c
as it should. To check the sign for B consider a particle moving in a magnetic
50 CHAPTER 3. THE ELECTROMAGNETIC TENSOR

field (without electric field) with 4-velocity u = γ(c, v) . Then

e
fk = Fkj uj
c
e
= γFkj v j
c
e e
= γ εkjm B m v j = γ εkjm v j B m
c c
e
= γ (v × B)k
c
For a slow particle γ ≈ 1 we recover Lorentz force.

3.2 The electromagnetic potential

The homogeneous Maxwell equation:

∇·B =0

says that there are no magnetic monopoles. In Euclidean space this statement
is equivalent to the fact that B is derived from a vector potential A:

B=∇×A

A second (vector valued) Maxwell equation is Faraday law of induction3

 
1 ∂t A
∇ × E = − ∂t B =⇒ ∇ × E + =0 (3.12)
c c

and this says that E + ∂tcA is the gradient of a scalar potential −φ. E and B
are therefore derived from potentials
1
B = ∇ × A, E = −∇φ − ∂t A (3.13)
c

3.2.1 Covarinat formulation

The scalar and vector potentials φ and A can be amalgamated to a 4-vector
potential
Aµ = (−φ, A) (3.14)
(recall η = (−1, 1, 1, 1)). Given a 4-vector field, we can construct from it the
anti-symmetric tensor field

Fµν = ∂µ Aν − ∂ν Aµ (3.15)

In components:
1
Ej = Fj0 = ∂j A0 − ∂0 Aj = − ∂t Aj − ∂j φ,
c
3 In SI units c is absorbed in B.
3.2. THE ELECTROMAGNETIC POTENTIAL 51

and
B i = 21 εijk Fjk = (∇ × A)i
We have reproduced Eq. 3.13.
Exercise 3.11. Show that if A is a 4-vector field then F defined through the
4-potential transforms like a second rank tensor under arbitrary (possibly curvi-
linear) coordinate transformation.

3.2.2 Gauge freedom

The field tensor F does not determine the 4-potential A uniquely. For any
(scalar) function Λ of space-time
Aµ → Aµ + ∂µ Λ (3.16)
share the same fields. In particular, the vector potential ∂µ Λ generates no fields:
∂µ (∂ν Λ) − ∂ν (∂µ Λ) = (∂µ ∂ν )Λ − (∂ν ∂µ )Λ = 0
since derivatives commute. Fields can be measured at point in space time and
so have a direct physical meaning. Potentials are tools for computations.
Exercise 3.12. Suppose that someone tells you that Aµ (x) is time-like (or
space-like) at a given point x. What information does this give on the fields.
(Answer: None)

3.2.3 Non-local gauge invariants Lorentz scalars

xΜ

xΝ
Figure 3.1: Loop and surface element for Stokes.

Although A is not gauge invariant, line integral of A over closed loops in

space-time is a gauge invariant scalar. By Stokes:
I Z
Aµ dx = Fµν dS µν
µ

where dS is the area element spanned by the loop. Both sides are manifestly
Lorentz scalars, and the right hand side is manifestly gauge invariant.
52 CHAPTER 3. THE ELECTROMAGNETIC TENSOR

Exercise 3.13. Prove Stokes for the square planar loop in Fig 3.1.

A familiar, special case of the formula is a closed loop at fixed time, dxµ =
(0, dxk ), where
I I Z Z
k
Ak dx = A · d` = ∇ × A · dS = B · dS = Φ

gives the magnetic flux through the loop.

Exercise 3.14. Suppose that γ is a curve in Euclidean spaceR and consider the
surface S spanned
R by the curve for t ∈ [0, t0 ]. Show that F dS is the emf
action, i.e. Eemf dt.

Quantum mechanics gives a fundamental unit of magnetic flux Φ0 = 2 ×

10−15 [W eber]. In SI [W eber] = [~/e]. The unit is such that there are about 10
quantum flux quanta flux through a 1 [µ2 ] area of the earth magnetic field. So for
a bacterium a quantum flux is a natural magnetic flux scale. It is interesting,
and even mysterious, that when quantum mechanics meets special relativity,
the Lorentz scalars give rise to quantized objects. Here are some examples
whey Φ0 shows up: It is the quantized magnetic flux in vortices that thread
a
R superconductor; The charge of magnetic monopoles; and it is the emf action
E · dt that shows up in the quantum Hall effect.

Exercise 3.15. What is Φ0 = 2 × 10−15 [W eber] in c.s.g in terms of ,̄e and c .

3.2.4 What does a voltmeter measure?

Fields are measurable. Potentials are not. No instruments can measure the
potential at a point. What does a voltmeter measure?

a b
V

R
Figure 3.2: Voltmeter measures the electromotive force: Eab = − E · d`, which
is gauge invariant but, in general, path dependent (and non-local). In contrast
with A0 which is gauge dependent, but path independent (and local). It is path
independent provided the loop defined by the path (black) and its variation
(gray) enclose no time-dependent magnetic fields.
3.2. THE ELECTROMAGNETIC POTENTIAL 53

R
The emf E = γ E · d` is manifestly gauge invariant and so, in principle,
measurable. In general, it is non-local4 and path dependent. It is not a prop-
erty of a single event but the pair (a, b) and, in general, depends on the path
connecting them. The integral form of Faraday induction law is
I
Φ̇
E · dx = −
c

where Φ is the magnetic field enclosed by a loop (at a given time slice). If the
variation of the path γ does not enclose any time dependent magnetic fields
Faraday induction law, Eq. 3.12, says the emf is (at least locally) path indepen-
dent. We may then define E = −∇φ. The difference in φ is what a voltmeter
measures.

3.2.5 Generalizations
Manifolds

Since we wrote the equations in a tensorial form, they have a natural translation
to space-time manifolds that are only locally Minkowski.

d + 1 dimensions

It may well be that the apparent dimension of space-time as 3+1 is the dimension
we perceive on macroscopic scales whereas the dimension is different (larger) in
the microscopic scale. It is a nice feature of the formalism that one can formulate
electrodynamics in d+1 space time dimensions. The vector potential Aµ has d+1
components, and the field F is still a second rank antisymmetric tensor. It has
d+1


2 components. d components “electric” and the remaining are “magnetic”.
In d = 3 the number of electric and magnetic component coincide. In lower
dimension there are more electric components andd in higher dimensions more
magnetic.

Exercise 3.16 (1+1 dimension). How many components does the electric field
have in one dimension and how many the magnetic field?. Show that the electric
field is a scalar.

Non Abelian gauge fields

QCD, and other nonabelian gauge theories, are the non-commutative general-
izations of electrodynamics in the sense that the (real, commutative) 4-vector
potential Aµ , is replaced by a Hermitian (matrix valued) 4-vector potential.
4 To measure the emf oen needs to measure E simultaneously along the Euclidean path γ.

One could argue that a simple instrument, like a AVOmeter can’t possibly do that exactly. It
is at best an idealization.
54 CHAPTER 3. THE ELECTROMAGNETIC TENSOR

3.3 Lorentz scalars

One can construct interesting Lorentz scalars from the field tensor F .

3.3.1 Local scalars

By local scalars I mean scalars one can construct from F at a given event. There
are no interesting scalars that are linear in the field at a given point since

Fµν η µν = Fµ µ = 0

We can however construct interesting quadratic scalars:

Fµν F µν = F0j F 0j + Fj0 F j0 + Fjk F jk

= −2Ej Ej + 2Bi Bi
= −2(E2 − B2 )

You might be worried that we have made a sign error: E2 + B2 is proportional

to the energy density of the field. Why the minus sign? This is the same minus
sign you find in Lagrangian mechanics: The Lagrangian is the difference of the
kinetic and potential energies, not their sum.

3.3.2 Dual: F ∗
Duality, denoted by ∗, is an operation whose square in the identity: ∗∗ = 1.
This means that taking a dual involves no loss of information. In n dimensions,
the Levi-Civita tensor allows us to define a duality for anti-symmetric tensors of
rank r and anti-symmetric tensors of rank n − r. Note first that
anti-symmetric
tensors of rank r make a linear space whose dimension nr (the number of
independent components). The contraction of the Levi-Civita with an anti-
symmetric
n

n−r

tensor of rank r gives an anti-symmetric tensor of rank n − r. Since
r = r the operation can be used to define a duality.
In 4-dimensions the dual of an anti-symmetric second rank tensor is a second
rank tensor. The dual of F is then defined by
p
∗ µν εµναβ ∗ |g|εµναβ αβ
(F ) = p Fαβ ⇐⇒ (F )µν = F (3.17)
2 |g| 2

In Lorentz-Cartesian coordinates,
p g = η and |g| = 1, of course. To avoid writing
ugly formulas involving |g| is shall sacrifice generality for transparency and
write formulas in Lorentz-Cartesian coordinates.

Exercise 3.17. Show that ∗ is indeed a duality. Namely, show first that

1.
εαβγδ εαβµν = 2(δα µ δβ ν − δα ν δβ µ )
3.4. THE HOMOGENEOUS MAXWELL EQUATIONS 55

2. Using this show that

F ∗∗ = F

F ∗ effectively interchanges E and −B. In Minkowski-Cartesian coordinates:

(F ∗ )0i = 21 ε0ijk Fjk = εijk Fjk = B i = Bi

Similarly,

(F ∗ )jk = 21 εjkαβ Fαβ = εjk0i F0i = ε0jki F0i = −εjki Ei

3.3.3 Second Lorentz scalar: E · B

Evidently,
(F ∗ )µν Fµν = 2(F ∗ )0j F0j + (F ∗ )jk Fjk = −4E · B (3.18)
is a Lorentz scalar.
It follows that there are nine classes of Lorentz distinct fields:
 
> 0
 > 0

E2 − B2 = = 0 E·B= =0
 
<0 <0
 

For example, the field of a charge moving at uniform velocity has E2 − B2 >
0 and E · B = 0. Similarly, The field of a plane electromagnetic wave has
E2 − B2 = E · B = 0, in any frame.

3.4 The homogeneous Maxwell equations

The 4 homogeneous Maxwell’s equations express the fact that F is derived from
the potential A. Write
1
∂β Aα = 2 (∂β Aα − ∂α Aβ ) + 12 (∂β Aα + ∂α Aβ )
| {z } | {z }
anti−symmetric symmetric

Since Levi-Civita is completely anti-symmetric, any5 vector field, Aµ , trivially

satisfies
0 = 2εµναβ ∂νβ Aα = εµναβ ∂ν Fβα
since ∂µν is symmetric. Since the Levi-Civita symbols are fixed numbers we can
rewrite this as  µναβ 
ε p 
0 = ∂ν Fβα = ∂ν |g|(F ∗ )µν (3.19)
2
by the definition of the dual 3.17.
5 Twice differentiable
56 CHAPTER 3. THE ELECTROMAGNETIC TENSOR

Exercise 3.18 (Alternate form). Show that 4 equations (3.19) can also be
written as
∂α Fβγ + ∂β Fγα + ∂γ Fαβ = 0
Exercise
p 3.19.Show that if F is an anti-symmetric second rank tensor then
∂µ |g|(F ∗ )µν transforms like a 4-vector under curvilinear coordinate trans-
formations.
In mechanics the price for using the a non-inertial coordinate system, such as
earth, leads to the price of the emergence of new forces: Coriolis and Centrifugal.
You may wonder if there is an analog in electrodynamics. The next exercise
explains why there is none.
Exercise 3.20. Consider the coordinate system attached to the rotating earth
introduced in the previous chapter.
1. Compute det g for the earth rotating coordinate system. (Answer: det g =
−1)
2. What does this imply about the homogeneous Maxwell equations in the
earth frame?

3.4.1 Maxwell form

The 0 component of Eq. (3.19) is the statement that magnetic fields are source-
less:
0 = ∂µ (F ∗ )0µ = ∂j (F ∗ )0j = −∇ · B (3.20)
The spatial components give Faraday law:
1
0= ∂t B + ∇ × E (3.21)
c
The two equations have different physical character. The first equation can
be viewed as a constraint equation on the admissible magnetic fields at any given
time. The second may be viewed as an equation that determines the evolution
in time of the magnetic fields (given the electric field). This system must be
consistent. Namely, if B starts divergence-less, it must evolve in a way that its
stays divergence-less. This is indeed the case:
1


∂t ∇ · B = −∇ · ∇ × E = 0 (3.22)
c
Exercise 3.21 (Harmonic A). Given the vector potential
ν
Aµ (x) = Aµ eikν x

Compute F . Show that

Fµν F µν = −2 kµ k µ Aν Aν − (kµ Aµ )2 , Fµν (F ∗ )µν = 0

Chapter 4

Variational principle

The equations of motion of a relativistic (charged) particle are formulated

through a variational principle .

4.1 The action begins

Lagrangian mechanics reformulates Newtonian mechanics: Newton equations
are interpreted as the Euler-Lagrange equations for the minimizers of the action.
One advantage of the Lagrangian formulation is that it facilitates implementing
basic symmetry principles, such as Lorentz invariance.
The property of being minimizer does not depend of the the choice of co-
ordinates and so the Lagrangian formulation guarantees the tensorial character
of the equations. Since Lorentz transformations are special coordinate trans-
formations of Minkowski space, a theory is guaranteed to be Lorentz invariant
once the action is a Lorentz scalar. The action S associates a scalar to an orbit.

Figure 4.1: The blue curve shows S(x) = x2 . The black curve shows the action
under a change of coordinate x → tan x, a change of scale and shift S →
2S − 1/10. The action changes, but the physical point p where the minimum
occurs is the same. In the example also the coordinates of the points are the
same: x = 0.

Requiring that the action be a Lorentz scalar which reduces to the Newtonian
form for non-relativistic motions fixes the action uniquely.

57
58 CHAPTER 4. VARIATIONAL PRINCIPLE

Figure 4.2: The world line is required to have time-like tangents–otherwise the
action is complex. The blue curve represents the variation of a world line with
fixed end points. The black line maximizes the proper-time τ . The red lines are
light-like. The red path between the end points has zero proper-time.

4.1.1 Action for a free particle

Let us start with a free particle with mass m > 0 (a scalar). A natural Lorentz
scalar associated to the orbit in the elapse off proper time τ . Since the action
has units of [p][x], we append scalars c and m to fix the dimension:
Z Z
2
Sp = −mc dτ
|{z} = L dt
|{z} (4.1)
proper time coordinate time

and
r
mc dxα dxβ mc2
q
L=− −gαβ (x) =− −g00 c2 − 2g0j cv j gjk (x)v j v k
γ dt dt γ

In particular, in Minkowski space-time with metric

 
−1 0 0 0
 0 
g=  0

gjk 
0

and a particle moving non-relativistically, γ → 1 and

m
mc2 +
L → − |{z} gjk (x)v j v k
const |2 {z }
kinetic energy

This is the same as the classical Lagrangian of free particle, up to a (large)

negative constant −mc2 . Adding a constant to the action does not affect the
minimizing path, of course.
4.2. VARIATION OF THE ACTION 59

Example 4.1. In rotating (earth) coordinates, 2.3.3,

1 1  
≈ 1 − 2 v2 − Ω2 ρ2 + Ω φ̇ρ2
γ 2c
 
1  2
=1− v − (Ω × x)2 + 2 Ω · v × x

2c2
 |{z}
| {z } | {z }
Kinetic centrif ugal Coriolis

4.1.2 Interaction
The action in Mechanics is constructed from the potentials, not the fields. The
electromagnetic potential is a 4-vector, which we can pair with the path element
dxµ to form a scalar. Electromagnetic interaction is, of course, proportional to
the charge. In c.g.s we also need c to fix the units and the overall sign is fixed
by the choice of the metric η = (−1, 1, 1, 1) so
Z Z Z
e µ e
Sint = Aµ dx = −e φ(x)dt + A(x) · v dt (4.2)
c c
This is precisely the terms one adds to the Kinetic energy in classical mechanics
to describe the interaction with E and B.

4.1.3 Gauge invariance

Since the action is constructed from the potentials, one should worry about the
gauge invariance of the minimizing orbit. Under change of gauge

A0µ = Aµ + ∂µ Λ (4.3)

This leads to
Z
0 e e
(∂µ Λ) dxµ = Sint −


Sint = Sint − Λ(xf ) − Λ(xi ) (4.4)
c c
This means that although Sint changes under a gauge transformation, its vari-
ation δSint does not (so long as the end points are fixed). This guarantees that
the Euler Lagrange equaltions are gauge invariant.

4.2 Variation of the action

The action is a function on paths: It associates a number with a given path
xµ (τ ). We may think of the path as parametrized by its proper time. The
action is assumed to have the form
Z xf
S= f (x, u)dτ
xi
60 CHAPTER 4. VARIATIONAL PRINCIPLE

i.e. it is a function of the position and velocity. The end point events xi
and xf are fixed. The action associates a number for every path x(τ ) =
{x0 (τ ), . . . x3 (τ )}. We shall denote the variation of the path by

δx = { δx0 (τ ), . . . δx3 (τ ) }
| {z }
inf initesimal f unctions

The rule of the game is that the end points are fixed: δx vanishes at the end
points events. (The events which are fixed not the proper-times at the end
points.) The strategy is to use integration by parts to bring δS to the form

x
Z xf
δS = hµ (x, u)δxµ xfi + gµ (x, u, u̇)δxµ dτ
xi

Since δxµ vanish at the end points the first term drops. And since δxµ are
arbitrary,1 the only way for the integral to vanish is if

gµ (x, u, u̇) = 0

These are differential equation that the optimal path must satisfy. They are
known as Euler-Lagrange. Lets consider examples.

4.2.1 Variation of Sp
In Minkowski Cartesian coordinates2 , g = η, the variation in the proper-time
with is given by

δ(cdτ )2 = 2c2 dτ δ(dτ ) = δ(−dxµ dxµ ) = −2dxµ δ(dxµ )

This can be written in terms of the 4-velocity as

c2 δ(dτ ) = −uµ δ(dxµ ) = −d uµ δxµ + duµ (δxµ )

The variation of the free action is then

Z
− δSp = −m uµ δxµ end u̇µ δxµ dτ

pts
+m (4.5)

By assumption, the variation vanishes at the boundary so the first term drops.
since δx is arbitrary, so long as it is small, if follows that δSp = 0 is equivalent
to conservation of 4-momentum:

ṗµ = mu̇µ = 0
1 Viewed as functions of τ the variations satisfy one constraint: (dδxµ )(dδxµ ) = −(cdτ )2
2 The case of a general metric g(x) is treated in a supplement
4.2. VARIATION OF THE ACTION 61

4.2.2 Variation of Sint

For Sint write

δ(Aµ dxµ ) = (δA)µ dxµ + Aµ δ(dx)µ = (∂ν Aµ ) (δxν )dxµ + Aµ δ(dx)µ

The basic idea in the calculus of variation is to use integration by parts to get
rid of terms of the form δdx. Hence, rewrite the last term

Aµ δ(dx)µ = d Aµ δxµ − (dA)µ δxµ = d Aµ δxµ − (∂ν Aµ )dxν δxµ

Combining the two expressions and changing summation indices where needed
we find

δ(Aµ dxµ ) = d Aµ δxµ + (∂µ Aν ) (δxµ )dxν − (∂ν Aµ )dxν δxµ

 
= d Aµ δxµ + ∂µ Aν − ∂ν Aµ uν dτ δxµ
= d Aµ δxµ + Fµν uν dτ δxµ

Hence,
Z
e e
Aµ δxµ |bdry + Fµν uν dτ δxν


δSint = (4.6)
c c

4.2.3 Euler-Lagrange equation

The variation of the total action vanishes for any δxµ provided the integrand in
Sf ree + Sint add to zero. This gives the Euler-Lagrange equation

e
m u̇µ = Fµν uν (4.7)
c

4.2.4 The non-relativistic limit

By general principles, we know that Eq. 4.7 must reduce to the standard formu-
las of non-relativistic classical mechanics. It is also easy to check this directly:
For a slow particle has
uµ ≈ (c, v), u̇j = aj

(recall that η = (−1, 1, 1, 1)). The Euler-Lagrange equations reduce to Newton

equations of motions
e
ma = eE + v × B (4.8)
c

Remark 4.2 (Sign conventions). You can use the non-relativistic limit to fix
and verify the signs convention in 3.9
62 CHAPTER 4. VARIATIONAL PRINCIPLE

4.2.5 Existence and uniqueness

In general the action does not have an honest minimizer and when it does it
not be unique. The (non-relativistic) Harmonic oscillator is an example. The
Lagrangian is
2L = ẋ2 − x2 (4.9)
Now consider paths that starts and terminates at the origin in half the period,
t = π. There are many that satisfy the Euler-Lagrange equation

xn (t) = A sin t

for arbitrary amplitude A. All solve the equation of motion ẍ = −x and so are
local minimizers. For all the action vanishes:
Z π Z π
S= Ldt = A2 (n2 cos2 nt − sin2 nt)dt = 0,
0 0

At the same time there is no nice minimizer connecting the origin and any other
point at half the period.

Figure 4.3: There are infinitely many paths connecting the origin when the time
difference is half the period. But there is no honest minimizer connecting the
origin to any other point on the red line at half the period. The minimizer ”goes
through infinity”.

Exercise 4.3. Consider the family of paths parametrized by A > x0 ≥ 0

(
x0 0 < sin t < x0 /A
xA (t) =
A sin t, sin t > x0 /A and t ≤ π

Note that that the second segment is a solution of the Euler-Lgrange equation
for the Harmonic oscillator Eq. 4.9. Compute the action of the two segments
for A/x0  1. Show that when A → ∞, the total action S(xA ) ≈ −x0 A and
hence diverges to −∞ if x0 6= 0.

Convexity and uniqueness

A function S is called convex if

S(λx + λ0 y) ≤ λS(x) + λ0 S(y), λ + λ0 = 1, 1 ≥ λ, λ0 ≥ 0

4.3. SUPPLEMENT 63

For example, the functions in Fig. 4.1 are convex. It is evident that if a function
is convex its minimum is unique. (It may, however, lie at infinity).
In the case at hand x, the argument of S, is itself a function–a path. Func-
tions naturally form a vector space so the notion of convex function also extends
to this case. γ is a convex function of v. This then implies that the minimizer
for a free relativistic particle is unique.

Figure 4.4: γ is a convex function of v . This implies that S is a convex

functional of the path.

4.2.6 Consistency check

The 4- acceleration is always (Lorentz) perpendicular to the 4-velocity (by the
constancy of uµ uµ ). This is indeed respected by the equation of motion since
Fµν is antisymmetric:
e
m u̇µ uµ = Fµν uµ uν = 0
c
Exercise 4.4 (Charged particle in a constant fields). Solve the equations of
motion of a charge particle in constant parallel electric and magnetic fields

Exercise 4.5 (Charged particle in a radiation field). Show that the equations
of motion of a charge particle in the radiation field of a circularly polarized
plane wave admit solutions that are circular orbit in the plane orthogonal to the
direction of propagation of the light.

4.3 Supplement
4.3.1 Fermat principle
The mother of variational principles is Fermat principle. It formulates geometric
optics at the the minimizer of the time of propagation between two points. The
ray propagates in a medium with index of refraction n(x). The propagation
time dt is c dt = n(x)|dx|. We can think of n as inducing a metric in Euclidean
space–one that measures the propagation time:

(c dt)2 = n2 (x) dx · dx = n2 (x) (d`)2

Such a metric is called conformal.

64 CHAPTER 4. VARIATIONAL PRINCIPLE
√
Consider a ray x(`) parametrized by it Euclidean length: d` = dx · dx.
The tangent to the ray
dx
t=
d`
is a unit vector. For a variation of the path δx:

dx · dδx
δ(d`) = = t · dδx, δn = (δx · ∇)n
d`
The corresponding time variation is (the integral of)
 


δ(c dt) = (δx · ∇)nd` + nt · dδx = δx · ∇n d` − d nt + d nt · δx)

The integral of variation vanishes provided the brackets (and boundary terms)
vanish:
d(nt)
= ∇n
d`
In particular, in a region where the refraction index is a constant, ∇n = 0, the
ray keeps it direction of propagation: t is a constant.

Exercise 4.6. Show that the equation of motion is consistent with the t being
a unite vector.

Θ1 n1

Figure 4.5: The change in direction of a ray when n jumps is ddetermined by

Snell’s law

Exercise 4.7. Derive Snell’s law, fig.4.5,

n1 sin θ1 = n2 sin θ2

from Fermat principle.

4.3.2 Rainbow
The simplest features of the rainbow can be understood from Snell’s law.
4.3. SUPPLEMENT 65

Φ Α=2Φ-Θ
Θ

Figure 4.6: The blue line shows a ray undergoing one internal reflection in a
drop of water. The impact angle θ is defined in the figure. The outgoing ray
is focused near the maximum 2φ − θ. This partial focusing is called a caustic.
This gives the direction of the rainbow.

Exercise 4.8. Use Snell law and show that a light ray in air (na = 1) hitting a
water droplet, nw > 1, at lattitude θ is reflected back at angle 2α(θ) = 4φ(θ)−2θ,
see figure. The function φ(θ) is defined by Snell law: nw sin φ = sin θ.

The intensity of the light reflected at angle I(2α) is proportional to the

intensity of the incoming light:

d(sin θ) = I(2α) |dα|

A computation gives
dα cos θ
= −1 + 2 √ 2
dθ n − sin2 θ
The derivative vanishes for

3 cos2 θ = n2 − 1

which gives a real value for θ provided 1 < n < 2. This gives the maximal
value of 2α. Evidently, I(2α) = ∞ there. The divergence implies focusing of
the reflected light. This is called caustic in geometrical optics.

Exercise 4.9. Show that for water (n = 1.33) the caustic occurs for 2α = 42◦ .
This is the main angle of the rainbow, first found by Bacon in 1268. (Different
colors have slightly different angles due to the slight frequency dependence of n).

4.3.3 Geodesics in Curved space-time

This section is an aside. We set c = 1. In a general space-time (not necessarily
Minkowski) the notion of proper-time is defined, as usual, by

(dτ )2 = −gµν (x) dxµ dxν

The 4-velocity is
dxµ
uµ =
dτ
66 CHAPTER 4. VARIATIONAL PRINCIPLE

Figure 4.7: The cyan arrows represent light rays from the sun. The two small
light-blue balls represent two water droplets. The red arrow are the reflected
light rays in the direction of the rainbow caustics. The green eye represents the
observer. Pilots sometimes see rainbow that are circular.

which is normalized to −1 since

(dτ )2 = −gµν (x) uµ uν (dτ )2 = −uµ uµ (dτ )2

Exercise 4.10 (Acceleration). When the metric is posiytion dependent it is not

true that the acceleration is orthogonal to the velocity. Show that

2u̇µ uµ + (∂α gµν )uµ uν uα = 0

We want to find the path that minimizes the action (equivalently, maximizes
the proper time) Z
S=− dτ

The variation of (dτ )2 is

δ(dτ )2 = 2(dτ )δ(dτ ) = −(δgµν ) dxµ dxν − 2gµν δ(dxµ )dxν

Hence
−2δ(dτ ) = (δgµν ) uν dxµ + 2gµν uν δ(dxµ )
Rewrite the first therm as

(δgµν ) uν dxµ = (∂α gµν ) uν dxµ δxα = (∂µ gαν ) uν dxα δxµ

The second term can be rewritten as

gµν δ(dxµ )uν = d(gµν δxµ uν ) − d(gµν uν ) δxµ

= d(gµν δxµ uν ) − (∂α gµν ) dxα uν δxµ − gµν duν δxµ

collecting
 
− δ(dτ ) = d(gµν δxµ uν ) + 1 ν
2 (∂µ gαν ) u dx
α
− (∂α gµν ) uν dxα − gµν duν δxµ
4.3. SUPPLEMENT 67

The first term is a boundary term so vanishes upon integration. The vanishing
of the variation gives the Euler-Lagrange which is of the form

u̇µ + Γµαβ uα uβ = 0 (4.10)

and Γµαβ may be assumed to be symmetric in α, β without loss. One then finds
 
Γµαβ = 21 g µν ∂α gνβ + ∂β gνα − ∂ν gαβ (4.11)

aka as the Christoffel symbol.

Exercise 4.11. Show that great circles on the sphere are geodesics.
Exercise 4.12 (Geodesic equation in covariant components). Show that the
geodesic equation for the covariant components satisfies the equation
 
2u̇µ = − ∂µ g αβ uα uβ
68 CHAPTER 4. VARIATIONAL PRINCIPLE
Chapter 5

Maxwell Equations

The non-relativistic notions of charge and currents is amalgamated into a single

notion in space-time the 4-current and the inhomogeneous Maxwell equations
are derived from a variational principle.

5.1 Technical preliminaries

5.1.1 Space-time volume element
In Riemannian geometry the volume element is1
√ Y j
dV = g dx , g = det g

We stick with the same definition in Minkowski geometry paying the small price
g → |g|, taking care of the fact that the metric is non-definite. We shall denote
by dΩ the space-time volume element.
Lorentz transformations are isometrics of Minkowski space-time. In Carte-
sian coordinates det η = −1. It follows that the Cartesian expression for the
space-time volume is a Lorentz scalar:
p
dΩ = |η|dx0 dx1 dx2 dx3 = dx 0 1 2 3
| dx {zdx dx} = (cdt) dV (5.1)
M inkowski cartesian

in agreement with the fact that time-dilation and space contraction have com-
pensating factors.

5.1.2 Densities
Under a change of coordinates x0 ↔ x a scalar function ϕ(x) has the transfor-
mation rule
ϕ0 (x0 ) = ϕ(x)
1 This is clearly the right notion of volume in orthogonal coordinates where g is diagonal.

The tensorial properties then guarantee that this expression holds in general.

69
70 CHAPTER 5. MAXWELL EQUATIONS

Figure 5.1: The volume of the parallelepiped spanned by the vectors vj is their
triple product.

It retains the value at the image point. Densities, ρ(x) are different. The
associated scalar is ρ(x)dV . It may count the charge in the volume, or the
number of particles there. The rule is therefore

ρ0 (x0 )dV 0 = ρ(x)dV

But dV 0 is related to dV be the Jacobian of the transformation. As a conse-

quence ρ(x) does not transform like a scalar, but like a density:
ρ
p
|g|

transforms like a scalar.

5.1.3 Distributions
We want to introduce Dirac delta function in Minkowsky space-time. The defin-
ing property is Z
f (0) = dΩf (x)δ (4) (x) (5.2)

for any smooth function (scalar) f .

Remark 5.1. In Minkowski-Cartesian coordinates dΩ = cdtdV is a Lorentz

scalar. It follows that δ (4) (x) is a Lorentz scalar.

Exercise 5.2. Write the delta function δ 3 (x − x0 ) in spherical coordinates.

Remark 5.3 (Functios and Distributions). Dirac delta function is an example

distributions. Distributions and functions are distinct objects:
5.1. TECHNICAL PRELIMINARIES 71

• Smooth functions form an algebra: You can add and multiply them in the
obvious way. Distributions are a vector space: You can add them but not
multiply. There is not such thing as a square of a delta function, or a root
of a delta function.
• Distributions are naturally viewed as the dual vector space to the space of
smooth and localized functions. With a pair we associate the number given
by integration.
• A distribution D(x) is the zero distribution if for any smooth and localized
function f (x) Z
f (x)D(x)dx = 0

An example is xδ(x). In contrast, the useful notion of the trivial (zero)

function in the theory of integration is: f (x) = 0 for all x except for a set
of measure zero.
• All linear operations on distributions are allowed. In particular, distribu-
tions D can be differentiated according to the rules of integration by parts,
as many times as one pleases
Z Z
f (x)D(n) (x) dx = (−)n−1 f (n) (x)D(x)dx

where f is a smooth (infinitely differentiable) function. In contrast, func-

tions may or may not be differentiable. For example, the derivative of the
step function, θ, is the Dirac distribution.
Exercise 5.4 (Delta of f ). Suppose f (x) is a nice function, such that f (xj ) = 0
and f 0 (xj ) 6= 0. Show that

X δ(x − xj )
δ f (x) =
j
|f 0 (xj )|

Figure 5.2: The function f (x) associated with exercise 5.4.

72 CHAPTER 5. MAXWELL EQUATIONS

5.2 Charge densities and currents

For a charges e moving on the trajectories ξ(t) in 3 dimensions, the non-
relativistic notion of charge density ρ and current j is defined by
ρ(x, t) = eδ (3) x − ξ(t) , j(x, t) = eδ (3) x − ξ(t) v(t)



(5.3)

where v = ξ˙ is the velocity. For several particles, of charges ea , moving of

different trajectories ξa , the generalization is obviously
X X
ea δ (3) x − ξa (t) , j(x, t) = ea δ (3) x − ξa (t) va (t)



ρ(x, t) = (5.4)
a a

We would like to amalgamate ρ and j into a n option of a 4-current-density.

However, neither v nor δ (3) (x) are natural objects in space-time and, a-priori,
you may well worry that the non-relativistic notions of charge and currents are,
at best, a non-relativistic approximations so we will need to tinker with them
before being able to amalgamate them. It will turn out that these expressions
are fine as they stand. (We still need to adjust dimensions so we can fit both ρ
and j in a 4-vector with identical dimensions.)
It is clear that is is enough to formulate the notions of current and density
for a single particle. This simplifies the notation.

5.2.1 4-current-density
Consider a point particle whose trajectory is given ξ µ (τ ) as a function of its
proper-time. For the sake of simplicity, we work in Minkowski Cartesian co-
ordinates. We can make a 4-current density using only scalars and 4-vectors,
objects that behave nicely under Lorentz transformations:
Z
j µ (x) = ec dτ δ (4) x − ξ(τ ) ξ˙µ (τ )


(5.5)

where dot is a derivative with respect to the proper time. The scalar factor ec
fixes the dimensions to the dimensions of current density.
To relate this expression to Eq. (5.4) integrate over τ , getting rid of one of
the delta functions. Since ξ is a real orbit, there is a 1-1 correspondence between
coordinate time ξ 0 (τ )/c and the proper time τ . Changing variables from cτ to
ξ0
Z Z

dτ
c dτ δ (4) x − ξ(τ ) ξ˙µ (τ ) = c dξ 0 δ (4) x − ξ ξ˙µ 0


dξ
Z
= dξ 0 δ (3) x − ξ δ(ct − ξ 0 )v µ (ξ 0 )

= δ (3) x − ξ(t) v µ (t)

where t0 is the solution of ct0 = ξ 0 (τ ) and v µ = (c, v). The result is a pleasant
surprise because it coincides with
(cρ, j)
5.2. CHARGE DENSITIES AND CURRENTS 73

ct

Figure 5.3: The parametrized orbit ξ µ (τ ).

derived before we knew anything about Lorentz invariance. There are no rela-
tivistic corrections one needs to make to the classical formulas for charge den-
sities and currents. δ (3) is not a density in Minkowsky space, and v µ is not a
4-vector. However, together they conspire to give the 4-vector (density) j µ .

5.2.2 Charge conservation

The total charge, at any given instant,
Z Z
dV ρ(x, t) = e dV δ (3) (x − ξ(t)) = e

is independent of the time slice t2 . This is charge conservation.

The local, differential equations, expressing this is the continuity equation
which we now derive: Consider the rigid transport of a bump function: ρ(x −
ξ(t)). Write ξ˙ = v. Then



∂t ρ(x − ξ(t) = −(v · ∇)ρ x − ξ(t)

Evidently,



∇ · v(t)ρ(x − ξ(t) = (v · ∇)ρ x − ξ(t)
and thus
0 = ∂t ρ +∇ · (vρ) = ∂µ j µ , j µ = (c, v)ρ
|{z} |{z}
density current

A point charge is the limit ρ → δ.

2 And in which frame the slice is taken. You may take this to be the definition of δ (3) .
74 CHAPTER 5. MAXWELL EQUATIONS

ct

Figure 5.4: Charge conservation expresses the fact that the orbit is a continuous
curve which does not terminate and moves always into the future. If it enters
a box in space-time it also leaves it. If the orbit enters the box at the bottom
leaves it at the top we say that the charge in the box is conserved. If it leaves
and enters on the sides we say that incoming current balances the outgoing
current.

Once this equation holds for one charge, it holds for any number. When
we consider huge numbers of charges with poor spatial resolution we may then
think of j µ (x) as a smooth function on space time, which satisfies the continuity
equation.

Example 5.5 (Orders of magnitudes).

1. A current of 1 Ampere transports 6 × 1018 electrons per second

2. A copper wire of cross section S = 1 [mm2 ] has S8 × 1020 [cm−1 ] atoms

per unit length. Since it copper has valence 2, the number of electrons
per unit length is Sn ≈ 1.6 × 1021 [cm−1 ]. If yo use the classical formula
for the current I = env you find that the velocity associated with ampere
in such a wire is very small, about 40 [µ/sec]. In reality, electrons move
in copper with Femi velocity which is of the order of 1/137 of the speed
of light. The small velocity we have computed may then be interpreted as
representing the tiny displacement of the Fermi sphere from the origin.

5.2.3 Current conservation and gauge invariance

We want to generalize the expression for Sint form a finite collection of charges
to a continuous distribution. For a single charge the action representing the
5.3. LAGRANGIAN FIELD THEORY 75

interaction is
Z
e
Sint = Aµ (ξ)uµ (τ ) dτ
c
Z
e
= 2 Aµ (ξ)uµ (τ ) δ (3) (x − ξ(τ )) dV d(cτ )
c
Z
e
Aµ ξ 0 , x v µ (ξ 0 ) δ (3) (x − ξ(t)) dV dξ 0


= 2
c
Z
e
Aµ x v µ (t) δ (3) (x − ξ(t)) dΩ


= 2
c
The middle two terms are the 4-current, hence
Z
1
Sint = 2 Aµ (x)j µ (x) dΩ (5.6)
c
describing both smooth and discrete 4-current distributions.

5.2.4 Gauge invariance and the continuity equation

We have seen that under a change of gauge Sint for a single charge changed
by a boundary term. This then implied the gauge invariance of the Euler-
Lagrange equations. It is interesting to reconsider this issue for smooth current
distributions. We will lean something. Under a change of gauge
Aµ → Aµ + ∂µ Λ
and Z
1
Sint → Sint − (∂µ Λ) j µ dΩ
c2
Allowing for curvilinear coordinates, the integrand can be rearranged as
p p p 
|g|(∂µ Λ) j µ = ∂µ ( |g|Λ j µ ) − Λ ∂µ |g|j µ )

The first term gives a boundary term and vanishes if Λ → 0 at infinity. Sint is
therefore guaranteed to be gauge invariance provided the current satisfies the
continuity equation holds
1 p 
p ∂µ |g|j µ = 0 (5.7)
|g|

5.3 Lagrangian field theory

In Lagrangian mechanics the basic object is the Lagrangian, L(qj , q̇j , t), a func-
tion of the “generalized coordinates” qj and their velocities q̇j and j labels the
degrees of freedom. Lagrangian field theory can be viewed as a generalization of
Lagrangian mechanics to infinitely many degrees of freedom where the discrete
indexR j is replaced by x, a point in space. The Lagrangian is then of the form
L = dx LF where LF is a suitable Lagrangian density for the field: a function
of the fields and their time derivatives.
76 CHAPTER 5. MAXWELL EQUATIONS

Figure 5.5: The action associated a number with a given field configuration and
a box in space-time. We allow variation of Aµ inside the box: The variation
vanishes outside the box and on its boundary. This is the analog of what we do
when we vary the path.

5.3.1 The Lagrangian of the electromagnetic field

Now we come to deciding what replaces the qj and q̇j for the electromagnetic
field. Two natural choices are Fµν and Aµ . The right choice is, of course, the
one that reproduces Maxwell equations and turns out to be
qj ↔ Aµ (x), q̇j ↔ Ȧµ (x)
This agrees with Sint being a function of A not F .
Lorentz invariance of the Euler-Lagrange equations is automatically guar-
anteed if the action was a Lorentz scalar. We have at our disposal two Lorentz
scalars
F · F = Fµν F µν , F · F ∗ = Fµν (F ∗ )µν ,
whose dimensions are energy density. Since the volume element in space-time
dΩ is a Lorentz scalar and since the action must3 have dimension [Et], two
candidates for the field action are suitable numerical multiples of
Z Z
1 1
dΩ F · F, dΩF ∗ · F
c |{z} c
volume element
3 So we can add it to Sp and Sint
5.4. VARIATION OF THE FIELD: RULES OF THE GAME 77

However, using the homogeneous Maxwell equation

(F ∗ ) · F = 2(F ∗ )µν (∂µ Aν )

= 2∂µ ((F ∗ )µν Aν )

This means that the associated action is a boundary term: Its variation vanishes
identically.
We are left with the first candidate. We need first to decide on the sign so
that the action will have a minimum rather than a maximum. Recall that in
Lagrangian mechanics the kinetic energy comes with a positive sign. Now E is
linear in Ȧ so the E2 must come with a positive coefficients.
Z
1
SF = − F ·F dΩ (5.8)
16πc √ Q µ
|{z}
|g| µ dx

The 16π is the choice4 that gives Maxwell equations in c.g.s units and in par-
ticular leads to the Coulomb potential re .

5.4 Variation of the field: Rules of the game

The actions SF assigns a number for any given field Aµ (x). The action is then a
functions whose arguments are functions too. Such objects are sometimes called
functionals.
We consider variation of the action due to variations δAµ . We shall consider
local variations only, namely, variations in a finite region of space time: δAµ = 0
outside some large space-time box, so we do not need to worry about infinite
variations that can come with infinite boxes.

5.4.1 Variation of the field: Calculations

The variation of A causes a variation of F · F which is

δ(Fµν F µν ) = 2F µν δ(Fµν )

and
δ(Fµν ) = ∂µ δAν − ∂ν δAµ
By the anti-symmetry of F

δ(Fµν F µν ) = 2F µν (∂µ δAν − ∂ν δAµ ) = 4F µν (∂µ δAν )

In variational calculation one wants to end up with an expression proportional

to δAµ : We need to get rid of terms of the form δ∂A by integrating by parts.
4 A natural choice that gets rid of the 4π which is special to 3-dimensions, is to replace 16π

by 4 where Coulomb potential is e/4πr.

78 CHAPTER 5. MAXWELL EQUATIONS

However, in the case of curvilinear coordinates, we must also pay attention to

det |g| in the volume element5
p p p
|g|δ(Fµν F µν ) = 4∂µ ( |g|F µν δAν ) − 4( |g|∂µ F µν ) δAν
p
The first term looks like the divergence of the vector field |g|F µν δAν and by
a Gauss type theorem can be converted to a 4-surface integral on the boundary
of the box where δA = 0. Hence,
Z
1 1 p
δSF = bdry terms + dΩ p (∂µ |g|F µν ) δAν (5.9)
4πc |g|
| {z }
Πα dxα

5.4.2 Variation of the interaction

We have already determined the action associated with the interaction when we
studied the dynamics of relativistic charged particles as
Z
1
Sint = Aν j ν dΩ
c2

To get the field equations we consider variations δA for a given source term j.
This variation gives
Z
1
δSint = 2 δAν j ν dΩ (5.10)
c

5.4.3 The inhomogeneous Maxwell equations

The Euler-Lagrange equations for the fields are those that minimize the action
Sf ield + Sint . The minimizer is the stationary point of the variation:
Z  
1 µν 4π ν
0 = δSf ield + δSint = dΩ ∂µ F + j δAν dΩ
4πc c

This will vanish for arbitrary variation δA provided

1 p  4π
p ∂µ |g|F νµ = jν (5.11)
|g| c

These are the 4-inhomogeneous Maxwell equations in curvilinear coordinates.

The space-time formalism encapsulate the inhomogeneous Maxwell’s equations
in a neat a concise form.
5 g is a function of the coordinats. It is not a function of A and is not affected by the

variation.
5.4. VARIATION OF THE FIELD: RULES OF THE GAME 79

5.4.4 Current conservation

We have derived Maxwell equations as the Euler-Lagrange equations for the
field Aµ for a given source term j µ . This derivation did not assume that the
source j µ is a reasonable physical current and did not explicitly require that it
be a conserved current. However, a-posterior, Maxwell equation enforce current
conservation on j as a direct consequence of the fact that F is an antisymmetric
tensor:
p  4π p 
0 = ∂µν |g|F µν = ∂ν |g|j ν
| {z } c
0 by symmetry

in accordance wirh Eq. 5.7. If the source j was not current conserving, Maxwell
equations would not form a consistent set of equation.

5.4.5 3-D form

To translate back Maxwell equations from their covariant space-time form to
3-D form, consider first the ν = 0 equation ∂µ F µ0 = j 0 . Since

−Ej = F0j = F j0 j 0 = cρ

we get Gauss-Coulomb law

4π 0
∇ · E = 4πρ ⇐⇒ ∂µ F 0µ = j (5.12)
c
The spatial components are:

1 4π j
∂µ F jµ = ∂0 F j0 + ∂k F jk = − ∂t Ej − ∂k (εikj Bi ) = j
c c
Using
∂k (εikj Bi ) = −εjki ∂k Bi = −(∇ × B)j

This gives Ampere-Maxwell equation:

4π 4π k
− Ė + ∇ × B = J ⇐⇒ ∂µ F kµ = j (5.13)
c c

5.4.6 Time reversal

Time-reversal of the orbits of the sources ξa (t) 7→ ξa (−t), sends ρ(t, x) 7→
ρ(−t, x) but flips the currents J(t, x) 7→ −J(−t, x). It follows that solutions
to Maxwell’s equations transfrom under time-reversal as E(t, x) 7→ E(−t, x)
and B(t, x) 7→ −B(−t, x). We say that E is even under time reversal and B is
odd.
80 CHAPTER 5. MAXWELL EQUATIONS

5.4.7 Maxwell equations: Structure

The traditional form of Maxwell’ equations, as equations for E and B, is a set
of two scalar equations and two vector equations. The two scalar equations are
Gauss laws:
∇ · E = 4πρ , ∇ · B = 0 (5.14)
| {z }
Gauss
The two vector equations are Faraday and Maxwell-Ampere laws
4π
Ė = ∇ × B − J, Ḃ = −∇ × E (5.15)
c } | {z }
F araday
| {z
Ampere

and dot denotes partial derivative with respect to ct. The vector equations are
written in the form of first order evolution equations that allow to propagate E
and B in time, given their initial values and the source J.
In total, there are 8 Maxwell equations for the 6 unknown fields. This looks
like an over constrained system. It is better to view them as two evolution
vector equation for two vectors and view the scalar equations as a constraint on
the initial data. This constraint is preserved by the evolution provided (ρ, J)
satisfy the continuity equation.
Exercise 5.6. Show that the evolution respects the constraint.
Example 5.7 (Current carrying wire). An electrically neutral, infinitely long,
metallic straight wire (along the z-axis) with circular cross section of radius a
carries a stationary current I. Suppose Ohm’s law in the form J = σE with σ a
constant in the wire. Find the profiles of the electric and magnetic fields inside
and outside the wire. Assume cylindrical symmetry, translational symmetry in
the z direction and stationarity. Analyze the problem in cylindrical coordinates
(ρ, θ, z).
By Gauss and the symmetry Eρ = 0 Since the magnetic field is assumed
stationary, by Faraday and the symmetry Eθ = 0. You might then be tempted
to say that outside the wire, one should have Ez = 0. This, however, leads to a
contradiction: Combinig Gauss and Faraday



0 = ∇ × E ⇒ 0 = ∇ × ∇ × E = −∆E + ∇ ∇ · E ⇒ ∆E = 0
Which says that E is harmonic everywhere. Hence, if it is zero outside the wire,
it is zero everywhere. This, together with Ohm’s law, contradicts the assumption
that the wire carries current.
Let us then retreat to the next line of defense and take E = E0 ẑ with E0 a
constant. This is still harmonic By the integral form of Ampere
(
2I 1 ρ>a
B= θ̂ × ρ 2


cρ a ρ<a

where I is the total current. We have used the fact that inside the wire, the
constancy of E implies the constancy of J.
5.5. DIELECTRIC AND MAGNETIC MEDIA 81

It may be a little shocking at first that a neutral current carrying wire bundles
with it an electric field that does not decay as you get far from the wire. This is
a pathology due to the assumed infinite length of the wire.

5.5 Dielectric and magnetic media

So far we considered the action of the field in Minkowski space where the action
was essentially determined by requiring Lorentz invariance. A medium breaks
Lorentz invariance: The dielectric and permeability tensors that characterize a
medium are Lorentz invariant only in the trivial case where both are the identity
tensors. These tensors represent microscopic charge densities and currents in the
medium which may not be directly accessible. I want to outline the underlying
theory.
The homogeneous Maxwell equations express the fact that E and B are
derived from the potential and as such, are oblivious to the: They remain
intact:
∇ · B = 0, Ḃ + ∇ × E = 0
For the inhomogeneous equations we want to encode the (possible unknown)
sources in tw fields which called polarization P and magnetization M. They are
defined as the solutions of equations which are formally like Gauss and Ampere
law up to funny normalization6 :

ρ = −∇ · P, J = ∇ × M − Ṗ

As these are supposed to represent the local sources in the material, we impose
the boundary condition

P(x) = M(x) = 0, x ∈ {outside body}

(We allow for currents and surface charges.) Define the auxiliary fields D and
H

D = E + 4πP, H = B − 4πM

They satisfy the homogeneous Gauss and Ampere equations:

∇ · D = 0, −Ḋ + ∇ × H = 0

So far, we have not done anything: We made no approximation and no assump-

tion. We have just recast the source terms in terms of the polarization and
magnetization field and introduced the auxiliary fields D and H.

Exercise 5.8. Show that the if you treat Gauss law as a constraint, Ampere’s
evolution law respects the constraint.
6 The choice of normalization is unit dependent. The choice made here is standard in cgs.
82 CHAPTER 5. MAXWELL EQUATIONS

The sources respond to the fields and vice versa. If, in the absence of ex-
ternal driving fields, the body in question is in equilibrium, one would expect a
linear relation between cause and effect. Moreover, if we are only interested in
macroscopic length and memoryless response, the relation should be local:
Dj = εjk Ek , B j = µjk Hk (5.16)
This replace the unknown microscopic sources by material specific functions ε
and µ. For a homogeneous system, these are constant matrices. For isotropic
bodies, they are material specific constant numbers. Eq. 5.16 is known as con-
stitutive relation.
Exercise 5.9. Derive Maxwell’s equation from the variation of the action
Z
1
S= dΩ(Dj Ej − B j Hj )
16πc
subject to the constitutive relations.

Figure 5.6: Sphere with constant polarization

Exercise 5.10. Explain why one might expect ε to be a positive matrix (i.e.
real symmetric with positive eigenvalues) and why there is no similar expectation
from µ.
Exercise 5.11. Show that the charge distribution of a homogeneously polarized
sphere of radius a is concentrated on the surface with surface density
P · x̂ δ(|x| − a)
Exercise 5.12. Show that the current distribution of a homogeneously magne-
tized sphere of radius a is concentrated on the surface with surface density
M × x̂ δ(|x| − a)
5.6. NEW PHYSICS 83

5.6 New Physics

Lagrangian field theory is a framework that allows one to repackage existing
theories such as Maxwell’s electrodynamics in an elegant formalism, but perhaps
more importantly, allows to explore new theoretical models. Most models are,
at the end, just models, and their best application is as questions in homework
sets and exams. But occasionally, some turn out to capture new physics. Here
are three examples.

5.6.1 The quantum Hall effect and Chern-Simon action

The Integer Quantum Hall effect, was discovered by Klaus von Klitzing in 1980.
This discovery ushered a new era of research now called the study of topological
phases. These phases are are intrinsically quantum and labeled by integers7 .
In the case of the integer quantum Hall effect the discovery of phases labeled
a quantized value of conductance. In two dimensions that conductance is, in
general, a matrix. If, in addition, the system is isotropic the conductance matrix
is necessarily of the form
 
σ σH
J= E
−σH σ

Exercise 5.13. Show this

The diagonal part is the dissipative conductance and the off-diagonal is the
Hall conductance. von Klitzing found that in certain two dimensional systems,
at sufficiently low temperatures, and with sufficiently strong magnetic field, the
system is characterized by non-dissipative topological phases where

e2
σ = 0, σH ∈ Z
h
Planck constant is an indication that the phenomenon is quantum.
A field theory that encapsulates the Hall effect relies on the Chern Simon
action. In 2+1 dimensions we can construct a the scalar F ∗ · A. The Chern
Simon Lagrangian density:

νσ0 νσ0 e2
LCS = − (F ∗ )α Aα = − 2 εαβγ Fβγ Aα , σ0 =
2c2 4c h
α runs over 0, 1, 2. ν is a dimensionless number h is Planck constant. σ0 is
the quantum unit of conductance with dimensions of velocity (in c.g.s) . So
the prefactor the same dimensions as in the corresponding Maxwell term. This
guarantees that the action has dimensions [Energy × time]. Planck constant
naturally appears since the origin of the term is quantum.

Exercise 5.14. Show that F ∗ = (B, −E2 , E1 ).

7A rational number is a pair of integers.
84 CHAPTER 5. MAXWELL EQUATIONS

Figure 5.7: The phase diagram for the Integer quantum Hall effect for the Hofs-
tadter model on the triangular lattice at T = 0. The vertical axis is the magnetic
flux through the unit cell. The horizontal axis is the chemical potential. Figure
made by Gal Yehoshua for an undergrad project.

Exercise 5.15. Explain why LCS has the right dimensions.

You may worry that the Chern Simon Lagrangian density is not gauge in-
variant. This causes no problem for the variation of the action since:
Exercise 5.16. Show that under a change of gauge Aµ → Aµ +∂µ Λ, the Chern-
Simon Lagrangian density LCS changes by a boundary term.
The variation of the action is

δ(F ∗ A) = εαβγ (δ∂β Aγ )Aα + (F ∗ )α δAα

= −εαβγ (δAγ )(∂β Aα ) + (F ∗ )α δAα + ∂β (. . . )
= 2(F ∗ )α δAα + ∂β (. . . )

As a model for the quantum Hall effect take the Lagrangian density
1
LCS + Aα j α
c2
The corresponding Euler-Lagrange equations are
νσ0 ∗ α 1
− (F ) + 2 j α = 0 =⇒ νσ0 F ∗ = j
c2 c
Unlike Maxwell’s equations, this is not a differential equation, but an algebraic
relation between the fields and sources. In components

νσ0 B = cρ, −νσ0 E2 = j1 νσ0 E1 = j2

We have reproduced some of the known features of the Hall effect, namely
• Ohms law: The current is proportional to the fields
5.6. NEW PHYSICS 85

• There is no dissipation

• The conductivity tensor is isotropic

• The magnetic field determines the charge density

To explain why ν must be quantized one needs to input quantum mechanics,

which is good, but also somewhat artificial assumptions about the topology of
space time. Here is a sketch.
In quantum mechanics one is not looking at the extremal actions, but rather
one allows the system to explore all path configurations, weighted by

eiS/~

Gauge invariance then means that under a gauge transfromation Λ

Z
νσ0
F ∗ ΛdS α = 0 M od (2π~)
c2 bdry α

Now comes a shady trick. We assume that time is a circle we take for Λ the
transformation corresponding to unit of emf-action: A jump in Λ across the
time cut (also, the quantum unit of flux)

hc
∆Λ =
e
The quantization condition is then reduced to spatial integration
Z Z
νσ0 νσ0
F0∗ dS 0 = B dxdy = 0 M od 1
ce ce
Now, by an argument of Dirac, if space is a close manifold, say a torus then the
total flux is quantized
Z
hc
B dxdy = m , m ∈ Z
e
We finally get
h
νm σ0 = mν = 0 M od 1
e2
which quantizes ν to a rational number.

5.6.2 Axion electrodynamics

In 3+1 dimensions F ∗ · F is a boundary term, and as such it does not affect
the equations of motion. However, this terms can do something interesting if
its coupling constant is replaced by a function. The function is called the Axion
field φ(x):
1 σ0
L=− F ·F − φ(x)F ∗ · F
16πc 8π
86 CHAPTER 5. MAXWELL EQUATIONS

Evidently, this Lagrangian is gauge invariant8 . Since

φF ∗ · F = 2φ (F ∗ )µν ∂µ An = 2φ∂µ (F ∗ )µν Aν = −2(∂µ φ) (F ∗ )µν Aν + ∂µ (. . . )

we can then replace it by (a formally gauge dependent Lagrangian density)

1 σ0
L=− F ·F − (∂µ φ)(F ∗ )µν Aν
16πc 4π
Up to boundary terms, the variation of the action is

4πc δL = (∂µ F µν ) δAν + α (∂µ φ)(F ∗ )µν δAν ,

2
e
Exercise 5.17. Determine α (Answer:α = −3 hc )
The Euler-Lagrange equations are

(∂µ F µν ) = −α (∂µ φ)(F ∗ )µν

It is instructive to write the equations in terms of E and B. Gauss law (without

external sources) now takes the form

∇ · E = α(∇φ) · B

Ampere law
∂0 E − ∇ × B = αφ̇B + α∇φ × E
Exercise 5.18. Verify.
When φ is a constant one recovers the sourceless Maxwell equations. In
general, ∂µ φ acts like a source term in Maxwell equations.

5.6.3 Quantum interface

Axion electrodynamics started as a speculative model of an elementary particle:
The Axion. A different perspective was taken by Qi et. al. who proposed looking
at the interface between topologically distinct quantum phases. In the bulk of
the two insulators Maxwell theory applies. This says that φ is constant in
each. The constant is quantized to be 0 or π, by an gauge invariance argument
similar to the one in CS theory of the quantum Hall effect. By definition, the
two insulators are topologically distinct if the constant is different in each. This
means that ∇φ = δ (2) (x)n. Gauss law is replaced by

∇ · E = αδ (2 (x)n · B

The magnetic field on the surface acts as if there was a charge on the interface.
This is something we have already encountered in the CS theory of the quantum
Hall effect.
8 Since E is even and B odd under time reversal the Lagrangian breaks time-reversal unless

φ is also odd under time reversal. The notion of time reversal in the quantum case is subtle.
5.6. NEW PHYSICS 87

Φ=0

Φ=Π

Figure 5.8: The interface between two insulators that are topologically distinct
gives rise to a singular Axion field.

Ampere law is replaced by

∂0 E − ∇ × B = αδ (2) (x)n × E

This means that electric field on the surface acts as if there were currents at the
interface. In particular in Axion electro-Magneto-statics

∇ · E = αδ (2 (x)n · B, −∇ × B = αδ (2) (x)n × E

B
E

Figure 5.9: A cylinder of a (non-trivial) topological insulator is immersed in

vacuum (trivial insulator). A uniform electric field E in the axial direction
leads to response in a magnetic field inside the cylinder
88 CHAPTER 5. MAXWELL EQUATIONS

5.6.4 Magnetic response to an electric field

Consider a non-trivial insulator in the form of an infinitely long cylinder of
radius a immersed in the trivial vacuum and z-oriented. Take uniform electric
field everywhere and uniform magnetic field inside the cylinder
E = E0 ẑ, B = B0 θ(a − ρ) ẑ
Clearly
∇ · E = 0, n·B=0
so Axion Gauss law is satisfied.
∇ × B vanishes inside the cylinder and outside the cylinder, but has a delta
jump on the boundary. The magnetic field in the wire is proportional to the
constant electric field:
Exercise 5.19. Use Stokes theorem for the (blue) rectangle shown in the figure
to show that
B0 = αE0

5.6.5 Phantom monopoles

In electrostatics an electric charge near a (grounded) conductor has an oppo-
sitely charged image. I want to show that in Axion electrodynamics you can
create image which is a magnetic monopole. We want to find consistent solution

E, B
Φ=0

Φ=Π
E=0

Figure 5.10: A electric charge, (red dot) is placed near a different topological
insulator with zero fields. On the left, the physical setup. On the right the
image method.

as if there was a magnetic monopole in the lower half space. Namely

x + dẑ
B(x) = g θ(z) + (yet unknown f unction)θ(−z)
|x + dẑ|3
5.6. NEW PHYSICS 89

Exercise 5.20. Explain why ∇ · B = ∇ × B = 0 in the half space z > 0

The magnetic provides a source term for the electric field. The source term
is precisely the same as the source term in the corresponding electrostatic image
charge problem provided
gα = 2e
Now, if we add to this field the electric field given by electric monopole of charge
e above the x-y plane we obtain the same electric field configuration as in the
electrostatic image charge problem, everywhere, i.e.
 
x − dẑ x + dẑ
E=e − θ(z)
|x − dẑ|3 |x + dẑ|3

This describes the electric field everywhere. It remains to see what values B
takes in the lower half-space. Now E · n = 0 on the boundary and so we see
that B is the solution of
∇×B=∇·B=0
everywhere subject to the boundary condition that fixes B on the plane z = 0.
We introduce a scalar potential for B in the lower half pace

B = ∇φ, ∆φ = 0

subject to the boundary condition ∇φ = B on the plane z = 0. Evidently

g d
φ(x, y, z = 0) = p , Bz = ∂z φ = g
x2 + y2 + d2 |x2 + y 2 + d2 |3/2

The problem then reduces to solving Laplace equation with two types of bound-
ary conditions.

E B

Image Image

Figure 5.11: The red curve shows the surface charge density that allows the
field to terminate at the surface. On the right one sees the response in the form
of a magnetic field that seems to have a magnetic monopole at the image point.
There is no real magnetic monopole anywhere, of course.
90 CHAPTER 5. MAXWELL EQUATIONS

5.6.6 Electrodynamics in 1+1 dimensions

It is not possible to explicitly solve Maxwell’s equation for the fields for general
motion of the sources. However, in 1+1 dimensions this is possible.
In 1+1 dimensions F is an anti-symmetric 2 × 2 matrix. Its single entry is
the electric field, which is a Lorentz scalar.
Exercise 5.21 (Solution of Maxwell equations for arbitrary motion of a point
charge). Show that for a charged particle with a given, arbitrary, orbit, the
solution of Maxwell equations for E takes two constant values in the space-time
plane separated by the world line of the particle. Determine the jump across the
world line.

-1

-2
-2 -1 0 1 2

Figure 5.12: Maxwell’s equation in 1+1 dimensions can be solved geometrically

for arbitrary motion of the source. The figure illustrates the solution for two
point sources undergoing constant acceleration. The field takes constant values
in the different regions delineated by the orbits of the charges.

In Maxwell’s theory the source term is a the vector field of currents j µ . In

1+1 dimensions there is a different option for a source term, namely a scalar
field φ(x) 9 :
1 1
L = − Fµν F µν + φ(x)εµν Fµν
4 2
This looks first like a different theory, but it is actually equivalent to Maxwell’s.
The variation of A gives, up to boundary terms,
δL = (∂µ F µν )δAν − ∂µ φ(x)εµν δAν
The Euler-Lagrange equations for this model are then
∂µ F µν = j ν , j ν = (∂µ φ)εµν (5.17)
9 Frank Wilczek, who invented this field, called it Axion field.
5.6. NEW PHYSICS 91

Note that ∂ν j µ = 0 so the current is conserved. We have recovered Maxwell

theory except that the current is interpreted as the gradient of a scalar.
Bibliography Xiao-Liang Qi, et al, Inducing a Magnetic Monopole with
Topological Surface States, Science 323, 1184 (2009);
92 CHAPTER 5. MAXWELL EQUATIONS
Chapter 6

Conservation laws and the

Stress-Energy tensor

The energy, momentum and angular momentum of the electromagnetic field are
identified as the conservation laws associated with symmetries of space-time.
Maxwell stress tensor is derived from variation of action due to variations of
the metric.

6.1 Maxwell stress energy tensor

Our aim here is to identify the energy, momentum, and angular momentum of
the electromagnetic field. They express conservation laws. The local form of
the conservation law is
∂µ T µν = 0 (6.1)

where T is known as Maxwell stress-energy tensor

1  αµ β 
T αβ = F F µ − 41 g αβ F · F (6.2)
4π

T is chosen so that its physical dimension is [energy density], or, equivalently

[pressure]. Explicitly, T is
 
0 B×E
E2 + B2
1
 
4πT = − 
 B×E

2 |{z} Ei Ej + Bi Bj 
relative sign

In the first row and columns you recognize the energy density and Poynting
vector. This explains the 4π normalization. We shall discuss the other terms
below.

93
94CHAPTER 6. CONSERVATION LAWS AND THE STRESS-ENERGY TENSOR

6.1.1 Symmetric and traceless

The tensor is symmetric and traceless. Symmetry is obvious from the definition,
as g is symmetric. Traceless follows from the fact that g α α = 4 and so
1  αµ 
T αα = F Fαµ − 14 g α α F µν Fµν = 0
4π

ct

Figure 6.1: The green rectangle represent the field. The red box is a box in
space-time. The symmetries of Minkowski space allow to shift (and rotate the
box) and fields without affecting the action. For infinitesimal shifts of physical
fields, this triviality statement translates to a conservation law.

6.1.2 Conservation laws

Consider a space-time box as in Fig. 6.1. Suppose the spatial box is large
enough to embrace all the fields at any given time. Then
Z Z Z  t2 Z t2
0 = dΩ ∂µ T µν = dSµ T µν = dS0 T 0ν  = dV T 0ν 
 
t1 t1

It follows that the 4 energy momentum vector

Z
dV T 0ν = (E, cP j )
6.1. MAXWELL STRESS ENERGY TENSOR 95

is conserved. This identifies the first row (and column) with the energy (mo-
mentum) densities.

6.1.3 Nöther currents

Eq. 6.1 is a consequence of the (free) Maxwell equations. However, this alone
does not offer insight.

Exercise 6.1. Show that Eq.6.1 follows from Maxwell’s

The missing insight is that conservation laws expresses the homogeneity of

space time, a relation due to Nöther. If you compute the action of a given field
configuration in a box, you get the same number if you translate (and rotate)
the coordinates. This is a trivial statement, following from the homogeneity
of Minkowski space-time. and you would not expect to get any interesting
identities from it. The genius of Nöther was to realize that if the field is not an
arbitrary field configuration, but one that solves the Euler Lagrange equations,
then the statement reduces to statement about the fields on the boundaries of
the box. This is what you expect form a conservation law: What comes in
through one boundary must leave through another.
To do that Nöther introduced a technical step which is amusing: She split
the real operation of coordinate shift into two virtual operations: One that shifts
the box–but not the fields–and one that shifts the fields–but not the box. Think
of moving your bag in two steps: First you move the content of the bag and
second you move the empty bag. For infinitesimal shifts, for fields that satisfy
Euler-Lagrange equations, these contributions live on the boundary.

6.1.4 Shifting the field

Consider the change in action due to a shift of the field without a shift of the
box. The variation of the action SF due to arbitrary variation δAµ of the fields
in a a fixed box, has been computed in Eq. 5.9 and was found to be
Z Z
4πc (δSF ) = − dΩ ∂µ (F µν δAν ) + dΩ(∂µ F µν )δAν
| {z } | {z }
bdry term E−L

(6.3)

It follows that for fields that satisfy the Euler-Lagrane equation, the variation
is a boundary term
Z
4πc (δSF ) = − dΩ ∂µ (F µν δAν ) (6.4)

A uniform space-time shift by δξ α leads to variation in the fields

δAµ (x) = −(∂α Aµ )δξ α (6.5)

96CHAPTER 6. CONSERVATION LAWS AND THE STRESS-ENERGY TENSOR

t t t

x x x

Figure 6.2: The action remains the same when the fields and the integration
box are both shifted in space-time. For a small shift the change in action can be
split into two virtual shifts: A shift of the field with the box held fixed, shown
in the middle figure, and a shift of the box with the field held fixed field, shown
on the right.

Exercise 6.2 (Signs-Sigh). Explain the minus sign.

Inserting the variation into Eq. (6.3) and using Maxwell equation1 one finds
for the integrand
∂µ (F µν δAν ) = ∂µ (F µν ∂α Aν )δξ α
= ∂µ F µν Fαν δξ α + ∂µ F µν ∂ν Aα δξ α

= ∂µ F µν Fαν δξ α + ∂µν |{z}

F µν Aα δξ α



|{z}
symm anti
µν α


= ∂µ F Fαν δξ
The change in action has been reduced to a boundary term
Z Z
µν

α
4πc (δSF ) = dΩ ∂µ F Fαν δξ = dSµ F µν Fαν δξ α (6.6)
|{z}
bdry

6.1.5 Shifting the box

As a warmup consider the variation of a one dimensional integral upon shifting
boundary points by δξ
Z b !


δ f (x)dx = δξ f (b) − f (a)
a

1 Note that ξ α is a constant, not a function.

6.1. MAXWELL STRESS ENERGY TENSOR 97

The case at hand is the multidimensional version of this.

Shifting the box without shifting the fields changes the action by boundary
terms of the form
Z 
4πc (δSF ) = − 41 dSα F · F δξ α
Z
= − 41 dΩ ∂α (F · F ) δξ α

6.1.6 Joint box and field shift

For the joint shift we get
Z  
0 = dΩ ∂µ F µν Fαν − 14 ∂α (F · F ) δξ α

Z  
= dΩ ∂µ F µν Fαν − 14 gαµ (F · F ) δξ α
Z  
= dΩ ∂µ T µ α δξ α

Since this is supposed to hold for any (infinitesimal) box and any shift, we get
the conservation law.

6.1.7 The momentum of the field

Conservation of momentum reflects the homogeneity of space. A spatial shift
and a large box is no shift at all. So that we only need to consider Eq. (6.6).
Call the contribution from the surfaces at the final time Pj . The contribution
from the initial time must then be −Pj , since the two cancel each other. We
are still free to choose the normalization, and we choose
Z
4π
Pj = dS0 F 0k Fjk
c
Z
= dS0 εijk Ek Bi

which we recognize as the j-th component of the Poynting vector

Z
c
P= dV E × B (6.7)
4π
Exercise 6.3. Suppose ∇ × E = 0 and B = ∇ × A show that

E × B = ∇(E · A) − (E · ∇)A − (A · ∇)E

Show that in the Coulomb gauge ∇ · A = 0

∇ × (E × A) = −A(∇ · E) + (A · ∇)E − (E · ∇)A

98CHAPTER 6. CONSERVATION LAWS AND THE STRESS-ENERGY TENSOR

Exercise 6.4 (Courtesy of O. Kenneth). A charged point particle moves in a

magnetic field B. Show that, if A is the vector potential in the Coulomb gauge,
∇ · A = 0. Show that adding the field momentum to the particle kinetic mo-
mentum mv gives the minimal coupling expression for the particle momentum
in the field Z
1 e
mv + E × B = mv + A
8π c
where B = ∇ × A

6.1.8 The energy of a field

Energy is the conserved quantity associated to time translation. Since the box
is finite in the time direction, need to pay attention to the fact that the box is
shifted and the variation the action, and we need also to take care of Eq. (??).
Again, for a large spatial box only the dS0 surfaces contribute and combining
Eq. 6.6 and ?? we get the conserved quantity2
Z  
4πE = − dV F 0j F0j − 14 Fµν F µν
Z  
= dV E · E − 12 (E · E − B · B)
Z  
1
= 2 dV E · E + B · B)

We then identify the energy of the field as

Z
1
dV E2 + B2


E= (6.8)
8π
The integrand is the energy density that you recognize.
Example 6.5. The magnetic field of earth is about 1 [Gauss]. The pressure
associates with magnetic field of 100 [Gauss] is the same as .5 [cm] of water.
In contrast, the pressure associated with 10 [KV olt/Cm] is 1 [mm] of water..
Exercise 6.6. Calculate the energy in that lies outside a ball of radius r0 of
charge e placed in its center. What should be the radius of the electron if its
mass had purely electromagnetic origin? (Answer: Energy: E = 4πe2 /r0 . The
4πe2 −15
classical radius of the electron is r0 = m 2 = 2. × 10 [m]).
ec

6.1.9 Angular momentum

The angular momentum of the field is the conserved quantity
Z
1
J = dV x × (E × B) (6.9)
4πc
2 Conserved quantities are fixed up to normalization. Here we picked different normaliza-

tions for the energy and the momentum so each has the appropriate dimensions. A more
convincing normalization procedure will be given when we discuss Maxwell stress tensor.
6.2. T AND VARIATION OF THE METRIC 99

This can be seen ass follows. Under infinitesimal spatial rotation by δθ (a vector)
the coordinates transform like
x0 → x0 , xj → xj + Rkj xk , Rkj = εj km (δθ)m
We are interested in the variation of δAn (x). The rotation acts both on the
coordinates xm and also on the vector Am . Together, we have
δAn = Rj n Aj + (Rj k xk )(∂j An )
= Rj n Aj + (Rj k xk )(Fjn + ∂n Aj )
= Rj k xk Fjn + Rj n Aj + Rj k ∂n (xk Aj ) − Rj k δnk Aj
= Rj k xk Fjn + Rj k ∂n (xk Aj )
By Eq. (6.6) the conservation law is determined by
F 0n δAn = F 0n Rj k xk Fjn + Rj k ∂n (xk Aj )

= Rj k xk F 0n Fjn + Rj k ∂n (xk F 0n Aj )
| {z }
bdry term

and we have used Gauss law in the last step. Now

 
Rkj xk F 0n Fjn = x × (E × B) · δθ

from which the result follows.

Exercise 6.7. Conservation of angular momentum is easier to see in cylindrical
coordinates. Can you do that.
Exercise 6.8. Generalize the result to Lorentz transformations.

6.2 T and variation of the metric

6.2.1 Blah
In the previous section we got the Maxwell tensor from conservation laws and
symmetries: Deformations that did not affect the metric. It is interesting that
one can also recover the tensor doing the complementary thing: Looking at how
the action varies when we make a deformation that changes the metric. What
should we expect?
When Maxwell constructed his theory the queen of science was, of course,
mechanics. In particular, he understood well elasticity and fluid mechanics. In
elasticity theory the concepts of stress and strain are important, and it was
natural for Maxwell to ask what is their analogs in electrodynamics. One can
think of a strain as a deformation of the metric. For example, the strain shown
in the figure 3 can be represented by deformation of the Euclidean metric
   
1 0 1+ 0
g= →g=
0 1 0 1−
3 The vector field is divergence-less and curl-free.
100CHAPTER 6. CONSERVATION LAWS AND THE STRESS-ENERGY TENSOR

Figure 6.3: In the theory of elasticity a strain is described by a vector field.

The figure shows the vector field associated with (uniform) contraction of y and
dilation of x: Namely (x, −y). The strain causes stress in the material. Energy
is stored in it like in a compressed spring.

6.2.2 Variation of the metric in mechanics

dS

Figure 6.4: The momentum flux is a second rank tensor made from teh two
vectors: Momentum and velocity.

The Lagrangian of a free classical particle is

m ∂L
L= gij q̇ i q̇ j , pj = = mgjk q̇ k = mq̇j
2 ∂ q̇ j
6.2. T AND VARIATION OF THE METRIC 101

The variation of the metric gives

∂L
= 1
2 mq̇ i q̇ j = 12 T ij
∂gij
The symmetric tensor
T ij = pi q̇ j
|{z} |{z}
momentum velocity

is interpreted as the momentum flux.

6.2.3 Variation of the metric

The action depends on the metric in two places. First in the volume element.
So let us reorganize
p p
dΩ
|{z} F · F = |g|dx0 dV · F}
|F {z = dx 0
| {zdV} |g|F · F
| {z }
volume element Lorentz scalar metric indep Lorentz density

The metric is also buried in the scalar product

F · F = F αβ gβγ F γδ gδα (6.10)

Hence
p p p
|g|F · F = δ( |g|) F · F + |g|F αβ F γδ δ(gβγ gδα )


δ
p p
= δ( |g|) F · F + 2 |g|F αβ F γδ gδα δgβγ
p p
= δ( |g|) F · F + 2 |g|F αβ F γ α δgβγ

6.2.4 Matrix calculus

To compute the variation of det g we need tools from matrix calculus. Being
symmetric g can be diagonalized
X
g= Pµ γµ , Pµ = |µi hµ|

where γµ are its (real) eigenvalues and Pµ are orthogonal projections. By defi-
nition Y X
det g = γµ =⇒ log |g| = log γµ
and so
p X δγµ
δ log det g = 21 δ log(|g|) = 1
2 γµ
We want to express the right hand side in terms of g and its variationδg. To do
that observe that, by the functional calculus of operators,
X Pµ
g −1 =
γµ
102CHAPTER 6. CONSERVATION LAWS AND THE STRESS-ENERGY TENSOR

and so
X Pµ (Pν δγν + γν δPν )
g −1 δg =
γµ
Exercise 6.9 (Projections). Show that if Pµ are orthogonal projections, Pµ Pν =
δµν Pν , follows that
T r(Pµ δPν ) = 0

We have then showed that

p
δ log − det g = 21 T r(g −1 δg) = 12 g γβ δgβγ

and so finally p p p p
δ |g| = |g| δ log |g| = 12 |g|g γβ δgβγ (6.11)

6.2.5 The stress tensor

Collecting terms we get
p
∂( |g|F · F ) p  βα γ 
= |g| F F α − 14 g βγ F µν Fµν (6.12)
∂gβγ
√
We now shift back g into the volume element and recover the energy momen-
tum tensor.

6.3 Energy momentum conservation

The most interesting property of T is that, in the presence of sources, it expresses
the fact that the particles provide the source of energy and momentum to the
fields. This is the content of the following elegant formula4
1
∂α T βα = jα F βα (6.13)
c
Before discussing its content, let us comment on the derivation. Normally,
identities have straightforward derivation. This one is tricky. The first step is
easy enough
1  
∂α T αβ = ∂α F αµ F β µ + F αµ ∂α F β µ − 21 g αβ F µν ∂α Fµν
4π
1 1  αµ 
= jµF β µ + F ∂α F β µ − 21 g αβ F µν ∂α Fµν
c 4π
The tricky part is to show that the brackets vanish. To see that note first that
the homogeneous Maxwell equations implies

0 = ∂γ (F ∗ )γν = − 12 ενγαβ ∂γ Fαβ =⇒ ∂α Fβγ + ∂β Fγα + ∂γ Fαβ = 0 (6.14)

4 In curvilinear coordinates replace the partial derivative with covariant derivative.
6.3. ENERGY MOMENTUM CONSERVATION 103

Using this we can now manipulate (twice the covariant β components) of the
vector in the brackets
2F νµ ∂ν Fβµ − F µν ∂β Fµν
 
= −F µν 2∂ν Fβµ + ∂β Fµν

= −F µν ∂ν Fβµ + ∂β Fµν +∂ν Fβµ


| {z }
use hom

µν


= −F − ∂µ Fνβ + ∂ν Fβµ

= −F µν ∂µ Fβν + ∂ν Fβµ


| {z }
µ−ν symmetric

=0
Exercise 6.10 (Plane waves). Show that the stress tensor for plane electromag-
netic waves
Aµ = aµ eik·x , k µ Aµ = 0, kµ k µ = 0
is
4πT µν = a · a k µ k ν

6.3.1 The source term j · F and conservation of energy

The source term j · F is a 4-vector. Consider first its time-component
jα F 0α = jk F 0k = −j k F0k = J · E
which we identify (up to signs) either as the power (per unit volume) that the
field puts into accelerating the charges or alternatively, the energy that moving
charges impart to the field. The time component of the conservation law gives
 2
E + B2
  
E×B 1
∂t + c∇ · = J·E
8π 4π c
The first term on the left is the (field) energy density. The second term on the
left is the energy current, and the right hand side says that currents in a field
are a source term for the equation: The field energy is not conserved, it can
exchange energy with the particles.
You see that the Poynting vector admits two different interpretations: As
the momentum density in the field adn as the energy current.
The spatial components of the source gives the force density
1 1 1 1
− jα F mα = − j0 F m0 − jk F mk = −ρEm + emkj jk Bj =
c c c  c
1
= − ρE + J × B (6.15)
c m
Since a force causes a rate of momentum change, the rhs determines the rate at
which momentum is transferred from the sources to the field.
104CHAPTER 6. CONSERVATION LAWS AND THE STRESS-ENERGY TENSOR

6.3.2 The interpretation of T jk : Stress

Eq. 6.13 allows us to give an interpretation of the remaining terms in T αβ and
see why the tensor is also called Maxwell stress tensor. To do so, let us look at
an example.
Simple settings are stationary, where both the fields and sources are time
independent. Consider the attraction of two ±e charges. This, by itself, can
not be a steady state since the two charges attract. One needs to apply non-
electromagnetic force to keep the charges at rest. Think about this from the
perspective of conservation of momentum. Bob holds the charge on the left for
time T and Alice holds the charge on the right for time T . Since Bob is applying
a constant force for time T he expects, by Newton, to transfer momentum

e2
δP = −F × T x̂, F =
4d2
to the system. The charge is at rest, so it is not the charge that is absorbing
the momentum. It mus be the field.

Exercise 6.11. Why minus?

Figure 6.5: You need to apply a (non electromagnetic) force to hold two op-
positely charged particles apart. This force can be computed from the surface
integral of T 11 on the relevant surface. The arrow, the force and the surface are
all in the 1 direction.

We now use Eq.6.13 to see where the momentum of the field sits in T jk .
Consider the space-time box associated with the left half space of time duration
T , i.e.
Ω = {t, x|0 < t < T, x1 < 0}
Now consider what Eq. 6.13 implies for this box. Lets focus on the 1 component:
Z Z Z
1 1
∂α T 1α dΩ = jα F 1α dΩ = j0 F 10 dΩ = −F T
Ω c Ω c Ω
6.3. ENERGY MOMENTUM CONSERVATION 105

The right hand side was easy to evaluate since there is a single, stationary a
point charge in the box which only feels the electric field of the other side5 This
is interpreted the momentum Bob transfers to the box.

Exercise 6.12. Check the sign on the right hand side.

Now let us see how the field accommodates the momentum. For this consider
the left hand side of the equation
Z Z
dΩ ∂µ T 1µ = dSµ T 1µ
Ω ∂Ω

The momentum transfer of Bob to the box is expressed as an property of the

fields alone one the boundary of the box. The boundary of the box is the three
dimensional space time

∂Ω = {t, x|x1 = 0, 0 < t < T }

So
Z Z T Z
1µ 1  2 
dSµ T = dt dx2 dx3 T 11 , T 11 = E2 + E32 − E12
∂Ω 0 8π

For the case at hand, E2 = E3 = 0 and E1 6= 0, but this is not crucial. The
point is that we can now identify T 11 with the momentum that is either leaving
or entering the space-time box through its boundary. If we now focus on the
spatial boundary of the box, namely, the plane, we get the interpretation that
T 11 is the force density on the boundary,
Z
F1 = dx2 dx3 T 11

Force per unit area is the standard notion of stress in elasticity. A similar
interpretation applies to the other components by considering more complicated
boxes.

6.3.3 Field lines as rubber bands

Let us look at the signs in

8πT 11 = E22 + E33 − E12 + B22 + B33 − B12

The parallel and perp components come with opposite signs. The stress can be
positive or negative, but the sign has nothing to do with the signs of E.

5ρ is a delta function on the world line of charge.

106CHAPTER 6. CONSERVATION LAWS AND THE STRESS-ENERGY TENSOR

Figure 6.6: You need to apply a force to hold two oppositely charged capacitor
plates apart (red arrows). If the arrow is in the z-direction, T zz < 0. You get
this sign if you think of the electric field lines as rubber band: As if the pressure
is negative.

Figure 6.7: The magnetic field lines of a solenoid. The stress in the radial
direction T ρρ > 0 inside the solenoid. This is because Bz 6= 0 while Bρ ≈ 0.
You get the right sign if you replaced the field lines by rubber bands: Stretched
rubber bands along the z-axis, that fan out in the radial direction, will lead
to a positive pressure in the radial direction and negative pressure in the axial
direction.

6.4 Applications
6.4.1 Radiation pressure
The luminosity of the sun L = 3.6 × 1026 [W ], giving a stream of 1045 pho-
tons/sec. The radial component of Maxwell energy momentum tensor at a
distance R from the sun is then
L
T0r̂ = (6.16)
4πR2 c
Consider a macroscopic (black) particle of radius r that perfectly absorbs radi-
ation. The force on the particle at a distance R from the sun is then
r2
Fradiation = L (6.17)
4R2 c
The gravitational force on a particle with density ρ is
4πρr3 M G
Fgravity = (6.18)
3R2
6.4. APPLICATIONS 107

Where M = 2 × 1033 [gram] and Newton constant G = 6.7 × 10−8 [cgs]. The
ratio of the two is then

Fradiation 3L 0.06 [gr/cm2 ]

= ≈ (6.19)
Fgravity 16πrρcM G ρr

For water ρ = 1 [gm/cm3 ]. For earth, r = 6 × 108 [cm], the ratio is minuscule:
10−10 . However, for very small grains, of radius less than 6×10−2 [cm] ≈ 600 [µ]
radiation dominates.
Radiation pressure cleans the solar neighborhood from fine dust. This could
be a mechanism of transporting viruses from our solar system to distant parts
of the universe6 .

Exercise 6.13 (Comet tails). Can you figure out the shape of a comet tail?
Suppose the tail is associated with a planet in circular non-relativistic orbit.
Hint: Figure out the tail in the rotation frame.

Figure 6.8: A planet encircling a star and the tail of dust it sprays (tail)

6.4.2 Solar sails

The computation above can be applied to solar sails. A sail of area A and width
d can be used to sail away from the sun provided

A ρAdM G
Fradiation = L > Fgravity = (6.20)
4πR2 c R2

Cancelling the similar terms we get

1
L > ρdM G (6.21)
4πc

Up to factors of order unity we get, the same estimate as above. You need very
thin sails to build solar sails.
6 The assumption that the particle is black is not reasonable when the radiation penetrates

a distance comparable to the size.

108CHAPTER 6. CONSERVATION LAWS AND THE STRESS-ENERGY TENSOR

6.4.3 Halbach array

Exercise 6.14 (Paradox). The Halbach array is shown in the fig 6.4.3. Make a
qualitative plot of its field lines. You will find a proper plot here. Since the field
is large and essentially parallel to the array on one side of the array, and small
on the other. You may then wonder if the array is a Baron von Munchhausen:
Properly oriented, it will float in gravitational field. Discuss the Maxwell stress
tensor and resolve this apparent paradox.

Figure 6.9: Halbach array gives a large magnetic field above the array and small
one below it.

Figure 6.10: Halbach field (Wikipedia)

Chapter 7

Poisson equation, Cloaking

Vector fields in 3 dimensions are determined by their vorticity and sources.

The basic techniques for solving such problems is described in the context of
static electric and magnetic fields. Maxwell equation in dielectric media can
be reinterpreted as Maxwell equations in curvilinear coordinates. This is the
basis for cloaking.

7.1 Vector fields in 3D: Source and vorticity

The source ρ and vorticity ω of a vector field V in 3 dimensions are defined by

∇ · V = 4πρ, ∇ × V = 4πω

The vorticity is always sourcesless

∇·ω =0

Radial vector fields are vorticity free. x is vector field with uniform source,
∇ · x = 3.
The converse is also true: The sources ρ and ω with ∇ · ω = 0 determine the
field V. By linearity, we can decompose the problem in to two problems:

V =E+B

where E is irrotational (conservative)

∇ · E = 4πρ, ∇×E=0

and B is sourceless
∇ · B = 0, ∇ × B = 4πω

As we shall see the equations for E and b are solved by the same technique.

109
110 CHAPTER 7. POISSON EQUATION, CLOAKING

Figure 7.1: Left: An irrotational field with a source at the origin. A sourceless
field with vorticity along the z-axis

7.1.1 Static electric fields: Ḃ = 0

The electric field E is determined by Gauss law and Faraday inductions law
∇ · E = 4πρ, ∇ × E + Ḃ = 0
and dot is a derivative with respect to x0 = ct. When Ḃ = 0 the electric field is
conservative and is the gradient of a potential that solves Poisson’s equation
∆φ = −4πρ (7.1)
Remark 7.1 (Time dependent ρ). ρ may be time independent. However, time
dependent ρ would normally entail Ḃ 6= 0

7.1.2 Harmonic functions and Poisson’s equation

Solutions of
∆φ = 0 (7.2)
are called Harmonic functions. A fundamental fact about Harmonic functions
is:
Theorem 7.2 (Harmonic functions). If φ is Harmonic, then φ(x) is the average
value of φ on a sphere which has x as its center.
The theorem is evident in one dimension, because a harmonic function is a
linear function–clearly the average of equidistant neighbors. We shall postpone
the proof in the general case after we assemble some more tools.
Corollary 7.3 (Boundary). On a given set, harmonic functions assume their
maxima and minima of its boundary.
This is also known as Ernshaw’s theorem: A charge can not be trapped by
electrostatic fields alone.
Exercise 7.4. Suppose that locally
Ej (x) = Ej (0) + gjk xk + O(x2 )
Show that ∇ · E = 0 if T r g = 0.
7.1. VECTOR FIELDS IN 3D: SOURCE AND VORTICITY 111

7.1.3 The case of two dimensions

A point in the plane can be described by a the complex number z = x + iy (or,
alternatively, by z̄ = x − iy). A general function is a function of both f (z, z̄)
while analytic functions are functions of z only. The real and imaginary parts
of a complex valued function described a vector field.
Exercise 7.5. Show that
2∂ = 2∂z = ∂x − i∂y , 2∂¯ = 2∂z̄ = ∂x + i∂y ,
Show that the Laplacian is
∆ = 4∂z ∂¯z ,
Using this show that f (z) + g(z̄) for analytic f and g is Harmonic. Show that
grad, div and curl are
¯
∇f =⇒ 2∂f,
∇ · V =⇒ ∂V + ∂¯V̄
∇ × V =⇒ i ∂V − ∂¯V̄

7.1.4 The Green function of the Laplacian

For unit point charge e in 3-dimensions the electric field, by symmetry must be
radial and to satisfy Gauss law must be
 
x x e
∇· = 4πδ(x) =⇒ E = e 3 , φ(x) = (7.3)
|x|3 |x| |x|
A useful identity
1
∆x G(x − y) = δ (3) (x − y), G(x) = − (7.4)
4π|x|
This is aka Poisson kernel, aka the Green function. G is a radial function and
1
G0 (r) = (7.5)
4πr2
is the inverse area of the sphere of radius r.
Remark 7.6 (Operator interpretation). It is useful to interpret this equation
as an operator identity: When interpreted as an integral operator, the delta
function represents the identity. The Laplacian in a differential operator and G
is an integral operator. From this point of view G = ∆−1 : The inverse of the
Laplacian.
By linearity, the solution of Poisson equation for a source term ρ(x) reduces
to an integration: Z
φ(x) = −4π G(x − y)ρ(y)dy, (7.6)

The solution is unique up to Harmonic function.

112 CHAPTER 7. POISSON EQUATION, CLOAKING

Exercise 7.7 (Generalization to arbitrary dimension).

1. Show that in dimension d
1
G0 (r) = , r = |x|
ωd rd−1
where ωd is the area of the d dimensional unit sphere.
2. Compute ωd using the Gaussian integral
Z Z ∞
2 2
dd xe−x = ωd rd−1 dre−r
0

7.1.5 Proof of theorem 9.32

Let G(x) be the Green function of the Laplacian in d-dimensions and φ Har-
monic:

∆ (G(x)φ(x)) = G∆φ + 2∇φ · ∇G + φ∆ (G)

= 2∇φ · ∇G + φ(x) δ(x)

Integrate this identity on a ball at the origin. The last term (on the right) gives
φ(0). Since G is a radial function, the middle term can be written as
Z Z R Z
d−1 0
2∇φ · ∇GdV = ωd r dr G (r) dS · ∇φ
|x|≤R 0 |x|=r
| {z }
0 by Gauss

(φ is Harmonic, the flux through any closed surface of ∇φ vanishes so the integral
on the right vanishes for any r > 0.)
It remains to integrate the term on the left
Z Z
∆ (φG) dV = dS · ∇ (φ(x)G(r))
|x|≤R |x|=R
Z  
= dS · (∇φ)G(r) +φ(x) G0 (r)r̂
|x|=R | {z }
0 by Gauss
Z
= G0 (R) dS · r̂ φ(x)
|x|=R

which is precisely the average of φ over the sphere of radius R.

Exercise 7.8. Analyze the stability of a dipole d in electrostatic fields:
1. Show that its is
E = −d · E

2. Suppose first that d is a fixed vector. Show that E is harmoic

∆E = (d · ∇)∆φ = 0
7.1. VECTOR FIELDS IN 3D: SOURCE AND VORTICITY 113

3. Using the principle of virtual work show that the torque on the dipole is
T=d×E

4. Suppose that at every point x the dipole is co-oriented with the field
d(x) = ± d Ê(x)
and the dipole is placed initially at a point where E(0) = 0 (no sources
near the origin). Show that one of E± = ±d|E(x)| always has a minimizer
at x = 0.
5. What is the form of the minimizer?
6. Can you relate this to the stability of the Levitron?

7.1.6 Uniformly moving charge

For a charge e at rest at the origin
xj
Ej = e , Bj = 0 (7.7)
r3
Everything is time independent so we can compute this at any time we want.
Consider the Lorentz boost
 
C S 0 0
 S C 0 0 
Λ=  0 0 1
, C = cosh φ, S = sinh φ (7.8)
0 
0 0 0 1
where φ is the rapidity connected with the usual γ and β by
γ = cosh φ, β = tanh φ (7.9)
The two Lorentz scalars are
e2
E2 − B2 = , E·B =0 (7.10)
r4
The transformation rules are
F 0 µν (x0 ) = Λµ α Λν β Fαβ (x0 ) = Λµ α Λν β Fαβ (Λx) (7.11)
In the frame where we see a moving charge, everything depends on time. So let
us compute everything at t0 = 0 when the charge is at the origin. We have
x = Cx0 , y = y0 , z = z0 (7.12)
In particular
r2 = C 2 x02 + y 02 + z 02 = C 2 x02 + (C 2 − S 2 )(y 02 + z 02 )
= C 2 r02 − S 2 (y 02 + z 02 ) (7.13)
114 CHAPTER 7. POISSON EQUATION, CLOAKING

Figure 7.2: The vector field of a moving charge with rapidity φ = 1. The field
is manifestly radial but not spherically symmetric.

Let us turn to the fields. Ex does not change

Ex0 = F 0 01 = Λ0 α Λ1 β Fαβ = (Λ0 0 Λ1 1 −Λ0 1 Λ1 0 )F01 = (C 2 −S 2 )Ex = Ex (7.14)
Hence, at t0 = 0
x x0
Ex (x) = e = eγ = Ex0 (x0 ) (7.15)
r3 r3 (r0 )
where r(r0 ) is the ugly expression Eq. (7.13).
For the transverse directions
Ey0 = F 0 02 = Λ0 α Λ2 β Fαβ = Λ0 α Fα2 = CF02 = γEy (7.16)
and so
y y y0
Ey (x) = e , = Ey0 (x0 ) = γe = γe (7.17)
r3 r3 r3 (r0 )
The formula is the same but for different reasons. In one case γ came form the
field transformation and in the other from the coordinates.
It now follows that in both frames the field is radial, because
x Ex (x) E 0 (x) x0
= = x0 = 0 (7.18)
y Ey (x) Ey (x) y
Remark 7.9. This is a bit surprising. One could have argued that since the
field is radial in the rest frame, you may expect it to point in the direction of
the particle at the retarded time, not now.
The total strength of the field is
2 γ2
E 0 = e2 = γ2E2 (7.19)
r4
It is stronger in the frame where the charge is seen moving (computed for the
same event).
7.2. COULOMB GAUGE 115

7.2 Coulomb gauge

The Coulomb gauge is aka the radiation gauge aka the transverse gauge is:

Theorem 7.10 (Coulomb gauge). It is always possible to choose the vector

potential Aµ = (−φ, A) so that

∇ · A = 0, ∆φ = −4πρ

The theorem is a consequence what we have learned about Poisson equation.

Suppose ∇ · A 6= 0. Let Λ be a solution of the Poisson’s equation

∆Λ = ∇ · A

The gauge transformation

A0µ = Aµ − ∂µ Λ
reproduces the the same field F with A0 satisfying the Coulomb gauge condition

∇ · A0 = ∇ · A − ∆Λ = 0

What about φ0 ? It is determined by

E = −∇φ0 − Ȧ0

Taking the divergence of this we see that φ0 as a solution of Poisson’s equation:

− ∆φ0 = ∇ · E = 4πρ (7.20)

Remark 7.11 (Causality). The Coulomb gauge is a-causal: The scalar potential
φ is fixed by the instantaneous charge distribution. You move a charge here and
the potential φ changes immediately everywhere. The fact that the potential
changes faster than light has no use for transferring information because the
fields are still causal.

Remark 7.12 (Free space). When ρ = 0 we may take φ = 0 together with

∇ · A = 0.

7.3 Magnetic fields in the case Ė = 0

The magnetic field is determined by Ampere’s law and “Gauss” law for the
magnetic field
4π
∇ · B = 0, ∇ × B + Ė = J
c
In the case Ė = 0 this reduces to
4π
∇ · B = 0, ∇×B= J (7.21)
c
116 CHAPTER 7. POISSON EQUATION, CLOAKING

Note that Ė = 0 implies ρ̇ = 0 and then ∇ · J = 0, which is a consistency

condition for the Ampere equation in this case.
Since B is sourceless it is given by a vector potential for which we are allowed
to impose a Coulomb gauge condition:
B = ∇ × A, ∇·A=0
Using the identity
∇ × (∇ × A) = −∆A + ∇(∇ · A)
we find that A is a solution of (the vector valued) Poisson’s equation
4π
∆A = − J
c
Remark 7.13 (Consistency). The equation is consistent with the gauge condi-
tion ∇ · A = 0 since ∇ · J = 0.

7.3.1 Biot-Savart law

We have seen that in the case that Ė = 0 the magnetic field is determined by a
Poisson equation for the vector potential. It follows that
Z
1 J(y)
A(x) = dy (7.22)
c |x − y|
This gives gives B as an explicit line integral over the current:
B(x) = ∇ × A(x)
Z  
1 1
= ∇x ×J(y) dy
c |x − y|
| {z }
Coulomb
J(y) × (x − y)
Z
1
= dy
c |x − y|3
This is the Biot-Savart law.
Exercise 7.14 (A straight line of current). Consider a cylindrically symmetric
tube carrying constant current I along the z-axis. Using the cylindrical symmetry
of the problem and the integral version of Ampere’s law show that
ẑ × x
B = 2I
|x|2
Exercise 7.15 (Constant magnetic fields). The vector potential of a constant
magnetic field is a linear vector values function and so of the form
A = a × x + (b · x)c
Show that
B = 2a + b × c, ∇·A=c·b
7.3. MAGNETIC FIELDS IN THE CASE Ė = 0 117

7.3.2 Magnetic dipole

The current associated with thin loop of radius a in the x − y plane carrying
current I is

2J = (−y, x, 0) Iδ(x2 + y 2 − a2 )δ(z)

= 12 (Iẑ × ∇)θ(a2 − x2 − y 2 ) δ(z)

and θ(x) = 1 for x > 0 and 0 otherwise–the standard step function. Now
consider the limit a → 0 and I → ∞ so that that Ia2 is fixed.
Exercise 7.16 (Delta function). Show that

θ(a2 − x2 − y 2 )
lim δ(z) = δ(x)
a→0 πa2
The a → 0 limit represents a point dipole, characterized by a vector
 2 2
π Ia
m= ẑ
c
Ampere equation takes the form

(∇ × B)(x) = 4π(m × ∇)δ(x)

To find B we could plug the source into Biot-Savart. However, this is not much
simpler then retracing the derivation. For A we find
(m × ∇)y δ(y)
Z
A(x) = dy
|x − y|
 
Z  
 δ(y) δ(y) 
= dy (m × ∇)y −(m × ∇)x
 
|x − y| |x − y| 


| {z }
bdry term
Z
δ(y)
= −(m × ∇)x dy
|x − y|
1
= −(m × ∇)
|x|
m×x
=
|x|3
To compute B we need a version of the vector identity

a × (b × c) = b(a · c) − c(a · b)

where a = ∇ is a differential operator,b = m a fixed vector and c = xr−3 a

vector valued function. Reflection shows that the right form is

∇ × (m × c) = m(∇ · c) − (m · ∇)c
118 CHAPTER 7. POISSON EQUATION, CLOAKING

From the solution of the Coulomb problem we know that

∇ · (xr−3 ) = 4πδ(x)

Hence x
B(x) = 4πmδ(x) − (m · ∇)
r3
It follows that the magnetic field of a dipole is

−m + 3(m · x̂)x̂
B= + 4πmδ(x)
|x|3

Exercise 7.17. Verify all steps.

Remark 7.18 (Singularity). The magnetic field has a bad (non-integrable) sin-
gularity at the origin. One way to see this is to consider the total flux through
the origin. Take the plane oriented with m through the dipole. The flux through
such a plane is
=0
z }| {
(m · m)(x · x) − 3((m · x))2
B·m=− + 4πm2 δ(x)
|x|5
(m × x)2
=− + 4πm2 δ(x)
|x|5

The first term has an non-integrable singularity at the origin. At the same time,
we know that the total flux through any surface must be zero

Exercise 7.19 (Vanishing flux). Show that the total flux through any such plane
at distance ε from the origin vanishes.

Figure 7.3: Dipole field

7.3. MAGNETIC FIELDS IN THE CASE Ė = 0 119

7.3.3 Dirac monopole

Dirac monopoles were invented, by Dirac of course, to explain the quantization
of charge: The charge of the proton is exactly minus that of the electron, in
contrast with, say, their their mass ratio which does not look like a simple
fraction. Dirac realized that if there was even a single monopole of magnetic
charge em anywhere in the universe, say, behind Andromeda, then charged
quantization will be a consequence of quantum mechanics: The electric charge
of any quantum particle e will be constrained by
2em e
∈Z
~c
This is Dirac charge quantization. This Dirac quantization can be viewed as a
consequence of the Aharonov-Bohm effect.

Figure 7.4: Dirac string and monopole

We look for a vector potential whose flux is the monopole charge em , and
with nice Coulombic field
x
B = em 3
|x|
There is no smooth A that does it. So imagine we allow A with a singularity
along the negative z axis. (But B is still smoth.) Consider the sphere minus
the south pole. Then, by Stokes,
Z Z
4πem = ∇ × A dS = A · d`
`

Suppose we A is azimuthal, i.e. has only one covariant component Aφ . Then

4πem = Aφ (θ = π)2π =⇒ Aφ (θ = π) = 2em

120 CHAPTER 7. POISSON EQUATION, CLOAKING

You will see the singularity in the normalized coordinate

Aφ 2em
Aφ̂ (θ = π) = √ =
gφφ r sin θ

Exercise 7.20. Show this. (Recall that Aφ eφ = Aφ̂ φ̂, with eφ · eφ = gφφ )

How do we extend this to any θ? Note that on the negative z-axis 2 =

1 − cos θ. This regularizes the potential on the positive real axis and maintains
the total flux through the sphere. Hence

1 − cos θ
A = em φ̂
r sin θ

is nicely behaved along the positive z axis. This singularity is called the Dirac
string. A computation you are asked to do in the next exercise shows that A is
the vector potential of a monopole
x
∇ × A = em
|x|3

Exercise 7.21 (Monopole). Show that the covariant component of the vector
potential is Aφ = g(1 − cos θ). Show that

em
Br = B r = Br̂ =
r2

You may think of the string along the negative z-axis as a flux tube that
brings in the magnetic flux that emanates from the monopole. By the Aharonov-
Bohm effect, such a string is invisible if the flux satisfies Dirac quantization rule.

7.3.4 Application to geometry: Linking number

Suppose you have two loops γ1 and γ2 in space and you want to know if they
link. Imagine that the loop γ1 carries a unit current. Then, if the loop γ2 links
n times the loop γ1 we have
Z
4πn
B(x2 ) · dx2 =
γ2 c

Now plug B from the solution of Poisson’s equation to get

dx2 · (x2 − x1 ) × dx1

Z Z
= 4πn
γ2 γ1 |x2 − x1 |3

If n 6= 0 the loops link. The converse is, however, not always true.
7.4. CLOAKING 121

Figure 7.5: Linking curves with linking number 4, (Wikipedia)

7.4 Cloaking
7.4.1 Maxwell equation in a dielectric medium
In a dielectric medium, the homogeneous equations, Faraday’s law and no-
monoples, are the same as in free space1

∇ · B = 0, Ḃ + ∇ × E = 0, (homogenous)

In the absence of external sources, the inhomogeneous Maxwell equations, Gauss

and Ampere laws, are modified to

∇ · D = 0, Ḋ − ∇ × H = 0 (inhomogeneous)

Dot stands for derivative with respect to x0 and D and H are defined through
the constitutive relations2

D = εE, B = µH

where ε and µ are tensors. In a fixed, narrow, band of frequencies, these tensors
can be viewed as functions of the spatial coordinates alone. This will be assumed
from now on. In free (Euclidean) space

ε=µ=g

where g is the (Euclidean) metric. (Recall that g i j = δji .)

1 E and B then represent averages over a macroscopically small, but microscopically large,

ball.
2 Hopefully no confusion will arise between ε as dielectric constant and ε as Levi-Civita

tensor.
122 CHAPTER 7. POISSON EQUATION, CLOAKING

7.4.2 Maxwell in curvilinear coordinates

Maxwell equations are made from div and curl. These are geometric and have
a relatively simple expression in curvilinear coordinates:

1 √ εijk
∇ · E = √ ∂j ( gE j ), (∇ × E)i = √ ∂j Ek
g g

The homogeneous Maxwell equations in curvilinear coordinates3 , are

√ √
∂j ( gB j ) = ∂j ( g µjk Hk ) = 0,
| {z }
Hk 6=Bk
√ j ijk √ jk ijk
g Ḃ + ε ∂j Ek = gµ Ḣk + ε ∂j Ek = 0 (F araday)

We write the equations in terms of Ej and Hj both with lower indexes. In the
case of vacuum µ = g.
The inhomogeneous equations (without sources) are
√ √
∂j ( gDj ) = ∂j ( g εjm Em ) =0, (Gauss)
| {z }
Em 6=Dm
√ j ijk √
g Ḋ − ε ∂j Hk = g ε Ėm −εijk ∂j Hj =0
jm
(Ampere)
| {z }
Dm 6=Em

This shows that in a dielectric material where the dielectric and permeability
tensors are the same (and time independent) hen we can reinterpret the consti-
tutive relations as metric:

εjm = µkm ⇐⇒ g jm

Now in most materials µ ≈ 1 while ε 6= 1, so this is not a general prescription.

You need to engineer the material to make ε = µ.

7.4.3 Metric and dielectric medium

The geometric description of the constitutive relation is due to Ulf Leonhardt
(Now at Weizmann). This of what we have foudn like this: You have a dielectric
with ε, = µ in a space with metric is g 6= ε. Maxwell equations are precisely
the same as those in a (different) space, with metric g 0 , but no dielectric, i.e
g 0 = ε0 = µ0 , provided
√ ij √ ij p 0 0 ij
g µ = g ε = g (g ) ,

as an identity between functions. This means:

√ ij
√ ij

g 0 (g 0 )ij (x)
p

g µ (x) = g ε (x) = (7.23)
|{z}
not x0
3 In section Tensors for a discussion of div and curl in curvilinear coordinates.
7.4. CLOAKING 123

So far, this is just a reinterpretation of the symbols. It says that whenever

µ = ε you can reinterpret dielectrics media in one space as the geometry of an
empty, but in general different space. This is not cloaking: You got a new space,
possibly curved, and the curvature now replaces the dielectric.
You get cloaking provided the two spaces are geometrically the same, and
the change in g reflects a coordinate change. In this case

∂(x0 )i ∂(x0 )j
(g 0 )ij = g αβ (7.24)
∂xα ∂xβ
This translates a dielectric into empty space in curved coordinates. This is the
basis of cloaking.

Exercise 7.22. Consider the scaling (x0 )j = αxj . What is the value of ε and
µ that behaves like vacuum? (Note that time has not been scaled.).

7.4.4 Dirichlet to Neuman: Calderon problem

Cloaking means that you can not tell what is inside an inaccessible region by
manipulating and measuring in its exterior. Cloaking means that you can not
do CT, MRI, and can’t explore for oil.
Here is a prototype of cloaking. You are given a medium with unknown
ε, and are allowed to manipulate and measure only electrostatic potential and
normal field on the surface. Can you determine ε(x) inside?

Exercise 7.23. Show that if ε(x) is known, then the field on the surface, En (x)
is a linear functional of φ(x) on the surface.

The equation for the potential is

√
gεjm ∂m φ = 0


∂j

(The metric g is known.) This is a tensorial equation, it retains its form under
coordinate change. Consider a coordinate transformation that only affects the
interior of the body. This will not affect anything you can do or measure, , but
it will scamble ε and φ inside the body. It follows that you can not determine
ε from (static) boundary data.

7.4.5 An invisible dielectric ball

Take g to be the usual spherical coordinate metric with diagonal covariant com-
ponents
√
gr,θ,φ = (1, r2 , r2 sin2 θ), g = r2 sin θ
Consider change of the radial coordinate, r, alone
(
r r0 > 1
h(r0 ) = , h(1) = 1, h(0) = 0, h0 (r0 ) > 0
h(r0 ) r0 < 1
124 CHAPTER 7. POISSON EQUATION, CLOAKING

Figure 7.6: Straight lines, geodesics in the Euclidean metric, look like deformed
curves when plotted in the coordinates x0 associated with the deformed metric
g 0 . In cloaking you use this fact, and Fermat idea, to force the light avoid the
hiding place.

h(r0 ) is a nice, monotonic, 1-1. This introduces a dielectric medium inside the
unit ball. We want to find the dielectric media.
Using the covariant version of Eq. 7.24:
2 !
h2 (r0 )

0 0 1 2 0 2 0 2
p
0
(g )r,θ,φ (x ) = , h (r ), h (r ) sin θ , g 0 (x ) = 0 0 sin θ
h (r0 )
0 h (r )

The right hand side of Eq. (7.23) is the diagonal matrix

sin θ
2
(hh0 (r0 ) , 1, sin−2 θ


h0 (r0 )

viewed as a function of (r0 , θ). Replace the argument r0 → r gives the function

sin θ
2
(hh0 (r) , 1, sin−2 θ


0
h (r)

Plugging in Eq. 7.23 gives for the contravariant components of the (diagonal)
tensors !
2
h(r)h0 (r)

r,θ,φ r,θ,φ 1 1 1
µ =ε = 0 , 2, 2 2
h (r) r r r sin θ

The dielectric tensor looks nicer, and may have clearer physical meaning, in
normalized components
2 !
h(r)h0 (r)

1
µr̂,θ̂φ̂ = εr̂,θ̂φ̂ = 0 , 1, 1
h (r) r

h(r) = r for r > 1 one gets µ = ε = 1 outside, as one must. However, inside the
ball we have a non-trivial dielectric medium. Such a medium is invisible.

Exercise 7.24. Explain why the tensor represents and isotropic medium.
7.4. CLOAKING 125

7.4.6 Cloaking
In cloaking you want more than an invisible ball. You want to use its interior to
hide something. This can be achieved with an h that creates a protected cavity
inside ball of radius 1/2.
1+r
r0 = θ(1 − r) + rθ(r − 1)
2

r'
2.0

1.5

1.0

0.5

0.0 0.5 1.0 1.5 2.0

r

Figure 7.7: Coordinate change which creates a hiding cache in a ball of radius
1/2. The origin The surface of the sphere r0 = 12 is mapped to the origin r = 0.
The interior of the sphere is mapped into a different world r < 0. The real world
with a hole r0 > 12 is mapped into a fictitious world without a hole which looks
empty.

h is a piecewise linear function which inside the unit ball is simply

r = h(r0 ) = 2r0 − 1

Plugging in the equation for the dielectric functions we get in the coordinates
of physical space (with the standard spherical coordinates)
 2 !
1 2(2r − 1)
µr̂,θ̂,φ̂ = εr̂,θ̂,φ̂ = , 1, 1 , 1/2 < r < 1
2 r

The radial component of µ = ε vanishes at the boundary of the protected acvity:

r = 1/2. Not the discontinuity on the boundary: The limit from inside the ball
gives ε = 21 (4, 1, 1) while the limit from outside is (1, 1, 1).
Bibliography :
J. B. Pendry et. al .“Controlling electromagnetic fields”, Science 312, (2006)
126 CHAPTER 7. POISSON EQUATION, CLOAKING
Chapter 8

Electromagnetic waves

Maxwell equations in vacuum are reduced to the wave equation for the poten-
tials. We discuss the notions of polarization, Green’s function, retarded and
advanced solutions.

8.1 Electromagnetic waves

In the absence of sources, the inhomogeneous Maxwell equations are:

0 = ∂ µ Fµν = ∂ µ (∂µ Aν − ∂ν Aµ ) (8.1)

In the absence of charges the Coulomb gauge allows us to choose φ = A0 = 0

and ∇ · A = 0. In particular, it says that1

∂ µ Aµ = 0

This is known as the Lorenz gauge condition (see section 9.1.2). Here we derived
it as a special case of the Coulomb gauge in the absence of charges. In fact, one
can always impose the Lorentz gauge condition (even when there are charges),
but the proof that one can do that shall only be given later. It is manifestly
gauge invariant.
In the Lorentz gauge the potentials satisfy the wave equation

1
Aµ = 0,  = ∂ µ ∂µ = − ∂tt + ∆
c2

Remark 8.1 (Lorentz invariance). Since the Dalambertian,  and the Lorenz
gauge conditions are manifestly a Lorentz invariants, Lorentz transformations
of electromagnetic waves are electromagnetic waves.

1 Assuming det |g| = 1

127
128 CHAPTER 8. ELECTROMAGNETIC WAVES

8.1.1 Electric and Magnetic fields

The electric and magnetic fields in vacuum satisfy the wave equations. This
follows from Faraday and Ampere laws:

Ė + ∇ × B = 0, Ḃ − ∇ × E = 0

Substituting one into the other gives

Ë + ∇ × (∇ × E) = 0, B̈ + ∇ × (∇ × B) = 0

The constraints are Gauss laws

∇ · E = 0, ∇·B=0 (8.2)

Combining we get that the electric and magnetic fields satisfy the wave equation

Ë − ∆E = 0, B̈ − ∆B = 0 (8.3)

Waves that satisfy the wave equation, Eq. (8.3), and the divergence-less con-
straint, Eq. (8.2) are called transverse waves.

8.2 Plane waves

Plane waves are (the real part of)

Aµ (x) = aµ eik·x , k · x = kµ xµ (8.4)

with aµ a 4-vector of fixed amplitudes and kµ = (−ω, k) a fixed 4-wave vector.

It follows
Fµν = i(kµ aν − kν aµ )eik·x
Exercise 8.2. Show that F is invariant under

aµ → aµ + λkµ (8.5)

Explain why this is a gauge transformation.

Maxwell equations, Eq. (8.1), reduce to an algebraic equation for k and a
linear equation for aν
(k · k) aν − kν (k · a) = 0 (8.6)
This, together with the Lorentz gauge condition, Eq. (8.5), allows us to impose
that the two terms in Eq.(8.6) separately vanish

k · k = 0, k·a=0

This says that k is a light-like vector and fixes the dispersion relation

ω = ±c|k|
8.2. PLANE WAVES 129

The second condition, k · a = 0, the Lorenz gauge condition, is a transversality

condition on the amplitude.
We are still free to orient the Euclidean frame so that the wave propagates
in the z-direction. The light-like vector kµ and the amplitude aµ are then

kµ = ω(−1, 0, 0, 1), aµ = (a0 , a1 , a2 , −a0 )

| {z }
Lorentz gauge

This is how a plane wave looks in the Lorenz gauge.

There is a remnant gauge freedom that allows us to choose a0 . Using the
fact that k is light-like, k0 6= 0 hence the gauge freedom, Eq. (8.5), allows us to
set a0 = 0. This reduces the Lorenz gauge to the Coulomb gauge:

kµ = ω(−1, 0, 0, 1), aµ = (0, a1 , a2 , 0) (8.7)

| {z }
Coulomb gauge

In the Coulomb gauge, the amplitudes are orthogonal (in Euclidean space) to
the direction of propagation.

8.2.1 Electric and magnetic fields

For plane waves
ω
E = −i A, B = −ik × A
c
Since k · A = 0 this implies that E and B are orthogonal and have equal mag-
nitudes. E, B and k form an orthogonal triad.

Figure 8.1: The triad of E, B, k for a plane wave. The wave propagates in the
k̂ = Ê × B̂ direction
130 CHAPTER 8. ELECTROMAGNETIC WAVES

8.2.2 Doppler
Since k · x is a Lorentz scalar

k · x = k 0 · x0

Lorentz transformation of a plane wave is a still a plane wave. The wave has,
in general, different wave vectors and amplitudes in different frames:

kµ0 = Λµ ν kν , a0µ = Λµ ν aν

Longitudinal Doppler
Consider a plane wave propagating in the z-direction, Eq. (8.7). Boosting the
wave with rapidity φ in the same direction is the same as viewing the wave from
an inertial frame boosted in the opposite direction. The associated Lorentz
transformation is

Λ0 3 = Λ3 0 = − sinh φ, Λ0 0 = Λ3 3 = cosh φ, Λ1 1 = Λ2 2 = 1 (8.8)

The Lorentz transformation of the light-like vector kµ = ω(1, 0, 0, 1) gives kµ0 =

ω 0 (1, 0, 0, 1) where
s
0 φ 1+β
ω = ω(cosh φ + sinh φ) = ωe = ω (8.9)
1−β

This is linear in the velocities for small speeds. a1,2 are not affected by the
boost.

Transverse Doppler
Consider, as before, a wave propagating in the z-direction, but a boost in the
x-direction so that

Λ0 1 = Λ1 0 = − sinh φ, Λ0 0 = Λ1 1 = cosh φ, Λ2 2 = Λ3 3 = 1 (8.10)

The wave vector kµ = ω(1, 0, 0, 1) is transformed to a light like wave vector

kµ0 = ω(cosh φ, − sinh φ, 0, 1) in the x − z plane. The new frequency is

ω 0 = ω cosh φ = ω γ

This is quadratic in the velocities for small speeds.

8.2.3 Particle production: GZK limit

The cosmic microwave background (CMB) provides a shield that screens ultra
high energy cosmic rays: The GZK limit says that protons with energies above
5 × 1013 M eV are screened by the 3◦ K thermal photons of the CMB. As an
exercise let us compute the threshold for pion production.
8.2. PLANE WAVES 131

Let us compute the threshold for particle production. The total energy-
momentum of a proton with rapidity φ and counter-propagating photon in the
plane is
pµ = mP (cosh φ, sinh φ) + ~ω(1, −1) = (p0 , p1 )
The energy in the center of mass frame, Ecm , is the scalar
q q
Ecm = p20 − p11 = m2P + 2mP ~ωeφ = mP + mπ

and the equality on the right expresses The threshold for pion production.

m2π + 2mP mπ = 2mP ~ω eφ

Since mπ  mP one finds a simple formula for the rapidity

mπ
eφ ≈
~ω
The corresponding energy threshold is
mp mπ
mp cosh φ ≈ 12 mp eφ ≈ ≈ 2.5 × 1014 M ev
2 × 3kB

which is factor 5 too large from the estimate given above.

Exercise 8.3. Can you figure out why the estimate is too big?

8.2.4 Laser cooling and optical molasses

Laser cooling is a cool application of the Doppler effect to slow down atoms.
Think of the atom as a two level system with energy gap E. Suppose you point
a laser beam with frequency ~ω < E at the atom. Atoms that move towards
the light source will see bluer light and if they move fast enough, they will be
able to absorb the light. This will slow the atom down. At the same time, slow
atoms will be transparent to the light.
The remnant velocity is the one whose Doppler shift is comparable to the
natural line width Γ of the atomic a level. This gives the velocity cΓ/E. The
corresponding temperature is of the order
 2
Γ
kb T ≈ 12 M c2 (8.11)
E

where M is the mass of the atom. This leads to low temperatures whenerevr
the energy level has long life time so Γ is small. Indeed,
 2
Γ
= O(α6 ) = O(10−12 )
E
132 CHAPTER 8. ELECTROMAGNETIC WAVES

8.3 Polarization
8.3.1 Amplitude and phase
Scalar plane waves are simply characterized by their frequency ω, wave vector
k, amplitude and phase. Electromagnetic waves, being vector valued, are more
complicated. In addition to the amplitude and phase they are also characterized
by their polarization.
The electric field of an electromagnetic wave propagating is the real part of

E0 eiφ , φ = k · x − ωt (8.12)

The amplitude, E0 , is a complex vector in the plane perpendicular to direction

of propagation, k · E0 = 0. E0 has 4 real amplitudes. What is the physical
interpretation of these four amplitudes?

Figure 8.2: Four Stokes parameters describe elliptically polarized light. Three
numbers identify the size of the ellipse, its tilt to the axes, its eccentricity. A
fourth number gives the purity (the coherence) of the light. The plane of the
ellipse is perpendicular to the direction of propagation k.

8.3.2 Polarization
Let x̂ and ŷ denote orthogonal unit vectors in the plane perpendicular to k.
Write √
E0 = E+ z+ + E− z− , 2z± = x̂ ± iŷ
We may formally identify E± with the components of a an (un-normalized)
spin 1/2 namely, |ψi ⇐⇒ (E+ , E− )t . Quantum mechanics provides us with a
canonical procedure for factoring out the normalization and the overall phase
in of a quantum state: The density matrix:
∗
|E+ |2 E+ E−
 
1
ρ = |ψi hψ| = ∗ (8.13)
|E+ |2 + |E− |2 E+ E− |E− |2
8.3. POLARIZATION 133
p
ρ does not care about the overall amplitude, |E+ |2 + |E− |2 , and overall phase
of the wave. Since T rρ = 1 while det ρ = 0 the two eigenvalues of ρ are 1 and
0: ρ is a projection ρ2 = ρ

Figure 8.3: The Poincare sphere associates with every point on the sphere a
polarization. The north and south poles represent right and left circularly po-
larized light and the equator with linearly polarized light.

8.3.3 Poincare sphere

Any 2 × 2 hermitian matrix of unit trace can be written as
 
1 1 1 + s3 s1 + is2
ρ = 2 (1 + s · σ) = (8.14)
2 s1 − is2 1 − s3

where s = (s1 , s2 , s3 ) ∈ R3 and σ = (σ1 , σ2 , σ3 ) is the vector of Pauli matrices:

     
1 0 0 1 0 −i
σ3 = , σ1 = , σ2 = − (8.15)
0 −1 1 0 i 0

From Eq.( 8.14)

4 det ρ = 1 − s · s (8.16)
so ρ is a projection if s is a unit vector ŝ. This allows us to identify state of
polarization with points on the unit sphere.

Exercise 8.4. Show that antipodal points on the Poincare sphere represent
orthogonal polarization in the sense that ρŝ · ρ−ŝ = 0.

8.3.4 Stokes parameters

Since we are oblivious to the overall phase, we may write (E+ , E− ) = (cos χ, eiψ sin χ),
with cos χ ≥ 0, i.e. 0 ≤ χ ≤ π/2. The corresponding points on the unit sphere
134 CHAPTER 8. ELECTROMAGNETIC WAVES

are
s3 = cos2 χ − sin2 χ = cos 2χ,
s1 + is2 = 2e−iψ cos χ sin χ = e−iψ sin 2χ
This makes 2χ and ψ the standard spherical coordinates.

Circular polarization
The north poles correspond to cos χ = 1 so that E0 =⇒ z+ . The electric field
is–up to an overall amplitude and phase–
√
2E = x̂ cos φ − ŷ sin φ, φ = k · z − ωt (8.17)
As φ increases from 0 to 2π the vector E describes a circle in the x-y plane which
is turning clockwise. At a fixed z and as a function of time the field rotates
counter clockwise while propagating: It behaves like a left handed screw and so
is called left circularly polarized.
Exercise 8.5 (South pole). Show that the south pole represent right circular
polarization.

Linear polarization
The equator is s3 = 0 =⇒ cos2 χ = sin2 χ. As the overall
√ phase does not play a
role we may take for the amplitude (e−iψ/2 , eiψ/2 )/ 2. The corresponding E0
is:
E0 = 12 (x̂ + iŷ)e−iψ/2 + 12 (x̂ − iŷ)eiψ/2
= x̂ cos(ψ/2) + ŷ sin(ψ/2)
As φ increases from 0 to 2π this describes a line element in the x-y plane at
angle ψ/2 to the x axis. ψ = 0 corresponds to x̂ polarized wave and ψ = π to
ŷ polarized wave.

8.3.5 Partially polarized light

The discussion so far addressed ideal plane waves where k is sharply defined and
the amplitude E0 is an honest constant. Such a wave is coherent. Many light
sources are incoherent. One way to model incoherence is as a statistical average
X
E0 = pj Ej , pj ≥ 0,

where Ej represent independent (normalized) light sources, namely

h(Ej )a (Ek )b i = 0 j 6= k, ∀a, b
The polarization of such a mixture is naturally defined as the mixture of polar-
izations X
ρ= p j ρj (8.18)
8.3. POLARIZATION 135

Exercise 8.6. Show that

∗

2

 
1
|E∗+ | 
E+ E−
ρ=  (8.19)
h|E+ |2 i + h|E− |2 i E+ E− |E− |2

It is still true that T rρ = 1. But now ρ is not a projection, because det ρ

need not vanish:

2
∗

∗
det ρ |E+ |2 + |E− |2 = |E+ |2 |E− |2 − E+ E−





E+ E− ≥ 0 (8.20)

(Schwartz inequality was used here). By 8.16 this implies that the vector |s| ≤ 1;
the vector lies in the unit ball. The light we get from the sun is completely
unpolarized. It is associated with s = 0, the center of the Poincare ball.

Remark 8.7 (Combing a tennis ball). One amusing, essentially topological,

property of the transverse nature of electromagnetic waves is that it is not pos-
sible to have a fully spherically symmetric electromagnetic wave. The point is
that a spherical wave, with k pointing radially, has E tangent to the sphere. It
is a basic fact in topology that any vector field on the sphere must vanish at (at
least) two points. The field can not be “the same” everywhere.

8.3.6 3D glasses
When you view a 3D movie, the 3D glasses transmit a picture with right circular
polarization to, say, the right eye and left circular polarization to the left eye.

Figure 8.4: An arrangement that transmits right circular polarization.

Exercise 8.8. Can you give an (ergonomic) argument why spectators would
prefer circular to linear polarization?

Exercise 8.9. Define quarter wave plate as the rotation of the Poincare sphere
that turns circular polarization to linear. Show that it is represented by Hadamard
gate H √
2H = σ3 + σ1
136 CHAPTER 8. ELECTROMAGNETIC WAVES

Exercise 8.10. Explain why the filtering associated by linear polarizers can be
described by the projections
1 ± σ1
PH,V =
2
It follows from the two exercises above that the right and left glasses can be
represented by 2 × 2 matrices

g1 = PH H, g 2 = PV H

Exercise 8.11. Give a physical interpretation of the identity

PV σ 3 = σ 3 PH

in terms of rotation of the glasses.

Exercise 8.12. Explain why holding the glasses backwards is represented by
transposition:
gj ⇐⇒ gjt
If you place glass 1 rotated by π/2 behind glass 2 inverted the joint system
is represented by represented by the matrix product σ3 g1 g2t . A computation
gives 0 which means that no light passes through.
Exercise 8.13. Show that there are 64 ways of arranging the pair of glasses.
How many of these let no light through.

8.4 The wave equation

So far, we have discussed plane wave solutions of the electromagnetic wave
equation. It is instructive to study some basic properties of the wave equation
in general.

8.4.1 The wave equation in one dimension

The one dimensional wave equation, for a scalar field φ, in light-cone coordinates
u = x1 − x0 , v = x1 + x0 takes the form

φ = 4∂uv φ = 0

The general solution of which is

φ(u, v) = f (u) + g(v)

with (essentially) arbitrary f and g. f describes a wave rigidly propagating to

the right at speed c and g a wave rigidly propagating to the left at speed c.
The functions f and g are determined by the initial (Cauchy) data

φ0 (x1 , x0 = 0) = f (x) + g(x), φ̇0 (x1 , x0 = 0) = − f 0 (x) − g 0 (x)

8.4. THE WAVE EQUATION 137

Figure 8.5: If the initial data, φ0 and φ̇0 are localized in the red interval, the
solution at later times lives in the cone. This is called the domain of influence
of the initial (red) data.

This allows for reconstructing f and g from the initial data by integration. As
we shall discuss in more detail below, the localization of φ0 and φ̇0 does not
imply that f and g are localized, see Fig. 8.6.
Any solution can be thought of a a linear combination of the solution with
vanishing initial data for φ̇ = 0 and the complementary case, where the initial
data for φ vanish. We are, of course interested in the case where the initial data
are localized bump functions.

Figure 8.6: The initial data is φ0 = 0 and bump function φ̇0 = 2 cosh−2 x shown
on the left. Also shown are f (x) = tanh(x) and g(x) = tanh(−x). On the
right you see the initial data φ̇(0) again, and the wave φ(x) at time 2. This is
supposed to illustrate that wave lingers near the origin forever and the failure
of Huygens principle in one dimension.

Consider first the initial data φ̇0 = 0

φ̇0 = 0 =⇒ f − g = const =⇒ φ0 = f + 12 const = g − 21 const

By assumption, φ0 is a (localized) bump function. It follows that so are 2f +

const and 2g − const. One propagates to the right and the other to the left.
The wave lives inside the light cone and after a while there is no remnant of the
wave in the interval of the initial data. The wave propagated to infinity.
In contrast, when φ0 (x, 0) = 0

φ0 = 0 =⇒ f = −g =⇒ 2g 0 = φ̇0

Now, the localized initial data φ̇0 do not guarantee that f and g are localized
(only that f 0 and g 0 are bump function). See fig. 8.6. As a consequence Huygens
principle fails in one dimension.
138 CHAPTER 8. ELECTROMAGNETIC WAVES

Exercise 8.14. The energy density of a scalar wave is

φ̇2 + (∇φ)2

1. Express the energy density in terms of f and g.

2. Explain why the failure of Huygens principle in one dimension does not
lead to conflict with energy conservation.

8.4.2 Huygens principle

The solution of the wave equation in d dimensions is better behaved than the
solution in one dimension: Huygens principle hold. This means that the value
of the wave at a space time point is fully determined by the intersection of the
backward light cone with the initial data.

Figure 8.7: The domain of dependence of the initial red data. When Huygens
principle holds, the value of the wave at the space time point is determined by
the intersection of the backward light-cone with the initial data. In general, it is
determined by the intersection of the interior of the cone with the initial data.

Explicit integral representation for the initial value problem of the wave
equation in three dimensions, is given e.g. Wikipedia see also Pinchover and
Rubinstein.

8.4.3 Covariant superposition

The wave equation is an algebraic equation in Fourier space. A solution can be
written as Z
1
φ(x) = d4 k δ(k · k) φ̃(k) eik·x (8.21)
(2π)2

with φ̃(k) an arbitrary function of the 4-vector k. Of course, because of the

δ function only the values that the function takes on the light-cone are rele-
vant. This expression for a scalar wave φ is manifestly Lorentz invariant. The
generalization to vector waves Aµ and tensor waves is clear.
It is instructive to split the solution to the forward and backward light cone
we have
φ(k) = θ(−t)φ< (k) + θ(t)φ> (k) (8.22)
8.4. THE WAVE EQUATION 139

We can carry out the time integration to get:

Z
1 dk
φ> (t, x) = φ̃(|k|, k) ei(|k|t−k·x) (8.23)
(2π)2 2|k|

The forward light cone is associated with out going (retarded) waves. Similarly
Z
1 dk
φ< (t, x) = φ̃(−|k|, k) e−i(|k|t+k·x) (8.24)
(2π)2 2|k|

the backward light-cone can be associated with incoming (advanced) waves.

Note that in both cases, Lorentz invariance induces a weight on the three di-
mensional k space.

8.4.4 Waves with Gaussian waists

Exercise 8.15 (Light cone). Show that the wave equation in 3+1 dimensions
can be written as
¯
(∂uv + ∂∂)φ(u, v, z, z̄) = 0, z = x1 + ix2 , z̄ = x1 − ix2

where U, v are light cone coordinates, v = x3 − x0 , u = x3 + x0

One is often interested in narrow pencils of light. Here is an example of a

monochromatic wave with a Gaussian waist

Exercise 8.16 (Gaussian waists). Show that

φ = c(u)e−c(u)zz̄ eiv/λ

solves the wave equation provides

λc2 (u) + ic0 (u) = 0

whose solution is
1
c(u) =
`2 + iλu

8.4.5 Monochromatic waves

Monochromatic waves are solutions of the wave equation whose time dependence
is eiωt . Hence
∆φ = −k0 2 φ (8.25)
In Fourier space the solution is supported on the sphere

k · k = k02 (8.26)

The smallest wave length that such a wave can accommodate is 2π/k0 : The
frequency limis the spatial resolution.
140 CHAPTER 8. ELECTROMAGNETIC WAVES

8.4.6 Evanescent waves

Near a planar boundary between two media one can can sometimes arrange for
monochromatic, evanescent plane wave solution to the wave equation: Plane
waves in the half-space which decay in x and propagate in the z-direction. Namly

e−κx eikz , k 2 = k02 + κ2

The noteworthy fact about this waves is that the frequency k0 , does not limit
anymore the spatial resolution 1/k. Near x = 0 one can find waves with k  k0 .

Example 8.17 (Transversal waves). The notion of transversality for evanescent

waves is different from ordinary plane waves. For the wave

E = E0 e−κx eikz

Gauss law ∇ · E = 0 reduces to

κ(E0 )1 + ik(E0 )3 = 0

which allows (E0 )3 6= 0 for a wave propagating the in z-direction.

8.4.7 The eikonal equation: Geometric optics

8.5 Waves in dielectric media: Birefringence:

In the absence of external sources, the time evolution of the fields in a dielectric
is dictated by Faraday and Ampere laws

Ė + ∇ × B = 0, Ḣ − ∇ × D = 0

subject to the constraints

∇ · D = 0, ∇·B=0 (8.27)

In Fourier space (x, t) ↔ (ω, k) the differential equations reduce to algebraic

equations

ωε−1 D + k × µH = 0, ωH − k × D = 0, k · D = 0, k · µH = 0

where µ and ε are the constitutive relations. We assume that µ, ε are positive
matrices. (Possibly functions of ω.) Substitution gives for D

ω 2 ε−1 D + k × (µk × D) = 0 (8.28)

which can be written as a (generalized) eigenvalue problem for the 3 × 3 sym-

metric matrix

Mjk (ω, k) = ω 2 (ε−1 )jk + εjmn εabk µna km kb

8.6. GREEN’S FUNCTION FOR THE WAVE EQUATION 141

Given k a non-trivial solution for the eigenvector D exists provided det M = 0.

This fixes the dispersion relation ωj2 (k) with j = 0, 1, 2, the three eigenvalues
of Eq. (8.28). One eigenvalue is always trivial since the matrix always 0 as an
eigenvalue, corresponding to the eigenvector k. There are, therefore, in general,
only two non-trivial eigenvalues
Exercise 8.18. Show that that if ε and µ are symmetric, so is M .
In the frame where ε and µ are diagonal

Mjk (ω, k) = ω 2 (ε−1 )j δjk + µn εnjm εnbk km kb

In the special case that µ is a scalar

Mjk (ω, k) = ω 2 (ε−1 )j − µk · k δjk + µkj kk

For k in the principal direction j we get

ω 2 = εj µ k 2
√
The wave propagates at different speeds ej µ along the principal directions of
ε. This is birefringence.

8.6 Green’s function for the wave equation

The Green function, aka fundamental solution, is the solution to the wave equa-
tion with a point source term

G = 4πδ (4) (x) (8.29)

This equation merits some discussion regarding the existence and the uniqueness
of the solution. The issue of existence has to do with the fact that the source
terms is singular: A delta function (a distribution). What regularity properties
shall we require of the solutions? When we studied Poisson’s equation–an elliptic
equation–a singular source had a Greens function that was a pretty regular
function (the Coulomb singularity is integrable). This is a feature of elliptic
equations. We will not have this luxury in the case of the wave equation which
is hyperbolic: We shall have to allow for solutions G that are distributions if we
insist on causality.
The second issue is uniqueness. As usual, we have the freedom of adding
solutions of the free wave equation. We fix the solution by imposing causality:
We shall denote by G> the solutions which is created by the source. This is
also called a retarded solution. Formally, the wave equation with a source also
admits advanced solutions where the wave is fully absorbed by the source (and
solutions of mixed type). We select the retarded solutions because it is causal.
As we shall show below, the retarded Green function in 3+1 dimensions is
δ(|x| − ct)
G3+1
> (x) = 2θ(t)δ(x · x) = , x = (ct, x), (8.30)
|x|
142 CHAPTER 8. ELECTROMAGNETIC WAVES

G> lives on the light-cone–rather than in its interior. This is the Huygens
principle.

Exercise 8.19. Show the second identity (on the right).

x0
@s = 0 6 s=0
@ s<0
@
@
@
@ s>0
s>0 @
@ - x1
@
@
@
@
s<0 @
@
@

The case of d + 1 dimensions

.
Although our main interest is in 3+1 dimensions, it is instructive to look at
the general case of d + 1. Causality means that Gd+1> lives in the forward light
cone. Since the source terms is isotropic let us seek a solution of the form

Gd+1
> (x) = θ(t)f (s), s = xµ xµ (8.31)

We need to ascertain that G> solves the homogeneous wave equation for t > 0.
To do so, let us compute the two derivatives in the Dalambertian one at a time:

∂µ f (s) = (∂µ s) f 0 (s) = 2xµ f 0 (s) (8.32)

It follows
1 µ
2 ∂ ∂µ f (s) = ∂ µ (xµ f 0 (s))
= (∂ µ xµ )f 0 (s) + 2(xµ xµ ) f 00 (s)
= (d + 1)f 0 (s) + 2s f 00 (s)
= (d − 1)f 0 + 2(s f 0 )0

Hence

∂ µ ∂µ f = 2(d − 1)f 0 + 4(s f 0 )0 = 2(d − 3)f 0 + 4(sf )00

8.6. GREEN’S FUNCTION FOR THE WAVE EQUATION 143

This is a simple, second order ordinary differential equation which can be inte-
grated directly to give
(d − 1)f + 2(s f 0 ) = c1
To compute the constant consider the left hand side outside of the forward light
cone. Since we insist on causal solutions both f and f 0 must vanish outside the
light cone.
Exercise 8.20. Why?
This says that c1 = 0. We are left with a first order equation
(d − 1)f + 2s f 0 = (d − 3)f + 2(sf )0 = 0 (8.33)
This simple differential equation can be integrated in any dimension d.
Exercise 8.21 (Regular solutions). Show that the regular solutions of the dif-
ferential equation 8.33 are (the real and imaginary parts of )
1−d
f (s) = c2 (−s) 2 d 6= 1,
The solution found in the exercise distinguished d even and d odd. For even
d it looks ok for the (real part) lives in the forward light cone. However, for
odd d the solution is not causal: Causality forces c2 = 0 and we are left empty
handed. What have we missed? We missed solutions that are distributions.
Example 8.22. In 2+1 dimensions
 
1
G2+1
> = λ2 θ(t) Re √
−s
is causal but does not satisfy the Huygens principle.

Back to 3 + 1
Let us now focus on the case d = 3. The differential equation 8.33 reduces to
(sf )0 = 0
which is easily integrated to
sf (s) = c2
with c2 the integration constant above. We use causality to conclude that
c1 2 = 0 and we are left with
sf (s) = 0
If you treat this as an equation for a function f then f = 0. Have we done
something wrong?
The subtle point here is that f need not be a function. After all the source
was a distribution so we should allow also distributional solutions. Now we can
find a non-trivial solution, namely
sδ(s) = 0
144 CHAPTER 8. ELECTROMAGNETIC WAVES

We conclude that
G3+1
> (x) = λθ(t)δ(s)
and Huygens principle holds.
Exercise 8.23 (Solution in 1 + 1). Using the similar arguments show that
G1+1
> (x) = λ1 θ(t)θ(−s)

The solution lives in the forward light-cone rather than its boundary. It does not
satisfy Huygens principle.
It remains to determine λ3 in 3+1 dimensions. We do that by Gauss law.
We integrate the wave equation with a source at the origin of a space-time box
Ω 2 as in fig. 8.6
Z Z
G> dΩ = 4π δ (4) (x)dΩ = 4π (8.34)
Ω Ω
The left-hand side can be converted to an integral on the boundary of the box
Z Z
G> dΩ = (∂µ G> ) dS µ
Ω ∂Ω

Exercise 8.24 (Signs-Sigh). On the early and late face of the box, determine
the sign in dS 0 = ±dV .

t
@ 6
@
@
T
' @ $
@
@
@
x
@ -
@
& %
@
@
@
@
@
@
Inspecting the figure one realizes that only the intersection of the top face of
the box with the forward light cone contributes. Everywhere else the retarded
solution vanishes. Using dS 0 = dV and Eq. 8.32
Z Z Z
µ
(∂µ G> ) dS = dV ∂0 G> = λ3 dV (−2cT ) δ 0 (s)
∂Ω t=T t=T
2 Minkowski cartesian coordinates with |g| = 1 are assumed.
8.6. GREEN’S FUNCTION FOR THE WAVE EQUATION 145

With s = r2 − c2 T 2
Z Z ∞
dδ dr
dV δ 0 (s) = 4π dr r2
t=T dr ds
Z0 ∞
dδ
= 2π dr r
0 dr
Z ∞  
d(rδ)
= 2π dr −δ
0 dr
Z
= −2π drδ(r2 − c2 T 2 )
π
=−
cT
This and Eq.8.34 fixes λ3
π
λ3 (2cT ) = 2λ3 π = 4π =⇒ λ3 = 2
cT
Exercise 8.25. Compute the normalization constants λ1+1 and λ2+1 .
Bibliography
F. John, Partial differential equations,
Y. Pinchover and Y. Rubinstein, —Introduction to Partial Differential Equa-
tions, Cambridge
146 CHAPTER 8. ELECTROMAGNETIC WAVES
Chapter 9

Radiation
Retarded solution of the wave equation. Retarded potentials, retarded electro-
magnetic fields. Lienard Wiechert potentials and fields; Dipole radiation.

9.1 Wave equation with a source term

We have seen that the retarded Green function for the wave equation in 3+1
dimensions is

δ(ct − r)
G> (x) = 2 θ(x0 )δ x · x = , r = |x| ≥ 0 (9.1)
r
By the homogeneity of Minkowski space, for source term located at the space-
time point y rather than the origin, the Green function is G> (x − y)

9.1.1 Retarded waves for an arbitrary source

By the linearity of the wave equation, and the homogeneity of Minkowski space-
time, the retarded solution of the wave equation, generated by an arbitrary1
source ρ
φ = −4πρ(x) (9.2)
is Z
φρ (x) = d4 y G> (x − y)ρ(y) (9.3)

9.1.2 Application: Lorenz gauge

Let us show that one can always impose the Lorenz2 gauge condition
∂ µ Aµ = 0 (9.4)
1 Some condition on the localization of the sources should be imposed. This is related to
Olber’s paradox: If you assume constant density of stars, and that intensity of radiation falls
like r−2 the night sky shoudl be as bright as the sun.
2 This is the Danish Ludvig Lorenz, a contemporary of the Dutch, Nobel laureate Hendrik

Lorentz of the Lorentz transformation.

147
148 CHAPTER 9. RADIATION

Suppose ∂ µ Aµ 6= 0. Let Λ be the (retarded) solution of the wave equation with

a source term
Λ = −∂ µ Aµ (9.5)
Then, the gauge transformation,
A0µ = Aµ + ∂µ Λ (9.6)
satisfies the Lorenz gauge condition:
0 = ∂ µ A0µ = ∂ µ Aµ + ∂ µ ∂µ Λ (9.7)

9.1.3 Scalar wave generated by a moving point source

As a preparation for studying the radiation of electromagnetic waves, consider
the simpler problem of radiation of scalar

waves generated by a point source
moving on a world line z = ct, z(ct) , −∞ < t < ∞. The motion is assumed
to be that of a real particle so the velocity is time-like. The source density in
space time is
ρ(y, t) = δ (3) y − z(t)


(9.8)

t
zHtL

x0
y0

Figure 9.1: The world like z(t) of the point charge is the blue line. The wave at
the point x is determined by the intersection of the backward light cone with
the orbit at time y 0 . There is one such point since the velocity is time like.

The retarded wave that the source generates is (y 0 = ct):

Z
φρ (x) = 2 d4 y δ (3) y − z(y 0 ) δ (x − y) · (x − y) θ(x0 − y 0 )



| {z }| {z }
source Green
Z
= 2 dy 0 δ R · R θ(x0 − y 0 ), R = (x0 − y 0 , x − z(y 0 ))


9.2. MAXWELL EQUATION IN THE LORENZ GAUGE 149

The θ function guarantees causality: The past influences the present. The delta
function says that signals propagate with the speed of light–Huygens principle
holds. If the orbit z(z 0 ) is that of a physical particle, a single point contributes:
a single particle has a single image (in the absence of mirrors) .
To compute the remaining integral use
Z

1
δ f (y) dy = 0 , f (y0 ) = 0. (9.9)
|f (y0 )|
Here f = R · R whose derivative is related to the the 4-velocity of the source
u = −Ṙ:   
df df dτ R · Ṙ R·u
= =2 = −2
dy 0 dτ dy 0 γ γ
We obtain a simple looking formula for the wave φ(x) at the observing point x:
γ(y 0 )
φ(x) = , R = x − y, R·R=0 (9.10)
|R · u(y 0 )|
which is manifestly causal and satisfies Huygens principle.
The formula looks simple but at the price of being implicit. The right hand
side is not an explicit function of the argument x. To compute γ(y 0 ), u(y 0 ) and
R = x − y on the right hand side you need first to evaluated at the earlier time
y 0 (see Fig. 9.1.3), which is not explicitly given. This time is determined as the
solution of the equation

2
(x0 − y 0 )2 = x − z(y 0 )
which may be arbitrarily complicated if the orbit z(z 0 ) is complicated.
As expected, the amplitude of scalar waves decays like 1/R. However, you
may argue if you want to call the general solution we have found a wave in all
cases. For example you would probably not call the solution for a source at rest
a wave.

9.2 Maxwell equation in the Lorenz gauge

The inhomogeneous Maxwell equations are
4π
∂ µ Fµν = − jν (9.11)
c
Expressed in terms of the potentials, (this guarantees the homogeneous equa-
tions) one gets a system of second order PDE
4π
∂ µ µ Aν − ∂ µ ν Aµ = − jν (9.12)
c
In the Lorenz gauge, ∂ µ Aµ = 0, Maxwell equations reduce to 4 decoupled wave
equations
4π
Aν = ∂ µ µ Aν = − jν (9.13)
c
150 CHAPTER 9. RADIATION

The equations are coupled through the Lorenz gauge condition. If the current
j µ is not conserved then the derivation is inconsistent with the Lorentz gauge.
Conversely, if current is conserved, then Lorentz gauge condition follows for all
times provided the initial data for Aµ and Ȧµ satisfy it.
Exercise 9.1. Show that if one imposes the Lorentz gague condition as initial
data for the wave equation, then the Lorenz gauge condition holds for all times
provided current is conserved.

9.3 Lienard-Wiechert: Retarded potentials

Our aim here is to compute the potential Aµ generated by a point charge moving
with a given orbit. The Maxwell equations, Eqs. (9.13), can be viewed as 4
independent copies of the scalar wave equation with given source terms. We
can therefore transcribe the solution from the previous section to this case. The
source is now


(3) 0 0 δ (3) y − z(y 0 )
uν (y 0 )


jν (y) = eδ y − z(y ) vν (y ) = e (9.14)
| {z } γ | {z }
velocity 4−velocity

Comparing with Eq. (10.6) for scalar waves we see that the retarded potentials
are:
uν uν (y 0 )
Aν (x) = e = −e (9.15)
|R · u| R · u(y 0 )
where removed the absolute value by taking into account that R is forward
light-like and u forward time-like so R · u < 0.
The result admits the following interpretation: The vector potential, being
a 4-vector, must be of the form
(scalar)(vector)µ
We have (at least) two 4-vectors at our disposal: R and u. Between these
we can form 3 scalars: Two uninteresting u · u = −c2 and R · R = 0and one
interesting R · u. This, plus dimension analysis and the limit case of a charge
at rest determines Eq. (9.1).
The result can also be viewed as the covariant form of Coulomb law:
e e uν
Aν (x) = (1, 0, 0, 0) = − (−c, 0, 0, 0) =⇒ −e
|x| c|x| | {z } R·u
uν

9.3.1 The Lorenz Gauge condition

We still need to verify the Lorenz gauge condition.
A clever argument: Since the condition is Lorentz invariant, it is sufficient
to verify it in some Lorentz frame. So let us do that in the frame where the
charge is instantaneously at rest at the early time. Then,
∂ µ Aµ = ∂ 0 A0
9.3. LIENARD-WIECHERT: RETARDED POTENTIALS 151

with
e
A0 = −
|x|
when you change t of the event x the distance x does not change because the
particle is at rest. Hence
∂ 0 A0 = 0

A sneaky argument: Suppose in the distant past the source was a rest. The
coulomb solution then satisfies the Lorenz gauge in the past. By conservation
of charge, the Lorenz gauge is preserved by the evolution.
An honest Computation: The reason for doing also an honest computation
is that this will force us to derive the identity that describes how the retarded
time depends upon variation of the observation event x:

Rµ
∂µ τ = (9.16)
R·u
This follows by differentiating R · R = 0

0 = 21 ∂µ (R · R) = Rα ∂µ (xα − z α ) = Rµ − R · u (∂µ τ )

Back to the honest verification of the Lorenz gauge condition:

uµ
 
µ
0 = ∂µ A = −e∂µ
R·u

This will hold provided

?
uµ
  z}|{
2 µ µ
(R · u) ∂µ = (R · u) ∂µ u − u ∂µ (R · u) = 0 (9.17)
R·u

To verify that this is indeed so, let us prepare

∂µ uα = u̇α (∂µ τ ), ∂µ Rα = δµα − uα (∂µ τ )

Substituting this in Eq. (9.17) and using Eq. (9.16) we find

(R · u) ∂µ uµ − uµ ∂µ (R · u) = (R · u) u̇µ (∂µ τ ) − uµ uα ∂µ Rα − uµ Rα u̇α (∂µ τ )

u̇ · R Rµ
= (R · u) − uµ uα ∂µ Rα − uµ Rα u̇α
R·u R·u
= −uµ uα ∂µ Rα
= −u · u + uµ ∂µ τ
= −u · u + u· u = 0

We have verified that the solution indeed satisfies the Lorenz condition.
152 CHAPTER 9. RADIATION

Figure 9.2: The self time τ parametrizes the blue orbit. It can be extended to
a function on space time by pushing the value of τ to the forward light-cone.
The figure illustrates how τ changes when the point of observation x changes.
The red lines are light-like.

9.4 Retarded Fields

In the Lorenz gauge the potentials we found are causal. This implies the causal-
ity of the fields. We want now to derive direct formulas for the fields.
To find the fields we need to differentiate the potentials with respect to the
space-time coordinates xµ . Formally,

Fµν = ∂µ Aν − ∂ν Aµ = ∂[µ Aν] (9.18)

and the right hand side is a convenient notation. The word formal above refers
to the fact that in taking the partial derivatives we need to remember that τ ,
the retarded time, is a function of the point of observation x. If we want to
treat x and τ as independent variables, ∂µ needs to be interpreted as
   
∂ ∂τ ∂ ∂ Rµ ∂
∂µ =⇒ + = + (9.19)
∂xµ ∂xµ ∂τ ∂xµ R·u ∂τ
9.4. RETARDED FIELDS 153

The xµ differentiation sees only the first term in R, the location of the observer,
while the τ differentiation only sees the second term, the location of the charge.
Using the explicit form of the potential:
 u   
ν uµ uν Rµ  u 
ν
− ∂µ Aν = e ∂µ = −e 2
+ e ∂τ (9.20)
R·u (R · u) R·u R·u
Because F is anti-symmetric, the first term in Eq. (9.20) drops upon anti-
symmetrization and only the second term contributes
The field F depend on the location, velocity u and the acceleration u̇ of the
charge at the early time. It does not depend on any higher derivatives, e.g. the
jerk ü. Now compute:
 u  u̇ν uν
ν
∂τ = − ∂τ (R · u) (9.21)
R·u R · u (R · u)2
u̇ν uν uν
= + 2
u·u− R · u̇
R·u (R · u) (R · u)2
u̇ν uν uν
= − c2 − R · u̇
R·u (R · u)2 (R · u)2
Consequently
 
Rµ  u 
ν Rµ u̇ν Rµ uν Rµ uν
∂τ = − c2 − R · u̇ (9.22)
R·u R·u (R · u)2 (R · u)3 (R · u)3
We get F by anti-symmetrizing:
 
R[µ u̇ν] R[µ uν] R[µ uν]
Fµν = −e 2
− 3
R · u̇ +e c2 (9.23)
(R · u) (R · u) (R · u)3
| {z } | {z }
radiation ”Coulomb”

Since u is normalized to c the last term, is order O(c0 ). It decays with distance
like R−2 . This is, essentially, the Coulomb term. The first two terms are
proportional to the acceleration and so formally of order O(c−2 ). They decay
more slowly at large distance, like R−1 . These are the radiating terms.

9.4.1 Interpretation
The formula is complicated and at first also opaque. It may be useful to view
it from general principles.
We have three vectors in the problem: R, the light-like vector connecting
the point of observation and the source, u the particle 4-velocity and u̇ its
4-acceleration. From these we can make three interesting scalars

R · u, R · u̇, u̇ · u̇

The remaining scalars are not interesting

R · R = 0, u · u = −c2 , u · u̇ = 0 (9.24)
154 CHAPTER 9. RADIATION

F must be linear in the acceleration. This is because the potential did not
depend on the acceleration at all. As a consequence, that the scalar u̇ · u̇ should
not appear and the scalar R · u̇ can only appear in the numerator. Given this
we can reconstruct all the three terms in F , up to sign, just by the fact that F
is a tensor and dimension analysis. From the tensorial properties of F it must
be of the form
(tensor)µν = (scalar) (vector)µ (vector)ν
Since F has dimension of [charge][length−2 ] and u has the dimension of c,
one possible term is
R[µ uν]
ec2
(R · u)3
which gives the last term in Eq. (10.5). You can even get the numerical factor
(and the sign) by looking at the limiting case of the Coulomb field of a particle
at rest where uµ = (−c, 0, 0, 0).
Looking at the terms proportional to the acceleration u̇ one possibility is
R[µ u̇ν]
e
(R · u)2

which gives the first term up to sign and numerical factors. The middle term is
obtained similarly.

9.5 Particle instantaneously at rest

The formula for F simplifies for a particle instantaneously at rest at the origin
(at the early time), i.e. y 0 = 0: uµ = (c, 0, 0, 0) and R · u = −c|x|. This
determines F on the forward light-cone (|x|, x).

9.5.1 The Magnetic field:

Consider first the magnetic field. Since the spatial components of u vanish, we
have (in Cartesian coordinates)

Fij (|x|, x) = εijk B k (|x|, x)

 =0  =0
z }| { z }| {
 R[i u̇j] R[i uj]  2 R[i uj]
= −e 
 (R · u)2 − (R · u)3 R · u̇ + ec (R · u)3


e R[i u̇j]
= −
c2 |x|2
e (a(0) × x)k
= (9.25)
c2 |x|2

where a, the 3-vector of acceleration u̇µ = (0, a), is orthogonal to u = (c, 0).
9.5. PARTICLE INSTANTANEOUSLY AT REST 155

It follows that, 3 vector of magnetic field on the light-cone emanating from

teh origin, (|x|, x) is
e a(0) × x
B(|x|, x) = 2 (9.26)
c |x|2
The main conclusions we draw from this are
1. The field decays like the inverse distance
2. The field is perpendicular to both the line of sight and the acceleration
vector: there is no radiation in the direction of acceleration.
3. You can use the homogeneity of space-time to write a slightly more com-
plicated formula for the case where the particle is at rest at an arbitrary
point.

9.5.2 The electric field

Recall that F0j = −Ej . For a particle at rest at the origin at lab time y 0 = 0
R · u = −c|x| and R0 = −|x|. Hence, on the light-cone emanating from the
origin
 
R[0 u̇j] R[0 uj] R[0 uj]
Ej = −F0j = e 2
− 3
R · u̇ − ec2
(R · u) (R · u) (R · u)3
 
e 1 e
= 2 3
|x| R[0 u̇j] − R[0 uj] R · u̇ + R[0 uj]
c |x| c c|x|3
e e
= 2 3
(−(x · x) aj + (x · a) xj ) + 2 xj
c |x| r
Since

(x · x)a − (x · a) x = − x × (x × a)
We can collect the above to a vector identity
e
e
E(|x|, x) = x̂ × (x̂ × a) + x̂
c2 |x| |x|2
e
= B(|x|, x) × x̂ + x̂ (9.27)
|x|2
The longitudinal part is Coulomb and the transversal part is radiation. E and
B are mutually orthogonal. The two Lorentz scalars are

e2
E · B = 0, E2 − B2 = −
|x|4
Exercise 9.2. Show that the Poynting vector is

c e2
E×B= x̂
4π 4πc3 |x|2
156 CHAPTER 9. RADIATION

r/c

t-r/c
t-r/c+r.z/c

Figure 9.3: A space time diagram showing the geometric meaning of the succes-
sive approximations leading to Eq. (9.36). The world-line of the moving charge
is the blue path. It is confined to a narrow corridor between the two red lines
a distance 2` apart. The retardation time is the intersection of the back light
cone (green line) with the blue orbit. The first approximation t − r/c is the
intersection of the back light cone with the origin (gray vertical line). The next
correction in Eq. (9.36) is the corner of the small green square.

9.5.3 Slow particles

We found the electric and magnetic field of a particle at rest at the origin. If
the particle is at the origin but not at rest, we can get the fields by making a
Lorentz transformation. In particular, if the particle is slow the field is
e a(x0 − |x|) × x  v 
B(x) = 1 + O (9.28)
c2 |x|2 c

(We have use the homogeneity of Minkowski space time to shift y 0 = 0 to any
other time.) and similarly for E.

9.6 Retardation from a distant source

To compute the Fµν we need to compute the retardation: x0 − y 0 . This time is
determined as the solution of the equation

x0 − y 0 = |x − z(y 0 )|
9.6. RETARDATION FROM A DISTANT SOURCE 157

where z(y 0 ) is the obit of the charge. What we seek is an explicit, approximate,
formula for y 0 (x).
In general, there is little one can say for arbitrary xµ and general orbits
0
z(y ). In practice, one often is interested in the field far from the source: The
light from a star; from radiating atom. In these cases the observer is far from
the source:|x|  |z|. Assume the orbit z is confined to a ball of radius `  |x|.
Under this condition we can write
 2 !
2 2 2 2 x·z `
|x − z| = |x| + |z| − 2x · z = |x| 1 − 2 2 + O
|x| |x|

Hence the implicit equation for the retardation for a distant source reduces to:
 2
0 0 0 `
x − y = |x| − x̂ · z(y ) + O (9.29)
| {z } |{z} | {z } |x|
large O(x) O(x) O(`) | {z }
negligible

Although this is still an implicit equation, which is in general, as difficult as the

original one, it is good starting point for getting explicit approximate solution
when one has additional dimensionless parameters.
How accurately do we actually need to compute y 0 ? The answer depends on
the characteristic frequency of the source. The accuracy we need is such that
it allows to locate the source to better than the characteristic wave length. To
see this consider an oscillating charge

z(y 0 ) = ` n̂ cos(ky 0 )

So, if
ky 0 = |123456789
{z } .987654321 × (2π)
irrelevant

and we want an accuracy of 1%, then the all we care are the three digits in blue,
just after the decimal point. In other words, cos is a periodic function, we need
only the fractional part of ky 0 . In particular we need to compute the fractional
part of the retardation. We can now write Eq. (9.29) in dimensionless form
 2
0 0 0 k`
k(x − z ) = k|x| − kx̂ · z(z ) + O (9.30)
| {z } |{z} | {z } |x|
1 1 O(k`) | {z }
negligible

Whether the term O(k`) can be neglected or not does not depend anymore on
how far the source is but rather on how big k` is.

9.6.1 The dipole approximation: k`  1

The dipole, approximation is concerned with the case k`  1: the wave length
of the emitted radiation is much larger than the size of the source. This is, for
158 CHAPTER 9. RADIATION

example, the case for the light emitted by a single atom where the characteristic
wave length is thousands of times larger than the atom. In this case the adequate
approximation it to keep only the big terms in Eq. (9.30) to get the an explicit,
approximate, expression for the early-time y 0 :

y 0 (x) = x0 − |x| (9.31)

The dipole approximation describes small antennas.

Harmonic motion
As an application consider a charge undergoing non-relativistic harmonic motion
with acceleration
0
a(y 0 ) = a0 eiky (9.32)
√
a is a vector with complex amplitudes. Thus, for example ao = `(1, i, 0)/ 2
describes circular motion with radius ` in the plane. Non-relativistic means
ω`  c ⇔ k`  1 which is the condition for the dipole approximation to apply.
This sets
ky 0 = k(x0 − r), r = |x|
Using Eq. (9.28) for the magnetic field of a charge moving non-relativistically
we get
0
eik(x −r)
 
e a0 × x̂
B(x) = (9.33)
c2 | {z r }
spehrical wave

From this point of view, an outgoing spherical wave is a consequence of retar-

dation.

Many particles
The radiation fields from many particles with prescribed orbits is, by linear-
ity of the Maxwell equation, the sum of the radiation of the individuals ones.
In general, each particle will have its own retarded time, and the formulas re-
main implicit. A simplification occurs in the dipole approximation for “ small
antenna” where all the charges share the same retardation.
The dipole moment of a large collection of charges is:
X
d(t) = ej zj (t) (9.34)

and we assume that all the orbits zj are such that the dipole approximation
applies. In this case all the charges have the same retardation and the magnetic
field is simple
d̈(t − r/c) × x̂
B(r, t) = (9.35)
c2 |x|
9.7. POWER 159

9.6.2 Beyond dipole: Slow motion

The dipole approximation applies provided k`  1. In the case that the charges
move slowly, one can go beyond the dipole approximation and get explicit ap-
proximation for the retardation that apply when k` = O(1) This is often prac-
tically important: Your cell phone works at ω ≈ 2 [GHz] associated with wave-
length 15 [cm] which is comparable to the cell phone antenna. When k` = O(1)
one needs to take the O(k`) term in Eq. (9.30) into account. In general, this
gives, once again an intractable implicit equation which depends on the details
of the motion z(y 0 ). We can get an explicit simple formula in the special case
that the charge moves slowly |ż| = v  c. This is illustrated in the Fig. 9.3.
Since the world-line is almost vertical, we may solve Eq. (9.30) iteratively,
y 0 (x) = x0 − |x| + x̂ · z(x0 − |x|) (9.36)
(Eq. (9.31) has been substituted in Eq. 9.29). It should be clear that the error
we make is proportional to v/c  1.
,

9.7 Power

B
E
a

Figure 9.4: Dipole oriented along the z-axis and the associated fields. φ is the
angle between the z-axis and the blue arrrow.

The power emitted by dipole can be computed from Poynting

E×B
P=c (9.37)
4π
160 CHAPTER 9. RADIATION

Both E and B lie in the plane perpendicular to the line of sight so P is parallel
to r̂. Suppose that a = aẑ. The magnitude of P in the direction φ relative to
the z-axis is
E2 e2 a2 sin2 φ
P (φ) = c = (9.38)
4π 4πc3 r2
The power through a spherical shell of radius r is then
Z
2
PT = 2πr dφ sin φ P (φ) (9.39)

Evidently
Z Z

4
dφ sin φ sin2 φ = − d(cos φ) (1 − cos2 φ) = 2 1 − 13 = (9.40)
3

Hence, the total power

2 e2 a2 2 d̈2
P = = (9.41)
3 c3 3 c3

Figure 9.5: Polar plot of the power radiated by a dipole antenna as function
of the angle. The maximal power is radiated in the plane perpendicular to the
dipole.

A dipole antenna is not isotropic: It does not radiate at all in the directions
of the dipole.

Remark 9.3. You can not make an isotropic antenna. No matter how com-
plicated an antenna you make the Poynting vector must vanish it at east two
directions. This is a consequence of topology: The vector B is tangent to the
sphere. It is a fact that every vector field tangent to the sphere must vanish at
two points, at least. This is sometimes expressed as you can not comb a tennis
ball. Hence P must vanish at two points at least.

9.8 Classical instability of atoms

I now want to explain a puzzle in classical electrodynamics that turned out to
be a window that opened the way to quantum mechanics. In classical physics
atoms are unstable, and should collapse in no time by radiation.
9.9. ORIENTING AN ANTENNA 161

Consider a charge e in Keplerian orbit around a nucleus of charge e in a

circular orbit. The energy (non-relativistic) of the system is

1 e2
E=−
2 |x|

while the acceleration is

e2
a=
m|x|2
The rate of loss of energy by radiation is

2 e2 2
−Ė = a
3 c3
We can now eliminate a and obtain a differential equation for the energy

25
Ė = −kE 4 , k=
3(me)2 c3

Integrating the differential equation we find a blow up at finite time:

E(t) = E0 (1 − γt)−1/3 , γ = −3E03 k > 0

In other words, the charge would collapse on the nucleus in finite time 1/γ.
The ground state energy of hydrogen-like atom is,
2
e2

2
2E0 = −mc = −mc2 α2
~c

and its period is 2π~/E0 . Therefore, the decay time, counted in periods, is

E0 1
= × α−3 = 6400
2π~γ 128π
This means that the electron in hydrogen would fall on the on the proton in
2 × 10−12 seconds. The classical world is unprotected against collapse to the
nucleus.
The apparent instability of the atoms in classical physics is one of the reasons
for quantum mechanics.

9.9 Orienting an antenna

9.10 Reciprocity
162 CHAPTER 9. RADIATION

E
t

Figure 9.6: Blowup at finite time: The energy of a charged particle encircling
the nucleus goes to −∞ in finite time.
Chapter 10

Radiation reaction
Electrodynamics is a practical theory that proved itself in huge number of appli-
cation. It is a consistent theory when the sources are continuous distributions.
However, it is not a self-consistent fundamental theory of point particles.

10.1 Interacting systems

In studying interacting systems one needs to solve Newton equations for the
motion of charged particles in a given electromagnetic field and simultaneously
solve Maxwell equations for a given motion of the sources. Newton equations
are non-linear and Maxwell equations are partial differential equations. One
therefore ends up with a set of non-linear partial differential equation. This is
a hard problem with few general methods other than simulations.

10.1.1 Non-relativistic interacting particles

The theory of interacting point charges simplifies in the limit of slowly moving
point charges. This limit is perfectly consistent and practically useful. Let us
discuss this first. Slow particles interact weakly with magnetic fields, and the
forces is dominated by Coulomb forces:
e
mu̇µ + Fµν uν = 0
c
If the particles move sufficiently slowly, uµ ≈ −cδ0µ , and the charges are affected
mostly by the electric field given by
1
E = −∇Φ − Ȧ
c
In the Coulomb gauge, the scalar potential Φ is the instantaneous solution of
Poisson equation while A, (with ∇ · A = 0), solves the wave equation with a
source term
1 4π 1
−∆A − 2 Ä = J − ∇Φ̇
c |c {z c }
negligible

163
164 CHAPTER 10. RADIATION REACTION

which is small when the particles move slowly. The problem of interacting
charges is then encapsulated by the Hamiltonian
X 1 1 X ej ek
H= vj2 +
2mj 2 |xj − xk |

The electromagnetic field has been integrated out of the problem. This is the
starting point of musch of atomic physics.

10.1.2 Plasma and Magnetohydrodynamics

In plasma physics and magnetohydrodynamics Newton equations take the form
of the partial differential equations similar to those of fluid mechanics1 . Com-
bined with Maxwell equations, one gets a theory of coupled, non-linear partial
differential equations. The theory is technically hard, but conceptually is fine.
As a consequence there are relatively simple phenomena in plasma physics,
that are only poorly understoon. For example the current understanding of the
spontaneous generation of the magnetic field of the earth–driven by the earth
rotation–is largely based on computer simulations.

10.1.3 Infinities
Electrodynamics of point-like charges is not a fully consistent theory. The source
of trouble can already be seen by looking at the the field energy of a point charge:

e2
Z
1
dx = ∞
8π |x|4

The divergence comes from the singularity at x = 0. (The integral at infinity

is perfectly convergent in 3 spatial dimensions). Modern physics tinkers with
the dimension of space time, and allows for elementary objects that are strings
and branes, these affect the field energy. It may be that this divergence is the
window for a future (string) theory free from such divergences.

10.2 Radiation reaction

Consider a charged particle moving non-relativistically in a circle due to the
action of a central force. We allow for non-electromagnetic forces that make
sure that the particle moves in a stationary orbit. The accelerating charge
radiates and we want to compute the force that acts back on the charge because
of this process. We have a formula for the power radiated by the charge and we
want to convert the information in the power to information about the force.
For circular the acceleration is perpendicular to the velocity and we have

a2 = (a ·˙ v) − ȧ · v = −ȧ · v
1 Thermodynamics needs to be thrown in to close the equations.
10.2. RADIATION REACTION 165

The radiated power can now be written as

2 e2 2 2 e2
P = a = − ȧ · v = −FAL · v
3 c3 3 c3
If the particle is moving at constant speed, an opposite (non-electromagnetic)
force must be applied to feed back the radiated energy. This means that the
radiation acts back on the charge and produces a force

2 e2
FAL = ȧ
3 c3
The jerk ȧ is opposite to the velocity, Fig 10.1. The force acts like a friction
force. −FAL is then the non-electromagnetic force that we need to apply to keep
the orbit stationary. It is known as the Abraham-Lorentz force. It is a somewhat
unusual force, in two ways. Unlike the usual friction it is not proportional to the
velocity but rather to the jerk. Second, it is ultimately a force that a particle
applies on itself. Something Newton would not approve of.
v a 

Figure 10.1: For a cicrular motion every derivative is a rotation by π/2

Exercise 10.1 (Covariant form of Abraham Lorentz force). Using the fact that
Newton law
maµ = fµ
is consistent with u · u = −c2 provided f · u = 0 and a · a + u · ȧ = 0, show that
the covariant form of Abaraham-Lorentz force is

2e2  a·a 
maµ = ȧ µ − uµ
3c3 c2

10.2.1 When is radiation reaction important?

In MKS units
e2 e2
⇐⇒ = 8 × 10−54 [Kg s]
c3
|{z} 4πε0 c3
| {z }
Gaussian M KS
166 CHAPTER 10. RADIATION REACTION

a small number. You’d need huge jerk to get tiny forces.

If ω is the characteristic frequency of the problem, the radiation reaction
force becomes comparable to the inertia at frequencies:

mc3
ω
e2
If you take for m the mass and charge of the electron, then the left hand side
is of the order of 1023 [Hz]. For most practical purposes, radiation reaction is
negligible.

10.2.2 Friction
Small forces can still do something if they act for long time. Consider the
equation of motion in a force field f with radiation reaction like friction:

2 e2
ma = f (x) + k ȧ, k= (10.1)
3 c3
From a conceptual point of view, this is a major modification of Newton law,
since the order of the equation changed: It is not enough to specify initial
conditions in the form of position and velocity, one also needs to specify the
initial acceleration. This is contrary to much evidence we have.
However, since k is so small, a reasonable strategy for interpreting the equa-
tion is look at it iteratively. Namely, replace the equation by
k k
ma = f (x) + ḟ = f (x) + (v · ∇)f
m m
In this form the equation of motion is still second order and the radiation reac-
tion looks like a friction term.
As an example consider the harmonic oscillator where f (x) = −κx. The
equation of motion is now
κk
ma = −κx − v
m
which is solved by x(t) = x(0)e±iω± t where ω is a solution of the quadratic
equation
2 κ κk
ω± − ∓ iω± 2 = 0
m m
Since k is very small, this is solved by
r 
κ κ
ω± = 1 ± ik
m m
This describes reasonable solutions of slowly decaying oscillations.
Exercise 10.2. Write the equations of motion for the Kepler problem with
radiation reaction friction term.
10.2. RADIATION REACTION 167

10.2.3 The Dumbbell

A more careful derivation of the radiation reaction force starts with the forces
an ex extended object of size ε. In this cased different parts of the object apply
mutual forces that add up to the radiation reaction force. In the limit ε → 0
one recovers Abaraham Lorentz with an extra bonus: The mass is related to
the energy of the field.
We shall re-derive the radiation reaction force thinking of a point charge as
the ε → 0 of an object of size ε.
We want to compute the forces that the dumbbell applies on itself when it
moves in a prescribed way. The world-lines of the dumbbell are shown in the
figure (red) and the light-cone is drawn blue. We shall show that Newton third
law is violated, so there is a net force acting on an extended body.





hhhh 
 hh
  x
 










The dumbbell is made of two point charges separated by a rod of length ε.

The dumbbell moves along the x-axis and is aligned with the y-axis. So the
world line of the two charges is

x± (t) = (ct, q(t), ±ε/2, 0)

The two charges communicate when x± (t + τ ) − x∓ (t) is light like. The time
delay τ = O(ε) since the dumbbell is small.
R± (τ ) = x± (τ ) − x∓ (0) is a light like vector. We choose a Lorentz frame so
that the dumbbell is at rest at time zero, i.e. q(0) = q̇(0) = 0 and so

R± (τ ) = (cτ, q(τ ), ±ε, 0) (10.2)

168 CHAPTER 10. RADIATION REACTION

We want to find the mutual forces on the dumbbell at time τ and express
them in terms of the acceleration and jerk at time τ , not at time 0. This gives a
funny situation where we specify the position and velocity at time 0 but specify
the acceleration at time τ . (We take the jerk to be constant.)
Let a and ȧ be the acceleration and jerk at time τ . Clearly
a(τ ) = a = a(0) + ȧτ
From this it follows that
q̇(τ ) = a(0)τ + 21 ȧτ 2 = aτ − 21 ȧτ 2 (10.3)
Integrating that gives the position
q(τ ) = 12 a(0)τ 2 + 16 ȧτ 3 = 12 aτ 2 + 1 1
ȧτ 3 = 12 aτ 2 − 31 ȧτ 3


6 − 2 (10.4)
This fixes the function q(τ ) in R± .
τ and ε are related by the condition that R is light like:
(cτ )2 = ε2 + q 2 (τ )
which is a polynomial equation for τ of order 6. However, as q(τ ) is quadratic
in τ to leading order
cτ ≈ ε
The retarded field at time τ is determined by the velocity and acceleration
of the particle at time 0:
 
R[µ aν] R[µ uν] 2 R[µ uν]
Fµν = −e − R·a−c (10.5)
(R · u)2 (R · u)3 (R · u)3
The four velocity

is uµ (0) = (−c, 0, 0, 0), and the four acceleration is aµ (0) =
0, a(0), 0, 0 . We therefore have
R · u = −c2 τ
The electric field in the direction of motion on one of the charges due to the
other at time τ is then
!
R0 a1 R1 u0 2 R1 u0
− Ex = F01 = −e + R · a +c
(c2 τ )2 (−c2 τ )3 (−c2 τ )3
| {z }
O(τ )

We are interested in the terms that do not vanish when cτ = ε → 0. Since

R1 = q(τ ) = O(τ 2 ) the middle terms in the formula above tends to zero with τ .
We drop it. The remaining terms are
 
e a(0) q(τ )
Ex = F10 = − +
c3 τ τ3
1 1
2 a − 3 ȧτ
 
e a − ȧτ
= − + (10.6)
c3 τ τ
 
e a 2
= − 3 − ȧ
c 2τ 3
10.2. RADIATION REACTION 169

and we dropped terms that scale to zero with τ

The retarded force that one charge applies on the other in the x direction is
e2 2e2
F = eE = − 3
a+ ȧ
|2τ{zc } 3c3
e.m. mass

Disregarding sel-forces, we can now write Newton law assuming bare dumbbell
mass mb , in an external force Fex and radiation self-force as
e2
   2
2e
mef a = mb + 3
a = ȧ + Fex
2τ c 3c3
The effective mass gets a contribution from the electromagnetic interaction.
This is very nice except that as we send τ → 0 we need to take the bare mass
large and negative.
Remark 10.3. One can get the self radiation reaction using the following trick
Amos Ori taught me. Let f (e) be the self force. Then, in the limit τ → 0 the
total force on the dumbbell is
4ε2
Ft = 2f (e) + ȧ = f (2e) = 4f (e)
3c3
2e2
This says that the radiation reaction force is f (e) = 3c3 ȧ as we have seen before.
When is it important? Small forces are important in two cases. The first
is when they are all that there is and the second if they operate for very long
time.

10.2.4 Conceptual difficulties

Radiation reaction is conceptually problematic for several reasons. First, it
changes the order of Newton equations of motion from second order to third,
see Eq. (10.1). This means that fixing the initial position and velocity is not
sufficient to determine the orbit. This is in contrast with common experience.
Another problem is that the equations of motion , Eq. (10.1), admit non-
physical solutions. For example, with f = the equation is
2e2
ma = k ȧ , k=
3c3
It admit the solution
a(0)m kt/m
a(t) = a(0)ekt/m =⇒ v(t) = v(0) + e
k
A particle, initially at rest, self accelerate to large velocities. This can only be
avoided if you tune a(0) = 0 precisely. In practice, we only tune x(0) and v(0),
so how come we do not see these self-accelerations? The theory of self-force for
point particles, must be fundamentally flawed.